Commit | Line | Data |
---|---|---|
81819f0f CL |
1 | /* |
2 | * SLUB: A slab allocator that limits cache line use instead of queuing | |
3 | * objects in per cpu and per node lists. | |
4 | * | |
5 | * The allocator synchronizes using per slab locks and only | |
6 | * uses a centralized lock to manage a pool of partial slabs. | |
7 | * | |
8 | * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com> | |
9 | */ | |
10 | ||
11 | #include <linux/mm.h> | |
12 | #include <linux/module.h> | |
13 | #include <linux/bit_spinlock.h> | |
14 | #include <linux/interrupt.h> | |
15 | #include <linux/bitops.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/seq_file.h> | |
18 | #include <linux/cpu.h> | |
19 | #include <linux/cpuset.h> | |
20 | #include <linux/mempolicy.h> | |
21 | #include <linux/ctype.h> | |
22 | #include <linux/kallsyms.h> | |
b9049e23 | 23 | #include <linux/memory.h> |
81819f0f CL |
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 | * | |
672bba3a CL |
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 | |
81819f0f | 72 | * freed then the slab will show up again on the partial lists. |
672bba3a CL |
73 | * We track full slabs for debugging purposes though because otherwise we |
74 | * cannot scan all objects. | |
81819f0f CL |
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 | * | |
4b6f0750 CL |
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 | |
dfb4f096 | 94 | * freelist that allows lockless access to |
894b8788 CL |
95 | * free objects in addition to the regular freelist |
96 | * that requires the slab lock. | |
81819f0f CL |
97 | * |
98 | * PageError Slab requires special handling due to debug | |
99 | * options set. This moves slab handling out of | |
894b8788 | 100 | * the fast path and disables lockless freelists. |
81819f0f CL |
101 | */ |
102 | ||
5577bd8a CL |
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 | ||
4b6f0750 CL |
111 | static inline int SlabFrozen(struct page *page) |
112 | { | |
5577bd8a | 113 | return page->flags & FROZEN; |
4b6f0750 CL |
114 | } |
115 | ||
116 | static inline void SetSlabFrozen(struct page *page) | |
117 | { | |
5577bd8a | 118 | page->flags |= FROZEN; |
4b6f0750 CL |
119 | } |
120 | ||
121 | static inline void ClearSlabFrozen(struct page *page) | |
122 | { | |
5577bd8a | 123 | page->flags &= ~FROZEN; |
4b6f0750 CL |
124 | } |
125 | ||
35e5d7ee CL |
126 | static inline int SlabDebug(struct page *page) |
127 | { | |
5577bd8a | 128 | return page->flags & SLABDEBUG; |
35e5d7ee CL |
129 | } |
130 | ||
131 | static inline void SetSlabDebug(struct page *page) | |
132 | { | |
5577bd8a | 133 | page->flags |= SLABDEBUG; |
35e5d7ee CL |
134 | } |
135 | ||
136 | static inline void ClearSlabDebug(struct page *page) | |
137 | { | |
5577bd8a | 138 | page->flags &= ~SLABDEBUG; |
35e5d7ee CL |
139 | } |
140 | ||
81819f0f CL |
141 | /* |
142 | * Issues still to be resolved: | |
143 | * | |
81819f0f CL |
144 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. |
145 | * | |
81819f0f CL |
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 | ||
2086d26a CL |
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 | */ | |
76be8950 | 175 | #define MIN_PARTIAL 5 |
e95eed57 | 176 | |
2086d26a CL |
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 | ||
81819f0f CL |
184 | #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ |
185 | SLAB_POISON | SLAB_STORE_USER) | |
672bba3a | 186 | |
81819f0f CL |
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 | |
47bfdc0d | 197 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
198 | #endif |
199 | ||
200 | #ifndef ARCH_SLAB_MINALIGN | |
47bfdc0d | 201 | #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) |
81819f0f CL |
202 | #endif |
203 | ||
204 | /* Internal SLUB flags */ | |
1ceef402 CL |
205 | #define __OBJECT_POISON 0x80000000 /* Poison object */ |
206 | #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ | |
81819f0f | 207 | |
65c02d4c CL |
208 | /* Not all arches define cache_line_size */ |
209 | #ifndef cache_line_size | |
210 | #define cache_line_size() L1_CACHE_BYTES | |
211 | #endif | |
212 | ||
81819f0f CL |
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 */ | |
672bba3a | 222 | UP, /* Everything works but does not show up in sysfs */ |
81819f0f CL |
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); | |
5af328a5 | 228 | static LIST_HEAD(slab_caches); |
81819f0f | 229 | |
02cbc874 CL |
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 | ||
41ecc55b | 242 | #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) |
81819f0f CL |
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 | #else | |
0c710013 CL |
247 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } |
248 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) | |
249 | { return 0; } | |
151c602f CL |
250 | static inline void sysfs_slab_remove(struct kmem_cache *s) |
251 | { | |
252 | kfree(s); | |
253 | } | |
81819f0f CL |
254 | #endif |
255 | ||
256 | /******************************************************************** | |
257 | * Core slab cache functions | |
258 | *******************************************************************/ | |
259 | ||
260 | int slab_is_available(void) | |
261 | { | |
262 | return slab_state >= UP; | |
263 | } | |
264 | ||
265 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) | |
266 | { | |
267 | #ifdef CONFIG_NUMA | |
268 | return s->node[node]; | |
269 | #else | |
270 | return &s->local_node; | |
271 | #endif | |
272 | } | |
273 | ||
dfb4f096 CL |
274 | static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) |
275 | { | |
4c93c355 CL |
276 | #ifdef CONFIG_SMP |
277 | return s->cpu_slab[cpu]; | |
278 | #else | |
279 | return &s->cpu_slab; | |
280 | #endif | |
dfb4f096 CL |
281 | } |
282 | ||
02cbc874 CL |
283 | static inline int check_valid_pointer(struct kmem_cache *s, |
284 | struct page *page, const void *object) | |
285 | { | |
286 | void *base; | |
287 | ||
288 | if (!object) | |
289 | return 1; | |
290 | ||
291 | base = page_address(page); | |
292 | if (object < base || object >= base + s->objects * s->size || | |
293 | (object - base) % s->size) { | |
294 | return 0; | |
295 | } | |
296 | ||
297 | return 1; | |
298 | } | |
299 | ||
7656c72b CL |
300 | /* |
301 | * Slow version of get and set free pointer. | |
302 | * | |
303 | * This version requires touching the cache lines of kmem_cache which | |
304 | * we avoid to do in the fast alloc free paths. There we obtain the offset | |
305 | * from the page struct. | |
306 | */ | |
307 | static inline void *get_freepointer(struct kmem_cache *s, void *object) | |
308 | { | |
309 | return *(void **)(object + s->offset); | |
310 | } | |
311 | ||
312 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) | |
313 | { | |
314 | *(void **)(object + s->offset) = fp; | |
315 | } | |
316 | ||
317 | /* Loop over all objects in a slab */ | |
318 | #define for_each_object(__p, __s, __addr) \ | |
319 | for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\ | |
320 | __p += (__s)->size) | |
321 | ||
322 | /* Scan freelist */ | |
323 | #define for_each_free_object(__p, __s, __free) \ | |
324 | for (__p = (__free); __p; __p = get_freepointer((__s), __p)) | |
325 | ||
326 | /* Determine object index from a given position */ | |
327 | static inline int slab_index(void *p, struct kmem_cache *s, void *addr) | |
328 | { | |
329 | return (p - addr) / s->size; | |
330 | } | |
331 | ||
41ecc55b CL |
332 | #ifdef CONFIG_SLUB_DEBUG |
333 | /* | |
334 | * Debug settings: | |
335 | */ | |
f0630fff CL |
336 | #ifdef CONFIG_SLUB_DEBUG_ON |
337 | static int slub_debug = DEBUG_DEFAULT_FLAGS; | |
338 | #else | |
41ecc55b | 339 | static int slub_debug; |
f0630fff | 340 | #endif |
41ecc55b CL |
341 | |
342 | static char *slub_debug_slabs; | |
343 | ||
81819f0f CL |
344 | /* |
345 | * Object debugging | |
346 | */ | |
347 | static void print_section(char *text, u8 *addr, unsigned int length) | |
348 | { | |
349 | int i, offset; | |
350 | int newline = 1; | |
351 | char ascii[17]; | |
352 | ||
353 | ascii[16] = 0; | |
354 | ||
355 | for (i = 0; i < length; i++) { | |
356 | if (newline) { | |
24922684 | 357 | printk(KERN_ERR "%8s 0x%p: ", text, addr + i); |
81819f0f CL |
358 | newline = 0; |
359 | } | |
360 | printk(" %02x", addr[i]); | |
361 | offset = i % 16; | |
362 | ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; | |
363 | if (offset == 15) { | |
364 | printk(" %s\n",ascii); | |
365 | newline = 1; | |
366 | } | |
367 | } | |
368 | if (!newline) { | |
369 | i %= 16; | |
370 | while (i < 16) { | |
371 | printk(" "); | |
372 | ascii[i] = ' '; | |
373 | i++; | |
374 | } | |
375 | printk(" %s\n", ascii); | |
376 | } | |
377 | } | |
378 | ||
81819f0f CL |
379 | static struct track *get_track(struct kmem_cache *s, void *object, |
380 | enum track_item alloc) | |
381 | { | |
382 | struct track *p; | |
383 | ||
384 | if (s->offset) | |
385 | p = object + s->offset + sizeof(void *); | |
386 | else | |
387 | p = object + s->inuse; | |
388 | ||
389 | return p + alloc; | |
390 | } | |
391 | ||
392 | static void set_track(struct kmem_cache *s, void *object, | |
393 | enum track_item alloc, void *addr) | |
394 | { | |
395 | struct track *p; | |
396 | ||
397 | if (s->offset) | |
398 | p = object + s->offset + sizeof(void *); | |
399 | else | |
400 | p = object + s->inuse; | |
401 | ||
402 | p += alloc; | |
403 | if (addr) { | |
404 | p->addr = addr; | |
405 | p->cpu = smp_processor_id(); | |
406 | p->pid = current ? current->pid : -1; | |
407 | p->when = jiffies; | |
408 | } else | |
409 | memset(p, 0, sizeof(struct track)); | |
410 | } | |
411 | ||
81819f0f CL |
412 | static void init_tracking(struct kmem_cache *s, void *object) |
413 | { | |
24922684 CL |
414 | if (!(s->flags & SLAB_STORE_USER)) |
415 | return; | |
416 | ||
417 | set_track(s, object, TRACK_FREE, NULL); | |
418 | set_track(s, object, TRACK_ALLOC, NULL); | |
81819f0f CL |
419 | } |
420 | ||
421 | static void print_track(const char *s, struct track *t) | |
422 | { | |
423 | if (!t->addr) | |
424 | return; | |
425 | ||
24922684 | 426 | printk(KERN_ERR "INFO: %s in ", s); |
81819f0f | 427 | __print_symbol("%s", (unsigned long)t->addr); |
24922684 CL |
428 | printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); |
429 | } | |
430 | ||
431 | static void print_tracking(struct kmem_cache *s, void *object) | |
432 | { | |
433 | if (!(s->flags & SLAB_STORE_USER)) | |
434 | return; | |
435 | ||
436 | print_track("Allocated", get_track(s, object, TRACK_ALLOC)); | |
437 | print_track("Freed", get_track(s, object, TRACK_FREE)); | |
438 | } | |
439 | ||
440 | static void print_page_info(struct page *page) | |
441 | { | |
442 | printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n", | |
443 | page, page->inuse, page->freelist, page->flags); | |
444 | ||
445 | } | |
446 | ||
447 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) | |
448 | { | |
449 | va_list args; | |
450 | char buf[100]; | |
451 | ||
452 | va_start(args, fmt); | |
453 | vsnprintf(buf, sizeof(buf), fmt, args); | |
454 | va_end(args); | |
455 | printk(KERN_ERR "========================================" | |
456 | "=====================================\n"); | |
457 | printk(KERN_ERR "BUG %s: %s\n", s->name, buf); | |
458 | printk(KERN_ERR "----------------------------------------" | |
459 | "-------------------------------------\n\n"); | |
81819f0f CL |
460 | } |
461 | ||
24922684 CL |
462 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) |
463 | { | |
464 | va_list args; | |
465 | char buf[100]; | |
466 | ||
467 | va_start(args, fmt); | |
468 | vsnprintf(buf, sizeof(buf), fmt, args); | |
469 | va_end(args); | |
470 | printk(KERN_ERR "FIX %s: %s\n", s->name, buf); | |
471 | } | |
472 | ||
473 | static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) | |
81819f0f CL |
474 | { |
475 | unsigned int off; /* Offset of last byte */ | |
24922684 CL |
476 | u8 *addr = page_address(page); |
477 | ||
478 | print_tracking(s, p); | |
479 | ||
480 | print_page_info(page); | |
481 | ||
482 | printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", | |
483 | p, p - addr, get_freepointer(s, p)); | |
484 | ||
485 | if (p > addr + 16) | |
486 | print_section("Bytes b4", p - 16, 16); | |
487 | ||
488 | print_section("Object", p, min(s->objsize, 128)); | |
81819f0f CL |
489 | |
490 | if (s->flags & SLAB_RED_ZONE) | |
491 | print_section("Redzone", p + s->objsize, | |
492 | s->inuse - s->objsize); | |
493 | ||
81819f0f CL |
494 | if (s->offset) |
495 | off = s->offset + sizeof(void *); | |
496 | else | |
497 | off = s->inuse; | |
498 | ||
24922684 | 499 | if (s->flags & SLAB_STORE_USER) |
81819f0f | 500 | off += 2 * sizeof(struct track); |
81819f0f CL |
501 | |
502 | if (off != s->size) | |
503 | /* Beginning of the filler is the free pointer */ | |
24922684 CL |
504 | print_section("Padding", p + off, s->size - off); |
505 | ||
506 | dump_stack(); | |
81819f0f CL |
507 | } |
508 | ||
509 | static void object_err(struct kmem_cache *s, struct page *page, | |
510 | u8 *object, char *reason) | |
511 | { | |
24922684 CL |
512 | slab_bug(s, reason); |
513 | print_trailer(s, page, object); | |
81819f0f CL |
514 | } |
515 | ||
24922684 | 516 | static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) |
81819f0f CL |
517 | { |
518 | va_list args; | |
519 | char buf[100]; | |
520 | ||
24922684 CL |
521 | va_start(args, fmt); |
522 | vsnprintf(buf, sizeof(buf), fmt, args); | |
81819f0f | 523 | va_end(args); |
24922684 CL |
524 | slab_bug(s, fmt); |
525 | print_page_info(page); | |
81819f0f CL |
526 | dump_stack(); |
527 | } | |
528 | ||
529 | static void init_object(struct kmem_cache *s, void *object, int active) | |
530 | { | |
531 | u8 *p = object; | |
532 | ||
533 | if (s->flags & __OBJECT_POISON) { | |
534 | memset(p, POISON_FREE, s->objsize - 1); | |
535 | p[s->objsize -1] = POISON_END; | |
536 | } | |
537 | ||
538 | if (s->flags & SLAB_RED_ZONE) | |
539 | memset(p + s->objsize, | |
540 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, | |
541 | s->inuse - s->objsize); | |
542 | } | |
543 | ||
24922684 | 544 | static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) |
81819f0f CL |
545 | { |
546 | while (bytes) { | |
547 | if (*start != (u8)value) | |
24922684 | 548 | return start; |
81819f0f CL |
549 | start++; |
550 | bytes--; | |
551 | } | |
24922684 CL |
552 | return NULL; |
553 | } | |
554 | ||
555 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, | |
556 | void *from, void *to) | |
557 | { | |
558 | slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); | |
559 | memset(from, data, to - from); | |
560 | } | |
561 | ||
562 | static int check_bytes_and_report(struct kmem_cache *s, struct page *page, | |
563 | u8 *object, char *what, | |
564 | u8* start, unsigned int value, unsigned int bytes) | |
565 | { | |
566 | u8 *fault; | |
567 | u8 *end; | |
568 | ||
569 | fault = check_bytes(start, value, bytes); | |
570 | if (!fault) | |
571 | return 1; | |
572 | ||
573 | end = start + bytes; | |
574 | while (end > fault && end[-1] == value) | |
575 | end--; | |
576 | ||
577 | slab_bug(s, "%s overwritten", what); | |
578 | printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", | |
579 | fault, end - 1, fault[0], value); | |
580 | print_trailer(s, page, object); | |
581 | ||
582 | restore_bytes(s, what, value, fault, end); | |
583 | return 0; | |
81819f0f CL |
584 | } |
585 | ||
81819f0f CL |
586 | /* |
587 | * Object layout: | |
588 | * | |
589 | * object address | |
590 | * Bytes of the object to be managed. | |
591 | * If the freepointer may overlay the object then the free | |
592 | * pointer is the first word of the object. | |
672bba3a | 593 | * |
81819f0f CL |
594 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
595 | * 0xa5 (POISON_END) | |
596 | * | |
597 | * object + s->objsize | |
598 | * Padding to reach word boundary. This is also used for Redzoning. | |
672bba3a CL |
599 | * Padding is extended by another word if Redzoning is enabled and |
600 | * objsize == inuse. | |
601 | * | |
81819f0f CL |
602 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
603 | * 0xcc (RED_ACTIVE) for objects in use. | |
604 | * | |
605 | * object + s->inuse | |
672bba3a CL |
606 | * Meta data starts here. |
607 | * | |
81819f0f CL |
608 | * A. Free pointer (if we cannot overwrite object on free) |
609 | * B. Tracking data for SLAB_STORE_USER | |
672bba3a CL |
610 | * C. Padding to reach required alignment boundary or at mininum |
611 | * one word if debuggin is on to be able to detect writes | |
612 | * before the word boundary. | |
613 | * | |
614 | * Padding is done using 0x5a (POISON_INUSE) | |
81819f0f CL |
615 | * |
616 | * object + s->size | |
672bba3a | 617 | * Nothing is used beyond s->size. |
81819f0f | 618 | * |
672bba3a CL |
619 | * If slabcaches are merged then the objsize and inuse boundaries are mostly |
620 | * ignored. And therefore no slab options that rely on these boundaries | |
81819f0f CL |
621 | * may be used with merged slabcaches. |
622 | */ | |
623 | ||
81819f0f CL |
624 | static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) |
625 | { | |
626 | unsigned long off = s->inuse; /* The end of info */ | |
627 | ||
628 | if (s->offset) | |
629 | /* Freepointer is placed after the object. */ | |
630 | off += sizeof(void *); | |
631 | ||
632 | if (s->flags & SLAB_STORE_USER) | |
633 | /* We also have user information there */ | |
634 | off += 2 * sizeof(struct track); | |
635 | ||
636 | if (s->size == off) | |
637 | return 1; | |
638 | ||
24922684 CL |
639 | return check_bytes_and_report(s, page, p, "Object padding", |
640 | p + off, POISON_INUSE, s->size - off); | |
81819f0f CL |
641 | } |
642 | ||
643 | static int slab_pad_check(struct kmem_cache *s, struct page *page) | |
644 | { | |
