]>
Commit | Line | Data |
---|---|---|
1 | /* memcontrol.c - Memory Controller | |
2 | * | |
3 | * Copyright IBM Corporation, 2007 | |
4 | * Author Balbir Singh <balbir@linux.vnet.ibm.com> | |
5 | * | |
6 | * Copyright 2007 OpenVZ SWsoft Inc | |
7 | * Author: Pavel Emelianov <xemul@openvz.org> | |
8 | * | |
9 | * Memory thresholds | |
10 | * Copyright (C) 2009 Nokia Corporation | |
11 | * Author: Kirill A. Shutemov | |
12 | * | |
13 | * Kernel Memory Controller | |
14 | * Copyright (C) 2012 Parallels Inc. and Google Inc. | |
15 | * Authors: Glauber Costa and Suleiman Souhlal | |
16 | * | |
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> | |
31 | #include <linux/mm.h> | |
32 | #include <linux/hugetlb.h> | |
33 | #include <linux/pagemap.h> | |
34 | #include <linux/smp.h> | |
35 | #include <linux/page-flags.h> | |
36 | #include <linux/backing-dev.h> | |
37 | #include <linux/bit_spinlock.h> | |
38 | #include <linux/rcupdate.h> | |
39 | #include <linux/limits.h> | |
40 | #include <linux/export.h> | |
41 | #include <linux/mutex.h> | |
42 | #include <linux/rbtree.h> | |
43 | #include <linux/slab.h> | |
44 | #include <linux/swap.h> | |
45 | #include <linux/swapops.h> | |
46 | #include <linux/spinlock.h> | |
47 | #include <linux/eventfd.h> | |
48 | #include <linux/poll.h> | |
49 | #include <linux/sort.h> | |
50 | #include <linux/fs.h> | |
51 | #include <linux/seq_file.h> | |
52 | #include <linux/vmpressure.h> | |
53 | #include <linux/mm_inline.h> | |
54 | #include <linux/page_cgroup.h> | |
55 | #include <linux/cpu.h> | |
56 | #include <linux/oom.h> | |
57 | #include <linux/lockdep.h> | |
58 | #include <linux/file.h> | |
59 | #include "internal.h" | |
60 | #include <net/sock.h> | |
61 | #include <net/ip.h> | |
62 | #include <net/tcp_memcontrol.h> | |
63 | #include "slab.h" | |
64 | ||
65 | #include <asm/uaccess.h> | |
66 | ||
67 | #include <trace/events/vmscan.h> | |
68 | ||
69 | struct cgroup_subsys memory_cgrp_subsys __read_mostly; | |
70 | EXPORT_SYMBOL(memory_cgrp_subsys); | |
71 | ||
72 | #define MEM_CGROUP_RECLAIM_RETRIES 5 | |
73 | static struct mem_cgroup *root_mem_cgroup __read_mostly; | |
74 | ||
75 | #ifdef CONFIG_MEMCG_SWAP | |
76 | /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ | |
77 | int do_swap_account __read_mostly; | |
78 | ||
79 | /* for remember boot option*/ | |
80 | #ifdef CONFIG_MEMCG_SWAP_ENABLED | |
81 | static int really_do_swap_account __initdata = 1; | |
82 | #else | |
83 | static int really_do_swap_account __initdata; | |
84 | #endif | |
85 | ||
86 | #else | |
87 | #define do_swap_account 0 | |
88 | #endif | |
89 | ||
90 | ||
91 | static const char * const mem_cgroup_stat_names[] = { | |
92 | "cache", | |
93 | "rss", | |
94 | "rss_huge", | |
95 | "mapped_file", | |
96 | "writeback", | |
97 | "swap", | |
98 | }; | |
99 | ||
100 | enum mem_cgroup_events_index { | |
101 | MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ | |
102 | MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ | |
103 | MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */ | |
104 | MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */ | |
105 | MEM_CGROUP_EVENTS_NSTATS, | |
106 | }; | |
107 | ||
108 | static const char * const mem_cgroup_events_names[] = { | |
109 | "pgpgin", | |
110 | "pgpgout", | |
111 | "pgfault", | |
112 | "pgmajfault", | |
113 | }; | |
114 | ||
115 | static const char * const mem_cgroup_lru_names[] = { | |
116 | "inactive_anon", | |
117 | "active_anon", | |
118 | "inactive_file", | |
119 | "active_file", | |
120 | "unevictable", | |
121 | }; | |
122 | ||
123 | /* | |
124 | * Per memcg event counter is incremented at every pagein/pageout. With THP, | |
125 | * it will be incremated by the number of pages. This counter is used for | |
126 | * for trigger some periodic events. This is straightforward and better | |
127 | * than using jiffies etc. to handle periodic memcg event. | |
128 | */ | |
129 | enum mem_cgroup_events_target { | |
130 | MEM_CGROUP_TARGET_THRESH, | |
131 | MEM_CGROUP_TARGET_SOFTLIMIT, | |
132 | MEM_CGROUP_TARGET_NUMAINFO, | |
133 | MEM_CGROUP_NTARGETS, | |
134 | }; | |
135 | #define THRESHOLDS_EVENTS_TARGET 128 | |
136 | #define SOFTLIMIT_EVENTS_TARGET 1024 | |
137 | #define NUMAINFO_EVENTS_TARGET 1024 | |
138 | ||
139 | struct mem_cgroup_stat_cpu { | |
140 | long count[MEM_CGROUP_STAT_NSTATS]; | |
141 | unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; | |
142 | unsigned long nr_page_events; | |
143 | unsigned long targets[MEM_CGROUP_NTARGETS]; | |
144 | }; | |
145 | ||
146 | struct mem_cgroup_reclaim_iter { | |
147 | /* | |
148 | * last scanned hierarchy member. Valid only if last_dead_count | |
149 | * matches memcg->dead_count of the hierarchy root group. | |
150 | */ | |
151 | struct mem_cgroup *last_visited; | |
152 | int last_dead_count; | |
153 | ||
154 | /* scan generation, increased every round-trip */ | |
155 | unsigned int generation; | |
156 | }; | |
157 | ||
158 | /* | |
159 | * per-zone information in memory controller. | |
160 | */ | |
161 | struct mem_cgroup_per_zone { | |
162 | struct lruvec lruvec; | |
163 | unsigned long lru_size[NR_LRU_LISTS]; | |
164 | ||
165 | struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1]; | |
166 | ||
167 | struct rb_node tree_node; /* RB tree node */ | |
168 | unsigned long long usage_in_excess;/* Set to the value by which */ | |
169 | /* the soft limit is exceeded*/ | |
170 | bool on_tree; | |
171 | struct mem_cgroup *memcg; /* Back pointer, we cannot */ | |
172 | /* use container_of */ | |
173 | }; | |
174 | ||
175 | struct mem_cgroup_per_node { | |
176 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; | |
177 | }; | |
178 | ||
179 | /* | |
180 | * Cgroups above their limits are maintained in a RB-Tree, independent of | |
181 | * their hierarchy representation | |
182 | */ | |
183 | ||
184 | struct mem_cgroup_tree_per_zone { | |
185 | struct rb_root rb_root; | |
186 | spinlock_t lock; | |
187 | }; | |
188 | ||
189 | struct mem_cgroup_tree_per_node { | |
190 | struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; | |
191 | }; | |
192 | ||
193 | struct mem_cgroup_tree { | |
194 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; | |
195 | }; | |
196 | ||
197 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; | |
198 | ||
199 | struct mem_cgroup_threshold { | |
200 | struct eventfd_ctx *eventfd; | |
201 | u64 threshold; | |
202 | }; | |
203 | ||
204 | /* For threshold */ | |
205 | struct mem_cgroup_threshold_ary { | |
206 | /* An array index points to threshold just below or equal to usage. */ | |
207 | int current_threshold; | |
208 | /* Size of entries[] */ | |
209 | unsigned int size; | |
210 | /* Array of thresholds */ | |
211 | struct mem_cgroup_threshold entries[0]; | |
212 | }; | |
213 | ||
214 | struct mem_cgroup_thresholds { | |
215 | /* Primary thresholds array */ | |
216 | struct mem_cgroup_threshold_ary *primary; | |
217 | /* | |
218 | * Spare threshold array. | |
219 | * This is needed to make mem_cgroup_unregister_event() "never fail". | |
220 | * It must be able to store at least primary->size - 1 entries. | |
221 | */ | |
222 | struct mem_cgroup_threshold_ary *spare; | |
223 | }; | |
224 | ||
225 | /* for OOM */ | |
226 | struct mem_cgroup_eventfd_list { | |
227 | struct list_head list; | |
228 | struct eventfd_ctx *eventfd; | |
229 | }; | |
230 | ||
231 | /* | |
232 | * cgroup_event represents events which userspace want to receive. | |
233 | */ | |
234 | struct mem_cgroup_event { | |
235 | /* | |
236 | * memcg which the event belongs to. | |
237 | */ | |
238 | struct mem_cgroup *memcg; | |
239 | /* | |
240 | * eventfd to signal userspace about the event. | |
241 | */ | |
242 | struct eventfd_ctx *eventfd; | |
243 | /* | |
244 | * Each of these stored in a list by the cgroup. | |
245 | */ | |
246 | struct list_head list; | |
247 | /* | |
248 | * register_event() callback will be used to add new userspace | |
249 | * waiter for changes related to this event. Use eventfd_signal() | |
250 | * on eventfd to send notification to userspace. | |
251 | */ | |
252 | int (*register_event)(struct mem_cgroup *memcg, | |
253 | struct eventfd_ctx *eventfd, const char *args); | |
254 | /* | |
255 | * unregister_event() callback will be called when userspace closes | |
256 | * the eventfd or on cgroup removing. This callback must be set, | |
257 | * if you want provide notification functionality. | |
258 | */ | |
259 | void (*unregister_event)(struct mem_cgroup *memcg, | |
260 | struct eventfd_ctx *eventfd); | |
261 | /* | |
262 | * All fields below needed to unregister event when | |
263 | * userspace closes eventfd. | |
264 | */ | |
265 | poll_table pt; | |
266 | wait_queue_head_t *wqh; | |
267 | wait_queue_t wait; | |
268 | struct work_struct remove; | |
269 | }; | |
270 | ||
271 | static void mem_cgroup_threshold(struct mem_cgroup *memcg); | |
272 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg); | |
273 | ||
274 | /* | |
275 | * The memory controller data structure. The memory controller controls both | |
276 | * page cache and RSS per cgroup. We would eventually like to provide | |
277 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, | |
278 | * to help the administrator determine what knobs to tune. | |
279 | * | |
280 | * TODO: Add a water mark for the memory controller. Reclaim will begin when | |
281 | * we hit the water mark. May be even add a low water mark, such that | |
282 | * no reclaim occurs from a cgroup at it's low water mark, this is | |
283 | * a feature that will be implemented much later in the future. | |
284 | */ | |
285 | struct mem_cgroup { | |
286 | struct cgroup_subsys_state css; | |
287 | /* | |
288 | * the counter to account for memory usage | |
289 | */ | |
290 | struct res_counter res; | |
291 | ||
292 | /* vmpressure notifications */ | |
293 | struct vmpressure vmpressure; | |
294 | ||
295 | /* | |
296 | * the counter to account for mem+swap usage. | |
297 | */ | |
298 | struct res_counter memsw; | |
299 | ||
300 | /* | |
301 | * the counter to account for kernel memory usage. | |
302 | */ | |
303 | struct res_counter kmem; | |
304 | /* | |
305 | * Should the accounting and control be hierarchical, per subtree? | |
306 | */ | |
307 | bool use_hierarchy; | |
308 | unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */ | |
309 | ||
310 | bool oom_lock; | |
311 | atomic_t under_oom; | |
312 | atomic_t oom_wakeups; | |
313 | ||
314 | int swappiness; | |
315 | /* OOM-Killer disable */ | |
316 | int oom_kill_disable; | |
317 | ||
318 | /* set when res.limit == memsw.limit */ | |
319 | bool memsw_is_minimum; | |
320 | ||
321 | /* protect arrays of thresholds */ | |
322 | struct mutex thresholds_lock; | |
323 | ||
324 | /* thresholds for memory usage. RCU-protected */ | |
325 | struct mem_cgroup_thresholds thresholds; | |
326 | ||
327 | /* thresholds for mem+swap usage. RCU-protected */ | |
328 | struct mem_cgroup_thresholds memsw_thresholds; | |
329 | ||
330 | /* For oom notifier event fd */ | |
331 | struct list_head oom_notify; | |
332 | ||
333 | /* | |
334 | * Should we move charges of a task when a task is moved into this | |
335 | * mem_cgroup ? And what type of charges should we move ? | |
336 | */ | |
337 | unsigned long move_charge_at_immigrate; | |
338 | /* | |
339 | * set > 0 if pages under this cgroup are moving to other cgroup. | |
340 | */ | |
341 | atomic_t moving_account; | |
342 | /* taken only while moving_account > 0 */ | |
343 | spinlock_t move_lock; | |
344 | /* | |
345 | * percpu counter. | |
346 | */ | |
347 | struct mem_cgroup_stat_cpu __percpu *stat; | |
348 | /* | |
349 | * used when a cpu is offlined or other synchronizations | |
350 | * See mem_cgroup_read_stat(). | |
351 | */ | |
352 | struct mem_cgroup_stat_cpu nocpu_base; | |
353 | spinlock_t pcp_counter_lock; | |
354 | ||
355 | atomic_t dead_count; | |
356 | #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET) | |
357 | struct cg_proto tcp_mem; | |
358 | #endif | |
359 | #if defined(CONFIG_MEMCG_KMEM) | |
360 | /* analogous to slab_common's slab_caches list, but per-memcg; | |
361 | * protected by memcg_slab_mutex */ | |
362 | struct list_head memcg_slab_caches; | |
363 | /* Index in the kmem_cache->memcg_params->memcg_caches array */ | |
364 | int kmemcg_id; | |
365 | #endif | |
366 | ||
367 | int last_scanned_node; | |
368 | #if MAX_NUMNODES > 1 | |
369 | nodemask_t scan_nodes; | |
370 | atomic_t numainfo_events; | |
371 | atomic_t numainfo_updating; | |
372 | #endif | |
373 | ||
374 | /* List of events which userspace want to receive */ | |
375 | struct list_head event_list; | |
376 | spinlock_t event_list_lock; | |
377 | ||
378 | struct mem_cgroup_per_node *nodeinfo[0]; | |
379 | /* WARNING: nodeinfo must be the last member here */ | |
380 | }; | |
381 | ||
382 | /* internal only representation about the status of kmem accounting. */ | |
383 | enum { | |
384 | KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */ | |
385 | KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */ | |
386 | }; | |
387 | ||
388 | #ifdef CONFIG_MEMCG_KMEM | |
389 | static inline void memcg_kmem_set_active(struct mem_cgroup *memcg) | |
390 | { | |
391 | set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags); | |
392 | } | |
393 | ||
394 | static bool memcg_kmem_is_active(struct mem_cgroup *memcg) | |
395 | { | |
396 | return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags); | |
397 | } | |
398 | ||
399 | static void memcg_kmem_mark_dead(struct mem_cgroup *memcg) | |
400 | { | |
401 | /* | |
402 | * Our caller must use css_get() first, because memcg_uncharge_kmem() | |
403 | * will call css_put() if it sees the memcg is dead. | |
404 | */ | |
405 | smp_wmb(); | |
406 | if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags)) | |
407 | set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags); | |
408 | } | |
409 | ||
410 | static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg) | |
411 | { | |
412 | return test_and_clear_bit(KMEM_ACCOUNTED_DEAD, | |
413 | &memcg->kmem_account_flags); | |
414 | } | |
415 | #endif | |
416 | ||
417 | /* Stuffs for move charges at task migration. */ | |
418 | /* | |
419 | * Types of charges to be moved. "move_charge_at_immitgrate" and | |
420 | * "immigrate_flags" are treated as a left-shifted bitmap of these types. | |
421 | */ | |
422 | enum move_type { | |
423 | MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ | |
424 | MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ | |
425 | NR_MOVE_TYPE, | |
426 | }; | |
427 | ||
428 | /* "mc" and its members are protected by cgroup_mutex */ | |
429 | static struct move_charge_struct { | |
430 | spinlock_t lock; /* for from, to */ | |
431 | struct mem_cgroup *from; | |
432 | struct mem_cgroup *to; | |
433 | unsigned long immigrate_flags; | |
434 | unsigned long precharge; | |
435 | unsigned long moved_charge; | |
436 | unsigned long moved_swap; | |
437 | struct task_struct *moving_task; /* a task moving charges */ | |
438 | wait_queue_head_t waitq; /* a waitq for other context */ | |
439 | } mc = { | |
440 | .lock = __SPIN_LOCK_UNLOCKED(mc.lock), | |
441 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), | |
442 | }; | |
443 | ||
444 | static bool move_anon(void) | |
445 | { | |
446 | return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags); | |
447 | } | |
448 | ||
449 | static bool move_file(void) | |
450 | { | |
451 | return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags); | |
452 | } | |
453 | ||
454 | /* | |
455 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft | |
456 | * limit reclaim to prevent infinite loops, if they ever occur. | |
457 | */ | |
458 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100 | |
459 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2 | |
460 | ||
461 | enum charge_type { | |
462 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, | |
463 | MEM_CGROUP_CHARGE_TYPE_ANON, | |
464 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ | |
465 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ | |
466 | NR_CHARGE_TYPE, | |
467 | }; | |
468 | ||
469 | /* for encoding cft->private value on file */ | |
470 | enum res_type { | |
471 | _MEM, | |
472 | _MEMSWAP, | |
473 | _OOM_TYPE, | |
474 | _KMEM, | |
475 | }; | |
476 | ||
477 | #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val)) | |
478 | #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff) | |
479 | #define MEMFILE_ATTR(val) ((val) & 0xffff) | |
480 | /* Used for OOM nofiier */ | |
481 | #define OOM_CONTROL (0) | |
482 | ||
483 | /* | |
484 | * Reclaim flags for mem_cgroup_hierarchical_reclaim | |
485 | */ | |
486 | #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 | |
487 | #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) | |
488 | #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 | |
489 | #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) | |
490 | ||
491 | /* | |
492 | * The memcg_create_mutex will be held whenever a new cgroup is created. | |
493 | * As a consequence, any change that needs to protect against new child cgroups | |
494 | * appearing has to hold it as well. | |
495 | */ | |
496 | static DEFINE_MUTEX(memcg_create_mutex); | |
497 | ||
498 | struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s) | |
499 | { | |
500 | return s ? container_of(s, struct mem_cgroup, css) : NULL; | |
501 | } | |
502 | ||
503 | /* Some nice accessors for the vmpressure. */ | |
504 | struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg) | |
505 | { | |
506 | if (!memcg) | |
507 | memcg = root_mem_cgroup; | |
508 | return &memcg->vmpressure; | |
509 | } | |
510 | ||
511 | struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr) | |
512 | { | |
513 | return &container_of(vmpr, struct mem_cgroup, vmpressure)->css; | |
514 | } | |
515 | ||
516 | static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg) | |
517 | { | |
518 | return (memcg == root_mem_cgroup); | |
519 | } | |
520 | ||
521 | /* | |
522 | * We restrict the id in the range of [1, 65535], so it can fit into | |
523 | * an unsigned short. | |
524 | */ | |
525 | #define MEM_CGROUP_ID_MAX USHRT_MAX | |
526 | ||
527 | static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg) | |
528 | { | |
529 | /* | |
530 | * The ID of the root cgroup is 0, but memcg treat 0 as an | |
531 | * invalid ID, so we return (cgroup_id + 1). | |
532 | */ | |
533 | return memcg->css.cgroup->id + 1; | |
534 | } | |
535 | ||
536 | static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id) | |
537 | { | |
538 | struct cgroup_subsys_state *css; | |
539 | ||
540 | css = css_from_id(id - 1, &memory_cgrp_subsys); | |
541 | return mem_cgroup_from_css(css); | |
542 | } | |
543 | ||
544 | /* Writing them here to avoid exposing memcg's inner layout */ | |
545 | #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM) | |
546 | ||
547 | void sock_update_memcg(struct sock *sk) | |
548 | { | |
549 | if (mem_cgroup_sockets_enabled) { | |
550 | struct mem_cgroup *memcg; | |
551 | struct cg_proto *cg_proto; | |
552 | ||
553 | BUG_ON(!sk->sk_prot->proto_cgroup); | |
554 | ||
555 | /* Socket cloning can throw us here with sk_cgrp already | |
556 | * filled. It won't however, necessarily happen from | |
557 | * process context. So the test for root memcg given | |
558 | * the current task's memcg won't help us in this case. | |
559 | * | |
560 | * Respecting the original socket's memcg is a better | |
561 | * decision in this case. | |
562 | */ | |
563 | if (sk->sk_cgrp) { | |
564 | BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg)); | |
565 | css_get(&sk->sk_cgrp->memcg->css); | |
566 | return; | |
567 | } | |
568 | ||
569 | rcu_read_lock(); | |
570 | memcg = mem_cgroup_from_task(current); | |
571 | cg_proto = sk->sk_prot->proto_cgroup(memcg); | |
572 | if (!mem_cgroup_is_root(memcg) && | |
573 | memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) { | |
574 | sk->sk_cgrp = cg_proto; | |
575 | } | |
576 | rcu_read_unlock(); | |
577 | } | |
578 | } | |
579 | EXPORT_SYMBOL(sock_update_memcg); | |
580 | ||
581 | void sock_release_memcg(struct sock *sk) | |
582 | { | |
583 | if (mem_cgroup_sockets_enabled && sk->sk_cgrp) { | |
584 | struct mem_cgroup *memcg; | |
585 | WARN_ON(!sk->sk_cgrp->memcg); | |
586 | memcg = sk->sk_cgrp->memcg; | |
587 | css_put(&sk->sk_cgrp->memcg->css); | |
588 | } | |
589 | } | |
590 | ||
591 | struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg) | |
592 | { | |
593 | if (!memcg || mem_cgroup_is_root(memcg)) | |
594 | return NULL; | |
595 | ||
596 | return &memcg->tcp_mem; | |
597 | } | |
598 | EXPORT_SYMBOL(tcp_proto_cgroup); | |
599 | ||
600 | static void disarm_sock_keys(struct mem_cgroup *memcg) | |
601 | { | |
602 | if (!memcg_proto_activated(&memcg->tcp_mem)) | |
603 | return; | |
604 | static_key_slow_dec(&memcg_socket_limit_enabled); | |
605 | } | |
606 | #else | |
607 | static void disarm_sock_keys(struct mem_cgroup *memcg) | |
608 | { | |
609 | } | |
610 | #endif | |
611 | ||
612 | #ifdef CONFIG_MEMCG_KMEM | |
613 | /* | |
614 | * This will be the memcg's index in each cache's ->memcg_params->memcg_caches. | |
615 | * The main reason for not using cgroup id for this: | |
616 | * this works better in sparse environments, where we have a lot of memcgs, | |
617 | * but only a few kmem-limited. Or also, if we have, for instance, 200 | |
618 | * memcgs, and none but the 200th is kmem-limited, we'd have to have a | |
619 | * 200 entry array for that. | |
620 | * | |
621 | * The current size of the caches array is stored in | |
622 | * memcg_limited_groups_array_size. It will double each time we have to | |
623 | * increase it. | |
624 | */ | |
625 | static DEFINE_IDA(kmem_limited_groups); | |
626 | int memcg_limited_groups_array_size; | |
627 | ||
628 | /* | |
629 | * MIN_SIZE is different than 1, because we would like to avoid going through | |
630 | * the alloc/free process all the time. In a small machine, 4 kmem-limited | |
631 | * cgroups is a reasonable guess. In the future, it could be a parameter or | |
632 | * tunable, but that is strictly not necessary. | |
633 | * | |
634 | * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get | |
635 | * this constant directly from cgroup, but it is understandable that this is | |
636 | * better kept as an internal representation in cgroup.c. In any case, the | |
637 | * cgrp_id space is not getting any smaller, and we don't have to necessarily | |
638 | * increase ours as well if it increases. | |
639 | */ | |
640 | #define MEMCG_CACHES_MIN_SIZE 4 | |
641 | #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX | |
642 | ||
643 | /* | |
644 | * A lot of the calls to the cache allocation functions are expected to be | |
645 | * inlined by the compiler. Since the calls to memcg_kmem_get_cache are | |
646 | * conditional to this static branch, we'll have to allow modules that does | |
647 | * kmem_cache_alloc and the such to see this symbol as well | |
648 | */ | |
649 | struct static_key memcg_kmem_enabled_key; | |
650 | EXPORT_SYMBOL(memcg_kmem_enabled_key); | |
651 | ||
652 | static void disarm_kmem_keys(struct mem_cgroup *memcg) | |
653 | { | |
654 | if (memcg_kmem_is_active(memcg)) { | |
655 | static_key_slow_dec(&memcg_kmem_enabled_key); | |
656 | ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id); | |
657 | } | |
658 | /* | |
659 | * This check can't live in kmem destruction function, | |
660 | * since the charges will outlive the cgroup | |
661 | */ | |
662 | WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0); | |
663 | } | |
664 | #else | |
665 | static void disarm_kmem_keys(struct mem_cgroup *memcg) | |
666 | { | |
667 | } | |
668 | #endif /* CONFIG_MEMCG_KMEM */ | |
669 | ||
670 | static void disarm_static_keys(struct mem_cgroup *memcg) | |
671 | { | |
672 | disarm_sock_keys(memcg); | |
673 | disarm_kmem_keys(memcg); | |
674 | } | |
675 | ||
676 | static void drain_all_stock_async(struct mem_cgroup *memcg); | |
677 | ||
678 | static struct mem_cgroup_per_zone * | |
679 | mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid) | |
680 | { | |
681 | VM_BUG_ON((unsigned)nid >= nr_node_ids); | |
682 | return &memcg->nodeinfo[nid]->zoneinfo[zid]; | |
683 | } | |
684 | ||
685 | struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg) | |
686 | { | |
687 | return &memcg->css; | |
688 | } | |
689 | ||
690 | static struct mem_cgroup_per_zone * | |
691 | page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page) | |
692 | { | |
693 | int nid = page_to_nid(page); | |
694 | int zid = page_zonenum(page); | |
695 | ||
696 | return mem_cgroup_zoneinfo(memcg, nid, zid); | |
697 | } | |
698 | ||
699 | static struct mem_cgroup_tree_per_zone * | |
700 | soft_limit_tree_node_zone(int nid, int zid) | |
701 | { | |
702 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
703 | } | |
704 | ||
705 | static struct mem_cgroup_tree_per_zone * | |
706 | soft_limit_tree_from_page(struct page *page) | |
707 | { | |
708 | int nid = page_to_nid(page); | |
709 | int zid = page_zonenum(page); | |
710 | ||
711 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
712 | } | |
713 | ||
714 | static void | |
715 | __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg, | |
716 | struct mem_cgroup_per_zone *mz, | |
717 | struct mem_cgroup_tree_per_zone *mctz, | |
718 | unsigned long long new_usage_in_excess) | |
719 | { | |
720 | struct rb_node **p = &mctz->rb_root.rb_node; | |
721 | struct rb_node *parent = NULL; | |
722 | struct mem_cgroup_per_zone *mz_node; | |
723 | ||
724 | if (mz->on_tree) | |
725 | return; | |
726 | ||
727 | mz->usage_in_excess = new_usage_in_excess; | |
728 | if (!mz->usage_in_excess) | |
729 | return; | |
730 | while (*p) { | |
731 | parent = *p; | |
732 | mz_node = rb_entry(parent, struct mem_cgroup_per_zone, | |
733 | tree_node); | |
734 | if (mz->usage_in_excess < mz_node->usage_in_excess) | |
735 | p = &(*p)->rb_left; | |
736 | /* | |
737 | * We can't avoid mem cgroups that are over their soft | |
738 | * limit by the same amount | |
739 | */ | |
740 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) | |
741 | p = &(*p)->rb_right; | |
742 | } | |
743 | rb_link_node(&mz->tree_node, parent, p); | |
744 | rb_insert_color(&mz->tree_node, &mctz->rb_root); | |
745 | mz->on_tree = true; | |
746 | } | |
747 | ||
748 | static void | |
749 | __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, | |
750 | struct mem_cgroup_per_zone *mz, | |
751 | struct mem_cgroup_tree_per_zone *mctz) | |
752 | { | |
753 | if (!mz->on_tree) | |
754 | return; | |
755 | rb_erase(&mz->tree_node, &mctz->rb_root); | |
756 | mz->on_tree = false; | |
757 | } | |
758 | ||
759 | static void | |
760 | mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, | |
761 | struct mem_cgroup_per_zone *mz, | |
762 | struct mem_cgroup_tree_per_zone *mctz) | |
763 | { | |
764 | spin_lock(&mctz->lock); | |
765 | __mem_cgroup_remove_exceeded(memcg, mz, mctz); | |
766 | spin_unlock(&mctz->lock); | |
767 | } | |
768 | ||
769 | ||
770 | static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page) | |
771 | { | |
772 | unsigned long long excess; | |
773 | struct mem_cgroup_per_zone *mz; | |
774 | struct mem_cgroup_tree_per_zone *mctz; | |
775 | int nid = page_to_nid(page); | |
776 | int zid = page_zonenum(page); | |
777 | mctz = soft_limit_tree_from_page(page); | |
778 | ||
779 | /* | |
780 | * Necessary to update all ancestors when hierarchy is used. | |
781 | * because their event counter is not touched. | |
782 | */ | |
783 | for (; memcg; memcg = parent_mem_cgroup(memcg)) { | |
784 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
785 | excess = res_counter_soft_limit_excess(&memcg->res); | |
786 | /* | |
787 | * We have to update the tree if mz is on RB-tree or | |
788 | * mem is over its softlimit. | |
789 | */ | |
790 | if (excess || mz->on_tree) { | |
791 | spin_lock(&mctz->lock); | |
792 | /* if on-tree, remove it */ | |
793 | if (mz->on_tree) | |
794 | __mem_cgroup_remove_exceeded(memcg, mz, mctz); | |
795 | /* | |
796 | * Insert again. mz->usage_in_excess will be updated. | |
797 | * If excess is 0, no tree ops. | |
798 | */ | |
799 | __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess); | |
800 | spin_unlock(&mctz->lock); | |
801 | } | |
802 | } | |
803 | } | |
804 | ||
805 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg) | |
806 | { | |
807 | int node, zone; | |
808 | struct mem_cgroup_per_zone *mz; | |
809 | struct mem_cgroup_tree_per_zone *mctz; | |
810 | ||
811 | for_each_node(node) { | |
812 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
813 | mz = mem_cgroup_zoneinfo(memcg, node, zone); | |
814 | mctz = soft_limit_tree_node_zone(node, zone); | |
815 | mem_cgroup_remove_exceeded(memcg, mz, mctz); | |
816 | } | |
817 | } | |
818 | } | |
819 | ||
820 | static struct mem_cgroup_per_zone * | |
821 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
822 | { | |
823 | struct rb_node *rightmost = NULL; | |
824 | struct mem_cgroup_per_zone *mz; | |
825 | ||
826 | retry: | |
827 | mz = NULL; | |
828 | rightmost = rb_last(&mctz->rb_root); | |
829 | if (!rightmost) | |
830 | goto done; /* Nothing to reclaim from */ | |
831 | ||
832 | mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); | |
833 | /* | |
834 | * Remove the node now but someone else can add it back, | |
835 | * we will to add it back at the end of reclaim to its correct | |
836 | * position in the tree. | |
837 | */ | |
838 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); | |
839 | if (!res_counter_soft_limit_excess(&mz->memcg->res) || | |
840 | !css_tryget(&mz->memcg->css)) | |
841 | goto retry; | |
842 | done: | |
843 | return mz; | |
844 | } | |
845 | ||
846 | static struct mem_cgroup_per_zone * | |
847 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
848 | { | |
849 | struct mem_cgroup_per_zone *mz; | |
850 | ||
851 | spin_lock(&mctz->lock); | |
852 | mz = __mem_cgroup_largest_soft_limit_node(mctz); | |
853 | spin_unlock(&mctz->lock); | |
854 | return mz; | |
855 | } | |
856 | ||
857 | /* | |
858 | * Implementation Note: reading percpu statistics for memcg. | |
859 | * | |
860 | * Both of vmstat[] and percpu_counter has threshold and do periodic | |
861 | * synchronization to implement "quick" read. There are trade-off between | |
862 | * reading cost and precision of value. Then, we may have a chance to implement | |
863 | * a periodic synchronizion of counter in memcg's counter. | |
864 | * | |
865 | * But this _read() function is used for user interface now. The user accounts | |
866 | * memory usage by memory cgroup and he _always_ requires exact value because | |
867 | * he accounts memory. Even if we provide quick-and-fuzzy read, we always | |
868 | * have to visit all online cpus and make sum. So, for now, unnecessary | |
869 | * synchronization is not implemented. (just implemented for cpu hotplug) | |
870 | * | |
871 | * If there are kernel internal actions which can make use of some not-exact | |
872 | * value, and reading all cpu value can be performance bottleneck in some | |
873 | * common workload, threashold and synchonization as vmstat[] should be | |
874 | * implemented. | |
875 | */ | |
876 | static long mem_cgroup_read_stat(struct mem_cgroup *memcg, | |
877 | enum mem_cgroup_stat_index idx) | |
878 | { | |
879 | long val = 0; | |
880 | int cpu; | |
881 | ||
882 | get_online_cpus(); | |
883 | for_each_online_cpu(cpu) | |
884 | val += per_cpu(memcg->stat->count[idx], cpu); | |
885 | #ifdef CONFIG_HOTPLUG_CPU | |
886 | spin_lock(&memcg->pcp_counter_lock); | |
887 | val += memcg->nocpu_base.count[idx]; | |
888 | spin_unlock(&memcg->pcp_counter_lock); | |
889 | #endif | |
890 | put_online_cpus(); | |
891 | return val; | |
892 | } | |
893 | ||
894 | static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg, | |
895 | bool charge) | |
896 | { | |
