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/slab.h> 90 #include <linux/mm.h> 91 #include <linux/poison.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/uaccess.h> 107 #include <linux/nodemask.h> 108 #include <linux/mempolicy.h> 109 #include <linux/mutex.h> 110 #include <linux/fault-inject.h> 111 #include <linux/rtmutex.h> 112 #include <linux/reciprocal_div.h> 113 114 #include <asm/cacheflush.h> 115 #include <asm/tlbflush.h> 116 #include <asm/page.h> 117 118 /* 119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, 120 * SLAB_RED_ZONE & SLAB_POISON. 121 * 0 for faster, smaller code (especially in the critical paths). 122 * 123 * STATS - 1 to collect stats for /proc/slabinfo. 124 * 0 for faster, smaller code (especially in the critical paths). 125 * 126 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 127 */ 128 129 #ifdef CONFIG_DEBUG_SLAB 130 #define DEBUG 1 131 #define STATS 1 132 #define FORCED_DEBUG 1 133 #else 134 #define DEBUG 0 135 #define STATS 0 136 #define FORCED_DEBUG 0 137 #endif 138 139 /* Shouldn't this be in a header file somewhere? */ 140 #define BYTES_PER_WORD sizeof(void *) 141 142 #ifndef cache_line_size 143 #define cache_line_size() L1_CACHE_BYTES 144 #endif 145 146 #ifndef ARCH_KMALLOC_MINALIGN 147 /* 148 * Enforce a minimum alignment for the kmalloc caches. 149 * Usually, the kmalloc caches are cache_line_size() aligned, except when 150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. 151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed 152 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. 153 * Note that this flag disables some debug features. 154 */ 155 #define ARCH_KMALLOC_MINALIGN 0 156 #endif 157 158 #ifndef ARCH_SLAB_MINALIGN 159 /* 160 * Enforce a minimum alignment for all caches. 161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD 162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. 163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables 164 * some debug features. 165 */ 166 #define ARCH_SLAB_MINALIGN 0 167 #endif 168 169 #ifndef ARCH_KMALLOC_FLAGS 170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 171 #endif 172 173 /* Legal flag mask for kmem_cache_create(). */ 174 #if DEBUG 175 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ 176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 177 SLAB_CACHE_DMA | \ 178 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ 179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 181 #else 182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 183 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ 184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 186 #endif 187 188 /* 189 * kmem_bufctl_t: 190 * 191 * Bufctl's are used for linking objs within a slab 192 * linked offsets. 193 * 194 * This implementation relies on "struct page" for locating the cache & 195 * slab an object belongs to. 196 * This allows the bufctl structure to be small (one int), but limits 197 * the number of objects a slab (not a cache) can contain when off-slab 198 * bufctls are used. The limit is the size of the largest general cache 199 * that does not use off-slab slabs. 200 * For 32bit archs with 4 kB pages, is this 56. 201 * This is not serious, as it is only for large objects, when it is unwise 202 * to have too many per slab. 203 * Note: This limit can be raised by introducing a general cache whose size 204 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 205 */ 206 207 typedef unsigned int kmem_bufctl_t; 208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 212 213 /* 214 * struct slab 215 * 216 * Manages the objs in a slab. Placed either at the beginning of mem allocated 217 * for a slab, or allocated from an general cache. 218 * Slabs are chained into three list: fully used, partial, fully free slabs. 219 */ 220 struct slab { 221 struct list_head list; 222 unsigned long colouroff; 223 void *s_mem; /* including colour offset */ 224 unsigned int inuse; /* num of objs active in slab */ 225 kmem_bufctl_t free; 226 unsigned short nodeid; 227 }; 228 229 /* 230 * struct slab_rcu 231 * 232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 233 * arrange for kmem_freepages to be called via RCU. This is useful if 234 * we need to approach a kernel structure obliquely, from its address 235 * obtained without the usual locking. We can lock the structure to 236 * stabilize it and check it's still at the given address, only if we 237 * can be sure that the memory has not been meanwhile reused for some 238 * other kind of object (which our subsystem's lock might corrupt). 239 * 240 * rcu_read_lock before reading the address, then rcu_read_unlock after 241 * taking the spinlock within the structure expected at that address. 242 * 243 * We assume struct slab_rcu can overlay struct slab when destroying. 244 */ 245 struct slab_rcu { 246 struct rcu_head head; 247 struct kmem_cache *cachep; 248 void *addr; 249 }; 250 251 /* 252 * struct array_cache 253 * 254 * Purpose: 255 * - LIFO ordering, to hand out cache-warm objects from _alloc 256 * - reduce the number of linked list operations 257 * - reduce spinlock operations 258 * 259 * The limit is stored in the per-cpu structure to reduce the data cache 260 * footprint. 261 * 262 */ 263 struct array_cache { 264 unsigned int avail; 265 unsigned int limit; 266 unsigned int batchcount; 267 unsigned int touched; 268 spinlock_t lock; 269 void *entry[0]; /* 270 * Must have this definition in here for the proper 271 * alignment of array_cache. Also simplifies accessing 272 * the entries. 273 * [0] is for gcc 2.95. It should really be []. 274 */ 275 }; 276 277 /* 278 * bootstrap: The caches do not work without cpuarrays anymore, but the 279 * cpuarrays are allocated from the generic caches... 280 */ 281 #define BOOT_CPUCACHE_ENTRIES 1 282 struct arraycache_init { 283 struct array_cache cache; 284 void *entries[BOOT_CPUCACHE_ENTRIES]; 285 }; 286 287 /* 288 * The slab lists for all objects. 289 */ 290 struct kmem_list3 { 291 struct list_head slabs_partial; /* partial list first, better asm code */ 292 struct list_head slabs_full; 293 struct list_head slabs_free; 294 unsigned long free_objects; 295 unsigned int free_limit; 296 unsigned int colour_next; /* Per-node cache coloring */ 297 spinlock_t list_lock; 298 struct array_cache *shared; /* shared per node */ 299 struct array_cache **alien; /* on other nodes */ 300 unsigned long next_reap; /* updated without locking */ 301 int free_touched; /* updated without locking */ 302 }; 303 304 /* 305 * Need this for bootstrapping a per node allocator. 306 */ 307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) 308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 309 #define CACHE_CACHE 0 310 #define SIZE_AC 1 311 #define SIZE_L3 (1 + MAX_NUMNODES) 312 313 static int drain_freelist(struct kmem_cache *cache, 314 struct kmem_list3 *l3, int tofree); 315 static void free_block(struct kmem_cache *cachep, void **objpp, int len, 316 int node); 317 static int enable_cpucache(struct kmem_cache *cachep); 318 static void cache_reap(struct work_struct *unused); 319 320 /* 321 * This function must be completely optimized away if a constant is passed to 322 * it. Mostly the same as what is in linux/slab.h except it returns an index. 323 */ 324 static __always_inline int index_of(const size_t size) 325 { 326 extern void __bad_size(void); 327 328 if (__builtin_constant_p(size)) { 329 int i = 0; 330 331 #define CACHE(x) \ 332 if (size <=x) \ 333 return i; \ 334 else \ 335 i++; 336 #include "linux/kmalloc_sizes.h" 337 #undef CACHE 338 __bad_size(); 339 } else 340 __bad_size(); 341 return 0; 342 } 343 344 static int slab_early_init = 1; 345 346 #define INDEX_AC index_of(sizeof(struct arraycache_init)) 347 #define INDEX_L3 index_of(sizeof(struct kmem_list3)) 348 349 static void kmem_list3_init(struct kmem_list3 *parent) 350 { 351 INIT_LIST_HEAD(&parent->slabs_full); 352 INIT_LIST_HEAD(&parent->slabs_partial); 353 INIT_LIST_HEAD(&parent->slabs_free); 354 parent->shared = NULL; 355 parent->alien = NULL; 356 parent->colour_next = 0; 357 spin_lock_init(&parent->list_lock); 358 parent->free_objects = 0; 359 parent->free_touched = 0; 360 } 361 362 #define MAKE_LIST(cachep, listp, slab, nodeid) \ 363 do { \ 364 INIT_LIST_HEAD(listp); \ 365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 366 } while (0) 367 368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 369 do { \ 370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 373 } while (0) 374 375 /* 376 * struct kmem_cache 377 * 378 * manages a cache. 379 */ 380 381 struct kmem_cache { 382 /* 1) per-cpu data, touched during every alloc/free */ 383 struct array_cache *array[NR_CPUS]; 384 /* 2) Cache tunables. Protected by cache_chain_mutex */ 385 unsigned int batchcount; 386 unsigned int limit; 387 unsigned int shared; 388 389 unsigned int buffer_size; 390 u32 reciprocal_buffer_size; 391 /* 3) touched by every alloc & free from the backend */ 392 struct kmem_list3 *nodelists[MAX_NUMNODES]; 393 394 unsigned int flags; /* constant flags */ 395 unsigned int num; /* # of objs per slab */ 396 397 /* 4) cache_grow/shrink */ 398 /* order of pgs per slab (2^n) */ 399 unsigned int gfporder; 400 401 /* force GFP flags, e.g. GFP_DMA */ 402 gfp_t gfpflags; 403 404 size_t colour; /* cache colouring range */ 405 unsigned int colour_off; /* colour offset */ 406 struct kmem_cache *slabp_cache; 407 unsigned int slab_size; 408 unsigned int dflags; /* dynamic flags */ 409 410 /* constructor func */ 411 void (*ctor) (void *, struct kmem_cache *, unsigned long); 412 413 /* de-constructor func */ 414 void (*dtor) (void *, struct kmem_cache *, unsigned long); 415 416 /* 5) cache creation/removal */ 417 const char *name; 418 struct list_head next; 419 420 /* 6) statistics */ 421 #if STATS 422 unsigned long num_active; 423 unsigned long num_allocations; 424 unsigned long high_mark; 425 unsigned long grown; 426 unsigned long reaped; 427 unsigned long errors; 428 unsigned long max_freeable; 429 unsigned long node_allocs; 430 unsigned long node_frees; 431 unsigned long node_overflow; 432 atomic_t allochit; 433 atomic_t allocmiss; 434 atomic_t freehit; 435 atomic_t freemiss; 436 #endif 437 #if DEBUG 438 /* 439 * If debugging is enabled, then the allocator can add additional 440 * fields and/or padding to every object. buffer_size contains the total 441 * object size including these internal fields, the following two 442 * variables contain the offset to the user object and its size. 443 */ 444 int obj_offset; 445 int obj_size; 446 #endif 447 }; 448 449 #define CFLGS_OFF_SLAB (0x80000000UL) 450 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 451 452 #define BATCHREFILL_LIMIT 16 453 /* 454 * Optimization question: fewer reaps means less probability for unnessary 455 * cpucache drain/refill cycles. 456 * 457 * OTOH the cpuarrays can contain lots of objects, 458 * which could lock up otherwise freeable slabs. 459 */ 460 #define REAPTIMEOUT_CPUC (2*HZ) 461 #define REAPTIMEOUT_LIST3 (4*HZ) 462 463 #if STATS 464 #define STATS_INC_ACTIVE(x) ((x)->num_active++) 465 #define STATS_DEC_ACTIVE(x) ((x)->num_active--) 466 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 467 #define STATS_INC_GROWN(x) ((x)->grown++) 468 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 469 #define STATS_SET_HIGH(x) \ 470 do { \ 471 if ((x)->num_active > (x)->high_mark) \ 472 (x)->high_mark = (x)->num_active; \ 473 } while (0) 474 #define STATS_INC_ERR(x) ((x)->errors++) 475 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 476 #define STATS_INC_NODEFREES(x) ((x)->node_frees++) 477 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 478 #define STATS_SET_FREEABLE(x, i) \ 479 do { \ 480 if ((x)->max_freeable < i) \ 481 (x)->max_freeable = i; \ 482 } while (0) 483 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 484 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 485 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 486 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 487 #else 488 #define STATS_INC_ACTIVE(x) do { } while (0) 489 #define STATS_DEC_ACTIVE(x) do { } while (0) 490 #define STATS_INC_ALLOCED(x) do { } while (0) 491 #define STATS_INC_GROWN(x) do { } while (0) 492 #define STATS_ADD_REAPED(x,y) do { } while (0) 493 #define STATS_SET_HIGH(x) do { } while (0) 494 #define STATS_INC_ERR(x) do { } while (0) 495 #define STATS_INC_NODEALLOCS(x) do { } while (0) 496 #define STATS_INC_NODEFREES(x) do { } while (0) 497 #define STATS_INC_ACOVERFLOW(x) do { } while (0) 498 #define STATS_SET_FREEABLE(x, i) do { } while (0) 499 #define STATS_INC_ALLOCHIT(x) do { } while (0) 500 #define STATS_INC_ALLOCMISS(x) do { } while (0) 501 #define STATS_INC_FREEHIT(x) do { } while (0) 502 #define STATS_INC_FREEMISS(x) do { } while (0) 503 #endif 504 505 #if DEBUG 506 507 /* 508 * memory layout of objects: 509 * 0 : objp 510 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 511 * the end of an object is aligned with the end of the real 512 * allocation. Catches writes behind the end of the allocation. 