24922684 CL |
645 | u8 *start; |
646 | u8 *fault; | |
647 | u8 *end; | |
648 | int length; | |
649 | int remainder; | |
81819f0f CL |
650 | |
651 | if (!(s->flags & SLAB_POISON)) | |
652 | return 1; | |
653 | ||
24922684 CL |
654 | start = page_address(page); |
655 | end = start + (PAGE_SIZE << s->order); | |
81819f0f | 656 | length = s->objects * s->size; |
24922684 | 657 | remainder = end - (start + length); |
81819f0f CL |
658 | if (!remainder) |
659 | return 1; | |
660 | ||
24922684 CL |
661 | fault = check_bytes(start + length, POISON_INUSE, remainder); |
662 | if (!fault) | |
663 | return 1; | |
664 | while (end > fault && end[-1] == POISON_INUSE) | |
665 | end--; | |
666 | ||
667 | slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); | |
668 | print_section("Padding", start, length); | |
669 | ||
670 | restore_bytes(s, "slab padding", POISON_INUSE, start, end); | |
671 | return 0; | |
81819f0f CL |
672 | } |
673 | ||
674 | static int check_object(struct kmem_cache *s, struct page *page, | |
675 | void *object, int active) | |
676 | { | |
677 | u8 *p = object; | |
678 | u8 *endobject = object + s->objsize; | |
679 | ||
680 | if (s->flags & SLAB_RED_ZONE) { | |
681 | unsigned int red = | |
682 | active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; | |
683 | ||
24922684 CL |
684 | if (!check_bytes_and_report(s, page, object, "Redzone", |
685 | endobject, red, s->inuse - s->objsize)) | |
81819f0f | 686 | return 0; |
81819f0f | 687 | } else { |
24922684 CL |
688 | if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) |
689 | check_bytes_and_report(s, page, p, "Alignment padding", endobject, | |
690 | POISON_INUSE, s->inuse - s->objsize); | |
81819f0f CL |
691 | } |
692 | ||
693 | if (s->flags & SLAB_POISON) { | |
694 | if (!active && (s->flags & __OBJECT_POISON) && | |
24922684 CL |
695 | (!check_bytes_and_report(s, page, p, "Poison", p, |
696 | POISON_FREE, s->objsize - 1) || | |
697 | !check_bytes_and_report(s, page, p, "Poison", | |
698 | p + s->objsize -1, POISON_END, 1))) | |
81819f0f | 699 | return 0; |
81819f0f CL |
700 | /* |
701 | * check_pad_bytes cleans up on its own. | |
702 | */ | |
703 | check_pad_bytes(s, page, p); | |
704 | } | |
705 | ||
706 | if (!s->offset && active) | |
707 | /* | |
708 | * Object and freepointer overlap. Cannot check | |
709 | * freepointer while object is allocated. | |
710 | */ | |
711 | return 1; | |
712 | ||
713 | /* Check free pointer validity */ | |
714 | if (!check_valid_pointer(s, page, get_freepointer(s, p))) { | |
715 | object_err(s, page, p, "Freepointer corrupt"); | |
716 | /* | |
717 | * No choice but to zap it and thus loose the remainder | |
718 | * of the free objects in this slab. May cause | |
672bba3a | 719 | * another error because the object count is now wrong. |
81819f0f CL |
720 | */ |
721 | set_freepointer(s, p, NULL); | |
722 | return 0; | |
723 | } | |
724 | return 1; | |
725 | } | |
726 | ||
727 | static int check_slab(struct kmem_cache *s, struct page *page) | |
728 | { | |
729 | VM_BUG_ON(!irqs_disabled()); | |
730 | ||
731 | if (!PageSlab(page)) { | |
24922684 | 732 | slab_err(s, page, "Not a valid slab page"); |
81819f0f CL |
733 | return 0; |
734 | } | |
81819f0f | 735 | if (page->inuse > s->objects) { |
24922684 CL |
736 | slab_err(s, page, "inuse %u > max %u", |
737 | s->name, page->inuse, s->objects); | |
81819f0f CL |
738 | return 0; |
739 | } | |
740 | /* Slab_pad_check fixes things up after itself */ | |
741 | slab_pad_check(s, page); | |
742 | return 1; | |
743 | } | |
744 | ||
745 | /* | |
672bba3a CL |
746 | * Determine if a certain object on a page is on the freelist. Must hold the |
747 | * slab lock to guarantee that the chains are in a consistent state. | |
81819f0f CL |
748 | */ |
749 | static int on_freelist(struct kmem_cache *s, struct page *page, void *search) | |
750 | { | |
751 | int nr = 0; | |
752 | void *fp = page->freelist; | |
753 | void *object = NULL; | |
754 | ||
755 | while (fp && nr <= s->objects) { | |
756 | if (fp == search) | |
757 | return 1; | |
758 | if (!check_valid_pointer(s, page, fp)) { | |
759 | if (object) { | |
760 | object_err(s, page, object, | |
761 | "Freechain corrupt"); | |
762 | set_freepointer(s, object, NULL); | |
763 | break; | |
764 | } else { | |
24922684 | 765 | slab_err(s, page, "Freepointer corrupt"); |
81819f0f CL |
766 | page->freelist = NULL; |
767 | page->inuse = s->objects; | |
24922684 | 768 | slab_fix(s, "Freelist cleared"); |
81819f0f CL |
769 | return 0; |
770 | } | |
771 | break; | |
772 | } | |
773 | object = fp; | |
774 | fp = get_freepointer(s, object); | |
775 | nr++; | |
776 | } | |
777 | ||
778 | if (page->inuse != s->objects - nr) { | |
70d71228 | 779 | slab_err(s, page, "Wrong object count. Counter is %d but " |
24922684 | 780 | "counted were %d", page->inuse, s->objects - nr); |
81819f0f | 781 | page->inuse = s->objects - nr; |
24922684 | 782 | slab_fix(s, "Object count adjusted."); |
81819f0f CL |
783 | } |
784 | return search == NULL; | |
785 | } | |
786 | ||
3ec09742 CL |
787 | static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc) |
788 | { | |
789 | if (s->flags & SLAB_TRACE) { | |
790 | printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", | |
791 | s->name, | |
792 | alloc ? "alloc" : "free", | |
793 | object, page->inuse, | |
794 | page->freelist); | |
795 | ||
796 | if (!alloc) | |
797 | print_section("Object", (void *)object, s->objsize); | |
798 | ||
799 | dump_stack(); | |
800 | } | |
801 | } | |
802 | ||
643b1138 | 803 | /* |
672bba3a | 804 | * Tracking of fully allocated slabs for debugging purposes. |
643b1138 | 805 | */ |
e95eed57 | 806 | static void add_full(struct kmem_cache_node *n, struct page *page) |
643b1138 | 807 | { |
643b1138 CL |
808 | spin_lock(&n->list_lock); |
809 | list_add(&page->lru, &n->full); | |
810 | spin_unlock(&n->list_lock); | |
811 | } | |
812 | ||
813 | static void remove_full(struct kmem_cache *s, struct page *page) | |
814 | { | |
815 | struct kmem_cache_node *n; | |
816 | ||
817 | if (!(s->flags & SLAB_STORE_USER)) | |
818 | return; | |
819 | ||
820 | n = get_node(s, page_to_nid(page)); | |
821 | ||
822 | spin_lock(&n->list_lock); | |
823 | list_del(&page->lru); | |
824 | spin_unlock(&n->list_lock); | |
825 | } | |
826 | ||
3ec09742 CL |
827 | static void setup_object_debug(struct kmem_cache *s, struct page *page, |
828 | void *object) | |
829 | { | |
830 | if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) | |
831 | return; | |
832 | ||
833 | init_object(s, object, 0); | |
834 | init_tracking(s, object); | |
835 | } | |
836 | ||
837 | static int alloc_debug_processing(struct kmem_cache *s, struct page *page, | |
838 | void *object, void *addr) | |
81819f0f CL |
839 | { |
840 | if (!check_slab(s, page)) | |
841 | goto bad; | |
842 | ||
843 | if (object && !on_freelist(s, page, object)) { | |
24922684 | 844 | object_err(s, page, object, "Object already allocated"); |
70d71228 | 845 | goto bad; |
81819f0f CL |
846 | } |
847 | ||
848 | if (!check_valid_pointer(s, page, object)) { | |
849 | object_err(s, page, object, "Freelist Pointer check fails"); | |
70d71228 | 850 | goto bad; |
81819f0f CL |
851 | } |
852 | ||
3ec09742 | 853 | if (object && !check_object(s, page, object, 0)) |
81819f0f | 854 | goto bad; |
81819f0f | 855 | |
3ec09742 CL |
856 | /* Success perform special debug activities for allocs */ |
857 | if (s->flags & SLAB_STORE_USER) | |
858 | set_track(s, object, TRACK_ALLOC, addr); | |
859 | trace(s, page, object, 1); | |
860 | init_object(s, object, 1); | |
81819f0f | 861 | return 1; |
3ec09742 | 862 | |
81819f0f CL |
863 | bad: |
864 | if (PageSlab(page)) { | |
865 | /* | |
866 | * If this is a slab page then lets do the best we can | |
867 | * to avoid issues in the future. Marking all objects | |
672bba3a | 868 | * as used avoids touching the remaining objects. |
81819f0f | 869 | */ |
24922684 | 870 | slab_fix(s, "Marking all objects used"); |
81819f0f CL |
871 | page->inuse = s->objects; |
872 | page->freelist = NULL; | |
81819f0f CL |
873 | } |
874 | return 0; | |
875 | } | |
876 | ||
3ec09742 CL |
877 | static int free_debug_processing(struct kmem_cache *s, struct page *page, |
878 | void *object, void *addr) | |
81819f0f CL |
879 | { |
880 | if (!check_slab(s, page)) | |
881 | goto fail; | |
882 | ||
883 | if (!check_valid_pointer(s, page, object)) { | |
70d71228 | 884 | slab_err(s, page, "Invalid object pointer 0x%p", object); |
81819f0f CL |
885 | goto fail; |
886 | } | |
887 | ||
888 | if (on_freelist(s, page, object)) { | |
24922684 | 889 | object_err(s, page, object, "Object already free"); |
81819f0f CL |
890 | goto fail; |
891 | } | |
892 | ||
893 | if (!check_object(s, page, object, 1)) | |
894 | return 0; | |
895 | ||
896 | if (unlikely(s != page->slab)) { | |
897 | if (!PageSlab(page)) | |
70d71228 CL |
898 | slab_err(s, page, "Attempt to free object(0x%p) " |
899 | "outside of slab", object); | |
81819f0f | 900 | else |
70d71228 | 901 | if (!page->slab) { |
81819f0f | 902 | printk(KERN_ERR |
70d71228 | 903 | "SLUB <none>: no slab for object 0x%p.\n", |
81819f0f | 904 | object); |
70d71228 CL |
905 | dump_stack(); |
906 | } | |
81819f0f | 907 | else |
24922684 CL |
908 | object_err(s, page, object, |
909 | "page slab pointer corrupt."); | |
81819f0f CL |
910 | goto fail; |
911 | } | |
3ec09742 CL |
912 | |
913 | /* Special debug activities for freeing objects */ | |
914 | if (!SlabFrozen(page) && !page->freelist) | |
915 | remove_full(s, page); | |
916 | if (s->flags & SLAB_STORE_USER) | |
917 | set_track(s, object, TRACK_FREE, addr); | |
918 | trace(s, page, object, 0); | |
919 | init_object(s, object, 0); | |
81819f0f | 920 | return 1; |
3ec09742 | 921 | |
81819f0f | 922 | fail: |
24922684 | 923 | slab_fix(s, "Object at 0x%p not freed", object); |
81819f0f CL |
924 | return 0; |
925 | } | |
926 | ||
41ecc55b CL |
927 | static int __init setup_slub_debug(char *str) |
928 | { | |
f0630fff CL |
929 | slub_debug = DEBUG_DEFAULT_FLAGS; |
930 | if (*str++ != '=' || !*str) | |
931 | /* | |
932 | * No options specified. Switch on full debugging. | |
933 | */ | |
934 | goto out; | |
935 | ||
936 | if (*str == ',') | |
937 | /* | |
938 | * No options but restriction on slabs. This means full | |
939 | * debugging for slabs matching a pattern. | |
940 | */ | |
941 | goto check_slabs; | |
942 | ||
943 | slub_debug = 0; | |
944 | if (*str == '-') | |
945 | /* | |
946 | * Switch off all debugging measures. | |
947 | */ | |
948 | goto out; | |
949 | ||
950 | /* | |
951 | * Determine which debug features should be switched on | |
952 | */ | |
953 | for ( ;*str && *str != ','; str++) { | |
954 | switch (tolower(*str)) { | |
955 | case 'f': | |
956 | slub_debug |= SLAB_DEBUG_FREE; | |
957 | break; | |
958 | case 'z': | |
959 | slub_debug |= SLAB_RED_ZONE; | |
960 | break; | |
961 | case 'p': | |
962 | slub_debug |= SLAB_POISON; | |
963 | break; | |
964 | case 'u': | |
965 | slub_debug |= SLAB_STORE_USER; | |
966 | break; | |
967 | case 't': | |
968 | slub_debug |= SLAB_TRACE; | |
969 | break; | |
970 | default: | |
971 | printk(KERN_ERR "slub_debug option '%c' " | |
972 | "unknown. skipped\n",*str); | |
973 | } | |
41ecc55b CL |
974 | } |
975 | ||
f0630fff | 976 | check_slabs: |
41ecc55b CL |
977 | if (*str == ',') |
978 | slub_debug_slabs = str + 1; | |
f0630fff | 979 | out: |
41ecc55b CL |
980 | return 1; |
981 | } | |
982 | ||
983 | __setup("slub_debug", setup_slub_debug); | |
984 | ||
ba0268a8 CL |
985 | static unsigned long kmem_cache_flags(unsigned long objsize, |
986 | unsigned long flags, const char *name, | |
4ba9b9d0 | 987 | void (*ctor)(struct kmem_cache *, void *)) |
41ecc55b CL |
988 | { |
989 | /* | |
990 | * The page->offset field is only 16 bit wide. This is an offset | |
991 | * in units of words from the beginning of an object. If the slab | |
992 | * size is bigger then we cannot move the free pointer behind the | |
993 | * object anymore. | |
994 | * | |
995 | * On 32 bit platforms the limit is 256k. On 64bit platforms | |
996 | * the limit is 512k. | |
997 | * | |
c59def9f | 998 | * Debugging or ctor may create a need to move the free |
41ecc55b CL |
999 | * pointer. Fail if this happens. |
1000 | */ | |
ba0268a8 CL |
1001 | if (objsize >= 65535 * sizeof(void *)) { |
1002 | BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON | | |
41ecc55b | 1003 | SLAB_STORE_USER | SLAB_DESTROY_BY_RCU)); |
ba0268a8 CL |
1004 | BUG_ON(ctor); |
1005 | } else { | |
41ecc55b CL |
1006 | /* |
1007 | * Enable debugging if selected on the kernel commandline. | |
1008 | */ | |
1009 | if (slub_debug && (!slub_debug_slabs || | |
ba0268a8 | 1010 | strncmp(slub_debug_slabs, name, |
41ecc55b | 1011 | strlen(slub_debug_slabs)) == 0)) |
ba0268a8 CL |
1012 | flags |= slub_debug; |
1013 | } | |
1014 | ||
1015 | return flags; | |
41ecc55b CL |
1016 | } |
1017 | #else | |
3ec09742 CL |
1018 | static inline void setup_object_debug(struct kmem_cache *s, |
1019 | struct page *page, void *object) {} | |
41ecc55b | 1020 | |
3ec09742 CL |
1021 | static inline int alloc_debug_processing(struct kmem_cache *s, |
1022 | struct page *page, void *object, void *addr) { return 0; } | |
41ecc55b | 1023 | |
3ec09742 CL |
1024 | static inline int free_debug_processing(struct kmem_cache *s, |
1025 | struct page *page, void *object, void *addr) { return 0; } | |
41ecc55b | 1026 | |
41ecc55b CL |
1027 | static inline int slab_pad_check(struct kmem_cache *s, struct page *page) |
1028 | { return 1; } | |
1029 | static inline int check_object(struct kmem_cache *s, struct page *page, | |
1030 | void *object, int active) { return 1; } | |
3ec09742 | 1031 | static inline void add_full(struct kmem_cache_node *n, struct page *page) {} |
ba0268a8 CL |
1032 | static inline unsigned long kmem_cache_flags(unsigned long objsize, |
1033 | unsigned long flags, const char *name, | |
4ba9b9d0 | 1034 | void (*ctor)(struct kmem_cache *, void *)) |
ba0268a8 CL |
1035 | { |
1036 | return flags; | |
1037 | } | |
41ecc55b CL |
1038 | #define slub_debug 0 |
1039 | #endif | |
81819f0f CL |
1040 | /* |
1041 | * Slab allocation and freeing | |
1042 | */ | |
1043 | static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1044 | { | |
1045 | struct page * page; | |
1046 | int pages = 1 << s->order; | |
1047 | ||
1048 | if (s->order) | |
1049 | flags |= __GFP_COMP; | |
1050 | ||
1051 | if (s->flags & SLAB_CACHE_DMA) | |
1052 | flags |= SLUB_DMA; | |
1053 | ||
e12ba74d MG |
1054 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
1055 | flags |= __GFP_RECLAIMABLE; | |
1056 | ||
81819f0f CL |
1057 | if (node == -1) |
1058 | page = alloc_pages(flags, s->order); | |
1059 | else | |
1060 | page = alloc_pages_node(node, flags, s->order); | |
1061 | ||
1062 | if (!page) | |
1063 | return NULL; | |
1064 | ||
1065 | mod_zone_page_state(page_zone(page), | |
1066 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1067 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1068 | pages); | |
1069 | ||
1070 | return page; | |
1071 | } | |
1072 | ||
1073 | static void setup_object(struct kmem_cache *s, struct page *page, | |
1074 | void *object) | |
1075 | { | |
3ec09742 | 1076 | setup_object_debug(s, page, object); |
4f104934 | 1077 | if (unlikely(s->ctor)) |
4ba9b9d0 | 1078 | s->ctor(s, object); |
81819f0f CL |
1079 | } |
1080 | ||
1081 | static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) | |
1082 | { | |
1083 | struct page *page; | |
1084 | struct kmem_cache_node *n; | |
1085 | void *start; | |
81819f0f CL |
1086 | void *last; |
1087 | void *p; | |
1088 | ||
6cb06229 | 1089 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
81819f0f | 1090 | |
6cb06229 CL |
1091 | page = allocate_slab(s, |
1092 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); | |
81819f0f CL |
1093 | if (!page) |
1094 | goto out; | |
1095 | ||
1096 | n = get_node(s, page_to_nid(page)); | |
1097 | if (n) | |
1098 | atomic_long_inc(&n->nr_slabs); | |
81819f0f CL |
1099 | page->slab = s; |
1100 | page->flags |= 1 << PG_slab; | |
1101 | if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | | |
1102 | SLAB_STORE_USER | SLAB_TRACE)) | |
35e5d7ee | 1103 | SetSlabDebug(page); |
81819f0f CL |
1104 | |
1105 | start = page_address(page); | |
81819f0f CL |
1106 | |
1107 | if (unlikely(s->flags & SLAB_POISON)) | |
1108 | memset(start, POISON_INUSE, PAGE_SIZE << s->order); | |
1109 | ||
1110 | last = start; | |
7656c72b | 1111 | for_each_object(p, s, start) { |
81819f0f CL |
1112 | setup_object(s, page, last); |
1113 | set_freepointer(s, last, p); | |
1114 | last = p; | |
1115 | } | |
1116 | setup_object(s, page, last); | |
1117 | set_freepointer(s, last, NULL); | |
1118 | ||
1119 | page->freelist = start; | |
1120 | page->inuse = 0; | |
1121 | out: | |
81819f0f CL |
1122 | return page; |
1123 | } | |
1124 | ||
1125 | static void __free_slab(struct kmem_cache *s, struct page *page) | |
1126 | { | |
1127 | int pages = 1 << s->order; | |
1128 | ||
c59def9f | 1129 | if (unlikely(SlabDebug(page))) { |
81819f0f CL |
1130 | void *p; |
1131 | ||
1132 | slab_pad_check(s, page); | |
c59def9f | 1133 | for_each_object(p, s, page_address(page)) |
81819f0f | 1134 | check_object(s, page, p, 0); |
2208b764 | 1135 | ClearSlabDebug(page); |
81819f0f CL |
1136 | } |
1137 | ||
1138 | mod_zone_page_state(page_zone(page), | |
1139 | (s->flags & SLAB_RECLAIM_ACCOUNT) ? | |
1140 | NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, | |
1141 | - pages); | |
1142 | ||
81819f0f CL |
1143 | __free_pages(page, s->order); |
1144 | } | |
1145 | ||
1146 | static void rcu_free_slab(struct rcu_head *h) | |
1147 | { | |
1148 | struct page *page; | |
1149 | ||
1150 | page = container_of((struct list_head *)h, struct page, lru); | |
1151 | __free_slab(page->slab, page); | |
1152 | } | |
1153 | ||
