1 /* 2 * linux/mm/slab.c 3 * Written by Mark Hemment, 1996/97. 4 * (markhe@nextd.demon.co.uk) 5 * 6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 7 * 8 * Major cleanup, different bufctl logic, per-cpu arrays 9 * (c) 2000 Manfred Spraul 10 * 11 * Cleanup, make the head arrays unconditional, preparation for NUMA 12 * (c) 2002 Manfred Spraul 13 * 14 * An implementation of the Slab Allocator as described in outline in; 15 * UNIX Internals: The New Frontiers by Uresh Vahalia 16 * Pub: Prentice Hall ISBN 0-13-101908-2 17 * or with a little more detail in; 18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 19 * Jeff Bonwick (Sun Microsystems). 20 * Presented at: USENIX Summer 1994 Technical Conference 21 * 22 * The memory is organized in caches, one cache for each object type. 23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 24 * Each cache consists out of many slabs (they are small (usually one 25 * page long) and always contiguous), and each slab contains multiple 26 * initialized objects. 27 * 28 * This means, that your constructor is used only for newly allocated 29 * slabs and you must pass objects with the same intializations to 30 * kmem_cache_free. 31 * 32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 33 * normal). If you need a special memory type, then must create a new 34 * cache for that memory type. 35 * 36 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 37 * full slabs with 0 free objects 38 * partial slabs 39 * empty slabs with no allocated objects 40 * 41 * If partial slabs exist, then new allocations come from these slabs, 42 * otherwise from empty slabs or new slabs are allocated. 43 * 44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 46 * 47 * Each cache has a short per-cpu head array, most allocs 48 * and frees go into that array, and if that array overflows, then 1/2 49 * of the entries in the array are given back into the global cache. 50 * The head array is strictly LIFO and should improve the cache hit rates. 51 * On SMP, it additionally reduces the spinlock operations. 52 * 53 * The c_cpuarray may not be read with enabled local interrupts - 54 * it's changed with a smp_call_function(). 55 * 56 * SMP synchronization: 57 * constructors and destructors are called without any locking. 58 * Several members in struct kmem_cache and struct slab never change, they 59 * are accessed without any locking. 60 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 61 * and local interrupts are disabled so slab code is preempt-safe. 62 * The non-constant members are protected with a per-cache irq spinlock. 63 * 64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 65 * in 2000 - many ideas in the current implementation are derived from 66 * his patch. 67 * 68 * Further notes from the original documentation: 69 * 70 * 11 April '97. Started multi-threading - markhe 71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'. 72 * The sem is only needed when accessing/extending the cache-chain, which 73 * can never happen inside an interrupt (kmem_cache_create(), 74 * kmem_cache_shrink() and kmem_cache_reap()). 75 * 76 * At present, each engine can be growing a cache. This should be blocked. 77 * 78 * 15 March 2005. NUMA slab allocator. 79 * Shai Fultheim <shai@scalex86.org>. 80 * Shobhit Dayal <shobhit@calsoftinc.com> 81 * Alok N Kataria <alokk@calsoftinc.com> 82 * Christoph Lameter <christoph@lameter.com> 83 * 84 * Modified the slab allocator to be node aware on NUMA systems. 85 * Each node has its own list of partial, free and full slabs. 86 * All object allocations for a node occur from node specific slab lists. 87 */ 88 89 #include <linux/config.h> 90 #include <linux/slab.h> 91 #include <linux/mm.h> 92 #include <linux/swap.h> 93 #include <linux/cache.h> 94 #include <linux/interrupt.h> 95 #include <linux/init.h> 96 #include <linux/compiler.h> 97 #include <linux/cpuset.h> 98 #include <linux/seq_file.h> 99 #include <linux/notifier.h> 100 #include <linux/kallsyms.h> 101 #include <linux/cpu.h> 102 #include <linux/sysctl.h> 103 #include <linux/module.h> 104 #include <linux/rcupdate.h> 105 #include <linux/string.h> 106 #include <linux/nodemask.h> 107 #include <linux/mempolicy.h> 108 #include <linux/mutex.h> 109 110 #include <asm/uaccess.h> 111 #include <asm/cacheflush.h> 112 #include <asm/tlbflush.h> 113 #include <asm/page.h> 114 115 /* 116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, 117 * SLAB_RED_ZONE & SLAB_POISON. 118 * 0 for faster, smaller code (especially in the critical paths). 119 * 120 * STATS - 1 to collect stats for /proc/slabinfo. 121 * 0 for faster, smaller code (especially in the critical paths). 122 * 123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 124 */ 125 126 #ifdef CONFIG_DEBUG_SLAB 127 #define DEBUG 1 128 #define STATS 1 129 #define FORCED_DEBUG 1 130 #else 131 #define DEBUG 0 132 #define STATS 0 133 #define FORCED_DEBUG 0 134 #endif 135 136 /* Shouldn't this be in a header file somewhere? */ 137 #define BYTES_PER_WORD sizeof(void *) 138 139 #ifndef cache_line_size 140 #define cache_line_size() L1_CACHE_BYTES 141 #endif 142 143 #ifndef ARCH_KMALLOC_MINALIGN 144 /* 145 * Enforce a minimum alignment for the kmalloc caches. 146 * Usually, the kmalloc caches are cache_line_size() aligned, except when 147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. 148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed 149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. 150 * Note that this flag disables some debug features. 151 */ 152 #define ARCH_KMALLOC_MINALIGN 0 153 #endif 154 155 #ifndef ARCH_SLAB_MINALIGN 156 /* 157 * Enforce a minimum alignment for all caches. 158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD 159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. 160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables 161 * some debug features. 162 */ 163 #define ARCH_SLAB_MINALIGN 0 164 #endif 165 166 #ifndef ARCH_KMALLOC_FLAGS 167 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 168 #endif 169 170 /* Legal flag mask for kmem_cache_create(). */ 171 #if DEBUG 172 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ 173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 174 SLAB_CACHE_DMA | \ 175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ 176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 178 #else 179 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ 181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 183 #endif 184 185 /* 186 * kmem_bufctl_t: 187 * 188 * Bufctl's are used for linking objs within a slab 189 * linked offsets. 190 * 191 * This implementation relies on "struct page" for locating the cache & 192 * slab an object belongs to. 193 * This allows the bufctl structure to be small (one int), but limits 194 * the number of objects a slab (not a cache) can contain when off-slab 195 * bufctls are used. The limit is the size of the largest general cache 196 * that does not use off-slab slabs. 197 * For 32bit archs with 4 kB pages, is this 56. 198 * This is not serious, as it is only for large objects, when it is unwise 199 * to have too many per slab. 200 * Note: This limit can be raised by introducing a general cache whose size 201 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 202 */ 203 204 typedef unsigned int kmem_bufctl_t; 205 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 206 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 207 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 208 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 209 210 /* 211 * struct slab 212 * 213 * Manages the objs in a slab. Placed either at the beginning of mem allocated 214 * for a slab, or allocated from an general cache. 215 * Slabs are chained into three list: fully used, partial, fully free slabs. 216 */ 217 struct slab { 218 struct list_head list; 219 unsigned long colouroff; 220 void *s_mem; /* including colour offset */ 221 unsigned int inuse; /* num of objs active in slab */ 222 kmem_bufctl_t free; 223 unsigned short nodeid; 224 }; 225 226 /* 227 * struct slab_rcu 228 * 229 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 230 * arrange for kmem_freepages to be called via RCU. This is useful if 231 * we need to approach a kernel structure obliquely, from its address 232 * obtained without the usual locking. We can lock the structure to 233 * stabilize it and check it's still at the given address, only if we 234 * can be sure that the memory has not been meanwhile reused for some 235 * other kind of object (which our subsystem's lock might corrupt). 236 * 237 * rcu_read_lock before reading the address, then rcu_read_unlock after 238 * taking the spinlock within the structure expected at that address. 239 * 240 * We assume struct slab_rcu can overlay struct slab when destroying. 241 */ 242 struct slab_rcu { 243 struct rcu_head head; 244 struct kmem_cache *cachep; 245 void *addr; 246 }; 247 248 /* 249 * struct array_cache 250 * 251 * Purpose: 252 * - LIFO ordering, to hand out cache-warm objects from _alloc 253 * - reduce the number of linked list operations 254 * - reduce spinlock operations 255 * 256 * The limit is stored in the per-cpu structure to reduce the data cache 257 * footprint. 258 * 259 */ 260 struct array_cache { 261 unsigned int avail; 262 unsigned int limit; 263 unsigned int batchcount; 264 unsigned int touched; 265 spinlock_t lock; 266 void *entry[0]; /* 267 * Must have this definition in here for the proper 268 * alignment of array_cache. Also simplifies accessing 269 * the entries. 270 * [0] is for gcc 2.95. It should really be []. 271 */ 272 }; 273 274 /* 275 * bootstrap: The caches do not work without cpuarrays anymore, but the 276 * cpuarrays are allocated from the generic caches... 277 */ 278 #define BOOT_CPUCACHE_ENTRIES 1 279 struct arraycache_init { 280 struct array_cache cache; 281 void *entries[BOOT_CPUCACHE_ENTRIES]; 282 }; 283 284 /* 285 * The slab lists for all objects. 286 */ 287 struct kmem_list3 { 288 struct list_head slabs_partial; /* partial list first, better asm code */ 289 struct list_head slabs_full; 290 struct list_head slabs_free; 291 unsigned long free_objects; 292 unsigned int free_limit; 293 unsigned int colour_next; /* Per-node cache coloring */ 294 spinlock_t list_lock; 295 struct array_cache *shared; /* shared per node */ 296 struct array_cache **alien; /* on other nodes */ 297 unsigned long next_reap; /* updated without locking */ 298 int free_touched; /* updated without locking */ 299 }; 300 301 /* 302 * Need this for bootstrapping a per node allocator. 303 */ 304 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) 305 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 306 #define CACHE_CACHE 0 307 #define SIZE_AC 1 308 #define SIZE_L3 (1 + MAX_NUMNODES) 309 310 /* 311 * This function must be completely optimized away if a constant is passed to 312 * it. Mostly the same as what is in linux/slab.h except it returns an index. 313 */ 314 static __always_inline int index_of(const size_t size) 315 { 316 extern void __bad_size(void); 317 318 if (__builtin_constant_p(size)) { 319 int i = 0; 320 321 #define CACHE(x) \ 322 if (size <=x) \ 323 return i; \ 324 else \ 325 i++; 326 #include "linux/kmalloc_sizes.h" 327 #undef CACHE 328 __bad_size(); 329 } else 330 __bad_size(); 331 return 0; 332 } 333 334 #define INDEX_AC index_of(sizeof(struct arraycache_init)) 335 #define INDEX_L3 index_of(sizeof(struct kmem_list3)) 336 337 static void kmem_list3_init(struct kmem_list3 *parent) 338 { 339 INIT_LIST_HEAD(&parent->slabs_full); 340 INIT_LIST_HEAD(&parent->slabs_partial); 341 INIT_LIST_HEAD(&parent->slabs_free); 342 parent->shared = NULL; 343 parent->alien = NULL; 344 parent->colour_next = 0; 345 spin_lock_init(&parent->list_lock); 346 parent->free_objects = 0; 347 parent->free_touched = 0; 348 } 349 350 #define MAKE_LIST(cachep, listp, slab, nodeid) \ 351 do { \ 352 INIT_LIST_HEAD(listp); \ 353 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 354 } while (0) 355 356 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 357 do { \ 358 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 359 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 360 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 361 } while (0) 362 363 /* 364 * struct kmem_cache 365 * 366 * manages a cache. 367 */ 368 369 struct kmem_cache { 370 /* 1) per-cpu data, touched during every alloc/free */ 371 struct array_cache *array[NR_CPUS]; 372 /* 2) Cache tunables. Protected by cache_chain_mutex */ 373 unsigned int batchcount; 374 unsigned int limit; 375 unsigned int shared; 376 377 unsigned int buffer_size; 378 /* 3) touched by every alloc & free from the backend */ 379 struct kmem_list3 *nodelists[MAX_NUMNODES]; 380 381 unsigned int flags; /* constant flags */ 382 unsigned int num; /* # of objs per slab */ 383 384 /* 4) cache_grow/shrink */ 385 /* order of pgs per slab (2^n) */ 386 unsigned int gfporder; 387 388 /* force GFP flags, e.g. GFP_DMA */ 389 gfp_t gfpflags; 390 391 size_t colour; /* cache colouring range */ 392 unsigned int colour_off; /* colour offset */ 393 struct kmem_cache *slabp_cache; 394 unsigned int slab_size; 395 unsigned int dflags; /* dynamic flags */ 396 397 /* constructor func */ 398 void (*ctor) (void *, struct kmem_cache *, unsigned long); 399 400 /* de-constructor func */ 401 void (*dtor) (void *, struct kmem_cache *, unsigned long); 402 403 /* 5) cache creation/removal */ 404 const char *name; 405 struct list_head next; 406 407 /* 6) statistics */ 408 #if STATS 409 unsigned long num_active; 410 unsigned long num_allocations; 411 unsigned long high_mark; 412 unsigned long grown; 413 unsigned long reaped; 414 unsigned long errors; 415 unsigned long max_freeable; 416 unsigned long node_allocs; 417 unsigned long node_frees; 418 unsigned long node_overflow; 419 atomic_t allochit; 420 atomic_t allocmiss; 421 atomic_t freehit; 422 atomic_t freemiss; 423 #endif 424 #if DEBUG 425 /* 426 * If debugging is enabled, then the allocator can add additional 427 * fields and/or padding to every object. buffer_size contains the total 428 * object size including these internal fields, the following two 429 * variables contain the offset to the user object and its size. 430 */ 431 int obj_offset; 432 int obj_size; 433 #endif 434 }; 435 436 #define CFLGS_OFF_SLAB (0x80000000UL) 437 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 438 439 #define BATCHREFILL_LIMIT 16 440 /* 441 * Optimization question: fewer reaps means less probability for unnessary 442 * cpucache drain/refill cycles. 443 * 444 * OTOH the cpuarrays can contain lots of objects, 445 * which could lock up otherwise freeable slabs. 