513 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 514 * redzone word. 515 * cachep->obj_offset: The real object. 516 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 517 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 518 * [BYTES_PER_WORD long] 519 */ 520 static int obj_offset(struct kmem_cache *cachep) 521 { 522 return cachep->obj_offset; 523 } 524 525 static int obj_size(struct kmem_cache *cachep) 526 { 527 return cachep->obj_size; 528 } 529 530 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 531 { 532 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 533 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); 534 } 535 536 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 537 { 538 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 539 if (cachep->flags & SLAB_STORE_USER) 540 return (unsigned long *)(objp + cachep->buffer_size - 541 2 * BYTES_PER_WORD); 542 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); 543 } 544 545 static void **dbg_userword(struct kmem_cache *cachep, void *objp) 546 { 547 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 548 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 549 } 550 551 #else 552 553 #define obj_offset(x) 0 554 #define obj_size(cachep) (cachep->buffer_size) 555 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) 556 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) 557 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 558 559 #endif 560 561 /* 562 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp 563 * order. 564 */ 565 #if defined(CONFIG_LARGE_ALLOCS) 566 #define MAX_OBJ_ORDER 13 /* up to 32Mb */ 567 #define MAX_GFP_ORDER 13 /* up to 32Mb */ 568 #elif defined(CONFIG_MMU) 569 #define MAX_OBJ_ORDER 5 /* 32 pages */ 570 #define MAX_GFP_ORDER 5 /* 32 pages */ 571 #else 572 #define MAX_OBJ_ORDER 8 /* up to 1Mb */ 573 #define MAX_GFP_ORDER 8 /* up to 1Mb */ 574 #endif 575 576 /* 577 * Do not go above this order unless 0 objects fit into the slab. 578 */ 579 #define BREAK_GFP_ORDER_HI 1 580 #define BREAK_GFP_ORDER_LO 0 581 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 582 583 /* 584 * Functions for storing/retrieving the cachep and or slab from the page 585 * allocator. These are used to find the slab an obj belongs to. With kfree(), 586 * these are used to find the cache which an obj belongs to. 587 */ 588 static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 589 { 590 page->lru.next = (struct list_head *)cache; 591 } 592 593 static inline struct kmem_cache *page_get_cache(struct page *page) 594 { 595 if (unlikely(PageCompound(page))) 596 page = (struct page *)page_private(page); 597 BUG_ON(!PageSlab(page)); 598 return (struct kmem_cache *)page->lru.next; 599 } 600 601 static inline void page_set_slab(struct page *page, struct slab *slab) 602 { 603 page->lru.prev = (struct list_head *)slab; 604 } 605 606 static inline struct slab *page_get_slab(struct page *page) 607 { 608 if (unlikely(PageCompound(page))) 609 page = (struct page *)page_private(page); 610 BUG_ON(!PageSlab(page)); 611 return (struct slab *)page->lru.prev; 612 } 613 614 static inline struct kmem_cache *virt_to_cache(const void *obj) 615 { 616 struct page *page = virt_to_page(obj); 617 return page_get_cache(page); 618 } 619 620 static inline struct slab *virt_to_slab(const void *obj) 621 { 622 struct page *page = virt_to_page(obj); 623 return page_get_slab(page); 624 } 625 626 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 627 unsigned int idx) 628 { 629 return slab->s_mem + cache->buffer_size * idx; 630 } 631 632 /* 633 * We want to avoid an expensive divide : (offset / cache->buffer_size) 634 * Using the fact that buffer_size is a constant for a particular cache, 635 * we can replace (offset / cache->buffer_size) by 636 * reciprocal_divide(offset, cache->reciprocal_buffer_size) 637 */ 638 static inline unsigned int obj_to_index(const struct kmem_cache *cache, 639 const struct slab *slab, void *obj) 640 { 641 u32 offset = (obj - slab->s_mem); 642 return reciprocal_divide(offset, cache->reciprocal_buffer_size); 643 } 644 645 /* 646 * These are the default caches for kmalloc. Custom caches can have other sizes. 647 */ 648 struct cache_sizes malloc_sizes[] = { 649 #define CACHE(x) { .cs_size = (x) }, 650 #include <linux/kmalloc_sizes.h> 651 CACHE(ULONG_MAX) 652 #undef CACHE 653 }; 654 EXPORT_SYMBOL(malloc_sizes); 655 656 /* Must match cache_sizes above. Out of line to keep cache footprint low. */ 657 struct cache_names { 658 char *name; 659 char *name_dma; 660 }; 661 662 static struct cache_names __initdata cache_names[] = { 663 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 664 #include <linux/kmalloc_sizes.h> 665 {NULL,} 666 #undef CACHE 667 }; 668 669 static struct arraycache_init initarray_cache __initdata = 670 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 671 static struct arraycache_init initarray_generic = 672 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 673 674 /* internal cache of cache description objs */ 675 static struct kmem_cache cache_cache = { 676 .batchcount = 1, 677 .limit = BOOT_CPUCACHE_ENTRIES, 678 .shared = 1, 679 .buffer_size = sizeof(struct kmem_cache), 680 .name = "kmem_cache", 681 #if DEBUG 682 .obj_size = sizeof(struct kmem_cache), 683 #endif 684 }; 685 686 #define BAD_ALIEN_MAGIC 0x01020304ul 687 688 #ifdef CONFIG_LOCKDEP 689 690 /* 691 * Slab sometimes uses the kmalloc slabs to store the slab headers 692 * for other slabs "off slab". 693 * The locking for this is tricky in that it nests within the locks 694 * of all other slabs in a few places; to deal with this special 695 * locking we put on-slab caches into a separate lock-class. 696 * 697 * We set lock class for alien array caches which are up during init. 698 * The lock annotation will be lost if all cpus of a node goes down and 699 * then comes back up during hotplug 700 */ 701 static struct lock_class_key on_slab_l3_key; 702 static struct lock_class_key on_slab_alc_key; 703 704 static inline void init_lock_keys(void) 705 706 { 707 int q; 708 struct cache_sizes *s = malloc_sizes; 709 710 while (s->cs_size != ULONG_MAX) { 711 for_each_node(q) { 712 struct array_cache **alc; 713 int r; 714 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; 715 if (!l3 || OFF_SLAB(s->cs_cachep)) 716 continue; 717 lockdep_set_class(&l3->list_lock, &on_slab_l3_key); 718 alc = l3->alien; 719 /* 720 * FIXME: This check for BAD_ALIEN_MAGIC 721 * should go away when common slab code is taught to 722 * work even without alien caches. 723 * Currently, non NUMA code returns BAD_ALIEN_MAGIC 724 * for alloc_alien_cache, 725 */ 726 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) 727 continue; 728 for_each_node(r) { 729 if (alc[r]) 730 lockdep_set_class(&alc[r]->lock, 731 &on_slab_alc_key); 732 } 733 } 734 s++; 735 } 736 } 737 #else 738 static inline void init_lock_keys(void) 739 { 740 } 741 #endif 742 743 /* 744 * 1. Guard access to the cache-chain. 745 * 2. Protect sanity of cpu_online_map against cpu hotplug events 746 */ 747 static DEFINE_MUTEX(cache_chain_mutex); 748 static struct list_head cache_chain; 749 750 /* 751 * chicken and egg problem: delay the per-cpu array allocation 752 * until the general caches are up. 753 */ 754 static enum { 755 NONE, 756 PARTIAL_AC, 757 PARTIAL_L3, 758 FULL 759 } g_cpucache_up; 760 761 /* 762 * used by boot code to determine if it can use slab based allocator 763 */ 764 int slab_is_available(void) 765 { 766 return g_cpucache_up == FULL; 767 } 768 769 static DEFINE_PER_CPU(struct delayed_work, reap_work); 770 771 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 772 { 773 return cachep->array[smp_processor_id()]; 774 } 775 776 static inline struct kmem_cache *__find_general_cachep(size_t size, 777 gfp_t gfpflags) 778 { 779 struct cache_sizes *csizep = malloc_sizes; 780 781 #if DEBUG 782 /* This happens if someone tries to call 783 * kmem_cache_create(), or __kmalloc(), before 784 * the generic caches are initialized. 785 */ 786 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 787 #endif 788 while (size > csizep->cs_size) 789 csizep++; 790 791 /* 792 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 793 * has cs_{dma,}cachep==NULL. Thus no special case 794 * for large kmalloc calls required. 795 */ 796 if (unlikely(gfpflags & GFP_DMA)) 797 return csizep->cs_dmacachep; 798 return csizep->cs_cachep; 799 } 800 801 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 802 { 803 return __find_general_cachep(size, gfpflags); 804 } 805 806 static size_t slab_mgmt_size(size_t nr_objs, size_t align) 807 { 808 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 809 } 810 811 /* 812 * Calculate the number of objects and left-over bytes for a given buffer size. 813 */ 814 static void cache_estimate(unsigned long gfporder, size_t buffer_size, 815 size_t align, int flags, size_t *left_over, 816 unsigned int *num) 817 { 818 int nr_objs; 819 size_t mgmt_size; 820 size_t slab_size = PAGE_SIZE << gfporder; 821 822 /* 823 * The slab management structure can be either off the slab or 824 * on it. For the latter case, the memory allocated for a 825 * slab is used for: 826 * 827 * - The struct slab 828 * - One kmem_bufctl_t for each object 829 * - Padding to respect alignment of @align 830 * - @buffer_size bytes for each object 831 * 832 * If the slab management structure is off the slab, then the 833 * alignment will already be calculated into the size. Because 834 * the slabs are all pages aligned, the objects will be at the 835 * correct alignment when allocated. 836 */ 837 if (flags & CFLGS_OFF_SLAB) { 838 mgmt_size = 0; 839 nr_objs = slab_size / buffer_size; 840 841 if (nr_objs > SLAB_LIMIT) 842 nr_objs = SLAB_LIMIT; 843 } else { 844 /* 845 * Ignore padding for the initial guess. The padding 846 * is at most @align-1 bytes, and @buffer_size is at 847 * least @align. In the worst case, this result will 848 * be one greater than the number of objects that fit 849 * into the memory allocation when taking the padding 850 * into account. 851 */ 852 nr_objs = (slab_size - sizeof(struct slab)) / 853 (buffer_size + sizeof(kmem_bufctl_t)); 854 855 /* 856 * This calculated number will be either the right 857 * amount, or one greater than what we want. 858 */ 859 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 860 > slab_size) 861 nr_objs--; 862 863 if (nr_objs > SLAB_LIMIT) 864 nr_objs = SLAB_LIMIT; 865 866 mgmt_size = slab_mgmt_size(nr_objs, align); 867 } 868 *num = nr_objs; 869 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 870 } 871 872 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) 873 874 static void __slab_error(const char *function, struct kmem_cache *cachep, 875 char *msg) 876 { 877 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 878 function, cachep->name, msg); 879 dump_stack(); 880 } 881 882 /* 883 * By default on NUMA we use alien caches to stage the freeing of 884 * objects allocated from other nodes. This causes massive memory 885 * inefficiencies when using fake NUMA setup to split memory into a 886 * large number of small nodes, so it can be disabled on the command 887 * line 888 */ 889 890 static int use_alien_caches __read_mostly = 1; 891 static int __init noaliencache_setup(char *s) 892 { 893 use_alien_caches = 0; 894 return 1; 895 } 896 __setup("noaliencache", noaliencache_setup); 897 898 #ifdef CONFIG_NUMA 899 /* 900 * Special reaping functions for NUMA systems called from cache_reap(). 901 * These take care of doing round robin flushing of alien caches (containing 902 * objects freed on different nodes from which they were allocated) and the 903 * flushing of remote pcps by calling drain_node_pages. 