]>
Commit | Line | Data |
---|---|---|
81819f0f CL |
1 | /* |
2 | * SLUB: A slab allocator that limits cache line use instead of queuing | |
3 | * objects in per cpu and per node lists. | |
4 | * | |
5 | * The allocator synchronizes using per slab locks and only | |
6 | * uses a centralized lock to manage a pool of partial slabs. | |
7 | * | |
8 | * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com> | |
9 | */ | |
10 | ||
11 | #include <linux/mm.h> | |
12 | #include <linux/module.h> | |
13 | #include <linux/bit_spinlock.h> | |
14 | #include <linux/interrupt.h> | |
15 | #include <linux/bitops.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/seq_file.h> | |
18 | #include <linux/cpu.h> | |
19 | #include <linux/cpuset.h> | |
20 | #include <linux/mempolicy.h> | |
21 | #include <linux/ctype.h> | |
22 | #include <linux/kallsyms.h> | |
23 | ||
24 | /* | |
25 | * Lock order: | |
26 | * 1. slab_lock(page) | |
27 | * 2. slab->list_lock | |
28 | * | |
29 | * The slab_lock protects operations on the object of a particular | |
30 | * slab and its metadata in the page struct. If the slab lock | |
31 | * has been taken then no allocations nor frees can be performed | |
32 | * on the objects in the slab nor can the slab be added or removed | |
33 | * from the partial or full lists since this would mean modifying | |
34 | * the page_struct of the slab. | |
35 | * | |
36 | * The list_lock protects the partial and full list on each node and | |
37 | * the partial slab counter. If taken then no new slabs may be added or | |
38 | * removed from the lists nor make the number of partial slabs be modified. | |
39 | * (Note that the total number of slabs is an atomic value that may be | |
40 | * modified without taking the list lock). | |
41 | * | |
42 | * The list_lock is a centralized lock and thus we avoid taking it as | |
43 | * much as possible. As long as SLUB does not have to handle partial | |
44 | * slabs, operations can continue without any centralized lock. F.e. | |
45 | * allocating a long series of objects that fill up slabs does not require | |
46 | * the list lock. | |
47 | * | |
48 | * The lock order is sometimes inverted when we are trying to get a slab | |
49 | * off a list. We take the list_lock and then look for a page on the list | |
50 | * to use. While we do that objects in the slabs may be freed. We can | |
51 | * only operate on the slab if we have also taken the slab_lock. So we use | |
52 | * a slab_trylock() on the slab. If trylock was successful then no frees | |
53 | * can occur anymore and we can use the slab for allocations etc. If the | |
54 | * slab_trylock() does not succeed then frees are in progress in the slab and | |
55 | * we must stay away from it for a while since we may cause a bouncing | |
56 | * cacheline if we try to acquire the lock. So go onto the next slab. | |
57 | * If all pages are busy then we may allocate a new slab instead of reusing | |
58 | * a partial slab. A new slab has noone operating on it and thus there is | |
59 | * no danger of cacheline contention. | |
60 | * | |
61 | * Interrupts are disabled during allocation and deallocation in order to | |
62 | * make the slab allocator safe to use in the context of an irq. In addition | |
63 | * interrupts are disabled to ensure that the processor does not change | |
64 | * while handling per_cpu slabs, due to kernel preemption. | |
65 | * | |
66 | * SLUB assigns one slab for allocation to each processor. | |
67 | * Allocations only occur from these slabs called cpu slabs. | |
68 | * | |
672bba3a CL |
69 | * Slabs with free elements are kept on a partial list and during regular |
70 | * operations no list for full slabs is used. If an object in a full slab is | |
81819f0f | 71 | * freed then the slab will show up again on the partial lists. |
672bba3a CL |
72 | * We track full slabs for debugging purposes though because otherwise we |
73 | * cannot scan all objects. | |
81819f0f CL |
74 | * |
75 | * Slabs are freed when they become empty. Teardown and setup is | |
76 | * minimal so we rely on the page allocators per cpu caches for | |
77 | * fast frees and allocs. | |
78 | * | |
79 | * Overloading of page flags that are otherwise used for LRU management. | |
80 | * | |
81 | * PageActive The slab is used as a cpu cache. Allocations | |
82 | * may be performed from the slab. The slab is not | |
83 | * on any slab list and cannot be moved onto one. | |
84 | * | |
85 | * PageError Slab requires special handling due to debug | |
86 | * options set. This moves slab handling out of | |
87 | * the fast path. | |
88 | */ | |
89 | ||
90 | /* | |
91 | * Issues still to be resolved: | |
92 | * | |
93 | * - The per cpu array is updated for each new slab and and is a remote | |
94 | * cacheline for most nodes. This could become a bouncing cacheline given | |
672bba3a CL |
95 | * enough frequent updates. There are 16 pointers in a cacheline, so at |
96 | * max 16 cpus could compete for the cacheline which may be okay. | |
81819f0f CL |
97 | * |
98 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
99 | * | |
81819f0f CL |
100 | * - Variable sizing of the per node arrays |
101 | */ | |
102 | ||
103 | /* Enable to test recovery from slab corruption on boot */ | |
104 | #undef SLUB_RESILIENCY_TEST | |
105 | ||
106 | #if PAGE_SHIFT <= 12 | |
107 | ||
108 | /* | |
109 | * Small page size. Make sure that we do not fragment memory | |
110 | */ | |
111 | #define DEFAULT_MAX_ORDER 1 | |
112 | #define DEFAULT_MIN_OBJECTS 4 | |
113 | ||
114 | #else | |
115 | ||
116 | /* | |
117 | * Large page machines are customarily able to handle larger | |
118 | * page orders. | |
119 | */ | |
120 | #define DEFAULT_MAX_ORDER 2 | |
121 | #define DEFAULT_MIN_OBJECTS 8 | |
122 | ||
123 | #endif | |
124 | ||
2086d26a CL |
125 | /* |
126 | * Mininum number of partial slabs. These will be left on the partial | |
127 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
128 | */ | |
e95eed57 CL |
129 | #define MIN_PARTIAL 2 |
130 | ||
2086d26a CL |
131 | /* |
132 | * Maximum number of desirable partial slabs. | |
133 | * The existence of more partial slabs makes kmem_cache_shrink | |
134 | * sort the partial list by the number of objects in the. | |
135 | */ | |
136 | #define MAX_PARTIAL 10 | |
137 | ||
81819f0f CL |
138 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ |
139 | SLAB_POISON | SLAB_STORE_USER) | |
672bba3a | 140 | |
81819f0f CL |
141 | /* |
142 | * Set of flags that will prevent slab merging | |
143 | */ | |
144 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
145 | SLAB_TRACE | SLAB_DESTROY_BY_RCU) | |
146 | ||
147 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ | |
148 | SLAB_CACHE_DMA) | |
149 | ||
150 | #ifndef ARCH_KMALLOC_MINALIGN | |
47bfdc0d | 151 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
152 | #endif |
153 | ||
154 | #ifndef ARCH_SLAB_MINALIGN | |
47bfdc0d | 155 | #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
156 | #endif |
157 | ||
158 | /* Internal SLUB flags */ | |
159 | #define __OBJECT_POISON 0x80000000 /* Poison object */ | |
160 | ||
65c02d4c CL |
161 | /* Not all arches define cache_line_size */ |
162 | #ifndef cache_line_size | |
163 | #define cache_line_size() L1_CACHE_BYTES | |
164 | #endif | |
165 | ||
81819f0f CL |
166 | static int kmem_size = sizeof(struct kmem_cache); |
167 | ||
168 | #ifdef CONFIG_SMP | |
169 | static struct notifier_block slab_notifier; | |
170 | #endif | |
171 | ||
172 | static enum { | |
173 | DOWN, /* No slab functionality available */ | |
174 | PARTIAL, /* kmem_cache_open() works but kmalloc does not */ | |
672bba3a | 175 | UP, /* Everything works but does not show up in sysfs */ |
81819f0f CL |
176 | SYSFS /* Sysfs up */ |
177 | } slab_state = DOWN; | |
178 | ||
179 | /* A list of all slab caches on the system */ | |
180 | static DECLARE_RWSEM(slub_lock); | |
181 | LIST_HEAD(slab_caches); | |
182 | ||
183 | #ifdef CONFIG_SYSFS | |
184 | static int sysfs_slab_add(struct kmem_cache *); | |
185 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
186 | static void sysfs_slab_remove(struct kmem_cache *); | |
187 | #else | |
188 | static int sysfs_slab_add(struct kmem_cache *s) { return 0; } | |
189 | static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; } | |
190 | static void sysfs_slab_remove(struct kmem_cache *s) {} | |
191 | #endif | |
192 | ||
193 | /******************************************************************** | |
194 | * Core slab cache functions | |
195 | *******************************************************************/ | |
196 | ||
197 | int slab_is_available(void) | |
198 | { | |
199 | return slab_state >= UP; | |
200 | } | |
201 | ||
202 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
203 | { | |
204 | #ifdef CONFIG_NUMA | |
205 | return s->node[node]; | |
206 | #else | |
207 | return &s->local_node; | |
208 | #endif | |
209 | } | |
210 | ||
211 | /* | |
212 | * Object debugging | |
213 | */ | |
214 | static void print_section(char *text, u8 *addr, unsigned int length) | |
215 | { | |
216 | int i, offset; | |
217 | int newline = 1; | |
218 | char ascii[17]; | |
219 | ||
220 | ascii[16] = 0; | |
221 | ||
222 | for (i = 0; i < length; i++) { | |
223 | if (newline) { | |
224 | printk(KERN_ERR "%10s 0x%p: ", text, addr + i); | |
225 | newline = 0; | |
226 | } | |
227 | printk(" %02x", addr[i]); | |
228 | offset = i % 16; | |
229 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; | |
230 | if (offset == 15) { | |
231 | printk(" %s\n",ascii); | |
232 | newline = 1; | |
233 | } | |
234 | } | |
235 | if (!newline) { | |
236 | i %= 16; | |
237 | while (i < 16) { | |
238 | printk(" "); | |
239 | ascii[i] = ' '; | |
240 | i++; | |
241 | } | |
242 | printk(" %s\n", ascii); | |
243 | } | |
244 | } | |
245 | ||
246 | /* | |
247 | * Slow version of get and set free pointer. | |
248 | * | |
672bba3a CL |
249 | * This version requires touching the cache lines of kmem_cache which |
250 | * we avoid to do in the fast alloc free paths. There we obtain the offset | |
251 | * from the page struct. | |
81819f0f CL |
252 | */ |
253 | static void *get_freepointer(struct kmem_cache *s, void *object) | |
254 | { | |
255 | return *(void **)(object + s->offset); | |
256 | } | |
257 | ||
258 | static void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
259 | { | |
260 | *(void **)(object + s->offset) = fp; | |
261 | } | |
262 | ||
263 | /* | |
264 | * Tracking user of a slab. | |
265 | */ | |
266 | struct track { | |
267 | void *addr; /* Called from address */ | |
268 | int cpu; /* Was running on cpu */ | |
269 | int pid; /* Pid context */ | |
270 | unsigned long when; /* When did the operation occur */ | |
271 | }; | |
272 | ||
273 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
274 | ||
275 | static struct track *get_track(struct kmem_cache *s, void *object, | |
276 | enum track_item alloc) | |
277 | { | |
278 | struct track *p; | |
279 | ||
280 | if (s->offset) | |
281 | p = object + s->offset + sizeof(void *); | |
282 | else | |
283 | p = object + s->inuse; | |
284 | ||
285 | return p + alloc; | |
286 | } | |
287 | ||
288 | static void set_track(struct kmem_cache *s, void *object, | |
289 | enum track_item alloc, void *addr) | |
290 | { | |
291 | struct track *p; | |
292 | ||
293 | if (s->offset) | |
294 | p = object + s->offset + sizeof(void *); | |
295 | else | |
296 | p = object + s->inuse; | |
297 | ||
298 | p += alloc; | |
299 | if (addr) { | |
300 | p->addr = addr; | |
301 | p->cpu = smp_processor_id(); | |
302 | p->pid = current ? current->pid : -1; | |
303 | p->when = jiffies; | |
304 | } else | |
305 | memset(p, 0, sizeof(struct track)); | |
306 | } | |
307 | ||
81819f0f CL |
308 | static void init_tracking(struct kmem_cache *s, void *object) |
309 | { | |
310 | if (s->flags & SLAB_STORE_USER) { | |
311 | set_track(s, object, TRACK_FREE, NULL); | |
312 | set_track(s, object, TRACK_ALLOC, NULL); | |
313 | } | |
314 | } | |
315 | ||
316 | static void print_track(const char *s, struct track *t) | |
317 | { | |
318 | if (!t->addr) | |
319 | return; | |
320 | ||
321 | printk(KERN_ERR "%s: ", s); | |
322 | __print_symbol("%s", (unsigned long)t->addr); | |
323 | printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); | |
324 | } | |
325 | ||
326 | static void print_trailer(struct kmem_cache *s, u8 *p) | |
327 | { | |
328 | unsigned int off; /* Offset of last byte */ | |
329 | ||
330 | if (s->flags & SLAB_RED_ZONE) | |
331 | print_section("Redzone", p + s->objsize, | |
332 | s->inuse - s->objsize); | |
333 | ||
334 | printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n", | |
335 | p + s->offset, | |
336 | get_freepointer(s, p)); | |
337 | ||
338 | if (s->offset) | |
339 | off = s->offset + sizeof(void *); | |
340 | else | |
341 | off = s->inuse; | |
342 | ||
343 | if (s->flags & SLAB_STORE_USER) { | |
344 | print_track("Last alloc", get_track(s, p, TRACK_ALLOC)); | |
345 | print_track("Last free ", get_track(s, p, TRACK_FREE)); | |
346 | off += 2 * sizeof(struct track); | |
347 | } | |
348 | ||
349 | if (off != s->size) | |
350 | /* Beginning of the filler is the free pointer */ | |
351 | print_section("Filler", p + off, s->size - off); | |
352 | } | |
353 | ||
354 | static void object_err(struct kmem_cache *s, struct page *page, | |
355 | u8 *object, char *reason) | |
356 | { | |
357 | u8 *addr = page_address(page); | |
358 | ||
359 | printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n", | |
360 | s->name, reason, object, page); | |
361 | printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n", | |
362 | object - addr, page->flags, page->inuse, page->freelist); | |
363 | if (object > addr + 16) | |
364 | print_section("Bytes b4", object - 16, 16); | |
365 | print_section("Object", object, min(s->objsize, 128)); | |
366 | print_trailer(s, object); | |
367 | dump_stack(); | |
368 | } | |
369 | ||
370 | static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...) | |
371 | { | |
372 | va_list args; | |
373 | char buf[100]; | |
374 | ||
375 | va_start(args, reason); | |
376 | vsnprintf(buf, sizeof(buf), reason, args); | |
377 | va_end(args); | |
378 | printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf, | |
379 | page); | |
380 | dump_stack(); | |
381 | } | |
382 | ||
383 | static void init_object(struct kmem_cache *s, void *object, int active) | |
384 | { | |
385 | u8 *p = object; | |
386 | ||
387 | if (s->flags & __OBJECT_POISON) { | |
388 | memset(p, POISON_FREE, s->objsize - 1); | |
389 | p[s->objsize -1] = POISON_END; | |
390 | } | |
391 | ||
