]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - mm/memcontrol.c
drivers/video: implement a simple framebuffer driver
[mirror_ubuntu-artful-kernel.git] / mm / memcontrol.c
CommitLineData
8cdea7c0
BS
1/* memcontrol.c - Memory Controller
2 *
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
5 *
78fb7466
PE
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
8 *
2e72b634
KS
9 * Memory thresholds
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
12 *
7ae1e1d0
GC
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
16 *
8cdea7c0
BS
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
21 *
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
26 */
27
28#include <linux/res_counter.h>
29#include <linux/memcontrol.h>
30#include <linux/cgroup.h>
78fb7466 31#include <linux/mm.h>
4ffef5fe 32#include <linux/hugetlb.h>
d13d1443 33#include <linux/pagemap.h>
d52aa412 34#include <linux/smp.h>
8a9f3ccd 35#include <linux/page-flags.h>
66e1707b 36#include <linux/backing-dev.h>
8a9f3ccd
BS
37#include <linux/bit_spinlock.h>
38#include <linux/rcupdate.h>
e222432b 39#include <linux/limits.h>
b9e15baf 40#include <linux/export.h>
8c7c6e34 41#include <linux/mutex.h>
f64c3f54 42#include <linux/rbtree.h>
b6ac57d5 43#include <linux/slab.h>
66e1707b 44#include <linux/swap.h>
02491447 45#include <linux/swapops.h>
66e1707b 46#include <linux/spinlock.h>
2e72b634
KS
47#include <linux/eventfd.h>
48#include <linux/sort.h>
66e1707b 49#include <linux/fs.h>
d2ceb9b7 50#include <linux/seq_file.h>
33327948 51#include <linux/vmalloc.h>
70ddf637 52#include <linux/vmpressure.h>
b69408e8 53#include <linux/mm_inline.h>
52d4b9ac 54#include <linux/page_cgroup.h>
cdec2e42 55#include <linux/cpu.h>
158e0a2d 56#include <linux/oom.h>
08e552c6 57#include "internal.h"
d1a4c0b3 58#include <net/sock.h>
4bd2c1ee 59#include <net/ip.h>
d1a4c0b3 60#include <net/tcp_memcontrol.h>
8cdea7c0 61
8697d331
BS
62#include <asm/uaccess.h>
63
cc8e970c
KM
64#include <trace/events/vmscan.h>
65
a181b0e8 66struct cgroup_subsys mem_cgroup_subsys __read_mostly;
68ae564b
DR
67EXPORT_SYMBOL(mem_cgroup_subsys);
68
a181b0e8 69#define MEM_CGROUP_RECLAIM_RETRIES 5
6bbda35c 70static struct mem_cgroup *root_mem_cgroup __read_mostly;
8cdea7c0 71
c255a458 72#ifdef CONFIG_MEMCG_SWAP
338c8431 73/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
c077719b 74int do_swap_account __read_mostly;
a42c390c
MH
75
76/* for remember boot option*/
c255a458 77#ifdef CONFIG_MEMCG_SWAP_ENABLED
a42c390c
MH
78static int really_do_swap_account __initdata = 1;
79#else
80static int really_do_swap_account __initdata = 0;
81#endif
82
c077719b 83#else
a0db00fc 84#define do_swap_account 0
c077719b
KH
85#endif
86
87
d52aa412
KH
88/*
89 * Statistics for memory cgroup.
90 */
91enum mem_cgroup_stat_index {
92 /*
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 */
b070e65c
DR
95 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
d52aa412
KH
100 MEM_CGROUP_STAT_NSTATS,
101};
102
af7c4b0e
JW
103static const char * const mem_cgroup_stat_names[] = {
104 "cache",
105 "rss",
b070e65c 106 "rss_huge",
af7c4b0e
JW
107 "mapped_file",
108 "swap",
109};
110
e9f8974f
JW
111enum mem_cgroup_events_index {
112 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
456f998e
YH
114 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
e9f8974f
JW
116 MEM_CGROUP_EVENTS_NSTATS,
117};
af7c4b0e
JW
118
119static const char * const mem_cgroup_events_names[] = {
120 "pgpgin",
121 "pgpgout",
122 "pgfault",
123 "pgmajfault",
124};
125
58cf188e
SZ
126static const char * const mem_cgroup_lru_names[] = {
127 "inactive_anon",
128 "active_anon",
129 "inactive_file",
130 "active_file",
131 "unevictable",
132};
133
7a159cc9
JW
134/*
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
139 */
140enum mem_cgroup_events_target {
141 MEM_CGROUP_TARGET_THRESH,
142 MEM_CGROUP_TARGET_SOFTLIMIT,
453a9bf3 143 MEM_CGROUP_TARGET_NUMAINFO,
7a159cc9
JW
144 MEM_CGROUP_NTARGETS,
145};
a0db00fc
KS
146#define THRESHOLDS_EVENTS_TARGET 128
147#define SOFTLIMIT_EVENTS_TARGET 1024
148#define NUMAINFO_EVENTS_TARGET 1024
e9f8974f 149
d52aa412 150struct mem_cgroup_stat_cpu {
7a159cc9 151 long count[MEM_CGROUP_STAT_NSTATS];
e9f8974f 152 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
13114716 153 unsigned long nr_page_events;
7a159cc9 154 unsigned long targets[MEM_CGROUP_NTARGETS];
d52aa412
KH
155};
156
527a5ec9 157struct mem_cgroup_reclaim_iter {
5f578161
MH
158 /*
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
161 */
542f85f9 162 struct mem_cgroup *last_visited;
5f578161
MH
163 unsigned long last_dead_count;
164
527a5ec9
JW
165 /* scan generation, increased every round-trip */
166 unsigned int generation;
167};
168
6d12e2d8
KH
169/*
170 * per-zone information in memory controller.
171 */
6d12e2d8 172struct mem_cgroup_per_zone {
6290df54 173 struct lruvec lruvec;
1eb49272 174 unsigned long lru_size[NR_LRU_LISTS];
3e2f41f1 175
527a5ec9
JW
176 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
177
f64c3f54
BS
178 struct rb_node tree_node; /* RB tree node */
179 unsigned long long usage_in_excess;/* Set to the value by which */
180 /* the soft limit is exceeded*/
181 bool on_tree;
d79154bb 182 struct mem_cgroup *memcg; /* Back pointer, we cannot */
4e416953 183 /* use container_of */
6d12e2d8 184};
6d12e2d8
KH
185
186struct mem_cgroup_per_node {
187 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
188};
189
190struct mem_cgroup_lru_info {
45cf7ebd 191 struct mem_cgroup_per_node *nodeinfo[0];
6d12e2d8
KH
192};
193
f64c3f54
BS
194/*
195 * Cgroups above their limits are maintained in a RB-Tree, independent of
196 * their hierarchy representation
197 */
198
199struct mem_cgroup_tree_per_zone {
200 struct rb_root rb_root;
201 spinlock_t lock;
202};
203
204struct mem_cgroup_tree_per_node {
205 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
206};
207
208struct mem_cgroup_tree {
209 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
210};
211
212static struct mem_cgroup_tree soft_limit_tree __read_mostly;
213
2e72b634
KS
214struct mem_cgroup_threshold {
215 struct eventfd_ctx *eventfd;
216 u64 threshold;
217};
218
9490ff27 219/* For threshold */
2e72b634 220struct mem_cgroup_threshold_ary {
748dad36 221 /* An array index points to threshold just below or equal to usage. */
5407a562 222 int current_threshold;
2e72b634
KS
223 /* Size of entries[] */
224 unsigned int size;
225 /* Array of thresholds */
226 struct mem_cgroup_threshold entries[0];
227};
2c488db2
KS
228
229struct mem_cgroup_thresholds {
230 /* Primary thresholds array */
231 struct mem_cgroup_threshold_ary *primary;
232 /*
233 * Spare threshold array.
234 * This is needed to make mem_cgroup_unregister_event() "never fail".
235 * It must be able to store at least primary->size - 1 entries.
236 */
237 struct mem_cgroup_threshold_ary *spare;
238};
239
9490ff27
KH
240/* for OOM */
241struct mem_cgroup_eventfd_list {
242 struct list_head list;
243 struct eventfd_ctx *eventfd;
244};
2e72b634 245
c0ff4b85
R
246static void mem_cgroup_threshold(struct mem_cgroup *memcg);
247static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
2e72b634 248
8cdea7c0
BS
249/*
250 * The memory controller data structure. The memory controller controls both
251 * page cache and RSS per cgroup. We would eventually like to provide
252 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
253 * to help the administrator determine what knobs to tune.
254 *
255 * TODO: Add a water mark for the memory controller. Reclaim will begin when
8a9f3ccd
BS
256 * we hit the water mark. May be even add a low water mark, such that
257 * no reclaim occurs from a cgroup at it's low water mark, this is
258 * a feature that will be implemented much later in the future.
8cdea7c0
BS
259 */
260struct mem_cgroup {
261 struct cgroup_subsys_state css;
262 /*
263 * the counter to account for memory usage
264 */
265 struct res_counter res;
59927fb9 266
70ddf637
AV
267 /* vmpressure notifications */
268 struct vmpressure vmpressure;
269
59927fb9
HD
270 union {
271 /*
272 * the counter to account for mem+swap usage.
273 */
274 struct res_counter memsw;
275
276 /*
277 * rcu_freeing is used only when freeing struct mem_cgroup,
278 * so put it into a union to avoid wasting more memory.
279 * It must be disjoint from the css field. It could be
280 * in a union with the res field, but res plays a much
281 * larger part in mem_cgroup life than memsw, and might
282 * be of interest, even at time of free, when debugging.
283 * So share rcu_head with the less interesting memsw.
284 */
285 struct rcu_head rcu_freeing;
286 /*
3afe36b1
GC
287 * We also need some space for a worker in deferred freeing.
288 * By the time we call it, rcu_freeing is no longer in use.
59927fb9
HD
289 */
290 struct work_struct work_freeing;
291 };
292
510fc4e1
GC
293 /*
294 * the counter to account for kernel memory usage.
295 */
296 struct res_counter kmem;
18f59ea7
BS
297 /*
298 * Should the accounting and control be hierarchical, per subtree?
299 */
300 bool use_hierarchy;
510fc4e1 301 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
79dfdacc
MH
302
303 bool oom_lock;
304 atomic_t under_oom;
305
8c7c6e34 306 atomic_t refcnt;
14797e23 307
1f4c025b 308 int swappiness;
3c11ecf4
KH
309 /* OOM-Killer disable */
310 int oom_kill_disable;
a7885eb8 311
22a668d7
KH
312 /* set when res.limit == memsw.limit */
313 bool memsw_is_minimum;
314
2e72b634
KS
315 /* protect arrays of thresholds */
316 struct mutex thresholds_lock;
317
318 /* thresholds for memory usage. RCU-protected */
2c488db2 319 struct mem_cgroup_thresholds thresholds;
907860ed 320
2e72b634 321 /* thresholds for mem+swap usage. RCU-protected */
2c488db2 322 struct mem_cgroup_thresholds memsw_thresholds;
907860ed 323
9490ff27
KH
324 /* For oom notifier event fd */
325 struct list_head oom_notify;
185efc0f 326
7dc74be0
DN
327 /*
328 * Should we move charges of a task when a task is moved into this
329 * mem_cgroup ? And what type of charges should we move ?
330 */
331 unsigned long move_charge_at_immigrate;
619d094b
KH
332 /*
333 * set > 0 if pages under this cgroup are moving to other cgroup.
334 */
335 atomic_t moving_account;
312734c0
KH
336 /* taken only while moving_account > 0 */
337 spinlock_t move_lock;
d52aa412 338 /*
c62b1a3b 339 * percpu counter.
d52aa412 340 */
3a7951b4 341 struct mem_cgroup_stat_cpu __percpu *stat;
711d3d2c
KH
342 /*
343 * used when a cpu is offlined or other synchronizations
344 * See mem_cgroup_read_stat().
345 */
346 struct mem_cgroup_stat_cpu nocpu_base;
347 spinlock_t pcp_counter_lock;
d1a4c0b3 348
5f578161 349 atomic_t dead_count;
4bd2c1ee 350#if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
d1a4c0b3
GC
351 struct tcp_memcontrol tcp_mem;
352#endif
2633d7a0
GC
353#if defined(CONFIG_MEMCG_KMEM)
354 /* analogous to slab_common's slab_caches list. per-memcg */
355 struct list_head memcg_slab_caches;
356 /* Not a spinlock, we can take a lot of time walking the list */
357 struct mutex slab_caches_mutex;
358 /* Index in the kmem_cache->memcg_params->memcg_caches array */
359 int kmemcg_id;
360#endif
45cf7ebd
GC
361
362 int last_scanned_node;
363#if MAX_NUMNODES > 1
364 nodemask_t scan_nodes;
365 atomic_t numainfo_events;
366 atomic_t numainfo_updating;
367#endif
70ddf637 368
45cf7ebd
GC
369 /*
370 * Per cgroup active and inactive list, similar to the
371 * per zone LRU lists.
372 *
373 * WARNING: This has to be the last element of the struct. Don't
374 * add new fields after this point.
375 */
376 struct mem_cgroup_lru_info info;
8cdea7c0
BS
377};
378
45cf7ebd
GC
379static size_t memcg_size(void)
380{
381 return sizeof(struct mem_cgroup) +
382 nr_node_ids * sizeof(struct mem_cgroup_per_node);
383}
384
510fc4e1
GC
385/* internal only representation about the status of kmem accounting. */
386enum {
387 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
a8964b9b 388 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
7de37682 389 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
510fc4e1
GC
390};
391
a8964b9b
GC
392/* We account when limit is on, but only after call sites are patched */
393#define KMEM_ACCOUNTED_MASK \
394 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
510fc4e1
GC
395
396#ifdef CONFIG_MEMCG_KMEM
397static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
398{
399 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400}
7de37682
GC
401
402static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
403{
404 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
405}
406
a8964b9b
GC
407static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
408{
409 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
410}
411
55007d84
GC
412static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
413{
414 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
415}
416
7de37682
GC
417static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
418{
419 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
420 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
421}
422
423static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
424{
425 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
426 &memcg->kmem_account_flags);
427}
510fc4e1
GC
428#endif
429
7dc74be0
DN
430/* Stuffs for move charges at task migration. */
431/*
ee5e8472
GC
432 * Types of charges to be moved. "move_charge_at_immitgrate" and
433 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
7dc74be0
DN
434 */
435enum move_type {
4ffef5fe 436 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
87946a72 437 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
7dc74be0
DN
438 NR_MOVE_TYPE,
439};
440
4ffef5fe
DN
441/* "mc" and its members are protected by cgroup_mutex */
442static struct move_charge_struct {
b1dd693e 443 spinlock_t lock; /* for from, to */
4ffef5fe
DN
444 struct mem_cgroup *from;
445 struct mem_cgroup *to;
ee5e8472 446 unsigned long immigrate_flags;
4ffef5fe 447 unsigned long precharge;
854ffa8d 448 unsigned long moved_charge;
483c30b5 449 unsigned long moved_swap;
8033b97c
DN
450 struct task_struct *moving_task; /* a task moving charges */
451 wait_queue_head_t waitq; /* a waitq for other context */
452} mc = {
2bd9bb20 453 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
8033b97c
DN
454 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
455};
4ffef5fe 456
90254a65
DN
457static bool move_anon(void)
458{
ee5e8472 459 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
90254a65
DN
460}
461
87946a72
DN
462static bool move_file(void)
463{
ee5e8472 464 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
87946a72
DN
465}
466
4e416953
BS
467/*
468 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
469 * limit reclaim to prevent infinite loops, if they ever occur.
470 */
a0db00fc
KS
471#define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
472#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
4e416953 473
217bc319
KH
474enum charge_type {
475 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
41326c17 476 MEM_CGROUP_CHARGE_TYPE_ANON,
d13d1443 477 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
8a9478ca 478 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
c05555b5
KH
479 NR_CHARGE_TYPE,
480};
481
8c7c6e34 482/* for encoding cft->private value on file */
86ae53e1
GC
483enum res_type {
484 _MEM,
485 _MEMSWAP,
486 _OOM_TYPE,
510fc4e1 487 _KMEM,
86ae53e1
GC
488};
489
a0db00fc
KS
490#define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
491#define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
8c7c6e34 492#define MEMFILE_ATTR(val) ((val) & 0xffff)
9490ff27
KH
493/* Used for OOM nofiier */
494#define OOM_CONTROL (0)
8c7c6e34 495
75822b44
BS
496/*
497 * Reclaim flags for mem_cgroup_hierarchical_reclaim
498 */
499#define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
500#define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
501#define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
502#define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
503
0999821b
GC
504/*
505 * The memcg_create_mutex will be held whenever a new cgroup is created.
506 * As a consequence, any change that needs to protect against new child cgroups
507 * appearing has to hold it as well.
508 */
509static DEFINE_MUTEX(memcg_create_mutex);
510
c0ff4b85
R
511static void mem_cgroup_get(struct mem_cgroup *memcg);
512static void mem_cgroup_put(struct mem_cgroup *memcg);
e1aab161 513
b2145145
WL
514static inline
515struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
516{
517 return container_of(s, struct mem_cgroup, css);
518}
519
70ddf637
AV
520/* Some nice accessors for the vmpressure. */
521struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
522{
523 if (!memcg)
524 memcg = root_mem_cgroup;
525 return &memcg->vmpressure;
526}
527
528struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
529{
530 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
531}
532
533struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
534{
535 return &mem_cgroup_from_css(css)->vmpressure;
536}
537
7ffc0edc
MH
538static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
539{
540 return (memcg == root_mem_cgroup);
541}
542
e1aab161 543/* Writing them here to avoid exposing memcg's inner layout */
4bd2c1ee 544#if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
e1aab161 545
e1aab161
GC
546void sock_update_memcg(struct sock *sk)
547{
376be5ff 548 if (mem_cgroup_sockets_enabled) {
e1aab161 549 struct mem_cgroup *memcg;
3f134619 550 struct cg_proto *cg_proto;
e1aab161
GC
551
552 BUG_ON(!sk->sk_prot->proto_cgroup);
553
f3f511e1
GC
554 /* Socket cloning can throw us here with sk_cgrp already
555 * filled. It won't however, necessarily happen from
556 * process context. So the test for root memcg given
557 * the current task's memcg won't help us in this case.
558 *
559 * Respecting the original socket's memcg is a better
560 * decision in this case.
561 */
562 if (sk->sk_cgrp) {
563 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
564 mem_cgroup_get(sk->sk_cgrp->memcg);
565 return;
566 }
567
e1aab161
GC
568 rcu_read_lock();
569 memcg = mem_cgroup_from_task(current);
3f134619
GC
570 cg_proto = sk->sk_prot->proto_cgroup(memcg);
571 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
e1aab161 572 mem_cgroup_get(memcg);
3f134619 573 sk->sk_cgrp = cg_proto;
e1aab161
GC
574 }
575 rcu_read_unlock();
576 }
577}
578EXPORT_SYMBOL(sock_update_memcg);
579
580void sock_release_memcg(struct sock *sk)
581{
376be5ff 582 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
e1aab161
GC
583 struct mem_cgroup *memcg;
584 WARN_ON(!sk->sk_cgrp->memcg);
585 memcg = sk->sk_cgrp->memcg;
586 mem_cgroup_put(memcg);
587 }
588}
d1a4c0b3
GC
589
590struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
591{
592 if (!memcg || mem_cgroup_is_root(memcg))
593 return NULL;
594
595 return &memcg->tcp_mem.cg_proto;
596}
597EXPORT_SYMBOL(tcp_proto_cgroup);
e1aab161 598
3f134619
GC
599static void disarm_sock_keys(struct mem_cgroup *memcg)
600{
601 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
602 return;
603 static_key_slow_dec(&memcg_socket_limit_enabled);
604}
605#else
606static void disarm_sock_keys(struct mem_cgroup *memcg)
607{
608}
609#endif
610
a8964b9b 611#ifdef CONFIG_MEMCG_KMEM
55007d84
GC
612/*
613 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
614 * There are two main reasons for not using the css_id for this:
615 * 1) this works better in sparse environments, where we have a lot of memcgs,
616 * but only a few kmem-limited. Or also, if we have, for instance, 200
617 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
618 * 200 entry array for that.
619 *
620 * 2) In order not to violate the cgroup API, we would like to do all memory
621 * allocation in ->create(). At that point, we haven't yet allocated the
622 * css_id. Having a separate index prevents us from messing with the cgroup
623 * core for this
624 *
625 * The current size of the caches array is stored in
626 * memcg_limited_groups_array_size. It will double each time we have to
627 * increase it.
628 */
629static DEFINE_IDA(kmem_limited_groups);
749c5415
GC
630int memcg_limited_groups_array_size;
631
55007d84
GC
632/*
633 * MIN_SIZE is different than 1, because we would like to avoid going through
634 * the alloc/free process all the time. In a small machine, 4 kmem-limited
635 * cgroups is a reasonable guess. In the future, it could be a parameter or
636 * tunable, but that is strictly not necessary.
637 *
638 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
639 * this constant directly from cgroup, but it is understandable that this is
640 * better kept as an internal representation in cgroup.c. In any case, the
641 * css_id space is not getting any smaller, and we don't have to necessarily
642 * increase ours as well if it increases.
643 */
644#define MEMCG_CACHES_MIN_SIZE 4
645#define MEMCG_CACHES_MAX_SIZE 65535
646
d7f25f8a
GC
647/*
648 * A lot of the calls to the cache allocation functions are expected to be
649 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
650 * conditional to this static branch, we'll have to allow modules that does
651 * kmem_cache_alloc and the such to see this symbol as well
652 */
a8964b9b 653struct static_key memcg_kmem_enabled_key;
d7f25f8a 654EXPORT_SYMBOL(memcg_kmem_enabled_key);
a8964b9b
GC
655
656static void disarm_kmem_keys(struct mem_cgroup *memcg)
657{
55007d84 658 if (memcg_kmem_is_active(memcg)) {
a8964b9b 659 static_key_slow_dec(&memcg_kmem_enabled_key);
55007d84
GC
660 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
661 }
bea207c8
GC
662 /*
663 * This check can't live in kmem destruction function,
664 * since the charges will outlive the cgroup
665 */
666 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
a8964b9b
GC
667}
668#else
669static void disarm_kmem_keys(struct mem_cgroup *memcg)
670{
671}
672#endif /* CONFIG_MEMCG_KMEM */
673
674static void disarm_static_keys(struct mem_cgroup *memcg)
675{
676 disarm_sock_keys(memcg);
677 disarm_kmem_keys(memcg);
678}
679
c0ff4b85 680static void drain_all_stock_async(struct mem_cgroup *memcg);
8c7c6e34 681
f64c3f54 682static struct mem_cgroup_per_zone *
c0ff4b85 683mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
f64c3f54 684{
45cf7ebd 685 VM_BUG_ON((unsigned)nid >= nr_node_ids);
c0ff4b85 686 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
f64c3f54
BS
687}
688
c0ff4b85 689struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
d324236b 690{
c0ff4b85 691 return &memcg->css;
d324236b
WF
692}
693
f64c3f54 694static struct mem_cgroup_per_zone *
c0ff4b85 695page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
f64c3f54 696{
97a6c37b
JW
697 int nid = page_to_nid(page);
698 int zid = page_zonenum(page);
f64c3f54 699
c0ff4b85 700 return mem_cgroup_zoneinfo(memcg, nid, zid);
f64c3f54
BS
701}
702
703static struct mem_cgroup_tree_per_zone *
704soft_limit_tree_node_zone(int nid, int zid)
705{
706 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
707}
708
709static struct mem_cgroup_tree_per_zone *
710soft_limit_tree_from_page(struct page *page)
711{
712 int nid = page_to_nid(page);
713 int zid = page_zonenum(page);
714
715 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
716}
717
718static void
c0ff4b85 719__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
f64c3f54 720 struct mem_cgroup_per_zone *mz,
ef8745c1
KH
721 struct mem_cgroup_tree_per_zone *mctz,
722 unsigned long long new_usage_in_excess)
f64c3f54
BS
723{
724 struct rb_node **p = &mctz->rb_root.rb_node;
725 struct rb_node *parent = NULL;
726 struct mem_cgroup_per_zone *mz_node;
727
728 if (mz->on_tree)
729 return;
730
ef8745c1
KH
731 mz->usage_in_excess = new_usage_in_excess;
732 if (!mz->usage_in_excess)
733 return;
f64c3f54
BS
734 while (*p) {
735 parent = *p;
736 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
737 tree_node);
738 if (mz->usage_in_excess < mz_node->usage_in_excess)
739 p = &(*p)->rb_left;
740 /*
741 * We can't avoid mem cgroups that are over their soft
742 * limit by the same amount
743 */
744 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
745 p = &(*p)->rb_right;
746 }
747 rb_link_node(&mz->tree_node, parent, p);
748 rb_insert_color(&mz->tree_node, &mctz->rb_root);
749 mz->on_tree = true;
4e416953
BS
750}
751
752static void
c0ff4b85 753__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
4e416953
BS
754 struct mem_cgroup_per_zone *mz,
755 struct mem_cgroup_tree_per_zone *mctz)
756{
757 if (!mz->on_tree)
758 return;
759 rb_erase(&mz->tree_node, &mctz->rb_root);
760 mz->on_tree = false;
761}
762
f64c3f54 763static void
c0ff4b85 764mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
f64c3f54
BS
765 struct mem_cgroup_per_zone *mz,
766 struct mem_cgroup_tree_per_zone *mctz)
767{
768 spin_lock(&mctz->lock);
c0ff4b85 769 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
f64c3f54
BS
770 spin_unlock(&mctz->lock);
771}
772
f64c3f54 773
c0ff4b85 774static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
f64c3f54 775{
ef8745c1 776 unsigned long long excess;
f64c3f54
BS
777 struct mem_cgroup_per_zone *mz;
778 struct mem_cgroup_tree_per_zone *mctz;
4e649152
KH
779 int nid = page_to_nid(page);
780 int zid = page_zonenum(page);
f64c3f54
BS
781 mctz = soft_limit_tree_from_page(page);
782
783 /*
4e649152
KH
784 * Necessary to update all ancestors when hierarchy is used.
785 * because their event counter is not touched.
f64c3f54 786 */
c0ff4b85
R
787 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
788 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
789 excess = res_counter_soft_limit_excess(&memcg->res);
4e649152
KH
790 /*
791 * We have to update the tree if mz is on RB-tree or
792 * mem is over its softlimit.
793 */
ef8745c1 794 if (excess || mz->on_tree) {
4e649152
KH
795 spin_lock(&mctz->lock);
796 /* if on-tree, remove it */
797 if (mz->on_tree)
c0ff4b85 798 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
4e649152 799 /*
ef8745c1
KH
800 * Insert again. mz->usage_in_excess will be updated.
801 * If excess is 0, no tree ops.
4e649152 802 */
c0ff4b85 803 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
4e649152
KH
804 spin_unlock(&mctz->lock);
805 }
f64c3f54
BS
806 }
807}
808
c0ff4b85 809static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
f64c3f54
BS
810{
811 int node, zone;
812 struct mem_cgroup_per_zone *mz;
813 struct mem_cgroup_tree_per_zone *mctz;
814
3ed28fa1 815 for_each_node(node) {
f64c3f54 816 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
c0ff4b85 817 mz = mem_cgroup_zoneinfo(memcg, node, zone);
f64c3f54 818 mctz = soft_limit_tree_node_zone(node, zone);
c0ff4b85 819 mem_cgroup_remove_exceeded(memcg, mz, mctz);
f64c3f54
BS
820 }
821 }
822}
823
4e416953
BS
824static struct mem_cgroup_per_zone *
825__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
826{
827 struct rb_node *rightmost = NULL;
26251eaf 828 struct mem_cgroup_per_zone *mz;
4e416953
BS
829
830retry:
26251eaf 831 mz = NULL;
4e416953
BS
832 rightmost = rb_last(&mctz->rb_root);
833 if (!rightmost)
834 goto done; /* Nothing to reclaim from */
835
836 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
837 /*
838 * Remove the node now but someone else can add it back,
839 * we will to add it back at the end of reclaim to its correct
840 * position in the tree.
841 */
d79154bb
HD
842 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
843 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
844 !css_tryget(&mz->memcg->css))
4e416953
BS
845 goto retry;
846done:
847 return mz;
848}
849
850static struct mem_cgroup_per_zone *
851mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
852{
853 struct mem_cgroup_per_zone *mz;
854
855 spin_lock(&mctz->lock);
856 mz = __mem_cgroup_largest_soft_limit_node(mctz);
857 spin_unlock(&mctz->lock);
858 return mz;
859}
860
711d3d2c
KH
861/*
862 * Implementation Note: reading percpu statistics for memcg.
863 *
864 * Both of vmstat[] and percpu_counter has threshold and do periodic
865 * synchronization to implement "quick" read. There are trade-off between
866 * reading cost and precision of value. Then, we may have a chance to implement
867 * a periodic synchronizion of counter in memcg's counter.
868 *
869 * But this _read() function is used for user interface now. The user accounts
870 * memory usage by memory cgroup and he _always_ requires exact value because
871 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
872 * have to visit all online cpus and make sum. So, for now, unnecessary
873 * synchronization is not implemented. (just implemented for cpu hotplug)
874 *
875 * If there are kernel internal actions which can make use of some not-exact
876 * value, and reading all cpu value can be performance bottleneck in some
877 * common workload, threashold and synchonization as vmstat[] should be
878 * implemented.
879 */
c0ff4b85 880static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
7a159cc9 881 enum mem_cgroup_stat_index idx)
c62b1a3b 882{
7a159cc9 883 long val = 0;
c62b1a3b 884 int cpu;
c62b1a3b 885
711d3d2c
KH
886 get_online_cpus();
887 for_each_online_cpu(cpu)
c0ff4b85 888 val += per_cpu(memcg->stat->count[idx], cpu);
711d3d2c 889#ifdef CONFIG_HOTPLUG_CPU
c0ff4b85
R
890 spin_lock(&memcg->pcp_counter_lock);
891 val += memcg->nocpu_base.count[idx];
892 spin_unlock(&memcg->pcp_counter_lock);
711d3d2c
KH
893#endif
894 put_online_cpus();
c62b1a3b
KH
895 return val;
896}
897
c0ff4b85 898static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
0c3e73e8
BS
899 bool charge)
900{
901 int val = (charge) ? 1 : -1;
bff6bb83 902 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
0c3e73e8
BS
903}
904
c0ff4b85 905static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
e9f8974f
JW
906 enum mem_cgroup_events_index idx)
907{
908 unsigned long val = 0;
909 int cpu;
910
911 for_each_online_cpu(cpu)
c0ff4b85 912 val += per_cpu(memcg->stat->events[idx], cpu);
e9f8974f 913#ifdef CONFIG_HOTPLUG_CPU
c0ff4b85
R
914 spin_lock(&memcg->pcp_counter_lock);
915 val += memcg->nocpu_base.events[idx];
916 spin_unlock(&memcg->pcp_counter_lock);
e9f8974f
JW
917#endif
918 return val;
919}
920
c0ff4b85 921static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
b070e65c 922 struct page *page,
b2402857 923 bool anon, int nr_pages)
d52aa412 924{
c62b1a3b
KH
925 preempt_disable();
926
b2402857
KH
927 /*
928 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
929 * counted as CACHE even if it's on ANON LRU.
