]>
Commit | Line | Data |
---|---|---|
1 | /* memcontrol.c - Memory Controller | |
2 | * | |
3 | * Copyright IBM Corporation, 2007 | |
4 | * Author Balbir Singh <balbir@linux.vnet.ibm.com> | |
5 | * | |
6 | * Copyright 2007 OpenVZ SWsoft Inc | |
7 | * Author: Pavel Emelianov <xemul@openvz.org> | |
8 | * | |
9 | * Memory thresholds | |
10 | * Copyright (C) 2009 Nokia Corporation | |
11 | * Author: Kirill A. Shutemov | |
12 | * | |
13 | * This program is free software; you can redistribute it and/or modify | |
14 | * it under the terms of the GNU General Public License as published by | |
15 | * the Free Software Foundation; either version 2 of the License, or | |
16 | * (at your option) any later version. | |
17 | * | |
18 | * This program is distributed in the hope that it will be useful, | |
19 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
20 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
21 | * GNU General Public License for more details. | |
22 | */ | |
23 | ||
24 | #include <linux/res_counter.h> | |
25 | #include <linux/memcontrol.h> | |
26 | #include <linux/cgroup.h> | |
27 | #include <linux/mm.h> | |
28 | #include <linux/hugetlb.h> | |
29 | #include <linux/pagemap.h> | |
30 | #include <linux/smp.h> | |
31 | #include <linux/page-flags.h> | |
32 | #include <linux/backing-dev.h> | |
33 | #include <linux/bit_spinlock.h> | |
34 | #include <linux/rcupdate.h> | |
35 | #include <linux/limits.h> | |
36 | #include <linux/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 <linux/oom.h> | |
51 | #include "internal.h" | |
52 | ||
53 | #include <asm/uaccess.h> | |
54 | ||
55 | #include <trace/events/vmscan.h> | |
56 | ||
57 | struct cgroup_subsys mem_cgroup_subsys __read_mostly; | |
58 | #define MEM_CGROUP_RECLAIM_RETRIES 5 | |
59 | struct mem_cgroup *root_mem_cgroup __read_mostly; | |
60 | ||
61 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
62 | /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ | |
63 | int do_swap_account __read_mostly; | |
64 | static int really_do_swap_account __initdata = 1; /* for remember boot option*/ | |
65 | #else | |
66 | #define do_swap_account (0) | |
67 | #endif | |
68 | ||
69 | /* | |
70 | * Per memcg event counter is incremented at every pagein/pageout. This counter | |
71 | * is used for trigger some periodic events. This is straightforward and better | |
72 | * than using jiffies etc. to handle periodic memcg event. | |
73 | * | |
74 | * These values will be used as !((event) & ((1 <<(thresh)) - 1)) | |
75 | */ | |
76 | #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */ | |
77 | #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */ | |
78 | ||
79 | /* | |
80 | * Statistics for memory cgroup. | |
81 | */ | |
82 | enum mem_cgroup_stat_index { | |
83 | /* | |
84 | * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. | |
85 | */ | |
86 | MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ | |
87 | MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ | |
88 | MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ | |
89 | MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */ | |
90 | MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */ | |
91 | MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ | |
92 | MEM_CGROUP_EVENTS, /* incremented at every pagein/pageout */ | |
93 | ||
94 | MEM_CGROUP_STAT_NSTATS, | |
95 | }; | |
96 | ||
97 | struct mem_cgroup_stat_cpu { | |
98 | s64 count[MEM_CGROUP_STAT_NSTATS]; | |
99 | }; | |
100 | ||
101 | /* | |
102 | * per-zone information in memory controller. | |
103 | */ | |
104 | struct mem_cgroup_per_zone { | |
105 | /* | |
106 | * spin_lock to protect the per cgroup LRU | |
107 | */ | |
108 | struct list_head lists[NR_LRU_LISTS]; | |
109 | unsigned long count[NR_LRU_LISTS]; | |
110 | ||
111 | struct zone_reclaim_stat reclaim_stat; | |
112 | struct rb_node tree_node; /* RB tree node */ | |
113 | unsigned long long usage_in_excess;/* Set to the value by which */ | |
114 | /* the soft limit is exceeded*/ | |
115 | bool on_tree; | |
116 | struct mem_cgroup *mem; /* Back pointer, we cannot */ | |
117 | /* use container_of */ | |
118 | }; | |
119 | /* Macro for accessing counter */ | |
120 | #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) | |
121 | ||
122 | struct mem_cgroup_per_node { | |
123 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; | |
124 | }; | |
125 | ||
126 | struct mem_cgroup_lru_info { | |
127 | struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; | |
128 | }; | |
129 | ||
130 | /* | |
131 | * Cgroups above their limits are maintained in a RB-Tree, independent of | |
132 | * their hierarchy representation | |
133 | */ | |
134 | ||
135 | struct mem_cgroup_tree_per_zone { | |
136 | struct rb_root rb_root; | |
137 | spinlock_t lock; | |
138 | }; | |
139 | ||
140 | struct mem_cgroup_tree_per_node { | |
141 | struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; | |
142 | }; | |
143 | ||
144 | struct mem_cgroup_tree { | |
145 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; | |
146 | }; | |
147 | ||
148 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; | |
149 | ||
150 | struct mem_cgroup_threshold { | |
151 | struct eventfd_ctx *eventfd; | |
152 | u64 threshold; | |
153 | }; | |
154 | ||
155 | /* For threshold */ | |
156 | struct mem_cgroup_threshold_ary { | |
157 | /* An array index points to threshold just below usage. */ | |
158 | int current_threshold; | |
159 | /* Size of entries[] */ | |
160 | unsigned int size; | |
161 | /* Array of thresholds */ | |
162 | struct mem_cgroup_threshold entries[0]; | |
163 | }; | |
164 | ||
165 | struct mem_cgroup_thresholds { | |
166 | /* Primary thresholds array */ | |
167 | struct mem_cgroup_threshold_ary *primary; | |
168 | /* | |
169 | * Spare threshold array. | |
170 | * This is needed to make mem_cgroup_unregister_event() "never fail". | |
171 | * It must be able to store at least primary->size - 1 entries. | |
172 | */ | |
173 | struct mem_cgroup_threshold_ary *spare; | |
174 | }; | |
175 | ||
176 | /* for OOM */ | |
177 | struct mem_cgroup_eventfd_list { | |
178 | struct list_head list; | |
179 | struct eventfd_ctx *eventfd; | |
180 | }; | |
181 | ||
182 | static void mem_cgroup_threshold(struct mem_cgroup *mem); | |
183 | static void mem_cgroup_oom_notify(struct mem_cgroup *mem); | |
184 | ||
185 | /* | |
186 | * The memory controller data structure. The memory controller controls both | |
187 | * page cache and RSS per cgroup. We would eventually like to provide | |
188 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, | |
189 | * to help the administrator determine what knobs to tune. | |
190 | * | |
191 | * TODO: Add a water mark for the memory controller. Reclaim will begin when | |
192 | * we hit the water mark. May be even add a low water mark, such that | |
193 | * no reclaim occurs from a cgroup at it's low water mark, this is | |
194 | * a feature that will be implemented much later in the future. | |
195 | */ | |
196 | struct mem_cgroup { | |
197 | struct cgroup_subsys_state css; | |
198 | /* | |
199 | * the counter to account for memory usage | |
200 | */ | |
201 | struct res_counter res; | |
202 | /* | |
203 | * the counter to account for mem+swap usage. | |
204 | */ | |
205 | struct res_counter memsw; | |
206 | /* | |
207 | * Per cgroup active and inactive list, similar to the | |
208 | * per zone LRU lists. | |
209 | */ | |
210 | struct mem_cgroup_lru_info info; | |
211 | ||
212 | /* | |
213 | protect against reclaim related member. | |
214 | */ | |
215 | spinlock_t reclaim_param_lock; | |
216 | ||
217 | /* | |
218 | * While reclaiming in a hierarchy, we cache the last child we | |
219 | * reclaimed from. | |
220 | */ | |
221 | int last_scanned_child; | |
222 | /* | |
223 | * Should the accounting and control be hierarchical, per subtree? | |
224 | */ | |
225 | bool use_hierarchy; | |
226 | atomic_t oom_lock; | |
227 | atomic_t refcnt; | |
228 | ||
229 | unsigned int swappiness; | |
230 | /* OOM-Killer disable */ | |
231 | int oom_kill_disable; | |
232 | ||
233 | /* set when res.limit == memsw.limit */ | |
234 | bool memsw_is_minimum; | |
235 | ||
236 | /* protect arrays of thresholds */ | |
237 | struct mutex thresholds_lock; | |
238 | ||
239 | /* thresholds for memory usage. RCU-protected */ | |
240 | struct mem_cgroup_thresholds thresholds; | |
241 | ||
242 | /* thresholds for mem+swap usage. RCU-protected */ | |
243 | struct mem_cgroup_thresholds memsw_thresholds; | |
244 | ||
245 | /* For oom notifier event fd */ | |
246 | struct list_head oom_notify; | |
247 | ||
248 | /* | |
249 | * Should we move charges of a task when a task is moved into this | |
250 | * mem_cgroup ? And what type of charges should we move ? | |
251 | */ | |
252 | unsigned long move_charge_at_immigrate; | |
253 | /* | |
254 | * percpu counter. | |
255 | */ | |
256 | struct mem_cgroup_stat_cpu *stat; | |
257 | }; | |
258 | ||
259 | /* Stuffs for move charges at task migration. */ | |
260 | /* | |
261 | * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a | |
262 | * left-shifted bitmap of these types. | |
263 | */ | |
264 | enum move_type { | |
265 | MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ | |
266 | MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ | |
267 | NR_MOVE_TYPE, | |
268 | }; | |
269 | ||
270 | /* "mc" and its members are protected by cgroup_mutex */ | |
271 | static struct move_charge_struct { | |
272 | spinlock_t lock; /* for from, to, moving_task */ | |
273 | struct mem_cgroup *from; | |
274 | struct mem_cgroup *to; | |
275 | unsigned long precharge; | |
276 | unsigned long moved_charge; | |
277 | unsigned long moved_swap; | |
278 | struct task_struct *moving_task; /* a task moving charges */ | |
279 | wait_queue_head_t waitq; /* a waitq for other context */ | |
280 | } mc = { | |
281 | .lock = __SPIN_LOCK_UNLOCKED(mc.lock), | |
282 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), | |
283 | }; | |
284 | ||
285 | static bool move_anon(void) | |
286 | { | |
287 | return test_bit(MOVE_CHARGE_TYPE_ANON, | |
288 | &mc.to->move_charge_at_immigrate); | |
289 | } | |
290 | ||
291 | static bool move_file(void) | |
292 | { | |
293 | return test_bit(MOVE_CHARGE_TYPE_FILE, | |
294 | &mc.to->move_charge_at_immigrate); | |
295 | } | |
296 | ||
297 | /* | |
298 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft | |
299 | * limit reclaim to prevent infinite loops, if they ever occur. | |
300 | */ | |
301 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) | |
302 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) | |
303 | ||
304 | enum charge_type { | |
305 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, | |
306 | MEM_CGROUP_CHARGE_TYPE_MAPPED, | |
307 | MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ | |
308 | MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ | |
309 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ | |
310 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ | |
311 | NR_CHARGE_TYPE, | |
312 | }; | |
313 | ||
314 | /* only for here (for easy reading.) */ | |
315 | #define PCGF_CACHE (1UL << PCG_CACHE) | |
316 | #define PCGF_USED (1UL << PCG_USED) | |
317 | #define PCGF_LOCK (1UL << PCG_LOCK) | |
318 | /* Not used, but added here for completeness */ | |
319 | #define PCGF_ACCT (1UL << PCG_ACCT) | |
320 | ||
321 | /* for encoding cft->private value on file */ | |
322 | #define _MEM (0) | |
323 | #define _MEMSWAP (1) | |
324 | #define _OOM_TYPE (2) | |
325 | #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) | |
326 | #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) | |
327 | #define MEMFILE_ATTR(val) ((val) & 0xffff) | |
328 | /* Used for OOM nofiier */ | |
329 | #define OOM_CONTROL (0) | |
330 | ||
331 | /* | |
332 | * Reclaim flags for mem_cgroup_hierarchical_reclaim | |
333 | */ | |
334 | #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 | |
335 | #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) | |
336 | #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 | |
337 | #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) | |
338 | #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2 | |
339 | #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT) | |
340 | ||
341 | static void mem_cgroup_get(struct mem_cgroup *mem); | |
342 | static void mem_cgroup_put(struct mem_cgroup *mem); | |
343 | static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); | |
344 | static void drain_all_stock_async(void); | |
345 | ||
346 | static struct mem_cgroup_per_zone * | |
347 | mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) | |
348 | { | |
349 | return &mem->info.nodeinfo[nid]->zoneinfo[zid]; | |
350 | } | |
351 | ||
352 | struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem) | |
353 | { | |
354 | return &mem->css; | |
355 | } | |
356 | ||
357 | static struct mem_cgroup_per_zone * | |
358 | page_cgroup_zoneinfo(struct page_cgroup *pc) | |
359 | { | |
360 | struct mem_cgroup *mem = pc->mem_cgroup; | |
361 | int nid = page_cgroup_nid(pc); | |
362 | int zid = page_cgroup_zid(pc); | |
363 | ||
364 | if (!mem) | |
365 | return NULL; | |
366 | ||
367 | return mem_cgroup_zoneinfo(mem, nid, zid); | |
368 | } | |
369 | ||
370 | static struct mem_cgroup_tree_per_zone * | |
371 | soft_limit_tree_node_zone(int nid, int zid) | |
372 | { | |
373 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
374 | } | |
375 | ||
376 | static struct mem_cgroup_tree_per_zone * | |
377 | soft_limit_tree_from_page(struct page *page) | |
378 | { | |
379 | int nid = page_to_nid(page); | |
380 | int zid = page_zonenum(page); | |
381 | ||
382 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
383 | } | |
384 | ||
385 | static void | |
386 | __mem_cgroup_insert_exceeded(struct mem_cgroup *mem, | |
387 | struct mem_cgroup_per_zone *mz, | |
388 | struct mem_cgroup_tree_per_zone *mctz, | |
389 | unsigned long long new_usage_in_excess) | |
390 | { | |
391 | struct rb_node **p = &mctz->rb_root.rb_node; | |
392 | struct rb_node *parent = NULL; | |
393 | struct mem_cgroup_per_zone *mz_node; | |
394 | ||
395 | if (mz->on_tree) | |
396 | return; | |
397 | ||
398 | mz->usage_in_excess = new_usage_in_excess; | |
399 | if (!mz->usage_in_excess) | |
400 | return; | |
401 | while (*p) { | |
402 | parent = *p; | |
403 | mz_node = rb_entry(parent, struct mem_cgroup_per_zone, | |
404 | tree_node); | |
405 | if (mz->usage_in_excess < mz_node->usage_in_excess) | |
406 | p = &(*p)->rb_left; | |
407 | /* | |
408 | * We can't avoid mem cgroups that are over their soft | |
409 | * limit by the same amount | |
410 | */ | |
411 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) | |
412 | p = &(*p)->rb_right; | |
413 | } | |
414 | rb_link_node(&mz->tree_node, parent, p); | |
415 | rb_insert_color(&mz->tree_node, &mctz->rb_root); | |
416 | mz->on_tree = true; | |
417 | } | |
418 | ||
419 | static void | |
420 | __mem_cgroup_remove_exceeded(struct mem_cgroup *mem, | |
421 | struct mem_cgroup_per_zone *mz, | |
422 | struct mem_cgroup_tree_per_zone *mctz) | |
423 | { | |
424 | if (!mz->on_tree) | |
425 | return; | |
426 | rb_erase(&mz->tree_node, &mctz->rb_root); | |
427 | mz->on_tree = false; | |
428 | } | |
429 | ||
430 | static void | |
431 | mem_cgroup_remove_exceeded(struct mem_cgroup *mem, | |
432 | struct mem_cgroup_per_zone *mz, | |
433 | struct mem_cgroup_tree_per_zone *mctz) | |
434 | { | |
435 | spin_lock(&mctz->lock); | |
436 | __mem_cgroup_remove_exceeded(mem, mz, mctz); | |
437 | spin_unlock(&mctz->lock); | |
438 | } | |
439 | ||
440 | ||
441 | static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page) | |
442 | { | |
443 | unsigned long long excess; | |
444 | struct mem_cgroup_per_zone *mz; | |
445 | struct mem_cgroup_tree_per_zone *mctz; | |
446 | int nid = page_to_nid(page); | |
447 | int zid = page_zonenum(page); | |
448 | mctz = soft_limit_tree_from_page(page); | |
449 | ||
450 | /* | |
451 | * Necessary to update all ancestors when hierarchy is used. | |
452 | * because their event counter is not touched. | |
453 | */ | |
454 | for (; mem; mem = parent_mem_cgroup(mem)) { | |
455 | mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
456 | excess = res_counter_soft_limit_excess(&mem->res); | |
457 | /* | |
458 | * We have to update the tree if mz is on RB-tree or | |
459 | * mem is over its softlimit. | |
460 | */ | |
461 | if (excess || mz->on_tree) { | |
462 | spin_lock(&mctz->lock); | |
463 | /* if on-tree, remove it */ | |
464 | if (mz->on_tree) | |
465 | __mem_cgroup_remove_exceeded(mem, mz, mctz); | |
466 | /* | |
467 | * Insert again. mz->usage_in_excess will be updated. | |
468 | * If excess is 0, no tree ops. | |
469 | */ | |
470 | __mem_cgroup_insert_exceeded(mem, mz, mctz, excess); | |
471 | spin_unlock(&mctz->lock); | |
472 | } | |
473 | } | |
474 | } | |
475 | ||
476 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem) | |
477 | { | |
478 | int node, zone; | |
479 | struct mem_cgroup_per_zone *mz; | |
480 | struct mem_cgroup_tree_per_zone *mctz; | |
481 | ||
482 | for_each_node_state(node, N_POSSIBLE) { | |
483 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
484 | mz = mem_cgroup_zoneinfo(mem, node, zone); | |
485 | mctz = soft_limit_tree_node_zone(node, zone); | |
486 | mem_cgroup_remove_exceeded(mem, mz, mctz); | |
487 | } | |
488 | } | |
489 | } | |
490 | ||
491 | static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem) | |
492 | { | |
493 | return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT; | |
494 | } | |
495 | ||
496 | static struct mem_cgroup_per_zone * | |
497 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
498 | { | |
499 | struct rb_node *rightmost = NULL; | |
500 | struct mem_cgroup_per_zone *mz; | |
501 | ||
502 | retry: | |
503 | mz = NULL; | |
504 | rightmost = rb_last(&mctz->rb_root); | |
505 | if (!rightmost) | |
506 | goto done; /* Nothing to reclaim from */ | |
507 | ||
508 | mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); | |
509 | /* | |
510 | * Remove the node now but someone else can add it back, | |
511 | * we will to add it back at the end of reclaim to its correct | |
512 | * position in the tree. | |
513 | */ | |
514 | __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); | |
515 | if (!res_counter_soft_limit_excess(&mz->mem->res) || | |
516 | !css_tryget(&mz->mem->css)) | |
517 | goto retry; | |
518 | done: | |
519 | return mz; | |
520 | } | |
521 | ||
522 | static struct mem_cgroup_per_zone * | |
523 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
524 | { | |
525 | struct mem_cgroup_per_zone *mz; | |
526 | ||
527 | spin_lock(&mctz->lock); | |
528 | mz = __mem_cgroup_largest_soft_limit_node(mctz); | |
529 | spin_unlock(&mctz->lock); | |
530 | return mz; | |
531 | } | |
532 | ||
533 | static s64 mem_cgroup_read_stat(struct mem_cgroup *mem, | |
534 | enum mem_cgroup_stat_index idx) | |
535 | { | |
536 | int cpu; | |
537 | s64 val = 0; | |
538 | ||
539 | for_each_possible_cpu(cpu) | |
540 | val += per_cpu(mem->stat->count[idx], cpu); | |
541 | return val; | |
542 | } | |
543 | ||
544 | static s64 mem_cgroup_local_usage(struct mem_cgroup *mem) | |
545 | { | |
546 | s64 ret; | |
547 | ||
548 | ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); | |
549 | ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); | |
550 | return ret; | |
551 | } | |
552 | ||
553 | static void mem_cgroup_swap_statistics(struct mem_cgroup *mem, | |
554 | bool charge) | |
555 | { | |
556 | int val = (charge) ? 1 : -1; | |
557 | this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); | |
558 | } | |
559 | ||
560 | static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, | |
561 | struct page_cgroup *pc, | |
562 | bool charge) | |
563 | { | |
564 | int val = (charge) ? 1 : -1; | |
565 | ||
566 | preempt_disable(); | |
567 | ||
568 | if (PageCgroupCache(pc)) | |
569 | __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val); | |
570 | else | |
571 | __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val); | |
572 | ||
573 | if (charge) | |
574 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]); | |
575 | else | |
576 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]); | |
577 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]); | |
578 | ||
579 | preempt_enable(); | |
580 | } | |
581 | ||
582 | static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, | |
583 | enum lru_list idx) | |
584 | { | |
585 | int nid, zid; | |
586 | struct mem_cgroup_per_zone *mz; | |
587 | u64 total = 0; | |
588 | ||
589 | for_each_online_node(nid) | |
