]>
Commit | Line | Data |
---|---|---|
81819f0f CL |
1 | /* |
2 | * SLUB: A slab allocator that limits cache line use instead of queuing | |
3 | * objects in per cpu and per node lists. | |
4 | * | |
5 | * The allocator synchronizes using per slab locks and only | |
6 | * uses a centralized lock to manage a pool of partial slabs. | |
7 | * | |
8 | * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com> | |
9 | */ | |
10 | ||
11 | #include <linux/mm.h> | |
12 | #include <linux/module.h> | |
13 | #include <linux/bit_spinlock.h> | |
14 | #include <linux/interrupt.h> | |
15 | #include <linux/bitops.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/seq_file.h> | |
18 | #include <linux/cpu.h> | |
19 | #include <linux/cpuset.h> | |
20 | #include <linux/mempolicy.h> | |
21 | #include <linux/ctype.h> | |
22 | #include <linux/kallsyms.h> | |
23 | ||
24 | /* | |
25 | * Lock order: | |
26 | * 1. slab_lock(page) | |
27 | * 2. slab->list_lock | |
28 | * | |
29 | * The slab_lock protects operations on the object of a particular | |
30 | * slab and its metadata in the page struct. If the slab lock | |
31 | * has been taken then no allocations nor frees can be performed | |
32 | * on the objects in the slab nor can the slab be added or removed | |
33 | * from the partial or full lists since this would mean modifying | |
34 | * the page_struct of the slab. | |
35 | * | |
36 | * The list_lock protects the partial and full list on each node and | |
37 | * the partial slab counter. If taken then no new slabs may be added or | |
38 | * removed from the lists nor make the number of partial slabs be modified. | |
39 | * (Note that the total number of slabs is an atomic value that may be | |
40 | * modified without taking the list lock). | |
41 | * | |
42 | * The list_lock is a centralized lock and thus we avoid taking it as | |
43 | * much as possible. As long as SLUB does not have to handle partial | |
44 | * slabs, operations can continue without any centralized lock. F.e. | |
45 | * allocating a long series of objects that fill up slabs does not require | |
46 | * the list lock. | |
47 | * | |
48 | * The lock order is sometimes inverted when we are trying to get a slab | |
49 | * off a list. We take the list_lock and then look for a page on the list | |
50 | * to use. While we do that objects in the slabs may be freed. We can | |
51 | * only operate on the slab if we have also taken the slab_lock. So we use | |
52 | * a slab_trylock() on the slab. If trylock was successful then no frees | |
53 | * can occur anymore and we can use the slab for allocations etc. If the | |
54 | * slab_trylock() does not succeed then frees are in progress in the slab and | |
55 | * we must stay away from it for a while since we may cause a bouncing | |
56 | * cacheline if we try to acquire the lock. So go onto the next slab. | |
57 | * If all pages are busy then we may allocate a new slab instead of reusing | |
58 | * a partial slab. A new slab has noone operating on it and thus there is | |
59 | * no danger of cacheline contention. | |
60 | * | |
61 | * Interrupts are disabled during allocation and deallocation in order to | |
62 | * make the slab allocator safe to use in the context of an irq. In addition | |
63 | * interrupts are disabled to ensure that the processor does not change | |
64 | * while handling per_cpu slabs, due to kernel preemption. | |
65 | * | |
66 | * SLUB assigns one slab for allocation to each processor. | |
67 | * Allocations only occur from these slabs called cpu slabs. | |
68 | * | |
672bba3a CL |
69 | * Slabs with free elements are kept on a partial list and during regular |
70 | * operations no list for full slabs is used. If an object in a full slab is | |
81819f0f | 71 | * freed then the slab will show up again on the partial lists. |
672bba3a CL |
72 | * We track full slabs for debugging purposes though because otherwise we |
73 | * cannot scan all objects. | |
81819f0f CL |
74 | * |
75 | * Slabs are freed when they become empty. Teardown and setup is | |
76 | * minimal so we rely on the page allocators per cpu caches for | |
77 | * fast frees and allocs. | |
78 | * | |
79 | * Overloading of page flags that are otherwise used for LRU management. | |
80 | * | |
4b6f0750 CL |
81 | * PageActive The slab is frozen and exempt from list processing. |
82 | * This means that the slab is dedicated to a purpose | |
83 | * such as satisfying allocations for a specific | |
84 | * processor. Objects may be freed in the slab while | |
85 | * it is frozen but slab_free will then skip the usual | |
86 | * list operations. It is up to the processor holding | |
87 | * the slab to integrate the slab into the slab lists | |
88 | * when the slab is no longer needed. | |
89 | * | |
90 | * One use of this flag is to mark slabs that are | |
91 | * used for allocations. Then such a slab becomes a cpu | |
92 | * slab. The cpu slab may be equipped with an additional | |
894b8788 CL |
93 | * lockless_freelist that allows lockless access to |
94 | * free objects in addition to the regular freelist | |
95 | * that requires the slab lock. | |
81819f0f CL |
96 | * |
97 | * PageError Slab requires special handling due to debug | |
98 | * options set. This moves slab handling out of | |
894b8788 | 99 | * the fast path and disables lockless freelists. |
81819f0f CL |
100 | */ |
101 | ||
5577bd8a CL |
102 | #define FROZEN (1 << PG_active) |
103 | ||
104 | #ifdef CONFIG_SLUB_DEBUG | |
105 | #define SLABDEBUG (1 << PG_error) | |
106 | #else | |
107 | #define SLABDEBUG 0 | |
108 | #endif | |
109 | ||
4b6f0750 CL |
110 | static inline int SlabFrozen(struct page *page) |
111 | { | |
5577bd8a | 112 | return page->flags & FROZEN; |
4b6f0750 CL |
113 | } |
114 | ||
115 | static inline void SetSlabFrozen(struct page *page) | |
116 | { | |
5577bd8a | 117 | page->flags |= FROZEN; |
4b6f0750 CL |
118 | } |
119 | ||
120 | static inline void ClearSlabFrozen(struct page *page) | |
121 | { | |
5577bd8a | 122 | page->flags &= ~FROZEN; |
4b6f0750 CL |
123 | } |
124 | ||
35e5d7ee CL |
125 | static inline int SlabDebug(struct page *page) |
126 | { | |
5577bd8a | 127 | return page->flags & SLABDEBUG; |
35e5d7ee CL |
128 | } |
129 | ||
130 | static inline void SetSlabDebug(struct page *page) | |
131 | { | |
5577bd8a | 132 | page->flags |= SLABDEBUG; |
35e5d7ee CL |
133 | } |
134 | ||
135 | static inline void ClearSlabDebug(struct page *page) | |
136 | { | |
5577bd8a | 137 | page->flags &= ~SLABDEBUG; |
35e5d7ee CL |
138 | } |
139 | ||
81819f0f CL |
140 | /* |
141 | * Issues still to be resolved: | |
142 | * | |
143 | * - The per cpu array is updated for each new slab and and is a remote | |
144 | * cacheline for most nodes. This could become a bouncing cacheline given | |
672bba3a CL |
145 | * enough frequent updates. There are 16 pointers in a cacheline, so at |
146 | * max 16 cpus could compete for the cacheline which may be okay. | |
81819f0f CL |
147 | * |
148 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
149 | * | |
81819f0f CL |
150 | * - Variable sizing of the per node arrays |
151 | */ | |
152 | ||
153 | /* Enable to test recovery from slab corruption on boot */ | |
154 | #undef SLUB_RESILIENCY_TEST | |
155 | ||
156 | #if PAGE_SHIFT <= 12 | |
157 | ||
158 | /* | |
159 | * Small page size. Make sure that we do not fragment memory | |
160 | */ | |
161 | #define DEFAULT_MAX_ORDER 1 | |
162 | #define DEFAULT_MIN_OBJECTS 4 | |
163 | ||
164 | #else | |
165 | ||
166 | /* | |
167 | * Large page machines are customarily able to handle larger | |
168 | * page orders. | |
169 | */ | |
170 | #define DEFAULT_MAX_ORDER 2 | |
171 | #define DEFAULT_MIN_OBJECTS 8 | |
172 | ||
173 | #endif | |
174 | ||
2086d26a CL |
175 | /* |
176 | * Mininum number of partial slabs. These will be left on the partial | |
177 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
178 | */ | |
e95eed57 CL |
179 | #define MIN_PARTIAL 2 |
180 | ||
2086d26a CL |
181 | /* |
182 | * Maximum number of desirable partial slabs. | |
183 | * The existence of more partial slabs makes kmem_cache_shrink | |
184 | * sort the partial list by the number of objects in the. | |
185 | */ | |
186 | #define MAX_PARTIAL 10 | |
187 | ||
81819f0f CL |
188 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ |
189 | SLAB_POISON | SLAB_STORE_USER) | |
672bba3a | 190 | |
81819f0f CL |
191 | /* |
192 | * Set of flags that will prevent slab merging | |
193 | */ | |
194 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
195 | SLAB_TRACE | SLAB_DESTROY_BY_RCU) | |
196 | ||
197 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ | |
198 | SLAB_CACHE_DMA) | |
199 | ||
200 | #ifndef ARCH_KMALLOC_MINALIGN | |
47bfdc0d | 201 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
202 | #endif |
203 | ||
204 | #ifndef ARCH_SLAB_MINALIGN | |
47bfdc0d | 205 | #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
206 | #endif |
207 | ||
6300ea75 CL |
208 | /* |
209 | * The page->inuse field is 16 bit thus we have this limitation | |
210 | */ | |
211 | #define MAX_OBJECTS_PER_SLAB 65535 | |
212 | ||
81819f0f CL |
213 | /* Internal SLUB flags */ |
214 | #define __OBJECT_POISON 0x80000000 /* Poison object */ | |
215 | ||
65c02d4c CL |
216 | /* Not all arches define cache_line_size */ |
217 | #ifndef cache_line_size | |
218 | #define cache_line_size() L1_CACHE_BYTES | |
219 | #endif | |
220 | ||
81819f0f CL |
221 | static int kmem_size = sizeof(struct kmem_cache); |
222 | ||
223 | #ifdef CONFIG_SMP | |
224 | static struct notifier_block slab_notifier; | |
225 | #endif | |
226 | ||
227 | static enum { | |
228 | DOWN, /* No slab functionality available */ | |
229 | PARTIAL, /* kmem_cache_open() works but kmalloc does not */ | |
672bba3a | 230 | UP, /* Everything works but does not show up in sysfs */ |
81819f0f CL |
231 | SYSFS /* Sysfs up */ |
232 | } slab_state = DOWN; | |
233 | ||
234 | /* A list of all slab caches on the system */ | |
235 | static DECLARE_RWSEM(slub_lock); | |
5af328a5 | 236 | static LIST_HEAD(slab_caches); |
81819f0f | 237 | |
02cbc874 CL |
238 | /* |
239 | * Tracking user of a slab. | |
240 | */ | |
241 | struct track { | |
242 | void *addr; /* Called from address */ | |
243 | int cpu; /* Was running on cpu */ | |
244 | int pid; /* Pid context */ | |
245 | unsigned long when; /* When did the operation occur */ | |
246 | }; | |
247 | ||
248 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
249 | ||
41ecc55b | 250 | #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) |
81819f0f CL |
251 | static int sysfs_slab_add(struct kmem_cache *); |
252 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
253 | static void sysfs_slab_remove(struct kmem_cache *); | |
254 | #else | |
0c710013 CL |
255 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } |
256 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) | |
257 | { return 0; } | |
258 | static inline void sysfs_slab_remove(struct kmem_cache *s) {} | |
81819f0f CL |
259 | #endif |
260 | ||
261 | /******************************************************************** | |
262 | * Core slab cache functions | |
263 | *******************************************************************/ | |
264 | ||
265 | int slab_is_available(void) | |
266 | { | |
267 | return slab_state >= UP; | |
268 | } | |
269 | ||
270 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
271 | { | |
272 | #ifdef CONFIG_NUMA | |
273 | return s->node[node]; | |
274 | #else | |
275 | return &s->local_node; | |
276 | #endif | |
277 | } | |
278 | ||
02cbc874 CL |
279 | static inline int check_valid_pointer(struct kmem_cache *s, |
280 | struct page *page, const void *object) | |
281 | { | |
282 | void *base; | |
283 | ||
284 | if (!object) | |
285 | return 1; | |
286 | ||
287 | base = page_address(page); | |
288 | if (object < base || object >= base + s->objects * s->size || | |
289 | (object - base) % s->size) { | |
290 | return 0; | |
291 | } | |
292 | ||
293 | return 1; | |
294 | } | |
295 | ||
7656c72b CL |
296 | /* |
297 | * Slow version of get and set free pointer. | |
298 | * | |
299 | * This version requires touching the cache lines of kmem_cache which | |
300 | * we avoid to do in the fast alloc free paths. There we obtain the offset | |
301 | * from the page struct. | |
302 | */ | |
303 | static inline void *get_freepointer(struct kmem_cache *s, void *object) | |
304 | { | |
305 | return *(void **)(object + s->offset); | |
306 | } | |
307 | ||
308 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
309 | { | |
310 | *(void **)(object + s->offset) = fp; | |
311 | } | |
312 | ||
313 | /* Loop over all objects in a slab */ | |
314 | #define for_each_object(__p, __s, __addr) \ | |
315 | for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\ | |
316 | __p += (__s)->size) | |
317 | ||
318 | /* Scan freelist */ | |
319 | #define for_each_free_object(__p, __s, __free) \ | |
320 | for (__p = (__free); __p; __p = get_freepointer((__s), __p)) | |
321 | ||
322 | /* Determine object index from a given position */ | |
323 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) | |
324 | { | |
325 | return (p - addr) / s->size; | |
326 | } | |
327 | ||
41ecc55b CL |
328 | #ifdef CONFIG_SLUB_DEBUG |
329 | /* | |
330 | * Debug settings: | |
331 | */ | |
f0630fff CL |
332 | #ifdef CONFIG_SLUB_DEBUG_ON |
333 | static int slub_debug = DEBUG_DEFAULT_FLAGS; | |
334 | #else | |
41ecc55b | 335 | static int slub_debug; |
f0630fff | 336 | #endif |
41ecc55b CL |
337 | |
338 | static char *slub_debug_slabs; | |
339 | ||
81819f0f CL |
340 | /* |
341 | * Object debugging | |
342 | */ | |
343 | static void print_section(char *text, u8 *addr, unsigned int length) | |
344 | { | |
345 | int i, offset; | |
346 | int newline = 1; | |
347 | char ascii[17]; | |
348 | ||
349 | ascii[16] = 0; | |
350 | ||
351 | for (i = 0; i < length; i++) { | |
352 | if (newline) { | |
24922684 | 353 | printk(KERN_ERR "%8s 0x%p: ", text, addr + i); |
81819f0f CL |
354 | newline = 0; |
355 | } | |
356 | printk(" %02x", addr[i]); | |
357 | offset = i % 16; | |
358 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; | |
359 | if (offset == 15) { | |
360 | printk(" %s\n",ascii); | |
361 | newline = 1; | |
362 | } | |
363 | } | |
364 | if (!newline) { | |
365 | i %= 16; | |
366 | while (i < 16) { | |
367 | printk(" "); | |
368 | ascii[i] = ' '; | |
369 | i++; | |
370 | } | |
371 | printk(" %s\n", ascii); | |
372 | } | |
373 | } | |
374 | ||
81819f0f CL |
375 | static struct track *get_track(struct kmem_cache *s, void *object, |
376 | enum track_item alloc) | |
377 | { | |
378 | struct track *p; | |
379 | ||
380 | if (s->offset) | |
381 | p = object + s->offset + sizeof(void *); | |
382 | else | |
383 | p = object + s->inuse; | |
384 | ||
385 | return p + alloc; | |
386 | } | |
387 | ||
388 | static void set_track(struct kmem_cache *s, void *object, | |
389 | enum track_item alloc, void *addr) | |
390 | { | |
391 | struct track *p; | |
392 | ||
393 | if (s->offset) | |
394 | p = object + s->offset + sizeof(void *); | |
395 | else | |
396 | p = object + s->inuse; | |
397 | ||
398 | p += alloc; | |
399 | if (addr) { | |
400 | p->addr = addr; | |
401 | p->cpu = smp_processor_id(); | |
402 | p->pid = current ? current->pid : -1; | |
403 | p->when = jiffies; | |
404 | } else | |
405 | memset(p, 0, sizeof(struct track)); | |
406 | } | |
407 | ||
81819f0f CL |
408 | static void init_tracking(struct kmem_cache *s, void *object) |
409 | { | |
24922684 CL |
410 | if (!(s->flags & SLAB_STORE_USER)) |
411 | return; | |
412 | ||
413 | set_track(s, object, TRACK_FREE, NULL); | |
414 | set_track(s, object, TRACK_ALLOC, NULL); | |
81819f0f CL |
415 | } |
416 | ||
417 | static void print_track(const char *s, struct track *t) | |
418 | { | |
419 | if (!t->addr) | |
420 | return; | |
421 | ||
24922684 | 422 | printk(KERN_ERR "INFO: %s in ", s); |
81819f0f | 423 | __print_symbol("%s", (unsigned long)t->addr); |
24922684 CL |
424 | printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); |
425 | } | |
426 | ||
427 | static void print_tracking(struct kmem_cache *s, void *object) | |
428 | { | |
429 | if (!(s->flags & SLAB_STORE_USER)) | |
430 | return; | |
431 | ||
432 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); | |
433 | print_track("Freed", get_track(s, object, TRACK_FREE)); | |
434 | } | |
435 | ||
436 | static void print_page_info(struct page *page) | |
437 | { | |
438 | printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n", | |
