]>
Commit | Line | Data |
---|---|---|
1 | /* memcontrol.c - Memory Controller | |
2 | * | |
3 | * Copyright IBM Corporation, 2007 | |
4 | * Author Balbir Singh <balbir@linux.vnet.ibm.com> | |
5 | * | |
6 | * Copyright 2007 OpenVZ SWsoft Inc | |
7 | * Author: Pavel Emelianov <xemul@openvz.org> | |
8 | * | |
9 | * This program is free software; you can redistribute it and/or modify | |
10 | * it under the terms of the GNU General Public License as published by | |
11 | * the Free Software Foundation; either version 2 of the License, or | |
12 | * (at your option) any later version. | |
13 | * | |
14 | * This program is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
17 | * GNU General Public License for more details. | |
18 | */ | |
19 | ||
20 | #include <linux/res_counter.h> | |
21 | #include <linux/memcontrol.h> | |
22 | #include <linux/cgroup.h> | |
23 | #include <linux/mm.h> | |
24 | #include <linux/smp.h> | |
25 | #include <linux/page-flags.h> | |
26 | #include <linux/backing-dev.h> | |
27 | #include <linux/bit_spinlock.h> | |
28 | #include <linux/rcupdate.h> | |
29 | #include <linux/slab.h> | |
30 | #include <linux/swap.h> | |
31 | #include <linux/spinlock.h> | |
32 | #include <linux/fs.h> | |
33 | #include <linux/seq_file.h> | |
34 | ||
35 | #include <asm/uaccess.h> | |
36 | ||
37 | struct cgroup_subsys mem_cgroup_subsys; | |
38 | static const int MEM_CGROUP_RECLAIM_RETRIES = 5; | |
39 | static struct kmem_cache *page_cgroup_cache; | |
40 | ||
41 | /* | |
42 | * Statistics for memory cgroup. | |
43 | */ | |
44 | enum mem_cgroup_stat_index { | |
45 | /* | |
46 | * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. | |
47 | */ | |
48 | MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ | |
49 | MEM_CGROUP_STAT_RSS, /* # of pages charged as rss */ | |
50 | ||
51 | MEM_CGROUP_STAT_NSTATS, | |
52 | }; | |
53 | ||
54 | struct mem_cgroup_stat_cpu { | |
55 | s64 count[MEM_CGROUP_STAT_NSTATS]; | |
56 | } ____cacheline_aligned_in_smp; | |
57 | ||
58 | struct mem_cgroup_stat { | |
59 | struct mem_cgroup_stat_cpu cpustat[NR_CPUS]; | |
60 | }; | |
61 | ||
62 | /* | |
63 | * For accounting under irq disable, no need for increment preempt count. | |
64 | */ | |
65 | static void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat *stat, | |
66 | enum mem_cgroup_stat_index idx, int val) | |
67 | { | |
68 | int cpu = smp_processor_id(); | |
69 | stat->cpustat[cpu].count[idx] += val; | |
70 | } | |
71 | ||
72 | static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat, | |
73 | enum mem_cgroup_stat_index idx) | |
74 | { | |
75 | int cpu; | |
76 | s64 ret = 0; | |
77 | for_each_possible_cpu(cpu) | |
78 | ret += stat->cpustat[cpu].count[idx]; | |
79 | return ret; | |
80 | } | |
81 | ||
82 | /* | |
83 | * per-zone information in memory controller. | |
84 | */ | |
85 | ||
86 | enum mem_cgroup_zstat_index { | |
87 | MEM_CGROUP_ZSTAT_ACTIVE, | |
88 | MEM_CGROUP_ZSTAT_INACTIVE, | |
89 | ||
90 | NR_MEM_CGROUP_ZSTAT, | |
91 | }; | |
92 | ||
93 | struct mem_cgroup_per_zone { | |
94 | /* | |
95 | * spin_lock to protect the per cgroup LRU | |
96 | */ | |
97 | spinlock_t lru_lock; | |
98 | struct list_head active_list; | |
99 | struct list_head inactive_list; | |
100 | unsigned long count[NR_MEM_CGROUP_ZSTAT]; | |
101 | }; | |
102 | /* Macro for accessing counter */ | |
103 | #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) | |
104 | ||
105 | struct mem_cgroup_per_node { | |
106 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; | |
107 | }; | |
108 | ||
109 | struct mem_cgroup_lru_info { | |
110 | struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; | |
111 | }; | |
112 | ||
113 | /* | |
114 | * The memory controller data structure. The memory controller controls both | |
115 | * page cache and RSS per cgroup. We would eventually like to provide | |
116 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, | |
117 | * to help the administrator determine what knobs to tune. | |
118 | * | |
119 | * TODO: Add a water mark for the memory controller. Reclaim will begin when | |
120 | * we hit the water mark. May be even add a low water mark, such that | |
121 | * no reclaim occurs from a cgroup at it's low water mark, this is | |
122 | * a feature that will be implemented much later in the future. | |
123 | */ | |
124 | struct mem_cgroup { | |
125 | struct cgroup_subsys_state css; | |
126 | /* | |
127 | * the counter to account for memory usage | |
128 | */ | |
129 | struct res_counter res; | |
130 | /* | |
131 | * Per cgroup active and inactive list, similar to the | |
132 | * per zone LRU lists. | |
