Root/
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 | static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, |
1153 | struct mem_cgroup *memcg) |
1154 | { |
1155 | if (root_memcg != memcg) { |
1156 | return (root_memcg->use_hierarchy && |
1157 | css_is_ancestor(&memcg->css, &root_memcg->css)); |
1158 | } |
1159 | |
1160 | return true; |
1161 | } |
1162 | |
1163 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) |
1164 | { |
1165 | int ret; |
1166 | struct mem_cgroup *curr = NULL; |
1167 | struct task_struct *p; |
1168 | |
1169 | p = find_lock_task_mm(task); |
1170 | if (p) { |
1171 | curr = try_get_mem_cgroup_from_mm(p->mm); |
1172 | task_unlock(p); |
1173 | } else { |
1174 | /* |
1175 | * All threads may have already detached their mm's, but the oom |
1176 | * killer still needs to detect if they have already been oom |
1177 | * killed to prevent needlessly killing additional tasks. |
1178 | */ |
1179 | task_lock(task); |
1180 | curr = mem_cgroup_from_task(task); |
1181 | if (curr) |
1182 | css_get(&curr->css); |
1183 | task_unlock(task); |
1184 | } |
1185 | if (!curr) |
1186 | return 0; |
1187 | /* |
1188 | * We should check use_hierarchy of "memcg" not "curr". Because checking |
1189 | * use_hierarchy of "curr" here make this function true if hierarchy is |
1190 | * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* |
1191 | * hierarchy(even if use_hierarchy is disabled in "memcg"). |
1192 | */ |
1193 | ret = mem_cgroup_same_or_subtree(memcg, curr); |
1194 | css_put(&curr->css); |
1195 | return ret; |
1196 | } |
1197 | |
1198 | int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) |
1199 | { |
1200 | unsigned long inactive_ratio; |
1201 | int nid = zone_to_nid(zone); |
1202 | int zid = zone_idx(zone); |
1203 | unsigned long inactive; |
1204 | unsigned long active; |
1205 | unsigned long gb; |
1206 | |
1207 | inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
1208 | BIT(LRU_INACTIVE_ANON)); |
1209 | active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
1210 | BIT(LRU_ACTIVE_ANON)); |
1211 | |
1212 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
1213 | if (gb) |
1214 | inactive_ratio = int_sqrt(10 * gb); |
1215 | else |
1216 | inactive_ratio = 1; |
1217 | |
1218 | return inactive * inactive_ratio < active; |
1219 | } |
1220 | |
1221 | int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) |
1222 | { |
1223 | unsigned long active; |
1224 | unsigned long inactive; |
1225 | int zid = zone_idx(zone); |
1226 | int nid = zone_to_nid(zone); |
1227 | |
1228 | inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
1229 | BIT(LRU_INACTIVE_FILE)); |
1230 | active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, |
1231 | BIT(LRU_ACTIVE_FILE)); |
1232 | |
1233 | return (active > inactive); |
1234 | } |
1235 | |
1236 | struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, |
1237 | struct zone *zone) |
1238 | { |
1239 | int nid = zone_to_nid(zone); |
1240 | int zid = zone_idx(zone); |
1241 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
1242 | |
1243 | return &mz->reclaim_stat; |
1244 | } |
1245 | |
1246 | struct zone_reclaim_stat * |
1247 | mem_cgroup_get_reclaim_stat_from_page(struct page *page) |
1248 | { |
1249 | struct page_cgroup *pc; |
1250 | struct mem_cgroup_per_zone *mz; |
1251 | |
1252 | if (mem_cgroup_disabled()) |
1253 | return NULL; |
1254 | |
1255 | pc = lookup_page_cgroup(page); |
1256 | if (!PageCgroupUsed(pc)) |
1257 | return NULL; |
1258 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ |
1259 | smp_rmb(); |
1260 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); |
1261 | return &mz->reclaim_stat; |
1262 | } |
1263 | |
1264 | #define mem_cgroup_from_res_counter(counter, member) \ |
1265 | container_of(counter, struct mem_cgroup, member) |
1266 | |
1267 | /** |
1268 | * mem_cgroup_margin - calculate chargeable space of a memory cgroup |
1269 | * @mem: the memory cgroup |
1270 | * |
1271 | * Returns the maximum amount of memory @mem can be charged with, in |
1272 | * pages. |
1273 | */ |
1274 | static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) |
1275 | { |
1276 | unsigned long long margin; |
1277 | |
1278 | margin = res_counter_margin(&memcg->res); |
1279 | if (do_swap_account) |
1280 | margin = min(margin, res_counter_margin(&memcg->memsw)); |
1281 | return margin >> PAGE_SHIFT; |
1282 | } |
1283 | |
1284 | int mem_cgroup_swappiness(struct mem_cgroup *memcg) |
1285 | { |
1286 | struct cgroup *cgrp = memcg->css.cgroup; |
1287 | |
1288 | /* root ? */ |
1289 | if (cgrp->parent == NULL) |
1290 | return vm_swappiness; |
1291 | |
1292 | return memcg->swappiness; |
1293 | } |
1294 | |
1295 | /* |
1296 | * memcg->moving_account is used for checking possibility that some thread is |
1297 | * calling move_account(). When a thread on CPU-A starts moving pages under |
1298 | * a memcg, other threads should check memcg->moving_account under |
1299 | * rcu_read_lock(), like this: |
1300 | * |
1301 | * CPU-A CPU-B |
1302 | * rcu_read_lock() |
1303 | * memcg->moving_account+1 if (memcg->mocing_account) |
1304 | * take heavy locks. |
1305 | * synchronize_rcu() update something. |
1306 | * rcu_read_unlock() |
1307 | * start move here. |
1308 | */ |
1309 | |
1310 | /* for quick checking without looking up memcg */ |
1311 | atomic_t memcg_moving __read_mostly; |
1312 | |
1313 | static void mem_cgroup_start_move(struct mem_cgroup *memcg) |
1314 | { |
1315 | atomic_inc(&memcg_moving); |
1316 | atomic_inc(&memcg->moving_account); |
1317 | synchronize_rcu(); |
1318 | } |
1319 | |
1320 | static void mem_cgroup_end_move(struct mem_cgroup *memcg) |
1321 | { |
1322 | /* |
1323 | * Now, mem_cgroup_clear_mc() may call this function with NULL. |
1324 | * We check NULL in callee rather than caller. |
1325 | */ |
1326 | if (memcg) { |
1327 | atomic_dec(&memcg_moving); |
1328 | atomic_dec(&memcg->moving_account); |
1329 | } |
1330 | } |
1331 | |
1332 | /* |
1333 | * 2 routines for checking "mem" is under move_account() or not. |
1334 | * |
1335 | * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This |
1336 | * is used for avoiding races in accounting. If true, |
1337 | * pc->mem_cgroup may be overwritten. |
1338 | * |
1339 | * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or |
1340 | * under hierarchy of moving cgroups. This is for |
1341 | * waiting at hith-memory prressure caused by "move". |
1342 | */ |
1343 | |
1344 | static bool mem_cgroup_stolen(struct mem_cgroup *memcg) |
1345 | { |
1346 | VM_BUG_ON(!rcu_read_lock_held()); |
1347 | return atomic_read(&memcg->moving_account) > 0; |
1348 | } |
1349 | |
1350 | static bool mem_cgroup_under_move(struct mem_cgroup *memcg) |
1351 | { |
1352 | struct mem_cgroup *from; |
1353 | struct mem_cgroup *to; |
1354 | bool ret = false; |
1355 | /* |
1356 | * Unlike task_move routines, we access mc.to, mc.from not under |
1357 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. |
1358 | */ |
1359 | spin_lock(&mc.lock); |
1360 | from = mc.from; |
1361 | to = mc.to; |
1362 | if (!from) |
1363 | goto unlock; |
1364 | |
1365 | ret = mem_cgroup_same_or_subtree(memcg, from) |
1366 | || mem_cgroup_same_or_subtree(memcg, to); |
1367 | unlock: |
1368 | spin_unlock(&mc.lock); |
1369 | return ret; |
1370 | } |
1371 | |
1372 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) |
1373 | { |
1374 | if (mc.moving_task && current != mc.moving_task) { |
1375 | if (mem_cgroup_under_move(memcg)) { |
1376 | DEFINE_WAIT(wait); |
1377 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); |
1378 | /* moving charge context might have finished. */ |
1379 | if (mc.moving_task) |
1380 | schedule(); |
1381 | finish_wait(&mc.waitq, &wait); |
1382 | return true; |
1383 | } |
1384 | } |
1385 | return false; |
1386 | } |
1387 | |
1388 | /* |
1389 | * Take this lock when |
1390 | * - a code tries to modify page's memcg while it's USED. |
1391 | * - a code tries to modify page state accounting in a memcg. |
1392 | * see mem_cgroup_stolen(), too. |
1393 | */ |
1394 | static void move_lock_mem_cgroup(struct mem_cgroup *memcg, |
1395 | unsigned long *flags) |
1396 | { |
1397 | spin_lock_irqsave(&memcg->move_lock, *flags); |
1398 | } |
1399 | |
1400 | static void move_unlock_mem_cgroup(struct mem_cgroup *memcg, |
1401 | unsigned long *flags) |
1402 | { |
1403 | spin_unlock_irqrestore(&memcg->move_lock, *flags); |
1404 | } |
1405 | |
1406 | /** |
1407 | * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. |
1408 | * @memcg: The memory cgroup that went over limit |
1409 | * @p: Task that is going to be killed |
1410 | * |
1411 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is |
1412 | * enabled |
1413 | */ |
1414 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) |
1415 | { |
1416 | struct cgroup *task_cgrp; |
1417 | struct cgroup *mem_cgrp; |
1418 | /* |
1419 | * Need a buffer in BSS, can't rely on allocations. The code relies |
1420 | * on the assumption that OOM is serialized for memory controller. |
1421 | * If this assumption is broken, revisit this code. |
1422 | */ |
1423 | static char memcg_name[PATH_MAX]; |
1424 | int ret; |
1425 | |
1426 | if (!memcg || !p) |
1427 | return; |
1428 | |
1429 | rcu_read_lock(); |
1430 | |
1431 | mem_cgrp = memcg->css.cgroup; |
1432 | task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); |
1433 | |
1434 | ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); |
1435 | if (ret < 0) { |
1436 | /* |
1437 | * Unfortunately, we are unable to convert to a useful name |
1438 | * But we'll still print out the usage information |
1439 | */ |
1440 | rcu_read_unlock(); |
1441 | goto done; |
1442 | } |
1443 | rcu_read_unlock(); |
1444 | |
1445 | printk(KERN_INFO "Task in %s killed", memcg_name); |
1446 | |
1447 | rcu_read_lock(); |
1448 | ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); |
1449 | if (ret < 0) { |
1450 | rcu_read_unlock(); |
1451 | goto done; |
1452 | } |
1453 | rcu_read_unlock(); |
1454 | |
1455 | /* |
1456 | * Continues from above, so we don't need an KERN_ level |
1457 | */ |
1458 | printk(KERN_CONT " as a result of limit of %s\n", memcg_name); |
1459 | done: |
1460 | |
1461 | printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", |
1462 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, |
1463 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, |
1464 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); |
1465 | printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " |
1466 | "failcnt %llu\n", |
1467 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, |
1468 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, |
1469 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); |
1470 | } |
1471 | |
1472 | /* |
1473 | * This function returns the number of memcg under hierarchy tree. Returns |
1474 | * 1(self count) if no children. |
1475 | */ |
1476 | static int mem_cgroup_count_children(struct mem_cgroup *memcg) |
1477 | { |
1478 | int num = 0; |
1479 | struct mem_cgroup *iter; |
1480 | |
1481 | for_each_mem_cgroup_tree(iter, memcg) |
1482 | num++; |
1483 | return num; |
1484 | } |
1485 | |
1486 | /* |
1487 | * Return the memory (and swap, if configured) limit for a memcg. |
1488 | */ |
1489 | u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) |
1490 | { |
1491 | u64 limit; |
1492 | u64 memsw; |
1493 | |
1494 | limit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
1495 | limit += total_swap_pages << PAGE_SHIFT; |
1496 | |
1497 | memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
1498 | /* |
1499 | * If memsw is finite and limits the amount of swap space available |
1500 | * to this memcg, return that limit. |
1501 | */ |
1502 | return min(limit, memsw); |
1503 | } |
1504 | |
1505 | static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, |
1506 | gfp_t gfp_mask, |
1507 | unsigned long flags) |
1508 | { |
1509 | unsigned long total = 0; |
1510 | bool noswap = false; |
1511 | int loop; |
1512 | |
1513 | if (flags & MEM_CGROUP_RECLAIM_NOSWAP) |
1514 | noswap = true; |
1515 | if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) |
1516 | noswap = true; |
1517 | |
1518 | for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { |
1519 | if (loop) |
1520 | drain_all_stock_async(memcg); |
1521 | total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); |
1522 | /* |
1523 | * Allow limit shrinkers, which are triggered directly |
1524 | * by userspace, to catch signals and stop reclaim |
1525 | * after minimal progress, regardless of the margin. |
1526 | */ |
1527 | if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) |
1528 | break; |
1529 | if (mem_cgroup_margin(memcg)) |
1530 | break; |
1531 | /* |
1532 | * If nothing was reclaimed after two attempts, there |
1533 | * may be no reclaimable pages in this hierarchy. |
1534 | */ |
1535 | if (loop && !total) |
1536 | break; |
1537 | } |
1538 | return total; |
1539 | } |
1540 | |
1541 | /** |
1542 | * test_mem_cgroup_node_reclaimable |
1543 | * @mem: the target memcg |
1544 | * @nid: the node ID to be checked. |
1545 | * @noswap : specify true here if the user wants flle only information. |
1546 | * |
1547 | * This function returns whether the specified memcg contains any |
1548 | * reclaimable pages on a node. Returns true if there are any reclaimable |
1549 | * pages in the node. |
1550 | */ |
1551 | static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, |
1552 | int nid, bool noswap) |
1553 | { |
1554 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) |
1555 | return true; |
1556 | if (noswap || !total_swap_pages) |
1557 | return false; |
1558 | if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) |
1559 | return true; |
1560 | return false; |
1561 | |
1562 | } |
1563 | #if MAX_NUMNODES > 1 |
1564 | |
1565 | /* |
1566 | * Always updating the nodemask is not very good - even if we have an empty |
1567 | * list or the wrong list here, we can start from some node and traverse all |
1568 | * nodes based on the zonelist. So update the list loosely once per 10 secs. |
1569 | * |
1570 | */ |
1571 | static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) |
1572 | { |
1573 | int nid; |
1574 | /* |
1575 | * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET |
1576 | * pagein/pageout changes since the last update. |
1577 | */ |
1578 | if (!atomic_read(&memcg->numainfo_events)) |
1579 | return; |
1580 | if (atomic_inc_return(&memcg->numainfo_updating) > 1) |
1581 | return; |
1582 | |
1583 | /* make a nodemask where this memcg uses memory from */ |
1584 | memcg->scan_nodes = node_states[N_HIGH_MEMORY]; |
1585 | |
1586 | for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { |
1587 | |
1588 | if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) |
1589 | node_clear(nid, memcg->scan_nodes); |
1590 | } |
1591 | |
1592 | atomic_set(&memcg->numainfo_events, 0); |
1593 | atomic_set(&memcg->numainfo_updating, 0); |
1594 | } |
1595 | |
1596 | /* |
1597 | * Selecting a node where we start reclaim from. Because what we need is just |
1598 | * reducing usage counter, start from anywhere is O,K. Considering |
1599 | * memory reclaim from current node, there are pros. and cons. |
1600 | * |
1601 | * Freeing memory from current node means freeing memory from a node which |
1602 | * we'll use or we've used. So, it may make LRU bad. And if several threads |
1603 | * hit limits, it will see a contention on a node. But freeing from remote |
1604 | * node means more costs for memory reclaim because of memory latency. |
1605 | * |
1606 | * Now, we use round-robin. Better algorithm is welcomed. |
1607 | */ |
1608 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) |
1609 | { |
1610 | int node; |
1611 | |
1612 | mem_cgroup_may_update_nodemask(memcg); |
1613 | node = memcg->last_scanned_node; |
1614 | |
1615 | node = next_node(node, memcg->scan_nodes); |
1616 | if (node == MAX_NUMNODES) |
1617 | node = first_node(memcg->scan_nodes); |
1618 | /* |
1619 | * We call this when we hit limit, not when pages are added to LRU. |
1620 | * No LRU may hold pages because all pages are UNEVICTABLE or |
1621 | * memcg is too small and all pages are not on LRU. In that case, |
1622 | * we use curret node. |
1623 | */ |
1624 | if (unlikely(node == MAX_NUMNODES)) |
1625 | node = numa_node_id(); |
1626 | |
1627 | memcg->last_scanned_node = node; |
1628 | return node; |
1629 | } |
1630 | |
1631 | /* |
1632 | * Check all nodes whether it contains reclaimable pages or not. |
1633 | * For quick scan, we make use of scan_nodes. This will allow us to skip |
1634 | * unused nodes. But scan_nodes is lazily updated and may not cotain |
1635 | * enough new information. We need to do double check. |
1636 | */ |
1637 | bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) |
1638 | { |
1639 | int nid; |
1640 | |
1641 | /* |
1642 | * quick check...making use of scan_node. |
1643 | * We can skip unused nodes. |
1644 | */ |
1645 | if (!nodes_empty(memcg->scan_nodes)) { |
1646 | for (nid = first_node(memcg->scan_nodes); |
1647 | nid < MAX_NUMNODES; |
1648 | nid = next_node(nid, memcg->scan_nodes)) { |
1649 | |
1650 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) |
1651 | return true; |
1652 | } |
1653 | } |
1654 | /* |
1655 | * Check rest of nodes. |
1656 | */ |
1657 | for_each_node_state(nid, N_HIGH_MEMORY) { |
1658 | if (node_isset(nid, memcg->scan_nodes)) |
1659 | continue; |
1660 | if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) |
1661 | return true; |
1662 | } |
1663 | return false; |
1664 | } |
1665 | |
1666 | #else |
1667 | int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) |
1668 | { |
1669 | return 0; |
1670 | } |
1671 | |
1672 | bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) |
1673 | { |
1674 | return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); |
1675 | } |
1676 | #endif |
1677 | |
1678 | static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, |
1679 | struct zone *zone, |
1680 | gfp_t gfp_mask, |
1681 | unsigned long *total_scanned) |
1682 | { |
1683 | struct mem_cgroup *victim = NULL; |
1684 | int total = 0; |
1685 | int loop = 0; |
1686 | unsigned long excess; |
1687 | unsigned long nr_scanned; |
1688 | struct mem_cgroup_reclaim_cookie reclaim = { |
1689 | .zone = zone, |
1690 | .priority = 0, |
1691 | }; |
1692 | |
1693 | excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; |
1694 | |
1695 | while (1) { |
1696 | victim = mem_cgroup_iter(root_memcg, victim, &reclaim); |
1697 | if (!victim) { |
1698 | loop++; |
1699 | if (loop >= 2) { |
1700 | /* |
1701 | * If we have not been able to reclaim |