439 | page, page->inuse, page->freelist, page->flags); | |
440 | ||
441 | } | |
442 | ||
443 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) | |
444 | { | |
445 | va_list args; | |
446 | char buf[100]; | |
447 | ||
448 | va_start(args, fmt); | |
449 | vsnprintf(buf, sizeof(buf), fmt, args); | |
450 | va_end(args); | |
451 | printk(KERN_ERR "========================================" | |
452 | "=====================================\n"); | |
453 | printk(KERN_ERR "BUG %s: %s\n", s->name, buf); | |
454 | printk(KERN_ERR "----------------------------------------" | |
455 | "-------------------------------------\n\n"); | |
81819f0f CL |
456 | } |
457 | ||
24922684 CL |
458 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) |
459 | { | |
460 | va_list args; | |
461 | char buf[100]; | |
462 | ||
463 | va_start(args, fmt); | |
464 | vsnprintf(buf, sizeof(buf), fmt, args); | |
465 | va_end(args); | |
466 | printk(KERN_ERR "FIX %s: %s\n", s->name, buf); | |
467 | } | |
468 | ||
469 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) | |
81819f0f CL |
470 | { |
471 | unsigned int off; /* Offset of last byte */ | |
24922684 CL |
472 | u8 *addr = page_address(page); |
473 | ||
474 | print_tracking(s, p); | |
475 | ||
476 | print_page_info(page); | |
477 | ||
478 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", | |
479 | p, p - addr, get_freepointer(s, p)); | |
480 | ||
481 | if (p > addr + 16) | |
482 | print_section("Bytes b4", p - 16, 16); | |
483 | ||
484 | print_section("Object", p, min(s->objsize, 128)); | |
81819f0f CL |
485 | |
486 | if (s->flags & SLAB_RED_ZONE) | |
487 | print_section("Redzone", p + s->objsize, | |
488 | s->inuse - s->objsize); | |
489 | ||
81819f0f CL |
490 | if (s->offset) |
491 | off = s->offset + sizeof(void *); | |
492 | else | |
493 | off = s->inuse; | |
494 | ||
24922684 | 495 | if (s->flags & SLAB_STORE_USER) |
81819f0f | 496 | off += 2 * sizeof(struct track); |
81819f0f CL |
497 | |
498 | if (off != s->size) | |
499 | /* Beginning of the filler is the free pointer */ | |
24922684 CL |
500 | print_section("Padding", p + off, s->size - off); |
501 | ||
502 | dump_stack(); | |
81819f0f CL |
503 | } |
504 | ||
505 | static void object_err(struct kmem_cache *s, struct page *page, | |
506 | u8 *object, char *reason) | |
507 | { | |
24922684 CL |
508 | slab_bug(s, reason); |
509 | print_trailer(s, page, object); | |
81819f0f CL |
510 | } |
511 | ||
24922684 | 512 | static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) |
81819f0f CL |
513 | { |
514 | va_list args; | |
515 | char buf[100]; | |
516 | ||
24922684 CL |
517 | va_start(args, fmt); |
518 | vsnprintf(buf, sizeof(buf), fmt, args); | |
81819f0f | 519 | va_end(args); |
24922684 CL |
520 | slab_bug(s, fmt); |
521 | print_page_info(page); | |
81819f0f CL |
522 | dump_stack(); |
523 | } | |
524 | ||
525 | static void init_object(struct kmem_cache *s, void *object, int active) | |
526 | { | |
527 | u8 *p = object; | |
528 | ||
529 | if (s->flags & __OBJECT_POISON) { | |
530 | memset(p, POISON_FREE, s->objsize - 1); | |
531 | p[s->objsize -1] = POISON_END; | |
532 | } | |
533 | ||
534 | if (s->flags & SLAB_RED_ZONE) | |
535 | memset(p + s->objsize, | |
536 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, | |
537 | s->inuse - s->objsize); | |
538 | } | |
539 | ||
24922684 | 540 | static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) |
81819f0f CL |
541 | { |
542 | while (bytes) { | |
543 | if (*start != (u8)value) | |
24922684 | 544 | return start; |
81819f0f CL |
545 | start++; |
546 | bytes--; | |
547 | } | |
24922684 CL |
548 | return NULL; |
549 | } | |
550 | ||
551 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
552 | void *from, void *to) | |
553 | { | |
554 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); | |
555 | memset(from, data, to - from); | |
556 | } | |
557 | ||
558 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, | |
559 | u8 *object, char *what, | |
560 | u8* start, unsigned int value, unsigned int bytes) | |
561 | { | |
562 | u8 *fault; | |
563 | u8 *end; | |
564 | ||
565 | fault = check_bytes(start, value, bytes); | |
566 | if (!fault) | |
567 | return 1; | |
568 | ||
569 | end = start + bytes; | |
570 | while (end > fault && end[-1] == value) | |
571 | end--; | |
572 | ||
573 | slab_bug(s, "%s overwritten", what); | |
574 | printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", | |
575 | fault, end - 1, fault[0], value); | |
576 | print_trailer(s, page, object); | |
577 | ||
578 | restore_bytes(s, what, value, fault, end); | |
579 | return 0; | |
81819f0f CL |
580 | } |
581 | ||
81819f0f CL |
582 | /* |
583 | * Object layout: | |
584 | * | |
585 | * object address | |
586 | * Bytes of the object to be managed. | |
587 | * If the freepointer may overlay the object then the free | |
588 | * pointer is the first word of the object. | |
672bba3a | 589 | * |
81819f0f CL |
590 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
591 | * 0xa5 (POISON_END) | |
592 | * | |
593 | * object + s->objsize | |
594 | * Padding to reach word boundary. This is also used for Redzoning. | |
672bba3a CL |
595 | * Padding is extended by another word if Redzoning is enabled and |
596 | * objsize == inuse. | |
597 | * | |
81819f0f CL |
598 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
599 | * 0xcc (RED_ACTIVE) for objects in use. | |
600 | * | |
601 | * object + s->inuse | |
672bba3a CL |
602 | * Meta data starts here. |
603 | * | |
81819f0f CL |
604 | * A. Free pointer (if we cannot overwrite object on free) |
605 | * B. Tracking data for SLAB_STORE_USER | |
672bba3a CL |
606 | * C. Padding to reach required alignment boundary or at mininum |
607 | * one word if debuggin is on to be able to detect writes | |
608 | * before the word boundary. | |
609 | * | |
610 | * Padding is done using 0x5a (POISON_INUSE) | |
81819f0f CL |
611 | * |
612 | * object + s->size | |
672bba3a | 613 | * Nothing is used beyond s->size. |
81819f0f | 614 | * |
672bba3a CL |
615 | * If slabcaches are merged then the objsize and inuse boundaries are mostly |
616 | * ignored. And therefore no slab options that rely on these boundaries | |
81819f0f CL |
617 | * may be used with merged slabcaches. |
618 | */ | |
619 | ||
81819f0f CL |
620 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) |
621 | { | |
622 | unsigned long off = s->inuse; /* The end of info */ | |
623 | ||
624 | if (s->offset) | |
625 | /* Freepointer is placed after the object. */ | |
626 | off += sizeof(void *); | |
627 | ||
628 | if (s->flags & SLAB_STORE_USER) | |
629 | /* We also have user information there */ | |
630 | off += 2 * sizeof(struct track); | |
631 | ||
632 | if (s->size == off) | |
633 | return 1; | |
634 | ||
24922684 CL |
635 | return check_bytes_and_report(s, page, p, "Object padding", |
636 | p + off, POISON_INUSE, s->size - off); | |
81819f0f CL |
637 | } |
638 | ||
639 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
640 | { | |
24922684 CL |
641 | u8 *start; |
642 | u8 *fault; | |
643 | u8 *end; | |
644 | int length; | |
645 | int remainder; | |
81819f0f CL |
646 | |
647 | if (!(s->flags & SLAB_POISON)) | |
648 | return 1; | |
649 | ||
24922684 CL |
650 | start = page_address(page); |
651 | end = start + (PAGE_SIZE << s->order); | |
81819f0f | 652 | length = s->objects * s->size; |
24922684 | 653 | remainder = end - (start + length); |
81819f0f CL |
654 | if (!remainder) |
655 | return 1; | |
656 | ||
24922684 CL |
657 | fault = check_bytes(start + length, POISON_INUSE, remainder); |
658 | if (!fault) | |
659 | return 1; | |
660 | while (end > fault && end[-1] == POISON_INUSE) | |
661 | end--; | |
662 | ||
663 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); | |
664 | print_section("Padding", start, length); | |
665 | ||
666 | restore_bytes(s, "slab padding", POISON_INUSE, start, end); | |
667 | return 0; | |
81819f0f CL |
668 | } |
669 | ||
670 | static int check_object(struct kmem_cache *s, struct page *page, | |
671 | void *object, int active) | |
672 | { | |
673 | u8 *p = object; | |
674 | u8 *endobject = object + s->objsize; | |
675 | ||
676 | if (s->flags & SLAB_RED_ZONE) { | |
677 | unsigned int red = | |
678 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; | |
679 | ||
24922684 CL |
680 | if (!check_bytes_and_report(s, page, object, "Redzone", |
681 | endobject, red, s->inuse - s->objsize)) | |
81819f0f | 682 | return 0; |
81819f0f | 683 | } else { |
24922684 CL |
684 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) |
685 | check_bytes_and_report(s, page, p, "Alignment padding", endobject, | |
686 | POISON_INUSE, s->inuse - s->objsize); | |
81819f0f CL |
687 | } |
688 | ||
689 | if (s->flags & SLAB_POISON) { | |
690 | if (!active && (s->flags & __OBJECT_POISON) && | |
24922684 CL |
691 | (!check_bytes_and_report(s, page, p, "Poison", p, |
692 | POISON_FREE, s->objsize - 1) || | |
693 | !check_bytes_and_report(s, page, p, "Poison", | |
694 | p + s->objsize -1, POISON_END, 1))) | |
81819f0f | 695 | return 0; |
81819f0f CL |
696 | /* |
697 | * check_pad_bytes cleans up on its own. | |
698 | */ | |
699 | check_pad_bytes(s, page, p); | |
700 | } | |
701 | ||
702 | if (!s->offset && active) | |
703 | /* | |
704 | * Object and freepointer overlap. Cannot check | |
705 | * freepointer while object is allocated. | |
706 | */ | |
707 | return 1; | |
708 | ||
709 | /* Check free pointer validity */ | |
710 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
711 | object_err(s, page, p, "Freepointer corrupt"); | |
712 | /* | |
713 | * No choice but to zap it and thus loose the remainder | |
714 | * of the free objects in this slab. May cause | |
672bba3a | 715 | * another error because the object count is now wrong. |
81819f0f CL |
716 | */ |
717 | set_freepointer(s, p, NULL); | |
718 | return 0; | |
719 | } | |
720 | return 1; | |
721 | } | |
722 | ||
723 | static int check_slab(struct kmem_cache *s, struct page *page) | |
724 | { | |
725 | VM_BUG_ON(!irqs_disabled()); | |
726 | ||
727 | if (!PageSlab(page)) { | |
24922684 | 728 | slab_err(s, page, "Not a valid slab page"); |
81819f0f CL |
729 | return 0; |
730 | } | |
731 | if (page->offset * sizeof(void *) != s->offset) { | |
24922684 CL |
732 | slab_err(s, page, "Corrupted offset %lu", |
733 | (unsigned long)(page->offset * sizeof(void *))); | |
81819f0f CL |
734 | return 0; |
735 | } | |
736 | if (page->inuse > s->objects) { | |
24922684 CL |
737 | slab_err(s, page, "inuse %u > max %u", |
738 | s->name, page->inuse, s->objects); | |
81819f0f CL |
739 | return 0; |
740 | } | |
741 | /* Slab_pad_check fixes things up after itself */ | |
742 | slab_pad_check(s, page); | |
743 | return 1; | |
744 | } | |
745 | ||
746 | /* | |
672bba3a CL |
747 | * Determine if a certain object on a page is on the freelist. Must hold the |
748 | * slab lock to guarantee that the chains are in a consistent state. | |
81819f0f CL |
749 | */ |
750 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
751 | { | |
752 | int nr = 0; | |
753 | void *fp = page->freelist; | |
754 | void *object = NULL; | |
755 | ||
756 | while (fp && nr <= s->objects) { | |
757 | if (fp == search) | |
758 | return 1; | |
759 | if (!check_valid_pointer(s, page, fp)) { | |
760 | if (object) { | |
761 | object_err(s, page, object, | |
762 | "Freechain corrupt"); | |
763 | set_freepointer(s, object, NULL); | |
764 | break; | |
765 | } else { | |
24922684 | 766 | slab_err(s, page, "Freepointer corrupt"); |
81819f0f CL |
767 | page->freelist = NULL; |
768 | page->inuse = s->objects; | |
24922684 | 769 | slab_fix(s, "Freelist cleared"); |
81819f0f CL |
770 | return 0; |
771 | } | |
772 | break; | |
773 | } | |
774 | object = fp; | |
775 | fp = get_freepointer(s, object); | |
776 | nr++; | |
777 | } | |
778 | ||
779 | if (page->inuse != s->objects - nr) { | |
70d71228 | 780 | slab_err(s, page, "Wrong object count. Counter is %d but " |
24922684 | 781 | "counted were %d", page->inuse, s->objects - nr); |
81819f0f | 782 | page->inuse = s->objects - nr; |
24922684 | 783 | slab_fix(s, "Object count adjusted."); |
81819f0f CL |
784 | } |
785 | return search == NULL; | |
786 | } | |
787 | ||
3ec09742 CL |
788 | static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc) |
789 | { | |
790 | if (s->flags & SLAB_TRACE) { | |
791 | printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", | |
792 | s->name, | |
793 | alloc ? "alloc" : "free", | |
794 | object, page->inuse, | |
795 | page->freelist); | |
796 | ||
797 | if (!alloc) | |
798 | print_section("Object", (void *)object, s->objsize); | |
799 | ||
800 | dump_stack(); | |
801 | } | |
802 | } | |
803 | ||
643b1138 | 804 | /* |
672bba3a | 805 | * Tracking of fully allocated slabs for debugging purposes. |
643b1138 | 806 | */ |
e95eed57 | 807 | static void add_full(struct kmem_cache_node *n, struct page *page) |
643b1138 | 808 | { |
643b1138 CL |
809 | spin_lock(&n->list_lock); |
810 | list_add(&page->lru, &n->full); | |
811 | spin_unlock(&n->list_lock); | |
812 | } | |
813 | ||
814 | static void remove_full(struct kmem_cache *s, struct page *page) | |
815 | { | |
816 | struct kmem_cache_node *n; | |
817 | ||
818 | if (!(s->flags & SLAB_STORE_USER)) | |
819 | return; | |
820 | ||
821 | n = get_node(s, page_to_nid(page)); | |
822 | ||
823 | spin_lock(&n->list_lock); | |
824 | list_del(&page->lru); | |
825 | spin_unlock(&n->list_lock); | |
826 | } | |
827 | ||
3ec09742 CL |
828 | static void setup_object_debug(struct kmem_cache *s, struct page *page, |
829 | void *object) | |
830 | { | |
831 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) | |
832 | return; | |
833 | ||
834 | init_object(s, object, 0); | |
835 | init_tracking(s, object); | |
836 | } | |
837 | ||
838 | static int alloc_debug_processing(struct kmem_cache *s, struct page *page, | |
839 | void *object, void *addr) | |
81819f0f CL |
840 | { |
841 | if (!check_slab(s, page)) | |
842 | goto bad; | |
843 | ||
844 | if (object && !on_freelist(s, page, object)) { | |
24922684 | 845 | object_err(s, page, object, "Object already allocated"); |
70d71228 | 846 | goto bad; |
81819f0f CL |
847 | } |
848 | ||
849 | if (!check_valid_pointer(s, page, object)) { | |
850 | object_err(s, page, object, "Freelist Pointer check fails"); | |
70d71228 | 851 | goto bad; |
81819f0f CL |
852 | } |
853 | ||
3ec09742 | 854 | if (object && !check_object(s, page, object, 0)) |
81819f0f | 855 | goto bad; |
81819f0f | 856 | |
3ec09742 CL |
857 | /* Success perform special debug activities for allocs */ |
858 | if (s->flags & SLAB_STORE_USER) | |
859 | set_track(s, object, TRACK_ALLOC, addr); | |
860 | trace(s, page, object, 1); | |
861 | init_object(s, object, 1); | |
81819f0f | 862 | return 1; |
3ec09742 | 863 | |
81819f0f CL |
864 | bad: |
865 | if (PageSlab(page)) { | |
866 | /* | |
867 | * If this is a slab page then lets do the best we can | |
868 | * to avoid issues in the future. Marking all objects | |
672bba3a | 869 | * as used avoids touching the remaining objects. |
81819f0f | 870 | */ |
24922684 | 871 | slab_fix(s, "Marking all objects used"); |
81819f0f CL |
872 | page->inuse = s->objects; |
873 | page->freelist = NULL; | |
874 | /* Fix up fields that may be corrupted */ | |
875 | page->offset = s->offset / sizeof(void *); | |
876 | } | |
877 | return 0; | |
878 | } | |
879 | ||
3ec09742 CL |
880 | static int free_debug_processing(struct kmem_cache *s, struct page *page, |
881 | void *object, void *addr) | |
81819f0f CL |
882 | { |
883 | if (!check_slab(s, page)) | |
884 | goto fail; | |
885 | ||
886 | if (!check_valid_pointer(s, page, object)) { | |
70d71228 | 887 | slab_err(s, page, "Invalid object pointer 0x%p", object); |
81819f0f CL |
888 | goto fail; |
889 | } | |
890 | ||
891 | if (on_freelist(s, page, object)) { | |
24922684 | 892 | object_err(s, page, object, "Object already free"); |
81819f0f CL |
893 | goto fail; |
894 | } | |
895 | ||
896 | if (!check_object(s, page, object, 1)) | |
897 | return 0; | |
898 | ||
899 | if (unlikely(s != page->slab)) { | |
900 | if (!PageSlab(page)) | |
70d71228 CL |
901 | slab_err(s, page, "Attempt to free object(0x%p) " |
902 | "outside of slab", object); | |
81819f0f | 903 | else |
70d71228 | 904 | if (!page->slab) { |
81819f0f | 905 | printk(KERN_ERR |
70d71228 | 906 | "SLUB <none>: no slab for object 0x%p.\n", |
81819f0f | 907 | object); |
70d71228 CL |
908 | dump_stack(); |
909 | } | |
81819f0f | 910 | else |
24922684 CL |