133 | */ | |
134 | struct mem_cgroup_lru_info info; | |
135 | ||
136 | int prev_priority; /* for recording reclaim priority */ | |
137 | /* | |
138 | * statistics. | |
139 | */ | |
140 | struct mem_cgroup_stat stat; | |
141 | }; | |
142 | static struct mem_cgroup init_mem_cgroup; | |
143 | ||
144 | /* | |
145 | * We use the lower bit of the page->page_cgroup pointer as a bit spin | |
146 | * lock. We need to ensure that page->page_cgroup is at least two | |
147 | * byte aligned (based on comments from Nick Piggin). But since | |
148 | * bit_spin_lock doesn't actually set that lock bit in a non-debug | |
149 | * uniprocessor kernel, we should avoid setting it here too. | |
150 | */ | |
151 | #define PAGE_CGROUP_LOCK_BIT 0x0 | |
152 | #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK) | |
153 | #define PAGE_CGROUP_LOCK (1 << PAGE_CGROUP_LOCK_BIT) | |
154 | #else | |
155 | #define PAGE_CGROUP_LOCK 0x0 | |
156 | #endif | |
157 | ||
158 | /* | |
159 | * A page_cgroup page is associated with every page descriptor. The | |
160 | * page_cgroup helps us identify information about the cgroup | |
161 | */ | |
162 | struct page_cgroup { | |
163 | struct list_head lru; /* per cgroup LRU list */ | |
164 | struct page *page; | |
165 | struct mem_cgroup *mem_cgroup; | |
166 | int ref_cnt; /* cached, mapped, migrating */ | |
167 | int flags; | |
168 | }; | |
169 | #define PAGE_CGROUP_FLAG_CACHE (0x1) /* charged as cache */ | |
170 | #define PAGE_CGROUP_FLAG_ACTIVE (0x2) /* page is active in this cgroup */ | |
171 | ||
172 | static int page_cgroup_nid(struct page_cgroup *pc) | |
173 | { | |
174 | return page_to_nid(pc->page); | |
175 | } | |
176 | ||
177 | static enum zone_type page_cgroup_zid(struct page_cgroup *pc) | |
178 | { | |
179 | return page_zonenum(pc->page); | |
180 | } | |
181 | ||
182 | enum charge_type { | |
183 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, | |
184 | MEM_CGROUP_CHARGE_TYPE_MAPPED, | |
185 | }; | |
186 | ||
187 | /* | |
188 | * Always modified under lru lock. Then, not necessary to preempt_disable() | |
189 | */ | |
190 | static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags, | |
191 | bool charge) | |
192 | { | |
193 | int val = (charge)? 1 : -1; | |
194 | struct mem_cgroup_stat *stat = &mem->stat; | |
195 | ||
196 | VM_BUG_ON(!irqs_disabled()); | |
197 | if (flags & PAGE_CGROUP_FLAG_CACHE) | |
198 | __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_CACHE, val); | |
199 | else | |
200 | __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val); | |
201 | } | |
202 | ||
203 | static struct mem_cgroup_per_zone * | |
204 | mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) | |
205 | { | |
206 | return &mem->info.nodeinfo[nid]->zoneinfo[zid]; | |
207 | } | |
208 | ||
209 | static struct mem_cgroup_per_zone * | |
210 | page_cgroup_zoneinfo(struct page_cgroup *pc) | |
211 | { | |
212 | struct mem_cgroup *mem = pc->mem_cgroup; | |
213 | int nid = page_cgroup_nid(pc); | |
214 | int zid = page_cgroup_zid(pc); | |
215 | ||
216 | return mem_cgroup_zoneinfo(mem, nid, zid); | |
217 | } | |
218 | ||
219 | static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem, | |
220 | enum mem_cgroup_zstat_index idx) | |
221 | { | |
222 | int nid, zid; | |
223 | struct mem_cgroup_per_zone *mz; | |
224 | u64 total = 0; | |
225 | ||
226 | for_each_online_node(nid) | |
227 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
228 | mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
229 | total += MEM_CGROUP_ZSTAT(mz, idx); | |
230 | } | |
231 | return total; | |
232 | } | |
233 | ||
234 | static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) | |
235 | { | |
236 | return container_of(cgroup_subsys_state(cont, | |
237 | mem_cgroup_subsys_id), struct mem_cgroup, | |
238 | css); | |
239 | } | |
240 | ||
241 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) | |
242 | { | |
243 | return container_of(task_subsys_state(p, mem_cgroup_subsys_id), | |
244 | struct mem_cgroup, css); | |
245 | } | |
246 | ||
247 | static inline int page_cgroup_locked(struct page *page) | |
248 | { | |
249 | return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup); | |
250 | } | |
251 | ||
252 | static void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc) | |
253 | { | |
254 | VM_BUG_ON(!page_cgroup_locked(page)); | |
255 | page->page_cgroup = ((unsigned long)pc | PAGE_CGROUP_LOCK); | |
256 | } | |
257 | ||
258 | struct page_cgroup *page_get_page_cgroup(struct page *page) | |
259 | { | |
260 | return (struct page_cgroup *) (page->page_cgroup & ~PAGE_CGROUP_LOCK); | |
261 | } | |
262 | ||
