]>
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 | ||
65 | /* for remember boot option*/ | |
66 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED | |
67 | static int really_do_swap_account __initdata = 1; | |
68 | #else | |
69 | static int really_do_swap_account __initdata = 0; | |
70 | #endif | |
71 | ||
72 | #else | |
73 | #define do_swap_account (0) | |
74 | #endif | |
75 | ||
76 | /* | |
77 | * Per memcg event counter is incremented at every pagein/pageout. This counter | |
78 | * is used for trigger some periodic events. This is straightforward and better | |
79 | * than using jiffies etc. to handle periodic memcg event. | |
80 | * | |
81 | * These values will be used as !((event) & ((1 <<(thresh)) - 1)) | |
82 | */ | |
83 | #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */ | |
84 | #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */ | |
85 | ||
86 | /* | |
87 | * Statistics for memory cgroup. | |
88 | */ | |
89 | enum mem_cgroup_stat_index { | |
90 | /* | |
91 | * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. | |
92 | */ | |
93 | MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ | |
94 | MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ | |
95 | MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ | |
96 | MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ | |
97 | MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ | |
98 | MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */ | |
99 | MEM_CGROUP_STAT_NSTATS, | |
100 | }; | |
101 | ||
102 | enum mem_cgroup_events_index { | |
103 | MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ | |
104 | MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ | |
105 | MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */ | |
106 | MEM_CGROUP_EVENTS_NSTATS, | |
107 | }; | |
108 | ||
109 | struct mem_cgroup_stat_cpu { | |
110 | s64 count[MEM_CGROUP_STAT_NSTATS]; | |
111 | unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; | |
112 | }; | |
113 | ||
114 | /* | |
115 | * per-zone information in memory controller. | |
116 | */ | |
117 | struct mem_cgroup_per_zone { | |
118 | /* | |
119 | * spin_lock to protect the per cgroup LRU | |
120 | */ | |
121 | struct list_head lists[NR_LRU_LISTS]; | |
122 | unsigned long count[NR_LRU_LISTS]; | |
123 | ||
124 | struct zone_reclaim_stat reclaim_stat; | |
125 | struct rb_node tree_node; /* RB tree node */ | |
126 | unsigned long long usage_in_excess;/* Set to the value by which */ | |
127 | /* the soft limit is exceeded*/ | |
128 | bool on_tree; | |
129 | struct mem_cgroup *mem; /* Back pointer, we cannot */ | |
130 | /* use container_of */ | |
131 | }; | |
132 | /* Macro for accessing counter */ | |
133 | #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) | |
134 | ||
135 | struct mem_cgroup_per_node { | |
136 | struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; | |
137 | }; | |
138 | ||
139 | struct mem_cgroup_lru_info { | |
140 | struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; | |
141 | }; | |
142 | ||
143 | /* | |
144 | * Cgroups above their limits are maintained in a RB-Tree, independent of | |
145 | * their hierarchy representation | |
146 | */ | |
147 | ||
148 | struct mem_cgroup_tree_per_zone { | |
149 | struct rb_root rb_root; | |
150 | spinlock_t lock; | |
151 | }; | |
152 | ||
153 | struct mem_cgroup_tree_per_node { | |
154 | struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; | |
155 | }; | |
156 | ||
157 | struct mem_cgroup_tree { | |
158 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; | |
159 | }; | |
160 | ||
161 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; | |
162 | ||
163 | struct mem_cgroup_threshold { | |
164 | struct eventfd_ctx *eventfd; | |
165 | u64 threshold; | |
166 | }; | |
167 | ||
168 | /* For threshold */ | |
169 | struct mem_cgroup_threshold_ary { | |
170 | /* An array index points to threshold just below usage. */ | |
171 | int current_threshold; | |
172 | /* Size of entries[] */ | |
173 | unsigned int size; | |
174 | /* Array of thresholds */ | |
175 | struct mem_cgroup_threshold entries[0]; | |
176 | }; | |
177 | ||
178 | struct mem_cgroup_thresholds { | |
179 | /* Primary thresholds array */ | |
180 | struct mem_cgroup_threshold_ary *primary; | |
181 | /* | |
182 | * Spare threshold array. | |
183 | * This is needed to make mem_cgroup_unregister_event() "never fail". | |
184 | * It must be able to store at least primary->size - 1 entries. | |
185 | */ | |
186 | struct mem_cgroup_threshold_ary *spare; | |
187 | }; | |
188 | ||
189 | /* for OOM */ | |
190 | struct mem_cgroup_eventfd_list { | |
191 | struct list_head list; | |
192 | struct eventfd_ctx *eventfd; | |
193 | }; | |
194 | ||
195 | static void mem_cgroup_threshold(struct mem_cgroup *mem); | |
196 | static void mem_cgroup_oom_notify(struct mem_cgroup *mem); | |
197 | ||
198 | /* | |
199 | * The memory controller data structure. The memory controller controls both | |
200 | * page cache and RSS per cgroup. We would eventually like to provide | |
201 | * statistics based on the statistics developed by Rik Van Riel for clock-pro, | |
202 | * to help the administrator determine what knobs to tune. | |
203 | * | |
204 | * TODO: Add a water mark for the memory controller. Reclaim will begin when | |
205 | * we hit the water mark. May be even add a low water mark, such that | |
206 | * no reclaim occurs from a cgroup at it's low water mark, this is | |
207 | * a feature that will be implemented much later in the future. | |
208 | */ | |
209 | struct mem_cgroup { | |
210 | struct cgroup_subsys_state css; | |
211 | /* | |
212 | * the counter to account for memory usage | |
213 | */ | |
214 | struct res_counter res; | |
215 | /* | |
216 | * the counter to account for mem+swap usage. | |
217 | */ | |
218 | struct res_counter memsw; | |
219 | /* | |
220 | * Per cgroup active and inactive list, similar to the | |
221 | * per zone LRU lists. | |
222 | */ | |
223 | struct mem_cgroup_lru_info info; | |
224 | /* | |
225 | * While reclaiming in a hierarchy, we cache the last child we | |
226 | * reclaimed from. | |
227 | */ | |
228 | int last_scanned_child; | |
229 | /* | |
230 | * Should the accounting and control be hierarchical, per subtree? | |
231 | */ | |
232 | bool use_hierarchy; | |
233 | atomic_t oom_lock; | |
234 | atomic_t refcnt; | |
235 | ||
236 | unsigned int swappiness; | |
237 | /* OOM-Killer disable */ | |
238 | int oom_kill_disable; | |
239 | ||
240 | /* set when res.limit == memsw.limit */ | |
241 | bool memsw_is_minimum; | |
242 | ||
243 | /* protect arrays of thresholds */ | |
244 | struct mutex thresholds_lock; | |
245 | ||
246 | /* thresholds for memory usage. RCU-protected */ | |
247 | struct mem_cgroup_thresholds thresholds; | |
248 | ||
249 | /* thresholds for mem+swap usage. RCU-protected */ | |
250 | struct mem_cgroup_thresholds memsw_thresholds; | |
251 | ||
252 | /* For oom notifier event fd */ | |
253 | struct list_head oom_notify; | |
254 | ||
255 | /* | |
256 | * Should we move charges of a task when a task is moved into this | |
257 | * mem_cgroup ? And what type of charges should we move ? | |
258 | */ | |
259 | unsigned long move_charge_at_immigrate; | |
260 | /* | |
261 | * percpu counter. | |
262 | */ | |
263 | struct mem_cgroup_stat_cpu *stat; | |
264 | /* | |
265 | * used when a cpu is offlined or other synchronizations | |
266 | * See mem_cgroup_read_stat(). | |
267 | */ | |
268 | struct mem_cgroup_stat_cpu nocpu_base; | |
269 | spinlock_t pcp_counter_lock; | |
270 | }; | |
271 | ||
272 | /* Stuffs for move charges at task migration. */ | |
273 | /* | |
274 | * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a | |
275 | * left-shifted bitmap of these types. | |
276 | */ | |
277 | enum move_type { | |
278 | MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ | |
279 | MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ | |
280 | NR_MOVE_TYPE, | |
281 | }; | |
282 | ||
283 | /* "mc" and its members are protected by cgroup_mutex */ | |
284 | static struct move_charge_struct { | |
285 | spinlock_t lock; /* for from, to */ | |
286 | struct mem_cgroup *from; | |
287 | struct mem_cgroup *to; | |
288 | unsigned long precharge; | |
289 | unsigned long moved_charge; | |
290 | unsigned long moved_swap; | |
291 | struct task_struct *moving_task; /* a task moving charges */ | |
292 | wait_queue_head_t waitq; /* a waitq for other context */ | |
293 | } mc = { | |
294 | .lock = __SPIN_LOCK_UNLOCKED(mc.lock), | |
295 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), | |
296 | }; | |
297 | ||
298 | static bool move_anon(void) | |
299 | { | |
300 | return test_bit(MOVE_CHARGE_TYPE_ANON, | |
301 | &mc.to->move_charge_at_immigrate); | |
302 | } | |
303 | ||
304 | static bool move_file(void) | |
305 | { | |
306 | return test_bit(MOVE_CHARGE_TYPE_FILE, | |
307 | &mc.to->move_charge_at_immigrate); | |
308 | } | |
309 | ||
310 | /* | |
311 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft | |
312 | * limit reclaim to prevent infinite loops, if they ever occur. | |
313 | */ | |
314 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) | |
315 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) | |
316 | ||
317 | enum charge_type { | |
318 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, | |
319 | MEM_CGROUP_CHARGE_TYPE_MAPPED, | |
320 | MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ | |
321 | MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ | |
322 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ | |
323 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ | |
324 | NR_CHARGE_TYPE, | |
325 | }; | |
326 | ||
327 | /* for encoding cft->private value on file */ | |
328 | #define _MEM (0) | |
329 | #define _MEMSWAP (1) | |
330 | #define _OOM_TYPE (2) | |
331 | #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) | |
332 | #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) | |
333 | #define MEMFILE_ATTR(val) ((val) & 0xffff) | |
334 | /* Used for OOM nofiier */ | |
335 | #define OOM_CONTROL (0) | |
336 | ||
337 | /* | |
338 | * Reclaim flags for mem_cgroup_hierarchical_reclaim | |
339 | */ | |
340 | #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 | |
341 | #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) | |
342 | #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 | |
343 | #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) | |
344 | #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2 | |
345 | #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT) | |
346 | ||
347 | static void mem_cgroup_get(struct mem_cgroup *mem); | |
348 | static void mem_cgroup_put(struct mem_cgroup *mem); | |
349 | static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); | |
350 | static void drain_all_stock_async(void); | |
351 | ||
352 | static struct mem_cgroup_per_zone * | |
353 | mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) | |
354 | { | |
355 | return &mem->info.nodeinfo[nid]->zoneinfo[zid]; | |
356 | } | |
357 | ||
358 | struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem) | |
359 | { | |
360 | return &mem->css; | |
361 | } | |
362 | ||
363 | static struct mem_cgroup_per_zone * | |
364 | page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page) | |
365 | { | |
366 | int nid = page_to_nid(page); | |
367 | int zid = page_zonenum(page); | |
368 | ||
369 | return mem_cgroup_zoneinfo(mem, nid, zid); | |
370 | } | |
371 | ||
372 | static struct mem_cgroup_tree_per_zone * | |
373 | soft_limit_tree_node_zone(int nid, int zid) | |
374 | { | |
375 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
376 | } | |
377 | ||
378 | static struct mem_cgroup_tree_per_zone * | |
379 | soft_limit_tree_from_page(struct page *page) | |
380 | { | |
381 | int nid = page_to_nid(page); | |
382 | int zid = page_zonenum(page); | |
383 | ||
384 | return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; | |
385 | } | |
386 | ||
387 | static void | |
388 | __mem_cgroup_insert_exceeded(struct mem_cgroup *mem, | |
389 | struct mem_cgroup_per_zone *mz, | |
390 | struct mem_cgroup_tree_per_zone *mctz, | |
391 | unsigned long long new_usage_in_excess) | |
392 | { | |
393 | struct rb_node **p = &mctz->rb_root.rb_node; | |
394 | struct rb_node *parent = NULL; | |
395 | struct mem_cgroup_per_zone *mz_node; | |
396 | ||
397 | if (mz->on_tree) | |
398 | return; | |
399 | ||
400 | mz->usage_in_excess = new_usage_in_excess; | |
401 | if (!mz->usage_in_excess) | |
402 | return; | |
403 | while (*p) { | |
404 | parent = *p; | |
405 | mz_node = rb_entry(parent, struct mem_cgroup_per_zone, | |
406 | tree_node); | |
407 | if (mz->usage_in_excess < mz_node->usage_in_excess) | |
408 | p = &(*p)->rb_left; | |
409 | /* | |
410 | * We can't avoid mem cgroups that are over their soft | |
411 | * limit by the same amount | |
412 | */ | |
413 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) | |
414 | p = &(*p)->rb_right; | |
415 | } | |
416 | rb_link_node(&mz->tree_node, parent, p); | |
417 | rb_insert_color(&mz->tree_node, &mctz->rb_root); | |
418 | mz->on_tree = true; | |
419 | } | |
420 | ||
421 | static void | |
422 | __mem_cgroup_remove_exceeded(struct mem_cgroup *mem, | |
423 | struct mem_cgroup_per_zone *mz, | |
424 | struct mem_cgroup_tree_per_zone *mctz) | |
425 | { | |
426 | if (!mz->on_tree) | |
427 | return; | |
428 | rb_erase(&mz->tree_node, &mctz->rb_root); | |
429 | mz->on_tree = false; | |
430 | } | |
431 | ||
432 | static void | |
433 | mem_cgroup_remove_exceeded(struct mem_cgroup *mem, | |
434 | struct mem_cgroup_per_zone *mz, | |
435 | struct mem_cgroup_tree_per_zone *mctz) | |
436 | { | |
437 | spin_lock(&mctz->lock); | |
438 | __mem_cgroup_remove_exceeded(mem, mz, mctz); | |
439 | spin_unlock(&mctz->lock); | |
440 | } | |
441 | ||
442 | ||
443 | static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page) | |
444 | { | |
445 | unsigned long long excess; | |
446 | struct mem_cgroup_per_zone *mz; | |
447 | struct mem_cgroup_tree_per_zone *mctz; | |
448 | int nid = page_to_nid(page); | |
449 | int zid = page_zonenum(page); | |
450 | mctz = soft_limit_tree_from_page(page); | |
451 | ||
452 | /* | |
453 | * Necessary to update all ancestors when hierarchy is used. | |
454 | * because their event counter is not touched. | |
455 | */ | |
456 | for (; mem; mem = parent_mem_cgroup(mem)) { | |
457 | mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
458 | excess = res_counter_soft_limit_excess(&mem->res); | |
459 | /* | |
460 | * We have to update the tree if mz is on RB-tree or | |
461 | * mem is over its softlimit. | |
462 | */ | |
463 | if (excess || mz->on_tree) { | |
464 | spin_lock(&mctz->lock); | |
465 | /* if on-tree, remove it */ | |
466 | if (mz->on_tree) | |
467 | __mem_cgroup_remove_exceeded(mem, mz, mctz); | |
468 | /* | |
469 | * Insert again. mz->usage_in_excess will be updated. | |
470 | * If excess is 0, no tree ops. | |
471 | */ | |
472 | __mem_cgroup_insert_exceeded(mem, mz, mctz, excess); | |
473 | spin_unlock(&mctz->lock); | |
474 | } | |
475 | } | |
476 | } | |
477 | ||
478 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem) | |
479 | { | |
480 | int node, zone; | |
481 | struct mem_cgroup_per_zone *mz; | |
482 | struct mem_cgroup_tree_per_zone *mctz; | |
483 | ||
484 | for_each_node_state(node, N_POSSIBLE) { | |
485 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
486 | mz = mem_cgroup_zoneinfo(mem, node, zone); | |
487 | mctz = soft_limit_tree_node_zone(node, zone); | |
488 | mem_cgroup_remove_exceeded(mem, mz, mctz); | |
489 | } | |
490 | } | |
491 | } | |
492 | ||
493 | static struct mem_cgroup_per_zone * | |
494 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
495 | { | |
496 | struct rb_node *rightmost = NULL; | |
497 | struct mem_cgroup_per_zone *mz; | |
498 | ||
499 | retry: | |
500 | mz = NULL; | |
501 | rightmost = rb_last(&mctz->rb_root); | |
502 | if (!rightmost) | |
503 | goto done; /* Nothing to reclaim from */ | |
504 | ||
505 | mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); | |
506 | /* | |
507 | * Remove the node now but someone else can add it back, | |
508 | * we will to add it back at the end of reclaim to its correct | |
509 | * position in the tree. | |
510 | */ | |
511 | __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); | |
512 | if (!res_counter_soft_limit_excess(&mz->mem->res) || | |
513 | !css_tryget(&mz->mem->css)) | |
514 | goto retry; | |
515 | done: | |
516 | return mz; | |
517 | } | |
518 | ||
519 | static struct mem_cgroup_per_zone * | |
520 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) | |
521 | { | |
522 | struct mem_cgroup_per_zone *mz; | |
523 | ||
524 | spin_lock(&mctz->lock); | |
525 | mz = __mem_cgroup_largest_soft_limit_node(mctz); | |
526 | spin_unlock(&mctz->lock); | |
527 | return mz; | |
528 | } | |
529 | ||
530 | /* | |
531 | * Implementation Note: reading percpu statistics for memcg. | |
532 | * | |
533 | * Both of vmstat[] and percpu_counter has threshold and do periodic | |
534 | * synchronization to implement "quick" read. There are trade-off between | |
535 | * reading cost and precision of value. Then, we may have a chance to implement | |
536 | * a periodic synchronizion of counter in memcg's counter. | |
537 | * | |
538 | * But this _read() function is used for user interface now. The user accounts | |
539 | * memory usage by memory cgroup and he _always_ requires exact value because | |
540 | * he accounts memory. Even if we provide quick-and-fuzzy read, we always | |
541 | * have to visit all online cpus and make sum. So, for now, unnecessary | |
542 | * synchronization is not implemented. (just implemented for cpu hotplug) | |
543 | * | |
544 | * If there are kernel internal actions which can make use of some not-exact | |
545 | * value, and reading all cpu value can be performance bottleneck in some | |
546 | * common workload, threashold and synchonization as vmstat[] should be | |
547 | * implemented. | |
548 | */ | |
549 | static s64 mem_cgroup_read_stat(struct mem_cgroup *mem, | |
550 | enum mem_cgroup_stat_index idx) | |
551 | { | |
552 | int cpu; | |
553 | s64 val = 0; | |
554 | ||
555 | get_online_cpus(); | |
556 | for_each_online_cpu(cpu) | |
557 | val += per_cpu(mem->stat->count[idx], cpu); | |
558 | #ifdef CONFIG_HOTPLUG_CPU | |
559 | spin_lock(&mem->pcp_counter_lock); | |
560 | val += mem->nocpu_base.count[idx]; | |
561 | spin_unlock(&mem->pcp_counter_lock); | |
562 | #endif | |
563 | put_online_cpus(); | |
564 | return val; | |
565 | } | |
566 | ||
567 | static s64 mem_cgroup_local_usage(struct mem_cgroup *mem) | |
568 | { | |
569 | s64 ret; | |
570 | ||
571 | ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); | |
572 | ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); | |
573 | return ret; | |
574 | } | |
575 | ||
576 | static void mem_cgroup_swap_statistics(struct mem_cgroup *mem, | |
577 | bool charge) | |
578 | { | |
579 | int val = (charge) ? 1 : -1; | |
580 | this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); | |
581 | } | |
582 | ||
583 | static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem, | |
584 | enum mem_cgroup_events_index idx) | |
585 | { | |
586 | unsigned long val = 0; | |
587 | int cpu; | |
588 | ||
589 | for_each_online_cpu(cpu) | |
590 | val += per_cpu(mem->stat->events[idx], cpu); | |
591 | #ifdef CONFIG_HOTPLUG_CPU | |
592 | spin_lock(&mem->pcp_counter_lock); | |
593 | val += mem->nocpu_base.events[idx]; | |
594 | spin_unlock(&mem->pcp_counter_lock); | |
595 | #endif | |
596 | return val; | |
597 | } | |
598 | ||
599 | static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, | |
600 | bool file, int nr_pages) | |
601 | { | |
602 | preempt_disable(); | |
603 | ||
604 | if (file) | |
605 | __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages); | |
606 | else | |
607 | __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages); | |
608 | ||
609 | /* pagein of a big page is an event. So, ignore page size */ | |
610 | if (nr_pages > 0) | |
611 | __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); | |
612 | else { | |
613 | __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); | |
614 | nr_pages = -nr_pages; /* for event */ | |
615 | } | |
616 | ||
617 | __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); | |
618 | ||
619 | preempt_enable(); | |
620 | } | |
621 | ||
622 | static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, | |
623 | enum lru_list idx) | |
624 | { | |
625 | int nid, zid; | |
626 | struct mem_cgroup_per_zone *mz; | |
627 | u64 total = 0; | |
628 | ||
629 | for_each_online_node(nid) | |
630 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
631 | mz = mem_cgroup_zoneinfo(mem, nid, zid); | |
632 | total += MEM_CGROUP_ZSTAT(mz, idx); | |
633 | } | |
634 | return total; | |
635 | } | |
636 | ||
637 | static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift) | |
638 | { | |
639 | unsigned long val; | |
640 | ||
641 | val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]); | |
642 | ||
643 | return !(val & ((1 << event_mask_shift) - 1)); | |
644 | } | |
645 | ||
646 | /* | |
647 | * Check events in order. | |
648 | * | |
649 | */ | |
650 | static void memcg_check_events(struct mem_cgroup *mem, struct page *page) | |
651 | { | |
652 | /* threshold event is triggered in finer grain than soft limit */ | |
653 | if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) { | |
654 | mem_cgroup_threshold(mem); | |
655 | if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH))) | |
656 | mem_cgroup_update_tree(mem, page); | |
657 | } | |
658 | } | |
659 | ||
660 | static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) | |
661 | { | |
662 | return container_of(cgroup_subsys_state(cont, | |
663 | mem_cgroup_subsys_id), struct mem_cgroup, | |
664 | css); | |
665 | } | |
666 | ||
667 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) | |
668 | { | |
669 | /* | |
670 | * mm_update_next_owner() may clear mm->owner to NULL | |
671 | * if it races with swapoff, page migration, etc. | |
672 | * So this can be called with p == NULL. | |
673 | */ | |
674 | if (unlikely(!p)) | |
675 | return NULL; | |
676 | ||
677 | return container_of(task_subsys_state(p, mem_cgroup_subsys_id), | |
678 | struct mem_cgroup, css); | |
679 | } | |
680 | ||
