blob: 8c3dd645ac1083c8b3dd2abdeba4b96cec0000e2 [file] [log] [blame]
lh9ed821d2023-04-07 01:36:19 -07001/* 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
60struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61#define MEM_CGROUP_RECLAIM_RETRIES 5
62struct 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 */
66int do_swap_account __read_mostly;
67
68/* for remember boot option*/
69#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70static int really_do_swap_account __initdata = 1;
71#else
72static 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 */
83enum 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_STAT_NSTATS,
93};
94
95enum mem_cgroup_events_index {
96 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
97 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
98 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
99 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
100 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
101 MEM_CGROUP_EVENTS_NSTATS,
102};
103/*
104 * Per memcg event counter is incremented at every pagein/pageout. With THP,
105 * it will be incremated by the number of pages. This counter is used for
106 * for trigger some periodic events. This is straightforward and better
107 * than using jiffies etc. to handle periodic memcg event.
108 */
109enum mem_cgroup_events_target {
110 MEM_CGROUP_TARGET_THRESH,
111 MEM_CGROUP_TARGET_SOFTLIMIT,
112 MEM_CGROUP_TARGET_NUMAINFO,
113 MEM_CGROUP_NTARGETS,
114};
115#define THRESHOLDS_EVENTS_TARGET (128)
116#define SOFTLIMIT_EVENTS_TARGET (1024)
117#define NUMAINFO_EVENTS_TARGET (1024)
118
119struct mem_cgroup_stat_cpu {
120 long count[MEM_CGROUP_STAT_NSTATS];
121 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
122 unsigned long targets[MEM_CGROUP_NTARGETS];
123};
124
125struct mem_cgroup_reclaim_iter {
126 /* css_id of the last scanned hierarchy member */
127 int position;
128 /* scan generation, increased every round-trip */
129 unsigned int generation;
130};
131
132/*
133 * per-zone information in memory controller.
134 */
135struct mem_cgroup_per_zone {
136 struct lruvec lruvec;
137 unsigned long lru_size[NR_LRU_LISTS];
138
139 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
140
141 struct zone_reclaim_stat reclaim_stat;
142 struct rb_node tree_node; /* RB tree node */
143 unsigned long long usage_in_excess;/* Set to the value by which */
144 /* the soft limit is exceeded*/
145 bool on_tree;
146 struct mem_cgroup *memcg; /* Back pointer, we cannot */
147 /* use container_of */
148};
149
150struct mem_cgroup_per_node {
151 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
152};
153
154struct mem_cgroup_lru_info {
155 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
156};
157
158/*
159 * Cgroups above their limits are maintained in a RB-Tree, independent of
160 * their hierarchy representation
161 */
162
163struct mem_cgroup_tree_per_zone {
164 struct rb_root rb_root;
165 spinlock_t lock;
166};
167
168struct mem_cgroup_tree_per_node {
169 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
170};
171
172struct mem_cgroup_tree {
173 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
174};
175
176static struct mem_cgroup_tree soft_limit_tree __read_mostly;
177
178struct mem_cgroup_threshold {
179 struct eventfd_ctx *eventfd;
180 u64 threshold;
181};
182
183/* For threshold */
184struct mem_cgroup_threshold_ary {
185 /* An array index points to threshold just below usage. */
186 int current_threshold;
187 /* Size of entries[] */
188 unsigned int size;
189 /* Array of thresholds */
190 struct mem_cgroup_threshold entries[0];
191};
192
193struct mem_cgroup_thresholds {
194 /* Primary thresholds array */
195 struct mem_cgroup_threshold_ary *primary;
196 /*
197 * Spare threshold array.
198 * This is needed to make mem_cgroup_unregister_event() "never fail".
199 * It must be able to store at least primary->size - 1 entries.
200 */
201 struct mem_cgroup_threshold_ary *spare;
202};
203
204/* for OOM */
205struct mem_cgroup_eventfd_list {
206 struct list_head list;
207 struct eventfd_ctx *eventfd;
208};
209
210static void mem_cgroup_threshold(struct mem_cgroup *memcg);
211static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
212
213/*
214 * The memory controller data structure. The memory controller controls both
215 * page cache and RSS per cgroup. We would eventually like to provide
216 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
217 * to help the administrator determine what knobs to tune.
218 *
219 * TODO: Add a water mark for the memory controller. Reclaim will begin when
220 * we hit the water mark. May be even add a low water mark, such that
221 * no reclaim occurs from a cgroup at it's low water mark, this is
222 * a feature that will be implemented much later in the future.
223 */
224struct mem_cgroup {
225 struct cgroup_subsys_state css;
226 /*
227 * the counter to account for memory usage
228 */
229 struct res_counter res;
230
231 union {
232 /*
233 * the counter to account for mem+swap usage.
234 */
235 struct res_counter memsw;
236
237 /*
238 * rcu_freeing is used only when freeing struct mem_cgroup,
239 * so put it into a union to avoid wasting more memory.
240 * It must be disjoint from the css field. It could be
241 * in a union with the res field, but res plays a much
242 * larger part in mem_cgroup life than memsw, and might
243 * be of interest, even at time of free, when debugging.
244 * So share rcu_head with the less interesting memsw.
245 */
246 struct rcu_head rcu_freeing;
247 /*
248 * But when using vfree(), that cannot be done at
249 * interrupt time, so we must then queue the work.
250 */
251 struct work_struct work_freeing;
252 };
253
254 /*
255 * Per cgroup active and inactive list, similar to the
256 * per zone LRU lists.
257 */
258 struct mem_cgroup_lru_info info;
259 int last_scanned_node;
260#if MAX_NUMNODES > 1
261 nodemask_t scan_nodes;
262 atomic_t numainfo_events;
263 atomic_t numainfo_updating;
264#endif
265 /*
266 * Should the accounting and control be hierarchical, per subtree?
267 */
268 bool use_hierarchy;
269
270 bool oom_lock;
271 atomic_t under_oom;
272
273 atomic_t refcnt;
274
275 int swappiness;
276 /* OOM-Killer disable */
277 int oom_kill_disable;
278
279 /* set when res.limit == memsw.limit */
280 bool memsw_is_minimum;
281
282 /* protect arrays of thresholds */
283 struct mutex thresholds_lock;
284
285 /* thresholds for memory usage. RCU-protected */
286 struct mem_cgroup_thresholds thresholds;
287
288 /* thresholds for mem+swap usage. RCU-protected */
289 struct mem_cgroup_thresholds memsw_thresholds;
290
291 /* For oom notifier event fd */
292 struct list_head oom_notify;
293
294 /*
295 * Should we move charges of a task when a task is moved into this
296 * mem_cgroup ? And what type of charges should we move ?
297 */
298 unsigned long move_charge_at_immigrate;
299 /*
300 * set > 0 if pages under this cgroup are moving to other cgroup.
301 */
302 atomic_t moving_account;
303 /* taken only while moving_account > 0 */
304 spinlock_t move_lock;
305 /*
306 * percpu counter.
307 */
308 struct mem_cgroup_stat_cpu *stat;
309 /*
310 * used when a cpu is offlined or other synchronizations
311 * See mem_cgroup_read_stat().
312 */
313 struct mem_cgroup_stat_cpu nocpu_base;
314 spinlock_t pcp_counter_lock;
315
316#ifdef CONFIG_INET
317 struct tcp_memcontrol tcp_mem;
318#endif
319};
320
321/* Stuffs for move charges at task migration. */
322/*
323 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
324 * left-shifted bitmap of these types.
325 */
326enum move_type {
327 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
328 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
329 NR_MOVE_TYPE,
330};
331
332/* "mc" and its members are protected by cgroup_mutex */
333static struct move_charge_struct {
334 spinlock_t lock; /* for from, to */
335 struct mem_cgroup *from;
336 struct mem_cgroup *to;
337 unsigned long precharge;
338 unsigned long moved_charge;
339 unsigned long moved_swap;
340 struct task_struct *moving_task; /* a task moving charges */
341 wait_queue_head_t waitq; /* a waitq for other context */
342} mc = {
343 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
344 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
345};
346
347static bool move_anon(void)
348{
349 return test_bit(MOVE_CHARGE_TYPE_ANON,
350 &mc.to->move_charge_at_immigrate);
351}
352
353static bool move_file(void)
354{
355 return test_bit(MOVE_CHARGE_TYPE_FILE,
356 &mc.to->move_charge_at_immigrate);
357}
358
359/*
360 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
361 * limit reclaim to prevent infinite loops, if they ever occur.
362 */
363#define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
364#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
365
366enum charge_type {
367 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
368 MEM_CGROUP_CHARGE_TYPE_MAPPED,
369 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
370 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
371 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
372 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
373 NR_CHARGE_TYPE,
374};
375
376/* for encoding cft->private value on file */
377#define _MEM (0)
378#define _MEMSWAP (1)
379#define _OOM_TYPE (2)
380#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
381#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
382#define MEMFILE_ATTR(val) ((val) & 0xffff)
383/* Used for OOM nofiier */
384#define OOM_CONTROL (0)
385
386/*
387 * Reclaim flags for mem_cgroup_hierarchical_reclaim
388 */
389#define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
390#define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
391#define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
392#define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
393
394static void mem_cgroup_get(struct mem_cgroup *memcg);
395static void mem_cgroup_put(struct mem_cgroup *memcg);
396
397/* Writing them here to avoid exposing memcg's inner layout */
398#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
399#include <net/sock.h>
400#include <net/ip.h>
401
402static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
403void sock_update_memcg(struct sock *sk)
404{
405 if (mem_cgroup_sockets_enabled) {
406 struct mem_cgroup *memcg;
407
408 BUG_ON(!sk->sk_prot->proto_cgroup);
409
410 /* Socket cloning can throw us here with sk_cgrp already
411 * filled. It won't however, necessarily happen from
412 * process context. So the test for root memcg given
413 * the current task's memcg won't help us in this case.
414 *
415 * Respecting the original socket's memcg is a better
416 * decision in this case.
417 */
418 if (sk->sk_cgrp) {
419 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
420 mem_cgroup_get(sk->sk_cgrp->memcg);
421 return;
422 }
423
424 rcu_read_lock();
425 memcg = mem_cgroup_from_task(current);
426 if (!mem_cgroup_is_root(memcg)) {
427 mem_cgroup_get(memcg);
428 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
429 }
430 rcu_read_unlock();
431 }
432}
433EXPORT_SYMBOL(sock_update_memcg);
434
435void sock_release_memcg(struct sock *sk)
436{
437 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
438 struct mem_cgroup *memcg;
439 WARN_ON(!sk->sk_cgrp->memcg);
440 memcg = sk->sk_cgrp->memcg;
441 mem_cgroup_put(memcg);
442 }
443}
444
445#ifdef CONFIG_INET
446struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
447{
448 if (!memcg || mem_cgroup_is_root(memcg))
449 return NULL;
450
451 return &memcg->tcp_mem.cg_proto;
452}
453EXPORT_SYMBOL(tcp_proto_cgroup);
454#endif /* CONFIG_INET */
455#endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
456
457static void drain_all_stock_async(struct mem_cgroup *memcg);
458
459static struct mem_cgroup_per_zone *
460mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
461{
462 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
463}
464
465struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
466{
467 return &memcg->css;
468}
469
470static struct mem_cgroup_per_zone *
471page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
472{
473 int nid = page_to_nid(page);
474 int zid = page_zonenum(page);
475
476 return mem_cgroup_zoneinfo(memcg, nid, zid);
477}
478
479static struct mem_cgroup_tree_per_zone *
480soft_limit_tree_node_zone(int nid, int zid)
481{
482 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
483}
484
485static struct mem_cgroup_tree_per_zone *
486soft_limit_tree_from_page(struct page *page)
487{
488 int nid = page_to_nid(page);
489 int zid = page_zonenum(page);
490
491 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
492}
493
494static void
495__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
496 struct mem_cgroup_per_zone *mz,
497 struct mem_cgroup_tree_per_zone *mctz,
498 unsigned long long new_usage_in_excess)
499{
500 struct rb_node **p = &mctz->rb_root.rb_node;
501 struct rb_node *parent = NULL;
502 struct mem_cgroup_per_zone *mz_node;
503
504 if (mz->on_tree)
505 return;
506
507 mz->usage_in_excess = new_usage_in_excess;
508 if (!mz->usage_in_excess)
509 return;
510 while (*p) {
511 parent = *p;
512 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
513 tree_node);
514 if (mz->usage_in_excess < mz_node->usage_in_excess)
515 p = &(*p)->rb_left;
516 /*
517 * We can't avoid mem cgroups that are over their soft
518 * limit by the same amount
519 */
520 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
521 p = &(*p)->rb_right;
522 }
523 rb_link_node(&mz->tree_node, parent, p);
524 rb_insert_color(&mz->tree_node, &mctz->rb_root);
525 mz->on_tree = true;
526}
527
528static void
529__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
530 struct mem_cgroup_per_zone *mz,
531 struct mem_cgroup_tree_per_zone *mctz)
532{
533 if (!mz->on_tree)
534 return;
535 rb_erase(&mz->tree_node, &mctz->rb_root);
536 mz->on_tree = false;
537}
538
539static void
540mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
541 struct mem_cgroup_per_zone *mz,
542 struct mem_cgroup_tree_per_zone *mctz)
543{
544 spin_lock(&mctz->lock);
545 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
546 spin_unlock(&mctz->lock);
547}
548
549
550static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
551{
552 unsigned long long excess;
553 struct mem_cgroup_per_zone *mz;
554 struct mem_cgroup_tree_per_zone *mctz;
555 int nid = page_to_nid(page);
556 int zid = page_zonenum(page);
557 mctz = soft_limit_tree_from_page(page);
558
559 /*
560 * Necessary to update all ancestors when hierarchy is used.
561 * because their event counter is not touched.
562 */
563 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
564 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
565 excess = res_counter_soft_limit_excess(&memcg->res);
566 /*
567 * We have to update the tree if mz is on RB-tree or
568 * mem is over its softlimit.
569 */
570 if (excess || mz->on_tree) {
571 spin_lock(&mctz->lock);
572 /* if on-tree, remove it */
573 if (mz->on_tree)
574 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
575 /*
576 * Insert again. mz->usage_in_excess will be updated.
577 * If excess is 0, no tree ops.
578 */
579 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
580 spin_unlock(&mctz->lock);
581 }
582 }
583}
584
585static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
586{
587 int node, zone;
588 struct mem_cgroup_per_zone *mz;
589 struct mem_cgroup_tree_per_zone *mctz;
590
591 for_each_node(node) {
592 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
593 mz = mem_cgroup_zoneinfo(memcg, node, zone);
594 mctz = soft_limit_tree_node_zone(node, zone);
595 mem_cgroup_remove_exceeded(memcg, mz, mctz);
596 }
597 }
598}
599
600static struct mem_cgroup_per_zone *
601__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
602{
603 struct rb_node *rightmost = NULL;
604 struct mem_cgroup_per_zone *mz;
605
606retry:
607 mz = NULL;
608 rightmost = rb_last(&mctz->rb_root);
609 if (!rightmost)
610 goto done; /* Nothing to reclaim from */
611
612 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
613 /*
614 * Remove the node now but someone else can add it back,
615 * we will to add it back at the end of reclaim to its correct
616 * position in the tree.
617 */
618 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
619 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
620 !css_tryget(&mz->memcg->css))
621 goto retry;
622done:
623 return mz;
624}
625
626static struct mem_cgroup_per_zone *
627mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
628{
629 struct mem_cgroup_per_zone *mz;
630
631 spin_lock(&mctz->lock);
632 mz = __mem_cgroup_largest_soft_limit_node(mctz);
633 spin_unlock(&mctz->lock);
634 return mz;
635}
636
637/*
638 * Implementation Note: reading percpu statistics for memcg.
639 *
640 * Both of vmstat[] and percpu_counter has threshold and do periodic
641 * synchronization to implement "quick" read. There are trade-off between
642 * reading cost and precision of value. Then, we may have a chance to implement
643 * a periodic synchronizion of counter in memcg's counter.
644 *
645 * But this _read() function is used for user interface now. The user accounts
646 * memory usage by memory cgroup and he _always_ requires exact value because
647 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
648 * have to visit all online cpus and make sum. So, for now, unnecessary
649 * synchronization is not implemented. (just implemented for cpu hotplug)
650 *
651 * If there are kernel internal actions which can make use of some not-exact
652 * value, and reading all cpu value can be performance bottleneck in some
653 * common workload, threashold and synchonization as vmstat[] should be
654 * implemented.
655 */
656static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
657 enum mem_cgroup_stat_index idx)
658{
659 long val = 0;
660 int cpu;
661
662 get_online_cpus();
663 for_each_online_cpu(cpu)
664 val += per_cpu(memcg->stat->count[idx], cpu);
665#ifdef CONFIG_HOTPLUG_CPU
666 spin_lock(&memcg->pcp_counter_lock);
667 val += memcg->nocpu_base.count[idx];
668 spin_unlock(&memcg->pcp_counter_lock);
669#endif
670 put_online_cpus();
671 return val;
672}
673
674static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
675 bool charge)
676{
677 int val = (charge) ? 1 : -1;
678 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
679}
680
681static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
682 enum mem_cgroup_events_index idx)
683{
684 unsigned long val = 0;
685 int cpu;
686
687 for_each_online_cpu(cpu)
688 val += per_cpu(memcg->stat->events[idx], cpu);
689#ifdef CONFIG_HOTPLUG_CPU
690 spin_lock(&memcg->pcp_counter_lock);
691 val += memcg->nocpu_base.events[idx];
692 spin_unlock(&memcg->pcp_counter_lock);
693#endif
694 return val;
695}
696
697static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
698 bool anon, int nr_pages)
699{
700 preempt_disable();
701
702 /*
703 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
704 * counted as CACHE even if it's on ANON LRU.
705 */
706 if (anon)
707 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
708 nr_pages);
709 else
710 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
711 nr_pages);
712
713 /* pagein of a big page is an event. So, ignore page size */
714 if (nr_pages > 0)
715 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
716 else {
717 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
718 nr_pages = -nr_pages; /* for event */
719 }
720
721 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
722
723 preempt_enable();
724}
725
726unsigned long
727mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
728 unsigned int lru_mask)
729{
730 struct mem_cgroup_per_zone *mz;
731 enum lru_list lru;
732 unsigned long ret = 0;
733
734 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
735
736 for_each_lru(lru) {
737 if (BIT(lru) & lru_mask)
738 ret += mz->lru_size[lru];
739 }
740 return ret;
741}
742
743static unsigned long
744mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
745 int nid, unsigned int lru_mask)
746{
747 u64 total = 0;
748 int zid;
749
750 for (zid = 0; zid < MAX_NR_ZONES; zid++)
751 total += mem_cgroup_zone_nr_lru_pages(memcg,
752 nid, zid, lru_mask);
753
754 return total;
755}
756
757static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
758 unsigned int lru_mask)
759{
760 int nid;
761 u64 total = 0;
762
763 for_each_node_state(nid, N_HIGH_MEMORY)
764 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
765 return total;
766}
767
768static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
769 enum mem_cgroup_events_target target)
770{
771 unsigned long val, next;
772
773 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
774 next = __this_cpu_read(memcg->stat->targets[target]);
775 /* from time_after() in jiffies.h */
776 if ((long)next - (long)val < 0) {
777 switch (target) {
778 case MEM_CGROUP_TARGET_THRESH:
779 next = val + THRESHOLDS_EVENTS_TARGET;
780 break;
781 case MEM_CGROUP_TARGET_SOFTLIMIT:
782 next = val + SOFTLIMIT_EVENTS_TARGET;
783 break;
784 case MEM_CGROUP_TARGET_NUMAINFO:
785 next = val + NUMAINFO_EVENTS_TARGET;
786 break;
787 default:
788 break;
789 }
790 __this_cpu_write(memcg->stat->targets[target], next);
791 return true;
792 }
793 return false;
794}
795
796/*
797 * Check events in order.
