]>
Commit | Line | Data |
---|---|---|
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 | |
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/proc_fs.h> | |
18 | #include <linux/seq_file.h> | |
19 | #include <linux/cpu.h> | |
20 | #include <linux/cpuset.h> | |
21 | #include <linux/mempolicy.h> | |
22 | #include <linux/ctype.h> | |
23 | #include <linux/debugobjects.h> | |
24 | #include <linux/kallsyms.h> | |
25 | #include <linux/memory.h> | |
26 | #include <linux/math64.h> | |
27 | ||
28 | /* | |
29 | * Lock order: | |
30 | * 1. slab_lock(page) | |
31 | * 2. slab->list_lock | |
32 | * | |
33 | * The slab_lock protects operations on the object of a particular | |
34 | * slab and its metadata in the page struct. If the slab lock | |
35 | * has been taken then no allocations nor frees can be performed | |
36 | * on the objects in the slab nor can the slab be added or removed | |
37 | * from the partial or full lists since this would mean modifying | |
38 | * the page_struct of the slab. | |
39 | * | |
40 | * The list_lock protects the partial and full list on each node and | |
41 | * the partial slab counter. If taken then no new slabs may be added or | |
42 | * removed from the lists nor make the number of partial slabs be modified. | |
43 | * (Note that the total number of slabs is an atomic value that may be | |
44 | * modified without taking the list lock). | |
45 | * | |
46 | * The list_lock is a centralized lock and thus we avoid taking it as | |
47 | * much as possible. As long as SLUB does not have to handle partial | |
48 | * slabs, operations can continue without any centralized lock. F.e. | |
49 | * allocating a long series of objects that fill up slabs does not require | |
50 | * the list lock. | |
51 | * | |
52 | * The lock order is sometimes inverted when we are trying to get a slab | |
53 | * off a list. We take the list_lock and then look for a page on the list | |
54 | * to use. While we do that objects in the slabs may be freed. We can | |
55 | * only operate on the slab if we have also taken the slab_lock. So we use | |
56 | * a slab_trylock() on the slab. If trylock was successful then no frees | |
57 | * can occur anymore and we can use the slab for allocations etc. If the | |
58 | * slab_trylock() does not succeed then frees are in progress in the slab and | |
59 | * we must stay away from it for a while since we may cause a bouncing | |
60 | * cacheline if we try to acquire the lock. So go onto the next slab. | |
61 | * If all pages are busy then we may allocate a new slab instead of reusing | |
62 | * a partial slab. A new slab has noone operating on it and thus there is | |
63 | * no danger of cacheline contention. | |
64 | * | |
65 | * Interrupts are disabled during allocation and deallocation in order to | |
66 | * make the slab allocator safe to use in the context of an irq. In addition | |
67 | * interrupts are disabled to ensure that the processor does not change | |
68 | * while handling per_cpu slabs, due to kernel preemption. | |
69 | * | |
70 | * SLUB assigns one slab for allocation to each processor. | |
71 | * Allocations only occur from these slabs called cpu slabs. | |
72 | * | |
73 | * Slabs with free elements are kept on a partial list and during regular | |
74 | * operations no list for full slabs is used. If an object in a full slab is | |
75 | * freed then the slab will show up again on the partial lists. | |
76 | * We track full slabs for debugging purposes though because otherwise we | |
77 | * cannot scan all objects. | |
78 | * | |
79 | * Slabs are freed when they become empty. Teardown and setup is | |
80 | * minimal so we rely on the page allocators per cpu caches for | |
81 | * fast frees and allocs. | |
82 | * | |
83 | * Overloading of page flags that are otherwise used for LRU management. | |
84 | * | |
85 | * PageActive The slab is frozen and exempt from list processing. | |
86 | * This means that the slab is dedicated to a purpose | |
87 | * such as satisfying allocations for a specific | |
88 | * processor. Objects may be freed in the slab while | |
89 | * it is frozen but slab_free will then skip the usual | |
90 | * list operations. It is up to the processor holding | |
91 | * the slab to integrate the slab into the slab lists | |
92 | * when the slab is no longer needed. | |
93 | * | |
94 | * One use of this flag is to mark slabs that are | |
95 | * used for allocations. Then such a slab becomes a cpu | |
96 | * slab. The cpu slab may be equipped with an additional | |
97 | * freelist that allows lockless access to | |
98 | * free objects in addition to the regular freelist | |
99 | * that requires the slab lock. | |
100 | * | |
101 | * PageError Slab requires special handling due to debug | |
102 | * options set. This moves slab handling out of | |
103 | * the fast path and disables lockless freelists. | |
104 | */ | |
105 | ||
106 | #ifdef CONFIG_SLUB_DEBUG | |
107 | #define SLABDEBUG 1 | |
108 | #else | |
109 | #define SLABDEBUG 0 | |
110 | #endif | |
111 | ||
112 | /* | |
113 | * Issues still to be resolved: | |
114 | * | |
115 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
116 | * | |
117 | * - Variable sizing of the per node arrays | |
118 | */ | |
119 | ||
120 | /* Enable to test recovery from slab corruption on boot */ | |
121 | #undef SLUB_RESILIENCY_TEST | |
122 | ||
123 | /* | |
124 | * Mininum number of partial slabs. These will be left on the partial | |
125 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
126 | */ | |
127 | #define MIN_PARTIAL 5 | |
128 | ||
129 | /* | |
130 | * Maximum number of desirable partial slabs. | |
131 | * The existence of more partial slabs makes kmem_cache_shrink | |
132 | * sort the partial list by the number of objects in the. | |
133 | */ | |
134 | #define MAX_PARTIAL 10 | |
135 | ||
136 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ | |
137 | SLAB_POISON | SLAB_STORE_USER) | |
138 | ||
139 | /* | |
140 | * Set of flags that will prevent slab merging | |
141 | */ | |
142 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
143 | SLAB_TRACE | SLAB_DESTROY_BY_RCU) | |
144 | ||
145 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ | |
146 | SLAB_CACHE_DMA) | |
147 | ||
148 | #ifndef ARCH_KMALLOC_MINALIGN | |
149 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) | |
150 | #endif | |
151 | ||
152 | #ifndef ARCH_SLAB_MINALIGN | |
153 | #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) | |
154 | #endif | |
155 | ||
156 | #define OO_SHIFT 16 | |
157 | #define OO_MASK ((1 << OO_SHIFT) - 1) | |
158 | #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ | |
159 | ||
160 | /* Internal SLUB flags */ | |
161 | #define __OBJECT_POISON 0x80000000 /* Poison object */ | |
162 | #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ | |
163 | ||
164 | static int kmem_size = sizeof(struct kmem_cache); | |
165 | ||
166 | #ifdef CONFIG_SMP | |
167 | static struct notifier_block slab_notifier; | |
168 | #endif | |
169 | ||
170 | static enum { | |
171 | DOWN, /* No slab functionality available */ | |
172 | PARTIAL, /* kmem_cache_open() works but kmalloc does not */ | |
173 | UP, /* Everything works but does not show up in sysfs */ | |
174 | SYSFS /* Sysfs up */ | |
175 | } slab_state = DOWN; | |
176 | ||
177 | /* A list of all slab caches on the system */ | |
178 | static DECLARE_RWSEM(slub_lock); | |
179 | static LIST_HEAD(slab_caches); | |
180 | ||
181 | /* | |
182 | * Tracking user of a slab. | |
183 | */ | |
184 | struct track { | |
185 | unsigned long addr; /* Called from address */ | |
186 | int cpu; /* Was running on cpu */ | |
187 | int pid; /* Pid context */ | |
188 | unsigned long when; /* When did the operation occur */ | |
189 | }; | |
190 | ||
191 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
192 | ||
193 | #ifdef CONFIG_SLUB_DEBUG | |
194 | static int sysfs_slab_add(struct kmem_cache *); | |
195 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
196 | static void sysfs_slab_remove(struct kmem_cache *); | |
197 | ||
198 | #else | |
199 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } | |
200 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) | |
201 | { return 0; } | |
202 | static inline void sysfs_slab_remove(struct kmem_cache *s) | |
203 | { | |
204 | kfree(s); | |
205 | } | |
206 | ||
207 | #endif | |
208 | ||
209 | static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) | |
210 | { | |
211 | #ifdef CONFIG_SLUB_STATS | |
212 | c->stat[si]++; | |
213 | #endif | |
214 | } | |
215 | ||
216 | /******************************************************************** | |
217 | * Core slab cache functions | |
218 | *******************************************************************/ | |
219 | ||
220 | int slab_is_available(void) | |
221 | { | |
222 | return slab_state >= UP; | |
223 | } | |
224 | ||
225 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
226 | { | |
227 | #ifdef CONFIG_NUMA | |
228 | return s->node[node]; | |
229 | #else | |
230 | return &s->local_node; | |
231 | #endif | |
232 | } | |
233 | ||
234 | static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) | |
235 | { | |
236 | #ifdef CONFIG_SMP | |
237 | return s->cpu_slab[cpu]; | |
238 | #else | |
239 | return &s->cpu_slab; | |
240 | #endif | |
241 | } | |
242 | ||
243 | /* Verify that a pointer has an address that is valid within a slab page */ | |
244 | static inline int check_valid_pointer(struct kmem_cache *s, | |
245 | struct page *page, const void *object) | |
246 | { | |
247 | void *base; | |
248 | ||
249 | if (!object) | |
250 | return 1; | |
251 | ||
252 | base = page_address(page); | |
253 | if (object < base || object >= base + page->objects * s->size || | |
254 | (object - base) % s->size) { | |
255 | return 0; | |
256 | } | |
257 | ||
258 | return 1; | |
259 | } | |
260 | ||
261 | /* | |
262 | * Slow version of get and set free pointer. | |
263 | * | |
264 | * This version requires touching the cache lines of kmem_cache which | |
265 | * we avoid to do in the fast alloc free paths. There we obtain the offset | |
266 | * from the page struct. | |
267 | */ | |
268 | static inline void *get_freepointer(struct kmem_cache *s, void *object) | |
269 | { | |
270 | return *(void **)(object + s->offset); | |
271 | } | |
272 | ||
273 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
274 | { | |
275 | *(void **)(object + s->offset) = fp; | |
276 | } | |
277 | ||
278 | /* Loop over all objects in a slab */ | |
279 | #define for_each_object(__p, __s, __addr, __objects) \ | |
280 | for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ | |
281 | __p += (__s)->size) | |
282 | ||
283 | /* Scan freelist */ | |
284 | #define for_each_free_object(__p, __s, __free) \ | |
285 | for (__p = (__free); __p; __p = get_freepointer((__s), __p)) | |
286 | ||
287 | /* Determine object index from a given position */ | |
288 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) | |
289 | { | |
290 | return (p - addr) / s->size; | |
291 | } | |
292 | ||
293 | static inline struct kmem_cache_order_objects oo_make(int order, | |
294 | unsigned long size) | |
295 | { | |
296 | struct kmem_cache_order_objects x = { | |
297 | (order << OO_SHIFT) + (PAGE_SIZE << order) / size | |
298 | }; | |
299 | ||
300 | return x; | |
301 | } | |
302 | ||
303 | static inline int oo_order(struct kmem_cache_order_objects x) | |
304 | { | |
305 | return x.x >> OO_SHIFT; | |
306 | } | |
307 | ||
308 | static inline int oo_objects(struct kmem_cache_order_objects x) | |
309 | { | |
310 | return x.x & OO_MASK; | |
311 | } | |
312 | ||
313 | #ifdef CONFIG_SLUB_DEBUG | |
314 | /* | |
315 | * Debug settings: | |
316 | */ | |
317 | #ifdef CONFIG_SLUB_DEBUG_ON | |
318 | static int slub_debug = DEBUG_DEFAULT_FLAGS; | |
319 | #else | |
320 | static int slub_debug; | |
321 | #endif | |
322 | ||
323 | static char *slub_debug_slabs; | |
324 | ||
325 | /* | |
326 | * Object debugging | |
327 | */ | |
328 | static void print_section(char *text, u8 *addr, unsigned int length) | |
329 | { | |
330 | int i, offset; | |
331 | int newline = 1; | |
332 | char ascii[17]; | |
333 | ||
334 | ascii[16] = 0; | |
335 | ||
336 | for (i = 0; i < length; i++) { | |
337 | if (newline) { | |
338 | printk(KERN_ERR "%8s 0x%p: ", text, addr + i); | |
339 | newline = 0; | |
340 | } | |
341 | printk(KERN_CONT " %02x", addr[i]); | |
342 | offset = i % 16; | |
343 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; | |
344 | if (offset == 15) { | |
345 | printk(KERN_CONT " %s\n", ascii); | |
346 | newline = 1; | |
347 | } | |
348 | } | |
349 | if (!newline) { | |
350 | i %= 16; | |
351 | while (i < 16) { | |
352 | printk(KERN_CONT " "); | |
353 | ascii[i] = ' '; | |
354 | i++; | |
355 | } | |
356 | printk(KERN_CONT " %s\n", ascii); | |
357 | } | |
358 | } | |
359 | ||
360 | static struct track *get_track(struct kmem_cache *s, void *object, | |
361 | enum track_item alloc) | |
362 | { | |
363 | struct track *p; | |
364 | ||
365 | if (s->offset) | |
366 | p = object + s->offset + sizeof(void *); | |
367 | else | |
368 | p = object + s->inuse; | |
369 | ||
370 | return p + alloc; | |
371 | } | |
372 | ||
373 | static void set_track(struct kmem_cache *s, void *object, | |
374 | enum track_item alloc, unsigned long addr) | |
375 | { | |
376 | struct track *p; | |
377 | ||
378 | if (s->offset) | |
379 | p = object + s->offset + sizeof(void *); | |
380 | else | |
381 | p = object + s->inuse; | |
382 | ||
383 | p += alloc; | |
384 | if (addr) { | |
385 | p->addr = addr; | |
386 | p->cpu = smp_processor_id(); | |
387 | p->pid = current->pid; | |
388 | p->when = jiffies; | |
389 | } else | |
390 | memset(p, 0, sizeof(struct track)); | |
391 | } | |
392 | ||
393 | static void init_tracking(struct kmem_cache *s, void *object) | |
394 | { | |
395 | if (!(s->flags & SLAB_STORE_USER)) | |
396 | return; | |
397 | ||
398 | set_track(s, object, TRACK_FREE, 0UL); | |
399 | set_track(s, object, TRACK_ALLOC, 0UL); | |
400 | } | |
401 | ||
402 | static void print_track(const char *s, struct track *t) | |
403 | { | |
404 | if (!t->addr) | |
405 | return; | |
406 | ||
407 | printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", | |
408 | s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); | |
409 | } | |
410 | ||
411 | static void print_tracking(struct kmem_cache *s, void *object) | |
412 | { | |
413 | if (!(s->flags & SLAB_STORE_USER)) | |
414 | return; | |
415 | ||
416 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); | |
417 | print_track("Freed", get_track(s, object, TRACK_FREE)); | |
418 | } | |
419 | ||
420 | static void print_page_info(struct page *page) | |
421 | { | |
422 | printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", | |
423 | page, page->objects, page->inuse, page->freelist, page->flags); | |
424 | ||
425 | } | |
426 | ||
427 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) | |
428 | { | |
429 | va_list args; | |
430 | char buf[100]; | |
431 | ||
432 | va_start(args, fmt); | |
433 | vsnprintf(buf, sizeof(buf), fmt, args); | |
434 | va_end(args); | |
435 | printk(KERN_ERR "========================================" | |
436 | "=====================================\n"); | |
437 | printk(KERN_ERR "BUG %s: %s\n", s->name, buf); | |
438 | printk(KERN_ERR "----------------------------------------" | |
439 | "-------------------------------------\n\n"); | |
440 | } | |
441 | ||
442 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) | |
443 | { | |
444 | va_list args; | |
445 | char buf[100]; | |
446 | ||
447 | va_start(args, fmt); | |
448 | vsnprintf(buf, sizeof(buf), fmt, args); | |
449 | va_end(args); | |
450 | printk(KERN_ERR "FIX %s: %s\n", s->name, buf); | |
451 | } | |
452 | ||
453 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) | |
454 | { | |
455 | unsigned int off; /* Offset of last byte */ | |
456 | u8 *addr = page_address(page); | |
457 | ||
458 | print_tracking(s, p); | |
459 | ||
460 | print_page_info(page); | |
461 | ||
462 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", | |
463 | p, p - addr, get_freepointer(s, p)); | |
464 | ||
465 | if (p > addr + 16) | |
466 | print_section("Bytes b4", p - 16, 16); | |
467 | ||
468 | print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); | |
469 | ||
470 | if (s->flags & SLAB_RED_ZONE) | |
471 | print_section("Redzone", p + s->objsize, | |
472 | s->inuse - s->objsize); | |
473 | ||
474 | if (s->offset) | |
475 | off = s->offset + sizeof(void *); | |
476 | else | |
477 | off = s->inuse; | |
478 | ||
479 | if (s->flags & SLAB_STORE_USER) | |
480 | off += 2 * sizeof(struct track); | |
481 | ||
482 | if (off != s->size) | |
483 | /* Beginning of the filler is the free pointer */ | |
484 | print_section("Padding", p + off, s->size - off); | |
485 | ||
486 | dump_stack(); | |
487 | } | |
488 | ||
489 | static void object_err(struct kmem_cache *s, struct page *page, | |
490 | u8 *object, char *reason) | |
491 | { | |
492 | slab_bug(s, "%s", reason); | |
493 | print_trailer(s, page, object); | |
494 | } | |
495 | ||
496 | static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) | |
497 | { | |
498 | va_list args; | |
499 | char buf[100]; | |
500 | ||
501 | va_start(args, fmt); | |
502 | vsnprintf(buf, sizeof(buf), fmt, args); | |
503 | va_end(args); | |
504 | slab_bug(s, "%s", buf); | |
505 | print_page_info(page); | |
506 | dump_stack(); | |
507 | } | |
508 | ||
509 | static void init_object(struct kmem_cache *s, void *object, int active) | |
510 | { | |
511 | u8 *p = object; | |
512 | ||
513 | if (s->flags & __OBJECT_POISON) { | |
514 | memset(p, POISON_FREE, s->objsize - 1); | |
515 | p[s->objsize - 1] = POISON_END; | |
516 | } | |
517 | ||
518 | if (s->flags & SLAB_RED_ZONE) | |
519 | memset(p + s->objsize, | |
520 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, | |
521 | s->inuse - s->objsize); | |
522 | } | |
523 | ||
524 | static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) | |
525 | { | |
526 | while (bytes) { | |
527 | if (*start != (u8)value) | |
528 | return start; | |
529 | start++; | |
530 | bytes--; | |
531 | } | |
532 | return NULL; | |
533 | } | |
534 | ||
535 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
536 | void *from, void *to) | |
537 | { | |
538 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); | |
539 | memset(from, data, to - from); | |
540 | } | |
541 | ||
542 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, | |
543 | u8 *object, char *what, | |
544 | u8 *start, unsigned int value, unsigned int bytes) | |
545 | { | |