681 | static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) | |
682 | { | |
683 | struct mem_cgroup *mem = NULL; | |
684 | ||
685 | if (!mm) | |
686 | return NULL; | |
687 | /* | |
688 | * Because we have no locks, mm->owner's may be being moved to other | |
689 | * cgroup. We use css_tryget() here even if this looks | |
690 | * pessimistic (rather than adding locks here). | |
691 | */ | |
692 | rcu_read_lock(); | |
693 | do { | |
694 | mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
695 | if (unlikely(!mem)) | |
696 | break; | |
697 | } while (!css_tryget(&mem->css)); | |
698 | rcu_read_unlock(); | |
699 | return mem; | |
700 | } | |
701 | ||
702 | /* The caller has to guarantee "mem" exists before calling this */ | |
703 | static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem) | |
704 | { | |
705 | struct cgroup_subsys_state *css; | |
706 | int found; | |
707 | ||
708 | if (!mem) /* ROOT cgroup has the smallest ID */ | |
709 | return root_mem_cgroup; /*css_put/get against root is ignored*/ | |
710 | if (!mem->use_hierarchy) { | |
711 | if (css_tryget(&mem->css)) | |
712 | return mem; | |
713 | return NULL; | |
714 | } | |
715 | rcu_read_lock(); | |
716 | /* | |
717 | * searching a memory cgroup which has the smallest ID under given | |
718 | * ROOT cgroup. (ID >= 1) | |
719 | */ | |
720 | css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found); | |
721 | if (css && css_tryget(css)) | |
722 | mem = container_of(css, struct mem_cgroup, css); | |
723 | else | |
724 | mem = NULL; | |
725 | rcu_read_unlock(); | |
726 | return mem; | |
727 | } | |
728 | ||
729 | static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter, | |
730 | struct mem_cgroup *root, | |
731 | bool cond) | |
732 | { | |
733 | int nextid = css_id(&iter->css) + 1; | |
734 | int found; | |
735 | int hierarchy_used; | |
736 | struct cgroup_subsys_state *css; | |
737 | ||
738 | hierarchy_used = iter->use_hierarchy; | |
739 | ||
740 | css_put(&iter->css); | |
741 | /* If no ROOT, walk all, ignore hierarchy */ | |
742 | if (!cond || (root && !hierarchy_used)) | |
743 | return NULL; | |
744 | ||
745 | if (!root) | |
746 | root = root_mem_cgroup; | |
747 | ||
748 | do { | |
749 | iter = NULL; | |
750 | rcu_read_lock(); | |
751 | ||
752 | css = css_get_next(&mem_cgroup_subsys, nextid, | |
753 | &root->css, &found); | |
754 | if (css && css_tryget(css)) | |
755 | iter = container_of(css, struct mem_cgroup, css); | |
756 | rcu_read_unlock(); | |
757 | /* If css is NULL, no more cgroups will be found */ | |
758 | nextid = found + 1; | |
759 | } while (css && !iter); | |
760 | ||
761 | return iter; | |
762 | } | |
763 | /* | |
764 | * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please | |
765 | * be careful that "break" loop is not allowed. We have reference count. | |
766 | * Instead of that modify "cond" to be false and "continue" to exit the loop. | |
767 | */ | |
768 | #define for_each_mem_cgroup_tree_cond(iter, root, cond) \ | |
769 | for (iter = mem_cgroup_start_loop(root);\ | |
770 | iter != NULL;\ | |
771 | iter = mem_cgroup_get_next(iter, root, cond)) | |
772 | ||
773 | #define for_each_mem_cgroup_tree(iter, root) \ | |
774 | for_each_mem_cgroup_tree_cond(iter, root, true) | |
775 | ||
776 | #define for_each_mem_cgroup_all(iter) \ | |
777 | for_each_mem_cgroup_tree_cond(iter, NULL, true) | |
778 | ||
779 | ||
780 | static inline bool mem_cgroup_is_root(struct mem_cgroup *mem) | |
781 | { | |
782 | return (mem == root_mem_cgroup); | |
783 | } | |
784 | ||
785 | /* | |
786 | * Following LRU functions are allowed to be used without PCG_LOCK. | |
787 | * Operations are called by routine of global LRU independently from memcg. | |
788 | * What we have to take care of here is validness of pc->mem_cgroup. | |
789 | * | |
790 | * Changes to pc->mem_cgroup happens when | |
791 | * 1. charge | |
792 | * 2. moving account | |
793 | * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. | |
794 | * It is added to LRU before charge. | |
795 | * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. | |
796 | * When moving account, the page is not on LRU. It's isolated. | |
797 | */ | |
798 | ||
799 | void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) | |
800 | { | |
801 | struct page_cgroup *pc; | |
802 | struct mem_cgroup_per_zone *mz; | |
803 | ||
804 | if (mem_cgroup_disabled()) | |
805 | return; | |
806 | pc = lookup_page_cgroup(page); | |
807 | /* can happen while we handle swapcache. */ | |
808 | if (!TestClearPageCgroupAcctLRU(pc)) | |
809 | return; | |
810 | VM_BUG_ON(!pc->mem_cgroup); | |
811 | /* | |
812 | * We don't check PCG_USED bit. It's cleared when the "page" is finally | |
813 | * removed from global LRU. | |
814 | */ | |
815 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); | |
816 | /* huge page split is done under lru_lock. so, we have no races. */ | |
817 | MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); | |
818 | if (mem_cgroup_is_root(pc->mem_cgroup)) | |
819 | return; | |
820 | VM_BUG_ON(list_empty(&pc->lru)); | |
821 | list_del_init(&pc->lru); | |
822 | } | |
823 | ||
824 | void mem_cgroup_del_lru(struct page *page) | |
825 | { | |
826 | mem_cgroup_del_lru_list(page, page_lru(page)); | |
827 | } | |
828 | ||
829 | /* | |
830 | * Writeback is about to end against a page which has been marked for immediate | |
831 | * reclaim. If it still appears to be reclaimable, move it to the tail of the | |
832 | * inactive list. | |
833 | */ | |
834 | void mem_cgroup_rotate_reclaimable_page(struct page *page) | |
835 | { | |
836 | struct mem_cgroup_per_zone *mz; | |
837 | struct page_cgroup *pc; | |
838 | enum lru_list lru = page_lru(page); | |
839 | ||
840 | if (mem_cgroup_disabled()) | |
841 | return; | |
842 | ||
843 | pc = lookup_page_cgroup(page); | |
844 | /* unused or root page is not rotated. */ | |
845 | if (!PageCgroupUsed(pc)) | |
846 | return; | |
847 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ | |
848 | smp_rmb(); | |
849 | if (mem_cgroup_is_root(pc->mem_cgroup)) | |
850 | return; | |
851 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); | |
852 | list_move_tail(&pc->lru, &mz->lists[lru]); | |
853 | } | |
854 | ||
855 | void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru) | |
856 | { | |
857 | struct mem_cgroup_per_zone *mz; | |
858 | struct page_cgroup *pc; | |
859 | ||
860 | if (mem_cgroup_disabled()) | |
861 | return; | |
862 | ||
863 | pc = lookup_page_cgroup(page); | |
864 | /* unused or root page is not rotated. */ | |
865 | if (!PageCgroupUsed(pc)) | |
866 | return; | |
867 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ | |
868 | smp_rmb(); | |
869 | if (mem_cgroup_is_root(pc->mem_cgroup)) | |
870 | return; | |
871 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); | |
872 | list_move(&pc->lru, &mz->lists[lru]); | |
873 | } | |
874 | ||
875 | void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru) | |
876 | { | |
877 | struct page_cgroup *pc; | |
878 | struct mem_cgroup_per_zone *mz; | |
879 | ||
880 | if (mem_cgroup_disabled()) | |
881 | return; | |
882 | pc = lookup_page_cgroup(page); | |
883 | VM_BUG_ON(PageCgroupAcctLRU(pc)); | |
884 | if (!PageCgroupUsed(pc)) | |
885 | return; | |
886 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ | |
887 | smp_rmb(); | |
888 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); | |
889 | /* huge page split is done under lru_lock. so, we have no races. */ | |
890 | MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); | |
891 | SetPageCgroupAcctLRU(pc); | |
892 | if (mem_cgroup_is_root(pc->mem_cgroup)) | |
893 | return; | |
894 | list_add(&pc->lru, &mz->lists[lru]); | |
895 | } | |
896 | ||
897 | /* | |
898 | * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to | |
899 | * lru because the page may.be reused after it's fully uncharged (because of | |
900 | * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge | |
901 | * it again. This function is only used to charge SwapCache. It's done under | |
902 | * lock_page and expected that zone->lru_lock is never held. | |
903 | */ | |
904 | static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page) | |
905 | { | |
906 | unsigned long flags; | |
907 | struct zone *zone = page_zone(page); | |
908 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
909 | ||
910 | spin_lock_irqsave(&zone->lru_lock, flags); | |
911 | /* | |
912 | * Forget old LRU when this page_cgroup is *not* used. This Used bit | |
913 | * is guarded by lock_page() because the page is SwapCache. | |
914 | */ | |
915 | if (!PageCgroupUsed(pc)) | |
916 | mem_cgroup_del_lru_list(page, page_lru(page)); | |
917 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
918 | } | |
919 | ||
920 | static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page) | |
921 | { | |
922 | unsigned long flags; | |
923 | struct zone *zone = page_zone(page); | |
924 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
925 | ||
926 | spin_lock_irqsave(&zone->lru_lock, flags); | |
927 | /* link when the page is linked to LRU but page_cgroup isn't */ | |
928 | if (PageLRU(page) && !PageCgroupAcctLRU(pc)) | |
929 | mem_cgroup_add_lru_list(page, page_lru(page)); | |
930 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
931 | } | |
932 | ||
933 | ||
934 | void mem_cgroup_move_lists(struct page *page, | |
935 | enum lru_list from, enum lru_list to) | |
936 | { | |
937 | if (mem_cgroup_disabled()) | |
938 | return; | |
939 | mem_cgroup_del_lru_list(page, from); | |
940 | mem_cgroup_add_lru_list(page, to); | |
941 | } | |
942 | ||
943 | int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) | |
944 | { | |
945 | int ret; | |
946 | struct mem_cgroup *curr = NULL; | |
947 | struct task_struct *p; | |
948 | ||
949 | p = find_lock_task_mm(task); | |
950 | if (!p) | |
951 | return 0; | |
952 | curr = try_get_mem_cgroup_from_mm(p->mm); | |
953 | task_unlock(p); | |
954 | if (!curr) | |
955 | return 0; | |
956 | /* | |
957 | * We should check use_hierarchy of "mem" not "curr". Because checking | |
958 | * use_hierarchy of "curr" here make this function true if hierarchy is | |
959 | * enabled in "curr" and "curr" is a child of "mem" in *cgroup* | |
960 | * hierarchy(even if use_hierarchy is disabled in "mem"). | |
961 | */ | |
962 | if (mem->use_hierarchy) | |
963 | ret = css_is_ancestor(&curr->css, &mem->css); | |
964 | else | |
965 | ret = (curr == mem); | |
966 | css_put(&curr->css); | |
967 | return ret; | |
968 | } | |
969 | ||
970 | static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) | |
971 | { | |
972 | unsigned long active; | |
973 | unsigned long inactive; | |
974 | unsigned long gb; | |
975 | unsigned long inactive_ratio; | |
976 | ||
977 | inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); | |
978 | active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); | |
979 | ||
980 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
981 | if (gb) | |
982 | inactive_ratio = int_sqrt(10 * gb); | |
983 | else | |
984 | inactive_ratio = 1; | |
985 | ||
986 | if (present_pages) { | |
987 | present_pages[0] = inactive; | |
988 | present_pages[1] = active; | |
989 | } | |
990 | ||
991 | return inactive_ratio; | |
992 | } | |
993 | ||
994 | int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) | |
995 | { | |
996 | unsigned long active; | |
997 | unsigned long inactive; | |
998 | unsigned long present_pages[2]; | |
999 | unsigned long inactive_ratio; | |
1000 | ||
1001 | inactive_ratio = calc_inactive_ratio(memcg, present_pages); | |
1002 | ||
1003 | inactive = present_pages[0]; | |
1004 | active = present_pages[1]; | |
1005 | ||
1006 | if (inactive * inactive_ratio < active) | |
1007 | return 1; | |
1008 | ||
1009 | return 0; | |
1010 | } | |
1011 | ||
1012 | int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) | |
1013 | { | |
1014 | unsigned long active; | |
1015 | unsigned long inactive; | |
1016 | ||
1017 | inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); | |
1018 | active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); | |
1019 | ||
1020 | return (active > inactive); | |
1021 | } | |
1022 | ||
1023 | unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg, | |
1024 | struct zone *zone, | |
1025 | enum lru_list lru) | |
1026 | { | |
1027 | int nid = zone_to_nid(zone); | |
1028 | int zid = zone_idx(zone); | |
1029 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
1030 | ||
1031 | return MEM_CGROUP_ZSTAT(mz, lru); | |
1032 | } | |
1033 | ||
1034 | struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, | |
1035 | struct zone *zone) | |
1036 | { | |
1037 | int nid = zone_to_nid(zone); | |
1038 | int zid = zone_idx(zone); | |
1039 | struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); | |
1040 | ||
1041 | return &mz->reclaim_stat; | |
1042 | } | |
1043 | ||
1044 | struct zone_reclaim_stat * | |
1045 | mem_cgroup_get_reclaim_stat_from_page(struct page *page) | |
1046 | { | |
1047 | struct page_cgroup *pc; | |
1048 | struct mem_cgroup_per_zone *mz; | |
1049 | ||
1050 | if (mem_cgroup_disabled()) | |
1051 | return NULL; | |
1052 | ||
1053 | pc = lookup_page_cgroup(page); | |
1054 | if (!PageCgroupUsed(pc)) | |
1055 | return NULL; | |
1056 | /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ | |
1057 | smp_rmb(); | |
1058 | mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); | |
1059 | return &mz->reclaim_stat; | |
1060 | } | |
1061 | ||
1062 | unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, | |
1063 | struct list_head *dst, | |
1064 | unsigned long *scanned, int order, | |
1065 | int mode, struct zone *z, | |
1066 | struct mem_cgroup *mem_cont, | |
1067 | int active, int file) | |
1068 | { | |
1069 | unsigned long nr_taken = 0; | |
1070 | struct page *page; | |
1071 | unsigned long scan; | |
1072 | LIST_HEAD(pc_list); | |
1073 | struct list_head *src; | |
1074 | struct page_cgroup *pc, *tmp; | |
1075 | int nid = zone_to_nid(z); | |
1076 | int zid = zone_idx(z); | |
1077 | struct mem_cgroup_per_zone *mz; | |
1078 | int lru = LRU_FILE * file + active; | |
1079 | int ret; | |
1080 | ||
1081 | BUG_ON(!mem_cont); | |
1082 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); | |
1083 | src = &mz->lists[lru]; | |
1084 | ||
1085 | scan = 0; | |
1086 | list_for_each_entry_safe_reverse(pc, tmp, src, lru) { | |
1087 | if (scan >= nr_to_scan) | |
1088 | break; | |
1089 | ||
1090 | if (unlikely(!PageCgroupUsed(pc))) | |
1091 | continue; | |
1092 | ||
1093 | page = lookup_cgroup_page(pc); | |
1094 | ||
1095 | if (unlikely(!PageLRU(page))) | |
1096 | continue; | |
1097 | ||
1098 | scan++; | |
1099 | ret = __isolate_lru_page(page, mode, file); | |
1100 | switch (ret) { | |
1101 | case 0: | |
1102 | list_move(&page->lru, dst); | |
1103 | mem_cgroup_del_lru(page); | |
1104 | nr_taken += hpage_nr_pages(page); | |
1105 | break; | |
1106 | case -EBUSY: | |
1107 | /* we don't affect global LRU but rotate in our LRU */ | |
1108 | mem_cgroup_rotate_lru_list(page, page_lru(page)); | |
1109 | break; | |
1110 | default: | |
1111 | break; | |
1112 | } | |
1113 | } | |
1114 | ||
1115 | *scanned = scan; | |
1116 | ||
1117 | trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken, | |
1118 | 0, 0, 0, mode); | |
1119 | ||
1120 | return nr_taken; | |
1121 | } | |
1122 | ||
1123 | #define mem_cgroup_from_res_counter(counter, member) \ | |
1124 | container_of(counter, struct mem_cgroup, member) | |
1125 | ||
1126 | /** | |
1127 | * mem_cgroup_margin - calculate chargeable space of a memory cgroup | |
1128 | * @mem: the memory cgroup | |
1129 | * | |
1130 | * Returns the maximum amount of memory @mem can be charged with, in | |
1131 | * pages. | |
1132 | */ | |
1133 | static unsigned long mem_cgroup_margin(struct mem_cgroup *mem) | |
1134 | { | |
1135 | unsigned long long margin; | |
1136 | ||
1137 | margin = res_counter_margin(&mem->res); | |
1138 | if (do_swap_account) | |
1139 | margin = min(margin, res_counter_margin(&mem->memsw)); | |
1140 | return margin >> PAGE_SHIFT; | |
1141 | } | |
1142 | ||
1143 | static unsigned int get_swappiness(struct mem_cgroup *memcg) | |
1144 | { | |
1145 | struct cgroup *cgrp = memcg->css.cgroup; | |
1146 | ||
1147 | /* root ? */ | |
1148 | if (cgrp->parent == NULL) | |
1149 | return vm_swappiness; | |
1150 | ||
1151 | return memcg->swappiness; | |
1152 | } | |
1153 | ||
1154 | static void mem_cgroup_start_move(struct mem_cgroup *mem) | |
1155 | { | |
1156 | int cpu; | |
1157 | ||
1158 | get_online_cpus(); | |
1159 | spin_lock(&mem->pcp_counter_lock); | |
1160 | for_each_online_cpu(cpu) | |
1161 | per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; | |
1162 | mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; | |
1163 | spin_unlock(&mem->pcp_counter_lock); | |
1164 | put_online_cpus(); | |
1165 | ||
1166 | synchronize_rcu(); | |
1167 | } | |
1168 | ||
1169 | static void mem_cgroup_end_move(struct mem_cgroup *mem) | |
1170 | { | |
1171 | int cpu; | |
1172 | ||
1173 | if (!mem) | |
1174 | return; | |
1175 | get_online_cpus(); | |
1176 | spin_lock(&mem->pcp_counter_lock); | |
1177 | for_each_online_cpu(cpu) | |
1178 | per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; | |
1179 | mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; | |
1180 | spin_unlock(&mem->pcp_counter_lock); | |
1181 | put_online_cpus(); | |
1182 | } | |
1183 | /* | |
1184 | * 2 routines for checking "mem" is under move_account() or not. | |
1185 | * | |
1186 | * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used | |
1187 | * for avoiding race in accounting. If true, | |
1188 | * pc->mem_cgroup may be overwritten. | |
1189 | * | |
1190 | * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or | |
1191 | * under hierarchy of moving cgroups. This is for | |
1192 | * waiting at hith-memory prressure caused by "move". | |
1193 | */ | |
1194 | ||
1195 | static bool mem_cgroup_stealed(struct mem_cgroup *mem) | |
1196 | { | |
1197 | VM_BUG_ON(!rcu_read_lock_held()); | |
1198 | return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0; | |
1199 | } | |
1200 | ||
1201 | static bool mem_cgroup_under_move(struct mem_cgroup *mem) | |
1202 | { | |
1203 | struct mem_cgroup *from; | |
1204 | struct mem_cgroup *to; | |
1205 | bool ret = false; | |
1206 | /* | |
1207 | * Unlike task_move routines, we access mc.to, mc.from not under | |
1208 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. | |
1209 | */ | |
1210 | spin_lock(&mc.lock); | |
1211 | from = mc.from; | |
1212 | to = mc.to; | |
1213 | if (!from) | |
1214 | goto unlock; | |
1215 | if (from == mem || to == mem | |
1216 | || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css)) | |
1217 | || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css))) | |
1218 | ret = true; | |
1219 | unlock: | |
1220 | spin_unlock(&mc.lock); | |
1221 | return ret; | |
1222 | } | |
1223 | ||
1224 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem) | |
1225 | { | |
1226 | if (mc.moving_task && current != mc.moving_task) { | |
1227 | if (mem_cgroup_under_move(mem)) { | |
1228 | DEFINE_WAIT(wait); | |
1229 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); | |
1230 | /* moving charge context might have finished. */ | |
1231 | if (mc.moving_task) | |
1232 | schedule(); | |
1233 | finish_wait(&mc.waitq, &wait); | |
1234 | return true; | |
1235 | } | |
1236 | } | |
1237 | return false; | |
1238 | } | |
1239 | ||
1240 | /** | |
1241 | * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. | |
1242 | * @memcg: The memory cgroup that went over limit | |
1243 | * @p: Task that is going to be killed | |
1244 | * | |
1245 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is | |
1246 | * enabled | |
1247 | */ | |
1248 | void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) | |
1249 | { | |
1250 | struct cgroup *task_cgrp; | |
1251 | struct cgroup *mem_cgrp; | |
1252 | /* | |
1253 | * Need a buffer in BSS, can't rely on allocations. The code relies | |
1254 | * on the assumption that OOM is serialized for memory controller. | |
1255 | * If this assumption is broken, revisit this code. | |
1256 | */ | |
1257 | static char memcg_name[PATH_MAX]; | |
1258 | int ret; | |
1259 | ||
1260 | if (!memcg || !p) | |
1261 | return; | |
1262 | ||
1263 | ||
1264 | rcu_read_lock(); | |
1265 | ||
1266 | mem_cgrp = memcg->css.cgroup; | |
1267 | task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); | |
1268 | ||
1269 | ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); | |
1270 | if (ret < 0) { | |
1271 | /* | |
1272 | * Unfortunately, we are unable to convert to a useful name | |
1273 | * But we'll still print out the usage information | |
1274 | */ | |
1275 | rcu_read_unlock(); | |
1276 | goto done; | |
1277 | } | |
1278 | rcu_read_unlock(); | |
1279 | ||
1280 | printk(KERN_INFO "Task in %s killed", memcg_name); | |
1281 | ||
1282 | rcu_read_lock(); | |
1283 | ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); | |
1284 | if (ret < 0) { | |
1285 | rcu_read_unlock(); | |
1286 | goto done; | |
1287 | } | |
1288 | rcu_read_unlock(); | |
1289 | ||
1290 | /* | |
1291 | * Continues from above, so we don't need an KERN_ level | |
1292 | */ | |
1293 | printk(KERN_CONT " as a result of limit of %s\n", memcg_name); | |
1294 | done: | |
1295 | ||
1296 | printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", | |
1297 | res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, | |
1298 | res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, | |
1299 | res_counter_read_u64(&memcg->res, RES_FAILCNT)); | |
1300 | printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " | |
1301 | "failcnt %llu\n", | |
1302 | res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, | |
1303 | res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, | |
1304 | res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); | |
1305 | } | |
1306 | ||
1307 | /* | |
1308 | * This function returns the number of memcg under hierarchy tree. Returns | |
1309 | * 1(self count) if no children. | |
1310 | */ | |
1311 | static int mem_cgroup_count_children(struct mem_cgroup *mem) | |
1312 | { | |
1313 | int num = 0; | |
1314 | struct mem_cgroup *iter; | |
1315 | ||
1316 | for_each_mem_cgroup_tree(iter, mem) | |
1317 | num++; | |
1318 | return num; | |
1319 | } | |
1320 | ||
1321 | /* | |