911 | object_err(s, page, object, |
912 | "page slab pointer corrupt."); | |
81819f0f CL |
913 | goto fail; |
914 | } | |
3ec09742 CL |
915 | |
916 | /* Special debug activities for freeing objects */ | |
917 | if (!SlabFrozen(page) && !page->freelist) | |
918 | remove_full(s, page); | |
919 | if (s->flags & SLAB_STORE_USER) | |
920 | set_track(s, object, TRACK_FREE, addr); | |
921 | trace(s, page, object, 0); | |
922 | init_object(s, object, 0); | |
81819f0f | 923 | return 1; |
3ec09742 | 924 | |
81819f0f | 925 | fail: |
24922684 | 926 | slab_fix(s, "Object at 0x%p not freed", object); |
81819f0f CL |
927 | return 0; |
928 | } | |
929 | ||
41ecc55b CL |
930 | static int __init setup_slub_debug(char *str) |
931 | { | |
f0630fff CL |
932 | slub_debug = DEBUG_DEFAULT_FLAGS; |
933 | if (*str++ != '=' || !*str) | |
934 | /* | |
935 | * No options specified. Switch on full debugging. | |
936 | */ | |
937 | goto out; | |
938 | ||
939 | if (*str == ',') | |
940 | /* | |
941 | * No options but restriction on slabs. This means full | |
942 | * debugging for slabs matching a pattern. | |
943 | */ | |
944 | goto check_slabs; | |
945 | ||
946 | slub_debug = 0; | |
947 | if (*str == '-') | |
948 | /* | |
949 | * Switch off all debugging measures. | |
950 | */ | |
951 | goto out; | |
952 | ||
953 | /* | |
954 | * Determine which debug features should be switched on | |
955 | */ | |
956 | for ( ;*str && *str != ','; str++) { | |
957 | switch (tolower(*str)) { | |
958 | case 'f': | |
959 | slub_debug |= SLAB_DEBUG_FREE; | |
960 | break; | |
961 | case 'z': | |
962 | slub_debug |= SLAB_RED_ZONE; | |
963 | break; | |
964 | case 'p': | |
965 | slub_debug |= SLAB_POISON; | |
966 | break; | |
967 | case 'u': | |
968 | slub_debug |= SLAB_STORE_USER; | |
969 | break; | |
970 | case 't': | |
971 | slub_debug |= SLAB_TRACE; | |
972 | break; | |
973 | default: | |
974 | printk(KERN_ERR "slub_debug option '%c' " | |
975 | "unknown. skipped\n",*str); | |
976 | } | |
41ecc55b CL |
977 | } |
978 | ||
f0630fff | 979 | check_slabs: |
41ecc55b CL |
980 | if (*str == ',') |
981 | slub_debug_slabs = str + 1; | |
f0630fff | 982 | out: |
41ecc55b CL |
983 | return 1; |
984 | } | |
985 | ||
986 | __setup("slub_debug", setup_slub_debug); | |
987 | ||
988 | static void kmem_cache_open_debug_check(struct kmem_cache *s) | |
989 | { | |
990 | /* | |
991 | * The page->offset field is only 16 bit wide. This is an offset | |
992 | * in units of words from the beginning of an object. If the slab | |
993 | * size is bigger then we cannot move the free pointer behind the | |
994 | * object anymore. | |
995 | * | |
996 | * On 32 bit platforms the limit is 256k. On 64bit platforms | |
997 | * the limit is 512k. | |
998 | * | |
c59def9f | 999 | * Debugging or ctor may create a need to move the free |
41ecc55b CL |
1000 | * pointer. Fail if this happens. |
1001 | */ | |
33e9e241 | 1002 | if (s->objsize >= 65535 * sizeof(void *)) { |
41ecc55b CL |
1003 | BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON | |
1004 | SLAB_STORE_USER | SLAB_DESTROY_BY_RCU)); | |
c59def9f | 1005 | BUG_ON(s->ctor); |
41ecc55b CL |
1006 | } |
1007 | else | |
1008 | /* | |
1009 | * Enable debugging if selected on the kernel commandline. | |
1010 | */ | |
1011 | if (slub_debug && (!slub_debug_slabs || | |
1012 | strncmp(slub_debug_slabs, s->name, | |
1013 | strlen(slub_debug_slabs)) == 0)) | |
1014 | s->flags |= slub_debug; | |
1015 | } | |
1016 | #else | |
3ec09742 CL |
1017 | static inline void setup_object_debug(struct kmem_cache *s, |
1018 | struct page *page, void *object) {} | |
41ecc55b | 1019 | |
3ec09742 CL |
1020 | static inline int alloc_debug_processing(struct kmem_cache *s, |
1021 | struct page *page, void *object, void *addr) { return 0; } | |
41ecc55b | 1022 | |
3ec09742 CL |
1023 | static inline int free_debug_processing(struct kmem_cache *s, |
1024 | struct page *page, void *object, void *addr) { return 0; } | |
41ecc55b | 1025 | |
41ecc55b CL |
1026 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) |
1027 | { return 1; } | |
1028 | static inline int check_object(struct kmem_cache *s, struct page *page, | |
1029 | void *object, int active) { return 1; } | |
3ec09742 | 1030 | static inline void add_full(struct kmem_cache_node *n, struct page *page) {} |
41ecc55b CL |
1031 | static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {} |
1032 | #define slub_debug 0 | |
1033 | #endif | |
81819f0f CL |
1034 | /* |
1035 | * Slab allocation and freeing | |
1036 | */ | |
1037 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1038 | { | |
1039 | struct page * page; | |
1040 | int pages = 1 << s->order; | |
1041 | ||
1042 | if (s->order) | |
1043 | flags |= __GFP_COMP; | |
1044 | ||
1045 | if (s->flags & SLAB_CACHE_DMA) | |
1046 | flags |= SLUB_DMA; | |
1047 | ||
1048 | if (node == -1) | |
1049 | page = alloc_pages(flags, s->order); | |
1050 | else | |
1051 | page = alloc_pages_node(node, flags, s->order); | |
1052 | ||
1053 | if (!page) | |
1054 | return NULL; | |
1055 | ||
1056 | mod_zone_page_state(page_zone(page), | |
1057 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1058 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1059 | pages); | |
1060 | ||
1061 | return page; | |
1062 | } | |
1063 | ||
1064 | static void setup_object(struct kmem_cache *s, struct page *page, | |
1065 | void *object) | |
1066 | { | |
3ec09742 | 1067 | setup_object_debug(s, page, object); |
4f104934 | 1068 | if (unlikely(s->ctor)) |
a35afb83 | 1069 | s->ctor(object, s, 0); |
81819f0f CL |
1070 | } |
1071 | ||
1072 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1073 | { | |
1074 | struct page *page; | |
1075 | struct kmem_cache_node *n; | |
1076 | void *start; | |
1077 | void *end; | |
1078 | void *last; | |
1079 | void *p; | |
1080 | ||
d07dbea4 | 1081 | BUG_ON(flags & ~(GFP_DMA | __GFP_ZERO | GFP_LEVEL_MASK)); |
81819f0f CL |
1082 | |
1083 | if (flags & __GFP_WAIT) | |
1084 | local_irq_enable(); | |
1085 | ||
1086 | page = allocate_slab(s, flags & GFP_LEVEL_MASK, node); | |
1087 | if (!page) | |
1088 | goto out; | |
1089 | ||
1090 | n = get_node(s, page_to_nid(page)); | |
1091 | if (n) | |
1092 | atomic_long_inc(&n->nr_slabs); | |
1093 | page->offset = s->offset / sizeof(void *); | |
1094 | page->slab = s; | |
1095 | page->flags |= 1 << PG_slab; | |
1096 | if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | | |
1097 | SLAB_STORE_USER | SLAB_TRACE)) | |
35e5d7ee | 1098 | SetSlabDebug(page); |
81819f0f CL |
1099 | |
1100 | start = page_address(page); | |
1101 | end = start + s->objects * s->size; | |
1102 | ||
1103 | if (unlikely(s->flags & SLAB_POISON)) | |
1104 | memset(start, POISON_INUSE, PAGE_SIZE << s->order); | |
1105 | ||
1106 | last = start; | |
7656c72b | 1107 | for_each_object(p, s, start) { |
81819f0f CL |
1108 | setup_object(s, page, last); |
1109 | set_freepointer(s, last, p); | |
1110 | last = p; | |
1111 | } | |
1112 | setup_object(s, page, last); | |
1113 | set_freepointer(s, last, NULL); | |
1114 | ||
1115 | page->freelist = start; | |
894b8788 | 1116 | page->lockless_freelist = NULL; |
81819f0f CL |
1117 | page->inuse = 0; |
1118 | out: | |
1119 | if (flags & __GFP_WAIT) | |
1120 | local_irq_disable(); | |
1121 | return page; | |
1122 | } | |
1123 | ||
1124 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
1125 | { | |
1126 | int pages = 1 << s->order; | |
1127 | ||
c59def9f | 1128 | if (unlikely(SlabDebug(page))) { |
81819f0f CL |
1129 | void *p; |
1130 | ||
1131 | slab_pad_check(s, page); | |
c59def9f | 1132 | for_each_object(p, s, page_address(page)) |
81819f0f | 1133 | check_object(s, page, p, 0); |
81819f0f CL |
1134 | } |
1135 | ||
1136 | mod_zone_page_state(page_zone(page), | |
1137 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1138 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1139 | - pages); | |
1140 | ||
1141 | page->mapping = NULL; | |
1142 | __free_pages(page, s->order); | |
1143 | } | |
1144 | ||
1145 | static void rcu_free_slab(struct rcu_head *h) | |
1146 | { | |
1147 | struct page *page; | |
1148 | ||
1149 | page = container_of((struct list_head *)h, struct page, lru); | |
1150 | __free_slab(page->slab, page); | |
1151 | } | |
1152 | ||
1153 | static void free_slab(struct kmem_cache *s, struct page *page) | |
1154 | { | |
1155 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
1156 | /* | |
1157 | * RCU free overloads the RCU head over the LRU | |
1158 | */ | |
1159 | struct rcu_head *head = (void *)&page->lru; | |
1160 | ||
1161 | call_rcu(head, rcu_free_slab); | |
1162 | } else | |
1163 | __free_slab(s, page); | |
1164 | } | |
1165 | ||
1166 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
1167 | { | |
1168 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1169 | ||
1170 | atomic_long_dec(&n->nr_slabs); | |
1171 | reset_page_mapcount(page); | |
35e5d7ee CL |
1172 | ClearSlabDebug(page); |
1173 | __ClearPageSlab(page); | |
81819f0f CL |
1174 | free_slab(s, page); |
1175 | } | |
1176 | ||
1177 | /* | |
1178 | * Per slab locking using the pagelock | |
1179 | */ | |
1180 | static __always_inline void slab_lock(struct page *page) | |
1181 | { | |
1182 | bit_spin_lock(PG_locked, &page->flags); | |
1183 | } | |
1184 | ||
1185 | static __always_inline void slab_unlock(struct page *page) | |
1186 | { | |
1187 | bit_spin_unlock(PG_locked, &page->flags); | |
1188 | } | |
1189 | ||
1190 | static __always_inline int slab_trylock(struct page *page) | |
1191 | { | |
1192 | int rc = 1; | |
1193 | ||
1194 | rc = bit_spin_trylock(PG_locked, &page->flags); | |
1195 | return rc; | |
1196 | } | |
1197 | ||
1198 | /* | |
1199 | * Management of partially allocated slabs | |
1200 | */ | |
e95eed57 | 1201 | static void add_partial_tail(struct kmem_cache_node *n, struct page *page) |
81819f0f | 1202 | { |
e95eed57 CL |
1203 | spin_lock(&n->list_lock); |
1204 | n->nr_partial++; | |
1205 | list_add_tail(&page->lru, &n->partial); | |
1206 | spin_unlock(&n->list_lock); | |
1207 | } | |
81819f0f | 1208 | |
e95eed57 CL |
1209 | static void add_partial(struct kmem_cache_node *n, struct page *page) |
1210 | { | |
81819f0f CL |
1211 | spin_lock(&n->list_lock); |
1212 | n->nr_partial++; | |
1213 | list_add(&page->lru, &n->partial); | |
1214 | spin_unlock(&n->list_lock); | |
1215 | } | |
1216 | ||
1217 | static void remove_partial(struct kmem_cache *s, | |
1218 | struct page *page) | |
1219 | { | |
1220 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1221 | ||
1222 | spin_lock(&n->list_lock); | |
1223 | list_del(&page->lru); | |
1224 | n->nr_partial--; | |
1225 | spin_unlock(&n->list_lock); | |
1226 | } | |
1227 | ||
1228 | /* | |
672bba3a | 1229 | * Lock slab and remove from the partial list. |
81819f0f | 1230 | * |
672bba3a | 1231 | * Must hold list_lock. |
81819f0f | 1232 | */ |
4b6f0750 | 1233 | static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page) |
81819f0f CL |
1234 | { |
1235 | if (slab_trylock(page)) { | |
1236 | list_del(&page->lru); | |
1237 | n->nr_partial--; | |
4b6f0750 | 1238 | SetSlabFrozen(page); |
81819f0f CL |
1239 | return 1; |
1240 | } | |
1241 | return 0; | |
1242 | } | |
1243 | ||
1244 | /* | |
672bba3a | 1245 | * Try to allocate a partial slab from a specific node. |
81819f0f CL |
1246 | */ |
1247 | static struct page *get_partial_node(struct kmem_cache_node *n) | |
1248 | { | |
1249 | struct page *page; | |
1250 | ||
1251 | /* | |
1252 | * Racy check. If we mistakenly see no partial slabs then we | |
1253 | * just allocate an empty slab. If we mistakenly try to get a | |
672bba3a CL |
1254 | * partial slab and there is none available then get_partials() |
1255 | * will return NULL. | |
81819f0f CL |
1256 | */ |
1257 | if (!n || !n->nr_partial) | |
1258 | return NULL; | |
1259 | ||
1260 | spin_lock(&n->list_lock); | |
1261 | list_for_each_entry(page, &n->partial, lru) | |
4b6f0750 | 1262 | if (lock_and_freeze_slab(n, page)) |
81819f0f CL |
1263 | goto out; |
1264 | page = NULL; | |
1265 | out: | |
1266 | spin_unlock(&n->list_lock); | |
1267 | return page; | |
1268 | } | |
1269 | ||
1270 | /* | |
672bba3a | 1271 | * Get a page from somewhere. Search in increasing NUMA distances. |
81819f0f CL |
1272 | */ |
1273 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) | |
1274 | { | |
1275 | #ifdef CONFIG_NUMA | |
1276 | struct zonelist *zonelist; | |
1277 | struct zone **z; | |
1278 | struct page *page; | |
1279 | ||
1280 | /* | |
672bba3a CL |
1281 | * The defrag ratio allows a configuration of the tradeoffs between |
1282 | * inter node defragmentation and node local allocations. A lower | |
1283 | * defrag_ratio increases the tendency to do local allocations | |
1284 | * instead of attempting to obtain partial slabs from other nodes. | |
81819f0f | 1285 | * |
672bba3a CL |
1286 | * If the defrag_ratio is set to 0 then kmalloc() always |
1287 | * returns node local objects. If the ratio is higher then kmalloc() | |
1288 | * may return off node objects because partial slabs are obtained | |
1289 | * from other nodes and filled up. | |
81819f0f CL |
1290 | * |
1291 | * If /sys/slab/xx/defrag_ratio is set to 100 (which makes | |
672bba3a CL |
1292 | * defrag_ratio = 1000) then every (well almost) allocation will |
1293 | * first attempt to defrag slab caches on other nodes. This means | |
1294 | * scanning over all nodes to look for partial slabs which may be | |
1295 | * expensive if we do it every time we are trying to find a slab | |
1296 | * with available objects. | |
81819f0f CL |
1297 | */ |
1298 | if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio) | |
1299 | return NULL; | |
1300 | ||
1301 | zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
1302 | ->node_zonelists[gfp_zone(flags)]; | |
1303 | for (z = zonelist->zones; *z; z++) { | |
1304 | struct kmem_cache_node *n; | |
1305 | ||
1306 | n = get_node(s, zone_to_nid(*z)); | |
1307 | ||
1308 | if (n && cpuset_zone_allowed_hardwall(*z, flags) && | |
e95eed57 | 1309 | n->nr_partial > MIN_PARTIAL) { |
81819f0f CL |
1310 | page = get_partial_node(n); |
1311 | if (page) | |
1312 | return page; | |
1313 | } | |
1314 | } | |
1315 | #endif | |
1316 | return NULL; | |
1317 | } | |
1318 | ||
1319 | /* | |
1320 | * Get a partial page, lock it and return it. | |
1321 | */ | |
1322 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) | |
1323 | { | |
1324 | struct page *page; | |
1325 | int searchnode = (node == -1) ? numa_node_id() : node; | |
1326 | ||
1327 | page = get_partial_node(get_node(s, searchnode)); | |
1328 | if (page || (flags & __GFP_THISNODE)) | |
1329 | return page; | |
1330 | ||
1331 | return get_any_partial(s, flags); | |
1332 | } | |
1333 | ||
1334 | /* | |
1335 | * Move a page back to the lists. | |
1336 | * | |
1337 | * Must be called with the slab lock held. | |
1338 | * | |
1339 | * On exit the slab lock will have been dropped. | |
1340 | */ | |
4b6f0750 | 1341 | static void unfreeze_slab(struct kmem_cache *s, struct page *page) |
81819f0f | 1342 | { |
e95eed57 CL |
1343 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1344 | ||
4b6f0750 | 1345 | ClearSlabFrozen(page); |
81819f0f | 1346 | if (page->inuse) { |
e95eed57 | 1347 | |
81819f0f | 1348 | if (page->freelist) |
e95eed57 | 1349 | add_partial(n, page); |
35e5d7ee | 1350 | else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER)) |
e95eed57 | 1351 | add_full(n, page); |
81819f0f | 1352 | slab_unlock(page); |
e95eed57 | 1353 | |
81819f0f | 1354 | } else { |
e95eed57 CL |
1355 | if (n->nr_partial < MIN_PARTIAL) { |
1356 | /* | |
672bba3a CL |
1357 | * Adding an empty slab to the partial slabs in order |
1358 | * to avoid page allocator overhead. This slab needs | |
1359 | * to come after the other slabs with objects in | |
1360 | * order to fill them up. That way the size of the | |
1361 | * partial list stays small. kmem_cache_shrink can | |
1362 | * reclaim empty slabs from the partial list. | |
e95eed57 CL |
1363 | */ |
1364 | add_partial_tail(n, page); | |
1365 | slab_unlock(page); | |
1366 | } else { | |
1367 | slab_unlock(page); | |
1368 | discard_slab(s, page); | |
1369 | } | |
81819f0f CL |
1370 | } |
1371 | } | |
1372 | ||
1373 | /* | |
1374 | * Remove the cpu slab | |
1375 | */ | |
1376 | static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu) | |
1377 | { | |
894b8788 CL |
1378 | /* |
1379 | * Merge cpu freelist into freelist. Typically we get here | |
1380 | * because both freelists are empty. So this is unlikely | |
1381 | * to occur. | |