1702 | * anything, it might because there are |
1703 | * no reclaimable pages under this hierarchy |
1704 | */ |
1705 | if (!total) |
1706 | break; |
1707 | /* |
1708 | * We want to do more targeted reclaim. |
1709 | * excess >> 2 is not to excessive so as to |
1710 | * reclaim too much, nor too less that we keep |
1711 | * coming back to reclaim from this cgroup |
1712 | */ |
1713 | if (total >= (excess >> 2) || |
1714 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) |
1715 | break; |
1716 | } |
1717 | continue; |
1718 | } |
1719 | if (!mem_cgroup_reclaimable(victim, false)) |
1720 | continue; |
1721 | total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, |
1722 | zone, &nr_scanned); |
1723 | *total_scanned += nr_scanned; |
1724 | if (!res_counter_soft_limit_excess(&root_memcg->res)) |
1725 | break; |
1726 | } |
1727 | mem_cgroup_iter_break(root_memcg, victim); |
1728 | return total; |
1729 | } |
1730 | |
1731 | /* |
1732 | * Check OOM-Killer is already running under our hierarchy. |
1733 | * If someone is running, return false. |
1734 | * Has to be called with memcg_oom_lock |
1735 | */ |
1736 | static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) |
1737 | { |
1738 | struct mem_cgroup *iter, *failed = NULL; |
1739 | |
1740 | for_each_mem_cgroup_tree(iter, memcg) { |
1741 | if (iter->oom_lock) { |
1742 | /* |
1743 | * this subtree of our hierarchy is already locked |
1744 | * so we cannot give a lock. |
1745 | */ |
1746 | failed = iter; |
1747 | mem_cgroup_iter_break(memcg, iter); |
1748 | break; |
1749 | } else |
1750 | iter->oom_lock = true; |
1751 | } |
1752 | |
1753 | if (!failed) |
1754 | return true; |
1755 | |
1756 | /* |
1757 | * OK, we failed to lock the whole subtree so we have to clean up |
1758 | * what we set up to the failing subtree |
1759 | */ |
1760 | for_each_mem_cgroup_tree(iter, memcg) { |
1761 | if (iter == failed) { |
1762 | mem_cgroup_iter_break(memcg, iter); |
1763 | break; |
1764 | } |
1765 | iter->oom_lock = false; |
1766 | } |
1767 | return false; |
1768 | } |
1769 | |
1770 | /* |
1771 | * Has to be called with memcg_oom_lock |
1772 | */ |
1773 | static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) |
1774 | { |
1775 | struct mem_cgroup *iter; |
1776 | |
1777 | for_each_mem_cgroup_tree(iter, memcg) |
1778 | iter->oom_lock = false; |
1779 | return 0; |
1780 | } |
1781 | |
1782 | static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) |
1783 | { |
1784 | struct mem_cgroup *iter; |
1785 | |
1786 | for_each_mem_cgroup_tree(iter, memcg) |
1787 | atomic_inc(&iter->under_oom); |
1788 | } |
1789 | |
1790 | static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) |
1791 | { |
1792 | struct mem_cgroup *iter; |
1793 | |
1794 | /* |
1795 | * When a new child is created while the hierarchy is under oom, |
1796 | * mem_cgroup_oom_lock() may not be called. We have to use |
1797 | * atomic_add_unless() here. |
1798 | */ |
1799 | for_each_mem_cgroup_tree(iter, memcg) |
1800 | atomic_add_unless(&iter->under_oom, -1, 0); |
1801 | } |
1802 | |
1803 | static DEFINE_SPINLOCK(memcg_oom_lock); |
1804 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); |
1805 | |
1806 | struct oom_wait_info { |
1807 | struct mem_cgroup *memcg; |
1808 | wait_queue_t wait; |
1809 | }; |
1810 | |
1811 | static int memcg_oom_wake_function(wait_queue_t *wait, |
1812 | unsigned mode, int sync, void *arg) |
1813 | { |
1814 | struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg; |
1815 | struct mem_cgroup *oom_wait_memcg; |
1816 | struct oom_wait_info *oom_wait_info; |
1817 | |
1818 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); |
1819 | oom_wait_memcg = oom_wait_info->memcg; |
1820 | |
1821 | /* |
1822 | * Both of oom_wait_info->memcg and wake_memcg are stable under us. |
1823 | * Then we can use css_is_ancestor without taking care of RCU. |
1824 | */ |
1825 | if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) |
1826 | && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) |
1827 | return 0; |
1828 | return autoremove_wake_function(wait, mode, sync, arg); |
1829 | } |
1830 | |
1831 | static void memcg_wakeup_oom(struct mem_cgroup *memcg) |
1832 | { |
1833 | /* for filtering, pass "memcg" as argument. */ |
1834 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); |
1835 | } |
1836 | |
1837 | static void memcg_oom_recover(struct mem_cgroup *memcg) |
1838 | { |
1839 | if (memcg && atomic_read(&memcg->under_oom)) |
1840 | memcg_wakeup_oom(memcg); |
1841 | } |
1842 | |
1843 | /* |
1844 | * try to call OOM killer. returns false if we should exit memory-reclaim loop. |
1845 | */ |
1846 | bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask, int order) |
1847 | { |
1848 | struct oom_wait_info owait; |
1849 | bool locked, need_to_kill; |
1850 | |
1851 | owait.memcg = memcg; |
1852 | owait.wait.flags = 0; |
1853 | owait.wait.func = memcg_oom_wake_function; |
1854 | owait.wait.private = current; |
1855 | INIT_LIST_HEAD(&owait.wait.task_list); |
1856 | need_to_kill = true; |
1857 | mem_cgroup_mark_under_oom(memcg); |
1858 | |
1859 | /* At first, try to OOM lock hierarchy under memcg.*/ |
1860 | spin_lock(&memcg_oom_lock); |
1861 | locked = mem_cgroup_oom_lock(memcg); |
1862 | /* |
1863 | * Even if signal_pending(), we can't quit charge() loop without |
1864 | * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL |
1865 | * under OOM is always welcomed, use TASK_KILLABLE here. |
1866 | */ |
1867 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); |
1868 | if (!locked || memcg->oom_kill_disable) |
1869 | need_to_kill = false; |
1870 | if (locked) |
1871 | mem_cgroup_oom_notify(memcg); |
1872 | spin_unlock(&memcg_oom_lock); |
1873 | |
1874 | if (need_to_kill) { |
1875 | finish_wait(&memcg_oom_waitq, &owait.wait); |
1876 | mem_cgroup_out_of_memory(memcg, mask, order); |
1877 | } else { |
1878 | schedule(); |
1879 | finish_wait(&memcg_oom_waitq, &owait.wait); |
1880 | } |
1881 | spin_lock(&memcg_oom_lock); |
1882 | if (locked) |
1883 | mem_cgroup_oom_unlock(memcg); |
1884 | memcg_wakeup_oom(memcg); |
1885 | spin_unlock(&memcg_oom_lock); |
1886 | |
1887 | mem_cgroup_unmark_under_oom(memcg); |
1888 | |
1889 | if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) |
1890 | return false; |
1891 | /* Give chance to dying process */ |
1892 | schedule_timeout_uninterruptible(1); |
1893 | return true; |
1894 | } |
1895 | |
1896 | /* |
1897 | * Currently used to update mapped file statistics, but the routine can be |
1898 | * generalized to update other statistics as well. |
1899 | * |
1900 | * Notes: Race condition |
1901 | * |
1902 | * We usually use page_cgroup_lock() for accessing page_cgroup member but |
1903 | * it tends to be costly. But considering some conditions, we doesn't need |
1904 | * to do so _always_. |
1905 | * |
1906 | * Considering "charge", lock_page_cgroup() is not required because all |
1907 | * file-stat operations happen after a page is attached to radix-tree. There |
1908 | * are no race with "charge". |
1909 | * |
1910 | * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup |
1911 | * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even |
1912 | * if there are race with "uncharge". Statistics itself is properly handled |
1913 | * by flags. |
1914 | * |
1915 | * Considering "move", this is an only case we see a race. To make the race |
1916 | * small, we check mm->moving_account and detect there are possibility of race |
1917 | * If there is, we take a lock. |
1918 | */ |
1919 | |
1920 | void __mem_cgroup_begin_update_page_stat(struct page *page, |
1921 | bool *locked, unsigned long *flags) |
1922 | { |
1923 | struct mem_cgroup *memcg; |
1924 | struct page_cgroup *pc; |
1925 | |
1926 | pc = lookup_page_cgroup(page); |
1927 | again: |
1928 | memcg = pc->mem_cgroup; |
1929 | if (unlikely(!memcg || !PageCgroupUsed(pc))) |
1930 | return; |
1931 | /* |
1932 | * If this memory cgroup is not under account moving, we don't |
1933 | * need to take move_lock_page_cgroup(). Because we already hold |
1934 | * rcu_read_lock(), any calls to move_account will be delayed until |
1935 | * rcu_read_unlock() if mem_cgroup_stolen() == true. |
1936 | */ |
1937 | if (!mem_cgroup_stolen(memcg)) |
1938 | return; |
1939 | |
1940 | move_lock_mem_cgroup(memcg, flags); |
1941 | if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) { |
1942 | move_unlock_mem_cgroup(memcg, flags); |
1943 | goto again; |
1944 | } |
1945 | *locked = true; |
1946 | } |
1947 | |
1948 | void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags) |
1949 | { |
1950 | struct page_cgroup *pc = lookup_page_cgroup(page); |
1951 | |
1952 | /* |
1953 | * It's guaranteed that pc->mem_cgroup never changes while |
1954 | * lock is held because a routine modifies pc->mem_cgroup |
1955 | * should take move_lock_page_cgroup(). |
1956 | */ |
1957 | move_unlock_mem_cgroup(pc->mem_cgroup, flags); |
1958 | } |
1959 | |
1960 | void mem_cgroup_update_page_stat(struct page *page, |
1961 | enum mem_cgroup_page_stat_item idx, int val) |
1962 | { |
1963 | struct mem_cgroup *memcg; |
1964 | struct page_cgroup *pc = lookup_page_cgroup(page); |
1965 | unsigned long uninitialized_var(flags); |
1966 | |
1967 | if (mem_cgroup_disabled()) |
1968 | return; |
1969 | |
1970 | memcg = pc->mem_cgroup; |
1971 | if (unlikely(!memcg || !PageCgroupUsed(pc))) |
1972 | return; |
1973 | |
1974 | switch (idx) { |
1975 | case MEMCG_NR_FILE_MAPPED: |
1976 | idx = MEM_CGROUP_STAT_FILE_MAPPED; |
1977 | break; |
1978 | default: |
1979 | BUG(); |
1980 | } |
1981 | |
1982 | this_cpu_add(memcg->stat->count[idx], val); |
1983 | } |
1984 | |
1985 | /* |
1986 | * size of first charge trial. "32" comes from vmscan.c's magic value. |
1987 | * TODO: maybe necessary to use big numbers in big irons. |
1988 | */ |
1989 | #define CHARGE_BATCH 32U |
1990 | struct memcg_stock_pcp { |
1991 | struct mem_cgroup *cached; /* this never be root cgroup */ |
1992 | unsigned int nr_pages; |
1993 | struct work_struct work; |
1994 | unsigned long flags; |
1995 | #define FLUSHING_CACHED_CHARGE (0) |
1996 | }; |
1997 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); |
1998 | static DEFINE_MUTEX(percpu_charge_mutex); |
1999 | |
2000 | /* |
2001 | * Try to consume stocked charge on this cpu. If success, one page is consumed |
2002 | * from local stock and true is returned. If the stock is 0 or charges from a |
2003 | * cgroup which is not current target, returns false. This stock will be |
2004 | * refilled. |
2005 | */ |
2006 | static bool consume_stock(struct mem_cgroup *memcg) |
2007 | { |
2008 | struct memcg_stock_pcp *stock; |
2009 | bool ret = true; |
2010 | |
2011 | stock = &get_cpu_var(memcg_stock); |
2012 | if (memcg == stock->cached && stock->nr_pages) |
2013 | stock->nr_pages--; |
2014 | else /* need to call res_counter_charge */ |
2015 | ret = false; |
2016 | put_cpu_var(memcg_stock); |
2017 | return ret; |
2018 | } |
2019 | |
2020 | /* |
2021 | * Returns stocks cached in percpu to res_counter and reset cached information. |
2022 | */ |
2023 | static void drain_stock(struct memcg_stock_pcp *stock) |
2024 | { |
2025 | struct mem_cgroup *old = stock->cached; |
2026 | |
2027 | if (stock->nr_pages) { |
2028 | unsigned long bytes = stock->nr_pages * PAGE_SIZE; |
2029 | |
2030 | res_counter_uncharge(&old->res, bytes); |
2031 | if (do_swap_account) |
2032 | res_counter_uncharge(&old->memsw, bytes); |
2033 | stock->nr_pages = 0; |
2034 | } |
2035 | stock->cached = NULL; |
2036 | } |
2037 | |
2038 | /* |
2039 | * This must be called under preempt disabled or must be called by |
2040 | * a thread which is pinned to local cpu. |
2041 | */ |
2042 | static void drain_local_stock(struct work_struct *dummy) |
2043 | { |
2044 | struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); |
2045 | drain_stock(stock); |
2046 | clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); |
2047 | } |
2048 | |
2049 | /* |
2050 | * Cache charges(val) which is from res_counter, to local per_cpu area. |
2051 | * This will be consumed by consume_stock() function, later. |
2052 | */ |
2053 | static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) |
2054 | { |
2055 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); |
2056 | |
2057 | if (stock->cached != memcg) { /* reset if necessary */ |
2058 | drain_stock(stock); |
2059 | stock->cached = memcg; |
2060 | } |
2061 | stock->nr_pages += nr_pages; |
2062 | put_cpu_var(memcg_stock); |
2063 | } |
2064 | |
2065 | /* |
2066 | * Drains all per-CPU charge caches for given root_memcg resp. subtree |
2067 | * of the hierarchy under it. sync flag says whether we should block |
2068 | * until the work is done. |
2069 | */ |
2070 | static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) |
2071 | { |
2072 | int cpu, curcpu; |
2073 | |
2074 | /* Notify other cpus that system-wide "drain" is running */ |
2075 | get_online_cpus(); |
2076 | curcpu = get_cpu(); |
2077 | for_each_online_cpu(cpu) { |
2078 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); |
2079 | struct mem_cgroup *memcg; |
2080 | |
2081 | memcg = stock->cached; |
2082 | if (!memcg || !stock->nr_pages) |
2083 | continue; |
2084 | if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) |
2085 | continue; |
2086 | if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { |
2087 | if (cpu == curcpu) |
2088 | drain_local_stock(&stock->work); |
2089 | else |
2090 | schedule_work_on(cpu, &stock->work); |
2091 | } |
2092 | } |
2093 | put_cpu(); |
2094 | |
2095 | if (!sync) |
2096 | goto out; |
2097 | |
2098 | for_each_online_cpu(cpu) { |
2099 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); |
2100 | if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) |
2101 | flush_work(&stock->work); |
2102 | } |
2103 | out: |
2104 | put_online_cpus(); |
2105 | } |
2106 | |
2107 | /* |
2108 | * Tries to drain stocked charges in other cpus. This function is asynchronous |
2109 | * and just put a work per cpu for draining localy on each cpu. Caller can |
2110 | * expects some charges will be back to res_counter later but cannot wait for |
2111 | * it. |
2112 | */ |
2113 | static void drain_all_stock_async(struct mem_cgroup *root_memcg) |
2114 | { |
2115 | /* |
2116 | * If someone calls draining, avoid adding more kworker runs. |
2117 | */ |
2118 | if (!mutex_trylock(&percpu_charge_mutex)) |
2119 | return; |
2120 | drain_all_stock(root_memcg, false); |
2121 | mutex_unlock(&percpu_charge_mutex); |
2122 | } |
2123 | |
2124 | /* This is a synchronous drain interface. */ |
2125 | static void drain_all_stock_sync(struct mem_cgroup *root_memcg) |
2126 | { |
2127 | /* called when force_empty is called */ |
2128 | mutex_lock(&percpu_charge_mutex); |
2129 | drain_all_stock(root_memcg, true); |
2130 | mutex_unlock(&percpu_charge_mutex); |
2131 | } |
2132 | |
2133 | /* |
2134 | * This function drains percpu counter value from DEAD cpu and |
2135 | * move it to local cpu. Note that this function can be preempted. |
2136 | */ |
2137 | static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) |
2138 | { |
2139 | int i; |
2140 | |
2141 | spin_lock(&memcg->pcp_counter_lock); |
2142 | for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { |
2143 | long x = per_cpu(memcg->stat->count[i], cpu); |
2144 | |
2145 | per_cpu(memcg->stat->count[i], cpu) = 0; |
2146 | memcg->nocpu_base.count[i] += x; |
2147 | } |
2148 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { |
2149 | unsigned long x = per_cpu(memcg->stat->events[i], cpu); |
2150 | |
2151 | per_cpu(memcg->stat->events[i], cpu) = 0; |
2152 | memcg->nocpu_base.events[i] += x; |
2153 | } |
2154 | spin_unlock(&memcg->pcp_counter_lock); |
2155 | } |
2156 | |
2157 | static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, |
2158 | unsigned long action, |
2159 | void *hcpu) |
2160 | { |
2161 | int cpu = (unsigned long)hcpu; |
2162 | struct memcg_stock_pcp *stock; |
2163 | struct mem_cgroup *iter; |
2164 | |
2165 | if (action == CPU_ONLINE) |
2166 | return NOTIFY_OK; |
2167 | |
2168 | if (action != CPU_DEAD && action != CPU_DEAD_FROZEN) |
2169 | return NOTIFY_OK; |
2170 | |
2171 | for_each_mem_cgroup(iter) |
2172 | mem_cgroup_drain_pcp_counter(iter, cpu); |
2173 | |
2174 | stock = &per_cpu(memcg_stock, cpu); |
2175 | drain_stock(stock); |
2176 | return NOTIFY_OK; |
2177 | } |
2178 | |
2179 | |
2180 | /* See __mem_cgroup_try_charge() for details */ |
2181 | enum { |
2182 | CHARGE_OK, /* success */ |
2183 | CHARGE_RETRY, /* need to retry but retry is not bad */ |
2184 | CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ |
2185 | CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ |
2186 | CHARGE_OOM_DIE, /* the current is killed because of OOM */ |
2187 | }; |
2188 | |
2189 | static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, |
2190 | unsigned int nr_pages, bool oom_check) |
2191 | { |
2192 | unsigned long csize = nr_pages * PAGE_SIZE; |
2193 | struct mem_cgroup *mem_over_limit; |
2194 | struct res_counter *fail_res; |
2195 | unsigned long flags = 0; |
2196 | int ret; |
2197 | |
2198 | ret = res_counter_charge(&memcg->res, csize, &fail_res); |
2199 | |
2200 | if (likely(!ret)) { |
2201 | if (!do_swap_account) |
2202 | return CHARGE_OK; |
2203 | ret = res_counter_charge(&memcg->memsw, csize, &fail_res); |
2204 | if (likely(!ret)) |
2205 | return CHARGE_OK; |
2206 | |
2207 | res_counter_uncharge(&memcg->res, csize); |
2208 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); |
2209 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; |
2210 | } else |
2211 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); |
2212 | /* |
2213 | * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch |
2214 | * of regular pages (CHARGE_BATCH), or a single regular page (1). |
2215 | * |
2216 | * Never reclaim on behalf of optional batching, retry with a |
2217 | * single page instead. |
2218 | */ |
2219 | if (nr_pages == CHARGE_BATCH) |
2220 | return CHARGE_RETRY; |
2221 | |
2222 | if (!(gfp_mask & __GFP_WAIT)) |
2223 | return CHARGE_WOULDBLOCK; |
2224 | |
2225 | ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); |
2226 | if (mem_cgroup_margin(mem_over_limit) >= nr_pages) |
2227 | return CHARGE_RETRY; |
2228 | /* |
2229 | * Even though the limit is exceeded at this point, reclaim |
2230 | * may have been able to free some pages. Retry the charge |
2231 | * before killing the task. |
2232 | * |
2233 | * Only for regular pages, though: huge pages are rather |
2234 | * unlikely to succeed so close to the limit, and we fall back |
2235 | * to regular pages anyway in case of failure. |