798 *
799 */
800static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
801{
802 preempt_disable();
803 /* threshold event is triggered in finer grain than soft limit */
804 if (unlikely(mem_cgroup_event_ratelimit(memcg,
805 MEM_CGROUP_TARGET_THRESH))) {
806 bool do_softlimit;
807 bool do_numainfo __maybe_unused;
808
809 do_softlimit = mem_cgroup_event_ratelimit(memcg,
810 MEM_CGROUP_TARGET_SOFTLIMIT);
811#if MAX_NUMNODES > 1
812 do_numainfo = mem_cgroup_event_ratelimit(memcg,
813 MEM_CGROUP_TARGET_NUMAINFO);
814#endif
815 preempt_enable();
816
817 mem_cgroup_threshold(memcg);
818 if (unlikely(do_softlimit))
819 mem_cgroup_update_tree(memcg, page);
820#if MAX_NUMNODES > 1
821 if (unlikely(do_numainfo))
822 atomic_inc(&memcg->numainfo_events);
823#endif
824 } else
825 preempt_enable();
826}
827
828struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
829{
830 return container_of(cgroup_subsys_state(cont,
831 mem_cgroup_subsys_id), struct mem_cgroup,
832 css);
833}
834
835struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
836{
837 /*
838 * mm_update_next_owner() may clear mm->owner to NULL
839 * if it races with swapoff, page migration, etc.
840 * So this can be called with p == NULL.
841 */
842 if (unlikely(!p))
843 return NULL;
844
845 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
846 struct mem_cgroup, css);
847}
848
849struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
850{
851 struct mem_cgroup *memcg = NULL;
852
853 if (!mm)
854 return NULL;
855 /*
856 * Because we have no locks, mm->owner's may be being moved to other
857 * cgroup. We use css_tryget() here even if this looks
858 * pessimistic (rather than adding locks here).
859 */
860 rcu_read_lock();
861 do {
862 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
863 if (unlikely(!memcg))
864 break;
865 } while (!css_tryget(&memcg->css));
866 rcu_read_unlock();
867 return memcg;
868}
869
870/**
871 * mem_cgroup_iter - iterate over memory cgroup hierarchy
872 * @root: hierarchy root
873 * @prev: previously returned memcg, NULL on first invocation
874 * @reclaim: cookie for shared reclaim walks, NULL for full walks
875 *
876 * Returns references to children of the hierarchy below @root, or
877 * @root itself, or %NULL after a full round-trip.
878 *
879 * Caller must pass the return value in @prev on subsequent
880 * invocations for reference counting, or use mem_cgroup_iter_break()
881 * to cancel a hierarchy walk before the round-trip is complete.
882 *
883 * Reclaimers can specify a zone and a priority level in @reclaim to
884 * divide up the memcgs in the hierarchy among all concurrent
885 * reclaimers operating on the same zone and priority.
886 */
887struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
888 struct mem_cgroup *prev,
889 struct mem_cgroup_reclaim_cookie *reclaim)
890{
891 struct mem_cgroup *memcg = NULL;
892 int id = 0;
893
894 if (mem_cgroup_disabled())
895 return NULL;
896
897 if (!root)
898 root = root_mem_cgroup;
899
900 if (prev && !reclaim)
901 id = css_id(&prev->css);
902
903 if (prev && prev != root)
904 css_put(&prev->css);
905
906 if (!root->use_hierarchy && root != root_mem_cgroup) {
907 if (prev)
908 return NULL;
909 return root;
910 }
911
912 while (!memcg) {
913 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
914 struct cgroup_subsys_state *css;
915
916 if (reclaim) {
917 int nid = zone_to_nid(reclaim->zone);
918 int zid = zone_idx(reclaim->zone);
919 struct mem_cgroup_per_zone *mz;
920
921 mz = mem_cgroup_zoneinfo(root, nid, zid);
922 iter = &mz->reclaim_iter[reclaim->priority];
923 if (prev && reclaim->generation != iter->generation)
924 return NULL;
925 id = iter->position;
926 }
927
928 rcu_read_lock();
929 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
930 if (css) {
931 if (css == &root->css || css_tryget(css))
932 memcg = container_of(css,
933 struct mem_cgroup, css);
934 } else
935 id = 0;
936 rcu_read_unlock();
937
938 if (reclaim) {
939 iter->position = id;
940 if (!css)
941 iter->generation++;
942 else if (!prev && memcg)
943 reclaim->generation = iter->generation;
944 }
945
946 if (prev && !css)
947 return NULL;
948 }
949 return memcg;
950}
951
952/**
953 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
954 * @root: hierarchy root
955 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
956 */
957void mem_cgroup_iter_break(struct mem_cgroup *root,
958 struct mem_cgroup *prev)
959{
960 if (!root)
961 root = root_mem_cgroup;
962 if (prev && prev != root)
963 css_put(&prev->css);
964}
965
966/*
967 * Iteration constructs for visiting all cgroups (under a tree). If
968 * loops are exited prematurely (break), mem_cgroup_iter_break() must
969 * be used for reference counting.
970 */
971#define for_each_mem_cgroup_tree(iter, root) \
972 for (iter = mem_cgroup_iter(root, NULL, NULL); \
973 iter != NULL; \
974 iter = mem_cgroup_iter(root, iter, NULL))
975
976#define for_each_mem_cgroup(iter) \
977 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
978 iter != NULL; \
979 iter = mem_cgroup_iter(NULL, iter, NULL))
980
981static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
982{
983 return (memcg == root_mem_cgroup);
984}
985
986void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
987{
988 struct mem_cgroup *memcg;
989
990 if (!mm)
991 return;
992
993 rcu_read_lock();
994 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
995 if (unlikely(!memcg))
996 goto out;
997
998 switch (idx) {
999 case PGFAULT:
1000 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1001 break;
1002 case PGMAJFAULT:
1003 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1004 break;
1005 default:
1006 BUG();
1007 }
1008out:
1009 rcu_read_unlock();
1010}
1011EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1012
1013/**
1014 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1015 * @zone: zone of the wanted lruvec
1016 * @mem: memcg of the wanted lruvec
1017 *
1018 * Returns the lru list vector holding pages for the given @zone and
1019 * @mem. This can be the global zone lruvec, if the memory controller
1020 * is disabled.
1021 */
1022struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1023 struct mem_cgroup *memcg)
1024{
1025 struct mem_cgroup_per_zone *mz;
1026
1027 if (mem_cgroup_disabled())
1028 return &zone->lruvec;
1029
1030 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1031 return &mz->lruvec;
1032}
1033
1034/*
1035 * Following LRU functions are allowed to be used without PCG_LOCK.
1036 * Operations are called by routine of global LRU independently from memcg.
1037 * What we have to take care of here is validness of pc->mem_cgroup.
1038 *
1039 * Changes to pc->mem_cgroup happens when
1040 * 1. charge
1041 * 2. moving account
1042 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1043 * It is added to LRU before charge.
1044 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1045 * When moving account, the page is not on LRU. It's isolated.
1046 */
1047
1048/**
1049 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1050 * @zone: zone of the page
1051 * @page: the page
1052 * @lru: current lru
1053 *
1054 * This function accounts for @page being added to @lru, and returns
1055 * the lruvec for the given @zone and the memcg @page is charged to.
1056 *
1057 * The callsite is then responsible for physically linking the page to
1058 * the returned lruvec->lists[@lru].
1059 */
1060struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1061 enum lru_list lru)
1062{
1063 struct mem_cgroup_per_zone *mz;
1064 struct mem_cgroup *memcg;
1065 struct page_cgroup *pc;
1066
1067 if (mem_cgroup_disabled())
1068 return &zone->lruvec;
1069
1070 pc = lookup_page_cgroup(page);
1071 memcg = pc->mem_cgroup;
1072
1073 /*
1074 * Surreptitiously switch any uncharged page to root:
1075 * an uncharged page off lru does nothing to secure
1076 * its former mem_cgroup from sudden removal.
1077 *
1078 * Our caller holds lru_lock, and PageCgroupUsed is updated
1079 * under page_cgroup lock: between them, they make all uses
1080 * of pc->mem_cgroup safe.
1081 */
1082 if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1083 pc->mem_cgroup = memcg = root_mem_cgroup;
1084
1085 mz = page_cgroup_zoneinfo(memcg, page);
1086 /* compound_order() is stabilized through lru_lock */
1087 mz->lru_size[lru] += 1 << compound_order(page);
1088 return &mz->lruvec;
1089}
1090
1091/**
1092 * mem_cgroup_lru_del_list - account for removing an lru page
1093 * @page: the page
1094 * @lru: target lru
1095 *
1096 * This function accounts for @page being removed from @lru.
1097 *
1098 * The callsite is then responsible for physically unlinking
1099 * @page->lru.
1100 */
1101void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1102{
1103 struct mem_cgroup_per_zone *mz;
1104 struct mem_cgroup *memcg;
1105 struct page_cgroup *pc;
1106
1107 if (mem_cgroup_disabled())
1108 return;
1109
1110 pc = lookup_page_cgroup(page);
1111 memcg = pc->mem_cgroup;
1112 VM_BUG_ON(!memcg);
1113 mz = page_cgroup_zoneinfo(memcg, page);
1114 /* huge page split is done under lru_lock. so, we have no races. */
1115 VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page)));
1116 mz->lru_size[lru] -= 1 << compound_order(page);
1117}
1118
1119void mem_cgroup_lru_del(struct page *page)
1120{
1121 mem_cgroup_lru_del_list(page, page_lru(page));
1122}
1123
1124/**
1125 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1126 * @zone: zone of the page
1127 * @page: the page
1128 * @from: current lru
1129 * @to: target lru
1130 *
1131 * This function accounts for @page being moved between the lrus @from
1132 * and @to, and returns the lruvec for the given @zone and the memcg
1133 * @page is charged to.
1134 *
1135 * The callsite is then responsible for physically relinking
1136 * @page->lru to the returned lruvec->lists[@to].
1137 */
1138struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1139 struct page *page,
1140 enum lru_list from,
1141 enum lru_list to)
1142{
1143 /* XXX: Optimize this, especially for @from == @to */
1144 mem_cgroup_lru_del_list(page, from);
1145 return mem_cgroup_lru_add_list(zone, page, to);
1146}
1147
1148/*
1149 * Checks whether given mem is same or in the root_mem_cgroup's
1150 * hierarchy subtree
1151 */
1152static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1153 struct mem_cgroup *memcg)
1154{
1155 if (root_memcg != memcg) {
1156 return (root_memcg->use_hierarchy &&
1157 css_is_ancestor(&memcg->css, &root_memcg->css));
1158 }
1159
1160 return true;
1161}
1162
1163int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1164{
1165 int ret;
1166 struct mem_cgroup *curr = NULL;
1167 struct task_struct *p;
1168
1169 p = find_lock_task_mm(task);
1170 if (p) {
1171 curr = try_get_mem_cgroup_from_mm(p->mm);
1172 task_unlock(p);
1173 } else {
1174 /*
1175 * All threads may have already detached their mm's, but the oom
1176 * killer still needs to detect if they have already been oom
1177 * killed to prevent needlessly killing additional tasks.
1178 */
1179 task_lock(task);
1180 curr = mem_cgroup_from_task(task);
1181 if (curr)
1182 css_get(&curr->css);
1183 task_unlock(task);
1184 }
1185 if (!curr)
1186 return 0;
1187 /*
1188 * We should check use_hierarchy of "memcg" not "curr". Because checking
1189 * use_hierarchy of "curr" here make this function true if hierarchy is
1190 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1191 * hierarchy(even if use_hierarchy is disabled in "memcg").
1192 */
1193 ret = mem_cgroup_same_or_subtree(memcg, curr);
1194 css_put(&curr->css);
1195 return ret;
1196}
1197
1198int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1199{
1200 unsigned long inactive_ratio;
1201 int nid = zone_to_nid(zone);
1202 int zid = zone_idx(zone);
1203 unsigned long inactive;
1204 unsigned long active;
1205 unsigned long gb;
1206
1207 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1208 BIT(LRU_INACTIVE_ANON));
1209 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1210 BIT(LRU_ACTIVE_ANON));
1211
1212 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1213 if (gb)
1214 inactive_ratio = int_sqrt(10 * gb);
1215 else
1216 inactive_ratio = 1;
1217
1218 return inactive * inactive_ratio < active;
1219}
1220
1221int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1222{
1223 unsigned long active;
1224 unsigned long inactive;
1225 int zid = zone_idx(zone);
1226 int nid = zone_to_nid(zone);
1227
1228 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1229 BIT(LRU_INACTIVE_FILE));
1230 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1231 BIT(LRU_ACTIVE_FILE));
1232
1233 return (active > inactive);
1234}
1235
1236struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1237 struct zone *zone)
1238{
1239 int nid = zone_to_nid(zone);
1240 int zid = zone_idx(zone);
1241 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1242
1243 return &mz->reclaim_stat;
1244}
1245
1246struct zone_reclaim_stat *
1247mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1248{
1249 struct page_cgroup *pc;
1250 struct mem_cgroup_per_zone *mz;
1251
1252 if (mem_cgroup_disabled())
1253 return NULL;
1254
1255 pc = lookup_page_cgroup(page);
1256 if (!PageCgroupUsed(pc))
1257 return NULL;
1258 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1259 smp_rmb();
1260 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1261 return &mz->reclaim_stat;
1262}
1263
1264#define mem_cgroup_from_res_counter(counter, member) \
1265 container_of(counter, struct mem_cgroup, member)
1266
1267/**
1268 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1269 * @mem: the memory cgroup
1270 *
1271 * Returns the maximum amount of memory @mem can be charged with, in
1272 * pages.
1273 */
1274static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1275{
1276 unsigned long long margin;
1277
1278 margin = res_counter_margin(&memcg->res);
1279 if (do_swap_account)
1280 margin = min(margin, res_counter_margin(&memcg->memsw));
1281 return margin >> PAGE_SHIFT;
1282}
1283
1284int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1285{
1286 struct cgroup *cgrp = memcg->css.cgroup;
1287
1288 /* root ? */
1289 if (cgrp->parent == NULL)
1290 return vm_swappiness;
1291
1292 return memcg->swappiness;
1293}
1294
1295/*
1296 * memcg->moving_account is used for checking possibility that some thread is
1297 * calling move_account(). When a thread on CPU-A starts moving pages under
1298 * a memcg, other threads should check memcg->moving_account under
1299 * rcu_read_lock(), like this:
1300 *
1301 * CPU-A CPU-B
1302 * rcu_read_lock()
1303 * memcg->moving_account+1 if (memcg->mocing_account)
1304 * take heavy locks.
1305 * synchronize_rcu() update something.
1306 * rcu_read_unlock()
1307 * start move here.
1308 */
1309
1310/* for quick checking without looking up memcg */
1311atomic_t memcg_moving __read_mostly;
1312
1313static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1314{
1315 atomic_inc(&memcg_moving);
1316 atomic_inc(&memcg->moving_account);
1317 synchronize_rcu();
1318}
1319
1320static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1321{
1322 /*
1323 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1324 * We check NULL in callee rather than caller.
1325 */
1326 if (memcg) {
1327 atomic_dec(&memcg_moving);
1328 atomic_dec(&memcg->moving_account);
1329 }
1330}
1331
1332/*
1333 * 2 routines for checking "mem" is under move_account() or not.
1334 *
1335 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1336 * is used for avoiding races in accounting. If true,
1337 * pc->mem_cgroup may be overwritten.
1338 *
1339 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1340 * under hierarchy of moving cgroups. This is for
1341 * waiting at hith-memory prressure caused by "move".
1342 */
1343
1344static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1345{
1346 VM_BUG_ON(!rcu_read_lock_held());
1347 return atomic_read(&memcg->moving_account) > 0;
1348}
1349
1350static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1351{
1352 struct mem_cgroup *from;
1353 struct mem_cgroup *to;
1354 bool ret = false;
1355 /*
1356 * Unlike task_move routines, we access mc.to, mc.from not under
1357 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1358 */
1359 spin_lock(&mc.lock);
1360 from = mc.from;
1361 to = mc.to;
1362 if (!from)
1363 goto unlock;
1364
1365 ret = mem_cgroup_same_or_subtree(memcg, from)
1366 || mem_cgroup_same_or_subtree(memcg, to);
1367unlock:
1368 spin_unlock(&mc.lock);
1369 return ret;
1370}
1371
1372static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1373{
1374 if (mc.moving_task && current != mc.moving_task) {
1375 if (mem_cgroup_under_move(memcg)) {
1376 DEFINE_WAIT(wait);
1377 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1378 /* moving charge context might have finished. */
1379 if (mc.moving_task)
1380 schedule();
1381 finish_wait(&mc.waitq, &wait);
1382 return true;
1383 }
1384 }
1385 return false;
1386}
1387
1388/*
1389 * Take this lock when
1390 * - a code tries to modify page's memcg while it's USED.
1391 * - a code tries to modify page state accounting in a memcg.
1392 * see mem_cgroup_stolen(), too.
1393 */
1394static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1395 unsigned long *flags)
1396{
1397 spin_lock_irqsave(&memcg->move_lock, *flags);
1398}
1399
1400static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1401 unsigned long *flags)
1402{
1403 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1404}
1405
1406/**
1407 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1408 * @memcg: The memory cgroup that went over limit
1409 * @p: Task that is going to be killed
1410 *
1411 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1412 * enabled
1413 */
1414void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1415{
1416 struct cgroup *task_cgrp;
1417 struct cgroup *mem_cgrp;
1418 /*
1419 * Need a buffer in BSS, can't rely on allocations. The code relies
1420 * on the assumption that OOM is serialized for memory controller.
1421 * If this assumption is broken, revisit this code.
1422 */
1423 static char memcg_name[PATH_MAX];
1424 int ret;
1425
1426 if (!memcg || !p)
1427 return;
1428
1429 rcu_read_lock();
1430
1431 mem_cgrp = memcg->css.cgroup;
1432 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1433
1434 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1435 if (ret < 0) {
1436 /*
1437 * Unfortunately, we are unable to convert to a useful name
1438 * But we'll still print out the usage information
1439 */
1440 rcu_read_unlock();
1441 goto done;
1442 }
1443 rcu_read_unlock();
1444
1445 printk(KERN_INFO "Task in %s killed", memcg_name);
1446
1447 rcu_read_lock();
1448 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1449 if (ret < 0) {
1450 rcu_read_unlock();
1451 goto done;
1452 }
1453 rcu_read_unlock();
1454
1455 /*
1456 * Continues from above, so we don't need an KERN_ level
1457 */
1458 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1459done:
1460
1461 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1462 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1463 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1464 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1465 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1466 "failcnt %llu\n",
1467 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1468 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1469 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1470}
1471
1472/*
1473 * This function returns the number of memcg under hierarchy tree. Returns
1474 * 1(self count) if no children.
1475 */
1476static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1477{
1478 int num = 0;
1479 struct mem_cgroup *iter;
1480
1481 for_each_mem_cgroup_tree(iter, memcg)
1482 num++;
1483 return num;
1484}
1485
1486/*
1487 * Return the memory (and swap, if configured) limit for a memcg.
1488 */
1489u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1490{
1491 u64 limit;
1492
1493 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1494
1495 /*
1496 * Do not consider swap space if we cannot swap due to swappiness
1497 */
1498 if (mem_cgroup_swappiness(memcg)) {
1499 u64 memsw;
1500
1501 limit += total_swap_pages << PAGE_SHIFT;
1502 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1503
1504 /*
1505 * If memsw is finite and limits the amount of swap space
1506 * available to this memcg, return that limit.
1507 */
1508 limit = min(limit, memsw);
1509 }
1510
1511 return limit;
1512}
1513
1514static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1515 gfp_t gfp_mask,
1516 unsigned long flags)
1517{
1518 unsigned long total = 0;
1519 bool noswap = false;
1520 int loop;
1521
1522 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1523 noswap = true;
1524 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1525 noswap = true;
1526
1527 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1528 if (loop)
1529 drain_all_stock_async(memcg);
1530 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1531 /*
1532 * Allow limit shrinkers, which are triggered directly
1533 * by userspace, to catch signals and stop reclaim
1534 * after minimal progress, regardless of the margin.
1535 */
1536 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1537 break;
1538 if (mem_cgroup_margin(memcg))
1539 break;
1540 /*
1541 * If nothing was reclaimed after two attempts, there
1542 * may be no reclaimable pages in this hierarchy.
1543 */
1544 if (loop && !total)
1545 break;
1546 }
1547 return total;
1548}
1549
1550/**
1551 * test_mem_cgroup_node_reclaimable
1552 * @mem: the target memcg
1553 * @nid: the node ID to be checked.
1554 * @noswap : specify true here if the user wants flle only information.
1555 *
1556 * This function returns whether the specified memcg contains any
1557 * reclaimable pages on a node. Returns true if there are any reclaimable
1558 * pages in the node.