1322 | * Return the memory (and swap, if configured) limit for a memcg. | |
1323 | */ | |
1324 | u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) | |
1325 | { | |
1326 | u64 limit; | |
1327 | u64 memsw; | |
1328 | ||
1329 | limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
1330 | limit += total_swap_pages << PAGE_SHIFT; | |
1331 | ||
1332 | memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
1333 | /* | |
1334 | * If memsw is finite and limits the amount of swap space available | |
1335 | * to this memcg, return that limit. | |
1336 | */ | |
1337 | return min(limit, memsw); | |
1338 | } | |
1339 | ||
1340 | /* | |
1341 | * Visit the first child (need not be the first child as per the ordering | |
1342 | * of the cgroup list, since we track last_scanned_child) of @mem and use | |
1343 | * that to reclaim free pages from. | |
1344 | */ | |
1345 | static struct mem_cgroup * | |
1346 | mem_cgroup_select_victim(struct mem_cgroup *root_mem) | |
1347 | { | |
1348 | struct mem_cgroup *ret = NULL; | |
1349 | struct cgroup_subsys_state *css; | |
1350 | int nextid, found; | |
1351 | ||
1352 | if (!root_mem->use_hierarchy) { | |
1353 | css_get(&root_mem->css); | |
1354 | ret = root_mem; | |
1355 | } | |
1356 | ||
1357 | while (!ret) { | |
1358 | rcu_read_lock(); | |
1359 | nextid = root_mem->last_scanned_child + 1; | |
1360 | css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, | |
1361 | &found); | |
1362 | if (css && css_tryget(css)) | |
1363 | ret = container_of(css, struct mem_cgroup, css); | |
1364 | ||
1365 | rcu_read_unlock(); | |
1366 | /* Updates scanning parameter */ | |
1367 | if (!css) { | |
1368 | /* this means start scan from ID:1 */ | |
1369 | root_mem->last_scanned_child = 0; | |
1370 | } else | |
1371 | root_mem->last_scanned_child = found; | |
1372 | } | |
1373 | ||
1374 | return ret; | |
1375 | } | |
1376 | ||
1377 | /* | |
1378 | * Scan the hierarchy if needed to reclaim memory. We remember the last child | |
1379 | * we reclaimed from, so that we don't end up penalizing one child extensively | |
1380 | * based on its position in the children list. | |
1381 | * | |
1382 | * root_mem is the original ancestor that we've been reclaim from. | |
1383 | * | |
1384 | * We give up and return to the caller when we visit root_mem twice. | |
1385 | * (other groups can be removed while we're walking....) | |
1386 | * | |
1387 | * If shrink==true, for avoiding to free too much, this returns immedieately. | |
1388 | */ | |
1389 | static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, | |
1390 | struct zone *zone, | |
1391 | gfp_t gfp_mask, | |
1392 | unsigned long reclaim_options) | |
1393 | { | |
1394 | struct mem_cgroup *victim; | |
1395 | int ret, total = 0; | |
1396 | int loop = 0; | |
1397 | bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP; | |
1398 | bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK; | |
1399 | bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT; | |
1400 | unsigned long excess; | |
1401 | ||
1402 | excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT; | |
1403 | ||
1404 | /* If memsw_is_minimum==1, swap-out is of-no-use. */ | |
1405 | if (root_mem->memsw_is_minimum) | |
1406 | noswap = true; | |
1407 | ||
1408 | while (1) { | |
1409 | victim = mem_cgroup_select_victim(root_mem); | |
1410 | if (victim == root_mem) { | |
1411 | loop++; | |
1412 | if (loop >= 1) | |
1413 | drain_all_stock_async(); | |
1414 | if (loop >= 2) { | |
1415 | /* | |
1416 | * If we have not been able to reclaim | |
1417 | * anything, it might because there are | |
1418 | * no reclaimable pages under this hierarchy | |
1419 | */ | |
1420 | if (!check_soft || !total) { | |
1421 | css_put(&victim->css); | |
1422 | break; | |
1423 | } | |
1424 | /* | |
1425 | * We want to do more targetted reclaim. | |
1426 | * excess >> 2 is not to excessive so as to | |
1427 | * reclaim too much, nor too less that we keep | |
1428 | * coming back to reclaim from this cgroup | |
1429 | */ | |
1430 | if (total >= (excess >> 2) || | |
1431 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) { | |
1432 | css_put(&victim->css); | |
1433 | break; | |
1434 | } | |
1435 | } | |
1436 | } | |
1437 | if (!mem_cgroup_local_usage(victim)) { | |
1438 | /* this cgroup's local usage == 0 */ | |
1439 | css_put(&victim->css); | |
1440 | continue; | |
1441 | } | |
1442 | /* we use swappiness of local cgroup */ | |
1443 | if (check_soft) | |
1444 | ret = mem_cgroup_shrink_node_zone(victim, gfp_mask, | |
1445 | noswap, get_swappiness(victim), zone); | |
1446 | else | |
1447 | ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, | |
1448 | noswap, get_swappiness(victim)); | |
1449 | css_put(&victim->css); | |
1450 | /* | |
1451 | * At shrinking usage, we can't check we should stop here or | |
1452 | * reclaim more. It's depends on callers. last_scanned_child | |
1453 | * will work enough for keeping fairness under tree. | |
1454 | */ | |
1455 | if (shrink) | |
1456 | return ret; | |
1457 | total += ret; | |
1458 | if (check_soft) { | |
1459 | if (!res_counter_soft_limit_excess(&root_mem->res)) | |
1460 | return total; | |
1461 | } else if (mem_cgroup_margin(root_mem)) | |
1462 | return 1 + total; | |
1463 | } | |
1464 | return total; | |
1465 | } | |
1466 | ||
1467 | /* | |
1468 | * Check OOM-Killer is already running under our hierarchy. | |
1469 | * If someone is running, return false. | |
1470 | */ | |
1471 | static bool mem_cgroup_oom_lock(struct mem_cgroup *mem) | |
1472 | { | |
1473 | int x, lock_count = 0; | |
1474 | struct mem_cgroup *iter; | |
1475 | ||
1476 | for_each_mem_cgroup_tree(iter, mem) { | |
1477 | x = atomic_inc_return(&iter->oom_lock); | |
1478 | lock_count = max(x, lock_count); | |
1479 | } | |
1480 | ||
1481 | if (lock_count == 1) | |
1482 | return true; | |
1483 | return false; | |
1484 | } | |
1485 | ||
1486 | static int mem_cgroup_oom_unlock(struct mem_cgroup *mem) | |
1487 | { | |
1488 | struct mem_cgroup *iter; | |
1489 | ||
1490 | /* | |
1491 | * When a new child is created while the hierarchy is under oom, | |
1492 | * mem_cgroup_oom_lock() may not be called. We have to use | |
1493 | * atomic_add_unless() here. | |
1494 | */ | |
1495 | for_each_mem_cgroup_tree(iter, mem) | |
1496 | atomic_add_unless(&iter->oom_lock, -1, 0); | |
1497 | return 0; | |
1498 | } | |
1499 | ||
1500 | ||
1501 | static DEFINE_MUTEX(memcg_oom_mutex); | |
1502 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); | |
1503 | ||
1504 | struct oom_wait_info { | |
1505 | struct mem_cgroup *mem; | |
1506 | wait_queue_t wait; | |
1507 | }; | |
1508 | ||
1509 | static int memcg_oom_wake_function(wait_queue_t *wait, | |
1510 | unsigned mode, int sync, void *arg) | |
1511 | { | |
1512 | struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg; | |
1513 | struct oom_wait_info *oom_wait_info; | |
1514 | ||
1515 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); | |
1516 | ||
1517 | if (oom_wait_info->mem == wake_mem) | |
1518 | goto wakeup; | |
1519 | /* if no hierarchy, no match */ | |
1520 | if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy) | |
1521 | return 0; | |
1522 | /* | |
1523 | * Both of oom_wait_info->mem and wake_mem are stable under us. | |
1524 | * Then we can use css_is_ancestor without taking care of RCU. | |
1525 | */ | |
1526 | if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) && | |
1527 | !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css)) | |
1528 | return 0; | |
1529 | ||
1530 | wakeup: | |
1531 | return autoremove_wake_function(wait, mode, sync, arg); | |
1532 | } | |
1533 | ||
1534 | static void memcg_wakeup_oom(struct mem_cgroup *mem) | |
1535 | { | |
1536 | /* for filtering, pass "mem" as argument. */ | |
1537 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem); | |
1538 | } | |
1539 | ||
1540 | static void memcg_oom_recover(struct mem_cgroup *mem) | |
1541 | { | |
1542 | if (mem && atomic_read(&mem->oom_lock)) | |
1543 | memcg_wakeup_oom(mem); | |
1544 | } | |
1545 | ||
1546 | /* | |
1547 | * try to call OOM killer. returns false if we should exit memory-reclaim loop. | |
1548 | */ | |
1549 | bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask) | |
1550 | { | |
1551 | struct oom_wait_info owait; | |
1552 | bool locked, need_to_kill; | |
1553 | ||
1554 | owait.mem = mem; | |
1555 | owait.wait.flags = 0; | |
1556 | owait.wait.func = memcg_oom_wake_function; | |
1557 | owait.wait.private = current; | |
1558 | INIT_LIST_HEAD(&owait.wait.task_list); | |
1559 | need_to_kill = true; | |
1560 | /* At first, try to OOM lock hierarchy under mem.*/ | |
1561 | mutex_lock(&memcg_oom_mutex); | |
1562 | locked = mem_cgroup_oom_lock(mem); | |
1563 | /* | |
1564 | * Even if signal_pending(), we can't quit charge() loop without | |
1565 | * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL | |
1566 | * under OOM is always welcomed, use TASK_KILLABLE here. | |
1567 | */ | |
1568 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); | |
1569 | if (!locked || mem->oom_kill_disable) | |
1570 | need_to_kill = false; | |
1571 | if (locked) | |
1572 | mem_cgroup_oom_notify(mem); | |
1573 | mutex_unlock(&memcg_oom_mutex); | |
1574 | ||
1575 | if (need_to_kill) { | |
1576 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1577 | mem_cgroup_out_of_memory(mem, mask); | |
1578 | } else { | |
1579 | schedule(); | |
1580 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1581 | } | |
1582 | mutex_lock(&memcg_oom_mutex); | |
1583 | mem_cgroup_oom_unlock(mem); | |
1584 | memcg_wakeup_oom(mem); | |
1585 | mutex_unlock(&memcg_oom_mutex); | |
1586 | ||
1587 | if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) | |
1588 | return false; | |
1589 | /* Give chance to dying process */ | |
1590 | schedule_timeout(1); | |
1591 | return true; | |
1592 | } | |
1593 | ||
1594 | /* | |
1595 | * Currently used to update mapped file statistics, but the routine can be | |
1596 | * generalized to update other statistics as well. | |
1597 | * | |
1598 | * Notes: Race condition | |
1599 | * | |
1600 | * We usually use page_cgroup_lock() for accessing page_cgroup member but | |
1601 | * it tends to be costly. But considering some conditions, we doesn't need | |
1602 | * to do so _always_. | |
1603 | * | |
1604 | * Considering "charge", lock_page_cgroup() is not required because all | |
1605 | * file-stat operations happen after a page is attached to radix-tree. There | |
1606 | * are no race with "charge". | |
1607 | * | |
1608 | * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup | |
1609 | * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even | |
1610 | * if there are race with "uncharge". Statistics itself is properly handled | |
1611 | * by flags. | |
1612 | * | |
1613 | * Considering "move", this is an only case we see a race. To make the race | |
1614 | * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are | |
1615 | * possibility of race condition. If there is, we take a lock. | |
1616 | */ | |
1617 | ||
1618 | void mem_cgroup_update_page_stat(struct page *page, | |
1619 | enum mem_cgroup_page_stat_item idx, int val) | |
1620 | { | |
1621 | struct mem_cgroup *mem; | |
1622 | struct page_cgroup *pc = lookup_page_cgroup(page); | |
1623 | bool need_unlock = false; | |
1624 | unsigned long uninitialized_var(flags); | |
1625 | ||
1626 | if (unlikely(!pc)) | |
1627 | return; | |
1628 | ||
1629 | rcu_read_lock(); | |
1630 | mem = pc->mem_cgroup; | |
1631 | if (unlikely(!mem || !PageCgroupUsed(pc))) | |
1632 | goto out; | |
1633 | /* pc->mem_cgroup is unstable ? */ | |
1634 | if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) { | |
1635 | /* take a lock against to access pc->mem_cgroup */ | |
1636 | move_lock_page_cgroup(pc, &flags); | |
1637 | need_unlock = true; | |
1638 | mem = pc->mem_cgroup; | |
1639 | if (!mem || !PageCgroupUsed(pc)) | |
1640 | goto out; | |
1641 | } | |
1642 | ||
1643 | switch (idx) { | |
1644 | case MEMCG_NR_FILE_MAPPED: | |
1645 | if (val > 0) | |
1646 | SetPageCgroupFileMapped(pc); | |
1647 | else if (!page_mapped(page)) | |
1648 | ClearPageCgroupFileMapped(pc); | |
1649 | idx = MEM_CGROUP_STAT_FILE_MAPPED; | |
1650 | break; | |
1651 | default: | |
1652 | BUG(); | |
1653 | } | |
1654 | ||
1655 | this_cpu_add(mem->stat->count[idx], val); | |
1656 | ||
1657 | out: | |
1658 | if (unlikely(need_unlock)) | |
1659 | move_unlock_page_cgroup(pc, &flags); | |
1660 | rcu_read_unlock(); | |
1661 | return; | |
1662 | } | |
1663 | EXPORT_SYMBOL(mem_cgroup_update_page_stat); | |
1664 | ||
1665 | /* | |
1666 | * size of first charge trial. "32" comes from vmscan.c's magic value. | |
1667 | * TODO: maybe necessary to use big numbers in big irons. | |
1668 | */ | |
1669 | #define CHARGE_BATCH 32U | |
1670 | struct memcg_stock_pcp { | |
1671 | struct mem_cgroup *cached; /* this never be root cgroup */ | |
1672 | unsigned int nr_pages; | |
1673 | struct work_struct work; | |
1674 | }; | |
1675 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); | |
1676 | static atomic_t memcg_drain_count; | |
1677 | ||
1678 | /* | |
1679 | * Try to consume stocked charge on this cpu. If success, one page is consumed | |
1680 | * from local stock and true is returned. If the stock is 0 or charges from a | |
1681 | * cgroup which is not current target, returns false. This stock will be | |
1682 | * refilled. | |
1683 | */ | |
1684 | static bool consume_stock(struct mem_cgroup *mem) | |
1685 | { | |
1686 | struct memcg_stock_pcp *stock; | |
1687 | bool ret = true; | |
1688 | ||
1689 | stock = &get_cpu_var(memcg_stock); | |
1690 | if (mem == stock->cached && stock->nr_pages) | |
1691 | stock->nr_pages--; | |
1692 | else /* need to call res_counter_charge */ | |
1693 | ret = false; | |
1694 | put_cpu_var(memcg_stock); | |
1695 | return ret; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Returns stocks cached in percpu to res_counter and reset cached information. | |
1700 | */ | |
1701 | static void drain_stock(struct memcg_stock_pcp *stock) | |
1702 | { | |
1703 | struct mem_cgroup *old = stock->cached; | |
1704 | ||
1705 | if (stock->nr_pages) { | |
1706 | unsigned long bytes = stock->nr_pages * PAGE_SIZE; | |
1707 | ||
1708 | res_counter_uncharge(&old->res, bytes); | |
1709 | if (do_swap_account) | |
1710 | res_counter_uncharge(&old->memsw, bytes); | |
1711 | stock->nr_pages = 0; | |
1712 | } | |
1713 | stock->cached = NULL; | |
1714 | } | |
1715 | ||
1716 | /* | |
1717 | * This must be called under preempt disabled or must be called by | |
1718 | * a thread which is pinned to local cpu. | |
1719 | */ | |
1720 | static void drain_local_stock(struct work_struct *dummy) | |
1721 | { | |
1722 | struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); | |
1723 | drain_stock(stock); | |
1724 | } | |
1725 | ||
1726 | /* | |
1727 | * Cache charges(val) which is from res_counter, to local per_cpu area. | |
1728 | * This will be consumed by consume_stock() function, later. | |
1729 | */ | |
1730 | static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages) | |
1731 | { | |
1732 | struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); | |
1733 | ||
1734 | if (stock->cached != mem) { /* reset if necessary */ | |
1735 | drain_stock(stock); | |
1736 | stock->cached = mem; | |
1737 | } | |
1738 | stock->nr_pages += nr_pages; | |
1739 | put_cpu_var(memcg_stock); | |
1740 | } | |
1741 | ||
1742 | /* | |
1743 | * Tries to drain stocked charges in other cpus. This function is asynchronous | |
1744 | * and just put a work per cpu for draining localy on each cpu. Caller can | |
1745 | * expects some charges will be back to res_counter later but cannot wait for | |
1746 | * it. | |
1747 | */ | |
1748 | static void drain_all_stock_async(void) | |
1749 | { | |
1750 | int cpu; | |
1751 | /* This function is for scheduling "drain" in asynchronous way. | |
1752 | * The result of "drain" is not directly handled by callers. Then, | |
1753 | * if someone is calling drain, we don't have to call drain more. | |
1754 | * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if | |
1755 | * there is a race. We just do loose check here. | |
1756 | */ | |
1757 | if (atomic_read(&memcg_drain_count)) | |
1758 | return; | |
1759 | /* Notify other cpus that system-wide "drain" is running */ | |
1760 | atomic_inc(&memcg_drain_count); | |
1761 | get_online_cpus(); | |
1762 | for_each_online_cpu(cpu) { | |
1763 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
1764 | schedule_work_on(cpu, &stock->work); | |
1765 | } | |
1766 | put_online_cpus(); | |
1767 | atomic_dec(&memcg_drain_count); | |
1768 | /* We don't wait for flush_work */ | |
1769 | } | |
1770 | ||
1771 | /* This is a synchronous drain interface. */ | |
1772 | static void drain_all_stock_sync(void) | |
1773 | { | |
1774 | /* called when force_empty is called */ | |
1775 | atomic_inc(&memcg_drain_count); | |
1776 | schedule_on_each_cpu(drain_local_stock); | |
1777 | atomic_dec(&memcg_drain_count); | |
1778 | } | |
1779 | ||
1780 | /* | |
1781 | * This function drains percpu counter value from DEAD cpu and | |
1782 | * move it to local cpu. Note that this function can be preempted. | |
1783 | */ | |
1784 | static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu) | |
1785 | { | |
1786 | int i; | |
1787 | ||
1788 | spin_lock(&mem->pcp_counter_lock); | |
1789 | for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { | |
1790 | s64 x = per_cpu(mem->stat->count[i], cpu); | |
1791 | ||
1792 | per_cpu(mem->stat->count[i], cpu) = 0; | |
1793 | mem->nocpu_base.count[i] += x; | |
1794 | } | |
1795 | for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { | |
1796 | unsigned long x = per_cpu(mem->stat->events[i], cpu); | |
1797 | ||
1798 | per_cpu(mem->stat->events[i], cpu) = 0; | |
1799 | mem->nocpu_base.events[i] += x; | |
1800 | } | |
1801 | /* need to clear ON_MOVE value, works as a kind of lock. */ | |
1802 | per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; | |
1803 | spin_unlock(&mem->pcp_counter_lock); | |
1804 | } | |
1805 | ||
1806 | static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu) | |
1807 | { | |
1808 | int idx = MEM_CGROUP_ON_MOVE; | |
1809 | ||
1810 | spin_lock(&mem->pcp_counter_lock); | |
1811 | per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx]; | |
1812 | spin_unlock(&mem->pcp_counter_lock); | |
1813 | } | |
1814 | ||
1815 | static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, | |
1816 | unsigned long action, | |
1817 | void *hcpu) | |
1818 | { | |
1819 | int cpu = (unsigned long)hcpu; | |
1820 | struct memcg_stock_pcp *stock; | |
1821 | struct mem_cgroup *iter; | |
1822 | ||
1823 | if ((action == CPU_ONLINE)) { | |
1824 | for_each_mem_cgroup_all(iter) | |
1825 | synchronize_mem_cgroup_on_move(iter, cpu); | |
1826 | return NOTIFY_OK; | |
1827 | } | |
1828 | ||
1829 | if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) | |
1830 | return NOTIFY_OK; | |
1831 | ||
1832 | for_each_mem_cgroup_all(iter) | |
1833 | mem_cgroup_drain_pcp_counter(iter, cpu); | |
1834 | ||
1835 | stock = &per_cpu(memcg_stock, cpu); | |
1836 | drain_stock(stock); | |
1837 | return NOTIFY_OK; | |
1838 | } | |
1839 | ||
1840 | ||
1841 | /* See __mem_cgroup_try_charge() for details */ | |
1842 | enum { | |
1843 | CHARGE_OK, /* success */ | |
1844 | CHARGE_RETRY, /* need to retry but retry is not bad */ | |
1845 | CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ | |
1846 | CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ | |
1847 | CHARGE_OOM_DIE, /* the current is killed because of OOM */ | |
1848 | }; | |
1849 | ||
1850 | static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask, | |
1851 | unsigned int nr_pages, bool oom_check) | |
1852 | { | |
1853 | unsigned long csize = nr_pages * PAGE_SIZE; | |
1854 | struct mem_cgroup *mem_over_limit; | |
1855 | struct res_counter *fail_res; | |
1856 | unsigned long flags = 0; | |
1857 | int ret; | |
1858 | ||
1859 | ret = res_counter_charge(&mem->res, csize, &fail_res); | |
1860 | ||
1861 | if (likely(!ret)) { | |
1862 | if (!do_swap_account) | |
1863 | return CHARGE_OK; | |
1864 | ret = res_counter_charge(&mem->memsw, csize, &fail_res); | |
1865 | if (likely(!ret)) | |
1866 | return CHARGE_OK; | |
1867 | ||
1868 | res_counter_uncharge(&mem->res, csize); | |
1869 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); | |
1870 | flags |= MEM_CGROUP_RECLAIM_NOSWAP; | |
1871 | } else | |
1872 | mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); | |
1873 | /* | |
1874 | * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch | |
1875 | * of regular pages (CHARGE_BATCH), or a single regular page (1). | |
1876 | * | |
1877 | * Never reclaim on behalf of optional batching, retry with a | |
1878 | * single page instead. | |
1879 | */ | |
1880 | if (nr_pages == CHARGE_BATCH) | |
1881 | return CHARGE_RETRY; | |
1882 | ||
1883 | if (!(gfp_mask & __GFP_WAIT)) | |
1884 | return CHARGE_WOULDBLOCK; | |
1885 | ||
1886 | ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL, | |
1887 | gfp_mask, flags); | |
1888 | if (mem_cgroup_margin(mem_over_limit) >= nr_pages) | |
1889 | return CHARGE_RETRY; | |
1890 | /* | |
1891 | * Even though the limit is exceeded at this point, reclaim | |
1892 | * may have been able to free some pages. Retry the charge | |
1893 | * before killing the task. | |
1894 | * | |
1895 | * Only for regular pages, though: huge pages are rather | |
1896 | * unlikely to succeed so close to the limit, and we fall back | |
1897 | * to regular pages anyway in case of failure. | |
1898 | */ | |
1899 | if (nr_pages == 1 && ret) | |
1900 | return CHARGE_RETRY; | |
1901 | ||
1902 | /* | |
1903 | * At task move, charge accounts can be doubly counted. So, it's | |
1904 | * better to wait until the end of task_move if something is going on. | |
1905 | */ | |
1906 | if (mem_cgroup_wait_acct_move(mem_over_limit)) | |
1907 | return CHARGE_RETRY; | |
1908 | ||
1909 | /* If we don't need to call oom-killer at el, return immediately */ | |
1910 | if (!oom_check) | |
1911 | return CHARGE_NOMEM; | |
1912 | /* check OOM */ | |
1913 | if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) | |
1914 | return CHARGE_OOM_DIE; | |
1915 | ||
1916 | return CHARGE_RETRY; | |
1917 | } | |
1918 | ||
1919 | /* | |
1920 | * Unlike exported interface, "oom" parameter is added. if oom==true, | |
