]>
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 | * This program is free software; you can redistribute it and/or modify | |
14 | * it under the terms of the GNU General Public License as published by | |
15 | * the Free Software Foundation; either version 2 of the License, or | |
16 | * (at your option) any later version. | |
17 | * | |
18 | * This program is distributed in the hope that it will be useful, | |
19 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
20 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
21 | * GNU General Public License for more details. | |
22 | */ | |
23 | ||
24 | #include <linux/res_counter.h> | |
25 | #include <linux/memcontrol.h> | |
26 | #include <linux/cgroup.h> | |
27 | #include <linux/mm.h> | |
28 | #include <linux/hugetlb.h> | |
29 | #include <linux/pagemap.h> | |
30 | #include <linux/smp.h> | |
31 | #include <linux/page-flags.h> | |
32 | #include <linux/backing-dev.h> | |
33 | #include <linux/bit_spinlock.h> | |
34 | #include <linux/rcupdate.h> | |
35 | #include <linux/limits.h> | |
36 | #include <linux/export.h> | |
37 | #include <linux/mutex.h> | |
38 | #include <linux/rbtree.h> | |
39 | #include <linux/slab.h> | |
40 | #include <linux/swap.h> | |
41 | #include <linux/swapops.h> | |
42 | #include <linux/spinlock.h> | |
43 | #include <linux/eventfd.h> | |
44 | #include <linux/sort.h> | |
45 | #include <linux/fs.h> | |
46 | #include <linux/seq_file.h> | |
47 | #include <linux/vmalloc.h> | |
48 | #include <linux/mm_inline.h> | |
49 | #include <linux/page_cgroup.h> | |
50 | #include <linux/cpu.h> | |
51 | #include <linux/oom.h> | |
52 | #include "internal.h" | |
53 | #include <net/sock.h> | |
54 | #include <net/tcp_memcontrol.h> | |
55 | ||
56 | #include <asm/uaccess.h> | |
57 | ||
58 | #include <trace/events/vmscan.h> | |
59 | ||
60 | struct cgroup_subsys mem_cgroup_subsys __read_mostly; | |
61 | #define MEM_CGROUP_RECLAIM_RETRIES 5 | |
62 | struct mem_cgroup *root_mem_cgroup __read_mostly; | |
63 | ||
64 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
65 | /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ | |
66 | int do_swap_account __read_mostly; | |
67 | ||
68 | /* for remember boot option*/ | |
69 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED | |
70 | static int really_do_swap_account __initdata = 1; | |
71 | #else | |
72 | static int really_do_swap_account __initdata = 0; | |
73 | #endif | |
74 | ||
75 | #else | |
76 | #define do_swap_account (0) | |
77 | #endif | |
78 | ||
79 | ||
80 | /* | |
81 | * Statistics for memory cgroup. | |
82 | */ | |
83 | enum mem_cgroup_stat_index { | |
84 | /* | |
85 | * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. | |
86 | */ | |
87 | MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ | |
88 | MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ | |
89 | MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ | |
90 | MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ | |
91 | MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ | |
92 | MEM_CGROUP_STAT_NSTATS, | |
93 | }; | |
94 | ||
95 | enum mem_cgroup_events_index { | |
96 | MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ | |
97 | MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ | |
98 | MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */ | |
99 | MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */ | |
100 | MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */ | |
101 | MEM_CGROUP_EVENTS_NSTATS, | |
102 | }; | |
103 | /* | |
104 | * Per memcg event counter is incremented at every pagein/pageout. With THP, | |
105 | * it will be incremated by the number of pages. This counter is used for | |
106 | * for trigger some periodic events. This is straightforward and better | |
107 | * than using jiffies etc. to handle periodic memcg event. | |
108 | */ | |
109 | enum mem_cgroup_events_target { | |
110 | MEM_CGROUP_TARGET_THRESH, | |
111 | MEM_CGROUP_TARGET_SOFTLIMIT, | |
112 | MEM_CGROUP_TARGET_NUMAINFO, | |
113 | MEM_CGROUP_NTARGETS, | |
114 | }; | |
115 | #define THRESHOLDS_EVENTS_TARGET (128) | |
116 | #define SOFTLIMIT_EVENTS_TARGET (1024) | |
117 | #define NUMAINFO_EVENTS_TARGET (1024) | |
118 | ||
119 | struct mem_cgroup_stat_cpu { | |
120 | long count[MEM_CGROUP_STAT_NSTATS]; | |
121 | unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; | |
122 | unsigned long targets[MEM_CGROUP_NTARGETS]; | |
123 | }; | |
124 | ||
125 | struct mem_cgroup_reclaim_iter { | |
126 | /* css_id of the last scanned hierarchy member */ | |
127 | int position; | |
128 | /* scan generation, increased every round-trip */ | |
129 | unsigned int generation; | |
130 | }; | |
131 | ||
132 | /* | |
133 | * per-zone information in memory controller. | |
134 | */ | |
135 | struct mem_cgroup_per_zone { | |
136 | struct lruvec lruvec; | |
137 | unsigned long lru_size[NR_LRU_LISTS]; | |
138 | ||
139 | struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1]; | |
140 | ||
141 | struct zone_reclaim_stat reclaim_stat; | |
142 | struct rb_node tree_node; /* RB tree node */ | |
143 | unsigned long long usage_in_excess;/* Set to the value by which */ | |
144 | /* the soft limit is exceeded*/ | |
145 | bool on_tree; | |
146 | struct mem_cgroup *memcg; /* Back pointer, we cannot */ | |
147 | /* use container_of */ | |
148 | }; | |
149 | ||
150 | struct mem_cgroup_per_node { | |
151 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; | |
152 | }; | |
153 | ||
154 | struct mem_cgroup_lru_info { | |
155 | struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; | |
156 | }; | |
157 | ||
158 | /* | |
159 | * Cgroups above their limits are maintained in a RB-Tree, independent of | |
160 | * their hierarchy representation | |
161 | */ | |
162 | ||
163 | struct mem_cgroup_tree_per_zone { | |
164 | struct rb_root rb_root; | |
165 | spinlock_t lock; | |
166 | }; | |
167 | ||
168 | struct mem_cgroup_tree_per_node { | |
169 | struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; | |
170 | }; | |
171 | ||
172 | struct mem_cgroup_tree { | |
173 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; | |
174 | }; | |
175 | ||
176 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; | |
177 | ||
178 | struct mem_cgroup_threshold { | |
179 | struct eventfd_ctx *eventfd; | |
180 | u64 threshold; | |
181 | }; | |
182 | ||
183 | /* For threshold */ | |
184 | struct mem_cgroup_threshold_ary { | |
185 | /* An array index points to threshold just below usage. */ | |
186 | int current_threshold; | |
187 | /* Size of entries[] */ | |
188 | unsigned int size; | |
189 | /* Array of thresholds */ | |
190 | struct mem_cgroup_threshold entries[0]; | |
191 | }; | |
192 | ||
193 | struct mem_cgroup_thresholds { | |
194 | /* Primary thresholds array */ | |
195 | struct mem_cgroup_threshold_ary *primary; | |
196 | /* | |
197 | * Spare threshold array. | |
198 | * This is needed to make mem_cgroup_unregister_event() "never fail". | |
199 | * It must be able to store at least primary->size - 1 entries. | |
200 | */ | |
201 | struct mem_cgroup_threshold_ary *spare; | |
202 | }; | |
203 | ||
204 | /* for OOM */ | |
205 | struct mem_cgroup_eventfd_list { | |
206 | struct list_head list; | |
207 | struct eventfd_ctx *eventfd; | |
208 | }; | |
209 | ||
210 | static void mem_cgroup_threshold(struct mem_cgroup *memcg); | |
211 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg); | |
212 | ||
213 | /* | |
214 | * The memory controller data structure. The memory controller controls both | |
215 | * page cache and RSS per cgroup. We would eventually like to provide | |
216 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, | |
217 | * to help the administrator determine what knobs to tune. | |
218 | * | |
219 | * TODO: Add a water mark for the memory controller. Reclaim will begin when | |
220 | * we hit the water mark. May be even add a low water mark, such that | |
221 | * no reclaim occurs from a cgroup at it's low water mark, this is | |
222 | * a feature that will be implemented much later in the future. | |
223 | */ | |
224 | struct mem_cgroup { | |
225 | struct cgroup_subsys_state css; | |
226 | /* | |
227 | * the counter to account for memory usage | |
228 | */ | |
229 | struct res_counter res; | |
230 | ||
231 | union { | |
232 | /* | |
233 | * the counter to account for mem+swap usage. | |
234 | */ | |
235 | struct res_counter memsw; | |
236 | ||
237 | /* | |
238 | * rcu_freeing is used only when freeing struct mem_cgroup, | |
239 | * so put it into a union to avoid wasting more memory. | |
240 | * It must be disjoint from the css field. It could be | |
241 | * in a union with the res field, but res plays a much | |
242 | * larger part in mem_cgroup life than memsw, and might | |
243 | * be of interest, even at time of free, when debugging. | |
244 | * So share rcu_head with the less interesting memsw. | |
245 | */ | |
246 | struct rcu_head rcu_freeing; | |
247 | /* | |
248 | * But when using vfree(), that cannot be done at | |
249 | * interrupt time, so we must then queue the work. | |
250 | */ | |
251 | struct work_struct work_freeing; | |
252 | }; | |
253 | ||
254 | /* | |
255 | * Per cgroup active and inactive list, similar to the | |
256 | * per zone LRU lists. | |
257 | */ | |
258 | struct mem_cgroup_lru_info info; | |
259 | int last_scanned_node; | |
260 | #if MAX_NUMNODES > 1 | |
261 | nodemask_t scan_nodes; | |
262 | atomic_t numainfo_events; | |
263 | atomic_t numainfo_updating; | |
264 | #endif | |
265 | /* | |
266 | * Should the accounting and control be hierarchical, per subtree? | |
267 | */ | |
268 | bool use_hierarchy; | |
269 | ||
270 | bool oom_lock; | |
271 | atomic_t under_oom; | |
272 | ||
273 | atomic_t refcnt; | |
274 | ||
275 | int swappiness; | |
276 | /* OOM-Killer disable */ | |
277 | int oom_kill_disable; | |
278 | ||
279 | /* set when res.limit == memsw.limit */ | |
280 | bool memsw_is_minimum; | |
281 | ||
282 | /* protect arrays of thresholds */ | |
283 | struct mutex thresholds_lock; | |
284 | ||
285 | /* thresholds for memory usage. RCU-protected */ | |
286 | struct mem_cgroup_thresholds thresholds; | |
287 | ||
288 | /* thresholds for mem+swap usage. RCU-protected */ | |
289 | struct mem_cgroup_thresholds memsw_thresholds; | |
290 | ||
291 | /* For oom notifier event fd */ | |
292 | struct list_head oom_notify; | |
293 | ||
294 | /* | |
295 | * Should we move charges of a task when a task is moved into this | |
296 | * mem_cgroup ? And what type of charges should we move ? | |
297 | */ | |
298 | unsigned long move_charge_at_immigrate; | |
299 | /* | |
300 | * set > 0 if pages under this cgroup are moving to other cgroup. | |
301 | */ | |
302 | atomic_t moving_account; | |
303 | /* taken only while moving_account > 0 */ | |
304 | spinlock_t move_lock; | |
305 | /* | |
306 | * percpu counter. | |
307 | */ | |
308 | struct mem_cgroup_stat_cpu *stat; | |
309 | /* | |
310 | * used when a cpu is offlined or other synchronizations | |
311 | * See mem_cgroup_read_stat(). | |
312 | */ | |
313 | struct mem_cgroup_stat_cpu nocpu_base; | |
314 | spinlock_t pcp_counter_lock; | |
315 | ||
316 | #ifdef CONFIG_INET | |
317 | struct tcp_memcontrol tcp_mem; | |
318 | #endif | |
319 | }; | |
320 | ||
321 | /* Stuffs for move charges at task migration. */ | |
322 | /* | |
323 | * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a | |
324 | * left-shifted bitmap of these types. | |
325 | */ | |
326 | enum move_type { | |
327 | MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ | |
328 | MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ | |
329 | NR_MOVE_TYPE, | |
330 | }; | |
331 | ||
332 | /* "mc" and its members are protected by cgroup_mutex */ | |
333 | static struct move_charge_struct { | |
334 | spinlock_t lock; /* for from, to */ | |
335 | struct mem_cgroup *from; | |
336 | struct mem_cgroup *to; | |
337 | unsigned long precharge; | |
338 | unsigned long moved_charge; | |
339 | unsigned long moved_swap; | |
340 | struct task_struct *moving_task; /* a task moving charges */ | |
341 | wait_queue_head_t waitq; /* a waitq for other context */ | |
342 | } mc = { | |
343 | .lock = __SPIN_LOCK_UNLOCKED(mc.lock), | |
344 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), | |
345 | }; | |
346 | ||
347 | static bool move_anon(void) | |
348 | { | |
349 | return test_bit(MOVE_CHARGE_TYPE_ANON, | |
350 | &mc.to->move_charge_at_immigrate); | |
351 | } | |
352 | ||
353 | static bool move_file(void) | |
354 | { | |
355 | return test_bit(MOVE_CHARGE_TYPE_FILE, | |
356 | &mc.to->move_charge_at_immigrate); | |
357 | } | |
358 | ||
359 | /* | |
360 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft | |
361 | * limit reclaim to prevent infinite loops, if they ever occur. | |
362 | */ | |
363 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) | |
364 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) | |
365 | ||
366 | enum charge_type { | |
367 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, | |
368 | MEM_CGROUP_CHARGE_TYPE_MAPPED, | |
369 | MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ | |
370 | MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ | |
371 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ | |
372 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ | |
373 | NR_CHARGE_TYPE, | |
374 | }; | |
375 | ||
376 | /* for encoding cft->private value on file */ | |
377 | #define _MEM (0) | |
378 | #define _MEMSWAP (1) | |
379 | #define _OOM_TYPE (2) | |
380 | #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) | |
381 | #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) | |
382 | #define MEMFILE_ATTR(val) ((val) & 0xffff) | |
383 | /* Used for OOM nofiier */ | |
384 | #define OOM_CONTROL (0) | |
385 | ||
386 | /* | |
387 | * Reclaim flags for mem_cgroup_hierarchical_reclaim | |
388 | */ | |
389 | #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 | |
390 | #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) | |
391 | #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 | |
392 | #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) | |
393 | ||
394 | static void mem_cgroup_get(struct mem_cgroup *memcg); | |
395 | static void mem_cgroup_put(struct mem_cgroup *memcg); | |
396 | ||
397 | /* Writing them here to avoid exposing memcg's inner layout */ | |
398 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM | |
399 | #include <net/sock.h> | |
400 | #include <net/ip.h> | |
401 | ||
402 | static bool mem_cgroup_is_root(struct mem_cgroup *memcg); | |
403 | void sock_update_memcg(struct sock *sk) | |
404 | { | |
405 | if (mem_cgroup_sockets_enabled) { | |
406 | struct mem_cgroup *memcg; | |
407 | ||
408 | BUG_ON(!sk->sk_prot->proto_cgroup); | |
409 | ||
410 | /* Socket cloning can throw us here with sk_cgrp already | |
411 | * filled. It won't however, necessarily happen from | |
412 | * process context. So the test for root memcg given | |
413 | * the current task's memcg won't help us in this case. | |
414 | * | |
415 | * Respecting the original socket's memcg is a better | |
416 | * decision in this case. | |
417 | */ | |
418 | if (sk->sk_cgrp) { | |
419 | BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg)); | |
420 | mem_cgroup_get(sk->sk_cgrp->memcg); | |
421 | return; | |
422 | } | |
423 | ||
424 | rcu_read_lock(); | |
425 | memcg = mem_cgroup_from_task(current); | |
426 | if (!mem_cgroup_is_root(memcg)) { | |
427 | mem_cgroup_get(memcg); | |
428 | sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg); | |
429 | } | |
430 | rcu_read_unlock(); | |
431 | } | |
432 | } | |
433 | EXPORT_SYMBOL(sock_update_memcg); | |
434 | ||
435 | void sock_release_memcg(struct sock *sk) | |
436 | { | |
437 | if (mem_cgroup_sockets_enabled && sk->sk_cgrp) { | |
438 | struct mem_cgroup *memcg; | |
439 | WARN_ON(!sk->sk_cgrp->memcg); | |
440 | memcg = sk->sk_cgrp->memcg; | |
441 | mem_cgroup_put(memcg); | |
442 | } | |
443 | } | |
444 | ||
445 | #ifdef CONFIG_INET | |
446 | struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg) | |
447 | { | |
448 | if (!memcg || mem_cgroup_is_root(memcg)) | |
449 | return NULL; | |
450 | ||
451 | return &memcg->tcp_mem.cg_proto; | |
452 | } | |
453 | EXPORT_SYMBOL(tcp_proto_cgroup); | |
454 | #endif /* CONFIG_INET */ | |
455 | #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */ | |
456 | ||
457 | static void drain_all_stock_async(struct mem_cgroup *memcg); | |
458 | ||
459 | static struct mem_cgroup_per_zone * | |
460 | mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid) | |
461 | { | |
462 | return &memcg->info.nodeinfo[nid]->zoneinfo[zid]; | |
463 | } | |
464 | ||
465 | struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg) | |
466 | { | |
467 | return &memcg->css; | |
468 | } | |
469 | ||
470 | static struct mem_cgroup_per_zone * | |
471 | page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page) | |
472 | { | |
473 | int nid = page_to_nid(page); | |
474 | int zid = page_zonenum(page); | |
475 | ||
476 | return mem_cgroup_zoneinfo(memcg, nid, zid); | |
477 | } | |
478 | ||
479 | static struct mem_cgroup_tree_per_zone * | |
480 | soft_limit_tree_node_zone(int nid, int zid) | |
481 | { | |
482 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
483 | } | |
484 | ||
485 | static struct mem_cgroup_tree_per_zone * | |
486 | soft_limit_tree_from_page(struct page *page) | |
487 | { | |
488 | int nid = page_to_nid(page); | |
489 | int zid = page_zonenum(page); | |
490 | ||
491 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
492 | } | |
493 | ||
494 | static void | |
495 | __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg, | |
496 | struct mem_cgroup_per_zone *mz, | |
497 | struct mem_cgroup_tree_per_zone *mctz, | |
498 | unsigned long long new_usage_in_excess) | |
499 | { | |
500 | struct rb_node **p = &mctz->rb_root.rb_node; | |
501 | struct rb_node *parent = NULL; | |
502 | struct mem_cgroup_per_zone *mz_node; | |
503 | ||
504 | if (mz->on_tree) | |
505 | return; | |
506 | ||
507 | mz->usage_in_excess = new_usage_in_excess; | |
508 | if (!mz->usage_in_excess) | |
509 | return; | |
510 | while (*p) { | |
511 | parent = *p; | |
512 | mz_node = rb_entry(parent, struct mem_cgroup_per_zone, | |
513 | tree_node); | |
514 | if (mz->usage_in_excess < mz_node->usage_in_excess) | |
515 | p = &(*p)->rb_left; | |
516 | /* | |
517 | * We can't avoid mem cgroups that are over their soft | |
518 | * limit by the same amount | |
519 | */ | |
520 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) | |
521 | p = &(*p)->rb_right; | |
522 | } | |
523 | rb_link_node(&mz->tree_node, parent, p); | |
524 | rb_insert_color(&mz->tree_node, &mctz->rb_root); | |
525 | mz->on_tree = true; | |
526 | } | |
527 | ||
528 | static void | |
529 | __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, | |
530 | struct mem_cgroup_per_zone *mz, | |
531 | struct mem_cgroup_tree_per_zone *mctz) | |
532 | { | |
533 | if (!mz->on_tree) | |
534 | return; | |
535 | rb_erase(&mz->tree_node, &mctz->rb_root); | |
536 | mz->on_tree = false; | |
537 | } | |
538 | ||
539 | static void | |
540 | mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, | |
541 | struct mem_cgroup_per_zone *mz, | |
542 | struct mem_cgroup_tree_per_zone *mctz) | |
543 | { | |
544 | spin_lock(&mctz->lock); | |
545 | __mem_cgroup_remove_exceeded(memcg, mz, mctz); | |
546 | spin_unlock(&mctz->lock); | |
547 | } | |
548 | ||
549 | ||
550 | static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page) | |
551 | { | |
552 | unsigned long long excess; | |
553 | struct mem_cgroup_per_zone *mz; | |
554 | struct mem_cgroup_tree_per_zone *mctz; | |
555 | int nid = page_to_nid(page); | |
556 | int zid = page_zonenum(page); | |
557 | mctz = soft_limit_tree_from_page(page); | |
558 | ||
559 | /* | |
560 | * Necessary to update all ancestors when hierarchy is used. | |
561 | * because their event counter is not touched. | |
562 | */ | |
563 | for (; memcg; memcg = parent_mem_cgroup(memcg)) { | |
564 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
565 | excess = res_counter_soft_limit_excess(&memcg->res); | |
566 | /* | |
567 | * We have to update the tree if mz is on RB-tree or | |
568 | * mem is over its softlimit. | |
569 | */ | |
570 | if (excess || mz->on_tree) { | |
571 | spin_lock(&mctz->lock); | |
572 | /* if on-tree, remove it */ | |
573 | if (mz->on_tree) | |
574 | __mem_cgroup_remove_exceeded(memcg, mz, mctz); | |
575 | /* | |
576 | * Insert again. mz->usage_in_excess will be updated. | |
577 | * If excess is 0, no tree ops. | |
578 | */ | |
579 | __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess); | |
580 | spin_unlock(&mctz->lock); | |
581 | } | |
582 | } | |
583 | } | |
584 | ||
585 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg) | |
586 | { | |
587 | int node, zone; | |
588 | struct mem_cgroup_per_zone *mz; | |
589 | struct mem_cgroup_tree_per_zone *mctz; | |
590 | ||
591 | for_each_node(node) { | |
592 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
593 | mz = mem_cgroup_zoneinfo(memcg, node, zone); | |
594 | mctz = soft_limit_tree_node_zone(node, zone); | |
595 | mem_cgroup_remove_exceeded(memcg, mz, mctz); | |
596 | } | |
597 | } | |
598 | } | |
599 | ||
600 | static struct mem_cgroup_per_zone * | |
601 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
602 | { | |
603 | struct rb_node *rightmost = NULL; | |
604 | struct mem_cgroup_per_zone *mz; | |
605 | ||
606 | retry: | |
607 | mz = NULL; | |
608 | rightmost = rb_last(&mctz->rb_root); | |
609 | if (!rightmost) | |
610 | goto done; /* Nothing to reclaim from */ | |
611 | ||
612 | mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); | |
613 | /* | |
614 | * Remove the node now but someone else can add it back, | |
615 | * we will to add it back at the end of reclaim to its correct | |
616 | * position in the tree. | |
617 | */ | |
618 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); | |
619 | if (!res_counter_soft_limit_excess(&mz->memcg->res) || | |
620 | !css_tryget(&mz->memcg->css)) | |
621 | goto retry; | |
622 | done: | |
623 | return mz; | |
624 | } | |
625 | ||
626 | static struct mem_cgroup_per_zone * | |
627 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
628 | { | |
629 | struct mem_cgroup_per_zone *mz; | |
630 | ||
631 | spin_lock(&mctz->lock); | |
632 | mz = __mem_cgroup_largest_soft_limit_node(mctz); | |
633 | spin_unlock(&mctz->lock); | |
634 | return mz; | |
635 | } | |
636 | ||
637 | /* | |
638 | * Implementation Note: reading percpu statistics for memcg. | |
639 | * | |
640 | * Both of vmstat[] and percpu_counter has threshold and do periodic | |
641 | * synchronization to implement "quick" read. There are trade-off between | |
642 | * reading cost and precision of value. Then, we may have a chance to implement | |
643 | * a periodic synchronizion of counter in memcg's counter. | |
644 | * | |
645 | * But this _read() function is used for user interface now. The user accounts | |
646 | * memory usage by memory cgroup and he _always_ requires exact value because | |
647 | * he accounts memory. Even if we provide quick-and-fuzzy read, we always | |
648 | * have to visit all online cpus and make sum. So, for now, unnecessary | |
649 | * synchronization is not implemented. (just implemented for cpu hotplug) | |
650 | * | |
651 | * If there are kernel internal actions which can make use of some not-exact | |
652 | * value, and reading all cpu value can be performance bottleneck in some | |
653 | * common workload, threashold and synchonization as vmstat[] should be | |
654 | * implemented. | |
655 | */ | |
656 | static long mem_cgroup_read_stat(struct mem_cgroup *memcg, | |
657 | enum mem_cgroup_stat_index idx) | |
658 | { | |
659 | long val = 0; | |
660 | int cpu; | |
661 | ||
662 | get_online_cpus(); | |
663 | for_each_online_cpu(cpu) | |
664 | val += per_cpu(memcg->stat->count[idx], cpu); | |
665 | #ifdef CONFIG_HOTPLUG_CPU | |
666 | spin_lock(&memcg->pcp_counter_lock); | |
667 | val += memcg->nocpu_base.count[idx]; | |
668 | spin_unlock(&memcg->pcp_counter_lock); | |
669 | #endif | |
670 | put_online_cpus(); | |
671 | return val; | |
672 | } | |
673 | ||
674 | static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg, | |
675 | bool charge) | |
676 | { | |
677 | int val = (charge) ? 1 : -1; | |
678 | this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); | |
679 | } | |
680 | ||
681 | static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg, | |
682 | enum mem_cgroup_events_index idx) | |
683 | { | |
684 | unsigned long val = 0; | |
685 | int cpu; | |
686 | ||
687 | for_each_online_cpu(cpu) | |
688 | val += per_cpu(memcg->stat->events[idx], cpu); | |
689 | #ifdef CONFIG_HOTPLUG_CPU | |
690 | spin_lock(&memcg->pcp_counter_lock); | |
691 | val += memcg->nocpu_base.events[idx]; | |
692 | spin_unlock(&memcg->pcp_counter_lock); | |
693 | #endif | |
694 | return val; | |
695 | } | |
696 | ||
697 | static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, | |