1154 | static void free_slab(struct kmem_cache *s, struct page *page) | |
1155 | { | |
1156 | if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { | |
1157 | /* | |
1158 | * RCU free overloads the RCU head over the LRU | |
1159 | */ | |
1160 | struct rcu_head *head = (void *)&page->lru; | |
1161 | ||
1162 | call_rcu(head, rcu_free_slab); | |
1163 | } else | |
1164 | __free_slab(s, page); | |
1165 | } | |
1166 | ||
1167 | static void discard_slab(struct kmem_cache *s, struct page *page) | |
1168 | { | |
1169 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1170 | ||
1171 | atomic_long_dec(&n->nr_slabs); | |
1172 | reset_page_mapcount(page); | |
35e5d7ee | 1173 | __ClearPageSlab(page); |
81819f0f CL |
1174 | free_slab(s, page); |
1175 | } | |
1176 | ||
1177 | /* | |
1178 | * Per slab locking using the pagelock | |
1179 | */ | |
1180 | static __always_inline void slab_lock(struct page *page) | |
1181 | { | |
1182 | bit_spin_lock(PG_locked, &page->flags); | |
1183 | } | |
1184 | ||
1185 | static __always_inline void slab_unlock(struct page *page) | |
1186 | { | |
1187 | bit_spin_unlock(PG_locked, &page->flags); | |
1188 | } | |
1189 | ||
1190 | static __always_inline int slab_trylock(struct page *page) | |
1191 | { | |
1192 | int rc = 1; | |
1193 | ||
1194 | rc = bit_spin_trylock(PG_locked, &page->flags); | |
1195 | return rc; | |
1196 | } | |
1197 | ||
1198 | /* | |
1199 | * Management of partially allocated slabs | |
1200 | */ | |
7c2e132c CL |
1201 | static void add_partial(struct kmem_cache_node *n, |
1202 | struct page *page, int tail) | |
81819f0f | 1203 | { |
e95eed57 CL |
1204 | spin_lock(&n->list_lock); |
1205 | n->nr_partial++; | |
7c2e132c CL |
1206 | if (tail) |
1207 | list_add_tail(&page->lru, &n->partial); | |
1208 | else | |
1209 | list_add(&page->lru, &n->partial); | |
81819f0f CL |
1210 | spin_unlock(&n->list_lock); |
1211 | } | |
1212 | ||
1213 | static void remove_partial(struct kmem_cache *s, | |
1214 | struct page *page) | |
1215 | { | |
1216 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); | |
1217 | ||
1218 | spin_lock(&n->list_lock); | |
1219 | list_del(&page->lru); | |
1220 | n->nr_partial--; | |
1221 | spin_unlock(&n->list_lock); | |
1222 | } | |
1223 | ||
1224 | /* | |
672bba3a | 1225 | * Lock slab and remove from the partial list. |
81819f0f | 1226 | * |
672bba3a | 1227 | * Must hold list_lock. |
81819f0f | 1228 | */ |
4b6f0750 | 1229 | static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page) |
81819f0f CL |
1230 | { |
1231 | if (slab_trylock(page)) { | |
1232 | list_del(&page->lru); | |
1233 | n->nr_partial--; | |
4b6f0750 | 1234 | SetSlabFrozen(page); |
81819f0f CL |
1235 | return 1; |
1236 | } | |
1237 | return 0; | |
1238 | } | |
1239 | ||
1240 | /* | |
672bba3a | 1241 | * Try to allocate a partial slab from a specific node. |
81819f0f CL |
1242 | */ |
1243 | static struct page *get_partial_node(struct kmem_cache_node *n) | |
1244 | { | |
1245 | struct page *page; | |
1246 | ||
1247 | /* | |
1248 | * Racy check. If we mistakenly see no partial slabs then we | |
1249 | * just allocate an empty slab. If we mistakenly try to get a | |
672bba3a CL |
1250 | * partial slab and there is none available then get_partials() |
1251 | * will return NULL. | |
81819f0f CL |
1252 | */ |
1253 | if (!n || !n->nr_partial) | |
1254 | return NULL; | |
1255 | ||
1256 | spin_lock(&n->list_lock); | |
1257 | list_for_each_entry(page, &n->partial, lru) | |
4b6f0750 | 1258 | if (lock_and_freeze_slab(n, page)) |
81819f0f CL |
1259 | goto out; |
1260 | page = NULL; | |
1261 | out: | |
1262 | spin_unlock(&n->list_lock); | |
1263 | return page; | |
1264 | } | |
1265 | ||
1266 | /* | |
672bba3a | 1267 | * Get a page from somewhere. Search in increasing NUMA distances. |
81819f0f CL |
1268 | */ |
1269 | static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) | |
1270 | { | |
1271 | #ifdef CONFIG_NUMA | |
1272 | struct zonelist *zonelist; | |
1273 | struct zone **z; | |
1274 | struct page *page; | |
1275 | ||
1276 | /* | |
672bba3a CL |
1277 | * The defrag ratio allows a configuration of the tradeoffs between |
1278 | * inter node defragmentation and node local allocations. A lower | |
1279 | * defrag_ratio increases the tendency to do local allocations | |
1280 | * instead of attempting to obtain partial slabs from other nodes. | |
81819f0f | 1281 | * |
672bba3a CL |
1282 | * If the defrag_ratio is set to 0 then kmalloc() always |
1283 | * returns node local objects. If the ratio is higher then kmalloc() | |
1284 | * may return off node objects because partial slabs are obtained | |
1285 | * from other nodes and filled up. | |
81819f0f CL |
1286 | * |
1287 | * If /sys/slab/xx/defrag_ratio is set to 100 (which makes | |
672bba3a CL |
1288 | * defrag_ratio = 1000) then every (well almost) allocation will |
1289 | * first attempt to defrag slab caches on other nodes. This means | |
1290 | * scanning over all nodes to look for partial slabs which may be | |
1291 | * expensive if we do it every time we are trying to find a slab | |
1292 | * with available objects. | |
81819f0f | 1293 | */ |
9824601e CL |
1294 | if (!s->remote_node_defrag_ratio || |
1295 | get_cycles() % 1024 > s->remote_node_defrag_ratio) | |
81819f0f CL |
1296 | return NULL; |
1297 | ||
1298 | zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
1299 | ->node_zonelists[gfp_zone(flags)]; | |
1300 | for (z = zonelist->zones; *z; z++) { | |
1301 | struct kmem_cache_node *n; | |
1302 | ||
1303 | n = get_node(s, zone_to_nid(*z)); | |
1304 | ||
1305 | if (n && cpuset_zone_allowed_hardwall(*z, flags) && | |
e95eed57 | 1306 | n->nr_partial > MIN_PARTIAL) { |
81819f0f CL |
1307 | page = get_partial_node(n); |
1308 | if (page) | |
1309 | return page; | |
1310 | } | |
1311 | } | |
1312 | #endif | |
1313 | return NULL; | |
1314 | } | |
1315 | ||
1316 | /* | |
1317 | * Get a partial page, lock it and return it. | |
1318 | */ | |
1319 | static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) | |
1320 | { | |
1321 | struct page *page; | |
1322 | int searchnode = (node == -1) ? numa_node_id() : node; | |
1323 | ||
1324 | page = get_partial_node(get_node(s, searchnode)); | |
1325 | if (page || (flags & __GFP_THISNODE)) | |
1326 | return page; | |
1327 | ||
1328 | return get_any_partial(s, flags); | |
1329 | } | |
1330 | ||
1331 | /* | |
1332 | * Move a page back to the lists. | |
1333 | * | |
1334 | * Must be called with the slab lock held. | |
1335 | * | |
1336 | * On exit the slab lock will have been dropped. | |
1337 | */ | |
7c2e132c | 1338 | static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) |
81819f0f | 1339 | { |
e95eed57 CL |
1340 | struct kmem_cache_node *n = get_node(s, page_to_nid(page)); |
1341 | ||
4b6f0750 | 1342 | ClearSlabFrozen(page); |
81819f0f | 1343 | if (page->inuse) { |
e95eed57 | 1344 | |
81819f0f | 1345 | if (page->freelist) |
7c2e132c | 1346 | add_partial(n, page, tail); |
35e5d7ee | 1347 | else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER)) |
e95eed57 | 1348 | add_full(n, page); |
81819f0f | 1349 | slab_unlock(page); |
e95eed57 | 1350 | |
81819f0f | 1351 | } else { |
e95eed57 CL |
1352 | if (n->nr_partial < MIN_PARTIAL) { |
1353 | /* | |
672bba3a CL |
1354 | * Adding an empty slab to the partial slabs in order |
1355 | * to avoid page allocator overhead. This slab needs | |
1356 | * to come after the other slabs with objects in | |
1357 | * order to fill them up. That way the size of the | |
1358 | * partial list stays small. kmem_cache_shrink can | |
1359 | * reclaim empty slabs from the partial list. | |
e95eed57 | 1360 | */ |
7c2e132c | 1361 | add_partial(n, page, 1); |
e95eed57 CL |
1362 | slab_unlock(page); |
1363 | } else { | |
1364 | slab_unlock(page); | |
1365 | discard_slab(s, page); | |
1366 | } | |
81819f0f CL |
1367 | } |
1368 | } | |
1369 | ||
1370 | /* | |
1371 | * Remove the cpu slab | |
1372 | */ | |
dfb4f096 | 1373 | static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
81819f0f | 1374 | { |
dfb4f096 | 1375 | struct page *page = c->page; |
7c2e132c | 1376 | int tail = 1; |
894b8788 CL |
1377 | /* |
1378 | * Merge cpu freelist into freelist. Typically we get here | |
1379 | * because both freelists are empty. So this is unlikely | |
1380 | * to occur. | |
1381 | */ | |
dfb4f096 | 1382 | while (unlikely(c->freelist)) { |
894b8788 CL |
1383 | void **object; |
1384 | ||
7c2e132c CL |
1385 | tail = 0; /* Hot objects. Put the slab first */ |
1386 | ||
894b8788 | 1387 | /* Retrieve object from cpu_freelist */ |
dfb4f096 | 1388 | object = c->freelist; |
b3fba8da | 1389 | c->freelist = c->freelist[c->offset]; |
894b8788 CL |
1390 | |
1391 | /* And put onto the regular freelist */ | |
b3fba8da | 1392 | object[c->offset] = page->freelist; |
894b8788 CL |
1393 | page->freelist = object; |
1394 | page->inuse--; | |
1395 | } | |
dfb4f096 | 1396 | c->page = NULL; |
7c2e132c | 1397 | unfreeze_slab(s, page, tail); |
81819f0f CL |
1398 | } |
1399 | ||
dfb4f096 | 1400 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
81819f0f | 1401 | { |
dfb4f096 CL |
1402 | slab_lock(c->page); |
1403 | deactivate_slab(s, c); | |
81819f0f CL |
1404 | } |
1405 | ||
1406 | /* | |
1407 | * Flush cpu slab. | |
1408 | * Called from IPI handler with interrupts disabled. | |
1409 | */ | |
0c710013 | 1410 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) |
81819f0f | 1411 | { |
dfb4f096 | 1412 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); |
81819f0f | 1413 | |
dfb4f096 CL |
1414 | if (likely(c && c->page)) |
1415 | flush_slab(s, c); | |
81819f0f CL |
1416 | } |
1417 | ||
1418 | static void flush_cpu_slab(void *d) | |
1419 | { | |
1420 | struct kmem_cache *s = d; | |
81819f0f | 1421 | |
dfb4f096 | 1422 | __flush_cpu_slab(s, smp_processor_id()); |
81819f0f CL |
1423 | } |
1424 | ||
1425 | static void flush_all(struct kmem_cache *s) | |
1426 | { | |
1427 | #ifdef CONFIG_SMP | |
1428 | on_each_cpu(flush_cpu_slab, s, 1, 1); | |
1429 | #else | |
1430 | unsigned long flags; | |
1431 | ||
1432 | local_irq_save(flags); | |
1433 | flush_cpu_slab(s); | |
1434 | local_irq_restore(flags); | |
1435 | #endif | |
1436 | } | |
1437 | ||
dfb4f096 CL |
1438 | /* |
1439 | * Check if the objects in a per cpu structure fit numa | |
1440 | * locality expectations. | |
1441 | */ | |
1442 | static inline int node_match(struct kmem_cache_cpu *c, int node) | |
1443 | { | |
1444 | #ifdef CONFIG_NUMA | |
1445 | if (node != -1 && c->node != node) | |
1446 | return 0; | |
1447 | #endif | |
1448 | return 1; | |
1449 | } | |
1450 | ||
81819f0f | 1451 | /* |
894b8788 CL |
1452 | * Slow path. The lockless freelist is empty or we need to perform |
1453 | * debugging duties. | |
1454 | * | |
1455 | * Interrupts are disabled. | |
81819f0f | 1456 | * |
894b8788 CL |
1457 | * Processing is still very fast if new objects have been freed to the |
1458 | * regular freelist. In that case we simply take over the regular freelist | |
1459 | * as the lockless freelist and zap the regular freelist. | |
81819f0f | 1460 | * |
894b8788 CL |
1461 | * If that is not working then we fall back to the partial lists. We take the |
1462 | * first element of the freelist as the object to allocate now and move the | |
1463 | * rest of the freelist to the lockless freelist. | |
81819f0f | 1464 | * |
894b8788 CL |
1465 | * And if we were unable to get a new slab from the partial slab lists then |
1466 | * we need to allocate a new slab. This is slowest path since we may sleep. | |
81819f0f | 1467 | */ |
894b8788 | 1468 | static void *__slab_alloc(struct kmem_cache *s, |
dfb4f096 | 1469 | gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c) |
81819f0f | 1470 | { |
81819f0f | 1471 | void **object; |
dfb4f096 | 1472 | struct page *new; |
81819f0f | 1473 | |
dfb4f096 | 1474 | if (!c->page) |
81819f0f CL |
1475 | goto new_slab; |
1476 | ||
dfb4f096 CL |
1477 | slab_lock(c->page); |
1478 | if (unlikely(!node_match(c, node))) | |
81819f0f | 1479 | goto another_slab; |
894b8788 | 1480 | load_freelist: |
dfb4f096 | 1481 | object = c->page->freelist; |
81819f0f CL |
1482 | if (unlikely(!object)) |
1483 | goto another_slab; | |
dfb4f096 | 1484 | if (unlikely(SlabDebug(c->page))) |
81819f0f CL |
1485 | goto debug; |
1486 | ||
dfb4f096 | 1487 | object = c->page->freelist; |
b3fba8da | 1488 | c->freelist = object[c->offset]; |
dfb4f096 CL |
1489 | c->page->inuse = s->objects; |
1490 | c->page->freelist = NULL; | |
1491 | c->node = page_to_nid(c->page); | |
1492 | slab_unlock(c->page); | |
81819f0f CL |
1493 | return object; |
1494 | ||
1495 | another_slab: | |
dfb4f096 | 1496 | deactivate_slab(s, c); |
81819f0f CL |
1497 | |
1498 | new_slab: | |
dfb4f096 CL |
1499 | new = get_partial(s, gfpflags, node); |
1500 | if (new) { | |
1501 | c->page = new; | |
894b8788 | 1502 | goto load_freelist; |
81819f0f CL |
1503 | } |
1504 | ||
b811c202 CL |
1505 | if (gfpflags & __GFP_WAIT) |
1506 | local_irq_enable(); | |
1507 | ||
dfb4f096 | 1508 | new = new_slab(s, gfpflags, node); |
b811c202 CL |
1509 | |
1510 | if (gfpflags & __GFP_WAIT) | |
1511 | local_irq_disable(); | |
1512 | ||
dfb4f096 CL |
1513 | if (new) { |
1514 | c = get_cpu_slab(s, smp_processor_id()); | |
05aa3450 | 1515 | if (c->page) |
dfb4f096 | 1516 | flush_slab(s, c); |
dfb4f096 CL |
1517 | slab_lock(new); |
1518 | SetSlabFrozen(new); | |
1519 | c->page = new; | |
4b6f0750 | 1520 | goto load_freelist; |
81819f0f | 1521 | } |
81819f0f CL |
1522 | return NULL; |
1523 | debug: | |
dfb4f096 CL |
1524 | object = c->page->freelist; |
1525 | if (!alloc_debug_processing(s, c->page, object, addr)) | |
81819f0f | 1526 | goto another_slab; |
894b8788 | 1527 | |
dfb4f096 | 1528 | c->page->inuse++; |
b3fba8da | 1529 | c->page->freelist = object[c->offset]; |
ee3c72a1 | 1530 | c->node = -1; |
dfb4f096 | 1531 | slab_unlock(c->page); |
894b8788 CL |
1532 | return object; |
1533 | } | |
1534 | ||
1535 | /* | |
1536 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) | |
1537 | * have the fastpath folded into their functions. So no function call | |
1538 | * overhead for requests that can be satisfied on the fastpath. | |
1539 | * | |
1540 | * The fastpath works by first checking if the lockless freelist can be used. | |
1541 | * If not then __slab_alloc is called for slow processing. | |
1542 | * | |
1543 | * Otherwise we can simply pick the next object from the lockless free list. | |
1544 | */ | |
1545 | static void __always_inline *slab_alloc(struct kmem_cache *s, | |
ce15fea8 | 1546 | gfp_t gfpflags, int node, void *addr) |
894b8788 | 1547 | { |
894b8788 CL |
1548 | void **object; |
1549 | unsigned long flags; | |
dfb4f096 | 1550 | struct kmem_cache_cpu *c; |
894b8788 CL |
1551 | |
1552 | local_irq_save(flags); | |
dfb4f096 | 1553 | c = get_cpu_slab(s, smp_processor_id()); |
ee3c72a1 | 1554 | if (unlikely(!c->freelist || !node_match(c, node))) |
894b8788 | 1555 | |
dfb4f096 | 1556 | object = __slab_alloc(s, gfpflags, node, addr, c); |
894b8788 CL |
1557 | |
1558 | else { | |
dfb4f096 | 1559 | object = c->freelist; |
b3fba8da | 1560 | c->freelist = object[c->offset]; |
894b8788 CL |
1561 | } |
1562 | local_irq_restore(flags); | |
d07dbea4 CL |
1563 | |
1564 | if (unlikely((gfpflags & __GFP_ZERO) && object)) | |
42a9fdbb | 1565 | memset(object, 0, c->objsize); |
d07dbea4 | 1566 | |
894b8788 | 1567 | return object; |
81819f0f CL |
1568 | } |
1569 | ||
1570 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) | |
1571 | { | |
ce15fea8 | 1572 | return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); |
81819f0f CL |
1573 | } |
1574 | EXPORT_SYMBOL(kmem_cache_alloc); | |
1575 | ||
1576 | #ifdef CONFIG_NUMA | |
1577 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) | |
1578 | { | |
ce15fea8 | 1579 | return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); |
81819f0f CL |
1580 | } |
1581 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
1582 | #endif | |
1583 | ||
1584 | /* | |
894b8788 CL |
1585 | * Slow patch handling. This may still be called frequently since objects |
1586 | * have a longer lifetime than the cpu slabs in most processing loads. | |
81819f0f | 1587 | * |
894b8788 CL |
1588 | * So we still attempt to reduce cache line usage. Just take the slab |
1589 | * lock and free the item. If there is no additional partial page | |
1590 | * handling required then we can return immediately. | |
81819f0f | 1591 | */ |
894b8788 | 1592 | static void __slab_free(struct kmem_cache *s, struct page *page, |
b3fba8da | 1593 | void *x, void *addr, unsigned int offset) |
81819f0f CL |
1594 | { |
1595 | void *prior; | |
1596 | void **object = (void *)x; | |
81819f0f | 1597 | |
81819f0f CL |
1598 | slab_lock(page); |
1599 | ||
35e5d7ee | 1600 | if (unlikely(SlabDebug(page))) |
81819f0f CL |
1601 | goto debug; |
1602 | checks_ok: | |
b3fba8da | 1603 | prior = object[offset] = page->freelist; |
81819f0f CL |
1604 | page->freelist = object; |
1605 | page->inuse--; | |
1606 | ||
4b6f0750 | 1607 | if (unlikely(SlabFrozen(page))) |
81819f0f CL |
1608 | goto out_unlock; |
1609 | ||
1610 | if (unlikely(!page->inuse)) | |
1611 | goto slab_empty; | |
1612 | ||
1613 | /* | |
1614 | * Objects left in the slab. If it | |
1615 | * was not on the partial list before | |
1616 | * then add it. | |
1617 | */ | |
1618 | if (unlikely(!prior)) | |
7c2e132c | 1619 | add_partial(get_node(s, page_to_nid(page)), page, 1); |
81819f0f CL |
1620 | |
1621 | out_unlock: | |
1622 | slab_unlock(page); | |
81819f0f CL |
1623 | return; |
1624 | ||
1625 | slab_empty: | |
1626 | if (prior) | |
1627 | /* | |
672bba3a | 1628 | * Slab still on the partial list. |
81819f0f CL |
1629 | */ |
1630 | remove_partial(s, page); | |
1631 | ||
1632 | slab_unlock(page); | |
1633 | discard_slab(s, page); | |
81819f0f CL |
1634 | return; |
1635 | ||