392 | if (s->flags & SLAB_RED_ZONE) | |
393 | memset(p + s->objsize, | |
394 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, | |
395 | s->inuse - s->objsize); | |
396 | } | |
397 | ||
398 | static int check_bytes(u8 *start, unsigned int value, unsigned int bytes) | |
399 | { | |
400 | while (bytes) { | |
401 | if (*start != (u8)value) | |
402 | return 0; | |
403 | start++; | |
404 | bytes--; | |
405 | } | |
406 | return 1; | |
407 | } | |
408 | ||
abcd08a6 CL |
409 | static inline int check_valid_pointer(struct kmem_cache *s, |
410 | struct page *page, const void *object) | |
81819f0f CL |
411 | { |
412 | void *base; | |
413 | ||
414 | if (!object) | |
415 | return 1; | |
416 | ||
417 | base = page_address(page); | |
418 | if (object < base || object >= base + s->objects * s->size || | |
419 | (object - base) % s->size) { | |
420 | return 0; | |
421 | } | |
422 | ||
423 | return 1; | |
424 | } | |
425 | ||
426 | /* | |
427 | * Object layout: | |
428 | * | |
429 | * object address | |
430 | * Bytes of the object to be managed. | |
431 | * If the freepointer may overlay the object then the free | |
432 | * pointer is the first word of the object. | |
672bba3a | 433 | * |
81819f0f CL |
434 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
435 | * 0xa5 (POISON_END) | |
436 | * | |
437 | * object + s->objsize | |
438 | * Padding to reach word boundary. This is also used for Redzoning. | |
672bba3a CL |
439 | * Padding is extended by another word if Redzoning is enabled and |
440 | * objsize == inuse. | |
441 | * | |
81819f0f CL |
442 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
443 | * 0xcc (RED_ACTIVE) for objects in use. | |
444 | * | |
445 | * object + s->inuse | |
672bba3a CL |
446 | * Meta data starts here. |
447 | * | |
81819f0f CL |
448 | * A. Free pointer (if we cannot overwrite object on free) |
449 | * B. Tracking data for SLAB_STORE_USER | |
672bba3a CL |
450 | * C. Padding to reach required alignment boundary or at mininum |
451 | * one word if debuggin is on to be able to detect writes | |
452 | * before the word boundary. | |
453 | * | |
454 | * Padding is done using 0x5a (POISON_INUSE) | |
81819f0f CL |
455 | * |
456 | * object + s->size | |
672bba3a | 457 | * Nothing is used beyond s->size. |
81819f0f | 458 | * |
672bba3a CL |
459 | * If slabcaches are merged then the objsize and inuse boundaries are mostly |
460 | * ignored. And therefore no slab options that rely on these boundaries | |
81819f0f CL |
461 | * may be used with merged slabcaches. |
462 | */ | |
463 | ||
464 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
465 | void *from, void *to) | |
466 | { | |
70d71228 | 467 | printk(KERN_ERR "@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n", |
81819f0f CL |
468 | s->name, message, data, from, to - 1); |
469 | memset(from, data, to - from); | |
470 | } | |
471 | ||
472 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) | |
473 | { | |
474 | unsigned long off = s->inuse; /* The end of info */ | |
475 | ||
476 | if (s->offset) | |
477 | /* Freepointer is placed after the object. */ | |
478 | off += sizeof(void *); | |
479 | ||
480 | if (s->flags & SLAB_STORE_USER) | |
481 | /* We also have user information there */ | |
482 | off += 2 * sizeof(struct track); | |
483 | ||
484 | if (s->size == off) | |
485 | return 1; | |
486 | ||
487 | if (check_bytes(p + off, POISON_INUSE, s->size - off)) | |
488 | return 1; | |
489 | ||
490 | object_err(s, page, p, "Object padding check fails"); | |
491 | ||
492 | /* | |
493 | * Restore padding | |
494 | */ | |
495 | restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size); | |
496 | return 0; | |
497 | } | |
498 | ||
499 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
500 | { | |
501 | u8 *p; | |
502 | int length, remainder; | |
503 | ||
504 | if (!(s->flags & SLAB_POISON)) | |
505 | return 1; | |
506 | ||
507 | p = page_address(page); | |
508 | length = s->objects * s->size; | |
509 | remainder = (PAGE_SIZE << s->order) - length; | |
510 | if (!remainder) | |
511 | return 1; | |
512 | ||
513 | if (!check_bytes(p + length, POISON_INUSE, remainder)) { | |
70d71228 | 514 | slab_err(s, page, "Padding check failed"); |
81819f0f CL |
515 | restore_bytes(s, "slab padding", POISON_INUSE, p + length, |
516 | p + length + remainder); | |
517 | return 0; | |
518 | } | |
519 | return 1; | |
520 | } | |
521 | ||
522 | static int check_object(struct kmem_cache *s, struct page *page, | |
523 | void *object, int active) | |
524 | { | |
525 | u8 *p = object; | |
526 | u8 *endobject = object + s->objsize; | |
527 | ||
528 | if (s->flags & SLAB_RED_ZONE) { | |
529 | unsigned int red = | |
530 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; | |
531 | ||
532 | if (!check_bytes(endobject, red, s->inuse - s->objsize)) { | |
533 | object_err(s, page, object, | |
534 | active ? "Redzone Active" : "Redzone Inactive"); | |
535 | restore_bytes(s, "redzone", red, | |
536 | endobject, object + s->inuse); | |
537 | return 0; | |
538 | } | |
539 | } else { | |
540 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse && | |
541 | !check_bytes(endobject, POISON_INUSE, | |
542 | s->inuse - s->objsize)) { | |
543 | object_err(s, page, p, "Alignment padding check fails"); | |
544 | /* | |
545 | * Fix it so that there will not be another report. | |
546 | * | |
547 | * Hmmm... We may be corrupting an object that now expects | |
548 | * to be longer than allowed. | |
549 | */ | |
550 | restore_bytes(s, "alignment padding", POISON_INUSE, | |
551 | endobject, object + s->inuse); | |
552 | } | |
553 | } | |
554 | ||
555 | if (s->flags & SLAB_POISON) { | |
556 | if (!active && (s->flags & __OBJECT_POISON) && | |
557 | (!check_bytes(p, POISON_FREE, s->objsize - 1) || | |
558 | p[s->objsize - 1] != POISON_END)) { | |
559 | ||
560 | object_err(s, page, p, "Poison check failed"); | |
561 | restore_bytes(s, "Poison", POISON_FREE, | |
562 | p, p + s->objsize -1); | |
563 | restore_bytes(s, "Poison", POISON_END, | |
564 | p + s->objsize - 1, p + s->objsize); | |
565 | return 0; | |
566 | } | |
567 | /* | |
568 | * check_pad_bytes cleans up on its own. | |
569 | */ | |
570 | check_pad_bytes(s, page, p); | |
571 | } | |
572 | ||
573 | if (!s->offset && active) | |
574 | /* | |
575 | * Object and freepointer overlap. Cannot check | |
576 | * freepointer while object is allocated. | |
577 | */ | |
578 | return 1; | |
579 | ||
580 | /* Check free pointer validity */ | |
581 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
582 | object_err(s, page, p, "Freepointer corrupt"); | |
583 | /* | |
584 | * No choice but to zap it and thus loose the remainder | |
585 | * of the free objects in this slab. May cause | |
672bba3a | 586 | * another error because the object count is now wrong. |
81819f0f CL |
587 | */ |
588 | set_freepointer(s, p, NULL); | |
589 | return 0; | |
590 | } | |
591 | return 1; | |
592 | } | |
593 | ||
594 | static int check_slab(struct kmem_cache *s, struct page *page) | |
595 | { | |
596 | VM_BUG_ON(!irqs_disabled()); | |
597 | ||
598 | if (!PageSlab(page)) { | |
70d71228 CL |
599 | slab_err(s, page, "Not a valid slab page flags=%lx " |
600 | "mapping=0x%p count=%d", page->flags, page->mapping, | |
81819f0f CL |
601 | page_count(page)); |
602 | return 0; | |
603 | } | |
604 | if (page->offset * sizeof(void *) != s->offset) { | |
70d71228 CL |
605 | slab_err(s, page, "Corrupted offset %lu flags=0x%lx " |
606 | "mapping=0x%p count=%d", | |
81819f0f | 607 | (unsigned long)(page->offset * sizeof(void *)), |
81819f0f CL |
608 | page->flags, |
609 | page->mapping, | |
610 | page_count(page)); | |
81819f0f CL |
611 | return 0; |
612 | } | |
613 | if (page->inuse > s->objects) { | |
70d71228 CL |
614 | slab_err(s, page, "inuse %u > max %u @0x%p flags=%lx " |
615 | "mapping=0x%p count=%d", | |
616 | s->name, page->inuse, s->objects, page->flags, | |
81819f0f | 617 | page->mapping, page_count(page)); |
81819f0f CL |
618 | return 0; |
619 | } | |
620 | /* Slab_pad_check fixes things up after itself */ | |
621 | slab_pad_check(s, page); | |
622 | return 1; | |
623 | } | |
624 | ||
625 | /* | |
672bba3a CL |
626 | * Determine if a certain object on a page is on the freelist. Must hold the |
627 | * slab lock to guarantee that the chains are in a consistent state. | |
81819f0f CL |
628 | */ |
629 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
630 | { | |
631 | int nr = 0; | |
632 | void *fp = page->freelist; | |
633 | void *object = NULL; | |
634 | ||
635 | while (fp && nr <= s->objects) { | |
636 | if (fp == search) | |
637 | return 1; | |
638 | if (!check_valid_pointer(s, page, fp)) { | |
639 | if (object) { | |
640 | object_err(s, page, object, | |
641 | "Freechain corrupt"); | |
642 | set_freepointer(s, object, NULL); | |
643 | break; | |
644 | } else { | |
70d71228 CL |
645 | slab_err(s, page, "Freepointer 0x%p corrupt", |
646 | fp); | |
81819f0f CL |
647 | page->freelist = NULL; |
648 | page->inuse = s->objects; | |
70d71228 CL |
649 | printk(KERN_ERR "@@@ SLUB %s: Freelist " |
650 | "cleared. Slab 0x%p\n", | |
651 | s->name, page); | |
81819f0f CL |
652 | return 0; |
653 | } | |
654 | break; | |
655 | } | |
656 | object = fp; | |
657 | fp = get_freepointer(s, object); | |
658 | nr++; | |
659 | } | |
660 | ||
661 | if (page->inuse != s->objects - nr) { | |
70d71228 CL |
662 | slab_err(s, page, "Wrong object count. Counter is %d but " |
663 | "counted were %d", s, page, page->inuse, | |
664 | s->objects - nr); | |
81819f0f | 665 | page->inuse = s->objects - nr; |
70d71228 CL |
666 | printk(KERN_ERR "@@@ SLUB %s: Object count adjusted. " |
667 | "Slab @0x%p\n", s->name, page); | |
81819f0f CL |
668 | } |
669 | return search == NULL; | |
670 | } | |
671 | ||
643b1138 | 672 | /* |
672bba3a | 673 | * Tracking of fully allocated slabs for debugging purposes. |
643b1138 | 674 | */ |
e95eed57 | 675 | static void add_full(struct kmem_cache_node *n, struct page *page) |
643b1138 | 676 | { |
643b1138 CL |
677 | spin_lock(&n->list_lock); |
678 | list_add(&page->lru, &n->full); | |
679 | spin_unlock(&n->list_lock); | |
680 | } | |
681 | ||
682 | static void remove_full(struct kmem_cache *s, struct page *page) | |
683 | { | |
684 | struct kmem_cache_node *n; | |
685 | ||
686 | if (!(s->flags & SLAB_STORE_USER)) | |
687 | return; | |
688 | ||
689 | n = get_node(s, page_to_nid(page)); | |
690 | ||
691 | spin_lock(&n->list_lock); | |
692 | list_del(&page->lru); | |
693 | spin_unlock(&n->list_lock); | |
694 | } | |
695 | ||
81819f0f CL |
696 | static int alloc_object_checks(struct kmem_cache *s, struct page *page, |
697 | void *object) | |
698 | { | |
699 | if (!check_slab(s, page)) | |
700 | goto bad; | |
701 | ||
702 | if (object && !on_freelist(s, page, object)) { | |
70d71228 CL |
703 | slab_err(s, page, "Object 0x%p already allocated", object); |
704 | goto bad; | |
81819f0f CL |
705 | } |
706 | ||
707 | if (!check_valid_pointer(s, page, object)) { | |
708 | object_err(s, page, object, "Freelist Pointer check fails"); | |
70d71228 | 709 | goto bad; |
81819f0f CL |
710 | } |
711 | ||
712 | if (!object) | |
713 | return 1; | |
714 | ||
715 | if (!check_object(s, page, object, 0)) | |
716 | goto bad; | |
81819f0f | 717 | |
81819f0f | 718 | return 1; |
81819f0f CL |
719 | bad: |
720 | if (PageSlab(page)) { | |
721 | /* | |
722 | * If this is a slab page then lets do the best we can | |
723 | * to avoid issues in the future. Marking all objects | |
672bba3a | 724 | * as used avoids touching the remaining objects. |
81819f0f CL |
725 | */ |
726 | printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n", | |
727 | s->name, page); | |
728 | page->inuse = s->objects; | |
729 | page->freelist = NULL; | |
730 | /* Fix up fields that may be corrupted */ | |
731 | page->offset = s->offset / sizeof(void *); | |
732 | } | |
733 | return 0; | |
734 | } | |
735 | ||
736 | static int free_object_checks(struct kmem_cache *s, struct page *page, | |
737 | void *object) | |
738 | { | |
739 | if (!check_slab(s, page)) | |
740 | goto fail; | |
741 | ||
742 | if (!check_valid_pointer(s, page, object)) { | |
70d71228 | 743 | slab_err(s, page, "Invalid object pointer 0x%p", object); |
81819f0f CL |
744 | goto fail; |
745 | } | |
746 | ||
747 | if (on_freelist(s, page, object)) { | |
70d71228 | 748 | slab_err(s, page, "Object 0x%p already free", object); |
81819f0f CL |
749 | goto fail; |
750 | } | |
751 | ||
752 | if (!check_object(s, page, object, 1)) | |
753 | return 0; | |
754 | ||
755 | if (unlikely(s != page->slab)) { | |
756 | if (!PageSlab(page)) | |
70d71228 CL |
757 | slab_err(s, page, "Attempt to free object(0x%p) " |
758 | "outside of slab", object); | |
81819f0f | 759 | else |
70d71228 | 760 | if (!page->slab) { |
81819f0f | 761 | printk(KERN_ERR |
70d71228 | 762 | "SLUB <none>: no slab for object 0x%p.\n", |
81819f0f | 763 | object); |
70d71228 CL |
764 | dump_stack(); |
765 | } | |
81819f0f | 766 | else |
70d71228 CL |
767 | slab_err(s, page, "object at 0x%p belongs " |
768 | "to slab %s", object, page->slab->name); | |
81819f0f CL |
769 | goto fail; |
770 | } | |
81819f0f CL |
771 | return 1; |
772 | fail: | |
81819f0f CL |
773 | printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n", |
774 | s->name, page, object); | |
775 | return 0; | |
776 | } | |
777 | ||
778 | /* | |
779 | * Slab allocation and freeing | |
780 | */ | |
781 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
782 | { | |
783 | struct page * page; | |
784 | int pages = 1 << s->order; | |
785 | ||
786 | if (s->order) | |
787 | flags |= __GFP_COMP; | |
788 | ||
789 | if (s->flags & SLAB_CACHE_DMA) | |
790 | flags |= SLUB_DMA; | |
791 | ||
792 | if (node == -1) | |
793 | page = alloc_pages(flags, s->order); | |
794 | else | |
795 | page = alloc_pages_node(node, flags, s->order); | |
796 | ||
797 | if (!page) | |
798 | return NULL; | |
799 | ||
800 | mod_zone_page_state(page_zone(page), | |
801 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
802 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
803 | pages); | |
804 | ||
805 | return page; | |
806 | } | |
807 | ||
808 | static void setup_object(struct kmem_cache *s, struct page *page, | |
809 | void *object) | |
810 | { | |
811 | if (PageError(page)) { | |
812 | init_object(s, object, 0); | |
813 | init_tracking(s, object); | |
814 | } | |
815 | ||
4f104934 CL |
816 | if (unlikely(s->ctor)) |
817 | s->ctor(object, s, SLAB_CTOR_CONSTRUCTOR); | |
81819f0f CL |
818 | } |
819 | ||
820 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
821 | { | |
