]>
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 <clameter@sgi.com> | |
9 | */ | |
10 | ||
11 | #include <linux/mm.h> | |
12 | #include <linux/module.h> | |
13 | #include <linux/bit_spinlock.h> | |
14 | #include <linux/interrupt.h> | |
15 | #include <linux/bitops.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/seq_file.h> | |
18 | #include <linux/cpu.h> | |
19 | #include <linux/cpuset.h> | |
20 | #include <linux/mempolicy.h> | |
21 | #include <linux/ctype.h> | |
22 | #include <linux/kallsyms.h> | |
23 | #include <linux/memory.h> | |
24 | ||
25 | /* | |
26 | * Lock order: | |
27 | * 1. slab_lock(page) | |
28 | * 2. slab->list_lock | |
29 | * | |
30 | * The slab_lock protects operations on the object of a particular | |
31 | * slab and its metadata in the page struct. If the slab lock | |
32 | * has been taken then no allocations nor frees can be performed | |
33 | * on the objects in the slab nor can the slab be added or removed | |
34 | * from the partial or full lists since this would mean modifying | |
35 | * the page_struct of the slab. | |
36 | * | |
37 | * The list_lock protects the partial and full list on each node and | |
38 | * the partial slab counter. If taken then no new slabs may be added or | |
39 | * removed from the lists nor make the number of partial slabs be modified. | |
40 | * (Note that the total number of slabs is an atomic value that may be | |
41 | * modified without taking the list lock). | |
42 | * | |
43 | * The list_lock is a centralized lock and thus we avoid taking it as | |
44 | * much as possible. As long as SLUB does not have to handle partial | |
45 | * slabs, operations can continue without any centralized lock. F.e. | |
46 | * allocating a long series of objects that fill up slabs does not require | |
47 | * the list lock. | |
48 | * | |
49 | * The lock order is sometimes inverted when we are trying to get a slab | |
50 | * off a list. We take the list_lock and then look for a page on the list | |
51 | * to use. While we do that objects in the slabs may be freed. We can | |
52 | * only operate on the slab if we have also taken the slab_lock. So we use | |
53 | * a slab_trylock() on the slab. If trylock was successful then no frees | |
54 | * can occur anymore and we can use the slab for allocations etc. If the | |
55 | * slab_trylock() does not succeed then frees are in progress in the slab and | |
56 | * we must stay away from it for a while since we may cause a bouncing | |
57 | * cacheline if we try to acquire the lock. So go onto the next slab. | |
58 | * If all pages are busy then we may allocate a new slab instead of reusing | |
59 | * a partial slab. A new slab has noone operating on it and thus there is | |
60 | * no danger of cacheline contention. | |
61 | * | |
62 | * Interrupts are disabled during allocation and deallocation in order to | |
63 | * make the slab allocator safe to use in the context of an irq. In addition | |
64 | * interrupts are disabled to ensure that the processor does not change | |
65 | * while handling per_cpu slabs, due to kernel preemption. | |
66 | * | |
67 | * SLUB assigns one slab for allocation to each processor. | |
68 | * Allocations only occur from these slabs called cpu slabs. | |
69 | * | |
70 | * Slabs with free elements are kept on a partial list and during regular | |
71 | * operations no list for full slabs is used. If an object in a full slab is | |
72 | * freed then the slab will show up again on the partial lists. | |
73 | * We track full slabs for debugging purposes though because otherwise we | |
74 | * cannot scan all objects. | |
75 | * | |
76 | * Slabs are freed when they become empty. Teardown and setup is | |
77 | * minimal so we rely on the page allocators per cpu caches for | |
78 | * fast frees and allocs. | |
79 | * | |
80 | * Overloading of page flags that are otherwise used for LRU management. | |
81 | * | |
82 | * PageActive The slab is frozen and exempt from list processing. | |
83 | * This means that the slab is dedicated to a purpose | |
84 | * such as satisfying allocations for a specific | |
85 | * processor. Objects may be freed in the slab while | |
86 | * it is frozen but slab_free will then skip the usual | |
87 | * list operations. It is up to the processor holding | |
88 | * the slab to integrate the slab into the slab lists | |
89 | * when the slab is no longer needed. | |
90 | * | |
91 | * One use of this flag is to mark slabs that are | |
92 | * used for allocations. Then such a slab becomes a cpu | |
93 | * slab. The cpu slab may be equipped with an additional | |
94 | * freelist that allows lockless access to | |
95 | * free objects in addition to the regular freelist | |
96 | * that requires the slab lock. | |
97 | * | |
98 | * PageError Slab requires special handling due to debug | |
99 | * options set. This moves slab handling out of | |
100 | * the fast path and disables lockless freelists. | |
101 | */ | |
102 | ||
103 | #define FROZEN (1 << PG_active) | |
104 | ||
105 | #ifdef CONFIG_SLUB_DEBUG | |
106 | #define SLABDEBUG (1 << PG_error) | |
107 | #else | |
108 | #define SLABDEBUG 0 | |
109 | #endif | |
110 | ||
111 | static inline int SlabFrozen(struct page *page) | |
112 | { | |
113 | return page->flags & FROZEN; | |
114 | } | |
115 | ||
116 | static inline void SetSlabFrozen(struct page *page) | |
117 | { | |
118 | page->flags |= FROZEN; | |
119 | } | |
120 | ||
121 | static inline void ClearSlabFrozen(struct page *page) | |
122 | { | |
123 | page->flags &= ~FROZEN; | |
124 | } | |
125 | ||
126 | static inline int SlabDebug(struct page *page) | |
127 | { | |
128 | return page->flags & SLABDEBUG; | |
129 | } | |
130 | ||
131 | static inline void SetSlabDebug(struct page *page) | |
132 | { | |
133 | page->flags |= SLABDEBUG; | |
134 | } | |
135 | ||
136 | static inline void ClearSlabDebug(struct page *page) | |
137 | { | |
138 | page->flags &= ~SLABDEBUG; | |
139 | } | |
140 | ||
141 | /* | |
142 | * Issues still to be resolved: | |
143 | * | |
144 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. | |
145 | * | |
146 | * - Variable sizing of the per node arrays | |
147 | */ | |
148 | ||
149 | /* Enable to test recovery from slab corruption on boot */ | |
150 | #undef SLUB_RESILIENCY_TEST | |
151 | ||
152 | #if PAGE_SHIFT <= 12 | |
153 | ||
154 | /* | |
155 | * Small page size. Make sure that we do not fragment memory | |
156 | */ | |
157 | #define DEFAULT_MAX_ORDER 1 | |
158 | #define DEFAULT_MIN_OBJECTS 4 | |
159 | ||
160 | #else | |
161 | ||
162 | /* | |
163 | * Large page machines are customarily able to handle larger | |
164 | * page orders. | |
165 | */ | |
166 | #define DEFAULT_MAX_ORDER 2 | |
167 | #define DEFAULT_MIN_OBJECTS 8 | |
168 | ||
169 | #endif | |
170 | ||
171 | /* | |
172 | * Mininum number of partial slabs. These will be left on the partial | |
173 | * lists even if they are empty. kmem_cache_shrink may reclaim them. | |
174 | */ | |
175 | #define MIN_PARTIAL 5 | |
176 | ||
177 | /* | |
178 | * Maximum number of desirable partial slabs. | |
179 | * The existence of more partial slabs makes kmem_cache_shrink | |
180 | * sort the partial list by the number of objects in the. | |
181 | */ | |
182 | #define MAX_PARTIAL 10 | |
183 | ||
184 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ | |
185 | SLAB_POISON | SLAB_STORE_USER) | |
186 | ||
187 | /* | |
188 | * Set of flags that will prevent slab merging | |
189 | */ | |
190 | #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ | |
191 | SLAB_TRACE | SLAB_DESTROY_BY_RCU) | |
192 | ||
193 | #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ | |
194 | SLAB_CACHE_DMA) | |
195 | ||
196 | #ifndef ARCH_KMALLOC_MINALIGN | |
197 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) | |
198 | #endif | |
199 | ||
200 | #ifndef ARCH_SLAB_MINALIGN | |
201 | #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) | |
202 | #endif | |
203 | ||
204 | /* Internal SLUB flags */ | |
205 | #define __OBJECT_POISON 0x80000000 /* Poison object */ | |
206 | #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ | |
207 | ||
208 | /* Not all arches define cache_line_size */ | |
209 | #ifndef cache_line_size | |
210 | #define cache_line_size() L1_CACHE_BYTES | |
211 | #endif | |
212 | ||
213 | static int kmem_size = sizeof(struct kmem_cache); | |
214 | ||
215 | #ifdef CONFIG_SMP | |
216 | static struct notifier_block slab_notifier; | |
217 | #endif | |
218 | ||
219 | static enum { | |
220 | DOWN, /* No slab functionality available */ | |
221 | PARTIAL, /* kmem_cache_open() works but kmalloc does not */ | |
222 | UP, /* Everything works but does not show up in sysfs */ | |
223 | SYSFS /* Sysfs up */ | |
224 | } slab_state = DOWN; | |
225 | ||
226 | /* A list of all slab caches on the system */ | |
227 | static DECLARE_RWSEM(slub_lock); | |
228 | static LIST_HEAD(slab_caches); | |
229 | ||
230 | /* | |
231 | * Tracking user of a slab. | |
232 | */ | |
233 | struct track { | |
234 | void *addr; /* Called from address */ | |
235 | int cpu; /* Was running on cpu */ | |
236 | int pid; /* Pid context */ | |
237 | unsigned long when; /* When did the operation occur */ | |
238 | }; | |
239 | ||
240 | enum track_item { TRACK_ALLOC, TRACK_FREE }; | |
241 | ||
242 | #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) | |
243 | static int sysfs_slab_add(struct kmem_cache *); | |
244 | static int sysfs_slab_alias(struct kmem_cache *, const char *); | |
245 | static void sysfs_slab_remove(struct kmem_cache *); | |
246 | ||
247 | #else | |
248 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } | |
249 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) | |
250 | { return 0; } | |
251 | static inline void sysfs_slab_remove(struct kmem_cache *s) | |
252 | { | |
253 | kfree(s); | |
254 | } | |
255 | ||
256 | #endif | |
257 | ||
258 | static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) | |
259 | { | |
260 | #ifdef CONFIG_SLUB_STATS | |
261 | c->stat[si]++; | |
262 | #endif | |
263 | } | |
264 | ||
265 | /******************************************************************** | |
266 | * Core slab cache functions | |
267 | *******************************************************************/ | |
268 | ||
269 | int slab_is_available(void) | |
270 | { | |
271 | return slab_state >= UP; | |
272 | } | |
273 | ||
274 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
275 | { | |
276 | #ifdef CONFIG_NUMA | |
277 | return s->node[node]; | |
278 | #else | |
279 | return &s->local_node; | |
280 | #endif | |
281 | } | |
282 | ||
283 | static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) | |
284 | { | |
285 | #ifdef CONFIG_SMP | |
286 | return s->cpu_slab[cpu]; | |
287 | #else | |
288 | return &s->cpu_slab; | |
289 | #endif | |
290 | } | |
291 | ||
292 | /* Verify that a pointer has an address that is valid within a slab page */ | |
293 | static inline int check_valid_pointer(struct kmem_cache *s, | |
294 | struct page *page, const void *object) | |
295 | { | |
296 | void *base; | |
297 | ||
298 | if (!object) | |
299 | return 1; | |
300 | ||
301 | base = page_address(page); | |
302 | if (object < base || object >= base + page->objects * s->size || | |
303 | (object - base) % s->size) { | |
304 | return 0; | |
305 | } | |
306 | ||
307 | return 1; | |
308 | } | |
309 | ||
310 | /* | |
311 | * Slow version of get and set free pointer. | |
312 | * | |
313 | * This version requires touching the cache lines of kmem_cache which | |
314 | * we avoid to do in the fast alloc free paths. There we obtain the offset | |
315 | * from the page struct. | |
316 | */ | |
317 | static inline void *get_freepointer(struct kmem_cache *s, void *object) | |
318 | { | |
319 | return *(void **)(object + s->offset); | |
320 | } | |
321 | ||
322 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
323 | { | |
324 | *(void **)(object + s->offset) = fp; | |
325 | } | |
326 | ||
327 | /* Loop over all objects in a slab */ | |
328 | #define for_each_object(__p, __s, __addr, __objects) \ | |
329 | for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ | |
330 | __p += (__s)->size) | |
331 | ||
332 | /* Scan freelist */ | |
333 | #define for_each_free_object(__p, __s, __free) \ | |
334 | for (__p = (__free); __p; __p = get_freepointer((__s), __p)) | |
335 | ||
336 | /* Determine object index from a given position */ | |
337 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) | |
338 | { | |
339 | return (p - addr) / s->size; | |
340 | } | |
341 | ||
342 | static inline struct kmem_cache_order_objects oo_make(int order, | |
343 | unsigned long size) | |
344 | { | |
345 | struct kmem_cache_order_objects x = { | |
346 | (order << 16) + (PAGE_SIZE << order) / size | |
347 | }; | |
348 | ||
349 | return x; | |
350 | } | |
351 | ||
352 | static inline int oo_order(struct kmem_cache_order_objects x) | |
353 | { | |
354 | return x.x >> 16; | |
355 | } | |
356 | ||
357 | static inline int oo_objects(struct kmem_cache_order_objects x) | |
358 | { | |
359 | return x.x & ((1 << 16) - 1); | |
360 | } | |
361 | ||
362 | #ifdef CONFIG_SLUB_DEBUG | |
363 | /* | |
364 | * Debug settings: | |
365 | */ | |
366 | #ifdef CONFIG_SLUB_DEBUG_ON | |
367 | static int slub_debug = DEBUG_DEFAULT_FLAGS; | |
368 | #else | |
369 | static int slub_debug; | |
370 | #endif | |
371 | ||
372 | static char *slub_debug_slabs; | |
373 | ||
374 | /* | |
375 | * Object debugging | |
376 | */ | |
377 | static void print_section(char *text, u8 *addr, unsigned int length) | |
378 | { | |
379 | int i, offset; | |
380 | int newline = 1; | |
381 | char ascii[17]; | |
382 | ||
383 | ascii[16] = 0; | |
384 | ||
385 | for (i = 0; i < length; i++) { | |
386 | if (newline) { | |
387 | printk(KERN_ERR "%8s 0x%p: ", text, addr + i); | |
388 | newline = 0; | |
389 | } | |
390 | printk(KERN_CONT " %02x", addr[i]); | |
391 | offset = i % 16; | |
392 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; | |
393 | if (offset == 15) { | |
394 | printk(KERN_CONT " %s\n", ascii); | |
395 | newline = 1; | |
396 | } | |
397 | } | |
398 | if (!newline) { | |
399 | i %= 16; | |
400 | while (i < 16) { | |
401 | printk(KERN_CONT " "); | |
402 | ascii[i] = ' '; | |
403 | i++; | |
404 | } | |
405 | printk(KERN_CONT " %s\n", ascii); | |
406 | } | |
407 | } | |
408 | ||
409 | static struct track *get_track(struct kmem_cache *s, void *object, | |
410 | enum track_item alloc) | |
411 | { | |
412 | struct track *p; | |
413 | ||
414 | if (s->offset) | |
415 | p = object + s->offset + sizeof(void *); | |
416 | else | |
417 | p = object + s->inuse; | |
418 | ||
419 | return p + alloc; | |
420 | } | |
421 | ||
422 | static void set_track(struct kmem_cache *s, void *object, | |
423 | enum track_item alloc, void *addr) | |
424 | { | |
425 | struct track *p; | |
426 | ||
427 | if (s->offset) | |
428 | p = object + s->offset + sizeof(void *); | |
429 | else | |
430 | p = object + s->inuse; | |
431 | ||
432 | p += alloc; | |
433 | if (addr) { | |
434 | p->addr = addr; | |
435 | p->cpu = smp_processor_id(); | |
436 | p->pid = current ? current->pid : -1; | |
437 | p->when = jiffies; | |
438 | } else | |
439 | memset(p, 0, sizeof(struct track)); | |
440 | } | |
441 | ||
442 | static void init_tracking(struct kmem_cache *s, void *object) | |
443 | { | |
444 | if (!(s->flags & SLAB_STORE_USER)) | |
445 | return; | |
446 | ||
447 | set_track(s, object, TRACK_FREE, NULL); | |
448 | set_track(s, object, TRACK_ALLOC, NULL); | |
449 | } | |
450 | ||
451 | static void print_track(const char *s, struct track *t) | |
452 | { | |
453 | if (!t->addr) | |
454 | return; | |
455 | ||
456 | printk(KERN_ERR "INFO: %s in ", s); | |
457 | __print_symbol("%s", (unsigned long)t->addr); | |
458 | printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); | |
459 | } | |
460 | ||
461 | static void print_tracking(struct kmem_cache *s, void *object) | |
462 | { | |
463 | if (!(s->flags & SLAB_STORE_USER)) | |
464 | return; | |
465 | ||
466 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); | |
467 | print_track("Freed", get_track(s, object, TRACK_FREE)); | |
468 | } | |
469 | ||
470 | static void print_page_info(struct page *page) | |
471 | { | |
472 | printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", | |
473 | page, page->objects, page->inuse, page->freelist, page->flags); | |
474 | ||
475 | } | |
476 | ||
477 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) | |
478 | { | |
479 | va_list args; | |
480 | char buf[100]; | |
481 | ||
482 | va_start(args, fmt); | |
483 | vsnprintf(buf, sizeof(buf), fmt, args); | |
484 | va_end(args); | |
485 | printk(KERN_ERR "========================================" | |
486 | "=====================================\n"); | |
487 | printk(KERN_ERR "BUG %s: %s\n", s->name, buf); | |
488 | printk(KERN_ERR "----------------------------------------" | |
489 | "-------------------------------------\n\n"); | |
490 | } | |
491 | ||
492 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) | |
493 | { | |
494 | va_list args; | |
495 | char buf[100]; | |
496 | ||
497 | va_start(args, fmt); | |
498 | vsnprintf(buf, sizeof(buf), fmt, args); | |
499 | va_end(args); | |
500 | printk(KERN_ERR "FIX %s: %s\n", s->name, buf); | |
501 | } | |
502 | ||
503 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) | |
504 | { | |
505 | unsigned int off; /* Offset of last byte */ | |
506 | u8 *addr = page_address(page); | |
507 | ||
508 | print_tracking(s, p); | |
509 | ||
510 | print_page_info(page); | |
511 | ||
512 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", | |
513 | p, p - addr, get_freepointer(s, p)); | |
514 | ||
515 | if (p > addr + 16) | |
516 | print_section("Bytes b4", p - 16, 16); | |
517 | ||
518 | print_section("Object", p, min(s->objsize, 128)); | |
519 | ||
520 | if (s->flags & SLAB_RED_ZONE) | |
521 | print_section("Redzone", p + s->objsize, | |
522 | s->inuse - s->objsize); | |
523 | ||
524 | if (s->offset) | |
525 | off = s->offset + sizeof(void *); | |
526 | else | |
527 | off = s->inuse; | |
528 | ||
529 | if (s->flags & SLAB_STORE_USER) | |
530 | off += 2 * sizeof(struct track); | |
531 | ||
532 | if (off != s->size) | |
533 | /* Beginning of the filler is the free pointer */ | |
534 | print_section("Padding", p + off, s->size - off); | |
535 | ||
536 | dump_stack(); | |
537 | } | |
538 | ||
539 | static void object_err(struct kmem_cache *s, struct page *page, | |
540 | u8 *object, char *reason) | |
541 | { | |
542 | slab_bug(s, "%s", reason); | |
543 | print_trailer(s, page, object); | |