446 */ 447 #define REAPTIMEOUT_CPUC (2*HZ) 448 #define REAPTIMEOUT_LIST3 (4*HZ) 449 450 #if STATS 451 #define STATS_INC_ACTIVE(x) ((x)->num_active++) 452 #define STATS_DEC_ACTIVE(x) ((x)->num_active--) 453 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 454 #define STATS_INC_GROWN(x) ((x)->grown++) 455 #define STATS_INC_REAPED(x) ((x)->reaped++) 456 #define STATS_SET_HIGH(x) \ 457 do { \ 458 if ((x)->num_active > (x)->high_mark) \ 459 (x)->high_mark = (x)->num_active; \ 460 } while (0) 461 #define STATS_INC_ERR(x) ((x)->errors++) 462 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 463 #define STATS_INC_NODEFREES(x) ((x)->node_frees++) 464 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 465 #define STATS_SET_FREEABLE(x, i) \ 466 do { \ 467 if ((x)->max_freeable < i) \ 468 (x)->max_freeable = i; \ 469 } while (0) 470 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 471 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 472 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 473 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 474 #else 475 #define STATS_INC_ACTIVE(x) do { } while (0) 476 #define STATS_DEC_ACTIVE(x) do { } while (0) 477 #define STATS_INC_ALLOCED(x) do { } while (0) 478 #define STATS_INC_GROWN(x) do { } while (0) 479 #define STATS_INC_REAPED(x) do { } while (0) 480 #define STATS_SET_HIGH(x) do { } while (0) 481 #define STATS_INC_ERR(x) do { } while (0) 482 #define STATS_INC_NODEALLOCS(x) do { } while (0) 483 #define STATS_INC_NODEFREES(x) do { } while (0) 484 #define STATS_INC_ACOVERFLOW(x) do { } while (0) 485 #define STATS_SET_FREEABLE(x, i) do { } while (0) 486 #define STATS_INC_ALLOCHIT(x) do { } while (0) 487 #define STATS_INC_ALLOCMISS(x) do { } while (0) 488 #define STATS_INC_FREEHIT(x) do { } while (0) 489 #define STATS_INC_FREEMISS(x) do { } while (0) 490 #endif 491 492 #if DEBUG 493 /* 494 * Magic nums for obj red zoning. 495 * Placed in the first word before and the first word after an obj. 496 */ 497 #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ 498 #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ 499 500 /* ...and for poisoning */ 501 #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ 502 #define POISON_FREE 0x6b /* for use-after-free poisoning */ 503 #define POISON_END 0xa5 /* end-byte of poisoning */ 504 505 /* 506 * memory layout of objects: 507 * 0 : objp 508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 509 * the end of an object is aligned with the end of the real 510 * allocation. Catches writes behind the end of the allocation. 511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 512 * redzone word. 513 * cachep->obj_offset: The real object. 514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 516 * [BYTES_PER_WORD long] 517 */ 518 static int obj_offset(struct kmem_cache *cachep) 519 { 520 return cachep->obj_offset; 521 } 522 523 static int obj_size(struct kmem_cache *cachep) 524 { 525 return cachep->obj_size; 526 } 527 528 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 529 { 530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); 532 } 533 534 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 535 { 536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 537 if (cachep->flags & SLAB_STORE_USER) 538 return (unsigned long *)(objp + cachep->buffer_size - 539 2 * BYTES_PER_WORD); 540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); 541 } 542 543 static void **dbg_userword(struct kmem_cache *cachep, void *objp) 544 { 545 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 547 } 548 549 #else 550 551 #define obj_offset(x) 0 552 #define obj_size(cachep) (cachep->buffer_size) 553 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) 554 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) 555 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 556 557 #endif 558 559 /* 560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp 561 * order. 562 */ 563 #if defined(CONFIG_LARGE_ALLOCS) 564 #define MAX_OBJ_ORDER 13 /* up to 32Mb */ 565 #define MAX_GFP_ORDER 13 /* up to 32Mb */ 566 #elif defined(CONFIG_MMU) 567 #define MAX_OBJ_ORDER 5 /* 32 pages */ 568 #define MAX_GFP_ORDER 5 /* 32 pages */ 569 #else 570 #define MAX_OBJ_ORDER 8 /* up to 1Mb */ 571 #define MAX_GFP_ORDER 8 /* up to 1Mb */ 572 #endif 573 574 /* 575 * Do not go above this order unless 0 objects fit into the slab. 576 */ 577 #define BREAK_GFP_ORDER_HI 1 578 #define BREAK_GFP_ORDER_LO 0 579 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 580 581 /* 582 * Functions for storing/retrieving the cachep and or slab from the page 583 * allocator. These are used to find the slab an obj belongs to. With kfree(), 584 * these are used to find the cache which an obj belongs to. 585 */ 586 static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 587 { 588 page->lru.next = (struct list_head *)cache; 589 } 590 591 static inline struct kmem_cache *page_get_cache(struct page *page) 592 { 593 if (unlikely(PageCompound(page))) 594 page = (struct page *)page_private(page); 595 return (struct kmem_cache *)page->lru.next; 596 } 597 598 static inline void page_set_slab(struct page *page, struct slab *slab) 599 { 600 page->lru.prev = (struct list_head *)slab; 601 } 602 603 static inline struct slab *page_get_slab(struct page *page) 604 { 605 if (unlikely(PageCompound(page))) 606 page = (struct page *)page_private(page); 607 return (struct slab *)page->lru.prev; 608 } 609 610 static inline struct kmem_cache *virt_to_cache(const void *obj) 611 { 612 struct page *page = virt_to_page(obj); 613 return page_get_cache(page); 614 } 615 616 static inline struct slab *virt_to_slab(const void *obj) 617 { 618 struct page *page = virt_to_page(obj); 619 return page_get_slab(page); 620 } 621 622 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 623 unsigned int idx) 624 { 625 return slab->s_mem + cache->buffer_size * idx; 626 } 627 628 static inline unsigned int obj_to_index(struct kmem_cache *cache, 629 struct slab *slab, void *obj) 630 { 631 return (unsigned)(obj - slab->s_mem) / cache->buffer_size; 632 } 633 634 /* 635 * These are the default caches for kmalloc. Custom caches can have other sizes. 636 */ 637 struct cache_sizes malloc_sizes[] = { 638 #define CACHE(x) { .cs_size = (x) }, 639 #include <linux/kmalloc_sizes.h> 640 CACHE(ULONG_MAX) 641 #undef CACHE 642 }; 643 EXPORT_SYMBOL(malloc_sizes); 644 645 /* Must match cache_sizes above. Out of line to keep cache footprint low. */ 646 struct cache_names { 647 char *name; 648 char *name_dma; 649 }; 650 651 static struct cache_names __initdata cache_names[] = { 652 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 653 #include <linux/kmalloc_sizes.h> 654 {NULL,} 655 #undef CACHE 656 }; 657 658 static struct arraycache_init initarray_cache __initdata = 659 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 660 static struct arraycache_init initarray_generic = 661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 662 663 /* internal cache of cache description objs */ 664 static struct kmem_cache cache_cache = { 665 .batchcount = 1, 666 .limit = BOOT_CPUCACHE_ENTRIES, 667 .shared = 1, 668 .buffer_size = sizeof(struct kmem_cache), 669 .name = "kmem_cache", 670 #if DEBUG 671 .obj_size = sizeof(struct kmem_cache), 672 #endif 673 }; 674 675 /* Guard access to the cache-chain. */ 676 static DEFINE_MUTEX(cache_chain_mutex); 677 static struct list_head cache_chain; 678 679 /* 680 * vm_enough_memory() looks at this to determine how many slab-allocated pages 681 * are possibly freeable under pressure 682 * 683 * SLAB_RECLAIM_ACCOUNT turns this on per-slab 684 */ 685 atomic_t slab_reclaim_pages; 686 687 /* 688 * chicken and egg problem: delay the per-cpu array allocation 689 * until the general caches are up. 690 */ 691 static enum { 692 NONE, 693 PARTIAL_AC, 694 PARTIAL_L3, 695 FULL 696 } g_cpucache_up; 697 698 /* 699 * used by boot code to determine if it can use slab based allocator 700 */ 701 int slab_is_available(void) 702 { 703 return g_cpucache_up == FULL; 704 } 705 706 static DEFINE_PER_CPU(struct work_struct, reap_work); 707 708 static void free_block(struct kmem_cache *cachep, void **objpp, int len, 709 int node); 710 static void enable_cpucache(struct kmem_cache *cachep); 711 static void cache_reap(void *unused); 712 static int __node_shrink(struct kmem_cache *cachep, int node); 713 714 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 715 { 716 return cachep->array[smp_processor_id()]; 717 } 718 719 static inline struct kmem_cache *__find_general_cachep(size_t size, 720 gfp_t gfpflags) 721 { 722 struct cache_sizes *csizep = malloc_sizes; 723 724 #if DEBUG 725 /* This happens if someone tries to call 726 * kmem_cache_create(), or __kmalloc(), before 727 * the generic caches are initialized. 728 */ 729 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 730 #endif 731 while (size > csizep->cs_size) 732 csizep++; 733 734 /* 735 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 736 * has cs_{dma,}cachep==NULL. Thus no special case 737 * for large kmalloc calls required. 738 */ 739 if (unlikely(gfpflags & GFP_DMA)) 740 return csizep->cs_dmacachep; 741 return csizep->cs_cachep; 742 } 743 744 struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 745 { 746 return __find_general_cachep(size, gfpflags); 747 } 748 EXPORT_SYMBOL(kmem_find_general_cachep); 749 750 static size_t slab_mgmt_size(size_t nr_objs, size_t align) 751 { 752 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 753 } 754 755 /* 756 * Calculate the number of objects and left-over bytes for a given buffer size. 757 */ 758 static void cache_estimate(unsigned long gfporder, size_t buffer_size, 759 size_t align, int flags, size_t *left_over, 760 unsigned int *num) 761 { 762 int nr_objs; 763 size_t mgmt_size; 764 size_t slab_size = PAGE_SIZE << gfporder; 765 766 /* 767 * The slab management structure can be either off the slab or 768 * on it. For the latter case, the memory allocated for a 769 * slab is used for: 770 * 771 * - The struct slab 772 * - One kmem_bufctl_t for each object 773 * - Padding to respect alignment of @align 774 * - @buffer_size bytes for each object 775 * 776 * If the slab management structure is off the slab, then the 777 * alignment will already be calculated into the size. Because 778 * the slabs are all pages aligned, the objects will be at the 779 * correct alignment when allocated. 780 */ 781 if (flags & CFLGS_OFF_SLAB) { 782 mgmt_size = 0; 783 nr_objs = slab_size / buffer_size; 784 785 if (nr_objs > SLAB_LIMIT) 786 nr_objs = SLAB_LIMIT; 787 } else { 788 /* 789 * Ignore padding for the initial guess. The padding 790 * is at most @align-1 bytes, and @buffer_size is at 791 * least @align. In the worst case, this result will 792 * be one greater than the number of objects that fit 793 * into the memory allocation when taking the padding 794 * into account. 795 */ 796 nr_objs = (slab_size - sizeof(struct slab)) / 797 (buffer_size + sizeof(kmem_bufctl_t)); 798 799 /* 800 * This calculated number will be either the right 801 * amount, or one greater than what we want. 802 */ 803 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 804 > slab_size) 805 nr_objs--; 806 807 if (nr_objs > SLAB_LIMIT) 808 nr_objs = SLAB_LIMIT; 809 810 mgmt_size = slab_mgmt_size(nr_objs, align); 811 } 812 *num = nr_objs; 813 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 814 } 815 816 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) 817 818 static void __slab_error(const char *function, struct kmem_cache *cachep, 819 char *msg) 820 { 821 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 822 function, cachep->name, msg); 823 dump_stack(); 824 } 825 826 #ifdef CONFIG_NUMA 827 /* 828 * Special reaping functions for NUMA systems called from cache_reap(). 829 * These take care of doing round robin flushing of alien caches (containing 830 * objects freed on different nodes from which they were allocated) and the 831 * flushing of remote pcps by calling drain_node_pages. 832 */ 833 static DEFINE_PER_CPU(unsigned long, reap_node); 834 835 static void init_reap_node(int cpu) 836 { 837 int node; 838 839 node = next_node(cpu_to_node(cpu), node_online_map); 840 if (node == MAX_NUMNODES) 841 node = first_node(node_online_map); 842 843 __get_cpu_var(reap_node) = node; 844 } 845 846 static void next_reap_node(void) 847 { 848 int node = __get_cpu_var(reap_node); 849 850 /* 851 * Also drain per cpu pages on remote zones 852 */ 853 if (node != numa_node_id()) 854 drain_node_pages(node); 855 856 node = next_node(node, node_online_map); 857 if (unlikely(node >= MAX_NUMNODES)) 858 node = first_node(node_online_map); 859 __get_cpu_var(reap_node) = node; 860 } 861 862 #else 863 #define init_reap_node(cpu) do { } while (0) 864 #define next_reap_node(void) do { } while (0) 865 #endif 866 867 /* 868 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 869 * via the workqueue/eventd. 870 * Add the CPU number into the expiration time to minimize the possibility of 871 * the CPUs getting into lockstep and contending for the global cache chain 872 * lock. 873 */ 874 static void __devinit start_cpu_timer(int cpu) 875 { 876 struct work_struct *reap_work = &per_cpu(reap_work, cpu); 877 878 /* 879 * When this gets called from do_initcalls via cpucache_init(), 880 * init_workqueues() has already run, so keventd will be setup 881 * at that time. 882 */ 883 if (keventd_up() && reap_work->func == NULL) { 884 init_reap_node(cpu); 885 INIT_WORK(reap_work, cache_reap, NULL); 886 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); 887 } 888 } 889 890 static struct array_cache *alloc_arraycache(int node, int entries, 891 int batchcount) 892 { 893 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 894 struct array_cache *nc = NULL; 895 896 nc = kmalloc_node(memsize, GFP_KERNEL, node); 897 if (nc) { 898 nc->avail = 0; 899 nc->limit = entries; 900 nc->batchcount = batchcount; 901 nc->touched = 0; 902 spin_lock_init(&nc->lock); 903 } 904 return nc; 905 } 906 907 /* 908 * Transfer objects in one arraycache to another. 909 * Locking must be handled by the caller. 910 * 911 * Return the number of entries transferred. 912 */ 913 static int transfer_objects(struct array_cache *to, 914 struct array_cache *from, unsigned int max) 915 { 916 /* Figure out how many entries to transfer */ 917 int nr = min(min(from->avail, max), to->limit - to->avail); 918 919 if (!nr) 920 return 0; 921 922 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 923 sizeof(void *) *nr); 924 925 from->avail -= nr; 926 to->avail += nr; 927 to->touched = 1; 928 return nr; 929 } 930 931 #ifdef CONFIG_NUMA 932 static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); 933 static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 934 935 static struct array_cache **alloc_alien_cache(int node, int limit) 936 { 937 struct array_cache **ac_ptr; 938 int memsize = sizeof(void *) * MAX_NUMNODES; 939 int i; 940 941 if (limit > 1) 942 limit = 12; 943 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); 944 if (ac_ptr) { 945 for_each_node(i) { 946 if (i == node || !node_online(i)) { 947 ac_ptr[i] = NULL; 948 continue; 949 } 950 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); 951 if (!ac_ptr[i]) { 952 for (i--; i <= 0; i--) 953 kfree(ac_ptr[i]); 954 kfree(ac_ptr); 955 return NULL; 956 } 957 } 958 } 959 return ac_ptr; 960 } 961 962 static void free_alien_cache(struct array_cache **ac_ptr) 963 { 964 int i; 965 966 if (!ac_ptr) 967 return; 968 for_each_node(i) 969 kfree(ac_ptr[i]); 970 kfree(ac_ptr); 971 } 972 973 static void __drain_alien_cache(struct kmem_cache *cachep, 974 struct array_cache *ac, int node) 975 { 976 struct kmem_list3 *rl3 = cachep->nodelists[node]; 977 978 if (ac->avail) { 979 spin_lock(&rl3->list_lock); 980 /* 981 * Stuff objects into the remote nodes shared array first. 