590 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
591 | mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
592 | total += MEM_CGROUP_ZSTAT(mz, idx); | |
593 | } | |
594 | return total; | |
595 | } | |
596 | ||
597 | static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift) | |
598 | { | |
599 | s64 val; | |
600 | ||
601 | val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]); | |
602 | ||
603 | return !(val & ((1 << event_mask_shift) - 1)); | |
604 | } | |
605 | ||
606 | /* | |
607 | * Check events in order. | |
608 | * | |
609 | */ | |
610 | static void memcg_check_events(struct mem_cgroup *mem, struct page *page) | |
611 | { | |
612 | /* threshold event is triggered in finer grain than soft limit */ | |
613 | if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) { | |
614 | mem_cgroup_threshold(mem); | |
615 | if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH))) | |
616 | mem_cgroup_update_tree(mem, page); | |
617 | } | |
618 | } | |
619 | ||
620 | static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) | |
621 | { | |
622 | return container_of(cgroup_subsys_state(cont, | |
623 | mem_cgroup_subsys_id), struct mem_cgroup, | |
624 | css); | |
625 | } | |
626 | ||
627 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) | |
628 | { | |
629 | /* | |
630 | * mm_update_next_owner() may clear mm->owner to NULL | |
631 | * if it races with swapoff, page migration, etc. | |
632 | * So this can be called with p == NULL. | |
633 | */ | |
634 | if (unlikely(!p)) | |
635 | return NULL; | |
636 | ||
637 | return container_of(task_subsys_state(p, mem_cgroup_subsys_id), | |
638 | struct mem_cgroup, css); | |
639 | } | |
640 | ||
641 | static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) | |
642 | { | |
643 | struct mem_cgroup *mem = NULL; | |
644 | ||
645 | if (!mm) | |
646 | return NULL; | |
647 | /* | |
648 | * Because we have no locks, mm->owner's may be being moved to other | |
649 | * cgroup. We use css_tryget() here even if this looks | |
650 | * pessimistic (rather than adding locks here). | |
651 | */ | |
652 | rcu_read_lock(); | |
653 | do { | |
654 | mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
655 | if (unlikely(!mem)) | |
656 | break; | |
657 | } while (!css_tryget(&mem->css)); | |
658 | rcu_read_unlock(); | |
659 | return mem; | |
660 | } | |
661 | ||
662 | /* | |
663 | * Call callback function against all cgroup under hierarchy tree. | |
664 | */ | |
665 | static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data, | |
666 | int (*func)(struct mem_cgroup *, void *)) | |
667 | { | |
668 | int found, ret, nextid; | |
669 | struct cgroup_subsys_state *css; | |
670 | struct mem_cgroup *mem; | |
671 | ||
672 | if (!root->use_hierarchy) | |
673 | return (*func)(root, data); | |
674 | ||
675 | nextid = 1; | |
676 | do { | |
677 | ret = 0; | |
678 | mem = NULL; | |
679 | ||
680 | rcu_read_lock(); | |
681 | css = css_get_next(&mem_cgroup_subsys, nextid, &root->css, | |
682 | &found); | |
683 | if (css && css_tryget(css)) | |
684 | mem = container_of(css, struct mem_cgroup, css); | |
685 | rcu_read_unlock(); | |
686 | ||
687 | if (mem) { | |
688 | ret = (*func)(mem, data); | |
689 | css_put(&mem->css); | |
690 | } | |
691 | nextid = found + 1; | |
692 | } while (!ret && css); | |
693 | ||
694 | return ret; | |
695 | } | |
696 | ||
697 | static inline bool mem_cgroup_is_root(struct mem_cgroup *mem) | |
698 | { | |
699 | return (mem == root_mem_cgroup); | |
700 | } | |
701 | ||
702 | /* | |
703 | * Following LRU functions are allowed to be used without PCG_LOCK. | |
704 | * Operations are called by routine of global LRU independently from memcg. | |
705 | * What we have to take care of here is validness of pc->mem_cgroup. | |
706 | * | |
707 | * Changes to pc->mem_cgroup happens when | |
708 | * 1. charge | |
709 | * 2. moving account | |
710 | * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. | |
711 | * It is added to LRU before charge. | |
712 | * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. | |
713 | * When moving account, the page is not on LRU. It's isolated. | |
714 | */ | |
715 | ||
716 | void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) | |
717 | { | |
718 | struct page_cgroup *pc; | |
719 | struct mem_cgroup_per_zone *mz; | |
720 | ||
721 | if (mem_cgroup_disabled()) | |
722 | return; | |
723 | pc = lookup_page_cgroup(page); | |
724 | /* can happen while we handle swapcache. */ | |
725 | if (!TestClearPageCgroupAcctLRU(pc)) | |
726 | return; | |
727 | VM_BUG_ON(!pc->mem_cgroup); | |
728 | /* | |
729 | * We don't check PCG_USED bit. It's cleared when the "page" is finally | |
730 | * removed from global LRU. | |
731 | */ | |
732 | mz = page_cgroup_zoneinfo(pc); | |
733 | MEM_CGROUP_ZSTAT(mz, lru) -= 1; | |
734 | if (mem_cgroup_is_root(pc->mem_cgroup)) | |
735 | return; | |
736 | VM_BUG_ON(list_empty(&pc->lru)); | |
737 | list_del_init(&pc->lru); | |
738 | return; | |
739 | } | |
740 | ||
741 | void mem_cgroup_del_lru(struct page *page) | |
742 | { | |
743 | mem_cgroup_del_lru_list(page, page_lru(page)); | |
744 | } | |
745 | ||
746 | void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru) | |
747 | { | |
748 | struct mem_cgroup_per_zone *mz; | |
749 | struct page_cgroup *pc; | |
750 | ||
751 | if (mem_cgroup_disabled()) | |
752 | return; | |
753 | ||
754 | pc = lookup_page_cgroup(page); | |
755 | /* | |
756 | * Used bit is set without atomic ops but after smp_wmb(). | |
757 | * For making pc->mem_cgroup visible, insert smp_rmb() here. | |
758 | */ | |
759 | smp_rmb(); | |
760 | /* unused or root page is not rotated. */ | |
761 | if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup)) | |
762 | return; | |
763 | mz = page_cgroup_zoneinfo(pc); | |
764 | list_move(&pc->lru, &mz->lists[lru]); | |
765 | } | |
766 | ||
767 | void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru) | |
768 | { | |
769 | struct page_cgroup *pc; | |
770 | struct mem_cgroup_per_zone *mz; | |
771 | ||
772 | if (mem_cgroup_disabled()) | |
773 | return; | |
774 | pc = lookup_page_cgroup(page); | |
775 | VM_BUG_ON(PageCgroupAcctLRU(pc)); | |
776 | /* | |
777 | * Used bit is set without atomic ops but after smp_wmb(). | |
778 | * For making pc->mem_cgroup visible, insert smp_rmb() here. | |
779 | */ | |
780 | smp_rmb(); | |
781 | if (!PageCgroupUsed(pc)) | |
782 | return; | |
783 | ||
784 | mz = page_cgroup_zoneinfo(pc); | |
785 | MEM_CGROUP_ZSTAT(mz, lru) += 1; | |
786 | SetPageCgroupAcctLRU(pc); | |
787 | if (mem_cgroup_is_root(pc->mem_cgroup)) | |
788 | return; | |
789 | list_add(&pc->lru, &mz->lists[lru]); | |
790 | } | |
791 | ||
792 | /* | |
793 | * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to | |
794 | * lru because the page may.be reused after it's fully uncharged (because of | |
795 | * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge | |
796 | * it again. This function is only used to charge SwapCache. It's done under | |
797 | * lock_page and expected that zone->lru_lock is never held. | |
798 | */ | |
799 | static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page) | |
800 | { | |
801 | unsigned long flags; | |
802 | struct zone *zone = page_zone(page); | |
803 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
804 | ||
805 | spin_lock_irqsave(&zone->lru_lock, flags); | |
806 | /* | |
807 | * Forget old LRU when this page_cgroup is *not* used. This Used bit | |
808 | * is guarded by lock_page() because the page is SwapCache. | |
809 | */ | |
810 | if (!PageCgroupUsed(pc)) | |
811 | mem_cgroup_del_lru_list(page, page_lru(page)); | |
812 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
813 | } | |
814 | ||
815 | static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page) | |
816 | { | |
817 | unsigned long flags; | |
818 | struct zone *zone = page_zone(page); | |
819 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
820 | ||
821 | spin_lock_irqsave(&zone->lru_lock, flags); | |
822 | /* link when the page is linked to LRU but page_cgroup isn't */ | |
823 | if (PageLRU(page) && !PageCgroupAcctLRU(pc)) | |
824 | mem_cgroup_add_lru_list(page, page_lru(page)); | |
825 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
826 | } | |
827 | ||
828 | ||
829 | void mem_cgroup_move_lists(struct page *page, | |
830 | enum lru_list from, enum lru_list to) | |
831 | { | |
832 | if (mem_cgroup_disabled()) | |
833 | return; | |
834 | mem_cgroup_del_lru_list(page, from); | |
835 | mem_cgroup_add_lru_list(page, to); | |
836 | } | |
837 | ||
838 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) | |
839 | { | |
840 | int ret; | |
841 | struct mem_cgroup *curr = NULL; | |
842 | struct task_struct *p; | |
843 | ||
844 | p = find_lock_task_mm(task); | |
845 | if (!p) | |
846 | return 0; | |
847 | curr = try_get_mem_cgroup_from_mm(p->mm); | |
848 | task_unlock(p); | |
849 | if (!curr) | |
850 | return 0; | |
851 | /* | |
852 | * We should check use_hierarchy of "mem" not "curr". Because checking | |
853 | * use_hierarchy of "curr" here make this function true if hierarchy is | |
854 | * enabled in "curr" and "curr" is a child of "mem" in *cgroup* | |
855 | * hierarchy(even if use_hierarchy is disabled in "mem"). | |
856 | */ | |
857 | if (mem->use_hierarchy) | |
858 | ret = css_is_ancestor(&curr->css, &mem->css); | |
859 | else | |
860 | ret = (curr == mem); | |
861 | css_put(&curr->css); | |
862 | return ret; | |
863 | } | |
864 | ||
865 | static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) | |
866 | { | |
867 | unsigned long active; | |
868 | unsigned long inactive; | |
869 | unsigned long gb; | |
870 | unsigned long inactive_ratio; | |
871 | ||
872 | inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); | |
873 | active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); | |
874 | ||
875 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
876 | if (gb) | |
877 | inactive_ratio = int_sqrt(10 * gb); | |
878 | else | |
879 | inactive_ratio = 1; | |
880 | ||
881 | if (present_pages) { | |
882 | present_pages[0] = inactive; | |
883 | present_pages[1] = active; | |
884 | } | |
885 | ||
886 | return inactive_ratio; | |
887 | } | |
888 | ||
889 | int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) | |
890 | { | |
891 | unsigned long active; | |
892 | unsigned long inactive; | |
893 | unsigned long present_pages[2]; | |
894 | unsigned long inactive_ratio; | |
895 | ||
896 | inactive_ratio = calc_inactive_ratio(memcg, present_pages); | |
897 | ||
898 | inactive = present_pages[0]; | |
899 | active = present_pages[1]; | |
900 | ||
901 | if (inactive * inactive_ratio < active) | |
902 | return 1; | |
903 | ||
904 | return 0; | |
905 | } | |
906 | ||
907 | int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) | |
908 | { | |
909 | unsigned long active; | |
910 | unsigned long inactive; | |
911 | ||
912 | inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); | |
913 | active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); | |
914 | ||
915 | return (active > inactive); | |
916 | } | |
917 | ||
918 | unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg, | |
919 | struct zone *zone, | |
920 | enum lru_list lru) | |
921 | { | |
922 | int nid = zone_to_nid(zone); | |
923 | int zid = zone_idx(zone); | |
924 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
925 | ||
926 | return MEM_CGROUP_ZSTAT(mz, lru); | |
927 | } | |
928 | ||
929 | struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, | |
930 | struct zone *zone) | |
931 | { | |
932 | int nid = zone_to_nid(zone); | |
933 | int zid = zone_idx(zone); | |
934 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
935 | ||
936 | return &mz->reclaim_stat; | |
937 | } | |
938 | ||
939 | struct zone_reclaim_stat * | |
940 | mem_cgroup_get_reclaim_stat_from_page(struct page *page) | |
941 | { | |
942 | struct page_cgroup *pc; | |
943 | struct mem_cgroup_per_zone *mz; | |
944 | ||
945 | if (mem_cgroup_disabled()) | |
946 | return NULL; | |
947 | ||
948 | pc = lookup_page_cgroup(page); | |
949 | /* | |
950 | * Used bit is set without atomic ops but after smp_wmb(). | |
951 | * For making pc->mem_cgroup visible, insert smp_rmb() here. | |
952 | */ | |
953 | smp_rmb(); | |
954 | if (!PageCgroupUsed(pc)) | |
955 | return NULL; | |
956 | ||
957 | mz = page_cgroup_zoneinfo(pc); | |
958 | if (!mz) | |
959 | return NULL; | |
960 | ||
961 | return &mz->reclaim_stat; | |
962 | } | |
963 | ||
964 | unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, | |
965 | struct list_head *dst, | |
966 | unsigned long *scanned, int order, | |
967 | int mode, struct zone *z, | |
968 | struct mem_cgroup *mem_cont, | |
969 | int active, int file) | |
970 | { | |
971 | unsigned long nr_taken = 0; | |
972 | struct page *page; | |
973 | unsigned long scan; | |
974 | LIST_HEAD(pc_list); | |
975 | struct list_head *src; | |
976 | struct page_cgroup *pc, *tmp; | |
977 | int nid = zone_to_nid(z); | |
978 | int zid = zone_idx(z); | |
979 | struct mem_cgroup_per_zone *mz; | |
980 | int lru = LRU_FILE * file + active; | |
981 | int ret; | |
982 | ||
983 | BUG_ON(!mem_cont); | |
984 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); | |
985 | src = &mz->lists[lru]; | |
986 | ||
987 | scan = 0; | |
988 | list_for_each_entry_safe_reverse(pc, tmp, src, lru) { | |
989 | if (scan >= nr_to_scan) | |
990 | break; | |
991 | ||
992 | page = pc->page; | |
993 | if (unlikely(!PageCgroupUsed(pc))) | |
994 | continue; | |
995 | if (unlikely(!PageLRU(page))) | |
996 | continue; | |
997 | ||
998 | scan++; | |
999 | ret = __isolate_lru_page(page, mode, file); | |
1000 | switch (ret) { | |
1001 | case 0: | |
1002 | list_move(&page->lru, dst); | |
1003 | mem_cgroup_del_lru(page); | |
1004 | nr_taken++; | |
1005 | break; | |
1006 | case -EBUSY: | |
1007 | /* we don't affect global LRU but rotate in our LRU */ | |
1008 | mem_cgroup_rotate_lru_list(page, page_lru(page)); | |
1009 | break; | |
1010 | default: | |
1011 | break; | |
1012 | } | |
1013 | } | |
1014 | ||
1015 | *scanned = scan; | |
1016 | ||
1017 | trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken, | |
1018 | 0, 0, 0, mode); | |
1019 | ||
1020 | return nr_taken; | |
1021 | } | |
1022 | ||
1023 | #define mem_cgroup_from_res_counter(counter, member) \ | |
1024 | container_of(counter, struct mem_cgroup, member) | |
1025 | ||
1026 | static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem) | |
1027 | { | |
1028 | if (do_swap_account) { | |
1029 | if (res_counter_check_under_limit(&mem->res) && | |
1030 | res_counter_check_under_limit(&mem->memsw)) | |
1031 | return true; | |
1032 | } else | |
1033 | if (res_counter_check_under_limit(&mem->res)) | |
1034 | return true; | |
1035 | return false; | |
1036 | } | |
1037 | ||
1038 | static unsigned int get_swappiness(struct mem_cgroup *memcg) | |
1039 | { | |
1040 | struct cgroup *cgrp = memcg->css.cgroup; | |
1041 | unsigned int swappiness; | |
1042 | ||
1043 | /* root ? */ | |
1044 | if (cgrp->parent == NULL) | |
1045 | return vm_swappiness; | |
1046 | ||
1047 | spin_lock(&memcg->reclaim_param_lock); | |
1048 | swappiness = memcg->swappiness; | |
1049 | spin_unlock(&memcg->reclaim_param_lock); | |
1050 | ||
1051 | return swappiness; | |
1052 | } | |
1053 | ||
1054 | /* A routine for testing mem is not under move_account */ | |
1055 | ||
1056 | static bool mem_cgroup_under_move(struct mem_cgroup *mem) | |
1057 | { | |
1058 | struct mem_cgroup *from; | |
1059 | struct mem_cgroup *to; | |
1060 | bool ret = false; | |
1061 | /* | |
1062 | * Unlike task_move routines, we access mc.to, mc.from not under | |
1063 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. | |
1064 | */ | |
1065 | spin_lock(&mc.lock); | |
1066 | from = mc.from; | |
1067 | to = mc.to; | |
1068 | if (!from) | |
1069 | goto unlock; | |
1070 | if (from == mem || to == mem | |
1071 | || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css)) | |
1072 | || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css))) | |
1073 | ret = true; | |
1074 | unlock: | |
1075 | spin_unlock(&mc.lock); | |
1076 | return ret; | |
1077 | } | |
1078 | ||
1079 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem) | |
1080 | { | |
1081 | if (mc.moving_task && current != mc.moving_task) { | |
1082 | if (mem_cgroup_under_move(mem)) { | |
1083 | DEFINE_WAIT(wait); | |
1084 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); | |
1085 | /* moving charge context might have finished. */ | |
1086 | if (mc.moving_task) | |
1087 | schedule(); | |
1088 | finish_wait(&mc.waitq, &wait); | |
1089 | return true; | |
1090 | } | |
1091 | } | |
1092 | return false; | |
1093 | } | |
1094 | ||
1095 | static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data) | |
1096 | { | |
1097 | int *val = data; | |
1098 | (*val)++; | |
1099 | return 0; | |
1100 | } | |
1101 | ||
1102 | /** | |
1103 | * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. | |
1104 | * @memcg: The memory cgroup that went over limit | |
1105 | * @p: Task that is going to be killed | |
1106 | * | |
1107 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is | |
1108 | * enabled | |
1109 | */ | |
1110 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) | |
1111 | { | |
1112 | struct cgroup *task_cgrp; | |
1113 | struct cgroup *mem_cgrp; | |
1114 | /* | |
1115 | * Need a buffer in BSS, can't rely on allocations. The code relies | |
1116 | * on the assumption that OOM is serialized for memory controller. | |
1117 | * If this assumption is broken, revisit this code. | |
1118 | */ | |
1119 | static char memcg_name[PATH_MAX]; | |
1120 | int ret; | |
1121 | ||
1122 | if (!memcg || !p) | |
1123 | return; | |
1124 | ||
1125 | ||
1126 | rcu_read_lock(); | |
1127 | ||
1128 | mem_cgrp = memcg->css.cgroup; | |
1129 | task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); | |
1130 | ||
1131 | ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); | |
1132 | if (ret < 0) { | |
1133 | /* | |
1134 | * Unfortunately, we are unable to convert to a useful name | |
1135 | * But we'll still print out the usage information | |
1136 | */ | |
1137 | rcu_read_unlock(); | |
1138 | goto done; | |
1139 | } | |
1140 | rcu_read_unlock(); | |
1141 | ||
1142 | printk(KERN_INFO "Task in %s killed", memcg_name); | |
1143 | ||
1144 | rcu_read_lock(); | |
1145 | ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); | |
1146 | if (ret < 0) { | |
1147 | rcu_read_unlock(); | |
1148 | goto done; | |
1149 | } | |
1150 | rcu_read_unlock(); | |
1151 | ||
1152 | /* | |
1153 | * Continues from above, so we don't need an KERN_ level | |
1154 | */ | |
1155 | printk(KERN_CONT " as a result of limit of %s\n", memcg_name); | |
1156 | done: | |
1157 | ||
1158 | printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", | |
1159 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, | |
1160 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, | |
1161 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); | |
1162 | printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " | |
1163 | "failcnt %llu\n", | |
1164 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, | |
1165 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, | |
1166 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); | |
1167 | } | |
1168 | ||
1169 | /* | |
1170 | * This function returns the number of memcg under hierarchy tree. Returns | |
1171 | * 1(self count) if no children. | |
1172 | */ | |
1173 | static int mem_cgroup_count_children(struct mem_cgroup *mem) | |
1174 | { | |
1175 | int num = 0; | |
1176 | mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb); | |
1177 | return num; | |
1178 | } | |
1179 | ||
1180 | /* | |
1181 | * Return the memory (and swap, if configured) limit for a memcg. | |
1182 | */ | |
1183 | u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) | |
1184 | { | |
1185 | u64 limit; | |
1186 | u64 memsw; | |
1187 | ||
1188 | limit = res_counter_read_u64(&memcg->res, RES_LIMIT) + | |
1189 | total_swap_pages; | |
1190 | memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
1191 | /* | |
1192 | * If memsw is finite and limits the amount of swap space available | |
1193 | * to this memcg, return that limit. | |
1194 | */ | |
1195 | return min(limit, memsw); | |
1196 | } | |
1197 | ||
1198 | /* | |
1199 | * Visit the first child (need not be the first child as per the ordering | |
1200 | * of the cgroup list, since we track last_scanned_child) of @mem and use | |
1201 | * that to reclaim free pages from. | |
1202 | */ | |
1203 | static struct mem_cgroup * | |
1204 | mem_cgroup_select_victim(struct mem_cgroup *root_mem) | |