930 */
931 if (anon)
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
c0ff4b85 933 nr_pages);
d52aa412 934 else
b2402857 935 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
c0ff4b85 936 nr_pages);
55e462b0 937
b070e65c
DR
938 if (PageTransHuge(page))
939 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
940 nr_pages);
941
e401f176
KH
942 /* pagein of a big page is an event. So, ignore page size */
943 if (nr_pages > 0)
c0ff4b85 944 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
3751d604 945 else {
c0ff4b85 946 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
3751d604
KH
947 nr_pages = -nr_pages; /* for event */
948 }
e401f176 949
13114716 950 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
2e72b634 951
c62b1a3b 952 preempt_enable();
6d12e2d8
KH
953}
954
bb2a0de9 955unsigned long
4d7dcca2 956mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
074291fe
KK
957{
958 struct mem_cgroup_per_zone *mz;
959
960 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
961 return mz->lru_size[lru];
962}
963
964static unsigned long
c0ff4b85 965mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
bb2a0de9 966 unsigned int lru_mask)
889976db
YH
967{
968 struct mem_cgroup_per_zone *mz;
f156ab93 969 enum lru_list lru;
bb2a0de9
KH
970 unsigned long ret = 0;
971
c0ff4b85 972 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
bb2a0de9 973
f156ab93
HD
974 for_each_lru(lru) {
975 if (BIT(lru) & lru_mask)
976 ret += mz->lru_size[lru];
bb2a0de9
KH
977 }
978 return ret;
979}
980
981static unsigned long
c0ff4b85 982mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
bb2a0de9
KH
983 int nid, unsigned int lru_mask)
984{
889976db
YH
985 u64 total = 0;
986 int zid;
987
bb2a0de9 988 for (zid = 0; zid < MAX_NR_ZONES; zid++)
c0ff4b85
R
989 total += mem_cgroup_zone_nr_lru_pages(memcg,
990 nid, zid, lru_mask);
bb2a0de9 991
889976db
YH
992 return total;
993}
bb2a0de9 994
c0ff4b85 995static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
bb2a0de9 996 unsigned int lru_mask)
6d12e2d8 997{
889976db 998 int nid;
6d12e2d8
KH
999 u64 total = 0;
1000
31aaea4a 1001 for_each_node_state(nid, N_MEMORY)
c0ff4b85 1002 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
6d12e2d8 1003 return total;
d52aa412
KH
1004}
1005
f53d7ce3
JW
1006static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1007 enum mem_cgroup_events_target target)
7a159cc9
JW
1008{
1009 unsigned long val, next;
1010
13114716 1011 val = __this_cpu_read(memcg->stat->nr_page_events);
4799401f 1012 next = __this_cpu_read(memcg->stat->targets[target]);
7a159cc9 1013 /* from time_after() in jiffies.h */
f53d7ce3
JW
1014 if ((long)next - (long)val < 0) {
1015 switch (target) {
1016 case MEM_CGROUP_TARGET_THRESH:
1017 next = val + THRESHOLDS_EVENTS_TARGET;
1018 break;
1019 case MEM_CGROUP_TARGET_SOFTLIMIT:
1020 next = val + SOFTLIMIT_EVENTS_TARGET;
1021 break;
1022 case MEM_CGROUP_TARGET_NUMAINFO:
1023 next = val + NUMAINFO_EVENTS_TARGET;
1024 break;
1025 default:
1026 break;
1027 }
1028 __this_cpu_write(memcg->stat->targets[target], next);
1029 return true;
7a159cc9 1030 }
f53d7ce3 1031 return false;
d2265e6f
KH
1032}
1033
1034/*
1035 * Check events in order.
1036 *
1037 */
c0ff4b85 1038static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
d2265e6f 1039{
4799401f 1040 preempt_disable();
d2265e6f 1041 /* threshold event is triggered in finer grain than soft limit */
f53d7ce3
JW
1042 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_THRESH))) {
82b3f2a7
AM
1044 bool do_softlimit;
1045 bool do_numainfo __maybe_unused;
f53d7ce3
JW
1046
1047 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1048 MEM_CGROUP_TARGET_SOFTLIMIT);
1049#if MAX_NUMNODES > 1
1050 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1051 MEM_CGROUP_TARGET_NUMAINFO);
1052#endif
1053 preempt_enable();
1054
c0ff4b85 1055 mem_cgroup_threshold(memcg);
f53d7ce3 1056 if (unlikely(do_softlimit))
c0ff4b85 1057 mem_cgroup_update_tree(memcg, page);
453a9bf3 1058#if MAX_NUMNODES > 1
f53d7ce3 1059 if (unlikely(do_numainfo))
c0ff4b85 1060 atomic_inc(&memcg->numainfo_events);
453a9bf3 1061#endif
f53d7ce3
JW
1062 } else
1063 preempt_enable();
d2265e6f
KH
1064}
1065
d1a4c0b3 1066struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
8cdea7c0 1067{
b2145145
WL
1068 return mem_cgroup_from_css(
1069 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
8cdea7c0
BS
1070}
1071
cf475ad2 1072struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
78fb7466 1073{
31a78f23
BS
1074 /*
1075 * mm_update_next_owner() may clear mm->owner to NULL
1076 * if it races with swapoff, page migration, etc.
1077 * So this can be called with p == NULL.
1078 */
1079 if (unlikely(!p))
1080 return NULL;
1081
b2145145 1082 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
78fb7466
PE
1083}
1084
a433658c 1085struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
54595fe2 1086{
c0ff4b85 1087 struct mem_cgroup *memcg = NULL;
0b7f569e
KH
1088
1089 if (!mm)
1090 return NULL;
54595fe2
KH
1091 /*
1092 * Because we have no locks, mm->owner's may be being moved to other
1093 * cgroup. We use css_tryget() here even if this looks
1094 * pessimistic (rather than adding locks here).
1095 */
1096 rcu_read_lock();
1097 do {
c0ff4b85
R
1098 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1099 if (unlikely(!memcg))
54595fe2 1100 break;
c0ff4b85 1101 } while (!css_tryget(&memcg->css));
54595fe2 1102 rcu_read_unlock();
c0ff4b85 1103 return memcg;
54595fe2
KH
1104}
1105
16248d8f
MH
1106/*
1107 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1108 * ref. count) or NULL if the whole root's subtree has been visited.
1109 *
1110 * helper function to be used by mem_cgroup_iter
1111 */
1112static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1113 struct mem_cgroup *last_visited)
1114{
1115 struct cgroup *prev_cgroup, *next_cgroup;
1116
1117 /*
1118 * Root is not visited by cgroup iterators so it needs an
1119 * explicit visit.
1120 */
1121 if (!last_visited)
1122 return root;
1123
1124 prev_cgroup = (last_visited == root) ? NULL
1125 : last_visited->css.cgroup;
1126skip_node:
1127 next_cgroup = cgroup_next_descendant_pre(
1128 prev_cgroup, root->css.cgroup);
1129
1130 /*
1131 * Even if we found a group we have to make sure it is
1132 * alive. css && !memcg means that the groups should be
1133 * skipped and we should continue the tree walk.
1134 * last_visited css is safe to use because it is
1135 * protected by css_get and the tree walk is rcu safe.
1136 */
1137 if (next_cgroup) {
1138 struct mem_cgroup *mem = mem_cgroup_from_cont(
1139 next_cgroup);
1140 if (css_tryget(&mem->css))
1141 return mem;
1142 else {
1143 prev_cgroup = next_cgroup;
1144 goto skip_node;
1145 }
1146 }
1147
1148 return NULL;
1149}
1150
5660048c
JW
1151/**
1152 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1153 * @root: hierarchy root
1154 * @prev: previously returned memcg, NULL on first invocation
1155 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1156 *
1157 * Returns references to children of the hierarchy below @root, or
1158 * @root itself, or %NULL after a full round-trip.
1159 *
1160 * Caller must pass the return value in @prev on subsequent
1161 * invocations for reference counting, or use mem_cgroup_iter_break()
1162 * to cancel a hierarchy walk before the round-trip is complete.
1163 *
1164 * Reclaimers can specify a zone and a priority level in @reclaim to
1165 * divide up the memcgs in the hierarchy among all concurrent
1166 * reclaimers operating on the same zone and priority.
1167 */
1168struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1169 struct mem_cgroup *prev,
1170 struct mem_cgroup_reclaim_cookie *reclaim)
14067bb3 1171{
9f3a0d09 1172 struct mem_cgroup *memcg = NULL;
542f85f9 1173 struct mem_cgroup *last_visited = NULL;
5f578161 1174 unsigned long uninitialized_var(dead_count);
711d3d2c 1175
5660048c
JW
1176 if (mem_cgroup_disabled())
1177 return NULL;
1178
9f3a0d09
JW
1179 if (!root)
1180 root = root_mem_cgroup;
7d74b06f 1181
9f3a0d09 1182 if (prev && !reclaim)
542f85f9 1183 last_visited = prev;
14067bb3 1184
9f3a0d09
JW
1185 if (!root->use_hierarchy && root != root_mem_cgroup) {
1186 if (prev)
c40046f3 1187 goto out_css_put;
9f3a0d09
JW
1188 return root;
1189 }
14067bb3 1190
542f85f9 1191 rcu_read_lock();
9f3a0d09 1192 while (!memcg) {
527a5ec9 1193 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
711d3d2c 1194
527a5ec9
JW
1195 if (reclaim) {
1196 int nid = zone_to_nid(reclaim->zone);
1197 int zid = zone_idx(reclaim->zone);
1198 struct mem_cgroup_per_zone *mz;
1199
1200 mz = mem_cgroup_zoneinfo(root, nid, zid);
1201 iter = &mz->reclaim_iter[reclaim->priority];
542f85f9
MH
1202 last_visited = iter->last_visited;
1203 if (prev && reclaim->generation != iter->generation) {
5f578161 1204 iter->last_visited = NULL;
542f85f9
MH
1205 goto out_unlock;
1206 }
5f578161
MH
1207
1208 /*
1209 * If the dead_count mismatches, a destruction
1210 * has happened or is happening concurrently.
1211 * If the dead_count matches, a destruction
1212 * might still happen concurrently, but since
1213 * we checked under RCU, that destruction
1214 * won't free the object until we release the
1215 * RCU reader lock. Thus, the dead_count
1216 * check verifies the pointer is still valid,
1217 * css_tryget() verifies the cgroup pointed to
1218 * is alive.
1219 */
1220 dead_count = atomic_read(&root->dead_count);
1221 smp_rmb();
1222 last_visited = iter->last_visited;
1223 if (last_visited) {
1224 if ((dead_count != iter->last_dead_count) ||
1225 !css_tryget(&last_visited->css)) {
1226 last_visited = NULL;
1227 }
1228 }
527a5ec9 1229 }
7d74b06f 1230
16248d8f 1231 memcg = __mem_cgroup_iter_next(root, last_visited);
14067bb3 1232
527a5ec9 1233 if (reclaim) {
542f85f9
MH
1234 if (last_visited)
1235 css_put(&last_visited->css);
1236
19f39402 1237 iter->last_visited = memcg;
5f578161
MH
1238 smp_wmb();
1239 iter->last_dead_count = dead_count;
542f85f9 1240
19f39402 1241 if (!memcg)
527a5ec9
JW
1242 iter->generation++;
1243 else if (!prev && memcg)
1244 reclaim->generation = iter->generation;
1245 }
9f3a0d09 1246
19f39402 1247 if (prev && !memcg)
542f85f9 1248 goto out_unlock;
9f3a0d09 1249 }
542f85f9
MH
1250out_unlock:
1251 rcu_read_unlock();
c40046f3
MH
1252out_css_put:
1253 if (prev && prev != root)
1254 css_put(&prev->css);
1255
9f3a0d09 1256 return memcg;
14067bb3 1257}
7d74b06f 1258
5660048c
JW
1259/**
1260 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1261 * @root: hierarchy root
1262 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1263 */
1264void mem_cgroup_iter_break(struct mem_cgroup *root,
1265 struct mem_cgroup *prev)
9f3a0d09
JW
1266{
1267 if (!root)
1268 root = root_mem_cgroup;
1269 if (prev && prev != root)
1270 css_put(&prev->css);
1271}
7d74b06f 1272
9f3a0d09
JW
1273/*
1274 * Iteration constructs for visiting all cgroups (under a tree). If
1275 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1276 * be used for reference counting.
1277 */
1278#define for_each_mem_cgroup_tree(iter, root) \
527a5ec9 1279 for (iter = mem_cgroup_iter(root, NULL, NULL); \
9f3a0d09 1280 iter != NULL; \
527a5ec9 1281 iter = mem_cgroup_iter(root, iter, NULL))
711d3d2c 1282
9f3a0d09 1283#define for_each_mem_cgroup(iter) \
527a5ec9 1284 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
9f3a0d09 1285 iter != NULL; \
527a5ec9 1286 iter = mem_cgroup_iter(NULL, iter, NULL))
14067bb3 1287
68ae564b 1288void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
456f998e 1289{
c0ff4b85 1290 struct mem_cgroup *memcg;
456f998e 1291
456f998e 1292 rcu_read_lock();
c0ff4b85
R
1293 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1294 if (unlikely(!memcg))
456f998e
YH
1295 goto out;
1296
1297 switch (idx) {
456f998e 1298 case PGFAULT:
0e574a93
JW
1299 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1300 break;
1301 case PGMAJFAULT:
1302 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
456f998e
YH
1303 break;
1304 default:
1305 BUG();
1306 }
1307out:
1308 rcu_read_unlock();
1309}
68ae564b 1310EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
456f998e 1311
925b7673
JW
1312/**
1313 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1314 * @zone: zone of the wanted lruvec
fa9add64 1315 * @memcg: memcg of the wanted lruvec
925b7673
JW
1316 *
1317 * Returns the lru list vector holding pages for the given @zone and
1318 * @mem. This can be the global zone lruvec, if the memory controller
1319 * is disabled.
1320 */
1321struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1322 struct mem_cgroup *memcg)
1323{
1324 struct mem_cgroup_per_zone *mz;
bea8c150 1325 struct lruvec *lruvec;
925b7673 1326
bea8c150
HD
1327 if (mem_cgroup_disabled()) {
1328 lruvec = &zone->lruvec;
1329 goto out;
1330 }
925b7673
JW
1331
1332 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
bea8c150
HD
1333 lruvec = &mz->lruvec;
1334out:
1335 /*
1336 * Since a node can be onlined after the mem_cgroup was created,
1337 * we have to be prepared to initialize lruvec->zone here;
1338 * and if offlined then reonlined, we need to reinitialize it.
1339 */
1340 if (unlikely(lruvec->zone != zone))
1341 lruvec->zone = zone;
1342 return lruvec;
925b7673
JW
1343}
1344
08e552c6
KH
1345/*
1346 * Following LRU functions are allowed to be used without PCG_LOCK.
1347 * Operations are called by routine of global LRU independently from memcg.
1348 * What we have to take care of here is validness of pc->mem_cgroup.
1349 *
1350 * Changes to pc->mem_cgroup happens when
1351 * 1. charge
1352 * 2. moving account
1353 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1354 * It is added to LRU before charge.
1355 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1356 * When moving account, the page is not on LRU. It's isolated.
1357 */
4f98a2fe 1358
925b7673 1359/**
fa9add64 1360 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
925b7673 1361 * @page: the page
fa9add64 1362 * @zone: zone of the page
925b7673 1363 */
fa9add64 1364struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
08e552c6 1365{
08e552c6 1366 struct mem_cgroup_per_zone *mz;
925b7673
JW
1367 struct mem_cgroup *memcg;
1368 struct page_cgroup *pc;
bea8c150 1369 struct lruvec *lruvec;
6d12e2d8 1370
bea8c150
HD
1371 if (mem_cgroup_disabled()) {
1372 lruvec = &zone->lruvec;
1373 goto out;
1374 }
925b7673 1375
08e552c6 1376 pc = lookup_page_cgroup(page);
38c5d72f 1377 memcg = pc->mem_cgroup;
7512102c
HD
1378
1379 /*
fa9add64 1380 * Surreptitiously switch any uncharged offlist page to root:
7512102c
HD
1381 * an uncharged page off lru does nothing to secure
1382 * its former mem_cgroup from sudden removal.
1383 *
1384 * Our caller holds lru_lock, and PageCgroupUsed is updated
1385 * under page_cgroup lock: between them, they make all uses
1386 * of pc->mem_cgroup safe.
1387 */
fa9add64 1388 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
7512102c
HD
1389 pc->mem_cgroup = memcg = root_mem_cgroup;
1390
925b7673 1391 mz = page_cgroup_zoneinfo(memcg, page);
bea8c150
HD
1392 lruvec = &mz->lruvec;
1393out:
1394 /*
1395 * Since a node can be onlined after the mem_cgroup was created,
1396 * we have to be prepared to initialize lruvec->zone here;
1397 * and if offlined then reonlined, we need to reinitialize it.
1398 */
1399 if (unlikely(lruvec->zone != zone))
1400 lruvec->zone = zone;
1401 return lruvec;
08e552c6 1402}
b69408e8 1403
925b7673 1404/**
fa9add64
HD
1405 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1406 * @lruvec: mem_cgroup per zone lru vector
1407 * @lru: index of lru list the page is sitting on
1408 * @nr_pages: positive when adding or negative when removing
925b7673 1409 *
fa9add64
HD
1410 * This function must be called when a page is added to or removed from an
1411 * lru list.
3f58a829 1412 */
fa9add64
HD
1413void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1414 int nr_pages)
3f58a829
MK
1415{
1416 struct mem_cgroup_per_zone *mz;
fa9add64 1417 unsigned long *lru_size;
3f58a829
MK
1418
1419 if (mem_cgroup_disabled())
1420 return;
1421
fa9add64
HD
1422 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1423 lru_size = mz->lru_size + lru;
1424 *lru_size += nr_pages;
1425 VM_BUG_ON((long)(*lru_size) < 0);
08e552c6 1426}
544122e5 1427
3e92041d 1428/*
c0ff4b85 1429 * Checks whether given mem is same or in the root_mem_cgroup's
3e92041d
MH
1430 * hierarchy subtree
1431 */
c3ac9a8a
JW
1432bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1433 struct mem_cgroup *memcg)
3e92041d 1434{
91c63734
JW
1435 if (root_memcg == memcg)
1436 return true;
3a981f48 1437 if (!root_memcg->use_hierarchy || !memcg)
91c63734 1438 return false;
c3ac9a8a
JW
1439 return css_is_ancestor(&memcg->css, &root_memcg->css);
1440}
1441
1442static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1443 struct mem_cgroup *memcg)
1444{
1445 bool ret;
1446
91c63734 1447 rcu_read_lock();
c3ac9a8a 1448 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
91c63734
JW
1449 rcu_read_unlock();
1450 return ret;
3e92041d
MH
1451}
1452
c0ff4b85 1453int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
4c4a2214
DR
1454{
1455 int ret;
0b7f569e 1456 struct mem_cgroup *curr = NULL;
158e0a2d 1457 struct task_struct *p;
4c4a2214 1458
158e0a2d 1459 p = find_lock_task_mm(task);
de077d22
DR
1460 if (p) {
1461 curr = try_get_mem_cgroup_from_mm(p->mm);
1462 task_unlock(p);
1463 } else {
1464 /*
1465 * All threads may have already detached their mm's, but the oom
1466 * killer still needs to detect if they have already been oom
1467 * killed to prevent needlessly killing additional tasks.
1468 */
1469 task_lock(task);
1470 curr = mem_cgroup_from_task(task);
1471 if (curr)
1472 css_get(&curr->css);
1473 task_unlock(task);
1474 }
0b7f569e
KH
1475 if (!curr)
1476 return 0;
d31f56db 1477 /*
c0ff4b85 1478 * We should check use_hierarchy of "memcg" not "curr". Because checking
d31f56db 1479 * use_hierarchy of "curr" here make this function true if hierarchy is
c0ff4b85
R
1480 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1481 * hierarchy(even if use_hierarchy is disabled in "memcg").
d31f56db 1482 */
c0ff4b85 1483 ret = mem_cgroup_same_or_subtree(memcg, curr);
0b7f569e 1484 css_put(&curr->css);
4c4a2214
DR
1485 return ret;
1486}
1487
c56d5c7d 1488int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
14797e23 1489{
9b272977 1490 unsigned long inactive_ratio;
14797e23 1491 unsigned long inactive;
9b272977 1492 unsigned long active;
c772be93 1493 unsigned long gb;
14797e23 1494
4d7dcca2
HD
1495 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1496 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
14797e23 1497
c772be93
KM
1498 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1499 if (gb)
1500 inactive_ratio = int_sqrt(10 * gb);
1501 else
1502 inactive_ratio = 1;
1503
9b272977 1504 return inactive * inactive_ratio < active;
14797e23
KM
1505}
1506
6d61ef40
BS
1507#define mem_cgroup_from_res_counter(counter, member) \
1508 container_of(counter, struct mem_cgroup, member)
1509
19942822 1510/**
9d11ea9f 1511 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
dad7557e 1512 * @memcg: the memory cgroup
19942822 1513 *
9d11ea9f 1514 * Returns the maximum amount of memory @mem can be charged with, in
7ec99d62 1515 * pages.
19942822 1516 */
c0ff4b85 1517static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
19942822 1518{
9d11ea9f
JW
1519 unsigned long long margin;
1520
c0ff4b85 1521 margin = res_counter_margin(&memcg->res);
9d11ea9f 1522 if (do_swap_account)
c0ff4b85 1523 margin = min(margin, res_counter_margin(&memcg->memsw));
7ec99d62 1524 return margin >> PAGE_SHIFT;
19942822
JW
1525}
1526
1f4c025b 1527int mem_cgroup_swappiness(struct mem_cgroup *memcg)
a7885eb8
KM
1528{
1529 struct cgroup *cgrp = memcg->css.cgroup;
a7885eb8
KM
1530
1531 /* root ? */
1532 if (cgrp->parent == NULL)
1533 return vm_swappiness;
1534
bf1ff263 1535 return memcg->swappiness;
a7885eb8
KM
1536}
1537
619d094b
KH
1538/*
1539 * memcg->moving_account is used for checking possibility that some thread is
1540 * calling move_account(). When a thread on CPU-A starts moving pages under
1541 * a memcg, other threads should check memcg->moving_account under
1542 * rcu_read_lock(), like this:
1543 *
1544 * CPU-A CPU-B
1545 * rcu_read_lock()
1546 * memcg->moving_account+1 if (memcg->mocing_account)
1547 * take heavy locks.
1548 * synchronize_rcu() update something.
1549 * rcu_read_unlock()
1550 * start move here.
1551 */
4331f7d3
KH
1552
1553/* for quick checking without looking up memcg */
1554atomic_t memcg_moving __read_mostly;
1555
c0ff4b85 1556static void mem_cgroup_start_move(struct mem_cgroup *memcg)
32047e2a 1557{
4331f7d3 1558 atomic_inc(&memcg_moving);
619d094b 1559 atomic_inc(&memcg->moving_account);
32047e2a
KH
1560 synchronize_rcu();
1561}
1562
c0ff4b85 1563static void mem_cgroup_end_move(struct mem_cgroup *memcg)
32047e2a 1564{
619d094b
KH
1565 /*
1566 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1567 * We check NULL in callee rather than caller.
1568 */
4331f7d3
KH
1569 if (memcg) {
1570 atomic_dec(&memcg_moving);
619d094b 1571 atomic_dec(&memcg->moving_account);
4331f7d3 1572 }
32047e2a 1573}
619d094b 1574
32047e2a
KH
1575/*
1576 * 2 routines for checking "mem" is under move_account() or not.
1577 *
13fd1dd9
AM
1578 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1579 * is used for avoiding races in accounting. If true,
32047e2a
KH
1580 * pc->mem_cgroup may be overwritten.
1581 *
1582 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1583 * under hierarchy of moving cgroups. This is for
1584 * waiting at hith-memory prressure caused by "move".
1585 */
1586
13fd1dd9 1587static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
32047e2a
KH
1588{
1589 VM_BUG_ON(!rcu_read_lock_held());
619d094b 1590 return atomic_read(&memcg->moving_account) > 0;
32047e2a 1591}
4b534334 1592
c0ff4b85 1593static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
4b534334 1594{
2bd9bb20
KH
1595 struct mem_cgroup *from;
1596 struct mem_cgroup *to;
4b534334 1597 bool ret = false;
2bd9bb20
KH
1598 /*
1599 * Unlike task_move routines, we access mc.to, mc.from not under
1600 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1601 */
1602 spin_lock(&mc.lock);
1603 from = mc.from;
1604 to = mc.to;
1605 if (!from)
1606 goto unlock;
3e92041d 1607
c0ff4b85
R
1608 ret = mem_cgroup_same_or_subtree(memcg, from)
1609 || mem_cgroup_same_or_subtree(memcg, to);
2bd9bb20
KH
1610unlock:
1611 spin_unlock(&mc.lock);
4b534334
KH
1612 return ret;
1613}
1614
c0ff4b85 1615static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
4b534334
KH
1616{
1617 if (mc.moving_task && current != mc.moving_task) {
c0ff4b85 1618 if (mem_cgroup_under_move(memcg)) {
4b534334
KH
1619 DEFINE_WAIT(wait);
1620 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1621 /* moving charge context might have finished. */
1622 if (mc.moving_task)
1623 schedule();
1624 finish_wait(&mc.waitq, &wait);
1625 return true;
1626 }
1627 }
1628 return false;
1629}
1630
312734c0
KH
1631/*
1632 * Take this lock when
1633 * - a code tries to modify page's memcg while it's USED.
1634 * - a code tries to modify page state accounting in a memcg.
13fd1dd9 1635 * see mem_cgroup_stolen(), too.
312734c0
KH
1636 */
1637static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1638 unsigned long *flags)
1639{
1640 spin_lock_irqsave(&memcg->move_lock, *flags);
1641}
1642
1643static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1644 unsigned long *flags)
1645{
1646 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1647}
1648
58cf188e 1649#define K(x) ((x) << (PAGE_SHIFT-10))
e222432b 1650/**
58cf188e 1651 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
e222432b
BS
1652 * @memcg: The memory cgroup that went over limit
1653 * @p: Task that is going to be killed
1654 *
1655 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1656 * enabled
1657 */
1658void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1659{
1660 struct cgroup *task_cgrp;
1661 struct cgroup *mem_cgrp;
1662 /*
1663 * Need a buffer in BSS, can't rely on allocations. The code relies
1664 * on the assumption that OOM is serialized for memory controller.
1665 * If this assumption is broken, revisit this code.
1666 */
1667 static char memcg_name[PATH_MAX];
1668 int ret;
58cf188e
SZ
1669 struct mem_cgroup *iter;
1670 unsigned int i;
e222432b 1671
58cf188e 1672 if (!p)
e222432b
BS
1673 return;
1674
e222432b
BS
1675 rcu_read_lock();
1676
1677 mem_cgrp = memcg->css.cgroup;
1678 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1679
1680 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1681 if (ret < 0) {
1682 /*
1683 * Unfortunately, we are unable to convert to a useful name
1684 * But we'll still print out the usage information
1685 */
1686 rcu_read_unlock();
1687 goto done;
1688 }
1689 rcu_read_unlock();
1690
d045197f 1691 pr_info("Task in %s killed", memcg_name);
e222432b
BS
1692
1693 rcu_read_lock();
1694 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1695 if (ret < 0) {
1696 rcu_read_unlock();
1697 goto done;
1698 }
1699 rcu_read_unlock();
1700
1701 /*
1702 * Continues from above, so we don't need an KERN_ level
1703 */
d045197f 1704 pr_cont(" as a result of limit of %s\n", memcg_name);
e222432b
BS
1705done:
1706
d045197f 1707 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
e222432b
BS
1708 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1709 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1710 res_counter_read_u64(&memcg->res, RES_FAILCNT));
d045197f 1711 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
e222432b
BS
1712 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1713 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1714 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
d045197f 1715 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
510fc4e1
GC
1716 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1717 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1718 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
58cf188e
SZ
1719
1720 for_each_mem_cgroup_tree(iter, memcg) {
1721 pr_info("Memory cgroup stats");
1722
1723 rcu_read_lock();
1724 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1725 if (!ret)
1726 pr_cont(" for %s", memcg_name);
1727 rcu_read_unlock();
1728 pr_cont(":");
1729
1730 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1731 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1732 continue;
1733 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1734 K(mem_cgroup_read_stat(iter, i)));
1735 }
1736
1737 for (i = 0; i < NR_LRU_LISTS; i++)
1738 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1739 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1740
1741 pr_cont("\n");
1742 }
e222432b
BS
1743}
1744
81d39c20
KH
1745/*
1746 * This function returns the number of memcg under hierarchy tree. Returns
1747 * 1(self count) if no children.
1748 */
c0ff4b85 1749static int mem_cgroup_count_children(struct mem_cgroup *memcg)
81d39c20
KH
1750{
1751 int num = 0;
7d74b06f
KH
1752 struct mem_cgroup *iter;
1753
c0ff4b85 1754 for_each_mem_cgroup_tree(iter, memcg)
7d74b06f 1755 num++;
81d39c20
KH
1756 return num;
1757}
1758
a63d83f4
DR
1759/*
1760 * Return the memory (and swap, if configured) limit for a memcg.
1761 */
9cbb78bb 1762static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
a63d83f4
DR
1763{
1764 u64 limit;
a63d83f4 1765
f3e8eb70 1766 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
f3e8eb70 1767
a63d83f4 1768 /*
9a5a8f19 1769 * Do not consider swap space if we cannot swap due to swappiness
a63d83f4 1770 */
9a5a8f19
MH
1771 if (mem_cgroup_swappiness(memcg)) {
1772 u64 memsw;
1773
1774 limit += total_swap_pages << PAGE_SHIFT;
1775 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1776
1777 /*
1778 * If memsw is finite and limits the amount of swap space
1779 * available to this memcg, return that limit.
1780 */
1781 limit = min(limit, memsw);
1782 }
1783
1784 return limit;
a63d83f4
DR
1785}
1786
19965460
DR
1787static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1788 int order)
9cbb78bb
DR
1789{
1790 struct mem_cgroup *iter;
1791 unsigned long chosen_points = 0;
1792 unsigned long totalpages;
1793 unsigned int points = 0;
1794 struct task_struct *chosen = NULL;
1795
876aafbf 1796 /*
465adcf1
DR
1797 * If current has a pending SIGKILL or is exiting, then automatically
1798 * select it. The goal is to allow it to allocate so that it may
1799 * quickly exit and free its memory.
876aafbf 1800 */
465adcf1 1801 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
876aafbf
DR
1802 set_thread_flag(TIF_MEMDIE);
1803 return;
1804 }
1805
1806 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
9cbb78bb
DR
1807 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1808 for_each_mem_cgroup_tree(iter, memcg) {
1809 struct cgroup *cgroup = iter->css.cgroup;
1810 struct cgroup_iter it;
1811 struct task_struct *task;
1812
1813 cgroup_iter_start(cgroup, &it);
1814 while ((task = cgroup_iter_next(cgroup, &it))) {
1815 switch (oom_scan_process_thread(task, totalpages, NULL,
1816 false)) {
1817 case OOM_SCAN_SELECT:
1818 if (chosen)
1819 put_task_struct(chosen);
1820 chosen = task;
1821 chosen_points = ULONG_MAX;
1822 get_task_struct(chosen);
1823 /* fall through */
1824 case OOM_SCAN_CONTINUE:
1825 continue;
1826 case OOM_SCAN_ABORT:
1827 cgroup_iter_end(cgroup, &it);
1828 mem_cgroup_iter_break(memcg, iter);
1829 if (chosen)
1830 put_task_struct(chosen);
1831 return;
1832 case OOM_SCAN_OK:
1833 break;
1834 };
1835 points = oom_badness(task, memcg, NULL, totalpages);
1836 if (points > chosen_points) {
1837 if (chosen)
1838 put_task_struct(chosen);
1839 chosen = task;
1840 chosen_points = points;
1841 get_task_struct(chosen);
1842 }
1843 }
1844 cgroup_iter_end(cgroup, &it);
1845 }
1846
1847 if (!chosen)
1848 return;
1849 points = chosen_points * 1000 / totalpages;
9cbb78bb
DR
1850 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1851 NULL, "Memory cgroup out of memory");
9cbb78bb
DR
1852}
1853
5660048c
JW
1854static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1855 gfp_t gfp_mask,
1856 unsigned long flags)
1857{
1858 unsigned long total = 0;
1859 bool noswap = false;
1860 int loop;
1861
1862 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1863 noswap = true;
1864 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1865 noswap = true;
1866
1867 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1868 if (loop)
1869 drain_all_stock_async(memcg);
1870 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1871 /*
1872 * Allow limit shrinkers, which are triggered directly
1873 * by userspace, to catch signals and stop reclaim
1874 * after minimal progress, regardless of the margin.