1559 */
1560static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1561 int nid, bool noswap)
1562{
1563 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1564 return true;
1565 if (noswap || !total_swap_pages)
1566 return false;
1567 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1568 return true;
1569 return false;
1570
1571}
1572#if MAX_NUMNODES > 1
1573
1574/*
1575 * Always updating the nodemask is not very good - even if we have an empty
1576 * list or the wrong list here, we can start from some node and traverse all
1577 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1578 *
1579 */
1580static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1581{
1582 int nid;
1583 /*
1584 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1585 * pagein/pageout changes since the last update.
1586 */
1587 if (!atomic_read(&memcg->numainfo_events))
1588 return;
1589 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1590 return;
1591
1592 /* make a nodemask where this memcg uses memory from */
1593 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1594
1595 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1596
1597 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1598 node_clear(nid, memcg->scan_nodes);
1599 }
1600
1601 atomic_set(&memcg->numainfo_events, 0);
1602 atomic_set(&memcg->numainfo_updating, 0);
1603}
1604
1605/*
1606 * Selecting a node where we start reclaim from. Because what we need is just
1607 * reducing usage counter, start from anywhere is O,K. Considering
1608 * memory reclaim from current node, there are pros. and cons.
1609 *
1610 * Freeing memory from current node means freeing memory from a node which
1611 * we'll use or we've used. So, it may make LRU bad. And if several threads
1612 * hit limits, it will see a contention on a node. But freeing from remote
1613 * node means more costs for memory reclaim because of memory latency.
1614 *
1615 * Now, we use round-robin. Better algorithm is welcomed.
1616 */
1617int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1618{
1619 int node;
1620
1621 mem_cgroup_may_update_nodemask(memcg);
1622 node = memcg->last_scanned_node;
1623
1624 node = next_node(node, memcg->scan_nodes);
1625 if (node == MAX_NUMNODES)
1626 node = first_node(memcg->scan_nodes);
1627 /*
1628 * We call this when we hit limit, not when pages are added to LRU.
1629 * No LRU may hold pages because all pages are UNEVICTABLE or
1630 * memcg is too small and all pages are not on LRU. In that case,
1631 * we use curret node.
1632 */
1633 if (unlikely(node == MAX_NUMNODES))
1634 node = numa_node_id();
1635
1636 memcg->last_scanned_node = node;
1637 return node;
1638}
1639
1640/*
1641 * Check all nodes whether it contains reclaimable pages or not.
1642 * For quick scan, we make use of scan_nodes. This will allow us to skip
1643 * unused nodes. But scan_nodes is lazily updated and may not cotain
1644 * enough new information. We need to do double check.
1645 */
1646bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1647{
1648 int nid;
1649
1650 /*
1651 * quick check...making use of scan_node.
1652 * We can skip unused nodes.
1653 */
1654 if (!nodes_empty(memcg->scan_nodes)) {
1655 for (nid = first_node(memcg->scan_nodes);
1656 nid < MAX_NUMNODES;
1657 nid = next_node(nid, memcg->scan_nodes)) {
1658
1659 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1660 return true;
1661 }
1662 }
1663 /*
1664 * Check rest of nodes.
1665 */
1666 for_each_node_state(nid, N_HIGH_MEMORY) {
1667 if (node_isset(nid, memcg->scan_nodes))
1668 continue;
1669 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1670 return true;
1671 }
1672 return false;
1673}
1674
1675#else
1676int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1677{
1678 return 0;
1679}
1680
1681bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1682{
1683 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1684}
1685#endif
1686
1687static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1688 struct zone *zone,
1689 gfp_t gfp_mask,
1690 unsigned long *total_scanned)
1691{
1692 struct mem_cgroup *victim = NULL;
1693 int total = 0;
1694 int loop = 0;
1695 unsigned long excess;
1696 unsigned long nr_scanned;
1697 struct mem_cgroup_reclaim_cookie reclaim = {
1698 .zone = zone,
1699 .priority = 0,
1700 };
1701
1702 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1703
1704 while (1) {
1705 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1706 if (!victim) {
1707 loop++;
1708 if (loop >= 2) {
1709 /*
1710 * If we have not been able to reclaim
1711 * anything, it might because there are
1712 * no reclaimable pages under this hierarchy
1713 */
1714 if (!total)
1715 break;
1716 /*
1717 * We want to do more targeted reclaim.
1718 * excess >> 2 is not to excessive so as to
1719 * reclaim too much, nor too less that we keep
1720 * coming back to reclaim from this cgroup
1721 */
1722 if (total >= (excess >> 2) ||
1723 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1724 break;
1725 }
1726 continue;
1727 }
1728 if (!mem_cgroup_reclaimable(victim, false))
1729 continue;
1730 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1731 zone, &nr_scanned);
1732 *total_scanned += nr_scanned;
1733 if (!res_counter_soft_limit_excess(&root_memcg->res))
1734 break;
1735 }
1736 mem_cgroup_iter_break(root_memcg, victim);
1737 return total;
1738}
1739
1740/*
1741 * Check OOM-Killer is already running under our hierarchy.
1742 * If someone is running, return false.
1743 * Has to be called with memcg_oom_lock
1744 */
1745static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1746{
1747 struct mem_cgroup *iter, *failed = NULL;
1748
1749 for_each_mem_cgroup_tree(iter, memcg) {
1750 if (iter->oom_lock) {
1751 /*
1752 * this subtree of our hierarchy is already locked
1753 * so we cannot give a lock.
1754 */
1755 failed = iter;
1756 mem_cgroup_iter_break(memcg, iter);
1757 break;
1758 } else
1759 iter->oom_lock = true;
1760 }
1761
1762 if (!failed)
1763 return true;
1764
1765 /*
1766 * OK, we failed to lock the whole subtree so we have to clean up
1767 * what we set up to the failing subtree
1768 */
1769 for_each_mem_cgroup_tree(iter, memcg) {
1770 if (iter == failed) {
1771 mem_cgroup_iter_break(memcg, iter);
1772 break;
1773 }
1774 iter->oom_lock = false;
1775 }
1776 return false;
1777}
1778
1779/*
1780 * Has to be called with memcg_oom_lock
1781 */
1782static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1783{
1784 struct mem_cgroup *iter;
1785
1786 for_each_mem_cgroup_tree(iter, memcg)
1787 iter->oom_lock = false;
1788 return 0;
1789}
1790
1791static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1792{
1793 struct mem_cgroup *iter;
1794
1795 for_each_mem_cgroup_tree(iter, memcg)
1796 atomic_inc(&iter->under_oom);
1797}
1798
1799static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1800{
1801 struct mem_cgroup *iter;
1802
1803 /*
1804 * When a new child is created while the hierarchy is under oom,
1805 * mem_cgroup_oom_lock() may not be called. We have to use
1806 * atomic_add_unless() here.
1807 */
1808 for_each_mem_cgroup_tree(iter, memcg)
1809 atomic_add_unless(&iter->under_oom, -1, 0);
1810}
1811
1812static DEFINE_SPINLOCK(memcg_oom_lock);
1813static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1814
1815struct oom_wait_info {
1816 struct mem_cgroup *memcg;
1817 wait_queue_t wait;
1818};
1819
1820static int memcg_oom_wake_function(wait_queue_t *wait,
1821 unsigned mode, int sync, void *arg)
1822{
1823 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1824 struct mem_cgroup *oom_wait_memcg;
1825 struct oom_wait_info *oom_wait_info;
1826
1827 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1828 oom_wait_memcg = oom_wait_info->memcg;
1829
1830 /*
1831 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1832 * Then we can use css_is_ancestor without taking care of RCU.
1833 */
1834 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1835 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1836 return 0;
1837 return autoremove_wake_function(wait, mode, sync, arg);
1838}
1839
1840static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1841{
1842 /* for filtering, pass "memcg" as argument. */
1843 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1844}
1845
1846static void memcg_oom_recover(struct mem_cgroup *memcg)
1847{
1848 if (memcg && atomic_read(&memcg->under_oom))
1849 memcg_wakeup_oom(memcg);
1850}
1851
1852/*
1853 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1854 */
1855bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1856{
1857 struct oom_wait_info owait;
1858 bool locked, need_to_kill;
1859
1860 owait.memcg = memcg;
1861 owait.wait.flags = 0;
1862 owait.wait.func = memcg_oom_wake_function;
1863 owait.wait.private = current;
1864 INIT_LIST_HEAD(&owait.wait.task_list);
1865 need_to_kill = true;
1866 mem_cgroup_mark_under_oom(memcg);
1867
1868 /* At first, try to OOM lock hierarchy under memcg.*/
1869 spin_lock(&memcg_oom_lock);
1870 locked = mem_cgroup_oom_lock(memcg);
1871 /*
1872 * Even if signal_pending(), we can't quit charge() loop without
1873 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1874 * under OOM is always welcomed, use TASK_KILLABLE here.
1875 */
1876 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1877 if (!locked || memcg->oom_kill_disable)
1878 need_to_kill = false;
1879 if (locked)
1880 mem_cgroup_oom_notify(memcg);
1881 spin_unlock(&memcg_oom_lock);
1882
1883 if (need_to_kill) {
1884 finish_wait(&memcg_oom_waitq, &owait.wait);
1885 mem_cgroup_out_of_memory(memcg, mask, order);
1886 } else {
1887 schedule();
1888 finish_wait(&memcg_oom_waitq, &owait.wait);
1889 }
1890 spin_lock(&memcg_oom_lock);
1891 if (locked)
1892 mem_cgroup_oom_unlock(memcg);
1893 memcg_wakeup_oom(memcg);
1894 spin_unlock(&memcg_oom_lock);
1895
1896 mem_cgroup_unmark_under_oom(memcg);
1897
1898 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1899 return false;
1900 /* Give chance to dying process */
1901 schedule_timeout_uninterruptible(1);
1902 return true;
1903}
1904
1905/*
1906 * Currently used to update mapped file statistics, but the routine can be
1907 * generalized to update other statistics as well.
1908 *
1909 * Notes: Race condition
1910 *
1911 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1912 * it tends to be costly. But considering some conditions, we doesn't need
1913 * to do so _always_.
1914 *
1915 * Considering "charge", lock_page_cgroup() is not required because all
1916 * file-stat operations happen after a page is attached to radix-tree. There
1917 * are no race with "charge".
1918 *
1919 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1920 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1921 * if there are race with "uncharge". Statistics itself is properly handled
1922 * by flags.
1923 *
1924 * Considering "move", this is an only case we see a race. To make the race
1925 * small, we check mm->moving_account and detect there are possibility of race
1926 * If there is, we take a lock.
1927 */
1928
1929void __mem_cgroup_begin_update_page_stat(struct page *page,
1930 bool *locked, unsigned long *flags)
1931{
1932 struct mem_cgroup *memcg;
1933 struct page_cgroup *pc;
1934
1935 pc = lookup_page_cgroup(page);
1936again:
1937 memcg = pc->mem_cgroup;
1938 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1939 return;
1940 /*
1941 * If this memory cgroup is not under account moving, we don't
1942 * need to take move_lock_page_cgroup(). Because we already hold
1943 * rcu_read_lock(), any calls to move_account will be delayed until
1944 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1945 */
1946 if (!mem_cgroup_stolen(memcg))
1947 return;
1948
1949 move_lock_mem_cgroup(memcg, flags);
1950 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1951 move_unlock_mem_cgroup(memcg, flags);
1952 goto again;
1953 }
1954 *locked = true;
1955}
1956
1957void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1958{
1959 struct page_cgroup *pc = lookup_page_cgroup(page);
1960
1961 /*
1962 * It's guaranteed that pc->mem_cgroup never changes while
1963 * lock is held because a routine modifies pc->mem_cgroup
1964 * should take move_lock_page_cgroup().
1965 */
1966 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1967}
1968
1969void mem_cgroup_update_page_stat(struct page *page,
1970 enum mem_cgroup_page_stat_item idx, int val)
1971{
1972 struct mem_cgroup *memcg;
1973 struct page_cgroup *pc = lookup_page_cgroup(page);
1974 unsigned long uninitialized_var(flags);
1975
1976 if (mem_cgroup_disabled())
1977 return;
1978
1979 memcg = pc->mem_cgroup;
1980 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1981 return;
1982
1983 switch (idx) {
1984 case MEMCG_NR_FILE_MAPPED:
1985 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1986 break;
1987 default:
1988 BUG();
1989 }
1990
1991 this_cpu_add(memcg->stat->count[idx], val);
1992}
1993
1994/*
1995 * size of first charge trial. "32" comes from vmscan.c's magic value.
1996 * TODO: maybe necessary to use big numbers in big irons.
1997 */
1998#define CHARGE_BATCH 32U
1999struct memcg_stock_pcp {
2000 struct mem_cgroup *cached; /* this never be root cgroup */
2001 unsigned int nr_pages;
2002 struct work_struct work;
2003 unsigned long flags;
2004#define FLUSHING_CACHED_CHARGE (0)
2005};
2006static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2007static DEFINE_MUTEX(percpu_charge_mutex);
2008
2009/*
2010 * Try to consume stocked charge on this cpu. If success, one page is consumed
2011 * from local stock and true is returned. If the stock is 0 or charges from a
2012 * cgroup which is not current target, returns false. This stock will be
2013 * refilled.
2014 */
2015static bool consume_stock(struct mem_cgroup *memcg)
2016{
2017 struct memcg_stock_pcp *stock;
2018 bool ret = true;
2019
2020 stock = &get_cpu_var(memcg_stock);
2021 if (memcg == stock->cached && stock->nr_pages)
2022 stock->nr_pages--;
2023 else /* need to call res_counter_charge */
2024 ret = false;
2025 put_cpu_var(memcg_stock);
2026 return ret;
2027}
2028
2029/*
2030 * Returns stocks cached in percpu to res_counter and reset cached information.
2031 */
2032static void drain_stock(struct memcg_stock_pcp *stock)
2033{
2034 struct mem_cgroup *old = stock->cached;
2035
2036 if (stock->nr_pages) {
2037 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2038
2039 res_counter_uncharge(&old->res, bytes);
2040 if (do_swap_account)
2041 res_counter_uncharge(&old->memsw, bytes);
2042 stock->nr_pages = 0;
2043 }
2044 stock->cached = NULL;
2045}
2046
2047/*
2048 * This must be called under preempt disabled or must be called by
2049 * a thread which is pinned to local cpu.
2050 */
2051static void drain_local_stock(struct work_struct *dummy)
2052{
2053 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2054 drain_stock(stock);
2055 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2056}
2057
2058/*
2059 * Cache charges(val) which is from res_counter, to local per_cpu area.
2060 * This will be consumed by consume_stock() function, later.
2061 */
2062static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2063{
2064 struct memcg_stock_pcp *stock;
2065 int cpu = get_cpu_light();
2066
2067 stock = &per_cpu(memcg_stock, cpu);
2068
2069 if (stock->cached != memcg) { /* reset if necessary */
2070 drain_stock(stock);
2071 stock->cached = memcg;
2072 }
2073 stock->nr_pages += nr_pages;
2074 put_cpu_light();
2075}
2076
2077/*
2078 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2079 * of the hierarchy under it. sync flag says whether we should block
2080 * until the work is done.
2081 */
2082static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2083{
2084 int cpu, curcpu;
2085
2086 /* Notify other cpus that system-wide "drain" is running */
2087 get_online_cpus();
2088 curcpu = get_cpu_light();
2089 for_each_online_cpu(cpu) {
2090 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2091 struct mem_cgroup *memcg;
2092
2093 memcg = stock->cached;
2094 if (!memcg || !stock->nr_pages)
2095 continue;
2096 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2097 continue;
2098 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2099 if (cpu == curcpu)
2100 drain_local_stock(&stock->work);
2101 else
2102 schedule_work_on(cpu, &stock->work);
2103 }
2104 }
2105 put_cpu_light();
2106
2107 if (!sync)
2108 goto out;
2109
2110 for_each_online_cpu(cpu) {
2111 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2112 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2113 flush_work(&stock->work);
2114 }
2115out:
2116 put_online_cpus();
2117}
2118
2119/*
2120 * Tries to drain stocked charges in other cpus. This function is asynchronous
2121 * and just put a work per cpu for draining localy on each cpu. Caller can
2122 * expects some charges will be back to res_counter later but cannot wait for
2123 * it.
2124 */
2125static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2126{
2127 /*
2128 * If someone calls draining, avoid adding more kworker runs.
2129 */
2130 if (!mutex_trylock(&percpu_charge_mutex))
2131 return;
2132 drain_all_stock(root_memcg, false);
2133 mutex_unlock(&percpu_charge_mutex);
2134}
2135
2136/* This is a synchronous drain interface. */
2137static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2138{
2139 /* called when force_empty is called */
2140 mutex_lock(&percpu_charge_mutex);
2141 drain_all_stock(root_memcg, true);
2142 mutex_unlock(&percpu_charge_mutex);
2143}
2144
2145/*
2146 * This function drains percpu counter value from DEAD cpu and
2147 * move it to local cpu. Note that this function can be preempted.
2148 */
2149static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2150{
2151 int i;
2152
2153 spin_lock(&memcg->pcp_counter_lock);
2154 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2155 long x = per_cpu(memcg->stat->count[i], cpu);
2156
2157 per_cpu(memcg->stat->count[i], cpu) = 0;
2158 memcg->nocpu_base.count[i] += x;
2159 }
2160 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2161 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2162
2163 per_cpu(memcg->stat->events[i], cpu) = 0;
2164 memcg->nocpu_base.events[i] += x;
2165 }
2166 spin_unlock(&memcg->pcp_counter_lock);
2167}
2168
2169static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2170 unsigned long action,
2171 void *hcpu)
2172{
2173 int cpu = (unsigned long)hcpu;
2174 struct memcg_stock_pcp *stock;
2175 struct mem_cgroup *iter;
2176
2177 if (action == CPU_ONLINE)
2178 return NOTIFY_OK;
2179
2180 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2181 return NOTIFY_OK;
2182
2183 for_each_mem_cgroup(iter)
2184 mem_cgroup_drain_pcp_counter(iter, cpu);
2185
2186 stock = &per_cpu(memcg_stock, cpu);
2187 drain_stock(stock);
2188 return NOTIFY_OK;
2189}
2190
2191
2192/* See __mem_cgroup_try_charge() for details */
2193enum {
2194 CHARGE_OK, /* success */
2195 CHARGE_RETRY, /* need to retry but retry is not bad */
2196 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2197 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2198 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2199};
2200
2201static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2202 unsigned int nr_pages, bool oom_check)
2203{
2204 unsigned long csize = nr_pages * PAGE_SIZE;
2205 struct mem_cgroup *mem_over_limit;
2206 struct res_counter *fail_res;
2207 unsigned long flags = 0;
2208 int ret;
2209
2210 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2211
2212 if (likely(!ret)) {
2213 if (!do_swap_account)
2214 return CHARGE_OK;
2215 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2216 if (likely(!ret))
2217 return CHARGE_OK;
2218
2219 res_counter_uncharge(&memcg->res, csize);
2220 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2221 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2222 } else
2223 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2224 /*
2225 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2226 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2227 *
2228 * Never reclaim on behalf of optional batching, retry with a
2229 * single page instead.
2230 */
2231 if (nr_pages == CHARGE_BATCH)
2232 return CHARGE_RETRY;
2233
2234 if (!(gfp_mask & __GFP_WAIT))
2235 return CHARGE_WOULDBLOCK;
2236
2237 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2238 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2239 return CHARGE_RETRY;
2240 /*
2241 * Even though the limit is exceeded at this point, reclaim
2242 * may have been able to free some pages. Retry the charge
2243 * before killing the task.
2244 *
2245 * Only for regular pages, though: huge pages are rather
2246 * unlikely to succeed so close to the limit, and we fall back
2247 * to regular pages anyway in case of failure.
2248 */
2249 if (nr_pages == 1 && ret)
2250 return CHARGE_RETRY;
2251
2252 /*
2253 * At task move, charge accounts can be doubly counted. So, it's
2254 * better to wait until the end of task_move if something is going on.
2255 */
2256 if (mem_cgroup_wait_acct_move(mem_over_limit))
2257 return CHARGE_RETRY;
2258
2259 /* If we don't need to call oom-killer at el, return immediately */
2260 if (!oom_check)
2261 return CHARGE_NOMEM;
2262 /* check OOM */
2263 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2264 return CHARGE_OOM_DIE;
2265
2266 return CHARGE_RETRY;
2267}
2268
2269/*
2270 * __mem_cgroup_try_charge() does
2271 * 1. detect memcg to be charged against from passed *mm and *ptr,
2272 * 2. update res_counter
2273 * 3. call memory reclaim if necessary.
