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/mutex.h> |
37 | #include <linux/rbtree.h> |
38 | #include <linux/slab.h> |
39 | #include <linux/swap.h> |
40 | #include <linux/swapops.h> |
41 | #include <linux/spinlock.h> |
42 | #include <linux/eventfd.h> |
43 | #include <linux/sort.h> |
44 | #include <linux/fs.h> |
45 | #include <linux/seq_file.h> |
46 | #include <linux/vmalloc.h> |
47 | #include <linux/mm_inline.h> |
48 | #include <linux/page_cgroup.h> |
49 | #include <linux/cpu.h> |
50 | #include "internal.h" |
51 | |
52 | #include <asm/uaccess.h> |
53 | |
54 | struct cgroup_subsys mem_cgroup_subsys __read_mostly; |
55 | #define MEM_CGROUP_RECLAIM_RETRIES 5 |
56 | struct mem_cgroup *root_mem_cgroup __read_mostly; |
57 | |
58 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
59 | /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ |
60 | int do_swap_account __read_mostly; |
61 | static int really_do_swap_account __initdata = 1; /* for remember boot option*/ |
62 | #else |
63 | #define do_swap_account (0) |
64 | #endif |
65 | |
66 | /* |
67 | * Per memcg event counter is incremented at every pagein/pageout. This counter |
68 | * is used for trigger some periodic events. This is straightforward and better |
69 | * than using jiffies etc. to handle periodic memcg event. |
70 | * |
71 | * These values will be used as !((event) & ((1 <<(thresh)) - 1)) |
72 | */ |
73 | #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */ |
74 | #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */ |
75 | |
76 | /* |
77 | * Statistics for memory cgroup. |
78 | */ |
79 | enum mem_cgroup_stat_index { |
80 | /* |
81 | * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. |
82 | */ |
83 | MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ |
84 | MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ |
85 | MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ |
86 | MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */ |
87 | MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */ |
88 | MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ |
89 | MEM_CGROUP_EVENTS, /* incremented at every pagein/pageout */ |
90 | |
91 | MEM_CGROUP_STAT_NSTATS, |
92 | }; |
93 | |
94 | struct mem_cgroup_stat_cpu { |
95 | s64 count[MEM_CGROUP_STAT_NSTATS]; |
96 | }; |
97 | |
98 | /* |
99 | * per-zone information in memory controller. |
100 | */ |
101 | struct mem_cgroup_per_zone { |
102 | /* |
103 | * spin_lock to protect the per cgroup LRU |
104 | */ |
105 | struct list_head lists[NR_LRU_LISTS]; |
106 | unsigned long count[NR_LRU_LISTS]; |
107 | |
108 | struct zone_reclaim_stat reclaim_stat; |
109 | struct rb_node tree_node; /* RB tree node */ |
110 | unsigned long long usage_in_excess;/* Set to the value by which */ |
111 | /* the soft limit is exceeded*/ |
112 | bool on_tree; |
113 | struct mem_cgroup *mem; /* Back pointer, we cannot */ |
114 | /* use container_of */ |
115 | }; |
116 | /* Macro for accessing counter */ |
117 | #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) |
118 | |
119 | struct mem_cgroup_per_node { |
120 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; |
121 | }; |
122 | |
123 | struct mem_cgroup_lru_info { |
124 | struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; |
125 | }; |
126 | |
127 | /* |
128 | * Cgroups above their limits are maintained in a RB-Tree, independent of |
129 | * their hierarchy representation |
130 | */ |
131 | |
132 | struct mem_cgroup_tree_per_zone { |
133 | struct rb_root rb_root; |
134 | spinlock_t lock; |
135 | }; |
136 | |
137 | struct mem_cgroup_tree_per_node { |
138 | struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; |
139 | }; |
140 | |
141 | struct mem_cgroup_tree { |
142 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; |
143 | }; |
144 | |
145 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; |
146 | |
147 | struct mem_cgroup_threshold { |
148 | struct eventfd_ctx *eventfd; |
149 | u64 threshold; |
150 | }; |
151 | |
152 | struct mem_cgroup_threshold_ary { |
153 | /* An array index points to threshold just below usage. */ |
154 | atomic_t current_threshold; |
155 | /* Size of entries[] */ |
156 | unsigned int size; |
157 | /* Array of thresholds */ |
158 | struct mem_cgroup_threshold entries[0]; |
159 | }; |
160 | |
161 | static void mem_cgroup_threshold(struct mem_cgroup *mem); |
162 | |
163 | /* |
164 | * The memory controller data structure. The memory controller controls both |
165 | * page cache and RSS per cgroup. We would eventually like to provide |
166 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, |
167 | * to help the administrator determine what knobs to tune. |
168 | * |
169 | * TODO: Add a water mark for the memory controller. Reclaim will begin when |
170 | * we hit the water mark. May be even add a low water mark, such that |
171 | * no reclaim occurs from a cgroup at it's low water mark, this is |
172 | * a feature that will be implemented much later in the future. |
173 | */ |
174 | struct mem_cgroup { |
175 | struct cgroup_subsys_state css; |
176 | /* |
177 | * the counter to account for memory usage |
178 | */ |
179 | struct res_counter res; |
180 | /* |
181 | * the counter to account for mem+swap usage. |
182 | */ |
183 | struct res_counter memsw; |
184 | /* |
185 | * Per cgroup active and inactive list, similar to the |
186 | * per zone LRU lists. |
187 | */ |
188 | struct mem_cgroup_lru_info info; |
189 | |
190 | /* |
191 | protect against reclaim related member. |
192 | */ |
193 | spinlock_t reclaim_param_lock; |
194 | |
195 | int prev_priority; /* for recording reclaim priority */ |
196 | |
197 | /* |
198 | * While reclaiming in a hierarchy, we cache the last child we |
199 | * reclaimed from. |
200 | */ |
201 | int last_scanned_child; |
202 | /* |
203 | * Should the accounting and control be hierarchical, per subtree? |
204 | */ |
205 | bool use_hierarchy; |
206 | atomic_t oom_lock; |
207 | atomic_t refcnt; |
208 | |
209 | unsigned int swappiness; |
210 | |
211 | /* set when res.limit == memsw.limit */ |
212 | bool memsw_is_minimum; |
213 | |
214 | /* protect arrays of thresholds */ |
215 | struct mutex thresholds_lock; |
216 | |
217 | /* thresholds for memory usage. RCU-protected */ |
218 | struct mem_cgroup_threshold_ary *thresholds; |
219 | |
220 | /* thresholds for mem+swap usage. RCU-protected */ |
221 | struct mem_cgroup_threshold_ary *memsw_thresholds; |
222 | |
223 | /* |
224 | * Should we move charges of a task when a task is moved into this |
225 | * mem_cgroup ? And what type of charges should we move ? |
226 | */ |
227 | unsigned long move_charge_at_immigrate; |
228 | |
229 | /* |
230 | * percpu counter. |
231 | */ |
232 | struct mem_cgroup_stat_cpu *stat; |
233 | }; |
234 | |
235 | /* Stuffs for move charges at task migration. */ |
236 | /* |
237 | * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a |
238 | * left-shifted bitmap of these types. |
239 | */ |
240 | enum move_type { |
241 | MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ |
242 | NR_MOVE_TYPE, |
243 | }; |
244 | |
245 | /* "mc" and its members are protected by cgroup_mutex */ |
246 | static struct move_charge_struct { |
247 | struct mem_cgroup *from; |
248 | struct mem_cgroup *to; |
249 | unsigned long precharge; |
250 | unsigned long moved_charge; |
251 | unsigned long moved_swap; |
252 | struct task_struct *moving_task; /* a task moving charges */ |
253 | wait_queue_head_t waitq; /* a waitq for other context */ |
254 | } mc = { |
255 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), |
256 | }; |
257 | |
258 | /* |
259 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft |
260 | * limit reclaim to prevent infinite loops, if they ever occur. |
261 | */ |
262 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) |
263 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) |
264 | |
265 | enum charge_type { |
266 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, |
267 | MEM_CGROUP_CHARGE_TYPE_MAPPED, |
268 | MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ |
269 | MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ |
270 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ |
271 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ |
272 | NR_CHARGE_TYPE, |
273 | }; |
274 | |
275 | /* only for here (for easy reading.) */ |
276 | #define PCGF_CACHE (1UL << PCG_CACHE) |
277 | #define PCGF_USED (1UL << PCG_USED) |
278 | #define PCGF_LOCK (1UL << PCG_LOCK) |
279 | /* Not used, but added here for completeness */ |
280 | #define PCGF_ACCT (1UL << PCG_ACCT) |
281 | |
282 | /* for encoding cft->private value on file */ |
283 | #define _MEM (0) |
284 | #define _MEMSWAP (1) |
285 | #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) |
286 | #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) |
287 | #define MEMFILE_ATTR(val) ((val) & 0xffff) |
288 | |
289 | /* |
290 | * Reclaim flags for mem_cgroup_hierarchical_reclaim |
291 | */ |
292 | #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 |
293 | #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) |
294 | #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 |
295 | #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) |
296 | #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2 |
297 | #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT) |
298 | |
299 | static void mem_cgroup_get(struct mem_cgroup *mem); |
300 | static void mem_cgroup_put(struct mem_cgroup *mem); |
301 | static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); |
302 | static void drain_all_stock_async(void); |
303 | |
304 | static struct mem_cgroup_per_zone * |
305 | mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) |
306 | { |
307 | return &mem->info.nodeinfo[nid]->zoneinfo[zid]; |
308 | } |
309 | |
310 | struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem) |
311 | { |
312 | return &mem->css; |
313 | } |
314 | |
315 | static struct mem_cgroup_per_zone * |
316 | page_cgroup_zoneinfo(struct page_cgroup *pc) |
317 | { |
318 | struct mem_cgroup *mem = pc->mem_cgroup; |
319 | int nid = page_cgroup_nid(pc); |
320 | int zid = page_cgroup_zid(pc); |
321 | |
322 | if (!mem) |
323 | return NULL; |
324 | |
325 | return mem_cgroup_zoneinfo(mem, nid, zid); |
326 | } |
327 | |
328 | static struct mem_cgroup_tree_per_zone * |
329 | soft_limit_tree_node_zone(int nid, int zid) |
330 | { |
331 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; |
332 | } |
333 | |
334 | static struct mem_cgroup_tree_per_zone * |
335 | soft_limit_tree_from_page(struct page *page) |
336 | { |
337 | int nid = page_to_nid(page); |
338 | int zid = page_zonenum(page); |
339 | |
340 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; |
341 | } |
342 | |
343 | static void |
344 | __mem_cgroup_insert_exceeded(struct mem_cgroup *mem, |
345 | struct mem_cgroup_per_zone *mz, |
346 | struct mem_cgroup_tree_per_zone *mctz, |
347 | unsigned long long new_usage_in_excess) |
348 | { |
349 | struct rb_node **p = &mctz->rb_root.rb_node; |
350 | struct rb_node *parent = NULL; |
351 | struct mem_cgroup_per_zone *mz_node; |
352 | |
353 | if (mz->on_tree) |
354 | return; |
355 | |
356 | mz->usage_in_excess = new_usage_in_excess; |
357 | if (!mz->usage_in_excess) |
358 | return; |
359 | while (*p) { |
360 | parent = *p; |
361 | mz_node = rb_entry(parent, struct mem_cgroup_per_zone, |
362 | tree_node); |
363 | if (mz->usage_in_excess < mz_node->usage_in_excess) |
364 | p = &(*p)->rb_left; |
365 | /* |
366 | * We can't avoid mem cgroups that are over their soft |
367 | * limit by the same amount |
368 | */ |
369 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) |
370 | p = &(*p)->rb_right; |
371 | } |
372 | rb_link_node(&mz->tree_node, parent, p); |
373 | rb_insert_color(&mz->tree_node, &mctz->rb_root); |
374 | mz->on_tree = true; |
375 | } |
376 | |
377 | static void |
378 | __mem_cgroup_remove_exceeded(struct mem_cgroup *mem, |
379 | struct mem_cgroup_per_zone *mz, |
380 | struct mem_cgroup_tree_per_zone *mctz) |
381 | { |
382 | if (!mz->on_tree) |
383 | return; |
384 | rb_erase(&mz->tree_node, &mctz->rb_root); |
385 | mz->on_tree = false; |
386 | } |
387 | |
388 | static void |
389 | mem_cgroup_remove_exceeded(struct mem_cgroup *mem, |
390 | struct mem_cgroup_per_zone *mz, |
391 | struct mem_cgroup_tree_per_zone *mctz) |
392 | { |
393 | spin_lock(&mctz->lock); |
394 | __mem_cgroup_remove_exceeded(mem, mz, mctz); |
395 | spin_unlock(&mctz->lock); |
396 | } |
397 | |
398 | |
399 | static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page) |
400 | { |
401 | unsigned long long excess; |
402 | struct mem_cgroup_per_zone *mz; |
403 | struct mem_cgroup_tree_per_zone *mctz; |
404 | int nid = page_to_nid(page); |
405 | int zid = page_zonenum(page); |
406 | mctz = soft_limit_tree_from_page(page); |
407 | |
408 | /* |
409 | * Necessary to update all ancestors when hierarchy is used. |
410 | * because their event counter is not touched. |
411 | */ |
412 | for (; mem; mem = parent_mem_cgroup(mem)) { |
413 | mz = mem_cgroup_zoneinfo(mem, nid, zid); |
414 | excess = res_counter_soft_limit_excess(&mem->res); |
415 | /* |
416 | * We have to update the tree if mz is on RB-tree or |
417 | * mem is over its softlimit. |
418 | */ |
419 | if (excess || mz->on_tree) { |
420 | spin_lock(&mctz->lock); |
421 | /* if on-tree, remove it */ |
422 | if (mz->on_tree) |
423 | __mem_cgroup_remove_exceeded(mem, mz, mctz); |
424 | /* |
425 | * Insert again. mz->usage_in_excess will be updated. |
426 | * If excess is 0, no tree ops. |
427 | */ |
428 | __mem_cgroup_insert_exceeded(mem, mz, mctz, excess); |
429 | spin_unlock(&mctz->lock); |
430 | } |
431 | } |
432 | } |
433 | |
434 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem) |
435 | { |
436 | int node, zone; |
437 | struct mem_cgroup_per_zone *mz; |
438 | struct mem_cgroup_tree_per_zone *mctz; |
439 | |
440 | for_each_node_state(node, N_POSSIBLE) { |
441 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
442 | mz = mem_cgroup_zoneinfo(mem, node, zone); |
443 | mctz = soft_limit_tree_node_zone(node, zone); |
444 | mem_cgroup_remove_exceeded(mem, mz, mctz); |
445 | } |
446 | } |
447 | } |
448 | |
449 | static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem) |
450 | { |
451 | return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT; |
452 | } |
453 | |
454 | static struct mem_cgroup_per_zone * |
455 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) |
456 | { |
457 | struct rb_node *rightmost = NULL; |
458 | struct mem_cgroup_per_zone *mz; |
459 | |
460 | retry: |
461 | mz = NULL; |
462 | rightmost = rb_last(&mctz->rb_root); |
463 | if (!rightmost) |
464 | goto done; /* Nothing to reclaim from */ |
465 | |
466 | mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); |
467 | /* |
468 | * Remove the node now but someone else can add it back, |
469 | * we will to add it back at the end of reclaim to its correct |
470 | * position in the tree. |
471 | */ |
472 | __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); |
473 | if (!res_counter_soft_limit_excess(&mz->mem->res) || |
474 | !css_tryget(&mz->mem->css)) |
475 | goto retry; |
476 | done: |
477 | return mz; |
478 | } |
479 | |
480 | static struct mem_cgroup_per_zone * |
481 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) |
482 | { |
483 | struct mem_cgroup_per_zone *mz; |
484 | |
485 | spin_lock(&mctz->lock); |
486 | mz = __mem_cgroup_largest_soft_limit_node(mctz); |
487 | spin_unlock(&mctz->lock); |
488 | return mz; |
489 | } |
490 | |
491 | static s64 mem_cgroup_read_stat(struct mem_cgroup *mem, |
492 | enum mem_cgroup_stat_index idx) |
493 | { |
494 | int cpu; |
495 | s64 val = 0; |
496 | |
497 | for_each_possible_cpu(cpu) |
498 | val += per_cpu(mem->stat->count[idx], cpu); |
499 | return val; |
500 | } |
501 | |
502 | static s64 mem_cgroup_local_usage(struct mem_cgroup *mem) |
503 | { |
504 | s64 ret; |
505 | |
506 | ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); |
507 | ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); |
508 | return ret; |
509 | } |
510 | |
511 | static void mem_cgroup_swap_statistics(struct mem_cgroup *mem, |
512 | bool charge) |
513 | { |
514 | int val = (charge) ? 1 : -1; |
515 | this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); |
516 | } |
517 | |
518 | static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, |
519 | struct page_cgroup *pc, |
520 | bool charge) |
521 | { |
522 | int val = (charge) ? 1 : -1; |
523 | |
524 | preempt_disable(); |
525 | |
526 | if (PageCgroupCache(pc)) |
527 | __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val); |
528 | else |
529 | __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val); |
530 | |
531 | if (charge) |
532 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]); |
533 | else |
534 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]); |
535 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]); |
536 | |
537 | preempt_enable(); |
538 | } |
539 | |
540 | static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, |
541 | enum lru_list idx) |
542 | { |
543 | int nid, zid; |
544 | struct mem_cgroup_per_zone *mz; |
545 | u64 total = 0; |
546 | |
547 | for_each_online_node(nid) |
548 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
549 | mz = mem_cgroup_zoneinfo(mem, nid, zid); |
550 | total += MEM_CGROUP_ZSTAT(mz, idx); |
551 | } |
552 | return total; |
553 | } |
554 | |
555 | static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift) |
556 | { |
557 | s64 val; |
558 | |
559 | val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]); |
560 | |
561 | return !(val & ((1 << event_mask_shift) - 1)); |
562 | } |
563 | |
564 | /* |
565 | * Check events in order. |
566 | * |
567 | */ |