1921 | * oom-killer can be invoked. | |
1922 | */ | |
1923 | static int __mem_cgroup_try_charge(struct mm_struct *mm, | |
1924 | gfp_t gfp_mask, | |
1925 | unsigned int nr_pages, | |
1926 | struct mem_cgroup **memcg, | |
1927 | bool oom) | |
1928 | { | |
1929 | unsigned int batch = max(CHARGE_BATCH, nr_pages); | |
1930 | int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
1931 | struct mem_cgroup *mem = NULL; | |
1932 | int ret; | |
1933 | ||
1934 | /* | |
1935 | * Unlike gloval-vm's OOM-kill, we're not in memory shortage | |
1936 | * in system level. So, allow to go ahead dying process in addition to | |
1937 | * MEMDIE process. | |
1938 | */ | |
1939 | if (unlikely(test_thread_flag(TIF_MEMDIE) | |
1940 | || fatal_signal_pending(current))) | |
1941 | goto bypass; | |
1942 | ||
1943 | /* | |
1944 | * We always charge the cgroup the mm_struct belongs to. | |
1945 | * The mm_struct's mem_cgroup changes on task migration if the | |
1946 | * thread group leader migrates. It's possible that mm is not | |
1947 | * set, if so charge the init_mm (happens for pagecache usage). | |
1948 | */ | |
1949 | if (!*memcg && !mm) | |
1950 | goto bypass; | |
1951 | again: | |
1952 | if (*memcg) { /* css should be a valid one */ | |
1953 | mem = *memcg; | |
1954 | VM_BUG_ON(css_is_removed(&mem->css)); | |
1955 | if (mem_cgroup_is_root(mem)) | |
1956 | goto done; | |
1957 | if (nr_pages == 1 && consume_stock(mem)) | |
1958 | goto done; | |
1959 | css_get(&mem->css); | |
1960 | } else { | |
1961 | struct task_struct *p; | |
1962 | ||
1963 | rcu_read_lock(); | |
1964 | p = rcu_dereference(mm->owner); | |
1965 | /* | |
1966 | * Because we don't have task_lock(), "p" can exit. | |
1967 | * In that case, "mem" can point to root or p can be NULL with | |
1968 | * race with swapoff. Then, we have small risk of mis-accouning. | |
1969 | * But such kind of mis-account by race always happens because | |
1970 | * we don't have cgroup_mutex(). It's overkill and we allo that | |
1971 | * small race, here. | |
1972 | * (*) swapoff at el will charge against mm-struct not against | |
1973 | * task-struct. So, mm->owner can be NULL. | |
1974 | */ | |
1975 | mem = mem_cgroup_from_task(p); | |
1976 | if (!mem || mem_cgroup_is_root(mem)) { | |
1977 | rcu_read_unlock(); | |
1978 | goto done; | |
1979 | } | |
1980 | if (nr_pages == 1 && consume_stock(mem)) { | |
1981 | /* | |
1982 | * It seems dagerous to access memcg without css_get(). | |
1983 | * But considering how consume_stok works, it's not | |
1984 | * necessary. If consume_stock success, some charges | |
1985 | * from this memcg are cached on this cpu. So, we | |
1986 | * don't need to call css_get()/css_tryget() before | |
1987 | * calling consume_stock(). | |
1988 | */ | |
1989 | rcu_read_unlock(); | |
1990 | goto done; | |
1991 | } | |
1992 | /* after here, we may be blocked. we need to get refcnt */ | |
1993 | if (!css_tryget(&mem->css)) { | |
1994 | rcu_read_unlock(); | |
1995 | goto again; | |
1996 | } | |
1997 | rcu_read_unlock(); | |
1998 | } | |
1999 | ||
2000 | do { | |
2001 | bool oom_check; | |
2002 | ||
2003 | /* If killed, bypass charge */ | |
2004 | if (fatal_signal_pending(current)) { | |
2005 | css_put(&mem->css); | |
2006 | goto bypass; | |
2007 | } | |
2008 | ||
2009 | oom_check = false; | |
2010 | if (oom && !nr_oom_retries) { | |
2011 | oom_check = true; | |
2012 | nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
2013 | } | |
2014 | ||
2015 | ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check); | |
2016 | switch (ret) { | |
2017 | case CHARGE_OK: | |
2018 | break; | |
2019 | case CHARGE_RETRY: /* not in OOM situation but retry */ | |
2020 | batch = nr_pages; | |
2021 | css_put(&mem->css); | |
2022 | mem = NULL; | |
2023 | goto again; | |
2024 | case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ | |
2025 | css_put(&mem->css); | |
2026 | goto nomem; | |
2027 | case CHARGE_NOMEM: /* OOM routine works */ | |
2028 | if (!oom) { | |
2029 | css_put(&mem->css); | |
2030 | goto nomem; | |
2031 | } | |
2032 | /* If oom, we never return -ENOMEM */ | |
2033 | nr_oom_retries--; | |
2034 | break; | |
2035 | case CHARGE_OOM_DIE: /* Killed by OOM Killer */ | |
2036 | css_put(&mem->css); | |
2037 | goto bypass; | |
2038 | } | |
2039 | } while (ret != CHARGE_OK); | |
2040 | ||
2041 | if (batch > nr_pages) | |
2042 | refill_stock(mem, batch - nr_pages); | |
2043 | css_put(&mem->css); | |
2044 | done: | |
2045 | *memcg = mem; | |
2046 | return 0; | |
2047 | nomem: | |
2048 | *memcg = NULL; | |
2049 | return -ENOMEM; | |
2050 | bypass: | |
2051 | *memcg = NULL; | |
2052 | return 0; | |
2053 | } | |
2054 | ||
2055 | /* | |
2056 | * Somemtimes we have to undo a charge we got by try_charge(). | |
2057 | * This function is for that and do uncharge, put css's refcnt. | |
2058 | * gotten by try_charge(). | |
2059 | */ | |
2060 | static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem, | |
2061 | unsigned int nr_pages) | |
2062 | { | |
2063 | if (!mem_cgroup_is_root(mem)) { | |
2064 | unsigned long bytes = nr_pages * PAGE_SIZE; | |
2065 | ||
2066 | res_counter_uncharge(&mem->res, bytes); | |
2067 | if (do_swap_account) | |
2068 | res_counter_uncharge(&mem->memsw, bytes); | |
2069 | } | |
2070 | } | |
2071 | ||
2072 | /* | |
2073 | * A helper function to get mem_cgroup from ID. must be called under | |
2074 | * rcu_read_lock(). The caller must check css_is_removed() or some if | |
2075 | * it's concern. (dropping refcnt from swap can be called against removed | |
2076 | * memcg.) | |
2077 | */ | |
2078 | static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) | |
2079 | { | |
2080 | struct cgroup_subsys_state *css; | |
2081 | ||
2082 | /* ID 0 is unused ID */ | |
2083 | if (!id) | |
2084 | return NULL; | |
2085 | css = css_lookup(&mem_cgroup_subsys, id); | |
2086 | if (!css) | |
2087 | return NULL; | |
2088 | return container_of(css, struct mem_cgroup, css); | |
2089 | } | |
2090 | ||
2091 | struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) | |
2092 | { | |
2093 | struct mem_cgroup *mem = NULL; | |
2094 | struct page_cgroup *pc; | |
2095 | unsigned short id; | |
2096 | swp_entry_t ent; | |
2097 | ||
2098 | VM_BUG_ON(!PageLocked(page)); | |
2099 | ||
2100 | pc = lookup_page_cgroup(page); | |
2101 | lock_page_cgroup(pc); | |
2102 | if (PageCgroupUsed(pc)) { | |
2103 | mem = pc->mem_cgroup; | |
2104 | if (mem && !css_tryget(&mem->css)) | |
2105 | mem = NULL; | |
2106 | } else if (PageSwapCache(page)) { | |
2107 | ent.val = page_private(page); | |
2108 | id = lookup_swap_cgroup(ent); | |
2109 | rcu_read_lock(); | |
2110 | mem = mem_cgroup_lookup(id); | |
2111 | if (mem && !css_tryget(&mem->css)) | |
2112 | mem = NULL; | |
2113 | rcu_read_unlock(); | |
2114 | } | |
2115 | unlock_page_cgroup(pc); | |
2116 | return mem; | |
2117 | } | |
2118 | ||
2119 | static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, | |
2120 | struct page *page, | |
2121 | unsigned int nr_pages, | |
2122 | struct page_cgroup *pc, | |
2123 | enum charge_type ctype) | |
2124 | { | |
2125 | lock_page_cgroup(pc); | |
2126 | if (unlikely(PageCgroupUsed(pc))) { | |
2127 | unlock_page_cgroup(pc); | |
2128 | __mem_cgroup_cancel_charge(mem, nr_pages); | |
2129 | return; | |
2130 | } | |
2131 | /* | |
2132 | * we don't need page_cgroup_lock about tail pages, becase they are not | |
2133 | * accessed by any other context at this point. | |
2134 | */ | |
2135 | pc->mem_cgroup = mem; | |
2136 | /* | |
2137 | * We access a page_cgroup asynchronously without lock_page_cgroup(). | |
2138 | * Especially when a page_cgroup is taken from a page, pc->mem_cgroup | |
2139 | * is accessed after testing USED bit. To make pc->mem_cgroup visible | |
2140 | * before USED bit, we need memory barrier here. | |
2141 | * See mem_cgroup_add_lru_list(), etc. | |
2142 | */ | |
2143 | smp_wmb(); | |
2144 | switch (ctype) { | |
2145 | case MEM_CGROUP_CHARGE_TYPE_CACHE: | |
2146 | case MEM_CGROUP_CHARGE_TYPE_SHMEM: | |
2147 | SetPageCgroupCache(pc); | |
2148 | SetPageCgroupUsed(pc); | |
2149 | break; | |
2150 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: | |
2151 | ClearPageCgroupCache(pc); | |
2152 | SetPageCgroupUsed(pc); | |
2153 | break; | |
2154 | default: | |
2155 | break; | |
2156 | } | |
2157 | ||
2158 | mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages); | |
2159 | unlock_page_cgroup(pc); | |
2160 | /* | |
2161 | * "charge_statistics" updated event counter. Then, check it. | |
2162 | * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. | |
2163 | * if they exceeds softlimit. | |
2164 | */ | |
2165 | memcg_check_events(mem, page); | |
2166 | } | |
2167 | ||
2168 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
2169 | ||
2170 | #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ | |
2171 | (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION)) | |
2172 | /* | |
2173 | * Because tail pages are not marked as "used", set it. We're under | |
2174 | * zone->lru_lock, 'splitting on pmd' and compund_lock. | |
2175 | */ | |
2176 | void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail) | |
2177 | { | |
2178 | struct page_cgroup *head_pc = lookup_page_cgroup(head); | |
2179 | struct page_cgroup *tail_pc = lookup_page_cgroup(tail); | |
2180 | unsigned long flags; | |
2181 | ||
2182 | if (mem_cgroup_disabled()) | |
2183 | return; | |
2184 | /* | |
2185 | * We have no races with charge/uncharge but will have races with | |
2186 | * page state accounting. | |
2187 | */ | |
2188 | move_lock_page_cgroup(head_pc, &flags); | |
2189 | ||
2190 | tail_pc->mem_cgroup = head_pc->mem_cgroup; | |
2191 | smp_wmb(); /* see __commit_charge() */ | |
2192 | if (PageCgroupAcctLRU(head_pc)) { | |
2193 | enum lru_list lru; | |
2194 | struct mem_cgroup_per_zone *mz; | |
2195 | ||
2196 | /* | |
2197 | * LRU flags cannot be copied because we need to add tail | |
2198 | *.page to LRU by generic call and our hook will be called. | |
2199 | * We hold lru_lock, then, reduce counter directly. | |
2200 | */ | |
2201 | lru = page_lru(head); | |
2202 | mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head); | |
2203 | MEM_CGROUP_ZSTAT(mz, lru) -= 1; | |
2204 | } | |
2205 | tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; | |
2206 | move_unlock_page_cgroup(head_pc, &flags); | |
2207 | } | |
2208 | #endif | |
2209 | ||
2210 | /** | |
2211 | * mem_cgroup_move_account - move account of the page | |
2212 | * @page: the page | |
2213 | * @nr_pages: number of regular pages (>1 for huge pages) | |
2214 | * @pc: page_cgroup of the page. | |
2215 | * @from: mem_cgroup which the page is moved from. | |
2216 | * @to: mem_cgroup which the page is moved to. @from != @to. | |
2217 | * @uncharge: whether we should call uncharge and css_put against @from. | |
2218 | * | |
2219 | * The caller must confirm following. | |
2220 | * - page is not on LRU (isolate_page() is useful.) | |
2221 | * - compound_lock is held when nr_pages > 1 | |
2222 | * | |
2223 | * This function doesn't do "charge" nor css_get to new cgroup. It should be | |
2224 | * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is | |
2225 | * true, this function does "uncharge" from old cgroup, but it doesn't if | |
2226 | * @uncharge is false, so a caller should do "uncharge". | |
2227 | */ | |
2228 | static int mem_cgroup_move_account(struct page *page, | |
2229 | unsigned int nr_pages, | |
2230 | struct page_cgroup *pc, | |
2231 | struct mem_cgroup *from, | |
2232 | struct mem_cgroup *to, | |
2233 | bool uncharge) | |
2234 | { | |
2235 | unsigned long flags; | |
2236 | int ret; | |
2237 | ||
2238 | VM_BUG_ON(from == to); | |
2239 | VM_BUG_ON(PageLRU(page)); | |
2240 | /* | |
2241 | * The page is isolated from LRU. So, collapse function | |
2242 | * will not handle this page. But page splitting can happen. | |
2243 | * Do this check under compound_page_lock(). The caller should | |
2244 | * hold it. | |
2245 | */ | |
2246 | ret = -EBUSY; | |
2247 | if (nr_pages > 1 && !PageTransHuge(page)) | |
2248 | goto out; | |
2249 | ||
2250 | lock_page_cgroup(pc); | |
2251 | ||
2252 | ret = -EINVAL; | |
2253 | if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) | |
2254 | goto unlock; | |
2255 | ||
2256 | move_lock_page_cgroup(pc, &flags); | |
2257 | ||
2258 | if (PageCgroupFileMapped(pc)) { | |
2259 | /* Update mapped_file data for mem_cgroup */ | |
2260 | preempt_disable(); | |
2261 | __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
2262 | __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); | |
2263 | preempt_enable(); | |
2264 | } | |
2265 | mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); | |
2266 | if (uncharge) | |
2267 | /* This is not "cancel", but cancel_charge does all we need. */ | |
2268 | __mem_cgroup_cancel_charge(from, nr_pages); | |
2269 | ||
2270 | /* caller should have done css_get */ | |
2271 | pc->mem_cgroup = to; | |
2272 | mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); | |
2273 | /* | |
2274 | * We charges against "to" which may not have any tasks. Then, "to" | |
2275 | * can be under rmdir(). But in current implementation, caller of | |
2276 | * this function is just force_empty() and move charge, so it's | |
2277 | * garanteed that "to" is never removed. So, we don't check rmdir | |
2278 | * status here. | |
2279 | */ | |
2280 | move_unlock_page_cgroup(pc, &flags); | |
2281 | ret = 0; | |
2282 | unlock: | |
2283 | unlock_page_cgroup(pc); | |
2284 | /* | |
2285 | * check events | |
2286 | */ | |
2287 | memcg_check_events(to, page); | |
2288 | memcg_check_events(from, page); | |
2289 | out: | |
2290 | return ret; | |
2291 | } | |
2292 | ||
2293 | /* | |
2294 | * move charges to its parent. | |
2295 | */ | |
2296 | ||
2297 | static int mem_cgroup_move_parent(struct page *page, | |
2298 | struct page_cgroup *pc, | |
2299 | struct mem_cgroup *child, | |
2300 | gfp_t gfp_mask) | |
2301 | { | |
2302 | struct cgroup *cg = child->css.cgroup; | |
2303 | struct cgroup *pcg = cg->parent; | |
2304 | struct mem_cgroup *parent; | |
2305 | unsigned int nr_pages; | |
2306 | unsigned long flags; | |
2307 | int ret; | |
2308 | ||
2309 | /* Is ROOT ? */ | |
2310 | if (!pcg) | |
2311 | return -EINVAL; | |
2312 | ||
2313 | ret = -EBUSY; | |
2314 | if (!get_page_unless_zero(page)) | |
2315 | goto out; | |
2316 | if (isolate_lru_page(page)) | |
2317 | goto put; | |
2318 | ||
2319 | nr_pages = hpage_nr_pages(page); | |
2320 | ||
2321 | parent = mem_cgroup_from_cont(pcg); | |
2322 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); | |
2323 | if (ret || !parent) | |
2324 | goto put_back; | |
2325 | ||
2326 | if (nr_pages > 1) | |
2327 | flags = compound_lock_irqsave(page); | |
2328 | ||
2329 | ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); | |
2330 | if (ret) | |
2331 | __mem_cgroup_cancel_charge(parent, nr_pages); | |
2332 | ||
2333 | if (nr_pages > 1) | |
2334 | compound_unlock_irqrestore(page, flags); | |
2335 | put_back: | |
2336 | putback_lru_page(page); | |
2337 | put: | |
2338 | put_page(page); | |
2339 | out: | |
2340 | return ret; | |
2341 | } | |
2342 | ||
2343 | /* | |
2344 | * Charge the memory controller for page usage. | |
2345 | * Return | |
2346 | * 0 if the charge was successful | |
2347 | * < 0 if the cgroup is over its limit | |
2348 | */ | |
2349 | static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, | |
2350 | gfp_t gfp_mask, enum charge_type ctype) | |
2351 | { | |
2352 | struct mem_cgroup *mem = NULL; | |
2353 | unsigned int nr_pages = 1; | |
2354 | struct page_cgroup *pc; | |
2355 | bool oom = true; | |
2356 | int ret; | |
2357 | ||
2358 | if (PageTransHuge(page)) { | |
2359 | nr_pages <<= compound_order(page); | |
2360 | VM_BUG_ON(!PageTransHuge(page)); | |
2361 | /* | |
2362 | * Never OOM-kill a process for a huge page. The | |
2363 | * fault handler will fall back to regular pages. | |
2364 | */ | |
2365 | oom = false; | |
2366 | } | |
2367 | ||
2368 | pc = lookup_page_cgroup(page); | |
2369 | BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */ | |
2370 | ||
2371 | ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom); | |
2372 | if (ret || !mem) | |
2373 | return ret; | |
2374 | ||
2375 | __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype); | |
2376 | return 0; | |
2377 | } | |
2378 | ||
2379 | int mem_cgroup_newpage_charge(struct page *page, | |
2380 | struct mm_struct *mm, gfp_t gfp_mask) | |
2381 | { | |
2382 | if (mem_cgroup_disabled()) | |
2383 | return 0; | |
2384 | /* | |
2385 | * If already mapped, we don't have to account. | |
2386 | * If page cache, page->mapping has address_space. | |
2387 | * But page->mapping may have out-of-use anon_vma pointer, | |
2388 | * detecit it by PageAnon() check. newly-mapped-anon's page->mapping | |
2389 | * is NULL. | |
2390 | */ | |
2391 | if (page_mapped(page) || (page->mapping && !PageAnon(page))) | |
2392 | return 0; | |
2393 | if (unlikely(!mm)) | |
2394 | mm = &init_mm; | |
2395 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
2396 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2397 | } | |
2398 | ||
2399 | static void | |
2400 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, | |
2401 | enum charge_type ctype); | |
2402 | ||
2403 | int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, | |
2404 | gfp_t gfp_mask) | |
2405 | { | |
2406 | int ret; | |
2407 | ||
2408 | if (mem_cgroup_disabled()) | |
2409 | return 0; | |
2410 | if (PageCompound(page)) | |
2411 | return 0; | |
2412 | /* | |
2413 | * Corner case handling. This is called from add_to_page_cache() | |
2414 | * in usual. But some FS (shmem) precharges this page before calling it | |
2415 | * and call add_to_page_cache() with GFP_NOWAIT. | |
2416 | * | |
2417 | * For GFP_NOWAIT case, the page may be pre-charged before calling | |
2418 | * add_to_page_cache(). (See shmem.c) check it here and avoid to call | |
2419 | * charge twice. (It works but has to pay a bit larger cost.) | |
2420 | * And when the page is SwapCache, it should take swap information | |
2421 | * into account. This is under lock_page() now. | |
2422 | */ | |
2423 | if (!(gfp_mask & __GFP_WAIT)) { | |
2424 | struct page_cgroup *pc; | |
2425 | ||
2426 | pc = lookup_page_cgroup(page); | |
2427 | if (!pc) | |
2428 | return 0; | |
2429 | lock_page_cgroup(pc); | |
2430 | if (PageCgroupUsed(pc)) { | |
2431 | unlock_page_cgroup(pc); | |
2432 | return 0; | |
2433 | } | |
2434 | unlock_page_cgroup(pc); | |
2435 | } | |
2436 | ||
2437 | if (unlikely(!mm)) | |
2438 | mm = &init_mm; | |
2439 | ||
2440 | if (page_is_file_cache(page)) | |
2441 | return mem_cgroup_charge_common(page, mm, gfp_mask, | |
2442 | MEM_CGROUP_CHARGE_TYPE_CACHE); | |
2443 | ||
2444 | /* shmem */ | |
2445 | if (PageSwapCache(page)) { | |
2446 | struct mem_cgroup *mem; | |
2447 | ||
2448 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); | |
2449 | if (!ret) | |
2450 | __mem_cgroup_commit_charge_swapin(page, mem, | |
2451 | MEM_CGROUP_CHARGE_TYPE_SHMEM); | |
2452 | } else | |
2453 | ret = mem_cgroup_charge_common(page, mm, gfp_mask, | |
2454 | MEM_CGROUP_CHARGE_TYPE_SHMEM); | |
2455 | ||
2456 | return ret; | |
2457 | } | |
2458 | ||
2459 | /* | |
2460 | * While swap-in, try_charge -> commit or cancel, the page is locked. | |
2461 | * And when try_charge() successfully returns, one refcnt to memcg without | |
2462 | * struct page_cgroup is acquired. This refcnt will be consumed by | |
2463 | * "commit()" or removed by "cancel()" | |
2464 | */ | |
2465 | int mem_cgroup_try_charge_swapin(struct mm_struct *mm, | |
2466 | struct page *page, | |
2467 | gfp_t mask, struct mem_cgroup **ptr) | |
2468 | { | |
2469 | struct mem_cgroup *mem; | |
2470 | int ret; | |
2471 | ||
2472 | *ptr = NULL; | |
2473 | ||
2474 | if (mem_cgroup_disabled()) | |
2475 | return 0; | |
2476 | ||
2477 | if (!do_swap_account) | |
2478 | goto charge_cur_mm; | |
2479 | /* | |
2480 | * A racing thread's fault, or swapoff, may have already updated | |
2481 | * the pte, and even removed page from swap cache: in those cases | |
2482 | * do_swap_page()'s pte_same() test will fail; but there's also a | |
2483 | * KSM case which does need to charge the page. | |
2484 | */ | |
2485 | if (!PageSwapCache(page)) | |
2486 | goto charge_cur_mm; | |
2487 | mem = try_get_mem_cgroup_from_page(page); | |
2488 | if (!mem) | |
2489 | goto charge_cur_mm; | |
2490 | *ptr = mem; | |
2491 | ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true); | |
2492 | css_put(&mem->css); | |
2493 | return ret; | |
2494 | charge_cur_mm: | |
2495 | if (unlikely(!mm)) | |
2496 | mm = &init_mm; | |
2497 | return __mem_cgroup_try_charge(mm, mask, 1, ptr, true); | |
2498 | } | |
2499 | ||
2500 | static void | |
2501 | __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, | |
2502 | enum charge_type ctype) | |
2503 | { | |
2504 | struct page_cgroup *pc; | |
2505 | ||
2506 | if (mem_cgroup_disabled()) | |
2507 | return; | |
2508 | if (!ptr) | |
2509 | return; | |
2510 | cgroup_exclude_rmdir(&ptr->css); | |
2511 | pc = lookup_page_cgroup(page); | |
2512 | mem_cgroup_lru_del_before_commit_swapcache(page); | |
2513 | __mem_cgroup_commit_charge(ptr, page, 1, pc, ctype); | |
2514 | mem_cgroup_lru_add_after_commit_swapcache(page); | |
2515 | /* | |
2516 | * Now swap is on-memory. This means this page may be | |
2517 | * counted both as mem and swap....double count. | |
2518 | * Fix it by uncharging from memsw. Basically, this SwapCache is stable | |
2519 | * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() | |
2520 | * may call delete_from_swap_cache() before reach here. | |
2521 | */ | |
2522 | if (do_swap_account && PageSwapCache(page)) { | |
2523 | swp_entry_t ent = {.val = page_private(page)}; | |
2524 | unsigned short id; | |
2525 | struct mem_cgroup *memcg; | |
2526 | ||
2527 | id = swap_cgroup_record(ent, 0); | |
2528 | rcu_read_lock(); | |
2529 | memcg = mem_cgroup_lookup(id); | |
2530 | if (memcg) { | |
2531 | /* | |
2532 | * This recorded memcg can be obsolete one. So, avoid | |
2533 | * calling css_tryget | |
2534 | */ | |
2535 | if (!mem_cgroup_is_root(memcg)) | |
2536 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); | |
2537 | mem_cgroup_swap_statistics(memcg, false); | |
2538 | mem_cgroup_put(memcg); | |
2539 | } | |
2540 | rcu_read_unlock(); | |
2541 | } | |
2542 | /* | |