1875 */
1876 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1877 break;
1878 if (mem_cgroup_margin(memcg))
1879 break;
1880 /*
1881 * If nothing was reclaimed after two attempts, there
1882 * may be no reclaimable pages in this hierarchy.
1883 */
1884 if (loop && !total)
1885 break;
1886 }
1887 return total;
1888}
1889
4d0c066d
KH
1890/**
1891 * test_mem_cgroup_node_reclaimable
dad7557e 1892 * @memcg: the target memcg
4d0c066d
KH
1893 * @nid: the node ID to be checked.
1894 * @noswap : specify true here if the user wants flle only information.
1895 *
1896 * This function returns whether the specified memcg contains any
1897 * reclaimable pages on a node. Returns true if there are any reclaimable
1898 * pages in the node.
1899 */
c0ff4b85 1900static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
4d0c066d
KH
1901 int nid, bool noswap)
1902{
c0ff4b85 1903 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
4d0c066d
KH
1904 return true;
1905 if (noswap || !total_swap_pages)
1906 return false;
c0ff4b85 1907 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
4d0c066d
KH
1908 return true;
1909 return false;
1910
1911}
889976db
YH
1912#if MAX_NUMNODES > 1
1913
1914/*
1915 * Always updating the nodemask is not very good - even if we have an empty
1916 * list or the wrong list here, we can start from some node and traverse all
1917 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1918 *
1919 */
c0ff4b85 1920static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
889976db
YH
1921{
1922 int nid;
453a9bf3
KH
1923 /*
1924 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1925 * pagein/pageout changes since the last update.
1926 */
c0ff4b85 1927 if (!atomic_read(&memcg->numainfo_events))
453a9bf3 1928 return;
c0ff4b85 1929 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
889976db
YH
1930 return;
1931
889976db 1932 /* make a nodemask where this memcg uses memory from */
31aaea4a 1933 memcg->scan_nodes = node_states[N_MEMORY];
889976db 1934
31aaea4a 1935 for_each_node_mask(nid, node_states[N_MEMORY]) {
889976db 1936
c0ff4b85
R
1937 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1938 node_clear(nid, memcg->scan_nodes);
889976db 1939 }
453a9bf3 1940
c0ff4b85
R
1941 atomic_set(&memcg->numainfo_events, 0);
1942 atomic_set(&memcg->numainfo_updating, 0);
889976db
YH
1943}
1944
1945/*
1946 * Selecting a node where we start reclaim from. Because what we need is just
1947 * reducing usage counter, start from anywhere is O,K. Considering
1948 * memory reclaim from current node, there are pros. and cons.
1949 *
1950 * Freeing memory from current node means freeing memory from a node which
1951 * we'll use or we've used. So, it may make LRU bad. And if several threads
1952 * hit limits, it will see a contention on a node. But freeing from remote
1953 * node means more costs for memory reclaim because of memory latency.
1954 *
1955 * Now, we use round-robin. Better algorithm is welcomed.
1956 */
c0ff4b85 1957int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
889976db
YH
1958{
1959 int node;
1960
c0ff4b85
R
1961 mem_cgroup_may_update_nodemask(memcg);
1962 node = memcg->last_scanned_node;
889976db 1963
c0ff4b85 1964 node = next_node(node, memcg->scan_nodes);
889976db 1965 if (node == MAX_NUMNODES)
c0ff4b85 1966 node = first_node(memcg->scan_nodes);
889976db
YH
1967 /*
1968 * We call this when we hit limit, not when pages are added to LRU.
1969 * No LRU may hold pages because all pages are UNEVICTABLE or
1970 * memcg is too small and all pages are not on LRU. In that case,
1971 * we use curret node.
1972 */
1973 if (unlikely(node == MAX_NUMNODES))
1974 node = numa_node_id();
1975
c0ff4b85 1976 memcg->last_scanned_node = node;
889976db
YH
1977 return node;
1978}
1979
4d0c066d
KH
1980/*
1981 * Check all nodes whether it contains reclaimable pages or not.
1982 * For quick scan, we make use of scan_nodes. This will allow us to skip
1983 * unused nodes. But scan_nodes is lazily updated and may not cotain
1984 * enough new information. We need to do double check.
1985 */
6bbda35c 1986static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
4d0c066d
KH
1987{
1988 int nid;
1989
1990 /*
1991 * quick check...making use of scan_node.
1992 * We can skip unused nodes.
1993 */
c0ff4b85
R
1994 if (!nodes_empty(memcg->scan_nodes)) {
1995 for (nid = first_node(memcg->scan_nodes);
4d0c066d 1996 nid < MAX_NUMNODES;
c0ff4b85 1997 nid = next_node(nid, memcg->scan_nodes)) {
4d0c066d 1998
c0ff4b85 1999 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
4d0c066d
KH
2000 return true;
2001 }
2002 }
2003 /*
2004 * Check rest of nodes.
2005 */
31aaea4a 2006 for_each_node_state(nid, N_MEMORY) {
c0ff4b85 2007 if (node_isset(nid, memcg->scan_nodes))
4d0c066d 2008 continue;
c0ff4b85 2009 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
4d0c066d
KH
2010 return true;
2011 }
2012 return false;
2013}
2014
889976db 2015#else
c0ff4b85 2016int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
889976db
YH
2017{
2018 return 0;
2019}
4d0c066d 2020
6bbda35c 2021static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
4d0c066d 2022{
c0ff4b85 2023 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
4d0c066d 2024}
889976db
YH
2025#endif
2026
5660048c
JW
2027static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2028 struct zone *zone,
2029 gfp_t gfp_mask,
2030 unsigned long *total_scanned)
6d61ef40 2031{
9f3a0d09 2032 struct mem_cgroup *victim = NULL;
5660048c 2033 int total = 0;
04046e1a 2034 int loop = 0;
9d11ea9f 2035 unsigned long excess;
185efc0f 2036 unsigned long nr_scanned;
527a5ec9
JW
2037 struct mem_cgroup_reclaim_cookie reclaim = {
2038 .zone = zone,
2039 .priority = 0,
2040 };
9d11ea9f 2041
c0ff4b85 2042 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
04046e1a 2043
4e416953 2044 while (1) {
527a5ec9 2045 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
9f3a0d09 2046 if (!victim) {
04046e1a 2047 loop++;
4e416953
BS
2048 if (loop >= 2) {
2049 /*
2050 * If we have not been able to reclaim
2051 * anything, it might because there are
2052 * no reclaimable pages under this hierarchy
2053 */
5660048c 2054 if (!total)
4e416953 2055 break;
4e416953 2056 /*
25985edc 2057 * We want to do more targeted reclaim.
4e416953
BS
2058 * excess >> 2 is not to excessive so as to
2059 * reclaim too much, nor too less that we keep
2060 * coming back to reclaim from this cgroup
2061 */
2062 if (total >= (excess >> 2) ||
9f3a0d09 2063 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
4e416953 2064 break;
4e416953 2065 }
9f3a0d09 2066 continue;
4e416953 2067 }
5660048c 2068 if (!mem_cgroup_reclaimable(victim, false))
6d61ef40 2069 continue;
5660048c
JW
2070 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2071 zone, &nr_scanned);
2072 *total_scanned += nr_scanned;
2073 if (!res_counter_soft_limit_excess(&root_memcg->res))
9f3a0d09 2074 break;
6d61ef40 2075 }
9f3a0d09 2076 mem_cgroup_iter_break(root_memcg, victim);
04046e1a 2077 return total;
6d61ef40
BS
2078}
2079
867578cb
KH
2080/*
2081 * Check OOM-Killer is already running under our hierarchy.
2082 * If someone is running, return false.
1af8efe9 2083 * Has to be called with memcg_oom_lock
867578cb 2084 */
c0ff4b85 2085static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
867578cb 2086{
79dfdacc 2087 struct mem_cgroup *iter, *failed = NULL;
a636b327 2088
9f3a0d09 2089 for_each_mem_cgroup_tree(iter, memcg) {
23751be0 2090 if (iter->oom_lock) {
79dfdacc
MH
2091 /*
2092 * this subtree of our hierarchy is already locked
2093 * so we cannot give a lock.
2094 */
79dfdacc 2095 failed = iter;
9f3a0d09
JW
2096 mem_cgroup_iter_break(memcg, iter);
2097 break;
23751be0
JW
2098 } else
2099 iter->oom_lock = true;
7d74b06f 2100 }
867578cb 2101
79dfdacc 2102 if (!failed)
23751be0 2103 return true;
79dfdacc
MH
2104
2105 /*
2106 * OK, we failed to lock the whole subtree so we have to clean up
2107 * what we set up to the failing subtree
2108 */
9f3a0d09 2109 for_each_mem_cgroup_tree(iter, memcg) {
79dfdacc 2110 if (iter == failed) {
9f3a0d09
JW
2111 mem_cgroup_iter_break(memcg, iter);
2112 break;
79dfdacc
MH
2113 }
2114 iter->oom_lock = false;
2115 }
23751be0 2116 return false;
a636b327 2117}
0b7f569e 2118
79dfdacc 2119/*
1af8efe9 2120 * Has to be called with memcg_oom_lock
79dfdacc 2121 */
c0ff4b85 2122static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
0b7f569e 2123{
7d74b06f
KH
2124 struct mem_cgroup *iter;
2125
c0ff4b85 2126 for_each_mem_cgroup_tree(iter, memcg)
79dfdacc
MH
2127 iter->oom_lock = false;
2128 return 0;
2129}
2130
c0ff4b85 2131static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
79dfdacc
MH
2132{
2133 struct mem_cgroup *iter;
2134
c0ff4b85 2135 for_each_mem_cgroup_tree(iter, memcg)
79dfdacc
MH
2136 atomic_inc(&iter->under_oom);
2137}
2138
c0ff4b85 2139static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
79dfdacc
MH
2140{
2141 struct mem_cgroup *iter;
2142
867578cb
KH
2143 /*
2144 * When a new child is created while the hierarchy is under oom,
2145 * mem_cgroup_oom_lock() may not be called. We have to use
2146 * atomic_add_unless() here.
2147 */
c0ff4b85 2148 for_each_mem_cgroup_tree(iter, memcg)
79dfdacc 2149 atomic_add_unless(&iter->under_oom, -1, 0);
0b7f569e
KH
2150}
2151
1af8efe9 2152static DEFINE_SPINLOCK(memcg_oom_lock);
867578cb
KH
2153static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2154
dc98df5a 2155struct oom_wait_info {
d79154bb 2156 struct mem_cgroup *memcg;
dc98df5a
KH
2157 wait_queue_t wait;
2158};
2159
2160static int memcg_oom_wake_function(wait_queue_t *wait,
2161 unsigned mode, int sync, void *arg)
2162{
d79154bb
HD
2163 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2164 struct mem_cgroup *oom_wait_memcg;
dc98df5a
KH
2165 struct oom_wait_info *oom_wait_info;
2166
2167 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
d79154bb 2168 oom_wait_memcg = oom_wait_info->memcg;
dc98df5a 2169
dc98df5a 2170 /*
d79154bb 2171 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
dc98df5a
KH
2172 * Then we can use css_is_ancestor without taking care of RCU.
2173 */
c0ff4b85
R
2174 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2175 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
dc98df5a 2176 return 0;
dc98df5a
KH
2177 return autoremove_wake_function(wait, mode, sync, arg);
2178}
2179
c0ff4b85 2180static void memcg_wakeup_oom(struct mem_cgroup *memcg)
dc98df5a 2181{
c0ff4b85
R
2182 /* for filtering, pass "memcg" as argument. */
2183 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
dc98df5a
KH
2184}
2185
c0ff4b85 2186static void memcg_oom_recover(struct mem_cgroup *memcg)
3c11ecf4 2187{
c0ff4b85
R
2188 if (memcg && atomic_read(&memcg->under_oom))
2189 memcg_wakeup_oom(memcg);
3c11ecf4
KH
2190}
2191
867578cb
KH
2192/*
2193 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2194 */
6bbda35c
KS
2195static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2196 int order)
0b7f569e 2197{
dc98df5a 2198 struct oom_wait_info owait;
3c11ecf4 2199 bool locked, need_to_kill;
867578cb 2200
d79154bb 2201 owait.memcg = memcg;
dc98df5a
KH
2202 owait.wait.flags = 0;
2203 owait.wait.func = memcg_oom_wake_function;
2204 owait.wait.private = current;
2205 INIT_LIST_HEAD(&owait.wait.task_list);
3c11ecf4 2206 need_to_kill = true;
c0ff4b85 2207 mem_cgroup_mark_under_oom(memcg);
79dfdacc 2208
c0ff4b85 2209 /* At first, try to OOM lock hierarchy under memcg.*/
1af8efe9 2210 spin_lock(&memcg_oom_lock);
c0ff4b85 2211 locked = mem_cgroup_oom_lock(memcg);
867578cb
KH
2212 /*
2213 * Even if signal_pending(), we can't quit charge() loop without
2214 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2215 * under OOM is always welcomed, use TASK_KILLABLE here.
2216 */
3c11ecf4 2217 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
c0ff4b85 2218 if (!locked || memcg->oom_kill_disable)
3c11ecf4
KH
2219 need_to_kill = false;
2220 if (locked)
c0ff4b85 2221 mem_cgroup_oom_notify(memcg);
1af8efe9 2222 spin_unlock(&memcg_oom_lock);
867578cb 2223
3c11ecf4
KH
2224 if (need_to_kill) {
2225 finish_wait(&memcg_oom_waitq, &owait.wait);
e845e199 2226 mem_cgroup_out_of_memory(memcg, mask, order);
3c11ecf4 2227 } else {
867578cb 2228 schedule();
dc98df5a 2229 finish_wait(&memcg_oom_waitq, &owait.wait);
867578cb 2230 }
1af8efe9 2231 spin_lock(&memcg_oom_lock);
79dfdacc 2232 if (locked)
c0ff4b85
R
2233 mem_cgroup_oom_unlock(memcg);
2234 memcg_wakeup_oom(memcg);
1af8efe9 2235 spin_unlock(&memcg_oom_lock);
867578cb 2236
c0ff4b85 2237 mem_cgroup_unmark_under_oom(memcg);
79dfdacc 2238
867578cb
KH
2239 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2240 return false;
2241 /* Give chance to dying process */
715a5ee8 2242 schedule_timeout_uninterruptible(1);
867578cb 2243 return true;
0b7f569e
KH
2244}
2245
d69b042f
BS
2246/*
2247 * Currently used to update mapped file statistics, but the routine can be
2248 * generalized to update other statistics as well.
32047e2a
KH
2249 *
2250 * Notes: Race condition
2251 *
2252 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2253 * it tends to be costly. But considering some conditions, we doesn't need
2254 * to do so _always_.
2255 *
2256 * Considering "charge", lock_page_cgroup() is not required because all
2257 * file-stat operations happen after a page is attached to radix-tree. There
2258 * are no race with "charge".
2259 *
2260 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2261 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2262 * if there are race with "uncharge". Statistics itself is properly handled
2263 * by flags.
2264 *
2265 * Considering "move", this is an only case we see a race. To make the race
619d094b
KH
2266 * small, we check mm->moving_account and detect there are possibility of race
2267 * If there is, we take a lock.
d69b042f 2268 */
26174efd 2269
89c06bd5
KH
2270void __mem_cgroup_begin_update_page_stat(struct page *page,
2271 bool *locked, unsigned long *flags)
2272{
2273 struct mem_cgroup *memcg;
2274 struct page_cgroup *pc;
2275
2276 pc = lookup_page_cgroup(page);
2277again:
2278 memcg = pc->mem_cgroup;
2279 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2280 return;
2281 /*
2282 * If this memory cgroup is not under account moving, we don't
da92c47d 2283 * need to take move_lock_mem_cgroup(). Because we already hold
89c06bd5 2284 * rcu_read_lock(), any calls to move_account will be delayed until
13fd1dd9 2285 * rcu_read_unlock() if mem_cgroup_stolen() == true.
89c06bd5 2286 */
13fd1dd9 2287 if (!mem_cgroup_stolen(memcg))
89c06bd5
KH
2288 return;
2289
2290 move_lock_mem_cgroup(memcg, flags);
2291 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2292 move_unlock_mem_cgroup(memcg, flags);
2293 goto again;
2294 }
2295 *locked = true;
2296}
2297
2298void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2299{
2300 struct page_cgroup *pc = lookup_page_cgroup(page);
2301
2302 /*
2303 * It's guaranteed that pc->mem_cgroup never changes while
2304 * lock is held because a routine modifies pc->mem_cgroup
da92c47d 2305 * should take move_lock_mem_cgroup().
89c06bd5
KH
2306 */
2307 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2308}
2309
2a7106f2
GT
2310void mem_cgroup_update_page_stat(struct page *page,
2311 enum mem_cgroup_page_stat_item idx, int val)
d69b042f 2312{
c0ff4b85 2313 struct mem_cgroup *memcg;
32047e2a 2314 struct page_cgroup *pc = lookup_page_cgroup(page);
dbd4ea78 2315 unsigned long uninitialized_var(flags);
d69b042f 2316
cfa44946 2317 if (mem_cgroup_disabled())
d69b042f 2318 return;
89c06bd5 2319
c0ff4b85
R
2320 memcg = pc->mem_cgroup;
2321 if (unlikely(!memcg || !PageCgroupUsed(pc)))
89c06bd5 2322 return;
26174efd 2323
26174efd 2324 switch (idx) {
2a7106f2 2325 case MEMCG_NR_FILE_MAPPED:
2a7106f2 2326 idx = MEM_CGROUP_STAT_FILE_MAPPED;
26174efd
KH
2327 break;
2328 default:
2329 BUG();
8725d541 2330 }
d69b042f 2331
c0ff4b85 2332 this_cpu_add(memcg->stat->count[idx], val);
d69b042f 2333}
26174efd 2334
cdec2e42
KH
2335/*
2336 * size of first charge trial. "32" comes from vmscan.c's magic value.
2337 * TODO: maybe necessary to use big numbers in big irons.
2338 */
7ec99d62 2339#define CHARGE_BATCH 32U
cdec2e42
KH
2340struct memcg_stock_pcp {
2341 struct mem_cgroup *cached; /* this never be root cgroup */
11c9ea4e 2342 unsigned int nr_pages;
cdec2e42 2343 struct work_struct work;
26fe6168 2344 unsigned long flags;
a0db00fc 2345#define FLUSHING_CACHED_CHARGE 0
cdec2e42
KH
2346};
2347static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
9f50fad6 2348static DEFINE_MUTEX(percpu_charge_mutex);
cdec2e42 2349
a0956d54
SS
2350/**
2351 * consume_stock: Try to consume stocked charge on this cpu.
2352 * @memcg: memcg to consume from.
2353 * @nr_pages: how many pages to charge.
2354 *
2355 * The charges will only happen if @memcg matches the current cpu's memcg
2356 * stock, and at least @nr_pages are available in that stock. Failure to
2357 * service an allocation will refill the stock.
2358 *
2359 * returns true if successful, false otherwise.
cdec2e42 2360 */
a0956d54 2361static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
cdec2e42
KH
2362{
2363 struct memcg_stock_pcp *stock;
2364 bool ret = true;
2365
a0956d54
SS
2366 if (nr_pages > CHARGE_BATCH)
2367 return false;
2368
cdec2e42 2369 stock = &get_cpu_var(memcg_stock);
a0956d54
SS
2370 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2371 stock->nr_pages -= nr_pages;
cdec2e42
KH
2372 else /* need to call res_counter_charge */
2373 ret = false;
2374 put_cpu_var(memcg_stock);
2375 return ret;
2376}
2377
2378/*
2379 * Returns stocks cached in percpu to res_counter and reset cached information.
2380 */
2381static void drain_stock(struct memcg_stock_pcp *stock)
2382{
2383 struct mem_cgroup *old = stock->cached;
2384
11c9ea4e
JW
2385 if (stock->nr_pages) {
2386 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2387
2388 res_counter_uncharge(&old->res, bytes);
cdec2e42 2389 if (do_swap_account)
11c9ea4e
JW
2390 res_counter_uncharge(&old->memsw, bytes);
2391 stock->nr_pages = 0;
cdec2e42
KH
2392 }
2393 stock->cached = NULL;
cdec2e42
KH
2394}
2395
2396/*
2397 * This must be called under preempt disabled or must be called by
2398 * a thread which is pinned to local cpu.
2399 */
2400static void drain_local_stock(struct work_struct *dummy)
2401{
2402 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2403 drain_stock(stock);
26fe6168 2404 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
cdec2e42
KH
2405}
2406
e4777496
MH
2407static void __init memcg_stock_init(void)
2408{
2409 int cpu;
2410
2411 for_each_possible_cpu(cpu) {
2412 struct memcg_stock_pcp *stock =
2413 &per_cpu(memcg_stock, cpu);
2414 INIT_WORK(&stock->work, drain_local_stock);
2415 }
2416}
2417
cdec2e42
KH
2418/*
2419 * Cache charges(val) which is from res_counter, to local per_cpu area.
320cc51d 2420 * This will be consumed by consume_stock() function, later.
cdec2e42 2421 */
c0ff4b85 2422static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
cdec2e42
KH
2423{
2424 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2425
c0ff4b85 2426 if (stock->cached != memcg) { /* reset if necessary */
cdec2e42 2427 drain_stock(stock);
c0ff4b85 2428 stock->cached = memcg;
cdec2e42 2429 }
11c9ea4e 2430 stock->nr_pages += nr_pages;
cdec2e42
KH
2431 put_cpu_var(memcg_stock);
2432}
2433
2434/*
c0ff4b85 2435 * Drains all per-CPU charge caches for given root_memcg resp. subtree
d38144b7
MH
2436 * of the hierarchy under it. sync flag says whether we should block
2437 * until the work is done.
cdec2e42 2438 */
c0ff4b85 2439static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
cdec2e42 2440{
26fe6168 2441 int cpu, curcpu;
d38144b7 2442
cdec2e42 2443 /* Notify other cpus that system-wide "drain" is running */
cdec2e42 2444 get_online_cpus();
5af12d0e 2445 curcpu = get_cpu();
cdec2e42
KH
2446 for_each_online_cpu(cpu) {
2447 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
c0ff4b85 2448 struct mem_cgroup *memcg;
26fe6168 2449
c0ff4b85
R
2450 memcg = stock->cached;
2451 if (!memcg || !stock->nr_pages)
26fe6168 2452 continue;
c0ff4b85 2453 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
3e92041d 2454 continue;
d1a05b69
MH
2455 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2456 if (cpu == curcpu)
2457 drain_local_stock(&stock->work);
2458 else
2459 schedule_work_on(cpu, &stock->work);
2460 }
cdec2e42 2461 }
5af12d0e 2462 put_cpu();
d38144b7
MH
2463
2464 if (!sync)
2465 goto out;
2466
2467 for_each_online_cpu(cpu) {
2468 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
9f50fad6 2469 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
d38144b7
MH
2470 flush_work(&stock->work);
2471 }
2472out:
cdec2e42 2473 put_online_cpus();
d38144b7
MH
2474}
2475
2476/*
2477 * Tries to drain stocked charges in other cpus. This function is asynchronous
2478 * and just put a work per cpu for draining localy on each cpu. Caller can
2479 * expects some charges will be back to res_counter later but cannot wait for
2480 * it.
2481 */
c0ff4b85 2482static void drain_all_stock_async(struct mem_cgroup *root_memcg)
d38144b7 2483{
9f50fad6
MH
2484 /*
2485 * If someone calls draining, avoid adding more kworker runs.
2486 */
2487 if (!mutex_trylock(&percpu_charge_mutex))
2488 return;
c0ff4b85 2489 drain_all_stock(root_memcg, false);
9f50fad6 2490 mutex_unlock(&percpu_charge_mutex);
cdec2e42
KH
2491}
2492
2493/* This is a synchronous drain interface. */
c0ff4b85 2494static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
cdec2e42
KH
2495{
2496 /* called when force_empty is called */
9f50fad6 2497 mutex_lock(&percpu_charge_mutex);
c0ff4b85 2498 drain_all_stock(root_memcg, true);
9f50fad6 2499 mutex_unlock(&percpu_charge_mutex);
cdec2e42
KH
2500}
2501
711d3d2c
KH
2502/*
2503 * This function drains percpu counter value from DEAD cpu and
2504 * move it to local cpu. Note that this function can be preempted.
2505 */
c0ff4b85 2506static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
711d3d2c
KH
2507{
2508 int i;
2509
c0ff4b85 2510 spin_lock(&memcg->pcp_counter_lock);
6104621d 2511 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
c0ff4b85 2512 long x = per_cpu(memcg->stat->count[i], cpu);
711d3d2c 2513
c0ff4b85
R
2514 per_cpu(memcg->stat->count[i], cpu) = 0;
2515 memcg->nocpu_base.count[i] += x;
711d3d2c 2516 }
e9f8974f 2517 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
c0ff4b85 2518 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
e9f8974f 2519
c0ff4b85
R
2520 per_cpu(memcg->stat->events[i], cpu) = 0;
2521 memcg->nocpu_base.events[i] += x;
e9f8974f 2522 }
c0ff4b85 2523 spin_unlock(&memcg->pcp_counter_lock);
711d3d2c
KH
2524}
2525
2526static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
cdec2e42
KH
2527 unsigned long action,
2528 void *hcpu)
2529{
2530 int cpu = (unsigned long)hcpu;
2531 struct memcg_stock_pcp *stock;
711d3d2c 2532 struct mem_cgroup *iter;
cdec2e42 2533
619d094b 2534 if (action == CPU_ONLINE)
1489ebad 2535 return NOTIFY_OK;
1489ebad 2536
d833049b 2537 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
cdec2e42 2538 return NOTIFY_OK;
711d3d2c 2539
9f3a0d09 2540 for_each_mem_cgroup(iter)
711d3d2c
KH
2541 mem_cgroup_drain_pcp_counter(iter, cpu);
2542
cdec2e42
KH
2543 stock = &per_cpu(memcg_stock, cpu);
2544 drain_stock(stock);
2545 return NOTIFY_OK;
2546}
2547
4b534334
KH
2548
2549/* See __mem_cgroup_try_charge() for details */
2550enum {
2551 CHARGE_OK, /* success */
2552 CHARGE_RETRY, /* need to retry but retry is not bad */
2553 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2554 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2555 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2556};
2557
c0ff4b85 2558static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
4c9c5359
SS
2559 unsigned int nr_pages, unsigned int min_pages,
2560 bool oom_check)
4b534334 2561{
7ec99d62 2562 unsigned long csize = nr_pages * PAGE_SIZE;
4b534334
KH
2563 struct mem_cgroup *mem_over_limit;
2564 struct res_counter *fail_res;
2565 unsigned long flags = 0;
2566 int ret;
2567
c0ff4b85 2568 ret = res_counter_charge(&memcg->res, csize, &fail_res);
4b534334
KH
2569
2570 if (likely(!ret)) {
2571 if (!do_swap_account)
2572 return CHARGE_OK;
c0ff4b85 2573 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
4b534334
KH
2574 if (likely(!ret))
2575 return CHARGE_OK;
2576
c0ff4b85 2577 res_counter_uncharge(&memcg->res, csize);
4b534334
KH
2578 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2579 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2580 } else
2581 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
9221edb7 2582 /*
9221edb7
JW
2583 * Never reclaim on behalf of optional batching, retry with a
2584 * single page instead.
2585 */
4c9c5359 2586 if (nr_pages > min_pages)
4b534334
KH
2587 return CHARGE_RETRY;
2588
2589 if (!(gfp_mask & __GFP_WAIT))
2590 return CHARGE_WOULDBLOCK;
2591
4c9c5359
SS
2592 if (gfp_mask & __GFP_NORETRY)
2593 return CHARGE_NOMEM;
2594
5660048c 2595 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
7ec99d62 2596 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
19942822 2597 return CHARGE_RETRY;
4b534334 2598 /*
19942822
JW
2599 * Even though the limit is exceeded at this point, reclaim
2600 * may have been able to free some pages. Retry the charge
2601 * before killing the task.
2602 *
2603 * Only for regular pages, though: huge pages are rather
2604 * unlikely to succeed so close to the limit, and we fall back
2605 * to regular pages anyway in case of failure.
4b534334 2606 */
4c9c5359 2607 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
4b534334
KH
2608 return CHARGE_RETRY;
2609
2610 /*
2611 * At task move, charge accounts can be doubly counted. So, it's
2612 * better to wait until the end of task_move if something is going on.
2613 */
2614 if (mem_cgroup_wait_acct_move(mem_over_limit))
2615 return CHARGE_RETRY;
2616
2617 /* If we don't need to call oom-killer at el, return immediately */
2618 if (!oom_check)
2619 return CHARGE_NOMEM;
2620 /* check OOM */
e845e199 2621 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
4b534334
KH
2622 return CHARGE_OOM_DIE;
2623
2624 return CHARGE_RETRY;
2625}
2626
f817ed48 2627/*
38c5d72f
KH
2628 * __mem_cgroup_try_charge() does
2629 * 1. detect memcg to be charged against from passed *mm and *ptr,
2630 * 2. update res_counter
2631 * 3. call memory reclaim if necessary.
2632 *
2633 * In some special case, if the task is fatal, fatal_signal_pending() or
2634 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2635 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2636 * as possible without any hazards. 2: all pages should have a valid
2637 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2638 * pointer, that is treated as a charge to root_mem_cgroup.
2639 *
2640 * So __mem_cgroup_try_charge() will return
2641 * 0 ... on success, filling *ptr with a valid memcg pointer.
2642 * -ENOMEM ... charge failure because of resource limits.
2643 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2644 *
2645 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2646 * the oom-killer can be invoked.
8a9f3ccd 2647 */
f817ed48 2648static int __mem_cgroup_try_charge(struct mm_struct *mm,
ec168510 2649 gfp_t gfp_mask,
7ec99d62 2650 unsigned int nr_pages,
c0ff4b85 2651 struct mem_cgroup **ptr,
7ec99d62 2652 bool oom)
8a9f3ccd 2653{
7ec99d62 2654 unsigned int batch = max(CHARGE_BATCH, nr_pages);
4b534334 2655 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
c0ff4b85 2656 struct mem_cgroup *memcg = NULL;
4b534334 2657 int ret;
a636b327 2658
867578cb
KH
2659 /*
2660 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2661 * in system level. So, allow to go ahead dying process in addition to
2662 * MEMDIE process.
2663 */
2664 if (unlikely(test_thread_flag(TIF_MEMDIE)
2665 || fatal_signal_pending(current)))
2666 goto bypass;
a636b327 2667
8a9f3ccd 2668 /*
3be91277
HD
2669 * We always charge the cgroup the mm_struct belongs to.
2670 * The mm_struct's mem_cgroup changes on task migration if the
8a9f3ccd 2671 * thread group leader migrates. It's possible that mm is not
24467cac 2672 * set, if so charge the root memcg (happens for pagecache usage).
8a9f3ccd 2673 */
c0ff4b85 2674 if (!*ptr && !mm)
38c5d72f 2675 *ptr = root_mem_cgroup;
f75ca962 2676again:
c0ff4b85
R
2677 if (*ptr) { /* css should be a valid one */
2678 memcg = *ptr;
c0ff4b85 2679 if (mem_cgroup_is_root(memcg))
f75ca962 2680 goto done;
a0956d54 2681 if (consume_stock(memcg, nr_pages))
f75ca962 2682 goto done;
c0ff4b85 2683 css_get(&memcg->css);
4b534334 2684 } else {
f75ca962 2685 struct task_struct *p;
54595fe2 2686
f75ca962
KH
2687 rcu_read_lock();
2688 p = rcu_dereference(mm->owner);
f75ca962 2689 /*
ebb76ce1 2690 * Because we don't have task_lock(), "p" can exit.
c0ff4b85 2691 * In that case, "memcg" can point to root or p can be NULL with
ebb76ce1
KH
2692 * race with swapoff. Then, we have small risk of mis-accouning.
2693 * But such kind of mis-account by race always happens because
2694 * we don't have cgroup_mutex(). It's overkill and we allo that
2695 * small race, here.