1382 | */ | |
1383 | while (unlikely(page->lockless_freelist)) { | |
1384 | void **object; | |
1385 | ||
1386 | /* Retrieve object from cpu_freelist */ | |
1387 | object = page->lockless_freelist; | |
1388 | page->lockless_freelist = page->lockless_freelist[page->offset]; | |
1389 | ||
1390 | /* And put onto the regular freelist */ | |
1391 | object[page->offset] = page->freelist; | |
1392 | page->freelist = object; | |
1393 | page->inuse--; | |
1394 | } | |
81819f0f | 1395 | s->cpu_slab[cpu] = NULL; |
4b6f0750 | 1396 | unfreeze_slab(s, page); |
81819f0f CL |
1397 | } |
1398 | ||
0c710013 | 1399 | static inline void flush_slab(struct kmem_cache *s, struct page *page, int cpu) |
81819f0f CL |
1400 | { |
1401 | slab_lock(page); | |
1402 | deactivate_slab(s, page, cpu); | |
1403 | } | |
1404 | ||
1405 | /* | |
1406 | * Flush cpu slab. | |
1407 | * Called from IPI handler with interrupts disabled. | |
1408 | */ | |
0c710013 | 1409 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) |
81819f0f CL |
1410 | { |
1411 | struct page *page = s->cpu_slab[cpu]; | |
1412 | ||
1413 | if (likely(page)) | |
1414 | flush_slab(s, page, cpu); | |
1415 | } | |
1416 | ||
1417 | static void flush_cpu_slab(void *d) | |
1418 | { | |
1419 | struct kmem_cache *s = d; | |
1420 | int cpu = smp_processor_id(); | |
1421 | ||
1422 | __flush_cpu_slab(s, cpu); | |
1423 | } | |
1424 | ||
1425 | static void flush_all(struct kmem_cache *s) | |
1426 | { | |
1427 | #ifdef CONFIG_SMP | |
1428 | on_each_cpu(flush_cpu_slab, s, 1, 1); | |
1429 | #else | |
1430 | unsigned long flags; | |
1431 | ||
1432 | local_irq_save(flags); | |
1433 | flush_cpu_slab(s); | |
1434 | local_irq_restore(flags); | |
1435 | #endif | |
1436 | } | |
1437 | ||
1438 | /* | |
894b8788 CL |
1439 | * Slow path. The lockless freelist is empty or we need to perform |
1440 | * debugging duties. | |
1441 | * | |
1442 | * Interrupts are disabled. | |
81819f0f | 1443 | * |
894b8788 CL |
1444 | * Processing is still very fast if new objects have been freed to the |
1445 | * regular freelist. In that case we simply take over the regular freelist | |
1446 | * as the lockless freelist and zap the regular freelist. | |
81819f0f | 1447 | * |
894b8788 CL |
1448 | * If that is not working then we fall back to the partial lists. We take the |
1449 | * first element of the freelist as the object to allocate now and move the | |
1450 | * rest of the freelist to the lockless freelist. | |
81819f0f | 1451 | * |
894b8788 CL |
1452 | * And if we were unable to get a new slab from the partial slab lists then |
1453 | * we need to allocate a new slab. This is slowest path since we may sleep. | |
81819f0f | 1454 | */ |
894b8788 CL |
1455 | static void *__slab_alloc(struct kmem_cache *s, |
1456 | gfp_t gfpflags, int node, void *addr, struct page *page) | |
81819f0f | 1457 | { |
81819f0f | 1458 | void **object; |
894b8788 | 1459 | int cpu = smp_processor_id(); |
81819f0f | 1460 | |
81819f0f CL |
1461 | if (!page) |
1462 | goto new_slab; | |
1463 | ||
1464 | slab_lock(page); | |
1465 | if (unlikely(node != -1 && page_to_nid(page) != node)) | |
1466 | goto another_slab; | |
894b8788 | 1467 | load_freelist: |
81819f0f CL |
1468 | object = page->freelist; |
1469 | if (unlikely(!object)) | |
1470 | goto another_slab; | |
35e5d7ee | 1471 | if (unlikely(SlabDebug(page))) |
81819f0f CL |
1472 | goto debug; |
1473 | ||
894b8788 CL |
1474 | object = page->freelist; |
1475 | page->lockless_freelist = object[page->offset]; | |
1476 | page->inuse = s->objects; | |
1477 | page->freelist = NULL; | |
81819f0f | 1478 | slab_unlock(page); |
81819f0f CL |
1479 | return object; |
1480 | ||
1481 | another_slab: | |
1482 | deactivate_slab(s, page, cpu); | |
1483 | ||
1484 | new_slab: | |
1485 | page = get_partial(s, gfpflags, node); | |
894b8788 | 1486 | if (page) { |
81819f0f | 1487 | s->cpu_slab[cpu] = page; |
894b8788 | 1488 | goto load_freelist; |
81819f0f CL |
1489 | } |
1490 | ||
1491 | page = new_slab(s, gfpflags, node); | |
1492 | if (page) { | |
1493 | cpu = smp_processor_id(); | |
1494 | if (s->cpu_slab[cpu]) { | |
1495 | /* | |
672bba3a CL |
1496 | * Someone else populated the cpu_slab while we |
1497 | * enabled interrupts, or we have gotten scheduled | |
1498 | * on another cpu. The page may not be on the | |
1499 | * requested node even if __GFP_THISNODE was | |
1500 | * specified. So we need to recheck. | |
81819f0f CL |
1501 | */ |
1502 | if (node == -1 || | |
1503 | page_to_nid(s->cpu_slab[cpu]) == node) { | |
1504 | /* | |
1505 | * Current cpuslab is acceptable and we | |
1506 | * want the current one since its cache hot | |
1507 | */ | |
1508 | discard_slab(s, page); | |
1509 | page = s->cpu_slab[cpu]; | |
1510 | slab_lock(page); | |
894b8788 | 1511 | goto load_freelist; |
81819f0f | 1512 | } |
672bba3a | 1513 | /* New slab does not fit our expectations */ |
81819f0f CL |
1514 | flush_slab(s, s->cpu_slab[cpu], cpu); |
1515 | } | |
1516 | slab_lock(page); | |
4b6f0750 CL |
1517 | SetSlabFrozen(page); |
1518 | s->cpu_slab[cpu] = page; | |
1519 | goto load_freelist; | |
81819f0f | 1520 | } |
81819f0f CL |
1521 | return NULL; |
1522 | debug: | |
894b8788 | 1523 | object = page->freelist; |
3ec09742 | 1524 | if (!alloc_debug_processing(s, page, object, addr)) |
81819f0f | 1525 | goto another_slab; |
894b8788 CL |
1526 | |
1527 | page->inuse++; | |
1528 | page->freelist = object[page->offset]; | |
1529 | slab_unlock(page); | |
1530 | return object; | |
1531 | } | |
1532 | ||
1533 | /* | |
1534 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) | |
1535 | * have the fastpath folded into their functions. So no function call | |
1536 | * overhead for requests that can be satisfied on the fastpath. | |
1537 | * | |
1538 | * The fastpath works by first checking if the lockless freelist can be used. | |
1539 | * If not then __slab_alloc is called for slow processing. | |
1540 | * | |
1541 | * Otherwise we can simply pick the next object from the lockless free list. | |
1542 | */ | |
1543 | static void __always_inline *slab_alloc(struct kmem_cache *s, | |
ce15fea8 | 1544 | gfp_t gfpflags, int node, void *addr) |
894b8788 CL |
1545 | { |
1546 | struct page *page; | |
1547 | void **object; | |
1548 | unsigned long flags; | |
1549 | ||
1550 | local_irq_save(flags); | |
1551 | page = s->cpu_slab[smp_processor_id()]; | |
1552 | if (unlikely(!page || !page->lockless_freelist || | |
1553 | (node != -1 && page_to_nid(page) != node))) | |
1554 | ||
1555 | object = __slab_alloc(s, gfpflags, node, addr, page); | |
1556 | ||
1557 | else { | |
1558 | object = page->lockless_freelist; | |
1559 | page->lockless_freelist = object[page->offset]; | |
1560 | } | |
1561 | local_irq_restore(flags); | |
d07dbea4 CL |
1562 | |
1563 | if (unlikely((gfpflags & __GFP_ZERO) && object)) | |
ce15fea8 | 1564 | memset(object, 0, s->objsize); |
d07dbea4 | 1565 | |
894b8788 | 1566 | return object; |
81819f0f CL |
1567 | } |
1568 | ||
1569 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
1570 | { | |
ce15fea8 | 1571 | return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); |
81819f0f CL |
1572 | } |
1573 | EXPORT_SYMBOL(kmem_cache_alloc); | |
1574 | ||
1575 | #ifdef CONFIG_NUMA | |
1576 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
1577 | { | |
ce15fea8 | 1578 | return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); |
81819f0f CL |
1579 | } |
1580 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
1581 | #endif | |
1582 | ||
1583 | /* | |
894b8788 CL |
1584 | * Slow patch handling. This may still be called frequently since objects |
1585 | * have a longer lifetime than the cpu slabs in most processing loads. | |
81819f0f | 1586 | * |
894b8788 CL |
1587 | * So we still attempt to reduce cache line usage. Just take the slab |
1588 | * lock and free the item. If there is no additional partial page | |
1589 | * handling required then we can return immediately. | |
81819f0f | 1590 | */ |
894b8788 | 1591 | static void __slab_free(struct kmem_cache *s, struct page *page, |
77c5e2d0 | 1592 | void *x, void *addr) |
81819f0f CL |
1593 | { |
1594 | void *prior; | |
1595 | void **object = (void *)x; | |
81819f0f | 1596 | |
81819f0f CL |
1597 | slab_lock(page); |
1598 | ||
35e5d7ee | 1599 | if (unlikely(SlabDebug(page))) |
81819f0f CL |
1600 | goto debug; |
1601 | checks_ok: | |
1602 | prior = object[page->offset] = page->freelist; | |
1603 | page->freelist = object; | |
1604 | page->inuse--; | |
1605 | ||
4b6f0750 | 1606 | if (unlikely(SlabFrozen(page))) |
81819f0f CL |
1607 | goto out_unlock; |
1608 | ||
1609 | if (unlikely(!page->inuse)) | |
1610 | goto slab_empty; | |
1611 | ||
1612 | /* | |
1613 | * Objects left in the slab. If it | |
1614 | * was not on the partial list before | |
1615 | * then add it. | |
1616 | */ | |
1617 | if (unlikely(!prior)) | |
e95eed57 | 1618 | add_partial(get_node(s, page_to_nid(page)), page); |
81819f0f CL |
1619 | |
1620 | out_unlock: | |
1621 | slab_unlock(page); | |
81819f0f CL |
1622 | return; |
1623 | ||
1624 | slab_empty: | |
1625 | if (prior) | |
1626 | /* | |
672bba3a | 1627 | * Slab still on the partial list. |
81819f0f CL |
1628 | */ |
1629 | remove_partial(s, page); | |
1630 | ||
1631 | slab_unlock(page); | |
1632 | discard_slab(s, page); | |
81819f0f CL |
1633 | return; |
1634 | ||
1635 | debug: | |
3ec09742 | 1636 | if (!free_debug_processing(s, page, x, addr)) |
77c5e2d0 | 1637 | goto out_unlock; |
77c5e2d0 | 1638 | goto checks_ok; |
81819f0f CL |
1639 | } |
1640 | ||
894b8788 CL |
1641 | /* |
1642 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that | |
1643 | * can perform fastpath freeing without additional function calls. | |
1644 | * | |
1645 | * The fastpath is only possible if we are freeing to the current cpu slab | |
1646 | * of this processor. This typically the case if we have just allocated | |
1647 | * the item before. | |
1648 | * | |
1649 | * If fastpath is not possible then fall back to __slab_free where we deal | |
1650 | * with all sorts of special processing. | |
1651 | */ | |
1652 | static void __always_inline slab_free(struct kmem_cache *s, | |
1653 | struct page *page, void *x, void *addr) | |
1654 | { | |
1655 | void **object = (void *)x; | |
1656 | unsigned long flags; | |
1657 | ||
1658 | local_irq_save(flags); | |
1659 | if (likely(page == s->cpu_slab[smp_processor_id()] && | |
1660 | !SlabDebug(page))) { | |
1661 | object[page->offset] = page->lockless_freelist; | |
1662 | page->lockless_freelist = object; | |
1663 | } else | |
1664 | __slab_free(s, page, x, addr); | |
1665 | ||
1666 | local_irq_restore(flags); | |
1667 | } | |
1668 | ||
81819f0f CL |
1669 | void kmem_cache_free(struct kmem_cache *s, void *x) |
1670 | { | |
77c5e2d0 | 1671 | struct page *page; |
81819f0f | 1672 | |
b49af68f | 1673 | page = virt_to_head_page(x); |
81819f0f | 1674 | |
77c5e2d0 | 1675 | slab_free(s, page, x, __builtin_return_address(0)); |
81819f0f CL |
1676 | } |
1677 | EXPORT_SYMBOL(kmem_cache_free); | |
1678 | ||
1679 | /* Figure out on which slab object the object resides */ | |
1680 | static struct page *get_object_page(const void *x) | |
1681 | { | |
b49af68f | 1682 | struct page *page = virt_to_head_page(x); |
81819f0f CL |
1683 | |
1684 | if (!PageSlab(page)) | |
1685 | return NULL; | |
1686 | ||
1687 | return page; | |
1688 | } | |
1689 | ||
1690 | /* | |
672bba3a CL |
1691 | * Object placement in a slab is made very easy because we always start at |
1692 | * offset 0. If we tune the size of the object to the alignment then we can | |
1693 | * get the required alignment by putting one properly sized object after | |
1694 | * another. | |
81819f0f CL |
1695 | * |
1696 | * Notice that the allocation order determines the sizes of the per cpu | |
1697 | * caches. Each processor has always one slab available for allocations. | |
1698 | * Increasing the allocation order reduces the number of times that slabs | |
672bba3a | 1699 | * must be moved on and off the partial lists and is therefore a factor in |
81819f0f | 1700 | * locking overhead. |
81819f0f CL |
1701 | */ |
1702 | ||
1703 | /* | |
1704 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
1705 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
1706 | * and increases the number of allocations possible without having to | |
1707 | * take the list_lock. | |
1708 | */ | |
1709 | static int slub_min_order; | |
1710 | static int slub_max_order = DEFAULT_MAX_ORDER; | |
81819f0f CL |
1711 | static int slub_min_objects = DEFAULT_MIN_OBJECTS; |
1712 | ||
1713 | /* | |
1714 | * Merge control. If this is set then no merging of slab caches will occur. | |
672bba3a | 1715 | * (Could be removed. This was introduced to pacify the merge skeptics.) |
81819f0f CL |
1716 | */ |
1717 | static int slub_nomerge; | |
1718 | ||
81819f0f CL |
1719 | /* |
1720 | * Calculate the order of allocation given an slab object size. | |
1721 | * | |
672bba3a CL |
1722 | * The order of allocation has significant impact on performance and other |
1723 | * system components. Generally order 0 allocations should be preferred since | |
1724 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
1725 | * be problematic to put into order 0 slabs because there may be too much | |
1726 | * unused space left. We go to a higher order if more than 1/8th of the slab | |
1727 | * would be wasted. | |
1728 | * | |
1729 | * In order to reach satisfactory performance we must ensure that a minimum | |
1730 | * number of objects is in one slab. Otherwise we may generate too much | |
1731 | * activity on the partial lists which requires taking the list_lock. This is | |
1732 | * less a concern for large slabs though which are rarely used. | |
81819f0f | 1733 | * |
672bba3a CL |
1734 | * slub_max_order specifies the order where we begin to stop considering the |
1735 | * number of objects in a slab as critical. If we reach slub_max_order then | |
1736 | * we try to keep the page order as low as possible. So we accept more waste | |
1737 | * of space in favor of a small page order. | |
81819f0f | 1738 | * |
672bba3a CL |
1739 | * Higher order allocations also allow the placement of more objects in a |
1740 | * slab and thereby reduce object handling overhead. If the user has | |
1741 | * requested a higher mininum order then we start with that one instead of | |
1742 | * the smallest order which will fit the object. | |
81819f0f | 1743 | */ |
5e6d444e CL |
1744 | static inline int slab_order(int size, int min_objects, |
1745 | int max_order, int fract_leftover) | |
81819f0f CL |
1746 | { |
1747 | int order; | |
1748 | int rem; | |
6300ea75 | 1749 | int min_order = slub_min_order; |
81819f0f | 1750 | |
6300ea75 CL |
1751 | /* |
1752 | * If we would create too many object per slab then reduce | |
1753 | * the slab order even if it goes below slub_min_order. | |
1754 | */ | |
1755 | while (min_order > 0 && | |
1756 | (PAGE_SIZE << min_order) >= MAX_OBJECTS_PER_SLAB * size) | |
1757 | min_order--; | |
1758 | ||
1759 | for (order = max(min_order, | |
5e6d444e CL |
1760 | fls(min_objects * size - 1) - PAGE_SHIFT); |
1761 | order <= max_order; order++) { | |
81819f0f | 1762 | |
5e6d444e | 1763 | unsigned long slab_size = PAGE_SIZE << order; |
81819f0f | 1764 | |
5e6d444e | 1765 | if (slab_size < min_objects * size) |
81819f0f CL |
1766 | continue; |
1767 | ||
1768 | rem = slab_size % size; | |
1769 | ||
5e6d444e | 1770 | if (rem <= slab_size / fract_leftover) |
81819f0f CL |
1771 | break; |
1772 | ||
6300ea75 CL |
1773 | /* If the next size is too high then exit now */ |
1774 | if (slab_size * 2 >= MAX_OBJECTS_PER_SLAB * size) | |
1775 | break; | |
81819f0f | 1776 | } |
672bba3a | 1777 | |
81819f0f CL |
1778 | return order; |
1779 | } | |
1780 | ||
5e6d444e CL |
1781 | static inline int calculate_order(int size) |
1782 | { | |
1783 | int order; | |
1784 | int min_objects; | |
1785 | int fraction; | |
1786 | ||
1787 | /* | |
1788 | * Attempt to find best configuration for a slab. This | |
1789 | * works by first attempting to generate a layout with | |
1790 | * the best configuration and backing off gradually. | |
1791 | * | |
1792 | * First we reduce the acceptable waste in a slab. Then | |
1793 | * we reduce the minimum objects required in a slab. | |
1794 | */ | |
1795 | min_objects = slub_min_objects; | |
1796 | while (min_objects > 1) { | |
1797 | fraction = 8; | |
1798 | while (fraction >= 4) { | |
1799 | order = slab_order(size, min_objects, | |