2543 | * At swapin, we may charge account against cgroup which has no tasks. | |
2544 | * So, rmdir()->pre_destroy() can be called while we do this charge. | |
2545 | * In that case, we need to call pre_destroy() again. check it here. | |
2546 | */ | |
2547 | cgroup_release_and_wakeup_rmdir(&ptr->css); | |
2548 | } | |
2549 | ||
2550 | void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) | |
2551 | { | |
2552 | __mem_cgroup_commit_charge_swapin(page, ptr, | |
2553 | MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2554 | } | |
2555 | ||
2556 | void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) | |
2557 | { | |
2558 | if (mem_cgroup_disabled()) | |
2559 | return; | |
2560 | if (!mem) | |
2561 | return; | |
2562 | __mem_cgroup_cancel_charge(mem, 1); | |
2563 | } | |
2564 | ||
2565 | static void mem_cgroup_do_uncharge(struct mem_cgroup *mem, | |
2566 | unsigned int nr_pages, | |
2567 | const enum charge_type ctype) | |
2568 | { | |
2569 | struct memcg_batch_info *batch = NULL; | |
2570 | bool uncharge_memsw = true; | |
2571 | ||
2572 | /* If swapout, usage of swap doesn't decrease */ | |
2573 | if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) | |
2574 | uncharge_memsw = false; | |
2575 | ||
2576 | batch = ¤t->memcg_batch; | |
2577 | /* | |
2578 | * In usual, we do css_get() when we remember memcg pointer. | |
2579 | * But in this case, we keep res->usage until end of a series of | |
2580 | * uncharges. Then, it's ok to ignore memcg's refcnt. | |
2581 | */ | |
2582 | if (!batch->memcg) | |
2583 | batch->memcg = mem; | |
2584 | /* | |
2585 | * do_batch > 0 when unmapping pages or inode invalidate/truncate. | |
2586 | * In those cases, all pages freed continously can be expected to be in | |
2587 | * the same cgroup and we have chance to coalesce uncharges. | |
2588 | * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) | |
2589 | * because we want to do uncharge as soon as possible. | |
2590 | */ | |
2591 | ||
2592 | if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) | |
2593 | goto direct_uncharge; | |
2594 | ||
2595 | if (nr_pages > 1) | |
2596 | goto direct_uncharge; | |
2597 | ||
2598 | /* | |
2599 | * In typical case, batch->memcg == mem. This means we can | |
2600 | * merge a series of uncharges to an uncharge of res_counter. | |
2601 | * If not, we uncharge res_counter ony by one. | |
2602 | */ | |
2603 | if (batch->memcg != mem) | |
2604 | goto direct_uncharge; | |
2605 | /* remember freed charge and uncharge it later */ | |
2606 | batch->nr_pages++; | |
2607 | if (uncharge_memsw) | |
2608 | batch->memsw_nr_pages++; | |
2609 | return; | |
2610 | direct_uncharge: | |
2611 | res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE); | |
2612 | if (uncharge_memsw) | |
2613 | res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE); | |
2614 | if (unlikely(batch->memcg != mem)) | |
2615 | memcg_oom_recover(mem); | |
2616 | return; | |
2617 | } | |
2618 | ||
2619 | /* | |
2620 | * uncharge if !page_mapped(page) | |
2621 | */ | |
2622 | static struct mem_cgroup * | |
2623 | __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) | |
2624 | { | |
2625 | struct mem_cgroup *mem = NULL; | |
2626 | unsigned int nr_pages = 1; | |
2627 | struct page_cgroup *pc; | |
2628 | ||
2629 | if (mem_cgroup_disabled()) | |
2630 | return NULL; | |
2631 | ||
2632 | if (PageSwapCache(page)) | |
2633 | return NULL; | |
2634 | ||
2635 | if (PageTransHuge(page)) { | |
2636 | nr_pages <<= compound_order(page); | |
2637 | VM_BUG_ON(!PageTransHuge(page)); | |
2638 | } | |
2639 | /* | |
2640 | * Check if our page_cgroup is valid | |
2641 | */ | |
2642 | pc = lookup_page_cgroup(page); | |
2643 | if (unlikely(!pc || !PageCgroupUsed(pc))) | |
2644 | return NULL; | |
2645 | ||
2646 | lock_page_cgroup(pc); | |
2647 | ||
2648 | mem = pc->mem_cgroup; | |
2649 | ||
2650 | if (!PageCgroupUsed(pc)) | |
2651 | goto unlock_out; | |
2652 | ||
2653 | switch (ctype) { | |
2654 | case MEM_CGROUP_CHARGE_TYPE_MAPPED: | |
2655 | case MEM_CGROUP_CHARGE_TYPE_DROP: | |
2656 | /* See mem_cgroup_prepare_migration() */ | |
2657 | if (page_mapped(page) || PageCgroupMigration(pc)) | |
2658 | goto unlock_out; | |
2659 | break; | |
2660 | case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: | |
2661 | if (!PageAnon(page)) { /* Shared memory */ | |
2662 | if (page->mapping && !page_is_file_cache(page)) | |
2663 | goto unlock_out; | |
2664 | } else if (page_mapped(page)) /* Anon */ | |
2665 | goto unlock_out; | |
2666 | break; | |
2667 | default: | |
2668 | break; | |
2669 | } | |
2670 | ||
2671 | mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages); | |
2672 | ||
2673 | ClearPageCgroupUsed(pc); | |
2674 | /* | |
2675 | * pc->mem_cgroup is not cleared here. It will be accessed when it's | |
2676 | * freed from LRU. This is safe because uncharged page is expected not | |
2677 | * to be reused (freed soon). Exception is SwapCache, it's handled by | |
2678 | * special functions. | |
2679 | */ | |
2680 | ||
2681 | unlock_page_cgroup(pc); | |
2682 | /* | |
2683 | * even after unlock, we have mem->res.usage here and this memcg | |
2684 | * will never be freed. | |
2685 | */ | |
2686 | memcg_check_events(mem, page); | |
2687 | if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { | |
2688 | mem_cgroup_swap_statistics(mem, true); | |
2689 | mem_cgroup_get(mem); | |
2690 | } | |
2691 | if (!mem_cgroup_is_root(mem)) | |
2692 | mem_cgroup_do_uncharge(mem, nr_pages, ctype); | |
2693 | ||
2694 | return mem; | |
2695 | ||
2696 | unlock_out: | |
2697 | unlock_page_cgroup(pc); | |
2698 | return NULL; | |
2699 | } | |
2700 | ||
2701 | void mem_cgroup_uncharge_page(struct page *page) | |
2702 | { | |
2703 | /* early check. */ | |
2704 | if (page_mapped(page)) | |
2705 | return; | |
2706 | if (page->mapping && !PageAnon(page)) | |
2707 | return; | |
2708 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); | |
2709 | } | |
2710 | ||
2711 | void mem_cgroup_uncharge_cache_page(struct page *page) | |
2712 | { | |
2713 | VM_BUG_ON(page_mapped(page)); | |
2714 | VM_BUG_ON(page->mapping); | |
2715 | __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); | |
2716 | } | |
2717 | ||
2718 | /* | |
2719 | * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. | |
2720 | * In that cases, pages are freed continuously and we can expect pages | |
2721 | * are in the same memcg. All these calls itself limits the number of | |
2722 | * pages freed at once, then uncharge_start/end() is called properly. | |
2723 | * This may be called prural(2) times in a context, | |
2724 | */ | |
2725 | ||
2726 | void mem_cgroup_uncharge_start(void) | |
2727 | { | |
2728 | current->memcg_batch.do_batch++; | |
2729 | /* We can do nest. */ | |
2730 | if (current->memcg_batch.do_batch == 1) { | |
2731 | current->memcg_batch.memcg = NULL; | |
2732 | current->memcg_batch.nr_pages = 0; | |
2733 | current->memcg_batch.memsw_nr_pages = 0; | |
2734 | } | |
2735 | } | |
2736 | ||
2737 | void mem_cgroup_uncharge_end(void) | |
2738 | { | |
2739 | struct memcg_batch_info *batch = ¤t->memcg_batch; | |
2740 | ||
2741 | if (!batch->do_batch) | |
2742 | return; | |
2743 | ||
2744 | batch->do_batch--; | |
2745 | if (batch->do_batch) /* If stacked, do nothing. */ | |
2746 | return; | |
2747 | ||
2748 | if (!batch->memcg) | |
2749 | return; | |
2750 | /* | |
2751 | * This "batch->memcg" is valid without any css_get/put etc... | |
2752 | * bacause we hide charges behind us. | |
2753 | */ | |
2754 | if (batch->nr_pages) | |
2755 | res_counter_uncharge(&batch->memcg->res, | |
2756 | batch->nr_pages * PAGE_SIZE); | |
2757 | if (batch->memsw_nr_pages) | |
2758 | res_counter_uncharge(&batch->memcg->memsw, | |
2759 | batch->memsw_nr_pages * PAGE_SIZE); | |
2760 | memcg_oom_recover(batch->memcg); | |
2761 | /* forget this pointer (for sanity check) */ | |
2762 | batch->memcg = NULL; | |
2763 | } | |
2764 | ||
2765 | #ifdef CONFIG_SWAP | |
2766 | /* | |
2767 | * called after __delete_from_swap_cache() and drop "page" account. | |
2768 | * memcg information is recorded to swap_cgroup of "ent" | |
2769 | */ | |
2770 | void | |
2771 | mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) | |
2772 | { | |
2773 | struct mem_cgroup *memcg; | |
2774 | int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; | |
2775 | ||
2776 | if (!swapout) /* this was a swap cache but the swap is unused ! */ | |
2777 | ctype = MEM_CGROUP_CHARGE_TYPE_DROP; | |
2778 | ||
2779 | memcg = __mem_cgroup_uncharge_common(page, ctype); | |
2780 | ||
2781 | /* | |
2782 | * record memcg information, if swapout && memcg != NULL, | |
2783 | * mem_cgroup_get() was called in uncharge(). | |
2784 | */ | |
2785 | if (do_swap_account && swapout && memcg) | |
2786 | swap_cgroup_record(ent, css_id(&memcg->css)); | |
2787 | } | |
2788 | #endif | |
2789 | ||
2790 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
2791 | /* | |
2792 | * called from swap_entry_free(). remove record in swap_cgroup and | |
2793 | * uncharge "memsw" account. | |
2794 | */ | |
2795 | void mem_cgroup_uncharge_swap(swp_entry_t ent) | |
2796 | { | |
2797 | struct mem_cgroup *memcg; | |
2798 | unsigned short id; | |
2799 | ||
2800 | if (!do_swap_account) | |
2801 | return; | |
2802 | ||
2803 | id = swap_cgroup_record(ent, 0); | |
2804 | rcu_read_lock(); | |
2805 | memcg = mem_cgroup_lookup(id); | |
2806 | if (memcg) { | |
2807 | /* | |
2808 | * We uncharge this because swap is freed. | |
2809 | * This memcg can be obsolete one. We avoid calling css_tryget | |
2810 | */ | |
2811 | if (!mem_cgroup_is_root(memcg)) | |
2812 | res_counter_uncharge(&memcg->memsw, PAGE_SIZE); | |
2813 | mem_cgroup_swap_statistics(memcg, false); | |
2814 | mem_cgroup_put(memcg); | |
2815 | } | |
2816 | rcu_read_unlock(); | |
2817 | } | |
2818 | ||
2819 | /** | |
2820 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. | |
2821 | * @entry: swap entry to be moved | |
2822 | * @from: mem_cgroup which the entry is moved from | |
2823 | * @to: mem_cgroup which the entry is moved to | |
2824 | * @need_fixup: whether we should fixup res_counters and refcounts. | |
2825 | * | |
2826 | * It succeeds only when the swap_cgroup's record for this entry is the same | |
2827 | * as the mem_cgroup's id of @from. | |
2828 | * | |
2829 | * Returns 0 on success, -EINVAL on failure. | |
2830 | * | |
2831 | * The caller must have charged to @to, IOW, called res_counter_charge() about | |
2832 | * both res and memsw, and called css_get(). | |
2833 | */ | |
2834 | static int mem_cgroup_move_swap_account(swp_entry_t entry, | |
2835 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) | |
2836 | { | |
2837 | unsigned short old_id, new_id; | |
2838 | ||
2839 | old_id = css_id(&from->css); | |
2840 | new_id = css_id(&to->css); | |
2841 | ||
2842 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { | |
2843 | mem_cgroup_swap_statistics(from, false); | |
2844 | mem_cgroup_swap_statistics(to, true); | |
2845 | /* | |
2846 | * This function is only called from task migration context now. | |
2847 | * It postpones res_counter and refcount handling till the end | |
2848 | * of task migration(mem_cgroup_clear_mc()) for performance | |
2849 | * improvement. But we cannot postpone mem_cgroup_get(to) | |
2850 | * because if the process that has been moved to @to does | |
2851 | * swap-in, the refcount of @to might be decreased to 0. | |
2852 | */ | |
2853 | mem_cgroup_get(to); | |
2854 | if (need_fixup) { | |
2855 | if (!mem_cgroup_is_root(from)) | |
2856 | res_counter_uncharge(&from->memsw, PAGE_SIZE); | |
2857 | mem_cgroup_put(from); | |
2858 | /* | |
2859 | * we charged both to->res and to->memsw, so we should | |
2860 | * uncharge to->res. | |
2861 | */ | |
2862 | if (!mem_cgroup_is_root(to)) | |
2863 | res_counter_uncharge(&to->res, PAGE_SIZE); | |
2864 | } | |
2865 | return 0; | |
2866 | } | |
2867 | return -EINVAL; | |
2868 | } | |
2869 | #else | |
2870 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, | |
2871 | struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) | |
2872 | { | |
2873 | return -EINVAL; | |
2874 | } | |
2875 | #endif | |
2876 | ||
2877 | /* | |
2878 | * Before starting migration, account PAGE_SIZE to mem_cgroup that the old | |
2879 | * page belongs to. | |
2880 | */ | |
2881 | int mem_cgroup_prepare_migration(struct page *page, | |
2882 | struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask) | |
2883 | { | |
2884 | struct mem_cgroup *mem = NULL; | |
2885 | struct page_cgroup *pc; | |
2886 | enum charge_type ctype; | |
2887 | int ret = 0; | |
2888 | ||
2889 | *ptr = NULL; | |
2890 | ||
2891 | VM_BUG_ON(PageTransHuge(page)); | |
2892 | if (mem_cgroup_disabled()) | |
2893 | return 0; | |
2894 | ||
2895 | pc = lookup_page_cgroup(page); | |
2896 | lock_page_cgroup(pc); | |
2897 | if (PageCgroupUsed(pc)) { | |
2898 | mem = pc->mem_cgroup; | |
2899 | css_get(&mem->css); | |
2900 | /* | |
2901 | * At migrating an anonymous page, its mapcount goes down | |
2902 | * to 0 and uncharge() will be called. But, even if it's fully | |
2903 | * unmapped, migration may fail and this page has to be | |
2904 | * charged again. We set MIGRATION flag here and delay uncharge | |
2905 | * until end_migration() is called | |
2906 | * | |
2907 | * Corner Case Thinking | |
2908 | * A) | |
2909 | * When the old page was mapped as Anon and it's unmap-and-freed | |
2910 | * while migration was ongoing. | |
2911 | * If unmap finds the old page, uncharge() of it will be delayed | |
2912 | * until end_migration(). If unmap finds a new page, it's | |
2913 | * uncharged when it make mapcount to be 1->0. If unmap code | |
2914 | * finds swap_migration_entry, the new page will not be mapped | |
2915 | * and end_migration() will find it(mapcount==0). | |
2916 | * | |
2917 | * B) | |
2918 | * When the old page was mapped but migraion fails, the kernel | |
2919 | * remaps it. A charge for it is kept by MIGRATION flag even | |
2920 | * if mapcount goes down to 0. We can do remap successfully | |
2921 | * without charging it again. | |
2922 | * | |
2923 | * C) | |
2924 | * The "old" page is under lock_page() until the end of | |
2925 | * migration, so, the old page itself will not be swapped-out. | |
2926 | * If the new page is swapped out before end_migraton, our | |
2927 | * hook to usual swap-out path will catch the event. | |
2928 | */ | |
2929 | if (PageAnon(page)) | |
2930 | SetPageCgroupMigration(pc); | |
2931 | } | |
2932 | unlock_page_cgroup(pc); | |
2933 | /* | |
2934 | * If the page is not charged at this point, | |
2935 | * we return here. | |
2936 | */ | |
2937 | if (!mem) | |
2938 | return 0; | |
2939 | ||
2940 | *ptr = mem; | |
2941 | ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false); | |
2942 | css_put(&mem->css);/* drop extra refcnt */ | |
2943 | if (ret || *ptr == NULL) { | |
2944 | if (PageAnon(page)) { | |
2945 | lock_page_cgroup(pc); | |
2946 | ClearPageCgroupMigration(pc); | |
2947 | unlock_page_cgroup(pc); | |
2948 | /* | |
2949 | * The old page may be fully unmapped while we kept it. | |
2950 | */ | |
2951 | mem_cgroup_uncharge_page(page); | |
2952 | } | |
2953 | return -ENOMEM; | |
2954 | } | |
2955 | /* | |
2956 | * We charge new page before it's used/mapped. So, even if unlock_page() | |
2957 | * is called before end_migration, we can catch all events on this new | |
2958 | * page. In the case new page is migrated but not remapped, new page's | |
2959 | * mapcount will be finally 0 and we call uncharge in end_migration(). | |
2960 | */ | |
2961 | pc = lookup_page_cgroup(newpage); | |
2962 | if (PageAnon(page)) | |
2963 | ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; | |
2964 | else if (page_is_file_cache(page)) | |
2965 | ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; | |
2966 | else | |
2967 | ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; | |
2968 | __mem_cgroup_commit_charge(mem, page, 1, pc, ctype); | |
2969 | return ret; | |
2970 | } | |
2971 | ||
2972 | /* remove redundant charge if migration failed*/ | |
2973 | void mem_cgroup_end_migration(struct mem_cgroup *mem, | |
2974 | struct page *oldpage, struct page *newpage, bool migration_ok) | |
2975 | { | |
2976 | struct page *used, *unused; | |
2977 | struct page_cgroup *pc; | |
2978 | ||
2979 | if (!mem) | |
2980 | return; | |
2981 | /* blocks rmdir() */ | |
2982 | cgroup_exclude_rmdir(&mem->css); | |
2983 | if (!migration_ok) { | |
2984 | used = oldpage; | |
2985 | unused = newpage; | |
2986 | } else { | |
2987 | used = newpage; | |
2988 | unused = oldpage; | |
2989 | } | |
2990 | /* | |
2991 | * We disallowed uncharge of pages under migration because mapcount | |
2992 | * of the page goes down to zero, temporarly. | |
2993 | * Clear the flag and check the page should be charged. | |
2994 | */ | |
2995 | pc = lookup_page_cgroup(oldpage); | |
2996 | lock_page_cgroup(pc); | |
2997 | ClearPageCgroupMigration(pc); | |
2998 | unlock_page_cgroup(pc); | |
2999 | ||
3000 | __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); | |
3001 | ||
3002 | /* | |
3003 | * If a page is a file cache, radix-tree replacement is very atomic | |
3004 | * and we can skip this check. When it was an Anon page, its mapcount | |
3005 | * goes down to 0. But because we added MIGRATION flage, it's not | |
3006 | * uncharged yet. There are several case but page->mapcount check | |
3007 | * and USED bit check in mem_cgroup_uncharge_page() will do enough | |
3008 | * check. (see prepare_charge() also) | |
3009 | */ | |
3010 | if (PageAnon(used)) | |
3011 | mem_cgroup_uncharge_page(used); | |
3012 | /* | |
3013 | * At migration, we may charge account against cgroup which has no | |
3014 | * tasks. | |
3015 | * So, rmdir()->pre_destroy() can be called while we do this charge. | |
3016 | * In that case, we need to call pre_destroy() again. check it here. | |
3017 | */ | |
3018 | cgroup_release_and_wakeup_rmdir(&mem->css); | |
3019 | } | |
3020 | ||
3021 | /* | |
3022 | * A call to try to shrink memory usage on charge failure at shmem's swapin. | |
3023 | * Calling hierarchical_reclaim is not enough because we should update | |
3024 | * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM. | |
3025 | * Moreover considering hierarchy, we should reclaim from the mem_over_limit, | |
3026 | * not from the memcg which this page would be charged to. | |
3027 | * try_charge_swapin does all of these works properly. | |
3028 | */ | |
3029 | int mem_cgroup_shmem_charge_fallback(struct page *page, | |
3030 | struct mm_struct *mm, | |
3031 | gfp_t gfp_mask) | |
3032 | { | |
3033 | struct mem_cgroup *mem; | |
3034 | int ret; | |
3035 | ||
3036 | if (mem_cgroup_disabled()) | |
3037 | return 0; | |
3038 | ||
3039 | ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); | |
3040 | if (!ret) | |
3041 | mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */ | |
3042 | ||
3043 | return ret; | |
3044 | } | |
3045 | ||
3046 | #ifdef CONFIG_DEBUG_VM | |
3047 | static struct page_cgroup *lookup_page_cgroup_used(struct page *page) | |
3048 | { | |
3049 | struct page_cgroup *pc; | |
3050 | ||
3051 | pc = lookup_page_cgroup(page); | |
3052 | if (likely(pc) && PageCgroupUsed(pc)) | |
3053 | return pc; | |
3054 | return NULL; | |
3055 | } | |
3056 | ||
3057 | bool mem_cgroup_bad_page_check(struct page *page) | |
3058 | { | |
3059 | if (mem_cgroup_disabled()) | |
3060 | return false; | |
3061 | ||
3062 | return lookup_page_cgroup_used(page) != NULL; | |
3063 | } | |
3064 | ||
3065 | void mem_cgroup_print_bad_page(struct page *page) | |
3066 | { | |
3067 | struct page_cgroup *pc; | |
3068 | ||
3069 | pc = lookup_page_cgroup_used(page); | |
3070 | if (pc) { | |
3071 | int ret = -1; | |
3072 | char *path; | |
3073 | ||
3074 | printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p", | |
3075 | pc, pc->flags, pc->mem_cgroup); | |
3076 | ||
3077 | path = kmalloc(PATH_MAX, GFP_KERNEL); | |
3078 | if (path) { | |
3079 | rcu_read_lock(); | |
3080 | ret = cgroup_path(pc->mem_cgroup->css.cgroup, | |
3081 | path, PATH_MAX); | |
3082 | rcu_read_unlock(); | |
3083 | } | |
3084 | ||
3085 | printk(KERN_CONT "(%s)\n", | |
3086 | (ret < 0) ? "cannot get the path" : path); | |
3087 | kfree(path); | |
3088 | } | |
3089 | } | |
3090 | #endif | |
3091 | ||
3092 | static DEFINE_MUTEX(set_limit_mutex); | |
3093 | ||
3094 | static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, | |
3095 | unsigned long long val) | |
3096 | { | |
3097 | int retry_count; | |
3098 | u64 memswlimit, memlimit; | |
3099 | int ret = 0; | |
3100 | int children = mem_cgroup_count_children(memcg); | |
3101 | u64 curusage, oldusage; | |
3102 | int enlarge; | |
3103 | ||
3104 | /* | |
3105 | * For keeping hierarchical_reclaim simple, how long we should retry | |
3106 | * is depends on callers. We set our retry-count to be function | |
3107 | * of # of children which we should visit in this loop. | |
3108 | */ | |
3109 | retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; | |
3110 | ||
3111 | oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
3112 | ||
3113 | enlarge = 0; | |
3114 | while (retry_count) { | |
3115 | if (signal_pending(current)) { | |
3116 | ret = -EINTR; | |
3117 | break; | |
3118 | } | |
3119 | /* | |
3120 | * Rather than hide all in some function, I do this in | |
3121 | * open coded manner. You see what this really does. | |
3122 | * We have to guarantee mem->res.limit < mem->memsw.limit. | |
3123 | */ | |
3124 | mutex_lock(&set_limit_mutex); | |
3125 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3126 | if (memswlimit < val) { | |
3127 | ret = -EINVAL; | |
3128 | mutex_unlock(&set_limit_mutex); | |
3129 | break; | |
3130 | } | |
3131 | ||
3132 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3133 | if (memlimit < val) | |
3134 | enlarge = 1; | |
3135 | ||
3136 | ret = res_counter_set_limit(&memcg->res, val); | |
3137 | if (!ret) { | |
3138 | if (memswlimit == val) | |
3139 | memcg->memsw_is_minimum = true; | |
3140 | else | |
3141 | memcg->memsw_is_minimum = false; | |
3142 | } | |
3143 | mutex_unlock(&set_limit_mutex); | |
3144 | ||
3145 | if (!ret) | |
3146 | break; | |
3147 | ||
3148 | mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, | |
3149 | MEM_CGROUP_RECLAIM_SHRINK); | |
3150 | curusage = res_counter_read_u64(&memcg->res, RES_USAGE); | |