698 | bool anon, int nr_pages) | |
699 | { | |
700 | preempt_disable(); | |
701 | ||
702 | /* | |
703 | * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is | |
704 | * counted as CACHE even if it's on ANON LRU. | |
705 | */ | |
706 | if (anon) | |
707 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS], | |
708 | nr_pages); | |
709 | else | |
710 | __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE], | |
711 | nr_pages); | |
712 | ||
713 | /* pagein of a big page is an event. So, ignore page size */ | |
714 | if (nr_pages > 0) | |
715 | __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); | |
716 | else { | |
717 | __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); | |
718 | nr_pages = -nr_pages; /* for event */ | |
719 | } | |
720 | ||
721 | __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); | |
722 | ||
723 | preempt_enable(); | |
724 | } | |
725 | ||
726 | unsigned long | |
727 | mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid, | |
728 | unsigned int lru_mask) | |
729 | { | |
730 | struct mem_cgroup_per_zone *mz; | |
731 | enum lru_list lru; | |
732 | unsigned long ret = 0; | |
733 | ||
734 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
735 | ||
736 | for_each_lru(lru) { | |
737 | if (BIT(lru) & lru_mask) | |
738 | ret += mz->lru_size[lru]; | |
739 | } | |
740 | return ret; | |
741 | } | |
742 | ||
743 | static unsigned long | |
744 | mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, | |
745 | int nid, unsigned int lru_mask) | |
746 | { | |
747 | u64 total = 0; | |
748 | int zid; | |
749 | ||
750 | for (zid = 0; zid < MAX_NR_ZONES; zid++) | |
751 | total += mem_cgroup_zone_nr_lru_pages(memcg, | |
752 | nid, zid, lru_mask); | |
753 | ||
754 | return total; | |
755 | } | |
756 | ||
757 | static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, | |
758 | unsigned int lru_mask) | |
759 | { | |
760 | int nid; | |
761 | u64 total = 0; | |
762 | ||
763 | for_each_node_state(nid, N_HIGH_MEMORY) | |
764 | total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask); | |
765 | return total; | |
766 | } | |
767 | ||
768 | static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, | |
769 | enum mem_cgroup_events_target target) | |
770 | { | |
771 | unsigned long val, next; | |
772 | ||
773 | val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]); | |
774 | next = __this_cpu_read(memcg->stat->targets[target]); | |
775 | /* from time_after() in jiffies.h */ | |
776 | if ((long)next - (long)val < 0) { | |
777 | switch (target) { | |
778 | case MEM_CGROUP_TARGET_THRESH: | |
779 | next = val + THRESHOLDS_EVENTS_TARGET; | |
780 | break; | |
781 | case MEM_CGROUP_TARGET_SOFTLIMIT: | |
782 | next = val + SOFTLIMIT_EVENTS_TARGET; | |
783 | break; | |
784 | case MEM_CGROUP_TARGET_NUMAINFO: | |
785 | next = val + NUMAINFO_EVENTS_TARGET; | |
786 | break; | |
787 | default: | |
788 | break; | |
789 | } | |
790 | __this_cpu_write(memcg->stat->targets[target], next); | |
791 | return true; | |
792 | } | |
793 | return false; | |
794 | } | |
795 | ||
796 | /* | |
797 | * Check events in order. | |
798 | * | |
799 | */ | |
800 | static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) | |
801 | { | |
802 | preempt_disable(); | |
803 | /* threshold event is triggered in finer grain than soft limit */ | |
804 | if (unlikely(mem_cgroup_event_ratelimit(memcg, | |
805 | MEM_CGROUP_TARGET_THRESH))) { | |
806 | bool do_softlimit; | |
807 | bool do_numainfo __maybe_unused; | |
808 | ||
809 | do_softlimit = mem_cgroup_event_ratelimit(memcg, | |
810 | MEM_CGROUP_TARGET_SOFTLIMIT); | |
811 | #if MAX_NUMNODES > 1 | |
812 | do_numainfo = mem_cgroup_event_ratelimit(memcg, | |
813 | MEM_CGROUP_TARGET_NUMAINFO); | |
814 | #endif | |
815 | preempt_enable(); | |
816 | ||
817 | mem_cgroup_threshold(memcg); | |
818 | if (unlikely(do_softlimit)) | |
819 | mem_cgroup_update_tree(memcg, page); | |
820 | #if MAX_NUMNODES > 1 | |
821 | if (unlikely(do_numainfo)) | |
822 | atomic_inc(&memcg->numainfo_events); | |
823 | #endif | |
824 | } else | |
825 | preempt_enable(); | |
826 | } | |
827 | ||
828 | struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) | |
829 | { | |
830 | return container_of(cgroup_subsys_state(cont, | |
831 | mem_cgroup_subsys_id), struct mem_cgroup, | |
832 | css); | |
833 | } | |
834 | ||
835 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) | |
836 | { | |
837 | /* | |
838 | * mm_update_next_owner() may clear mm->owner to NULL | |
839 | * if it races with swapoff, page migration, etc. | |
840 | * So this can be called with p == NULL. | |
841 | */ | |
842 | if (unlikely(!p)) | |
843 | return NULL; | |
844 | ||
845 | return container_of(task_subsys_state(p, mem_cgroup_subsys_id), | |
846 | struct mem_cgroup, css); | |
847 | } | |
848 | ||
849 | struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) | |
850 | { | |
851 | struct mem_cgroup *memcg = NULL; | |
852 | ||
853 | if (!mm) | |
854 | return NULL; | |
855 | /* | |
856 | * Because we have no locks, mm->owner's may be being moved to other | |
857 | * cgroup. We use css_tryget() here even if this looks | |
858 | * pessimistic (rather than adding locks here). | |
859 | */ | |
860 | rcu_read_lock(); | |
861 | do { | |
862 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
863 | if (unlikely(!memcg)) | |
864 | break; | |
865 | } while (!css_tryget(&memcg->css)); | |
866 | rcu_read_unlock(); | |
867 | return memcg; | |
868 | } | |
869 | ||
870 | /** | |
871 | * mem_cgroup_iter - iterate over memory cgroup hierarchy | |
872 | * @root: hierarchy root | |
873 | * @prev: previously returned memcg, NULL on first invocation | |
874 | * @reclaim: cookie for shared reclaim walks, NULL for full walks | |
875 | * | |
876 | * Returns references to children of the hierarchy below @root, or | |
877 | * @root itself, or %NULL after a full round-trip. | |
878 | * | |
879 | * Caller must pass the return value in @prev on subsequent | |
880 | * invocations for reference counting, or use mem_cgroup_iter_break() | |
881 | * to cancel a hierarchy walk before the round-trip is complete. | |
882 | * | |
883 | * Reclaimers can specify a zone and a priority level in @reclaim to | |
884 | * divide up the memcgs in the hierarchy among all concurrent | |
885 | * reclaimers operating on the same zone and priority. | |
886 | */ | |
887 | struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, | |
888 | struct mem_cgroup *prev, | |
889 | struct mem_cgroup_reclaim_cookie *reclaim) | |
890 | { | |
891 | struct mem_cgroup *memcg = NULL; | |
892 | int id = 0; | |
893 | ||
894 | if (mem_cgroup_disabled()) | |
895 | return NULL; | |
896 | ||
897 | if (!root) | |
898 | root = root_mem_cgroup; | |
899 | ||
900 | if (prev && !reclaim) | |
901 | id = css_id(&prev->css); | |
902 | ||
903 | if (prev && prev != root) | |
904 | css_put(&prev->css); | |
905 | ||
906 | if (!root->use_hierarchy && root != root_mem_cgroup) { | |
907 | if (prev) | |
908 | return NULL; | |
909 | return root; | |
910 | } | |
911 | ||
912 | while (!memcg) { | |
913 | struct mem_cgroup_reclaim_iter *uninitialized_var(iter); | |
914 | struct cgroup_subsys_state *css; | |
915 | ||
916 | if (reclaim) { | |
917 | int nid = zone_to_nid(reclaim->zone); | |
918 | int zid = zone_idx(reclaim->zone); | |
919 | struct mem_cgroup_per_zone *mz; | |
920 | ||
921 | mz = mem_cgroup_zoneinfo(root, nid, zid); | |
922 | iter = &mz->reclaim_iter[reclaim->priority]; | |
923 | if (prev && reclaim->generation != iter->generation) | |
924 | return NULL; | |
925 | id = iter->position; | |
926 | } | |
927 | ||
928 | rcu_read_lock(); | |
929 | css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id); | |
930 | if (css) { | |
931 | if (css == &root->css || css_tryget(css)) | |
932 | memcg = container_of(css, | |
933 | struct mem_cgroup, css); | |
934 | } else | |
935 | id = 0; | |
936 | rcu_read_unlock(); | |
937 | ||
938 | if (reclaim) { | |
939 | iter->position = id; | |
940 | if (!css) | |
941 | iter->generation++; | |
942 | else if (!prev && memcg) | |
943 | reclaim->generation = iter->generation; | |
944 | } | |
945 | ||
946 | if (prev && !css) | |
947 | return NULL; | |
948 | } | |
949 | return memcg; | |
950 | } | |
951 | ||
952 | /** | |
953 | * mem_cgroup_iter_break - abort a hierarchy walk prematurely | |
954 | * @root: hierarchy root | |
955 | * @prev: last visited hierarchy member as returned by mem_cgroup_iter() | |
956 | */ | |
957 | void mem_cgroup_iter_break(struct mem_cgroup *root, | |
958 | struct mem_cgroup *prev) | |
959 | { | |
960 | if (!root) | |
961 | root = root_mem_cgroup; | |
962 | if (prev && prev != root) | |
963 | css_put(&prev->css); | |
964 | } | |
965 | ||
966 | /* | |
967 | * Iteration constructs for visiting all cgroups (under a tree). If | |
968 | * loops are exited prematurely (break), mem_cgroup_iter_break() must | |
969 | * be used for reference counting. | |
970 | */ | |
971 | #define for_each_mem_cgroup_tree(iter, root) \ | |
972 | for (iter = mem_cgroup_iter(root, NULL, NULL); \ | |
973 | iter != NULL; \ | |
974 | iter = mem_cgroup_iter(root, iter, NULL)) | |
975 | ||
976 | #define for_each_mem_cgroup(iter) \ | |
977 | for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ | |
978 | iter != NULL; \ | |
979 | iter = mem_cgroup_iter(NULL, iter, NULL)) | |
980 | ||
981 | static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg) | |
982 | { | |
983 | return (memcg == root_mem_cgroup); | |
984 | } | |
985 | ||
986 | void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) | |
987 | { | |
988 | struct mem_cgroup *memcg; | |
989 | ||
990 | if (!mm) | |
991 | return; | |
992 | ||
993 | rcu_read_lock(); | |
994 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
995 | if (unlikely(!memcg)) | |
996 | goto out; | |
997 | ||
998 | switch (idx) { | |
999 | case PGFAULT: | |
1000 | this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]); | |
1001 | break; | |
1002 | case PGMAJFAULT: | |
1003 | this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]); | |
1004 | break; | |
1005 | default: | |
1006 | BUG(); | |
1007 | } | |
1008 | out: | |
1009 | rcu_read_unlock(); | |
1010 | } | |
1011 | EXPORT_SYMBOL(mem_cgroup_count_vm_event); | |
1012 | ||
1013 | /** | |
1014 | * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg | |
1015 | * @zone: zone of the wanted lruvec | |
1016 | * @mem: memcg of the wanted lruvec | |
1017 | * | |
1018 | * Returns the lru list vector holding pages for the given @zone and | |
1019 | * @mem. This can be the global zone lruvec, if the memory controller | |
1020 | * is disabled. | |
1021 | */ | |
1022 | struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone, | |
1023 | struct mem_cgroup *memcg) | |
1024 | { | |
1025 | struct mem_cgroup_per_zone *mz; | |
1026 | ||
1027 | if (mem_cgroup_disabled()) | |
1028 | return &zone->lruvec; | |
1029 | ||
1030 | mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone)); | |
1031 | return &mz->lruvec; | |
1032 | } | |
1033 | ||
1034 | /* | |
1035 | * Following LRU functions are allowed to be used without PCG_LOCK. | |
1036 | * Operations are called by routine of global LRU independently from memcg. | |
1037 | * What we have to take care of here is validness of pc->mem_cgroup. | |
1038 | * | |
1039 | * Changes to pc->mem_cgroup happens when | |
1040 | * 1. charge | |
1041 | * 2. moving account | |
1042 | * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. | |
1043 | * It is added to LRU before charge. | |
1044 | * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. | |
1045 | * When moving account, the page is not on LRU. It's isolated. | |
1046 | */ | |
1047 | ||
1048 | /** | |
1049 | * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec | |
1050 | * @zone: zone of the page | |
1051 | * @page: the page | |
1052 | * @lru: current lru | |
1053 | * | |
1054 | * This function accounts for @page being added to @lru, and returns | |
1055 | * the lruvec for the given @zone and the memcg @page is charged to. | |
1056 | * | |
1057 | * The callsite is then responsible for physically linking the page to | |
1058 | * the returned lruvec->lists[@lru]. | |
1059 | */ | |
1060 | struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page, | |
1061 | enum lru_list lru) | |
1062 | { | |
1063 | struct mem_cgroup_per_zone *mz; | |
1064 | struct mem_cgroup *memcg; | |
1065 | struct page_cgroup *pc; | |
1066 | ||
1067 | if (mem_cgroup_disabled()) | |
1068 | return &zone->lruvec; | |
1069 | ||
1070 | pc = lookup_page_cgroup(page); | |
1071 | memcg = pc->mem_cgroup; | |
1072 | ||
1073 | /* | |
1074 | * Surreptitiously switch any uncharged page to root: | |
1075 | * an uncharged page off lru does nothing to secure | |
1076 | * its former mem_cgroup from sudden removal. | |
1077 | * | |
1078 | * Our caller holds lru_lock, and PageCgroupUsed is updated | |
1079 | * under page_cgroup lock: between them, they make all uses | |
1080 | * of pc->mem_cgroup safe. | |
1081 | */ | |
1082 | if (!PageCgroupUsed(pc) && memcg != root_mem_cgroup) | |
1083 | pc->mem_cgroup = memcg = root_mem_cgroup; | |
1084 | ||
1085 | mz = page_cgroup_zoneinfo(memcg, page); | |
1086 | /* compound_order() is stabilized through lru_lock */ | |
1087 | mz->lru_size[lru] += 1 << compound_order(page); | |
1088 | return &mz->lruvec; | |
1089 | } | |
1090 | ||
1091 | /** | |
1092 | * mem_cgroup_lru_del_list - account for removing an lru page | |
1093 | * @page: the page | |
1094 | * @lru: target lru | |
1095 | * | |
1096 | * This function accounts for @page being removed from @lru. | |
1097 | * | |
1098 | * The callsite is then responsible for physically unlinking | |
1099 | * @page->lru. | |
1100 | */ | |
1101 | void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru) | |
1102 | { | |
1103 | struct mem_cgroup_per_zone *mz; | |
1104 | struct mem_cgroup *memcg; | |
1105 | struct page_cgroup *pc; | |
1106 | ||
1107 | if (mem_cgroup_disabled()) | |
1108 | return; | |
1109 | ||
1110 | pc = lookup_page_cgroup(page); | |
1111 | memcg = pc->mem_cgroup; | |
1112 | VM_BUG_ON(!memcg); | |
1113 | mz = page_cgroup_zoneinfo(memcg, page); | |
1114 | /* huge page split is done under lru_lock. so, we have no races. */ | |
1115 | VM_BUG_ON(mz->lru_size[lru] < (1 << compound_order(page))); | |
1116 | mz->lru_size[lru] -= 1 << compound_order(page); | |
1117 | } | |
1118 | ||
1119 | void mem_cgroup_lru_del(struct page *page) | |
1120 | { | |
1121 | mem_cgroup_lru_del_list(page, page_lru(page)); | |
1122 | } | |
1123 | ||
1124 | /** | |
1125 | * mem_cgroup_lru_move_lists - account for moving a page between lrus | |
1126 | * @zone: zone of the page | |
1127 | * @page: the page | |
1128 | * @from: current lru | |
1129 | * @to: target lru | |
1130 | * | |
1131 | * This function accounts for @page being moved between the lrus @from | |
1132 | * and @to, and returns the lruvec for the given @zone and the memcg | |
1133 | * @page is charged to. | |
1134 | * | |
1135 | * The callsite is then responsible for physically relinking | |
1136 | * @page->lru to the returned lruvec->lists[@to]. | |
1137 | */ | |
1138 | struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone, | |
1139 | struct page *page, | |
1140 | enum lru_list from, | |
1141 | enum lru_list to) | |
1142 | { | |
1143 | /* XXX: Optimize this, especially for @from == @to */ | |
1144 | mem_cgroup_lru_del_list(page, from); | |
1145 | return mem_cgroup_lru_add_list(zone, page, to); | |
1146 | } | |
1147 | ||
1148 | /* | |
1149 | * Checks whether given mem is same or in the root_mem_cgroup's | |
1150 | * hierarchy subtree | |
1151 | */ | |
1152 | bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, | |
1153 | struct mem_cgroup *memcg) | |
1154 | { | |
1155 | if (root_memcg == memcg) | |
1156 | return true; | |
1157 | if (!root_memcg->use_hierarchy) | |
1158 | return false; | |
1159 | return css_is_ancestor(&memcg->css, &root_memcg->css); | |
1160 | } | |
1161 | ||
1162 | static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, | |
1163 | struct mem_cgroup *memcg) | |
1164 | { | |
1165 | bool ret; | |
1166 | ||
1167 | rcu_read_lock(); | |
1168 | ret = __mem_cgroup_same_or_subtree(root_memcg, memcg); | |
1169 | rcu_read_unlock(); | |
1170 | return ret; | |
1171 | } | |
1172 | ||
1173 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) | |
1174 | { | |
1175 | int ret; | |
1176 | struct mem_cgroup *curr = NULL; | |
1177 | struct task_struct *p; | |
1178 | ||
1179 | p = find_lock_task_mm(task); | |
1180 | if (p) { | |
1181 | curr = try_get_mem_cgroup_from_mm(p->mm); | |
1182 | task_unlock(p); | |
1183 | } else { | |
1184 | /* | |
1185 | * All threads may have already detached their mm's, but the oom | |
1186 | * killer still needs to detect if they have already been oom | |
1187 | * killed to prevent needlessly killing additional tasks. | |
1188 | */ | |
1189 | task_lock(task); | |
1190 | curr = mem_cgroup_from_task(task); | |
1191 | if (curr) | |
1192 | css_get(&curr->css); | |
1193 | task_unlock(task); | |
1194 | } | |
1195 | if (!curr) | |
1196 | return 0; | |
1197 | /* | |
1198 | * We should check use_hierarchy of "memcg" not "curr". Because checking | |
1199 | * use_hierarchy of "curr" here make this function true if hierarchy is | |
1200 | * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* | |
1201 | * hierarchy(even if use_hierarchy is disabled in "memcg"). | |
1202 | */ | |
1203 | ret = mem_cgroup_same_or_subtree(memcg, curr); | |
1204 | css_put(&curr->css); | |
1205 | return ret; | |
1206 | } | |
1207 | ||
1208 | int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) | |
1209 | { | |
1210 | unsigned long inactive_ratio; | |
1211 | int nid = zone_to_nid(zone); | |
1212 | int zid = zone_idx(zone); | |
1213 | unsigned long inactive; | |
1214 | unsigned long active; | |
1215 | unsigned long gb; | |
1216 | ||
1217 | inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, | |
1218 | BIT(LRU_INACTIVE_ANON)); | |
1219 | active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, | |
1220 | BIT(LRU_ACTIVE_ANON)); | |
1221 | ||
1222 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
1223 | if (gb) | |
1224 | inactive_ratio = int_sqrt(10 * gb); | |
1225 | else | |
1226 | inactive_ratio = 1; | |
1227 | ||
1228 | return inactive * inactive_ratio < active; | |
1229 | } | |
1230 | ||
1231 | int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) | |
1232 | { | |
1233 | unsigned long active; | |
1234 | unsigned long inactive; | |
1235 | int zid = zone_idx(zone); | |
1236 | int nid = zone_to_nid(zone); | |
1237 | ||
1238 | inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, | |
1239 | BIT(LRU_INACTIVE_FILE)); | |
1240 | active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, | |
1241 | BIT(LRU_ACTIVE_FILE)); | |
1242 | ||
1243 | return (active > inactive); | |
1244 | } | |
1245 | ||
1246 | struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, | |
1247 | struct zone *zone) | |
1248 | { | |
1249 | int nid = zone_to_nid(zone); | |
1250 | int zid = zone_idx(zone); | |
1251 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
1252 | ||
1253 | return &mz->reclaim_stat; | |
1254 | } | |
1255 | ||
1256 | struct zone_reclaim_stat * | |
1257 | mem_cgroup_get_reclaim_stat_from_page(struct page *page) | |
1258 | { | |
1259 | struct page_cgroup *pc; | |
1260 | struct mem_cgroup_per_zone *mz; | |
1261 | ||
1262 | if (mem_cgroup_disabled()) | |
1263 | return NULL; | |
1264 | ||
1265 | pc = lookup_page_cgroup(page); | |
1266 | if (!PageCgroupUsed(pc)) | |
1267 | return NULL; | |
1268 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ | |
1269 | smp_rmb(); | |
1270 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); | |
1271 | return &mz->reclaim_stat; | |
1272 | } | |
1273 | ||
1274 | #define mem_cgroup_from_res_counter(counter, member) \ | |
1275 | container_of(counter, struct mem_cgroup, member) | |
1276 | ||
1277 | /** | |
1278 | * mem_cgroup_margin - calculate chargeable space of a memory cgroup | |
1279 | * @mem: the memory cgroup | |
1280 | * | |
1281 | * Returns the maximum amount of memory @mem can be charged with, in | |
1282 | * pages. | |
1283 | */ | |
1284 | static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) | |
1285 | { | |
1286 | unsigned long long margin; | |
1287 | ||
1288 | margin = res_counter_margin(&memcg->res); | |
1289 | if (do_swap_account) | |
1290 | margin = min(margin, res_counter_margin(&memcg->memsw)); | |
1291 | return margin >> PAGE_SHIFT; | |
1292 | } | |
1293 | ||
1294 | int mem_cgroup_swappiness(struct mem_cgroup *memcg) | |
1295 | { | |
1296 | struct cgroup *cgrp = memcg->css.cgroup; | |
1297 | ||
1298 | /* root ? */ | |
1299 | if (cgrp->parent == NULL) | |
1300 | return vm_swappiness; | |
1301 | ||
1302 | return memcg->swappiness; | |
1303 | } | |
1304 | ||
1305 | /* | |
1306 | * memcg->moving_account is used for checking possibility that some thread is | |
1307 | * calling move_account(). When a thread on CPU-A starts moving pages under | |
1308 | * a memcg, other threads should check memcg->moving_account under | |
1309 | * rcu_read_lock(), like this: | |
1310 | * | |
1311 | * CPU-A CPU-B | |
1312 | * rcu_read_lock() | |
1313 | * memcg->moving_account+1 if (memcg->mocing_account) | |
1314 | * take heavy locks. | |
1315 | * synchronize_rcu() update something. | |
1316 | * rcu_read_unlock() | |
1317 | * start move here. | |
1318 | */ | |
1319 | ||
1320 | /* for quick checking without looking up memcg */ | |
1321 | atomic_t memcg_moving __read_mostly; | |
1322 | ||
1323 | static void mem_cgroup_start_move(struct mem_cgroup *memcg) | |
1324 | { | |
1325 | atomic_inc(&memcg_moving); | |
1326 | atomic_inc(&memcg->moving_account); | |
1327 | synchronize_rcu(); | |
1328 | } | |
1329 | ||
1330 | static void mem_cgroup_end_move(struct mem_cgroup *memcg) | |
1331 | { | |
1332 | /* | |
1333 | * Now, mem_cgroup_clear_mc() may call this function with NULL. | |
1334 | * We check NULL in callee rather than caller. | |
1335 | */ | |
1336 | if (memcg) { | |
1337 | atomic_dec(&memcg_moving); | |
1338 | atomic_dec(&memcg->moving_account); | |
1339 | } | |
1340 | } | |
1341 | ||
1342 | /* | |
1343 | * 2 routines for checking "mem" is under move_account() or not. | |
1344 | * | |
1345 | * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This | |
1346 | * is used for avoiding races in accounting. If true, | |
1347 | * pc->mem_cgroup may be overwritten. | |
1348 | * | |
1349 | * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or | |
1350 | * under hierarchy of moving cgroups. This is for | |
1351 | * waiting at hith-memory prressure caused by "move". | |
1352 | */ | |
1353 | ||
1354 | static bool mem_cgroup_stolen(struct mem_cgroup *memcg) | |
1355 | { | |
1356 | VM_BUG_ON(!rcu_read_lock_held()); | |
1357 | return atomic_read(&memcg->moving_account) > 0; | |
1358 | } | |
1359 | ||
1360 | static bool mem_cgroup_under_move(struct mem_cgroup *memcg) | |
1361 | { | |
1362 | struct mem_cgroup *from; | |
1363 | struct mem_cgroup *to; | |
1364 | bool ret = false; | |
1365 | /* | |
1366 | * Unlike task_move routines, we access mc.to, mc.from not under | |
1367 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. | |
1368 | */ | |
1369 | spin_lock(&mc.lock); | |
1370 | from = mc.from; | |
1371 | to = mc.to; | |
1372 | if (!from) | |
1373 | goto unlock; | |
1374 | ||
1375 | ret = mem_cgroup_same_or_subtree(memcg, from) | |
1376 | || mem_cgroup_same_or_subtree(memcg, to); | |
1377 | unlock: | |
1378 | spin_unlock(&mc.lock); | |
1379 | return ret; | |
1380 | } | |
1381 | ||
1382 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) | |
1383 | { | |
1384 | if (mc.moving_task && current != mc.moving_task) { | |
1385 | if (mem_cgroup_under_move(memcg)) { | |
1386 | DEFINE_WAIT(wait); | |
1387 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); | |
1388 | /* moving charge context might have finished. */ | |
1389 | if (mc.moving_task) | |
1390 | schedule(); | |
1391 | finish_wait(&mc.waitq, &wait); | |
1392 | return true; | |
1393 | } | |
1394 | } | |
1395 | return false; | |
1396 | } | |
1397 | ||
1398 | /* | |
1399 | * Take this lock when | |
1400 | * - a code tries to modify page's memcg while it's USED. | |
1401 | * - a code tries to modify page state accounting in a memcg. | |
1402 | * see mem_cgroup_stolen(), too. | |
1403 | */ | |
1404 | static void move_lock_mem_cgroup(struct mem_cgroup *memcg, | |
1405 | unsigned long *flags) | |
1406 | { | |
1407 | spin_lock_irqsave(&memcg->move_lock, *flags); | |
1408 | } | |
1409 | ||
1410 | static void move_unlock_mem_cgroup(struct mem_cgroup *memcg, | |
1411 | unsigned long *flags) | |
1412 | { | |
1413 | spin_unlock_irqrestore(&memcg->move_lock, *flags); | |
1414 | } | |
1415 | ||
1416 | /** | |
1417 | * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. | |
1418 | * @memcg: The memory cgroup that went over limit | |
1419 | * @p: Task that is going to be killed | |
1420 | * | |