546 | u8 *fault; | |
547 | u8 *end; | |
548 | ||
549 | fault = check_bytes(start, value, bytes); | |
550 | if (!fault) | |
551 | return 1; | |
552 | ||
553 | end = start + bytes; | |
554 | while (end > fault && end[-1] == value) | |
555 | end--; | |
556 | ||
557 | slab_bug(s, "%s overwritten", what); | |
558 | printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", | |
559 | fault, end - 1, fault[0], value); | |
560 | print_trailer(s, page, object); | |
561 | ||
562 | restore_bytes(s, what, value, fault, end); | |
563 | return 0; | |
564 | } | |
565 | ||
566 | /* | |
567 | * Object layout: | |
568 | * | |
569 | * object address | |
570 | * Bytes of the object to be managed. | |
571 | * If the freepointer may overlay the object then the free | |
572 | * pointer is the first word of the object. | |
573 | * | |
574 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is | |
575 | * 0xa5 (POISON_END) | |
576 | * | |
577 | * object + s->objsize | |
578 | * Padding to reach word boundary. This is also used for Redzoning. | |
579 | * Padding is extended by another word if Redzoning is enabled and | |
580 | * objsize == inuse. | |
581 | * | |
582 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with | |
583 | * 0xcc (RED_ACTIVE) for objects in use. | |
584 | * | |
585 | * object + s->inuse | |
586 | * Meta data starts here. | |
587 | * | |
588 | * A. Free pointer (if we cannot overwrite object on free) | |
589 | * B. Tracking data for SLAB_STORE_USER | |
590 | * C. Padding to reach required alignment boundary or at mininum | |
591 | * one word if debugging is on to be able to detect writes | |
592 | * before the word boundary. | |
593 | * | |
594 | * Padding is done using 0x5a (POISON_INUSE) | |
595 | * | |
596 | * object + s->size | |
597 | * Nothing is used beyond s->size. | |
598 | * | |
599 | * If slabcaches are merged then the objsize and inuse boundaries are mostly | |
600 | * ignored. And therefore no slab options that rely on these boundaries | |
601 | * may be used with merged slabcaches. | |
602 | */ | |
603 | ||
604 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) | |
605 | { | |
606 | unsigned long off = s->inuse; /* The end of info */ | |
607 | ||
608 | if (s->offset) | |
609 | /* Freepointer is placed after the object. */ | |
610 | off += sizeof(void *); | |
611 | ||
612 | if (s->flags & SLAB_STORE_USER) | |
613 | /* We also have user information there */ | |
614 | off += 2 * sizeof(struct track); | |
615 | ||
616 | if (s->size == off) | |
617 | return 1; | |
618 | ||
619 | return check_bytes_and_report(s, page, p, "Object padding", | |
620 | p + off, POISON_INUSE, s->size - off); | |
621 | } | |
622 | ||
623 | /* Check the pad bytes at the end of a slab page */ | |
624 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
625 | { | |
626 | u8 *start; | |
627 | u8 *fault; | |
628 | u8 *end; | |
629 | int length; | |
630 | int remainder; | |
631 | ||
632 | if (!(s->flags & SLAB_POISON)) | |
633 | return 1; | |
634 | ||
635 | start = page_address(page); | |
636 | length = (PAGE_SIZE << compound_order(page)); | |
637 | end = start + length; | |
638 | remainder = length % s->size; | |
639 | if (!remainder) | |
640 | return 1; | |
641 | ||
642 | fault = check_bytes(end - remainder, POISON_INUSE, remainder); | |
643 | if (!fault) | |
644 | return 1; | |
645 | while (end > fault && end[-1] == POISON_INUSE) | |
646 | end--; | |
647 | ||
648 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); | |
649 | print_section("Padding", end - remainder, remainder); | |
650 | ||
651 | restore_bytes(s, "slab padding", POISON_INUSE, start, end); | |
652 | return 0; | |
653 | } | |
654 | ||
655 | static int check_object(struct kmem_cache *s, struct page *page, | |
656 | void *object, int active) | |
657 | { | |
658 | u8 *p = object; | |
659 | u8 *endobject = object + s->objsize; | |
660 | ||
661 | if (s->flags & SLAB_RED_ZONE) { | |
662 | unsigned int red = | |
663 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; | |
664 | ||
665 | if (!check_bytes_and_report(s, page, object, "Redzone", | |
666 | endobject, red, s->inuse - s->objsize)) | |
667 | return 0; | |
668 | } else { | |
669 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { | |
670 | check_bytes_and_report(s, page, p, "Alignment padding", | |
671 | endobject, POISON_INUSE, s->inuse - s->objsize); | |
672 | } | |
673 | } | |
674 | ||
675 | if (s->flags & SLAB_POISON) { | |
676 | if (!active && (s->flags & __OBJECT_POISON) && | |
677 | (!check_bytes_and_report(s, page, p, "Poison", p, | |
678 | POISON_FREE, s->objsize - 1) || | |
679 | !check_bytes_and_report(s, page, p, "Poison", | |
680 | p + s->objsize - 1, POISON_END, 1))) | |
681 | return 0; | |
682 | /* | |
683 | * check_pad_bytes cleans up on its own. | |
684 | */ | |
685 | check_pad_bytes(s, page, p); | |
686 | } | |
687 | ||
688 | if (!s->offset && active) | |
689 | /* | |
690 | * Object and freepointer overlap. Cannot check | |
691 | * freepointer while object is allocated. | |
692 | */ | |
693 | return 1; | |
694 | ||
695 | /* Check free pointer validity */ | |
696 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
697 | object_err(s, page, p, "Freepointer corrupt"); | |
698 | /* | |
699 | * No choice but to zap it and thus lose the remainder | |
700 | * of the free objects in this slab. May cause | |
701 | * another error because the object count is now wrong. | |
702 | */ | |
703 | set_freepointer(s, p, NULL); | |
704 | return 0; | |
705 | } | |
706 | return 1; | |
707 | } | |
708 | ||
709 | static int check_slab(struct kmem_cache *s, struct page *page) | |
710 | { | |
711 | int maxobj; | |
712 | ||
713 | VM_BUG_ON(!irqs_disabled()); | |
714 | ||
715 | if (!PageSlab(page)) { | |
716 | slab_err(s, page, "Not a valid slab page"); | |
717 | return 0; | |
718 | } | |
719 | ||
720 | maxobj = (PAGE_SIZE << compound_order(page)) / s->size; | |
721 | if (page->objects > maxobj) { | |
722 | slab_err(s, page, "objects %u > max %u", | |
723 | s->name, page->objects, maxobj); | |
724 | return 0; | |
725 | } | |
726 | if (page->inuse > page->objects) { | |
727 | slab_err(s, page, "inuse %u > max %u", | |
728 | s->name, page->inuse, page->objects); | |
729 | return 0; | |
730 | } | |
731 | /* Slab_pad_check fixes things up after itself */ | |
732 | slab_pad_check(s, page); | |
733 | return 1; | |
734 | } | |
735 | ||
736 | /* | |
737 | * Determine if a certain object on a page is on the freelist. Must hold the | |
738 | * slab lock to guarantee that the chains are in a consistent state. | |
739 | */ | |
740 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
741 | { | |
742 | int nr = 0; | |
743 | void *fp = page->freelist; | |
744 | void *object = NULL; | |
745 | unsigned long max_objects; | |
746 | ||
747 | while (fp && nr <= page->objects) { | |
748 | if (fp == search) | |
749 | return 1; | |
750 | if (!check_valid_pointer(s, page, fp)) { | |
751 | if (object) { | |
752 | object_err(s, page, object, | |
753 | "Freechain corrupt"); | |
754 | set_freepointer(s, object, NULL); | |
755 | break; | |
756 | } else { | |
757 | slab_err(s, page, "Freepointer corrupt"); | |
758 | page->freelist = NULL; | |
759 | page->inuse = page->objects; | |
760 | slab_fix(s, "Freelist cleared"); | |
761 | return 0; | |
762 | } | |
763 | break; | |
764 | } | |
765 | object = fp; | |
766 | fp = get_freepointer(s, object); | |
767 | nr++; | |
768 | } | |
769 | ||
770 | max_objects = (PAGE_SIZE << compound_order(page)) / s->size; | |
771 | if (max_objects > MAX_OBJS_PER_PAGE) | |
772 | max_objects = MAX_OBJS_PER_PAGE; | |
773 | ||
774 | if (page->objects != max_objects) { | |
775 | slab_err(s, page, "Wrong number of objects. Found %d but " | |
776 | "should be %d", page->objects, max_objects); | |
777 | page->objects = max_objects; | |
778 | slab_fix(s, "Number of objects adjusted."); | |
779 | } | |
780 | if (page->inuse != page->objects - nr) { | |
781 | slab_err(s, page, "Wrong object count. Counter is %d but " | |
782 | "counted were %d", page->inuse, page->objects - nr); | |
783 | page->inuse = page->objects - nr; | |
784 | slab_fix(s, "Object count adjusted."); | |
785 | } | |
786 | return search == NULL; | |
787 | } | |
788 | ||
789 | static void trace(struct kmem_cache *s, struct page *page, void *object, | |
790 | int alloc) | |
791 | { | |
792 | if (s->flags & SLAB_TRACE) { | |
793 | printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", | |
794 | s->name, | |
795 | alloc ? "alloc" : "free", | |
796 | object, page->inuse, | |
797 | page->freelist); | |
798 | ||
799 | if (!alloc) | |
800 | print_section("Object", (void *)object, s->objsize); | |
801 | ||
802 | dump_stack(); | |
803 | } | |
804 | } | |
805 | ||
806 | /* | |
807 | * Tracking of fully allocated slabs for debugging purposes. | |
808 | */ | |
809 | static void add_full(struct kmem_cache_node *n, struct page *page) | |
810 | { | |
811 | spin_lock(&n->list_lock); | |
812 | list_add(&page->lru, &n->full); | |
813 | spin_unlock(&n->list_lock); | |
814 | } | |
815 | ||
816 | static void remove_full(struct kmem_cache *s, struct page *page) | |
817 | { | |
818 | struct kmem_cache_node *n; | |
819 | ||
820 | if (!(s->flags & SLAB_STORE_USER)) | |
821 | return; | |
822 | ||
823 | n = get_node(s, page_to_nid(page)); | |
824 | ||
825 | spin_lock(&n->list_lock); | |
826 | list_del(&page->lru); | |
827 | spin_unlock(&n->list_lock); | |
828 | } | |
829 | ||
830 | /* Tracking of the number of slabs for debugging purposes */ | |
831 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
832 | { | |
833 | struct kmem_cache_node *n = get_node(s, node); | |
834 | ||
835 | return atomic_long_read(&n->nr_slabs); | |
836 | } | |
837 | ||
838 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) | |
839 | { | |
840 | struct kmem_cache_node *n = get_node(s, node); | |
841 | ||
842 | /* | |
843 | * May be called early in order to allocate a slab for the | |
844 | * kmem_cache_node structure. Solve the chicken-egg | |
845 | * dilemma by deferring the increment of the count during | |
846 | * bootstrap (see early_kmem_cache_node_alloc). | |
847 | */ | |
848 | if (!NUMA_BUILD || n) { | |
849 | atomic_long_inc(&n->nr_slabs); | |
850 | atomic_long_add(objects, &n->total_objects); | |
851 | } | |
852 | } | |
853 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) | |
854 | { | |
855 | struct kmem_cache_node *n = get_node(s, node); | |
856 | ||
857 | atomic_long_dec(&n->nr_slabs); | |
858 | atomic_long_sub(objects, &n->total_objects); | |
859 | } | |
860 | ||
861 | /* Object debug checks for alloc/free paths */ | |
862 | static void setup_object_debug(struct kmem_cache *s, struct page *page, | |
863 | void *object) | |
864 | { | |
865 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) | |
866 | return; | |
867 | ||
868 | init_object(s, object, 0); | |
869 | init_tracking(s, object); | |
870 | } | |
871 | ||
872 | static int alloc_debug_processing(struct kmem_cache *s, struct page *page, | |
873 | void *object, unsigned long addr) | |
874 | { | |
875 | if (!check_slab(s, page)) | |
876 | goto bad; | |
877 | ||
878 | if (!on_freelist(s, page, object)) { | |
879 | object_err(s, page, object, "Object already allocated"); | |
880 | goto bad; | |
881 | } | |
882 | ||
883 | if (!check_valid_pointer(s, page, object)) { | |
884 | object_err(s, page, object, "Freelist Pointer check fails"); | |
885 | goto bad; | |
886 | } | |
887 | ||
888 | if (!check_object(s, page, object, 0)) | |
889 | goto bad; | |
890 | ||
891 | /* Success perform special debug activities for allocs */ | |
892 | if (s->flags & SLAB_STORE_USER) | |
893 | set_track(s, object, TRACK_ALLOC, addr); | |
894 | trace(s, page, object, 1); | |
895 | init_object(s, object, 1); | |
896 | return 1; | |
897 | ||
898 | bad: | |
899 | if (PageSlab(page)) { | |
900 | /* | |
901 | * If this is a slab page then lets do the best we can | |
902 | * to avoid issues in the future. Marking all objects | |
903 | * as used avoids touching the remaining objects. | |
904 | */ | |
905 | slab_fix(s, "Marking all objects used"); | |
906 | page->inuse = page->objects; | |
907 | page->freelist = NULL; | |
908 | } | |
909 | return 0; | |
910 | } | |
911 | ||
912 | static int free_debug_processing(struct kmem_cache *s, struct page *page, | |
913 | void *object, unsigned long addr) | |
914 | { | |
915 | if (!check_slab(s, page)) | |
916 | goto fail; | |
917 | ||
918 | if (!check_valid_pointer(s, page, object)) { | |
919 | slab_err(s, page, "Invalid object pointer 0x%p", object); | |
920 | goto fail; | |
921 | } | |
922 | ||
923 | if (on_freelist(s, page, object)) { | |
924 | object_err(s, page, object, "Object already free"); | |
925 | goto fail; | |
926 | } | |
927 | ||
928 | if (!check_object(s, page, object, 1)) | |
929 | return 0; | |
930 | ||
931 | if (unlikely(s != page->slab)) { | |
932 | if (!PageSlab(page)) { | |
933 | slab_err(s, page, "Attempt to free object(0x%p) " | |
934 | "outside of slab", object); | |
935 | } else if (!page->slab) { | |
936 | printk(KERN_ERR | |
937 | "SLUB <none>: no slab for object 0x%p.\n", | |
938 | object); | |
939 | dump_stack(); | |
940 | } else | |
941 | object_err(s, page, object, | |
942 | "page slab pointer corrupt."); | |
943 | goto fail; | |
944 | } | |
945 | ||
946 | /* Special debug activities for freeing objects */ | |
947 | if (!PageSlubFrozen(page) && !page->freelist) | |
948 | remove_full(s, page); | |
949 | if (s->flags & SLAB_STORE_USER) | |
950 | set_track(s, object, TRACK_FREE, addr); | |
951 | trace(s, page, object, 0); | |
952 | init_object(s, object, 0); | |
953 | return 1; | |
954 | ||
955 | fail: | |
956 | slab_fix(s, "Object at 0x%p not freed", object); | |
957 | return 0; | |
958 | } | |
959 | ||
960 | static int __init setup_slub_debug(char *str) | |
961 | { | |
962 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
963 | if (*str++ != '=' || !*str) | |
964 | /* | |
965 | * No options specified. Switch on full debugging. | |
966 | */ | |
967 | goto out; | |
968 | ||
969 | if (*str == ',') | |
970 | /* | |
971 | * No options but restriction on slabs. This means full | |
972 | * debugging for slabs matching a pattern. | |
973 | */ | |
974 | goto check_slabs; | |
975 | ||
976 | slub_debug = 0; | |
977 | if (*str == '-') | |
978 | /* | |
979 | * Switch off all debugging measures. | |
980 | */ | |
981 | goto out; | |
982 | ||
983 | /* | |
984 | * Determine which debug features should be switched on | |
985 | */ | |
986 | for (; *str && *str != ','; str++) { | |
987 | switch (tolower(*str)) { | |
988 | case 'f': | |
989 | slub_debug |= SLAB_DEBUG_FREE; | |
990 | break; | |
991 | case 'z': | |
992 | slub_debug |= SLAB_RED_ZONE; | |
993 | break; | |
994 | case 'p': | |
995 | slub_debug |= SLAB_POISON; | |
996 | break; | |
997 | case 'u': | |
998 | slub_debug |= SLAB_STORE_USER; | |
999 | break; | |
1000 | case 't': | |
1001 | slub_debug |= SLAB_TRACE; | |
1002 | break; | |
1003 | default: | |
1004 | printk(KERN_ERR "slub_debug option '%c' " | |
1005 | "unknown. skipped\n", *str); | |
1006 | } | |
1007 | } | |
1008 | ||
1009 | check_slabs: | |
1010 | if (*str == ',') | |
1011 | slub_debug_slabs = str + 1; | |
1012 | out: | |
1013 | return 1; | |
1014 | } | |
1015 | ||
1016 | __setup("slub_debug", setup_slub_debug); | |
1017 | ||
1018 | static unsigned long kmem_cache_flags(unsigned long objsize, | |
1019 | unsigned long flags, const char *name, | |
1020 | void (*ctor)(void *)) | |
1021 | { | |
1022 | /* | |
1023 | * Enable debugging if selected on the kernel commandline. | |
1024 | */ | |
1025 | if (slub_debug && (!slub_debug_slabs || | |
1026 | strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0)) | |
1027 | flags |= slub_debug; | |
1028 | ||
1029 | return flags; | |
1030 | } | |
1031 | #else | |
1032 | static inline void setup_object_debug(struct kmem_cache *s, | |
1033 | struct page *page, void *object) {} | |
1034 | ||
1035 | static inline int alloc_debug_processing(struct kmem_cache *s, | |
1036 | struct page *page, void *object, unsigned long addr) { return 0; } | |
1037 | ||
1038 | static inline int free_debug_processing(struct kmem_cache *s, | |
1039 | struct page *page, void *object, unsigned long addr) { return 0; } | |
1040 | ||
1041 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) | |
1042 | { return 1; } | |
1043 | static inline int check_object(struct kmem_cache *s, struct page *page, | |
1044 | void *object, int active) { return 1; } | |
1045 | static inline void add_full(struct kmem_cache_node *n, struct page *page) {} | |
1046 | static inline unsigned long kmem_cache_flags(unsigned long objsize, | |
1047 | unsigned long flags, const char *name, | |
1048 | void (*ctor)(void *)) | |
1049 | { | |
1050 | return flags; | |
1051 | } | |
1052 | #define slub_debug 0 | |
1053 | ||
1054 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
1055 | { return 0; } | |
1056 | static inline void inc_slabs_node(struct kmem_cache *s, int node, | |
1057 | int objects) {} | |
1058 | static inline void dec_slabs_node(struct kmem_cache *s, int node, | |
1059 | int objects) {} | |
1060 | #endif | |
1061 | ||
1062 | /* | |
1063 | * Slab allocation and freeing | |
1064 | */ | |
1065 | static inline struct page *alloc_slab_page(gfp_t flags, int node, | |
1066 | struct kmem_cache_order_objects oo) | |
1067 | { | |
1068 | int order = oo_order(oo); | |
1069 | ||
1070 | if (node == -1) | |
1071 | return alloc_pages(flags, order); | |
1072 | else | |
1073 | return alloc_pages_node(node, flags, order); | |
1074 | } | |
1075 | ||
1076 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1077 | { | |
1078 | struct page *page; | |
1079 | struct kmem_cache_order_objects oo = s->oo; | |
1080 | ||
1081 | flags |= s->allocflags; | |
1082 | ||
1083 | page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node, | |
1084 | oo); | |
1085 | if (unlikely(!page)) { | |
1086 | oo = s->min; | |
1087 | /* | |
1088 | * Allocation may have failed due to fragmentation. | |
1089 | * Try a lower order alloc if possible | |
1090 | */ | |
1091 | page = alloc_slab_page(flags, node, oo); | |
1092 | if (!page) | |
1093 | return NULL; | |
1094 | ||
1095 | stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK); | |
1096 | } | |
1097 | page->objects = oo_objects(oo); | |
1098 | mod_zone_page_state(page_zone(page), | |
1099 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1100 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1101 | 1 << oo_order(oo)); | |
1102 | ||
1103 | return page; | |
1104 | } | |
1105 | ||
