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 | * | |
69 | * Slabs with free elements are kept on a partial list. | |
70 | * There is no list for full slabs. If an object in a full slab is | |
71 | * freed then the slab will show up again on the partial lists. | |
72 | * Otherwise there is no need to track full slabs unless we have to | |
73 | * track full slabs for debugging purposes. | |
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 | |
95 | * enough frequent updates. There are 16 pointers in a cacheline.so at | |
96 | * max 16 cpus could compete. Likely okay. | |
97 | * | |
98 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
99 | * | |
81819f0f CL |
100 | * - SLAB_DEBUG_INITIAL is not supported but I have never seen a use of |
101 | * it. | |
102 | * | |
103 | * - Variable sizing of the per node arrays | |
104 | */ | |
105 | ||
106 | /* Enable to test recovery from slab corruption on boot */ | |
107 | #undef SLUB_RESILIENCY_TEST | |
108 | ||
109 | #if PAGE_SHIFT <= 12 | |
110 | ||
111 | /* | |
112 | * Small page size. Make sure that we do not fragment memory | |
113 | */ | |
114 | #define DEFAULT_MAX_ORDER 1 | |
115 | #define DEFAULT_MIN_OBJECTS 4 | |
116 | ||
117 | #else | |
118 | ||
119 | /* | |
120 | * Large page machines are customarily able to handle larger | |
121 | * page orders. | |
122 | */ | |
123 | #define DEFAULT_MAX_ORDER 2 | |
124 | #define DEFAULT_MIN_OBJECTS 8 | |
125 | ||
126 | #endif | |
127 | ||
128 | /* | |
129 | * Flags from the regular SLAB that SLUB does not support: | |
130 | */ | |
131 | #define SLUB_UNIMPLEMENTED (SLAB_DEBUG_INITIAL) | |
132 | ||
2086d26a CL |
133 | /* |
134 | * Mininum number of partial slabs. These will be left on the partial | |
135 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
136 | */ | |
e95eed57 CL |
137 | #define MIN_PARTIAL 2 |
138 | ||
2086d26a CL |
139 | /* |
140 | * Maximum number of desirable partial slabs. | |
141 | * The existence of more partial slabs makes kmem_cache_shrink | |
142 | * sort the partial list by the number of objects in the. | |
143 | */ | |
144 | #define MAX_PARTIAL 10 | |
145 | ||
81819f0f CL |
146 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ |
147 | SLAB_POISON | SLAB_STORE_USER) | |
148 | /* | |
149 | * Set of flags that will prevent slab merging | |
150 | */ | |
151 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
152 | SLAB_TRACE | SLAB_DESTROY_BY_RCU) | |
153 | ||
154 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ | |
155 | SLAB_CACHE_DMA) | |
156 | ||
157 | #ifndef ARCH_KMALLOC_MINALIGN | |
47bfdc0d | 158 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
159 | #endif |
160 | ||
161 | #ifndef ARCH_SLAB_MINALIGN | |
47bfdc0d | 162 | #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
163 | #endif |
164 | ||
165 | /* Internal SLUB flags */ | |
166 | #define __OBJECT_POISON 0x80000000 /* Poison object */ | |
167 | ||
168 | static int kmem_size = sizeof(struct kmem_cache); | |
169 | ||
170 | #ifdef CONFIG_SMP | |
171 | static struct notifier_block slab_notifier; | |
172 | #endif | |
173 | ||
174 | static enum { | |
175 | DOWN, /* No slab functionality available */ | |
176 | PARTIAL, /* kmem_cache_open() works but kmalloc does not */ | |
177 | UP, /* Everything works */ | |
178 | SYSFS /* Sysfs up */ | |
179 | } slab_state = DOWN; | |
180 | ||
181 | /* A list of all slab caches on the system */ | |
182 | static DECLARE_RWSEM(slub_lock); | |
183 | LIST_HEAD(slab_caches); | |
184 | ||
185 | #ifdef CONFIG_SYSFS | |
186 | static int sysfs_slab_add(struct kmem_cache *); | |
187 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
188 | static void sysfs_slab_remove(struct kmem_cache *); | |
189 | #else | |
190 | static int sysfs_slab_add(struct kmem_cache *s) { return 0; } | |
191 | static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; } | |
192 | static void sysfs_slab_remove(struct kmem_cache *s) {} | |
193 | #endif | |
194 | ||
195 | /******************************************************************** | |
196 | * Core slab cache functions | |
197 | *******************************************************************/ | |
198 | ||
199 | int slab_is_available(void) | |
200 | { | |
201 | return slab_state >= UP; | |
202 | } | |
203 | ||
204 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
205 | { | |
206 | #ifdef CONFIG_NUMA | |
207 | return s->node[node]; | |
208 | #else | |
209 | return &s->local_node; | |
210 | #endif | |
211 | } | |
212 | ||
213 | /* | |
214 | * Object debugging | |
215 | */ | |
216 | static void print_section(char *text, u8 *addr, unsigned int length) | |
217 | { | |
218 | int i, offset; | |
219 | int newline = 1; | |
220 | char ascii[17]; | |
221 | ||
222 | ascii[16] = 0; | |
223 | ||
224 | for (i = 0; i < length; i++) { | |
225 | if (newline) { | |
226 | printk(KERN_ERR "%10s 0x%p: ", text, addr + i); | |
227 | newline = 0; | |
228 | } | |
229 | printk(" %02x", addr[i]); | |
230 | offset = i % 16; | |
231 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; | |
232 | if (offset == 15) { | |
233 | printk(" %s\n",ascii); | |
234 | newline = 1; | |
235 | } | |
236 | } | |
237 | if (!newline) { | |
238 | i %= 16; | |
239 | while (i < 16) { | |
240 | printk(" "); | |
241 | ascii[i] = ' '; | |
242 | i++; | |
243 | } | |
244 | printk(" %s\n", ascii); | |
245 | } | |
246 | } | |
247 | ||
248 | /* | |
249 | * Slow version of get and set free pointer. | |
250 | * | |
251 | * This requires touching the cache lines of kmem_cache. | |
252 | * The offset can also be obtained from the page. In that | |
253 | * case it is in the cacheline that we already need to touch. | |
254 | */ | |
255 | static void *get_freepointer(struct kmem_cache *s, void *object) | |
256 | { | |
257 | return *(void **)(object + s->offset); | |
258 | } | |
259 | ||
260 | static void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
261 | { | |
262 | *(void **)(object + s->offset) = fp; | |
263 | } | |
264 | ||
265 | /* | |
266 | * Tracking user of a slab. | |
267 | */ | |
268 | struct track { | |
269 | void *addr; /* Called from address */ | |
270 | int cpu; /* Was running on cpu */ | |
271 | int pid; /* Pid context */ | |
272 | unsigned long when; /* When did the operation occur */ | |
273 | }; | |
274 | ||
275 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
276 | ||
277 | static struct track *get_track(struct kmem_cache *s, void *object, | |
278 | enum track_item alloc) | |
279 | { | |
280 | struct track *p; | |
281 | ||
282 | if (s->offset) | |
283 | p = object + s->offset + sizeof(void *); | |
284 | else | |
285 | p = object + s->inuse; | |
286 | ||
287 | return p + alloc; | |
288 | } | |
289 | ||
290 | static void set_track(struct kmem_cache *s, void *object, | |
291 | enum track_item alloc, void *addr) | |
292 | { | |
293 | struct track *p; | |
294 | ||
295 | if (s->offset) | |
296 | p = object + s->offset + sizeof(void *); | |
297 | else | |
298 | p = object + s->inuse; | |
299 | ||
300 | p += alloc; | |
301 | if (addr) { | |
302 | p->addr = addr; | |
303 | p->cpu = smp_processor_id(); | |
304 | p->pid = current ? current->pid : -1; | |
305 | p->when = jiffies; | |
306 | } else | |
307 | memset(p, 0, sizeof(struct track)); | |
308 | } | |
309 | ||
81819f0f CL |
310 | static void init_tracking(struct kmem_cache *s, void *object) |
311 | { | |
312 | if (s->flags & SLAB_STORE_USER) { | |
313 | set_track(s, object, TRACK_FREE, NULL); | |
314 | set_track(s, object, TRACK_ALLOC, NULL); | |
315 | } | |
316 | } | |
317 | ||
318 | static void print_track(const char *s, struct track *t) | |
319 | { | |
320 | if (!t->addr) | |
321 | return; | |
322 | ||
323 | printk(KERN_ERR "%s: ", s); | |
324 | __print_symbol("%s", (unsigned long)t->addr); | |
325 | printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); | |
326 | } | |
327 | ||
328 | static void print_trailer(struct kmem_cache *s, u8 *p) | |
329 | { | |
330 | unsigned int off; /* Offset of last byte */ | |
331 | ||
332 | if (s->flags & SLAB_RED_ZONE) | |
333 | print_section("Redzone", p + s->objsize, | |
334 | s->inuse - s->objsize); | |
335 | ||
336 | printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n", | |
337 | p + s->offset, | |
338 | get_freepointer(s, p)); | |
339 | ||
340 | if (s->offset) | |
341 | off = s->offset + sizeof(void *); | |
342 | else | |
343 | off = s->inuse; | |
344 | ||
345 | if (s->flags & SLAB_STORE_USER) { | |
346 | print_track("Last alloc", get_track(s, p, TRACK_ALLOC)); | |
347 | print_track("Last free ", get_track(s, p, TRACK_FREE)); | |
348 | off += 2 * sizeof(struct track); | |
349 | } | |
350 | ||
351 | if (off != s->size) | |
352 | /* Beginning of the filler is the free pointer */ | |
353 | print_section("Filler", p + off, s->size - off); | |
354 | } | |
355 | ||
356 | static void object_err(struct kmem_cache *s, struct page *page, | |
357 | u8 *object, char *reason) | |
358 | { | |
359 | u8 *addr = page_address(page); | |
360 | ||
361 | printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n", | |
362 | s->name, reason, object, page); | |
363 | printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n", | |
364 | object - addr, page->flags, page->inuse, page->freelist); | |
365 | if (object > addr + 16) | |
366 | print_section("Bytes b4", object - 16, 16); | |
367 | print_section("Object", object, min(s->objsize, 128)); | |
368 | print_trailer(s, object); | |
369 | dump_stack(); | |
370 | } | |
371 | ||
372 | static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...) | |
373 | { | |
374 | va_list args; | |
375 | char buf[100]; | |
376 | ||
377 | va_start(args, reason); | |
378 | vsnprintf(buf, sizeof(buf), reason, args); | |
379 | va_end(args); | |
380 | printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf, | |
381 | page); | |
382 | dump_stack(); | |
383 | } | |
384 | ||
385 | static void init_object(struct kmem_cache *s, void *object, int active) | |
386 | { | |
387 | u8 *p = object; | |
388 | ||
389 | if (s->flags & __OBJECT_POISON) { | |
390 | memset(p, POISON_FREE, s->objsize - 1); | |
391 | p[s->objsize -1] = POISON_END; | |
392 | } | |
393 | ||
394 | if (s->flags & SLAB_RED_ZONE) | |
395 | memset(p + s->objsize, | |
396 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, | |
397 | s->inuse - s->objsize); | |
398 | } | |
399 | ||
400 | static int check_bytes(u8 *start, unsigned int value, unsigned int bytes) | |
401 | { | |
402 | while (bytes) { | |
403 | if (*start != (u8)value) | |
404 | return 0; | |
405 | start++; | |
406 | bytes--; | |
407 | } | |
408 | return 1; | |
409 | } | |
410 | ||
411 | ||
412 | static int check_valid_pointer(struct kmem_cache *s, struct page *page, | |
413 | void *object) | |
414 | { | |
415 | void *base; | |
416 | ||
417 | if (!object) | |
418 | return 1; | |
419 | ||
420 | base = page_address(page); | |
421 | if (object < base || object >= base + s->objects * s->size || | |
422 | (object - base) % s->size) { | |
423 | return 0; | |
424 | } | |
425 | ||
426 | return 1; | |
427 | } | |
428 | ||
429 | /* | |
430 | * Object layout: | |
431 | * | |
432 | * object address | |
433 | * Bytes of the object to be managed. | |
434 | * If the freepointer may overlay the object then the free | |
435 | * pointer is the first word of the object. | |
436 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is | |
437 | * 0xa5 (POISON_END) | |
438 | * | |
439 | * object + s->objsize | |
440 | * Padding to reach word boundary. This is also used for Redzoning. | |
441 | * Padding is extended to word size if Redzoning is enabled | |
442 | * and objsize == inuse. | |
443 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with | |
444 | * 0xcc (RED_ACTIVE) for objects in use. | |
445 | * | |
446 | * object + s->inuse | |
447 | * A. Free pointer (if we cannot overwrite object on free) | |
448 | * B. Tracking data for SLAB_STORE_USER | |
449 | * C. Padding to reach required alignment boundary | |
450 | * Padding is done using 0x5a (POISON_INUSE) | |
451 | * | |
452 | * object + s->size | |
453 | * | |
454 | * If slabcaches are merged then the objsize and inuse boundaries are to | |
455 | * be ignored. And therefore no slab options that rely on these boundaries | |
456 | * may be used with merged slabcaches. | |
457 | */ | |
458 | ||
459 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
460 | void *from, void *to) | |
461 | { | |
462 | printk(KERN_ERR "@@@ SLUB: %s Restoring %s (0x%x) from 0x%p-0x%p\n", | |
463 | s->name, message, data, from, to - 1); | |
464 | memset(from, data, to - from); | |
465 | } | |
466 | ||
467 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) | |
468 | { | |
469 | unsigned long off = s->inuse; /* The end of info */ | |
470 | ||
471 | if (s->offset) | |
472 | /* Freepointer is placed after the object. */ | |
473 | off += sizeof(void *); | |
474 | ||
475 | if (s->flags & SLAB_STORE_USER) | |
476 | /* We also have user information there */ | |
477 | off += 2 * sizeof(struct track); | |
478 | ||
479 | if (s->size == off) | |
480 | return 1; | |
481 | ||
482 | if (check_bytes(p + off, POISON_INUSE, s->size - off)) | |
483 | return 1; | |
484 | ||
485 | object_err(s, page, p, "Object padding check fails"); | |
486 | ||
487 | /* | |
488 | * Restore padding | |
489 | */ | |
490 | restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size); | |
491 | return 0; | |
492 | } | |
493 | ||
494 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
495 | { | |
496 | u8 *p; | |
497 | int length, remainder; | |
498 | ||
499 | if (!(s->flags & SLAB_POISON)) | |
500 | return 1; | |
501 | ||
502 | p = page_address(page); | |
503 | length = s->objects * s->size; | |
504 | remainder = (PAGE_SIZE << s->order) - length; | |
505 | if (!remainder) | |
506 | return 1; | |
507 | ||
508 | if (!check_bytes(p + length, POISON_INUSE, remainder)) { | |
509 | printk(KERN_ERR "SLUB: %s slab 0x%p: Padding fails check\n", | |
510 | s->name, p); | |
511 | dump_stack(); | |