568 | static void memcg_check_events(struct mem_cgroup *mem, struct page *page) |
569 | { |
570 | /* threshold event is triggered in finer grain than soft limit */ |
571 | if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) { |
572 | mem_cgroup_threshold(mem); |
573 | if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH))) |
574 | mem_cgroup_update_tree(mem, page); |
575 | } |
576 | } |
577 | |
578 | static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) |
579 | { |
580 | return container_of(cgroup_subsys_state(cont, |
581 | mem_cgroup_subsys_id), struct mem_cgroup, |
582 | css); |
583 | } |
584 | |
585 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) |
586 | { |
587 | /* |
588 | * mm_update_next_owner() may clear mm->owner to NULL |
589 | * if it races with swapoff, page migration, etc. |
590 | * So this can be called with p == NULL. |
591 | */ |
592 | if (unlikely(!p)) |
593 | return NULL; |
594 | |
595 | return container_of(task_subsys_state(p, mem_cgroup_subsys_id), |
596 | struct mem_cgroup, css); |
597 | } |
598 | |
599 | static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) |
600 | { |
601 | struct mem_cgroup *mem = NULL; |
602 | |
603 | if (!mm) |
604 | return NULL; |
605 | /* |
606 | * Because we have no locks, mm->owner's may be being moved to other |
607 | * cgroup. We use css_tryget() here even if this looks |
608 | * pessimistic (rather than adding locks here). |
609 | */ |
610 | rcu_read_lock(); |
611 | do { |
612 | mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); |
613 | if (unlikely(!mem)) |
614 | break; |
615 | } while (!css_tryget(&mem->css)); |
616 | rcu_read_unlock(); |
617 | return mem; |
618 | } |
619 | |
620 | /* |
621 | * Call callback function against all cgroup under hierarchy tree. |
622 | */ |
623 | static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data, |
624 | int (*func)(struct mem_cgroup *, void *)) |
625 | { |
626 | int found, ret, nextid; |
627 | struct cgroup_subsys_state *css; |
628 | struct mem_cgroup *mem; |
629 | |
630 | if (!root->use_hierarchy) |
631 | return (*func)(root, data); |
632 | |
633 | nextid = 1; |
634 | do { |
635 | ret = 0; |
636 | mem = NULL; |
637 | |
638 | rcu_read_lock(); |
639 | css = css_get_next(&mem_cgroup_subsys, nextid, &root->css, |
640 | &found); |
641 | if (css && css_tryget(css)) |
642 | mem = container_of(css, struct mem_cgroup, css); |
643 | rcu_read_unlock(); |
644 | |
645 | if (mem) { |
646 | ret = (*func)(mem, data); |
647 | css_put(&mem->css); |
648 | } |
649 | nextid = found + 1; |
650 | } while (!ret && css); |
651 | |
652 | return ret; |
653 | } |
654 | |
655 | static inline bool mem_cgroup_is_root(struct mem_cgroup *mem) |
656 | { |
657 | return (mem == root_mem_cgroup); |
658 | } |
659 | |
660 | /* |
661 | * Following LRU functions are allowed to be used without PCG_LOCK. |
662 | * Operations are called by routine of global LRU independently from memcg. |
663 | * What we have to take care of here is validness of pc->mem_cgroup. |
664 | * |
665 | * Changes to pc->mem_cgroup happens when |
666 | * 1. charge |
667 | * 2. moving account |
668 | * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. |
669 | * It is added to LRU before charge. |
670 | * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. |
671 | * When moving account, the page is not on LRU. It's isolated. |
672 | */ |
673 | |
674 | void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) |
675 | { |
676 | struct page_cgroup *pc; |
677 | struct mem_cgroup_per_zone *mz; |
678 | |
679 | if (mem_cgroup_disabled()) |
680 | return; |
681 | pc = lookup_page_cgroup(page); |
682 | /* can happen while we handle swapcache. */ |
683 | if (!TestClearPageCgroupAcctLRU(pc)) |
684 | return; |
685 | VM_BUG_ON(!pc->mem_cgroup); |
686 | /* |
687 | * We don't check PCG_USED bit. It's cleared when the "page" is finally |
688 | * removed from global LRU. |
689 | */ |
690 | mz = page_cgroup_zoneinfo(pc); |
691 | MEM_CGROUP_ZSTAT(mz, lru) -= 1; |
692 | if (mem_cgroup_is_root(pc->mem_cgroup)) |
693 | return; |
694 | VM_BUG_ON(list_empty(&pc->lru)); |
695 | list_del_init(&pc->lru); |
696 | return; |
697 | } |
698 | |
699 | void mem_cgroup_del_lru(struct page *page) |
700 | { |
701 | mem_cgroup_del_lru_list(page, page_lru(page)); |
702 | } |
703 | |
704 | void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru) |
705 | { |
706 | struct mem_cgroup_per_zone *mz; |
707 | struct page_cgroup *pc; |
708 | |
709 | if (mem_cgroup_disabled()) |
710 | return; |
711 | |
712 | pc = lookup_page_cgroup(page); |
713 | /* |
714 | * Used bit is set without atomic ops but after smp_wmb(). |
715 | * For making pc->mem_cgroup visible, insert smp_rmb() here. |
716 | */ |
717 | smp_rmb(); |
718 | /* unused or root page is not rotated. */ |
719 | if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup)) |
720 | return; |
721 | mz = page_cgroup_zoneinfo(pc); |
722 | list_move(&pc->lru, &mz->lists[lru]); |
723 | } |
724 | |
725 | void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru) |
726 | { |
727 | struct page_cgroup *pc; |
728 | struct mem_cgroup_per_zone *mz; |
729 | |
730 | if (mem_cgroup_disabled()) |
731 | return; |
732 | pc = lookup_page_cgroup(page); |
733 | VM_BUG_ON(PageCgroupAcctLRU(pc)); |
734 | /* |
735 | * Used bit is set without atomic ops but after smp_wmb(). |
736 | * For making pc->mem_cgroup visible, insert smp_rmb() here. |
737 | */ |
738 | smp_rmb(); |
739 | if (!PageCgroupUsed(pc)) |
740 | return; |
741 | |
742 | mz = page_cgroup_zoneinfo(pc); |
743 | MEM_CGROUP_ZSTAT(mz, lru) += 1; |
744 | SetPageCgroupAcctLRU(pc); |
745 | if (mem_cgroup_is_root(pc->mem_cgroup)) |
746 | return; |
747 | list_add(&pc->lru, &mz->lists[lru]); |
748 | } |
749 | |
750 | /* |
751 | * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to |
752 | * lru because the page may.be reused after it's fully uncharged (because of |
753 | * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge |
754 | * it again. This function is only used to charge SwapCache. It's done under |
755 | * lock_page and expected that zone->lru_lock is never held. |
756 | */ |
757 | static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page) |
758 | { |
759 | unsigned long flags; |
760 | struct zone *zone = page_zone(page); |
761 | struct page_cgroup *pc = lookup_page_cgroup(page); |
762 | |
763 | spin_lock_irqsave(&zone->lru_lock, flags); |
764 | /* |
765 | * Forget old LRU when this page_cgroup is *not* used. This Used bit |
766 | * is guarded by lock_page() because the page is SwapCache. |
767 | */ |
768 | if (!PageCgroupUsed(pc)) |
769 | mem_cgroup_del_lru_list(page, page_lru(page)); |
770 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
771 | } |
772 | |
773 | static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page) |
774 | { |
775 | unsigned long flags; |
776 | struct zone *zone = page_zone(page); |
777 | struct page_cgroup *pc = lookup_page_cgroup(page); |
778 | |
779 | spin_lock_irqsave(&zone->lru_lock, flags); |
780 | /* link when the page is linked to LRU but page_cgroup isn't */ |
781 | if (PageLRU(page) && !PageCgroupAcctLRU(pc)) |
782 | mem_cgroup_add_lru_list(page, page_lru(page)); |
783 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
784 | } |
785 | |
786 | |
787 | void mem_cgroup_move_lists(struct page *page, |
788 | enum lru_list from, enum lru_list to) |
789 | { |
790 | if (mem_cgroup_disabled()) |
791 | return; |
792 | mem_cgroup_del_lru_list(page, from); |
793 | mem_cgroup_add_lru_list(page, to); |
794 | } |
795 | |
796 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) |
797 | { |
798 | int ret; |
799 | struct mem_cgroup *curr = NULL; |
800 | |
801 | task_lock(task); |
802 | rcu_read_lock(); |
803 | curr = try_get_mem_cgroup_from_mm(task->mm); |
804 | rcu_read_unlock(); |
805 | task_unlock(task); |
806 | if (!curr) |
807 | return 0; |
808 | /* |
809 | * We should check use_hierarchy of "mem" not "curr". Because checking |
810 | * use_hierarchy of "curr" here make this function true if hierarchy is |
811 | * enabled in "curr" and "curr" is a child of "mem" in *cgroup* |
812 | * hierarchy(even if use_hierarchy is disabled in "mem"). |
813 | */ |
814 | if (mem->use_hierarchy) |
815 | ret = css_is_ancestor(&curr->css, &mem->css); |
816 | else |
817 | ret = (curr == mem); |
818 | css_put(&curr->css); |
819 | return ret; |
820 | } |
821 | |
822 | /* |
823 | * prev_priority control...this will be used in memory reclaim path. |
824 | */ |
825 | int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem) |
826 | { |
827 | int prev_priority; |
828 | |
829 | spin_lock(&mem->reclaim_param_lock); |
830 | prev_priority = mem->prev_priority; |
831 | spin_unlock(&mem->reclaim_param_lock); |
832 | |
833 | return prev_priority; |
834 | } |
835 | |
836 | void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority) |
837 | { |
838 | spin_lock(&mem->reclaim_param_lock); |
839 | if (priority < mem->prev_priority) |
840 | mem->prev_priority = priority; |
841 | spin_unlock(&mem->reclaim_param_lock); |
842 | } |
843 | |
844 | void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority) |
845 | { |
846 | spin_lock(&mem->reclaim_param_lock); |
847 | mem->prev_priority = priority; |
848 | spin_unlock(&mem->reclaim_param_lock); |
849 | } |
850 | |
851 | static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) |
852 | { |
853 | unsigned long active; |
854 | unsigned long inactive; |
855 | unsigned long gb; |
856 | unsigned long inactive_ratio; |
857 | |
858 | inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); |
859 | active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); |
860 | |
861 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
862 | if (gb) |
863 | inactive_ratio = int_sqrt(10 * gb); |
864 | else |
865 | inactive_ratio = 1; |
866 | |
867 | if (present_pages) { |
868 | present_pages[0] = inactive; |
869 | present_pages[1] = active; |
870 | } |
871 | |
872 | return inactive_ratio; |
873 | } |
874 | |
875 | int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) |
876 | { |
877 | unsigned long active; |
878 | unsigned long inactive; |
879 | unsigned long present_pages[2]; |
880 | unsigned long inactive_ratio; |
881 | |
882 | inactive_ratio = calc_inactive_ratio(memcg, present_pages); |
883 | |
884 | inactive = present_pages[0]; |
885 | active = present_pages[1]; |
886 | |
887 | if (inactive * inactive_ratio < active) |
888 | return 1; |
889 | |
890 | return 0; |
891 | } |
892 | |
893 | int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) |
894 | { |
895 | unsigned long active; |
896 | unsigned long inactive; |
897 | |
898 | inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); |
899 | active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); |
900 | |
901 | return (active > inactive); |
902 | } |
903 | |
904 | unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg, |
905 | struct zone *zone, |
906 | enum lru_list lru) |
907 | { |
908 | int nid = zone->zone_pgdat->node_id; |
909 | int zid = zone_idx(zone); |
910 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
911 | |
912 | return MEM_CGROUP_ZSTAT(mz, lru); |
913 | } |
914 | |
915 | struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, |
916 | struct zone *zone) |
917 | { |
918 | int nid = zone->zone_pgdat->node_id; |
919 | int zid = zone_idx(zone); |
920 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); |
921 | |
922 | return &mz->reclaim_stat; |
923 | } |
924 | |
925 | struct zone_reclaim_stat * |
926 | mem_cgroup_get_reclaim_stat_from_page(struct page *page) |
927 | { |
928 | struct page_cgroup *pc; |
929 | struct mem_cgroup_per_zone *mz; |
930 | |
931 | if (mem_cgroup_disabled()) |
932 | return NULL; |
933 | |
934 | pc = lookup_page_cgroup(page); |
935 | /* |
936 | * Used bit is set without atomic ops but after smp_wmb(). |
937 | * For making pc->mem_cgroup visible, insert smp_rmb() here. |
938 | */ |
939 | smp_rmb(); |
940 | if (!PageCgroupUsed(pc)) |
941 | return NULL; |
942 | |
943 | mz = page_cgroup_zoneinfo(pc); |
944 | if (!mz) |
945 | return NULL; |
946 | |
947 | return &mz->reclaim_stat; |
948 | } |
949 | |
950 | unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, |
951 | struct list_head *dst, |
952 | unsigned long *scanned, int order, |
953 | int mode, struct zone *z, |
954 | struct mem_cgroup *mem_cont, |
955 | int active, int file) |
956 | { |
957 | unsigned long nr_taken = 0; |
958 | struct page *page; |
959 | unsigned long scan; |
960 | LIST_HEAD(pc_list); |
961 | struct list_head *src; |
962 | struct page_cgroup *pc, *tmp; |
963 | int nid = z->zone_pgdat->node_id; |
964 | int zid = zone_idx(z); |
965 | struct mem_cgroup_per_zone *mz; |
966 | int lru = LRU_FILE * file + active; |
967 | int ret; |
968 | |
969 | BUG_ON(!mem_cont); |
970 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); |
971 | src = &mz->lists[lru]; |
972 | |
973 | scan = 0; |
974 | list_for_each_entry_safe_reverse(pc, tmp, src, lru) { |
975 | if (scan >= nr_to_scan) |
976 | break; |
977 | |
978 | page = pc->page; |
979 | if (unlikely(!PageCgroupUsed(pc))) |
980 | continue; |
981 | if (unlikely(!PageLRU(page))) |
982 | continue; |
983 | |
984 | scan++; |
985 | ret = __isolate_lru_page(page, mode, file); |
986 | switch (ret) { |
987 | case 0: |
988 | list_move(&page->lru, dst); |
989 | mem_cgroup_del_lru(page); |
990 | nr_taken++; |
991 | break; |
992 | case -EBUSY: |
993 | /* we don't affect global LRU but rotate in our LRU */ |
994 | mem_cgroup_rotate_lru_list(page, page_lru(page)); |
995 | break; |
996 | default: |
997 | break; |
998 | } |
999 | } |
1000 | |
1001 | *scanned = scan; |
1002 | return nr_taken; |
1003 | } |
1004 | |
1005 | #define mem_cgroup_from_res_counter(counter, member) \ |
1006 | container_of(counter, struct mem_cgroup, member) |
1007 | |
1008 | static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem) |
1009 | { |
1010 | if (do_swap_account) { |
1011 | if (res_counter_check_under_limit(&mem->res) && |
1012 | res_counter_check_under_limit(&mem->memsw)) |
1013 | return true; |
1014 | } else |
1015 | if (res_counter_check_under_limit(&mem->res)) |
1016 | return true; |
1017 | return false; |
1018 | } |
1019 | |
1020 | static unsigned int get_swappiness(struct mem_cgroup *memcg) |
1021 | { |
1022 | struct cgroup *cgrp = memcg->css.cgroup; |
1023 | unsigned int swappiness; |
1024 | |
1025 | /* root ? */ |
1026 | if (cgrp->parent == NULL) |
1027 | return vm_swappiness; |
1028 | |
1029 | spin_lock(&memcg->reclaim_param_lock); |
1030 | swappiness = memcg->swappiness; |
1031 | spin_unlock(&memcg->reclaim_param_lock); |
1032 | |
1033 | return swappiness; |
1034 | } |
1035 | |
1036 | static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data) |
1037 | { |
1038 | int *val = data; |
1039 | (*val)++; |
1040 | return 0; |
1041 | } |
1042 | |
1043 | /** |
1044 | * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. |
1045 | * @memcg: The memory cgroup that went over limit |
1046 | * @p: Task that is going to be killed |
1047 | * |
1048 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is |
1049 | * enabled |
1050 | */ |
1051 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) |
1052 | { |
1053 | struct cgroup *task_cgrp; |
1054 | struct cgroup *mem_cgrp; |
1055 | /* |
1056 | * Need a buffer in BSS, can't rely on allocations. The code relies |
1057 | * on the assumption that OOM is serialized for memory controller. |
1058 | * If this assumption is broken, revisit this code. |
1059 | */ |
1060 | static char memcg_name[PATH_MAX]; |
1061 | int ret; |
1062 | |
1063 | if (!memcg || !p) |
1064 | return; |
1065 | |
1066 | |
1067 | rcu_read_lock(); |
1068 | |
1069 | mem_cgrp = memcg->css.cgroup; |
1070 | task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); |
1071 | |
1072 | ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); |
1073 | if (ret < 0) { |
1074 | /* |
1075 | * Unfortunately, we are unable to convert to a useful name |
1076 | * But we'll still print out the usage information |
1077 | */ |
1078 | rcu_read_unlock(); |
1079 | goto done; |
1080 | } |
1081 | rcu_read_unlock(); |
1082 | |
1083 | printk(KERN_INFO "Task in %s killed", memcg_name); |
1084 | |
1085 | rcu_read_lock(); |
1086 | ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); |
1087 | if (ret < 0) { |
1088 | rcu_read_unlock(); |
1089 | goto done; |
1090 | } |
1091 | rcu_read_unlock(); |
1092 | |
1093 | /* |
1094 | * Continues from above, so we don't need an KERN_ level |
1095 | */ |
1096 | printk(KERN_CONT " as a result of limit of %s\n", memcg_name); |
1097 | done: |
1098 | |
1099 | printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", |
1100 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, |
1101 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, |
1102 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); |
1103 | printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " |
1104 | "failcnt %llu\n", |
1105 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, |
1106 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, |
1107 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); |
1108 | } |
1109 | |
1110 | /* |
1111 | * This function returns the number of memcg under hierarchy tree. Returns |
1112 | * 1(self count) if no children. |
1113 | */ |
1114 | static int mem_cgroup_count_children(struct mem_cgroup *mem) |
1115 | { |
1116 | int num = 0; |
1117 | mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb); |
1118 | return num; |
1119 | } |
1120 | |
1121 | /* |
1122 | * Visit the first child (need not be the first child as per the ordering |
1123 | * of the cgroup list, since we track last_scanned_child) of @mem and use |
1124 | * that to reclaim free pages from. |
1125 | */ |
1126 | static struct mem_cgroup * |
1127 | mem_cgroup_select_victim(struct mem_cgroup *root_mem) |