1106 | static void setup_object(struct kmem_cache *s, struct page *page, | |
1107 | void *object) | |
1108 | { | |
1109 | setup_object_debug(s, page, object); | |
1110 | if (unlikely(s->ctor)) | |
1111 | s->ctor(object); | |
1112 | } | |
1113 | ||
1114 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1115 | { | |
1116 | struct page *page; | |
1117 | void *start; | |
1118 | void *last; | |
1119 | void *p; | |
1120 | ||
1121 | BUG_ON(flags & GFP_SLAB_BUG_MASK); | |
1122 | ||
1123 | page = allocate_slab(s, | |
1124 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); | |
1125 | if (!page) | |
1126 | goto out; | |
1127 | ||
1128 | inc_slabs_node(s, page_to_nid(page), page->objects); | |
1129 | page->slab = s; | |
1130 | page->flags |= 1 << PG_slab; | |
1131 | if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | | |
1132 | SLAB_STORE_USER | SLAB_TRACE)) | |
1133 | __SetPageSlubDebug(page); | |
1134 | ||
1135 | start = page_address(page); | |
1136 | ||
1137 | if (unlikely(s->flags & SLAB_POISON)) | |
1138 | memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); | |
1139 | ||
1140 | last = start; | |
1141 | for_each_object(p, s, start, page->objects) { | |
1142 | setup_object(s, page, last); | |
1143 | set_freepointer(s, last, p); | |
1144 | last = p; | |
1145 | } | |
1146 | setup_object(s, page, last); | |
1147 | set_freepointer(s, last, NULL); | |
1148 | ||
1149 | page->freelist = start; | |
1150 | page->inuse = 0; | |
1151 | out: | |
1152 | return page; | |
1153 | } | |
1154 | ||
1155 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
1156 | { | |
1157 | int order = compound_order(page); | |
1158 | int pages = 1 << order; | |
1159 | ||
1160 | if (unlikely(SLABDEBUG && PageSlubDebug(page))) { | |
1161 | void *p; | |
1162 | ||
1163 | slab_pad_check(s, page); | |
1164 | for_each_object(p, s, page_address(page), | |
1165 | page->objects) | |
1166 | check_object(s, page, p, 0); | |
1167 | __ClearPageSlubDebug(page); | |
1168 | } | |
1169 | ||
1170 | mod_zone_page_state(page_zone(page), | |
1171 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1172 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1173 | -pages); | |
1174 | ||
1175 | __ClearPageSlab(page); | |
1176 | reset_page_mapcount(page); | |
1177 | __free_pages(page, order); | |
1178 | } | |
1179 | ||
1180 | static void rcu_free_slab(struct rcu_head *h) | |
1181 | { | |
1182 | struct page *page; | |
1183 | ||
1184 | page = container_of((struct list_head *)h, struct page, lru); | |
1185 | __free_slab(page->slab, page); | |
1186 | } | |
1187 | ||
1188 | static void free_slab(struct kmem_cache *s, struct page *page) | |
1189 | { | |
1190 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
1191 | /* | |
1192 | * RCU free overloads the RCU head over the LRU | |
1193 | */ | |
1194 | struct rcu_head *head = (void *)&page->lru; | |
1195 | ||
1196 | call_rcu(head, rcu_free_slab); | |
1197 | } else | |
1198 | __free_slab(s, page); | |
1199 | } | |
1200 | ||
1201 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
1202 | { | |
1203 | dec_slabs_node(s, page_to_nid(page), page->objects); | |
1204 | free_slab(s, page); | |
1205 | } | |
1206 | ||
1207 | /* | |
1208 | * Per slab locking using the pagelock | |
1209 | */ | |
1210 | static __always_inline void slab_lock(struct page *page) | |
1211 | { | |
1212 | bit_spin_lock(PG_locked, &page->flags); | |
1213 | } | |
1214 | ||
1215 | static __always_inline void slab_unlock(struct page *page) | |
1216 | { | |
1217 | __bit_spin_unlock(PG_locked, &page->flags); | |
1218 | } | |
1219 | ||
1220 | static __always_inline int slab_trylock(struct page *page) | |
1221 | { | |
1222 | int rc = 1; | |
1223 | ||
1224 | rc = bit_spin_trylock(PG_locked, &page->flags); | |
1225 | return rc; | |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * Management of partially allocated slabs | |
1230 | */ | |
1231 | static void add_partial(struct kmem_cache_node *n, | |
1232 | struct page *page, int tail) | |
1233 | { | |
1234 | spin_lock(&n->list_lock); | |
1235 | n->nr_partial++; | |
1236 | if (tail) | |
1237 | list_add_tail(&page->lru, &n->partial); | |
1238 | else | |
1239 | list_add(&page->lru, &n->partial); | |
1240 | spin_unlock(&n->list_lock); | |
1241 | } | |
1242 | ||
1243 | static void remove_partial(struct kmem_cache *s, struct page *page) | |
1244 | { | |
1245 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1246 | ||
1247 | spin_lock(&n->list_lock); | |
1248 | list_del(&page->lru); | |
1249 | n->nr_partial--; | |
1250 | spin_unlock(&n->list_lock); | |
1251 | } | |
1252 | ||
1253 | /* | |
1254 | * Lock slab and remove from the partial list. | |
1255 | * | |
1256 | * Must hold list_lock. | |
1257 | */ | |
1258 | static inline int lock_and_freeze_slab(struct kmem_cache_node *n, | |
1259 | struct page *page) | |
1260 | { | |
1261 | if (slab_trylock(page)) { | |
1262 | list_del(&page->lru); | |
1263 | n->nr_partial--; | |
1264 | __SetPageSlubFrozen(page); | |
1265 | return 1; | |
1266 | } | |
1267 | return 0; | |
1268 | } | |
1269 | ||
1270 | /* | |
1271 | * Try to allocate a partial slab from a specific node. | |
1272 | */ | |
1273 | static struct page *get_partial_node(struct kmem_cache_node *n) | |
1274 | { | |
1275 | struct page *page; | |
1276 | ||
1277 | /* | |
1278 | * Racy check. If we mistakenly see no partial slabs then we | |
1279 | * just allocate an empty slab. If we mistakenly try to get a | |
1280 | * partial slab and there is none available then get_partials() | |
1281 | * will return NULL. | |
1282 | */ | |
1283 | if (!n || !n->nr_partial) | |
1284 | return NULL; | |
1285 | ||
1286 | spin_lock(&n->list_lock); | |
1287 | list_for_each_entry(page, &n->partial, lru) | |
1288 | if (lock_and_freeze_slab(n, page)) | |
1289 | goto out; | |
1290 | page = NULL; | |
1291 | out: | |
1292 | spin_unlock(&n->list_lock); | |
1293 | return page; | |
1294 | } | |
1295 | ||
1296 | /* | |
1297 | * Get a page from somewhere. Search in increasing NUMA distances. | |
1298 | */ | |
1299 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) | |
1300 | { | |
1301 | #ifdef CONFIG_NUMA | |
1302 | struct zonelist *zonelist; | |
1303 | struct zoneref *z; | |
1304 | struct zone *zone; | |
1305 | enum zone_type high_zoneidx = gfp_zone(flags); | |
1306 | struct page *page; | |
1307 | ||
1308 | /* | |
1309 | * The defrag ratio allows a configuration of the tradeoffs between | |
1310 | * inter node defragmentation and node local allocations. A lower | |
1311 | * defrag_ratio increases the tendency to do local allocations | |
1312 | * instead of attempting to obtain partial slabs from other nodes. | |
1313 | * | |
1314 | * If the defrag_ratio is set to 0 then kmalloc() always | |
1315 | * returns node local objects. If the ratio is higher then kmalloc() | |
1316 | * may return off node objects because partial slabs are obtained | |
1317 | * from other nodes and filled up. | |
1318 | * | |
1319 | * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes | |
1320 | * defrag_ratio = 1000) then every (well almost) allocation will | |
1321 | * first attempt to defrag slab caches on other nodes. This means | |
1322 | * scanning over all nodes to look for partial slabs which may be | |
1323 | * expensive if we do it every time we are trying to find a slab | |
1324 | * with available objects. | |
1325 | */ | |
1326 | if (!s->remote_node_defrag_ratio || | |
1327 | get_cycles() % 1024 > s->remote_node_defrag_ratio) | |
1328 | return NULL; | |
1329 | ||
1330 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); | |
1331 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { | |
1332 | struct kmem_cache_node *n; | |
1333 | ||
1334 | n = get_node(s, zone_to_nid(zone)); | |
1335 | ||
1336 | if (n && cpuset_zone_allowed_hardwall(zone, flags) && | |
1337 | n->nr_partial > n->min_partial) { | |
1338 | page = get_partial_node(n); | |
1339 | if (page) | |
1340 | return page; | |
1341 | } | |
1342 | } | |
1343 | #endif | |
1344 | return NULL; | |
1345 | } | |
1346 | ||
1347 | /* | |
1348 | * Get a partial page, lock it and return it. | |
1349 | */ | |
1350 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) | |
1351 | { | |
1352 | struct page *page; | |
1353 | int searchnode = (node == -1) ? numa_node_id() : node; | |
1354 | ||
1355 | page = get_partial_node(get_node(s, searchnode)); | |
1356 | if (page || (flags & __GFP_THISNODE)) | |
1357 | return page; | |
1358 | ||
1359 | return get_any_partial(s, flags); | |
1360 | } | |
1361 | ||
1362 | /* | |
1363 | * Move a page back to the lists. | |
1364 | * | |
1365 | * Must be called with the slab lock held. | |
1366 | * | |
1367 | * On exit the slab lock will have been dropped. | |
1368 | */ | |
1369 | static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) | |
1370 | { | |
1371 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1372 | struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id()); | |
1373 | ||
1374 | __ClearPageSlubFrozen(page); | |
1375 | if (page->inuse) { | |
1376 | ||
1377 | if (page->freelist) { | |
1378 | add_partial(n, page, tail); | |
1379 | stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); | |
1380 | } else { | |
1381 | stat(c, DEACTIVATE_FULL); | |
1382 | if (SLABDEBUG && PageSlubDebug(page) && | |
1383 | (s->flags & SLAB_STORE_USER)) | |
1384 | add_full(n, page); | |
1385 | } | |
1386 | slab_unlock(page); | |
1387 | } else { | |
1388 | stat(c, DEACTIVATE_EMPTY); | |
1389 | if (n->nr_partial < n->min_partial) { | |
1390 | /* | |
1391 | * Adding an empty slab to the partial slabs in order | |
1392 | * to avoid page allocator overhead. This slab needs | |
1393 | * to come after the other slabs with objects in | |
1394 | * so that the others get filled first. That way the | |
1395 | * size of the partial list stays small. | |
1396 | * | |
1397 | * kmem_cache_shrink can reclaim any empty slabs from | |
1398 | * the partial list. | |
1399 | */ | |
1400 | add_partial(n, page, 1); | |
1401 | slab_unlock(page); | |
1402 | } else { | |
1403 | slab_unlock(page); | |
1404 | stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB); | |
1405 | discard_slab(s, page); | |
1406 | } | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | /* | |
1411 | * Remove the cpu slab | |
1412 | */ | |
1413 | static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) | |
1414 | { | |
1415 | struct page *page = c->page; | |
1416 | int tail = 1; | |
1417 | ||
1418 | if (page->freelist) | |
1419 | stat(c, DEACTIVATE_REMOTE_FREES); | |
1420 | /* | |
1421 | * Merge cpu freelist into slab freelist. Typically we get here | |
1422 | * because both freelists are empty. So this is unlikely | |
1423 | * to occur. | |
1424 | */ | |
1425 | while (unlikely(c->freelist)) { | |
1426 | void **object; | |
1427 | ||
1428 | tail = 0; /* Hot objects. Put the slab first */ | |
1429 | ||
1430 | /* Retrieve object from cpu_freelist */ | |
1431 | object = c->freelist; | |
1432 | c->freelist = c->freelist[c->offset]; | |
1433 | ||
1434 | /* And put onto the regular freelist */ | |
1435 | object[c->offset] = page->freelist; | |
1436 | page->freelist = object; | |
1437 | page->inuse--; | |
1438 | } | |
1439 | c->page = NULL; | |
1440 | unfreeze_slab(s, page, tail); | |
1441 | } | |
1442 | ||
1443 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) | |
1444 | { | |
1445 | stat(c, CPUSLAB_FLUSH); | |
1446 | slab_lock(c->page); | |
1447 | deactivate_slab(s, c); | |
1448 | } | |
1449 | ||
1450 | /* | |
1451 | * Flush cpu slab. | |
1452 | * | |
1453 | * Called from IPI handler with interrupts disabled. | |
1454 | */ | |
1455 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) | |
1456 | { | |
1457 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
1458 | ||
1459 | if (likely(c && c->page)) | |
1460 | flush_slab(s, c); | |
1461 | } | |
1462 | ||
1463 | static void flush_cpu_slab(void *d) | |
1464 | { | |
1465 | struct kmem_cache *s = d; | |
1466 | ||
1467 | __flush_cpu_slab(s, smp_processor_id()); | |
1468 | } | |
1469 | ||
1470 | static void flush_all(struct kmem_cache *s) | |
1471 | { | |
1472 | on_each_cpu(flush_cpu_slab, s, 1); | |
1473 | } | |
1474 | ||
1475 | /* | |
1476 | * Check if the objects in a per cpu structure fit numa | |
1477 | * locality expectations. | |
1478 | */ | |
1479 | static inline int node_match(struct kmem_cache_cpu *c, int node) | |
1480 | { | |
1481 | #ifdef CONFIG_NUMA | |
1482 | if (node != -1 && c->node != node) | |
1483 | return 0; | |
1484 | #endif | |
1485 | return 1; | |
1486 | } | |
1487 | ||
1488 | /* | |
1489 | * Slow path. The lockless freelist is empty or we need to perform | |
1490 | * debugging duties. | |
1491 | * | |
1492 | * Interrupts are disabled. | |
1493 | * | |
1494 | * Processing is still very fast if new objects have been freed to the | |
1495 | * regular freelist. In that case we simply take over the regular freelist | |
1496 | * as the lockless freelist and zap the regular freelist. | |
1497 | * | |
1498 | * If that is not working then we fall back to the partial lists. We take the | |
1499 | * first element of the freelist as the object to allocate now and move the | |
1500 | * rest of the freelist to the lockless freelist. | |
1501 | * | |
1502 | * And if we were unable to get a new slab from the partial slab lists then | |
1503 | * we need to allocate a new slab. This is the slowest path since it involves | |
1504 | * a call to the page allocator and the setup of a new slab. | |
1505 | */ | |
1506 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, | |
1507 | unsigned long addr, struct kmem_cache_cpu *c) | |
1508 | { | |
1509 | void **object; | |
1510 | struct page *new; | |
1511 | ||
1512 | /* We handle __GFP_ZERO in the caller */ | |
1513 | gfpflags &= ~__GFP_ZERO; | |
1514 | ||
1515 | if (!c->page) | |
1516 | goto new_slab; | |
1517 | ||
1518 | slab_lock(c->page); | |
1519 | if (unlikely(!node_match(c, node))) | |
1520 | goto another_slab; | |
1521 | ||
1522 | stat(c, ALLOC_REFILL); | |
1523 | ||
1524 | load_freelist: | |
1525 | object = c->page->freelist; | |
1526 | if (unlikely(!object)) | |
1527 | goto another_slab; | |
1528 | if (unlikely(SLABDEBUG && PageSlubDebug(c->page))) | |
1529 | goto debug; | |
1530 | ||
1531 | c->freelist = object[c->offset]; | |
1532 | c->page->inuse = c->page->objects; | |
1533 | c->page->freelist = NULL; | |
1534 | c->node = page_to_nid(c->page); | |
1535 | unlock_out: | |
1536 | slab_unlock(c->page); | |
1537 | stat(c, ALLOC_SLOWPATH); | |
1538 | return object; | |
1539 | ||
1540 | another_slab: | |
1541 | deactivate_slab(s, c); | |
1542 | ||
1543 | new_slab: | |
1544 | new = get_partial(s, gfpflags, node); | |
1545 | if (new) { | |
1546 | c->page = new; | |
1547 | stat(c, ALLOC_FROM_PARTIAL); | |
1548 | goto load_freelist; | |
1549 | } | |
1550 | ||
1551 | if (gfpflags & __GFP_WAIT) | |
1552 | local_irq_enable(); | |
1553 | ||
1554 | new = new_slab(s, gfpflags, node); | |
1555 | ||
1556 | if (gfpflags & __GFP_WAIT) | |
1557 | local_irq_disable(); | |
1558 | ||
1559 | if (new) { | |
1560 | c = get_cpu_slab(s, smp_processor_id()); | |
1561 | stat(c, ALLOC_SLAB); | |
1562 | if (c->page) | |
1563 | flush_slab(s, c); | |
1564 | slab_lock(new); | |
1565 | __SetPageSlubFrozen(new); | |
1566 | c->page = new; | |
1567 | goto load_freelist; | |
1568 | } | |
1569 | return NULL; | |
1570 | debug: | |
1571 | if (!alloc_debug_processing(s, c->page, object, addr)) | |
1572 | goto another_slab; | |
1573 | ||
1574 | c->page->inuse++; | |
1575 | c->page->freelist = object[c->offset]; | |
1576 | c->node = -1; | |
1577 | goto unlock_out; | |
1578 | } | |
1579 | ||
1580 | /* | |
1581 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) | |
1582 | * have the fastpath folded into their functions. So no function call | |
1583 | * overhead for requests that can be satisfied on the fastpath. | |
1584 | * | |
1585 | * The fastpath works by first checking if the lockless freelist can be used. | |
1586 | * If not then __slab_alloc is called for slow processing. | |
1587 | * | |
1588 | * Otherwise we can simply pick the next object from the lockless free list. | |
1589 | */ | |
1590 | static __always_inline void *slab_alloc(struct kmem_cache *s, | |
1591 | gfp_t gfpflags, int node, unsigned long addr) | |
1592 | { | |
1593 | void **object; | |
1594 | struct kmem_cache_cpu *c; | |
1595 | unsigned long flags; | |
1596 | unsigned int objsize; | |
1597 | ||
1598 | might_sleep_if(gfpflags & __GFP_WAIT); | |
1599 | local_irq_save(flags); | |
1600 | c = get_cpu_slab(s, smp_processor_id()); | |
1601 | objsize = c->objsize; | |
1602 | if (unlikely(!c->freelist || !node_match(c, node))) | |
1603 | ||
1604 | object = __slab_alloc(s, gfpflags, node, addr, c); | |
1605 | ||
1606 | else { | |
1607 | object = c->freelist; | |
1608 | c->freelist = object[c->offset]; | |
1609 | stat(c, ALLOC_FASTPATH); | |
1610 | } | |
1611 | local_irq_restore(flags); | |
1612 | ||
1613 | if (unlikely((gfpflags & __GFP_ZERO) && object)) | |
1614 | memset(object, 0, objsize); | |
1615 | ||
1616 | return object; | |
1617 | } | |
1618 | ||
1619 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
1620 | { | |
1621 | return slab_alloc(s, gfpflags, -1, _RET_IP_); | |
1622 | } | |
1623 | EXPORT_SYMBOL(kmem_cache_alloc); | |
1624 | ||
1625 | #ifdef CONFIG_NUMA | |
1626 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
1627 | { | |
1628 | return slab_alloc(s, gfpflags, node, _RET_IP_); | |
1629 | } | |
1630 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
1631 | #endif | |
1632 | ||
1633 | /* | |
1634 | * Slow patch handling. This may still be called frequently since objects | |
1635 | * have a longer lifetime than the cpu slabs in most processing loads. | |
1636 | * | |
1637 | * So we still attempt to reduce cache line usage. Just take the slab | |
1638 | * lock and free the item. If there is no additional partial page | |
1639 | * handling required then we can return immediately. | |
1640 | */ | |
1641 | static void __slab_free(struct kmem_cache *s, struct page *page, | |
1642 | void *x, unsigned long addr, unsigned int offset) | |
1643 | { | |
1644 | void *prior; | |
1645 | void **object = (void *)x; | |
1646 | struct kmem_cache_cpu *c; | |
1647 | ||
1648 | c = get_cpu_slab(s, raw_smp_processor_id()); | |
1649 | stat(c, FREE_SLOWPATH); | |
1650 | slab_lock(page); | |
1651 | ||
1652 | if (unlikely(SLABDEBUG && PageSlubDebug(page))) | |
1653 | goto debug; | |
1654 | ||
1655 | checks_ok: | |
1656 | prior = object[offset] = page->freelist; | |
1657 | page->freelist = object; | |