982 * That way we could avoid the overhead of putting the objects 983 * into the free lists and getting them back later. 984 */ 985 if (rl3->shared) 986 transfer_objects(rl3->shared, ac, ac->limit); 987 988 free_block(cachep, ac->entry, ac->avail, node); 989 ac->avail = 0; 990 spin_unlock(&rl3->list_lock); 991 } 992 } 993 994 /* 995 * Called from cache_reap() to regularly drain alien caches round robin. 996 */ 997 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 998 { 999 int node = __get_cpu_var(reap_node); 1000 1001 if (l3->alien) { 1002 struct array_cache *ac = l3->alien[node]; 1003 1004 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1005 __drain_alien_cache(cachep, ac, node); 1006 spin_unlock_irq(&ac->lock); 1007 } 1008 } 1009 } 1010 1011 static void drain_alien_cache(struct kmem_cache *cachep, 1012 struct array_cache **alien) 1013 { 1014 int i = 0; 1015 struct array_cache *ac; 1016 unsigned long flags; 1017 1018 for_each_online_node(i) { 1019 ac = alien[i]; 1020 if (ac) { 1021 spin_lock_irqsave(&ac->lock, flags); 1022 __drain_alien_cache(cachep, ac, i); 1023 spin_unlock_irqrestore(&ac->lock, flags); 1024 } 1025 } 1026 } 1027 #else 1028 1029 #define drain_alien_cache(cachep, alien) do { } while (0) 1030 #define reap_alien(cachep, l3) do { } while (0) 1031 1032 static inline struct array_cache **alloc_alien_cache(int node, int limit) 1033 { 1034 return (struct array_cache **) 0x01020304ul; 1035 } 1036 1037 static inline void free_alien_cache(struct array_cache **ac_ptr) 1038 { 1039 } 1040 1041 #endif 1042 1043 static int cpuup_callback(struct notifier_block *nfb, 1044 unsigned long action, void *hcpu) 1045 { 1046 long cpu = (long)hcpu; 1047 struct kmem_cache *cachep; 1048 struct kmem_list3 *l3 = NULL; 1049 int node = cpu_to_node(cpu); 1050 int memsize = sizeof(struct kmem_list3); 1051 1052 switch (action) { 1053 case CPU_UP_PREPARE: 1054 mutex_lock(&cache_chain_mutex); 1055 /* 1056 * We need to do this right in the beginning since 1057 * alloc_arraycache's are going to use this list. 1058 * kmalloc_node allows us to add the slab to the right 1059 * kmem_list3 and not this cpu's kmem_list3 1060 */ 1061 1062 list_for_each_entry(cachep, &cache_chain, next) { 1063 /* 1064 * Set up the size64 kmemlist for cpu before we can 1065 * begin anything. Make sure some other cpu on this 1066 * node has not already allocated this 1067 */ 1068 if (!cachep->nodelists[node]) { 1069 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1070 if (!l3) 1071 goto bad; 1072 kmem_list3_init(l3); 1073 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1074 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1075 1076 /* 1077 * The l3s don't come and go as CPUs come and 1078 * go. cache_chain_mutex is sufficient 1079 * protection here. 1080 */ 1081 cachep->nodelists[node] = l3; 1082 } 1083 1084 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1085 cachep->nodelists[node]->free_limit = 1086 (1 + nr_cpus_node(node)) * 1087 cachep->batchcount + cachep->num; 1088 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1089 } 1090 1091 /* 1092 * Now we can go ahead with allocating the shared arrays and 1093 * array caches 1094 */ 1095 list_for_each_entry(cachep, &cache_chain, next) { 1096 struct array_cache *nc; 1097 struct array_cache *shared; 1098 struct array_cache **alien; 1099 1100 nc = alloc_arraycache(node, cachep->limit, 1101 cachep->batchcount); 1102 if (!nc) 1103 goto bad; 1104 shared = alloc_arraycache(node, 1105 cachep->shared * cachep->batchcount, 1106 0xbaadf00d); 1107 if (!shared) 1108 goto bad; 1109 1110 alien = alloc_alien_cache(node, cachep->limit); 1111 if (!alien) 1112 goto bad; 1113 cachep->array[cpu] = nc; 1114 l3 = cachep->nodelists[node]; 1115 BUG_ON(!l3); 1116 1117 spin_lock_irq(&l3->list_lock); 1118 if (!l3->shared) { 1119 /* 1120 * We are serialised from CPU_DEAD or 1121 * CPU_UP_CANCELLED by the cpucontrol lock 1122 */ 1123 l3->shared = shared; 1124 shared = NULL; 1125 } 1126 #ifdef CONFIG_NUMA 1127 if (!l3->alien) { 1128 l3->alien = alien; 1129 alien = NULL; 1130 } 1131 #endif 1132 spin_unlock_irq(&l3->list_lock); 1133 kfree(shared); 1134 free_alien_cache(alien); 1135 } 1136 mutex_unlock(&cache_chain_mutex); 1137 break; 1138 case CPU_ONLINE: 1139 start_cpu_timer(cpu); 1140 break; 1141 #ifdef CONFIG_HOTPLUG_CPU 1142 case CPU_DEAD: 1143 /* 1144 * Even if all the cpus of a node are down, we don't free the 1145 * kmem_list3 of any cache. This to avoid a race between 1146 * cpu_down, and a kmalloc allocation from another cpu for 1147 * memory from the node of the cpu going down. The list3 1148 * structure is usually allocated from kmem_cache_create() and 1149 * gets destroyed at kmem_cache_destroy(). 1150 */ 1151 /* fall thru */ 1152 case CPU_UP_CANCELED: 1153 mutex_lock(&cache_chain_mutex); 1154 list_for_each_entry(cachep, &cache_chain, next) { 1155 struct array_cache *nc; 1156 struct array_cache *shared; 1157 struct array_cache **alien; 1158 cpumask_t mask; 1159 1160 mask = node_to_cpumask(node); 1161 /* cpu is dead; no one can alloc from it. */ 1162 nc = cachep->array[cpu]; 1163 cachep->array[cpu] = NULL; 1164 l3 = cachep->nodelists[node]; 1165 1166 if (!l3) 1167 goto free_array_cache; 1168 1169 spin_lock_irq(&l3->list_lock); 1170 1171 /* Free limit for this kmem_list3 */ 1172 l3->free_limit -= cachep->batchcount; 1173 if (nc) 1174 free_block(cachep, nc->entry, nc->avail, node); 1175 1176 if (!cpus_empty(mask)) { 1177 spin_unlock_irq(&l3->list_lock); 1178 goto free_array_cache; 1179 } 1180 1181 shared = l3->shared; 1182 if (shared) { 1183 free_block(cachep, l3->shared->entry, 1184 l3->shared->avail, node); 1185 l3->shared = NULL; 1186 } 1187 1188 alien = l3->alien; 1189 l3->alien = NULL; 1190 1191 spin_unlock_irq(&l3->list_lock); 1192 1193 kfree(shared); 1194 if (alien) { 1195 drain_alien_cache(cachep, alien); 1196 free_alien_cache(alien); 1197 } 1198 free_array_cache: 1199 kfree(nc); 1200 } 1201 /* 1202 * In the previous loop, all the objects were freed to 1203 * the respective cache's slabs, now we can go ahead and 1204 * shrink each nodelist to its limit. 1205 */ 1206 list_for_each_entry(cachep, &cache_chain, next) { 1207 l3 = cachep->nodelists[node]; 1208 if (!l3) 1209 continue; 1210 spin_lock_irq(&l3->list_lock); 1211 /* free slabs belonging to this node */ 1212 __node_shrink(cachep, node); 1213 spin_unlock_irq(&l3->list_lock); 1214 } 1215 mutex_unlock(&cache_chain_mutex); 1216 break; 1217 #endif 1218 } 1219 return NOTIFY_OK; 1220 bad: 1221 mutex_unlock(&cache_chain_mutex); 1222 return NOTIFY_BAD; 1223 } 1224 1225 static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; 1226 1227 /* 1228 * swap the static kmem_list3 with kmalloced memory 1229 */ 1230 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1231 int nodeid) 1232 { 1233 struct kmem_list3 *ptr; 1234 1235 BUG_ON(cachep->nodelists[nodeid] != list); 1236 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); 1237 BUG_ON(!ptr); 1238 1239 local_irq_disable(); 1240 memcpy(ptr, list, sizeof(struct kmem_list3)); 1241 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1242 cachep->nodelists[nodeid] = ptr; 1243 local_irq_enable(); 1244 } 1245 1246 /* 1247 * Initialisation. Called after the page allocator have been initialised and 1248 * before smp_init(). 1249 */ 1250 void __init kmem_cache_init(void) 1251 { 1252 size_t left_over; 1253 struct cache_sizes *sizes; 1254 struct cache_names *names; 1255 int i; 1256 int order; 1257 1258 for (i = 0; i < NUM_INIT_LISTS; i++) { 1259 kmem_list3_init(&initkmem_list3[i]); 1260 if (i < MAX_NUMNODES) 1261 cache_cache.nodelists[i] = NULL; 1262 } 1263 1264 /* 1265 * Fragmentation resistance on low memory - only use bigger 1266 * page orders on machines with more than 32MB of memory. 1267 */ 1268 if (num_physpages > (32 << 20) >> PAGE_SHIFT) 1269 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1270 1271 /* Bootstrap is tricky, because several objects are allocated 1272 * from caches that do not exist yet: 1273 * 1) initialize the cache_cache cache: it contains the struct 1274 * kmem_cache structures of all caches, except cache_cache itself: 1275 * cache_cache is statically allocated. 1276 * Initially an __init data area is used for the head array and the 1277 * kmem_list3 structures, it's replaced with a kmalloc allocated 1278 * array at the end of the bootstrap. 1279 * 2) Create the first kmalloc cache. 1280 * The struct kmem_cache for the new cache is allocated normally. 1281 * An __init data area is used for the head array. 1282 * 3) Create the remaining kmalloc caches, with minimally sized 1283 * head arrays. 1284 * 4) Replace the __init data head arrays for cache_cache and the first 1285 * kmalloc cache with kmalloc allocated arrays. 1286 * 5) Replace the __init data for kmem_list3 for cache_cache and 1287 * the other cache's with kmalloc allocated memory. 1288 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1289 */ 1290 1291 /* 1) create the cache_cache */ 1292 INIT_LIST_HEAD(&cache_chain); 1293 list_add(&cache_cache.next, &cache_chain); 1294 cache_cache.colour_off = cache_line_size(); 1295 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1296 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; 1297 1298 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1299 cache_line_size()); 1300 1301 for (order = 0; order < MAX_ORDER; order++) { 1302 cache_estimate(order, cache_cache.buffer_size, 1303 cache_line_size(), 0, &left_over, &cache_cache.num); 1304 if (cache_cache.num) 1305 break; 1306 } 1307 BUG_ON(!cache_cache.num); 1308 cache_cache.gfporder = order; 1309 cache_cache.colour = left_over / cache_cache.colour_off; 1310 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1311 sizeof(struct slab), cache_line_size()); 1312 1313 /* 2+3) create the kmalloc caches */ 1314 sizes = malloc_sizes; 1315 names = cache_names; 1316 1317 /* 1318 * Initialize the caches that provide memory for the array cache and the 1319 * kmem_list3 structures first. Without this, further allocations will 1320 * bug. 1321 */ 1322 1323 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1324 sizes[INDEX_AC].cs_size, 1325 ARCH_KMALLOC_MINALIGN, 1326 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1327 NULL, NULL); 1328 1329 if (INDEX_AC != INDEX_L3) { 1330 sizes[INDEX_L3].cs_cachep = 1331 kmem_cache_create(names[INDEX_L3].name, 1332 sizes[INDEX_L3].cs_size, 1333 ARCH_KMALLOC_MINALIGN, 1334 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1335 NULL, NULL); 1336 } 1337 1338 while (sizes->cs_size != ULONG_MAX) { 1339 /* 1340 * For performance, all the general caches are L1 aligned. 1341 * This should be particularly beneficial on SMP boxes, as it 1342 * eliminates "false sharing". 1343 * Note for systems short on memory removing the alignment will 1344 * allow tighter packing of the smaller caches. 1345 */ 1346 if (!sizes->cs_cachep) { 1347 sizes->cs_cachep = kmem_cache_create(names->name, 1348 sizes->cs_size, 1349 ARCH_KMALLOC_MINALIGN, 1350 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1351 NULL, NULL); 1352 } 1353 1354 sizes->cs_dmacachep = kmem_cache_create(names->name_dma, 1355 sizes->cs_size, 1356 ARCH_KMALLOC_MINALIGN, 1357 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1358 SLAB_PANIC, 1359 NULL, NULL); 1360 sizes++; 1361 names++; 1362 } 1363 /* 4) Replace the bootstrap head arrays */ 1364 { 1365 void *ptr; 1366 1367 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1368 1369 local_irq_disable(); 1370 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1371 memcpy(ptr, cpu_cache_get(&cache_cache), 1372 sizeof(struct arraycache_init)); 1373 cache_cache.array[smp_processor_id()] = ptr; 1374 local_irq_enable(); 1375 1376 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1377 1378 local_irq_disable(); 1379 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1380 != &initarray_generic.cache); 1381 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1382 sizeof(struct arraycache_init)); 1383 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1384 ptr; 1385 local_irq_enable(); 1386 } 1387 /* 5) Replace the bootstrap kmem_list3's */ 1388 { 1389 int node; 1390 /* Replace the static kmem_list3 structures for the boot cpu */ 1391 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], 1392 numa_node_id()); 1393 1394 for_each_online_node(node) { 1395 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1396 &initkmem_list3[SIZE_AC + node], node); 1397 1398 if (INDEX_AC != INDEX_L3) { 1399 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1400 &initkmem_list3[SIZE_L3 + node], 1401 node); 1402 } 1403 } 1404 } 1405 1406 /* 6) resize the head arrays to their final sizes */ 1407 { 1408 struct kmem_cache *cachep; 1409 mutex_lock(&cache_chain_mutex); 1410 list_for_each_entry(cachep, &cache_chain, next) 1411 enable_cpucache(cachep); 1412 mutex_unlock(&cache_chain_mutex); 1413 } 1414 1415 /* Done! */ 1416 g_cpucache_up = FULL; 1417 1418 /* 1419 * Register a cpu startup notifier callback that initializes 1420 * cpu_cache_get for all new cpus 1421 */ 1422 register_cpu_notifier(&cpucache_notifier); 1423 1424 /* 1425 * The reap timers are started later, with a module init call: That part 1426 * of the kernel is not yet operational. 1427 */ 1428 } 1429 1430 static int __init cpucache_init(void) 1431 { 1432 int cpu; 1433 1434 /* 1435 * Register the timers that return unneeded pages to the page allocator 1436 */ 1437 for_each_online_cpu(cpu) 1438 start_cpu_timer(cpu); 1439 return 0; 1440 } 1441 __initcall(cpucache_init); 1442 1443 /* 1444 * Interface to system's page allocator. No need to hold the cache-lock. 1445 * 1446 * If we requested dmaable memory, we will get it. Even if we 1447 * did not request dmaable memory, we might get it, but that 1448 * would be relatively rare and ignorable. 1449 */ 1450 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1451 { 1452 struct page *page; 1453 void *addr; 1454 int i; 1455 1456 flags |= cachep->gfpflags; 1457 #ifndef CONFIG_MMU 1458 /* nommu uses slab's for process anonymous memory allocations, so 1459 * requires __GFP_COMP to properly refcount higher order allocations" 1460 */ 1461 page = alloc_pages_node(nodeid, (flags | __GFP_COMP), cachep->gfporder); 1462 #else 1463 page = alloc_pages_node(nodeid, flags, cachep->gfporder); 1464 #endif 1465 if (!page) 1466 return NULL; 1467 addr = page_address(page); 1468 1469 i = (1 << cachep->gfporder); 1470 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1471 atomic_add(i, &slab_reclaim_pages); 1472 add_page_state(nr_slab, i); 1473 while (i--) { 1474 __SetPageSlab(page); 1475 page++; 1476 } 1477 return addr; 1478 } 1479 1480 /* 1481 * Interface to system's page release. 