897 | int val = (charge) ? 1 : -1; | |
898 | this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val); | |
899 | } | |
900 | ||
901 | static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg, | |
902 | enum mem_cgroup_events_index idx) | |
903 | { | |
904 | unsigned long val = 0; | |
905 | int cpu; | |
906 | ||
907 | get_online_cpus(); | |
908 | for_each_online_cpu(cpu) | |
909 | val += per_cpu(memcg->stat->events[idx], cpu); | |
910 | #ifdef CONFIG_HOTPLUG_CPU | |
911 | spin_lock(&memcg->pcp_counter_lock); | |
912 | val += memcg->nocpu_base.events[idx]; | |
913 | spin_unlock(&memcg->pcp_counter_lock); | |
914 | #endif | |
915 | put_online_cpus(); | |
916 | return val; | |
917 | } | |
918 | ||
919 | static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, | |
920 | struct page *page, | |
921 | bool anon, int nr_pages) | |
922 | { | |
923 | /* | |
924 | * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is | |
925 | * counted as CACHE even if it's on ANON LRU. | |
926 | */ | |
927 | if (anon) | |
928 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS], | |
929 | nr_pages); | |
930 | else | |
931 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE], | |
932 | nr_pages); | |
933 | ||
934 | if (PageTransHuge(page)) | |
935 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], | |
936 | nr_pages); | |
937 | ||
938 | /* pagein of a big page is an event. So, ignore page size */ | |
939 | if (nr_pages > 0) | |
940 | __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); | |
941 | else { | |
942 | __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); | |
943 | nr_pages = -nr_pages; /* for event */ | |
944 | } | |
945 | ||
946 | __this_cpu_add(memcg->stat->nr_page_events, nr_pages); | |
947 | } | |
948 | ||
949 | unsigned long | |
950 | mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru) | |
951 | { | |
952 | struct mem_cgroup_per_zone *mz; | |
953 | ||
954 | mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec); | |
955 | return mz->lru_size[lru]; | |
956 | } | |
957 | ||
958 | static unsigned long | |
959 | mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid, | |
960 | unsigned int lru_mask) | |
961 | { | |
962 | struct mem_cgroup_per_zone *mz; | |
963 | enum lru_list lru; | |
964 | unsigned long ret = 0; | |
965 | ||
966 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
967 | ||
968 | for_each_lru(lru) { | |
969 | if (BIT(lru) & lru_mask) | |
970 | ret += mz->lru_size[lru]; | |
971 | } | |
972 | return ret; | |
973 | } | |
974 | ||
975 | static unsigned long | |
976 | mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, | |
977 | int nid, unsigned int lru_mask) | |
978 | { | |
979 | u64 total = 0; | |
980 | int zid; | |
981 | ||
982 | for (zid = 0; zid < MAX_NR_ZONES; zid++) | |
983 | total += mem_cgroup_zone_nr_lru_pages(memcg, | |
984 | nid, zid, lru_mask); | |
985 | ||
986 | return total; | |
987 | } | |
988 | ||
989 | static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, | |
990 | unsigned int lru_mask) | |
991 | { | |
992 | int nid; | |
993 | u64 total = 0; | |
994 | ||
995 | for_each_node_state(nid, N_MEMORY) | |
996 | total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask); | |
997 | return total; | |
998 | } | |
999 | ||
1000 | static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, | |
1001 | enum mem_cgroup_events_target target) | |
1002 | { | |
1003 | unsigned long val, next; | |
1004 | ||
1005 | val = __this_cpu_read(memcg->stat->nr_page_events); | |
1006 | next = __this_cpu_read(memcg->stat->targets[target]); | |
1007 | /* from time_after() in jiffies.h */ | |
1008 | if ((long)next - (long)val < 0) { | |
1009 | switch (target) { | |
1010 | case MEM_CGROUP_TARGET_THRESH: | |
1011 | next = val + THRESHOLDS_EVENTS_TARGET; | |
1012 | break; | |
1013 | case MEM_CGROUP_TARGET_SOFTLIMIT: | |
1014 | next = val + SOFTLIMIT_EVENTS_TARGET; | |
1015 | break; | |
1016 | case MEM_CGROUP_TARGET_NUMAINFO: | |
1017 | next = val + NUMAINFO_EVENTS_TARGET; | |
1018 | break; | |
1019 | default: | |
1020 | break; | |
1021 | } | |
1022 | __this_cpu_write(memcg->stat->targets[target], next); | |
1023 | return true; | |
1024 | } | |
1025 | return false; | |
1026 | } | |
1027 | ||
1028 | /* | |
1029 | * Check events in order. | |
1030 | * | |
1031 | */ | |
1032 | static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) | |
1033 | { | |
1034 | preempt_disable(); | |
1035 | /* threshold event is triggered in finer grain than soft limit */ | |
1036 | if (unlikely(mem_cgroup_event_ratelimit(memcg, | |
1037 | MEM_CGROUP_TARGET_THRESH))) { | |
1038 | bool do_softlimit; | |
1039 | bool do_numainfo __maybe_unused; | |
1040 | ||
1041 | do_softlimit = mem_cgroup_event_ratelimit(memcg, | |
1042 | MEM_CGROUP_TARGET_SOFTLIMIT); | |
1043 | #if MAX_NUMNODES > 1 | |
1044 | do_numainfo = mem_cgroup_event_ratelimit(memcg, | |
1045 | MEM_CGROUP_TARGET_NUMAINFO); | |
1046 | #endif | |
1047 | preempt_enable(); | |
1048 | ||
1049 | mem_cgroup_threshold(memcg); | |
1050 | if (unlikely(do_softlimit)) | |
1051 | mem_cgroup_update_tree(memcg, page); | |
1052 | #if MAX_NUMNODES > 1 | |
1053 | if (unlikely(do_numainfo)) | |
1054 | atomic_inc(&memcg->numainfo_events); | |
1055 | #endif | |
1056 | } else | |
1057 | preempt_enable(); | |
1058 | } | |
1059 | ||
1060 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) | |
1061 | { | |
1062 | /* | |
1063 | * mm_update_next_owner() may clear mm->owner to NULL | |
1064 | * if it races with swapoff, page migration, etc. | |
1065 | * So this can be called with p == NULL. | |
1066 | */ | |
1067 | if (unlikely(!p)) | |
1068 | return NULL; | |
1069 | ||
1070 | return mem_cgroup_from_css(task_css(p, memory_cgrp_id)); | |
1071 | } | |
1072 | ||
1073 | static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm) | |
1074 | { | |
1075 | struct mem_cgroup *memcg = NULL; | |
1076 | ||
1077 | rcu_read_lock(); | |
1078 | do { | |
1079 | /* | |
1080 | * Page cache insertions can happen withou an | |
1081 | * actual mm context, e.g. during disk probing | |
1082 | * on boot, loopback IO, acct() writes etc. | |
1083 | */ | |
1084 | if (unlikely(!mm)) | |
1085 | memcg = root_mem_cgroup; | |
1086 | else { | |
1087 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
1088 | if (unlikely(!memcg)) | |
1089 | memcg = root_mem_cgroup; | |
1090 | } | |
1091 | } while (!css_tryget(&memcg->css)); | |
1092 | rcu_read_unlock(); | |
1093 | return memcg; | |
1094 | } | |
1095 | ||
1096 | /* | |
1097 | * Returns a next (in a pre-order walk) alive memcg (with elevated css | |
1098 | * ref. count) or NULL if the whole root's subtree has been visited. | |
1099 | * | |
1100 | * helper function to be used by mem_cgroup_iter | |
1101 | */ | |
1102 | static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root, | |
1103 | struct mem_cgroup *last_visited) | |
1104 | { | |
1105 | struct cgroup_subsys_state *prev_css, *next_css; | |
1106 | ||
1107 | prev_css = last_visited ? &last_visited->css : NULL; | |
1108 | skip_node: | |
1109 | next_css = css_next_descendant_pre(prev_css, &root->css); | |
1110 | ||
1111 | /* | |
1112 | * Even if we found a group we have to make sure it is | |
1113 | * alive. css && !memcg means that the groups should be | |
1114 | * skipped and we should continue the tree walk. | |
1115 | * last_visited css is safe to use because it is | |
1116 | * protected by css_get and the tree walk is rcu safe. | |
1117 | * | |
1118 | * We do not take a reference on the root of the tree walk | |
1119 | * because we might race with the root removal when it would | |
1120 | * be the only node in the iterated hierarchy and mem_cgroup_iter | |
1121 | * would end up in an endless loop because it expects that at | |
1122 | * least one valid node will be returned. Root cannot disappear | |
1123 | * because caller of the iterator should hold it already so | |
1124 | * skipping css reference should be safe. | |
1125 | */ | |
1126 | if (next_css) { | |
1127 | if ((next_css == &root->css) || | |
1128 | ((next_css->flags & CSS_ONLINE) && css_tryget(next_css))) | |
1129 | return mem_cgroup_from_css(next_css); | |
1130 | ||
1131 | prev_css = next_css; | |
1132 | goto skip_node; | |
1133 | } | |
1134 | ||
1135 | return NULL; | |
1136 | } | |
1137 | ||
1138 | static void mem_cgroup_iter_invalidate(struct mem_cgroup *root) | |
1139 | { | |
1140 | /* | |
1141 | * When a group in the hierarchy below root is destroyed, the | |
1142 | * hierarchy iterator can no longer be trusted since it might | |
1143 | * have pointed to the destroyed group. Invalidate it. | |
1144 | */ | |
1145 | atomic_inc(&root->dead_count); | |
1146 | } | |
1147 | ||
1148 | static struct mem_cgroup * | |
1149 | mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter, | |
1150 | struct mem_cgroup *root, | |
1151 | int *sequence) | |
1152 | { | |
1153 | struct mem_cgroup *position = NULL; | |
1154 | /* | |
1155 | * A cgroup destruction happens in two stages: offlining and | |
1156 | * release. They are separated by a RCU grace period. | |
1157 | * | |
1158 | * If the iterator is valid, we may still race with an | |
1159 | * offlining. The RCU lock ensures the object won't be | |
1160 | * released, tryget will fail if we lost the race. | |
1161 | */ | |
1162 | *sequence = atomic_read(&root->dead_count); | |
1163 | if (iter->last_dead_count == *sequence) { | |
1164 | smp_rmb(); | |
1165 | position = iter->last_visited; | |
1166 | ||
1167 | /* | |
1168 | * We cannot take a reference to root because we might race | |
1169 | * with root removal and returning NULL would end up in | |
1170 | * an endless loop on the iterator user level when root | |
1171 | * would be returned all the time. | |
1172 | */ | |
1173 | if (position && position != root && | |
1174 | !css_tryget(&position->css)) | |
1175 | position = NULL; | |
1176 | } | |
1177 | return position; | |
1178 | } | |
1179 | ||
1180 | static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter, | |
1181 | struct mem_cgroup *last_visited, | |
1182 | struct mem_cgroup *new_position, | |
1183 | struct mem_cgroup *root, | |
1184 | int sequence) | |
1185 | { | |
1186 | /* root reference counting symmetric to mem_cgroup_iter_load */ | |
1187 | if (last_visited && last_visited != root) | |
1188 | css_put(&last_visited->css); | |
1189 | /* | |
1190 | * We store the sequence count from the time @last_visited was | |
1191 | * loaded successfully instead of rereading it here so that we | |
1192 | * don't lose destruction events in between. We could have | |
1193 | * raced with the destruction of @new_position after all. | |
1194 | */ | |
1195 | iter->last_visited = new_position; | |
1196 | smp_wmb(); | |
1197 | iter->last_dead_count = sequence; | |
1198 | } | |
1199 | ||
1200 | /** | |
1201 | * mem_cgroup_iter - iterate over memory cgroup hierarchy | |
1202 | * @root: hierarchy root | |
1203 | * @prev: previously returned memcg, NULL on first invocation | |
1204 | * @reclaim: cookie for shared reclaim walks, NULL for full walks | |
1205 | * | |
1206 | * Returns references to children of the hierarchy below @root, or | |
1207 | * @root itself, or %NULL after a full round-trip. | |
1208 | * | |
1209 | * Caller must pass the return value in @prev on subsequent | |
1210 | * invocations for reference counting, or use mem_cgroup_iter_break() | |
1211 | * to cancel a hierarchy walk before the round-trip is complete. | |
1212 | * | |
1213 | * Reclaimers can specify a zone and a priority level in @reclaim to | |
1214 | * divide up the memcgs in the hierarchy among all concurrent | |
1215 | * reclaimers operating on the same zone and priority. | |
1216 | */ | |
1217 | struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, | |
1218 | struct mem_cgroup *prev, | |
1219 | struct mem_cgroup_reclaim_cookie *reclaim) | |
1220 | { | |
1221 | struct mem_cgroup *memcg = NULL; | |
1222 | struct mem_cgroup *last_visited = NULL; | |
1223 | ||
1224 | if (mem_cgroup_disabled()) | |
1225 | return NULL; | |
1226 | ||
1227 | if (!root) | |
1228 | root = root_mem_cgroup; | |
1229 | ||
1230 | if (prev && !reclaim) | |
1231 | last_visited = prev; | |
1232 | ||
1233 | if (!root->use_hierarchy && root != root_mem_cgroup) { | |
1234 | if (prev) | |
1235 | goto out_css_put; | |
1236 | return root; | |
1237 | } | |
1238 | ||
1239 | rcu_read_lock(); | |
1240 | while (!memcg) { | |
1241 | struct mem_cgroup_reclaim_iter *uninitialized_var(iter); | |
1242 | int uninitialized_var(seq); | |
1243 | ||
1244 | if (reclaim) { | |
1245 | int nid = zone_to_nid(reclaim->zone); | |
1246 | int zid = zone_idx(reclaim->zone); | |
1247 | struct mem_cgroup_per_zone *mz; | |
1248 | ||
1249 | mz = mem_cgroup_zoneinfo(root, nid, zid); | |
1250 | iter = &mz->reclaim_iter[reclaim->priority]; | |
1251 | if (prev && reclaim->generation != iter->generation) { | |
1252 | iter->last_visited = NULL; | |
1253 | goto out_unlock; | |
1254 | } | |
1255 | ||
1256 | last_visited = mem_cgroup_iter_load(iter, root, &seq); | |
1257 | } | |
1258 | ||
1259 | memcg = __mem_cgroup_iter_next(root, last_visited); | |
1260 | ||
1261 | if (reclaim) { | |
1262 | mem_cgroup_iter_update(iter, last_visited, memcg, root, | |
1263 | seq); | |
1264 | ||
1265 | if (!memcg) | |
1266 | iter->generation++; | |
1267 | else if (!prev && memcg) | |
1268 | reclaim->generation = iter->generation; | |
1269 | } | |
1270 | ||
1271 | if (prev && !memcg) | |
1272 | goto out_unlock; | |
1273 | } | |
1274 | out_unlock: | |
1275 | rcu_read_unlock(); | |
1276 | out_css_put: | |
1277 | if (prev && prev != root) | |
1278 | css_put(&prev->css); | |
1279 | ||
1280 | return memcg; | |
1281 | } | |
1282 | ||
1283 | /** | |
1284 | * mem_cgroup_iter_break - abort a hierarchy walk prematurely | |
1285 | * @root: hierarchy root | |
1286 | * @prev: last visited hierarchy member as returned by mem_cgroup_iter() | |
1287 | */ | |
1288 | void mem_cgroup_iter_break(struct mem_cgroup *root, | |
1289 | struct mem_cgroup *prev) | |
1290 | { | |
1291 | if (!root) | |
1292 | root = root_mem_cgroup; | |
1293 | if (prev && prev != root) | |
1294 | css_put(&prev->css); | |
1295 | } | |
1296 | ||
1297 | /* | |
1298 | * Iteration constructs for visiting all cgroups (under a tree). If | |
1299 | * loops are exited prematurely (break), mem_cgroup_iter_break() must | |
1300 | * be used for reference counting. | |
1301 | */ | |
1302 | #define for_each_mem_cgroup_tree(iter, root) \ | |
1303 | for (iter = mem_cgroup_iter(root, NULL, NULL); \ | |
1304 | iter != NULL; \ | |
1305 | iter = mem_cgroup_iter(root, iter, NULL)) | |
1306 | ||
1307 | #define for_each_mem_cgroup(iter) \ | |
1308 | for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ | |
1309 | iter != NULL; \ | |
1310 | iter = mem_cgroup_iter(NULL, iter, NULL)) | |
1311 | ||
1312 | void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) | |
1313 | { | |
1314 | struct mem_cgroup *memcg; | |
1315 | ||
1316 | rcu_read_lock(); | |
1317 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
1318 | if (unlikely(!memcg)) | |
1319 | goto out; | |
1320 | ||
1321 | switch (idx) { | |
1322 | case PGFAULT: | |
1323 | this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]); | |
1324 | break; | |
1325 | case PGMAJFAULT: | |
1326 | this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]); | |
1327 | break; | |
1328 | default: | |
1329 | BUG(); | |
1330 | } | |
1331 | out: | |
1332 | rcu_read_unlock(); | |
1333 | } | |
1334 | EXPORT_SYMBOL(__mem_cgroup_count_vm_event); | |
1335 | ||
1336 | /** | |
1337 | * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg | |
1338 | * @zone: zone of the wanted lruvec | |
1339 | * @memcg: memcg of the wanted lruvec | |
1340 | * | |
1341 | * Returns the lru list vector holding pages for the given @zone and | |
1342 | * @mem. This can be the global zone lruvec, if the memory controller | |
1343 | * is disabled. | |
1344 | */ | |
1345 | struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone, | |
1346 | struct mem_cgroup *memcg) | |
1347 | { | |
1348 | struct mem_cgroup_per_zone *mz; | |
1349 | struct lruvec *lruvec; | |
1350 | ||
1351 | if (mem_cgroup_disabled()) { | |
1352 | lruvec = &zone->lruvec; | |
1353 | goto out; | |
1354 | } | |
1355 | ||
1356 | mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone)); | |
1357 | lruvec = &mz->lruvec; | |
1358 | out: | |
1359 | /* | |
1360 | * Since a node can be onlined after the mem_cgroup was created, | |
1361 | * we have to be prepared to initialize lruvec->zone here; | |
1362 | * and if offlined then reonlined, we need to reinitialize it. | |
1363 | */ | |
1364 | if (unlikely(lruvec->zone != zone)) | |
1365 | lruvec->zone = zone; | |
1366 | return lruvec; | |
1367 | } | |
1368 | ||
1369 | /* | |
1370 | * Following LRU functions are allowed to be used without PCG_LOCK. | |
1371 | * Operations are called by routine of global LRU independently from memcg. | |
1372 | * What we have to take care of here is validness of pc->mem_cgroup. | |
1373 | * | |
1374 | * Changes to pc->mem_cgroup happens when | |
1375 | * 1. charge | |
1376 | * 2. moving account | |
1377 | * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. | |
1378 | * It is added to LRU before charge. | |
1379 | * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. | |
1380 | * When moving account, the page is not on LRU. It's isolated. | |
1381 | */ | |
1382 | ||
1383 | /** | |
1384 | * mem_cgroup_page_lruvec - return lruvec for adding an lru page | |
1385 | * @page: the page | |
1386 | * @zone: zone of the page | |
1387 | */ | |
1388 | struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone) | |
1389 | { | |
1390 | struct mem_cgroup_per_zone *mz; | |
1391 | struct mem_cgroup *memcg; | |
1392 | struct page_cgroup *pc; | |
1393 | struct lruvec *lruvec; | |
1394 | ||
1395 | if (mem_cgroup_disabled()) { | |
1396 | lruvec = &zone->lruvec; | |
1397 | goto out; | |
1398 | } | |
1399 | ||
1400 | pc = lookup_page_cgroup(page); | |
1401 | memcg = pc->mem_cgroup; | |
1402 | ||
1403 | /* | |
1404 | * Surreptitiously switch any uncharged offlist page to root: | |
1405 | * an uncharged page off lru does nothing to secure | |
1406 | * its former mem_cgroup from sudden removal. | |
1407 | * | |
1408 | * Our caller holds lru_lock, and PageCgroupUsed is updated | |
1409 | * under page_cgroup lock: between them, they make all uses | |
1410 | * of pc->mem_cgroup safe. | |
1411 | */ | |
1412 | if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup) | |
1413 | pc->mem_cgroup = memcg = root_mem_cgroup; | |
1414 | ||
1415 | mz = page_cgroup_zoneinfo(memcg, page); | |
1416 | lruvec = &mz->lruvec; | |
1417 | out: | |
1418 | /* | |
1419 | * Since a node can be onlined after the mem_cgroup was created, | |
1420 | * we have to be prepared to initialize lruvec->zone here; | |
1421 | * and if offlined then reonlined, we need to reinitialize it. | |
1422 | */ | |
1423 | if (unlikely(lruvec->zone != zone)) | |
1424 | lruvec->zone = zone; | |
1425 | return lruvec; | |
1426 | } | |
1427 | ||
1428 | /** | |
1429 | * mem_cgroup_update_lru_size - account for adding or removing an lru page | |
1430 | * @lruvec: mem_cgroup per zone lru vector | |
1431 | * @lru: index of lru list the page is sitting on | |
1432 | * @nr_pages: positive when adding or negative when removing | |
1433 | * | |
1434 | * This function must be called when a page is added to or removed from an | |
1435 | * lru list. | |
1436 | */ | |
1437 | void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru, | |
1438 | int nr_pages) | |
1439 | { | |
1440 | struct mem_cgroup_per_zone *mz; | |
1441 | unsigned long *lru_size; | |
1442 | ||
1443 | if (mem_cgroup_disabled()) | |
1444 | return; | |
1445 | ||
1446 | mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec); | |
1447 | lru_size = mz->lru_size + lru; | |
1448 | *lru_size += nr_pages; | |
1449 | VM_BUG_ON((long)(*lru_size) < 0); | |
1450 | } | |
1451 | ||
1452 | /* | |
1453 | * Checks whether given mem is same or in the root_mem_cgroup's | |
1454 | * hierarchy subtree | |
1455 | */ | |
1456 | bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, | |
1457 | struct mem_cgroup *memcg) | |
1458 | { | |
1459 | if (root_memcg == memcg) | |
1460 | return true; | |
1461 | if (!root_memcg->use_hierarchy || !memcg) | |
1462 | return false; | |
1463 | return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup); | |
1464 | } | |
1465 | ||
1466 | static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, | |
1467 | struct mem_cgroup *memcg) | |
1468 | { | |
1469 | bool ret; | |
1470 | ||
1471 | rcu_read_lock(); | |
1472 | ret = __mem_cgroup_same_or_subtree(root_memcg, memcg); | |
1473 | rcu_read_unlock(); | |
1474 | return ret; | |
1475 | } | |
1476 | ||
1477 | bool task_in_mem_cgroup(struct task_struct *task, | |
1478 | const struct mem_cgroup *memcg) | |
1479 | { | |
1480 | struct mem_cgroup *curr = NULL; | |
1481 | struct task_struct *p; | |
1482 | bool ret; | |
1483 | ||
1484 | p = find_lock_task_mm(task); | |
1485 | if (p) { | |
1486 | curr = get_mem_cgroup_from_mm(p->mm); | |
1487 | task_unlock(p); | |
1488 | } else { | |
1489 | /* | |
1490 | * All threads may have already detached their mm's, but the oom | |
1491 | * killer still needs to detect if they have already been oom | |
1492 | * killed to prevent needlessly killing additional tasks. | |
1493 | */ | |
1494 | rcu_read_lock(); | |
1495 | curr = mem_cgroup_from_task(task); | |
1496 | if (curr) | |
1497 | css_get(&curr->css); | |
1498 | rcu_read_unlock(); | |
1499 | } | |
1500 | /* | |
1501 | * We should check use_hierarchy of "memcg" not "curr". Because checking | |
1502 | * use_hierarchy of "curr" here make this function true if hierarchy is | |
1503 | * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* | |
1504 | * hierarchy(even if use_hierarchy is disabled in "memcg"). | |
1505 | */ | |
1506 | ret = mem_cgroup_same_or_subtree(memcg, curr); | |
1507 | css_put(&curr->css); | |
1508 | return ret; | |
1509 | } | |
1510 | ||
1511 | int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec) | |
1512 | { | |
1513 | unsigned long inactive_ratio; | |
1514 | unsigned long inactive; | |
1515 | unsigned long active; | |
1516 | unsigned long gb; | |
1517 | ||
1518 | inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON); | |
1519 | active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON); | |
1520 | ||
1521 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
1522 | if (gb) | |
1523 | inactive_ratio = int_sqrt(10 * gb); | |
1524 | else | |
1525 | inactive_ratio = 1; | |
1526 | ||
1527 | return inactive * inactive_ratio < active; | |
1528 | } | |
1529 | ||
1530 | #define mem_cgroup_from_res_counter(counter, member) \ | |
1531 | container_of(counter, struct mem_cgroup, member) | |
1532 | ||
1533 | /** | |
1534 | * mem_cgroup_margin - calculate chargeable space of a memory cgroup | |
1535 | * @memcg: the memory cgroup | |
1536 | * | |
1537 | * Returns the maximum amount of memory @mem can be charged with, in | |
1538 | * pages. | |
1539 | */ | |
1540 | static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) | |
1541 | { | |
1542 | unsigned long long margin; | |
1543 | ||
1544 | margin = res_counter_margin(&memcg->res); | |
1545 | if (do_swap_account) | |
1546 | margin = min(margin, res_counter_margin(&memcg->memsw)); | |
1547 | return margin >> PAGE_SHIFT; | |
1548 | } | |
1549 | ||
1550 | int mem_cgroup_swappiness(struct mem_cgroup *memcg) | |
1551 | { | |
1552 | /* root ? */ | |
1553 | if (!css_parent(&memcg->css)) | |
1554 | return vm_swappiness; | |
1555 | ||
1556 | return memcg->swappiness; | |
1557 | } | |
1558 | ||
1559 | /* | |
1560 | * memcg->moving_account is used for checking possibility that some thread is | |
1561 | * calling move_account(). When a thread on CPU-A starts moving pages under | |
1562 | * a memcg, other threads should check memcg->moving_account under | |
1563 | * rcu_read_lock(), like this: | |
1564 | * | |
1565 | * CPU-A CPU-B | |
1566 | * rcu_read_lock() | |
1567 | * memcg->moving_account+1 if (memcg->mocing_account) | |
1568 | * take heavy locks. | |
1569 | * synchronize_rcu() update something. | |
1570 | * rcu_read_unlock() | |
1571 | * start move here. | |
1572 | */ | |
1573 | ||
1574 | /* for quick checking without looking up memcg */ | |
1575 | atomic_t memcg_moving __read_mostly; | |
1576 | ||
1577 | static void mem_cgroup_start_move(struct mem_cgroup *memcg) | |
1578 | { | |
1579 | atomic_inc(&memcg_moving); | |
1580 | atomic_inc(&memcg->moving_account); | |
1581 | synchronize_rcu(); | |
1582 | } | |
1583 | ||
1584 | static void mem_cgroup_end_move(struct mem_cgroup *memcg) | |
1585 | { | |
1586 | /* | |
1587 | * Now, mem_cgroup_clear_mc() may call this function with NULL. | |
1588 | * We check NULL in callee rather than caller. | |
1589 | */ | |
1590 | if (memcg) { | |
1591 | atomic_dec(&memcg_moving); | |
1592 | atomic_dec(&memcg->moving_account); | |
1593 | } | |
1594 | } | |
1595 | ||
1596 | /* | |
1597 | * 2 routines for checking "mem" is under move_account() or not. | |
1598 | * | |
1599 | * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This | |
1600 | * is used for avoiding races in accounting. If true, | |
1601 | * pc->mem_cgroup may be overwritten. | |
1602 | * | |
1603 | * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or | |
1604 | * under hierarchy of moving cgroups. This is for | |
1605 | * waiting at hith-memory prressure caused by "move". | |
1606 | */ | |
1607 | ||
1608 | static bool mem_cgroup_stolen(struct mem_cgroup *memcg) | |
1609 | { | |
1610 | VM_BUG_ON(!rcu_read_lock_held()); | |
1611 | return atomic_read(&memcg->moving_account) > 0; | |
1612 | } | |
1613 | ||
1614 | static bool mem_cgroup_under_move(struct mem_cgroup *memcg) | |
1615 | { | |
1616 | struct mem_cgroup *from; | |
1617 | struct mem_cgroup *to; | |
1618 | bool ret = false; | |
1619 | /* | |
1620 | * Unlike task_move routines, we access mc.to, mc.from not under | |
1621 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. | |
1622 | */ | |
1623 | spin_lock(&mc.lock); | |
1624 | from = mc.from; | |
1625 | to = mc.to; | |
1626 | if (!from) | |
1627 | goto unlock; | |
1628 | ||
1629 | ret = mem_cgroup_same_or_subtree(memcg, from) | |
1630 | || mem_cgroup_same_or_subtree(memcg, to); | |
1631 | unlock: | |
1632 | spin_unlock(&mc.lock); | |
1633 | return ret; | |
1634 | } | |
1635 | ||
1636 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) | |
1637 | { | |
1638 | if (mc.moving_task && current != mc.moving_task) { | |
1639 | if (mem_cgroup_under_move(memcg)) { | |
1640 | DEFINE_WAIT(wait); | |
1641 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); | |
1642 | /* moving charge context might have finished. */ | |
1643 | if (mc.moving_task) | |
1644 | schedule(); | |
1645 | finish_wait(&mc.waitq, &wait); | |
1646 | return true; | |
1647 | } | |
1648 | } | |
1649 | return false; | |
1650 | } | |
1651 | ||
1652 | /* | |
1653 | * Take this lock when | |
1654 | * - a code tries to modify page's memcg while it's USED. | |
1655 | * - a code tries to modify page state accounting in a memcg. | |
1656 | * see mem_cgroup_stolen(), too. | |
1657 | */ | |
1658 | static void move_lock_mem_cgroup(struct mem_cgroup *memcg, | |
1659 | unsigned long *flags) | |
1660 | { | |
1661 | spin_lock_irqsave(&memcg->move_lock, *flags); | |
1662 | } | |
1663 | ||
1664 | static void move_unlock_mem_cgroup(struct mem_cgroup *memcg, | |
1665 | unsigned long *flags) | |
1666 | { | |
1667 | spin_unlock_irqrestore(&memcg->move_lock, *flags); | |
1668 | } | |
1669 | ||
1670 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
1671 | /** | |
1672 | * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller. | |
1673 | * @memcg: The memory cgroup that went over limit | |
1674 | * @p: Task that is going to be killed | |
1675 | * | |
1676 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is | |
1677 | * enabled | |
1678 | */ | |
1679 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) | |
1680 | { | |
1681 | /* oom_info_lock ensures that parallel ooms do not interleave */ | |
1682 | static DEFINE_MUTEX(oom_info_lock); | |
1683 | struct mem_cgroup *iter; | |
1684 | unsigned int i; | |
1685 | ||
1686 | if (!p) | |
1687 | return; | |
1688 | ||
1689 | mutex_lock(&oom_info_lock); | |
1690 | rcu_read_lock(); | |
1691 | ||
1692 | pr_info("Task in "); | |
1693 | pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id)); | |
1694 | pr_info(" killed as a result of limit of "); | |
1695 | pr_cont_cgroup_path(memcg->css.cgroup); | |
1696 | pr_info("\n"); | |
1697 | ||
1698 | rcu_read_unlock(); | |
1699 | ||
1700 | pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n", | |
1701 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, | |
1702 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, | |
1703 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); | |
1704 | pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n", | |
1705 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, | |
1706 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, | |
1707 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); | |
1708 | pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n", | |
1709 | res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10, | |
1710 | res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10, | |
1711 | res_counter_read_u64(&memcg->kmem, RES_FAILCNT)); | |
1712 | ||
1713 | for_each_mem_cgroup_tree(iter, memcg) { | |
1714 | pr_info("Memory cgroup stats for "); | |
1715 | pr_cont_cgroup_path(iter->css.cgroup); | |
1716 | pr_cont(":"); | |
1717 | ||
1718 | for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) { | |
1719 | if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account) | |
1720 | continue; | |
1721 | pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i], | |
1722 | K(mem_cgroup_read_stat(iter, i))); | |
1723 | } | |
1724 | ||
1725 | for (i = 0; i < NR_LRU_LISTS; i++) | |
1726 | pr_cont(" %s:%luKB", mem_cgroup_lru_names[i], | |
1727 | K(mem_cgroup_nr_lru_pages(iter, BIT(i)))); | |
1728 | ||
1729 | pr_cont("\n"); | |
1730 | } | |
1731 | mutex_unlock(&oom_info_lock); | |
1732 | } | |
1733 | ||
1734 | /* | |
1735 | * This function returns the number of memcg under hierarchy tree. Returns | |
1736 | * 1(self count) if no children. | |
1737 | */ | |
1738 | static int mem_cgroup_count_children(struct mem_cgroup *memcg) | |
1739 | { | |
1740 | int num = 0; | |
1741 | struct mem_cgroup *iter; | |
1742 | ||
1743 | for_each_mem_cgroup_tree(iter, memcg) | |
1744 | num++; | |
1745 | return num; | |
1746 | } | |
1747 | ||
1748 | /* | |
1749 | * Return the memory (and swap, if configured) limit for a memcg. | |
1750 | */ | |
1751 | static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) | |
1752 | { | |
1753 | u64 limit; | |
1754 | ||
1755 | limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
1756 | ||
1757 | /* | |
1758 | * Do not consider swap space if we cannot swap due to swappiness | |
1759 | */ | |
1760 | if (mem_cgroup_swappiness(memcg)) { | |
1761 | u64 memsw; | |
1762 | ||
1763 | limit += total_swap_pages << PAGE_SHIFT; | |
1764 | memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
1765 | ||
1766 | /* | |
1767 | * If memsw is finite and limits the amount of swap space | |
1768 | * available to this memcg, return that limit. | |
1769 | */ | |
1770 | limit = min(limit, memsw); | |
1771 | } | |
1772 | ||
1773 | return limit; | |
1774 | } | |
1775 | ||
1776 | static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask, | |
1777 | int order) | |
1778 | { | |
1779 | struct mem_cgroup *iter; | |
1780 | unsigned long chosen_points = 0; | |
1781 | unsigned long totalpages; | |
1782 | unsigned int points = 0; | |
1783 | struct task_struct *chosen = NULL; | |
1784 | ||
1785 | /* | |
1786 | * If current has a pending SIGKILL or is exiting, then automatically | |