1636 | debug: | |
3ec09742 | 1637 | if (!free_debug_processing(s, page, x, addr)) |
77c5e2d0 | 1638 | goto out_unlock; |
77c5e2d0 | 1639 | goto checks_ok; |
81819f0f CL |
1640 | } |
1641 | ||
894b8788 CL |
1642 | /* |
1643 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that | |
1644 | * can perform fastpath freeing without additional function calls. | |
1645 | * | |
1646 | * The fastpath is only possible if we are freeing to the current cpu slab | |
1647 | * of this processor. This typically the case if we have just allocated | |
1648 | * the item before. | |
1649 | * | |
1650 | * If fastpath is not possible then fall back to __slab_free where we deal | |
1651 | * with all sorts of special processing. | |
1652 | */ | |
1653 | static void __always_inline slab_free(struct kmem_cache *s, | |
1654 | struct page *page, void *x, void *addr) | |
1655 | { | |
1656 | void **object = (void *)x; | |
1657 | unsigned long flags; | |
dfb4f096 | 1658 | struct kmem_cache_cpu *c; |
894b8788 CL |
1659 | |
1660 | local_irq_save(flags); | |
02febdf7 | 1661 | debug_check_no_locks_freed(object, s->objsize); |
dfb4f096 | 1662 | c = get_cpu_slab(s, smp_processor_id()); |
ee3c72a1 | 1663 | if (likely(page == c->page && c->node >= 0)) { |
b3fba8da | 1664 | object[c->offset] = c->freelist; |
dfb4f096 | 1665 | c->freelist = object; |
894b8788 | 1666 | } else |
b3fba8da | 1667 | __slab_free(s, page, x, addr, c->offset); |
894b8788 CL |
1668 | |
1669 | local_irq_restore(flags); | |
1670 | } | |
1671 | ||
81819f0f CL |
1672 | void kmem_cache_free(struct kmem_cache *s, void *x) |
1673 | { | |
77c5e2d0 | 1674 | struct page *page; |
81819f0f | 1675 | |
b49af68f | 1676 | page = virt_to_head_page(x); |
81819f0f | 1677 | |
77c5e2d0 | 1678 | slab_free(s, page, x, __builtin_return_address(0)); |
81819f0f CL |
1679 | } |
1680 | EXPORT_SYMBOL(kmem_cache_free); | |
1681 | ||
1682 | /* Figure out on which slab object the object resides */ | |
1683 | static struct page *get_object_page(const void *x) | |
1684 | { | |
b49af68f | 1685 | struct page *page = virt_to_head_page(x); |
81819f0f CL |
1686 | |
1687 | if (!PageSlab(page)) | |
1688 | return NULL; | |
1689 | ||
1690 | return page; | |
1691 | } | |
1692 | ||
1693 | /* | |
672bba3a CL |
1694 | * Object placement in a slab is made very easy because we always start at |
1695 | * offset 0. If we tune the size of the object to the alignment then we can | |
1696 | * get the required alignment by putting one properly sized object after | |
1697 | * another. | |
81819f0f CL |
1698 | * |
1699 | * Notice that the allocation order determines the sizes of the per cpu | |
1700 | * caches. Each processor has always one slab available for allocations. | |
1701 | * Increasing the allocation order reduces the number of times that slabs | |
672bba3a | 1702 | * must be moved on and off the partial lists and is therefore a factor in |
81819f0f | 1703 | * locking overhead. |
81819f0f CL |
1704 | */ |
1705 | ||
1706 | /* | |
1707 | * Mininum / Maximum order of slab pages. This influences locking overhead | |
1708 | * and slab fragmentation. A higher order reduces the number of partial slabs | |
1709 | * and increases the number of allocations possible without having to | |
1710 | * take the list_lock. | |
1711 | */ | |
1712 | static int slub_min_order; | |
1713 | static int slub_max_order = DEFAULT_MAX_ORDER; | |
81819f0f CL |
1714 | static int slub_min_objects = DEFAULT_MIN_OBJECTS; |
1715 | ||
1716 | /* | |
1717 | * Merge control. If this is set then no merging of slab caches will occur. | |
672bba3a | 1718 | * (Could be removed. This was introduced to pacify the merge skeptics.) |
81819f0f CL |
1719 | */ |
1720 | static int slub_nomerge; | |
1721 | ||
81819f0f CL |
1722 | /* |
1723 | * Calculate the order of allocation given an slab object size. | |
1724 | * | |
672bba3a CL |
1725 | * The order of allocation has significant impact on performance and other |
1726 | * system components. Generally order 0 allocations should be preferred since | |
1727 | * order 0 does not cause fragmentation in the page allocator. Larger objects | |
1728 | * be problematic to put into order 0 slabs because there may be too much | |
1729 | * unused space left. We go to a higher order if more than 1/8th of the slab | |
1730 | * would be wasted. | |
1731 | * | |
1732 | * In order to reach satisfactory performance we must ensure that a minimum | |
1733 | * number of objects is in one slab. Otherwise we may generate too much | |
1734 | * activity on the partial lists which requires taking the list_lock. This is | |
1735 | * less a concern for large slabs though which are rarely used. | |
81819f0f | 1736 | * |
672bba3a CL |
1737 | * slub_max_order specifies the order where we begin to stop considering the |
1738 | * number of objects in a slab as critical. If we reach slub_max_order then | |
1739 | * we try to keep the page order as low as possible. So we accept more waste | |
1740 | * of space in favor of a small page order. | |
81819f0f | 1741 | * |
672bba3a CL |
1742 | * Higher order allocations also allow the placement of more objects in a |
1743 | * slab and thereby reduce object handling overhead. If the user has | |
1744 | * requested a higher mininum order then we start with that one instead of | |
1745 | * the smallest order which will fit the object. | |
81819f0f | 1746 | */ |
5e6d444e CL |
1747 | static inline int slab_order(int size, int min_objects, |
1748 | int max_order, int fract_leftover) | |
81819f0f CL |
1749 | { |
1750 | int order; | |
1751 | int rem; | |
6300ea75 | 1752 | int min_order = slub_min_order; |
81819f0f | 1753 | |
6300ea75 | 1754 | for (order = max(min_order, |
5e6d444e CL |
1755 | fls(min_objects * size - 1) - PAGE_SHIFT); |
1756 | order <= max_order; order++) { | |
81819f0f | 1757 | |
5e6d444e | 1758 | unsigned long slab_size = PAGE_SIZE << order; |
81819f0f | 1759 | |
5e6d444e | 1760 | if (slab_size < min_objects * size) |
81819f0f CL |
1761 | continue; |
1762 | ||
1763 | rem = slab_size % size; | |
1764 | ||
5e6d444e | 1765 | if (rem <= slab_size / fract_leftover) |
81819f0f CL |
1766 | break; |
1767 | ||
1768 | } | |
672bba3a | 1769 | |
81819f0f CL |
1770 | return order; |
1771 | } | |
1772 | ||
5e6d444e CL |
1773 | static inline int calculate_order(int size) |
1774 | { | |
1775 | int order; | |
1776 | int min_objects; | |
1777 | int fraction; | |
1778 | ||
1779 | /* | |
1780 | * Attempt to find best configuration for a slab. This | |
1781 | * works by first attempting to generate a layout with | |
1782 | * the best configuration and backing off gradually. | |
1783 | * | |
1784 | * First we reduce the acceptable waste in a slab. Then | |
1785 | * we reduce the minimum objects required in a slab. | |
1786 | */ | |
1787 | min_objects = slub_min_objects; | |
1788 | while (min_objects > 1) { | |
1789 | fraction = 8; | |
1790 | while (fraction >= 4) { | |
1791 | order = slab_order(size, min_objects, | |
1792 | slub_max_order, fraction); | |
1793 | if (order <= slub_max_order) | |
1794 | return order; | |
1795 | fraction /= 2; | |
1796 | } | |
1797 | min_objects /= 2; | |
1798 | } | |
1799 | ||
1800 | /* | |
1801 | * We were unable to place multiple objects in a slab. Now | |
1802 | * lets see if we can place a single object there. | |
1803 | */ | |
1804 | order = slab_order(size, 1, slub_max_order, 1); | |
1805 | if (order <= slub_max_order) | |
1806 | return order; | |
1807 | ||
1808 | /* | |
1809 | * Doh this slab cannot be placed using slub_max_order. | |
1810 | */ | |
1811 | order = slab_order(size, 1, MAX_ORDER, 1); | |
1812 | if (order <= MAX_ORDER) | |
1813 | return order; | |
1814 | return -ENOSYS; | |
1815 | } | |
1816 | ||
81819f0f | 1817 | /* |
672bba3a | 1818 | * Figure out what the alignment of the objects will be. |
81819f0f CL |
1819 | */ |
1820 | static unsigned long calculate_alignment(unsigned long flags, | |
1821 | unsigned long align, unsigned long size) | |
1822 | { | |
1823 | /* | |
1824 | * If the user wants hardware cache aligned objects then | |
1825 | * follow that suggestion if the object is sufficiently | |
1826 | * large. | |
1827 | * | |
1828 | * The hardware cache alignment cannot override the | |
1829 | * specified alignment though. If that is greater | |
1830 | * then use it. | |
1831 | */ | |
5af60839 | 1832 | if ((flags & SLAB_HWCACHE_ALIGN) && |
65c02d4c CL |
1833 | size > cache_line_size() / 2) |
1834 | return max_t(unsigned long, align, cache_line_size()); | |
81819f0f CL |
1835 | |
1836 | if (align < ARCH_SLAB_MINALIGN) | |
1837 | return ARCH_SLAB_MINALIGN; | |
1838 | ||
1839 | return ALIGN(align, sizeof(void *)); | |
1840 | } | |
1841 | ||
dfb4f096 CL |
1842 | static void init_kmem_cache_cpu(struct kmem_cache *s, |
1843 | struct kmem_cache_cpu *c) | |
1844 | { | |
1845 | c->page = NULL; | |
1846 | c->freelist = NULL; | |
1847 | c->node = 0; | |
42a9fdbb CL |
1848 | c->offset = s->offset / sizeof(void *); |
1849 | c->objsize = s->objsize; | |
dfb4f096 CL |
1850 | } |
1851 | ||
81819f0f CL |
1852 | static void init_kmem_cache_node(struct kmem_cache_node *n) |
1853 | { | |
1854 | n->nr_partial = 0; | |
1855 | atomic_long_set(&n->nr_slabs, 0); | |
1856 | spin_lock_init(&n->list_lock); | |
1857 | INIT_LIST_HEAD(&n->partial); | |
8ab1372f | 1858 | #ifdef CONFIG_SLUB_DEBUG |
643b1138 | 1859 | INIT_LIST_HEAD(&n->full); |
8ab1372f | 1860 | #endif |
81819f0f CL |
1861 | } |
1862 | ||
4c93c355 CL |
1863 | #ifdef CONFIG_SMP |
1864 | /* | |
1865 | * Per cpu array for per cpu structures. | |
1866 | * | |
1867 | * The per cpu array places all kmem_cache_cpu structures from one processor | |
1868 | * close together meaning that it becomes possible that multiple per cpu | |
1869 | * structures are contained in one cacheline. This may be particularly | |
1870 | * beneficial for the kmalloc caches. | |
1871 | * | |
1872 | * A desktop system typically has around 60-80 slabs. With 100 here we are | |
1873 | * likely able to get per cpu structures for all caches from the array defined | |
1874 | * here. We must be able to cover all kmalloc caches during bootstrap. | |
1875 | * | |
1876 | * If the per cpu array is exhausted then fall back to kmalloc | |
1877 | * of individual cachelines. No sharing is possible then. | |
1878 | */ | |
1879 | #define NR_KMEM_CACHE_CPU 100 | |
1880 | ||
1881 | static DEFINE_PER_CPU(struct kmem_cache_cpu, | |
1882 | kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; | |
1883 | ||
1884 | static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); | |
1885 | static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; | |
1886 | ||
1887 | static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, | |
1888 | int cpu, gfp_t flags) | |
1889 | { | |
1890 | struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); | |
1891 | ||
1892 | if (c) | |
1893 | per_cpu(kmem_cache_cpu_free, cpu) = | |
1894 | (void *)c->freelist; | |
1895 | else { | |
1896 | /* Table overflow: So allocate ourselves */ | |
1897 | c = kmalloc_node( | |
1898 | ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), | |
1899 | flags, cpu_to_node(cpu)); | |
1900 | if (!c) | |
1901 | return NULL; | |
1902 | } | |
1903 | ||
1904 | init_kmem_cache_cpu(s, c); | |
1905 | return c; | |
1906 | } | |
1907 | ||
1908 | static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) | |
1909 | { | |
1910 | if (c < per_cpu(kmem_cache_cpu, cpu) || | |
1911 | c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { | |
1912 | kfree(c); | |
1913 | return; | |
1914 | } | |
1915 | c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); | |
1916 | per_cpu(kmem_cache_cpu_free, cpu) = c; | |
1917 | } | |
1918 | ||
1919 | static void free_kmem_cache_cpus(struct kmem_cache *s) | |
1920 | { | |
1921 | int cpu; | |
1922 | ||
1923 | for_each_online_cpu(cpu) { | |
1924 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
1925 | ||
1926 | if (c) { | |
1927 | s->cpu_slab[cpu] = NULL; | |
1928 | free_kmem_cache_cpu(c, cpu); | |
1929 | } | |
1930 | } | |
1931 | } | |
1932 | ||
1933 | static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) | |
1934 | { | |
1935 | int cpu; | |
1936 | ||
1937 | for_each_online_cpu(cpu) { | |
1938 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
1939 | ||
1940 | if (c) | |
1941 | continue; | |
1942 | ||
1943 | c = alloc_kmem_cache_cpu(s, cpu, flags); | |
1944 | if (!c) { | |
1945 | free_kmem_cache_cpus(s); | |
1946 | return 0; | |
1947 | } | |
1948 | s->cpu_slab[cpu] = c; | |
1949 | } | |
1950 | return 1; | |
1951 | } | |
1952 | ||
1953 | /* | |
1954 | * Initialize the per cpu array. | |
1955 | */ | |
1956 | static void init_alloc_cpu_cpu(int cpu) | |
1957 | { | |
1958 | int i; | |
1959 | ||
1960 | if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) | |
1961 | return; | |
1962 | ||
1963 | for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) | |
1964 | free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); | |
1965 | ||
1966 | cpu_set(cpu, kmem_cach_cpu_free_init_once); | |
1967 | } | |
1968 | ||
1969 | static void __init init_alloc_cpu(void) | |
1970 | { | |
1971 | int cpu; | |
1972 | ||
1973 | for_each_online_cpu(cpu) | |
1974 | init_alloc_cpu_cpu(cpu); | |
1975 | } | |
1976 | ||
1977 | #else | |
1978 | static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} | |
1979 | static inline void init_alloc_cpu(void) {} | |
1980 | ||
1981 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) | |
1982 | { | |
1983 | init_kmem_cache_cpu(s, &s->cpu_slab); | |
1984 | return 1; | |
1985 | } | |
1986 | #endif | |
1987 | ||
81819f0f CL |
1988 | #ifdef CONFIG_NUMA |
1989 | /* | |
1990 | * No kmalloc_node yet so do it by hand. We know that this is the first | |
1991 | * slab on the node for this slabcache. There are no concurrent accesses | |
1992 | * possible. | |
1993 | * | |
1994 | * Note that this function only works on the kmalloc_node_cache | |
4c93c355 CL |
1995 | * when allocating for the kmalloc_node_cache. This is used for bootstrapping |
1996 | * memory on a fresh node that has no slab structures yet. | |
81819f0f | 1997 | */ |
1cd7daa5 AB |
1998 | static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags, |
1999 | int node) | |
81819f0f CL |
2000 | { |
2001 | struct page *page; | |
2002 | struct kmem_cache_node *n; | |
2003 | ||
2004 | BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); | |
2005 | ||
a2f92ee7 | 2006 | page = new_slab(kmalloc_caches, gfpflags, node); |
81819f0f CL |
2007 | |
2008 | BUG_ON(!page); | |
a2f92ee7 CL |
2009 | if (page_to_nid(page) != node) { |
2010 | printk(KERN_ERR "SLUB: Unable to allocate memory from " | |
2011 | "node %d\n", node); | |
2012 | printk(KERN_ERR "SLUB: Allocating a useless per node structure " | |
2013 | "in order to be able to continue\n"); | |
2014 | } | |
2015 | ||
81819f0f CL |
2016 | n = page->freelist; |
2017 | BUG_ON(!n); | |
2018 | page->freelist = get_freepointer(kmalloc_caches, n); | |
2019 | page->inuse++; | |
2020 | kmalloc_caches->node[node] = n; | |
8ab1372f | 2021 | #ifdef CONFIG_SLUB_DEBUG |
d45f39cb CL |
2022 | init_object(kmalloc_caches, n, 1); |
2023 | init_tracking(kmalloc_caches, n); | |
8ab1372f | 2024 | #endif |
81819f0f CL |
2025 | init_kmem_cache_node(n); |
2026 | atomic_long_inc(&n->nr_slabs); | |
7c2e132c | 2027 | add_partial(n, page, 0); |
81819f0f CL |
2028 | return n; |
2029 | } | |
2030 | ||
2031 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2032 | { | |
2033 | int node; | |
2034 | ||
f64dc58c | 2035 | for_each_node_state(node, N_NORMAL_MEMORY) { |
81819f0f CL |
2036 | struct kmem_cache_node *n = s->node[node]; |
2037 | if (n && n != &s->local_node) | |
2038 | kmem_cache_free(kmalloc_caches, n); | |
2039 | s->node[node] = NULL; | |
2040 | } | |
2041 | } | |
2042 | ||
2043 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
2044 | { | |
2045 | int node; | |
2046 | int local_node; | |
2047 | ||
2048 | if (slab_state >= UP) | |
2049 | local_node = page_to_nid(virt_to_page(s)); | |
2050 | else | |
2051 | local_node = 0; | |
2052 | ||
f64dc58c | 2053 | for_each_node_state(node, N_NORMAL_MEMORY) { |
81819f0f CL |
2054 | struct kmem_cache_node *n; |
2055 | ||
2056 | if (local_node == node) | |
2057 | n = &s->local_node; | |
2058 | else { | |
2059 | if (slab_state == DOWN) { | |
2060 | n = early_kmem_cache_node_alloc(gfpflags, | |
2061 | node); | |
2062 | continue; | |
2063 | } | |
2064 | n = kmem_cache_alloc_node(kmalloc_caches, | |
2065 | gfpflags, node); | |
2066 | ||
2067 | if (!n) { | |
2068 | free_kmem_cache_nodes(s); | |
2069 | return 0; | |
2070 | } | |
2071 | ||
2072 | } | |
2073 | s->node[node] = n; | |
2074 | init_kmem_cache_node(n); | |
2075 | } | |
2076 | return 1; | |
2077 | } | |
2078 | #else | |
2079 | static void free_kmem_cache_nodes(struct kmem_cache *s) | |
2080 | { | |
2081 | } | |
2082 | ||
2083 | static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) | |
2084 | { | |
2085 | init_kmem_cache_node(&s->local_node); | |
2086 | return 1; | |
2087 | } | |
2088 | #endif | |
2089 | ||
2090 | /* | |
2091 | * calculate_sizes() determines the order and the distribution of data within | |
2092 | * a slab object. | |
2093 | */ | |
2094 | static int calculate_sizes(struct kmem_cache *s) | |
2095 | { | |
2096 | unsigned long flags = s->flags; | |
2097 | unsigned long size = s->objsize; | |
2098 | unsigned long align = s->align; | |
2099 | ||
2100 | /* | |
2101 | * Determine if we can poison the object itself. If the user of | |
2102 | * the slab may touch the object after free or before allocation | |
2103 | * then we should never poison the object itself. | |
2104 | */ | |
2105 | if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && | |
c59def9f | 2106 | !s->ctor) |
81819f0f CL |
2107 | s->flags |= __OBJECT_POISON; |
2108 | else | |
2109 | s->flags &= ~__OBJECT_POISON; | |