2236 | */ |
2237 | if (nr_pages == 1 && ret) |
2238 | return CHARGE_RETRY; |
2239 | |
2240 | /* |
2241 | * At task move, charge accounts can be doubly counted. So, it's |
2242 | * better to wait until the end of task_move if something is going on. |
2243 | */ |
2244 | if (mem_cgroup_wait_acct_move(mem_over_limit)) |
2245 | return CHARGE_RETRY; |
2246 | |
2247 | /* If we don't need to call oom-killer at el, return immediately */ |
2248 | if (!oom_check) |
2249 | return CHARGE_NOMEM; |
2250 | /* check OOM */ |
2251 | if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize))) |
2252 | return CHARGE_OOM_DIE; |
2253 | |
2254 | return CHARGE_RETRY; |
2255 | } |
2256 | |
2257 | /* |
2258 | * __mem_cgroup_try_charge() does |
2259 | * 1. detect memcg to be charged against from passed *mm and *ptr, |
2260 | * 2. update res_counter |
2261 | * 3. call memory reclaim if necessary. |
2262 | * |
2263 | * In some special case, if the task is fatal, fatal_signal_pending() or |
2264 | * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup |
2265 | * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon |
2266 | * as possible without any hazards. 2: all pages should have a valid |
2267 | * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg |
2268 | * pointer, that is treated as a charge to root_mem_cgroup. |
2269 | * |
2270 | * So __mem_cgroup_try_charge() will return |
2271 | * 0 ... on success, filling *ptr with a valid memcg pointer. |
2272 | * -ENOMEM ... charge failure because of resource limits. |
2273 | * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup. |
2274 | * |
2275 | * Unlike the exported interface, an "oom" parameter is added. if oom==true, |
2276 | * the oom-killer can be invoked. |
2277 | */ |
2278 | static int __mem_cgroup_try_charge(struct mm_struct *mm, |
2279 | gfp_t gfp_mask, |
2280 | unsigned int nr_pages, |
2281 | struct mem_cgroup **ptr, |
2282 | bool oom) |
2283 | { |
2284 | unsigned int batch = max(CHARGE_BATCH, nr_pages); |
2285 | int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; |
2286 | struct mem_cgroup *memcg = NULL; |
2287 | int ret; |
2288 | |
2289 | /* |
2290 | * Unlike gloval-vm's OOM-kill, we're not in memory shortage |
2291 | * in system level. So, allow to go ahead dying process in addition to |
2292 | * MEMDIE process. |
2293 | */ |
2294 | if (unlikely(test_thread_flag(TIF_MEMDIE) |
2295 | || fatal_signal_pending(current))) |
2296 | goto bypass; |
2297 | |
2298 | /* |
2299 | * We always charge the cgroup the mm_struct belongs to. |
2300 | * The mm_struct's mem_cgroup changes on task migration if the |
2301 | * thread group leader migrates. It's possible that mm is not |
2302 | * set, if so charge the init_mm (happens for pagecache usage). |
2303 | */ |
2304 | if (!*ptr && !mm) |
2305 | *ptr = root_mem_cgroup; |
2306 | again: |
2307 | if (*ptr) { /* css should be a valid one */ |
2308 | memcg = *ptr; |
2309 | VM_BUG_ON(css_is_removed(&memcg->css)); |
2310 | if (mem_cgroup_is_root(memcg)) |
2311 | goto done; |
2312 | if (nr_pages == 1 && consume_stock(memcg)) |
2313 | goto done; |
2314 | css_get(&memcg->css); |
2315 | } else { |
2316 | struct task_struct *p; |
2317 | |
2318 | rcu_read_lock(); |
2319 | p = rcu_dereference(mm->owner); |
2320 | /* |
2321 | * Because we don't have task_lock(), "p" can exit. |
2322 | * In that case, "memcg" can point to root or p can be NULL with |
2323 | * race with swapoff. Then, we have small risk of mis-accouning. |
2324 | * But such kind of mis-account by race always happens because |
2325 | * we don't have cgroup_mutex(). It's overkill and we allo that |
2326 | * small race, here. |
2327 | * (*) swapoff at el will charge against mm-struct not against |
2328 | * task-struct. So, mm->owner can be NULL. |
2329 | */ |
2330 | memcg = mem_cgroup_from_task(p); |
2331 | if (!memcg) |
2332 | memcg = root_mem_cgroup; |
2333 | if (mem_cgroup_is_root(memcg)) { |
2334 | rcu_read_unlock(); |
2335 | goto done; |
2336 | } |
2337 | if (nr_pages == 1 && consume_stock(memcg)) { |
2338 | /* |
2339 | * It seems dagerous to access memcg without css_get(). |
2340 | * But considering how consume_stok works, it's not |
2341 | * necessary. If consume_stock success, some charges |
2342 | * from this memcg are cached on this cpu. So, we |
2343 | * don't need to call css_get()/css_tryget() before |
2344 | * calling consume_stock(). |
2345 | */ |
2346 | rcu_read_unlock(); |
2347 | goto done; |
2348 | } |
2349 | /* after here, we may be blocked. we need to get refcnt */ |
2350 | if (!css_tryget(&memcg->css)) { |
2351 | rcu_read_unlock(); |
2352 | goto again; |
2353 | } |
2354 | rcu_read_unlock(); |
2355 | } |
2356 | |
2357 | do { |
2358 | bool oom_check; |
2359 | |
2360 | /* If killed, bypass charge */ |
2361 | if (fatal_signal_pending(current)) { |
2362 | css_put(&memcg->css); |
2363 | goto bypass; |
2364 | } |
2365 | |
2366 | oom_check = false; |
2367 | if (oom && !nr_oom_retries) { |
2368 | oom_check = true; |
2369 | nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; |
2370 | } |
2371 | |
2372 | ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); |
2373 | switch (ret) { |
2374 | case CHARGE_OK: |
2375 | break; |
2376 | case CHARGE_RETRY: /* not in OOM situation but retry */ |
2377 | batch = nr_pages; |
2378 | css_put(&memcg->css); |
2379 | memcg = NULL; |
2380 | goto again; |
2381 | case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ |
2382 | css_put(&memcg->css); |
2383 | goto nomem; |
2384 | case CHARGE_NOMEM: /* OOM routine works */ |
2385 | if (!oom) { |
2386 | css_put(&memcg->css); |
2387 | goto nomem; |
2388 | } |
2389 | /* If oom, we never return -ENOMEM */ |
2390 | nr_oom_retries--; |
2391 | break; |
2392 | case CHARGE_OOM_DIE: /* Killed by OOM Killer */ |
2393 | css_put(&memcg->css); |
2394 | goto bypass; |
2395 | } |
2396 | } while (ret != CHARGE_OK); |
2397 | |
2398 | if (batch > nr_pages) |
2399 | refill_stock(memcg, batch - nr_pages); |
2400 | css_put(&memcg->css); |
2401 | done: |
2402 | *ptr = memcg; |
2403 | return 0; |
2404 | nomem: |
2405 | *ptr = NULL; |
2406 | return -ENOMEM; |
2407 | bypass: |
2408 | *ptr = root_mem_cgroup; |
2409 | return -EINTR; |
2410 | } |
2411 | |
2412 | /* |
2413 | * Somemtimes we have to undo a charge we got by try_charge(). |
2414 | * This function is for that and do uncharge, put css's refcnt. |
2415 | * gotten by try_charge(). |
2416 | */ |
2417 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, |
2418 | unsigned int nr_pages) |
2419 | { |
2420 | if (!mem_cgroup_is_root(memcg)) { |
2421 | unsigned long bytes = nr_pages * PAGE_SIZE; |
2422 | |
2423 | res_counter_uncharge(&memcg->res, bytes); |
2424 | if (do_swap_account) |
2425 | res_counter_uncharge(&memcg->memsw, bytes); |
2426 | } |
2427 | } |
2428 | |
2429 | /* |
2430 | * A helper function to get mem_cgroup from ID. must be called under |
2431 | * rcu_read_lock(). The caller must check css_is_removed() or some if |
2432 | * it's concern. (dropping refcnt from swap can be called against removed |
2433 | * memcg.) |
2434 | */ |
2435 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) |
2436 | { |
2437 | struct cgroup_subsys_state *css; |
2438 | |
2439 | /* ID 0 is unused ID */ |
2440 | if (!id) |
2441 | return NULL; |
2442 | css = css_lookup(&mem_cgroup_subsys, id); |
2443 | if (!css) |
2444 | return NULL; |
2445 | return container_of(css, struct mem_cgroup, css); |
2446 | } |
2447 | |
2448 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) |
2449 | { |
2450 | struct mem_cgroup *memcg = NULL; |
2451 | struct page_cgroup *pc; |
2452 | unsigned short id; |
2453 | swp_entry_t ent; |
2454 | |
2455 | VM_BUG_ON(!PageLocked(page)); |
2456 | |
2457 | pc = lookup_page_cgroup(page); |
2458 | lock_page_cgroup(pc); |
2459 | if (PageCgroupUsed(pc)) { |
2460 | memcg = pc->mem_cgroup; |
2461 | if (memcg && !css_tryget(&memcg->css)) |
2462 | memcg = NULL; |
2463 | } else if (PageSwapCache(page)) { |
2464 | ent.val = page_private(page); |
2465 | id = lookup_swap_cgroup_id(ent); |
2466 | rcu_read_lock(); |
2467 | memcg = mem_cgroup_lookup(id); |
2468 | if (memcg && !css_tryget(&memcg->css)) |
2469 | memcg = NULL; |
2470 | rcu_read_unlock(); |
2471 | } |
2472 | unlock_page_cgroup(pc); |
2473 | return memcg; |
2474 | } |
2475 | |
2476 | static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, |
2477 | struct page *page, |
2478 | unsigned int nr_pages, |
2479 | enum charge_type ctype, |
2480 | bool lrucare) |
2481 | { |
2482 | struct page_cgroup *pc = lookup_page_cgroup(page); |
2483 | struct zone *uninitialized_var(zone); |
2484 | bool was_on_lru = false; |
2485 | bool anon; |
2486 | |
2487 | lock_page_cgroup(pc); |
2488 | if (unlikely(PageCgroupUsed(pc))) { |
2489 | unlock_page_cgroup(pc); |
2490 | __mem_cgroup_cancel_charge(memcg, nr_pages); |
2491 | return; |
2492 | } |
2493 | /* |
2494 | * we don't need page_cgroup_lock about tail pages, becase they are not |
2495 | * accessed by any other context at this point. |
2496 | */ |
2497 | |
2498 | /* |
2499 | * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page |
2500 | * may already be on some other mem_cgroup's LRU. Take care of it. |
2501 | */ |
2502 | if (lrucare) { |
2503 | zone = page_zone(page); |
2504 | spin_lock_irq(&zone->lru_lock); |
2505 | if (PageLRU(page)) { |
2506 | ClearPageLRU(page); |
2507 | del_page_from_lru_list(zone, page, page_lru(page)); |
2508 | was_on_lru = true; |
2509 | } |
2510 | } |
2511 | |
2512 | pc->mem_cgroup = memcg; |
2513 | /* |
2514 | * We access a page_cgroup asynchronously without lock_page_cgroup(). |
2515 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup |
2516 | * is accessed after testing USED bit. To make pc->mem_cgroup visible |
2517 | * before USED bit, we need memory barrier here. |
2518 | * See mem_cgroup_add_lru_list(), etc. |
2519 | */ |
2520 | smp_wmb(); |
2521 | SetPageCgroupUsed(pc); |
2522 | |
2523 | if (lrucare) { |
2524 | if (was_on_lru) { |
2525 | VM_BUG_ON(PageLRU(page)); |
2526 | SetPageLRU(page); |
2527 | add_page_to_lru_list(zone, page, page_lru(page)); |
2528 | } |
2529 | spin_unlock_irq(&zone->lru_lock); |
2530 | } |
2531 | |
2532 | if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) |
2533 | anon = true; |
2534 | else |
2535 | anon = false; |
2536 | |
2537 | mem_cgroup_charge_statistics(memcg, anon, nr_pages); |
2538 | unlock_page_cgroup(pc); |
2539 | |
2540 | /* |
2541 | * "charge_statistics" updated event counter. Then, check it. |
2542 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. |
2543 | * if they exceeds softlimit. |
2544 | */ |
2545 | memcg_check_events(memcg, page); |
2546 | } |
2547 | |
2548 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
2549 | |
2550 | #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MIGRATION)) |
2551 | /* |
2552 | * Because tail pages are not marked as "used", set it. We're under |
2553 | * zone->lru_lock, 'splitting on pmd' and compound_lock. |
2554 | * charge/uncharge will be never happen and move_account() is done under |
2555 | * compound_lock(), so we don't have to take care of races. |
2556 | */ |
2557 | void mem_cgroup_split_huge_fixup(struct page *head) |
2558 | { |
2559 | struct page_cgroup *head_pc = lookup_page_cgroup(head); |
2560 | struct page_cgroup *pc; |
2561 | int i; |
2562 | |
2563 | if (mem_cgroup_disabled()) |
2564 | return; |
2565 | for (i = 1; i < HPAGE_PMD_NR; i++) { |
2566 | pc = head_pc + i; |
2567 | pc->mem_cgroup = head_pc->mem_cgroup; |
2568 | smp_wmb();/* see __commit_charge() */ |
2569 | pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; |
2570 | } |
2571 | } |
2572 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |
2573 | |
2574 | /** |
2575 | * mem_cgroup_move_account - move account of the page |
2576 | * @page: the page |
2577 | * @nr_pages: number of regular pages (>1 for huge pages) |
2578 | * @pc: page_cgroup of the page. |
2579 | * @from: mem_cgroup which the page is moved from. |
2580 | * @to: mem_cgroup which the page is moved to. @from != @to. |
2581 | * @uncharge: whether we should call uncharge and css_put against @from. |
2582 | * |
2583 | * The caller must confirm following. |
2584 | * - page is not on LRU (isolate_page() is useful.) |
2585 | * - compound_lock is held when nr_pages > 1 |
2586 | * |
2587 | * This function doesn't do "charge" nor css_get to new cgroup. It should be |
2588 | * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is |
2589 | * true, this function does "uncharge" from old cgroup, but it doesn't if |
2590 | * @uncharge is false, so a caller should do "uncharge". |
2591 | */ |
2592 | static int mem_cgroup_move_account(struct page *page, |
2593 | unsigned int nr_pages, |
2594 | struct page_cgroup *pc, |
2595 | struct mem_cgroup *from, |
2596 | struct mem_cgroup *to, |
2597 | bool uncharge) |
2598 | { |
2599 | unsigned long flags; |
2600 | int ret; |
2601 | bool anon = PageAnon(page); |
2602 | |
2603 | VM_BUG_ON(from == to); |
2604 | VM_BUG_ON(PageLRU(page)); |
2605 | /* |
2606 | * The page is isolated from LRU. So, collapse function |
2607 | * will not handle this page. But page splitting can happen. |
2608 | * Do this check under compound_page_lock(). The caller should |
2609 | * hold it. |
2610 | */ |
2611 | ret = -EBUSY; |
2612 | if (nr_pages > 1 && !PageTransHuge(page)) |
2613 | goto out; |
2614 | |
2615 | lock_page_cgroup(pc); |
2616 | |
2617 | ret = -EINVAL; |
2618 | if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) |
2619 | goto unlock; |
2620 | |
2621 | move_lock_mem_cgroup(from, &flags); |
2622 | |
2623 | if (!anon && page_mapped(page)) { |
2624 | /* Update mapped_file data for mem_cgroup */ |
2625 | preempt_disable(); |
2626 | __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
2627 | __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
2628 | preempt_enable(); |
2629 | } |
2630 | mem_cgroup_charge_statistics(from, anon, -nr_pages); |
2631 | if (uncharge) |
2632 | /* This is not "cancel", but cancel_charge does all we need. */ |
2633 | __mem_cgroup_cancel_charge(from, nr_pages); |
2634 | |
2635 | /* caller should have done css_get */ |
2636 | pc->mem_cgroup = to; |
2637 | mem_cgroup_charge_statistics(to, anon, nr_pages); |
2638 | /* |
2639 | * We charges against "to" which may not have any tasks. Then, "to" |
2640 | * can be under rmdir(). But in current implementation, caller of |
2641 | * this function is just force_empty() and move charge, so it's |
2642 | * guaranteed that "to" is never removed. So, we don't check rmdir |
2643 | * status here. |
2644 | */ |
2645 | move_unlock_mem_cgroup(from, &flags); |
2646 | ret = 0; |
2647 | unlock: |
2648 | unlock_page_cgroup(pc); |
2649 | /* |
2650 | * check events |
2651 | */ |
2652 | memcg_check_events(to, page); |
2653 | memcg_check_events(from, page); |
2654 | out: |
2655 | return ret; |
2656 | } |
2657 | |
2658 | /* |
2659 | * move charges to its parent. |
2660 | */ |
2661 | |
2662 | static int mem_cgroup_move_parent(struct page *page, |
2663 | struct page_cgroup *pc, |
2664 | struct mem_cgroup *child, |
2665 | gfp_t gfp_mask) |
2666 | { |
2667 | struct cgroup *cg = child->css.cgroup; |
2668 | struct cgroup *pcg = cg->parent; |
2669 | struct mem_cgroup *parent; |
2670 | unsigned int nr_pages; |
2671 | unsigned long uninitialized_var(flags); |
2672 | int ret; |
2673 | |
2674 | /* Is ROOT ? */ |
2675 | if (!pcg) |
2676 | return -EINVAL; |
2677 | |
2678 | ret = -EBUSY; |
2679 | if (!get_page_unless_zero(page)) |
2680 | goto out; |
2681 | if (isolate_lru_page(page)) |
2682 | goto put; |
2683 | |
2684 | nr_pages = hpage_nr_pages(page); |
2685 | |
2686 | parent = mem_cgroup_from_cont(pcg); |
2687 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); |
2688 | if (ret) |
2689 | goto put_back; |
2690 | |
2691 | if (nr_pages > 1) |
2692 | flags = compound_lock_irqsave(page); |
2693 | |
2694 | ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); |
2695 | if (ret) |
2696 | __mem_cgroup_cancel_charge(parent, nr_pages); |
2697 | |
2698 | if (nr_pages > 1) |
2699 | compound_unlock_irqrestore(page, flags); |
2700 | put_back: |
2701 | putback_lru_page(page); |
2702 | put: |
2703 | put_page(page); |
2704 | out: |
2705 | return ret; |
2706 | } |
2707 | |
2708 | /* |
2709 | * Charge the memory controller for page usage. |
2710 | * Return |
2711 | * 0 if the charge was successful |
2712 | * < 0 if the cgroup is over its limit |
2713 | */ |
2714 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, |
2715 | gfp_t gfp_mask, enum charge_type ctype) |
2716 | { |
2717 | struct mem_cgroup *memcg = NULL; |
2718 | unsigned int nr_pages = 1; |
2719 | bool oom = true; |
2720 | int ret; |
2721 | |
2722 | if (PageTransHuge(page)) { |
2723 | nr_pages <<= compound_order(page); |
2724 | VM_BUG_ON(!PageTransHuge(page)); |
2725 | /* |
2726 | * Never OOM-kill a process for a huge page. The |
2727 | * fault handler will fall back to regular pages. |
2728 | */ |
2729 | oom = false; |
2730 | } |
2731 | |
2732 | ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); |
2733 | if (ret == -ENOMEM) |
2734 | return ret; |
2735 | __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false); |
2736 | return 0; |
2737 | } |
2738 | |
2739 | int mem_cgroup_newpage_charge(struct page *page, |
2740 | struct mm_struct *mm, gfp_t gfp_mask) |
2741 | { |
2742 | if (mem_cgroup_disabled()) |
2743 | return 0; |
2744 | VM_BUG_ON(page_mapped(page)); |
2745 | VM_BUG_ON(page->mapping && !PageAnon(page)); |
2746 | VM_BUG_ON(!mm); |
2747 | return mem_cgroup_charge_common(page, mm, gfp_mask, |
2748 | MEM_CGROUP_CHARGE_TYPE_MAPPED); |
2749 | } |
2750 | |
2751 | static void |
2752 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, |
2753 | enum charge_type ctype); |
2754 | |
2755 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, |
2756 | gfp_t gfp_mask) |
2757 | { |
2758 | struct mem_cgroup *memcg = NULL; |
2759 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; |
2760 | int ret; |
2761 | |
2762 | if (mem_cgroup_disabled()) |
2763 | return 0; |
2764 | if (PageCompound(page)) |
2765 | return 0; |
2766 | |
2767 | if (unlikely(!mm)) |
2768 | mm = &init_mm; |
2769 | if (!page_is_file_cache(page)) |
2770 | type = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
2771 | |
2772 | if (!PageSwapCache(page)) |
2773 | ret = mem_cgroup_charge_common(page, mm, gfp_mask, type); |
2774 | else { /* page is swapcache/shmem */ |
2775 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); |
2776 | if (!ret) |
2777 | __mem_cgroup_commit_charge_swapin(page, memcg, type); |
2778 | } |
2779 | return ret; |