1787 | * select it. The goal is to allow it to allocate so that it may | |
1788 | * quickly exit and free its memory. | |
1789 | */ | |
1790 | if (fatal_signal_pending(current) || current->flags & PF_EXITING) { | |
1791 | set_thread_flag(TIF_MEMDIE); | |
1792 | return; | |
1793 | } | |
1794 | ||
1795 | check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL); | |
1796 | totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1; | |
1797 | for_each_mem_cgroup_tree(iter, memcg) { | |
1798 | struct css_task_iter it; | |
1799 | struct task_struct *task; | |
1800 | ||
1801 | css_task_iter_start(&iter->css, &it); | |
1802 | while ((task = css_task_iter_next(&it))) { | |
1803 | switch (oom_scan_process_thread(task, totalpages, NULL, | |
1804 | false)) { | |
1805 | case OOM_SCAN_SELECT: | |
1806 | if (chosen) | |
1807 | put_task_struct(chosen); | |
1808 | chosen = task; | |
1809 | chosen_points = ULONG_MAX; | |
1810 | get_task_struct(chosen); | |
1811 | /* fall through */ | |
1812 | case OOM_SCAN_CONTINUE: | |
1813 | continue; | |
1814 | case OOM_SCAN_ABORT: | |
1815 | css_task_iter_end(&it); | |
1816 | mem_cgroup_iter_break(memcg, iter); | |
1817 | if (chosen) | |
1818 | put_task_struct(chosen); | |
1819 | return; | |
1820 | case OOM_SCAN_OK: | |
1821 | break; | |
1822 | }; | |
1823 | points = oom_badness(task, memcg, NULL, totalpages); | |
1824 | if (!points || points < chosen_points) | |
1825 | continue; | |
1826 | /* Prefer thread group leaders for display purposes */ | |
1827 | if (points == chosen_points && | |
1828 | thread_group_leader(chosen)) | |
1829 | continue; | |
1830 | ||
1831 | if (chosen) | |
1832 | put_task_struct(chosen); | |
1833 | chosen = task; | |
1834 | chosen_points = points; | |
1835 | get_task_struct(chosen); | |
1836 | } | |
1837 | css_task_iter_end(&it); | |
1838 | } | |
1839 | ||
1840 | if (!chosen) | |
1841 | return; | |
1842 | points = chosen_points * 1000 / totalpages; | |
1843 | oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg, | |
1844 | NULL, "Memory cgroup out of memory"); | |
1845 | } | |
1846 | ||
1847 | static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, | |
1848 | gfp_t gfp_mask, | |
1849 | unsigned long flags) | |
1850 | { | |
1851 | unsigned long total = 0; | |
1852 | bool noswap = false; | |
1853 | int loop; | |
1854 | ||
1855 | if (flags & MEM_CGROUP_RECLAIM_NOSWAP) | |
1856 | noswap = true; | |
1857 | if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) | |
1858 | noswap = true; | |
1859 | ||
1860 | for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { | |
1861 | if (loop) | |
1862 | drain_all_stock_async(memcg); | |
1863 | total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); | |
1864 | /* | |
1865 | * Allow limit shrinkers, which are triggered directly | |
1866 | * by userspace, to catch signals and stop reclaim | |
1867 | * after minimal progress, regardless of the margin. | |
1868 | */ | |
1869 | if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) | |
1870 | break; | |
1871 | if (mem_cgroup_margin(memcg)) | |
1872 | break; | |
1873 | /* | |
1874 | * If nothing was reclaimed after two attempts, there | |
1875 | * may be no reclaimable pages in this hierarchy. | |
1876 | */ | |
1877 | if (loop && !total) | |
1878 | break; | |
1879 | } | |
1880 | return total; | |
1881 | } | |
1882 | ||
1883 | /** | |
1884 | * test_mem_cgroup_node_reclaimable | |
1885 | * @memcg: the target memcg | |
1886 | * @nid: the node ID to be checked. | |
1887 | * @noswap : specify true here if the user wants flle only information. | |
1888 | * | |
1889 | * This function returns whether the specified memcg contains any | |
1890 | * reclaimable pages on a node. Returns true if there are any reclaimable | |
1891 | * pages in the node. | |
1892 | */ | |
1893 | static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, | |
1894 | int nid, bool noswap) | |
1895 | { | |
1896 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) | |
1897 | return true; | |
1898 | if (noswap || !total_swap_pages) | |
1899 | return false; | |
1900 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) | |
1901 | return true; | |
1902 | return false; | |
1903 | ||
1904 | } | |
1905 | #if MAX_NUMNODES > 1 | |
1906 | ||
1907 | /* | |
1908 | * Always updating the nodemask is not very good - even if we have an empty | |
1909 | * list or the wrong list here, we can start from some node and traverse all | |
1910 | * nodes based on the zonelist. So update the list loosely once per 10 secs. | |
1911 | * | |
1912 | */ | |
1913 | static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) | |
1914 | { | |
1915 | int nid; | |
1916 | /* | |
1917 | * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET | |
1918 | * pagein/pageout changes since the last update. | |
1919 | */ | |
1920 | if (!atomic_read(&memcg->numainfo_events)) | |
1921 | return; | |
1922 | if (atomic_inc_return(&memcg->numainfo_updating) > 1) | |
1923 | return; | |
1924 | ||
1925 | /* make a nodemask where this memcg uses memory from */ | |
1926 | memcg->scan_nodes = node_states[N_MEMORY]; | |
1927 | ||
1928 | for_each_node_mask(nid, node_states[N_MEMORY]) { | |
1929 | ||
1930 | if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) | |
1931 | node_clear(nid, memcg->scan_nodes); | |
1932 | } | |
1933 | ||
1934 | atomic_set(&memcg->numainfo_events, 0); | |
1935 | atomic_set(&memcg->numainfo_updating, 0); | |
1936 | } | |
1937 | ||
1938 | /* | |
1939 | * Selecting a node where we start reclaim from. Because what we need is just | |
1940 | * reducing usage counter, start from anywhere is O,K. Considering | |
1941 | * memory reclaim from current node, there are pros. and cons. | |
1942 | * | |
1943 | * Freeing memory from current node means freeing memory from a node which | |
1944 | * we'll use or we've used. So, it may make LRU bad. And if several threads | |
1945 | * hit limits, it will see a contention on a node. But freeing from remote | |
1946 | * node means more costs for memory reclaim because of memory latency. | |
1947 | * | |
1948 | * Now, we use round-robin. Better algorithm is welcomed. | |
1949 | */ | |
1950 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) | |
1951 | { | |
1952 | int node; | |
1953 | ||
1954 | mem_cgroup_may_update_nodemask(memcg); | |
1955 | node = memcg->last_scanned_node; | |
1956 | ||
1957 | node = next_node(node, memcg->scan_nodes); | |
1958 | if (node == MAX_NUMNODES) | |
1959 | node = first_node(memcg->scan_nodes); | |
1960 | /* | |
1961 | * We call this when we hit limit, not when pages are added to LRU. | |
1962 | * No LRU may hold pages because all pages are UNEVICTABLE or | |
1963 | * memcg is too small and all pages are not on LRU. In that case, | |
1964 | * we use curret node. | |
1965 | */ | |
1966 | if (unlikely(node == MAX_NUMNODES)) | |
1967 | node = numa_node_id(); | |
1968 | ||
1969 | memcg->last_scanned_node = node; | |
1970 | return node; | |
1971 | } | |
1972 | ||
1973 | /* | |
1974 | * Check all nodes whether it contains reclaimable pages or not. | |
1975 | * For quick scan, we make use of scan_nodes. This will allow us to skip | |
1976 | * unused nodes. But scan_nodes is lazily updated and may not cotain | |
1977 | * enough new information. We need to do double check. | |
1978 | */ | |
1979 | static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) | |
1980 | { | |
1981 | int nid; | |
1982 | ||
1983 | /* | |
1984 | * quick check...making use of scan_node. | |
1985 | * We can skip unused nodes. | |
1986 | */ | |
1987 | if (!nodes_empty(memcg->scan_nodes)) { | |
1988 | for (nid = first_node(memcg->scan_nodes); | |
1989 | nid < MAX_NUMNODES; | |
1990 | nid = next_node(nid, memcg->scan_nodes)) { | |
1991 | ||
1992 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) | |
1993 | return true; | |
1994 | } | |
1995 | } | |
1996 | /* | |
1997 | * Check rest of nodes. | |
1998 | */ | |
1999 | for_each_node_state(nid, N_MEMORY) { | |
2000 | if (node_isset(nid, memcg->scan_nodes)) | |
2001 | continue; | |
2002 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) | |
2003 | return true; | |
2004 | } | |
2005 | return false; | |
2006 | } | |
2007 | ||
2008 | #else | |
2009 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) | |
2010 | { | |
2011 | return 0; | |
2012 | } | |
2013 | ||
2014 | static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) | |
2015 | { | |
2016 | return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); | |
2017 | } | |
2018 | #endif | |
2019 | ||
2020 | static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, | |
2021 | struct zone *zone, | |
2022 | gfp_t gfp_mask, | |
2023 | unsigned long *total_scanned) | |
2024 | { | |
2025 | struct mem_cgroup *victim = NULL; | |
2026 | int total = 0; | |
2027 | int loop = 0; | |
2028 | unsigned long excess; | |
2029 | unsigned long nr_scanned; | |
2030 | struct mem_cgroup_reclaim_cookie reclaim = { | |
2031 | .zone = zone, | |
2032 | .priority = 0, | |
2033 | }; | |
2034 | ||
2035 | excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; | |
2036 | ||
2037 | while (1) { | |
2038 | victim = mem_cgroup_iter(root_memcg, victim, &reclaim); | |
2039 | if (!victim) { | |
2040 | loop++; | |
2041 | if (loop >= 2) { | |
2042 | /* | |
2043 | * If we have not been able to reclaim | |
2044 | * anything, it might because there are | |
2045 | * no reclaimable pages under this hierarchy | |
2046 | */ | |
2047 | if (!total) | |
2048 | break; | |
2049 | /* | |
2050 | * We want to do more targeted reclaim. | |
2051 | * excess >> 2 is not to excessive so as to | |
2052 | * reclaim too much, nor too less that we keep | |
2053 | * coming back to reclaim from this cgroup | |
2054 | */ | |
2055 | if (total >= (excess >> 2) || | |
2056 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) | |
2057 | break; | |
2058 | } | |
2059 | continue; | |
2060 | } | |
2061 | if (!mem_cgroup_reclaimable(victim, false)) | |
2062 | continue; | |
2063 | total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, | |
2064 | zone, &nr_scanned); | |
2065 | *total_scanned += nr_scanned; | |
2066 | if (!res_counter_soft_limit_excess(&root_memcg->res)) | |
2067 | break; | |
2068 | } | |
2069 | mem_cgroup_iter_break(root_memcg, victim); | |
2070 | return total; | |
2071 | } | |
2072 | ||
2073 | #ifdef CONFIG_LOCKDEP | |
2074 | static struct lockdep_map memcg_oom_lock_dep_map = { | |
2075 | .name = "memcg_oom_lock", | |
2076 | }; | |
2077 | #endif | |
2078 | ||
2079 | static DEFINE_SPINLOCK(memcg_oom_lock); | |
2080 | ||
2081 | /* | |
2082 | * Check OOM-Killer is already running under our hierarchy. | |
2083 | * If someone is running, return false. | |
2084 | */ | |
2085 | static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg) | |
2086 | { | |
2087 | struct mem_cgroup *iter, *failed = NULL; | |
2088 | ||
2089 | spin_lock(&memcg_oom_lock); | |
2090 | ||
2091 | for_each_mem_cgroup_tree(iter, memcg) { | |
2092 | if (iter->oom_lock) { | |
2093 | /* | |
2094 | * this subtree of our hierarchy is already locked | |
2095 | * so we cannot give a lock. | |
2096 | */ | |
2097 | failed = iter; | |
2098 | mem_cgroup_iter_break(memcg, iter); | |
2099 | break; | |
2100 | } else | |
2101 | iter->oom_lock = true; | |
2102 | } | |
2103 | ||
2104 | if (failed) { | |
2105 | /* | |
2106 | * OK, we failed to lock the whole subtree so we have | |
2107 | * to clean up what we set up to the failing subtree | |
2108 | */ | |
2109 | for_each_mem_cgroup_tree(iter, memcg) { | |
2110 | if (iter == failed) { | |
2111 | mem_cgroup_iter_break(memcg, iter); | |
2112 | break; | |
2113 | } | |
2114 | iter->oom_lock = false; | |
2115 | } | |
2116 | } else | |
2117 | mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_); | |
2118 | ||
2119 | spin_unlock(&memcg_oom_lock); | |
2120 | ||
2121 | return !failed; | |
2122 | } | |
2123 | ||
2124 | static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg) | |
2125 | { | |
2126 | struct mem_cgroup *iter; | |
2127 | ||
2128 | spin_lock(&memcg_oom_lock); | |
2129 | mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_); | |
2130 | for_each_mem_cgroup_tree(iter, memcg) | |
2131 | iter->oom_lock = false; | |
2132 | spin_unlock(&memcg_oom_lock); | |
2133 | } | |
2134 | ||
2135 | static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) | |
2136 | { | |
2137 | struct mem_cgroup *iter; | |
2138 | ||
2139 | for_each_mem_cgroup_tree(iter, memcg) | |
2140 | atomic_inc(&iter->under_oom); | |
2141 | } | |
2142 | ||
2143 | static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) | |
2144 | { | |
2145 | struct mem_cgroup *iter; | |
2146 | ||
2147 | /* | |
2148 | * When a new child is created while the hierarchy is under oom, | |
2149 | * mem_cgroup_oom_lock() may not be called. We have to use | |
2150 | * atomic_add_unless() here. | |
2151 | */ | |
2152 | for_each_mem_cgroup_tree(iter, memcg) | |
2153 | atomic_add_unless(&iter->under_oom, -1, 0); | |
2154 | } | |
2155 | ||
2156 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); | |
2157 | ||
2158 | struct oom_wait_info { | |
2159 | struct mem_cgroup *memcg; | |
2160 | wait_queue_t wait; | |
2161 | }; | |
2162 | ||
2163 | static int memcg_oom_wake_function(wait_queue_t *wait, | |
2164 | unsigned mode, int sync, void *arg) | |
2165 | { | |
2166 | struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg; | |
2167 | struct mem_cgroup *oom_wait_memcg; | |
2168 | struct oom_wait_info *oom_wait_info; | |
2169 | ||
2170 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); | |
2171 | oom_wait_memcg = oom_wait_info->memcg; | |
2172 | ||
2173 | /* | |
2174 | * Both of oom_wait_info->memcg and wake_memcg are stable under us. | |
2175 | * Then we can use css_is_ancestor without taking care of RCU. | |
2176 | */ | |
2177 | if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) | |
2178 | && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) | |
2179 | return 0; | |
2180 | return autoremove_wake_function(wait, mode, sync, arg); | |
2181 | } | |
2182 | ||
2183 | static void memcg_wakeup_oom(struct mem_cgroup *memcg) | |
2184 | { | |
2185 | atomic_inc(&memcg->oom_wakeups); | |
2186 | /* for filtering, pass "memcg" as argument. */ | |
2187 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); | |
2188 | } | |
2189 | ||
2190 | static void memcg_oom_recover(struct mem_cgroup *memcg) | |
2191 | { | |
2192 | if (memcg && atomic_read(&memcg->under_oom)) | |
2193 | memcg_wakeup_oom(memcg); | |
2194 | } | |
2195 | ||
2196 | static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order) | |
2197 | { | |
2198 | if (!current->memcg_oom.may_oom) | |
2199 | return; | |
2200 | /* | |
2201 | * We are in the middle of the charge context here, so we | |
2202 | * don't want to block when potentially sitting on a callstack | |
2203 | * that holds all kinds of filesystem and mm locks. | |
2204 | * | |
2205 | * Also, the caller may handle a failed allocation gracefully | |
2206 | * (like optional page cache readahead) and so an OOM killer | |
2207 | * invocation might not even be necessary. | |
2208 | * | |
2209 | * That's why we don't do anything here except remember the | |
2210 | * OOM context and then deal with it at the end of the page | |
2211 | * fault when the stack is unwound, the locks are released, | |
2212 | * and when we know whether the fault was overall successful. | |
2213 | */ | |
2214 | css_get(&memcg->css); | |
2215 | current->memcg_oom.memcg = memcg; | |
2216 | current->memcg_oom.gfp_mask = mask; | |
2217 | current->memcg_oom.order = order; | |
2218 | } | |
2219 | ||
2220 | /** | |
2221 | * mem_cgroup_oom_synchronize - complete memcg OOM handling | |
2222 | * @handle: actually kill/wait or just clean up the OOM state | |
2223 | * | |
2224 | * This has to be called at the end of a page fault if the memcg OOM | |
2225 | * handler was enabled. | |
2226 | * | |
2227 | * Memcg supports userspace OOM handling where failed allocations must | |
2228 | * sleep on a waitqueue until the userspace task resolves the | |
2229 | * situation. Sleeping directly in the charge context with all kinds | |
2230 | * of locks held is not a good idea, instead we remember an OOM state | |
2231 | * in the task and mem_cgroup_oom_synchronize() has to be called at | |
2232 | * the end of the page fault to complete the OOM handling. | |
2233 | * | |
2234 | * Returns %true if an ongoing memcg OOM situation was detected and | |
2235 | * completed, %false otherwise. | |
2236 | */ | |
2237 | bool mem_cgroup_oom_synchronize(bool handle) | |
2238 | { | |
2239 | struct mem_cgroup *memcg = current->memcg_oom.memcg; | |
2240 | struct oom_wait_info owait; | |
2241 | bool locked; | |
2242 | ||
2243 | /* OOM is global, do not handle */ | |
2244 | if (!memcg) | |
2245 | return false; | |
2246 | ||
2247 | if (!handle) | |
2248 | goto cleanup; | |
2249 | ||
2250 | owait.memcg = memcg; | |
2251 | owait.wait.flags = 0; | |
2252 | owait.wait.func = memcg_oom_wake_function; | |
2253 | owait.wait.private = current; | |
2254 | INIT_LIST_HEAD(&owait.wait.task_list); | |
2255 | ||
2256 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); | |
2257 | mem_cgroup_mark_under_oom(memcg); | |
2258 | ||
2259 | locked = mem_cgroup_oom_trylock(memcg); | |
2260 | ||
2261 | if (locked) | |
2262 | mem_cgroup_oom_notify(memcg); | |
2263 | ||
2264 | if (locked && !memcg->oom_kill_disable) { | |
2265 | mem_cgroup_unmark_under_oom(memcg); | |
2266 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
2267 | mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask, | |
2268 | current->memcg_oom.order); | |
2269 | } else { | |
2270 | schedule(); | |
2271 | mem_cgroup_unmark_under_oom(memcg); | |
2272 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
2273 | } | |
2274 | ||
2275 | if (locked) { | |
2276 | mem_cgroup_oom_unlock(memcg); | |
2277 | /* | |
2278 | * There is no guarantee that an OOM-lock contender | |
2279 | * sees the wakeups triggered by the OOM kill | |
2280 | * uncharges. Wake any sleepers explicitely. | |
2281 | */ | |
2282 | memcg_oom_recover(memcg); | |
2283 | } | |
2284 | cleanup: | |
2285 | current->memcg_oom.memcg = NULL; | |
2286 | css_put(&memcg->css); | |
2287 | return true; | |
2288 | } | |
2289 | ||
2290 | /* | |
2291 | * Currently used to update mapped file statistics, but the routine can be | |
2292 | * generalized to update other statistics as well. | |
2293 | * | |
2294 | * Notes: Race condition | |
2295 | * | |
2296 | * We usually use page_cgroup_lock() for accessing page_cgroup member but | |
2297 | * it tends to be costly. But considering some conditions, we doesn't need | |
2298 | * to do so _always_. | |
2299 | * | |
2300 | * Considering "charge", lock_page_cgroup() is not required because all | |
2301 | * file-stat operations happen after a page is attached to radix-tree. There | |
2302 | * are no race with "charge". | |
2303 | * | |
2304 | * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup | |
2305 | * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even | |
2306 | * if there are race with "uncharge". Statistics itself is properly handled | |
2307 | * by flags. | |
2308 | * | |
2309 | * Considering "move", this is an only case we see a race. To make the race | |
2310 | * small, we check mm->moving_account and detect there are possibility of race | |
2311 | * If there is, we take a lock. | |
2312 | */ | |
2313 | ||
2314 | void __mem_cgroup_begin_update_page_stat(struct page *page, | |
2315 | bool *locked, unsigned long *flags) | |
2316 | { | |
2317 | struct mem_cgroup *memcg; | |
2318 | struct page_cgroup *pc; | |
2319 | ||
2320 | pc = lookup_page_cgroup(page); | |
2321 | again: | |
2322 | memcg = pc->mem_cgroup; | |
2323 | if (unlikely(!memcg || !PageCgroupUsed(pc))) | |
2324 | return; | |
2325 | /* | |
2326 | * If this memory cgroup is not under account moving, we don't | |
2327 | * need to take move_lock_mem_cgroup(). Because we already hold | |
2328 | * rcu_read_lock(), any calls to move_account will be delayed until | |
2329 | * rcu_read_unlock() if mem_cgroup_stolen() == true. | |
2330 | */ | |
2331 | if (!mem_cgroup_stolen(memcg)) | |
2332 | return; | |
2333 | ||
2334 | move_lock_mem_cgroup(memcg, flags); | |
2335 | if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) { | |
2336 | move_unlock_mem_cgroup(memcg, flags); | |
2337 | goto again; | |
2338 | } | |
2339 | *locked = true; | |
2340 | } | |
2341 | ||
2342 | void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags) | |
2343 | { | |
2344 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
2345 | ||
2346 | /* | |
2347 | * It's guaranteed that pc->mem_cgroup never changes while | |
2348 | * lock is held because a routine modifies pc->mem_cgroup | |
2349 | * should take move_lock_mem_cgroup(). | |
2350 | */ | |
2351 | move_unlock_mem_cgroup(pc->mem_cgroup, flags); | |
2352 | } | |
2353 | ||
2354 | void mem_cgroup_update_page_stat(struct page *page, | |
2355 | enum mem_cgroup_stat_index idx, int val) | |
2356 | { | |
2357 | struct mem_cgroup *memcg; | |
2358 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
2359 | unsigned long uninitialized_var(flags); | |
2360 | ||
2361 | if (mem_cgroup_disabled()) | |
2362 | return; | |
2363 | ||
2364 | VM_BUG_ON(!rcu_read_lock_held()); | |
2365 | memcg = pc->mem_cgroup; | |
2366 | if (unlikely(!memcg || !PageCgroupUsed(pc))) | |
2367 | return; | |
2368 | ||
2369 | this_cpu_add(memcg->stat->count[idx], val); | |
2370 | } | |
2371 | ||
2372 | /* | |
2373 | * size of first charge trial. "32" comes from vmscan.c's magic value. | |
2374 | * TODO: maybe necessary to use big numbers in big irons. | |
2375 | */ | |
2376 | #define CHARGE_BATCH 32U | |
2377 | struct memcg_stock_pcp { | |
2378 | struct mem_cgroup *cached; /* this never be root cgroup */ | |
2379 | unsigned int nr_pages; | |
2380 | struct work_struct work; | |
2381 | unsigned long flags; | |
2382 | #define FLUSHING_CACHED_CHARGE 0 | |
2383 | }; | |
2384 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); | |
2385 | static DEFINE_MUTEX(percpu_charge_mutex); | |
2386 | ||
2387 | /** | |
2388 | * consume_stock: Try to consume stocked charge on this cpu. | |
2389 | * @memcg: memcg to consume from. | |
2390 | * @nr_pages: how many pages to charge. | |
2391 | * | |
2392 | * The charges will only happen if @memcg matches the current cpu's memcg | |
2393 | * stock, and at least @nr_pages are available in that stock. Failure to | |
2394 | * service an allocation will refill the stock. | |
2395 | * | |
2396 | * returns true if successful, false otherwise. | |
2397 | */ | |
2398 | static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages) | |
2399 | { | |
2400 | struct memcg_stock_pcp *stock; | |
2401 | bool ret = true; | |
2402 | ||
2403 | if (nr_pages > CHARGE_BATCH) | |
2404 | return false; | |
2405 | ||
2406 | stock = &get_cpu_var(memcg_stock); | |
2407 | if (memcg == stock->cached && stock->nr_pages >= nr_pages) | |
2408 | stock->nr_pages -= nr_pages; | |
2409 | else /* need to call res_counter_charge */ | |
2410 | ret = false; | |
2411 | put_cpu_var(memcg_stock); | |
2412 | return ret; | |
2413 | } | |
2414 | ||
2415 | /* | |
2416 | * Returns stocks cached in percpu to res_counter and reset cached information. | |
2417 | */ | |
2418 | static void drain_stock(struct memcg_stock_pcp *stock) | |
2419 | { | |
2420 | struct mem_cgroup *old = stock->cached; | |
2421 | ||
2422 | if (stock->nr_pages) { | |
2423 | unsigned long bytes = stock->nr_pages * PAGE_SIZE; | |
2424 | ||
2425 | res_counter_uncharge(&old->res, bytes); | |
2426 | if (do_swap_account) | |
2427 | res_counter_uncharge(&old->memsw, bytes); | |
2428 | stock->nr_pages = 0; | |
2429 | } | |
2430 | stock->cached = NULL; | |
2431 | } | |
2432 | ||
2433 | /* | |
2434 | * This must be called under preempt disabled or must be called by | |
2435 | * a thread which is pinned to local cpu. | |
2436 | */ | |
2437 | static void drain_local_stock(struct work_struct *dummy) | |
2438 | { | |
2439 | struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock); | |
2440 | drain_stock(stock); | |
2441 | clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); | |
2442 | } | |
2443 | ||
2444 | static void __init memcg_stock_init(void) | |
2445 | { | |
2446 | int cpu; | |
2447 | ||
2448 | for_each_possible_cpu(cpu) { | |
2449 | struct memcg_stock_pcp *stock = | |
2450 | &per_cpu(memcg_stock, cpu); | |
2451 | INIT_WORK(&stock->work, drain_local_stock); | |
2452 | } | |
2453 | } | |
2454 | ||
2455 | /* | |
2456 | * Cache charges(val) which is from res_counter, to local per_cpu area. | |
2457 | * This will be consumed by consume_stock() function, later. | |
2458 | */ | |
2459 | static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) | |
2460 | { | |
2461 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); | |
2462 | ||
2463 | if (stock->cached != memcg) { /* reset if necessary */ | |
2464 | drain_stock(stock); | |
2465 | stock->cached = memcg; | |
2466 | } | |
2467 | stock->nr_pages += nr_pages; | |
2468 | put_cpu_var(memcg_stock); | |
2469 | } | |
2470 | ||
2471 | /* | |
2472 | * Drains all per-CPU charge caches for given root_memcg resp. subtree | |
2473 | * of the hierarchy under it. sync flag says whether we should block | |
2474 | * until the work is done. | |
2475 | */ | |
2476 | static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) | |
2477 | { | |
2478 | int cpu, curcpu; | |
2479 | ||
2480 | /* Notify other cpus that system-wide "drain" is running */ | |
2481 | get_online_cpus(); | |
2482 | curcpu = get_cpu(); | |
2483 | for_each_online_cpu(cpu) { | |
2484 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
2485 | struct mem_cgroup *memcg; | |
2486 | ||
2487 | memcg = stock->cached; | |
2488 | if (!memcg || !stock->nr_pages) | |
2489 | continue; | |
2490 | if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) | |
2491 | continue; | |
2492 | if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { | |
2493 | if (cpu == curcpu) | |
2494 | drain_local_stock(&stock->work); | |
2495 | else | |
2496 | schedule_work_on(cpu, &stock->work); | |
2497 | } | |
2498 | } | |
2499 | put_cpu(); | |
2500 | ||
2501 | if (!sync) | |
2502 | goto out; | |
2503 | ||
2504 | for_each_online_cpu(cpu) { | |
2505 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
2506 | if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) | |
2507 | flush_work(&stock->work); | |
2508 | } | |
2509 | out: | |
2510 | put_online_cpus(); | |
2511 | } | |
2512 | ||
2513 | /* | |
2514 | * Tries to drain stocked charges in other cpus. This function is asynchronous | |
2515 | * and just put a work per cpu for draining localy on each cpu. Caller can | |
2516 | * expects some charges will be back to res_counter later but cannot wait for | |
2517 | * it. | |
2518 | */ | |
2519 | static void drain_all_stock_async(struct mem_cgroup *root_memcg) | |
2520 | { | |
2521 | /* | |
2522 | * If someone calls draining, avoid adding more kworker runs. | |
2523 | */ | |
2524 | if (!mutex_trylock(&percpu_charge_mutex)) | |
2525 | return; | |
2526 | drain_all_stock(root_memcg, false); | |
2527 | mutex_unlock(&percpu_charge_mutex); | |
2528 | } | |
2529 | ||
2530 | /* This is a synchronous drain interface. */ | |
2531 | static void drain_all_stock_sync(struct mem_cgroup *root_memcg) | |
2532 | { | |
2533 | /* called when force_empty is called */ | |
2534 | mutex_lock(&percpu_charge_mutex); | |
2535 | drain_all_stock(root_memcg, true); | |
2536 | mutex_unlock(&percpu_charge_mutex); | |
2537 | } | |
2538 | ||
2539 | /* | |
2540 | * This function drains percpu counter value from DEAD cpu and | |
2541 | * move it to local cpu. Note that this function can be preempted. | |
2542 | */ | |
2543 | static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) | |
2544 | { | |
2545 | int i; | |
2546 | ||
2547 | spin_lock(&memcg->pcp_counter_lock); | |
2548 | for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) { | |
2549 | long x = per_cpu(memcg->stat->count[i], cpu); | |
2550 | ||
2551 | per_cpu(memcg->stat->count[i], cpu) = 0; | |
2552 | memcg->nocpu_base.count[i] += x; | |
2553 | } | |
2554 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { | |
2555 | unsigned long x = per_cpu(memcg->stat->events[i], cpu); | |
2556 | ||
2557 | per_cpu(memcg->stat->events[i], cpu) = 0; | |
2558 | memcg->nocpu_base.events[i] += x; | |
2559 | } | |
2560 | spin_unlock(&memcg->pcp_counter_lock); | |
2561 | } | |
2562 | ||
2563 | static int memcg_cpu_hotplug_callback(struct notifier_block *nb, | |
2564 | unsigned long action, | |
2565 | void *hcpu) | |
2566 | { | |
2567 | int cpu = (unsigned long)hcpu; | |
2568 | struct memcg_stock_pcp *stock; | |
2569 | struct mem_cgroup *iter; | |
2570 | ||
2571 | if (action == CPU_ONLINE) | |
2572 | return NOTIFY_OK; | |
2573 | ||
2574 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) | |
2575 | return NOTIFY_OK; | |
2576 | ||
2577 | for_each_mem_cgroup(iter) | |
2578 | mem_cgroup_drain_pcp_counter(iter, cpu); | |
2579 | ||
2580 | stock = &per_cpu(memcg_stock, cpu); | |
2581 | drain_stock(stock); | |
2582 | return NOTIFY_OK; | |
2583 | } | |
2584 | ||
2585 | ||
2586 | /* See mem_cgroup_try_charge() for details */ | |
2587 | enum { | |
2588 | CHARGE_OK, /* success */ | |
2589 | CHARGE_RETRY, /* need to retry but retry is not bad */ | |
2590 | CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ | |
2591 | CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ | |
2592 | }; | |
2593 | ||
2594 | static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, | |
2595 | unsigned int nr_pages, unsigned int min_pages, | |
2596 | bool invoke_oom) | |
2597 | { | |
2598 | unsigned long csize = nr_pages * PAGE_SIZE; | |
2599 | struct mem_cgroup *mem_over_limit; | |
2600 | struct res_counter *fail_res; | |
2601 | unsigned long flags = 0; | |
2602 | int ret; | |
2603 | ||
2604 | ret = res_counter_charge(&memcg->res, csize, &fail_res); | |
2605 | ||
2606 | if (likely(!ret)) { | |
2607 | if (!do_swap_account) | |
2608 | return CHARGE_OK; | |
2609 | ret = res_counter_charge(&memcg->memsw, csize, &fail_res); | |
2610 | if (likely(!ret)) | |
2611 | return CHARGE_OK; | |
2612 | ||
2613 | res_counter_uncharge(&memcg->res, csize); | |
2614 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); | |
2615 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; | |
2616 | } else | |
2617 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); | |
2618 | /* | |
2619 | * Never reclaim on behalf of optional batching, retry with a | |
2620 | * single page instead. | |
2621 | */ | |
2622 | if (nr_pages > min_pages) | |
2623 | return CHARGE_RETRY; | |
2624 | ||
2625 | if (!(gfp_mask & __GFP_WAIT)) | |
2626 | return CHARGE_WOULDBLOCK; | |
2627 | ||
2628 | if (gfp_mask & __GFP_NORETRY) | |
2629 | return CHARGE_NOMEM; | |
2630 | ||
2631 | ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); | |
2632 | if (mem_cgroup_margin(mem_over_limit) >= nr_pages) | |
2633 | return CHARGE_RETRY; | |
2634 | /* | |
2635 | * Even though the limit is exceeded at this point, reclaim | |
2636 | * may have been able to free some pages. Retry the charge | |
2637 | * before killing the task. | |
2638 | * | |
2639 | * Only for regular pages, though: huge pages are rather | |
2640 | * unlikely to succeed so close to the limit, and we fall back | |
2641 | * to regular pages anyway in case of failure. | |
2642 | */ | |
2643 | if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret) | |
2644 | return CHARGE_RETRY; | |
2645 | ||
2646 | /* | |
2647 | * At task move, charge accounts can be doubly counted. So, it's | |
2648 | * better to wait until the end of task_move if something is going on. | |
2649 | */ | |
2650 | if (mem_cgroup_wait_acct_move(mem_over_limit)) | |
2651 | return CHARGE_RETRY; | |
2652 | ||
2653 | if (invoke_oom) | |
2654 | mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize)); | |
2655 | ||
2656 | return CHARGE_NOMEM; | |