1205 | { | |
1206 | struct mem_cgroup *ret = NULL; | |
1207 | struct cgroup_subsys_state *css; | |
1208 | int nextid, found; | |
1209 | ||
1210 | if (!root_mem->use_hierarchy) { | |
1211 | css_get(&root_mem->css); | |
1212 | ret = root_mem; | |
1213 | } | |
1214 | ||
1215 | while (!ret) { | |
1216 | rcu_read_lock(); | |
1217 | nextid = root_mem->last_scanned_child + 1; | |
1218 | css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, | |
1219 | &found); | |
1220 | if (css && css_tryget(css)) | |
1221 | ret = container_of(css, struct mem_cgroup, css); | |
1222 | ||
1223 | rcu_read_unlock(); | |
1224 | /* Updates scanning parameter */ | |
1225 | spin_lock(&root_mem->reclaim_param_lock); | |
1226 | if (!css) { | |
1227 | /* this means start scan from ID:1 */ | |
1228 | root_mem->last_scanned_child = 0; | |
1229 | } else | |
1230 | root_mem->last_scanned_child = found; | |
1231 | spin_unlock(&root_mem->reclaim_param_lock); | |
1232 | } | |
1233 | ||
1234 | return ret; | |
1235 | } | |
1236 | ||
1237 | /* | |
1238 | * Scan the hierarchy if needed to reclaim memory. We remember the last child | |
1239 | * we reclaimed from, so that we don't end up penalizing one child extensively | |
1240 | * based on its position in the children list. | |
1241 | * | |
1242 | * root_mem is the original ancestor that we've been reclaim from. | |
1243 | * | |
1244 | * We give up and return to the caller when we visit root_mem twice. | |
1245 | * (other groups can be removed while we're walking....) | |
1246 | * | |
1247 | * If shrink==true, for avoiding to free too much, this returns immedieately. | |
1248 | */ | |
1249 | static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, | |
1250 | struct zone *zone, | |
1251 | gfp_t gfp_mask, | |
1252 | unsigned long reclaim_options) | |
1253 | { | |
1254 | struct mem_cgroup *victim; | |
1255 | int ret, total = 0; | |
1256 | int loop = 0; | |
1257 | bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP; | |
1258 | bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK; | |
1259 | bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT; | |
1260 | unsigned long excess = mem_cgroup_get_excess(root_mem); | |
1261 | ||
1262 | /* If memsw_is_minimum==1, swap-out is of-no-use. */ | |
1263 | if (root_mem->memsw_is_minimum) | |
1264 | noswap = true; | |
1265 | ||
1266 | while (1) { | |
1267 | victim = mem_cgroup_select_victim(root_mem); | |
1268 | if (victim == root_mem) { | |
1269 | loop++; | |
1270 | if (loop >= 1) | |
1271 | drain_all_stock_async(); | |
1272 | if (loop >= 2) { | |
1273 | /* | |
1274 | * If we have not been able to reclaim | |
1275 | * anything, it might because there are | |
1276 | * no reclaimable pages under this hierarchy | |
1277 | */ | |
1278 | if (!check_soft || !total) { | |
1279 | css_put(&victim->css); | |
1280 | break; | |
1281 | } | |
1282 | /* | |
1283 | * We want to do more targetted reclaim. | |
1284 | * excess >> 2 is not to excessive so as to | |
1285 | * reclaim too much, nor too less that we keep | |
1286 | * coming back to reclaim from this cgroup | |
1287 | */ | |
1288 | if (total >= (excess >> 2) || | |
1289 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) { | |
1290 | css_put(&victim->css); | |
1291 | break; | |
1292 | } | |
1293 | } | |
1294 | } | |
1295 | if (!mem_cgroup_local_usage(victim)) { | |
1296 | /* this cgroup's local usage == 0 */ | |
1297 | css_put(&victim->css); | |
1298 | continue; | |
1299 | } | |
1300 | /* we use swappiness of local cgroup */ | |
1301 | if (check_soft) | |
1302 | ret = mem_cgroup_shrink_node_zone(victim, gfp_mask, | |
1303 | noswap, get_swappiness(victim), zone); | |
1304 | else | |
1305 | ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, | |
1306 | noswap, get_swappiness(victim)); | |
1307 | css_put(&victim->css); | |
1308 | /* | |
1309 | * At shrinking usage, we can't check we should stop here or | |
1310 | * reclaim more. It's depends on callers. last_scanned_child | |
1311 | * will work enough for keeping fairness under tree. | |
1312 | */ | |
1313 | if (shrink) | |
1314 | return ret; | |
1315 | total += ret; | |
1316 | if (check_soft) { | |
1317 | if (res_counter_check_under_soft_limit(&root_mem->res)) | |
1318 | return total; | |
1319 | } else if (mem_cgroup_check_under_limit(root_mem)) | |
1320 | return 1 + total; | |
1321 | } | |
1322 | return total; | |
1323 | } | |
1324 | ||
1325 | static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data) | |
1326 | { | |
1327 | int *val = (int *)data; | |
1328 | int x; | |
1329 | /* | |
1330 | * Logically, we can stop scanning immediately when we find | |
1331 | * a memcg is already locked. But condidering unlock ops and | |
1332 | * creation/removal of memcg, scan-all is simple operation. | |
1333 | */ | |
1334 | x = atomic_inc_return(&mem->oom_lock); | |
1335 | *val = max(x, *val); | |
1336 | return 0; | |
1337 | } | |
1338 | /* | |
1339 | * Check OOM-Killer is already running under our hierarchy. | |
1340 | * If someone is running, return false. | |
1341 | */ | |
1342 | static bool mem_cgroup_oom_lock(struct mem_cgroup *mem) | |
1343 | { | |
1344 | int lock_count = 0; | |
1345 | ||
1346 | mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb); | |
1347 | ||
1348 | if (lock_count == 1) | |
1349 | return true; | |
1350 | return false; | |
1351 | } | |
1352 | ||
1353 | static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data) | |
1354 | { | |
1355 | /* | |
1356 | * When a new child is created while the hierarchy is under oom, | |
1357 | * mem_cgroup_oom_lock() may not be called. We have to use | |
1358 | * atomic_add_unless() here. | |
1359 | */ | |
1360 | atomic_add_unless(&mem->oom_lock, -1, 0); | |
1361 | return 0; | |
1362 | } | |
1363 | ||
1364 | static void mem_cgroup_oom_unlock(struct mem_cgroup *mem) | |
1365 | { | |
1366 | mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb); | |
1367 | } | |
1368 | ||
1369 | static DEFINE_MUTEX(memcg_oom_mutex); | |
1370 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); | |
1371 | ||
1372 | struct oom_wait_info { | |
1373 | struct mem_cgroup *mem; | |
1374 | wait_queue_t wait; | |
1375 | }; | |
1376 | ||
1377 | static int memcg_oom_wake_function(wait_queue_t *wait, | |
1378 | unsigned mode, int sync, void *arg) | |
1379 | { | |
1380 | struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg; | |
1381 | struct oom_wait_info *oom_wait_info; | |
1382 | ||
1383 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); | |
1384 | ||
1385 | if (oom_wait_info->mem == wake_mem) | |
1386 | goto wakeup; | |
1387 | /* if no hierarchy, no match */ | |
1388 | if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy) | |
1389 | return 0; | |
1390 | /* | |
1391 | * Both of oom_wait_info->mem and wake_mem are stable under us. | |
1392 | * Then we can use css_is_ancestor without taking care of RCU. | |
1393 | */ | |
1394 | if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) && | |
1395 | !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css)) | |
1396 | return 0; | |
1397 | ||
1398 | wakeup: | |
1399 | return autoremove_wake_function(wait, mode, sync, arg); | |
1400 | } | |
1401 | ||
1402 | static void memcg_wakeup_oom(struct mem_cgroup *mem) | |
1403 | { | |
1404 | /* for filtering, pass "mem" as argument. */ | |
1405 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem); | |
1406 | } | |
1407 | ||
1408 | static void memcg_oom_recover(struct mem_cgroup *mem) | |
1409 | { | |
1410 | if (mem && atomic_read(&mem->oom_lock)) | |
1411 | memcg_wakeup_oom(mem); | |
1412 | } | |
1413 | ||
1414 | /* | |
1415 | * try to call OOM killer. returns false if we should exit memory-reclaim loop. | |
1416 | */ | |
1417 | bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask) | |
1418 | { | |
1419 | struct oom_wait_info owait; | |
1420 | bool locked, need_to_kill; | |
1421 | ||
1422 | owait.mem = mem; | |
1423 | owait.wait.flags = 0; | |
1424 | owait.wait.func = memcg_oom_wake_function; | |
1425 | owait.wait.private = current; | |
1426 | INIT_LIST_HEAD(&owait.wait.task_list); | |
1427 | need_to_kill = true; | |
1428 | /* At first, try to OOM lock hierarchy under mem.*/ | |
1429 | mutex_lock(&memcg_oom_mutex); | |
1430 | locked = mem_cgroup_oom_lock(mem); | |
1431 | /* | |
1432 | * Even if signal_pending(), we can't quit charge() loop without | |
1433 | * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL | |
1434 | * under OOM is always welcomed, use TASK_KILLABLE here. | |
1435 | */ | |
1436 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); | |
1437 | if (!locked || mem->oom_kill_disable) | |
1438 | need_to_kill = false; | |
1439 | if (locked) | |
1440 | mem_cgroup_oom_notify(mem); | |
1441 | mutex_unlock(&memcg_oom_mutex); | |
1442 | ||
1443 | if (need_to_kill) { | |
1444 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1445 | mem_cgroup_out_of_memory(mem, mask); | |
1446 | } else { | |
1447 | schedule(); | |
1448 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1449 | } | |
1450 | mutex_lock(&memcg_oom_mutex); | |
1451 | mem_cgroup_oom_unlock(mem); | |
1452 | memcg_wakeup_oom(mem); | |
1453 | mutex_unlock(&memcg_oom_mutex); | |
1454 | ||
1455 | if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) | |
1456 | return false; | |
1457 | /* Give chance to dying process */ | |
1458 | schedule_timeout(1); | |
1459 | return true; | |
1460 | } | |
1461 | ||
1462 | /* | |
1463 | * Currently used to update mapped file statistics, but the routine can be | |
1464 | * generalized to update other statistics as well. | |
1465 | */ | |
1466 | void mem_cgroup_update_file_mapped(struct page *page, int val) | |
1467 | { | |
1468 | struct mem_cgroup *mem; | |
1469 | struct page_cgroup *pc; | |
1470 | ||
1471 | pc = lookup_page_cgroup(page); | |
1472 | if (unlikely(!pc)) | |
1473 | return; | |
1474 | ||
1475 | lock_page_cgroup(pc); | |
1476 | mem = pc->mem_cgroup; | |
1477 | if (!mem || !PageCgroupUsed(pc)) | |
1478 | goto done; | |
1479 | ||
1480 | /* | |
1481 | * Preemption is already disabled. We can use __this_cpu_xxx | |
1482 | */ | |
1483 | if (val > 0) { | |
1484 | __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
1485 | SetPageCgroupFileMapped(pc); | |
1486 | } else { | |
1487 | __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
1488 | ClearPageCgroupFileMapped(pc); | |
1489 | } | |
1490 | ||
1491 | done: | |
1492 | unlock_page_cgroup(pc); | |
1493 | } | |
1494 | ||
1495 | /* | |
1496 | * size of first charge trial. "32" comes from vmscan.c's magic value. | |
1497 | * TODO: maybe necessary to use big numbers in big irons. | |
1498 | */ | |
1499 | #define CHARGE_SIZE (32 * PAGE_SIZE) | |
1500 | struct memcg_stock_pcp { | |
1501 | struct mem_cgroup *cached; /* this never be root cgroup */ | |
1502 | int charge; | |
1503 | struct work_struct work; | |
1504 | }; | |
1505 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); | |
1506 | static atomic_t memcg_drain_count; | |
1507 | ||
1508 | /* | |
1509 | * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed | |
1510 | * from local stock and true is returned. If the stock is 0 or charges from a | |
1511 | * cgroup which is not current target, returns false. This stock will be | |
1512 | * refilled. | |
1513 | */ | |
1514 | static bool consume_stock(struct mem_cgroup *mem) | |
1515 | { | |
1516 | struct memcg_stock_pcp *stock; | |
1517 | bool ret = true; | |
1518 | ||
1519 | stock = &get_cpu_var(memcg_stock); | |
1520 | if (mem == stock->cached && stock->charge) | |
1521 | stock->charge -= PAGE_SIZE; | |
1522 | else /* need to call res_counter_charge */ | |
1523 | ret = false; | |
1524 | put_cpu_var(memcg_stock); | |
1525 | return ret; | |
1526 | } | |
1527 | ||
1528 | /* | |
1529 | * Returns stocks cached in percpu to res_counter and reset cached information. | |
1530 | */ | |
1531 | static void drain_stock(struct memcg_stock_pcp *stock) | |
1532 | { | |
1533 | struct mem_cgroup *old = stock->cached; | |
1534 | ||
1535 | if (stock->charge) { | |
1536 | res_counter_uncharge(&old->res, stock->charge); | |
1537 | if (do_swap_account) | |
1538 | res_counter_uncharge(&old->memsw, stock->charge); | |
1539 | } | |
1540 | stock->cached = NULL; | |
1541 | stock->charge = 0; | |
1542 | } | |
1543 | ||
1544 | /* | |
1545 | * This must be called under preempt disabled or must be called by | |
1546 | * a thread which is pinned to local cpu. | |
1547 | */ | |
1548 | static void drain_local_stock(struct work_struct *dummy) | |
1549 | { | |
1550 | struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); | |
1551 | drain_stock(stock); | |
1552 | } | |
1553 | ||
1554 | /* | |
1555 | * Cache charges(val) which is from res_counter, to local per_cpu area. | |
1556 | * This will be consumed by consume_stock() function, later. | |
1557 | */ | |
1558 | static void refill_stock(struct mem_cgroup *mem, int val) | |
1559 | { | |
1560 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); | |
1561 | ||
1562 | if (stock->cached != mem) { /* reset if necessary */ | |
1563 | drain_stock(stock); | |
1564 | stock->cached = mem; | |
1565 | } | |
1566 | stock->charge += val; | |
1567 | put_cpu_var(memcg_stock); | |
1568 | } | |
1569 | ||
1570 | /* | |
1571 | * Tries to drain stocked charges in other cpus. This function is asynchronous | |
1572 | * and just put a work per cpu for draining localy on each cpu. Caller can | |
1573 | * expects some charges will be back to res_counter later but cannot wait for | |
1574 | * it. | |
1575 | */ | |
1576 | static void drain_all_stock_async(void) | |
1577 | { | |
1578 | int cpu; | |
1579 | /* This function is for scheduling "drain" in asynchronous way. | |
1580 | * The result of "drain" is not directly handled by callers. Then, | |
1581 | * if someone is calling drain, we don't have to call drain more. | |
1582 | * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if | |
1583 | * there is a race. We just do loose check here. | |
1584 | */ | |
1585 | if (atomic_read(&memcg_drain_count)) | |
1586 | return; | |
1587 | /* Notify other cpus that system-wide "drain" is running */ | |
1588 | atomic_inc(&memcg_drain_count); | |
1589 | get_online_cpus(); | |
1590 | for_each_online_cpu(cpu) { | |
1591 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
1592 | schedule_work_on(cpu, &stock->work); | |
1593 | } | |
1594 | put_online_cpus(); | |
1595 | atomic_dec(&memcg_drain_count); | |
1596 | /* We don't wait for flush_work */ | |
1597 | } | |
1598 | ||
1599 | /* This is a synchronous drain interface. */ | |
1600 | static void drain_all_stock_sync(void) | |
1601 | { | |
1602 | /* called when force_empty is called */ | |
1603 | atomic_inc(&memcg_drain_count); | |
1604 | schedule_on_each_cpu(drain_local_stock); | |
1605 | atomic_dec(&memcg_drain_count); | |
1606 | } | |
1607 | ||
1608 | static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb, | |
1609 | unsigned long action, | |
1610 | void *hcpu) | |
1611 | { | |
1612 | int cpu = (unsigned long)hcpu; | |
1613 | struct memcg_stock_pcp *stock; | |
1614 | ||
1615 | if (action != CPU_DEAD) | |
1616 | return NOTIFY_OK; | |
1617 | stock = &per_cpu(memcg_stock, cpu); | |
1618 | drain_stock(stock); | |
1619 | return NOTIFY_OK; | |
1620 | } | |
1621 | ||
1622 | ||
1623 | /* See __mem_cgroup_try_charge() for details */ | |
1624 | enum { | |
1625 | CHARGE_OK, /* success */ | |
1626 | CHARGE_RETRY, /* need to retry but retry is not bad */ | |
1627 | CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ | |
1628 | CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ | |
1629 | CHARGE_OOM_DIE, /* the current is killed because of OOM */ | |
1630 | }; | |
1631 | ||
1632 | static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask, | |
1633 | int csize, bool oom_check) | |
1634 | { | |
1635 | struct mem_cgroup *mem_over_limit; | |
1636 | struct res_counter *fail_res; | |
1637 | unsigned long flags = 0; | |
1638 | int ret; | |
1639 | ||
1640 | ret = res_counter_charge(&mem->res, csize, &fail_res); | |
1641 | ||
1642 | if (likely(!ret)) { | |
1643 | if (!do_swap_account) | |
1644 | return CHARGE_OK; | |
1645 | ret = res_counter_charge(&mem->memsw, csize, &fail_res); | |
1646 | if (likely(!ret)) | |
1647 | return CHARGE_OK; | |
1648 | ||
1649 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); | |
1650 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; | |
1651 | } else | |
1652 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); | |
1653 | ||
1654 | if (csize > PAGE_SIZE) /* change csize and retry */ | |
1655 | return CHARGE_RETRY; | |
1656 | ||
1657 | if (!(gfp_mask & __GFP_WAIT)) | |
1658 | return CHARGE_WOULDBLOCK; | |
1659 | ||
1660 | ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL, | |
1661 | gfp_mask, flags); | |
1662 | /* | |
1663 | * try_to_free_mem_cgroup_pages() might not give us a full | |
1664 | * picture of reclaim. Some pages are reclaimed and might be | |
1665 | * moved to swap cache or just unmapped from the cgroup. | |
1666 | * Check the limit again to see if the reclaim reduced the | |
1667 | * current usage of the cgroup before giving up | |
1668 | */ | |
1669 | if (ret || mem_cgroup_check_under_limit(mem_over_limit)) | |
1670 | return CHARGE_RETRY; | |
1671 | ||
1672 | /* | |
1673 | * At task move, charge accounts can be doubly counted. So, it's | |
1674 | * better to wait until the end of task_move if something is going on. | |
1675 | */ | |
1676 | if (mem_cgroup_wait_acct_move(mem_over_limit)) | |
1677 | return CHARGE_RETRY; | |
1678 | ||
1679 | /* If we don't need to call oom-killer at el, return immediately */ | |
1680 | if (!oom_check) | |
1681 | return CHARGE_NOMEM; | |
1682 | /* check OOM */ | |
1683 | if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) | |
1684 | return CHARGE_OOM_DIE; | |
1685 | ||
1686 | return CHARGE_RETRY; | |
1687 | } | |
1688 | ||
1689 | /* | |
1690 | * Unlike exported interface, "oom" parameter is added. if oom==true, | |
1691 | * oom-killer can be invoked. | |
1692 | */ | |
1693 | static int __mem_cgroup_try_charge(struct mm_struct *mm, | |
1694 | gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom) | |
1695 | { | |
1696 | int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
1697 | struct mem_cgroup *mem = NULL; | |
1698 | int ret; | |
1699 | int csize = CHARGE_SIZE; | |
1700 | ||
1701 | /* | |
1702 | * Unlike gloval-vm's OOM-kill, we're not in memory shortage | |
1703 | * in system level. So, allow to go ahead dying process in addition to | |
1704 | * MEMDIE process. | |
1705 | */ | |
1706 | if (unlikely(test_thread_flag(TIF_MEMDIE) | |
1707 | || fatal_signal_pending(current))) | |
1708 | goto bypass; | |
1709 | ||
1710 | /* | |
1711 | * We always charge the cgroup the mm_struct belongs to. | |
1712 | * The mm_struct's mem_cgroup changes on task migration if the | |
1713 | * thread group leader migrates. It's possible that mm is not | |
1714 | * set, if so charge the init_mm (happens for pagecache usage). | |
1715 | */ | |
1716 | if (!*memcg && !mm) | |
1717 | goto bypass; | |
1718 | again: | |
1719 | if (*memcg) { /* css should be a valid one */ | |
1720 | mem = *memcg; | |
1721 | VM_BUG_ON(css_is_removed(&mem->css)); | |
1722 | if (mem_cgroup_is_root(mem)) | |
1723 | goto done; | |
1724 | if (consume_stock(mem)) | |
1725 | goto done; | |
1726 | css_get(&mem->css); | |
1727 | } else { | |
1728 | struct task_struct *p; | |
1729 | ||
1730 | rcu_read_lock(); | |
1731 | p = rcu_dereference(mm->owner); | |
1732 | VM_BUG_ON(!p); | |
1733 | /* | |
1734 | * because we don't have task_lock(), "p" can exit while | |
1735 | * we're here. In that case, "mem" can point to root | |
1736 | * cgroup but never be NULL. (and task_struct itself is freed | |
1737 | * by RCU, cgroup itself is RCU safe.) Then, we have small | |
1738 | * risk here to get wrong cgroup. But such kind of mis-account | |
1739 | * by race always happens because we don't have cgroup_mutex(). | |
1740 | * It's overkill and we allow that small race, here. | |
1741 | */ | |
1742 | mem = mem_cgroup_from_task(p); | |
1743 | VM_BUG_ON(!mem); | |
1744 | if (mem_cgroup_is_root(mem)) { | |
1745 | rcu_read_unlock(); | |
1746 | goto done; | |
1747 | } | |
1748 | if (consume_stock(mem)) { | |
1749 | /* | |
1750 | * It seems dagerous to access memcg without css_get(). | |
1751 | * But considering how consume_stok works, it's not | |
1752 | * necessary. If consume_stock success, some charges | |
1753 | * from this memcg are cached on this cpu. So, we | |
1754 | * don't need to call css_get()/css_tryget() before | |
1755 | * calling consume_stock(). | |
1756 | */ | |