3151 | /* Usage is reduced ? */ | |
3152 | if (curusage >= oldusage) | |
3153 | retry_count--; | |
3154 | else | |
3155 | oldusage = curusage; | |
3156 | } | |
3157 | if (!ret && enlarge) | |
3158 | memcg_oom_recover(memcg); | |
3159 | ||
3160 | return ret; | |
3161 | } | |
3162 | ||
3163 | static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, | |
3164 | unsigned long long val) | |
3165 | { | |
3166 | int retry_count; | |
3167 | u64 memlimit, memswlimit, oldusage, curusage; | |
3168 | int children = mem_cgroup_count_children(memcg); | |
3169 | int ret = -EBUSY; | |
3170 | int enlarge = 0; | |
3171 | ||
3172 | /* see mem_cgroup_resize_res_limit */ | |
3173 | retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; | |
3174 | oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
3175 | while (retry_count) { | |
3176 | if (signal_pending(current)) { | |
3177 | ret = -EINTR; | |
3178 | break; | |
3179 | } | |
3180 | /* | |
3181 | * Rather than hide all in some function, I do this in | |
3182 | * open coded manner. You see what this really does. | |
3183 | * We have to guarantee mem->res.limit < mem->memsw.limit. | |
3184 | */ | |
3185 | mutex_lock(&set_limit_mutex); | |
3186 | memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3187 | if (memlimit > val) { | |
3188 | ret = -EINVAL; | |
3189 | mutex_unlock(&set_limit_mutex); | |
3190 | break; | |
3191 | } | |
3192 | memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3193 | if (memswlimit < val) | |
3194 | enlarge = 1; | |
3195 | ret = res_counter_set_limit(&memcg->memsw, val); | |
3196 | if (!ret) { | |
3197 | if (memlimit == val) | |
3198 | memcg->memsw_is_minimum = true; | |
3199 | else | |
3200 | memcg->memsw_is_minimum = false; | |
3201 | } | |
3202 | mutex_unlock(&set_limit_mutex); | |
3203 | ||
3204 | if (!ret) | |
3205 | break; | |
3206 | ||
3207 | mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, | |
3208 | MEM_CGROUP_RECLAIM_NOSWAP | | |
3209 | MEM_CGROUP_RECLAIM_SHRINK); | |
3210 | curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); | |
3211 | /* Usage is reduced ? */ | |
3212 | if (curusage >= oldusage) | |
3213 | retry_count--; | |
3214 | else | |
3215 | oldusage = curusage; | |
3216 | } | |
3217 | if (!ret && enlarge) | |
3218 | memcg_oom_recover(memcg); | |
3219 | return ret; | |
3220 | } | |
3221 | ||
3222 | unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, | |
3223 | gfp_t gfp_mask) | |
3224 | { | |
3225 | unsigned long nr_reclaimed = 0; | |
3226 | struct mem_cgroup_per_zone *mz, *next_mz = NULL; | |
3227 | unsigned long reclaimed; | |
3228 | int loop = 0; | |
3229 | struct mem_cgroup_tree_per_zone *mctz; | |
3230 | unsigned long long excess; | |
3231 | ||
3232 | if (order > 0) | |
3233 | return 0; | |
3234 | ||
3235 | mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); | |
3236 | /* | |
3237 | * This loop can run a while, specially if mem_cgroup's continuously | |
3238 | * keep exceeding their soft limit and putting the system under | |
3239 | * pressure | |
3240 | */ | |
3241 | do { | |
3242 | if (next_mz) | |
3243 | mz = next_mz; | |
3244 | else | |
3245 | mz = mem_cgroup_largest_soft_limit_node(mctz); | |
3246 | if (!mz) | |
3247 | break; | |
3248 | ||
3249 | reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone, | |
3250 | gfp_mask, | |
3251 | MEM_CGROUP_RECLAIM_SOFT); | |
3252 | nr_reclaimed += reclaimed; | |
3253 | spin_lock(&mctz->lock); | |
3254 | ||
3255 | /* | |
3256 | * If we failed to reclaim anything from this memory cgroup | |
3257 | * it is time to move on to the next cgroup | |
3258 | */ | |
3259 | next_mz = NULL; | |
3260 | if (!reclaimed) { | |
3261 | do { | |
3262 | /* | |
3263 | * Loop until we find yet another one. | |
3264 | * | |
3265 | * By the time we get the soft_limit lock | |
3266 | * again, someone might have aded the | |
3267 | * group back on the RB tree. Iterate to | |
3268 | * make sure we get a different mem. | |
3269 | * mem_cgroup_largest_soft_limit_node returns | |
3270 | * NULL if no other cgroup is present on | |
3271 | * the tree | |
3272 | */ | |
3273 | next_mz = | |
3274 | __mem_cgroup_largest_soft_limit_node(mctz); | |
3275 | if (next_mz == mz) { | |
3276 | css_put(&next_mz->mem->css); | |
3277 | next_mz = NULL; | |
3278 | } else /* next_mz == NULL or other memcg */ | |
3279 | break; | |
3280 | } while (1); | |
3281 | } | |
3282 | __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); | |
3283 | excess = res_counter_soft_limit_excess(&mz->mem->res); | |
3284 | /* | |
3285 | * One school of thought says that we should not add | |
3286 | * back the node to the tree if reclaim returns 0. | |
3287 | * But our reclaim could return 0, simply because due | |
3288 | * to priority we are exposing a smaller subset of | |
3289 | * memory to reclaim from. Consider this as a longer | |
3290 | * term TODO. | |
3291 | */ | |
3292 | /* If excess == 0, no tree ops */ | |
3293 | __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); | |
3294 | spin_unlock(&mctz->lock); | |
3295 | css_put(&mz->mem->css); | |
3296 | loop++; | |
3297 | /* | |
3298 | * Could not reclaim anything and there are no more | |
3299 | * mem cgroups to try or we seem to be looping without | |
3300 | * reclaiming anything. | |
3301 | */ | |
3302 | if (!nr_reclaimed && | |
3303 | (next_mz == NULL || | |
3304 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) | |
3305 | break; | |
3306 | } while (!nr_reclaimed); | |
3307 | if (next_mz) | |
3308 | css_put(&next_mz->mem->css); | |
3309 | return nr_reclaimed; | |
3310 | } | |
3311 | ||
3312 | /* | |
3313 | * This routine traverse page_cgroup in given list and drop them all. | |
3314 | * *And* this routine doesn't reclaim page itself, just removes page_cgroup. | |
3315 | */ | |
3316 | static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, | |
3317 | int node, int zid, enum lru_list lru) | |
3318 | { | |
3319 | struct zone *zone; | |
3320 | struct mem_cgroup_per_zone *mz; | |
3321 | struct page_cgroup *pc, *busy; | |
3322 | unsigned long flags, loop; | |
3323 | struct list_head *list; | |
3324 | int ret = 0; | |
3325 | ||
3326 | zone = &NODE_DATA(node)->node_zones[zid]; | |
3327 | mz = mem_cgroup_zoneinfo(mem, node, zid); | |
3328 | list = &mz->lists[lru]; | |
3329 | ||
3330 | loop = MEM_CGROUP_ZSTAT(mz, lru); | |
3331 | /* give some margin against EBUSY etc...*/ | |
3332 | loop += 256; | |
3333 | busy = NULL; | |
3334 | while (loop--) { | |
3335 | struct page *page; | |
3336 | ||
3337 | ret = 0; | |
3338 | spin_lock_irqsave(&zone->lru_lock, flags); | |
3339 | if (list_empty(list)) { | |
3340 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3341 | break; | |
3342 | } | |
3343 | pc = list_entry(list->prev, struct page_cgroup, lru); | |
3344 | if (busy == pc) { | |
3345 | list_move(&pc->lru, list); | |
3346 | busy = NULL; | |
3347 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3348 | continue; | |
3349 | } | |
3350 | spin_unlock_irqrestore(&zone->lru_lock, flags); | |
3351 | ||
3352 | page = lookup_cgroup_page(pc); | |
3353 | ||
3354 | ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL); | |
3355 | if (ret == -ENOMEM) | |
3356 | break; | |
3357 | ||
3358 | if (ret == -EBUSY || ret == -EINVAL) { | |
3359 | /* found lock contention or "pc" is obsolete. */ | |
3360 | busy = pc; | |
3361 | cond_resched(); | |
3362 | } else | |
3363 | busy = NULL; | |
3364 | } | |
3365 | ||
3366 | if (!ret && !list_empty(list)) | |
3367 | return -EBUSY; | |
3368 | return ret; | |
3369 | } | |
3370 | ||
3371 | /* | |
3372 | * make mem_cgroup's charge to be 0 if there is no task. | |
3373 | * This enables deleting this mem_cgroup. | |
3374 | */ | |
3375 | static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) | |
3376 | { | |
3377 | int ret; | |
3378 | int node, zid, shrink; | |
3379 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
3380 | struct cgroup *cgrp = mem->css.cgroup; | |
3381 | ||
3382 | css_get(&mem->css); | |
3383 | ||
3384 | shrink = 0; | |
3385 | /* should free all ? */ | |
3386 | if (free_all) | |
3387 | goto try_to_free; | |
3388 | move_account: | |
3389 | do { | |
3390 | ret = -EBUSY; | |
3391 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) | |
3392 | goto out; | |
3393 | ret = -EINTR; | |
3394 | if (signal_pending(current)) | |
3395 | goto out; | |
3396 | /* This is for making all *used* pages to be on LRU. */ | |
3397 | lru_add_drain_all(); | |
3398 | drain_all_stock_sync(); | |
3399 | ret = 0; | |
3400 | mem_cgroup_start_move(mem); | |
3401 | for_each_node_state(node, N_HIGH_MEMORY) { | |
3402 | for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { | |
3403 | enum lru_list l; | |
3404 | for_each_lru(l) { | |
3405 | ret = mem_cgroup_force_empty_list(mem, | |
3406 | node, zid, l); | |
3407 | if (ret) | |
3408 | break; | |
3409 | } | |
3410 | } | |
3411 | if (ret) | |
3412 | break; | |
3413 | } | |
3414 | mem_cgroup_end_move(mem); | |
3415 | memcg_oom_recover(mem); | |
3416 | /* it seems parent cgroup doesn't have enough mem */ | |
3417 | if (ret == -ENOMEM) | |
3418 | goto try_to_free; | |
3419 | cond_resched(); | |
3420 | /* "ret" should also be checked to ensure all lists are empty. */ | |
3421 | } while (mem->res.usage > 0 || ret); | |
3422 | out: | |
3423 | css_put(&mem->css); | |
3424 | return ret; | |
3425 | ||
3426 | try_to_free: | |
3427 | /* returns EBUSY if there is a task or if we come here twice. */ | |
3428 | if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { | |
3429 | ret = -EBUSY; | |
3430 | goto out; | |
3431 | } | |
3432 | /* we call try-to-free pages for make this cgroup empty */ | |
3433 | lru_add_drain_all(); | |
3434 | /* try to free all pages in this cgroup */ | |
3435 | shrink = 1; | |
3436 | while (nr_retries && mem->res.usage > 0) { | |
3437 | int progress; | |
3438 | ||
3439 | if (signal_pending(current)) { | |
3440 | ret = -EINTR; | |
3441 | goto out; | |
3442 | } | |
3443 | progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, | |
3444 | false, get_swappiness(mem)); | |
3445 | if (!progress) { | |
3446 | nr_retries--; | |
3447 | /* maybe some writeback is necessary */ | |
3448 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
3449 | } | |
3450 | ||
3451 | } | |
3452 | lru_add_drain(); | |
3453 | /* try move_account...there may be some *locked* pages. */ | |
3454 | goto move_account; | |
3455 | } | |
3456 | ||
3457 | int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) | |
3458 | { | |
3459 | return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); | |
3460 | } | |
3461 | ||
3462 | ||
3463 | static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) | |
3464 | { | |
3465 | return mem_cgroup_from_cont(cont)->use_hierarchy; | |
3466 | } | |
3467 | ||
3468 | static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, | |
3469 | u64 val) | |
3470 | { | |
3471 | int retval = 0; | |
3472 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
3473 | struct cgroup *parent = cont->parent; | |
3474 | struct mem_cgroup *parent_mem = NULL; | |
3475 | ||
3476 | if (parent) | |
3477 | parent_mem = mem_cgroup_from_cont(parent); | |
3478 | ||
3479 | cgroup_lock(); | |
3480 | /* | |
3481 | * If parent's use_hierarchy is set, we can't make any modifications | |
3482 | * in the child subtrees. If it is unset, then the change can | |
3483 | * occur, provided the current cgroup has no children. | |
3484 | * | |
3485 | * For the root cgroup, parent_mem is NULL, we allow value to be | |
3486 | * set if there are no children. | |
3487 | */ | |
3488 | if ((!parent_mem || !parent_mem->use_hierarchy) && | |
3489 | (val == 1 || val == 0)) { | |
3490 | if (list_empty(&cont->children)) | |
3491 | mem->use_hierarchy = val; | |
3492 | else | |
3493 | retval = -EBUSY; | |
3494 | } else | |
3495 | retval = -EINVAL; | |
3496 | cgroup_unlock(); | |
3497 | ||
3498 | return retval; | |
3499 | } | |
3500 | ||
3501 | ||
3502 | static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem, | |
3503 | enum mem_cgroup_stat_index idx) | |
3504 | { | |
3505 | struct mem_cgroup *iter; | |
3506 | s64 val = 0; | |
3507 | ||
3508 | /* each per cpu's value can be minus.Then, use s64 */ | |
3509 | for_each_mem_cgroup_tree(iter, mem) | |
3510 | val += mem_cgroup_read_stat(iter, idx); | |
3511 | ||
3512 | if (val < 0) /* race ? */ | |
3513 | val = 0; | |
3514 | return val; | |
3515 | } | |
3516 | ||
3517 | static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap) | |
3518 | { | |
3519 | u64 val; | |
3520 | ||
3521 | if (!mem_cgroup_is_root(mem)) { | |
3522 | if (!swap) | |
3523 | return res_counter_read_u64(&mem->res, RES_USAGE); | |
3524 | else | |
3525 | return res_counter_read_u64(&mem->memsw, RES_USAGE); | |
3526 | } | |
3527 | ||
3528 | val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE); | |
3529 | val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS); | |
3530 | ||
3531 | if (swap) | |
3532 | val += mem_cgroup_get_recursive_idx_stat(mem, | |
3533 | MEM_CGROUP_STAT_SWAPOUT); | |
3534 | ||
3535 | return val << PAGE_SHIFT; | |
3536 | } | |
3537 | ||
3538 | static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) | |
3539 | { | |
3540 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
3541 | u64 val; | |
3542 | int type, name; | |
3543 | ||
3544 | type = MEMFILE_TYPE(cft->private); | |
3545 | name = MEMFILE_ATTR(cft->private); | |
3546 | switch (type) { | |
3547 | case _MEM: | |
3548 | if (name == RES_USAGE) | |
3549 | val = mem_cgroup_usage(mem, false); | |
3550 | else | |
3551 | val = res_counter_read_u64(&mem->res, name); | |
3552 | break; | |
3553 | case _MEMSWAP: | |
3554 | if (name == RES_USAGE) | |
3555 | val = mem_cgroup_usage(mem, true); | |
3556 | else | |
3557 | val = res_counter_read_u64(&mem->memsw, name); | |
3558 | break; | |
3559 | default: | |
3560 | BUG(); | |
3561 | break; | |
3562 | } | |
3563 | return val; | |
3564 | } | |
3565 | /* | |
3566 | * The user of this function is... | |
3567 | * RES_LIMIT. | |
3568 | */ | |
3569 | static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, | |
3570 | const char *buffer) | |
3571 | { | |
3572 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); | |
3573 | int type, name; | |
3574 | unsigned long long val; | |
3575 | int ret; | |
3576 | ||
3577 | type = MEMFILE_TYPE(cft->private); | |
3578 | name = MEMFILE_ATTR(cft->private); | |
3579 | switch (name) { | |
3580 | case RES_LIMIT: | |
3581 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ | |
3582 | ret = -EINVAL; | |
3583 | break; | |
3584 | } | |
3585 | /* This function does all necessary parse...reuse it */ | |
3586 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
3587 | if (ret) | |
3588 | break; | |
3589 | if (type == _MEM) | |
3590 | ret = mem_cgroup_resize_limit(memcg, val); | |
3591 | else | |
3592 | ret = mem_cgroup_resize_memsw_limit(memcg, val); | |
3593 | break; | |
3594 | case RES_SOFT_LIMIT: | |
3595 | ret = res_counter_memparse_write_strategy(buffer, &val); | |
3596 | if (ret) | |
3597 | break; | |
3598 | /* | |
3599 | * For memsw, soft limits are hard to implement in terms | |
3600 | * of semantics, for now, we support soft limits for | |
3601 | * control without swap | |
3602 | */ | |
3603 | if (type == _MEM) | |
3604 | ret = res_counter_set_soft_limit(&memcg->res, val); | |
3605 | else | |
3606 | ret = -EINVAL; | |
3607 | break; | |
3608 | default: | |
3609 | ret = -EINVAL; /* should be BUG() ? */ | |
3610 | break; | |
3611 | } | |
3612 | return ret; | |
3613 | } | |
3614 | ||
3615 | static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, | |
3616 | unsigned long long *mem_limit, unsigned long long *memsw_limit) | |
3617 | { | |
3618 | struct cgroup *cgroup; | |
3619 | unsigned long long min_limit, min_memsw_limit, tmp; | |
3620 | ||
3621 | min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3622 | min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3623 | cgroup = memcg->css.cgroup; | |
3624 | if (!memcg->use_hierarchy) | |
3625 | goto out; | |
3626 | ||
3627 | while (cgroup->parent) { | |
3628 | cgroup = cgroup->parent; | |
3629 | memcg = mem_cgroup_from_cont(cgroup); | |
3630 | if (!memcg->use_hierarchy) | |
3631 | break; | |
3632 | tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); | |
3633 | min_limit = min(min_limit, tmp); | |
3634 | tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); | |
3635 | min_memsw_limit = min(min_memsw_limit, tmp); | |
3636 | } | |
3637 | out: | |
3638 | *mem_limit = min_limit; | |
3639 | *memsw_limit = min_memsw_limit; | |
3640 | return; | |
3641 | } | |
3642 | ||
3643 | static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) | |
3644 | { | |
3645 | struct mem_cgroup *mem; | |
3646 | int type, name; | |
3647 | ||
3648 | mem = mem_cgroup_from_cont(cont); | |
3649 | type = MEMFILE_TYPE(event); | |
3650 | name = MEMFILE_ATTR(event); | |
3651 | switch (name) { | |
3652 | case RES_MAX_USAGE: | |
3653 | if (type == _MEM) | |
3654 | res_counter_reset_max(&mem->res); | |
3655 | else | |
3656 | res_counter_reset_max(&mem->memsw); | |
3657 | break; | |
3658 | case RES_FAILCNT: | |
3659 | if (type == _MEM) | |
3660 | res_counter_reset_failcnt(&mem->res); | |
3661 | else | |
3662 | res_counter_reset_failcnt(&mem->memsw); | |
3663 | break; | |
3664 | } | |
3665 | ||
3666 | return 0; | |
3667 | } | |
3668 | ||
3669 | static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, | |
3670 | struct cftype *cft) | |
3671 | { | |
3672 | return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; | |
3673 | } | |
3674 | ||
3675 | #ifdef CONFIG_MMU | |
3676 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, | |
3677 | struct cftype *cft, u64 val) | |
3678 | { | |
3679 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
3680 | ||
3681 | if (val >= (1 << NR_MOVE_TYPE)) | |
3682 | return -EINVAL; | |
3683 | /* | |
3684 | * We check this value several times in both in can_attach() and | |
3685 | * attach(), so we need cgroup lock to prevent this value from being | |
3686 | * inconsistent. | |
3687 | */ | |
3688 | cgroup_lock(); | |
3689 | mem->move_charge_at_immigrate = val; | |
3690 | cgroup_unlock(); | |
3691 | ||
3692 | return 0; | |
3693 | } | |
3694 | #else | |
3695 | static int mem_cgroup_move_charge_write(struct cgroup *cgrp, | |
3696 | struct cftype *cft, u64 val) | |
3697 | { | |
3698 | return -ENOSYS; | |
3699 | } | |
3700 | #endif | |
3701 | ||
3702 | ||
3703 | /* For read statistics */ | |
3704 | enum { | |
3705 | MCS_CACHE, | |
3706 | MCS_RSS, | |
3707 | MCS_FILE_MAPPED, | |
3708 | MCS_PGPGIN, | |
3709 | MCS_PGPGOUT, | |
3710 | MCS_SWAP, | |
3711 | MCS_INACTIVE_ANON, | |
3712 | MCS_ACTIVE_ANON, | |
3713 | MCS_INACTIVE_FILE, | |
3714 | MCS_ACTIVE_FILE, | |
3715 | MCS_UNEVICTABLE, | |
3716 | NR_MCS_STAT, | |
3717 | }; | |
3718 | ||
3719 | struct mcs_total_stat { | |
3720 | s64 stat[NR_MCS_STAT]; | |
3721 | }; | |
3722 | ||
3723 | struct { | |
3724 | char *local_name; | |
3725 | char *total_name; | |
3726 | } memcg_stat_strings[NR_MCS_STAT] = { | |
3727 | {"cache", "total_cache"}, | |
3728 | {"rss", "total_rss"}, | |
3729 | {"mapped_file", "total_mapped_file"}, | |
3730 | {"pgpgin", "total_pgpgin"}, | |
3731 | {"pgpgout", "total_pgpgout"}, | |
3732 | {"swap", "total_swap"}, | |
3733 | {"inactive_anon", "total_inactive_anon"}, | |
3734 | {"active_anon", "total_active_anon"}, | |
3735 | {"inactive_file", "total_inactive_file"}, | |
3736 | {"active_file", "total_active_file"}, | |
3737 | {"unevictable", "total_unevictable"} | |
3738 | }; | |
3739 | ||
3740 | ||
3741 | static void | |
3742 | mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) | |
3743 | { | |
3744 | s64 val; | |
3745 | ||
3746 | /* per cpu stat */ | |
3747 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); | |
3748 | s->stat[MCS_CACHE] += val * PAGE_SIZE; | |
3749 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); | |
3750 | s->stat[MCS_RSS] += val * PAGE_SIZE; | |
3751 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED); | |
3752 | s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; | |
3753 | val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN); | |
3754 | s->stat[MCS_PGPGIN] += val; | |
3755 | val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT); | |
3756 | s->stat[MCS_PGPGOUT] += val; | |
3757 | if (do_swap_account) { | |
3758 | val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT); | |
3759 | s->stat[MCS_SWAP] += val * PAGE_SIZE; | |
3760 | } | |
3761 | ||
3762 | /* per zone stat */ | |
3763 | val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); | |
3764 | s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; | |
3765 | val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); | |
3766 | s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; | |
3767 | val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); | |
3768 | s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; | |
3769 | val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); | |
3770 | s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; | |
3771 | val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); | |
3772 | s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; | |
3773 | } | |
3774 | ||
3775 | static void | |
3776 | mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) | |
3777 | { | |
3778 | struct mem_cgroup *iter; | |
3779 | ||
3780 | for_each_mem_cgroup_tree(iter, mem) | |
3781 | mem_cgroup_get_local_stat(iter, s); | |
3782 | } | |
3783 | ||
3784 | static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, | |
3785 | struct cgroup_map_cb *cb) | |
3786 | { | |
3787 | struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); | |
3788 | struct mcs_total_stat mystat; | |
3789 | int i; | |
3790 | ||
3791 | memset(&mystat, 0, sizeof(mystat)); | |
3792 | mem_cgroup_get_local_stat(mem_cont, &mystat); | |
3793 | ||
3794 | for (i = 0; i < NR_MCS_STAT; i++) { | |
3795 | if (i == MCS_SWAP && !do_swap_account) | |
3796 | continue; | |
3797 | cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); | |
3798 | } | |
3799 | ||
3800 | /* Hierarchical information */ | |
3801 | { | |
3802 | unsigned long long limit, memsw_limit; | |
3803 | memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); | |
3804 | cb->fill(cb, "hierarchical_memory_limit", limit); | |
3805 | if (do_swap_account) | |