1482 */ 1483 static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1484 { 1485 unsigned long i = (1 << cachep->gfporder); 1486 struct page *page = virt_to_page(addr); 1487 const unsigned long nr_freed = i; 1488 1489 while (i--) { 1490 BUG_ON(!PageSlab(page)); 1491 __ClearPageSlab(page); 1492 page++; 1493 } 1494 sub_page_state(nr_slab, nr_freed); 1495 if (current->reclaim_state) 1496 current->reclaim_state->reclaimed_slab += nr_freed; 1497 free_pages((unsigned long)addr, cachep->gfporder); 1498 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1499 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); 1500 } 1501 1502 static void kmem_rcu_free(struct rcu_head *head) 1503 { 1504 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1505 struct kmem_cache *cachep = slab_rcu->cachep; 1506 1507 kmem_freepages(cachep, slab_rcu->addr); 1508 if (OFF_SLAB(cachep)) 1509 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1510 } 1511 1512 #if DEBUG 1513 1514 #ifdef CONFIG_DEBUG_PAGEALLOC 1515 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1516 unsigned long caller) 1517 { 1518 int size = obj_size(cachep); 1519 1520 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1521 1522 if (size < 5 * sizeof(unsigned long)) 1523 return; 1524 1525 *addr++ = 0x12345678; 1526 *addr++ = caller; 1527 *addr++ = smp_processor_id(); 1528 size -= 3 * sizeof(unsigned long); 1529 { 1530 unsigned long *sptr = &caller; 1531 unsigned long svalue; 1532 1533 while (!kstack_end(sptr)) { 1534 svalue = *sptr++; 1535 if (kernel_text_address(svalue)) { 1536 *addr++ = svalue; 1537 size -= sizeof(unsigned long); 1538 if (size <= sizeof(unsigned long)) 1539 break; 1540 } 1541 } 1542 1543 } 1544 *addr++ = 0x87654321; 1545 } 1546 #endif 1547 1548 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1549 { 1550 int size = obj_size(cachep); 1551 addr = &((char *)addr)[obj_offset(cachep)]; 1552 1553 memset(addr, val, size); 1554 *(unsigned char *)(addr + size - 1) = POISON_END; 1555 } 1556 1557 static void dump_line(char *data, int offset, int limit) 1558 { 1559 int i; 1560 printk(KERN_ERR "%03x:", offset); 1561 for (i = 0; i < limit; i++) 1562 printk(" %02x", (unsigned char)data[offset + i]); 1563 printk("\n"); 1564 } 1565 #endif 1566 1567 #if DEBUG 1568 1569 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1570 { 1571 int i, size; 1572 char *realobj; 1573 1574 if (cachep->flags & SLAB_RED_ZONE) { 1575 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", 1576 *dbg_redzone1(cachep, objp), 1577 *dbg_redzone2(cachep, objp)); 1578 } 1579 1580 if (cachep->flags & SLAB_STORE_USER) { 1581 printk(KERN_ERR "Last user: [<%p>]", 1582 *dbg_userword(cachep, objp)); 1583 print_symbol("(%s)", 1584 (unsigned long)*dbg_userword(cachep, objp)); 1585 printk("\n"); 1586 } 1587 realobj = (char *)objp + obj_offset(cachep); 1588 size = obj_size(cachep); 1589 for (i = 0; i < size && lines; i += 16, lines--) { 1590 int limit; 1591 limit = 16; 1592 if (i + limit > size) 1593 limit = size - i; 1594 dump_line(realobj, i, limit); 1595 } 1596 } 1597 1598 static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1599 { 1600 char *realobj; 1601 int size, i; 1602 int lines = 0; 1603 1604 realobj = (char *)objp + obj_offset(cachep); 1605 size = obj_size(cachep); 1606 1607 for (i = 0; i < size; i++) { 1608 char exp = POISON_FREE; 1609 if (i == size - 1) 1610 exp = POISON_END; 1611 if (realobj[i] != exp) { 1612 int limit; 1613 /* Mismatch ! */ 1614 /* Print header */ 1615 if (lines == 0) { 1616 printk(KERN_ERR 1617 "Slab corruption: start=%p, len=%d\n", 1618 realobj, size); 1619 print_objinfo(cachep, objp, 0); 1620 } 1621 /* Hexdump the affected line */ 1622 i = (i / 16) * 16; 1623 limit = 16; 1624 if (i + limit > size) 1625 limit = size - i; 1626 dump_line(realobj, i, limit); 1627 i += 16; 1628 lines++; 1629 /* Limit to 5 lines */ 1630 if (lines > 5) 1631 break; 1632 } 1633 } 1634 if (lines != 0) { 1635 /* Print some data about the neighboring objects, if they 1636 * exist: 1637 */ 1638 struct slab *slabp = virt_to_slab(objp); 1639 unsigned int objnr; 1640 1641 objnr = obj_to_index(cachep, slabp, objp); 1642 if (objnr) { 1643 objp = index_to_obj(cachep, slabp, objnr - 1); 1644 realobj = (char *)objp + obj_offset(cachep); 1645 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1646 realobj, size); 1647 print_objinfo(cachep, objp, 2); 1648 } 1649 if (objnr + 1 < cachep->num) { 1650 objp = index_to_obj(cachep, slabp, objnr + 1); 1651 realobj = (char *)objp + obj_offset(cachep); 1652 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1653 realobj, size); 1654 print_objinfo(cachep, objp, 2); 1655 } 1656 } 1657 } 1658 #endif 1659 1660 #if DEBUG 1661 /** 1662 * slab_destroy_objs - destroy a slab and its objects 1663 * @cachep: cache pointer being destroyed 1664 * @slabp: slab pointer being destroyed 1665 * 1666 * Call the registered destructor for each object in a slab that is being 1667 * destroyed. 1668 */ 1669 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1670 { 1671 int i; 1672 for (i = 0; i < cachep->num; i++) { 1673 void *objp = index_to_obj(cachep, slabp, i); 1674 1675 if (cachep->flags & SLAB_POISON) { 1676 #ifdef CONFIG_DEBUG_PAGEALLOC 1677 if (cachep->buffer_size % PAGE_SIZE == 0 && 1678 OFF_SLAB(cachep)) 1679 kernel_map_pages(virt_to_page(objp), 1680 cachep->buffer_size / PAGE_SIZE, 1); 1681 else 1682 check_poison_obj(cachep, objp); 1683 #else 1684 check_poison_obj(cachep, objp); 1685 #endif 1686 } 1687 if (cachep->flags & SLAB_RED_ZONE) { 1688 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1689 slab_error(cachep, "start of a freed object " 1690 "was overwritten"); 1691 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1692 slab_error(cachep, "end of a freed object " 1693 "was overwritten"); 1694 } 1695 if (cachep->dtor && !(cachep->flags & SLAB_POISON)) 1696 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); 1697 } 1698 } 1699 #else 1700 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1701 { 1702 if (cachep->dtor) { 1703 int i; 1704 for (i = 0; i < cachep->num; i++) { 1705 void *objp = index_to_obj(cachep, slabp, i); 1706 (cachep->dtor) (objp, cachep, 0); 1707 } 1708 } 1709 } 1710 #endif 1711 1712 /** 1713 * slab_destroy - destroy and release all objects in a slab 1714 * @cachep: cache pointer being destroyed 1715 * @slabp: slab pointer being destroyed 1716 * 1717 * Destroy all the objs in a slab, and release the mem back to the system. 1718 * Before calling the slab must have been unlinked from the cache. The 1719 * cache-lock is not held/needed. 1720 */ 1721 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1722 { 1723 void *addr = slabp->s_mem - slabp->colouroff; 1724 1725 slab_destroy_objs(cachep, slabp); 1726 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1727 struct slab_rcu *slab_rcu; 1728 1729 slab_rcu = (struct slab_rcu *)slabp; 1730 slab_rcu->cachep = cachep; 1731 slab_rcu->addr = addr; 1732 call_rcu(&slab_rcu->head, kmem_rcu_free); 1733 } else { 1734 kmem_freepages(cachep, addr); 1735 if (OFF_SLAB(cachep)) 1736 kmem_cache_free(cachep->slabp_cache, slabp); 1737 } 1738 } 1739 1740 /* 1741 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1742 * size of kmem_list3. 1743 */ 1744 static void set_up_list3s(struct kmem_cache *cachep, int index) 1745 { 1746 int node; 1747 1748 for_each_online_node(node) { 1749 cachep->nodelists[node] = &initkmem_list3[index + node]; 1750 cachep->nodelists[node]->next_reap = jiffies + 1751 REAPTIMEOUT_LIST3 + 1752 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1753 } 1754 } 1755 1756 /** 1757 * calculate_slab_order - calculate size (page order) of slabs 1758 * @cachep: pointer to the cache that is being created 1759 * @size: size of objects to be created in this cache. 1760 * @align: required alignment for the objects. 1761 * @flags: slab allocation flags 1762 * 1763 * Also calculates the number of objects per slab. 1764 * 1765 * This could be made much more intelligent. For now, try to avoid using 1766 * high order pages for slabs. When the gfp() functions are more friendly 1767 * towards high-order requests, this should be changed. 1768 */ 1769 static size_t calculate_slab_order(struct kmem_cache *cachep, 1770 size_t size, size_t align, unsigned long flags) 1771 { 1772 unsigned long offslab_limit; 1773 size_t left_over = 0; 1774 int gfporder; 1775 1776 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { 1777 unsigned int num; 1778 size_t remainder; 1779 1780 cache_estimate(gfporder, size, align, flags, &remainder, &num); 1781 if (!num) 1782 continue; 1783 1784 if (flags & CFLGS_OFF_SLAB) { 1785 /* 1786 * Max number of objs-per-slab for caches which 1787 * use off-slab slabs. Needed to avoid a possible 1788 * looping condition in cache_grow(). 1789 */ 1790 offslab_limit = size - sizeof(struct slab); 1791 offslab_limit /= sizeof(kmem_bufctl_t); 1792 1793 if (num > offslab_limit) 1794 break; 1795 } 1796 1797 /* Found something acceptable - save it away */ 1798 cachep->num = num; 1799 cachep->gfporder = gfporder; 1800 left_over = remainder; 1801 1802 /* 1803 * A VFS-reclaimable slab tends to have most allocations 1804 * as GFP_NOFS and we really don't want to have to be allocating 1805 * higher-order pages when we are unable to shrink dcache. 1806 */ 1807 if (flags & SLAB_RECLAIM_ACCOUNT) 1808 break; 1809 1810 /* 1811 * Large number of objects is good, but very large slabs are 1812 * currently bad for the gfp()s. 1813 */ 1814 if (gfporder >= slab_break_gfp_order) 1815 break; 1816 1817 /* 1818 * Acceptable internal fragmentation? 1819 */ 1820 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 1821 break; 1822 } 1823 return left_over; 1824 } 1825 1826 static void setup_cpu_cache(struct kmem_cache *cachep) 1827 { 1828 if (g_cpucache_up == FULL) { 1829 enable_cpucache(cachep); 1830 return; 1831 } 1832 if (g_cpucache_up == NONE) { 1833 /* 1834 * Note: the first kmem_cache_create must create the cache 1835 * that's used by kmalloc(24), otherwise the creation of 1836 * further caches will BUG(). 1837 */ 1838 cachep->array[smp_processor_id()] = &initarray_generic.cache; 1839 1840 /* 1841 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 1842 * the first cache, then we need to set up all its list3s, 1843 * otherwise the creation of further caches will BUG(). 1844 */ 1845 set_up_list3s(cachep, SIZE_AC); 1846 if (INDEX_AC == INDEX_L3) 1847 g_cpucache_up = PARTIAL_L3; 1848 else 1849 g_cpucache_up = PARTIAL_AC; 1850 } else { 1851 cachep->array[smp_processor_id()] = 1852 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1853 1854 if (g_cpucache_up == PARTIAL_AC) { 1855 set_up_list3s(cachep, SIZE_L3); 1856 g_cpucache_up = PARTIAL_L3; 1857 } else { 1858 int node; 1859 for_each_online_node(node) { 1860 cachep->nodelists[node] = 1861 kmalloc_node(sizeof(struct kmem_list3), 1862 GFP_KERNEL, node); 1863 BUG_ON(!cachep->nodelists[node]); 1864 kmem_list3_init(cachep->nodelists[node]); 1865 } 1866 } 1867 } 1868 cachep->nodelists[numa_node_id()]->next_reap = 1869 jiffies + REAPTIMEOUT_LIST3 + 1870 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1871 1872 cpu_cache_get(cachep)->avail = 0; 1873 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 1874 cpu_cache_get(cachep)->batchcount = 1; 1875 cpu_cache_get(cachep)->touched = 0; 1876 cachep->batchcount = 1; 1877 cachep->limit = BOOT_CPUCACHE_ENTRIES; 1878 } 1879 1880 /** 1881 * kmem_cache_create - Create a cache. 1882 * @name: A string which is used in /proc/slabinfo to identify this cache. 1883 * @size: The size of objects to be created in this cache. 1884 * @align: The required alignment for the objects. 1885 * @flags: SLAB flags 1886 * @ctor: A constructor for the objects. 1887 * @dtor: A destructor for the objects. 1888 * 1889 * Returns a ptr to the cache on success, NULL on failure. 1890 * Cannot be called within a int, but can be interrupted. 1891 * The @ctor is run when new pages are allocated by the cache 1892 * and the @dtor is run before the pages are handed back. 1893 * 1894 * @name must be valid until the cache is destroyed. This implies that 1895 * the module calling this has to destroy the cache before getting unloaded. 1896 * 1897 * The flags are 1898 * 1899 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 1900 * to catch references to uninitialised memory. 1901 * 1902 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 1903 * for buffer overruns. 1904 * 1905 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 1906 * cacheline. This can be beneficial if you're counting cycles as closely 1907 * as davem. 1908 */ 1909 struct kmem_cache * 1910 kmem_cache_create (const char *name, size_t size, size_t align, 1911 unsigned long flags, 1912 void (*ctor)(void*, struct kmem_cache *, unsigned long), 1913 void (*dtor)(void*, struct kmem_cache *, unsigned long)) 1914 { 1915 size_t left_over, slab_size, ralign; 1916 struct kmem_cache *cachep = NULL; 1917 struct list_head *p; 1918 1919 /* 1920 * Sanity checks... these are all serious usage bugs. 1921 */ 1922 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 1923 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { 1924 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, 1925 name); 1926 BUG(); 1927 } 1928 1929 /* 1930 * Prevent CPUs from coming and going. 1931 * lock_cpu_hotplug() nests outside cache_chain_mutex 1932 */ 1933 lock_cpu_hotplug(); 1934 1935 mutex_lock(&cache_chain_mutex); 1936 1937 list_for_each(p, &cache_chain) { 1938 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next); 1939 mm_segment_t old_fs = get_fs(); 1940 char tmp; 1941 int res; 1942 1943 /* 1944 * This happens when the module gets unloaded and doesn't 1945 * destroy its slab cache and no-one else reuses the vmalloc 1946 * area of the module. Print a warning. 1947 */ 1948 set_fs(KERNEL_DS); 1949 res = __get_user(tmp, pc->name); 1950 set_fs(old_fs); 1951 if (res) { 1952 printk("SLAB: cache with size %d has lost its name\n", 1953 pc->buffer_size); 1954 continue; 1955 } 1956 1957 if (!strcmp(pc->name, name)) { 1958 printk("kmem_cache_create: duplicate cache %s\n", name); 1959 dump_stack(); 1960 goto oops; 1961 } 1962 } 1963 1964 #if DEBUG 1965 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 1966 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { 1967 /* No constructor, but inital state check requested */ 1968 printk(KERN_ERR "%s: No con, but init state check " 1969 "requested - %s\n", __FUNCTION__, name); 1970 flags &= ~SLAB_DEBUG_INITIAL; 1971 } 1972 #if FORCED_DEBUG 1973 /* 1974 * Enable redzoning and last user accounting, except for caches with 1975 * large objects, if the increased size would increase the object size 1976 * above the next power of two: caches with object sizes just above a 1977 * power of two have a significant amount of internal fragmentation. 1978 */ 1979 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) 1980 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 1981 if (!(flags & SLAB_DESTROY_BY_RCU)) 1982 flags |= SLAB_POISON; 1983 #endif 1984 if (flags & SLAB_DESTROY_BY_RCU) 1985 BUG_ON(flags & SLAB_POISON); 1986 #endif 1987 if (flags & SLAB_DESTROY_BY_RCU) 1988 BUG_ON(dtor); 1989 1990 /* 1991 * Always checks flags, a caller might be expecting debug support which 1992 * isn't available. 1993 */ 1994 BUG_ON(flags & ~CREATE_MASK); 1995 1996 /* 1997 * Check that size is in terms of words. This is needed to avoid 1998 * unaligned accesses for some archs when redzoning is used, and makes 1999 * sure any on-slab bufctl's are also correctly aligned. 2000 */ 2001 if (size & (BYTES_PER_WORD - 1)) { 2002 size += (BYTES_PER_WORD - 1); 2003 size &= ~(BYTES_PER_WORD - 1); 2004 } 2005 2006 /* calculate the final buffer alignment: */ 2007 2008 /* 1) arch recommendation: can be overridden for debug */ 2009 if (flags & SLAB_HWCACHE_ALIGN) { 2010 /* 2011 * Default alignment: as specified by the arch code. Except if 2012 * an object is really small, then squeeze multiple objects into 2013 * one cacheline. 