2696 * (*) swapoff at el will charge against mm-struct not against
2697 * task-struct. So, mm->owner can be NULL.
f75ca962 2698 */
c0ff4b85 2699 memcg = mem_cgroup_from_task(p);
38c5d72f
KH
2700 if (!memcg)
2701 memcg = root_mem_cgroup;
2702 if (mem_cgroup_is_root(memcg)) {
f75ca962
KH
2703 rcu_read_unlock();
2704 goto done;
2705 }
a0956d54 2706 if (consume_stock(memcg, nr_pages)) {
f75ca962
KH
2707 /*
2708 * It seems dagerous to access memcg without css_get().
2709 * But considering how consume_stok works, it's not
2710 * necessary. If consume_stock success, some charges
2711 * from this memcg are cached on this cpu. So, we
2712 * don't need to call css_get()/css_tryget() before
2713 * calling consume_stock().
2714 */
2715 rcu_read_unlock();
2716 goto done;
2717 }
2718 /* after here, we may be blocked. we need to get refcnt */
c0ff4b85 2719 if (!css_tryget(&memcg->css)) {
f75ca962
KH
2720 rcu_read_unlock();
2721 goto again;
2722 }
2723 rcu_read_unlock();
2724 }
8a9f3ccd 2725
4b534334
KH
2726 do {
2727 bool oom_check;
7a81b88c 2728
4b534334 2729 /* If killed, bypass charge */
f75ca962 2730 if (fatal_signal_pending(current)) {
c0ff4b85 2731 css_put(&memcg->css);
4b534334 2732 goto bypass;
f75ca962 2733 }
6d61ef40 2734
4b534334
KH
2735 oom_check = false;
2736 if (oom && !nr_oom_retries) {
2737 oom_check = true;
2738 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
cdec2e42 2739 }
66e1707b 2740
4c9c5359
SS
2741 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2742 oom_check);
4b534334
KH
2743 switch (ret) {
2744 case CHARGE_OK:
2745 break;
2746 case CHARGE_RETRY: /* not in OOM situation but retry */
7ec99d62 2747 batch = nr_pages;
c0ff4b85
R
2748 css_put(&memcg->css);
2749 memcg = NULL;
f75ca962 2750 goto again;
4b534334 2751 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
c0ff4b85 2752 css_put(&memcg->css);
4b534334
KH
2753 goto nomem;
2754 case CHARGE_NOMEM: /* OOM routine works */
f75ca962 2755 if (!oom) {
c0ff4b85 2756 css_put(&memcg->css);
867578cb 2757 goto nomem;
f75ca962 2758 }
4b534334
KH
2759 /* If oom, we never return -ENOMEM */
2760 nr_oom_retries--;
2761 break;
2762 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
c0ff4b85 2763 css_put(&memcg->css);
867578cb 2764 goto bypass;
66e1707b 2765 }
4b534334
KH
2766 } while (ret != CHARGE_OK);
2767
7ec99d62 2768 if (batch > nr_pages)
c0ff4b85
R
2769 refill_stock(memcg, batch - nr_pages);
2770 css_put(&memcg->css);
0c3e73e8 2771done:
c0ff4b85 2772 *ptr = memcg;
7a81b88c
KH
2773 return 0;
2774nomem:
c0ff4b85 2775 *ptr = NULL;
7a81b88c 2776 return -ENOMEM;
867578cb 2777bypass:
38c5d72f
KH
2778 *ptr = root_mem_cgroup;
2779 return -EINTR;
7a81b88c 2780}
8a9f3ccd 2781
a3032a2c
DN
2782/*
2783 * Somemtimes we have to undo a charge we got by try_charge().
2784 * This function is for that and do uncharge, put css's refcnt.
2785 * gotten by try_charge().
2786 */
c0ff4b85 2787static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
e7018b8d 2788 unsigned int nr_pages)
a3032a2c 2789{
c0ff4b85 2790 if (!mem_cgroup_is_root(memcg)) {
e7018b8d
JW
2791 unsigned long bytes = nr_pages * PAGE_SIZE;
2792
c0ff4b85 2793 res_counter_uncharge(&memcg->res, bytes);
a3032a2c 2794 if (do_swap_account)
c0ff4b85 2795 res_counter_uncharge(&memcg->memsw, bytes);
a3032a2c 2796 }
854ffa8d
DN
2797}
2798
d01dd17f
KH
2799/*
2800 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2801 * This is useful when moving usage to parent cgroup.
2802 */
2803static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2804 unsigned int nr_pages)
2805{
2806 unsigned long bytes = nr_pages * PAGE_SIZE;
2807
2808 if (mem_cgroup_is_root(memcg))
2809 return;
2810
2811 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2812 if (do_swap_account)
2813 res_counter_uncharge_until(&memcg->memsw,
2814 memcg->memsw.parent, bytes);
2815}
2816
a3b2d692
KH
2817/*
2818 * A helper function to get mem_cgroup from ID. must be called under
e9316080
TH
2819 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2820 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2821 * called against removed memcg.)
a3b2d692
KH
2822 */
2823static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2824{
2825 struct cgroup_subsys_state *css;
2826
2827 /* ID 0 is unused ID */
2828 if (!id)
2829 return NULL;
2830 css = css_lookup(&mem_cgroup_subsys, id);
2831 if (!css)
2832 return NULL;
b2145145 2833 return mem_cgroup_from_css(css);
a3b2d692
KH
2834}
2835
e42d9d5d 2836struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
b5a84319 2837{
c0ff4b85 2838 struct mem_cgroup *memcg = NULL;
3c776e64 2839 struct page_cgroup *pc;
a3b2d692 2840 unsigned short id;
b5a84319
KH
2841 swp_entry_t ent;
2842
3c776e64
DN
2843 VM_BUG_ON(!PageLocked(page));
2844
3c776e64 2845 pc = lookup_page_cgroup(page);
c0bd3f63 2846 lock_page_cgroup(pc);
a3b2d692 2847 if (PageCgroupUsed(pc)) {
c0ff4b85
R
2848 memcg = pc->mem_cgroup;
2849 if (memcg && !css_tryget(&memcg->css))
2850 memcg = NULL;
e42d9d5d 2851 } else if (PageSwapCache(page)) {
3c776e64 2852 ent.val = page_private(page);
9fb4b7cc 2853 id = lookup_swap_cgroup_id(ent);
a3b2d692 2854 rcu_read_lock();
c0ff4b85
R
2855 memcg = mem_cgroup_lookup(id);
2856 if (memcg && !css_tryget(&memcg->css))
2857 memcg = NULL;
a3b2d692 2858 rcu_read_unlock();
3c776e64 2859 }
c0bd3f63 2860 unlock_page_cgroup(pc);
c0ff4b85 2861 return memcg;
b5a84319
KH
2862}
2863
c0ff4b85 2864static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
5564e88b 2865 struct page *page,
7ec99d62 2866 unsigned int nr_pages,
9ce70c02
HD
2867 enum charge_type ctype,
2868 bool lrucare)
7a81b88c 2869{
ce587e65 2870 struct page_cgroup *pc = lookup_page_cgroup(page);
9ce70c02 2871 struct zone *uninitialized_var(zone);
fa9add64 2872 struct lruvec *lruvec;
9ce70c02 2873 bool was_on_lru = false;
b2402857 2874 bool anon;
9ce70c02 2875
ca3e0214 2876 lock_page_cgroup(pc);
90deb788 2877 VM_BUG_ON(PageCgroupUsed(pc));
ca3e0214
KH
2878 /*
2879 * we don't need page_cgroup_lock about tail pages, becase they are not
2880 * accessed by any other context at this point.
2881 */
9ce70c02
HD
2882
2883 /*
2884 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2885 * may already be on some other mem_cgroup's LRU. Take care of it.
2886 */
2887 if (lrucare) {
2888 zone = page_zone(page);
2889 spin_lock_irq(&zone->lru_lock);
2890 if (PageLRU(page)) {
fa9add64 2891 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
9ce70c02 2892 ClearPageLRU(page);
fa9add64 2893 del_page_from_lru_list(page, lruvec, page_lru(page));
9ce70c02
HD
2894 was_on_lru = true;
2895 }
2896 }
2897
c0ff4b85 2898 pc->mem_cgroup = memcg;
261fb61a
KH
2899 /*
2900 * We access a page_cgroup asynchronously without lock_page_cgroup().
2901 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2902 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2903 * before USED bit, we need memory barrier here.
2904 * See mem_cgroup_add_lru_list(), etc.
2905 */
08e552c6 2906 smp_wmb();
b2402857 2907 SetPageCgroupUsed(pc);
3be91277 2908
9ce70c02
HD
2909 if (lrucare) {
2910 if (was_on_lru) {
fa9add64 2911 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
9ce70c02
HD
2912 VM_BUG_ON(PageLRU(page));
2913 SetPageLRU(page);
fa9add64 2914 add_page_to_lru_list(page, lruvec, page_lru(page));
9ce70c02
HD
2915 }
2916 spin_unlock_irq(&zone->lru_lock);
2917 }
2918
41326c17 2919 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
b2402857
KH
2920 anon = true;
2921 else
2922 anon = false;
2923
b070e65c 2924 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
52d4b9ac 2925 unlock_page_cgroup(pc);
9ce70c02 2926
430e4863
KH
2927 /*
2928 * "charge_statistics" updated event counter. Then, check it.
2929 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2930 * if they exceeds softlimit.
2931 */
c0ff4b85 2932 memcg_check_events(memcg, page);
7a81b88c 2933}
66e1707b 2934
7cf27982
GC
2935static DEFINE_MUTEX(set_limit_mutex);
2936
7ae1e1d0
GC
2937#ifdef CONFIG_MEMCG_KMEM
2938static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2939{
2940 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2941 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2942}
2943
1f458cbf
GC
2944/*
2945 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2946 * in the memcg_cache_params struct.
2947 */
2948static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2949{
2950 struct kmem_cache *cachep;
2951
2952 VM_BUG_ON(p->is_root_cache);
2953 cachep = p->root_cache;
2954 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2955}
2956
749c5415
GC
2957#ifdef CONFIG_SLABINFO
2958static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2959 struct seq_file *m)
2960{
2961 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2962 struct memcg_cache_params *params;
2963
2964 if (!memcg_can_account_kmem(memcg))
2965 return -EIO;
2966
2967 print_slabinfo_header(m);
2968
2969 mutex_lock(&memcg->slab_caches_mutex);
2970 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2971 cache_show(memcg_params_to_cache(params), m);
2972 mutex_unlock(&memcg->slab_caches_mutex);
2973
2974 return 0;
2975}
2976#endif
2977
7ae1e1d0
GC
2978static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2979{
2980 struct res_counter *fail_res;
2981 struct mem_cgroup *_memcg;
2982 int ret = 0;
2983 bool may_oom;
2984
2985 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2986 if (ret)
2987 return ret;
2988
2989 /*
2990 * Conditions under which we can wait for the oom_killer. Those are
2991 * the same conditions tested by the core page allocator
2992 */
2993 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2994
2995 _memcg = memcg;
2996 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2997 &_memcg, may_oom);
2998
2999 if (ret == -EINTR) {
3000 /*
3001 * __mem_cgroup_try_charge() chosed to bypass to root due to
3002 * OOM kill or fatal signal. Since our only options are to
3003 * either fail the allocation or charge it to this cgroup, do
3004 * it as a temporary condition. But we can't fail. From a
3005 * kmem/slab perspective, the cache has already been selected,
3006 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3007 * our minds.
3008 *
3009 * This condition will only trigger if the task entered
3010 * memcg_charge_kmem in a sane state, but was OOM-killed during
3011 * __mem_cgroup_try_charge() above. Tasks that were already
3012 * dying when the allocation triggers should have been already
3013 * directed to the root cgroup in memcontrol.h
3014 */
3015 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3016 if (do_swap_account)
3017 res_counter_charge_nofail(&memcg->memsw, size,
3018 &fail_res);
3019 ret = 0;
3020 } else if (ret)
3021 res_counter_uncharge(&memcg->kmem, size);
3022
3023 return ret;
3024}
3025
3026static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3027{
7ae1e1d0
GC
3028 res_counter_uncharge(&memcg->res, size);
3029 if (do_swap_account)
3030 res_counter_uncharge(&memcg->memsw, size);
7de37682
GC
3031
3032 /* Not down to 0 */
3033 if (res_counter_uncharge(&memcg->kmem, size))
3034 return;
3035
3036 if (memcg_kmem_test_and_clear_dead(memcg))
3037 mem_cgroup_put(memcg);
7ae1e1d0
GC
3038}
3039
2633d7a0
GC
3040void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3041{
3042 if (!memcg)
3043 return;
3044
3045 mutex_lock(&memcg->slab_caches_mutex);
3046 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3047 mutex_unlock(&memcg->slab_caches_mutex);
3048}
3049
3050/*
3051 * helper for acessing a memcg's index. It will be used as an index in the
3052 * child cache array in kmem_cache, and also to derive its name. This function
3053 * will return -1 when this is not a kmem-limited memcg.
3054 */
3055int memcg_cache_id(struct mem_cgroup *memcg)
3056{
3057 return memcg ? memcg->kmemcg_id : -1;
3058}
3059
55007d84
GC
3060/*
3061 * This ends up being protected by the set_limit mutex, during normal
3062 * operation, because that is its main call site.
3063 *
3064 * But when we create a new cache, we can call this as well if its parent
3065 * is kmem-limited. That will have to hold set_limit_mutex as well.
3066 */
3067int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3068{
3069 int num, ret;
3070
3071 num = ida_simple_get(&kmem_limited_groups,
3072 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3073 if (num < 0)
3074 return num;
3075 /*
3076 * After this point, kmem_accounted (that we test atomically in
3077 * the beginning of this conditional), is no longer 0. This
3078 * guarantees only one process will set the following boolean
3079 * to true. We don't need test_and_set because we're protected
3080 * by the set_limit_mutex anyway.
3081 */
3082 memcg_kmem_set_activated(memcg);
3083
3084 ret = memcg_update_all_caches(num+1);
3085 if (ret) {
3086 ida_simple_remove(&kmem_limited_groups, num);
3087 memcg_kmem_clear_activated(memcg);
3088 return ret;
3089 }
3090
3091 memcg->kmemcg_id = num;
3092 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3093 mutex_init(&memcg->slab_caches_mutex);
3094 return 0;
3095}
3096
3097static size_t memcg_caches_array_size(int num_groups)
3098{
3099 ssize_t size;
3100 if (num_groups <= 0)
3101 return 0;
3102
3103 size = 2 * num_groups;
3104 if (size < MEMCG_CACHES_MIN_SIZE)
3105 size = MEMCG_CACHES_MIN_SIZE;
3106 else if (size > MEMCG_CACHES_MAX_SIZE)
3107 size = MEMCG_CACHES_MAX_SIZE;
3108
3109 return size;
3110}
3111
3112/*
3113 * We should update the current array size iff all caches updates succeed. This
3114 * can only be done from the slab side. The slab mutex needs to be held when
3115 * calling this.
3116 */
3117void memcg_update_array_size(int num)
3118{
3119 if (num > memcg_limited_groups_array_size)
3120 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3121}
3122
15cf17d2
KK
3123static void kmem_cache_destroy_work_func(struct work_struct *w);
3124
55007d84
GC
3125int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3126{
3127 struct memcg_cache_params *cur_params = s->memcg_params;
3128
3129 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3130
3131 if (num_groups > memcg_limited_groups_array_size) {
3132 int i;
3133 ssize_t size = memcg_caches_array_size(num_groups);
3134
3135 size *= sizeof(void *);
3136 size += sizeof(struct memcg_cache_params);
3137
3138 s->memcg_params = kzalloc(size, GFP_KERNEL);
3139 if (!s->memcg_params) {
3140 s->memcg_params = cur_params;
3141 return -ENOMEM;
3142 }
3143
15cf17d2
KK
3144 INIT_WORK(&s->memcg_params->destroy,
3145 kmem_cache_destroy_work_func);
55007d84
GC
3146 s->memcg_params->is_root_cache = true;
3147
3148 /*
3149 * There is the chance it will be bigger than
3150 * memcg_limited_groups_array_size, if we failed an allocation
3151 * in a cache, in which case all caches updated before it, will
3152 * have a bigger array.
3153 *
3154 * But if that is the case, the data after
3155 * memcg_limited_groups_array_size is certainly unused
3156 */
3157 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3158 if (!cur_params->memcg_caches[i])
3159 continue;
3160 s->memcg_params->memcg_caches[i] =
3161 cur_params->memcg_caches[i];
3162 }
3163
3164 /*
3165 * Ideally, we would wait until all caches succeed, and only
3166 * then free the old one. But this is not worth the extra
3167 * pointer per-cache we'd have to have for this.
3168 *
3169 * It is not a big deal if some caches are left with a size
3170 * bigger than the others. And all updates will reset this
3171 * anyway.
3172 */
3173 kfree(cur_params);
3174 }
3175 return 0;
3176}
3177
943a451a
GC
3178int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3179 struct kmem_cache *root_cache)
2633d7a0
GC
3180{
3181 size_t size = sizeof(struct memcg_cache_params);
3182
3183 if (!memcg_kmem_enabled())
3184 return 0;
3185
55007d84
GC
3186 if (!memcg)
3187 size += memcg_limited_groups_array_size * sizeof(void *);
3188
2633d7a0
GC
3189 s->memcg_params = kzalloc(size, GFP_KERNEL);
3190 if (!s->memcg_params)
3191 return -ENOMEM;
3192
15cf17d2
KK
3193 INIT_WORK(&s->memcg_params->destroy,
3194 kmem_cache_destroy_work_func);
943a451a 3195 if (memcg) {
2633d7a0 3196 s->memcg_params->memcg = memcg;
943a451a 3197 s->memcg_params->root_cache = root_cache;
4ba902b5
GC
3198 } else
3199 s->memcg_params->is_root_cache = true;
3200
2633d7a0
GC
3201 return 0;
3202}
3203
3204void memcg_release_cache(struct kmem_cache *s)
3205{
d7f25f8a
GC
3206 struct kmem_cache *root;
3207 struct mem_cgroup *memcg;
3208 int id;
3209
3210 /*
3211 * This happens, for instance, when a root cache goes away before we
3212 * add any memcg.
3213 */
3214 if (!s->memcg_params)
3215 return;
3216
3217 if (s->memcg_params->is_root_cache)
3218 goto out;
3219
3220 memcg = s->memcg_params->memcg;
3221 id = memcg_cache_id(memcg);
3222
3223 root = s->memcg_params->root_cache;
3224 root->memcg_params->memcg_caches[id] = NULL;
d7f25f8a
GC
3225
3226 mutex_lock(&memcg->slab_caches_mutex);
3227 list_del(&s->memcg_params->list);
3228 mutex_unlock(&memcg->slab_caches_mutex);
3229
fd0ccaf2 3230 mem_cgroup_put(memcg);
d7f25f8a 3231out:
2633d7a0
GC
3232 kfree(s->memcg_params);
3233}
3234
0e9d92f2
GC
3235/*
3236 * During the creation a new cache, we need to disable our accounting mechanism
3237 * altogether. This is true even if we are not creating, but rather just
3238 * enqueing new caches to be created.
3239 *
3240 * This is because that process will trigger allocations; some visible, like
3241 * explicit kmallocs to auxiliary data structures, name strings and internal
3242 * cache structures; some well concealed, like INIT_WORK() that can allocate
3243 * objects during debug.
3244 *
3245 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3246 * to it. This may not be a bounded recursion: since the first cache creation
3247 * failed to complete (waiting on the allocation), we'll just try to create the
3248 * cache again, failing at the same point.
3249 *
3250 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3251 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3252 * inside the following two functions.
3253 */
3254static inline void memcg_stop_kmem_account(void)
3255{
3256 VM_BUG_ON(!current->mm);
3257 current->memcg_kmem_skip_account++;
3258}
3259
3260static inline void memcg_resume_kmem_account(void)
3261{
3262 VM_BUG_ON(!current->mm);
3263 current->memcg_kmem_skip_account--;
3264}
3265
1f458cbf
GC
3266static void kmem_cache_destroy_work_func(struct work_struct *w)
3267{
3268 struct kmem_cache *cachep;
3269 struct memcg_cache_params *p;
3270
3271 p = container_of(w, struct memcg_cache_params, destroy);
3272
3273 cachep = memcg_params_to_cache(p);
3274
22933152
GC
3275 /*
3276 * If we get down to 0 after shrink, we could delete right away.
3277 * However, memcg_release_pages() already puts us back in the workqueue
3278 * in that case. If we proceed deleting, we'll get a dangling
3279 * reference, and removing the object from the workqueue in that case
3280 * is unnecessary complication. We are not a fast path.
3281 *
3282 * Note that this case is fundamentally different from racing with
3283 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3284 * kmem_cache_shrink, not only we would be reinserting a dead cache
3285 * into the queue, but doing so from inside the worker racing to
3286 * destroy it.
3287 *
3288 * So if we aren't down to zero, we'll just schedule a worker and try
3289 * again
3290 */
3291 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3292 kmem_cache_shrink(cachep);
3293 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3294 return;
3295 } else
1f458cbf
GC
3296 kmem_cache_destroy(cachep);
3297}
3298
3299void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3300{
3301 if (!cachep->memcg_params->dead)
3302 return;
3303
22933152
GC
3304 /*
3305 * There are many ways in which we can get here.
3306 *
3307 * We can get to a memory-pressure situation while the delayed work is
3308 * still pending to run. The vmscan shrinkers can then release all
3309 * cache memory and get us to destruction. If this is the case, we'll
3310 * be executed twice, which is a bug (the second time will execute over
3311 * bogus data). In this case, cancelling the work should be fine.
3312 *
3313 * But we can also get here from the worker itself, if
3314 * kmem_cache_shrink is enough to shake all the remaining objects and
3315 * get the page count to 0. In this case, we'll deadlock if we try to
3316 * cancel the work (the worker runs with an internal lock held, which
3317 * is the same lock we would hold for cancel_work_sync().)
3318 *
3319 * Since we can't possibly know who got us here, just refrain from
3320 * running if there is already work pending
3321 */
3322 if (work_pending(&cachep->memcg_params->destroy))
3323 return;
1f458cbf
GC
3324 /*
3325 * We have to defer the actual destroying to a workqueue, because
3326 * we might currently be in a context that cannot sleep.
3327 */
3328 schedule_work(&cachep->memcg_params->destroy);
3329}
3330
d9c10ddd
MH
3331/*
3332 * This lock protects updaters, not readers. We want readers to be as fast as
3333 * they can, and they will either see NULL or a valid cache value. Our model
3334 * allow them to see NULL, in which case the root memcg will be selected.
3335 *
3336 * We need this lock because multiple allocations to the same cache from a non
3337 * will span more than one worker. Only one of them can create the cache.
3338 */
3339static DEFINE_MUTEX(memcg_cache_mutex);
d7f25f8a 3340
d9c10ddd
MH
3341/*
3342 * Called with memcg_cache_mutex held
3343 */
d7f25f8a
GC
3344static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3345 struct kmem_cache *s)
3346{
d7f25f8a 3347 struct kmem_cache *new;
d9c10ddd 3348 static char *tmp_name = NULL;
d7f25f8a 3349
d9c10ddd
MH
3350 lockdep_assert_held(&memcg_cache_mutex);
3351
3352 /*
3353 * kmem_cache_create_memcg duplicates the given name and
3354 * cgroup_name for this name requires RCU context.
3355 * This static temporary buffer is used to prevent from
3356 * pointless shortliving allocation.
3357 */
3358 if (!tmp_name) {
3359 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3360 if (!tmp_name)
3361 return NULL;
3362 }
3363
3364 rcu_read_lock();
3365 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3366 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3367 rcu_read_unlock();
d7f25f8a 3368
d9c10ddd 3369 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
943a451a 3370 (s->flags & ~SLAB_PANIC), s->ctor, s);
d7f25f8a 3371
d79923fa
GC
3372 if (new)
3373 new->allocflags |= __GFP_KMEMCG;
3374
d7f25f8a
GC
3375 return new;
3376}
3377
d7f25f8a
GC
3378static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3379 struct kmem_cache *cachep)
3380{
3381 struct kmem_cache *new_cachep;
3382 int idx;
3383
3384 BUG_ON(!memcg_can_account_kmem(memcg));
3385
3386 idx = memcg_cache_id(memcg);
3387
3388 mutex_lock(&memcg_cache_mutex);
3389 new_cachep = cachep->memcg_params->memcg_caches[idx];
3390 if (new_cachep)
3391 goto out;
3392
3393 new_cachep = kmem_cache_dup(memcg, cachep);
d7f25f8a
GC
3394 if (new_cachep == NULL) {
3395 new_cachep = cachep;
3396 goto out;
3397 }
3398
3399 mem_cgroup_get(memcg);
1f458cbf 3400 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
d7f25f8a
GC
3401
3402 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3403 /*
3404 * the readers won't lock, make sure everybody sees the updated value,
3405 * so they won't put stuff in the queue again for no reason
3406 */
3407 wmb();
3408out:
3409 mutex_unlock(&memcg_cache_mutex);
3410 return new_cachep;
3411}
3412
7cf27982
GC
3413void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3414{
3415 struct kmem_cache *c;
3416 int i;
3417
3418 if (!s->memcg_params)
3419 return;
3420 if (!s->memcg_params->is_root_cache)
3421 return;
3422
3423 /*
3424 * If the cache is being destroyed, we trust that there is no one else
3425 * requesting objects from it. Even if there are, the sanity checks in
3426 * kmem_cache_destroy should caught this ill-case.
3427 *
3428 * Still, we don't want anyone else freeing memcg_caches under our
3429 * noses, which can happen if a new memcg comes to life. As usual,
3430 * we'll take the set_limit_mutex to protect ourselves against this.
3431 */
3432 mutex_lock(&set_limit_mutex);
3433 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3434 c = s->memcg_params->memcg_caches[i];
3435 if (!c)
3436 continue;
3437
3438 /*
3439 * We will now manually delete the caches, so to avoid races
3440 * we need to cancel all pending destruction workers and
3441 * proceed with destruction ourselves.
3442 *
3443 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3444 * and that could spawn the workers again: it is likely that
3445 * the cache still have active pages until this very moment.
3446 * This would lead us back to mem_cgroup_destroy_cache.
3447 *
3448 * But that will not execute at all if the "dead" flag is not
3449 * set, so flip it down to guarantee we are in control.
3450 */
3451 c->memcg_params->dead = false;
22933152 3452 cancel_work_sync(&c->memcg_params->destroy);
7cf27982
GC
3453 kmem_cache_destroy(c);
3454 }
3455 mutex_unlock(&set_limit_mutex);
3456}
3457
d7f25f8a
GC
3458struct create_work {
3459 struct mem_cgroup *memcg;
3460 struct kmem_cache *cachep;
3461 struct work_struct work;
3462};
3463
1f458cbf
GC
3464static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3465{
3466 struct kmem_cache *cachep;
3467 struct memcg_cache_params *params;
3468
3469 if (!memcg_kmem_is_active(memcg))
3470 return;
3471
3472 mutex_lock(&memcg->slab_caches_mutex);
3473 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3474 cachep = memcg_params_to_cache(params);
3475 cachep->memcg_params->dead = true;
1f458cbf
GC
3476 schedule_work(&cachep->memcg_params->destroy);
3477 }
3478 mutex_unlock(&memcg->slab_caches_mutex);
3479}
3480
d7f25f8a
GC
3481static void memcg_create_cache_work_func(struct work_struct *w)
3482{
3483 struct create_work *cw;
3484
3485 cw = container_of(w, struct create_work, work);
3486 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3487 /* Drop the reference gotten when we enqueued. */
3488 css_put(&cw->memcg->css);
3489 kfree(cw);
3490}
3491
3492/*
3493 * Enqueue the creation of a per-memcg kmem_cache.
d7f25f8a 3494 */
0e9d92f2
GC
3495static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3496 struct kmem_cache *cachep)
d7f25f8a
GC
3497{
3498 struct create_work *cw;
3499
3500 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
ca0dde97
LZ
3501 if (cw == NULL) {
3502 css_put(&memcg->css);
d7f25f8a
GC
3503 return;
3504 }
3505
3506 cw->memcg = memcg;
3507 cw->cachep = cachep;
3508
3509 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3510 schedule_work(&cw->work);
3511}
3512
0e9d92f2
GC
3513static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3514 struct kmem_cache *cachep)
3515{
3516 /*
3517 * We need to stop accounting when we kmalloc, because if the
3518 * corresponding kmalloc cache is not yet created, the first allocation
3519 * in __memcg_create_cache_enqueue will recurse.
3520 *
3521 * However, it is better to enclose the whole function. Depending on
3522 * the debugging options enabled, INIT_WORK(), for instance, can
3523 * trigger an allocation. This too, will make us recurse. Because at
3524 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3525 * the safest choice is to do it like this, wrapping the whole function.
3526 */
3527 memcg_stop_kmem_account();
3528 __memcg_create_cache_enqueue(memcg, cachep);
3529 memcg_resume_kmem_account();
3530}
d7f25f8a
GC
3531/*
3532 * Return the kmem_cache we're supposed to use for a slab allocation.
3533 * We try to use the current memcg's version of the cache.
3534 *
3535 * If the cache does not exist yet, if we are the first user of it,
3536 * we either create it immediately, if possible, or create it asynchronously
3537 * in a workqueue.
3538 * In the latter case, we will let the current allocation go through with
3539 * the original cache.
3540 *
3541 * Can't be called in interrupt context or from kernel threads.
3542 * This function needs to be called with rcu_read_lock() held.
3543 */
3544struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3545 gfp_t gfp)
3546{
3547 struct mem_cgroup *memcg;
3548 int idx;
3549
3550 VM_BUG_ON(!cachep->memcg_params);
3551 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3552
0e9d92f2
GC
3553 if (!current->mm || current->memcg_kmem_skip_account)
3554 return cachep;
3555
d7f25f8a
GC
3556 rcu_read_lock();
3557 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
d7f25f8a
GC
3558
3559 if (!memcg_can_account_kmem(memcg))
ca0dde97 3560 goto out;
d7f25f8a
GC
3561
3562 idx = memcg_cache_id(memcg);
3563
3564 /*
3565 * barrier to mare sure we're always seeing the up to date value. The
3566 * code updating memcg_caches will issue a write barrier to match this.
3567 */
3568 read_barrier_depends();
ca0dde97
LZ
3569 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3570 cachep = cachep->memcg_params->memcg_caches[idx];
3571 goto out;
d7f25f8a
GC
3572 }
3573
ca0dde97
LZ
3574 /* The corresponding put will be done in the workqueue. */
3575 if (!css_tryget(&memcg->css))
3576 goto out;
3577 rcu_read_unlock();
3578
3579 /*
3580 * If we are in a safe context (can wait, and not in interrupt
3581 * context), we could be be predictable and return right away.
3582 * This would guarantee that the allocation being performed
3583 * already belongs in the new cache.
3584 *
3585 * However, there are some clashes that can arrive from locking.
3586 * For instance, because we acquire the slab_mutex while doing
3587 * kmem_cache_dup, this means no further allocation could happen
3588 * with the slab_mutex held.
3589 *
3590 * Also, because cache creation issue get_online_cpus(), this
3591 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3592 * that ends up reversed during cpu hotplug. (cpuset allocates
3593 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3594 * better to defer everything.
3595 */
3596 memcg_create_cache_enqueue(memcg, cachep);
3597 return cachep;
3598out:
3599 rcu_read_unlock();
3600 return cachep;
d7f25f8a
GC
3601}
3602EXPORT_SYMBOL(__memcg_kmem_get_cache);
3603
7ae1e1d0
GC
3604/*
3605 * We need to verify if the allocation against current->mm->owner's memcg is
3606 * possible for the given order. But the page is not allocated yet, so we'll
3607 * need a further commit step to do the final arrangements.
3608 *
3609 * It is possible for the task to switch cgroups in this mean time, so at
3610 * commit time, we can't rely on task conversion any longer. We'll then use
3611 * the handle argument to return to the caller which cgroup we should commit
3612 * against. We could also return the memcg directly and avoid the pointer
3613 * passing, but a boolean return value gives better semantics considering
3614 * the compiled-out case as well.
3615 *
3616 * Returning true means the allocation is possible.