2274 *
2275 * In some special case, if the task is fatal, fatal_signal_pending() or
2276 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2277 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2278 * as possible without any hazards. 2: all pages should have a valid
2279 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2280 * pointer, that is treated as a charge to root_mem_cgroup.
2281 *
2282 * So __mem_cgroup_try_charge() will return
2283 * 0 ... on success, filling *ptr with a valid memcg pointer.
2284 * -ENOMEM ... charge failure because of resource limits.
2285 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2286 *
2287 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2288 * the oom-killer can be invoked.
2289 */
2290static int __mem_cgroup_try_charge(struct mm_struct *mm,
2291 gfp_t gfp_mask,
2292 unsigned int nr_pages,
2293 struct mem_cgroup **ptr,
2294 bool oom)
2295{
2296 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2297 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2298 struct mem_cgroup *memcg = NULL;
2299 int ret;
2300
2301 /*
2302 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2303 * in system level. So, allow to go ahead dying process in addition to
2304 * MEMDIE process.
2305 */
2306 if (unlikely(test_thread_flag(TIF_MEMDIE)
2307 || fatal_signal_pending(current)))
2308 goto bypass;
2309
2310 /*
2311 * We always charge the cgroup the mm_struct belongs to.
2312 * The mm_struct's mem_cgroup changes on task migration if the
2313 * thread group leader migrates. It's possible that mm is not
2314 * set, if so charge the init_mm (happens for pagecache usage).
2315 */
2316 if (!*ptr && !mm)
2317 *ptr = root_mem_cgroup;
2318again:
2319 if (*ptr) { /* css should be a valid one */
2320 memcg = *ptr;
2321 VM_BUG_ON(css_is_removed(&memcg->css));
2322 if (mem_cgroup_is_root(memcg))
2323 goto done;
2324 if (nr_pages == 1 && consume_stock(memcg))
2325 goto done;
2326 css_get(&memcg->css);
2327 } else {
2328 struct task_struct *p;
2329
2330 rcu_read_lock();
2331 p = rcu_dereference(mm->owner);
2332 /*
2333 * Because we don't have task_lock(), "p" can exit.
2334 * In that case, "memcg" can point to root or p can be NULL with
2335 * race with swapoff. Then, we have small risk of mis-accouning.
2336 * But such kind of mis-account by race always happens because
2337 * we don't have cgroup_mutex(). It's overkill and we allo that
2338 * small race, here.
2339 * (*) swapoff at el will charge against mm-struct not against
2340 * task-struct. So, mm->owner can be NULL.
2341 */
2342 memcg = mem_cgroup_from_task(p);
2343 if (!memcg)
2344 memcg = root_mem_cgroup;
2345 if (mem_cgroup_is_root(memcg)) {
2346 rcu_read_unlock();
2347 goto done;
2348 }
2349 if (nr_pages == 1 && consume_stock(memcg)) {
2350 /*
2351 * It seems dagerous to access memcg without css_get().
2352 * But considering how consume_stok works, it's not
2353 * necessary. If consume_stock success, some charges
2354 * from this memcg are cached on this cpu. So, we
2355 * don't need to call css_get()/css_tryget() before
2356 * calling consume_stock().
2357 */
2358 rcu_read_unlock();
2359 goto done;
2360 }
2361 /* after here, we may be blocked. we need to get refcnt */
2362 if (!css_tryget(&memcg->css)) {
2363 rcu_read_unlock();
2364 goto again;
2365 }
2366 rcu_read_unlock();
2367 }
2368
2369 do {
2370 bool oom_check;
2371
2372 /* If killed, bypass charge */
2373 if (fatal_signal_pending(current)) {
2374 css_put(&memcg->css);
2375 goto bypass;
2376 }
2377
2378 oom_check = false;
2379 if (oom && !nr_oom_retries) {
2380 oom_check = true;
2381 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2382 }
2383
2384 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2385 switch (ret) {
2386 case CHARGE_OK:
2387 break;
2388 case CHARGE_RETRY: /* not in OOM situation but retry */
2389 batch = nr_pages;
2390 css_put(&memcg->css);
2391 memcg = NULL;
2392 goto again;
2393 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2394 css_put(&memcg->css);
2395 goto nomem;
2396 case CHARGE_NOMEM: /* OOM routine works */
2397 if (!oom) {
2398 css_put(&memcg->css);
2399 goto nomem;
2400 }
2401 /* If oom, we never return -ENOMEM */
2402 nr_oom_retries--;
2403 break;
2404 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2405 css_put(&memcg->css);
2406 goto bypass;
2407 }
2408 } while (ret != CHARGE_OK);
2409
2410 if (batch > nr_pages)
2411 refill_stock(memcg, batch - nr_pages);
2412 css_put(&memcg->css);
2413done:
2414 *ptr = memcg;
2415 return 0;
2416nomem:
2417 *ptr = NULL;
2418 return -ENOMEM;
2419bypass:
2420 *ptr = root_mem_cgroup;
2421 return -EINTR;
2422}
2423
2424/*
2425 * Somemtimes we have to undo a charge we got by try_charge().
2426 * This function is for that and do uncharge, put css's refcnt.
2427 * gotten by try_charge().
2428 */
2429static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2430 unsigned int nr_pages)
2431{
2432 if (!mem_cgroup_is_root(memcg)) {
2433 unsigned long bytes = nr_pages * PAGE_SIZE;
2434
2435 res_counter_uncharge(&memcg->res, bytes);
2436 if (do_swap_account)
2437 res_counter_uncharge(&memcg->memsw, bytes);
2438 }
2439}
2440
2441/*
2442 * A helper function to get mem_cgroup from ID. must be called under
2443 * rcu_read_lock(). The caller must check css_is_removed() or some if
2444 * it's concern. (dropping refcnt from swap can be called against removed
2445 * memcg.)
2446 */
2447static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2448{
2449 struct cgroup_subsys_state *css;
2450
2451 /* ID 0 is unused ID */
2452 if (!id)
2453 return NULL;
2454 css = css_lookup(&mem_cgroup_subsys, id);
2455 if (!css)
2456 return NULL;
2457 return container_of(css, struct mem_cgroup, css);
2458}
2459
2460struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2461{
2462 struct mem_cgroup *memcg = NULL;
2463 struct page_cgroup *pc;
2464 unsigned short id;
2465 swp_entry_t ent;
2466
2467 VM_BUG_ON(!PageLocked(page));
2468
2469 pc = lookup_page_cgroup(page);
2470 lock_page_cgroup(pc);
2471 if (PageCgroupUsed(pc)) {
2472 memcg = pc->mem_cgroup;
2473 if (memcg && !css_tryget(&memcg->css))
2474 memcg = NULL;
2475 } else if (PageSwapCache(page)) {
2476 ent.val = page_private(page);
2477 id = lookup_swap_cgroup_id(ent);
2478 rcu_read_lock();
2479 memcg = mem_cgroup_lookup(id);
2480 if (memcg && !css_tryget(&memcg->css))
2481 memcg = NULL;
2482 rcu_read_unlock();
2483 }
2484 unlock_page_cgroup(pc);
2485 return memcg;
2486}
2487
2488static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2489 struct page *page,
2490 unsigned int nr_pages,
2491 enum charge_type ctype,
2492 bool lrucare)
2493{
2494 struct page_cgroup *pc = lookup_page_cgroup(page);
2495 struct zone *uninitialized_var(zone);
2496 bool was_on_lru = false;
2497 bool anon;
2498
2499 lock_page_cgroup(pc);
2500 if (unlikely(PageCgroupUsed(pc))) {
2501 unlock_page_cgroup(pc);
2502 __mem_cgroup_cancel_charge(memcg, nr_pages);
2503 return;
2504 }
2505 /*
2506 * we don't need page_cgroup_lock about tail pages, becase they are not
2507 * accessed by any other context at this point.
2508 */
2509
2510 /*
2511 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2512 * may already be on some other mem_cgroup's LRU. Take care of it.
2513 */
2514 if (lrucare) {
2515 zone = page_zone(page);
2516 spin_lock_irq(&zone->lru_lock);
2517 if (PageLRU(page)) {
2518 ClearPageLRU(page);
2519 del_page_from_lru_list(zone, page, page_lru(page));
2520 was_on_lru = true;
2521 }
2522 }
2523
2524 pc->mem_cgroup = memcg;
2525 /*
2526 * We access a page_cgroup asynchronously without lock_page_cgroup().
2527 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2528 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2529 * before USED bit, we need memory barrier here.
2530 * See mem_cgroup_add_lru_list(), etc.
2531 */
2532 smp_wmb();
2533 SetPageCgroupUsed(pc);
2534
2535 if (lrucare) {
2536 if (was_on_lru) {
2537 VM_BUG_ON(PageLRU(page));
2538 SetPageLRU(page);
2539 add_page_to_lru_list(zone, page, page_lru(page));
2540 }
2541 spin_unlock_irq(&zone->lru_lock);
2542 }
2543
2544 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2545 anon = true;
2546 else
2547 anon = false;
2548
2549 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2550 unlock_page_cgroup(pc);
2551
2552 /*
2553 * "charge_statistics" updated event counter. Then, check it.
2554 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2555 * if they exceeds softlimit.
2556 */
2557 memcg_check_events(memcg, page);
2558}
2559
2560#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2561
2562#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION))
2563/*
2564 * Because tail pages are not marked as "used", set it. We're under
2565 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2566 * charge/uncharge will be never happen and move_account() is done under
2567 * compound_lock(), so we don't have to take care of races.
2568 */
2569void mem_cgroup_split_huge_fixup(struct page *head)
2570{
2571 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2572 struct page_cgroup *pc;
2573 int i;
2574
2575 if (mem_cgroup_disabled())
2576 return;
2577 for (i = 1; i < HPAGE_PMD_NR; i++) {
2578 pc = head_pc + i;
2579 pc->mem_cgroup = head_pc->mem_cgroup;
2580 smp_wmb();/* see __commit_charge() */
2581 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2582 }
2583}
2584#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2585
2586/**
2587 * mem_cgroup_move_account - move account of the page
2588 * @page: the page
2589 * @nr_pages: number of regular pages (>1 for huge pages)
2590 * @pc: page_cgroup of the page.
2591 * @from: mem_cgroup which the page is moved from.
2592 * @to: mem_cgroup which the page is moved to. @from != @to.
2593 * @uncharge: whether we should call uncharge and css_put against @from.
2594 *
2595 * The caller must confirm following.
2596 * - page is not on LRU (isolate_page() is useful.)
2597 * - compound_lock is held when nr_pages > 1
2598 *
2599 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2600 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2601 * true, this function does "uncharge" from old cgroup, but it doesn't if
2602 * @uncharge is false, so a caller should do "uncharge".
2603 */
2604static int mem_cgroup_move_account(struct page *page,
2605 unsigned int nr_pages,
2606 struct page_cgroup *pc,
2607 struct mem_cgroup *from,
2608 struct mem_cgroup *to,
2609 bool uncharge)
2610{
2611 unsigned long flags;
2612 int ret;
2613 bool anon = PageAnon(page);
2614
2615 VM_BUG_ON(from == to);
2616 VM_BUG_ON(PageLRU(page));
2617 /*
2618 * The page is isolated from LRU. So, collapse function
2619 * will not handle this page. But page splitting can happen.
2620 * Do this check under compound_page_lock(). The caller should
2621 * hold it.
2622 */
2623 ret = -EBUSY;
2624 if (nr_pages > 1 && !PageTransHuge(page))
2625 goto out;
2626
2627 lock_page_cgroup(pc);
2628
2629 ret = -EINVAL;
2630 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2631 goto unlock;
2632
2633 move_lock_mem_cgroup(from, &flags);
2634
2635 if (!anon && page_mapped(page)) {
2636 /* Update mapped_file data for mem_cgroup */
2637 preempt_disable();
2638 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2639 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2640 preempt_enable();
2641 }
2642 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2643 if (uncharge)
2644 /* This is not "cancel", but cancel_charge does all we need. */
2645 __mem_cgroup_cancel_charge(from, nr_pages);
2646
2647 /* caller should have done css_get */
2648 pc->mem_cgroup = to;
2649 mem_cgroup_charge_statistics(to, anon, nr_pages);
2650 /*
2651 * We charges against "to" which may not have any tasks. Then, "to"
2652 * can be under rmdir(). But in current implementation, caller of
2653 * this function is just force_empty() and move charge, so it's
2654 * guaranteed that "to" is never removed. So, we don't check rmdir
2655 * status here.
2656 */
2657 move_unlock_mem_cgroup(from, &flags);
2658 ret = 0;
2659unlock:
2660 unlock_page_cgroup(pc);
2661 /*
2662 * check events
2663 */
2664 memcg_check_events(to, page);
2665 memcg_check_events(from, page);
2666out:
2667 return ret;
2668}
2669
2670/*
2671 * move charges to its parent.
2672 */
2673
2674static int mem_cgroup_move_parent(struct page *page,
2675 struct page_cgroup *pc,
2676 struct mem_cgroup *child,
2677 gfp_t gfp_mask)
2678{
2679 struct cgroup *cg = child->css.cgroup;
2680 struct cgroup *pcg = cg->parent;
2681 struct mem_cgroup *parent;
2682 unsigned int nr_pages;
2683 unsigned long uninitialized_var(flags);
2684 int ret;
2685
2686 /* Is ROOT ? */
2687 if (!pcg)
2688 return -EINVAL;
2689
2690 ret = -EBUSY;
2691 if (!get_page_unless_zero(page))
2692 goto out;
2693 if (isolate_lru_page(page))
2694 goto put;
2695
2696 nr_pages = hpage_nr_pages(page);
2697
2698 parent = mem_cgroup_from_cont(pcg);
2699 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2700 if (ret)
2701 goto put_back;
2702
2703 if (nr_pages > 1)
2704 flags = compound_lock_irqsave(page);
2705
2706 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2707 if (ret)
2708 __mem_cgroup_cancel_charge(parent, nr_pages);
2709
2710 if (nr_pages > 1)
2711 compound_unlock_irqrestore(page, flags);
2712put_back:
2713 putback_lru_page(page);
2714put:
2715 put_page(page);
2716out:
2717 return ret;
2718}
2719
2720/*
2721 * Charge the memory controller for page usage.
2722 * Return
2723 * 0 if the charge was successful
2724 * < 0 if the cgroup is over its limit
2725 */
2726static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2727 gfp_t gfp_mask, enum charge_type ctype)
2728{
2729 struct mem_cgroup *memcg = NULL;
2730 unsigned int nr_pages = 1;
2731 bool oom = true;
2732 int ret;
2733
2734 if (PageTransHuge(page)) {
2735 nr_pages <<= compound_order(page);
2736 VM_BUG_ON(!PageTransHuge(page));
2737 /*
2738 * Never OOM-kill a process for a huge page. The
2739 * fault handler will fall back to regular pages.
2740 */
2741 oom = false;
2742 }
2743
2744 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2745 if (ret == -ENOMEM)
2746 return ret;
2747 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2748 return 0;
2749}
2750
2751int mem_cgroup_newpage_charge(struct page *page,
2752 struct mm_struct *mm, gfp_t gfp_mask)
2753{
2754 if (mem_cgroup_disabled())
2755 return 0;
2756 VM_BUG_ON(page_mapped(page));
2757 VM_BUG_ON(page->mapping && !PageAnon(page));
2758 VM_BUG_ON(!mm);
2759 return mem_cgroup_charge_common(page, mm, gfp_mask,
2760 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2761}
2762
2763static void
2764__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2765 enum charge_type ctype);
2766
2767int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2768 gfp_t gfp_mask)
2769{
2770 struct mem_cgroup *memcg = NULL;
2771 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2772 int ret;
2773
2774 if (mem_cgroup_disabled())
2775 return 0;
2776 if (PageCompound(page))
2777 return 0;
2778
2779 if (unlikely(!mm))
2780 mm = &init_mm;
2781 if (!page_is_file_cache(page))
2782 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2783
2784 if (!PageSwapCache(page))
2785 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2786 else { /* page is swapcache/shmem */
2787 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2788 if (!ret)
2789 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2790 }
2791 return ret;
2792}
2793
2794/*
2795 * While swap-in, try_charge -> commit or cancel, the page is locked.
2796 * And when try_charge() successfully returns, one refcnt to memcg without
2797 * struct page_cgroup is acquired. This refcnt will be consumed by
2798 * "commit()" or removed by "cancel()"
2799 */
2800int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2801 struct page *page,
2802 gfp_t mask, struct mem_cgroup **memcgp)
2803{
2804 struct mem_cgroup *memcg;
2805 int ret;
2806
2807 *memcgp = NULL;
2808
2809 if (mem_cgroup_disabled())
2810 return 0;
2811
2812 if (!do_swap_account)
2813 goto charge_cur_mm;
2814 /*
2815 * A racing thread's fault, or swapoff, may have already updated
2816 * the pte, and even removed page from swap cache: in those cases
2817 * do_swap_page()'s pte_same() test will fail; but there's also a
2818 * KSM case which does need to charge the page.
2819 */
2820 if (!PageSwapCache(page))
2821 goto charge_cur_mm;
2822 memcg = try_get_mem_cgroup_from_page(page);
2823 if (!memcg)
2824 goto charge_cur_mm;
2825 *memcgp = memcg;
2826 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2827 css_put(&memcg->css);
2828 if (ret == -EINTR)
2829 ret = 0;
2830 return ret;
2831charge_cur_mm:
2832 if (unlikely(!mm))
2833 mm = &init_mm;
2834 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2835 if (ret == -EINTR)
2836 ret = 0;
2837 return ret;
2838}
2839
2840static void
2841__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2842 enum charge_type ctype)
2843{
2844 if (mem_cgroup_disabled())
2845 return;
2846 if (!memcg)
2847 return;
2848 cgroup_exclude_rmdir(&memcg->css);
2849
2850 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2851 /*
2852 * Now swap is on-memory. This means this page may be
2853 * counted both as mem and swap....double count.
2854 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2855 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2856 * may call delete_from_swap_cache() before reach here.
2857 */
2858 if (do_swap_account && PageSwapCache(page)) {
2859 swp_entry_t ent = {.val = page_private(page)};
2860 struct mem_cgroup *swap_memcg;
2861 unsigned short id;
2862
2863 id = swap_cgroup_record(ent, 0);
2864 rcu_read_lock();
2865 swap_memcg = mem_cgroup_lookup(id);
2866 if (swap_memcg) {
2867 /*
2868 * This recorded memcg can be obsolete one. So, avoid
2869 * calling css_tryget
2870 */
2871 if (!mem_cgroup_is_root(swap_memcg))
2872 res_counter_uncharge(&swap_memcg->memsw,
2873 PAGE_SIZE);
2874 mem_cgroup_swap_statistics(swap_memcg, false);
2875 mem_cgroup_put(swap_memcg);
2876 }
2877 rcu_read_unlock();
2878 }
2879 /*
2880 * At swapin, we may charge account against cgroup which has no tasks.
2881 * So, rmdir()->pre_destroy() can be called while we do this charge.
2882 * In that case, we need to call pre_destroy() again. check it here.
2883 */
2884 cgroup_release_and_wakeup_rmdir(&memcg->css);
2885}
2886
2887void mem_cgroup_commit_charge_swapin(struct page *page,
2888 struct mem_cgroup *memcg)
2889{
2890 __mem_cgroup_commit_charge_swapin(page, memcg,
2891 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2892}
2893
2894void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2895{
2896 if (mem_cgroup_disabled())
2897 return;
2898 if (!memcg)
2899 return;
2900 __mem_cgroup_cancel_charge(memcg, 1);
2901}
2902
2903static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2904 unsigned int nr_pages,
2905 const enum charge_type ctype)
2906{
2907 struct memcg_batch_info *batch = NULL;
2908 bool uncharge_memsw = true;
2909
2910 /* If swapout, usage of swap doesn't decrease */
2911 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2912 uncharge_memsw = false;
2913
2914 batch = &current->memcg_batch;
2915 /*
2916 * In usual, we do css_get() when we remember memcg pointer.
2917 * But in this case, we keep res->usage until end of a series of
2918 * uncharges. Then, it's ok to ignore memcg's refcnt.