2014 */ 2015 ralign = cache_line_size(); 2016 while (size <= ralign / 2) 2017 ralign /= 2; 2018 } else { 2019 ralign = BYTES_PER_WORD; 2020 } 2021 /* 2) arch mandated alignment: disables debug if necessary */ 2022 if (ralign < ARCH_SLAB_MINALIGN) { 2023 ralign = ARCH_SLAB_MINALIGN; 2024 if (ralign > BYTES_PER_WORD) 2025 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2026 } 2027 /* 3) caller mandated alignment: disables debug if necessary */ 2028 if (ralign < align) { 2029 ralign = align; 2030 if (ralign > BYTES_PER_WORD) 2031 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2032 } 2033 /* 2034 * 4) Store it. Note that the debug code below can reduce 2035 * the alignment to BYTES_PER_WORD. 2036 */ 2037 align = ralign; 2038 2039 /* Get cache's description obj. */ 2040 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); 2041 if (!cachep) 2042 goto oops; 2043 2044 #if DEBUG 2045 cachep->obj_size = size; 2046 2047 if (flags & SLAB_RED_ZONE) { 2048 /* redzoning only works with word aligned caches */ 2049 align = BYTES_PER_WORD; 2050 2051 /* add space for red zone words */ 2052 cachep->obj_offset += BYTES_PER_WORD; 2053 size += 2 * BYTES_PER_WORD; 2054 } 2055 if (flags & SLAB_STORE_USER) { 2056 /* user store requires word alignment and 2057 * one word storage behind the end of the real 2058 * object. 2059 */ 2060 align = BYTES_PER_WORD; 2061 size += BYTES_PER_WORD; 2062 } 2063 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2064 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2065 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { 2066 cachep->obj_offset += PAGE_SIZE - size; 2067 size = PAGE_SIZE; 2068 } 2069 #endif 2070 #endif 2071 2072 /* Determine if the slab management is 'on' or 'off' slab. */ 2073 if (size >= (PAGE_SIZE >> 3)) 2074 /* 2075 * Size is large, assume best to place the slab management obj 2076 * off-slab (should allow better packing of objs). 2077 */ 2078 flags |= CFLGS_OFF_SLAB; 2079 2080 size = ALIGN(size, align); 2081 2082 left_over = calculate_slab_order(cachep, size, align, flags); 2083 2084 if (!cachep->num) { 2085 printk("kmem_cache_create: couldn't create cache %s.\n", name); 2086 kmem_cache_free(&cache_cache, cachep); 2087 cachep = NULL; 2088 goto oops; 2089 } 2090 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2091 + sizeof(struct slab), align); 2092 2093 /* 2094 * If the slab has been placed off-slab, and we have enough space then 2095 * move it on-slab. This is at the expense of any extra colouring. 2096 */ 2097 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2098 flags &= ~CFLGS_OFF_SLAB; 2099 left_over -= slab_size; 2100 } 2101 2102 if (flags & CFLGS_OFF_SLAB) { 2103 /* really off slab. No need for manual alignment */ 2104 slab_size = 2105 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2106 } 2107 2108 cachep->colour_off = cache_line_size(); 2109 /* Offset must be a multiple of the alignment. */ 2110 if (cachep->colour_off < align) 2111 cachep->colour_off = align; 2112 cachep->colour = left_over / cachep->colour_off; 2113 cachep->slab_size = slab_size; 2114 cachep->flags = flags; 2115 cachep->gfpflags = 0; 2116 if (flags & SLAB_CACHE_DMA) 2117 cachep->gfpflags |= GFP_DMA; 2118 cachep->buffer_size = size; 2119 2120 if (flags & CFLGS_OFF_SLAB) 2121 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2122 cachep->ctor = ctor; 2123 cachep->dtor = dtor; 2124 cachep->name = name; 2125 2126 2127 setup_cpu_cache(cachep); 2128 2129 /* cache setup completed, link it into the list */ 2130 list_add(&cachep->next, &cache_chain); 2131 oops: 2132 if (!cachep && (flags & SLAB_PANIC)) 2133 panic("kmem_cache_create(): failed to create slab `%s'\n", 2134 name); 2135 mutex_unlock(&cache_chain_mutex); 2136 unlock_cpu_hotplug(); 2137 return cachep; 2138 } 2139 EXPORT_SYMBOL(kmem_cache_create); 2140 2141 #if DEBUG 2142 static void check_irq_off(void) 2143 { 2144 BUG_ON(!irqs_disabled()); 2145 } 2146 2147 static void check_irq_on(void) 2148 { 2149 BUG_ON(irqs_disabled()); 2150 } 2151 2152 static void check_spinlock_acquired(struct kmem_cache *cachep) 2153 { 2154 #ifdef CONFIG_SMP 2155 check_irq_off(); 2156 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); 2157 #endif 2158 } 2159 2160 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2161 { 2162 #ifdef CONFIG_SMP 2163 check_irq_off(); 2164 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2165 #endif 2166 } 2167 2168 #else 2169 #define check_irq_off() do { } while(0) 2170 #define check_irq_on() do { } while(0) 2171 #define check_spinlock_acquired(x) do { } while(0) 2172 #define check_spinlock_acquired_node(x, y) do { } while(0) 2173 #endif 2174 2175 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2176 struct array_cache *ac, 2177 int force, int node); 2178 2179 static void do_drain(void *arg) 2180 { 2181 struct kmem_cache *cachep = arg; 2182 struct array_cache *ac; 2183 int node = numa_node_id(); 2184 2185 check_irq_off(); 2186 ac = cpu_cache_get(cachep); 2187 spin_lock(&cachep->nodelists[node]->list_lock); 2188 free_block(cachep, ac->entry, ac->avail, node); 2189 spin_unlock(&cachep->nodelists[node]->list_lock); 2190 ac->avail = 0; 2191 } 2192 2193 static void drain_cpu_caches(struct kmem_cache *cachep) 2194 { 2195 struct kmem_list3 *l3; 2196 int node; 2197 2198 on_each_cpu(do_drain, cachep, 1, 1); 2199 check_irq_on(); 2200 for_each_online_node(node) { 2201 l3 = cachep->nodelists[node]; 2202 if (l3 && l3->alien) 2203 drain_alien_cache(cachep, l3->alien); 2204 } 2205 2206 for_each_online_node(node) { 2207 l3 = cachep->nodelists[node]; 2208 if (l3) 2209 drain_array(cachep, l3, l3->shared, 1, node); 2210 } 2211 } 2212 2213 static int __node_shrink(struct kmem_cache *cachep, int node) 2214 { 2215 struct slab *slabp; 2216 struct kmem_list3 *l3 = cachep->nodelists[node]; 2217 int ret; 2218 2219 for (;;) { 2220 struct list_head *p; 2221 2222 p = l3->slabs_free.prev; 2223 if (p == &l3->slabs_free) 2224 break; 2225 2226 slabp = list_entry(l3->slabs_free.prev, struct slab, list); 2227 #if DEBUG 2228 BUG_ON(slabp->inuse); 2229 #endif 2230 list_del(&slabp->list); 2231 2232 l3->free_objects -= cachep->num; 2233 spin_unlock_irq(&l3->list_lock); 2234 slab_destroy(cachep, slabp); 2235 spin_lock_irq(&l3->list_lock); 2236 } 2237 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); 2238 return ret; 2239 } 2240 2241 static int __cache_shrink(struct kmem_cache *cachep) 2242 { 2243 int ret = 0, i = 0; 2244 struct kmem_list3 *l3; 2245 2246 drain_cpu_caches(cachep); 2247 2248 check_irq_on(); 2249 for_each_online_node(i) { 2250 l3 = cachep->nodelists[i]; 2251 if (l3) { 2252 spin_lock_irq(&l3->list_lock); 2253 ret += __node_shrink(cachep, i); 2254 spin_unlock_irq(&l3->list_lock); 2255 } 2256 } 2257 return (ret ? 1 : 0); 2258 } 2259 2260 /** 2261 * kmem_cache_shrink - Shrink a cache. 2262 * @cachep: The cache to shrink. 2263 * 2264 * Releases as many slabs as possible for a cache. 2265 * To help debugging, a zero exit status indicates all slabs were released. 2266 */ 2267 int kmem_cache_shrink(struct kmem_cache *cachep) 2268 { 2269 BUG_ON(!cachep || in_interrupt()); 2270 2271 return __cache_shrink(cachep); 2272 } 2273 EXPORT_SYMBOL(kmem_cache_shrink); 2274 2275 /** 2276 * kmem_cache_destroy - delete a cache 2277 * @cachep: the cache to destroy 2278 * 2279 * Remove a struct kmem_cache object from the slab cache. 2280 * Returns 0 on success. 2281 * 2282 * It is expected this function will be called by a module when it is 2283 * unloaded. This will remove the cache completely, and avoid a duplicate 2284 * cache being allocated each time a module is loaded and unloaded, if the 2285 * module doesn't have persistent in-kernel storage across loads and unloads. 2286 * 2287 * The cache must be empty before calling this function. 2288 * 2289 * The caller must guarantee that noone will allocate memory from the cache 2290 * during the kmem_cache_destroy(). 2291 */ 2292 int kmem_cache_destroy(struct kmem_cache *cachep) 2293 { 2294 int i; 2295 struct kmem_list3 *l3; 2296 2297 BUG_ON(!cachep || in_interrupt()); 2298 2299 /* Don't let CPUs to come and go */ 2300 lock_cpu_hotplug(); 2301 2302 /* Find the cache in the chain of caches. */ 2303 mutex_lock(&cache_chain_mutex); 2304 /* 2305 * the chain is never empty, cache_cache is never destroyed 2306 */ 2307 list_del(&cachep->next); 2308 mutex_unlock(&cache_chain_mutex); 2309 2310 if (__cache_shrink(cachep)) { 2311 slab_error(cachep, "Can't free all objects"); 2312 mutex_lock(&cache_chain_mutex); 2313 list_add(&cachep->next, &cache_chain); 2314 mutex_unlock(&cache_chain_mutex); 2315 unlock_cpu_hotplug(); 2316 return 1; 2317 } 2318 2319 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2320 synchronize_rcu(); 2321 2322 for_each_online_cpu(i) 2323 kfree(cachep->array[i]); 2324 2325 /* NUMA: free the list3 structures */ 2326 for_each_online_node(i) { 2327 l3 = cachep->nodelists[i]; 2328 if (l3) { 2329 kfree(l3->shared); 2330 free_alien_cache(l3->alien); 2331 kfree(l3); 2332 } 2333 } 2334 kmem_cache_free(&cache_cache, cachep); 2335 unlock_cpu_hotplug(); 2336 return 0; 2337 } 2338 EXPORT_SYMBOL(kmem_cache_destroy); 2339 2340 /* Get the memory for a slab management obj. */ 2341 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2342 int colour_off, gfp_t local_flags, 2343 int nodeid) 2344 { 2345 struct slab *slabp; 2346 2347 if (OFF_SLAB(cachep)) { 2348 /* Slab management obj is off-slab. */ 2349 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2350 local_flags, nodeid); 2351 if (!slabp) 2352 return NULL; 2353 } else { 2354 slabp = objp + colour_off; 2355 colour_off += cachep->slab_size; 2356 } 2357 slabp->inuse = 0; 2358 slabp->colouroff = colour_off; 2359 slabp->s_mem = objp + colour_off; 2360 slabp->nodeid = nodeid; 2361 return slabp; 2362 } 2363 2364 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2365 { 2366 return (kmem_bufctl_t *) (slabp + 1); 2367 } 2368 2369 static void cache_init_objs(struct kmem_cache *cachep, 2370 struct slab *slabp, unsigned long ctor_flags) 2371 { 2372 int i; 2373 2374 for (i = 0; i < cachep->num; i++) { 2375 void *objp = index_to_obj(cachep, slabp, i); 2376 #if DEBUG 2377 /* need to poison the objs? */ 2378 if (cachep->flags & SLAB_POISON) 2379 poison_obj(cachep, objp, POISON_FREE); 2380 if (cachep->flags & SLAB_STORE_USER) 2381 *dbg_userword(cachep, objp) = NULL; 2382 2383 if (cachep->flags & SLAB_RED_ZONE) { 2384 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2385 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2386 } 2387 /* 2388 * Constructors are not allowed to allocate memory from the same 2389 * cache which they are a constructor for. Otherwise, deadlock. 2390 * They must also be threaded. 2391 */ 2392 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2393 cachep->ctor(objp + obj_offset(cachep), cachep, 2394 ctor_flags); 2395 2396 if (cachep->flags & SLAB_RED_ZONE) { 2397 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2398 slab_error(cachep, "constructor overwrote the" 2399 " end of an object"); 2400 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2401 slab_error(cachep, "constructor overwrote the" 2402 " start of an object"); 2403 } 2404 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2405 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2406 kernel_map_pages(virt_to_page(objp), 2407 cachep->buffer_size / PAGE_SIZE, 0); 2408 #else 2409 if (cachep->ctor) 2410 cachep->ctor(objp, cachep, ctor_flags); 2411 #endif 2412 slab_bufctl(slabp)[i] = i + 1; 2413 } 2414 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2415 slabp->free = 0; 2416 } 2417 2418 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2419 { 2420 if (flags & SLAB_DMA) 2421 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2422 else 2423 BUG_ON(cachep->gfpflags & GFP_DMA); 2424 } 2425 2426 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2427 int nodeid) 2428 { 2429 void *objp = index_to_obj(cachep, slabp, slabp->free); 2430 kmem_bufctl_t next; 2431 2432 slabp->inuse++; 2433 next = slab_bufctl(slabp)[slabp->free]; 2434 #if DEBUG 2435 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2436 WARN_ON(slabp->nodeid != nodeid); 2437 #endif 2438 slabp->free = next; 2439 2440 return objp; 2441 } 2442 2443 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2444 void *objp, int nodeid) 2445 { 2446 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2447 2448 #if DEBUG 2449 /* Verify that the slab belongs to the intended node */ 2450 WARN_ON(slabp->nodeid != nodeid); 2451 2452 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2453 printk(KERN_ERR "slab: double free detected in cache " 2454 "'%s', objp %p\n", cachep->name, objp); 2455 BUG(); 2456 } 2457 #endif 2458 slab_bufctl(slabp)[objnr] = slabp->free; 2459 slabp->free = objnr; 2460 slabp->inuse--; 2461 } 2462 2463 static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, 2464 void *objp) 2465 { 2466 int i; 2467 struct page *page; 2468 2469 /* Nasty!!!!!! I hope this is OK. */ 2470 page = virt_to_page(objp); 2471 2472 i = 1; 2473 if (likely(!PageCompound(page))) 2474 i <<= cachep->gfporder; 2475 do { 2476 page_set_cache(page, cachep); 2477 page_set_slab(page, slabp); 2478 page++; 2479 } while (--i); 2480 } 2481 2482 /* 2483 * Grow (by 1) the number of slabs within a cache. This is called by 2484 * kmem_cache_alloc() when there are no active objs left in a cache. 2485 */ 2486 static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) 2487 { 2488 struct slab *slabp; 2489 void *objp; 2490 size_t offset; 2491 gfp_t local_flags; 2492 unsigned long ctor_flags; 2493 struct kmem_list3 *l3; 2494 2495 /* 2496 * Be lazy and only check for valid flags here, keeping it out of the 2497 * critical path in kmem_cache_alloc(). 2498 */ 2499 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); 2500 if (flags & SLAB_NO_GROW) 2501 return 0; 2502 2503 ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2504 local_flags = (flags & SLAB_LEVEL_MASK); 2505 if (!(local_flags & __GFP_WAIT)) 2506 /* 2507 * Not allowed to sleep. Need to tell a constructor about 2508 * this - it might need to know... 2509 */ 2510 ctor_flags |= SLAB_CTOR_ATOMIC; 2511 2512 /* Take the l3 list lock to change the colour_next on this node */ 2513 check_irq_off(); 2514 l3 = cachep->nodelists[nodeid]; 2515 spin_lock(&l3->list_lock); 2516 2517 /* Get colour for the slab, and cal the next value. */ 2518 offset = l3->colour_next; 2519 l3->colour_next++; 2520 if (l3->colour_next >= cachep->colour) 2521 l3->colour_next = 0; 2522 spin_unlock(&l3->list_lock); 2523 2524 offset *= cachep->colour_off; 2525 2526 if (local_flags & __GFP_WAIT) 2527 local_irq_enable(); 2528 2529 /* 2530 * The test for missing atomic flag is performed here, rather than 2531 * the more obvious place, simply to reduce the critical path length 2532 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2533 * will eventually be caught here (where it matters). 2534 */ 2535 kmem_flagcheck(cachep, flags); 2536 2537 /* 2538 * Get mem for the objs. Attempt to allocate a physical page from 2539 * 'nodeid'. 