1658 | page->inuse--; | |
1659 | ||
1660 | if (unlikely(PageSlubFrozen(page))) { | |
1661 | stat(c, FREE_FROZEN); | |
1662 | goto out_unlock; | |
1663 | } | |
1664 | ||
1665 | if (unlikely(!page->inuse)) | |
1666 | goto slab_empty; | |
1667 | ||
1668 | /* | |
1669 | * Objects left in the slab. If it was not on the partial list before | |
1670 | * then add it. | |
1671 | */ | |
1672 | if (unlikely(!prior)) { | |
1673 | add_partial(get_node(s, page_to_nid(page)), page, 1); | |
1674 | stat(c, FREE_ADD_PARTIAL); | |
1675 | } | |
1676 | ||
1677 | out_unlock: | |
1678 | slab_unlock(page); | |
1679 | return; | |
1680 | ||
1681 | slab_empty: | |
1682 | if (prior) { | |
1683 | /* | |
1684 | * Slab still on the partial list. | |
1685 | */ | |
1686 | remove_partial(s, page); | |
1687 | stat(c, FREE_REMOVE_PARTIAL); | |
1688 | } | |
1689 | slab_unlock(page); | |
1690 | stat(c, FREE_SLAB); | |
1691 | discard_slab(s, page); | |
1692 | return; | |
1693 | ||
1694 | debug: | |
1695 | if (!free_debug_processing(s, page, x, addr)) | |
1696 | goto out_unlock; | |
1697 | goto checks_ok; | |
1698 | } | |
1699 | ||
1700 | /* | |
1701 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that | |
1702 | * can perform fastpath freeing without additional function calls. | |
1703 | * | |
1704 | * The fastpath is only possible if we are freeing to the current cpu slab | |
1705 | * of this processor. This typically the case if we have just allocated | |
1706 | * the item before. | |
1707 | * | |
1708 | * If fastpath is not possible then fall back to __slab_free where we deal | |
1709 | * with all sorts of special processing. | |
1710 | */ | |
1711 | static __always_inline void slab_free(struct kmem_cache *s, | |
1712 | struct page *page, void *x, unsigned long addr) | |
1713 | { | |
1714 | void **object = (void *)x; | |
1715 | struct kmem_cache_cpu *c; | |
1716 | unsigned long flags; | |
1717 | ||
1718 | local_irq_save(flags); | |
1719 | c = get_cpu_slab(s, smp_processor_id()); | |
1720 | debug_check_no_locks_freed(object, c->objsize); | |
1721 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) | |
1722 | debug_check_no_obj_freed(object, s->objsize); | |
1723 | if (likely(page == c->page && c->node >= 0)) { | |
1724 | object[c->offset] = c->freelist; | |
1725 | c->freelist = object; | |
1726 | stat(c, FREE_FASTPATH); | |
1727 | } else | |
1728 | __slab_free(s, page, x, addr, c->offset); | |
1729 | ||
1730 | local_irq_restore(flags); | |
1731 | } | |
1732 | ||
1733 | void kmem_cache_free(struct kmem_cache *s, void *x) | |
1734 | { | |
1735 | struct page *page; | |
1736 | ||
1737 | page = virt_to_head_page(x); | |
1738 | ||
1739 | slab_free(s, page, x, _RET_IP_); | |
1740 | } | |
1741 | EXPORT_SYMBOL(kmem_cache_free); | |
1742 | ||
1743 | /* Figure out on which slab page the object resides */ | |
1744 | static struct page *get_object_page(const void *x) | |
1745 | { | |
1746 | struct page *page = virt_to_head_page(x); | |
1747 | ||
1748 | if (!PageSlab(page)) | |
1749 | return NULL; | |
1750 | ||
1751 | return page; | |
1752 | } | |
1753 | ||
1754 | /* | |
1755 | * Object placement in a slab is made very easy because we always start at | |
1756 | * offset 0. If we tune the size of the object to the alignment then we can | |
1757 | * get the required alignment by putting one properly sized object after | |
1758 | * another. | |
1759 | * | |
1760 | * Notice that the allocation order determines the sizes of the per cpu | |
1761 | * caches. Each processor has always one slab available for allocations. | |
1762 | * Increasing the allocation order reduces the number of times that slabs | |
1763 | * must be moved on and off the partial lists and is therefore a factor in | |
1764 | * locking overhead. | |
1765 | */ | |
1766 | ||
1767 | /* | |
1768 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
1769 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
1770 | * and increases the number of allocations possible without having to | |
1771 | * take the list_lock. | |
1772 | */ | |
1773 | static int slub_min_order; | |
1774 | static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; | |
1775 | static int slub_min_objects; | |
1776 | ||
1777 | /* | |
1778 | * Merge control. If this is set then no merging of slab caches will occur. | |
1779 | * (Could be removed. This was introduced to pacify the merge skeptics.) | |
1780 | */ | |
1781 | static int slub_nomerge; | |
1782 | ||
1783 | /* | |
1784 | * Calculate the order of allocation given an slab object size. | |
1785 | * | |
1786 | * The order of allocation has significant impact on performance and other | |
1787 | * system components. Generally order 0 allocations should be preferred since | |
1788 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
1789 | * be problematic to put into order 0 slabs because there may be too much | |
1790 | * unused space left. We go to a higher order if more than 1/16th of the slab | |
1791 | * would be wasted. | |
1792 | * | |
1793 | * In order to reach satisfactory performance we must ensure that a minimum | |
1794 | * number of objects is in one slab. Otherwise we may generate too much | |
1795 | * activity on the partial lists which requires taking the list_lock. This is | |
1796 | * less a concern for large slabs though which are rarely used. | |
1797 | * | |
1798 | * slub_max_order specifies the order where we begin to stop considering the | |
1799 | * number of objects in a slab as critical. If we reach slub_max_order then | |
1800 | * we try to keep the page order as low as possible. So we accept more waste | |
1801 | * of space in favor of a small page order. | |
1802 | * | |
1803 | * Higher order allocations also allow the placement of more objects in a | |
1804 | * slab and thereby reduce object handling overhead. If the user has | |
1805 | * requested a higher mininum order then we start with that one instead of | |
1806 | * the smallest order which will fit the object. | |
1807 | */ | |
1808 | static inline int slab_order(int size, int min_objects, | |
1809 | int max_order, int fract_leftover) | |
1810 | { | |
1811 | int order; | |
1812 | int rem; | |
1813 | int min_order = slub_min_order; | |
1814 | ||
1815 | if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE) | |
1816 | return get_order(size * MAX_OBJS_PER_PAGE) - 1; | |
1817 | ||
1818 | for (order = max(min_order, | |
1819 | fls(min_objects * size - 1) - PAGE_SHIFT); | |
1820 | order <= max_order; order++) { | |
1821 | ||
1822 | unsigned long slab_size = PAGE_SIZE << order; | |
1823 | ||
1824 | if (slab_size < min_objects * size) | |
1825 | continue; | |
1826 | ||
1827 | rem = slab_size % size; | |
1828 | ||
1829 | if (rem <= slab_size / fract_leftover) | |
1830 | break; | |
1831 | ||
1832 | } | |
1833 | ||
1834 | return order; | |
1835 | } | |
1836 | ||
1837 | static inline int calculate_order(int size) | |
1838 | { | |
1839 | int order; | |
1840 | int min_objects; | |
1841 | int fraction; | |
1842 | ||
1843 | /* | |
1844 | * Attempt to find best configuration for a slab. This | |
1845 | * works by first attempting to generate a layout with | |
1846 | * the best configuration and backing off gradually. | |
1847 | * | |
1848 | * First we reduce the acceptable waste in a slab. Then | |
1849 | * we reduce the minimum objects required in a slab. | |
1850 | */ | |
1851 | min_objects = slub_min_objects; | |
1852 | if (!min_objects) | |
1853 | min_objects = 4 * (fls(nr_cpu_ids) + 1); | |
1854 | while (min_objects > 1) { | |
1855 | fraction = 16; | |
1856 | while (fraction >= 4) { | |
1857 | order = slab_order(size, min_objects, | |
1858 | slub_max_order, fraction); | |
1859 | if (order <= slub_max_order) | |
1860 | return order; | |
1861 | fraction /= 2; | |
1862 | } | |
1863 | min_objects /= 2; | |
1864 | } | |
1865 | ||
1866 | /* | |
1867 | * We were unable to place multiple objects in a slab. Now | |
1868 | * lets see if we can place a single object there. | |
1869 | */ | |
1870 | order = slab_order(size, 1, slub_max_order, 1); | |
1871 | if (order <= slub_max_order) | |
1872 | return order; | |
1873 | ||
1874 | /* | |
1875 | * Doh this slab cannot be placed using slub_max_order. | |
1876 | */ | |
1877 | order = slab_order(size, 1, MAX_ORDER, 1); | |
1878 | if (order <= MAX_ORDER) | |
1879 | return order; | |
1880 | return -ENOSYS; | |
1881 | } | |
1882 | ||
1883 | /* | |
1884 | * Figure out what the alignment of the objects will be. | |
1885 | */ | |
1886 | static unsigned long calculate_alignment(unsigned long flags, | |
1887 | unsigned long align, unsigned long size) | |
1888 | { | |
1889 | /* | |
1890 | * If the user wants hardware cache aligned objects then follow that | |
1891 | * suggestion if the object is sufficiently large. | |
1892 | * | |
1893 | * The hardware cache alignment cannot override the specified | |
1894 | * alignment though. If that is greater then use it. | |
1895 | */ | |
1896 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1897 | unsigned long ralign = cache_line_size(); | |
1898 | while (size <= ralign / 2) | |
1899 | ralign /= 2; | |
1900 | align = max(align, ralign); | |
1901 | } | |
1902 | ||
1903 | if (align < ARCH_SLAB_MINALIGN) | |
1904 | align = ARCH_SLAB_MINALIGN; | |
1905 | ||
1906 | return ALIGN(align, sizeof(void *)); | |
1907 | } | |
1908 | ||
1909 | static void init_kmem_cache_cpu(struct kmem_cache *s, | |
1910 | struct kmem_cache_cpu *c) | |
1911 | { | |
1912 | c->page = NULL; | |
1913 | c->freelist = NULL; | |
1914 | c->node = 0; | |
1915 | c->offset = s->offset / sizeof(void *); | |
1916 | c->objsize = s->objsize; | |
1917 | #ifdef CONFIG_SLUB_STATS | |
1918 | memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); | |
1919 | #endif | |
1920 | } | |
1921 | ||
1922 | static void | |
1923 | init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) | |
1924 | { | |
1925 | n->nr_partial = 0; | |
1926 | ||
1927 | /* | |
1928 | * The larger the object size is, the more pages we want on the partial | |
1929 | * list to avoid pounding the page allocator excessively. | |
1930 | */ | |
1931 | n->min_partial = ilog2(s->size); | |
1932 | if (n->min_partial < MIN_PARTIAL) | |
1933 | n->min_partial = MIN_PARTIAL; | |
1934 | else if (n->min_partial > MAX_PARTIAL) | |
1935 | n->min_partial = MAX_PARTIAL; | |
1936 | ||
1937 | spin_lock_init(&n->list_lock); | |
1938 | INIT_LIST_HEAD(&n->partial); | |
1939 | #ifdef CONFIG_SLUB_DEBUG | |
1940 | atomic_long_set(&n->nr_slabs, 0); | |
1941 | atomic_long_set(&n->total_objects, 0); | |
1942 | INIT_LIST_HEAD(&n->full); | |
1943 | #endif | |
1944 | } | |
1945 | ||
1946 | #ifdef CONFIG_SMP | |
1947 | /* | |
1948 | * Per cpu array for per cpu structures. | |
1949 | * | |
1950 | * The per cpu array places all kmem_cache_cpu structures from one processor | |
1951 | * close together meaning that it becomes possible that multiple per cpu | |
1952 | * structures are contained in one cacheline. This may be particularly | |
1953 | * beneficial for the kmalloc caches. | |
1954 | * | |
1955 | * A desktop system typically has around 60-80 slabs. With 100 here we are | |
1956 | * likely able to get per cpu structures for all caches from the array defined | |
1957 | * here. We must be able to cover all kmalloc caches during bootstrap. | |
1958 | * | |
1959 | * If the per cpu array is exhausted then fall back to kmalloc | |
1960 | * of individual cachelines. No sharing is possible then. | |
1961 | */ | |
1962 | #define NR_KMEM_CACHE_CPU 100 | |
1963 | ||
1964 | static DEFINE_PER_CPU(struct kmem_cache_cpu, | |
1965 | kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; | |
1966 | ||
1967 | static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); | |
1968 | static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; | |
1969 | ||
1970 | static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, | |
1971 | int cpu, gfp_t flags) | |
1972 | { | |
1973 | struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); | |
1974 | ||
1975 | if (c) | |
1976 | per_cpu(kmem_cache_cpu_free, cpu) = | |
1977 | (void *)c->freelist; | |
1978 | else { | |
1979 | /* Table overflow: So allocate ourselves */ | |
1980 | c = kmalloc_node( | |
1981 | ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), | |
1982 | flags, cpu_to_node(cpu)); | |
1983 | if (!c) | |
1984 | return NULL; | |
1985 | } | |
1986 | ||
1987 | init_kmem_cache_cpu(s, c); | |
1988 | return c; | |
1989 | } | |
1990 | ||
1991 | static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) | |
1992 | { | |
1993 | if (c < per_cpu(kmem_cache_cpu, cpu) || | |
1994 | c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { | |
1995 | kfree(c); | |
1996 | return; | |
1997 | } | |
1998 | c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); | |
1999 | per_cpu(kmem_cache_cpu_free, cpu) = c; | |
2000 | } | |
2001 | ||
2002 | static void free_kmem_cache_cpus(struct kmem_cache *s) | |
2003 | { | |
2004 | int cpu; | |
2005 | ||
2006 | for_each_online_cpu(cpu) { | |
2007 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
2008 | ||
2009 | if (c) { | |
2010 | s->cpu_slab[cpu] = NULL; | |
2011 | free_kmem_cache_cpu(c, cpu); | |
2012 | } | |
2013 | } | |
2014 | } | |
2015 | ||
2016 | static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) | |
2017 | { | |
2018 | int cpu; | |
2019 | ||
2020 | for_each_online_cpu(cpu) { | |
2021 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
2022 | ||
2023 | if (c) | |
2024 | continue; | |
2025 | ||
2026 | c = alloc_kmem_cache_cpu(s, cpu, flags); | |
2027 | if (!c) { | |
2028 | free_kmem_cache_cpus(s); | |
2029 | return 0; | |
2030 | } | |
2031 | s->cpu_slab[cpu] = c; | |
2032 | } | |
2033 | return 1; | |
2034 | } | |
2035 | ||
2036 | /* | |
2037 | * Initialize the per cpu array. | |
2038 | */ | |
2039 | static void init_alloc_cpu_cpu(int cpu) | |
2040 | { | |
2041 | int i; | |
2042 | ||
2043 | if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) | |
2044 | return; | |
2045 | ||
2046 | for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) | |
2047 | free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); | |
2048 | ||
2049 | cpu_set(cpu, kmem_cach_cpu_free_init_once); | |
2050 | } | |
2051 | ||
2052 | static void __init init_alloc_cpu(void) | |
2053 | { | |
2054 | int cpu; | |
2055 | ||
2056 | for_each_online_cpu(cpu) | |
2057 | init_alloc_cpu_cpu(cpu); | |
2058 | } | |
2059 | ||
2060 | #else | |
2061 | static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} | |
2062 | static inline void init_alloc_cpu(void) {} | |
2063 | ||
2064 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) | |
2065 | { | |
2066 | init_kmem_cache_cpu(s, &s->cpu_slab); | |
2067 | return 1; | |
2068 | } | |
2069 | #endif | |
2070 | ||
2071 | #ifdef CONFIG_NUMA | |
2072 | /* | |
2073 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
2074 | * slab on the node for this slabcache. There are no concurrent accesses | |
2075 | * possible. | |
2076 | * | |
2077 | * Note that this function only works on the kmalloc_node_cache | |
2078 | * when allocating for the kmalloc_node_cache. This is used for bootstrapping | |
2079 | * memory on a fresh node that has no slab structures yet. | |
2080 | */ | |
2081 | static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node) | |
2082 | { | |
2083 | struct page *page; | |
2084 | struct kmem_cache_node *n; | |
2085 | unsigned long flags; | |
2086 | ||
2087 | BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); | |
2088 | ||
2089 | page = new_slab(kmalloc_caches, gfpflags, node); | |
2090 | ||
2091 | BUG_ON(!page); | |
2092 | if (page_to_nid(page) != node) { | |
2093 | printk(KERN_ERR "SLUB: Unable to allocate memory from " | |
2094 | "node %d\n", node); | |
2095 | printk(KERN_ERR "SLUB: Allocating a useless per node structure " | |
2096 | "in order to be able to continue\n"); | |
2097 | } | |
2098 | ||
2099 | n = page->freelist; | |
2100 | BUG_ON(!n); | |
2101 | page->freelist = get_freepointer(kmalloc_caches, n); | |
2102 | page->inuse++; | |
2103 | kmalloc_caches->node[node] = n; | |
2104 | #ifdef CONFIG_SLUB_DEBUG | |
2105 | init_object(kmalloc_caches, n, 1); | |
2106 | init_tracking(kmalloc_caches, n); | |
2107 | #endif | |
2108 | init_kmem_cache_node(n, kmalloc_caches); | |
2109 | inc_slabs_node(kmalloc_caches, node, page->objects); | |
2110 | ||
2111 | /* | |
2112 | * lockdep requires consistent irq usage for each lock | |
2113 | * so even though there cannot be a race this early in | |
2114 | * the boot sequence, we still disable irqs. | |
2115 | */ | |
2116 | local_irq_save(flags); | |
2117 | add_partial(n, page, 0); | |
2118 | local_irq_restore(flags); | |
2119 | } | |
2120 | ||
2121 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2122 | { | |
2123 | int node; | |
2124 | ||
2125 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2126 | struct kmem_cache_node *n = s->node[node]; | |
2127 | if (n && n != &s->local_node) | |
2128 | kmem_cache_free(kmalloc_caches, n); | |
2129 | s->node[node] = NULL; | |
2130 | } | |
2131 | } | |
2132 | ||
2133 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
2134 | { | |
2135 | int node; | |
2136 | int local_node; | |
2137 | ||
2138 | if (slab_state >= UP) | |
2139 | local_node = page_to_nid(virt_to_page(s)); | |
2140 | else | |
2141 | local_node = 0; | |
2142 | ||
2143 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2144 | struct kmem_cache_node *n; | |
2145 | ||
2146 | if (local_node == node) | |
2147 | n = &s->local_node; | |
2148 | else { | |
2149 | if (slab_state == DOWN) { | |
2150 | early_kmem_cache_node_alloc(gfpflags, node); | |
2151 | continue; | |
2152 | } | |
2153 | n = kmem_cache_alloc_node(kmalloc_caches, | |
2154 | gfpflags, node); | |
2155 | ||
2156 | if (!n) { | |
2157 | free_kmem_cache_nodes(s); | |
2158 | return 0; | |
2159 | } | |
2160 | ||
2161 | } | |
2162 | s->node[node] = n; | |
2163 | init_kmem_cache_node(n, s); | |
2164 | } | |
2165 | return 1; | |
2166 | } | |
2167 | #else | |
2168 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2169 | { | |
2170 | } | |
2171 | ||
2172 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
2173 | { | |
2174 | init_kmem_cache_node(&s->local_node, s); | |
2175 | return 1; | |
2176 | } | |
2177 | #endif | |
2178 | ||
2179 | /* | |
2180 | * calculate_sizes() determines the order and the distribution of data within | |
2181 | * a slab object. | |
2182 | */ | |
2183 | static int calculate_sizes(struct kmem_cache *s, int forced_order) | |
2184 | { | |
2185 | unsigned long flags = s->flags; | |