512 | restore_bytes(s, "slab padding", POISON_INUSE, p + length, | |
513 | p + length + remainder); | |
514 | return 0; | |
515 | } | |
516 | return 1; | |
517 | } | |
518 | ||
519 | static int check_object(struct kmem_cache *s, struct page *page, | |
520 | void *object, int active) | |
521 | { | |
522 | u8 *p = object; | |
523 | u8 *endobject = object + s->objsize; | |
524 | ||
525 | if (s->flags & SLAB_RED_ZONE) { | |
526 | unsigned int red = | |
527 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; | |
528 | ||
529 | if (!check_bytes(endobject, red, s->inuse - s->objsize)) { | |
530 | object_err(s, page, object, | |
531 | active ? "Redzone Active" : "Redzone Inactive"); | |
532 | restore_bytes(s, "redzone", red, | |
533 | endobject, object + s->inuse); | |
534 | return 0; | |
535 | } | |
536 | } else { | |
537 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse && | |
538 | !check_bytes(endobject, POISON_INUSE, | |
539 | s->inuse - s->objsize)) { | |
540 | object_err(s, page, p, "Alignment padding check fails"); | |
541 | /* | |
542 | * Fix it so that there will not be another report. | |
543 | * | |
544 | * Hmmm... We may be corrupting an object that now expects | |
545 | * to be longer than allowed. | |
546 | */ | |
547 | restore_bytes(s, "alignment padding", POISON_INUSE, | |
548 | endobject, object + s->inuse); | |
549 | } | |
550 | } | |
551 | ||
552 | if (s->flags & SLAB_POISON) { | |
553 | if (!active && (s->flags & __OBJECT_POISON) && | |
554 | (!check_bytes(p, POISON_FREE, s->objsize - 1) || | |
555 | p[s->objsize - 1] != POISON_END)) { | |
556 | ||
557 | object_err(s, page, p, "Poison check failed"); | |
558 | restore_bytes(s, "Poison", POISON_FREE, | |
559 | p, p + s->objsize -1); | |
560 | restore_bytes(s, "Poison", POISON_END, | |
561 | p + s->objsize - 1, p + s->objsize); | |
562 | return 0; | |
563 | } | |
564 | /* | |
565 | * check_pad_bytes cleans up on its own. | |
566 | */ | |
567 | check_pad_bytes(s, page, p); | |
568 | } | |
569 | ||
570 | if (!s->offset && active) | |
571 | /* | |
572 | * Object and freepointer overlap. Cannot check | |
573 | * freepointer while object is allocated. | |
574 | */ | |
575 | return 1; | |
576 | ||
577 | /* Check free pointer validity */ | |
578 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
579 | object_err(s, page, p, "Freepointer corrupt"); | |
580 | /* | |
581 | * No choice but to zap it and thus loose the remainder | |
582 | * of the free objects in this slab. May cause | |
583 | * another error because the object count maybe | |
584 | * wrong now. | |
585 | */ | |
586 | set_freepointer(s, p, NULL); | |
587 | return 0; | |
588 | } | |
589 | return 1; | |
590 | } | |
591 | ||
592 | static int check_slab(struct kmem_cache *s, struct page *page) | |
593 | { | |
594 | VM_BUG_ON(!irqs_disabled()); | |
595 | ||
596 | if (!PageSlab(page)) { | |
597 | printk(KERN_ERR "SLUB: %s Not a valid slab page @0x%p " | |
598 | "flags=%lx mapping=0x%p count=%d \n", | |
599 | s->name, page, page->flags, page->mapping, | |
600 | page_count(page)); | |
601 | return 0; | |
602 | } | |
603 | if (page->offset * sizeof(void *) != s->offset) { | |
604 | printk(KERN_ERR "SLUB: %s Corrupted offset %lu in slab @0x%p" | |
605 | " flags=0x%lx mapping=0x%p count=%d\n", | |
606 | s->name, | |
607 | (unsigned long)(page->offset * sizeof(void *)), | |
608 | page, | |
609 | page->flags, | |
610 | page->mapping, | |
611 | page_count(page)); | |
612 | dump_stack(); | |
613 | return 0; | |
614 | } | |
615 | if (page->inuse > s->objects) { | |
616 | printk(KERN_ERR "SLUB: %s Inuse %u > max %u in slab " | |
617 | "page @0x%p flags=%lx mapping=0x%p count=%d\n", | |
618 | s->name, page->inuse, s->objects, page, page->flags, | |
619 | page->mapping, page_count(page)); | |
620 | dump_stack(); | |
621 | return 0; | |
622 | } | |
623 | /* Slab_pad_check fixes things up after itself */ | |
624 | slab_pad_check(s, page); | |
625 | return 1; | |
626 | } | |
627 | ||
628 | /* | |
629 | * Determine if a certain object on a page is on the freelist and | |
630 | * therefore free. Must hold the slab lock for cpu slabs to | |
631 | * guarantee that the chains are consistent. | |
632 | */ | |
633 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
634 | { | |
635 | int nr = 0; | |
636 | void *fp = page->freelist; | |
637 | void *object = NULL; | |
638 | ||
639 | while (fp && nr <= s->objects) { | |
640 | if (fp == search) | |
641 | return 1; | |
642 | if (!check_valid_pointer(s, page, fp)) { | |
643 | if (object) { | |
644 | object_err(s, page, object, | |
645 | "Freechain corrupt"); | |
646 | set_freepointer(s, object, NULL); | |
647 | break; | |
648 | } else { | |
649 | printk(KERN_ERR "SLUB: %s slab 0x%p " | |
650 | "freepointer 0x%p corrupted.\n", | |
651 | s->name, page, fp); | |
652 | dump_stack(); | |
653 | page->freelist = NULL; | |
654 | page->inuse = s->objects; | |
655 | return 0; | |
656 | } | |
657 | break; | |
658 | } | |
659 | object = fp; | |
660 | fp = get_freepointer(s, object); | |
661 | nr++; | |
662 | } | |
663 | ||
664 | if (page->inuse != s->objects - nr) { | |
665 | printk(KERN_ERR "slab %s: page 0x%p wrong object count." | |
666 | " counter is %d but counted were %d\n", | |
667 | s->name, page, page->inuse, | |
668 | s->objects - nr); | |
669 | page->inuse = s->objects - nr; | |
670 | } | |
671 | return search == NULL; | |
672 | } | |
673 | ||
643b1138 CL |
674 | /* |
675 | * Tracking of fully allocated slabs for debugging | |
676 | */ | |
e95eed57 | 677 | static void add_full(struct kmem_cache_node *n, struct page *page) |
643b1138 | 678 | { |
643b1138 CL |
679 | spin_lock(&n->list_lock); |
680 | list_add(&page->lru, &n->full); | |
681 | spin_unlock(&n->list_lock); | |
682 | } | |
683 | ||
684 | static void remove_full(struct kmem_cache *s, struct page *page) | |
685 | { | |
686 | struct kmem_cache_node *n; | |
687 | ||
688 | if (!(s->flags & SLAB_STORE_USER)) | |
689 | return; | |
690 | ||
691 | n = get_node(s, page_to_nid(page)); | |
692 | ||
693 | spin_lock(&n->list_lock); | |
694 | list_del(&page->lru); | |
695 | spin_unlock(&n->list_lock); | |
696 | } | |
697 | ||
81819f0f CL |
698 | static int alloc_object_checks(struct kmem_cache *s, struct page *page, |
699 | void *object) | |
700 | { | |
701 | if (!check_slab(s, page)) | |
702 | goto bad; | |
703 | ||
704 | if (object && !on_freelist(s, page, object)) { | |
705 | printk(KERN_ERR "SLUB: %s Object 0x%p@0x%p " | |
706 | "already allocated.\n", | |
707 | s->name, object, page); | |
708 | goto dump; | |
709 | } | |
710 | ||
711 | if (!check_valid_pointer(s, page, object)) { | |
712 | object_err(s, page, object, "Freelist Pointer check fails"); | |
713 | goto dump; | |
714 | } | |
715 | ||
716 | if (!object) | |
717 | return 1; | |
718 | ||
719 | if (!check_object(s, page, object, 0)) | |
720 | goto bad; | |
721 | init_object(s, object, 1); | |
722 | ||
723 | if (s->flags & SLAB_TRACE) { | |
724 | printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n", | |
725 | s->name, object, page->inuse, | |
726 | page->freelist); | |
727 | dump_stack(); | |
728 | } | |
729 | return 1; | |
730 | dump: | |
731 | dump_stack(); | |
732 | bad: | |
733 | if (PageSlab(page)) { | |
734 | /* | |
735 | * If this is a slab page then lets do the best we can | |
736 | * to avoid issues in the future. Marking all objects | |
737 | * as used avoids touching the remainder. | |
738 | */ | |
739 | printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n", | |
740 | s->name, page); | |
741 | page->inuse = s->objects; | |
742 | page->freelist = NULL; | |
743 | /* Fix up fields that may be corrupted */ | |
744 | page->offset = s->offset / sizeof(void *); | |
745 | } | |
746 | return 0; | |
747 | } | |
748 | ||
749 | static int free_object_checks(struct kmem_cache *s, struct page *page, | |
750 | void *object) | |
751 | { | |
752 | if (!check_slab(s, page)) | |
753 | goto fail; | |
754 | ||
755 | if (!check_valid_pointer(s, page, object)) { | |
756 | printk(KERN_ERR "SLUB: %s slab 0x%p invalid " | |
757 | "object pointer 0x%p\n", | |
758 | s->name, page, object); | |
759 | goto fail; | |
760 | } | |
761 | ||
762 | if (on_freelist(s, page, object)) { | |
763 | printk(KERN_ERR "SLUB: %s slab 0x%p object " | |
764 | "0x%p already free.\n", s->name, page, object); | |
765 | goto fail; | |
766 | } | |
767 | ||
768 | if (!check_object(s, page, object, 1)) | |
769 | return 0; | |
770 | ||
771 | if (unlikely(s != page->slab)) { | |
772 | if (!PageSlab(page)) | |
773 | printk(KERN_ERR "slab_free %s size %d: attempt to" | |
774 | "free object(0x%p) outside of slab.\n", | |
775 | s->name, s->size, object); | |
776 | else | |
777 | if (!page->slab) | |
778 | printk(KERN_ERR | |
779 | "slab_free : no slab(NULL) for object 0x%p.\n", | |
780 | object); | |
781 | else | |
782 | printk(KERN_ERR "slab_free %s(%d): object at 0x%p" | |
783 | " belongs to slab %s(%d)\n", | |
784 | s->name, s->size, object, | |
785 | page->slab->name, page->slab->size); | |
786 | goto fail; | |
787 | } | |
788 | if (s->flags & SLAB_TRACE) { | |
789 | printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n", | |
790 | s->name, object, page->inuse, | |
791 | page->freelist); | |
792 | print_section("Object", object, s->objsize); | |
793 | dump_stack(); | |
794 | } | |
795 | init_object(s, object, 0); | |
796 | return 1; | |
797 | fail: | |
798 | dump_stack(); | |
799 | printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n", | |
800 | s->name, page, object); | |
801 | return 0; | |
802 | } | |
803 | ||
804 | /* | |
805 | * Slab allocation and freeing | |
806 | */ | |
807 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
808 | { | |
809 | struct page * page; | |
810 | int pages = 1 << s->order; | |
811 | ||
812 | if (s->order) | |
813 | flags |= __GFP_COMP; | |
814 | ||
815 | if (s->flags & SLAB_CACHE_DMA) | |
816 | flags |= SLUB_DMA; | |
817 | ||
818 | if (node == -1) | |
819 | page = alloc_pages(flags, s->order); | |
820 | else | |
821 | page = alloc_pages_node(node, flags, s->order); | |
822 | ||
823 | if (!page) | |
824 | return NULL; | |
825 | ||
826 | mod_zone_page_state(page_zone(page), | |
827 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
828 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
829 | pages); | |
830 | ||
831 | return page; | |
832 | } | |
833 | ||
834 | static void setup_object(struct kmem_cache *s, struct page *page, | |
835 | void *object) | |
836 | { | |
837 | if (PageError(page)) { | |
838 | init_object(s, object, 0); | |
839 | init_tracking(s, object); | |
840 | } | |
841 | ||
842 | if (unlikely(s->ctor)) { | |
843 | int mode = SLAB_CTOR_CONSTRUCTOR; | |
844 | ||
845 | if (!(s->flags & __GFP_WAIT)) | |
846 | mode |= SLAB_CTOR_ATOMIC; | |
847 | ||
848 | s->ctor(object, s, mode); | |
849 | } | |
850 | } | |
851 | ||
852 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
853 | { | |
854 | struct page *page; | |
855 | struct kmem_cache_node *n; | |
856 | void *start; | |
857 | void *end; | |
858 | void *last; | |
859 | void *p; | |
860 | ||
861 | if (flags & __GFP_NO_GROW) | |
862 | return NULL; | |
863 | ||
864 | BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK)); | |
865 | ||
866 | if (flags & __GFP_WAIT) | |
867 | local_irq_enable(); | |
868 | ||
869 | page = allocate_slab(s, flags & GFP_LEVEL_MASK, node); | |
870 | if (!page) | |
871 | goto out; | |
872 | ||
873 | n = get_node(s, page_to_nid(page)); | |
874 | if (n) | |
875 | atomic_long_inc(&n->nr_slabs); | |
876 | page->offset = s->offset / sizeof(void *); | |
877 | page->slab = s; | |
878 | page->flags |= 1 << PG_slab; | |
879 | if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | | |
880 | SLAB_STORE_USER | SLAB_TRACE)) | |
881 | page->flags |= 1 << PG_error; | |
882 | ||
883 | start = page_address(page); | |
884 | end = start + s->objects * s->size; | |
885 | ||
886 | if (unlikely(s->flags & SLAB_POISON)) | |
887 | memset(start, POISON_INUSE, PAGE_SIZE << s->order); | |
888 | ||
889 | last = start; | |
890 | for (p = start + s->size; p < end; p += s->size) { | |
891 | setup_object(s, page, last); | |
892 | set_freepointer(s, last, p); | |
893 | last = p; | |
894 | } | |
895 | setup_object(s, page, last); | |
896 | set_freepointer(s, last, NULL); | |
897 | ||
898 | page->freelist = start; | |
899 | page->inuse = 0; | |
900 | out: | |
901 | if (flags & __GFP_WAIT) | |
902 | local_irq_disable(); | |
903 | return page; | |
904 | } | |
905 | ||
906 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
907 | { | |
908 | int pages = 1 << s->order; | |
909 | ||
910 | if (unlikely(PageError(page) || s->dtor)) { | |
911 | void *start = page_address(page); | |
912 | void *end = start + (pages << PAGE_SHIFT); | |
913 | void *p; | |
914 | ||
915 | slab_pad_check(s, page); | |
916 | for (p = start; p <= end - s->size; p += s->size) { | |
917 | if (s->dtor) | |
918 | s->dtor(p, s, 0); | |
919 | check_object(s, page, p, 0); | |
920 | } | |
921 | } | |
922 | ||
923 | mod_zone_page_state(page_zone(page), | |
924 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
925 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
926 | - pages); | |
927 | ||
928 | page->mapping = NULL; | |
929 | __free_pages(page, s->order); | |
930 | } | |
931 | ||
932 | static void rcu_free_slab(struct rcu_head *h) | |
933 | { | |
934 | struct page *page; | |
935 | ||
936 | page = container_of((struct list_head *)h, struct page, lru); | |
937 | __free_slab(page->slab, page); | |
938 | } | |
939 | ||
940 | static void free_slab(struct kmem_cache *s, struct page *page) | |
941 | { | |
942 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
943 | /* | |
944 | * RCU free overloads the RCU head over the LRU | |
945 | */ | |
946 | struct rcu_head *head = (void *)&page->lru; | |
947 | ||
948 | call_rcu(head, rcu_free_slab); | |
949 | } else | |
950 | __free_slab(s, page); | |
951 | } | |
952 | ||
953 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
954 | { | |