1421 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is | |
1422 | * enabled | |
1423 | */ | |
1424 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) | |
1425 | { | |
1426 | struct cgroup *task_cgrp; | |
1427 | struct cgroup *mem_cgrp; | |
1428 | /* | |
1429 | * Need a buffer in BSS, can't rely on allocations. The code relies | |
1430 | * on the assumption that OOM is serialized for memory controller. | |
1431 | * If this assumption is broken, revisit this code. | |
1432 | */ | |
1433 | static char memcg_name[PATH_MAX]; | |
1434 | int ret; | |
1435 | ||
1436 | if (!memcg || !p) | |
1437 | return; | |
1438 | ||
1439 | rcu_read_lock(); | |
1440 | ||
1441 | mem_cgrp = memcg->css.cgroup; | |
1442 | task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); | |
1443 | ||
1444 | ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); | |
1445 | if (ret < 0) { | |
1446 | /* | |
1447 | * Unfortunately, we are unable to convert to a useful name | |
1448 | * But we'll still print out the usage information | |
1449 | */ | |
1450 | rcu_read_unlock(); | |
1451 | goto done; | |
1452 | } | |
1453 | rcu_read_unlock(); | |
1454 | ||
1455 | printk(KERN_INFO "Task in %s killed", memcg_name); | |
1456 | ||
1457 | rcu_read_lock(); | |
1458 | ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); | |
1459 | if (ret < 0) { | |
1460 | rcu_read_unlock(); | |
1461 | goto done; | |
1462 | } | |
1463 | rcu_read_unlock(); | |
1464 | ||
1465 | /* | |
1466 | * Continues from above, so we don't need an KERN_ level | |
1467 | */ | |
1468 | printk(KERN_CONT " as a result of limit of %s\n", memcg_name); | |
1469 | done: | |
1470 | ||
1471 | printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", | |
1472 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, | |
1473 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, | |
1474 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); | |
1475 | printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " | |
1476 | "failcnt %llu\n", | |
1477 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, | |
1478 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, | |
1479 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); | |
1480 | } | |
1481 | ||
1482 | /* | |
1483 | * This function returns the number of memcg under hierarchy tree. Returns | |
1484 | * 1(self count) if no children. | |
1485 | */ | |
1486 | static int mem_cgroup_count_children(struct mem_cgroup *memcg) | |
1487 | { | |
1488 | int num = 0; | |
1489 | struct mem_cgroup *iter; | |
1490 | ||
1491 | for_each_mem_cgroup_tree(iter, memcg) | |
1492 | num++; | |
1493 | return num; | |
1494 | } | |
1495 | ||
1496 | /* | |
1497 | * Return the memory (and swap, if configured) limit for a memcg. | |
1498 | */ | |
1499 | u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) | |
1500 | { | |
1501 | u64 limit; | |
1502 | u64 memsw; | |
1503 | ||
1504 | limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
1505 | limit += total_swap_pages << PAGE_SHIFT; | |
1506 | ||
1507 | memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
1508 | /* | |
1509 | * If memsw is finite and limits the amount of swap space available | |
1510 | * to this memcg, return that limit. | |
1511 | */ | |
1512 | return min(limit, memsw); | |
1513 | } | |
1514 | ||
1515 | static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, | |
1516 | gfp_t gfp_mask, | |
1517 | unsigned long flags) | |
1518 | { | |
1519 | unsigned long total = 0; | |
1520 | bool noswap = false; | |
1521 | int loop; | |
1522 | ||
1523 | if (flags & MEM_CGROUP_RECLAIM_NOSWAP) | |
1524 | noswap = true; | |
1525 | if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) | |
1526 | noswap = true; | |
1527 | ||
1528 | for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { | |
1529 | if (loop) | |
1530 | drain_all_stock_async(memcg); | |
1531 | total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); | |
1532 | /* | |
1533 | * Allow limit shrinkers, which are triggered directly | |
1534 | * by userspace, to catch signals and stop reclaim | |
1535 | * after minimal progress, regardless of the margin. | |
1536 | */ | |
1537 | if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) | |
1538 | break; | |
1539 | if (mem_cgroup_margin(memcg)) | |
1540 | break; | |
1541 | /* | |
1542 | * If nothing was reclaimed after two attempts, there | |
1543 | * may be no reclaimable pages in this hierarchy. | |
1544 | */ | |
1545 | if (loop && !total) | |
1546 | break; | |
1547 | } | |
1548 | return total; | |
1549 | } | |
1550 | ||
1551 | /** | |
1552 | * test_mem_cgroup_node_reclaimable | |
1553 | * @mem: the target memcg | |
1554 | * @nid: the node ID to be checked. | |
1555 | * @noswap : specify true here if the user wants flle only information. | |
1556 | * | |
1557 | * This function returns whether the specified memcg contains any | |
1558 | * reclaimable pages on a node. Returns true if there are any reclaimable | |
1559 | * pages in the node. | |
1560 | */ | |
1561 | static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, | |
1562 | int nid, bool noswap) | |
1563 | { | |
1564 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) | |
1565 | return true; | |
1566 | if (noswap || !total_swap_pages) | |
1567 | return false; | |
1568 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) | |
1569 | return true; | |
1570 | return false; | |
1571 | ||
1572 | } | |
1573 | #if MAX_NUMNODES > 1 | |
1574 | ||
1575 | /* | |
1576 | * Always updating the nodemask is not very good - even if we have an empty | |
1577 | * list or the wrong list here, we can start from some node and traverse all | |
1578 | * nodes based on the zonelist. So update the list loosely once per 10 secs. | |
1579 | * | |
1580 | */ | |
1581 | static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) | |
1582 | { | |
1583 | int nid; | |
1584 | /* | |
1585 | * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET | |
1586 | * pagein/pageout changes since the last update. | |
1587 | */ | |
1588 | if (!atomic_read(&memcg->numainfo_events)) | |
1589 | return; | |
1590 | if (atomic_inc_return(&memcg->numainfo_updating) > 1) | |
1591 | return; | |
1592 | ||
1593 | /* make a nodemask where this memcg uses memory from */ | |
1594 | memcg->scan_nodes = node_states[N_HIGH_MEMORY]; | |
1595 | ||
1596 | for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { | |
1597 | ||
1598 | if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) | |
1599 | node_clear(nid, memcg->scan_nodes); | |
1600 | } | |
1601 | ||
1602 | atomic_set(&memcg->numainfo_events, 0); | |
1603 | atomic_set(&memcg->numainfo_updating, 0); | |
1604 | } | |
1605 | ||
1606 | /* | |
1607 | * Selecting a node where we start reclaim from. Because what we need is just | |
1608 | * reducing usage counter, start from anywhere is O,K. Considering | |
1609 | * memory reclaim from current node, there are pros. and cons. | |
1610 | * | |
1611 | * Freeing memory from current node means freeing memory from a node which | |
1612 | * we'll use or we've used. So, it may make LRU bad. And if several threads | |
1613 | * hit limits, it will see a contention on a node. But freeing from remote | |
1614 | * node means more costs for memory reclaim because of memory latency. | |
1615 | * | |
1616 | * Now, we use round-robin. Better algorithm is welcomed. | |
1617 | */ | |
1618 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) | |
1619 | { | |
1620 | int node; | |
1621 | ||
1622 | mem_cgroup_may_update_nodemask(memcg); | |
1623 | node = memcg->last_scanned_node; | |
1624 | ||
1625 | node = next_node(node, memcg->scan_nodes); | |
1626 | if (node == MAX_NUMNODES) | |
1627 | node = first_node(memcg->scan_nodes); | |
1628 | /* | |
1629 | * We call this when we hit limit, not when pages are added to LRU. | |
1630 | * No LRU may hold pages because all pages are UNEVICTABLE or | |
1631 | * memcg is too small and all pages are not on LRU. In that case, | |
1632 | * we use curret node. | |
1633 | */ | |
1634 | if (unlikely(node == MAX_NUMNODES)) | |
1635 | node = numa_node_id(); | |
1636 | ||
1637 | memcg->last_scanned_node = node; | |
1638 | return node; | |
1639 | } | |
1640 | ||
1641 | /* | |
1642 | * Check all nodes whether it contains reclaimable pages or not. | |
1643 | * For quick scan, we make use of scan_nodes. This will allow us to skip | |
1644 | * unused nodes. But scan_nodes is lazily updated and may not cotain | |
1645 | * enough new information. We need to do double check. | |
1646 | */ | |
1647 | bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) | |
1648 | { | |
1649 | int nid; | |
1650 | ||
1651 | /* | |
1652 | * quick check...making use of scan_node. | |
1653 | * We can skip unused nodes. | |
1654 | */ | |
1655 | if (!nodes_empty(memcg->scan_nodes)) { | |
1656 | for (nid = first_node(memcg->scan_nodes); | |
1657 | nid < MAX_NUMNODES; | |
1658 | nid = next_node(nid, memcg->scan_nodes)) { | |
1659 | ||
1660 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) | |
1661 | return true; | |
1662 | } | |
1663 | } | |
1664 | /* | |
1665 | * Check rest of nodes. | |
1666 | */ | |
1667 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
1668 | if (node_isset(nid, memcg->scan_nodes)) | |
1669 | continue; | |
1670 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) | |
1671 | return true; | |
1672 | } | |
1673 | return false; | |
1674 | } | |
1675 | ||
1676 | #else | |
1677 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) | |
1678 | { | |
1679 | return 0; | |
1680 | } | |
1681 | ||
1682 | bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) | |
1683 | { | |
1684 | return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); | |
1685 | } | |
1686 | #endif | |
1687 | ||
1688 | static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, | |
1689 | struct zone *zone, | |
1690 | gfp_t gfp_mask, | |
1691 | unsigned long *total_scanned) | |
1692 | { | |
1693 | struct mem_cgroup *victim = NULL; | |
1694 | int total = 0; | |
1695 | int loop = 0; | |
1696 | unsigned long excess; | |
1697 | unsigned long nr_scanned; | |
1698 | struct mem_cgroup_reclaim_cookie reclaim = { | |
1699 | .zone = zone, | |
1700 | .priority = 0, | |
1701 | }; | |
1702 | ||
1703 | excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; | |
1704 | ||
1705 | while (1) { | |
1706 | victim = mem_cgroup_iter(root_memcg, victim, &reclaim); | |
1707 | if (!victim) { | |
1708 | loop++; | |
1709 | if (loop >= 2) { | |
1710 | /* | |
1711 | * If we have not been able to reclaim | |
1712 | * anything, it might because there are | |
1713 | * no reclaimable pages under this hierarchy | |
1714 | */ | |
1715 | if (!total) | |
1716 | break; | |
1717 | /* | |
1718 | * We want to do more targeted reclaim. | |
1719 | * excess >> 2 is not to excessive so as to | |
1720 | * reclaim too much, nor too less that we keep | |
1721 | * coming back to reclaim from this cgroup | |
1722 | */ | |
1723 | if (total >= (excess >> 2) || | |
1724 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) | |
1725 | break; | |
1726 | } | |
1727 | continue; | |
1728 | } | |
1729 | if (!mem_cgroup_reclaimable(victim, false)) | |
1730 | continue; | |
1731 | total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, | |
1732 | zone, &nr_scanned); | |
1733 | *total_scanned += nr_scanned; | |
1734 | if (!res_counter_soft_limit_excess(&root_memcg->res)) | |
1735 | break; | |
1736 | } | |
1737 | mem_cgroup_iter_break(root_memcg, victim); | |
1738 | return total; | |
1739 | } | |
1740 | ||
1741 | /* | |
1742 | * Check OOM-Killer is already running under our hierarchy. | |
1743 | * If someone is running, return false. | |
1744 | * Has to be called with memcg_oom_lock | |
1745 | */ | |
1746 | static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) | |
1747 | { | |
1748 | struct mem_cgroup *iter, *failed = NULL; | |
1749 | ||
1750 | for_each_mem_cgroup_tree(iter, memcg) { | |
1751 | if (iter->oom_lock) { | |
1752 | /* | |
1753 | * this subtree of our hierarchy is already locked | |
1754 | * so we cannot give a lock. | |
1755 | */ | |
1756 | failed = iter; | |
1757 | mem_cgroup_iter_break(memcg, iter); | |
1758 | break; | |
1759 | } else | |
1760 | iter->oom_lock = true; | |
1761 | } | |
1762 | ||
1763 | if (!failed) | |
1764 | return true; | |
1765 | ||
1766 | /* | |
1767 | * OK, we failed to lock the whole subtree so we have to clean up | |
1768 | * what we set up to the failing subtree | |
1769 | */ | |
1770 | for_each_mem_cgroup_tree(iter, memcg) { | |
1771 | if (iter == failed) { | |
1772 | mem_cgroup_iter_break(memcg, iter); | |
1773 | break; | |
1774 | } | |
1775 | iter->oom_lock = false; | |
1776 | } | |
1777 | return false; | |
1778 | } | |
1779 | ||
1780 | /* | |
1781 | * Has to be called with memcg_oom_lock | |
1782 | */ | |
1783 | static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) | |
1784 | { | |
1785 | struct mem_cgroup *iter; | |
1786 | ||
1787 | for_each_mem_cgroup_tree(iter, memcg) | |
1788 | iter->oom_lock = false; | |
1789 | return 0; | |
1790 | } | |
1791 | ||
1792 | static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) | |
1793 | { | |
1794 | struct mem_cgroup *iter; | |
1795 | ||
1796 | for_each_mem_cgroup_tree(iter, memcg) | |
1797 | atomic_inc(&iter->under_oom); | |
1798 | } | |
1799 | ||
1800 | static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) | |
1801 | { | |
1802 | struct mem_cgroup *iter; | |
1803 | ||
1804 | /* | |
1805 | * When a new child is created while the hierarchy is under oom, | |
1806 | * mem_cgroup_oom_lock() may not be called. We have to use | |
1807 | * atomic_add_unless() here. | |
1808 | */ | |
1809 | for_each_mem_cgroup_tree(iter, memcg) | |
1810 | atomic_add_unless(&iter->under_oom, -1, 0); | |
1811 | } | |
1812 | ||
1813 | static DEFINE_SPINLOCK(memcg_oom_lock); | |
1814 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); | |
1815 | ||
1816 | struct oom_wait_info { | |
1817 | struct mem_cgroup *memcg; | |
1818 | wait_queue_t wait; | |
1819 | }; | |
1820 | ||
1821 | static int memcg_oom_wake_function(wait_queue_t *wait, | |
1822 | unsigned mode, int sync, void *arg) | |
1823 | { | |
1824 | struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg; | |
1825 | struct mem_cgroup *oom_wait_memcg; | |
1826 | struct oom_wait_info *oom_wait_info; | |
1827 | ||
1828 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); | |
1829 | oom_wait_memcg = oom_wait_info->memcg; | |
1830 | ||
1831 | /* | |
1832 | * Both of oom_wait_info->memcg and wake_memcg are stable under us. | |
1833 | * Then we can use css_is_ancestor without taking care of RCU. | |
1834 | */ | |
1835 | if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) | |
1836 | && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) | |
1837 | return 0; | |
1838 | return autoremove_wake_function(wait, mode, sync, arg); | |
1839 | } | |
1840 | ||
1841 | static void memcg_wakeup_oom(struct mem_cgroup *memcg) | |
1842 | { | |
1843 | /* for filtering, pass "memcg" as argument. */ | |
1844 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); | |
1845 | } | |
1846 | ||
1847 | static void memcg_oom_recover(struct mem_cgroup *memcg) | |
1848 | { | |
1849 | if (memcg && atomic_read(&memcg->under_oom)) | |
1850 | memcg_wakeup_oom(memcg); | |
1851 | } | |
1852 | ||
1853 | /* | |
1854 | * try to call OOM killer. returns false if we should exit memory-reclaim loop. | |
1855 | */ | |
1856 | bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order) | |
1857 | { | |
1858 | struct oom_wait_info owait; | |
1859 | bool locked, need_to_kill; | |
1860 | ||
1861 | owait.memcg = memcg; | |
1862 | owait.wait.flags = 0; | |
1863 | owait.wait.func = memcg_oom_wake_function; | |
1864 | owait.wait.private = current; | |
1865 | INIT_LIST_HEAD(&owait.wait.task_list); | |
1866 | need_to_kill = true; | |
1867 | mem_cgroup_mark_under_oom(memcg); | |
1868 | ||
1869 | /* At first, try to OOM lock hierarchy under memcg.*/ | |
1870 | spin_lock(&memcg_oom_lock); | |
1871 | locked = mem_cgroup_oom_lock(memcg); | |
1872 | /* | |
1873 | * Even if signal_pending(), we can't quit charge() loop without | |
1874 | * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL | |
1875 | * under OOM is always welcomed, use TASK_KILLABLE here. | |
1876 | */ | |
1877 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); | |
1878 | if (!locked || memcg->oom_kill_disable) | |
1879 | need_to_kill = false; | |
1880 | if (locked) | |
1881 | mem_cgroup_oom_notify(memcg); | |
1882 | spin_unlock(&memcg_oom_lock); | |
1883 | ||
1884 | if (need_to_kill) { | |
1885 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1886 | mem_cgroup_out_of_memory(memcg, mask, order); | |
1887 | } else { | |
1888 | schedule(); | |
1889 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1890 | } | |
1891 | spin_lock(&memcg_oom_lock); | |
1892 | if (locked) | |
1893 | mem_cgroup_oom_unlock(memcg); | |
1894 | memcg_wakeup_oom(memcg); | |
1895 | spin_unlock(&memcg_oom_lock); | |
1896 | ||
1897 | mem_cgroup_unmark_under_oom(memcg); | |
1898 | ||
1899 | if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) | |
1900 | return false; | |
1901 | /* Give chance to dying process */ | |
1902 | schedule_timeout_uninterruptible(1); | |
1903 | return true; | |
1904 | } | |
1905 | ||
1906 | /* | |
1907 | * Currently used to update mapped file statistics, but the routine can be | |
1908 | * generalized to update other statistics as well. | |
1909 | * | |
1910 | * Notes: Race condition | |
1911 | * | |
1912 | * We usually use page_cgroup_lock() for accessing page_cgroup member but | |
1913 | * it tends to be costly. But considering some conditions, we doesn't need | |
1914 | * to do so _always_. | |
1915 | * | |
1916 | * Considering "charge", lock_page_cgroup() is not required because all | |
1917 | * file-stat operations happen after a page is attached to radix-tree. There | |
1918 | * are no race with "charge". | |
1919 | * | |
1920 | * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup | |
1921 | * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even | |
1922 | * if there are race with "uncharge". Statistics itself is properly handled | |
1923 | * by flags. | |
1924 | * | |
1925 | * Considering "move", this is an only case we see a race. To make the race | |
1926 | * small, we check mm->moving_account and detect there are possibility of race | |
1927 | * If there is, we take a lock. | |
1928 | */ | |
1929 | ||
1930 | void __mem_cgroup_begin_update_page_stat(struct page *page, | |
1931 | bool *locked, unsigned long *flags) | |
1932 | { | |
1933 | struct mem_cgroup *memcg; | |
1934 | struct page_cgroup *pc; | |
1935 | ||
1936 | pc = lookup_page_cgroup(page); | |
1937 | again: | |
1938 | memcg = pc->mem_cgroup; | |
1939 | if (unlikely(!memcg || !PageCgroupUsed(pc))) | |
1940 | return; | |
1941 | /* | |
1942 | * If this memory cgroup is not under account moving, we don't | |
1943 | * need to take move_lock_page_cgroup(). Because we already hold | |
1944 | * rcu_read_lock(), any calls to move_account will be delayed until | |
1945 | * rcu_read_unlock() if mem_cgroup_stolen() == true. | |
1946 | */ | |
1947 | if (!mem_cgroup_stolen(memcg)) | |
1948 | return; | |
1949 | ||
1950 | move_lock_mem_cgroup(memcg, flags); | |
1951 | if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) { | |
1952 | move_unlock_mem_cgroup(memcg, flags); | |
1953 | goto again; | |
1954 | } | |
1955 | *locked = true; | |
1956 | } | |
1957 | ||
1958 | void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags) | |
1959 | { | |
1960 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
1961 | ||
1962 | /* | |
1963 | * It's guaranteed that pc->mem_cgroup never changes while | |
1964 | * lock is held because a routine modifies pc->mem_cgroup | |
1965 | * should take move_lock_page_cgroup(). | |
1966 | */ | |
1967 | move_unlock_mem_cgroup(pc->mem_cgroup, flags); | |
1968 | } | |
1969 | ||
1970 | void mem_cgroup_update_page_stat(struct page *page, | |
1971 | enum mem_cgroup_page_stat_item idx, int val) | |
1972 | { | |
1973 | struct mem_cgroup *memcg; | |
1974 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
1975 | unsigned long uninitialized_var(flags); | |
1976 | ||
1977 | if (mem_cgroup_disabled()) | |
1978 | return; | |
1979 | ||
1980 | memcg = pc->mem_cgroup; | |
1981 | if (unlikely(!memcg || !PageCgroupUsed(pc))) | |
1982 | return; | |
1983 | ||
1984 | switch (idx) { | |
1985 | case MEMCG_NR_FILE_MAPPED: | |
1986 | idx = MEM_CGROUP_STAT_FILE_MAPPED; | |
1987 | break; | |
1988 | default: | |
1989 | BUG(); | |
1990 | } | |
1991 | ||
1992 | this_cpu_add(memcg->stat->count[idx], val); | |
1993 | } | |
1994 | ||
1995 | /* | |
1996 | * size of first charge trial. "32" comes from vmscan.c's magic value. | |
1997 | * TODO: maybe necessary to use big numbers in big irons. | |
1998 | */ | |
1999 | #define CHARGE_BATCH 32U | |
2000 | struct memcg_stock_pcp { | |
2001 | struct mem_cgroup *cached; /* this never be root cgroup */ | |
2002 | unsigned int nr_pages; | |
2003 | struct work_struct work; | |
2004 | unsigned long flags; | |
2005 | #define FLUSHING_CACHED_CHARGE (0) | |
2006 | }; | |
2007 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); | |
2008 | static DEFINE_MUTEX(percpu_charge_mutex); | |
2009 | ||
2010 | /* | |
2011 | * Try to consume stocked charge on this cpu. If success, one page is consumed | |
2012 | * from local stock and true is returned. If the stock is 0 or charges from a | |
2013 | * cgroup which is not current target, returns false. This stock will be | |
2014 | * refilled. | |
2015 | */ | |
2016 | static bool consume_stock(struct mem_cgroup *memcg) | |
2017 | { | |
2018 | struct memcg_stock_pcp *stock; | |
2019 | bool ret = true; | |
2020 | ||
2021 | stock = &get_cpu_var(memcg_stock); | |
2022 | if (memcg == stock->cached && stock->nr_pages) | |
2023 | stock->nr_pages--; | |
2024 | else /* need to call res_counter_charge */ | |
2025 | ret = false; | |
2026 | put_cpu_var(memcg_stock); | |
2027 | return ret; | |
2028 | } | |
2029 | ||
2030 | /* | |
2031 | * Returns stocks cached in percpu to res_counter and reset cached information. | |
2032 | */ | |
2033 | static void drain_stock(struct memcg_stock_pcp *stock) | |
2034 | { | |
2035 | struct mem_cgroup *old = stock->cached; | |
2036 | ||
2037 | if (stock->nr_pages) { | |
2038 | unsigned long bytes = stock->nr_pages * PAGE_SIZE; | |
2039 | ||
2040 | res_counter_uncharge(&old->res, bytes); | |
2041 | if (do_swap_account) | |
2042 | res_counter_uncharge(&old->memsw, bytes); | |
2043 | stock->nr_pages = 0; | |
2044 | } | |
2045 | stock->cached = NULL; | |
2046 | } | |
2047 | ||
2048 | /* | |
2049 | * This must be called under preempt disabled or must be called by | |
2050 | * a thread which is pinned to local cpu. | |
2051 | */ | |
2052 | static void drain_local_stock(struct work_struct *dummy) | |
2053 | { | |
2054 | struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); | |
2055 | drain_stock(stock); | |
2056 | clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); | |
2057 | } | |
2058 | ||
2059 | /* | |
2060 | * Cache charges(val) which is from res_counter, to local per_cpu area. | |
2061 | * This will be consumed by consume_stock() function, later. | |
2062 | */ | |
2063 | static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) | |
2064 | { | |
2065 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); | |
2066 | ||
2067 | if (stock->cached != memcg) { /* reset if necessary */ | |
2068 | drain_stock(stock); | |
2069 | stock->cached = memcg; | |
2070 | } | |
2071 | stock->nr_pages += nr_pages; | |
2072 | put_cpu_var(memcg_stock); | |
2073 | } | |
2074 | ||
2075 | /* | |
2076 | * Drains all per-CPU charge caches for given root_memcg resp. subtree | |
2077 | * of the hierarchy under it. sync flag says whether we should block | |
2078 | * until the work is done. | |
2079 | */ | |
2080 | static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) | |
2081 | { | |
2082 | int cpu, curcpu; | |
2083 | ||
2084 | /* Notify other cpus that system-wide "drain" is running */ | |
2085 | get_online_cpus(); | |
2086 | curcpu = get_cpu(); | |
2087 | for_each_online_cpu(cpu) { | |
2088 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
2089 | struct mem_cgroup *memcg; | |
2090 | ||
2091 | memcg = stock->cached; | |
2092 | if (!memcg || !stock->nr_pages) | |
2093 | continue; | |
2094 | if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) | |
2095 | continue; | |
2096 | if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { | |
2097 | if (cpu == curcpu) | |
2098 | drain_local_stock(&stock->work); | |
2099 | else | |
2100 | schedule_work_on(cpu, &stock->work); | |
2101 | } | |
2102 | } | |
2103 | put_cpu(); | |
2104 | ||
2105 | if (!sync) | |
2106 | goto out; | |
2107 | ||
2108 | for_each_online_cpu(cpu) { | |
2109 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
2110 | if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) | |
2111 | flush_work(&stock->work); | |