2540 */ 2541 objp = kmem_getpages(cachep, flags, nodeid); 2542 if (!objp) 2543 goto failed; 2544 2545 /* Get slab management. */ 2546 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); 2547 if (!slabp) 2548 goto opps1; 2549 2550 slabp->nodeid = nodeid; 2551 set_slab_attr(cachep, slabp, objp); 2552 2553 cache_init_objs(cachep, slabp, ctor_flags); 2554 2555 if (local_flags & __GFP_WAIT) 2556 local_irq_disable(); 2557 check_irq_off(); 2558 spin_lock(&l3->list_lock); 2559 2560 /* Make slab active. */ 2561 list_add_tail(&slabp->list, &(l3->slabs_free)); 2562 STATS_INC_GROWN(cachep); 2563 l3->free_objects += cachep->num; 2564 spin_unlock(&l3->list_lock); 2565 return 1; 2566 opps1: 2567 kmem_freepages(cachep, objp); 2568 failed: 2569 if (local_flags & __GFP_WAIT) 2570 local_irq_disable(); 2571 return 0; 2572 } 2573 2574 #if DEBUG 2575 2576 /* 2577 * Perform extra freeing checks: 2578 * - detect bad pointers. 2579 * - POISON/RED_ZONE checking 2580 * - destructor calls, for caches with POISON+dtor 2581 */ 2582 static void kfree_debugcheck(const void *objp) 2583 { 2584 struct page *page; 2585 2586 if (!virt_addr_valid(objp)) { 2587 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2588 (unsigned long)objp); 2589 BUG(); 2590 } 2591 page = virt_to_page(objp); 2592 if (!PageSlab(page)) { 2593 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", 2594 (unsigned long)objp); 2595 BUG(); 2596 } 2597 } 2598 2599 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2600 void *caller) 2601 { 2602 struct page *page; 2603 unsigned int objnr; 2604 struct slab *slabp; 2605 2606 objp -= obj_offset(cachep); 2607 kfree_debugcheck(objp); 2608 page = virt_to_page(objp); 2609 2610 if (page_get_cache(page) != cachep) { 2611 printk(KERN_ERR "mismatch in kmem_cache_free: expected " 2612 "cache %p, got %p\n", 2613 page_get_cache(page), cachep); 2614 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); 2615 printk(KERN_ERR "%p is %s.\n", page_get_cache(page), 2616 page_get_cache(page)->name); 2617 WARN_ON(1); 2618 } 2619 slabp = page_get_slab(page); 2620 2621 if (cachep->flags & SLAB_RED_ZONE) { 2622 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || 2623 *dbg_redzone2(cachep, objp) != RED_ACTIVE) { 2624 slab_error(cachep, "double free, or memory outside" 2625 " object was overwritten"); 2626 printk(KERN_ERR "%p: redzone 1:0x%lx, " 2627 "redzone 2:0x%lx.\n", 2628 objp, *dbg_redzone1(cachep, objp), 2629 *dbg_redzone2(cachep, objp)); 2630 } 2631 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2632 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2633 } 2634 if (cachep->flags & SLAB_STORE_USER) 2635 *dbg_userword(cachep, objp) = caller; 2636 2637 objnr = obj_to_index(cachep, slabp, objp); 2638 2639 BUG_ON(objnr >= cachep->num); 2640 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2641 2642 if (cachep->flags & SLAB_DEBUG_INITIAL) { 2643 /* 2644 * Need to call the slab's constructor so the caller can 2645 * perform a verify of its state (debugging). Called without 2646 * the cache-lock held. 2647 */ 2648 cachep->ctor(objp + obj_offset(cachep), 2649 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); 2650 } 2651 if (cachep->flags & SLAB_POISON && cachep->dtor) { 2652 /* we want to cache poison the object, 2653 * call the destruction callback 2654 */ 2655 cachep->dtor(objp + obj_offset(cachep), cachep, 0); 2656 } 2657 #ifdef CONFIG_DEBUG_SLAB_LEAK 2658 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2659 #endif 2660 if (cachep->flags & SLAB_POISON) { 2661 #ifdef CONFIG_DEBUG_PAGEALLOC 2662 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2663 store_stackinfo(cachep, objp, (unsigned long)caller); 2664 kernel_map_pages(virt_to_page(objp), 2665 cachep->buffer_size / PAGE_SIZE, 0); 2666 } else { 2667 poison_obj(cachep, objp, POISON_FREE); 2668 } 2669 #else 2670 poison_obj(cachep, objp, POISON_FREE); 2671 #endif 2672 } 2673 return objp; 2674 } 2675 2676 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2677 { 2678 kmem_bufctl_t i; 2679 int entries = 0; 2680 2681 /* Check slab's freelist to see if this obj is there. */ 2682 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2683 entries++; 2684 if (entries > cachep->num || i >= cachep->num) 2685 goto bad; 2686 } 2687 if (entries != cachep->num - slabp->inuse) { 2688 bad: 2689 printk(KERN_ERR "slab: Internal list corruption detected in " 2690 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2691 cachep->name, cachep->num, slabp, slabp->inuse); 2692 for (i = 0; 2693 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2694 i++) { 2695 if (i % 16 == 0) 2696 printk("\n%03x:", i); 2697 printk(" %02x", ((unsigned char *)slabp)[i]); 2698 } 2699 printk("\n"); 2700 BUG(); 2701 } 2702 } 2703 #else 2704 #define kfree_debugcheck(x) do { } while(0) 2705 #define cache_free_debugcheck(x,objp,z) (objp) 2706 #define check_slabp(x,y) do { } while(0) 2707 #endif 2708 2709 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2710 { 2711 int batchcount; 2712 struct kmem_list3 *l3; 2713 struct array_cache *ac; 2714 2715 check_irq_off(); 2716 ac = cpu_cache_get(cachep); 2717 retry: 2718 batchcount = ac->batchcount; 2719 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2720 /* 2721 * If there was little recent activity on this cache, then 2722 * perform only a partial refill. Otherwise we could generate 2723 * refill bouncing. 2724 */ 2725 batchcount = BATCHREFILL_LIMIT; 2726 } 2727 l3 = cachep->nodelists[numa_node_id()]; 2728 2729 BUG_ON(ac->avail > 0 || !l3); 2730 spin_lock(&l3->list_lock); 2731 2732 /* See if we can refill from the shared array */ 2733 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) 2734 goto alloc_done; 2735 2736 while (batchcount > 0) { 2737 struct list_head *entry; 2738 struct slab *slabp; 2739 /* Get slab alloc is to come from. */ 2740 entry = l3->slabs_partial.next; 2741 if (entry == &l3->slabs_partial) { 2742 l3->free_touched = 1; 2743 entry = l3->slabs_free.next; 2744 if (entry == &l3->slabs_free) 2745 goto must_grow; 2746 } 2747 2748 slabp = list_entry(entry, struct slab, list); 2749 check_slabp(cachep, slabp); 2750 check_spinlock_acquired(cachep); 2751 while (slabp->inuse < cachep->num && batchcount--) { 2752 STATS_INC_ALLOCED(cachep); 2753 STATS_INC_ACTIVE(cachep); 2754 STATS_SET_HIGH(cachep); 2755 2756 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 2757 numa_node_id()); 2758 } 2759 check_slabp(cachep, slabp); 2760 2761 /* move slabp to correct slabp list: */ 2762 list_del(&slabp->list); 2763 if (slabp->free == BUFCTL_END) 2764 list_add(&slabp->list, &l3->slabs_full); 2765 else 2766 list_add(&slabp->list, &l3->slabs_partial); 2767 } 2768 2769 must_grow: 2770 l3->free_objects -= ac->avail; 2771 alloc_done: 2772 spin_unlock(&l3->list_lock); 2773 2774 if (unlikely(!ac->avail)) { 2775 int x; 2776 x = cache_grow(cachep, flags, numa_node_id()); 2777 2778 /* cache_grow can reenable interrupts, then ac could change. */ 2779 ac = cpu_cache_get(cachep); 2780 if (!x && ac->avail == 0) /* no objects in sight? abort */ 2781 return NULL; 2782 2783 if (!ac->avail) /* objects refilled by interrupt? */ 2784 goto retry; 2785 } 2786 ac->touched = 1; 2787 return ac->entry[--ac->avail]; 2788 } 2789 2790 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 2791 gfp_t flags) 2792 { 2793 might_sleep_if(flags & __GFP_WAIT); 2794 #if DEBUG 2795 kmem_flagcheck(cachep, flags); 2796 #endif 2797 } 2798 2799 #if DEBUG 2800 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 2801 gfp_t flags, void *objp, void *caller) 2802 { 2803 if (!objp) 2804 return objp; 2805 if (cachep->flags & SLAB_POISON) { 2806 #ifdef CONFIG_DEBUG_PAGEALLOC 2807 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 2808 kernel_map_pages(virt_to_page(objp), 2809 cachep->buffer_size / PAGE_SIZE, 1); 2810 else 2811 check_poison_obj(cachep, objp); 2812 #else 2813 check_poison_obj(cachep, objp); 2814 #endif 2815 poison_obj(cachep, objp, POISON_INUSE); 2816 } 2817 if (cachep->flags & SLAB_STORE_USER) 2818 *dbg_userword(cachep, objp) = caller; 2819 2820 if (cachep->flags & SLAB_RED_ZONE) { 2821 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 2822 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 2823 slab_error(cachep, "double free, or memory outside" 2824 " object was overwritten"); 2825 printk(KERN_ERR 2826 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", 2827 objp, *dbg_redzone1(cachep, objp), 2828 *dbg_redzone2(cachep, objp)); 2829 } 2830 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 2831 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 2832 } 2833 #ifdef CONFIG_DEBUG_SLAB_LEAK 2834 { 2835 struct slab *slabp; 2836 unsigned objnr; 2837 2838 slabp = page_get_slab(virt_to_page(objp)); 2839 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 2840 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 2841 } 2842 #endif 2843 objp += obj_offset(cachep); 2844 if (cachep->ctor && cachep->flags & SLAB_POISON) { 2845 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2846 2847 if (!(flags & __GFP_WAIT)) 2848 ctor_flags |= SLAB_CTOR_ATOMIC; 2849 2850 cachep->ctor(objp, cachep, ctor_flags); 2851 } 2852 return objp; 2853 } 2854 #else 2855 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 2856 #endif 2857 2858 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 2859 { 2860 void *objp; 2861 struct array_cache *ac; 2862 2863 #ifdef CONFIG_NUMA 2864 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { 2865 objp = alternate_node_alloc(cachep, flags); 2866 if (objp != NULL) 2867 return objp; 2868 } 2869 #endif 2870 2871 check_irq_off(); 2872 ac = cpu_cache_get(cachep); 2873 if (likely(ac->avail)) { 2874 STATS_INC_ALLOCHIT(cachep); 2875 ac->touched = 1; 2876 objp = ac->entry[--ac->avail]; 2877 } else { 2878 STATS_INC_ALLOCMISS(cachep); 2879 objp = cache_alloc_refill(cachep, flags); 2880 } 2881 return objp; 2882 } 2883 2884 static __always_inline void *__cache_alloc(struct kmem_cache *cachep, 2885 gfp_t flags, void *caller) 2886 { 2887 unsigned long save_flags; 2888 void *objp; 2889 2890 cache_alloc_debugcheck_before(cachep, flags); 2891 2892 local_irq_save(save_flags); 2893 objp = ____cache_alloc(cachep, flags); 2894 local_irq_restore(save_flags); 2895 objp = cache_alloc_debugcheck_after(cachep, flags, objp, 2896 caller); 2897 prefetchw(objp); 2898 return objp; 2899 } 2900 2901 #ifdef CONFIG_NUMA 2902 /* 2903 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 2904 * 2905 * If we are in_interrupt, then process context, including cpusets and 2906 * mempolicy, may not apply and should not be used for allocation policy. 2907 */ 2908 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 2909 { 2910 int nid_alloc, nid_here; 2911 2912 if (in_interrupt()) 2913 return NULL; 2914 nid_alloc = nid_here = numa_node_id(); 2915 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 2916 nid_alloc = cpuset_mem_spread_node(); 2917 else if (current->mempolicy) 2918 nid_alloc = slab_node(current->mempolicy); 2919 if (nid_alloc != nid_here) 2920 return __cache_alloc_node(cachep, flags, nid_alloc); 2921 return NULL; 2922 } 2923 2924 /* 2925 * A interface to enable slab creation on nodeid 2926 */ 2927 static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 2928 int nodeid) 2929 { 2930 struct list_head *entry; 2931 struct slab *slabp; 2932 struct kmem_list3 *l3; 2933 void *obj; 2934 int x; 2935 2936 l3 = cachep->nodelists[nodeid]; 2937 BUG_ON(!l3); 2938 2939 retry: 2940 check_irq_off(); 2941 spin_lock(&l3->list_lock); 2942 entry = l3->slabs_partial.next; 2943 if (entry == &l3->slabs_partial) { 2944 l3->free_touched = 1; 2945 entry = l3->slabs_free.next; 2946 if (entry == &l3->slabs_free) 2947 goto must_grow; 2948 } 2949 2950 slabp = list_entry(entry, struct slab, list); 2951 check_spinlock_acquired_node(cachep, nodeid); 2952 check_slabp(cachep, slabp); 2953 2954 STATS_INC_NODEALLOCS(cachep); 2955 STATS_INC_ACTIVE(cachep); 2956 STATS_SET_HIGH(cachep); 2957 2958 BUG_ON(slabp->inuse == cachep->num); 2959 2960 obj = slab_get_obj(cachep, slabp, nodeid); 2961 check_slabp(cachep, slabp); 2962 l3->free_objects--; 2963 /* move slabp to correct slabp list: */ 2964 list_del(&slabp->list); 2965 2966 if (slabp->free == BUFCTL_END) 2967 list_add(&slabp->list, &l3->slabs_full); 2968 else 2969 list_add(&slabp->list, &l3->slabs_partial); 2970 2971 spin_unlock(&l3->list_lock); 2972 goto done; 2973 2974 must_grow: 2975 spin_unlock(&l3->list_lock); 2976 x = cache_grow(cachep, flags, nodeid); 2977 2978 if (!x) 2979 return NULL; 2980 2981 goto retry; 2982 done: 2983 return obj; 2984 } 2985 #endif 2986 2987 /* 2988 * Caller needs to acquire correct kmem_list's list_lock 2989 */ 2990 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 2991 int node) 2992 { 2993 int i; 2994 struct kmem_list3 *l3; 2995 2996 for (i = 0; i < nr_objects; i++) { 2997 void *objp = objpp[i]; 2998 struct slab *slabp; 2999 3000 slabp = virt_to_slab(objp); 3001 l3 = cachep->nodelists[node]; 3002 list_del(&slabp->list); 3003 check_spinlock_acquired_node(cachep, node); 3004 check_slabp(cachep, slabp); 3005 slab_put_obj(cachep, slabp, objp, node); 3006 STATS_DEC_ACTIVE(cachep); 3007 l3->free_objects++; 3008 check_slabp(cachep, slabp); 3009 3010 /* fixup slab chains */ 3011 if (slabp->inuse == 0) { 3012 if (l3->free_objects > l3->free_limit) { 3013 l3->free_objects -= cachep->num; 3014 slab_destroy(cachep, slabp); 3015 } else { 3016 list_add(&slabp->list, &l3->slabs_free); 3017 } 3018 } else { 3019 /* Unconditionally move a slab to the end of the 3020 * partial list on free - maximum time for the 3021 * other objects to be freed, too. 3022 */ 3023 list_add_tail(&slabp->list, &l3->slabs_partial); 3024 } 3025 } 3026 } 3027 3028 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3029 { 3030 int batchcount; 3031 struct kmem_list3 *l3; 3032 int node = numa_node_id(); 3033 3034 batchcount = ac->batchcount; 3035 #if DEBUG 3036 BUG_ON(!batchcount || batchcount > ac->avail); 3037 #endif 3038 check_irq_off(); 3039 l3 = cachep->nodelists[node]; 3040 spin_lock(&l3->list_lock); 3041 if (l3->shared) { 3042 struct array_cache *shared_array = l3->shared; 3043 int max = shared_array->limit - shared_array->avail; 3044 if (max) { 3045 if (batchcount > max) 3046 batchcount = max; 3047 memcpy(&(shared_array->entry[shared_array->avail]), 3048 ac->entry, sizeof(void *) * batchcount); 3049 shared_array->avail += batchcount; 3050 goto free_done; 3051 } 3052 } 3053 3054 free_block(cachep, ac->entry, batchcount, node); 3055 free_done: 3056 #if STATS 3057 { 3058 int i = 0; 3059 struct list_head *p; 3060 3061 p = l3->slabs_free.next; 3062 while (p != &(l3->slabs_free)) { 3063 struct slab *slabp; 3064 3065 slabp = list_entry(p, struct slab, list); 3066 BUG_ON(slabp->inuse); 3067 3068 i++; 3069 p = p->next; 3070 } 3071 STATS_SET_FREEABLE(cachep, i); 3072 } 3073 #endif 3074 spin_unlock(&l3->list_lock); 3075 ac->avail -= batchcount; 3076 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3077 } 3078 3079 /* 3080 * Release an obj back to its cache. If the obj has a constructed state, it must 3081 * be in this state _before_ it is released. Called with disabled ints. 3082 */ 3083 static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3084 { 3085 struct array_cache *ac = cpu_cache_get(cachep); 3086 3087 check_irq_off(); 3088 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3089 3090 /* Make sure we are not freeing a object from another 3091 * node to the array cache on this cpu. 3092 */ 3093 #ifdef CONFIG_NUMA 3094 { 3095 struct slab *slabp; 3096 slabp = virt_to_slab(objp); 3097 if (unlikely(slabp->nodeid != numa_node_id())) { 3098 struct array_cache *alien = NULL; 3099 int nodeid = slabp->nodeid; 3100 struct kmem_list3 *l3; 3101 3102 l3 = cachep->nodelists[numa_node_id()]; 3103 STATS_INC_NODEFREES(cachep); 3104 if (l3->alien && l3->alien[nodeid]) { 3105 alien = l3->alien[nodeid]; 3106 spin_lock(&alien->lock); 3107 if (unlikely(alien->avail == alien->limit)) { 3108 STATS_INC_ACOVERFLOW(cachep); 3109 __drain_alien_cache(cachep, 3110 alien, nodeid); 3111 } 3112 alien->entry[alien->avail++] = objp; 3113 spin_unlock(&alien->lock); 3114 } else { 3115 spin_lock(&(cachep->nodelists[nodeid])-> 3116 list_lock); 3117 free_block(cachep, &objp, 1, nodeid); 3118 spin_unlock(&(cachep->nodelists[nodeid])-> 3119 list_lock); 3120 } 3121 return; 3122 } 3123 } 3124 #endif 3125 if (likely(ac->avail < ac->limit)) { 3126 STATS_INC_FREEHIT(cachep); 3127 ac->entry[ac->avail++] = objp; 3128 return; 3129 } else { 3130 STATS_INC_FREEMISS(cachep); 3131 cache_flusharray(cachep, ac); 3132 ac->entry[ac->avail++] = objp; 3133 } 3134 } 3135 3136 /** 3137 * kmem_cache_alloc - Allocate an object 3138 * @cachep: The cache to allocate from. 3139 * @flags: See kmalloc(). 