2186 | unsigned long size = s->objsize; | |
2187 | unsigned long align = s->align; | |
2188 | int order; | |
2189 | ||
2190 | /* | |
2191 | * Round up object size to the next word boundary. We can only | |
2192 | * place the free pointer at word boundaries and this determines | |
2193 | * the possible location of the free pointer. | |
2194 | */ | |
2195 | size = ALIGN(size, sizeof(void *)); | |
2196 | ||
2197 | #ifdef CONFIG_SLUB_DEBUG | |
2198 | /* | |
2199 | * Determine if we can poison the object itself. If the user of | |
2200 | * the slab may touch the object after free or before allocation | |
2201 | * then we should never poison the object itself. | |
2202 | */ | |
2203 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
2204 | !s->ctor) | |
2205 | s->flags |= __OBJECT_POISON; | |
2206 | else | |
2207 | s->flags &= ~__OBJECT_POISON; | |
2208 | ||
2209 | ||
2210 | /* | |
2211 | * If we are Redzoning then check if there is some space between the | |
2212 | * end of the object and the free pointer. If not then add an | |
2213 | * additional word to have some bytes to store Redzone information. | |
2214 | */ | |
2215 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
2216 | size += sizeof(void *); | |
2217 | #endif | |
2218 | ||
2219 | /* | |
2220 | * With that we have determined the number of bytes in actual use | |
2221 | * by the object. This is the potential offset to the free pointer. | |
2222 | */ | |
2223 | s->inuse = size; | |
2224 | ||
2225 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
2226 | s->ctor)) { | |
2227 | /* | |
2228 | * Relocate free pointer after the object if it is not | |
2229 | * permitted to overwrite the first word of the object on | |
2230 | * kmem_cache_free. | |
2231 | * | |
2232 | * This is the case if we do RCU, have a constructor or | |
2233 | * destructor or are poisoning the objects. | |
2234 | */ | |
2235 | s->offset = size; | |
2236 | size += sizeof(void *); | |
2237 | } | |
2238 | ||
2239 | #ifdef CONFIG_SLUB_DEBUG | |
2240 | if (flags & SLAB_STORE_USER) | |
2241 | /* | |
2242 | * Need to store information about allocs and frees after | |
2243 | * the object. | |
2244 | */ | |
2245 | size += 2 * sizeof(struct track); | |
2246 | ||
2247 | if (flags & SLAB_RED_ZONE) | |
2248 | /* | |
2249 | * Add some empty padding so that we can catch | |
2250 | * overwrites from earlier objects rather than let | |
2251 | * tracking information or the free pointer be | |
2252 | * corrupted if an user writes before the start | |
2253 | * of the object. | |
2254 | */ | |
2255 | size += sizeof(void *); | |
2256 | #endif | |
2257 | ||
2258 | /* | |
2259 | * Determine the alignment based on various parameters that the | |
2260 | * user specified and the dynamic determination of cache line size | |
2261 | * on bootup. | |
2262 | */ | |
2263 | align = calculate_alignment(flags, align, s->objsize); | |
2264 | ||
2265 | /* | |
2266 | * SLUB stores one object immediately after another beginning from | |
2267 | * offset 0. In order to align the objects we have to simply size | |
2268 | * each object to conform to the alignment. | |
2269 | */ | |
2270 | size = ALIGN(size, align); | |
2271 | s->size = size; | |
2272 | if (forced_order >= 0) | |
2273 | order = forced_order; | |
2274 | else | |
2275 | order = calculate_order(size); | |
2276 | ||
2277 | if (order < 0) | |
2278 | return 0; | |
2279 | ||
2280 | s->allocflags = 0; | |
2281 | if (order) | |
2282 | s->allocflags |= __GFP_COMP; | |
2283 | ||
2284 | if (s->flags & SLAB_CACHE_DMA) | |
2285 | s->allocflags |= SLUB_DMA; | |
2286 | ||
2287 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
2288 | s->allocflags |= __GFP_RECLAIMABLE; | |
2289 | ||
2290 | /* | |
2291 | * Determine the number of objects per slab | |
2292 | */ | |
2293 | s->oo = oo_make(order, size); | |
2294 | s->min = oo_make(get_order(size), size); | |
2295 | if (oo_objects(s->oo) > oo_objects(s->max)) | |
2296 | s->max = s->oo; | |
2297 | ||
2298 | return !!oo_objects(s->oo); | |
2299 | ||
2300 | } | |
2301 | ||
2302 | static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, | |
2303 | const char *name, size_t size, | |
2304 | size_t align, unsigned long flags, | |
2305 | void (*ctor)(void *)) | |
2306 | { | |
2307 | memset(s, 0, kmem_size); | |
2308 | s->name = name; | |
2309 | s->ctor = ctor; | |
2310 | s->objsize = size; | |
2311 | s->align = align; | |
2312 | s->flags = kmem_cache_flags(size, flags, name, ctor); | |
2313 | ||
2314 | if (!calculate_sizes(s, -1)) | |
2315 | goto error; | |
2316 | ||
2317 | s->refcount = 1; | |
2318 | #ifdef CONFIG_NUMA | |
2319 | s->remote_node_defrag_ratio = 1000; | |
2320 | #endif | |
2321 | if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) | |
2322 | goto error; | |
2323 | ||
2324 | if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) | |
2325 | return 1; | |
2326 | free_kmem_cache_nodes(s); | |
2327 | error: | |
2328 | if (flags & SLAB_PANIC) | |
2329 | panic("Cannot create slab %s size=%lu realsize=%u " | |
2330 | "order=%u offset=%u flags=%lx\n", | |
2331 | s->name, (unsigned long)size, s->size, oo_order(s->oo), | |
2332 | s->offset, flags); | |
2333 | return 0; | |
2334 | } | |
2335 | ||
2336 | /* | |
2337 | * Check if a given pointer is valid | |
2338 | */ | |
2339 | int kmem_ptr_validate(struct kmem_cache *s, const void *object) | |
2340 | { | |
2341 | struct page *page; | |
2342 | ||
2343 | page = get_object_page(object); | |
2344 | ||
2345 | if (!page || s != page->slab) | |
2346 | /* No slab or wrong slab */ | |
2347 | return 0; | |
2348 | ||
2349 | if (!check_valid_pointer(s, page, object)) | |
2350 | return 0; | |
2351 | ||
2352 | /* | |
2353 | * We could also check if the object is on the slabs freelist. | |
2354 | * But this would be too expensive and it seems that the main | |
2355 | * purpose of kmem_ptr_valid() is to check if the object belongs | |
2356 | * to a certain slab. | |
2357 | */ | |
2358 | return 1; | |
2359 | } | |
2360 | EXPORT_SYMBOL(kmem_ptr_validate); | |
2361 | ||
2362 | /* | |
2363 | * Determine the size of a slab object | |
2364 | */ | |
2365 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
2366 | { | |
2367 | return s->objsize; | |
2368 | } | |
2369 | EXPORT_SYMBOL(kmem_cache_size); | |
2370 | ||
2371 | const char *kmem_cache_name(struct kmem_cache *s) | |
2372 | { | |
2373 | return s->name; | |
2374 | } | |
2375 | EXPORT_SYMBOL(kmem_cache_name); | |
2376 | ||
2377 | static void list_slab_objects(struct kmem_cache *s, struct page *page, | |
2378 | const char *text) | |
2379 | { | |
2380 | #ifdef CONFIG_SLUB_DEBUG | |
2381 | void *addr = page_address(page); | |
2382 | void *p; | |
2383 | DECLARE_BITMAP(map, page->objects); | |
2384 | ||
2385 | bitmap_zero(map, page->objects); | |
2386 | slab_err(s, page, "%s", text); | |
2387 | slab_lock(page); | |
2388 | for_each_free_object(p, s, page->freelist) | |
2389 | set_bit(slab_index(p, s, addr), map); | |
2390 | ||
2391 | for_each_object(p, s, addr, page->objects) { | |
2392 | ||
2393 | if (!test_bit(slab_index(p, s, addr), map)) { | |
2394 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", | |
2395 | p, p - addr); | |
2396 | print_tracking(s, p); | |
2397 | } | |
2398 | } | |
2399 | slab_unlock(page); | |
2400 | #endif | |
2401 | } | |
2402 | ||
2403 | /* | |
2404 | * Attempt to free all partial slabs on a node. | |
2405 | */ | |
2406 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) | |
2407 | { | |
2408 | unsigned long flags; | |
2409 | struct page *page, *h; | |
2410 | ||
2411 | spin_lock_irqsave(&n->list_lock, flags); | |
2412 | list_for_each_entry_safe(page, h, &n->partial, lru) { | |
2413 | if (!page->inuse) { | |
2414 | list_del(&page->lru); | |
2415 | discard_slab(s, page); | |
2416 | n->nr_partial--; | |
2417 | } else { | |
2418 | list_slab_objects(s, page, | |
2419 | "Objects remaining on kmem_cache_close()"); | |
2420 | } | |
2421 | } | |
2422 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2423 | } | |
2424 | ||
2425 | /* | |
2426 | * Release all resources used by a slab cache. | |
2427 | */ | |
2428 | static inline int kmem_cache_close(struct kmem_cache *s) | |
2429 | { | |
2430 | int node; | |
2431 | ||
2432 | flush_all(s); | |
2433 | ||
2434 | /* Attempt to free all objects */ | |
2435 | free_kmem_cache_cpus(s); | |
2436 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2437 | struct kmem_cache_node *n = get_node(s, node); | |
2438 | ||
2439 | free_partial(s, n); | |
2440 | if (n->nr_partial || slabs_node(s, node)) | |
2441 | return 1; | |
2442 | } | |
2443 | free_kmem_cache_nodes(s); | |
2444 | return 0; | |
2445 | } | |
2446 | ||
2447 | /* | |
2448 | * Close a cache and release the kmem_cache structure | |
2449 | * (must be used for caches created using kmem_cache_create) | |
2450 | */ | |
2451 | void kmem_cache_destroy(struct kmem_cache *s) | |
2452 | { | |
2453 | down_write(&slub_lock); | |
2454 | s->refcount--; | |
2455 | if (!s->refcount) { | |
2456 | list_del(&s->list); | |
2457 | up_write(&slub_lock); | |
2458 | if (kmem_cache_close(s)) { | |
2459 | printk(KERN_ERR "SLUB %s: %s called for cache that " | |
2460 | "still has objects.\n", s->name, __func__); | |
2461 | dump_stack(); | |
2462 | } | |
2463 | sysfs_slab_remove(s); | |
2464 | } else | |
2465 | up_write(&slub_lock); | |
2466 | } | |
2467 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2468 | ||
2469 | /******************************************************************** | |
2470 | * Kmalloc subsystem | |
2471 | *******************************************************************/ | |
2472 | ||
2473 | struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; | |
2474 | EXPORT_SYMBOL(kmalloc_caches); | |
2475 | ||
2476 | static int __init setup_slub_min_order(char *str) | |
2477 | { | |
2478 | get_option(&str, &slub_min_order); | |
2479 | ||
2480 | return 1; | |
2481 | } | |
2482 | ||
2483 | __setup("slub_min_order=", setup_slub_min_order); | |
2484 | ||
2485 | static int __init setup_slub_max_order(char *str) | |
2486 | { | |
2487 | get_option(&str, &slub_max_order); | |
2488 | ||
2489 | return 1; | |
2490 | } | |
2491 | ||
2492 | __setup("slub_max_order=", setup_slub_max_order); | |
2493 | ||
2494 | static int __init setup_slub_min_objects(char *str) | |
2495 | { | |
2496 | get_option(&str, &slub_min_objects); | |
2497 | ||
2498 | return 1; | |
2499 | } | |
2500 | ||
2501 | __setup("slub_min_objects=", setup_slub_min_objects); | |
2502 | ||
2503 | static int __init setup_slub_nomerge(char *str) | |
2504 | { | |
2505 | slub_nomerge = 1; | |
2506 | return 1; | |
2507 | } | |
2508 | ||
2509 | __setup("slub_nomerge", setup_slub_nomerge); | |
2510 | ||
2511 | static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, | |
2512 | const char *name, int size, gfp_t gfp_flags) | |
2513 | { | |
2514 | unsigned int flags = 0; | |
2515 | ||
2516 | if (gfp_flags & SLUB_DMA) | |
2517 | flags = SLAB_CACHE_DMA; | |
2518 | ||
2519 | down_write(&slub_lock); | |
2520 | if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, | |
2521 | flags, NULL)) | |
2522 | goto panic; | |
2523 | ||
2524 | list_add(&s->list, &slab_caches); | |
2525 | up_write(&slub_lock); | |
2526 | if (sysfs_slab_add(s)) | |
2527 | goto panic; | |
2528 | return s; | |
2529 | ||
2530 | panic: | |
2531 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
2532 | } | |
2533 | ||
2534 | #ifdef CONFIG_ZONE_DMA | |
2535 | static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; | |
2536 | ||
2537 | static void sysfs_add_func(struct work_struct *w) | |
2538 | { | |
2539 | struct kmem_cache *s; | |
2540 | ||
2541 | down_write(&slub_lock); | |
2542 | list_for_each_entry(s, &slab_caches, list) { | |
2543 | if (s->flags & __SYSFS_ADD_DEFERRED) { | |
2544 | s->flags &= ~__SYSFS_ADD_DEFERRED; | |
2545 | sysfs_slab_add(s); | |
2546 | } | |
2547 | } | |
2548 | up_write(&slub_lock); | |
2549 | } | |
2550 | ||
2551 | static DECLARE_WORK(sysfs_add_work, sysfs_add_func); | |
2552 | ||
2553 | static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) | |
2554 | { | |
2555 | struct kmem_cache *s; | |
2556 | char *text; | |
2557 | size_t realsize; | |
2558 | ||
2559 | s = kmalloc_caches_dma[index]; | |
2560 | if (s) | |
2561 | return s; | |
2562 | ||
2563 | /* Dynamically create dma cache */ | |
2564 | if (flags & __GFP_WAIT) | |
2565 | down_write(&slub_lock); | |
2566 | else { | |
2567 | if (!down_write_trylock(&slub_lock)) | |
2568 | goto out; | |
2569 | } | |
2570 | ||
2571 | if (kmalloc_caches_dma[index]) | |
2572 | goto unlock_out; | |
2573 | ||
2574 | realsize = kmalloc_caches[index].objsize; | |
2575 | text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", | |
2576 | (unsigned int)realsize); | |
2577 | s = kmalloc(kmem_size, flags & ~SLUB_DMA); | |
2578 | ||
2579 | if (!s || !text || !kmem_cache_open(s, flags, text, | |
2580 | realsize, ARCH_KMALLOC_MINALIGN, | |
2581 | SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { | |
2582 | kfree(s); | |
2583 | kfree(text); | |
2584 | goto unlock_out; | |
2585 | } | |
2586 | ||
2587 | list_add(&s->list, &slab_caches); | |
2588 | kmalloc_caches_dma[index] = s; | |
2589 | ||
2590 | schedule_work(&sysfs_add_work); | |
2591 | ||
2592 | unlock_out: | |
2593 | up_write(&slub_lock); | |
2594 | out: | |
2595 | return kmalloc_caches_dma[index]; | |
2596 | } | |
2597 | #endif | |
2598 | ||
2599 | /* | |
2600 | * Conversion table for small slabs sizes / 8 to the index in the | |
2601 | * kmalloc array. This is necessary for slabs < 192 since we have non power | |
2602 | * of two cache sizes there. The size of larger slabs can be determined using | |
2603 | * fls. | |
2604 | */ | |
2605 | static s8 size_index[24] = { | |
2606 | 3, /* 8 */ | |
2607 | 4, /* 16 */ | |
2608 | 5, /* 24 */ | |
2609 | 5, /* 32 */ | |
2610 | 6, /* 40 */ | |
2611 | 6, /* 48 */ | |
2612 | 6, /* 56 */ | |
2613 | 6, /* 64 */ | |
2614 | 1, /* 72 */ | |
2615 | 1, /* 80 */ | |
2616 | 1, /* 88 */ | |
2617 | 1, /* 96 */ | |
2618 | 7, /* 104 */ | |
2619 | 7, /* 112 */ | |
2620 | 7, /* 120 */ | |
2621 | 7, /* 128 */ | |
2622 | 2, /* 136 */ | |
2623 | 2, /* 144 */ | |
2624 | 2, /* 152 */ | |
2625 | 2, /* 160 */ | |
2626 | 2, /* 168 */ | |
2627 | 2, /* 176 */ | |
2628 | 2, /* 184 */ | |
2629 | 2 /* 192 */ | |
2630 | }; | |
2631 | ||
2632 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) | |
2633 | { | |
2634 | int index; | |
2635 | ||
2636 | if (size <= 192) { | |
2637 | if (!size) | |
2638 | return ZERO_SIZE_PTR; | |
2639 | ||
2640 | index = size_index[(size - 1) / 8]; | |
2641 | } else | |
2642 | index = fls(size - 1); | |
2643 | ||
2644 | #ifdef CONFIG_ZONE_DMA | |
2645 | if (unlikely((flags & SLUB_DMA))) | |
2646 | return dma_kmalloc_cache(index, flags); | |
2647 | ||
2648 | #endif | |
2649 | return &kmalloc_caches[index]; | |
2650 | } | |
2651 | ||
2652 | void *__kmalloc(size_t size, gfp_t flags) | |
2653 | { | |
2654 | struct kmem_cache *s; | |
2655 | ||
2656 | if (unlikely(size > PAGE_SIZE)) | |
2657 | return kmalloc_large(size, flags); | |
2658 | ||
2659 | s = get_slab(size, flags); | |
2660 | ||
2661 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
2662 | return s; | |
2663 | ||
2664 | return slab_alloc(s, flags, -1, _RET_IP_); | |
2665 | } | |
2666 | EXPORT_SYMBOL(__kmalloc); | |
2667 | ||
2668 | static void *kmalloc_large_node(size_t size, gfp_t flags, int node) | |
2669 | { | |
2670 | struct page *page = alloc_pages_node(node, flags | __GFP_COMP, | |
2671 | get_order(size)); | |
2672 | ||
2673 | if (page) | |
2674 | return page_address(page); | |
2675 | else | |
2676 | return NULL; | |
2677 | } | |
2678 | ||
2679 | #ifdef CONFIG_NUMA | |
2680 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
2681 | { | |
2682 | struct kmem_cache *s; | |
2683 | ||
2684 | if (unlikely(size > PAGE_SIZE)) | |
2685 | return kmalloc_large_node(size, flags, node); | |
2686 | ||
2687 | s = get_slab(size, flags); | |
2688 | ||
2689 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
2690 | return s; | |
2691 | ||
2692 | return slab_alloc(s, flags, node, _RET_IP_); | |
2693 | } | |
2694 | EXPORT_SYMBOL(__kmalloc_node); | |
2695 | #endif | |
2696 | ||
2697 | size_t ksize(const void *object) | |
2698 | { | |
2699 | struct page *page; | |
2700 | struct kmem_cache *s; | |
2701 | ||
2702 | if (unlikely(object == ZERO_SIZE_PTR)) | |
2703 | return 0; | |
2704 | ||
2705 | page = virt_to_head_page(object); | |
2706 | ||
2707 | if (unlikely(!PageSlab(page))) { | |
2708 | WARN_ON(!PageCompound(page)); | |
2709 | return PAGE_SIZE << compound_order(page); | |
2710 | } | |
2711 | s = page->slab; | |
2712 | ||
2713 | #ifdef CONFIG_SLUB_DEBUG | |
2714 | /* | |
2715 | * Debugging requires use of the padding between object | |
2716 | * and whatever may come after it. | |
2717 | */ | |
2718 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
2719 | return s->objsize; | |
2720 | ||
2721 | #endif | |
2722 | /* | |
2723 | * If we have the need to store the freelist pointer | |
2724 | * back there or track user information then we can | |
2725 | * only use the space before that information. | |
2726 | */ | |
2727 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
2728 | return s->inuse; | |
2729 | /* | |
2730 | * Else we can use all the padding etc for the allocation | |
2731 | */ | |
2732 | return s->size; | |
2733 | } | |
2734 | ||
2735 | void kfree(const void *x) | |
2736 | { | |
2737 | struct page *page; | |
2738 | void *object = (void *)x; | |
2739 | ||
2740 | if (unlikely(ZERO_OR_NULL_PTR(x))) | |
2741 | return; | |
2742 | ||
2743 | page = virt_to_head_page(x); | |
2744 | if (unlikely(!PageSlab(page))) { | |
2745 | BUG_ON(!PageCompound(page)); | |
2746 | put_page(page); | |
2747 | return; | |
2748 | } | |
2749 | slab_free(page->slab, page, object, _RET_IP_); | |
2750 | } | |
2751 | EXPORT_SYMBOL(kfree); | |
2752 | ||
2753 | /* | |
2754 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts | |
2755 | * the remaining slabs by the number of items in use. The slabs with the | |
2756 | * most items in use come first. New allocations will then fill those up | |
2757 | * and thus they can be removed from the partial lists. | |
2758 | * | |
2759 | * The slabs with the least items are placed last. This results in them | |
2760 | * being allocated from last increasing the chance that the last objects | |
2761 | * are freed in them. | |
2762 | */ | |
2763 | int kmem_cache_shrink(struct kmem_cache *s) | |
2764 | { | |
2765 | int node; | |
2766 | int i; | |
2767 | struct kmem_cache_node *n; | |
2768 | struct page *page; | |