955 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
956 | ||
957 | atomic_long_dec(&n->nr_slabs); | |
958 | reset_page_mapcount(page); | |
959 | page->flags &= ~(1 << PG_slab | 1 << PG_error); | |
960 | free_slab(s, page); | |
961 | } | |
962 | ||
963 | /* | |
964 | * Per slab locking using the pagelock | |
965 | */ | |
966 | static __always_inline void slab_lock(struct page *page) | |
967 | { | |
968 | bit_spin_lock(PG_locked, &page->flags); | |
969 | } | |
970 | ||
971 | static __always_inline void slab_unlock(struct page *page) | |
972 | { | |
973 | bit_spin_unlock(PG_locked, &page->flags); | |
974 | } | |
975 | ||
976 | static __always_inline int slab_trylock(struct page *page) | |
977 | { | |
978 | int rc = 1; | |
979 | ||
980 | rc = bit_spin_trylock(PG_locked, &page->flags); | |
981 | return rc; | |
982 | } | |
983 | ||
984 | /* | |
985 | * Management of partially allocated slabs | |
986 | */ | |
e95eed57 | 987 | static void add_partial_tail(struct kmem_cache_node *n, struct page *page) |
81819f0f | 988 | { |
e95eed57 CL |
989 | spin_lock(&n->list_lock); |
990 | n->nr_partial++; | |
991 | list_add_tail(&page->lru, &n->partial); | |
992 | spin_unlock(&n->list_lock); | |
993 | } | |
81819f0f | 994 | |
e95eed57 CL |
995 | static void add_partial(struct kmem_cache_node *n, struct page *page) |
996 | { | |
81819f0f CL |
997 | spin_lock(&n->list_lock); |
998 | n->nr_partial++; | |
999 | list_add(&page->lru, &n->partial); | |
1000 | spin_unlock(&n->list_lock); | |
1001 | } | |
1002 | ||
1003 | static void remove_partial(struct kmem_cache *s, | |
1004 | struct page *page) | |
1005 | { | |
1006 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1007 | ||
1008 | spin_lock(&n->list_lock); | |
1009 | list_del(&page->lru); | |
1010 | n->nr_partial--; | |
1011 | spin_unlock(&n->list_lock); | |
1012 | } | |
1013 | ||
1014 | /* | |
1015 | * Lock page and remove it from the partial list | |
1016 | * | |
1017 | * Must hold list_lock | |
1018 | */ | |
1019 | static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page) | |
1020 | { | |
1021 | if (slab_trylock(page)) { | |
1022 | list_del(&page->lru); | |
1023 | n->nr_partial--; | |
1024 | return 1; | |
1025 | } | |
1026 | return 0; | |
1027 | } | |
1028 | ||
1029 | /* | |
1030 | * Try to get a partial slab from a specific node | |
1031 | */ | |
1032 | static struct page *get_partial_node(struct kmem_cache_node *n) | |
1033 | { | |
1034 | struct page *page; | |
1035 | ||
1036 | /* | |
1037 | * Racy check. If we mistakenly see no partial slabs then we | |
1038 | * just allocate an empty slab. If we mistakenly try to get a | |
1039 | * partial slab then get_partials() will return NULL. | |
1040 | */ | |
1041 | if (!n || !n->nr_partial) | |
1042 | return NULL; | |
1043 | ||
1044 | spin_lock(&n->list_lock); | |
1045 | list_for_each_entry(page, &n->partial, lru) | |
1046 | if (lock_and_del_slab(n, page)) | |
1047 | goto out; | |
1048 | page = NULL; | |
1049 | out: | |
1050 | spin_unlock(&n->list_lock); | |
1051 | return page; | |
1052 | } | |
1053 | ||
1054 | /* | |
1055 | * Get a page from somewhere. Search in increasing NUMA | |
1056 | * distances. | |
1057 | */ | |
1058 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) | |
1059 | { | |
1060 | #ifdef CONFIG_NUMA | |
1061 | struct zonelist *zonelist; | |
1062 | struct zone **z; | |
1063 | struct page *page; | |
1064 | ||
1065 | /* | |
1066 | * The defrag ratio allows to configure the tradeoffs between | |
1067 | * inter node defragmentation and node local allocations. | |
1068 | * A lower defrag_ratio increases the tendency to do local | |
1069 | * allocations instead of scanning throught the partial | |
1070 | * lists on other nodes. | |
1071 | * | |
1072 | * If defrag_ratio is set to 0 then kmalloc() always | |
1073 | * returns node local objects. If its higher then kmalloc() | |
1074 | * may return off node objects in order to avoid fragmentation. | |
1075 | * | |
1076 | * A higher ratio means slabs may be taken from other nodes | |
1077 | * thus reducing the number of partial slabs on those nodes. | |
1078 | * | |
1079 | * If /sys/slab/xx/defrag_ratio is set to 100 (which makes | |
1080 | * defrag_ratio = 1000) then every (well almost) allocation | |
1081 | * will first attempt to defrag slab caches on other nodes. This | |
1082 | * means scanning over all nodes to look for partial slabs which | |
1083 | * may be a bit expensive to do on every slab allocation. | |
1084 | */ | |
1085 | if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio) | |
1086 | return NULL; | |
1087 | ||
1088 | zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
1089 | ->node_zonelists[gfp_zone(flags)]; | |
1090 | for (z = zonelist->zones; *z; z++) { | |
1091 | struct kmem_cache_node *n; | |
1092 | ||
1093 | n = get_node(s, zone_to_nid(*z)); | |
1094 | ||
1095 | if (n && cpuset_zone_allowed_hardwall(*z, flags) && | |
e95eed57 | 1096 | n->nr_partial > MIN_PARTIAL) { |
81819f0f CL |
1097 | page = get_partial_node(n); |
1098 | if (page) | |
1099 | return page; | |
1100 | } | |
1101 | } | |
1102 | #endif | |
1103 | return NULL; | |
1104 | } | |
1105 | ||
1106 | /* | |
1107 | * Get a partial page, lock it and return it. | |
1108 | */ | |
1109 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) | |
1110 | { | |
1111 | struct page *page; | |
1112 | int searchnode = (node == -1) ? numa_node_id() : node; | |
1113 | ||
1114 | page = get_partial_node(get_node(s, searchnode)); | |
1115 | if (page || (flags & __GFP_THISNODE)) | |
1116 | return page; | |
1117 | ||
1118 | return get_any_partial(s, flags); | |
1119 | } | |
1120 | ||
1121 | /* | |
1122 | * Move a page back to the lists. | |
1123 | * | |
1124 | * Must be called with the slab lock held. | |
1125 | * | |
1126 | * On exit the slab lock will have been dropped. | |
1127 | */ | |
1128 | static void putback_slab(struct kmem_cache *s, struct page *page) | |
1129 | { | |
e95eed57 CL |
1130 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1131 | ||
81819f0f | 1132 | if (page->inuse) { |
e95eed57 | 1133 | |
81819f0f | 1134 | if (page->freelist) |
e95eed57 CL |
1135 | add_partial(n, page); |
1136 | else if (PageError(page) && (s->flags & SLAB_STORE_USER)) | |
1137 | add_full(n, page); | |
81819f0f | 1138 | slab_unlock(page); |
e95eed57 | 1139 | |
81819f0f | 1140 | } else { |
e95eed57 CL |
1141 | if (n->nr_partial < MIN_PARTIAL) { |
1142 | /* | |
1143 | * Adding an empty page to the partial slabs in order | |
1144 | * to avoid page allocator overhead. This page needs to | |
1145 | * come after all the others that are not fully empty | |
1146 | * in order to make sure that we do maximum | |
1147 | * defragmentation. | |
1148 | */ | |
1149 | add_partial_tail(n, page); | |
1150 | slab_unlock(page); | |
1151 | } else { | |
1152 | slab_unlock(page); | |
1153 | discard_slab(s, page); | |
1154 | } | |
81819f0f CL |
1155 | } |
1156 | } | |
1157 | ||
1158 | /* | |
1159 | * Remove the cpu slab | |
1160 | */ | |
1161 | static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu) | |
1162 | { | |
1163 | s->cpu_slab[cpu] = NULL; | |
1164 | ClearPageActive(page); | |
1165 | ||
1166 | putback_slab(s, page); | |
1167 | } | |
1168 | ||
1169 | static void flush_slab(struct kmem_cache *s, struct page *page, int cpu) | |
1170 | { | |
1171 | slab_lock(page); | |
1172 | deactivate_slab(s, page, cpu); | |
1173 | } | |
1174 | ||
1175 | /* | |
1176 | * Flush cpu slab. | |
1177 | * Called from IPI handler with interrupts disabled. | |
1178 | */ | |
1179 | static void __flush_cpu_slab(struct kmem_cache *s, int cpu) | |
1180 | { | |
1181 | struct page *page = s->cpu_slab[cpu]; | |
1182 | ||
1183 | if (likely(page)) | |
1184 | flush_slab(s, page, cpu); | |
1185 | } | |
1186 | ||
1187 | static void flush_cpu_slab(void *d) | |
1188 | { | |
1189 | struct kmem_cache *s = d; | |
1190 | int cpu = smp_processor_id(); | |
1191 | ||
1192 | __flush_cpu_slab(s, cpu); | |
1193 | } | |
1194 | ||
1195 | static void flush_all(struct kmem_cache *s) | |
1196 | { | |
1197 | #ifdef CONFIG_SMP | |
1198 | on_each_cpu(flush_cpu_slab, s, 1, 1); | |
1199 | #else | |
1200 | unsigned long flags; | |
1201 | ||
1202 | local_irq_save(flags); | |
1203 | flush_cpu_slab(s); | |
1204 | local_irq_restore(flags); | |
1205 | #endif | |
1206 | } | |
1207 | ||
1208 | /* | |
1209 | * slab_alloc is optimized to only modify two cachelines on the fast path | |
1210 | * (aside from the stack): | |
1211 | * | |
1212 | * 1. The page struct | |
1213 | * 2. The first cacheline of the object to be allocated. | |
1214 | * | |
1215 | * The only cache lines that are read (apart from code) is the | |
1216 | * per cpu array in the kmem_cache struct. | |
1217 | * | |
1218 | * Fastpath is not possible if we need to get a new slab or have | |
1219 | * debugging enabled (which means all slabs are marked with PageError) | |
1220 | */ | |
77c5e2d0 CL |
1221 | static void *slab_alloc(struct kmem_cache *s, |
1222 | gfp_t gfpflags, int node, void *addr) | |
81819f0f CL |
1223 | { |
1224 | struct page *page; | |
1225 | void **object; | |
1226 | unsigned long flags; | |
1227 | int cpu; | |
1228 | ||
1229 | local_irq_save(flags); | |
1230 | cpu = smp_processor_id(); | |
1231 | page = s->cpu_slab[cpu]; | |
1232 | if (!page) | |
1233 | goto new_slab; | |
1234 | ||
1235 | slab_lock(page); | |
1236 | if (unlikely(node != -1 && page_to_nid(page) != node)) | |
1237 | goto another_slab; | |
1238 | redo: | |
1239 | object = page->freelist; | |
1240 | if (unlikely(!object)) | |
1241 | goto another_slab; | |
1242 | if (unlikely(PageError(page))) | |
1243 | goto debug; | |
1244 | ||
1245 | have_object: | |
1246 | page->inuse++; | |
1247 | page->freelist = object[page->offset]; | |
1248 | slab_unlock(page); | |
1249 | local_irq_restore(flags); | |
1250 | return object; | |
1251 | ||
1252 | another_slab: | |
1253 | deactivate_slab(s, page, cpu); | |
1254 | ||
1255 | new_slab: | |
1256 | page = get_partial(s, gfpflags, node); | |
1257 | if (likely(page)) { | |
1258 | have_slab: | |
1259 | s->cpu_slab[cpu] = page; | |
1260 | SetPageActive(page); | |
1261 | goto redo; | |
1262 | } | |
1263 | ||
1264 | page = new_slab(s, gfpflags, node); | |
1265 | if (page) { | |
1266 | cpu = smp_processor_id(); | |
1267 | if (s->cpu_slab[cpu]) { | |
1268 | /* | |
1269 | * Someone else populated the cpu_slab while we enabled | |
1270 | * interrupts, or we have got scheduled on another cpu. | |
1271 | * The page may not be on the requested node. | |
1272 | */ | |
1273 | if (node == -1 || | |
1274 | page_to_nid(s->cpu_slab[cpu]) == node) { | |
1275 | /* | |
1276 | * Current cpuslab is acceptable and we | |
1277 | * want the current one since its cache hot | |
1278 | */ | |
1279 | discard_slab(s, page); | |
1280 | page = s->cpu_slab[cpu]; | |
1281 | slab_lock(page); | |
1282 | goto redo; | |
1283 | } | |
1284 | /* Dump the current slab */ | |
1285 | flush_slab(s, s->cpu_slab[cpu], cpu); | |
1286 | } | |
1287 | slab_lock(page); | |
1288 | goto have_slab; | |
1289 | } | |
1290 | local_irq_restore(flags); | |
1291 | return NULL; | |
1292 | debug: | |
1293 | if (!alloc_object_checks(s, page, object)) | |
1294 | goto another_slab; | |
1295 | if (s->flags & SLAB_STORE_USER) | |
77c5e2d0 | 1296 | set_track(s, object, TRACK_ALLOC, addr); |
81819f0f CL |
1297 | goto have_object; |
1298 | } | |
1299 | ||
1300 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
1301 | { | |
77c5e2d0 | 1302 | return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); |
81819f0f CL |
1303 | } |
1304 | EXPORT_SYMBOL(kmem_cache_alloc); | |
1305 | ||
1306 | #ifdef CONFIG_NUMA | |
1307 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
1308 | { | |
77c5e2d0 | 1309 | return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); |
81819f0f CL |
1310 | } |
1311 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
1312 | #endif | |
1313 | ||
1314 | /* | |
1315 | * The fastpath only writes the cacheline of the page struct and the first | |
1316 | * cacheline of the object. | |
1317 | * | |
1318 | * No special cachelines need to be read | |
1319 | */ | |
77c5e2d0 CL |
1320 | static void slab_free(struct kmem_cache *s, struct page *page, |
1321 | void *x, void *addr) | |
81819f0f CL |
1322 | { |
1323 | void *prior; | |
1324 | void **object = (void *)x; | |
1325 | unsigned long flags; | |
1326 | ||
1327 | local_irq_save(flags); | |
1328 | slab_lock(page); | |
1329 | ||
1330 | if (unlikely(PageError(page))) | |
1331 | goto debug; | |
1332 | checks_ok: | |
1333 | prior = object[page->offset] = page->freelist; | |
1334 | page->freelist = object; | |
1335 | page->inuse--; | |
1336 | ||
1337 | if (unlikely(PageActive(page))) | |
1338 | /* | |
1339 | * Cpu slabs are never on partial lists and are | |
1340 | * never freed. | |
1341 | */ | |
1342 | goto out_unlock; | |
1343 | ||
1344 | if (unlikely(!page->inuse)) | |
1345 | goto slab_empty; | |
1346 | ||
1347 | /* | |
1348 | * Objects left in the slab. If it | |
1349 | * was not on the partial list before | |
1350 | * then add it. | |
1351 | */ | |
1352 | if (unlikely(!prior)) | |
e95eed57 | 1353 | add_partial(get_node(s, page_to_nid(page)), page); |
81819f0f CL |
1354 | |
1355 | out_unlock: | |
1356 | slab_unlock(page); | |
1357 | local_irq_restore(flags); | |
1358 | return; | |
1359 | ||
1360 | slab_empty: | |
1361 | if (prior) | |
1362 | /* | |
643b1138 | 1363 | * Slab on the partial list. |
81819f0f CL |
1364 | */ |
1365 | remove_partial(s, page); | |
1366 | ||
1367 | slab_unlock(page); | |
1368 | discard_slab(s, page); | |
1369 | local_irq_restore(flags); | |
1370 | return; | |
1371 | ||
1372 | debug: | |
77c5e2d0 CL |
1373 | if (!free_object_checks(s, page, x)) |
1374 | goto out_unlock; | |
643b1138 CL |
1375 | if (!PageActive(page) && !page->freelist) |
1376 | remove_full(s, page); | |
77c5e2d0 CL |
1377 | if (s->flags & SLAB_STORE_USER) |
1378 | set_track(s, x, TRACK_FREE, addr); | |
1379 | goto checks_ok; | |
81819f0f CL |
1380 | } |
