Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 1 | /* |
| 2 | * SLOB Allocator: Simple List Of Blocks |
| 3 | * |
| 4 | * Matt Mackall <mpm@selenic.com> 12/30/03 |
| 5 | * |
| 6 | * How SLOB works: |
| 7 | * |
| 8 | * The core of SLOB is a traditional K&R style heap allocator, with |
| 9 | * support for returning aligned objects. The granularity of this |
| 10 | * allocator is 8 bytes on x86, though it's perhaps possible to reduce |
| 11 | * this to 4 if it's deemed worth the effort. The slob heap is a |
| 12 | * singly-linked list of pages from __get_free_page, grown on demand |
| 13 | * and allocation from the heap is currently first-fit. |
| 14 | * |
| 15 | * Above this is an implementation of kmalloc/kfree. Blocks returned |
| 16 | * from kmalloc are 8-byte aligned and prepended with a 8-byte header. |
| 17 | * If kmalloc is asked for objects of PAGE_SIZE or larger, it calls |
| 18 | * __get_free_pages directly so that it can return page-aligned blocks |
| 19 | * and keeps a linked list of such pages and their orders. These |
| 20 | * objects are detected in kfree() by their page alignment. |
| 21 | * |
| 22 | * SLAB is emulated on top of SLOB by simply calling constructors and |
| 23 | * destructors for every SLAB allocation. Objects are returned with |
| 24 | * the 8-byte alignment unless the SLAB_MUST_HWCACHE_ALIGN flag is |
| 25 | * set, in which case the low-level allocator will fragment blocks to |
| 26 | * create the proper alignment. Again, objects of page-size or greater |
| 27 | * are allocated by calling __get_free_pages. As SLAB objects know |
| 28 | * their size, no separate size bookkeeping is necessary and there is |
| 29 | * essentially no allocation space overhead. |
| 30 | */ |
| 31 | |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 32 | #include <linux/slab.h> |
| 33 | #include <linux/mm.h> |
| 34 | #include <linux/cache.h> |
| 35 | #include <linux/init.h> |
| 36 | #include <linux/module.h> |
| 37 | #include <linux/timer.h> |
| 38 | |
| 39 | struct slob_block { |
| 40 | int units; |
| 41 | struct slob_block *next; |
| 42 | }; |
| 43 | typedef struct slob_block slob_t; |
| 44 | |
| 45 | #define SLOB_UNIT sizeof(slob_t) |
| 46 | #define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT) |
| 47 | #define SLOB_ALIGN L1_CACHE_BYTES |
| 48 | |
| 49 | struct bigblock { |
| 50 | int order; |
| 51 | void *pages; |
| 52 | struct bigblock *next; |
| 53 | }; |
| 54 | typedef struct bigblock bigblock_t; |
| 55 | |
| 56 | static slob_t arena = { .next = &arena, .units = 1 }; |
| 57 | static slob_t *slobfree = &arena; |
| 58 | static bigblock_t *bigblocks; |
| 59 | static DEFINE_SPINLOCK(slob_lock); |
| 60 | static DEFINE_SPINLOCK(block_lock); |
| 61 | |
| 62 | static void slob_free(void *b, int size); |
Dimitri Gorokhovik | bcb4ddb | 2006-12-29 16:48:28 -0800 | [diff] [blame] | 63 | static void slob_timer_cbk(void); |
| 64 | |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 65 | |
| 66 | static void *slob_alloc(size_t size, gfp_t gfp, int align) |
| 67 | { |
| 68 | slob_t *prev, *cur, *aligned = 0; |
| 69 | int delta = 0, units = SLOB_UNITS(size); |
| 70 | unsigned long flags; |
| 71 | |
| 72 | spin_lock_irqsave(&slob_lock, flags); |
| 73 | prev = slobfree; |
| 74 | for (cur = prev->next; ; prev = cur, cur = cur->next) { |
| 75 | if (align) { |
| 76 | aligned = (slob_t *)ALIGN((unsigned long)cur, align); |
| 77 | delta = aligned - cur; |
| 78 | } |
| 79 | if (cur->units >= units + delta) { /* room enough? */ |
| 80 | if (delta) { /* need to fragment head to align? */ |