904 */ 905 static DEFINE_PER_CPU(unsigned long, reap_node); 906 907 static void init_reap_node(int cpu) 908 { 909 int node; 910 911 node = next_node(cpu_to_node(cpu), node_online_map); 912 if (node == MAX_NUMNODES) 913 node = first_node(node_online_map); 914 915 per_cpu(reap_node, cpu) = node; 916 } 917 918 static void next_reap_node(void) 919 { 920 int node = __get_cpu_var(reap_node); 921 922 /* 923 * Also drain per cpu pages on remote zones 924 */ 925 if (node != numa_node_id()) 926 drain_node_pages(node); 927 928 node = next_node(node, node_online_map); 929 if (unlikely(node >= MAX_NUMNODES)) 930 node = first_node(node_online_map); 931 __get_cpu_var(reap_node) = node; 932 } 933 934 #else 935 #define init_reap_node(cpu) do { } while (0) 936 #define next_reap_node(void) do { } while (0) 937 #endif 938 939 /* 940 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 941 * via the workqueue/eventd. 942 * Add the CPU number into the expiration time to minimize the possibility of 943 * the CPUs getting into lockstep and contending for the global cache chain 944 * lock. 945 */ 946 static void __devinit start_cpu_timer(int cpu) 947 { 948 struct delayed_work *reap_work = &per_cpu(reap_work, cpu); 949 950 /* 951 * When this gets called from do_initcalls via cpucache_init(), 952 * init_workqueues() has already run, so keventd will be setup 953 * at that time. 954 */ 955 if (keventd_up() && reap_work->work.func == NULL) { 956 init_reap_node(cpu); 957 INIT_DELAYED_WORK(reap_work, cache_reap); 958 schedule_delayed_work_on(cpu, reap_work, 959 __round_jiffies_relative(HZ, cpu)); 960 } 961 } 962 963 static struct array_cache *alloc_arraycache(int node, int entries, 964 int batchcount) 965 { 966 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 967 struct array_cache *nc = NULL; 968 969 nc = kmalloc_node(memsize, GFP_KERNEL, node); 970 if (nc) { 971 nc->avail = 0; 972 nc->limit = entries; 973 nc->batchcount = batchcount; 974 nc->touched = 0; 975 spin_lock_init(&nc->lock); 976 } 977 return nc; 978 } 979 980 /* 981 * Transfer objects in one arraycache to another. 982 * Locking must be handled by the caller. 983 * 984 * Return the number of entries transferred. 985 */ 986 static int transfer_objects(struct array_cache *to, 987 struct array_cache *from, unsigned int max) 988 { 989 /* Figure out how many entries to transfer */ 990 int nr = min(min(from->avail, max), to->limit - to->avail); 991 992 if (!nr) 993 return 0; 994 995 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 996 sizeof(void *) *nr); 997 998 from->avail -= nr; 999 to->avail += nr; 1000 to->touched = 1; 1001 return nr; 1002 } 1003 1004 #ifndef CONFIG_NUMA 1005 1006 #define drain_alien_cache(cachep, alien) do { } while (0) 1007 #define reap_alien(cachep, l3) do { } while (0) 1008 1009 static inline struct array_cache **alloc_alien_cache(int node, int limit) 1010 { 1011 return (struct array_cache **)BAD_ALIEN_MAGIC; 1012 } 1013 1014 static inline void free_alien_cache(struct array_cache **ac_ptr) 1015 { 1016 } 1017 1018 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1019 { 1020 return 0; 1021 } 1022 1023 static inline void *alternate_node_alloc(struct kmem_cache *cachep, 1024 gfp_t flags) 1025 { 1026 return NULL; 1027 } 1028 1029 static inline void *____cache_alloc_node(struct kmem_cache *cachep, 1030 gfp_t flags, int nodeid) 1031 { 1032 return NULL; 1033 } 1034 1035 #else /* CONFIG_NUMA */ 1036 1037 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 1038 static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 1039 1040 static struct array_cache **alloc_alien_cache(int node, int limit) 1041 { 1042 struct array_cache **ac_ptr; 1043 int memsize = sizeof(void *) * MAX_NUMNODES; 1044 int i; 1045 1046 if (limit > 1) 1047 limit = 12; 1048 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); 1049 if (ac_ptr) { 1050 for_each_node(i) { 1051 if (i == node || !node_online(i)) { 1052 ac_ptr[i] = NULL; 1053 continue; 1054 } 1055 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); 1056 if (!ac_ptr[i]) { 1057 for (i--; i <= 0; i--) 1058 kfree(ac_ptr[i]); 1059 kfree(ac_ptr); 1060 return NULL; 1061 } 1062 } 1063 } 1064 return ac_ptr; 1065 } 1066 1067 static void free_alien_cache(struct array_cache **ac_ptr) 1068 { 1069 int i; 1070 1071 if (!ac_ptr) 1072 return; 1073 for_each_node(i) 1074 kfree(ac_ptr[i]); 1075 kfree(ac_ptr); 1076 } 1077 1078 static void __drain_alien_cache(struct kmem_cache *cachep, 1079 struct array_cache *ac, int node) 1080 { 1081 struct kmem_list3 *rl3 = cachep->nodelists[node]; 1082 1083 if (ac->avail) { 1084 spin_lock(&rl3->list_lock); 1085 /* 1086 * Stuff objects into the remote nodes shared array first. 1087 * That way we could avoid the overhead of putting the objects 1088 * into the free lists and getting them back later. 1089 */ 1090 if (rl3->shared) 1091 transfer_objects(rl3->shared, ac, ac->limit); 1092 1093 free_block(cachep, ac->entry, ac->avail, node); 1094 ac->avail = 0; 1095 spin_unlock(&rl3->list_lock); 1096 } 1097 } 1098 1099 /* 1100 * Called from cache_reap() to regularly drain alien caches round robin. 1101 */ 1102 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 1103 { 1104 int node = __get_cpu_var(reap_node); 1105 1106 if (l3->alien) { 1107 struct array_cache *ac = l3->alien[node]; 1108 1109 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1110 __drain_alien_cache(cachep, ac, node); 1111 spin_unlock_irq(&ac->lock); 1112 } 1113 } 1114 } 1115 1116 static void drain_alien_cache(struct kmem_cache *cachep, 1117 struct array_cache **alien) 1118 { 1119 int i = 0; 1120 struct array_cache *ac; 1121 unsigned long flags; 1122 1123 for_each_online_node(i) { 1124 ac = alien[i]; 1125 if (ac) { 1126 spin_lock_irqsave(&ac->lock, flags); 1127 __drain_alien_cache(cachep, ac, i); 1128 spin_unlock_irqrestore(&ac->lock, flags); 1129 } 1130 } 1131 } 1132 1133 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1134 { 1135 struct slab *slabp = virt_to_slab(objp); 1136 int nodeid = slabp->nodeid; 1137 struct kmem_list3 *l3; 1138 struct array_cache *alien = NULL; 1139 int node; 1140 1141 node = numa_node_id(); 1142 1143 /* 1144 * Make sure we are not freeing a object from another node to the array 1145 * cache on this cpu. 1146 */ 1147 if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches)) 1148 return 0; 1149 1150 l3 = cachep->nodelists[node]; 1151 STATS_INC_NODEFREES(cachep); 1152 if (l3->alien && l3->alien[nodeid]) { 1153 alien = l3->alien[nodeid]; 1154 spin_lock(&alien->lock); 1155 if (unlikely(alien->avail == alien->limit)) { 1156 STATS_INC_ACOVERFLOW(cachep); 1157 __drain_alien_cache(cachep, alien, nodeid); 1158 } 1159 alien->entry[alien->avail++] = objp; 1160 spin_unlock(&alien->lock); 1161 } else { 1162 spin_lock(&(cachep->nodelists[nodeid])->list_lock); 1163 free_block(cachep, &objp, 1, nodeid); 1164 spin_unlock(&(cachep->nodelists[nodeid])->list_lock); 1165 } 1166 return 1; 1167 } 1168 #endif 1169 1170 static int __cpuinit cpuup_callback(struct notifier_block *nfb, 1171 unsigned long action, void *hcpu) 1172 { 1173 long cpu = (long)hcpu; 1174 struct kmem_cache *cachep; 1175 struct kmem_list3 *l3 = NULL; 1176 int node = cpu_to_node(cpu); 1177 int memsize = sizeof(struct kmem_list3); 1178 1179 switch (action) { 1180 case CPU_UP_PREPARE: 1181 mutex_lock(&cache_chain_mutex); 1182 /* 1183 * We need to do this right in the beginning since 1184 * alloc_arraycache's are going to use this list. 1185 * kmalloc_node allows us to add the slab to the right 1186 * kmem_list3 and not this cpu's kmem_list3 1187 */ 1188 1189 list_for_each_entry(cachep, &cache_chain, next) { 1190 /* 1191 * Set up the size64 kmemlist for cpu before we can 1192 * begin anything. Make sure some other cpu on this 1193 * node has not already allocated this 1194 */ 1195 if (!cachep->nodelists[node]) { 1196 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1197 if (!l3) 1198 goto bad; 1199 kmem_list3_init(l3); 1200 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1201 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1202 1203 /* 1204 * The l3s don't come and go as CPUs come and 1205 * go. cache_chain_mutex is sufficient 1206 * protection here. 1207 */ 1208 cachep->nodelists[node] = l3; 1209 } 1210 1211 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1212 cachep->nodelists[node]->free_limit = 1213 (1 + nr_cpus_node(node)) * 1214 cachep->batchcount + cachep->num; 1215 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1216 } 1217 1218 /* 1219 * Now we can go ahead with allocating the shared arrays and 1220 * array caches 1221 */ 1222 list_for_each_entry(cachep, &cache_chain, next) { 1223 struct array_cache *nc; 1224 struct array_cache *shared; 1225 struct array_cache **alien = NULL; 1226 1227 nc = alloc_arraycache(node, cachep->limit, 1228 cachep->batchcount); 1229 if (!nc) 1230 goto bad; 1231 shared = alloc_arraycache(node, 1232 cachep->shared * cachep->batchcount, 1233 0xbaadf00d); 1234 if (!shared) 1235 goto bad; 1236 1237 if (use_alien_caches) { 1238 alien = alloc_alien_cache(node, cachep->limit); 1239 if (!alien) 1240 goto bad; 1241 } 1242 cachep->array[cpu] = nc; 1243 l3 = cachep->nodelists[node]; 1244 BUG_ON(!l3); 1245 1246 spin_lock_irq(&l3->list_lock); 1247 if (!l3->shared) { 1248 /* 1249 * We are serialised from CPU_DEAD or 1250 * CPU_UP_CANCELLED by the cpucontrol lock 1251 */ 1252 l3->shared = shared; 1253 shared = NULL; 1254 } 1255 #ifdef CONFIG_NUMA 1256 if (!l3->alien) { 1257 l3->alien = alien; 1258 alien = NULL; 1259 } 1260 #endif 1261 spin_unlock_irq(&l3->list_lock); 1262 kfree(shared); 1263 free_alien_cache(alien); 1264 } 1265 break; 1266 case CPU_ONLINE: 1267 mutex_unlock(&cache_chain_mutex); 1268 start_cpu_timer(cpu); 1269 break; 1270 #ifdef CONFIG_HOTPLUG_CPU 1271 case CPU_DOWN_PREPARE: 1272 mutex_lock(&cache_chain_mutex); 1273 break; 1274 case CPU_DOWN_FAILED: 1275 mutex_unlock(&cache_chain_mutex); 1276 break; 1277 case CPU_DEAD: 1278 /* 1279 * Even if all the cpus of a node are down, we don't free the 1280 * kmem_list3 of any cache. This to avoid a race between 1281 * cpu_down, and a kmalloc allocation from another cpu for 1282 * memory from the node of the cpu going down. The list3 1283 * structure is usually allocated from kmem_cache_create() and 1284 * gets destroyed at kmem_cache_destroy(). 1285 */ 1286 /* fall thru */ 1287 #endif 1288 case CPU_UP_CANCELED: 1289 list_for_each_entry(cachep, &cache_chain, next) { 1290 struct array_cache *nc; 1291 struct array_cache *shared; 1292 struct array_cache **alien; 1293 cpumask_t mask; 1294 1295 mask = node_to_cpumask(node); 1296 /* cpu is dead; no one can alloc from it. */ 1297 nc = cachep->array[cpu]; 1298 cachep->array[cpu] = NULL; 1299 l3 = cachep->nodelists[node]; 1300 1301 if (!l3) 1302 goto free_array_cache; 1303 1304 spin_lock_irq(&l3->list_lock); 1305 1306 /* Free limit for this kmem_list3 */ 1307 l3->free_limit -= cachep->batchcount; 1308 if (nc) 1309 free_block(cachep, nc->entry, nc->avail, node); 1310 1311 if (!cpus_empty(mask)) { 1312 spin_unlock_irq(&l3->list_lock); 1313 goto free_array_cache; 1314 } 1315 1316 shared = l3->shared; 1317 if (shared) { 1318 free_block(cachep, l3->shared->entry, 1319 l3->shared->avail, node); 1320 l3->shared = NULL; 1321 } 1322 1323 alien = l3->alien; 1324 l3->alien = NULL; 1325 1326 spin_unlock_irq(&l3->list_lock); 1327 1328 kfree(shared); 1329 if (alien) { 1330 drain_alien_cache(cachep, alien); 1331 free_alien_cache(alien); 1332 } 1333 free_array_cache: 1334 kfree(nc); 1335 } 1336 /* 1337 * In the previous loop, all the objects were freed to 1338 * the respective cache's slabs, now we can go ahead and 1339 * shrink each nodelist to its limit. 1340 */ 1341 list_for_each_entry(cachep, &cache_chain, next) { 1342 l3 = cachep->nodelists[node]; 1343 if (!l3) 1344 continue; 1345 drain_freelist(cachep, l3, l3->free_objects); 1346 } 1347 mutex_unlock(&cache_chain_mutex); 1348 break; 1349 } 1350 return NOTIFY_OK; 1351 bad: 1352 return NOTIFY_BAD; 1353 } 1354 1355 static struct notifier_block __cpuinitdata cpucache_notifier = { 1356 &cpuup_callback, NULL, 0 1357 }; 1358 1359 /* 1360 * swap the static kmem_list3 with kmalloced memory 1361 */ 1362 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1363 int nodeid) 1364 { 1365 struct kmem_list3 *ptr; 1366 1367 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); 1368 BUG_ON(!ptr); 1369 1370 local_irq_disable(); 1371 memcpy(ptr, list, sizeof(struct kmem_list3)); 1372 /* 1373 * Do not assume that spinlocks can be initialized via memcpy: 1374 */ 1375 spin_lock_init(&ptr->list_lock); 1376 1377 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1378 cachep->nodelists[nodeid] = ptr; 1379 local_irq_enable(); 1380 } 1381 1382 /* 1383 * Initialisation. Called after the page allocator have been initialised and 1384 * before smp_init(). 