1381 | ||
1382 | void kmem_cache_free(struct kmem_cache *s, void *x) | |
1383 | { | |
77c5e2d0 | 1384 | struct page *page; |
81819f0f | 1385 | |
b49af68f | 1386 | page = virt_to_head_page(x); |
81819f0f | 1387 | |
77c5e2d0 | 1388 | slab_free(s, page, x, __builtin_return_address(0)); |
81819f0f CL |
1389 | } |
1390 | EXPORT_SYMBOL(kmem_cache_free); | |
1391 | ||
1392 | /* Figure out on which slab object the object resides */ | |
1393 | static struct page *get_object_page(const void *x) | |
1394 | { | |
b49af68f | 1395 | struct page *page = virt_to_head_page(x); |
81819f0f CL |
1396 | |
1397 | if (!PageSlab(page)) | |
1398 | return NULL; | |
1399 | ||
1400 | return page; | |
1401 | } | |
1402 | ||
1403 | /* | |
1404 | * kmem_cache_open produces objects aligned at "size" and the first object | |
1405 | * is placed at offset 0 in the slab (We have no metainformation on the | |
1406 | * slab, all slabs are in essence "off slab"). | |
1407 | * | |
1408 | * In order to get the desired alignment one just needs to align the | |
1409 | * size. | |
1410 | * | |
1411 | * Notice that the allocation order determines the sizes of the per cpu | |
1412 | * caches. Each processor has always one slab available for allocations. | |
1413 | * Increasing the allocation order reduces the number of times that slabs | |
1414 | * must be moved on and off the partial lists and therefore may influence | |
1415 | * locking overhead. | |
1416 | * | |
1417 | * The offset is used to relocate the free list link in each object. It is | |
1418 | * therefore possible to move the free list link behind the object. This | |
1419 | * is necessary for RCU to work properly and also useful for debugging. | |
1420 | */ | |
1421 | ||
1422 | /* | |
1423 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
1424 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
1425 | * and increases the number of allocations possible without having to | |
1426 | * take the list_lock. | |
1427 | */ | |
1428 | static int slub_min_order; | |
1429 | static int slub_max_order = DEFAULT_MAX_ORDER; | |
1430 | ||
1431 | /* | |
1432 | * Minimum number of objects per slab. This is necessary in order to | |
1433 | * reduce locking overhead. Similar to the queue size in SLAB. | |
1434 | */ | |
1435 | static int slub_min_objects = DEFAULT_MIN_OBJECTS; | |
1436 | ||
1437 | /* | |
1438 | * Merge control. If this is set then no merging of slab caches will occur. | |
1439 | */ | |
1440 | static int slub_nomerge; | |
1441 | ||
1442 | /* | |
1443 | * Debug settings: | |
1444 | */ | |
1445 | static int slub_debug; | |
1446 | ||
1447 | static char *slub_debug_slabs; | |
1448 | ||
1449 | /* | |
1450 | * Calculate the order of allocation given an slab object size. | |
1451 | * | |
1452 | * The order of allocation has significant impact on other elements | |
1453 | * of the system. Generally order 0 allocations should be preferred | |
1454 | * since they do not cause fragmentation in the page allocator. Larger | |
1455 | * objects may have problems with order 0 because there may be too much | |
1456 | * space left unused in a slab. We go to a higher order if more than 1/8th | |
1457 | * of the slab would be wasted. | |
1458 | * | |
1459 | * In order to reach satisfactory performance we must ensure that | |
1460 | * a minimum number of objects is in one slab. Otherwise we may | |
1461 | * generate too much activity on the partial lists. This is less a | |
1462 | * concern for large slabs though. slub_max_order specifies the order | |
1463 | * where we begin to stop considering the number of objects in a slab. | |
1464 | * | |
1465 | * Higher order allocations also allow the placement of more objects | |
1466 | * in a slab and thereby reduce object handling overhead. If the user | |
1467 | * has requested a higher mininum order then we start with that one | |
1468 | * instead of zero. | |
1469 | */ | |
1470 | static int calculate_order(int size) | |
1471 | { | |
1472 | int order; | |
1473 | int rem; | |
1474 | ||
1475 | for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT); | |
1476 | order < MAX_ORDER; order++) { | |
1477 | unsigned long slab_size = PAGE_SIZE << order; | |
1478 | ||
1479 | if (slub_max_order > order && | |
1480 | slab_size < slub_min_objects * size) | |
1481 | continue; | |
1482 | ||
1483 | if (slab_size < size) | |
1484 | continue; | |
1485 | ||
1486 | rem = slab_size % size; | |
1487 | ||
1488 | if (rem <= (PAGE_SIZE << order) / 8) | |
1489 | break; | |
1490 | ||
1491 | } | |
1492 | if (order >= MAX_ORDER) | |
1493 | return -E2BIG; | |
1494 | return order; | |
1495 | } | |
1496 | ||
1497 | /* | |
1498 | * Function to figure out which alignment to use from the | |
1499 | * various ways of specifying it. | |
1500 | */ | |
1501 | static unsigned long calculate_alignment(unsigned long flags, | |
1502 | unsigned long align, unsigned long size) | |
1503 | { | |
1504 | /* | |
1505 | * If the user wants hardware cache aligned objects then | |
1506 | * follow that suggestion if the object is sufficiently | |
1507 | * large. | |
1508 | * | |
1509 | * The hardware cache alignment cannot override the | |
1510 | * specified alignment though. If that is greater | |
1511 | * then use it. | |
1512 | */ | |
1513 | if ((flags & (SLAB_MUST_HWCACHE_ALIGN | SLAB_HWCACHE_ALIGN)) && | |
1514 | size > L1_CACHE_BYTES / 2) | |
1515 | return max_t(unsigned long, align, L1_CACHE_BYTES); | |
1516 | ||
1517 | if (align < ARCH_SLAB_MINALIGN) | |
1518 | return ARCH_SLAB_MINALIGN; | |
1519 | ||
1520 | return ALIGN(align, sizeof(void *)); | |
1521 | } | |
1522 | ||
1523 | static void init_kmem_cache_node(struct kmem_cache_node *n) | |
1524 | { | |
1525 | n->nr_partial = 0; | |
1526 | atomic_long_set(&n->nr_slabs, 0); | |
1527 | spin_lock_init(&n->list_lock); | |
1528 | INIT_LIST_HEAD(&n->partial); | |
643b1138 | 1529 | INIT_LIST_HEAD(&n->full); |
81819f0f CL |
1530 | } |
1531 | ||
1532 | #ifdef CONFIG_NUMA | |
1533 | /* | |
1534 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
1535 | * slab on the node for this slabcache. There are no concurrent accesses | |
1536 | * possible. | |
1537 | * | |
1538 | * Note that this function only works on the kmalloc_node_cache | |
1539 | * when allocating for the kmalloc_node_cache. | |
1540 | */ | |
1541 | static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags, | |
1542 | int node) | |
1543 | { | |
1544 | struct page *page; | |
1545 | struct kmem_cache_node *n; | |
1546 | ||
1547 | BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); | |
1548 | ||
1549 | page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node); | |
1550 | /* new_slab() disables interupts */ | |
1551 | local_irq_enable(); | |
1552 | ||
1553 | BUG_ON(!page); | |
1554 | n = page->freelist; | |
1555 | BUG_ON(!n); | |
1556 | page->freelist = get_freepointer(kmalloc_caches, n); | |
1557 | page->inuse++; | |
1558 | kmalloc_caches->node[node] = n; | |
1559 | init_object(kmalloc_caches, n, 1); | |
1560 | init_kmem_cache_node(n); | |
1561 | atomic_long_inc(&n->nr_slabs); | |
e95eed57 | 1562 | add_partial(n, page); |
81819f0f CL |
1563 | return n; |
1564 | } | |
1565 | ||
1566 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
1567 | { | |
1568 | int node; | |
1569 | ||
1570 | for_each_online_node(node) { | |
1571 | struct kmem_cache_node *n = s->node[node]; | |
1572 | if (n && n != &s->local_node) | |
1573 | kmem_cache_free(kmalloc_caches, n); | |
1574 | s->node[node] = NULL; | |
1575 | } | |
1576 | } | |
1577 | ||
1578 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
1579 | { | |
1580 | int node; | |
1581 | int local_node; | |
1582 | ||
1583 | if (slab_state >= UP) | |
1584 | local_node = page_to_nid(virt_to_page(s)); | |
1585 | else | |
1586 | local_node = 0; | |
1587 | ||
1588 | for_each_online_node(node) { | |
1589 | struct kmem_cache_node *n; | |
1590 | ||
1591 | if (local_node == node) | |
1592 | n = &s->local_node; | |
1593 | else { | |
1594 | if (slab_state == DOWN) { | |
1595 | n = early_kmem_cache_node_alloc(gfpflags, | |
1596 | node); | |
1597 | continue; | |
1598 | } | |
1599 | n = kmem_cache_alloc_node(kmalloc_caches, | |
1600 | gfpflags, node); | |
1601 | ||
1602 | if (!n) { | |
1603 | free_kmem_cache_nodes(s); | |
1604 | return 0; | |
1605 | } | |
1606 | ||
1607 | } | |
1608 | s->node[node] = n; | |
1609 | init_kmem_cache_node(n); | |
1610 | } | |
1611 | return 1; | |
1612 | } | |
1613 | #else | |
1614 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
1615 | { | |
1616 | } | |
1617 | ||
1618 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
1619 | { | |
1620 | init_kmem_cache_node(&s->local_node); | |
1621 | return 1; | |
1622 | } | |
1623 | #endif | |
1624 | ||
1625 | /* | |
1626 | * calculate_sizes() determines the order and the distribution of data within | |
1627 | * a slab object. | |
1628 | */ | |
1629 | static int calculate_sizes(struct kmem_cache *s) | |
1630 | { | |
1631 | unsigned long flags = s->flags; | |
1632 | unsigned long size = s->objsize; | |
1633 | unsigned long align = s->align; | |
1634 | ||
1635 | /* | |
1636 | * Determine if we can poison the object itself. If the user of | |
1637 | * the slab may touch the object after free or before allocation | |
1638 | * then we should never poison the object itself. | |
1639 | */ | |
1640 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
1641 | !s->ctor && !s->dtor) | |
1642 | s->flags |= __OBJECT_POISON; | |
1643 | else | |
1644 | s->flags &= ~__OBJECT_POISON; | |
1645 | ||
1646 | /* | |
1647 | * Round up object size to the next word boundary. We can only | |
1648 | * place the free pointer at word boundaries and this determines | |
1649 | * the possible location of the free pointer. | |
1650 | */ | |
1651 | size = ALIGN(size, sizeof(void *)); | |
1652 | ||
1653 | /* | |
1654 | * If we are redzoning then check if there is some space between the | |
1655 | * end of the object and the free pointer. If not then add an | |
1656 | * additional word, so that we can establish a redzone between | |
1657 | * the object and the freepointer to be able to check for overwrites. | |
1658 | */ | |
1659 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
1660 | size += sizeof(void *); | |
1661 | ||
1662 | /* | |
1663 | * With that we have determined how much of the slab is in actual | |
1664 | * use by the object. This is the potential offset to the free | |
1665 | * pointer. | |
1666 | */ | |
1667 | s->inuse = size; | |
1668 | ||
1669 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
1670 | s->ctor || s->dtor)) { | |
1671 | /* | |
1672 | * Relocate free pointer after the object if it is not | |
1673 | * permitted to overwrite the first word of the object on | |
1674 | * kmem_cache_free. | |
1675 | * | |
1676 | * This is the case if we do RCU, have a constructor or | |
1677 | * destructor or are poisoning the objects. | |
1678 | */ | |
1679 | s->offset = size; | |
1680 | size += sizeof(void *); | |
1681 | } | |
1682 | ||
1683 | if (flags & SLAB_STORE_USER) | |
1684 | /* | |
1685 | * Need to store information about allocs and frees after | |
1686 | * the object. | |
1687 | */ | |
1688 | size += 2 * sizeof(struct track); | |
1689 | ||
1690 | if (flags & DEBUG_DEFAULT_FLAGS) | |
1691 | /* | |
1692 | * Add some empty padding so that we can catch | |
1693 | * overwrites from earlier objects rather than let | |
1694 | * tracking information or the free pointer be | |
1695 | * corrupted if an user writes before the start | |
1696 | * of the object. | |
1697 | */ | |
1698 | size += sizeof(void *); | |
1699 | /* | |
1700 | * Determine the alignment based on various parameters that the | |
1701 | * user specified (this is unecessarily complex due to the attempt | |
1702 | * to be compatible with SLAB. Should be cleaned up some day). | |
1703 | */ | |
1704 | align = calculate_alignment(flags, align, s->objsize); | |
1705 | ||
1706 | /* | |
1707 | * SLUB stores one object immediately after another beginning from | |
1708 | * offset 0. In order to align the objects we have to simply size | |
1709 | * each object to conform to the alignment. | |
1710 | */ | |
1711 | size = ALIGN(size, align); | |
1712 | s->size = size; | |
1713 | ||
1714 | s->order = calculate_order(size); | |
1715 | if (s->order < 0) | |
1716 | return 0; | |
1717 | ||
1718 | /* | |
1719 | * Determine the number of objects per slab | |
1720 | */ | |
1721 | s->objects = (PAGE_SIZE << s->order) / size; | |
1722 | ||
1723 | /* | |
1724 | * Verify that the number of objects is within permitted limits. | |
1725 | * The page->inuse field is only 16 bit wide! So we cannot have | |
1726 | * more than 64k objects per slab. | |
1727 | */ | |
1728 | if (!s->objects || s->objects > 65535) | |
1729 | return 0; | |
1730 | return 1; | |
1731 | ||
1732 | } | |
1733 | ||
1734 | static int __init finish_bootstrap(void) | |
1735 | { | |
1736 | struct list_head *h; | |
1737 | int err; | |
1738 | ||
1739 | slab_state = SYSFS; | |
1740 | ||
1741 | list_for_each(h, &slab_caches) { | |
1742 | struct kmem_cache *s = | |
1743 | container_of(h, struct kmem_cache, list); | |
1744 | ||
1745 | err = sysfs_slab_add(s); | |
1746 | BUG_ON(err); | |
1747 | } | |
1748 | return 0; | |
1749 | } | |
1750 | ||
1751 | static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, | |
1752 | const char *name, size_t size, | |
1753 | size_t align, unsigned long flags, | |
1754 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
1755 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
1756 | { | |
1757 | memset(s, 0, kmem_size); | |
1758 | s->name = name; | |
1759 | s->ctor = ctor; | |
1760 | s->dtor = dtor; | |
1761 | s->objsize = size; | |
1762 | s->flags = flags; | |
1763 | s->align = align; | |
1764 | ||
1765 | BUG_ON(flags & SLUB_UNIMPLEMENTED); | |
1766 | ||
1767 | /* | |
1768 | * The page->offset field is only 16 bit wide. This is an offset | |
1769 | * in units of words from the beginning of an object. If the slab | |