544 | } | |
545 | ||
546 | static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) | |
547 | { | |
548 | va_list args; | |
549 | char buf[100]; | |
550 | ||
551 | va_start(args, fmt); | |
552 | vsnprintf(buf, sizeof(buf), fmt, args); | |
553 | va_end(args); | |
554 | slab_bug(s, "%s", buf); | |
555 | print_page_info(page); | |
556 | dump_stack(); | |
557 | } | |
558 | ||
559 | static void init_object(struct kmem_cache *s, void *object, int active) | |
560 | { | |
561 | u8 *p = object; | |
562 | ||
563 | if (s->flags & __OBJECT_POISON) { | |
564 | memset(p, POISON_FREE, s->objsize - 1); | |
565 | p[s->objsize - 1] = POISON_END; | |
566 | } | |
567 | ||
568 | if (s->flags & SLAB_RED_ZONE) | |
569 | memset(p + s->objsize, | |
570 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, | |
571 | s->inuse - s->objsize); | |
572 | } | |
573 | ||
574 | static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) | |
575 | { | |
576 | while (bytes) { | |
577 | if (*start != (u8)value) | |
578 | return start; | |
579 | start++; | |
580 | bytes--; | |
581 | } | |
582 | return NULL; | |
583 | } | |
584 | ||
585 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
586 | void *from, void *to) | |
587 | { | |
588 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); | |
589 | memset(from, data, to - from); | |
590 | } | |
591 | ||
592 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, | |
593 | u8 *object, char *what, | |
594 | u8 *start, unsigned int value, unsigned int bytes) | |
595 | { | |
596 | u8 *fault; | |
597 | u8 *end; | |
598 | ||
599 | fault = check_bytes(start, value, bytes); | |
600 | if (!fault) | |
601 | return 1; | |
602 | ||
603 | end = start + bytes; | |
604 | while (end > fault && end[-1] == value) | |
605 | end--; | |
606 | ||
607 | slab_bug(s, "%s overwritten", what); | |
608 | printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", | |
609 | fault, end - 1, fault[0], value); | |
610 | print_trailer(s, page, object); | |
611 | ||
612 | restore_bytes(s, what, value, fault, end); | |
613 | return 0; | |
614 | } | |
615 | ||
616 | /* | |
617 | * Object layout: | |
618 | * | |
619 | * object address | |
620 | * Bytes of the object to be managed. | |
621 | * If the freepointer may overlay the object then the free | |
622 | * pointer is the first word of the object. | |
623 | * | |
624 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is | |
625 | * 0xa5 (POISON_END) | |
626 | * | |
627 | * object + s->objsize | |
628 | * Padding to reach word boundary. This is also used for Redzoning. | |
629 | * Padding is extended by another word if Redzoning is enabled and | |
630 | * objsize == inuse. | |
631 | * | |
632 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with | |
633 | * 0xcc (RED_ACTIVE) for objects in use. | |
634 | * | |
635 | * object + s->inuse | |
636 | * Meta data starts here. | |
637 | * | |
638 | * A. Free pointer (if we cannot overwrite object on free) | |
639 | * B. Tracking data for SLAB_STORE_USER | |
640 | * C. Padding to reach required alignment boundary or at mininum | |
641 | * one word if debugging is on to be able to detect writes | |
642 | * before the word boundary. | |
643 | * | |
644 | * Padding is done using 0x5a (POISON_INUSE) | |
645 | * | |
646 | * object + s->size | |
647 | * Nothing is used beyond s->size. | |
648 | * | |
649 | * If slabcaches are merged then the objsize and inuse boundaries are mostly | |
650 | * ignored. And therefore no slab options that rely on these boundaries | |
651 | * may be used with merged slabcaches. | |
652 | */ | |
653 | ||
654 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) | |
655 | { | |
656 | unsigned long off = s->inuse; /* The end of info */ | |
657 | ||
658 | if (s->offset) | |
659 | /* Freepointer is placed after the object. */ | |
660 | off += sizeof(void *); | |
661 | ||
662 | if (s->flags & SLAB_STORE_USER) | |
663 | /* We also have user information there */ | |
664 | off += 2 * sizeof(struct track); | |
665 | ||
666 | if (s->size == off) | |
667 | return 1; | |
668 | ||
669 | return check_bytes_and_report(s, page, p, "Object padding", | |
670 | p + off, POISON_INUSE, s->size - off); | |
671 | } | |
672 | ||
673 | /* Check the pad bytes at the end of a slab page */ | |
674 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
675 | { | |
676 | u8 *start; | |
677 | u8 *fault; | |
678 | u8 *end; | |
679 | int length; | |
680 | int remainder; | |
681 | ||
682 | if (!(s->flags & SLAB_POISON)) | |
683 | return 1; | |
684 | ||
685 | start = page_address(page); | |
686 | length = (PAGE_SIZE << compound_order(page)); | |
687 | end = start + length; | |
688 | remainder = length % s->size; | |
689 | if (!remainder) | |
690 | return 1; | |
691 | ||
692 | fault = check_bytes(end - remainder, POISON_INUSE, remainder); | |
693 | if (!fault) | |
694 | return 1; | |
695 | while (end > fault && end[-1] == POISON_INUSE) | |
696 | end--; | |
697 | ||
698 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); | |
699 | print_section("Padding", end - remainder, remainder); | |
700 | ||
701 | restore_bytes(s, "slab padding", POISON_INUSE, start, end); | |
702 | return 0; | |
703 | } | |
704 | ||
705 | static int check_object(struct kmem_cache *s, struct page *page, | |
706 | void *object, int active) | |
707 | { | |
708 | u8 *p = object; | |
709 | u8 *endobject = object + s->objsize; | |
710 | ||
711 | if (s->flags & SLAB_RED_ZONE) { | |
712 | unsigned int red = | |
713 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; | |
714 | ||
715 | if (!check_bytes_and_report(s, page, object, "Redzone", | |
716 | endobject, red, s->inuse - s->objsize)) | |
717 | return 0; | |
718 | } else { | |
719 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { | |
720 | check_bytes_and_report(s, page, p, "Alignment padding", | |
721 | endobject, POISON_INUSE, s->inuse - s->objsize); | |
722 | } | |
723 | } | |
724 | ||
725 | if (s->flags & SLAB_POISON) { | |
726 | if (!active && (s->flags & __OBJECT_POISON) && | |
727 | (!check_bytes_and_report(s, page, p, "Poison", p, | |
728 | POISON_FREE, s->objsize - 1) || | |
729 | !check_bytes_and_report(s, page, p, "Poison", | |
730 | p + s->objsize - 1, POISON_END, 1))) | |
731 | return 0; | |
732 | /* | |
733 | * check_pad_bytes cleans up on its own. | |
734 | */ | |
735 | check_pad_bytes(s, page, p); | |
736 | } | |
737 | ||
738 | if (!s->offset && active) | |
739 | /* | |
740 | * Object and freepointer overlap. Cannot check | |
741 | * freepointer while object is allocated. | |
742 | */ | |
743 | return 1; | |
744 | ||
745 | /* Check free pointer validity */ | |
746 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
747 | object_err(s, page, p, "Freepointer corrupt"); | |
748 | /* | |
749 | * No choice but to zap it and thus loose the remainder | |
750 | * of the free objects in this slab. May cause | |
751 | * another error because the object count is now wrong. | |
752 | */ | |
753 | set_freepointer(s, p, NULL); | |
754 | return 0; | |
755 | } | |
756 | return 1; | |
757 | } | |
758 | ||
759 | static int check_slab(struct kmem_cache *s, struct page *page) | |
760 | { | |
761 | int maxobj; | |
762 | ||
763 | VM_BUG_ON(!irqs_disabled()); | |
764 | ||
765 | if (!PageSlab(page)) { | |
766 | slab_err(s, page, "Not a valid slab page"); | |
767 | return 0; | |
768 | } | |
769 | ||
770 | maxobj = (PAGE_SIZE << compound_order(page)) / s->size; | |
771 | if (page->objects > maxobj) { | |
772 | slab_err(s, page, "objects %u > max %u", | |
773 | s->name, page->objects, maxobj); | |
774 | return 0; | |
775 | } | |
776 | if (page->inuse > page->objects) { | |
777 | slab_err(s, page, "inuse %u > max %u", | |
778 | s->name, page->inuse, page->objects); | |
779 | return 0; | |
780 | } | |
781 | /* Slab_pad_check fixes things up after itself */ | |
782 | slab_pad_check(s, page); | |
783 | return 1; | |
784 | } | |
785 | ||
786 | /* | |
787 | * Determine if a certain object on a page is on the freelist. Must hold the | |
788 | * slab lock to guarantee that the chains are in a consistent state. | |
789 | */ | |
790 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
791 | { | |
792 | int nr = 0; | |
793 | void *fp = page->freelist; | |
794 | void *object = NULL; | |
795 | unsigned long max_objects; | |
796 | ||
797 | while (fp && nr <= page->objects) { | |
798 | if (fp == search) | |
799 | return 1; | |
800 | if (!check_valid_pointer(s, page, fp)) { | |
801 | if (object) { | |
802 | object_err(s, page, object, | |
803 | "Freechain corrupt"); | |
804 | set_freepointer(s, object, NULL); | |
805 | break; | |
806 | } else { | |
807 | slab_err(s, page, "Freepointer corrupt"); | |
808 | page->freelist = NULL; | |
809 | page->inuse = page->objects; | |
810 | slab_fix(s, "Freelist cleared"); | |
811 | return 0; | |
812 | } | |
813 | break; | |
814 | } | |
815 | object = fp; | |
816 | fp = get_freepointer(s, object); | |
817 | nr++; | |
818 | } | |
819 | ||
820 | max_objects = (PAGE_SIZE << compound_order(page)) / s->size; | |
821 | if (max_objects > 65535) | |
822 | max_objects = 65535; | |
823 | ||
824 | if (page->objects != max_objects) { | |
825 | slab_err(s, page, "Wrong number of objects. Found %d but " | |
826 | "should be %d", page->objects, max_objects); | |
827 | page->objects = max_objects; | |
828 | slab_fix(s, "Number of objects adjusted."); | |
829 | } | |
830 | if (page->inuse != page->objects - nr) { | |
831 | slab_err(s, page, "Wrong object count. Counter is %d but " | |
832 | "counted were %d", page->inuse, page->objects - nr); | |
833 | page->inuse = page->objects - nr; | |
834 | slab_fix(s, "Object count adjusted."); | |
835 | } | |
836 | return search == NULL; | |
837 | } | |
838 | ||
839 | static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc) | |
840 | { | |
841 | if (s->flags & SLAB_TRACE) { | |
842 | printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", | |
843 | s->name, | |
844 | alloc ? "alloc" : "free", | |
845 | object, page->inuse, | |
846 | page->freelist); | |
847 | ||
848 | if (!alloc) | |
849 | print_section("Object", (void *)object, s->objsize); | |
850 | ||
851 | dump_stack(); | |
852 | } | |
853 | } | |
854 | ||
855 | /* | |
856 | * Tracking of fully allocated slabs for debugging purposes. | |
857 | */ | |
858 | static void add_full(struct kmem_cache_node *n, struct page *page) | |
859 | { | |
860 | spin_lock(&n->list_lock); | |
861 | list_add(&page->lru, &n->full); | |
862 | spin_unlock(&n->list_lock); | |
863 | } | |
864 | ||
865 | static void remove_full(struct kmem_cache *s, struct page *page) | |
866 | { | |
867 | struct kmem_cache_node *n; | |
868 | ||
869 | if (!(s->flags & SLAB_STORE_USER)) | |
870 | return; | |
871 | ||
872 | n = get_node(s, page_to_nid(page)); | |
873 | ||
874 | spin_lock(&n->list_lock); | |
875 | list_del(&page->lru); | |
876 | spin_unlock(&n->list_lock); | |
877 | } | |
878 | ||
879 | /* Tracking of the number of slabs for debugging purposes */ | |
880 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
881 | { | |
882 | struct kmem_cache_node *n = get_node(s, node); | |
883 | ||
884 | return atomic_long_read(&n->nr_slabs); | |
885 | } | |
886 | ||
887 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) | |
888 | { | |
889 | struct kmem_cache_node *n = get_node(s, node); | |
890 | ||
891 | /* | |
892 | * May be called early in order to allocate a slab for the | |
893 | * kmem_cache_node structure. Solve the chicken-egg | |
894 | * dilemma by deferring the increment of the count during | |
895 | * bootstrap (see early_kmem_cache_node_alloc). | |
896 | */ | |
897 | if (!NUMA_BUILD || n) { | |
898 | atomic_long_inc(&n->nr_slabs); | |
899 | atomic_long_add(objects, &n->total_objects); | |
900 | } | |
901 | } | |
902 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) | |
903 | { | |
904 | struct kmem_cache_node *n = get_node(s, node); | |
905 | ||
906 | atomic_long_dec(&n->nr_slabs); | |
907 | atomic_long_sub(objects, &n->total_objects); | |
908 | } | |
909 | ||
910 | /* Object debug checks for alloc/free paths */ | |
911 | static void setup_object_debug(struct kmem_cache *s, struct page *page, | |
912 | void *object) | |
913 | { | |
914 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) | |
915 | return; | |
916 | ||
917 | init_object(s, object, 0); | |
918 | init_tracking(s, object); | |
919 | } | |
920 | ||
921 | static int alloc_debug_processing(struct kmem_cache *s, struct page *page, | |
922 | void *object, void *addr) | |
923 | { | |
924 | if (!check_slab(s, page)) | |
925 | goto bad; | |
926 | ||
927 | if (!on_freelist(s, page, object)) { | |
928 | object_err(s, page, object, "Object already allocated"); | |
929 | goto bad; | |
930 | } | |
931 | ||
932 | if (!check_valid_pointer(s, page, object)) { | |
933 | object_err(s, page, object, "Freelist Pointer check fails"); | |
934 | goto bad; | |
935 | } | |
936 | ||
937 | if (!check_object(s, page, object, 0)) | |
938 | goto bad; | |
939 | ||
940 | /* Success perform special debug activities for allocs */ | |
941 | if (s->flags & SLAB_STORE_USER) | |
942 | set_track(s, object, TRACK_ALLOC, addr); | |
943 | trace(s, page, object, 1); | |
944 | init_object(s, object, 1); | |
945 | return 1; | |
946 | ||
947 | bad: | |
948 | if (PageSlab(page)) { | |
949 | /* | |
950 | * If this is a slab page then lets do the best we can | |
951 | * to avoid issues in the future. Marking all objects | |
952 | * as used avoids touching the remaining objects. | |
953 | */ | |
954 | slab_fix(s, "Marking all objects used"); | |
955 | page->inuse = page->objects; | |
956 | page->freelist = NULL; | |
957 | } | |
958 | return 0; | |
959 | } | |
960 | ||
961 | static int free_debug_processing(struct kmem_cache *s, struct page *page, | |
962 | void *object, void *addr) | |
963 | { | |
964 | if (!check_slab(s, page)) | |
965 | goto fail; | |
966 | ||
967 | if (!check_valid_pointer(s, page, object)) { | |
968 | slab_err(s, page, "Invalid object pointer 0x%p", object); | |
969 | goto fail; | |
970 | } | |
971 | ||
972 | if (on_freelist(s, page, object)) { | |
973 | object_err(s, page, object, "Object already free"); | |
974 | goto fail; | |
975 | } | |
976 | ||
977 | if (!check_object(s, page, object, 1)) | |
978 | return 0; | |
979 | ||
980 | if (unlikely(s != page->slab)) { | |
981 | if (!PageSlab(page)) { | |
982 | slab_err(s, page, "Attempt to free object(0x%p) " | |
983 | "outside of slab", object); | |
984 | } else if (!page->slab) { | |
985 | printk(KERN_ERR | |
986 | "SLUB <none>: no slab for object 0x%p.\n", | |
987 | object); | |
988 | dump_stack(); | |
989 | } else | |
990 | object_err(s, page, object, | |
991 | "page slab pointer corrupt."); | |
992 | goto fail; | |
993 | } | |
994 | ||
995 | /* Special debug activities for freeing objects */ | |
996 | if (!SlabFrozen(page) && !page->freelist) | |
997 | remove_full(s, page); | |
998 | if (s->flags & SLAB_STORE_USER) | |
999 | set_track(s, object, TRACK_FREE, addr); | |
1000 | trace(s, page, object, 0); | |
1001 | init_object(s, object, 0); | |
1002 | return 1; | |
1003 | ||
1004 | fail: | |
1005 | slab_fix(s, "Object at 0x%p not freed", object); | |
1006 | return 0; | |
1007 | } | |
1008 | ||
1009 | static int __init setup_slub_debug(char *str) | |
1010 | { | |
1011 | slub_debug = DEBUG_DEFAULT_FLAGS; | |
1012 | if (*str++ != '=' || !*str) | |
1013 | /* | |
1014 | * No options specified. Switch on full debugging. | |
1015 | */ | |
1016 | goto out; | |
1017 | ||
1018 | if (*str == ',') | |
1019 | /* | |
1020 | * No options but restriction on slabs. This means full | |
1021 | * debugging for slabs matching a pattern. | |
1022 | */ | |
1023 | goto check_slabs; | |
1024 | ||
1025 | slub_debug = 0; | |
1026 | if (*str == '-') | |
1027 | /* | |
1028 | * Switch off all debugging measures. | |
1029 | */ | |
1030 | goto out; | |
1031 | ||
1032 | /* | |
1033 | * Determine which debug features should be switched on | |
1034 | */ | |
1035 | for (; *str && *str != ','; str++) { | |
1036 | switch (tolower(*str)) { | |
1037 | case 'f': | |
1038 | slub_debug |= SLAB_DEBUG_FREE; | |
1039 | break; | |
1040 | case 'z': | |
1041 | slub_debug |= SLAB_RED_ZONE; | |
1042 | break; | |
1043 | case 'p': | |
1044 | slub_debug |= SLAB_POISON; | |
1045 | break; | |
1046 | case 'u': | |
1047 | slub_debug |= SLAB_STORE_USER; | |
1048 | break; | |
1049 | case 't': | |
1050 | slub_debug |= SLAB_TRACE; | |
1051 | break; | |
1052 | default: | |
1053 | printk(KERN_ERR "slub_debug option '%c' " | |
1054 | "unknown. skipped\n", *str); | |
1055 | } | |
1056 | } | |
1057 | ||
1058 | check_slabs: | |
1059 | if (*str == ',') | |
1060 | slub_debug_slabs = str + 1; | |
1061 | out: | |
1062 | return 1; | |
1063 | } | |
1064 | ||
1065 | __setup("slub_debug", setup_slub_debug); | |
1066 | ||
1067 | static unsigned long kmem_cache_flags(unsigned long objsize, | |
1068 | unsigned long flags, const char *name, | |
1069 | void (*ctor)(struct kmem_cache *, void *)) | |
1070 | { | |
1071 | /* | |
1072 | * Enable debugging if selected on the kernel commandline. | |
1073 | */ | |
1074 | if (slub_debug && (!slub_debug_slabs || | |
1075 | strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0)) | |
1076 | flags |= slub_debug; | |
1077 | ||
1078 | return flags; | |
1079 | } | |
1080 | #else | |
1081 | static inline void setup_object_debug(struct kmem_cache *s, | |
1082 | struct page *page, void *object) {} | |
1083 | ||
1084 | static inline int alloc_debug_processing(struct kmem_cache *s, | |
1085 | struct page *page, void *object, void *addr) { return 0; } | |
1086 | ||
1087 | static inline int free_debug_processing(struct kmem_cache *s, | |
1088 | struct page *page, void *object, void *addr) { return 0; } | |
1089 | ||
1090 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) | |
1091 | { return 1; } | |
1092 | static inline int check_object(struct kmem_cache *s, struct page *page, | |
1093 | void *object, int active) { return 1; } | |
1094 | static inline void add_full(struct kmem_cache_node *n, struct page *page) {} | |
1095 | static inline unsigned long kmem_cache_flags(unsigned long objsize, | |
1096 | unsigned long flags, const char *name, | |
1097 | void (*ctor)(struct kmem_cache *, void *)) | |
1098 | { | |
1099 | return flags; | |
1100 | } | |
1101 | #define slub_debug 0 | |
1102 | ||
1103 | static inline unsigned long slabs_node(struct kmem_cache *s, int node) | |
1104 | { return 0; } | |
1105 | static inline void inc_slabs_node(struct kmem_cache *s, int node, | |