3617 */
3618bool
3619__memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3620{
3621 struct mem_cgroup *memcg;
3622 int ret;
3623
3624 *_memcg = NULL;
3625 memcg = try_get_mem_cgroup_from_mm(current->mm);
3626
3627 /*
3628 * very rare case described in mem_cgroup_from_task. Unfortunately there
3629 * isn't much we can do without complicating this too much, and it would
3630 * be gfp-dependent anyway. Just let it go
3631 */
3632 if (unlikely(!memcg))
3633 return true;
3634
3635 if (!memcg_can_account_kmem(memcg)) {
3636 css_put(&memcg->css);
3637 return true;
3638 }
3639
7ae1e1d0
GC
3640 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3641 if (!ret)
3642 *_memcg = memcg;
7ae1e1d0
GC
3643
3644 css_put(&memcg->css);
3645 return (ret == 0);
3646}
3647
3648void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3649 int order)
3650{
3651 struct page_cgroup *pc;
3652
3653 VM_BUG_ON(mem_cgroup_is_root(memcg));
3654
3655 /* The page allocation failed. Revert */
3656 if (!page) {
3657 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
7ae1e1d0
GC
3658 return;
3659 }
3660
3661 pc = lookup_page_cgroup(page);
3662 lock_page_cgroup(pc);
3663 pc->mem_cgroup = memcg;
3664 SetPageCgroupUsed(pc);
3665 unlock_page_cgroup(pc);
3666}
3667
3668void __memcg_kmem_uncharge_pages(struct page *page, int order)
3669{
3670 struct mem_cgroup *memcg = NULL;
3671 struct page_cgroup *pc;
3672
3673
3674 pc = lookup_page_cgroup(page);
3675 /*
3676 * Fast unlocked return. Theoretically might have changed, have to
3677 * check again after locking.
3678 */
3679 if (!PageCgroupUsed(pc))
3680 return;
3681
3682 lock_page_cgroup(pc);
3683 if (PageCgroupUsed(pc)) {
3684 memcg = pc->mem_cgroup;
3685 ClearPageCgroupUsed(pc);
3686 }
3687 unlock_page_cgroup(pc);
3688
3689 /*
3690 * We trust that only if there is a memcg associated with the page, it
3691 * is a valid allocation
3692 */
3693 if (!memcg)
3694 return;
3695
3696 VM_BUG_ON(mem_cgroup_is_root(memcg));
3697 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
7ae1e1d0 3698}
1f458cbf
GC
3699#else
3700static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3701{
3702}
7ae1e1d0
GC
3703#endif /* CONFIG_MEMCG_KMEM */
3704
ca3e0214
KH
3705#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3706
a0db00fc 3707#define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
ca3e0214
KH
3708/*
3709 * Because tail pages are not marked as "used", set it. We're under
e94c8a9c
KH
3710 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3711 * charge/uncharge will be never happen and move_account() is done under
3712 * compound_lock(), so we don't have to take care of races.
ca3e0214 3713 */
e94c8a9c 3714void mem_cgroup_split_huge_fixup(struct page *head)
ca3e0214
KH
3715{
3716 struct page_cgroup *head_pc = lookup_page_cgroup(head);
e94c8a9c 3717 struct page_cgroup *pc;
b070e65c 3718 struct mem_cgroup *memcg;
e94c8a9c 3719 int i;
ca3e0214 3720
3d37c4a9
KH
3721 if (mem_cgroup_disabled())
3722 return;
b070e65c
DR
3723
3724 memcg = head_pc->mem_cgroup;
e94c8a9c
KH
3725 for (i = 1; i < HPAGE_PMD_NR; i++) {
3726 pc = head_pc + i;
b070e65c 3727 pc->mem_cgroup = memcg;
e94c8a9c 3728 smp_wmb();/* see __commit_charge() */
e94c8a9c
KH
3729 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3730 }
b070e65c
DR
3731 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3732 HPAGE_PMD_NR);
ca3e0214 3733}
12d27107 3734#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
ca3e0214 3735
f817ed48 3736/**
de3638d9 3737 * mem_cgroup_move_account - move account of the page
5564e88b 3738 * @page: the page
7ec99d62 3739 * @nr_pages: number of regular pages (>1 for huge pages)
f817ed48
KH
3740 * @pc: page_cgroup of the page.
3741 * @from: mem_cgroup which the page is moved from.
3742 * @to: mem_cgroup which the page is moved to. @from != @to.
3743 *
3744 * The caller must confirm following.
08e552c6 3745 * - page is not on LRU (isolate_page() is useful.)
7ec99d62 3746 * - compound_lock is held when nr_pages > 1
f817ed48 3747 *
2f3479b1
KH
3748 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3749 * from old cgroup.
f817ed48 3750 */
7ec99d62
JW
3751static int mem_cgroup_move_account(struct page *page,
3752 unsigned int nr_pages,
3753 struct page_cgroup *pc,
3754 struct mem_cgroup *from,
2f3479b1 3755 struct mem_cgroup *to)
f817ed48 3756{
de3638d9
JW
3757 unsigned long flags;
3758 int ret;
b2402857 3759 bool anon = PageAnon(page);
987eba66 3760
f817ed48 3761 VM_BUG_ON(from == to);
5564e88b 3762 VM_BUG_ON(PageLRU(page));
de3638d9
JW
3763 /*
3764 * The page is isolated from LRU. So, collapse function
3765 * will not handle this page. But page splitting can happen.
3766 * Do this check under compound_page_lock(). The caller should
3767 * hold it.
3768 */
3769 ret = -EBUSY;
7ec99d62 3770 if (nr_pages > 1 && !PageTransHuge(page))
de3638d9
JW
3771 goto out;
3772
3773 lock_page_cgroup(pc);
3774
3775 ret = -EINVAL;
3776 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3777 goto unlock;
3778
312734c0 3779 move_lock_mem_cgroup(from, &flags);
f817ed48 3780
2ff76f11 3781 if (!anon && page_mapped(page)) {
c62b1a3b
KH
3782 /* Update mapped_file data for mem_cgroup */
3783 preempt_disable();
3784 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3785 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3786 preempt_enable();
d69b042f 3787 }
b070e65c 3788 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
d69b042f 3789
854ffa8d 3790 /* caller should have done css_get */
08e552c6 3791 pc->mem_cgroup = to;
b070e65c 3792 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
312734c0 3793 move_unlock_mem_cgroup(from, &flags);
de3638d9
JW
3794 ret = 0;
3795unlock:
57f9fd7d 3796 unlock_page_cgroup(pc);
d2265e6f
KH
3797 /*
3798 * check events
3799 */
5564e88b
JW
3800 memcg_check_events(to, page);
3801 memcg_check_events(from, page);
de3638d9 3802out:
f817ed48
KH
3803 return ret;
3804}
3805
2ef37d3f
MH
3806/**
3807 * mem_cgroup_move_parent - moves page to the parent group
3808 * @page: the page to move
3809 * @pc: page_cgroup of the page
3810 * @child: page's cgroup
3811 *
3812 * move charges to its parent or the root cgroup if the group has no
3813 * parent (aka use_hierarchy==0).
3814 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3815 * mem_cgroup_move_account fails) the failure is always temporary and
3816 * it signals a race with a page removal/uncharge or migration. In the
3817 * first case the page is on the way out and it will vanish from the LRU
3818 * on the next attempt and the call should be retried later.
3819 * Isolation from the LRU fails only if page has been isolated from
3820 * the LRU since we looked at it and that usually means either global
3821 * reclaim or migration going on. The page will either get back to the
3822 * LRU or vanish.
3823 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3824 * (!PageCgroupUsed) or moved to a different group. The page will
3825 * disappear in the next attempt.
f817ed48 3826 */
5564e88b
JW
3827static int mem_cgroup_move_parent(struct page *page,
3828 struct page_cgroup *pc,
6068bf01 3829 struct mem_cgroup *child)
f817ed48 3830{
f817ed48 3831 struct mem_cgroup *parent;
7ec99d62 3832 unsigned int nr_pages;
4be4489f 3833 unsigned long uninitialized_var(flags);
f817ed48
KH
3834 int ret;
3835
d8423011 3836 VM_BUG_ON(mem_cgroup_is_root(child));
f817ed48 3837
57f9fd7d
DN
3838 ret = -EBUSY;
3839 if (!get_page_unless_zero(page))
3840 goto out;
3841 if (isolate_lru_page(page))
3842 goto put;
52dbb905 3843
7ec99d62 3844 nr_pages = hpage_nr_pages(page);
08e552c6 3845
cc926f78
KH
3846 parent = parent_mem_cgroup(child);
3847 /*
3848 * If no parent, move charges to root cgroup.
3849 */
3850 if (!parent)
3851 parent = root_mem_cgroup;
f817ed48 3852
2ef37d3f
MH
3853 if (nr_pages > 1) {
3854 VM_BUG_ON(!PageTransHuge(page));
987eba66 3855 flags = compound_lock_irqsave(page);
2ef37d3f 3856 }
987eba66 3857
cc926f78 3858 ret = mem_cgroup_move_account(page, nr_pages,
2f3479b1 3859 pc, child, parent);
cc926f78
KH
3860 if (!ret)
3861 __mem_cgroup_cancel_local_charge(child, nr_pages);
8dba474f 3862
7ec99d62 3863 if (nr_pages > 1)
987eba66 3864 compound_unlock_irqrestore(page, flags);
08e552c6 3865 putback_lru_page(page);
57f9fd7d 3866put:
40d58138 3867 put_page(page);
57f9fd7d 3868out:
f817ed48
KH
3869 return ret;
3870}
3871
7a81b88c
KH
3872/*
3873 * Charge the memory controller for page usage.
3874 * Return
3875 * 0 if the charge was successful
3876 * < 0 if the cgroup is over its limit
3877 */
3878static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
73045c47 3879 gfp_t gfp_mask, enum charge_type ctype)
7a81b88c 3880{
c0ff4b85 3881 struct mem_cgroup *memcg = NULL;
7ec99d62 3882 unsigned int nr_pages = 1;
8493ae43 3883 bool oom = true;
7a81b88c 3884 int ret;
ec168510 3885
37c2ac78 3886 if (PageTransHuge(page)) {
7ec99d62 3887 nr_pages <<= compound_order(page);
37c2ac78 3888 VM_BUG_ON(!PageTransHuge(page));
8493ae43
JW
3889 /*
3890 * Never OOM-kill a process for a huge page. The
3891 * fault handler will fall back to regular pages.
3892 */
3893 oom = false;
37c2ac78 3894 }
7a81b88c 3895
c0ff4b85 3896 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
38c5d72f 3897 if (ret == -ENOMEM)
7a81b88c 3898 return ret;
ce587e65 3899 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
8a9f3ccd 3900 return 0;
8a9f3ccd
BS
3901}
3902
7a81b88c
KH
3903int mem_cgroup_newpage_charge(struct page *page,
3904 struct mm_struct *mm, gfp_t gfp_mask)
217bc319 3905{
f8d66542 3906 if (mem_cgroup_disabled())
cede86ac 3907 return 0;
7a0524cf
JW
3908 VM_BUG_ON(page_mapped(page));
3909 VM_BUG_ON(page->mapping && !PageAnon(page));
3910 VM_BUG_ON(!mm);
217bc319 3911 return mem_cgroup_charge_common(page, mm, gfp_mask,
41326c17 3912 MEM_CGROUP_CHARGE_TYPE_ANON);
217bc319
KH
3913}
3914
54595fe2
KH
3915/*
3916 * While swap-in, try_charge -> commit or cancel, the page is locked.
3917 * And when try_charge() successfully returns, one refcnt to memcg without
21ae2956 3918 * struct page_cgroup is acquired. This refcnt will be consumed by
54595fe2
KH
3919 * "commit()" or removed by "cancel()"
3920 */
0435a2fd
JW
3921static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3922 struct page *page,
3923 gfp_t mask,
3924 struct mem_cgroup **memcgp)
8c7c6e34 3925{
c0ff4b85 3926 struct mem_cgroup *memcg;
90deb788 3927 struct page_cgroup *pc;
54595fe2 3928 int ret;
8c7c6e34 3929
90deb788
JW
3930 pc = lookup_page_cgroup(page);
3931 /*
3932 * Every swap fault against a single page tries to charge the
3933 * page, bail as early as possible. shmem_unuse() encounters
3934 * already charged pages, too. The USED bit is protected by
3935 * the page lock, which serializes swap cache removal, which
3936 * in turn serializes uncharging.
3937 */
3938 if (PageCgroupUsed(pc))
3939 return 0;
8c7c6e34
KH
3940 if (!do_swap_account)
3941 goto charge_cur_mm;
c0ff4b85
R
3942 memcg = try_get_mem_cgroup_from_page(page);
3943 if (!memcg)
54595fe2 3944 goto charge_cur_mm;
72835c86
JW
3945 *memcgp = memcg;
3946 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
c0ff4b85 3947 css_put(&memcg->css);
38c5d72f
KH
3948 if (ret == -EINTR)
3949 ret = 0;
54595fe2 3950 return ret;
8c7c6e34 3951charge_cur_mm:
38c5d72f
KH
3952 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3953 if (ret == -EINTR)
3954 ret = 0;
3955 return ret;
8c7c6e34
KH
3956}
3957
0435a2fd
JW
3958int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3959 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3960{
3961 *memcgp = NULL;
3962 if (mem_cgroup_disabled())
3963 return 0;
bdf4f4d2
JW
3964 /*
3965 * A racing thread's fault, or swapoff, may have already
3966 * updated the pte, and even removed page from swap cache: in
3967 * those cases unuse_pte()'s pte_same() test will fail; but
3968 * there's also a KSM case which does need to charge the page.
3969 */
3970 if (!PageSwapCache(page)) {
3971 int ret;
3972
3973 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3974 if (ret == -EINTR)
3975 ret = 0;
3976 return ret;
3977 }
0435a2fd
JW
3978 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3979}
3980
827a03d2
JW
3981void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3982{
3983 if (mem_cgroup_disabled())
3984 return;
3985 if (!memcg)
3986 return;
3987 __mem_cgroup_cancel_charge(memcg, 1);
3988}
3989
83aae4c7 3990static void
72835c86 3991__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
83aae4c7 3992 enum charge_type ctype)
7a81b88c 3993{
f8d66542 3994 if (mem_cgroup_disabled())
7a81b88c 3995 return;
72835c86 3996 if (!memcg)
7a81b88c 3997 return;
5a6475a4 3998
ce587e65 3999 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
8c7c6e34
KH
4000 /*
4001 * Now swap is on-memory. This means this page may be
4002 * counted both as mem and swap....double count.
03f3c433
KH
4003 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4004 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4005 * may call delete_from_swap_cache() before reach here.
8c7c6e34 4006 */
03f3c433 4007 if (do_swap_account && PageSwapCache(page)) {
8c7c6e34 4008 swp_entry_t ent = {.val = page_private(page)};
86493009 4009 mem_cgroup_uncharge_swap(ent);
8c7c6e34 4010 }
7a81b88c
KH
4011}
4012
72835c86
JW
4013void mem_cgroup_commit_charge_swapin(struct page *page,
4014 struct mem_cgroup *memcg)
83aae4c7 4015{
72835c86 4016 __mem_cgroup_commit_charge_swapin(page, memcg,
41326c17 4017 MEM_CGROUP_CHARGE_TYPE_ANON);
83aae4c7
DN
4018}
4019
827a03d2
JW
4020int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4021 gfp_t gfp_mask)
7a81b88c 4022{
827a03d2
JW
4023 struct mem_cgroup *memcg = NULL;
4024 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4025 int ret;
4026
f8d66542 4027 if (mem_cgroup_disabled())
827a03d2
JW
4028 return 0;
4029 if (PageCompound(page))
4030 return 0;
4031
827a03d2
JW
4032 if (!PageSwapCache(page))
4033 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4034 else { /* page is swapcache/shmem */
0435a2fd
JW
4035 ret = __mem_cgroup_try_charge_swapin(mm, page,
4036 gfp_mask, &memcg);
827a03d2
JW
4037 if (!ret)
4038 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4039 }
4040 return ret;
7a81b88c
KH
4041}
4042
c0ff4b85 4043static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
7ec99d62
JW
4044 unsigned int nr_pages,
4045 const enum charge_type ctype)
569b846d
KH
4046{
4047 struct memcg_batch_info *batch = NULL;
4048 bool uncharge_memsw = true;
7ec99d62 4049
569b846d
KH
4050 /* If swapout, usage of swap doesn't decrease */
4051 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4052 uncharge_memsw = false;
569b846d
KH
4053
4054 batch = &current->memcg_batch;
4055 /*
4056 * In usual, we do css_get() when we remember memcg pointer.
4057 * But in this case, we keep res->usage until end of a series of
4058 * uncharges. Then, it's ok to ignore memcg's refcnt.
4059 */
4060 if (!batch->memcg)
c0ff4b85 4061 batch->memcg = memcg;
3c11ecf4
KH
4062 /*
4063 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
25985edc 4064 * In those cases, all pages freed continuously can be expected to be in
3c11ecf4
KH
4065 * the same cgroup and we have chance to coalesce uncharges.
4066 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4067 * because we want to do uncharge as soon as possible.
4068 */
4069
4070 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4071 goto direct_uncharge;
4072
7ec99d62 4073 if (nr_pages > 1)
ec168510
AA
4074 goto direct_uncharge;
4075
569b846d
KH
4076 /*
4077 * In typical case, batch->memcg == mem. This means we can
4078 * merge a series of uncharges to an uncharge of res_counter.
4079 * If not, we uncharge res_counter ony by one.
4080 */
c0ff4b85 4081 if (batch->memcg != memcg)
569b846d
KH
4082 goto direct_uncharge;
4083 /* remember freed charge and uncharge it later */
7ffd4ca7 4084 batch->nr_pages++;
569b846d 4085 if (uncharge_memsw)
7ffd4ca7 4086 batch->memsw_nr_pages++;
569b846d
KH
4087 return;
4088direct_uncharge:
c0ff4b85 4089 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
569b846d 4090 if (uncharge_memsw)
c0ff4b85
R
4091 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4092 if (unlikely(batch->memcg != memcg))
4093 memcg_oom_recover(memcg);
569b846d 4094}
7a81b88c 4095
8a9f3ccd 4096/*
69029cd5 4097 * uncharge if !page_mapped(page)
8a9f3ccd 4098 */
8c7c6e34 4099static struct mem_cgroup *
0030f535
JW
4100__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4101 bool end_migration)
8a9f3ccd 4102{
c0ff4b85 4103 struct mem_cgroup *memcg = NULL;
7ec99d62
JW
4104 unsigned int nr_pages = 1;
4105 struct page_cgroup *pc;
b2402857 4106 bool anon;
8a9f3ccd 4107
f8d66542 4108 if (mem_cgroup_disabled())
8c7c6e34 4109 return NULL;
4077960e 4110
0c59b89c 4111 VM_BUG_ON(PageSwapCache(page));
d13d1443 4112
37c2ac78 4113 if (PageTransHuge(page)) {
7ec99d62 4114 nr_pages <<= compound_order(page);
37c2ac78
AA
4115 VM_BUG_ON(!PageTransHuge(page));
4116 }
8697d331 4117 /*
3c541e14 4118 * Check if our page_cgroup is valid
8697d331 4119 */
52d4b9ac 4120 pc = lookup_page_cgroup(page);
cfa44946 4121 if (unlikely(!PageCgroupUsed(pc)))
8c7c6e34 4122 return NULL;
b9c565d5 4123
52d4b9ac 4124 lock_page_cgroup(pc);
d13d1443 4125
c0ff4b85 4126 memcg = pc->mem_cgroup;
8c7c6e34 4127
d13d1443
KH
4128 if (!PageCgroupUsed(pc))
4129 goto unlock_out;
4130
b2402857
KH
4131 anon = PageAnon(page);
4132
d13d1443 4133 switch (ctype) {
41326c17 4134 case MEM_CGROUP_CHARGE_TYPE_ANON:
2ff76f11
KH
4135 /*
4136 * Generally PageAnon tells if it's the anon statistics to be
4137 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4138 * used before page reached the stage of being marked PageAnon.
4139 */
b2402857
KH
4140 anon = true;
4141 /* fallthrough */
8a9478ca 4142 case MEM_CGROUP_CHARGE_TYPE_DROP:
ac39cf8c 4143 /* See mem_cgroup_prepare_migration() */
0030f535
JW
4144 if (page_mapped(page))
4145 goto unlock_out;
4146 /*
4147 * Pages under migration may not be uncharged. But
4148 * end_migration() /must/ be the one uncharging the
4149 * unused post-migration page and so it has to call
4150 * here with the migration bit still set. See the
4151 * res_counter handling below.
4152 */
4153 if (!end_migration && PageCgroupMigration(pc))
d13d1443
KH
4154 goto unlock_out;
4155 break;
4156 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4157 if (!PageAnon(page)) { /* Shared memory */
4158 if (page->mapping && !page_is_file_cache(page))
4159 goto unlock_out;
4160 } else if (page_mapped(page)) /* Anon */
4161 goto unlock_out;
4162 break;
4163 default:
4164 break;
52d4b9ac 4165 }
d13d1443 4166
b070e65c 4167 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
04046e1a 4168
52d4b9ac 4169 ClearPageCgroupUsed(pc);
544122e5
KH
4170 /*
4171 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4172 * freed from LRU. This is safe because uncharged page is expected not
4173 * to be reused (freed soon). Exception is SwapCache, it's handled by
4174 * special functions.
4175 */
b9c565d5 4176
52d4b9ac 4177 unlock_page_cgroup(pc);
f75ca962 4178 /*
c0ff4b85 4179 * even after unlock, we have memcg->res.usage here and this memcg
f75ca962
KH
4180 * will never be freed.
4181 */
c0ff4b85 4182 memcg_check_events(memcg, page);
f75ca962 4183 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
c0ff4b85
R
4184 mem_cgroup_swap_statistics(memcg, true);
4185 mem_cgroup_get(memcg);
f75ca962 4186 }
0030f535
JW
4187 /*
4188 * Migration does not charge the res_counter for the
4189 * replacement page, so leave it alone when phasing out the
4190 * page that is unused after the migration.
4191 */
4192 if (!end_migration && !mem_cgroup_is_root(memcg))
c0ff4b85 4193 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
6d12e2d8 4194
c0ff4b85 4195 return memcg;
d13d1443
KH
4196
4197unlock_out:
4198 unlock_page_cgroup(pc);
8c7c6e34 4199 return NULL;
3c541e14
BS
4200}
4201
69029cd5
KH
4202void mem_cgroup_uncharge_page(struct page *page)
4203{
52d4b9ac
KH
4204 /* early check. */
4205 if (page_mapped(page))
4206 return;
40f23a21 4207 VM_BUG_ON(page->mapping && !PageAnon(page));
0c59b89c
JW
4208 if (PageSwapCache(page))
4209 return;
0030f535 4210 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
69029cd5
KH
4211}
4212
4213void mem_cgroup_uncharge_cache_page(struct page *page)
4214{
4215 VM_BUG_ON(page_mapped(page));
b7abea96 4216 VM_BUG_ON(page->mapping);
0030f535 4217 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
69029cd5
KH
4218}
4219
569b846d
KH
4220/*
4221 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4222 * In that cases, pages are freed continuously and we can expect pages
4223 * are in the same memcg. All these calls itself limits the number of
4224 * pages freed at once, then uncharge_start/end() is called properly.
4225 * This may be called prural(2) times in a context,
4226 */
4227
4228void mem_cgroup_uncharge_start(void)
4229{
4230 current->memcg_batch.do_batch++;
4231 /* We can do nest. */
4232 if (current->memcg_batch.do_batch == 1) {
4233 current->memcg_batch.memcg = NULL;
7ffd4ca7
JW
4234 current->memcg_batch.nr_pages = 0;
4235 current->memcg_batch.memsw_nr_pages = 0;
569b846d
KH
4236 }
4237}
4238
4239void mem_cgroup_uncharge_end(void)
4240{
4241 struct memcg_batch_info *batch = &current->memcg_batch;
4242
4243 if (!batch->do_batch)
4244 return;
4245
4246 batch->do_batch--;
4247 if (batch->do_batch) /* If stacked, do nothing. */
4248 return;
4249
4250 if (!batch->memcg)
4251 return;
4252 /*
4253 * This "batch->memcg" is valid without any css_get/put etc...
4254 * bacause we hide charges behind us.
4255 */
7ffd4ca7
JW
4256 if (batch->nr_pages)
4257 res_counter_uncharge(&batch->memcg->res,
4258 batch->nr_pages * PAGE_SIZE);
4259 if (batch->memsw_nr_pages)
4260 res_counter_uncharge(&batch->memcg->memsw,
4261 batch->memsw_nr_pages * PAGE_SIZE);
3c11ecf4 4262 memcg_oom_recover(batch->memcg);
569b846d
KH
4263 /* forget this pointer (for sanity check) */
4264 batch->memcg = NULL;
4265}
4266
e767e056 4267#ifdef CONFIG_SWAP
8c7c6e34 4268/*
e767e056 4269 * called after __delete_from_swap_cache() and drop "page" account.
8c7c6e34
KH
4270 * memcg information is recorded to swap_cgroup of "ent"
4271 */
8a9478ca
KH
4272void
4273mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
8c7c6e34
KH
4274{
4275 struct mem_cgroup *memcg;
8a9478ca
KH
4276 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4277
4278 if (!swapout) /* this was a swap cache but the swap is unused ! */
4279 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4280
0030f535 4281 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
8c7c6e34 4282
f75ca962
KH
4283 /*
4284 * record memcg information, if swapout && memcg != NULL,
4285 * mem_cgroup_get() was called in uncharge().
4286 */
4287 if (do_swap_account && swapout && memcg)
a3b2d692 4288 swap_cgroup_record(ent, css_id(&memcg->css));
8c7c6e34 4289}
e767e056 4290#endif
8c7c6e34 4291
c255a458 4292#ifdef CONFIG_MEMCG_SWAP
8c7c6e34
KH
4293/*
4294 * called from swap_entry_free(). remove record in swap_cgroup and
4295 * uncharge "memsw" account.
4296 */
4297void mem_cgroup_uncharge_swap(swp_entry_t ent)
d13d1443 4298{
8c7c6e34 4299 struct mem_cgroup *memcg;
a3b2d692 4300 unsigned short id;
8c7c6e34
KH
4301
4302 if (!do_swap_account)
4303 return;
4304
a3b2d692
KH
4305 id = swap_cgroup_record(ent, 0);
4306 rcu_read_lock();
4307 memcg = mem_cgroup_lookup(id);
8c7c6e34 4308 if (memcg) {
a3b2d692
KH
4309 /*
4310 * We uncharge this because swap is freed.
4311 * This memcg can be obsolete one. We avoid calling css_tryget
4312 */
0c3e73e8 4313 if (!mem_cgroup_is_root(memcg))
4e649152 4314 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
0c3e73e8 4315 mem_cgroup_swap_statistics(memcg, false);
8c7c6e34
KH
4316 mem_cgroup_put(memcg);
4317 }
a3b2d692 4318 rcu_read_unlock();
d13d1443 4319}
02491447
DN
4320
4321/**
4322 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4323 * @entry: swap entry to be moved
4324 * @from: mem_cgroup which the entry is moved from
4325 * @to: mem_cgroup which the entry is moved to
4326 *
4327 * It succeeds only when the swap_cgroup's record for this entry is the same
4328 * as the mem_cgroup's id of @from.
4329 *
4330 * Returns 0 on success, -EINVAL on failure.
4331 *
4332 * The caller must have charged to @to, IOW, called res_counter_charge() about
4333 * both res and memsw, and called css_get().
4334 */
4335static int mem_cgroup_move_swap_account(swp_entry_t entry,
e91cbb42 4336 struct mem_cgroup *from, struct mem_cgroup *to)
02491447
DN
4337{
4338 unsigned short old_id, new_id;
4339
4340 old_id = css_id(&from->css);
4341 new_id = css_id(&to->css);
4342
4343 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
02491447 4344 mem_cgroup_swap_statistics(from, false);
483c30b5 4345 mem_cgroup_swap_statistics(to, true);
02491447 4346 /*
483c30b5
DN
4347 * This function is only called from task migration context now.
4348 * It postpones res_counter and refcount handling till the end
4349 * of task migration(mem_cgroup_clear_mc()) for performance
4350 * improvement. But we cannot postpone mem_cgroup_get(to)
4351 * because if the process that has been moved to @to does
4352 * swap-in, the refcount of @to might be decreased to 0.
02491447 4353 */
02491447 4354 mem_cgroup_get(to);
02491447
DN
4355 return 0;
4356 }
4357 return -EINVAL;
4358}
4359#else
4360static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
e91cbb42 4361 struct mem_cgroup *from, struct mem_cgroup *to)
02491447
DN
4362{
4363 return -EINVAL;
4364}
8c7c6e34 4365#endif
d13d1443 4366
ae41be37 4367/*
01b1ae63
KH
4368 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4369 * page belongs to.
ae41be37 4370 */
0030f535
JW
4371void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4372 struct mem_cgroup **memcgp)
ae41be37 4373{
c0ff4b85 4374 struct mem_cgroup *memcg = NULL;
b32967ff 4375 unsigned int nr_pages = 1;
7ec99d62 4376 struct page_cgroup *pc;
ac39cf8c 4377 enum charge_type ctype;
8869b8f6 4378
72835c86 4379 *memcgp = NULL;
56039efa 4380
f8d66542 4381 if (mem_cgroup_disabled())
0030f535 4382 return;
4077960e 4383
b32967ff
MG
4384 if (PageTransHuge(page))
4385 nr_pages <<= compound_order(page);
4386
52d4b9ac
KH
4387 pc = lookup_page_cgroup(page);
4388 lock_page_cgroup(pc);
4389 if (PageCgroupUsed(pc)) {
c0ff4b85
R
4390 memcg = pc->mem_cgroup;
4391 css_get(&memcg->css);
ac39cf8c
AM
4392 /*
4393 * At migrating an anonymous page, its mapcount goes down
4394 * to 0 and uncharge() will be called. But, even if it's fully
4395 * unmapped, migration may fail and this page has to be
4396 * charged again. We set MIGRATION flag here and delay uncharge
4397 * until end_migration() is called
4398 *
4399 * Corner Case Thinking
4400 * A)
4401 * When the old page was mapped as Anon and it's unmap-and-freed
4402 * while migration was ongoing.
4403 * If unmap finds the old page, uncharge() of it will be delayed
4404 * until end_migration(). If unmap finds a new page, it's
4405 * uncharged when it make mapcount to be 1->0. If unmap code
4406 * finds swap_migration_entry, the new page will not be mapped
4407 * and end_migration() will find it(mapcount==0).
4408 *
4409 * B)
4410 * When the old page was mapped but migraion fails, the kernel
4411 * remaps it. A charge for it is kept by MIGRATION flag even
4412 * if mapcount goes down to 0. We can do remap successfully
4413 * without charging it again.
4414 *
4415 * C)
4416 * The "old" page is under lock_page() until the end of
4417 * migration, so, the old page itself will not be swapped-out.
4418 * If the new page is swapped out before end_migraton, our
4419 * hook to usual swap-out path will catch the event.
4420 */
4421 if (PageAnon(page))
4422 SetPageCgroupMigration(pc);
e8589cc1 4423 }
52d4b9ac 4424 unlock_page_cgroup(pc);
ac39cf8c
AM
4425 /*
4426 * If the page is not charged at this point,
4427 * we return here.
4428 */
c0ff4b85 4429 if (!memcg)
0030f535 4430 return;
01b1ae63 4431
72835c86 4432 *memcgp = memcg;
ac39cf8c
AM
4433 /*
4434 * We charge new page before it's used/mapped. So, even if unlock_page()
4435 * is called before end_migration, we can catch all events on this new
4436 * page. In the case new page is migrated but not remapped, new page's
4437 * mapcount will be finally 0 and we call uncharge in end_migration().
4438 */
ac39cf8c 4439 if (PageAnon(page))
41326c17 4440 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
ac39cf8c 4441 else
62ba7442 4442 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
0030f535
JW
4443 /*
4444 * The page is committed to the memcg, but it's not actually
4445 * charged to the res_counter since we plan on replacing the
4446 * old one and only one page is going to be left afterwards.