2110 | ||
2111 | /* | |
2112 | * Round up object size to the next word boundary. We can only | |
2113 | * place the free pointer at word boundaries and this determines | |
2114 | * the possible location of the free pointer. | |
2115 | */ | |
2116 | size = ALIGN(size, sizeof(void *)); | |
2117 | ||
41ecc55b | 2118 | #ifdef CONFIG_SLUB_DEBUG |
81819f0f | 2119 | /* |
672bba3a | 2120 | * If we are Redzoning then check if there is some space between the |
81819f0f | 2121 | * end of the object and the free pointer. If not then add an |
672bba3a | 2122 | * additional word to have some bytes to store Redzone information. |
81819f0f CL |
2123 | */ |
2124 | if ((flags & SLAB_RED_ZONE) && size == s->objsize) | |
2125 | size += sizeof(void *); | |
41ecc55b | 2126 | #endif |
81819f0f CL |
2127 | |
2128 | /* | |
672bba3a CL |
2129 | * With that we have determined the number of bytes in actual use |
2130 | * by the object. This is the potential offset to the free pointer. | |
81819f0f CL |
2131 | */ |
2132 | s->inuse = size; | |
2133 | ||
2134 | if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || | |
c59def9f | 2135 | s->ctor)) { |
81819f0f CL |
2136 | /* |
2137 | * Relocate free pointer after the object if it is not | |
2138 | * permitted to overwrite the first word of the object on | |
2139 | * kmem_cache_free. | |
2140 | * | |
2141 | * This is the case if we do RCU, have a constructor or | |
2142 | * destructor or are poisoning the objects. | |
2143 | */ | |
2144 | s->offset = size; | |
2145 | size += sizeof(void *); | |
2146 | } | |
2147 | ||
c12b3c62 | 2148 | #ifdef CONFIG_SLUB_DEBUG |
81819f0f CL |
2149 | if (flags & SLAB_STORE_USER) |
2150 | /* | |
2151 | * Need to store information about allocs and frees after | |
2152 | * the object. | |
2153 | */ | |
2154 | size += 2 * sizeof(struct track); | |
2155 | ||
be7b3fbc | 2156 | if (flags & SLAB_RED_ZONE) |
81819f0f CL |
2157 | /* |
2158 | * Add some empty padding so that we can catch | |
2159 | * overwrites from earlier objects rather than let | |
2160 | * tracking information or the free pointer be | |
2161 | * corrupted if an user writes before the start | |
2162 | * of the object. | |
2163 | */ | |
2164 | size += sizeof(void *); | |
41ecc55b | 2165 | #endif |
672bba3a | 2166 | |
81819f0f CL |
2167 | /* |
2168 | * Determine the alignment based on various parameters that the | |
65c02d4c CL |
2169 | * user specified and the dynamic determination of cache line size |
2170 | * on bootup. | |
81819f0f CL |
2171 | */ |
2172 | align = calculate_alignment(flags, align, s->objsize); | |
2173 | ||
2174 | /* | |
2175 | * SLUB stores one object immediately after another beginning from | |
2176 | * offset 0. In order to align the objects we have to simply size | |
2177 | * each object to conform to the alignment. | |
2178 | */ | |
2179 | size = ALIGN(size, align); | |
2180 | s->size = size; | |
2181 | ||
2182 | s->order = calculate_order(size); | |
2183 | if (s->order < 0) | |
2184 | return 0; | |
2185 | ||
2186 | /* | |
2187 | * Determine the number of objects per slab | |
2188 | */ | |
2189 | s->objects = (PAGE_SIZE << s->order) / size; | |
2190 | ||
b3fba8da | 2191 | return !!s->objects; |
81819f0f CL |
2192 | |
2193 | } | |
2194 | ||
81819f0f CL |
2195 | static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, |
2196 | const char *name, size_t size, | |
2197 | size_t align, unsigned long flags, | |
4ba9b9d0 | 2198 | void (*ctor)(struct kmem_cache *, void *)) |
81819f0f CL |
2199 | { |
2200 | memset(s, 0, kmem_size); | |
2201 | s->name = name; | |
2202 | s->ctor = ctor; | |
81819f0f | 2203 | s->objsize = size; |
81819f0f | 2204 | s->align = align; |
ba0268a8 | 2205 | s->flags = kmem_cache_flags(size, flags, name, ctor); |
81819f0f CL |
2206 | |
2207 | if (!calculate_sizes(s)) | |
2208 | goto error; | |
2209 | ||
2210 | s->refcount = 1; | |
2211 | #ifdef CONFIG_NUMA | |
9824601e | 2212 | s->remote_node_defrag_ratio = 100; |
81819f0f | 2213 | #endif |
dfb4f096 CL |
2214 | if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) |
2215 | goto error; | |
81819f0f | 2216 | |
dfb4f096 | 2217 | if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) |
81819f0f | 2218 | return 1; |
4c93c355 | 2219 | free_kmem_cache_nodes(s); |
81819f0f CL |
2220 | error: |
2221 | if (flags & SLAB_PANIC) | |
2222 | panic("Cannot create slab %s size=%lu realsize=%u " | |
2223 | "order=%u offset=%u flags=%lx\n", | |
2224 | s->name, (unsigned long)size, s->size, s->order, | |
2225 | s->offset, flags); | |
2226 | return 0; | |
2227 | } | |
81819f0f CL |
2228 | |
2229 | /* | |
2230 | * Check if a given pointer is valid | |
2231 | */ | |
2232 | int kmem_ptr_validate(struct kmem_cache *s, const void *object) | |
2233 | { | |
2234 | struct page * page; | |
81819f0f CL |
2235 | |
2236 | page = get_object_page(object); | |
2237 | ||
2238 | if (!page || s != page->slab) | |
2239 | /* No slab or wrong slab */ | |
2240 | return 0; | |
2241 | ||
abcd08a6 | 2242 | if (!check_valid_pointer(s, page, object)) |
81819f0f CL |
2243 | return 0; |
2244 | ||
2245 | /* | |
2246 | * We could also check if the object is on the slabs freelist. | |
2247 | * But this would be too expensive and it seems that the main | |
2248 | * purpose of kmem_ptr_valid is to check if the object belongs | |
2249 | * to a certain slab. | |
2250 | */ | |
2251 | return 1; | |
2252 | } | |
2253 | EXPORT_SYMBOL(kmem_ptr_validate); | |
2254 | ||
2255 | /* | |
2256 | * Determine the size of a slab object | |
2257 | */ | |
2258 | unsigned int kmem_cache_size(struct kmem_cache *s) | |
2259 | { | |
2260 | return s->objsize; | |
2261 | } | |
2262 | EXPORT_SYMBOL(kmem_cache_size); | |
2263 | ||
2264 | const char *kmem_cache_name(struct kmem_cache *s) | |
2265 | { | |
2266 | return s->name; | |
2267 | } | |
2268 | EXPORT_SYMBOL(kmem_cache_name); | |
2269 | ||
2270 | /* | |
672bba3a CL |
2271 | * Attempt to free all slabs on a node. Return the number of slabs we |
2272 | * were unable to free. | |
81819f0f CL |
2273 | */ |
2274 | static int free_list(struct kmem_cache *s, struct kmem_cache_node *n, | |
2275 | struct list_head *list) | |
2276 | { | |
2277 | int slabs_inuse = 0; | |
2278 | unsigned long flags; | |
2279 | struct page *page, *h; | |
2280 | ||
2281 | spin_lock_irqsave(&n->list_lock, flags); | |
2282 | list_for_each_entry_safe(page, h, list, lru) | |
2283 | if (!page->inuse) { | |
2284 | list_del(&page->lru); | |
2285 | discard_slab(s, page); | |
2286 | } else | |
2287 | slabs_inuse++; | |
2288 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2289 | return slabs_inuse; | |
2290 | } | |
2291 | ||
2292 | /* | |
672bba3a | 2293 | * Release all resources used by a slab cache. |
81819f0f | 2294 | */ |
0c710013 | 2295 | static inline int kmem_cache_close(struct kmem_cache *s) |
81819f0f CL |
2296 | { |
2297 | int node; | |
2298 | ||
2299 | flush_all(s); | |
2300 | ||
2301 | /* Attempt to free all objects */ | |
4c93c355 | 2302 | free_kmem_cache_cpus(s); |
f64dc58c | 2303 | for_each_node_state(node, N_NORMAL_MEMORY) { |
81819f0f CL |
2304 | struct kmem_cache_node *n = get_node(s, node); |
2305 | ||
2086d26a | 2306 | n->nr_partial -= free_list(s, n, &n->partial); |
81819f0f CL |
2307 | if (atomic_long_read(&n->nr_slabs)) |
2308 | return 1; | |
2309 | } | |
2310 | free_kmem_cache_nodes(s); | |
2311 | return 0; | |
2312 | } | |
2313 | ||
2314 | /* | |
2315 | * Close a cache and release the kmem_cache structure | |
2316 | * (must be used for caches created using kmem_cache_create) | |
2317 | */ | |
2318 | void kmem_cache_destroy(struct kmem_cache *s) | |
2319 | { | |
2320 | down_write(&slub_lock); | |
2321 | s->refcount--; | |
2322 | if (!s->refcount) { | |
2323 | list_del(&s->list); | |
a0e1d1be | 2324 | up_write(&slub_lock); |
81819f0f CL |
2325 | if (kmem_cache_close(s)) |
2326 | WARN_ON(1); | |
2327 | sysfs_slab_remove(s); | |
a0e1d1be CL |
2328 | } else |
2329 | up_write(&slub_lock); | |
81819f0f CL |
2330 | } |
2331 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2332 | ||
2333 | /******************************************************************** | |
2334 | * Kmalloc subsystem | |
2335 | *******************************************************************/ | |
2336 | ||
aadb4bc4 | 2337 | struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned; |
81819f0f CL |
2338 | EXPORT_SYMBOL(kmalloc_caches); |
2339 | ||
2340 | #ifdef CONFIG_ZONE_DMA | |
aadb4bc4 | 2341 | static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT]; |
81819f0f CL |
2342 | #endif |
2343 | ||
2344 | static int __init setup_slub_min_order(char *str) | |
2345 | { | |
2346 | get_option (&str, &slub_min_order); | |
2347 | ||
2348 | return 1; | |
2349 | } | |
2350 | ||
2351 | __setup("slub_min_order=", setup_slub_min_order); | |
2352 | ||
2353 | static int __init setup_slub_max_order(char *str) | |
2354 | { | |
2355 | get_option (&str, &slub_max_order); | |
2356 | ||
2357 | return 1; | |
2358 | } | |
2359 | ||
2360 | __setup("slub_max_order=", setup_slub_max_order); | |
2361 | ||
2362 | static int __init setup_slub_min_objects(char *str) | |
2363 | { | |
2364 | get_option (&str, &slub_min_objects); | |
2365 | ||
2366 | return 1; | |
2367 | } | |
2368 | ||
2369 | __setup("slub_min_objects=", setup_slub_min_objects); | |
2370 | ||
2371 | static int __init setup_slub_nomerge(char *str) | |
2372 | { | |
2373 | slub_nomerge = 1; | |
2374 | return 1; | |
2375 | } | |
2376 | ||
2377 | __setup("slub_nomerge", setup_slub_nomerge); | |
2378 | ||
81819f0f CL |
2379 | static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, |
2380 | const char *name, int size, gfp_t gfp_flags) | |
2381 | { | |
2382 | unsigned int flags = 0; | |
2383 | ||
2384 | if (gfp_flags & SLUB_DMA) | |
2385 | flags = SLAB_CACHE_DMA; | |
2386 | ||
2387 | down_write(&slub_lock); | |
2388 | if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, | |
c59def9f | 2389 | flags, NULL)) |
81819f0f CL |
2390 | goto panic; |
2391 | ||
2392 | list_add(&s->list, &slab_caches); | |
2393 | up_write(&slub_lock); | |
2394 | if (sysfs_slab_add(s)) | |
2395 | goto panic; | |
2396 | return s; | |
2397 | ||
2398 | panic: | |
2399 | panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); | |
2400 | } | |
2401 | ||
2e443fd0 | 2402 | #ifdef CONFIG_ZONE_DMA |
1ceef402 CL |
2403 | |
2404 | static void sysfs_add_func(struct work_struct *w) | |
2405 | { | |
2406 | struct kmem_cache *s; | |
2407 | ||
2408 | down_write(&slub_lock); | |
2409 | list_for_each_entry(s, &slab_caches, list) { | |
2410 | if (s->flags & __SYSFS_ADD_DEFERRED) { | |
2411 | s->flags &= ~__SYSFS_ADD_DEFERRED; | |
2412 | sysfs_slab_add(s); | |
2413 | } | |
2414 | } | |
2415 | up_write(&slub_lock); | |
2416 | } | |
2417 | ||
2418 | static DECLARE_WORK(sysfs_add_work, sysfs_add_func); | |
2419 | ||
2e443fd0 CL |
2420 | static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) |
2421 | { | |
2422 | struct kmem_cache *s; | |
2e443fd0 CL |
2423 | char *text; |
2424 | size_t realsize; | |
2425 | ||
2426 | s = kmalloc_caches_dma[index]; | |
2427 | if (s) | |
2428 | return s; | |
2429 | ||
2430 | /* Dynamically create dma cache */ | |
1ceef402 CL |
2431 | if (flags & __GFP_WAIT) |
2432 | down_write(&slub_lock); | |
2433 | else { | |
2434 | if (!down_write_trylock(&slub_lock)) | |
2435 | goto out; | |
2436 | } | |
2437 | ||
2438 | if (kmalloc_caches_dma[index]) | |
2439 | goto unlock_out; | |
2e443fd0 | 2440 | |
7b55f620 | 2441 | realsize = kmalloc_caches[index].objsize; |
1ceef402 CL |
2442 | text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize), |
2443 | s = kmalloc(kmem_size, flags & ~SLUB_DMA); | |
2444 | ||
2445 | if (!s || !text || !kmem_cache_open(s, flags, text, | |
2446 | realsize, ARCH_KMALLOC_MINALIGN, | |
2447 | SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { | |
2448 | kfree(s); | |
2449 | kfree(text); | |
2450 | goto unlock_out; | |
dfce8648 | 2451 | } |
1ceef402 CL |
2452 | |
2453 | list_add(&s->list, &slab_caches); | |
2454 | kmalloc_caches_dma[index] = s; | |
2455 | ||
2456 | schedule_work(&sysfs_add_work); | |
2457 | ||
2458 | unlock_out: | |
dfce8648 | 2459 | up_write(&slub_lock); |
1ceef402 | 2460 | out: |
dfce8648 | 2461 | return kmalloc_caches_dma[index]; |
2e443fd0 CL |
2462 | } |
2463 | #endif | |
2464 | ||
f1b26339 CL |
2465 | /* |
2466 | * Conversion table for small slabs sizes / 8 to the index in the | |
2467 | * kmalloc array. This is necessary for slabs < 192 since we have non power | |
2468 | * of two cache sizes there. The size of larger slabs can be determined using | |
2469 | * fls. | |
2470 | */ | |
2471 | static s8 size_index[24] = { | |
2472 | 3, /* 8 */ | |
2473 | 4, /* 16 */ | |
2474 | 5, /* 24 */ | |
2475 | 5, /* 32 */ | |
2476 | 6, /* 40 */ | |
2477 | 6, /* 48 */ | |
2478 | 6, /* 56 */ | |
2479 | 6, /* 64 */ | |
2480 | 1, /* 72 */ | |
2481 | 1, /* 80 */ | |
2482 | 1, /* 88 */ | |
2483 | 1, /* 96 */ | |
2484 | 7, /* 104 */ | |
2485 | 7, /* 112 */ | |
2486 | 7, /* 120 */ | |
2487 | 7, /* 128 */ | |
2488 | 2, /* 136 */ | |
2489 | 2, /* 144 */ | |
2490 | 2, /* 152 */ | |
2491 | 2, /* 160 */ | |
2492 | 2, /* 168 */ | |
2493 | 2, /* 176 */ | |
2494 | 2, /* 184 */ | |
2495 | 2 /* 192 */ | |
2496 | }; | |
2497 | ||
81819f0f CL |
2498 | static struct kmem_cache *get_slab(size_t size, gfp_t flags) |
2499 | { | |
f1b26339 | 2500 | int index; |
81819f0f | 2501 | |
f1b26339 CL |
2502 | if (size <= 192) { |
2503 | if (!size) | |
2504 | return ZERO_SIZE_PTR; | |
81819f0f | 2505 | |
f1b26339 | 2506 | index = size_index[(size - 1) / 8]; |
aadb4bc4 | 2507 | } else |
f1b26339 | 2508 | index = fls(size - 1); |
81819f0f CL |
2509 | |
2510 | #ifdef CONFIG_ZONE_DMA | |
f1b26339 | 2511 | if (unlikely((flags & SLUB_DMA))) |
2e443fd0 | 2512 | return dma_kmalloc_cache(index, flags); |
f1b26339 | 2513 | |
81819f0f CL |
2514 | #endif |
2515 | return &kmalloc_caches[index]; | |
2516 | } | |
2517 | ||
2518 | void *__kmalloc(size_t size, gfp_t flags) | |
2519 | { | |
aadb4bc4 | 2520 | struct kmem_cache *s; |
81819f0f | 2521 | |
aadb4bc4 CL |
2522 | if (unlikely(size > PAGE_SIZE / 2)) |
2523 | return (void *)__get_free_pages(flags | __GFP_COMP, | |
2524 | get_order(size)); | |
2525 | ||
2526 | s = get_slab(size, flags); | |
2527 | ||
2528 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
6cb8f913 CL |
2529 | return s; |
2530 | ||
ce15fea8 | 2531 | return slab_alloc(s, flags, -1, __builtin_return_address(0)); |
81819f0f CL |
2532 | } |
2533 | EXPORT_SYMBOL(__kmalloc); | |
2534 | ||
2535 | #ifdef CONFIG_NUMA | |
2536 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
2537 | { | |
aadb4bc4 | 2538 | struct kmem_cache *s; |
81819f0f | 2539 | |
aadb4bc4 CL |
2540 | if (unlikely(size > PAGE_SIZE / 2)) |
2541 | return (void *)__get_free_pages(flags | __GFP_COMP, | |
2542 | get_order(size)); | |
2543 | ||
2544 | s = get_slab(size, flags); | |
2545 | ||
2546 | if (unlikely(ZERO_OR_NULL_PTR(s))) | |
6cb8f913 CL |
2547 | return s; |
2548 | ||
ce15fea8 | 2549 | return slab_alloc(s, flags, node, __builtin_return_address(0)); |
81819f0f CL |
2550 | } |
2551 | EXPORT_SYMBOL(__kmalloc_node); | |
2552 | #endif | |
2553 | ||
2554 | size_t ksize(const void *object) | |
2555 | { | |
272c1d21 | 2556 | struct page *page; |
81819f0f CL |
2557 | struct kmem_cache *s; |
2558 | ||
ef8b4520 CL |
2559 | BUG_ON(!object); |
2560 | if (unlikely(object == ZERO_SIZE_PTR)) | |
272c1d21 CL |
2561 | return 0; |
2562 | ||
294a80a8 | 2563 | page = virt_to_head_page(object); |
81819f0f | 2564 | BUG_ON(!page); |
294a80a8 VN |
2565 | |
2566 | if (unlikely(!PageSlab(page))) | |
2567 | return PAGE_SIZE << compound_order(page); | |
2568 | ||
81819f0f CL |
2569 | s = page->slab; |
2570 | BUG_ON(!s); | |
2571 | ||
2572 | /* | |
2573 | * Debugging requires use of the padding between object | |
2574 | * and whatever may come after it. | |
2575 | */ | |
2576 | if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) | |
2577 | return s->objsize; | |
2578 | ||
2579 | /* | |
2580 | * If we have the need to store the freelist pointer | |
2581 | * back there or track user information then we can | |
2582 | * only use the space before that information. | |
2583 | */ | |
2584 | if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) | |
2585 | return s->inuse; | |
2586 | ||
2587 | /* | |
2588 | * Else we can use all the padding etc for the allocation | |
2589 | */ | |
2590 | return s->size; | |
2591 | } | |
2592 | EXPORT_SYMBOL(ksize); | |
2593 | ||
2594 | void kfree(const void *x) | |
2595 | { | |
81819f0f CL |
2596 | struct page *page; |
2597 | ||
2408c550 | 2598 | if (unlikely(ZERO_OR_NULL_PTR(x))) |
81819f0f CL |
2599 | return; |
2600 | ||
b49af68f | 2601 | page = virt_to_head_page(x); |
aadb4bc4 CL |
2602 | if (unlikely(!PageSlab(page))) { |
2603 | put_page(page); | |
2604 | return; | |
2605 | } | |
2606 | slab_free(page->slab, page, (void *)x, __builtin_return_address(0)); | |
81819f0f CL |
2607 | } |
2608 | EXPORT_SYMBOL(kfree); | |
2609 | ||
f61396ae CL |
2610 | static unsigned long count_partial(struct kmem_cache_node *n) |
2611 | { | |
2612 | unsigned long flags; | |
2613 | unsigned long x = 0; | |
2614 | struct page *page; | |
2615 | ||
2616 | spin_lock_irqsave(&n->list_lock, flags); | |
2617 | list_for_each_entry(page, &n->partial, lru) | |
2618 | x += page->inuse; | |
2619 | spin_unlock_irqrestore(&n->list_lock, flags); | |
2620 | return x; | |
2621 | } | |
2622 | ||
2086d26a | 2623 | /* |
672bba3a CL |
2624 | * kmem_cache_shrink removes empty slabs from the partial lists and sorts |
2625 | * the remaining slabs by the number of items in use. The slabs with the | |
2626 | * most items in use come first. New allocations will then fill those up | |
2627 | * and thus they can be removed from the partial lists. | |
2628 | * | |
2629 | * The slabs with the least items are placed last. This results in them | |