2919 */
2920 if (!batch->memcg)
2921 batch->memcg = memcg;
2922 /*
2923 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2924 * In those cases, all pages freed continuously can be expected to be in
2925 * the same cgroup and we have chance to coalesce uncharges.
2926 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2927 * because we want to do uncharge as soon as possible.
2928 */
2929
2930 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2931 goto direct_uncharge;
2932
2933 if (nr_pages > 1)
2934 goto direct_uncharge;
2935
2936 /*
2937 * In typical case, batch->memcg == mem. This means we can
2938 * merge a series of uncharges to an uncharge of res_counter.
2939 * If not, we uncharge res_counter ony by one.
2940 */
2941 if (batch->memcg != memcg)
2942 goto direct_uncharge;
2943 /* remember freed charge and uncharge it later */
2944 batch->nr_pages++;
2945 if (uncharge_memsw)
2946 batch->memsw_nr_pages++;
2947 return;
2948direct_uncharge:
2949 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2950 if (uncharge_memsw)
2951 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2952 if (unlikely(batch->memcg != memcg))
2953 memcg_oom_recover(memcg);
2954}
2955
2956/*
2957 * uncharge if !page_mapped(page)
2958 */
2959static struct mem_cgroup *
2960__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2961{
2962 struct mem_cgroup *memcg = NULL;
2963 unsigned int nr_pages = 1;
2964 struct page_cgroup *pc;
2965 bool anon;
2966
2967 if (mem_cgroup_disabled())
2968 return NULL;
2969
2970 if (PageSwapCache(page))
2971 return NULL;
2972
2973 if (PageTransHuge(page)) {
2974 nr_pages <<= compound_order(page);
2975 VM_BUG_ON(!PageTransHuge(page));
2976 }
2977 /*
2978 * Check if our page_cgroup is valid
2979 */
2980 pc = lookup_page_cgroup(page);
2981 if (unlikely(!PageCgroupUsed(pc)))
2982 return NULL;
2983
2984 lock_page_cgroup(pc);
2985
2986 memcg = pc->mem_cgroup;
2987
2988 if (!PageCgroupUsed(pc))
2989 goto unlock_out;
2990
2991 anon = PageAnon(page);
2992
2993 switch (ctype) {
2994 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2995 /*
2996 * Generally PageAnon tells if it's the anon statistics to be
2997 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
2998 * used before page reached the stage of being marked PageAnon.
2999 */
3000 anon = true;
3001 /* fallthrough */
3002 case MEM_CGROUP_CHARGE_TYPE_DROP:
3003 /* See mem_cgroup_prepare_migration() */
3004 if (page_mapped(page) || PageCgroupMigration(pc))
3005 goto unlock_out;
3006 break;
3007 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3008 if (!PageAnon(page)) { /* Shared memory */
3009 if (page->mapping && !page_is_file_cache(page))
3010 goto unlock_out;
3011 } else if (page_mapped(page)) /* Anon */
3012 goto unlock_out;
3013 break;
3014 default:
3015 break;
3016 }
3017
3018 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3019
3020 ClearPageCgroupUsed(pc);
3021 /*
3022 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3023 * freed from LRU. This is safe because uncharged page is expected not
3024 * to be reused (freed soon). Exception is SwapCache, it's handled by
3025 * special functions.
3026 */
3027
3028 unlock_page_cgroup(pc);
3029 /*
3030 * even after unlock, we have memcg->res.usage here and this memcg
3031 * will never be freed.
3032 */
3033 memcg_check_events(memcg, page);
3034 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3035 mem_cgroup_swap_statistics(memcg, true);
3036 mem_cgroup_get(memcg);
3037 }
3038 if (!mem_cgroup_is_root(memcg))
3039 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3040
3041 return memcg;
3042
3043unlock_out:
3044 unlock_page_cgroup(pc);
3045 return NULL;
3046}
3047
3048void mem_cgroup_uncharge_page(struct page *page)
3049{
3050 /* early check. */
3051 if (page_mapped(page))
3052 return;
3053 VM_BUG_ON(page->mapping && !PageAnon(page));
3054 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3055}
3056
3057void mem_cgroup_uncharge_cache_page(struct page *page)
3058{
3059 VM_BUG_ON(page_mapped(page));
3060 VM_BUG_ON(page->mapping);
3061 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3062}
3063
3064/*
3065 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3066 * In that cases, pages are freed continuously and we can expect pages
3067 * are in the same memcg. All these calls itself limits the number of
3068 * pages freed at once, then uncharge_start/end() is called properly.
3069 * This may be called prural(2) times in a context,
3070 */
3071
3072void mem_cgroup_uncharge_start(void)
3073{
3074 current->memcg_batch.do_batch++;
3075 /* We can do nest. */
3076 if (current->memcg_batch.do_batch == 1) {
3077 current->memcg_batch.memcg = NULL;
3078 current->memcg_batch.nr_pages = 0;
3079 current->memcg_batch.memsw_nr_pages = 0;
3080 }
3081}
3082
3083void mem_cgroup_uncharge_end(void)
3084{
3085 struct memcg_batch_info *batch = &current->memcg_batch;
3086
3087 if (!batch->do_batch)
3088 return;
3089
3090 batch->do_batch--;
3091 if (batch->do_batch) /* If stacked, do nothing. */
3092 return;
3093
3094 if (!batch->memcg)
3095 return;
3096 /*
3097 * This "batch->memcg" is valid without any css_get/put etc...
3098 * bacause we hide charges behind us.
3099 */
3100 if (batch->nr_pages)
3101 res_counter_uncharge(&batch->memcg->res,
3102 batch->nr_pages * PAGE_SIZE);
3103 if (batch->memsw_nr_pages)
3104 res_counter_uncharge(&batch->memcg->memsw,
3105 batch->memsw_nr_pages * PAGE_SIZE);
3106 memcg_oom_recover(batch->memcg);
3107 /* forget this pointer (for sanity check) */
3108 batch->memcg = NULL;
3109}
3110
3111#ifdef CONFIG_SWAP
3112/*
3113 * called after __delete_from_swap_cache() and drop "page" account.
3114 * memcg information is recorded to swap_cgroup of "ent"
3115 */
3116void
3117mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3118{
3119 struct mem_cgroup *memcg;
3120 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3121
3122 if (!swapout) /* this was a swap cache but the swap is unused ! */
3123 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3124
3125 memcg = __mem_cgroup_uncharge_common(page, ctype);
3126
3127 /*
3128 * record memcg information, if swapout && memcg != NULL,
3129 * mem_cgroup_get() was called in uncharge().
3130 */
3131 if (do_swap_account && swapout && memcg)
3132 swap_cgroup_record(ent, css_id(&memcg->css));
3133}
3134#endif
3135
3136#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3137/*
3138 * called from swap_entry_free(). remove record in swap_cgroup and
3139 * uncharge "memsw" account.
3140 */
3141void mem_cgroup_uncharge_swap(swp_entry_t ent)
3142{
3143 struct mem_cgroup *memcg;
3144 unsigned short id;
3145
3146 if (!do_swap_account)
3147 return;
3148
3149 id = swap_cgroup_record(ent, 0);
3150 rcu_read_lock();
3151 memcg = mem_cgroup_lookup(id);
3152 if (memcg) {
3153 /*
3154 * We uncharge this because swap is freed.
3155 * This memcg can be obsolete one. We avoid calling css_tryget
3156 */
3157 if (!mem_cgroup_is_root(memcg))
3158 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3159 mem_cgroup_swap_statistics(memcg, false);
3160 mem_cgroup_put(memcg);
3161 }
3162 rcu_read_unlock();
3163}
3164
3165/**
3166 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3167 * @entry: swap entry to be moved
3168 * @from: mem_cgroup which the entry is moved from
3169 * @to: mem_cgroup which the entry is moved to
3170 * @need_fixup: whether we should fixup res_counters and refcounts.
3171 *
3172 * It succeeds only when the swap_cgroup's record for this entry is the same
3173 * as the mem_cgroup's id of @from.
3174 *
3175 * Returns 0 on success, -EINVAL on failure.
3176 *
3177 * The caller must have charged to @to, IOW, called res_counter_charge() about
3178 * both res and memsw, and called css_get().
3179 */
3180static int mem_cgroup_move_swap_account(swp_entry_t entry,
3181 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3182{
3183 unsigned short old_id, new_id;
3184
3185 old_id = css_id(&from->css);
3186 new_id = css_id(&to->css);
3187
3188 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3189 mem_cgroup_swap_statistics(from, false);
3190 mem_cgroup_swap_statistics(to, true);
3191 /*
3192 * This function is only called from task migration context now.
3193 * It postpones res_counter and refcount handling till the end
3194 * of task migration(mem_cgroup_clear_mc()) for performance
3195 * improvement. But we cannot postpone mem_cgroup_get(to)
3196 * because if the process that has been moved to @to does
3197 * swap-in, the refcount of @to might be decreased to 0.
3198 */
3199 mem_cgroup_get(to);
3200 if (need_fixup) {
3201 if (!mem_cgroup_is_root(from))
3202 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3203 mem_cgroup_put(from);
3204 /*
3205 * we charged both to->res and to->memsw, so we should
3206 * uncharge to->res.
3207 */
3208 if (!mem_cgroup_is_root(to))
3209 res_counter_uncharge(&to->res, PAGE_SIZE);
3210 }
3211 return 0;
3212 }
3213 return -EINVAL;
3214}
3215#else
3216static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3217 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3218{
3219 return -EINVAL;
3220}
3221#endif
3222
3223/*
3224 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3225 * page belongs to.
3226 */
3227int mem_cgroup_prepare_migration(struct page *page,
3228 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3229{
3230 struct mem_cgroup *memcg = NULL;
3231 struct page_cgroup *pc;
3232 enum charge_type ctype;
3233 int ret = 0;
3234
3235 *memcgp = NULL;
3236
3237 VM_BUG_ON(PageTransHuge(page));
3238 if (mem_cgroup_disabled())
3239 return 0;
3240
3241 pc = lookup_page_cgroup(page);
3242 lock_page_cgroup(pc);
3243 if (PageCgroupUsed(pc)) {
3244 memcg = pc->mem_cgroup;
3245 css_get(&memcg->css);
3246 /*
3247 * At migrating an anonymous page, its mapcount goes down
3248 * to 0 and uncharge() will be called. But, even if it's fully
3249 * unmapped, migration may fail and this page has to be
3250 * charged again. We set MIGRATION flag here and delay uncharge
3251 * until end_migration() is called
3252 *
3253 * Corner Case Thinking
3254 * A)
3255 * When the old page was mapped as Anon and it's unmap-and-freed
3256 * while migration was ongoing.
3257 * If unmap finds the old page, uncharge() of it will be delayed
3258 * until end_migration(). If unmap finds a new page, it's
3259 * uncharged when it make mapcount to be 1->0. If unmap code
3260 * finds swap_migration_entry, the new page will not be mapped
3261 * and end_migration() will find it(mapcount==0).
3262 *
3263 * B)
3264 * When the old page was mapped but migraion fails, the kernel
3265 * remaps it. A charge for it is kept by MIGRATION flag even
3266 * if mapcount goes down to 0. We can do remap successfully
3267 * without charging it again.
3268 *
3269 * C)
3270 * The "old" page is under lock_page() until the end of
3271 * migration, so, the old page itself will not be swapped-out.
3272 * If the new page is swapped out before end_migraton, our
3273 * hook to usual swap-out path will catch the event.
3274 */
3275 if (PageAnon(page))
3276 SetPageCgroupMigration(pc);
3277 }
3278 unlock_page_cgroup(pc);
3279 /*
3280 * If the page is not charged at this point,
3281 * we return here.
3282 */
3283 if (!memcg)
3284 return 0;
3285
3286 *memcgp = memcg;
3287 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3288 css_put(&memcg->css);/* drop extra refcnt */
3289 if (ret) {
3290 if (PageAnon(page)) {
3291 lock_page_cgroup(pc);
3292 ClearPageCgroupMigration(pc);
3293 unlock_page_cgroup(pc);
3294 /*
3295 * The old page may be fully unmapped while we kept it.
3296 */
3297 mem_cgroup_uncharge_page(page);
3298 }
3299 /* we'll need to revisit this error code (we have -EINTR) */
3300 return -ENOMEM;
3301 }
3302 /*
3303 * We charge new page before it's used/mapped. So, even if unlock_page()
3304 * is called before end_migration, we can catch all events on this new
3305 * page. In the case new page is migrated but not remapped, new page's
3306 * mapcount will be finally 0 and we call uncharge in end_migration().
3307 */
3308 if (PageAnon(page))
3309 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3310 else if (page_is_file_cache(page))
3311 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3312 else
3313 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3314 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3315 return ret;
3316}
3317
3318/* remove redundant charge if migration failed*/
3319void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3320 struct page *oldpage, struct page *newpage, bool migration_ok)
3321{
3322 struct page *used, *unused;
3323 struct page_cgroup *pc;
3324 bool anon;
3325
3326 if (!memcg)
3327 return;
3328 /* blocks rmdir() */
3329 cgroup_exclude_rmdir(&memcg->css);
3330 if (!migration_ok) {
3331 used = oldpage;
3332 unused = newpage;
3333 } else {
3334 used = newpage;
3335 unused = oldpage;
3336 }
3337 /*
3338 * We disallowed uncharge of pages under migration because mapcount
3339 * of the page goes down to zero, temporarly.
3340 * Clear the flag and check the page should be charged.
3341 */
3342 pc = lookup_page_cgroup(oldpage);
3343 lock_page_cgroup(pc);
3344 ClearPageCgroupMigration(pc);
3345 unlock_page_cgroup(pc);
3346 anon = PageAnon(used);
3347 __mem_cgroup_uncharge_common(unused,
3348 anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED
3349 : MEM_CGROUP_CHARGE_TYPE_CACHE);
3350
3351 /*
3352 * If a page is a file cache, radix-tree replacement is very atomic
3353 * and we can skip this check. When it was an Anon page, its mapcount
3354 * goes down to 0. But because we added MIGRATION flage, it's not
3355 * uncharged yet. There are several case but page->mapcount check
3356 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3357 * check. (see prepare_charge() also)
3358 */
3359 if (anon)
3360 mem_cgroup_uncharge_page(used);
3361 /*
3362 * At migration, we may charge account against cgroup which has no
3363 * tasks.
3364 * So, rmdir()->pre_destroy() can be called while we do this charge.
3365 * In that case, we need to call pre_destroy() again. check it here.
3366 */
3367 cgroup_release_and_wakeup_rmdir(&memcg->css);
3368}
3369
3370/*
3371 * At replace page cache, newpage is not under any memcg but it's on
3372 * LRU. So, this function doesn't touch res_counter but handles LRU
3373 * in correct way. Both pages are locked so we cannot race with uncharge.
3374 */
3375void mem_cgroup_replace_page_cache(struct page *oldpage,
3376 struct page *newpage)
3377{
3378 struct mem_cgroup *memcg;
3379 struct page_cgroup *pc;
3380 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3381
3382 if (mem_cgroup_disabled())
3383 return;
3384
3385 pc = lookup_page_cgroup(oldpage);
3386 /* fix accounting on old pages */
3387 lock_page_cgroup(pc);
3388 memcg = pc->mem_cgroup;
3389 mem_cgroup_charge_statistics(memcg, false, -1);
3390 ClearPageCgroupUsed(pc);
3391 unlock_page_cgroup(pc);
3392
3393 if (PageSwapBacked(oldpage))
3394 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3395
3396 /*
3397 * Even if newpage->mapping was NULL before starting replacement,
3398 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3399 * LRU while we overwrite pc->mem_cgroup.
3400 */
3401 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3402}
3403
3404#ifdef CONFIG_DEBUG_VM
3405static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3406{
3407 struct page_cgroup *pc;
3408
3409 pc = lookup_page_cgroup(page);
3410 /*
3411 * Can be NULL while feeding pages into the page allocator for
3412 * the first time, i.e. during boot or memory hotplug;
3413 * or when mem_cgroup_disabled().
3414 */
3415 if (likely(pc) && PageCgroupUsed(pc))
3416 return pc;
3417 return NULL;
3418}
3419
3420bool mem_cgroup_bad_page_check(struct page *page)
3421{
3422 if (mem_cgroup_disabled())
3423 return false;
3424
3425 return lookup_page_cgroup_used(page) != NULL;
3426}
3427
3428void mem_cgroup_print_bad_page(struct page *page)
3429{
3430 struct page_cgroup *pc;
3431
3432 pc = lookup_page_cgroup_used(page);
3433 if (pc) {
3434 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3435 pc, pc->flags, pc->mem_cgroup);
3436 }
3437}
3438#endif
3439
3440static DEFINE_MUTEX(set_limit_mutex);
3441
3442static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3443 unsigned long long val)
3444{
3445 int retry_count;
3446 u64 memswlimit, memlimit;
3447 int ret = 0;
3448 int children = mem_cgroup_count_children(memcg);
3449 u64 curusage, oldusage;
3450 int enlarge;
3451
3452 /*
3453 * For keeping hierarchical_reclaim simple, how long we should retry
3454 * is depends on callers. We set our retry-count to be function
3455 * of # of children which we should visit in this loop.
3456 */
3457 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3458
3459 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3460
3461 enlarge = 0;
3462 while (retry_count) {
3463 if (signal_pending(current)) {
3464 ret = -EINTR;
3465 break;
3466 }
3467 /*
3468 * Rather than hide all in some function, I do this in
3469 * open coded manner. You see what this really does.
3470 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3471 */
3472 mutex_lock(&set_limit_mutex);
3473 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3474 if (memswlimit < val) {
3475 ret = -EINVAL;
3476 mutex_unlock(&set_limit_mutex);
3477 break;
3478 }
3479
3480 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3481 if (memlimit < val)
3482 enlarge = 1;
3483
3484 ret = res_counter_set_limit(&memcg->res, val);
3485 if (!ret) {
3486 if (memswlimit == val)
3487 memcg->memsw_is_minimum = true;
3488 else
3489 memcg->memsw_is_minimum = false;
3490 }
3491 mutex_unlock(&set_limit_mutex);
3492
3493 if (!ret)
3494 break;
3495
3496 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3497 MEM_CGROUP_RECLAIM_SHRINK);
3498 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3499 /* Usage is reduced ? */
3500 if (curusage >= oldusage)
3501 retry_count--;
3502 else
3503 oldusage = curusage;
3504 }
3505 if (!ret && enlarge)
3506 memcg_oom_recover(memcg);
3507
3508 return ret;
3509}
3510
3511static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3512 unsigned long long val)
3513{
3514 int retry_count;
3515 u64 memlimit, memswlimit, oldusage, curusage;
3516 int children = mem_cgroup_count_children(memcg);
3517 int ret = -EBUSY;
3518 int enlarge = 0;
3519
3520 /* see mem_cgroup_resize_res_limit */
3521 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3522 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3523 while (retry_count) {
3524 if (signal_pending(current)) {
3525 ret = -EINTR;
3526 break;
3527 }
3528 /*
3529 * Rather than hide all in some function, I do this in
3530 * open coded manner. You see what this really does.
3531 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3532 */
3533 mutex_lock(&set_limit_mutex);
3534 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3535 if (memlimit > val) {
3536 ret = -EINVAL;
3537 mutex_unlock(&set_limit_mutex);
3538 break;
3539 }
3540 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3541 if (memswlimit < val)
3542 enlarge = 1;
3543 ret = res_counter_set_limit(&memcg->memsw, val);
3544 if (!ret) {
3545 if (memlimit == val)
3546 memcg->memsw_is_minimum = true;
3547 else
3548 memcg->memsw_is_minimum = false;
3549 }
3550 mutex_unlock(&set_limit_mutex);
3551
3552 if (!ret)
3553 break;
3554
3555 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3556 MEM_CGROUP_RECLAIM_NOSWAP |
3557 MEM_CGROUP_RECLAIM_SHRINK);
3558 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3559 /* Usage is reduced ? */
3560 if (curusage >= oldusage)
3561 retry_count--;
3562 else
3563 oldusage = curusage;
3564 }
3565 if (!ret && enlarge)
3566 memcg_oom_recover(memcg);
3567 return ret;
3568}
3569
3570unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3571 gfp_t gfp_mask,
3572 unsigned long *total_scanned)
3573{
3574 unsigned long nr_reclaimed = 0;
3575 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3576 unsigned long reclaimed;
3577 int loop = 0;
3578 struct mem_cgroup_tree_per_zone *mctz;
3579 unsigned long long excess;
3580 unsigned long nr_scanned;
3581
3582 if (order > 0)
3583 return 0;
3584
3585 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3586 /*
3587 * This loop can run a while, specially if mem_cgroup's continuously
3588 * keep exceeding their soft limit and putting the system under
3589 * pressure
3590 */
3591 do {
3592 if (next_mz)
3593 mz = next_mz;
3594 else
3595 mz = mem_cgroup_largest_soft_limit_node(mctz);
3596 if (!mz)
3597 break;
3598
3599 nr_scanned = 0;
3600 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3601 gfp_mask, &nr_scanned);
3602 nr_reclaimed += reclaimed;
3603 *total_scanned += nr_scanned;
3604 spin_lock(&mctz->lock);
3605
3606 /*
3607 * If we failed to reclaim anything from this memory cgroup
3608 * it is time to move on to the next cgroup
3609 */
3610 next_mz = NULL;
3611 if (!reclaimed) {
3612 do {
3613 /*
3614 * Loop until we find yet another one.