1128 | { |
1129 | struct mem_cgroup *ret = NULL; |
1130 | struct cgroup_subsys_state *css; |
1131 | int nextid, found; |
1132 | |
1133 | if (!root_mem->use_hierarchy) { |
1134 | css_get(&root_mem->css); |
1135 | ret = root_mem; |
1136 | } |
1137 | |
1138 | while (!ret) { |
1139 | rcu_read_lock(); |
1140 | nextid = root_mem->last_scanned_child + 1; |
1141 | css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, |
1142 | &found); |
1143 | if (css && css_tryget(css)) |
1144 | ret = container_of(css, struct mem_cgroup, css); |
1145 | |
1146 | rcu_read_unlock(); |
1147 | /* Updates scanning parameter */ |
1148 | spin_lock(&root_mem->reclaim_param_lock); |
1149 | if (!css) { |
1150 | /* this means start scan from ID:1 */ |
1151 | root_mem->last_scanned_child = 0; |
1152 | } else |
1153 | root_mem->last_scanned_child = found; |
1154 | spin_unlock(&root_mem->reclaim_param_lock); |
1155 | } |
1156 | |
1157 | return ret; |
1158 | } |
1159 | |
1160 | /* |
1161 | * Scan the hierarchy if needed to reclaim memory. We remember the last child |
1162 | * we reclaimed from, so that we don't end up penalizing one child extensively |
1163 | * based on its position in the children list. |
1164 | * |
1165 | * root_mem is the original ancestor that we've been reclaim from. |
1166 | * |
1167 | * We give up and return to the caller when we visit root_mem twice. |
1168 | * (other groups can be removed while we're walking....) |
1169 | * |
1170 | * If shrink==true, for avoiding to free too much, this returns immedieately. |
1171 | */ |
1172 | static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, |
1173 | struct zone *zone, |
1174 | gfp_t gfp_mask, |
1175 | unsigned long reclaim_options) |
1176 | { |
1177 | struct mem_cgroup *victim; |
1178 | int ret, total = 0; |
1179 | int loop = 0; |
1180 | bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP; |
1181 | bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK; |
1182 | bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT; |
1183 | unsigned long excess = mem_cgroup_get_excess(root_mem); |
1184 | |
1185 | /* If memsw_is_minimum==1, swap-out is of-no-use. */ |
1186 | if (root_mem->memsw_is_minimum) |
1187 | noswap = true; |
1188 | |
1189 | while (1) { |
1190 | victim = mem_cgroup_select_victim(root_mem); |
1191 | if (victim == root_mem) { |
1192 | loop++; |
1193 | if (loop >= 1) |
1194 | drain_all_stock_async(); |
1195 | if (loop >= 2) { |
1196 | /* |
1197 | * If we have not been able to reclaim |
1198 | * anything, it might because there are |
1199 | * no reclaimable pages under this hierarchy |
1200 | */ |
1201 | if (!check_soft || !total) { |
1202 | css_put(&victim->css); |
1203 | break; |
1204 | } |
1205 | /* |
1206 | * We want to do more targetted reclaim. |
1207 | * excess >> 2 is not to excessive so as to |
1208 | * reclaim too much, nor too less that we keep |
1209 | * coming back to reclaim from this cgroup |
1210 | */ |
1211 | if (total >= (excess >> 2) || |
1212 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) { |
1213 | css_put(&victim->css); |
1214 | break; |
1215 | } |
1216 | } |
1217 | } |
1218 | if (!mem_cgroup_local_usage(victim)) { |
1219 | /* this cgroup's local usage == 0 */ |
1220 | css_put(&victim->css); |
1221 | continue; |
1222 | } |
1223 | /* we use swappiness of local cgroup */ |
1224 | if (check_soft) |
1225 | ret = mem_cgroup_shrink_node_zone(victim, gfp_mask, |
1226 | noswap, get_swappiness(victim), zone, |
1227 | zone->zone_pgdat->node_id); |
1228 | else |
1229 | ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, |
1230 | noswap, get_swappiness(victim)); |
1231 | css_put(&victim->css); |
1232 | /* |
1233 | * At shrinking usage, we can't check we should stop here or |
1234 | * reclaim more. It's depends on callers. last_scanned_child |
1235 | * will work enough for keeping fairness under tree. |
1236 | */ |
1237 | if (shrink) |
1238 | return ret; |
1239 | total += ret; |
1240 | if (check_soft) { |
1241 | if (res_counter_check_under_soft_limit(&root_mem->res)) |
1242 | return total; |
1243 | } else if (mem_cgroup_check_under_limit(root_mem)) |
1244 | return 1 + total; |
1245 | } |
1246 | return total; |
1247 | } |
1248 | |
1249 | static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data) |
1250 | { |
1251 | int *val = (int *)data; |
1252 | int x; |
1253 | /* |
1254 | * Logically, we can stop scanning immediately when we find |
1255 | * a memcg is already locked. But condidering unlock ops and |
1256 | * creation/removal of memcg, scan-all is simple operation. |
1257 | */ |
1258 | x = atomic_inc_return(&mem->oom_lock); |
1259 | *val = max(x, *val); |
1260 | return 0; |
1261 | } |
1262 | /* |
1263 | * Check OOM-Killer is already running under our hierarchy. |
1264 | * If someone is running, return false. |
1265 | */ |
1266 | static bool mem_cgroup_oom_lock(struct mem_cgroup *mem) |
1267 | { |
1268 | int lock_count = 0; |
1269 | |
1270 | mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb); |
1271 | |
1272 | if (lock_count == 1) |
1273 | return true; |
1274 | return false; |
1275 | } |
1276 | |
1277 | static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data) |
1278 | { |
1279 | /* |
1280 | * When a new child is created while the hierarchy is under oom, |
1281 | * mem_cgroup_oom_lock() may not be called. We have to use |
1282 | * atomic_add_unless() here. |
1283 | */ |
1284 | atomic_add_unless(&mem->oom_lock, -1, 0); |
1285 | return 0; |
1286 | } |
1287 | |
1288 | static void mem_cgroup_oom_unlock(struct mem_cgroup *mem) |
1289 | { |
1290 | mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb); |
1291 | } |
1292 | |
1293 | static DEFINE_MUTEX(memcg_oom_mutex); |
1294 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); |
1295 | |
1296 | /* |
1297 | * try to call OOM killer. returns false if we should exit memory-reclaim loop. |
1298 | */ |
1299 | bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask) |
1300 | { |
1301 | DEFINE_WAIT(wait); |
1302 | bool locked; |
1303 | |
1304 | /* At first, try to OOM lock hierarchy under mem.*/ |
1305 | mutex_lock(&memcg_oom_mutex); |
1306 | locked = mem_cgroup_oom_lock(mem); |
1307 | /* |
1308 | * Even if signal_pending(), we can't quit charge() loop without |
1309 | * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL |
1310 | * under OOM is always welcomed, use TASK_KILLABLE here. |
1311 | */ |
1312 | if (!locked) |
1313 | prepare_to_wait(&memcg_oom_waitq, &wait, TASK_KILLABLE); |
1314 | mutex_unlock(&memcg_oom_mutex); |
1315 | |
1316 | if (locked) |
1317 | mem_cgroup_out_of_memory(mem, mask); |
1318 | else { |
1319 | schedule(); |
1320 | finish_wait(&memcg_oom_waitq, &wait); |
1321 | } |
1322 | mutex_lock(&memcg_oom_mutex); |
1323 | mem_cgroup_oom_unlock(mem); |
1324 | /* |
1325 | * Here, we use global waitq .....more fine grained waitq ? |
1326 | * Assume following hierarchy. |
1327 | * A/ |
1328 | * 01 |
1329 | * 02 |
1330 | * assume OOM happens both in A and 01 at the same time. Tthey are |
1331 | * mutually exclusive by lock. (kill in 01 helps A.) |
1332 | * When we use per memcg waitq, we have to wake up waiters on A and 02 |
1333 | * in addtion to waiters on 01. We use global waitq for avoiding mess. |
1334 | * It will not be a big problem. |
1335 | * (And a task may be moved to other groups while it's waiting for OOM.) |
1336 | */ |
1337 | wake_up_all(&memcg_oom_waitq); |
1338 | mutex_unlock(&memcg_oom_mutex); |
1339 | |
1340 | if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) |
1341 | return false; |
1342 | /* Give chance to dying process */ |
1343 | schedule_timeout(1); |
1344 | return true; |
1345 | } |
1346 | |
1347 | /* |
1348 | * Currently used to update mapped file statistics, but the routine can be |
1349 | * generalized to update other statistics as well. |
1350 | */ |
1351 | void mem_cgroup_update_file_mapped(struct page *page, int val) |
1352 | { |
1353 | struct mem_cgroup *mem; |
1354 | struct page_cgroup *pc; |
1355 | |
1356 | pc = lookup_page_cgroup(page); |
1357 | if (unlikely(!pc)) |
1358 | return; |
1359 | |
1360 | lock_page_cgroup(pc); |
1361 | mem = pc->mem_cgroup; |
1362 | if (!mem || !PageCgroupUsed(pc)) |
1363 | goto done; |
1364 | |
1365 | /* |
1366 | * Preemption is already disabled. We can use __this_cpu_xxx |
1367 | */ |
1368 | if (val > 0) { |
1369 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
1370 | SetPageCgroupFileMapped(pc); |
1371 | } else { |
1372 | __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
1373 | ClearPageCgroupFileMapped(pc); |
1374 | } |
1375 | |
1376 | done: |
1377 | unlock_page_cgroup(pc); |
1378 | } |
1379 | |
1380 | /* |
1381 | * size of first charge trial. "32" comes from vmscan.c's magic value. |
1382 | * TODO: maybe necessary to use big numbers in big irons. |
1383 | */ |
1384 | #define CHARGE_SIZE (32 * PAGE_SIZE) |
1385 | struct memcg_stock_pcp { |
1386 | struct mem_cgroup *cached; /* this never be root cgroup */ |
1387 | int charge; |
1388 | struct work_struct work; |
1389 | }; |
1390 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); |
1391 | static atomic_t memcg_drain_count; |
1392 | |
1393 | /* |
1394 | * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed |
1395 | * from local stock and true is returned. If the stock is 0 or charges from a |
1396 | * cgroup which is not current target, returns false. This stock will be |
1397 | * refilled. |
1398 | */ |
1399 | static bool consume_stock(struct mem_cgroup *mem) |
1400 | { |
1401 | struct memcg_stock_pcp *stock; |
1402 | bool ret = true; |
1403 | |
1404 | stock = &get_cpu_var(memcg_stock); |
1405 | if (mem == stock->cached && stock->charge) |
1406 | stock->charge -= PAGE_SIZE; |
1407 | else /* need to call res_counter_charge */ |
1408 | ret = false; |
1409 | put_cpu_var(memcg_stock); |
1410 | return ret; |
1411 | } |
1412 | |
1413 | /* |
1414 | * Returns stocks cached in percpu to res_counter and reset cached information. |
1415 | */ |
1416 | static void drain_stock(struct memcg_stock_pcp *stock) |
1417 | { |
1418 | struct mem_cgroup *old = stock->cached; |
1419 | |
1420 | if (stock->charge) { |
1421 | res_counter_uncharge(&old->res, stock->charge); |
1422 | if (do_swap_account) |
1423 | res_counter_uncharge(&old->memsw, stock->charge); |
1424 | } |
1425 | stock->cached = NULL; |
1426 | stock->charge = 0; |
1427 | } |
1428 | |
1429 | /* |
1430 | * This must be called under preempt disabled or must be called by |
1431 | * a thread which is pinned to local cpu. |
1432 | */ |
1433 | static void drain_local_stock(struct work_struct *dummy) |
1434 | { |
1435 | struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); |
1436 | drain_stock(stock); |
1437 | } |
1438 | |
1439 | /* |
1440 | * Cache charges(val) which is from res_counter, to local per_cpu area. |
1441 | * This will be consumed by consumt_stock() function, later. |
1442 | */ |
1443 | static void refill_stock(struct mem_cgroup *mem, int val) |
1444 | { |
1445 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); |
1446 | |
1447 | if (stock->cached != mem) { /* reset if necessary */ |
1448 | drain_stock(stock); |
1449 | stock->cached = mem; |
1450 | } |
1451 | stock->charge += val; |
1452 | put_cpu_var(memcg_stock); |
1453 | } |
1454 | |
1455 | /* |
1456 | * Tries to drain stocked charges in other cpus. This function is asynchronous |
1457 | * and just put a work per cpu for draining localy on each cpu. Caller can |
1458 | * expects some charges will be back to res_counter later but cannot wait for |
1459 | * it. |
1460 | */ |
1461 | static void drain_all_stock_async(void) |
1462 | { |
1463 | int cpu; |
1464 | /* This function is for scheduling "drain" in asynchronous way. |
1465 | * The result of "drain" is not directly handled by callers. Then, |
1466 | * if someone is calling drain, we don't have to call drain more. |
1467 | * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if |
1468 | * there is a race. We just do loose check here. |
1469 | */ |
1470 | if (atomic_read(&memcg_drain_count)) |
1471 | return; |
1472 | /* Notify other cpus that system-wide "drain" is running */ |
1473 | atomic_inc(&memcg_drain_count); |
1474 | get_online_cpus(); |
1475 | for_each_online_cpu(cpu) { |
1476 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); |
1477 | schedule_work_on(cpu, &stock->work); |
1478 | } |
1479 | put_online_cpus(); |
1480 | atomic_dec(&memcg_drain_count); |
1481 | /* We don't wait for flush_work */ |
1482 | } |
1483 | |
1484 | /* This is a synchronous drain interface. */ |
1485 | static void drain_all_stock_sync(void) |
1486 | { |
1487 | /* called when force_empty is called */ |
1488 | atomic_inc(&memcg_drain_count); |
1489 | schedule_on_each_cpu(drain_local_stock); |
1490 | atomic_dec(&memcg_drain_count); |
1491 | } |
1492 | |
1493 | static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb, |
1494 | unsigned long action, |
1495 | void *hcpu) |
1496 | { |
1497 | int cpu = (unsigned long)hcpu; |
1498 | struct memcg_stock_pcp *stock; |
1499 | |
1500 | if (action != CPU_DEAD) |
1501 | return NOTIFY_OK; |
1502 | stock = &per_cpu(memcg_stock, cpu); |
1503 | drain_stock(stock); |
1504 | return NOTIFY_OK; |
1505 | } |
1506 | |
1507 | /* |
1508 | * Unlike exported interface, "oom" parameter is added. if oom==true, |
1509 | * oom-killer can be invoked. |
1510 | */ |
1511 | static int __mem_cgroup_try_charge(struct mm_struct *mm, |
1512 | gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom) |
1513 | { |
1514 | struct mem_cgroup *mem, *mem_over_limit; |
1515 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; |
1516 | struct res_counter *fail_res; |
1517 | int csize = CHARGE_SIZE; |
1518 | |
1519 | /* |
1520 | * Unlike gloval-vm's OOM-kill, we're not in memory shortage |
1521 | * in system level. So, allow to go ahead dying process in addition to |
1522 | * MEMDIE process. |
1523 | */ |
1524 | if (unlikely(test_thread_flag(TIF_MEMDIE) |
1525 | || fatal_signal_pending(current))) |
1526 | goto bypass; |
1527 | |
1528 | /* |
1529 | * We always charge the cgroup the mm_struct belongs to. |
1530 | * The mm_struct's mem_cgroup changes on task migration if the |
1531 | * thread group leader migrates. It's possible that mm is not |
1532 | * set, if so charge the init_mm (happens for pagecache usage). |
1533 | */ |
1534 | mem = *memcg; |
1535 | if (likely(!mem)) { |
1536 | mem = try_get_mem_cgroup_from_mm(mm); |
1537 | *memcg = mem; |
1538 | } else { |
1539 | css_get(&mem->css); |
1540 | } |
1541 | if (unlikely(!mem)) |
1542 | return 0; |
1543 | |
1544 | VM_BUG_ON(css_is_removed(&mem->css)); |
1545 | if (mem_cgroup_is_root(mem)) |
1546 | goto done; |
1547 | |
1548 | while (1) { |
1549 | int ret = 0; |
1550 | unsigned long flags = 0; |
1551 | |
1552 | if (consume_stock(mem)) |
1553 | goto done; |
1554 | |
1555 | ret = res_counter_charge(&mem->res, csize, &fail_res); |
1556 | if (likely(!ret)) { |
1557 | if (!do_swap_account) |
1558 | break; |
1559 | ret = res_counter_charge(&mem->memsw, csize, &fail_res); |
1560 | if (likely(!ret)) |
1561 | break; |
1562 | /* mem+swap counter fails */ |
1563 | res_counter_uncharge(&mem->res, csize); |
1564 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; |
1565 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, |
1566 | memsw); |
1567 | } else |
1568 | /* mem counter fails */ |
1569 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, |
1570 | res); |
1571 | |
1572 | /* reduce request size and retry */ |
1573 | if (csize > PAGE_SIZE) { |
1574 | csize = PAGE_SIZE; |
1575 | continue; |
1576 | } |
1577 | if (!(gfp_mask & __GFP_WAIT)) |
1578 | goto nomem; |
1579 | |
1580 | ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL, |
1581 | gfp_mask, flags); |
1582 | if (ret) |
1583 | continue; |
1584 | |
1585 | /* |
1586 | * try_to_free_mem_cgroup_pages() might not give us a full |
1587 | * picture of reclaim. Some pages are reclaimed and might be |
1588 | * moved to swap cache or just unmapped from the cgroup. |
1589 | * Check the limit again to see if the reclaim reduced the |
1590 | * current usage of the cgroup before giving up |
1591 | * |
1592 | */ |
1593 | if (mem_cgroup_check_under_limit(mem_over_limit)) |
1594 | continue; |
1595 | |
1596 | /* try to avoid oom while someone is moving charge */ |
1597 | if (mc.moving_task && current != mc.moving_task) { |
1598 | struct mem_cgroup *from, *to; |
1599 | bool do_continue = false; |
1600 | /* |
1601 | * There is a small race that "from" or "to" can be |
1602 | * freed by rmdir, so we use css_tryget(). |
1603 | */ |
1604 | from = mc.from; |
1605 | to = mc.to; |
1606 | if (from && css_tryget(&from->css)) { |
1607 | if (mem_over_limit->use_hierarchy) |
1608 | do_continue = css_is_ancestor( |
1609 | &from->css, |
1610 | &mem_over_limit->css); |
1611 | else |
1612 | do_continue = (from == mem_over_limit); |
1613 | css_put(&from->css); |
1614 | } |
1615 | if (!do_continue && to && css_tryget(&to->css)) { |
1616 | if (mem_over_limit->use_hierarchy) |
1617 | do_continue = css_is_ancestor( |
1618 | &to->css, |
1619 | &mem_over_limit->css); |
1620 | else |
1621 | do_continue = (to == mem_over_limit); |
1622 | css_put(&to->css); |
1623 | } |
1624 | if (do_continue) { |
1625 | DEFINE_WAIT(wait); |
1626 | prepare_to_wait(&mc.waitq, &wait, |
1627 | TASK_INTERRUPTIBLE); |
1628 | /* moving charge context might have finished. */ |
1629 | if (mc.moving_task) |
1630 | schedule(); |
1631 | finish_wait(&mc.waitq, &wait); |
1632 | continue; |
1633 | } |
1634 | } |
1635 | |
1636 | if (!nr_retries--) { |
1637 | if (!oom) |
1638 | goto nomem; |
1639 | if (mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) { |
1640 | nr_retries = MEM_CGROUP_RECLAIM_RETRIES; |
1641 | continue; |
1642 | } |
1643 | /* When we reach here, current task is dying .*/ |