2657 | } | |
2658 | ||
2659 | /** | |
2660 | * mem_cgroup_try_charge - try charging a memcg | |
2661 | * @memcg: memcg to charge | |
2662 | * @nr_pages: number of pages to charge | |
2663 | * @oom: trigger OOM if reclaim fails | |
2664 | * | |
2665 | * Returns 0 if @memcg was charged successfully, -EINTR if the charge | |
2666 | * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed. | |
2667 | */ | |
2668 | static int mem_cgroup_try_charge(struct mem_cgroup *memcg, | |
2669 | gfp_t gfp_mask, | |
2670 | unsigned int nr_pages, | |
2671 | bool oom) | |
2672 | { | |
2673 | unsigned int batch = max(CHARGE_BATCH, nr_pages); | |
2674 | int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
2675 | int ret; | |
2676 | ||
2677 | if (mem_cgroup_is_root(memcg)) | |
2678 | goto done; | |
2679 | /* | |
2680 | * Unlike in global OOM situations, memcg is not in a physical | |
2681 | * memory shortage. Allow dying and OOM-killed tasks to | |
2682 | * bypass the last charges so that they can exit quickly and | |
2683 | * free their memory. | |
2684 | */ | |
2685 | if (unlikely(test_thread_flag(TIF_MEMDIE) || | |
2686 | fatal_signal_pending(current) || | |
2687 | current->flags & PF_EXITING)) | |
2688 | goto bypass; | |
2689 | ||
2690 | if (unlikely(task_in_memcg_oom(current))) | |
2691 | goto nomem; | |
2692 | ||
2693 | if (gfp_mask & __GFP_NOFAIL) | |
2694 | oom = false; | |
2695 | again: | |
2696 | if (consume_stock(memcg, nr_pages)) | |
2697 | goto done; | |
2698 | ||
2699 | do { | |
2700 | bool invoke_oom = oom && !nr_oom_retries; | |
2701 | ||
2702 | /* If killed, bypass charge */ | |
2703 | if (fatal_signal_pending(current)) | |
2704 | goto bypass; | |
2705 | ||
2706 | ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, | |
2707 | nr_pages, invoke_oom); | |
2708 | switch (ret) { | |
2709 | case CHARGE_OK: | |
2710 | break; | |
2711 | case CHARGE_RETRY: /* not in OOM situation but retry */ | |
2712 | batch = nr_pages; | |
2713 | goto again; | |
2714 | case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ | |
2715 | goto nomem; | |
2716 | case CHARGE_NOMEM: /* OOM routine works */ | |
2717 | if (!oom || invoke_oom) | |
2718 | goto nomem; | |
2719 | nr_oom_retries--; | |
2720 | break; | |
2721 | } | |
2722 | } while (ret != CHARGE_OK); | |
2723 | ||
2724 | if (batch > nr_pages) | |
2725 | refill_stock(memcg, batch - nr_pages); | |
2726 | done: | |
2727 | return 0; | |
2728 | nomem: | |
2729 | if (!(gfp_mask & __GFP_NOFAIL)) | |
2730 | return -ENOMEM; | |
2731 | bypass: | |
2732 | return -EINTR; | |
2733 | } | |
2734 | ||
2735 | /** | |
2736 | * mem_cgroup_try_charge_mm - try charging a mm | |
2737 | * @mm: mm_struct to charge | |
2738 | * @nr_pages: number of pages to charge | |
2739 | * @oom: trigger OOM if reclaim fails | |
2740 | * | |
2741 | * Returns the charged mem_cgroup associated with the given mm_struct or | |
2742 | * NULL the charge failed. | |
2743 | */ | |
2744 | static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm, | |
2745 | gfp_t gfp_mask, | |
2746 | unsigned int nr_pages, | |
2747 | bool oom) | |
2748 | ||
2749 | { | |
2750 | struct mem_cgroup *memcg; | |
2751 | int ret; | |
2752 | ||
2753 | memcg = get_mem_cgroup_from_mm(mm); | |
2754 | ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom); | |
2755 | css_put(&memcg->css); | |
2756 | if (ret == -EINTR) | |
2757 | memcg = root_mem_cgroup; | |
2758 | else if (ret) | |
2759 | memcg = NULL; | |
2760 | ||
2761 | return memcg; | |
2762 | } | |
2763 | ||
2764 | /* | |
2765 | * Somemtimes we have to undo a charge we got by try_charge(). | |
2766 | * This function is for that and do uncharge, put css's refcnt. | |
2767 | * gotten by try_charge(). | |
2768 | */ | |
2769 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, | |
2770 | unsigned int nr_pages) | |
2771 | { | |
2772 | if (!mem_cgroup_is_root(memcg)) { | |
2773 | unsigned long bytes = nr_pages * PAGE_SIZE; | |
2774 | ||
2775 | res_counter_uncharge(&memcg->res, bytes); | |
2776 | if (do_swap_account) | |
2777 | res_counter_uncharge(&memcg->memsw, bytes); | |
2778 | } | |
2779 | } | |
2780 | ||
2781 | /* | |
2782 | * Cancel chrages in this cgroup....doesn't propagate to parent cgroup. | |
2783 | * This is useful when moving usage to parent cgroup. | |
2784 | */ | |
2785 | static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg, | |
2786 | unsigned int nr_pages) | |
2787 | { | |
2788 | unsigned long bytes = nr_pages * PAGE_SIZE; | |
2789 | ||
2790 | if (mem_cgroup_is_root(memcg)) | |
2791 | return; | |
2792 | ||
2793 | res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes); | |
2794 | if (do_swap_account) | |
2795 | res_counter_uncharge_until(&memcg->memsw, | |
2796 | memcg->memsw.parent, bytes); | |
2797 | } | |
2798 | ||
2799 | /* | |
2800 | * A helper function to get mem_cgroup from ID. must be called under | |
2801 | * rcu_read_lock(). The caller is responsible for calling css_tryget if | |
2802 | * the mem_cgroup is used for charging. (dropping refcnt from swap can be | |
2803 | * called against removed memcg.) | |
2804 | */ | |
2805 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) | |
2806 | { | |
2807 | /* ID 0 is unused ID */ | |
2808 | if (!id) | |
2809 | return NULL; | |
2810 | return mem_cgroup_from_id(id); | |
2811 | } | |
2812 | ||
2813 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) | |
2814 | { | |
2815 | struct mem_cgroup *memcg = NULL; | |
2816 | struct page_cgroup *pc; | |
2817 | unsigned short id; | |
2818 | swp_entry_t ent; | |
2819 | ||
2820 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
2821 | ||
2822 | pc = lookup_page_cgroup(page); | |
2823 | lock_page_cgroup(pc); | |
2824 | if (PageCgroupUsed(pc)) { | |
2825 | memcg = pc->mem_cgroup; | |
2826 | if (memcg && !css_tryget(&memcg->css)) | |
2827 | memcg = NULL; | |
2828 | } else if (PageSwapCache(page)) { | |
2829 | ent.val = page_private(page); | |
2830 | id = lookup_swap_cgroup_id(ent); | |
2831 | rcu_read_lock(); | |
2832 | memcg = mem_cgroup_lookup(id); | |
2833 | if (memcg && !css_tryget(&memcg->css)) | |
2834 | memcg = NULL; | |
2835 | rcu_read_unlock(); | |
2836 | } | |
2837 | unlock_page_cgroup(pc); | |
2838 | return memcg; | |
2839 | } | |
2840 | ||
2841 | static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, | |
2842 | struct page *page, | |
2843 | unsigned int nr_pages, | |
2844 | enum charge_type ctype, | |
2845 | bool lrucare) | |
2846 | { | |
2847 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
2848 | struct zone *uninitialized_var(zone); | |
2849 | struct lruvec *lruvec; | |
2850 | bool was_on_lru = false; | |
2851 | bool anon; | |
2852 | ||
2853 | lock_page_cgroup(pc); | |
2854 | VM_BUG_ON_PAGE(PageCgroupUsed(pc), page); | |
2855 | /* | |
2856 | * we don't need page_cgroup_lock about tail pages, becase they are not | |
2857 | * accessed by any other context at this point. | |
2858 | */ | |
2859 | ||
2860 | /* | |
2861 | * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page | |
2862 | * may already be on some other mem_cgroup's LRU. Take care of it. | |
2863 | */ | |
2864 | if (lrucare) { | |
2865 | zone = page_zone(page); | |
2866 | spin_lock_irq(&zone->lru_lock); | |
2867 | if (PageLRU(page)) { | |
2868 | lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup); | |
2869 | ClearPageLRU(page); | |
2870 | del_page_from_lru_list(page, lruvec, page_lru(page)); | |
2871 | was_on_lru = true; | |
2872 | } | |
2873 | } | |
2874 | ||
2875 | pc->mem_cgroup = memcg; | |
2876 | /* | |
2877 | * We access a page_cgroup asynchronously without lock_page_cgroup(). | |
2878 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup | |
2879 | * is accessed after testing USED bit. To make pc->mem_cgroup visible | |
2880 | * before USED bit, we need memory barrier here. | |
2881 | * See mem_cgroup_add_lru_list(), etc. | |
2882 | */ | |
2883 | smp_wmb(); | |
2884 | SetPageCgroupUsed(pc); | |
2885 | ||
2886 | if (lrucare) { | |
2887 | if (was_on_lru) { | |
2888 | lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup); | |
2889 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
2890 | SetPageLRU(page); | |
2891 | add_page_to_lru_list(page, lruvec, page_lru(page)); | |
2892 | } | |
2893 | spin_unlock_irq(&zone->lru_lock); | |
2894 | } | |
2895 | ||
2896 | if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON) | |
2897 | anon = true; | |
2898 | else | |
2899 | anon = false; | |
2900 | ||
2901 | mem_cgroup_charge_statistics(memcg, page, anon, nr_pages); | |
2902 | unlock_page_cgroup(pc); | |
2903 | ||
2904 | /* | |
2905 | * "charge_statistics" updated event counter. Then, check it. | |
2906 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. | |
2907 | * if they exceeds softlimit. | |
2908 | */ | |
2909 | memcg_check_events(memcg, page); | |
2910 | } | |
2911 | ||
2912 | static DEFINE_MUTEX(set_limit_mutex); | |
2913 | ||
2914 | #ifdef CONFIG_MEMCG_KMEM | |
2915 | /* | |
2916 | * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or | |
2917 | * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists. | |
2918 | */ | |
2919 | static DEFINE_MUTEX(memcg_slab_mutex); | |
2920 | ||
2921 | static DEFINE_MUTEX(activate_kmem_mutex); | |
2922 | ||
2923 | static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg) | |
2924 | { | |
2925 | return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) && | |
2926 | memcg_kmem_is_active(memcg); | |
2927 | } | |
2928 | ||
2929 | /* | |
2930 | * This is a bit cumbersome, but it is rarely used and avoids a backpointer | |
2931 | * in the memcg_cache_params struct. | |
2932 | */ | |
2933 | static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p) | |
2934 | { | |
2935 | struct kmem_cache *cachep; | |
2936 | ||
2937 | VM_BUG_ON(p->is_root_cache); | |
2938 | cachep = p->root_cache; | |
2939 | return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg)); | |
2940 | } | |
2941 | ||
2942 | #ifdef CONFIG_SLABINFO | |
2943 | static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v) | |
2944 | { | |
2945 | struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m)); | |
2946 | struct memcg_cache_params *params; | |
2947 | ||
2948 | if (!memcg_can_account_kmem(memcg)) | |
2949 | return -EIO; | |
2950 | ||
2951 | print_slabinfo_header(m); | |
2952 | ||
2953 | mutex_lock(&memcg_slab_mutex); | |
2954 | list_for_each_entry(params, &memcg->memcg_slab_caches, list) | |
2955 | cache_show(memcg_params_to_cache(params), m); | |
2956 | mutex_unlock(&memcg_slab_mutex); | |
2957 | ||
2958 | return 0; | |
2959 | } | |
2960 | #endif | |
2961 | ||
2962 | static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size) | |
2963 | { | |
2964 | struct res_counter *fail_res; | |
2965 | int ret = 0; | |
2966 | ||
2967 | ret = res_counter_charge(&memcg->kmem, size, &fail_res); | |
2968 | if (ret) | |
2969 | return ret; | |
2970 | ||
2971 | ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT, | |
2972 | oom_gfp_allowed(gfp)); | |
2973 | if (ret == -EINTR) { | |
2974 | /* | |
2975 | * mem_cgroup_try_charge() chosed to bypass to root due to | |
2976 | * OOM kill or fatal signal. Since our only options are to | |
2977 | * either fail the allocation or charge it to this cgroup, do | |
2978 | * it as a temporary condition. But we can't fail. From a | |
2979 | * kmem/slab perspective, the cache has already been selected, | |
2980 | * by mem_cgroup_kmem_get_cache(), so it is too late to change | |
2981 | * our minds. | |
2982 | * | |
2983 | * This condition will only trigger if the task entered | |
2984 | * memcg_charge_kmem in a sane state, but was OOM-killed during | |
2985 | * mem_cgroup_try_charge() above. Tasks that were already | |
2986 | * dying when the allocation triggers should have been already | |
2987 | * directed to the root cgroup in memcontrol.h | |
2988 | */ | |
2989 | res_counter_charge_nofail(&memcg->res, size, &fail_res); | |
2990 | if (do_swap_account) | |
2991 | res_counter_charge_nofail(&memcg->memsw, size, | |
2992 | &fail_res); | |
2993 | ret = 0; | |
2994 | } else if (ret) | |
2995 | res_counter_uncharge(&memcg->kmem, size); | |
2996 | ||
2997 | return ret; | |
2998 | } | |
2999 | ||
3000 | static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size) | |
3001 | { | |
3002 | res_counter_uncharge(&memcg->res, size); | |
3003 | if (do_swap_account) | |
3004 | res_counter_uncharge(&memcg->memsw, size); | |
3005 | ||
3006 | /* Not down to 0 */ | |
3007 | if (res_counter_uncharge(&memcg->kmem, size)) | |
3008 | return; | |
3009 | ||
3010 | /* | |
3011 | * Releases a reference taken in kmem_cgroup_css_offline in case | |
3012 | * this last uncharge is racing with the offlining code or it is | |
3013 | * outliving the memcg existence. | |
3014 | * | |
3015 | * The memory barrier imposed by test&clear is paired with the | |
3016 | * explicit one in memcg_kmem_mark_dead(). | |
3017 | */ | |
3018 | if (memcg_kmem_test_and_clear_dead(memcg)) | |
3019 | css_put(&memcg->css); | |
3020 | } | |
3021 | ||
3022 | /* | |
3023 | * helper for acessing a memcg's index. It will be used as an index in the | |
3024 | * child cache array in kmem_cache, and also to derive its name. This function | |
3025 | * will return -1 when this is not a kmem-limited memcg. | |
3026 | */ | |
3027 | int memcg_cache_id(struct mem_cgroup *memcg) | |
3028 | { | |
3029 | return memcg ? memcg->kmemcg_id : -1; | |
3030 | } | |
3031 | ||
3032 | static size_t memcg_caches_array_size(int num_groups) | |
3033 | { | |
3034 | ssize_t size; | |
3035 | if (num_groups <= 0) | |
3036 | return 0; | |
3037 | ||
3038 | size = 2 * num_groups; | |
3039 | if (size < MEMCG_CACHES_MIN_SIZE) | |
3040 | size = MEMCG_CACHES_MIN_SIZE; | |
3041 | else if (size > MEMCG_CACHES_MAX_SIZE) | |
3042 | size = MEMCG_CACHES_MAX_SIZE; | |
3043 | ||
3044 | return size; | |
3045 | } | |
3046 | ||
3047 | /* | |
3048 | * We should update the current array size iff all caches updates succeed. This | |
3049 | * can only be done from the slab side. The slab mutex needs to be held when | |
3050 | * calling this. | |
3051 | */ | |
3052 | void memcg_update_array_size(int num) | |
3053 | { | |
3054 | if (num > memcg_limited_groups_array_size) | |
3055 | memcg_limited_groups_array_size = memcg_caches_array_size(num); | |
3056 | } | |
3057 | ||
3058 | int memcg_update_cache_size(struct kmem_cache *s, int num_groups) | |
3059 | { | |
3060 | struct memcg_cache_params *cur_params = s->memcg_params; | |
3061 | ||
3062 | VM_BUG_ON(!is_root_cache(s)); | |
3063 | ||
3064 | if (num_groups > memcg_limited_groups_array_size) { | |
3065 | int i; | |
3066 | struct memcg_cache_params *new_params; | |
3067 | ssize_t size = memcg_caches_array_size(num_groups); | |
3068 | ||
3069 | size *= sizeof(void *); | |
3070 | size += offsetof(struct memcg_cache_params, memcg_caches); | |
3071 | ||
3072 | new_params = kzalloc(size, GFP_KERNEL); | |
3073 | if (!new_params) | |
3074 | return -ENOMEM; | |
3075 | ||
3076 | new_params->is_root_cache = true; | |
3077 | ||
3078 | /* | |
3079 | * There is the chance it will be bigger than | |
3080 | * memcg_limited_groups_array_size, if we failed an allocation | |
3081 | * in a cache, in which case all caches updated before it, will | |
3082 | * have a bigger array. | |
3083 | * | |
3084 | * But if that is the case, the data after | |
3085 | * memcg_limited_groups_array_size is certainly unused | |
3086 | */ | |
3087 | for (i = 0; i < memcg_limited_groups_array_size; i++) { | |
3088 | if (!cur_params->memcg_caches[i]) | |
3089 | continue; | |
3090 | new_params->memcg_caches[i] = | |
3091 | cur_params->memcg_caches[i]; | |
3092 | } | |
3093 | ||
3094 | /* | |
3095 | * Ideally, we would wait until all caches succeed, and only | |
3096 | * then free the old one. But this is not worth the extra | |
3097 | * pointer per-cache we'd have to have for this. | |
3098 | * | |
3099 | * It is not a big deal if some caches are left with a size | |
3100 | * bigger than the others. And all updates will reset this | |
3101 | * anyway. | |
3102 | */ | |
3103 | rcu_assign_pointer(s->memcg_params, new_params); | |
3104 | if (cur_params) | |
3105 | kfree_rcu(cur_params, rcu_head); | |
3106 | } | |
3107 | return 0; | |
3108 | } | |
3109 | ||
3110 | char *memcg_create_cache_name(struct mem_cgroup *memcg, | |
3111 | struct kmem_cache *root_cache) | |
3112 | { | |
3113 | static char *buf; | |
3114 | ||
3115 | /* | |
3116 | * We need a mutex here to protect the shared buffer. Since this is | |
3117 | * expected to be called only on cache creation, we can employ the | |
3118 | * slab_mutex for that purpose. | |
3119 | */ | |
3120 | lockdep_assert_held(&slab_mutex); | |
3121 | ||
3122 | if (!buf) { | |
3123 | buf = kmalloc(NAME_MAX + 1, GFP_KERNEL); | |
3124 | if (!buf) | |
3125 | return NULL; | |
3126 | } | |
3127 | ||
3128 | cgroup_name(memcg->css.cgroup, buf, NAME_MAX + 1); | |
3129 | return kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name, | |
3130 | memcg_cache_id(memcg), buf); | |
3131 | } | |
3132 | ||
3133 | int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s, | |
3134 | struct kmem_cache *root_cache) | |
3135 | { | |
3136 | size_t size; | |
3137 | ||
3138 | if (!memcg_kmem_enabled()) | |
3139 | return 0; | |
3140 | ||
3141 | if (!memcg) { | |
3142 | size = offsetof(struct memcg_cache_params, memcg_caches); | |
3143 | size += memcg_limited_groups_array_size * sizeof(void *); | |
3144 | } else | |
3145 | size = sizeof(struct memcg_cache_params); | |
3146 | ||
3147 | s->memcg_params = kzalloc(size, GFP_KERNEL); | |
3148 | if (!s->memcg_params) | |
3149 | return -ENOMEM; | |
3150 | ||
3151 | if (memcg) { | |
3152 | s->memcg_params->memcg = memcg; | |
3153 | s->memcg_params->root_cache = root_cache; | |
3154 | css_get(&memcg->css); | |
3155 | } else | |
3156 | s->memcg_params->is_root_cache = true; | |
3157 | ||
3158 | return 0; | |
3159 | } | |
3160 | ||
3161 | void memcg_free_cache_params(struct kmem_cache *s) | |
3162 | { | |
3163 | if (!s->memcg_params) | |
3164 | return; | |
3165 | if (!s->memcg_params->is_root_cache) | |
3166 | css_put(&s->memcg_params->memcg->css); | |
3167 | kfree(s->memcg_params); | |
3168 | } | |
3169 | ||
3170 | static void memcg_kmem_create_cache(struct mem_cgroup *memcg, | |
3171 | struct kmem_cache *root_cache) | |
3172 | { | |
3173 | struct kmem_cache *cachep; | |
3174 | int id; | |
3175 | ||
3176 | lockdep_assert_held(&memcg_slab_mutex); | |
3177 | ||
3178 | id = memcg_cache_id(memcg); | |
3179 | ||
3180 | /* | |
3181 | * Since per-memcg caches are created asynchronously on first | |
3182 | * allocation (see memcg_kmem_get_cache()), several threads can try to | |
3183 | * create the same cache, but only one of them may succeed. | |
3184 | */ | |
3185 | if (cache_from_memcg_idx(root_cache, id)) | |
3186 | return; | |
3187 | ||
3188 | cachep = kmem_cache_create_memcg(memcg, root_cache); | |
3189 | /* | |
3190 | * If we could not create a memcg cache, do not complain, because | |
3191 | * that's not critical at all as we can always proceed with the root | |
3192 | * cache. | |
3193 | */ | |
3194 | if (!cachep) | |
3195 | return; | |
3196 | ||
3197 | list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches); | |
3198 | ||
3199 | /* | |
3200 | * Since readers won't lock (see cache_from_memcg_idx()), we need a | |
3201 | * barrier here to ensure nobody will see the kmem_cache partially | |
3202 | * initialized. | |
3203 | */ | |
3204 | smp_wmb(); | |
3205 | ||
3206 | BUG_ON(root_cache->memcg_params->memcg_caches[id]); | |
3207 | root_cache->memcg_params->memcg_caches[id] = cachep; | |
3208 | } | |
3209 | ||
3210 | static void memcg_kmem_destroy_cache(struct kmem_cache *cachep) | |
3211 | { | |
3212 | struct kmem_cache *root_cache; | |
3213 | struct mem_cgroup *memcg; | |
3214 | int id; | |
3215 | ||
3216 | lockdep_assert_held(&memcg_slab_mutex); | |
3217 | ||
3218 | BUG_ON(is_root_cache(cachep)); | |
3219 | ||
3220 | root_cache = cachep->memcg_params->root_cache; | |
3221 | memcg = cachep->memcg_params->memcg; | |
3222 | id = memcg_cache_id(memcg); | |
3223 | ||
3224 | BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep); | |
3225 | root_cache->memcg_params->memcg_caches[id] = NULL; | |
3226 | ||
3227 | list_del(&cachep->memcg_params->list); | |
3228 | ||
3229 | kmem_cache_destroy(cachep); | |
3230 | } | |
3231 | ||
3232 | /* | |
3233 | * During the creation a new cache, we need to disable our accounting mechanism | |
3234 | * altogether. This is true even if we are not creating, but rather just | |
3235 | * enqueing new caches to be created. | |
3236 | * | |
3237 | * This is because that process will trigger allocations; some visible, like | |
3238 | * explicit kmallocs to auxiliary data structures, name strings and internal | |
3239 | * cache structures; some well concealed, like INIT_WORK() that can allocate | |
3240 | * objects during debug. | |
3241 | * | |
3242 | * If any allocation happens during memcg_kmem_get_cache, we will recurse back | |
3243 | * to it. This may not be a bounded recursion: since the first cache creation | |
3244 | * failed to complete (waiting on the allocation), we'll just try to create the | |
3245 | * cache again, failing at the same point. | |
3246 | * | |
3247 | * memcg_kmem_get_cache is prepared to abort after seeing a positive count of | |
3248 | * memcg_kmem_skip_account. So we enclose anything that might allocate memory | |
3249 | * inside the following two functions. | |
3250 | */ | |
3251 | static inline void memcg_stop_kmem_account(void) | |
3252 | { | |
3253 | VM_BUG_ON(!current->mm); | |
3254 | current->memcg_kmem_skip_account++; | |
3255 | } | |
3256 | ||
3257 | static inline void memcg_resume_kmem_account(void) | |
3258 | { | |
3259 | VM_BUG_ON(!current->mm); | |
3260 | current->memcg_kmem_skip_account--; | |
3261 | } | |
3262 | ||
3263 | int __kmem_cache_destroy_memcg_children(struct kmem_cache *s) | |
3264 | { | |
3265 | struct kmem_cache *c; | |
3266 | int i, failed = 0; | |
3267 | ||
3268 | mutex_lock(&memcg_slab_mutex); | |
3269 | for_each_memcg_cache_index(i) { | |
3270 | c = cache_from_memcg_idx(s, i); | |
3271 | if (!c) | |
3272 | continue; | |
3273 | ||
3274 | memcg_kmem_destroy_cache(c); | |
3275 | ||
3276 | if (cache_from_memcg_idx(s, i)) | |
3277 | failed++; | |
3278 | } | |
3279 | mutex_unlock(&memcg_slab_mutex); | |
3280 | return failed; | |
3281 | } | |
3282 | ||
3283 | static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg) | |
3284 | { | |
3285 | struct kmem_cache *cachep; | |
3286 | struct memcg_cache_params *params, *tmp; | |
3287 | ||
3288 | if (!memcg_kmem_is_active(memcg)) | |
3289 | return; | |
3290 | ||
3291 | mutex_lock(&memcg_slab_mutex); | |
3292 | list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) { | |
3293 | cachep = memcg_params_to_cache(params); | |
3294 | kmem_cache_shrink(cachep); | |
3295 | if (atomic_read(&cachep->memcg_params->nr_pages) == 0) | |
3296 | memcg_kmem_destroy_cache(cachep); | |
3297 | } | |
3298 | mutex_unlock(&memcg_slab_mutex); | |
3299 | } | |
3300 | ||
3301 | struct create_work { | |
3302 | struct mem_cgroup *memcg; | |
3303 | struct kmem_cache *cachep; | |
3304 | struct work_struct work; | |
3305 | }; | |
3306 | ||
3307 | static void memcg_create_cache_work_func(struct work_struct *w) | |
3308 | { | |
3309 | struct create_work *cw = container_of(w, struct create_work, work); | |
3310 | struct mem_cgroup *memcg = cw->memcg; | |
3311 | struct kmem_cache *cachep = cw->cachep; | |
3312 | ||
3313 | mutex_lock(&memcg_slab_mutex); | |
3314 | memcg_kmem_create_cache(memcg, cachep); | |
3315 | mutex_unlock(&memcg_slab_mutex); | |
3316 | ||
3317 | css_put(&memcg->css); | |
3318 | kfree(cw); | |
3319 | } | |
3320 | ||
3321 | /* | |
3322 | * Enqueue the creation of a per-memcg kmem_cache. | |
3323 | */ | |
3324 | static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg, | |
3325 | struct kmem_cache *cachep) | |
3326 | { | |
3327 | struct create_work *cw; | |
3328 | ||
3329 | cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT); | |
3330 | if (cw == NULL) { | |
3331 | css_put(&memcg->css); | |
3332 | return; | |
3333 | } | |
3334 | ||
3335 | cw->memcg = memcg; | |
3336 | cw->cachep = cachep; | |
3337 | ||
3338 | INIT_WORK(&cw->work, memcg_create_cache_work_func); | |
3339 | schedule_work(&cw->work); | |
3340 | } | |
3341 | ||
3342 | static void memcg_create_cache_enqueue(struct mem_cgroup *memcg, | |
3343 | struct kmem_cache *cachep) | |
3344 | { | |
3345 | /* | |
3346 | * We need to stop accounting when we kmalloc, because if the | |
3347 | * corresponding kmalloc cache is not yet created, the first allocation | |
3348 | * in __memcg_create_cache_enqueue will recurse. | |
3349 | * | |
3350 | * However, it is better to enclose the whole function. Depending on | |
3351 | * the debugging options enabled, INIT_WORK(), for instance, can | |
3352 | * trigger an allocation. This too, will make us recurse. Because at | |
3353 | * this point we can't allow ourselves back into memcg_kmem_get_cache, | |
3354 | * the safest choice is to do it like this, wrapping the whole function. | |
3355 | */ | |
3356 | memcg_stop_kmem_account(); | |
3357 | __memcg_create_cache_enqueue(memcg, cachep); | |
3358 | memcg_resume_kmem_account(); | |
3359 | } | |
3360 | ||
3361 | int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order) | |
3362 | { | |
3363 | int res; | |
3364 | ||
3365 | res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, | |
3366 | PAGE_SIZE << order); | |
3367 | if (!res) | |
3368 | atomic_add(1 << order, &cachep->memcg_params->nr_pages); | |
3369 | return res; | |
3370 | } | |
3371 | ||
3372 | void __memcg_uncharge_slab(struct kmem_cache *cachep, int order) | |
3373 | { | |
3374 | memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order); | |
3375 | atomic_sub(1 << order, &cachep->memcg_params->nr_pages); | |
3376 | } | |
3377 | ||
3378 | /* | |
3379 | * Return the kmem_cache we're supposed to use for a slab allocation. | |
3380 | * We try to use the current memcg's version of the cache. | |
3381 | * | |
3382 | * If the cache does not exist yet, if we are the first user of it, | |
3383 | * we either create it immediately, if possible, or create it asynchronously | |
3384 | * in a workqueue. | |
3385 | * In the latter case, we will let the current allocation go through with | |
3386 | * the original cache. | |
3387 | * | |
3388 | * Can't be called in interrupt context or from kernel threads. | |
3389 | * This function needs to be called with rcu_read_lock() held. | |
3390 | */ | |
3391 | struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, | |
3392 | gfp_t gfp) | |
3393 | { | |
3394 | struct mem_cgroup *memcg; | |
3395 | struct kmem_cache *memcg_cachep; | |
3396 | ||
3397 | VM_BUG_ON(!cachep->memcg_params); | |
3398 | VM_BUG_ON(!cachep->memcg_params->is_root_cache); | |
3399 | ||
3400 | if (!current->mm || current->memcg_kmem_skip_account) | |
3401 | return cachep; | |
3402 | ||
3403 | rcu_read_lock(); | |
3404 | memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner)); | |
3405 | ||
3406 | if (!memcg_can_account_kmem(memcg)) | |
3407 | goto out; | |
3408 | ||
3409 | memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg)); | |
3410 | if (likely(memcg_cachep)) { | |
3411 | cachep = memcg_cachep; | |
3412 | goto out; | |
3413 | } | |
3414 | ||
3415 | /* The corresponding put will be done in the workqueue. */ | |
3416 | if (!css_tryget(&memcg->css)) | |
3417 | goto out; | |
3418 | rcu_read_unlock(); | |
3419 | ||
3420 | /* | |
3421 | * If we are in a safe context (can wait, and not in interrupt | |
3422 | * context), we could be be predictable and return right away. | |
3423 | * This would guarantee that the allocation being performed | |
3424 | * already belongs in the new cache. | |
3425 | * | |
3426 | * However, there are some clashes that can arrive from locking. | |
3427 | * For instance, because we acquire the slab_mutex while doing | |
3428 | * kmem_cache_dup, this means no further allocation could happen | |
3429 | * with the slab_mutex held. | |
3430 | * | |
3431 | * Also, because cache creation issue get_online_cpus(), this | |
3432 | * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex, | |
3433 | * that ends up reversed during cpu hotplug. (cpuset allocates | |
3434 | * a bunch of GFP_KERNEL memory during cpuup). Due to all that, | |
3435 | * better to defer everything. | |
3436 | */ | |
3437 | memcg_create_cache_enqueue(memcg, cachep); | |
3438 | return cachep; | |
3439 | out: | |
3440 | rcu_read_unlock(); | |
3441 | return cachep; | |
3442 | } | |
3443 | ||
3444 | /* | |
3445 | * We need to verify if the allocation against current->mm->owner's memcg is | |
3446 | * possible for the given order. But the page is not allocated yet, so we'll | |
3447 | * need a further commit step to do the final arrangements. | |
3448 | * | |
3449 | * It is possible for the task to switch cgroups in this mean time, so at | |
3450 | * commit time, we can't rely on task conversion any longer. We'll then use | |
3451 | * the handle argument to return to the caller which cgroup we should commit | |
3452 | * against. We could also return the memcg directly and avoid the pointer | |
3453 | * passing, but a boolean return value gives better semantics considering | |
3454 | * the compiled-out case as well. | |
3455 | * | |
3456 | * Returning true means the allocation is possible. | |
3457 | */ | |
3458 | bool | |
3459 | __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order) | |
3460 | { | |
3461 | struct mem_cgroup *memcg; | |
3462 | int ret; | |
3463 | ||
3464 | *_memcg = NULL; | |
3465 | ||
3466 | /* | |
3467 | * Disabling accounting is only relevant for some specific memcg | |
3468 | * internal allocations. Therefore we would initially not have such | |
3469 | * check here, since direct calls to the page allocator that are | |
3470 | * accounted to kmemcg (alloc_kmem_pages and friends) only happen | |
3471 | * outside memcg core. We are mostly concerned with cache allocations, | |
3472 | * and by having this test at memcg_kmem_get_cache, we are already able | |
3473 | * to relay the allocation to the root cache and bypass the memcg cache | |
3474 | * altogether. | |
3475 | * | |
3476 | * There is one exception, though: the SLUB allocator does not create | |
3477 | * large order caches, but rather service large kmallocs directly from | |
3478 | * the page allocator. Therefore, the following sequence when backed by | |
3479 | * the SLUB allocator: | |
3480 | * | |
3481 | * memcg_stop_kmem_account(); | |
3482 | * kmalloc(<large_number>) | |
3483 | * memcg_resume_kmem_account(); | |
3484 | * | |
3485 | * would effectively ignore the fact that we should skip accounting, | |
3486 | * since it will drive us directly to this function without passing | |
3487 | * through the cache selector memcg_kmem_get_cache. Such large | |
3488 | * allocations are extremely rare but can happen, for instance, for the | |
3489 | * cache arrays. We bring this test here. | |
3490 | */ | |
3491 | if (!current->mm || current->memcg_kmem_skip_account) | |
3492 | return true; | |
3493 | ||
3494 | memcg = get_mem_cgroup_from_mm(current->mm); | |
3495 | ||
3496 | if (!memcg_can_account_kmem(memcg)) { | |
3497 | css_put(&memcg->css); | |
3498 | return true; | |
3499 | } | |
3500 | ||
3501 | ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order); | |
3502 | if (!ret) | |
3503 | *_memcg = memcg; | |
3504 | ||
3505 | css_put(&memcg->css); | |
3506 | return (ret == 0); | |
3507 | } | |
3508 | ||
3509 | void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg, | |
3510 | int order) | |
3511 | { | |
3512 | struct page_cgroup *pc; | |
3513 | ||
3514 | VM_BUG_ON(mem_cgroup_is_root(memcg)); | |
3515 | ||
3516 | /* The page allocation failed. Revert */ | |
3517 | if (!page) { | |
3518 | memcg_uncharge_kmem(memcg, PAGE_SIZE << order); | |
3519 | return; | |