1385 */ 1386 void __init kmem_cache_init(void) 1387 { 1388 size_t left_over; 1389 struct cache_sizes *sizes; 1390 struct cache_names *names; 1391 int i; 1392 int order; 1393 int node; 1394 1395 for (i = 0; i < NUM_INIT_LISTS; i++) { 1396 kmem_list3_init(&initkmem_list3[i]); 1397 if (i < MAX_NUMNODES) 1398 cache_cache.nodelists[i] = NULL; 1399 } 1400 1401 /* 1402 * Fragmentation resistance on low memory - only use bigger 1403 * page orders on machines with more than 32MB of memory. 1404 */ 1405 if (num_physpages > (32 << 20) >> PAGE_SHIFT) 1406 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1407 1408 /* Bootstrap is tricky, because several objects are allocated 1409 * from caches that do not exist yet: 1410 * 1) initialize the cache_cache cache: it contains the struct 1411 * kmem_cache structures of all caches, except cache_cache itself: 1412 * cache_cache is statically allocated. 1413 * Initially an __init data area is used for the head array and the 1414 * kmem_list3 structures, it's replaced with a kmalloc allocated 1415 * array at the end of the bootstrap. 1416 * 2) Create the first kmalloc cache. 1417 * The struct kmem_cache for the new cache is allocated normally. 1418 * An __init data area is used for the head array. 1419 * 3) Create the remaining kmalloc caches, with minimally sized 1420 * head arrays. 1421 * 4) Replace the __init data head arrays for cache_cache and the first 1422 * kmalloc cache with kmalloc allocated arrays. 1423 * 5) Replace the __init data for kmem_list3 for cache_cache and 1424 * the other cache's with kmalloc allocated memory. 1425 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1426 */ 1427 1428 node = numa_node_id(); 1429 1430 /* 1) create the cache_cache */ 1431 INIT_LIST_HEAD(&cache_chain); 1432 list_add(&cache_cache.next, &cache_chain); 1433 cache_cache.colour_off = cache_line_size(); 1434 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1435 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; 1436 1437 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1438 cache_line_size()); 1439 cache_cache.reciprocal_buffer_size = 1440 reciprocal_value(cache_cache.buffer_size); 1441 1442 for (order = 0; order < MAX_ORDER; order++) { 1443 cache_estimate(order, cache_cache.buffer_size, 1444 cache_line_size(), 0, &left_over, &cache_cache.num); 1445 if (cache_cache.num) 1446 break; 1447 } 1448 BUG_ON(!cache_cache.num); 1449 cache_cache.gfporder = order; 1450 cache_cache.colour = left_over / cache_cache.colour_off; 1451 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1452 sizeof(struct slab), cache_line_size()); 1453 1454 /* 2+3) create the kmalloc caches */ 1455 sizes = malloc_sizes; 1456 names = cache_names; 1457 1458 /* 1459 * Initialize the caches that provide memory for the array cache and the 1460 * kmem_list3 structures first. Without this, further allocations will 1461 * bug. 1462 */ 1463 1464 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1465 sizes[INDEX_AC].cs_size, 1466 ARCH_KMALLOC_MINALIGN, 1467 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1468 NULL, NULL); 1469 1470 if (INDEX_AC != INDEX_L3) { 1471 sizes[INDEX_L3].cs_cachep = 1472 kmem_cache_create(names[INDEX_L3].name, 1473 sizes[INDEX_L3].cs_size, 1474 ARCH_KMALLOC_MINALIGN, 1475 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1476 NULL, NULL); 1477 } 1478 1479 slab_early_init = 0; 1480 1481 while (sizes->cs_size != ULONG_MAX) { 1482 /* 1483 * For performance, all the general caches are L1 aligned. 1484 * This should be particularly beneficial on SMP boxes, as it 1485 * eliminates "false sharing". 1486 * Note for systems short on memory removing the alignment will 1487 * allow tighter packing of the smaller caches. 1488 */ 1489 if (!sizes->cs_cachep) { 1490 sizes->cs_cachep = kmem_cache_create(names->name, 1491 sizes->cs_size, 1492 ARCH_KMALLOC_MINALIGN, 1493 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1494 NULL, NULL); 1495 } 1496 1497 sizes->cs_dmacachep = kmem_cache_create(names->name_dma, 1498 sizes->cs_size, 1499 ARCH_KMALLOC_MINALIGN, 1500 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1501 SLAB_PANIC, 1502 NULL, NULL); 1503 sizes++; 1504 names++; 1505 } 1506 /* 4) Replace the bootstrap head arrays */ 1507 { 1508 struct array_cache *ptr; 1509 1510 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1511 1512 local_irq_disable(); 1513 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1514 memcpy(ptr, cpu_cache_get(&cache_cache), 1515 sizeof(struct arraycache_init)); 1516 /* 1517 * Do not assume that spinlocks can be initialized via memcpy: 1518 */ 1519 spin_lock_init(&ptr->lock); 1520 1521 cache_cache.array[smp_processor_id()] = ptr; 1522 local_irq_enable(); 1523 1524 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1525 1526 local_irq_disable(); 1527 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1528 != &initarray_generic.cache); 1529 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1530 sizeof(struct arraycache_init)); 1531 /* 1532 * Do not assume that spinlocks can be initialized via memcpy: 1533 */ 1534 spin_lock_init(&ptr->lock); 1535 1536 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1537 ptr; 1538 local_irq_enable(); 1539 } 1540 /* 5) Replace the bootstrap kmem_list3's */ 1541 { 1542 int nid; 1543 1544 /* Replace the static kmem_list3 structures for the boot cpu */ 1545 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); 1546 1547 for_each_online_node(nid) { 1548 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1549 &initkmem_list3[SIZE_AC + nid], nid); 1550 1551 if (INDEX_AC != INDEX_L3) { 1552 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1553 &initkmem_list3[SIZE_L3 + nid], nid); 1554 } 1555 } 1556 } 1557 1558 /* 6) resize the head arrays to their final sizes */ 1559 { 1560 struct kmem_cache *cachep; 1561 mutex_lock(&cache_chain_mutex); 1562 list_for_each_entry(cachep, &cache_chain, next) 1563 if (enable_cpucache(cachep)) 1564 BUG(); 1565 mutex_unlock(&cache_chain_mutex); 1566 } 1567 1568 /* Annotate slab for lockdep -- annotate the malloc caches */ 1569 init_lock_keys(); 1570 1571 1572 /* Done! */ 1573 g_cpucache_up = FULL; 1574 1575 /* 1576 * Register a cpu startup notifier callback that initializes 1577 * cpu_cache_get for all new cpus 1578 */ 1579 register_cpu_notifier(&cpucache_notifier); 1580 1581 /* 1582 * The reap timers are started later, with a module init call: That part 1583 * of the kernel is not yet operational. 1584 */ 1585 } 1586 1587 static int __init cpucache_init(void) 1588 { 1589 int cpu; 1590 1591 /* 1592 * Register the timers that return unneeded pages to the page allocator 1593 */ 1594 for_each_online_cpu(cpu) 1595 start_cpu_timer(cpu); 1596 return 0; 1597 } 1598 __initcall(cpucache_init); 1599 1600 /* 1601 * Interface to system's page allocator. No need to hold the cache-lock. 1602 * 1603 * If we requested dmaable memory, we will get it. Even if we 1604 * did not request dmaable memory, we might get it, but that 1605 * would be relatively rare and ignorable. 1606 */ 1607 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1608 { 1609 struct page *page; 1610 int nr_pages; 1611 int i; 1612 1613 #ifndef CONFIG_MMU 1614 /* 1615 * Nommu uses slab's for process anonymous memory allocations, and thus 1616 * requires __GFP_COMP to properly refcount higher order allocations 1617 */ 1618 flags |= __GFP_COMP; 1619 #endif 1620 1621 flags |= cachep->gfpflags; 1622 1623 page = alloc_pages_node(nodeid, flags, cachep->gfporder); 1624 if (!page) 1625 return NULL; 1626 1627 nr_pages = (1 << cachep->gfporder); 1628 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1629 add_zone_page_state(page_zone(page), 1630 NR_SLAB_RECLAIMABLE, nr_pages); 1631 else 1632 add_zone_page_state(page_zone(page), 1633 NR_SLAB_UNRECLAIMABLE, nr_pages); 1634 for (i = 0; i < nr_pages; i++) 1635 __SetPageSlab(page + i); 1636 return page_address(page); 1637 } 1638 1639 /* 1640 * Interface to system's page release. 1641 */ 1642 static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1643 { 1644 unsigned long i = (1 << cachep->gfporder); 1645 struct page *page = virt_to_page(addr); 1646 const unsigned long nr_freed = i; 1647 1648 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1649 sub_zone_page_state(page_zone(page), 1650 NR_SLAB_RECLAIMABLE, nr_freed); 1651 else 1652 sub_zone_page_state(page_zone(page), 1653 NR_SLAB_UNRECLAIMABLE, nr_freed); 1654 while (i--) { 1655 BUG_ON(!PageSlab(page)); 1656 __ClearPageSlab(page); 1657 page++; 1658 } 1659 if (current->reclaim_state) 1660 current->reclaim_state->reclaimed_slab += nr_freed; 1661 free_pages((unsigned long)addr, cachep->gfporder); 1662 } 1663 1664 static void kmem_rcu_free(struct rcu_head *head) 1665 { 1666 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1667 struct kmem_cache *cachep = slab_rcu->cachep; 1668 1669 kmem_freepages(cachep, slab_rcu->addr); 1670 if (OFF_SLAB(cachep)) 1671 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1672 } 1673 1674 #if DEBUG 1675 1676 #ifdef CONFIG_DEBUG_PAGEALLOC 1677 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1678 unsigned long caller) 1679 { 1680 int size = obj_size(cachep); 1681 1682 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1683 1684 if (size < 5 * sizeof(unsigned long)) 1685 return; 1686 1687 *addr++ = 0x12345678; 1688 *addr++ = caller; 1689 *addr++ = smp_processor_id(); 1690 size -= 3 * sizeof(unsigned long); 1691 { 1692 unsigned long *sptr = &caller; 1693 unsigned long svalue; 1694 1695 while (!kstack_end(sptr)) { 1696 svalue = *sptr++; 1697 if (kernel_text_address(svalue)) { 1698 *addr++ = svalue; 1699 size -= sizeof(unsigned long); 1700 if (size <= sizeof(unsigned long)) 1701 break; 1702 } 1703 } 1704 1705 } 1706 *addr++ = 0x87654321; 1707 } 1708 #endif 1709 1710 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1711 { 1712 int size = obj_size(cachep); 1713 addr = &((char *)addr)[obj_offset(cachep)]; 1714 1715 memset(addr, val, size); 1716 *(unsigned char *)(addr + size - 1) = POISON_END; 1717 } 1718 1719 static void dump_line(char *data, int offset, int limit) 1720 { 1721 int i; 1722 unsigned char error = 0; 1723 int bad_count = 0; 1724 1725 printk(KERN_ERR "%03x:", offset); 1726 for (i = 0; i < limit; i++) { 1727 if (data[offset + i] != POISON_FREE) { 1728 error = data[offset + i]; 1729 bad_count++; 1730 } 1731 printk(" %02x", (unsigned char)data[offset + i]); 1732 } 1733 printk("\n"); 1734 1735 if (bad_count == 1) { 1736 error ^= POISON_FREE; 1737 if (!(error & (error - 1))) { 1738 printk(KERN_ERR "Single bit error detected. Probably " 1739 "bad RAM.\n"); 1740 #ifdef CONFIG_X86 1741 printk(KERN_ERR "Run memtest86+ or a similar memory " 1742 "test tool.\n"); 1743 #else 1744 printk(KERN_ERR "Run a memory test tool.