263 | static void lock_page_cgroup(struct page *page) | |
264 | { | |
265 | bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup); | |
266 | } | |
267 | ||
268 | static int try_lock_page_cgroup(struct page *page) | |
269 | { | |
270 | return bit_spin_trylock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup); | |
271 | } | |
272 | ||
273 | static void unlock_page_cgroup(struct page *page) | |
274 | { | |
275 | bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup); | |
276 | } | |
277 | ||
278 | static void __mem_cgroup_remove_list(struct page_cgroup *pc) | |
279 | { | |
280 | int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE; | |
281 | struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc); | |
282 | ||
283 | if (from) | |
284 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1; | |
285 | else | |
286 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1; | |
287 | ||
288 | mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, false); | |
289 | list_del_init(&pc->lru); | |
290 | } | |
291 | ||
292 | static void __mem_cgroup_add_list(struct page_cgroup *pc) | |
293 | { | |
294 | int to = pc->flags & PAGE_CGROUP_FLAG_ACTIVE; | |
295 | struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc); | |
296 | ||
297 | if (!to) { | |
298 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1; | |
299 | list_add(&pc->lru, &mz->inactive_list); | |
300 | } else { | |
301 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1; | |
302 | list_add(&pc->lru, &mz->active_list); | |
303 | } | |
304 | mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, true); | |
305 | } | |
306 | ||
307 | static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active) | |
308 | { | |
309 | int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE; | |
310 | struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc); | |
311 | ||
312 | if (from) | |
313 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1; | |
314 | else | |
315 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1; | |
316 | ||
317 | if (active) { | |
318 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1; | |
319 | pc->flags |= PAGE_CGROUP_FLAG_ACTIVE; | |
320 | list_move(&pc->lru, &mz->active_list); | |
321 | } else { | |
322 | MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1; | |
323 | pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE; | |
324 | list_move(&pc->lru, &mz->inactive_list); | |
325 | } | |
326 | } | |
327 | ||
328 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) | |
329 | { | |
330 | int ret; | |
331 | ||
332 | task_lock(task); | |
333 | ret = task->mm && mm_match_cgroup(task->mm, mem); | |
334 | task_unlock(task); | |
335 | return ret; | |
336 | } | |
337 | ||
338 | /* | |
339 | * This routine assumes that the appropriate zone's lru lock is already held | |
340 | */ | |
341 | void mem_cgroup_move_lists(struct page *page, bool active) | |
342 | { | |
343 | struct page_cgroup *pc; | |
344 | struct mem_cgroup_per_zone *mz; | |
345 | unsigned long flags; | |
346 | ||
347 | /* | |
348 | * We cannot lock_page_cgroup while holding zone's lru_lock, | |
349 | * because other holders of lock_page_cgroup can be interrupted | |
350 | * with an attempt to rotate_reclaimable_page. But we cannot | |
351 | * safely get to page_cgroup without it, so just try_lock it: | |
352 | * mem_cgroup_isolate_pages allows for page left on wrong list. | |
353 | */ | |
354 | if (!try_lock_page_cgroup(page)) | |
355 | return; | |
356 | ||
357 | pc = page_get_page_cgroup(page); | |
358 | if (pc) { | |
359 | mz = page_cgroup_zoneinfo(pc); | |
360 | spin_lock_irqsave(&mz->lru_lock, flags); | |
361 | __mem_cgroup_move_lists(pc, active); | |
362 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
363 | } | |
364 | unlock_page_cgroup(page); | |
365 | } | |
366 | ||
367 | /* | |
368 | * Calculate mapped_ratio under memory controller. This will be used in | |
369 | * vmscan.c for deteremining we have to reclaim mapped pages. | |
370 | */ | |
371 | int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem) | |
372 | { | |
373 | long total, rss; | |
374 | ||
375 | /* | |
376 | * usage is recorded in bytes. But, here, we assume the number of | |
377 | * physical pages can be represented by "long" on any arch. | |
378 | */ | |
379 | total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L; | |
380 | rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS); | |
381 | return (int)((rss * 100L) / total); | |
382 | } | |
383 | ||
384 | /* | |
385 | * This function is called from vmscan.c. In page reclaiming loop. balance | |