822 | struct page *page; | |
823 | struct kmem_cache_node *n; | |
824 | void *start; | |
825 | void *end; | |
826 | void *last; | |
827 | void *p; | |
828 | ||
81819f0f CL |
829 | BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK)); |
830 | ||
831 | if (flags & __GFP_WAIT) | |
832 | local_irq_enable(); | |
833 | ||
834 | page = allocate_slab(s, flags & GFP_LEVEL_MASK, node); | |
835 | if (!page) | |
836 | goto out; | |
837 | ||
838 | n = get_node(s, page_to_nid(page)); | |
839 | if (n) | |
840 | atomic_long_inc(&n->nr_slabs); | |
841 | page->offset = s->offset / sizeof(void *); | |
842 | page->slab = s; | |
843 | page->flags |= 1 << PG_slab; | |
844 | if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | | |
845 | SLAB_STORE_USER | SLAB_TRACE)) | |
846 | page->flags |= 1 << PG_error; | |
847 | ||
848 | start = page_address(page); | |
849 | end = start + s->objects * s->size; | |
850 | ||
851 | if (unlikely(s->flags & SLAB_POISON)) | |
852 | memset(start, POISON_INUSE, PAGE_SIZE << s->order); | |
853 | ||
854 | last = start; | |
855 | for (p = start + s->size; p < end; p += s->size) { | |
856 | setup_object(s, page, last); | |
857 | set_freepointer(s, last, p); | |
858 | last = p; | |
859 | } | |
860 | setup_object(s, page, last); | |
861 | set_freepointer(s, last, NULL); | |
862 | ||
863 | page->freelist = start; | |
864 | page->inuse = 0; | |
865 | out: | |
866 | if (flags & __GFP_WAIT) | |
867 | local_irq_disable(); | |
868 | return page; | |
869 | } | |
870 | ||
871 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
872 | { | |
873 | int pages = 1 << s->order; | |
874 | ||
875 | if (unlikely(PageError(page) || s->dtor)) { | |
876 | void *start = page_address(page); | |
877 | void *end = start + (pages << PAGE_SHIFT); | |
878 | void *p; | |
879 | ||
880 | slab_pad_check(s, page); | |
881 | for (p = start; p <= end - s->size; p += s->size) { | |
882 | if (s->dtor) | |
883 | s->dtor(p, s, 0); | |
884 | check_object(s, page, p, 0); | |
885 | } | |
886 | } | |
887 | ||
888 | mod_zone_page_state(page_zone(page), | |
889 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
890 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
891 | - pages); | |
892 | ||
893 | page->mapping = NULL; | |
894 | __free_pages(page, s->order); | |
895 | } | |
896 | ||
897 | static void rcu_free_slab(struct rcu_head *h) | |
898 | { | |
899 | struct page *page; | |
900 | ||
901 | page = container_of((struct list_head *)h, struct page, lru); | |
902 | __free_slab(page->slab, page); | |
903 | } | |
904 | ||
905 | static void free_slab(struct kmem_cache *s, struct page *page) | |
906 | { | |
907 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
908 | /* | |
909 | * RCU free overloads the RCU head over the LRU | |
910 | */ | |
911 | struct rcu_head *head = (void *)&page->lru; | |
912 | ||
913 | call_rcu(head, rcu_free_slab); | |
914 | } else | |
915 | __free_slab(s, page); | |
916 | } | |
917 | ||
918 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
919 | { | |
920 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
921 | ||
922 | atomic_long_dec(&n->nr_slabs); | |
923 | reset_page_mapcount(page); | |
924 | page->flags &= ~(1 << PG_slab | 1 << PG_error); | |
925 | free_slab(s, page); | |
926 | } | |
927 | ||
928 | /* | |
929 | * Per slab locking using the pagelock | |
930 | */ | |
931 | static __always_inline void slab_lock(struct page *page) | |
932 | { | |
933 | bit_spin_lock(PG_locked, &page->flags); | |
934 | } | |
935 | ||
936 | static __always_inline void slab_unlock(struct page *page) | |
937 | { | |
938 | bit_spin_unlock(PG_locked, &page->flags); | |
939 | } | |
940 | ||
941 | static __always_inline int slab_trylock(struct page *page) | |
942 | { | |
943 | int rc = 1; | |
944 | ||
945 | rc = bit_spin_trylock(PG_locked, &page->flags); | |
946 | return rc; | |
947 | } | |
948 | ||
949 | /* | |
950 | * Management of partially allocated slabs | |
951 | */ | |
e95eed57 | 952 | static void add_partial_tail(struct kmem_cache_node *n, struct page *page) |
81819f0f | 953 | { |
e95eed57 CL |
954 | spin_lock(&n->list_lock); |
955 | n->nr_partial++; | |
956 | list_add_tail(&page->lru, &n->partial); | |
957 | spin_unlock(&n->list_lock); | |
958 | } | |
81819f0f | 959 | |
e95eed57 CL |
960 | static void add_partial(struct kmem_cache_node *n, struct page *page) |
961 | { | |
81819f0f CL |
962 | spin_lock(&n->list_lock); |
963 | n->nr_partial++; | |
964 | list_add(&page->lru, &n->partial); | |
965 | spin_unlock(&n->list_lock); | |
966 | } | |
967 | ||
968 | static void remove_partial(struct kmem_cache *s, | |
969 | struct page *page) | |
970 | { | |
971 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
972 | ||
973 | spin_lock(&n->list_lock); | |
974 | list_del(&page->lru); | |
975 | n->nr_partial--; | |
976 | spin_unlock(&n->list_lock); | |
977 | } | |
978 | ||
979 | /* | |
672bba3a | 980 | * Lock slab and remove from the partial list. |
81819f0f | 981 | * |
672bba3a | 982 | * Must hold list_lock. |
81819f0f CL |
983 | */ |
984 | static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page) | |
985 | { | |
986 | if (slab_trylock(page)) { | |
987 | list_del(&page->lru); | |
988 | n->nr_partial--; | |
989 | return 1; | |
990 | } | |
991 | return 0; | |
992 | } | |
993 | ||
994 | /* | |
672bba3a | 995 | * Try to allocate a partial slab from a specific node. |
81819f0f CL |
996 | */ |
997 | static struct page *get_partial_node(struct kmem_cache_node *n) | |
998 | { | |
999 | struct page *page; | |
1000 | ||
1001 | /* | |
1002 | * Racy check. If we mistakenly see no partial slabs then we | |
1003 | * just allocate an empty slab. If we mistakenly try to get a | |
672bba3a CL |
1004 | * partial slab and there is none available then get_partials() |
1005 | * will return NULL. | |
81819f0f CL |
1006 | */ |
1007 | if (!n || !n->nr_partial) | |
1008 | return NULL; | |
1009 | ||
1010 | spin_lock(&n->list_lock); | |
1011 | list_for_each_entry(page, &n->partial, lru) | |
1012 | if (lock_and_del_slab(n, page)) | |
1013 | goto out; | |
1014 | page = NULL; | |
1015 | out: | |
1016 | spin_unlock(&n->list_lock); | |
1017 | return page; | |
1018 | } | |
1019 | ||
1020 | /* | |
672bba3a | 1021 | * Get a page from somewhere. Search in increasing NUMA distances. |
81819f0f CL |
1022 | */ |
1023 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) | |
1024 | { | |
1025 | #ifdef CONFIG_NUMA | |
1026 | struct zonelist *zonelist; | |
1027 | struct zone **z; | |
1028 | struct page *page; | |
1029 | ||
1030 | /* | |
672bba3a CL |
1031 | * The defrag ratio allows a configuration of the tradeoffs between |
1032 | * inter node defragmentation and node local allocations. A lower | |
1033 | * defrag_ratio increases the tendency to do local allocations | |
1034 | * instead of attempting to obtain partial slabs from other nodes. | |
81819f0f | 1035 | * |
672bba3a CL |
1036 | * If the defrag_ratio is set to 0 then kmalloc() always |
1037 | * returns node local objects. If the ratio is higher then kmalloc() | |
1038 | * may return off node objects because partial slabs are obtained | |
1039 | * from other nodes and filled up. | |
81819f0f CL |
1040 | * |
1041 | * If /sys/slab/xx/defrag_ratio is set to 100 (which makes | |
672bba3a CL |
1042 | * defrag_ratio = 1000) then every (well almost) allocation will |
1043 | * first attempt to defrag slab caches on other nodes. This means | |
1044 | * scanning over all nodes to look for partial slabs which may be | |
1045 | * expensive if we do it every time we are trying to find a slab | |
1046 | * with available objects. | |
81819f0f CL |
1047 | */ |
1048 | if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio) | |
1049 | return NULL; | |
1050 | ||
1051 | zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
1052 | ->node_zonelists[gfp_zone(flags)]; | |
1053 | for (z = zonelist->zones; *z; z++) { | |
1054 | struct kmem_cache_node *n; | |
1055 | ||
1056 | n = get_node(s, zone_to_nid(*z)); | |
1057 | ||
1058 | if (n && cpuset_zone_allowed_hardwall(*z, flags) && | |
e95eed57 | 1059 | n->nr_partial > MIN_PARTIAL) { |
81819f0f CL |
1060 | page = get_partial_node(n); |
1061 | if (page) | |
1062 | return page; | |
1063 | } | |
1064 | } | |
1065 | #endif | |
1066 | return NULL; | |
1067 | } | |
1068 | ||
1069 | /* | |
1070 | * Get a partial page, lock it and return it. | |
1071 | */ | |
1072 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) | |
1073 | { | |
1074 | struct page *page; | |
1075 | int searchnode = (node == -1) ? numa_node_id() : node; | |
1076 | ||
1077 | page = get_partial_node(get_node(s, searchnode)); | |
1078 | if (page || (flags & __GFP_THISNODE)) | |
1079 | return page; | |
1080 | ||
1081 | return get_any_partial(s, flags); | |
1082 | } | |
1083 | ||
1084 | /* | |
1085 | * Move a page back to the lists. | |
1086 | * | |
1087 | * Must be called with the slab lock held. | |
1088 | * | |
1089 | * On exit the slab lock will have been dropped. | |
1090 | */ | |
1091 | static void putback_slab(struct kmem_cache *s, struct page *page) | |
1092 | { | |
e95eed57 CL |
1093 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1094 | ||
81819f0f | 1095 | if (page->inuse) { |
e95eed57 | 1096 | |
81819f0f | 1097 | if (page->freelist) |
e95eed57 CL |
1098 | add_partial(n, page); |
1099 | else if (PageError(page) && (s->flags & SLAB_STORE_USER)) | |
1100 | add_full(n, page); | |
81819f0f | 1101 | slab_unlock(page); |
e95eed57 | 1102 | |
81819f0f | 1103 | } else { |
e95eed57 CL |
1104 | if (n->nr_partial < MIN_PARTIAL) { |
1105 | /* | |
672bba3a CL |
1106 | * Adding an empty slab to the partial slabs in order |
1107 | * to avoid page allocator overhead. This slab needs | |
1108 | * to come after the other slabs with objects in | |
1109 | * order to fill them up. That way the size of the | |
1110 | * partial list stays small. kmem_cache_shrink can | |
1111 | * reclaim empty slabs from the partial list. | |
e95eed57 CL |
1112 | */ |
1113 | add_partial_tail(n, page); | |
1114 | slab_unlock(page); | |
1115 | } else { | |
1116 | slab_unlock(page); | |
1117 | discard_slab(s, page); | |
1118 | } | |
81819f0f CL |
1119 | } |
1120 | } | |
1121 | ||
1122 | /* | |
1123 | * Remove the cpu slab | |
1124 | */ | |
1125 | static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu) | |
1126 | { | |
1127 | s->cpu_slab[cpu] = NULL; | |
1128 | ClearPageActive(page); | |
1129 | ||
1130 | putback_slab(s, page); | |
1131 | } | |
1132 | ||
1133 | static void flush_slab(struct kmem_cache *s, struct page *page, int cpu) | |
1134 | { | |
1135 | slab_lock(page); | |
1136 | deactivate_slab(s, page, cpu); | |
1137 | } | |
1138 | ||
1139 | /* | |
1140 | * Flush cpu slab. | |
1141 | * Called from IPI handler with interrupts disabled. | |
1142 | */ | |
1143 | static void __flush_cpu_slab(struct kmem_cache *s, int cpu) | |
1144 | { | |
1145 | struct page *page = s->cpu_slab[cpu]; | |
1146 | ||
1147 | if (likely(page)) | |
1148 | flush_slab(s, page, cpu); | |
1149 | } | |
1150 | ||
1151 | static void flush_cpu_slab(void *d) | |
1152 | { | |
1153 | struct kmem_cache *s = d; | |
1154 | int cpu = smp_processor_id(); | |
1155 | ||
1156 | __flush_cpu_slab(s, cpu); | |
1157 | } | |
1158 | ||
1159 | static void flush_all(struct kmem_cache *s) | |
1160 | { | |
1161 | #ifdef CONFIG_SMP | |
1162 | on_each_cpu(flush_cpu_slab, s, 1, 1); | |
1163 | #else | |
1164 | unsigned long flags; | |
1165 | ||
1166 | local_irq_save(flags); | |
1167 | flush_cpu_slab(s); | |
1168 | local_irq_restore(flags); | |
1169 | #endif | |
1170 | } | |
1171 | ||
1172 | /* | |
1173 | * slab_alloc is optimized to only modify two cachelines on the fast path | |
1174 | * (aside from the stack): | |
1175 | * | |
1176 | * 1. The page struct | |
1177 | * 2. The first cacheline of the object to be allocated. | |
1178 | * | |
672bba3a | 1179 | * The only other cache lines that are read (apart from code) is the |
81819f0f CL |
1180 | * per cpu array in the kmem_cache struct. |
1181 | * | |
1182 | * Fastpath is not possible if we need to get a new slab or have | |
1183 | * debugging enabled (which means all slabs are marked with PageError) | |
1184 | */ | |
77c5e2d0 CL |
1185 | static void *slab_alloc(struct kmem_cache *s, |
1186 | gfp_t gfpflags, int node, void *addr) | |
81819f0f CL |
1187 | { |
1188 | struct page *page; | |
1189 | void **object; | |
1190 | unsigned long flags; | |
1191 | int cpu; | |
1192 | ||
1193 | local_irq_save(flags); | |
1194 | cpu = smp_processor_id(); | |
1195 | page = s->cpu_slab[cpu]; | |
1196 | if (!page) | |
1197 | goto new_slab; | |
1198 | ||
1199 | slab_lock(page); | |
1200 | if (unlikely(node != -1 && page_to_nid(page) != node)) | |
1201 | goto another_slab; | |
1202 | redo: | |
1203 | object = page->freelist; | |
1204 | if (unlikely(!object)) | |
1205 | goto another_slab; | |
1206 | if (unlikely(PageError(page))) | |
1207 | goto debug; | |
1208 | ||
1209 | have_object: | |
1210 | page->inuse++; | |
1211 | page->freelist = object[page->offset]; | |
1212 | slab_unlock(page); | |
1213 | local_irq_restore(flags); | |
1214 | return object; | |
1215 | ||
1216 | another_slab: | |
1217 | deactivate_slab(s, page, cpu); | |
1218 | ||
1219 | new_slab: | |
1220 | page = get_partial(s, gfpflags, node); | |
1221 | if (likely(page)) { | |
1222 | have_slab: | |
1223 | s->cpu_slab[cpu] = page; | |
1224 | SetPageActive(page); | |
1225 | goto redo; | |
1226 | } | |
1227 | ||
1228 | page = new_slab(s, gfpflags, node); | |
1229 | if (page) { | |
1230 | cpu = smp_processor_id(); | |
1231 | if (s->cpu_slab[cpu]) { | |
1232 | /* | |
672bba3a CL |
1233 | * Someone else populated the cpu_slab while we |
1234 | * enabled interrupts, or we have gotten scheduled | |
1235 | * on another cpu. The page may not be on the | |
1236 | * requested node even if __GFP_THISNODE was | |
1237 | * specified. So we need to recheck. | |
81819f0f CL |
1238 | */ |
1239 | if (node == -1 || | |
1240 | page_to_nid(s->cpu_slab[cpu]) == node) { | |
1241 | /* | |
1242 | * Current cpuslab is acceptable and we | |
1243 | * want the current one since its cache hot | |
1244 | */ | |
1245 | discard_slab(s, page); | |
1246 | page = s->cpu_slab[cpu]; | |
1247 | slab_lock(page); | |
1248 | goto redo; | |
1249 | } | |
672bba3a | 1250 | /* New slab does not fit our expectations */ |
81819f0f CL |
1251 | flush_slab(s, s->cpu_slab[cpu], cpu); |
1252 | } | |
1253 | slab_lock(page); | |
1254 | goto have_slab; | |
1255 | } | |
1256 | local_irq_restore(flags); | |
1257 | return NULL; | |
1258 | debug: | |
1259 | if (!alloc_object_checks(s, page, object)) | |