\n"); 1745 #endif 1746 } 1747 } 1748 } 1749 #endif 1750 1751 #if DEBUG 1752 1753 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1754 { 1755 int i, size; 1756 char *realobj; 1757 1758 if (cachep->flags & SLAB_RED_ZONE) { 1759 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", 1760 *dbg_redzone1(cachep, objp), 1761 *dbg_redzone2(cachep, objp)); 1762 } 1763 1764 if (cachep->flags & SLAB_STORE_USER) { 1765 printk(KERN_ERR "Last user: [<%p>]", 1766 *dbg_userword(cachep, objp)); 1767 print_symbol("(%s)", 1768 (unsigned long)*dbg_userword(cachep, objp)); 1769 printk("\n"); 1770 } 1771 realobj = (char *)objp + obj_offset(cachep); 1772 size = obj_size(cachep); 1773 for (i = 0; i < size && lines; i += 16, lines--) { 1774 int limit; 1775 limit = 16; 1776 if (i + limit > size) 1777 limit = size - i; 1778 dump_line(realobj, i, limit); 1779 } 1780 } 1781 1782 static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1783 { 1784 char *realobj; 1785 int size, i; 1786 int lines = 0; 1787 1788 realobj = (char *)objp + obj_offset(cachep); 1789 size = obj_size(cachep); 1790 1791 for (i = 0; i < size; i++) { 1792 char exp = POISON_FREE; 1793 if (i == size - 1) 1794 exp = POISON_END; 1795 if (realobj[i] != exp) { 1796 int limit; 1797 /* Mismatch ! */ 1798 /* Print header */ 1799 if (lines == 0) { 1800 printk(KERN_ERR 1801 "Slab corruption: start=%p, len=%d\n", 1802 realobj, size); 1803 print_objinfo(cachep, objp, 0); 1804 } 1805 /* Hexdump the affected line */ 1806 i = (i / 16) * 16; 1807 limit = 16; 1808 if (i + limit > size) 1809 limit = size - i; 1810 dump_line(realobj, i, limit); 1811 i += 16; 1812 lines++; 1813 /* Limit to 5 lines */ 1814 if (lines > 5) 1815 break; 1816 } 1817 } 1818 if (lines != 0) { 1819 /* Print some data about the neighboring objects, if they 1820 * exist: 1821 */ 1822 struct slab *slabp = virt_to_slab(objp); 1823 unsigned int objnr; 1824 1825 objnr = obj_to_index(cachep, slabp, objp); 1826 if (objnr) { 1827 objp = index_to_obj(cachep, slabp, objnr - 1); 1828 realobj = (char *)objp + obj_offset(cachep); 1829 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1830 realobj, size); 1831 print_objinfo(cachep, objp, 2); 1832 } 1833 if (objnr + 1 < cachep->num) { 1834 objp = index_to_obj(cachep, slabp, objnr + 1); 1835 realobj = (char *)objp + obj_offset(cachep); 1836 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1837 realobj, size); 1838 print_objinfo(cachep, objp, 2); 1839 } 1840 } 1841 } 1842 #endif 1843 1844 #if DEBUG 1845 /** 1846 * slab_destroy_objs - destroy a slab and its objects 1847 * @cachep: cache pointer being destroyed 1848 * @slabp: slab pointer being destroyed 1849 * 1850 * Call the registered destructor for each object in a slab that is being 1851 * destroyed. 1852 */ 1853 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1854 { 1855 int i; 1856 for (i = 0; i < cachep->num; i++) { 1857 void *objp = index_to_obj(cachep, slabp, i); 1858 1859 if (cachep->flags & SLAB_POISON) { 1860 #ifdef CONFIG_DEBUG_PAGEALLOC 1861 if (cachep->buffer_size % PAGE_SIZE == 0 && 1862 OFF_SLAB(cachep)) 1863 kernel_map_pages(virt_to_page(objp), 1864 cachep->buffer_size / PAGE_SIZE, 1); 1865 else 1866 check_poison_obj(cachep, objp); 1867 #else 1868 check_poison_obj(cachep, objp); 1869 #endif 1870 } 1871 if (cachep->flags & SLAB_RED_ZONE) { 1872 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1873 slab_error(cachep, "start of a freed object " 1874 "was overwritten"); 1875 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1876 slab_error(cachep, "end of a freed object " 1877 "was overwritten"); 1878 } 1879 if (cachep->dtor && !(cachep->flags & SLAB_POISON)) 1880 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); 1881 } 1882 } 1883 #else 1884 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1885 { 1886 if (cachep->dtor) { 1887 int i; 1888 for (i = 0; i < cachep->num; i++) { 1889 void *objp = index_to_obj(cachep, slabp, i); 1890 (cachep->dtor) (objp, cachep, 0); 1891 } 1892 } 1893 } 1894 #endif 1895 1896 /** 1897 * slab_destroy - destroy and release all objects in a slab 1898 * @cachep: cache pointer being destroyed 1899 * @slabp: slab pointer being destroyed 1900 * 1901 * Destroy all the objs in a slab, and release the mem back to the system. 1902 * Before calling the slab must have been unlinked from the cache. The 1903 * cache-lock is not held/needed. 1904 */ 1905 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1906 { 1907 void *addr = slabp->s_mem - slabp->colouroff; 1908 1909 slab_destroy_objs(cachep, slabp); 1910 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1911 struct slab_rcu *slab_rcu; 1912 1913 slab_rcu = (struct slab_rcu *)slabp; 1914 slab_rcu->cachep = cachep; 1915 slab_rcu->addr = addr; 1916 call_rcu(&slab_rcu->head, kmem_rcu_free); 1917 } else { 1918 kmem_freepages(cachep, addr); 1919 if (OFF_SLAB(cachep)) 1920 kmem_cache_free(cachep->slabp_cache, slabp); 1921 } 1922 } 1923 1924 /* 1925 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1926 * size of kmem_list3. 1927 */ 1928 static void set_up_list3s(struct kmem_cache *cachep, int index) 1929 { 1930 int node; 1931 1932 for_each_online_node(node) { 1933 cachep->nodelists[node] = &initkmem_list3[index + node]; 1934 cachep->nodelists[node]->next_reap = jiffies + 1935 REAPTIMEOUT_LIST3 + 1936 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1937 } 1938 } 1939 1940 static void __kmem_cache_destroy(struct kmem_cache *cachep) 1941 { 1942 int i; 1943 struct kmem_list3 *l3; 1944 1945 for_each_online_cpu(i) 1946 kfree(cachep->array[i]); 1947 1948 /* NUMA: free the list3 structures */ 1949 for_each_online_node(i) { 1950 l3 = cachep->nodelists[i]; 1951 if (l3) { 1952 kfree(l3->shared); 1953 free_alien_cache(l3->alien); 1954 kfree(l3); 1955 } 1956 } 1957 kmem_cache_free(&cache_cache, cachep); 1958 } 1959 1960 1961 /** 1962 * calculate_slab_order - calculate size (page order) of slabs 1963 * @cachep: pointer to the cache that is being created 1964 * @size: size of objects to be created in this cache. 1965 * @align: required alignment for the objects. 1966 * @flags: slab allocation flags 1967 * 1968 * Also calculates the number of objects per slab. 1969 * 1970 * This could be made much more intelligent. For now, try to avoid using 1971 * high order pages for slabs. When the gfp() functions are more friendly 1972 * towards high-order requests, this should be changed. 1973 */ 1974 static size_t calculate_slab_order(struct kmem_cache *cachep, 1975 size_t size, size_t align, unsigned long flags) 1976 { 1977 unsigned long offslab_limit; 1978 size_t left_over = 0; 1979 int gfporder; 1980 1981 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { 1982 unsigned int num; 1983 size_t remainder; 1984 1985 cache_estimate(gfporder, size, align, flags, &remainder, &num); 1986 if (!num) 1987 continue; 1988 1989 if (flags & CFLGS_OFF_SLAB) { 1990 /* 1991 * Max number of objs-per-slab for caches which 1992 * use off-slab slabs. Needed to avoid a possible 1993 * looping condition in cache_grow(). 1994 */ 1995 offslab_limit = size - sizeof(struct slab); 1996 offslab_limit /= sizeof(kmem_bufctl_t); 1997 1998 if (num > offslab_limit) 1999 break; 2000 } 2001 2002 /* Found something acceptable - save it away */ 2003 cachep->num = num; 2004 cachep->gfporder = gfporder; 2005 left_over = remainder; 2006 2007 /* 2008 * A VFS-reclaimable slab tends to have most allocations 2009 * as GFP_NOFS and we really don't want to have to be allocating 2010 * higher-order pages when we are unable to shrink dcache. 2011 */ 2012 if (flags & SLAB_RECLAIM_ACCOUNT) 2013 break; 2014 2015 /* 2016 * Large number of objects is good, but very large slabs are 2017 * currently bad for the gfp()s. 2018 */ 2019 if (gfporder >= slab_break_gfp_order) 2020 break; 2021 2022 /* 2023 * Acceptable internal fragmentation? 2024 */ 2025 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 2026 break; 2027 } 2028 return left_over; 2029 } 2030 2031 static int setup_cpu_cache(struct kmem_cache *cachep) 2032 { 2033 if (g_cpucache_up == FULL) 2034 return enable_cpucache(cachep); 2035 2036 if (g_cpucache_up == NONE) { 2037 /* 2038 * Note: the first kmem_cache_create must create the cache 2039 * that's used by kmalloc(24), otherwise the creation of 2040 * further caches will BUG(). 2041 */ 2042 cachep->array[smp_processor_id()] = &initarray_generic.cache; 2043 2044 /* 2045 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 2046 * the first cache, then we need to set up all its list3s, 2047 * otherwise the creation of further caches will BUG(). 2048 */ 2049 set_up_list3s(cachep, SIZE_AC); 2050 if (INDEX_AC == INDEX_L3) 2051 g_cpucache_up = PARTIAL_L3; 2052 else 2053 g_cpucache_up = PARTIAL_AC; 2054 } else { 2055 cachep->array[smp_processor_id()] = 2056 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 2057 2058 if (g_cpucache_up == PARTIAL_AC) { 2059 set_up_list3s(cachep, SIZE_L3); 2060 g_cpucache_up = PARTIAL_L3; 2061 } else { 2062 int node; 2063 for_each_online_node(node) { 2064 cachep->nodelists[node] = 2065 kmalloc_node(sizeof(struct kmem_list3), 2066 GFP_KERNEL, node); 2067 BUG_ON(!cachep->nodelists[node]); 2068 kmem_list3_init(cachep->nodelists[node]); 2069 } 2070 } 2071 } 2072 cachep->nodelists[numa_node_id()]->next_reap = 2073 jiffies + REAPTIMEOUT_LIST3 + 2074 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 2075 2076 cpu_cache_get(cachep)->avail = 0; 2077 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 2078 cpu_cache_get(cachep)->batchcount = 1; 2079 cpu_cache_get(cachep)->touched = 0; 2080 cachep->batchcount = 1; 2081 cachep->limit = BOOT_CPUCACHE_ENTRIES; 2082 return 0; 2083 } 2084 2085 /** 2086 * kmem_cache_create - Create a cache. 2087 * @name: A string which is used in /proc/slabinfo to identify this cache. 2088 * @size: The size of objects to be created in this cache. 2089 * @align: The required alignment for the objects. 2090 * @flags: SLAB flags 2091 * @ctor: A constructor for the objects. 2092 * @dtor: A destructor for the objects. 2093 * 2094 * Returns a ptr to the cache on success, NULL on failure. 2095 * Cannot be called within a int, but can be interrupted. 2096 * The @ctor is run when new pages are allocated by the cache 2097 * and the @dtor is run before the pages are handed back. 2098 * 2099 * @name must be valid until the cache is destroyed. This implies that 2100 * the module calling this has to destroy the cache before getting unloaded. 2101 * 2102 * The flags are 2103 * 2104 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 2105 * to catch references to uninitialised memory. 2106 * 2107 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 2108 * for buffer overruns. 2109 * 2110 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 2111 * cacheline. This can be beneficial if you're counting cycles as closely 2112 * as davem. 2113 */ 2114 struct kmem_cache * 2115 kmem_cache_create (const char *name, size_t size, size_t align, 2116 unsigned long flags, 2117 void (*ctor)(void*, struct kmem_cache *, unsigned long), 2118 void (*dtor)(void*, struct kmem_cache *, unsigned long)) 2119 { 2120 size_t left_over, slab_size, ralign; 2121 struct kmem_cache *cachep = NULL, *pc; 2122 2123 /* 2124 * Sanity checks... these are all serious usage bugs. 2125 */ 2126 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 2127 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { 2128 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, 2129 name); 2130 BUG(); 2131 } 2132 2133 /* 2134 * We use cache_chain_mutex to ensure a consistent view of 2135 * cpu_online_map as well. Please see cpuup_callback 2136 */ 2137 mutex_lock(&cache_chain_mutex); 2138 2139 list_for_each_entry(pc, &cache_chain, next) { 2140 char tmp; 2141 int res; 2142 2143 /* 2144 * This happens when the module gets unloaded and doesn't 2145 * destroy its slab cache and no-one else reuses the vmalloc 2146 * area of the module. Print a warning. 2147 */ 2148 res = probe_kernel_address(pc->name, tmp); 2149 if (res) { 2150 printk("SLAB: cache with size %d has lost its name\n", 2151 pc->buffer_size); 2152 continue; 2153 } 2154 2155 if (!strcmp(pc->name, name)) { 2156 printk("kmem_cache_create: duplicate cache %s\n", name); 2157 dump_stack(); 2158 goto oops; 2159 } 2160 } 2161 2162 #if DEBUG 2163 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 2164 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { 2165 /* No constructor, but inital state check requested */ 2166 printk(KERN_ERR "%s: No con, but init state check " 2167 "requested - %s\n", __FUNCTION__, name); 2168 flags &= ~SLAB_DEBUG_INITIAL; 2169 } 2170 #if FORCED_DEBUG 2171 /* 2172 * Enable redzoning and last user accounting, except for caches with 2173 * large objects, if the increased size would increase the object size 2174 * above the next power of two: caches with object sizes just above a 2175 * power of two have a significant amount of internal fragmentation. 2176 */ 2177 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) 2178 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2179 if (!(flags & SLAB_DESTROY_BY_RCU)) 2180 flags |= SLAB_POISON; 2181 #endif 2182 if (flags & SLAB_DESTROY_BY_RCU) 2183 BUG_ON(flags & SLAB_POISON); 2184 #endif 2185 if (flags & SLAB_DESTROY_BY_RCU) 2186 BUG_ON(dtor); 2187 2188 /* 2189 * Always checks flags, a caller might be expecting debug support which 2190 * isn't available. 