2780 | } |
2781 | |
2782 | /* |
2783 | * While swap-in, try_charge -> commit or cancel, the page is locked. |
2784 | * And when try_charge() successfully returns, one refcnt to memcg without |
2785 | * struct page_cgroup is acquired. This refcnt will be consumed by |
2786 | * "commit()" or removed by "cancel()" |
2787 | */ |
2788 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, |
2789 | struct page *page, |
2790 | gfp_t mask, struct mem_cgroup **memcgp) |
2791 | { |
2792 | struct mem_cgroup *memcg; |
2793 | int ret; |
2794 | |
2795 | *memcgp = NULL; |
2796 | |
2797 | if (mem_cgroup_disabled()) |
2798 | return 0; |
2799 | |
2800 | if (!do_swap_account) |
2801 | goto charge_cur_mm; |
2802 | /* |
2803 | * A racing thread's fault, or swapoff, may have already updated |
2804 | * the pte, and even removed page from swap cache: in those cases |
2805 | * do_swap_page()'s pte_same() test will fail; but there's also a |
2806 | * KSM case which does need to charge the page. |
2807 | */ |
2808 | if (!PageSwapCache(page)) |
2809 | goto charge_cur_mm; |
2810 | memcg = try_get_mem_cgroup_from_page(page); |
2811 | if (!memcg) |
2812 | goto charge_cur_mm; |
2813 | *memcgp = memcg; |
2814 | ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true); |
2815 | css_put(&memcg->css); |
2816 | if (ret == -EINTR) |
2817 | ret = 0; |
2818 | return ret; |
2819 | charge_cur_mm: |
2820 | if (unlikely(!mm)) |
2821 | mm = &init_mm; |
2822 | ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true); |
2823 | if (ret == -EINTR) |
2824 | ret = 0; |
2825 | return ret; |
2826 | } |
2827 | |
2828 | static void |
2829 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, |
2830 | enum charge_type ctype) |
2831 | { |
2832 | if (mem_cgroup_disabled()) |
2833 | return; |
2834 | if (!memcg) |
2835 | return; |
2836 | cgroup_exclude_rmdir(&memcg->css); |
2837 | |
2838 | __mem_cgroup_commit_charge(memcg, page, 1, ctype, true); |
2839 | /* |
2840 | * Now swap is on-memory. This means this page may be |
2841 | * counted both as mem and swap....double count. |
2842 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable |
2843 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() |
2844 | * may call delete_from_swap_cache() before reach here. |
2845 | */ |
2846 | if (do_swap_account && PageSwapCache(page)) { |
2847 | swp_entry_t ent = {.val = page_private(page)}; |
2848 | struct mem_cgroup *swap_memcg; |
2849 | unsigned short id; |
2850 | |
2851 | id = swap_cgroup_record(ent, 0); |
2852 | rcu_read_lock(); |
2853 | swap_memcg = mem_cgroup_lookup(id); |
2854 | if (swap_memcg) { |
2855 | /* |
2856 | * This recorded memcg can be obsolete one. So, avoid |
2857 | * calling css_tryget |
2858 | */ |
2859 | if (!mem_cgroup_is_root(swap_memcg)) |
2860 | res_counter_uncharge(&swap_memcg->memsw, |
2861 | PAGE_SIZE); |
2862 | mem_cgroup_swap_statistics(swap_memcg, false); |
2863 | mem_cgroup_put(swap_memcg); |
2864 | } |
2865 | rcu_read_unlock(); |
2866 | } |
2867 | /* |
2868 | * At swapin, we may charge account against cgroup which has no tasks. |
2869 | * So, rmdir()->pre_destroy() can be called while we do this charge. |
2870 | * In that case, we need to call pre_destroy() again. check it here. |
2871 | */ |
2872 | cgroup_release_and_wakeup_rmdir(&memcg->css); |
2873 | } |
2874 | |
2875 | void mem_cgroup_commit_charge_swapin(struct page *page, |
2876 | struct mem_cgroup *memcg) |
2877 | { |
2878 | __mem_cgroup_commit_charge_swapin(page, memcg, |
2879 | MEM_CGROUP_CHARGE_TYPE_MAPPED); |
2880 | } |
2881 | |
2882 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) |
2883 | { |
2884 | if (mem_cgroup_disabled()) |
2885 | return; |
2886 | if (!memcg) |
2887 | return; |
2888 | __mem_cgroup_cancel_charge(memcg, 1); |
2889 | } |
2890 | |
2891 | static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, |
2892 | unsigned int nr_pages, |
2893 | const enum charge_type ctype) |
2894 | { |
2895 | struct memcg_batch_info *batch = NULL; |
2896 | bool uncharge_memsw = true; |
2897 | |
2898 | /* If swapout, usage of swap doesn't decrease */ |
2899 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) |
2900 | uncharge_memsw = false; |
2901 | |
2902 | batch = ¤t->memcg_batch; |
2903 | /* |
2904 | * In usual, we do css_get() when we remember memcg pointer. |
2905 | * But in this case, we keep res->usage until end of a series of |
2906 | * uncharges. Then, it's ok to ignore memcg's refcnt. |
2907 | */ |
2908 | if (!batch->memcg) |
2909 | batch->memcg = memcg; |
2910 | /* |
2911 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. |
2912 | * In those cases, all pages freed continuously can be expected to be in |
2913 | * the same cgroup and we have chance to coalesce uncharges. |
2914 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) |
2915 | * because we want to do uncharge as soon as possible. |
2916 | */ |
2917 | |
2918 | if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) |
2919 | goto direct_uncharge; |
2920 | |
2921 | if (nr_pages > 1) |
2922 | goto direct_uncharge; |
2923 | |
2924 | /* |
2925 | * In typical case, batch->memcg == mem. This means we can |
2926 | * merge a series of uncharges to an uncharge of res_counter. |
2927 | * If not, we uncharge res_counter ony by one. |
2928 | */ |
2929 | if (batch->memcg != memcg) |
2930 | goto direct_uncharge; |
2931 | /* remember freed charge and uncharge it later */ |
2932 | batch->nr_pages++; |
2933 | if (uncharge_memsw) |
2934 | batch->memsw_nr_pages++; |
2935 | return; |
2936 | direct_uncharge: |
2937 | res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); |
2938 | if (uncharge_memsw) |
2939 | res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); |
2940 | if (unlikely(batch->memcg != memcg)) |
2941 | memcg_oom_recover(memcg); |
2942 | } |
2943 | |
2944 | /* |
2945 | * uncharge if !page_mapped(page) |
2946 | */ |
2947 | static struct mem_cgroup * |
2948 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) |
2949 | { |
2950 | struct mem_cgroup *memcg = NULL; |
2951 | unsigned int nr_pages = 1; |
2952 | struct page_cgroup *pc; |
2953 | bool anon; |
2954 | |
2955 | if (mem_cgroup_disabled()) |
2956 | return NULL; |
2957 | |
2958 | if (PageSwapCache(page)) |
2959 | return NULL; |
2960 | |
2961 | if (PageTransHuge(page)) { |
2962 | nr_pages <<= compound_order(page); |
2963 | VM_BUG_ON(!PageTransHuge(page)); |
2964 | } |
2965 | /* |
2966 | * Check if our page_cgroup is valid |
2967 | */ |
2968 | pc = lookup_page_cgroup(page); |
2969 | if (unlikely(!PageCgroupUsed(pc))) |
2970 | return NULL; |
2971 | |
2972 | lock_page_cgroup(pc); |
2973 | |
2974 | memcg = pc->mem_cgroup; |
2975 | |
2976 | if (!PageCgroupUsed(pc)) |
2977 | goto unlock_out; |
2978 | |
2979 | anon = PageAnon(page); |
2980 | |
2981 | switch (ctype) { |
2982 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: |
2983 | /* |
2984 | * Generally PageAnon tells if it's the anon statistics to be |
2985 | * updated; but sometimes e.g. mem_cgroup_uncharge_page() is |
2986 | * used before page reached the stage of being marked PageAnon. |
2987 | */ |
2988 | anon = true; |
2989 | /* fallthrough */ |
2990 | case MEM_CGROUP_CHARGE_TYPE_DROP: |
2991 | /* See mem_cgroup_prepare_migration() */ |
2992 | if (page_mapped(page) || PageCgroupMigration(pc)) |
2993 | goto unlock_out; |
2994 | break; |
2995 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: |
2996 | if (!PageAnon(page)) { /* Shared memory */ |
2997 | if (page->mapping && !page_is_file_cache(page)) |
2998 | goto unlock_out; |
2999 | } else if (page_mapped(page)) /* Anon */ |
3000 | goto unlock_out; |
3001 | break; |
3002 | default: |
3003 | break; |
3004 | } |
3005 | |
3006 | mem_cgroup_charge_statistics(memcg, anon, -nr_pages); |
3007 | |
3008 | ClearPageCgroupUsed(pc); |
3009 | /* |
3010 | * pc->mem_cgroup is not cleared here. It will be accessed when it's |
3011 | * freed from LRU. This is safe because uncharged page is expected not |
3012 | * to be reused (freed soon). Exception is SwapCache, it's handled by |
3013 | * special functions. |
3014 | */ |
3015 | |
3016 | unlock_page_cgroup(pc); |
3017 | /* |
3018 | * even after unlock, we have memcg->res.usage here and this memcg |
3019 | * will never be freed. |
3020 | */ |
3021 | memcg_check_events(memcg, page); |
3022 | if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { |
3023 | mem_cgroup_swap_statistics(memcg, true); |
3024 | mem_cgroup_get(memcg); |
3025 | } |
3026 | if (!mem_cgroup_is_root(memcg)) |
3027 | mem_cgroup_do_uncharge(memcg, nr_pages, ctype); |
3028 | |
3029 | return memcg; |
3030 | |
3031 | unlock_out: |
3032 | unlock_page_cgroup(pc); |
3033 | return NULL; |
3034 | } |
3035 | |
3036 | void mem_cgroup_uncharge_page(struct page *page) |
3037 | { |
3038 | /* early check. */ |
3039 | if (page_mapped(page)) |
3040 | return; |
3041 | VM_BUG_ON(page->mapping && !PageAnon(page)); |
3042 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); |
3043 | } |
3044 | |
3045 | void mem_cgroup_uncharge_cache_page(struct page *page) |
3046 | { |
3047 | VM_BUG_ON(page_mapped(page)); |
3048 | VM_BUG_ON(page->mapping); |
3049 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); |
3050 | } |
3051 | |
3052 | /* |
3053 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. |
3054 | * In that cases, pages are freed continuously and we can expect pages |
3055 | * are in the same memcg. All these calls itself limits the number of |
3056 | * pages freed at once, then uncharge_start/end() is called properly. |
3057 | * This may be called prural(2) times in a context, |
3058 | */ |
3059 | |
3060 | void mem_cgroup_uncharge_start(void) |
3061 | { |
3062 | current->memcg_batch.do_batch++; |
3063 | /* We can do nest. */ |
3064 | if (current->memcg_batch.do_batch == 1) { |
3065 | current->memcg_batch.memcg = NULL; |
3066 | current->memcg_batch.nr_pages = 0; |
3067 | current->memcg_batch.memsw_nr_pages = 0; |
3068 | } |
3069 | } |
3070 | |
3071 | void mem_cgroup_uncharge_end(void) |
3072 | { |
3073 | struct memcg_batch_info *batch = ¤t->memcg_batch; |
3074 | |
3075 | if (!batch->do_batch) |
3076 | return; |
3077 | |
3078 | batch->do_batch--; |
3079 | if (batch->do_batch) /* If stacked, do nothing. */ |
3080 | return; |
3081 | |
3082 | if (!batch->memcg) |
3083 | return; |
3084 | /* |
3085 | * This "batch->memcg" is valid without any css_get/put etc... |
3086 | * bacause we hide charges behind us. |
3087 | */ |
3088 | if (batch->nr_pages) |
3089 | res_counter_uncharge(&batch->memcg->res, |
3090 | batch->nr_pages * PAGE_SIZE); |
3091 | if (batch->memsw_nr_pages) |
3092 | res_counter_uncharge(&batch->memcg->memsw, |
3093 | batch->memsw_nr_pages * PAGE_SIZE); |
3094 | memcg_oom_recover(batch->memcg); |
3095 | /* forget this pointer (for sanity check) */ |
3096 | batch->memcg = NULL; |
3097 | } |
3098 | |
3099 | #ifdef CONFIG_SWAP |
3100 | /* |
3101 | * called after __delete_from_swap_cache() and drop "page" account. |
3102 | * memcg information is recorded to swap_cgroup of "ent" |
3103 | */ |
3104 | void |
3105 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) |
3106 | { |
3107 | struct mem_cgroup *memcg; |
3108 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; |
3109 | |
3110 | if (!swapout) /* this was a swap cache but the swap is unused ! */ |
3111 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; |
3112 | |
3113 | memcg = __mem_cgroup_uncharge_common(page, ctype); |
3114 | |
3115 | /* |
3116 | * record memcg information, if swapout && memcg != NULL, |
3117 | * mem_cgroup_get() was called in uncharge(). |
3118 | */ |
3119 | if (do_swap_account && swapout && memcg) |
3120 | swap_cgroup_record(ent, css_id(&memcg->css)); |
3121 | } |
3122 | #endif |
3123 | |
3124 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
3125 | /* |
3126 | * called from swap_entry_free(). remove record in swap_cgroup and |
3127 | * uncharge "memsw" account. |
3128 | */ |
3129 | void mem_cgroup_uncharge_swap(swp_entry_t ent) |
3130 | { |
3131 | struct mem_cgroup *memcg; |
3132 | unsigned short id; |
3133 | |
3134 | if (!do_swap_account) |
3135 | return; |
3136 | |
3137 | id = swap_cgroup_record(ent, 0); |
3138 | rcu_read_lock(); |
3139 | memcg = mem_cgroup_lookup(id); |
3140 | if (memcg) { |
3141 | /* |
3142 | * We uncharge this because swap is freed. |
3143 | * This memcg can be obsolete one. We avoid calling css_tryget |
3144 | */ |
3145 | if (!mem_cgroup_is_root(memcg)) |
3146 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); |
3147 | mem_cgroup_swap_statistics(memcg, false); |
3148 | mem_cgroup_put(memcg); |
3149 | } |
3150 | rcu_read_unlock(); |
3151 | } |
3152 | |
3153 | /** |
3154 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. |
3155 | * @entry: swap entry to be moved |
3156 | * @from: mem_cgroup which the entry is moved from |
3157 | * @to: mem_cgroup which the entry is moved to |
3158 | * @need_fixup: whether we should fixup res_counters and refcounts. |
3159 | * |
3160 | * It succeeds only when the swap_cgroup's record for this entry is the same |
3161 | * as the mem_cgroup's id of @from. |
3162 | * |
3163 | * Returns 0 on success, -EINVAL on failure. |
3164 | * |
3165 | * The caller must have charged to @to, IOW, called res_counter_charge() about |
3166 | * both res and memsw, and called css_get(). |
3167 | */ |
3168 | static int mem_cgroup_move_swap_account(swp_entry_t entry, |
3169 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) |
3170 | { |
3171 | unsigned short old_id, new_id; |
3172 | |
3173 | old_id = css_id(&from->css); |
3174 | new_id = css_id(&to->css); |
3175 | |
3176 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { |
3177 | mem_cgroup_swap_statistics(from, false); |
3178 | mem_cgroup_swap_statistics(to, true); |
3179 | /* |
3180 | * This function is only called from task migration context now. |
3181 | * It postpones res_counter and refcount handling till the end |
3182 | * of task migration(mem_cgroup_clear_mc()) for performance |
3183 | * improvement. But we cannot postpone mem_cgroup_get(to) |
3184 | * because if the process that has been moved to @to does |
3185 | * swap-in, the refcount of @to might be decreased to 0. |
3186 | */ |
3187 | mem_cgroup_get(to); |
3188 | if (need_fixup) { |
3189 | if (!mem_cgroup_is_root(from)) |
3190 | res_counter_uncharge(&from->memsw, PAGE_SIZE); |
3191 | mem_cgroup_put(from); |
3192 | /* |
3193 | * we charged both to->res and to->memsw, so we should |
3194 | * uncharge to->res. |
3195 | */ |
3196 | if (!mem_cgroup_is_root(to)) |
3197 | res_counter_uncharge(&to->res, PAGE_SIZE); |
3198 | } |
3199 | return 0; |
3200 | } |
3201 | return -EINVAL; |
3202 | } |
3203 | #else |
3204 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, |
3205 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) |
3206 | { |
3207 | return -EINVAL; |
3208 | } |
3209 | #endif |
3210 | |
3211 | /* |
3212 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old |
3213 | * page belongs to. |
3214 | */ |
3215 | int mem_cgroup_prepare_migration(struct page *page, |
3216 | struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask) |
3217 | { |
3218 | struct mem_cgroup *memcg = NULL; |
3219 | struct page_cgroup *pc; |
3220 | enum charge_type ctype; |
3221 | int ret = 0; |
3222 | |
3223 | *memcgp = NULL; |
3224 | |
3225 | VM_BUG_ON(PageTransHuge(page)); |
3226 | if (mem_cgroup_disabled()) |
3227 | return 0; |
3228 | |
3229 | pc = lookup_page_cgroup(page); |
3230 | lock_page_cgroup(pc); |
3231 | if (PageCgroupUsed(pc)) { |
3232 | memcg = pc->mem_cgroup; |
3233 | css_get(&memcg->css); |
3234 | /* |
3235 | * At migrating an anonymous page, its mapcount goes down |
3236 | * to 0 and uncharge() will be called. But, even if it's fully |
3237 | * unmapped, migration may fail and this page has to be |
3238 | * charged again. We set MIGRATION flag here and delay uncharge |
3239 | * until end_migration() is called |
3240 | * |
3241 | * Corner Case Thinking |
3242 | * A) |
3243 | * When the old page was mapped as Anon and it's unmap-and-freed |
3244 | * while migration was ongoing. |
3245 | * If unmap finds the old page, uncharge() of it will be delayed |
3246 | * until end_migration(). If unmap finds a new page, it's |
3247 | * uncharged when it make mapcount to be 1->0. If unmap code |
3248 | * finds swap_migration_entry, the new page will not be mapped |
3249 | * and end_migration() will find it(mapcount==0). |
3250 | * |
3251 | * B) |
3252 | * When the old page was mapped but migraion fails, the kernel |
3253 | * remaps it. A charge for it is kept by MIGRATION flag even |
3254 | * if mapcount goes down to 0. We can do remap successfully |
3255 | * without charging it again. |
3256 | * |
3257 | * C) |
3258 | * The "old" page is under lock_page() until the end of |
3259 | * migration, so, the old page itself will not be swapped-out. |
3260 | * If the new page is swapped out before end_migraton, our |
3261 | * hook to usual swap-out path will catch the event. |
3262 | */ |
3263 | if (PageAnon(page)) |
3264 | SetPageCgroupMigration(pc); |
3265 | } |
3266 | unlock_page_cgroup(pc); |
3267 | /* |
3268 | * If the page is not charged at this point, |
3269 | * we return here. |
3270 | */ |
3271 | if (!memcg) |
3272 | return 0; |
3273 | |
3274 | *memcgp = memcg; |
3275 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false); |
3276 | css_put(&memcg->css);/* drop extra refcnt */ |
3277 | if (ret) { |
3278 | if (PageAnon(page)) { |
3279 | lock_page_cgroup(pc); |
3280 | ClearPageCgroupMigration(pc); |
3281 | unlock_page_cgroup(pc); |
3282 | /* |
3283 | * The old page may be fully unmapped while we kept it. |
3284 | */ |
3285 | mem_cgroup_uncharge_page(page); |
3286 | } |
3287 | /* we'll need to revisit this error code (we have -EINTR) */ |
3288 | return -ENOMEM; |
3289 | } |
3290 | /* |
3291 | * We charge new page before it's used/mapped. So, even if unlock_page() |
3292 | * is called before end_migration, we can catch all events on this new |
3293 | * page. In the case new page is migrated but not remapped, new page's |
3294 | * mapcount will be finally 0 and we call uncharge in end_migration(). |
3295 | */ |
3296 | if (PageAnon(page)) |
3297 | ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; |
3298 | else if (page_is_file_cache(page)) |
3299 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; |
3300 | else |
3301 | ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
3302 | __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false); |
3303 | return ret; |
3304 | } |
3305 | |
3306 | /* remove redundant charge if migration failed*/ |