386 | * between active and inactive list is calculated. For memory controller | |
387 | * page reclaiming, we should use using mem_cgroup's imbalance rather than | |
388 | * zone's global lru imbalance. | |
389 | */ | |
390 | long mem_cgroup_reclaim_imbalance(struct mem_cgroup *mem) | |
391 | { | |
392 | unsigned long active, inactive; | |
393 | /* active and inactive are the number of pages. 'long' is ok.*/ | |
394 | active = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_ACTIVE); | |
395 | inactive = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_INACTIVE); | |
396 | return (long) (active / (inactive + 1)); | |
397 | } | |
398 | ||
399 | /* | |
400 | * prev_priority control...this will be used in memory reclaim path. | |
401 | */ | |
402 | int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem) | |
403 | { | |
404 | return mem->prev_priority; | |
405 | } | |
406 | ||
407 | void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority) | |
408 | { | |
409 | if (priority < mem->prev_priority) | |
410 | mem->prev_priority = priority; | |
411 | } | |
412 | ||
413 | void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority) | |
414 | { | |
415 | mem->prev_priority = priority; | |
416 | } | |
417 | ||
418 | /* | |
419 | * Calculate # of pages to be scanned in this priority/zone. | |
420 | * See also vmscan.c | |
421 | * | |
422 | * priority starts from "DEF_PRIORITY" and decremented in each loop. | |
423 | * (see include/linux/mmzone.h) | |
424 | */ | |
425 | ||
426 | long mem_cgroup_calc_reclaim_active(struct mem_cgroup *mem, | |
427 | struct zone *zone, int priority) | |
428 | { | |
429 | long nr_active; | |
430 | int nid = zone->zone_pgdat->node_id; | |
431 | int zid = zone_idx(zone); | |
432 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
433 | ||
434 | nr_active = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE); | |
435 | return (nr_active >> priority); | |
436 | } | |
437 | ||
438 | long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup *mem, | |
439 | struct zone *zone, int priority) | |
440 | { | |
441 | long nr_inactive; | |
442 | int nid = zone->zone_pgdat->node_id; | |
443 | int zid = zone_idx(zone); | |
444 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
445 | ||
446 | nr_inactive = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE); | |
447 | return (nr_inactive >> priority); | |
448 | } | |
449 | ||
450 | unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, | |
451 | struct list_head *dst, | |
452 | unsigned long *scanned, int order, | |
453 | int mode, struct zone *z, | |
454 | struct mem_cgroup *mem_cont, | |
455 | int active) | |
456 | { | |
457 | unsigned long nr_taken = 0; | |
458 | struct page *page; | |
459 | unsigned long scan; | |
460 | LIST_HEAD(pc_list); | |
461 | struct list_head *src; | |
462 | struct page_cgroup *pc, *tmp; | |
463 | int nid = z->zone_pgdat->node_id; | |
464 | int zid = zone_idx(z); | |
465 | struct mem_cgroup_per_zone *mz; | |
466 | ||
467 | BUG_ON(!mem_cont); | |
468 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); | |
469 | if (active) | |
470 | src = &mz->active_list; | |
471 | else | |
472 | src = &mz->inactive_list; | |
473 | ||
474 | ||
475 | spin_lock(&mz->lru_lock); | |
476 | scan = 0; | |
477 | list_for_each_entry_safe_reverse(pc, tmp, src, lru) { | |
478 | if (scan >= nr_to_scan) | |
479 | break; | |
480 | page = pc->page; | |
481 | ||
482 | if (unlikely(!PageLRU(page))) | |
483 | continue; | |
484 | ||
485 | if (PageActive(page) && !active) { | |
486 | __mem_cgroup_move_lists(pc, true); | |
487 | continue; | |
488 | } | |
489 | if (!PageActive(page) && active) { | |
490 | __mem_cgroup_move_lists(pc, false); | |
491 | continue; | |
492 | } | |
493 | ||
494 | scan++; | |
495 | list_move(&pc->lru, &pc_list); | |
496 | ||
497 | if (__isolate_lru_page(page, mode) == 0) { | |
498 | list_move(&page->lru, dst); | |
499 | nr_taken++; | |
500 | } | |
501 | } | |
502 | ||
503 | list_splice(&pc_list, src); | |
504 | spin_unlock(&mz->lru_lock); | |
505 | ||
506 | *scanned = scan; | |
507 | return nr_taken; | |
508 | } | |
509 | ||
510 | /* | |
511 | * Charge the memory controller for page usage. | |
512 | * Return | |
513 | * 0 if the charge was successful | |
514 | * < 0 if the cgroup is over its limit | |
515 | */ | |
516 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, | |
517 | gfp_t gfp_mask, enum charge_type ctype) | |
518 | { | |
519 | struct mem_cgroup *mem; | |
520 | struct page_cgroup *pc; | |
521 | unsigned long flags; | |
522 | unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
523 | struct mem_cgroup_per_zone *mz; | |