1800 | slub_max_order, fraction); | |
1801 | if (order <= slub_max_order) | |
1802 | return order; | |
1803 | fraction /= 2; | |
1804 | } | |
1805 | min_objects /= 2; | |
1806 | } | |
1807 | ||
1808 | /* | |
1809 | * We were unable to place multiple objects in a slab. Now | |
1810 | * lets see if we can place a single object there. | |
1811 | */ | |
1812 | order = slab_order(size, 1, slub_max_order, 1); | |
1813 | if (order <= slub_max_order) | |
1814 | return order; | |
1815 | ||
1816 | /* | |
1817 | * Doh this slab cannot be placed using slub_max_order. | |
1818 | */ | |
1819 | order = slab_order(size, 1, MAX_ORDER, 1); | |
1820 | if (order <= MAX_ORDER) | |
1821 | return order; | |
1822 | return -ENOSYS; | |
1823 | } | |
1824 | ||
81819f0f | 1825 | /* |
672bba3a | 1826 | * Figure out what the alignment of the objects will be. |
81819f0f CL |
1827 | */ |
1828 | static unsigned long calculate_alignment(unsigned long flags, | |
1829 | unsigned long align, unsigned long size) | |
1830 | { | |
1831 | /* | |
1832 | * If the user wants hardware cache aligned objects then | |
1833 | * follow that suggestion if the object is sufficiently | |
1834 | * large. | |
1835 | * | |
1836 | * The hardware cache alignment cannot override the | |
1837 | * specified alignment though. If that is greater | |
1838 | * then use it. | |
1839 | */ | |
5af60839 | 1840 | if ((flags & SLAB_HWCACHE_ALIGN) && |
65c02d4c CL |
1841 | size > cache_line_size() / 2) |
1842 | return max_t(unsigned long, align, cache_line_size()); | |
81819f0f CL |
1843 | |
1844 | if (align < ARCH_SLAB_MINALIGN) | |
1845 | return ARCH_SLAB_MINALIGN; | |
1846 | ||
1847 | return ALIGN(align, sizeof(void *)); | |
1848 | } | |
1849 | ||
1850 | static void init_kmem_cache_node(struct kmem_cache_node *n) | |
1851 | { | |
1852 | n->nr_partial = 0; | |
1853 | atomic_long_set(&n->nr_slabs, 0); | |
1854 | spin_lock_init(&n->list_lock); | |
1855 | INIT_LIST_HEAD(&n->partial); | |
643b1138 | 1856 | INIT_LIST_HEAD(&n->full); |
81819f0f CL |
1857 | } |
1858 | ||
1859 | #ifdef CONFIG_NUMA | |
1860 | /* | |
1861 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
1862 | * slab on the node for this slabcache. There are no concurrent accesses | |
1863 | * possible. | |
1864 | * | |
1865 | * Note that this function only works on the kmalloc_node_cache | |
1866 | * when allocating for the kmalloc_node_cache. | |
1867 | */ | |
1868 | static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags, | |
1869 | int node) | |
1870 | { | |
1871 | struct page *page; | |
1872 | struct kmem_cache_node *n; | |
1873 | ||
1874 | BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); | |
1875 | ||
1876 | page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node); | |
81819f0f CL |
1877 | |
1878 | BUG_ON(!page); | |
1879 | n = page->freelist; | |
1880 | BUG_ON(!n); | |
1881 | page->freelist = get_freepointer(kmalloc_caches, n); | |
1882 | page->inuse++; | |
1883 | kmalloc_caches->node[node] = n; | |
d45f39cb CL |
1884 | init_object(kmalloc_caches, n, 1); |
1885 | init_tracking(kmalloc_caches, n); | |
81819f0f CL |
1886 | init_kmem_cache_node(n); |
1887 | atomic_long_inc(&n->nr_slabs); | |
e95eed57 | 1888 | add_partial(n, page); |
dbc55faa CL |
1889 | |
1890 | /* | |
1891 | * new_slab() disables interupts. If we do not reenable interrupts here | |
1892 | * then bootup would continue with interrupts disabled. | |
1893 | */ | |
1894 | local_irq_enable(); | |
81819f0f CL |
1895 | return n; |
1896 | } | |
1897 | ||
1898 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
1899 | { | |
1900 | int node; | |
1901 | ||
1902 | for_each_online_node(node) { | |
1903 | struct kmem_cache_node *n = s->node[node]; | |
1904 | if (n && n != &s->local_node) | |
1905 | kmem_cache_free(kmalloc_caches, n); | |
1906 | s->node[node] = NULL; | |
1907 | } | |
1908 | } | |
1909 | ||
1910 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
1911 | { | |
1912 | int node; | |
1913 | int local_node; | |
1914 | ||
1915 | if (slab_state >= UP) | |
1916 | local_node = page_to_nid(virt_to_page(s)); | |
1917 | else | |
1918 | local_node = 0; | |
1919 | ||
1920 | for_each_online_node(node) { | |
1921 | struct kmem_cache_node *n; | |
1922 | ||
1923 | if (local_node == node) | |
1924 | n = &s->local_node; | |
1925 | else { | |
1926 | if (slab_state == DOWN) { | |
1927 | n = early_kmem_cache_node_alloc(gfpflags, | |
1928 | node); | |
1929 | continue; | |
1930 | } | |
1931 | n = kmem_cache_alloc_node(kmalloc_caches, | |
1932 | gfpflags, node); | |
1933 | ||
1934 | if (!n) { | |
1935 | free_kmem_cache_nodes(s); | |
1936 | return 0; | |
1937 | } | |
1938 | ||
1939 | } | |
1940 | s->node[node] = n; | |
1941 | init_kmem_cache_node(n); | |
1942 | } | |
1943 | return 1; | |
1944 | } | |
1945 | #else | |
1946 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
1947 | { | |
1948 | } | |
1949 | ||
1950 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
1951 | { | |
1952 | init_kmem_cache_node(&s->local_node); | |
1953 | return 1; | |
1954 | } | |
1955 | #endif | |
1956 | ||
1957 | /* | |
1958 | * calculate_sizes() determines the order and the distribution of data within | |
1959 | * a slab object. | |
1960 | */ | |
1961 | static int calculate_sizes(struct kmem_cache *s) | |
1962 | { | |
1963 | unsigned long flags = s->flags; | |
1964 | unsigned long size = s->objsize; | |
1965 | unsigned long align = s->align; | |
1966 | ||
1967 | /* | |
1968 | * Determine if we can poison the object itself. If the user of | |
1969 | * the slab may touch the object after free or before allocation | |
1970 | * then we should never poison the object itself. | |
1971 | */ | |
1972 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
c59def9f | 1973 | !s->ctor) |
81819f0f CL |
1974 | s->flags |= __OBJECT_POISON; |
1975 | else | |
1976 | s->flags &= ~__OBJECT_POISON; | |
1977 | ||
1978 | /* | |
1979 | * Round up object size to the next word boundary. We can only | |
1980 | * place the free pointer at word boundaries and this determines | |
1981 | * the possible location of the free pointer. | |
1982 | */ | |
1983 | size = ALIGN(size, sizeof(void *)); | |
1984 | ||
41ecc55b | 1985 | #ifdef CONFIG_SLUB_DEBUG |
81819f0f | 1986 | /* |
672bba3a | 1987 | * If we are Redzoning then check if there is some space between the |
81819f0f | 1988 | * end of the object and the free pointer. If not then add an |
672bba3a | 1989 | * additional word to have some bytes to store Redzone information. |
81819f0f CL |
1990 | */ |
1991 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
1992 | size += sizeof(void *); | |
41ecc55b | 1993 | #endif |
81819f0f CL |
1994 | |
1995 | /* | |
672bba3a CL |
1996 | * With that we have determined the number of bytes in actual use |
1997 | * by the object. This is the potential offset to the free pointer. | |
81819f0f CL |
1998 | */ |
1999 | s->inuse = size; | |
2000 | ||
2001 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
c59def9f | 2002 | s->ctor)) { |
81819f0f CL |
2003 | /* |
2004 | * Relocate free pointer after the object if it is not | |
2005 | * permitted to overwrite the first word of the object on | |
2006 | * kmem_cache_free. | |
2007 | * | |
2008 | * This is the case if we do RCU, have a constructor or | |
2009 | * destructor or are poisoning the objects. | |
2010 | */ | |
2011 | s->offset = size; | |
2012 | size += sizeof(void *); | |
2013 | } | |
2014 | ||
c12b3c62 | 2015 | #ifdef CONFIG_SLUB_DEBUG |
81819f0f CL |
2016 | if (flags & SLAB_STORE_USER) |
2017 | /* | |
2018 | * Need to store information about allocs and frees after | |
2019 | * the object. | |
2020 | */ | |
2021 | size += 2 * sizeof(struct track); | |
2022 | ||
be7b3fbc | 2023 | if (flags & SLAB_RED_ZONE) |
81819f0f CL |
2024 | /* |
2025 | * Add some empty padding so that we can catch | |
2026 | * overwrites from earlier objects rather than let | |
2027 | * tracking information or the free pointer be | |
2028 | * corrupted if an user writes before the start | |
2029 | * of the object. | |
2030 | */ | |
2031 | size += sizeof(void *); | |
41ecc55b | 2032 | #endif |
672bba3a | 2033 | |
81819f0f CL |
2034 | /* |
2035 | * Determine the alignment based on various parameters that the | |
65c02d4c CL |
2036 | * user specified and the dynamic determination of cache line size |
2037 | * on bootup. | |
81819f0f CL |
2038 | */ |
2039 | align = calculate_alignment(flags, align, s->objsize); | |
2040 | ||
2041 | /* | |
2042 | * SLUB stores one object immediately after another beginning from | |
2043 | * offset 0. In order to align the objects we have to simply size | |
2044 | * each object to conform to the alignment. | |
2045 | */ | |
2046 | size = ALIGN(size, align); | |
2047 | s->size = size; | |
2048 | ||
2049 | s->order = calculate_order(size); | |
2050 | if (s->order < 0) | |
2051 | return 0; | |
2052 | ||
2053 | /* | |
2054 | * Determine the number of objects per slab | |
2055 | */ | |
2056 | s->objects = (PAGE_SIZE << s->order) / size; | |
2057 | ||
2058 | /* | |
2059 | * Verify that the number of objects is within permitted limits. | |
2060 | * The page->inuse field is only 16 bit wide! So we cannot have | |
2061 | * more than 64k objects per slab. | |
2062 | */ | |
6300ea75 | 2063 | if (!s->objects || s->objects > MAX_OBJECTS_PER_SLAB) |
81819f0f CL |
2064 | return 0; |
2065 | return 1; | |
2066 | ||
2067 | } | |
2068 | ||
81819f0f CL |
2069 | static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, |
2070 | const char *name, size_t size, | |
2071 | size_t align, unsigned long flags, | |
c59def9f | 2072 | void (*ctor)(void *, struct kmem_cache *, unsigned long)) |
81819f0f CL |
2073 | { |
2074 | memset(s, 0, kmem_size); | |
2075 | s->name = name; | |
2076 | s->ctor = ctor; | |
81819f0f CL |
2077 | s->objsize = size; |
2078 | s->flags = flags; | |
2079 | s->align = align; | |
41ecc55b | 2080 | kmem_cache_open_debug_check(s); |
81819f0f CL |
2081 | |
2082 | if (!calculate_sizes(s)) | |
2083 | goto error; | |
2084 | ||
2085 | s->refcount = 1; | |
2086 | #ifdef CONFIG_NUMA | |
2087 | s->defrag_ratio = 100; | |
2088 | #endif | |
2089 | ||
2090 | if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) | |
2091 | return 1; | |
2092 | error: | |
2093 | if (flags & SLAB_PANIC) | |
2094 | panic("Cannot create slab %s size=%lu realsize=%u " | |
2095 | "order=%u offset=%u flags=%lx\n", | |
2096 | s->name, (unsigned long)size, s->size, s->order, | |
2097 | s->offset, flags); | |
2098 | return 0; | |
2099 | } | |
81819f0f CL |
2100 | |
2101 | /* | |
2102 | * Check if a given pointer is valid | |
2103 | */ | |
2104 | int kmem_ptr_validate(struct kmem_cache *s, const void *object) | |
2105 | { | |
2106 | struct page * page; | |
81819f0f CL |
2107 | |
2108 | page = get_object_page(object); | |
2109 | ||
2110 | if (!page || s != page->slab) | |
2111 | /* No slab or wrong slab */ | |
2112 | return 0; | |
2113 | ||
abcd08a6 | 2114 | if (!check_valid_pointer(s, page, object)) |
81819f0f CL |
2115 | return 0; |
2116 | ||
2117 | /* | |
2118 | * We could also check if the object is on the slabs freelist. | |
2119 | * But this would be too expensive and it seems that the main | |
2120 | * purpose of kmem_ptr_valid is to check if the object belongs | |
2121 | * to a certain slab. | |
2122 | */ | |
2123 | return 1; | |
2124 | } | |
2125 | EXPORT_SYMBOL(kmem_ptr_validate); | |
2126 | ||
2127 | /* | |
2128 | * Determine the size of a slab object | |
2129 | */ | |
2130 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
2131 | { | |
2132 | return s->objsize; | |
2133 | } | |
2134 | EXPORT_SYMBOL(kmem_cache_size); | |
2135 | ||
2136 | const char *kmem_cache_name(struct kmem_cache *s) | |
2137 | { | |
2138 | return s->name; | |
2139 | } | |
2140 | EXPORT_SYMBOL(kmem_cache_name); | |
2141 | ||
2142 | /* | |
672bba3a CL |
2143 | * Attempt to free all slabs on a node. Return the number of slabs we |
2144 | * were unable to free. | |
81819f0f CL |
2145 | */ |
2146 | static int free_list(struct kmem_cache *s, struct kmem_cache_node *n, | |
2147 | struct list_head *list) | |
2148 | { | |
2149 | int slabs_inuse = 0; | |
2150 | unsigned long flags; | |
2151 | struct page *page, *h; | |
2152 | ||
2153 | spin_lock_irqsave(&n->list_lock, flags); | |
2154 | list_for_each_entry_safe(page, h, list, lru) | |
2155 | if (!page->inuse) { | |
2156 | list_del(&page->lru); | |
2157 | discard_slab(s, page); | |
2158 | } else | |
2159 | slabs_inuse++; | |
2160 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2161 | return slabs_inuse; | |
2162 | } | |
2163 | ||
2164 | /* | |
672bba3a | 2165 | * Release all resources used by a slab cache. |
81819f0f | 2166 | */ |
0c710013 | 2167 | static inline int kmem_cache_close(struct kmem_cache *s) |
81819f0f CL |
2168 | { |
2169 | int node; | |
2170 | ||
2171 | flush_all(s); | |
2172 | ||
2173 | /* Attempt to free all objects */ | |
2174 | for_each_online_node(node) { | |
2175 | struct kmem_cache_node *n = get_node(s, node); | |
2176 | ||
2086d26a | 2177 | n->nr_partial -= free_list(s, n, &n->partial); |
81819f0f CL |
2178 | if (atomic_long_read(&n->nr_slabs)) |
2179 | return 1; | |
2180 | } | |
2181 | free_kmem_cache_nodes(s); | |
2182 | return 0; | |
2183 | } | |
2184 | ||
2185 | /* | |
2186 | * Close a cache and release the kmem_cache structure | |
2187 | * (must be used for caches created using kmem_cache_create) | |
2188 | */ | |
2189 | void kmem_cache_destroy(struct kmem_cache *s) | |
2190 | { | |
2191 | down_write(&slub_lock); | |
2192 | s->refcount--; | |
2193 | if (!s->refcount) { | |
2194 | list_del(&s->list); | |
a0e1d1be | 2195 | up_write(&slub_lock); |
81819f0f CL |
2196 | if (kmem_cache_close(s)) |
2197 | WARN_ON(1); | |
2198 | sysfs_slab_remove(s); | |
2199 | kfree(s); | |
a0e1d1be CL |
2200 | } else |
2201 | up_write(&slub_lock); | |
81819f0f CL |
2202 | } |
2203 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2204 | ||
2205 | /******************************************************************** | |
2206 | * Kmalloc subsystem | |
2207 | *******************************************************************/ | |
2208 | ||
2209 | struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned; | |
2210 | EXPORT_SYMBOL(kmalloc_caches); | |
2211 | ||
2212 | #ifdef CONFIG_ZONE_DMA | |
2213 | static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1]; | |
2214 | #endif | |
2215 | ||
2216 | static int __init setup_slub_min_order(char *str) | |
2217 | { | |
2218 | get_option (&str, &slub_min_order); | |
2219 | ||
2220 | return 1; | |
2221 | } | |
2222 | ||
2223 | __setup("slub_min_order=", setup_slub_min_order); | |
2224 | ||
2225 | static int __init setup_slub_max_order(char *str) | |
2226 | { | |
2227 | get_option (&str, &slub_max_order); | |
2228 | ||
2229 | return 1; | |
2230 | } | |
2231 | ||
2232 | __setup("slub_max_order=", setup_slub_max_order); | |
2233 | ||
2234 | static int __init setup_slub_min_objects(char *str) | |
2235 | { | |
2236 | get_option (&str, &slub_min_objects); | |
2237 | ||
2238 | return 1; | |
2239 | } | |
2240 | ||
2241 | __setup("slub_min_objects=", setup_slub_min_objects); | |
2242 | ||
2243 | static int __init setup_slub_nomerge(char *str) | |
2244 | { | |
2245 | slub_nomerge = 1; | |
2246 | return 1; | |
2247 | } | |
2248 | ||
2249 | __setup("slub_nomerge", setup_slub_nomerge); | |
2250 | ||
81819f0f CL |
2251 | static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, |
2252 | const char *name, int size, gfp_t gfp_flags) | |
2253 | { | |
2254 | unsigned int flags = 0; | |
2255 | ||
2256 | if (gfp_flags & SLUB_DMA) | |
2257 | flags = SLAB_CACHE_DMA; | |
2258 | ||
2259 | down_write(&slub_lock); | |
2260 | if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, | |
c59def9f | 2261 | flags, NULL)) |
81819f0f CL |
2262 | goto panic; |
2263 | ||
2264 | list_add(&s->list, &slab_caches); | |
2265 | up_write(&slub_lock); | |
2266 | if (sysfs_slab_add(s)) | |
2267 | goto panic; | |
2268 | return s; | |
2269 | ||
2270 | panic: | |
2271 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
2272 | } | |
2273 | ||
2e443fd0 CL |
2274 | #ifdef CONFIG_ZONE_DMA |
2275 | static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) | |
2276 | { | |
2277 | struct kmem_cache *s; | |
2278 | struct kmem_cache *x; | |
2279 | char *text; | |
2280 | size_t realsize; | |
2281 | ||
2282 | s = kmalloc_caches_dma[index]; | |
2283 | if (s) | |
2284 | return s; | |
2285 | ||
2286 | /* Dynamically create dma cache */ | |