4447 */
b32967ff 4448 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
ae41be37 4449}
8869b8f6 4450
69029cd5 4451/* remove redundant charge if migration failed*/
c0ff4b85 4452void mem_cgroup_end_migration(struct mem_cgroup *memcg,
50de1dd9 4453 struct page *oldpage, struct page *newpage, bool migration_ok)
ae41be37 4454{
ac39cf8c 4455 struct page *used, *unused;
01b1ae63 4456 struct page_cgroup *pc;
b2402857 4457 bool anon;
01b1ae63 4458
c0ff4b85 4459 if (!memcg)
01b1ae63 4460 return;
b25ed609 4461
50de1dd9 4462 if (!migration_ok) {
ac39cf8c
AM
4463 used = oldpage;
4464 unused = newpage;
01b1ae63 4465 } else {
ac39cf8c 4466 used = newpage;
01b1ae63
KH
4467 unused = oldpage;
4468 }
0030f535 4469 anon = PageAnon(used);
7d188958
JW
4470 __mem_cgroup_uncharge_common(unused,
4471 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4472 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4473 true);
0030f535 4474 css_put(&memcg->css);
69029cd5 4475 /*
ac39cf8c
AM
4476 * We disallowed uncharge of pages under migration because mapcount
4477 * of the page goes down to zero, temporarly.
4478 * Clear the flag and check the page should be charged.
01b1ae63 4479 */
ac39cf8c
AM
4480 pc = lookup_page_cgroup(oldpage);
4481 lock_page_cgroup(pc);
4482 ClearPageCgroupMigration(pc);
4483 unlock_page_cgroup(pc);
ac39cf8c 4484
01b1ae63 4485 /*
ac39cf8c
AM
4486 * If a page is a file cache, radix-tree replacement is very atomic
4487 * and we can skip this check. When it was an Anon page, its mapcount
4488 * goes down to 0. But because we added MIGRATION flage, it's not
4489 * uncharged yet. There are several case but page->mapcount check
4490 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4491 * check. (see prepare_charge() also)
69029cd5 4492 */
b2402857 4493 if (anon)
ac39cf8c 4494 mem_cgroup_uncharge_page(used);
ae41be37 4495}
78fb7466 4496
ab936cbc
KH
4497/*
4498 * At replace page cache, newpage is not under any memcg but it's on
4499 * LRU. So, this function doesn't touch res_counter but handles LRU
4500 * in correct way. Both pages are locked so we cannot race with uncharge.
4501 */
4502void mem_cgroup_replace_page_cache(struct page *oldpage,
4503 struct page *newpage)
4504{
bde05d1c 4505 struct mem_cgroup *memcg = NULL;
ab936cbc 4506 struct page_cgroup *pc;
ab936cbc 4507 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
ab936cbc
KH
4508
4509 if (mem_cgroup_disabled())
4510 return;
4511
4512 pc = lookup_page_cgroup(oldpage);
4513 /* fix accounting on old pages */
4514 lock_page_cgroup(pc);
bde05d1c
HD
4515 if (PageCgroupUsed(pc)) {
4516 memcg = pc->mem_cgroup;
b070e65c 4517 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
bde05d1c
HD
4518 ClearPageCgroupUsed(pc);
4519 }
ab936cbc
KH
4520 unlock_page_cgroup(pc);
4521
bde05d1c
HD
4522 /*
4523 * When called from shmem_replace_page(), in some cases the
4524 * oldpage has already been charged, and in some cases not.
4525 */
4526 if (!memcg)
4527 return;
ab936cbc
KH
4528 /*
4529 * Even if newpage->mapping was NULL before starting replacement,
4530 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4531 * LRU while we overwrite pc->mem_cgroup.
4532 */
ce587e65 4533 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
ab936cbc
KH
4534}
4535
f212ad7c
DN
4536#ifdef CONFIG_DEBUG_VM
4537static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4538{
4539 struct page_cgroup *pc;
4540
4541 pc = lookup_page_cgroup(page);
cfa44946
JW
4542 /*
4543 * Can be NULL while feeding pages into the page allocator for
4544 * the first time, i.e. during boot or memory hotplug;
4545 * or when mem_cgroup_disabled().
4546 */
f212ad7c
DN
4547 if (likely(pc) && PageCgroupUsed(pc))
4548 return pc;
4549 return NULL;
4550}
4551
4552bool mem_cgroup_bad_page_check(struct page *page)
4553{
4554 if (mem_cgroup_disabled())
4555 return false;
4556
4557 return lookup_page_cgroup_used(page) != NULL;
4558}
4559
4560void mem_cgroup_print_bad_page(struct page *page)
4561{
4562 struct page_cgroup *pc;
4563
4564 pc = lookup_page_cgroup_used(page);
4565 if (pc) {
d045197f
AM
4566 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4567 pc, pc->flags, pc->mem_cgroup);
f212ad7c
DN
4568 }
4569}
4570#endif
4571
d38d2a75 4572static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
8c7c6e34 4573 unsigned long long val)
628f4235 4574{
81d39c20 4575 int retry_count;
3c11ecf4 4576 u64 memswlimit, memlimit;
628f4235 4577 int ret = 0;
81d39c20
KH
4578 int children = mem_cgroup_count_children(memcg);
4579 u64 curusage, oldusage;
3c11ecf4 4580 int enlarge;
81d39c20
KH
4581
4582 /*
4583 * For keeping hierarchical_reclaim simple, how long we should retry
4584 * is depends on callers. We set our retry-count to be function
4585 * of # of children which we should visit in this loop.
4586 */
4587 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4588
4589 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
628f4235 4590
3c11ecf4 4591 enlarge = 0;
8c7c6e34 4592 while (retry_count) {
628f4235
KH
4593 if (signal_pending(current)) {
4594 ret = -EINTR;
4595 break;
4596 }
8c7c6e34
KH
4597 /*
4598 * Rather than hide all in some function, I do this in
4599 * open coded manner. You see what this really does.
aaad153e 4600 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
8c7c6e34
KH
4601 */
4602 mutex_lock(&set_limit_mutex);
4603 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4604 if (memswlimit < val) {
4605 ret = -EINVAL;
4606 mutex_unlock(&set_limit_mutex);
628f4235
KH
4607 break;
4608 }
3c11ecf4
KH
4609
4610 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4611 if (memlimit < val)
4612 enlarge = 1;
4613
8c7c6e34 4614 ret = res_counter_set_limit(&memcg->res, val);
22a668d7
KH
4615 if (!ret) {
4616 if (memswlimit == val)
4617 memcg->memsw_is_minimum = true;
4618 else
4619 memcg->memsw_is_minimum = false;
4620 }
8c7c6e34
KH
4621 mutex_unlock(&set_limit_mutex);
4622
4623 if (!ret)
4624 break;
4625
5660048c
JW
4626 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4627 MEM_CGROUP_RECLAIM_SHRINK);
81d39c20
KH
4628 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4629 /* Usage is reduced ? */
4630 if (curusage >= oldusage)
4631 retry_count--;
4632 else
4633 oldusage = curusage;
8c7c6e34 4634 }
3c11ecf4
KH
4635 if (!ret && enlarge)
4636 memcg_oom_recover(memcg);
14797e23 4637
8c7c6e34
KH
4638 return ret;
4639}
4640
338c8431
LZ
4641static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4642 unsigned long long val)
8c7c6e34 4643{
81d39c20 4644 int retry_count;
3c11ecf4 4645 u64 memlimit, memswlimit, oldusage, curusage;
81d39c20
KH
4646 int children = mem_cgroup_count_children(memcg);
4647 int ret = -EBUSY;
3c11ecf4 4648 int enlarge = 0;
8c7c6e34 4649
81d39c20
KH
4650 /* see mem_cgroup_resize_res_limit */
4651 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4652 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
8c7c6e34
KH
4653 while (retry_count) {
4654 if (signal_pending(current)) {
4655 ret = -EINTR;
4656 break;
4657 }
4658 /*
4659 * Rather than hide all in some function, I do this in
4660 * open coded manner. You see what this really does.
aaad153e 4661 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
8c7c6e34
KH
4662 */
4663 mutex_lock(&set_limit_mutex);
4664 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4665 if (memlimit > val) {
4666 ret = -EINVAL;
4667 mutex_unlock(&set_limit_mutex);
4668 break;
4669 }
3c11ecf4
KH
4670 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4671 if (memswlimit < val)
4672 enlarge = 1;
8c7c6e34 4673 ret = res_counter_set_limit(&memcg->memsw, val);
22a668d7
KH
4674 if (!ret) {
4675 if (memlimit == val)
4676 memcg->memsw_is_minimum = true;
4677 else
4678 memcg->memsw_is_minimum = false;
4679 }
8c7c6e34
KH
4680 mutex_unlock(&set_limit_mutex);
4681
4682 if (!ret)
4683 break;
4684
5660048c
JW
4685 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4686 MEM_CGROUP_RECLAIM_NOSWAP |
4687 MEM_CGROUP_RECLAIM_SHRINK);
8c7c6e34 4688 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
81d39c20 4689 /* Usage is reduced ? */
8c7c6e34 4690 if (curusage >= oldusage)
628f4235 4691 retry_count--;
81d39c20
KH
4692 else
4693 oldusage = curusage;
628f4235 4694 }
3c11ecf4
KH
4695 if (!ret && enlarge)
4696 memcg_oom_recover(memcg);
628f4235
KH
4697 return ret;
4698}
4699
4e416953 4700unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
0ae5e89c
YH
4701 gfp_t gfp_mask,
4702 unsigned long *total_scanned)
4e416953
BS
4703{
4704 unsigned long nr_reclaimed = 0;
4705 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4706 unsigned long reclaimed;
4707 int loop = 0;
4708 struct mem_cgroup_tree_per_zone *mctz;
ef8745c1 4709 unsigned long long excess;
0ae5e89c 4710 unsigned long nr_scanned;
4e416953
BS
4711
4712 if (order > 0)
4713 return 0;
4714
00918b6a 4715 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4e416953
BS
4716 /*
4717 * This loop can run a while, specially if mem_cgroup's continuously
4718 * keep exceeding their soft limit and putting the system under
4719 * pressure
4720 */
4721 do {
4722 if (next_mz)
4723 mz = next_mz;
4724 else
4725 mz = mem_cgroup_largest_soft_limit_node(mctz);
4726 if (!mz)
4727 break;
4728
0ae5e89c 4729 nr_scanned = 0;
d79154bb 4730 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
5660048c 4731 gfp_mask, &nr_scanned);
4e416953 4732 nr_reclaimed += reclaimed;
0ae5e89c 4733 *total_scanned += nr_scanned;
4e416953
BS
4734 spin_lock(&mctz->lock);
4735
4736 /*
4737 * If we failed to reclaim anything from this memory cgroup
4738 * it is time to move on to the next cgroup
4739 */
4740 next_mz = NULL;
4741 if (!reclaimed) {
4742 do {
4743 /*
4744 * Loop until we find yet another one.
4745 *
4746 * By the time we get the soft_limit lock
4747 * again, someone might have aded the
4748 * group back on the RB tree. Iterate to
4749 * make sure we get a different mem.
4750 * mem_cgroup_largest_soft_limit_node returns
4751 * NULL if no other cgroup is present on
4752 * the tree
4753 */
4754 next_mz =
4755 __mem_cgroup_largest_soft_limit_node(mctz);
39cc98f1 4756 if (next_mz == mz)
d79154bb 4757 css_put(&next_mz->memcg->css);
39cc98f1 4758 else /* next_mz == NULL or other memcg */
4e416953
BS
4759 break;
4760 } while (1);
4761 }
d79154bb
HD
4762 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4763 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4e416953
BS
4764 /*
4765 * One school of thought says that we should not add
4766 * back the node to the tree if reclaim returns 0.
4767 * But our reclaim could return 0, simply because due
4768 * to priority we are exposing a smaller subset of
4769 * memory to reclaim from. Consider this as a longer
4770 * term TODO.
4771 */
ef8745c1 4772 /* If excess == 0, no tree ops */
d79154bb 4773 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4e416953 4774 spin_unlock(&mctz->lock);
d79154bb 4775 css_put(&mz->memcg->css);
4e416953
BS
4776 loop++;
4777 /*
4778 * Could not reclaim anything and there are no more
4779 * mem cgroups to try or we seem to be looping without
4780 * reclaiming anything.
4781 */
4782 if (!nr_reclaimed &&
4783 (next_mz == NULL ||
4784 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4785 break;
4786 } while (!nr_reclaimed);
4787 if (next_mz)
d79154bb 4788 css_put(&next_mz->memcg->css);
4e416953
BS
4789 return nr_reclaimed;
4790}
4791
2ef37d3f
MH
4792/**
4793 * mem_cgroup_force_empty_list - clears LRU of a group
4794 * @memcg: group to clear
4795 * @node: NUMA node
4796 * @zid: zone id
4797 * @lru: lru to to clear
4798 *
3c935d18 4799 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
2ef37d3f
MH
4800 * reclaim the pages page themselves - pages are moved to the parent (or root)
4801 * group.
cc847582 4802 */
2ef37d3f 4803static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
08e552c6 4804 int node, int zid, enum lru_list lru)
cc847582 4805{
bea8c150 4806 struct lruvec *lruvec;
2ef37d3f 4807 unsigned long flags;
072c56c1 4808 struct list_head *list;
925b7673
JW
4809 struct page *busy;
4810 struct zone *zone;
072c56c1 4811
08e552c6 4812 zone = &NODE_DATA(node)->node_zones[zid];
bea8c150
HD
4813 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4814 list = &lruvec->lists[lru];
cc847582 4815
f817ed48 4816 busy = NULL;
2ef37d3f 4817 do {
925b7673 4818 struct page_cgroup *pc;
5564e88b
JW
4819 struct page *page;
4820
08e552c6 4821 spin_lock_irqsave(&zone->lru_lock, flags);
f817ed48 4822 if (list_empty(list)) {
08e552c6 4823 spin_unlock_irqrestore(&zone->lru_lock, flags);
52d4b9ac 4824 break;
f817ed48 4825 }
925b7673
JW
4826 page = list_entry(list->prev, struct page, lru);
4827 if (busy == page) {
4828 list_move(&page->lru, list);
648bcc77 4829 busy = NULL;
08e552c6 4830 spin_unlock_irqrestore(&zone->lru_lock, flags);
f817ed48
KH
4831 continue;
4832 }
08e552c6 4833 spin_unlock_irqrestore(&zone->lru_lock, flags);
f817ed48 4834
925b7673 4835 pc = lookup_page_cgroup(page);
5564e88b 4836
3c935d18 4837 if (mem_cgroup_move_parent(page, pc, memcg)) {
f817ed48 4838 /* found lock contention or "pc" is obsolete. */
925b7673 4839 busy = page;
f817ed48
KH
4840 cond_resched();
4841 } else
4842 busy = NULL;
2ef37d3f 4843 } while (!list_empty(list));
cc847582
KH
4844}
4845
4846/*
c26251f9
MH
4847 * make mem_cgroup's charge to be 0 if there is no task by moving
4848 * all the charges and pages to the parent.
cc847582 4849 * This enables deleting this mem_cgroup.
c26251f9
MH
4850 *
4851 * Caller is responsible for holding css reference on the memcg.
cc847582 4852 */
ab5196c2 4853static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
cc847582 4854{
c26251f9 4855 int node, zid;
bea207c8 4856 u64 usage;
f817ed48 4857
fce66477 4858 do {
52d4b9ac
KH
4859 /* This is for making all *used* pages to be on LRU. */
4860 lru_add_drain_all();
c0ff4b85 4861 drain_all_stock_sync(memcg);
c0ff4b85 4862 mem_cgroup_start_move(memcg);
31aaea4a 4863 for_each_node_state(node, N_MEMORY) {
2ef37d3f 4864 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
f156ab93
HD
4865 enum lru_list lru;
4866 for_each_lru(lru) {
2ef37d3f 4867 mem_cgroup_force_empty_list(memcg,
f156ab93 4868 node, zid, lru);
f817ed48 4869 }
1ecaab2b 4870 }
f817ed48 4871 }
c0ff4b85
R
4872 mem_cgroup_end_move(memcg);
4873 memcg_oom_recover(memcg);
52d4b9ac 4874 cond_resched();
f817ed48 4875
2ef37d3f 4876 /*
bea207c8
GC
4877 * Kernel memory may not necessarily be trackable to a specific
4878 * process. So they are not migrated, and therefore we can't
4879 * expect their value to drop to 0 here.
4880 * Having res filled up with kmem only is enough.
4881 *
2ef37d3f
MH
4882 * This is a safety check because mem_cgroup_force_empty_list
4883 * could have raced with mem_cgroup_replace_page_cache callers
4884 * so the lru seemed empty but the page could have been added
4885 * right after the check. RES_USAGE should be safe as we always
4886 * charge before adding to the LRU.
4887 */
bea207c8
GC
4888 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4889 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4890 } while (usage > 0);
c26251f9
MH
4891}
4892
b5f99b53
GC
4893/*
4894 * This mainly exists for tests during the setting of set of use_hierarchy.
4895 * Since this is the very setting we are changing, the current hierarchy value
4896 * is meaningless
4897 */
4898static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4899{
4900 struct cgroup *pos;
4901
4902 /* bounce at first found */
4903 cgroup_for_each_child(pos, memcg->css.cgroup)
4904 return true;
4905 return false;
4906}
4907
4908/*
0999821b
GC
4909 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4910 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
b5f99b53
GC
4911 * from mem_cgroup_count_children(), in the sense that we don't really care how
4912 * many children we have; we only need to know if we have any. It also counts
4913 * any memcg without hierarchy as infertile.
4914 */
4915static inline bool memcg_has_children(struct mem_cgroup *memcg)
4916{
4917 return memcg->use_hierarchy && __memcg_has_children(memcg);
4918}
4919
c26251f9
MH
4920/*
4921 * Reclaims as many pages from the given memcg as possible and moves
4922 * the rest to the parent.
4923 *
4924 * Caller is responsible for holding css reference for memcg.
4925 */
4926static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4927{
4928 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4929 struct cgroup *cgrp = memcg->css.cgroup;
f817ed48 4930
c1e862c1 4931 /* returns EBUSY if there is a task or if we come here twice. */
c26251f9
MH
4932 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4933 return -EBUSY;
4934
c1e862c1
KH
4935 /* we call try-to-free pages for make this cgroup empty */
4936 lru_add_drain_all();
f817ed48 4937 /* try to free all pages in this cgroup */
569530fb 4938 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
f817ed48 4939 int progress;
c1e862c1 4940
c26251f9
MH
4941 if (signal_pending(current))
4942 return -EINTR;
4943
c0ff4b85 4944 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
185efc0f 4945 false);
c1e862c1 4946 if (!progress) {
f817ed48 4947 nr_retries--;
c1e862c1 4948 /* maybe some writeback is necessary */
8aa7e847 4949 congestion_wait(BLK_RW_ASYNC, HZ/10);
c1e862c1 4950 }
f817ed48
KH
4951
4952 }
08e552c6 4953 lru_add_drain();
ab5196c2
MH
4954 mem_cgroup_reparent_charges(memcg);
4955
4956 return 0;
cc847582
KH
4957}
4958
6bbda35c 4959static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
c1e862c1 4960{
c26251f9
MH
4961 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4962 int ret;
4963
d8423011
MH
4964 if (mem_cgroup_is_root(memcg))
4965 return -EINVAL;
c26251f9
MH
4966 css_get(&memcg->css);
4967 ret = mem_cgroup_force_empty(memcg);
4968 css_put(&memcg->css);
4969
4970 return ret;
c1e862c1
KH
4971}
4972
4973
18f59ea7
BS
4974static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4975{
4976 return mem_cgroup_from_cont(cont)->use_hierarchy;
4977}
4978
4979static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4980 u64 val)
4981{
4982 int retval = 0;
c0ff4b85 4983 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
18f59ea7 4984 struct cgroup *parent = cont->parent;
c0ff4b85 4985 struct mem_cgroup *parent_memcg = NULL;
18f59ea7
BS
4986
4987 if (parent)
c0ff4b85 4988 parent_memcg = mem_cgroup_from_cont(parent);
18f59ea7 4989
0999821b 4990 mutex_lock(&memcg_create_mutex);
567fb435
GC
4991
4992 if (memcg->use_hierarchy == val)
4993 goto out;
4994
18f59ea7 4995 /*
af901ca1 4996 * If parent's use_hierarchy is set, we can't make any modifications
18f59ea7
BS
4997 * in the child subtrees. If it is unset, then the change can
4998 * occur, provided the current cgroup has no children.
4999 *
5000 * For the root cgroup, parent_mem is NULL, we allow value to be
5001 * set if there are no children.
5002 */
c0ff4b85 5003 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
18f59ea7 5004 (val == 1 || val == 0)) {
b5f99b53 5005 if (!__memcg_has_children(memcg))
c0ff4b85 5006 memcg->use_hierarchy = val;
18f59ea7
BS
5007 else
5008 retval = -EBUSY;
5009 } else
5010 retval = -EINVAL;
567fb435
GC
5011
5012out:
0999821b 5013 mutex_unlock(&memcg_create_mutex);
18f59ea7
BS
5014
5015 return retval;
5016}
5017
0c3e73e8 5018
c0ff4b85 5019static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
7a159cc9 5020 enum mem_cgroup_stat_index idx)
0c3e73e8 5021{
7d74b06f 5022 struct mem_cgroup *iter;
7a159cc9 5023 long val = 0;
0c3e73e8 5024
7a159cc9 5025 /* Per-cpu values can be negative, use a signed accumulator */
c0ff4b85 5026 for_each_mem_cgroup_tree(iter, memcg)
7d74b06f
KH
5027 val += mem_cgroup_read_stat(iter, idx);
5028
5029 if (val < 0) /* race ? */
5030 val = 0;
5031 return val;
0c3e73e8
BS
5032}
5033
c0ff4b85 5034static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
104f3928 5035{
7d74b06f 5036 u64 val;
104f3928 5037
c0ff4b85 5038 if (!mem_cgroup_is_root(memcg)) {
104f3928 5039 if (!swap)
65c64ce8 5040 return res_counter_read_u64(&memcg->res, RES_USAGE);
104f3928 5041 else
65c64ce8 5042 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
104f3928
KS
5043 }
5044
b070e65c
DR
5045 /*
5046 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5047 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5048 */
c0ff4b85
R
5049 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5050 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
104f3928 5051
7d74b06f 5052 if (swap)
bff6bb83 5053 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
104f3928
KS
5054
5055 return val << PAGE_SHIFT;
5056}
5057
af36f906
TH
5058static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5059 struct file *file, char __user *buf,
5060 size_t nbytes, loff_t *ppos)
8cdea7c0 5061{
c0ff4b85 5062 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
af36f906 5063 char str[64];
104f3928 5064 u64 val;
86ae53e1
GC
5065 int name, len;
5066 enum res_type type;
8c7c6e34
KH
5067
5068 type = MEMFILE_TYPE(cft->private);
5069 name = MEMFILE_ATTR(cft->private);
af36f906 5070
8c7c6e34
KH
5071 switch (type) {
5072 case _MEM:
104f3928 5073 if (name == RES_USAGE)
c0ff4b85 5074 val = mem_cgroup_usage(memcg, false);
104f3928 5075 else
c0ff4b85 5076 val = res_counter_read_u64(&memcg->res, name);
8c7c6e34
KH
5077 break;
5078 case _MEMSWAP:
104f3928 5079 if (name == RES_USAGE)
c0ff4b85 5080 val = mem_cgroup_usage(memcg, true);
104f3928 5081 else
c0ff4b85 5082 val = res_counter_read_u64(&memcg->memsw, name);
8c7c6e34 5083 break;
510fc4e1
GC
5084 case _KMEM:
5085 val = res_counter_read_u64(&memcg->kmem, name);
5086 break;
8c7c6e34
KH
5087 default:
5088 BUG();
8c7c6e34 5089 }
af36f906
TH
5090
5091 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5092 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
8cdea7c0 5093}
510fc4e1
GC
5094
5095static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5096{
5097 int ret = -EINVAL;
5098#ifdef CONFIG_MEMCG_KMEM
5099 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5100 /*
5101 * For simplicity, we won't allow this to be disabled. It also can't
5102 * be changed if the cgroup has children already, or if tasks had
5103 * already joined.
5104 *
5105 * If tasks join before we set the limit, a person looking at
5106 * kmem.usage_in_bytes will have no way to determine when it took
5107 * place, which makes the value quite meaningless.
5108 *
5109 * After it first became limited, changes in the value of the limit are
5110 * of course permitted.
510fc4e1 5111 */
0999821b 5112 mutex_lock(&memcg_create_mutex);
510fc4e1
GC
5113 mutex_lock(&set_limit_mutex);
5114 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
b5f99b53 5115 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
510fc4e1
GC
5116 ret = -EBUSY;
5117 goto out;
5118 }
5119 ret = res_counter_set_limit(&memcg->kmem, val);
5120 VM_BUG_ON(ret);
5121
55007d84
GC
5122 ret = memcg_update_cache_sizes(memcg);
5123 if (ret) {
5124 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5125 goto out;
5126 }
692e89ab
GC
5127 static_key_slow_inc(&memcg_kmem_enabled_key);
5128 /*
5129 * setting the active bit after the inc will guarantee no one
5130 * starts accounting before all call sites are patched
5131 */
5132 memcg_kmem_set_active(memcg);
5133
7de37682
GC
5134 /*
5135 * kmem charges can outlive the cgroup. In the case of slab
5136 * pages, for instance, a page contain objects from various
5137 * processes, so it is unfeasible to migrate them away. We
5138 * need to reference count the memcg because of that.
5139 */
5140 mem_cgroup_get(memcg);
510fc4e1
GC
5141 } else
5142 ret = res_counter_set_limit(&memcg->kmem, val);
5143out:
5144 mutex_unlock(&set_limit_mutex);
0999821b 5145 mutex_unlock(&memcg_create_mutex);
510fc4e1
GC
5146#endif
5147 return ret;
5148}
5149
6d043990 5150#ifdef CONFIG_MEMCG_KMEM
55007d84 5151static int memcg_propagate_kmem(struct mem_cgroup *memcg)
510fc4e1 5152{
55007d84 5153 int ret = 0;
510fc4e1
GC
5154 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5155 if (!parent)
55007d84
GC
5156 goto out;
5157
510fc4e1 5158 memcg->kmem_account_flags = parent->kmem_account_flags;
a8964b9b
GC
5159 /*
5160 * When that happen, we need to disable the static branch only on those
5161 * memcgs that enabled it. To achieve this, we would be forced to
5162 * complicate the code by keeping track of which memcgs were the ones
5163 * that actually enabled limits, and which ones got it from its
5164 * parents.
5165 *
5166 * It is a lot simpler just to do static_key_slow_inc() on every child
5167 * that is accounted.
5168 */
55007d84
GC
5169 if (!memcg_kmem_is_active(memcg))
5170 goto out;
5171
5172 /*
5173 * destroy(), called if we fail, will issue static_key_slow_inc() and
5174 * mem_cgroup_put() if kmem is enabled. We have to either call them
5175 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5176 * this more consistent, since it always leads to the same destroy path
5177 */
5178 mem_cgroup_get(memcg);
5179 static_key_slow_inc(&memcg_kmem_enabled_key);
5180
5181 mutex_lock(&set_limit_mutex);
5182 ret = memcg_update_cache_sizes(memcg);
5183 mutex_unlock(&set_limit_mutex);
55007d84
GC
5184out:
5185 return ret;
510fc4e1 5186}
6d043990 5187#endif /* CONFIG_MEMCG_KMEM */
510fc4e1 5188
628f4235
KH
5189/*
5190 * The user of this function is...
5191 * RES_LIMIT.
5192 */
856c13aa
PM
5193static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5194 const char *buffer)
8cdea7c0 5195{
628f4235 5196 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
86ae53e1
GC
5197 enum res_type type;
5198 int name;
628f4235
KH
5199 unsigned long long val;
5200 int ret;
5201
8c7c6e34
KH
5202 type = MEMFILE_TYPE(cft->private);
5203 name = MEMFILE_ATTR(cft->private);
af36f906 5204
8c7c6e34 5205 switch (name) {
628f4235 5206 case RES_LIMIT:
4b3bde4c
BS
5207 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5208 ret = -EINVAL;
5209 break;
5210 }
628f4235
KH
5211 /* This function does all necessary parse...reuse it */
5212 ret = res_counter_memparse_write_strategy(buffer, &val);
8c7c6e34
KH
5213 if (ret)
5214 break;
5215 if (type == _MEM)
628f4235 5216 ret = mem_cgroup_resize_limit(memcg, val);
510fc4e1 5217 else if (type == _MEMSWAP)
8c7c6e34 5218 ret = mem_cgroup_resize_memsw_limit(memcg, val);
510fc4e1
GC
5219 else if (type == _KMEM)
5220 ret = memcg_update_kmem_limit(cont, val);
5221 else
5222 return -EINVAL;
628f4235 5223 break;
296c81d8
BS
5224 case RES_SOFT_LIMIT:
5225 ret = res_counter_memparse_write_strategy(buffer, &val);
5226 if (ret)
5227 break;
5228 /*
5229 * For memsw, soft limits are hard to implement in terms
5230 * of semantics, for now, we support soft limits for
5231 * control without swap
5232 */
5233 if (type == _MEM)
5234 ret = res_counter_set_soft_limit(&memcg->res, val);
5235 else
5236 ret = -EINVAL;
5237 break;
628f4235
KH
5238 default:
5239 ret = -EINVAL; /* should be BUG() ? */
5240 break;
5241 }
5242 return ret;
8cdea7c0
BS
5243}
5244
fee7b548
KH
5245static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5246 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5247{
5248 struct cgroup *cgroup;
5249 unsigned long long min_limit, min_memsw_limit, tmp;
5250
5251 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5252 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5253 cgroup = memcg->css.cgroup;
5254 if (!memcg->use_hierarchy)
5255 goto out;
5256
5257 while (cgroup->parent) {
5258 cgroup = cgroup->parent;
5259 memcg = mem_cgroup_from_cont(cgroup);
5260 if (!memcg->use_hierarchy)
5261 break;
5262 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5263 min_limit = min(min_limit, tmp);
5264 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5265 min_memsw_limit = min(min_memsw_limit, tmp);
5266 }
5267out:
5268 *mem_limit = min_limit;
5269 *memsw_limit = min_memsw_limit;
fee7b548
KH
5270}
5271
29f2a4da 5272static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
c84872e1 5273{
af36f906 5274 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
86ae53e1
GC
5275 int name;
5276 enum res_type type;
c84872e1 5277
8c7c6e34
KH
5278 type = MEMFILE_TYPE(event);
5279 name = MEMFILE_ATTR(event);
af36f906 5280
8c7c6e34 5281 switch (name) {
29f2a4da 5282 case RES_MAX_USAGE:
8c7c6e34 5283 if (type == _MEM)
c0ff4b85 5284 res_counter_reset_max(&memcg->res);
510fc4e1 5285 else if (type == _MEMSWAP)
c0ff4b85 5286 res_counter_reset_max(&memcg->memsw);
510fc4e1
GC
5287 else if (type == _KMEM)
5288 res_counter_reset_max(&memcg->kmem);
5289 else
5290 return -EINVAL;
29f2a4da
PE
5291 break;
5292 case RES_FAILCNT:
8c7c6e34 5293 if (type == _MEM)
c0ff4b85 5294 res_counter_reset_failcnt(&memcg->res);
510fc4e1 5295 else if (type == _MEMSWAP)
c0ff4b85 5296 res_counter_reset_failcnt(&memcg->memsw);
510fc4e1
GC
5297 else if (type == _KMEM)
5298 res_counter_reset_failcnt(&memcg->kmem);
5299 else
5300 return -EINVAL;
29f2a4da
PE
5301 break;
5302 }
f64c3f54 5303
85cc59db 5304 return 0;
c84872e1
PE
5305}
5306
7dc74be0
DN
5307static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5308 struct cftype *cft)
5309{
5310 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5311}
5312
02491447 5313#ifdef CONFIG_MMU
7dc74be0
DN
5314static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5315 struct cftype *cft, u64 val)
5316{
c0ff4b85 5317 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
7dc74be0
DN
5318
5319 if (val >= (1 << NR_MOVE_TYPE))
5320 return -EINVAL;
ee5e8472 5321
7dc74be0 5322 /*
ee5e8472
GC
5323 * No kind of locking is needed in here, because ->can_attach() will
5324 * check this value once in the beginning of the process, and then carry
5325 * on with stale data. This means that changes to this value will only
5326 * affect task migrations starting after the change.