| 81 | aligned->units = cur->units - delta; |
| 82 | aligned->next = cur->next; |
| 83 | cur->next = aligned; |
| 84 | cur->units = delta; |
| 85 | prev = cur; |
| 86 | cur = aligned; |
| 87 | } |
| 88 | |
| 89 | if (cur->units == units) /* exact fit? */ |
| 90 | prev->next = cur->next; /* unlink */ |
| 91 | else { /* fragment */ |
| 92 | prev->next = cur + units; |
| 93 | prev->next->units = cur->units - units; |
| 94 | prev->next->next = cur->next; |
| 95 | cur->units = units; |
| 96 | } |
| 97 | |
| 98 | slobfree = prev; |
| 99 | spin_unlock_irqrestore(&slob_lock, flags); |
| 100 | return cur; |
| 101 | } |
| 102 | if (cur == slobfree) { |
| 103 | spin_unlock_irqrestore(&slob_lock, flags); |
| 104 | |
| 105 | if (size == PAGE_SIZE) /* trying to shrink arena? */ |
| 106 | return 0; |
| 107 | |
| 108 | cur = (slob_t *)__get_free_page(gfp); |
| 109 | if (!cur) |
| 110 | return 0; |
| 111 | |
| 112 | slob_free(cur, PAGE_SIZE); |
| 113 | spin_lock_irqsave(&slob_lock, flags); |
| 114 | cur = slobfree; |
| 115 | } |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | static void slob_free(void *block, int size) |
| 120 | { |
| 121 | slob_t *cur, *b = (slob_t *)block; |
| 122 | unsigned long flags; |
| 123 | |
| 124 | if (!block) |
| 125 | return; |
| 126 | |
| 127 | if (size) |
| 128 | b->units = SLOB_UNITS(size); |
| 129 | |
| 130 | /* Find reinsertion point */ |
| 131 | spin_lock_irqsave(&slob_lock, flags); |
| 132 | for (cur = slobfree; !(b > cur && b < cur->next); cur = cur->next) |
| 133 | if (cur >= cur->next && (b > cur || b < cur->next)) |
| 134 | break; |
| 135 | |
| 136 | if (b + b->units == cur->next) { |
| 137 | b->units += cur->next->units; |
| 138 | b->next = cur->next->next; |
| 139 | } else |
| 140 | b->next = cur->next; |
| 141 | |
| 142 | if (cur + cur->units == b) { |
| 143 | cur->units += b->units; |
| 144 | cur->next = b->next; |
| 145 | } else |
| 146 | cur->next = b; |
| 147 | |
| 148 | slobfree = cur; |
| 149 | |
| 150 | spin_unlock_irqrestore(&slob_lock, flags); |
| 151 | } |
| 152 | |
| 153 | static int FASTCALL(find_order(int size)); |
| 154 | static int fastcall find_order(int size) |
| 155 | { |
| 156 | int order = 0; |
| 157 | for ( ; size > 4096 ; size >>=1) |
| 158 | order++; |
| 159 | return order; |
| 160 | } |
| 161 | |
Christoph Lameter | 2e892f4 | 2006-12-13 00:34:23 -0800 | [diff] [blame] | 162 | void *__kmalloc(size_t size, gfp_t gfp) |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 163 | { |
| 164 | slob_t *m; |
| 165 | bigblock_t *bb; |
| 166 | unsigned long flags; |
| 167 | |
| 168 | if (size < PAGE_SIZE - SLOB_UNIT) { |
| 169 | m = slob_alloc(size + SLOB_UNIT, gfp, 0); |
| 170 | return m ? (void *)(m + 1) : 0; |
| 171 | } |
| 172 | |
| 173 | bb = slob_alloc(sizeof(bigblock_t), gfp, 0); |
| 174 | if (!bb) |
| 175 | return 0; |
| 176 | |
| 177 | bb->order = find_order(size); |
| 178 | bb->pages = (void *)__get_free_pages(gfp, bb->order); |
| 179 | |
| 180 | if (bb->pages) { |
| 181 | spin_lock_irqsave(&block_lock, flags); |
| 182 | bb->next = bigblocks; |
| 183 | bigblocks = bb; |
| 184 | spin_unlock_irqrestore(&block_lock, flags); |
| 185 | return bb->pages; |
| 186 | } |
| 187 | |
| 188 | slob_free(bb, sizeof(bigblock_t)); |
| 189 | return 0; |
| 190 | } |
Christoph Lameter | 2e892f4 | 2006-12-13 00:34:23 -0800 | [diff] [blame] | 191 | EXPORT_SYMBOL(__kmalloc); |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 192 | |
Pekka Enberg | fd76bab | 2007-05-06 14:48:40 -0700 | [diff] [blame] | 193 | /** |
| 194 | * krealloc - reallocate memory. The contents will remain unchanged. |