3806 | cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); | |
3807 | } | |
3808 | ||
3809 | memset(&mystat, 0, sizeof(mystat)); | |
3810 | mem_cgroup_get_total_stat(mem_cont, &mystat); | |
3811 | for (i = 0; i < NR_MCS_STAT; i++) { | |
3812 | if (i == MCS_SWAP && !do_swap_account) | |
3813 | continue; | |
3814 | cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); | |
3815 | } | |
3816 | ||
3817 | #ifdef CONFIG_DEBUG_VM | |
3818 | cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); | |
3819 | ||
3820 | { | |
3821 | int nid, zid; | |
3822 | struct mem_cgroup_per_zone *mz; | |
3823 | unsigned long recent_rotated[2] = {0, 0}; | |
3824 | unsigned long recent_scanned[2] = {0, 0}; | |
3825 | ||
3826 | for_each_online_node(nid) | |
3827 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { | |
3828 | mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); | |
3829 | ||
3830 | recent_rotated[0] += | |
3831 | mz->reclaim_stat.recent_rotated[0]; | |
3832 | recent_rotated[1] += | |
3833 | mz->reclaim_stat.recent_rotated[1]; | |
3834 | recent_scanned[0] += | |
3835 | mz->reclaim_stat.recent_scanned[0]; | |
3836 | recent_scanned[1] += | |
3837 | mz->reclaim_stat.recent_scanned[1]; | |
3838 | } | |
3839 | cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); | |
3840 | cb->fill(cb, "recent_rotated_file", recent_rotated[1]); | |
3841 | cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); | |
3842 | cb->fill(cb, "recent_scanned_file", recent_scanned[1]); | |
3843 | } | |
3844 | #endif | |
3845 | ||
3846 | return 0; | |
3847 | } | |
3848 | ||
3849 | static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) | |
3850 | { | |
3851 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3852 | ||
3853 | return get_swappiness(memcg); | |
3854 | } | |
3855 | ||
3856 | static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, | |
3857 | u64 val) | |
3858 | { | |
3859 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3860 | struct mem_cgroup *parent; | |
3861 | ||
3862 | if (val > 100) | |
3863 | return -EINVAL; | |
3864 | ||
3865 | if (cgrp->parent == NULL) | |
3866 | return -EINVAL; | |
3867 | ||
3868 | parent = mem_cgroup_from_cont(cgrp->parent); | |
3869 | ||
3870 | cgroup_lock(); | |
3871 | ||
3872 | /* If under hierarchy, only empty-root can set this value */ | |
3873 | if ((parent->use_hierarchy) || | |
3874 | (memcg->use_hierarchy && !list_empty(&cgrp->children))) { | |
3875 | cgroup_unlock(); | |
3876 | return -EINVAL; | |
3877 | } | |
3878 | ||
3879 | memcg->swappiness = val; | |
3880 | ||
3881 | cgroup_unlock(); | |
3882 | ||
3883 | return 0; | |
3884 | } | |
3885 | ||
3886 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) | |
3887 | { | |
3888 | struct mem_cgroup_threshold_ary *t; | |
3889 | u64 usage; | |
3890 | int i; | |
3891 | ||
3892 | rcu_read_lock(); | |
3893 | if (!swap) | |
3894 | t = rcu_dereference(memcg->thresholds.primary); | |
3895 | else | |
3896 | t = rcu_dereference(memcg->memsw_thresholds.primary); | |
3897 | ||
3898 | if (!t) | |
3899 | goto unlock; | |
3900 | ||
3901 | usage = mem_cgroup_usage(memcg, swap); | |
3902 | ||
3903 | /* | |
3904 | * current_threshold points to threshold just below usage. | |
3905 | * If it's not true, a threshold was crossed after last | |
3906 | * call of __mem_cgroup_threshold(). | |
3907 | */ | |
3908 | i = t->current_threshold; | |
3909 | ||
3910 | /* | |
3911 | * Iterate backward over array of thresholds starting from | |
3912 | * current_threshold and check if a threshold is crossed. | |
3913 | * If none of thresholds below usage is crossed, we read | |
3914 | * only one element of the array here. | |
3915 | */ | |
3916 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) | |
3917 | eventfd_signal(t->entries[i].eventfd, 1); | |
3918 | ||
3919 | /* i = current_threshold + 1 */ | |
3920 | i++; | |
3921 | ||
3922 | /* | |
3923 | * Iterate forward over array of thresholds starting from | |
3924 | * current_threshold+1 and check if a threshold is crossed. | |
3925 | * If none of thresholds above usage is crossed, we read | |
3926 | * only one element of the array here. | |
3927 | */ | |
3928 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) | |
3929 | eventfd_signal(t->entries[i].eventfd, 1); | |
3930 | ||
3931 | /* Update current_threshold */ | |
3932 | t->current_threshold = i - 1; | |
3933 | unlock: | |
3934 | rcu_read_unlock(); | |
3935 | } | |
3936 | ||
3937 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) | |
3938 | { | |
3939 | while (memcg) { | |
3940 | __mem_cgroup_threshold(memcg, false); | |
3941 | if (do_swap_account) | |
3942 | __mem_cgroup_threshold(memcg, true); | |
3943 | ||
3944 | memcg = parent_mem_cgroup(memcg); | |
3945 | } | |
3946 | } | |
3947 | ||
3948 | static int compare_thresholds(const void *a, const void *b) | |
3949 | { | |
3950 | const struct mem_cgroup_threshold *_a = a; | |
3951 | const struct mem_cgroup_threshold *_b = b; | |
3952 | ||
3953 | return _a->threshold - _b->threshold; | |
3954 | } | |
3955 | ||
3956 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem) | |
3957 | { | |
3958 | struct mem_cgroup_eventfd_list *ev; | |
3959 | ||
3960 | list_for_each_entry(ev, &mem->oom_notify, list) | |
3961 | eventfd_signal(ev->eventfd, 1); | |
3962 | return 0; | |
3963 | } | |
3964 | ||
3965 | static void mem_cgroup_oom_notify(struct mem_cgroup *mem) | |
3966 | { | |
3967 | struct mem_cgroup *iter; | |
3968 | ||
3969 | for_each_mem_cgroup_tree(iter, mem) | |
3970 | mem_cgroup_oom_notify_cb(iter); | |
3971 | } | |
3972 | ||
3973 | static int mem_cgroup_usage_register_event(struct cgroup *cgrp, | |
3974 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) | |
3975 | { | |
3976 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
3977 | struct mem_cgroup_thresholds *thresholds; | |
3978 | struct mem_cgroup_threshold_ary *new; | |
3979 | int type = MEMFILE_TYPE(cft->private); | |
3980 | u64 threshold, usage; | |
3981 | int i, size, ret; | |
3982 | ||
3983 | ret = res_counter_memparse_write_strategy(args, &threshold); | |
3984 | if (ret) | |
3985 | return ret; | |
3986 | ||
3987 | mutex_lock(&memcg->thresholds_lock); | |
3988 | ||
3989 | if (type == _MEM) | |
3990 | thresholds = &memcg->thresholds; | |
3991 | else if (type == _MEMSWAP) | |
3992 | thresholds = &memcg->memsw_thresholds; | |
3993 | else | |
3994 | BUG(); | |
3995 | ||
3996 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
3997 | ||
3998 | /* Check if a threshold crossed before adding a new one */ | |
3999 | if (thresholds->primary) | |
4000 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
4001 | ||
4002 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; | |
4003 | ||
4004 | /* Allocate memory for new array of thresholds */ | |
4005 | new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), | |
4006 | GFP_KERNEL); | |
4007 | if (!new) { | |
4008 | ret = -ENOMEM; | |
4009 | goto unlock; | |
4010 | } | |
4011 | new->size = size; | |
4012 | ||
4013 | /* Copy thresholds (if any) to new array */ | |
4014 | if (thresholds->primary) { | |
4015 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * | |
4016 | sizeof(struct mem_cgroup_threshold)); | |
4017 | } | |
4018 | ||
4019 | /* Add new threshold */ | |
4020 | new->entries[size - 1].eventfd = eventfd; | |
4021 | new->entries[size - 1].threshold = threshold; | |
4022 | ||
4023 | /* Sort thresholds. Registering of new threshold isn't time-critical */ | |
4024 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), | |
4025 | compare_thresholds, NULL); | |
4026 | ||
4027 | /* Find current threshold */ | |
4028 | new->current_threshold = -1; | |
4029 | for (i = 0; i < size; i++) { | |
4030 | if (new->entries[i].threshold < usage) { | |
4031 | /* | |
4032 | * new->current_threshold will not be used until | |
4033 | * rcu_assign_pointer(), so it's safe to increment | |
4034 | * it here. | |
4035 | */ | |
4036 | ++new->current_threshold; | |
4037 | } | |
4038 | } | |
4039 | ||
4040 | /* Free old spare buffer and save old primary buffer as spare */ | |
4041 | kfree(thresholds->spare); | |
4042 | thresholds->spare = thresholds->primary; | |
4043 | ||
4044 | rcu_assign_pointer(thresholds->primary, new); | |
4045 | ||
4046 | /* To be sure that nobody uses thresholds */ | |
4047 | synchronize_rcu(); | |
4048 | ||
4049 | unlock: | |
4050 | mutex_unlock(&memcg->thresholds_lock); | |
4051 | ||
4052 | return ret; | |
4053 | } | |
4054 | ||
4055 | static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, | |
4056 | struct cftype *cft, struct eventfd_ctx *eventfd) | |
4057 | { | |
4058 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4059 | struct mem_cgroup_thresholds *thresholds; | |
4060 | struct mem_cgroup_threshold_ary *new; | |
4061 | int type = MEMFILE_TYPE(cft->private); | |
4062 | u64 usage; | |
4063 | int i, j, size; | |
4064 | ||
4065 | mutex_lock(&memcg->thresholds_lock); | |
4066 | if (type == _MEM) | |
4067 | thresholds = &memcg->thresholds; | |
4068 | else if (type == _MEMSWAP) | |
4069 | thresholds = &memcg->memsw_thresholds; | |
4070 | else | |
4071 | BUG(); | |
4072 | ||
4073 | /* | |
4074 | * Something went wrong if we trying to unregister a threshold | |
4075 | * if we don't have thresholds | |
4076 | */ | |
4077 | BUG_ON(!thresholds); | |
4078 | ||
4079 | usage = mem_cgroup_usage(memcg, type == _MEMSWAP); | |
4080 | ||
4081 | /* Check if a threshold crossed before removing */ | |
4082 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
4083 | ||
4084 | /* Calculate new number of threshold */ | |
4085 | size = 0; | |
4086 | for (i = 0; i < thresholds->primary->size; i++) { | |
4087 | if (thresholds->primary->entries[i].eventfd != eventfd) | |
4088 | size++; | |
4089 | } | |
4090 | ||
4091 | new = thresholds->spare; | |
4092 | ||
4093 | /* Set thresholds array to NULL if we don't have thresholds */ | |
4094 | if (!size) { | |
4095 | kfree(new); | |
4096 | new = NULL; | |
4097 | goto swap_buffers; | |
4098 | } | |
4099 | ||
4100 | new->size = size; | |
4101 | ||
4102 | /* Copy thresholds and find current threshold */ | |
4103 | new->current_threshold = -1; | |
4104 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { | |
4105 | if (thresholds->primary->entries[i].eventfd == eventfd) | |
4106 | continue; | |
4107 | ||
4108 | new->entries[j] = thresholds->primary->entries[i]; | |
4109 | if (new->entries[j].threshold < usage) { | |
4110 | /* | |
4111 | * new->current_threshold will not be used | |
4112 | * until rcu_assign_pointer(), so it's safe to increment | |
4113 | * it here. | |
4114 | */ | |
4115 | ++new->current_threshold; | |
4116 | } | |
4117 | j++; | |
4118 | } | |
4119 | ||
4120 | swap_buffers: | |
4121 | /* Swap primary and spare array */ | |
4122 | thresholds->spare = thresholds->primary; | |
4123 | rcu_assign_pointer(thresholds->primary, new); | |
4124 | ||
4125 | /* To be sure that nobody uses thresholds */ | |
4126 | synchronize_rcu(); | |
4127 | ||
4128 | mutex_unlock(&memcg->thresholds_lock); | |
4129 | } | |
4130 | ||
4131 | static int mem_cgroup_oom_register_event(struct cgroup *cgrp, | |
4132 | struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) | |
4133 | { | |
4134 | struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); | |
4135 | struct mem_cgroup_eventfd_list *event; | |
4136 | int type = MEMFILE_TYPE(cft->private); | |
4137 | ||
4138 | BUG_ON(type != _OOM_TYPE); | |
4139 | event = kmalloc(sizeof(*event), GFP_KERNEL); | |
4140 | if (!event) | |
4141 | return -ENOMEM; | |
4142 | ||
4143 | mutex_lock(&memcg_oom_mutex); | |
4144 | ||
4145 | event->eventfd = eventfd; | |
4146 | list_add(&event->list, &memcg->oom_notify); | |
4147 | ||
4148 | /* already in OOM ? */ | |
4149 | if (atomic_read(&memcg->oom_lock)) | |
4150 | eventfd_signal(eventfd, 1); | |
4151 | mutex_unlock(&memcg_oom_mutex); | |
4152 | ||
4153 | return 0; | |
4154 | } | |
4155 | ||
4156 | static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, | |
4157 | struct cftype *cft, struct eventfd_ctx *eventfd) | |
4158 | { | |
4159 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
4160 | struct mem_cgroup_eventfd_list *ev, *tmp; | |
4161 | int type = MEMFILE_TYPE(cft->private); | |
4162 | ||
4163 | BUG_ON(type != _OOM_TYPE); | |
4164 | ||
4165 | mutex_lock(&memcg_oom_mutex); | |
4166 | ||
4167 | list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) { | |
4168 | if (ev->eventfd == eventfd) { | |
4169 | list_del(&ev->list); | |
4170 | kfree(ev); | |
4171 | } | |
4172 | } | |
4173 | ||
4174 | mutex_unlock(&memcg_oom_mutex); | |
4175 | } | |
4176 | ||
4177 | static int mem_cgroup_oom_control_read(struct cgroup *cgrp, | |
4178 | struct cftype *cft, struct cgroup_map_cb *cb) | |
4179 | { | |
4180 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
4181 | ||
4182 | cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable); | |
4183 | ||
4184 | if (atomic_read(&mem->oom_lock)) | |
4185 | cb->fill(cb, "under_oom", 1); | |
4186 | else | |
4187 | cb->fill(cb, "under_oom", 0); | |
4188 | return 0; | |
4189 | } | |
4190 | ||
4191 | static int mem_cgroup_oom_control_write(struct cgroup *cgrp, | |
4192 | struct cftype *cft, u64 val) | |
4193 | { | |
4194 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); | |
4195 | struct mem_cgroup *parent; | |
4196 | ||
4197 | /* cannot set to root cgroup and only 0 and 1 are allowed */ | |
4198 | if (!cgrp->parent || !((val == 0) || (val == 1))) | |
4199 | return -EINVAL; | |
4200 | ||
4201 | parent = mem_cgroup_from_cont(cgrp->parent); | |
4202 | ||
4203 | cgroup_lock(); | |
4204 | /* oom-kill-disable is a flag for subhierarchy. */ | |
4205 | if ((parent->use_hierarchy) || | |
4206 | (mem->use_hierarchy && !list_empty(&cgrp->children))) { | |
4207 | cgroup_unlock(); | |
4208 | return -EINVAL; | |
4209 | } | |
4210 | mem->oom_kill_disable = val; | |
4211 | if (!val) | |
4212 | memcg_oom_recover(mem); | |
4213 | cgroup_unlock(); | |
4214 | return 0; | |
4215 | } | |
4216 | ||
4217 | static struct cftype mem_cgroup_files[] = { | |
4218 | { | |
4219 | .name = "usage_in_bytes", | |
4220 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), | |
4221 | .read_u64 = mem_cgroup_read, | |
4222 | .register_event = mem_cgroup_usage_register_event, | |
4223 | .unregister_event = mem_cgroup_usage_unregister_event, | |
4224 | }, | |
4225 | { | |
4226 | .name = "max_usage_in_bytes", | |
4227 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), | |
4228 | .trigger = mem_cgroup_reset, | |
4229 | .read_u64 = mem_cgroup_read, | |
4230 | }, | |
4231 | { | |
4232 | .name = "limit_in_bytes", | |
4233 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), | |
4234 | .write_string = mem_cgroup_write, | |
4235 | .read_u64 = mem_cgroup_read, | |
4236 | }, | |
4237 | { | |
4238 | .name = "soft_limit_in_bytes", | |
4239 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), | |
4240 | .write_string = mem_cgroup_write, | |
4241 | .read_u64 = mem_cgroup_read, | |
4242 | }, | |
4243 | { | |
4244 | .name = "failcnt", | |
4245 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), | |
4246 | .trigger = mem_cgroup_reset, | |
4247 | .read_u64 = mem_cgroup_read, | |
4248 | }, | |
4249 | { | |
4250 | .name = "stat", | |
4251 | .read_map = mem_control_stat_show, | |
4252 | }, | |
4253 | { | |
4254 | .name = "force_empty", | |
4255 | .trigger = mem_cgroup_force_empty_write, | |
4256 | }, | |
4257 | { | |
4258 | .name = "use_hierarchy", | |
4259 | .write_u64 = mem_cgroup_hierarchy_write, | |
4260 | .read_u64 = mem_cgroup_hierarchy_read, | |
4261 | }, | |
4262 | { | |
4263 | .name = "swappiness", | |
4264 | .read_u64 = mem_cgroup_swappiness_read, | |
4265 | .write_u64 = mem_cgroup_swappiness_write, | |
4266 | }, | |
4267 | { | |
4268 | .name = "move_charge_at_immigrate", | |
4269 | .read_u64 = mem_cgroup_move_charge_read, | |
4270 | .write_u64 = mem_cgroup_move_charge_write, | |
4271 | }, | |
4272 | { | |
4273 | .name = "oom_control", | |
4274 | .read_map = mem_cgroup_oom_control_read, | |
4275 | .write_u64 = mem_cgroup_oom_control_write, | |
4276 | .register_event = mem_cgroup_oom_register_event, | |
4277 | .unregister_event = mem_cgroup_oom_unregister_event, | |
4278 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), | |
4279 | }, | |
4280 | }; | |
4281 | ||
4282 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
4283 | static struct cftype memsw_cgroup_files[] = { | |
4284 | { | |
4285 | .name = "memsw.usage_in_bytes", | |
4286 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), | |
4287 | .read_u64 = mem_cgroup_read, | |
4288 | .register_event = mem_cgroup_usage_register_event, | |
4289 | .unregister_event = mem_cgroup_usage_unregister_event, | |
4290 | }, | |
4291 | { | |
4292 | .name = "memsw.max_usage_in_bytes", | |
4293 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), | |
4294 | .trigger = mem_cgroup_reset, | |
4295 | .read_u64 = mem_cgroup_read, | |
4296 | }, | |
4297 | { | |
4298 | .name = "memsw.limit_in_bytes", | |
4299 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), | |
4300 | .write_string = mem_cgroup_write, | |
4301 | .read_u64 = mem_cgroup_read, | |
4302 | }, | |
4303 | { | |
4304 | .name = "memsw.failcnt", | |
4305 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), | |
4306 | .trigger = mem_cgroup_reset, | |
4307 | .read_u64 = mem_cgroup_read, | |
4308 | }, | |
4309 | }; | |
4310 | ||
4311 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) | |
4312 | { | |
4313 | if (!do_swap_account) | |
4314 | return 0; | |
4315 | return cgroup_add_files(cont, ss, memsw_cgroup_files, | |
4316 | ARRAY_SIZE(memsw_cgroup_files)); | |
4317 | }; | |
4318 | #else | |
4319 | static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) | |
4320 | { | |
4321 | return 0; | |
4322 | } | |
4323 | #endif | |
4324 | ||
4325 | static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) | |
4326 | { | |
4327 | struct mem_cgroup_per_node *pn; | |
4328 | struct mem_cgroup_per_zone *mz; | |
4329 | enum lru_list l; | |
4330 | int zone, tmp = node; | |
4331 | /* | |
4332 | * This routine is called against possible nodes. | |
4333 | * But it's BUG to call kmalloc() against offline node. | |
4334 | * | |
4335 | * TODO: this routine can waste much memory for nodes which will | |
4336 | * never be onlined. It's better to use memory hotplug callback | |
4337 | * function. | |
4338 | */ | |
4339 | if (!node_state(node, N_NORMAL_MEMORY)) | |
4340 | tmp = -1; | |
4341 | pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); | |
4342 | if (!pn) | |
4343 | return 1; | |
4344 | ||
4345 | mem->info.nodeinfo[node] = pn; | |
4346 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4347 | mz = &pn->zoneinfo[zone]; | |
4348 | for_each_lru(l) | |
4349 | INIT_LIST_HEAD(&mz->lists[l]); | |
4350 | mz->usage_in_excess = 0; | |
4351 | mz->on_tree = false; | |
4352 | mz->mem = mem; | |
4353 | } | |
4354 | return 0; | |
4355 | } | |
4356 | ||
4357 | static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) | |
4358 | { | |
4359 | kfree(mem->info.nodeinfo[node]); | |
4360 | } | |
4361 | ||
4362 | static struct mem_cgroup *mem_cgroup_alloc(void) | |
4363 | { | |
4364 | struct mem_cgroup *mem; | |
4365 | int size = sizeof(struct mem_cgroup); | |
4366 | ||
4367 | /* Can be very big if MAX_NUMNODES is very big */ | |
4368 | if (size < PAGE_SIZE) | |
4369 | mem = kzalloc(size, GFP_KERNEL); | |
4370 | else | |
4371 | mem = vzalloc(size); | |
4372 | ||
4373 | if (!mem) | |
4374 | return NULL; | |
4375 | ||
4376 | mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); | |
4377 | if (!mem->stat) | |
4378 | goto out_free; | |
4379 | spin_lock_init(&mem->pcp_counter_lock); | |
4380 | return mem; | |
4381 | ||
4382 | out_free: | |
4383 | if (size < PAGE_SIZE) | |
4384 | kfree(mem); | |
4385 | else | |
4386 | vfree(mem); | |
4387 | return NULL; | |
4388 | } | |
4389 | ||
4390 | /* | |
4391 | * At destroying mem_cgroup, references from swap_cgroup can remain. | |
4392 | * (scanning all at force_empty is too costly...) | |
4393 | * | |
4394 | * Instead of clearing all references at force_empty, we remember | |
4395 | * the number of reference from swap_cgroup and free mem_cgroup when | |
4396 | * it goes down to 0. | |
4397 | * | |
4398 | * Removal of cgroup itself succeeds regardless of refs from swap. | |
4399 | */ | |
4400 | ||
4401 | static void __mem_cgroup_free(struct mem_cgroup *mem) | |
4402 | { | |
4403 | int node; | |
4404 | ||
4405 | mem_cgroup_remove_from_trees(mem); | |
4406 | free_css_id(&mem_cgroup_subsys, &mem->css); | |
4407 | ||
4408 | for_each_node_state(node, N_POSSIBLE) | |
4409 | free_mem_cgroup_per_zone_info(mem, node); | |
4410 | ||
4411 | free_percpu(mem->stat); | |
4412 | if (sizeof(struct mem_cgroup) < PAGE_SIZE) | |
4413 | kfree(mem); | |
4414 | else | |
4415 | vfree(mem); | |
4416 | } | |
4417 | ||
4418 | static void mem_cgroup_get(struct mem_cgroup *mem) | |
4419 | { | |
4420 | atomic_inc(&mem->refcnt); | |
4421 | } | |
4422 | ||
4423 | static void __mem_cgroup_put(struct mem_cgroup *mem, int count) | |
4424 | { | |
4425 | if (atomic_sub_and_test(count, &mem->refcnt)) { | |
4426 | struct mem_cgroup *parent = parent_mem_cgroup(mem); | |
4427 | __mem_cgroup_free(mem); | |
4428 | if (parent) | |
4429 | mem_cgroup_put(parent); | |
4430 | } | |
4431 | } | |
4432 | ||
4433 | static void mem_cgroup_put(struct mem_cgroup *mem) | |
4434 | { | |
4435 | __mem_cgroup_put(mem, 1); | |
4436 | } | |
4437 | ||
4438 | /* | |
4439 | * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. | |
4440 | */ | |
4441 | static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) | |
4442 | { | |
4443 | if (!mem->res.parent) | |
4444 | return NULL; | |
4445 | return mem_cgroup_from_res_counter(mem->res.parent, res); | |
4446 | } | |
4447 | ||
4448 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
4449 | static void __init enable_swap_cgroup(void) | |
4450 | { | |
4451 | if (!mem_cgroup_disabled() && really_do_swap_account) | |
4452 | do_swap_account = 1; | |
4453 | } | |
4454 | #else | |
4455 | static void __init enable_swap_cgroup(void) | |
4456 | { | |
4457 | } | |
4458 | #endif | |
4459 | ||
4460 | static int mem_cgroup_soft_limit_tree_init(void) | |
4461 | { | |
4462 | struct mem_cgroup_tree_per_node *rtpn; | |
4463 | struct mem_cgroup_tree_per_zone *rtpz; | |
4464 | int tmp, node, zone; | |
4465 | ||