2112 | } | |
2113 | out: | |
2114 | put_online_cpus(); | |
2115 | } | |
2116 | ||
2117 | /* | |
2118 | * Tries to drain stocked charges in other cpus. This function is asynchronous | |
2119 | * and just put a work per cpu for draining localy on each cpu. Caller can | |
2120 | * expects some charges will be back to res_counter later but cannot wait for | |
2121 | * it. | |
2122 | */ | |
2123 | static void drain_all_stock_async(struct mem_cgroup *root_memcg) | |
2124 | { | |
2125 | /* | |
2126 | * If someone calls draining, avoid adding more kworker runs. | |
2127 | */ | |
2128 | if (!mutex_trylock(&percpu_charge_mutex)) | |
2129 | return; | |
2130 | drain_all_stock(root_memcg, false); | |
2131 | mutex_unlock(&percpu_charge_mutex); | |
2132 | } | |
2133 | ||
2134 | /* This is a synchronous drain interface. */ | |
2135 | static void drain_all_stock_sync(struct mem_cgroup *root_memcg) | |
2136 | { | |
2137 | /* called when force_empty is called */ | |
2138 | mutex_lock(&percpu_charge_mutex); | |
2139 | drain_all_stock(root_memcg, true); | |
2140 | mutex_unlock(&percpu_charge_mutex); | |
2141 | } | |
2142 | ||
2143 | /* | |
2144 | * This function drains percpu counter value from DEAD cpu and | |
2145 | * move it to local cpu. Note that this function can be preempted. | |
2146 | */ | |
2147 | static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) | |
2148 | { | |
2149 | int i; | |
2150 | ||
2151 | spin_lock(&memcg->pcp_counter_lock); | |
2152 | for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { | |
2153 | long x = per_cpu(memcg->stat->count[i], cpu); | |
2154 | ||
2155 | per_cpu(memcg->stat->count[i], cpu) = 0; | |
2156 | memcg->nocpu_base.count[i] += x; | |
2157 | } | |
2158 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { | |
2159 | unsigned long x = per_cpu(memcg->stat->events[i], cpu); | |
2160 | ||
2161 | per_cpu(memcg->stat->events[i], cpu) = 0; | |
2162 | memcg->nocpu_base.events[i] += x; | |
2163 | } | |
2164 | spin_unlock(&memcg->pcp_counter_lock); | |
2165 | } | |
2166 | ||
2167 | static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, | |
2168 | unsigned long action, | |
2169 | void *hcpu) | |
2170 | { | |
2171 | int cpu = (unsigned long)hcpu; | |
2172 | struct memcg_stock_pcp *stock; | |
2173 | struct mem_cgroup *iter; | |
2174 | ||
2175 | if (action == CPU_ONLINE) | |
2176 | return NOTIFY_OK; | |
2177 | ||
2178 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) | |
2179 | return NOTIFY_OK; | |
2180 | ||
2181 | for_each_mem_cgroup(iter) | |
2182 | mem_cgroup_drain_pcp_counter(iter, cpu); | |
2183 | ||
2184 | stock = &per_cpu(memcg_stock, cpu); | |
2185 | drain_stock(stock); | |
2186 | return NOTIFY_OK; | |
2187 | } | |
2188 | ||
2189 | ||
2190 | /* See __mem_cgroup_try_charge() for details */ | |
2191 | enum { | |
2192 | CHARGE_OK, /* success */ | |
2193 | CHARGE_RETRY, /* need to retry but retry is not bad */ | |
2194 | CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ | |
2195 | CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ | |
2196 | CHARGE_OOM_DIE, /* the current is killed because of OOM */ | |
2197 | }; | |
2198 | ||
2199 | static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, | |
2200 | unsigned int nr_pages, bool oom_check) | |
2201 | { | |
2202 | unsigned long csize = nr_pages * PAGE_SIZE; | |
2203 | struct mem_cgroup *mem_over_limit; | |
2204 | struct res_counter *fail_res; | |
2205 | unsigned long flags = 0; | |
2206 | int ret; | |
2207 | ||
2208 | ret = res_counter_charge(&memcg->res, csize, &fail_res); | |
2209 | ||
2210 | if (likely(!ret)) { | |
2211 | if (!do_swap_account) | |
2212 | return CHARGE_OK; | |
2213 | ret = res_counter_charge(&memcg->memsw, csize, &fail_res); | |
2214 | if (likely(!ret)) | |
2215 | return CHARGE_OK; | |
2216 | ||
2217 | res_counter_uncharge(&memcg->res, csize); | |
2218 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); | |
2219 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; | |
2220 | } else | |
2221 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); | |
2222 | /* | |
2223 | * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch | |
2224 | * of regular pages (CHARGE_BATCH), or a single regular page (1). | |
2225 | * | |
2226 | * Never reclaim on behalf of optional batching, retry with a | |
2227 | * single page instead. | |
2228 | */ | |
2229 | if (nr_pages == CHARGE_BATCH) | |
2230 | return CHARGE_RETRY; | |
2231 | ||
2232 | if (!(gfp_mask & __GFP_WAIT)) | |
2233 | return CHARGE_WOULDBLOCK; | |
2234 | ||
2235 | ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); | |
2236 | if (mem_cgroup_margin(mem_over_limit) >= nr_pages) | |
2237 | return CHARGE_RETRY; | |
2238 | /* | |
2239 | * Even though the limit is exceeded at this point, reclaim | |
2240 | * may have been able to free some pages. Retry the charge | |
2241 | * before killing the task. | |
2242 | * | |
2243 | * Only for regular pages, though: huge pages are rather | |
2244 | * unlikely to succeed so close to the limit, and we fall back | |
2245 | * to regular pages anyway in case of failure. | |
2246 | */ | |
2247 | if (nr_pages == 1 && ret) | |
2248 | return CHARGE_RETRY; | |
2249 | ||
2250 | /* | |
2251 | * At task move, charge accounts can be doubly counted. So, it's | |
2252 | * better to wait until the end of task_move if something is going on. | |
2253 | */ | |
2254 | if (mem_cgroup_wait_acct_move(mem_over_limit)) | |
2255 | return CHARGE_RETRY; | |
2256 | ||
2257 | /* If we don't need to call oom-killer at el, return immediately */ | |
2258 | if (!oom_check) | |
2259 | return CHARGE_NOMEM; | |
2260 | /* check OOM */ | |
2261 | if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize))) | |
2262 | return CHARGE_OOM_DIE; | |
2263 | ||
2264 | return CHARGE_RETRY; | |
2265 | } | |
2266 | ||
2267 | /* | |
2268 | * __mem_cgroup_try_charge() does | |
2269 | * 1. detect memcg to be charged against from passed *mm and *ptr, | |
2270 | * 2. update res_counter | |
2271 | * 3. call memory reclaim if necessary. | |
2272 | * | |
2273 | * In some special case, if the task is fatal, fatal_signal_pending() or | |
2274 | * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup | |
2275 | * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon | |
2276 | * as possible without any hazards. 2: all pages should have a valid | |
2277 | * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg | |
2278 | * pointer, that is treated as a charge to root_mem_cgroup. | |
2279 | * | |
2280 | * So __mem_cgroup_try_charge() will return | |
2281 | * 0 ... on success, filling *ptr with a valid memcg pointer. | |
2282 | * -ENOMEM ... charge failure because of resource limits. | |
2283 | * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup. | |
2284 | * | |
2285 | * Unlike the exported interface, an "oom" parameter is added. if oom==true, | |
2286 | * the oom-killer can be invoked. | |
2287 | */ | |
2288 | static int __mem_cgroup_try_charge(struct mm_struct *mm, | |
2289 | gfp_t gfp_mask, | |
2290 | unsigned int nr_pages, | |
2291 | struct mem_cgroup **ptr, | |
2292 | bool oom) | |
2293 | { | |
2294 | unsigned int batch = max(CHARGE_BATCH, nr_pages); | |
2295 | int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
2296 | struct mem_cgroup *memcg = NULL; | |
2297 | int ret; | |
2298 | ||
2299 | /* | |
2300 | * Unlike gloval-vm's OOM-kill, we're not in memory shortage | |
2301 | * in system level. So, allow to go ahead dying process in addition to | |
2302 | * MEMDIE process. | |
2303 | */ | |
2304 | if (unlikely(test_thread_flag(TIF_MEMDIE) | |
2305 | || fatal_signal_pending(current))) | |
2306 | goto bypass; | |
2307 | ||
2308 | /* | |
2309 | * We always charge the cgroup the mm_struct belongs to. | |
2310 | * The mm_struct's mem_cgroup changes on task migration if the | |
2311 | * thread group leader migrates. It's possible that mm is not | |
2312 | * set, if so charge the init_mm (happens for pagecache usage). | |
2313 | */ | |
2314 | if (!*ptr && !mm) | |
2315 | *ptr = root_mem_cgroup; | |
2316 | again: | |
2317 | if (*ptr) { /* css should be a valid one */ | |
2318 | memcg = *ptr; | |
2319 | VM_BUG_ON(css_is_removed(&memcg->css)); | |
2320 | if (mem_cgroup_is_root(memcg)) | |
2321 | goto done; | |
2322 | if (nr_pages == 1 && consume_stock(memcg)) | |
2323 | goto done; | |
2324 | css_get(&memcg->css); | |
2325 | } else { | |
2326 | struct task_struct *p; | |
2327 | ||
2328 | rcu_read_lock(); | |
2329 | p = rcu_dereference(mm->owner); | |
2330 | /* | |
2331 | * Because we don't have task_lock(), "p" can exit. | |
2332 | * In that case, "memcg" can point to root or p can be NULL with | |
2333 | * race with swapoff. Then, we have small risk of mis-accouning. | |
2334 | * But such kind of mis-account by race always happens because | |
2335 | * we don't have cgroup_mutex(). It's overkill and we allo that | |
2336 | * small race, here. | |
2337 | * (*) swapoff at el will charge against mm-struct not against | |
2338 | * task-struct. So, mm->owner can be NULL. | |
2339 | */ | |
2340 | memcg = mem_cgroup_from_task(p); | |
2341 | if (!memcg) | |
2342 | memcg = root_mem_cgroup; | |
2343 | if (mem_cgroup_is_root(memcg)) { | |
2344 | rcu_read_unlock(); | |
2345 | goto done; | |
2346 | } | |
2347 | if (nr_pages == 1 && consume_stock(memcg)) { | |
2348 | /* | |
2349 | * It seems dagerous to access memcg without css_get(). | |
2350 | * But considering how consume_stok works, it's not | |
2351 | * necessary. If consume_stock success, some charges | |
2352 | * from this memcg are cached on this cpu. So, we | |
2353 | * don't need to call css_get()/css_tryget() before | |
2354 | * calling consume_stock(). | |
2355 | */ | |
2356 | rcu_read_unlock(); | |
2357 | goto done; | |
2358 | } | |
2359 | /* after here, we may be blocked. we need to get refcnt */ | |
2360 | if (!css_tryget(&memcg->css)) { | |
2361 | rcu_read_unlock(); | |
2362 | goto again; | |
2363 | } | |
2364 | rcu_read_unlock(); | |
2365 | } | |
2366 | ||
2367 | do { | |
2368 | bool oom_check; | |
2369 | ||
2370 | /* If killed, bypass charge */ | |
2371 | if (fatal_signal_pending(current)) { | |
2372 | css_put(&memcg->css); | |
2373 | goto bypass; | |
2374 | } | |
2375 | ||
2376 | oom_check = false; | |
2377 | if (oom && !nr_oom_retries) { | |
2378 | oom_check = true; | |
2379 | nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
2380 | } | |
2381 | ||
2382 | ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); | |
2383 | switch (ret) { | |
2384 | case CHARGE_OK: | |
2385 | break; | |
2386 | case CHARGE_RETRY: /* not in OOM situation but retry */ | |
2387 | batch = nr_pages; | |
2388 | css_put(&memcg->css); | |
2389 | memcg = NULL; | |
2390 | goto again; | |
2391 | case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ | |
2392 | css_put(&memcg->css); | |
2393 | goto nomem; | |
2394 | case CHARGE_NOMEM: /* OOM routine works */ | |
2395 | if (!oom) { | |
2396 | css_put(&memcg->css); | |
2397 | goto nomem; | |
2398 | } | |
2399 | /* If oom, we never return -ENOMEM */ | |
2400 | nr_oom_retries--; | |
2401 | break; | |
2402 | case CHARGE_OOM_DIE: /* Killed by OOM Killer */ | |
2403 | css_put(&memcg->css); | |
2404 | goto bypass; | |
2405 | } | |
2406 | } while (ret != CHARGE_OK); | |
2407 | ||
2408 | if (batch > nr_pages) | |
2409 | refill_stock(memcg, batch - nr_pages); | |
2410 | css_put(&memcg->css); | |
2411 | done: | |
2412 | *ptr = memcg; | |
2413 | return 0; | |
2414 | nomem: | |
2415 | *ptr = NULL; | |
2416 | return -ENOMEM; | |
2417 | bypass: | |
2418 | *ptr = root_mem_cgroup; | |
2419 | return -EINTR; | |
2420 | } | |
2421 | ||
2422 | /* | |
2423 | * Somemtimes we have to undo a charge we got by try_charge(). | |
2424 | * This function is for that and do uncharge, put css's refcnt. | |
2425 | * gotten by try_charge(). | |
2426 | */ | |
2427 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, | |
2428 | unsigned int nr_pages) | |
2429 | { | |
2430 | if (!mem_cgroup_is_root(memcg)) { | |
2431 | unsigned long bytes = nr_pages * PAGE_SIZE; | |
2432 | ||
2433 | res_counter_uncharge(&memcg->res, bytes); | |
2434 | if (do_swap_account) | |
2435 | res_counter_uncharge(&memcg->memsw, bytes); | |
2436 | } | |
2437 | } | |
2438 | ||
2439 | /* | |
2440 | * A helper function to get mem_cgroup from ID. must be called under | |
2441 | * rcu_read_lock(). The caller must check css_is_removed() or some if | |
2442 | * it's concern. (dropping refcnt from swap can be called against removed | |
2443 | * memcg.) | |
2444 | */ | |
2445 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) | |
2446 | { | |
2447 | struct cgroup_subsys_state *css; | |
2448 | ||
2449 | /* ID 0 is unused ID */ | |
2450 | if (!id) | |
2451 | return NULL; | |
2452 | css = css_lookup(&mem_cgroup_subsys, id); | |
2453 | if (!css) | |
2454 | return NULL; | |
2455 | return container_of(css, struct mem_cgroup, css); | |
2456 | } | |
2457 | ||
2458 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) | |
2459 | { | |
2460 | struct mem_cgroup *memcg = NULL; | |
2461 | struct page_cgroup *pc; | |
2462 | unsigned short id; | |
2463 | swp_entry_t ent; | |
2464 | ||
2465 | VM_BUG_ON(!PageLocked(page)); | |
2466 | ||
2467 | pc = lookup_page_cgroup(page); | |
2468 | lock_page_cgroup(pc); | |
2469 | if (PageCgroupUsed(pc)) { | |
2470 | memcg = pc->mem_cgroup; | |
2471 | if (memcg && !css_tryget(&memcg->css)) | |
2472 | memcg = NULL; | |
2473 | } else if (PageSwapCache(page)) { | |
2474 | ent.val = page_private(page); | |
2475 | id = lookup_swap_cgroup_id(ent); | |
2476 | rcu_read_lock(); | |
2477 | memcg = mem_cgroup_lookup(id); | |
2478 | if (memcg && !css_tryget(&memcg->css)) | |
2479 | memcg = NULL; | |
2480 | rcu_read_unlock(); | |
2481 | } | |
2482 | unlock_page_cgroup(pc); | |
2483 | return memcg; | |
2484 | } | |
2485 | ||
2486 | static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, | |
2487 | struct page *page, | |
2488 | unsigned int nr_pages, | |
2489 | enum charge_type ctype, | |
2490 | bool lrucare) | |
2491 | { | |
2492 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
2493 | struct zone *uninitialized_var(zone); | |
2494 | bool was_on_lru = false; | |
2495 | bool anon; | |
2496 | ||
2497 | lock_page_cgroup(pc); | |
2498 | if (unlikely(PageCgroupUsed(pc))) { | |
2499 | unlock_page_cgroup(pc); | |
2500 | __mem_cgroup_cancel_charge(memcg, nr_pages); | |
2501 | return; | |
2502 | } | |
2503 | /* | |
2504 | * we don't need page_cgroup_lock about tail pages, becase they are not | |
2505 | * accessed by any other context at this point. | |
2506 | */ | |
2507 | ||
2508 | /* | |
2509 | * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page | |
2510 | * may already be on some other mem_cgroup's LRU. Take care of it. | |
2511 | */ | |
2512 | if (lrucare) { | |
2513 | zone = page_zone(page); | |
2514 | spin_lock_irq(&zone->lru_lock); | |
2515 | if (PageLRU(page)) { | |
2516 | ClearPageLRU(page); | |
2517 | del_page_from_lru_list(zone, page, page_lru(page)); | |
2518 | was_on_lru = true; | |
2519 | } | |
2520 | } | |
2521 | ||
2522 | pc->mem_cgroup = memcg; | |
2523 | /* | |
2524 | * We access a page_cgroup asynchronously without lock_page_cgroup(). | |
2525 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup | |
2526 | * is accessed after testing USED bit. To make pc->mem_cgroup visible | |
2527 | * before USED bit, we need memory barrier here. | |
2528 | * See mem_cgroup_add_lru_list(), etc. | |
2529 | */ | |
2530 | smp_wmb(); | |
2531 | SetPageCgroupUsed(pc); | |
2532 | ||
2533 | if (lrucare) { | |
2534 | if (was_on_lru) { | |
2535 | VM_BUG_ON(PageLRU(page)); | |
2536 | SetPageLRU(page); | |
2537 | add_page_to_lru_list(zone, page, page_lru(page)); | |
2538 | } | |
2539 | spin_unlock_irq(&zone->lru_lock); | |
2540 | } | |
2541 | ||
2542 | if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) | |
2543 | anon = true; | |
2544 | else | |
2545 | anon = false; | |
2546 | ||
2547 | mem_cgroup_charge_statistics(memcg, anon, nr_pages); | |
2548 | unlock_page_cgroup(pc); | |
2549 | ||
2550 | /* | |
2551 | * "charge_statistics" updated event counter. Then, check it. | |
2552 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. | |
2553 | * if they exceeds softlimit. | |
2554 | */ | |
2555 | memcg_check_events(memcg, page); | |
2556 | } | |
2557 | ||
2558 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
2559 | ||
2560 | #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION)) | |
2561 | /* | |
2562 | * Because tail pages are not marked as "used", set it. We're under | |
2563 | * zone->lru_lock, 'splitting on pmd' and compound_lock. | |
2564 | * charge/uncharge will be never happen and move_account() is done under | |
2565 | * compound_lock(), so we don't have to take care of races. | |
2566 | */ | |
2567 | void mem_cgroup_split_huge_fixup(struct page *head) | |
2568 | { | |
2569 | struct page_cgroup *head_pc = lookup_page_cgroup(head); | |
2570 | struct page_cgroup *pc; | |
2571 | int i; | |
2572 | ||
2573 | if (mem_cgroup_disabled()) | |
2574 | return; | |
2575 | for (i = 1; i < HPAGE_PMD_NR; i++) { | |
2576 | pc = head_pc + i; | |
2577 | pc->mem_cgroup = head_pc->mem_cgroup; | |
2578 | smp_wmb();/* see __commit_charge() */ | |
2579 | pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; | |
2580 | } | |
2581 | } | |
2582 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
2583 | ||
2584 | /** | |
2585 | * mem_cgroup_move_account - move account of the page | |
2586 | * @page: the page | |
2587 | * @nr_pages: number of regular pages (>1 for huge pages) | |
2588 | * @pc: page_cgroup of the page. | |
2589 | * @from: mem_cgroup which the page is moved from. | |
2590 | * @to: mem_cgroup which the page is moved to. @from != @to. | |
2591 | * @uncharge: whether we should call uncharge and css_put against @from. | |
2592 | * | |
2593 | * The caller must confirm following. | |
2594 | * - page is not on LRU (isolate_page() is useful.) | |
2595 | * - compound_lock is held when nr_pages > 1 | |
2596 | * | |
2597 | * This function doesn't do "charge" nor css_get to new cgroup. It should be | |
2598 | * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is | |
2599 | * true, this function does "uncharge" from old cgroup, but it doesn't if | |
2600 | * @uncharge is false, so a caller should do "uncharge". | |
2601 | */ | |
2602 | static int mem_cgroup_move_account(struct page *page, | |
2603 | unsigned int nr_pages, | |
2604 | struct page_cgroup *pc, | |
2605 | struct mem_cgroup *from, | |
2606 | struct mem_cgroup *to, | |
2607 | bool uncharge) | |
2608 | { | |
2609 | unsigned long flags; | |
2610 | int ret; | |
2611 | bool anon = PageAnon(page); | |
2612 | ||
2613 | VM_BUG_ON(from == to); | |
2614 | VM_BUG_ON(PageLRU(page)); | |
2615 | /* | |
2616 | * The page is isolated from LRU. So, collapse function | |
2617 | * will not handle this page. But page splitting can happen. | |
2618 | * Do this check under compound_page_lock(). The caller should | |
2619 | * hold it. | |
2620 | */ | |
2621 | ret = -EBUSY; | |
2622 | if (nr_pages > 1 && !PageTransHuge(page)) | |
2623 | goto out; | |
2624 | ||
2625 | lock_page_cgroup(pc); | |
2626 | ||
2627 | ret = -EINVAL; | |
2628 | if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) | |
2629 | goto unlock; | |
2630 | ||
2631 | move_lock_mem_cgroup(from, &flags); | |
2632 | ||
2633 | if (!anon && page_mapped(page)) { | |
2634 | /* Update mapped_file data for mem_cgroup */ | |
2635 | preempt_disable(); | |
2636 | __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
2637 | __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
2638 | preempt_enable(); | |
2639 | } | |
2640 | mem_cgroup_charge_statistics(from, anon, -nr_pages); | |
2641 | if (uncharge) | |
2642 | /* This is not "cancel", but cancel_charge does all we need. */ | |
2643 | __mem_cgroup_cancel_charge(from, nr_pages); | |
2644 | ||
2645 | /* caller should have done css_get */ | |
2646 | pc->mem_cgroup = to; | |
2647 | mem_cgroup_charge_statistics(to, anon, nr_pages); | |
2648 | /* | |
2649 | * We charges against "to" which may not have any tasks. Then, "to" | |
2650 | * can be under rmdir(). But in current implementation, caller of | |
2651 | * this function is just force_empty() and move charge, so it's | |
2652 | * guaranteed that "to" is never removed. So, we don't check rmdir | |
2653 | * status here. | |
2654 | */ | |
2655 | move_unlock_mem_cgroup(from, &flags); | |
2656 | ret = 0; | |
2657 | unlock: | |
2658 | unlock_page_cgroup(pc); | |
2659 | /* | |
2660 | * check events | |
2661 | */ | |
2662 | memcg_check_events(to, page); | |
2663 | memcg_check_events(from, page); | |
2664 | out: | |
2665 | return ret; | |
2666 | } | |
2667 | ||
2668 | /* | |
2669 | * move charges to its parent. | |
2670 | */ | |
2671 | ||
2672 | static int mem_cgroup_move_parent(struct page *page, | |
2673 | struct page_cgroup *pc, | |
2674 | struct mem_cgroup *child, | |
2675 | gfp_t gfp_mask) | |
2676 | { | |
2677 | struct cgroup *cg = child->css.cgroup; | |
2678 | struct cgroup *pcg = cg->parent; | |
2679 | struct mem_cgroup *parent; | |
2680 | unsigned int nr_pages; | |
2681 | unsigned long uninitialized_var(flags); | |
2682 | int ret; | |
2683 | ||
2684 | /* Is ROOT ? */ | |
2685 | if (!pcg) | |
2686 | return -EINVAL; | |
2687 | ||
2688 | ret = -EBUSY; | |
2689 | if (!get_page_unless_zero(page)) | |
2690 | goto out; | |
2691 | if (isolate_lru_page(page)) | |
2692 | goto put; | |
2693 | ||
2694 | nr_pages = hpage_nr_pages(page); | |
2695 | ||
2696 | parent = mem_cgroup_from_cont(pcg); | |
2697 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); | |
2698 | if (ret) | |
2699 | goto put_back; | |
2700 | ||
2701 | if (nr_pages > 1) | |
2702 | flags = compound_lock_irqsave(page); | |
2703 | ||
2704 | ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); | |
2705 | if (ret) | |
2706 | __mem_cgroup_cancel_charge(parent, nr_pages); | |
2707 | ||
2708 | if (nr_pages > 1) | |
2709 | compound_unlock_irqrestore(page, flags); | |
2710 | put_back: | |
2711 | putback_lru_page(page); | |
2712 | put: | |
2713 | put_page(page); | |
2714 | out: | |
2715 | return ret; | |
2716 | } | |
2717 | ||
2718 | /* | |
2719 | * Charge the memory controller for page usage. | |
2720 | * Return | |
2721 | * 0 if the charge was successful | |
2722 | * < 0 if the cgroup is over its limit | |
2723 | */ | |
2724 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, | |
2725 | gfp_t gfp_mask, enum charge_type ctype) | |
2726 | { | |
2727 | struct mem_cgroup *memcg = NULL; | |
2728 | unsigned int nr_pages = 1; | |
2729 | bool oom = true; | |
2730 | int ret; | |
2731 | ||
2732 | if (PageTransHuge(page)) { | |
2733 | nr_pages <<= compound_order(page); | |
2734 | VM_BUG_ON(!PageTransHuge(page)); | |
2735 | /* | |
2736 | * Never OOM-kill a process for a huge page. The | |
2737 | * fault handler will fall back to regular pages. | |
2738 | */ | |
2739 | oom = false; | |
2740 | } | |
2741 | ||
2742 | ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); | |
2743 | if (ret == -ENOMEM) | |
2744 | return ret; | |
2745 | __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false); | |
2746 | return 0; | |
2747 | } | |
2748 | ||
2749 | int mem_cgroup_newpage_charge(struct page *page, | |
2750 | struct mm_struct *mm, gfp_t gfp_mask) | |
2751 | { | |
2752 | if (mem_cgroup_disabled()) | |
2753 | return 0; | |
2754 | VM_BUG_ON(page_mapped(page)); | |
2755 | VM_BUG_ON(page->mapping && !PageAnon(page)); | |
2756 | VM_BUG_ON(!mm); | |
2757 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
2758 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2759 | } | |
2760 | ||
2761 | static void | |
2762 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, | |
2763 | enum charge_type ctype); | |
2764 | ||
2765 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, | |
2766 | gfp_t gfp_mask) | |
2767 | { | |
2768 | struct mem_cgroup *memcg = NULL; | |
2769 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
2770 | int ret; | |
2771 | ||
2772 | if (mem_cgroup_disabled()) | |
2773 | return 0; | |
2774 | if (PageCompound(page)) | |
2775 | return 0; | |
2776 | ||
2777 | if (unlikely(!mm)) | |
2778 | mm = &init_mm; | |
2779 | if (!page_is_file_cache(page)) | |
2780 | type = MEM_CGROUP_CHARGE_TYPE_SHMEM; | |
2781 | ||
2782 | if (!PageSwapCache(page)) | |
2783 | ret = mem_cgroup_charge_common(page, mm, gfp_mask, type); | |
2784 | else { /* page is swapcache/shmem */ | |
2785 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); | |
2786 | if (!ret) | |
2787 | __mem_cgroup_commit_charge_swapin(page, memcg, type); | |
2788 | } | |
2789 | return ret; | |
2790 | } | |
2791 | ||
2792 | /* | |
2793 | * While swap-in, try_charge -> commit or cancel, the page is locked. | |
2794 | * And when try_charge() successfully returns, one refcnt to memcg without | |
2795 | * struct page_cgroup is acquired. This refcnt will be consumed by | |
2796 | * "commit()" or removed by "cancel()" | |
2797 | */ | |