1770 | * size is bigger then we cannot move the free pointer behind the | |
1771 | * object anymore. | |
1772 | * | |
1773 | * On 32 bit platforms the limit is 256k. On 64bit platforms | |
1774 | * the limit is 512k. | |
1775 | * | |
1776 | * Debugging or ctor/dtors may create a need to move the free | |
1777 | * pointer. Fail if this happens. | |
1778 | */ | |
1779 | if (s->size >= 65535 * sizeof(void *)) { | |
1780 | BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON | | |
1781 | SLAB_STORE_USER | SLAB_DESTROY_BY_RCU)); | |
1782 | BUG_ON(ctor || dtor); | |
1783 | } | |
1784 | else | |
1785 | /* | |
1786 | * Enable debugging if selected on the kernel commandline. | |
1787 | */ | |
1788 | if (slub_debug && (!slub_debug_slabs || | |
1789 | strncmp(slub_debug_slabs, name, | |
1790 | strlen(slub_debug_slabs)) == 0)) | |
1791 | s->flags |= slub_debug; | |
1792 | ||
1793 | if (!calculate_sizes(s)) | |
1794 | goto error; | |
1795 | ||
1796 | s->refcount = 1; | |
1797 | #ifdef CONFIG_NUMA | |
1798 | s->defrag_ratio = 100; | |
1799 | #endif | |
1800 | ||
1801 | if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) | |
1802 | return 1; | |
1803 | error: | |
1804 | if (flags & SLAB_PANIC) | |
1805 | panic("Cannot create slab %s size=%lu realsize=%u " | |
1806 | "order=%u offset=%u flags=%lx\n", | |
1807 | s->name, (unsigned long)size, s->size, s->order, | |
1808 | s->offset, flags); | |
1809 | return 0; | |
1810 | } | |
1811 | EXPORT_SYMBOL(kmem_cache_open); | |
1812 | ||
1813 | /* | |
1814 | * Check if a given pointer is valid | |
1815 | */ | |
1816 | int kmem_ptr_validate(struct kmem_cache *s, const void *object) | |
1817 | { | |
1818 | struct page * page; | |
1819 | void *addr; | |
1820 | ||
1821 | page = get_object_page(object); | |
1822 | ||
1823 | if (!page || s != page->slab) | |
1824 | /* No slab or wrong slab */ | |
1825 | return 0; | |
1826 | ||
1827 | addr = page_address(page); | |
1828 | if (object < addr || object >= addr + s->objects * s->size) | |
1829 | /* Out of bounds */ | |
1830 | return 0; | |
1831 | ||
1832 | if ((object - addr) % s->size) | |
1833 | /* Improperly aligned */ | |
1834 | return 0; | |
1835 | ||
1836 | /* | |
1837 | * We could also check if the object is on the slabs freelist. | |
1838 | * But this would be too expensive and it seems that the main | |
1839 | * purpose of kmem_ptr_valid is to check if the object belongs | |
1840 | * to a certain slab. | |
1841 | */ | |
1842 | return 1; | |
1843 | } | |
1844 | EXPORT_SYMBOL(kmem_ptr_validate); | |
1845 | ||
1846 | /* | |
1847 | * Determine the size of a slab object | |
1848 | */ | |
1849 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
1850 | { | |
1851 | return s->objsize; | |
1852 | } | |
1853 | EXPORT_SYMBOL(kmem_cache_size); | |
1854 | ||
1855 | const char *kmem_cache_name(struct kmem_cache *s) | |
1856 | { | |
1857 | return s->name; | |
1858 | } | |
1859 | EXPORT_SYMBOL(kmem_cache_name); | |
1860 | ||
1861 | /* | |
1862 | * Attempt to free all slabs on a node | |
1863 | */ | |
1864 | static int free_list(struct kmem_cache *s, struct kmem_cache_node *n, | |
1865 | struct list_head *list) | |
1866 | { | |
1867 | int slabs_inuse = 0; | |
1868 | unsigned long flags; | |
1869 | struct page *page, *h; | |
1870 | ||
1871 | spin_lock_irqsave(&n->list_lock, flags); | |
1872 | list_for_each_entry_safe(page, h, list, lru) | |
1873 | if (!page->inuse) { | |
1874 | list_del(&page->lru); | |
1875 | discard_slab(s, page); | |
1876 | } else | |
1877 | slabs_inuse++; | |
1878 | spin_unlock_irqrestore(&n->list_lock, flags); | |
1879 | return slabs_inuse; | |
1880 | } | |
1881 | ||
1882 | /* | |
1883 | * Release all resources used by slab cache | |
1884 | */ | |
1885 | static int kmem_cache_close(struct kmem_cache *s) | |
1886 | { | |
1887 | int node; | |
1888 | ||
1889 | flush_all(s); | |
1890 | ||
1891 | /* Attempt to free all objects */ | |
1892 | for_each_online_node(node) { | |
1893 | struct kmem_cache_node *n = get_node(s, node); | |
1894 | ||
2086d26a | 1895 | n->nr_partial -= free_list(s, n, &n->partial); |
81819f0f CL |
1896 | if (atomic_long_read(&n->nr_slabs)) |
1897 | return 1; | |
1898 | } | |
1899 | free_kmem_cache_nodes(s); | |
1900 | return 0; | |
1901 | } | |
1902 | ||
1903 | /* | |
1904 | * Close a cache and release the kmem_cache structure | |
1905 | * (must be used for caches created using kmem_cache_create) | |
1906 | */ | |
1907 | void kmem_cache_destroy(struct kmem_cache *s) | |
1908 | { | |
1909 | down_write(&slub_lock); | |
1910 | s->refcount--; | |
1911 | if (!s->refcount) { | |
1912 | list_del(&s->list); | |
1913 | if (kmem_cache_close(s)) | |
1914 | WARN_ON(1); | |
1915 | sysfs_slab_remove(s); | |
1916 | kfree(s); | |
1917 | } | |
1918 | up_write(&slub_lock); | |
1919 | } | |
1920 | EXPORT_SYMBOL(kmem_cache_destroy); | |
1921 | ||
1922 | /******************************************************************** | |
1923 | * Kmalloc subsystem | |
1924 | *******************************************************************/ | |
1925 | ||
1926 | struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned; | |
1927 | EXPORT_SYMBOL(kmalloc_caches); | |
1928 | ||
1929 | #ifdef CONFIG_ZONE_DMA | |
1930 | static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1]; | |
1931 | #endif | |
1932 | ||
1933 | static int __init setup_slub_min_order(char *str) | |
1934 | { | |
1935 | get_option (&str, &slub_min_order); | |
1936 | ||
1937 | return 1; | |
1938 | } | |
1939 | ||
1940 | __setup("slub_min_order=", setup_slub_min_order); | |
1941 | ||
1942 | static int __init setup_slub_max_order(char *str) | |
1943 | { | |
1944 | get_option (&str, &slub_max_order); | |
1945 | ||
1946 | return 1; | |
1947 | } | |
1948 | ||
1949 | __setup("slub_max_order=", setup_slub_max_order); | |
1950 | ||
1951 | static int __init setup_slub_min_objects(char *str) | |
1952 | { | |
1953 | get_option (&str, &slub_min_objects); | |
1954 | ||
1955 | return 1; | |
1956 | } | |
1957 | ||
1958 | __setup("slub_min_objects=", setup_slub_min_objects); | |
1959 | ||
1960 | static int __init setup_slub_nomerge(char *str) | |
1961 | { | |
1962 | slub_nomerge = 1; | |
1963 | return 1; | |
1964 | } | |
1965 | ||
1966 | __setup("slub_nomerge", setup_slub_nomerge); | |
1967 | ||
1968 | static int __init setup_slub_debug(char *str) | |
1969 | { | |
1970 | if (!str || *str != '=') | |
1971 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1972 | else { | |
1973 | str++; | |
1974 | if (*str == 0 || *str == ',') | |
1975 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1976 | else | |
1977 | for( ;*str && *str != ','; str++) | |
1978 | switch (*str) { | |
1979 | case 'f' : case 'F' : | |
1980 | slub_debug |= SLAB_DEBUG_FREE; | |
1981 | break; | |
1982 | case 'z' : case 'Z' : | |
1983 | slub_debug |= SLAB_RED_ZONE; | |
1984 | break; | |
1985 | case 'p' : case 'P' : | |
1986 | slub_debug |= SLAB_POISON; | |
1987 | break; | |
1988 | case 'u' : case 'U' : | |
1989 | slub_debug |= SLAB_STORE_USER; | |
1990 | break; | |
1991 | case 't' : case 'T' : | |
1992 | slub_debug |= SLAB_TRACE; | |
1993 | break; | |
1994 | default: | |
1995 | printk(KERN_ERR "slub_debug option '%c' " | |
1996 | "unknown. skipped\n",*str); | |
1997 | } | |
1998 | } | |
1999 | ||
2000 | if (*str == ',') | |
2001 | slub_debug_slabs = str + 1; | |
2002 | return 1; | |
2003 | } | |
2004 | ||
2005 | __setup("slub_debug", setup_slub_debug); | |
2006 | ||
2007 | static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, | |
2008 | const char *name, int size, gfp_t gfp_flags) | |
2009 | { | |
2010 | unsigned int flags = 0; | |
2011 | ||
2012 | if (gfp_flags & SLUB_DMA) | |
2013 | flags = SLAB_CACHE_DMA; | |
2014 | ||
2015 | down_write(&slub_lock); | |
2016 | if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, | |
2017 | flags, NULL, NULL)) | |
2018 | goto panic; | |
2019 | ||
2020 | list_add(&s->list, &slab_caches); | |
2021 | up_write(&slub_lock); | |
2022 | if (sysfs_slab_add(s)) | |
2023 | goto panic; | |
2024 | return s; | |
2025 | ||
2026 | panic: | |
2027 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
2028 | } | |
2029 | ||
2030 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) | |
2031 | { | |
2032 | int index = kmalloc_index(size); | |
2033 | ||
614410d5 | 2034 | if (!index) |
81819f0f CL |
2035 | return NULL; |
2036 | ||
2037 | /* Allocation too large? */ | |
2038 | BUG_ON(index < 0); | |
2039 | ||
2040 | #ifdef CONFIG_ZONE_DMA | |
2041 | if ((flags & SLUB_DMA)) { | |
2042 | struct kmem_cache *s; | |
2043 | struct kmem_cache *x; | |
2044 | char *text; | |
2045 | size_t realsize; | |
2046 | ||
2047 | s = kmalloc_caches_dma[index]; | |
2048 | if (s) | |
2049 | return s; | |
2050 | ||
2051 | /* Dynamically create dma cache */ | |
2052 | x = kmalloc(kmem_size, flags & ~SLUB_DMA); | |
2053 | if (!x) | |
2054 | panic("Unable to allocate memory for dma cache\n"); | |
2055 | ||
2056 | if (index <= KMALLOC_SHIFT_HIGH) | |
2057 | realsize = 1 << index; | |
2058 | else { | |
2059 | if (index == 1) | |
2060 | realsize = 96; | |
2061 | else | |
2062 | realsize = 192; | |
2063 | } | |
2064 | ||
2065 | text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", | |
2066 | (unsigned int)realsize); | |
2067 | s = create_kmalloc_cache(x, text, realsize, flags); | |
2068 | kmalloc_caches_dma[index] = s; | |
2069 | return s; | |
2070 | } | |
2071 | #endif | |
2072 | return &kmalloc_caches[index]; | |
2073 | } | |
2074 | ||
2075 | void *__kmalloc(size_t size, gfp_t flags) | |
2076 | { | |
2077 | struct kmem_cache *s = get_slab(size, flags); | |
2078 | ||
2079 | if (s) | |
77c5e2d0 | 2080 | return slab_alloc(s, flags, -1, __builtin_return_address(0)); |
81819f0f CL |
2081 | return NULL; |
2082 | } | |
2083 | EXPORT_SYMBOL(__kmalloc); | |
2084 | ||
2085 | #ifdef CONFIG_NUMA | |
2086 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
2087 | { | |
2088 | struct kmem_cache *s = get_slab(size, flags); | |
2089 | ||
2090 | if (s) | |
77c5e2d0 | 2091 | return slab_alloc(s, flags, node, __builtin_return_address(0)); |
81819f0f CL |
2092 | return NULL; |
2093 | } | |
2094 | EXPORT_SYMBOL(__kmalloc_node); | |
2095 | #endif | |
2096 | ||
2097 | size_t ksize(const void *object) | |
2098 | { | |
2099 | struct page *page = get_object_page(object); | |
2100 | struct kmem_cache *s; | |
2101 | ||
2102 | BUG_ON(!page); | |
2103 | s = page->slab; | |
2104 | BUG_ON(!s); | |
2105 | ||
2106 | /* | |
2107 | * Debugging requires use of the padding between object | |
2108 | * and whatever may come after it. | |
2109 | */ | |
2110 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
2111 | return s->objsize; | |
2112 | ||
2113 | /* | |
2114 | * If we have the need to store the freelist pointer | |
2115 | * back there or track user information then we can | |
2116 | * only use the space before that information. | |
2117 | */ | |
2118 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
2119 | return s->inuse; | |
2120 | ||
2121 | /* | |
2122 | * Else we can use all the padding etc for the allocation | |
2123 | */ | |
2124 | return s->size; | |
2125 | } | |
2126 | EXPORT_SYMBOL(ksize); | |
2127 | ||
2128 | void kfree(const void *x) | |
2129 | { | |
2130 | struct kmem_cache *s; | |
2131 | struct page *page; | |
2132 | ||
2133 | if (!x) | |
2134 | return; | |
2135 | ||
b49af68f | 2136 | page = virt_to_head_page(x); |
81819f0f CL |
2137 | s = page->slab; |
2138 | ||
77c5e2d0 | 2139 | slab_free(s, page, (void *)x, __builtin_return_address(0)); |
81819f0f CL |
2140 | } |
2141 | EXPORT_SYMBOL(kfree); | |
2142 | ||
2086d26a CL |
2143 | /* |
2144 | * kmem_cache_shrink removes empty slabs from the partial lists | |
2145 | * and then sorts the partially allocated slabs by the number | |
2146 | * of items in use. The slabs with the most items in use | |
2147 | * come first. New allocations will remove these from the | |
2148 | * partial list because they are full. The slabs with the | |
2149 | * least items are placed last. If it happens that the objects | |
2150 | * are freed then the page can be returned to the page allocator. | |
2151 | */ | |
2152 | int kmem_cache_shrink(struct kmem_cache *s) | |
2153 | { | |
2154 | int node; | |
2155 | int i; | |
2156 | struct kmem_cache_node *n; | |
2157 | struct page *page; | |
2158 | struct page *t; | |
2159 | struct list_head *slabs_by_inuse = | |
2160 | kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL); | |
2161 | unsigned long flags; | |
2162 | ||
2163 | if (!slabs_by_inuse) | |
2164 | return -ENOMEM; | |
2165 | ||
2166 | flush_all(s); | |
2167 | for_each_online_node(node) { | |
2168 | n = get_node(s, node); | |
2169 | ||
2170 | if (!n->nr_partial) | |
2171 | continue; | |
2172 | ||
2173 | for (i = 0; i < s->objects; i++) | |
2174 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
2175 | ||
2176 | spin_lock_irqsave(&n->list_lock, flags); | |
2177 | ||
2178 | /* | |
2179 | * Build lists indexed by the items in use in | |
2180 | * each slab or free slabs if empty. | |
2181 | * | |
2182 | * Note that concurrent frees may occur while | |
2183 | * we hold the list_lock. page->inuse here is | |
2184 | * the upper limit. | |
2185 | */ | |
2186 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
2187 | if (!page->inuse && slab_trylock(page)) { | |
2188 | /* | |
2189 | * Must hold slab lock here because slab_free | |
2190 | * may have freed the last object and be | |
2191 | * waiting to release the slab. | |
2192 | */ | |
2193 | list_del(&page->lru); | |
2194 | n->nr_partial--; | |
2195 | slab_unlock(page); | |
2196 | discard_slab(s, page); | |
2197 | } else { | |
2198 | if (n->nr_partial > MAX_PARTIAL) | |
2199 | list_move(&page->lru, | |
2200 | slabs_by_inuse + page->inuse); | |
2201 | } | |
2202 | } | |
2203 | ||
2204 | if (n->nr_partial <= MAX_PARTIAL) | |
2205 | goto out; | |
2206 | ||
2207 | /* | |
2208 | * Rebuild the partial list with the slabs filled up | |
2209 | * most first and the least used slabs at the end. | |