3520 | } | |
3521 | ||
3522 | pc = lookup_page_cgroup(page); | |
3523 | lock_page_cgroup(pc); | |
3524 | pc->mem_cgroup = memcg; | |
3525 | SetPageCgroupUsed(pc); | |
3526 | unlock_page_cgroup(pc); | |
3527 | } | |
3528 | ||
3529 | void __memcg_kmem_uncharge_pages(struct page *page, int order) | |
3530 | { | |
3531 | struct mem_cgroup *memcg = NULL; | |
3532 | struct page_cgroup *pc; | |
3533 | ||
3534 | ||
3535 | pc = lookup_page_cgroup(page); | |
3536 | /* | |
3537 | * Fast unlocked return. Theoretically might have changed, have to | |
3538 | * check again after locking. | |
3539 | */ | |
3540 | if (!PageCgroupUsed(pc)) | |
3541 | return; | |
3542 | ||
3543 | lock_page_cgroup(pc); | |
3544 | if (PageCgroupUsed(pc)) { | |
3545 | memcg = pc->mem_cgroup; | |
3546 | ClearPageCgroupUsed(pc); | |
3547 | } | |
3548 | unlock_page_cgroup(pc); | |
3549 | ||
3550 | /* | |
3551 | * We trust that only if there is a memcg associated with the page, it | |
3552 | * is a valid allocation | |
3553 | */ | |
3554 | if (!memcg) | |
3555 | return; | |
3556 | ||
3557 | VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page); | |
3558 | memcg_uncharge_kmem(memcg, PAGE_SIZE << order); | |
3559 | } | |
3560 | #else | |
3561 | static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg) | |
3562 | { | |
3563 | } | |
3564 | #endif /* CONFIG_MEMCG_KMEM */ | |
3565 | ||
3566 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
3567 | ||
3568 | #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION) | |
3569 | /* | |
3570 | * Because tail pages are not marked as "used", set it. We're under | |
3571 | * zone->lru_lock, 'splitting on pmd' and compound_lock. | |
3572 | * charge/uncharge will be never happen and move_account() is done under | |
3573 | * compound_lock(), so we don't have to take care of races. | |
3574 | */ | |
3575 | void mem_cgroup_split_huge_fixup(struct page *head) | |
3576 | { | |
3577 | struct page_cgroup *head_pc = lookup_page_cgroup(head); | |
3578 | struct page_cgroup *pc; | |
3579 | struct mem_cgroup *memcg; | |
3580 | int i; | |
3581 | ||
3582 | if (mem_cgroup_disabled()) | |
3583 | return; | |
3584 | ||
3585 | memcg = head_pc->mem_cgroup; | |
3586 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
3587 | pc = head_pc + i; | |
3588 | pc->mem_cgroup = memcg; | |
3589 | smp_wmb();/* see __commit_charge() */ | |
3590 | pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; | |
3591 | } | |
3592 | __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], | |
3593 | HPAGE_PMD_NR); | |
3594 | } | |
3595 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
3596 | ||
3597 | /** | |
3598 | * mem_cgroup_move_account - move account of the page | |
3599 | * @page: the page | |
3600 | * @nr_pages: number of regular pages (>1 for huge pages) | |
3601 | * @pc: page_cgroup of the page. | |
3602 | * @from: mem_cgroup which the page is moved from. | |
3603 | * @to: mem_cgroup which the page is moved to. @from != @to. | |
3604 | * | |
3605 | * The caller must confirm following. | |
3606 | * - page is not on LRU (isolate_page() is useful.) | |
3607 | * - compound_lock is held when nr_pages > 1 | |
3608 | * | |
3609 | * This function doesn't do "charge" to new cgroup and doesn't do "uncharge" | |
3610 | * from old cgroup. | |
3611 | */ | |
3612 | static int mem_cgroup_move_account(struct page *page, | |
3613 | unsigned int nr_pages, | |
3614 | struct page_cgroup *pc, | |
3615 | struct mem_cgroup *from, | |
3616 | struct mem_cgroup *to) | |
3617 | { | |
3618 | unsigned long flags; | |
3619 | int ret; | |
3620 | bool anon = PageAnon(page); | |
3621 | ||
3622 | VM_BUG_ON(from == to); | |
3623 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
3624 | /* | |
3625 | * The page is isolated from LRU. So, collapse function | |
3626 | * will not handle this page. But page splitting can happen. | |
3627 | * Do this check under compound_page_lock(). The caller should | |
3628 | * hold it. | |
3629 | */ | |
3630 | ret = -EBUSY; | |
3631 | if (nr_pages > 1 && !PageTransHuge(page)) | |
3632 | goto out; | |
3633 | ||
3634 | lock_page_cgroup(pc); | |
3635 | ||
3636 | ret = -EINVAL; | |
3637 | if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) | |
3638 | goto unlock; | |
3639 | ||
3640 | move_lock_mem_cgroup(from, &flags); | |
3641 | ||
3642 | if (!anon && page_mapped(page)) { | |
3643 | __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED], | |
3644 | nr_pages); | |
3645 | __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED], | |
3646 | nr_pages); | |
3647 | } | |
3648 | ||
3649 | if (PageWriteback(page)) { | |
3650 | __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK], | |
3651 | nr_pages); | |
3652 | __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK], | |
3653 | nr_pages); | |
3654 | } | |
3655 | ||
3656 | mem_cgroup_charge_statistics(from, page, anon, -nr_pages); | |
3657 | ||
3658 | /* caller should have done css_get */ | |
3659 | pc->mem_cgroup = to; | |
3660 | mem_cgroup_charge_statistics(to, page, anon, nr_pages); | |
3661 | move_unlock_mem_cgroup(from, &flags); | |
3662 | ret = 0; | |
3663 | unlock: | |
3664 | unlock_page_cgroup(pc); | |
3665 | /* | |
3666 | * check events | |
3667 | */ | |
3668 | memcg_check_events(to, page); | |
3669 | memcg_check_events(from, page); | |
3670 | out: | |
3671 | return ret; | |
3672 | } | |
3673 | ||
3674 | /** | |
3675 | * mem_cgroup_move_parent - moves page to the parent group | |
3676 | * @page: the page to move | |
3677 | * @pc: page_cgroup of the page | |
3678 | * @child: page's cgroup | |
3679 | * | |
3680 | * move charges to its parent or the root cgroup if the group has no | |
3681 | * parent (aka use_hierarchy==0). | |
3682 | * Although this might fail (get_page_unless_zero, isolate_lru_page or | |
3683 | * mem_cgroup_move_account fails) the failure is always temporary and | |
3684 | * it signals a race with a page removal/uncharge or migration. In the | |
3685 | * first case the page is on the way out and it will vanish from the LRU | |
3686 | * on the next attempt and the call should be retried later. | |
3687 | * Isolation from the LRU fails only if page has been isolated from | |
3688 | * the LRU since we looked at it and that usually means either global | |
3689 | * reclaim or migration going on. The page will either get back to the | |
3690 | * LRU or vanish. | |
3691 | * Finaly mem_cgroup_move_account fails only if the page got uncharged | |
3692 | * (!PageCgroupUsed) or moved to a different group. The page will | |
3693 | * disappear in the next attempt. | |
3694 | */ | |
3695 | static int mem_cgroup_move_parent(struct page *page, | |
3696 | struct page_cgroup *pc, | |
3697 | struct mem_cgroup *child) | |
3698 | { | |
3699 | struct mem_cgroup *parent; | |
3700 | unsigned int nr_pages; | |
3701 | unsigned long uninitialized_var(flags); | |
3702 | int ret; | |
3703 | ||
3704 | VM_BUG_ON(mem_cgroup_is_root(child)); | |
3705 | ||
3706 | ret = -EBUSY; | |
3707 | if (!get_page_unless_zero(page)) | |
3708 | goto out; | |
3709 | if (isolate_lru_page(page)) | |
3710 | goto put; | |
3711 | ||
3712 | nr_pages = hpage_nr_pages(page); | |
3713 | ||
3714 | parent = parent_mem_cgroup(child); | |
3715 | /* | |
3716 | * If no parent, move charges to root cgroup. | |
3717 | */ | |
3718 | if (!parent) | |
3719 | parent = root_mem_cgroup; | |
3720 | ||
3721 | if (nr_pages > 1) { | |
3722 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
3723 | flags = compound_lock_irqsave(page); | |
3724 | } | |
3725 | ||
3726 | ret = mem_cgroup_move_account(page, nr_pages, | |
3727 | pc, child, parent); | |
3728 | if (!ret) | |
3729 | __mem_cgroup_cancel_local_charge(child, nr_pages); | |
3730 | ||
3731 | if (nr_pages > 1) | |
3732 | compound_unlock_irqrestore(page, flags); | |
3733 | putback_lru_page(page); | |
3734 | put: | |
3735 | put_page(page); | |
3736 | out: | |
3737 | return ret; | |
3738 | } | |
3739 | ||
3740 | int mem_cgroup_charge_anon(struct page *page, | |
3741 | struct mm_struct *mm, gfp_t gfp_mask) | |
3742 | { | |
3743 | unsigned int nr_pages = 1; | |
3744 | struct mem_cgroup *memcg; | |
3745 | bool oom = true; | |
3746 | ||
3747 | if (mem_cgroup_disabled()) | |
3748 | return 0; | |
3749 | ||
3750 | VM_BUG_ON_PAGE(page_mapped(page), page); | |
3751 | VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page); | |
3752 | VM_BUG_ON(!mm); | |
3753 | ||
3754 | if (PageTransHuge(page)) { | |
3755 | nr_pages <<= compound_order(page); | |
3756 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
3757 | /* | |
3758 | * Never OOM-kill a process for a huge page. The | |
3759 | * fault handler will fall back to regular pages. | |
3760 | */ | |
3761 | oom = false; | |
3762 | } | |
3763 | ||
3764 | memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom); | |
3765 | if (!memcg) | |
3766 | return -ENOMEM; | |
3767 | __mem_cgroup_commit_charge(memcg, page, nr_pages, | |
3768 | MEM_CGROUP_CHARGE_TYPE_ANON, false); | |
3769 | return 0; | |
3770 | } | |
3771 | ||
3772 | /* | |
3773 | * While swap-in, try_charge -> commit or cancel, the page is locked. | |
3774 | * And when try_charge() successfully returns, one refcnt to memcg without | |
3775 | * struct page_cgroup is acquired. This refcnt will be consumed by | |
3776 | * "commit()" or removed by "cancel()" | |
3777 | */ | |
3778 | static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm, | |
3779 | struct page *page, | |
3780 | gfp_t mask, | |
3781 | struct mem_cgroup **memcgp) | |
3782 | { | |
3783 | struct mem_cgroup *memcg = NULL; | |
3784 | struct page_cgroup *pc; | |
3785 | int ret; | |
3786 | ||
3787 | pc = lookup_page_cgroup(page); | |
3788 | /* | |
3789 | * Every swap fault against a single page tries to charge the | |
3790 | * page, bail as early as possible. shmem_unuse() encounters | |
3791 | * already charged pages, too. The USED bit is protected by | |
3792 | * the page lock, which serializes swap cache removal, which | |
3793 | * in turn serializes uncharging. | |
3794 | */ | |
3795 | if (PageCgroupUsed(pc)) | |
3796 | goto out; | |
3797 | if (do_swap_account) | |
3798 | memcg = try_get_mem_cgroup_from_page(page); | |
3799 | if (!memcg) | |
3800 | memcg = get_mem_cgroup_from_mm(mm); | |
3801 | ret = mem_cgroup_try_charge(memcg, mask, 1, true); | |
3802 | css_put(&memcg->css); | |
3803 | if (ret == -EINTR) | |
3804 | memcg = root_mem_cgroup; | |
3805 | else if (ret) | |
3806 | return ret; | |
3807 | out: | |
3808 | *memcgp = memcg; | |
3809 | return 0; | |
3810 | } | |
3811 | ||
3812 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page, | |
3813 | gfp_t gfp_mask, struct mem_cgroup **memcgp) | |
3814 | { | |
3815 | if (mem_cgroup_disabled()) { | |
3816 | *memcgp = NULL; | |
3817 | return 0; | |
3818 | } | |
3819 | /* | |
3820 | * A racing thread's fault, or swapoff, may have already | |
3821 | * updated the pte, and even removed page from swap cache: in | |
3822 | * those cases unuse_pte()'s pte_same() test will fail; but | |
3823 | * there's also a KSM case which does need to charge the page. | |
3824 | */ | |
3825 | if (!PageSwapCache(page)) { | |
3826 | struct mem_cgroup *memcg; | |
3827 | ||
3828 | memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true); | |
3829 | if (!memcg) | |
3830 | return -ENOMEM; | |
3831 | *memcgp = memcg; | |
3832 | return 0; | |
3833 | } | |
3834 | return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp); | |
3835 | } | |
3836 | ||
3837 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) | |
3838 | { | |
3839 | if (mem_cgroup_disabled()) | |
3840 | return; | |
3841 | if (!memcg) | |
3842 | return; | |
3843 | __mem_cgroup_cancel_charge(memcg, 1); | |
3844 | } | |
3845 | ||
3846 | static void | |
3847 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, | |
3848 | enum charge_type ctype) | |
3849 | { | |
3850 | if (mem_cgroup_disabled()) | |
3851 | return; | |
3852 | if (!memcg) | |
3853 | return; | |
3854 | ||
3855 | __mem_cgroup_commit_charge(memcg, page, 1, ctype, true); | |
3856 | /* | |
3857 | * Now swap is on-memory. This means this page may be | |
3858 | * counted both as mem and swap....double count. | |
3859 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable | |
3860 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() | |
3861 | * may call delete_from_swap_cache() before reach here. | |
3862 | */ | |
3863 | if (do_swap_account && PageSwapCache(page)) { | |
3864 | swp_entry_t ent = {.val = page_private(page)}; | |
3865 | mem_cgroup_uncharge_swap(ent); | |
3866 | } | |
3867 | } | |
3868 | ||
3869 | void mem_cgroup_commit_charge_swapin(struct page *page, | |
3870 | struct mem_cgroup *memcg) | |
3871 | { | |
3872 | __mem_cgroup_commit_charge_swapin(page, memcg, | |
3873 | MEM_CGROUP_CHARGE_TYPE_ANON); | |
3874 | } | |
3875 | ||
3876 | int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm, | |
3877 | gfp_t gfp_mask) | |
3878 | { | |
3879 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
3880 | struct mem_cgroup *memcg; | |
3881 | int ret; | |
3882 | ||
3883 | if (mem_cgroup_disabled()) | |
3884 | return 0; | |
3885 | if (PageCompound(page)) | |
3886 | return 0; | |
3887 | ||
3888 | if (PageSwapCache(page)) { /* shmem */ | |
3889 | ret = __mem_cgroup_try_charge_swapin(mm, page, | |
3890 | gfp_mask, &memcg); | |
3891 | if (ret) | |
3892 | return ret; | |
3893 | __mem_cgroup_commit_charge_swapin(page, memcg, type); | |
3894 | return 0; | |
3895 | } | |
3896 | ||
3897 | memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true); | |
3898 | if (!memcg) | |
3899 | return -ENOMEM; | |
3900 | __mem_cgroup_commit_charge(memcg, page, 1, type, false); | |
3901 | return 0; | |
3902 | } | |
3903 | ||
3904 | static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, | |
3905 | unsigned int nr_pages, | |
3906 | const enum charge_type ctype) | |
3907 | { | |
3908 | struct memcg_batch_info *batch = NULL; | |
3909 | bool uncharge_memsw = true; | |
3910 | ||
3911 | /* If swapout, usage of swap doesn't decrease */ | |
3912 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) | |
3913 | uncharge_memsw = false; | |
3914 | ||
3915 | batch = ¤t->memcg_batch; | |
3916 | /* | |
3917 | * In usual, we do css_get() when we remember memcg pointer. | |
3918 | * But in this case, we keep res->usage until end of a series of | |
3919 | * uncharges. Then, it's ok to ignore memcg's refcnt. | |
3920 | */ | |
3921 | if (!batch->memcg) | |
3922 | batch->memcg = memcg; | |
3923 | /* | |
3924 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. | |
3925 | * In those cases, all pages freed continuously can be expected to be in | |
3926 | * the same cgroup and we have chance to coalesce uncharges. | |
3927 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) | |
3928 | * because we want to do uncharge as soon as possible. | |
3929 | */ | |
3930 | ||
3931 | if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) | |
3932 | goto direct_uncharge; | |
3933 | ||
3934 | if (nr_pages > 1) | |
3935 | goto direct_uncharge; | |
3936 | ||
3937 | /* | |
3938 | * In typical case, batch->memcg == mem. This means we can | |
3939 | * merge a series of uncharges to an uncharge of res_counter. | |
3940 | * If not, we uncharge res_counter ony by one. | |
3941 | */ | |
3942 | if (batch->memcg != memcg) | |
3943 | goto direct_uncharge; | |
3944 | /* remember freed charge and uncharge it later */ | |
3945 | batch->nr_pages++; | |
3946 | if (uncharge_memsw) | |
3947 | batch->memsw_nr_pages++; | |
3948 | return; | |
3949 | direct_uncharge: | |
3950 | res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); | |
3951 | if (uncharge_memsw) | |
3952 | res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); | |
3953 | if (unlikely(batch->memcg != memcg)) | |
3954 | memcg_oom_recover(memcg); | |
3955 | } | |
3956 | ||
3957 | /* | |
3958 | * uncharge if !page_mapped(page) | |
3959 | */ | |
3960 | static struct mem_cgroup * | |
3961 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype, | |
3962 | bool end_migration) | |
3963 | { | |
3964 | struct mem_cgroup *memcg = NULL; | |
3965 | unsigned int nr_pages = 1; | |
3966 | struct page_cgroup *pc; | |
3967 | bool anon; | |
3968 | ||
3969 | if (mem_cgroup_disabled()) | |
3970 | return NULL; | |
3971 | ||
3972 | if (PageTransHuge(page)) { | |
3973 | nr_pages <<= compound_order(page); | |
3974 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
3975 | } | |
3976 | /* | |
3977 | * Check if our page_cgroup is valid | |
3978 | */ | |
3979 | pc = lookup_page_cgroup(page); | |
3980 | if (unlikely(!PageCgroupUsed(pc))) | |
3981 | return NULL; | |
3982 | ||
3983 | lock_page_cgroup(pc); | |
3984 | ||
3985 | memcg = pc->mem_cgroup; | |
3986 | ||
3987 | if (!PageCgroupUsed(pc)) | |
3988 | goto unlock_out; | |
3989 | ||
3990 | anon = PageAnon(page); | |
3991 | ||
3992 | switch (ctype) { | |
3993 | case MEM_CGROUP_CHARGE_TYPE_ANON: | |
3994 | /* | |
3995 | * Generally PageAnon tells if it's the anon statistics to be | |
3996 | * updated; but sometimes e.g. mem_cgroup_uncharge_page() is | |
3997 | * used before page reached the stage of being marked PageAnon. | |
3998 | */ | |
3999 | anon = true; | |
4000 | /* fallthrough */ | |
4001 | case MEM_CGROUP_CHARGE_TYPE_DROP: | |
4002 | /* See mem_cgroup_prepare_migration() */ | |
4003 | if (page_mapped(page)) | |
4004 | goto unlock_out; | |
4005 | /* | |
4006 | * Pages under migration may not be uncharged. But | |
4007 | * end_migration() /must/ be the one uncharging the | |
4008 | * unused post-migration page and so it has to call | |
4009 | * here with the migration bit still set. See the | |
4010 | * res_counter handling below. | |
4011 | */ | |
4012 | if (!end_migration && PageCgroupMigration(pc)) | |
4013 | goto unlock_out; | |
4014 | break; | |
4015 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: | |
4016 | if (!PageAnon(page)) { /* Shared memory */ | |
4017 | if (page->mapping && !page_is_file_cache(page)) | |
4018 | goto unlock_out; | |
4019 | } else if (page_mapped(page)) /* Anon */ | |
4020 | goto unlock_out; | |
4021 | break; | |
4022 | default: | |
4023 | break; | |
4024 | } | |
4025 | ||
4026 | mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages); | |
4027 | ||
4028 | ClearPageCgroupUsed(pc); | |
4029 | /* | |
4030 | * pc->mem_cgroup is not cleared here. It will be accessed when it's | |
4031 | * freed from LRU. This is safe because uncharged page is expected not | |
4032 | * to be reused (freed soon). Exception is SwapCache, it's handled by | |
4033 | * special functions. | |
4034 | */ | |
4035 | ||
4036 | unlock_page_cgroup(pc); | |
4037 | /* | |
4038 | * even after unlock, we have memcg->res.usage here and this memcg | |
4039 | * will never be freed, so it's safe to call css_get(). | |
4040 | */ | |
4041 | memcg_check_events(memcg, page); | |
4042 | if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { | |
4043 | mem_cgroup_swap_statistics(memcg, true); | |
4044 | css_get(&memcg->css); | |
4045 | } | |
4046 | /* | |
4047 | * Migration does not charge the res_counter for the | |
4048 | * replacement page, so leave it alone when phasing out the | |
4049 | * page that is unused after the migration. | |
4050 | */ | |
4051 | if (!end_migration && !mem_cgroup_is_root(memcg)) | |
4052 | mem_cgroup_do_uncharge(memcg, nr_pages, ctype); | |
4053 | ||
4054 | return memcg; | |
4055 | ||
4056 | unlock_out: | |
4057 | unlock_page_cgroup(pc); | |
4058 | return NULL; | |
4059 | } | |
4060 | ||
4061 | void mem_cgroup_uncharge_page(struct page *page) | |
4062 | { | |
4063 | /* early check. */ | |
4064 | if (page_mapped(page)) | |
4065 | return; | |
4066 | VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page); | |
4067 | /* | |
4068 | * If the page is in swap cache, uncharge should be deferred | |
4069 | * to the swap path, which also properly accounts swap usage | |
4070 | * and handles memcg lifetime. | |
4071 | * | |
4072 | * Note that this check is not stable and reclaim may add the | |
4073 | * page to swap cache at any time after this. However, if the | |
4074 | * page is not in swap cache by the time page->mapcount hits | |
4075 | * 0, there won't be any page table references to the swap | |
4076 | * slot, and reclaim will free it and not actually write the | |
4077 | * page to disk. | |
4078 | */ | |
4079 | if (PageSwapCache(page)) | |
4080 | return; | |
4081 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false); | |
4082 | } | |
4083 | ||
4084 | void mem_cgroup_uncharge_cache_page(struct page *page) | |
4085 | { | |
4086 | VM_BUG_ON_PAGE(page_mapped(page), page); | |
4087 | VM_BUG_ON_PAGE(page->mapping, page); | |
4088 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false); | |
4089 | } | |
4090 | ||
4091 | /* | |
4092 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. | |
4093 | * In that cases, pages are freed continuously and we can expect pages | |
4094 | * are in the same memcg. All these calls itself limits the number of | |
4095 | * pages freed at once, then uncharge_start/end() is called properly. | |
4096 | * This may be called prural(2) times in a context, | |
4097 | */ | |
4098 | ||
4099 | void mem_cgroup_uncharge_start(void) | |
4100 | { | |
4101 | current->memcg_batch.do_batch++; | |
4102 | /* We can do nest. */ | |
4103 | if (current->memcg_batch.do_batch == 1) { | |
4104 | current->memcg_batch.memcg = NULL; | |
4105 | current->memcg_batch.nr_pages = 0; | |
4106 | current->memcg_batch.memsw_nr_pages = 0; | |
4107 | } | |
4108 | } | |
4109 | ||
4110 | void mem_cgroup_uncharge_end(void) | |
4111 | { | |
4112 | struct memcg_batch_info *batch = ¤t->memcg_batch; | |
4113 | ||
4114 | if (!batch->do_batch) | |
4115 | return; | |
4116 | ||
4117 | batch->do_batch--; | |
4118 | if (batch->do_batch) /* If stacked, do nothing. */ | |
4119 | return; | |
4120 | ||
4121 | if (!batch->memcg) | |
4122 | return; | |
4123 | /* | |
4124 | * This "batch->memcg" is valid without any css_get/put etc... | |
4125 | * bacause we hide charges behind us. | |
4126 | */ | |
4127 | if (batch->nr_pages) | |
4128 | res_counter_uncharge(&batch->memcg->res, | |
4129 | batch->nr_pages * PAGE_SIZE); | |
4130 | if (batch->memsw_nr_pages) | |
4131 | res_counter_uncharge(&batch->memcg->memsw, | |
4132 | batch->memsw_nr_pages * PAGE_SIZE); | |
4133 | memcg_oom_recover(batch->memcg); | |
4134 | /* forget this pointer (for sanity check) */ | |
4135 | batch->memcg = NULL; | |
4136 | } | |
4137 | ||
4138 | #ifdef CONFIG_SWAP | |
4139 | /* | |
4140 | * called after __delete_from_swap_cache() and drop "page" account. | |
4141 | * memcg information is recorded to swap_cgroup of "ent" | |
4142 | */ | |
4143 | void | |
4144 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) | |
4145 | { | |
4146 | struct mem_cgroup *memcg; | |
4147 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; | |
4148 | ||
4149 | if (!swapout) /* this was a swap cache but the swap is unused ! */ | |
4150 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; | |
4151 | ||
4152 | memcg = __mem_cgroup_uncharge_common(page, ctype, false); | |
4153 | ||
4154 | /* | |
4155 | * record memcg information, if swapout && memcg != NULL, | |
4156 | * css_get() was called in uncharge(). | |
4157 | */ | |
4158 | if (do_swap_account && swapout && memcg) | |
4159 | swap_cgroup_record(ent, mem_cgroup_id(memcg)); | |
4160 | } | |
4161 | #endif | |
4162 | ||
4163 | #ifdef CONFIG_MEMCG_SWAP | |
4164 | /* | |
4165 | * called from swap_entry_free(). remove record in swap_cgroup and | |
4166 | * uncharge "memsw" account. | |
4167 | */ | |
4168 | void mem_cgroup_uncharge_swap(swp_entry_t ent) | |
4169 | { | |
4170 | struct mem_cgroup *memcg; | |
4171 | unsigned short id; | |
4172 | ||
4173 | if (!do_swap_account) | |
4174 | return; | |
4175 | ||
4176 | id = swap_cgroup_record(ent, 0); | |
4177 | rcu_read_lock(); | |
4178 | memcg = mem_cgroup_lookup(id); | |
4179 | if (memcg) { | |
4180 | /* | |
4181 | * We uncharge this because swap is freed. | |
4182 | * This memcg can be obsolete one. We avoid calling css_tryget | |
4183 | */ | |
4184 | if (!mem_cgroup_is_root(memcg)) | |
4185 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); | |
4186 | mem_cgroup_swap_statistics(memcg, false); | |
4187 | css_put(&memcg->css); | |
4188 | } | |
4189 | rcu_read_unlock(); | |
4190 | } | |
4191 | ||
4192 | /** | |
4193 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. | |
4194 | * @entry: swap entry to be moved | |
4195 | * @from: mem_cgroup which the entry is moved from | |
4196 | * @to: mem_cgroup which the entry is moved to | |
4197 | * | |
4198 | * It succeeds only when the swap_cgroup's record for this entry is the same | |
4199 | * as the mem_cgroup's id of @from. | |
4200 | * | |
4201 | * Returns 0 on success, -EINVAL on failure. | |
4202 | * | |
4203 | * The caller must have charged to @to, IOW, called res_counter_charge() about | |
4204 | * both res and memsw, and called css_get(). | |
4205 | */ | |
4206 | static int mem_cgroup_move_swap_account(swp_entry_t entry, | |
4207 | struct mem_cgroup *from, struct mem_cgroup *to) | |
4208 | { | |
4209 | unsigned short old_id, new_id; | |
4210 | ||
4211 | old_id = mem_cgroup_id(from); | |
4212 | new_id = mem_cgroup_id(to); | |
4213 | ||
4214 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { | |
4215 | mem_cgroup_swap_statistics(from, false); | |
4216 | mem_cgroup_swap_statistics(to, true); | |
4217 | /* | |
4218 | * This function is only called from task migration context now. | |
4219 | * It postpones res_counter and refcount handling till the end | |
4220 | * of task migration(mem_cgroup_clear_mc()) for performance | |
4221 | * improvement. But we cannot postpone css_get(to) because if | |
4222 | * the process that has been moved to @to does swap-in, the | |
4223 | * refcount of @to might be decreased to 0. | |
4224 | * | |
4225 | * We are in attach() phase, so the cgroup is guaranteed to be | |
4226 | * alive, so we can just call css_get(). | |
4227 | */ | |
4228 | css_get(&to->css); | |
4229 | return 0; | |
4230 | } | |
4231 | return -EINVAL; | |
4232 | } | |
4233 | #else | |
4234 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, | |
4235 | struct mem_cgroup *from, struct mem_cgroup *to) | |
4236 | { | |
4237 | return -EINVAL; | |
4238 | } | |
4239 | #endif | |
4240 | ||
4241 | /* | |
4242 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old | |
4243 | * page belongs to. | |
4244 | */ | |
4245 | void mem_cgroup_prepare_migration(struct page *page, struct page *newpage, | |
4246 | struct mem_cgroup **memcgp) | |
4247 | { | |
4248 | struct mem_cgroup *memcg = NULL; | |
4249 | unsigned int nr_pages = 1; | |
4250 | struct page_cgroup *pc; | |
4251 | enum charge_type ctype; | |
4252 | ||
4253 | *memcgp = NULL; | |
4254 | ||
4255 | if (mem_cgroup_disabled()) | |
4256 | return; | |
4257 | ||
4258 | if (PageTransHuge(page)) | |
4259 | nr_pages <<= compound_order(page); | |
4260 | ||
4261 | pc = lookup_page_cgroup(page); | |
4262 | lock_page_cgroup(pc); | |
4263 | if (PageCgroupUsed(pc)) { | |
4264 | memcg = pc->mem_cgroup; | |
4265 | css_get(&memcg->css); | |
4266 | /* | |
4267 | * At migrating an anonymous page, its mapcount goes down | |
4268 | * to 0 and uncharge() will be called. But, even if it's fully | |
4269 | * unmapped, migration may fail and this page has to be | |
4270 | * charged again. We set MIGRATION flag here and delay uncharge | |
4271 | * until end_migration() is called | |
4272 | * | |
4273 | * Corner Case Thinking | |
4274 | * A) | |
4275 | * When the old page was mapped as Anon and it's unmap-and-freed | |
4276 | * while migration was ongoing. | |
4277 | * If unmap finds the old page, uncharge() of it will be delayed | |
4278 | * until end_migration(). If unmap finds a new page, it's | |
4279 | * uncharged when it make mapcount to be 1->0. If unmap code | |
4280 | * finds swap_migration_entry, the new page will not be mapped | |
4281 | * and end_migration() will find it(mapcount==0). | |
4282 | * | |
4283 | * B) | |
4284 | * When the old page was mapped but migraion fails, the kernel | |
4285 | * remaps it. A charge for it is kept by MIGRATION flag even | |
4286 | * if mapcount goes down to 0. We can do remap successfully | |
4287 | * without charging it again. | |
4288 | * | |
4289 | * C) | |
4290 | * The "old" page is under lock_page() until the end of | |
4291 | * migration, so, the old page itself will not be swapped-out. | |
4292 | * If the new page is swapped out before end_migraton, our | |
4293 | * hook to usual swap-out path will catch the event. | |
4294 | */ | |
4295 | if (PageAnon(page)) | |
4296 | SetPageCgroupMigration(pc); | |
4297 | } | |
4298 | unlock_page_cgroup(pc); | |
4299 | /* | |
4300 | * If the page is not charged at this point, | |
4301 | * we return here. | |
4302 | */ | |
4303 | if (!memcg) | |
4304 | return; | |
4305 | ||
4306 | *memcgp = memcg; | |
4307 | /* | |
4308 | * We charge new page before it's used/mapped. So, even if unlock_page() | |
4309 | * is called before end_migration, we can catch all events on this new | |
4310 | * page. In the case new page is migrated but not remapped, new page's | |
4311 | * mapcount will be finally 0 and we call uncharge in end_migration(). | |
4312 | */ | |
4313 | if (PageAnon(page)) | |
4314 | ctype = MEM_CGROUP_CHARGE_TYPE_ANON; | |
4315 | else | |
4316 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
4317 | /* | |
4318 | * The page is committed to the memcg, but it's not actually | |
4319 | * charged to the res_counter since we plan on replacing the | |
4320 | * old one and only one page is going to be left afterwards. | |
4321 | */ | |
4322 | __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false); | |
4323 | } | |
4324 | ||
4325 | /* remove redundant charge if migration failed*/ | |
4326 | void mem_cgroup_end_migration(struct mem_cgroup *memcg, | |
4327 | struct page *oldpage, struct page *newpage, bool migration_ok) | |
4328 | { | |
4329 | struct page *used, *unused; | |
4330 | struct page_cgroup *pc; | |
4331 | bool anon; | |
4332 | ||
4333 | if (!memcg) | |
4334 | return; | |
4335 | ||
4336 | if (!migration_ok) { | |
4337 | used = oldpage; | |
4338 | unused = newpage; | |
4339 | } else { | |
4340 | used = newpage; | |
4341 | unused = oldpage; | |
4342 | } | |
4343 | anon = PageAnon(used); | |
4344 | __mem_cgroup_uncharge_common(unused, | |
4345 | anon ? MEM_CGROUP_CHARGE_TYPE_ANON | |
4346 | : MEM_CGROUP_CHARGE_TYPE_CACHE, | |
4347 | true); | |
4348 | css_put(&memcg->css); | |
4349 | /* | |
4350 | * We disallowed uncharge of pages under migration because mapcount | |
4351 | * of the page goes down to zero, temporarly. | |
4352 | * Clear the flag and check the page should be charged. | |
4353 | */ | |
4354 | pc = lookup_page_cgroup(oldpage); | |
4355 | lock_page_cgroup(pc); | |
4356 | ClearPageCgroupMigration(pc); | |
4357 | unlock_page_cgroup(pc); | |
4358 | ||
4359 | /* | |
4360 | * If a page is a file cache, radix-tree replacement is very atomic | |
4361 | * and we can skip this check. When it was an Anon page, its mapcount | |
4362 | * goes down to 0. But because we added MIGRATION flage, it's not | |
4363 | * uncharged yet. There are several case but page->mapcount check | |
4364 | * and USED bit check in mem_cgroup_uncharge_page() will do enough | |
4365 | * check. (see prepare_charge() also) | |
4366 | */ | |
4367 | if (anon) | |
4368 | mem_cgroup_uncharge_page(used); | |
4369 | } | |
4370 | ||
4371 | /* | |
4372 | * At replace page cache, newpage is not under any memcg but it's on | |
4373 | * LRU. So, this function doesn't touch res_counter but handles LRU | |
4374 | * in correct way. Both pages are locked so we cannot race with uncharge. | |
4375 | */ | |
4376 | void mem_cgroup_replace_page_cache(struct page *oldpage, | |
4377 | struct page *newpage) | |
4378 | { | |
4379 | struct mem_cgroup *memcg = NULL; | |
4380 | struct page_cgroup *pc; | |