1106 | int objects) {} | |
1107 | static inline void dec_slabs_node(struct kmem_cache *s, int node, | |
1108 | int objects) {} | |
1109 | #endif | |
1110 | ||
1111 | /* | |
1112 | * Slab allocation and freeing | |
1113 | */ | |
1114 | static inline struct page *alloc_slab_page(gfp_t flags, int node, | |
1115 | struct kmem_cache_order_objects oo) | |
1116 | { | |
1117 | int order = oo_order(oo); | |
1118 | ||
1119 | if (node == -1) | |
1120 | return alloc_pages(flags, order); | |
1121 | else | |
1122 | return alloc_pages_node(node, flags, order); | |
1123 | } | |
1124 | ||
1125 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1126 | { | |
1127 | struct page *page; | |
1128 | struct kmem_cache_order_objects oo = s->oo; | |
1129 | ||
1130 | flags |= s->allocflags; | |
1131 | ||
1132 | page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node, | |
1133 | oo); | |
1134 | if (unlikely(!page)) { | |
1135 | oo = s->min; | |
1136 | /* | |
1137 | * Allocation may have failed due to fragmentation. | |
1138 | * Try a lower order alloc if possible | |
1139 | */ | |
1140 | page = alloc_slab_page(flags, node, oo); | |
1141 | if (!page) | |
1142 | return NULL; | |
1143 | ||
1144 | stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK); | |
1145 | } | |
1146 | page->objects = oo_objects(oo); | |
1147 | mod_zone_page_state(page_zone(page), | |
1148 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1149 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1150 | 1 << oo_order(oo)); | |
1151 | ||
1152 | return page; | |
1153 | } | |
1154 | ||
1155 | static void setup_object(struct kmem_cache *s, struct page *page, | |
1156 | void *object) | |
1157 | { | |
1158 | setup_object_debug(s, page, object); | |
1159 | if (unlikely(s->ctor)) | |
1160 | s->ctor(s, object); | |
1161 | } | |
1162 | ||
1163 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1164 | { | |
1165 | struct page *page; | |
1166 | void *start; | |
1167 | void *last; | |
1168 | void *p; | |
1169 | ||
1170 | BUG_ON(flags & GFP_SLAB_BUG_MASK); | |
1171 | ||
1172 | page = allocate_slab(s, | |
1173 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); | |
1174 | if (!page) | |
1175 | goto out; | |
1176 | ||
1177 | inc_slabs_node(s, page_to_nid(page), page->objects); | |
1178 | page->slab = s; | |
1179 | page->flags |= 1 << PG_slab; | |
1180 | if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | | |
1181 | SLAB_STORE_USER | SLAB_TRACE)) | |
1182 | SetSlabDebug(page); | |
1183 | ||
1184 | start = page_address(page); | |
1185 | ||
1186 | if (unlikely(s->flags & SLAB_POISON)) | |
1187 | memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); | |
1188 | ||
1189 | last = start; | |
1190 | for_each_object(p, s, start, page->objects) { | |
1191 | setup_object(s, page, last); | |
1192 | set_freepointer(s, last, p); | |
1193 | last = p; | |
1194 | } | |
1195 | setup_object(s, page, last); | |
1196 | set_freepointer(s, last, NULL); | |
1197 | ||
1198 | page->freelist = start; | |
1199 | page->inuse = 0; | |
1200 | out: | |
1201 | return page; | |
1202 | } | |
1203 | ||
1204 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
1205 | { | |
1206 | int order = compound_order(page); | |
1207 | int pages = 1 << order; | |
1208 | ||
1209 | if (unlikely(SlabDebug(page))) { | |
1210 | void *p; | |
1211 | ||
1212 | slab_pad_check(s, page); | |
1213 | for_each_object(p, s, page_address(page), | |
1214 | page->objects) | |
1215 | check_object(s, page, p, 0); | |
1216 | ClearSlabDebug(page); | |
1217 | } | |
1218 | ||
1219 | mod_zone_page_state(page_zone(page), | |
1220 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1221 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1222 | -pages); | |
1223 | ||
1224 | __ClearPageSlab(page); | |
1225 | reset_page_mapcount(page); | |
1226 | __free_pages(page, order); | |
1227 | } | |
1228 | ||
1229 | static void rcu_free_slab(struct rcu_head *h) | |
1230 | { | |
1231 | struct page *page; | |
1232 | ||
1233 | page = container_of((struct list_head *)h, struct page, lru); | |
1234 | __free_slab(page->slab, page); | |
1235 | } | |
1236 | ||
1237 | static void free_slab(struct kmem_cache *s, struct page *page) | |
1238 | { | |
1239 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
1240 | /* | |
1241 | * RCU free overloads the RCU head over the LRU | |
1242 | */ | |
1243 | struct rcu_head *head = (void *)&page->lru; | |
1244 | ||
1245 | call_rcu(head, rcu_free_slab); | |
1246 | } else | |
1247 | __free_slab(s, page); | |
1248 | } | |
1249 | ||
1250 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
1251 | { | |
1252 | dec_slabs_node(s, page_to_nid(page), page->objects); | |
1253 | free_slab(s, page); | |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * Per slab locking using the pagelock | |
1258 | */ | |
1259 | static __always_inline void slab_lock(struct page *page) | |
1260 | { | |
1261 | bit_spin_lock(PG_locked, &page->flags); | |
1262 | } | |
1263 | ||
1264 | static __always_inline void slab_unlock(struct page *page) | |
1265 | { | |
1266 | __bit_spin_unlock(PG_locked, &page->flags); | |
1267 | } | |
1268 | ||
1269 | static __always_inline int slab_trylock(struct page *page) | |
1270 | { | |
1271 | int rc = 1; | |
1272 | ||
1273 | rc = bit_spin_trylock(PG_locked, &page->flags); | |
1274 | return rc; | |
1275 | } | |
1276 | ||
1277 | /* | |
1278 | * Management of partially allocated slabs | |
1279 | */ | |
1280 | static void add_partial(struct kmem_cache_node *n, | |
1281 | struct page *page, int tail) | |
1282 | { | |
1283 | spin_lock(&n->list_lock); | |
1284 | n->nr_partial++; | |
1285 | if (tail) | |
1286 | list_add_tail(&page->lru, &n->partial); | |
1287 | else | |
1288 | list_add(&page->lru, &n->partial); | |
1289 | spin_unlock(&n->list_lock); | |
1290 | } | |
1291 | ||
1292 | static void remove_partial(struct kmem_cache *s, | |
1293 | struct page *page) | |
1294 | { | |
1295 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1296 | ||
1297 | spin_lock(&n->list_lock); | |
1298 | list_del(&page->lru); | |
1299 | n->nr_partial--; | |
1300 | spin_unlock(&n->list_lock); | |
1301 | } | |
1302 | ||
1303 | /* | |
1304 | * Lock slab and remove from the partial list. | |
1305 | * | |
1306 | * Must hold list_lock. | |
1307 | */ | |
1308 | static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page) | |
1309 | { | |
1310 | if (slab_trylock(page)) { | |
1311 | list_del(&page->lru); | |
1312 | n->nr_partial--; | |
1313 | SetSlabFrozen(page); | |
1314 | return 1; | |
1315 | } | |
1316 | return 0; | |
1317 | } | |
1318 | ||
1319 | /* | |
1320 | * Try to allocate a partial slab from a specific node. | |
1321 | */ | |
1322 | static struct page *get_partial_node(struct kmem_cache_node *n) | |
1323 | { | |
1324 | struct page *page; | |
1325 | ||
1326 | /* | |
1327 | * Racy check. If we mistakenly see no partial slabs then we | |
1328 | * just allocate an empty slab. If we mistakenly try to get a | |
1329 | * partial slab and there is none available then get_partials() | |
1330 | * will return NULL. | |
1331 | */ | |
1332 | if (!n || !n->nr_partial) | |
1333 | return NULL; | |
1334 | ||
1335 | spin_lock(&n->list_lock); | |
1336 | list_for_each_entry(page, &n->partial, lru) | |
1337 | if (lock_and_freeze_slab(n, page)) | |
1338 | goto out; | |
1339 | page = NULL; | |
1340 | out: | |
1341 | spin_unlock(&n->list_lock); | |
1342 | return page; | |
1343 | } | |
1344 | ||
1345 | /* | |
1346 | * Get a page from somewhere. Search in increasing NUMA distances. | |
1347 | */ | |
1348 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) | |
1349 | { | |
1350 | #ifdef CONFIG_NUMA | |
1351 | struct zonelist *zonelist; | |
1352 | struct zone **z; | |
1353 | struct page *page; | |
1354 | ||
1355 | /* | |
1356 | * The defrag ratio allows a configuration of the tradeoffs between | |
1357 | * inter node defragmentation and node local allocations. A lower | |
1358 | * defrag_ratio increases the tendency to do local allocations | |
1359 | * instead of attempting to obtain partial slabs from other nodes. | |
1360 | * | |
1361 | * If the defrag_ratio is set to 0 then kmalloc() always | |
1362 | * returns node local objects. If the ratio is higher then kmalloc() | |
1363 | * may return off node objects because partial slabs are obtained | |
1364 | * from other nodes and filled up. | |
1365 | * | |
1366 | * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes | |
1367 | * defrag_ratio = 1000) then every (well almost) allocation will | |
1368 | * first attempt to defrag slab caches on other nodes. This means | |
1369 | * scanning over all nodes to look for partial slabs which may be | |
1370 | * expensive if we do it every time we are trying to find a slab | |
1371 | * with available objects. | |
1372 | */ | |
1373 | if (!s->remote_node_defrag_ratio || | |
1374 | get_cycles() % 1024 > s->remote_node_defrag_ratio) | |
1375 | return NULL; | |
1376 | ||
1377 | zonelist = &NODE_DATA( | |
1378 | slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)]; | |
1379 | for (z = zonelist->zones; *z; z++) { | |
1380 | struct kmem_cache_node *n; | |
1381 | ||
1382 | n = get_node(s, zone_to_nid(*z)); | |
1383 | ||
1384 | if (n && cpuset_zone_allowed_hardwall(*z, flags) && | |
1385 | n->nr_partial > MIN_PARTIAL) { | |
1386 | page = get_partial_node(n); | |
1387 | if (page) | |
1388 | return page; | |
1389 | } | |
1390 | } | |
1391 | #endif | |
1392 | return NULL; | |
1393 | } | |
1394 | ||
1395 | /* | |
1396 | * Get a partial page, lock it and return it. | |
1397 | */ | |
1398 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) | |
1399 | { | |
1400 | struct page *page; | |
1401 | int searchnode = (node == -1) ? numa_node_id() : node; | |
1402 | ||
1403 | page = get_partial_node(get_node(s, searchnode)); | |
1404 | if (page || (flags & __GFP_THISNODE)) | |
1405 | return page; | |
1406 | ||
1407 | return get_any_partial(s, flags); | |
1408 | } | |
1409 | ||
1410 | /* | |
1411 | * Move a page back to the lists. | |
1412 | * | |
1413 | * Must be called with the slab lock held. | |
1414 | * | |
1415 | * On exit the slab lock will have been dropped. | |
1416 | */ | |
1417 | static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) | |
1418 | { | |
1419 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1420 | struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id()); | |
1421 | ||
1422 | ClearSlabFrozen(page); | |
1423 | if (page->inuse) { | |
1424 | ||
1425 | if (page->freelist) { | |
1426 | add_partial(n, page, tail); | |
1427 | stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); | |
1428 | } else { | |
1429 | stat(c, DEACTIVATE_FULL); | |
1430 | if (SlabDebug(page) && (s->flags & SLAB_STORE_USER)) | |
1431 | add_full(n, page); | |
1432 | } | |
1433 | slab_unlock(page); | |
1434 | } else { | |
1435 | stat(c, DEACTIVATE_EMPTY); | |
1436 | if (n->nr_partial < MIN_PARTIAL) { | |
1437 | /* | |
1438 | * Adding an empty slab to the partial slabs in order | |
1439 | * to avoid page allocator overhead. This slab needs | |
1440 | * to come after the other slabs with objects in | |
1441 | * so that the others get filled first. That way the | |
1442 | * size of the partial list stays small. | |
1443 | * | |
1444 | * kmem_cache_shrink can reclaim any empty slabs from the | |
1445 | * partial list. | |
1446 | */ | |
1447 | add_partial(n, page, 1); | |
1448 | slab_unlock(page); | |
1449 | } else { | |
1450 | slab_unlock(page); | |
1451 | stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB); | |
1452 | discard_slab(s, page); | |
1453 | } | |
1454 | } | |
1455 | } | |
1456 | ||
1457 | /* | |
1458 | * Remove the cpu slab | |
1459 | */ | |
1460 | static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) | |
1461 | { | |
1462 | struct page *page = c->page; | |
1463 | int tail = 1; | |
1464 | ||
1465 | if (page->freelist) | |
1466 | stat(c, DEACTIVATE_REMOTE_FREES); | |
1467 | /* | |
1468 | * Merge cpu freelist into slab freelist. Typically we get here | |
1469 | * because both freelists are empty. So this is unlikely | |
1470 | * to occur. | |
1471 | */ | |
1472 | while (unlikely(c->freelist)) { | |
1473 | void **object; | |
1474 | ||
1475 | tail = 0; /* Hot objects. Put the slab first */ | |
1476 | ||
1477 | /* Retrieve object from cpu_freelist */ | |
1478 | object = c->freelist; | |
1479 | c->freelist = c->freelist[c->offset]; | |
1480 | ||
1481 | /* And put onto the regular freelist */ | |
1482 | object[c->offset] = page->freelist; | |
1483 | page->freelist = object; | |
1484 | page->inuse--; | |
1485 | } | |
1486 | c->page = NULL; | |
1487 | unfreeze_slab(s, page, tail); | |
1488 | } | |
1489 | ||
1490 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) | |
1491 | { | |
1492 | stat(c, CPUSLAB_FLUSH); | |
1493 | slab_lock(c->page); | |
1494 | deactivate_slab(s, c); | |
1495 | } | |
1496 | ||
1497 | /* | |
1498 | * Flush cpu slab. | |
1499 | * | |
1500 | * Called from IPI handler with interrupts disabled. | |
1501 | */ | |
1502 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) | |
1503 | { | |
1504 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
1505 | ||
1506 | if (likely(c && c->page)) | |
1507 | flush_slab(s, c); | |
1508 | } | |
1509 | ||
1510 | static void flush_cpu_slab(void *d) | |
1511 | { | |
1512 | struct kmem_cache *s = d; | |
1513 | ||
1514 | __flush_cpu_slab(s, smp_processor_id()); | |
1515 | } | |
1516 | ||
1517 | static void flush_all(struct kmem_cache *s) | |
1518 | { | |
1519 | #ifdef CONFIG_SMP | |
1520 | on_each_cpu(flush_cpu_slab, s, 1, 1); | |
1521 | #else | |
1522 | unsigned long flags; | |
1523 | ||
1524 | local_irq_save(flags); | |
1525 | flush_cpu_slab(s); | |
1526 | local_irq_restore(flags); | |
1527 | #endif | |
1528 | } | |
1529 | ||
1530 | /* | |
1531 | * Check if the objects in a per cpu structure fit numa | |
1532 | * locality expectations. | |
1533 | */ | |
1534 | static inline int node_match(struct kmem_cache_cpu *c, int node) | |
1535 | { | |
1536 | #ifdef CONFIG_NUMA | |
1537 | if (node != -1 && c->node != node) | |
1538 | return 0; | |
1539 | #endif | |
1540 | return 1; | |
1541 | } | |
1542 | ||
1543 | /* | |
1544 | * Slow path. The lockless freelist is empty or we need to perform | |
1545 | * debugging duties. | |
1546 | * | |
1547 | * Interrupts are disabled. | |
1548 | * | |
1549 | * Processing is still very fast if new objects have been freed to the | |
1550 | * regular freelist. In that case we simply take over the regular freelist | |
1551 | * as the lockless freelist and zap the regular freelist. | |
1552 | * | |
1553 | * If that is not working then we fall back to the partial lists. We take the | |
1554 | * first element of the freelist as the object to allocate now and move the | |
1555 | * rest of the freelist to the lockless freelist. | |
1556 | * | |
1557 | * And if we were unable to get a new slab from the partial slab lists then | |
1558 | * we need to allocate a new slab. This is the slowest path since it involves | |
1559 | * a call to the page allocator and the setup of a new slab. | |
1560 | */ | |
1561 | static void *__slab_alloc(struct kmem_cache *s, | |
1562 | gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c) | |
1563 | { | |
1564 | void **object; | |
1565 | struct page *new; | |
1566 | ||
1567 | /* We handle __GFP_ZERO in the caller */ | |
1568 | gfpflags &= ~__GFP_ZERO; | |
1569 | ||
1570 | if (!c->page) | |
1571 | goto new_slab; | |
1572 | ||
1573 | slab_lock(c->page); | |
1574 | if (unlikely(!node_match(c, node))) | |
1575 | goto another_slab; | |
1576 | ||
1577 | stat(c, ALLOC_REFILL); | |
1578 | ||
1579 | load_freelist: | |
1580 | object = c->page->freelist; | |
1581 | if (unlikely(!object)) | |
1582 | goto another_slab; | |
1583 | if (unlikely(SlabDebug(c->page))) | |
1584 | goto debug; | |
1585 | ||
1586 | c->freelist = object[c->offset]; | |
1587 | c->page->inuse = c->page->objects; | |
1588 | c->page->freelist = NULL; | |
1589 | c->node = page_to_nid(c->page); | |
1590 | unlock_out: | |
1591 | slab_unlock(c->page); | |
1592 | stat(c, ALLOC_SLOWPATH); | |
1593 | return object; | |
1594 | ||
1595 | another_slab: | |
1596 | deactivate_slab(s, c); | |
1597 | ||
1598 | new_slab: | |
1599 | new = get_partial(s, gfpflags, node); | |
1600 | if (new) { | |
1601 | c->page = new; | |
1602 | stat(c, ALLOC_FROM_PARTIAL); | |
1603 | goto load_freelist; | |
1604 | } | |
1605 | ||
1606 | if (gfpflags & __GFP_WAIT) | |
1607 | local_irq_enable(); | |
1608 | ||
1609 | new = new_slab(s, gfpflags, node); | |
1610 | ||
1611 | if (gfpflags & __GFP_WAIT) | |
1612 | local_irq_disable(); | |
1613 | ||
1614 | if (new) { | |
1615 | c = get_cpu_slab(s, smp_processor_id()); | |
1616 | stat(c, ALLOC_SLAB); | |
1617 | if (c->page) | |
1618 | flush_slab(s, c); | |
1619 | slab_lock(new); | |
1620 | SetSlabFrozen(new); | |
1621 | c->page = new; | |
1622 | goto load_freelist; | |
1623 | } | |
1624 | return NULL; | |
1625 | debug: | |
1626 | if (!alloc_debug_processing(s, c->page, object, addr)) | |
1627 | goto another_slab; | |
1628 | ||
1629 | c->page->inuse++; | |
1630 | c->page->freelist = object[c->offset]; | |
1631 | c->node = -1; | |
1632 | goto unlock_out; | |
1633 | } | |
1634 | ||
1635 | /* | |
1636 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) | |
1637 | * have the fastpath folded into their functions. So no function call | |
1638 | * overhead for requests that can be satisfied on the fastpath. | |
1639 | * | |
1640 | * The fastpath works by first checking if the lockless freelist can be used. | |
1641 | * If not then __slab_alloc is called for slow processing. | |
1642 | * | |
1643 | * Otherwise we can simply pick the next object from the lockless free list. | |
1644 | */ | |
1645 | static __always_inline void *slab_alloc(struct kmem_cache *s, | |
1646 | gfp_t gfpflags, int node, void *addr) | |
1647 | { | |
1648 | void **object; | |
1649 | struct kmem_cache_cpu *c; | |
1650 | unsigned long flags; | |
1651 | ||
1652 | local_irq_save(flags); | |
1653 | c = get_cpu_slab(s, smp_processor_id()); | |
1654 | if (unlikely(!c->freelist || !node_match(c, node))) | |
1655 | ||
1656 | object = __slab_alloc(s, gfpflags, node, addr, c); | |
1657 | ||
1658 | else { | |
1659 | object = c->freelist; | |
1660 | c->freelist = object[c->offset]; | |
1661 | stat(c, ALLOC_FASTPATH); | |
1662 | } | |
1663 | local_irq_restore(flags); | |
1664 | ||
1665 | if (unlikely((gfpflags & __GFP_ZERO) && object)) | |
1666 | memset(object, 0, c->objsize); | |
1667 | ||