1757 | rcu_read_unlock(); | |
1758 | goto done; | |
1759 | } | |
1760 | /* after here, we may be blocked. we need to get refcnt */ | |
1761 | if (!css_tryget(&mem->css)) { | |
1762 | rcu_read_unlock(); | |
1763 | goto again; | |
1764 | } | |
1765 | rcu_read_unlock(); | |
1766 | } | |
1767 | ||
1768 | do { | |
1769 | bool oom_check; | |
1770 | ||
1771 | /* If killed, bypass charge */ | |
1772 | if (fatal_signal_pending(current)) { | |
1773 | css_put(&mem->css); | |
1774 | goto bypass; | |
1775 | } | |
1776 | ||
1777 | oom_check = false; | |
1778 | if (oom && !nr_oom_retries) { | |
1779 | oom_check = true; | |
1780 | nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
1781 | } | |
1782 | ||
1783 | ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check); | |
1784 | ||
1785 | switch (ret) { | |
1786 | case CHARGE_OK: | |
1787 | break; | |
1788 | case CHARGE_RETRY: /* not in OOM situation but retry */ | |
1789 | csize = PAGE_SIZE; | |
1790 | css_put(&mem->css); | |
1791 | mem = NULL; | |
1792 | goto again; | |
1793 | case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ | |
1794 | css_put(&mem->css); | |
1795 | goto nomem; | |
1796 | case CHARGE_NOMEM: /* OOM routine works */ | |
1797 | if (!oom) { | |
1798 | css_put(&mem->css); | |
1799 | goto nomem; | |
1800 | } | |
1801 | /* If oom, we never return -ENOMEM */ | |
1802 | nr_oom_retries--; | |
1803 | break; | |
1804 | case CHARGE_OOM_DIE: /* Killed by OOM Killer */ | |
1805 | css_put(&mem->css); | |
1806 | goto bypass; | |
1807 | } | |
1808 | } while (ret != CHARGE_OK); | |
1809 | ||
1810 | if (csize > PAGE_SIZE) | |
1811 | refill_stock(mem, csize - PAGE_SIZE); | |
1812 | css_put(&mem->css); | |
1813 | done: | |
1814 | *memcg = mem; | |
1815 | return 0; | |
1816 | nomem: | |
1817 | *memcg = NULL; | |
1818 | return -ENOMEM; | |
1819 | bypass: | |
1820 | *memcg = NULL; | |
1821 | return 0; | |
1822 | } | |
1823 | ||
1824 | /* | |
1825 | * Somemtimes we have to undo a charge we got by try_charge(). | |
1826 | * This function is for that and do uncharge, put css's refcnt. | |
1827 | * gotten by try_charge(). | |
1828 | */ | |
1829 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem, | |
1830 | unsigned long count) | |
1831 | { | |
1832 | if (!mem_cgroup_is_root(mem)) { | |
1833 | res_counter_uncharge(&mem->res, PAGE_SIZE * count); | |
1834 | if (do_swap_account) | |
1835 | res_counter_uncharge(&mem->memsw, PAGE_SIZE * count); | |
1836 | } | |
1837 | } | |
1838 | ||
1839 | static void mem_cgroup_cancel_charge(struct mem_cgroup *mem) | |
1840 | { | |
1841 | __mem_cgroup_cancel_charge(mem, 1); | |
1842 | } | |
1843 | ||
1844 | /* | |
1845 | * A helper function to get mem_cgroup from ID. must be called under | |
1846 | * rcu_read_lock(). The caller must check css_is_removed() or some if | |
1847 | * it's concern. (dropping refcnt from swap can be called against removed | |
1848 | * memcg.) | |
1849 | */ | |
1850 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) | |
1851 | { | |
1852 | struct cgroup_subsys_state *css; | |
1853 | ||
1854 | /* ID 0 is unused ID */ | |
1855 | if (!id) | |
1856 | return NULL; | |
1857 | css = css_lookup(&mem_cgroup_subsys, id); | |
1858 | if (!css) | |
1859 | return NULL; | |
1860 | return container_of(css, struct mem_cgroup, css); | |
1861 | } | |
1862 | ||
1863 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) | |
1864 | { | |
1865 | struct mem_cgroup *mem = NULL; | |
1866 | struct page_cgroup *pc; | |
1867 | unsigned short id; | |
1868 | swp_entry_t ent; | |
1869 | ||
1870 | VM_BUG_ON(!PageLocked(page)); | |
1871 | ||
1872 | pc = lookup_page_cgroup(page); | |
1873 | lock_page_cgroup(pc); | |
1874 | if (PageCgroupUsed(pc)) { | |
1875 | mem = pc->mem_cgroup; | |
1876 | if (mem && !css_tryget(&mem->css)) | |
1877 | mem = NULL; | |
1878 | } else if (PageSwapCache(page)) { | |
1879 | ent.val = page_private(page); | |
1880 | id = lookup_swap_cgroup(ent); | |
1881 | rcu_read_lock(); | |
1882 | mem = mem_cgroup_lookup(id); | |
1883 | if (mem && !css_tryget(&mem->css)) | |
1884 | mem = NULL; | |
1885 | rcu_read_unlock(); | |
1886 | } | |
1887 | unlock_page_cgroup(pc); | |
1888 | return mem; | |
1889 | } | |
1890 | ||
1891 | /* | |
1892 | * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be | |
1893 | * USED state. If already USED, uncharge and return. | |
1894 | */ | |
1895 | ||
1896 | static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, | |
1897 | struct page_cgroup *pc, | |
1898 | enum charge_type ctype) | |
1899 | { | |
1900 | /* try_charge() can return NULL to *memcg, taking care of it. */ | |
1901 | if (!mem) | |
1902 | return; | |
1903 | ||
1904 | lock_page_cgroup(pc); | |
1905 | if (unlikely(PageCgroupUsed(pc))) { | |
1906 | unlock_page_cgroup(pc); | |
1907 | mem_cgroup_cancel_charge(mem); | |
1908 | return; | |
1909 | } | |
1910 | ||
1911 | pc->mem_cgroup = mem; | |
1912 | /* | |
1913 | * We access a page_cgroup asynchronously without lock_page_cgroup(). | |
1914 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup | |
1915 | * is accessed after testing USED bit. To make pc->mem_cgroup visible | |
1916 | * before USED bit, we need memory barrier here. | |
1917 | * See mem_cgroup_add_lru_list(), etc. | |
1918 | */ | |
1919 | smp_wmb(); | |
1920 | switch (ctype) { | |
1921 | case MEM_CGROUP_CHARGE_TYPE_CACHE: | |
1922 | case MEM_CGROUP_CHARGE_TYPE_SHMEM: | |
1923 | SetPageCgroupCache(pc); | |
1924 | SetPageCgroupUsed(pc); | |
1925 | break; | |
1926 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: | |
1927 | ClearPageCgroupCache(pc); | |
1928 | SetPageCgroupUsed(pc); | |
1929 | break; | |
1930 | default: | |
1931 | break; | |
1932 | } | |
1933 | ||
1934 | mem_cgroup_charge_statistics(mem, pc, true); | |
1935 | ||
1936 | unlock_page_cgroup(pc); | |
1937 | /* | |
1938 | * "charge_statistics" updated event counter. Then, check it. | |
1939 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. | |
1940 | * if they exceeds softlimit. | |
1941 | */ | |
1942 | memcg_check_events(mem, pc->page); | |
1943 | } | |
1944 | ||
1945 | /** | |
1946 | * __mem_cgroup_move_account - move account of the page | |
1947 | * @pc: page_cgroup of the page. | |
1948 | * @from: mem_cgroup which the page is moved from. | |
1949 | * @to: mem_cgroup which the page is moved to. @from != @to. | |
1950 | * @uncharge: whether we should call uncharge and css_put against @from. | |
1951 | * | |
1952 | * The caller must confirm following. | |
1953 | * - page is not on LRU (isolate_page() is useful.) | |
1954 | * - the pc is locked, used, and ->mem_cgroup points to @from. | |
1955 | * | |
1956 | * This function doesn't do "charge" nor css_get to new cgroup. It should be | |
1957 | * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is | |
1958 | * true, this function does "uncharge" from old cgroup, but it doesn't if | |
1959 | * @uncharge is false, so a caller should do "uncharge". | |
1960 | */ | |
1961 | ||
1962 | static void __mem_cgroup_move_account(struct page_cgroup *pc, | |
1963 | struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge) | |
1964 | { | |
1965 | VM_BUG_ON(from == to); | |
1966 | VM_BUG_ON(PageLRU(pc->page)); | |
1967 | VM_BUG_ON(!PageCgroupLocked(pc)); | |
1968 | VM_BUG_ON(!PageCgroupUsed(pc)); | |
1969 | VM_BUG_ON(pc->mem_cgroup != from); | |
1970 | ||
1971 | if (PageCgroupFileMapped(pc)) { | |
1972 | /* Update mapped_file data for mem_cgroup */ | |
1973 | preempt_disable(); | |
1974 | __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
1975 | __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
1976 | preempt_enable(); | |
1977 | } | |
1978 | mem_cgroup_charge_statistics(from, pc, false); | |
1979 | if (uncharge) | |
1980 | /* This is not "cancel", but cancel_charge does all we need. */ | |
1981 | mem_cgroup_cancel_charge(from); | |
1982 | ||
1983 | /* caller should have done css_get */ | |
1984 | pc->mem_cgroup = to; | |
1985 | mem_cgroup_charge_statistics(to, pc, true); | |
1986 | /* | |
1987 | * We charges against "to" which may not have any tasks. Then, "to" | |
1988 | * can be under rmdir(). But in current implementation, caller of | |
1989 | * this function is just force_empty() and move charge, so it's | |
1990 | * garanteed that "to" is never removed. So, we don't check rmdir | |
1991 | * status here. | |
1992 | */ | |
1993 | } | |
1994 | ||
1995 | /* | |
1996 | * check whether the @pc is valid for moving account and call | |
1997 | * __mem_cgroup_move_account() | |
1998 | */ | |
1999 | static int mem_cgroup_move_account(struct page_cgroup *pc, | |
2000 | struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge) | |
2001 | { | |
2002 | int ret = -EINVAL; | |
2003 | lock_page_cgroup(pc); | |
2004 | if (PageCgroupUsed(pc) && pc->mem_cgroup == from) { | |
2005 | __mem_cgroup_move_account(pc, from, to, uncharge); | |
2006 | ret = 0; | |
2007 | } | |
2008 | unlock_page_cgroup(pc); | |
2009 | /* | |
2010 | * check events | |
2011 | */ | |
2012 | memcg_check_events(to, pc->page); | |
2013 | memcg_check_events(from, pc->page); | |
2014 | return ret; | |
2015 | } | |
2016 | ||
2017 | /* | |
2018 | * move charges to its parent. | |
2019 | */ | |
2020 | ||
2021 | static int mem_cgroup_move_parent(struct page_cgroup *pc, | |
2022 | struct mem_cgroup *child, | |
2023 | gfp_t gfp_mask) | |
2024 | { | |
2025 | struct page *page = pc->page; | |
2026 | struct cgroup *cg = child->css.cgroup; | |
2027 | struct cgroup *pcg = cg->parent; | |
2028 | struct mem_cgroup *parent; | |
2029 | int ret; | |
2030 | ||
2031 | /* Is ROOT ? */ | |
2032 | if (!pcg) | |
2033 | return -EINVAL; | |
2034 | ||
2035 | ret = -EBUSY; | |
2036 | if (!get_page_unless_zero(page)) | |
2037 | goto out; | |
2038 | if (isolate_lru_page(page)) | |
2039 | goto put; | |
2040 | ||
2041 | parent = mem_cgroup_from_cont(pcg); | |
2042 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false); | |
2043 | if (ret || !parent) | |
2044 | goto put_back; | |
2045 | ||
2046 | ret = mem_cgroup_move_account(pc, child, parent, true); | |
2047 | if (ret) | |
2048 | mem_cgroup_cancel_charge(parent); | |
2049 | put_back: | |
2050 | putback_lru_page(page); | |
2051 | put: | |
2052 | put_page(page); | |
2053 | out: | |
2054 | return ret; | |
2055 | } | |
2056 | ||
2057 | /* | |
2058 | * Charge the memory controller for page usage. | |
2059 | * Return | |
2060 | * 0 if the charge was successful | |
2061 | * < 0 if the cgroup is over its limit | |
2062 | */ | |
2063 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, | |
2064 | gfp_t gfp_mask, enum charge_type ctype) | |
2065 | { | |
2066 | struct mem_cgroup *mem = NULL; | |
2067 | struct page_cgroup *pc; | |
2068 | int ret; | |
2069 | ||
2070 | pc = lookup_page_cgroup(page); | |
2071 | /* can happen at boot */ | |
2072 | if (unlikely(!pc)) | |
2073 | return 0; | |
2074 | prefetchw(pc); | |
2075 | ||
2076 | ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true); | |
2077 | if (ret || !mem) | |
2078 | return ret; | |
2079 | ||
2080 | __mem_cgroup_commit_charge(mem, pc, ctype); | |
2081 | return 0; | |
2082 | } | |
2083 | ||
2084 | int mem_cgroup_newpage_charge(struct page *page, | |
2085 | struct mm_struct *mm, gfp_t gfp_mask) | |
2086 | { | |
2087 | if (mem_cgroup_disabled()) | |
2088 | return 0; | |
2089 | if (PageCompound(page)) | |
2090 | return 0; | |
2091 | /* | |
2092 | * If already mapped, we don't have to account. | |
2093 | * If page cache, page->mapping has address_space. | |
2094 | * But page->mapping may have out-of-use anon_vma pointer, | |
2095 | * detecit it by PageAnon() check. newly-mapped-anon's page->mapping | |
2096 | * is NULL. | |
2097 | */ | |
2098 | if (page_mapped(page) || (page->mapping && !PageAnon(page))) | |
2099 | return 0; | |
2100 | if (unlikely(!mm)) | |
2101 | mm = &init_mm; | |
2102 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
2103 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2104 | } | |
2105 | ||
2106 | static void | |
2107 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, | |
2108 | enum charge_type ctype); | |
2109 | ||
2110 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, | |
2111 | gfp_t gfp_mask) | |
2112 | { | |
2113 | int ret; | |
2114 | ||
2115 | if (mem_cgroup_disabled()) | |
2116 | return 0; | |
2117 | if (PageCompound(page)) | |
2118 | return 0; | |
2119 | /* | |
2120 | * Corner case handling. This is called from add_to_page_cache() | |
2121 | * in usual. But some FS (shmem) precharges this page before calling it | |
2122 | * and call add_to_page_cache() with GFP_NOWAIT. | |
2123 | * | |
2124 | * For GFP_NOWAIT case, the page may be pre-charged before calling | |
2125 | * add_to_page_cache(). (See shmem.c) check it here and avoid to call | |
2126 | * charge twice. (It works but has to pay a bit larger cost.) | |
2127 | * And when the page is SwapCache, it should take swap information | |
2128 | * into account. This is under lock_page() now. | |
2129 | */ | |
2130 | if (!(gfp_mask & __GFP_WAIT)) { | |
2131 | struct page_cgroup *pc; | |
2132 | ||
2133 | pc = lookup_page_cgroup(page); | |
2134 | if (!pc) | |
2135 | return 0; | |
2136 | lock_page_cgroup(pc); | |
2137 | if (PageCgroupUsed(pc)) { | |
2138 | unlock_page_cgroup(pc); | |
2139 | return 0; | |
2140 | } | |
2141 | unlock_page_cgroup(pc); | |
2142 | } | |
2143 | ||
2144 | if (unlikely(!mm)) | |
2145 | mm = &init_mm; | |
2146 | ||
2147 | if (page_is_file_cache(page)) | |
2148 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
2149 | MEM_CGROUP_CHARGE_TYPE_CACHE); | |
2150 | ||
2151 | /* shmem */ | |
2152 | if (PageSwapCache(page)) { | |
2153 | struct mem_cgroup *mem = NULL; | |
2154 | ||
2155 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); | |
2156 | if (!ret) | |
2157 | __mem_cgroup_commit_charge_swapin(page, mem, | |
2158 | MEM_CGROUP_CHARGE_TYPE_SHMEM); | |
2159 | } else | |
2160 | ret = mem_cgroup_charge_common(page, mm, gfp_mask, | |
2161 | MEM_CGROUP_CHARGE_TYPE_SHMEM); | |
2162 | ||
2163 | return ret; | |
2164 | } | |
2165 | ||
2166 | /* | |
2167 | * While swap-in, try_charge -> commit or cancel, the page is locked. | |
2168 | * And when try_charge() successfully returns, one refcnt to memcg without | |
2169 | * struct page_cgroup is acquired. This refcnt will be consumed by | |
2170 | * "commit()" or removed by "cancel()" | |
2171 | */ | |
2172 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, | |
2173 | struct page *page, | |
2174 | gfp_t mask, struct mem_cgroup **ptr) | |
2175 | { | |
2176 | struct mem_cgroup *mem; | |
2177 | int ret; | |
2178 | ||
2179 | if (mem_cgroup_disabled()) | |
2180 | return 0; | |
2181 | ||
2182 | if (!do_swap_account) | |
2183 | goto charge_cur_mm; | |
2184 | /* | |
2185 | * A racing thread's fault, or swapoff, may have already updated | |
2186 | * the pte, and even removed page from swap cache: in those cases | |
2187 | * do_swap_page()'s pte_same() test will fail; but there's also a | |
2188 | * KSM case which does need to charge the page. | |
2189 | */ | |
2190 | if (!PageSwapCache(page)) | |
2191 | goto charge_cur_mm; | |
2192 | mem = try_get_mem_cgroup_from_page(page); | |
2193 | if (!mem) | |
2194 | goto charge_cur_mm; | |
2195 | *ptr = mem; | |
2196 | ret = __mem_cgroup_try_charge(NULL, mask, ptr, true); | |
2197 | css_put(&mem->css); | |
2198 | return ret; | |
2199 | charge_cur_mm: | |
2200 | if (unlikely(!mm)) | |
2201 | mm = &init_mm; | |
2202 | return __mem_cgroup_try_charge(mm, mask, ptr, true); | |
2203 | } | |
2204 | ||
2205 | static void | |
2206 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, | |
2207 | enum charge_type ctype) | |
2208 | { | |
2209 | struct page_cgroup *pc; | |
2210 | ||
2211 | if (mem_cgroup_disabled()) | |
2212 | return; | |
2213 | if (!ptr) | |
2214 | return; | |
2215 | cgroup_exclude_rmdir(&ptr->css); | |
2216 | pc = lookup_page_cgroup(page); | |
2217 | mem_cgroup_lru_del_before_commit_swapcache(page); | |
2218 | __mem_cgroup_commit_charge(ptr, pc, ctype); | |
2219 | mem_cgroup_lru_add_after_commit_swapcache(page); | |
2220 | /* | |
2221 | * Now swap is on-memory. This means this page may be | |
2222 | * counted both as mem and swap....double count. | |
2223 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable | |
2224 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() | |
2225 | * may call delete_from_swap_cache() before reach here. | |
2226 | */ | |
2227 | if (do_swap_account && PageSwapCache(page)) { | |
2228 | swp_entry_t ent = {.val = page_private(page)}; | |
2229 | unsigned short id; | |
2230 | struct mem_cgroup *memcg; | |
2231 | ||
2232 | id = swap_cgroup_record(ent, 0); | |
2233 | rcu_read_lock(); | |
2234 | memcg = mem_cgroup_lookup(id); | |
2235 | if (memcg) { | |
2236 | /* | |
2237 | * This recorded memcg can be obsolete one. So, avoid | |
2238 | * calling css_tryget | |
2239 | */ | |
2240 | if (!mem_cgroup_is_root(memcg)) | |
2241 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); | |
2242 | mem_cgroup_swap_statistics(memcg, false); | |
2243 | mem_cgroup_put(memcg); | |
2244 | } | |
2245 | rcu_read_unlock(); | |
2246 | } | |
2247 | /* | |
2248 | * At swapin, we may charge account against cgroup which has no tasks. | |
2249 | * So, rmdir()->pre_destroy() can be called while we do this charge. | |
2250 | * In that case, we need to call pre_destroy() again. check it here. | |
2251 | */ | |
2252 | cgroup_release_and_wakeup_rmdir(&ptr->css); | |
2253 | } | |
2254 | ||
2255 | void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) | |
2256 | { | |
2257 | __mem_cgroup_commit_charge_swapin(page, ptr, | |
2258 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2259 | } | |
2260 | ||
2261 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) | |
2262 | { | |
2263 | if (mem_cgroup_disabled()) | |
2264 | return; | |
2265 | if (!mem) | |
2266 | return; | |
2267 | mem_cgroup_cancel_charge(mem); | |
2268 | } | |
2269 | ||
2270 | static void | |
2271 | __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype) | |
2272 | { | |
2273 | struct memcg_batch_info *batch = NULL; | |
2274 | bool uncharge_memsw = true; | |
2275 | /* If swapout, usage of swap doesn't decrease */ | |
2276 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) | |
2277 | uncharge_memsw = false; | |
2278 | ||
2279 | batch = ¤t->memcg_batch; | |
2280 | /* | |
2281 | * In usual, we do css_get() when we remember memcg pointer. | |
2282 | * But in this case, we keep res->usage until end of a series of | |
2283 | * uncharges. Then, it's ok to ignore memcg's refcnt. | |
2284 | */ | |
2285 | if (!batch->memcg) | |
2286 | batch->memcg = mem; | |
2287 | /* | |
2288 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. | |
2289 | * In those cases, all pages freed continously can be expected to be in | |
2290 | * the same cgroup and we have chance to coalesce uncharges. | |
2291 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) | |
2292 | * because we want to do uncharge as soon as possible. | |
2293 | */ | |
2294 | ||
2295 | if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) | |
2296 | goto direct_uncharge; | |
2297 | ||
2298 | /* | |
2299 | * In typical case, batch->memcg == mem. This means we can | |
2300 | * merge a series of uncharges to an uncharge of res_counter. | |
2301 | * If not, we uncharge res_counter ony by one. | |
2302 | */ | |
2303 | if (batch->memcg != mem) | |
2304 | goto direct_uncharge; | |
2305 | /* remember freed charge and uncharge it later */ | |
2306 | batch->bytes += PAGE_SIZE; | |
2307 | if (uncharge_memsw) | |
2308 | batch->memsw_bytes += PAGE_SIZE; | |
2309 | return; | |
2310 | direct_uncharge: | |
2311 | res_counter_uncharge(&mem->res, PAGE_SIZE); | |
2312 | if (uncharge_memsw) | |
2313 | res_counter_uncharge(&mem->memsw, PAGE_SIZE); | |
2314 | if (unlikely(batch->memcg != mem)) | |
2315 | memcg_oom_recover(mem); | |
2316 | return; | |
2317 | } | |
2318 | ||
2319 | /* | |
2320 | * uncharge if !page_mapped(page) | |
2321 | */ | |
2322 | static struct mem_cgroup * | |
2323 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) | |
2324 | { | |
2325 | struct page_cgroup *pc; | |
2326 | struct mem_cgroup *mem = NULL; | |
2327 | ||
2328 | if (mem_cgroup_disabled()) | |
2329 | return NULL; | |
2330 | ||
2331 | if (PageSwapCache(page)) | |
2332 | return NULL; | |
2333 | ||
2334 | /* | |
2335 | * Check if our page_cgroup is valid | |
2336 | */ | |
2337 | pc = lookup_page_cgroup(page); | |
2338 | if (unlikely(!pc || !PageCgroupUsed(pc))) | |
2339 | return NULL; | |
2340 | ||
2341 | lock_page_cgroup(pc); | |
2342 | ||
2343 | mem = pc->mem_cgroup; | |
2344 | ||
2345 | if (!PageCgroupUsed(pc)) | |