3140 * 3141 * Allocate an object from this cache. The flags are only relevant 3142 * if the cache has no available objects. 3143 */ 3144 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3145 { 3146 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3147 } 3148 EXPORT_SYMBOL(kmem_cache_alloc); 3149 3150 /** 3151 * kmem_cache_alloc - Allocate an object. The memory is set to zero. 3152 * @cache: The cache to allocate from. 3153 * @flags: See kmalloc(). 3154 * 3155 * Allocate an object from this cache and set the allocated memory to zero. 3156 * The flags are only relevant if the cache has no available objects. 3157 */ 3158 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) 3159 { 3160 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); 3161 if (ret) 3162 memset(ret, 0, obj_size(cache)); 3163 return ret; 3164 } 3165 EXPORT_SYMBOL(kmem_cache_zalloc); 3166 3167 /** 3168 * kmem_ptr_validate - check if an untrusted pointer might 3169 * be a slab entry. 3170 * @cachep: the cache we're checking against 3171 * @ptr: pointer to validate 3172 * 3173 * This verifies that the untrusted pointer looks sane: 3174 * it is _not_ a guarantee that the pointer is actually 3175 * part of the slab cache in question, but it at least 3176 * validates that the pointer can be dereferenced and 3177 * looks half-way sane. 3178 * 3179 * Currently only used for dentry validation. 3180 */ 3181 int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) 3182 { 3183 unsigned long addr = (unsigned long)ptr; 3184 unsigned long min_addr = PAGE_OFFSET; 3185 unsigned long align_mask = BYTES_PER_WORD - 1; 3186 unsigned long size = cachep->buffer_size; 3187 struct page *page; 3188 3189 if (unlikely(addr < min_addr)) 3190 goto out; 3191 if (unlikely(addr > (unsigned long)high_memory - size)) 3192 goto out; 3193 if (unlikely(addr & align_mask)) 3194 goto out; 3195 if (unlikely(!kern_addr_valid(addr))) 3196 goto out; 3197 if (unlikely(!kern_addr_valid(addr + size - 1))) 3198 goto out; 3199 page = virt_to_page(ptr); 3200 if (unlikely(!PageSlab(page))) 3201 goto out; 3202 if (unlikely(page_get_cache(page) != cachep)) 3203 goto out; 3204 return 1; 3205 out: 3206 return 0; 3207 } 3208 3209 #ifdef CONFIG_NUMA 3210 /** 3211 * kmem_cache_alloc_node - Allocate an object on the specified node 3212 * @cachep: The cache to allocate from. 3213 * @flags: See kmalloc(). 3214 * @nodeid: node number of the target node. 3215 * 3216 * Identical to kmem_cache_alloc, except that this function is slow 3217 * and can sleep. And it will allocate memory on the given node, which 3218 * can improve the performance for cpu bound structures. 3219 * New and improved: it will now make sure that the object gets 3220 * put on the correct node list so that there is no false sharing. 3221 */ 3222 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3223 { 3224 unsigned long save_flags; 3225 void *ptr; 3226 3227 cache_alloc_debugcheck_before(cachep, flags); 3228 local_irq_save(save_flags); 3229 3230 if (nodeid == -1 || nodeid == numa_node_id() || 3231 !cachep->nodelists[nodeid]) 3232 ptr = ____cache_alloc(cachep, flags); 3233 else 3234 ptr = __cache_alloc_node(cachep, flags, nodeid); 3235 local_irq_restore(save_flags); 3236 3237 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, 3238 __builtin_return_address(0)); 3239 3240 return ptr; 3241 } 3242 EXPORT_SYMBOL(kmem_cache_alloc_node); 3243 3244 void *kmalloc_node(size_t size, gfp_t flags, int node) 3245 { 3246 struct kmem_cache *cachep; 3247 3248 cachep = kmem_find_general_cachep(size, flags); 3249 if (unlikely(cachep == NULL)) 3250 return NULL; 3251 return kmem_cache_alloc_node(cachep, flags, node); 3252 } 3253 EXPORT_SYMBOL(kmalloc_node); 3254 #endif 3255 3256 /** 3257 * kmalloc - allocate memory 3258 * @size: how many bytes of memory are required. 3259 * @flags: the type of memory to allocate. 3260 * @caller: function caller for debug tracking of the caller 3261 * 3262 * kmalloc is the normal method of allocating memory 3263 * in the kernel. 3264 * 3265 * The @flags argument may be one of: 3266 * 3267 * %GFP_USER - Allocate memory on behalf of user. May sleep. 3268 * 3269 * %GFP_KERNEL - Allocate normal kernel ram. May sleep. 3270 * 3271 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. 3272 * 3273 * Additionally, the %GFP_DMA flag may be set to indicate the memory 3274 * must be suitable for DMA. This can mean different things on different 3275 * platforms. For example, on i386, it means that the memory must come 3276 * from the first 16MB. 3277 */ 3278 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3279 void *caller) 3280 { 3281 struct kmem_cache *cachep; 3282 3283 /* If you want to save a few bytes .text space: replace 3284 * __ with kmem_. 3285 * Then kmalloc uses the uninlined functions instead of the inline 3286 * functions. 3287 */ 3288 cachep = __find_general_cachep(size, flags); 3289 if (unlikely(cachep == NULL)) 3290 return NULL; 3291 return __cache_alloc(cachep, flags, caller); 3292 } 3293 3294 3295 void *__kmalloc(size_t size, gfp_t flags) 3296 { 3297 #ifndef CONFIG_DEBUG_SLAB 3298 return __do_kmalloc(size, flags, NULL); 3299 #else 3300 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3301 #endif 3302 } 3303 EXPORT_SYMBOL(__kmalloc); 3304 3305 #ifdef CONFIG_DEBUG_SLAB 3306 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) 3307 { 3308 return __do_kmalloc(size, flags, caller); 3309 } 3310 EXPORT_SYMBOL(__kmalloc_track_caller); 3311 #endif 3312 3313 #ifdef CONFIG_SMP 3314 /** 3315 * __alloc_percpu - allocate one copy of the object for every present 3316 * cpu in the system, zeroing them. 3317 * Objects should be dereferenced using the per_cpu_ptr macro only. 3318 * 3319 * @size: how many bytes of memory are required. 3320 */ 3321 void *__alloc_percpu(size_t size) 3322 { 3323 int i; 3324 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); 3325 3326 if (!pdata) 3327 return NULL; 3328 3329 /* 3330 * Cannot use for_each_online_cpu since a cpu may come online 3331 * and we have no way of figuring out how to fix the array 3332 * that we have allocated then.... 3333 */ 3334 for_each_possible_cpu(i) { 3335 int node = cpu_to_node(i); 3336 3337 if (node_online(node)) 3338 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); 3339 else 3340 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); 3341 3342 if (!pdata->ptrs[i]) 3343 goto unwind_oom; 3344 memset(pdata->ptrs[i], 0, size); 3345 } 3346 3347 /* Catch derefs w/o wrappers */ 3348 return (void *)(~(unsigned long)pdata); 3349 3350 unwind_oom: 3351 while (--i >= 0) { 3352 if (!cpu_possible(i)) 3353 continue; 3354 kfree(pdata->ptrs[i]); 3355 } 3356 kfree(pdata); 3357 return NULL; 3358 } 3359 EXPORT_SYMBOL(__alloc_percpu); 3360 #endif 3361 3362 /** 3363 * kmem_cache_free - Deallocate an object 3364 * @cachep: The cache the allocation was from. 3365 * @objp: The previously allocated object. 3366 * 3367 * Free an object which was previously allocated from this 3368 * cache. 3369 */ 3370 void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3371 { 3372 unsigned long flags; 3373 3374 local_irq_save(flags); 3375 __cache_free(cachep, objp); 3376 local_irq_restore(flags); 3377 } 3378 EXPORT_SYMBOL(kmem_cache_free); 3379 3380 /** 3381 * kfree - free previously allocated memory 3382 * @objp: pointer returned by kmalloc. 3383 * 3384 * If @objp is NULL, no operation is performed. 3385 * 3386 * Don't free memory not originally allocated by kmalloc() 3387 * or you will run into trouble. 3388 */ 3389 void kfree(const void *objp) 3390 { 3391 struct kmem_cache *c; 3392 unsigned long flags; 3393 3394 if (unlikely(!objp)) 3395 return; 3396 local_irq_save(flags); 3397 kfree_debugcheck(objp); 3398 c = virt_to_cache(objp); 3399 mutex_debug_check_no_locks_freed(objp, obj_size(c)); 3400 __cache_free(c, (void *)objp); 3401 local_irq_restore(flags); 3402 } 3403 EXPORT_SYMBOL(kfree); 3404 3405 #ifdef CONFIG_SMP 3406 /** 3407 * free_percpu - free previously allocated percpu memory 3408 * @objp: pointer returned by alloc_percpu. 3409 * 3410 * Don't free memory not originally allocated by alloc_percpu() 3411 * The complemented objp is to check for that. 3412 */ 3413 void free_percpu(const void *objp) 3414 { 3415 int i; 3416 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); 3417 3418 /* 3419 * We allocate for all cpus so we cannot use for online cpu here. 3420 */ 3421 for_each_possible_cpu(i) 3422 kfree(p->ptrs[i]); 3423 kfree(p); 3424 } 3425 EXPORT_SYMBOL(free_percpu); 3426 #endif 3427 3428 unsigned int kmem_cache_size(struct kmem_cache *cachep) 3429 { 3430 return obj_size(cachep); 3431 } 3432 EXPORT_SYMBOL(kmem_cache_size); 3433 3434 const char *kmem_cache_name(struct kmem_cache *cachep) 3435 { 3436 return cachep->name; 3437 } 3438 EXPORT_SYMBOL_GPL(kmem_cache_name); 3439 3440 /* 3441 * This initializes kmem_list3 or resizes varioius caches for all nodes. 3442 */ 3443 static int alloc_kmemlist(struct kmem_cache *cachep) 3444 { 3445 int node; 3446 struct kmem_list3 *l3; 3447 struct array_cache *new_shared; 3448 struct array_cache **new_alien; 3449 3450 for_each_online_node(node) { 3451 3452 new_alien = alloc_alien_cache(node, cachep->limit); 3453 if (!new_alien) 3454 goto fail; 3455 3456 new_shared = alloc_arraycache(node, 3457 cachep->shared*cachep->batchcount, 3458 0xbaadf00d); 3459 if (!new_shared) { 3460 free_alien_cache(new_alien); 3461 goto fail; 3462 } 3463 3464 l3 = cachep->nodelists[node]; 3465 if (l3) { 3466 struct array_cache *shared = l3->shared; 3467 3468 spin_lock_irq(&l3->list_lock); 3469 3470 if (shared) 3471 free_block(cachep, shared->entry, 3472 shared->avail, node); 3473 3474 l3->shared = new_shared; 3475 if (!l3->alien) { 3476 l3->alien = new_alien; 3477 new_alien = NULL; 3478 } 3479 l3->free_limit = (1 + nr_cpus_node(node)) * 3480 cachep->batchcount + cachep->num; 3481 spin_unlock_irq(&l3->list_lock); 3482 kfree(shared); 3483 free_alien_cache(new_alien); 3484 continue; 3485 } 3486 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); 3487 if (!l3) { 3488 free_alien_cache(new_alien); 3489 kfree(new_shared); 3490 goto fail; 3491 } 3492 3493 kmem_list3_init(l3); 3494 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3495 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3496 l3->shared = new_shared; 3497 l3->alien = new_alien; 3498 l3->free_limit = (1 + nr_cpus_node(node)) * 3499 cachep->batchcount + cachep->num; 3500 cachep->nodelists[node] = l3; 3501 } 3502 return 0; 3503 3504 fail: 3505 if (!cachep->next.next) { 3506 /* Cache is not active yet. Roll back what we did */ 3507 node--; 3508 while (node >= 0) { 3509 if (cachep->nodelists[node]) { 3510 l3 = cachep->nodelists[node]; 3511 3512 kfree(l3->shared); 3513 free_alien_cache(l3->alien); 3514 kfree(l3); 3515 cachep->nodelists[node] = NULL; 3516 } 3517 node--; 3518 } 3519 } 3520 return -ENOMEM; 3521 } 3522 3523 struct ccupdate_struct { 3524 struct kmem_cache *cachep; 3525 struct array_cache *new[NR_CPUS]; 3526 }; 3527 3528 static void do_ccupdate_local(void *info) 3529 { 3530 struct ccupdate_struct *new = info; 3531 struct array_cache *old; 3532 3533 check_irq_off(); 3534 old = cpu_cache_get(new->cachep); 3535 3536 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3537 new->new[smp_processor_id()] = old; 3538 } 3539 3540 /* Always called with the cache_chain_mutex held */ 3541 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3542 int batchcount, int shared) 3543 { 3544 struct ccupdate_struct new; 3545 int i, err; 3546 3547 memset(&new.new, 0, sizeof(new.new)); 3548 for_each_online_cpu(i) { 3549 new.new[i] = alloc_arraycache(cpu_to_node(i), limit, 3550 batchcount); 3551 if (!new.new[i]) { 3552 for (i--; i >= 0; i--) 3553 kfree(new.new[i]); 3554 return -ENOMEM; 3555 } 3556 } 3557 new.cachep = cachep; 3558 3559 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1); 3560 3561 check_irq_on(); 3562 cachep->batchcount = batchcount; 3563 cachep->limit = limit; 3564 cachep->shared = shared; 3565 3566 for_each_online_cpu(i) { 3567 struct array_cache *ccold = new.new[i]; 3568 if (!ccold) 3569 continue; 3570 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3571 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 3572 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3573 kfree(ccold); 3574 } 3575 3576 err = alloc_kmemlist(cachep); 3577 if (err) { 3578 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", 3579 cachep->name, -err); 3580 BUG(); 3581 } 3582 return 0; 3583 } 3584 3585 /* Called with cache_chain_mutex held always */ 3586 static void enable_cpucache(struct kmem_cache *cachep) 3587 { 3588 int err; 3589 int limit, shared; 3590 3591 /* 3592 * The head array serves three purposes: 3593 * - create a LIFO ordering, i.e. return objects that are cache-warm 3594 * - reduce the number of spinlock operations. 3595 * - reduce the number of linked list operations on the slab and 3596 * bufctl chains: array operations are cheaper. 3597 * The numbers are guessed, we should auto-tune as described by 3598 * Bonwick. 3599 */ 3600 if (cachep->buffer_size > 131072) 3601 limit = 1; 3602 else if (cachep->buffer_size > PAGE_SIZE) 3603 limit = 8; 3604 else if (cachep->buffer_size > 1024) 3605 limit = 24; 3606 else if (cachep->buffer_size > 256) 3607 limit = 54; 3608 else 3609 limit = 120; 3610 3611 /* 3612 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 3613 * allocation behaviour: Most allocs on one cpu, most free operations 3614 * on another cpu. For these cases, an efficient object passing between 3615 * cpus is necessary. This is provided by a shared array. The array 3616 * replaces Bonwick's magazine layer. 3617 * On uniprocessor, it's functionally equivalent (but less efficient) 3618 * to a larger limit. Thus disabled by default. 