3615 *
3616 * By the time we get the soft_limit lock
3617 * again, someone might have aded the
3618 * group back on the RB tree. Iterate to
3619 * make sure we get a different mem.
3620 * mem_cgroup_largest_soft_limit_node returns
3621 * NULL if no other cgroup is present on
3622 * the tree
3623 */
3624 next_mz =
3625 __mem_cgroup_largest_soft_limit_node(mctz);
3626 if (next_mz == mz)
3627 css_put(&next_mz->memcg->css);
3628 else /* next_mz == NULL or other memcg */
3629 break;
3630 } while (1);
3631 }
3632 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3633 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3634 /*
3635 * One school of thought says that we should not add
3636 * back the node to the tree if reclaim returns 0.
3637 * But our reclaim could return 0, simply because due
3638 * to priority we are exposing a smaller subset of
3639 * memory to reclaim from. Consider this as a longer
3640 * term TODO.
3641 */
3642 /* If excess == 0, no tree ops */
3643 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3644 spin_unlock(&mctz->lock);
3645 css_put(&mz->memcg->css);
3646 loop++;
3647 /*
3648 * Could not reclaim anything and there are no more
3649 * mem cgroups to try or we seem to be looping without
3650 * reclaiming anything.
3651 */
3652 if (!nr_reclaimed &&
3653 (next_mz == NULL ||
3654 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3655 break;
3656 } while (!nr_reclaimed);
3657 if (next_mz)
3658 css_put(&next_mz->memcg->css);
3659 return nr_reclaimed;
3660}
3661
3662/*
3663 * This routine traverse page_cgroup in given list and drop them all.
3664 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3665 */
3666static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3667 int node, int zid, enum lru_list lru)
3668{
3669 struct mem_cgroup_per_zone *mz;
3670 unsigned long flags, loop;
3671 struct list_head *list;
3672 struct page *busy;
3673 struct zone *zone;
3674 int ret = 0;
3675
3676 zone = &NODE_DATA(node)->node_zones[zid];
3677 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3678 list = &mz->lruvec.lists[lru];
3679
3680 loop = mz->lru_size[lru];
3681 /* give some margin against EBUSY etc...*/
3682 loop += 256;
3683 busy = NULL;
3684 while (loop--) {
3685 struct page_cgroup *pc;
3686 struct page *page;
3687
3688 ret = 0;
3689 spin_lock_irqsave(&zone->lru_lock, flags);
3690 if (list_empty(list)) {
3691 spin_unlock_irqrestore(&zone->lru_lock, flags);
3692 break;
3693 }
3694 page = list_entry(list->prev, struct page, lru);
3695 if (busy == page) {
3696 list_move(&page->lru, list);
3697 busy = NULL;
3698 spin_unlock_irqrestore(&zone->lru_lock, flags);
3699 continue;
3700 }
3701 spin_unlock_irqrestore(&zone->lru_lock, flags);
3702
3703 pc = lookup_page_cgroup(page);
3704
3705 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3706 if (ret == -ENOMEM || ret == -EINTR)
3707 break;
3708
3709 if (ret == -EBUSY || ret == -EINVAL) {
3710 /* found lock contention or "pc" is obsolete. */
3711 busy = page;
3712 cond_resched();
3713 } else
3714 busy = NULL;
3715 }
3716
3717 if (!ret && !list_empty(list))
3718 return -EBUSY;
3719 return ret;
3720}
3721
3722/*
3723 * make mem_cgroup's charge to be 0 if there is no task.
3724 * This enables deleting this mem_cgroup.
3725 */
3726static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3727{
3728 int ret;
3729 int node, zid, shrink;
3730 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3731 struct cgroup *cgrp = memcg->css.cgroup;
3732
3733 css_get(&memcg->css);
3734
3735 shrink = 0;
3736 /* should free all ? */
3737 if (free_all)
3738 goto try_to_free;
3739move_account:
3740 do {
3741 ret = -EBUSY;
3742 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3743 goto out;
3744 ret = -EINTR;
3745 if (signal_pending(current))
3746 goto out;
3747 /* This is for making all *used* pages to be on LRU. */
3748 lru_add_drain_all();
3749 drain_all_stock_sync(memcg);
3750 ret = 0;
3751 mem_cgroup_start_move(memcg);
3752 for_each_node_state(node, N_HIGH_MEMORY) {
3753 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3754 enum lru_list lru;
3755 for_each_lru(lru) {
3756 ret = mem_cgroup_force_empty_list(memcg,
3757 node, zid, lru);
3758 if (ret)
3759 break;
3760 }
3761 }
3762 if (ret)
3763 break;
3764 }
3765 mem_cgroup_end_move(memcg);
3766 memcg_oom_recover(memcg);
3767 /* it seems parent cgroup doesn't have enough mem */
3768 if (ret == -ENOMEM)
3769 goto try_to_free;
3770 cond_resched();
3771 /* "ret" should also be checked to ensure all lists are empty. */
3772 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3773out:
3774 css_put(&memcg->css);
3775 return ret;
3776
3777try_to_free:
3778 /* returns EBUSY if there is a task or if we come here twice. */
3779 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3780 ret = -EBUSY;
3781 goto out;
3782 }
3783 /* we call try-to-free pages for make this cgroup empty */
3784 lru_add_drain_all();
3785 /* try to free all pages in this cgroup */
3786 shrink = 1;
3787 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3788 int progress;
3789
3790 if (signal_pending(current)) {
3791 ret = -EINTR;
3792 goto out;
3793 }
3794 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3795 false);
3796 if (!progress) {
3797 nr_retries--;
3798 /* maybe some writeback is necessary */
3799 congestion_wait(BLK_RW_ASYNC, HZ/10);
3800 }
3801
3802 }
3803 lru_add_drain();
3804 /* try move_account...there may be some *locked* pages. */
3805 goto move_account;
3806}
3807
3808int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3809{
3810 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3811}
3812
3813
3814static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3815{
3816 return mem_cgroup_from_cont(cont)->use_hierarchy;
3817}
3818
3819static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3820 u64 val)
3821{
3822 int retval = 0;
3823 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3824 struct cgroup *parent = cont->parent;
3825 struct mem_cgroup *parent_memcg = NULL;
3826
3827 if (parent)
3828 parent_memcg = mem_cgroup_from_cont(parent);
3829
3830 cgroup_lock();
3831 /*
3832 * If parent's use_hierarchy is set, we can't make any modifications
3833 * in the child subtrees. If it is unset, then the change can
3834 * occur, provided the current cgroup has no children.
3835 *
3836 * For the root cgroup, parent_mem is NULL, we allow value to be
3837 * set if there are no children.
3838 */
3839 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3840 (val == 1 || val == 0)) {
3841 if (list_empty(&cont->children))
3842 memcg->use_hierarchy = val;
3843 else
3844 retval = -EBUSY;
3845 } else
3846 retval = -EINVAL;
3847 cgroup_unlock();
3848
3849 return retval;
3850}
3851
3852
3853static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3854 enum mem_cgroup_stat_index idx)
3855{
3856 struct mem_cgroup *iter;
3857 long val = 0;
3858
3859 /* Per-cpu values can be negative, use a signed accumulator */
3860 for_each_mem_cgroup_tree(iter, memcg)
3861 val += mem_cgroup_read_stat(iter, idx);
3862
3863 if (val < 0) /* race ? */
3864 val = 0;
3865 return val;
3866}
3867
3868static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3869{
3870 u64 val;
3871
3872 if (!mem_cgroup_is_root(memcg)) {
3873 if (!swap)
3874 return res_counter_read_u64(&memcg->res, RES_USAGE);
3875 else
3876 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3877 }
3878
3879 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3880 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3881
3882 if (swap)
3883 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3884
3885 return val << PAGE_SHIFT;
3886}
3887
3888static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3889{
3890 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3891 u64 val;
3892 int type, name;
3893
3894 type = MEMFILE_TYPE(cft->private);
3895 name = MEMFILE_ATTR(cft->private);
3896 switch (type) {
3897 case _MEM:
3898 if (name == RES_USAGE)
3899 val = mem_cgroup_usage(memcg, false);
3900 else
3901 val = res_counter_read_u64(&memcg->res, name);
3902 break;
3903 case _MEMSWAP:
3904 if (name == RES_USAGE)
3905 val = mem_cgroup_usage(memcg, true);
3906 else
3907 val = res_counter_read_u64(&memcg->memsw, name);
3908 break;
3909 default:
3910 BUG();
3911 }
3912 return val;
3913}
3914/*
3915 * The user of this function is...
3916 * RES_LIMIT.
3917 */
3918static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3919 const char *buffer)
3920{
3921 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3922 int type, name;
3923 unsigned long long val;
3924 int ret;
3925
3926 type = MEMFILE_TYPE(cft->private);
3927 name = MEMFILE_ATTR(cft->private);
3928 switch (name) {
3929 case RES_LIMIT:
3930 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3931 ret = -EINVAL;
3932 break;
3933 }
3934 /* This function does all necessary parse...reuse it */
3935 ret = res_counter_memparse_write_strategy(buffer, &val);
3936 if (ret)
3937 break;
3938 if (type == _MEM)
3939 ret = mem_cgroup_resize_limit(memcg, val);
3940 else
3941 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3942 break;
3943 case RES_SOFT_LIMIT:
3944 ret = res_counter_memparse_write_strategy(buffer, &val);
3945 if (ret)
3946 break;
3947 /*
3948 * For memsw, soft limits are hard to implement in terms
3949 * of semantics, for now, we support soft limits for
3950 * control without swap
3951 */
3952 if (type == _MEM)
3953 ret = res_counter_set_soft_limit(&memcg->res, val);
3954 else
3955 ret = -EINVAL;
3956 break;
3957 default:
3958 ret = -EINVAL; /* should be BUG() ? */
3959 break;
3960 }
3961 return ret;
3962}
3963
3964static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3965 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3966{
3967 struct cgroup *cgroup;
3968 unsigned long long min_limit, min_memsw_limit, tmp;
3969
3970 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3971 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3972 cgroup = memcg->css.cgroup;
3973 if (!memcg->use_hierarchy)
3974 goto out;
3975
3976 while (cgroup->parent) {
3977 cgroup = cgroup->parent;
3978 memcg = mem_cgroup_from_cont(cgroup);
3979 if (!memcg->use_hierarchy)
3980 break;
3981 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3982 min_limit = min(min_limit, tmp);
3983 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3984 min_memsw_limit = min(min_memsw_limit, tmp);
3985 }
3986out:
3987 *mem_limit = min_limit;
3988 *memsw_limit = min_memsw_limit;
3989}
3990
3991static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3992{
3993 struct mem_cgroup *memcg;
3994 int type, name;
3995
3996 memcg = mem_cgroup_from_cont(cont);
3997 type = MEMFILE_TYPE(event);
3998 name = MEMFILE_ATTR(event);
3999 switch (name) {
4000 case RES_MAX_USAGE:
4001 if (type == _MEM)
4002 res_counter_reset_max(&memcg->res);
4003 else
4004 res_counter_reset_max(&memcg->memsw);
4005 break;
4006 case RES_FAILCNT:
4007 if (type == _MEM)
4008 res_counter_reset_failcnt(&memcg->res);
4009 else
4010 res_counter_reset_failcnt(&memcg->memsw);
4011 break;
4012 }
4013
4014 return 0;
4015}
4016
4017static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4018 struct cftype *cft)
4019{
4020 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4021}
4022
4023#ifdef CONFIG_MMU
4024static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4025 struct cftype *cft, u64 val)
4026{
4027 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4028
4029 if (val >= (1 << NR_MOVE_TYPE))
4030 return -EINVAL;
4031 /*
4032 * We check this value several times in both in can_attach() and
4033 * attach(), so we need cgroup lock to prevent this value from being
4034 * inconsistent.
4035 */
4036 cgroup_lock();
4037 memcg->move_charge_at_immigrate = val;
4038 cgroup_unlock();
4039
4040 return 0;
4041}
4042#else
4043static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4044 struct cftype *cft, u64 val)
4045{
4046 return -ENOSYS;
4047}
4048#endif
4049
4050
4051/* For read statistics */
4052enum {
4053 MCS_CACHE,
4054 MCS_RSS,
4055 MCS_FILE_MAPPED,
4056 MCS_PGPGIN,
4057 MCS_PGPGOUT,
4058 MCS_SWAP,
4059 MCS_PGFAULT,
4060 MCS_PGMAJFAULT,
4061 MCS_INACTIVE_ANON,
4062 MCS_ACTIVE_ANON,
4063 MCS_INACTIVE_FILE,
4064 MCS_ACTIVE_FILE,
4065 MCS_UNEVICTABLE,
4066 NR_MCS_STAT,
4067};
4068
4069struct mcs_total_stat {
4070 s64 stat[NR_MCS_STAT];
4071};
4072
4073struct {
4074 char *local_name;
4075 char *total_name;
4076} memcg_stat_strings[NR_MCS_STAT] = {
4077 {"cache", "total_cache"},
4078 {"rss", "total_rss"},
4079 {"mapped_file", "total_mapped_file"},
4080 {"pgpgin", "total_pgpgin"},
4081 {"pgpgout", "total_pgpgout"},
4082 {"swap", "total_swap"},
4083 {"pgfault", "total_pgfault"},
4084 {"pgmajfault", "total_pgmajfault"},
4085 {"inactive_anon", "total_inactive_anon"},
4086 {"active_anon", "total_active_anon"},
4087 {"inactive_file", "total_inactive_file"},
4088 {"active_file", "total_active_file"},
4089 {"unevictable", "total_unevictable"}
4090};
4091
4092
4093static void
4094mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4095{
4096 s64 val;
4097
4098 /* per cpu stat */
4099 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4100 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4101 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4102 s->stat[MCS_RSS] += val * PAGE_SIZE;
4103 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4104 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4105 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4106 s->stat[MCS_PGPGIN] += val;
4107 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4108 s->stat[MCS_PGPGOUT] += val;
4109 if (do_swap_account) {
4110 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4111 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4112 }
4113 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4114 s->stat[MCS_PGFAULT] += val;
4115 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4116 s->stat[MCS_PGMAJFAULT] += val;
4117
4118 /* per zone stat */
4119 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4120 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4121 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4122 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4123 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4124 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4125 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4126 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4127 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4128 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4129}
4130
4131static void
4132mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4133{
4134 struct mem_cgroup *iter;
4135
4136 for_each_mem_cgroup_tree(iter, memcg)
4137 mem_cgroup_get_local_stat(iter, s);
4138}
4139
4140#ifdef CONFIG_NUMA
4141static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4142{
4143 int nid;
4144 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4145 unsigned long node_nr;
4146 struct cgroup *cont = m->private;
4147 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4148
4149 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4150 seq_printf(m, "total=%lu", total_nr);
4151 for_each_node_state(nid, N_HIGH_MEMORY) {
4152 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4153 seq_printf(m, " N%d=%lu", nid, node_nr);
4154 }
4155 seq_putc(m, '\n');
4156
4157 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4158 seq_printf(m, "file=%lu", file_nr);
4159 for_each_node_state(nid, N_HIGH_MEMORY) {
4160 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4161 LRU_ALL_FILE);
4162 seq_printf(m, " N%d=%lu", nid, node_nr);
4163 }
4164 seq_putc(m, '\n');
4165
4166 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4167 seq_printf(m, "anon=%lu", anon_nr);
4168 for_each_node_state(nid, N_HIGH_MEMORY) {
4169 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4170 LRU_ALL_ANON);
4171 seq_printf(m, " N%d=%lu", nid, node_nr);
4172 }
4173 seq_putc(m, '\n');
4174
4175 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4176 seq_printf(m, "unevictable=%lu", unevictable_nr);
4177 for_each_node_state(nid, N_HIGH_MEMORY) {
4178 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4179 BIT(LRU_UNEVICTABLE));
4180 seq_printf(m, " N%d=%lu", nid, node_nr);
4181 }
4182 seq_putc(m, '\n');
4183 return 0;
4184}
4185#endif /* CONFIG_NUMA */
4186
4187static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4188 struct cgroup_map_cb *cb)
4189{
4190 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4191 struct mcs_total_stat mystat;
4192 int i;
4193
4194 memset(&mystat, 0, sizeof(mystat));
4195 mem_cgroup_get_local_stat(memcg, &mystat);
4196
4197
4198 for (i = 0; i < NR_MCS_STAT; i++) {
4199 if (i == MCS_SWAP && !do_swap_account)
4200 continue;
4201 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4202 }
4203
4204 /* Hierarchical information */
4205 {
4206 unsigned long long limit, memsw_limit;
4207 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4208 cb->fill(cb, "hierarchical_memory_limit", limit);
4209 if (do_swap_account)
4210 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4211 }
4212
4213 memset(&mystat, 0, sizeof(mystat));
4214 mem_cgroup_get_total_stat(memcg, &mystat);
4215 for (i = 0; i < NR_MCS_STAT; i++) {
4216 if (i == MCS_SWAP && !do_swap_account)
4217 continue;
4218 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4219 }
4220
4221#ifdef CONFIG_DEBUG_VM
4222 {
4223 int nid, zid;
4224 struct mem_cgroup_per_zone *mz;
4225 unsigned long recent_rotated[2] = {0, 0};
4226 unsigned long recent_scanned[2] = {0, 0};
4227
4228 for_each_online_node(nid)
4229 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4230 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4231
4232 recent_rotated[0] +=
4233 mz->reclaim_stat.recent_rotated[0];
4234 recent_rotated[1] +=
4235 mz->reclaim_stat.recent_rotated[1];
4236 recent_scanned[0] +=
4237 mz->reclaim_stat.recent_scanned[0];
4238 recent_scanned[1] +=
4239 mz->reclaim_stat.recent_scanned[1];
4240 }
4241 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4242 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4243 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4244 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4245 }
4246#endif
4247
4248 return 0;
4249}
4250
4251static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4252{
4253 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4254
4255 return mem_cgroup_swappiness(memcg);
4256}
4257
4258static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4259 u64 val)
4260{
4261 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4262 struct mem_cgroup *parent;
4263
4264 if (val > 100)
4265 return -EINVAL;
4266
4267 if (cgrp->parent == NULL)
4268 return -EINVAL;
4269
4270 parent = mem_cgroup_from_cont(cgrp->parent);
4271
4272 cgroup_lock();
4273
4274 /* If under hierarchy, only empty-root can set this value */
4275 if ((parent->use_hierarchy) ||
4276 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4277 cgroup_unlock();
4278 return -EINVAL;
4279 }
4280
4281 memcg->swappiness = val;
4282
4283 cgroup_unlock();
4284
4285 return 0;
4286}
4287
4288static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4289{
4290 struct mem_cgroup_threshold_ary *t;
4291 u64 usage;
4292 int i;
4293
4294 rcu_read_lock();
4295 if (!swap)
4296 t = rcu_dereference(memcg->thresholds.primary);
4297 else
4298 t = rcu_dereference(memcg->memsw_thresholds.primary);
4299
4300 if (!t)
4301 goto unlock;
4302
4303 usage = mem_cgroup_usage(memcg, swap);
4304
4305 /*
4306 * current_threshold points to threshold just below usage.
4307 * If it's not true, a threshold was crossed after last
4308 * call of __mem_cgroup_threshold().
4309 */
4310 i = t->current_threshold;
4311
4312 /*
4313 * Iterate backward over array of thresholds starting from
4314 * current_threshold and check if a threshold is crossed.