7dc74be0 5327 */
c0ff4b85 5328 memcg->move_charge_at_immigrate = val;
7dc74be0
DN
5329 return 0;
5330}
02491447
DN
5331#else
5332static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5333 struct cftype *cft, u64 val)
5334{
5335 return -ENOSYS;
5336}
5337#endif
7dc74be0 5338
406eb0c9 5339#ifdef CONFIG_NUMA
ab215884 5340static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
fada52ca 5341 struct seq_file *m)
406eb0c9
YH
5342{
5343 int nid;
5344 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5345 unsigned long node_nr;
d79154bb 5346 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
406eb0c9 5347
d79154bb 5348 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
406eb0c9 5349 seq_printf(m, "total=%lu", total_nr);
31aaea4a 5350 for_each_node_state(nid, N_MEMORY) {
d79154bb 5351 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
406eb0c9
YH
5352 seq_printf(m, " N%d=%lu", nid, node_nr);
5353 }
5354 seq_putc(m, '\n');
5355
d79154bb 5356 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
406eb0c9 5357 seq_printf(m, "file=%lu", file_nr);
31aaea4a 5358 for_each_node_state(nid, N_MEMORY) {
d79154bb 5359 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
bb2a0de9 5360 LRU_ALL_FILE);
406eb0c9
YH
5361 seq_printf(m, " N%d=%lu", nid, node_nr);
5362 }
5363 seq_putc(m, '\n');
5364
d79154bb 5365 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
406eb0c9 5366 seq_printf(m, "anon=%lu", anon_nr);
31aaea4a 5367 for_each_node_state(nid, N_MEMORY) {
d79154bb 5368 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
bb2a0de9 5369 LRU_ALL_ANON);
406eb0c9
YH
5370 seq_printf(m, " N%d=%lu", nid, node_nr);
5371 }
5372 seq_putc(m, '\n');
5373
d79154bb 5374 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
406eb0c9 5375 seq_printf(m, "unevictable=%lu", unevictable_nr);
31aaea4a 5376 for_each_node_state(nid, N_MEMORY) {
d79154bb 5377 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
bb2a0de9 5378 BIT(LRU_UNEVICTABLE));
406eb0c9
YH
5379 seq_printf(m, " N%d=%lu", nid, node_nr);
5380 }
5381 seq_putc(m, '\n');
5382 return 0;
5383}
5384#endif /* CONFIG_NUMA */
5385
af7c4b0e
JW
5386static inline void mem_cgroup_lru_names_not_uptodate(void)
5387{
5388 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5389}
5390
ab215884 5391static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
78ccf5b5 5392 struct seq_file *m)
d2ceb9b7 5393{
d79154bb 5394 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
af7c4b0e
JW
5395 struct mem_cgroup *mi;
5396 unsigned int i;
406eb0c9 5397
af7c4b0e 5398 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
bff6bb83 5399 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1dd3a273 5400 continue;
af7c4b0e
JW
5401 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5402 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
1dd3a273 5403 }
7b854121 5404
af7c4b0e
JW
5405 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5406 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5407 mem_cgroup_read_events(memcg, i));
5408
5409 for (i = 0; i < NR_LRU_LISTS; i++)
5410 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5411 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5412
14067bb3 5413 /* Hierarchical information */
fee7b548
KH
5414 {
5415 unsigned long long limit, memsw_limit;
d79154bb 5416 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
78ccf5b5 5417 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
fee7b548 5418 if (do_swap_account)
78ccf5b5
JW
5419 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5420 memsw_limit);
fee7b548 5421 }
7f016ee8 5422
af7c4b0e
JW
5423 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5424 long long val = 0;
5425
bff6bb83 5426 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1dd3a273 5427 continue;
af7c4b0e
JW
5428 for_each_mem_cgroup_tree(mi, memcg)
5429 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5430 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5431 }
5432
5433 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5434 unsigned long long val = 0;
5435
5436 for_each_mem_cgroup_tree(mi, memcg)
5437 val += mem_cgroup_read_events(mi, i);
5438 seq_printf(m, "total_%s %llu\n",
5439 mem_cgroup_events_names[i], val);
5440 }
5441
5442 for (i = 0; i < NR_LRU_LISTS; i++) {
5443 unsigned long long val = 0;
5444
5445 for_each_mem_cgroup_tree(mi, memcg)
5446 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5447 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
1dd3a273 5448 }
14067bb3 5449
7f016ee8 5450#ifdef CONFIG_DEBUG_VM
7f016ee8
KM
5451 {
5452 int nid, zid;
5453 struct mem_cgroup_per_zone *mz;
89abfab1 5454 struct zone_reclaim_stat *rstat;
7f016ee8
KM
5455 unsigned long recent_rotated[2] = {0, 0};
5456 unsigned long recent_scanned[2] = {0, 0};
5457
5458 for_each_online_node(nid)
5459 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
d79154bb 5460 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
89abfab1 5461 rstat = &mz->lruvec.reclaim_stat;
7f016ee8 5462
89abfab1
HD
5463 recent_rotated[0] += rstat->recent_rotated[0];
5464 recent_rotated[1] += rstat->recent_rotated[1];
5465 recent_scanned[0] += rstat->recent_scanned[0];
5466 recent_scanned[1] += rstat->recent_scanned[1];
7f016ee8 5467 }
78ccf5b5
JW
5468 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5469 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5470 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5471 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
7f016ee8
KM
5472 }
5473#endif
5474
d2ceb9b7
KH
5475 return 0;
5476}
5477
a7885eb8
KM
5478static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5479{
5480 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5481
1f4c025b 5482 return mem_cgroup_swappiness(memcg);
a7885eb8
KM
5483}
5484
5485static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5486 u64 val)
5487{
5488 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5489 struct mem_cgroup *parent;
068b38c1 5490
a7885eb8
KM
5491 if (val > 100)
5492 return -EINVAL;
5493
5494 if (cgrp->parent == NULL)
5495 return -EINVAL;
5496
5497 parent = mem_cgroup_from_cont(cgrp->parent);
068b38c1 5498
0999821b 5499 mutex_lock(&memcg_create_mutex);
068b38c1 5500
a7885eb8 5501 /* If under hierarchy, only empty-root can set this value */
b5f99b53 5502 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
0999821b 5503 mutex_unlock(&memcg_create_mutex);
a7885eb8 5504 return -EINVAL;
068b38c1 5505 }
a7885eb8 5506
a7885eb8 5507 memcg->swappiness = val;
a7885eb8 5508
0999821b 5509 mutex_unlock(&memcg_create_mutex);
068b38c1 5510
a7885eb8
KM
5511 return 0;
5512}
5513
2e72b634
KS
5514static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5515{
5516 struct mem_cgroup_threshold_ary *t;
5517 u64 usage;
5518 int i;
5519
5520 rcu_read_lock();
5521 if (!swap)
2c488db2 5522 t = rcu_dereference(memcg->thresholds.primary);
2e72b634 5523 else
2c488db2 5524 t = rcu_dereference(memcg->memsw_thresholds.primary);
2e72b634
KS
5525
5526 if (!t)
5527 goto unlock;
5528
5529 usage = mem_cgroup_usage(memcg, swap);
5530
5531 /*
748dad36 5532 * current_threshold points to threshold just below or equal to usage.
2e72b634
KS
5533 * If it's not true, a threshold was crossed after last
5534 * call of __mem_cgroup_threshold().
5535 */
5407a562 5536 i = t->current_threshold;
2e72b634
KS
5537
5538 /*
5539 * Iterate backward over array of thresholds starting from
5540 * current_threshold and check if a threshold is crossed.
5541 * If none of thresholds below usage is crossed, we read
5542 * only one element of the array here.
5543 */
5544 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5545 eventfd_signal(t->entries[i].eventfd, 1);
5546
5547 /* i = current_threshold + 1 */
5548 i++;
5549
5550 /*
5551 * Iterate forward over array of thresholds starting from
5552 * current_threshold+1 and check if a threshold is crossed.
5553 * If none of thresholds above usage is crossed, we read
5554 * only one element of the array here.
5555 */
5556 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5557 eventfd_signal(t->entries[i].eventfd, 1);
5558
5559 /* Update current_threshold */
5407a562 5560 t->current_threshold = i - 1;
2e72b634
KS
5561unlock:
5562 rcu_read_unlock();
5563}
5564
5565static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5566{
ad4ca5f4
KS
5567 while (memcg) {
5568 __mem_cgroup_threshold(memcg, false);
5569 if (do_swap_account)
5570 __mem_cgroup_threshold(memcg, true);
5571
5572 memcg = parent_mem_cgroup(memcg);
5573 }
2e72b634
KS
5574}
5575
5576static int compare_thresholds(const void *a, const void *b)
5577{
5578 const struct mem_cgroup_threshold *_a = a;
5579 const struct mem_cgroup_threshold *_b = b;
5580
5581 return _a->threshold - _b->threshold;
5582}
5583
c0ff4b85 5584static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
9490ff27
KH
5585{
5586 struct mem_cgroup_eventfd_list *ev;
5587
c0ff4b85 5588 list_for_each_entry(ev, &memcg->oom_notify, list)
9490ff27
KH
5589 eventfd_signal(ev->eventfd, 1);
5590 return 0;
5591}
5592
c0ff4b85 5593static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
9490ff27 5594{
7d74b06f
KH
5595 struct mem_cgroup *iter;
5596
c0ff4b85 5597 for_each_mem_cgroup_tree(iter, memcg)
7d74b06f 5598 mem_cgroup_oom_notify_cb(iter);
9490ff27
KH
5599}
5600
5601static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5602 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
2e72b634
KS
5603{
5604 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2c488db2
KS
5605 struct mem_cgroup_thresholds *thresholds;
5606 struct mem_cgroup_threshold_ary *new;
86ae53e1 5607 enum res_type type = MEMFILE_TYPE(cft->private);
2e72b634 5608 u64 threshold, usage;
2c488db2 5609 int i, size, ret;
2e72b634
KS
5610
5611 ret = res_counter_memparse_write_strategy(args, &threshold);
5612 if (ret)
5613 return ret;
5614
5615 mutex_lock(&memcg->thresholds_lock);
2c488db2 5616
2e72b634 5617 if (type == _MEM)
2c488db2 5618 thresholds = &memcg->thresholds;
2e72b634 5619 else if (type == _MEMSWAP)
2c488db2 5620 thresholds = &memcg->memsw_thresholds;
2e72b634
KS
5621 else
5622 BUG();
5623
5624 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5625
5626 /* Check if a threshold crossed before adding a new one */
2c488db2 5627 if (thresholds->primary)
2e72b634
KS
5628 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5629
2c488db2 5630 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
2e72b634
KS
5631
5632 /* Allocate memory for new array of thresholds */
2c488db2 5633 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
2e72b634 5634 GFP_KERNEL);
2c488db2 5635 if (!new) {
2e72b634
KS
5636 ret = -ENOMEM;
5637 goto unlock;
5638 }
2c488db2 5639 new->size = size;
2e72b634
KS
5640
5641 /* Copy thresholds (if any) to new array */
2c488db2
KS
5642 if (thresholds->primary) {
5643 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
2e72b634 5644 sizeof(struct mem_cgroup_threshold));
2c488db2
KS
5645 }
5646
2e72b634 5647 /* Add new threshold */
2c488db2
KS
5648 new->entries[size - 1].eventfd = eventfd;
5649 new->entries[size - 1].threshold = threshold;
2e72b634
KS
5650
5651 /* Sort thresholds. Registering of new threshold isn't time-critical */
2c488db2 5652 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
2e72b634
KS
5653 compare_thresholds, NULL);
5654
5655 /* Find current threshold */
2c488db2 5656 new->current_threshold = -1;
2e72b634 5657 for (i = 0; i < size; i++) {
748dad36 5658 if (new->entries[i].threshold <= usage) {
2e72b634 5659 /*
2c488db2
KS
5660 * new->current_threshold will not be used until
5661 * rcu_assign_pointer(), so it's safe to increment
2e72b634
KS
5662 * it here.
5663 */
2c488db2 5664 ++new->current_threshold;
748dad36
SZ
5665 } else
5666 break;
2e72b634
KS
5667 }
5668
2c488db2
KS
5669 /* Free old spare buffer and save old primary buffer as spare */
5670 kfree(thresholds->spare);
5671 thresholds->spare = thresholds->primary;
5672
5673 rcu_assign_pointer(thresholds->primary, new);
2e72b634 5674
907860ed 5675 /* To be sure that nobody uses thresholds */
2e72b634
KS
5676 synchronize_rcu();
5677
2e72b634
KS
5678unlock:
5679 mutex_unlock(&memcg->thresholds_lock);
5680
5681 return ret;
5682}
5683
907860ed 5684static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
9490ff27 5685 struct cftype *cft, struct eventfd_ctx *eventfd)
2e72b634
KS
5686{
5687 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2c488db2
KS
5688 struct mem_cgroup_thresholds *thresholds;
5689 struct mem_cgroup_threshold_ary *new;
86ae53e1 5690 enum res_type type = MEMFILE_TYPE(cft->private);
2e72b634 5691 u64 usage;
2c488db2 5692 int i, j, size;
2e72b634
KS
5693
5694 mutex_lock(&memcg->thresholds_lock);
5695 if (type == _MEM)
2c488db2 5696 thresholds = &memcg->thresholds;
2e72b634 5697 else if (type == _MEMSWAP)
2c488db2 5698 thresholds = &memcg->memsw_thresholds;
2e72b634
KS
5699 else
5700 BUG();
5701
371528ca
AV
5702 if (!thresholds->primary)
5703 goto unlock;
5704
2e72b634
KS
5705 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5706
5707 /* Check if a threshold crossed before removing */
5708 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5709
5710 /* Calculate new number of threshold */
2c488db2
KS
5711 size = 0;
5712 for (i = 0; i < thresholds->primary->size; i++) {
5713 if (thresholds->primary->entries[i].eventfd != eventfd)
2e72b634
KS
5714 size++;
5715 }
5716
2c488db2 5717 new = thresholds->spare;
907860ed 5718
2e72b634
KS
5719 /* Set thresholds array to NULL if we don't have thresholds */
5720 if (!size) {
2c488db2
KS
5721 kfree(new);
5722 new = NULL;
907860ed 5723 goto swap_buffers;
2e72b634
KS
5724 }
5725
2c488db2 5726 new->size = size;
2e72b634
KS
5727
5728 /* Copy thresholds and find current threshold */
2c488db2
KS
5729 new->current_threshold = -1;
5730 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5731 if (thresholds->primary->entries[i].eventfd == eventfd)
2e72b634
KS
5732 continue;
5733
2c488db2 5734 new->entries[j] = thresholds->primary->entries[i];
748dad36 5735 if (new->entries[j].threshold <= usage) {
2e72b634 5736 /*
2c488db2 5737 * new->current_threshold will not be used
2e72b634
KS
5738 * until rcu_assign_pointer(), so it's safe to increment
5739 * it here.
5740 */
2c488db2 5741 ++new->current_threshold;
2e72b634
KS
5742 }
5743 j++;
5744 }
5745
907860ed 5746swap_buffers:
2c488db2
KS
5747 /* Swap primary and spare array */
5748 thresholds->spare = thresholds->primary;
8c757763
SZ
5749 /* If all events are unregistered, free the spare array */
5750 if (!new) {
5751 kfree(thresholds->spare);
5752 thresholds->spare = NULL;
5753 }
5754
2c488db2 5755 rcu_assign_pointer(thresholds->primary, new);
2e72b634 5756
907860ed 5757 /* To be sure that nobody uses thresholds */
2e72b634 5758 synchronize_rcu();
371528ca 5759unlock:
2e72b634 5760 mutex_unlock(&memcg->thresholds_lock);
2e72b634 5761}
c1e862c1 5762
9490ff27
KH
5763static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5764 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5765{
5766 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5767 struct mem_cgroup_eventfd_list *event;
86ae53e1 5768 enum res_type type = MEMFILE_TYPE(cft->private);
9490ff27
KH
5769
5770 BUG_ON(type != _OOM_TYPE);
5771 event = kmalloc(sizeof(*event), GFP_KERNEL);
5772 if (!event)
5773 return -ENOMEM;
5774
1af8efe9 5775 spin_lock(&memcg_oom_lock);
9490ff27
KH
5776
5777 event->eventfd = eventfd;
5778 list_add(&event->list, &memcg->oom_notify);
5779
5780 /* already in OOM ? */
79dfdacc 5781 if (atomic_read(&memcg->under_oom))
9490ff27 5782 eventfd_signal(eventfd, 1);
1af8efe9 5783 spin_unlock(&memcg_oom_lock);
9490ff27
KH
5784
5785 return 0;
5786}
5787
907860ed 5788static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
9490ff27
KH
5789 struct cftype *cft, struct eventfd_ctx *eventfd)
5790{
c0ff4b85 5791 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
9490ff27 5792 struct mem_cgroup_eventfd_list *ev, *tmp;
86ae53e1 5793 enum res_type type = MEMFILE_TYPE(cft->private);
9490ff27
KH
5794
5795 BUG_ON(type != _OOM_TYPE);
5796
1af8efe9 5797 spin_lock(&memcg_oom_lock);
9490ff27 5798
c0ff4b85 5799 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
9490ff27
KH
5800 if (ev->eventfd == eventfd) {
5801 list_del(&ev->list);
5802 kfree(ev);
5803 }
5804 }
5805
1af8efe9 5806 spin_unlock(&memcg_oom_lock);
9490ff27
KH
5807}
5808
3c11ecf4
KH
5809static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5810 struct cftype *cft, struct cgroup_map_cb *cb)
5811{
c0ff4b85 5812 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3c11ecf4 5813
c0ff4b85 5814 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
3c11ecf4 5815
c0ff4b85 5816 if (atomic_read(&memcg->under_oom))
3c11ecf4
KH
5817 cb->fill(cb, "under_oom", 1);
5818 else
5819 cb->fill(cb, "under_oom", 0);
5820 return 0;
5821}
5822
3c11ecf4
KH
5823static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5824 struct cftype *cft, u64 val)
5825{
c0ff4b85 5826 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3c11ecf4
KH
5827 struct mem_cgroup *parent;
5828
5829 /* cannot set to root cgroup and only 0 and 1 are allowed */
5830 if (!cgrp->parent || !((val == 0) || (val == 1)))
5831 return -EINVAL;
5832
5833 parent = mem_cgroup_from_cont(cgrp->parent);
5834
0999821b 5835 mutex_lock(&memcg_create_mutex);
3c11ecf4 5836 /* oom-kill-disable is a flag for subhierarchy. */
b5f99b53 5837 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
0999821b 5838 mutex_unlock(&memcg_create_mutex);
3c11ecf4
KH
5839 return -EINVAL;
5840 }
c0ff4b85 5841 memcg->oom_kill_disable = val;
4d845ebf 5842 if (!val)
c0ff4b85 5843 memcg_oom_recover(memcg);
0999821b 5844 mutex_unlock(&memcg_create_mutex);
3c11ecf4
KH
5845 return 0;
5846}
5847
c255a458 5848#ifdef CONFIG_MEMCG_KMEM
cbe128e3 5849static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
e5671dfa 5850{
55007d84
GC
5851 int ret;
5852
2633d7a0 5853 memcg->kmemcg_id = -1;
55007d84
GC
5854 ret = memcg_propagate_kmem(memcg);
5855 if (ret)
5856 return ret;
2633d7a0 5857
1d62e436 5858 return mem_cgroup_sockets_init(memcg, ss);
573b400d 5859}
e5671dfa 5860
1d62e436 5861static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
d1a4c0b3 5862{
1d62e436 5863 mem_cgroup_sockets_destroy(memcg);
7de37682
GC
5864
5865 memcg_kmem_mark_dead(memcg);
5866
5867 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5868 return;
5869
5870 /*
5871 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5872 * path here, being careful not to race with memcg_uncharge_kmem: it is
5873 * possible that the charges went down to 0 between mark_dead and the
5874 * res_counter read, so in that case, we don't need the put
5875 */
5876 if (memcg_kmem_test_and_clear_dead(memcg))
5877 mem_cgroup_put(memcg);
d1a4c0b3 5878}
e5671dfa 5879#else
cbe128e3 5880static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
e5671dfa
GC
5881{
5882 return 0;
5883}
d1a4c0b3 5884
1d62e436 5885static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
d1a4c0b3
GC
5886{
5887}
e5671dfa
GC
5888#endif
5889
8cdea7c0
BS
5890static struct cftype mem_cgroup_files[] = {
5891 {
0eea1030 5892 .name = "usage_in_bytes",
8c7c6e34 5893 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
af36f906 5894 .read = mem_cgroup_read,
9490ff27
KH
5895 .register_event = mem_cgroup_usage_register_event,
5896 .unregister_event = mem_cgroup_usage_unregister_event,
8cdea7c0 5897 },
c84872e1
PE
5898 {
5899 .name = "max_usage_in_bytes",
8c7c6e34 5900 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
29f2a4da 5901 .trigger = mem_cgroup_reset,
af36f906 5902 .read = mem_cgroup_read,
c84872e1 5903 },
8cdea7c0 5904 {
0eea1030 5905 .name = "limit_in_bytes",
8c7c6e34 5906 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
856c13aa 5907 .write_string = mem_cgroup_write,
af36f906 5908 .read = mem_cgroup_read,
8cdea7c0 5909 },
296c81d8
BS
5910 {
5911 .name = "soft_limit_in_bytes",
5912 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5913 .write_string = mem_cgroup_write,
af36f906 5914 .read = mem_cgroup_read,
296c81d8 5915 },
8cdea7c0
BS
5916 {
5917 .name = "failcnt",
8c7c6e34 5918 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
29f2a4da 5919 .trigger = mem_cgroup_reset,
af36f906 5920 .read = mem_cgroup_read,
8cdea7c0 5921 },
d2ceb9b7
KH
5922 {
5923 .name = "stat",
ab215884 5924 .read_seq_string = memcg_stat_show,
d2ceb9b7 5925 },
c1e862c1
KH
5926 {
5927 .name = "force_empty",
5928 .trigger = mem_cgroup_force_empty_write,
5929 },
18f59ea7
BS
5930 {
5931 .name = "use_hierarchy",
f00baae7 5932 .flags = CFTYPE_INSANE,
18f59ea7
BS
5933 .write_u64 = mem_cgroup_hierarchy_write,
5934 .read_u64 = mem_cgroup_hierarchy_read,
5935 },
a7885eb8
KM
5936 {
5937 .name = "swappiness",
5938 .read_u64 = mem_cgroup_swappiness_read,
5939 .write_u64 = mem_cgroup_swappiness_write,
5940 },
7dc74be0
DN
5941 {
5942 .name = "move_charge_at_immigrate",
5943 .read_u64 = mem_cgroup_move_charge_read,
5944 .write_u64 = mem_cgroup_move_charge_write,
5945 },
9490ff27
KH
5946 {
5947 .name = "oom_control",
3c11ecf4
KH
5948 .read_map = mem_cgroup_oom_control_read,
5949 .write_u64 = mem_cgroup_oom_control_write,
9490ff27
KH
5950 .register_event = mem_cgroup_oom_register_event,
5951 .unregister_event = mem_cgroup_oom_unregister_event,
5952 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5953 },
70ddf637
AV
5954 {
5955 .name = "pressure_level",
5956 .register_event = vmpressure_register_event,
5957 .unregister_event = vmpressure_unregister_event,
5958 },
406eb0c9
YH
5959#ifdef CONFIG_NUMA
5960 {
5961 .name = "numa_stat",
ab215884 5962 .read_seq_string = memcg_numa_stat_show,
406eb0c9
YH
5963 },
5964#endif
510fc4e1
GC
5965#ifdef CONFIG_MEMCG_KMEM
5966 {
5967 .name = "kmem.limit_in_bytes",
5968 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5969 .write_string = mem_cgroup_write,
5970 .read = mem_cgroup_read,
5971 },
5972 {
5973 .name = "kmem.usage_in_bytes",
5974 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5975 .read = mem_cgroup_read,
5976 },
5977 {
5978 .name = "kmem.failcnt",
5979 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5980 .trigger = mem_cgroup_reset,
5981 .read = mem_cgroup_read,
5982 },
5983 {
5984 .name = "kmem.max_usage_in_bytes",
5985 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5986 .trigger = mem_cgroup_reset,
5987 .read = mem_cgroup_read,
5988 },
749c5415
GC
5989#ifdef CONFIG_SLABINFO
5990 {
5991 .name = "kmem.slabinfo",
5992 .read_seq_string = mem_cgroup_slabinfo_read,
5993 },
5994#endif
8c7c6e34 5995#endif
6bc10349 5996 { }, /* terminate */
af36f906 5997};
8c7c6e34 5998
2d11085e
MH
5999#ifdef CONFIG_MEMCG_SWAP
6000static struct cftype memsw_cgroup_files[] = {
6001 {
6002 .name = "memsw.usage_in_bytes",
6003 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6004 .read = mem_cgroup_read,
6005 .register_event = mem_cgroup_usage_register_event,
6006 .unregister_event = mem_cgroup_usage_unregister_event,
6007 },
6008 {
6009 .name = "memsw.max_usage_in_bytes",
6010 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6011 .trigger = mem_cgroup_reset,
6012 .read = mem_cgroup_read,
6013 },
6014 {
6015 .name = "memsw.limit_in_bytes",
6016 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6017 .write_string = mem_cgroup_write,
6018 .read = mem_cgroup_read,
6019 },
6020 {
6021 .name = "memsw.failcnt",
6022 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6023 .trigger = mem_cgroup_reset,
6024 .read = mem_cgroup_read,
6025 },
6026 { }, /* terminate */
6027};
6028#endif
c0ff4b85 6029static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6d12e2d8
KH
6030{
6031 struct mem_cgroup_per_node *pn;
1ecaab2b 6032 struct mem_cgroup_per_zone *mz;
41e3355d 6033 int zone, tmp = node;
1ecaab2b
KH
6034 /*
6035 * This routine is called against possible nodes.
6036 * But it's BUG to call kmalloc() against offline node.
6037 *
6038 * TODO: this routine can waste much memory for nodes which will
6039 * never be onlined. It's better to use memory hotplug callback
6040 * function.
6041 */
41e3355d
KH
6042 if (!node_state(node, N_NORMAL_MEMORY))
6043 tmp = -1;
17295c88 6044 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6d12e2d8
KH
6045 if (!pn)
6046 return 1;
1ecaab2b 6047
1ecaab2b
KH
6048 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6049 mz = &pn->zoneinfo[zone];
bea8c150 6050 lruvec_init(&mz->lruvec);
f64c3f54 6051 mz->usage_in_excess = 0;
4e416953 6052 mz->on_tree = false;
d79154bb 6053 mz->memcg = memcg;
1ecaab2b 6054 }
0a619e58 6055 memcg->info.nodeinfo[node] = pn;
6d12e2d8
KH
6056 return 0;
6057}
6058
c0ff4b85 6059static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
1ecaab2b 6060{
c0ff4b85 6061 kfree(memcg->info.nodeinfo[node]);
1ecaab2b
KH
6062}
6063
33327948
KH
6064static struct mem_cgroup *mem_cgroup_alloc(void)
6065{
d79154bb 6066 struct mem_cgroup *memcg;
45cf7ebd 6067 size_t size = memcg_size();
33327948 6068
45cf7ebd 6069 /* Can be very big if nr_node_ids is very big */
c8dad2bb 6070 if (size < PAGE_SIZE)
d79154bb 6071 memcg = kzalloc(size, GFP_KERNEL);
33327948 6072 else
d79154bb 6073 memcg = vzalloc(size);
33327948 6074
d79154bb 6075 if (!memcg)
e7bbcdf3
DC
6076 return NULL;
6077
d79154bb
HD
6078 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6079 if (!memcg->stat)
d2e61b8d 6080 goto out_free;
d79154bb
HD
6081 spin_lock_init(&memcg->pcp_counter_lock);
6082 return memcg;
d2e61b8d
DC
6083
6084out_free:
6085 if (size < PAGE_SIZE)
d79154bb 6086 kfree(memcg);
d2e61b8d 6087 else
d79154bb 6088 vfree(memcg);
d2e61b8d 6089 return NULL;
33327948
KH
6090}
6091
59927fb9 6092/*
c8b2a36f
GC
6093 * At destroying mem_cgroup, references from swap_cgroup can remain.
6094 * (scanning all at force_empty is too costly...)
6095 *
6096 * Instead of clearing all references at force_empty, we remember
6097 * the number of reference from swap_cgroup and free mem_cgroup when
6098 * it goes down to 0.
6099 *
6100 * Removal of cgroup itself succeeds regardless of refs from swap.
59927fb9 6101 */
c8b2a36f
GC
6102
6103static void __mem_cgroup_free(struct mem_cgroup *memcg)
59927fb9 6104{
c8b2a36f 6105 int node;
45cf7ebd 6106 size_t size = memcg_size();
59927fb9 6107
c8b2a36f
GC
6108 mem_cgroup_remove_from_trees(memcg);
6109 free_css_id(&mem_cgroup_subsys, &memcg->css);
6110
6111 for_each_node(node)
6112 free_mem_cgroup_per_zone_info(memcg, node);
6113
6114 free_percpu(memcg->stat);
6115
3f134619
GC
6116 /*
6117 * We need to make sure that (at least for now), the jump label
6118 * destruction code runs outside of the cgroup lock. This is because
6119 * get_online_cpus(), which is called from the static_branch update,
6120 * can't be called inside the cgroup_lock. cpusets are the ones
6121 * enforcing this dependency, so if they ever change, we might as well.
6122 *
6123 * schedule_work() will guarantee this happens. Be careful if you need
6124 * to move this code around, and make sure it is outside
6125 * the cgroup_lock.
6126 */
a8964b9b 6127 disarm_static_keys(memcg);
3afe36b1
GC
6128 if (size < PAGE_SIZE)
6129 kfree(memcg);
6130 else
6131 vfree(memcg);
59927fb9 6132}
3afe36b1 6133
59927fb9 6134
8c7c6e34 6135/*
c8b2a36f
GC
6136 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6137 * but in process context. The work_freeing structure is overlaid
6138 * on the rcu_freeing structure, which itself is overlaid on memsw.