2346 | goto unlock_out; | |
2347 | ||
2348 | switch (ctype) { | |
2349 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: | |
2350 | case MEM_CGROUP_CHARGE_TYPE_DROP: | |
2351 | /* See mem_cgroup_prepare_migration() */ | |
2352 | if (page_mapped(page) || PageCgroupMigration(pc)) | |
2353 | goto unlock_out; | |
2354 | break; | |
2355 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: | |
2356 | if (!PageAnon(page)) { /* Shared memory */ | |
2357 | if (page->mapping && !page_is_file_cache(page)) | |
2358 | goto unlock_out; | |
2359 | } else if (page_mapped(page)) /* Anon */ | |
2360 | goto unlock_out; | |
2361 | break; | |
2362 | default: | |
2363 | break; | |
2364 | } | |
2365 | ||
2366 | mem_cgroup_charge_statistics(mem, pc, false); | |
2367 | ||
2368 | ClearPageCgroupUsed(pc); | |
2369 | /* | |
2370 | * pc->mem_cgroup is not cleared here. It will be accessed when it's | |
2371 | * freed from LRU. This is safe because uncharged page is expected not | |
2372 | * to be reused (freed soon). Exception is SwapCache, it's handled by | |
2373 | * special functions. | |
2374 | */ | |
2375 | ||
2376 | unlock_page_cgroup(pc); | |
2377 | /* | |
2378 | * even after unlock, we have mem->res.usage here and this memcg | |
2379 | * will never be freed. | |
2380 | */ | |
2381 | memcg_check_events(mem, page); | |
2382 | if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { | |
2383 | mem_cgroup_swap_statistics(mem, true); | |
2384 | mem_cgroup_get(mem); | |
2385 | } | |
2386 | if (!mem_cgroup_is_root(mem)) | |
2387 | __do_uncharge(mem, ctype); | |
2388 | ||
2389 | return mem; | |
2390 | ||
2391 | unlock_out: | |
2392 | unlock_page_cgroup(pc); | |
2393 | return NULL; | |
2394 | } | |
2395 | ||
2396 | void mem_cgroup_uncharge_page(struct page *page) | |
2397 | { | |
2398 | /* early check. */ | |
2399 | if (page_mapped(page)) | |
2400 | return; | |
2401 | if (page->mapping && !PageAnon(page)) | |
2402 | return; | |
2403 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2404 | } | |
2405 | ||
2406 | void mem_cgroup_uncharge_cache_page(struct page *page) | |
2407 | { | |
2408 | VM_BUG_ON(page_mapped(page)); | |
2409 | VM_BUG_ON(page->mapping); | |
2410 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); | |
2411 | } | |
2412 | ||
2413 | /* | |
2414 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. | |
2415 | * In that cases, pages are freed continuously and we can expect pages | |
2416 | * are in the same memcg. All these calls itself limits the number of | |
2417 | * pages freed at once, then uncharge_start/end() is called properly. | |
2418 | * This may be called prural(2) times in a context, | |
2419 | */ | |
2420 | ||
2421 | void mem_cgroup_uncharge_start(void) | |
2422 | { | |
2423 | current->memcg_batch.do_batch++; | |
2424 | /* We can do nest. */ | |
2425 | if (current->memcg_batch.do_batch == 1) { | |
2426 | current->memcg_batch.memcg = NULL; | |
2427 | current->memcg_batch.bytes = 0; | |
2428 | current->memcg_batch.memsw_bytes = 0; | |
2429 | } | |
2430 | } | |
2431 | ||
2432 | void mem_cgroup_uncharge_end(void) | |
2433 | { | |
2434 | struct memcg_batch_info *batch = ¤t->memcg_batch; | |
2435 | ||
2436 | if (!batch->do_batch) | |
2437 | return; | |
2438 | ||
2439 | batch->do_batch--; | |
2440 | if (batch->do_batch) /* If stacked, do nothing. */ | |
2441 | return; | |
2442 | ||
2443 | if (!batch->memcg) | |
2444 | return; | |
2445 | /* | |
2446 | * This "batch->memcg" is valid without any css_get/put etc... | |
2447 | * bacause we hide charges behind us. | |
2448 | */ | |
2449 | if (batch->bytes) | |
2450 | res_counter_uncharge(&batch->memcg->res, batch->bytes); | |
2451 | if (batch->memsw_bytes) | |
2452 | res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes); | |
2453 | memcg_oom_recover(batch->memcg); | |
2454 | /* forget this pointer (for sanity check) */ | |
2455 | batch->memcg = NULL; | |
2456 | } | |
2457 | ||
2458 | #ifdef CONFIG_SWAP | |
2459 | /* | |
2460 | * called after __delete_from_swap_cache() and drop "page" account. | |
2461 | * memcg information is recorded to swap_cgroup of "ent" | |
2462 | */ | |
2463 | void | |
2464 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) | |
2465 | { | |
2466 | struct mem_cgroup *memcg; | |
2467 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; | |
2468 | ||
2469 | if (!swapout) /* this was a swap cache but the swap is unused ! */ | |
2470 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; | |
2471 | ||
2472 | memcg = __mem_cgroup_uncharge_common(page, ctype); | |
2473 | ||
2474 | /* | |
2475 | * record memcg information, if swapout && memcg != NULL, | |
2476 | * mem_cgroup_get() was called in uncharge(). | |
2477 | */ | |
2478 | if (do_swap_account && swapout && memcg) | |
2479 | swap_cgroup_record(ent, css_id(&memcg->css)); | |
2480 | } | |
2481 | #endif | |
2482 | ||
2483 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
2484 | /* | |
2485 | * called from swap_entry_free(). remove record in swap_cgroup and | |
2486 | * uncharge "memsw" account. | |
2487 | */ | |
2488 | void mem_cgroup_uncharge_swap(swp_entry_t ent) | |
2489 | { | |
2490 | struct mem_cgroup *memcg; | |
2491 | unsigned short id; | |
2492 | ||
2493 | if (!do_swap_account) | |
2494 | return; | |
2495 | ||
2496 | id = swap_cgroup_record(ent, 0); | |
2497 | rcu_read_lock(); | |
2498 | memcg = mem_cgroup_lookup(id); | |
2499 | if (memcg) { | |
2500 | /* | |
2501 | * We uncharge this because swap is freed. | |
2502 | * This memcg can be obsolete one. We avoid calling css_tryget | |
2503 | */ | |
2504 | if (!mem_cgroup_is_root(memcg)) | |
2505 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); | |
2506 | mem_cgroup_swap_statistics(memcg, false); | |
2507 | mem_cgroup_put(memcg); | |
2508 | } | |
2509 | rcu_read_unlock(); | |
2510 | } | |
2511 | ||
2512 | /** | |
2513 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. | |
2514 | * @entry: swap entry to be moved | |
2515 | * @from: mem_cgroup which the entry is moved from | |
2516 | * @to: mem_cgroup which the entry is moved to | |
2517 | * @need_fixup: whether we should fixup res_counters and refcounts. | |
2518 | * | |
2519 | * It succeeds only when the swap_cgroup's record for this entry is the same | |
2520 | * as the mem_cgroup's id of @from. | |
2521 | * | |
2522 | * Returns 0 on success, -EINVAL on failure. | |
2523 | * | |
2524 | * The caller must have charged to @to, IOW, called res_counter_charge() about | |
2525 | * both res and memsw, and called css_get(). | |
2526 | */ | |
2527 | static int mem_cgroup_move_swap_account(swp_entry_t entry, | |
2528 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) | |
2529 | { | |
2530 | unsigned short old_id, new_id; | |
2531 | ||
2532 | old_id = css_id(&from->css); | |
2533 | new_id = css_id(&to->css); | |
2534 | ||
2535 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { | |
2536 | mem_cgroup_swap_statistics(from, false); | |
2537 | mem_cgroup_swap_statistics(to, true); | |
2538 | /* | |
2539 | * This function is only called from task migration context now. | |
2540 | * It postpones res_counter and refcount handling till the end | |
2541 | * of task migration(mem_cgroup_clear_mc()) for performance | |
2542 | * improvement. But we cannot postpone mem_cgroup_get(to) | |
2543 | * because if the process that has been moved to @to does | |
2544 | * swap-in, the refcount of @to might be decreased to 0. | |
2545 | */ | |
2546 | mem_cgroup_get(to); | |
2547 | if (need_fixup) { | |
2548 | if (!mem_cgroup_is_root(from)) | |
2549 | res_counter_uncharge(&from->memsw, PAGE_SIZE); | |
2550 | mem_cgroup_put(from); | |
2551 | /* | |
2552 | * we charged both to->res and to->memsw, so we should | |
2553 | * uncharge to->res. | |
2554 | */ | |
2555 | if (!mem_cgroup_is_root(to)) | |
2556 | res_counter_uncharge(&to->res, PAGE_SIZE); | |
2557 | } | |
2558 | return 0; | |
2559 | } | |
2560 | return -EINVAL; | |
2561 | } | |
2562 | #else | |
2563 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, | |
2564 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) | |
2565 | { | |
2566 | return -EINVAL; | |
2567 | } | |
2568 | #endif | |
2569 | ||
2570 | /* | |
2571 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old | |
2572 | * page belongs to. | |
2573 | */ | |
2574 | int mem_cgroup_prepare_migration(struct page *page, | |
2575 | struct page *newpage, struct mem_cgroup **ptr) | |
2576 | { | |
2577 | struct page_cgroup *pc; | |
2578 | struct mem_cgroup *mem = NULL; | |
2579 | enum charge_type ctype; | |
2580 | int ret = 0; | |
2581 | ||
2582 | if (mem_cgroup_disabled()) | |
2583 | return 0; | |
2584 | ||
2585 | pc = lookup_page_cgroup(page); | |
2586 | lock_page_cgroup(pc); | |
2587 | if (PageCgroupUsed(pc)) { | |
2588 | mem = pc->mem_cgroup; | |
2589 | css_get(&mem->css); | |
2590 | /* | |
2591 | * At migrating an anonymous page, its mapcount goes down | |
2592 | * to 0 and uncharge() will be called. But, even if it's fully | |
2593 | * unmapped, migration may fail and this page has to be | |
2594 | * charged again. We set MIGRATION flag here and delay uncharge | |
2595 | * until end_migration() is called | |
2596 | * | |
2597 | * Corner Case Thinking | |
2598 | * A) | |
2599 | * When the old page was mapped as Anon and it's unmap-and-freed | |
2600 | * while migration was ongoing. | |
2601 | * If unmap finds the old page, uncharge() of it will be delayed | |
2602 | * until end_migration(). If unmap finds a new page, it's | |
2603 | * uncharged when it make mapcount to be 1->0. If unmap code | |
2604 | * finds swap_migration_entry, the new page will not be mapped | |
2605 | * and end_migration() will find it(mapcount==0). | |
2606 | * | |
2607 | * B) | |
2608 | * When the old page was mapped but migraion fails, the kernel | |
2609 | * remaps it. A charge for it is kept by MIGRATION flag even | |
2610 | * if mapcount goes down to 0. We can do remap successfully | |
2611 | * without charging it again. | |
2612 | * | |
2613 | * C) | |
2614 | * The "old" page is under lock_page() until the end of | |
2615 | * migration, so, the old page itself will not be swapped-out. | |
2616 | * If the new page is swapped out before end_migraton, our | |
2617 | * hook to usual swap-out path will catch the event. | |
2618 | */ | |
2619 | if (PageAnon(page)) | |
2620 | SetPageCgroupMigration(pc); | |
2621 | } | |
2622 | unlock_page_cgroup(pc); | |
2623 | /* | |
2624 | * If the page is not charged at this point, | |
2625 | * we return here. | |
2626 | */ | |
2627 | if (!mem) | |
2628 | return 0; | |
2629 | ||
2630 | *ptr = mem; | |
2631 | ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false); | |
2632 | css_put(&mem->css);/* drop extra refcnt */ | |
2633 | if (ret || *ptr == NULL) { | |
2634 | if (PageAnon(page)) { | |
2635 | lock_page_cgroup(pc); | |
2636 | ClearPageCgroupMigration(pc); | |
2637 | unlock_page_cgroup(pc); | |
2638 | /* | |
2639 | * The old page may be fully unmapped while we kept it. | |
2640 | */ | |
2641 | mem_cgroup_uncharge_page(page); | |
2642 | } | |
2643 | return -ENOMEM; | |
2644 | } | |
2645 | /* | |
2646 | * We charge new page before it's used/mapped. So, even if unlock_page() | |
2647 | * is called before end_migration, we can catch all events on this new | |
2648 | * page. In the case new page is migrated but not remapped, new page's | |
2649 | * mapcount will be finally 0 and we call uncharge in end_migration(). | |
2650 | */ | |
2651 | pc = lookup_page_cgroup(newpage); | |
2652 | if (PageAnon(page)) | |
2653 | ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; | |
2654 | else if (page_is_file_cache(page)) | |
2655 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
2656 | else | |
2657 | ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; | |
2658 | __mem_cgroup_commit_charge(mem, pc, ctype); | |
2659 | return ret; | |
2660 | } | |
2661 | ||
2662 | /* remove redundant charge if migration failed*/ | |
2663 | void mem_cgroup_end_migration(struct mem_cgroup *mem, | |
2664 | struct page *oldpage, struct page *newpage) | |
2665 | { | |
2666 | struct page *used, *unused; | |
2667 | struct page_cgroup *pc; | |
2668 | ||
2669 | if (!mem) | |
2670 | return; | |
2671 | /* blocks rmdir() */ | |
2672 | cgroup_exclude_rmdir(&mem->css); | |
2673 | /* at migration success, oldpage->mapping is NULL. */ | |
2674 | if (oldpage->mapping) { | |
2675 | used = oldpage; | |
2676 | unused = newpage; | |
2677 | } else { | |
2678 | used = newpage; | |
2679 | unused = oldpage; | |
2680 | } | |
2681 | /* | |
2682 | * We disallowed uncharge of pages under migration because mapcount | |
2683 | * of the page goes down to zero, temporarly. | |
2684 | * Clear the flag and check the page should be charged. | |
2685 | */ | |
2686 | pc = lookup_page_cgroup(oldpage); | |
2687 | lock_page_cgroup(pc); | |
2688 | ClearPageCgroupMigration(pc); | |
2689 | unlock_page_cgroup(pc); | |
2690 | ||
2691 | __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); | |
2692 | ||
2693 | /* | |
2694 | * If a page is a file cache, radix-tree replacement is very atomic | |
2695 | * and we can skip this check. When it was an Anon page, its mapcount | |
2696 | * goes down to 0. But because we added MIGRATION flage, it's not | |
2697 | * uncharged yet. There are several case but page->mapcount check | |
2698 | * and USED bit check in mem_cgroup_uncharge_page() will do enough | |
2699 | * check. (see prepare_charge() also) | |
2700 | */ | |
2701 | if (PageAnon(used)) | |
2702 | mem_cgroup_uncharge_page(used); | |
2703 | /* | |
2704 | * At migration, we may charge account against cgroup which has no | |
2705 | * tasks. | |
2706 | * So, rmdir()->pre_destroy() can be called while we do this charge. | |
2707 | * In that case, we need to call pre_destroy() again. check it here. | |
2708 | */ | |
2709 | cgroup_release_and_wakeup_rmdir(&mem->css); | |
2710 | } | |
2711 | ||
2712 | /* | |
2713 | * A call to try to shrink memory usage on charge failure at shmem's swapin. | |
2714 | * Calling hierarchical_reclaim is not enough because we should update | |
2715 | * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM. | |
2716 | * Moreover considering hierarchy, we should reclaim from the mem_over_limit, | |
2717 | * not from the memcg which this page would be charged to. | |
2718 | * try_charge_swapin does all of these works properly. | |
2719 | */ | |
2720 | int mem_cgroup_shmem_charge_fallback(struct page *page, | |
2721 | struct mm_struct *mm, | |
2722 | gfp_t gfp_mask) | |
2723 | { | |
2724 | struct mem_cgroup *mem = NULL; | |
2725 | int ret; | |
2726 | ||
2727 | if (mem_cgroup_disabled()) | |
2728 | return 0; | |
2729 | ||
2730 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); | |
2731 | if (!ret) | |
2732 | mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */ | |
2733 | ||
2734 | return ret; | |
2735 | } | |
2736 | ||
2737 | static DEFINE_MUTEX(set_limit_mutex); | |
2738 | ||
2739 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, | |
2740 | unsigned long long val) | |
2741 | { | |
2742 | int retry_count; | |
2743 | u64 memswlimit, memlimit; | |
2744 | int ret = 0; | |
2745 | int children = mem_cgroup_count_children(memcg); | |
2746 | u64 curusage, oldusage; | |
2747 | int enlarge; | |
2748 | ||
2749 | /* | |
2750 | * For keeping hierarchical_reclaim simple, how long we should retry | |
2751 | * is depends on callers. We set our retry-count to be function | |
2752 | * of # of children which we should visit in this loop. | |
2753 | */ | |
2754 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; | |
2755 | ||
2756 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
2757 | ||
2758 | enlarge = 0; | |
2759 | while (retry_count) { | |
2760 | if (signal_pending(current)) { | |
2761 | ret = -EINTR; | |
2762 | break; | |
2763 | } | |
2764 | /* | |
2765 | * Rather than hide all in some function, I do this in | |
2766 | * open coded manner. You see what this really does. | |
2767 | * We have to guarantee mem->res.limit < mem->memsw.limit. | |
2768 | */ | |
2769 | mutex_lock(&set_limit_mutex); | |
2770 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
2771 | if (memswlimit < val) { | |
2772 | ret = -EINVAL; | |
2773 | mutex_unlock(&set_limit_mutex); | |
2774 | break; | |
2775 | } | |
2776 | ||
2777 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
2778 | if (memlimit < val) | |
2779 | enlarge = 1; | |
2780 | ||
2781 | ret = res_counter_set_limit(&memcg->res, val); | |
2782 | if (!ret) { | |
2783 | if (memswlimit == val) | |
2784 | memcg->memsw_is_minimum = true; | |
2785 | else | |
2786 | memcg->memsw_is_minimum = false; | |
2787 | } | |
2788 | mutex_unlock(&set_limit_mutex); | |
2789 | ||
2790 | if (!ret) | |
2791 | break; | |
2792 | ||
2793 | mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, | |
2794 | MEM_CGROUP_RECLAIM_SHRINK); | |
2795 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
2796 | /* Usage is reduced ? */ | |
2797 | if (curusage >= oldusage) | |
2798 | retry_count--; | |
2799 | else | |
2800 | oldusage = curusage; | |
2801 | } | |
2802 | if (!ret && enlarge) | |
2803 | memcg_oom_recover(memcg); | |
2804 | ||
2805 | return ret; | |
2806 | } | |
2807 | ||
2808 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, | |
2809 | unsigned long long val) | |
2810 | { | |
2811 | int retry_count; | |
2812 | u64 memlimit, memswlimit, oldusage, curusage; | |
2813 | int children = mem_cgroup_count_children(memcg); | |
2814 | int ret = -EBUSY; | |
2815 | int enlarge = 0; | |
2816 | ||
2817 | /* see mem_cgroup_resize_res_limit */ | |
2818 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; | |
2819 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
2820 | while (retry_count) { | |
2821 | if (signal_pending(current)) { | |
2822 | ret = -EINTR; | |
2823 | break; | |
2824 | } | |
2825 | /* | |
2826 | * Rather than hide all in some function, I do this in | |
2827 | * open coded manner. You see what this really does. | |
2828 | * We have to guarantee mem->res.limit < mem->memsw.limit. | |
2829 | */ | |
2830 | mutex_lock(&set_limit_mutex); | |
2831 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
2832 | if (memlimit > val) { | |
2833 | ret = -EINVAL; | |
2834 | mutex_unlock(&set_limit_mutex); | |
2835 | break; | |
2836 | } | |
2837 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
2838 | if (memswlimit < val) | |
2839 | enlarge = 1; | |
2840 | ret = res_counter_set_limit(&memcg->memsw, val); | |
2841 | if (!ret) { | |
2842 | if (memlimit == val) | |
2843 | memcg->memsw_is_minimum = true; | |
2844 | else | |
2845 | memcg->memsw_is_minimum = false; | |
2846 | } | |
2847 | mutex_unlock(&set_limit_mutex); | |
2848 | ||
2849 | if (!ret) | |
2850 | break; | |
2851 | ||
2852 | mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, | |
2853 | MEM_CGROUP_RECLAIM_NOSWAP | | |
2854 | MEM_CGROUP_RECLAIM_SHRINK); | |
2855 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
2856 | /* Usage is reduced ? */ | |
2857 | if (curusage >= oldusage) | |
2858 | retry_count--; | |
2859 | else | |
2860 | oldusage = curusage; | |
2861 | } | |
2862 | if (!ret && enlarge) | |
2863 | memcg_oom_recover(memcg); | |
2864 | return ret; | |
2865 | } | |
2866 | ||
2867 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, | |
2868 | gfp_t gfp_mask) | |
2869 | { | |
2870 | unsigned long nr_reclaimed = 0; | |
2871 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; | |
2872 | unsigned long reclaimed; | |
2873 | int loop = 0; | |
2874 | struct mem_cgroup_tree_per_zone *mctz; | |
2875 | unsigned long long excess; | |
2876 | ||
2877 | if (order > 0) | |
2878 | return 0; | |
2879 | ||
2880 | mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); | |
2881 | /* | |
2882 | * This loop can run a while, specially if mem_cgroup's continuously | |
2883 | * keep exceeding their soft limit and putting the system under | |
2884 | * pressure | |
2885 | */ | |
2886 | do { | |
2887 | if (next_mz) | |
2888 | mz = next_mz; | |
2889 | else | |
2890 | mz = mem_cgroup_largest_soft_limit_node(mctz); | |
2891 | if (!mz) | |
2892 | break; | |
2893 | ||
2894 | reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone, | |
2895 | gfp_mask, | |
2896 | MEM_CGROUP_RECLAIM_SOFT); | |
2897 | nr_reclaimed += reclaimed; | |
2898 | spin_lock(&mctz->lock); | |
2899 | ||
2900 | /* | |
2901 | * If we failed to reclaim anything from this memory cgroup | |
2902 | * it is time to move on to the next cgroup | |
2903 | */ | |
2904 | next_mz = NULL; | |
2905 | if (!reclaimed) { | |
2906 | do { | |
2907 | /* | |
2908 | * Loop until we find yet another one. | |