2210 | */ | |
2211 | for (i = s->objects - 1; i >= 0; i--) | |
2212 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
2213 | ||
2214 | out: | |
2215 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2216 | } | |
2217 | ||
2218 | kfree(slabs_by_inuse); | |
2219 | return 0; | |
2220 | } | |
2221 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2222 | ||
81819f0f CL |
2223 | /** |
2224 | * krealloc - reallocate memory. The contents will remain unchanged. | |
2225 | * | |
2226 | * @p: object to reallocate memory for. | |
2227 | * @new_size: how many bytes of memory are required. | |
2228 | * @flags: the type of memory to allocate. | |
2229 | * | |
2230 | * The contents of the object pointed to are preserved up to the | |
2231 | * lesser of the new and old sizes. If @p is %NULL, krealloc() | |
2232 | * behaves exactly like kmalloc(). If @size is 0 and @p is not a | |
2233 | * %NULL pointer, the object pointed to is freed. | |
2234 | */ | |
2235 | void *krealloc(const void *p, size_t new_size, gfp_t flags) | |
2236 | { | |
2237 | struct kmem_cache *new_cache; | |
2238 | void *ret; | |
2239 | struct page *page; | |
2240 | ||
2241 | if (unlikely(!p)) | |
2242 | return kmalloc(new_size, flags); | |
2243 | ||
2244 | if (unlikely(!new_size)) { | |
2245 | kfree(p); | |
2246 | return NULL; | |
2247 | } | |
2248 | ||
b49af68f | 2249 | page = virt_to_head_page(p); |
81819f0f CL |
2250 | |
2251 | new_cache = get_slab(new_size, flags); | |
2252 | ||
2253 | /* | |
2254 | * If new size fits in the current cache, bail out. | |
2255 | */ | |
2256 | if (likely(page->slab == new_cache)) | |
2257 | return (void *)p; | |
2258 | ||
2259 | ret = kmalloc(new_size, flags); | |
2260 | if (ret) { | |
2261 | memcpy(ret, p, min(new_size, ksize(p))); | |
2262 | kfree(p); | |
2263 | } | |
2264 | return ret; | |
2265 | } | |
2266 | EXPORT_SYMBOL(krealloc); | |
2267 | ||
2268 | /******************************************************************** | |
2269 | * Basic setup of slabs | |
2270 | *******************************************************************/ | |
2271 | ||
2272 | void __init kmem_cache_init(void) | |
2273 | { | |
2274 | int i; | |
2275 | ||
2276 | #ifdef CONFIG_NUMA | |
2277 | /* | |
2278 | * Must first have the slab cache available for the allocations of the | |
2279 | * struct kmalloc_cache_node's. There is special bootstrap code in | |
2280 | * kmem_cache_open for slab_state == DOWN. | |
2281 | */ | |
2282 | create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", | |
2283 | sizeof(struct kmem_cache_node), GFP_KERNEL); | |
2284 | #endif | |
2285 | ||
2286 | /* Able to allocate the per node structures */ | |
2287 | slab_state = PARTIAL; | |
2288 | ||
2289 | /* Caches that are not of the two-to-the-power-of size */ | |
2290 | create_kmalloc_cache(&kmalloc_caches[1], | |
2291 | "kmalloc-96", 96, GFP_KERNEL); | |
2292 | create_kmalloc_cache(&kmalloc_caches[2], | |
2293 | "kmalloc-192", 192, GFP_KERNEL); | |
2294 | ||
2295 | for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) | |
2296 | create_kmalloc_cache(&kmalloc_caches[i], | |
2297 | "kmalloc", 1 << i, GFP_KERNEL); | |
2298 | ||
2299 | slab_state = UP; | |
2300 | ||
2301 | /* Provide the correct kmalloc names now that the caches are up */ | |
2302 | for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) | |
2303 | kmalloc_caches[i]. name = | |
2304 | kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); | |
2305 | ||
2306 | #ifdef CONFIG_SMP | |
2307 | register_cpu_notifier(&slab_notifier); | |
2308 | #endif | |
2309 | ||
2310 | if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */ | |
2311 | kmem_size = offsetof(struct kmem_cache, cpu_slab) | |
2312 | + nr_cpu_ids * sizeof(struct page *); | |
2313 | ||
2314 | printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
2315 | " Processors=%d, Nodes=%d\n", | |
2316 | KMALLOC_SHIFT_HIGH, L1_CACHE_BYTES, | |
2317 | slub_min_order, slub_max_order, slub_min_objects, | |
2318 | nr_cpu_ids, nr_node_ids); | |
2319 | } | |
2320 | ||
2321 | /* | |
2322 | * Find a mergeable slab cache | |
2323 | */ | |
2324 | static int slab_unmergeable(struct kmem_cache *s) | |
2325 | { | |
2326 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
2327 | return 1; | |
2328 | ||
2329 | if (s->ctor || s->dtor) | |
2330 | return 1; | |
2331 | ||
2332 | return 0; | |
2333 | } | |
2334 | ||
2335 | static struct kmem_cache *find_mergeable(size_t size, | |
2336 | size_t align, unsigned long flags, | |
2337 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
2338 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
2339 | { | |
2340 | struct list_head *h; | |
2341 | ||
2342 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
2343 | return NULL; | |
2344 | ||
2345 | if (ctor || dtor) | |
2346 | return NULL; | |
2347 | ||
2348 | size = ALIGN(size, sizeof(void *)); | |
2349 | align = calculate_alignment(flags, align, size); | |
2350 | size = ALIGN(size, align); | |
2351 | ||
2352 | list_for_each(h, &slab_caches) { | |
2353 | struct kmem_cache *s = | |
2354 | container_of(h, struct kmem_cache, list); | |
2355 | ||
2356 | if (slab_unmergeable(s)) | |
2357 | continue; | |
2358 | ||
2359 | if (size > s->size) | |
2360 | continue; | |
2361 | ||
2362 | if (((flags | slub_debug) & SLUB_MERGE_SAME) != | |
2363 | (s->flags & SLUB_MERGE_SAME)) | |
2364 | continue; | |
2365 | /* | |
2366 | * Check if alignment is compatible. | |
2367 | * Courtesy of Adrian Drzewiecki | |
2368 | */ | |
2369 | if ((s->size & ~(align -1)) != s->size) | |
2370 | continue; | |
2371 | ||
2372 | if (s->size - size >= sizeof(void *)) | |
2373 | continue; | |
2374 | ||
2375 | return s; | |
2376 | } | |
2377 | return NULL; | |
2378 | } | |
2379 | ||
2380 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
2381 | size_t align, unsigned long flags, | |
2382 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
2383 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
2384 | { | |
2385 | struct kmem_cache *s; | |
2386 | ||
2387 | down_write(&slub_lock); | |
2388 | s = find_mergeable(size, align, flags, dtor, ctor); | |
2389 | if (s) { | |
2390 | s->refcount++; | |
2391 | /* | |
2392 | * Adjust the object sizes so that we clear | |
2393 | * the complete object on kzalloc. | |
2394 | */ | |
2395 | s->objsize = max(s->objsize, (int)size); | |
2396 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
2397 | if (sysfs_slab_alias(s, name)) | |
2398 | goto err; | |
2399 | } else { | |
2400 | s = kmalloc(kmem_size, GFP_KERNEL); | |
2401 | if (s && kmem_cache_open(s, GFP_KERNEL, name, | |
2402 | size, align, flags, ctor, dtor)) { | |
2403 | if (sysfs_slab_add(s)) { | |
2404 | kfree(s); | |
2405 | goto err; | |
2406 | } | |
2407 | list_add(&s->list, &slab_caches); | |
2408 | } else | |
2409 | kfree(s); | |
2410 | } | |
2411 | up_write(&slub_lock); | |
2412 | return s; | |
2413 | ||
2414 | err: | |
2415 | up_write(&slub_lock); | |
2416 | if (flags & SLAB_PANIC) | |
2417 | panic("Cannot create slabcache %s\n", name); | |
2418 | else | |
2419 | s = NULL; | |
2420 | return s; | |
2421 | } | |
2422 | EXPORT_SYMBOL(kmem_cache_create); | |
2423 | ||
2424 | void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags) | |
2425 | { | |
2426 | void *x; | |
2427 | ||
77c5e2d0 | 2428 | x = slab_alloc(s, flags, -1, __builtin_return_address(0)); |
81819f0f CL |
2429 | if (x) |
2430 | memset(x, 0, s->objsize); | |
2431 | return x; | |
2432 | } | |
2433 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
2434 | ||
2435 | #ifdef CONFIG_SMP | |
2436 | static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu) | |
2437 | { | |
2438 | struct list_head *h; | |
2439 | ||
2440 | down_read(&slub_lock); | |
2441 | list_for_each(h, &slab_caches) { | |
2442 | struct kmem_cache *s = | |
2443 | container_of(h, struct kmem_cache, list); | |
2444 | ||
2445 | func(s, cpu); | |
2446 | } | |
2447 | up_read(&slub_lock); | |
2448 | } | |
2449 | ||
2450 | /* | |
2451 | * Use the cpu notifier to insure that the slab are flushed | |
2452 | * when necessary. | |
2453 | */ | |
2454 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
2455 | unsigned long action, void *hcpu) | |
2456 | { | |
2457 | long cpu = (long)hcpu; | |
2458 | ||
2459 | switch (action) { | |
2460 | case CPU_UP_CANCELED: | |
2461 | case CPU_DEAD: | |
2462 | for_all_slabs(__flush_cpu_slab, cpu); | |
2463 | break; | |
2464 | default: | |
2465 | break; | |
2466 | } | |
2467 | return NOTIFY_OK; | |
2468 | } | |
2469 | ||
2470 | static struct notifier_block __cpuinitdata slab_notifier = | |
2471 | { &slab_cpuup_callback, NULL, 0 }; | |
2472 | ||
2473 | #endif | |
2474 | ||
81819f0f CL |
2475 | #ifdef CONFIG_NUMA |
2476 | ||
2477 | /***************************************************************** | |
2478 | * Generic reaper used to support the page allocator | |
2479 | * (the cpu slabs are reaped by a per slab workqueue). | |
2480 | * | |
2481 | * Maybe move this to the page allocator? | |
2482 | ****************************************************************/ | |
2483 | ||
2484 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
2485 | ||
2486 | static void init_reap_node(int cpu) | |
2487 | { | |
2488 | int node; | |
2489 | ||
2490 | node = next_node(cpu_to_node(cpu), node_online_map); | |
2491 | if (node == MAX_NUMNODES) | |
2492 | node = first_node(node_online_map); | |
2493 | ||
2494 | __get_cpu_var(reap_node) = node; | |
2495 | } | |
2496 | ||
2497 | static void next_reap_node(void) | |
2498 | { | |
2499 | int node = __get_cpu_var(reap_node); | |
2500 | ||
2501 | /* | |
2502 | * Also drain per cpu pages on remote zones | |
2503 | */ | |
2504 | if (node != numa_node_id()) | |
2505 | drain_node_pages(node); | |
2506 | ||
2507 | node = next_node(node, node_online_map); | |
2508 | if (unlikely(node >= MAX_NUMNODES)) | |
2509 | node = first_node(node_online_map); | |
2510 | __get_cpu_var(reap_node) = node; | |
2511 | } | |
2512 | #else | |
2513 | #define init_reap_node(cpu) do { } while (0) | |
2514 | #define next_reap_node(void) do { } while (0) | |
2515 | #endif | |
2516 | ||
2517 | #define REAPTIMEOUT_CPUC (2*HZ) | |
2518 | ||
2519 | #ifdef CONFIG_SMP | |
2520 | static DEFINE_PER_CPU(struct delayed_work, reap_work); | |
2521 | ||
2522 | static void cache_reap(struct work_struct *unused) | |
2523 | { | |
2524 | next_reap_node(); | |
2525 | refresh_cpu_vm_stats(smp_processor_id()); | |
2526 | schedule_delayed_work(&__get_cpu_var(reap_work), | |
2527 | REAPTIMEOUT_CPUC); | |
2528 | } | |
2529 | ||
2530 | static void __devinit start_cpu_timer(int cpu) | |
2531 | { | |
2532 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); | |
2533 | ||
2534 | /* | |
2535 | * When this gets called from do_initcalls via cpucache_init(), | |
2536 | * init_workqueues() has already run, so keventd will be setup | |
2537 | * at that time. | |
2538 | */ | |
2539 | if (keventd_up() && reap_work->work.func == NULL) { | |
2540 | init_reap_node(cpu); | |
2541 | INIT_DELAYED_WORK(reap_work, cache_reap); | |
2542 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
2543 | } | |
2544 | } | |
2545 | ||
2546 | static int __init cpucache_init(void) | |
2547 | { | |
2548 | int cpu; | |
2549 | ||
2550 | /* | |
2551 | * Register the timers that drain pcp pages and update vm statistics | |
2552 | */ | |
2553 | for_each_online_cpu(cpu) | |
2554 | start_cpu_timer(cpu); | |
2555 | return 0; | |
2556 | } | |
2557 | __initcall(cpucache_init); | |
2558 | #endif | |
2559 | ||
2560 | #ifdef SLUB_RESILIENCY_TEST | |
2561 | static unsigned long validate_slab_cache(struct kmem_cache *s); | |
2562 | ||
2563 | static void resiliency_test(void) | |
2564 | { | |
2565 | u8 *p; | |
2566 | ||
2567 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
2568 | printk(KERN_ERR "-----------------------\n"); | |
2569 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
2570 | ||
2571 | p = kzalloc(16, GFP_KERNEL); | |
2572 | p[16] = 0x12; | |
2573 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
2574 | " 0x12->0x%p\n\n", p + 16); | |
2575 | ||
2576 | validate_slab_cache(kmalloc_caches + 4); | |
2577 | ||
2578 | /* Hmmm... The next two are dangerous */ | |
2579 | p = kzalloc(32, GFP_KERNEL); | |
2580 | p[32 + sizeof(void *)] = 0x34; | |
2581 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
2582 | " 0x34 -> -0x%p\n", p); | |
2583 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
2584 | ||
2585 | validate_slab_cache(kmalloc_caches + 5); | |
2586 | p = kzalloc(64, GFP_KERNEL); | |
2587 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
2588 | *p = 0x56; | |
2589 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
2590 | p); | |
2591 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
2592 | validate_slab_cache(kmalloc_caches + 6); | |
2593 | ||
2594 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
2595 | p = kzalloc(128, GFP_KERNEL); | |
2596 | kfree(p); | |
2597 | *p = 0x78; | |
2598 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
2599 | validate_slab_cache(kmalloc_caches + 7); | |
2600 | ||
2601 | p = kzalloc(256, GFP_KERNEL); | |
2602 | kfree(p); | |
2603 | p[50] = 0x9a; | |
2604 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); | |
2605 | validate_slab_cache(kmalloc_caches + 8); | |
2606 | ||
2607 | p = kzalloc(512, GFP_KERNEL); | |
2608 | kfree(p); | |
2609 | p[512] = 0xab; | |
2610 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
2611 | validate_slab_cache(kmalloc_caches + 9); | |
2612 | } | |
2613 | #else | |
2614 | static void resiliency_test(void) {}; | |
2615 | #endif | |
2616 | ||
2617 | /* | |
2618 | * These are not as efficient as kmalloc for the non debug case. | |
2619 | * We do not have the page struct available so we have to touch one | |
2620 | * cacheline in struct kmem_cache to check slab flags. | |
2621 | */ | |
2622 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) | |
2623 | { | |
2624 | struct kmem_cache *s = get_slab(size, gfpflags); | |
81819f0f CL |
2625 | |
2626 | if (!s) | |