1644 | css_put(&mem->css); |
1645 | goto bypass; |
1646 | } |
1647 | } |
1648 | if (csize > PAGE_SIZE) |
1649 | refill_stock(mem, csize - PAGE_SIZE); |
1650 | done: |
1651 | return 0; |
1652 | nomem: |
1653 | css_put(&mem->css); |
1654 | return -ENOMEM; |
1655 | bypass: |
1656 | *memcg = NULL; |
1657 | return 0; |
1658 | } |
1659 | |
1660 | /* |
1661 | * Somemtimes we have to undo a charge we got by try_charge(). |
1662 | * This function is for that and do uncharge, put css's refcnt. |
1663 | * gotten by try_charge(). |
1664 | */ |
1665 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem, |
1666 | unsigned long count) |
1667 | { |
1668 | if (!mem_cgroup_is_root(mem)) { |
1669 | res_counter_uncharge(&mem->res, PAGE_SIZE * count); |
1670 | if (do_swap_account) |
1671 | res_counter_uncharge(&mem->memsw, PAGE_SIZE * count); |
1672 | VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags)); |
1673 | WARN_ON_ONCE(count > INT_MAX); |
1674 | __css_put(&mem->css, (int)count); |
1675 | } |
1676 | /* we don't need css_put for root */ |
1677 | } |
1678 | |
1679 | static void mem_cgroup_cancel_charge(struct mem_cgroup *mem) |
1680 | { |
1681 | __mem_cgroup_cancel_charge(mem, 1); |
1682 | } |
1683 | |
1684 | /* |
1685 | * A helper function to get mem_cgroup from ID. must be called under |
1686 | * rcu_read_lock(). The caller must check css_is_removed() or some if |
1687 | * it's concern. (dropping refcnt from swap can be called against removed |
1688 | * memcg.) |
1689 | */ |
1690 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) |
1691 | { |
1692 | struct cgroup_subsys_state *css; |
1693 | |
1694 | /* ID 0 is unused ID */ |
1695 | if (!id) |
1696 | return NULL; |
1697 | css = css_lookup(&mem_cgroup_subsys, id); |
1698 | if (!css) |
1699 | return NULL; |
1700 | return container_of(css, struct mem_cgroup, css); |
1701 | } |
1702 | |
1703 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) |
1704 | { |
1705 | struct mem_cgroup *mem = NULL; |
1706 | struct page_cgroup *pc; |
1707 | unsigned short id; |
1708 | swp_entry_t ent; |
1709 | |
1710 | VM_BUG_ON(!PageLocked(page)); |
1711 | |
1712 | pc = lookup_page_cgroup(page); |
1713 | lock_page_cgroup(pc); |
1714 | if (PageCgroupUsed(pc)) { |
1715 | mem = pc->mem_cgroup; |
1716 | if (mem && !css_tryget(&mem->css)) |
1717 | mem = NULL; |
1718 | } else if (PageSwapCache(page)) { |
1719 | ent.val = page_private(page); |
1720 | id = lookup_swap_cgroup(ent); |
1721 | rcu_read_lock(); |
1722 | mem = mem_cgroup_lookup(id); |
1723 | if (mem && !css_tryget(&mem->css)) |
1724 | mem = NULL; |
1725 | rcu_read_unlock(); |
1726 | } |
1727 | unlock_page_cgroup(pc); |
1728 | return mem; |
1729 | } |
1730 | |
1731 | /* |
1732 | * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be |
1733 | * USED state. If already USED, uncharge and return. |
1734 | */ |
1735 | |
1736 | static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, |
1737 | struct page_cgroup *pc, |
1738 | enum charge_type ctype) |
1739 | { |
1740 | /* try_charge() can return NULL to *memcg, taking care of it. */ |
1741 | if (!mem) |
1742 | return; |
1743 | |
1744 | lock_page_cgroup(pc); |
1745 | if (unlikely(PageCgroupUsed(pc))) { |
1746 | unlock_page_cgroup(pc); |
1747 | mem_cgroup_cancel_charge(mem); |
1748 | return; |
1749 | } |
1750 | |
1751 | pc->mem_cgroup = mem; |
1752 | /* |
1753 | * We access a page_cgroup asynchronously without lock_page_cgroup(). |
1754 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup |
1755 | * is accessed after testing USED bit. To make pc->mem_cgroup visible |
1756 | * before USED bit, we need memory barrier here. |
1757 | * See mem_cgroup_add_lru_list(), etc. |
1758 | */ |
1759 | smp_wmb(); |
1760 | switch (ctype) { |
1761 | case MEM_CGROUP_CHARGE_TYPE_CACHE: |
1762 | case MEM_CGROUP_CHARGE_TYPE_SHMEM: |
1763 | SetPageCgroupCache(pc); |
1764 | SetPageCgroupUsed(pc); |
1765 | break; |
1766 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: |
1767 | ClearPageCgroupCache(pc); |
1768 | SetPageCgroupUsed(pc); |
1769 | break; |
1770 | default: |
1771 | break; |
1772 | } |
1773 | |
1774 | mem_cgroup_charge_statistics(mem, pc, true); |
1775 | |
1776 | unlock_page_cgroup(pc); |
1777 | /* |
1778 | * "charge_statistics" updated event counter. Then, check it. |
1779 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. |
1780 | * if they exceeds softlimit. |
1781 | */ |
1782 | memcg_check_events(mem, pc->page); |
1783 | } |
1784 | |
1785 | /** |
1786 | * __mem_cgroup_move_account - move account of the page |
1787 | * @pc: page_cgroup of the page. |
1788 | * @from: mem_cgroup which the page is moved from. |
1789 | * @to: mem_cgroup which the page is moved to. @from != @to. |
1790 | * @uncharge: whether we should call uncharge and css_put against @from. |
1791 | * |
1792 | * The caller must confirm following. |
1793 | * - page is not on LRU (isolate_page() is useful.) |
1794 | * - the pc is locked, used, and ->mem_cgroup points to @from. |
1795 | * |
1796 | * This function doesn't do "charge" nor css_get to new cgroup. It should be |
1797 | * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is |
1798 | * true, this function does "uncharge" from old cgroup, but it doesn't if |
1799 | * @uncharge is false, so a caller should do "uncharge". |
1800 | */ |
1801 | |
1802 | static void __mem_cgroup_move_account(struct page_cgroup *pc, |
1803 | struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge) |
1804 | { |
1805 | VM_BUG_ON(from == to); |
1806 | VM_BUG_ON(PageLRU(pc->page)); |
1807 | VM_BUG_ON(!PageCgroupLocked(pc)); |
1808 | VM_BUG_ON(!PageCgroupUsed(pc)); |
1809 | VM_BUG_ON(pc->mem_cgroup != from); |
1810 | |
1811 | if (PageCgroupFileMapped(pc)) { |
1812 | /* Update mapped_file data for mem_cgroup */ |
1813 | preempt_disable(); |
1814 | __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
1815 | __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); |
1816 | preempt_enable(); |
1817 | } |
1818 | mem_cgroup_charge_statistics(from, pc, false); |
1819 | if (uncharge) |
1820 | /* This is not "cancel", but cancel_charge does all we need. */ |
1821 | mem_cgroup_cancel_charge(from); |
1822 | |
1823 | /* caller should have done css_get */ |
1824 | pc->mem_cgroup = to; |
1825 | mem_cgroup_charge_statistics(to, pc, true); |
1826 | /* |
1827 | * We charges against "to" which may not have any tasks. Then, "to" |
1828 | * can be under rmdir(). But in current implementation, caller of |
1829 | * this function is just force_empty() and move charge, so it's |
1830 | * garanteed that "to" is never removed. So, we don't check rmdir |
1831 | * status here. |
1832 | */ |
1833 | } |
1834 | |
1835 | /* |
1836 | * check whether the @pc is valid for moving account and call |
1837 | * __mem_cgroup_move_account() |
1838 | */ |
1839 | static int mem_cgroup_move_account(struct page_cgroup *pc, |
1840 | struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge) |
1841 | { |
1842 | int ret = -EINVAL; |
1843 | lock_page_cgroup(pc); |
1844 | if (PageCgroupUsed(pc) && pc->mem_cgroup == from) { |
1845 | __mem_cgroup_move_account(pc, from, to, uncharge); |
1846 | ret = 0; |
1847 | } |
1848 | unlock_page_cgroup(pc); |
1849 | /* |
1850 | * check events |
1851 | */ |
1852 | memcg_check_events(to, pc->page); |
1853 | memcg_check_events(from, pc->page); |
1854 | return ret; |
1855 | } |
1856 | |
1857 | /* |
1858 | * move charges to its parent. |
1859 | */ |
1860 | |
1861 | static int mem_cgroup_move_parent(struct page_cgroup *pc, |
1862 | struct mem_cgroup *child, |
1863 | gfp_t gfp_mask) |
1864 | { |
1865 | struct page *page = pc->page; |
1866 | struct cgroup *cg = child->css.cgroup; |
1867 | struct cgroup *pcg = cg->parent; |
1868 | struct mem_cgroup *parent; |
1869 | int ret; |
1870 | |
1871 | /* Is ROOT ? */ |
1872 | if (!pcg) |
1873 | return -EINVAL; |
1874 | |
1875 | ret = -EBUSY; |
1876 | if (!get_page_unless_zero(page)) |
1877 | goto out; |
1878 | if (isolate_lru_page(page)) |
1879 | goto put; |
1880 | |
1881 | parent = mem_cgroup_from_cont(pcg); |
1882 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false); |
1883 | if (ret || !parent) |
1884 | goto put_back; |
1885 | |
1886 | ret = mem_cgroup_move_account(pc, child, parent, true); |
1887 | if (ret) |
1888 | mem_cgroup_cancel_charge(parent); |
1889 | put_back: |
1890 | putback_lru_page(page); |
1891 | put: |
1892 | put_page(page); |
1893 | out: |
1894 | return ret; |
1895 | } |
1896 | |
1897 | /* |
1898 | * Charge the memory controller for page usage. |
1899 | * Return |
1900 | * 0 if the charge was successful |
1901 | * < 0 if the cgroup is over its limit |
1902 | */ |
1903 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, |
1904 | gfp_t gfp_mask, enum charge_type ctype, |
1905 | struct mem_cgroup *memcg) |
1906 | { |
1907 | struct mem_cgroup *mem; |
1908 | struct page_cgroup *pc; |
1909 | int ret; |
1910 | |
1911 | pc = lookup_page_cgroup(page); |
1912 | /* can happen at boot */ |
1913 | if (unlikely(!pc)) |
1914 | return 0; |
1915 | prefetchw(pc); |
1916 | |
1917 | mem = memcg; |
1918 | ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true); |
1919 | if (ret || !mem) |
1920 | return ret; |
1921 | |
1922 | __mem_cgroup_commit_charge(mem, pc, ctype); |
1923 | return 0; |
1924 | } |
1925 | |
1926 | int mem_cgroup_newpage_charge(struct page *page, |
1927 | struct mm_struct *mm, gfp_t gfp_mask) |
1928 | { |
1929 | if (mem_cgroup_disabled()) |
1930 | return 0; |
1931 | if (PageCompound(page)) |
1932 | return 0; |
1933 | /* |
1934 | * If already mapped, we don't have to account. |
1935 | * If page cache, page->mapping has address_space. |
1936 | * But page->mapping may have out-of-use anon_vma pointer, |
1937 | * detecit it by PageAnon() check. newly-mapped-anon's page->mapping |
1938 | * is NULL. |
1939 | */ |
1940 | if (page_mapped(page) || (page->mapping && !PageAnon(page))) |
1941 | return 0; |
1942 | if (unlikely(!mm)) |
1943 | mm = &init_mm; |
1944 | return mem_cgroup_charge_common(page, mm, gfp_mask, |
1945 | MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL); |
1946 | } |
1947 | |
1948 | static void |
1949 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, |
1950 | enum charge_type ctype); |
1951 | |
1952 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, |
1953 | gfp_t gfp_mask) |
1954 | { |
1955 | struct mem_cgroup *mem = NULL; |
1956 | int ret; |
1957 | |
1958 | if (mem_cgroup_disabled()) |
1959 | return 0; |
1960 | if (PageCompound(page)) |
1961 | return 0; |
1962 | /* |
1963 | * Corner case handling. This is called from add_to_page_cache() |
1964 | * in usual. But some FS (shmem) precharges this page before calling it |
1965 | * and call add_to_page_cache() with GFP_NOWAIT. |
1966 | * |
1967 | * For GFP_NOWAIT case, the page may be pre-charged before calling |
1968 | * add_to_page_cache(). (See shmem.c) check it here and avoid to call |
1969 | * charge twice. (It works but has to pay a bit larger cost.) |
1970 | * And when the page is SwapCache, it should take swap information |
1971 | * into account. This is under lock_page() now. |
1972 | */ |
1973 | if (!(gfp_mask & __GFP_WAIT)) { |
1974 | struct page_cgroup *pc; |
1975 | |
1976 | |
1977 | pc = lookup_page_cgroup(page); |
1978 | if (!pc) |
1979 | return 0; |
1980 | lock_page_cgroup(pc); |
1981 | if (PageCgroupUsed(pc)) { |
1982 | unlock_page_cgroup(pc); |
1983 | return 0; |
1984 | } |
1985 | unlock_page_cgroup(pc); |
1986 | } |
1987 | |
1988 | if (unlikely(!mm && !mem)) |
1989 | mm = &init_mm; |
1990 | |
1991 | if (page_is_file_cache(page)) |
1992 | return mem_cgroup_charge_common(page, mm, gfp_mask, |
1993 | MEM_CGROUP_CHARGE_TYPE_CACHE, NULL); |
1994 | |
1995 | /* shmem */ |
1996 | if (PageSwapCache(page)) { |
1997 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); |
1998 | if (!ret) |
1999 | __mem_cgroup_commit_charge_swapin(page, mem, |
2000 | MEM_CGROUP_CHARGE_TYPE_SHMEM); |
2001 | } else |
2002 | ret = mem_cgroup_charge_common(page, mm, gfp_mask, |
2003 | MEM_CGROUP_CHARGE_TYPE_SHMEM, mem); |
2004 | |
2005 | return ret; |
2006 | } |
2007 | |
2008 | /* |
2009 | * While swap-in, try_charge -> commit or cancel, the page is locked. |
2010 | * And when try_charge() successfully returns, one refcnt to memcg without |
2011 | * struct page_cgroup is acquired. This refcnt will be consumed by |
2012 | * "commit()" or removed by "cancel()" |
2013 | */ |
2014 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, |
2015 | struct page *page, |
2016 | gfp_t mask, struct mem_cgroup **ptr) |
2017 | { |
2018 | struct mem_cgroup *mem; |
2019 | int ret; |
2020 | |
2021 | if (mem_cgroup_disabled()) |
2022 | return 0; |
2023 | |
2024 | if (!do_swap_account) |
2025 | goto charge_cur_mm; |
2026 | /* |
2027 | * A racing thread's fault, or swapoff, may have already updated |
2028 | * the pte, and even removed page from swap cache: in those cases |
2029 | * do_swap_page()'s pte_same() test will fail; but there's also a |
2030 | * KSM case which does need to charge the page. |
2031 | */ |
2032 | if (!PageSwapCache(page)) |
2033 | goto charge_cur_mm; |
2034 | mem = try_get_mem_cgroup_from_page(page); |
2035 | if (!mem) |
2036 | goto charge_cur_mm; |
2037 | *ptr = mem; |
2038 | ret = __mem_cgroup_try_charge(NULL, mask, ptr, true); |
2039 | /* drop extra refcnt from tryget */ |
2040 | css_put(&mem->css); |
2041 | return ret; |
2042 | charge_cur_mm: |
2043 | if (unlikely(!mm)) |
2044 | mm = &init_mm; |
2045 | return __mem_cgroup_try_charge(mm, mask, ptr, true); |
2046 | } |
2047 | |
2048 | static void |
2049 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, |
2050 | enum charge_type ctype) |
2051 | { |
2052 | struct page_cgroup *pc; |
2053 | |
2054 | if (mem_cgroup_disabled()) |
2055 | return; |
2056 | if (!ptr) |
2057 | return; |
2058 | cgroup_exclude_rmdir(&ptr->css); |
2059 | pc = lookup_page_cgroup(page); |
2060 | mem_cgroup_lru_del_before_commit_swapcache(page); |
2061 | __mem_cgroup_commit_charge(ptr, pc, ctype); |
2062 | mem_cgroup_lru_add_after_commit_swapcache(page); |
2063 | /* |
2064 | * Now swap is on-memory. This means this page may be |
2065 | * counted both as mem and swap....double count. |
2066 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable |
2067 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() |
2068 | * may call delete_from_swap_cache() before reach here. |
2069 | */ |
2070 | if (do_swap_account && PageSwapCache(page)) { |
2071 | swp_entry_t ent = {.val = page_private(page)}; |
2072 | unsigned short id; |
2073 | struct mem_cgroup *memcg; |
2074 | |
2075 | id = swap_cgroup_record(ent, 0); |
2076 | rcu_read_lock(); |
2077 | memcg = mem_cgroup_lookup(id); |
2078 | if (memcg) { |
2079 | /* |
2080 | * This recorded memcg can be obsolete one. So, avoid |
2081 | * calling css_tryget |
2082 | */ |
2083 | if (!mem_cgroup_is_root(memcg)) |
2084 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); |
2085 | mem_cgroup_swap_statistics(memcg, false); |
2086 | mem_cgroup_put(memcg); |
2087 | } |
2088 | rcu_read_unlock(); |
2089 | } |
2090 | /* |
2091 | * At swapin, we may charge account against cgroup which has no tasks. |
2092 | * So, rmdir()->pre_destroy() can be called while we do this charge. |
2093 | * In that case, we need to call pre_destroy() again. check it here. |
2094 | */ |
2095 | cgroup_release_and_wakeup_rmdir(&ptr->css); |
2096 | } |
2097 | |
2098 | void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) |
2099 | { |
2100 | __mem_cgroup_commit_charge_swapin(page, ptr, |
2101 | MEM_CGROUP_CHARGE_TYPE_MAPPED); |
2102 | } |
2103 | |
2104 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) |
2105 | { |
2106 | if (mem_cgroup_disabled()) |
2107 | return; |
2108 | if (!mem) |
2109 | return; |
2110 | mem_cgroup_cancel_charge(mem); |
2111 | } |
2112 | |
2113 | static void |
2114 | __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype) |
2115 | { |
2116 | struct memcg_batch_info *batch = NULL; |
2117 | bool uncharge_memsw = true; |
2118 | /* If swapout, usage of swap doesn't decrease */ |
2119 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) |
2120 | uncharge_memsw = false; |
2121 | /* |
2122 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. |
2123 | * In those cases, all pages freed continously can be expected to be in |
2124 | * the same cgroup and we have chance to coalesce uncharges. |
2125 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) |
2126 | * because we want to do uncharge as soon as possible. |
2127 | */ |
2128 | if (!current->memcg_batch.do_batch || test_thread_flag(TIF_MEMDIE)) |
2129 | goto direct_uncharge; |
2130 | |
2131 | batch = ¤t->memcg_batch; |
2132 | /* |
2133 | * In usual, we do css_get() when we remember memcg pointer. |
2134 | * But in this case, we keep res->usage until end of a series of |
2135 | * uncharges. Then, it's ok to ignore memcg's refcnt. |
2136 | */ |
2137 | if (!batch->memcg) |
2138 | batch->memcg = mem; |
2139 | /* |
2140 | * In typical case, batch->memcg == mem. This means we can |
2141 | * merge a series of uncharges to an uncharge of res_counter. |
2142 | * If not, we uncharge res_counter ony by one. |
2143 | */ |
2144 | if (batch->memcg != mem) |
2145 | goto direct_uncharge; |
2146 | /* remember freed charge and uncharge it later */ |
2147 | batch->bytes += PAGE_SIZE; |
2148 | if (uncharge_memsw) |
2149 | batch->memsw_bytes += PAGE_SIZE; |
2150 | return; |
2151 | direct_uncharge: |
2152 | res_counter_uncharge(&mem->res, PAGE_SIZE); |
2153 | if (uncharge_memsw) |
2154 | res_counter_uncharge(&mem->memsw, PAGE_SIZE); |
2155 | return; |
2156 | } |
2157 | |
2158 | /* |
2159 | * uncharge if !page_mapped(page) |
2160 | */ |
2161 | static struct mem_cgroup * |
2162 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) |
2163 | { |
2164 | struct page_cgroup *pc; |