2909 | * | |
2910 | * By the time we get the soft_limit lock | |
2911 | * again, someone might have aded the | |
2912 | * group back on the RB tree. Iterate to | |
2913 | * make sure we get a different mem. | |
2914 | * mem_cgroup_largest_soft_limit_node returns | |
2915 | * NULL if no other cgroup is present on | |
2916 | * the tree | |
2917 | */ | |
2918 | next_mz = | |
2919 | __mem_cgroup_largest_soft_limit_node(mctz); | |
2920 | if (next_mz == mz) { | |
2921 | css_put(&next_mz->mem->css); | |
2922 | next_mz = NULL; | |
2923 | } else /* next_mz == NULL or other memcg */ | |
2924 | break; | |
2925 | } while (1); | |
2926 | } | |
2927 | __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); | |
2928 | excess = res_counter_soft_limit_excess(&mz->mem->res); | |
2929 | /* | |
2930 | * One school of thought says that we should not add | |
2931 | * back the node to the tree if reclaim returns 0. | |
2932 | * But our reclaim could return 0, simply because due | |
2933 | * to priority we are exposing a smaller subset of | |
2934 | * memory to reclaim from. Consider this as a longer | |
2935 | * term TODO. | |
2936 | */ | |
2937 | /* If excess == 0, no tree ops */ | |
2938 | __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); | |
2939 | spin_unlock(&mctz->lock); | |
2940 | css_put(&mz->mem->css); | |
2941 | loop++; | |
2942 | /* | |
2943 | * Could not reclaim anything and there are no more | |
2944 | * mem cgroups to try or we seem to be looping without | |
2945 | * reclaiming anything. | |
2946 | */ | |
2947 | if (!nr_reclaimed && | |
2948 | (next_mz == NULL || | |
2949 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) | |
2950 | break; | |
2951 | } while (!nr_reclaimed); | |
2952 | if (next_mz) | |
2953 | css_put(&next_mz->mem->css); | |
2954 | return nr_reclaimed; | |
2955 | } | |
2956 | ||
2957 | /* | |
2958 | * This routine traverse page_cgroup in given list and drop them all. | |
2959 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. | |
2960 | */ | |
2961 | static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, | |
2962 | int node, int zid, enum lru_list lru) | |
2963 | { | |
2964 | struct zone *zone; | |
2965 | struct mem_cgroup_per_zone *mz; | |
2966 | struct page_cgroup *pc, *busy; | |
2967 | unsigned long flags, loop; | |
2968 | struct list_head *list; | |
2969 | int ret = 0; | |
2970 | ||
2971 | zone = &NODE_DATA(node)->node_zones[zid]; | |
2972 | mz = mem_cgroup_zoneinfo(mem, node, zid); | |
2973 | list = &mz->lists[lru]; | |
2974 | ||
2975 | loop = MEM_CGROUP_ZSTAT(mz, lru); | |
2976 | /* give some margin against EBUSY etc...*/ | |
2977 | loop += 256; | |
2978 | busy = NULL; | |
2979 | while (loop--) { | |
2980 | ret = 0; | |
2981 | spin_lock_irqsave(&zone->lru_lock, flags); | |
2982 | if (list_empty(list)) { | |
2983 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
2984 | break; | |
2985 | } | |
2986 | pc = list_entry(list->prev, struct page_cgroup, lru); | |
2987 | if (busy == pc) { | |
2988 | list_move(&pc->lru, list); | |
2989 | busy = NULL; | |
2990 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
2991 | continue; | |
2992 | } | |
2993 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
2994 | ||
2995 | ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL); | |
2996 | if (ret == -ENOMEM) | |
2997 | break; | |
2998 | ||
2999 | if (ret == -EBUSY || ret == -EINVAL) { | |
3000 | /* found lock contention or "pc" is obsolete. */ | |
3001 | busy = pc; | |
3002 | cond_resched(); | |
3003 | } else | |
3004 | busy = NULL; | |
3005 | } | |
3006 | ||
3007 | if (!ret && !list_empty(list)) | |
3008 | return -EBUSY; | |
3009 | return ret; | |
3010 | } | |
3011 | ||
3012 | /* | |
3013 | * make mem_cgroup's charge to be 0 if there is no task. | |
3014 | * This enables deleting this mem_cgroup. | |
3015 | */ | |
3016 | static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) | |
3017 | { | |
3018 | int ret; | |
3019 | int node, zid, shrink; | |
3020 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
3021 | struct cgroup *cgrp = mem->css.cgroup; | |
3022 | ||
3023 | css_get(&mem->css); | |
3024 | ||
3025 | shrink = 0; | |
3026 | /* should free all ? */ | |
3027 | if (free_all) | |
3028 | goto try_to_free; | |
3029 | move_account: | |
3030 | do { | |
3031 | ret = -EBUSY; | |
3032 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) | |
3033 | goto out; | |
3034 | ret = -EINTR; | |
3035 | if (signal_pending(current)) | |
3036 | goto out; | |
3037 | /* This is for making all *used* pages to be on LRU. */ | |
3038 | lru_add_drain_all(); | |
3039 | drain_all_stock_sync(); | |
3040 | ret = 0; | |
3041 | for_each_node_state(node, N_HIGH_MEMORY) { | |
3042 | for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { | |
3043 | enum lru_list l; | |
3044 | for_each_lru(l) { | |
3045 | ret = mem_cgroup_force_empty_list(mem, | |
3046 | node, zid, l); | |
3047 | if (ret) | |
3048 | break; | |
3049 | } | |
3050 | } | |
3051 | if (ret) | |
3052 | break; | |
3053 | } | |
3054 | memcg_oom_recover(mem); | |
3055 | /* it seems parent cgroup doesn't have enough mem */ | |
3056 | if (ret == -ENOMEM) | |
3057 | goto try_to_free; | |
3058 | cond_resched(); | |
3059 | /* "ret" should also be checked to ensure all lists are empty. */ | |
3060 | } while (mem->res.usage > 0 || ret); | |
3061 | out: | |
3062 | css_put(&mem->css); | |
3063 | return ret; | |
3064 | ||
3065 | try_to_free: | |
3066 | /* returns EBUSY if there is a task or if we come here twice. */ | |
3067 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { | |
3068 | ret = -EBUSY; | |
3069 | goto out; | |
3070 | } | |
3071 | /* we call try-to-free pages for make this cgroup empty */ | |
3072 | lru_add_drain_all(); | |
3073 | /* try to free all pages in this cgroup */ | |
3074 | shrink = 1; | |
3075 | while (nr_retries && mem->res.usage > 0) { | |
3076 | int progress; | |
3077 | ||
3078 | if (signal_pending(current)) { | |
3079 | ret = -EINTR; | |
3080 | goto out; | |
3081 | } | |
3082 | progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, | |
3083 | false, get_swappiness(mem)); | |
3084 | if (!progress) { | |
3085 | nr_retries--; | |
3086 | /* maybe some writeback is necessary */ | |
3087 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
3088 | } | |
3089 | ||
3090 | } | |
3091 | lru_add_drain(); | |
3092 | /* try move_account...there may be some *locked* pages. */ | |
3093 | goto move_account; | |
3094 | } | |
3095 | ||
3096 | int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) | |
3097 | { | |
3098 | return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); | |
3099 | } | |
3100 | ||
3101 | ||
3102 | static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) | |
3103 | { | |
3104 | return mem_cgroup_from_cont(cont)->use_hierarchy; | |
3105 | } | |
3106 | ||
3107 | static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, | |
3108 | u64 val) | |
3109 | { | |
3110 | int retval = 0; | |
3111 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
3112 | struct cgroup *parent = cont->parent; | |
3113 | struct mem_cgroup *parent_mem = NULL; | |
3114 | ||
3115 | if (parent) | |
3116 | parent_mem = mem_cgroup_from_cont(parent); | |
3117 | ||
3118 | cgroup_lock(); | |
3119 | /* | |
3120 | * If parent's use_hierarchy is set, we can't make any modifications | |
3121 | * in the child subtrees. If it is unset, then the change can | |
3122 | * occur, provided the current cgroup has no children. | |
3123 | * | |
3124 | * For the root cgroup, parent_mem is NULL, we allow value to be | |
3125 | * set if there are no children. | |
3126 | */ | |
3127 | if ((!parent_mem || !parent_mem->use_hierarchy) && | |
3128 | (val == 1 || val == 0)) { | |
3129 | if (list_empty(&cont->children)) | |
3130 | mem->use_hierarchy = val; | |
3131 | else | |
3132 | retval = -EBUSY; | |
3133 | } else | |
3134 | retval = -EINVAL; | |
3135 | cgroup_unlock(); | |
3136 | ||
3137 | return retval; | |
3138 | } | |
3139 | ||
3140 | struct mem_cgroup_idx_data { | |
3141 | s64 val; | |
3142 | enum mem_cgroup_stat_index idx; | |
3143 | }; | |
3144 | ||
3145 | static int | |
3146 | mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data) | |
3147 | { | |
3148 | struct mem_cgroup_idx_data *d = data; | |
3149 | d->val += mem_cgroup_read_stat(mem, d->idx); | |
3150 | return 0; | |
3151 | } | |
3152 | ||
3153 | static void | |
3154 | mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem, | |
3155 | enum mem_cgroup_stat_index idx, s64 *val) | |
3156 | { | |
3157 | struct mem_cgroup_idx_data d; | |
3158 | d.idx = idx; | |
3159 | d.val = 0; | |
3160 | mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat); | |
3161 | *val = d.val; | |
3162 | } | |
3163 | ||
3164 | static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap) | |
3165 | { | |
3166 | u64 idx_val, val; | |
3167 | ||
3168 | if (!mem_cgroup_is_root(mem)) { | |
3169 | if (!swap) | |
3170 | return res_counter_read_u64(&mem->res, RES_USAGE); | |
3171 | else | |
3172 | return res_counter_read_u64(&mem->memsw, RES_USAGE); | |
3173 | } | |
3174 | ||
3175 | mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val); | |
3176 | val = idx_val; | |
3177 | mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val); | |
3178 | val += idx_val; | |
3179 | ||
3180 | if (swap) { | |
3181 | mem_cgroup_get_recursive_idx_stat(mem, | |
3182 | MEM_CGROUP_STAT_SWAPOUT, &idx_val); | |
3183 | val += idx_val; | |
3184 | } | |
3185 | ||
3186 | return val << PAGE_SHIFT; | |
3187 | } | |
3188 | ||
3189 | static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) | |
3190 | { | |
3191 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
3192 | u64 val; | |
3193 | int type, name; | |
3194 | ||
3195 | type = MEMFILE_TYPE(cft->private); | |
3196 | name = MEMFILE_ATTR(cft->private); | |
3197 | switch (type) { | |
3198 | case _MEM: | |
3199 | if (name == RES_USAGE) | |
3200 | val = mem_cgroup_usage(mem, false); | |
3201 | else | |
3202 | val = res_counter_read_u64(&mem->res, name); | |
3203 | break; | |
3204 | case _MEMSWAP: | |
3205 | if (name == RES_USAGE) | |
3206 | val = mem_cgroup_usage(mem, true); | |
3207 | else | |
3208 | val = res_counter_read_u64(&mem->memsw, name); | |
3209 | break; | |
3210 | default: | |
3211 | BUG(); | |
3212 | break; | |
3213 | } | |
3214 | return val; | |
3215 | } | |
3216 | /* | |
3217 | * The user of this function is... | |
3218 | * RES_LIMIT. | |
3219 | */ | |
3220 | static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, | |
3221 | const char *buffer) | |
3222 | { | |
3223 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
3224 | int type, name; | |
3225 | unsigned long long val; | |
3226 | int ret; | |
3227 | ||
3228 | type = MEMFILE_TYPE(cft->private); | |
3229 | name = MEMFILE_ATTR(cft->private); | |
3230 | switch (name) { | |
3231 | case RES_LIMIT: | |
3232 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ | |
3233 | ret = -EINVAL; | |
3234 | break; | |
3235 | } | |
3236 | /* This function does all necessary parse...reuse it */ | |
3237 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
3238 | if (ret) | |
3239 | break; | |
3240 | if (type == _MEM) | |
3241 | ret = mem_cgroup_resize_limit(memcg, val); | |
3242 | else | |
3243 | ret = mem_cgroup_resize_memsw_limit(memcg, val); | |
3244 | break; | |
3245 | case RES_SOFT_LIMIT: | |
3246 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
3247 | if (ret) | |
3248 | break; | |
3249 | /* | |
3250 | * For memsw, soft limits are hard to implement in terms | |
3251 | * of semantics, for now, we support soft limits for | |
3252 | * control without swap | |
3253 | */ | |
3254 | if (type == _MEM) | |
3255 | ret = res_counter_set_soft_limit(&memcg->res, val); | |
3256 | else | |
3257 | ret = -EINVAL; | |
3258 | break; | |
3259 | default: | |
3260 | ret = -EINVAL; /* should be BUG() ? */ | |
3261 | break; | |
3262 | } | |
3263 | return ret; | |
3264 | } | |
3265 | ||
3266 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, | |
3267 | unsigned long long *mem_limit, unsigned long long *memsw_limit) | |
3268 | { | |
3269 | struct cgroup *cgroup; | |
3270 | unsigned long long min_limit, min_memsw_limit, tmp; | |
3271 | ||
3272 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3273 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3274 | cgroup = memcg->css.cgroup; | |
3275 | if (!memcg->use_hierarchy) | |
3276 | goto out; | |
3277 | ||
3278 | while (cgroup->parent) { | |
3279 | cgroup = cgroup->parent; | |
3280 | memcg = mem_cgroup_from_cont(cgroup); | |
3281 | if (!memcg->use_hierarchy) | |
3282 | break; | |
3283 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3284 | min_limit = min(min_limit, tmp); | |
3285 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3286 | min_memsw_limit = min(min_memsw_limit, tmp); | |
3287 | } | |
3288 | out: | |
3289 | *mem_limit = min_limit; | |
3290 | *memsw_limit = min_memsw_limit; | |
3291 | return; | |
3292 | } | |
3293 | ||
3294 | static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) | |
3295 | { | |
3296 | struct mem_cgroup *mem; | |
3297 | int type, name; | |
3298 | ||
3299 | mem = mem_cgroup_from_cont(cont); | |
3300 | type = MEMFILE_TYPE(event); | |
3301 | name = MEMFILE_ATTR(event); | |
3302 | switch (name) { | |
3303 | case RES_MAX_USAGE: | |
3304 | if (type == _MEM) | |
3305 | res_counter_reset_max(&mem->res); | |
3306 | else | |
3307 | res_counter_reset_max(&mem->memsw); | |
3308 | break; | |
3309 | case RES_FAILCNT: | |
3310 | if (type == _MEM) | |
3311 | res_counter_reset_failcnt(&mem->res); | |
3312 | else | |
3313 | res_counter_reset_failcnt(&mem->memsw); | |
3314 | break; | |
3315 | } | |
3316 | ||
3317 | return 0; | |
3318 | } | |
3319 | ||
3320 | static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, | |
3321 | struct cftype *cft) | |
3322 | { | |
3323 | return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; | |
3324 | } | |
3325 | ||
3326 | #ifdef CONFIG_MMU | |
3327 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, | |
3328 | struct cftype *cft, u64 val) | |
3329 | { | |
3330 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
3331 | ||
3332 | if (val >= (1 << NR_MOVE_TYPE)) | |
3333 | return -EINVAL; | |
3334 | /* | |
3335 | * We check this value several times in both in can_attach() and | |
3336 | * attach(), so we need cgroup lock to prevent this value from being | |
3337 | * inconsistent. | |
3338 | */ | |
3339 | cgroup_lock(); | |
3340 | mem->move_charge_at_immigrate = val; | |
3341 | cgroup_unlock(); | |
3342 | ||
3343 | return 0; | |
3344 | } | |
3345 | #else | |
3346 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, | |
3347 | struct cftype *cft, u64 val) | |
3348 | { | |
3349 | return -ENOSYS; | |
3350 | } | |
3351 | #endif | |
3352 | ||
3353 | ||
3354 | /* For read statistics */ | |
3355 | enum { | |
3356 | MCS_CACHE, | |
3357 | MCS_RSS, | |
3358 | MCS_FILE_MAPPED, | |
3359 | MCS_PGPGIN, | |
3360 | MCS_PGPGOUT, | |
3361 | MCS_SWAP, | |
3362 | MCS_INACTIVE_ANON, | |
3363 | MCS_ACTIVE_ANON, | |
3364 | MCS_INACTIVE_FILE, | |
3365 | MCS_ACTIVE_FILE, | |
3366 | MCS_UNEVICTABLE, | |
3367 | NR_MCS_STAT, | |
3368 | }; | |
3369 | ||
3370 | struct mcs_total_stat { | |
3371 | s64 stat[NR_MCS_STAT]; | |
3372 | }; | |
3373 | ||
3374 | struct { | |
3375 | char *local_name; | |
3376 | char *total_name; | |
3377 | } memcg_stat_strings[NR_MCS_STAT] = { | |
3378 | {"cache", "total_cache"}, | |
3379 | {"rss", "total_rss"}, | |
3380 | {"mapped_file", "total_mapped_file"}, | |
3381 | {"pgpgin", "total_pgpgin"}, | |
3382 | {"pgpgout", "total_pgpgout"}, | |
3383 | {"swap", "total_swap"}, | |
3384 | {"inactive_anon", "total_inactive_anon"}, | |
3385 | {"active_anon", "total_active_anon"}, | |
3386 | {"inactive_file", "total_inactive_file"}, | |
3387 | {"active_file", "total_active_file"}, | |
3388 | {"unevictable", "total_unevictable"} | |
3389 | }; | |
3390 | ||
3391 | ||
3392 | static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data) | |
3393 | { | |
3394 | struct mcs_total_stat *s = data; | |
3395 | s64 val; | |
3396 | ||
3397 | /* per cpu stat */ | |
3398 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); | |
3399 | s->stat[MCS_CACHE] += val * PAGE_SIZE; | |
3400 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); | |
3401 | s->stat[MCS_RSS] += val * PAGE_SIZE; | |
3402 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED); | |
3403 | s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; | |
3404 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT); | |
3405 | s->stat[MCS_PGPGIN] += val; | |
3406 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT); | |
3407 | s->stat[MCS_PGPGOUT] += val; | |
3408 | if (do_swap_account) { | |
3409 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT); | |
3410 | s->stat[MCS_SWAP] += val * PAGE_SIZE; | |
3411 | } | |
3412 | ||
3413 | /* per zone stat */ | |
3414 | val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); | |
3415 | s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; | |
3416 | val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); | |
3417 | s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; | |
3418 | val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); | |
3419 | s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; | |
3420 | val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); | |
3421 | s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; | |
3422 | val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); | |
3423 | s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; | |
3424 | return 0; | |
3425 | } | |
3426 | ||
3427 | static void | |
3428 | mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) | |
3429 | { | |
3430 | mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat); | |
3431 | } | |
3432 | ||
3433 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, | |
3434 | struct cgroup_map_cb *cb) | |
3435 | { | |
3436 | struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); | |
3437 | struct mcs_total_stat mystat; | |
3438 | int i; | |
3439 | ||
3440 | memset(&mystat, 0, sizeof(mystat)); | |
3441 | mem_cgroup_get_local_stat(mem_cont, &mystat); | |
3442 | ||
3443 | for (i = 0; i < NR_MCS_STAT; i++) { | |
3444 | if (i == MCS_SWAP && !do_swap_account) | |
3445 | continue; | |
3446 | cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); | |
3447 | } | |
3448 | ||
3449 | /* Hierarchical information */ | |
3450 | { | |
3451 | unsigned long long limit, memsw_limit; | |
3452 | memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); | |
3453 | cb->fill(cb, "hierarchical_memory_limit", limit); | |
3454 | if (do_swap_account) | |
3455 | cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); | |
3456 | } | |
3457 | ||
3458 | memset(&mystat, 0, sizeof(mystat)); | |
3459 | mem_cgroup_get_total_stat(mem_cont, &mystat); | |
3460 | for (i = 0; i < NR_MCS_STAT; i++) { | |
3461 | if (i == MCS_SWAP && !do_swap_account) | |
3462 | continue; | |
3463 | cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); | |
3464 | } | |
3465 | ||
3466 | #ifdef CONFIG_DEBUG_VM | |
3467 | cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); | |
3468 | ||
3469 | { | |
3470 | int nid, zid; | |
3471 | struct mem_cgroup_per_zone *mz; | |
3472 | unsigned long recent_rotated[2] = {0, 0}; | |
3473 | unsigned long recent_scanned[2] = {0, 0}; | |
3474 | ||
3475 | for_each_online_node(nid) | |
3476 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
3477 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); | |
3478 | ||
3479 | recent_rotated[0] += | |
3480 | mz->reclaim_stat.recent_rotated[0]; | |
3481 | recent_rotated[1] += | |
3482 | mz->reclaim_stat.recent_rotated[1]; | |
3483 | recent_scanned[0] += | |
3484 | mz->reclaim_stat.recent_scanned[0]; | |
3485 | recent_scanned[1] += | |
3486 | mz->reclaim_stat.recent_scanned[1]; | |
3487 | } | |
3488 | cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); | |
3489 | cb->fill(cb, "recent_rotated_file", recent_rotated[1]); | |
3490 | cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); | |
3491 | cb->fill(cb, "recent_scanned_file", recent_scanned[1]); | |
3492 | } | |
3493 | #endif | |
3494 | ||
3495 | return 0; | |
3496 | } | |
3497 | ||
3498 | static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) | |
3499 | { | |
3500 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3501 | ||
3502 | return get_swappiness(memcg); | |
3503 | } | |
3504 | ||
3505 | static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, | |
3506 | u64 val) | |
3507 | { | |