524 | ||
525 | if (mem_cgroup_subsys.disabled) | |
526 | return 0; | |
527 | ||
528 | /* | |
529 | * Should page_cgroup's go to their own slab? | |
530 | * One could optimize the performance of the charging routine | |
531 | * by saving a bit in the page_flags and using it as a lock | |
532 | * to see if the cgroup page already has a page_cgroup associated | |
533 | * with it | |
534 | */ | |
535 | retry: | |
536 | lock_page_cgroup(page); | |
537 | pc = page_get_page_cgroup(page); | |
538 | /* | |
539 | * The page_cgroup exists and | |
540 | * the page has already been accounted. | |
541 | */ | |
542 | if (pc) { | |
543 | VM_BUG_ON(pc->page != page); | |
544 | VM_BUG_ON(pc->ref_cnt <= 0); | |
545 | ||
546 | pc->ref_cnt++; | |
547 | unlock_page_cgroup(page); | |
548 | goto done; | |
549 | } | |
550 | unlock_page_cgroup(page); | |
551 | ||
552 | pc = kmem_cache_zalloc(page_cgroup_cache, gfp_mask); | |
553 | if (pc == NULL) | |
554 | goto err; | |
555 | ||
556 | /* | |
557 | * We always charge the cgroup the mm_struct belongs to. | |
558 | * The mm_struct's mem_cgroup changes on task migration if the | |
559 | * thread group leader migrates. It's possible that mm is not | |
560 | * set, if so charge the init_mm (happens for pagecache usage). | |
561 | */ | |
562 | if (!mm) | |
563 | mm = &init_mm; | |
564 | ||
565 | rcu_read_lock(); | |
566 | mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
567 | /* | |
568 | * For every charge from the cgroup, increment reference count | |
569 | */ | |
570 | css_get(&mem->css); | |
571 | rcu_read_unlock(); | |
572 | ||
573 | while (res_counter_charge(&mem->res, PAGE_SIZE)) { | |
574 | if (!(gfp_mask & __GFP_WAIT)) | |
575 | goto out; | |
576 | ||
577 | if (try_to_free_mem_cgroup_pages(mem, gfp_mask)) | |
578 | continue; | |
579 | ||
580 | /* | |
581 | * try_to_free_mem_cgroup_pages() might not give us a full | |
582 | * picture of reclaim. Some pages are reclaimed and might be | |
583 | * moved to swap cache or just unmapped from the cgroup. | |
584 | * Check the limit again to see if the reclaim reduced the | |
585 | * current usage of the cgroup before giving up | |
586 | */ | |
587 | if (res_counter_check_under_limit(&mem->res)) | |
588 | continue; | |
589 | ||
590 | if (!nr_retries--) { | |
591 | mem_cgroup_out_of_memory(mem, gfp_mask); | |
592 | goto out; | |
593 | } | |
594 | congestion_wait(WRITE, HZ/10); | |
595 | } | |
596 | ||
597 | pc->ref_cnt = 1; | |
598 | pc->mem_cgroup = mem; | |
599 | pc->page = page; | |
600 | pc->flags = PAGE_CGROUP_FLAG_ACTIVE; | |
601 | if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE) | |
602 | pc->flags |= PAGE_CGROUP_FLAG_CACHE; | |
603 | ||
604 | lock_page_cgroup(page); | |
605 | if (page_get_page_cgroup(page)) { | |
606 | unlock_page_cgroup(page); | |
607 | /* | |
608 | * Another charge has been added to this page already. | |
609 | * We take lock_page_cgroup(page) again and read | |
610 | * page->cgroup, increment refcnt.... just retry is OK. | |
611 | */ | |
612 | res_counter_uncharge(&mem->res, PAGE_SIZE); | |
613 | css_put(&mem->css); | |
614 | kmem_cache_free(page_cgroup_cache, pc); | |
615 | goto retry; | |
616 | } | |
617 | page_assign_page_cgroup(page, pc); | |
618 | ||
619 | mz = page_cgroup_zoneinfo(pc); | |
620 | spin_lock_irqsave(&mz->lru_lock, flags); | |
621 | __mem_cgroup_add_list(pc); | |
622 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
623 | ||
624 | unlock_page_cgroup(page); | |
625 | done: | |
626 | return 0; | |
627 | out: | |
628 | css_put(&mem->css); | |
629 | kmem_cache_free(page_cgroup_cache, pc); | |
630 | err: | |
631 | return -ENOMEM; | |
632 | } | |
633 | ||
634 | int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask) | |
635 | { | |
636 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
637 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
638 | } | |
639 | ||
640 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, | |
641 | gfp_t gfp_mask) | |
642 | { | |
643 | if (!mm) | |
644 | mm = &init_mm; | |
645 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
646 | MEM_CGROUP_CHARGE_TYPE_CACHE); | |
647 | } | |
648 | ||
649 | /* | |
650 | * Uncharging is always a welcome operation, we never complain, simply | |
651 | * uncharge. | |
652 | */ | |
653 | void mem_cgroup_uncharge_page(struct page *page) | |
654 | { | |
655 | struct page_cgroup *pc; | |
656 | struct mem_cgroup *mem; | |
657 | struct mem_cgroup_per_zone *mz; | |
658 | unsigned long flags; | |
659 | ||
660 | if (mem_cgroup_subsys.disabled) | |
661 | return; | |
662 | ||
663 | /* | |