4381 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
4382 | ||
4383 | if (mem_cgroup_disabled()) | |
4384 | return; | |
4385 | ||
4386 | pc = lookup_page_cgroup(oldpage); | |
4387 | /* fix accounting on old pages */ | |
4388 | lock_page_cgroup(pc); | |
4389 | if (PageCgroupUsed(pc)) { | |
4390 | memcg = pc->mem_cgroup; | |
4391 | mem_cgroup_charge_statistics(memcg, oldpage, false, -1); | |
4392 | ClearPageCgroupUsed(pc); | |
4393 | } | |
4394 | unlock_page_cgroup(pc); | |
4395 | ||
4396 | /* | |
4397 | * When called from shmem_replace_page(), in some cases the | |
4398 | * oldpage has already been charged, and in some cases not. | |
4399 | */ | |
4400 | if (!memcg) | |
4401 | return; | |
4402 | /* | |
4403 | * Even if newpage->mapping was NULL before starting replacement, | |
4404 | * the newpage may be on LRU(or pagevec for LRU) already. We lock | |
4405 | * LRU while we overwrite pc->mem_cgroup. | |
4406 | */ | |
4407 | __mem_cgroup_commit_charge(memcg, newpage, 1, type, true); | |
4408 | } | |
4409 | ||
4410 | #ifdef CONFIG_DEBUG_VM | |
4411 | static struct page_cgroup *lookup_page_cgroup_used(struct page *page) | |
4412 | { | |
4413 | struct page_cgroup *pc; | |
4414 | ||
4415 | pc = lookup_page_cgroup(page); | |
4416 | /* | |
4417 | * Can be NULL while feeding pages into the page allocator for | |
4418 | * the first time, i.e. during boot or memory hotplug; | |
4419 | * or when mem_cgroup_disabled(). | |
4420 | */ | |
4421 | if (likely(pc) && PageCgroupUsed(pc)) | |
4422 | return pc; | |
4423 | return NULL; | |
4424 | } | |
4425 | ||
4426 | bool mem_cgroup_bad_page_check(struct page *page) | |
4427 | { | |
4428 | if (mem_cgroup_disabled()) | |
4429 | return false; | |
4430 | ||
4431 | return lookup_page_cgroup_used(page) != NULL; | |
4432 | } | |
4433 | ||
4434 | void mem_cgroup_print_bad_page(struct page *page) | |
4435 | { | |
4436 | struct page_cgroup *pc; | |
4437 | ||
4438 | pc = lookup_page_cgroup_used(page); | |
4439 | if (pc) { | |
4440 | pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n", | |
4441 | pc, pc->flags, pc->mem_cgroup); | |
4442 | } | |
4443 | } | |
4444 | #endif | |
4445 | ||
4446 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, | |
4447 | unsigned long long val) | |
4448 | { | |
4449 | int retry_count; | |
4450 | u64 memswlimit, memlimit; | |
4451 | int ret = 0; | |
4452 | int children = mem_cgroup_count_children(memcg); | |
4453 | u64 curusage, oldusage; | |
4454 | int enlarge; | |
4455 | ||
4456 | /* | |
4457 | * For keeping hierarchical_reclaim simple, how long we should retry | |
4458 | * is depends on callers. We set our retry-count to be function | |
4459 | * of # of children which we should visit in this loop. | |
4460 | */ | |
4461 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; | |
4462 | ||
4463 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
4464 | ||
4465 | enlarge = 0; | |
4466 | while (retry_count) { | |
4467 | if (signal_pending(current)) { | |
4468 | ret = -EINTR; | |
4469 | break; | |
4470 | } | |
4471 | /* | |
4472 | * Rather than hide all in some function, I do this in | |
4473 | * open coded manner. You see what this really does. | |
4474 | * We have to guarantee memcg->res.limit <= memcg->memsw.limit. | |
4475 | */ | |
4476 | mutex_lock(&set_limit_mutex); | |
4477 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
4478 | if (memswlimit < val) { | |
4479 | ret = -EINVAL; | |
4480 | mutex_unlock(&set_limit_mutex); | |
4481 | break; | |
4482 | } | |
4483 | ||
4484 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
4485 | if (memlimit < val) | |
4486 | enlarge = 1; | |
4487 | ||
4488 | ret = res_counter_set_limit(&memcg->res, val); | |
4489 | if (!ret) { | |
4490 | if (memswlimit == val) | |
4491 | memcg->memsw_is_minimum = true; | |
4492 | else | |
4493 | memcg->memsw_is_minimum = false; | |
4494 | } | |
4495 | mutex_unlock(&set_limit_mutex); | |
4496 | ||
4497 | if (!ret) | |
4498 | break; | |
4499 | ||
4500 | mem_cgroup_reclaim(memcg, GFP_KERNEL, | |
4501 | MEM_CGROUP_RECLAIM_SHRINK); | |
4502 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
4503 | /* Usage is reduced ? */ | |
4504 | if (curusage >= oldusage) | |
4505 | retry_count--; | |
4506 | else | |
4507 | oldusage = curusage; | |
4508 | } | |
4509 | if (!ret && enlarge) | |
4510 | memcg_oom_recover(memcg); | |
4511 | ||
4512 | return ret; | |
4513 | } | |
4514 | ||
4515 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, | |
4516 | unsigned long long val) | |
4517 | { | |
4518 | int retry_count; | |
4519 | u64 memlimit, memswlimit, oldusage, curusage; | |
4520 | int children = mem_cgroup_count_children(memcg); | |
4521 | int ret = -EBUSY; | |
4522 | int enlarge = 0; | |
4523 | ||
4524 | /* see mem_cgroup_resize_res_limit */ | |
4525 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; | |
4526 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
4527 | while (retry_count) { | |
4528 | if (signal_pending(current)) { | |
4529 | ret = -EINTR; | |
4530 | break; | |
4531 | } | |
4532 | /* | |
4533 | * Rather than hide all in some function, I do this in | |
4534 | * open coded manner. You see what this really does. | |
4535 | * We have to guarantee memcg->res.limit <= memcg->memsw.limit. | |
4536 | */ | |
4537 | mutex_lock(&set_limit_mutex); | |
4538 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
4539 | if (memlimit > val) { | |
4540 | ret = -EINVAL; | |
4541 | mutex_unlock(&set_limit_mutex); | |
4542 | break; | |
4543 | } | |
4544 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
4545 | if (memswlimit < val) | |
4546 | enlarge = 1; | |
4547 | ret = res_counter_set_limit(&memcg->memsw, val); | |
4548 | if (!ret) { | |
4549 | if (memlimit == val) | |
4550 | memcg->memsw_is_minimum = true; | |
4551 | else | |
4552 | memcg->memsw_is_minimum = false; | |
4553 | } | |
4554 | mutex_unlock(&set_limit_mutex); | |
4555 | ||
4556 | if (!ret) | |
4557 | break; | |
4558 | ||
4559 | mem_cgroup_reclaim(memcg, GFP_KERNEL, | |
4560 | MEM_CGROUP_RECLAIM_NOSWAP | | |
4561 | MEM_CGROUP_RECLAIM_SHRINK); | |
4562 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
4563 | /* Usage is reduced ? */ | |
4564 | if (curusage >= oldusage) | |
4565 | retry_count--; | |
4566 | else | |
4567 | oldusage = curusage; | |
4568 | } | |
4569 | if (!ret && enlarge) | |
4570 | memcg_oom_recover(memcg); | |
4571 | return ret; | |
4572 | } | |
4573 | ||
4574 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, | |
4575 | gfp_t gfp_mask, | |
4576 | unsigned long *total_scanned) | |
4577 | { | |
4578 | unsigned long nr_reclaimed = 0; | |
4579 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; | |
4580 | unsigned long reclaimed; | |
4581 | int loop = 0; | |
4582 | struct mem_cgroup_tree_per_zone *mctz; | |
4583 | unsigned long long excess; | |
4584 | unsigned long nr_scanned; | |
4585 | ||
4586 | if (order > 0) | |
4587 | return 0; | |
4588 | ||
4589 | mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); | |
4590 | /* | |
4591 | * This loop can run a while, specially if mem_cgroup's continuously | |
4592 | * keep exceeding their soft limit and putting the system under | |
4593 | * pressure | |
4594 | */ | |
4595 | do { | |
4596 | if (next_mz) | |
4597 | mz = next_mz; | |
4598 | else | |
4599 | mz = mem_cgroup_largest_soft_limit_node(mctz); | |
4600 | if (!mz) | |
4601 | break; | |
4602 | ||
4603 | nr_scanned = 0; | |
4604 | reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone, | |
4605 | gfp_mask, &nr_scanned); | |
4606 | nr_reclaimed += reclaimed; | |
4607 | *total_scanned += nr_scanned; | |
4608 | spin_lock(&mctz->lock); | |
4609 | ||
4610 | /* | |
4611 | * If we failed to reclaim anything from this memory cgroup | |
4612 | * it is time to move on to the next cgroup | |
4613 | */ | |
4614 | next_mz = NULL; | |
4615 | if (!reclaimed) { | |
4616 | do { | |
4617 | /* | |
4618 | * Loop until we find yet another one. | |
4619 | * | |
4620 | * By the time we get the soft_limit lock | |
4621 | * again, someone might have aded the | |
4622 | * group back on the RB tree. Iterate to | |
4623 | * make sure we get a different mem. | |
4624 | * mem_cgroup_largest_soft_limit_node returns | |
4625 | * NULL if no other cgroup is present on | |
4626 | * the tree | |
4627 | */ | |
4628 | next_mz = | |
4629 | __mem_cgroup_largest_soft_limit_node(mctz); | |
4630 | if (next_mz == mz) | |
4631 | css_put(&next_mz->memcg->css); | |
4632 | else /* next_mz == NULL or other memcg */ | |
4633 | break; | |
4634 | } while (1); | |
4635 | } | |
4636 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); | |
4637 | excess = res_counter_soft_limit_excess(&mz->memcg->res); | |
4638 | /* | |
4639 | * One school of thought says that we should not add | |
4640 | * back the node to the tree if reclaim returns 0. | |
4641 | * But our reclaim could return 0, simply because due | |
4642 | * to priority we are exposing a smaller subset of | |
4643 | * memory to reclaim from. Consider this as a longer | |
4644 | * term TODO. | |
4645 | */ | |
4646 | /* If excess == 0, no tree ops */ | |
4647 | __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess); | |
4648 | spin_unlock(&mctz->lock); | |
4649 | css_put(&mz->memcg->css); | |
4650 | loop++; | |
4651 | /* | |
4652 | * Could not reclaim anything and there are no more | |
4653 | * mem cgroups to try or we seem to be looping without | |
4654 | * reclaiming anything. | |
4655 | */ | |
4656 | if (!nr_reclaimed && | |
4657 | (next_mz == NULL || | |
4658 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) | |
4659 | break; | |
4660 | } while (!nr_reclaimed); | |
4661 | if (next_mz) | |
4662 | css_put(&next_mz->memcg->css); | |
4663 | return nr_reclaimed; | |
4664 | } | |
4665 | ||
4666 | /** | |
4667 | * mem_cgroup_force_empty_list - clears LRU of a group | |
4668 | * @memcg: group to clear | |
4669 | * @node: NUMA node | |
4670 | * @zid: zone id | |
4671 | * @lru: lru to to clear | |
4672 | * | |
4673 | * Traverse a specified page_cgroup list and try to drop them all. This doesn't | |
4674 | * reclaim the pages page themselves - pages are moved to the parent (or root) | |
4675 | * group. | |
4676 | */ | |
4677 | static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg, | |
4678 | int node, int zid, enum lru_list lru) | |
4679 | { | |
4680 | struct lruvec *lruvec; | |
4681 | unsigned long flags; | |
4682 | struct list_head *list; | |
4683 | struct page *busy; | |
4684 | struct zone *zone; | |
4685 | ||
4686 | zone = &NODE_DATA(node)->node_zones[zid]; | |
4687 | lruvec = mem_cgroup_zone_lruvec(zone, memcg); | |
4688 | list = &lruvec->lists[lru]; | |
4689 | ||
4690 | busy = NULL; | |
4691 | do { | |
4692 | struct page_cgroup *pc; | |
4693 | struct page *page; | |
4694 | ||
4695 | spin_lock_irqsave(&zone->lru_lock, flags); | |
4696 | if (list_empty(list)) { | |
4697 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
4698 | break; | |
4699 | } | |
4700 | page = list_entry(list->prev, struct page, lru); | |
4701 | if (busy == page) { | |
4702 | list_move(&page->lru, list); | |
4703 | busy = NULL; | |
4704 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
4705 | continue; | |
4706 | } | |
4707 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
4708 | ||
4709 | pc = lookup_page_cgroup(page); | |
4710 | ||
4711 | if (mem_cgroup_move_parent(page, pc, memcg)) { | |
4712 | /* found lock contention or "pc" is obsolete. */ | |
4713 | busy = page; | |
4714 | cond_resched(); | |
4715 | } else | |
4716 | busy = NULL; | |
4717 | } while (!list_empty(list)); | |
4718 | } | |
4719 | ||
4720 | /* | |
4721 | * make mem_cgroup's charge to be 0 if there is no task by moving | |
4722 | * all the charges and pages to the parent. | |
4723 | * This enables deleting this mem_cgroup. | |
4724 | * | |
4725 | * Caller is responsible for holding css reference on the memcg. | |
4726 | */ | |
4727 | static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg) | |
4728 | { | |
4729 | int node, zid; | |
4730 | u64 usage; | |
4731 | ||
4732 | do { | |
4733 | /* This is for making all *used* pages to be on LRU. */ | |
4734 | lru_add_drain_all(); | |
4735 | drain_all_stock_sync(memcg); | |
4736 | mem_cgroup_start_move(memcg); | |
4737 | for_each_node_state(node, N_MEMORY) { | |
4738 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
4739 | enum lru_list lru; | |
4740 | for_each_lru(lru) { | |
4741 | mem_cgroup_force_empty_list(memcg, | |
4742 | node, zid, lru); | |
4743 | } | |
4744 | } | |
4745 | } | |
4746 | mem_cgroup_end_move(memcg); | |
4747 | memcg_oom_recover(memcg); | |
4748 | cond_resched(); | |
4749 | ||
4750 | /* | |
4751 | * Kernel memory may not necessarily be trackable to a specific | |
4752 | * process. So they are not migrated, and therefore we can't | |
4753 | * expect their value to drop to 0 here. | |
4754 | * Having res filled up with kmem only is enough. | |
4755 | * | |
4756 | * This is a safety check because mem_cgroup_force_empty_list | |
4757 | * could have raced with mem_cgroup_replace_page_cache callers | |
4758 | * so the lru seemed empty but the page could have been added | |
4759 | * right after the check. RES_USAGE should be safe as we always | |
4760 | * charge before adding to the LRU. | |
4761 | */ | |
4762 | usage = res_counter_read_u64(&memcg->res, RES_USAGE) - | |
4763 | res_counter_read_u64(&memcg->kmem, RES_USAGE); | |
4764 | } while (usage > 0); | |
4765 | } | |
4766 | ||
4767 | static inline bool memcg_has_children(struct mem_cgroup *memcg) | |
4768 | { | |
4769 | lockdep_assert_held(&memcg_create_mutex); | |
4770 | /* | |
4771 | * The lock does not prevent addition or deletion to the list | |
4772 | * of children, but it prevents a new child from being | |
4773 | * initialized based on this parent in css_online(), so it's | |
4774 | * enough to decide whether hierarchically inherited | |
4775 | * attributes can still be changed or not. | |
4776 | */ | |
4777 | return memcg->use_hierarchy && | |
4778 | !list_empty(&memcg->css.cgroup->children); | |
4779 | } | |
4780 | ||
4781 | /* | |
4782 | * Reclaims as many pages from the given memcg as possible and moves | |
4783 | * the rest to the parent. | |
4784 | * | |
4785 | * Caller is responsible for holding css reference for memcg. | |
4786 | */ | |
4787 | static int mem_cgroup_force_empty(struct mem_cgroup *memcg) | |
4788 | { | |
4789 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
4790 | struct cgroup *cgrp = memcg->css.cgroup; | |
4791 | ||
4792 | /* returns EBUSY if there is a task or if we come here twice. */ | |
4793 | if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children)) | |
4794 | return -EBUSY; | |
4795 | ||
4796 | /* we call try-to-free pages for make this cgroup empty */ | |
4797 | lru_add_drain_all(); | |
4798 | /* try to free all pages in this cgroup */ | |
4799 | while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) { | |
4800 | int progress; | |
4801 | ||
4802 | if (signal_pending(current)) | |
4803 | return -EINTR; | |
4804 | ||
4805 | progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, | |
4806 | false); | |
4807 | if (!progress) { | |
4808 | nr_retries--; | |
4809 | /* maybe some writeback is necessary */ | |
4810 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
4811 | } | |
4812 | ||
4813 | } | |
4814 | lru_add_drain(); | |
4815 | mem_cgroup_reparent_charges(memcg); | |
4816 | ||
4817 | return 0; | |
4818 | } | |
4819 | ||
4820 | static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css, | |
4821 | unsigned int event) | |
4822 | { | |
4823 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
4824 | ||
4825 | if (mem_cgroup_is_root(memcg)) | |
4826 | return -EINVAL; | |
4827 | return mem_cgroup_force_empty(memcg); | |
4828 | } | |
4829 | ||
4830 | static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css, | |
4831 | struct cftype *cft) | |
4832 | { | |
4833 | return mem_cgroup_from_css(css)->use_hierarchy; | |
4834 | } | |
4835 | ||
4836 | static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css, | |
4837 | struct cftype *cft, u64 val) | |
4838 | { | |
4839 | int retval = 0; | |
4840 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
4841 | struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css)); | |
4842 | ||
4843 | mutex_lock(&memcg_create_mutex); | |
4844 | ||
4845 | if (memcg->use_hierarchy == val) | |
4846 | goto out; | |
4847 | ||
4848 | /* | |
4849 | * If parent's use_hierarchy is set, we can't make any modifications | |
4850 | * in the child subtrees. If it is unset, then the change can | |
4851 | * occur, provided the current cgroup has no children. | |
4852 | * | |
4853 | * For the root cgroup, parent_mem is NULL, we allow value to be | |
4854 | * set if there are no children. | |
4855 | */ | |
4856 | if ((!parent_memcg || !parent_memcg->use_hierarchy) && | |
4857 | (val == 1 || val == 0)) { | |
4858 | if (list_empty(&memcg->css.cgroup->children)) | |
4859 | memcg->use_hierarchy = val; | |
4860 | else | |
4861 | retval = -EBUSY; | |
4862 | } else | |
4863 | retval = -EINVAL; | |
4864 | ||
4865 | out: | |
4866 | mutex_unlock(&memcg_create_mutex); | |
4867 | ||
4868 | return retval; | |
4869 | } | |
4870 | ||
4871 | ||
4872 | static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, | |
4873 | enum mem_cgroup_stat_index idx) | |
4874 | { | |
4875 | struct mem_cgroup *iter; | |
4876 | long val = 0; | |
4877 | ||
4878 | /* Per-cpu values can be negative, use a signed accumulator */ | |
4879 | for_each_mem_cgroup_tree(iter, memcg) | |
4880 | val += mem_cgroup_read_stat(iter, idx); | |
4881 | ||
4882 | if (val < 0) /* race ? */ | |
4883 | val = 0; | |
4884 | return val; | |
4885 | } | |
4886 | ||
4887 | static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) | |
4888 | { | |
4889 | u64 val; | |
4890 | ||
4891 | if (!mem_cgroup_is_root(memcg)) { | |
4892 | if (!swap) | |
4893 | return res_counter_read_u64(&memcg->res, RES_USAGE); | |
4894 | else | |
4895 | return res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
4896 | } | |
4897 | ||
4898 | /* | |
4899 | * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS | |
4900 | * as well as in MEM_CGROUP_STAT_RSS_HUGE. | |
4901 | */ | |
4902 | val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); | |
4903 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); | |
4904 | ||
4905 | if (swap) | |
4906 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP); | |
4907 | ||
4908 | return val << PAGE_SHIFT; | |
4909 | } | |
4910 | ||
4911 | static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css, | |
4912 | struct cftype *cft) | |
4913 | { | |
4914 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
4915 | u64 val; | |
4916 | int name; | |
4917 | enum res_type type; | |
4918 | ||
4919 | type = MEMFILE_TYPE(cft->private); | |
4920 | name = MEMFILE_ATTR(cft->private); | |
4921 | ||
4922 | switch (type) { | |
4923 | case _MEM: | |
4924 | if (name == RES_USAGE) | |
4925 | val = mem_cgroup_usage(memcg, false); | |
4926 | else | |
4927 | val = res_counter_read_u64(&memcg->res, name); | |
4928 | break; | |
4929 | case _MEMSWAP: | |
4930 | if (name == RES_USAGE) | |
4931 | val = mem_cgroup_usage(memcg, true); | |
4932 | else | |
4933 | val = res_counter_read_u64(&memcg->memsw, name); | |
4934 | break; | |
4935 | case _KMEM: | |
4936 | val = res_counter_read_u64(&memcg->kmem, name); | |
4937 | break; | |
4938 | default: | |
4939 | BUG(); | |
4940 | } | |
4941 | ||
4942 | return val; | |
4943 | } | |
4944 | ||
4945 | #ifdef CONFIG_MEMCG_KMEM | |
4946 | /* should be called with activate_kmem_mutex held */ | |
4947 | static int __memcg_activate_kmem(struct mem_cgroup *memcg, | |
4948 | unsigned long long limit) | |
4949 | { | |
4950 | int err = 0; | |
4951 | int memcg_id; | |
4952 | ||
4953 | if (memcg_kmem_is_active(memcg)) | |
4954 | return 0; | |
4955 | ||
4956 | /* | |
4957 | * We are going to allocate memory for data shared by all memory | |
4958 | * cgroups so let's stop accounting here. | |
4959 | */ | |
4960 | memcg_stop_kmem_account(); | |
4961 | ||
4962 | /* | |
4963 | * For simplicity, we won't allow this to be disabled. It also can't | |
4964 | * be changed if the cgroup has children already, or if tasks had | |
4965 | * already joined. | |
4966 | * | |
4967 | * If tasks join before we set the limit, a person looking at | |
4968 | * kmem.usage_in_bytes will have no way to determine when it took | |
4969 | * place, which makes the value quite meaningless. | |
4970 | * | |
4971 | * After it first became limited, changes in the value of the limit are | |
4972 | * of course permitted. | |
4973 | */ | |
4974 | mutex_lock(&memcg_create_mutex); | |
4975 | if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg)) | |
4976 | err = -EBUSY; | |
4977 | mutex_unlock(&memcg_create_mutex); | |
4978 | if (err) | |
4979 | goto out; | |
4980 | ||
4981 | memcg_id = ida_simple_get(&kmem_limited_groups, | |
4982 | 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL); | |
4983 | if (memcg_id < 0) { | |
4984 | err = memcg_id; | |
4985 | goto out; | |
4986 | } | |
4987 | ||
4988 | /* | |
4989 | * Make sure we have enough space for this cgroup in each root cache's | |
4990 | * memcg_params. | |
4991 | */ | |
4992 | mutex_lock(&memcg_slab_mutex); | |
4993 | err = memcg_update_all_caches(memcg_id + 1); | |
4994 | mutex_unlock(&memcg_slab_mutex); | |
4995 | if (err) | |
4996 | goto out_rmid; | |
4997 | ||
4998 | memcg->kmemcg_id = memcg_id; | |
4999 | INIT_LIST_HEAD(&memcg->memcg_slab_caches); | |
5000 | ||
5001 | /* | |
5002 | * We couldn't have accounted to this cgroup, because it hasn't got the | |
5003 | * active bit set yet, so this should succeed. | |
5004 | */ | |
5005 | err = res_counter_set_limit(&memcg->kmem, limit); | |
5006 | VM_BUG_ON(err); | |
5007 | ||
5008 | static_key_slow_inc(&memcg_kmem_enabled_key); | |
5009 | /* | |
5010 | * Setting the active bit after enabling static branching will | |
5011 | * guarantee no one starts accounting before all call sites are | |
5012 | * patched. | |
5013 | */ | |
5014 | memcg_kmem_set_active(memcg); | |
5015 | out: | |
5016 | memcg_resume_kmem_account(); | |
5017 | return err; | |
5018 | ||
5019 | out_rmid: | |
5020 | ida_simple_remove(&kmem_limited_groups, memcg_id); | |
5021 | goto out; | |
5022 | } | |
5023 | ||
5024 | static int memcg_activate_kmem(struct mem_cgroup *memcg, | |
5025 | unsigned long long limit) | |
5026 | { | |
5027 | int ret; | |
5028 | ||
5029 | mutex_lock(&activate_kmem_mutex); | |
5030 | ret = __memcg_activate_kmem(memcg, limit); | |
5031 | mutex_unlock(&activate_kmem_mutex); | |
5032 | return ret; | |
5033 | } | |
5034 | ||
5035 | static int memcg_update_kmem_limit(struct mem_cgroup *memcg, | |
5036 | unsigned long long val) | |
5037 | { | |
5038 | int ret; | |
5039 | ||
5040 | if (!memcg_kmem_is_active(memcg)) | |
5041 | ret = memcg_activate_kmem(memcg, val); | |
5042 | else | |
5043 | ret = res_counter_set_limit(&memcg->kmem, val); | |
5044 | return ret; | |
5045 | } | |
5046 | ||
5047 | static int memcg_propagate_kmem(struct mem_cgroup *memcg) | |
5048 | { | |
5049 | int ret = 0; | |
5050 | struct mem_cgroup *parent = parent_mem_cgroup(memcg); | |
5051 | ||
5052 | if (!parent) | |
5053 | return 0; | |
5054 | ||
5055 | mutex_lock(&activate_kmem_mutex); | |
5056 | /* | |
5057 | * If the parent cgroup is not kmem-active now, it cannot be activated | |
5058 | * after this point, because it has at least one child already. | |
5059 | */ | |
5060 | if (memcg_kmem_is_active(parent)) | |
5061 | ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX); | |
5062 | mutex_unlock(&activate_kmem_mutex); | |
5063 | return ret; | |
5064 | } | |
5065 | #else | |
5066 | static int memcg_update_kmem_limit(struct mem_cgroup *memcg, | |
5067 | unsigned long long val) | |
5068 | { | |
5069 | return -EINVAL; | |
5070 | } | |
5071 | #endif /* CONFIG_MEMCG_KMEM */ | |
5072 | ||
5073 | /* | |
5074 | * The user of this function is... | |
5075 | * RES_LIMIT. | |
5076 | */ | |
5077 | static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft, | |
5078 | char *buffer) | |
5079 | { | |
5080 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5081 | enum res_type type; | |
5082 | int name; | |
5083 | unsigned long long val; | |
5084 | int ret; | |
5085 | ||
5086 | type = MEMFILE_TYPE(cft->private); | |
5087 | name = MEMFILE_ATTR(cft->private); | |
5088 | ||
5089 | switch (name) { | |
5090 | case RES_LIMIT: | |
5091 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ | |
5092 | ret = -EINVAL; | |
5093 | break; | |
5094 | } | |
5095 | /* This function does all necessary parse...reuse it */ | |
5096 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
5097 | if (ret) | |
5098 | break; | |
5099 | if (type == _MEM) | |
5100 | ret = mem_cgroup_resize_limit(memcg, val); | |
5101 | else if (type == _MEMSWAP) | |
5102 | ret = mem_cgroup_resize_memsw_limit(memcg, val); | |
5103 | else if (type == _KMEM) | |
5104 | ret = memcg_update_kmem_limit(memcg, val); | |
5105 | else | |
5106 | return -EINVAL; | |
5107 | break; | |
5108 | case RES_SOFT_LIMIT: | |
5109 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
5110 | if (ret) | |
5111 | break; | |
5112 | /* | |
5113 | * For memsw, soft limits are hard to implement in terms | |
5114 | * of semantics, for now, we support soft limits for | |
5115 | * control without swap | |
5116 | */ | |
5117 | if (type == _MEM) | |
5118 | ret = res_counter_set_soft_limit(&memcg->res, val); | |
5119 | else | |
5120 | ret = -EINVAL; | |
5121 | break; | |
5122 | default: | |
5123 | ret = -EINVAL; /* should be BUG() ? */ | |
5124 | break; | |
5125 | } | |
5126 | return ret; | |
5127 | } | |
5128 | ||
5129 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, | |
5130 | unsigned long long *mem_limit, unsigned long long *memsw_limit) | |
5131 | { | |
5132 | unsigned long long min_limit, min_memsw_limit, tmp; | |
5133 | ||
5134 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
5135 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
5136 | if (!memcg->use_hierarchy) | |
5137 | goto out; | |
5138 | ||
5139 | while (css_parent(&memcg->css)) { | |
5140 | memcg = mem_cgroup_from_css(css_parent(&memcg->css)); | |
5141 | if (!memcg->use_hierarchy) | |
5142 | break; | |
5143 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
5144 | min_limit = min(min_limit, tmp); | |
5145 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
5146 | min_memsw_limit = min(min_memsw_limit, tmp); | |
5147 | } | |
5148 | out: | |
5149 | *mem_limit = min_limit; | |
5150 | *memsw_limit = min_memsw_limit; | |
5151 | } | |
5152 | ||
5153 | static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event) | |
5154 | { | |
5155 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5156 | int name; | |
5157 | enum res_type type; | |
5158 | ||
5159 | type = MEMFILE_TYPE(event); | |
5160 | name = MEMFILE_ATTR(event); | |
5161 | ||
5162 | switch (name) { | |
5163 | case RES_MAX_USAGE: | |
5164 | if (type == _MEM) | |
5165 | res_counter_reset_max(&memcg->res); | |
5166 | else if (type == _MEMSWAP) | |
5167 | res_counter_reset_max(&memcg->memsw); | |
5168 | else if (type == _KMEM) | |
5169 | res_counter_reset_max(&memcg->kmem); | |
5170 | else | |
5171 | return -EINVAL; | |
5172 | break; | |
5173 | case RES_FAILCNT: | |
5174 | if (type == _MEM) | |
5175 | res_counter_reset_failcnt(&memcg->res); | |
5176 | else if (type == _MEMSWAP) | |
5177 | res_counter_reset_failcnt(&memcg->memsw); | |
5178 | else if (type == _KMEM) | |
5179 | res_counter_reset_failcnt(&memcg->kmem); | |
5180 | else | |
5181 | return -EINVAL; | |
5182 | break; | |
5183 | } | |
5184 | ||
5185 | return 0; | |
5186 | } | |
5187 | ||
5188 | static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css, | |
5189 | struct cftype *cft) | |
5190 | { | |
5191 | return mem_cgroup_from_css(css)->move_charge_at_immigrate; | |
5192 | } | |
5193 | ||
5194 | #ifdef CONFIG_MMU | |
5195 | static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css, | |
5196 | struct cftype *cft, u64 val) | |
5197 | { | |
5198 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5199 | ||
5200 | if (val >= (1 << NR_MOVE_TYPE)) | |
5201 | return -EINVAL; | |
5202 | ||
5203 | /* | |
5204 | * No kind of locking is needed in here, because ->can_attach() will | |
5205 | * check this value once in the beginning of the process, and then carry | |
5206 | * on with stale data. This means that changes to this value will only | |
5207 | * affect task migrations starting after the change. | |
5208 | */ | |
5209 | memcg->move_charge_at_immigrate = val; | |
5210 | return 0; | |
5211 | } | |
5212 | #else | |
5213 | static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css, | |
5214 | struct cftype *cft, u64 val) | |
5215 | { | |
5216 | return -ENOSYS; | |
5217 | } | |
5218 | #endif | |
5219 | ||
5220 | #ifdef CONFIG_NUMA | |
5221 | static int memcg_numa_stat_show(struct seq_file *m, void *v) | |
5222 | { | |
5223 | struct numa_stat { | |
5224 | const char *name; | |
5225 | unsigned int lru_mask; | |
5226 | }; | |
5227 | ||
5228 | static const struct numa_stat stats[] = { | |
5229 | { "total", LRU_ALL }, | |
5230 | { "file", LRU_ALL_FILE }, | |
5231 | { "anon", LRU_ALL_ANON }, | |
5232 | { "unevictable", BIT(LRU_UNEVICTABLE) }, | |
5233 | }; | |
5234 | const struct numa_stat *stat; | |
5235 | int nid; | |
5236 | unsigned long nr; | |
5237 | struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m)); | |
5238 | ||
5239 | for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) { | |
5240 | nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask); | |
5241 | seq_printf(m, "%s=%lu", stat->name, nr); | |
5242 | for_each_node_state(nid, N_MEMORY) { | |
5243 | nr = mem_cgroup_node_nr_lru_pages(memcg, nid, | |
5244 | stat->lru_mask); | |
5245 | seq_printf(m, " N%d=%lu", nid, nr); | |
5246 | } | |
5247 | seq_putc(m, '\n'); | |
5248 | } | |
5249 | ||
5250 | for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) { | |
5251 | struct mem_cgroup *iter; | |
5252 | ||
5253 | nr = 0; | |
5254 | for_each_mem_cgroup_tree(iter, memcg) | |
5255 | nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask); | |
5256 | seq_printf(m, "hierarchical_%s=%lu", stat->name, nr); | |
5257 | for_each_node_state(nid, N_MEMORY) { | |
5258 | nr = 0; | |
5259 | for_each_mem_cgroup_tree(iter, memcg) | |
5260 | nr += mem_cgroup_node_nr_lru_pages( | |
5261 | iter, nid, stat->lru_mask); | |
5262 | seq_printf(m, " N%d=%lu", nid, nr); | |
5263 | } | |
5264 | seq_putc(m, '\n'); | |
5265 | } | |
5266 | ||
5267 | return 0; | |
5268 | } | |
5269 | #endif /* CONFIG_NUMA */ | |
5270 | ||
5271 | static inline void mem_cgroup_lru_names_not_uptodate(void) | |
5272 | { | |
5273 | BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS); | |
5274 | } | |
5275 | ||
5276 | static int memcg_stat_show(struct seq_file *m, void *v) | |
5277 | { | |
5278 | struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m)); | |
5279 | struct mem_cgroup *mi; | |
5280 | unsigned int i; | |
5281 | ||
5282 | for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) { | |
5283 | if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account) | |
5284 | continue; | |
5285 | seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i], | |
5286 | mem_cgroup_read_stat(memcg, i) * PAGE_SIZE); | |
5287 | } | |
5288 | ||
5289 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) | |
5290 | seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i], | |
5291 | mem_cgroup_read_events(memcg, i)); | |
5292 | ||
5293 | for (i = 0; i < NR_LRU_LISTS; i++) | |
5294 | seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i], | |
5295 | mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE); | |
5296 | ||
5297 | /* Hierarchical information */ | |
5298 | { | |
5299 | unsigned long long limit, memsw_limit; | |
5300 | memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit); | |