2165 | struct mem_cgroup *mem = NULL; |
2166 | struct mem_cgroup_per_zone *mz; |
2167 | |
2168 | if (mem_cgroup_disabled()) |
2169 | return NULL; |
2170 | |
2171 | if (PageSwapCache(page)) |
2172 | return NULL; |
2173 | |
2174 | /* |
2175 | * Check if our page_cgroup is valid |
2176 | */ |
2177 | pc = lookup_page_cgroup(page); |
2178 | if (unlikely(!pc || !PageCgroupUsed(pc))) |
2179 | return NULL; |
2180 | |
2181 | lock_page_cgroup(pc); |
2182 | |
2183 | mem = pc->mem_cgroup; |
2184 | |
2185 | if (!PageCgroupUsed(pc)) |
2186 | goto unlock_out; |
2187 | |
2188 | switch (ctype) { |
2189 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: |
2190 | case MEM_CGROUP_CHARGE_TYPE_DROP: |
2191 | if (page_mapped(page)) |
2192 | goto unlock_out; |
2193 | break; |
2194 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: |
2195 | if (!PageAnon(page)) { /* Shared memory */ |
2196 | if (page->mapping && !page_is_file_cache(page)) |
2197 | goto unlock_out; |
2198 | } else if (page_mapped(page)) /* Anon */ |
2199 | goto unlock_out; |
2200 | break; |
2201 | default: |
2202 | break; |
2203 | } |
2204 | |
2205 | if (!mem_cgroup_is_root(mem)) |
2206 | __do_uncharge(mem, ctype); |
2207 | if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) |
2208 | mem_cgroup_swap_statistics(mem, true); |
2209 | mem_cgroup_charge_statistics(mem, pc, false); |
2210 | |
2211 | ClearPageCgroupUsed(pc); |
2212 | /* |
2213 | * pc->mem_cgroup is not cleared here. It will be accessed when it's |
2214 | * freed from LRU. This is safe because uncharged page is expected not |
2215 | * to be reused (freed soon). Exception is SwapCache, it's handled by |
2216 | * special functions. |
2217 | */ |
2218 | |
2219 | mz = page_cgroup_zoneinfo(pc); |
2220 | unlock_page_cgroup(pc); |
2221 | |
2222 | memcg_check_events(mem, page); |
2223 | /* at swapout, this memcg will be accessed to record to swap */ |
2224 | if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT) |
2225 | css_put(&mem->css); |
2226 | |
2227 | return mem; |
2228 | |
2229 | unlock_out: |
2230 | unlock_page_cgroup(pc); |
2231 | return NULL; |
2232 | } |
2233 | |
2234 | void mem_cgroup_uncharge_page(struct page *page) |
2235 | { |
2236 | /* early check. */ |
2237 | if (page_mapped(page)) |
2238 | return; |
2239 | if (page->mapping && !PageAnon(page)) |
2240 | return; |
2241 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); |
2242 | } |
2243 | |
2244 | void mem_cgroup_uncharge_cache_page(struct page *page) |
2245 | { |
2246 | VM_BUG_ON(page_mapped(page)); |
2247 | VM_BUG_ON(page->mapping); |
2248 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); |
2249 | } |
2250 | |
2251 | /* |
2252 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. |
2253 | * In that cases, pages are freed continuously and we can expect pages |
2254 | * are in the same memcg. All these calls itself limits the number of |
2255 | * pages freed at once, then uncharge_start/end() is called properly. |
2256 | * This may be called prural(2) times in a context, |
2257 | */ |
2258 | |
2259 | void mem_cgroup_uncharge_start(void) |
2260 | { |
2261 | current->memcg_batch.do_batch++; |
2262 | /* We can do nest. */ |
2263 | if (current->memcg_batch.do_batch == 1) { |
2264 | current->memcg_batch.memcg = NULL; |
2265 | current->memcg_batch.bytes = 0; |
2266 | current->memcg_batch.memsw_bytes = 0; |
2267 | } |
2268 | } |
2269 | |
2270 | void mem_cgroup_uncharge_end(void) |
2271 | { |
2272 | struct memcg_batch_info *batch = ¤t->memcg_batch; |
2273 | |
2274 | if (!batch->do_batch) |
2275 | return; |
2276 | |
2277 | batch->do_batch--; |
2278 | if (batch->do_batch) /* If stacked, do nothing. */ |
2279 | return; |
2280 | |
2281 | if (!batch->memcg) |
2282 | return; |
2283 | /* |
2284 | * This "batch->memcg" is valid without any css_get/put etc... |
2285 | * bacause we hide charges behind us. |
2286 | */ |
2287 | if (batch->bytes) |
2288 | res_counter_uncharge(&batch->memcg->res, batch->bytes); |
2289 | if (batch->memsw_bytes) |
2290 | res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes); |
2291 | /* forget this pointer (for sanity check) */ |
2292 | batch->memcg = NULL; |
2293 | } |
2294 | |
2295 | #ifdef CONFIG_SWAP |
2296 | /* |
2297 | * called after __delete_from_swap_cache() and drop "page" account. |
2298 | * memcg information is recorded to swap_cgroup of "ent" |
2299 | */ |
2300 | void |
2301 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) |
2302 | { |
2303 | struct mem_cgroup *memcg; |
2304 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; |
2305 | |
2306 | if (!swapout) /* this was a swap cache but the swap is unused ! */ |
2307 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; |
2308 | |
2309 | memcg = __mem_cgroup_uncharge_common(page, ctype); |
2310 | |
2311 | /* record memcg information */ |
2312 | if (do_swap_account && swapout && memcg) { |
2313 | swap_cgroup_record(ent, css_id(&memcg->css)); |
2314 | mem_cgroup_get(memcg); |
2315 | } |
2316 | if (swapout && memcg) |
2317 | css_put(&memcg->css); |
2318 | } |
2319 | #endif |
2320 | |
2321 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
2322 | /* |
2323 | * called from swap_entry_free(). remove record in swap_cgroup and |
2324 | * uncharge "memsw" account. |
2325 | */ |
2326 | void mem_cgroup_uncharge_swap(swp_entry_t ent) |
2327 | { |
2328 | struct mem_cgroup *memcg; |
2329 | unsigned short id; |
2330 | |
2331 | if (!do_swap_account) |
2332 | return; |
2333 | |
2334 | id = swap_cgroup_record(ent, 0); |
2335 | rcu_read_lock(); |
2336 | memcg = mem_cgroup_lookup(id); |
2337 | if (memcg) { |
2338 | /* |
2339 | * We uncharge this because swap is freed. |
2340 | * This memcg can be obsolete one. We avoid calling css_tryget |
2341 | */ |
2342 | if (!mem_cgroup_is_root(memcg)) |
2343 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); |
2344 | mem_cgroup_swap_statistics(memcg, false); |
2345 | mem_cgroup_put(memcg); |
2346 | } |
2347 | rcu_read_unlock(); |
2348 | } |
2349 | |
2350 | /** |
2351 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. |
2352 | * @entry: swap entry to be moved |
2353 | * @from: mem_cgroup which the entry is moved from |
2354 | * @to: mem_cgroup which the entry is moved to |
2355 | * @need_fixup: whether we should fixup res_counters and refcounts. |
2356 | * |
2357 | * It succeeds only when the swap_cgroup's record for this entry is the same |
2358 | * as the mem_cgroup's id of @from. |
2359 | * |
2360 | * Returns 0 on success, -EINVAL on failure. |
2361 | * |
2362 | * The caller must have charged to @to, IOW, called res_counter_charge() about |
2363 | * both res and memsw, and called css_get(). |
2364 | */ |
2365 | static int mem_cgroup_move_swap_account(swp_entry_t entry, |
2366 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) |
2367 | { |
2368 | unsigned short old_id, new_id; |
2369 | |
2370 | old_id = css_id(&from->css); |
2371 | new_id = css_id(&to->css); |
2372 | |
2373 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { |
2374 | mem_cgroup_swap_statistics(from, false); |
2375 | mem_cgroup_swap_statistics(to, true); |
2376 | /* |
2377 | * This function is only called from task migration context now. |
2378 | * It postpones res_counter and refcount handling till the end |
2379 | * of task migration(mem_cgroup_clear_mc()) for performance |
2380 | * improvement. But we cannot postpone mem_cgroup_get(to) |
2381 | * because if the process that has been moved to @to does |
2382 | * swap-in, the refcount of @to might be decreased to 0. |
2383 | */ |
2384 | mem_cgroup_get(to); |
2385 | if (need_fixup) { |
2386 | if (!mem_cgroup_is_root(from)) |
2387 | res_counter_uncharge(&from->memsw, PAGE_SIZE); |
2388 | mem_cgroup_put(from); |
2389 | /* |
2390 | * we charged both to->res and to->memsw, so we should |
2391 | * uncharge to->res. |
2392 | */ |
2393 | if (!mem_cgroup_is_root(to)) |
2394 | res_counter_uncharge(&to->res, PAGE_SIZE); |
2395 | css_put(&to->css); |
2396 | } |
2397 | return 0; |
2398 | } |
2399 | return -EINVAL; |
2400 | } |
2401 | #else |
2402 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, |
2403 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) |
2404 | { |
2405 | return -EINVAL; |
2406 | } |
2407 | #endif |
2408 | |
2409 | /* |
2410 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old |
2411 | * page belongs to. |
2412 | */ |
2413 | int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr) |
2414 | { |
2415 | struct page_cgroup *pc; |
2416 | struct mem_cgroup *mem = NULL; |
2417 | int ret = 0; |
2418 | |
2419 | if (mem_cgroup_disabled()) |
2420 | return 0; |
2421 | |
2422 | pc = lookup_page_cgroup(page); |
2423 | lock_page_cgroup(pc); |
2424 | if (PageCgroupUsed(pc)) { |
2425 | mem = pc->mem_cgroup; |
2426 | css_get(&mem->css); |
2427 | } |
2428 | unlock_page_cgroup(pc); |
2429 | |
2430 | *ptr = mem; |
2431 | if (mem) { |
2432 | ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false); |
2433 | css_put(&mem->css); |
2434 | } |
2435 | return ret; |
2436 | } |
2437 | |
2438 | /* remove redundant charge if migration failed*/ |
2439 | void mem_cgroup_end_migration(struct mem_cgroup *mem, |
2440 | struct page *oldpage, struct page *newpage) |
2441 | { |
2442 | struct page *target, *unused; |
2443 | struct page_cgroup *pc; |
2444 | enum charge_type ctype; |
2445 | |
2446 | if (!mem) |
2447 | return; |
2448 | cgroup_exclude_rmdir(&mem->css); |
2449 | /* at migration success, oldpage->mapping is NULL. */ |
2450 | if (oldpage->mapping) { |
2451 | target = oldpage; |
2452 | unused = NULL; |
2453 | } else { |
2454 | target = newpage; |
2455 | unused = oldpage; |
2456 | } |
2457 | |
2458 | if (PageAnon(target)) |
2459 | ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; |
2460 | else if (page_is_file_cache(target)) |
2461 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; |
2462 | else |
2463 | ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; |
2464 | |
2465 | /* unused page is not on radix-tree now. */ |
2466 | if (unused) |
2467 | __mem_cgroup_uncharge_common(unused, ctype); |
2468 | |
2469 | pc = lookup_page_cgroup(target); |
2470 | /* |
2471 | * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup. |
2472 | * So, double-counting is effectively avoided. |
2473 | */ |
2474 | __mem_cgroup_commit_charge(mem, pc, ctype); |
2475 | |
2476 | /* |
2477 | * Both of oldpage and newpage are still under lock_page(). |
2478 | * Then, we don't have to care about race in radix-tree. |
2479 | * But we have to be careful that this page is unmapped or not. |
2480 | * |
2481 | * There is a case for !page_mapped(). At the start of |
2482 | * migration, oldpage was mapped. But now, it's zapped. |
2483 | * But we know *target* page is not freed/reused under us. |
2484 | * mem_cgroup_uncharge_page() does all necessary checks. |
2485 | */ |
2486 | if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) |
2487 | mem_cgroup_uncharge_page(target); |
2488 | /* |
2489 | * At migration, we may charge account against cgroup which has no tasks |
2490 | * So, rmdir()->pre_destroy() can be called while we do this charge. |
2491 | * In that case, we need to call pre_destroy() again. check it here. |
2492 | */ |
2493 | cgroup_release_and_wakeup_rmdir(&mem->css); |
2494 | } |
2495 | |
2496 | /* |
2497 | * A call to try to shrink memory usage on charge failure at shmem's swapin. |
2498 | * Calling hierarchical_reclaim is not enough because we should update |
2499 | * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM. |
2500 | * Moreover considering hierarchy, we should reclaim from the mem_over_limit, |
2501 | * not from the memcg which this page would be charged to. |
2502 | * try_charge_swapin does all of these works properly. |
2503 | */ |
2504 | int mem_cgroup_shmem_charge_fallback(struct page *page, |
2505 | struct mm_struct *mm, |
2506 | gfp_t gfp_mask) |
2507 | { |
2508 | struct mem_cgroup *mem = NULL; |
2509 | int ret; |
2510 | |
2511 | if (mem_cgroup_disabled()) |
2512 | return 0; |
2513 | |
2514 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); |
2515 | if (!ret) |
2516 | mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */ |
2517 | |
2518 | return ret; |
2519 | } |
2520 | |
2521 | static DEFINE_MUTEX(set_limit_mutex); |
2522 | |
2523 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, |
2524 | unsigned long long val) |
2525 | { |
2526 | int retry_count; |
2527 | u64 memswlimit; |
2528 | int ret = 0; |
2529 | int children = mem_cgroup_count_children(memcg); |
2530 | u64 curusage, oldusage; |
2531 | |
2532 | /* |
2533 | * For keeping hierarchical_reclaim simple, how long we should retry |
2534 | * is depends on callers. We set our retry-count to be function |
2535 | * of # of children which we should visit in this loop. |
2536 | */ |
2537 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; |
2538 | |
2539 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); |
2540 | |
2541 | while (retry_count) { |
2542 | if (signal_pending(current)) { |
2543 | ret = -EINTR; |
2544 | break; |
2545 | } |
2546 | /* |
2547 | * Rather than hide all in some function, I do this in |
2548 | * open coded manner. You see what this really does. |
2549 | * We have to guarantee mem->res.limit < mem->memsw.limit. |
2550 | */ |
2551 | mutex_lock(&set_limit_mutex); |
2552 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
2553 | if (memswlimit < val) { |
2554 | ret = -EINVAL; |
2555 | mutex_unlock(&set_limit_mutex); |
2556 | break; |
2557 | } |
2558 | ret = res_counter_set_limit(&memcg->res, val); |
2559 | if (!ret) { |
2560 | if (memswlimit == val) |
2561 | memcg->memsw_is_minimum = true; |
2562 | else |
2563 | memcg->memsw_is_minimum = false; |
2564 | } |
2565 | mutex_unlock(&set_limit_mutex); |
2566 | |
2567 | if (!ret) |
2568 | break; |
2569 | |
2570 | mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, |
2571 | MEM_CGROUP_RECLAIM_SHRINK); |
2572 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); |
2573 | /* Usage is reduced ? */ |
2574 | if (curusage >= oldusage) |
2575 | retry_count--; |
2576 | else |
2577 | oldusage = curusage; |
2578 | } |
2579 | |
2580 | return ret; |
2581 | } |
2582 | |
2583 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, |
2584 | unsigned long long val) |
2585 | { |
2586 | int retry_count; |
2587 | u64 memlimit, oldusage, curusage; |
2588 | int children = mem_cgroup_count_children(memcg); |
2589 | int ret = -EBUSY; |
2590 | |
2591 | /* see mem_cgroup_resize_res_limit */ |
2592 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; |
2593 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); |
2594 | while (retry_count) { |
2595 | if (signal_pending(current)) { |
2596 | ret = -EINTR; |
2597 | break; |
2598 | } |
2599 | /* |
2600 | * Rather than hide all in some function, I do this in |
2601 | * open coded manner. You see what this really does. |
2602 | * We have to guarantee mem->res.limit < mem->memsw.limit. |
2603 | */ |
2604 | mutex_lock(&set_limit_mutex); |
2605 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
2606 | if (memlimit > val) { |
2607 | ret = -EINVAL; |
2608 | mutex_unlock(&set_limit_mutex); |
2609 | break; |
2610 | } |
2611 | ret = res_counter_set_limit(&memcg->memsw, val); |
2612 | if (!ret) { |
2613 | if (memlimit == val) |
2614 | memcg->memsw_is_minimum = true; |
2615 | else |
2616 | memcg->memsw_is_minimum = false; |
2617 | } |
2618 | mutex_unlock(&set_limit_mutex); |
2619 | |
2620 | if (!ret) |
2621 | break; |
2622 | |
2623 | mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, |
2624 | MEM_CGROUP_RECLAIM_NOSWAP | |
2625 | MEM_CGROUP_RECLAIM_SHRINK); |
2626 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); |
2627 | /* Usage is reduced ? */ |
2628 | if (curusage >= oldusage) |
2629 | retry_count--; |
2630 | else |
2631 | oldusage = curusage; |
2632 | } |
2633 | return ret; |
2634 | } |
2635 | |
2636 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, |
2637 | gfp_t gfp_mask, int nid, |
2638 | int zid) |
2639 | { |
2640 | unsigned long nr_reclaimed = 0; |
2641 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; |
2642 | unsigned long reclaimed; |
2643 | int loop = 0; |
2644 | struct mem_cgroup_tree_per_zone *mctz; |
2645 | unsigned long long excess; |
2646 | |
2647 | if (order > 0) |
2648 | return 0; |
2649 | |
2650 | mctz = soft_limit_tree_node_zone(nid, zid); |
2651 | /* |
2652 | * This loop can run a while, specially if mem_cgroup's continuously |
2653 | * keep exceeding their soft limit and putting the system under |
2654 | * pressure |
2655 | */ |
2656 | do { |
2657 | if (next_mz) |
2658 | mz = next_mz; |
2659 | else |
2660 | mz = mem_cgroup_largest_soft_limit_node(mctz); |
2661 | if (!mz) |
2662 | break; |
2663 | |
2664 | reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone, |
2665 | gfp_mask, |
2666 | MEM_CGROUP_RECLAIM_SOFT); |
2667 | nr_reclaimed += reclaimed; |
2668 | spin_lock(&mctz->lock); |
2669 | |
2670 | /* |
2671 | * If we failed to reclaim anything from this memory cgroup |
2672 | * it is time to move on to the next cgroup |
2673 | */ |
2674 | next_mz = NULL; |
2675 | if (!reclaimed) { |
2676 | do { |
2677 | /* |
2678 | * Loop until we find yet another one. |
2679 | * |
2680 | * By the time we get the soft_limit lock |
2681 | * again, someone might have aded the |
2682 | * group back on the RB tree. Iterate to |
2683 | * make sure we get a different mem. |
2684 | * mem_cgroup_largest_soft_limit_node returns |
2685 | * NULL if no other cgroup is present on |
2686 | * the tree |
2687 | */ |
2688 | next_mz = |
2689 | __mem_cgroup_largest_soft_limit_node(mctz); |
2690 | if (next_mz == mz) { |
2691 | css_put(&next_mz->mem->css); |
2692 | next_mz = NULL; |
2693 | } else /* next_mz == NULL or other memcg */ |
2694 | break; |
2695 | } while (1); |
2696 | } |