1668 | return object; | |
1669 | } | |
1670 | ||
1671 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
1672 | { | |
1673 | return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); | |
1674 | } | |
1675 | EXPORT_SYMBOL(kmem_cache_alloc); | |
1676 | ||
1677 | #ifdef CONFIG_NUMA | |
1678 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
1679 | { | |
1680 | return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); | |
1681 | } | |
1682 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
1683 | #endif | |
1684 | ||
1685 | /* | |
1686 | * Slow patch handling. This may still be called frequently since objects | |
1687 | * have a longer lifetime than the cpu slabs in most processing loads. | |
1688 | * | |
1689 | * So we still attempt to reduce cache line usage. Just take the slab | |
1690 | * lock and free the item. If there is no additional partial page | |
1691 | * handling required then we can return immediately. | |
1692 | */ | |
1693 | static void __slab_free(struct kmem_cache *s, struct page *page, | |
1694 | void *x, void *addr, unsigned int offset) | |
1695 | { | |
1696 | void *prior; | |
1697 | void **object = (void *)x; | |
1698 | struct kmem_cache_cpu *c; | |
1699 | ||
1700 | c = get_cpu_slab(s, raw_smp_processor_id()); | |
1701 | stat(c, FREE_SLOWPATH); | |
1702 | slab_lock(page); | |
1703 | ||
1704 | if (unlikely(SlabDebug(page))) | |
1705 | goto debug; | |
1706 | ||
1707 | checks_ok: | |
1708 | prior = object[offset] = page->freelist; | |
1709 | page->freelist = object; | |
1710 | page->inuse--; | |
1711 | ||
1712 | if (unlikely(SlabFrozen(page))) { | |
1713 | stat(c, FREE_FROZEN); | |
1714 | goto out_unlock; | |
1715 | } | |
1716 | ||
1717 | if (unlikely(!page->inuse)) | |
1718 | goto slab_empty; | |
1719 | ||
1720 | /* | |
1721 | * Objects left in the slab. If it was not on the partial list before | |
1722 | * then add it. | |
1723 | */ | |
1724 | if (unlikely(!prior)) { | |
1725 | add_partial(get_node(s, page_to_nid(page)), page, 1); | |
1726 | stat(c, FREE_ADD_PARTIAL); | |
1727 | } | |
1728 | ||
1729 | out_unlock: | |
1730 | slab_unlock(page); | |
1731 | return; | |
1732 | ||
1733 | slab_empty: | |
1734 | if (prior) { | |
1735 | /* | |
1736 | * Slab still on the partial list. | |
1737 | */ | |
1738 | remove_partial(s, page); | |
1739 | stat(c, FREE_REMOVE_PARTIAL); | |
1740 | } | |
1741 | slab_unlock(page); | |
1742 | stat(c, FREE_SLAB); | |
1743 | discard_slab(s, page); | |
1744 | return; | |
1745 | ||
1746 | debug: | |
1747 | if (!free_debug_processing(s, page, x, addr)) | |
1748 | goto out_unlock; | |
1749 | goto checks_ok; | |
1750 | } | |
1751 | ||
1752 | /* | |
1753 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that | |
1754 | * can perform fastpath freeing without additional function calls. | |
1755 | * | |
1756 | * The fastpath is only possible if we are freeing to the current cpu slab | |
1757 | * of this processor. This typically the case if we have just allocated | |
1758 | * the item before. | |
1759 | * | |
1760 | * If fastpath is not possible then fall back to __slab_free where we deal | |
1761 | * with all sorts of special processing. | |
1762 | */ | |
1763 | static __always_inline void slab_free(struct kmem_cache *s, | |
1764 | struct page *page, void *x, void *addr) | |
1765 | { | |
1766 | void **object = (void *)x; | |
1767 | struct kmem_cache_cpu *c; | |
1768 | unsigned long flags; | |
1769 | ||
1770 | local_irq_save(flags); | |
1771 | c = get_cpu_slab(s, smp_processor_id()); | |
1772 | debug_check_no_locks_freed(object, c->objsize); | |
1773 | if (likely(page == c->page && c->node >= 0)) { | |
1774 | object[c->offset] = c->freelist; | |
1775 | c->freelist = object; | |
1776 | stat(c, FREE_FASTPATH); | |
1777 | } else | |
1778 | __slab_free(s, page, x, addr, c->offset); | |
1779 | ||
1780 | local_irq_restore(flags); | |
1781 | } | |
1782 | ||
1783 | void kmem_cache_free(struct kmem_cache *s, void *x) | |
1784 | { | |
1785 | struct page *page; | |
1786 | ||
1787 | page = virt_to_head_page(x); | |
1788 | ||
1789 | slab_free(s, page, x, __builtin_return_address(0)); | |
1790 | } | |
1791 | EXPORT_SYMBOL(kmem_cache_free); | |
1792 | ||
1793 | /* Figure out on which slab object the object resides */ | |
1794 | static struct page *get_object_page(const void *x) | |
1795 | { | |
1796 | struct page *page = virt_to_head_page(x); | |
1797 | ||
1798 | if (!PageSlab(page)) | |
1799 | return NULL; | |
1800 | ||
1801 | return page; | |
1802 | } | |
1803 | ||
1804 | /* | |
1805 | * Object placement in a slab is made very easy because we always start at | |
1806 | * offset 0. If we tune the size of the object to the alignment then we can | |
1807 | * get the required alignment by putting one properly sized object after | |
1808 | * another. | |
1809 | * | |
1810 | * Notice that the allocation order determines the sizes of the per cpu | |
1811 | * caches. Each processor has always one slab available for allocations. | |
1812 | * Increasing the allocation order reduces the number of times that slabs | |
1813 | * must be moved on and off the partial lists and is therefore a factor in | |
1814 | * locking overhead. | |
1815 | */ | |
1816 | ||
1817 | /* | |
1818 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
1819 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
1820 | * and increases the number of allocations possible without having to | |
1821 | * take the list_lock. | |
1822 | */ | |
1823 | static int slub_min_order; | |
1824 | static int slub_max_order = DEFAULT_MAX_ORDER; | |
1825 | static int slub_min_objects = DEFAULT_MIN_OBJECTS; | |
1826 | ||
1827 | /* | |
1828 | * Merge control. If this is set then no merging of slab caches will occur. | |
1829 | * (Could be removed. This was introduced to pacify the merge skeptics.) | |
1830 | */ | |
1831 | static int slub_nomerge; | |
1832 | ||
1833 | /* | |
1834 | * Calculate the order of allocation given an slab object size. | |
1835 | * | |
1836 | * The order of allocation has significant impact on performance and other | |
1837 | * system components. Generally order 0 allocations should be preferred since | |
1838 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
1839 | * be problematic to put into order 0 slabs because there may be too much | |
1840 | * unused space left. We go to a higher order if more than 1/8th of the slab | |
1841 | * would be wasted. | |
1842 | * | |
1843 | * In order to reach satisfactory performance we must ensure that a minimum | |
1844 | * number of objects is in one slab. Otherwise we may generate too much | |
1845 | * activity on the partial lists which requires taking the list_lock. This is | |
1846 | * less a concern for large slabs though which are rarely used. | |
1847 | * | |
1848 | * slub_max_order specifies the order where we begin to stop considering the | |
1849 | * number of objects in a slab as critical. If we reach slub_max_order then | |
1850 | * we try to keep the page order as low as possible. So we accept more waste | |
1851 | * of space in favor of a small page order. | |
1852 | * | |
1853 | * Higher order allocations also allow the placement of more objects in a | |
1854 | * slab and thereby reduce object handling overhead. If the user has | |
1855 | * requested a higher mininum order then we start with that one instead of | |
1856 | * the smallest order which will fit the object. | |
1857 | */ | |
1858 | static inline int slab_order(int size, int min_objects, | |
1859 | int max_order, int fract_leftover) | |
1860 | { | |
1861 | int order; | |
1862 | int rem; | |
1863 | int min_order = slub_min_order; | |
1864 | ||
1865 | if ((PAGE_SIZE << min_order) / size > 65535) | |
1866 | return get_order(size * 65535) - 1; | |
1867 | ||
1868 | for (order = max(min_order, | |
1869 | fls(min_objects * size - 1) - PAGE_SHIFT); | |
1870 | order <= max_order; order++) { | |
1871 | ||
1872 | unsigned long slab_size = PAGE_SIZE << order; | |
1873 | ||
1874 | if (slab_size < min_objects * size) | |
1875 | continue; | |
1876 | ||
1877 | rem = slab_size % size; | |
1878 | ||
1879 | if (rem <= slab_size / fract_leftover) | |
1880 | break; | |
1881 | ||
1882 | } | |
1883 | ||
1884 | return order; | |
1885 | } | |
1886 | ||
1887 | static inline int calculate_order(int size) | |
1888 | { | |
1889 | int order; | |
1890 | int min_objects; | |
1891 | int fraction; | |
1892 | ||
1893 | /* | |
1894 | * Attempt to find best configuration for a slab. This | |
1895 | * works by first attempting to generate a layout with | |
1896 | * the best configuration and backing off gradually. | |
1897 | * | |
1898 | * First we reduce the acceptable waste in a slab. Then | |
1899 | * we reduce the minimum objects required in a slab. | |
1900 | */ | |
1901 | min_objects = slub_min_objects; | |
1902 | while (min_objects > 1) { | |
1903 | fraction = 8; | |
1904 | while (fraction >= 4) { | |
1905 | order = slab_order(size, min_objects, | |
1906 | slub_max_order, fraction); | |
1907 | if (order <= slub_max_order) | |
1908 | return order; | |
1909 | fraction /= 2; | |
1910 | } | |
1911 | min_objects /= 2; | |
1912 | } | |
1913 | ||
1914 | /* | |
1915 | * We were unable to place multiple objects in a slab. Now | |
1916 | * lets see if we can place a single object there. | |
1917 | */ | |
1918 | order = slab_order(size, 1, slub_max_order, 1); | |
1919 | if (order <= slub_max_order) | |
1920 | return order; | |
1921 | ||
1922 | /* | |
1923 | * Doh this slab cannot be placed using slub_max_order. | |
1924 | */ | |
1925 | order = slab_order(size, 1, MAX_ORDER, 1); | |
1926 | if (order <= MAX_ORDER) | |
1927 | return order; | |
1928 | return -ENOSYS; | |
1929 | } | |
1930 | ||
1931 | /* | |
1932 | * Figure out what the alignment of the objects will be. | |
1933 | */ | |
1934 | static unsigned long calculate_alignment(unsigned long flags, | |
1935 | unsigned long align, unsigned long size) | |
1936 | { | |
1937 | /* | |
1938 | * If the user wants hardware cache aligned objects then follow that | |
1939 | * suggestion if the object is sufficiently large. | |
1940 | * | |
1941 | * The hardware cache alignment cannot override the specified | |
1942 | * alignment though. If that is greater then use it. | |
1943 | */ | |
1944 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1945 | unsigned long ralign = cache_line_size(); | |
1946 | while (size <= ralign / 2) | |
1947 | ralign /= 2; | |
1948 | align = max(align, ralign); | |
1949 | } | |
1950 | ||
1951 | if (align < ARCH_SLAB_MINALIGN) | |
1952 | align = ARCH_SLAB_MINALIGN; | |
1953 | ||
1954 | return ALIGN(align, sizeof(void *)); | |
1955 | } | |
1956 | ||
1957 | static void init_kmem_cache_cpu(struct kmem_cache *s, | |
1958 | struct kmem_cache_cpu *c) | |
1959 | { | |
1960 | c->page = NULL; | |
1961 | c->freelist = NULL; | |
1962 | c->node = 0; | |
1963 | c->offset = s->offset / sizeof(void *); | |
1964 | c->objsize = s->objsize; | |
1965 | #ifdef CONFIG_SLUB_STATS | |
1966 | memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); | |
1967 | #endif | |
1968 | } | |
1969 | ||
1970 | static void init_kmem_cache_node(struct kmem_cache_node *n) | |
1971 | { | |
1972 | n->nr_partial = 0; | |
1973 | spin_lock_init(&n->list_lock); | |
1974 | INIT_LIST_HEAD(&n->partial); | |
1975 | #ifdef CONFIG_SLUB_DEBUG | |
1976 | atomic_long_set(&n->nr_slabs, 0); | |
1977 | INIT_LIST_HEAD(&n->full); | |
1978 | #endif | |
1979 | } | |
1980 | ||
1981 | #ifdef CONFIG_SMP | |
1982 | /* | |
1983 | * Per cpu array for per cpu structures. | |
1984 | * | |
1985 | * The per cpu array places all kmem_cache_cpu structures from one processor | |
1986 | * close together meaning that it becomes possible that multiple per cpu | |
1987 | * structures are contained in one cacheline. This may be particularly | |
1988 | * beneficial for the kmalloc caches. | |
1989 | * | |
1990 | * A desktop system typically has around 60-80 slabs. With 100 here we are | |
1991 | * likely able to get per cpu structures for all caches from the array defined | |
1992 | * here. We must be able to cover all kmalloc caches during bootstrap. | |
1993 | * | |
1994 | * If the per cpu array is exhausted then fall back to kmalloc | |
1995 | * of individual cachelines. No sharing is possible then. | |
1996 | */ | |
1997 | #define NR_KMEM_CACHE_CPU 100 | |
1998 | ||
1999 | static DEFINE_PER_CPU(struct kmem_cache_cpu, | |
2000 | kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; | |
2001 | ||
2002 | static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); | |
2003 | static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; | |
2004 | ||
2005 | static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, | |
2006 | int cpu, gfp_t flags) | |
2007 | { | |
2008 | struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); | |
2009 | ||
2010 | if (c) | |
2011 | per_cpu(kmem_cache_cpu_free, cpu) = | |
2012 | (void *)c->freelist; | |
2013 | else { | |
2014 | /* Table overflow: So allocate ourselves */ | |
2015 | c = kmalloc_node( | |
2016 | ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), | |
2017 | flags, cpu_to_node(cpu)); | |
2018 | if (!c) | |
2019 | return NULL; | |
2020 | } | |
2021 | ||
2022 | init_kmem_cache_cpu(s, c); | |
2023 | return c; | |
2024 | } | |
2025 | ||
2026 | static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) | |
2027 | { | |
2028 | if (c < per_cpu(kmem_cache_cpu, cpu) || | |
2029 | c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { | |
2030 | kfree(c); | |
2031 | return; | |
2032 | } | |
2033 | c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); | |
2034 | per_cpu(kmem_cache_cpu_free, cpu) = c; | |
2035 | } | |
2036 | ||
2037 | static void free_kmem_cache_cpus(struct kmem_cache *s) | |
2038 | { | |
2039 | int cpu; | |
2040 | ||
2041 | for_each_online_cpu(cpu) { | |
2042 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
2043 | ||
2044 | if (c) { | |
2045 | s->cpu_slab[cpu] = NULL; | |
2046 | free_kmem_cache_cpu(c, cpu); | |
2047 | } | |
2048 | } | |
2049 | } | |
2050 | ||
2051 | static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) | |
2052 | { | |
2053 | int cpu; | |
2054 | ||
2055 | for_each_online_cpu(cpu) { | |
2056 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
2057 | ||
2058 | if (c) | |
2059 | continue; | |
2060 | ||
2061 | c = alloc_kmem_cache_cpu(s, cpu, flags); | |
2062 | if (!c) { | |
2063 | free_kmem_cache_cpus(s); | |
2064 | return 0; | |
2065 | } | |
2066 | s->cpu_slab[cpu] = c; | |
2067 | } | |
2068 | return 1; | |
2069 | } | |
2070 | ||
2071 | /* | |
2072 | * Initialize the per cpu array. | |
2073 | */ | |
2074 | static void init_alloc_cpu_cpu(int cpu) | |
2075 | { | |
2076 | int i; | |
2077 | ||
2078 | if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) | |
2079 | return; | |
2080 | ||
2081 | for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) | |
2082 | free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); | |
2083 | ||
2084 | cpu_set(cpu, kmem_cach_cpu_free_init_once); | |
2085 | } | |
2086 | ||
2087 | static void __init init_alloc_cpu(void) | |
2088 | { | |
2089 | int cpu; | |
2090 | ||
2091 | for_each_online_cpu(cpu) | |
2092 | init_alloc_cpu_cpu(cpu); | |
2093 | } | |
2094 | ||
2095 | #else | |
2096 | static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} | |
2097 | static inline void init_alloc_cpu(void) {} | |
2098 | ||
2099 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) | |
2100 | { | |
2101 | init_kmem_cache_cpu(s, &s->cpu_slab); | |
2102 | return 1; | |
2103 | } | |
2104 | #endif | |
2105 | ||
2106 | #ifdef CONFIG_NUMA | |
2107 | /* | |
2108 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
2109 | * slab on the node for this slabcache. There are no concurrent accesses | |
2110 | * possible. | |
2111 | * | |
2112 | * Note that this function only works on the kmalloc_node_cache | |
2113 | * when allocating for the kmalloc_node_cache. This is used for bootstrapping | |
2114 | * memory on a fresh node that has no slab structures yet. | |
2115 | */ | |
2116 | static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags, | |
2117 | int node) | |
2118 | { | |
2119 | struct page *page; | |
2120 | struct kmem_cache_node *n; | |
2121 | unsigned long flags; | |
2122 | ||
2123 | BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); | |
2124 | ||
2125 | page = new_slab(kmalloc_caches, gfpflags, node); | |
2126 | ||
2127 | BUG_ON(!page); | |
2128 | if (page_to_nid(page) != node) { | |
2129 | printk(KERN_ERR "SLUB: Unable to allocate memory from " | |
2130 | "node %d\n", node); | |
2131 | printk(KERN_ERR "SLUB: Allocating a useless per node structure " | |
2132 | "in order to be able to continue\n"); | |
2133 | } | |
2134 | ||
2135 | n = page->freelist; | |
2136 | BUG_ON(!n); | |
2137 | page->freelist = get_freepointer(kmalloc_caches, n); | |
2138 | page->inuse++; | |
2139 | kmalloc_caches->node[node] = n; | |
2140 | #ifdef CONFIG_SLUB_DEBUG | |
2141 | init_object(kmalloc_caches, n, 1); | |
2142 | init_tracking(kmalloc_caches, n); | |
2143 | #endif | |
2144 | init_kmem_cache_node(n); | |
2145 | inc_slabs_node(kmalloc_caches, node, page->objects); | |
2146 | ||
2147 | /* | |
2148 | * lockdep requires consistent irq usage for each lock | |
2149 | * so even though there cannot be a race this early in | |
2150 | * the boot sequence, we still disable irqs. | |
2151 | */ | |
2152 | local_irq_save(flags); | |
2153 | add_partial(n, page, 0); | |
2154 | local_irq_restore(flags); | |
2155 | return n; | |
2156 | } | |
2157 | ||
2158 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2159 | { | |
2160 | int node; | |
2161 | ||
2162 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2163 | struct kmem_cache_node *n = s->node[node]; | |
2164 | if (n && n != &s->local_node) | |
2165 | kmem_cache_free(kmalloc_caches, n); | |
2166 | s->node[node] = NULL; | |
2167 | } | |
2168 | } | |
2169 | ||
2170 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
2171 | { | |
2172 | int node; | |
2173 | int local_node; | |
2174 | ||
2175 | if (slab_state >= UP) | |
2176 | local_node = page_to_nid(virt_to_page(s)); | |
2177 | else | |
2178 | local_node = 0; | |
2179 | ||
2180 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2181 | struct kmem_cache_node *n; | |
2182 | ||
2183 | if (local_node == node) | |
2184 | n = &s->local_node; | |
2185 | else { | |
2186 | if (slab_state == DOWN) { | |
2187 | n = early_kmem_cache_node_alloc(gfpflags, | |
2188 | node); | |
2189 | continue; | |
2190 | } | |