664 | * Check if our page_cgroup is valid | |
665 | */ | |
666 | lock_page_cgroup(page); | |
667 | pc = page_get_page_cgroup(page); | |
668 | if (!pc) | |
669 | goto unlock; | |
670 | ||
671 | VM_BUG_ON(pc->page != page); | |
672 | VM_BUG_ON(pc->ref_cnt <= 0); | |
673 | ||
674 | if (--(pc->ref_cnt) == 0) { | |
675 | mz = page_cgroup_zoneinfo(pc); | |
676 | spin_lock_irqsave(&mz->lru_lock, flags); | |
677 | __mem_cgroup_remove_list(pc); | |
678 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
679 | ||
680 | page_assign_page_cgroup(page, NULL); | |
681 | unlock_page_cgroup(page); | |
682 | ||
683 | mem = pc->mem_cgroup; | |
684 | res_counter_uncharge(&mem->res, PAGE_SIZE); | |
685 | css_put(&mem->css); | |
686 | ||
687 | kmem_cache_free(page_cgroup_cache, pc); | |
688 | return; | |
689 | } | |
690 | ||
691 | unlock: | |
692 | unlock_page_cgroup(page); | |
693 | } | |
694 | ||
695 | /* | |
696 | * Returns non-zero if a page (under migration) has valid page_cgroup member. | |
697 | * Refcnt of page_cgroup is incremented. | |
698 | */ | |
699 | int mem_cgroup_prepare_migration(struct page *page) | |
700 | { | |
701 | struct page_cgroup *pc; | |
702 | ||
703 | if (mem_cgroup_subsys.disabled) | |
704 | return 0; | |
705 | ||
706 | lock_page_cgroup(page); | |
707 | pc = page_get_page_cgroup(page); | |
708 | if (pc) | |
709 | pc->ref_cnt++; | |
710 | unlock_page_cgroup(page); | |
711 | return pc != NULL; | |
712 | } | |
713 | ||
714 | void mem_cgroup_end_migration(struct page *page) | |
715 | { | |
716 | mem_cgroup_uncharge_page(page); | |
717 | } | |
718 | ||
719 | /* | |
720 | * We know both *page* and *newpage* are now not-on-LRU and PG_locked. | |
721 | * And no race with uncharge() routines because page_cgroup for *page* | |
722 | * has extra one reference by mem_cgroup_prepare_migration. | |
723 | */ | |
724 | void mem_cgroup_page_migration(struct page *page, struct page *newpage) | |
725 | { | |
726 | struct page_cgroup *pc; | |
727 | struct mem_cgroup_per_zone *mz; | |
728 | unsigned long flags; | |
729 | ||
730 | lock_page_cgroup(page); | |
731 | pc = page_get_page_cgroup(page); | |
732 | if (!pc) { | |
733 | unlock_page_cgroup(page); | |
734 | return; | |
735 | } | |
736 | ||
737 | mz = page_cgroup_zoneinfo(pc); | |
738 | spin_lock_irqsave(&mz->lru_lock, flags); | |
739 | __mem_cgroup_remove_list(pc); | |
740 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
741 | ||
742 | page_assign_page_cgroup(page, NULL); | |
743 | unlock_page_cgroup(page); | |
744 | ||
745 | pc->page = newpage; | |
746 | lock_page_cgroup(newpage); | |
747 | page_assign_page_cgroup(newpage, pc); | |
748 | ||
749 | mz = page_cgroup_zoneinfo(pc); | |
750 | spin_lock_irqsave(&mz->lru_lock, flags); | |
751 | __mem_cgroup_add_list(pc); | |
752 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
753 | ||
754 | unlock_page_cgroup(newpage); | |
755 | } | |
756 | ||
757 | /* | |
758 | * This routine traverse page_cgroup in given list and drop them all. | |
759 | * This routine ignores page_cgroup->ref_cnt. | |
760 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. | |
761 | */ | |
762 | #define FORCE_UNCHARGE_BATCH (128) | |
763 | static void mem_cgroup_force_empty_list(struct mem_cgroup *mem, | |
764 | struct mem_cgroup_per_zone *mz, | |
765 | int active) | |
766 | { | |
767 | struct page_cgroup *pc; | |
768 | struct page *page; | |
769 | int count = FORCE_UNCHARGE_BATCH; | |
770 | unsigned long flags; | |
771 | struct list_head *list; | |
772 | ||
773 | if (active) | |
774 | list = &mz->active_list; | |
775 | else | |
776 | list = &mz->inactive_list; | |
777 | ||
778 | spin_lock_irqsave(&mz->lru_lock, flags); | |
779 | while (!list_empty(list)) { | |
780 | pc = list_entry(list->prev, struct page_cgroup, lru); | |
781 | page = pc->page; | |
782 | get_page(page); | |
783 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
784 | mem_cgroup_uncharge_page(page); | |
785 | put_page(page); | |
786 | if (--count <= 0) { | |
787 | count = FORCE_UNCHARGE_BATCH; | |
788 | cond_resched(); | |
789 | } | |
790 | spin_lock_irqsave(&mz->lru_lock, flags); | |
791 | } | |
792 | spin_unlock_irqrestore(&mz->lru_lock, flags); | |
793 | } | |
794 | ||
795 | /* | |
796 | * make mem_cgroup's charge to be 0 if there is no task. | |
797 | * This enables deleting this mem_cgroup. | |
798 | */ | |
799 | static int mem_cgroup_force_empty(struct mem_cgroup *mem) | |
800 | { | |
801 | int ret = -EBUSY; | |
802 | int node, zid; | |
803 | ||
804 | if (mem_cgroup_subsys.disabled) | |
805 | return 0; | |
806 | ||
807 | css_get(&mem->css); | |
808 | /* | |