| 195 | * |
| 196 | * @p: object to reallocate memory for. |
| 197 | * @new_size: how many bytes of memory are required. |
| 198 | * @flags: the type of memory to allocate. |
| 199 | * |
| 200 | * The contents of the object pointed to are preserved up to the |
| 201 | * lesser of the new and old sizes. If @p is %NULL, krealloc() |
| 202 | * behaves exactly like kmalloc(). If @size is 0 and @p is not a |
| 203 | * %NULL pointer, the object pointed to is freed. |
| 204 | */ |
| 205 | void *krealloc(const void *p, size_t new_size, gfp_t flags) |
| 206 | { |
| 207 | void *ret; |
| 208 | |
| 209 | if (unlikely(!p)) |
| 210 | return kmalloc_track_caller(new_size, flags); |
| 211 | |
| 212 | if (unlikely(!new_size)) { |
| 213 | kfree(p); |
| 214 | return NULL; |
| 215 | } |
| 216 | |
| 217 | ret = kmalloc_track_caller(new_size, flags); |
| 218 | if (ret) { |
| 219 | memcpy(ret, p, min(new_size, ksize(p))); |
| 220 | kfree(p); |
| 221 | } |
| 222 | return ret; |
| 223 | } |
| 224 | EXPORT_SYMBOL(krealloc); |
| 225 | |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 226 | void kfree(const void *block) |
| 227 | { |
| 228 | bigblock_t *bb, **last = &bigblocks; |
| 229 | unsigned long flags; |
| 230 | |
| 231 | if (!block) |
| 232 | return; |
| 233 | |
| 234 | if (!((unsigned long)block & (PAGE_SIZE-1))) { |
| 235 | /* might be on the big block list */ |
| 236 | spin_lock_irqsave(&block_lock, flags); |
| 237 | for (bb = bigblocks; bb; last = &bb->next, bb = bb->next) { |
| 238 | if (bb->pages == block) { |
| 239 | *last = bb->next; |
| 240 | spin_unlock_irqrestore(&block_lock, flags); |
| 241 | free_pages((unsigned long)block, bb->order); |
| 242 | slob_free(bb, sizeof(bigblock_t)); |
| 243 | return; |
| 244 | } |
| 245 | } |
| 246 | spin_unlock_irqrestore(&block_lock, flags); |
| 247 | } |
| 248 | |
| 249 | slob_free((slob_t *)block - 1, 0); |
| 250 | return; |
| 251 | } |
| 252 | |
| 253 | EXPORT_SYMBOL(kfree); |
| 254 | |
Pekka Enberg | fd76bab | 2007-05-06 14:48:40 -0700 | [diff] [blame] | 255 | size_t ksize(const void *block) |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 256 | { |
| 257 | bigblock_t *bb; |
| 258 | unsigned long flags; |
| 259 | |
| 260 | if (!block) |
| 261 | return 0; |
| 262 | |
| 263 | if (!((unsigned long)block & (PAGE_SIZE-1))) { |
| 264 | spin_lock_irqsave(&block_lock, flags); |
| 265 | for (bb = bigblocks; bb; bb = bb->next) |
| 266 | if (bb->pages == block) { |
| 267 | spin_unlock_irqrestore(&slob_lock, flags); |
| 268 | return PAGE_SIZE << bb->order; |
| 269 | } |
| 270 | spin_unlock_irqrestore(&block_lock, flags); |
| 271 | } |
| 272 | |
| 273 | return ((slob_t *)block - 1)->units * SLOB_UNIT; |
| 274 | } |
| 275 | |
| 276 | struct kmem_cache { |
| 277 | unsigned int size, align; |
| 278 | const char *name; |
| 279 | void (*ctor)(void *, struct kmem_cache *, unsigned long); |
| 280 | void (*dtor)(void *, struct kmem_cache *, unsigned long); |
| 281 | }; |
| 282 | |
| 283 | struct kmem_cache *kmem_cache_create(const char *name, size_t size, |
| 284 | size_t align, unsigned long flags, |
| 285 | void (*ctor)(void*, struct kmem_cache *, unsigned long), |
| 286 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
| 287 | { |
| 288 | struct kmem_cache *c; |
| 289 | |
| 290 | c = slob_alloc(sizeof(struct kmem_cache), flags, 0); |
| 291 | |
| 292 | if (c) { |
| 293 | c->name = name; |
| 294 | c->size = size; |
| 295 | c->ctor = ctor; |
| 296 | c->dtor = dtor; |
| 297 | /* ignore alignment unless it's forced */ |