1260 | goto another_slab; | |
1261 | if (s->flags & SLAB_STORE_USER) | |
77c5e2d0 | 1262 | set_track(s, object, TRACK_ALLOC, addr); |
70d71228 CL |
1263 | if (s->flags & SLAB_TRACE) { |
1264 | printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n", | |
1265 | s->name, object, page->inuse, | |
1266 | page->freelist); | |
1267 | dump_stack(); | |
1268 | } | |
1269 | init_object(s, object, 1); | |
81819f0f CL |
1270 | goto have_object; |
1271 | } | |
1272 | ||
1273 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
1274 | { | |
77c5e2d0 | 1275 | return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); |
81819f0f CL |
1276 | } |
1277 | EXPORT_SYMBOL(kmem_cache_alloc); | |
1278 | ||
1279 | #ifdef CONFIG_NUMA | |
1280 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
1281 | { | |
77c5e2d0 | 1282 | return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); |
81819f0f CL |
1283 | } |
1284 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
1285 | #endif | |
1286 | ||
1287 | /* | |
1288 | * The fastpath only writes the cacheline of the page struct and the first | |
1289 | * cacheline of the object. | |
1290 | * | |
672bba3a CL |
1291 | * We read the cpu_slab cacheline to check if the slab is the per cpu |
1292 | * slab for this processor. | |
81819f0f | 1293 | */ |
77c5e2d0 CL |
1294 | static void slab_free(struct kmem_cache *s, struct page *page, |
1295 | void *x, void *addr) | |
81819f0f CL |
1296 | { |
1297 | void *prior; | |
1298 | void **object = (void *)x; | |
1299 | unsigned long flags; | |
1300 | ||
1301 | local_irq_save(flags); | |
1302 | slab_lock(page); | |
1303 | ||
1304 | if (unlikely(PageError(page))) | |
1305 | goto debug; | |
1306 | checks_ok: | |
1307 | prior = object[page->offset] = page->freelist; | |
1308 | page->freelist = object; | |
1309 | page->inuse--; | |
1310 | ||
1311 | if (unlikely(PageActive(page))) | |
1312 | /* | |
1313 | * Cpu slabs are never on partial lists and are | |
1314 | * never freed. | |
1315 | */ | |
1316 | goto out_unlock; | |
1317 | ||
1318 | if (unlikely(!page->inuse)) | |
1319 | goto slab_empty; | |
1320 | ||
1321 | /* | |
1322 | * Objects left in the slab. If it | |
1323 | * was not on the partial list before | |
1324 | * then add it. | |
1325 | */ | |
1326 | if (unlikely(!prior)) | |
e95eed57 | 1327 | add_partial(get_node(s, page_to_nid(page)), page); |
81819f0f CL |
1328 | |
1329 | out_unlock: | |
1330 | slab_unlock(page); | |
1331 | local_irq_restore(flags); | |
1332 | return; | |
1333 | ||
1334 | slab_empty: | |
1335 | if (prior) | |
1336 | /* | |
672bba3a | 1337 | * Slab still on the partial list. |
81819f0f CL |
1338 | */ |
1339 | remove_partial(s, page); | |
1340 | ||
1341 | slab_unlock(page); | |
1342 | discard_slab(s, page); | |
1343 | local_irq_restore(flags); | |
1344 | return; | |
1345 | ||
1346 | debug: | |
77c5e2d0 CL |
1347 | if (!free_object_checks(s, page, x)) |
1348 | goto out_unlock; | |
643b1138 CL |
1349 | if (!PageActive(page) && !page->freelist) |
1350 | remove_full(s, page); | |
77c5e2d0 CL |
1351 | if (s->flags & SLAB_STORE_USER) |
1352 | set_track(s, x, TRACK_FREE, addr); | |
70d71228 CL |
1353 | if (s->flags & SLAB_TRACE) { |
1354 | printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n", | |
1355 | s->name, object, page->inuse, | |
1356 | page->freelist); | |
1357 | print_section("Object", (void *)object, s->objsize); | |
1358 | dump_stack(); | |
1359 | } | |
1360 | init_object(s, object, 0); | |
77c5e2d0 | 1361 | goto checks_ok; |
81819f0f CL |
1362 | } |
1363 | ||
1364 | void kmem_cache_free(struct kmem_cache *s, void *x) | |
1365 | { | |
77c5e2d0 | 1366 | struct page *page; |
81819f0f | 1367 | |
b49af68f | 1368 | page = virt_to_head_page(x); |
81819f0f | 1369 | |
77c5e2d0 | 1370 | slab_free(s, page, x, __builtin_return_address(0)); |
81819f0f CL |
1371 | } |
1372 | EXPORT_SYMBOL(kmem_cache_free); | |
1373 | ||
1374 | /* Figure out on which slab object the object resides */ | |
1375 | static struct page *get_object_page(const void *x) | |
1376 | { | |
b49af68f | 1377 | struct page *page = virt_to_head_page(x); |
81819f0f CL |
1378 | |
1379 | if (!PageSlab(page)) | |
1380 | return NULL; | |
1381 | ||
1382 | return page; | |
1383 | } | |
1384 | ||
1385 | /* | |
672bba3a CL |
1386 | * Object placement in a slab is made very easy because we always start at |
1387 | * offset 0. If we tune the size of the object to the alignment then we can | |
1388 | * get the required alignment by putting one properly sized object after | |
1389 | * another. | |
81819f0f CL |
1390 | * |
1391 | * Notice that the allocation order determines the sizes of the per cpu | |
1392 | * caches. Each processor has always one slab available for allocations. | |
1393 | * Increasing the allocation order reduces the number of times that slabs | |
672bba3a | 1394 | * must be moved on and off the partial lists and is therefore a factor in |
81819f0f | 1395 | * locking overhead. |
81819f0f CL |
1396 | */ |
1397 | ||
1398 | /* | |
1399 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
1400 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
1401 | * and increases the number of allocations possible without having to | |
1402 | * take the list_lock. | |
1403 | */ | |
1404 | static int slub_min_order; | |
1405 | static int slub_max_order = DEFAULT_MAX_ORDER; | |
81819f0f CL |
1406 | static int slub_min_objects = DEFAULT_MIN_OBJECTS; |
1407 | ||
1408 | /* | |
1409 | * Merge control. If this is set then no merging of slab caches will occur. | |
672bba3a | 1410 | * (Could be removed. This was introduced to pacify the merge skeptics.) |
81819f0f CL |
1411 | */ |
1412 | static int slub_nomerge; | |
1413 | ||
1414 | /* | |
1415 | * Debug settings: | |
1416 | */ | |
1417 | static int slub_debug; | |
1418 | ||
1419 | static char *slub_debug_slabs; | |
1420 | ||
1421 | /* | |
1422 | * Calculate the order of allocation given an slab object size. | |
1423 | * | |
672bba3a CL |
1424 | * The order of allocation has significant impact on performance and other |
1425 | * system components. Generally order 0 allocations should be preferred since | |
1426 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
1427 | * be problematic to put into order 0 slabs because there may be too much | |
1428 | * unused space left. We go to a higher order if more than 1/8th of the slab | |
1429 | * would be wasted. | |
1430 | * | |
1431 | * In order to reach satisfactory performance we must ensure that a minimum | |
1432 | * number of objects is in one slab. Otherwise we may generate too much | |
1433 | * activity on the partial lists which requires taking the list_lock. This is | |
1434 | * less a concern for large slabs though which are rarely used. | |
81819f0f | 1435 | * |
672bba3a CL |
1436 | * slub_max_order specifies the order where we begin to stop considering the |
1437 | * number of objects in a slab as critical. If we reach slub_max_order then | |
1438 | * we try to keep the page order as low as possible. So we accept more waste | |
1439 | * of space in favor of a small page order. | |
81819f0f | 1440 | * |
672bba3a CL |
1441 | * Higher order allocations also allow the placement of more objects in a |
1442 | * slab and thereby reduce object handling overhead. If the user has | |
1443 | * requested a higher mininum order then we start with that one instead of | |
1444 | * the smallest order which will fit the object. | |
81819f0f CL |
1445 | */ |
1446 | static int calculate_order(int size) | |
1447 | { | |
1448 | int order; | |
1449 | int rem; | |
1450 | ||
1451 | for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT); | |
1452 | order < MAX_ORDER; order++) { | |
1453 | unsigned long slab_size = PAGE_SIZE << order; | |
1454 | ||
1455 | if (slub_max_order > order && | |
1456 | slab_size < slub_min_objects * size) | |
1457 | continue; | |
1458 | ||
1459 | if (slab_size < size) | |
1460 | continue; | |
1461 | ||
1462 | rem = slab_size % size; | |
1463 | ||
672bba3a | 1464 | if (rem <= slab_size / 8) |
81819f0f CL |
1465 | break; |
1466 | ||
1467 | } | |
1468 | if (order >= MAX_ORDER) | |
1469 | return -E2BIG; | |
672bba3a | 1470 | |
81819f0f CL |
1471 | return order; |
1472 | } | |
1473 | ||
1474 | /* | |
672bba3a | 1475 | * Figure out what the alignment of the objects will be. |
81819f0f CL |
1476 | */ |
1477 | static unsigned long calculate_alignment(unsigned long flags, | |
1478 | unsigned long align, unsigned long size) | |
1479 | { | |
1480 | /* | |
1481 | * If the user wants hardware cache aligned objects then | |
1482 | * follow that suggestion if the object is sufficiently | |
1483 | * large. | |
1484 | * | |
1485 | * The hardware cache alignment cannot override the | |
1486 | * specified alignment though. If that is greater | |
1487 | * then use it. | |
1488 | */ | |
5af60839 | 1489 | if ((flags & SLAB_HWCACHE_ALIGN) && |
65c02d4c CL |
1490 | size > cache_line_size() / 2) |
1491 | return max_t(unsigned long, align, cache_line_size()); | |
81819f0f CL |
1492 | |
1493 | if (align < ARCH_SLAB_MINALIGN) | |
1494 | return ARCH_SLAB_MINALIGN; | |
1495 | ||
1496 | return ALIGN(align, sizeof(void *)); | |
1497 | } | |
1498 | ||
1499 | static void init_kmem_cache_node(struct kmem_cache_node *n) | |
1500 | { | |
1501 | n->nr_partial = 0; | |
1502 | atomic_long_set(&n->nr_slabs, 0); | |
1503 | spin_lock_init(&n->list_lock); | |
1504 | INIT_LIST_HEAD(&n->partial); | |
643b1138 | 1505 | INIT_LIST_HEAD(&n->full); |
81819f0f CL |
1506 | } |
1507 | ||
1508 | #ifdef CONFIG_NUMA | |
1509 | /* | |
1510 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
1511 | * slab on the node for this slabcache. There are no concurrent accesses | |
1512 | * possible. | |
1513 | * | |
1514 | * Note that this function only works on the kmalloc_node_cache | |
1515 | * when allocating for the kmalloc_node_cache. | |
1516 | */ | |
1517 | static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags, | |
1518 | int node) | |
1519 | { | |
1520 | struct page *page; | |
1521 | struct kmem_cache_node *n; | |
1522 | ||
1523 | BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); | |
1524 | ||
1525 | page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node); | |
1526 | /* new_slab() disables interupts */ | |
1527 | local_irq_enable(); | |
1528 | ||
1529 | BUG_ON(!page); | |
1530 | n = page->freelist; | |
1531 | BUG_ON(!n); | |
1532 | page->freelist = get_freepointer(kmalloc_caches, n); | |
1533 | page->inuse++; | |
1534 | kmalloc_caches->node[node] = n; | |
1535 | init_object(kmalloc_caches, n, 1); | |
1536 | init_kmem_cache_node(n); | |
1537 | atomic_long_inc(&n->nr_slabs); | |
e95eed57 | 1538 | add_partial(n, page); |
81819f0f CL |
1539 | return n; |
1540 | } | |
1541 | ||
1542 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
1543 | { | |
1544 | int node; | |
1545 | ||
1546 | for_each_online_node(node) { | |
1547 | struct kmem_cache_node *n = s->node[node]; | |
1548 | if (n && n != &s->local_node) | |
1549 | kmem_cache_free(kmalloc_caches, n); | |
1550 | s->node[node] = NULL; | |
1551 | } | |
1552 | } | |
1553 | ||
1554 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
1555 | { | |
1556 | int node; | |
1557 | int local_node; | |
1558 | ||
1559 | if (slab_state >= UP) | |
1560 | local_node = page_to_nid(virt_to_page(s)); | |
1561 | else | |
1562 | local_node = 0; | |
1563 | ||
1564 | for_each_online_node(node) { | |
1565 | struct kmem_cache_node *n; | |
1566 | ||
1567 | if (local_node == node) | |
1568 | n = &s->local_node; | |
1569 | else { | |
1570 | if (slab_state == DOWN) { | |
1571 | n = early_kmem_cache_node_alloc(gfpflags, | |
1572 | node); | |
1573 | continue; | |
1574 | } | |
1575 | n = kmem_cache_alloc_node(kmalloc_caches, | |
1576 | gfpflags, node); | |
1577 | ||
1578 | if (!n) { | |
1579 | free_kmem_cache_nodes(s); | |
1580 | return 0; | |
1581 | } | |
1582 | ||
1583 | } | |
1584 | s->node[node] = n; | |
1585 | init_kmem_cache_node(n); | |
1586 | } | |
1587 | return 1; | |
1588 | } | |
1589 | #else | |
1590 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
1591 | { | |
1592 | } | |
1593 | ||
1594 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
1595 | { | |
1596 | init_kmem_cache_node(&s->local_node); | |
1597 | return 1; | |
1598 | } | |
1599 | #endif | |
1600 | ||
1601 | /* | |
1602 | * calculate_sizes() determines the order and the distribution of data within | |
1603 | * a slab object. | |
1604 | */ | |
1605 | static int calculate_sizes(struct kmem_cache *s) | |
1606 | { | |
1607 | unsigned long flags = s->flags; | |
1608 | unsigned long size = s->objsize; | |
1609 | unsigned long align = s->align; | |
1610 | ||
1611 | /* | |
1612 | * Determine if we can poison the object itself. If the user of | |
1613 | * the slab may touch the object after free or before allocation | |
1614 | * then we should never poison the object itself. | |
1615 | */ | |
1616 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
1617 | !s->ctor && !s->dtor) | |
1618 | s->flags |= __OBJECT_POISON; | |
1619 | else | |
1620 | s->flags &= ~__OBJECT_POISON; | |
1621 | ||
1622 | /* | |
1623 | * Round up object size to the next word boundary. We can only | |
1624 | * place the free pointer at word boundaries and this determines | |
1625 | * the possible location of the free pointer. | |
1626 | */ | |
1627 | size = ALIGN(size, sizeof(void *)); | |
1628 | ||
1629 | /* | |
672bba3a | 1630 | * If we are Redzoning then check if there is some space between the |
81819f0f | 1631 | * end of the object and the free pointer. If not then add an |
672bba3a | 1632 | * additional word to have some bytes to store Redzone information. |
81819f0f CL |
1633 | */ |
1634 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
1635 | size += sizeof(void *); | |
1636 | ||
1637 | /* | |
672bba3a CL |
1638 | * With that we have determined the number of bytes in actual use |
1639 | * by the object. This is the potential offset to the free pointer. | |
81819f0f CL |
1640 | */ |
1641 | s->inuse = size; | |
1642 | ||
1643 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
1644 | s->ctor || s->dtor)) { | |
1645 | /* | |
1646 | * Relocate free pointer after the object if it is not | |
1647 | * permitted to overwrite the first word of the object on | |
1648 | * kmem_cache_free. | |
1649 | * | |
1650 | * This is the case if we do RCU, have a constructor or | |
1651 | * destructor or are poisoning the objects. | |
1652 | */ | |
1653 | s->offset = size; | |
1654 | size += sizeof(void *); | |
1655 | } | |
1656 | ||
1657 | if (flags & SLAB_STORE_USER) | |