809 | * page reclaim code (kswapd etc..) will move pages between | |
810 | * active_list <-> inactive_list while we don't take a lock. | |
811 | * So, we have to do loop here until all lists are empty. | |
812 | */ | |
813 | while (mem->res.usage > 0) { | |
814 | if (atomic_read(&mem->css.cgroup->count) > 0) | |
815 | goto out; | |
816 | for_each_node_state(node, N_POSSIBLE) | |
817 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
818 | struct mem_cgroup_per_zone *mz; | |
819 | mz = mem_cgroup_zoneinfo(mem, node, zid); | |
820 | /* drop all page_cgroup in active_list */ | |
821 | mem_cgroup_force_empty_list(mem, mz, 1); | |
822 | /* drop all page_cgroup in inactive_list */ | |
823 | mem_cgroup_force_empty_list(mem, mz, 0); | |
824 | } | |
825 | } | |
826 | ret = 0; | |
827 | out: | |
828 | css_put(&mem->css); | |
829 | return ret; | |
830 | } | |
831 | ||
832 | static int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp) | |
833 | { | |
834 | *tmp = memparse(buf, &buf); | |
835 | if (*buf != '\0') | |
836 | return -EINVAL; | |
837 | ||
838 | /* | |
839 | * Round up the value to the closest page size | |
840 | */ | |
841 | *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT; | |
842 | return 0; | |
843 | } | |
844 | ||
845 | static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) | |
846 | { | |
847 | return res_counter_read_u64(&mem_cgroup_from_cont(cont)->res, | |
848 | cft->private); | |
849 | } | |
850 | ||
851 | static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft, | |
852 | struct file *file, const char __user *userbuf, | |
853 | size_t nbytes, loff_t *ppos) | |
854 | { | |
855 | return res_counter_write(&mem_cgroup_from_cont(cont)->res, | |
856 | cft->private, userbuf, nbytes, ppos, | |
857 | mem_cgroup_write_strategy); | |
858 | } | |
859 | ||
860 | static ssize_t mem_cgroup_max_reset(struct cgroup *cont, struct cftype *cft, | |
861 | struct file *file, const char __user *userbuf, | |
862 | size_t nbytes, loff_t *ppos) | |
863 | { | |
864 | struct mem_cgroup *mem; | |
865 | ||
866 | mem = mem_cgroup_from_cont(cont); | |
867 | res_counter_reset_max(&mem->res); | |
868 | return nbytes; | |
869 | } | |
870 | ||
871 | static ssize_t mem_force_empty_write(struct cgroup *cont, | |
872 | struct cftype *cft, struct file *file, | |
873 | const char __user *userbuf, | |
874 | size_t nbytes, loff_t *ppos) | |
875 | { | |
876 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
877 | int ret = mem_cgroup_force_empty(mem); | |
878 | if (!ret) | |
879 | ret = nbytes; | |
880 | return ret; | |
881 | } | |
882 | ||
883 | static const struct mem_cgroup_stat_desc { | |
884 | const char *msg; | |
885 | u64 unit; | |
886 | } mem_cgroup_stat_desc[] = { | |
887 | [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, }, | |
888 | [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, }, | |
889 | }; | |
890 | ||
891 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, | |
892 | struct cgroup_map_cb *cb) | |
893 | { | |
894 | struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); | |
895 | struct mem_cgroup_stat *stat = &mem_cont->stat; | |
896 | int i; | |
897 | ||
898 | for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) { | |
899 | s64 val; | |
900 | ||
901 | val = mem_cgroup_read_stat(stat, i); | |
902 | val *= mem_cgroup_stat_desc[i].unit; | |
903 | cb->fill(cb, mem_cgroup_stat_desc[i].msg, val); | |
904 | } | |
905 | /* showing # of active pages */ | |
906 | { | |
907 | unsigned long active, inactive; | |
908 | ||
909 | inactive = mem_cgroup_get_all_zonestat(mem_cont, | |
910 | MEM_CGROUP_ZSTAT_INACTIVE); | |
911 | active = mem_cgroup_get_all_zonestat(mem_cont, | |
912 | MEM_CGROUP_ZSTAT_ACTIVE); | |
913 | cb->fill(cb, "active", (active) * PAGE_SIZE); | |
914 | cb->fill(cb, "inactive", (inactive) * PAGE_SIZE); | |
915 | } | |
916 | return 0; | |
917 | } | |
918 | ||
919 | static struct cftype mem_cgroup_files[] = { | |
920 | { | |
921 | .name = "usage_in_bytes", | |
922 | .private = RES_USAGE, | |
923 | .read_u64 = mem_cgroup_read, | |
924 | }, | |
925 | { | |
926 | .name = "max_usage_in_bytes", | |
927 | .private = RES_MAX_USAGE, | |
928 | .write = mem_cgroup_max_reset, | |
929 | .read_u64 = mem_cgroup_read, | |
930 | }, | |
931 | { | |
932 | .name = "limit_in_bytes", | |
933 | .private = RES_LIMIT, | |
934 | .write = mem_cgroup_write, | |
935 | .read_u64 = mem_cgroup_read, | |
936 | }, | |
937 | { | |
938 | .name = "failcnt", | |
939 | .private = RES_FAILCNT, | |
940 | .read_u64 = mem_cgroup_read, | |
941 | }, | |
942 | { | |
943 | .name = "force_empty", | |
944 | .write = mem_force_empty_write, | |