2191 */ 2192 BUG_ON(flags & ~CREATE_MASK); 2193 2194 /* 2195 * Check that size is in terms of words. This is needed to avoid 2196 * unaligned accesses for some archs when redzoning is used, and makes 2197 * sure any on-slab bufctl's are also correctly aligned. 2198 */ 2199 if (size & (BYTES_PER_WORD - 1)) { 2200 size += (BYTES_PER_WORD - 1); 2201 size &= ~(BYTES_PER_WORD - 1); 2202 } 2203 2204 /* calculate the final buffer alignment: */ 2205 2206 /* 1) arch recommendation: can be overridden for debug */ 2207 if (flags & SLAB_HWCACHE_ALIGN) { 2208 /* 2209 * Default alignment: as specified by the arch code. Except if 2210 * an object is really small, then squeeze multiple objects into 2211 * one cacheline. 2212 */ 2213 ralign = cache_line_size(); 2214 while (size <= ralign / 2) 2215 ralign /= 2; 2216 } else { 2217 ralign = BYTES_PER_WORD; 2218 } 2219 2220 /* 2221 * Redzoning and user store require word alignment. Note this will be 2222 * overridden by architecture or caller mandated alignment if either 2223 * is greater than BYTES_PER_WORD. 2224 */ 2225 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) 2226 ralign = BYTES_PER_WORD; 2227 2228 /* 2) arch mandated alignment */ 2229 if (ralign < ARCH_SLAB_MINALIGN) { 2230 ralign = ARCH_SLAB_MINALIGN; 2231 } 2232 /* 3) caller mandated alignment */ 2233 if (ralign < align) { 2234 ralign = align; 2235 } 2236 /* disable debug if necessary */ 2237 if (ralign > BYTES_PER_WORD) 2238 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2239 /* 2240 * 4) Store it. 2241 */ 2242 align = ralign; 2243 2244 /* Get cache's description obj. */ 2245 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); 2246 if (!cachep) 2247 goto oops; 2248 2249 #if DEBUG 2250 cachep->obj_size = size; 2251 2252 /* 2253 * Both debugging options require word-alignment which is calculated 2254 * into align above. 2255 */ 2256 if (flags & SLAB_RED_ZONE) { 2257 /* add space for red zone words */ 2258 cachep->obj_offset += BYTES_PER_WORD; 2259 size += 2 * BYTES_PER_WORD; 2260 } 2261 if (flags & SLAB_STORE_USER) { 2262 /* user store requires one word storage behind the end of 2263 * the real object. 2264 */ 2265 size += BYTES_PER_WORD; 2266 } 2267 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2268 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2269 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { 2270 cachep->obj_offset += PAGE_SIZE - size; 2271 size = PAGE_SIZE; 2272 } 2273 #endif 2274 #endif 2275 2276 /* 2277 * Determine if the slab management is 'on' or 'off' slab. 2278 * (bootstrapping cannot cope with offslab caches so don't do 2279 * it too early on.) 2280 */ 2281 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) 2282 /* 2283 * Size is large, assume best to place the slab management obj 2284 * off-slab (should allow better packing of objs). 2285 */ 2286 flags |= CFLGS_OFF_SLAB; 2287 2288 size = ALIGN(size, align); 2289 2290 left_over = calculate_slab_order(cachep, size, align, flags); 2291 2292 if (!cachep->num) { 2293 printk("kmem_cache_create: couldn't create cache %s.\n", name); 2294 kmem_cache_free(&cache_cache, cachep); 2295 cachep = NULL; 2296 goto oops; 2297 } 2298 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2299 + sizeof(struct slab), align); 2300 2301 /* 2302 * If the slab has been placed off-slab, and we have enough space then 2303 * move it on-slab. This is at the expense of any extra colouring. 2304 */ 2305 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2306 flags &= ~CFLGS_OFF_SLAB; 2307 left_over -= slab_size; 2308 } 2309 2310 if (flags & CFLGS_OFF_SLAB) { 2311 /* really off slab. No need for manual alignment */ 2312 slab_size = 2313 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2314 } 2315 2316 cachep->colour_off = cache_line_size(); 2317 /* Offset must be a multiple of the alignment. */ 2318 if (cachep->colour_off < align) 2319 cachep->colour_off = align; 2320 cachep->colour = left_over / cachep->colour_off; 2321 cachep->slab_size = slab_size; 2322 cachep->flags = flags; 2323 cachep->gfpflags = 0; 2324 if (flags & SLAB_CACHE_DMA) 2325 cachep->gfpflags |= GFP_DMA; 2326 cachep->buffer_size = size; 2327 cachep->reciprocal_buffer_size = reciprocal_value(size); 2328 2329 if (flags & CFLGS_OFF_SLAB) { 2330 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2331 /* 2332 * This is a possibility for one of the malloc_sizes caches. 2333 * But since we go off slab only for object size greater than 2334 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, 2335 * this should not happen at all. 2336 * But leave a BUG_ON for some lucky dude. 2337 */ 2338 BUG_ON(!cachep->slabp_cache); 2339 } 2340 cachep->ctor = ctor; 2341 cachep->dtor = dtor; 2342 cachep->name = name; 2343 2344 if (setup_cpu_cache(cachep)) { 2345 __kmem_cache_destroy(cachep); 2346 cachep = NULL; 2347 goto oops; 2348 } 2349 2350 /* cache setup completed, link it into the list */ 2351 list_add(&cachep->next, &cache_chain); 2352 oops: 2353 if (!cachep && (flags & SLAB_PANIC)) 2354 panic("kmem_cache_create(): failed to create slab `%s'\n", 2355 name); 2356 mutex_unlock(&cache_chain_mutex); 2357 return cachep; 2358 } 2359 EXPORT_SYMBOL(kmem_cache_create); 2360 2361 #if DEBUG 2362 static void check_irq_off(void) 2363 { 2364 BUG_ON(!irqs_disabled()); 2365 } 2366 2367 static void check_irq_on(void) 2368 { 2369 BUG_ON(irqs_disabled()); 2370 } 2371 2372 static void check_spinlock_acquired(struct kmem_cache *cachep) 2373 { 2374 #ifdef CONFIG_SMP 2375 check_irq_off(); 2376 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); 2377 #endif 2378 } 2379 2380 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2381 { 2382 #ifdef CONFIG_SMP 2383 check_irq_off(); 2384 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2385 #endif 2386 } 2387 2388 #else 2389 #define check_irq_off() do { } while(0) 2390 #define check_irq_on() do { } while(0) 2391 #define check_spinlock_acquired(x) do { } while(0) 2392 #define check_spinlock_acquired_node(x, y) do { } while(0) 2393 #endif 2394 2395 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2396 struct array_cache *ac, 2397 int force, int node); 2398 2399 static void do_drain(void *arg) 2400 { 2401 struct kmem_cache *cachep = arg; 2402 struct array_cache *ac; 2403 int node = numa_node_id(); 2404 2405 check_irq_off(); 2406 ac = cpu_cache_get(cachep); 2407 spin_lock(&cachep->nodelists[node]->list_lock); 2408 free_block(cachep, ac->entry, ac->avail, node); 2409 spin_unlock(&cachep->nodelists[node]->list_lock); 2410 ac->avail = 0; 2411 } 2412 2413 static void drain_cpu_caches(struct kmem_cache *cachep) 2414 { 2415 struct kmem_list3 *l3; 2416 int node; 2417 2418 on_each_cpu(do_drain, cachep, 1, 1); 2419 check_irq_on(); 2420 for_each_online_node(node) { 2421 l3 = cachep->nodelists[node]; 2422 if (l3 && l3->alien) 2423 drain_alien_cache(cachep, l3->alien); 2424 } 2425 2426 for_each_online_node(node) { 2427 l3 = cachep->nodelists[node]; 2428 if (l3) 2429 drain_array(cachep, l3, l3->shared, 1, node); 2430 } 2431 } 2432 2433 /* 2434 * Remove slabs from the list of free slabs. 2435 * Specify the number of slabs to drain in tofree. 2436 * 2437 * Returns the actual number of slabs released. 2438 */ 2439 static int drain_freelist(struct kmem_cache *cache, 2440 struct kmem_list3 *l3, int tofree) 2441 { 2442 struct list_head *p; 2443 int nr_freed; 2444 struct slab *slabp; 2445 2446 nr_freed = 0; 2447 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { 2448 2449 spin_lock_irq(&l3->list_lock); 2450 p = l3->slabs_free.prev; 2451 if (p == &l3->slabs_free) { 2452 spin_unlock_irq(&l3->list_lock); 2453 goto out; 2454 } 2455 2456 slabp = list_entry(p, struct slab, list); 2457 #if DEBUG 2458 BUG_ON(slabp->inuse); 2459 #endif 2460 list_del(&slabp->list); 2461 /* 2462 * Safe to drop the lock. The slab is no longer linked 2463 * to the cache. 2464 */ 2465 l3->free_objects -= cache->num; 2466 spin_unlock_irq(&l3->list_lock); 2467 slab_destroy(cache, slabp); 2468 nr_freed++; 2469 } 2470 out: 2471 return nr_freed; 2472 } 2473 2474 /* Called with cache_chain_mutex held to protect against cpu hotplug */ 2475 static int __cache_shrink(struct kmem_cache *cachep) 2476 { 2477 int ret = 0, i = 0; 2478 struct kmem_list3 *l3; 2479 2480 drain_cpu_caches(cachep); 2481 2482 check_irq_on(); 2483 for_each_online_node(i) { 2484 l3 = cachep->nodelists[i]; 2485 if (!l3) 2486 continue; 2487 2488 drain_freelist(cachep, l3, l3->free_objects); 2489 2490 ret += !list_empty(&l3->slabs_full) || 2491 !list_empty(&l3->slabs_partial); 2492 } 2493 return (ret ? 1 : 0); 2494 } 2495 2496 /** 2497 * kmem_cache_shrink - Shrink a cache. 2498 * @cachep: The cache to shrink. 2499 * 2500 * Releases as many slabs as possible for a cache. 2501 * To help debugging, a zero exit status indicates all slabs were released. 2502 */ 2503 int kmem_cache_shrink(struct kmem_cache *cachep) 2504 { 2505 int ret; 2506 BUG_ON(!cachep || in_interrupt()); 2507 2508 mutex_lock(&cache_chain_mutex); 2509 ret = __cache_shrink(cachep); 2510 mutex_unlock(&cache_chain_mutex); 2511 return ret; 2512 } 2513 EXPORT_SYMBOL(kmem_cache_shrink); 2514 2515 /** 2516 * kmem_cache_destroy - delete a cache 2517 * @cachep: the cache to destroy 2518 * 2519 * Remove a struct kmem_cache object from the slab cache. 2520 * 2521 * It is expected this function will be called by a module when it is 2522 * unloaded. This will remove the cache completely, and avoid a duplicate 2523 * cache being allocated each time a module is loaded and unloaded, if the 2524 * module doesn't have persistent in-kernel storage across loads and unloads. 2525 * 2526 * The cache must be empty before calling this function. 2527 * 2528 * The caller must guarantee that noone will allocate memory from the cache 2529 * during the kmem_cache_destroy(). 2530 */ 2531 void kmem_cache_destroy(struct kmem_cache *cachep) 2532 { 2533 BUG_ON(!cachep || in_interrupt()); 2534 2535 /* Find the cache in the chain of caches. */ 2536 mutex_lock(&cache_chain_mutex); 2537 /* 2538 * the chain is never empty, cache_cache is never destroyed 2539 */ 2540 list_del(&cachep->next); 2541 if (__cache_shrink(cachep)) { 2542 slab_error(cachep, "Can't free all objects"); 2543 list_add(&cachep->next, &cache_chain); 2544 mutex_unlock(&cache_chain_mutex); 2545 return; 2546 } 2547 2548 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2549 synchronize_rcu(); 2550 2551 __kmem_cache_destroy(cachep); 2552 mutex_unlock(&cache_chain_mutex); 2553 } 2554 EXPORT_SYMBOL(kmem_cache_destroy); 2555 2556 /* 2557 * Get the memory for a slab management obj. 2558 * For a slab cache when the slab descriptor is off-slab, slab descriptors 2559 * always come from malloc_sizes caches. The slab descriptor cannot 2560 * come from the same cache which is getting created because, 2561 * when we are searching for an appropriate cache for these 2562 * descriptors in kmem_cache_create, we search through the malloc_sizes array. 2563 * If we are creating a malloc_sizes cache here it would not be visible to 2564 * kmem_find_general_cachep till the initialization is complete. 2565 * Hence we cannot have slabp_cache same as the original cache. 2566 */ 2567 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2568 int colour_off, gfp_t local_flags, 2569 int nodeid) 2570 { 2571 struct slab *slabp; 2572 2573 if (OFF_SLAB(cachep)) { 2574 /* Slab management obj is off-slab. */ 2575 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2576 local_flags & ~GFP_THISNODE, nodeid); 2577 if (!slabp) 2578 return NULL; 2579 } else { 2580 slabp = objp + colour_off; 2581 colour_off += cachep->slab_size; 2582 } 2583 slabp->inuse = 0; 2584 slabp->colouroff = colour_off; 2585 slabp->s_mem = objp + colour_off; 2586 slabp->nodeid = nodeid; 2587 return slabp; 2588 } 2589 2590 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2591 { 2592 return (kmem_bufctl_t *) (slabp + 1); 2593 } 2594 2595 static void cache_init_objs(struct kmem_cache *cachep, 2596 struct slab *slabp, unsigned long ctor_flags) 2597 { 2598 int i; 2599 2600 for (i = 0; i < cachep->num; i++) { 2601 void *objp = index_to_obj(cachep, slabp, i); 2602 #if DEBUG 2603 /* need to poison the objs? */ 2604 if (cachep->flags & SLAB_POISON) 2605 poison_obj(cachep, objp, POISON_FREE); 2606 if (cachep->flags & SLAB_STORE_USER) 2607 *dbg_userword(cachep, objp) = NULL; 2608 2609 if (cachep->flags & SLAB_RED_ZONE) { 2610 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2611 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2612 } 2613 /* 2614 * Constructors are not allowed to allocate memory from the same 2615 * cache which they are a constructor for. Otherwise, deadlock. 