1658 | /* | |
1659 | * Need to store information about allocs and frees after | |
1660 | * the object. | |
1661 | */ | |
1662 | size += 2 * sizeof(struct track); | |
1663 | ||
be7b3fbc | 1664 | if (flags & SLAB_RED_ZONE) |
81819f0f CL |
1665 | /* |
1666 | * Add some empty padding so that we can catch | |
1667 | * overwrites from earlier objects rather than let | |
1668 | * tracking information or the free pointer be | |
1669 | * corrupted if an user writes before the start | |
1670 | * of the object. | |
1671 | */ | |
1672 | size += sizeof(void *); | |
672bba3a | 1673 | |
81819f0f CL |
1674 | /* |
1675 | * Determine the alignment based on various parameters that the | |
65c02d4c CL |
1676 | * user specified and the dynamic determination of cache line size |
1677 | * on bootup. | |
81819f0f CL |
1678 | */ |
1679 | align = calculate_alignment(flags, align, s->objsize); | |
1680 | ||
1681 | /* | |
1682 | * SLUB stores one object immediately after another beginning from | |
1683 | * offset 0. In order to align the objects we have to simply size | |
1684 | * each object to conform to the alignment. | |
1685 | */ | |
1686 | size = ALIGN(size, align); | |
1687 | s->size = size; | |
1688 | ||
1689 | s->order = calculate_order(size); | |
1690 | if (s->order < 0) | |
1691 | return 0; | |
1692 | ||
1693 | /* | |
1694 | * Determine the number of objects per slab | |
1695 | */ | |
1696 | s->objects = (PAGE_SIZE << s->order) / size; | |
1697 | ||
1698 | /* | |
1699 | * Verify that the number of objects is within permitted limits. | |
1700 | * The page->inuse field is only 16 bit wide! So we cannot have | |
1701 | * more than 64k objects per slab. | |
1702 | */ | |
1703 | if (!s->objects || s->objects > 65535) | |
1704 | return 0; | |
1705 | return 1; | |
1706 | ||
1707 | } | |
1708 | ||
81819f0f CL |
1709 | static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, |
1710 | const char *name, size_t size, | |
1711 | size_t align, unsigned long flags, | |
1712 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
1713 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
1714 | { | |
1715 | memset(s, 0, kmem_size); | |
1716 | s->name = name; | |
1717 | s->ctor = ctor; | |
1718 | s->dtor = dtor; | |
1719 | s->objsize = size; | |
1720 | s->flags = flags; | |
1721 | s->align = align; | |
1722 | ||
81819f0f CL |
1723 | /* |
1724 | * The page->offset field is only 16 bit wide. This is an offset | |
1725 | * in units of words from the beginning of an object. If the slab | |
1726 | * size is bigger then we cannot move the free pointer behind the | |
1727 | * object anymore. | |
1728 | * | |
1729 | * On 32 bit platforms the limit is 256k. On 64bit platforms | |
1730 | * the limit is 512k. | |
1731 | * | |
1732 | * Debugging or ctor/dtors may create a need to move the free | |
1733 | * pointer. Fail if this happens. | |
1734 | */ | |
1735 | if (s->size >= 65535 * sizeof(void *)) { | |
1736 | BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON | | |
1737 | SLAB_STORE_USER | SLAB_DESTROY_BY_RCU)); | |
1738 | BUG_ON(ctor || dtor); | |
1739 | } | |
1740 | else | |
1741 | /* | |
1742 | * Enable debugging if selected on the kernel commandline. | |
1743 | */ | |
1744 | if (slub_debug && (!slub_debug_slabs || | |
1745 | strncmp(slub_debug_slabs, name, | |
1746 | strlen(slub_debug_slabs)) == 0)) | |
1747 | s->flags |= slub_debug; | |
1748 | ||
1749 | if (!calculate_sizes(s)) | |
1750 | goto error; | |
1751 | ||
1752 | s->refcount = 1; | |
1753 | #ifdef CONFIG_NUMA | |
1754 | s->defrag_ratio = 100; | |
1755 | #endif | |
1756 | ||
1757 | if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) | |
1758 | return 1; | |
1759 | error: | |
1760 | if (flags & SLAB_PANIC) | |
1761 | panic("Cannot create slab %s size=%lu realsize=%u " | |
1762 | "order=%u offset=%u flags=%lx\n", | |
1763 | s->name, (unsigned long)size, s->size, s->order, | |
1764 | s->offset, flags); | |
1765 | return 0; | |
1766 | } | |
1767 | EXPORT_SYMBOL(kmem_cache_open); | |
1768 | ||
1769 | /* | |
1770 | * Check if a given pointer is valid | |
1771 | */ | |
1772 | int kmem_ptr_validate(struct kmem_cache *s, const void *object) | |
1773 | { | |
1774 | struct page * page; | |
81819f0f CL |
1775 | |
1776 | page = get_object_page(object); | |
1777 | ||
1778 | if (!page || s != page->slab) | |
1779 | /* No slab or wrong slab */ | |
1780 | return 0; | |
1781 | ||
abcd08a6 | 1782 | if (!check_valid_pointer(s, page, object)) |
81819f0f CL |
1783 | return 0; |
1784 | ||
1785 | /* | |
1786 | * We could also check if the object is on the slabs freelist. | |
1787 | * But this would be too expensive and it seems that the main | |
1788 | * purpose of kmem_ptr_valid is to check if the object belongs | |
1789 | * to a certain slab. | |
1790 | */ | |
1791 | return 1; | |
1792 | } | |
1793 | EXPORT_SYMBOL(kmem_ptr_validate); | |
1794 | ||
1795 | /* | |
1796 | * Determine the size of a slab object | |
1797 | */ | |
1798 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
1799 | { | |
1800 | return s->objsize; | |
1801 | } | |
1802 | EXPORT_SYMBOL(kmem_cache_size); | |
1803 | ||
1804 | const char *kmem_cache_name(struct kmem_cache *s) | |
1805 | { | |
1806 | return s->name; | |
1807 | } | |
1808 | EXPORT_SYMBOL(kmem_cache_name); | |
1809 | ||
1810 | /* | |
672bba3a CL |
1811 | * Attempt to free all slabs on a node. Return the number of slabs we |
1812 | * were unable to free. | |
81819f0f CL |
1813 | */ |
1814 | static int free_list(struct kmem_cache *s, struct kmem_cache_node *n, | |
1815 | struct list_head *list) | |
1816 | { | |
1817 | int slabs_inuse = 0; | |
1818 | unsigned long flags; | |
1819 | struct page *page, *h; | |
1820 | ||
1821 | spin_lock_irqsave(&n->list_lock, flags); | |
1822 | list_for_each_entry_safe(page, h, list, lru) | |
1823 | if (!page->inuse) { | |
1824 | list_del(&page->lru); | |
1825 | discard_slab(s, page); | |
1826 | } else | |
1827 | slabs_inuse++; | |
1828 | spin_unlock_irqrestore(&n->list_lock, flags); | |
1829 | return slabs_inuse; | |
1830 | } | |
1831 | ||
1832 | /* | |
672bba3a | 1833 | * Release all resources used by a slab cache. |
81819f0f CL |
1834 | */ |
1835 | static int kmem_cache_close(struct kmem_cache *s) | |
1836 | { | |
1837 | int node; | |
1838 | ||
1839 | flush_all(s); | |
1840 | ||
1841 | /* Attempt to free all objects */ | |
1842 | for_each_online_node(node) { | |
1843 | struct kmem_cache_node *n = get_node(s, node); | |
1844 | ||
2086d26a | 1845 | n->nr_partial -= free_list(s, n, &n->partial); |
81819f0f CL |
1846 | if (atomic_long_read(&n->nr_slabs)) |
1847 | return 1; | |
1848 | } | |
1849 | free_kmem_cache_nodes(s); | |
1850 | return 0; | |
1851 | } | |
1852 | ||
1853 | /* | |
1854 | * Close a cache and release the kmem_cache structure | |
1855 | * (must be used for caches created using kmem_cache_create) | |
1856 | */ | |
1857 | void kmem_cache_destroy(struct kmem_cache *s) | |
1858 | { | |
1859 | down_write(&slub_lock); | |
1860 | s->refcount--; | |
1861 | if (!s->refcount) { | |
1862 | list_del(&s->list); | |
1863 | if (kmem_cache_close(s)) | |
1864 | WARN_ON(1); | |
1865 | sysfs_slab_remove(s); | |
1866 | kfree(s); | |
1867 | } | |
1868 | up_write(&slub_lock); | |
1869 | } | |
1870 | EXPORT_SYMBOL(kmem_cache_destroy); | |
1871 | ||
1872 | /******************************************************************** | |
1873 | * Kmalloc subsystem | |
1874 | *******************************************************************/ | |
1875 | ||
1876 | struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned; | |
1877 | EXPORT_SYMBOL(kmalloc_caches); | |
1878 | ||
1879 | #ifdef CONFIG_ZONE_DMA | |
1880 | static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1]; | |
1881 | #endif | |
1882 | ||
1883 | static int __init setup_slub_min_order(char *str) | |
1884 | { | |
1885 | get_option (&str, &slub_min_order); | |
1886 | ||
1887 | return 1; | |
1888 | } | |
1889 | ||
1890 | __setup("slub_min_order=", setup_slub_min_order); | |
1891 | ||
1892 | static int __init setup_slub_max_order(char *str) | |
1893 | { | |
1894 | get_option (&str, &slub_max_order); | |
1895 | ||
1896 | return 1; | |
1897 | } | |
1898 | ||
1899 | __setup("slub_max_order=", setup_slub_max_order); | |
1900 | ||
1901 | static int __init setup_slub_min_objects(char *str) | |
1902 | { | |
1903 | get_option (&str, &slub_min_objects); | |
1904 | ||
1905 | return 1; | |
1906 | } | |
1907 | ||
1908 | __setup("slub_min_objects=", setup_slub_min_objects); | |
1909 | ||
1910 | static int __init setup_slub_nomerge(char *str) | |
1911 | { | |
1912 | slub_nomerge = 1; | |
1913 | return 1; | |
1914 | } | |
1915 | ||
1916 | __setup("slub_nomerge", setup_slub_nomerge); | |
1917 | ||
1918 | static int __init setup_slub_debug(char *str) | |
1919 | { | |
1920 | if (!str || *str != '=') | |
1921 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1922 | else { | |
1923 | str++; | |
1924 | if (*str == 0 || *str == ',') | |
1925 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1926 | else | |
1927 | for( ;*str && *str != ','; str++) | |
1928 | switch (*str) { | |
1929 | case 'f' : case 'F' : | |
1930 | slub_debug |= SLAB_DEBUG_FREE; | |
1931 | break; | |
1932 | case 'z' : case 'Z' : | |
1933 | slub_debug |= SLAB_RED_ZONE; | |
1934 | break; | |
1935 | case 'p' : case 'P' : | |
1936 | slub_debug |= SLAB_POISON; | |
1937 | break; | |
1938 | case 'u' : case 'U' : | |
1939 | slub_debug |= SLAB_STORE_USER; | |
1940 | break; | |
1941 | case 't' : case 'T' : | |
1942 | slub_debug |= SLAB_TRACE; | |
1943 | break; | |
1944 | default: | |
1945 | printk(KERN_ERR "slub_debug option '%c' " | |
1946 | "unknown. skipped\n",*str); | |
1947 | } | |
1948 | } | |
1949 | ||
1950 | if (*str == ',') | |
1951 | slub_debug_slabs = str + 1; | |
1952 | return 1; | |
1953 | } | |
1954 | ||
1955 | __setup("slub_debug", setup_slub_debug); | |
1956 | ||
1957 | static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, | |
1958 | const char *name, int size, gfp_t gfp_flags) | |
1959 | { | |
1960 | unsigned int flags = 0; | |
1961 | ||
1962 | if (gfp_flags & SLUB_DMA) | |
1963 | flags = SLAB_CACHE_DMA; | |
1964 | ||
1965 | down_write(&slub_lock); | |
1966 | if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, | |
1967 | flags, NULL, NULL)) | |
1968 | goto panic; | |
1969 | ||
1970 | list_add(&s->list, &slab_caches); | |
1971 | up_write(&slub_lock); | |
1972 | if (sysfs_slab_add(s)) | |
1973 | goto panic; | |
1974 | return s; | |
1975 | ||
1976 | panic: | |
1977 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
1978 | } | |
1979 | ||
1980 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) | |
1981 | { | |
1982 | int index = kmalloc_index(size); | |
1983 | ||
614410d5 | 1984 | if (!index) |
81819f0f CL |
1985 | return NULL; |
1986 | ||
1987 | /* Allocation too large? */ | |
1988 | BUG_ON(index < 0); | |
1989 | ||
1990 | #ifdef CONFIG_ZONE_DMA | |
1991 | if ((flags & SLUB_DMA)) { | |
1992 | struct kmem_cache *s; | |
1993 | struct kmem_cache *x; | |
1994 | char *text; | |
1995 | size_t realsize; | |
1996 | ||
1997 | s = kmalloc_caches_dma[index]; | |
1998 | if (s) | |
1999 | return s; | |
2000 | ||
2001 | /* Dynamically create dma cache */ | |
2002 | x = kmalloc(kmem_size, flags & ~SLUB_DMA); | |
2003 | if (!x) | |
2004 | panic("Unable to allocate memory for dma cache\n"); | |
2005 | ||
2006 | if (index <= KMALLOC_SHIFT_HIGH) | |
2007 | realsize = 1 << index; | |
2008 | else { | |
2009 | if (index == 1) | |
2010 | realsize = 96; | |
2011 | else | |
2012 | realsize = 192; | |
2013 | } | |
2014 | ||
2015 | text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", | |
2016 | (unsigned int)realsize); | |
2017 | s = create_kmalloc_cache(x, text, realsize, flags); | |
2018 | kmalloc_caches_dma[index] = s; | |
2019 | return s; | |
2020 | } | |
2021 | #endif | |
2022 | return &kmalloc_caches[index]; | |
2023 | } | |
2024 | ||
2025 | void *__kmalloc(size_t size, gfp_t flags) | |
2026 | { | |
2027 | struct kmem_cache *s = get_slab(size, flags); | |
2028 | ||
2029 | if (s) | |
77c5e2d0 | 2030 | return slab_alloc(s, flags, -1, __builtin_return_address(0)); |
81819f0f CL |
2031 | return NULL; |
2032 | } | |
2033 | EXPORT_SYMBOL(__kmalloc); | |
2034 | ||
2035 | #ifdef CONFIG_NUMA | |
2036 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
2037 | { | |
2038 | struct kmem_cache *s = get_slab(size, flags); | |
2039 | ||
2040 | if (s) | |
77c5e2d0 | 2041 | return slab_alloc(s, flags, node, __builtin_return_address(0)); |
81819f0f CL |
2042 | return NULL; |
2043 | } | |
2044 | EXPORT_SYMBOL(__kmalloc_node); | |
2045 | #endif | |
2046 | ||
2047 | size_t ksize(const void *object) | |
2048 | { | |
2049 | struct page *page = get_object_page(object); | |
2050 | struct kmem_cache *s; | |
2051 | ||
2052 | BUG_ON(!page); | |
2053 | s = page->slab; | |
2054 | BUG_ON(!s); | |
2055 | ||
2056 | /* | |
2057 | * Debugging requires use of the padding between object | |
2058 | * and whatever may come after it. | |
2059 | */ | |
2060 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
2061 | return s->objsize; | |
2062 | ||
2063 | /* | |
2064 | * If we have the need to store the freelist pointer | |
2065 | * back there or track user information then we can | |
2066 | * only use the space before that information. | |
2067 | */ | |
2068 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
2069 | return s->inuse; | |
2070 | ||
2071 | /* | |
2072 | * Else we can use all the padding etc for the allocation | |
2073 | */ | |
2074 | return s->size; | |
2075 | } | |
2076 | EXPORT_SYMBOL(ksize); | |
2077 | ||
2078 | void kfree(const void *x) | |
2079 | { | |
2080 | struct kmem_cache *s; | |
2081 | struct page *page; | |
2082 | ||
2083 | if (!x) | |
2084 | return; | |
2085 | ||
b49af68f | 2086 | page = virt_to_head_page(x); |
81819f0f CL |
2087 | s = page->slab; |
2088 | ||
77c5e2d0 | 2089 | slab_free(s, page, (void *)x, __builtin_return_address(0)); |
81819f0f CL |
2090 | } |
2091 | EXPORT_SYMBOL(kfree); | |
2092 | ||
2086d26a | 2093 | /* |
672bba3a CL |
2094 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts |
2095 | * the remaining slabs by the number of items in use. The slabs with the | |
2096 | * most items in use come first. New allocations will then fill those up | |
2097 | * and thus they can be removed from the partial lists. | |
2098 | * | |
2099 | * The slabs with the least items are placed last. This results in them | |
2100 | * being allocated from last increasing the chance that the last objects | |
2101 | * are freed in them. | |
2086d26a CL |
2102 | */ |