| 298 | c->align = (flags & SLAB_MUST_HWCACHE_ALIGN) ? SLOB_ALIGN : 0; |
| 299 | if (c->align < align) |
| 300 | c->align = align; |
Akinobu Mita | bc0055a | 2007-05-06 14:49:52 -0700 | [diff] [blame^] | 301 | } else if (flags & SLAB_PANIC) |
| 302 | panic("Cannot create slab cache %s\n", name); |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 303 | |
| 304 | return c; |
| 305 | } |
| 306 | EXPORT_SYMBOL(kmem_cache_create); |
| 307 | |
Alexey Dobriyan | 133d205 | 2006-09-27 01:49:41 -0700 | [diff] [blame] | 308 | void kmem_cache_destroy(struct kmem_cache *c) |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 309 | { |
| 310 | slob_free(c, sizeof(struct kmem_cache)); |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 311 | } |
| 312 | EXPORT_SYMBOL(kmem_cache_destroy); |
| 313 | |
| 314 | void *kmem_cache_alloc(struct kmem_cache *c, gfp_t flags) |
| 315 | { |
| 316 | void *b; |
| 317 | |
| 318 | if (c->size < PAGE_SIZE) |
| 319 | b = slob_alloc(c->size, flags, c->align); |
| 320 | else |
| 321 | b = (void *)__get_free_pages(flags, find_order(c->size)); |
| 322 | |
| 323 | if (c->ctor) |
| 324 | c->ctor(b, c, SLAB_CTOR_CONSTRUCTOR); |
| 325 | |
| 326 | return b; |
| 327 | } |
| 328 | EXPORT_SYMBOL(kmem_cache_alloc); |
| 329 | |
Pekka Enberg | a8c0f9a | 2006-03-25 03:06:42 -0800 | [diff] [blame] | 330 | void *kmem_cache_zalloc(struct kmem_cache *c, gfp_t flags) |
| 331 | { |
| 332 | void *ret = kmem_cache_alloc(c, flags); |
| 333 | if (ret) |
| 334 | memset(ret, 0, c->size); |
| 335 | |
| 336 | return ret; |
| 337 | } |
| 338 | EXPORT_SYMBOL(kmem_cache_zalloc); |
| 339 | |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 340 | void kmem_cache_free(struct kmem_cache *c, void *b) |
| 341 | { |
| 342 | if (c->dtor) |
| 343 | c->dtor(b, c, 0); |
| 344 | |
| 345 | if (c->size < PAGE_SIZE) |
| 346 | slob_free(b, c->size); |
| 347 | else |
| 348 | free_pages((unsigned long)b, find_order(c->size)); |
| 349 | } |
| 350 | EXPORT_SYMBOL(kmem_cache_free); |
| 351 | |
| 352 | unsigned int kmem_cache_size(struct kmem_cache *c) |
| 353 | { |
| 354 | return c->size; |
| 355 | } |
| 356 | EXPORT_SYMBOL(kmem_cache_size); |
| 357 | |
| 358 | const char *kmem_cache_name(struct kmem_cache *c) |
| 359 | { |
| 360 | return c->name; |
| 361 | } |
| 362 | EXPORT_SYMBOL(kmem_cache_name); |
| 363 | |
| 364 | static struct timer_list slob_timer = TIMER_INITIALIZER( |
Dimitri Gorokhovik | bcb4ddb | 2006-12-29 16:48:28 -0800 | [diff] [blame] | 365 | (void (*)(unsigned long))slob_timer_cbk, 0, 0); |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 366 | |
Christoph Lameter | 2e892f4 | 2006-12-13 00:34:23 -0800 | [diff] [blame] | 367 | int kmem_cache_shrink(struct kmem_cache *d) |
| 368 | { |
| 369 | return 0; |
| 370 | } |
| 371 | EXPORT_SYMBOL(kmem_cache_shrink); |
| 372 | |
Christoph Lameter | 55935a3 | 2006-12-13 00:34:24 -0800 | [diff] [blame] | 373 | int kmem_ptr_validate(struct kmem_cache *a, const void *b) |
Christoph Lameter | 2e892f4 | 2006-12-13 00:34:23 -0800 | [diff] [blame] | 374 | { |
| 375 | return 0; |
| 376 | } |
| 377 | |
Dimitri Gorokhovik | bcb4ddb | 2006-12-29 16:48:28 -0800 | [diff] [blame] | 378 | void __init kmem_cache_init(void) |
| 379 | { |
| 380 | slob_timer_cbk(); |
| 381 | } |
| 382 | |
| 383 | static void slob_timer_cbk(void) |
Matt Mackall | 10cef60 | 2006-01-08 01:01:45 -0800 | [diff] [blame] | 384 | { |
| 385 | void *p = slob_alloc(PAGE_SIZE, 0, PAGE_SIZE-1); |
| 386 | |
| 387 | if (p) |
| 388 | free_page((unsigned long)p); |
| 389 | |
| 390 | mod_timer(&slob_timer, jiffies + HZ); |
| 391 | } |