4315 * If none of thresholds below usage is crossed, we read
4316 * only one element of the array here.
4317 */
4318 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4319 eventfd_signal(t->entries[i].eventfd, 1);
4320
4321 /* i = current_threshold + 1 */
4322 i++;
4323
4324 /*
4325 * Iterate forward over array of thresholds starting from
4326 * current_threshold+1 and check if a threshold is crossed.
4327 * If none of thresholds above usage is crossed, we read
4328 * only one element of the array here.
4329 */
4330 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4331 eventfd_signal(t->entries[i].eventfd, 1);
4332
4333 /* Update current_threshold */
4334 t->current_threshold = i - 1;
4335unlock:
4336 rcu_read_unlock();
4337}
4338
4339static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4340{
4341 while (memcg) {
4342 __mem_cgroup_threshold(memcg, false);
4343 if (do_swap_account)
4344 __mem_cgroup_threshold(memcg, true);
4345
4346 memcg = parent_mem_cgroup(memcg);
4347 }
4348}
4349
4350static int compare_thresholds(const void *a, const void *b)
4351{
4352 const struct mem_cgroup_threshold *_a = a;
4353 const struct mem_cgroup_threshold *_b = b;
4354
4355 if (_a->threshold > _b->threshold)
4356 return 1;
4357
4358 if (_a->threshold < _b->threshold)
4359 return -1;
4360
4361 return 0;
4362}
4363
4364static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4365{
4366 struct mem_cgroup_eventfd_list *ev;
4367
4368 list_for_each_entry(ev, &memcg->oom_notify, list)
4369 eventfd_signal(ev->eventfd, 1);
4370 return 0;
4371}
4372
4373static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4374{
4375 struct mem_cgroup *iter;
4376
4377 for_each_mem_cgroup_tree(iter, memcg)
4378 mem_cgroup_oom_notify_cb(iter);
4379}
4380
4381static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4382 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4383{
4384 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4385 struct mem_cgroup_thresholds *thresholds;
4386 struct mem_cgroup_threshold_ary *new;
4387 int type = MEMFILE_TYPE(cft->private);
4388 u64 threshold, usage;
4389 int i, size, ret;
4390
4391 ret = res_counter_memparse_write_strategy(args, &threshold);
4392 if (ret)
4393 return ret;
4394
4395 mutex_lock(&memcg->thresholds_lock);
4396
4397 if (type == _MEM)
4398 thresholds = &memcg->thresholds;
4399 else if (type == _MEMSWAP)
4400 thresholds = &memcg->memsw_thresholds;
4401 else
4402 BUG();
4403
4404 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4405
4406 /* Check if a threshold crossed before adding a new one */
4407 if (thresholds->primary)
4408 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4409
4410 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4411
4412 /* Allocate memory for new array of thresholds */
4413 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4414 GFP_KERNEL);
4415 if (!new) {
4416 ret = -ENOMEM;
4417 goto unlock;
4418 }
4419 new->size = size;
4420
4421 /* Copy thresholds (if any) to new array */
4422 if (thresholds->primary) {
4423 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4424 sizeof(struct mem_cgroup_threshold));
4425 }
4426
4427 /* Add new threshold */
4428 new->entries[size - 1].eventfd = eventfd;
4429 new->entries[size - 1].threshold = threshold;
4430
4431 /* Sort thresholds. Registering of new threshold isn't time-critical */
4432 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4433 compare_thresholds, NULL);
4434
4435 /* Find current threshold */
4436 new->current_threshold = -1;
4437 for (i = 0; i < size; i++) {
4438 if (new->entries[i].threshold < usage) {
4439 /*
4440 * new->current_threshold will not be used until
4441 * rcu_assign_pointer(), so it's safe to increment
4442 * it here.
4443 */
4444 ++new->current_threshold;
4445 }
4446 }
4447
4448 /* Free old spare buffer and save old primary buffer as spare */
4449 kfree(thresholds->spare);
4450 thresholds->spare = thresholds->primary;
4451
4452 rcu_assign_pointer(thresholds->primary, new);
4453
4454 /* To be sure that nobody uses thresholds */
4455 synchronize_rcu();
4456
4457unlock:
4458 mutex_unlock(&memcg->thresholds_lock);
4459
4460 return ret;
4461}
4462
4463static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4464 struct cftype *cft, struct eventfd_ctx *eventfd)
4465{
4466 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4467 struct mem_cgroup_thresholds *thresholds;
4468 struct mem_cgroup_threshold_ary *new;
4469 int type = MEMFILE_TYPE(cft->private);
4470 u64 usage;
4471 int i, j, size;
4472
4473 mutex_lock(&memcg->thresholds_lock);
4474 if (type == _MEM)
4475 thresholds = &memcg->thresholds;
4476 else if (type == _MEMSWAP)
4477 thresholds = &memcg->memsw_thresholds;
4478 else
4479 BUG();
4480
4481 if (!thresholds->primary)
4482 goto unlock;
4483
4484 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4485
4486 /* Check if a threshold crossed before removing */
4487 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4488
4489 /* Calculate new number of threshold */
4490 size = 0;
4491 for (i = 0; i < thresholds->primary->size; i++) {
4492 if (thresholds->primary->entries[i].eventfd != eventfd)
4493 size++;
4494 }
4495
4496 new = thresholds->spare;
4497
4498 /* Set thresholds array to NULL if we don't have thresholds */
4499 if (!size) {
4500 kfree(new);
4501 new = NULL;
4502 goto swap_buffers;
4503 }
4504
4505 new->size = size;
4506
4507 /* Copy thresholds and find current threshold */
4508 new->current_threshold = -1;
4509 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4510 if (thresholds->primary->entries[i].eventfd == eventfd)
4511 continue;
4512
4513 new->entries[j] = thresholds->primary->entries[i];
4514 if (new->entries[j].threshold < usage) {
4515 /*
4516 * new->current_threshold will not be used
4517 * until rcu_assign_pointer(), so it's safe to increment
4518 * it here.
4519 */
4520 ++new->current_threshold;
4521 }
4522 j++;
4523 }
4524
4525swap_buffers:
4526 /* Swap primary and spare array */
4527 thresholds->spare = thresholds->primary;
4528 /* If all events are unregistered, free the spare array */
4529 if (!new) {
4530 kfree(thresholds->spare);
4531 thresholds->spare = NULL;
4532 }
4533
4534 rcu_assign_pointer(thresholds->primary, new);
4535
4536 /* To be sure that nobody uses thresholds */
4537 synchronize_rcu();
4538unlock:
4539 mutex_unlock(&memcg->thresholds_lock);
4540}
4541
4542static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4543 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4544{
4545 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4546 struct mem_cgroup_eventfd_list *event;
4547 int type = MEMFILE_TYPE(cft->private);
4548
4549 BUG_ON(type != _OOM_TYPE);
4550 event = kmalloc(sizeof(*event), GFP_KERNEL);
4551 if (!event)
4552 return -ENOMEM;
4553
4554 spin_lock(&memcg_oom_lock);
4555
4556 event->eventfd = eventfd;
4557 list_add(&event->list, &memcg->oom_notify);
4558
4559 /* already in OOM ? */
4560 if (atomic_read(&memcg->under_oom))
4561 eventfd_signal(eventfd, 1);
4562 spin_unlock(&memcg_oom_lock);
4563
4564 return 0;
4565}
4566
4567static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4568 struct cftype *cft, struct eventfd_ctx *eventfd)
4569{
4570 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4571 struct mem_cgroup_eventfd_list *ev, *tmp;
4572 int type = MEMFILE_TYPE(cft->private);
4573
4574 BUG_ON(type != _OOM_TYPE);
4575
4576 spin_lock(&memcg_oom_lock);
4577
4578 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4579 if (ev->eventfd == eventfd) {
4580 list_del(&ev->list);
4581 kfree(ev);
4582 }
4583 }
4584
4585 spin_unlock(&memcg_oom_lock);
4586}
4587
4588static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4589 struct cftype *cft, struct cgroup_map_cb *cb)
4590{
4591 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4592
4593 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4594
4595 if (atomic_read(&memcg->under_oom))
4596 cb->fill(cb, "under_oom", 1);
4597 else
4598 cb->fill(cb, "under_oom", 0);
4599 return 0;
4600}
4601
4602static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4603 struct cftype *cft, u64 val)
4604{
4605 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4606 struct mem_cgroup *parent;
4607
4608 /* cannot set to root cgroup and only 0 and 1 are allowed */
4609 if (!cgrp->parent || !((val == 0) || (val == 1)))
4610 return -EINVAL;
4611
4612 parent = mem_cgroup_from_cont(cgrp->parent);
4613
4614 cgroup_lock();
4615 /* oom-kill-disable is a flag for subhierarchy. */
4616 if ((parent->use_hierarchy) ||
4617 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4618 cgroup_unlock();
4619 return -EINVAL;
4620 }
4621 memcg->oom_kill_disable = val;
4622 if (!val)
4623 memcg_oom_recover(memcg);
4624 cgroup_unlock();
4625 return 0;
4626}
4627
4628#ifdef CONFIG_NUMA
4629static const struct file_operations mem_control_numa_stat_file_operations = {
4630 .read = seq_read,
4631 .llseek = seq_lseek,
4632 .release = single_release,
4633};
4634
4635static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4636{
4637 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4638
4639 file->f_op = &mem_control_numa_stat_file_operations;
4640 return single_open(file, mem_control_numa_stat_show, cont);
4641}
4642#endif /* CONFIG_NUMA */
4643
4644#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4645static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4646{
4647 /*
4648 * Part of this would be better living in a separate allocation
4649 * function, leaving us with just the cgroup tree population work.
4650 * We, however, depend on state such as network's proto_list that
4651 * is only initialized after cgroup creation. I found the less
4652 * cumbersome way to deal with it to defer it all to populate time
4653 */
4654 return mem_cgroup_sockets_init(cont, ss);
4655};
4656
4657static void kmem_cgroup_destroy(struct cgroup *cont)
4658{
4659 mem_cgroup_sockets_destroy(cont);
4660}
4661#else
4662static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4663{
4664 return 0;
4665}
4666
4667static void kmem_cgroup_destroy(struct cgroup *cont)
4668{
4669}
4670#endif
4671
4672static struct cftype mem_cgroup_files[] = {
4673 {
4674 .name = "usage_in_bytes",
4675 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4676 .read_u64 = mem_cgroup_read,
4677 .register_event = mem_cgroup_usage_register_event,
4678 .unregister_event = mem_cgroup_usage_unregister_event,
4679 },
4680 {
4681 .name = "max_usage_in_bytes",
4682 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4683 .trigger = mem_cgroup_reset,
4684 .read_u64 = mem_cgroup_read,
4685 },
4686 {
4687 .name = "limit_in_bytes",
4688 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4689 .write_string = mem_cgroup_write,
4690 .read_u64 = mem_cgroup_read,
4691 },
4692 {
4693 .name = "soft_limit_in_bytes",
4694 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4695 .write_string = mem_cgroup_write,
4696 .read_u64 = mem_cgroup_read,
4697 },
4698 {
4699 .name = "failcnt",
4700 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4701 .trigger = mem_cgroup_reset,
4702 .read_u64 = mem_cgroup_read,
4703 },
4704 {
4705 .name = "stat",
4706 .read_map = mem_control_stat_show,
4707 },
4708 {
4709 .name = "force_empty",
4710 .trigger = mem_cgroup_force_empty_write,
4711 },
4712 {
4713 .name = "use_hierarchy",
4714 .write_u64 = mem_cgroup_hierarchy_write,
4715 .read_u64 = mem_cgroup_hierarchy_read,
4716 },
4717 {
4718 .name = "swappiness",
4719 .read_u64 = mem_cgroup_swappiness_read,
4720 .write_u64 = mem_cgroup_swappiness_write,
4721 },
4722 {
4723 .name = "move_charge_at_immigrate",
4724 .read_u64 = mem_cgroup_move_charge_read,
4725 .write_u64 = mem_cgroup_move_charge_write,
4726 },
4727 {
4728 .name = "oom_control",
4729 .read_map = mem_cgroup_oom_control_read,
4730 .write_u64 = mem_cgroup_oom_control_write,
4731 .register_event = mem_cgroup_oom_register_event,
4732 .unregister_event = mem_cgroup_oom_unregister_event,
4733 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4734 },
4735#ifdef CONFIG_NUMA
4736 {
4737 .name = "numa_stat",
4738 .open = mem_control_numa_stat_open,
4739 .mode = S_IRUGO,
4740 },
4741#endif
4742};
4743
4744#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4745static struct cftype memsw_cgroup_files[] = {
4746 {
4747 .name = "memsw.usage_in_bytes",
4748 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4749 .read_u64 = mem_cgroup_read,
4750 .register_event = mem_cgroup_usage_register_event,
4751 .unregister_event = mem_cgroup_usage_unregister_event,
4752 },
4753 {
4754 .name = "memsw.max_usage_in_bytes",
4755 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4756 .trigger = mem_cgroup_reset,
4757 .read_u64 = mem_cgroup_read,
4758 },
4759 {
4760 .name = "memsw.limit_in_bytes",
4761 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4762 .write_string = mem_cgroup_write,
4763 .read_u64 = mem_cgroup_read,
4764 },
4765 {
4766 .name = "memsw.failcnt",
4767 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4768 .trigger = mem_cgroup_reset,
4769 .read_u64 = mem_cgroup_read,
4770 },
4771};
4772
4773static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4774{
4775 if (!do_swap_account)
4776 return 0;
4777 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4778 ARRAY_SIZE(memsw_cgroup_files));
4779};
4780#else
4781static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4782{
4783 return 0;
4784}
4785#endif
4786
4787static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4788{
4789 struct mem_cgroup_per_node *pn;
4790 struct mem_cgroup_per_zone *mz;
4791 enum lru_list lru;
4792 int zone, tmp = node;
4793 /*
4794 * This routine is called against possible nodes.
4795 * But it's BUG to call kmalloc() against offline node.
4796 *
4797 * TODO: this routine can waste much memory for nodes which will
4798 * never be onlined. It's better to use memory hotplug callback
4799 * function.
4800 */
4801 if (!node_state(node, N_NORMAL_MEMORY))
4802 tmp = -1;
4803 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4804 if (!pn)
4805 return 1;
4806
4807 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4808 mz = &pn->zoneinfo[zone];
4809 for_each_lru(lru)
4810 INIT_LIST_HEAD(&mz->lruvec.lists[lru]);
4811 mz->usage_in_excess = 0;
4812 mz->on_tree = false;
4813 mz->memcg = memcg;
4814 }
4815 memcg->info.nodeinfo[node] = pn;
4816 return 0;
4817}
4818
4819static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4820{
4821 kfree(memcg->info.nodeinfo[node]);
4822}
4823
4824static struct mem_cgroup *mem_cgroup_alloc(void)
4825{
4826 struct mem_cgroup *memcg;
4827 int size = sizeof(struct mem_cgroup);
4828
4829 /* Can be very big if MAX_NUMNODES is very big */
4830 if (size < PAGE_SIZE)
4831 memcg = kzalloc(size, GFP_KERNEL);
4832 else
4833 memcg = vzalloc(size);
4834
4835 if (!memcg)
4836 return NULL;
4837
4838 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4839 if (!memcg->stat)
4840 goto out_free;
4841 spin_lock_init(&memcg->pcp_counter_lock);
4842 return memcg;
4843
4844out_free:
4845 if (size < PAGE_SIZE)
4846 kfree(memcg);
4847 else
4848 vfree(memcg);
4849 return NULL;
4850}
4851
4852/*
4853 * Helpers for freeing a vzalloc()ed mem_cgroup by RCU,
4854 * but in process context. The work_freeing structure is overlaid
4855 * on the rcu_freeing structure, which itself is overlaid on memsw.
4856 */
4857static void vfree_work(struct work_struct *work)
4858{
4859 struct mem_cgroup *memcg;
4860
4861 memcg = container_of(work, struct mem_cgroup, work_freeing);
4862 vfree(memcg);
4863}
4864static void vfree_rcu(struct rcu_head *rcu_head)
4865{
4866 struct mem_cgroup *memcg;
4867
4868 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4869 INIT_WORK(&memcg->work_freeing, vfree_work);
4870 schedule_work(&memcg->work_freeing);
4871}
4872
4873/*
4874 * At destroying mem_cgroup, references from swap_cgroup can remain.
4875 * (scanning all at force_empty is too costly...)
4876 *
4877 * Instead of clearing all references at force_empty, we remember
4878 * the number of reference from swap_cgroup and free mem_cgroup when
4879 * it goes down to 0.
4880 *
4881 * Removal of cgroup itself succeeds regardless of refs from swap.
4882 */
4883
4884static void __mem_cgroup_free(struct mem_cgroup *memcg)
4885{
4886 int node;
4887
4888 mem_cgroup_remove_from_trees(memcg);
4889 free_css_id(&mem_cgroup_subsys, &memcg->css);
4890
4891 for_each_node(node)
4892 free_mem_cgroup_per_zone_info(memcg, node);
4893
4894 free_percpu(memcg->stat);
4895 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4896 kfree_rcu(memcg, rcu_freeing);
4897 else
4898 call_rcu(&memcg->rcu_freeing, vfree_rcu);
4899}
4900
4901static void mem_cgroup_get(struct mem_cgroup *memcg)
4902{
4903 atomic_inc(&memcg->refcnt);
4904}
4905
4906static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4907{
4908 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4909 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4910 __mem_cgroup_free(memcg);
4911 if (parent)
4912 mem_cgroup_put(parent);
4913 }
4914}
4915
4916static void mem_cgroup_put(struct mem_cgroup *memcg)
4917{
4918 __mem_cgroup_put(memcg, 1);
4919}
4920
4921/*
4922 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4923 */
4924struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4925{
4926 if (!memcg->res.parent)
4927 return NULL;
4928 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4929}
4930EXPORT_SYMBOL(parent_mem_cgroup);
4931
4932#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4933static void __init enable_swap_cgroup(void)
4934{
4935 if (!mem_cgroup_disabled() && really_do_swap_account)
4936 do_swap_account = 1;
4937}
4938#else
4939static void __init enable_swap_cgroup(void)
4940{
4941}
4942#endif
4943
4944static int mem_cgroup_soft_limit_tree_init(void)
4945{
4946 struct mem_cgroup_tree_per_node *rtpn;
4947 struct mem_cgroup_tree_per_zone *rtpz;
4948 int tmp, node, zone;
4949
4950 for_each_node(node) {
4951 tmp = node;
4952 if (!node_state(node, N_NORMAL_MEMORY))
4953 tmp = -1;
4954 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4955 if (!rtpn)
4956 goto err_cleanup;
4957
4958 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4959
4960 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4961 rtpz = &rtpn->rb_tree_per_zone[zone];
4962 rtpz->rb_root = RB_ROOT;
4963 spin_lock_init(&rtpz->lock);
4964 }
4965 }
4966 return 0;
4967
4968err_cleanup:
4969 for_each_node(node) {
4970 if (!soft_limit_tree.rb_tree_per_node[node])
4971 break;
4972 kfree(soft_limit_tree.rb_tree_per_node[node]);
4973 soft_limit_tree.rb_tree_per_node[node] = NULL;
4974 }
4975 return 1;
4976
4977}
4978
4979static struct cgroup_subsys_state * __ref
4980mem_cgroup_create(struct cgroup *cont)
4981{
4982 struct mem_cgroup *memcg, *parent;
4983 long error = -ENOMEM;
4984 int node;
4985
4986 memcg = mem_cgroup_alloc();
4987 if (!memcg)
4988 return ERR_PTR(error);
4989
4990 for_each_node(node)
4991 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4992 goto free_out;
4993
4994 /* root ? */
4995 if (cont->parent == NULL) {
4996 int cpu;
4997 enable_swap_cgroup();
4998 parent = NULL;
4999 if (mem_cgroup_soft_limit_tree_init())
5000 goto free_out;
5001 root_mem_cgroup = memcg;
5002 for_each_possible_cpu(cpu) {
5003 struct memcg_stock_pcp *stock =
5004 &per_cpu(memcg_stock, cpu);
5005 INIT_WORK(&stock->work, drain_local_stock);
5006 }
5007 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5008 } else {
5009 parent = mem_cgroup_from_cont(cont->parent);
5010 memcg->use_hierarchy = parent->use_hierarchy;
5011 memcg->oom_kill_disable = parent->oom_kill_disable;
5012 }
5013
5014 if (parent && parent->use_hierarchy) {
5015 res_counter_init(&memcg->res, &parent->res);
5016 res_counter_init(&memcg->memsw, &parent->memsw);
5017 /*
5018 * We increment refcnt of the parent to ensure that we can
5019 * safely access it on res_counter_charge/uncharge.