8c7c6e34 6139 */
c8b2a36f 6140static void free_work(struct work_struct *work)
33327948 6141{
c8b2a36f 6142 struct mem_cgroup *memcg;
08e552c6 6143
c8b2a36f
GC
6144 memcg = container_of(work, struct mem_cgroup, work_freeing);
6145 __mem_cgroup_free(memcg);
6146}
04046e1a 6147
c8b2a36f
GC
6148static void free_rcu(struct rcu_head *rcu_head)
6149{
6150 struct mem_cgroup *memcg;
08e552c6 6151
c8b2a36f
GC
6152 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6153 INIT_WORK(&memcg->work_freeing, free_work);
6154 schedule_work(&memcg->work_freeing);
33327948
KH
6155}
6156
c0ff4b85 6157static void mem_cgroup_get(struct mem_cgroup *memcg)
8c7c6e34 6158{
c0ff4b85 6159 atomic_inc(&memcg->refcnt);
8c7c6e34
KH
6160}
6161
c0ff4b85 6162static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
8c7c6e34 6163{
c0ff4b85
R
6164 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6165 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
c8b2a36f 6166 call_rcu(&memcg->rcu_freeing, free_rcu);
7bcc1bb1
DN
6167 if (parent)
6168 mem_cgroup_put(parent);
6169 }
8c7c6e34
KH
6170}
6171
c0ff4b85 6172static void mem_cgroup_put(struct mem_cgroup *memcg)
483c30b5 6173{
c0ff4b85 6174 __mem_cgroup_put(memcg, 1);
483c30b5
DN
6175}
6176
7bcc1bb1
DN
6177/*
6178 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6179 */
e1aab161 6180struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
7bcc1bb1 6181{
c0ff4b85 6182 if (!memcg->res.parent)
7bcc1bb1 6183 return NULL;
c0ff4b85 6184 return mem_cgroup_from_res_counter(memcg->res.parent, res);
7bcc1bb1 6185}
e1aab161 6186EXPORT_SYMBOL(parent_mem_cgroup);
33327948 6187
8787a1df 6188static void __init mem_cgroup_soft_limit_tree_init(void)
f64c3f54
BS
6189{
6190 struct mem_cgroup_tree_per_node *rtpn;
6191 struct mem_cgroup_tree_per_zone *rtpz;
6192 int tmp, node, zone;
6193
3ed28fa1 6194 for_each_node(node) {
f64c3f54
BS
6195 tmp = node;
6196 if (!node_state(node, N_NORMAL_MEMORY))
6197 tmp = -1;
6198 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
8787a1df 6199 BUG_ON(!rtpn);
f64c3f54
BS
6200
6201 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6202
6203 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6204 rtpz = &rtpn->rb_tree_per_zone[zone];
6205 rtpz->rb_root = RB_ROOT;
6206 spin_lock_init(&rtpz->lock);
6207 }
6208 }
f64c3f54
BS
6209}
6210
0eb253e2 6211static struct cgroup_subsys_state * __ref
92fb9748 6212mem_cgroup_css_alloc(struct cgroup *cont)
8cdea7c0 6213{
d142e3e6 6214 struct mem_cgroup *memcg;
04046e1a 6215 long error = -ENOMEM;
6d12e2d8 6216 int node;
8cdea7c0 6217
c0ff4b85
R
6218 memcg = mem_cgroup_alloc();
6219 if (!memcg)
04046e1a 6220 return ERR_PTR(error);
78fb7466 6221
3ed28fa1 6222 for_each_node(node)
c0ff4b85 6223 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6d12e2d8 6224 goto free_out;
f64c3f54 6225
c077719b 6226 /* root ? */
28dbc4b6 6227 if (cont->parent == NULL) {
a41c58a6 6228 root_mem_cgroup = memcg;
d142e3e6
GC
6229 res_counter_init(&memcg->res, NULL);
6230 res_counter_init(&memcg->memsw, NULL);
6231 res_counter_init(&memcg->kmem, NULL);
18f59ea7 6232 }
28dbc4b6 6233
d142e3e6
GC
6234 memcg->last_scanned_node = MAX_NUMNODES;
6235 INIT_LIST_HEAD(&memcg->oom_notify);
6236 atomic_set(&memcg->refcnt, 1);
6237 memcg->move_charge_at_immigrate = 0;
6238 mutex_init(&memcg->thresholds_lock);
6239 spin_lock_init(&memcg->move_lock);
70ddf637 6240 vmpressure_init(&memcg->vmpressure);
d142e3e6
GC
6241
6242 return &memcg->css;
6243
6244free_out:
6245 __mem_cgroup_free(memcg);
6246 return ERR_PTR(error);
6247}
6248
6249static int
6250mem_cgroup_css_online(struct cgroup *cont)
6251{
6252 struct mem_cgroup *memcg, *parent;
6253 int error = 0;
6254
6255 if (!cont->parent)
6256 return 0;
6257
0999821b 6258 mutex_lock(&memcg_create_mutex);
d142e3e6
GC
6259 memcg = mem_cgroup_from_cont(cont);
6260 parent = mem_cgroup_from_cont(cont->parent);
6261
6262 memcg->use_hierarchy = parent->use_hierarchy;
6263 memcg->oom_kill_disable = parent->oom_kill_disable;
6264 memcg->swappiness = mem_cgroup_swappiness(parent);
6265
6266 if (parent->use_hierarchy) {
c0ff4b85
R
6267 res_counter_init(&memcg->res, &parent->res);
6268 res_counter_init(&memcg->memsw, &parent->memsw);
510fc4e1 6269 res_counter_init(&memcg->kmem, &parent->kmem);
55007d84 6270
7bcc1bb1
DN
6271 /*
6272 * We increment refcnt of the parent to ensure that we can
6273 * safely access it on res_counter_charge/uncharge.
6274 * This refcnt will be decremented when freeing this
6275 * mem_cgroup(see mem_cgroup_put).
6276 */
6277 mem_cgroup_get(parent);
18f59ea7 6278 } else {
c0ff4b85
R
6279 res_counter_init(&memcg->res, NULL);
6280 res_counter_init(&memcg->memsw, NULL);
510fc4e1 6281 res_counter_init(&memcg->kmem, NULL);
8c7f6edb
TH
6282 /*
6283 * Deeper hierachy with use_hierarchy == false doesn't make
6284 * much sense so let cgroup subsystem know about this
6285 * unfortunate state in our controller.
6286 */
d142e3e6 6287 if (parent != root_mem_cgroup)
8c7f6edb 6288 mem_cgroup_subsys.broken_hierarchy = true;
18f59ea7 6289 }
cbe128e3
GC
6290
6291 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
0999821b 6292 mutex_unlock(&memcg_create_mutex);
cbe128e3
GC
6293 if (error) {
6294 /*
6295 * We call put now because our (and parent's) refcnts
6296 * are already in place. mem_cgroup_put() will internally
6297 * call __mem_cgroup_free, so return directly
6298 */
6299 mem_cgroup_put(memcg);
e4715f01
GC
6300 if (parent->use_hierarchy)
6301 mem_cgroup_put(parent);
cbe128e3 6302 }
d142e3e6 6303 return error;
8cdea7c0
BS
6304}
6305
5f578161
MH
6306/*
6307 * Announce all parents that a group from their hierarchy is gone.
6308 */
6309static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6310{
6311 struct mem_cgroup *parent = memcg;
6312
6313 while ((parent = parent_mem_cgroup(parent)))
6314 atomic_inc(&parent->dead_count);
6315
6316 /*
6317 * if the root memcg is not hierarchical we have to check it
6318 * explicitely.
6319 */
6320 if (!root_mem_cgroup->use_hierarchy)
6321 atomic_inc(&root_mem_cgroup->dead_count);
6322}
6323
92fb9748 6324static void mem_cgroup_css_offline(struct cgroup *cont)
df878fb0 6325{
c0ff4b85 6326 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
ec64f515 6327
5f578161 6328 mem_cgroup_invalidate_reclaim_iterators(memcg);
ab5196c2 6329 mem_cgroup_reparent_charges(memcg);
1f458cbf 6330 mem_cgroup_destroy_all_caches(memcg);
df878fb0
KH
6331}
6332
92fb9748 6333static void mem_cgroup_css_free(struct cgroup *cont)
8cdea7c0 6334{
c0ff4b85 6335 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
c268e994 6336
1d62e436 6337 kmem_cgroup_destroy(memcg);
d1a4c0b3 6338
c0ff4b85 6339 mem_cgroup_put(memcg);
8cdea7c0
BS
6340}
6341
02491447 6342#ifdef CONFIG_MMU
7dc74be0 6343/* Handlers for move charge at task migration. */
854ffa8d
DN
6344#define PRECHARGE_COUNT_AT_ONCE 256
6345static int mem_cgroup_do_precharge(unsigned long count)
7dc74be0 6346{
854ffa8d
DN
6347 int ret = 0;
6348 int batch_count = PRECHARGE_COUNT_AT_ONCE;
c0ff4b85 6349 struct mem_cgroup *memcg = mc.to;
4ffef5fe 6350
c0ff4b85 6351 if (mem_cgroup_is_root(memcg)) {
854ffa8d
DN
6352 mc.precharge += count;
6353 /* we don't need css_get for root */
6354 return ret;
6355 }
6356 /* try to charge at once */
6357 if (count > 1) {
6358 struct res_counter *dummy;
6359 /*
c0ff4b85 6360 * "memcg" cannot be under rmdir() because we've already checked
854ffa8d
DN
6361 * by cgroup_lock_live_cgroup() that it is not removed and we
6362 * are still under the same cgroup_mutex. So we can postpone
6363 * css_get().
6364 */
c0ff4b85 6365 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
854ffa8d 6366 goto one_by_one;
c0ff4b85 6367 if (do_swap_account && res_counter_charge(&memcg->memsw,
854ffa8d 6368 PAGE_SIZE * count, &dummy)) {
c0ff4b85 6369 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
854ffa8d
DN
6370 goto one_by_one;
6371 }
6372 mc.precharge += count;
854ffa8d
DN
6373 return ret;
6374 }
6375one_by_one:
6376 /* fall back to one by one charge */
6377 while (count--) {
6378 if (signal_pending(current)) {
6379 ret = -EINTR;
6380 break;
6381 }
6382 if (!batch_count--) {
6383 batch_count = PRECHARGE_COUNT_AT_ONCE;
6384 cond_resched();
6385 }
c0ff4b85
R
6386 ret = __mem_cgroup_try_charge(NULL,
6387 GFP_KERNEL, 1, &memcg, false);
38c5d72f 6388 if (ret)
854ffa8d 6389 /* mem_cgroup_clear_mc() will do uncharge later */
38c5d72f 6390 return ret;
854ffa8d
DN
6391 mc.precharge++;
6392 }
4ffef5fe
DN
6393 return ret;
6394}
6395
6396/**
8d32ff84 6397 * get_mctgt_type - get target type of moving charge
4ffef5fe
DN
6398 * @vma: the vma the pte to be checked belongs
6399 * @addr: the address corresponding to the pte to be checked
6400 * @ptent: the pte to be checked
02491447 6401 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4ffef5fe
DN
6402 *
6403 * Returns
6404 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6405 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6406 * move charge. if @target is not NULL, the page is stored in target->page
6407 * with extra refcnt got(Callers should handle it).
02491447
DN
6408 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6409 * target for charge migration. if @target is not NULL, the entry is stored
6410 * in target->ent.
4ffef5fe
DN
6411 *
6412 * Called with pte lock held.
6413 */
4ffef5fe
DN
6414union mc_target {
6415 struct page *page;
02491447 6416 swp_entry_t ent;
4ffef5fe
DN
6417};
6418
4ffef5fe 6419enum mc_target_type {
8d32ff84 6420 MC_TARGET_NONE = 0,
4ffef5fe 6421 MC_TARGET_PAGE,
02491447 6422 MC_TARGET_SWAP,
4ffef5fe
DN
6423};
6424
90254a65
DN
6425static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6426 unsigned long addr, pte_t ptent)
4ffef5fe 6427{
90254a65 6428 struct page *page = vm_normal_page(vma, addr, ptent);
4ffef5fe 6429
90254a65
DN
6430 if (!page || !page_mapped(page))
6431 return NULL;
6432 if (PageAnon(page)) {
6433 /* we don't move shared anon */
4b91355e 6434 if (!move_anon())
90254a65 6435 return NULL;
87946a72
DN
6436 } else if (!move_file())
6437 /* we ignore mapcount for file pages */
90254a65
DN
6438 return NULL;
6439 if (!get_page_unless_zero(page))
6440 return NULL;
6441
6442 return page;
6443}
6444
4b91355e 6445#ifdef CONFIG_SWAP
90254a65
DN
6446static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6447 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6448{
90254a65
DN
6449 struct page *page = NULL;
6450 swp_entry_t ent = pte_to_swp_entry(ptent);
6451
6452 if (!move_anon() || non_swap_entry(ent))
6453 return NULL;
4b91355e
KH
6454 /*
6455 * Because lookup_swap_cache() updates some statistics counter,
6456 * we call find_get_page() with swapper_space directly.
6457 */
33806f06 6458 page = find_get_page(swap_address_space(ent), ent.val);
90254a65
DN
6459 if (do_swap_account)
6460 entry->val = ent.val;
6461
6462 return page;
6463}
4b91355e
KH
6464#else
6465static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6466 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6467{
6468 return NULL;
6469}
6470#endif
90254a65 6471
87946a72
DN
6472static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6473 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6474{
6475 struct page *page = NULL;
87946a72
DN
6476 struct address_space *mapping;
6477 pgoff_t pgoff;
6478
6479 if (!vma->vm_file) /* anonymous vma */
6480 return NULL;
6481 if (!move_file())
6482 return NULL;
6483
87946a72
DN
6484 mapping = vma->vm_file->f_mapping;
6485 if (pte_none(ptent))
6486 pgoff = linear_page_index(vma, addr);
6487 else /* pte_file(ptent) is true */
6488 pgoff = pte_to_pgoff(ptent);
6489
6490 /* page is moved even if it's not RSS of this task(page-faulted). */
aa3b1895
HD
6491 page = find_get_page(mapping, pgoff);
6492
6493#ifdef CONFIG_SWAP
6494 /* shmem/tmpfs may report page out on swap: account for that too. */
6495 if (radix_tree_exceptional_entry(page)) {
6496 swp_entry_t swap = radix_to_swp_entry(page);
87946a72 6497 if (do_swap_account)
aa3b1895 6498 *entry = swap;
33806f06 6499 page = find_get_page(swap_address_space(swap), swap.val);
87946a72 6500 }
aa3b1895 6501#endif
87946a72
DN
6502 return page;
6503}
6504
8d32ff84 6505static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
90254a65
DN
6506 unsigned long addr, pte_t ptent, union mc_target *target)
6507{
6508 struct page *page = NULL;
6509 struct page_cgroup *pc;
8d32ff84 6510 enum mc_target_type ret = MC_TARGET_NONE;
90254a65
DN
6511 swp_entry_t ent = { .val = 0 };
6512
6513 if (pte_present(ptent))
6514 page = mc_handle_present_pte(vma, addr, ptent);
6515 else if (is_swap_pte(ptent))
6516 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
87946a72
DN
6517 else if (pte_none(ptent) || pte_file(ptent))
6518 page = mc_handle_file_pte(vma, addr, ptent, &ent);
90254a65
DN
6519
6520 if (!page && !ent.val)
8d32ff84 6521 return ret;
02491447
DN
6522 if (page) {
6523 pc = lookup_page_cgroup(page);
6524 /*
6525 * Do only loose check w/o page_cgroup lock.
6526 * mem_cgroup_move_account() checks the pc is valid or not under
6527 * the lock.
6528 */
6529 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6530 ret = MC_TARGET_PAGE;
6531 if (target)
6532 target->page = page;
6533 }
6534 if (!ret || !target)
6535 put_page(page);
6536 }
90254a65
DN
6537 /* There is a swap entry and a page doesn't exist or isn't charged */
6538 if (ent.val && !ret &&
9fb4b7cc 6539 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
7f0f1546
KH
6540 ret = MC_TARGET_SWAP;
6541 if (target)
6542 target->ent = ent;
4ffef5fe 6543 }
4ffef5fe
DN
6544 return ret;
6545}
6546
12724850
NH
6547#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6548/*
6549 * We don't consider swapping or file mapped pages because THP does not
6550 * support them for now.
6551 * Caller should make sure that pmd_trans_huge(pmd) is true.
6552 */
6553static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6554 unsigned long addr, pmd_t pmd, union mc_target *target)
6555{
6556 struct page *page = NULL;
6557 struct page_cgroup *pc;
6558 enum mc_target_type ret = MC_TARGET_NONE;
6559
6560 page = pmd_page(pmd);
6561 VM_BUG_ON(!page || !PageHead(page));
6562 if (!move_anon())
6563 return ret;
6564 pc = lookup_page_cgroup(page);
6565 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6566 ret = MC_TARGET_PAGE;
6567 if (target) {
6568 get_page(page);
6569 target->page = page;
6570 }
6571 }
6572 return ret;
6573}
6574#else
6575static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6576 unsigned long addr, pmd_t pmd, union mc_target *target)
6577{
6578 return MC_TARGET_NONE;
6579}
6580#endif
6581
4ffef5fe
DN
6582static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6583 unsigned long addr, unsigned long end,
6584 struct mm_walk *walk)
6585{
6586 struct vm_area_struct *vma = walk->private;
6587 pte_t *pte;
6588 spinlock_t *ptl;
6589
12724850
NH
6590 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6591 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6592 mc.precharge += HPAGE_PMD_NR;
6593 spin_unlock(&vma->vm_mm->page_table_lock);
1a5a9906 6594 return 0;
12724850 6595 }
03319327 6596
45f83cef
AA
6597 if (pmd_trans_unstable(pmd))
6598 return 0;
4ffef5fe
DN
6599 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6600 for (; addr != end; pte++, addr += PAGE_SIZE)
8d32ff84 6601 if (get_mctgt_type(vma, addr, *pte, NULL))
4ffef5fe
DN
6602 mc.precharge++; /* increment precharge temporarily */
6603 pte_unmap_unlock(pte - 1, ptl);
6604 cond_resched();
6605
7dc74be0
DN
6606 return 0;
6607}
6608
4ffef5fe
DN
6609static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6610{
6611 unsigned long precharge;
6612 struct vm_area_struct *vma;
6613
dfe076b0 6614 down_read(&mm->mmap_sem);
4ffef5fe
DN
6615 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6616 struct mm_walk mem_cgroup_count_precharge_walk = {
6617 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6618 .mm = mm,
6619 .private = vma,
6620 };
6621 if (is_vm_hugetlb_page(vma))
6622 continue;
4ffef5fe
DN
6623 walk_page_range(vma->vm_start, vma->vm_end,
6624 &mem_cgroup_count_precharge_walk);
6625 }
dfe076b0 6626 up_read(&mm->mmap_sem);
4ffef5fe
DN
6627
6628 precharge = mc.precharge;
6629 mc.precharge = 0;
6630
6631 return precharge;
6632}
6633
4ffef5fe
DN
6634static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6635{
dfe076b0
DN
6636 unsigned long precharge = mem_cgroup_count_precharge(mm);
6637
6638 VM_BUG_ON(mc.moving_task);
6639 mc.moving_task = current;
6640 return mem_cgroup_do_precharge(precharge);
4ffef5fe
DN
6641}
6642
dfe076b0
DN
6643/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6644static void __mem_cgroup_clear_mc(void)
4ffef5fe 6645{
2bd9bb20
KH
6646 struct mem_cgroup *from = mc.from;
6647 struct mem_cgroup *to = mc.to;
6648
4ffef5fe 6649 /* we must uncharge all the leftover precharges from mc.to */
854ffa8d
DN
6650 if (mc.precharge) {
6651 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6652 mc.precharge = 0;
6653 }
6654 /*
6655 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6656 * we must uncharge here.
6657 */
6658 if (mc.moved_charge) {
6659 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6660 mc.moved_charge = 0;
4ffef5fe 6661 }
483c30b5
DN
6662 /* we must fixup refcnts and charges */
6663 if (mc.moved_swap) {
483c30b5
DN
6664 /* uncharge swap account from the old cgroup */
6665 if (!mem_cgroup_is_root(mc.from))
6666 res_counter_uncharge(&mc.from->memsw,
6667 PAGE_SIZE * mc.moved_swap);
6668 __mem_cgroup_put(mc.from, mc.moved_swap);
6669
6670 if (!mem_cgroup_is_root(mc.to)) {
6671 /*
6672 * we charged both to->res and to->memsw, so we should
6673 * uncharge to->res.
6674 */
6675 res_counter_uncharge(&mc.to->res,
6676 PAGE_SIZE * mc.moved_swap);
483c30b5
DN
6677 }
6678 /* we've already done mem_cgroup_get(mc.to) */
483c30b5
DN
6679 mc.moved_swap = 0;
6680 }
dfe076b0
DN
6681 memcg_oom_recover(from);
6682 memcg_oom_recover(to);
6683 wake_up_all(&mc.waitq);
6684}
6685
6686static void mem_cgroup_clear_mc(void)
6687{
6688 struct mem_cgroup *from = mc.from;
6689
6690 /*
6691 * we must clear moving_task before waking up waiters at the end of
6692 * task migration.
6693 */
6694 mc.moving_task = NULL;
6695 __mem_cgroup_clear_mc();
2bd9bb20 6696 spin_lock(&mc.lock);
4ffef5fe
DN
6697 mc.from = NULL;
6698 mc.to = NULL;
2bd9bb20 6699 spin_unlock(&mc.lock);
32047e2a 6700 mem_cgroup_end_move(from);
4ffef5fe
DN
6701}
6702
761b3ef5
LZ
6703static int mem_cgroup_can_attach(struct cgroup *cgroup,
6704 struct cgroup_taskset *tset)
7dc74be0 6705{
2f7ee569 6706 struct task_struct *p = cgroup_taskset_first(tset);
7dc74be0 6707 int ret = 0;
c0ff4b85 6708 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
ee5e8472 6709 unsigned long move_charge_at_immigrate;
7dc74be0 6710
ee5e8472
GC
6711 /*
6712 * We are now commited to this value whatever it is. Changes in this
6713 * tunable will only affect upcoming migrations, not the current one.
6714 * So we need to save it, and keep it going.
6715 */
6716 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6717 if (move_charge_at_immigrate) {
7dc74be0
DN
6718 struct mm_struct *mm;
6719 struct mem_cgroup *from = mem_cgroup_from_task(p);
6720
c0ff4b85 6721 VM_BUG_ON(from == memcg);
7dc74be0
DN
6722
6723 mm = get_task_mm(p);
6724 if (!mm)
6725 return 0;
7dc74be0 6726 /* We move charges only when we move a owner of the mm */
4ffef5fe
DN
6727 if (mm->owner == p) {
6728 VM_BUG_ON(mc.from);
6729 VM_BUG_ON(mc.to);
6730 VM_BUG_ON(mc.precharge);
854ffa8d 6731 VM_BUG_ON(mc.moved_charge);
483c30b5 6732 VM_BUG_ON(mc.moved_swap);
32047e2a 6733 mem_cgroup_start_move(from);
2bd9bb20 6734 spin_lock(&mc.lock);
4ffef5fe 6735 mc.from = from;
c0ff4b85 6736 mc.to = memcg;
ee5e8472 6737 mc.immigrate_flags = move_charge_at_immigrate;
2bd9bb20 6738 spin_unlock(&mc.lock);
dfe076b0 6739 /* We set mc.moving_task later */
4ffef5fe
DN
6740
6741 ret = mem_cgroup_precharge_mc(mm);
6742 if (ret)
6743 mem_cgroup_clear_mc();
dfe076b0
DN
6744 }
6745 mmput(mm);
7dc74be0
DN
6746 }
6747 return ret;
6748}
6749
761b3ef5
LZ
6750static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6751 struct cgroup_taskset *tset)
7dc74be0 6752{
4ffef5fe 6753 mem_cgroup_clear_mc();
7dc74be0
DN
6754}
6755
4ffef5fe
DN
6756static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6757 unsigned long addr, unsigned long end,
6758 struct mm_walk *walk)
7dc74be0 6759{
4ffef5fe
DN
6760 int ret = 0;
6761 struct vm_area_struct *vma = walk->private;
6762 pte_t *pte;
6763 spinlock_t *ptl;
12724850
NH
6764 enum mc_target_type target_type;
6765 union mc_target target;
6766 struct page *page;
6767 struct page_cgroup *pc;
4ffef5fe 6768
12724850
NH
6769 /*
6770 * We don't take compound_lock() here but no race with splitting thp
6771 * happens because:
6772 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6773 * under splitting, which means there's no concurrent thp split,
6774 * - if another thread runs into split_huge_page() just after we
6775 * entered this if-block, the thread must wait for page table lock
6776 * to be unlocked in __split_huge_page_splitting(), where the main
6777 * part of thp split is not executed yet.
6778 */
6779 if (pmd_trans_huge_lock(pmd, vma) == 1) {
62ade86a 6780 if (mc.precharge < HPAGE_PMD_NR) {
12724850
NH
6781 spin_unlock(&vma->vm_mm->page_table_lock);
6782 return 0;
6783 }
6784 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6785 if (target_type == MC_TARGET_PAGE) {
6786 page = target.page;
6787 if (!isolate_lru_page(page)) {
6788 pc = lookup_page_cgroup(page);
6789 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
2f3479b1 6790 pc, mc.from, mc.to)) {
12724850
NH
6791 mc.precharge -= HPAGE_PMD_NR;
6792 mc.moved_charge += HPAGE_PMD_NR;
6793 }
6794 putback_lru_page(page);
6795 }
6796 put_page(page);
6797 }
6798 spin_unlock(&vma->vm_mm->page_table_lock);
1a5a9906 6799 return 0;
12724850
NH
6800 }
6801
45f83cef
AA
6802 if (pmd_trans_unstable(pmd))
6803 return 0;
4ffef5fe
DN
6804retry:
6805 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6806 for (; addr != end; addr += PAGE_SIZE) {
6807 pte_t ptent = *(pte++);
02491447 6808 swp_entry_t ent;
4ffef5fe
DN
6809
6810 if (!mc.precharge)
6811 break;
6812
8d32ff84 6813 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4ffef5fe
DN
6814 case MC_TARGET_PAGE:
6815 page = target.page;
6816 if (isolate_lru_page(page))
6817 goto put;
6818 pc = lookup_page_cgroup(page);
7ec99d62 6819 if (!mem_cgroup_move_account(page, 1, pc,
2f3479b1 6820 mc.from, mc.to)) {
4ffef5fe 6821 mc.precharge--;
854ffa8d
DN
6822 /* we uncharge from mc.from later. */
6823 mc.moved_charge++;
4ffef5fe
DN
6824 }
6825 putback_lru_page(page);
8d32ff84 6826put: /* get_mctgt_type() gets the page */
4ffef5fe
DN
6827 put_page(page);
6828 break;
02491447
DN
6829 case MC_TARGET_SWAP:
6830 ent = target.ent;
e91cbb42 6831 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
02491447 6832 mc.precharge--;
483c30b5
DN
6833 /* we fixup refcnts and charges later. */
6834 mc.moved_swap++;
6835 }
02491447 6836 break;
4ffef5fe
DN
6837 default:
6838 break;
6839 }
6840 }
6841 pte_unmap_unlock(pte - 1, ptl);
6842 cond_resched();
6843
6844 if (addr != end) {
6845 /*
6846 * We have consumed all precharges we got in can_attach().
6847 * We try charge one by one, but don't do any additional
6848 * charges to mc.to if we have failed in charge once in attach()
6849 * phase.
6850 */
854ffa8d 6851 ret = mem_cgroup_do_precharge(1);
4ffef5fe
DN
6852 if (!ret)
6853 goto retry;
6854 }
6855
6856 return ret;
6857}
6858
6859static void mem_cgroup_move_charge(struct mm_struct *mm)
6860{
6861 struct vm_area_struct *vma;
6862
6863 lru_add_drain_all();
dfe076b0
DN
6864retry:
6865 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6866 /*
6867 * Someone who are holding the mmap_sem might be waiting in
6868 * waitq. So we cancel all extra charges, wake up all waiters,
6869 * and retry. Because we cancel precharges, we might not be able
6870 * to move enough charges, but moving charge is a best-effort
6871 * feature anyway, so it wouldn't be a big problem.
6872 */
6873 __mem_cgroup_clear_mc();
6874 cond_resched();
6875 goto retry;
6876 }
4ffef5fe
DN
6877 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6878 int ret;
6879 struct mm_walk mem_cgroup_move_charge_walk = {
6880 .pmd_entry = mem_cgroup_move_charge_pte_range,
6881 .mm = mm,
6882 .private = vma,
6883 };
6884 if (is_vm_hugetlb_page(vma))
6885 continue;
4ffef5fe
DN
6886 ret = walk_page_range(vma->vm_start, vma->vm_end,
6887 &mem_cgroup_move_charge_walk);
6888 if (ret)
6889 /*
6890 * means we have consumed all precharges and failed in
6891 * doing additional charge. Just abandon here.
6892 */
6893 break;
6894 }
dfe076b0 6895 up_read(&mm->mmap_sem);
7dc74be0
DN
6896}
6897
761b3ef5
LZ
6898static void mem_cgroup_move_task(struct cgroup *cont,
6899 struct cgroup_taskset *tset)
67e465a7 6900{
2f7ee569 6901 struct task_struct *p = cgroup_taskset_first(tset);
a433658c 6902 struct mm_struct *mm = get_task_mm(p);
dfe076b0 6903
dfe076b0 6904 if (mm) {
a433658c
KM
6905 if (mc.to)
6906 mem_cgroup_move_charge(mm);
dfe076b0
DN
6907 mmput(mm);
6908 }
a433658c
KM
6909 if (mc.to)
6910 mem_cgroup_clear_mc();
67e465a7 6911}
5cfb80a7 6912#else /* !CONFIG_MMU */
761b3ef5
LZ
6913static int mem_cgroup_can_attach(struct cgroup *cgroup,
6914 struct cgroup_taskset *tset)
5cfb80a7
DN
6915{
6916 return 0;
6917}
761b3ef5
LZ
6918static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6919 struct cgroup_taskset *tset)
5cfb80a7
DN
6920{
6921}
761b3ef5
LZ
6922static void mem_cgroup_move_task(struct cgroup *cont,
6923 struct cgroup_taskset *tset)
5cfb80a7
DN
6924{
6925}
6926#endif
67e465a7 6927
f00baae7
TH
6928/*
6929 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6930 * to verify sane_behavior flag on each mount attempt.
6931 */
6932static void mem_cgroup_bind(struct cgroup *root)
6933{
6934 /*
6935 * use_hierarchy is forced with sane_behavior. cgroup core
6936 * guarantees that @root doesn't have any children, so turning it
6937 * on for the root memcg is enough.
6938 */
6939 if (cgroup_sane_behavior(root))
6940 mem_cgroup_from_cont(root)->use_hierarchy = true;
6941}
6942
8cdea7c0
BS
6943struct cgroup_subsys mem_cgroup_subsys = {
6944 .name = "memory",
6945 .subsys_id = mem_cgroup_subsys_id,
92fb9748 6946 .css_alloc = mem_cgroup_css_alloc,
d142e3e6 6947 .css_online = mem_cgroup_css_online,
92fb9748
TH
6948 .css_offline = mem_cgroup_css_offline,
6949 .css_free = mem_cgroup_css_free,
7dc74be0
DN
6950 .can_attach = mem_cgroup_can_attach,
6951 .cancel_attach = mem_cgroup_cancel_attach,
67e465a7 6952 .attach = mem_cgroup_move_task,
f00baae7 6953 .bind = mem_cgroup_bind,
6bc10349 6954 .base_cftypes = mem_cgroup_files,
6d12e2d8 6955 .early_init = 0,
04046e1a 6956 .use_id = 1,
8cdea7c0 6957};
c077719b 6958
c255a458 6959#ifdef CONFIG_MEMCG_SWAP
a42c390c
MH
6960static int __init enable_swap_account(char *s)
6961{
6962 /* consider enabled if no parameter or 1 is given */
a2c8990a 6963 if (!strcmp(s, "1"))
a42c390c 6964 really_do_swap_account = 1;
a2c8990a 6965 else if (!strcmp(s, "0"))
a42c390c
MH
6966 really_do_swap_account = 0;
6967 return 1;
6968}
a2c8990a 6969__setup("swapaccount=", enable_swap_account);
c077719b 6970
2d11085e
MH
6971static void __init memsw_file_init(void)
6972{
6acc8b02
MH
6973 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6974}
6975
6976static void __init enable_swap_cgroup(void)
6977{
6978 if (!mem_cgroup_disabled() && really_do_swap_account) {
6979 do_swap_account = 1;
6980 memsw_file_init();
6981 }
2d11085e 6982}
6acc8b02 6983
2d11085e 6984#else
6acc8b02 6985static void __init enable_swap_cgroup(void)
2d11085e
MH
6986{
6987}
c077719b 6988#endif
2d11085e
MH
6989
6990/*
1081312f
MH
6991 * subsys_initcall() for memory controller.
6992 *
6993 * Some parts like hotcpu_notifier() have to be initialized from this context
6994 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6995 * everything that doesn't depend on a specific mem_cgroup structure should
6996 * be initialized from here.
2d11085e
MH
6997 */
6998static int __init mem_cgroup_init(void)
6999{
7000 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6acc8b02 7001 enable_swap_cgroup();
8787a1df 7002 mem_cgroup_soft_limit_tree_init();
e4777496 7003 memcg_stock_init();
2d11085e
MH
7004 return 0;
7005}
7006subsys_initcall(mem_cgroup_init);