2103 | int kmem_cache_shrink(struct kmem_cache *s) | |
2104 | { | |
2105 | int node; | |
2106 | int i; | |
2107 | struct kmem_cache_node *n; | |
2108 | struct page *page; | |
2109 | struct page *t; | |
2110 | struct list_head *slabs_by_inuse = | |
2111 | kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL); | |
2112 | unsigned long flags; | |
2113 | ||
2114 | if (!slabs_by_inuse) | |
2115 | return -ENOMEM; | |
2116 | ||
2117 | flush_all(s); | |
2118 | for_each_online_node(node) { | |
2119 | n = get_node(s, node); | |
2120 | ||
2121 | if (!n->nr_partial) | |
2122 | continue; | |
2123 | ||
2124 | for (i = 0; i < s->objects; i++) | |
2125 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
2126 | ||
2127 | spin_lock_irqsave(&n->list_lock, flags); | |
2128 | ||
2129 | /* | |
672bba3a | 2130 | * Build lists indexed by the items in use in each slab. |
2086d26a | 2131 | * |
672bba3a CL |
2132 | * Note that concurrent frees may occur while we hold the |
2133 | * list_lock. page->inuse here is the upper limit. | |
2086d26a CL |
2134 | */ |
2135 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
2136 | if (!page->inuse && slab_trylock(page)) { | |
2137 | /* | |
2138 | * Must hold slab lock here because slab_free | |
2139 | * may have freed the last object and be | |
2140 | * waiting to release the slab. | |
2141 | */ | |
2142 | list_del(&page->lru); | |
2143 | n->nr_partial--; | |
2144 | slab_unlock(page); | |
2145 | discard_slab(s, page); | |
2146 | } else { | |
2147 | if (n->nr_partial > MAX_PARTIAL) | |
2148 | list_move(&page->lru, | |
2149 | slabs_by_inuse + page->inuse); | |
2150 | } | |
2151 | } | |
2152 | ||
2153 | if (n->nr_partial <= MAX_PARTIAL) | |
2154 | goto out; | |
2155 | ||
2156 | /* | |
672bba3a CL |
2157 | * Rebuild the partial list with the slabs filled up most |
2158 | * first and the least used slabs at the end. | |
2086d26a CL |
2159 | */ |
2160 | for (i = s->objects - 1; i >= 0; i--) | |
2161 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
2162 | ||
2163 | out: | |
2164 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2165 | } | |
2166 | ||
2167 | kfree(slabs_by_inuse); | |
2168 | return 0; | |
2169 | } | |
2170 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2171 | ||
81819f0f CL |
2172 | /** |
2173 | * krealloc - reallocate memory. The contents will remain unchanged. | |
2174 | * | |
2175 | * @p: object to reallocate memory for. | |
2176 | * @new_size: how many bytes of memory are required. | |
2177 | * @flags: the type of memory to allocate. | |
2178 | * | |
2179 | * The contents of the object pointed to are preserved up to the | |
2180 | * lesser of the new and old sizes. If @p is %NULL, krealloc() | |
2181 | * behaves exactly like kmalloc(). If @size is 0 and @p is not a | |
2182 | * %NULL pointer, the object pointed to is freed. | |
2183 | */ | |
2184 | void *krealloc(const void *p, size_t new_size, gfp_t flags) | |
2185 | { | |
81819f0f | 2186 | void *ret; |
1f99a283 | 2187 | size_t ks; |
81819f0f CL |
2188 | |
2189 | if (unlikely(!p)) | |
2190 | return kmalloc(new_size, flags); | |
2191 | ||
2192 | if (unlikely(!new_size)) { | |
2193 | kfree(p); | |
2194 | return NULL; | |
2195 | } | |
2196 | ||
1f99a283 CL |
2197 | ks = ksize(p); |
2198 | if (ks >= new_size) | |
81819f0f CL |
2199 | return (void *)p; |
2200 | ||
2201 | ret = kmalloc(new_size, flags); | |
2202 | if (ret) { | |
1f99a283 | 2203 | memcpy(ret, p, min(new_size, ks)); |
81819f0f CL |
2204 | kfree(p); |
2205 | } | |
2206 | return ret; | |
2207 | } | |
2208 | EXPORT_SYMBOL(krealloc); | |
2209 | ||
2210 | /******************************************************************** | |
2211 | * Basic setup of slabs | |
2212 | *******************************************************************/ | |
2213 | ||
2214 | void __init kmem_cache_init(void) | |
2215 | { | |
2216 | int i; | |
2217 | ||
2218 | #ifdef CONFIG_NUMA | |
2219 | /* | |
2220 | * Must first have the slab cache available for the allocations of the | |
672bba3a | 2221 | * struct kmem_cache_node's. There is special bootstrap code in |
81819f0f CL |
2222 | * kmem_cache_open for slab_state == DOWN. |
2223 | */ | |
2224 | create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", | |
2225 | sizeof(struct kmem_cache_node), GFP_KERNEL); | |
2226 | #endif | |
2227 | ||
2228 | /* Able to allocate the per node structures */ | |
2229 | slab_state = PARTIAL; | |
2230 | ||
2231 | /* Caches that are not of the two-to-the-power-of size */ | |
2232 | create_kmalloc_cache(&kmalloc_caches[1], | |
2233 | "kmalloc-96", 96, GFP_KERNEL); | |
2234 | create_kmalloc_cache(&kmalloc_caches[2], | |
2235 | "kmalloc-192", 192, GFP_KERNEL); | |
2236 | ||
2237 | for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) | |
2238 | create_kmalloc_cache(&kmalloc_caches[i], | |
2239 | "kmalloc", 1 << i, GFP_KERNEL); | |
2240 | ||
2241 | slab_state = UP; | |
2242 | ||
2243 | /* Provide the correct kmalloc names now that the caches are up */ | |
2244 | for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) | |
2245 | kmalloc_caches[i]. name = | |
2246 | kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); | |
2247 | ||
2248 | #ifdef CONFIG_SMP | |
2249 | register_cpu_notifier(&slab_notifier); | |
2250 | #endif | |
2251 | ||
2252 | if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */ | |
2253 | kmem_size = offsetof(struct kmem_cache, cpu_slab) | |
2254 | + nr_cpu_ids * sizeof(struct page *); | |
2255 | ||
2256 | printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
2257 | " Processors=%d, Nodes=%d\n", | |
65c02d4c | 2258 | KMALLOC_SHIFT_HIGH, cache_line_size(), |
81819f0f CL |
2259 | slub_min_order, slub_max_order, slub_min_objects, |
2260 | nr_cpu_ids, nr_node_ids); | |
2261 | } | |
2262 | ||
2263 | /* | |
2264 | * Find a mergeable slab cache | |
2265 | */ | |
2266 | static int slab_unmergeable(struct kmem_cache *s) | |
2267 | { | |
2268 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
2269 | return 1; | |
2270 | ||
2271 | if (s->ctor || s->dtor) | |
2272 | return 1; | |
2273 | ||
2274 | return 0; | |
2275 | } | |
2276 | ||
2277 | static struct kmem_cache *find_mergeable(size_t size, | |
2278 | size_t align, unsigned long flags, | |
2279 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
2280 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
2281 | { | |
2282 | struct list_head *h; | |
2283 | ||
2284 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
2285 | return NULL; | |
2286 | ||
2287 | if (ctor || dtor) | |
2288 | return NULL; | |
2289 | ||
2290 | size = ALIGN(size, sizeof(void *)); | |
2291 | align = calculate_alignment(flags, align, size); | |
2292 | size = ALIGN(size, align); | |
2293 | ||
2294 | list_for_each(h, &slab_caches) { | |
2295 | struct kmem_cache *s = | |
2296 | container_of(h, struct kmem_cache, list); | |
2297 | ||
2298 | if (slab_unmergeable(s)) | |
2299 | continue; | |
2300 | ||
2301 | if (size > s->size) | |
2302 | continue; | |
2303 | ||
2304 | if (((flags | slub_debug) & SLUB_MERGE_SAME) != | |
2305 | (s->flags & SLUB_MERGE_SAME)) | |
2306 | continue; | |
2307 | /* | |
2308 | * Check if alignment is compatible. | |
2309 | * Courtesy of Adrian Drzewiecki | |
2310 | */ | |
2311 | if ((s->size & ~(align -1)) != s->size) | |
2312 | continue; | |
2313 | ||
2314 | if (s->size - size >= sizeof(void *)) | |
2315 | continue; | |
2316 | ||
2317 | return s; | |
2318 | } | |
2319 | return NULL; | |
2320 | } | |
2321 | ||
2322 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
2323 | size_t align, unsigned long flags, | |
2324 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
2325 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
2326 | { | |
2327 | struct kmem_cache *s; | |
2328 | ||
2329 | down_write(&slub_lock); | |
2330 | s = find_mergeable(size, align, flags, dtor, ctor); | |
2331 | if (s) { | |
2332 | s->refcount++; | |
2333 | /* | |
2334 | * Adjust the object sizes so that we clear | |
2335 | * the complete object on kzalloc. | |
2336 | */ | |
2337 | s->objsize = max(s->objsize, (int)size); | |
2338 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
2339 | if (sysfs_slab_alias(s, name)) | |
2340 | goto err; | |
2341 | } else { | |
2342 | s = kmalloc(kmem_size, GFP_KERNEL); | |
2343 | if (s && kmem_cache_open(s, GFP_KERNEL, name, | |
2344 | size, align, flags, ctor, dtor)) { | |
2345 | if (sysfs_slab_add(s)) { | |
2346 | kfree(s); | |
2347 | goto err; | |
2348 | } | |
2349 | list_add(&s->list, &slab_caches); | |
2350 | } else | |
2351 | kfree(s); | |
2352 | } | |
2353 | up_write(&slub_lock); | |
2354 | return s; | |
2355 | ||
2356 | err: | |
2357 | up_write(&slub_lock); | |
2358 | if (flags & SLAB_PANIC) | |
2359 | panic("Cannot create slabcache %s\n", name); | |
2360 | else | |
2361 | s = NULL; | |
2362 | return s; | |
2363 | } | |
2364 | EXPORT_SYMBOL(kmem_cache_create); | |
2365 | ||
2366 | void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags) | |
2367 | { | |
2368 | void *x; | |
2369 | ||
77c5e2d0 | 2370 | x = slab_alloc(s, flags, -1, __builtin_return_address(0)); |
81819f0f CL |
2371 | if (x) |
2372 | memset(x, 0, s->objsize); | |
2373 | return x; | |
2374 | } | |
2375 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
2376 | ||
2377 | #ifdef CONFIG_SMP | |
2378 | static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu) | |
2379 | { | |
2380 | struct list_head *h; | |
2381 | ||
2382 | down_read(&slub_lock); | |
2383 | list_for_each(h, &slab_caches) { | |
2384 | struct kmem_cache *s = | |
2385 | container_of(h, struct kmem_cache, list); | |
2386 | ||
2387 | func(s, cpu); | |
2388 | } | |
2389 | up_read(&slub_lock); | |
2390 | } | |
2391 | ||
2392 | /* | |
672bba3a CL |
2393 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
2394 | * necessary. | |
81819f0f CL |
2395 | */ |
2396 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
2397 | unsigned long action, void *hcpu) | |
2398 | { | |
2399 | long cpu = (long)hcpu; | |
2400 | ||
2401 | switch (action) { | |
2402 | case CPU_UP_CANCELED: | |
2403 | case CPU_DEAD: | |
2404 | for_all_slabs(__flush_cpu_slab, cpu); | |
2405 | break; | |
2406 | default: | |
2407 | break; | |
2408 | } | |
2409 | return NOTIFY_OK; | |
2410 | } | |
2411 | ||
2412 | static struct notifier_block __cpuinitdata slab_notifier = | |
2413 | { &slab_cpuup_callback, NULL, 0 }; | |
2414 | ||
2415 | #endif | |
2416 | ||
81819f0f CL |
2417 | #ifdef CONFIG_NUMA |
2418 | ||
2419 | /***************************************************************** | |
2420 | * Generic reaper used to support the page allocator | |
2421 | * (the cpu slabs are reaped by a per slab workqueue). | |
2422 | * | |
2423 | * Maybe move this to the page allocator? | |
2424 | ****************************************************************/ | |
2425 | ||
2426 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
2427 | ||
2428 | static void init_reap_node(int cpu) | |
2429 | { | |
2430 | int node; | |
2431 | ||
2432 | node = next_node(cpu_to_node(cpu), node_online_map); | |
2433 | if (node == MAX_NUMNODES) | |
2434 | node = first_node(node_online_map); | |
2435 | ||
2436 | __get_cpu_var(reap_node) = node; | |
2437 | } | |
2438 | ||
2439 | static void next_reap_node(void) | |
2440 | { | |
2441 | int node = __get_cpu_var(reap_node); | |
2442 | ||
2443 | /* | |
2444 | * Also drain per cpu pages on remote zones | |
2445 | */ | |
2446 | if (node != numa_node_id()) | |
2447 | drain_node_pages(node); | |
2448 | ||
2449 | node = next_node(node, node_online_map); | |
2450 | if (unlikely(node >= MAX_NUMNODES)) | |
2451 | node = first_node(node_online_map); | |
2452 | __get_cpu_var(reap_node) = node; | |
2453 | } | |
2454 | #else | |
2455 | #define init_reap_node(cpu) do { } while (0) | |
2456 | #define next_reap_node(void) do { } while (0) | |
2457 | #endif | |
2458 | ||
2459 | #define REAPTIMEOUT_CPUC (2*HZ) | |
2460 | ||
2461 | #ifdef CONFIG_SMP | |
2462 | static DEFINE_PER_CPU(struct delayed_work, reap_work); | |
2463 | ||
2464 | static void cache_reap(struct work_struct *unused) | |
2465 | { | |
2466 | next_reap_node(); | |
2467 | refresh_cpu_vm_stats(smp_processor_id()); | |
2468 | schedule_delayed_work(&__get_cpu_var(reap_work), | |
2469 | REAPTIMEOUT_CPUC); | |
2470 | } | |
2471 | ||
2472 | static void __devinit start_cpu_timer(int cpu) | |
2473 | { | |
2474 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); | |
2475 | ||
2476 | /* | |
2477 | * When this gets called from do_initcalls via cpucache_init(), | |
2478 | * init_workqueues() has already run, so keventd will be setup | |
2479 | * at that time. | |
2480 | */ | |
2481 | if (keventd_up() && reap_work->work.func == NULL) { | |
2482 | init_reap_node(cpu); | |
2483 | INIT_DELAYED_WORK(reap_work, cache_reap); | |
2484 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
2485 | } | |
2486 | } | |
2487 | ||
2488 | static int __init cpucache_init(void) | |
2489 | { | |
2490 | int cpu; | |
2491 | ||
2492 | /* | |
2493 | * Register the timers that drain pcp pages and update vm statistics | |
2494 | */ | |
2495 | for_each_online_cpu(cpu) | |
2496 | start_cpu_timer(cpu); | |
2497 | return 0; | |
2498 | } | |
2499 | __initcall(cpucache_init); | |
2500 | #endif | |
2501 | ||
2502 | #ifdef SLUB_RESILIENCY_TEST | |
2503 | static unsigned long validate_slab_cache(struct kmem_cache *s); | |
2504 | ||
2505 | static void resiliency_test(void) | |
2506 | { | |
2507 | u8 *p; | |
2508 | ||
2509 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
2510 | printk(KERN_ERR "-----------------------\n"); | |
2511 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
2512 | ||
2513 | p = kzalloc(16, GFP_KERNEL); | |
2514 | p[16] = 0x12; | |
2515 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
2516 | " 0x12->0x%p\n\n", p + 16); | |
2517 | ||
2518 | validate_slab_cache(kmalloc_caches + 4); | |
2519 | ||
2520 | /* Hmmm... The next two are dangerous */ | |
2521 | p = kzalloc(32, GFP_KERNEL); | |
2522 | p[32 + sizeof(void *)] = 0x34; | |
2523 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
2524 | " 0x34 -> -0x%p\n", p); | |
2525 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
2526 | ||
2527 | validate_slab_cache(kmalloc_caches + 5); | |
2528 | p = kzalloc(64, GFP_KERNEL); | |
2529 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
2530 | *p = 0x56; | |
2531 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
2532 | p); | |
2533 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
2534 | validate_slab_cache(kmalloc_caches + 6); | |
2535 | ||
2536 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