945 | }, | |
946 | { | |
947 | .name = "stat", | |
948 | .read_map = mem_control_stat_show, | |
949 | }, | |
950 | }; | |
951 | ||
952 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) | |
953 | { | |
954 | struct mem_cgroup_per_node *pn; | |
955 | struct mem_cgroup_per_zone *mz; | |
956 | int zone, tmp = node; | |
957 | /* | |
958 | * This routine is called against possible nodes. | |
959 | * But it's BUG to call kmalloc() against offline node. | |
960 | * | |
961 | * TODO: this routine can waste much memory for nodes which will | |
962 | * never be onlined. It's better to use memory hotplug callback | |
963 | * function. | |
964 | */ | |
965 | if (!node_state(node, N_NORMAL_MEMORY)) | |
966 | tmp = -1; | |
967 | pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp); | |
968 | if (!pn) | |
969 | return 1; | |
970 | ||
971 | mem->info.nodeinfo[node] = pn; | |
972 | memset(pn, 0, sizeof(*pn)); | |
973 | ||
974 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
975 | mz = &pn->zoneinfo[zone]; | |
976 | INIT_LIST_HEAD(&mz->active_list); | |
977 | INIT_LIST_HEAD(&mz->inactive_list); | |
978 | spin_lock_init(&mz->lru_lock); | |
979 | } | |
980 | return 0; | |
981 | } | |
982 | ||
983 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) | |
984 | { | |
985 | kfree(mem->info.nodeinfo[node]); | |
986 | } | |
987 | ||
988 | static struct cgroup_subsys_state * | |
989 | mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) | |
990 | { | |
991 | struct mem_cgroup *mem; | |
992 | int node; | |
993 | ||
994 | if (unlikely((cont->parent) == NULL)) { | |
995 | mem = &init_mem_cgroup; | |
996 | page_cgroup_cache = KMEM_CACHE(page_cgroup, SLAB_PANIC); | |
997 | } else { | |
998 | mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL); | |
999 | } | |
1000 | ||
1001 | if (mem == NULL) | |
1002 | return ERR_PTR(-ENOMEM); | |
1003 | ||
1004 | res_counter_init(&mem->res); | |
1005 | ||
1006 | memset(&mem->info, 0, sizeof(mem->info)); | |
1007 | ||
1008 | for_each_node_state(node, N_POSSIBLE) | |
1009 | if (alloc_mem_cgroup_per_zone_info(mem, node)) | |
1010 | goto free_out; | |
1011 | ||
1012 | return &mem->css; | |
1013 | free_out: | |
1014 | for_each_node_state(node, N_POSSIBLE) | |
1015 | free_mem_cgroup_per_zone_info(mem, node); | |
1016 | if (cont->parent != NULL) | |
1017 | kfree(mem); | |
1018 | return ERR_PTR(-ENOMEM); | |
1019 | } | |
1020 | ||
1021 | static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss, | |
1022 | struct cgroup *cont) | |
1023 | { | |
1024 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
1025 | mem_cgroup_force_empty(mem); | |
1026 | } | |
1027 | ||
1028 | static void mem_cgroup_destroy(struct cgroup_subsys *ss, | |
1029 | struct cgroup *cont) | |
1030 | { | |
1031 | int node; | |
1032 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
1033 | ||
1034 | for_each_node_state(node, N_POSSIBLE) | |
1035 | free_mem_cgroup_per_zone_info(mem, node); | |
1036 | ||
1037 | kfree(mem_cgroup_from_cont(cont)); | |
1038 | } | |
1039 | ||
1040 | static int mem_cgroup_populate(struct cgroup_subsys *ss, | |
1041 | struct cgroup *cont) | |
1042 | { | |
1043 | if (mem_cgroup_subsys.disabled) | |
1044 | return 0; | |
1045 | return cgroup_add_files(cont, ss, mem_cgroup_files, | |
1046 | ARRAY_SIZE(mem_cgroup_files)); | |
1047 | } | |
1048 | ||
1049 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, | |
1050 | struct cgroup *cont, | |
1051 | struct cgroup *old_cont, | |
1052 | struct task_struct *p) | |
1053 | { | |
1054 | struct mm_struct *mm; | |
1055 | struct mem_cgroup *mem, *old_mem; | |
1056 | ||
1057 | if (mem_cgroup_subsys.disabled) | |
1058 | return; | |
1059 | ||
1060 | mm = get_task_mm(p); | |
1061 | if (mm == NULL) | |
1062 | return; | |
1063 | ||
1064 | mem = mem_cgroup_from_cont(cont); | |
1065 | old_mem = mem_cgroup_from_cont(old_cont); | |
1066 | ||
1067 | if (mem == old_mem) | |
1068 | goto out; | |
1069 | ||
1070 | /* | |
1071 | * Only thread group leaders are allowed to migrate, the mm_struct is | |
1072 | * in effect owned by the leader | |
1073 | */ | |
1074 | if (!thread_group_leader(p)) | |
1075 | goto out; | |
1076 | ||
1077 | out: | |
1078 | mmput(mm); | |
1079 | } | |
1080 | ||
1081 | struct cgroup_subsys mem_cgroup_subsys = { | |
1082 | .name = "memory", | |
1083 | .subsys_id = mem_cgroup_subsys_id, | |
1084 | .create = mem_cgroup_create, | |
1085 | .pre_destroy = mem_cgroup_pre_destroy, | |
1086 | .destroy = mem_cgroup_destroy, | |
1087 | .populate = mem_cgroup_populate, | |
1088 | .attach = mem_cgroup_move_task, | |
1089 | .early_init = 0, | |
1090 | }; |