2627 | return NULL; | |
2628 | ||
77c5e2d0 | 2629 | return slab_alloc(s, gfpflags, -1, caller); |
81819f0f CL |
2630 | } |
2631 | ||
2632 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
2633 | int node, void *caller) | |
2634 | { | |
2635 | struct kmem_cache *s = get_slab(size, gfpflags); | |
81819f0f CL |
2636 | |
2637 | if (!s) | |
2638 | return NULL; | |
2639 | ||
77c5e2d0 | 2640 | return slab_alloc(s, gfpflags, node, caller); |
81819f0f CL |
2641 | } |
2642 | ||
2643 | #ifdef CONFIG_SYSFS | |
2644 | ||
53e15af0 CL |
2645 | static int validate_slab(struct kmem_cache *s, struct page *page) |
2646 | { | |
2647 | void *p; | |
2648 | void *addr = page_address(page); | |
2649 | unsigned long map[BITS_TO_LONGS(s->objects)]; | |
2650 | ||
2651 | if (!check_slab(s, page) || | |
2652 | !on_freelist(s, page, NULL)) | |
2653 | return 0; | |
2654 | ||
2655 | /* Now we know that a valid freelist exists */ | |
2656 | bitmap_zero(map, s->objects); | |
2657 | ||
2658 | for(p = page->freelist; p; p = get_freepointer(s, p)) { | |
2659 | set_bit((p - addr) / s->size, map); | |
2660 | if (!check_object(s, page, p, 0)) | |
2661 | return 0; | |
2662 | } | |
2663 | ||
2664 | for(p = addr; p < addr + s->objects * s->size; p += s->size) | |
2665 | if (!test_bit((p - addr) / s->size, map)) | |
2666 | if (!check_object(s, page, p, 1)) | |
2667 | return 0; | |
2668 | return 1; | |
2669 | } | |
2670 | ||
2671 | static void validate_slab_slab(struct kmem_cache *s, struct page *page) | |
2672 | { | |
2673 | if (slab_trylock(page)) { | |
2674 | validate_slab(s, page); | |
2675 | slab_unlock(page); | |
2676 | } else | |
2677 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", | |
2678 | s->name, page); | |
2679 | ||
2680 | if (s->flags & DEBUG_DEFAULT_FLAGS) { | |
2681 | if (!PageError(page)) | |
2682 | printk(KERN_ERR "SLUB %s: PageError not set " | |
2683 | "on slab 0x%p\n", s->name, page); | |
2684 | } else { | |
2685 | if (PageError(page)) | |
2686 | printk(KERN_ERR "SLUB %s: PageError set on " | |
2687 | "slab 0x%p\n", s->name, page); | |
2688 | } | |
2689 | } | |
2690 | ||
2691 | static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n) | |
2692 | { | |
2693 | unsigned long count = 0; | |
2694 | struct page *page; | |
2695 | unsigned long flags; | |
2696 | ||
2697 | spin_lock_irqsave(&n->list_lock, flags); | |
2698 | ||
2699 | list_for_each_entry(page, &n->partial, lru) { | |
2700 | validate_slab_slab(s, page); | |
2701 | count++; | |
2702 | } | |
2703 | if (count != n->nr_partial) | |
2704 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
2705 | "counter=%ld\n", s->name, count, n->nr_partial); | |
2706 | ||
2707 | if (!(s->flags & SLAB_STORE_USER)) | |
2708 | goto out; | |
2709 | ||
2710 | list_for_each_entry(page, &n->full, lru) { | |
2711 | validate_slab_slab(s, page); | |
2712 | count++; | |
2713 | } | |
2714 | if (count != atomic_long_read(&n->nr_slabs)) | |
2715 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
2716 | "counter=%ld\n", s->name, count, | |
2717 | atomic_long_read(&n->nr_slabs)); | |
2718 | ||
2719 | out: | |
2720 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2721 | return count; | |
2722 | } | |
2723 | ||
2724 | static unsigned long validate_slab_cache(struct kmem_cache *s) | |
2725 | { | |
2726 | int node; | |
2727 | unsigned long count = 0; | |
2728 | ||
2729 | flush_all(s); | |
2730 | for_each_online_node(node) { | |
2731 | struct kmem_cache_node *n = get_node(s, node); | |
2732 | ||
2733 | count += validate_slab_node(s, n); | |
2734 | } | |
2735 | return count; | |
2736 | } | |
2737 | ||
88a420e4 CL |
2738 | /* |
2739 | * Generate lists of locations where slabcache objects are allocated | |
2740 | * and freed. | |
2741 | */ | |
2742 | ||
2743 | struct location { | |
2744 | unsigned long count; | |
2745 | void *addr; | |
2746 | }; | |
2747 | ||
2748 | struct loc_track { | |
2749 | unsigned long max; | |
2750 | unsigned long count; | |
2751 | struct location *loc; | |
2752 | }; | |
2753 | ||
2754 | static void free_loc_track(struct loc_track *t) | |
2755 | { | |
2756 | if (t->max) | |
2757 | free_pages((unsigned long)t->loc, | |
2758 | get_order(sizeof(struct location) * t->max)); | |
2759 | } | |
2760 | ||
2761 | static int alloc_loc_track(struct loc_track *t, unsigned long max) | |
2762 | { | |
2763 | struct location *l; | |
2764 | int order; | |
2765 | ||
2766 | if (!max) | |
2767 | max = PAGE_SIZE / sizeof(struct location); | |
2768 | ||
2769 | order = get_order(sizeof(struct location) * max); | |
2770 | ||
2771 | l = (void *)__get_free_pages(GFP_KERNEL, order); | |
2772 | ||
2773 | if (!l) | |
2774 | return 0; | |
2775 | ||
2776 | if (t->count) { | |
2777 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
2778 | free_loc_track(t); | |
2779 | } | |
2780 | t->max = max; | |
2781 | t->loc = l; | |
2782 | return 1; | |
2783 | } | |
2784 | ||
2785 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
2786 | void *addr) | |
2787 | { | |
2788 | long start, end, pos; | |
2789 | struct location *l; | |
2790 | void *caddr; | |
2791 | ||
2792 | start = -1; | |
2793 | end = t->count; | |
2794 | ||
2795 | for ( ; ; ) { | |
2796 | pos = start + (end - start + 1) / 2; | |
2797 | ||
2798 | /* | |
2799 | * There is nothing at "end". If we end up there | |
2800 | * we need to add something to before end. | |
2801 | */ | |
2802 | if (pos == end) | |
2803 | break; | |
2804 | ||
2805 | caddr = t->loc[pos].addr; | |
2806 | if (addr == caddr) { | |
2807 | t->loc[pos].count++; | |
2808 | return 1; | |
2809 | } | |
2810 | ||
2811 | if (addr < caddr) | |
2812 | end = pos; | |
2813 | else | |
2814 | start = pos; | |
2815 | } | |
2816 | ||
2817 | /* | |
2818 | * Not found. Insert new tracking element | |
2819 | */ | |
2820 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max)) | |
2821 | return 0; | |
2822 | ||
2823 | l = t->loc + pos; | |
2824 | if (pos < t->count) | |
2825 | memmove(l + 1, l, | |
2826 | (t->count - pos) * sizeof(struct location)); | |
2827 | t->count++; | |
2828 | l->count = 1; | |
2829 | l->addr = addr; | |
2830 | return 1; | |
2831 | } | |
2832 | ||
2833 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
2834 | struct page *page, enum track_item alloc) | |
2835 | { | |
2836 | void *addr = page_address(page); | |
2837 | unsigned long map[BITS_TO_LONGS(s->objects)]; | |
2838 | void *p; | |
2839 | ||
2840 | bitmap_zero(map, s->objects); | |
2841 | for (p = page->freelist; p; p = get_freepointer(s, p)) | |
2842 | set_bit((p - addr) / s->size, map); | |
2843 | ||
2844 | for (p = addr; p < addr + s->objects * s->size; p += s->size) | |
2845 | if (!test_bit((p - addr) / s->size, map)) { | |
2846 | void *addr = get_track(s, p, alloc)->addr; | |
2847 | ||
2848 | add_location(t, s, addr); | |
2849 | } | |
2850 | } | |
2851 | ||
2852 | static int list_locations(struct kmem_cache *s, char *buf, | |
2853 | enum track_item alloc) | |
2854 | { | |
2855 | int n = 0; | |
2856 | unsigned long i; | |
2857 | struct loc_track t; | |
2858 | int node; | |
2859 | ||
2860 | t.count = 0; | |
2861 | t.max = 0; | |
2862 | ||
2863 | /* Push back cpu slabs */ | |
2864 | flush_all(s); | |
2865 | ||
2866 | for_each_online_node(node) { | |
2867 | struct kmem_cache_node *n = get_node(s, node); | |
2868 | unsigned long flags; | |
2869 | struct page *page; | |
2870 | ||
2871 | if (!atomic_read(&n->nr_slabs)) | |
2872 | continue; | |
2873 | ||
2874 | spin_lock_irqsave(&n->list_lock, flags); | |
2875 | list_for_each_entry(page, &n->partial, lru) | |
2876 | process_slab(&t, s, page, alloc); | |
2877 | list_for_each_entry(page, &n->full, lru) | |
2878 | process_slab(&t, s, page, alloc); | |
2879 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2880 | } | |
2881 | ||
2882 | for (i = 0; i < t.count; i++) { | |
2883 | void *addr = t.loc[i].addr; | |
2884 | ||
2885 | if (n > PAGE_SIZE - 100) | |
2886 | break; | |
2887 | n += sprintf(buf + n, "%7ld ", t.loc[i].count); | |
2888 | if (addr) | |
2889 | n += sprint_symbol(buf + n, (unsigned long)t.loc[i].addr); | |
2890 | else | |
2891 | n += sprintf(buf + n, "<not-available>"); | |
2892 | n += sprintf(buf + n, "\n"); | |
2893 | } | |
2894 | ||
2895 | free_loc_track(&t); | |
2896 | if (!t.count) | |
2897 | n += sprintf(buf, "No data\n"); | |
2898 | return n; | |
2899 | } | |
2900 | ||
81819f0f CL |
2901 | static unsigned long count_partial(struct kmem_cache_node *n) |
2902 | { | |
2903 | unsigned long flags; | |
2904 | unsigned long x = 0; | |
2905 | struct page *page; | |
2906 | ||
2907 | spin_lock_irqsave(&n->list_lock, flags); | |
2908 | list_for_each_entry(page, &n->partial, lru) | |
2909 | x += page->inuse; | |
2910 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2911 | return x; | |
2912 | } | |
2913 | ||
2914 | enum slab_stat_type { | |
2915 | SL_FULL, | |
2916 | SL_PARTIAL, | |
2917 | SL_CPU, | |
2918 | SL_OBJECTS | |
2919 | }; | |
2920 | ||
2921 | #define SO_FULL (1 << SL_FULL) | |
2922 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
2923 | #define SO_CPU (1 << SL_CPU) | |
2924 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
2925 | ||
2926 | static unsigned long slab_objects(struct kmem_cache *s, | |
2927 | char *buf, unsigned long flags) | |
2928 | { | |
2929 | unsigned long total = 0; | |
2930 | int cpu; | |
2931 | int node; | |
2932 | int x; | |
2933 | unsigned long *nodes; | |
2934 | unsigned long *per_cpu; | |
2935 | ||
2936 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
2937 | per_cpu = nodes + nr_node_ids; | |
2938 | ||
2939 | for_each_possible_cpu(cpu) { | |
2940 | struct page *page = s->cpu_slab[cpu]; | |
2941 | int node; | |
2942 | ||
2943 | if (page) { | |
2944 | node = page_to_nid(page); | |
2945 | if (flags & SO_CPU) { | |
2946 | int x = 0; | |
2947 | ||
2948 | if (flags & SO_OBJECTS) | |
2949 | x = page->inuse; | |
2950 | else | |
2951 | x = 1; | |
2952 | total += x; | |
2953 | nodes[node] += x; | |
2954 | } | |
2955 | per_cpu[node]++; | |
2956 | } | |
2957 | } | |
2958 | ||
2959 | for_each_online_node(node) { | |
2960 | struct kmem_cache_node *n = get_node(s, node); | |
2961 | ||
2962 | if (flags & SO_PARTIAL) { | |
2963 | if (flags & SO_OBJECTS) | |
2964 | x = count_partial(n); | |
2965 | else | |
2966 | x = n->nr_partial; | |
2967 | total += x; | |
2968 | nodes[node] += x; | |
2969 | } | |
2970 | ||
2971 | if (flags & SO_FULL) { | |
2972 | int full_slabs = atomic_read(&n->nr_slabs) | |
2973 | - per_cpu[node] | |
2974 | - n->nr_partial; | |
2975 | ||
2976 | if (flags & SO_OBJECTS) | |
2977 | x = full_slabs * s->objects; | |
2978 | else | |
2979 | x = full_slabs; | |
2980 | total += x; | |
2981 | nodes[node] += x; | |
2982 | } | |
2983 | } | |
2984 | ||
2985 | x = sprintf(buf, "%lu", total); | |
2986 | #ifdef CONFIG_NUMA | |
2987 | for_each_online_node(node) | |
2988 | if (nodes[node]) | |
2989 | x += sprintf(buf + x, " N%d=%lu", | |
2990 | node, nodes[node]); | |
2991 | #endif | |
2992 | kfree(nodes); | |
2993 | return x + sprintf(buf + x, "\n"); | |
2994 | } | |
2995 | ||
2996 | static int any_slab_objects(struct kmem_cache *s) | |
2997 | { | |
2998 | int node; | |
2999 | int cpu; | |
3000 | ||
3001 | for_each_possible_cpu(cpu) | |
3002 | if (s->cpu_slab[cpu]) | |
3003 | return 1; | |
3004 | ||
3005 | for_each_node(node) { | |
3006 | struct kmem_cache_node *n = get_node(s, node); | |
3007 | ||
3008 | if (n->nr_partial || atomic_read(&n->nr_slabs)) | |
3009 | return 1; | |
3010 | } | |
3011 | return 0; | |
3012 | } | |
3013 | ||
3014 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
3015 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); | |
3016 | ||
3017 | struct slab_attribute { | |
3018 | struct attribute attr; | |
3019 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
3020 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
3021 | }; | |
3022 | ||
3023 | #define SLAB_ATTR_RO(_name) \ | |
3024 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) | |
3025 | ||
3026 | #define SLAB_ATTR(_name) \ | |
3027 | static struct slab_attribute _name##_attr = \ | |
3028 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3029 | ||
81819f0f CL |
3030 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
3031 | { | |
3032 | return sprintf(buf, "%d\n", s->size); | |
3033 | } | |
3034 | SLAB_ATTR_RO(slab_size); | |
3035 | ||
3036 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
3037 | { | |
3038 | return sprintf(buf, "%d\n", s->align); | |
3039 | } | |
3040 | SLAB_ATTR_RO(align); | |
3041 | ||
3042 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
3043 | { | |
3044 | return sprintf(buf, "%d\n", s->objsize); | |
3045 | } | |
3046 | SLAB_ATTR_RO(object_size); | |
3047 | ||
3048 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
3049 | { | |
3050 | return sprintf(buf, "%d\n", s->objects); | |
3051 | } | |
3052 | SLAB_ATTR_RO(objs_per_slab); | |
3053 | ||
3054 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
3055 | { | |
3056 | return sprintf(buf, "%d\n", s->order); | |
3057 | } | |
3058 | SLAB_ATTR_RO(order); | |
3059 | ||
3060 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
3061 | { | |
3062 | if (s->ctor) { | |
3063 | int n = sprint_symbol(buf, (unsigned long)s->ctor); | |
3064 | ||
3065 | return n + sprintf(buf + n, "\n"); | |
3066 | } | |
3067 | return 0; | |
3068 | } | |
3069 | SLAB_ATTR_RO(ctor); | |
3070 | ||
3071 | static ssize_t dtor_show(struct kmem_cache *s, char *buf) | |
3072 | { | |
3073 | if (s->dtor) { | |
3074 | int n = sprint_symbol(buf, (unsigned long)s->dtor); | |
3075 | ||
3076 | return n + sprintf(buf + n, "\n"); | |
3077 | } | |
3078 | return 0; | |
3079 | } | |
3080 | SLAB_ATTR_RO(dtor); | |
3081 | ||
3082 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) | |
3083 | { | |
3084 | return sprintf(buf, "%d\n", s->refcount - 1); | |
3085 | } | |
3086 | SLAB_ATTR_RO(aliases); | |
3087 | ||
3088 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
3089 | { | |
3090 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU); | |