2798 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, | |
2799 | struct page *page, | |
2800 | gfp_t mask, struct mem_cgroup **memcgp) | |
2801 | { | |
2802 | struct mem_cgroup *memcg; | |
2803 | int ret; | |
2804 | ||
2805 | *memcgp = NULL; | |
2806 | ||
2807 | if (mem_cgroup_disabled()) | |
2808 | return 0; | |
2809 | ||
2810 | if (!do_swap_account) | |
2811 | goto charge_cur_mm; | |
2812 | /* | |
2813 | * A racing thread's fault, or swapoff, may have already updated | |
2814 | * the pte, and even removed page from swap cache: in those cases | |
2815 | * do_swap_page()'s pte_same() test will fail; but there's also a | |
2816 | * KSM case which does need to charge the page. | |
2817 | */ | |
2818 | if (!PageSwapCache(page)) | |
2819 | goto charge_cur_mm; | |
2820 | memcg = try_get_mem_cgroup_from_page(page); | |
2821 | if (!memcg) | |
2822 | goto charge_cur_mm; | |
2823 | *memcgp = memcg; | |
2824 | ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true); | |
2825 | css_put(&memcg->css); | |
2826 | if (ret == -EINTR) | |
2827 | ret = 0; | |
2828 | return ret; | |
2829 | charge_cur_mm: | |
2830 | if (unlikely(!mm)) | |
2831 | mm = &init_mm; | |
2832 | ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true); | |
2833 | if (ret == -EINTR) | |
2834 | ret = 0; | |
2835 | return ret; | |
2836 | } | |
2837 | ||
2838 | static void | |
2839 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, | |
2840 | enum charge_type ctype) | |
2841 | { | |
2842 | if (mem_cgroup_disabled()) | |
2843 | return; | |
2844 | if (!memcg) | |
2845 | return; | |
2846 | cgroup_exclude_rmdir(&memcg->css); | |
2847 | ||
2848 | __mem_cgroup_commit_charge(memcg, page, 1, ctype, true); | |
2849 | /* | |
2850 | * Now swap is on-memory. This means this page may be | |
2851 | * counted both as mem and swap....double count. | |
2852 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable | |
2853 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() | |
2854 | * may call delete_from_swap_cache() before reach here. | |
2855 | */ | |
2856 | if (do_swap_account && PageSwapCache(page)) { | |
2857 | swp_entry_t ent = {.val = page_private(page)}; | |
2858 | struct mem_cgroup *swap_memcg; | |
2859 | unsigned short id; | |
2860 | ||
2861 | id = swap_cgroup_record(ent, 0); | |
2862 | rcu_read_lock(); | |
2863 | swap_memcg = mem_cgroup_lookup(id); | |
2864 | if (swap_memcg) { | |
2865 | /* | |
2866 | * This recorded memcg can be obsolete one. So, avoid | |
2867 | * calling css_tryget | |
2868 | */ | |
2869 | if (!mem_cgroup_is_root(swap_memcg)) | |
2870 | res_counter_uncharge(&swap_memcg->memsw, | |
2871 | PAGE_SIZE); | |
2872 | mem_cgroup_swap_statistics(swap_memcg, false); | |
2873 | mem_cgroup_put(swap_memcg); | |
2874 | } | |
2875 | rcu_read_unlock(); | |
2876 | } | |
2877 | /* | |
2878 | * At swapin, we may charge account against cgroup which has no tasks. | |
2879 | * So, rmdir()->pre_destroy() can be called while we do this charge. | |
2880 | * In that case, we need to call pre_destroy() again. check it here. | |
2881 | */ | |
2882 | cgroup_release_and_wakeup_rmdir(&memcg->css); | |
2883 | } | |
2884 | ||
2885 | void mem_cgroup_commit_charge_swapin(struct page *page, | |
2886 | struct mem_cgroup *memcg) | |
2887 | { | |
2888 | __mem_cgroup_commit_charge_swapin(page, memcg, | |
2889 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2890 | } | |
2891 | ||
2892 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) | |
2893 | { | |
2894 | if (mem_cgroup_disabled()) | |
2895 | return; | |
2896 | if (!memcg) | |
2897 | return; | |
2898 | __mem_cgroup_cancel_charge(memcg, 1); | |
2899 | } | |
2900 | ||
2901 | static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, | |
2902 | unsigned int nr_pages, | |
2903 | const enum charge_type ctype) | |
2904 | { | |
2905 | struct memcg_batch_info *batch = NULL; | |
2906 | bool uncharge_memsw = true; | |
2907 | ||
2908 | /* If swapout, usage of swap doesn't decrease */ | |
2909 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) | |
2910 | uncharge_memsw = false; | |
2911 | ||
2912 | batch = ¤t->memcg_batch; | |
2913 | /* | |
2914 | * In usual, we do css_get() when we remember memcg pointer. | |
2915 | * But in this case, we keep res->usage until end of a series of | |
2916 | * uncharges. Then, it's ok to ignore memcg's refcnt. | |
2917 | */ | |
2918 | if (!batch->memcg) | |
2919 | batch->memcg = memcg; | |
2920 | /* | |
2921 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. | |
2922 | * In those cases, all pages freed continuously can be expected to be in | |
2923 | * the same cgroup and we have chance to coalesce uncharges. | |
2924 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) | |
2925 | * because we want to do uncharge as soon as possible. | |
2926 | */ | |
2927 | ||
2928 | if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) | |
2929 | goto direct_uncharge; | |
2930 | ||
2931 | if (nr_pages > 1) | |
2932 | goto direct_uncharge; | |
2933 | ||
2934 | /* | |
2935 | * In typical case, batch->memcg == mem. This means we can | |
2936 | * merge a series of uncharges to an uncharge of res_counter. | |
2937 | * If not, we uncharge res_counter ony by one. | |
2938 | */ | |
2939 | if (batch->memcg != memcg) | |
2940 | goto direct_uncharge; | |
2941 | /* remember freed charge and uncharge it later */ | |
2942 | batch->nr_pages++; | |
2943 | if (uncharge_memsw) | |
2944 | batch->memsw_nr_pages++; | |
2945 | return; | |
2946 | direct_uncharge: | |
2947 | res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); | |
2948 | if (uncharge_memsw) | |
2949 | res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); | |
2950 | if (unlikely(batch->memcg != memcg)) | |
2951 | memcg_oom_recover(memcg); | |
2952 | } | |
2953 | ||
2954 | /* | |
2955 | * uncharge if !page_mapped(page) | |
2956 | */ | |
2957 | static struct mem_cgroup * | |
2958 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) | |
2959 | { | |
2960 | struct mem_cgroup *memcg = NULL; | |
2961 | unsigned int nr_pages = 1; | |
2962 | struct page_cgroup *pc; | |
2963 | bool anon; | |
2964 | ||
2965 | if (mem_cgroup_disabled()) | |
2966 | return NULL; | |
2967 | ||
2968 | if (PageSwapCache(page)) | |
2969 | return NULL; | |
2970 | ||
2971 | if (PageTransHuge(page)) { | |
2972 | nr_pages <<= compound_order(page); | |
2973 | VM_BUG_ON(!PageTransHuge(page)); | |
2974 | } | |
2975 | /* | |
2976 | * Check if our page_cgroup is valid | |
2977 | */ | |
2978 | pc = lookup_page_cgroup(page); | |
2979 | if (unlikely(!PageCgroupUsed(pc))) | |
2980 | return NULL; | |
2981 | ||
2982 | lock_page_cgroup(pc); | |
2983 | ||
2984 | memcg = pc->mem_cgroup; | |
2985 | ||
2986 | if (!PageCgroupUsed(pc)) | |
2987 | goto unlock_out; | |
2988 | ||
2989 | anon = PageAnon(page); | |
2990 | ||
2991 | switch (ctype) { | |
2992 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: | |
2993 | /* | |
2994 | * Generally PageAnon tells if it's the anon statistics to be | |
2995 | * updated; but sometimes e.g. mem_cgroup_uncharge_page() is | |
2996 | * used before page reached the stage of being marked PageAnon. | |
2997 | */ | |
2998 | anon = true; | |
2999 | /* fallthrough */ | |
3000 | case MEM_CGROUP_CHARGE_TYPE_DROP: | |
3001 | /* See mem_cgroup_prepare_migration() */ | |
3002 | if (page_mapped(page) || PageCgroupMigration(pc)) | |
3003 | goto unlock_out; | |
3004 | break; | |
3005 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: | |
3006 | if (!PageAnon(page)) { /* Shared memory */ | |
3007 | if (page->mapping && !page_is_file_cache(page)) | |
3008 | goto unlock_out; | |
3009 | } else if (page_mapped(page)) /* Anon */ | |
3010 | goto unlock_out; | |
3011 | break; | |
3012 | default: | |
3013 | break; | |
3014 | } | |
3015 | ||
3016 | mem_cgroup_charge_statistics(memcg, anon, -nr_pages); | |
3017 | ||
3018 | ClearPageCgroupUsed(pc); | |
3019 | /* | |
3020 | * pc->mem_cgroup is not cleared here. It will be accessed when it's | |
3021 | * freed from LRU. This is safe because uncharged page is expected not | |
3022 | * to be reused (freed soon). Exception is SwapCache, it's handled by | |
3023 | * special functions. | |
3024 | */ | |
3025 | ||
3026 | unlock_page_cgroup(pc); | |
3027 | /* | |
3028 | * even after unlock, we have memcg->res.usage here and this memcg | |
3029 | * will never be freed. | |
3030 | */ | |
3031 | memcg_check_events(memcg, page); | |
3032 | if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { | |
3033 | mem_cgroup_swap_statistics(memcg, true); | |
3034 | mem_cgroup_get(memcg); | |
3035 | } | |
3036 | if (!mem_cgroup_is_root(memcg)) | |
3037 | mem_cgroup_do_uncharge(memcg, nr_pages, ctype); | |
3038 | ||
3039 | return memcg; | |
3040 | ||
3041 | unlock_out: | |
3042 | unlock_page_cgroup(pc); | |
3043 | return NULL; | |
3044 | } | |
3045 | ||
3046 | void mem_cgroup_uncharge_page(struct page *page) | |
3047 | { | |
3048 | /* early check. */ | |
3049 | if (page_mapped(page)) | |
3050 | return; | |
3051 | VM_BUG_ON(page->mapping && !PageAnon(page)); | |
3052 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
3053 | } | |
3054 | ||
3055 | void mem_cgroup_uncharge_cache_page(struct page *page) | |
3056 | { | |
3057 | VM_BUG_ON(page_mapped(page)); | |
3058 | VM_BUG_ON(page->mapping); | |
3059 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); | |
3060 | } | |
3061 | ||
3062 | /* | |
3063 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. | |
3064 | * In that cases, pages are freed continuously and we can expect pages | |
3065 | * are in the same memcg. All these calls itself limits the number of | |
3066 | * pages freed at once, then uncharge_start/end() is called properly. | |
3067 | * This may be called prural(2) times in a context, | |
3068 | */ | |
3069 | ||
3070 | void mem_cgroup_uncharge_start(void) | |
3071 | { | |
3072 | current->memcg_batch.do_batch++; | |
3073 | /* We can do nest. */ | |
3074 | if (current->memcg_batch.do_batch == 1) { | |
3075 | current->memcg_batch.memcg = NULL; | |
3076 | current->memcg_batch.nr_pages = 0; | |
3077 | current->memcg_batch.memsw_nr_pages = 0; | |
3078 | } | |
3079 | } | |
3080 | ||
3081 | void mem_cgroup_uncharge_end(void) | |
3082 | { | |
3083 | struct memcg_batch_info *batch = ¤t->memcg_batch; | |
3084 | ||
3085 | if (!batch->do_batch) | |
3086 | return; | |
3087 | ||
3088 | batch->do_batch--; | |
3089 | if (batch->do_batch) /* If stacked, do nothing. */ | |
3090 | return; | |
3091 | ||
3092 | if (!batch->memcg) | |
3093 | return; | |
3094 | /* | |
3095 | * This "batch->memcg" is valid without any css_get/put etc... | |
3096 | * bacause we hide charges behind us. | |
3097 | */ | |
3098 | if (batch->nr_pages) | |
3099 | res_counter_uncharge(&batch->memcg->res, | |
3100 | batch->nr_pages * PAGE_SIZE); | |
3101 | if (batch->memsw_nr_pages) | |
3102 | res_counter_uncharge(&batch->memcg->memsw, | |
3103 | batch->memsw_nr_pages * PAGE_SIZE); | |
3104 | memcg_oom_recover(batch->memcg); | |
3105 | /* forget this pointer (for sanity check) */ | |
3106 | batch->memcg = NULL; | |
3107 | } | |
3108 | ||
3109 | #ifdef CONFIG_SWAP | |
3110 | /* | |
3111 | * called after __delete_from_swap_cache() and drop "page" account. | |
3112 | * memcg information is recorded to swap_cgroup of "ent" | |
3113 | */ | |
3114 | void | |
3115 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) | |
3116 | { | |
3117 | struct mem_cgroup *memcg; | |
3118 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; | |
3119 | ||
3120 | if (!swapout) /* this was a swap cache but the swap is unused ! */ | |
3121 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; | |
3122 | ||
3123 | memcg = __mem_cgroup_uncharge_common(page, ctype); | |
3124 | ||
3125 | /* | |
3126 | * record memcg information, if swapout && memcg != NULL, | |
3127 | * mem_cgroup_get() was called in uncharge(). | |
3128 | */ | |
3129 | if (do_swap_account && swapout && memcg) | |
3130 | swap_cgroup_record(ent, css_id(&memcg->css)); | |
3131 | } | |
3132 | #endif | |
3133 | ||
3134 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
3135 | /* | |
3136 | * called from swap_entry_free(). remove record in swap_cgroup and | |
3137 | * uncharge "memsw" account. | |
3138 | */ | |
3139 | void mem_cgroup_uncharge_swap(swp_entry_t ent) | |
3140 | { | |
3141 | struct mem_cgroup *memcg; | |
3142 | unsigned short id; | |
3143 | ||
3144 | if (!do_swap_account) | |
3145 | return; | |
3146 | ||
3147 | id = swap_cgroup_record(ent, 0); | |
3148 | rcu_read_lock(); | |
3149 | memcg = mem_cgroup_lookup(id); | |
3150 | if (memcg) { | |
3151 | /* | |
3152 | * We uncharge this because swap is freed. | |
3153 | * This memcg can be obsolete one. We avoid calling css_tryget | |
3154 | */ | |
3155 | if (!mem_cgroup_is_root(memcg)) | |
3156 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); | |
3157 | mem_cgroup_swap_statistics(memcg, false); | |
3158 | mem_cgroup_put(memcg); | |
3159 | } | |
3160 | rcu_read_unlock(); | |
3161 | } | |
3162 | ||
3163 | /** | |
3164 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. | |
3165 | * @entry: swap entry to be moved | |
3166 | * @from: mem_cgroup which the entry is moved from | |
3167 | * @to: mem_cgroup which the entry is moved to | |
3168 | * | |
3169 | * It succeeds only when the swap_cgroup's record for this entry is the same | |
3170 | * as the mem_cgroup's id of @from. | |
3171 | * | |
3172 | * Returns 0 on success, -EINVAL on failure. | |
3173 | * | |
3174 | * The caller must have charged to @to, IOW, called res_counter_charge() about | |
3175 | * both res and memsw, and called css_get(). | |
3176 | */ | |
3177 | static int mem_cgroup_move_swap_account(swp_entry_t entry, | |
3178 | struct mem_cgroup *from, struct mem_cgroup *to) | |
3179 | { | |
3180 | unsigned short old_id, new_id; | |
3181 | ||
3182 | old_id = css_id(&from->css); | |
3183 | new_id = css_id(&to->css); | |
3184 | ||
3185 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { | |
3186 | mem_cgroup_swap_statistics(from, false); | |
3187 | mem_cgroup_swap_statistics(to, true); | |
3188 | /* | |
3189 | * This function is only called from task migration context now. | |
3190 | * It postpones res_counter and refcount handling till the end | |
3191 | * of task migration(mem_cgroup_clear_mc()) for performance | |
3192 | * improvement. But we cannot postpone mem_cgroup_get(to) | |
3193 | * because if the process that has been moved to @to does | |
3194 | * swap-in, the refcount of @to might be decreased to 0. | |
3195 | */ | |
3196 | mem_cgroup_get(to); | |
3197 | return 0; | |
3198 | } | |
3199 | return -EINVAL; | |
3200 | } | |
3201 | #else | |
3202 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, | |
3203 | struct mem_cgroup *from, struct mem_cgroup *to) | |
3204 | { | |
3205 | return -EINVAL; | |
3206 | } | |
3207 | #endif | |
3208 | ||
3209 | /* | |
3210 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old | |
3211 | * page belongs to. | |
3212 | */ | |
3213 | int mem_cgroup_prepare_migration(struct page *page, | |
3214 | struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask) | |
3215 | { | |
3216 | struct mem_cgroup *memcg = NULL; | |
3217 | struct page_cgroup *pc; | |
3218 | enum charge_type ctype; | |
3219 | int ret = 0; | |
3220 | ||
3221 | *memcgp = NULL; | |
3222 | ||
3223 | VM_BUG_ON(PageTransHuge(page)); | |
3224 | if (mem_cgroup_disabled()) | |
3225 | return 0; | |
3226 | ||
3227 | pc = lookup_page_cgroup(page); | |
3228 | lock_page_cgroup(pc); | |
3229 | if (PageCgroupUsed(pc)) { | |
3230 | memcg = pc->mem_cgroup; | |
3231 | css_get(&memcg->css); | |
3232 | /* | |
3233 | * At migrating an anonymous page, its mapcount goes down | |
3234 | * to 0 and uncharge() will be called. But, even if it's fully | |
3235 | * unmapped, migration may fail and this page has to be | |
3236 | * charged again. We set MIGRATION flag here and delay uncharge | |
3237 | * until end_migration() is called | |
3238 | * | |
3239 | * Corner Case Thinking | |
3240 | * A) | |
3241 | * When the old page was mapped as Anon and it's unmap-and-freed | |
3242 | * while migration was ongoing. | |
3243 | * If unmap finds the old page, uncharge() of it will be delayed | |
3244 | * until end_migration(). If unmap finds a new page, it's | |
3245 | * uncharged when it make mapcount to be 1->0. If unmap code | |
3246 | * finds swap_migration_entry, the new page will not be mapped | |
3247 | * and end_migration() will find it(mapcount==0). | |
3248 | * | |
3249 | * B) | |
3250 | * When the old page was mapped but migraion fails, the kernel | |
3251 | * remaps it. A charge for it is kept by MIGRATION flag even | |
3252 | * if mapcount goes down to 0. We can do remap successfully | |
3253 | * without charging it again. | |
3254 | * | |
3255 | * C) | |
3256 | * The "old" page is under lock_page() until the end of | |
3257 | * migration, so, the old page itself will not be swapped-out. | |
3258 | * If the new page is swapped out before end_migraton, our | |
3259 | * hook to usual swap-out path will catch the event. | |
3260 | */ | |
3261 | if (PageAnon(page)) | |
3262 | SetPageCgroupMigration(pc); | |
3263 | } | |
3264 | unlock_page_cgroup(pc); | |
3265 | /* | |
3266 | * If the page is not charged at this point, | |
3267 | * we return here. | |
3268 | */ | |
3269 | if (!memcg) | |
3270 | return 0; | |
3271 | ||
3272 | *memcgp = memcg; | |
3273 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false); | |
3274 | css_put(&memcg->css);/* drop extra refcnt */ | |
3275 | if (ret) { | |
3276 | if (PageAnon(page)) { | |
3277 | lock_page_cgroup(pc); | |
3278 | ClearPageCgroupMigration(pc); | |
3279 | unlock_page_cgroup(pc); | |
3280 | /* | |
3281 | * The old page may be fully unmapped while we kept it. | |
3282 | */ | |
3283 | mem_cgroup_uncharge_page(page); | |
3284 | } | |
3285 | /* we'll need to revisit this error code (we have -EINTR) */ | |
3286 | return -ENOMEM; | |
3287 | } | |
3288 | /* | |
3289 | * We charge new page before it's used/mapped. So, even if unlock_page() | |
3290 | * is called before end_migration, we can catch all events on this new | |
3291 | * page. In the case new page is migrated but not remapped, new page's | |
3292 | * mapcount will be finally 0 and we call uncharge in end_migration(). | |
3293 | */ | |
3294 | if (PageAnon(page)) | |
3295 | ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; | |
3296 | else if (page_is_file_cache(page)) | |
3297 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
3298 | else | |
3299 | ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; | |
3300 | __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false); | |
3301 | return ret; | |
3302 | } | |
3303 | ||
3304 | /* remove redundant charge if migration failed*/ | |
3305 | void mem_cgroup_end_migration(struct mem_cgroup *memcg, | |
3306 | struct page *oldpage, struct page *newpage, bool migration_ok) | |
3307 | { | |
3308 | struct page *used, *unused; | |
3309 | struct page_cgroup *pc; | |
3310 | bool anon; | |
3311 | ||
3312 | if (!memcg) | |
3313 | return; | |
3314 | /* blocks rmdir() */ | |
3315 | cgroup_exclude_rmdir(&memcg->css); | |
3316 | if (!migration_ok) { | |
3317 | used = oldpage; | |
3318 | unused = newpage; | |
3319 | } else { | |
3320 | used = newpage; | |
3321 | unused = oldpage; | |
3322 | } | |
3323 | /* | |
3324 | * We disallowed uncharge of pages under migration because mapcount | |
3325 | * of the page goes down to zero, temporarly. | |
3326 | * Clear the flag and check the page should be charged. | |
3327 | */ | |
3328 | pc = lookup_page_cgroup(oldpage); | |
3329 | lock_page_cgroup(pc); | |
3330 | ClearPageCgroupMigration(pc); | |
3331 | unlock_page_cgroup(pc); | |
3332 | anon = PageAnon(used); | |
3333 | __mem_cgroup_uncharge_common(unused, | |
3334 | anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED | |
3335 | : MEM_CGROUP_CHARGE_TYPE_CACHE); | |
3336 | ||
3337 | /* | |
3338 | * If a page is a file cache, radix-tree replacement is very atomic | |
3339 | * and we can skip this check. When it was an Anon page, its mapcount | |
3340 | * goes down to 0. But because we added MIGRATION flage, it's not | |
3341 | * uncharged yet. There are several case but page->mapcount check | |
3342 | * and USED bit check in mem_cgroup_uncharge_page() will do enough | |
3343 | * check. (see prepare_charge() also) | |
3344 | */ | |
3345 | if (anon) | |
3346 | mem_cgroup_uncharge_page(used); | |
3347 | /* | |
3348 | * At migration, we may charge account against cgroup which has no | |
3349 | * tasks. | |
3350 | * So, rmdir()->pre_destroy() can be called while we do this charge. | |
3351 | * In that case, we need to call pre_destroy() again. check it here. | |
3352 | */ | |
3353 | cgroup_release_and_wakeup_rmdir(&memcg->css); | |
3354 | } | |
3355 | ||
3356 | /* | |
3357 | * At replace page cache, newpage is not under any memcg but it's on | |
3358 | * LRU. So, this function doesn't touch res_counter but handles LRU | |
3359 | * in correct way. Both pages are locked so we cannot race with uncharge. | |
3360 | */ | |
3361 | void mem_cgroup_replace_page_cache(struct page *oldpage, | |
3362 | struct page *newpage) | |
3363 | { | |
3364 | struct mem_cgroup *memcg = NULL; | |
3365 | struct page_cgroup *pc; | |
3366 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
3367 | ||
3368 | if (mem_cgroup_disabled()) | |
3369 | return; | |
3370 | ||
3371 | pc = lookup_page_cgroup(oldpage); | |
3372 | /* fix accounting on old pages */ | |
3373 | lock_page_cgroup(pc); | |
3374 | if (PageCgroupUsed(pc)) { | |
3375 | memcg = pc->mem_cgroup; | |
3376 | mem_cgroup_charge_statistics(memcg, false, -1); | |
3377 | ClearPageCgroupUsed(pc); | |
3378 | } | |
3379 | unlock_page_cgroup(pc); | |
3380 | ||
3381 | /* | |
3382 | * When called from shmem_replace_page(), in some cases the | |
3383 | * oldpage has already been charged, and in some cases not. | |
3384 | */ | |
3385 | if (!memcg) | |
3386 | return; | |
3387 | ||
3388 | if (PageSwapBacked(oldpage)) | |
3389 | type = MEM_CGROUP_CHARGE_TYPE_SHMEM; | |
3390 | ||
3391 | /* | |
3392 | * Even if newpage->mapping was NULL before starting replacement, | |
3393 | * the newpage may be on LRU(or pagevec for LRU) already. We lock | |
3394 | * LRU while we overwrite pc->mem_cgroup. | |
3395 | */ | |
3396 | __mem_cgroup_commit_charge(memcg, newpage, 1, type, true); | |
3397 | } | |
3398 | ||
3399 | #ifdef CONFIG_DEBUG_VM | |
3400 | static struct page_cgroup *lookup_page_cgroup_used(struct page *page) | |
3401 | { | |
3402 | struct page_cgroup *pc; | |
3403 | ||
3404 | pc = lookup_page_cgroup(page); | |
3405 | /* | |
3406 | * Can be NULL while feeding pages into the page allocator for | |
3407 | * the first time, i.e. during boot or memory hotplug; | |
3408 | * or when mem_cgroup_disabled(). | |
3409 | */ | |
3410 | if (likely(pc) && PageCgroupUsed(pc)) | |
3411 | return pc; | |
3412 | return NULL; | |
3413 | } | |
3414 | ||
3415 | bool mem_cgroup_bad_page_check(struct page *page) | |
3416 | { | |
3417 | if (mem_cgroup_disabled()) | |
3418 | return false; | |
3419 | ||
3420 | return lookup_page_cgroup_used(page) != NULL; | |
3421 | } | |
3422 | ||
3423 | void mem_cgroup_print_bad_page(struct page *page) | |
3424 | { | |
3425 | struct page_cgroup *pc; | |
3426 | ||
3427 | pc = lookup_page_cgroup_used(page); | |
3428 | if (pc) { | |
3429 | printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n", | |
3430 | pc, pc->flags, pc->mem_cgroup); | |
3431 | } | |
3432 | } | |
3433 | #endif | |
3434 | ||
3435 | static DEFINE_MUTEX(set_limit_mutex); | |
3436 | ||
3437 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, | |
3438 | unsigned long long val) | |
3439 | { | |
3440 | int retry_count; | |
3441 | u64 memswlimit, memlimit; | |
3442 | int ret = 0; | |
3443 | int children = mem_cgroup_count_children(memcg); | |
3444 | u64 curusage, oldusage; | |
3445 | int enlarge; | |
3446 | ||
3447 | /* | |
3448 | * For keeping hierarchical_reclaim simple, how long we should retry | |
3449 | * is depends on callers. We set our retry-count to be function | |
3450 | * of # of children which we should visit in this loop. | |
3451 | */ | |
3452 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; | |
3453 | ||
3454 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
3455 | ||
3456 | enlarge = 0; | |
3457 | while (retry_count) { | |
3458 | if (signal_pending(current)) { | |
3459 | ret = -EINTR; | |
3460 | break; | |
3461 | } | |
3462 | /* | |
3463 | * Rather than hide all in some function, I do this in | |
3464 | * open coded manner. You see what this really does. | |
3465 | * We have to guarantee memcg->res.limit < memcg->memsw.limit. | |
3466 | */ | |
3467 | mutex_lock(&set_limit_mutex); | |
3468 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3469 | if (memswlimit < val) { | |
3470 | ret = -EINVAL; | |
3471 | mutex_unlock(&set_limit_mutex); | |
3472 | break; | |
3473 | } | |
3474 | ||
3475 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3476 | if (memlimit < val) | |