3619 */ 3620 shared = 0; 3621 #ifdef CONFIG_SMP 3622 if (cachep->buffer_size <= PAGE_SIZE) 3623 shared = 8; 3624 #endif 3625 3626 #if DEBUG 3627 /* 3628 * With debugging enabled, large batchcount lead to excessively long 3629 * periods with disabled local interrupts. Limit the batchcount 3630 */ 3631 if (limit > 32) 3632 limit = 32; 3633 #endif 3634 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); 3635 if (err) 3636 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 3637 cachep->name, -err); 3638 } 3639 3640 /* 3641 * Drain an array if it contains any elements taking the l3 lock only if 3642 * necessary. Note that the l3 listlock also protects the array_cache 3643 * if drain_array() is used on the shared array. 3644 */ 3645 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 3646 struct array_cache *ac, int force, int node) 3647 { 3648 int tofree; 3649 3650 if (!ac || !ac->avail) 3651 return; 3652 if (ac->touched && !force) { 3653 ac->touched = 0; 3654 } else { 3655 spin_lock_irq(&l3->list_lock); 3656 if (ac->avail) { 3657 tofree = force ? ac->avail : (ac->limit + 4) / 5; 3658 if (tofree > ac->avail) 3659 tofree = (ac->avail + 1) / 2; 3660 free_block(cachep, ac->entry, tofree, node); 3661 ac->avail -= tofree; 3662 memmove(ac->entry, &(ac->entry[tofree]), 3663 sizeof(void *) * ac->avail); 3664 } 3665 spin_unlock_irq(&l3->list_lock); 3666 } 3667 } 3668 3669 /** 3670 * cache_reap - Reclaim memory from caches. 3671 * @unused: unused parameter 3672 * 3673 * Called from workqueue/eventd every few seconds. 3674 * Purpose: 3675 * - clear the per-cpu caches for this CPU. 3676 * - return freeable pages to the main free memory pool. 3677 * 3678 * If we cannot acquire the cache chain mutex then just give up - we'll try 3679 * again on the next iteration. 3680 */ 3681 static void cache_reap(void *unused) 3682 { 3683 struct list_head *walk; 3684 struct kmem_list3 *l3; 3685 int node = numa_node_id(); 3686 3687 if (!mutex_trylock(&cache_chain_mutex)) { 3688 /* Give up. Setup the next iteration. */ 3689 schedule_delayed_work(&__get_cpu_var(reap_work), 3690 REAPTIMEOUT_CPUC); 3691 return; 3692 } 3693 3694 list_for_each(walk, &cache_chain) { 3695 struct kmem_cache *searchp; 3696 struct list_head *p; 3697 int tofree; 3698 struct slab *slabp; 3699 3700 searchp = list_entry(walk, struct kmem_cache, next); 3701 check_irq_on(); 3702 3703 /* 3704 * We only take the l3 lock if absolutely necessary and we 3705 * have established with reasonable certainty that 3706 * we can do some work if the lock was obtained. 3707 */ 3708 l3 = searchp->nodelists[node]; 3709 3710 reap_alien(searchp, l3); 3711 3712 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 3713 3714 /* 3715 * These are racy checks but it does not matter 3716 * if we skip one check or scan twice. 3717 */ 3718 if (time_after(l3->next_reap, jiffies)) 3719 goto next; 3720 3721 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 3722 3723 drain_array(searchp, l3, l3->shared, 0, node); 3724 3725 if (l3->free_touched) { 3726 l3->free_touched = 0; 3727 goto next; 3728 } 3729 3730 tofree = (l3->free_limit + 5 * searchp->num - 1) / 3731 (5 * searchp->num); 3732 do { 3733 /* 3734 * Do not lock if there are no free blocks. 3735 */ 3736 if (list_empty(&l3->slabs_free)) 3737 break; 3738 3739 spin_lock_irq(&l3->list_lock); 3740 p = l3->slabs_free.next; 3741 if (p == &(l3->slabs_free)) { 3742 spin_unlock_irq(&l3->list_lock); 3743 break; 3744 } 3745 3746 slabp = list_entry(p, struct slab, list); 3747 BUG_ON(slabp->inuse); 3748 list_del(&slabp->list); 3749 STATS_INC_REAPED(searchp); 3750 3751 /* 3752 * Safe to drop the lock. The slab is no longer linked 3753 * to the cache. searchp cannot disappear, we hold 3754 * cache_chain_lock 3755 */ 3756 l3->free_objects -= searchp->num; 3757 spin_unlock_irq(&l3->list_lock); 3758 slab_destroy(searchp, slabp); 3759 } while (--tofree > 0); 3760 next: 3761 cond_resched(); 3762 } 3763 check_irq_on(); 3764 mutex_unlock(&cache_chain_mutex); 3765 next_reap_node(); 3766 /* Set up the next iteration */ 3767 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); 3768 } 3769 3770 #ifdef CONFIG_PROC_FS 3771 3772 static void print_slabinfo_header(struct seq_file *m) 3773 { 3774 /* 3775 * Output format version, so at least we can change it 3776 * without _too_ many complaints. 3777 */ 3778 #if STATS 3779 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 3780 #else 3781 seq_puts(m, "slabinfo - version: 2.1\n"); 3782 #endif 3783 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 3784 "<objperslab> <pagesperslab>"); 3785 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 3786 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 3787 #if STATS 3788 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 3789 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 3790 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 3791 #endif 3792 seq_putc(m, '\n'); 3793 } 3794 3795 static void *s_start(struct seq_file *m, loff_t *pos) 3796 { 3797 loff_t n = *pos; 3798 struct list_head *p; 3799 3800 mutex_lock(&cache_chain_mutex); 3801 if (!n) 3802 print_slabinfo_header(m); 3803 p = cache_chain.next; 3804 while (n--) { 3805 p = p->next; 3806 if (p == &cache_chain) 3807 return NULL; 3808 } 3809 return list_entry(p, struct kmem_cache, next); 3810 } 3811 3812 static void *s_next(struct seq_file *m, void *p, loff_t *pos) 3813 { 3814 struct kmem_cache *cachep = p; 3815 ++*pos; 3816 return cachep->next.next == &cache_chain ? 3817 NULL : list_entry(cachep->next.next, struct kmem_cache, next); 3818 } 3819 3820 static void s_stop(struct seq_file *m, void *p) 3821 { 3822 mutex_unlock(&cache_chain_mutex); 3823 } 3824 3825 static int s_show(struct seq_file *m, void *p) 3826 { 3827 struct kmem_cache *cachep = p; 3828 struct list_head *q; 3829 struct slab *slabp; 3830 unsigned long active_objs; 3831 unsigned long num_objs; 3832 unsigned long active_slabs = 0; 3833 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 3834 const char *name; 3835 char *error = NULL; 3836 int node; 3837 struct kmem_list3 *l3; 3838 3839 active_objs = 0; 3840 num_slabs = 0; 3841 for_each_online_node(node) { 3842 l3 = cachep->nodelists[node]; 3843 if (!l3) 3844 continue; 3845 3846 check_irq_on(); 3847 spin_lock_irq(&l3->list_lock); 3848 3849 list_for_each(q, &l3->slabs_full) { 3850 slabp = list_entry(q, struct slab, list); 3851 if (slabp->inuse != cachep->num && !error) 3852 error = "slabs_full accounting error"; 3853 active_objs += cachep->num; 3854 active_slabs++; 3855 } 3856 list_for_each(q, &l3->slabs_partial) { 3857 slabp = list_entry(q, struct slab, list); 3858 if (slabp->inuse == cachep->num && !error) 3859 error = "slabs_partial inuse accounting error"; 3860 if (!slabp->inuse && !error) 3861 error = "slabs_partial/inuse accounting error"; 3862 active_objs += slabp->inuse; 3863 active_slabs++; 3864 } 3865 list_for_each(q, &l3->slabs_free) { 3866 slabp = list_entry(q, struct slab, list); 3867 if (slabp->inuse && !error) 3868 error = "slabs_free/inuse accounting error"; 3869 num_slabs++; 3870 } 3871 free_objects += l3->free_objects; 3872 if (l3->shared) 3873 shared_avail += l3->shared->avail; 3874 3875 spin_unlock_irq(&l3->list_lock); 3876 } 3877 num_slabs += active_slabs; 3878 num_objs = num_slabs * cachep->num; 3879 if (num_objs - active_objs != free_objects && !error) 3880 error = "free_objects accounting error"; 3881 3882 name = cachep->name; 3883 if (error) 3884 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 3885 3886 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 3887 name, active_objs, num_objs, cachep->buffer_size, 3888 cachep->num, (1 << cachep->gfporder)); 3889 seq_printf(m, " : tunables %4u %4u %4u", 3890 cachep->limit, cachep->batchcount, cachep->shared); 3891 seq_printf(m, " : slabdata %6lu %6lu %6lu", 3892 active_slabs, num_slabs, shared_avail); 3893 #if STATS 3894 { /* list3 stats */ 3895 unsigned long high = cachep->high_mark; 3896 unsigned long allocs = cachep->num_allocations; 3897 unsigned long grown = cachep->grown; 3898 unsigned long reaped = cachep->reaped; 3899 unsigned long errors = cachep->errors; 3900 unsigned long max_freeable = cachep->max_freeable; 3901 unsigned long node_allocs = cachep->node_allocs; 3902 unsigned long node_frees = cachep->node_frees; 3903 unsigned long overflows = cachep->node_overflow; 3904 3905 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ 3906 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, 3907 reaped, errors, max_freeable, node_allocs, 3908 node_frees, overflows); 3909 } 3910 /* cpu stats */ 3911 { 3912 unsigned long allochit = atomic_read(&cachep->allochit); 3913 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 3914 unsigned long freehit = atomic_read(&cachep->freehit); 3915 unsigned long freemiss = atomic_read(&cachep->freemiss); 3916 3917 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 3918 allochit, allocmiss, freehit, freemiss); 3919 } 3920 #endif 3921 seq_putc(m, '\n'); 3922 return 0; 3923 } 3924 3925 /* 3926 * slabinfo_op - iterator that generates /proc/slabinfo 3927 * 3928 * Output layout: 3929 * cache-name 3930 * num-active-objs 3931 * total-objs 3932 * object size 3933 * num-active-slabs 3934 * total-slabs 3935 * num-pages-per-slab 3936 * + further values on SMP and with statistics enabled 3937 */ 3938 3939 struct seq_operations slabinfo_op = { 3940 .start = s_start, 3941 .next = s_next, 3942 .stop = s_stop, 3943 .show = s_show, 3944 }; 3945 3946 #define MAX_SLABINFO_WRITE 128 3947 /** 3948 * slabinfo_write - Tuning for the slab allocator 3949 * @file: unused 3950 * @buffer: user buffer 3951 * @count: data length 3952 * @ppos: unused 3953 */ 3954 ssize_t slabinfo_write(struct file *file, const char __user * buffer, 3955 size_t count, loff_t *ppos) 3956 { 3957 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 3958 int limit, batchcount, shared, res; 3959 struct list_head *p; 3960 3961 if (count > MAX_SLABINFO_WRITE) 3962 return -EINVAL; 3963 if (copy_from_user(&kbuf, buffer, count)) 3964 return -EFAULT; 3965 kbuf[MAX_SLABINFO_WRITE] = '\0'; 3966 3967 tmp = strchr(kbuf, ' '); 3968 if (!tmp) 3969 return -EINVAL; 3970 *tmp = '\0'; 3971 tmp++; 3972 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 3973 return -EINVAL; 3974 3975 /* Find the cache in the chain of caches. */ 3976 mutex_lock(&cache_chain_mutex); 3977 res = -EINVAL; 3978 list_for_each(p, &cache_chain) { 3979 struct kmem_cache *cachep; 3980 3981 cachep = list_entry(p, struct kmem_cache, next); 3982 if (!strcmp(cachep->name, kbuf)) { 3983 if (limit < 1 || batchcount < 1 || 3984 batchcount > limit || shared < 0) { 3985 res = 0; 3986 } else { 3987 res = do_tune_cpucache(cachep, limit, 3988 batchcount, shared); 3989 } 3990 break; 3991 } 3992 } 3993 mutex_unlock(&cache_chain_mutex); 3994 if (res >= 0) 3995 res = count; 3996 return res; 3997 } 3998 3999 #ifdef CONFIG_DEBUG_SLAB_LEAK 4000 4001 static void *leaks_start(struct seq_file *m, loff_t *pos) 4002 { 4003 loff_t n = *pos; 4004 struct list_head *p; 4005 4006 mutex_lock(&cache_chain_mutex); 4007 p = cache_chain.next; 4008 while (n--) { 4009 p = p->next; 4010 if (p == &cache_chain) 4011 return NULL; 4012 } 4013 return list_entry(p, struct kmem_cache, next); 4014 } 4015 4016 static inline int add_caller(unsigned long *n, unsigned long v) 4017 { 4018 unsigned long *p; 4019 int l; 4020 if (!v) 4021 return 1; 4022 l = n[1]; 4023 p = n + 2; 4024 while (l) { 4025 int i = l/2; 4026 unsigned long *q = p + 2 * i; 4027 if (*q == v) { 4028 q[1]++; 4029 return 1; 4030 } 4031 if (*q > v) { 4032 l = i; 4033 } else { 4034 p = q + 2; 4035 l -= i + 1; 4036 } 4037 } 4038 if (++n[1] == n[0]) 4039 return 0; 4040 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4041 p[0] = v; 4042 p[1] = 1; 4043 return 1; 4044 } 4045 4046 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4047 { 4048 void *p; 4049 int i; 4050 if (n[0] == n[1]) 4051 return; 4052 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4053 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4054 continue; 4055 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4056 return; 4057 } 4058 } 4059 4060 static void show_symbol(struct seq_file *m, unsigned long address) 4061 { 4062 #ifdef CONFIG_KALLSYMS 4063 char *modname; 4064 const char *name; 4065 unsigned long offset, size; 4066 char namebuf[KSYM_NAME_LEN+1]; 4067 4068 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); 4069 4070 if (name) { 4071 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4072 if (modname) 4073 seq_printf(m, " [%s]", modname); 4074 return; 4075 } 4076 #endif 4077 seq_printf(m, "%p", (void *)address); 4078 } 4079 4080 static int leaks_show(struct seq_file *m, void *p) 4081 { 4082 struct kmem_cache *cachep = p; 4083 struct list_head *q; 4084 struct slab *slabp; 4085 struct kmem_list3 *l3; 4086 const char *name; 4087 unsigned long *n = m->private; 4088 int node; 4089 int i; 4090 4091 if (!(cachep->flags & SLAB_STORE_USER)) 4092 return 0; 4093 if (!(cachep->flags & SLAB_RED_ZONE)) 4094 return 0; 4095 4096 /* OK, we can do it */ 4097 4098 n[1] = 0; 4099 4100 for_each_online_node(node) { 4101 l3 = cachep->nodelists[node]; 4102 if (!l3) 4103 continue; 4104 4105 check_irq_on(); 4106 spin_lock_irq(&l3->list_lock); 4107 4108 list_for_each(q, &l3->slabs_full) { 4109 slabp = list_entry(q, struct slab, list); 4110 handle_slab(n, cachep, slabp); 4111 } 4112 list_for_each(q, &l3->slabs_partial) { 4113 slabp = list_entry(q, struct slab, list); 4114 handle_slab(n, cachep, slabp); 4115 } 4116 spin_unlock_irq(&l3->list_lock); 4117 } 4118 name = cachep->name; 4119 if (n[0] == n[1]) { 4120 /* Increase the buffer size */ 4121 mutex_unlock(&cache_chain_mutex); 4122 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4123 if (!m->private) { 4124 /* Too bad, we are really out */ 4125 m->private = n; 4126 mutex_lock(&cache_chain_mutex); 4127 return -ENOMEM; 4128 } 4129 *(unsigned long *)m->private = n[0] * 2; 4130 kfree(n); 4131 mutex_lock(&cache_chain_mutex); 4132 /* Now make sure this entry will be retried */ 4133 m->count = m->size; 4134 return 0; 4135 } 4136 for (i = 0; i < n[1]; i++) { 4137 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4138 show_symbol(m, n[2*i+2]); 4139 seq_putc(m, '\n'); 4140 } 4141 return 0; 4142 } 4143 4144 struct seq_operations slabstats_op = { 4145 .start = leaks_start, 4146 .next = s_next, 4147 .stop = s_stop, 4148 .show = leaks_show, 4149 }; 4150 #endif 4151 #endif 4152 4153 /** 4154 * ksize - get the actual amount of memory allocated for a given object 4155 * @objp: Pointer to the object 4156 * 4157 * kmalloc may internally round up allocations and return more memory 4158 * than requested. ksize() can be used to determine the actual amount of 4159 * memory allocated. The caller may use this additional memory, even though 4160 * a smaller amount of memory was initially specified with the kmalloc call. 4161 * The caller must guarantee that objp points to a valid object previously 4162 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4163 * must not be freed during the duration of the call. 4164 */ 4165 unsigned int ksize(const void *objp) 4166 { 4167 if (unlikely(objp == NULL)) 4168 return 0; 4169 4170 return obj_size(virt_to_cache(objp)); 4171 } 4172