2191 | n = kmem_cache_alloc_node(kmalloc_caches, | |
2192 | gfpflags, node); | |
2193 | ||
2194 | if (!n) { | |
2195 | free_kmem_cache_nodes(s); | |
2196 | return 0; | |
2197 | } | |
2198 | ||
2199 | } | |
2200 | s->node[node] = n; | |
2201 | init_kmem_cache_node(n); | |
2202 | } | |
2203 | return 1; | |
2204 | } | |
2205 | #else | |
2206 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2207 | { | |
2208 | } | |
2209 | ||
2210 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
2211 | { | |
2212 | init_kmem_cache_node(&s->local_node); | |
2213 | return 1; | |
2214 | } | |
2215 | #endif | |
2216 | ||
2217 | /* | |
2218 | * calculate_sizes() determines the order and the distribution of data within | |
2219 | * a slab object. | |
2220 | */ | |
2221 | static int calculate_sizes(struct kmem_cache *s, int forced_order) | |
2222 | { | |
2223 | unsigned long flags = s->flags; | |
2224 | unsigned long size = s->objsize; | |
2225 | unsigned long align = s->align; | |
2226 | int order; | |
2227 | ||
2228 | /* | |
2229 | * Round up object size to the next word boundary. We can only | |
2230 | * place the free pointer at word boundaries and this determines | |
2231 | * the possible location of the free pointer. | |
2232 | */ | |
2233 | size = ALIGN(size, sizeof(void *)); | |
2234 | ||
2235 | #ifdef CONFIG_SLUB_DEBUG | |
2236 | /* | |
2237 | * Determine if we can poison the object itself. If the user of | |
2238 | * the slab may touch the object after free or before allocation | |
2239 | * then we should never poison the object itself. | |
2240 | */ | |
2241 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
2242 | !s->ctor) | |
2243 | s->flags |= __OBJECT_POISON; | |
2244 | else | |
2245 | s->flags &= ~__OBJECT_POISON; | |
2246 | ||
2247 | ||
2248 | /* | |
2249 | * If we are Redzoning then check if there is some space between the | |
2250 | * end of the object and the free pointer. If not then add an | |
2251 | * additional word to have some bytes to store Redzone information. | |
2252 | */ | |
2253 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
2254 | size += sizeof(void *); | |
2255 | #endif | |
2256 | ||
2257 | /* | |
2258 | * With that we have determined the number of bytes in actual use | |
2259 | * by the object. This is the potential offset to the free pointer. | |
2260 | */ | |
2261 | s->inuse = size; | |
2262 | ||
2263 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
2264 | s->ctor)) { | |
2265 | /* | |
2266 | * Relocate free pointer after the object if it is not | |
2267 | * permitted to overwrite the first word of the object on | |
2268 | * kmem_cache_free. | |
2269 | * | |
2270 | * This is the case if we do RCU, have a constructor or | |
2271 | * destructor or are poisoning the objects. | |
2272 | */ | |
2273 | s->offset = size; | |
2274 | size += sizeof(void *); | |
2275 | } | |
2276 | ||
2277 | #ifdef CONFIG_SLUB_DEBUG | |
2278 | if (flags & SLAB_STORE_USER) | |
2279 | /* | |
2280 | * Need to store information about allocs and frees after | |
2281 | * the object. | |
2282 | */ | |
2283 | size += 2 * sizeof(struct track); | |
2284 | ||
2285 | if (flags & SLAB_RED_ZONE) | |
2286 | /* | |
2287 | * Add some empty padding so that we can catch | |
2288 | * overwrites from earlier objects rather than let | |
2289 | * tracking information or the free pointer be | |
2290 | * corrupted if an user writes before the start | |
2291 | * of the object. | |
2292 | */ | |
2293 | size += sizeof(void *); | |
2294 | #endif | |
2295 | ||
2296 | /* | |
2297 | * Determine the alignment based on various parameters that the | |
2298 | * user specified and the dynamic determination of cache line size | |
2299 | * on bootup. | |
2300 | */ | |
2301 | align = calculate_alignment(flags, align, s->objsize); | |
2302 | ||
2303 | /* | |
2304 | * SLUB stores one object immediately after another beginning from | |
2305 | * offset 0. In order to align the objects we have to simply size | |
2306 | * each object to conform to the alignment. | |
2307 | */ | |
2308 | size = ALIGN(size, align); | |
2309 | s->size = size; | |
2310 | if (forced_order >= 0) | |
2311 | order = forced_order; | |
2312 | else | |
2313 | order = calculate_order(size); | |
2314 | ||
2315 | if (order < 0) | |
2316 | return 0; | |
2317 | ||
2318 | s->allocflags = 0; | |
2319 | if (order) | |
2320 | s->allocflags |= __GFP_COMP; | |
2321 | ||
2322 | if (s->flags & SLAB_CACHE_DMA) | |
2323 | s->allocflags |= SLUB_DMA; | |
2324 | ||
2325 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
2326 | s->allocflags |= __GFP_RECLAIMABLE; | |
2327 | ||
2328 | /* | |
2329 | * Determine the number of objects per slab | |
2330 | */ | |
2331 | s->oo = oo_make(order, size); | |
2332 | s->min = oo_make(get_order(size), size); | |
2333 | if (oo_objects(s->oo) > oo_objects(s->max)) | |
2334 | s->max = s->oo; | |
2335 | ||
2336 | return !!oo_objects(s->oo); | |
2337 | ||
2338 | } | |
2339 | ||
2340 | static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, | |
2341 | const char *name, size_t size, | |
2342 | size_t align, unsigned long flags, | |
2343 | void (*ctor)(struct kmem_cache *, void *)) | |
2344 | { | |
2345 | memset(s, 0, kmem_size); | |
2346 | s->name = name; | |
2347 | s->ctor = ctor; | |
2348 | s->objsize = size; | |
2349 | s->align = align; | |
2350 | s->flags = kmem_cache_flags(size, flags, name, ctor); | |
2351 | ||
2352 | if (!calculate_sizes(s, -1)) | |
2353 | goto error; | |
2354 | ||
2355 | s->refcount = 1; | |
2356 | #ifdef CONFIG_NUMA | |
2357 | s->remote_node_defrag_ratio = 100; | |
2358 | #endif | |
2359 | if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) | |
2360 | goto error; | |
2361 | ||
2362 | if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) | |
2363 | return 1; | |
2364 | free_kmem_cache_nodes(s); | |
2365 | error: | |
2366 | if (flags & SLAB_PANIC) | |
2367 | panic("Cannot create slab %s size=%lu realsize=%u " | |
2368 | "order=%u offset=%u flags=%lx\n", | |
2369 | s->name, (unsigned long)size, s->size, oo_order(s->oo), | |
2370 | s->offset, flags); | |
2371 | return 0; | |
2372 | } | |
2373 | ||
2374 | /* | |
2375 | * Check if a given pointer is valid | |
2376 | */ | |
2377 | int kmem_ptr_validate(struct kmem_cache *s, const void *object) | |
2378 | { | |
2379 | struct page *page; | |
2380 | ||
2381 | page = get_object_page(object); | |
2382 | ||
2383 | if (!page || s != page->slab) | |
2384 | /* No slab or wrong slab */ | |
2385 | return 0; | |
2386 | ||
2387 | if (!check_valid_pointer(s, page, object)) | |
2388 | return 0; | |
2389 | ||
2390 | /* | |
2391 | * We could also check if the object is on the slabs freelist. | |
2392 | * But this would be too expensive and it seems that the main | |
2393 | * purpose of kmem_ptr_valid() is to check if the object belongs | |
2394 | * to a certain slab. | |
2395 | */ | |
2396 | return 1; | |
2397 | } | |
2398 | EXPORT_SYMBOL(kmem_ptr_validate); | |
2399 | ||
2400 | /* | |
2401 | * Determine the size of a slab object | |
2402 | */ | |
2403 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
2404 | { | |
2405 | return s->objsize; | |
2406 | } | |
2407 | EXPORT_SYMBOL(kmem_cache_size); | |
2408 | ||
2409 | const char *kmem_cache_name(struct kmem_cache *s) | |
2410 | { | |
2411 | return s->name; | |
2412 | } | |
2413 | EXPORT_SYMBOL(kmem_cache_name); | |
2414 | ||
2415 | static void list_slab_objects(struct kmem_cache *s, struct page *page, | |
2416 | const char *text) | |
2417 | { | |
2418 | #ifdef CONFIG_SLUB_DEBUG | |
2419 | void *addr = page_address(page); | |
2420 | void *p; | |
2421 | DECLARE_BITMAP(map, page->objects); | |
2422 | ||
2423 | bitmap_zero(map, page->objects); | |
2424 | slab_err(s, page, "%s", text); | |
2425 | slab_lock(page); | |
2426 | for_each_free_object(p, s, page->freelist) | |
2427 | set_bit(slab_index(p, s, addr), map); | |
2428 | ||
2429 | for_each_object(p, s, addr, page->objects) { | |
2430 | ||
2431 | if (!test_bit(slab_index(p, s, addr), map)) { | |
2432 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", | |
2433 | p, p - addr); | |
2434 | print_tracking(s, p); | |
2435 | } | |
2436 | } | |
2437 | slab_unlock(page); | |
2438 | #endif | |
2439 | } | |
2440 | ||
2441 | /* | |
2442 | * Attempt to free all partial slabs on a node. | |
2443 | */ | |
2444 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) | |
2445 | { | |
2446 | unsigned long flags; | |
2447 | struct page *page, *h; | |
2448 | ||
2449 | spin_lock_irqsave(&n->list_lock, flags); | |
2450 | list_for_each_entry_safe(page, h, &n->partial, lru) { | |
2451 | if (!page->inuse) { | |
2452 | list_del(&page->lru); | |
2453 | discard_slab(s, page); | |
2454 | n->nr_partial--; | |
2455 | } else { | |
2456 | list_slab_objects(s, page, | |
2457 | "Objects remaining on kmem_cache_close()"); | |
2458 | } | |
2459 | } | |
2460 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2461 | } | |
2462 | ||
2463 | /* | |
2464 | * Release all resources used by a slab cache. | |
2465 | */ | |
2466 | static inline int kmem_cache_close(struct kmem_cache *s) | |
2467 | { | |
2468 | int node; | |
2469 | ||
2470 | flush_all(s); | |
2471 | ||
2472 | /* Attempt to free all objects */ | |
2473 | free_kmem_cache_cpus(s); | |
2474 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2475 | struct kmem_cache_node *n = get_node(s, node); | |
2476 | ||
2477 | free_partial(s, n); | |
2478 | if (n->nr_partial || slabs_node(s, node)) | |
2479 | return 1; | |
2480 | } | |
2481 | free_kmem_cache_nodes(s); | |
2482 | return 0; | |
2483 | } | |
2484 | ||
2485 | /* | |
2486 | * Close a cache and release the kmem_cache structure | |
2487 | * (must be used for caches created using kmem_cache_create) | |
2488 | */ | |
2489 | void kmem_cache_destroy(struct kmem_cache *s) | |
2490 | { | |
2491 | down_write(&slub_lock); | |
2492 | s->refcount--; | |
2493 | if (!s->refcount) { | |
2494 | list_del(&s->list); | |
2495 | up_write(&slub_lock); | |
2496 | if (kmem_cache_close(s)) { | |
2497 | printk(KERN_ERR "SLUB %s: %s called for cache that " | |
2498 | "still has objects.\n", s->name, __func__); | |
2499 | dump_stack(); | |
2500 | } | |
2501 | sysfs_slab_remove(s); | |
2502 | } else | |
2503 | up_write(&slub_lock); | |
2504 | } | |
2505 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2506 | ||
2507 | /******************************************************************** | |
2508 | * Kmalloc subsystem | |
2509 | *******************************************************************/ | |
2510 | ||
2511 | struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; | |
2512 | EXPORT_SYMBOL(kmalloc_caches); | |
2513 | ||
2514 | static int __init setup_slub_min_order(char *str) | |
2515 | { | |
2516 | get_option(&str, &slub_min_order); | |
2517 | ||
2518 | return 1; | |
2519 | } | |
2520 | ||
2521 | __setup("slub_min_order=", setup_slub_min_order); | |
2522 | ||
2523 | static int __init setup_slub_max_order(char *str) | |
2524 | { | |
2525 | get_option(&str, &slub_max_order); | |
2526 | ||
2527 | return 1; | |
2528 | } | |
2529 | ||
2530 | __setup("slub_max_order=", setup_slub_max_order); | |
2531 | ||
2532 | static int __init setup_slub_min_objects(char *str) | |
2533 | { | |
2534 | get_option(&str, &slub_min_objects); | |
2535 | ||
2536 | return 1; | |
2537 | } | |
2538 | ||
2539 | __setup("slub_min_objects=", setup_slub_min_objects); | |
2540 | ||
2541 | static int __init setup_slub_nomerge(char *str) | |
2542 | { | |
2543 | slub_nomerge = 1; | |
2544 | return 1; | |
2545 | } | |
2546 | ||
2547 | __setup("slub_nomerge", setup_slub_nomerge); | |
2548 | ||
2549 | static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, | |
2550 | const char *name, int size, gfp_t gfp_flags) | |
2551 | { | |
2552 | unsigned int flags = 0; | |
2553 | ||
2554 | if (gfp_flags & SLUB_DMA) | |
2555 | flags = SLAB_CACHE_DMA; | |
2556 | ||
2557 | down_write(&slub_lock); | |
2558 | if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, | |
2559 | flags, NULL)) | |
2560 | goto panic; | |
2561 | ||
2562 | list_add(&s->list, &slab_caches); | |
2563 | up_write(&slub_lock); | |
2564 | if (sysfs_slab_add(s)) | |
2565 | goto panic; | |
2566 | return s; | |
2567 | ||
2568 | panic: | |
2569 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
2570 | } | |
2571 | ||
2572 | #ifdef CONFIG_ZONE_DMA | |
2573 | static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; | |
2574 | ||
2575 | static void sysfs_add_func(struct work_struct *w) | |
2576 | { | |
2577 | struct kmem_cache *s; | |
2578 | ||
2579 | down_write(&slub_lock); | |
2580 | list_for_each_entry(s, &slab_caches, list) { | |
2581 | if (s->flags & __SYSFS_ADD_DEFERRED) { | |
2582 | s->flags &= ~__SYSFS_ADD_DEFERRED; | |
2583 | sysfs_slab_add(s); | |
2584 | } | |
2585 | } | |
2586 | up_write(&slub_lock); | |
2587 | } | |
2588 | ||
2589 | static DECLARE_WORK(sysfs_add_work, sysfs_add_func); | |
2590 | ||
2591 | static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) | |
2592 | { | |
2593 | struct kmem_cache *s; | |
2594 | char *text; | |
2595 | size_t realsize; | |
2596 | ||
2597 | s = kmalloc_caches_dma[index]; | |
2598 | if (s) | |
2599 | return s; | |
2600 | ||
2601 | /* Dynamically create dma cache */ | |
2602 | if (flags & __GFP_WAIT) | |
2603 | down_write(&slub_lock); | |
2604 | else { | |
2605 | if (!down_write_trylock(&slub_lock)) | |
2606 | goto out; | |
2607 | } | |
2608 | ||
2609 | if (kmalloc_caches_dma[index]) | |
2610 | goto unlock_out; | |
2611 | ||
2612 | realsize = kmalloc_caches[index].objsize; | |
2613 | text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", | |
2614 | (unsigned int)realsize); | |
2615 | s = kmalloc(kmem_size, flags & ~SLUB_DMA); | |
2616 | ||
2617 | if (!s || !text || !kmem_cache_open(s, flags, text, | |
2618 | realsize, ARCH_KMALLOC_MINALIGN, | |
2619 | SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { | |
2620 | kfree(s); | |
2621 | kfree(text); | |
2622 | goto unlock_out; | |
2623 | } | |
2624 | ||
2625 | list_add(&s->list, &slab_caches); | |
2626 | kmalloc_caches_dma[index] = s; | |
2627 | ||
2628 | schedule_work(&sysfs_add_work); | |
2629 | ||
2630 | unlock_out: | |
2631 | up_write(&slub_lock); | |
2632 | out: | |
2633 | return kmalloc_caches_dma[index]; | |
2634 | } | |
2635 | #endif | |
2636 | ||
2637 | /* | |
2638 | * Conversion table for small slabs sizes / 8 to the index in the | |
2639 | * kmalloc array. This is necessary for slabs < 192 since we have non power | |
2640 | * of two cache sizes there. The size of larger slabs can be determined using | |
2641 | * fls. | |
2642 | */ | |
2643 | static s8 size_index[24] = { | |
2644 | 3, /* 8 */ | |
2645 | 4, /* 16 */ | |
2646 | 5, /* 24 */ | |
2647 | 5, /* 32 */ | |
2648 | 6, /* 40 */ | |
2649 | 6, /* 48 */ | |
2650 | 6, /* 56 */ | |
2651 | 6, /* 64 */ | |
2652 | 1, /* 72 */ | |
2653 | 1, /* 80 */ | |
2654 | 1, /* 88 */ | |
2655 | 1, /* 96 */ | |
2656 | 7, /* 104 */ | |
2657 | 7, /* 112 */ | |
2658 | 7, /* 120 */ | |
2659 | 7, /* 128 */ | |
2660 | 2, /* 136 */ | |
2661 | 2, /* 144 */ | |
2662 | 2, /* 152 */ | |
2663 | 2, /* 160 */ | |
2664 | 2, /* 168 */ | |
2665 | 2, /* 176 */ | |
2666 | 2, /* 184 */ | |
2667 | 2 /* 192 */ | |
2668 | }; | |
2669 | ||
2670 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) | |
2671 | { | |
2672 | int index; | |
2673 | ||
2674 | if (size <= 192) { | |
2675 | if (!size) | |
2676 | return ZERO_SIZE_PTR; | |
2677 | ||
2678 | index = size_index[(size - 1) / 8]; | |
2679 | } else | |
2680 | index = fls(size - 1); | |
2681 | ||
2682 | #ifdef CONFIG_ZONE_DMA | |
2683 | if (unlikely((flags & SLUB_DMA))) | |
2684 | return dma_kmalloc_cache(index, flags); | |
2685 | ||
2686 | #endif | |
2687 | return &kmalloc_caches[index]; | |
2688 | } | |
2689 | ||
2690 | void *__kmalloc(size_t size, gfp_t flags) | |
2691 | { | |
2692 | struct kmem_cache *s; | |
2693 | ||
2694 | if (unlikely(size > PAGE_SIZE)) | |
2695 | return kmalloc_large(size, flags); | |
2696 | ||
2697 | s = get_slab(size, flags); | |
2698 | ||
2699 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
2700 | return s; | |
2701 | ||
2702 | return slab_alloc(s, flags, -1, __builtin_return_address(0)); | |
2703 | } | |
2704 | EXPORT_SYMBOL(__kmalloc); | |
2705 | ||
2706 | static void *kmalloc_large_node(size_t size, gfp_t flags, int node) | |
2707 | { | |
2708 | struct page *page = alloc_pages_node(node, flags | __GFP_COMP, | |
2709 | get_order(size)); | |
2710 | ||
2711 | if (page) | |
2712 | return page_address(page); | |
2713 | else | |
2714 | return NULL; | |
2715 | } | |
2716 | ||
2717 | #ifdef CONFIG_NUMA | |
2718 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
2719 | { | |
2720 | struct kmem_cache *s; | |
2721 | ||
2722 | if (unlikely(size > PAGE_SIZE)) | |
2723 | return kmalloc_large_node(size, flags, node); | |
2724 | ||
2725 | s = get_slab(size, flags); | |
2726 | ||
2727 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
2728 | return s; | |
2729 | ||
2730 | return slab_alloc(s, flags, node, __builtin_return_address(0)); | |
2731 | } | |
2732 | EXPORT_SYMBOL(__kmalloc_node); | |
2733 | #endif | |
2734 | ||
2735 | size_t ksize(const void *object) | |
2736 | { | |
2737 | struct page *page; | |
2738 | struct kmem_cache *s; | |
2739 | ||
2740 | if (unlikely(object == ZERO_SIZE_PTR)) | |
2741 | return 0; | |
2742 | ||
2743 | page = virt_to_head_page(object); | |
2744 | ||
2745 | if (unlikely(!PageSlab(page))) | |
2746 | return PAGE_SIZE << compound_order(page); | |
2747 | ||
2748 | s = page->slab; | |
2749 | ||
2750 | #ifdef CONFIG_SLUB_DEBUG | |
2751 | /* | |
2752 | * Debugging requires use of the padding between object | |
2753 | * and whatever may come after it. | |
2754 | */ | |
2755 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
2756 | return s->objsize; | |
2757 | ||
2758 | #endif | |
2759 | /* | |
2760 | * If we have the need to store the freelist pointer | |
2761 | * back there or track user information then we can | |
2762 | * only use the space before that information. | |
2763 | */ | |
2764 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
2765 | return s->inuse; | |
2766 | /* | |
2767 | * Else we can use all the padding etc for the allocation | |
2768 | */ | |
2769 | return s->size; | |
2770 | } | |
2771 | EXPORT_SYMBOL(ksize); | |
2772 | ||
2773 | void kfree(const void *x) | |
2774 | { | |
2775 | struct page *page; | |
2776 | void *object = (void *)x; | |
2777 | ||
2778 | if (unlikely(ZERO_OR_NULL_PTR(x))) | |
2779 | return; | |
2780 | ||
2781 | page = virt_to_head_page(x); | |