2630 | * being allocated from last increasing the chance that the last objects | |
2631 | * are freed in them. | |
2086d26a CL |
2632 | */ |
2633 | int kmem_cache_shrink(struct kmem_cache *s) | |
2634 | { | |
2635 | int node; | |
2636 | int i; | |
2637 | struct kmem_cache_node *n; | |
2638 | struct page *page; | |
2639 | struct page *t; | |
2640 | struct list_head *slabs_by_inuse = | |
2641 | kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL); | |
2642 | unsigned long flags; | |
2643 | ||
2644 | if (!slabs_by_inuse) | |
2645 | return -ENOMEM; | |
2646 | ||
2647 | flush_all(s); | |
f64dc58c | 2648 | for_each_node_state(node, N_NORMAL_MEMORY) { |
2086d26a CL |
2649 | n = get_node(s, node); |
2650 | ||
2651 | if (!n->nr_partial) | |
2652 | continue; | |
2653 | ||
2654 | for (i = 0; i < s->objects; i++) | |
2655 | INIT_LIST_HEAD(slabs_by_inuse + i); | |
2656 | ||
2657 | spin_lock_irqsave(&n->list_lock, flags); | |
2658 | ||
2659 | /* | |
672bba3a | 2660 | * Build lists indexed by the items in use in each slab. |
2086d26a | 2661 | * |
672bba3a CL |
2662 | * Note that concurrent frees may occur while we hold the |
2663 | * list_lock. page->inuse here is the upper limit. | |
2086d26a CL |
2664 | */ |
2665 | list_for_each_entry_safe(page, t, &n->partial, lru) { | |
2666 | if (!page->inuse && slab_trylock(page)) { | |
2667 | /* | |
2668 | * Must hold slab lock here because slab_free | |
2669 | * may have freed the last object and be | |
2670 | * waiting to release the slab. | |
2671 | */ | |
2672 | list_del(&page->lru); | |
2673 | n->nr_partial--; | |
2674 | slab_unlock(page); | |
2675 | discard_slab(s, page); | |
2676 | } else { | |
fcda3d89 CL |
2677 | list_move(&page->lru, |
2678 | slabs_by_inuse + page->inuse); | |
2086d26a CL |
2679 | } |
2680 | } | |
2681 | ||
2086d26a | 2682 | /* |
672bba3a CL |
2683 | * Rebuild the partial list with the slabs filled up most |
2684 | * first and the least used slabs at the end. | |
2086d26a CL |
2685 | */ |
2686 | for (i = s->objects - 1; i >= 0; i--) | |
2687 | list_splice(slabs_by_inuse + i, n->partial.prev); | |
2688 | ||
2086d26a CL |
2689 | spin_unlock_irqrestore(&n->list_lock, flags); |
2690 | } | |
2691 | ||
2692 | kfree(slabs_by_inuse); | |
2693 | return 0; | |
2694 | } | |
2695 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2696 | ||
b9049e23 YG |
2697 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
2698 | static int slab_mem_going_offline_callback(void *arg) | |
2699 | { | |
2700 | struct kmem_cache *s; | |
2701 | ||
2702 | down_read(&slub_lock); | |
2703 | list_for_each_entry(s, &slab_caches, list) | |
2704 | kmem_cache_shrink(s); | |
2705 | up_read(&slub_lock); | |
2706 | ||
2707 | return 0; | |
2708 | } | |
2709 | ||
2710 | static void slab_mem_offline_callback(void *arg) | |
2711 | { | |
2712 | struct kmem_cache_node *n; | |
2713 | struct kmem_cache *s; | |
2714 | struct memory_notify *marg = arg; | |
2715 | int offline_node; | |
2716 | ||
2717 | offline_node = marg->status_change_nid; | |
2718 | ||
2719 | /* | |
2720 | * If the node still has available memory. we need kmem_cache_node | |
2721 | * for it yet. | |
2722 | */ | |
2723 | if (offline_node < 0) | |
2724 | return; | |
2725 | ||
2726 | down_read(&slub_lock); | |
2727 | list_for_each_entry(s, &slab_caches, list) { | |
2728 | n = get_node(s, offline_node); | |
2729 | if (n) { | |
2730 | /* | |
2731 | * if n->nr_slabs > 0, slabs still exist on the node | |
2732 | * that is going down. We were unable to free them, | |
2733 | * and offline_pages() function shoudn't call this | |
2734 | * callback. So, we must fail. | |
2735 | */ | |
27bb628a | 2736 | BUG_ON(atomic_long_read(&n->nr_slabs)); |
b9049e23 YG |
2737 | |
2738 | s->node[offline_node] = NULL; | |
2739 | kmem_cache_free(kmalloc_caches, n); | |
2740 | } | |
2741 | } | |
2742 | up_read(&slub_lock); | |
2743 | } | |
2744 | ||
2745 | static int slab_mem_going_online_callback(void *arg) | |
2746 | { | |
2747 | struct kmem_cache_node *n; | |
2748 | struct kmem_cache *s; | |
2749 | struct memory_notify *marg = arg; | |
2750 | int nid = marg->status_change_nid; | |
2751 | int ret = 0; | |
2752 | ||
2753 | /* | |
2754 | * If the node's memory is already available, then kmem_cache_node is | |
2755 | * already created. Nothing to do. | |
2756 | */ | |
2757 | if (nid < 0) | |
2758 | return 0; | |
2759 | ||
2760 | /* | |
2761 | * We are bringing a node online. No memory is availabe yet. We must | |
2762 | * allocate a kmem_cache_node structure in order to bring the node | |
2763 | * online. | |
2764 | */ | |
2765 | down_read(&slub_lock); | |
2766 | list_for_each_entry(s, &slab_caches, list) { | |
2767 | /* | |
2768 | * XXX: kmem_cache_alloc_node will fallback to other nodes | |
2769 | * since memory is not yet available from the node that | |
2770 | * is brought up. | |
2771 | */ | |
2772 | n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); | |
2773 | if (!n) { | |
2774 | ret = -ENOMEM; | |
2775 | goto out; | |
2776 | } | |
2777 | init_kmem_cache_node(n); | |
2778 | s->node[nid] = n; | |
2779 | } | |
2780 | out: | |
2781 | up_read(&slub_lock); | |
2782 | return ret; | |
2783 | } | |
2784 | ||
2785 | static int slab_memory_callback(struct notifier_block *self, | |
2786 | unsigned long action, void *arg) | |
2787 | { | |
2788 | int ret = 0; | |
2789 | ||
2790 | switch (action) { | |
2791 | case MEM_GOING_ONLINE: | |
2792 | ret = slab_mem_going_online_callback(arg); | |
2793 | break; | |
2794 | case MEM_GOING_OFFLINE: | |
2795 | ret = slab_mem_going_offline_callback(arg); | |
2796 | break; | |
2797 | case MEM_OFFLINE: | |
2798 | case MEM_CANCEL_ONLINE: | |
2799 | slab_mem_offline_callback(arg); | |
2800 | break; | |
2801 | case MEM_ONLINE: | |
2802 | case MEM_CANCEL_OFFLINE: | |
2803 | break; | |
2804 | } | |
2805 | ||
2806 | ret = notifier_from_errno(ret); | |
2807 | return ret; | |
2808 | } | |
2809 | ||
2810 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
2811 | ||
81819f0f CL |
2812 | /******************************************************************** |
2813 | * Basic setup of slabs | |
2814 | *******************************************************************/ | |
2815 | ||
2816 | void __init kmem_cache_init(void) | |
2817 | { | |
2818 | int i; | |
4b356be0 | 2819 | int caches = 0; |
81819f0f | 2820 | |
4c93c355 CL |
2821 | init_alloc_cpu(); |
2822 | ||
81819f0f CL |
2823 | #ifdef CONFIG_NUMA |
2824 | /* | |
2825 | * Must first have the slab cache available for the allocations of the | |
672bba3a | 2826 | * struct kmem_cache_node's. There is special bootstrap code in |
81819f0f CL |
2827 | * kmem_cache_open for slab_state == DOWN. |
2828 | */ | |
2829 | create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", | |
2830 | sizeof(struct kmem_cache_node), GFP_KERNEL); | |
8ffa6875 | 2831 | kmalloc_caches[0].refcount = -1; |
4b356be0 | 2832 | caches++; |
b9049e23 YG |
2833 | |
2834 | hotplug_memory_notifier(slab_memory_callback, 1); | |
81819f0f CL |
2835 | #endif |
2836 | ||
2837 | /* Able to allocate the per node structures */ | |
2838 | slab_state = PARTIAL; | |
2839 | ||
2840 | /* Caches that are not of the two-to-the-power-of size */ | |
4b356be0 CL |
2841 | if (KMALLOC_MIN_SIZE <= 64) { |
2842 | create_kmalloc_cache(&kmalloc_caches[1], | |
81819f0f | 2843 | "kmalloc-96", 96, GFP_KERNEL); |
4b356be0 CL |
2844 | caches++; |
2845 | } | |
2846 | if (KMALLOC_MIN_SIZE <= 128) { | |
2847 | create_kmalloc_cache(&kmalloc_caches[2], | |
81819f0f | 2848 | "kmalloc-192", 192, GFP_KERNEL); |
4b356be0 CL |
2849 | caches++; |
2850 | } | |
81819f0f | 2851 | |
aadb4bc4 | 2852 | for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) { |
81819f0f CL |
2853 | create_kmalloc_cache(&kmalloc_caches[i], |
2854 | "kmalloc", 1 << i, GFP_KERNEL); | |
4b356be0 CL |
2855 | caches++; |
2856 | } | |
81819f0f | 2857 | |
f1b26339 CL |
2858 | |
2859 | /* | |
2860 | * Patch up the size_index table if we have strange large alignment | |
2861 | * requirements for the kmalloc array. This is only the case for | |
2862 | * mips it seems. The standard arches will not generate any code here. | |
2863 | * | |
2864 | * Largest permitted alignment is 256 bytes due to the way we | |
2865 | * handle the index determination for the smaller caches. | |
2866 | * | |
2867 | * Make sure that nothing crazy happens if someone starts tinkering | |
2868 | * around with ARCH_KMALLOC_MINALIGN | |
2869 | */ | |
2870 | BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || | |
2871 | (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); | |
2872 | ||
12ad6843 | 2873 | for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) |
f1b26339 CL |
2874 | size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; |
2875 | ||
81819f0f CL |
2876 | slab_state = UP; |
2877 | ||
2878 | /* Provide the correct kmalloc names now that the caches are up */ | |
aadb4bc4 | 2879 | for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) |
81819f0f CL |
2880 | kmalloc_caches[i]. name = |
2881 | kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); | |
2882 | ||
2883 | #ifdef CONFIG_SMP | |
2884 | register_cpu_notifier(&slab_notifier); | |
4c93c355 CL |
2885 | kmem_size = offsetof(struct kmem_cache, cpu_slab) + |
2886 | nr_cpu_ids * sizeof(struct kmem_cache_cpu *); | |
2887 | #else | |
2888 | kmem_size = sizeof(struct kmem_cache); | |
81819f0f CL |
2889 | #endif |
2890 | ||
81819f0f CL |
2891 | |
2892 | printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," | |
4b356be0 CL |
2893 | " CPUs=%d, Nodes=%d\n", |
2894 | caches, cache_line_size(), | |
81819f0f CL |
2895 | slub_min_order, slub_max_order, slub_min_objects, |
2896 | nr_cpu_ids, nr_node_ids); | |
2897 | } | |
2898 | ||
2899 | /* | |
2900 | * Find a mergeable slab cache | |
2901 | */ | |
2902 | static int slab_unmergeable(struct kmem_cache *s) | |
2903 | { | |
2904 | if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) | |
2905 | return 1; | |
2906 | ||
c59def9f | 2907 | if (s->ctor) |
81819f0f CL |
2908 | return 1; |
2909 | ||
8ffa6875 CL |
2910 | /* |
2911 | * We may have set a slab to be unmergeable during bootstrap. | |
2912 | */ | |
2913 | if (s->refcount < 0) | |
2914 | return 1; | |
2915 | ||
81819f0f CL |
2916 | return 0; |
2917 | } | |
2918 | ||
2919 | static struct kmem_cache *find_mergeable(size_t size, | |
ba0268a8 | 2920 | size_t align, unsigned long flags, const char *name, |
4ba9b9d0 | 2921 | void (*ctor)(struct kmem_cache *, void *)) |
81819f0f | 2922 | { |
5b95a4ac | 2923 | struct kmem_cache *s; |
81819f0f CL |
2924 | |
2925 | if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) | |
2926 | return NULL; | |
2927 | ||
c59def9f | 2928 | if (ctor) |
81819f0f CL |
2929 | return NULL; |
2930 | ||
2931 | size = ALIGN(size, sizeof(void *)); | |
2932 | align = calculate_alignment(flags, align, size); | |
2933 | size = ALIGN(size, align); | |
ba0268a8 | 2934 | flags = kmem_cache_flags(size, flags, name, NULL); |
81819f0f | 2935 | |
5b95a4ac | 2936 | list_for_each_entry(s, &slab_caches, list) { |
81819f0f CL |
2937 | if (slab_unmergeable(s)) |
2938 | continue; | |
2939 | ||
2940 | if (size > s->size) | |
2941 | continue; | |
2942 | ||
ba0268a8 | 2943 | if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) |
81819f0f CL |
2944 | continue; |
2945 | /* | |
2946 | * Check if alignment is compatible. | |
2947 | * Courtesy of Adrian Drzewiecki | |
2948 | */ | |
2949 | if ((s->size & ~(align -1)) != s->size) | |
2950 | continue; | |
2951 | ||
2952 | if (s->size - size >= sizeof(void *)) | |
2953 | continue; | |
2954 | ||
2955 | return s; | |
2956 | } | |
2957 | return NULL; | |
2958 | } | |
2959 | ||
2960 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, | |
2961 | size_t align, unsigned long flags, | |
4ba9b9d0 | 2962 | void (*ctor)(struct kmem_cache *, void *)) |
81819f0f CL |
2963 | { |
2964 | struct kmem_cache *s; | |
2965 | ||
2966 | down_write(&slub_lock); | |
ba0268a8 | 2967 | s = find_mergeable(size, align, flags, name, ctor); |
81819f0f | 2968 | if (s) { |
42a9fdbb CL |
2969 | int cpu; |
2970 | ||
81819f0f CL |
2971 | s->refcount++; |
2972 | /* | |
2973 | * Adjust the object sizes so that we clear | |
2974 | * the complete object on kzalloc. | |
2975 | */ | |
2976 | s->objsize = max(s->objsize, (int)size); | |
42a9fdbb CL |
2977 | |
2978 | /* | |
2979 | * And then we need to update the object size in the | |
2980 | * per cpu structures | |
2981 | */ | |
2982 | for_each_online_cpu(cpu) | |
2983 | get_cpu_slab(s, cpu)->objsize = s->objsize; | |
81819f0f | 2984 | s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); |
a0e1d1be | 2985 | up_write(&slub_lock); |
81819f0f CL |
2986 | if (sysfs_slab_alias(s, name)) |
2987 | goto err; | |
a0e1d1be CL |
2988 | return s; |
2989 | } | |
2990 | s = kmalloc(kmem_size, GFP_KERNEL); | |
2991 | if (s) { | |
2992 | if (kmem_cache_open(s, GFP_KERNEL, name, | |
c59def9f | 2993 | size, align, flags, ctor)) { |
81819f0f | 2994 | list_add(&s->list, &slab_caches); |
a0e1d1be CL |
2995 | up_write(&slub_lock); |
2996 | if (sysfs_slab_add(s)) | |
2997 | goto err; | |
2998 | return s; | |
2999 | } | |
3000 | kfree(s); | |
81819f0f CL |
3001 | } |
3002 | up_write(&slub_lock); | |
81819f0f CL |
3003 | |
3004 | err: | |
81819f0f CL |
3005 | if (flags & SLAB_PANIC) |
3006 | panic("Cannot create slabcache %s\n", name); | |
3007 | else | |
3008 | s = NULL; | |
3009 | return s; | |
3010 | } | |
3011 | EXPORT_SYMBOL(kmem_cache_create); | |
3012 | ||
81819f0f | 3013 | #ifdef CONFIG_SMP |
81819f0f | 3014 | /* |
672bba3a CL |
3015 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
3016 | * necessary. | |
81819f0f CL |
3017 | */ |
3018 | static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, | |
3019 | unsigned long action, void *hcpu) | |
3020 | { | |
3021 | long cpu = (long)hcpu; | |
5b95a4ac CL |
3022 | struct kmem_cache *s; |
3023 | unsigned long flags; | |
81819f0f CL |
3024 | |
3025 | switch (action) { | |
4c93c355 CL |
3026 | case CPU_UP_PREPARE: |
3027 | case CPU_UP_PREPARE_FROZEN: | |
3028 | init_alloc_cpu_cpu(cpu); | |
3029 | down_read(&slub_lock); | |
3030 | list_for_each_entry(s, &slab_caches, list) | |
3031 | s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, | |
3032 | GFP_KERNEL); | |
3033 | up_read(&slub_lock); | |
3034 | break; | |
3035 | ||
81819f0f | 3036 | case CPU_UP_CANCELED: |
8bb78442 | 3037 | case CPU_UP_CANCELED_FROZEN: |
81819f0f | 3038 | case CPU_DEAD: |
8bb78442 | 3039 | case CPU_DEAD_FROZEN: |
5b95a4ac CL |
3040 | down_read(&slub_lock); |
3041 | list_for_each_entry(s, &slab_caches, list) { | |
4c93c355 CL |
3042 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); |
3043 | ||
5b95a4ac CL |
3044 | local_irq_save(flags); |
3045 | __flush_cpu_slab(s, cpu); | |
3046 | local_irq_restore(flags); | |
4c93c355 CL |
3047 | free_kmem_cache_cpu(c, cpu); |
3048 | s->cpu_slab[cpu] = NULL; | |
5b95a4ac CL |
3049 | } |
3050 | up_read(&slub_lock); | |
81819f0f CL |
3051 | break; |
3052 | default: | |
3053 | break; | |
3054 | } | |
3055 | return NOTIFY_OK; | |
3056 | } | |
3057 | ||
3058 | static struct notifier_block __cpuinitdata slab_notifier = | |
3059 | { &slab_cpuup_callback, NULL, 0 }; | |
3060 | ||
3061 | #endif | |
3062 | ||
81819f0f CL |
3063 | void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) |
3064 | { | |
aadb4bc4 CL |
3065 | struct kmem_cache *s; |
3066 | ||
3067 | if (unlikely(size > PAGE_SIZE / 2)) | |
3068 | return (void *)__get_free_pages(gfpflags | __GFP_COMP, | |
3069 | get_order(size)); | |
3070 | s = get_slab(size, gfpflags); | |
81819f0f | 3071 | |
2408c550 | 3072 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
6cb8f913 | 3073 | return s; |
81819f0f | 3074 | |
ce15fea8 | 3075 | return slab_alloc(s, gfpflags, -1, caller); |
81819f0f CL |
3076 | } |
3077 | ||
3078 | void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, | |
3079 | int node, void *caller) | |
3080 | { | |
aadb4bc4 CL |
3081 | struct kmem_cache *s; |
3082 | ||
3083 | if (unlikely(size > PAGE_SIZE / 2)) | |
3084 | return (void *)__get_free_pages(gfpflags | __GFP_COMP, | |
3085 | get_order(size)); | |
3086 | s = get_slab(size, gfpflags); | |
81819f0f | 3087 | |
2408c550 | 3088 | if (unlikely(ZERO_OR_NULL_PTR(s))) |
6cb8f913 | 3089 | return s; |
81819f0f | 3090 | |
ce15fea8 | 3091 | return slab_alloc(s, gfpflags, node, caller); |
81819f0f CL |
3092 | } |
3093 | ||
41ecc55b | 3094 | #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) |
434e245d CL |
3095 | static int validate_slab(struct kmem_cache *s, struct page *page, |
3096 | unsigned long *map) | |
53e15af0 CL |
3097 | { |
3098 | void *p; | |
3099 | void *addr = page_address(page); | |
53e15af0 CL |
3100 | |
3101 | if (!check_slab(s, page) || | |
3102 | !on_freelist(s, page, NULL)) | |
3103 | return 0; | |
3104 | ||
3105 | /* Now we know that a valid freelist exists */ | |
3106 | bitmap_zero(map, s->objects); | |
3107 | ||
7656c72b CL |
3108 | for_each_free_object(p, s, page->freelist) { |
3109 | set_bit(slab_index(p, s, addr), map); | |
53e15af0 CL |
3110 | if (!check_object(s, page, p, 0)) |
3111 | return 0; | |
3112 | } | |
3113 | ||
7656c72b CL |
3114 | for_each_object(p, s, addr) |
3115 | if (!test_bit(slab_index(p, s, addr), map)) | |
53e15af0 CL |
3116 | if (!check_object(s, page, p, 1)) |
3117 | return 0; | |
3118 | return 1; | |
3119 | } | |
3120 | ||
434e245d CL |
3121 | static void validate_slab_slab(struct kmem_cache *s, struct page *page, |
3122 | unsigned long *map) | |
53e15af0 CL |
3123 | { |
3124 | if (slab_trylock(page)) { | |
434e245d | 3125 | validate_slab(s, page, map); |
53e15af0 CL |
3126 | slab_unlock(page); |
3127 | } else | |
3128 | printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", | |
3129 | s->name, page); | |
3130 | ||