3508 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3509 | struct mem_cgroup *parent; | |
3510 | ||
3511 | if (val > 100) | |
3512 | return -EINVAL; | |
3513 | ||
3514 | if (cgrp->parent == NULL) | |
3515 | return -EINVAL; | |
3516 | ||
3517 | parent = mem_cgroup_from_cont(cgrp->parent); | |
3518 | ||
3519 | cgroup_lock(); | |
3520 | ||
3521 | /* If under hierarchy, only empty-root can set this value */ | |
3522 | if ((parent->use_hierarchy) || | |
3523 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { | |
3524 | cgroup_unlock(); | |
3525 | return -EINVAL; | |
3526 | } | |
3527 | ||
3528 | spin_lock(&memcg->reclaim_param_lock); | |
3529 | memcg->swappiness = val; | |
3530 | spin_unlock(&memcg->reclaim_param_lock); | |
3531 | ||
3532 | cgroup_unlock(); | |
3533 | ||
3534 | return 0; | |
3535 | } | |
3536 | ||
3537 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) | |
3538 | { | |
3539 | struct mem_cgroup_threshold_ary *t; | |
3540 | u64 usage; | |
3541 | int i; | |
3542 | ||
3543 | rcu_read_lock(); | |
3544 | if (!swap) | |
3545 | t = rcu_dereference(memcg->thresholds.primary); | |
3546 | else | |
3547 | t = rcu_dereference(memcg->memsw_thresholds.primary); | |
3548 | ||
3549 | if (!t) | |
3550 | goto unlock; | |
3551 | ||
3552 | usage = mem_cgroup_usage(memcg, swap); | |
3553 | ||
3554 | /* | |
3555 | * current_threshold points to threshold just below usage. | |
3556 | * If it's not true, a threshold was crossed after last | |
3557 | * call of __mem_cgroup_threshold(). | |
3558 | */ | |
3559 | i = t->current_threshold; | |
3560 | ||
3561 | /* | |
3562 | * Iterate backward over array of thresholds starting from | |
3563 | * current_threshold and check if a threshold is crossed. | |
3564 | * If none of thresholds below usage is crossed, we read | |
3565 | * only one element of the array here. | |
3566 | */ | |
3567 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) | |
3568 | eventfd_signal(t->entries[i].eventfd, 1); | |
3569 | ||
3570 | /* i = current_threshold + 1 */ | |
3571 | i++; | |
3572 | ||
3573 | /* | |
3574 | * Iterate forward over array of thresholds starting from | |
3575 | * current_threshold+1 and check if a threshold is crossed. | |
3576 | * If none of thresholds above usage is crossed, we read | |
3577 | * only one element of the array here. | |
3578 | */ | |
3579 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) | |
3580 | eventfd_signal(t->entries[i].eventfd, 1); | |
3581 | ||
3582 | /* Update current_threshold */ | |
3583 | t->current_threshold = i - 1; | |
3584 | unlock: | |
3585 | rcu_read_unlock(); | |
3586 | } | |
3587 | ||
3588 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) | |
3589 | { | |
3590 | while (memcg) { | |
3591 | __mem_cgroup_threshold(memcg, false); | |
3592 | if (do_swap_account) | |
3593 | __mem_cgroup_threshold(memcg, true); | |
3594 | ||
3595 | memcg = parent_mem_cgroup(memcg); | |
3596 | } | |
3597 | } | |
3598 | ||
3599 | static int compare_thresholds(const void *a, const void *b) | |
3600 | { | |
3601 | const struct mem_cgroup_threshold *_a = a; | |
3602 | const struct mem_cgroup_threshold *_b = b; | |
3603 | ||
3604 | return _a->threshold - _b->threshold; | |
3605 | } | |
3606 | ||
3607 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem, void *data) | |
3608 | { | |
3609 | struct mem_cgroup_eventfd_list *ev; | |
3610 | ||
3611 | list_for_each_entry(ev, &mem->oom_notify, list) | |
3612 | eventfd_signal(ev->eventfd, 1); | |
3613 | return 0; | |
3614 | } | |
3615 | ||
3616 | static void mem_cgroup_oom_notify(struct mem_cgroup *mem) | |
3617 | { | |
3618 | mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_notify_cb); | |
3619 | } | |
3620 | ||
3621 | static int mem_cgroup_usage_register_event(struct cgroup *cgrp, | |
3622 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) | |
3623 | { | |
3624 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3625 | struct mem_cgroup_thresholds *thresholds; | |
3626 | struct mem_cgroup_threshold_ary *new; | |
3627 | int type = MEMFILE_TYPE(cft->private); | |
3628 | u64 threshold, usage; | |
3629 | int i, size, ret; | |
3630 | ||
3631 | ret = res_counter_memparse_write_strategy(args, &threshold); | |
3632 | if (ret) | |
3633 | return ret; | |
3634 | ||
3635 | mutex_lock(&memcg->thresholds_lock); | |
3636 | ||
3637 | if (type == _MEM) | |
3638 | thresholds = &memcg->thresholds; | |
3639 | else if (type == _MEMSWAP) | |
3640 | thresholds = &memcg->memsw_thresholds; | |
3641 | else | |
3642 | BUG(); | |
3643 | ||
3644 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
3645 | ||
3646 | /* Check if a threshold crossed before adding a new one */ | |
3647 | if (thresholds->primary) | |
3648 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
3649 | ||
3650 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; | |
3651 | ||
3652 | /* Allocate memory for new array of thresholds */ | |
3653 | new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), | |
3654 | GFP_KERNEL); | |
3655 | if (!new) { | |
3656 | ret = -ENOMEM; | |
3657 | goto unlock; | |
3658 | } | |
3659 | new->size = size; | |
3660 | ||
3661 | /* Copy thresholds (if any) to new array */ | |
3662 | if (thresholds->primary) { | |
3663 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * | |
3664 | sizeof(struct mem_cgroup_threshold)); | |
3665 | } | |
3666 | ||
3667 | /* Add new threshold */ | |
3668 | new->entries[size - 1].eventfd = eventfd; | |
3669 | new->entries[size - 1].threshold = threshold; | |
3670 | ||
3671 | /* Sort thresholds. Registering of new threshold isn't time-critical */ | |
3672 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), | |
3673 | compare_thresholds, NULL); | |
3674 | ||
3675 | /* Find current threshold */ | |
3676 | new->current_threshold = -1; | |
3677 | for (i = 0; i < size; i++) { | |
3678 | if (new->entries[i].threshold < usage) { | |
3679 | /* | |
3680 | * new->current_threshold will not be used until | |
3681 | * rcu_assign_pointer(), so it's safe to increment | |
3682 | * it here. | |
3683 | */ | |
3684 | ++new->current_threshold; | |
3685 | } | |
3686 | } | |
3687 | ||
3688 | /* Free old spare buffer and save old primary buffer as spare */ | |
3689 | kfree(thresholds->spare); | |
3690 | thresholds->spare = thresholds->primary; | |
3691 | ||
3692 | rcu_assign_pointer(thresholds->primary, new); | |
3693 | ||
3694 | /* To be sure that nobody uses thresholds */ | |
3695 | synchronize_rcu(); | |
3696 | ||
3697 | unlock: | |
3698 | mutex_unlock(&memcg->thresholds_lock); | |
3699 | ||
3700 | return ret; | |
3701 | } | |
3702 | ||
3703 | static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, | |
3704 | struct cftype *cft, struct eventfd_ctx *eventfd) | |
3705 | { | |
3706 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3707 | struct mem_cgroup_thresholds *thresholds; | |
3708 | struct mem_cgroup_threshold_ary *new; | |
3709 | int type = MEMFILE_TYPE(cft->private); | |
3710 | u64 usage; | |
3711 | int i, j, size; | |
3712 | ||
3713 | mutex_lock(&memcg->thresholds_lock); | |
3714 | if (type == _MEM) | |
3715 | thresholds = &memcg->thresholds; | |
3716 | else if (type == _MEMSWAP) | |
3717 | thresholds = &memcg->memsw_thresholds; | |
3718 | else | |
3719 | BUG(); | |
3720 | ||
3721 | /* | |
3722 | * Something went wrong if we trying to unregister a threshold | |
3723 | * if we don't have thresholds | |
3724 | */ | |
3725 | BUG_ON(!thresholds); | |
3726 | ||
3727 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
3728 | ||
3729 | /* Check if a threshold crossed before removing */ | |
3730 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
3731 | ||
3732 | /* Calculate new number of threshold */ | |
3733 | size = 0; | |
3734 | for (i = 0; i < thresholds->primary->size; i++) { | |
3735 | if (thresholds->primary->entries[i].eventfd != eventfd) | |
3736 | size++; | |
3737 | } | |
3738 | ||
3739 | new = thresholds->spare; | |
3740 | ||
3741 | /* Set thresholds array to NULL if we don't have thresholds */ | |
3742 | if (!size) { | |
3743 | kfree(new); | |
3744 | new = NULL; | |
3745 | goto swap_buffers; | |
3746 | } | |
3747 | ||
3748 | new->size = size; | |
3749 | ||
3750 | /* Copy thresholds and find current threshold */ | |
3751 | new->current_threshold = -1; | |
3752 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { | |
3753 | if (thresholds->primary->entries[i].eventfd == eventfd) | |
3754 | continue; | |
3755 | ||
3756 | new->entries[j] = thresholds->primary->entries[i]; | |
3757 | if (new->entries[j].threshold < usage) { | |
3758 | /* | |
3759 | * new->current_threshold will not be used | |
3760 | * until rcu_assign_pointer(), so it's safe to increment | |
3761 | * it here. | |
3762 | */ | |
3763 | ++new->current_threshold; | |
3764 | } | |
3765 | j++; | |
3766 | } | |
3767 | ||
3768 | swap_buffers: | |
3769 | /* Swap primary and spare array */ | |
3770 | thresholds->spare = thresholds->primary; | |
3771 | rcu_assign_pointer(thresholds->primary, new); | |
3772 | ||
3773 | /* To be sure that nobody uses thresholds */ | |
3774 | synchronize_rcu(); | |
3775 | ||
3776 | mutex_unlock(&memcg->thresholds_lock); | |
3777 | } | |
3778 | ||
3779 | static int mem_cgroup_oom_register_event(struct cgroup *cgrp, | |
3780 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) | |
3781 | { | |
3782 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3783 | struct mem_cgroup_eventfd_list *event; | |
3784 | int type = MEMFILE_TYPE(cft->private); | |
3785 | ||
3786 | BUG_ON(type != _OOM_TYPE); | |
3787 | event = kmalloc(sizeof(*event), GFP_KERNEL); | |
3788 | if (!event) | |
3789 | return -ENOMEM; | |
3790 | ||
3791 | mutex_lock(&memcg_oom_mutex); | |
3792 | ||
3793 | event->eventfd = eventfd; | |
3794 | list_add(&event->list, &memcg->oom_notify); | |
3795 | ||
3796 | /* already in OOM ? */ | |
3797 | if (atomic_read(&memcg->oom_lock)) | |
3798 | eventfd_signal(eventfd, 1); | |
3799 | mutex_unlock(&memcg_oom_mutex); | |
3800 | ||
3801 | return 0; | |
3802 | } | |
3803 | ||
3804 | static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, | |
3805 | struct cftype *cft, struct eventfd_ctx *eventfd) | |
3806 | { | |
3807 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
3808 | struct mem_cgroup_eventfd_list *ev, *tmp; | |
3809 | int type = MEMFILE_TYPE(cft->private); | |
3810 | ||
3811 | BUG_ON(type != _OOM_TYPE); | |
3812 | ||
3813 | mutex_lock(&memcg_oom_mutex); | |
3814 | ||
3815 | list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) { | |
3816 | if (ev->eventfd == eventfd) { | |
3817 | list_del(&ev->list); | |
3818 | kfree(ev); | |
3819 | } | |
3820 | } | |
3821 | ||
3822 | mutex_unlock(&memcg_oom_mutex); | |
3823 | } | |
3824 | ||
3825 | static int mem_cgroup_oom_control_read(struct cgroup *cgrp, | |
3826 | struct cftype *cft, struct cgroup_map_cb *cb) | |
3827 | { | |
3828 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
3829 | ||
3830 | cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable); | |
3831 | ||
3832 | if (atomic_read(&mem->oom_lock)) | |
3833 | cb->fill(cb, "under_oom", 1); | |
3834 | else | |
3835 | cb->fill(cb, "under_oom", 0); | |
3836 | return 0; | |
3837 | } | |
3838 | ||
3839 | static int mem_cgroup_oom_control_write(struct cgroup *cgrp, | |
3840 | struct cftype *cft, u64 val) | |
3841 | { | |
3842 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
3843 | struct mem_cgroup *parent; | |
3844 | ||
3845 | /* cannot set to root cgroup and only 0 and 1 are allowed */ | |
3846 | if (!cgrp->parent || !((val == 0) || (val == 1))) | |
3847 | return -EINVAL; | |
3848 | ||
3849 | parent = mem_cgroup_from_cont(cgrp->parent); | |
3850 | ||
3851 | cgroup_lock(); | |
3852 | /* oom-kill-disable is a flag for subhierarchy. */ | |
3853 | if ((parent->use_hierarchy) || | |
3854 | (mem->use_hierarchy && !list_empty(&cgrp->children))) { | |
3855 | cgroup_unlock(); | |
3856 | return -EINVAL; | |
3857 | } | |
3858 | mem->oom_kill_disable = val; | |
3859 | if (!val) | |
3860 | memcg_oom_recover(mem); | |
3861 | cgroup_unlock(); | |
3862 | return 0; | |
3863 | } | |
3864 | ||
3865 | static struct cftype mem_cgroup_files[] = { | |
3866 | { | |
3867 | .name = "usage_in_bytes", | |
3868 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), | |
3869 | .read_u64 = mem_cgroup_read, | |
3870 | .register_event = mem_cgroup_usage_register_event, | |
3871 | .unregister_event = mem_cgroup_usage_unregister_event, | |
3872 | }, | |
3873 | { | |
3874 | .name = "max_usage_in_bytes", | |
3875 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), | |
3876 | .trigger = mem_cgroup_reset, | |
3877 | .read_u64 = mem_cgroup_read, | |
3878 | }, | |
3879 | { | |
3880 | .name = "limit_in_bytes", | |
3881 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), | |
3882 | .write_string = mem_cgroup_write, | |
3883 | .read_u64 = mem_cgroup_read, | |
3884 | }, | |
3885 | { | |
3886 | .name = "soft_limit_in_bytes", | |
3887 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), | |
3888 | .write_string = mem_cgroup_write, | |
3889 | .read_u64 = mem_cgroup_read, | |
3890 | }, | |
3891 | { | |
3892 | .name = "failcnt", | |
3893 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), | |
3894 | .trigger = mem_cgroup_reset, | |
3895 | .read_u64 = mem_cgroup_read, | |
3896 | }, | |
3897 | { | |
3898 | .name = "stat", | |
3899 | .read_map = mem_control_stat_show, | |
3900 | }, | |
3901 | { | |
3902 | .name = "force_empty", | |
3903 | .trigger = mem_cgroup_force_empty_write, | |
3904 | }, | |
3905 | { | |
3906 | .name = "use_hierarchy", | |
3907 | .write_u64 = mem_cgroup_hierarchy_write, | |
3908 | .read_u64 = mem_cgroup_hierarchy_read, | |
3909 | }, | |
3910 | { | |
3911 | .name = "swappiness", | |
3912 | .read_u64 = mem_cgroup_swappiness_read, | |
3913 | .write_u64 = mem_cgroup_swappiness_write, | |
3914 | }, | |
3915 | { | |
3916 | .name = "move_charge_at_immigrate", | |
3917 | .read_u64 = mem_cgroup_move_charge_read, | |
3918 | .write_u64 = mem_cgroup_move_charge_write, | |
3919 | }, | |
3920 | { | |
3921 | .name = "oom_control", | |
3922 | .read_map = mem_cgroup_oom_control_read, | |
3923 | .write_u64 = mem_cgroup_oom_control_write, | |
3924 | .register_event = mem_cgroup_oom_register_event, | |
3925 | .unregister_event = mem_cgroup_oom_unregister_event, | |
3926 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), | |
3927 | }, | |
3928 | }; | |
3929 | ||
3930 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
3931 | static struct cftype memsw_cgroup_files[] = { | |
3932 | { | |
3933 | .name = "memsw.usage_in_bytes", | |
3934 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), | |
3935 | .read_u64 = mem_cgroup_read, | |
3936 | .register_event = mem_cgroup_usage_register_event, | |
3937 | .unregister_event = mem_cgroup_usage_unregister_event, | |
3938 | }, | |
3939 | { | |
3940 | .name = "memsw.max_usage_in_bytes", | |
3941 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), | |
3942 | .trigger = mem_cgroup_reset, | |
3943 | .read_u64 = mem_cgroup_read, | |
3944 | }, | |
3945 | { | |
3946 | .name = "memsw.limit_in_bytes", | |
3947 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), | |
3948 | .write_string = mem_cgroup_write, | |
3949 | .read_u64 = mem_cgroup_read, | |
3950 | }, | |
3951 | { | |
3952 | .name = "memsw.failcnt", | |
3953 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), | |
3954 | .trigger = mem_cgroup_reset, | |
3955 | .read_u64 = mem_cgroup_read, | |
3956 | }, | |
3957 | }; | |
3958 | ||
3959 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) | |
3960 | { | |
3961 | if (!do_swap_account) | |
3962 | return 0; | |
3963 | return cgroup_add_files(cont, ss, memsw_cgroup_files, | |
3964 | ARRAY_SIZE(memsw_cgroup_files)); | |
3965 | }; | |
3966 | #else | |
3967 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) | |
3968 | { | |
3969 | return 0; | |
3970 | } | |
3971 | #endif | |
3972 | ||
3973 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) | |
3974 | { | |
3975 | struct mem_cgroup_per_node *pn; | |
3976 | struct mem_cgroup_per_zone *mz; | |
3977 | enum lru_list l; | |
3978 | int zone, tmp = node; | |
3979 | /* | |
3980 | * This routine is called against possible nodes. | |
3981 | * But it's BUG to call kmalloc() against offline node. | |
3982 | * | |
3983 | * TODO: this routine can waste much memory for nodes which will | |
3984 | * never be onlined. It's better to use memory hotplug callback | |
3985 | * function. | |
3986 | */ | |
3987 | if (!node_state(node, N_NORMAL_MEMORY)) | |
3988 | tmp = -1; | |
3989 | pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp); | |
3990 | if (!pn) | |
3991 | return 1; | |
3992 | ||
3993 | mem->info.nodeinfo[node] = pn; | |
3994 | memset(pn, 0, sizeof(*pn)); | |
3995 | ||
3996 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
3997 | mz = &pn->zoneinfo[zone]; | |
3998 | for_each_lru(l) | |
3999 | INIT_LIST_HEAD(&mz->lists[l]); | |
4000 | mz->usage_in_excess = 0; | |
4001 | mz->on_tree = false; | |
4002 | mz->mem = mem; | |
4003 | } | |
4004 | return 0; | |
4005 | } | |
4006 | ||
4007 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) | |
4008 | { | |
4009 | kfree(mem->info.nodeinfo[node]); | |
4010 | } | |
4011 | ||
4012 | static struct mem_cgroup *mem_cgroup_alloc(void) | |
4013 | { | |
4014 | struct mem_cgroup *mem; | |
4015 | int size = sizeof(struct mem_cgroup); | |
4016 | ||
4017 | /* Can be very big if MAX_NUMNODES is very big */ | |
4018 | if (size < PAGE_SIZE) | |
4019 | mem = kmalloc(size, GFP_KERNEL); | |
4020 | else | |
4021 | mem = vmalloc(size); | |
4022 | ||
4023 | if (!mem) | |
4024 | return NULL; | |
4025 | ||
4026 | memset(mem, 0, size); | |
4027 | mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); | |
4028 | if (!mem->stat) { | |
4029 | if (size < PAGE_SIZE) | |
4030 | kfree(mem); | |
4031 | else | |
4032 | vfree(mem); | |
4033 | mem = NULL; | |
4034 | } | |
4035 | return mem; | |
4036 | } | |
4037 | ||
4038 | /* | |
4039 | * At destroying mem_cgroup, references from swap_cgroup can remain. | |
4040 | * (scanning all at force_empty is too costly...) | |
4041 | * | |
4042 | * Instead of clearing all references at force_empty, we remember | |
4043 | * the number of reference from swap_cgroup and free mem_cgroup when | |
4044 | * it goes down to 0. | |
4045 | * | |
4046 | * Removal of cgroup itself succeeds regardless of refs from swap. | |
4047 | */ | |
4048 | ||
4049 | static void __mem_cgroup_free(struct mem_cgroup *mem) | |
4050 | { | |
4051 | int node; | |
4052 | ||
4053 | mem_cgroup_remove_from_trees(mem); | |
4054 | free_css_id(&mem_cgroup_subsys, &mem->css); | |
4055 | ||
4056 | for_each_node_state(node, N_POSSIBLE) | |
4057 | free_mem_cgroup_per_zone_info(mem, node); | |
4058 | ||
4059 | free_percpu(mem->stat); | |
4060 | if (sizeof(struct mem_cgroup) < PAGE_SIZE) | |
4061 | kfree(mem); | |
4062 | else | |
4063 | vfree(mem); | |
4064 | } | |
4065 | ||
4066 | static void mem_cgroup_get(struct mem_cgroup *mem) | |
4067 | { | |
4068 | atomic_inc(&mem->refcnt); | |
4069 | } | |
4070 | ||
4071 | static void __mem_cgroup_put(struct mem_cgroup *mem, int count) | |
4072 | { | |
4073 | if (atomic_sub_and_test(count, &mem->refcnt)) { | |
4074 | struct mem_cgroup *parent = parent_mem_cgroup(mem); | |
4075 | __mem_cgroup_free(mem); | |
4076 | if (parent) | |
4077 | mem_cgroup_put(parent); | |
4078 | } | |
4079 | } | |
4080 | ||
4081 | static void mem_cgroup_put(struct mem_cgroup *mem) | |
4082 | { | |
4083 | __mem_cgroup_put(mem, 1); | |
4084 | } | |
4085 | ||
4086 | /* | |
4087 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. | |
4088 | */ | |
4089 | static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) | |
4090 | { | |
4091 | if (!mem->res.parent) | |
4092 | return NULL; | |
4093 | return mem_cgroup_from_res_counter(mem->res.parent, res); | |
4094 | } | |
4095 | ||
4096 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
4097 | static void __init enable_swap_cgroup(void) | |
4098 | { | |
4099 | if (!mem_cgroup_disabled() && really_do_swap_account) | |
4100 | do_swap_account = 1; | |
4101 | } | |
4102 | #else | |
4103 | static void __init enable_swap_cgroup(void) | |
4104 | { | |
4105 | } | |
4106 | #endif | |
4107 | ||
4108 | static int mem_cgroup_soft_limit_tree_init(void) | |
4109 | { | |
4110 | struct mem_cgroup_tree_per_node *rtpn; | |
4111 | struct mem_cgroup_tree_per_zone *rtpz; | |
4112 | int tmp, node, zone; | |
4113 | ||
4114 | for_each_node_state(node, N_POSSIBLE) { | |
4115 | tmp = node; | |
4116 | if (!node_state(node, N_NORMAL_MEMORY)) | |
4117 | tmp = -1; | |
4118 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); | |