2616 * They must also be threaded. 2617 */ 2618 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2619 cachep->ctor(objp + obj_offset(cachep), cachep, 2620 ctor_flags); 2621 2622 if (cachep->flags & SLAB_RED_ZONE) { 2623 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2624 slab_error(cachep, "constructor overwrote the" 2625 " end of an object"); 2626 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2627 slab_error(cachep, "constructor overwrote the" 2628 " start of an object"); 2629 } 2630 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2631 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2632 kernel_map_pages(virt_to_page(objp), 2633 cachep->buffer_size / PAGE_SIZE, 0); 2634 #else 2635 if (cachep->ctor) 2636 cachep->ctor(objp, cachep, ctor_flags); 2637 #endif 2638 slab_bufctl(slabp)[i] = i + 1; 2639 } 2640 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2641 slabp->free = 0; 2642 } 2643 2644 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2645 { 2646 if (flags & GFP_DMA) 2647 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2648 else 2649 BUG_ON(cachep->gfpflags & GFP_DMA); 2650 } 2651 2652 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2653 int nodeid) 2654 { 2655 void *objp = index_to_obj(cachep, slabp, slabp->free); 2656 kmem_bufctl_t next; 2657 2658 slabp->inuse++; 2659 next = slab_bufctl(slabp)[slabp->free]; 2660 #if DEBUG 2661 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2662 WARN_ON(slabp->nodeid != nodeid); 2663 #endif 2664 slabp->free = next; 2665 2666 return objp; 2667 } 2668 2669 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2670 void *objp, int nodeid) 2671 { 2672 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2673 2674 #if DEBUG 2675 /* Verify that the slab belongs to the intended node */ 2676 WARN_ON(slabp->nodeid != nodeid); 2677 2678 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2679 printk(KERN_ERR "slab: double free detected in cache " 2680 "'%s', objp %p\n", cachep->name, objp); 2681 BUG(); 2682 } 2683 #endif 2684 slab_bufctl(slabp)[objnr] = slabp->free; 2685 slabp->free = objnr; 2686 slabp->inuse--; 2687 } 2688 2689 /* 2690 * Map pages beginning at addr to the given cache and slab. This is required 2691 * for the slab allocator to be able to lookup the cache and slab of a 2692 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. 2693 */ 2694 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, 2695 void *addr) 2696 { 2697 int nr_pages; 2698 struct page *page; 2699 2700 page = virt_to_page(addr); 2701 2702 nr_pages = 1; 2703 if (likely(!PageCompound(page))) 2704 nr_pages <<= cache->gfporder; 2705 2706 do { 2707 page_set_cache(page, cache); 2708 page_set_slab(page, slab); 2709 page++; 2710 } while (--nr_pages); 2711 } 2712 2713 /* 2714 * Grow (by 1) the number of slabs within a cache. This is called by 2715 * kmem_cache_alloc() when there are no active objs left in a cache. 2716 */ 2717 static int cache_grow(struct kmem_cache *cachep, 2718 gfp_t flags, int nodeid, void *objp) 2719 { 2720 struct slab *slabp; 2721 size_t offset; 2722 gfp_t local_flags; 2723 unsigned long ctor_flags; 2724 struct kmem_list3 *l3; 2725 2726 /* 2727 * Be lazy and only check for valid flags here, keeping it out of the 2728 * critical path in kmem_cache_alloc(). 2729 */ 2730 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW)); 2731 if (flags & __GFP_NO_GROW) 2732 return 0; 2733 2734 ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2735 local_flags = (flags & GFP_LEVEL_MASK); 2736 if (!(local_flags & __GFP_WAIT)) 2737 /* 2738 * Not allowed to sleep. Need to tell a constructor about 2739 * this - it might need to know... 2740 */ 2741 ctor_flags |= SLAB_CTOR_ATOMIC; 2742 2743 /* Take the l3 list lock to change the colour_next on this node */ 2744 check_irq_off(); 2745 l3 = cachep->nodelists[nodeid]; 2746 spin_lock(&l3->list_lock); 2747 2748 /* Get colour for the slab, and cal the next value. */ 2749 offset = l3->colour_next; 2750 l3->colour_next++; 2751 if (l3->colour_next >= cachep->colour) 2752 l3->colour_next = 0; 2753 spin_unlock(&l3->list_lock); 2754 2755 offset *= cachep->colour_off; 2756 2757 if (local_flags & __GFP_WAIT) 2758 local_irq_enable(); 2759 2760 /* 2761 * The test for missing atomic flag is performed here, rather than 2762 * the more obvious place, simply to reduce the critical path length 2763 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2764 * will eventually be caught here (where it matters). 2765 */ 2766 kmem_flagcheck(cachep, flags); 2767 2768 /* 2769 * Get mem for the objs. Attempt to allocate a physical page from 2770 * 'nodeid'. 2771 */ 2772 if (!objp) 2773 objp = kmem_getpages(cachep, flags, nodeid); 2774 if (!objp) 2775 goto failed; 2776 2777 /* Get slab management. */ 2778 slabp = alloc_slabmgmt(cachep, objp, offset, 2779 local_flags & ~GFP_THISNODE, nodeid); 2780 if (!slabp) 2781 goto opps1; 2782 2783 slabp->nodeid = nodeid; 2784 slab_map_pages(cachep, slabp, objp); 2785 2786 cache_init_objs(cachep, slabp, ctor_flags); 2787 2788 if (local_flags & __GFP_WAIT) 2789 local_irq_disable(); 2790 check_irq_off(); 2791 spin_lock(&l3->list_lock); 2792 2793 /* Make slab active. */ 2794 list_add_tail(&slabp->list, &(l3->slabs_free)); 2795 STATS_INC_GROWN(cachep); 2796 l3->free_objects += cachep->num; 2797 spin_unlock(&l3->list_lock); 2798 return 1; 2799 opps1: 2800 kmem_freepages(cachep, objp); 2801 failed: 2802 if (local_flags & __GFP_WAIT) 2803 local_irq_disable(); 2804 return 0; 2805 } 2806 2807 #if DEBUG 2808 2809 /* 2810 * Perform extra freeing checks: 2811 * - detect bad pointers. 2812 * - POISON/RED_ZONE checking 2813 * - destructor calls, for caches with POISON+dtor 2814 */ 2815 static void kfree_debugcheck(const void *objp) 2816 { 2817 struct page *page; 2818 2819 if (!virt_addr_valid(objp)) { 2820 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2821 (unsigned long)objp); 2822 BUG(); 2823 } 2824 page = virt_to_page(objp); 2825 if (!PageSlab(page)) { 2826 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", 2827 (unsigned long)objp); 2828 BUG(); 2829 } 2830 } 2831 2832 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2833 { 2834 unsigned long redzone1, redzone2; 2835 2836 redzone1 = *dbg_redzone1(cache, obj); 2837 redzone2 = *dbg_redzone2(cache, obj); 2838 2839 /* 2840 * Redzone is ok. 2841 */ 2842 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2843 return; 2844 2845 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2846 slab_error(cache, "double free detected"); 2847 else 2848 slab_error(cache, "memory outside object was overwritten"); 2849 2850 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n", 2851 obj, redzone1, redzone2); 2852 } 2853 2854 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2855 void *caller) 2856 { 2857 struct page *page; 2858 unsigned int objnr; 2859 struct slab *slabp; 2860 2861 objp -= obj_offset(cachep); 2862 kfree_debugcheck(objp); 2863 page = virt_to_page(objp); 2864 2865 slabp = page_get_slab(page); 2866 2867 if (cachep->flags & SLAB_RED_ZONE) { 2868 verify_redzone_free(cachep, objp); 2869 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2870 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2871 } 2872 if (cachep->flags & SLAB_STORE_USER) 2873 *dbg_userword(cachep, objp) = caller; 2874 2875 objnr = obj_to_index(cachep, slabp, objp); 2876 2877 BUG_ON(objnr >= cachep->num); 2878 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2879 2880 if (cachep->flags & SLAB_DEBUG_INITIAL) { 2881 /* 2882 * Need to call the slab's constructor so the caller can 2883 * perform a verify of its state (debugging). Called without 2884 * the cache-lock held. 2885 */ 2886 cachep->ctor(objp + obj_offset(cachep), 2887 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); 2888 } 2889 if (cachep->flags & SLAB_POISON && cachep->dtor) { 2890 /* we want to cache poison the object, 2891 * call the destruction callback 2892 */ 2893 cachep->dtor(objp + obj_offset(cachep), cachep, 0); 2894 } 2895 #ifdef CONFIG_DEBUG_SLAB_LEAK 2896 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2897 #endif 2898 if (cachep->flags & SLAB_POISON) { 2899 #ifdef CONFIG_DEBUG_PAGEALLOC 2900 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2901 store_stackinfo(cachep, objp, (unsigned long)caller); 2902 kernel_map_pages(virt_to_page(objp), 2903 cachep->buffer_size / PAGE_SIZE, 0); 2904 } else { 2905 poison_obj(cachep, objp, POISON_FREE); 2906 } 2907 #else 2908 poison_obj(cachep, objp, POISON_FREE); 2909 #endif 2910 } 2911 return objp; 2912 } 2913 2914 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2915 { 2916 kmem_bufctl_t i; 2917 int entries = 0; 2918 2919 /* Check slab's freelist to see if this obj is there. */ 2920 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2921 entries++; 2922 if (entries > cachep->num || i >= cachep->num) 2923 goto bad; 2924 } 2925 if (entries != cachep->num - slabp->inuse) { 2926 bad: 2927 printk(KERN_ERR "slab: Internal list corruption detected in " 2928 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2929 cachep->name, cachep->num, slabp, slabp->inuse); 2930 for (i = 0; 2931 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2932 i++) { 2933 if (i % 16 == 0) 2934 printk("\n%03x:", i); 2935 printk(" %02x", ((unsigned char *)slabp)[i]); 2936 } 2937 printk("\n"); 2938 BUG(); 2939 } 2940 } 2941 #else 2942 #define kfree_debugcheck(x) do { } while(0) 2943 #define cache_free_debugcheck(x,objp,z) (objp) 2944 #define check_slabp(x,y) do { } while(0) 2945 #endif 2946 2947 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2948 { 2949 int batchcount; 2950 struct kmem_list3 *l3; 2951 struct array_cache *ac; 2952 int node; 2953 2954 node = numa_node_id(); 2955 2956 check_irq_off(); 2957 ac = cpu_cache_get(cachep); 2958 retry: 2959 batchcount = ac->batchcount; 2960 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2961 /* 2962 * If there was little recent activity on this cache, then 2963 * perform only a partial refill. Otherwise we could generate 2964 * refill bouncing. 2965 */ 2966 batchcount = BATCHREFILL_LIMIT; 2967 } 2968 l3 = cachep->nodelists[node]; 2969 2970 BUG_ON(ac->avail > 0 || !l3); 2971 spin_lock(&l3->list_lock); 2972 2973 /* See if we can refill from the shared array */ 2974 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) 2975 goto alloc_done; 2976 2977 while (batchcount > 0) { 2978 struct list_head *entry; 2979 struct slab *slabp; 2980 /* Get slab alloc is to come from. */ 2981 entry = l3->slabs_partial.next; 2982 if (entry == &l3->slabs_partial) { 2983 l3->free_touched = 1; 2984 entry = l3->slabs_free.next; 2985 if (entry == &l3->slabs_free) 2986 goto must_grow; 2987 } 2988 2989 slabp = list_entry(entry, struct slab, list); 2990 check_slabp(cachep, slabp); 2991 check_spinlock_acquired(cachep); 2992 while (slabp->inuse < cachep->num && batchcount--) { 2993 STATS_INC_ALLOCED(cachep); 2994 STATS_INC_ACTIVE(cachep); 2995 STATS_SET_HIGH(cachep); 2996 2997 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 2998 node); 2999 } 3000 check_slabp(cachep, slabp); 3001 3002 /* move slabp to correct slabp list: */ 3003 list_del(&slabp->list); 3004 if (slabp->free == BUFCTL_END) 3005 list_add(&slabp->list, &l3->slabs_full); 3006 else 3007 list_add(&slabp->list, &l3->slabs_partial); 3008 } 3009 3010 must_grow: 3011 l3->free_objects -= ac->avail; 3012 alloc_done: 3013 spin_unlock(&l3->list_lock); 3014 3015 if (unlikely(!ac->avail)) { 3016 int x; 3017 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); 3018 3019 /* cache_grow can reenable interrupts, then ac could change. */ 3020 ac = cpu_cache_get(cachep); 3021 if (!x && ac->avail == 0) /* no objects in sight? abort */ 3022 return NULL; 3023 3024 if (!ac->avail) /* objects refilled by interrupt? */ 3025 goto retry; 3026 } 3027 ac->touched = 1; 3028 return ac->entry[--ac->avail]; 3029 } 3030 3031 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3032 gfp_t flags) 3033 { 3034 might_sleep_if(flags & __GFP_WAIT); 3035 #if DEBUG 3036 kmem_flagcheck(cachep, flags); 3037 #endif 3038 } 3039 3040 #if DEBUG 3041 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3042 gfp_t flags, void *objp, void *caller) 3043 { 3044 if (!objp) 3045 return objp; 3046 if (cachep->flags & SLAB_POISON) { 3047 #ifdef CONFIG_DEBUG_PAGEALLOC 3048 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 3049 kernel_map_pages(virt_to_page(objp), 3050 cachep->buffer_size / PAGE_SIZE, 1); 3051 else 3052 check_poison_obj(cachep, objp); 3053 #else 3054 check_poison_obj(cachep, objp); 3055 #endif 3056 poison_obj(cachep, objp, POISON_INUSE); 3057 } 3058 if (cachep->flags & SLAB_STORE_USER) 3059 *dbg_userword(cachep, objp) = caller; 3060 3061 if (cachep->flags & SLAB_RED_ZONE) { 3062 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3063 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3064 slab_error(cachep, "double free, or memory outside" 3065 " object was overwritten"); 3066 printk(KERN_ERR 3067 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", 3068 objp, *dbg_redzone1(cachep, objp), 3069 *dbg_redzone2(cachep, objp)); 3070 } 3071 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3072 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3073 } 3074 #ifdef CONFIG_DEBUG_SLAB_LEAK 3075 { 3076 struct slab *slabp; 3077 unsigned objnr; 3078 3079 slabp = page_get_slab(virt_to_page(objp)); 3080 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 3081 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 3082 } 3083 #endif 3084 objp += obj_offset(cachep); 3085 if (cachep->ctor && cachep->flags & SLAB_POISON) { 3086 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; 3087 3088 if (!