4466 | for_each_node_state(node, N_POSSIBLE) { | |
4467 | tmp = node; | |
4468 | if (!node_state(node, N_NORMAL_MEMORY)) | |
4469 | tmp = -1; | |
4470 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); | |
4471 | if (!rtpn) | |
4472 | return 1; | |
4473 | ||
4474 | soft_limit_tree.rb_tree_per_node[node] = rtpn; | |
4475 | ||
4476 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4477 | rtpz = &rtpn->rb_tree_per_zone[zone]; | |
4478 | rtpz->rb_root = RB_ROOT; | |
4479 | spin_lock_init(&rtpz->lock); | |
4480 | } | |
4481 | } | |
4482 | return 0; | |
4483 | } | |
4484 | ||
4485 | static struct cgroup_subsys_state * __ref | |
4486 | mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) | |
4487 | { | |
4488 | struct mem_cgroup *mem, *parent; | |
4489 | long error = -ENOMEM; | |
4490 | int node; | |
4491 | ||
4492 | mem = mem_cgroup_alloc(); | |
4493 | if (!mem) | |
4494 | return ERR_PTR(error); | |
4495 | ||
4496 | for_each_node_state(node, N_POSSIBLE) | |
4497 | if (alloc_mem_cgroup_per_zone_info(mem, node)) | |
4498 | goto free_out; | |
4499 | ||
4500 | /* root ? */ | |
4501 | if (cont->parent == NULL) { | |
4502 | int cpu; | |
4503 | enable_swap_cgroup(); | |
4504 | parent = NULL; | |
4505 | root_mem_cgroup = mem; | |
4506 | if (mem_cgroup_soft_limit_tree_init()) | |
4507 | goto free_out; | |
4508 | for_each_possible_cpu(cpu) { | |
4509 | struct memcg_stock_pcp *stock = | |
4510 | &per_cpu(memcg_stock, cpu); | |
4511 | INIT_WORK(&stock->work, drain_local_stock); | |
4512 | } | |
4513 | hotcpu_notifier(memcg_cpu_hotplug_callback, 0); | |
4514 | } else { | |
4515 | parent = mem_cgroup_from_cont(cont->parent); | |
4516 | mem->use_hierarchy = parent->use_hierarchy; | |
4517 | mem->oom_kill_disable = parent->oom_kill_disable; | |
4518 | } | |
4519 | ||
4520 | if (parent && parent->use_hierarchy) { | |
4521 | res_counter_init(&mem->res, &parent->res); | |
4522 | res_counter_init(&mem->memsw, &parent->memsw); | |
4523 | /* | |
4524 | * We increment refcnt of the parent to ensure that we can | |
4525 | * safely access it on res_counter_charge/uncharge. | |
4526 | * This refcnt will be decremented when freeing this | |
4527 | * mem_cgroup(see mem_cgroup_put). | |
4528 | */ | |
4529 | mem_cgroup_get(parent); | |
4530 | } else { | |
4531 | res_counter_init(&mem->res, NULL); | |
4532 | res_counter_init(&mem->memsw, NULL); | |
4533 | } | |
4534 | mem->last_scanned_child = 0; | |
4535 | INIT_LIST_HEAD(&mem->oom_notify); | |
4536 | ||
4537 | if (parent) | |
4538 | mem->swappiness = get_swappiness(parent); | |
4539 | atomic_set(&mem->refcnt, 1); | |
4540 | mem->move_charge_at_immigrate = 0; | |
4541 | mutex_init(&mem->thresholds_lock); | |
4542 | return &mem->css; | |
4543 | free_out: | |
4544 | __mem_cgroup_free(mem); | |
4545 | root_mem_cgroup = NULL; | |
4546 | return ERR_PTR(error); | |
4547 | } | |
4548 | ||
4549 | static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, | |
4550 | struct cgroup *cont) | |
4551 | { | |
4552 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
4553 | ||
4554 | return mem_cgroup_force_empty(mem, false); | |
4555 | } | |
4556 | ||
4557 | static void mem_cgroup_destroy(struct cgroup_subsys *ss, | |
4558 | struct cgroup *cont) | |
4559 | { | |
4560 | struct mem_cgroup *mem = mem_cgroup_from_cont(cont); | |
4561 | ||
4562 | mem_cgroup_put(mem); | |
4563 | } | |
4564 | ||
4565 | static int mem_cgroup_populate(struct cgroup_subsys *ss, | |
4566 | struct cgroup *cont) | |
4567 | { | |
4568 | int ret; | |
4569 | ||
4570 | ret = cgroup_add_files(cont, ss, mem_cgroup_files, | |
4571 | ARRAY_SIZE(mem_cgroup_files)); | |
4572 | ||
4573 | if (!ret) | |
4574 | ret = register_memsw_files(cont, ss); | |
4575 | return ret; | |
4576 | } | |
4577 | ||
4578 | #ifdef CONFIG_MMU | |
4579 | /* Handlers for move charge at task migration. */ | |
4580 | #define PRECHARGE_COUNT_AT_ONCE 256 | |
4581 | static int mem_cgroup_do_precharge(unsigned long count) | |
4582 | { | |
4583 | int ret = 0; | |
4584 | int batch_count = PRECHARGE_COUNT_AT_ONCE; | |
4585 | struct mem_cgroup *mem = mc.to; | |
4586 | ||
4587 | if (mem_cgroup_is_root(mem)) { | |
4588 | mc.precharge += count; | |
4589 | /* we don't need css_get for root */ | |
4590 | return ret; | |
4591 | } | |
4592 | /* try to charge at once */ | |
4593 | if (count > 1) { | |
4594 | struct res_counter *dummy; | |
4595 | /* | |
4596 | * "mem" cannot be under rmdir() because we've already checked | |
4597 | * by cgroup_lock_live_cgroup() that it is not removed and we | |
4598 | * are still under the same cgroup_mutex. So we can postpone | |
4599 | * css_get(). | |
4600 | */ | |
4601 | if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy)) | |
4602 | goto one_by_one; | |
4603 | if (do_swap_account && res_counter_charge(&mem->memsw, | |
4604 | PAGE_SIZE * count, &dummy)) { | |
4605 | res_counter_uncharge(&mem->res, PAGE_SIZE * count); | |
4606 | goto one_by_one; | |
4607 | } | |
4608 | mc.precharge += count; | |
4609 | return ret; | |
4610 | } | |
4611 | one_by_one: | |
4612 | /* fall back to one by one charge */ | |
4613 | while (count--) { | |
4614 | if (signal_pending(current)) { | |
4615 | ret = -EINTR; | |
4616 | break; | |
4617 | } | |
4618 | if (!batch_count--) { | |
4619 | batch_count = PRECHARGE_COUNT_AT_ONCE; | |
4620 | cond_resched(); | |
4621 | } | |
4622 | ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false); | |
4623 | if (ret || !mem) | |
4624 | /* mem_cgroup_clear_mc() will do uncharge later */ | |
4625 | return -ENOMEM; | |
4626 | mc.precharge++; | |
4627 | } | |
4628 | return ret; | |
4629 | } | |
4630 | ||
4631 | /** | |
4632 | * is_target_pte_for_mc - check a pte whether it is valid for move charge | |
4633 | * @vma: the vma the pte to be checked belongs | |
4634 | * @addr: the address corresponding to the pte to be checked | |
4635 | * @ptent: the pte to be checked | |
4636 | * @target: the pointer the target page or swap ent will be stored(can be NULL) | |
4637 | * | |
4638 | * Returns | |
4639 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. | |
4640 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for | |
4641 | * move charge. if @target is not NULL, the page is stored in target->page | |
4642 | * with extra refcnt got(Callers should handle it). | |
4643 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a | |
4644 | * target for charge migration. if @target is not NULL, the entry is stored | |
4645 | * in target->ent. | |
4646 | * | |
4647 | * Called with pte lock held. | |
4648 | */ | |
4649 | union mc_target { | |
4650 | struct page *page; | |
4651 | swp_entry_t ent; | |
4652 | }; | |
4653 | ||
4654 | enum mc_target_type { | |
4655 | MC_TARGET_NONE, /* not used */ | |
4656 | MC_TARGET_PAGE, | |
4657 | MC_TARGET_SWAP, | |
4658 | }; | |
4659 | ||
4660 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, | |
4661 | unsigned long addr, pte_t ptent) | |
4662 | { | |
4663 | struct page *page = vm_normal_page(vma, addr, ptent); | |
4664 | ||
4665 | if (!page || !page_mapped(page)) | |
4666 | return NULL; | |
4667 | if (PageAnon(page)) { | |
4668 | /* we don't move shared anon */ | |
4669 | if (!move_anon() || page_mapcount(page) > 2) | |
4670 | return NULL; | |
4671 | } else if (!move_file()) | |
4672 | /* we ignore mapcount for file pages */ | |
4673 | return NULL; | |
4674 | if (!get_page_unless_zero(page)) | |
4675 | return NULL; | |
4676 | ||
4677 | return page; | |
4678 | } | |
4679 | ||
4680 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
4681 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
4682 | { | |
4683 | int usage_count; | |
4684 | struct page *page = NULL; | |
4685 | swp_entry_t ent = pte_to_swp_entry(ptent); | |
4686 | ||
4687 | if (!move_anon() || non_swap_entry(ent)) | |
4688 | return NULL; | |
4689 | usage_count = mem_cgroup_count_swap_user(ent, &page); | |
4690 | if (usage_count > 1) { /* we don't move shared anon */ | |
4691 | if (page) | |
4692 | put_page(page); | |
4693 | return NULL; | |
4694 | } | |
4695 | if (do_swap_account) | |
4696 | entry->val = ent.val; | |
4697 | ||
4698 | return page; | |
4699 | } | |
4700 | ||
4701 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, | |
4702 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
4703 | { | |
4704 | struct page *page = NULL; | |
4705 | struct inode *inode; | |
4706 | struct address_space *mapping; | |
4707 | pgoff_t pgoff; | |
4708 | ||
4709 | if (!vma->vm_file) /* anonymous vma */ | |
4710 | return NULL; | |
4711 | if (!move_file()) | |
4712 | return NULL; | |
4713 | ||
4714 | inode = vma->vm_file->f_path.dentry->d_inode; | |
4715 | mapping = vma->vm_file->f_mapping; | |
4716 | if (pte_none(ptent)) | |
4717 | pgoff = linear_page_index(vma, addr); | |
4718 | else /* pte_file(ptent) is true */ | |
4719 | pgoff = pte_to_pgoff(ptent); | |
4720 | ||
4721 | /* page is moved even if it's not RSS of this task(page-faulted). */ | |
4722 | if (!mapping_cap_swap_backed(mapping)) { /* normal file */ | |
4723 | page = find_get_page(mapping, pgoff); | |
4724 | } else { /* shmem/tmpfs file. we should take account of swap too. */ | |
4725 | swp_entry_t ent; | |
4726 | mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent); | |
4727 | if (do_swap_account) | |
4728 | entry->val = ent.val; | |
4729 | } | |
4730 | ||
4731 | return page; | |
4732 | } | |
4733 | ||
4734 | static int is_target_pte_for_mc(struct vm_area_struct *vma, | |
4735 | unsigned long addr, pte_t ptent, union mc_target *target) | |
4736 | { | |
4737 | struct page *page = NULL; | |
4738 | struct page_cgroup *pc; | |
4739 | int ret = 0; | |
4740 | swp_entry_t ent = { .val = 0 }; | |
4741 | ||
4742 | if (pte_present(ptent)) | |
4743 | page = mc_handle_present_pte(vma, addr, ptent); | |
4744 | else if (is_swap_pte(ptent)) | |
4745 | page = mc_handle_swap_pte(vma, addr, ptent, &ent); | |
4746 | else if (pte_none(ptent) || pte_file(ptent)) | |
4747 | page = mc_handle_file_pte(vma, addr, ptent, &ent); | |
4748 | ||
4749 | if (!page && !ent.val) | |
4750 | return 0; | |
4751 | if (page) { | |
4752 | pc = lookup_page_cgroup(page); | |
4753 | /* | |
4754 | * Do only loose check w/o page_cgroup lock. | |
4755 | * mem_cgroup_move_account() checks the pc is valid or not under | |
4756 | * the lock. | |
4757 | */ | |
4758 | if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { | |
4759 | ret = MC_TARGET_PAGE; | |
4760 | if (target) | |
4761 | target->page = page; | |
4762 | } | |
4763 | if (!ret || !target) | |
4764 | put_page(page); | |
4765 | } | |
4766 | /* There is a swap entry and a page doesn't exist or isn't charged */ | |
4767 | if (ent.val && !ret && | |
4768 | css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { | |
4769 | ret = MC_TARGET_SWAP; | |
4770 | if (target) | |
4771 | target->ent = ent; | |
4772 | } | |
4773 | return ret; | |
4774 | } | |
4775 | ||
4776 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, | |
4777 | unsigned long addr, unsigned long end, | |
4778 | struct mm_walk *walk) | |
4779 | { | |
4780 | struct vm_area_struct *vma = walk->private; | |
4781 | pte_t *pte; | |
4782 | spinlock_t *ptl; | |
4783 | ||
4784 | split_huge_page_pmd(walk->mm, pmd); | |
4785 | ||
4786 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
4787 | for (; addr != end; pte++, addr += PAGE_SIZE) | |
4788 | if (is_target_pte_for_mc(vma, addr, *pte, NULL)) | |
4789 | mc.precharge++; /* increment precharge temporarily */ | |
4790 | pte_unmap_unlock(pte - 1, ptl); | |
4791 | cond_resched(); | |
4792 | ||
4793 | return 0; | |
4794 | } | |
4795 | ||
4796 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) | |
4797 | { | |
4798 | unsigned long precharge; | |
4799 | struct vm_area_struct *vma; | |
4800 | ||
4801 | down_read(&mm->mmap_sem); | |
4802 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
4803 | struct mm_walk mem_cgroup_count_precharge_walk = { | |
4804 | .pmd_entry = mem_cgroup_count_precharge_pte_range, | |
4805 | .mm = mm, | |
4806 | .private = vma, | |
4807 | }; | |
4808 | if (is_vm_hugetlb_page(vma)) | |
4809 | continue; | |
4810 | walk_page_range(vma->vm_start, vma->vm_end, | |
4811 | &mem_cgroup_count_precharge_walk); | |
4812 | } | |
4813 | up_read(&mm->mmap_sem); | |
4814 | ||
4815 | precharge = mc.precharge; | |
4816 | mc.precharge = 0; | |
4817 | ||
4818 | return precharge; | |
4819 | } | |
4820 | ||
4821 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) | |
4822 | { | |
4823 | unsigned long precharge = mem_cgroup_count_precharge(mm); | |
4824 | ||
4825 | VM_BUG_ON(mc.moving_task); | |
4826 | mc.moving_task = current; | |
4827 | return mem_cgroup_do_precharge(precharge); | |
4828 | } | |
4829 | ||
4830 | /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ | |
4831 | static void __mem_cgroup_clear_mc(void) | |
4832 | { | |
4833 | struct mem_cgroup *from = mc.from; | |
4834 | struct mem_cgroup *to = mc.to; | |
4835 | ||
4836 | /* we must uncharge all the leftover precharges from mc.to */ | |
4837 | if (mc.precharge) { | |
4838 | __mem_cgroup_cancel_charge(mc.to, mc.precharge); | |
4839 | mc.precharge = 0; | |
4840 | } | |
4841 | /* | |
4842 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so | |
4843 | * we must uncharge here. | |
4844 | */ | |
4845 | if (mc.moved_charge) { | |
4846 | __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); | |
4847 | mc.moved_charge = 0; | |
4848 | } | |
4849 | /* we must fixup refcnts and charges */ | |
4850 | if (mc.moved_swap) { | |
4851 | /* uncharge swap account from the old cgroup */ | |
4852 | if (!mem_cgroup_is_root(mc.from)) | |
4853 | res_counter_uncharge(&mc.from->memsw, | |
4854 | PAGE_SIZE * mc.moved_swap); | |
4855 | __mem_cgroup_put(mc.from, mc.moved_swap); | |
4856 | ||
4857 | if (!mem_cgroup_is_root(mc.to)) { | |
4858 | /* | |
4859 | * we charged both to->res and to->memsw, so we should | |
4860 | * uncharge to->res. | |
4861 | */ | |
4862 | res_counter_uncharge(&mc.to->res, | |
4863 | PAGE_SIZE * mc.moved_swap); | |
4864 | } | |
4865 | /* we've already done mem_cgroup_get(mc.to) */ | |
4866 | mc.moved_swap = 0; | |
4867 | } | |
4868 | memcg_oom_recover(from); | |
4869 | memcg_oom_recover(to); | |
4870 | wake_up_all(&mc.waitq); | |
4871 | } | |
4872 | ||
4873 | static void mem_cgroup_clear_mc(void) | |
4874 | { | |
4875 | struct mem_cgroup *from = mc.from; | |
4876 | ||
4877 | /* | |
4878 | * we must clear moving_task before waking up waiters at the end of | |
4879 | * task migration. | |
4880 | */ | |
4881 | mc.moving_task = NULL; | |
4882 | __mem_cgroup_clear_mc(); | |
4883 | spin_lock(&mc.lock); | |
4884 | mc.from = NULL; | |
4885 | mc.to = NULL; | |
4886 | spin_unlock(&mc.lock); | |
4887 | mem_cgroup_end_move(from); | |
4888 | } | |
4889 | ||
4890 | static int mem_cgroup_can_attach(struct cgroup_subsys *ss, | |
4891 | struct cgroup *cgroup, | |
4892 | struct task_struct *p, | |
4893 | bool threadgroup) | |
4894 | { | |
4895 | int ret = 0; | |
4896 | struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup); | |
4897 | ||
4898 | if (mem->move_charge_at_immigrate) { | |
4899 | struct mm_struct *mm; | |
4900 | struct mem_cgroup *from = mem_cgroup_from_task(p); | |
4901 | ||
4902 | VM_BUG_ON(from == mem); | |
4903 | ||
4904 | mm = get_task_mm(p); | |
4905 | if (!mm) | |
4906 | return 0; | |
4907 | /* We move charges only when we move a owner of the mm */ | |
4908 | if (mm->owner == p) { | |
4909 | VM_BUG_ON(mc.from); | |
4910 | VM_BUG_ON(mc.to); | |
4911 | VM_BUG_ON(mc.precharge); | |
4912 | VM_BUG_ON(mc.moved_charge); | |
4913 | VM_BUG_ON(mc.moved_swap); | |
4914 | mem_cgroup_start_move(from); | |
4915 | spin_lock(&mc.lock); | |
4916 | mc.from = from; | |
4917 | mc.to = mem; | |
4918 | spin_unlock(&mc.lock); | |
4919 | /* We set mc.moving_task later */ | |
4920 | ||
4921 | ret = mem_cgroup_precharge_mc(mm); | |
4922 | if (ret) | |
4923 | mem_cgroup_clear_mc(); | |
4924 | } | |
4925 | mmput(mm); | |
4926 | } | |
4927 | return ret; | |
4928 | } | |
4929 | ||
4930 | static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, | |
4931 | struct cgroup *cgroup, | |
4932 | struct task_struct *p, | |
4933 | bool threadgroup) | |
4934 | { | |
4935 | mem_cgroup_clear_mc(); | |
4936 | } | |
4937 | ||
4938 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, | |
4939 | unsigned long addr, unsigned long end, | |
4940 | struct mm_walk *walk) | |
4941 | { | |
4942 | int ret = 0; | |
4943 | struct vm_area_struct *vma = walk->private; | |
4944 | pte_t *pte; | |
4945 | spinlock_t *ptl; | |
4946 | ||
4947 | split_huge_page_pmd(walk->mm, pmd); | |
4948 | retry: | |
4949 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
4950 | for (; addr != end; addr += PAGE_SIZE) { | |
4951 | pte_t ptent = *(pte++); | |
4952 | union mc_target target; | |
4953 | int type; | |
4954 | struct page *page; | |
4955 | struct page_cgroup *pc; | |
4956 | swp_entry_t ent; | |
4957 | ||
4958 | if (!mc.precharge) | |
4959 | break; | |
4960 | ||
4961 | type = is_target_pte_for_mc(vma, addr, ptent, &target); | |
4962 | switch (type) { | |
4963 | case MC_TARGET_PAGE: | |
4964 | page = target.page; | |
4965 | if (isolate_lru_page(page)) | |
4966 | goto put; | |
4967 | pc = lookup_page_cgroup(page); | |
4968 | if (!mem_cgroup_move_account(page, 1, pc, | |
4969 | mc.from, mc.to, false)) { | |
4970 | mc.precharge--; | |
4971 | /* we uncharge from mc.from later. */ | |
4972 | mc.moved_charge++; | |
4973 | } | |
4974 | putback_lru_page(page); | |
4975 | put: /* is_target_pte_for_mc() gets the page */ | |
4976 | put_page(page); | |
4977 | break; | |
4978 | case MC_TARGET_SWAP: | |
4979 | ent = target.ent; | |
4980 | if (!mem_cgroup_move_swap_account(ent, | |
4981 | mc.from, mc.to, false)) { | |
4982 | mc.precharge--; | |
4983 | /* we fixup refcnts and charges later. */ | |
4984 | mc.moved_swap++; | |
4985 | } | |
4986 | break; | |
4987 | default: | |
4988 | break; | |
4989 | } | |
4990 | } | |
4991 | pte_unmap_unlock(pte - 1, ptl); | |
4992 | cond_resched(); | |
4993 | ||
4994 | if (addr != end) { | |
4995 | /* | |
4996 | * We have consumed all precharges we got in can_attach(). | |
4997 | * We try charge one by one, but don't do any additional | |
4998 | * charges to mc.to if we have failed in charge once in attach() | |
4999 | * phase. | |
5000 | */ | |
5001 | ret = mem_cgroup_do_precharge(1); | |
5002 | if (!ret) | |
5003 | goto retry; | |
5004 | } | |
5005 | ||
5006 | return ret; | |
5007 | } | |
5008 | ||
5009 | static void mem_cgroup_move_charge(struct mm_struct *mm) | |
5010 | { | |
5011 | struct vm_area_struct *vma; | |
5012 | ||
5013 | lru_add_drain_all(); | |
5014 | retry: | |
5015 | if (unlikely(!down_read_trylock(&mm->mmap_sem))) { | |
5016 | /* | |
5017 | * Someone who are holding the mmap_sem might be waiting in | |
5018 | * waitq. So we cancel all extra charges, wake up all waiters, | |
5019 | * and retry. Because we cancel precharges, we might not be able | |
5020 | * to move enough charges, but moving charge is a best-effort | |
5021 | * feature anyway, so it wouldn't be a big problem. | |
5022 | */ | |
5023 | __mem_cgroup_clear_mc(); | |
5024 | cond_resched(); | |
5025 | goto retry; | |
5026 | } | |
5027 | for (vma = mm->mmap; vma; vma = vma->vm_next) { | |
5028 | int ret; | |
5029 | struct mm_walk mem_cgroup_move_charge_walk = { | |
5030 | .pmd_entry = mem_cgroup_move_charge_pte_range, | |
5031 | .mm = mm, | |
5032 | .private = vma, | |
5033 | }; | |
5034 | if (is_vm_hugetlb_page(vma)) | |
5035 | continue; | |
5036 | ret = walk_page_range(vma->vm_start, vma->vm_end, | |
5037 | &mem_cgroup_move_charge_walk); | |
5038 | if (ret) | |
5039 | /* | |
5040 | * means we have consumed all precharges and failed in | |
5041 | * doing additional charge. Just abandon here. | |
5042 | */ | |
5043 | break; | |
5044 | } | |
5045 | up_read(&mm->mmap_sem); | |
5046 | } | |
5047 | ||
5048 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, | |
5049 | struct cgroup *cont, | |
5050 | struct cgroup *old_cont, | |
5051 | struct task_struct *p, | |
5052 | bool threadgroup) | |
5053 | { | |
5054 | struct mm_struct *mm; | |
5055 | ||
5056 | if (!mc.to) | |
5057 | /* no need to move charge */ | |
5058 | return; | |
5059 | ||
5060 | mm = get_task_mm(p); | |
5061 | if (mm) { | |
5062 | mem_cgroup_move_charge(mm); | |
5063 | mmput(mm); | |
5064 | } | |
5065 | mem_cgroup_clear_mc(); | |
5066 | } | |
5067 | #else /* !CONFIG_MMU */ | |
5068 | static int mem_cgroup_can_attach(struct cgroup_subsys *ss, | |
5069 | struct cgroup *cgroup, | |
5070 | struct task_struct *p, | |
5071 | bool threadgroup) | |
5072 | { | |
5073 | return 0; | |
5074 | } | |
5075 | static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, | |
5076 | struct cgroup *cgroup, | |
5077 | struct task_struct *p, | |
5078 | bool threadgroup) | |
5079 | { | |
5080 | } | |
5081 | static void mem_cgroup_move_task(struct cgroup_subsys *ss, | |
5082 | struct cgroup *cont, | |
5083 | struct cgroup *old_cont, | |
5084 | struct task_struct *p, | |
5085 | bool threadgroup) | |
5086 | { | |
5087 | } | |
5088 | #endif | |
5089 | ||
5090 | struct cgroup_subsys mem_cgroup_subsys = { | |
5091 | .name = "memory", | |
5092 | .subsys_id = mem_cgroup_subsys_id, | |
5093 | .create = mem_cgroup_create, | |
5094 | .pre_destroy = mem_cgroup_pre_destroy, | |
5095 | .destroy = mem_cgroup_destroy, | |
5096 | .populate = mem_cgroup_populate, | |
5097 | .can_attach = mem_cgroup_can_attach, | |
5098 | .cancel_attach = mem_cgroup_cancel_attach, | |
5099 | .attach = mem_cgroup_move_task, | |
5100 | .early_init = 0, | |
5101 | .use_id = 1, | |
5102 | }; | |
5103 | ||
5104 | #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP | |
5105 | static int __init enable_swap_account(char *s) | |
5106 | { | |
5107 | /* consider enabled if no parameter or 1 is given */ | |
5108 | if (!(*s) || !strcmp(s, "=1")) | |
5109 | really_do_swap_account = 1; | |
5110 | else if (!strcmp(s, "=0")) | |
5111 | really_do_swap_account = 0; | |
5112 | return 1; | |
5113 | } | |
5114 | __setup("swapaccount", enable_swap_account); | |
5115 | ||
5116 | static int __init disable_swap_account(char *s) | |
5117 | { | |
5118 | printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n"); | |
5119 | enable_swap_account("=0"); | |
5120 | return 1; | |
5121 | } | |
5122 | __setup("noswapaccount", disable_swap_account); | |
5123 | #endif |