3477 | enlarge = 1; | |
3478 | ||
3479 | ret = res_counter_set_limit(&memcg->res, val); | |
3480 | if (!ret) { | |
3481 | if (memswlimit == val) | |
3482 | memcg->memsw_is_minimum = true; | |
3483 | else | |
3484 | memcg->memsw_is_minimum = false; | |
3485 | } | |
3486 | mutex_unlock(&set_limit_mutex); | |
3487 | ||
3488 | if (!ret) | |
3489 | break; | |
3490 | ||
3491 | mem_cgroup_reclaim(memcg, GFP_KERNEL, | |
3492 | MEM_CGROUP_RECLAIM_SHRINK); | |
3493 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
3494 | /* Usage is reduced ? */ | |
3495 | if (curusage >= oldusage) | |
3496 | retry_count--; | |
3497 | else | |
3498 | oldusage = curusage; | |
3499 | } | |
3500 | if (!ret && enlarge) | |
3501 | memcg_oom_recover(memcg); | |
3502 | ||
3503 | return ret; | |
3504 | } | |
3505 | ||
3506 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, | |
3507 | unsigned long long val) | |
3508 | { | |
3509 | int retry_count; | |
3510 | u64 memlimit, memswlimit, oldusage, curusage; | |
3511 | int children = mem_cgroup_count_children(memcg); | |
3512 | int ret = -EBUSY; | |
3513 | int enlarge = 0; | |
3514 | ||
3515 | /* see mem_cgroup_resize_res_limit */ | |
3516 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; | |
3517 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
3518 | while (retry_count) { | |
3519 | if (signal_pending(current)) { | |
3520 | ret = -EINTR; | |
3521 | break; | |
3522 | } | |
3523 | /* | |
3524 | * Rather than hide all in some function, I do this in | |
3525 | * open coded manner. You see what this really does. | |
3526 | * We have to guarantee memcg->res.limit < memcg->memsw.limit. | |
3527 | */ | |
3528 | mutex_lock(&set_limit_mutex); | |
3529 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3530 | if (memlimit > val) { | |
3531 | ret = -EINVAL; | |
3532 | mutex_unlock(&set_limit_mutex); | |
3533 | break; | |
3534 | } | |
3535 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3536 | if (memswlimit < val) | |
3537 | enlarge = 1; | |
3538 | ret = res_counter_set_limit(&memcg->memsw, val); | |
3539 | if (!ret) { | |
3540 | if (memlimit == val) | |
3541 | memcg->memsw_is_minimum = true; | |
3542 | else | |
3543 | memcg->memsw_is_minimum = false; | |
3544 | } | |
3545 | mutex_unlock(&set_limit_mutex); | |
3546 | ||
3547 | if (!ret) | |
3548 | break; | |
3549 | ||
3550 | mem_cgroup_reclaim(memcg, GFP_KERNEL, | |
3551 | MEM_CGROUP_RECLAIM_NOSWAP | | |
3552 | MEM_CGROUP_RECLAIM_SHRINK); | |
3553 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
3554 | /* Usage is reduced ? */ | |
3555 | if (curusage >= oldusage) | |
3556 | retry_count--; | |
3557 | else | |
3558 | oldusage = curusage; | |
3559 | } | |
3560 | if (!ret && enlarge) | |
3561 | memcg_oom_recover(memcg); | |
3562 | return ret; | |
3563 | } | |
3564 | ||
3565 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, | |
3566 | gfp_t gfp_mask, | |
3567 | unsigned long *total_scanned) | |
3568 | { | |
3569 | unsigned long nr_reclaimed = 0; | |
3570 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; | |
3571 | unsigned long reclaimed; | |
3572 | int loop = 0; | |
3573 | struct mem_cgroup_tree_per_zone *mctz; | |
3574 | unsigned long long excess; | |
3575 | unsigned long nr_scanned; | |
3576 | ||
3577 | if (order > 0) | |
3578 | return 0; | |
3579 | ||
3580 | mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); | |
3581 | /* | |
3582 | * This loop can run a while, specially if mem_cgroup's continuously | |
3583 | * keep exceeding their soft limit and putting the system under | |
3584 | * pressure | |
3585 | */ | |
3586 | do { | |
3587 | if (next_mz) | |
3588 | mz = next_mz; | |
3589 | else | |
3590 | mz = mem_cgroup_largest_soft_limit_node(mctz); | |
3591 | if (!mz) | |
3592 | break; | |
3593 | ||
3594 | nr_scanned = 0; | |
3595 | reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone, | |
3596 | gfp_mask, &nr_scanned); | |
3597 | nr_reclaimed += reclaimed; | |
3598 | *total_scanned += nr_scanned; | |
3599 | spin_lock(&mctz->lock); | |
3600 | ||
3601 | /* | |
3602 | * If we failed to reclaim anything from this memory cgroup | |
3603 | * it is time to move on to the next cgroup | |
3604 | */ | |
3605 | next_mz = NULL; | |
3606 | if (!reclaimed) { | |
3607 | do { | |
3608 | /* | |
3609 | * Loop until we find yet another one. | |
3610 | * | |
3611 | * By the time we get the soft_limit lock | |
3612 | * again, someone might have aded the | |
3613 | * group back on the RB tree. Iterate to | |
3614 | * make sure we get a different mem. | |
3615 | * mem_cgroup_largest_soft_limit_node returns | |
3616 | * NULL if no other cgroup is present on | |
3617 | * the tree | |
3618 | */ | |
3619 | next_mz = | |
3620 | __mem_cgroup_largest_soft_limit_node(mctz); | |
3621 | if (next_mz == mz) | |
3622 | css_put(&next_mz->memcg->css); | |
3623 | else /* next_mz == NULL or other memcg */ | |
3624 | break; | |
3625 | } while (1); | |
3626 | } | |
3627 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); | |
3628 | excess = res_counter_soft_limit_excess(&mz->memcg->res); | |
3629 | /* | |
3630 | * One school of thought says that we should not add | |
3631 | * back the node to the tree if reclaim returns 0. | |
3632 | * But our reclaim could return 0, simply because due | |
3633 | * to priority we are exposing a smaller subset of | |
3634 | * memory to reclaim from. Consider this as a longer | |
3635 | * term TODO. | |
3636 | */ | |
3637 | /* If excess == 0, no tree ops */ | |
3638 | __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess); | |
3639 | spin_unlock(&mctz->lock); | |
3640 | css_put(&mz->memcg->css); | |
3641 | loop++; | |
3642 | /* | |
3643 | * Could not reclaim anything and there are no more | |
3644 | * mem cgroups to try or we seem to be looping without | |
3645 | * reclaiming anything. | |
3646 | */ | |
3647 | if (!nr_reclaimed && | |
3648 | (next_mz == NULL || | |
3649 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) | |
3650 | break; | |
3651 | } while (!nr_reclaimed); | |
3652 | if (next_mz) | |
3653 | css_put(&next_mz->memcg->css); | |
3654 | return nr_reclaimed; | |
3655 | } | |
3656 | ||
3657 | /* | |
3658 | * This routine traverse page_cgroup in given list and drop them all. | |
3659 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. | |
3660 | */ | |
3661 | static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, | |
3662 | int node, int zid, enum lru_list lru) | |
3663 | { | |
3664 | struct mem_cgroup_per_zone *mz; | |
3665 | unsigned long flags, loop; | |
3666 | struct list_head *list; | |
3667 | struct page *busy; | |
3668 | struct zone *zone; | |
3669 | int ret = 0; | |
3670 | ||
3671 | zone = &NODE_DATA(node)->node_zones[zid]; | |
3672 | mz = mem_cgroup_zoneinfo(memcg, node, zid); | |
3673 | list = &mz->lruvec.lists[lru]; | |
3674 | ||
3675 | loop = mz->lru_size[lru]; | |
3676 | /* give some margin against EBUSY etc...*/ | |
3677 | loop += 256; | |
3678 | busy = NULL; | |
3679 | while (loop--) { | |
3680 | struct page_cgroup *pc; | |
3681 | struct page *page; | |
3682 | ||
3683 | ret = 0; | |
3684 | spin_lock_irqsave(&zone->lru_lock, flags); | |
3685 | if (list_empty(list)) { | |
3686 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3687 | break; | |
3688 | } | |
3689 | page = list_entry(list->prev, struct page, lru); | |
3690 | if (busy == page) { | |
3691 | list_move(&page->lru, list); | |
3692 | busy = NULL; | |
3693 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3694 | continue; | |
3695 | } | |
3696 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3697 | ||
3698 | pc = lookup_page_cgroup(page); | |
3699 | ||
3700 | ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); | |
3701 | if (ret == -ENOMEM || ret == -EINTR) | |
3702 | break; | |
3703 | ||
3704 | if (ret == -EBUSY || ret == -EINVAL) { | |
3705 | /* found lock contention or "pc" is obsolete. */ | |
3706 | busy = page; | |
3707 | cond_resched(); | |
3708 | } else | |
3709 | busy = NULL; | |
3710 | } | |
3711 | ||
3712 | if (!ret && !list_empty(list)) | |
3713 | return -EBUSY; | |
3714 | return ret; | |
3715 | } | |
3716 | ||
3717 | /* | |
3718 | * make mem_cgroup's charge to be 0 if there is no task. | |
3719 | * This enables deleting this mem_cgroup. | |
3720 | */ | |
3721 | static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) | |
3722 | { | |
3723 | int ret; | |
3724 | int node, zid, shrink; | |
3725 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
3726 | struct cgroup *cgrp = memcg->css.cgroup; | |
3727 | ||
3728 | css_get(&memcg->css); | |
3729 | ||
3730 | shrink = 0; | |
3731 | /* should free all ? */ | |
3732 | if (free_all) | |
3733 | goto try_to_free; | |
3734 | move_account: | |
3735 | do { | |
3736 | ret = -EBUSY; | |
3737 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) | |
3738 | goto out; | |
3739 | ret = -EINTR; | |
3740 | if (signal_pending(current)) | |
3741 | goto out; | |
3742 | /* This is for making all *used* pages to be on LRU. */ | |
3743 | lru_add_drain_all(); | |
3744 | drain_all_stock_sync(memcg); | |
3745 | ret = 0; | |
3746 | mem_cgroup_start_move(memcg); | |
3747 | for_each_node_state(node, N_HIGH_MEMORY) { | |
3748 | for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { | |
3749 | enum lru_list lru; | |
3750 | for_each_lru(lru) { | |
3751 | ret = mem_cgroup_force_empty_list(memcg, | |
3752 | node, zid, lru); | |
3753 | if (ret) | |
3754 | break; | |
3755 | } | |
3756 | } | |
3757 | if (ret) | |
3758 | break; | |
3759 | } | |
3760 | mem_cgroup_end_move(memcg); | |
3761 | memcg_oom_recover(memcg); | |
3762 | /* it seems parent cgroup doesn't have enough mem */ | |
3763 | if (ret == -ENOMEM) | |
3764 | goto try_to_free; | |
3765 | cond_resched(); | |
3766 | /* "ret" should also be checked to ensure all lists are empty. */ | |
3767 | } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret); | |
3768 | out: | |
3769 | css_put(&memcg->css); | |
3770 | return ret; | |
3771 | ||
3772 | try_to_free: | |
3773 | /* returns EBUSY if there is a task or if we come here twice. */ | |
3774 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { | |
3775 | ret = -EBUSY; | |
3776 | goto out; | |
3777 | } | |
3778 | /* we call try-to-free pages for make this cgroup empty */ | |
3779 | lru_add_drain_all(); | |
3780 | /* try to free all pages in this cgroup */ | |
3781 | shrink = 1; | |
3782 | while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) { | |
3783 | int progress; | |
3784 | ||
3785 | if (signal_pending(current)) { | |
3786 | ret = -EINTR; | |
3787 | goto out; | |
3788 | } | |
3789 | progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, | |
3790 | false); | |
3791 | if (!progress) { | |
3792 | nr_retries--; | |
3793 | /* maybe some writeback is necessary */ | |
3794 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
3795 | } | |
3796 | ||
3797 | } | |
3798 | lru_add_drain(); | |
3799 | /* try move_account...there may be some *locked* pages. */ | |
3800 | goto move_account; | |
3801 | } | |
3802 | ||
3803 | int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) | |
3804 | { | |
3805 | return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); | |
3806 | } | |
3807 | ||
3808 | ||
3809 | static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) | |
3810 | { | |
3811 | return mem_cgroup_from_cont(cont)->use_hierarchy; | |
3812 | } | |
3813 | ||
3814 | static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, | |
3815 | u64 val) | |
3816 | { | |
3817 | int retval = 0; | |
3818 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
3819 | struct cgroup *parent = cont->parent; | |
3820 | struct mem_cgroup *parent_memcg = NULL; | |
3821 | ||
3822 | if (parent) | |
3823 | parent_memcg = mem_cgroup_from_cont(parent); | |
3824 | ||
3825 | cgroup_lock(); | |
3826 | /* | |
3827 | * If parent's use_hierarchy is set, we can't make any modifications | |
3828 | * in the child subtrees. If it is unset, then the change can | |
3829 | * occur, provided the current cgroup has no children. | |
3830 | * | |
3831 | * For the root cgroup, parent_mem is NULL, we allow value to be | |
3832 | * set if there are no children. | |
3833 | */ | |
3834 | if ((!parent_memcg || !parent_memcg->use_hierarchy) && | |
3835 | (val == 1 || val == 0)) { | |
3836 | if (list_empty(&cont->children)) | |
3837 | memcg->use_hierarchy = val; | |
3838 | else | |
3839 | retval = -EBUSY; | |
3840 | } else | |
3841 | retval = -EINVAL; | |
3842 | cgroup_unlock(); | |
3843 | ||
3844 | return retval; | |
3845 | } | |
3846 | ||
3847 | ||
3848 | static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, | |
3849 | enum mem_cgroup_stat_index idx) | |
3850 | { | |
3851 | struct mem_cgroup *iter; | |
3852 | long val = 0; | |
3853 | ||
3854 | /* Per-cpu values can be negative, use a signed accumulator */ | |
3855 | for_each_mem_cgroup_tree(iter, memcg) | |
3856 | val += mem_cgroup_read_stat(iter, idx); | |
3857 | ||
3858 | if (val < 0) /* race ? */ | |
3859 | val = 0; | |
3860 | return val; | |
3861 | } | |
3862 | ||
3863 | static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) | |
3864 | { | |
3865 | u64 val; | |
3866 | ||
3867 | if (!mem_cgroup_is_root(memcg)) { | |
3868 | if (!swap) | |
3869 | return res_counter_read_u64(&memcg->res, RES_USAGE); | |
3870 | else | |
3871 | return res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
3872 | } | |
3873 | ||
3874 | val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); | |
3875 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); | |
3876 | ||
3877 | if (swap) | |
3878 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); | |
3879 | ||
3880 | return val << PAGE_SHIFT; | |
3881 | } | |
3882 | ||
3883 | static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft, | |
3884 | struct file *file, char __user *buf, | |
3885 | size_t nbytes, loff_t *ppos) | |
3886 | { | |
3887 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
3888 | char str[64]; | |
3889 | u64 val; | |
3890 | int type, name, len; | |
3891 | ||
3892 | type = MEMFILE_TYPE(cft->private); | |
3893 | name = MEMFILE_ATTR(cft->private); | |
3894 | ||
3895 | if (!do_swap_account && type == _MEMSWAP) | |
3896 | return -EOPNOTSUPP; | |
3897 | ||
3898 | switch (type) { | |
3899 | case _MEM: | |
3900 | if (name == RES_USAGE) | |
3901 | val = mem_cgroup_usage(memcg, false); | |
3902 | else | |
3903 | val = res_counter_read_u64(&memcg->res, name); | |
3904 | break; | |
3905 | case _MEMSWAP: | |
3906 | if (name == RES_USAGE) | |
3907 | val = mem_cgroup_usage(memcg, true); | |
3908 | else | |
3909 | val = res_counter_read_u64(&memcg->memsw, name); | |
3910 | break; | |
3911 | default: | |
3912 | BUG(); | |
3913 | } | |
3914 | ||
3915 | len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val); | |
3916 | return simple_read_from_buffer(buf, nbytes, ppos, str, len); | |
3917 | } | |
3918 | /* | |
3919 | * The user of this function is... | |
3920 | * RES_LIMIT. | |
3921 | */ | |
3922 | static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, | |
3923 | const char *buffer) | |
3924 | { | |
3925 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
3926 | int type, name; | |
3927 | unsigned long long val; | |
3928 | int ret; | |
3929 | ||
3930 | type = MEMFILE_TYPE(cft->private); | |
3931 | name = MEMFILE_ATTR(cft->private); | |
3932 | ||
3933 | if (!do_swap_account && type == _MEMSWAP) | |
3934 | return -EOPNOTSUPP; | |
3935 | ||
3936 | switch (name) { | |
3937 | case RES_LIMIT: | |
3938 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ | |
3939 | ret = -EINVAL; | |
3940 | break; | |
3941 | } | |
3942 | /* This function does all necessary parse...reuse it */ | |
3943 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
3944 | if (ret) | |
3945 | break; | |
3946 | if (type == _MEM) | |
3947 | ret = mem_cgroup_resize_limit(memcg, val); | |
3948 | else | |
3949 | ret = mem_cgroup_resize_memsw_limit(memcg, val); | |
3950 | break; | |
3951 | case RES_SOFT_LIMIT: | |
3952 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
3953 | if (ret) | |
3954 | break; | |
3955 | /* | |
3956 | * For memsw, soft limits are hard to implement in terms | |
3957 | * of semantics, for now, we support soft limits for | |
3958 | * control without swap | |
3959 | */ | |
3960 | if (type == _MEM) | |
3961 | ret = res_counter_set_soft_limit(&memcg->res, val); | |
3962 | else | |
3963 | ret = -EINVAL; | |
3964 | break; | |
3965 | default: | |
3966 | ret = -EINVAL; /* should be BUG() ? */ | |
3967 | break; | |
3968 | } | |
3969 | return ret; | |
3970 | } | |
3971 | ||
3972 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, | |
3973 | unsigned long long *mem_limit, unsigned long long *memsw_limit) | |
3974 | { | |
3975 | struct cgroup *cgroup; | |
3976 | unsigned long long min_limit, min_memsw_limit, tmp; | |
3977 | ||
3978 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3979 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3980 | cgroup = memcg->css.cgroup; | |
3981 | if (!memcg->use_hierarchy) | |
3982 | goto out; | |
3983 | ||
3984 | while (cgroup->parent) { | |
3985 | cgroup = cgroup->parent; | |
3986 | memcg = mem_cgroup_from_cont(cgroup); | |
3987 | if (!memcg->use_hierarchy) | |
3988 | break; | |
3989 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3990 | min_limit = min(min_limit, tmp); | |
3991 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3992 | min_memsw_limit = min(min_memsw_limit, tmp); | |
3993 | } | |
3994 | out: | |
3995 | *mem_limit = min_limit; | |
3996 | *memsw_limit = min_memsw_limit; | |
3997 | } | |
3998 | ||
3999 | static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) | |
4000 | { | |
4001 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
4002 | int type, name; | |
4003 | ||
4004 | type = MEMFILE_TYPE(event); | |
4005 | name = MEMFILE_ATTR(event); | |
4006 | ||
4007 | if (!do_swap_account && type == _MEMSWAP) | |
4008 | return -EOPNOTSUPP; | |
4009 | ||
4010 | switch (name) { | |
4011 | case RES_MAX_USAGE: | |
4012 | if (type == _MEM) | |
4013 | res_counter_reset_max(&memcg->res); | |
4014 | else | |
4015 | res_counter_reset_max(&memcg->memsw); | |
4016 | break; | |
4017 | case RES_FAILCNT: | |
4018 | if (type == _MEM) | |
4019 | res_counter_reset_failcnt(&memcg->res); | |
4020 | else | |
4021 | res_counter_reset_failcnt(&memcg->memsw); | |
4022 | break; | |
4023 | } | |
4024 | ||
4025 | return 0; | |
4026 | } | |
4027 | ||
4028 | static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, | |
4029 | struct cftype *cft) | |
4030 | { | |
4031 | return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; | |
4032 | } | |
4033 | ||
4034 | #ifdef CONFIG_MMU | |
4035 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, | |
4036 | struct cftype *cft, u64 val) | |
4037 | { | |
4038 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4039 | ||
4040 | if (val >= (1 << NR_MOVE_TYPE)) | |
4041 | return -EINVAL; | |
4042 | /* | |
4043 | * We check this value several times in both in can_attach() and | |
4044 | * attach(), so we need cgroup lock to prevent this value from being | |
4045 | * inconsistent. | |
4046 | */ | |
4047 | cgroup_lock(); | |
4048 | memcg->move_charge_at_immigrate = val; | |
4049 | cgroup_unlock(); | |
4050 | ||
4051 | return 0; | |
4052 | } | |
4053 | #else | |
4054 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, | |
4055 | struct cftype *cft, u64 val) | |
4056 | { | |
4057 | return -ENOSYS; | |
4058 | } | |
4059 | #endif | |
4060 | ||
4061 | ||
4062 | /* For read statistics */ | |
4063 | enum { | |
4064 | MCS_CACHE, | |
4065 | MCS_RSS, | |
4066 | MCS_FILE_MAPPED, | |
4067 | MCS_PGPGIN, | |
4068 | MCS_PGPGOUT, | |
4069 | MCS_SWAP, | |
4070 | MCS_PGFAULT, | |
4071 | MCS_PGMAJFAULT, | |
4072 | MCS_INACTIVE_ANON, | |
4073 | MCS_ACTIVE_ANON, | |
4074 | MCS_INACTIVE_FILE, | |
4075 | MCS_ACTIVE_FILE, | |
4076 | MCS_UNEVICTABLE, | |
4077 | NR_MCS_STAT, | |
4078 | }; | |
4079 | ||
4080 | struct mcs_total_stat { | |
4081 | s64 stat[NR_MCS_STAT]; | |
4082 | }; | |
4083 | ||
4084 | struct { | |
4085 | char *local_name; | |
4086 | char *total_name; | |
4087 | } memcg_stat_strings[NR_MCS_STAT] = { | |
4088 | {"cache", "total_cache"}, | |
4089 | {"rss", "total_rss"}, | |
4090 | {"mapped_file", "total_mapped_file"}, | |
4091 | {"pgpgin", "total_pgpgin"}, | |
4092 | {"pgpgout", "total_pgpgout"}, | |
4093 | {"swap", "total_swap"}, | |
4094 | {"pgfault", "total_pgfault"}, | |
4095 | {"pgmajfault", "total_pgmajfault"}, | |
4096 | {"inactive_anon", "total_inactive_anon"}, | |
4097 | {"active_anon", "total_active_anon"}, | |
4098 | {"inactive_file", "total_inactive_file"}, | |
4099 | {"active_file", "total_active_file"}, | |
4100 | {"unevictable", "total_unevictable"} | |
4101 | }; | |
4102 | ||
4103 | ||
4104 | static void | |
4105 | mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) | |
4106 | { | |
4107 | s64 val; | |
4108 | ||
4109 | /* per cpu stat */ | |
4110 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); | |
4111 | s->stat[MCS_CACHE] += val * PAGE_SIZE; | |
4112 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); | |
4113 | s->stat[MCS_RSS] += val * PAGE_SIZE; | |
4114 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); | |
4115 | s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; | |
4116 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); | |
4117 | s->stat[MCS_PGPGIN] += val; | |
4118 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); | |
4119 | s->stat[MCS_PGPGOUT] += val; | |
4120 | if (do_swap_account) { | |
4121 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); | |
4122 | s->stat[MCS_SWAP] += val * PAGE_SIZE; | |
4123 | } | |
4124 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); | |
4125 | s->stat[MCS_PGFAULT] += val; | |
4126 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); | |
4127 | s->stat[MCS_PGMAJFAULT] += val; | |
4128 | ||
4129 | /* per zone stat */ | |
4130 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); | |
4131 | s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; | |
4132 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); | |
4133 | s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; | |
4134 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); | |
4135 | s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; | |
4136 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); | |
4137 | s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; | |
4138 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); | |
4139 | s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; | |
4140 | } | |
4141 | ||
4142 | static void | |
4143 | mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) | |
4144 | { | |
4145 | struct mem_cgroup *iter; | |
4146 | ||
4147 | for_each_mem_cgroup_tree(iter, memcg) | |
4148 | mem_cgroup_get_local_stat(iter, s); | |
4149 | } | |
4150 | ||
4151 | #ifdef CONFIG_NUMA | |
4152 | static int mem_control_numa_stat_show(struct seq_file *m, void *arg) | |
4153 | { | |
4154 | int nid; | |
4155 | unsigned long total_nr, file_nr, anon_nr, unevictable_nr; | |
4156 | unsigned long node_nr; | |
4157 | struct cgroup *cont = m->private; | |
4158 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
4159 | ||
4160 | total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL); | |
4161 | seq_printf(m, "total=%lu", total_nr); | |
4162 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
4163 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL); | |
4164 | seq_printf(m, " N%d=%lu", nid, node_nr); | |
4165 | } | |
4166 | seq_putc(m, '\n'); | |
4167 | ||
4168 | file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE); | |
4169 | seq_printf(m, "file=%lu", file_nr); | |
4170 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
4171 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, | |
4172 | LRU_ALL_FILE); | |
4173 | seq_printf(m, " N%d=%lu", nid, node_nr); | |
4174 | } | |
4175 | seq_putc(m, '\n'); | |
4176 | ||
4177 | anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON); | |
4178 | seq_printf(m, "anon=%lu", anon_nr); | |
4179 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
4180 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, | |
4181 | LRU_ALL_ANON); | |
4182 | seq_printf(m, " N%d=%lu", nid, node_nr); | |
4183 | } | |
4184 | seq_putc(m, '\n'); | |
4185 | ||
4186 | unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); | |
4187 | seq_printf(m, "unevictable=%lu", unevictable_nr); | |
4188 | for_each_node_state(nid, N_HIGH_MEMORY) { | |
4189 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, | |
4190 | BIT(LRU_UNEVICTABLE)); | |
4191 | seq_printf(m, " N%d=%lu", nid, node_nr); | |
4192 | } | |
4193 | seq_putc(m, '\n'); | |
4194 | return 0; | |
4195 | } | |
4196 | #endif /* CONFIG_NUMA */ | |
4197 | ||
4198 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, | |
4199 | struct cgroup_map_cb *cb) | |
4200 | { | |
4201 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
4202 | struct mcs_total_stat mystat; | |