2782 | if (unlikely(!PageSlab(page))) { | |
2783 | put_page(page); | |
2784 | return; | |
2785 | } | |
2786 | slab_free(page->slab, page, object, __builtin_return_address(0)); | |
2787 | } | |
2788 | EXPORT_SYMBOL(kfree); | |
2789 | ||
2790 | /* | |
2791 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts | |
2792 | * the remaining slabs by the number of items in use. The slabs with the | |
2793 | * most items in use come first. New allocations will then fill those up | |
2794 | * and thus they can be removed from the partial lists. | |
2795 | * | |
2796 | * The slabs with the least items are placed last. This results in them | |
2797 | * being allocated from last increasing the chance that the last objects | |
2798 | * are freed in them. | |
2799 | */ | |
2800 | int kmem_cache_shrink(struct kmem_cache *s) | |
2801 | { | |
2802 | int node; | |
2803 | int i; | |
2804 | struct kmem_cache_node *n; | |
2805 | struct page *page; | |
2806 | struct page *t; | |
2807 | int objects = oo_objects(s->max); | |
2808 | struct list_head *slabs_by_inuse = | |
2809 | kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); | |
2810 | unsigned long flags; | |
2811 | ||
2812 | if (!slabs_by_inuse) | |
2813 | return -ENOMEM; | |
2814 | ||
2815 | flush_all(s); | |
2816 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
2817 | n = get_node(s, node); | |
2818 | ||
2819 | if (!n->nr_partial) | |
2820 | continue; | |
2821 | ||
2822 | for (i = 0; i < objects; i++) | |
2823 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
2824 | ||
2825 | spin_lock_irqsave(&n->list_lock, flags); | |
2826 | ||
2827 | /* | |
2828 | * Build lists indexed by the items in use in each slab. | |
2829 | * | |
2830 | * Note that concurrent frees may occur while we hold the | |
2831 | * list_lock. page->inuse here is the upper limit. | |
2832 | */ | |
2833 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
2834 | if (!page->inuse && slab_trylock(page)) { | |
2835 | /* | |
2836 | * Must hold slab lock here because slab_free | |
2837 | * may have freed the last object and be | |
2838 | * waiting to release the slab. | |
2839 | */ | |
2840 | list_del(&page->lru); | |
2841 | n->nr_partial--; | |
2842 | slab_unlock(page); | |
2843 | discard_slab(s, page); | |
2844 | } else { | |
2845 | list_move(&page->lru, | |
2846 | slabs_by_inuse + page->inuse); | |
2847 | } | |
2848 | } | |
2849 | ||
2850 | /* | |
2851 | * Rebuild the partial list with the slabs filled up most | |
2852 | * first and the least used slabs at the end. | |
2853 | */ | |
2854 | for (i = objects - 1; i >= 0; i--) | |
2855 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
2856 | ||
2857 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2858 | } | |
2859 | ||
2860 | kfree(slabs_by_inuse); | |
2861 | return 0; | |
2862 | } | |
2863 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2864 | ||
2865 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) | |
2866 | static int slab_mem_going_offline_callback(void *arg) | |
2867 | { | |
2868 | struct kmem_cache *s; | |
2869 | ||
2870 | down_read(&slub_lock); | |
2871 | list_for_each_entry(s, &slab_caches, list) | |
2872 | kmem_cache_shrink(s); | |
2873 | up_read(&slub_lock); | |
2874 | ||
2875 | return 0; | |
2876 | } | |
2877 | ||
2878 | static void slab_mem_offline_callback(void *arg) | |
2879 | { | |
2880 | struct kmem_cache_node *n; | |
2881 | struct kmem_cache *s; | |
2882 | struct memory_notify *marg = arg; | |
2883 | int offline_node; | |
2884 | ||
2885 | offline_node = marg->status_change_nid; | |
2886 | ||
2887 | /* | |
2888 | * If the node still has available memory. we need kmem_cache_node | |
2889 | * for it yet. | |
2890 | */ | |
2891 | if (offline_node < 0) | |
2892 | return; | |
2893 | ||
2894 | down_read(&slub_lock); | |
2895 | list_for_each_entry(s, &slab_caches, list) { | |
2896 | n = get_node(s, offline_node); | |
2897 | if (n) { | |
2898 | /* | |
2899 | * if n->nr_slabs > 0, slabs still exist on the node | |
2900 | * that is going down. We were unable to free them, | |
2901 | * and offline_pages() function shoudn't call this | |
2902 | * callback. So, we must fail. | |
2903 | */ | |
2904 | BUG_ON(slabs_node(s, offline_node)); | |
2905 | ||
2906 | s->node[offline_node] = NULL; | |
2907 | kmem_cache_free(kmalloc_caches, n); | |
2908 | } | |
2909 | } | |
2910 | up_read(&slub_lock); | |
2911 | } | |
2912 | ||
2913 | static int slab_mem_going_online_callback(void *arg) | |
2914 | { | |
2915 | struct kmem_cache_node *n; | |
2916 | struct kmem_cache *s; | |
2917 | struct memory_notify *marg = arg; | |
2918 | int nid = marg->status_change_nid; | |
2919 | int ret = 0; | |
2920 | ||
2921 | /* | |
2922 | * If the node's memory is already available, then kmem_cache_node is | |
2923 | * already created. Nothing to do. | |
2924 | */ | |
2925 | if (nid < 0) | |
2926 | return 0; | |
2927 | ||
2928 | /* | |
2929 | * We are bringing a node online. No memory is availabe yet. We must | |
2930 | * allocate a kmem_cache_node structure in order to bring the node | |
2931 | * online. | |
2932 | */ | |
2933 | down_read(&slub_lock); | |
2934 | list_for_each_entry(s, &slab_caches, list) { | |
2935 | /* | |
2936 | * XXX: kmem_cache_alloc_node will fallback to other nodes | |
2937 | * since memory is not yet available from the node that | |
2938 | * is brought up. | |
2939 | */ | |
2940 | n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); | |
2941 | if (!n) { | |
2942 | ret = -ENOMEM; | |
2943 | goto out; | |
2944 | } | |
2945 | init_kmem_cache_node(n); | |
2946 | s->node[nid] = n; | |
2947 | } | |
2948 | out: | |
2949 | up_read(&slub_lock); | |
2950 | return ret; | |
2951 | } | |
2952 | ||
2953 | static int slab_memory_callback(struct notifier_block *self, | |
2954 | unsigned long action, void *arg) | |
2955 | { | |
2956 | int ret = 0; | |
2957 | ||
2958 | switch (action) { | |
2959 | case MEM_GOING_ONLINE: | |
2960 | ret = slab_mem_going_online_callback(arg); | |
2961 | break; | |
2962 | case MEM_GOING_OFFLINE: | |
2963 | ret = slab_mem_going_offline_callback(arg); | |
2964 | break; | |
2965 | case MEM_OFFLINE: | |
2966 | case MEM_CANCEL_ONLINE: | |
2967 | slab_mem_offline_callback(arg); | |
2968 | break; | |
2969 | case MEM_ONLINE: | |
2970 | case MEM_CANCEL_OFFLINE: | |
2971 | break; | |
2972 | } | |
2973 | ||
2974 | ret = notifier_from_errno(ret); | |
2975 | return ret; | |
2976 | } | |
2977 | ||
2978 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
2979 | ||
2980 | /******************************************************************** | |
2981 | * Basic setup of slabs | |
2982 | *******************************************************************/ | |
2983 | ||
2984 | void __init kmem_cache_init(void) | |
2985 | { | |
2986 | int i; | |
2987 | int caches = 0; | |
2988 | ||
2989 | init_alloc_cpu(); | |
2990 | ||
2991 | #ifdef CONFIG_NUMA | |
2992 | /* | |
2993 | * Must first have the slab cache available for the allocations of the | |
2994 | * struct kmem_cache_node's. There is special bootstrap code in | |
2995 | * kmem_cache_open for slab_state == DOWN. | |
2996 | */ | |
2997 | create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", | |
2998 | sizeof(struct kmem_cache_node), GFP_KERNEL); | |
2999 | kmalloc_caches[0].refcount = -1; | |
3000 | caches++; | |
3001 | ||
3002 | hotplug_memory_notifier(slab_memory_callback, 1); | |
3003 | #endif | |
3004 | ||
3005 | /* Able to allocate the per node structures */ | |
3006 | slab_state = PARTIAL; | |
3007 | ||
3008 | /* Caches that are not of the two-to-the-power-of size */ | |
3009 | if (KMALLOC_MIN_SIZE <= 64) { | |
3010 | create_kmalloc_cache(&kmalloc_caches[1], | |
3011 | "kmalloc-96", 96, GFP_KERNEL); | |
3012 | caches++; | |
3013 | } | |
3014 | if (KMALLOC_MIN_SIZE <= 128) { | |
3015 | create_kmalloc_cache(&kmalloc_caches[2], | |
3016 | "kmalloc-192", 192, GFP_KERNEL); | |
3017 | caches++; | |
3018 | } | |
3019 | ||
3020 | for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { | |
3021 | create_kmalloc_cache(&kmalloc_caches[i], | |
3022 | "kmalloc", 1 << i, GFP_KERNEL); | |
3023 | caches++; | |
3024 | } | |
3025 | ||
3026 | ||
3027 | /* | |
3028 | * Patch up the size_index table if we have strange large alignment | |
3029 | * requirements for the kmalloc array. This is only the case for | |
3030 | * MIPS it seems. The standard arches will not generate any code here. | |
3031 | * | |
3032 | * Largest permitted alignment is 256 bytes due to the way we | |
3033 | * handle the index determination for the smaller caches. | |
3034 | * | |
3035 | * Make sure that nothing crazy happens if someone starts tinkering | |
3036 | * around with ARCH_KMALLOC_MINALIGN | |
3037 | */ | |
3038 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || | |
3039 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); | |
3040 | ||
3041 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) | |
3042 | size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; | |
3043 | ||
3044 | slab_state = UP; | |
3045 | ||
3046 | /* Provide the correct kmalloc names now that the caches are up */ | |
3047 | for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) | |
3048 | kmalloc_caches[i]. name = | |
3049 | kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); | |
3050 | ||
3051 | #ifdef CONFIG_SMP | |
3052 | register_cpu_notifier(&slab_notifier); | |
3053 | kmem_size = offsetof(struct kmem_cache, cpu_slab) + | |
3054 | nr_cpu_ids * sizeof(struct kmem_cache_cpu *); | |
3055 | #else | |
3056 | kmem_size = sizeof(struct kmem_cache); | |
3057 | #endif | |
3058 | ||
3059 | printk(KERN_INFO | |
3060 | "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
3061 | " CPUs=%d, Nodes=%d\n", | |
3062 | caches, cache_line_size(), | |
3063 | slub_min_order, slub_max_order, slub_min_objects, | |
3064 | nr_cpu_ids, nr_node_ids); | |
3065 | } | |
3066 | ||
3067 | /* | |
3068 | * Find a mergeable slab cache | |
3069 | */ | |
3070 | static int slab_unmergeable(struct kmem_cache *s) | |
3071 | { | |
3072 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
3073 | return 1; | |
3074 | ||
3075 | if (s->ctor) | |
3076 | return 1; | |
3077 | ||
3078 | /* | |
3079 | * We may have set a slab to be unmergeable during bootstrap. | |
3080 | */ | |
3081 | if (s->refcount < 0) | |
3082 | return 1; | |
3083 | ||
3084 | return 0; | |
3085 | } | |
3086 | ||
3087 | static struct kmem_cache *find_mergeable(size_t size, | |
3088 | size_t align, unsigned long flags, const char *name, | |
3089 | void (*ctor)(struct kmem_cache *, void *)) | |
3090 | { | |
3091 | struct kmem_cache *s; | |
3092 | ||
3093 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
3094 | return NULL; | |
3095 | ||
3096 | if (ctor) | |
3097 | return NULL; | |
3098 | ||
3099 | size = ALIGN(size, sizeof(void *)); | |
3100 | align = calculate_alignment(flags, align, size); | |
3101 | size = ALIGN(size, align); | |
3102 | flags = kmem_cache_flags(size, flags, name, NULL); | |
3103 | ||
3104 | list_for_each_entry(s, &slab_caches, list) { | |
3105 | if (slab_unmergeable(s)) | |
3106 | continue; | |
3107 | ||
3108 | if (size > s->size) | |
3109 | continue; | |
3110 | ||
3111 | if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) | |
3112 | continue; | |
3113 | /* | |
3114 | * Check if alignment is compatible. | |
3115 | * Courtesy of Adrian Drzewiecki | |
3116 | */ | |
3117 | if ((s->size & ~(align - 1)) != s->size) | |
3118 | continue; | |
3119 | ||
3120 | if (s->size - size >= sizeof(void *)) | |
3121 | continue; | |
3122 | ||
3123 | return s; | |
3124 | } | |
3125 | return NULL; | |
3126 | } | |
3127 | ||
3128 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
3129 | size_t align, unsigned long flags, | |
3130 | void (*ctor)(struct kmem_cache *, void *)) | |
3131 | { | |
3132 | struct kmem_cache *s; | |
3133 | ||
3134 | down_write(&slub_lock); | |
3135 | s = find_mergeable(size, align, flags, name, ctor); | |
3136 | if (s) { | |
3137 | int cpu; | |
3138 | ||
3139 | s->refcount++; | |
3140 | /* | |
3141 | * Adjust the object sizes so that we clear | |
3142 | * the complete object on kzalloc. | |
3143 | */ | |
3144 | s->objsize = max(s->objsize, (int)size); | |
3145 | ||
3146 | /* | |
3147 | * And then we need to update the object size in the | |
3148 | * per cpu structures | |
3149 | */ | |
3150 | for_each_online_cpu(cpu) | |
3151 | get_cpu_slab(s, cpu)->objsize = s->objsize; | |
3152 | ||
3153 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); | |
3154 | up_write(&slub_lock); | |
3155 | ||
3156 | if (sysfs_slab_alias(s, name)) | |
3157 | goto err; | |
3158 | return s; | |
3159 | } | |
3160 | ||
3161 | s = kmalloc(kmem_size, GFP_KERNEL); | |
3162 | if (s) { | |
3163 | if (kmem_cache_open(s, GFP_KERNEL, name, | |
3164 | size, align, flags, ctor)) { | |
3165 | list_add(&s->list, &slab_caches); | |
3166 | up_write(&slub_lock); | |
3167 | if (sysfs_slab_add(s)) | |
3168 | goto err; | |
3169 | return s; | |
3170 | } | |
3171 | kfree(s); | |
3172 | } | |
3173 | up_write(&slub_lock); | |
3174 | ||
3175 | err: | |
3176 | if (flags & SLAB_PANIC) | |
3177 | panic("Cannot create slabcache %s\n", name); | |
3178 | else | |
3179 | s = NULL; | |
3180 | return s; | |
3181 | } | |
3182 | EXPORT_SYMBOL(kmem_cache_create); | |
3183 | ||
3184 | #ifdef CONFIG_SMP | |
3185 | /* | |
3186 | * Use the cpu notifier to insure that the cpu slabs are flushed when | |
3187 | * necessary. | |
3188 | */ | |
3189 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
3190 | unsigned long action, void *hcpu) | |
3191 | { | |
3192 | long cpu = (long)hcpu; | |
3193 | struct kmem_cache *s; | |
3194 | unsigned long flags; | |
3195 | ||
3196 | switch (action) { | |
3197 | case CPU_UP_PREPARE: | |
3198 | case CPU_UP_PREPARE_FROZEN: | |
3199 | init_alloc_cpu_cpu(cpu); | |
3200 | down_read(&slub_lock); | |
3201 | list_for_each_entry(s, &slab_caches, list) | |
3202 | s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, | |
3203 | GFP_KERNEL); | |
3204 | up_read(&slub_lock); | |
3205 | break; | |
3206 | ||
3207 | case CPU_UP_CANCELED: | |
3208 | case CPU_UP_CANCELED_FROZEN: | |
3209 | case CPU_DEAD: | |
3210 | case CPU_DEAD_FROZEN: | |
3211 | down_read(&slub_lock); | |
3212 | list_for_each_entry(s, &slab_caches, list) { | |
3213 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
3214 | ||
3215 | local_irq_save(flags); | |
3216 | __flush_cpu_slab(s, cpu); | |
3217 | local_irq_restore(flags); | |
3218 | free_kmem_cache_cpu(c, cpu); | |
3219 | s->cpu_slab[cpu] = NULL; | |
3220 | } | |
3221 | up_read(&slub_lock); | |
3222 | break; | |
3223 | default: | |
3224 | break; | |
3225 | } | |
3226 | return NOTIFY_OK; | |
3227 | } | |
3228 | ||
3229 | static struct notifier_block __cpuinitdata slab_notifier = { | |
3230 | .notifier_call = slab_cpuup_callback | |
3231 | }; | |
3232 | ||
3233 | #endif | |
3234 | ||
3235 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) | |
3236 | { | |
3237 | struct kmem_cache *s; | |
3238 | ||
3239 | if (unlikely(size > PAGE_SIZE)) | |
3240 | return kmalloc_large(size, gfpflags); | |
3241 | ||
3242 | s = get_slab(size, gfpflags); | |
3243 | ||
3244 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3245 | return s; | |
3246 | ||
3247 | return slab_alloc(s, gfpflags, -1, caller); | |
3248 | } | |
3249 | ||
3250 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
3251 | int node, void *caller) | |
3252 | { | |
3253 | struct kmem_cache *s; | |
3254 | ||
3255 | if (unlikely(size > PAGE_SIZE)) | |
3256 | return kmalloc_large_node(size, gfpflags, node); | |
3257 | ||
3258 | s = get_slab(size, gfpflags); | |
3259 | ||
3260 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
3261 | return s; | |
3262 | ||
3263 | return slab_alloc(s, gfpflags, node, caller); | |
3264 | } | |
3265 | ||
3266 | #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO) | |
3267 | static unsigned long count_partial(struct kmem_cache_node *n, | |
3268 | int (*get_count)(struct page *)) | |
3269 | { | |
3270 | unsigned long flags; | |
3271 | unsigned long x = 0; | |
3272 | struct page *page; | |
3273 | ||
3274 | spin_lock_irqsave(&n->list_lock, flags); | |
3275 | list_for_each_entry(page, &n->partial, lru) | |
3276 | x += get_count(page); | |
3277 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3278 | return x; | |
3279 | } | |
3280 | ||
3281 | static int count_inuse(struct page *page) | |
3282 | { | |
3283 | return page->inuse; | |
3284 | } | |
3285 | ||
3286 | static int count_total(struct page *page) | |
3287 | { | |
3288 | return page->objects; | |
3289 | } | |
3290 | ||
3291 | static int count_free(struct page *page) | |
3292 | { | |
3293 | return page->objects - page->inuse; | |
3294 | } | |
3295 | #endif | |
3296 | ||
3297 | #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) | |
3298 | static int validate_slab(struct kmem_cache *s, struct page *page, | |
3299 | unsigned long *map) | |
3300 | { | |
3301 | void *p; | |
3302 | void *addr = page_address(page); | |
3303 | ||
3304 | if (!check_slab(s, page) || | |
3305 | !on_freelist(s, page, NULL)) | |
3306 | return 0; | |
3307 | ||
3308 | /* Now we know that a valid freelist exists */ | |
3309 | bitmap_zero(map, page->objects); | |
3310 | ||
3311 | for_each_free_object(p, s, page->freelist) { | |
3312 | set_bit(slab_index(p, s, addr), map); | |
3313 | if (!check_object(s, page, p, 0)) | |
3314 | return 0; | |
3315 | } | |
3316 | ||
3317 | for_each_object(p, s, addr, page->objects) | |
3318 | if (!test_bit(slab_index(p, s, addr), map)) | |
3319 | if (!check_object(s, page, p, 1)) | |
3320 | return 0; | |
3321 | return 1; | |
3322 | } | |
3323 | ||
3324 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, | |
3325 | unsigned long *map) | |
3326 | { | |
3327 | if (slab_trylock(page)) { | |
3328 | validate_slab(s, page, map); | |
3329 | slab_unlock(page); | |
3330 | } else | |
3331 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", | |
3332 | s->name, page); | |
3333 | ||
3334 | if (s->flags & DEBUG_DEFAULT_FLAGS) { | |
3335 | if (!SlabDebug(page)) | |
3336 | printk(KERN_ERR "SLUB %s: SlabDebug not set " | |
3337 | "on slab 0x%p\n", s->name, page); | |
3338 | } else { | |
3339 | if (SlabDebug(page)) | |
3340 | printk(KERN_ERR "SLUB %s: SlabDebug set on " | |
3341 | "slab 0x%p\n", s->name, page); | |
3342 | } | |
3343 | } | |
3344 | ||
3345 | static int validate_slab_node(struct kmem_cache *s, | |
3346 | struct kmem_cache_node *n, unsigned long *map) | |
3347 | { | |
3348 | unsigned long count = 0; | |
3349 | struct page *page; | |