5020 * This refcnt will be decremented when freeing this
5021 * mem_cgroup(see mem_cgroup_put).
5022 */
5023 mem_cgroup_get(parent);
5024 } else {
5025 res_counter_init(&memcg->res, NULL);
5026 res_counter_init(&memcg->memsw, NULL);
5027 }
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 spin_lock_init(&memcg->move_lock);
5037 return &memcg->css;
5038free_out:
5039 __mem_cgroup_free(memcg);
5040 return ERR_PTR(error);
5041}
5042
5043static int mem_cgroup_pre_destroy(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
5050static void mem_cgroup_destroy(struct cgroup *cont)
5051{
5052 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5053
5054 kmem_cgroup_destroy(cont);
5055
5056 mem_cgroup_put(memcg);
5057}
5058
5059static int mem_cgroup_populate(struct cgroup_subsys *ss,
5060 struct cgroup *cont)
5061{
5062 int ret;
5063
5064 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5065 ARRAY_SIZE(mem_cgroup_files));
5066
5067 if (!ret)
5068 ret = register_memsw_files(cont, ss);
5069
5070 if (!ret)
5071 ret = register_kmem_files(cont, ss);
5072
5073 return ret;
5074}
5075
5076#ifdef CONFIG_MMU
5077/* Handlers for move charge at task migration. */
5078#define PRECHARGE_COUNT_AT_ONCE 256
5079static int mem_cgroup_do_precharge(unsigned long count)
5080{
5081 int ret = 0;
5082 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5083 struct mem_cgroup *memcg = mc.to;
5084
5085 if (mem_cgroup_is_root(memcg)) {
5086 mc.precharge += count;
5087 /* we don't need css_get for root */
5088 return ret;
5089 }
5090 /* try to charge at once */
5091 if (count > 1) {
5092 struct res_counter *dummy;
5093 /*
5094 * "memcg" cannot be under rmdir() because we've already checked
5095 * by cgroup_lock_live_cgroup() that it is not removed and we
5096 * are still under the same cgroup_mutex. So we can postpone
5097 * css_get().
5098 */
5099 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5100 goto one_by_one;
5101 if (do_swap_account && res_counter_charge(&memcg->memsw,
5102 PAGE_SIZE * count, &dummy)) {
5103 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5104 goto one_by_one;
5105 }
5106 mc.precharge += count;
5107 return ret;
5108 }
5109one_by_one:
5110 /* fall back to one by one charge */
5111 while (count--) {
5112 if (signal_pending(current)) {
5113 ret = -EINTR;
5114 break;
5115 }
5116 if (!batch_count--) {
5117 batch_count = PRECHARGE_COUNT_AT_ONCE;
5118 cond_resched();
5119 }
5120 ret = __mem_cgroup_try_charge(NULL,
5121 GFP_KERNEL, 1, &memcg, false);
5122 if (ret)
5123 /* mem_cgroup_clear_mc() will do uncharge later */
5124 return ret;
5125 mc.precharge++;
5126 }
5127 return ret;
5128}
5129
5130/**
5131 * get_mctgt_type - get target type of moving charge
5132 * @vma: the vma the pte to be checked belongs
5133 * @addr: the address corresponding to the pte to be checked
5134 * @ptent: the pte to be checked
5135 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5136 *
5137 * Returns
5138 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5139 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5140 * move charge. if @target is not NULL, the page is stored in target->page
5141 * with extra refcnt got(Callers should handle it).
5142 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5143 * target for charge migration. if @target is not NULL, the entry is stored
5144 * in target->ent.
5145 *
5146 * Called with pte lock held.
5147 */
5148union mc_target {
5149 struct page *page;
5150 swp_entry_t ent;
5151};
5152
5153enum mc_target_type {
5154 MC_TARGET_NONE = 0,
5155 MC_TARGET_PAGE,
5156 MC_TARGET_SWAP,
5157};
5158
5159static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5160 unsigned long addr, pte_t ptent)
5161{
5162 struct page *page = vm_normal_page(vma, addr, ptent);
5163
5164 if (!page || !page_mapped(page))
5165 return NULL;
5166 if (PageAnon(page)) {
5167 /* we don't move shared anon */
5168 if (!move_anon() || page_mapcount(page) > 2)
5169 return NULL;
5170 } else if (!move_file())
5171 /* we ignore mapcount for file pages */
5172 return NULL;
5173 if (!get_page_unless_zero(page))
5174 return NULL;
5175
5176 return page;
5177}
5178
5179static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5180 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5181{
5182 int usage_count;
5183 struct page *page = NULL;
5184 swp_entry_t ent = pte_to_swp_entry(ptent);
5185
5186 if (!move_anon() || non_swap_entry(ent))
5187 return NULL;
5188 usage_count = mem_cgroup_count_swap_user(ent, &page);
5189 if (usage_count > 1) { /* we don't move shared anon */
5190 if (page)
5191 put_page(page);
5192 return NULL;
5193 }
5194 if (do_swap_account)
5195 entry->val = ent.val;
5196
5197 return page;
5198}
5199
5200static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5201 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5202{
5203 struct page *page = NULL;
5204 struct inode *inode;
5205 struct address_space *mapping;
5206 pgoff_t pgoff;
5207
5208 if (!vma->vm_file) /* anonymous vma */
5209 return NULL;
5210 if (!move_file())
5211 return NULL;
5212
5213 inode = vma->vm_file->f_path.dentry->d_inode;
5214 mapping = vma->vm_file->f_mapping;
5215 if (pte_none(ptent))
5216 pgoff = linear_page_index(vma, addr);
5217 else /* pte_file(ptent) is true */
5218 pgoff = pte_to_pgoff(ptent);
5219
5220 /* page is moved even if it's not RSS of this task(page-faulted). */
5221 page = find_get_page(mapping, pgoff);
5222
5223#ifdef CONFIG_SWAP
5224 /* shmem/tmpfs may report page out on swap: account for that too. */
5225 if (radix_tree_exceptional_entry(page)) {
5226 swp_entry_t swap = radix_to_swp_entry(page);
5227 if (do_swap_account)
5228 *entry = swap;
5229 page = find_get_page(&swapper_space, swap.val);
5230 }
5231#endif
5232 return page;
5233}
5234
5235static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5236 unsigned long addr, pte_t ptent, union mc_target *target)
5237{
5238 struct page *page = NULL;
5239 struct page_cgroup *pc;
5240 enum mc_target_type ret = MC_TARGET_NONE;
5241 swp_entry_t ent = { .val = 0 };
5242
5243 if (pte_present(ptent))
5244 page = mc_handle_present_pte(vma, addr, ptent);
5245 else if (is_swap_pte(ptent))
5246 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5247 else if (pte_none(ptent) || pte_file(ptent))
5248 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5249
5250 if (!page && !ent.val)
5251 return ret;
5252 if (page) {
5253 pc = lookup_page_cgroup(page);
5254 /*
5255 * Do only loose check w/o page_cgroup lock.
5256 * mem_cgroup_move_account() checks the pc is valid or not under
5257 * the lock.
5258 */
5259 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5260 ret = MC_TARGET_PAGE;
5261 if (target)
5262 target->page = page;
5263 }
5264 if (!ret || !target)
5265 put_page(page);
5266 }
5267 /* There is a swap entry and a page doesn't exist or isn't charged */
5268 if (ent.val && !ret &&
5269 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5270 ret = MC_TARGET_SWAP;
5271 if (target)
5272 target->ent = ent;
5273 }
5274 return ret;
5275}
5276
5277#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5278/*
5279 * We don't consider swapping or file mapped pages because THP does not
5280 * support them for now.
5281 * Caller should make sure that pmd_trans_huge(pmd) is true.
5282 */
5283static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5284 unsigned long addr, pmd_t pmd, union mc_target *target)
5285{
5286 struct page *page = NULL;
5287 struct page_cgroup *pc;
5288 enum mc_target_type ret = MC_TARGET_NONE;
5289
5290 page = pmd_page(pmd);
5291 VM_BUG_ON(!page || !PageHead(page));
5292 if (!move_anon())
5293 return ret;
5294 pc = lookup_page_cgroup(page);
5295 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5296 ret = MC_TARGET_PAGE;
5297 if (target) {
5298 get_page(page);
5299 target->page = page;
5300 }
5301 }
5302 return ret;
5303}
5304#else
5305static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5306 unsigned long addr, pmd_t pmd, union mc_target *target)
5307{
5308 return MC_TARGET_NONE;
5309}
5310#endif
5311
5312static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5313 unsigned long addr, unsigned long end,
5314 struct mm_walk *walk)
5315{
5316 struct vm_area_struct *vma = walk->private;
5317 pte_t *pte;
5318 spinlock_t *ptl;
5319
5320 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5321 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5322 mc.precharge += HPAGE_PMD_NR;
5323 spin_unlock(&vma->vm_mm->page_table_lock);
5324 return 0;
5325 }
5326
5327 if (pmd_trans_unstable(pmd))
5328 return 0;
5329 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5330 for (; addr != end; pte++, addr += PAGE_SIZE)
5331 if (get_mctgt_type(vma, addr, *pte, NULL))
5332 mc.precharge++; /* increment precharge temporarily */
5333 pte_unmap_unlock(pte - 1, ptl);
5334 cond_resched();
5335
5336 return 0;
5337}
5338
5339static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5340{
5341 unsigned long precharge;
5342 struct vm_area_struct *vma;
5343
5344 down_read(&mm->mmap_sem);
5345 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5346 struct mm_walk mem_cgroup_count_precharge_walk = {
5347 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5348 .mm = mm,
5349 .private = vma,
5350 };
5351 if (is_vm_hugetlb_page(vma))
5352 continue;
5353 walk_page_range(vma->vm_start, vma->vm_end,
5354 &mem_cgroup_count_precharge_walk);
5355 }
5356 up_read(&mm->mmap_sem);
5357
5358 precharge = mc.precharge;
5359 mc.precharge = 0;
5360
5361 return precharge;
5362}
5363
5364static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5365{
5366 unsigned long precharge = mem_cgroup_count_precharge(mm);
5367
5368 VM_BUG_ON(mc.moving_task);
5369 mc.moving_task = current;
5370 return mem_cgroup_do_precharge(precharge);
5371}
5372
5373/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5374static void __mem_cgroup_clear_mc(void)
5375{
5376 struct mem_cgroup *from = mc.from;
5377 struct mem_cgroup *to = mc.to;
5378
5379 /* we must uncharge all the leftover precharges from mc.to */
5380 if (mc.precharge) {
5381 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5382 mc.precharge = 0;
5383 }
5384 /*
5385 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5386 * we must uncharge here.
5387 */
5388 if (mc.moved_charge) {
5389 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5390 mc.moved_charge = 0;
5391 }
5392 /* we must fixup refcnts and charges */
5393 if (mc.moved_swap) {
5394 /* uncharge swap account from the old cgroup */
5395 if (!mem_cgroup_is_root(mc.from))
5396 res_counter_uncharge(&mc.from->memsw,
5397 PAGE_SIZE * mc.moved_swap);
5398 __mem_cgroup_put(mc.from, mc.moved_swap);
5399
5400 if (!mem_cgroup_is_root(mc.to)) {
5401 /*
5402 * we charged both to->res and to->memsw, so we should
5403 * uncharge to->res.
5404 */
5405 res_counter_uncharge(&mc.to->res,
5406 PAGE_SIZE * mc.moved_swap);
5407 }
5408 /* we've already done mem_cgroup_get(mc.to) */
5409 mc.moved_swap = 0;
5410 }
5411 memcg_oom_recover(from);
5412 memcg_oom_recover(to);
5413 wake_up_all(&mc.waitq);
5414}
5415
5416static void mem_cgroup_clear_mc(void)
5417{
5418 struct mem_cgroup *from = mc.from;
5419
5420 /*
5421 * we must clear moving_task before waking up waiters at the end of
5422 * task migration.
5423 */
5424 mc.moving_task = NULL;
5425 __mem_cgroup_clear_mc();
5426 spin_lock(&mc.lock);
5427 mc.from = NULL;
5428 mc.to = NULL;
5429 spin_unlock(&mc.lock);
5430 mem_cgroup_end_move(from);
5431}
5432
5433static int mem_cgroup_can_attach(struct cgroup *cgroup,
5434 struct cgroup_taskset *tset)
5435{
5436 struct task_struct *p = cgroup_taskset_first(tset);
5437 int ret = 0;
5438 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5439
5440 if (memcg->move_charge_at_immigrate) {
5441 struct mm_struct *mm;
5442 struct mem_cgroup *from = mem_cgroup_from_task(p);
5443
5444 VM_BUG_ON(from == memcg);
5445
5446 mm = get_task_mm(p);
5447 if (!mm)
5448 return 0;
5449 /* We move charges only when we move a owner of the mm */
5450 if (mm->owner == p) {
5451 VM_BUG_ON(mc.from);
5452 VM_BUG_ON(mc.to);
5453 VM_BUG_ON(mc.precharge);
5454 VM_BUG_ON(mc.moved_charge);
5455 VM_BUG_ON(mc.moved_swap);
5456 mem_cgroup_start_move(from);
5457 spin_lock(&mc.lock);
5458 mc.from = from;
5459 mc.to = memcg;
5460 spin_unlock(&mc.lock);
5461 /* We set mc.moving_task later */
5462
5463 ret = mem_cgroup_precharge_mc(mm);
5464 if (ret)
5465 mem_cgroup_clear_mc();
5466 }
5467 mmput(mm);
5468 }
5469 return ret;
5470}
5471
5472static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5473 struct cgroup_taskset *tset)
5474{
5475 mem_cgroup_clear_mc();
5476}
5477
5478static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5479 unsigned long addr, unsigned long end,
5480 struct mm_walk *walk)
5481{
5482 int ret = 0;
5483 struct vm_area_struct *vma = walk->private;
5484 pte_t *pte;
5485 spinlock_t *ptl;
5486 enum mc_target_type target_type;
5487 union mc_target target;
5488 struct page *page;
5489 struct page_cgroup *pc;
5490
5491 /*
5492 * We don't take compound_lock() here but no race with splitting thp
5493 * happens because:
5494 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5495 * under splitting, which means there's no concurrent thp split,
5496 * - if another thread runs into split_huge_page() just after we
5497 * entered this if-block, the thread must wait for page table lock
5498 * to be unlocked in __split_huge_page_splitting(), where the main
5499 * part of thp split is not executed yet.
5500 */
5501 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5502 if (mc.precharge < HPAGE_PMD_NR) {
5503 spin_unlock(&vma->vm_mm->page_table_lock);
5504 return 0;
5505 }
5506 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5507 if (target_type == MC_TARGET_PAGE) {
5508 page = target.page;
5509 if (!isolate_lru_page(page)) {
5510 pc = lookup_page_cgroup(page);
5511 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5512 pc, mc.from, mc.to,
5513 false)) {
5514 mc.precharge -= HPAGE_PMD_NR;
5515 mc.moved_charge += HPAGE_PMD_NR;
5516 }
5517 putback_lru_page(page);
5518 }
5519 put_page(page);
5520 }
5521 spin_unlock(&vma->vm_mm->page_table_lock);
5522 return 0;
5523 }
5524
5525 if (pmd_trans_unstable(pmd))
5526 return 0;
5527retry:
5528 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5529 for (; addr != end; addr += PAGE_SIZE) {
5530 pte_t ptent = *(pte++);
5531 swp_entry_t ent;
5532
5533 if (!mc.precharge)
5534 break;
5535
5536 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5537 case MC_TARGET_PAGE:
5538 page = target.page;
5539 if (isolate_lru_page(page))
5540 goto put;
5541 pc = lookup_page_cgroup(page);
5542 if (!mem_cgroup_move_account(page, 1, pc,
5543 mc.from, mc.to, false)) {
5544 mc.precharge--;
5545 /* we uncharge from mc.from later. */
5546 mc.moved_charge++;
5547 }
5548 putback_lru_page(page);
5549put: /* get_mctgt_type() gets the page */
5550 put_page(page);
5551 break;
5552 case MC_TARGET_SWAP:
5553 ent = target.ent;
5554 if (!mem_cgroup_move_swap_account(ent,
5555 mc.from, mc.to, false)) {
5556 mc.precharge--;
5557 /* we fixup refcnts and charges later. */
5558 mc.moved_swap++;
5559 }
5560 break;
5561 default:
5562 break;
5563 }
5564 }
5565 pte_unmap_unlock(pte - 1, ptl);
5566 cond_resched();
5567
5568 if (addr != end) {
5569 /*
5570 * We have consumed all precharges we got in can_attach().
5571 * We try charge one by one, but don't do any additional
5572 * charges to mc.to if we have failed in charge once in attach()
5573 * phase.
5574 */
5575 ret = mem_cgroup_do_precharge(1);
5576 if (!ret)
5577 goto retry;
5578 }
5579
5580 return ret;
5581}
5582
5583static void mem_cgroup_move_charge(struct mm_struct *mm)
5584{
5585 struct vm_area_struct *vma;
5586
5587 lru_add_drain_all();
5588retry:
5589 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5590 /*
5591 * Someone who are holding the mmap_sem might be waiting in
5592 * waitq. So we cancel all extra charges, wake up all waiters,
5593 * and retry. Because we cancel precharges, we might not be able
5594 * to move enough charges, but moving charge is a best-effort
5595 * feature anyway, so it wouldn't be a big problem.
5596 */
5597 __mem_cgroup_clear_mc();
5598 cond_resched();
5599 goto retry;
5600 }
5601 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5602 int ret;
5603 struct mm_walk mem_cgroup_move_charge_walk = {
5604 .pmd_entry = mem_cgroup_move_charge_pte_range,
5605 .mm = mm,
5606 .private = vma,
5607 };
5608 if (is_vm_hugetlb_page(vma))
5609 continue;
5610 ret = walk_page_range(vma->vm_start, vma->vm_end,
5611 &mem_cgroup_move_charge_walk);
5612 if (ret)
5613 /*
5614 * means we have consumed all precharges and failed in
5615 * doing additional charge. Just abandon here.
5616 */
5617 break;
5618 }
5619 up_read(&mm->mmap_sem);
5620}
5621
5622static void mem_cgroup_move_task(struct cgroup *cont,
5623 struct cgroup_taskset *tset)
5624{
5625 struct task_struct *p = cgroup_taskset_first(tset);
5626 struct mm_struct *mm = get_task_mm(p);
5627
5628 if (mm) {
5629 if (mc.to)
5630 mem_cgroup_move_charge(mm);
5631 put_swap_token(mm);
5632 mmput(mm);
5633 }
5634 if (mc.to)
5635 mem_cgroup_clear_mc();
5636}
5637#else /* !CONFIG_MMU */
5638static int mem_cgroup_can_attach(struct cgroup *cgroup,
5639 struct cgroup_taskset *tset)
5640{
5641 return 0;
5642}
5643static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5644 struct cgroup_taskset *tset)
5645{
5646}
5647static void mem_cgroup_move_task(struct cgroup *cont,
5648 struct cgroup_taskset *tset)
5649{
5650}
5651#endif
5652
5653struct cgroup_subsys mem_cgroup_subsys = {
5654 .name = "memory",
5655 .subsys_id = mem_cgroup_subsys_id,
5656 .create = mem_cgroup_create,
5657 .pre_destroy = mem_cgroup_pre_destroy,
5658 .destroy = mem_cgroup_destroy,
5659 .populate = mem_cgroup_populate,
5660 .can_attach = mem_cgroup_can_attach,
5661 .cancel_attach = mem_cgroup_cancel_attach,
5662 .attach = mem_cgroup_move_task,
5663 .early_init = 0,
5664 .use_id = 1,
5665};
5666
5667#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5668static int __init enable_swap_account(char *s)
5669{
5670 /* consider enabled if no parameter or 1 is given */
5671 if (!strcmp(s, "1"))
5672 really_do_swap_account = 1;
5673 else if (!strcmp(s, "0"))
5674 really_do_swap_account = 0;
5675 return 1;
5676}
5677__setup("swapaccount=", enable_swap_account);
5678
5679#endif