4119 | if (!rtpn) | |
4120 | return 1; | |
4121 | ||
4122 | soft_limit_tree.rb_tree_per_node[node] = rtpn; | |
4123 | ||
4124 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4125 | rtpz = &rtpn->rb_tree_per_zone[zone]; | |
4126 | rtpz->rb_root = RB_ROOT; | |
4127 | spin_lock_init(&rtpz->lock); | |
4128 | } | |
4129 | } | |
4130 | return 0; | |
4131 | } | |
4132 | ||
4133 | static struct cgroup_subsys_state * __ref | |
4134 | mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) | |
4135 | { | |
4136 | struct mem_cgroup *mem, *parent; | |
4137 | long error = -ENOMEM; | |
4138 | int node; | |
4139 | ||
4140 | mem = mem_cgroup_alloc(); | |
4141 | if (!mem) | |
4142 | return ERR_PTR(error); | |
4143 | ||
4144 | for_each_node_state(node, N_POSSIBLE) | |
4145 | if (alloc_mem_cgroup_per_zone_info(mem, node)) | |
4146 | goto free_out; | |
4147 | ||
4148 | /* root ? */ | |
4149 | if (cont->parent == NULL) { | |
4150 | int cpu; | |
4151 | enable_swap_cgroup(); | |
4152 | parent = NULL; | |
4153 | root_mem_cgroup = mem; | |
4154 | if (mem_cgroup_soft_limit_tree_init()) | |
4155 | goto free_out; | |
4156 | for_each_possible_cpu(cpu) { | |
4157 | struct memcg_stock_pcp *stock = | |
4158 | &per_cpu(memcg_stock, cpu); | |
4159 | INIT_WORK(&stock->work, drain_local_stock); | |
4160 | } | |
4161 | hotcpu_notifier(memcg_stock_cpu_callback, 0); | |
4162 | } else { | |
4163 | parent = mem_cgroup_from_cont(cont->parent); | |
4164 | mem->use_hierarchy = parent->use_hierarchy; | |
4165 | mem->oom_kill_disable = parent->oom_kill_disable; | |
4166 | } | |
4167 | ||
4168 | if (parent && parent->use_hierarchy) { | |
4169 | res_counter_init(&mem->res, &parent->res); | |
4170 | res_counter_init(&mem->memsw, &parent->memsw); | |
4171 | /* | |
4172 | * We increment refcnt of the parent to ensure that we can | |
4173 | * safely access it on res_counter_charge/uncharge. | |
4174 | * This refcnt will be decremented when freeing this | |
4175 | * mem_cgroup(see mem_cgroup_put). | |
4176 | */ | |
4177 | mem_cgroup_get(parent); | |
4178 | } else { | |
4179 | res_counter_init(&mem->res, NULL); | |
4180 | res_counter_init(&mem->memsw, NULL); | |
4181 | } | |
4182 | mem->last_scanned_child = 0; | |
4183 | spin_lock_init(&mem->reclaim_param_lock); | |
4184 | INIT_LIST_HEAD(&mem->oom_notify); | |
4185 | ||
4186 | if (parent) | |
4187 | mem->swappiness = get_swappiness(parent); | |
4188 | atomic_set(&mem->refcnt, 1); | |
4189 | mem->move_charge_at_immigrate = 0; | |
4190 | mutex_init(&mem->thresholds_lock); | |
4191 | return &mem->css; | |
4192 | free_out: | |
4193 | __mem_cgroup_free(mem); | |
4194 | root_mem_cgroup = NULL; | |
4195 | return ERR_PTR(error); | |
4196 | } | |
4197 | ||
4198 | static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, | |
4199 | struct cgroup *cont) | |
4200 | { | |
4201 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
4202 | ||
4203 | return mem_cgroup_force_empty(mem, false); | |
4204 | } | |
4205 | ||
4206 | static void mem_cgroup_destroy(struct cgroup_subsys *ss, | |
4207 | struct cgroup *cont) | |
4208 | { | |
4209 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
4210 | ||
4211 | mem_cgroup_put(mem); | |
4212 | } | |
4213 | ||
4214 | static int mem_cgroup_populate(struct cgroup_subsys *ss, | |
4215 | struct cgroup *cont) | |
4216 | { | |
4217 | int ret; | |
4218 | ||
4219 | ret = cgroup_add_files(cont, ss, mem_cgroup_files, | |
4220 | ARRAY_SIZE(mem_cgroup_files)); | |
4221 | ||
4222 | if (!ret) | |
4223 | ret = register_memsw_files(cont, ss); | |
4224 | return ret; | |
4225 | } | |
4226 | ||
4227 | #ifdef CONFIG_MMU | |
4228 | /* Handlers for move charge at task migration. */ | |
4229 | #define PRECHARGE_COUNT_AT_ONCE 256 | |
4230 | static int mem_cgroup_do_precharge(unsigned long count) | |
4231 | { | |
4232 | int ret = 0; | |
4233 | int batch_count = PRECHARGE_COUNT_AT_ONCE; | |
4234 | struct mem_cgroup *mem = mc.to; | |
4235 | ||
4236 | if (mem_cgroup_is_root(mem)) { | |
4237 | mc.precharge += count; | |
4238 | /* we don't need css_get for root */ | |
4239 | return ret; | |
4240 | } | |
4241 | /* try to charge at once */ | |
4242 | if (count > 1) { | |
4243 | struct res_counter *dummy; | |
4244 | /* | |
4245 | * "mem" cannot be under rmdir() because we've already checked | |
4246 | * by cgroup_lock_live_cgroup() that it is not removed and we | |
4247 | * are still under the same cgroup_mutex. So we can postpone | |
4248 | * css_get(). | |
4249 | */ | |
4250 | if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy)) | |
4251 | goto one_by_one; | |
4252 | if (do_swap_account && res_counter_charge(&mem->memsw, | |
4253 | PAGE_SIZE * count, &dummy)) { | |
4254 | res_counter_uncharge(&mem->res, PAGE_SIZE * count); | |
4255 | goto one_by_one; | |
4256 | } | |
4257 | mc.precharge += count; | |
4258 | return ret; | |
4259 | } | |
4260 | one_by_one: | |
4261 | /* fall back to one by one charge */ | |
4262 | while (count--) { | |
4263 | if (signal_pending(current)) { | |
4264 | ret = -EINTR; | |
4265 | break; | |
4266 | } | |
4267 | if (!batch_count--) { | |
4268 | batch_count = PRECHARGE_COUNT_AT_ONCE; | |
4269 | cond_resched(); | |
4270 | } | |
4271 | ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false); | |
4272 | if (ret || !mem) | |
4273 | /* mem_cgroup_clear_mc() will do uncharge later */ | |
4274 | return -ENOMEM; | |
4275 | mc.precharge++; | |
4276 | } | |
4277 | return ret; | |
4278 | } | |
4279 | ||
4280 | /** | |
4281 | * is_target_pte_for_mc - check a pte whether it is valid for move charge | |
4282 | * @vma: the vma the pte to be checked belongs | |
4283 | * @addr: the address corresponding to the pte to be checked | |
4284 | * @ptent: the pte to be checked | |
4285 | * @target: the pointer the target page or swap ent will be stored(can be NULL) | |
4286 | * | |
4287 | * Returns | |
4288 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. | |
4289 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for | |
4290 | * move charge. if @target is not NULL, the page is stored in target->page | |
4291 | * with extra refcnt got(Callers should handle it). | |
4292 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a | |
4293 | * target for charge migration. if @target is not NULL, the entry is stored | |
4294 | * in target->ent. | |
4295 | * | |
4296 | * Called with pte lock held. | |
4297 | */ | |
4298 | union mc_target { | |
4299 | struct page *page; | |
4300 | swp_entry_t ent; | |
4301 | }; | |
4302 | ||
4303 | enum mc_target_type { | |
4304 | MC_TARGET_NONE, /* not used */ | |
4305 | MC_TARGET_PAGE, | |
4306 | MC_TARGET_SWAP, | |
4307 | }; | |
4308 | ||
4309 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, | |
4310 | unsigned long addr, pte_t ptent) | |
4311 | { | |
4312 | struct page *page = vm_normal_page(vma, addr, ptent); | |
4313 | ||
4314 | if (!page || !page_mapped(page)) | |
4315 | return NULL; | |
4316 | if (PageAnon(page)) { | |
4317 | /* we don't move shared anon */ | |
4318 | if (!move_anon() || page_mapcount(page) > 2) | |
4319 | return NULL; | |
4320 | } else if (!move_file()) | |
4321 | /* we ignore mapcount for file pages */ | |
4322 | return NULL; | |
4323 | if (!get_page_unless_zero(page)) | |
4324 | return NULL; | |
4325 | ||
4326 | return page; | |
4327 | } | |
4328 | ||
4329 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
4330 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
4331 | { | |
4332 | int usage_count; | |
4333 | struct page *page = NULL; | |
4334 | swp_entry_t ent = pte_to_swp_entry(ptent); | |
4335 | ||
4336 | if (!move_anon() || non_swap_entry(ent)) | |
4337 | return NULL; | |
4338 | usage_count = mem_cgroup_count_swap_user(ent, &page); | |
4339 | if (usage_count > 1) { /* we don't move shared anon */ | |
4340 | if (page) | |
4341 | put_page(page); | |
4342 | return NULL; | |
4343 | } | |
4344 | if (do_swap_account) | |
4345 | entry->val = ent.val; | |
4346 | ||
4347 | return page; | |
4348 | } | |
4349 | ||
4350 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, | |
4351 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
4352 | { | |
4353 | struct page *page = NULL; | |
4354 | struct inode *inode; | |
4355 | struct address_space *mapping; | |
4356 | pgoff_t pgoff; | |
4357 | ||
4358 | if (!vma->vm_file) /* anonymous vma */ | |
4359 | return NULL; | |
4360 | if (!move_file()) | |
4361 | return NULL; | |
4362 | ||
4363 | inode = vma->vm_file->f_path.dentry->d_inode; | |
4364 | mapping = vma->vm_file->f_mapping; | |
4365 | if (pte_none(ptent)) | |
4366 | pgoff = linear_page_index(vma, addr); | |
4367 | else /* pte_file(ptent) is true */ | |
4368 | pgoff = pte_to_pgoff(ptent); | |
4369 | ||
4370 | /* page is moved even if it's not RSS of this task(page-faulted). */ | |
4371 | if (!mapping_cap_swap_backed(mapping)) { /* normal file */ | |
4372 | page = find_get_page(mapping, pgoff); | |
4373 | } else { /* shmem/tmpfs file. we should take account of swap too. */ | |
4374 | swp_entry_t ent; | |
4375 | mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent); | |
4376 | if (do_swap_account) | |
4377 | entry->val = ent.val; | |
4378 | } | |
4379 | ||
4380 | return page; | |
4381 | } | |
4382 | ||
4383 | static int is_target_pte_for_mc(struct vm_area_struct *vma, | |
4384 | unsigned long addr, pte_t ptent, union mc_target *target) | |
4385 | { | |
4386 | struct page *page = NULL; | |
4387 | struct page_cgroup *pc; | |
4388 | int ret = 0; | |
4389 | swp_entry_t ent = { .val = 0 }; | |
4390 | ||
4391 | if (pte_present(ptent)) | |
4392 | page = mc_handle_present_pte(vma, addr, ptent); | |
4393 | else if (is_swap_pte(ptent)) | |
4394 | page = mc_handle_swap_pte(vma, addr, ptent, &ent); | |
4395 | else if (pte_none(ptent) || pte_file(ptent)) | |
4396 | page = mc_handle_file_pte(vma, addr, ptent, &ent); | |
4397 | ||
4398 | if (!page && !ent.val) | |
4399 | return 0; | |
4400 | if (page) { | |
4401 | pc = lookup_page_cgroup(page); | |
4402 | /* | |
4403 | * Do only loose check w/o page_cgroup lock. | |
4404 | * mem_cgroup_move_account() checks the pc is valid or not under | |
4405 | * the lock. | |
4406 | */ | |
4407 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { | |
4408 | ret = MC_TARGET_PAGE; | |
4409 | if (target) | |
4410 | target->page = page; | |
4411 | } | |
4412 | if (!ret || !target) | |
4413 | put_page(page); | |
4414 | } | |
4415 | /* There is a swap entry and a page doesn't exist or isn't charged */ | |
4416 | if (ent.val && !ret && | |
4417 | css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { | |
4418 | ret = MC_TARGET_SWAP; | |
4419 | if (target) | |
4420 | target->ent = ent; | |
4421 | } | |
4422 | return ret; | |
4423 | } | |
4424 | ||
4425 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, | |
4426 | unsigned long addr, unsigned long end, | |
4427 | struct mm_walk *walk) | |
4428 | { | |
4429 | struct vm_area_struct *vma = walk->private; | |
4430 | pte_t *pte; | |
4431 | spinlock_t *ptl; | |
4432 | ||
4433 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
4434 | for (; addr != end; pte++, addr += PAGE_SIZE) | |
4435 | if (is_target_pte_for_mc(vma, addr, *pte, NULL)) | |
4436 | mc.precharge++; /* increment precharge temporarily */ | |
4437 | pte_unmap_unlock(pte - 1, ptl); | |
4438 | cond_resched(); | |
4439 | ||
4440 | return 0; | |
4441 | } | |
4442 | ||
4443 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) | |
4444 | { | |
4445 | unsigned long precharge; | |
4446 | struct vm_area_struct *vma; | |
4447 | ||
4448 | down_read(&mm->mmap_sem); | |
4449 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
4450 | struct mm_walk mem_cgroup_count_precharge_walk = { | |
4451 | .pmd_entry = mem_cgroup_count_precharge_pte_range, | |
4452 | .mm = mm, | |
4453 | .private = vma, | |
4454 | }; | |
4455 | if (is_vm_hugetlb_page(vma)) | |
4456 | continue; | |
4457 | walk_page_range(vma->vm_start, vma->vm_end, | |
4458 | &mem_cgroup_count_precharge_walk); | |
4459 | } | |
4460 | up_read(&mm->mmap_sem); | |
4461 | ||
4462 | precharge = mc.precharge; | |
4463 | mc.precharge = 0; | |
4464 | ||
4465 | return precharge; | |
4466 | } | |
4467 | ||
4468 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) | |
4469 | { | |
4470 | return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm)); | |
4471 | } | |
4472 | ||
4473 | static void mem_cgroup_clear_mc(void) | |
4474 | { | |
4475 | struct mem_cgroup *from = mc.from; | |
4476 | struct mem_cgroup *to = mc.to; | |
4477 | ||
4478 | /* we must uncharge all the leftover precharges from mc.to */ | |
4479 | if (mc.precharge) { | |
4480 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); | |
4481 | mc.precharge = 0; | |
4482 | } | |
4483 | /* | |
4484 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so | |
4485 | * we must uncharge here. | |
4486 | */ | |
4487 | if (mc.moved_charge) { | |
4488 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); | |
4489 | mc.moved_charge = 0; | |
4490 | } | |
4491 | /* we must fixup refcnts and charges */ | |
4492 | if (mc.moved_swap) { | |
4493 | /* uncharge swap account from the old cgroup */ | |
4494 | if (!mem_cgroup_is_root(mc.from)) | |
4495 | res_counter_uncharge(&mc.from->memsw, | |
4496 | PAGE_SIZE * mc.moved_swap); | |
4497 | __mem_cgroup_put(mc.from, mc.moved_swap); | |
4498 | ||
4499 | if (!mem_cgroup_is_root(mc.to)) { | |
4500 | /* | |
4501 | * we charged both to->res and to->memsw, so we should | |
4502 | * uncharge to->res. | |
4503 | */ | |
4504 | res_counter_uncharge(&mc.to->res, | |
4505 | PAGE_SIZE * mc.moved_swap); | |
4506 | } | |
4507 | /* we've already done mem_cgroup_get(mc.to) */ | |
4508 | ||
4509 | mc.moved_swap = 0; | |
4510 | } | |
4511 | spin_lock(&mc.lock); | |
4512 | mc.from = NULL; | |
4513 | mc.to = NULL; | |
4514 | mc.moving_task = NULL; | |
4515 | spin_unlock(&mc.lock); | |
4516 | memcg_oom_recover(from); | |
4517 | memcg_oom_recover(to); | |
4518 | wake_up_all(&mc.waitq); | |
4519 | } | |
4520 | ||
4521 | static int mem_cgroup_can_attach(struct cgroup_subsys *ss, | |
4522 | struct cgroup *cgroup, | |
4523 | struct task_struct *p, | |
4524 | bool threadgroup) | |
4525 | { | |
4526 | int ret = 0; | |
4527 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup); | |
4528 | ||
4529 | if (mem->move_charge_at_immigrate) { | |
4530 | struct mm_struct *mm; | |
4531 | struct mem_cgroup *from = mem_cgroup_from_task(p); | |
4532 | ||
4533 | VM_BUG_ON(from == mem); | |
4534 | ||
4535 | mm = get_task_mm(p); | |
4536 | if (!mm) | |
4537 | return 0; | |
4538 | /* We move charges only when we move a owner of the mm */ | |
4539 | if (mm->owner == p) { | |
4540 | VM_BUG_ON(mc.from); | |
4541 | VM_BUG_ON(mc.to); | |
4542 | VM_BUG_ON(mc.precharge); | |
4543 | VM_BUG_ON(mc.moved_charge); | |
4544 | VM_BUG_ON(mc.moved_swap); | |
4545 | VM_BUG_ON(mc.moving_task); | |
4546 | spin_lock(&mc.lock); | |
4547 | mc.from = from; | |
4548 | mc.to = mem; | |
4549 | mc.precharge = 0; | |
4550 | mc.moved_charge = 0; | |
4551 | mc.moved_swap = 0; | |
4552 | mc.moving_task = current; | |
4553 | spin_unlock(&mc.lock); | |
4554 | ||
4555 | ret = mem_cgroup_precharge_mc(mm); | |
4556 | if (ret) | |
4557 | mem_cgroup_clear_mc(); | |
4558 | } | |
4559 | mmput(mm); | |
4560 | } | |
4561 | return ret; | |
4562 | } | |
4563 | ||
4564 | static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, | |
4565 | struct cgroup *cgroup, | |
4566 | struct task_struct *p, | |
4567 | bool threadgroup) | |
4568 | { | |
4569 | mem_cgroup_clear_mc(); | |
4570 | } | |
4571 | ||
4572 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, | |
4573 | unsigned long addr, unsigned long end, | |
4574 | struct mm_walk *walk) | |
4575 | { | |
4576 | int ret = 0; | |
4577 | struct vm_area_struct *vma = walk->private; | |
4578 | pte_t *pte; | |
4579 | spinlock_t *ptl; | |
4580 | ||
4581 | retry: | |
4582 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
4583 | for (; addr != end; addr += PAGE_SIZE) { | |
4584 | pte_t ptent = *(pte++); | |
4585 | union mc_target target; | |
4586 | int type; | |
4587 | struct page *page; | |
4588 | struct page_cgroup *pc; | |
4589 | swp_entry_t ent; | |
4590 | ||
4591 | if (!mc.precharge) | |
4592 | break; | |
4593 | ||
4594 | type = is_target_pte_for_mc(vma, addr, ptent, &target); | |
4595 | switch (type) { | |
4596 | case MC_TARGET_PAGE: | |
4597 | page = target.page; | |
4598 | if (isolate_lru_page(page)) | |
4599 | goto put; | |
4600 | pc = lookup_page_cgroup(page); | |
4601 | if (!mem_cgroup_move_account(pc, | |
4602 | mc.from, mc.to, false)) { | |
4603 | mc.precharge--; | |
4604 | /* we uncharge from mc.from later. */ | |
4605 | mc.moved_charge++; | |
4606 | } | |
4607 | putback_lru_page(page); | |
4608 | put: /* is_target_pte_for_mc() gets the page */ | |
4609 | put_page(page); | |
4610 | break; | |
4611 | case MC_TARGET_SWAP: | |
4612 | ent = target.ent; | |
4613 | if (!mem_cgroup_move_swap_account(ent, | |
4614 | mc.from, mc.to, false)) { | |
4615 | mc.precharge--; | |
4616 | /* we fixup refcnts and charges later. */ | |
4617 | mc.moved_swap++; | |
4618 | } | |
4619 | break; | |
4620 | default: | |
4621 | break; | |
4622 | } | |
4623 | } | |
4624 | pte_unmap_unlock(pte - 1, ptl); | |
4625 | cond_resched(); | |
4626 | ||
4627 | if (addr != end) { | |
4628 | /* | |
4629 | * We have consumed all precharges we got in can_attach(). | |
4630 | * We try charge one by one, but don't do any additional | |
4631 | * charges to mc.to if we have failed in charge once in attach() | |
4632 | * phase. | |
4633 | */ | |
4634 | ret = mem_cgroup_do_precharge(1); | |
4635 | if (!ret) | |
4636 | goto retry; | |
4637 | } | |
4638 | ||
4639 | return ret; | |
4640 | } | |
4641 | ||
4642 | static void mem_cgroup_move_charge(struct mm_struct *mm) | |
4643 | { | |
4644 | struct vm_area_struct *vma; | |
4645 | ||
4646 | lru_add_drain_all(); | |
4647 | down_read(&mm->mmap_sem); | |
4648 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
4649 | int ret; | |
4650 | struct mm_walk mem_cgroup_move_charge_walk = { | |
4651 | .pmd_entry = mem_cgroup_move_charge_pte_range, | |
4652 | .mm = mm, | |
4653 | .private = vma, | |
4654 | }; | |
4655 | if (is_vm_hugetlb_page(vma)) | |
4656 | continue; | |
4657 | ret = walk_page_range(vma->vm_start, vma->vm_end, | |
4658 | &mem_cgroup_move_charge_walk); | |
4659 | if (ret) | |
4660 | /* | |
4661 | * means we have consumed all precharges and failed in | |
4662 | * doing additional charge. Just abandon here. | |
4663 | */ | |
4664 | break; | |
4665 | } | |
4666 | up_read(&mm->mmap_sem); | |
4667 | } | |
4668 | ||
4669 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, | |
4670 | struct cgroup *cont, | |
4671 | struct cgroup *old_cont, | |
4672 | struct task_struct *p, | |
4673 | bool threadgroup) | |
4674 | { | |
4675 | struct mm_struct *mm; | |
4676 | ||
4677 | if (!mc.to) | |
4678 | /* no need to move charge */ | |
4679 | return; | |
4680 | ||
4681 | mm = get_task_mm(p); | |
4682 | if (mm) { | |
4683 | mem_cgroup_move_charge(mm); | |
4684 | mmput(mm); | |
4685 | } | |
4686 | mem_cgroup_clear_mc(); | |
4687 | } | |
4688 | #else /* !CONFIG_MMU */ | |
4689 | static int mem_cgroup_can_attach(struct cgroup_subsys *ss, | |
4690 | struct cgroup *cgroup, | |
4691 | struct task_struct *p, | |
4692 | bool threadgroup) | |
4693 | { | |
4694 | return 0; | |
4695 | } | |
4696 | static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, | |
4697 | struct cgroup *cgroup, | |
4698 | struct task_struct *p, | |
4699 | bool threadgroup) | |
4700 | { | |
4701 | } | |
4702 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, | |
4703 | struct cgroup *cont, | |
4704 | struct cgroup *old_cont, | |
4705 | struct task_struct *p, | |
4706 | bool threadgroup) | |
4707 | { | |
4708 | } | |
4709 | #endif | |
4710 | ||
4711 | struct cgroup_subsys mem_cgroup_subsys = { | |
4712 | .name = "memory", | |
4713 | .subsys_id = mem_cgroup_subsys_id, | |
4714 | .create = mem_cgroup_create, | |
4715 | .pre_destroy = mem_cgroup_pre_destroy, | |
4716 | .destroy = mem_cgroup_destroy, | |
4717 | .populate = mem_cgroup_populate, | |
4718 | .can_attach = mem_cgroup_can_attach, | |
4719 | .cancel_attach = mem_cgroup_cancel_attach, | |
4720 | .attach = mem_cgroup_move_task, | |
4721 | .early_init = 0, | |
4722 | .use_id = 1, | |
4723 | }; | |
4724 | ||
4725 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
4726 | ||
4727 | static int __init disable_swap_account(char *s) | |
4728 | { | |
4729 | really_do_swap_account = 0; | |
4730 | return 1; | |
4731 | } | |
4732 | __setup("noswapaccount", disable_swap_account); | |
4733 | #endif |