2537 | p = kzalloc(128, GFP_KERNEL); | |
2538 | kfree(p); | |
2539 | *p = 0x78; | |
2540 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
2541 | validate_slab_cache(kmalloc_caches + 7); | |
2542 | ||
2543 | p = kzalloc(256, GFP_KERNEL); | |
2544 | kfree(p); | |
2545 | p[50] = 0x9a; | |
2546 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); | |
2547 | validate_slab_cache(kmalloc_caches + 8); | |
2548 | ||
2549 | p = kzalloc(512, GFP_KERNEL); | |
2550 | kfree(p); | |
2551 | p[512] = 0xab; | |
2552 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
2553 | validate_slab_cache(kmalloc_caches + 9); | |
2554 | } | |
2555 | #else | |
2556 | static void resiliency_test(void) {}; | |
2557 | #endif | |
2558 | ||
81819f0f CL |
2559 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) |
2560 | { | |
2561 | struct kmem_cache *s = get_slab(size, gfpflags); | |
81819f0f CL |
2562 | |
2563 | if (!s) | |
2564 | return NULL; | |
2565 | ||
77c5e2d0 | 2566 | return slab_alloc(s, gfpflags, -1, caller); |
81819f0f CL |
2567 | } |
2568 | ||
2569 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
2570 | int node, void *caller) | |
2571 | { | |
2572 | struct kmem_cache *s = get_slab(size, gfpflags); | |
81819f0f CL |
2573 | |
2574 | if (!s) | |
2575 | return NULL; | |
2576 | ||
77c5e2d0 | 2577 | return slab_alloc(s, gfpflags, node, caller); |
81819f0f CL |
2578 | } |
2579 | ||
2580 | #ifdef CONFIG_SYSFS | |
2581 | ||
53e15af0 CL |
2582 | static int validate_slab(struct kmem_cache *s, struct page *page) |
2583 | { | |
2584 | void *p; | |
2585 | void *addr = page_address(page); | |
2586 | unsigned long map[BITS_TO_LONGS(s->objects)]; | |
2587 | ||
2588 | if (!check_slab(s, page) || | |
2589 | !on_freelist(s, page, NULL)) | |
2590 | return 0; | |
2591 | ||
2592 | /* Now we know that a valid freelist exists */ | |
2593 | bitmap_zero(map, s->objects); | |
2594 | ||
2595 | for(p = page->freelist; p; p = get_freepointer(s, p)) { | |
2596 | set_bit((p - addr) / s->size, map); | |
2597 | if (!check_object(s, page, p, 0)) | |
2598 | return 0; | |
2599 | } | |
2600 | ||
2601 | for(p = addr; p < addr + s->objects * s->size; p += s->size) | |
2602 | if (!test_bit((p - addr) / s->size, map)) | |
2603 | if (!check_object(s, page, p, 1)) | |
2604 | return 0; | |
2605 | return 1; | |
2606 | } | |
2607 | ||
2608 | static void validate_slab_slab(struct kmem_cache *s, struct page *page) | |
2609 | { | |
2610 | if (slab_trylock(page)) { | |
2611 | validate_slab(s, page); | |
2612 | slab_unlock(page); | |
2613 | } else | |
2614 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", | |
2615 | s->name, page); | |
2616 | ||
2617 | if (s->flags & DEBUG_DEFAULT_FLAGS) { | |
2618 | if (!PageError(page)) | |
2619 | printk(KERN_ERR "SLUB %s: PageError not set " | |
2620 | "on slab 0x%p\n", s->name, page); | |
2621 | } else { | |
2622 | if (PageError(page)) | |
2623 | printk(KERN_ERR "SLUB %s: PageError set on " | |
2624 | "slab 0x%p\n", s->name, page); | |
2625 | } | |
2626 | } | |
2627 | ||
2628 | static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n) | |
2629 | { | |
2630 | unsigned long count = 0; | |
2631 | struct page *page; | |
2632 | unsigned long flags; | |
2633 | ||
2634 | spin_lock_irqsave(&n->list_lock, flags); | |
2635 | ||
2636 | list_for_each_entry(page, &n->partial, lru) { | |
2637 | validate_slab_slab(s, page); | |
2638 | count++; | |
2639 | } | |
2640 | if (count != n->nr_partial) | |
2641 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
2642 | "counter=%ld\n", s->name, count, n->nr_partial); | |
2643 | ||
2644 | if (!(s->flags & SLAB_STORE_USER)) | |
2645 | goto out; | |
2646 | ||
2647 | list_for_each_entry(page, &n->full, lru) { | |
2648 | validate_slab_slab(s, page); | |
2649 | count++; | |
2650 | } | |
2651 | if (count != atomic_long_read(&n->nr_slabs)) | |
2652 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
2653 | "counter=%ld\n", s->name, count, | |
2654 | atomic_long_read(&n->nr_slabs)); | |
2655 | ||
2656 | out: | |
2657 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2658 | return count; | |
2659 | } | |
2660 | ||
2661 | static unsigned long validate_slab_cache(struct kmem_cache *s) | |
2662 | { | |
2663 | int node; | |
2664 | unsigned long count = 0; | |
2665 | ||
2666 | flush_all(s); | |
2667 | for_each_online_node(node) { | |
2668 | struct kmem_cache_node *n = get_node(s, node); | |
2669 | ||
2670 | count += validate_slab_node(s, n); | |
2671 | } | |
2672 | return count; | |
2673 | } | |
2674 | ||
88a420e4 | 2675 | /* |
672bba3a | 2676 | * Generate lists of code addresses where slabcache objects are allocated |
88a420e4 CL |
2677 | * and freed. |
2678 | */ | |
2679 | ||
2680 | struct location { | |
2681 | unsigned long count; | |
2682 | void *addr; | |
2683 | }; | |
2684 | ||
2685 | struct loc_track { | |
2686 | unsigned long max; | |
2687 | unsigned long count; | |
2688 | struct location *loc; | |
2689 | }; | |
2690 | ||
2691 | static void free_loc_track(struct loc_track *t) | |
2692 | { | |
2693 | if (t->max) | |
2694 | free_pages((unsigned long)t->loc, | |
2695 | get_order(sizeof(struct location) * t->max)); | |
2696 | } | |
2697 | ||
2698 | static int alloc_loc_track(struct loc_track *t, unsigned long max) | |
2699 | { | |
2700 | struct location *l; | |
2701 | int order; | |
2702 | ||
2703 | if (!max) | |
2704 | max = PAGE_SIZE / sizeof(struct location); | |
2705 | ||
2706 | order = get_order(sizeof(struct location) * max); | |
2707 | ||
2708 | l = (void *)__get_free_pages(GFP_KERNEL, order); | |
2709 | ||
2710 | if (!l) | |
2711 | return 0; | |
2712 | ||
2713 | if (t->count) { | |
2714 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
2715 | free_loc_track(t); | |
2716 | } | |
2717 | t->max = max; | |
2718 | t->loc = l; | |
2719 | return 1; | |
2720 | } | |
2721 | ||
2722 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
2723 | void *addr) | |
2724 | { | |
2725 | long start, end, pos; | |
2726 | struct location *l; | |
2727 | void *caddr; | |
2728 | ||
2729 | start = -1; | |
2730 | end = t->count; | |
2731 | ||
2732 | for ( ; ; ) { | |
2733 | pos = start + (end - start + 1) / 2; | |
2734 | ||
2735 | /* | |
2736 | * There is nothing at "end". If we end up there | |
2737 | * we need to add something to before end. | |
2738 | */ | |
2739 | if (pos == end) | |
2740 | break; | |
2741 | ||
2742 | caddr = t->loc[pos].addr; | |
2743 | if (addr == caddr) { | |
2744 | t->loc[pos].count++; | |
2745 | return 1; | |
2746 | } | |
2747 | ||
2748 | if (addr < caddr) | |
2749 | end = pos; | |
2750 | else | |
2751 | start = pos; | |
2752 | } | |
2753 | ||
2754 | /* | |
672bba3a | 2755 | * Not found. Insert new tracking element. |
88a420e4 CL |
2756 | */ |
2757 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max)) | |
2758 | return 0; | |
2759 | ||
2760 | l = t->loc + pos; | |
2761 | if (pos < t->count) | |
2762 | memmove(l + 1, l, | |
2763 | (t->count - pos) * sizeof(struct location)); | |
2764 | t->count++; | |
2765 | l->count = 1; | |
2766 | l->addr = addr; | |
2767 | return 1; | |
2768 | } | |
2769 | ||
2770 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
2771 | struct page *page, enum track_item alloc) | |
2772 | { | |
2773 | void *addr = page_address(page); | |
2774 | unsigned long map[BITS_TO_LONGS(s->objects)]; | |
2775 | void *p; | |
2776 | ||
2777 | bitmap_zero(map, s->objects); | |
2778 | for (p = page->freelist; p; p = get_freepointer(s, p)) | |
2779 | set_bit((p - addr) / s->size, map); | |
2780 | ||
2781 | for (p = addr; p < addr + s->objects * s->size; p += s->size) | |
2782 | if (!test_bit((p - addr) / s->size, map)) { | |
2783 | void *addr = get_track(s, p, alloc)->addr; | |
2784 | ||
2785 | add_location(t, s, addr); | |
2786 | } | |
2787 | } | |
2788 | ||
2789 | static int list_locations(struct kmem_cache *s, char *buf, | |
2790 | enum track_item alloc) | |
2791 | { | |
2792 | int n = 0; | |
2793 | unsigned long i; | |
2794 | struct loc_track t; | |
2795 | int node; | |
2796 | ||
2797 | t.count = 0; | |
2798 | t.max = 0; | |
2799 | ||
2800 | /* Push back cpu slabs */ | |
2801 | flush_all(s); | |
2802 | ||
2803 | for_each_online_node(node) { | |
2804 | struct kmem_cache_node *n = get_node(s, node); | |
2805 | unsigned long flags; | |
2806 | struct page *page; | |
2807 | ||
2808 | if (!atomic_read(&n->nr_slabs)) | |
2809 | continue; | |
2810 | ||
2811 | spin_lock_irqsave(&n->list_lock, flags); | |
2812 | list_for_each_entry(page, &n->partial, lru) | |
2813 | process_slab(&t, s, page, alloc); | |
2814 | list_for_each_entry(page, &n->full, lru) | |
2815 | process_slab(&t, s, page, alloc); | |
2816 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2817 | } | |
2818 | ||
2819 | for (i = 0; i < t.count; i++) { | |
2820 | void *addr = t.loc[i].addr; | |
2821 | ||
2822 | if (n > PAGE_SIZE - 100) | |
2823 | break; | |
2824 | n += sprintf(buf + n, "%7ld ", t.loc[i].count); | |
2825 | if (addr) | |
2826 | n += sprint_symbol(buf + n, (unsigned long)t.loc[i].addr); | |
2827 | else | |
2828 | n += sprintf(buf + n, "<not-available>"); | |
2829 | n += sprintf(buf + n, "\n"); | |
2830 | } | |
2831 | ||
2832 | free_loc_track(&t); | |
2833 | if (!t.count) | |
2834 | n += sprintf(buf, "No data\n"); | |
2835 | return n; | |
2836 | } | |
2837 | ||
81819f0f CL |
2838 | static unsigned long count_partial(struct kmem_cache_node *n) |
2839 | { | |
2840 | unsigned long flags; | |
2841 | unsigned long x = 0; | |
2842 | struct page *page; | |
2843 | ||
2844 | spin_lock_irqsave(&n->list_lock, flags); | |
2845 | list_for_each_entry(page, &n->partial, lru) | |
2846 | x += page->inuse; | |
2847 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2848 | return x; | |
2849 | } | |
2850 | ||
2851 | enum slab_stat_type { | |
2852 | SL_FULL, | |
2853 | SL_PARTIAL, | |
2854 | SL_CPU, | |
2855 | SL_OBJECTS | |
2856 | }; | |
2857 | ||
2858 | #define SO_FULL (1 << SL_FULL) | |
2859 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
2860 | #define SO_CPU (1 << SL_CPU) | |
2861 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
2862 | ||
2863 | static unsigned long slab_objects(struct kmem_cache *s, | |
2864 | char *buf, unsigned long flags) | |
2865 | { | |
2866 | unsigned long total = 0; | |
2867 | int cpu; | |
2868 | int node; | |
2869 | int x; | |
2870 | unsigned long *nodes; | |
2871 | unsigned long *per_cpu; | |
2872 | ||
2873 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
2874 | per_cpu = nodes + nr_node_ids; | |
2875 | ||
2876 | for_each_possible_cpu(cpu) { | |
2877 | struct page *page = s->cpu_slab[cpu]; | |
2878 | int node; | |
2879 | ||
2880 | if (page) { | |
2881 | node = page_to_nid(page); | |
2882 | if (flags & SO_CPU) { | |
2883 | int x = 0; | |
2884 | ||
2885 | if (flags & SO_OBJECTS) | |
2886 | x = page->inuse; | |
2887 | else | |
2888 | x = 1; | |
2889 | total += x; | |
2890 | nodes[node] += x; | |
2891 | } | |
2892 | per_cpu[node]++; | |
2893 | } | |
2894 | } | |
2895 | ||
2896 | for_each_online_node(node) { | |
2897 | struct kmem_cache_node *n = get_node(s, node); | |
2898 | ||
2899 | if (flags & SO_PARTIAL) { | |
2900 | if (flags & SO_OBJECTS) | |
2901 | x = count_partial(n); | |
2902 | else | |
2903 | x = n->nr_partial; | |
2904 | total += x; | |
2905 | nodes[node] += x; | |
2906 | } | |
2907 | ||
2908 | if (flags & SO_FULL) { | |
2909 | int full_slabs = atomic_read(&n->nr_slabs) | |
2910 | - per_cpu[node] | |
2911 | - n->nr_partial; | |
2912 | ||
2913 | if (flags & SO_OBJECTS) | |
2914 | x = full_slabs * s->objects; | |
2915 | else | |
2916 | x = full_slabs; | |
2917 | total += x; | |
2918 | nodes[node] += x; | |
2919 | } | |
2920 | } | |
2921 | ||
2922 | x = sprintf(buf, "%lu", total); | |
2923 | #ifdef CONFIG_NUMA | |
2924 | for_each_online_node(node) | |
2925 | if (nodes[node]) | |
2926 | x += sprintf(buf + x, " N%d=%lu", | |
2927 | node, nodes[node]); | |
2928 | #endif | |
2929 | kfree(nodes); | |
2930 | return x + sprintf(buf + x, "\n"); | |
2931 | } | |
2932 | ||
2933 | static int any_slab_objects(struct kmem_cache *s) | |
2934 | { | |
2935 | int node; | |
2936 | int cpu; | |
2937 | ||
2938 | for_each_possible_cpu(cpu) | |
2939 | if (s->cpu_slab[cpu]) | |
2940 | return 1; | |
2941 | ||
2942 | for_each_node(node) { | |
2943 | struct kmem_cache_node *n = get_node(s, node); | |
2944 | ||
2945 | if (n->nr_partial || atomic_read(&n->nr_slabs)) | |
2946 | return 1; | |
2947 | } | |
2948 | return 0; | |
2949 | } | |
2950 | ||
2951 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
2952 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); | |
2953 | ||
2954 | struct slab_attribute { | |
2955 | struct attribute attr; | |
2956 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
2957 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
2958 | }; | |
2959 | ||
2960 | #define SLAB_ATTR_RO(_name) \ | |
2961 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) | |
2962 | ||
2963 | #define SLAB_ATTR(_name) \ | |
2964 | static struct slab_attribute _name##_attr = \ | |
2965 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2966 | ||
81819f0f CL |
2967 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
2968 | { | |
2969 | return sprintf(buf, "%d\n", s->size); | |
2970 | } | |
2971 | SLAB_ATTR_RO(slab_size); | |
2972 | ||
2973 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
2974 | { | |
2975 | return sprintf(buf, "%d\n", s->align); | |
2976 | } | |
2977 | SLAB_ATTR_RO(align); | |
2978 | ||
2979 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
2980 | { | |
2981 | return sprintf(buf, "%d\n", s->objsize); | |
2982 | } | |
2983 | SLAB_ATTR_RO(object_size); | |
2984 | ||
2985 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
2986 | { | |
2987 | return sprintf(buf, "%d\n", s->objects); | |
2988 | } | |
2989 | SLAB_ATTR_RO(objs_per_slab); | |
2990 | ||
2991 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
2992 | { | |
2993 | return sprintf(buf, "%d\n", s->order); | |
2994 | } | |
2995 | SLAB_ATTR_RO(order); | |
2996 | ||
2997 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
2998 | { | |
2999 | if (s->ctor) { | |
3000 | int n = sprint_symbol(buf, (unsigned long)s->ctor); | |
3001 | ||
3002 | return n + sprintf(buf + n, "\n"); | |
3003 | } | |
3004 | return 0; | |
3005 | } | |
3006 | SLAB_ATTR_RO(ctor); | |
3007 | ||
3008 | static ssize_t dtor_show(struct kmem_cache *s, char *buf) | |
3009 | { | |
3010 | if (s->dtor) { | |
3011 | int n = sprint_symbol(buf, (unsigned long)s->dtor); | |
3012 | ||
3013 | return n + sprintf(buf + n, "\n"); | |
3014 | } | |
3015 | return 0; | |
3016 | } | |
3017 | SLAB_ATTR_RO(dtor); | |