5301 | seq_printf(m, "hierarchical_memory_limit %llu\n", limit); | |
5302 | if (do_swap_account) | |
5303 | seq_printf(m, "hierarchical_memsw_limit %llu\n", | |
5304 | memsw_limit); | |
5305 | } | |
5306 | ||
5307 | for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) { | |
5308 | long long val = 0; | |
5309 | ||
5310 | if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account) | |
5311 | continue; | |
5312 | for_each_mem_cgroup_tree(mi, memcg) | |
5313 | val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE; | |
5314 | seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val); | |
5315 | } | |
5316 | ||
5317 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { | |
5318 | unsigned long long val = 0; | |
5319 | ||
5320 | for_each_mem_cgroup_tree(mi, memcg) | |
5321 | val += mem_cgroup_read_events(mi, i); | |
5322 | seq_printf(m, "total_%s %llu\n", | |
5323 | mem_cgroup_events_names[i], val); | |
5324 | } | |
5325 | ||
5326 | for (i = 0; i < NR_LRU_LISTS; i++) { | |
5327 | unsigned long long val = 0; | |
5328 | ||
5329 | for_each_mem_cgroup_tree(mi, memcg) | |
5330 | val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE; | |
5331 | seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val); | |
5332 | } | |
5333 | ||
5334 | #ifdef CONFIG_DEBUG_VM | |
5335 | { | |
5336 | int nid, zid; | |
5337 | struct mem_cgroup_per_zone *mz; | |
5338 | struct zone_reclaim_stat *rstat; | |
5339 | unsigned long recent_rotated[2] = {0, 0}; | |
5340 | unsigned long recent_scanned[2] = {0, 0}; | |
5341 | ||
5342 | for_each_online_node(nid) | |
5343 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
5344 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
5345 | rstat = &mz->lruvec.reclaim_stat; | |
5346 | ||
5347 | recent_rotated[0] += rstat->recent_rotated[0]; | |
5348 | recent_rotated[1] += rstat->recent_rotated[1]; | |
5349 | recent_scanned[0] += rstat->recent_scanned[0]; | |
5350 | recent_scanned[1] += rstat->recent_scanned[1]; | |
5351 | } | |
5352 | seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]); | |
5353 | seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]); | |
5354 | seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]); | |
5355 | seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]); | |
5356 | } | |
5357 | #endif | |
5358 | ||
5359 | return 0; | |
5360 | } | |
5361 | ||
5362 | static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css, | |
5363 | struct cftype *cft) | |
5364 | { | |
5365 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5366 | ||
5367 | return mem_cgroup_swappiness(memcg); | |
5368 | } | |
5369 | ||
5370 | static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css, | |
5371 | struct cftype *cft, u64 val) | |
5372 | { | |
5373 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5374 | ||
5375 | if (val > 100) | |
5376 | return -EINVAL; | |
5377 | ||
5378 | if (css_parent(css)) | |
5379 | memcg->swappiness = val; | |
5380 | else | |
5381 | vm_swappiness = val; | |
5382 | ||
5383 | return 0; | |
5384 | } | |
5385 | ||
5386 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) | |
5387 | { | |
5388 | struct mem_cgroup_threshold_ary *t; | |
5389 | u64 usage; | |
5390 | int i; | |
5391 | ||
5392 | rcu_read_lock(); | |
5393 | if (!swap) | |
5394 | t = rcu_dereference(memcg->thresholds.primary); | |
5395 | else | |
5396 | t = rcu_dereference(memcg->memsw_thresholds.primary); | |
5397 | ||
5398 | if (!t) | |
5399 | goto unlock; | |
5400 | ||
5401 | usage = mem_cgroup_usage(memcg, swap); | |
5402 | ||
5403 | /* | |
5404 | * current_threshold points to threshold just below or equal to usage. | |
5405 | * If it's not true, a threshold was crossed after last | |
5406 | * call of __mem_cgroup_threshold(). | |
5407 | */ | |
5408 | i = t->current_threshold; | |
5409 | ||
5410 | /* | |
5411 | * Iterate backward over array of thresholds starting from | |
5412 | * current_threshold and check if a threshold is crossed. | |
5413 | * If none of thresholds below usage is crossed, we read | |
5414 | * only one element of the array here. | |
5415 | */ | |
5416 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) | |
5417 | eventfd_signal(t->entries[i].eventfd, 1); | |
5418 | ||
5419 | /* i = current_threshold + 1 */ | |
5420 | i++; | |
5421 | ||
5422 | /* | |
5423 | * Iterate forward over array of thresholds starting from | |
5424 | * current_threshold+1 and check if a threshold is crossed. | |
5425 | * If none of thresholds above usage is crossed, we read | |
5426 | * only one element of the array here. | |
5427 | */ | |
5428 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) | |
5429 | eventfd_signal(t->entries[i].eventfd, 1); | |
5430 | ||
5431 | /* Update current_threshold */ | |
5432 | t->current_threshold = i - 1; | |
5433 | unlock: | |
5434 | rcu_read_unlock(); | |
5435 | } | |
5436 | ||
5437 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) | |
5438 | { | |
5439 | while (memcg) { | |
5440 | __mem_cgroup_threshold(memcg, false); | |
5441 | if (do_swap_account) | |
5442 | __mem_cgroup_threshold(memcg, true); | |
5443 | ||
5444 | memcg = parent_mem_cgroup(memcg); | |
5445 | } | |
5446 | } | |
5447 | ||
5448 | static int compare_thresholds(const void *a, const void *b) | |
5449 | { | |
5450 | const struct mem_cgroup_threshold *_a = a; | |
5451 | const struct mem_cgroup_threshold *_b = b; | |
5452 | ||
5453 | if (_a->threshold > _b->threshold) | |
5454 | return 1; | |
5455 | ||
5456 | if (_a->threshold < _b->threshold) | |
5457 | return -1; | |
5458 | ||
5459 | return 0; | |
5460 | } | |
5461 | ||
5462 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) | |
5463 | { | |
5464 | struct mem_cgroup_eventfd_list *ev; | |
5465 | ||
5466 | list_for_each_entry(ev, &memcg->oom_notify, list) | |
5467 | eventfd_signal(ev->eventfd, 1); | |
5468 | return 0; | |
5469 | } | |
5470 | ||
5471 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) | |
5472 | { | |
5473 | struct mem_cgroup *iter; | |
5474 | ||
5475 | for_each_mem_cgroup_tree(iter, memcg) | |
5476 | mem_cgroup_oom_notify_cb(iter); | |
5477 | } | |
5478 | ||
5479 | static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg, | |
5480 | struct eventfd_ctx *eventfd, const char *args, enum res_type type) | |
5481 | { | |
5482 | struct mem_cgroup_thresholds *thresholds; | |
5483 | struct mem_cgroup_threshold_ary *new; | |
5484 | u64 threshold, usage; | |
5485 | int i, size, ret; | |
5486 | ||
5487 | ret = res_counter_memparse_write_strategy(args, &threshold); | |
5488 | if (ret) | |
5489 | return ret; | |
5490 | ||
5491 | mutex_lock(&memcg->thresholds_lock); | |
5492 | ||
5493 | if (type == _MEM) | |
5494 | thresholds = &memcg->thresholds; | |
5495 | else if (type == _MEMSWAP) | |
5496 | thresholds = &memcg->memsw_thresholds; | |
5497 | else | |
5498 | BUG(); | |
5499 | ||
5500 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
5501 | ||
5502 | /* Check if a threshold crossed before adding a new one */ | |
5503 | if (thresholds->primary) | |
5504 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
5505 | ||
5506 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; | |
5507 | ||
5508 | /* Allocate memory for new array of thresholds */ | |
5509 | new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), | |
5510 | GFP_KERNEL); | |
5511 | if (!new) { | |
5512 | ret = -ENOMEM; | |
5513 | goto unlock; | |
5514 | } | |
5515 | new->size = size; | |
5516 | ||
5517 | /* Copy thresholds (if any) to new array */ | |
5518 | if (thresholds->primary) { | |
5519 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * | |
5520 | sizeof(struct mem_cgroup_threshold)); | |
5521 | } | |
5522 | ||
5523 | /* Add new threshold */ | |
5524 | new->entries[size - 1].eventfd = eventfd; | |
5525 | new->entries[size - 1].threshold = threshold; | |
5526 | ||
5527 | /* Sort thresholds. Registering of new threshold isn't time-critical */ | |
5528 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), | |
5529 | compare_thresholds, NULL); | |
5530 | ||
5531 | /* Find current threshold */ | |
5532 | new->current_threshold = -1; | |
5533 | for (i = 0; i < size; i++) { | |
5534 | if (new->entries[i].threshold <= usage) { | |
5535 | /* | |
5536 | * new->current_threshold will not be used until | |
5537 | * rcu_assign_pointer(), so it's safe to increment | |
5538 | * it here. | |
5539 | */ | |
5540 | ++new->current_threshold; | |
5541 | } else | |
5542 | break; | |
5543 | } | |
5544 | ||
5545 | /* Free old spare buffer and save old primary buffer as spare */ | |
5546 | kfree(thresholds->spare); | |
5547 | thresholds->spare = thresholds->primary; | |
5548 | ||
5549 | rcu_assign_pointer(thresholds->primary, new); | |
5550 | ||
5551 | /* To be sure that nobody uses thresholds */ | |
5552 | synchronize_rcu(); | |
5553 | ||
5554 | unlock: | |
5555 | mutex_unlock(&memcg->thresholds_lock); | |
5556 | ||
5557 | return ret; | |
5558 | } | |
5559 | ||
5560 | static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg, | |
5561 | struct eventfd_ctx *eventfd, const char *args) | |
5562 | { | |
5563 | return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM); | |
5564 | } | |
5565 | ||
5566 | static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg, | |
5567 | struct eventfd_ctx *eventfd, const char *args) | |
5568 | { | |
5569 | return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP); | |
5570 | } | |
5571 | ||
5572 | static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg, | |
5573 | struct eventfd_ctx *eventfd, enum res_type type) | |
5574 | { | |
5575 | struct mem_cgroup_thresholds *thresholds; | |
5576 | struct mem_cgroup_threshold_ary *new; | |
5577 | u64 usage; | |
5578 | int i, j, size; | |
5579 | ||
5580 | mutex_lock(&memcg->thresholds_lock); | |
5581 | if (type == _MEM) | |
5582 | thresholds = &memcg->thresholds; | |
5583 | else if (type == _MEMSWAP) | |
5584 | thresholds = &memcg->memsw_thresholds; | |
5585 | else | |
5586 | BUG(); | |
5587 | ||
5588 | if (!thresholds->primary) | |
5589 | goto unlock; | |
5590 | ||
5591 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
5592 | ||
5593 | /* Check if a threshold crossed before removing */ | |
5594 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
5595 | ||
5596 | /* Calculate new number of threshold */ | |
5597 | size = 0; | |
5598 | for (i = 0; i < thresholds->primary->size; i++) { | |
5599 | if (thresholds->primary->entries[i].eventfd != eventfd) | |
5600 | size++; | |
5601 | } | |
5602 | ||
5603 | new = thresholds->spare; | |
5604 | ||
5605 | /* Set thresholds array to NULL if we don't have thresholds */ | |
5606 | if (!size) { | |
5607 | kfree(new); | |
5608 | new = NULL; | |
5609 | goto swap_buffers; | |
5610 | } | |
5611 | ||
5612 | new->size = size; | |
5613 | ||
5614 | /* Copy thresholds and find current threshold */ | |
5615 | new->current_threshold = -1; | |
5616 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { | |
5617 | if (thresholds->primary->entries[i].eventfd == eventfd) | |
5618 | continue; | |
5619 | ||
5620 | new->entries[j] = thresholds->primary->entries[i]; | |
5621 | if (new->entries[j].threshold <= usage) { | |
5622 | /* | |
5623 | * new->current_threshold will not be used | |
5624 | * until rcu_assign_pointer(), so it's safe to increment | |
5625 | * it here. | |
5626 | */ | |
5627 | ++new->current_threshold; | |
5628 | } | |
5629 | j++; | |
5630 | } | |
5631 | ||
5632 | swap_buffers: | |
5633 | /* Swap primary and spare array */ | |
5634 | thresholds->spare = thresholds->primary; | |
5635 | /* If all events are unregistered, free the spare array */ | |
5636 | if (!new) { | |
5637 | kfree(thresholds->spare); | |
5638 | thresholds->spare = NULL; | |
5639 | } | |
5640 | ||
5641 | rcu_assign_pointer(thresholds->primary, new); | |
5642 | ||
5643 | /* To be sure that nobody uses thresholds */ | |
5644 | synchronize_rcu(); | |
5645 | unlock: | |
5646 | mutex_unlock(&memcg->thresholds_lock); | |
5647 | } | |
5648 | ||
5649 | static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg, | |
5650 | struct eventfd_ctx *eventfd) | |
5651 | { | |
5652 | return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM); | |
5653 | } | |
5654 | ||
5655 | static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg, | |
5656 | struct eventfd_ctx *eventfd) | |
5657 | { | |
5658 | return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP); | |
5659 | } | |
5660 | ||
5661 | static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg, | |
5662 | struct eventfd_ctx *eventfd, const char *args) | |
5663 | { | |
5664 | struct mem_cgroup_eventfd_list *event; | |
5665 | ||
5666 | event = kmalloc(sizeof(*event), GFP_KERNEL); | |
5667 | if (!event) | |
5668 | return -ENOMEM; | |
5669 | ||
5670 | spin_lock(&memcg_oom_lock); | |
5671 | ||
5672 | event->eventfd = eventfd; | |
5673 | list_add(&event->list, &memcg->oom_notify); | |
5674 | ||
5675 | /* already in OOM ? */ | |
5676 | if (atomic_read(&memcg->under_oom)) | |
5677 | eventfd_signal(eventfd, 1); | |
5678 | spin_unlock(&memcg_oom_lock); | |
5679 | ||
5680 | return 0; | |
5681 | } | |
5682 | ||
5683 | static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg, | |
5684 | struct eventfd_ctx *eventfd) | |
5685 | { | |
5686 | struct mem_cgroup_eventfd_list *ev, *tmp; | |
5687 | ||
5688 | spin_lock(&memcg_oom_lock); | |
5689 | ||
5690 | list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { | |
5691 | if (ev->eventfd == eventfd) { | |
5692 | list_del(&ev->list); | |
5693 | kfree(ev); | |
5694 | } | |
5695 | } | |
5696 | ||
5697 | spin_unlock(&memcg_oom_lock); | |
5698 | } | |
5699 | ||
5700 | static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v) | |
5701 | { | |
5702 | struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf)); | |
5703 | ||
5704 | seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable); | |
5705 | seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom)); | |
5706 | return 0; | |
5707 | } | |
5708 | ||
5709 | static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css, | |
5710 | struct cftype *cft, u64 val) | |
5711 | { | |
5712 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5713 | ||
5714 | /* cannot set to root cgroup and only 0 and 1 are allowed */ | |
5715 | if (!css_parent(css) || !((val == 0) || (val == 1))) | |
5716 | return -EINVAL; | |
5717 | ||
5718 | memcg->oom_kill_disable = val; | |
5719 | if (!val) | |
5720 | memcg_oom_recover(memcg); | |
5721 | ||
5722 | return 0; | |
5723 | } | |
5724 | ||
5725 | #ifdef CONFIG_MEMCG_KMEM | |
5726 | static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss) | |
5727 | { | |
5728 | int ret; | |
5729 | ||
5730 | memcg->kmemcg_id = -1; | |
5731 | ret = memcg_propagate_kmem(memcg); | |
5732 | if (ret) | |
5733 | return ret; | |
5734 | ||
5735 | return mem_cgroup_sockets_init(memcg, ss); | |
5736 | } | |
5737 | ||
5738 | static void memcg_destroy_kmem(struct mem_cgroup *memcg) | |
5739 | { | |
5740 | mem_cgroup_sockets_destroy(memcg); | |
5741 | } | |
5742 | ||
5743 | static void kmem_cgroup_css_offline(struct mem_cgroup *memcg) | |
5744 | { | |
5745 | if (!memcg_kmem_is_active(memcg)) | |
5746 | return; | |
5747 | ||
5748 | /* | |
5749 | * kmem charges can outlive the cgroup. In the case of slab | |
5750 | * pages, for instance, a page contain objects from various | |
5751 | * processes. As we prevent from taking a reference for every | |
5752 | * such allocation we have to be careful when doing uncharge | |
5753 | * (see memcg_uncharge_kmem) and here during offlining. | |
5754 | * | |
5755 | * The idea is that that only the _last_ uncharge which sees | |
5756 | * the dead memcg will drop the last reference. An additional | |
5757 | * reference is taken here before the group is marked dead | |
5758 | * which is then paired with css_put during uncharge resp. here. | |
5759 | * | |
5760 | * Although this might sound strange as this path is called from | |
5761 | * css_offline() when the referencemight have dropped down to 0 | |
5762 | * and shouldn't be incremented anymore (css_tryget would fail) | |
5763 | * we do not have other options because of the kmem allocations | |
5764 | * lifetime. | |
5765 | */ | |
5766 | css_get(&memcg->css); | |
5767 | ||
5768 | memcg_kmem_mark_dead(memcg); | |
5769 | ||
5770 | if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0) | |
5771 | return; | |
5772 | ||
5773 | if (memcg_kmem_test_and_clear_dead(memcg)) | |
5774 | css_put(&memcg->css); | |
5775 | } | |
5776 | #else | |
5777 | static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss) | |
5778 | { | |
5779 | return 0; | |
5780 | } | |
5781 | ||
5782 | static void memcg_destroy_kmem(struct mem_cgroup *memcg) | |
5783 | { | |
5784 | } | |
5785 | ||
5786 | static void kmem_cgroup_css_offline(struct mem_cgroup *memcg) | |
5787 | { | |
5788 | } | |
5789 | #endif | |
5790 | ||
5791 | /* | |
5792 | * DO NOT USE IN NEW FILES. | |
5793 | * | |
5794 | * "cgroup.event_control" implementation. | |
5795 | * | |
5796 | * This is way over-engineered. It tries to support fully configurable | |
5797 | * events for each user. Such level of flexibility is completely | |
5798 | * unnecessary especially in the light of the planned unified hierarchy. | |
5799 | * | |
5800 | * Please deprecate this and replace with something simpler if at all | |
5801 | * possible. | |
5802 | */ | |
5803 | ||
5804 | /* | |
5805 | * Unregister event and free resources. | |
5806 | * | |
5807 | * Gets called from workqueue. | |
5808 | */ | |
5809 | static void memcg_event_remove(struct work_struct *work) | |
5810 | { | |
5811 | struct mem_cgroup_event *event = | |
5812 | container_of(work, struct mem_cgroup_event, remove); | |
5813 | struct mem_cgroup *memcg = event->memcg; | |
5814 | ||
5815 | remove_wait_queue(event->wqh, &event->wait); | |
5816 | ||
5817 | event->unregister_event(memcg, event->eventfd); | |
5818 | ||
5819 | /* Notify userspace the event is going away. */ | |
5820 | eventfd_signal(event->eventfd, 1); | |
5821 | ||
5822 | eventfd_ctx_put(event->eventfd); | |
5823 | kfree(event); | |
5824 | css_put(&memcg->css); | |
5825 | } | |
5826 | ||
5827 | /* | |
5828 | * Gets called on POLLHUP on eventfd when user closes it. | |
5829 | * | |
5830 | * Called with wqh->lock held and interrupts disabled. | |
5831 | */ | |
5832 | static int memcg_event_wake(wait_queue_t *wait, unsigned mode, | |
5833 | int sync, void *key) | |
5834 | { | |
5835 | struct mem_cgroup_event *event = | |
5836 | container_of(wait, struct mem_cgroup_event, wait); | |
5837 | struct mem_cgroup *memcg = event->memcg; | |
5838 | unsigned long flags = (unsigned long)key; | |
5839 | ||
5840 | if (flags & POLLHUP) { | |
5841 | /* | |
5842 | * If the event has been detached at cgroup removal, we | |
5843 | * can simply return knowing the other side will cleanup | |
5844 | * for us. | |
5845 | * | |
5846 | * We can't race against event freeing since the other | |
5847 | * side will require wqh->lock via remove_wait_queue(), | |
5848 | * which we hold. | |
5849 | */ | |
5850 | spin_lock(&memcg->event_list_lock); | |
5851 | if (!list_empty(&event->list)) { | |
5852 | list_del_init(&event->list); | |
5853 | /* | |
5854 | * We are in atomic context, but cgroup_event_remove() | |
5855 | * may sleep, so we have to call it in workqueue. | |
5856 | */ | |
5857 | schedule_work(&event->remove); | |
5858 | } | |
5859 | spin_unlock(&memcg->event_list_lock); | |
5860 | } | |
5861 | ||
5862 | return 0; | |
5863 | } | |
5864 | ||
5865 | static void memcg_event_ptable_queue_proc(struct file *file, | |
5866 | wait_queue_head_t *wqh, poll_table *pt) | |
5867 | { | |
5868 | struct mem_cgroup_event *event = | |
5869 | container_of(pt, struct mem_cgroup_event, pt); | |
5870 | ||
5871 | event->wqh = wqh; | |
5872 | add_wait_queue(wqh, &event->wait); | |
5873 | } | |
5874 | ||
5875 | /* | |
5876 | * DO NOT USE IN NEW FILES. | |
5877 | * | |
5878 | * Parse input and register new cgroup event handler. | |
5879 | * | |
5880 | * Input must be in format '<event_fd> <control_fd> <args>'. | |
5881 | * Interpretation of args is defined by control file implementation. | |
5882 | */ | |
5883 | static int memcg_write_event_control(struct cgroup_subsys_state *css, | |
5884 | struct cftype *cft, char *buffer) | |
5885 | { | |
5886 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5887 | struct mem_cgroup_event *event; | |
5888 | struct cgroup_subsys_state *cfile_css; | |
5889 | unsigned int efd, cfd; | |
5890 | struct fd efile; | |
5891 | struct fd cfile; | |
5892 | const char *name; | |
5893 | char *endp; | |
5894 | int ret; | |
5895 | ||
5896 | efd = simple_strtoul(buffer, &endp, 10); | |
5897 | if (*endp != ' ') | |
5898 | return -EINVAL; | |
5899 | buffer = endp + 1; | |
5900 | ||
5901 | cfd = simple_strtoul(buffer, &endp, 10); | |
5902 | if ((*endp != ' ') && (*endp != '\0')) | |
5903 | return -EINVAL; | |
5904 | buffer = endp + 1; | |
5905 | ||
5906 | event = kzalloc(sizeof(*event), GFP_KERNEL); | |
5907 | if (!event) | |
5908 | return -ENOMEM; | |
5909 | ||
5910 | event->memcg = memcg; | |
5911 | INIT_LIST_HEAD(&event->list); | |
5912 | init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc); | |
5913 | init_waitqueue_func_entry(&event->wait, memcg_event_wake); | |
5914 | INIT_WORK(&event->remove, memcg_event_remove); | |
5915 | ||
5916 | efile = fdget(efd); | |
5917 | if (!efile.file) { | |
5918 | ret = -EBADF; | |
5919 | goto out_kfree; | |
5920 | } | |
5921 | ||
5922 | event->eventfd = eventfd_ctx_fileget(efile.file); | |
5923 | if (IS_ERR(event->eventfd)) { | |
5924 | ret = PTR_ERR(event->eventfd); | |
5925 | goto out_put_efile; | |
5926 | } | |
5927 | ||
5928 | cfile = fdget(cfd); | |
5929 | if (!cfile.file) { | |
5930 | ret = -EBADF; | |
5931 | goto out_put_eventfd; | |
5932 | } | |
5933 | ||
5934 | /* the process need read permission on control file */ | |
5935 | /* AV: shouldn't we check that it's been opened for read instead? */ | |
5936 | ret = inode_permission(file_inode(cfile.file), MAY_READ); | |
5937 | if (ret < 0) | |
5938 | goto out_put_cfile; | |
5939 | ||
5940 | /* | |
5941 | * Determine the event callbacks and set them in @event. This used | |
5942 | * to be done via struct cftype but cgroup core no longer knows | |
5943 | * about these events. The following is crude but the whole thing | |
5944 | * is for compatibility anyway. | |
5945 | * | |
5946 | * DO NOT ADD NEW FILES. | |
5947 | */ | |
5948 | name = cfile.file->f_dentry->d_name.name; | |
5949 | ||
5950 | if (!strcmp(name, "memory.usage_in_bytes")) { | |
5951 | event->register_event = mem_cgroup_usage_register_event; | |
5952 | event->unregister_event = mem_cgroup_usage_unregister_event; | |
5953 | } else if (!strcmp(name, "memory.oom_control")) { | |
5954 | event->register_event = mem_cgroup_oom_register_event; | |
5955 | event->unregister_event = mem_cgroup_oom_unregister_event; | |
5956 | } else if (!strcmp(name, "memory.pressure_level")) { | |
5957 | event->register_event = vmpressure_register_event; | |
5958 | event->unregister_event = vmpressure_unregister_event; | |
5959 | } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) { | |
5960 | event->register_event = memsw_cgroup_usage_register_event; | |
5961 | event->unregister_event = memsw_cgroup_usage_unregister_event; | |
5962 | } else { | |
5963 | ret = -EINVAL; | |
5964 | goto out_put_cfile; | |
5965 | } | |
5966 | ||
5967 | /* | |
5968 | * Verify @cfile should belong to @css. Also, remaining events are | |
5969 | * automatically removed on cgroup destruction but the removal is | |
5970 | * asynchronous, so take an extra ref on @css. | |
5971 | */ | |
5972 | cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent, | |
5973 | &memory_cgrp_subsys); | |
5974 | ret = -EINVAL; | |
5975 | if (IS_ERR(cfile_css)) | |
5976 | goto out_put_cfile; | |
5977 | if (cfile_css != css) { | |
5978 | css_put(cfile_css); | |
5979 | goto out_put_cfile; | |
5980 | } | |
5981 | ||
5982 | ret = event->register_event(memcg, event->eventfd, buffer); | |
5983 | if (ret) | |
5984 | goto out_put_css; | |
5985 | ||
5986 | efile.file->f_op->poll(efile.file, &event->pt); | |
5987 | ||
5988 | spin_lock(&memcg->event_list_lock); | |
5989 | list_add(&event->list, &memcg->event_list); | |
5990 | spin_unlock(&memcg->event_list_lock); | |
5991 | ||
5992 | fdput(cfile); | |
5993 | fdput(efile); | |
5994 | ||
5995 | return 0; | |
5996 | ||
5997 | out_put_css: | |
5998 | css_put(css); | |
5999 | out_put_cfile: | |
6000 | fdput(cfile); | |
6001 | out_put_eventfd: | |
6002 | eventfd_ctx_put(event->eventfd); | |
6003 | out_put_efile: | |
6004 | fdput(efile); | |
6005 | out_kfree: | |
6006 | kfree(event); | |
6007 | ||
6008 | return ret; | |
6009 | } | |
6010 | ||
6011 | static struct cftype mem_cgroup_files[] = { | |
6012 | { | |
6013 | .name = "usage_in_bytes", | |
6014 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), | |
6015 | .read_u64 = mem_cgroup_read_u64, | |
6016 | }, | |
6017 | { | |
6018 | .name = "max_usage_in_bytes", | |
6019 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), | |
6020 | .trigger = mem_cgroup_reset, | |
6021 | .read_u64 = mem_cgroup_read_u64, | |
6022 | }, | |
6023 | { | |
6024 | .name = "limit_in_bytes", | |
6025 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), | |
6026 | .write_string = mem_cgroup_write, | |
6027 | .read_u64 = mem_cgroup_read_u64, | |
6028 | }, | |
6029 | { | |
6030 | .name = "soft_limit_in_bytes", | |
6031 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), | |
6032 | .write_string = mem_cgroup_write, | |
6033 | .read_u64 = mem_cgroup_read_u64, | |
6034 | }, | |
6035 | { | |
6036 | .name = "failcnt", | |
6037 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), | |
6038 | .trigger = mem_cgroup_reset, | |
6039 | .read_u64 = mem_cgroup_read_u64, | |
6040 | }, | |
6041 | { | |
6042 | .name = "stat", | |
6043 | .seq_show = memcg_stat_show, | |
6044 | }, | |
6045 | { | |
6046 | .name = "force_empty", | |
6047 | .trigger = mem_cgroup_force_empty_write, | |
6048 | }, | |
6049 | { | |
6050 | .name = "use_hierarchy", | |
6051 | .flags = CFTYPE_INSANE, | |
6052 | .write_u64 = mem_cgroup_hierarchy_write, | |
6053 | .read_u64 = mem_cgroup_hierarchy_read, | |
6054 | }, | |
6055 | { | |
6056 | .name = "cgroup.event_control", /* XXX: for compat */ | |
6057 | .write_string = memcg_write_event_control, | |
6058 | .flags = CFTYPE_NO_PREFIX, | |
6059 | .mode = S_IWUGO, | |
6060 | }, | |
6061 | { | |
6062 | .name = "swappiness", | |
6063 | .read_u64 = mem_cgroup_swappiness_read, | |
6064 | .write_u64 = mem_cgroup_swappiness_write, | |
6065 | }, | |
6066 | { | |
6067 | .name = "move_charge_at_immigrate", | |
6068 | .read_u64 = mem_cgroup_move_charge_read, | |
6069 | .write_u64 = mem_cgroup_move_charge_write, | |
6070 | }, | |
6071 | { | |
6072 | .name = "oom_control", | |
6073 | .seq_show = mem_cgroup_oom_control_read, | |
6074 | .write_u64 = mem_cgroup_oom_control_write, | |
6075 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), | |
6076 | }, | |
6077 | { | |
6078 | .name = "pressure_level", | |
6079 | }, | |
6080 | #ifdef CONFIG_NUMA | |
6081 | { | |
6082 | .name = "numa_stat", | |
6083 | .seq_show = memcg_numa_stat_show, | |
6084 | }, | |
6085 | #endif | |
6086 | #ifdef CONFIG_MEMCG_KMEM | |
6087 | { | |
6088 | .name = "kmem.limit_in_bytes", | |
6089 | .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT), | |
6090 | .write_string = mem_cgroup_write, | |
6091 | .read_u64 = mem_cgroup_read_u64, | |
6092 | }, | |
6093 | { | |
6094 | .name = "kmem.usage_in_bytes", | |
6095 | .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE), | |
6096 | .read_u64 = mem_cgroup_read_u64, | |
6097 | }, | |
6098 | { | |
6099 | .name = "kmem.failcnt", | |
6100 | .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT), | |
6101 | .trigger = mem_cgroup_reset, | |
6102 | .read_u64 = mem_cgroup_read_u64, | |
6103 | }, | |
6104 | { | |
6105 | .name = "kmem.max_usage_in_bytes", | |
6106 | .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE), | |
6107 | .trigger = mem_cgroup_reset, | |
6108 | .read_u64 = mem_cgroup_read_u64, | |
6109 | }, | |
6110 | #ifdef CONFIG_SLABINFO | |
6111 | { | |
6112 | .name = "kmem.slabinfo", | |
6113 | .seq_show = mem_cgroup_slabinfo_read, | |
6114 | }, | |
6115 | #endif | |
6116 | #endif | |
6117 | { }, /* terminate */ | |
6118 | }; | |
6119 | ||
6120 | #ifdef CONFIG_MEMCG_SWAP | |
6121 | static struct cftype memsw_cgroup_files[] = { | |
6122 | { | |
6123 | .name = "memsw.usage_in_bytes", | |
6124 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), | |
6125 | .read_u64 = mem_cgroup_read_u64, | |
6126 | }, | |
6127 | { | |
6128 | .name = "memsw.max_usage_in_bytes", | |
6129 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), | |
6130 | .trigger = mem_cgroup_reset, | |
6131 | .read_u64 = mem_cgroup_read_u64, | |
6132 | }, | |
6133 | { | |
6134 | .name = "memsw.limit_in_bytes", | |
6135 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), | |
6136 | .write_string = mem_cgroup_write, | |
6137 | .read_u64 = mem_cgroup_read_u64, | |
6138 | }, | |
6139 | { | |
6140 | .name = "memsw.failcnt", | |
6141 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), | |
6142 | .trigger = mem_cgroup_reset, | |
6143 | .read_u64 = mem_cgroup_read_u64, | |
6144 | }, | |
6145 | { }, /* terminate */ | |
6146 | }; | |
6147 | #endif | |
6148 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) | |
6149 | { | |
6150 | struct mem_cgroup_per_node *pn; | |
6151 | struct mem_cgroup_per_zone *mz; | |
6152 | int zone, tmp = node; | |
6153 | /* | |
6154 | * This routine is called against possible nodes. | |
6155 | * But it's BUG to call kmalloc() against offline node. | |
6156 | * | |
6157 | * TODO: this routine can waste much memory for nodes which will | |
6158 | * never be onlined. It's better to use memory hotplug callback | |
6159 | * function. | |
6160 | */ | |
6161 | if (!node_state(node, N_NORMAL_MEMORY)) | |
6162 | tmp = -1; | |
6163 | pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); | |
6164 | if (!pn) | |
6165 | return 1; | |
6166 | ||
6167 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
6168 | mz = &pn->zoneinfo[zone]; | |
6169 | lruvec_init(&mz->lruvec); | |
6170 | mz->usage_in_excess = 0; | |
6171 | mz->on_tree = false; | |
6172 | mz->memcg = memcg; | |
6173 | } | |
6174 | memcg->nodeinfo[node] = pn; | |
6175 | return 0; | |
6176 | } | |
6177 | ||
6178 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) | |
6179 | { | |
6180 | kfree(memcg->nodeinfo[node]); | |
6181 | } | |
6182 | ||
6183 | static struct mem_cgroup *mem_cgroup_alloc(void) | |
6184 | { | |
6185 | struct mem_cgroup *memcg; | |
6186 | size_t size; | |
6187 | ||
6188 | size = sizeof(struct mem_cgroup); | |
6189 | size += nr_node_ids * sizeof(struct mem_cgroup_per_node *); | |
6190 | ||
6191 | memcg = kzalloc(size, GFP_KERNEL); | |
6192 | if (!memcg) | |
6193 | return NULL; | |
6194 | ||
6195 | memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu); | |
6196 | if (!memcg->stat) | |
6197 | goto out_free; | |
6198 | spin_lock_init(&memcg->pcp_counter_lock); | |
6199 | return memcg; | |
6200 | ||
6201 | out_free: | |
6202 | kfree(memcg); | |
6203 | return NULL; | |
6204 | } | |
6205 | ||
6206 | /* | |
6207 | * At destroying mem_cgroup, references from swap_cgroup can remain. | |
6208 | * (scanning all at force_empty is too costly...) | |
6209 | * | |
6210 | * Instead of clearing all references at force_empty, we remember | |
6211 | * the number of reference from swap_cgroup and free mem_cgroup when | |
6212 | * it goes down to 0. | |
6213 | * | |
6214 | * Removal of cgroup itself succeeds regardless of refs from swap. | |
6215 | */ | |
6216 | ||
6217 | static void __mem_cgroup_free(struct mem_cgroup *memcg) | |
6218 | { | |
6219 | int node; | |
6220 | ||
6221 | mem_cgroup_remove_from_trees(memcg); | |
6222 | ||
6223 | for_each_node(node) | |