2287 | x = kmalloc(kmem_size, flags & ~SLUB_DMA); | |
2288 | if (!x) | |
2289 | panic("Unable to allocate memory for dma cache\n"); | |
2290 | ||
7b55f620 | 2291 | realsize = kmalloc_caches[index].objsize; |
2e443fd0 CL |
2292 | text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", |
2293 | (unsigned int)realsize); | |
2294 | s = create_kmalloc_cache(x, text, realsize, flags); | |
dfce8648 CL |
2295 | down_write(&slub_lock); |
2296 | if (!kmalloc_caches_dma[index]) { | |
2297 | kmalloc_caches_dma[index] = s; | |
2298 | up_write(&slub_lock); | |
2299 | return s; | |
2300 | } | |
2301 | up_write(&slub_lock); | |
2302 | kmem_cache_destroy(s); | |
2303 | return kmalloc_caches_dma[index]; | |
2e443fd0 CL |
2304 | } |
2305 | #endif | |
2306 | ||
f1b26339 CL |
2307 | /* |
2308 | * Conversion table for small slabs sizes / 8 to the index in the | |
2309 | * kmalloc array. This is necessary for slabs < 192 since we have non power | |
2310 | * of two cache sizes there. The size of larger slabs can be determined using | |
2311 | * fls. | |
2312 | */ | |
2313 | static s8 size_index[24] = { | |
2314 | 3, /* 8 */ | |
2315 | 4, /* 16 */ | |
2316 | 5, /* 24 */ | |
2317 | 5, /* 32 */ | |
2318 | 6, /* 40 */ | |
2319 | 6, /* 48 */ | |
2320 | 6, /* 56 */ | |
2321 | 6, /* 64 */ | |
2322 | 1, /* 72 */ | |
2323 | 1, /* 80 */ | |
2324 | 1, /* 88 */ | |
2325 | 1, /* 96 */ | |
2326 | 7, /* 104 */ | |
2327 | 7, /* 112 */ | |
2328 | 7, /* 120 */ | |
2329 | 7, /* 128 */ | |
2330 | 2, /* 136 */ | |
2331 | 2, /* 144 */ | |
2332 | 2, /* 152 */ | |
2333 | 2, /* 160 */ | |
2334 | 2, /* 168 */ | |
2335 | 2, /* 176 */ | |
2336 | 2, /* 184 */ | |
2337 | 2 /* 192 */ | |
2338 | }; | |
2339 | ||
81819f0f CL |
2340 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) |
2341 | { | |
f1b26339 | 2342 | int index; |
81819f0f | 2343 | |
f1b26339 CL |
2344 | if (size <= 192) { |
2345 | if (!size) | |
2346 | return ZERO_SIZE_PTR; | |
81819f0f | 2347 | |
f1b26339 CL |
2348 | index = size_index[(size - 1) / 8]; |
2349 | } else { | |
2350 | if (size > KMALLOC_MAX_SIZE) | |
2351 | return NULL; | |
2352 | ||
2353 | index = fls(size - 1); | |
2354 | } | |
81819f0f CL |
2355 | |
2356 | #ifdef CONFIG_ZONE_DMA | |
f1b26339 | 2357 | if (unlikely((flags & SLUB_DMA))) |
2e443fd0 | 2358 | return dma_kmalloc_cache(index, flags); |
f1b26339 | 2359 | |
81819f0f CL |
2360 | #endif |
2361 | return &kmalloc_caches[index]; | |
2362 | } | |
2363 | ||
2364 | void *__kmalloc(size_t size, gfp_t flags) | |
2365 | { | |
2366 | struct kmem_cache *s = get_slab(size, flags); | |
2367 | ||
6cb8f913 CL |
2368 | if (ZERO_OR_NULL_PTR(s)) |
2369 | return s; | |
2370 | ||
ce15fea8 | 2371 | return slab_alloc(s, flags, -1, __builtin_return_address(0)); |
81819f0f CL |
2372 | } |
2373 | EXPORT_SYMBOL(__kmalloc); | |
2374 | ||
2375 | #ifdef CONFIG_NUMA | |
2376 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
2377 | { | |
2378 | struct kmem_cache *s = get_slab(size, flags); | |
2379 | ||
6cb8f913 CL |
2380 | if (ZERO_OR_NULL_PTR(s)) |
2381 | return s; | |
2382 | ||
ce15fea8 | 2383 | return slab_alloc(s, flags, node, __builtin_return_address(0)); |
81819f0f CL |
2384 | } |
2385 | EXPORT_SYMBOL(__kmalloc_node); | |
2386 | #endif | |
2387 | ||
2388 | size_t ksize(const void *object) | |
2389 | { | |
272c1d21 | 2390 | struct page *page; |
81819f0f CL |
2391 | struct kmem_cache *s; |
2392 | ||
272c1d21 CL |
2393 | if (object == ZERO_SIZE_PTR) |
2394 | return 0; | |
2395 | ||
2396 | page = get_object_page(object); | |
81819f0f CL |
2397 | BUG_ON(!page); |
2398 | s = page->slab; | |
2399 | BUG_ON(!s); | |
2400 | ||
2401 | /* | |
2402 | * Debugging requires use of the padding between object | |
2403 | * and whatever may come after it. | |
2404 | */ | |
2405 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
2406 | return s->objsize; | |
2407 | ||
2408 | /* | |
2409 | * If we have the need to store the freelist pointer | |
2410 | * back there or track user information then we can | |
2411 | * only use the space before that information. | |
2412 | */ | |
2413 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
2414 | return s->inuse; | |
2415 | ||
2416 | /* | |
2417 | * Else we can use all the padding etc for the allocation | |
2418 | */ | |
2419 | return s->size; | |
2420 | } | |
2421 | EXPORT_SYMBOL(ksize); | |
2422 | ||
2423 | void kfree(const void *x) | |
2424 | { | |
2425 | struct kmem_cache *s; | |
2426 | struct page *page; | |
2427 | ||
272c1d21 CL |
2428 | /* |
2429 | * This has to be an unsigned comparison. According to Linus | |
2430 | * some gcc version treat a pointer as a signed entity. Then | |
2431 | * this comparison would be true for all "negative" pointers | |
2432 | * (which would cover the whole upper half of the address space). | |
2433 | */ | |
6cb8f913 | 2434 | if (ZERO_OR_NULL_PTR(x)) |
81819f0f CL |
2435 | return; |
2436 | ||
b49af68f | 2437 | page = virt_to_head_page(x); |
81819f0f CL |
2438 | s = page->slab; |
2439 | ||
77c5e2d0 | 2440 | slab_free(s, page, (void *)x, __builtin_return_address(0)); |
81819f0f CL |
2441 | } |
2442 | EXPORT_SYMBOL(kfree); | |
2443 | ||
2086d26a | 2444 | /* |
672bba3a CL |
2445 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts |
2446 | * the remaining slabs by the number of items in use. The slabs with the | |
2447 | * most items in use come first. New allocations will then fill those up | |
2448 | * and thus they can be removed from the partial lists. | |
2449 | * | |
2450 | * The slabs with the least items are placed last. This results in them | |
2451 | * being allocated from last increasing the chance that the last objects | |
2452 | * are freed in them. | |
2086d26a CL |
2453 | */ |
2454 | int kmem_cache_shrink(struct kmem_cache *s) | |
2455 | { | |
2456 | int node; | |
2457 | int i; | |
2458 | struct kmem_cache_node *n; | |
2459 | struct page *page; | |
2460 | struct page *t; | |
2461 | struct list_head *slabs_by_inuse = | |
2462 | kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL); | |
2463 | unsigned long flags; | |
2464 | ||
2465 | if (!slabs_by_inuse) | |
2466 | return -ENOMEM; | |
2467 | ||
2468 | flush_all(s); | |
2469 | for_each_online_node(node) { | |
2470 | n = get_node(s, node); | |
2471 | ||
2472 | if (!n->nr_partial) | |
2473 | continue; | |
2474 | ||
2475 | for (i = 0; i < s->objects; i++) | |
2476 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
2477 | ||
2478 | spin_lock_irqsave(&n->list_lock, flags); | |
2479 | ||
2480 | /* | |
672bba3a | 2481 | * Build lists indexed by the items in use in each slab. |
2086d26a | 2482 | * |
672bba3a CL |
2483 | * Note that concurrent frees may occur while we hold the |
2484 | * list_lock. page->inuse here is the upper limit. | |
2086d26a CL |
2485 | */ |
2486 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
2487 | if (!page->inuse && slab_trylock(page)) { | |
2488 | /* | |
2489 | * Must hold slab lock here because slab_free | |
2490 | * may have freed the last object and be | |
2491 | * waiting to release the slab. | |
2492 | */ | |
2493 | list_del(&page->lru); | |
2494 | n->nr_partial--; | |
2495 | slab_unlock(page); | |
2496 | discard_slab(s, page); | |
2497 | } else { | |
2498 | if (n->nr_partial > MAX_PARTIAL) | |
2499 | list_move(&page->lru, | |
2500 | slabs_by_inuse + page->inuse); | |
2501 | } | |
2502 | } | |
2503 | ||
2504 | if (n->nr_partial <= MAX_PARTIAL) | |
2505 | goto out; | |
2506 | ||
2507 | /* | |
672bba3a CL |
2508 | * Rebuild the partial list with the slabs filled up most |
2509 | * first and the least used slabs at the end. | |
2086d26a CL |
2510 | */ |
2511 | for (i = s->objects - 1; i >= 0; i--) | |
2512 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
2513 | ||
2514 | out: | |
2515 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2516 | } | |
2517 | ||
2518 | kfree(slabs_by_inuse); | |
2519 | return 0; | |
2520 | } | |
2521 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2522 | ||
81819f0f CL |
2523 | /******************************************************************** |
2524 | * Basic setup of slabs | |
2525 | *******************************************************************/ | |
2526 | ||
2527 | void __init kmem_cache_init(void) | |
2528 | { | |
2529 | int i; | |
4b356be0 | 2530 | int caches = 0; |
81819f0f CL |
2531 | |
2532 | #ifdef CONFIG_NUMA | |
2533 | /* | |
2534 | * Must first have the slab cache available for the allocations of the | |
672bba3a | 2535 | * struct kmem_cache_node's. There is special bootstrap code in |
81819f0f CL |
2536 | * kmem_cache_open for slab_state == DOWN. |
2537 | */ | |
2538 | create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", | |
2539 | sizeof(struct kmem_cache_node), GFP_KERNEL); | |
8ffa6875 | 2540 | kmalloc_caches[0].refcount = -1; |
4b356be0 | 2541 | caches++; |
81819f0f CL |
2542 | #endif |
2543 | ||
2544 | /* Able to allocate the per node structures */ | |
2545 | slab_state = PARTIAL; | |
2546 | ||
2547 | /* Caches that are not of the two-to-the-power-of size */ | |
4b356be0 CL |
2548 | if (KMALLOC_MIN_SIZE <= 64) { |
2549 | create_kmalloc_cache(&kmalloc_caches[1], | |
81819f0f | 2550 | "kmalloc-96", 96, GFP_KERNEL); |
4b356be0 CL |
2551 | caches++; |
2552 | } | |
2553 | if (KMALLOC_MIN_SIZE <= 128) { | |
2554 | create_kmalloc_cache(&kmalloc_caches[2], | |
81819f0f | 2555 | "kmalloc-192", 192, GFP_KERNEL); |
4b356be0 CL |
2556 | caches++; |
2557 | } | |
81819f0f | 2558 | |
4b356be0 | 2559 | for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) { |
81819f0f CL |
2560 | create_kmalloc_cache(&kmalloc_caches[i], |
2561 | "kmalloc", 1 << i, GFP_KERNEL); | |
4b356be0 CL |
2562 | caches++; |
2563 | } | |
81819f0f | 2564 | |
f1b26339 CL |
2565 | |
2566 | /* | |
2567 | * Patch up the size_index table if we have strange large alignment | |
2568 | * requirements for the kmalloc array. This is only the case for | |
2569 | * mips it seems. The standard arches will not generate any code here. | |
2570 | * | |
2571 | * Largest permitted alignment is 256 bytes due to the way we | |
2572 | * handle the index determination for the smaller caches. | |
2573 | * | |
2574 | * Make sure that nothing crazy happens if someone starts tinkering | |
2575 | * around with ARCH_KMALLOC_MINALIGN | |
2576 | */ | |
2577 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || | |
2578 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); | |
2579 | ||
12ad6843 | 2580 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) |
f1b26339 CL |
2581 | size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; |
2582 | ||
81819f0f CL |
2583 | slab_state = UP; |
2584 | ||
2585 | /* Provide the correct kmalloc names now that the caches are up */ | |
2586 | for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) | |
2587 | kmalloc_caches[i]. name = | |
2588 | kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); | |
2589 | ||
2590 | #ifdef CONFIG_SMP | |
2591 | register_cpu_notifier(&slab_notifier); | |
2592 | #endif | |
2593 | ||
bcf889f9 CL |
2594 | kmem_size = offsetof(struct kmem_cache, cpu_slab) + |
2595 | nr_cpu_ids * sizeof(struct page *); | |
81819f0f CL |
2596 | |
2597 | printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
4b356be0 CL |
2598 | " CPUs=%d, Nodes=%d\n", |
2599 | caches, cache_line_size(), | |
81819f0f CL |
2600 | slub_min_order, slub_max_order, slub_min_objects, |
2601 | nr_cpu_ids, nr_node_ids); | |
2602 | } | |
2603 | ||
2604 | /* | |
2605 | * Find a mergeable slab cache | |
2606 | */ | |
2607 | static int slab_unmergeable(struct kmem_cache *s) | |
2608 | { | |
2609 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
2610 | return 1; | |
2611 | ||
c59def9f | 2612 | if (s->ctor) |
81819f0f CL |
2613 | return 1; |
2614 | ||
8ffa6875 CL |
2615 | /* |
2616 | * We may have set a slab to be unmergeable during bootstrap. | |
2617 | */ | |
2618 | if (s->refcount < 0) | |
2619 | return 1; | |
2620 | ||
81819f0f CL |
2621 | return 0; |
2622 | } | |
2623 | ||
2624 | static struct kmem_cache *find_mergeable(size_t size, | |
2625 | size_t align, unsigned long flags, | |
c59def9f | 2626 | void (*ctor)(void *, struct kmem_cache *, unsigned long)) |
81819f0f | 2627 | { |
5b95a4ac | 2628 | struct kmem_cache *s; |
81819f0f CL |
2629 | |
2630 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
2631 | return NULL; | |
2632 | ||
c59def9f | 2633 | if (ctor) |
81819f0f CL |
2634 | return NULL; |
2635 | ||
2636 | size = ALIGN(size, sizeof(void *)); | |
2637 | align = calculate_alignment(flags, align, size); | |
2638 | size = ALIGN(size, align); | |
2639 | ||
5b95a4ac | 2640 | list_for_each_entry(s, &slab_caches, list) { |
81819f0f CL |
2641 | if (slab_unmergeable(s)) |
2642 | continue; | |
2643 | ||
2644 | if (size > s->size) | |
2645 | continue; | |
2646 | ||
2647 | if (((flags | slub_debug) & SLUB_MERGE_SAME) != | |
2648 | (s->flags & SLUB_MERGE_SAME)) | |
2649 | continue; | |
2650 | /* | |
2651 | * Check if alignment is compatible. | |
2652 | * Courtesy of Adrian Drzewiecki | |
2653 | */ | |
2654 | if ((s->size & ~(align -1)) != s->size) | |
2655 | continue; | |
2656 | ||
2657 | if (s->size - size >= sizeof(void *)) | |
2658 | continue; | |
2659 | ||
2660 | return s; | |
2661 | } | |
2662 | return NULL; | |
2663 | } | |
2664 | ||
2665 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
2666 | size_t align, unsigned long flags, | |
2667 | void (*ctor)(void *, struct kmem_cache *, unsigned long), | |
2668 | void (*dtor)(void *, struct kmem_cache *, unsigned long)) | |
2669 | { | |
2670 | struct kmem_cache *s; | |
2671 | ||
c59def9f | 2672 | BUG_ON(dtor); |
81819f0f | 2673 | down_write(&slub_lock); |
c59def9f | 2674 | s = find_mergeable(size, align, flags, ctor); |
81819f0f CL |
2675 | if (s) { |
2676 | s->refcount++; | |
2677 | /* | |
2678 | * Adjust the object sizes so that we clear | |
2679 | * the complete object on kzalloc. | |
2680 | */ | |
2681 | s->objsize = max(s->objsize, (int)size); | |
2682 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
a0e1d1be | 2683 | up_write(&slub_lock); |
81819f0f CL |
2684 | if (sysfs_slab_alias(s, name)) |
2685 | goto err; | |
a0e1d1be CL |
2686 | return s; |
2687 | } | |
2688 | s = kmalloc(kmem_size, GFP_KERNEL); | |
2689 | if (s) { | |
2690 | if (kmem_cache_open(s, GFP_KERNEL, name, | |
c59def9f | 2691 | size, align, flags, ctor)) { |
81819f0f | 2692 | list_add(&s->list, &slab_caches); |
a0e1d1be CL |
2693 | up_write(&slub_lock); |
2694 | if (sysfs_slab_add(s)) | |
2695 | goto err; | |
2696 | return s; | |
2697 | } | |
2698 | kfree(s); | |
81819f0f CL |
2699 | } |
2700 | up_write(&slub_lock); | |
81819f0f CL |
2701 | |
2702 | err: | |
81819f0f CL |
2703 | if (flags & SLAB_PANIC) |
2704 | panic("Cannot create slabcache %s\n", name); | |
2705 | else | |
2706 | s = NULL; | |
2707 | return s; | |
2708 | } | |
2709 | EXPORT_SYMBOL(kmem_cache_create); | |
2710 | ||
81819f0f | 2711 | #ifdef CONFIG_SMP |
81819f0f | 2712 | /* |
672bba3a CL |
2713 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
2714 | * necessary. | |
81819f0f CL |
2715 | */ |
2716 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
2717 | unsigned long action, void *hcpu) | |
2718 | { | |
2719 | long cpu = (long)hcpu; | |
5b95a4ac CL |
2720 | struct kmem_cache *s; |
2721 | unsigned long flags; | |
81819f0f CL |
2722 | |
2723 | switch (action) { | |
2724 | case CPU_UP_CANCELED: | |
8bb78442 | 2725 | case CPU_UP_CANCELED_FROZEN: |
81819f0f | 2726 | case CPU_DEAD: |
8bb78442 | 2727 | case CPU_DEAD_FROZEN: |
5b95a4ac CL |
2728 | down_read(&slub_lock); |
2729 | list_for_each_entry(s, &slab_caches, list) { | |
2730 | local_irq_save(flags); | |
2731 | __flush_cpu_slab(s, cpu); | |
2732 | local_irq_restore(flags); | |
2733 | } | |
2734 | up_read(&slub_lock); | |
81819f0f CL |
2735 | break; |
2736 | default: | |
2737 | break; | |
2738 | } | |
2739 | return NOTIFY_OK; | |
2740 | } | |
2741 | ||
2742 | static struct notifier_block __cpuinitdata slab_notifier = | |
2743 | { &slab_cpuup_callback, NULL, 0 }; | |
2744 | ||
2745 | #endif | |
2746 | ||
81819f0f CL |
2747 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) |
2748 | { | |
2749 | struct kmem_cache *s = get_slab(size, gfpflags); | |
81819f0f | 2750 | |
6cb8f913 CL |
2751 | if (ZERO_OR_NULL_PTR(s)) |
2752 | return s; | |
81819f0f | 2753 | |
ce15fea8 | 2754 | return slab_alloc(s, gfpflags, -1, caller); |
81819f0f CL |
2755 | } |
2756 | ||
2757 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
2758 | int node, void *caller) | |