2697 | __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); |
2698 | excess = res_counter_soft_limit_excess(&mz->mem->res); |
2699 | /* |
2700 | * One school of thought says that we should not add |
2701 | * back the node to the tree if reclaim returns 0. |
2702 | * But our reclaim could return 0, simply because due |
2703 | * to priority we are exposing a smaller subset of |
2704 | * memory to reclaim from. Consider this as a longer |
2705 | * term TODO. |
2706 | */ |
2707 | /* If excess == 0, no tree ops */ |
2708 | __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); |
2709 | spin_unlock(&mctz->lock); |
2710 | css_put(&mz->mem->css); |
2711 | loop++; |
2712 | /* |
2713 | * Could not reclaim anything and there are no more |
2714 | * mem cgroups to try or we seem to be looping without |
2715 | * reclaiming anything. |
2716 | */ |
2717 | if (!nr_reclaimed && |
2718 | (next_mz == NULL || |
2719 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) |
2720 | break; |
2721 | } while (!nr_reclaimed); |
2722 | if (next_mz) |
2723 | css_put(&next_mz->mem->css); |
2724 | return nr_reclaimed; |
2725 | } |
2726 | |
2727 | /* |
2728 | * This routine traverse page_cgroup in given list and drop them all. |
2729 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. |
2730 | */ |
2731 | static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, |
2732 | int node, int zid, enum lru_list lru) |
2733 | { |
2734 | struct zone *zone; |
2735 | struct mem_cgroup_per_zone *mz; |
2736 | struct page_cgroup *pc, *busy; |
2737 | unsigned long flags, loop; |
2738 | struct list_head *list; |
2739 | int ret = 0; |
2740 | |
2741 | zone = &NODE_DATA(node)->node_zones[zid]; |
2742 | mz = mem_cgroup_zoneinfo(mem, node, zid); |
2743 | list = &mz->lists[lru]; |
2744 | |
2745 | loop = MEM_CGROUP_ZSTAT(mz, lru); |
2746 | /* give some margin against EBUSY etc...*/ |
2747 | loop += 256; |
2748 | busy = NULL; |
2749 | while (loop--) { |
2750 | ret = 0; |
2751 | spin_lock_irqsave(&zone->lru_lock, flags); |
2752 | if (list_empty(list)) { |
2753 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
2754 | break; |
2755 | } |
2756 | pc = list_entry(list->prev, struct page_cgroup, lru); |
2757 | if (busy == pc) { |
2758 | list_move(&pc->lru, list); |
2759 | busy = NULL; |
2760 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
2761 | continue; |
2762 | } |
2763 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
2764 | |
2765 | ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL); |
2766 | if (ret == -ENOMEM) |
2767 | break; |
2768 | |
2769 | if (ret == -EBUSY || ret == -EINVAL) { |
2770 | /* found lock contention or "pc" is obsolete. */ |
2771 | busy = pc; |
2772 | cond_resched(); |
2773 | } else |
2774 | busy = NULL; |
2775 | } |
2776 | |
2777 | if (!ret && !list_empty(list)) |
2778 | return -EBUSY; |
2779 | return ret; |
2780 | } |
2781 | |
2782 | /* |
2783 | * make mem_cgroup's charge to be 0 if there is no task. |
2784 | * This enables deleting this mem_cgroup. |
2785 | */ |
2786 | static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) |
2787 | { |
2788 | int ret; |
2789 | int node, zid, shrink; |
2790 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; |
2791 | struct cgroup *cgrp = mem->css.cgroup; |
2792 | |
2793 | css_get(&mem->css); |
2794 | |
2795 | shrink = 0; |
2796 | /* should free all ? */ |
2797 | if (free_all) |
2798 | goto try_to_free; |
2799 | move_account: |
2800 | do { |
2801 | ret = -EBUSY; |
2802 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) |
2803 | goto out; |
2804 | ret = -EINTR; |
2805 | if (signal_pending(current)) |
2806 | goto out; |
2807 | /* This is for making all *used* pages to be on LRU. */ |
2808 | lru_add_drain_all(); |
2809 | drain_all_stock_sync(); |
2810 | ret = 0; |
2811 | for_each_node_state(node, N_HIGH_MEMORY) { |
2812 | for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { |
2813 | enum lru_list l; |
2814 | for_each_lru(l) { |
2815 | ret = mem_cgroup_force_empty_list(mem, |
2816 | node, zid, l); |
2817 | if (ret) |
2818 | break; |
2819 | } |
2820 | } |
2821 | if (ret) |
2822 | break; |
2823 | } |
2824 | /* it seems parent cgroup doesn't have enough mem */ |
2825 | if (ret == -ENOMEM) |
2826 | goto try_to_free; |
2827 | cond_resched(); |
2828 | /* "ret" should also be checked to ensure all lists are empty. */ |
2829 | } while (mem->res.usage > 0 || ret); |
2830 | out: |
2831 | css_put(&mem->css); |
2832 | return ret; |
2833 | |
2834 | try_to_free: |
2835 | /* returns EBUSY if there is a task or if we come here twice. */ |
2836 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { |
2837 | ret = -EBUSY; |
2838 | goto out; |
2839 | } |
2840 | /* we call try-to-free pages for make this cgroup empty */ |
2841 | lru_add_drain_all(); |
2842 | /* try to free all pages in this cgroup */ |
2843 | shrink = 1; |
2844 | while (nr_retries && mem->res.usage > 0) { |
2845 | int progress; |
2846 | |
2847 | if (signal_pending(current)) { |
2848 | ret = -EINTR; |
2849 | goto out; |
2850 | } |
2851 | progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, |
2852 | false, get_swappiness(mem)); |
2853 | if (!progress) { |
2854 | nr_retries--; |
2855 | /* maybe some writeback is necessary */ |
2856 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
2857 | } |
2858 | |
2859 | } |
2860 | lru_add_drain(); |
2861 | /* try move_account...there may be some *locked* pages. */ |
2862 | goto move_account; |
2863 | } |
2864 | |
2865 | int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) |
2866 | { |
2867 | return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); |
2868 | } |
2869 | |
2870 | |
2871 | static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) |
2872 | { |
2873 | return mem_cgroup_from_cont(cont)->use_hierarchy; |
2874 | } |
2875 | |
2876 | static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, |
2877 | u64 val) |
2878 | { |
2879 | int retval = 0; |
2880 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); |
2881 | struct cgroup *parent = cont->parent; |
2882 | struct mem_cgroup *parent_mem = NULL; |
2883 | |
2884 | if (parent) |
2885 | parent_mem = mem_cgroup_from_cont(parent); |
2886 | |
2887 | cgroup_lock(); |
2888 | /* |
2889 | * If parent's use_hierarchy is set, we can't make any modifications |
2890 | * in the child subtrees. If it is unset, then the change can |
2891 | * occur, provided the current cgroup has no children. |
2892 | * |
2893 | * For the root cgroup, parent_mem is NULL, we allow value to be |
2894 | * set if there are no children. |
2895 | */ |
2896 | if ((!parent_mem || !parent_mem->use_hierarchy) && |
2897 | (val == 1 || val == 0)) { |
2898 | if (list_empty(&cont->children)) |
2899 | mem->use_hierarchy = val; |
2900 | else |
2901 | retval = -EBUSY; |
2902 | } else |
2903 | retval = -EINVAL; |
2904 | cgroup_unlock(); |
2905 | |
2906 | return retval; |
2907 | } |
2908 | |
2909 | struct mem_cgroup_idx_data { |
2910 | s64 val; |
2911 | enum mem_cgroup_stat_index idx; |
2912 | }; |
2913 | |
2914 | static int |
2915 | mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data) |
2916 | { |
2917 | struct mem_cgroup_idx_data *d = data; |
2918 | d->val += mem_cgroup_read_stat(mem, d->idx); |
2919 | return 0; |
2920 | } |
2921 | |
2922 | static void |
2923 | mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem, |
2924 | enum mem_cgroup_stat_index idx, s64 *val) |
2925 | { |
2926 | struct mem_cgroup_idx_data d; |
2927 | d.idx = idx; |
2928 | d.val = 0; |
2929 | mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat); |
2930 | *val = d.val; |
2931 | } |
2932 | |
2933 | static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap) |
2934 | { |
2935 | u64 idx_val, val; |
2936 | |
2937 | if (!mem_cgroup_is_root(mem)) { |
2938 | if (!swap) |
2939 | return res_counter_read_u64(&mem->res, RES_USAGE); |
2940 | else |
2941 | return res_counter_read_u64(&mem->memsw, RES_USAGE); |
2942 | } |
2943 | |
2944 | mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val); |
2945 | val = idx_val; |
2946 | mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val); |
2947 | val += idx_val; |
2948 | |
2949 | if (swap) { |
2950 | mem_cgroup_get_recursive_idx_stat(mem, |
2951 | MEM_CGROUP_STAT_SWAPOUT, &idx_val); |
2952 | val += idx_val; |
2953 | } |
2954 | |
2955 | return val << PAGE_SHIFT; |
2956 | } |
2957 | |
2958 | static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) |
2959 | { |
2960 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); |
2961 | u64 val; |
2962 | int type, name; |
2963 | |
2964 | type = MEMFILE_TYPE(cft->private); |
2965 | name = MEMFILE_ATTR(cft->private); |
2966 | switch (type) { |
2967 | case _MEM: |
2968 | if (name == RES_USAGE) |
2969 | val = mem_cgroup_usage(mem, false); |
2970 | else |
2971 | val = res_counter_read_u64(&mem->res, name); |
2972 | break; |
2973 | case _MEMSWAP: |
2974 | if (name == RES_USAGE) |
2975 | val = mem_cgroup_usage(mem, true); |
2976 | else |
2977 | val = res_counter_read_u64(&mem->memsw, name); |
2978 | break; |
2979 | default: |
2980 | BUG(); |
2981 | break; |
2982 | } |
2983 | return val; |
2984 | } |
2985 | /* |
2986 | * The user of this function is... |
2987 | * RES_LIMIT. |
2988 | */ |
2989 | static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, |
2990 | const char *buffer) |
2991 | { |
2992 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); |
2993 | int type, name; |
2994 | unsigned long long val; |
2995 | int ret; |
2996 | |
2997 | type = MEMFILE_TYPE(cft->private); |
2998 | name = MEMFILE_ATTR(cft->private); |
2999 | switch (name) { |
3000 | case RES_LIMIT: |
3001 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ |
3002 | ret = -EINVAL; |
3003 | break; |
3004 | } |
3005 | /* This function does all necessary parse...reuse it */ |
3006 | ret = res_counter_memparse_write_strategy(buffer, &val); |
3007 | if (ret) |
3008 | break; |
3009 | if (type == _MEM) |
3010 | ret = mem_cgroup_resize_limit(memcg, val); |
3011 | else |
3012 | ret = mem_cgroup_resize_memsw_limit(memcg, val); |
3013 | break; |
3014 | case RES_SOFT_LIMIT: |
3015 | ret = res_counter_memparse_write_strategy(buffer, &val); |
3016 | if (ret) |
3017 | break; |
3018 | /* |
3019 | * For memsw, soft limits are hard to implement in terms |
3020 | * of semantics, for now, we support soft limits for |
3021 | * control without swap |
3022 | */ |
3023 | if (type == _MEM) |
3024 | ret = res_counter_set_soft_limit(&memcg->res, val); |
3025 | else |
3026 | ret = -EINVAL; |
3027 | break; |
3028 | default: |
3029 | ret = -EINVAL; /* should be BUG() ? */ |
3030 | break; |
3031 | } |
3032 | return ret; |
3033 | } |
3034 | |
3035 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, |
3036 | unsigned long long *mem_limit, unsigned long long *memsw_limit) |
3037 | { |
3038 | struct cgroup *cgroup; |
3039 | unsigned long long min_limit, min_memsw_limit, tmp; |
3040 | |
3041 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); |
3042 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
3043 | cgroup = memcg->css.cgroup; |
3044 | if (!memcg->use_hierarchy) |
3045 | goto out; |
3046 | |
3047 | while (cgroup->parent) { |
3048 | cgroup = cgroup->parent; |
3049 | memcg = mem_cgroup_from_cont(cgroup); |
3050 | if (!memcg->use_hierarchy) |
3051 | break; |
3052 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); |
3053 | min_limit = min(min_limit, tmp); |
3054 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); |
3055 | min_memsw_limit = min(min_memsw_limit, tmp); |
3056 | } |
3057 | out: |
3058 | *mem_limit = min_limit; |
3059 | *memsw_limit = min_memsw_limit; |
3060 | return; |
3061 | } |
3062 | |
3063 | static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) |
3064 | { |
3065 | struct mem_cgroup *mem; |
3066 | int type, name; |
3067 | |
3068 | mem = mem_cgroup_from_cont(cont); |
3069 | type = MEMFILE_TYPE(event); |
3070 | name = MEMFILE_ATTR(event); |
3071 | switch (name) { |
3072 | case RES_MAX_USAGE: |
3073 | if (type == _MEM) |
3074 | res_counter_reset_max(&mem->res); |
3075 | else |
3076 | res_counter_reset_max(&mem->memsw); |
3077 | break; |
3078 | case RES_FAILCNT: |
3079 | if (type == _MEM) |
3080 | res_counter_reset_failcnt(&mem->res); |
3081 | else |
3082 | res_counter_reset_failcnt(&mem->memsw); |
3083 | break; |
3084 | } |
3085 | |
3086 | return 0; |
3087 | } |
3088 | |
3089 | static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, |
3090 | struct cftype *cft) |
3091 | { |
3092 | return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; |
3093 | } |
3094 | |
3095 | #ifdef CONFIG_MMU |
3096 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, |
3097 | struct cftype *cft, u64 val) |
3098 | { |
3099 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); |
3100 | |
3101 | if (val >= (1 << NR_MOVE_TYPE)) |
3102 | return -EINVAL; |
3103 | /* |
3104 | * We check this value several times in both in can_attach() and |
3105 | * attach(), so we need cgroup lock to prevent this value from being |
3106 | * inconsistent. |
3107 | */ |
3108 | cgroup_lock(); |
3109 | mem->move_charge_at_immigrate = val; |
3110 | cgroup_unlock(); |
3111 | |
3112 | return 0; |
3113 | } |
3114 | #else |
3115 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, |
3116 | struct cftype *cft, u64 val) |
3117 | { |
3118 | return -ENOSYS; |
3119 | } |
3120 | #endif |
3121 | |
3122 | |
3123 | /* For read statistics */ |
3124 | enum { |
3125 | MCS_CACHE, |
3126 | MCS_RSS, |
3127 | MCS_FILE_MAPPED, |
3128 | MCS_PGPGIN, |
3129 | MCS_PGPGOUT, |
3130 | MCS_SWAP, |
3131 | MCS_INACTIVE_ANON, |
3132 | MCS_ACTIVE_ANON, |
3133 | MCS_INACTIVE_FILE, |
3134 | MCS_ACTIVE_FILE, |
3135 | MCS_UNEVICTABLE, |
3136 | NR_MCS_STAT, |
3137 | }; |
3138 | |
3139 | struct mcs_total_stat { |
3140 | s64 stat[NR_MCS_STAT]; |
3141 | }; |
3142 | |
3143 | struct { |
3144 | char *local_name; |
3145 | char *total_name; |
3146 | } memcg_stat_strings[NR_MCS_STAT] = { |
3147 | {"cache", "total_cache"}, |
3148 | {"rss", "total_rss"}, |
3149 | {"mapped_file", "total_mapped_file"}, |
3150 | {"pgpgin", "total_pgpgin"}, |
3151 | {"pgpgout", "total_pgpgout"}, |
3152 | {"swap", "total_swap"}, |
3153 | {"inactive_anon", "total_inactive_anon"}, |
3154 | {"active_anon", "total_active_anon"}, |
3155 | {"inactive_file", "total_inactive_file"}, |
3156 | {"active_file", "total_active_file"}, |
3157 | {"unevictable", "total_unevictable"} |
3158 | }; |
3159 | |
3160 | |
3161 | static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data) |
3162 | { |
3163 | struct mcs_total_stat *s = data; |
3164 | s64 val; |
3165 | |
3166 | /* per cpu stat */ |
3167 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); |
3168 | s->stat[MCS_CACHE] += val * PAGE_SIZE; |
3169 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); |
3170 | s->stat[MCS_RSS] += val * PAGE_SIZE; |
3171 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED); |
3172 | s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; |
3173 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT); |
3174 | s->stat[MCS_PGPGIN] += val; |
3175 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT); |
3176 | s->stat[MCS_PGPGOUT] += val; |
3177 | if (do_swap_account) { |
3178 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT); |
3179 | s->stat[MCS_SWAP] += val * PAGE_SIZE; |
3180 | } |
3181 | |
3182 | /* per zone stat */ |
3183 | val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); |
3184 | s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; |
3185 | val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); |
3186 | s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; |
3187 | val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); |
3188 | s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; |
3189 | val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); |
3190 | s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; |
3191 | val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); |
3192 | s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; |
3193 | return 0; |
3194 | } |
3195 | |
3196 | static void |
3197 | mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) |
3198 | { |
3199 | mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat); |
3200 | } |
3201 | |
3202 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, |
3203 | struct cgroup_map_cb *cb) |
3204 | { |
3205 | struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); |
3206 | struct mcs_total_stat mystat; |
3207 | int i; |
3208 | |
3209 | memset(&mystat, 0, sizeof(mystat)); |
3210 | mem_cgroup_get_local_stat(mem_cont, &mystat); |
3211 | |
3212 | for (i = 0; i < NR_MCS_STAT; i++) { |
3213 | if (i == MCS_SWAP && !do_swap_account) |
3214 | continue; |
3215 | cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); |
3216 | } |
3217 | |
3218 | /* Hierarchical information */ |
3219 | { |
3220 | unsigned long long limit, memsw_limit; |
3221 | memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); |
3222 | cb->fill(cb, "hierarchical_memory_limit", limit); |
3223 | if (do_swap_account) |
3224 | cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); |
3225 | } |
3226 | |
3227 | memset(&mystat, 0, sizeof(mystat)); |
3228 | mem_cgroup_get_total_stat(mem_cont, &mystat); |
3229 | for (i = 0; i < NR_MCS_STAT; i++) { |
3230 | if (i == MCS_SWAP && !do_swap_account) |
3231 | continue; |
3232 | cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); |
3233 | } |
3234 | |
3235 | #ifdef CONFIG_DEBUG_VM |
3236 | cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); |
3237 | |
3238 | { |
3239 | int nid, zid; |
3240 | struct mem_cgroup_per_zone *mz; |
3241 | unsigned long recent_rotated[2] = {0, 0}; |
3242 | unsigned long recent_scanned[2] = {0, 0}; |
3243 | |
3244 | for_each_online_node(nid) |