6224 | free_mem_cgroup_per_zone_info(memcg, node); | |
6225 | ||
6226 | free_percpu(memcg->stat); | |
6227 | ||
6228 | /* | |
6229 | * We need to make sure that (at least for now), the jump label | |
6230 | * destruction code runs outside of the cgroup lock. This is because | |
6231 | * get_online_cpus(), which is called from the static_branch update, | |
6232 | * can't be called inside the cgroup_lock. cpusets are the ones | |
6233 | * enforcing this dependency, so if they ever change, we might as well. | |
6234 | * | |
6235 | * schedule_work() will guarantee this happens. Be careful if you need | |
6236 | * to move this code around, and make sure it is outside | |
6237 | * the cgroup_lock. | |
6238 | */ | |
6239 | disarm_static_keys(memcg); | |
6240 | kfree(memcg); | |
6241 | } | |
6242 | ||
6243 | /* | |
6244 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. | |
6245 | */ | |
6246 | struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) | |
6247 | { | |
6248 | if (!memcg->res.parent) | |
6249 | return NULL; | |
6250 | return mem_cgroup_from_res_counter(memcg->res.parent, res); | |
6251 | } | |
6252 | EXPORT_SYMBOL(parent_mem_cgroup); | |
6253 | ||
6254 | static void __init mem_cgroup_soft_limit_tree_init(void) | |
6255 | { | |
6256 | struct mem_cgroup_tree_per_node *rtpn; | |
6257 | struct mem_cgroup_tree_per_zone *rtpz; | |
6258 | int tmp, node, zone; | |
6259 | ||
6260 | for_each_node(node) { | |
6261 | tmp = node; | |
6262 | if (!node_state(node, N_NORMAL_MEMORY)) | |
6263 | tmp = -1; | |
6264 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); | |
6265 | BUG_ON(!rtpn); | |
6266 | ||
6267 | soft_limit_tree.rb_tree_per_node[node] = rtpn; | |
6268 | ||
6269 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
6270 | rtpz = &rtpn->rb_tree_per_zone[zone]; | |
6271 | rtpz->rb_root = RB_ROOT; | |
6272 | spin_lock_init(&rtpz->lock); | |
6273 | } | |
6274 | } | |
6275 | } | |
6276 | ||
6277 | static struct cgroup_subsys_state * __ref | |
6278 | mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
6279 | { | |
6280 | struct mem_cgroup *memcg; | |
6281 | long error = -ENOMEM; | |
6282 | int node; | |
6283 | ||
6284 | memcg = mem_cgroup_alloc(); | |
6285 | if (!memcg) | |
6286 | return ERR_PTR(error); | |
6287 | ||
6288 | for_each_node(node) | |
6289 | if (alloc_mem_cgroup_per_zone_info(memcg, node)) | |
6290 | goto free_out; | |
6291 | ||
6292 | /* root ? */ | |
6293 | if (parent_css == NULL) { | |
6294 | root_mem_cgroup = memcg; | |
6295 | res_counter_init(&memcg->res, NULL); | |
6296 | res_counter_init(&memcg->memsw, NULL); | |
6297 | res_counter_init(&memcg->kmem, NULL); | |
6298 | } | |
6299 | ||
6300 | memcg->last_scanned_node = MAX_NUMNODES; | |
6301 | INIT_LIST_HEAD(&memcg->oom_notify); | |
6302 | memcg->move_charge_at_immigrate = 0; | |
6303 | mutex_init(&memcg->thresholds_lock); | |
6304 | spin_lock_init(&memcg->move_lock); | |
6305 | vmpressure_init(&memcg->vmpressure); | |
6306 | INIT_LIST_HEAD(&memcg->event_list); | |
6307 | spin_lock_init(&memcg->event_list_lock); | |
6308 | ||
6309 | return &memcg->css; | |
6310 | ||
6311 | free_out: | |
6312 | __mem_cgroup_free(memcg); | |
6313 | return ERR_PTR(error); | |
6314 | } | |
6315 | ||
6316 | static int | |
6317 | mem_cgroup_css_online(struct cgroup_subsys_state *css) | |
6318 | { | |
6319 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
6320 | struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css)); | |
6321 | ||
6322 | if (css->cgroup->id > MEM_CGROUP_ID_MAX) | |
6323 | return -ENOSPC; | |
6324 | ||
6325 | if (!parent) | |
6326 | return 0; | |
6327 | ||
6328 | mutex_lock(&memcg_create_mutex); | |
6329 | ||
6330 | memcg->use_hierarchy = parent->use_hierarchy; | |
6331 | memcg->oom_kill_disable = parent->oom_kill_disable; | |
6332 | memcg->swappiness = mem_cgroup_swappiness(parent); | |
6333 | ||
6334 | if (parent->use_hierarchy) { | |
6335 | res_counter_init(&memcg->res, &parent->res); | |
6336 | res_counter_init(&memcg->memsw, &parent->memsw); | |
6337 | res_counter_init(&memcg->kmem, &parent->kmem); | |
6338 | ||
6339 | /* | |
6340 | * No need to take a reference to the parent because cgroup | |
6341 | * core guarantees its existence. | |
6342 | */ | |
6343 | } else { | |
6344 | res_counter_init(&memcg->res, NULL); | |
6345 | res_counter_init(&memcg->memsw, NULL); | |
6346 | res_counter_init(&memcg->kmem, NULL); | |
6347 | /* | |
6348 | * Deeper hierachy with use_hierarchy == false doesn't make | |
6349 | * much sense so let cgroup subsystem know about this | |
6350 | * unfortunate state in our controller. | |
6351 | */ | |
6352 | if (parent != root_mem_cgroup) | |
6353 | memory_cgrp_subsys.broken_hierarchy = true; | |
6354 | } | |
6355 | mutex_unlock(&memcg_create_mutex); | |
6356 | ||
6357 | return memcg_init_kmem(memcg, &memory_cgrp_subsys); | |
6358 | } | |
6359 | ||
6360 | /* | |
6361 | * Announce all parents that a group from their hierarchy is gone. | |
6362 | */ | |
6363 | static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg) | |
6364 | { | |
6365 | struct mem_cgroup *parent = memcg; | |
6366 | ||
6367 | while ((parent = parent_mem_cgroup(parent))) | |
6368 | mem_cgroup_iter_invalidate(parent); | |
6369 | ||
6370 | /* | |
6371 | * if the root memcg is not hierarchical we have to check it | |
6372 | * explicitely. | |
6373 | */ | |
6374 | if (!root_mem_cgroup->use_hierarchy) | |
6375 | mem_cgroup_iter_invalidate(root_mem_cgroup); | |
6376 | } | |
6377 | ||
6378 | static void mem_cgroup_css_offline(struct cgroup_subsys_state *css) | |
6379 | { | |
6380 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
6381 | struct mem_cgroup_event *event, *tmp; | |
6382 | struct cgroup_subsys_state *iter; | |
6383 | ||
6384 | /* | |
6385 | * Unregister events and notify userspace. | |
6386 | * Notify userspace about cgroup removing only after rmdir of cgroup | |
6387 | * directory to avoid race between userspace and kernelspace. | |
6388 | */ | |
6389 | spin_lock(&memcg->event_list_lock); | |
6390 | list_for_each_entry_safe(event, tmp, &memcg->event_list, list) { | |
6391 | list_del_init(&event->list); | |
6392 | schedule_work(&event->remove); | |
6393 | } | |
6394 | spin_unlock(&memcg->event_list_lock); | |
6395 | ||
6396 | kmem_cgroup_css_offline(memcg); | |
6397 | ||
6398 | mem_cgroup_invalidate_reclaim_iterators(memcg); | |
6399 | ||
6400 | /* | |
6401 | * This requires that offlining is serialized. Right now that is | |
6402 | * guaranteed because css_killed_work_fn() holds the cgroup_mutex. | |
6403 | */ | |
6404 | css_for_each_descendant_post(iter, css) | |
6405 | mem_cgroup_reparent_charges(mem_cgroup_from_css(iter)); | |
6406 | ||
6407 | mem_cgroup_destroy_all_caches(memcg); | |
6408 | vmpressure_cleanup(&memcg->vmpressure); | |
6409 | } | |
6410 | ||
6411 | static void mem_cgroup_css_free(struct cgroup_subsys_state *css) | |
6412 | { | |
6413 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
6414 | /* | |
6415 | * XXX: css_offline() would be where we should reparent all | |
6416 | * memory to prepare the cgroup for destruction. However, | |
6417 | * memcg does not do css_tryget() and res_counter charging | |
6418 | * under the same RCU lock region, which means that charging | |
6419 | * could race with offlining. Offlining only happens to | |
6420 | * cgroups with no tasks in them but charges can show up | |
6421 | * without any tasks from the swapin path when the target | |
6422 | * memcg is looked up from the swapout record and not from the | |
6423 | * current task as it usually is. A race like this can leak | |
6424 | * charges and put pages with stale cgroup pointers into | |
6425 | * circulation: | |
6426 | * | |
6427 | * #0 #1 | |
6428 | * lookup_swap_cgroup_id() | |
6429 | * rcu_read_lock() | |
6430 | * mem_cgroup_lookup() | |
6431 | * css_tryget() | |
6432 | * rcu_read_unlock() | |
6433 | * disable css_tryget() | |
6434 | * call_rcu() | |
6435 | * offline_css() | |
6436 | * reparent_charges() | |
6437 | * res_counter_charge() | |
6438 | * css_put() | |
6439 | * css_free() | |
6440 | * pc->mem_cgroup = dead memcg | |
6441 | * add page to lru | |
6442 | * | |
6443 | * The bulk of the charges are still moved in offline_css() to | |
6444 | * avoid pinning a lot of pages in case a long-term reference | |
6445 | * like a swapout record is deferring the css_free() to long | |
6446 | * after offlining. But this makes sure we catch any charges | |
6447 | * made after offlining: | |
6448 | */ | |
6449 | mem_cgroup_reparent_charges(memcg); | |
6450 | ||
6451 | memcg_destroy_kmem(memcg); | |
6452 | __mem_cgroup_free(memcg); | |
6453 | } | |
6454 | ||
6455 | #ifdef CONFIG_MMU | |
6456 | /* Handlers for move charge at task migration. */ | |
6457 | #define PRECHARGE_COUNT_AT_ONCE 256 | |
6458 | static int mem_cgroup_do_precharge(unsigned long count) | |
6459 | { | |
6460 | int ret = 0; | |
6461 | int batch_count = PRECHARGE_COUNT_AT_ONCE; | |
6462 | struct mem_cgroup *memcg = mc.to; | |
6463 | ||
6464 | if (mem_cgroup_is_root(memcg)) { | |
6465 | mc.precharge += count; | |
6466 | /* we don't need css_get for root */ | |
6467 | return ret; | |
6468 | } | |
6469 | /* try to charge at once */ | |
6470 | if (count > 1) { | |
6471 | struct res_counter *dummy; | |
6472 | /* | |
6473 | * "memcg" cannot be under rmdir() because we've already checked | |
6474 | * by cgroup_lock_live_cgroup() that it is not removed and we | |
6475 | * are still under the same cgroup_mutex. So we can postpone | |
6476 | * css_get(). | |
6477 | */ | |
6478 | if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) | |
6479 | goto one_by_one; | |
6480 | if (do_swap_account && res_counter_charge(&memcg->memsw, | |
6481 | PAGE_SIZE * count, &dummy)) { | |
6482 | res_counter_uncharge(&memcg->res, PAGE_SIZE * count); | |
6483 | goto one_by_one; | |
6484 | } | |
6485 | mc.precharge += count; | |
6486 | return ret; | |
6487 | } | |
6488 | one_by_one: | |
6489 | /* fall back to one by one charge */ | |
6490 | while (count--) { | |
6491 | if (signal_pending(current)) { | |
6492 | ret = -EINTR; | |
6493 | break; | |
6494 | } | |
6495 | if (!batch_count--) { | |
6496 | batch_count = PRECHARGE_COUNT_AT_ONCE; | |
6497 | cond_resched(); | |
6498 | } | |
6499 | ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false); | |
6500 | if (ret) | |
6501 | /* mem_cgroup_clear_mc() will do uncharge later */ | |
6502 | return ret; | |
6503 | mc.precharge++; | |
6504 | } | |
6505 | return ret; | |
6506 | } | |
6507 | ||
6508 | /** | |
6509 | * get_mctgt_type - get target type of moving charge | |
6510 | * @vma: the vma the pte to be checked belongs | |
6511 | * @addr: the address corresponding to the pte to be checked | |
6512 | * @ptent: the pte to be checked | |
6513 | * @target: the pointer the target page or swap ent will be stored(can be NULL) | |
6514 | * | |
6515 | * Returns | |
6516 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. | |
6517 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for | |
6518 | * move charge. if @target is not NULL, the page is stored in target->page | |
6519 | * with extra refcnt got(Callers should handle it). | |
6520 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a | |
6521 | * target for charge migration. if @target is not NULL, the entry is stored | |
6522 | * in target->ent. | |
6523 | * | |
6524 | * Called with pte lock held. | |
6525 | */ | |
6526 | union mc_target { | |
6527 | struct page *page; | |
6528 | swp_entry_t ent; | |
6529 | }; | |
6530 | ||
6531 | enum mc_target_type { | |
6532 | MC_TARGET_NONE = 0, | |
6533 | MC_TARGET_PAGE, | |
6534 | MC_TARGET_SWAP, | |
6535 | }; | |
6536 | ||
6537 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, | |
6538 | unsigned long addr, pte_t ptent) | |
6539 | { | |
6540 | struct page *page = vm_normal_page(vma, addr, ptent); | |
6541 | ||
6542 | if (!page || !page_mapped(page)) | |
6543 | return NULL; | |
6544 | if (PageAnon(page)) { | |
6545 | /* we don't move shared anon */ | |
6546 | if (!move_anon()) | |
6547 | return NULL; | |
6548 | } else if (!move_file()) | |
6549 | /* we ignore mapcount for file pages */ | |
6550 | return NULL; | |
6551 | if (!get_page_unless_zero(page)) | |
6552 | return NULL; | |
6553 | ||
6554 | return page; | |
6555 | } | |
6556 | ||
6557 | #ifdef CONFIG_SWAP | |
6558 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
6559 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
6560 | { | |
6561 | struct page *page = NULL; | |
6562 | swp_entry_t ent = pte_to_swp_entry(ptent); | |
6563 | ||
6564 | if (!move_anon() || non_swap_entry(ent)) | |
6565 | return NULL; | |
6566 | /* | |
6567 | * Because lookup_swap_cache() updates some statistics counter, | |
6568 | * we call find_get_page() with swapper_space directly. | |
6569 | */ | |
6570 | page = find_get_page(swap_address_space(ent), ent.val); | |
6571 | if (do_swap_account) | |
6572 | entry->val = ent.val; | |
6573 | ||
6574 | return page; | |
6575 | } | |
6576 | #else | |
6577 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
6578 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
6579 | { | |
6580 | return NULL; | |
6581 | } | |
6582 | #endif | |
6583 | ||
6584 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, | |
6585 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
6586 | { | |
6587 | struct page *page = NULL; | |
6588 | struct address_space *mapping; | |
6589 | pgoff_t pgoff; | |
6590 | ||
6591 | if (!vma->vm_file) /* anonymous vma */ | |
6592 | return NULL; | |
6593 | if (!move_file()) | |
6594 | return NULL; | |
6595 | ||
6596 | mapping = vma->vm_file->f_mapping; | |
6597 | if (pte_none(ptent)) | |
6598 | pgoff = linear_page_index(vma, addr); | |
6599 | else /* pte_file(ptent) is true */ | |
6600 | pgoff = pte_to_pgoff(ptent); | |
6601 | ||
6602 | /* page is moved even if it's not RSS of this task(page-faulted). */ | |
6603 | #ifdef CONFIG_SWAP | |
6604 | /* shmem/tmpfs may report page out on swap: account for that too. */ | |
6605 | if (shmem_mapping(mapping)) { | |
6606 | page = find_get_entry(mapping, pgoff); | |
6607 | if (radix_tree_exceptional_entry(page)) { | |
6608 | swp_entry_t swp = radix_to_swp_entry(page); | |
6609 | if (do_swap_account) | |
6610 | *entry = swp; | |
6611 | page = find_get_page(swap_address_space(swp), swp.val); | |
6612 | } | |
6613 | } else | |
6614 | page = find_get_page(mapping, pgoff); | |
6615 | #else | |
6616 | page = find_get_page(mapping, pgoff); | |
6617 | #endif | |
6618 | return page; | |
6619 | } | |
6620 | ||
6621 | static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma, | |
6622 | unsigned long addr, pte_t ptent, union mc_target *target) | |
6623 | { | |
6624 | struct page *page = NULL; | |
6625 | struct page_cgroup *pc; | |
6626 | enum mc_target_type ret = MC_TARGET_NONE; | |
6627 | swp_entry_t ent = { .val = 0 }; | |
6628 | ||
6629 | if (pte_present(ptent)) | |
6630 | page = mc_handle_present_pte(vma, addr, ptent); | |
6631 | else if (is_swap_pte(ptent)) | |
6632 | page = mc_handle_swap_pte(vma, addr, ptent, &ent); | |
6633 | else if (pte_none(ptent) || pte_file(ptent)) | |
6634 | page = mc_handle_file_pte(vma, addr, ptent, &ent); | |
6635 | ||
6636 | if (!page && !ent.val) | |
6637 | return ret; | |
6638 | if (page) { | |
6639 | pc = lookup_page_cgroup(page); | |
6640 | /* | |
6641 | * Do only loose check w/o page_cgroup lock. | |
6642 | * mem_cgroup_move_account() checks the pc is valid or not under | |
6643 | * the lock. | |
6644 | */ | |
6645 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { | |
6646 | ret = MC_TARGET_PAGE; | |
6647 | if (target) | |
6648 | target->page = page; | |
6649 | } | |
6650 | if (!ret || !target) | |
6651 | put_page(page); | |
6652 | } | |
6653 | /* There is a swap entry and a page doesn't exist or isn't charged */ | |
6654 | if (ent.val && !ret && | |
6655 | mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) { | |
6656 | ret = MC_TARGET_SWAP; | |
6657 | if (target) | |
6658 | target->ent = ent; | |
6659 | } | |
6660 | return ret; | |
6661 | } | |
6662 | ||
6663 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
6664 | /* | |
6665 | * We don't consider swapping or file mapped pages because THP does not | |
6666 | * support them for now. | |
6667 | * Caller should make sure that pmd_trans_huge(pmd) is true. | |
6668 | */ | |
6669 | static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, | |
6670 | unsigned long addr, pmd_t pmd, union mc_target *target) | |
6671 | { | |
6672 | struct page *page = NULL; | |
6673 | struct page_cgroup *pc; | |
6674 | enum mc_target_type ret = MC_TARGET_NONE; | |
6675 | ||
6676 | page = pmd_page(pmd); | |
6677 | VM_BUG_ON_PAGE(!page || !PageHead(page), page); | |
6678 | if (!move_anon()) | |
6679 | return ret; | |
6680 | pc = lookup_page_cgroup(page); | |
6681 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { | |
6682 | ret = MC_TARGET_PAGE; | |
6683 | if (target) { | |
6684 | get_page(page); | |
6685 | target->page = page; | |
6686 | } | |
6687 | } | |
6688 | return ret; | |
6689 | } | |
6690 | #else | |
6691 | static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, | |
6692 | unsigned long addr, pmd_t pmd, union mc_target *target) | |
6693 | { | |
6694 | return MC_TARGET_NONE; | |
6695 | } | |
6696 | #endif | |
6697 | ||
6698 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, | |
6699 | unsigned long addr, unsigned long end, | |
6700 | struct mm_walk *walk) | |
6701 | { | |
6702 | struct vm_area_struct *vma = walk->private; | |
6703 | pte_t *pte; | |
6704 | spinlock_t *ptl; | |
6705 | ||
6706 | if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { | |
6707 | if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE) | |
6708 | mc.precharge += HPAGE_PMD_NR; | |
6709 | spin_unlock(ptl); | |
6710 | return 0; | |
6711 | } | |
6712 | ||
6713 | if (pmd_trans_unstable(pmd)) | |
6714 | return 0; | |
6715 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
6716 | for (; addr != end; pte++, addr += PAGE_SIZE) | |
6717 | if (get_mctgt_type(vma, addr, *pte, NULL)) | |
6718 | mc.precharge++; /* increment precharge temporarily */ | |
6719 | pte_unmap_unlock(pte - 1, ptl); | |
6720 | cond_resched(); | |
6721 | ||
6722 | return 0; | |
6723 | } | |
6724 | ||
6725 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) | |
6726 | { | |
6727 | unsigned long precharge; | |
6728 | struct vm_area_struct *vma; | |
6729 | ||
6730 | down_read(&mm->mmap_sem); | |
6731 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
6732 | struct mm_walk mem_cgroup_count_precharge_walk = { | |
6733 | .pmd_entry = mem_cgroup_count_precharge_pte_range, | |
6734 | .mm = mm, | |
6735 | .private = vma, | |
6736 | }; | |
6737 | if (is_vm_hugetlb_page(vma)) | |
6738 | continue; | |
6739 | walk_page_range(vma->vm_start, vma->vm_end, | |
6740 | &mem_cgroup_count_precharge_walk); | |
6741 | } | |
6742 | up_read(&mm->mmap_sem); | |
6743 | ||
6744 | precharge = mc.precharge; | |
6745 | mc.precharge = 0; | |
6746 | ||
6747 | return precharge; | |
6748 | } | |
6749 | ||
6750 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) | |
6751 | { | |
6752 | unsigned long precharge = mem_cgroup_count_precharge(mm); | |
6753 | ||
6754 | VM_BUG_ON(mc.moving_task); | |
6755 | mc.moving_task = current; | |
6756 | return mem_cgroup_do_precharge(precharge); | |
6757 | } | |
6758 | ||
6759 | /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ | |
6760 | static void __mem_cgroup_clear_mc(void) | |
6761 | { | |
6762 | struct mem_cgroup *from = mc.from; | |
6763 | struct mem_cgroup *to = mc.to; | |
6764 | int i; | |
6765 | ||
6766 | /* we must uncharge all the leftover precharges from mc.to */ | |
6767 | if (mc.precharge) { | |
6768 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); | |
6769 | mc.precharge = 0; | |
6770 | } | |
6771 | /* | |
6772 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so | |
6773 | * we must uncharge here. | |
6774 | */ | |
6775 | if (mc.moved_charge) { | |
6776 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); | |
6777 | mc.moved_charge = 0; | |
6778 | } | |
6779 | /* we must fixup refcnts and charges */ | |
6780 | if (mc.moved_swap) { | |
6781 | /* uncharge swap account from the old cgroup */ | |
6782 | if (!mem_cgroup_is_root(mc.from)) | |
6783 | res_counter_uncharge(&mc.from->memsw, | |
6784 | PAGE_SIZE * mc.moved_swap); | |
6785 | ||
6786 | for (i = 0; i < mc.moved_swap; i++) | |
6787 | css_put(&mc.from->css); | |
6788 | ||
6789 | if (!mem_cgroup_is_root(mc.to)) { | |
6790 | /* | |
6791 | * we charged both to->res and to->memsw, so we should | |
6792 | * uncharge to->res. | |
6793 | */ | |
6794 | res_counter_uncharge(&mc.to->res, | |
6795 | PAGE_SIZE * mc.moved_swap); | |
6796 | } | |
6797 | /* we've already done css_get(mc.to) */ | |
6798 | mc.moved_swap = 0; | |
6799 | } | |
6800 | memcg_oom_recover(from); | |
6801 | memcg_oom_recover(to); | |
6802 | wake_up_all(&mc.waitq); | |
6803 | } | |
6804 | ||
6805 | static void mem_cgroup_clear_mc(void) | |
6806 | { | |
6807 | struct mem_cgroup *from = mc.from; | |
6808 | ||
6809 | /* | |
6810 | * we must clear moving_task before waking up waiters at the end of | |
6811 | * task migration. | |
6812 | */ | |
6813 | mc.moving_task = NULL; | |
6814 | __mem_cgroup_clear_mc(); | |
6815 | spin_lock(&mc.lock); | |
6816 | mc.from = NULL; | |
6817 | mc.to = NULL; | |
6818 | spin_unlock(&mc.lock); | |
6819 | mem_cgroup_end_move(from); | |
6820 | } | |
6821 | ||
6822 | static int mem_cgroup_can_attach(struct cgroup_subsys_state *css, | |
6823 | struct cgroup_taskset *tset) | |
6824 | { | |
6825 | struct task_struct *p = cgroup_taskset_first(tset); | |
6826 | int ret = 0; | |
6827 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
6828 | unsigned long move_charge_at_immigrate; | |
6829 | ||
6830 | /* | |
6831 | * We are now commited to this value whatever it is. Changes in this | |
6832 | * tunable will only affect upcoming migrations, not the current one. | |
6833 | * So we need to save it, and keep it going. | |
6834 | */ | |
6835 | move_charge_at_immigrate = memcg->move_charge_at_immigrate; | |
6836 | if (move_charge_at_immigrate) { | |
6837 | struct mm_struct *mm; | |
6838 | struct mem_cgroup *from = mem_cgroup_from_task(p); | |
6839 | ||
6840 | VM_BUG_ON(from == memcg); | |
6841 | ||
6842 | mm = get_task_mm(p); | |
6843 | if (!mm) | |
6844 | return 0; | |
6845 | /* We move charges only when we move a owner of the mm */ | |
6846 | if (mm->owner == p) { | |
6847 | VM_BUG_ON(mc.from); | |
6848 | VM_BUG_ON(mc.to); | |
6849 | VM_BUG_ON(mc.precharge); | |
6850 | VM_BUG_ON(mc.moved_charge); | |
6851 | VM_BUG_ON(mc.moved_swap); | |
6852 | mem_cgroup_start_move(from); | |
6853 | spin_lock(&mc.lock); | |
6854 | mc.from = from; | |
6855 | mc.to = memcg; | |
6856 | mc.immigrate_flags = move_charge_at_immigrate; | |
6857 | spin_unlock(&mc.lock); | |
6858 | /* We set mc.moving_task later */ | |
6859 | ||
6860 | ret = mem_cgroup_precharge_mc(mm); | |
6861 | if (ret) | |
6862 | mem_cgroup_clear_mc(); | |
6863 | } | |
6864 | mmput(mm); | |
6865 | } | |
6866 | return ret; | |
6867 | } | |
6868 | ||
6869 | static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css, | |
6870 | struct cgroup_taskset *tset) | |
6871 | { | |
6872 | mem_cgroup_clear_mc(); | |
6873 | } | |
6874 | ||
6875 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, | |
6876 | unsigned long addr, unsigned long end, | |
6877 | struct mm_walk *walk) | |
6878 | { | |
6879 | int ret = 0; | |
6880 | struct vm_area_struct *vma = walk->private; | |
6881 | pte_t *pte; | |
6882 | spinlock_t *ptl; | |
6883 | enum mc_target_type target_type; | |
6884 | union mc_target target; | |
6885 | struct page *page; | |
6886 | struct page_cgroup *pc; | |
6887 | ||
6888 | /* | |
6889 | * We don't take compound_lock() here but no race with splitting thp | |
6890 | * happens because: | |
6891 | * - if pmd_trans_huge_lock() returns 1, the relevant thp is not | |
6892 | * under splitting, which means there's no concurrent thp split, | |
6893 | * - if another thread runs into split_huge_page() just after we | |
6894 | * entered this if-block, the thread must wait for page table lock | |
6895 | * to be unlocked in __split_huge_page_splitting(), where the main | |
6896 | * part of thp split is not executed yet. | |
6897 | */ | |
6898 | if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) { | |
6899 | if (mc.precharge < HPAGE_PMD_NR) { | |
6900 | spin_unlock(ptl); | |
6901 | return 0; | |
6902 | } | |
6903 | target_type = get_mctgt_type_thp(vma, addr, *pmd, &target); | |
6904 | if (target_type == MC_TARGET_PAGE) { | |
6905 | page = target.page; | |
6906 | if (!isolate_lru_page(page)) { | |
6907 | pc = lookup_page_cgroup(page); | |
6908 | if (!mem_cgroup_move_account(page, HPAGE_PMD_NR, | |
6909 | pc, mc.from, mc.to)) { | |
6910 | mc.precharge -= HPAGE_PMD_NR; | |
6911 | mc.moved_charge += HPAGE_PMD_NR; | |
6912 | } | |
6913 | putback_lru_page(page); | |
6914 | } | |
6915 | put_page(page); | |
6916 | } | |
6917 | spin_unlock(ptl); | |
6918 | return 0; | |
6919 | } | |
6920 | ||
6921 | if (pmd_trans_unstable(pmd)) | |
6922 | return 0; | |
6923 | retry: | |
6924 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
6925 | for (; addr != end; addr += PAGE_SIZE) { | |
6926 | pte_t ptent = *(pte++); | |
6927 | swp_entry_t ent; | |
6928 | ||
6929 | if (!mc.precharge) | |
6930 | break; | |
6931 | ||
6932 | switch (get_mctgt_type(vma, addr, ptent, &target)) { | |
6933 | case MC_TARGET_PAGE: | |
6934 | page = target.page; | |
6935 | if (isolate_lru_page(page)) | |
6936 | goto put; | |
6937 | pc = lookup_page_cgroup(page); | |
6938 | if (!mem_cgroup_move_account(page, 1, pc, | |
6939 | mc.from, mc.to)) { | |
6940 | mc.precharge--; | |
6941 | /* we uncharge from mc.from later. */ | |
6942 | mc.moved_charge++; | |
6943 | } | |
6944 | putback_lru_page(page); | |
6945 | put: /* get_mctgt_type() gets the page */ | |
6946 | put_page(page); | |
6947 | break; | |
6948 | case MC_TARGET_SWAP: | |
6949 | ent = target.ent; | |
6950 | if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) { | |
6951 | mc.precharge--; | |
6952 | /* we fixup refcnts and charges later. */ | |
6953 | mc.moved_swap++; | |
6954 | } | |
6955 | break; | |
6956 | default: | |
6957 | break; | |
6958 | } | |
6959 | } | |
6960 | pte_unmap_unlock(pte - 1, ptl); | |
6961 | cond_resched(); | |
6962 | ||
6963 | if (addr != end) { | |
6964 | /* | |
6965 | * We have consumed all precharges we got in can_attach(). | |
6966 | * We try charge one by one, but don't do any additional | |
6967 | * charges to mc.to if we have failed in charge once in attach() | |
6968 | * phase. | |
6969 | */ | |
6970 | ret = mem_cgroup_do_precharge(1); | |
6971 | if (!ret) | |
6972 | goto retry; | |
6973 | } | |
6974 | ||
6975 | return ret; | |
6976 | } | |
6977 | ||
6978 | static void mem_cgroup_move_charge(struct mm_struct *mm) | |
6979 | { | |
6980 | struct vm_area_struct *vma; | |
6981 | ||
6982 | lru_add_drain_all(); | |
6983 | retry: | |
6984 | if (unlikely(!down_read_trylock(&mm->mmap_sem))) { | |
6985 | /* | |
6986 | * Someone who are holding the mmap_sem might be waiting in | |
6987 | * waitq. So we cancel all extra charges, wake up all waiters, | |
6988 | * and retry. Because we cancel precharges, we might not be able | |
6989 | * to move enough charges, but moving charge is a best-effort | |
6990 | * feature anyway, so it wouldn't be a big problem. | |
6991 | */ | |
6992 | __mem_cgroup_clear_mc(); | |
6993 | cond_resched(); | |
6994 | goto retry; | |
6995 | } | |
6996 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
6997 | int ret; | |
6998 | struct mm_walk mem_cgroup_move_charge_walk = { | |
6999 | .pmd_entry = mem_cgroup_move_charge_pte_range, | |
7000 | .mm = mm, | |
7001 | .private = vma, | |
7002 | }; | |
7003 | if (is_vm_hugetlb_page(vma)) | |
7004 | continue; | |
7005 | ret = walk_page_range(vma->vm_start, vma->vm_end, | |
7006 | &mem_cgroup_move_charge_walk); | |
7007 | if (ret) | |
7008 | /* | |
7009 | * means we have consumed all precharges and failed in | |
7010 | * doing additional charge. Just abandon here. | |
7011 | */ | |
7012 | break; | |
7013 | } | |
7014 | up_read(&mm->mmap_sem); | |
7015 | } | |
7016 | ||
7017 | static void mem_cgroup_move_task(struct cgroup_subsys_state *css, | |
7018 | struct cgroup_taskset *tset) | |
7019 | { | |
7020 | struct task_struct *p = cgroup_taskset_first(tset); | |
7021 | struct mm_struct *mm = get_task_mm(p); | |
7022 | ||
7023 | if (mm) { | |
7024 | if (mc.to) | |
7025 | mem_cgroup_move_charge(mm); | |
7026 | mmput(mm); | |
7027 | } | |
7028 | if (mc.to) | |
7029 | mem_cgroup_clear_mc(); | |
7030 | } | |
7031 | #else /* !CONFIG_MMU */ | |
7032 | static int mem_cgroup_can_attach(struct cgroup_subsys_state *css, | |
7033 | struct cgroup_taskset *tset) | |
7034 | { | |
7035 | return 0; | |
7036 | } | |
7037 | static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css, | |
7038 | struct cgroup_taskset *tset) | |
7039 | { | |
7040 | } | |
7041 | static void mem_cgroup_move_task(struct cgroup_subsys_state *css, | |
7042 | struct cgroup_taskset *tset) | |
7043 | { | |
7044 | } | |
7045 | #endif | |
7046 | ||
7047 | /* | |
7048 | * Cgroup retains root cgroups across [un]mount cycles making it necessary | |
7049 | * to verify sane_behavior flag on each mount attempt. | |
7050 | */ | |
7051 | static void mem_cgroup_bind(struct cgroup_subsys_state *root_css) | |
7052 | { | |
7053 | /* | |
7054 | * use_hierarchy is forced with sane_behavior. cgroup core | |
7055 | * guarantees that @root doesn't have any children, so turning it | |
7056 | * on for the root memcg is enough. | |
7057 | */ | |
7058 | if (cgroup_sane_behavior(root_css->cgroup)) | |
7059 | mem_cgroup_from_css(root_css)->use_hierarchy = true; | |
7060 | } | |
7061 | ||
7062 | struct cgroup_subsys memory_cgrp_subsys = { | |
7063 | .css_alloc = mem_cgroup_css_alloc, | |
7064 | .css_online = mem_cgroup_css_online, | |
7065 | .css_offline = mem_cgroup_css_offline, | |
7066 | .css_free = mem_cgroup_css_free, | |
7067 | .can_attach = mem_cgroup_can_attach, | |
7068 | .cancel_attach = mem_cgroup_cancel_attach, | |
7069 | .attach = mem_cgroup_move_task, | |
7070 | .bind = mem_cgroup_bind, | |
7071 | .base_cftypes = mem_cgroup_files, | |
7072 | .early_init = 0, | |
7073 | }; | |
7074 | ||
7075 | #ifdef CONFIG_MEMCG_SWAP | |
7076 | static int __init enable_swap_account(char *s) | |
7077 | { | |
7078 | if (!strcmp(s, "1")) | |
7079 | really_do_swap_account = 1; | |
7080 | else if (!strcmp(s, "0")) | |
7081 | really_do_swap_account = 0; | |
7082 | return 1; | |
7083 | } | |
7084 | __setup("swapaccount=", enable_swap_account); | |
7085 | ||
7086 | static void __init memsw_file_init(void) | |
7087 | { | |
7088 | WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files)); | |
7089 | } | |
7090 | ||
7091 | static void __init enable_swap_cgroup(void) | |
7092 | { | |
7093 | if (!mem_cgroup_disabled() && really_do_swap_account) { | |
7094 | do_swap_account = 1; | |
7095 | memsw_file_init(); | |
7096 | } | |
7097 | } | |
7098 | ||
7099 | #else | |
7100 | static void __init enable_swap_cgroup(void) | |
7101 | { | |
7102 | } | |
7103 | #endif | |
7104 | ||
7105 | /* | |
7106 | * subsys_initcall() for memory controller. | |
7107 | * | |
7108 | * Some parts like hotcpu_notifier() have to be initialized from this context | |
7109 | * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically | |
7110 | * everything that doesn't depend on a specific mem_cgroup structure should | |
7111 | * be initialized from here. | |
7112 | */ | |
7113 | static int __init mem_cgroup_init(void) | |
7114 | { | |
7115 | hotcpu_notifier(memcg_cpu_hotplug_callback, 0); | |
7116 | enable_swap_cgroup(); | |
7117 | mem_cgroup_soft_limit_tree_init(); | |
7118 | memcg_stock_init(); | |
7119 | return 0; | |
7120 | } | |
7121 | subsys_initcall(mem_cgroup_init); |