2759 | { | |
2760 | struct kmem_cache *s = get_slab(size, gfpflags); | |
81819f0f | 2761 | |
6cb8f913 CL |
2762 | if (ZERO_OR_NULL_PTR(s)) |
2763 | return s; | |
81819f0f | 2764 | |
ce15fea8 | 2765 | return slab_alloc(s, gfpflags, node, caller); |
81819f0f CL |
2766 | } |
2767 | ||
41ecc55b | 2768 | #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) |
434e245d CL |
2769 | static int validate_slab(struct kmem_cache *s, struct page *page, |
2770 | unsigned long *map) | |
53e15af0 CL |
2771 | { |
2772 | void *p; | |
2773 | void *addr = page_address(page); | |
53e15af0 CL |
2774 | |
2775 | if (!check_slab(s, page) || | |
2776 | !on_freelist(s, page, NULL)) | |
2777 | return 0; | |
2778 | ||
2779 | /* Now we know that a valid freelist exists */ | |
2780 | bitmap_zero(map, s->objects); | |
2781 | ||
7656c72b CL |
2782 | for_each_free_object(p, s, page->freelist) { |
2783 | set_bit(slab_index(p, s, addr), map); | |
53e15af0 CL |
2784 | if (!check_object(s, page, p, 0)) |
2785 | return 0; | |
2786 | } | |
2787 | ||
7656c72b CL |
2788 | for_each_object(p, s, addr) |
2789 | if (!test_bit(slab_index(p, s, addr), map)) | |
53e15af0 CL |
2790 | if (!check_object(s, page, p, 1)) |
2791 | return 0; | |
2792 | return 1; | |
2793 | } | |
2794 | ||
434e245d CL |
2795 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, |
2796 | unsigned long *map) | |
53e15af0 CL |
2797 | { |
2798 | if (slab_trylock(page)) { | |
434e245d | 2799 | validate_slab(s, page, map); |
53e15af0 CL |
2800 | slab_unlock(page); |
2801 | } else | |
2802 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", | |
2803 | s->name, page); | |
2804 | ||
2805 | if (s->flags & DEBUG_DEFAULT_FLAGS) { | |
35e5d7ee CL |
2806 | if (!SlabDebug(page)) |
2807 | printk(KERN_ERR "SLUB %s: SlabDebug not set " | |
53e15af0 CL |
2808 | "on slab 0x%p\n", s->name, page); |
2809 | } else { | |
35e5d7ee CL |
2810 | if (SlabDebug(page)) |
2811 | printk(KERN_ERR "SLUB %s: SlabDebug set on " | |
53e15af0 CL |
2812 | "slab 0x%p\n", s->name, page); |
2813 | } | |
2814 | } | |
2815 | ||
434e245d CL |
2816 | static int validate_slab_node(struct kmem_cache *s, |
2817 | struct kmem_cache_node *n, unsigned long *map) | |
53e15af0 CL |
2818 | { |
2819 | unsigned long count = 0; | |
2820 | struct page *page; | |
2821 | unsigned long flags; | |
2822 | ||
2823 | spin_lock_irqsave(&n->list_lock, flags); | |
2824 | ||
2825 | list_for_each_entry(page, &n->partial, lru) { | |
434e245d | 2826 | validate_slab_slab(s, page, map); |
53e15af0 CL |
2827 | count++; |
2828 | } | |
2829 | if (count != n->nr_partial) | |
2830 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
2831 | "counter=%ld\n", s->name, count, n->nr_partial); | |
2832 | ||
2833 | if (!(s->flags & SLAB_STORE_USER)) | |
2834 | goto out; | |
2835 | ||
2836 | list_for_each_entry(page, &n->full, lru) { | |
434e245d | 2837 | validate_slab_slab(s, page, map); |
53e15af0 CL |
2838 | count++; |
2839 | } | |
2840 | if (count != atomic_long_read(&n->nr_slabs)) | |
2841 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
2842 | "counter=%ld\n", s->name, count, | |
2843 | atomic_long_read(&n->nr_slabs)); | |
2844 | ||
2845 | out: | |
2846 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2847 | return count; | |
2848 | } | |
2849 | ||
434e245d | 2850 | static long validate_slab_cache(struct kmem_cache *s) |
53e15af0 CL |
2851 | { |
2852 | int node; | |
2853 | unsigned long count = 0; | |
434e245d CL |
2854 | unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) * |
2855 | sizeof(unsigned long), GFP_KERNEL); | |
2856 | ||
2857 | if (!map) | |
2858 | return -ENOMEM; | |
53e15af0 CL |
2859 | |
2860 | flush_all(s); | |
2861 | for_each_online_node(node) { | |
2862 | struct kmem_cache_node *n = get_node(s, node); | |
2863 | ||
434e245d | 2864 | count += validate_slab_node(s, n, map); |
53e15af0 | 2865 | } |
434e245d | 2866 | kfree(map); |
53e15af0 CL |
2867 | return count; |
2868 | } | |
2869 | ||
b3459709 CL |
2870 | #ifdef SLUB_RESILIENCY_TEST |
2871 | static void resiliency_test(void) | |
2872 | { | |
2873 | u8 *p; | |
2874 | ||
2875 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
2876 | printk(KERN_ERR "-----------------------\n"); | |
2877 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
2878 | ||
2879 | p = kzalloc(16, GFP_KERNEL); | |
2880 | p[16] = 0x12; | |
2881 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
2882 | " 0x12->0x%p\n\n", p + 16); | |
2883 | ||
2884 | validate_slab_cache(kmalloc_caches + 4); | |
2885 | ||
2886 | /* Hmmm... The next two are dangerous */ | |
2887 | p = kzalloc(32, GFP_KERNEL); | |
2888 | p[32 + sizeof(void *)] = 0x34; | |
2889 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
2890 | " 0x34 -> -0x%p\n", p); | |
2891 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
2892 | ||
2893 | validate_slab_cache(kmalloc_caches + 5); | |
2894 | p = kzalloc(64, GFP_KERNEL); | |
2895 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
2896 | *p = 0x56; | |
2897 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
2898 | p); | |
2899 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
2900 | validate_slab_cache(kmalloc_caches + 6); | |
2901 | ||
2902 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
2903 | p = kzalloc(128, GFP_KERNEL); | |
2904 | kfree(p); | |
2905 | *p = 0x78; | |
2906 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
2907 | validate_slab_cache(kmalloc_caches + 7); | |
2908 | ||
2909 | p = kzalloc(256, GFP_KERNEL); | |
2910 | kfree(p); | |
2911 | p[50] = 0x9a; | |
2912 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); | |
2913 | validate_slab_cache(kmalloc_caches + 8); | |
2914 | ||
2915 | p = kzalloc(512, GFP_KERNEL); | |
2916 | kfree(p); | |
2917 | p[512] = 0xab; | |
2918 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
2919 | validate_slab_cache(kmalloc_caches + 9); | |
2920 | } | |
2921 | #else | |
2922 | static void resiliency_test(void) {}; | |
2923 | #endif | |
2924 | ||
88a420e4 | 2925 | /* |
672bba3a | 2926 | * Generate lists of code addresses where slabcache objects are allocated |
88a420e4 CL |
2927 | * and freed. |
2928 | */ | |
2929 | ||
2930 | struct location { | |
2931 | unsigned long count; | |
2932 | void *addr; | |
45edfa58 CL |
2933 | long long sum_time; |
2934 | long min_time; | |
2935 | long max_time; | |
2936 | long min_pid; | |
2937 | long max_pid; | |
2938 | cpumask_t cpus; | |
2939 | nodemask_t nodes; | |
88a420e4 CL |
2940 | }; |
2941 | ||
2942 | struct loc_track { | |
2943 | unsigned long max; | |
2944 | unsigned long count; | |
2945 | struct location *loc; | |
2946 | }; | |
2947 | ||
2948 | static void free_loc_track(struct loc_track *t) | |
2949 | { | |
2950 | if (t->max) | |
2951 | free_pages((unsigned long)t->loc, | |
2952 | get_order(sizeof(struct location) * t->max)); | |
2953 | } | |
2954 | ||
68dff6a9 | 2955 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
88a420e4 CL |
2956 | { |
2957 | struct location *l; | |
2958 | int order; | |
2959 | ||
88a420e4 CL |
2960 | order = get_order(sizeof(struct location) * max); |
2961 | ||
68dff6a9 | 2962 | l = (void *)__get_free_pages(flags, order); |
88a420e4 CL |
2963 | if (!l) |
2964 | return 0; | |
2965 | ||
2966 | if (t->count) { | |
2967 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
2968 | free_loc_track(t); | |
2969 | } | |
2970 | t->max = max; | |
2971 | t->loc = l; | |
2972 | return 1; | |
2973 | } | |
2974 | ||
2975 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
45edfa58 | 2976 | const struct track *track) |
88a420e4 CL |
2977 | { |
2978 | long start, end, pos; | |
2979 | struct location *l; | |
2980 | void *caddr; | |
45edfa58 | 2981 | unsigned long age = jiffies - track->when; |
88a420e4 CL |
2982 | |
2983 | start = -1; | |
2984 | end = t->count; | |
2985 | ||
2986 | for ( ; ; ) { | |
2987 | pos = start + (end - start + 1) / 2; | |
2988 | ||
2989 | /* | |
2990 | * There is nothing at "end". If we end up there | |
2991 | * we need to add something to before end. | |
2992 | */ | |
2993 | if (pos == end) | |
2994 | break; | |
2995 | ||
2996 | caddr = t->loc[pos].addr; | |
45edfa58 CL |
2997 | if (track->addr == caddr) { |
2998 | ||
2999 | l = &t->loc[pos]; | |
3000 | l->count++; | |
3001 | if (track->when) { | |
3002 | l->sum_time += age; | |
3003 | if (age < l->min_time) | |
3004 | l->min_time = age; | |
3005 | if (age > l->max_time) | |
3006 | l->max_time = age; | |
3007 | ||
3008 | if (track->pid < l->min_pid) | |
3009 | l->min_pid = track->pid; | |
3010 | if (track->pid > l->max_pid) | |
3011 | l->max_pid = track->pid; | |
3012 | ||
3013 | cpu_set(track->cpu, l->cpus); | |
3014 | } | |
3015 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
88a420e4 CL |
3016 | return 1; |
3017 | } | |
3018 | ||
45edfa58 | 3019 | if (track->addr < caddr) |
88a420e4 CL |
3020 | end = pos; |
3021 | else | |
3022 | start = pos; | |
3023 | } | |
3024 | ||
3025 | /* | |
672bba3a | 3026 | * Not found. Insert new tracking element. |
88a420e4 | 3027 | */ |
68dff6a9 | 3028 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
88a420e4 CL |
3029 | return 0; |
3030 | ||
3031 | l = t->loc + pos; | |
3032 | if (pos < t->count) | |
3033 | memmove(l + 1, l, | |
3034 | (t->count - pos) * sizeof(struct location)); | |
3035 | t->count++; | |
3036 | l->count = 1; | |
45edfa58 CL |
3037 | l->addr = track->addr; |
3038 | l->sum_time = age; | |
3039 | l->min_time = age; | |
3040 | l->max_time = age; | |
3041 | l->min_pid = track->pid; | |
3042 | l->max_pid = track->pid; | |
3043 | cpus_clear(l->cpus); | |
3044 | cpu_set(track->cpu, l->cpus); | |
3045 | nodes_clear(l->nodes); | |
3046 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
88a420e4 CL |
3047 | return 1; |
3048 | } | |
3049 | ||
3050 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
3051 | struct page *page, enum track_item alloc) | |
3052 | { | |
3053 | void *addr = page_address(page); | |
7656c72b | 3054 | DECLARE_BITMAP(map, s->objects); |
88a420e4 CL |
3055 | void *p; |
3056 | ||
3057 | bitmap_zero(map, s->objects); | |
7656c72b CL |
3058 | for_each_free_object(p, s, page->freelist) |
3059 | set_bit(slab_index(p, s, addr), map); | |
88a420e4 | 3060 | |
7656c72b | 3061 | for_each_object(p, s, addr) |
45edfa58 CL |
3062 | if (!test_bit(slab_index(p, s, addr), map)) |
3063 | add_location(t, s, get_track(s, p, alloc)); | |
88a420e4 CL |
3064 | } |
3065 | ||
3066 | static int list_locations(struct kmem_cache *s, char *buf, | |
3067 | enum track_item alloc) | |
3068 | { | |
3069 | int n = 0; | |
3070 | unsigned long i; | |
68dff6a9 | 3071 | struct loc_track t = { 0, 0, NULL }; |
88a420e4 CL |
3072 | int node; |
3073 | ||
68dff6a9 CL |
3074 | if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), |
3075 | GFP_KERNEL)) | |
3076 | return sprintf(buf, "Out of memory\n"); | |
88a420e4 CL |
3077 | |
3078 | /* Push back cpu slabs */ | |
3079 | flush_all(s); | |
3080 | ||
3081 | for_each_online_node(node) { | |
3082 | struct kmem_cache_node *n = get_node(s, node); | |
3083 | unsigned long flags; | |
3084 | struct page *page; | |
3085 | ||
3086 | if (!atomic_read(&n->nr_slabs)) | |
3087 | continue; | |
3088 | ||
3089 | spin_lock_irqsave(&n->list_lock, flags); | |
3090 | list_for_each_entry(page, &n->partial, lru) | |
3091 | process_slab(&t, s, page, alloc); | |
3092 | list_for_each_entry(page, &n->full, lru) | |
3093 | process_slab(&t, s, page, alloc); | |
3094 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3095 | } | |
3096 | ||
3097 | for (i = 0; i < t.count; i++) { | |
45edfa58 | 3098 | struct location *l = &t.loc[i]; |
88a420e4 CL |
3099 | |
3100 | if (n > PAGE_SIZE - 100) | |
3101 | break; | |
45edfa58 CL |
3102 | n += sprintf(buf + n, "%7ld ", l->count); |
3103 | ||
3104 | if (l->addr) | |
3105 | n += sprint_symbol(buf + n, (unsigned long)l->addr); | |
88a420e4 CL |
3106 | else |
3107 | n += sprintf(buf + n, "<not-available>"); | |
45edfa58 CL |
3108 | |
3109 | if (l->sum_time != l->min_time) { | |
3110 | unsigned long remainder; | |
3111 | ||
3112 | n += sprintf(buf + n, " age=%ld/%ld/%ld", | |
3113 | l->min_time, | |
3114 | div_long_long_rem(l->sum_time, l->count, &remainder), | |
3115 | l->max_time); | |
3116 | } else | |
3117 | n += sprintf(buf + n, " age=%ld", | |
3118 | l->min_time); | |
3119 | ||
3120 | if (l->min_pid != l->max_pid) | |
3121 | n += sprintf(buf + n, " pid=%ld-%ld", | |
3122 | l->min_pid, l->max_pid); | |
3123 | else | |
3124 | n += sprintf(buf + n, " pid=%ld", | |
3125 | l->min_pid); | |
3126 | ||
84966343 CL |
3127 | if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && |
3128 | n < PAGE_SIZE - 60) { | |
45edfa58 CL |
3129 | n += sprintf(buf + n, " cpus="); |
3130 | n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50, | |
3131 | l->cpus); | |
3132 | } | |
3133 | ||
84966343 CL |
3134 | if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && |
3135 | n < PAGE_SIZE - 60) { | |
45edfa58 CL |
3136 | n += sprintf(buf + n, " nodes="); |
3137 | n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50, | |
3138 | l->nodes); | |
3139 | } | |
3140 | ||
88a420e4 CL |
3141 | n += sprintf(buf + n, "\n"); |
3142 | } | |
3143 | ||
3144 | free_loc_track(&t); | |
3145 | if (!t.count) | |
3146 | n += sprintf(buf, "No data\n"); | |
3147 | return n; | |
3148 | } | |
3149 | ||
81819f0f CL |
3150 | static unsigned long count_partial(struct kmem_cache_node *n) |
3151 | { | |
3152 | unsigned long flags; | |
3153 | unsigned long x = 0; | |
3154 | struct page *page; | |
3155 | ||
3156 | spin_lock_irqsave(&n->list_lock, flags); | |
3157 | list_for_each_entry(page, &n->partial, lru) | |
3158 | x += page->inuse; | |
3159 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3160 | return x; | |
3161 | } | |
3162 | ||
3163 | enum slab_stat_type { | |
3164 | SL_FULL, | |
3165 | SL_PARTIAL, | |
3166 | SL_CPU, | |
3167 | SL_OBJECTS | |
3168 | }; | |
3169 | ||
3170 | #define SO_FULL (1 << SL_FULL) | |
3171 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
3172 | #define SO_CPU (1 << SL_CPU) | |
3173 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
3174 | ||
3175 | static unsigned long slab_objects(struct kmem_cache *s, | |
3176 | char *buf, unsigned long flags) | |
3177 | { | |
3178 | unsigned long total = 0; | |
3179 | int cpu; | |
3180 | int node; | |
3181 | int x; | |
3182 | unsigned long *nodes; | |
3183 | unsigned long *per_cpu; | |
3184 | ||
3185 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
3186 | per_cpu = nodes + nr_node_ids; | |
3187 | ||
3188 | for_each_possible_cpu(cpu) { | |
3189 | struct page *page = s->cpu_slab[cpu]; | |
3190 | int node; | |
3191 | ||
3192 | if (page) { | |
3193 | node = page_to_nid(page); | |
3194 | if (flags & SO_CPU) { | |
3195 | int x = 0; | |
3196 | ||
3197 | if (flags & SO_OBJECTS) | |
3198 | x = page->inuse; | |
3199 | else | |
3200 | x = 1; | |
3201 | total += x; | |
3202 | nodes[node] += x; | |
3203 | } | |
3204 | per_cpu[node]++; | |
3205 | } | |
3206 | } | |
3207 | ||
3208 | for_each_online_node(node) { | |
3209 | struct kmem_cache_node *n = get_node(s, node); | |
3210 | ||
3211 | if (flags & SO_PARTIAL) { | |
3212 | if (flags & SO_OBJECTS) | |
3213 | x = count_partial(n); | |
3214 | else | |
3215 | x = n->nr_partial; | |
3216 | total += x; | |
3217 | nodes[node] += x; | |
3218 | } | |
3219 | ||
3220 | if (flags & SO_FULL) { | |
3221 | int full_slabs = atomic_read(&n->nr_slabs) | |
3222 | - per_cpu[node] | |
3223 | - n->nr_partial; | |
3224 | ||
3225 | if (flags & SO_OBJECTS) | |
3226 | x = full_slabs * s->objects; | |
3227 | else | |
3228 | x = full_slabs; | |
3229 | total += x; | |
3230 | nodes[node] += x; | |
3231 | } | |
3232 | } | |
3233 | ||
3234 | x = sprintf(buf, "%lu", total); | |
3235 | #ifdef CONFIG_NUMA | |
3236 | for_each_online_node(node) | |
3237 | if (nodes[node]) | |
3238 | x += sprintf(buf + x, " N%d=%lu", | |
3239 | node, nodes[node]); | |
3240 | #endif | |
3241 | kfree(nodes); | |
3242 | return x + sprintf(buf + x, "\n"); | |
3243 | } | |
3244 | ||
3245 | static int any_slab_objects(struct kmem_cache *s) | |
3246 | { | |
3247 | int node; | |
3248 | int cpu; | |
3249 | ||
3250 | for_each_possible_cpu(cpu) | |
3251 | if (s->cpu_slab[cpu]) | |
3252 | return 1; | |
3253 | ||
3254 | for_each_node(node) { | |
3255 | struct kmem_cache_node *n = get_node(s, node); | |
3256 | ||
3257 | if (n->nr_partial || atomic_read(&n->nr_slabs)) | |
3258 | return 1; | |
3259 | } | |
3260 | return 0; | |
3261 | } | |
3262 | ||