3350 | unsigned long flags; | |
3351 | ||
3352 | spin_lock_irqsave(&n->list_lock, flags); | |
3353 | ||
3354 | list_for_each_entry(page, &n->partial, lru) { | |
3355 | validate_slab_slab(s, page, map); | |
3356 | count++; | |
3357 | } | |
3358 | if (count != n->nr_partial) | |
3359 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
3360 | "counter=%ld\n", s->name, count, n->nr_partial); | |
3361 | ||
3362 | if (!(s->flags & SLAB_STORE_USER)) | |
3363 | goto out; | |
3364 | ||
3365 | list_for_each_entry(page, &n->full, lru) { | |
3366 | validate_slab_slab(s, page, map); | |
3367 | count++; | |
3368 | } | |
3369 | if (count != atomic_long_read(&n->nr_slabs)) | |
3370 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
3371 | "counter=%ld\n", s->name, count, | |
3372 | atomic_long_read(&n->nr_slabs)); | |
3373 | ||
3374 | out: | |
3375 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3376 | return count; | |
3377 | } | |
3378 | ||
3379 | static long validate_slab_cache(struct kmem_cache *s) | |
3380 | { | |
3381 | int node; | |
3382 | unsigned long count = 0; | |
3383 | unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * | |
3384 | sizeof(unsigned long), GFP_KERNEL); | |
3385 | ||
3386 | if (!map) | |
3387 | return -ENOMEM; | |
3388 | ||
3389 | flush_all(s); | |
3390 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3391 | struct kmem_cache_node *n = get_node(s, node); | |
3392 | ||
3393 | count += validate_slab_node(s, n, map); | |
3394 | } | |
3395 | kfree(map); | |
3396 | return count; | |
3397 | } | |
3398 | ||
3399 | #ifdef SLUB_RESILIENCY_TEST | |
3400 | static void resiliency_test(void) | |
3401 | { | |
3402 | u8 *p; | |
3403 | ||
3404 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
3405 | printk(KERN_ERR "-----------------------\n"); | |
3406 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
3407 | ||
3408 | p = kzalloc(16, GFP_KERNEL); | |
3409 | p[16] = 0x12; | |
3410 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
3411 | " 0x12->0x%p\n\n", p + 16); | |
3412 | ||
3413 | validate_slab_cache(kmalloc_caches + 4); | |
3414 | ||
3415 | /* Hmmm... The next two are dangerous */ | |
3416 | p = kzalloc(32, GFP_KERNEL); | |
3417 | p[32 + sizeof(void *)] = 0x34; | |
3418 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
3419 | " 0x34 -> -0x%p\n", p); | |
3420 | printk(KERN_ERR | |
3421 | "If allocated object is overwritten then not detectable\n\n"); | |
3422 | ||
3423 | validate_slab_cache(kmalloc_caches + 5); | |
3424 | p = kzalloc(64, GFP_KERNEL); | |
3425 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
3426 | *p = 0x56; | |
3427 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
3428 | p); | |
3429 | printk(KERN_ERR | |
3430 | "If allocated object is overwritten then not detectable\n\n"); | |
3431 | validate_slab_cache(kmalloc_caches + 6); | |
3432 | ||
3433 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
3434 | p = kzalloc(128, GFP_KERNEL); | |
3435 | kfree(p); | |
3436 | *p = 0x78; | |
3437 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
3438 | validate_slab_cache(kmalloc_caches + 7); | |
3439 | ||
3440 | p = kzalloc(256, GFP_KERNEL); | |
3441 | kfree(p); | |
3442 | p[50] = 0x9a; | |
3443 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", | |
3444 | p); | |
3445 | validate_slab_cache(kmalloc_caches + 8); | |
3446 | ||
3447 | p = kzalloc(512, GFP_KERNEL); | |
3448 | kfree(p); | |
3449 | p[512] = 0xab; | |
3450 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
3451 | validate_slab_cache(kmalloc_caches + 9); | |
3452 | } | |
3453 | #else | |
3454 | static void resiliency_test(void) {}; | |
3455 | #endif | |
3456 | ||
3457 | /* | |
3458 | * Generate lists of code addresses where slabcache objects are allocated | |
3459 | * and freed. | |
3460 | */ | |
3461 | ||
3462 | struct location { | |
3463 | unsigned long count; | |
3464 | void *addr; | |
3465 | long long sum_time; | |
3466 | long min_time; | |
3467 | long max_time; | |
3468 | long min_pid; | |
3469 | long max_pid; | |
3470 | cpumask_t cpus; | |
3471 | nodemask_t nodes; | |
3472 | }; | |
3473 | ||
3474 | struct loc_track { | |
3475 | unsigned long max; | |
3476 | unsigned long count; | |
3477 | struct location *loc; | |
3478 | }; | |
3479 | ||
3480 | static void free_loc_track(struct loc_track *t) | |
3481 | { | |
3482 | if (t->max) | |
3483 | free_pages((unsigned long)t->loc, | |
3484 | get_order(sizeof(struct location) * t->max)); | |
3485 | } | |
3486 | ||
3487 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) | |
3488 | { | |
3489 | struct location *l; | |
3490 | int order; | |
3491 | ||
3492 | order = get_order(sizeof(struct location) * max); | |
3493 | ||
3494 | l = (void *)__get_free_pages(flags, order); | |
3495 | if (!l) | |
3496 | return 0; | |
3497 | ||
3498 | if (t->count) { | |
3499 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
3500 | free_loc_track(t); | |
3501 | } | |
3502 | t->max = max; | |
3503 | t->loc = l; | |
3504 | return 1; | |
3505 | } | |
3506 | ||
3507 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
3508 | const struct track *track) | |
3509 | { | |
3510 | long start, end, pos; | |
3511 | struct location *l; | |
3512 | void *caddr; | |
3513 | unsigned long age = jiffies - track->when; | |
3514 | ||
3515 | start = -1; | |
3516 | end = t->count; | |
3517 | ||
3518 | for ( ; ; ) { | |
3519 | pos = start + (end - start + 1) / 2; | |
3520 | ||
3521 | /* | |
3522 | * There is nothing at "end". If we end up there | |
3523 | * we need to add something to before end. | |
3524 | */ | |
3525 | if (pos == end) | |
3526 | break; | |
3527 | ||
3528 | caddr = t->loc[pos].addr; | |
3529 | if (track->addr == caddr) { | |
3530 | ||
3531 | l = &t->loc[pos]; | |
3532 | l->count++; | |
3533 | if (track->when) { | |
3534 | l->sum_time += age; | |
3535 | if (age < l->min_time) | |
3536 | l->min_time = age; | |
3537 | if (age > l->max_time) | |
3538 | l->max_time = age; | |
3539 | ||
3540 | if (track->pid < l->min_pid) | |
3541 | l->min_pid = track->pid; | |
3542 | if (track->pid > l->max_pid) | |
3543 | l->max_pid = track->pid; | |
3544 | ||
3545 | cpu_set(track->cpu, l->cpus); | |
3546 | } | |
3547 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
3548 | return 1; | |
3549 | } | |
3550 | ||
3551 | if (track->addr < caddr) | |
3552 | end = pos; | |
3553 | else | |
3554 | start = pos; | |
3555 | } | |
3556 | ||
3557 | /* | |
3558 | * Not found. Insert new tracking element. | |
3559 | */ | |
3560 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) | |
3561 | return 0; | |
3562 | ||
3563 | l = t->loc + pos; | |
3564 | if (pos < t->count) | |
3565 | memmove(l + 1, l, | |
3566 | (t->count - pos) * sizeof(struct location)); | |
3567 | t->count++; | |
3568 | l->count = 1; | |
3569 | l->addr = track->addr; | |
3570 | l->sum_time = age; | |
3571 | l->min_time = age; | |
3572 | l->max_time = age; | |
3573 | l->min_pid = track->pid; | |
3574 | l->max_pid = track->pid; | |
3575 | cpus_clear(l->cpus); | |
3576 | cpu_set(track->cpu, l->cpus); | |
3577 | nodes_clear(l->nodes); | |
3578 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
3579 | return 1; | |
3580 | } | |
3581 | ||
3582 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
3583 | struct page *page, enum track_item alloc) | |
3584 | { | |
3585 | void *addr = page_address(page); | |
3586 | DECLARE_BITMAP(map, page->objects); | |
3587 | void *p; | |
3588 | ||
3589 | bitmap_zero(map, page->objects); | |
3590 | for_each_free_object(p, s, page->freelist) | |
3591 | set_bit(slab_index(p, s, addr), map); | |
3592 | ||
3593 | for_each_object(p, s, addr, page->objects) | |
3594 | if (!test_bit(slab_index(p, s, addr), map)) | |
3595 | add_location(t, s, get_track(s, p, alloc)); | |
3596 | } | |
3597 | ||
3598 | static int list_locations(struct kmem_cache *s, char *buf, | |
3599 | enum track_item alloc) | |
3600 | { | |
3601 | int len = 0; | |
3602 | unsigned long i; | |
3603 | struct loc_track t = { 0, 0, NULL }; | |
3604 | int node; | |
3605 | ||
3606 | if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), | |
3607 | GFP_TEMPORARY)) | |
3608 | return sprintf(buf, "Out of memory\n"); | |
3609 | ||
3610 | /* Push back cpu slabs */ | |
3611 | flush_all(s); | |
3612 | ||
3613 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3614 | struct kmem_cache_node *n = get_node(s, node); | |
3615 | unsigned long flags; | |
3616 | struct page *page; | |
3617 | ||
3618 | if (!atomic_long_read(&n->nr_slabs)) | |
3619 | continue; | |
3620 | ||
3621 | spin_lock_irqsave(&n->list_lock, flags); | |
3622 | list_for_each_entry(page, &n->partial, lru) | |
3623 | process_slab(&t, s, page, alloc); | |
3624 | list_for_each_entry(page, &n->full, lru) | |
3625 | process_slab(&t, s, page, alloc); | |
3626 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3627 | } | |
3628 | ||
3629 | for (i = 0; i < t.count; i++) { | |
3630 | struct location *l = &t.loc[i]; | |
3631 | ||
3632 | if (len > PAGE_SIZE - 100) | |
3633 | break; | |
3634 | len += sprintf(buf + len, "%7ld ", l->count); | |
3635 | ||
3636 | if (l->addr) | |
3637 | len += sprint_symbol(buf + len, (unsigned long)l->addr); | |
3638 | else | |
3639 | len += sprintf(buf + len, "<not-available>"); | |
3640 | ||
3641 | if (l->sum_time != l->min_time) { | |
3642 | unsigned long remainder; | |
3643 | ||
3644 | len += sprintf(buf + len, " age=%ld/%ld/%ld", | |
3645 | l->min_time, | |
3646 | div_long_long_rem(l->sum_time, l->count, &remainder), | |
3647 | l->max_time); | |
3648 | } else | |
3649 | len += sprintf(buf + len, " age=%ld", | |
3650 | l->min_time); | |
3651 | ||
3652 | if (l->min_pid != l->max_pid) | |
3653 | len += sprintf(buf + len, " pid=%ld-%ld", | |
3654 | l->min_pid, l->max_pid); | |
3655 | else | |
3656 | len += sprintf(buf + len, " pid=%ld", | |
3657 | l->min_pid); | |
3658 | ||
3659 | if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && | |
3660 | len < PAGE_SIZE - 60) { | |
3661 | len += sprintf(buf + len, " cpus="); | |
3662 | len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
3663 | l->cpus); | |
3664 | } | |
3665 | ||
3666 | if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && | |
3667 | len < PAGE_SIZE - 60) { | |
3668 | len += sprintf(buf + len, " nodes="); | |
3669 | len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
3670 | l->nodes); | |
3671 | } | |
3672 | ||
3673 | len += sprintf(buf + len, "\n"); | |
3674 | } | |
3675 | ||
3676 | free_loc_track(&t); | |
3677 | if (!t.count) | |
3678 | len += sprintf(buf, "No data\n"); | |
3679 | return len; | |
3680 | } | |
3681 | ||
3682 | enum slab_stat_type { | |
3683 | SL_ALL, /* All slabs */ | |
3684 | SL_PARTIAL, /* Only partially allocated slabs */ | |
3685 | SL_CPU, /* Only slabs used for cpu caches */ | |
3686 | SL_OBJECTS, /* Determine allocated objects not slabs */ | |
3687 | SL_TOTAL /* Determine object capacity not slabs */ | |
3688 | }; | |
3689 | ||
3690 | #define SO_ALL (1 << SL_ALL) | |
3691 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
3692 | #define SO_CPU (1 << SL_CPU) | |
3693 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
3694 | #define SO_TOTAL (1 << SL_TOTAL) | |
3695 | ||
3696 | static ssize_t show_slab_objects(struct kmem_cache *s, | |
3697 | char *buf, unsigned long flags) | |
3698 | { | |
3699 | unsigned long total = 0; | |
3700 | int node; | |
3701 | int x; | |
3702 | unsigned long *nodes; | |
3703 | unsigned long *per_cpu; | |
3704 | ||
3705 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
3706 | if (!nodes) | |
3707 | return -ENOMEM; | |
3708 | per_cpu = nodes + nr_node_ids; | |
3709 | ||
3710 | if (flags & SO_CPU) { | |
3711 | int cpu; | |
3712 | ||
3713 | for_each_possible_cpu(cpu) { | |
3714 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
3715 | ||
3716 | if (!c || c->node < 0) | |
3717 | continue; | |
3718 | ||
3719 | if (c->page) { | |
3720 | if (flags & SO_TOTAL) | |
3721 | x = c->page->objects; | |
3722 | else if (flags & SO_OBJECTS) | |
3723 | x = c->page->inuse; | |
3724 | else | |
3725 | x = 1; | |
3726 | ||
3727 | total += x; | |
3728 | nodes[c->node] += x; | |
3729 | } | |
3730 | per_cpu[c->node]++; | |
3731 | } | |
3732 | } | |
3733 | ||
3734 | if (flags & SO_ALL) { | |
3735 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3736 | struct kmem_cache_node *n = get_node(s, node); | |
3737 | ||
3738 | if (flags & SO_TOTAL) | |
3739 | x = atomic_long_read(&n->total_objects); | |
3740 | else if (flags & SO_OBJECTS) | |
3741 | x = atomic_long_read(&n->total_objects) - | |
3742 | count_partial(n, count_free); | |
3743 | ||
3744 | else | |
3745 | x = atomic_long_read(&n->nr_slabs); | |
3746 | total += x; | |
3747 | nodes[node] += x; | |
3748 | } | |
3749 | ||
3750 | } else if (flags & SO_PARTIAL) { | |
3751 | for_each_node_state(node, N_NORMAL_MEMORY) { | |
3752 | struct kmem_cache_node *n = get_node(s, node); | |
3753 | ||
3754 | if (flags & SO_TOTAL) | |
3755 | x = count_partial(n, count_total); | |
3756 | else if (flags & SO_OBJECTS) | |
3757 | x = count_partial(n, count_inuse); | |
3758 | else | |
3759 | x = n->nr_partial; | |
3760 | total += x; | |
3761 | nodes[node] += x; | |
3762 | } | |
3763 | } | |
3764 | x = sprintf(buf, "%lu", total); | |
3765 | #ifdef CONFIG_NUMA | |
3766 | for_each_node_state(node, N_NORMAL_MEMORY) | |
3767 | if (nodes[node]) | |
3768 | x += sprintf(buf + x, " N%d=%lu", | |
3769 | node, nodes[node]); | |
3770 | #endif | |
3771 | kfree(nodes); | |
3772 | return x + sprintf(buf + x, "\n"); | |
3773 | } | |
3774 | ||
3775 | static int any_slab_objects(struct kmem_cache *s) | |
3776 | { | |
3777 | int node; | |
3778 | ||
3779 | for_each_online_node(node) { | |
3780 | struct kmem_cache_node *n = get_node(s, node); | |
3781 | ||
3782 | if (!n) | |
3783 | continue; | |
3784 | ||
3785 | if (atomic_read(&n->total_objects)) | |
3786 | return 1; | |
3787 | } | |
3788 | return 0; | |
3789 | } | |
3790 | ||
3791 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
3792 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); | |
3793 | ||
3794 | struct slab_attribute { | |
3795 | struct attribute attr; | |
3796 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
3797 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
3798 | }; | |
3799 | ||
3800 | #define SLAB_ATTR_RO(_name) \ | |
3801 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) | |
3802 | ||
3803 | #define SLAB_ATTR(_name) \ | |
3804 | static struct slab_attribute _name##_attr = \ | |
3805 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3806 | ||
3807 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) | |
3808 | { | |
3809 | return sprintf(buf, "%d\n", s->size); | |
3810 | } | |
3811 | SLAB_ATTR_RO(slab_size); | |
3812 | ||
3813 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
3814 | { | |
3815 | return sprintf(buf, "%d\n", s->align); | |
3816 | } | |
3817 | SLAB_ATTR_RO(align); | |
3818 | ||
3819 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
3820 | { | |
3821 | return sprintf(buf, "%d\n", s->objsize); | |
3822 | } | |
3823 | SLAB_ATTR_RO(object_size); | |
3824 | ||
3825 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
3826 | { | |
3827 | return sprintf(buf, "%d\n", oo_objects(s->oo)); | |
3828 | } | |
3829 | SLAB_ATTR_RO(objs_per_slab); | |
3830 | ||
3831 | static ssize_t order_store(struct kmem_cache *s, | |
3832 | const char *buf, size_t length) | |
3833 | { | |
3834 | int order = simple_strtoul(buf, NULL, 10); | |
3835 | ||
3836 | if (order > slub_max_order || order < slub_min_order) | |
3837 | return -EINVAL; | |
3838 | ||
3839 | calculate_sizes(s, order); | |
3840 | return length; | |
3841 | } | |
3842 | ||
3843 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
3844 | { | |
3845 | return sprintf(buf, "%d\n", oo_order(s->oo)); | |
3846 | } | |
3847 | SLAB_ATTR(order); | |
3848 | ||
3849 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
3850 | { | |
3851 | if (s->ctor) { | |
3852 | int n = sprint_symbol(buf, (unsigned long)s->ctor); | |
3853 | ||
3854 | return n + sprintf(buf + n, "\n"); | |
3855 | } | |
3856 | return 0; | |
3857 | } | |
3858 | SLAB_ATTR_RO(ctor); | |
3859 | ||
3860 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) | |
3861 | { | |
3862 | return sprintf(buf, "%d\n", s->refcount - 1); | |
3863 | } | |
3864 | SLAB_ATTR_RO(aliases); | |
3865 | ||
3866 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
3867 | { | |
3868 | return show_slab_objects(s, buf, SO_ALL); | |
3869 | } | |
3870 | SLAB_ATTR_RO(slabs); | |
3871 | ||
3872 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
3873 | { | |
3874 | return show_slab_objects(s, buf, SO_PARTIAL); | |
3875 | } | |
3876 | SLAB_ATTR_RO(partial); | |
3877 | ||
3878 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
3879 | { | |
3880 | return show_slab_objects(s, buf, SO_CPU); | |
3881 | } | |
3882 | SLAB_ATTR_RO(cpu_slabs); | |
3883 | ||
3884 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
3885 | { | |
3886 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); | |
3887 | } | |
3888 | SLAB_ATTR_RO(objects); | |
3889 | ||
3890 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) | |
3891 | { | |
3892 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); | |
3893 | } | |
3894 | SLAB_ATTR_RO(objects_partial); | |
3895 | ||
3896 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) | |
3897 | { | |
3898 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); | |
3899 | } | |
3900 | SLAB_ATTR_RO(total_objects); | |
3901 | ||
3902 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
3903 | { | |
3904 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
3905 | } | |
3906 | ||
3907 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
3908 | const char *buf, size_t length) | |
3909 | { | |
3910 | s->flags &= ~SLAB_DEBUG_FREE; | |
3911 | if (buf[0] == '1') | |
3912 | s->flags |= SLAB_DEBUG_FREE; | |
3913 | return length; | |
3914 | } | |
3915 | SLAB_ATTR(sanity_checks); | |
3916 | ||
3917 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
3918 | { | |
3919 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
3920 | } | |
3921 | ||
3922 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
3923 | size_t length) | |
3924 | { | |
3925 | s->flags &= ~SLAB_TRACE; | |
3926 | if (buf[0] == '1') | |
3927 | s->flags |= SLAB_TRACE; | |
3928 | return length; | |
3929 | } | |
3930 | SLAB_ATTR(trace); | |
3931 | ||
3932 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
3933 | { | |
3934 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