3131 | if (s->flags & DEBUG_DEFAULT_FLAGS) { | |
35e5d7ee CL |
3132 | if (!SlabDebug(page)) |
3133 | printk(KERN_ERR "SLUB %s: SlabDebug not set " | |
53e15af0 CL |
3134 | "on slab 0x%p\n", s->name, page); |
3135 | } else { | |
35e5d7ee CL |
3136 | if (SlabDebug(page)) |
3137 | printk(KERN_ERR "SLUB %s: SlabDebug set on " | |
53e15af0 CL |
3138 | "slab 0x%p\n", s->name, page); |
3139 | } | |
3140 | } | |
3141 | ||
434e245d CL |
3142 | static int validate_slab_node(struct kmem_cache *s, |
3143 | struct kmem_cache_node *n, unsigned long *map) | |
53e15af0 CL |
3144 | { |
3145 | unsigned long count = 0; | |
3146 | struct page *page; | |
3147 | unsigned long flags; | |
3148 | ||
3149 | spin_lock_irqsave(&n->list_lock, flags); | |
3150 | ||
3151 | list_for_each_entry(page, &n->partial, lru) { | |
434e245d | 3152 | validate_slab_slab(s, page, map); |
53e15af0 CL |
3153 | count++; |
3154 | } | |
3155 | if (count != n->nr_partial) | |
3156 | printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " | |
3157 | "counter=%ld\n", s->name, count, n->nr_partial); | |
3158 | ||
3159 | if (!(s->flags & SLAB_STORE_USER)) | |
3160 | goto out; | |
3161 | ||
3162 | list_for_each_entry(page, &n->full, lru) { | |
434e245d | 3163 | validate_slab_slab(s, page, map); |
53e15af0 CL |
3164 | count++; |
3165 | } | |
3166 | if (count != atomic_long_read(&n->nr_slabs)) | |
3167 | printk(KERN_ERR "SLUB: %s %ld slabs counted but " | |
3168 | "counter=%ld\n", s->name, count, | |
3169 | atomic_long_read(&n->nr_slabs)); | |
3170 | ||
3171 | out: | |
3172 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3173 | return count; | |
3174 | } | |
3175 | ||
434e245d | 3176 | static long validate_slab_cache(struct kmem_cache *s) |
53e15af0 CL |
3177 | { |
3178 | int node; | |
3179 | unsigned long count = 0; | |
434e245d CL |
3180 | unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) * |
3181 | sizeof(unsigned long), GFP_KERNEL); | |
3182 | ||
3183 | if (!map) | |
3184 | return -ENOMEM; | |
53e15af0 CL |
3185 | |
3186 | flush_all(s); | |
f64dc58c | 3187 | for_each_node_state(node, N_NORMAL_MEMORY) { |
53e15af0 CL |
3188 | struct kmem_cache_node *n = get_node(s, node); |
3189 | ||
434e245d | 3190 | count += validate_slab_node(s, n, map); |
53e15af0 | 3191 | } |
434e245d | 3192 | kfree(map); |
53e15af0 CL |
3193 | return count; |
3194 | } | |
3195 | ||
b3459709 CL |
3196 | #ifdef SLUB_RESILIENCY_TEST |
3197 | static void resiliency_test(void) | |
3198 | { | |
3199 | u8 *p; | |
3200 | ||
3201 | printk(KERN_ERR "SLUB resiliency testing\n"); | |
3202 | printk(KERN_ERR "-----------------------\n"); | |
3203 | printk(KERN_ERR "A. Corruption after allocation\n"); | |
3204 | ||
3205 | p = kzalloc(16, GFP_KERNEL); | |
3206 | p[16] = 0x12; | |
3207 | printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" | |
3208 | " 0x12->0x%p\n\n", p + 16); | |
3209 | ||
3210 | validate_slab_cache(kmalloc_caches + 4); | |
3211 | ||
3212 | /* Hmmm... The next two are dangerous */ | |
3213 | p = kzalloc(32, GFP_KERNEL); | |
3214 | p[32 + sizeof(void *)] = 0x34; | |
3215 | printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" | |
3216 | " 0x34 -> -0x%p\n", p); | |
3217 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
3218 | ||
3219 | validate_slab_cache(kmalloc_caches + 5); | |
3220 | p = kzalloc(64, GFP_KERNEL); | |
3221 | p += 64 + (get_cycles() & 0xff) * sizeof(void *); | |
3222 | *p = 0x56; | |
3223 | printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", | |
3224 | p); | |
3225 | printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); | |
3226 | validate_slab_cache(kmalloc_caches + 6); | |
3227 | ||
3228 | printk(KERN_ERR "\nB. Corruption after free\n"); | |
3229 | p = kzalloc(128, GFP_KERNEL); | |
3230 | kfree(p); | |
3231 | *p = 0x78; | |
3232 | printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); | |
3233 | validate_slab_cache(kmalloc_caches + 7); | |
3234 | ||
3235 | p = kzalloc(256, GFP_KERNEL); | |
3236 | kfree(p); | |
3237 | p[50] = 0x9a; | |
3238 | printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); | |
3239 | validate_slab_cache(kmalloc_caches + 8); | |
3240 | ||
3241 | p = kzalloc(512, GFP_KERNEL); | |
3242 | kfree(p); | |
3243 | p[512] = 0xab; | |
3244 | printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); | |
3245 | validate_slab_cache(kmalloc_caches + 9); | |
3246 | } | |
3247 | #else | |
3248 | static void resiliency_test(void) {}; | |
3249 | #endif | |
3250 | ||
88a420e4 | 3251 | /* |
672bba3a | 3252 | * Generate lists of code addresses where slabcache objects are allocated |
88a420e4 CL |
3253 | * and freed. |
3254 | */ | |
3255 | ||
3256 | struct location { | |
3257 | unsigned long count; | |
3258 | void *addr; | |
45edfa58 CL |
3259 | long long sum_time; |
3260 | long min_time; | |
3261 | long max_time; | |
3262 | long min_pid; | |
3263 | long max_pid; | |
3264 | cpumask_t cpus; | |
3265 | nodemask_t nodes; | |
88a420e4 CL |
3266 | }; |
3267 | ||
3268 | struct loc_track { | |
3269 | unsigned long max; | |
3270 | unsigned long count; | |
3271 | struct location *loc; | |
3272 | }; | |
3273 | ||
3274 | static void free_loc_track(struct loc_track *t) | |
3275 | { | |
3276 | if (t->max) | |
3277 | free_pages((unsigned long)t->loc, | |
3278 | get_order(sizeof(struct location) * t->max)); | |
3279 | } | |
3280 | ||
68dff6a9 | 3281 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
88a420e4 CL |
3282 | { |
3283 | struct location *l; | |
3284 | int order; | |
3285 | ||
88a420e4 CL |
3286 | order = get_order(sizeof(struct location) * max); |
3287 | ||
68dff6a9 | 3288 | l = (void *)__get_free_pages(flags, order); |
88a420e4 CL |
3289 | if (!l) |
3290 | return 0; | |
3291 | ||
3292 | if (t->count) { | |
3293 | memcpy(l, t->loc, sizeof(struct location) * t->count); | |
3294 | free_loc_track(t); | |
3295 | } | |
3296 | t->max = max; | |
3297 | t->loc = l; | |
3298 | return 1; | |
3299 | } | |
3300 | ||
3301 | static int add_location(struct loc_track *t, struct kmem_cache *s, | |
45edfa58 | 3302 | const struct track *track) |
88a420e4 CL |
3303 | { |
3304 | long start, end, pos; | |
3305 | struct location *l; | |
3306 | void *caddr; | |
45edfa58 | 3307 | unsigned long age = jiffies - track->when; |
88a420e4 CL |
3308 | |
3309 | start = -1; | |
3310 | end = t->count; | |
3311 | ||
3312 | for ( ; ; ) { | |
3313 | pos = start + (end - start + 1) / 2; | |
3314 | ||
3315 | /* | |
3316 | * There is nothing at "end". If we end up there | |
3317 | * we need to add something to before end. | |
3318 | */ | |
3319 | if (pos == end) | |
3320 | break; | |
3321 | ||
3322 | caddr = t->loc[pos].addr; | |
45edfa58 CL |
3323 | if (track->addr == caddr) { |
3324 | ||
3325 | l = &t->loc[pos]; | |
3326 | l->count++; | |
3327 | if (track->when) { | |
3328 | l->sum_time += age; | |
3329 | if (age < l->min_time) | |
3330 | l->min_time = age; | |
3331 | if (age > l->max_time) | |
3332 | l->max_time = age; | |
3333 | ||
3334 | if (track->pid < l->min_pid) | |
3335 | l->min_pid = track->pid; | |
3336 | if (track->pid > l->max_pid) | |
3337 | l->max_pid = track->pid; | |
3338 | ||
3339 | cpu_set(track->cpu, l->cpus); | |
3340 | } | |
3341 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
88a420e4 CL |
3342 | return 1; |
3343 | } | |
3344 | ||
45edfa58 | 3345 | if (track->addr < caddr) |
88a420e4 CL |
3346 | end = pos; |
3347 | else | |
3348 | start = pos; | |
3349 | } | |
3350 | ||
3351 | /* | |
672bba3a | 3352 | * Not found. Insert new tracking element. |
88a420e4 | 3353 | */ |
68dff6a9 | 3354 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
88a420e4 CL |
3355 | return 0; |
3356 | ||
3357 | l = t->loc + pos; | |
3358 | if (pos < t->count) | |
3359 | memmove(l + 1, l, | |
3360 | (t->count - pos) * sizeof(struct location)); | |
3361 | t->count++; | |
3362 | l->count = 1; | |
45edfa58 CL |
3363 | l->addr = track->addr; |
3364 | l->sum_time = age; | |
3365 | l->min_time = age; | |
3366 | l->max_time = age; | |
3367 | l->min_pid = track->pid; | |
3368 | l->max_pid = track->pid; | |
3369 | cpus_clear(l->cpus); | |
3370 | cpu_set(track->cpu, l->cpus); | |
3371 | nodes_clear(l->nodes); | |
3372 | node_set(page_to_nid(virt_to_page(track)), l->nodes); | |
88a420e4 CL |
3373 | return 1; |
3374 | } | |
3375 | ||
3376 | static void process_slab(struct loc_track *t, struct kmem_cache *s, | |
3377 | struct page *page, enum track_item alloc) | |
3378 | { | |
3379 | void *addr = page_address(page); | |
7656c72b | 3380 | DECLARE_BITMAP(map, s->objects); |
88a420e4 CL |
3381 | void *p; |
3382 | ||
3383 | bitmap_zero(map, s->objects); | |
7656c72b CL |
3384 | for_each_free_object(p, s, page->freelist) |
3385 | set_bit(slab_index(p, s, addr), map); | |
88a420e4 | 3386 | |
7656c72b | 3387 | for_each_object(p, s, addr) |
45edfa58 CL |
3388 | if (!test_bit(slab_index(p, s, addr), map)) |
3389 | add_location(t, s, get_track(s, p, alloc)); | |
88a420e4 CL |
3390 | } |
3391 | ||
3392 | static int list_locations(struct kmem_cache *s, char *buf, | |
3393 | enum track_item alloc) | |
3394 | { | |
e374d483 | 3395 | int len = 0; |
88a420e4 | 3396 | unsigned long i; |
68dff6a9 | 3397 | struct loc_track t = { 0, 0, NULL }; |
88a420e4 CL |
3398 | int node; |
3399 | ||
68dff6a9 | 3400 | if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), |
ea3061d2 | 3401 | GFP_TEMPORARY)) |
68dff6a9 | 3402 | return sprintf(buf, "Out of memory\n"); |
88a420e4 CL |
3403 | |
3404 | /* Push back cpu slabs */ | |
3405 | flush_all(s); | |
3406 | ||
f64dc58c | 3407 | for_each_node_state(node, N_NORMAL_MEMORY) { |
88a420e4 CL |
3408 | struct kmem_cache_node *n = get_node(s, node); |
3409 | unsigned long flags; | |
3410 | struct page *page; | |
3411 | ||
9e86943b | 3412 | if (!atomic_long_read(&n->nr_slabs)) |
88a420e4 CL |
3413 | continue; |
3414 | ||
3415 | spin_lock_irqsave(&n->list_lock, flags); | |
3416 | list_for_each_entry(page, &n->partial, lru) | |
3417 | process_slab(&t, s, page, alloc); | |
3418 | list_for_each_entry(page, &n->full, lru) | |
3419 | process_slab(&t, s, page, alloc); | |
3420 | spin_unlock_irqrestore(&n->list_lock, flags); | |
3421 | } | |
3422 | ||
3423 | for (i = 0; i < t.count; i++) { | |
45edfa58 | 3424 | struct location *l = &t.loc[i]; |
88a420e4 | 3425 | |
e374d483 | 3426 | if (len > PAGE_SIZE - 100) |
88a420e4 | 3427 | break; |
e374d483 | 3428 | len += sprintf(buf + len, "%7ld ", l->count); |
45edfa58 CL |
3429 | |
3430 | if (l->addr) | |
e374d483 | 3431 | len += sprint_symbol(buf + len, (unsigned long)l->addr); |
88a420e4 | 3432 | else |
e374d483 | 3433 | len += sprintf(buf + len, "<not-available>"); |
45edfa58 CL |
3434 | |
3435 | if (l->sum_time != l->min_time) { | |
3436 | unsigned long remainder; | |
3437 | ||
e374d483 | 3438 | len += sprintf(buf + len, " age=%ld/%ld/%ld", |
45edfa58 CL |
3439 | l->min_time, |
3440 | div_long_long_rem(l->sum_time, l->count, &remainder), | |
3441 | l->max_time); | |
3442 | } else | |
e374d483 | 3443 | len += sprintf(buf + len, " age=%ld", |
45edfa58 CL |
3444 | l->min_time); |
3445 | ||
3446 | if (l->min_pid != l->max_pid) | |
e374d483 | 3447 | len += sprintf(buf + len, " pid=%ld-%ld", |
45edfa58 CL |
3448 | l->min_pid, l->max_pid); |
3449 | else | |
e374d483 | 3450 | len += sprintf(buf + len, " pid=%ld", |
45edfa58 CL |
3451 | l->min_pid); |
3452 | ||
84966343 | 3453 | if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && |
e374d483 HH |
3454 | len < PAGE_SIZE - 60) { |
3455 | len += sprintf(buf + len, " cpus="); | |
3456 | len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
45edfa58 CL |
3457 | l->cpus); |
3458 | } | |
3459 | ||
84966343 | 3460 | if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && |
e374d483 HH |
3461 | len < PAGE_SIZE - 60) { |
3462 | len += sprintf(buf + len, " nodes="); | |
3463 | len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, | |
45edfa58 CL |
3464 | l->nodes); |
3465 | } | |
3466 | ||
e374d483 | 3467 | len += sprintf(buf + len, "\n"); |
88a420e4 CL |
3468 | } |
3469 | ||
3470 | free_loc_track(&t); | |
3471 | if (!t.count) | |
e374d483 HH |
3472 | len += sprintf(buf, "No data\n"); |
3473 | return len; | |
88a420e4 CL |
3474 | } |
3475 | ||
81819f0f CL |
3476 | enum slab_stat_type { |
3477 | SL_FULL, | |
3478 | SL_PARTIAL, | |
3479 | SL_CPU, | |
3480 | SL_OBJECTS | |
3481 | }; | |
3482 | ||
3483 | #define SO_FULL (1 << SL_FULL) | |
3484 | #define SO_PARTIAL (1 << SL_PARTIAL) | |
3485 | #define SO_CPU (1 << SL_CPU) | |
3486 | #define SO_OBJECTS (1 << SL_OBJECTS) | |
3487 | ||
3488 | static unsigned long slab_objects(struct kmem_cache *s, | |
3489 | char *buf, unsigned long flags) | |
3490 | { | |
3491 | unsigned long total = 0; | |
3492 | int cpu; | |
3493 | int node; | |
3494 | int x; | |
3495 | unsigned long *nodes; | |
3496 | unsigned long *per_cpu; | |
3497 | ||
3498 | nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); | |
3499 | per_cpu = nodes + nr_node_ids; | |
3500 | ||
3501 | for_each_possible_cpu(cpu) { | |
dfb4f096 CL |
3502 | struct page *page; |
3503 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
81819f0f | 3504 | |
dfb4f096 CL |
3505 | if (!c) |
3506 | continue; | |
3507 | ||
3508 | page = c->page; | |
ee3c72a1 CL |
3509 | node = c->node; |
3510 | if (node < 0) | |
3511 | continue; | |
81819f0f | 3512 | if (page) { |
81819f0f | 3513 | if (flags & SO_CPU) { |
81819f0f CL |
3514 | if (flags & SO_OBJECTS) |
3515 | x = page->inuse; | |
3516 | else | |
3517 | x = 1; | |
3518 | total += x; | |
ee3c72a1 | 3519 | nodes[node] += x; |
81819f0f | 3520 | } |
ee3c72a1 | 3521 | per_cpu[node]++; |
81819f0f CL |
3522 | } |
3523 | } | |
3524 | ||
f64dc58c | 3525 | for_each_node_state(node, N_NORMAL_MEMORY) { |
81819f0f CL |
3526 | struct kmem_cache_node *n = get_node(s, node); |
3527 | ||
3528 | if (flags & SO_PARTIAL) { | |
3529 | if (flags & SO_OBJECTS) | |
3530 | x = count_partial(n); | |
3531 | else | |
3532 | x = n->nr_partial; | |
3533 | total += x; | |
3534 | nodes[node] += x; | |
3535 | } | |
3536 | ||
3537 | if (flags & SO_FULL) { | |
9e86943b | 3538 | int full_slabs = atomic_long_read(&n->nr_slabs) |
81819f0f CL |
3539 | - per_cpu[node] |
3540 | - n->nr_partial; | |
3541 | ||
3542 | if (flags & SO_OBJECTS) | |
3543 | x = full_slabs * s->objects; | |
3544 | else | |
3545 | x = full_slabs; | |
3546 | total += x; | |
3547 | nodes[node] += x; | |
3548 | } | |
3549 | } | |
3550 | ||
3551 | x = sprintf(buf, "%lu", total); | |
3552 | #ifdef CONFIG_NUMA | |
f64dc58c | 3553 | for_each_node_state(node, N_NORMAL_MEMORY) |
81819f0f CL |
3554 | if (nodes[node]) |
3555 | x += sprintf(buf + x, " N%d=%lu", | |
3556 | node, nodes[node]); | |
3557 | #endif | |
3558 | kfree(nodes); | |
3559 | return x + sprintf(buf + x, "\n"); | |
3560 | } | |
3561 | ||
3562 | static int any_slab_objects(struct kmem_cache *s) | |
3563 | { | |
3564 | int node; | |
3565 | int cpu; | |
3566 | ||
dfb4f096 CL |
3567 | for_each_possible_cpu(cpu) { |
3568 | struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); | |
3569 | ||
3570 | if (c && c->page) | |
81819f0f | 3571 | return 1; |
dfb4f096 | 3572 | } |
81819f0f | 3573 | |
dfb4f096 | 3574 | for_each_online_node(node) { |
81819f0f CL |
3575 | struct kmem_cache_node *n = get_node(s, node); |
3576 | ||
dfb4f096 CL |
3577 | if (!n) |
3578 | continue; | |
3579 | ||
9e86943b | 3580 | if (n->nr_partial || atomic_long_read(&n->nr_slabs)) |
81819f0f CL |
3581 | return 1; |
3582 | } | |
3583 | return 0; | |
3584 | } | |
3585 | ||
3586 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) | |
3587 | #define to_slab(n) container_of(n, struct kmem_cache, kobj); | |
3588 | ||
3589 | struct slab_attribute { | |
3590 | struct attribute attr; | |
3591 | ssize_t (*show)(struct kmem_cache *s, char *buf); | |
3592 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); | |
3593 | }; | |
3594 | ||
3595 | #define SLAB_ATTR_RO(_name) \ | |
3596 | static struct slab_attribute _name##_attr = __ATTR_RO(_name) | |
3597 | ||
3598 | #define SLAB_ATTR(_name) \ | |
3599 | static struct slab_attribute _name##_attr = \ | |
3600 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3601 | ||
81819f0f CL |
3602 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
3603 | { | |
3604 | return sprintf(buf, "%d\n", s->size); | |
3605 | } | |
3606 | SLAB_ATTR_RO(slab_size); | |
3607 | ||
3608 | static ssize_t align_show(struct kmem_cache *s, char *buf) | |
3609 | { | |
3610 | return sprintf(buf, "%d\n", s->align); | |
3611 | } | |
3612 | SLAB_ATTR_RO(align); | |
3613 | ||
3614 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) | |
3615 | { | |
3616 | return sprintf(buf, "%d\n", s->objsize); | |
3617 | } | |
3618 | SLAB_ATTR_RO(object_size); | |
3619 | ||
3620 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) | |
3621 | { | |
3622 | return sprintf(buf, "%d\n", s->objects); | |
3623 | } | |
3624 | SLAB_ATTR_RO(objs_per_slab); | |
3625 | ||
3626 | static ssize_t order_show(struct kmem_cache *s, char *buf) | |
3627 | { | |
3628 | return sprintf(buf, "%d\n", s->order); | |
3629 | } | |
3630 | SLAB_ATTR_RO(order); | |
3631 | ||
3632 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) | |
3633 | { | |
3634 | if (s->ctor) { | |
3635 | int n = sprint_symbol(buf, (unsigned long)s->ctor); | |
3636 | ||
3637 | return n + sprintf(buf + n, "\n"); | |
3638 | } | |
3639 | return 0; | |
3640 | } | |
3641 | SLAB_ATTR_RO(ctor); | |
3642 | ||
81819f0f CL |
3643 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
3644 | { | |
3645 | return sprintf(buf, "%d\n", s->refcount - 1); | |
3646 | } | |
3647 | SLAB_ATTR_RO(aliases); | |
3648 | ||
3649 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) | |
3650 | { | |
3651 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU); | |
3652 | } | |
3653 | SLAB_ATTR_RO(slabs); | |
3654 | ||
3655 | static ssize_t partial_show(struct kmem_cache *s, char *buf) | |