3263 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
3264 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); | |
3265 | ||
3266 | struct slab_attribute { | |
3267 | struct attribute attr; | |
3268 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
3269 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
3270 | }; | |
3271 | ||
3272 | #define SLAB_ATTR_RO(_name) \ | |
3273 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) | |
3274 | ||
3275 | #define SLAB_ATTR(_name) \ | |
3276 | static struct slab_attribute _name##_attr = \ | |
3277 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3278 | ||
81819f0f CL |
3279 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
3280 | { | |
3281 | return sprintf(buf, "%d\n", s->size); | |
3282 | } | |
3283 | SLAB_ATTR_RO(slab_size); | |
3284 | ||
3285 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
3286 | { | |
3287 | return sprintf(buf, "%d\n", s->align); | |
3288 | } | |
3289 | SLAB_ATTR_RO(align); | |
3290 | ||
3291 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
3292 | { | |
3293 | return sprintf(buf, "%d\n", s->objsize); | |
3294 | } | |
3295 | SLAB_ATTR_RO(object_size); | |
3296 | ||
3297 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
3298 | { | |
3299 | return sprintf(buf, "%d\n", s->objects); | |
3300 | } | |
3301 | SLAB_ATTR_RO(objs_per_slab); | |
3302 | ||
3303 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
3304 | { | |
3305 | return sprintf(buf, "%d\n", s->order); | |
3306 | } | |
3307 | SLAB_ATTR_RO(order); | |
3308 | ||
3309 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
3310 | { | |
3311 | if (s->ctor) { | |
3312 | int n = sprint_symbol(buf, (unsigned long)s->ctor); | |
3313 | ||
3314 | return n + sprintf(buf + n, "\n"); | |
3315 | } | |
3316 | return 0; | |
3317 | } | |
3318 | SLAB_ATTR_RO(ctor); | |
3319 | ||
81819f0f CL |
3320 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
3321 | { | |
3322 | return sprintf(buf, "%d\n", s->refcount - 1); | |
3323 | } | |
3324 | SLAB_ATTR_RO(aliases); | |
3325 | ||
3326 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
3327 | { | |
3328 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU); | |
3329 | } | |
3330 | SLAB_ATTR_RO(slabs); | |
3331 | ||
3332 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
3333 | { | |
3334 | return slab_objects(s, buf, SO_PARTIAL); | |
3335 | } | |
3336 | SLAB_ATTR_RO(partial); | |
3337 | ||
3338 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
3339 | { | |
3340 | return slab_objects(s, buf, SO_CPU); | |
3341 | } | |
3342 | SLAB_ATTR_RO(cpu_slabs); | |
3343 | ||
3344 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
3345 | { | |
3346 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS); | |
3347 | } | |
3348 | SLAB_ATTR_RO(objects); | |
3349 | ||
3350 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
3351 | { | |
3352 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
3353 | } | |
3354 | ||
3355 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
3356 | const char *buf, size_t length) | |
3357 | { | |
3358 | s->flags &= ~SLAB_DEBUG_FREE; | |
3359 | if (buf[0] == '1') | |
3360 | s->flags |= SLAB_DEBUG_FREE; | |
3361 | return length; | |
3362 | } | |
3363 | SLAB_ATTR(sanity_checks); | |
3364 | ||
3365 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
3366 | { | |
3367 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
3368 | } | |
3369 | ||
3370 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
3371 | size_t length) | |
3372 | { | |
3373 | s->flags &= ~SLAB_TRACE; | |
3374 | if (buf[0] == '1') | |
3375 | s->flags |= SLAB_TRACE; | |
3376 | return length; | |
3377 | } | |
3378 | SLAB_ATTR(trace); | |
3379 | ||
3380 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
3381 | { | |
3382 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
3383 | } | |
3384 | ||
3385 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
3386 | const char *buf, size_t length) | |
3387 | { | |
3388 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
3389 | if (buf[0] == '1') | |
3390 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
3391 | return length; | |
3392 | } | |
3393 | SLAB_ATTR(reclaim_account); | |
3394 | ||
3395 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
3396 | { | |
5af60839 | 3397 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); |
81819f0f CL |
3398 | } |
3399 | SLAB_ATTR_RO(hwcache_align); | |
3400 | ||
3401 | #ifdef CONFIG_ZONE_DMA | |
3402 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
3403 | { | |
3404 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
3405 | } | |
3406 | SLAB_ATTR_RO(cache_dma); | |
3407 | #endif | |
3408 | ||
3409 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
3410 | { | |
3411 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
3412 | } | |
3413 | SLAB_ATTR_RO(destroy_by_rcu); | |
3414 | ||
3415 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
3416 | { | |
3417 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
3418 | } | |
3419 | ||
3420 | static ssize_t red_zone_store(struct kmem_cache *s, | |
3421 | const char *buf, size_t length) | |
3422 | { | |
3423 | if (any_slab_objects(s)) | |
3424 | return -EBUSY; | |
3425 | ||
3426 | s->flags &= ~SLAB_RED_ZONE; | |
3427 | if (buf[0] == '1') | |
3428 | s->flags |= SLAB_RED_ZONE; | |
3429 | calculate_sizes(s); | |
3430 | return length; | |
3431 | } | |
3432 | SLAB_ATTR(red_zone); | |
3433 | ||
3434 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
3435 | { | |
3436 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
3437 | } | |
3438 | ||
3439 | static ssize_t poison_store(struct kmem_cache *s, | |
3440 | const char *buf, size_t length) | |
3441 | { | |
3442 | if (any_slab_objects(s)) | |
3443 | return -EBUSY; | |
3444 | ||
3445 | s->flags &= ~SLAB_POISON; | |
3446 | if (buf[0] == '1') | |
3447 | s->flags |= SLAB_POISON; | |
3448 | calculate_sizes(s); | |
3449 | return length; | |
3450 | } | |
3451 | SLAB_ATTR(poison); | |
3452 | ||
3453 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
3454 | { | |
3455 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
3456 | } | |
3457 | ||
3458 | static ssize_t store_user_store(struct kmem_cache *s, | |
3459 | const char *buf, size_t length) | |
3460 | { | |
3461 | if (any_slab_objects(s)) | |
3462 | return -EBUSY; | |
3463 | ||
3464 | s->flags &= ~SLAB_STORE_USER; | |
3465 | if (buf[0] == '1') | |
3466 | s->flags |= SLAB_STORE_USER; | |
3467 | calculate_sizes(s); | |
3468 | return length; | |
3469 | } | |
3470 | SLAB_ATTR(store_user); | |
3471 | ||
53e15af0 CL |
3472 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
3473 | { | |
3474 | return 0; | |
3475 | } | |
3476 | ||
3477 | static ssize_t validate_store(struct kmem_cache *s, | |
3478 | const char *buf, size_t length) | |
3479 | { | |
434e245d CL |
3480 | int ret = -EINVAL; |
3481 | ||
3482 | if (buf[0] == '1') { | |
3483 | ret = validate_slab_cache(s); | |
3484 | if (ret >= 0) | |
3485 | ret = length; | |
3486 | } | |
3487 | return ret; | |
53e15af0 CL |
3488 | } |
3489 | SLAB_ATTR(validate); | |
3490 | ||
2086d26a CL |
3491 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
3492 | { | |
3493 | return 0; | |
3494 | } | |
3495 | ||
3496 | static ssize_t shrink_store(struct kmem_cache *s, | |
3497 | const char *buf, size_t length) | |
3498 | { | |
3499 | if (buf[0] == '1') { | |
3500 | int rc = kmem_cache_shrink(s); | |
3501 | ||
3502 | if (rc) | |
3503 | return rc; | |
3504 | } else | |
3505 | return -EINVAL; | |
3506 | return length; | |
3507 | } | |
3508 | SLAB_ATTR(shrink); | |
3509 | ||
88a420e4 CL |
3510 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) |
3511 | { | |
3512 | if (!(s->flags & SLAB_STORE_USER)) | |
3513 | return -ENOSYS; | |
3514 | return list_locations(s, buf, TRACK_ALLOC); | |
3515 | } | |
3516 | SLAB_ATTR_RO(alloc_calls); | |
3517 | ||
3518 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
3519 | { | |
3520 | if (!(s->flags & SLAB_STORE_USER)) | |
3521 | return -ENOSYS; | |
3522 | return list_locations(s, buf, TRACK_FREE); | |
3523 | } | |
3524 | SLAB_ATTR_RO(free_calls); | |
3525 | ||
81819f0f CL |
3526 | #ifdef CONFIG_NUMA |
3527 | static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf) | |
3528 | { | |
3529 | return sprintf(buf, "%d\n", s->defrag_ratio / 10); | |
3530 | } | |
3531 | ||
3532 | static ssize_t defrag_ratio_store(struct kmem_cache *s, | |
3533 | const char *buf, size_t length) | |
3534 | { | |
3535 | int n = simple_strtoul(buf, NULL, 10); | |
3536 | ||
3537 | if (n < 100) | |
3538 | s->defrag_ratio = n * 10; | |
3539 | return length; | |
3540 | } | |
3541 | SLAB_ATTR(defrag_ratio); | |
3542 | #endif | |
3543 | ||
3544 | static struct attribute * slab_attrs[] = { | |
3545 | &slab_size_attr.attr, | |
3546 | &object_size_attr.attr, | |
3547 | &objs_per_slab_attr.attr, | |
3548 | &order_attr.attr, | |
3549 | &objects_attr.attr, | |
3550 | &slabs_attr.attr, | |
3551 | &partial_attr.attr, | |
3552 | &cpu_slabs_attr.attr, | |
3553 | &ctor_attr.attr, | |
81819f0f CL |
3554 | &aliases_attr.attr, |
3555 | &align_attr.attr, | |
3556 | &sanity_checks_attr.attr, | |
3557 | &trace_attr.attr, | |
3558 | &hwcache_align_attr.attr, | |
3559 | &reclaim_account_attr.attr, | |
3560 | &destroy_by_rcu_attr.attr, | |
3561 | &red_zone_attr.attr, | |
3562 | &poison_attr.attr, | |
3563 | &store_user_attr.attr, | |
53e15af0 | 3564 | &validate_attr.attr, |
2086d26a | 3565 | &shrink_attr.attr, |
88a420e4 CL |
3566 | &alloc_calls_attr.attr, |
3567 | &free_calls_attr.attr, | |
81819f0f CL |
3568 | #ifdef CONFIG_ZONE_DMA |
3569 | &cache_dma_attr.attr, | |
3570 | #endif | |
3571 | #ifdef CONFIG_NUMA | |
3572 | &defrag_ratio_attr.attr, | |
3573 | #endif | |
3574 | NULL | |
3575 | }; | |
3576 | ||
3577 | static struct attribute_group slab_attr_group = { | |
3578 | .attrs = slab_attrs, | |
3579 | }; | |
3580 | ||
3581 | static ssize_t slab_attr_show(struct kobject *kobj, | |
3582 | struct attribute *attr, | |
3583 | char *buf) | |
3584 | { | |
3585 | struct slab_attribute *attribute; | |
3586 | struct kmem_cache *s; | |
3587 | int err; | |
3588 | ||
3589 | attribute = to_slab_attr(attr); | |
3590 | s = to_slab(kobj); | |
3591 | ||
3592 | if (!attribute->show) | |
3593 | return -EIO; | |
3594 | ||
3595 | err = attribute->show(s, buf); | |
3596 | ||
3597 | return err; | |
3598 | } | |
3599 | ||
3600 | static ssize_t slab_attr_store(struct kobject *kobj, | |
3601 | struct attribute *attr, | |
3602 | const char *buf, size_t len) | |
3603 | { | |
3604 | struct slab_attribute *attribute; | |
3605 | struct kmem_cache *s; | |
3606 | int err; | |
3607 | ||
3608 | attribute = to_slab_attr(attr); | |
3609 | s = to_slab(kobj); | |
3610 | ||
3611 | if (!attribute->store) | |
3612 | return -EIO; | |
3613 | ||
3614 | err = attribute->store(s, buf, len); | |
3615 | ||
3616 | return err; | |
3617 | } | |
3618 | ||
3619 | static struct sysfs_ops slab_sysfs_ops = { | |
3620 | .show = slab_attr_show, | |
3621 | .store = slab_attr_store, | |
3622 | }; | |
3623 | ||
3624 | static struct kobj_type slab_ktype = { | |
3625 | .sysfs_ops = &slab_sysfs_ops, | |
3626 | }; | |
3627 | ||
3628 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
3629 | { | |
3630 | struct kobj_type *ktype = get_ktype(kobj); | |
3631 | ||
3632 | if (ktype == &slab_ktype) | |
3633 | return 1; | |
3634 | return 0; | |
3635 | } | |
3636 | ||
3637 | static struct kset_uevent_ops slab_uevent_ops = { | |
3638 | .filter = uevent_filter, | |
3639 | }; | |
3640 | ||
5af328a5 | 3641 | static decl_subsys(slab, &slab_ktype, &slab_uevent_ops); |
81819f0f CL |
3642 | |
3643 | #define ID_STR_LENGTH 64 | |
3644 | ||
3645 | /* Create a unique string id for a slab cache: | |
3646 | * format | |
3647 | * :[flags-]size:[memory address of kmemcache] | |
3648 | */ | |
3649 | static char *create_unique_id(struct kmem_cache *s) | |
3650 | { | |
3651 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
3652 | char *p = name; | |
3653 | ||
3654 | BUG_ON(!name); | |
3655 | ||
3656 | *p++ = ':'; | |
3657 | /* | |
3658 | * First flags affecting slabcache operations. We will only | |
3659 | * get here for aliasable slabs so we do not need to support | |
3660 | * too many flags. The flags here must cover all flags that | |
3661 | * are matched during merging to guarantee that the id is | |
3662 | * unique. | |
3663 | */ | |
3664 | if (s->flags & SLAB_CACHE_DMA) | |
3665 | *p++ = 'd'; | |
3666 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
3667 | *p++ = 'a'; | |
3668 | if (s->flags & SLAB_DEBUG_FREE) | |
3669 | *p++ = 'F'; | |
3670 | if (p != name + 1) | |
3671 | *p++ = '-'; | |
3672 | p += sprintf(p, "%07d", s->size); | |
3673 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
3674 | return name; | |
3675 | } | |
3676 | ||
3677 | static int sysfs_slab_add(struct kmem_cache *s) | |
3678 | { | |
3679 | int err; | |
3680 | const char *name; | |
3681 | int unmergeable; | |
3682 | ||
3683 | if (slab_state < SYSFS) | |
3684 | /* Defer until later */ | |
3685 | return 0; | |
3686 | ||
3687 | unmergeable = slab_unmergeable(s); | |
3688 | if (unmergeable) { | |
3689 | /* | |
3690 | * Slabcache can never be merged so we can use the name proper. | |
3691 | * This is typically the case for debug situations. In that | |
3692 | * case we can catch duplicate names easily. | |
3693 | */ | |
0f9008ef | 3694 | sysfs_remove_link(&slab_subsys.kobj, s->name); |
81819f0f CL |
3695 | name = s->name; |
3696 | } else { | |
3697 | /* | |
3698 | * Create a unique name for the slab as a target | |
3699 | * for the symlinks. | |
3700 | */ | |
3701 | name = create_unique_id(s); | |
3702 | } | |
3703 | ||
3704 | kobj_set_kset_s(s, slab_subsys); | |
3705 | kobject_set_name(&s->kobj, name); | |
3706 | kobject_init(&s->kobj); | |
3707 | err = kobject_add(&s->kobj); | |
3708 | if (err) | |
3709 | return err; | |
3710 | ||
3711 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
3712 | if (err) | |
3713 | return err; | |
3714 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
3715 | if (!unmergeable) { | |
3716 | /* Setup first alias */ | |
3717 | sysfs_slab_alias(s, s->name); | |
3718 | kfree(name); | |
3719 | } | |
3720 | return 0; | |
3721 | } | |
3722 | ||
3723 | static void sysfs_slab_remove(struct kmem_cache *s) | |
3724 | { | |
3725 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
3726 | kobject_del(&s->kobj); | |
3727 | } | |
3728 | ||
3729 | /* | |
3730 | * Need to buffer aliases during bootup until sysfs becomes | |
3731 | * available lest we loose that information. | |
3732 | */ | |
3733 | struct saved_alias { | |
3734 | struct kmem_cache *s; | |
3735 | const char *name; | |
3736 | struct saved_alias *next; | |
3737 | }; | |
3738 | ||
5af328a5 | 3739 | static struct saved_alias *alias_list; |
81819f0f CL |
3740 | |
3741 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
3742 | { | |
3743 | struct saved_alias *al; | |
3744 | ||
3745 | if (slab_state == SYSFS) { | |
3746 | /* | |
3747 | * If we have a leftover link then remove it. | |
3748 | */ | |
0f9008ef LT |
3749 | sysfs_remove_link(&slab_subsys.kobj, name); |
3750 | return sysfs_create_link(&slab_subsys.kobj, | |
81819f0f CL |
3751 | &s->kobj, name); |
3752 | } | |
3753 | ||
3754 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
3755 | if (!al) | |
3756 | return -ENOMEM; | |
3757 | ||
3758 | al->s = s; | |
3759 | al->name = name; | |
3760 | al->next = alias_list; | |
3761 | alias_list = al; | |
3762 | return 0; | |
3763 | } | |
3764 | ||
3765 | static int __init slab_sysfs_init(void) | |
3766 | { | |
5b95a4ac | 3767 | struct kmem_cache *s; |
81819f0f CL |
3768 | int err; |
3769 | ||
3770 | err = subsystem_register(&slab_subsys); | |
3771 | if (err) { | |
3772 | printk(KERN_ERR "Cannot register slab subsystem.\n"); | |
3773 | return -ENOSYS; | |
3774 | } | |
3775 | ||
26a7bd03 CL |
3776 | slab_state = SYSFS; |
3777 | ||
5b95a4ac | 3778 | list_for_each_entry(s, &slab_caches, list) { |
26a7bd03 CL |
3779 | err = sysfs_slab_add(s); |
3780 | BUG_ON(err); | |
3781 | } | |
81819f0f CL |
3782 | |
3783 | while (alias_list) { | |
3784 | struct saved_alias *al = alias_list; | |
3785 | ||
3786 | alias_list = alias_list->next; | |
3787 | err = sysfs_slab_alias(al->s, al->name); | |
3788 | BUG_ON(err); | |
3789 | kfree(al); | |
3790 | } | |
3791 | ||
3792 | resiliency_test(); | |
3793 | return 0; | |
3794 | } | |
3795 | ||
3796 | __initcall(slab_sysfs_init); | |
81819f0f | 3797 | #endif |