3307 | void mem_cgroup_end_migration(struct mem_cgroup *memcg, |
3308 | struct page *oldpage, struct page *newpage, bool migration_ok) |
3309 | { |
3310 | struct page *used, *unused; |
3311 | struct page_cgroup *pc; |
3312 | bool anon; |
3313 | |
3314 | if (!memcg) |
3315 | return; |
3316 | /* blocks rmdir() */ |
3317 | cgroup_exclude_rmdir(&memcg->css); |
3318 | if (!migration_ok) { |
3319 | used = oldpage; |
3320 | unused = newpage; |
3321 | } else { |
3322 | used = newpage; |
3323 | unused = oldpage; |
3324 | } |
3325 | /* |
3326 | * We disallowed uncharge of pages under migration because mapcount |
3327 | * of the page goes down to zero, temporarly. |
3328 | * Clear the flag and check the page should be charged. |
3329 | */ |
3330 | pc = lookup_page_cgroup(oldpage); |
3331 | lock_page_cgroup(pc); |
3332 | ClearPageCgroupMigration(pc); |
3333 | unlock_page_cgroup(pc); |
3334 | anon = PageAnon(used); |
3335 | __mem_cgroup_uncharge_common(unused, |
3336 | anon ? MEM_CGROUP_CHARGE_TYPE_MAPPED |
3337 | : MEM_CGROUP_CHARGE_TYPE_CACHE); |
3338 | |
3339 | /* |
3340 | * If a page is a file cache, radix-tree replacement is very atomic |
3341 | * and we can skip this check. When it was an Anon page, its mapcount |
3342 | * goes down to 0. But because we added MIGRATION flage, it's not |
3343 | * uncharged yet. There are several case but page->mapcount check |
3344 | * and USED bit check in mem_cgroup_uncharge_page() will do enough |
3345 | * check. (see prepare_charge() also) |
3346 | */ |
3347 | if (anon) |
3348 | mem_cgroup_uncharge_page(used); |
3349 | /* |
3350 | * At migration, we may charge account against cgroup which has no |
3351 | * tasks. |
3352 | * So, rmdir()->pre_destroy() can be called while we do this charge. |
3353 | * In that case, we need to call pre_destroy() again. check it here. |
3354 | */ |
3355 | cgroup_release_and_wakeup_rmdir(&memcg->css); |
3356 | } |
3357 | |
3358 | /* |
3359 | * At replace page cache, newpage is not under any memcg but it's on |
3360 | * LRU. So, this function doesn't touch res_counter but handles LRU |
3361 | * in correct way. Both pages are locked so we cannot race with uncharge. |
3362 | */ |
3363 | void mem_cgroup_replace_page_cache(struct page *oldpage, |
3364 | struct page *newpage) |
3365 | { |
3366 | struct mem_cgroup *memcg; |
3367 | struct page_cgroup *pc; |
3368 | enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; |
3369 | |
3370 | if (mem_cgroup_disabled()) |
3371 | return; |
3372 | |
3373 | pc = lookup_page_cgroup(oldpage); |
3374 | /* fix accounting on old pages */ |
3375 | lock_page_cgroup(pc); |
3376 | memcg = pc->mem_cgroup; |
3377 | mem_cgroup_charge_statistics(memcg, false, -1); |
3378 | ClearPageCgroupUsed(pc); |
3379 | unlock_page_cgroup(pc); |
3380 | |
3381 | if (PageSwapBacked(oldpage)) |
3382 | type = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
3383 | |
3384 | /* |
3385 | * Even if newpage->mapping was NULL before starting replacement, |
3386 | * the newpage may be on LRU(or pagevec for LRU) already. We lock |
3387 | * LRU while we overwrite pc->mem_cgroup. |
3388 | */ |
3389 | __mem_cgroup_commit_charge(memcg, newpage, 1, type, true); |
3390 | } |
3391 | |
3392 | #ifdef CONFIG_DEBUG_VM |
3393 | static struct page_cgroup *lookup_page_cgroup_used(struct page *page) |
3394 | { |
3395 | struct page_cgroup *pc; |
3396 | |
3397 | pc = lookup_page_cgroup(page); |
3398 | /* |
3399 | * Can be NULL while feeding pages into the page allocator for |
3400 | * the first time, i.e. during boot or memory hotplug; |
3401 | * or when mem_cgroup_disabled(). |
3402 | */ |
3403 | if (likely(pc) && PageCgroupUsed(pc)) |
3404 | return pc; |
3405 | return NULL; |
3406 | } |
3407 | |
3408 | bool mem_cgroup_bad_page_check(struct page *page) |
3409 | { |
3410 | if (mem_cgroup_disabled()) |
3411 | return false; |
3412 | |
3413 | return lookup_page_cgroup_used(page) != NULL; |
3414 | } |
3415 | |
3416 | void mem_cgroup_print_bad_page(struct page *page) |
3417 | { |
3418 | struct page_cgroup *pc; |
3419 | |
3420 | pc = lookup_page_cgroup_used(page); |
3421 | if (pc) { |
3422 | printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n", |
3423 | pc, pc->flags, pc->mem_cgroup); |
3424 | } |
3425 | } |
3426 | #endif |
3427 | |
3428 | static DEFINE_MUTEX(set_limit_mutex); |
3429 | |
3430 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, |
3431 | unsigned long long val) |
3432 | { |
3433 | int retry_count; |
3434 | u64 memswlimit, memlimit; |
3435 | int ret = 0; |
3436 | int children = mem_cgroup_count_children(memcg); |
3437 | u64 curusage, oldusage; |
3438 | int enlarge; |
3439 | |
3440 | /* |
3441 | * For keeping hierarchical_reclaim simple, how long we should retry |
3442 | * is depends on callers. We set our retry-count to be function |
3443 | * of # of children which we should visit in this loop. |
3444 | */ |
3445 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; |
3446 | |
3447 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); |
3448 | |
3449 | enlarge = 0; |
3450 | while (retry_count) { |
3451 | if (signal_pending(current)) { |
3452 | ret = -EINTR; |
3453 | break; |
3454 | } |
3455 | /* |
3456 | * Rather than hide all in some function, I do this in |
3457 | * open coded manner. You see what this really does. |
3458 | * We have to guarantee memcg->res.limit < memcg->memsw.limit. |
3459 | */ |
3460 | mutex_lock(&set_limit_mutex); |
3461 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
3462 | if (memswlimit < val) { |
3463 | ret = -EINVAL; |
3464 | mutex_unlock(&set_limit_mutex); |
3465 | break; |
3466 | } |
3467 | |
3468 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
3469 | if (memlimit < val) |
3470 | enlarge = 1; |
3471 | |
3472 | ret = res_counter_set_limit(&memcg->res, val); |
3473 | if (!ret) { |
3474 | if (memswlimit == val) |
3475 | memcg->memsw_is_minimum = true; |
3476 | else |
3477 | memcg->memsw_is_minimum = false; |
3478 | } |
3479 | mutex_unlock(&set_limit_mutex); |
3480 | |
3481 | if (!ret) |
3482 | break; |
3483 | |
3484 | mem_cgroup_reclaim(memcg, GFP_KERNEL, |
3485 | MEM_CGROUP_RECLAIM_SHRINK); |
3486 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); |
3487 | /* Usage is reduced ? */ |
3488 | if (curusage >= oldusage) |
3489 | retry_count--; |
3490 | else |
3491 | oldusage = curusage; |
3492 | } |
3493 | if (!ret && enlarge) |
3494 | memcg_oom_recover(memcg); |
3495 | |
3496 | return ret; |
3497 | } |
3498 | |
3499 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, |
3500 | unsigned long long val) |
3501 | { |
3502 | int retry_count; |
3503 | u64 memlimit, memswlimit, oldusage, curusage; |
3504 | int children = mem_cgroup_count_children(memcg); |
3505 | int ret = -EBUSY; |
3506 | int enlarge = 0; |
3507 | |
3508 | /* see mem_cgroup_resize_res_limit */ |
3509 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; |
3510 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); |
3511 | while (retry_count) { |
3512 | if (signal_pending(current)) { |
3513 | ret = -EINTR; |
3514 | break; |
3515 | } |
3516 | /* |
3517 | * Rather than hide all in some function, I do this in |
3518 | * open coded manner. You see what this really does. |
3519 | * We have to guarantee memcg->res.limit < memcg->memsw.limit. |
3520 | */ |
3521 | mutex_lock(&set_limit_mutex); |
3522 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
3523 | if (memlimit > val) { |
3524 | ret = -EINVAL; |
3525 | mutex_unlock(&set_limit_mutex); |
3526 | break; |
3527 | } |
3528 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
3529 | if (memswlimit < val) |
3530 | enlarge = 1; |
3531 | ret = res_counter_set_limit(&memcg->memsw, val); |
3532 | if (!ret) { |
3533 | if (memlimit == val) |
3534 | memcg->memsw_is_minimum = true; |
3535 | else |
3536 | memcg->memsw_is_minimum = false; |
3537 | } |
3538 | mutex_unlock(&set_limit_mutex); |
3539 | |
3540 | if (!ret) |
3541 | break; |
3542 | |
3543 | mem_cgroup_reclaim(memcg, GFP_KERNEL, |
3544 | MEM_CGROUP_RECLAIM_NOSWAP | |
3545 | MEM_CGROUP_RECLAIM_SHRINK); |
3546 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); |
3547 | /* Usage is reduced ? */ |
3548 | if (curusage >= oldusage) |
3549 | retry_count--; |
3550 | else |
3551 | oldusage = curusage; |
3552 | } |
3553 | if (!ret && enlarge) |
3554 | memcg_oom_recover(memcg); |
3555 | return ret; |
3556 | } |
3557 | |
3558 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, |
3559 | gfp_t gfp_mask, |
3560 | unsigned long *total_scanned) |
3561 | { |
3562 | unsigned long nr_reclaimed = 0; |
3563 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; |
3564 | unsigned long reclaimed; |
3565 | int loop = 0; |
3566 | struct mem_cgroup_tree_per_zone *mctz; |
3567 | unsigned long long excess; |
3568 | unsigned long nr_scanned; |
3569 | |
3570 | if (order > 0) |
3571 | return 0; |
3572 | |
3573 | mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); |
3574 | /* |
3575 | * This loop can run a while, specially if mem_cgroup's continuously |
3576 | * keep exceeding their soft limit and putting the system under |
3577 | * pressure |
3578 | */ |
3579 | do { |
3580 | if (next_mz) |
3581 | mz = next_mz; |
3582 | else |
3583 | mz = mem_cgroup_largest_soft_limit_node(mctz); |
3584 | if (!mz) |
3585 | break; |
3586 | |
3587 | nr_scanned = 0; |
3588 | reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone, |
3589 | gfp_mask, &nr_scanned); |
3590 | nr_reclaimed += reclaimed; |
3591 | *total_scanned += nr_scanned; |
3592 | spin_lock(&mctz->lock); |
3593 | |
3594 | /* |
3595 | * If we failed to reclaim anything from this memory cgroup |
3596 | * it is time to move on to the next cgroup |
3597 | */ |
3598 | next_mz = NULL; |
3599 | if (!reclaimed) { |
3600 | do { |
3601 | /* |
3602 | * Loop until we find yet another one. |
3603 | * |
3604 | * By the time we get the soft_limit lock |
3605 | * again, someone might have aded the |
3606 | * group back on the RB tree. Iterate to |
3607 | * make sure we get a different mem. |
3608 | * mem_cgroup_largest_soft_limit_node returns |
3609 | * NULL if no other cgroup is present on |
3610 | * the tree |
3611 | */ |
3612 | next_mz = |
3613 | __mem_cgroup_largest_soft_limit_node(mctz); |
3614 | if (next_mz == mz) |
3615 | css_put(&next_mz->memcg->css); |
3616 | else /* next_mz == NULL or other memcg */ |
3617 | break; |
3618 | } while (1); |
3619 | } |
3620 | __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz); |
3621 | excess = res_counter_soft_limit_excess(&mz->memcg->res); |
3622 | /* |
3623 | * One school of thought says that we should not add |
3624 | * back the node to the tree if reclaim returns 0. |
3625 | * But our reclaim could return 0, simply because due |
3626 | * to priority we are exposing a smaller subset of |
3627 | * memory to reclaim from. Consider this as a longer |
3628 | * term TODO. |
3629 | */ |
3630 | /* If excess == 0, no tree ops */ |
3631 | __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess); |
3632 | spin_unlock(&mctz->lock); |
3633 | css_put(&mz->memcg->css); |
3634 | loop++; |
3635 | /* |
3636 | * Could not reclaim anything and there are no more |
3637 | * mem cgroups to try or we seem to be looping without |
3638 | * reclaiming anything. |
3639 | */ |
3640 | if (!nr_reclaimed && |
3641 | (next_mz == NULL || |
3642 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) |
3643 | break; |
3644 | } while (!nr_reclaimed); |
3645 | if (next_mz) |
3646 | css_put(&next_mz->memcg->css); |
3647 | return nr_reclaimed; |
3648 | } |
3649 | |
3650 | /* |
3651 | * This routine traverse page_cgroup in given list and drop them all. |
3652 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. |
3653 | */ |
3654 | static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, |
3655 | int node, int zid, enum lru_list lru) |
3656 | { |
3657 | struct mem_cgroup_per_zone *mz; |
3658 | unsigned long flags, loop; |
3659 | struct list_head *list; |
3660 | struct page *busy; |
3661 | struct zone *zone; |
3662 | int ret = 0; |
3663 | |
3664 | zone = &NODE_DATA(node)->node_zones[zid]; |
3665 | mz = mem_cgroup_zoneinfo(memcg, node, zid); |
3666 | list = &mz->lruvec.lists[lru]; |
3667 | |
3668 | loop = mz->lru_size[lru]; |
3669 | /* give some margin against EBUSY etc...*/ |
3670 | loop += 256; |
3671 | busy = NULL; |
3672 | while (loop--) { |
3673 | struct page_cgroup *pc; |
3674 | struct page *page; |
3675 | |
3676 | ret = 0; |
3677 | spin_lock_irqsave(&zone->lru_lock, flags); |
3678 | if (list_empty(list)) { |
3679 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
3680 | break; |
3681 | } |
3682 | page = list_entry(list->prev, struct page, lru); |
3683 | if (busy == page) { |
3684 | list_move(&page->lru, list); |
3685 | busy = NULL; |
3686 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
3687 | continue; |
3688 | } |
3689 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
3690 | |
3691 | pc = lookup_page_cgroup(page); |
3692 | |
3693 | ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); |
3694 | if (ret == -ENOMEM || ret == -EINTR) |
3695 | break; |
3696 | |
3697 | if (ret == -EBUSY || ret == -EINVAL) { |
3698 | /* found lock contention or "pc" is obsolete. */ |
3699 | busy = page; |
3700 | cond_resched(); |
3701 | } else |
3702 | busy = NULL; |
3703 | } |
3704 | |
3705 | if (!ret && !list_empty(list)) |
3706 | return -EBUSY; |
3707 | return ret; |
3708 | } |
3709 | |
3710 | /* |
3711 | * make mem_cgroup's charge to be 0 if there is no task. |
3712 | * This enables deleting this mem_cgroup. |
3713 | */ |
3714 | static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) |
3715 | { |
3716 | int ret; |
3717 | int node, zid, shrink; |
3718 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; |
3719 | struct cgroup *cgrp = memcg->css.cgroup; |
3720 | |
3721 | css_get(&memcg->css); |
3722 | |
3723 | shrink = 0; |
3724 | /* should free all ? */ |
3725 | if (free_all) |
3726 | goto try_to_free; |
3727 | move_account: |
3728 | do { |
3729 | ret = -EBUSY; |
3730 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) |
3731 | goto out; |
3732 | ret = -EINTR; |
3733 | if (signal_pending(current)) |
3734 | goto out; |
3735 | /* This is for making all *used* pages to be on LRU. */ |
3736 | lru_add_drain_all(); |
3737 | drain_all_stock_sync(memcg); |
3738 | ret = 0; |
3739 | mem_cgroup_start_move(memcg); |
3740 | for_each_node_state(node, N_HIGH_MEMORY) { |
3741 | for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { |
3742 | enum lru_list lru; |
3743 | for_each_lru(lru) { |
3744 | ret = mem_cgroup_force_empty_list(memcg, |
3745 | node, zid, lru); |
3746 | if (ret) |
3747 | break; |
3748 | } |
3749 | } |
3750 | if (ret) |
3751 | break; |
3752 | } |
3753 | mem_cgroup_end_move(memcg); |
3754 | memcg_oom_recover(memcg); |
3755 | /* it seems parent cgroup doesn't have enough mem */ |
3756 | if (ret == -ENOMEM) |
3757 | goto try_to_free; |
3758 | cond_resched(); |
3759 | /* "ret" should also be checked to ensure all lists are empty. */ |
3760 | } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret); |
3761 | out: |
3762 | css_put(&memcg->css); |
3763 | return ret; |
3764 | |
3765 | try_to_free: |
3766 | /* returns EBUSY if there is a task or if we come here twice. */ |
3767 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { |
3768 | ret = -EBUSY; |
3769 | goto out; |
3770 | } |
3771 | /* we call try-to-free pages for make this cgroup empty */ |
3772 | lru_add_drain_all(); |
3773 | /* try to free all pages in this cgroup */ |
3774 | shrink = 1; |
3775 | while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) { |
3776 | int progress; |
3777 | |
3778 | if (signal_pending(current)) { |
3779 | ret = -EINTR; |
3780 | goto out; |
3781 | } |
3782 | progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, |
3783 | false); |
3784 | if (!progress) { |
3785 | nr_retries--; |
3786 | /* maybe some writeback is necessary */ |
3787 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
3788 | } |
3789 | |
3790 | } |
3791 | lru_add_drain(); |
3792 | /* try move_account...there may be some *locked* pages. */ |
3793 | goto move_account; |
3794 | } |
3795 | |
3796 | int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) |
3797 | { |
3798 | return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); |
3799 | } |
3800 | |
3801 | |
3802 | static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) |
3803 | { |
3804 | return mem_cgroup_from_cont(cont)->use_hierarchy; |
3805 | } |
3806 | |
3807 | static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, |
3808 | u64 val) |
3809 | { |
3810 | int retval = 0; |
3811 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
3812 | struct cgroup *parent = cont->parent; |
3813 | struct mem_cgroup *parent_memcg = NULL; |
3814 | |
3815 | if (parent) |
3816 | parent_memcg = mem_cgroup_from_cont(parent); |
3817 | |
3818 | cgroup_lock(); |
3819 | /* |
3820 | * If parent's use_hierarchy is set, we can't make any modifications |
3821 | * in the child subtrees. If it is unset, then the change can |
3822 | * occur, provided the current cgroup has no children. |
3823 | * |
3824 | * For the root cgroup, parent_mem is NULL, we allow value to be |
3825 | * set if there are no children. |
3826 | */ |
3827 | if ((!parent_memcg || !parent_memcg->use_hierarchy) && |
3828 | (val == 1 || val == 0)) { |
3829 | if (list_empty(&cont->children)) |
3830 | memcg->use_hierarchy = val; |
3831 | else |
3832 | retval = -EBUSY; |
3833 | } else |
3834 | retval = -EINVAL; |
3835 | cgroup_unlock(); |
3836 | |
3837 | return retval; |
3838 | } |
3839 | |
3840 | |
3841 | static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, |
3842 | enum mem_cgroup_stat_index idx) |
3843 | { |
3844 | struct mem_cgroup *iter; |
3845 | long val = 0; |
3846 | |
3847 | /* Per-cpu values can be negative, use a signed accumulator */ |
3848 | for_each_mem_cgroup_tree(iter, memcg) |
3849 | val += mem_cgroup_read_stat(iter, idx); |
3850 | |
3851 | if (val < 0) /* race ? */ |
3852 | val = 0; |
3853 | return val; |
3854 | } |
3855 | |
3856 | static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) |
3857 | { |
3858 | u64 val; |
3859 | |
3860 | if (!mem_cgroup_is_root(memcg)) { |
3861 | if (!swap) |
3862 | return res_counter_read_u64(&memcg->res, RES_USAGE); |
3863 | else |
3864 | return res_counter_read_u64(&memcg->memsw, RES_USAGE); |
3865 | } |
3866 | |
3867 | val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); |
3868 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); |
3869 | |
3870 | if (swap) |
3871 | val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); |