4203 | int i; | |
4204 | ||
4205 | memset(&mystat, 0, sizeof(mystat)); | |
4206 | mem_cgroup_get_local_stat(memcg, &mystat); | |
4207 | ||
4208 | ||
4209 | for (i = 0; i < NR_MCS_STAT; i++) { | |
4210 | if (i == MCS_SWAP && !do_swap_account) | |
4211 | continue; | |
4212 | cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); | |
4213 | } | |
4214 | ||
4215 | /* Hierarchical information */ | |
4216 | { | |
4217 | unsigned long long limit, memsw_limit; | |
4218 | memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit); | |
4219 | cb->fill(cb, "hierarchical_memory_limit", limit); | |
4220 | if (do_swap_account) | |
4221 | cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); | |
4222 | } | |
4223 | ||
4224 | memset(&mystat, 0, sizeof(mystat)); | |
4225 | mem_cgroup_get_total_stat(memcg, &mystat); | |
4226 | for (i = 0; i < NR_MCS_STAT; i++) { | |
4227 | if (i == MCS_SWAP && !do_swap_account) | |
4228 | continue; | |
4229 | cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); | |
4230 | } | |
4231 | ||
4232 | #ifdef CONFIG_DEBUG_VM | |
4233 | { | |
4234 | int nid, zid; | |
4235 | struct mem_cgroup_per_zone *mz; | |
4236 | unsigned long recent_rotated[2] = {0, 0}; | |
4237 | unsigned long recent_scanned[2] = {0, 0}; | |
4238 | ||
4239 | for_each_online_node(nid) | |
4240 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
4241 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
4242 | ||
4243 | recent_rotated[0] += | |
4244 | mz->reclaim_stat.recent_rotated[0]; | |
4245 | recent_rotated[1] += | |
4246 | mz->reclaim_stat.recent_rotated[1]; | |
4247 | recent_scanned[0] += | |
4248 | mz->reclaim_stat.recent_scanned[0]; | |
4249 | recent_scanned[1] += | |
4250 | mz->reclaim_stat.recent_scanned[1]; | |
4251 | } | |
4252 | cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); | |
4253 | cb->fill(cb, "recent_rotated_file", recent_rotated[1]); | |
4254 | cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); | |
4255 | cb->fill(cb, "recent_scanned_file", recent_scanned[1]); | |
4256 | } | |
4257 | #endif | |
4258 | ||
4259 | return 0; | |
4260 | } | |
4261 | ||
4262 | static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) | |
4263 | { | |
4264 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4265 | ||
4266 | return mem_cgroup_swappiness(memcg); | |
4267 | } | |
4268 | ||
4269 | static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, | |
4270 | u64 val) | |
4271 | { | |
4272 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4273 | struct mem_cgroup *parent; | |
4274 | ||
4275 | if (val > 100) | |
4276 | return -EINVAL; | |
4277 | ||
4278 | if (cgrp->parent == NULL) | |
4279 | return -EINVAL; | |
4280 | ||
4281 | parent = mem_cgroup_from_cont(cgrp->parent); | |
4282 | ||
4283 | cgroup_lock(); | |
4284 | ||
4285 | /* If under hierarchy, only empty-root can set this value */ | |
4286 | if ((parent->use_hierarchy) || | |
4287 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { | |
4288 | cgroup_unlock(); | |
4289 | return -EINVAL; | |
4290 | } | |
4291 | ||
4292 | memcg->swappiness = val; | |
4293 | ||
4294 | cgroup_unlock(); | |
4295 | ||
4296 | return 0; | |
4297 | } | |
4298 | ||
4299 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) | |
4300 | { | |
4301 | struct mem_cgroup_threshold_ary *t; | |
4302 | u64 usage; | |
4303 | int i; | |
4304 | ||
4305 | rcu_read_lock(); | |
4306 | if (!swap) | |
4307 | t = rcu_dereference(memcg->thresholds.primary); | |
4308 | else | |
4309 | t = rcu_dereference(memcg->memsw_thresholds.primary); | |
4310 | ||
4311 | if (!t) | |
4312 | goto unlock; | |
4313 | ||
4314 | usage = mem_cgroup_usage(memcg, swap); | |
4315 | ||
4316 | /* | |
4317 | * current_threshold points to threshold just below usage. | |
4318 | * If it's not true, a threshold was crossed after last | |
4319 | * call of __mem_cgroup_threshold(). | |
4320 | */ | |
4321 | i = t->current_threshold; | |
4322 | ||
4323 | /* | |
4324 | * Iterate backward over array of thresholds starting from | |
4325 | * current_threshold and check if a threshold is crossed. | |
4326 | * If none of thresholds below usage is crossed, we read | |
4327 | * only one element of the array here. | |
4328 | */ | |
4329 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) | |
4330 | eventfd_signal(t->entries[i].eventfd, 1); | |
4331 | ||
4332 | /* i = current_threshold + 1 */ | |
4333 | i++; | |
4334 | ||
4335 | /* | |
4336 | * Iterate forward over array of thresholds starting from | |
4337 | * current_threshold+1 and check if a threshold is crossed. | |
4338 | * If none of thresholds above usage is crossed, we read | |
4339 | * only one element of the array here. | |
4340 | */ | |
4341 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) | |
4342 | eventfd_signal(t->entries[i].eventfd, 1); | |
4343 | ||
4344 | /* Update current_threshold */ | |
4345 | t->current_threshold = i - 1; | |
4346 | unlock: | |
4347 | rcu_read_unlock(); | |
4348 | } | |
4349 | ||
4350 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) | |
4351 | { | |
4352 | while (memcg) { | |
4353 | __mem_cgroup_threshold(memcg, false); | |
4354 | if (do_swap_account) | |
4355 | __mem_cgroup_threshold(memcg, true); | |
4356 | ||
4357 | memcg = parent_mem_cgroup(memcg); | |
4358 | } | |
4359 | } | |
4360 | ||
4361 | static int compare_thresholds(const void *a, const void *b) | |
4362 | { | |
4363 | const struct mem_cgroup_threshold *_a = a; | |
4364 | const struct mem_cgroup_threshold *_b = b; | |
4365 | ||
4366 | return _a->threshold - _b->threshold; | |
4367 | } | |
4368 | ||
4369 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) | |
4370 | { | |
4371 | struct mem_cgroup_eventfd_list *ev; | |
4372 | ||
4373 | list_for_each_entry(ev, &memcg->oom_notify, list) | |
4374 | eventfd_signal(ev->eventfd, 1); | |
4375 | return 0; | |
4376 | } | |
4377 | ||
4378 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) | |
4379 | { | |
4380 | struct mem_cgroup *iter; | |
4381 | ||
4382 | for_each_mem_cgroup_tree(iter, memcg) | |
4383 | mem_cgroup_oom_notify_cb(iter); | |
4384 | } | |
4385 | ||
4386 | static int mem_cgroup_usage_register_event(struct cgroup *cgrp, | |
4387 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) | |
4388 | { | |
4389 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4390 | struct mem_cgroup_thresholds *thresholds; | |
4391 | struct mem_cgroup_threshold_ary *new; | |
4392 | int type = MEMFILE_TYPE(cft->private); | |
4393 | u64 threshold, usage; | |
4394 | int i, size, ret; | |
4395 | ||
4396 | ret = res_counter_memparse_write_strategy(args, &threshold); | |
4397 | if (ret) | |
4398 | return ret; | |
4399 | ||
4400 | mutex_lock(&memcg->thresholds_lock); | |
4401 | ||
4402 | if (type == _MEM) | |
4403 | thresholds = &memcg->thresholds; | |
4404 | else if (type == _MEMSWAP) | |
4405 | thresholds = &memcg->memsw_thresholds; | |
4406 | else | |
4407 | BUG(); | |
4408 | ||
4409 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
4410 | ||
4411 | /* Check if a threshold crossed before adding a new one */ | |
4412 | if (thresholds->primary) | |
4413 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
4414 | ||
4415 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; | |
4416 | ||
4417 | /* Allocate memory for new array of thresholds */ | |
4418 | new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), | |
4419 | GFP_KERNEL); | |
4420 | if (!new) { | |
4421 | ret = -ENOMEM; | |
4422 | goto unlock; | |
4423 | } | |
4424 | new->size = size; | |
4425 | ||
4426 | /* Copy thresholds (if any) to new array */ | |
4427 | if (thresholds->primary) { | |
4428 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * | |
4429 | sizeof(struct mem_cgroup_threshold)); | |
4430 | } | |
4431 | ||
4432 | /* Add new threshold */ | |
4433 | new->entries[size - 1].eventfd = eventfd; | |
4434 | new->entries[size - 1].threshold = threshold; | |
4435 | ||
4436 | /* Sort thresholds. Registering of new threshold isn't time-critical */ | |
4437 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), | |
4438 | compare_thresholds, NULL); | |
4439 | ||
4440 | /* Find current threshold */ | |
4441 | new->current_threshold = -1; | |
4442 | for (i = 0; i < size; i++) { | |
4443 | if (new->entries[i].threshold < usage) { | |
4444 | /* | |
4445 | * new->current_threshold will not be used until | |
4446 | * rcu_assign_pointer(), so it's safe to increment | |
4447 | * it here. | |
4448 | */ | |
4449 | ++new->current_threshold; | |
4450 | } | |
4451 | } | |
4452 | ||
4453 | /* Free old spare buffer and save old primary buffer as spare */ | |
4454 | kfree(thresholds->spare); | |
4455 | thresholds->spare = thresholds->primary; | |
4456 | ||
4457 | rcu_assign_pointer(thresholds->primary, new); | |
4458 | ||
4459 | /* To be sure that nobody uses thresholds */ | |
4460 | synchronize_rcu(); | |
4461 | ||
4462 | unlock: | |
4463 | mutex_unlock(&memcg->thresholds_lock); | |
4464 | ||
4465 | return ret; | |
4466 | } | |
4467 | ||
4468 | static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, | |
4469 | struct cftype *cft, struct eventfd_ctx *eventfd) | |
4470 | { | |
4471 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4472 | struct mem_cgroup_thresholds *thresholds; | |
4473 | struct mem_cgroup_threshold_ary *new; | |
4474 | int type = MEMFILE_TYPE(cft->private); | |
4475 | u64 usage; | |
4476 | int i, j, size; | |
4477 | ||
4478 | mutex_lock(&memcg->thresholds_lock); | |
4479 | if (type == _MEM) | |
4480 | thresholds = &memcg->thresholds; | |
4481 | else if (type == _MEMSWAP) | |
4482 | thresholds = &memcg->memsw_thresholds; | |
4483 | else | |
4484 | BUG(); | |
4485 | ||
4486 | if (!thresholds->primary) | |
4487 | goto unlock; | |
4488 | ||
4489 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
4490 | ||
4491 | /* Check if a threshold crossed before removing */ | |
4492 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
4493 | ||
4494 | /* Calculate new number of threshold */ | |
4495 | size = 0; | |
4496 | for (i = 0; i < thresholds->primary->size; i++) { | |
4497 | if (thresholds->primary->entries[i].eventfd != eventfd) | |
4498 | size++; | |
4499 | } | |
4500 | ||
4501 | new = thresholds->spare; | |
4502 | ||
4503 | /* Set thresholds array to NULL if we don't have thresholds */ | |
4504 | if (!size) { | |
4505 | kfree(new); | |
4506 | new = NULL; | |
4507 | goto swap_buffers; | |
4508 | } | |
4509 | ||
4510 | new->size = size; | |
4511 | ||
4512 | /* Copy thresholds and find current threshold */ | |
4513 | new->current_threshold = -1; | |
4514 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { | |
4515 | if (thresholds->primary->entries[i].eventfd == eventfd) | |
4516 | continue; | |
4517 | ||
4518 | new->entries[j] = thresholds->primary->entries[i]; | |
4519 | if (new->entries[j].threshold < usage) { | |
4520 | /* | |
4521 | * new->current_threshold will not be used | |
4522 | * until rcu_assign_pointer(), so it's safe to increment | |
4523 | * it here. | |
4524 | */ | |
4525 | ++new->current_threshold; | |
4526 | } | |
4527 | j++; | |
4528 | } | |
4529 | ||
4530 | swap_buffers: | |
4531 | /* Swap primary and spare array */ | |
4532 | thresholds->spare = thresholds->primary; | |
4533 | /* If all events are unregistered, free the spare array */ | |
4534 | if (!new) { | |
4535 | kfree(thresholds->spare); | |
4536 | thresholds->spare = NULL; | |
4537 | } | |
4538 | ||
4539 | rcu_assign_pointer(thresholds->primary, new); | |
4540 | ||
4541 | /* To be sure that nobody uses thresholds */ | |
4542 | synchronize_rcu(); | |
4543 | unlock: | |
4544 | mutex_unlock(&memcg->thresholds_lock); | |
4545 | } | |
4546 | ||
4547 | static int mem_cgroup_oom_register_event(struct cgroup *cgrp, | |
4548 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) | |
4549 | { | |
4550 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4551 | struct mem_cgroup_eventfd_list *event; | |
4552 | int type = MEMFILE_TYPE(cft->private); | |
4553 | ||
4554 | BUG_ON(type != _OOM_TYPE); | |
4555 | event = kmalloc(sizeof(*event), GFP_KERNEL); | |
4556 | if (!event) | |
4557 | return -ENOMEM; | |
4558 | ||
4559 | spin_lock(&memcg_oom_lock); | |
4560 | ||
4561 | event->eventfd = eventfd; | |
4562 | list_add(&event->list, &memcg->oom_notify); | |
4563 | ||
4564 | /* already in OOM ? */ | |
4565 | if (atomic_read(&memcg->under_oom)) | |
4566 | eventfd_signal(eventfd, 1); | |
4567 | spin_unlock(&memcg_oom_lock); | |
4568 | ||
4569 | return 0; | |
4570 | } | |
4571 | ||
4572 | static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, | |
4573 | struct cftype *cft, struct eventfd_ctx *eventfd) | |
4574 | { | |
4575 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4576 | struct mem_cgroup_eventfd_list *ev, *tmp; | |
4577 | int type = MEMFILE_TYPE(cft->private); | |
4578 | ||
4579 | BUG_ON(type != _OOM_TYPE); | |
4580 | ||
4581 | spin_lock(&memcg_oom_lock); | |
4582 | ||
4583 | list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { | |
4584 | if (ev->eventfd == eventfd) { | |
4585 | list_del(&ev->list); | |
4586 | kfree(ev); | |
4587 | } | |
4588 | } | |
4589 | ||
4590 | spin_unlock(&memcg_oom_lock); | |
4591 | } | |
4592 | ||
4593 | static int mem_cgroup_oom_control_read(struct cgroup *cgrp, | |
4594 | struct cftype *cft, struct cgroup_map_cb *cb) | |
4595 | { | |
4596 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4597 | ||
4598 | cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); | |
4599 | ||
4600 | if (atomic_read(&memcg->under_oom)) | |
4601 | cb->fill(cb, "under_oom", 1); | |
4602 | else | |
4603 | cb->fill(cb, "under_oom", 0); | |
4604 | return 0; | |
4605 | } | |
4606 | ||
4607 | static int mem_cgroup_oom_control_write(struct cgroup *cgrp, | |
4608 | struct cftype *cft, u64 val) | |
4609 | { | |
4610 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4611 | struct mem_cgroup *parent; | |
4612 | ||
4613 | /* cannot set to root cgroup and only 0 and 1 are allowed */ | |
4614 | if (!cgrp->parent || !((val == 0) || (val == 1))) | |
4615 | return -EINVAL; | |
4616 | ||
4617 | parent = mem_cgroup_from_cont(cgrp->parent); | |
4618 | ||
4619 | cgroup_lock(); | |
4620 | /* oom-kill-disable is a flag for subhierarchy. */ | |
4621 | if ((parent->use_hierarchy) || | |
4622 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { | |
4623 | cgroup_unlock(); | |
4624 | return -EINVAL; | |
4625 | } | |
4626 | memcg->oom_kill_disable = val; | |
4627 | if (!val) | |
4628 | memcg_oom_recover(memcg); | |
4629 | cgroup_unlock(); | |
4630 | return 0; | |
4631 | } | |
4632 | ||
4633 | #ifdef CONFIG_NUMA | |
4634 | static const struct file_operations mem_control_numa_stat_file_operations = { | |
4635 | .read = seq_read, | |
4636 | .llseek = seq_lseek, | |
4637 | .release = single_release, | |
4638 | }; | |
4639 | ||
4640 | static int mem_control_numa_stat_open(struct inode *unused, struct file *file) | |
4641 | { | |
4642 | struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; | |
4643 | ||
4644 | file->f_op = &mem_control_numa_stat_file_operations; | |
4645 | return single_open(file, mem_control_numa_stat_show, cont); | |
4646 | } | |
4647 | #endif /* CONFIG_NUMA */ | |
4648 | ||
4649 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM | |
4650 | static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss) | |
4651 | { | |
4652 | return mem_cgroup_sockets_init(memcg, ss); | |
4653 | }; | |
4654 | ||
4655 | static void kmem_cgroup_destroy(struct mem_cgroup *memcg) | |
4656 | { | |
4657 | mem_cgroup_sockets_destroy(memcg); | |
4658 | } | |
4659 | #else | |
4660 | static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss) | |
4661 | { | |
4662 | return 0; | |
4663 | } | |
4664 | ||
4665 | static void kmem_cgroup_destroy(struct mem_cgroup *memcg) | |
4666 | { | |
4667 | } | |
4668 | #endif | |
4669 | ||
4670 | static struct cftype mem_cgroup_files[] = { | |
4671 | { | |
4672 | .name = "usage_in_bytes", | |
4673 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), | |
4674 | .read = mem_cgroup_read, | |
4675 | .register_event = mem_cgroup_usage_register_event, | |
4676 | .unregister_event = mem_cgroup_usage_unregister_event, | |
4677 | }, | |
4678 | { | |
4679 | .name = "max_usage_in_bytes", | |
4680 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), | |
4681 | .trigger = mem_cgroup_reset, | |
4682 | .read = mem_cgroup_read, | |
4683 | }, | |
4684 | { | |
4685 | .name = "limit_in_bytes", | |
4686 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), | |
4687 | .write_string = mem_cgroup_write, | |
4688 | .read = mem_cgroup_read, | |
4689 | }, | |
4690 | { | |
4691 | .name = "soft_limit_in_bytes", | |
4692 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), | |
4693 | .write_string = mem_cgroup_write, | |
4694 | .read = mem_cgroup_read, | |
4695 | }, | |
4696 | { | |
4697 | .name = "failcnt", | |
4698 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), | |
4699 | .trigger = mem_cgroup_reset, | |
4700 | .read = mem_cgroup_read, | |
4701 | }, | |
4702 | { | |
4703 | .name = "stat", | |
4704 | .read_map = mem_control_stat_show, | |
4705 | }, | |
4706 | { | |
4707 | .name = "force_empty", | |
4708 | .trigger = mem_cgroup_force_empty_write, | |
4709 | }, | |
4710 | { | |
4711 | .name = "use_hierarchy", | |
4712 | .write_u64 = mem_cgroup_hierarchy_write, | |
4713 | .read_u64 = mem_cgroup_hierarchy_read, | |
4714 | }, | |
4715 | { | |
4716 | .name = "swappiness", | |
4717 | .read_u64 = mem_cgroup_swappiness_read, | |
4718 | .write_u64 = mem_cgroup_swappiness_write, | |
4719 | }, | |
4720 | { | |
4721 | .name = "move_charge_at_immigrate", | |
4722 | .read_u64 = mem_cgroup_move_charge_read, | |
4723 | .write_u64 = mem_cgroup_move_charge_write, | |
4724 | }, | |
4725 | { | |
4726 | .name = "oom_control", | |
4727 | .read_map = mem_cgroup_oom_control_read, | |
4728 | .write_u64 = mem_cgroup_oom_control_write, | |
4729 | .register_event = mem_cgroup_oom_register_event, | |
4730 | .unregister_event = mem_cgroup_oom_unregister_event, | |
4731 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), | |
4732 | }, | |
4733 | #ifdef CONFIG_NUMA | |
4734 | { | |
4735 | .name = "numa_stat", | |
4736 | .open = mem_control_numa_stat_open, | |
4737 | .mode = S_IRUGO, | |
4738 | }, | |
4739 | #endif | |
4740 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
4741 | { | |
4742 | .name = "memsw.usage_in_bytes", | |
4743 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), | |
4744 | .read = mem_cgroup_read, | |
4745 | .register_event = mem_cgroup_usage_register_event, | |
4746 | .unregister_event = mem_cgroup_usage_unregister_event, | |
4747 | }, | |
4748 | { | |
4749 | .name = "memsw.max_usage_in_bytes", | |
4750 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), | |
4751 | .trigger = mem_cgroup_reset, | |
4752 | .read = mem_cgroup_read, | |
4753 | }, | |
4754 | { | |
4755 | .name = "memsw.limit_in_bytes", | |
4756 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), | |
4757 | .write_string = mem_cgroup_write, | |
4758 | .read = mem_cgroup_read, | |
4759 | }, | |
4760 | { | |
4761 | .name = "memsw.failcnt", | |
4762 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), | |
4763 | .trigger = mem_cgroup_reset, | |
4764 | .read = mem_cgroup_read, | |
4765 | }, | |
4766 | #endif | |
4767 | { }, /* terminate */ | |
4768 | }; | |
4769 | ||
4770 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) | |
4771 | { | |
4772 | struct mem_cgroup_per_node *pn; | |
4773 | struct mem_cgroup_per_zone *mz; | |
4774 | enum lru_list lru; | |
4775 | int zone, tmp = node; | |
4776 | /* | |
4777 | * This routine is called against possible nodes. | |
4778 | * But it's BUG to call kmalloc() against offline node. | |
4779 | * | |
4780 | * TODO: this routine can waste much memory for nodes which will | |
4781 | * never be onlined. It's better to use memory hotplug callback | |
4782 | * function. | |
4783 | */ | |
4784 | if (!node_state(node, N_NORMAL_MEMORY)) | |
4785 | tmp = -1; | |
4786 | pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); | |
4787 | if (!pn) | |
4788 | return 1; | |
4789 | ||
4790 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4791 | mz = &pn->zoneinfo[zone]; | |
4792 | for_each_lru(lru) | |
4793 | INIT_LIST_HEAD(&mz->lruvec.lists[lru]); | |
4794 | mz->usage_in_excess = 0; | |
4795 | mz->on_tree = false; | |
4796 | mz->memcg = memcg; | |
4797 | } | |
4798 | memcg->info.nodeinfo[node] = pn; | |
4799 | return 0; | |
4800 | } | |
4801 | ||
4802 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) | |
4803 | { | |
4804 | kfree(memcg->info.nodeinfo[node]); | |
4805 | } | |
4806 | ||
4807 | static struct mem_cgroup *mem_cgroup_alloc(void) | |
4808 | { | |
4809 | struct mem_cgroup *memcg; | |
4810 | int size = sizeof(struct mem_cgroup); | |
4811 | ||
4812 | /* Can be very big if MAX_NUMNODES is very big */ | |
4813 | if (size < PAGE_SIZE) | |
4814 | memcg = kzalloc(size, GFP_KERNEL); | |
4815 | else | |
4816 | memcg = vzalloc(size); | |
4817 | ||
4818 | if (!memcg) | |
4819 | return NULL; | |
4820 | ||
4821 | memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu); | |
4822 | if (!memcg->stat) | |
4823 | goto out_free; | |
4824 | spin_lock_init(&memcg->pcp_counter_lock); | |
4825 | return memcg; | |
4826 | ||
4827 | out_free: | |
4828 | if (size < PAGE_SIZE) | |
4829 | kfree(memcg); | |
4830 | else | |
4831 | vfree(memcg); | |
4832 | return NULL; | |
4833 | } | |
4834 | ||
4835 | /* | |
4836 | * Helpers for freeing a vzalloc()ed mem_cgroup by RCU, | |
4837 | * but in process context. The work_freeing structure is overlaid | |
4838 | * on the rcu_freeing structure, which itself is overlaid on memsw. | |
4839 | */ | |
4840 | static void vfree_work(struct work_struct *work) | |
4841 | { | |
4842 | struct mem_cgroup *memcg; | |
4843 | ||
4844 | memcg = container_of(work, struct mem_cgroup, work_freeing); | |
4845 | vfree(memcg); | |
4846 | } | |
4847 | static void vfree_rcu(struct rcu_head *rcu_head) | |
4848 | { | |
4849 | struct mem_cgroup *memcg; | |
4850 | ||
4851 | memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing); | |
4852 | INIT_WORK(&memcg->work_freeing, vfree_work); | |
4853 | schedule_work(&memcg->work_freeing); | |
4854 | } | |
4855 | ||
4856 | /* | |
4857 | * At destroying mem_cgroup, references from swap_cgroup can remain. | |
4858 | * (scanning all at force_empty is too costly...) | |
4859 | * | |
4860 | * Instead of clearing all references at force_empty, we remember | |
4861 | * the number of reference from swap_cgroup and free mem_cgroup when | |
4862 | * it goes down to 0. | |
4863 | * | |
4864 | * Removal of cgroup itself succeeds regardless of refs from swap. | |
4865 | */ | |
4866 | ||
4867 | static void __mem_cgroup_free(struct mem_cgroup *memcg) | |
4868 | { | |
4869 | int node; | |
4870 | ||
4871 | mem_cgroup_remove_from_trees(memcg); | |
4872 | free_css_id(&mem_cgroup_subsys, &memcg->css); | |
4873 | ||
4874 | for_each_node(node) | |
4875 | free_mem_cgroup_per_zone_info(memcg, node); | |
4876 | ||
4877 | free_percpu(memcg->stat); | |
4878 | if (sizeof(struct mem_cgroup) < PAGE_SIZE) | |
4879 | kfree_rcu(memcg, rcu_freeing); | |
4880 | else | |
4881 | call_rcu(&memcg->rcu_freeing, vfree_rcu); | |
4882 | } | |
4883 | ||
4884 | static void mem_cgroup_get(struct mem_cgroup *memcg) | |
4885 | { | |
4886 | atomic_inc(&memcg->refcnt); | |
4887 | } | |
4888 | ||
4889 | static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) | |
4890 | { | |
4891 | if (atomic_sub_and_test(count, &memcg->refcnt)) { | |
4892 | struct mem_cgroup *parent = parent_mem_cgroup(memcg); | |
4893 | __mem_cgroup_free(memcg); | |
4894 | if (parent) | |
4895 | mem_cgroup_put(parent); | |
4896 | } | |
4897 | } | |
4898 | ||
4899 | static void mem_cgroup_put(struct mem_cgroup *memcg) | |
4900 | { | |
4901 | __mem_cgroup_put(memcg, 1); | |
4902 | } | |
4903 | ||
4904 | /* | |
4905 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. | |
4906 | */ | |
4907 | struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) | |
4908 | { | |
4909 | if (!memcg->res.parent) | |
4910 | return NULL; | |
4911 | return mem_cgroup_from_res_counter(memcg->res.parent, res); | |
4912 | } | |
4913 | EXPORT_SYMBOL(parent_mem_cgroup); | |
4914 | ||
4915 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
4916 | static void __init enable_swap_cgroup(void) | |
4917 | { | |
4918 | if (!mem_cgroup_disabled() && really_do_swap_account) | |
4919 | do_swap_account = 1; | |
4920 | } | |
4921 | #else | |
4922 | static void __init enable_swap_cgroup(void) | |
4923 | { | |
4924 | } | |
4925 | #endif | |
4926 | ||
4927 | static int mem_cgroup_soft_limit_tree_init(void) | |
4928 | { | |
4929 | struct mem_cgroup_tree_per_node *rtpn; | |
4930 | struct mem_cgroup_tree_per_zone *rtpz; | |
4931 | int tmp, node, zone; | |
4932 | ||
4933 | for_each_node(node) { | |
4934 | tmp = node; | |
4935 | if (!node_state(node, N_NORMAL_MEMORY)) | |
4936 | tmp = -1; | |