3018 | ||
3019 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) | |
3020 | { | |
3021 | return sprintf(buf, "%d\n", s->refcount - 1); | |
3022 | } | |
3023 | SLAB_ATTR_RO(aliases); | |
3024 | ||
3025 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
3026 | { | |
3027 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU); | |
3028 | } | |
3029 | SLAB_ATTR_RO(slabs); | |
3030 | ||
3031 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
3032 | { | |
3033 | return slab_objects(s, buf, SO_PARTIAL); | |
3034 | } | |
3035 | SLAB_ATTR_RO(partial); | |
3036 | ||
3037 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
3038 | { | |
3039 | return slab_objects(s, buf, SO_CPU); | |
3040 | } | |
3041 | SLAB_ATTR_RO(cpu_slabs); | |
3042 | ||
3043 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
3044 | { | |
3045 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS); | |
3046 | } | |
3047 | SLAB_ATTR_RO(objects); | |
3048 | ||
3049 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
3050 | { | |
3051 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
3052 | } | |
3053 | ||
3054 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
3055 | const char *buf, size_t length) | |
3056 | { | |
3057 | s->flags &= ~SLAB_DEBUG_FREE; | |
3058 | if (buf[0] == '1') | |
3059 | s->flags |= SLAB_DEBUG_FREE; | |
3060 | return length; | |
3061 | } | |
3062 | SLAB_ATTR(sanity_checks); | |
3063 | ||
3064 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
3065 | { | |
3066 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
3067 | } | |
3068 | ||
3069 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
3070 | size_t length) | |
3071 | { | |
3072 | s->flags &= ~SLAB_TRACE; | |
3073 | if (buf[0] == '1') | |
3074 | s->flags |= SLAB_TRACE; | |
3075 | return length; | |
3076 | } | |
3077 | SLAB_ATTR(trace); | |
3078 | ||
3079 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
3080 | { | |
3081 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
3082 | } | |
3083 | ||
3084 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
3085 | const char *buf, size_t length) | |
3086 | { | |
3087 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
3088 | if (buf[0] == '1') | |
3089 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
3090 | return length; | |
3091 | } | |
3092 | SLAB_ATTR(reclaim_account); | |
3093 | ||
3094 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
3095 | { | |
5af60839 | 3096 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); |
81819f0f CL |
3097 | } |
3098 | SLAB_ATTR_RO(hwcache_align); | |
3099 | ||
3100 | #ifdef CONFIG_ZONE_DMA | |
3101 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
3102 | { | |
3103 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
3104 | } | |
3105 | SLAB_ATTR_RO(cache_dma); | |
3106 | #endif | |
3107 | ||
3108 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
3109 | { | |
3110 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
3111 | } | |
3112 | SLAB_ATTR_RO(destroy_by_rcu); | |
3113 | ||
3114 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
3115 | { | |
3116 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
3117 | } | |
3118 | ||
3119 | static ssize_t red_zone_store(struct kmem_cache *s, | |
3120 | const char *buf, size_t length) | |
3121 | { | |
3122 | if (any_slab_objects(s)) | |
3123 | return -EBUSY; | |
3124 | ||
3125 | s->flags &= ~SLAB_RED_ZONE; | |
3126 | if (buf[0] == '1') | |
3127 | s->flags |= SLAB_RED_ZONE; | |
3128 | calculate_sizes(s); | |
3129 | return length; | |
3130 | } | |
3131 | SLAB_ATTR(red_zone); | |
3132 | ||
3133 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
3134 | { | |
3135 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
3136 | } | |
3137 | ||
3138 | static ssize_t poison_store(struct kmem_cache *s, | |
3139 | const char *buf, size_t length) | |
3140 | { | |
3141 | if (any_slab_objects(s)) | |
3142 | return -EBUSY; | |
3143 | ||
3144 | s->flags &= ~SLAB_POISON; | |
3145 | if (buf[0] == '1') | |
3146 | s->flags |= SLAB_POISON; | |
3147 | calculate_sizes(s); | |
3148 | return length; | |
3149 | } | |
3150 | SLAB_ATTR(poison); | |
3151 | ||
3152 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
3153 | { | |
3154 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
3155 | } | |
3156 | ||
3157 | static ssize_t store_user_store(struct kmem_cache *s, | |
3158 | const char *buf, size_t length) | |
3159 | { | |
3160 | if (any_slab_objects(s)) | |
3161 | return -EBUSY; | |
3162 | ||
3163 | s->flags &= ~SLAB_STORE_USER; | |
3164 | if (buf[0] == '1') | |
3165 | s->flags |= SLAB_STORE_USER; | |
3166 | calculate_sizes(s); | |
3167 | return length; | |
3168 | } | |
3169 | SLAB_ATTR(store_user); | |
3170 | ||
53e15af0 CL |
3171 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
3172 | { | |
3173 | return 0; | |
3174 | } | |
3175 | ||
3176 | static ssize_t validate_store(struct kmem_cache *s, | |
3177 | const char *buf, size_t length) | |
3178 | { | |
3179 | if (buf[0] == '1') | |
3180 | validate_slab_cache(s); | |
3181 | else | |
3182 | return -EINVAL; | |
3183 | return length; | |
3184 | } | |
3185 | SLAB_ATTR(validate); | |
3186 | ||
2086d26a CL |
3187 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
3188 | { | |
3189 | return 0; | |
3190 | } | |
3191 | ||
3192 | static ssize_t shrink_store(struct kmem_cache *s, | |
3193 | const char *buf, size_t length) | |
3194 | { | |
3195 | if (buf[0] == '1') { | |
3196 | int rc = kmem_cache_shrink(s); | |
3197 | ||
3198 | if (rc) | |
3199 | return rc; | |
3200 | } else | |
3201 | return -EINVAL; | |
3202 | return length; | |
3203 | } | |
3204 | SLAB_ATTR(shrink); | |
3205 | ||
88a420e4 CL |
3206 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) |
3207 | { | |
3208 | if (!(s->flags & SLAB_STORE_USER)) | |
3209 | return -ENOSYS; | |
3210 | return list_locations(s, buf, TRACK_ALLOC); | |
3211 | } | |
3212 | SLAB_ATTR_RO(alloc_calls); | |
3213 | ||
3214 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
3215 | { | |
3216 | if (!(s->flags & SLAB_STORE_USER)) | |
3217 | return -ENOSYS; | |
3218 | return list_locations(s, buf, TRACK_FREE); | |
3219 | } | |
3220 | SLAB_ATTR_RO(free_calls); | |
3221 | ||
81819f0f CL |
3222 | #ifdef CONFIG_NUMA |
3223 | static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf) | |
3224 | { | |
3225 | return sprintf(buf, "%d\n", s->defrag_ratio / 10); | |
3226 | } | |
3227 | ||
3228 | static ssize_t defrag_ratio_store(struct kmem_cache *s, | |
3229 | const char *buf, size_t length) | |
3230 | { | |
3231 | int n = simple_strtoul(buf, NULL, 10); | |
3232 | ||
3233 | if (n < 100) | |
3234 | s->defrag_ratio = n * 10; | |
3235 | return length; | |
3236 | } | |
3237 | SLAB_ATTR(defrag_ratio); | |
3238 | #endif | |
3239 | ||
3240 | static struct attribute * slab_attrs[] = { | |
3241 | &slab_size_attr.attr, | |
3242 | &object_size_attr.attr, | |
3243 | &objs_per_slab_attr.attr, | |
3244 | &order_attr.attr, | |
3245 | &objects_attr.attr, | |
3246 | &slabs_attr.attr, | |
3247 | &partial_attr.attr, | |
3248 | &cpu_slabs_attr.attr, | |
3249 | &ctor_attr.attr, | |
3250 | &dtor_attr.attr, | |
3251 | &aliases_attr.attr, | |
3252 | &align_attr.attr, | |
3253 | &sanity_checks_attr.attr, | |
3254 | &trace_attr.attr, | |
3255 | &hwcache_align_attr.attr, | |
3256 | &reclaim_account_attr.attr, | |
3257 | &destroy_by_rcu_attr.attr, | |
3258 | &red_zone_attr.attr, | |
3259 | &poison_attr.attr, | |
3260 | &store_user_attr.attr, | |
53e15af0 | 3261 | &validate_attr.attr, |
2086d26a | 3262 | &shrink_attr.attr, |
88a420e4 CL |
3263 | &alloc_calls_attr.attr, |
3264 | &free_calls_attr.attr, | |
81819f0f CL |
3265 | #ifdef CONFIG_ZONE_DMA |
3266 | &cache_dma_attr.attr, | |
3267 | #endif | |
3268 | #ifdef CONFIG_NUMA | |
3269 | &defrag_ratio_attr.attr, | |
3270 | #endif | |
3271 | NULL | |
3272 | }; | |
3273 | ||
3274 | static struct attribute_group slab_attr_group = { | |
3275 | .attrs = slab_attrs, | |
3276 | }; | |
3277 | ||
3278 | static ssize_t slab_attr_show(struct kobject *kobj, | |
3279 | struct attribute *attr, | |
3280 | char *buf) | |
3281 | { | |
3282 | struct slab_attribute *attribute; | |
3283 | struct kmem_cache *s; | |
3284 | int err; | |
3285 | ||
3286 | attribute = to_slab_attr(attr); | |
3287 | s = to_slab(kobj); | |
3288 | ||
3289 | if (!attribute->show) | |
3290 | return -EIO; | |
3291 | ||
3292 | err = attribute->show(s, buf); | |
3293 | ||
3294 | return err; | |
3295 | } | |
3296 | ||
3297 | static ssize_t slab_attr_store(struct kobject *kobj, | |
3298 | struct attribute *attr, | |
3299 | const char *buf, size_t len) | |
3300 | { | |
3301 | struct slab_attribute *attribute; | |
3302 | struct kmem_cache *s; | |
3303 | int err; | |
3304 | ||
3305 | attribute = to_slab_attr(attr); | |
3306 | s = to_slab(kobj); | |
3307 | ||
3308 | if (!attribute->store) | |
3309 | return -EIO; | |
3310 | ||
3311 | err = attribute->store(s, buf, len); | |
3312 | ||
3313 | return err; | |
3314 | } | |
3315 | ||
3316 | static struct sysfs_ops slab_sysfs_ops = { | |
3317 | .show = slab_attr_show, | |
3318 | .store = slab_attr_store, | |
3319 | }; | |
3320 | ||
3321 | static struct kobj_type slab_ktype = { | |
3322 | .sysfs_ops = &slab_sysfs_ops, | |
3323 | }; | |
3324 | ||
3325 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
3326 | { | |
3327 | struct kobj_type *ktype = get_ktype(kobj); | |
3328 | ||
3329 | if (ktype == &slab_ktype) | |
3330 | return 1; | |
3331 | return 0; | |
3332 | } | |
3333 | ||
3334 | static struct kset_uevent_ops slab_uevent_ops = { | |
3335 | .filter = uevent_filter, | |
3336 | }; | |
3337 | ||
3338 | decl_subsys(slab, &slab_ktype, &slab_uevent_ops); | |
3339 | ||
3340 | #define ID_STR_LENGTH 64 | |
3341 | ||
3342 | /* Create a unique string id for a slab cache: | |
3343 | * format | |
3344 | * :[flags-]size:[memory address of kmemcache] | |
3345 | */ | |
3346 | static char *create_unique_id(struct kmem_cache *s) | |
3347 | { | |
3348 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
3349 | char *p = name; | |
3350 | ||
3351 | BUG_ON(!name); | |
3352 | ||
3353 | *p++ = ':'; | |
3354 | /* | |
3355 | * First flags affecting slabcache operations. We will only | |
3356 | * get here for aliasable slabs so we do not need to support | |
3357 | * too many flags. The flags here must cover all flags that | |
3358 | * are matched during merging to guarantee that the id is | |
3359 | * unique. | |
3360 | */ | |
3361 | if (s->flags & SLAB_CACHE_DMA) | |
3362 | *p++ = 'd'; | |
3363 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
3364 | *p++ = 'a'; | |
3365 | if (s->flags & SLAB_DEBUG_FREE) | |
3366 | *p++ = 'F'; | |
3367 | if (p != name + 1) | |
3368 | *p++ = '-'; | |
3369 | p += sprintf(p, "%07d", s->size); | |
3370 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
3371 | return name; | |
3372 | } | |
3373 | ||
3374 | static int sysfs_slab_add(struct kmem_cache *s) | |
3375 | { | |
3376 | int err; | |
3377 | const char *name; | |
3378 | int unmergeable; | |
3379 | ||
3380 | if (slab_state < SYSFS) | |
3381 | /* Defer until later */ | |
3382 | return 0; | |
3383 | ||
3384 | unmergeable = slab_unmergeable(s); | |
3385 | if (unmergeable) { | |
3386 | /* | |
3387 | * Slabcache can never be merged so we can use the name proper. | |
3388 | * This is typically the case for debug situations. In that | |
3389 | * case we can catch duplicate names easily. | |
3390 | */ | |
0f9008ef | 3391 | sysfs_remove_link(&slab_subsys.kobj, s->name); |
81819f0f CL |
3392 | name = s->name; |
3393 | } else { | |
3394 | /* | |
3395 | * Create a unique name for the slab as a target | |
3396 | * for the symlinks. | |
3397 | */ | |
3398 | name = create_unique_id(s); | |
3399 | } | |
3400 | ||
3401 | kobj_set_kset_s(s, slab_subsys); | |
3402 | kobject_set_name(&s->kobj, name); | |
3403 | kobject_init(&s->kobj); | |
3404 | err = kobject_add(&s->kobj); | |
3405 | if (err) | |
3406 | return err; | |
3407 | ||
3408 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
3409 | if (err) | |
3410 | return err; | |
3411 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
3412 | if (!unmergeable) { | |
3413 | /* Setup first alias */ | |
3414 | sysfs_slab_alias(s, s->name); | |
3415 | kfree(name); | |
3416 | } | |
3417 | return 0; | |
3418 | } | |
3419 | ||
3420 | static void sysfs_slab_remove(struct kmem_cache *s) | |
3421 | { | |
3422 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
3423 | kobject_del(&s->kobj); | |
3424 | } | |
3425 | ||
3426 | /* | |
3427 | * Need to buffer aliases during bootup until sysfs becomes | |
3428 | * available lest we loose that information. | |
3429 | */ | |
3430 | struct saved_alias { | |
3431 | struct kmem_cache *s; | |
3432 | const char *name; | |
3433 | struct saved_alias *next; | |
3434 | }; | |
3435 | ||
3436 | struct saved_alias *alias_list; | |
3437 | ||
3438 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
3439 | { | |
3440 | struct saved_alias *al; | |
3441 | ||
3442 | if (slab_state == SYSFS) { | |
3443 | /* | |
3444 | * If we have a leftover link then remove it. | |
3445 | */ | |
0f9008ef LT |
3446 | sysfs_remove_link(&slab_subsys.kobj, name); |
3447 | return sysfs_create_link(&slab_subsys.kobj, | |
81819f0f CL |
3448 | &s->kobj, name); |
3449 | } | |
3450 | ||
3451 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
3452 | if (!al) | |
3453 | return -ENOMEM; | |
3454 | ||
3455 | al->s = s; | |
3456 | al->name = name; | |
3457 | al->next = alias_list; | |
3458 | alias_list = al; | |
3459 | return 0; | |
3460 | } | |
3461 | ||
3462 | static int __init slab_sysfs_init(void) | |
3463 | { | |
26a7bd03 | 3464 | struct list_head *h; |
81819f0f CL |
3465 | int err; |
3466 | ||
3467 | err = subsystem_register(&slab_subsys); | |
3468 | if (err) { | |
3469 | printk(KERN_ERR "Cannot register slab subsystem.\n"); | |
3470 | return -ENOSYS; | |
3471 | } | |
3472 | ||
26a7bd03 CL |
3473 | slab_state = SYSFS; |
3474 | ||
3475 | list_for_each(h, &slab_caches) { | |
3476 | struct kmem_cache *s = | |
3477 | container_of(h, struct kmem_cache, list); | |
3478 | ||
3479 | err = sysfs_slab_add(s); | |
3480 | BUG_ON(err); | |
3481 | } | |
81819f0f CL |
3482 | |
3483 | while (alias_list) { | |
3484 | struct saved_alias *al = alias_list; | |
3485 | ||
3486 | alias_list = alias_list->next; | |
3487 | err = sysfs_slab_alias(al->s, al->name); | |
3488 | BUG_ON(err); | |
3489 | kfree(al); | |
3490 | } | |
3491 | ||
3492 | resiliency_test(); | |
3493 | return 0; | |
3494 | } | |
3495 | ||
3496 | __initcall(slab_sysfs_init); | |
81819f0f | 3497 | #endif |