3872 | |
3873 | return val << PAGE_SHIFT; |
3874 | } |
3875 | |
3876 | static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) |
3877 | { |
3878 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
3879 | u64 val; |
3880 | int type, name; |
3881 | |
3882 | type = MEMFILE_TYPE(cft->private); |
3883 | name = MEMFILE_ATTR(cft->private); |
3884 | switch (type) { |
3885 | case _MEM: |
3886 | if (name == RES_USAGE) |
3887 | val = mem_cgroup_usage(memcg, false); |
3888 | else |
3889 | val = res_counter_read_u64(&memcg->res, name); |
3890 | break; |
3891 | case _MEMSWAP: |
3892 | if (name == RES_USAGE) |
3893 | val = mem_cgroup_usage(memcg, true); |
3894 | else |
3895 | val = res_counter_read_u64(&memcg->memsw, name); |
3896 | break; |
3897 | default: |
3898 | BUG(); |
3899 | } |
3900 | return val; |
3901 | } |
3902 | /* |
3903 | * The user of this function is... |
3904 | * RES_LIMIT. |
3905 | */ |
3906 | static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, |
3907 | const char *buffer) |
3908 | { |
3909 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
3910 | int type, name; |
3911 | unsigned long long val; |
3912 | int ret; |
3913 | |
3914 | type = MEMFILE_TYPE(cft->private); |
3915 | name = MEMFILE_ATTR(cft->private); |
3916 | switch (name) { |
3917 | case RES_LIMIT: |
3918 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ |
3919 | ret = -EINVAL; |
3920 | break; |
3921 | } |
3922 | /* This function does all necessary parse...reuse it */ |
3923 | ret = res_counter_memparse_write_strategy(buffer, &val); |
3924 | if (ret) |
3925 | break; |
3926 | if (type == _MEM) |
3927 | ret = mem_cgroup_resize_limit(memcg, val); |
3928 | else |
3929 | ret = mem_cgroup_resize_memsw_limit(memcg, val); |
3930 | break; |
3931 | case RES_SOFT_LIMIT: |
3932 | ret = res_counter_memparse_write_strategy(buffer, &val); |
3933 | if (ret) |
3934 | break; |
3935 | /* |
3936 | * For memsw, soft limits are hard to implement in terms |
3937 | * of semantics, for now, we support soft limits for |
3938 | * control without swap |
3939 | */ |
3940 | if (type == _MEM) |
3941 | ret = res_counter_set_soft_limit(&memcg->res, val); |
3942 | else |
3943 | ret = -EINVAL; |
3944 | break; |
3945 | default: |
3946 | ret = -EINVAL; /* should be BUG() ? */ |
3947 | break; |
3948 | } |
3949 | return ret; |
3950 | } |
3951 | |
3952 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, |
3953 | unsigned long long *mem_limit, unsigned long long *memsw_limit) |
3954 | { |
3955 | struct cgroup *cgroup; |
3956 | unsigned long long min_limit, min_memsw_limit, tmp; |
3957 | |
3958 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
3959 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
3960 | cgroup = memcg->css.cgroup; |
3961 | if (!memcg->use_hierarchy) |
3962 | goto out; |
3963 | |
3964 | while (cgroup->parent) { |
3965 | cgroup = cgroup->parent; |
3966 | memcg = mem_cgroup_from_cont(cgroup); |
3967 | if (!memcg->use_hierarchy) |
3968 | break; |
3969 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); |
3970 | min_limit = min(min_limit, tmp); |
3971 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
3972 | min_memsw_limit = min(min_memsw_limit, tmp); |
3973 | } |
3974 | out: |
3975 | *mem_limit = min_limit; |
3976 | *memsw_limit = min_memsw_limit; |
3977 | } |
3978 | |
3979 | static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) |
3980 | { |
3981 | struct mem_cgroup *memcg; |
3982 | int type, name; |
3983 | |
3984 | memcg = mem_cgroup_from_cont(cont); |
3985 | type = MEMFILE_TYPE(event); |
3986 | name = MEMFILE_ATTR(event); |
3987 | switch (name) { |
3988 | case RES_MAX_USAGE: |
3989 | if (type == _MEM) |
3990 | res_counter_reset_max(&memcg->res); |
3991 | else |
3992 | res_counter_reset_max(&memcg->memsw); |
3993 | break; |
3994 | case RES_FAILCNT: |
3995 | if (type == _MEM) |
3996 | res_counter_reset_failcnt(&memcg->res); |
3997 | else |
3998 | res_counter_reset_failcnt(&memcg->memsw); |
3999 | break; |
4000 | } |
4001 | |
4002 | return 0; |
4003 | } |
4004 | |
4005 | static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, |
4006 | struct cftype *cft) |
4007 | { |
4008 | return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; |
4009 | } |
4010 | |
4011 | #ifdef CONFIG_MMU |
4012 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, |
4013 | struct cftype *cft, u64 val) |
4014 | { |
4015 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4016 | |
4017 | if (val >= (1 << NR_MOVE_TYPE)) |
4018 | return -EINVAL; |
4019 | /* |
4020 | * We check this value several times in both in can_attach() and |
4021 | * attach(), so we need cgroup lock to prevent this value from being |
4022 | * inconsistent. |
4023 | */ |
4024 | cgroup_lock(); |
4025 | memcg->move_charge_at_immigrate = val; |
4026 | cgroup_unlock(); |
4027 | |
4028 | return 0; |
4029 | } |
4030 | #else |
4031 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, |
4032 | struct cftype *cft, u64 val) |
4033 | { |
4034 | return -ENOSYS; |
4035 | } |
4036 | #endif |
4037 | |
4038 | |
4039 | /* For read statistics */ |
4040 | enum { |
4041 | MCS_CACHE, |
4042 | MCS_RSS, |
4043 | MCS_FILE_MAPPED, |
4044 | MCS_PGPGIN, |
4045 | MCS_PGPGOUT, |
4046 | MCS_SWAP, |
4047 | MCS_PGFAULT, |
4048 | MCS_PGMAJFAULT, |
4049 | MCS_INACTIVE_ANON, |
4050 | MCS_ACTIVE_ANON, |
4051 | MCS_INACTIVE_FILE, |
4052 | MCS_ACTIVE_FILE, |
4053 | MCS_UNEVICTABLE, |
4054 | NR_MCS_STAT, |
4055 | }; |
4056 | |
4057 | struct mcs_total_stat { |
4058 | s64 stat[NR_MCS_STAT]; |
4059 | }; |
4060 | |
4061 | struct { |
4062 | char *local_name; |
4063 | char *total_name; |
4064 | } memcg_stat_strings[NR_MCS_STAT] = { |
4065 | {"cache", "total_cache"}, |
4066 | {"rss", "total_rss"}, |
4067 | {"mapped_file", "total_mapped_file"}, |
4068 | {"pgpgin", "total_pgpgin"}, |
4069 | {"pgpgout", "total_pgpgout"}, |
4070 | {"swap", "total_swap"}, |
4071 | {"pgfault", "total_pgfault"}, |
4072 | {"pgmajfault", "total_pgmajfault"}, |
4073 | {"inactive_anon", "total_inactive_anon"}, |
4074 | {"active_anon", "total_active_anon"}, |
4075 | {"inactive_file", "total_inactive_file"}, |
4076 | {"active_file", "total_active_file"}, |
4077 | {"unevictable", "total_unevictable"} |
4078 | }; |
4079 | |
4080 | |
4081 | static void |
4082 | mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) |
4083 | { |
4084 | s64 val; |
4085 | |
4086 | /* per cpu stat */ |
4087 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); |
4088 | s->stat[MCS_CACHE] += val * PAGE_SIZE; |
4089 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); |
4090 | s->stat[MCS_RSS] += val * PAGE_SIZE; |
4091 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); |
4092 | s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; |
4093 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); |
4094 | s->stat[MCS_PGPGIN] += val; |
4095 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); |
4096 | s->stat[MCS_PGPGOUT] += val; |
4097 | if (do_swap_account) { |
4098 | val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); |
4099 | s->stat[MCS_SWAP] += val * PAGE_SIZE; |
4100 | } |
4101 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); |
4102 | s->stat[MCS_PGFAULT] += val; |
4103 | val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); |
4104 | s->stat[MCS_PGMAJFAULT] += val; |
4105 | |
4106 | /* per zone stat */ |
4107 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); |
4108 | s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; |
4109 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); |
4110 | s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; |
4111 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); |
4112 | s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; |
4113 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); |
4114 | s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; |
4115 | val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); |
4116 | s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; |
4117 | } |
4118 | |
4119 | static void |
4120 | mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) |
4121 | { |
4122 | struct mem_cgroup *iter; |
4123 | |
4124 | for_each_mem_cgroup_tree(iter, memcg) |
4125 | mem_cgroup_get_local_stat(iter, s); |
4126 | } |
4127 | |
4128 | #ifdef CONFIG_NUMA |
4129 | static int mem_control_numa_stat_show(struct seq_file *m, void *arg) |
4130 | { |
4131 | int nid; |
4132 | unsigned long total_nr, file_nr, anon_nr, unevictable_nr; |
4133 | unsigned long node_nr; |
4134 | struct cgroup *cont = m->private; |
4135 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
4136 | |
4137 | total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL); |
4138 | seq_printf(m, "total=%lu", total_nr); |
4139 | for_each_node_state(nid, N_HIGH_MEMORY) { |
4140 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL); |
4141 | seq_printf(m, " N%d=%lu", nid, node_nr); |
4142 | } |
4143 | seq_putc(m, '\n'); |
4144 | |
4145 | file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE); |
4146 | seq_printf(m, "file=%lu", file_nr); |
4147 | for_each_node_state(nid, N_HIGH_MEMORY) { |
4148 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, |
4149 | LRU_ALL_FILE); |
4150 | seq_printf(m, " N%d=%lu", nid, node_nr); |
4151 | } |
4152 | seq_putc(m, '\n'); |
4153 | |
4154 | anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON); |
4155 | seq_printf(m, "anon=%lu", anon_nr); |
4156 | for_each_node_state(nid, N_HIGH_MEMORY) { |
4157 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, |
4158 | LRU_ALL_ANON); |
4159 | seq_printf(m, " N%d=%lu", nid, node_nr); |
4160 | } |
4161 | seq_putc(m, '\n'); |
4162 | |
4163 | unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); |
4164 | seq_printf(m, "unevictable=%lu", unevictable_nr); |
4165 | for_each_node_state(nid, N_HIGH_MEMORY) { |
4166 | node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, |
4167 | BIT(LRU_UNEVICTABLE)); |
4168 | seq_printf(m, " N%d=%lu", nid, node_nr); |
4169 | } |
4170 | seq_putc(m, '\n'); |
4171 | return 0; |
4172 | } |
4173 | #endif /* CONFIG_NUMA */ |
4174 | |
4175 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, |
4176 | struct cgroup_map_cb *cb) |
4177 | { |
4178 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
4179 | struct mcs_total_stat mystat; |
4180 | int i; |
4181 | |
4182 | memset(&mystat, 0, sizeof(mystat)); |
4183 | mem_cgroup_get_local_stat(memcg, &mystat); |
4184 | |
4185 | |
4186 | for (i = 0; i < NR_MCS_STAT; i++) { |
4187 | if (i == MCS_SWAP && !do_swap_account) |
4188 | continue; |
4189 | cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); |
4190 | } |
4191 | |
4192 | /* Hierarchical information */ |
4193 | { |
4194 | unsigned long long limit, memsw_limit; |
4195 | memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit); |
4196 | cb->fill(cb, "hierarchical_memory_limit", limit); |
4197 | if (do_swap_account) |
4198 | cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); |
4199 | } |
4200 | |
4201 | memset(&mystat, 0, sizeof(mystat)); |
4202 | mem_cgroup_get_total_stat(memcg, &mystat); |
4203 | for (i = 0; i < NR_MCS_STAT; i++) { |
4204 | if (i == MCS_SWAP && !do_swap_account) |
4205 | continue; |
4206 | cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); |
4207 | } |
4208 | |
4209 | #ifdef CONFIG_DEBUG_VM |
4210 | { |
4211 | int nid, zid; |
4212 | struct mem_cgroup_per_zone *mz; |
4213 | unsigned long recent_rotated[2] = {0, 0}; |
4214 | unsigned long recent_scanned[2] = {0, 0}; |
4215 | |
4216 | for_each_online_node(nid) |
4217 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
4218 | mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
4219 | |
4220 | recent_rotated[0] += |
4221 | mz->reclaim_stat.recent_rotated[0]; |
4222 | recent_rotated[1] += |
4223 | mz->reclaim_stat.recent_rotated[1]; |
4224 | recent_scanned[0] += |
4225 | mz->reclaim_stat.recent_scanned[0]; |
4226 | recent_scanned[1] += |
4227 | mz->reclaim_stat.recent_scanned[1]; |
4228 | } |
4229 | cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); |
4230 | cb->fill(cb, "recent_rotated_file", recent_rotated[1]); |
4231 | cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); |
4232 | cb->fill(cb, "recent_scanned_file", recent_scanned[1]); |
4233 | } |
4234 | #endif |
4235 | |
4236 | return 0; |
4237 | } |
4238 | |
4239 | static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) |
4240 | { |
4241 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4242 | |
4243 | return mem_cgroup_swappiness(memcg); |
4244 | } |
4245 | |
4246 | static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, |
4247 | u64 val) |
4248 | { |
4249 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4250 | struct mem_cgroup *parent; |
4251 | |
4252 | if (val > 100) |
4253 | return -EINVAL; |
4254 | |
4255 | if (cgrp->parent == NULL) |
4256 | return -EINVAL; |
4257 | |
4258 | parent = mem_cgroup_from_cont(cgrp->parent); |
4259 | |
4260 | cgroup_lock(); |
4261 | |
4262 | /* If under hierarchy, only empty-root can set this value */ |
4263 | if ((parent->use_hierarchy) || |
4264 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { |
4265 | cgroup_unlock(); |
4266 | return -EINVAL; |
4267 | } |
4268 | |
4269 | memcg->swappiness = val; |
4270 | |
4271 | cgroup_unlock(); |
4272 | |
4273 | return 0; |
4274 | } |
4275 | |
4276 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) |
4277 | { |
4278 | struct mem_cgroup_threshold_ary *t; |
4279 | u64 usage; |
4280 | int i; |
4281 | |
4282 | rcu_read_lock(); |
4283 | if (!swap) |
4284 | t = rcu_dereference(memcg->thresholds.primary); |
4285 | else |
4286 | t = rcu_dereference(memcg->memsw_thresholds.primary); |
4287 | |
4288 | if (!t) |
4289 | goto unlock; |
4290 | |
4291 | usage = mem_cgroup_usage(memcg, swap); |
4292 | |
4293 | /* |
4294 | * current_threshold points to threshold just below usage. |
4295 | * If it's not true, a threshold was crossed after last |
4296 | * call of __mem_cgroup_threshold(). |
4297 | */ |
4298 | i = t->current_threshold; |
4299 | |
4300 | /* |
4301 | * Iterate backward over array of thresholds starting from |
4302 | * current_threshold and check if a threshold is crossed. |
4303 | * If none of thresholds below usage is crossed, we read |
4304 | * only one element of the array here. |
4305 | */ |
4306 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) |
4307 | eventfd_signal(t->entries[i].eventfd, 1); |
4308 | |
4309 | /* i = current_threshold + 1 */ |
4310 | i++; |
4311 | |
4312 | /* |
4313 | * Iterate forward over array of thresholds starting from |
4314 | * current_threshold+1 and check if a threshold is crossed. |
4315 | * If none of thresholds above usage is crossed, we read |
4316 | * only one element of the array here. |
4317 | */ |
4318 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) |
4319 | eventfd_signal(t->entries[i].eventfd, 1); |
4320 | |
4321 | /* Update current_threshold */ |
4322 | t->current_threshold = i - 1; |
4323 | unlock: |
4324 | rcu_read_unlock(); |
4325 | } |
4326 | |
4327 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) |
4328 | { |
4329 | while (memcg) { |
4330 | __mem_cgroup_threshold(memcg, false); |
4331 | if (do_swap_account) |
4332 | __mem_cgroup_threshold(memcg, true); |
4333 | |
4334 | memcg = parent_mem_cgroup(memcg); |
4335 | } |
4336 | } |
4337 | |
4338 | static int compare_thresholds(const void *a, const void *b) |
4339 | { |
4340 | const struct mem_cgroup_threshold *_a = a; |
4341 | const struct mem_cgroup_threshold *_b = b; |
4342 | |
4343 | return _a->threshold - _b->threshold; |
4344 | } |
4345 | |
4346 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) |
4347 | { |
4348 | struct mem_cgroup_eventfd_list *ev; |
4349 | |
4350 | list_for_each_entry(ev, &memcg->oom_notify, list) |
4351 | eventfd_signal(ev->eventfd, 1); |
4352 | return 0; |
4353 | } |
4354 | |
4355 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) |
4356 | { |
4357 | struct mem_cgroup *iter; |
4358 | |
4359 | for_each_mem_cgroup_tree(iter, memcg) |
4360 | mem_cgroup_oom_notify_cb(iter); |
4361 | } |
4362 | |
4363 | static int mem_cgroup_usage_register_event(struct cgroup *cgrp, |
4364 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) |
4365 | { |
4366 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4367 | struct mem_cgroup_thresholds *thresholds; |
4368 | struct mem_cgroup_threshold_ary *new; |
4369 | int type = MEMFILE_TYPE(cft->private); |
4370 | u64 threshold, usage; |
4371 | int i, size, ret; |
4372 | |
4373 | ret = res_counter_memparse_write_strategy(args, &threshold); |
4374 | if (ret) |
4375 | return ret; |
4376 | |
4377 | mutex_lock(&memcg->thresholds_lock); |
4378 | |
4379 | if (type == _MEM) |
4380 | thresholds = &memcg->thresholds; |
4381 | else if (type == _MEMSWAP) |
4382 | thresholds = &memcg->memsw_thresholds; |
4383 | else |
4384 | BUG(); |
4385 | |
4386 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); |
4387 | |
4388 | /* Check if a threshold crossed before adding a new one */ |
4389 | if (thresholds->primary) |
4390 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); |
4391 | |
4392 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; |
4393 | |
4394 | /* Allocate memory for new array of thresholds */ |
4395 | new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), |
4396 | GFP_KERNEL); |
4397 | if (!new) { |
4398 | ret = -ENOMEM; |
4399 | goto unlock; |
4400 | } |
4401 | new->size = size; |
4402 | |
4403 | /* Copy thresholds (if any) to new array */ |
4404 | if (thresholds->primary) { |
4405 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * |
4406 | sizeof(struct mem_cgroup_threshold)); |
4407 | } |
4408 | |
4409 | /* Add new threshold */ |
4410 | new->entries[size - 1].eventfd = eventfd; |
4411 | new->entries[size - 1].threshold = threshold; |
4412 | |
4413 | /* Sort thresholds. Registering of new threshold isn't time-critical */ |
4414 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), |
4415 | compare_thresholds, NULL); |
4416 | |
4417 | /* Find current threshold */ |
4418 | new->current_threshold = -1; |
4419 | for (i = 0; i < size; i++) { |
4420 | if (new->entries[i].threshold < usage) { |
4421 | /* |
4422 | * new->current_threshold will not be used until |
4423 | * rcu_assign_pointer(), so it's safe to increment |
4424 | * it here. |
4425 | */ |
4426 | ++new->current_threshold; |
4427 | } |
4428 | } |
4429 | |
4430 | /* Free old spare buffer and save old primary buffer as spare */ |
4431 | kfree(thresholds->spare); |
4432 | thresholds->spare = thresholds->primary; |
4433 | |