2769 | struct page *t; | |
2770 | int objects = oo_objects(s->max); | |
2771 | struct list_head *slabs_by_inuse = | |
2772 | kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); | |
2773 | unsigned long flags; | |
2774 | ||
2775 | if (!slabs_by_inuse) | |
2776 | return -ENOMEM; | |
2777 | ||
2778 | flush_all(s); | |
2779 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2780 | n = get_node(s, node); | |
2781 | ||
2782 | if (!n->nr_partial) | |
2783 | continue; | |
2784 | ||
2785 | for (i = 0; i < objects; i++) | |
2786 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
2787 | ||
2788 | spin_lock_irqsave(&n->list_lock, flags); | |
2789 | ||
2790 | /* | |
2791 | * Build lists indexed by the items in use in each slab. | |
2792 | * | |
2793 | * Note that concurrent frees may occur while we hold the | |
2794 | * list_lock. page->inuse here is the upper limit. | |
2795 | */ | |
2796 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
2797 | if (!page->inuse && slab_trylock(page)) { | |
2798 | /* | |
2799 | * Must hold slab lock here because slab_free | |
2800 | * may have freed the last object and be | |
2801 | * waiting to release the slab. | |
2802 | */ | |
2803 | list_del(&page->lru); | |
2804 | n->nr_partial--; | |
2805 | slab_unlock(page); | |
2806 | discard_slab(s, page); | |
2807 | } else { | |
2808 | list_move(&page->lru, | |
2809 | slabs_by_inuse + page->inuse); | |
2810 | } | |
2811 | } | |
2812 | ||
2813 | /* | |
2814 | * Rebuild the partial list with the slabs filled up most | |
2815 | * first and the least used slabs at the end. | |
2816 | */ | |
2817 | for (i = objects - 1; i >= 0; i--) | |
2818 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
2819 | ||
2820 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2821 | } | |
2822 | ||
2823 | kfree(slabs_by_inuse); | |
2824 | return 0; | |
2825 | } | |
2826 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2827 | ||
2828 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) | |
2829 | static int slab_mem_going_offline_callback(void *arg) | |
2830 | { | |
2831 | struct kmem_cache *s; | |
2832 | ||
2833 | down_read(&slub_lock); | |
2834 | list_for_each_entry(s, &slab_caches, list) | |
2835 | kmem_cache_shrink(s); | |
2836 | up_read(&slub_lock); | |
2837 | ||
2838 | return 0; | |
2839 | } | |
2840 | ||
2841 | static void slab_mem_offline_callback(void *arg) | |
2842 | { | |
2843 | struct kmem_cache_node *n; | |
2844 | struct kmem_cache *s; | |
2845 | struct memory_notify *marg = arg; | |
2846 | int offline_node; | |
2847 | ||
2848 | offline_node = marg->status_change_nid; | |
2849 | ||
2850 | /* | |
2851 | * If the node still has available memory. we need kmem_cache_node | |
2852 | * for it yet. | |
2853 | */ | |
2854 | if (offline_node < 0) | |
2855 | return; | |
2856 | ||
2857 | down_read(&slub_lock); | |
2858 | list_for_each_entry(s, &slab_caches, list) { | |
2859 | n = get_node(s, offline_node); | |
2860 | if (n) { | |
2861 | /* | |
2862 | * if n->nr_slabs > 0, slabs still exist on the node | |
2863 | * that is going down. We were unable to free them, | |
2864 | * and offline_pages() function shoudn't call this | |
2865 | * callback. So, we must fail. | |
2866 | */ | |
2867 | BUG_ON(slabs_node(s, offline_node)); | |
2868 | ||
2869 | s->node[offline_node] = NULL; | |
2870 | kmem_cache_free(kmalloc_caches, n); | |
2871 | } | |
2872 | } | |
2873 | up_read(&slub_lock); | |
2874 | } | |
2875 | ||
2876 | static int slab_mem_going_online_callback(void *arg) | |
2877 | { | |
2878 | struct kmem_cache_node *n; | |
2879 | struct kmem_cache *s; | |
2880 | struct memory_notify *marg = arg; | |
2881 | int nid = marg->status_change_nid; | |
2882 | int ret = 0; | |
2883 | ||
2884 | /* | |
2885 | * If the node's memory is already available, then kmem_cache_node is | |
2886 | * already created. Nothing to do. | |
2887 | */ | |
2888 | if (nid < 0) | |
2889 | return 0; | |
2890 | ||
2891 | /* | |
2892 | * We are bringing a node online. No memory is available yet. We must | |
2893 | * allocate a kmem_cache_node structure in order to bring the node | |
2894 | * online. | |
2895 | */ | |
2896 | down_read(&slub_lock); | |
2897 | list_for_each_entry(s, &slab_caches, list) { | |
2898 | /* | |
2899 | * XXX: kmem_cache_alloc_node will fallback to other nodes | |
2900 | * since memory is not yet available from the node that | |
2901 | * is brought up. | |
2902 | */ | |
2903 | n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); | |
2904 | if (!n) { | |
2905 | ret = -ENOMEM; | |
2906 | goto out; | |
2907 | } | |
2908 | init_kmem_cache_node(n, s); | |
2909 | s->node[nid] = n; | |
2910 | } | |
2911 | out: | |
2912 | up_read(&slub_lock); | |
2913 | return ret; | |
2914 | } | |
2915 | ||
2916 | static int slab_memory_callback(struct notifier_block *self, | |
2917 | unsigned long action, void *arg) | |
2918 | { | |
2919 | int ret = 0; | |
2920 | ||
2921 | switch (action) { | |
2922 | case MEM_GOING_ONLINE: | |
2923 | ret = slab_mem_going_online_callback(arg); | |
2924 | break; | |
2925 | case MEM_GOING_OFFLINE: | |
2926 | ret = slab_mem_going_offline_callback(arg); | |
2927 | break; | |
2928 | case MEM_OFFLINE: | |
2929 | case MEM_CANCEL_ONLINE: | |
2930 | slab_mem_offline_callback(arg); | |
2931 | break; | |
2932 | case MEM_ONLINE: | |
2933 | case MEM_CANCEL_OFFLINE: | |
2934 | break; | |
2935 | } | |
2936 | if (ret) | |
2937 | ret = notifier_from_errno(ret); | |
2938 | else | |
2939 | ret = NOTIFY_OK; | |
2940 | return ret; | |
2941 | } | |
2942 | ||
2943 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
2944 | ||
2945 | /******************************************************************** | |
2946 | * Basic setup of slabs | |
2947 | *******************************************************************/ | |
2948 | ||
2949 | void __init kmem_cache_init(void) | |
2950 | { | |
2951 | int i; | |
2952 | int caches = 0; | |
2953 | ||
2954 | init_alloc_cpu(); | |
2955 | ||
2956 | #ifdef CONFIG_NUMA | |
2957 | /* | |
2958 | * Must first have the slab cache available for the allocations of the | |
2959 | * struct kmem_cache_node's. There is special bootstrap code in | |
2960 | * kmem_cache_open for slab_state == DOWN. | |
2961 | */ | |
2962 | create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", | |
2963 | sizeof(struct kmem_cache_node), GFP_KERNEL); | |
2964 | kmalloc_caches[0].refcount = -1; | |
2965 | caches++; | |
2966 | ||
2967 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
2968 | #endif | |
2969 | ||
2970 | /* Able to allocate the per node structures */ | |
2971 | slab_state = PARTIAL; | |
2972 | ||
2973 | /* Caches that are not of the two-to-the-power-of size */ | |
2974 | if (KMALLOC_MIN_SIZE <= 64) { | |
2975 | create_kmalloc_cache(&kmalloc_caches[1], | |
2976 | "kmalloc-96", 96, GFP_KERNEL); | |
2977 | caches++; | |
2978 | create_kmalloc_cache(&kmalloc_caches[2], | |
2979 | "kmalloc-192", 192, GFP_KERNEL); | |
2980 | caches++; | |
2981 | } | |
2982 | ||
2983 | for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { | |
2984 | create_kmalloc_cache(&kmalloc_caches[i], | |
2985 | "kmalloc", 1 << i, GFP_KERNEL); | |
2986 | caches++; | |
2987 | } | |
2988 | ||
2989 | ||
2990 | /* | |
2991 | * Patch up the size_index table if we have strange large alignment | |
2992 | * requirements for the kmalloc array. This is only the case for | |
2993 | * MIPS it seems. The standard arches will not generate any code here. | |
2994 | * | |
2995 | * Largest permitted alignment is 256 bytes due to the way we | |
2996 | * handle the index determination for the smaller caches. | |
2997 | * | |
2998 | * Make sure that nothing crazy happens if someone starts tinkering | |
2999 | * around with ARCH_KMALLOC_MINALIGN | |
3000 | */ | |
3001 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || | |
3002 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); | |
3003 | ||
3004 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) | |
3005 | size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; | |
3006 | ||
3007 | if (KMALLOC_MIN_SIZE == 128) { | |
3008 | /* | |
3009 | * The 192 byte sized cache is not used if the alignment | |
3010 | * is 128 byte. Redirect kmalloc to use the 256 byte cache | |
3011 | * instead. | |
3012 | */ | |
3013 | for (i = 128 + 8; i <= 192; i += 8) | |
3014 | size_index[(i - 1) / 8] = 8; | |
3015 | } | |
3016 | ||
3017 | slab_state = UP; | |
3018 | ||
3019 | /* Provide the correct kmalloc names now that the caches are up */ | |
3020 | for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) | |
3021 | kmalloc_caches[i]. name = | |
3022 | kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); | |
3023 | ||
3024 | #ifdef CONFIG_SMP | |
3025 | register_cpu_notifier(&slab_notifier); | |
3026 | kmem_size = offsetof(struct kmem_cache, cpu_slab) + | |
3027 | nr_cpu_ids * sizeof(struct kmem_cache_cpu *); | |
3028 | #else | |
3029 | kmem_size = sizeof(struct kmem_cache); | |
3030 | #endif | |
3031 | ||
3032 | printk(KERN_INFO | |
3033 | "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
3034 | " CPUs=%d, Nodes=%d\n", | |
3035 | caches, cache_line_size(), | |
3036 | slub_min_order, slub_max_order, slub_min_objects, | |
3037 | nr_cpu_ids, nr_node_ids); | |
3038 | } | |
3039 | ||
3040 | /* | |
3041 | * Find a mergeable slab cache | |
3042 | */ | |
3043 | static int slab_unmergeable(struct kmem_cache *s) | |
3044 | { | |
3045 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
3046 | return 1; | |
3047 | ||
3048 | if (s->ctor) | |
3049 | return 1; | |
3050 | ||
3051 | /* | |
3052 | * We may have set a slab to be unmergeable during bootstrap. | |
3053 | */ | |
3054 | if (s->refcount < 0) | |
3055 | return 1; | |
3056 | ||
3057 | return 0; | |
3058 | } | |
3059 | ||
3060 | static struct kmem_cache *find_mergeable(size_t size, | |
3061 | size_t align, unsigned long flags, const char *name, | |
3062 | void (*ctor)(void *)) | |
3063 | { | |
3064 | struct kmem_cache *s; | |
3065 | ||
3066 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
3067 | return NULL; | |
3068 | ||
3069 | if (ctor) | |
3070 | return NULL; | |
3071 | ||
3072 | size = ALIGN(size, sizeof(void *)); | |
3073 | align = calculate_alignment(flags, align, size); | |
3074 | size = ALIGN(size, align); | |
3075 | flags = kmem_cache_flags(size, flags, name, NULL); | |
3076 | ||
3077 | list_for_each_entry(s, &slab_caches, list) { | |
3078 | if (slab_unmergeable(s)) | |
3079 | continue; | |
3080 | ||
3081 | if (size > s->size) | |
3082 | continue; | |
3083 | ||
3084 | if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) | |
3085 | continue; | |
3086 | /* | |
3087 | * Check if alignment is compatible. | |
3088 | * Courtesy of Adrian Drzewiecki | |
3089 | */ | |
3090 | if ((s->size & ~(align - 1)) != s->size) | |
3091 | continue; | |
3092 | ||
3093 | if (s->size - size >= sizeof(void *)) | |
3094 | continue; | |
3095 | ||
3096 | return s; | |
3097 | } | |
3098 | return NULL; | |
3099 | } | |
3100 | ||
3101 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
3102 | size_t align, unsigned long flags, void (*ctor)(void *)) | |
3103 | { | |
3104 | struct kmem_cache *s; | |
3105 | ||
3106 | down_write(&slub_lock); | |
3107 | s = find_mergeable(size, align, flags, name, ctor); | |
3108 | if (s) { | |
3109 | int cpu; | |
3110 | ||
3111 | s->refcount++; | |
3112 | /* | |
3113 | * Adjust the object sizes so that we clear | |
3114 | * the complete object on kzalloc. | |
3115 | */ | |
3116 | s->objsize = max(s->objsize, (int)size); | |
3117 | ||
3118 | /* | |
3119 | * And then we need to update the object size in the | |
3120 | * per cpu structures | |
3121 | */ | |
3122 | for_each_online_cpu(cpu) | |
3123 | get_cpu_slab(s, cpu)->objsize = s->objsize; | |
3124 | ||
3125 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
3126 | up_write(&slub_lock); | |
3127 | ||
3128 | if (sysfs_slab_alias(s, name)) { | |
3129 | down_write(&slub_lock); | |
3130 | s->refcount--; | |
3131 | up_write(&slub_lock); | |
3132 | goto err; | |
3133 | } | |
3134 | return s; | |
3135 | } | |
3136 | ||
3137 | s = kmalloc(kmem_size, GFP_KERNEL); | |
3138 | if (s) { | |
3139 | if (kmem_cache_open(s, GFP_KERNEL, name, | |
3140 | size, align, flags, ctor)) { | |
3141 | list_add(&s->list, &slab_caches); | |
3142 | up_write(&slub_lock); | |
3143 | if (sysfs_slab_add(s)) { | |
3144 | down_write(&slub_lock); | |
3145 | list_del(&s->list); | |
3146 | up_write(&slub_lock); | |
3147 | kfree(s); | |
3148 | goto err; | |
3149 | } | |
3150 | return s; | |
3151 | } | |
3152 | kfree(s); | |
3153 | } | |
3154 | up_write(&slub_lock); | |
3155 | ||
3156 | err: | |
3157 | if (flags & SLAB_PANIC) | |
3158 | panic("Cannot create slabcache %s\n", name); | |
3159 | else | |
3160 | s = NULL; | |
3161 | return s; | |
3162 | } | |
3163 | EXPORT_SYMBOL(kmem_cache_create); | |
3164 | ||
3165 | #ifdef CONFIG_SMP | |
3166 | /* | |
3167 | * Use the cpu notifier to insure that the cpu slabs are flushed when | |
3168 | * necessary. | |
3169 | */ | |
3170 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
3171 | unsigned long action, void *hcpu) | |
3172 | { | |
3173 | long cpu = (long)hcpu; | |
3174 | struct kmem_cache *s; | |
3175 | unsigned long flags; | |
3176 | ||
3177 | switch (action) { | |
3178 | case CPU_UP_PREPARE: | |
3179 | case CPU_UP_PREPARE_FROZEN: | |
3180 | init_alloc_cpu_cpu(cpu); | |
3181 | down_read(&slub_lock); | |
3182 | list_for_each_entry(s, &slab_caches, list) | |
3183 | s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, | |
3184 | GFP_KERNEL); | |
3185 | up_read(&slub_lock); | |
3186 | break; | |
3187 | ||
3188 | case CPU_UP_CANCELED: | |
3189 | case CPU_UP_CANCELED_FROZEN: | |
3190 | case CPU_DEAD: | |
3191 | case CPU_DEAD_FROZEN: | |
3192 | down_read(&slub_lock); | |
3193 | list_for_each_entry(s, &slab_caches, list) { | |
3194 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
3195 | ||
3196 | local_irq_save(flags); | |
3197 | __flush_cpu_slab(s, cpu); | |
3198 | local_irq_restore(flags); | |
3199 | free_kmem_cache_cpu(c, cpu); | |
3200 | s->cpu_slab[cpu] = NULL; | |
3201 | } | |
3202 | up_read(&slub_lock); | |
3203 | break; | |
3204 | default: | |
3205 | break; | |
3206 | } | |
3207 | return NOTIFY_OK; | |
3208 | } | |
3209 | ||
3210 | static struct notifier_block __cpuinitdata slab_notifier = { | |
3211 | .notifier_call = slab_cpuup_callback | |
3212 | }; | |
3213 | ||
3214 | #endif | |
3215 | ||
3216 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) | |
3217 | { | |
3218 | struct kmem_cache *s; | |
3219 | ||
3220 | if (unlikely(size > PAGE_SIZE)) | |
3221 | return kmalloc_large(size, gfpflags); | |
3222 | ||
3223 | s = get_slab(size, gfpflags); | |
3224 | ||
3225 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3226 | return s; | |
3227 | ||
3228 | return slab_alloc(s, gfpflags, -1, caller); | |
3229 | } | |
3230 | ||
3231 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
3232 | int node, unsigned long caller) | |
3233 | { | |
3234 | struct kmem_cache *s; | |
3235 | ||
3236 | if (unlikely(size > PAGE_SIZE)) | |
3237 | return kmalloc_large_node(size, gfpflags, node); | |
3238 | ||
3239 | s = get_slab(size, gfpflags); | |
3240 | ||
3241 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3242 | return s; | |
3243 | ||
3244 | return slab_alloc(s, gfpflags, node, caller); | |
3245 | } | |
3246 | ||
3247 | #ifdef CONFIG_SLUB_DEBUG | |
3248 | static unsigned long count_partial(struct kmem_cache_node *n, | |
3249 | int (*get_count)(struct page *)) | |
3250 | { | |
3251 | unsigned long flags; | |
3252 | unsigned long x = 0; | |
3253 | struct page *page; | |
3254 | ||
3255 | spin_lock_irqsave(&n->list_lock, flags); | |
3256 | list_for_each_entry(page, &n->partial, lru) | |
3257 | x += get_count(page); | |
3258 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3259 | return x; | |
3260 | } | |
3261 | ||
3262 | static int count_inuse(struct page *page) | |
3263 | { | |
3264 | return page->inuse; | |
3265 | } | |
3266 | ||
3267 | static int count_total(struct page *page) | |
3268 | { | |
3269 | return page->objects; | |
3270 | } | |
3271 | ||
3272 | static int count_free(struct page *page) | |
3273 | { | |
3274 | return page->objects - page->inuse; | |
3275 | } | |
3276 | ||
3277 | static int validate_slab(struct kmem_cache *s, struct page *page, | |
3278 | unsigned long *map) | |
3279 | { | |
3280 | void *p; | |
3281 | void *addr = page_address(page); | |
3282 | ||
3283 | if (!check_slab(s, page) || | |
3284 | !on_freelist(s, page, NULL)) | |
3285 | return 0; | |
3286 | ||
3287 | /* Now we know that a valid freelist exists */ | |
3288 | bitmap_zero(map, page->objects); | |
3289 | ||
3290 | for_each_free_object(p, s, page->freelist) { | |
3291 | set_bit(slab_index(p, s, addr), map); | |
3292 | if (!check_object(s, page, p, 0)) | |
3293 | return 0; | |
3294 | } | |
3295 | ||
3296 | for_each_object(p, s, addr, page->objects) | |
3297 | if (!test_bit(slab_index(p, s, addr), map)) | |
3298 | if (!check_object(s, page, p, 1)) | |
3299 | return 0; | |
3300 | return 1; | |
3301 | } | |
3302 | ||
3303 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, | |
3304 | unsigned long *map) | |
3305 | { | |
3306 | if (slab_trylock(page)) { | |
3307 | validate_slab(s, page, map); | |
3308 | slab_unlock(page); | |
3309 | } else | |
3310 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", | |
3311 | s->name, page); | |
3312 | ||
3313 | if (s->flags & DEBUG_DEFAULT_FLAGS) { | |
3314 | if (!PageSlubDebug(page)) | |
3315 | printk(KERN_ERR "SLUB %s: SlubDebug not set " | |
3316 | "on slab 0x%p\n", s->name, page); | |
3317 | } else { | |
3318 | if (PageSlubDebug(page)) | |
3319 | printk(KERN_ERR "SLUB %s: SlubDebug set on " | |
3320 | "slab 0x%p\n", s->name, page); | |
3321 | } | |
3322 | } | |
3323 | ||
3324 | static int validate_slab_node(struct kmem_cache *s, | |
3325 | struct kmem_cache_node *n, unsigned long *map) | |
3326 | { | |
3327 | unsigned long count = 0; | |
3328 | struct page *page; | |
3329 | unsigned long flags; | |
3330 | ||
3331 | spin_lock_irqsave(&n->list_lock, flags); | |
3332 | ||
3333 | list_for_each_entry(page, &n->partial, lru) { | |
3334 | validate_slab_slab(s, page, map); | |
3335 | count++; | |
3336 | } | |
3337 | if (count != n->nr_partial) | |
3338 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
3339 | "counter=%ld\n", s->name, count, n->nr_partial); | |
3340 | ||
3341 | if (!(s->flags & SLAB_STORE_USER)) | |
3342 | goto out; | |
3343 | ||
3344 | list_for_each_entry(page, &n->full, lru) { | |