4937 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); | |
4938 | if (!rtpn) | |
4939 | goto err_cleanup; | |
4940 | ||
4941 | soft_limit_tree.rb_tree_per_node[node] = rtpn; | |
4942 | ||
4943 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4944 | rtpz = &rtpn->rb_tree_per_zone[zone]; | |
4945 | rtpz->rb_root = RB_ROOT; | |
4946 | spin_lock_init(&rtpz->lock); | |
4947 | } | |
4948 | } | |
4949 | return 0; | |
4950 | ||
4951 | err_cleanup: | |
4952 | for_each_node(node) { | |
4953 | if (!soft_limit_tree.rb_tree_per_node[node]) | |
4954 | break; | |
4955 | kfree(soft_limit_tree.rb_tree_per_node[node]); | |
4956 | soft_limit_tree.rb_tree_per_node[node] = NULL; | |
4957 | } | |
4958 | return 1; | |
4959 | ||
4960 | } | |
4961 | ||
4962 | static struct cgroup_subsys_state * __ref | |
4963 | mem_cgroup_create(struct cgroup *cont) | |
4964 | { | |
4965 | struct mem_cgroup *memcg, *parent; | |
4966 | long error = -ENOMEM; | |
4967 | int node; | |
4968 | ||
4969 | memcg = mem_cgroup_alloc(); | |
4970 | if (!memcg) | |
4971 | return ERR_PTR(error); | |
4972 | ||
4973 | for_each_node(node) | |
4974 | if (alloc_mem_cgroup_per_zone_info(memcg, node)) | |
4975 | goto free_out; | |
4976 | ||
4977 | /* root ? */ | |
4978 | if (cont->parent == NULL) { | |
4979 | int cpu; | |
4980 | enable_swap_cgroup(); | |
4981 | parent = NULL; | |
4982 | if (mem_cgroup_soft_limit_tree_init()) | |
4983 | goto free_out; | |
4984 | root_mem_cgroup = memcg; | |
4985 | for_each_possible_cpu(cpu) { | |
4986 | struct memcg_stock_pcp *stock = | |
4987 | &per_cpu(memcg_stock, cpu); | |
4988 | INIT_WORK(&stock->work, drain_local_stock); | |
4989 | } | |
4990 | hotcpu_notifier(memcg_cpu_hotplug_callback, 0); | |
4991 | } else { | |
4992 | parent = mem_cgroup_from_cont(cont->parent); | |
4993 | memcg->use_hierarchy = parent->use_hierarchy; | |
4994 | memcg->oom_kill_disable = parent->oom_kill_disable; | |
4995 | } | |
4996 | ||
4997 | if (parent && parent->use_hierarchy) { | |
4998 | res_counter_init(&memcg->res, &parent->res); | |
4999 | res_counter_init(&memcg->memsw, &parent->memsw); | |
5000 | /* | |
5001 | * We increment refcnt of the parent to ensure that we can | |
5002 | * safely access it on res_counter_charge/uncharge. | |
5003 | * This refcnt will be decremented when freeing this | |
5004 | * mem_cgroup(see mem_cgroup_put). | |
5005 | */ | |
5006 | mem_cgroup_get(parent); | |
5007 | } else { | |
5008 | res_counter_init(&memcg->res, NULL); | |
5009 | res_counter_init(&memcg->memsw, NULL); | |
5010 | } | |
5011 | memcg->last_scanned_node = MAX_NUMNODES; | |
5012 | INIT_LIST_HEAD(&memcg->oom_notify); | |
5013 | ||
5014 | if (parent) | |
5015 | memcg->swappiness = mem_cgroup_swappiness(parent); | |
5016 | atomic_set(&memcg->refcnt, 1); | |
5017 | memcg->move_charge_at_immigrate = 0; | |
5018 | mutex_init(&memcg->thresholds_lock); | |
5019 | spin_lock_init(&memcg->move_lock); | |
5020 | ||
5021 | error = memcg_init_kmem(memcg, &mem_cgroup_subsys); | |
5022 | if (error) { | |
5023 | /* | |
5024 | * We call put now because our (and parent's) refcnts | |
5025 | * are already in place. mem_cgroup_put() will internally | |
5026 | * call __mem_cgroup_free, so return directly | |
5027 | */ | |
5028 | mem_cgroup_put(memcg); | |
5029 | return ERR_PTR(error); | |
5030 | } | |
5031 | return &memcg->css; | |
5032 | free_out: | |
5033 | __mem_cgroup_free(memcg); | |
5034 | return ERR_PTR(error); | |
5035 | } | |
5036 | ||
5037 | static int mem_cgroup_pre_destroy(struct cgroup *cont) | |
5038 | { | |
5039 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
5040 | ||
5041 | return mem_cgroup_force_empty(memcg, false); | |
5042 | } | |
5043 | ||
5044 | static void mem_cgroup_destroy(struct cgroup *cont) | |
5045 | { | |
5046 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
5047 | ||
5048 | kmem_cgroup_destroy(memcg); | |
5049 | ||
5050 | mem_cgroup_put(memcg); | |
5051 | } | |
5052 | ||
5053 | #ifdef CONFIG_MMU | |
5054 | /* Handlers for move charge at task migration. */ | |
5055 | #define PRECHARGE_COUNT_AT_ONCE 256 | |
5056 | static int mem_cgroup_do_precharge(unsigned long count) | |
5057 | { | |
5058 | int ret = 0; | |
5059 | int batch_count = PRECHARGE_COUNT_AT_ONCE; | |
5060 | struct mem_cgroup *memcg = mc.to; | |
5061 | ||
5062 | if (mem_cgroup_is_root(memcg)) { | |
5063 | mc.precharge += count; | |
5064 | /* we don't need css_get for root */ | |
5065 | return ret; | |
5066 | } | |
5067 | /* try to charge at once */ | |
5068 | if (count > 1) { | |
5069 | struct res_counter *dummy; | |
5070 | /* | |
5071 | * "memcg" cannot be under rmdir() because we've already checked | |
5072 | * by cgroup_lock_live_cgroup() that it is not removed and we | |
5073 | * are still under the same cgroup_mutex. So we can postpone | |
5074 | * css_get(). | |
5075 | */ | |
5076 | if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) | |
5077 | goto one_by_one; | |
5078 | if (do_swap_account && res_counter_charge(&memcg->memsw, | |
5079 | PAGE_SIZE * count, &dummy)) { | |
5080 | res_counter_uncharge(&memcg->res, PAGE_SIZE * count); | |
5081 | goto one_by_one; | |
5082 | } | |
5083 | mc.precharge += count; | |
5084 | return ret; | |
5085 | } | |
5086 | one_by_one: | |
5087 | /* fall back to one by one charge */ | |
5088 | while (count--) { | |
5089 | if (signal_pending(current)) { | |
5090 | ret = -EINTR; | |
5091 | break; | |
5092 | } | |
5093 | if (!batch_count--) { | |
5094 | batch_count = PRECHARGE_COUNT_AT_ONCE; | |
5095 | cond_resched(); | |
5096 | } | |
5097 | ret = __mem_cgroup_try_charge(NULL, | |
5098 | GFP_KERNEL, 1, &memcg, false); | |
5099 | if (ret) | |
5100 | /* mem_cgroup_clear_mc() will do uncharge later */ | |
5101 | return ret; | |
5102 | mc.precharge++; | |
5103 | } | |
5104 | return ret; | |
5105 | } | |
5106 | ||
5107 | /** | |
5108 | * get_mctgt_type - get target type of moving charge | |
5109 | * @vma: the vma the pte to be checked belongs | |
5110 | * @addr: the address corresponding to the pte to be checked | |
5111 | * @ptent: the pte to be checked | |
5112 | * @target: the pointer the target page or swap ent will be stored(can be NULL) | |
5113 | * | |
5114 | * Returns | |
5115 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. | |
5116 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for | |
5117 | * move charge. if @target is not NULL, the page is stored in target->page | |
5118 | * with extra refcnt got(Callers should handle it). | |
5119 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a | |
5120 | * target for charge migration. if @target is not NULL, the entry is stored | |
5121 | * in target->ent. | |
5122 | * | |
5123 | * Called with pte lock held. | |
5124 | */ | |
5125 | union mc_target { | |
5126 | struct page *page; | |
5127 | swp_entry_t ent; | |
5128 | }; | |
5129 | ||
5130 | enum mc_target_type { | |
5131 | MC_TARGET_NONE = 0, | |
5132 | MC_TARGET_PAGE, | |
5133 | MC_TARGET_SWAP, | |
5134 | }; | |
5135 | ||
5136 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, | |
5137 | unsigned long addr, pte_t ptent) | |
5138 | { | |
5139 | struct page *page = vm_normal_page(vma, addr, ptent); | |
5140 | ||
5141 | if (!page || !page_mapped(page)) | |
5142 | return NULL; | |
5143 | if (PageAnon(page)) { | |
5144 | /* we don't move shared anon */ | |
5145 | if (!move_anon()) | |
5146 | return NULL; | |
5147 | } else if (!move_file()) | |
5148 | /* we ignore mapcount for file pages */ | |
5149 | return NULL; | |
5150 | if (!get_page_unless_zero(page)) | |
5151 | return NULL; | |
5152 | ||
5153 | return page; | |
5154 | } | |
5155 | ||
5156 | #ifdef CONFIG_SWAP | |
5157 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
5158 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
5159 | { | |
5160 | struct page *page = NULL; | |
5161 | swp_entry_t ent = pte_to_swp_entry(ptent); | |
5162 | ||
5163 | if (!move_anon() || non_swap_entry(ent)) | |
5164 | return NULL; | |
5165 | /* | |
5166 | * Because lookup_swap_cache() updates some statistics counter, | |
5167 | * we call find_get_page() with swapper_space directly. | |
5168 | */ | |
5169 | page = find_get_page(&swapper_space, ent.val); | |
5170 | if (do_swap_account) | |
5171 | entry->val = ent.val; | |
5172 | ||
5173 | return page; | |
5174 | } | |
5175 | #else | |
5176 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
5177 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
5178 | { | |
5179 | return NULL; | |
5180 | } | |
5181 | #endif | |
5182 | ||
5183 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, | |
5184 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
5185 | { | |
5186 | struct page *page = NULL; | |
5187 | struct inode *inode; | |
5188 | struct address_space *mapping; | |
5189 | pgoff_t pgoff; | |
5190 | ||
5191 | if (!vma->vm_file) /* anonymous vma */ | |
5192 | return NULL; | |
5193 | if (!move_file()) | |
5194 | return NULL; | |
5195 | ||
5196 | inode = vma->vm_file->f_path.dentry->d_inode; | |
5197 | mapping = vma->vm_file->f_mapping; | |
5198 | if (pte_none(ptent)) | |
5199 | pgoff = linear_page_index(vma, addr); | |
5200 | else /* pte_file(ptent) is true */ | |
5201 | pgoff = pte_to_pgoff(ptent); | |
5202 | ||
5203 | /* page is moved even if it's not RSS of this task(page-faulted). */ | |
5204 | page = find_get_page(mapping, pgoff); | |
5205 | ||
5206 | #ifdef CONFIG_SWAP | |
5207 | /* shmem/tmpfs may report page out on swap: account for that too. */ | |
5208 | if (radix_tree_exceptional_entry(page)) { | |
5209 | swp_entry_t swap = radix_to_swp_entry(page); | |
5210 | if (do_swap_account) | |
5211 | *entry = swap; | |
5212 | page = find_get_page(&swapper_space, swap.val); | |
5213 | } | |
5214 | #endif | |
5215 | return page; | |
5216 | } | |
5217 | ||
5218 | static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma, | |
5219 | unsigned long addr, pte_t ptent, union mc_target *target) | |
5220 | { | |
5221 | struct page *page = NULL; | |
5222 | struct page_cgroup *pc; | |
5223 | enum mc_target_type ret = MC_TARGET_NONE; | |
5224 | swp_entry_t ent = { .val = 0 }; | |
5225 | ||
5226 | if (pte_present(ptent)) | |
5227 | page = mc_handle_present_pte(vma, addr, ptent); | |
5228 | else if (is_swap_pte(ptent)) | |
5229 | page = mc_handle_swap_pte(vma, addr, ptent, &ent); | |
5230 | else if (pte_none(ptent) || pte_file(ptent)) | |
5231 | page = mc_handle_file_pte(vma, addr, ptent, &ent); | |
5232 | ||
5233 | if (!page && !ent.val) | |
5234 | return ret; | |
5235 | if (page) { | |
5236 | pc = lookup_page_cgroup(page); | |
5237 | /* | |
5238 | * Do only loose check w/o page_cgroup lock. | |
5239 | * mem_cgroup_move_account() checks the pc is valid or not under | |
5240 | * the lock. | |
5241 | */ | |
5242 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { | |
5243 | ret = MC_TARGET_PAGE; | |
5244 | if (target) | |
5245 | target->page = page; | |
5246 | } | |
5247 | if (!ret || !target) | |
5248 | put_page(page); | |
5249 | } | |
5250 | /* There is a swap entry and a page doesn't exist or isn't charged */ | |
5251 | if (ent.val && !ret && | |
5252 | css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) { | |
5253 | ret = MC_TARGET_SWAP; | |
5254 | if (target) | |
5255 | target->ent = ent; | |
5256 | } | |
5257 | return ret; | |
5258 | } | |
5259 | ||
5260 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
5261 | /* | |
5262 | * We don't consider swapping or file mapped pages because THP does not | |
5263 | * support them for now. | |
5264 | * Caller should make sure that pmd_trans_huge(pmd) is true. | |
5265 | */ | |
5266 | static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, | |
5267 | unsigned long addr, pmd_t pmd, union mc_target *target) | |
5268 | { | |
5269 | struct page *page = NULL; | |
5270 | struct page_cgroup *pc; | |
5271 | enum mc_target_type ret = MC_TARGET_NONE; | |
5272 | ||
5273 | page = pmd_page(pmd); | |
5274 | VM_BUG_ON(!page || !PageHead(page)); | |
5275 | if (!move_anon()) | |
5276 | return ret; | |
5277 | pc = lookup_page_cgroup(page); | |
5278 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { | |
5279 | ret = MC_TARGET_PAGE; | |
5280 | if (target) { | |
5281 | get_page(page); | |
5282 | target->page = page; | |
5283 | } | |
5284 | } | |
5285 | return ret; | |
5286 | } | |
5287 | #else | |
5288 | static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, | |
5289 | unsigned long addr, pmd_t pmd, union mc_target *target) | |
5290 | { | |
5291 | return MC_TARGET_NONE; | |
5292 | } | |
5293 | #endif | |
5294 | ||
5295 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, | |
5296 | unsigned long addr, unsigned long end, | |
5297 | struct mm_walk *walk) | |
5298 | { | |
5299 | struct vm_area_struct *vma = walk->private; | |
5300 | pte_t *pte; | |
5301 | spinlock_t *ptl; | |
5302 | ||
5303 | if (pmd_trans_huge_lock(pmd, vma) == 1) { | |
5304 | if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE) | |
5305 | mc.precharge += HPAGE_PMD_NR; | |
5306 | spin_unlock(&vma->vm_mm->page_table_lock); | |
5307 | return 0; | |
5308 | } | |
5309 | ||
5310 | if (pmd_trans_unstable(pmd)) | |
5311 | return 0; | |
5312 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
5313 | for (; addr != end; pte++, addr += PAGE_SIZE) | |
5314 | if (get_mctgt_type(vma, addr, *pte, NULL)) | |
5315 | mc.precharge++; /* increment precharge temporarily */ | |
5316 | pte_unmap_unlock(pte - 1, ptl); | |
5317 | cond_resched(); | |
5318 | ||
5319 | return 0; | |
5320 | } | |
5321 | ||
5322 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) | |
5323 | { | |
5324 | unsigned long precharge; | |
5325 | struct vm_area_struct *vma; | |
5326 | ||
5327 | down_read(&mm->mmap_sem); | |
5328 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
5329 | struct mm_walk mem_cgroup_count_precharge_walk = { | |
5330 | .pmd_entry = mem_cgroup_count_precharge_pte_range, | |
5331 | .mm = mm, | |
5332 | .private = vma, | |
5333 | }; | |
5334 | if (is_vm_hugetlb_page(vma)) | |
5335 | continue; | |
5336 | walk_page_range(vma->vm_start, vma->vm_end, | |
5337 | &mem_cgroup_count_precharge_walk); | |
5338 | } | |
5339 | up_read(&mm->mmap_sem); | |
5340 | ||
5341 | precharge = mc.precharge; | |
5342 | mc.precharge = 0; | |
5343 | ||
5344 | return precharge; | |
5345 | } | |
5346 | ||
5347 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) | |
5348 | { | |
5349 | unsigned long precharge = mem_cgroup_count_precharge(mm); | |
5350 | ||
5351 | VM_BUG_ON(mc.moving_task); | |
5352 | mc.moving_task = current; | |
5353 | return mem_cgroup_do_precharge(precharge); | |
5354 | } | |
5355 | ||
5356 | /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ | |
5357 | static void __mem_cgroup_clear_mc(void) | |
5358 | { | |
5359 | struct mem_cgroup *from = mc.from; | |
5360 | struct mem_cgroup *to = mc.to; | |
5361 | ||
5362 | /* we must uncharge all the leftover precharges from mc.to */ | |
5363 | if (mc.precharge) { | |
5364 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); | |
5365 | mc.precharge = 0; | |
5366 | } | |
5367 | /* | |
5368 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so | |
5369 | * we must uncharge here. | |
5370 | */ | |
5371 | if (mc.moved_charge) { | |
5372 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); | |
5373 | mc.moved_charge = 0; | |
5374 | } | |
5375 | /* we must fixup refcnts and charges */ | |
5376 | if (mc.moved_swap) { | |
5377 | /* uncharge swap account from the old cgroup */ | |
5378 | if (!mem_cgroup_is_root(mc.from)) | |
5379 | res_counter_uncharge(&mc.from->memsw, | |
5380 | PAGE_SIZE * mc.moved_swap); | |
5381 | __mem_cgroup_put(mc.from, mc.moved_swap); | |
5382 | ||
5383 | if (!mem_cgroup_is_root(mc.to)) { | |
5384 | /* | |
5385 | * we charged both to->res and to->memsw, so we should | |
5386 | * uncharge to->res. | |
5387 | */ | |
5388 | res_counter_uncharge(&mc.to->res, | |
5389 | PAGE_SIZE * mc.moved_swap); | |
5390 | } | |
5391 | /* we've already done mem_cgroup_get(mc.to) */ | |
5392 | mc.moved_swap = 0; | |
5393 | } | |
5394 | memcg_oom_recover(from); | |
5395 | memcg_oom_recover(to); | |
5396 | wake_up_all(&mc.waitq); | |
5397 | } | |
5398 | ||
5399 | static void mem_cgroup_clear_mc(void) | |
5400 | { | |
5401 | struct mem_cgroup *from = mc.from; | |
5402 | ||
5403 | /* | |
5404 | * we must clear moving_task before waking up waiters at the end of | |
5405 | * task migration. | |
5406 | */ | |
5407 | mc.moving_task = NULL; | |
5408 | __mem_cgroup_clear_mc(); | |
5409 | spin_lock(&mc.lock); | |
5410 | mc.from = NULL; | |
5411 | mc.to = NULL; | |
5412 | spin_unlock(&mc.lock); | |
5413 | mem_cgroup_end_move(from); | |
5414 | } | |
5415 | ||
5416 | static int mem_cgroup_can_attach(struct cgroup *cgroup, | |
5417 | struct cgroup_taskset *tset) | |
5418 | { | |
5419 | struct task_struct *p = cgroup_taskset_first(tset); | |
5420 | int ret = 0; | |
5421 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); | |
5422 | ||
5423 | if (memcg->move_charge_at_immigrate) { | |
5424 | struct mm_struct *mm; | |
5425 | struct mem_cgroup *from = mem_cgroup_from_task(p); | |
5426 | ||
5427 | VM_BUG_ON(from == memcg); | |
5428 | ||
5429 | mm = get_task_mm(p); | |
5430 | if (!mm) | |
5431 | return 0; | |
5432 | /* We move charges only when we move a owner of the mm */ | |
5433 | if (mm->owner == p) { | |
5434 | VM_BUG_ON(mc.from); | |
5435 | VM_BUG_ON(mc.to); | |
5436 | VM_BUG_ON(mc.precharge); | |
5437 | VM_BUG_ON(mc.moved_charge); | |
5438 | VM_BUG_ON(mc.moved_swap); | |
5439 | mem_cgroup_start_move(from); | |
5440 | spin_lock(&mc.lock); | |
5441 | mc.from = from; | |
5442 | mc.to = memcg; | |
5443 | spin_unlock(&mc.lock); | |
5444 | /* We set mc.moving_task later */ | |
5445 | ||
5446 | ret = mem_cgroup_precharge_mc(mm); | |
5447 | if (ret) | |
5448 | mem_cgroup_clear_mc(); | |
5449 | } | |
5450 | mmput(mm); | |
5451 | } | |
5452 | return ret; | |
5453 | } | |
5454 | ||
5455 | static void mem_cgroup_cancel_attach(struct cgroup *cgroup, | |
5456 | struct cgroup_taskset *tset) | |
5457 | { | |
5458 | mem_cgroup_clear_mc(); | |
5459 | } | |
5460 | ||
5461 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, | |
5462 | unsigned long addr, unsigned long end, | |
5463 | struct mm_walk *walk) | |
5464 | { | |
5465 | int ret = 0; | |
5466 | struct vm_area_struct *vma = walk->private; | |
5467 | pte_t *pte; | |
5468 | spinlock_t *ptl; | |
5469 | enum mc_target_type target_type; | |
5470 | union mc_target target; | |
5471 | struct page *page; | |
5472 | struct page_cgroup *pc; | |
5473 | ||
5474 | /* | |
5475 | * We don't take compound_lock() here but no race with splitting thp | |
5476 | * happens because: | |
5477 | * - if pmd_trans_huge_lock() returns 1, the relevant thp is not | |
5478 | * under splitting, which means there's no concurrent thp split, | |
5479 | * - if another thread runs into split_huge_page() just after we | |
5480 | * entered this if-block, the thread must wait for page table lock | |
5481 | * to be unlocked in __split_huge_page_splitting(), where the main | |
5482 | * part of thp split is not executed yet. | |
5483 | */ | |
5484 | if (pmd_trans_huge_lock(pmd, vma) == 1) { | |
5485 | if (mc.precharge < HPAGE_PMD_NR) { | |
5486 | spin_unlock(&vma->vm_mm->page_table_lock); | |
5487 | return 0; | |
5488 | } | |
5489 | target_type = get_mctgt_type_thp(vma, addr, *pmd, &target); | |
5490 | if (target_type == MC_TARGET_PAGE) { | |
5491 | page = target.page; | |
5492 | if (!isolate_lru_page(page)) { | |
5493 | pc = lookup_page_cgroup(page); | |
5494 | if (!mem_cgroup_move_account(page, HPAGE_PMD_NR, | |
5495 | pc, mc.from, mc.to, | |
5496 | false)) { | |
5497 | mc.precharge -= HPAGE_PMD_NR; | |
5498 | mc.moved_charge += HPAGE_PMD_NR; | |
5499 | } | |
5500 | putback_lru_page(page); | |
5501 | } | |
5502 | put_page(page); | |
5503 | } | |
5504 | spin_unlock(&vma->vm_mm->page_table_lock); | |
5505 | return 0; | |
5506 | } | |
5507 | ||
5508 | if (pmd_trans_unstable(pmd)) | |
5509 | return 0; | |
5510 | retry: | |
5511 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
5512 | for (; addr != end; addr += PAGE_SIZE) { | |
5513 | pte_t ptent = *(pte++); | |
5514 | swp_entry_t ent; | |
5515 | ||
5516 | if (!mc.precharge) | |
5517 | break; | |
5518 | ||
5519 | switch (get_mctgt_type(vma, addr, ptent, &target)) { | |
5520 | case MC_TARGET_PAGE: | |
5521 | page = target.page; | |
5522 | if (isolate_lru_page(page)) | |
5523 | goto put; | |
5524 | pc = lookup_page_cgroup(page); | |
5525 | if (!mem_cgroup_move_account(page, 1, pc, | |
5526 | mc.from, mc.to, false)) { | |
5527 | mc.precharge--; | |
5528 | /* we uncharge from mc.from later. */ | |
5529 | mc.moved_charge++; | |
5530 | } | |
5531 | putback_lru_page(page); | |
5532 | put: /* get_mctgt_type() gets the page */ | |
5533 | put_page(page); | |
5534 | break; | |
5535 | case MC_TARGET_SWAP: | |
5536 | ent = target.ent; | |
5537 | if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) { | |
5538 | mc.precharge--; | |
5539 | /* we fixup refcnts and charges later. */ | |
5540 | mc.moved_swap++; | |
5541 | } | |
5542 | break; | |
5543 | default: | |
5544 | break; | |
5545 | } | |
5546 | } | |
5547 | pte_unmap_unlock(pte - 1, ptl); | |
5548 | cond_resched(); | |
5549 | ||
5550 | if (addr != end) { | |
5551 | /* | |
5552 | * We have consumed all precharges we got in can_attach(). | |
5553 | * We try charge one by one, but don't do any additional | |
5554 | * charges to mc.to if we have failed in charge once in attach() | |
5555 | * phase. | |
5556 | */ | |
5557 | ret = mem_cgroup_do_precharge(1); | |
5558 | if (!ret) | |
5559 | goto retry; | |
5560 | } | |
5561 | ||
5562 | return ret; | |
5563 | } | |
5564 | ||
5565 | static void mem_cgroup_move_charge(struct mm_struct *mm) | |
5566 | { | |
5567 | struct vm_area_struct *vma; | |
5568 | ||
5569 | lru_add_drain_all(); | |
5570 | retry: | |
5571 | if (unlikely(!down_read_trylock(&mm->mmap_sem))) { | |
5572 | /* | |
5573 | * Someone who are holding the mmap_sem might be waiting in | |
5574 | * waitq. So we cancel all extra charges, wake up all waiters, | |
5575 | * and retry. Because we cancel precharges, we might not be able | |
5576 | * to move enough charges, but moving charge is a best-effort | |
5577 | * feature anyway, so it wouldn't be a big problem. | |
5578 | */ | |
5579 | __mem_cgroup_clear_mc(); | |
5580 | cond_resched(); | |
5581 | goto retry; | |
5582 | } | |
5583 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
5584 | int ret; | |
5585 | struct mm_walk mem_cgroup_move_charge_walk = { | |
5586 | .pmd_entry = mem_cgroup_move_charge_pte_range, | |
5587 | .mm = mm, | |
5588 | .private = vma, | |
5589 | }; | |
5590 | if (is_vm_hugetlb_page(vma)) | |
5591 | continue; | |
5592 | ret = walk_page_range(vma->vm_start, vma->vm_end, | |
5593 | &mem_cgroup_move_charge_walk); | |
5594 | if (ret) | |
5595 | /* | |
5596 | * means we have consumed all precharges and failed in | |
5597 | * doing additional charge. Just abandon here. | |
5598 | */ | |
5599 | break; | |
5600 | } | |
5601 | up_read(&mm->mmap_sem); | |
5602 | } | |
5603 | ||
5604 | static void mem_cgroup_move_task(struct cgroup *cont, | |
5605 | struct cgroup_taskset *tset) | |
5606 | { | |
5607 | struct task_struct *p = cgroup_taskset_first(tset); | |
5608 | struct mm_struct *mm = get_task_mm(p); | |
5609 | ||
5610 | if (mm) { | |
5611 | if (mc.to) | |
5612 | mem_cgroup_move_charge(mm); | |
5613 | mmput(mm); | |
5614 | } | |
5615 | if (mc.to) | |
5616 | mem_cgroup_clear_mc(); | |
5617 | } | |
5618 | #else /* !CONFIG_MMU */ | |
5619 | static int mem_cgroup_can_attach(struct cgroup *cgroup, | |
5620 | struct cgroup_taskset *tset) | |
5621 | { | |
5622 | return 0; | |
5623 | } | |
5624 | static void mem_cgroup_cancel_attach(struct cgroup *cgroup, | |
5625 | struct cgroup_taskset *tset) | |
5626 | { | |
5627 | } | |
5628 | static void mem_cgroup_move_task(struct cgroup *cont, | |
5629 | struct cgroup_taskset *tset) | |
5630 | { | |
5631 | } | |
5632 | #endif | |
5633 | ||
5634 | struct cgroup_subsys mem_cgroup_subsys = { | |
5635 | .name = "memory", | |
5636 | .subsys_id = mem_cgroup_subsys_id, | |
5637 | .create = mem_cgroup_create, | |
5638 | .pre_destroy = mem_cgroup_pre_destroy, | |
5639 | .destroy = mem_cgroup_destroy, | |
5640 | .can_attach = mem_cgroup_can_attach, | |
5641 | .cancel_attach = mem_cgroup_cancel_attach, | |
5642 | .attach = mem_cgroup_move_task, | |
5643 | .base_cftypes = mem_cgroup_files, | |
5644 | .early_init = 0, | |
5645 | .use_id = 1, | |
5646 | .__DEPRECATED_clear_css_refs = true, | |
5647 | }; | |
5648 | ||
5649 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
5650 | static int __init enable_swap_account(char *s) | |
5651 | { | |
5652 | /* consider enabled if no parameter or 1 is given */ | |
5653 | if (!strcmp(s, "1")) | |
5654 | really_do_swap_account = 1; | |
5655 | else if (!strcmp(s, "0")) | |
5656 | really_do_swap_account = 0; | |
5657 | return 1; | |
5658 | } | |
5659 | __setup("swapaccount=", enable_swap_account); | |
5660 | ||
5661 | #endif |