4434 | rcu_assign_pointer(thresholds->primary, new); |
4435 | |
4436 | /* To be sure that nobody uses thresholds */ |
4437 | synchronize_rcu(); |
4438 | |
4439 | unlock: |
4440 | mutex_unlock(&memcg->thresholds_lock); |
4441 | |
4442 | return ret; |
4443 | } |
4444 | |
4445 | static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, |
4446 | struct cftype *cft, struct eventfd_ctx *eventfd) |
4447 | { |
4448 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4449 | struct mem_cgroup_thresholds *thresholds; |
4450 | struct mem_cgroup_threshold_ary *new; |
4451 | int type = MEMFILE_TYPE(cft->private); |
4452 | u64 usage; |
4453 | int i, j, size; |
4454 | |
4455 | mutex_lock(&memcg->thresholds_lock); |
4456 | if (type == _MEM) |
4457 | thresholds = &memcg->thresholds; |
4458 | else if (type == _MEMSWAP) |
4459 | thresholds = &memcg->memsw_thresholds; |
4460 | else |
4461 | BUG(); |
4462 | |
4463 | if (!thresholds->primary) |
4464 | goto unlock; |
4465 | |
4466 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); |
4467 | |
4468 | /* Check if a threshold crossed before removing */ |
4469 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); |
4470 | |
4471 | /* Calculate new number of threshold */ |
4472 | size = 0; |
4473 | for (i = 0; i < thresholds->primary->size; i++) { |
4474 | if (thresholds->primary->entries[i].eventfd != eventfd) |
4475 | size++; |
4476 | } |
4477 | |
4478 | new = thresholds->spare; |
4479 | |
4480 | /* Set thresholds array to NULL if we don't have thresholds */ |
4481 | if (!size) { |
4482 | kfree(new); |
4483 | new = NULL; |
4484 | goto swap_buffers; |
4485 | } |
4486 | |
4487 | new->size = size; |
4488 | |
4489 | /* Copy thresholds and find current threshold */ |
4490 | new->current_threshold = -1; |
4491 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { |
4492 | if (thresholds->primary->entries[i].eventfd == eventfd) |
4493 | continue; |
4494 | |
4495 | new->entries[j] = thresholds->primary->entries[i]; |
4496 | if (new->entries[j].threshold < usage) { |
4497 | /* |
4498 | * new->current_threshold will not be used |
4499 | * until rcu_assign_pointer(), so it's safe to increment |
4500 | * it here. |
4501 | */ |
4502 | ++new->current_threshold; |
4503 | } |
4504 | j++; |
4505 | } |
4506 | |
4507 | swap_buffers: |
4508 | /* Swap primary and spare array */ |
4509 | thresholds->spare = thresholds->primary; |
4510 | /* If all events are unregistered, free the spare array */ |
4511 | if (!new) { |
4512 | kfree(thresholds->spare); |
4513 | thresholds->spare = NULL; |
4514 | } |
4515 | |
4516 | rcu_assign_pointer(thresholds->primary, new); |
4517 | |
4518 | /* To be sure that nobody uses thresholds */ |
4519 | synchronize_rcu(); |
4520 | unlock: |
4521 | mutex_unlock(&memcg->thresholds_lock); |
4522 | } |
4523 | |
4524 | static int mem_cgroup_oom_register_event(struct cgroup *cgrp, |
4525 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) |
4526 | { |
4527 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4528 | struct mem_cgroup_eventfd_list *event; |
4529 | int type = MEMFILE_TYPE(cft->private); |
4530 | |
4531 | BUG_ON(type != _OOM_TYPE); |
4532 | event = kmalloc(sizeof(*event), GFP_KERNEL); |
4533 | if (!event) |
4534 | return -ENOMEM; |
4535 | |
4536 | spin_lock(&memcg_oom_lock); |
4537 | |
4538 | event->eventfd = eventfd; |
4539 | list_add(&event->list, &memcg->oom_notify); |
4540 | |
4541 | /* already in OOM ? */ |
4542 | if (atomic_read(&memcg->under_oom)) |
4543 | eventfd_signal(eventfd, 1); |
4544 | spin_unlock(&memcg_oom_lock); |
4545 | |
4546 | return 0; |
4547 | } |
4548 | |
4549 | static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, |
4550 | struct cftype *cft, struct eventfd_ctx *eventfd) |
4551 | { |
4552 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4553 | struct mem_cgroup_eventfd_list *ev, *tmp; |
4554 | int type = MEMFILE_TYPE(cft->private); |
4555 | |
4556 | BUG_ON(type != _OOM_TYPE); |
4557 | |
4558 | spin_lock(&memcg_oom_lock); |
4559 | |
4560 | list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { |
4561 | if (ev->eventfd == eventfd) { |
4562 | list_del(&ev->list); |
4563 | kfree(ev); |
4564 | } |
4565 | } |
4566 | |
4567 | spin_unlock(&memcg_oom_lock); |
4568 | } |
4569 | |
4570 | static int mem_cgroup_oom_control_read(struct cgroup *cgrp, |
4571 | struct cftype *cft, struct cgroup_map_cb *cb) |
4572 | { |
4573 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4574 | |
4575 | cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); |
4576 | |
4577 | if (atomic_read(&memcg->under_oom)) |
4578 | cb->fill(cb, "under_oom", 1); |
4579 | else |
4580 | cb->fill(cb, "under_oom", 0); |
4581 | return 0; |
4582 | } |
4583 | |
4584 | static int mem_cgroup_oom_control_write(struct cgroup *cgrp, |
4585 | struct cftype *cft, u64 val) |
4586 | { |
4587 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
4588 | struct mem_cgroup *parent; |
4589 | |
4590 | /* cannot set to root cgroup and only 0 and 1 are allowed */ |
4591 | if (!cgrp->parent || !((val == 0) || (val == 1))) |
4592 | return -EINVAL; |
4593 | |
4594 | parent = mem_cgroup_from_cont(cgrp->parent); |
4595 | |
4596 | cgroup_lock(); |
4597 | /* oom-kill-disable is a flag for subhierarchy. */ |
4598 | if ((parent->use_hierarchy) || |
4599 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { |
4600 | cgroup_unlock(); |
4601 | return -EINVAL; |
4602 | } |
4603 | memcg->oom_kill_disable = val; |
4604 | if (!val) |
4605 | memcg_oom_recover(memcg); |
4606 | cgroup_unlock(); |
4607 | return 0; |
4608 | } |
4609 | |
4610 | #ifdef CONFIG_NUMA |
4611 | static const struct file_operations mem_control_numa_stat_file_operations = { |
4612 | .read = seq_read, |
4613 | .llseek = seq_lseek, |
4614 | .release = single_release, |
4615 | }; |
4616 | |
4617 | static int mem_control_numa_stat_open(struct inode *unused, struct file *file) |
4618 | { |
4619 | struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; |
4620 | |
4621 | file->f_op = &mem_control_numa_stat_file_operations; |
4622 | return single_open(file, mem_control_numa_stat_show, cont); |
4623 | } |
4624 | #endif /* CONFIG_NUMA */ |
4625 | |
4626 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM |
4627 | static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) |
4628 | { |
4629 | /* |
4630 | * Part of this would be better living in a separate allocation |
4631 | * function, leaving us with just the cgroup tree population work. |
4632 | * We, however, depend on state such as network's proto_list that |
4633 | * is only initialized after cgroup creation. I found the less |
4634 | * cumbersome way to deal with it to defer it all to populate time |
4635 | */ |
4636 | return mem_cgroup_sockets_init(cont, ss); |
4637 | }; |
4638 | |
4639 | static void kmem_cgroup_destroy(struct cgroup *cont) |
4640 | { |
4641 | mem_cgroup_sockets_destroy(cont); |
4642 | } |
4643 | #else |
4644 | static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) |
4645 | { |
4646 | return 0; |
4647 | } |
4648 | |
4649 | static void kmem_cgroup_destroy(struct cgroup *cont) |
4650 | { |
4651 | } |
4652 | #endif |
4653 | |
4654 | static struct cftype mem_cgroup_files[] = { |
4655 | { |
4656 | .name = "usage_in_bytes", |
4657 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), |
4658 | .read_u64 = mem_cgroup_read, |
4659 | .register_event = mem_cgroup_usage_register_event, |
4660 | .unregister_event = mem_cgroup_usage_unregister_event, |
4661 | }, |
4662 | { |
4663 | .name = "max_usage_in_bytes", |
4664 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), |
4665 | .trigger = mem_cgroup_reset, |
4666 | .read_u64 = mem_cgroup_read, |
4667 | }, |
4668 | { |
4669 | .name = "limit_in_bytes", |
4670 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), |
4671 | .write_string = mem_cgroup_write, |
4672 | .read_u64 = mem_cgroup_read, |
4673 | }, |
4674 | { |
4675 | .name = "soft_limit_in_bytes", |
4676 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), |
4677 | .write_string = mem_cgroup_write, |
4678 | .read_u64 = mem_cgroup_read, |
4679 | }, |
4680 | { |
4681 | .name = "failcnt", |
4682 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), |
4683 | .trigger = mem_cgroup_reset, |
4684 | .read_u64 = mem_cgroup_read, |
4685 | }, |
4686 | { |
4687 | .name = "stat", |
4688 | .read_map = mem_control_stat_show, |
4689 | }, |
4690 | { |
4691 | .name = "force_empty", |
4692 | .trigger = mem_cgroup_force_empty_write, |
4693 | }, |
4694 | { |
4695 | .name = "use_hierarchy", |
4696 | .write_u64 = mem_cgroup_hierarchy_write, |
4697 | .read_u64 = mem_cgroup_hierarchy_read, |
4698 | }, |
4699 | { |
4700 | .name = "swappiness", |
4701 | .read_u64 = mem_cgroup_swappiness_read, |
4702 | .write_u64 = mem_cgroup_swappiness_write, |
4703 | }, |
4704 | { |
4705 | .name = "move_charge_at_immigrate", |
4706 | .read_u64 = mem_cgroup_move_charge_read, |
4707 | .write_u64 = mem_cgroup_move_charge_write, |
4708 | }, |
4709 | { |
4710 | .name = "oom_control", |
4711 | .read_map = mem_cgroup_oom_control_read, |
4712 | .write_u64 = mem_cgroup_oom_control_write, |
4713 | .register_event = mem_cgroup_oom_register_event, |
4714 | .unregister_event = mem_cgroup_oom_unregister_event, |
4715 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), |
4716 | }, |
4717 | #ifdef CONFIG_NUMA |
4718 | { |
4719 | .name = "numa_stat", |
4720 | .open = mem_control_numa_stat_open, |
4721 | .mode = S_IRUGO, |
4722 | }, |
4723 | #endif |
4724 | }; |
4725 | |
4726 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
4727 | static struct cftype memsw_cgroup_files[] = { |
4728 | { |
4729 | .name = "memsw.usage_in_bytes", |
4730 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), |
4731 | .read_u64 = mem_cgroup_read, |
4732 | .register_event = mem_cgroup_usage_register_event, |
4733 | .unregister_event = mem_cgroup_usage_unregister_event, |
4734 | }, |
4735 | { |
4736 | .name = "memsw.max_usage_in_bytes", |
4737 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), |
4738 | .trigger = mem_cgroup_reset, |
4739 | .read_u64 = mem_cgroup_read, |
4740 | }, |
4741 | { |
4742 | .name = "memsw.limit_in_bytes", |
4743 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), |
4744 | .write_string = mem_cgroup_write, |
4745 | .read_u64 = mem_cgroup_read, |
4746 | }, |
4747 | { |
4748 | .name = "memsw.failcnt", |
4749 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), |
4750 | .trigger = mem_cgroup_reset, |
4751 | .read_u64 = mem_cgroup_read, |
4752 | }, |
4753 | }; |
4754 | |
4755 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) |
4756 | { |
4757 | if (!do_swap_account) |
4758 | return 0; |
4759 | return cgroup_add_files(cont, ss, memsw_cgroup_files, |
4760 | ARRAY_SIZE(memsw_cgroup_files)); |
4761 | }; |
4762 | #else |
4763 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) |
4764 | { |
4765 | return 0; |
4766 | } |
4767 | #endif |
4768 | |
4769 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) |
4770 | { |
4771 | struct mem_cgroup_per_node *pn; |
4772 | struct mem_cgroup_per_zone *mz; |
4773 | enum lru_list lru; |
4774 | int zone, tmp = node; |
4775 | /* |
4776 | * This routine is called against possible nodes. |
4777 | * But it's BUG to call kmalloc() against offline node. |
4778 | * |
4779 | * TODO: this routine can waste much memory for nodes which will |
4780 | * never be onlined. It's better to use memory hotplug callback |
4781 | * function. |
4782 | */ |
4783 | if (!node_state(node, N_NORMAL_MEMORY)) |
4784 | tmp = -1; |
4785 | pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); |
4786 | if (!pn) |
4787 | return 1; |
4788 | |
4789 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
4790 | mz = &pn->zoneinfo[zone]; |
4791 | for_each_lru(lru) |
4792 | INIT_LIST_HEAD(&mz->lruvec.lists[lru]); |
4793 | mz->usage_in_excess = 0; |
4794 | mz->on_tree = false; |
4795 | mz->memcg = memcg; |
4796 | } |
4797 | memcg->info.nodeinfo[node] = pn; |
4798 | return 0; |
4799 | } |
4800 | |
4801 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) |
4802 | { |
4803 | kfree(memcg->info.nodeinfo[node]); |
4804 | } |
4805 | |
4806 | static struct mem_cgroup *mem_cgroup_alloc(void) |
4807 | { |
4808 | struct mem_cgroup *memcg; |
4809 | int size = sizeof(struct mem_cgroup); |
4810 | |
4811 | /* Can be very big if MAX_NUMNODES is very big */ |
4812 | if (size < PAGE_SIZE) |
4813 | memcg = kzalloc(size, GFP_KERNEL); |
4814 | else |
4815 | memcg = vzalloc(size); |
4816 | |
4817 | if (!memcg) |
4818 | return NULL; |
4819 | |
4820 | memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu); |
4821 | if (!memcg->stat) |
4822 | goto out_free; |
4823 | spin_lock_init(&memcg->pcp_counter_lock); |
4824 | return memcg; |
4825 | |
4826 | out_free: |
4827 | if (size < PAGE_SIZE) |
4828 | kfree(memcg); |
4829 | else |
4830 | vfree(memcg); |
4831 | return NULL; |
4832 | } |
4833 | |
4834 | /* |
4835 | * Helpers for freeing a vzalloc()ed mem_cgroup by RCU, |
4836 | * but in process context. The work_freeing structure is overlaid |
4837 | * on the rcu_freeing structure, which itself is overlaid on memsw. |
4838 | */ |
4839 | static void vfree_work(struct work_struct *work) |
4840 | { |
4841 | struct mem_cgroup *memcg; |
4842 | |
4843 | memcg = container_of(work, struct mem_cgroup, work_freeing); |
4844 | vfree(memcg); |
4845 | } |
4846 | static void vfree_rcu(struct rcu_head *rcu_head) |
4847 | { |
4848 | struct mem_cgroup *memcg; |
4849 | |
4850 | memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing); |
4851 | INIT_WORK(&memcg->work_freeing, vfree_work); |
4852 | schedule_work(&memcg->work_freeing); |
4853 | } |
4854 | |
4855 | /* |
4856 | * At destroying mem_cgroup, references from swap_cgroup can remain. |
4857 | * (scanning all at force_empty is too costly...) |
4858 | * |
4859 | * Instead of clearing all references at force_empty, we remember |
4860 | * the number of reference from swap_cgroup and free mem_cgroup when |
4861 | * it goes down to 0. |
4862 | * |
4863 | * Removal of cgroup itself succeeds regardless of refs from swap. |
4864 | */ |
4865 | |
4866 | static void __mem_cgroup_free(struct mem_cgroup *memcg) |
4867 | { |
4868 | int node; |
4869 | |
4870 | mem_cgroup_remove_from_trees(memcg); |
4871 | free_css_id(&mem_cgroup_subsys, &memcg->css); |
4872 | |
4873 | for_each_node(node) |
4874 | free_mem_cgroup_per_zone_info(memcg, node); |
4875 | |
4876 | free_percpu(memcg->stat); |
4877 | if (sizeof(struct mem_cgroup) < PAGE_SIZE) |
4878 | kfree_rcu(memcg, rcu_freeing); |
4879 | else |
4880 | call_rcu(&memcg->rcu_freeing, vfree_rcu); |
4881 | } |
4882 | |
4883 | static void mem_cgroup_get(struct mem_cgroup *memcg) |
4884 | { |
4885 | atomic_inc(&memcg->refcnt); |
4886 | } |
4887 | |
4888 | static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) |
4889 | { |
4890 | if (atomic_sub_and_test(count, &memcg->refcnt)) { |
4891 | struct mem_cgroup *parent = parent_mem_cgroup(memcg); |
4892 | __mem_cgroup_free(memcg); |
4893 | if (parent) |
4894 | mem_cgroup_put(parent); |
4895 | } |
4896 | } |
4897 | |
4898 | static void mem_cgroup_put(struct mem_cgroup *memcg) |
4899 | { |
4900 | __mem_cgroup_put(memcg, 1); |
4901 | } |
4902 | |
4903 | /* |
4904 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. |
4905 | */ |
4906 | struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) |
4907 | { |
4908 | if (!memcg->res.parent) |
4909 | return NULL; |
4910 | return mem_cgroup_from_res_counter(memcg->res.parent, res); |
4911 | } |
4912 | EXPORT_SYMBOL(parent_mem_cgroup); |
4913 | |
4914 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
4915 | static void __init enable_swap_cgroup(void) |
4916 | { |
4917 | if (!mem_cgroup_disabled() && really_do_swap_account) |
4918 | do_swap_account = 1; |
4919 | } |
4920 | #else |
4921 | static void __init enable_swap_cgroup(void) |
4922 | { |
4923 | } |
4924 | #endif |
4925 | |
4926 | static int mem_cgroup_soft_limit_tree_init(void) |
4927 | { |
4928 | struct mem_cgroup_tree_per_node *rtpn; |
4929 | struct mem_cgroup_tree_per_zone *rtpz; |
4930 | int tmp, node, zone; |
4931 | |
4932 | for_each_node(node) { |
4933 | tmp = node; |
4934 | if (!node_state(node, N_NORMAL_MEMORY)) |
4935 | tmp = -1; |
4936 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); |
4937 | if (!rtpn) |
4938 | goto err_cleanup; |
4939 | |
4940 | soft_limit_tree.rb_tree_per_node[node] = rtpn; |
4941 | |
4942 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
4943 | rtpz = &rtpn->rb_tree_per_zone[zone]; |
4944 | rtpz->rb_root = RB_ROOT; |
4945 | spin_lock_init(&rtpz->lock); |
4946 | } |
4947 | } |
4948 | return 0; |
4949 | |
4950 | err_cleanup: |
4951 | for_each_node(node) { |
4952 | if (!soft_limit_tree.rb_tree_per_node[node]) |
4953 | break; |
4954 | kfree(soft_limit_tree.rb_tree_per_node[node]); |
4955 | soft_limit_tree.rb_tree_per_node[node] = NULL; |
4956 | } |
4957 | return 1; |
4958 | |
4959 | } |
4960 | |
4961 | static struct cgroup_subsys_state * __ref |
4962 | mem_cgroup_create(struct cgroup *cont) |
4963 | { |
4964 | struct mem_cgroup *memcg, *parent; |
4965 | long error = -ENOMEM; |
4966 | int node; |
4967 | |
4968 | memcg = mem_cgroup_alloc(); |
4969 | if (!memcg) |
4970 | return ERR_PTR(error); |
4971 | |
4972 | for_each_node(node) |
4973 | if (alloc_mem_cgroup_per_zone_info(memcg, node)) |
4974 | goto free_out; |
4975 | |
4976 | /* root ? */ |
4977 | if (cont->parent == NULL) { |
4978 | int cpu; |
4979 | enable_swap_cgroup(); |
4980 | parent = NULL; |
4981 | if (mem_cgroup_soft_limit_tree_init()) |
4982 | goto free_out; |
4983 | root_mem_cgroup = memcg; |
4984 | for_each_possible_cpu(cpu) { |
4985 | struct memcg_stock_pcp *stock = |
4986 | &per_cpu(memcg_stock, cpu); |
4987 | INIT_WORK(&stock->work, drain_local_stock); |
4988 | } |
4989 | hotcpu_notifier(memcg_cpu_hotplug_callback, 0); |
4990 | } else { |
4991 | parent = mem_cgroup_from_cont(cont->parent); |
4992 | memcg->use_hierarchy = parent->use_hierarchy; |
4993 | memcg->oom_kill_disable = parent->oom_kill_disable; |
4994 | } |
4995 | |
4996 | if (parent && parent->use_hierarchy) { |
4997 | res_counter_init(&memcg->res, &parent->res); |
4998 | res_counter_init(&memcg->memsw, &parent->memsw); |
4999 | /* |
5000 | * We increment refcnt of the parent to ensure that we can |
5001 | * safely access it on res_counter_charge/uncharge. |
5002 | * This refcnt will be decremented when freeing this |
5003 | * mem_cgroup(see mem_cgroup_put). |
5004 | */ |
5005 | mem_cgroup_get(parent); |
5006 | } else { |
5007 | res_counter_init(&memcg->res, NULL); |