3345 | validate_slab_slab(s, page, map); | |
3346 | count++; | |
3347 | } | |
3348 | if (count != atomic_long_read(&n->nr_slabs)) | |
3349 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
3350 | "counter=%ld\n", s->name, count, | |
3351 | atomic_long_read(&n->nr_slabs)); | |
3352 | ||
3353 | out: | |
3354 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3355 | return count; | |
3356 | } | |
3357 | ||
3358 | static long validate_slab_cache(struct kmem_cache *s) | |
3359 | { | |
3360 | int node; | |
3361 | unsigned long count = 0; | |
3362 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * | |
3363 | sizeof(unsigned long), GFP_KERNEL); | |
3364 | ||
3365 | if (!map) | |
3366 | return -ENOMEM; | |
3367 | ||
3368 | flush_all(s); | |
3369 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3370 | struct kmem_cache_node *n = get_node(s, node); | |
3371 | ||
3372 | count += validate_slab_node(s, n, map); | |
3373 | } | |
3374 | kfree(map); | |
3375 | return count; | |
3376 | } | |
3377 | ||
3378 | #ifdef SLUB_RESILIENCY_TEST | |
3379 | static void resiliency_test(void) | |
3380 | { | |
3381 | u8 *p; | |
3382 | ||
3383 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
3384 | printk(KERN_ERR "-----------------------\n"); | |
3385 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
3386 | ||
3387 | p = kzalloc(16, GFP_KERNEL); | |
3388 | p[16] = 0x12; | |
3389 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
3390 | " 0x12->0x%p\n\n", p + 16); | |
3391 | ||
3392 | validate_slab_cache(kmalloc_caches + 4); | |
3393 | ||
3394 | /* Hmmm... The next two are dangerous */ | |
3395 | p = kzalloc(32, GFP_KERNEL); | |
3396 | p[32 + sizeof(void *)] = 0x34; | |
3397 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
3398 | " 0x34 -> -0x%p\n", p); | |
3399 | printk(KERN_ERR | |
3400 | "If allocated object is overwritten then not detectable\n\n"); | |
3401 | ||
3402 | validate_slab_cache(kmalloc_caches + 5); | |
3403 | p = kzalloc(64, GFP_KERNEL); | |
3404 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
3405 | *p = 0x56; | |
3406 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
3407 | p); | |
3408 | printk(KERN_ERR | |
3409 | "If allocated object is overwritten then not detectable\n\n"); | |
3410 | validate_slab_cache(kmalloc_caches + 6); | |
3411 | ||
3412 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
3413 | p = kzalloc(128, GFP_KERNEL); | |
3414 | kfree(p); | |
3415 | *p = 0x78; | |
3416 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
3417 | validate_slab_cache(kmalloc_caches + 7); | |
3418 | ||
3419 | p = kzalloc(256, GFP_KERNEL); | |
3420 | kfree(p); | |
3421 | p[50] = 0x9a; | |
3422 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", | |
3423 | p); | |
3424 | validate_slab_cache(kmalloc_caches + 8); | |
3425 | ||
3426 | p = kzalloc(512, GFP_KERNEL); | |
3427 | kfree(p); | |
3428 | p[512] = 0xab; | |
3429 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
3430 | validate_slab_cache(kmalloc_caches + 9); | |
3431 | } | |
3432 | #else | |
3433 | static void resiliency_test(void) {}; | |
3434 | #endif | |
3435 | ||
3436 | /* | |
3437 | * Generate lists of code addresses where slabcache objects are allocated | |
3438 | * and freed. | |
3439 | */ | |
3440 | ||
3441 | struct location { | |
3442 | unsigned long count; | |
3443 | unsigned long addr; | |
3444 | long long sum_time; | |
3445 | long min_time; | |
3446 | long max_time; | |
3447 | long min_pid; | |
3448 | long max_pid; | |
3449 | cpumask_t cpus; | |
3450 | nodemask_t nodes; | |
3451 | }; | |
3452 | ||
3453 | struct loc_track { | |
3454 | unsigned long max; | |
3455 | unsigned long count; | |
3456 | struct location *loc; | |
3457 | }; | |
3458 | ||
3459 | static void free_loc_track(struct loc_track *t) | |
3460 | { | |
3461 | if (t->max) | |
3462 | free_pages((unsigned long)t->loc, | |
3463 | get_order(sizeof(struct location) * t->max)); | |
3464 | } | |
3465 | ||
3466 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) | |
3467 | { | |
3468 | struct location *l; | |
3469 | int order; | |
3470 | ||
3471 | order = get_order(sizeof(struct location) * max); | |
3472 | ||
3473 | l = (void *)__get_free_pages(flags, order); | |
3474 | if (!l) | |
3475 | return 0; | |
3476 | ||
3477 | if (t->count) { | |
3478 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
3479 | free_loc_track(t); | |
3480 | } | |
3481 | t->max = max; | |
3482 | t->loc = l; | |
3483 | return 1; | |
3484 | } | |
3485 | ||
3486 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
3487 | const struct track *track) | |
3488 | { | |
3489 | long start, end, pos; | |
3490 | struct location *l; | |
3491 | unsigned long caddr; | |
3492 | unsigned long age = jiffies - track->when; | |
3493 | ||
3494 | start = -1; | |
3495 | end = t->count; | |
3496 | ||
3497 | for ( ; ; ) { | |
3498 | pos = start + (end - start + 1) / 2; | |
3499 | ||
3500 | /* | |
3501 | * There is nothing at "end". If we end up there | |
3502 | * we need to add something to before end. | |
3503 | */ | |
3504 | if (pos == end) | |
3505 | break; | |
3506 | ||
3507 | caddr = t->loc[pos].addr; | |
3508 | if (track->addr == caddr) { | |
3509 | ||
3510 | l = &t->loc[pos]; | |
3511 | l->count++; | |
3512 | if (track->when) { | |
3513 | l->sum_time += age; | |
3514 | if (age < l->min_time) | |
3515 | l->min_time = age; | |
3516 | if (age > l->max_time) | |
3517 | l->max_time = age; | |
3518 | ||
3519 | if (track->pid < l->min_pid) | |
3520 | l->min_pid = track->pid; | |
3521 | if (track->pid > l->max_pid) | |
3522 | l->max_pid = track->pid; | |
3523 | ||
3524 | cpu_set(track->cpu, l->cpus); | |
3525 | } | |
3526 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
3527 | return 1; | |
3528 | } | |
3529 | ||
3530 | if (track->addr < caddr) | |
3531 | end = pos; | |
3532 | else | |
3533 | start = pos; | |
3534 | } | |
3535 | ||
3536 | /* | |
3537 | * Not found. Insert new tracking element. | |
3538 | */ | |
3539 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) | |
3540 | return 0; | |
3541 | ||
3542 | l = t->loc + pos; | |
3543 | if (pos < t->count) | |
3544 | memmove(l + 1, l, | |
3545 | (t->count - pos) * sizeof(struct location)); | |
3546 | t->count++; | |
3547 | l->count = 1; | |
3548 | l->addr = track->addr; | |
3549 | l->sum_time = age; | |
3550 | l->min_time = age; | |
3551 | l->max_time = age; | |
3552 | l->min_pid = track->pid; | |
3553 | l->max_pid = track->pid; | |
3554 | cpus_clear(l->cpus); | |
3555 | cpu_set(track->cpu, l->cpus); | |
3556 | nodes_clear(l->nodes); | |
3557 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
3558 | return 1; | |
3559 | } | |
3560 | ||
3561 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
3562 | struct page *page, enum track_item alloc) | |
3563 | { | |
3564 | void *addr = page_address(page); | |
3565 | DECLARE_BITMAP(map, page->objects); | |
3566 | void *p; | |
3567 | ||
3568 | bitmap_zero(map, page->objects); | |
3569 | for_each_free_object(p, s, page->freelist) | |
3570 | set_bit(slab_index(p, s, addr), map); | |
3571 | ||
3572 | for_each_object(p, s, addr, page->objects) | |
3573 | if (!test_bit(slab_index(p, s, addr), map)) | |
3574 | add_location(t, s, get_track(s, p, alloc)); | |
3575 | } | |
3576 | ||
3577 | static int list_locations(struct kmem_cache *s, char *buf, | |
3578 | enum track_item alloc) | |
3579 | { | |
3580 | int len = 0; | |
3581 | unsigned long i; | |
3582 | struct loc_track t = { 0, 0, NULL }; | |
3583 | int node; | |
3584 | ||
3585 | if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), | |
3586 | GFP_TEMPORARY)) | |
3587 | return sprintf(buf, "Out of memory\n"); | |
3588 | ||
3589 | /* Push back cpu slabs */ | |
3590 | flush_all(s); | |
3591 | ||
3592 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3593 | struct kmem_cache_node *n = get_node(s, node); | |
3594 | unsigned long flags; | |
3595 | struct page *page; | |
3596 | ||
3597 | if (!atomic_long_read(&n->nr_slabs)) | |
3598 | continue; | |
3599 | ||
3600 | spin_lock_irqsave(&n->list_lock, flags); | |
3601 | list_for_each_entry(page, &n->partial, lru) | |
3602 | process_slab(&t, s, page, alloc); | |
3603 | list_for_each_entry(page, &n->full, lru) | |
3604 | process_slab(&t, s, page, alloc); | |
3605 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3606 | } | |
3607 | ||
3608 | for (i = 0; i < t.count; i++) { | |
3609 | struct location *l = &t.loc[i]; | |
3610 | ||
3611 | if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) | |
3612 | break; | |
3613 | len += sprintf(buf + len, "%7ld ", l->count); | |
3614 | ||
3615 | if (l->addr) | |
3616 | len += sprint_symbol(buf + len, (unsigned long)l->addr); | |
3617 | else | |
3618 | len += sprintf(buf + len, "<not-available>"); | |
3619 | ||
3620 | if (l->sum_time != l->min_time) { | |
3621 | len += sprintf(buf + len, " age=%ld/%ld/%ld", | |
3622 | l->min_time, | |
3623 | (long)div_u64(l->sum_time, l->count), | |
3624 | l->max_time); | |
3625 | } else | |
3626 | len += sprintf(buf + len, " age=%ld", | |
3627 | l->min_time); | |
3628 | ||
3629 | if (l->min_pid != l->max_pid) | |
3630 | len += sprintf(buf + len, " pid=%ld-%ld", | |
3631 | l->min_pid, l->max_pid); | |
3632 | else | |
3633 | len += sprintf(buf + len, " pid=%ld", | |
3634 | l->min_pid); | |
3635 | ||
3636 | if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && | |
3637 | len < PAGE_SIZE - 60) { | |
3638 | len += sprintf(buf + len, " cpus="); | |
3639 | len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
3640 | l->cpus); | |
3641 | } | |
3642 | ||
3643 | if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && | |
3644 | len < PAGE_SIZE - 60) { | |
3645 | len += sprintf(buf + len, " nodes="); | |
3646 | len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
3647 | l->nodes); | |
3648 | } | |
3649 | ||
3650 | len += sprintf(buf + len, "\n"); | |
3651 | } | |
3652 | ||
3653 | free_loc_track(&t); | |
3654 | if (!t.count) | |
3655 | len += sprintf(buf, "No data\n"); | |
3656 | return len; | |
3657 | } | |
3658 | ||
3659 | enum slab_stat_type { | |
3660 | SL_ALL, /* All slabs */ | |
3661 | SL_PARTIAL, /* Only partially allocated slabs */ | |
3662 | SL_CPU, /* Only slabs used for cpu caches */ | |
3663 | SL_OBJECTS, /* Determine allocated objects not slabs */ | |
3664 | SL_TOTAL /* Determine object capacity not slabs */ | |
3665 | }; | |
3666 | ||
3667 | #define SO_ALL (1 << SL_ALL) | |
3668 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
3669 | #define SO_CPU (1 << SL_CPU) | |
3670 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
3671 | #define SO_TOTAL (1 << SL_TOTAL) | |
3672 | ||
3673 | static ssize_t show_slab_objects(struct kmem_cache *s, | |
3674 | char *buf, unsigned long flags) | |
3675 | { | |
3676 | unsigned long total = 0; | |
3677 | int node; | |
3678 | int x; | |
3679 | unsigned long *nodes; | |
3680 | unsigned long *per_cpu; | |
3681 | ||
3682 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
3683 | if (!nodes) | |
3684 | return -ENOMEM; | |
3685 | per_cpu = nodes + nr_node_ids; | |
3686 | ||
3687 | if (flags & SO_CPU) { | |
3688 | int cpu; | |
3689 | ||
3690 | for_each_possible_cpu(cpu) { | |
3691 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
3692 | ||
3693 | if (!c || c->node < 0) | |
3694 | continue; | |
3695 | ||
3696 | if (c->page) { | |
3697 | if (flags & SO_TOTAL) | |
3698 | x = c->page->objects; | |
3699 | else if (flags & SO_OBJECTS) | |
3700 | x = c->page->inuse; | |
3701 | else | |
3702 | x = 1; | |
3703 | ||
3704 | total += x; | |
3705 | nodes[c->node] += x; | |
3706 | } | |
3707 | per_cpu[c->node]++; | |
3708 | } | |
3709 | } | |
3710 | ||
3711 | if (flags & SO_ALL) { | |
3712 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3713 | struct kmem_cache_node *n = get_node(s, node); | |
3714 | ||
3715 | if (flags & SO_TOTAL) | |
3716 | x = atomic_long_read(&n->total_objects); | |
3717 | else if (flags & SO_OBJECTS) | |
3718 | x = atomic_long_read(&n->total_objects) - | |
3719 | count_partial(n, count_free); | |
3720 | ||
3721 | else | |
3722 | x = atomic_long_read(&n->nr_slabs); | |
3723 | total += x; | |
3724 | nodes[node] += x; | |
3725 | } | |
3726 | ||
3727 | } else if (flags & SO_PARTIAL) { | |
3728 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3729 | struct kmem_cache_node *n = get_node(s, node); | |
3730 | ||
3731 | if (flags & SO_TOTAL) | |
3732 | x = count_partial(n, count_total); | |
3733 | else if (flags & SO_OBJECTS) | |
3734 | x = count_partial(n, count_inuse); | |
3735 | else | |
3736 | x = n->nr_partial; | |
3737 | total += x; | |
3738 | nodes[node] += x; | |
3739 | } | |
3740 | } | |
3741 | x = sprintf(buf, "%lu", total); | |
3742 | #ifdef CONFIG_NUMA | |
3743 | for_each_node_state(node, N_NORMAL_MEMORY) | |
3744 | if (nodes[node]) | |
3745 | x += sprintf(buf + x, " N%d=%lu", | |
3746 | node, nodes[node]); | |
3747 | #endif | |
3748 | kfree(nodes); | |
3749 | return x + sprintf(buf + x, "\n"); | |
3750 | } | |
3751 | ||
3752 | static int any_slab_objects(struct kmem_cache *s) | |
3753 | { | |
3754 | int node; | |
3755 | ||
3756 | for_each_online_node(node) { | |
3757 | struct kmem_cache_node *n = get_node(s, node); | |
3758 | ||
3759 | if (!n) | |
3760 | continue; | |
3761 | ||
3762 | if (atomic_long_read(&n->total_objects)) | |
3763 | return 1; | |
3764 | } | |
3765 | return 0; | |
3766 | } | |
3767 | ||
3768 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
3769 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); | |
3770 | ||
3771 | struct slab_attribute { | |
3772 | struct attribute attr; | |
3773 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
3774 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
3775 | }; | |
3776 | ||
3777 | #define SLAB_ATTR_RO(_name) \ | |
3778 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) | |
3779 | ||
3780 | #define SLAB_ATTR(_name) \ | |
3781 | static struct slab_attribute _name##_attr = \ | |
3782 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3783 | ||
3784 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) | |
3785 | { | |
3786 | return sprintf(buf, "%d\n", s->size); | |
3787 | } | |
3788 | SLAB_ATTR_RO(slab_size); | |
3789 | ||
3790 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
3791 | { | |
3792 | return sprintf(buf, "%d\n", s->align); | |
3793 | } | |
3794 | SLAB_ATTR_RO(align); | |
3795 | ||
3796 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
3797 | { | |
3798 | return sprintf(buf, "%d\n", s->objsize); | |
3799 | } | |
3800 | SLAB_ATTR_RO(object_size); | |
3801 | ||
3802 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
3803 | { | |
3804 | return sprintf(buf, "%d\n", oo_objects(s->oo)); | |
3805 | } | |
3806 | SLAB_ATTR_RO(objs_per_slab); | |
3807 | ||
3808 | static ssize_t order_store(struct kmem_cache *s, | |
3809 | const char *buf, size_t length) | |
3810 | { | |
3811 | unsigned long order; | |
3812 | int err; | |
3813 | ||
3814 | err = strict_strtoul(buf, 10, &order); | |
3815 | if (err) | |
3816 | return err; | |
3817 | ||
3818 | if (order > slub_max_order || order < slub_min_order) | |
3819 | return -EINVAL; | |
3820 | ||
3821 | calculate_sizes(s, order); | |
3822 | return length; | |
3823 | } | |
3824 | ||
3825 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
3826 | { | |
3827 | return sprintf(buf, "%d\n", oo_order(s->oo)); | |
3828 | } | |
3829 | SLAB_ATTR(order); | |
3830 | ||
3831 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
3832 | { | |
3833 | if (s->ctor) { | |
3834 | int n = sprint_symbol(buf, (unsigned long)s->ctor); | |
3835 | ||
3836 | return n + sprintf(buf + n, "\n"); | |
3837 | } | |
3838 | return 0; | |
3839 | } | |
3840 | SLAB_ATTR_RO(ctor); | |
3841 | ||
3842 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) | |
3843 | { | |
3844 | return sprintf(buf, "%d\n", s->refcount - 1); | |
3845 | } | |
3846 | SLAB_ATTR_RO(aliases); | |
3847 | ||
3848 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
3849 | { | |
3850 | return show_slab_objects(s, buf, SO_ALL); | |
3851 | } | |
3852 | SLAB_ATTR_RO(slabs); | |
3853 | ||
3854 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
3855 | { | |
3856 | return show_slab_objects(s, buf, SO_PARTIAL); | |
3857 | } | |
3858 | SLAB_ATTR_RO(partial); | |
3859 | ||
3860 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
3861 | { | |
3862 | return show_slab_objects(s, buf, SO_CPU); | |
3863 | } | |
3864 | SLAB_ATTR_RO(cpu_slabs); | |
3865 | ||
3866 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
3867 | { | |
3868 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); | |
3869 | } | |
3870 | SLAB_ATTR_RO(objects); | |
3871 | ||
3872 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) | |
3873 | { | |
3874 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); | |
3875 | } | |
3876 | SLAB_ATTR_RO(objects_partial); | |
3877 | ||
3878 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) | |
3879 | { | |
3880 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); | |
3881 | } | |
3882 | SLAB_ATTR_RO(total_objects); | |
3883 | ||
3884 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
3885 | { | |
3886 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
3887 | } | |
3888 | ||
3889 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
3890 | const char *buf, size_t length) | |
3891 | { | |
3892 | s->flags &= ~SLAB_DEBUG_FREE; | |
3893 | if (buf[0] == '1') | |
3894 | s->flags |= SLAB_DEBUG_FREE; | |
3895 | return length; | |
3896 | } | |
3897 | SLAB_ATTR(sanity_checks); | |
3898 | ||
3899 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
3900 | { | |
3901 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
3902 | } | |
3903 | ||
3904 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
3905 | size_t length) | |
3906 | { | |
3907 | s->flags &= ~SLAB_TRACE; | |
3908 | if (buf[0] == '1') | |
3909 | s->flags |= SLAB_TRACE; | |
3910 | return length; | |
3911 | } | |
3912 | SLAB_ATTR(trace); | |
3913 | ||
3914 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
3915 | { | |
3916 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
3917 | } | |
3918 | ||
3919 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
3920 | const char *buf, size_t length) | |
3921 | { | |
3922 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
3923 | if (buf[0] == '1') | |
3924 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
3925 | return length; | |
3926 | } | |
3927 | SLAB_ATTR(reclaim_account); | |
3928 | ||
3929 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
3930 | { | |
3931 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); | |