3935 | } | |
3936 | ||
3937 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
3938 | const char *buf, size_t length) | |
3939 | { | |
3940 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
3941 | if (buf[0] == '1') | |
3942 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
3943 | return length; | |
3944 | } | |
3945 | SLAB_ATTR(reclaim_account); | |
3946 | ||
3947 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
3948 | { | |
3949 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); | |
3950 | } | |
3951 | SLAB_ATTR_RO(hwcache_align); | |
3952 | ||
3953 | #ifdef CONFIG_ZONE_DMA | |
3954 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
3955 | { | |
3956 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
3957 | } | |
3958 | SLAB_ATTR_RO(cache_dma); | |
3959 | #endif | |
3960 | ||
3961 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
3962 | { | |
3963 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
3964 | } | |
3965 | SLAB_ATTR_RO(destroy_by_rcu); | |
3966 | ||
3967 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
3968 | { | |
3969 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
3970 | } | |
3971 | ||
3972 | static ssize_t red_zone_store(struct kmem_cache *s, | |
3973 | const char *buf, size_t length) | |
3974 | { | |
3975 | if (any_slab_objects(s)) | |
3976 | return -EBUSY; | |
3977 | ||
3978 | s->flags &= ~SLAB_RED_ZONE; | |
3979 | if (buf[0] == '1') | |
3980 | s->flags |= SLAB_RED_ZONE; | |
3981 | calculate_sizes(s, -1); | |
3982 | return length; | |
3983 | } | |
3984 | SLAB_ATTR(red_zone); | |
3985 | ||
3986 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
3987 | { | |
3988 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
3989 | } | |
3990 | ||
3991 | static ssize_t poison_store(struct kmem_cache *s, | |
3992 | const char *buf, size_t length) | |
3993 | { | |
3994 | if (any_slab_objects(s)) | |
3995 | return -EBUSY; | |
3996 | ||
3997 | s->flags &= ~SLAB_POISON; | |
3998 | if (buf[0] == '1') | |
3999 | s->flags |= SLAB_POISON; | |
4000 | calculate_sizes(s, -1); | |
4001 | return length; | |
4002 | } | |
4003 | SLAB_ATTR(poison); | |
4004 | ||
4005 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
4006 | { | |
4007 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
4008 | } | |
4009 | ||
4010 | static ssize_t store_user_store(struct kmem_cache *s, | |
4011 | const char *buf, size_t length) | |
4012 | { | |
4013 | if (any_slab_objects(s)) | |
4014 | return -EBUSY; | |
4015 | ||
4016 | s->flags &= ~SLAB_STORE_USER; | |
4017 | if (buf[0] == '1') | |
4018 | s->flags |= SLAB_STORE_USER; | |
4019 | calculate_sizes(s, -1); | |
4020 | return length; | |
4021 | } | |
4022 | SLAB_ATTR(store_user); | |
4023 | ||
4024 | static ssize_t validate_show(struct kmem_cache *s, char *buf) | |
4025 | { | |
4026 | return 0; | |
4027 | } | |
4028 | ||
4029 | static ssize_t validate_store(struct kmem_cache *s, | |
4030 | const char *buf, size_t length) | |
4031 | { | |
4032 | int ret = -EINVAL; | |
4033 | ||
4034 | if (buf[0] == '1') { | |
4035 | ret = validate_slab_cache(s); | |
4036 | if (ret >= 0) | |
4037 | ret = length; | |
4038 | } | |
4039 | return ret; | |
4040 | } | |
4041 | SLAB_ATTR(validate); | |
4042 | ||
4043 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) | |
4044 | { | |
4045 | return 0; | |
4046 | } | |
4047 | ||
4048 | static ssize_t shrink_store(struct kmem_cache *s, | |
4049 | const char *buf, size_t length) | |
4050 | { | |
4051 | if (buf[0] == '1') { | |
4052 | int rc = kmem_cache_shrink(s); | |
4053 | ||
4054 | if (rc) | |
4055 | return rc; | |
4056 | } else | |
4057 | return -EINVAL; | |
4058 | return length; | |
4059 | } | |
4060 | SLAB_ATTR(shrink); | |
4061 | ||
4062 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) | |
4063 | { | |
4064 | if (!(s->flags & SLAB_STORE_USER)) | |
4065 | return -ENOSYS; | |
4066 | return list_locations(s, buf, TRACK_ALLOC); | |
4067 | } | |
4068 | SLAB_ATTR_RO(alloc_calls); | |
4069 | ||
4070 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
4071 | { | |
4072 | if (!(s->flags & SLAB_STORE_USER)) | |
4073 | return -ENOSYS; | |
4074 | return list_locations(s, buf, TRACK_FREE); | |
4075 | } | |
4076 | SLAB_ATTR_RO(free_calls); | |
4077 | ||
4078 | #ifdef CONFIG_NUMA | |
4079 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) | |
4080 | { | |
4081 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); | |
4082 | } | |
4083 | ||
4084 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, | |
4085 | const char *buf, size_t length) | |
4086 | { | |
4087 | int n = simple_strtoul(buf, NULL, 10); | |
4088 | ||
4089 | if (n < 100) | |
4090 | s->remote_node_defrag_ratio = n * 10; | |
4091 | return length; | |
4092 | } | |
4093 | SLAB_ATTR(remote_node_defrag_ratio); | |
4094 | #endif | |
4095 | ||
4096 | #ifdef CONFIG_SLUB_STATS | |
4097 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) | |
4098 | { | |
4099 | unsigned long sum = 0; | |
4100 | int cpu; | |
4101 | int len; | |
4102 | int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); | |
4103 | ||
4104 | if (!data) | |
4105 | return -ENOMEM; | |
4106 | ||
4107 | for_each_online_cpu(cpu) { | |
4108 | unsigned x = get_cpu_slab(s, cpu)->stat[si]; | |
4109 | ||
4110 | data[cpu] = x; | |
4111 | sum += x; | |
4112 | } | |
4113 | ||
4114 | len = sprintf(buf, "%lu", sum); | |
4115 | ||
4116 | #ifdef CONFIG_SMP | |
4117 | for_each_online_cpu(cpu) { | |
4118 | if (data[cpu] && len < PAGE_SIZE - 20) | |
4119 | len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); | |
4120 | } | |
4121 | #endif | |
4122 | kfree(data); | |
4123 | return len + sprintf(buf + len, "\n"); | |
4124 | } | |
4125 | ||
4126 | #define STAT_ATTR(si, text) \ | |
4127 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ | |
4128 | { \ | |
4129 | return show_stat(s, buf, si); \ | |
4130 | } \ | |
4131 | SLAB_ATTR_RO(text); \ | |
4132 | ||
4133 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); | |
4134 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); | |
4135 | STAT_ATTR(FREE_FASTPATH, free_fastpath); | |
4136 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); | |
4137 | STAT_ATTR(FREE_FROZEN, free_frozen); | |
4138 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); | |
4139 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); | |
4140 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); | |
4141 | STAT_ATTR(ALLOC_SLAB, alloc_slab); | |
4142 | STAT_ATTR(ALLOC_REFILL, alloc_refill); | |
4143 | STAT_ATTR(FREE_SLAB, free_slab); | |
4144 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); | |
4145 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); | |
4146 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); | |
4147 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); | |
4148 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); | |
4149 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); | |
4150 | STAT_ATTR(ORDER_FALLBACK, order_fallback); | |
4151 | #endif | |
4152 | ||
4153 | static struct attribute *slab_attrs[] = { | |
4154 | &slab_size_attr.attr, | |
4155 | &object_size_attr.attr, | |
4156 | &objs_per_slab_attr.attr, | |
4157 | &order_attr.attr, | |
4158 | &objects_attr.attr, | |
4159 | &objects_partial_attr.attr, | |
4160 | &total_objects_attr.attr, | |
4161 | &slabs_attr.attr, | |
4162 | &partial_attr.attr, | |
4163 | &cpu_slabs_attr.attr, | |
4164 | &ctor_attr.attr, | |
4165 | &aliases_attr.attr, | |
4166 | &align_attr.attr, | |
4167 | &sanity_checks_attr.attr, | |
4168 | &trace_attr.attr, | |
4169 | &hwcache_align_attr.attr, | |
4170 | &reclaim_account_attr.attr, | |
4171 | &destroy_by_rcu_attr.attr, | |
4172 | &red_zone_attr.attr, | |
4173 | &poison_attr.attr, | |
4174 | &store_user_attr.attr, | |
4175 | &validate_attr.attr, | |
4176 | &shrink_attr.attr, | |
4177 | &alloc_calls_attr.attr, | |
4178 | &free_calls_attr.attr, | |
4179 | #ifdef CONFIG_ZONE_DMA | |
4180 | &cache_dma_attr.attr, | |
4181 | #endif | |
4182 | #ifdef CONFIG_NUMA | |
4183 | &remote_node_defrag_ratio_attr.attr, | |
4184 | #endif | |
4185 | #ifdef CONFIG_SLUB_STATS | |
4186 | &alloc_fastpath_attr.attr, | |
4187 | &alloc_slowpath_attr.attr, | |
4188 | &free_fastpath_attr.attr, | |
4189 | &free_slowpath_attr.attr, | |
4190 | &free_frozen_attr.attr, | |
4191 | &free_add_partial_attr.attr, | |
4192 | &free_remove_partial_attr.attr, | |
4193 | &alloc_from_partial_attr.attr, | |
4194 | &alloc_slab_attr.attr, | |
4195 | &alloc_refill_attr.attr, | |
4196 | &free_slab_attr.attr, | |
4197 | &cpuslab_flush_attr.attr, | |
4198 | &deactivate_full_attr.attr, | |
4199 | &deactivate_empty_attr.attr, | |
4200 | &deactivate_to_head_attr.attr, | |
4201 | &deactivate_to_tail_attr.attr, | |
4202 | &deactivate_remote_frees_attr.attr, | |
4203 | &order_fallback_attr.attr, | |
4204 | #endif | |
4205 | NULL | |
4206 | }; | |
4207 | ||
4208 | static struct attribute_group slab_attr_group = { | |
4209 | .attrs = slab_attrs, | |
4210 | }; | |
4211 | ||
4212 | static ssize_t slab_attr_show(struct kobject *kobj, | |
4213 | struct attribute *attr, | |
4214 | char *buf) | |
4215 | { | |
4216 | struct slab_attribute *attribute; | |
4217 | struct kmem_cache *s; | |
4218 | int err; | |
4219 | ||
4220 | attribute = to_slab_attr(attr); | |
4221 | s = to_slab(kobj); | |
4222 | ||
4223 | if (!attribute->show) | |
4224 | return -EIO; | |
4225 | ||
4226 | err = attribute->show(s, buf); | |
4227 | ||
4228 | return err; | |
4229 | } | |
4230 | ||
4231 | static ssize_t slab_attr_store(struct kobject *kobj, | |
4232 | struct attribute *attr, | |
4233 | const char *buf, size_t len) | |
4234 | { | |
4235 | struct slab_attribute *attribute; | |
4236 | struct kmem_cache *s; | |
4237 | int err; | |
4238 | ||
4239 | attribute = to_slab_attr(attr); | |
4240 | s = to_slab(kobj); | |
4241 | ||
4242 | if (!attribute->store) | |
4243 | return -EIO; | |
4244 | ||
4245 | err = attribute->store(s, buf, len); | |
4246 | ||
4247 | return err; | |
4248 | } | |
4249 | ||
4250 | static void kmem_cache_release(struct kobject *kobj) | |
4251 | { | |
4252 | struct kmem_cache *s = to_slab(kobj); | |
4253 | ||
4254 | kfree(s); | |
4255 | } | |
4256 | ||
4257 | static struct sysfs_ops slab_sysfs_ops = { | |
4258 | .show = slab_attr_show, | |
4259 | .store = slab_attr_store, | |
4260 | }; | |
4261 | ||
4262 | static struct kobj_type slab_ktype = { | |
4263 | .sysfs_ops = &slab_sysfs_ops, | |
4264 | .release = kmem_cache_release | |
4265 | }; | |
4266 | ||
4267 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
4268 | { | |
4269 | struct kobj_type *ktype = get_ktype(kobj); | |
4270 | ||
4271 | if (ktype == &slab_ktype) | |
4272 | return 1; | |
4273 | return 0; | |
4274 | } | |
4275 | ||
4276 | static struct kset_uevent_ops slab_uevent_ops = { | |
4277 | .filter = uevent_filter, | |
4278 | }; | |
4279 | ||
4280 | static struct kset *slab_kset; | |
4281 | ||
4282 | #define ID_STR_LENGTH 64 | |
4283 | ||
4284 | /* Create a unique string id for a slab cache: | |
4285 | * | |
4286 | * Format :[flags-]size | |
4287 | */ | |
4288 | static char *create_unique_id(struct kmem_cache *s) | |
4289 | { | |
4290 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
4291 | char *p = name; | |
4292 | ||
4293 | BUG_ON(!name); | |
4294 | ||
4295 | *p++ = ':'; | |
4296 | /* | |
4297 | * First flags affecting slabcache operations. We will only | |
4298 | * get here for aliasable slabs so we do not need to support | |
4299 | * too many flags. The flags here must cover all flags that | |
4300 | * are matched during merging to guarantee that the id is | |
4301 | * unique. | |
4302 | */ | |
4303 | if (s->flags & SLAB_CACHE_DMA) | |
4304 | *p++ = 'd'; | |
4305 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
4306 | *p++ = 'a'; | |
4307 | if (s->flags & SLAB_DEBUG_FREE) | |
4308 | *p++ = 'F'; | |
4309 | if (p != name + 1) | |
4310 | *p++ = '-'; | |
4311 | p += sprintf(p, "%07d", s->size); | |
4312 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
4313 | return name; | |
4314 | } | |
4315 | ||
4316 | static int sysfs_slab_add(struct kmem_cache *s) | |
4317 | { | |
4318 | int err; | |
4319 | const char *name; | |
4320 | int unmergeable; | |
4321 | ||
4322 | if (slab_state < SYSFS) | |
4323 | /* Defer until later */ | |
4324 | return 0; | |
4325 | ||
4326 | unmergeable = slab_unmergeable(s); | |
4327 | if (unmergeable) { | |
4328 | /* | |
4329 | * Slabcache can never be merged so we can use the name proper. | |
4330 | * This is typically the case for debug situations. In that | |
4331 | * case we can catch duplicate names easily. | |
4332 | */ | |
4333 | sysfs_remove_link(&slab_kset->kobj, s->name); | |
4334 | name = s->name; | |
4335 | } else { | |
4336 | /* | |
4337 | * Create a unique name for the slab as a target | |
4338 | * for the symlinks. | |
4339 | */ | |
4340 | name = create_unique_id(s); | |
4341 | } | |
4342 | ||
4343 | s->kobj.kset = slab_kset; | |
4344 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); | |
4345 | if (err) { | |
4346 | kobject_put(&s->kobj); | |
4347 | return err; | |
4348 | } | |
4349 | ||
4350 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
4351 | if (err) | |
4352 | return err; | |
4353 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
4354 | if (!unmergeable) { | |
4355 | /* Setup first alias */ | |
4356 | sysfs_slab_alias(s, s->name); | |
4357 | kfree(name); | |
4358 | } | |
4359 | return 0; | |
4360 | } | |
4361 | ||
4362 | static void sysfs_slab_remove(struct kmem_cache *s) | |
4363 | { | |
4364 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
4365 | kobject_del(&s->kobj); | |
4366 | kobject_put(&s->kobj); | |
4367 | } | |
4368 | ||
4369 | /* | |
4370 | * Need to buffer aliases during bootup until sysfs becomes | |
4371 | * available lest we loose that information. | |
4372 | */ | |
4373 | struct saved_alias { | |
4374 | struct kmem_cache *s; | |
4375 | const char *name; | |
4376 | struct saved_alias *next; | |
4377 | }; | |
4378 | ||
4379 | static struct saved_alias *alias_list; | |
4380 | ||
4381 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
4382 | { | |
4383 | struct saved_alias *al; | |
4384 | ||
4385 | if (slab_state == SYSFS) { | |
4386 | /* | |
4387 | * If we have a leftover link then remove it. | |
4388 | */ | |
4389 | sysfs_remove_link(&slab_kset->kobj, name); | |
4390 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); | |
4391 | } | |
4392 | ||
4393 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
4394 | if (!al) | |
4395 | return -ENOMEM; | |
4396 | ||
4397 | al->s = s; | |
4398 | al->name = name; | |
4399 | al->next = alias_list; | |
4400 | alias_list = al; | |
4401 | return 0; | |
4402 | } | |
4403 | ||
4404 | static int __init slab_sysfs_init(void) | |
4405 | { | |
4406 | struct kmem_cache *s; | |
4407 | int err; | |
4408 | ||
4409 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); | |
4410 | if (!slab_kset) { | |
4411 | printk(KERN_ERR "Cannot register slab subsystem.\n"); | |
4412 | return -ENOSYS; | |
4413 | } | |
4414 | ||
4415 | slab_state = SYSFS; | |
4416 | ||
4417 | list_for_each_entry(s, &slab_caches, list) { | |
4418 | err = sysfs_slab_add(s); | |
4419 | if (err) | |
4420 | printk(KERN_ERR "SLUB: Unable to add boot slab %s" | |
4421 | " to sysfs\n", s->name); | |
4422 | } | |
4423 | ||
4424 | while (alias_list) { | |
4425 | struct saved_alias *al = alias_list; | |
4426 | ||
4427 | alias_list = alias_list->next; | |
4428 | err = sysfs_slab_alias(al->s, al->name); | |
4429 | if (err) | |
4430 | printk(KERN_ERR "SLUB: Unable to add boot slab alias" | |
4431 | " %s to sysfs\n", s->name); | |
4432 | kfree(al); | |
4433 | } | |
4434 | ||
4435 | resiliency_test(); | |
4436 | return 0; | |
4437 | } | |
4438 | ||
4439 | __initcall(slab_sysfs_init); | |
4440 | #endif | |
4441 | ||
4442 | /* | |
4443 | * The /proc/slabinfo ABI | |
4444 | */ | |
4445 | #ifdef CONFIG_SLABINFO | |
4446 | ||
4447 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, | |
4448 | size_t count, loff_t *ppos) | |
4449 | { | |
4450 | return -EINVAL; | |
4451 | } | |
4452 | ||
4453 | ||
4454 | static void print_slabinfo_header(struct seq_file *m) | |
4455 | { | |
4456 | seq_puts(m, "slabinfo - version: 2.1\n"); | |
4457 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " | |
4458 | "<objperslab> <pagesperslab>"); | |
4459 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4460 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
4461 | seq_putc(m, '\n'); | |
4462 | } | |
4463 | ||
4464 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4465 | { | |
4466 | loff_t n = *pos; | |
4467 | ||
4468 | down_read(&slub_lock); | |
4469 | if (!n) | |
4470 | print_slabinfo_header(m); | |
4471 | ||
4472 | return seq_list_start(&slab_caches, *pos); | |
4473 | } | |
4474 | ||
4475 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4476 | { | |
4477 | return seq_list_next(p, &slab_caches, pos); | |
4478 | } | |
4479 | ||
4480 | static void s_stop(struct seq_file *m, void *p) | |
4481 | { | |
4482 | up_read(&slub_lock); | |
4483 | } | |
4484 | ||
4485 | static int s_show(struct seq_file *m, void *p) | |
4486 | { | |
4487 | unsigned long nr_partials = 0; | |
4488 | unsigned long nr_slabs = 0; | |
4489 | unsigned long nr_inuse = 0; | |
4490 | unsigned long nr_objs = 0; | |
4491 | unsigned long nr_free = 0; | |
4492 | struct kmem_cache *s; | |
4493 | int node; | |
4494 | ||
4495 | s = list_entry(p, struct kmem_cache, list); | |
4496 | ||
4497 | for_each_online_node(node) { | |
4498 | struct kmem_cache_node *n = get_node(s, node); | |
4499 | ||
4500 | if (!n) | |
4501 | continue; | |
4502 | ||
4503 | nr_partials += n->nr_partial; | |
4504 | nr_slabs += atomic_long_read(&n->nr_slabs); | |
4505 | nr_objs += atomic_long_read(&n->total_objects); | |
4506 | nr_free += count_partial(n, count_free); | |
4507 | } | |
4508 | ||
4509 | nr_inuse = nr_objs - nr_free; | |
4510 | ||
4511 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, | |
4512 | nr_objs, s->size, oo_objects(s->oo), | |
4513 | (1 << oo_order(s->oo))); | |
4514 | seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); | |
4515 | seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, | |
4516 | 0UL); | |
4517 | seq_putc(m, '\n'); | |
4518 | return 0; | |
4519 | } | |
4520 | ||
4521 | const struct seq_operations slabinfo_op = { | |
4522 | .start = s_start, | |
4523 | .next = s_next, | |
4524 | .stop = s_stop, | |
4525 | .show = s_show, | |
4526 | }; | |
4527 | ||
4528 | #endif /* CONFIG_SLABINFO */ |