3245 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
3246 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); |
3247 | |
3248 | recent_rotated[0] += |
3249 | mz->reclaim_stat.recent_rotated[0]; |
3250 | recent_rotated[1] += |
3251 | mz->reclaim_stat.recent_rotated[1]; |
3252 | recent_scanned[0] += |
3253 | mz->reclaim_stat.recent_scanned[0]; |
3254 | recent_scanned[1] += |
3255 | mz->reclaim_stat.recent_scanned[1]; |
3256 | } |
3257 | cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); |
3258 | cb->fill(cb, "recent_rotated_file", recent_rotated[1]); |
3259 | cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); |
3260 | cb->fill(cb, "recent_scanned_file", recent_scanned[1]); |
3261 | } |
3262 | #endif |
3263 | |
3264 | return 0; |
3265 | } |
3266 | |
3267 | static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) |
3268 | { |
3269 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
3270 | |
3271 | return get_swappiness(memcg); |
3272 | } |
3273 | |
3274 | static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, |
3275 | u64 val) |
3276 | { |
3277 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
3278 | struct mem_cgroup *parent; |
3279 | |
3280 | if (val > 100) |
3281 | return -EINVAL; |
3282 | |
3283 | if (cgrp->parent == NULL) |
3284 | return -EINVAL; |
3285 | |
3286 | parent = mem_cgroup_from_cont(cgrp->parent); |
3287 | |
3288 | cgroup_lock(); |
3289 | |
3290 | /* If under hierarchy, only empty-root can set this value */ |
3291 | if ((parent->use_hierarchy) || |
3292 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { |
3293 | cgroup_unlock(); |
3294 | return -EINVAL; |
3295 | } |
3296 | |
3297 | spin_lock(&memcg->reclaim_param_lock); |
3298 | memcg->swappiness = val; |
3299 | spin_unlock(&memcg->reclaim_param_lock); |
3300 | |
3301 | cgroup_unlock(); |
3302 | |
3303 | return 0; |
3304 | } |
3305 | |
3306 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) |
3307 | { |
3308 | struct mem_cgroup_threshold_ary *t; |
3309 | u64 usage; |
3310 | int i; |
3311 | |
3312 | rcu_read_lock(); |
3313 | if (!swap) |
3314 | t = rcu_dereference(memcg->thresholds); |
3315 | else |
3316 | t = rcu_dereference(memcg->memsw_thresholds); |
3317 | |
3318 | if (!t) |
3319 | goto unlock; |
3320 | |
3321 | usage = mem_cgroup_usage(memcg, swap); |
3322 | |
3323 | /* |
3324 | * current_threshold points to threshold just below usage. |
3325 | * If it's not true, a threshold was crossed after last |
3326 | * call of __mem_cgroup_threshold(). |
3327 | */ |
3328 | i = atomic_read(&t->current_threshold); |
3329 | |
3330 | /* |
3331 | * Iterate backward over array of thresholds starting from |
3332 | * current_threshold and check if a threshold is crossed. |
3333 | * If none of thresholds below usage is crossed, we read |
3334 | * only one element of the array here. |
3335 | */ |
3336 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) |
3337 | eventfd_signal(t->entries[i].eventfd, 1); |
3338 | |
3339 | /* i = current_threshold + 1 */ |
3340 | i++; |
3341 | |
3342 | /* |
3343 | * Iterate forward over array of thresholds starting from |
3344 | * current_threshold+1 and check if a threshold is crossed. |
3345 | * If none of thresholds above usage is crossed, we read |
3346 | * only one element of the array here. |
3347 | */ |
3348 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) |
3349 | eventfd_signal(t->entries[i].eventfd, 1); |
3350 | |
3351 | /* Update current_threshold */ |
3352 | atomic_set(&t->current_threshold, i - 1); |
3353 | unlock: |
3354 | rcu_read_unlock(); |
3355 | } |
3356 | |
3357 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) |
3358 | { |
3359 | __mem_cgroup_threshold(memcg, false); |
3360 | if (do_swap_account) |
3361 | __mem_cgroup_threshold(memcg, true); |
3362 | } |
3363 | |
3364 | static int compare_thresholds(const void *a, const void *b) |
3365 | { |
3366 | const struct mem_cgroup_threshold *_a = a; |
3367 | const struct mem_cgroup_threshold *_b = b; |
3368 | |
3369 | return _a->threshold - _b->threshold; |
3370 | } |
3371 | |
3372 | static int mem_cgroup_register_event(struct cgroup *cgrp, struct cftype *cft, |
3373 | struct eventfd_ctx *eventfd, const char *args) |
3374 | { |
3375 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
3376 | struct mem_cgroup_threshold_ary *thresholds, *thresholds_new; |
3377 | int type = MEMFILE_TYPE(cft->private); |
3378 | u64 threshold, usage; |
3379 | int size; |
3380 | int i, ret; |
3381 | |
3382 | ret = res_counter_memparse_write_strategy(args, &threshold); |
3383 | if (ret) |
3384 | return ret; |
3385 | |
3386 | mutex_lock(&memcg->thresholds_lock); |
3387 | if (type == _MEM) |
3388 | thresholds = memcg->thresholds; |
3389 | else if (type == _MEMSWAP) |
3390 | thresholds = memcg->memsw_thresholds; |
3391 | else |
3392 | BUG(); |
3393 | |
3394 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); |
3395 | |
3396 | /* Check if a threshold crossed before adding a new one */ |
3397 | if (thresholds) |
3398 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); |
3399 | |
3400 | if (thresholds) |
3401 | size = thresholds->size + 1; |
3402 | else |
3403 | size = 1; |
3404 | |
3405 | /* Allocate memory for new array of thresholds */ |
3406 | thresholds_new = kmalloc(sizeof(*thresholds_new) + |
3407 | size * sizeof(struct mem_cgroup_threshold), |
3408 | GFP_KERNEL); |
3409 | if (!thresholds_new) { |
3410 | ret = -ENOMEM; |
3411 | goto unlock; |
3412 | } |
3413 | thresholds_new->size = size; |
3414 | |
3415 | /* Copy thresholds (if any) to new array */ |
3416 | if (thresholds) |
3417 | memcpy(thresholds_new->entries, thresholds->entries, |
3418 | thresholds->size * |
3419 | sizeof(struct mem_cgroup_threshold)); |
3420 | /* Add new threshold */ |
3421 | thresholds_new->entries[size - 1].eventfd = eventfd; |
3422 | thresholds_new->entries[size - 1].threshold = threshold; |
3423 | |
3424 | /* Sort thresholds. Registering of new threshold isn't time-critical */ |
3425 | sort(thresholds_new->entries, size, |
3426 | sizeof(struct mem_cgroup_threshold), |
3427 | compare_thresholds, NULL); |
3428 | |
3429 | /* Find current threshold */ |
3430 | atomic_set(&thresholds_new->current_threshold, -1); |
3431 | for (i = 0; i < size; i++) { |
3432 | if (thresholds_new->entries[i].threshold < usage) { |
3433 | /* |
3434 | * thresholds_new->current_threshold will not be used |
3435 | * until rcu_assign_pointer(), so it's safe to increment |
3436 | * it here. |
3437 | */ |
3438 | atomic_inc(&thresholds_new->current_threshold); |
3439 | } |
3440 | } |
3441 | |
3442 | if (type == _MEM) |
3443 | rcu_assign_pointer(memcg->thresholds, thresholds_new); |
3444 | else |
3445 | rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new); |
3446 | |
3447 | /* To be sure that nobody uses thresholds before freeing it */ |
3448 | synchronize_rcu(); |
3449 | |
3450 | kfree(thresholds); |
3451 | unlock: |
3452 | mutex_unlock(&memcg->thresholds_lock); |
3453 | |
3454 | return ret; |
3455 | } |
3456 | |
3457 | static int mem_cgroup_unregister_event(struct cgroup *cgrp, struct cftype *cft, |
3458 | struct eventfd_ctx *eventfd) |
3459 | { |
3460 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); |
3461 | struct mem_cgroup_threshold_ary *thresholds, *thresholds_new; |
3462 | int type = MEMFILE_TYPE(cft->private); |
3463 | u64 usage; |
3464 | int size = 0; |
3465 | int i, j, ret; |
3466 | |
3467 | mutex_lock(&memcg->thresholds_lock); |
3468 | if (type == _MEM) |
3469 | thresholds = memcg->thresholds; |
3470 | else if (type == _MEMSWAP) |
3471 | thresholds = memcg->memsw_thresholds; |
3472 | else |
3473 | BUG(); |
3474 | |
3475 | /* |
3476 | * Something went wrong if we trying to unregister a threshold |
3477 | * if we don't have thresholds |
3478 | */ |
3479 | BUG_ON(!thresholds); |
3480 | |
3481 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); |
3482 | |
3483 | /* Check if a threshold crossed before removing */ |
3484 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); |
3485 | |
3486 | /* Calculate new number of threshold */ |
3487 | for (i = 0; i < thresholds->size; i++) { |
3488 | if (thresholds->entries[i].eventfd != eventfd) |
3489 | size++; |
3490 | } |
3491 | |
3492 | /* Set thresholds array to NULL if we don't have thresholds */ |
3493 | if (!size) { |
3494 | thresholds_new = NULL; |
3495 | goto assign; |
3496 | } |
3497 | |
3498 | /* Allocate memory for new array of thresholds */ |
3499 | thresholds_new = kmalloc(sizeof(*thresholds_new) + |
3500 | size * sizeof(struct mem_cgroup_threshold), |
3501 | GFP_KERNEL); |
3502 | if (!thresholds_new) { |
3503 | ret = -ENOMEM; |
3504 | goto unlock; |
3505 | } |
3506 | thresholds_new->size = size; |
3507 | |
3508 | /* Copy thresholds and find current threshold */ |
3509 | atomic_set(&thresholds_new->current_threshold, -1); |
3510 | for (i = 0, j = 0; i < thresholds->size; i++) { |
3511 | if (thresholds->entries[i].eventfd == eventfd) |
3512 | continue; |
3513 | |
3514 | thresholds_new->entries[j] = thresholds->entries[i]; |
3515 | if (thresholds_new->entries[j].threshold < usage) { |
3516 | /* |
3517 | * thresholds_new->current_threshold will not be used |
3518 | * until rcu_assign_pointer(), so it's safe to increment |
3519 | * it here. |
3520 | */ |
3521 | atomic_inc(&thresholds_new->current_threshold); |
3522 | } |
3523 | j++; |
3524 | } |
3525 | |
3526 | assign: |
3527 | if (type == _MEM) |
3528 | rcu_assign_pointer(memcg->thresholds, thresholds_new); |
3529 | else |
3530 | rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new); |
3531 | |
3532 | /* To be sure that nobody uses thresholds before freeing it */ |
3533 | synchronize_rcu(); |
3534 | |
3535 | kfree(thresholds); |
3536 | unlock: |
3537 | mutex_unlock(&memcg->thresholds_lock); |
3538 | |
3539 | return ret; |
3540 | } |
3541 | |
3542 | static struct cftype mem_cgroup_files[] = { |
3543 | { |
3544 | .name = "usage_in_bytes", |
3545 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), |
3546 | .read_u64 = mem_cgroup_read, |
3547 | .register_event = mem_cgroup_register_event, |
3548 | .unregister_event = mem_cgroup_unregister_event, |
3549 | }, |
3550 | { |
3551 | .name = "max_usage_in_bytes", |
3552 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), |
3553 | .trigger = mem_cgroup_reset, |
3554 | .read_u64 = mem_cgroup_read, |
3555 | }, |
3556 | { |
3557 | .name = "limit_in_bytes", |
3558 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), |
3559 | .write_string = mem_cgroup_write, |
3560 | .read_u64 = mem_cgroup_read, |
3561 | }, |
3562 | { |
3563 | .name = "soft_limit_in_bytes", |
3564 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), |
3565 | .write_string = mem_cgroup_write, |
3566 | .read_u64 = mem_cgroup_read, |
3567 | }, |
3568 | { |
3569 | .name = "failcnt", |
3570 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), |
3571 | .trigger = mem_cgroup_reset, |
3572 | .read_u64 = mem_cgroup_read, |
3573 | }, |
3574 | { |
3575 | .name = "stat", |
3576 | .read_map = mem_control_stat_show, |
3577 | }, |
3578 | { |
3579 | .name = "force_empty", |
3580 | .trigger = mem_cgroup_force_empty_write, |
3581 | }, |
3582 | { |
3583 | .name = "use_hierarchy", |
3584 | .write_u64 = mem_cgroup_hierarchy_write, |
3585 | .read_u64 = mem_cgroup_hierarchy_read, |
3586 | }, |
3587 | { |
3588 | .name = "swappiness", |
3589 | .read_u64 = mem_cgroup_swappiness_read, |
3590 | .write_u64 = mem_cgroup_swappiness_write, |
3591 | }, |
3592 | { |
3593 | .name = "move_charge_at_immigrate", |
3594 | .read_u64 = mem_cgroup_move_charge_read, |
3595 | .write_u64 = mem_cgroup_move_charge_write, |
3596 | }, |
3597 | }; |
3598 | |
3599 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
3600 | static struct cftype memsw_cgroup_files[] = { |
3601 | { |
3602 | .name = "memsw.usage_in_bytes", |
3603 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), |
3604 | .read_u64 = mem_cgroup_read, |
3605 | .register_event = mem_cgroup_register_event, |
3606 | .unregister_event = mem_cgroup_unregister_event, |
3607 | }, |
3608 | { |
3609 | .name = "memsw.max_usage_in_bytes", |
3610 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), |
3611 | .trigger = mem_cgroup_reset, |
3612 | .read_u64 = mem_cgroup_read, |
3613 | }, |
3614 | { |
3615 | .name = "memsw.limit_in_bytes", |
3616 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), |
3617 | .write_string = mem_cgroup_write, |
3618 | .read_u64 = mem_cgroup_read, |
3619 | }, |
3620 | { |
3621 | .name = "memsw.failcnt", |
3622 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), |
3623 | .trigger = mem_cgroup_reset, |
3624 | .read_u64 = mem_cgroup_read, |
3625 | }, |
3626 | }; |
3627 | |
3628 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) |
3629 | { |
3630 | if (!do_swap_account) |
3631 | return 0; |
3632 | return cgroup_add_files(cont, ss, memsw_cgroup_files, |
3633 | ARRAY_SIZE(memsw_cgroup_files)); |
3634 | }; |
3635 | #else |
3636 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) |
3637 | { |
3638 | return 0; |
3639 | } |
3640 | #endif |
3641 | |
3642 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) |
3643 | { |
3644 | struct mem_cgroup_per_node *pn; |
3645 | struct mem_cgroup_per_zone *mz; |
3646 | enum lru_list l; |
3647 | int zone, tmp = node; |
3648 | /* |
3649 | * This routine is called against possible nodes. |
3650 | * But it's BUG to call kmalloc() against offline node. |
3651 | * |
3652 | * TODO: this routine can waste much memory for nodes which will |
3653 | * never be onlined. It's better to use memory hotplug callback |
3654 | * function. |
3655 | */ |
3656 | if (!node_state(node, N_NORMAL_MEMORY)) |
3657 | tmp = -1; |
3658 | pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp); |
3659 | if (!pn) |
3660 | return 1; |
3661 | |
3662 | mem->info.nodeinfo[node] = pn; |
3663 | memset(pn, 0, sizeof(*pn)); |
3664 | |
3665 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
3666 | mz = &pn->zoneinfo[zone]; |
3667 | for_each_lru(l) |
3668 | INIT_LIST_HEAD(&mz->lists[l]); |
3669 | mz->usage_in_excess = 0; |
3670 | mz->on_tree = false; |
3671 | mz->mem = mem; |
3672 | } |
3673 | return 0; |
3674 | } |
3675 | |
3676 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) |
3677 | { |
3678 | kfree(mem->info.nodeinfo[node]); |
3679 | } |
3680 | |
3681 | static struct mem_cgroup *mem_cgroup_alloc(void) |
3682 | { |
3683 | struct mem_cgroup *mem; |
3684 | int size = sizeof(struct mem_cgroup); |
3685 | |
3686 | /* Can be very big if MAX_NUMNODES is very big */ |
3687 | if (size < PAGE_SIZE) |
3688 | mem = kmalloc(size, GFP_KERNEL); |
3689 | else |
3690 | mem = vmalloc(size); |
3691 | |
3692 | if (!mem) |
3693 | return NULL; |
3694 | |
3695 | memset(mem, 0, size); |
3696 | mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); |
3697 | if (!mem->stat) { |
3698 | if (size < PAGE_SIZE) |
3699 | kfree(mem); |
3700 | else |
3701 | vfree(mem); |
3702 | mem = NULL; |
3703 | } |
3704 | return mem; |
3705 | } |
3706 | |
3707 | /* |
3708 | * At destroying mem_cgroup, references from swap_cgroup can remain. |
3709 | * (scanning all at force_empty is too costly...) |
3710 | * |
3711 | * Instead of clearing all references at force_empty, we remember |
3712 | * the number of reference from swap_cgroup and free mem_cgroup when |
3713 | * it goes down to 0. |
3714 | * |
3715 | * Removal of cgroup itself succeeds regardless of refs from swap. |
3716 | */ |
3717 | |
3718 | static void __mem_cgroup_free(struct mem_cgroup *mem) |
3719 | { |
3720 | int node; |
3721 | |
3722 | mem_cgroup_remove_from_trees(mem); |
3723 | free_css_id(&mem_cgroup_subsys, &mem->css); |
3724 | |
3725 | for_each_node_state(node, N_POSSIBLE) |
3726 | free_mem_cgroup_per_zone_info(mem, node); |
3727 | |
3728 | free_percpu(mem->stat); |
3729 | if (sizeof(struct mem_cgroup) < PAGE_SIZE) |
3730 | kfree(mem); |
3731 | else |
3732 | vfree(mem); |
3733 | } |
3734 | |
3735 | static void mem_cgroup_get(struct mem_cgroup *mem) |
3736 | { |
3737 | atomic_inc(&mem->refcnt); |
3738 | } |
3739 | |
3740 | static void __mem_cgroup_put(struct mem_cgroup *mem, int count) |
3741 | { |
3742 | if (atomic_sub_and_test(count, &mem->refcnt)) { |
3743 | struct mem_cgroup *parent = parent_mem_cgroup(mem); |
3744 | __mem_cgroup_free(mem); |
3745 | if (parent) |
3746 | mem_cgroup_put(parent); |
3747 | } |
3748 | } |
3749 | |
3750 | static void mem_cgroup_put(struct mem_cgroup *mem) |
3751 | { |
3752 | __mem_cgroup_put(mem, 1); |
3753 | } |
3754 | |
3755 | /* |
3756 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. |
3757 | */ |
3758 | static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) |
3759 | { |
3760 | if (!mem->res.parent) |
3761 | return NULL; |
3762 | return mem_cgroup_from_res_counter(mem->res.parent, res); |
3763 | } |
3764 | |
3765 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
3766 | static void __init enable_swap_cgroup(void) |
3767 | { |
3768 | if (!mem_cgroup_disabled() && really_do_swap_account) |
3769 | do_swap_account = 1; |
3770 | } |
3771 | #else |
3772 | static void __init enable_swap_cgroup(void) |
3773 | { |
3774 | } |
3775 | #endif |
3776 | |
3777 | static int mem_cgroup_soft_limit_tree_init(void) |
3778 | { |
3779 | struct mem_cgroup_tree_per_node *rtpn; |
3780 | struct mem_cgroup_tree_per_zone *rtpz; |
3781 | int tmp, node, zone; |
3782 | |
3783 | for_each_node_state(node, N_POSSIBLE) { |
3784 | tmp = node; |
3785 | if (!node_state(node, N_NORMAL_MEMORY)) |
3786 | tmp = -1; |
3787 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); |
3788 | if (!rtpn) |
3789 | return 1; |
3790 | |
3791 | soft_limit_tree.rb_tree_per_node[node] = rtpn; |
3792 | |
3793 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { |
3794 | rtpz = &rtpn->rb_tree_per_zone[zone]; |
3795 | rtpz->rb_root = RB_ROOT; |