3932 | } | |
3933 | SLAB_ATTR_RO(hwcache_align); | |
3934 | ||
3935 | #ifdef CONFIG_ZONE_DMA | |
3936 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
3937 | { | |
3938 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
3939 | } | |
3940 | SLAB_ATTR_RO(cache_dma); | |
3941 | #endif | |
3942 | ||
3943 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
3944 | { | |
3945 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
3946 | } | |
3947 | SLAB_ATTR_RO(destroy_by_rcu); | |
3948 | ||
3949 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
3950 | { | |
3951 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
3952 | } | |
3953 | ||
3954 | static ssize_t red_zone_store(struct kmem_cache *s, | |
3955 | const char *buf, size_t length) | |
3956 | { | |
3957 | if (any_slab_objects(s)) | |
3958 | return -EBUSY; | |
3959 | ||
3960 | s->flags &= ~SLAB_RED_ZONE; | |
3961 | if (buf[0] == '1') | |
3962 | s->flags |= SLAB_RED_ZONE; | |
3963 | calculate_sizes(s, -1); | |
3964 | return length; | |
3965 | } | |
3966 | SLAB_ATTR(red_zone); | |
3967 | ||
3968 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
3969 | { | |
3970 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
3971 | } | |
3972 | ||
3973 | static ssize_t poison_store(struct kmem_cache *s, | |
3974 | const char *buf, size_t length) | |
3975 | { | |
3976 | if (any_slab_objects(s)) | |
3977 | return -EBUSY; | |
3978 | ||
3979 | s->flags &= ~SLAB_POISON; | |
3980 | if (buf[0] == '1') | |
3981 | s->flags |= SLAB_POISON; | |
3982 | calculate_sizes(s, -1); | |
3983 | return length; | |
3984 | } | |
3985 | SLAB_ATTR(poison); | |
3986 | ||
3987 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
3988 | { | |
3989 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
3990 | } | |
3991 | ||
3992 | static ssize_t store_user_store(struct kmem_cache *s, | |
3993 | const char *buf, size_t length) | |
3994 | { | |
3995 | if (any_slab_objects(s)) | |
3996 | return -EBUSY; | |
3997 | ||
3998 | s->flags &= ~SLAB_STORE_USER; | |
3999 | if (buf[0] == '1') | |
4000 | s->flags |= SLAB_STORE_USER; | |
4001 | calculate_sizes(s, -1); | |
4002 | return length; | |
4003 | } | |
4004 | SLAB_ATTR(store_user); | |
4005 | ||
4006 | static ssize_t validate_show(struct kmem_cache *s, char *buf) | |
4007 | { | |
4008 | return 0; | |
4009 | } | |
4010 | ||
4011 | static ssize_t validate_store(struct kmem_cache *s, | |
4012 | const char *buf, size_t length) | |
4013 | { | |
4014 | int ret = -EINVAL; | |
4015 | ||
4016 | if (buf[0] == '1') { | |
4017 | ret = validate_slab_cache(s); | |
4018 | if (ret >= 0) | |
4019 | ret = length; | |
4020 | } | |
4021 | return ret; | |
4022 | } | |
4023 | SLAB_ATTR(validate); | |
4024 | ||
4025 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) | |
4026 | { | |
4027 | return 0; | |
4028 | } | |
4029 | ||
4030 | static ssize_t shrink_store(struct kmem_cache *s, | |
4031 | const char *buf, size_t length) | |
4032 | { | |
4033 | if (buf[0] == '1') { | |
4034 | int rc = kmem_cache_shrink(s); | |
4035 | ||
4036 | if (rc) | |
4037 | return rc; | |
4038 | } else | |
4039 | return -EINVAL; | |
4040 | return length; | |
4041 | } | |
4042 | SLAB_ATTR(shrink); | |
4043 | ||
4044 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) | |
4045 | { | |
4046 | if (!(s->flags & SLAB_STORE_USER)) | |
4047 | return -ENOSYS; | |
4048 | return list_locations(s, buf, TRACK_ALLOC); | |
4049 | } | |
4050 | SLAB_ATTR_RO(alloc_calls); | |
4051 | ||
4052 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
4053 | { | |
4054 | if (!(s->flags & SLAB_STORE_USER)) | |
4055 | return -ENOSYS; | |
4056 | return list_locations(s, buf, TRACK_FREE); | |
4057 | } | |
4058 | SLAB_ATTR_RO(free_calls); | |
4059 | ||
4060 | #ifdef CONFIG_NUMA | |
4061 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) | |
4062 | { | |
4063 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); | |
4064 | } | |
4065 | ||
4066 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, | |
4067 | const char *buf, size_t length) | |
4068 | { | |
4069 | unsigned long ratio; | |
4070 | int err; | |
4071 | ||
4072 | err = strict_strtoul(buf, 10, &ratio); | |
4073 | if (err) | |
4074 | return err; | |
4075 | ||
4076 | if (ratio <= 100) | |
4077 | s->remote_node_defrag_ratio = ratio * 10; | |
4078 | ||
4079 | return length; | |
4080 | } | |
4081 | SLAB_ATTR(remote_node_defrag_ratio); | |
4082 | #endif | |
4083 | ||
4084 | #ifdef CONFIG_SLUB_STATS | |
4085 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) | |
4086 | { | |
4087 | unsigned long sum = 0; | |
4088 | int cpu; | |
4089 | int len; | |
4090 | int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); | |
4091 | ||
4092 | if (!data) | |
4093 | return -ENOMEM; | |
4094 | ||
4095 | for_each_online_cpu(cpu) { | |
4096 | unsigned x = get_cpu_slab(s, cpu)->stat[si]; | |
4097 | ||
4098 | data[cpu] = x; | |
4099 | sum += x; | |
4100 | } | |
4101 | ||
4102 | len = sprintf(buf, "%lu", sum); | |
4103 | ||
4104 | #ifdef CONFIG_SMP | |
4105 | for_each_online_cpu(cpu) { | |
4106 | if (data[cpu] && len < PAGE_SIZE - 20) | |
4107 | len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); | |
4108 | } | |
4109 | #endif | |
4110 | kfree(data); | |
4111 | return len + sprintf(buf + len, "\n"); | |
4112 | } | |
4113 | ||
4114 | #define STAT_ATTR(si, text) \ | |
4115 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ | |
4116 | { \ | |
4117 | return show_stat(s, buf, si); \ | |
4118 | } \ | |
4119 | SLAB_ATTR_RO(text); \ | |
4120 | ||
4121 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); | |
4122 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); | |
4123 | STAT_ATTR(FREE_FASTPATH, free_fastpath); | |
4124 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); | |
4125 | STAT_ATTR(FREE_FROZEN, free_frozen); | |
4126 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); | |
4127 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); | |
4128 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); | |
4129 | STAT_ATTR(ALLOC_SLAB, alloc_slab); | |
4130 | STAT_ATTR(ALLOC_REFILL, alloc_refill); | |
4131 | STAT_ATTR(FREE_SLAB, free_slab); | |
4132 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); | |
4133 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); | |
4134 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); | |
4135 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); | |
4136 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); | |
4137 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); | |
4138 | STAT_ATTR(ORDER_FALLBACK, order_fallback); | |
4139 | #endif | |
4140 | ||
4141 | static struct attribute *slab_attrs[] = { | |
4142 | &slab_size_attr.attr, | |
4143 | &object_size_attr.attr, | |
4144 | &objs_per_slab_attr.attr, | |
4145 | &order_attr.attr, | |
4146 | &objects_attr.attr, | |
4147 | &objects_partial_attr.attr, | |
4148 | &total_objects_attr.attr, | |
4149 | &slabs_attr.attr, | |
4150 | &partial_attr.attr, | |
4151 | &cpu_slabs_attr.attr, | |
4152 | &ctor_attr.attr, | |
4153 | &aliases_attr.attr, | |
4154 | &align_attr.attr, | |
4155 | &sanity_checks_attr.attr, | |
4156 | &trace_attr.attr, | |
4157 | &hwcache_align_attr.attr, | |
4158 | &reclaim_account_attr.attr, | |
4159 | &destroy_by_rcu_attr.attr, | |
4160 | &red_zone_attr.attr, | |
4161 | &poison_attr.attr, | |
4162 | &store_user_attr.attr, | |
4163 | &validate_attr.attr, | |
4164 | &shrink_attr.attr, | |
4165 | &alloc_calls_attr.attr, | |
4166 | &free_calls_attr.attr, | |
4167 | #ifdef CONFIG_ZONE_DMA | |
4168 | &cache_dma_attr.attr, | |
4169 | #endif | |
4170 | #ifdef CONFIG_NUMA | |
4171 | &remote_node_defrag_ratio_attr.attr, | |
4172 | #endif | |
4173 | #ifdef CONFIG_SLUB_STATS | |
4174 | &alloc_fastpath_attr.attr, | |
4175 | &alloc_slowpath_attr.attr, | |
4176 | &free_fastpath_attr.attr, | |
4177 | &free_slowpath_attr.attr, | |
4178 | &free_frozen_attr.attr, | |
4179 | &free_add_partial_attr.attr, | |
4180 | &free_remove_partial_attr.attr, | |
4181 | &alloc_from_partial_attr.attr, | |
4182 | &alloc_slab_attr.attr, | |
4183 | &alloc_refill_attr.attr, | |
4184 | &free_slab_attr.attr, | |
4185 | &cpuslab_flush_attr.attr, | |
4186 | &deactivate_full_attr.attr, | |
4187 | &deactivate_empty_attr.attr, | |
4188 | &deactivate_to_head_attr.attr, | |
4189 | &deactivate_to_tail_attr.attr, | |
4190 | &deactivate_remote_frees_attr.attr, | |
4191 | &order_fallback_attr.attr, | |
4192 | #endif | |
4193 | NULL | |
4194 | }; | |
4195 | ||
4196 | static struct attribute_group slab_attr_group = { | |
4197 | .attrs = slab_attrs, | |
4198 | }; | |
4199 | ||
4200 | static ssize_t slab_attr_show(struct kobject *kobj, | |
4201 | struct attribute *attr, | |
4202 | char *buf) | |
4203 | { | |
4204 | struct slab_attribute *attribute; | |
4205 | struct kmem_cache *s; | |
4206 | int err; | |
4207 | ||
4208 | attribute = to_slab_attr(attr); | |
4209 | s = to_slab(kobj); | |
4210 | ||
4211 | if (!attribute->show) | |
4212 | return -EIO; | |
4213 | ||
4214 | err = attribute->show(s, buf); | |
4215 | ||
4216 | return err; | |
4217 | } | |
4218 | ||
4219 | static ssize_t slab_attr_store(struct kobject *kobj, | |
4220 | struct attribute *attr, | |
4221 | const char *buf, size_t len) | |
4222 | { | |
4223 | struct slab_attribute *attribute; | |
4224 | struct kmem_cache *s; | |
4225 | int err; | |
4226 | ||
4227 | attribute = to_slab_attr(attr); | |
4228 | s = to_slab(kobj); | |
4229 | ||
4230 | if (!attribute->store) | |
4231 | return -EIO; | |
4232 | ||
4233 | err = attribute->store(s, buf, len); | |
4234 | ||
4235 | return err; | |
4236 | } | |
4237 | ||
4238 | static void kmem_cache_release(struct kobject *kobj) | |
4239 | { | |
4240 | struct kmem_cache *s = to_slab(kobj); | |
4241 | ||
4242 | kfree(s); | |
4243 | } | |
4244 | ||
4245 | static struct sysfs_ops slab_sysfs_ops = { | |
4246 | .show = slab_attr_show, | |
4247 | .store = slab_attr_store, | |
4248 | }; | |
4249 | ||
4250 | static struct kobj_type slab_ktype = { | |
4251 | .sysfs_ops = &slab_sysfs_ops, | |
4252 | .release = kmem_cache_release | |
4253 | }; | |
4254 | ||
4255 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
4256 | { | |
4257 | struct kobj_type *ktype = get_ktype(kobj); | |
4258 | ||
4259 | if (ktype == &slab_ktype) | |
4260 | return 1; | |
4261 | return 0; | |
4262 | } | |
4263 | ||
4264 | static struct kset_uevent_ops slab_uevent_ops = { | |
4265 | .filter = uevent_filter, | |
4266 | }; | |
4267 | ||
4268 | static struct kset *slab_kset; | |
4269 | ||
4270 | #define ID_STR_LENGTH 64 | |
4271 | ||
4272 | /* Create a unique string id for a slab cache: | |
4273 | * | |
4274 | * Format :[flags-]size | |
4275 | */ | |
4276 | static char *create_unique_id(struct kmem_cache *s) | |
4277 | { | |
4278 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
4279 | char *p = name; | |
4280 | ||
4281 | BUG_ON(!name); | |
4282 | ||
4283 | *p++ = ':'; | |
4284 | /* | |
4285 | * First flags affecting slabcache operations. We will only | |
4286 | * get here for aliasable slabs so we do not need to support | |
4287 | * too many flags. The flags here must cover all flags that | |
4288 | * are matched during merging to guarantee that the id is | |
4289 | * unique. | |
4290 | */ | |
4291 | if (s->flags & SLAB_CACHE_DMA) | |
4292 | *p++ = 'd'; | |
4293 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
4294 | *p++ = 'a'; | |
4295 | if (s->flags & SLAB_DEBUG_FREE) | |
4296 | *p++ = 'F'; | |
4297 | if (p != name + 1) | |
4298 | *p++ = '-'; | |
4299 | p += sprintf(p, "%07d", s->size); | |
4300 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
4301 | return name; | |
4302 | } | |
4303 | ||
4304 | static int sysfs_slab_add(struct kmem_cache *s) | |
4305 | { | |
4306 | int err; | |
4307 | const char *name; | |
4308 | int unmergeable; | |
4309 | ||
4310 | if (slab_state < SYSFS) | |
4311 | /* Defer until later */ | |
4312 | return 0; | |
4313 | ||
4314 | unmergeable = slab_unmergeable(s); | |
4315 | if (unmergeable) { | |
4316 | /* | |
4317 | * Slabcache can never be merged so we can use the name proper. | |
4318 | * This is typically the case for debug situations. In that | |
4319 | * case we can catch duplicate names easily. | |
4320 | */ | |
4321 | sysfs_remove_link(&slab_kset->kobj, s->name); | |
4322 | name = s->name; | |
4323 | } else { | |
4324 | /* | |
4325 | * Create a unique name for the slab as a target | |
4326 | * for the symlinks. | |
4327 | */ | |
4328 | name = create_unique_id(s); | |
4329 | } | |
4330 | ||
4331 | s->kobj.kset = slab_kset; | |
4332 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); | |
4333 | if (err) { | |
4334 | kobject_put(&s->kobj); | |
4335 | return err; | |
4336 | } | |
4337 | ||
4338 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
4339 | if (err) | |
4340 | return err; | |
4341 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
4342 | if (!unmergeable) { | |
4343 | /* Setup first alias */ | |
4344 | sysfs_slab_alias(s, s->name); | |
4345 | kfree(name); | |
4346 | } | |
4347 | return 0; | |
4348 | } | |
4349 | ||
4350 | static void sysfs_slab_remove(struct kmem_cache *s) | |
4351 | { | |
4352 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
4353 | kobject_del(&s->kobj); | |
4354 | kobject_put(&s->kobj); | |
4355 | } | |
4356 | ||
4357 | /* | |
4358 | * Need to buffer aliases during bootup until sysfs becomes | |
4359 | * available lest we lose that information. | |
4360 | */ | |
4361 | struct saved_alias { | |
4362 | struct kmem_cache *s; | |
4363 | const char *name; | |
4364 | struct saved_alias *next; | |
4365 | }; | |
4366 | ||
4367 | static struct saved_alias *alias_list; | |
4368 | ||
4369 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
4370 | { | |
4371 | struct saved_alias *al; | |
4372 | ||
4373 | if (slab_state == SYSFS) { | |
4374 | /* | |
4375 | * If we have a leftover link then remove it. | |
4376 | */ | |
4377 | sysfs_remove_link(&slab_kset->kobj, name); | |
4378 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); | |
4379 | } | |
4380 | ||
4381 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
4382 | if (!al) | |
4383 | return -ENOMEM; | |
4384 | ||
4385 | al->s = s; | |
4386 | al->name = name; | |
4387 | al->next = alias_list; | |
4388 | alias_list = al; | |
4389 | return 0; | |
4390 | } | |
4391 | ||
4392 | static int __init slab_sysfs_init(void) | |
4393 | { | |
4394 | struct kmem_cache *s; | |
4395 | int err; | |
4396 | ||
4397 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); | |
4398 | if (!slab_kset) { | |
4399 | printk(KERN_ERR "Cannot register slab subsystem.\n"); | |
4400 | return -ENOSYS; | |
4401 | } | |
4402 | ||
4403 | slab_state = SYSFS; | |
4404 | ||
4405 | list_for_each_entry(s, &slab_caches, list) { | |
4406 | err = sysfs_slab_add(s); | |
4407 | if (err) | |
4408 | printk(KERN_ERR "SLUB: Unable to add boot slab %s" | |
4409 | " to sysfs\n", s->name); | |
4410 | } | |
4411 | ||
4412 | while (alias_list) { | |
4413 | struct saved_alias *al = alias_list; | |
4414 | ||
4415 | alias_list = alias_list->next; | |
4416 | err = sysfs_slab_alias(al->s, al->name); | |
4417 | if (err) | |
4418 | printk(KERN_ERR "SLUB: Unable to add boot slab alias" | |
4419 | " %s to sysfs\n", s->name); | |
4420 | kfree(al); | |
4421 | } | |
4422 | ||
4423 | resiliency_test(); | |
4424 | return 0; | |
4425 | } | |
4426 | ||
4427 | __initcall(slab_sysfs_init); | |
4428 | #endif | |
4429 | ||
4430 | /* | |
4431 | * The /proc/slabinfo ABI | |
4432 | */ | |
4433 | #ifdef CONFIG_SLABINFO | |
4434 | static void print_slabinfo_header(struct seq_file *m) | |
4435 | { | |
4436 | seq_puts(m, "slabinfo - version: 2.1\n"); | |
4437 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " | |
4438 | "<objperslab> <pagesperslab>"); | |
4439 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4440 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
4441 | seq_putc(m, '\n'); | |
4442 | } | |
4443 | ||
4444 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4445 | { | |
4446 | loff_t n = *pos; | |
4447 | ||
4448 | down_read(&slub_lock); | |
4449 | if (!n) | |
4450 | print_slabinfo_header(m); | |
4451 | ||
4452 | return seq_list_start(&slab_caches, *pos); | |
4453 | } | |
4454 | ||
4455 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4456 | { | |
4457 | return seq_list_next(p, &slab_caches, pos); | |
4458 | } | |
4459 | ||
4460 | static void s_stop(struct seq_file *m, void *p) | |
4461 | { | |
4462 | up_read(&slub_lock); | |
4463 | } | |
4464 | ||
4465 | static int s_show(struct seq_file *m, void *p) | |
4466 | { | |
4467 | unsigned long nr_partials = 0; | |
4468 | unsigned long nr_slabs = 0; | |
4469 | unsigned long nr_inuse = 0; | |
4470 | unsigned long nr_objs = 0; | |
4471 | unsigned long nr_free = 0; | |
4472 | struct kmem_cache *s; | |
4473 | int node; | |
4474 | ||
4475 | s = list_entry(p, struct kmem_cache, list); | |
4476 | ||
4477 | for_each_online_node(node) { | |
4478 | struct kmem_cache_node *n = get_node(s, node); | |
4479 | ||
4480 | if (!n) | |
4481 | continue; | |
4482 | ||
4483 | nr_partials += n->nr_partial; | |
4484 | nr_slabs += atomic_long_read(&n->nr_slabs); | |
4485 | nr_objs += atomic_long_read(&n->total_objects); | |
4486 | nr_free += count_partial(n, count_free); | |
4487 | } | |
4488 | ||
4489 | nr_inuse = nr_objs - nr_free; | |
4490 | ||
4491 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, | |
4492 | nr_objs, s->size, oo_objects(s->oo), | |
4493 | (1 << oo_order(s->oo))); | |
4494 | seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); | |
4495 | seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, | |
4496 | 0UL); | |
4497 | seq_putc(m, '\n'); | |
4498 | return 0; | |
4499 | } | |
4500 | ||
4501 | static const struct seq_operations slabinfo_op = { | |
4502 | .start = s_start, | |
4503 | .next = s_next, | |
4504 | .stop = s_stop, | |
4505 | .show = s_show, | |
4506 | }; | |
4507 | ||
4508 | static int slabinfo_open(struct inode *inode, struct file *file) | |
4509 | { | |
4510 | return seq_open(file, &slabinfo_op); | |
4511 | } | |
4512 | ||
4513 | static const struct file_operations proc_slabinfo_operations = { | |
4514 | .open = slabinfo_open, | |
4515 | .read = seq_read, | |
4516 | .llseek = seq_lseek, | |
4517 | .release = seq_release, | |
4518 | }; | |
4519 | ||
4520 | static int __init slab_proc_init(void) | |
4521 | { | |
4522 | proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); | |
4523 | return 0; | |
4524 | } | |
4525 | module_init(slab_proc_init); | |
4526 | #endif /* CONFIG_SLABINFO */ |