5008 | res_counter_init(&memcg->memsw, NULL); |
5009 | } |
5010 | memcg->last_scanned_node = MAX_NUMNODES; |
5011 | INIT_LIST_HEAD(&memcg->oom_notify); |
5012 | |
5013 | if (parent) |
5014 | memcg->swappiness = mem_cgroup_swappiness(parent); |
5015 | atomic_set(&memcg->refcnt, 1); |
5016 | memcg->move_charge_at_immigrate = 0; |
5017 | mutex_init(&memcg->thresholds_lock); |
5018 | spin_lock_init(&memcg->move_lock); |
5019 | return &memcg->css; |
5020 | free_out: |
5021 | __mem_cgroup_free(memcg); |
5022 | return ERR_PTR(error); |
5023 | } |
5024 | |
5025 | static int mem_cgroup_pre_destroy(struct cgroup *cont) |
5026 | { |
5027 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
5028 | |
5029 | return mem_cgroup_force_empty(memcg, false); |
5030 | } |
5031 | |
5032 | static void mem_cgroup_destroy(struct cgroup *cont) |
5033 | { |
5034 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
5035 | |
5036 | kmem_cgroup_destroy(cont); |
5037 | |
5038 | mem_cgroup_put(memcg); |
5039 | } |
5040 | |
5041 | static int mem_cgroup_populate(struct cgroup_subsys *ss, |
5042 | struct cgroup *cont) |
5043 | { |
5044 | int ret; |
5045 | |
5046 | ret = cgroup_add_files(cont, ss, mem_cgroup_files, |
5047 | ARRAY_SIZE(mem_cgroup_files)); |
5048 | |
5049 | if (!ret) |
5050 | ret = register_memsw_files(cont, ss); |
5051 | |
5052 | if (!ret) |
5053 | ret = register_kmem_files(cont, ss); |
5054 | |
5055 | return ret; |
5056 | } |
5057 | |
5058 | #ifdef CONFIG_MMU |
5059 | /* Handlers for move charge at task migration. */ |
5060 | #define PRECHARGE_COUNT_AT_ONCE 256 |
5061 | static int mem_cgroup_do_precharge(unsigned long count) |
5062 | { |
5063 | int ret = 0; |
5064 | int batch_count = PRECHARGE_COUNT_AT_ONCE; |
5065 | struct mem_cgroup *memcg = mc.to; |
5066 | |
5067 | if (mem_cgroup_is_root(memcg)) { |
5068 | mc.precharge += count; |
5069 | /* we don't need css_get for root */ |
5070 | return ret; |
5071 | } |
5072 | /* try to charge at once */ |
5073 | if (count > 1) { |
5074 | struct res_counter *dummy; |
5075 | /* |
5076 | * "memcg" cannot be under rmdir() because we've already checked |
5077 | * by cgroup_lock_live_cgroup() that it is not removed and we |
5078 | * are still under the same cgroup_mutex. So we can postpone |
5079 | * css_get(). |
5080 | */ |
5081 | if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) |
5082 | goto one_by_one; |
5083 | if (do_swap_account && res_counter_charge(&memcg->memsw, |
5084 | PAGE_SIZE * count, &dummy)) { |
5085 | res_counter_uncharge(&memcg->res, PAGE_SIZE * count); |
5086 | goto one_by_one; |
5087 | } |
5088 | mc.precharge += count; |
5089 | return ret; |
5090 | } |
5091 | one_by_one: |
5092 | /* fall back to one by one charge */ |
5093 | while (count--) { |
5094 | if (signal_pending(current)) { |
5095 | ret = -EINTR; |
5096 | break; |
5097 | } |
5098 | if (!batch_count--) { |
5099 | batch_count = PRECHARGE_COUNT_AT_ONCE; |
5100 | cond_resched(); |
5101 | } |
5102 | ret = __mem_cgroup_try_charge(NULL, |
5103 | GFP_KERNEL, 1, &memcg, false); |
5104 | if (ret) |
5105 | /* mem_cgroup_clear_mc() will do uncharge later */ |
5106 | return ret; |
5107 | mc.precharge++; |
5108 | } |
5109 | return ret; |
5110 | } |
5111 | |
5112 | /** |
5113 | * get_mctgt_type - get target type of moving charge |
5114 | * @vma: the vma the pte to be checked belongs |
5115 | * @addr: the address corresponding to the pte to be checked |
5116 | * @ptent: the pte to be checked |
5117 | * @target: the pointer the target page or swap ent will be stored(can be NULL) |
5118 | * |
5119 | * Returns |
5120 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. |
5121 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for |
5122 | * move charge. if @target is not NULL, the page is stored in target->page |
5123 | * with extra refcnt got(Callers should handle it). |
5124 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a |
5125 | * target for charge migration. if @target is not NULL, the entry is stored |
5126 | * in target->ent. |
5127 | * |
5128 | * Called with pte lock held. |
5129 | */ |
5130 | union mc_target { |
5131 | struct page *page; |
5132 | swp_entry_t ent; |
5133 | }; |
5134 | |
5135 | enum mc_target_type { |
5136 | MC_TARGET_NONE = 0, |
5137 | MC_TARGET_PAGE, |
5138 | MC_TARGET_SWAP, |
5139 | }; |
5140 | |
5141 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, |
5142 | unsigned long addr, pte_t ptent) |
5143 | { |
5144 | struct page *page = vm_normal_page(vma, addr, ptent); |
5145 | |
5146 | if (!page || !page_mapped(page)) |
5147 | return NULL; |
5148 | if (PageAnon(page)) { |
5149 | /* we don't move shared anon */ |
5150 | if (!move_anon() || page_mapcount(page) > 2) |
5151 | return NULL; |
5152 | } else if (!move_file()) |
5153 | /* we ignore mapcount for file pages */ |
5154 | return NULL; |
5155 | if (!get_page_unless_zero(page)) |
5156 | return NULL; |
5157 | |
5158 | return page; |
5159 | } |
5160 | |
5161 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, |
5162 | unsigned long addr, pte_t ptent, swp_entry_t *entry) |
5163 | { |
5164 | int usage_count; |
5165 | struct page *page = NULL; |
5166 | swp_entry_t ent = pte_to_swp_entry(ptent); |
5167 | |
5168 | if (!move_anon() || non_swap_entry(ent)) |
5169 | return NULL; |
5170 | usage_count = mem_cgroup_count_swap_user(ent, &page); |
5171 | if (usage_count > 1) { /* we don't move shared anon */ |
5172 | if (page) |
5173 | put_page(page); |
5174 | return NULL; |
5175 | } |
5176 | if (do_swap_account) |
5177 | entry->val = ent.val; |
5178 | |
5179 | return page; |
5180 | } |
5181 | |
5182 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, |
5183 | unsigned long addr, pte_t ptent, swp_entry_t *entry) |
5184 | { |
5185 | struct page *page = NULL; |
5186 | struct inode *inode; |
5187 | struct address_space *mapping; |
5188 | pgoff_t pgoff; |
5189 | |
5190 | if (!vma->vm_file) /* anonymous vma */ |
5191 | return NULL; |
5192 | if (!move_file()) |
5193 | return NULL; |
5194 | |
5195 | inode = vma->vm_file->f_path.dentry->d_inode; |
5196 | mapping = vma->vm_file->f_mapping; |
5197 | if (pte_none(ptent)) |
5198 | pgoff = linear_page_index(vma, addr); |
5199 | else /* pte_file(ptent) is true */ |
5200 | pgoff = pte_to_pgoff(ptent); |
5201 | |
5202 | /* page is moved even if it's not RSS of this task(page-faulted). */ |
5203 | page = find_get_page(mapping, pgoff); |
5204 | |
5205 | #ifdef CONFIG_SWAP |
5206 | /* shmem/tmpfs may report page out on swap: account for that too. */ |
5207 | if (radix_tree_exceptional_entry(page)) { |
5208 | swp_entry_t swap = radix_to_swp_entry(page); |
5209 | if (do_swap_account) |
5210 | *entry = swap; |
5211 | page = find_get_page(&swapper_space, swap.val); |
5212 | } |
5213 | #endif |
5214 | return page; |
5215 | } |
5216 | |
5217 | static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma, |
5218 | unsigned long addr, pte_t ptent, union mc_target *target) |
5219 | { |
5220 | struct page *page = NULL; |
5221 | struct page_cgroup *pc; |
5222 | enum mc_target_type ret = MC_TARGET_NONE; |
5223 | swp_entry_t ent = { .val = 0 }; |
5224 | |
5225 | if (pte_present(ptent)) |
5226 | page = mc_handle_present_pte(vma, addr, ptent); |
5227 | else if (is_swap_pte(ptent)) |
5228 | page = mc_handle_swap_pte(vma, addr, ptent, &ent); |
5229 | else if (pte_none(ptent) || pte_file(ptent)) |
5230 | page = mc_handle_file_pte(vma, addr, ptent, &ent); |
5231 | |
5232 | if (!page && !ent.val) |
5233 | return ret; |
5234 | if (page) { |
5235 | pc = lookup_page_cgroup(page); |
5236 | /* |
5237 | * Do only loose check w/o page_cgroup lock. |
5238 | * mem_cgroup_move_account() checks the pc is valid or not under |
5239 | * the lock. |
5240 | */ |
5241 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { |
5242 | ret = MC_TARGET_PAGE; |
5243 | if (target) |
5244 | target->page = page; |
5245 | } |
5246 | if (!ret || !target) |
5247 | put_page(page); |
5248 | } |
5249 | /* There is a swap entry and a page doesn't exist or isn't charged */ |
5250 | if (ent.val && !ret && |
5251 | css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) { |
5252 | ret = MC_TARGET_SWAP; |
5253 | if (target) |
5254 | target->ent = ent; |
5255 | } |
5256 | return ret; |
5257 | } |
5258 | |
5259 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
5260 | /* |
5261 | * We don't consider swapping or file mapped pages because THP does not |
5262 | * support them for now. |
5263 | * Caller should make sure that pmd_trans_huge(pmd) is true. |
5264 | */ |
5265 | static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, |
5266 | unsigned long addr, pmd_t pmd, union mc_target *target) |
5267 | { |
5268 | struct page *page = NULL; |
5269 | struct page_cgroup *pc; |
5270 | enum mc_target_type ret = MC_TARGET_NONE; |
5271 | |
5272 | page = pmd_page(pmd); |
5273 | VM_BUG_ON(!page || !PageHead(page)); |
5274 | if (!move_anon()) |
5275 | return ret; |
5276 | pc = lookup_page_cgroup(page); |
5277 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { |
5278 | ret = MC_TARGET_PAGE; |
5279 | if (target) { |
5280 | get_page(page); |
5281 | target->page = page; |
5282 | } |
5283 | } |
5284 | return ret; |
5285 | } |
5286 | #else |
5287 | static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, |
5288 | unsigned long addr, pmd_t pmd, union mc_target *target) |
5289 | { |
5290 | return MC_TARGET_NONE; |
5291 | } |
5292 | #endif |
5293 | |
5294 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, |
5295 | unsigned long addr, unsigned long end, |
5296 | struct mm_walk *walk) |
5297 | { |
5298 | struct vm_area_struct *vma = walk->private; |
5299 | pte_t *pte; |
5300 | spinlock_t *ptl; |
5301 | |
5302 | if (pmd_trans_huge_lock(pmd, vma) == 1) { |
5303 | if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE) |
5304 | mc.precharge += HPAGE_PMD_NR; |
5305 | spin_unlock(&vma->vm_mm->page_table_lock); |
5306 | return 0; |
5307 | } |
5308 | |
5309 | if (pmd_trans_unstable(pmd)) |
5310 | return 0; |
5311 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
5312 | for (; addr != end; pte++, addr += PAGE_SIZE) |
5313 | if (get_mctgt_type(vma, addr, *pte, NULL)) |
5314 | mc.precharge++; /* increment precharge temporarily */ |
5315 | pte_unmap_unlock(pte - 1, ptl); |
5316 | cond_resched(); |
5317 | |
5318 | return 0; |
5319 | } |
5320 | |
5321 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) |
5322 | { |
5323 | unsigned long precharge; |
5324 | struct vm_area_struct *vma; |
5325 | |
5326 | down_read(&mm->mmap_sem); |
5327 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
5328 | struct mm_walk mem_cgroup_count_precharge_walk = { |
5329 | .pmd_entry = mem_cgroup_count_precharge_pte_range, |
5330 | .mm = mm, |
5331 | .private = vma, |
5332 | }; |
5333 | if (is_vm_hugetlb_page(vma)) |
5334 | continue; |
5335 | walk_page_range(vma->vm_start, vma->vm_end, |
5336 | &mem_cgroup_count_precharge_walk); |
5337 | } |
5338 | up_read(&mm->mmap_sem); |
5339 | |
5340 | precharge = mc.precharge; |
5341 | mc.precharge = 0; |
5342 | |
5343 | return precharge; |
5344 | } |
5345 | |
5346 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) |
5347 | { |
5348 | unsigned long precharge = mem_cgroup_count_precharge(mm); |
5349 | |
5350 | VM_BUG_ON(mc.moving_task); |
5351 | mc.moving_task = current; |
5352 | return mem_cgroup_do_precharge(precharge); |
5353 | } |
5354 | |
5355 | /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ |
5356 | static void __mem_cgroup_clear_mc(void) |
5357 | { |
5358 | struct mem_cgroup *from = mc.from; |
5359 | struct mem_cgroup *to = mc.to; |
5360 | |
5361 | /* we must uncharge all the leftover precharges from mc.to */ |
5362 | if (mc.precharge) { |
5363 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); |
5364 | mc.precharge = 0; |
5365 | } |
5366 | /* |
5367 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so |
5368 | * we must uncharge here. |
5369 | */ |
5370 | if (mc.moved_charge) { |
5371 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); |
5372 | mc.moved_charge = 0; |
5373 | } |
5374 | /* we must fixup refcnts and charges */ |
5375 | if (mc.moved_swap) { |
5376 | /* uncharge swap account from the old cgroup */ |
5377 | if (!mem_cgroup_is_root(mc.from)) |
5378 | res_counter_uncharge(&mc.from->memsw, |
5379 | PAGE_SIZE * mc.moved_swap); |
5380 | __mem_cgroup_put(mc.from, mc.moved_swap); |
5381 | |
5382 | if (!mem_cgroup_is_root(mc.to)) { |
5383 | /* |
5384 | * we charged both to->res and to->memsw, so we should |
5385 | * uncharge to->res. |
5386 | */ |
5387 | res_counter_uncharge(&mc.to->res, |
5388 | PAGE_SIZE * mc.moved_swap); |
5389 | } |
5390 | /* we've already done mem_cgroup_get(mc.to) */ |
5391 | mc.moved_swap = 0; |
5392 | } |
5393 | memcg_oom_recover(from); |
5394 | memcg_oom_recover(to); |
5395 | wake_up_all(&mc.waitq); |
5396 | } |
5397 | |
5398 | static void mem_cgroup_clear_mc(void) |
5399 | { |
5400 | struct mem_cgroup *from = mc.from; |
5401 | |
5402 | /* |
5403 | * we must clear moving_task before waking up waiters at the end of |
5404 | * task migration. |
5405 | */ |
5406 | mc.moving_task = NULL; |
5407 | __mem_cgroup_clear_mc(); |
5408 | spin_lock(&mc.lock); |
5409 | mc.from = NULL; |
5410 | mc.to = NULL; |
5411 | spin_unlock(&mc.lock); |
5412 | mem_cgroup_end_move(from); |
5413 | } |
5414 | |
5415 | static int mem_cgroup_can_attach(struct cgroup *cgroup, |
5416 | struct cgroup_taskset *tset) |
5417 | { |
5418 | struct task_struct *p = cgroup_taskset_first(tset); |
5419 | int ret = 0; |
5420 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); |
5421 | |
5422 | if (memcg->move_charge_at_immigrate) { |
5423 | struct mm_struct *mm; |
5424 | struct mem_cgroup *from = mem_cgroup_from_task(p); |
5425 | |
5426 | VM_BUG_ON(from == memcg); |
5427 | |
5428 | mm = get_task_mm(p); |
5429 | if (!mm) |
5430 | return 0; |
5431 | /* We move charges only when we move a owner of the mm */ |
5432 | if (mm->owner == p) { |
5433 | VM_BUG_ON(mc.from); |
5434 | VM_BUG_ON(mc.to); |
5435 | VM_BUG_ON(mc.precharge); |
5436 | VM_BUG_ON(mc.moved_charge); |
5437 | VM_BUG_ON(mc.moved_swap); |
5438 | mem_cgroup_start_move(from); |
5439 | spin_lock(&mc.lock); |
5440 | mc.from = from; |
5441 | mc.to = memcg; |
5442 | spin_unlock(&mc.lock); |
5443 | /* We set mc.moving_task later */ |
5444 | |
5445 | ret = mem_cgroup_precharge_mc(mm); |
5446 | if (ret) |
5447 | mem_cgroup_clear_mc(); |
5448 | } |
5449 | mmput(mm); |
5450 | } |
5451 | return ret; |
5452 | } |
5453 | |
5454 | static void mem_cgroup_cancel_attach(struct cgroup *cgroup, |
5455 | struct cgroup_taskset *tset) |
5456 | { |
5457 | mem_cgroup_clear_mc(); |
5458 | } |
5459 | |
5460 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, |
5461 | unsigned long addr, unsigned long end, |
5462 | struct mm_walk *walk) |
5463 | { |
5464 | int ret = 0; |
5465 | struct vm_area_struct *vma = walk->private; |
5466 | pte_t *pte; |
5467 | spinlock_t *ptl; |
5468 | enum mc_target_type target_type; |
5469 | union mc_target target; |
5470 | struct page *page; |
5471 | struct page_cgroup *pc; |
5472 | |
5473 | /* |
5474 | * We don't take compound_lock() here but no race with splitting thp |
5475 | * happens because: |
5476 | * - if pmd_trans_huge_lock() returns 1, the relevant thp is not |
5477 | * under splitting, which means there's no concurrent thp split, |
5478 | * - if another thread runs into split_huge_page() just after we |
5479 | * entered this if-block, the thread must wait for page table lock |
5480 | * to be unlocked in __split_huge_page_splitting(), where the main |
5481 | * part of thp split is not executed yet. |
5482 | */ |
5483 | if (pmd_trans_huge_lock(pmd, vma) == 1) { |
5484 | if (mc.precharge < HPAGE_PMD_NR) { |
5485 | spin_unlock(&vma->vm_mm->page_table_lock); |
5486 | return 0; |
5487 | } |
5488 | target_type = get_mctgt_type_thp(vma, addr, *pmd, &target); |
5489 | if (target_type == MC_TARGET_PAGE) { |
5490 | page = target.page; |
5491 | if (!isolate_lru_page(page)) { |
5492 | pc = lookup_page_cgroup(page); |
5493 | if (!mem_cgroup_move_account(page, HPAGE_PMD_NR, |
5494 | pc, mc.from, mc.to, |
5495 | false)) { |
5496 | mc.precharge -= HPAGE_PMD_NR; |
5497 | mc.moved_charge += HPAGE_PMD_NR; |
5498 | } |
5499 | putback_lru_page(page); |
5500 | } |
5501 | put_page(page); |
5502 | } |
5503 | spin_unlock(&vma->vm_mm->page_table_lock); |
5504 | return 0; |
5505 | } |
5506 | |
5507 | if (pmd_trans_unstable(pmd)) |
5508 | return 0; |
5509 | retry: |
5510 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
5511 | for (; addr != end; addr += PAGE_SIZE) { |
5512 | pte_t ptent = *(pte++); |
5513 | swp_entry_t ent; |
5514 | |
5515 | if (!mc.precharge) |
5516 | break; |
5517 | |
5518 | switch (get_mctgt_type(vma, addr, ptent, &target)) { |
5519 | case MC_TARGET_PAGE: |
5520 | page = target.page; |
5521 | if (isolate_lru_page(page)) |
5522 | goto put; |
5523 | pc = lookup_page_cgroup(page); |
5524 | if (!mem_cgroup_move_account(page, 1, pc, |
5525 | mc.from, mc.to, false)) { |
5526 | mc.precharge--; |
5527 | /* we uncharge from mc.from later. */ |
5528 | mc.moved_charge++; |
5529 | } |
5530 | putback_lru_page(page); |
5531 | put: /* get_mctgt_type() gets the page */ |
5532 | put_page(page); |
5533 | break; |
5534 | case MC_TARGET_SWAP: |
5535 | ent = target.ent; |
5536 | if (!mem_cgroup_move_swap_account(ent, |
5537 | mc.from, mc.to, false)) { |
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 | put_swap_token(mm); |
5614 | mmput(mm); |
5615 | } |
5616 | if (mc.to) |
5617 | mem_cgroup_clear_mc(); |
5618 | } |
5619 | #else /* !CONFIG_MMU */ |
5620 | static int mem_cgroup_can_attach(struct cgroup *cgroup, |
5621 | struct cgroup_taskset *tset) |
5622 | { |
5623 | return 0; |
5624 | } |
5625 | static void mem_cgroup_cancel_attach(struct cgroup *cgroup, |
5626 | struct cgroup_taskset *tset) |
5627 | { |
5628 | } |
5629 | static void mem_cgroup_move_task(struct cgroup *cont, |
5630 | struct cgroup_taskset *tset) |
5631 | { |
5632 | } |
5633 | #endif |
5634 | |
5635 | struct cgroup_subsys mem_cgroup_subsys = { |
5636 | .name = "memory", |
5637 | .subsys_id = mem_cgroup_subsys_id, |
5638 | .create = mem_cgroup_create, |
5639 | .pre_destroy = mem_cgroup_pre_destroy, |
5640 | .destroy = mem_cgroup_destroy, |
5641 | .populate = mem_cgroup_populate, |
5642 | .can_attach = mem_cgroup_can_attach, |
5643 | .cancel_attach = mem_cgroup_cancel_attach, |
5644 | .attach = mem_cgroup_move_task, |
5645 | .early_init = 0, |
5646 | .use_id = 1, |
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 |
5662 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9