3091 | } | |
3092 | SLAB_ATTR_RO(slabs); | |
3093 | ||
3094 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
3095 | { | |
3096 | return slab_objects(s, buf, SO_PARTIAL); | |
3097 | } | |
3098 | SLAB_ATTR_RO(partial); | |
3099 | ||
3100 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
3101 | { | |
3102 | return slab_objects(s, buf, SO_CPU); | |
3103 | } | |
3104 | SLAB_ATTR_RO(cpu_slabs); | |
3105 | ||
3106 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
3107 | { | |
3108 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS); | |
3109 | } | |
3110 | SLAB_ATTR_RO(objects); | |
3111 | ||
3112 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
3113 | { | |
3114 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
3115 | } | |
3116 | ||
3117 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
3118 | const char *buf, size_t length) | |
3119 | { | |
3120 | s->flags &= ~SLAB_DEBUG_FREE; | |
3121 | if (buf[0] == '1') | |
3122 | s->flags |= SLAB_DEBUG_FREE; | |
3123 | return length; | |
3124 | } | |
3125 | SLAB_ATTR(sanity_checks); | |
3126 | ||
3127 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
3128 | { | |
3129 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
3130 | } | |
3131 | ||
3132 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
3133 | size_t length) | |
3134 | { | |
3135 | s->flags &= ~SLAB_TRACE; | |
3136 | if (buf[0] == '1') | |
3137 | s->flags |= SLAB_TRACE; | |
3138 | return length; | |
3139 | } | |
3140 | SLAB_ATTR(trace); | |
3141 | ||
3142 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
3143 | { | |
3144 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
3145 | } | |
3146 | ||
3147 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
3148 | const char *buf, size_t length) | |
3149 | { | |
3150 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
3151 | if (buf[0] == '1') | |
3152 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
3153 | return length; | |
3154 | } | |
3155 | SLAB_ATTR(reclaim_account); | |
3156 | ||
3157 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
3158 | { | |
3159 | return sprintf(buf, "%d\n", !!(s->flags & | |
3160 | (SLAB_HWCACHE_ALIGN|SLAB_MUST_HWCACHE_ALIGN))); | |
3161 | } | |
3162 | SLAB_ATTR_RO(hwcache_align); | |
3163 | ||
3164 | #ifdef CONFIG_ZONE_DMA | |
3165 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
3166 | { | |
3167 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
3168 | } | |
3169 | SLAB_ATTR_RO(cache_dma); | |
3170 | #endif | |
3171 | ||
3172 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
3173 | { | |
3174 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
3175 | } | |
3176 | SLAB_ATTR_RO(destroy_by_rcu); | |
3177 | ||
3178 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
3179 | { | |
3180 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
3181 | } | |
3182 | ||
3183 | static ssize_t red_zone_store(struct kmem_cache *s, | |
3184 | const char *buf, size_t length) | |
3185 | { | |
3186 | if (any_slab_objects(s)) | |
3187 | return -EBUSY; | |
3188 | ||
3189 | s->flags &= ~SLAB_RED_ZONE; | |
3190 | if (buf[0] == '1') | |
3191 | s->flags |= SLAB_RED_ZONE; | |
3192 | calculate_sizes(s); | |
3193 | return length; | |
3194 | } | |
3195 | SLAB_ATTR(red_zone); | |
3196 | ||
3197 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
3198 | { | |
3199 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
3200 | } | |
3201 | ||
3202 | static ssize_t poison_store(struct kmem_cache *s, | |
3203 | const char *buf, size_t length) | |
3204 | { | |
3205 | if (any_slab_objects(s)) | |
3206 | return -EBUSY; | |
3207 | ||
3208 | s->flags &= ~SLAB_POISON; | |
3209 | if (buf[0] == '1') | |
3210 | s->flags |= SLAB_POISON; | |
3211 | calculate_sizes(s); | |
3212 | return length; | |
3213 | } | |
3214 | SLAB_ATTR(poison); | |
3215 | ||
3216 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
3217 | { | |
3218 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
3219 | } | |
3220 | ||
3221 | static ssize_t store_user_store(struct kmem_cache *s, | |
3222 | const char *buf, size_t length) | |
3223 | { | |
3224 | if (any_slab_objects(s)) | |
3225 | return -EBUSY; | |
3226 | ||
3227 | s->flags &= ~SLAB_STORE_USER; | |
3228 | if (buf[0] == '1') | |
3229 | s->flags |= SLAB_STORE_USER; | |
3230 | calculate_sizes(s); | |
3231 | return length; | |
3232 | } | |
3233 | SLAB_ATTR(store_user); | |
3234 | ||
53e15af0 CL |
3235 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
3236 | { | |
3237 | return 0; | |
3238 | } | |
3239 | ||
3240 | static ssize_t validate_store(struct kmem_cache *s, | |
3241 | const char *buf, size_t length) | |
3242 | { | |
3243 | if (buf[0] == '1') | |
3244 | validate_slab_cache(s); | |
3245 | else | |
3246 | return -EINVAL; | |
3247 | return length; | |
3248 | } | |
3249 | SLAB_ATTR(validate); | |
3250 | ||
2086d26a CL |
3251 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
3252 | { | |
3253 | return 0; | |
3254 | } | |
3255 | ||
3256 | static ssize_t shrink_store(struct kmem_cache *s, | |
3257 | const char *buf, size_t length) | |
3258 | { | |
3259 | if (buf[0] == '1') { | |
3260 | int rc = kmem_cache_shrink(s); | |
3261 | ||
3262 | if (rc) | |
3263 | return rc; | |
3264 | } else | |
3265 | return -EINVAL; | |
3266 | return length; | |
3267 | } | |
3268 | SLAB_ATTR(shrink); | |
3269 | ||
88a420e4 CL |
3270 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) |
3271 | { | |
3272 | if (!(s->flags & SLAB_STORE_USER)) | |
3273 | return -ENOSYS; | |
3274 | return list_locations(s, buf, TRACK_ALLOC); | |
3275 | } | |
3276 | SLAB_ATTR_RO(alloc_calls); | |
3277 | ||
3278 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
3279 | { | |
3280 | if (!(s->flags & SLAB_STORE_USER)) | |
3281 | return -ENOSYS; | |
3282 | return list_locations(s, buf, TRACK_FREE); | |
3283 | } | |
3284 | SLAB_ATTR_RO(free_calls); | |
3285 | ||
81819f0f CL |
3286 | #ifdef CONFIG_NUMA |
3287 | static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf) | |
3288 | { | |
3289 | return sprintf(buf, "%d\n", s->defrag_ratio / 10); | |
3290 | } | |
3291 | ||
3292 | static ssize_t defrag_ratio_store(struct kmem_cache *s, | |
3293 | const char *buf, size_t length) | |
3294 | { | |
3295 | int n = simple_strtoul(buf, NULL, 10); | |
3296 | ||
3297 | if (n < 100) | |
3298 | s->defrag_ratio = n * 10; | |
3299 | return length; | |
3300 | } | |
3301 | SLAB_ATTR(defrag_ratio); | |
3302 | #endif | |
3303 | ||
3304 | static struct attribute * slab_attrs[] = { | |
3305 | &slab_size_attr.attr, | |
3306 | &object_size_attr.attr, | |
3307 | &objs_per_slab_attr.attr, | |
3308 | &order_attr.attr, | |
3309 | &objects_attr.attr, | |
3310 | &slabs_attr.attr, | |
3311 | &partial_attr.attr, | |
3312 | &cpu_slabs_attr.attr, | |
3313 | &ctor_attr.attr, | |
3314 | &dtor_attr.attr, | |
3315 | &aliases_attr.attr, | |
3316 | &align_attr.attr, | |
3317 | &sanity_checks_attr.attr, | |
3318 | &trace_attr.attr, | |
3319 | &hwcache_align_attr.attr, | |
3320 | &reclaim_account_attr.attr, | |
3321 | &destroy_by_rcu_attr.attr, | |
3322 | &red_zone_attr.attr, | |
3323 | &poison_attr.attr, | |
3324 | &store_user_attr.attr, | |
53e15af0 | 3325 | &validate_attr.attr, |
2086d26a | 3326 | &shrink_attr.attr, |
88a420e4 CL |
3327 | &alloc_calls_attr.attr, |
3328 | &free_calls_attr.attr, | |
81819f0f CL |
3329 | #ifdef CONFIG_ZONE_DMA |
3330 | &cache_dma_attr.attr, | |
3331 | #endif | |
3332 | #ifdef CONFIG_NUMA | |
3333 | &defrag_ratio_attr.attr, | |
3334 | #endif | |
3335 | NULL | |
3336 | }; | |
3337 | ||
3338 | static struct attribute_group slab_attr_group = { | |
3339 | .attrs = slab_attrs, | |
3340 | }; | |
3341 | ||
3342 | static ssize_t slab_attr_show(struct kobject *kobj, | |
3343 | struct attribute *attr, | |
3344 | char *buf) | |
3345 | { | |
3346 | struct slab_attribute *attribute; | |
3347 | struct kmem_cache *s; | |
3348 | int err; | |
3349 | ||
3350 | attribute = to_slab_attr(attr); | |
3351 | s = to_slab(kobj); | |
3352 | ||
3353 | if (!attribute->show) | |
3354 | return -EIO; | |
3355 | ||
3356 | err = attribute->show(s, buf); | |
3357 | ||
3358 | return err; | |
3359 | } | |
3360 | ||
3361 | static ssize_t slab_attr_store(struct kobject *kobj, | |
3362 | struct attribute *attr, | |
3363 | const char *buf, size_t len) | |
3364 | { | |
3365 | struct slab_attribute *attribute; | |
3366 | struct kmem_cache *s; | |
3367 | int err; | |
3368 | ||
3369 | attribute = to_slab_attr(attr); | |
3370 | s = to_slab(kobj); | |
3371 | ||
3372 | if (!attribute->store) | |
3373 | return -EIO; | |
3374 | ||
3375 | err = attribute->store(s, buf, len); | |
3376 | ||
3377 | return err; | |
3378 | } | |
3379 | ||
3380 | static struct sysfs_ops slab_sysfs_ops = { | |
3381 | .show = slab_attr_show, | |
3382 | .store = slab_attr_store, | |
3383 | }; | |
3384 | ||
3385 | static struct kobj_type slab_ktype = { | |
3386 | .sysfs_ops = &slab_sysfs_ops, | |
3387 | }; | |
3388 | ||
3389 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
3390 | { | |
3391 | struct kobj_type *ktype = get_ktype(kobj); | |
3392 | ||
3393 | if (ktype == &slab_ktype) | |
3394 | return 1; | |
3395 | return 0; | |
3396 | } | |
3397 | ||
3398 | static struct kset_uevent_ops slab_uevent_ops = { | |
3399 | .filter = uevent_filter, | |
3400 | }; | |
3401 | ||
3402 | decl_subsys(slab, &slab_ktype, &slab_uevent_ops); | |
3403 | ||
3404 | #define ID_STR_LENGTH 64 | |
3405 | ||
3406 | /* Create a unique string id for a slab cache: | |
3407 | * format | |
3408 | * :[flags-]size:[memory address of kmemcache] | |
3409 | */ | |
3410 | static char *create_unique_id(struct kmem_cache *s) | |
3411 | { | |
3412 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
3413 | char *p = name; | |
3414 | ||
3415 | BUG_ON(!name); | |
3416 | ||
3417 | *p++ = ':'; | |
3418 | /* | |
3419 | * First flags affecting slabcache operations. We will only | |
3420 | * get here for aliasable slabs so we do not need to support | |
3421 | * too many flags. The flags here must cover all flags that | |
3422 | * are matched during merging to guarantee that the id is | |
3423 | * unique. | |
3424 | */ | |
3425 | if (s->flags & SLAB_CACHE_DMA) | |
3426 | *p++ = 'd'; | |
3427 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
3428 | *p++ = 'a'; | |
3429 | if (s->flags & SLAB_DEBUG_FREE) | |
3430 | *p++ = 'F'; | |
3431 | if (p != name + 1) | |
3432 | *p++ = '-'; | |
3433 | p += sprintf(p, "%07d", s->size); | |
3434 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
3435 | return name; | |
3436 | } | |
3437 | ||
3438 | static int sysfs_slab_add(struct kmem_cache *s) | |
3439 | { | |
3440 | int err; | |
3441 | const char *name; | |
3442 | int unmergeable; | |
3443 | ||
3444 | if (slab_state < SYSFS) | |
3445 | /* Defer until later */ | |
3446 | return 0; | |
3447 | ||
3448 | unmergeable = slab_unmergeable(s); | |
3449 | if (unmergeable) { | |
3450 | /* | |
3451 | * Slabcache can never be merged so we can use the name proper. | |
3452 | * This is typically the case for debug situations. In that | |
3453 | * case we can catch duplicate names easily. | |
3454 | */ | |
3455 | sysfs_remove_link(&slab_subsys.kset.kobj, s->name); | |
3456 | name = s->name; | |
3457 | } else { | |
3458 | /* | |
3459 | * Create a unique name for the slab as a target | |
3460 | * for the symlinks. | |
3461 | */ | |
3462 | name = create_unique_id(s); | |
3463 | } | |
3464 | ||
3465 | kobj_set_kset_s(s, slab_subsys); | |
3466 | kobject_set_name(&s->kobj, name); | |
3467 | kobject_init(&s->kobj); | |
3468 | err = kobject_add(&s->kobj); | |
3469 | if (err) | |
3470 | return err; | |
3471 | ||
3472 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
3473 | if (err) | |
3474 | return err; | |
3475 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
3476 | if (!unmergeable) { | |
3477 | /* Setup first alias */ | |
3478 | sysfs_slab_alias(s, s->name); | |
3479 | kfree(name); | |
3480 | } | |
3481 | return 0; | |
3482 | } | |
3483 | ||
3484 | static void sysfs_slab_remove(struct kmem_cache *s) | |
3485 | { | |
3486 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
3487 | kobject_del(&s->kobj); | |
3488 | } | |
3489 | ||
3490 | /* | |
3491 | * Need to buffer aliases during bootup until sysfs becomes | |
3492 | * available lest we loose that information. | |
3493 | */ | |
3494 | struct saved_alias { | |
3495 | struct kmem_cache *s; | |
3496 | const char *name; | |
3497 | struct saved_alias *next; | |
3498 | }; | |
3499 | ||
3500 | struct saved_alias *alias_list; | |
3501 | ||
3502 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
3503 | { | |
3504 | struct saved_alias *al; | |
3505 | ||
3506 | if (slab_state == SYSFS) { | |
3507 | /* | |
3508 | * If we have a leftover link then remove it. | |
3509 | */ | |
3510 | sysfs_remove_link(&slab_subsys.kset.kobj, name); | |
3511 | return sysfs_create_link(&slab_subsys.kset.kobj, | |
3512 | &s->kobj, name); | |
3513 | } | |
3514 | ||
3515 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
3516 | if (!al) | |
3517 | return -ENOMEM; | |
3518 | ||
3519 | al->s = s; | |
3520 | al->name = name; | |
3521 | al->next = alias_list; | |
3522 | alias_list = al; | |
3523 | return 0; | |
3524 | } | |
3525 | ||
3526 | static int __init slab_sysfs_init(void) | |
3527 | { | |
3528 | int err; | |
3529 | ||
3530 | err = subsystem_register(&slab_subsys); | |
3531 | if (err) { | |
3532 | printk(KERN_ERR "Cannot register slab subsystem.\n"); | |
3533 | return -ENOSYS; | |
3534 | } | |
3535 | ||
3536 | finish_bootstrap(); | |
3537 | ||
3538 | while (alias_list) { | |
3539 | struct saved_alias *al = alias_list; | |
3540 | ||
3541 | alias_list = alias_list->next; | |
3542 | err = sysfs_slab_alias(al->s, al->name); | |
3543 | BUG_ON(err); | |
3544 | kfree(al); | |
3545 | } | |
3546 | ||
3547 | resiliency_test(); | |
3548 | return 0; | |
3549 | } | |
3550 | ||
3551 | __initcall(slab_sysfs_init); | |
3552 | #else | |
3553 | __initcall(finish_bootstrap); | |
3554 | #endif |