3656 | { | |
3657 | return slab_objects(s, buf, SO_PARTIAL); | |
3658 | } | |
3659 | SLAB_ATTR_RO(partial); | |
3660 | ||
3661 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) | |
3662 | { | |
3663 | return slab_objects(s, buf, SO_CPU); | |
3664 | } | |
3665 | SLAB_ATTR_RO(cpu_slabs); | |
3666 | ||
3667 | static ssize_t objects_show(struct kmem_cache *s, char *buf) | |
3668 | { | |
3669 | return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS); | |
3670 | } | |
3671 | SLAB_ATTR_RO(objects); | |
3672 | ||
3673 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) | |
3674 | { | |
3675 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); | |
3676 | } | |
3677 | ||
3678 | static ssize_t sanity_checks_store(struct kmem_cache *s, | |
3679 | const char *buf, size_t length) | |
3680 | { | |
3681 | s->flags &= ~SLAB_DEBUG_FREE; | |
3682 | if (buf[0] == '1') | |
3683 | s->flags |= SLAB_DEBUG_FREE; | |
3684 | return length; | |
3685 | } | |
3686 | SLAB_ATTR(sanity_checks); | |
3687 | ||
3688 | static ssize_t trace_show(struct kmem_cache *s, char *buf) | |
3689 | { | |
3690 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); | |
3691 | } | |
3692 | ||
3693 | static ssize_t trace_store(struct kmem_cache *s, const char *buf, | |
3694 | size_t length) | |
3695 | { | |
3696 | s->flags &= ~SLAB_TRACE; | |
3697 | if (buf[0] == '1') | |
3698 | s->flags |= SLAB_TRACE; | |
3699 | return length; | |
3700 | } | |
3701 | SLAB_ATTR(trace); | |
3702 | ||
3703 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) | |
3704 | { | |
3705 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); | |
3706 | } | |
3707 | ||
3708 | static ssize_t reclaim_account_store(struct kmem_cache *s, | |
3709 | const char *buf, size_t length) | |
3710 | { | |
3711 | s->flags &= ~SLAB_RECLAIM_ACCOUNT; | |
3712 | if (buf[0] == '1') | |
3713 | s->flags |= SLAB_RECLAIM_ACCOUNT; | |
3714 | return length; | |
3715 | } | |
3716 | SLAB_ATTR(reclaim_account); | |
3717 | ||
3718 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) | |
3719 | { | |
5af60839 | 3720 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); |
81819f0f CL |
3721 | } |
3722 | SLAB_ATTR_RO(hwcache_align); | |
3723 | ||
3724 | #ifdef CONFIG_ZONE_DMA | |
3725 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) | |
3726 | { | |
3727 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); | |
3728 | } | |
3729 | SLAB_ATTR_RO(cache_dma); | |
3730 | #endif | |
3731 | ||
3732 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) | |
3733 | { | |
3734 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); | |
3735 | } | |
3736 | SLAB_ATTR_RO(destroy_by_rcu); | |
3737 | ||
3738 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) | |
3739 | { | |
3740 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); | |
3741 | } | |
3742 | ||
3743 | static ssize_t red_zone_store(struct kmem_cache *s, | |
3744 | const char *buf, size_t length) | |
3745 | { | |
3746 | if (any_slab_objects(s)) | |
3747 | return -EBUSY; | |
3748 | ||
3749 | s->flags &= ~SLAB_RED_ZONE; | |
3750 | if (buf[0] == '1') | |
3751 | s->flags |= SLAB_RED_ZONE; | |
3752 | calculate_sizes(s); | |
3753 | return length; | |
3754 | } | |
3755 | SLAB_ATTR(red_zone); | |
3756 | ||
3757 | static ssize_t poison_show(struct kmem_cache *s, char *buf) | |
3758 | { | |
3759 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); | |
3760 | } | |
3761 | ||
3762 | static ssize_t poison_store(struct kmem_cache *s, | |
3763 | const char *buf, size_t length) | |
3764 | { | |
3765 | if (any_slab_objects(s)) | |
3766 | return -EBUSY; | |
3767 | ||
3768 | s->flags &= ~SLAB_POISON; | |
3769 | if (buf[0] == '1') | |
3770 | s->flags |= SLAB_POISON; | |
3771 | calculate_sizes(s); | |
3772 | return length; | |
3773 | } | |
3774 | SLAB_ATTR(poison); | |
3775 | ||
3776 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) | |
3777 | { | |
3778 | return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); | |
3779 | } | |
3780 | ||
3781 | static ssize_t store_user_store(struct kmem_cache *s, | |
3782 | const char *buf, size_t length) | |
3783 | { | |
3784 | if (any_slab_objects(s)) | |
3785 | return -EBUSY; | |
3786 | ||
3787 | s->flags &= ~SLAB_STORE_USER; | |
3788 | if (buf[0] == '1') | |
3789 | s->flags |= SLAB_STORE_USER; | |
3790 | calculate_sizes(s); | |
3791 | return length; | |
3792 | } | |
3793 | SLAB_ATTR(store_user); | |
3794 | ||
53e15af0 CL |
3795 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
3796 | { | |
3797 | return 0; | |
3798 | } | |
3799 | ||
3800 | static ssize_t validate_store(struct kmem_cache *s, | |
3801 | const char *buf, size_t length) | |
3802 | { | |
434e245d CL |
3803 | int ret = -EINVAL; |
3804 | ||
3805 | if (buf[0] == '1') { | |
3806 | ret = validate_slab_cache(s); | |
3807 | if (ret >= 0) | |
3808 | ret = length; | |
3809 | } | |
3810 | return ret; | |
53e15af0 CL |
3811 | } |
3812 | SLAB_ATTR(validate); | |
3813 | ||
2086d26a CL |
3814 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
3815 | { | |
3816 | return 0; | |
3817 | } | |
3818 | ||
3819 | static ssize_t shrink_store(struct kmem_cache *s, | |
3820 | const char *buf, size_t length) | |
3821 | { | |
3822 | if (buf[0] == '1') { | |
3823 | int rc = kmem_cache_shrink(s); | |
3824 | ||
3825 | if (rc) | |
3826 | return rc; | |
3827 | } else | |
3828 | return -EINVAL; | |
3829 | return length; | |
3830 | } | |
3831 | SLAB_ATTR(shrink); | |
3832 | ||
88a420e4 CL |
3833 | static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) |
3834 | { | |
3835 | if (!(s->flags & SLAB_STORE_USER)) | |
3836 | return -ENOSYS; | |
3837 | return list_locations(s, buf, TRACK_ALLOC); | |
3838 | } | |
3839 | SLAB_ATTR_RO(alloc_calls); | |
3840 | ||
3841 | static ssize_t free_calls_show(struct kmem_cache *s, char *buf) | |
3842 | { | |
3843 | if (!(s->flags & SLAB_STORE_USER)) | |
3844 | return -ENOSYS; | |
3845 | return list_locations(s, buf, TRACK_FREE); | |
3846 | } | |
3847 | SLAB_ATTR_RO(free_calls); | |
3848 | ||
81819f0f | 3849 | #ifdef CONFIG_NUMA |
9824601e | 3850 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) |
81819f0f | 3851 | { |
9824601e | 3852 | return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); |
81819f0f CL |
3853 | } |
3854 | ||
9824601e | 3855 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, |
81819f0f CL |
3856 | const char *buf, size_t length) |
3857 | { | |
3858 | int n = simple_strtoul(buf, NULL, 10); | |
3859 | ||
3860 | if (n < 100) | |
9824601e | 3861 | s->remote_node_defrag_ratio = n * 10; |
81819f0f CL |
3862 | return length; |
3863 | } | |
9824601e | 3864 | SLAB_ATTR(remote_node_defrag_ratio); |
81819f0f CL |
3865 | #endif |
3866 | ||
3867 | static struct attribute * slab_attrs[] = { | |
3868 | &slab_size_attr.attr, | |
3869 | &object_size_attr.attr, | |
3870 | &objs_per_slab_attr.attr, | |
3871 | &order_attr.attr, | |
3872 | &objects_attr.attr, | |
3873 | &slabs_attr.attr, | |
3874 | &partial_attr.attr, | |
3875 | &cpu_slabs_attr.attr, | |
3876 | &ctor_attr.attr, | |
81819f0f CL |
3877 | &aliases_attr.attr, |
3878 | &align_attr.attr, | |
3879 | &sanity_checks_attr.attr, | |
3880 | &trace_attr.attr, | |
3881 | &hwcache_align_attr.attr, | |
3882 | &reclaim_account_attr.attr, | |
3883 | &destroy_by_rcu_attr.attr, | |
3884 | &red_zone_attr.attr, | |
3885 | &poison_attr.attr, | |
3886 | &store_user_attr.attr, | |
53e15af0 | 3887 | &validate_attr.attr, |
2086d26a | 3888 | &shrink_attr.attr, |
88a420e4 CL |
3889 | &alloc_calls_attr.attr, |
3890 | &free_calls_attr.attr, | |
81819f0f CL |
3891 | #ifdef CONFIG_ZONE_DMA |
3892 | &cache_dma_attr.attr, | |
3893 | #endif | |
3894 | #ifdef CONFIG_NUMA | |
9824601e | 3895 | &remote_node_defrag_ratio_attr.attr, |
81819f0f CL |
3896 | #endif |
3897 | NULL | |
3898 | }; | |
3899 | ||
3900 | static struct attribute_group slab_attr_group = { | |
3901 | .attrs = slab_attrs, | |
3902 | }; | |
3903 | ||
3904 | static ssize_t slab_attr_show(struct kobject *kobj, | |
3905 | struct attribute *attr, | |
3906 | char *buf) | |
3907 | { | |
3908 | struct slab_attribute *attribute; | |
3909 | struct kmem_cache *s; | |
3910 | int err; | |
3911 | ||
3912 | attribute = to_slab_attr(attr); | |
3913 | s = to_slab(kobj); | |
3914 | ||
3915 | if (!attribute->show) | |
3916 | return -EIO; | |
3917 | ||
3918 | err = attribute->show(s, buf); | |
3919 | ||
3920 | return err; | |
3921 | } | |
3922 | ||
3923 | static ssize_t slab_attr_store(struct kobject *kobj, | |
3924 | struct attribute *attr, | |
3925 | const char *buf, size_t len) | |
3926 | { | |
3927 | struct slab_attribute *attribute; | |
3928 | struct kmem_cache *s; | |
3929 | int err; | |
3930 | ||
3931 | attribute = to_slab_attr(attr); | |
3932 | s = to_slab(kobj); | |
3933 | ||
3934 | if (!attribute->store) | |
3935 | return -EIO; | |
3936 | ||
3937 | err = attribute->store(s, buf, len); | |
3938 | ||
3939 | return err; | |
3940 | } | |
3941 | ||
151c602f CL |
3942 | static void kmem_cache_release(struct kobject *kobj) |
3943 | { | |
3944 | struct kmem_cache *s = to_slab(kobj); | |
3945 | ||
3946 | kfree(s); | |
3947 | } | |
3948 | ||
81819f0f CL |
3949 | static struct sysfs_ops slab_sysfs_ops = { |
3950 | .show = slab_attr_show, | |
3951 | .store = slab_attr_store, | |
3952 | }; | |
3953 | ||
3954 | static struct kobj_type slab_ktype = { | |
3955 | .sysfs_ops = &slab_sysfs_ops, | |
151c602f | 3956 | .release = kmem_cache_release |
81819f0f CL |
3957 | }; |
3958 | ||
3959 | static int uevent_filter(struct kset *kset, struct kobject *kobj) | |
3960 | { | |
3961 | struct kobj_type *ktype = get_ktype(kobj); | |
3962 | ||
3963 | if (ktype == &slab_ktype) | |
3964 | return 1; | |
3965 | return 0; | |
3966 | } | |
3967 | ||
3968 | static struct kset_uevent_ops slab_uevent_ops = { | |
3969 | .filter = uevent_filter, | |
3970 | }; | |
3971 | ||
27c3a314 | 3972 | static struct kset *slab_kset; |
81819f0f CL |
3973 | |
3974 | #define ID_STR_LENGTH 64 | |
3975 | ||
3976 | /* Create a unique string id for a slab cache: | |
3977 | * format | |
3978 | * :[flags-]size:[memory address of kmemcache] | |
3979 | */ | |
3980 | static char *create_unique_id(struct kmem_cache *s) | |
3981 | { | |
3982 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); | |
3983 | char *p = name; | |
3984 | ||
3985 | BUG_ON(!name); | |
3986 | ||
3987 | *p++ = ':'; | |
3988 | /* | |
3989 | * First flags affecting slabcache operations. We will only | |
3990 | * get here for aliasable slabs so we do not need to support | |
3991 | * too many flags. The flags here must cover all flags that | |
3992 | * are matched during merging to guarantee that the id is | |
3993 | * unique. | |
3994 | */ | |
3995 | if (s->flags & SLAB_CACHE_DMA) | |
3996 | *p++ = 'd'; | |
3997 | if (s->flags & SLAB_RECLAIM_ACCOUNT) | |
3998 | *p++ = 'a'; | |
3999 | if (s->flags & SLAB_DEBUG_FREE) | |
4000 | *p++ = 'F'; | |
4001 | if (p != name + 1) | |
4002 | *p++ = '-'; | |
4003 | p += sprintf(p, "%07d", s->size); | |
4004 | BUG_ON(p > name + ID_STR_LENGTH - 1); | |
4005 | return name; | |
4006 | } | |
4007 | ||
4008 | static int sysfs_slab_add(struct kmem_cache *s) | |
4009 | { | |
4010 | int err; | |
4011 | const char *name; | |
4012 | int unmergeable; | |
4013 | ||
4014 | if (slab_state < SYSFS) | |
4015 | /* Defer until later */ | |
4016 | return 0; | |
4017 | ||
4018 | unmergeable = slab_unmergeable(s); | |
4019 | if (unmergeable) { | |
4020 | /* | |
4021 | * Slabcache can never be merged so we can use the name proper. | |
4022 | * This is typically the case for debug situations. In that | |
4023 | * case we can catch duplicate names easily. | |
4024 | */ | |
27c3a314 | 4025 | sysfs_remove_link(&slab_kset->kobj, s->name); |
81819f0f CL |
4026 | name = s->name; |
4027 | } else { | |
4028 | /* | |
4029 | * Create a unique name for the slab as a target | |
4030 | * for the symlinks. | |
4031 | */ | |
4032 | name = create_unique_id(s); | |
4033 | } | |
4034 | ||
27c3a314 | 4035 | s->kobj.kset = slab_kset; |
1eada11c GKH |
4036 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); |
4037 | if (err) { | |
4038 | kobject_put(&s->kobj); | |
81819f0f | 4039 | return err; |
1eada11c | 4040 | } |
81819f0f CL |
4041 | |
4042 | err = sysfs_create_group(&s->kobj, &slab_attr_group); | |
4043 | if (err) | |
4044 | return err; | |
4045 | kobject_uevent(&s->kobj, KOBJ_ADD); | |
4046 | if (!unmergeable) { | |
4047 | /* Setup first alias */ | |
4048 | sysfs_slab_alias(s, s->name); | |
4049 | kfree(name); | |
4050 | } | |
4051 | return 0; | |
4052 | } | |
4053 | ||
4054 | static void sysfs_slab_remove(struct kmem_cache *s) | |
4055 | { | |
4056 | kobject_uevent(&s->kobj, KOBJ_REMOVE); | |
4057 | kobject_del(&s->kobj); | |
151c602f | 4058 | kobject_put(&s->kobj); |
81819f0f CL |
4059 | } |
4060 | ||
4061 | /* | |
4062 | * Need to buffer aliases during bootup until sysfs becomes | |
4063 | * available lest we loose that information. | |
4064 | */ | |
4065 | struct saved_alias { | |
4066 | struct kmem_cache *s; | |
4067 | const char *name; | |
4068 | struct saved_alias *next; | |
4069 | }; | |
4070 | ||
5af328a5 | 4071 | static struct saved_alias *alias_list; |
81819f0f CL |
4072 | |
4073 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) | |
4074 | { | |
4075 | struct saved_alias *al; | |
4076 | ||
4077 | if (slab_state == SYSFS) { | |
4078 | /* | |
4079 | * If we have a leftover link then remove it. | |
4080 | */ | |
27c3a314 GKH |
4081 | sysfs_remove_link(&slab_kset->kobj, name); |
4082 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); | |
81819f0f CL |
4083 | } |
4084 | ||
4085 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); | |
4086 | if (!al) | |
4087 | return -ENOMEM; | |
4088 | ||
4089 | al->s = s; | |
4090 | al->name = name; | |
4091 | al->next = alias_list; | |
4092 | alias_list = al; | |
4093 | return 0; | |
4094 | } | |
4095 | ||
4096 | static int __init slab_sysfs_init(void) | |
4097 | { | |
5b95a4ac | 4098 | struct kmem_cache *s; |
81819f0f CL |
4099 | int err; |
4100 | ||
0ff21e46 | 4101 | slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); |
27c3a314 | 4102 | if (!slab_kset) { |
81819f0f CL |
4103 | printk(KERN_ERR "Cannot register slab subsystem.\n"); |
4104 | return -ENOSYS; | |
4105 | } | |
4106 | ||
26a7bd03 CL |
4107 | slab_state = SYSFS; |
4108 | ||
5b95a4ac | 4109 | list_for_each_entry(s, &slab_caches, list) { |
26a7bd03 | 4110 | err = sysfs_slab_add(s); |
5d540fb7 CL |
4111 | if (err) |
4112 | printk(KERN_ERR "SLUB: Unable to add boot slab %s" | |
4113 | " to sysfs\n", s->name); | |
26a7bd03 | 4114 | } |
81819f0f CL |
4115 | |
4116 | while (alias_list) { | |
4117 | struct saved_alias *al = alias_list; | |
4118 | ||
4119 | alias_list = alias_list->next; | |
4120 | err = sysfs_slab_alias(al->s, al->name); | |
5d540fb7 CL |
4121 | if (err) |
4122 | printk(KERN_ERR "SLUB: Unable to add boot slab alias" | |
4123 | " %s to sysfs\n", s->name); | |
81819f0f CL |
4124 | kfree(al); |
4125 | } | |
4126 | ||
4127 | resiliency_test(); | |
4128 | return 0; | |
4129 | } | |
4130 | ||
4131 | __initcall(slab_sysfs_init); | |
81819f0f | 4132 | #endif |
57ed3eda PE |
4133 | |
4134 | /* | |
4135 | * The /proc/slabinfo ABI | |
4136 | */ | |
158a9624 LT |
4137 | #ifdef CONFIG_SLABINFO |
4138 | ||
4139 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, | |
4140 | size_t count, loff_t *ppos) | |
4141 | { | |
4142 | return -EINVAL; | |
4143 | } | |
4144 | ||
57ed3eda PE |
4145 | |
4146 | static void print_slabinfo_header(struct seq_file *m) | |
4147 | { | |
4148 | seq_puts(m, "slabinfo - version: 2.1\n"); | |
4149 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " | |
4150 | "<objperslab> <pagesperslab>"); | |
4151 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4152 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
4153 | seq_putc(m, '\n'); | |
4154 | } | |
4155 | ||
4156 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4157 | { | |
4158 | loff_t n = *pos; | |
4159 | ||
4160 | down_read(&slub_lock); | |
4161 | if (!n) | |
4162 | print_slabinfo_header(m); | |
4163 | ||
4164 | return seq_list_start(&slab_caches, *pos); | |
4165 | } | |
4166 | ||
4167 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4168 | { | |
4169 | return seq_list_next(p, &slab_caches, pos); | |
4170 | } | |
4171 | ||
4172 | static void s_stop(struct seq_file *m, void *p) | |
4173 | { | |
4174 | up_read(&slub_lock); | |
4175 | } | |
4176 | ||
4177 | static int s_show(struct seq_file *m, void *p) | |
4178 | { | |
4179 | unsigned long nr_partials = 0; | |
4180 | unsigned long nr_slabs = 0; | |
4181 | unsigned long nr_inuse = 0; | |
4182 | unsigned long nr_objs; | |
4183 | struct kmem_cache *s; | |
4184 | int node; | |
4185 | ||
4186 | s = list_entry(p, struct kmem_cache, list); | |
4187 | ||
4188 | for_each_online_node(node) { | |
4189 | struct kmem_cache_node *n = get_node(s, node); | |
4190 | ||
4191 | if (!n) | |
4192 | continue; | |
4193 | ||
4194 | nr_partials += n->nr_partial; | |
4195 | nr_slabs += atomic_long_read(&n->nr_slabs); | |
4196 | nr_inuse += count_partial(n); | |
4197 | } | |
4198 | ||
4199 | nr_objs = nr_slabs * s->objects; | |
4200 | nr_inuse += (nr_slabs - nr_partials) * s->objects; | |
4201 | ||
4202 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, | |
4203 | nr_objs, s->size, s->objects, (1 << s->order)); | |
4204 | seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); | |
4205 | seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, | |
4206 | 0UL); | |
4207 | seq_putc(m, '\n'); | |
4208 | return 0; | |
4209 | } | |
4210 | ||
4211 | const struct seq_operations slabinfo_op = { | |
4212 | .start = s_start, | |
4213 | .next = s_next, | |
4214 | .stop = s_stop, | |
4215 | .show = s_show, | |
4216 | }; | |
4217 | ||
158a9624 | 4218 | #endif /* CONFIG_SLABINFO */ |