3796 | spin_lock_init(&rtpz->lock); |
3797 | } |
3798 | } |
3799 | return 0; |
3800 | } |
3801 | |
3802 | static struct cgroup_subsys_state * __ref |
3803 | mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) |
3804 | { |
3805 | struct mem_cgroup *mem, *parent; |
3806 | long error = -ENOMEM; |
3807 | int node; |
3808 | |
3809 | mem = mem_cgroup_alloc(); |
3810 | if (!mem) |
3811 | return ERR_PTR(error); |
3812 | |
3813 | for_each_node_state(node, N_POSSIBLE) |
3814 | if (alloc_mem_cgroup_per_zone_info(mem, node)) |
3815 | goto free_out; |
3816 | |
3817 | /* root ? */ |
3818 | if (cont->parent == NULL) { |
3819 | int cpu; |
3820 | enable_swap_cgroup(); |
3821 | parent = NULL; |
3822 | root_mem_cgroup = mem; |
3823 | if (mem_cgroup_soft_limit_tree_init()) |
3824 | goto free_out; |
3825 | for_each_possible_cpu(cpu) { |
3826 | struct memcg_stock_pcp *stock = |
3827 | &per_cpu(memcg_stock, cpu); |
3828 | INIT_WORK(&stock->work, drain_local_stock); |
3829 | } |
3830 | hotcpu_notifier(memcg_stock_cpu_callback, 0); |
3831 | } else { |
3832 | parent = mem_cgroup_from_cont(cont->parent); |
3833 | mem->use_hierarchy = parent->use_hierarchy; |
3834 | } |
3835 | |
3836 | if (parent && parent->use_hierarchy) { |
3837 | res_counter_init(&mem->res, &parent->res); |
3838 | res_counter_init(&mem->memsw, &parent->memsw); |
3839 | /* |
3840 | * We increment refcnt of the parent to ensure that we can |
3841 | * safely access it on res_counter_charge/uncharge. |
3842 | * This refcnt will be decremented when freeing this |
3843 | * mem_cgroup(see mem_cgroup_put). |
3844 | */ |
3845 | mem_cgroup_get(parent); |
3846 | } else { |
3847 | res_counter_init(&mem->res, NULL); |
3848 | res_counter_init(&mem->memsw, NULL); |
3849 | } |
3850 | mem->last_scanned_child = 0; |
3851 | spin_lock_init(&mem->reclaim_param_lock); |
3852 | |
3853 | if (parent) |
3854 | mem->swappiness = get_swappiness(parent); |
3855 | atomic_set(&mem->refcnt, 1); |
3856 | mem->move_charge_at_immigrate = 0; |
3857 | mutex_init(&mem->thresholds_lock); |
3858 | return &mem->css; |
3859 | free_out: |
3860 | __mem_cgroup_free(mem); |
3861 | root_mem_cgroup = NULL; |
3862 | return ERR_PTR(error); |
3863 | } |
3864 | |
3865 | static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, |
3866 | struct cgroup *cont) |
3867 | { |
3868 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); |
3869 | |
3870 | return mem_cgroup_force_empty(mem, false); |
3871 | } |
3872 | |
3873 | static void mem_cgroup_destroy(struct cgroup_subsys *ss, |
3874 | struct cgroup *cont) |
3875 | { |
3876 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); |
3877 | |
3878 | mem_cgroup_put(mem); |
3879 | } |
3880 | |
3881 | static int mem_cgroup_populate(struct cgroup_subsys *ss, |
3882 | struct cgroup *cont) |
3883 | { |
3884 | int ret; |
3885 | |
3886 | ret = cgroup_add_files(cont, ss, mem_cgroup_files, |
3887 | ARRAY_SIZE(mem_cgroup_files)); |
3888 | |
3889 | if (!ret) |
3890 | ret = register_memsw_files(cont, ss); |
3891 | return ret; |
3892 | } |
3893 | |
3894 | #ifdef CONFIG_MMU |
3895 | /* Handlers for move charge at task migration. */ |
3896 | #define PRECHARGE_COUNT_AT_ONCE 256 |
3897 | static int mem_cgroup_do_precharge(unsigned long count) |
3898 | { |
3899 | int ret = 0; |
3900 | int batch_count = PRECHARGE_COUNT_AT_ONCE; |
3901 | struct mem_cgroup *mem = mc.to; |
3902 | |
3903 | if (mem_cgroup_is_root(mem)) { |
3904 | mc.precharge += count; |
3905 | /* we don't need css_get for root */ |
3906 | return ret; |
3907 | } |
3908 | /* try to charge at once */ |
3909 | if (count > 1) { |
3910 | struct res_counter *dummy; |
3911 | /* |
3912 | * "mem" cannot be under rmdir() because we've already checked |
3913 | * by cgroup_lock_live_cgroup() that it is not removed and we |
3914 | * are still under the same cgroup_mutex. So we can postpone |
3915 | * css_get(). |
3916 | */ |
3917 | if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy)) |
3918 | goto one_by_one; |
3919 | if (do_swap_account && res_counter_charge(&mem->memsw, |
3920 | PAGE_SIZE * count, &dummy)) { |
3921 | res_counter_uncharge(&mem->res, PAGE_SIZE * count); |
3922 | goto one_by_one; |
3923 | } |
3924 | mc.precharge += count; |
3925 | VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags)); |
3926 | WARN_ON_ONCE(count > INT_MAX); |
3927 | __css_get(&mem->css, (int)count); |
3928 | return ret; |
3929 | } |
3930 | one_by_one: |
3931 | /* fall back to one by one charge */ |
3932 | while (count--) { |
3933 | if (signal_pending(current)) { |
3934 | ret = -EINTR; |
3935 | break; |
3936 | } |
3937 | if (!batch_count--) { |
3938 | batch_count = PRECHARGE_COUNT_AT_ONCE; |
3939 | cond_resched(); |
3940 | } |
3941 | ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false); |
3942 | if (ret || !mem) |
3943 | /* mem_cgroup_clear_mc() will do uncharge later */ |
3944 | return -ENOMEM; |
3945 | mc.precharge++; |
3946 | } |
3947 | return ret; |
3948 | } |
3949 | |
3950 | /** |
3951 | * is_target_pte_for_mc - check a pte whether it is valid for move charge |
3952 | * @vma: the vma the pte to be checked belongs |
3953 | * @addr: the address corresponding to the pte to be checked |
3954 | * @ptent: the pte to be checked |
3955 | * @target: the pointer the target page or swap ent will be stored(can be NULL) |
3956 | * |
3957 | * Returns |
3958 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. |
3959 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for |
3960 | * move charge. if @target is not NULL, the page is stored in target->page |
3961 | * with extra refcnt got(Callers should handle it). |
3962 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a |
3963 | * target for charge migration. if @target is not NULL, the entry is stored |
3964 | * in target->ent. |
3965 | * |
3966 | * Called with pte lock held. |
3967 | */ |
3968 | union mc_target { |
3969 | struct page *page; |
3970 | swp_entry_t ent; |
3971 | }; |
3972 | |
3973 | enum mc_target_type { |
3974 | MC_TARGET_NONE, /* not used */ |
3975 | MC_TARGET_PAGE, |
3976 | MC_TARGET_SWAP, |
3977 | }; |
3978 | |
3979 | static int is_target_pte_for_mc(struct vm_area_struct *vma, |
3980 | unsigned long addr, pte_t ptent, union mc_target *target) |
3981 | { |
3982 | struct page *page = NULL; |
3983 | struct page_cgroup *pc; |
3984 | int ret = 0; |
3985 | swp_entry_t ent = { .val = 0 }; |
3986 | int usage_count = 0; |
3987 | bool move_anon = test_bit(MOVE_CHARGE_TYPE_ANON, |
3988 | &mc.to->move_charge_at_immigrate); |
3989 | |
3990 | if (!pte_present(ptent)) { |
3991 | /* TODO: handle swap of shmes/tmpfs */ |
3992 | if (pte_none(ptent) || pte_file(ptent)) |
3993 | return 0; |
3994 | else if (is_swap_pte(ptent)) { |
3995 | ent = pte_to_swp_entry(ptent); |
3996 | if (!move_anon || non_swap_entry(ent)) |
3997 | return 0; |
3998 | usage_count = mem_cgroup_count_swap_user(ent, &page); |
3999 | } |
4000 | } else { |
4001 | page = vm_normal_page(vma, addr, ptent); |
4002 | if (!page || !page_mapped(page)) |
4003 | return 0; |
4004 | /* |
4005 | * TODO: We don't move charges of file(including shmem/tmpfs) |
4006 | * pages for now. |
4007 | */ |
4008 | if (!move_anon || !PageAnon(page)) |
4009 | return 0; |
4010 | if (!get_page_unless_zero(page)) |
4011 | return 0; |
4012 | usage_count = page_mapcount(page); |
4013 | } |
4014 | if (usage_count > 1) { |
4015 | /* |
4016 | * TODO: We don't move charges of shared(used by multiple |
4017 | * processes) pages for now. |
4018 | */ |
4019 | if (page) |
4020 | put_page(page); |
4021 | return 0; |
4022 | } |
4023 | if (page) { |
4024 | pc = lookup_page_cgroup(page); |
4025 | /* |
4026 | * Do only loose check w/o page_cgroup lock. |
4027 | * mem_cgroup_move_account() checks the pc is valid or not under |
4028 | * the lock. |
4029 | */ |
4030 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { |
4031 | ret = MC_TARGET_PAGE; |
4032 | if (target) |
4033 | target->page = page; |
4034 | } |
4035 | if (!ret || !target) |
4036 | put_page(page); |
4037 | } |
4038 | /* throught */ |
4039 | if (ent.val && do_swap_account && !ret && |
4040 | css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { |
4041 | ret = MC_TARGET_SWAP; |
4042 | if (target) |
4043 | target->ent = ent; |
4044 | } |
4045 | return ret; |
4046 | } |
4047 | |
4048 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, |
4049 | unsigned long addr, unsigned long end, |
4050 | struct mm_walk *walk) |
4051 | { |
4052 | struct vm_area_struct *vma = walk->private; |
4053 | pte_t *pte; |
4054 | spinlock_t *ptl; |
4055 | |
4056 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
4057 | for (; addr != end; pte++, addr += PAGE_SIZE) |
4058 | if (is_target_pte_for_mc(vma, addr, *pte, NULL)) |
4059 | mc.precharge++; /* increment precharge temporarily */ |
4060 | pte_unmap_unlock(pte - 1, ptl); |
4061 | cond_resched(); |
4062 | |
4063 | return 0; |
4064 | } |
4065 | |
4066 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) |
4067 | { |
4068 | unsigned long precharge; |
4069 | struct vm_area_struct *vma; |
4070 | |
4071 | down_read(&mm->mmap_sem); |
4072 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
4073 | struct mm_walk mem_cgroup_count_precharge_walk = { |
4074 | .pmd_entry = mem_cgroup_count_precharge_pte_range, |
4075 | .mm = mm, |
4076 | .private = vma, |
4077 | }; |
4078 | if (is_vm_hugetlb_page(vma)) |
4079 | continue; |
4080 | /* TODO: We don't move charges of shmem/tmpfs pages for now. */ |
4081 | if (vma->vm_flags & VM_SHARED) |
4082 | continue; |
4083 | walk_page_range(vma->vm_start, vma->vm_end, |
4084 | &mem_cgroup_count_precharge_walk); |
4085 | } |
4086 | up_read(&mm->mmap_sem); |
4087 | |
4088 | precharge = mc.precharge; |
4089 | mc.precharge = 0; |
4090 | |
4091 | return precharge; |
4092 | } |
4093 | |
4094 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) |
4095 | { |
4096 | return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm)); |
4097 | } |
4098 | |
4099 | static void mem_cgroup_clear_mc(void) |
4100 | { |
4101 | /* we must uncharge all the leftover precharges from mc.to */ |
4102 | if (mc.precharge) { |
4103 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); |
4104 | mc.precharge = 0; |
4105 | } |
4106 | /* |
4107 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so |
4108 | * we must uncharge here. |
4109 | */ |
4110 | if (mc.moved_charge) { |
4111 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); |
4112 | mc.moved_charge = 0; |
4113 | } |
4114 | /* we must fixup refcnts and charges */ |
4115 | if (mc.moved_swap) { |
4116 | WARN_ON_ONCE(mc.moved_swap > INT_MAX); |
4117 | /* uncharge swap account from the old cgroup */ |
4118 | if (!mem_cgroup_is_root(mc.from)) |
4119 | res_counter_uncharge(&mc.from->memsw, |
4120 | PAGE_SIZE * mc.moved_swap); |
4121 | __mem_cgroup_put(mc.from, mc.moved_swap); |
4122 | |
4123 | if (!mem_cgroup_is_root(mc.to)) { |
4124 | /* |
4125 | * we charged both to->res and to->memsw, so we should |
4126 | * uncharge to->res. |
4127 | */ |
4128 | res_counter_uncharge(&mc.to->res, |
4129 | PAGE_SIZE * mc.moved_swap); |
4130 | VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags)); |
4131 | __css_put(&mc.to->css, mc.moved_swap); |
4132 | } |
4133 | /* we've already done mem_cgroup_get(mc.to) */ |
4134 | |
4135 | mc.moved_swap = 0; |
4136 | } |
4137 | mc.from = NULL; |
4138 | mc.to = NULL; |
4139 | mc.moving_task = NULL; |
4140 | wake_up_all(&mc.waitq); |
4141 | } |
4142 | |
4143 | static int mem_cgroup_can_attach(struct cgroup_subsys *ss, |
4144 | struct cgroup *cgroup, |
4145 | struct task_struct *p, |
4146 | bool threadgroup) |
4147 | { |
4148 | int ret = 0; |
4149 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup); |
4150 | |
4151 | if (mem->move_charge_at_immigrate) { |
4152 | struct mm_struct *mm; |
4153 | struct mem_cgroup *from = mem_cgroup_from_task(p); |
4154 | |
4155 | VM_BUG_ON(from == mem); |
4156 | |
4157 | mm = get_task_mm(p); |
4158 | if (!mm) |
4159 | return 0; |
4160 | /* We move charges only when we move a owner of the mm */ |
4161 | if (mm->owner == p) { |
4162 | VM_BUG_ON(mc.from); |
4163 | VM_BUG_ON(mc.to); |
4164 | VM_BUG_ON(mc.precharge); |
4165 | VM_BUG_ON(mc.moved_charge); |
4166 | VM_BUG_ON(mc.moved_swap); |
4167 | VM_BUG_ON(mc.moving_task); |
4168 | mc.from = from; |
4169 | mc.to = mem; |
4170 | mc.precharge = 0; |
4171 | mc.moved_charge = 0; |
4172 | mc.moved_swap = 0; |
4173 | mc.moving_task = current; |
4174 | |
4175 | ret = mem_cgroup_precharge_mc(mm); |
4176 | if (ret) |
4177 | mem_cgroup_clear_mc(); |
4178 | } |
4179 | mmput(mm); |
4180 | } |
4181 | return ret; |
4182 | } |
4183 | |
4184 | static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, |
4185 | struct cgroup *cgroup, |
4186 | struct task_struct *p, |
4187 | bool threadgroup) |
4188 | { |
4189 | mem_cgroup_clear_mc(); |
4190 | } |
4191 | |
4192 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, |
4193 | unsigned long addr, unsigned long end, |
4194 | struct mm_walk *walk) |
4195 | { |
4196 | int ret = 0; |
4197 | struct vm_area_struct *vma = walk->private; |
4198 | pte_t *pte; |
4199 | spinlock_t *ptl; |
4200 | |
4201 | retry: |
4202 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); |
4203 | for (; addr != end; addr += PAGE_SIZE) { |
4204 | pte_t ptent = *(pte++); |
4205 | union mc_target target; |
4206 | int type; |
4207 | struct page *page; |
4208 | struct page_cgroup *pc; |
4209 | swp_entry_t ent; |
4210 | |
4211 | if (!mc.precharge) |
4212 | break; |
4213 | |
4214 | type = is_target_pte_for_mc(vma, addr, ptent, &target); |
4215 | switch (type) { |
4216 | case MC_TARGET_PAGE: |
4217 | page = target.page; |
4218 | if (isolate_lru_page(page)) |
4219 | goto put; |
4220 | pc = lookup_page_cgroup(page); |
4221 | if (!mem_cgroup_move_account(pc, |
4222 | mc.from, mc.to, false)) { |
4223 | mc.precharge--; |
4224 | /* we uncharge from mc.from later. */ |
4225 | mc.moved_charge++; |
4226 | } |
4227 | putback_lru_page(page); |
4228 | put: /* is_target_pte_for_mc() gets the page */ |
4229 | put_page(page); |
4230 | break; |
4231 | case MC_TARGET_SWAP: |
4232 | ent = target.ent; |
4233 | if (!mem_cgroup_move_swap_account(ent, |
4234 | mc.from, mc.to, false)) { |
4235 | mc.precharge--; |
4236 | /* we fixup refcnts and charges later. */ |
4237 | mc.moved_swap++; |
4238 | } |
4239 | break; |
4240 | default: |
4241 | break; |
4242 | } |
4243 | } |
4244 | pte_unmap_unlock(pte - 1, ptl); |
4245 | cond_resched(); |
4246 | |
4247 | if (addr != end) { |
4248 | /* |
4249 | * We have consumed all precharges we got in can_attach(). |
4250 | * We try charge one by one, but don't do any additional |
4251 | * charges to mc.to if we have failed in charge once in attach() |
4252 | * phase. |
4253 | */ |
4254 | ret = mem_cgroup_do_precharge(1); |
4255 | if (!ret) |
4256 | goto retry; |
4257 | } |
4258 | |
4259 | return ret; |
4260 | } |
4261 | |
4262 | static void mem_cgroup_move_charge(struct mm_struct *mm) |
4263 | { |
4264 | struct vm_area_struct *vma; |
4265 | |
4266 | lru_add_drain_all(); |
4267 | down_read(&mm->mmap_sem); |
4268 | for (vma = mm->mmap; vma; vma = vma->vm_next) { |
4269 | int ret; |
4270 | struct mm_walk mem_cgroup_move_charge_walk = { |
4271 | .pmd_entry = mem_cgroup_move_charge_pte_range, |
4272 | .mm = mm, |
4273 | .private = vma, |
4274 | }; |
4275 | if (is_vm_hugetlb_page(vma)) |
4276 | continue; |
4277 | /* TODO: We don't move charges of shmem/tmpfs pages for now. */ |
4278 | if (vma->vm_flags & VM_SHARED) |
4279 | continue; |
4280 | ret = walk_page_range(vma->vm_start, vma->vm_end, |
4281 | &mem_cgroup_move_charge_walk); |
4282 | if (ret) |
4283 | /* |
4284 | * means we have consumed all precharges and failed in |
4285 | * doing additional charge. Just abandon here. |
4286 | */ |
4287 | break; |
4288 | } |
4289 | up_read(&mm->mmap_sem); |
4290 | } |
4291 | |
4292 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, |
4293 | struct cgroup *cont, |
4294 | struct cgroup *old_cont, |
4295 | struct task_struct *p, |
4296 | bool threadgroup) |
4297 | { |
4298 | struct mm_struct *mm; |
4299 | |
4300 | if (!mc.to) |
4301 | /* no need to move charge */ |
4302 | return; |
4303 | |
4304 | mm = get_task_mm(p); |
4305 | if (mm) { |
4306 | mem_cgroup_move_charge(mm); |
4307 | mmput(mm); |
4308 | } |
4309 | mem_cgroup_clear_mc(); |
4310 | } |
4311 | #else /* !CONFIG_MMU */ |
4312 | static int mem_cgroup_can_attach(struct cgroup_subsys *ss, |
4313 | struct cgroup *cgroup, |
4314 | struct task_struct *p, |
4315 | bool threadgroup) |
4316 | { |
4317 | return 0; |
4318 | } |
4319 | static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, |
4320 | struct cgroup *cgroup, |
4321 | struct task_struct *p, |
4322 | bool threadgroup) |
4323 | { |
4324 | } |
4325 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, |
4326 | struct cgroup *cont, |
4327 | struct cgroup *old_cont, |
4328 | struct task_struct *p, |
4329 | bool threadgroup) |
4330 | { |
4331 | } |
4332 | #endif |
4333 | |
4334 | struct cgroup_subsys mem_cgroup_subsys = { |
4335 | .name = "memory", |
4336 | .subsys_id = mem_cgroup_subsys_id, |
4337 | .create = mem_cgroup_create, |
4338 | .pre_destroy = mem_cgroup_pre_destroy, |
4339 | .destroy = mem_cgroup_destroy, |
4340 | .populate = mem_cgroup_populate, |
4341 | .can_attach = mem_cgroup_can_attach, |
4342 | .cancel_attach = mem_cgroup_cancel_attach, |
4343 | .attach = mem_cgroup_move_task, |
4344 | .early_init = 0, |
4345 | .use_id = 1, |
4346 | }; |
4347 | |
4348 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP |
4349 | |
4350 | static int __init disable_swap_account(char *s) |
4351 | { |
4352 | really_do_swap_account = 0; |
4353 | return 1; |
4354 | } |
4355 | __setup("noswapaccount", disable_swap_account); |
4356 | #endif |
4357 |
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