(flags & __GFP_WAIT)) 3089 ctor_flags |= SLAB_CTOR_ATOMIC; 3090 3091 cachep->ctor(objp, cachep, ctor_flags); 3092 } 3093 #if ARCH_SLAB_MINALIGN 3094 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { 3095 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3096 objp, ARCH_SLAB_MINALIGN); 3097 } 3098 #endif 3099 return objp; 3100 } 3101 #else 3102 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3103 #endif 3104 3105 #ifdef CONFIG_FAILSLAB 3106 3107 static struct failslab_attr { 3108 3109 struct fault_attr attr; 3110 3111 u32 ignore_gfp_wait; 3112 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3113 struct dentry *ignore_gfp_wait_file; 3114 #endif 3115 3116 } failslab = { 3117 .attr = FAULT_ATTR_INITIALIZER, 3118 .ignore_gfp_wait = 1, 3119 }; 3120 3121 static int __init setup_failslab(char *str) 3122 { 3123 return setup_fault_attr(&failslab.attr, str); 3124 } 3125 __setup("failslab=", setup_failslab); 3126 3127 static int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3128 { 3129 if (cachep == &cache_cache) 3130 return 0; 3131 if (flags & __GFP_NOFAIL) 3132 return 0; 3133 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) 3134 return 0; 3135 3136 return should_fail(&failslab.attr, obj_size(cachep)); 3137 } 3138 3139 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3140 3141 static int __init failslab_debugfs(void) 3142 { 3143 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 3144 struct dentry *dir; 3145 int err; 3146 3147 err = init_fault_attr_dentries(&failslab.attr, "failslab"); 3148 if (err) 3149 return err; 3150 dir = failslab.attr.dentries.dir; 3151 3152 failslab.ignore_gfp_wait_file = 3153 debugfs_create_bool("ignore-gfp-wait", mode, dir, 3154 &failslab.ignore_gfp_wait); 3155 3156 if (!failslab.ignore_gfp_wait_file) { 3157 err = -ENOMEM; 3158 debugfs_remove(failslab.ignore_gfp_wait_file); 3159 cleanup_fault_attr_dentries(&failslab.attr); 3160 } 3161 3162 return err; 3163 } 3164 3165 late_initcall(failslab_debugfs); 3166 3167 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 3168 3169 #else /* CONFIG_FAILSLAB */ 3170 3171 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3172 { 3173 return 0; 3174 } 3175 3176 #endif /* CONFIG_FAILSLAB */ 3177 3178 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3179 { 3180 void *objp; 3181 struct array_cache *ac; 3182 3183 check_irq_off(); 3184 3185 if (should_failslab(cachep, flags)) 3186 return NULL; 3187 3188 ac = cpu_cache_get(cachep); 3189 if (likely(ac->avail)) { 3190 STATS_INC_ALLOCHIT(cachep); 3191 ac->touched = 1; 3192 objp = ac->entry[--ac->avail]; 3193 } else { 3194 STATS_INC_ALLOCMISS(cachep); 3195 objp = cache_alloc_refill(cachep, flags); 3196 } 3197 return objp; 3198 } 3199 3200 static __always_inline void *__cache_alloc(struct kmem_cache *cachep, 3201 gfp_t flags, void *caller) 3202 { 3203 unsigned long save_flags; 3204 void *objp = NULL; 3205 3206 cache_alloc_debugcheck_before(cachep, flags); 3207 3208 local_irq_save(save_flags); 3209 3210 if (unlikely(NUMA_BUILD && 3211 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) 3212 objp = alternate_node_alloc(cachep, flags); 3213 3214 if (!objp) 3215 objp = ____cache_alloc(cachep, flags); 3216 /* 3217 * We may just have run out of memory on the local node. 3218 * ____cache_alloc_node() knows how to locate memory on other nodes 3219 */ 3220 if (NUMA_BUILD && !objp) 3221 objp = ____cache_alloc_node(cachep, flags, numa_node_id()); 3222 local_irq_restore(save_flags); 3223 objp = cache_alloc_debugcheck_after(cachep, flags, objp, 3224 caller); 3225 prefetchw(objp); 3226 return objp; 3227 } 3228 3229 #ifdef CONFIG_NUMA 3230 /* 3231 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 3232 * 3233 * If we are in_interrupt, then process context, including cpusets and 3234 * mempolicy, may not apply and should not be used for allocation policy. 3235 */ 3236 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3237 { 3238 int nid_alloc, nid_here; 3239 3240 if (in_interrupt() || (flags & __GFP_THISNODE)) 3241 return NULL; 3242 nid_alloc = nid_here = numa_node_id(); 3243 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3244 nid_alloc = cpuset_mem_spread_node(); 3245 else if (current->mempolicy) 3246 nid_alloc = slab_node(current->mempolicy); 3247 if (nid_alloc != nid_here) 3248 return ____cache_alloc_node(cachep, flags, nid_alloc); 3249 return NULL; 3250 } 3251 3252 /* 3253 * Fallback function if there was no memory available and no objects on a 3254 * certain node and fall back is permitted. First we scan all the 3255 * available nodelists for available objects. If that fails then we 3256 * perform an allocation without specifying a node. This allows the page 3257 * allocator to do its reclaim / fallback magic. We then insert the 3258 * slab into the proper nodelist and then allocate from it. 3259 */ 3260 void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3261 { 3262 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy)) 3263 ->node_zonelists[gfp_zone(flags)]; 3264 struct zone **z; 3265 void *obj = NULL; 3266 int nid; 3267 gfp_t local_flags = (flags & GFP_LEVEL_MASK); 3268 3269 retry: 3270 /* 3271 * Look through allowed nodes for objects available 3272 * from existing per node queues. 3273 */ 3274 for (z = zonelist->zones; *z && !obj; z++) { 3275 nid = zone_to_nid(*z); 3276 3277 if (cpuset_zone_allowed_hardwall(*z, flags) && 3278 cache->nodelists[nid] && 3279 cache->nodelists[nid]->free_objects) 3280 obj = ____cache_alloc_node(cache, 3281 flags | GFP_THISNODE, nid); 3282 } 3283 3284 if (!obj && !(flags & __GFP_NO_GROW)) { 3285 /* 3286 * This allocation will be performed within the constraints 3287 * of the current cpuset / memory policy requirements. 3288 * We may trigger various forms of reclaim on the allowed 3289 * set and go into memory reserves if necessary. 3290 */ 3291 if (local_flags & __GFP_WAIT) 3292 local_irq_enable(); 3293 kmem_flagcheck(cache, flags); 3294 obj = kmem_getpages(cache, flags, -1); 3295 if (local_flags & __GFP_WAIT) 3296 local_irq_disable(); 3297 if (obj) { 3298 /* 3299 * Insert into the appropriate per node queues 3300 */ 3301 nid = page_to_nid(virt_to_page(obj)); 3302 if (cache_grow(cache, flags, nid, obj)) { 3303 obj = ____cache_alloc_node(cache, 3304 flags | GFP_THISNODE, nid); 3305 if (!obj) 3306 /* 3307 * Another processor may allocate the 3308 * objects in the slab since we are 3309 * not holding any locks. 3310 */ 3311 goto retry; 3312 } else { 3313 /* cache_grow already freed obj */ 3314 obj = NULL; 3315 } 3316 } 3317 } 3318 return obj; 3319 } 3320 3321 /* 3322 * A interface to enable slab creation on nodeid 3323 */ 3324 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3325 int nodeid) 3326 { 3327 struct list_head *entry; 3328 struct slab *slabp; 3329 struct kmem_list3 *l3; 3330 void *obj; 3331 int x; 3332 3333 l3 = cachep->nodelists[nodeid]; 3334 BUG_ON(!l3); 3335 3336 retry: 3337 check_irq_off(); 3338 spin_lock(&l3->list_lock); 3339 entry = l3->slabs_partial.next; 3340 if (entry == &l3->slabs_partial) { 3341 l3->free_touched = 1; 3342 entry = l3->slabs_free.next; 3343 if (entry == &l3->slabs_free) 3344 goto must_grow; 3345 } 3346 3347 slabp = list_entry(entry, struct slab, list); 3348 check_spinlock_acquired_node(cachep, nodeid); 3349 check_slabp(cachep, slabp); 3350 3351 STATS_INC_NODEALLOCS(cachep); 3352 STATS_INC_ACTIVE(cachep); 3353 STATS_SET_HIGH(cachep); 3354 3355 BUG_ON(slabp->inuse == cachep->num); 3356 3357 obj = slab_get_obj(cachep, slabp, nodeid); 3358 check_slabp(cachep, slabp); 3359 l3->free_objects--; 3360 /* move slabp to correct slabp list: */ 3361 list_del(&slabp->list); 3362 3363 if (slabp->free == BUFCTL_END) 3364 list_add(&slabp->list, &l3->slabs_full); 3365 else 3366 list_add(&slabp->list, &l3->slabs_partial); 3367 3368 spin_unlock(&l3->list_lock); 3369 goto done; 3370 3371 must_grow: 3372 spin_unlock(&l3->list_lock); 3373 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); 3374 if (x) 3375 goto retry; 3376 3377 if (!(flags & __GFP_THISNODE)) 3378 /* Unable to grow the cache. Fall back to other nodes. */ 3379 return fallback_alloc(cachep, flags); 3380 3381 return NULL; 3382 3383 done: 3384 return obj; 3385 } 3386 #endif 3387 3388 /* 3389 * Caller needs to acquire correct kmem_list's list_lock 3390 */ 3391 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 3392 int node) 3393 { 3394 int i; 3395 struct kmem_list3 *l3; 3396 3397 for (i = 0; i < nr_objects; i++) { 3398 void *objp = objpp[i]; 3399 struct slab *slabp; 3400 3401 slabp = virt_to_slab(objp); 3402 l3 = cachep->nodelists[node]; 3403 list_del(&slabp->list); 3404 check_spinlock_acquired_node(cachep, node); 3405 check_slabp(cachep, slabp); 3406 slab_put_obj(cachep, slabp, objp, node); 3407 STATS_DEC_ACTIVE(cachep); 3408 l3->free_objects++; 3409 check_slabp(cachep, slabp); 3410 3411 /* fixup slab chains */ 3412 if (slabp->inuse == 0) { 3413 if (l3->free_objects > l3->free_limit) { 3414 l3->free_objects -= cachep->num; 3415 /* No need to drop any previously held 3416 * lock here, even if we have a off-slab slab 3417 * descriptor it is guaranteed to come from 3418 * a different cache, refer to comments before 3419 * alloc_slabmgmt. 3420 */ 3421 slab_destroy(cachep, slabp); 3422 } else { 3423 list_add(&slabp->list, &l3->slabs_free); 3424 } 3425 } else { 3426 /* Unconditionally move a slab to the end of the 3427 * partial list on free - maximum time for the 3428 * other objects to be freed, too. 3429 */ 3430 list_add_tail(&slabp->list, &l3->slabs_partial); 3431 } 3432 } 3433 } 3434 3435 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3436 { 3437 int batchcount; 3438 struct kmem_list3 *l3; 3439 int node = numa_node_id(); 3440 3441 batchcount = ac->batchcount; 3442 #if DEBUG 3443 BUG_ON(!batchcount || batchcount > ac->avail); 3444 #endif 3445 check_irq_off(); 3446 l3 = cachep->nodelists[node]; 3447 spin_lock(&l3->list_lock); 3448 if (l3->shared) { 3449 struct array_cache *shared_array = l3->shared; 3450 int max = shared_array->limit - shared_array->avail; 3451 if (max) { 3452 if (batchcount > max) 3453 batchcount = max; 3454 memcpy(&(shared_array->entry[shared_array->avail]), 3455 ac->entry, sizeof(void *) * batchcount); 3456 shared_array->avail += batchcount; 3457 goto free_done; 3458 } 3459 } 3460 3461 free_block(cachep, ac->entry, batchcount, node); 3462 free_done: 3463 #if STATS 3464 { 3465 int i = 0; 3466 struct list_head *p; 3467 3468 p = l3->slabs_free.next; 3469 while (p != &(l3->slabs_free)) { 3470 struct slab *slabp; 3471 3472 slabp = list_entry(p, struct slab, list); 3473 BUG_ON(slabp->inuse); 3474 3475 i++; 3476 p = p->next; 3477 } 3478 STATS_SET_FREEABLE(cachep, i); 3479 } 3480 #endif 3481 spin_unlock(&l3->list_lock); 3482 ac->avail -= batchcount; 3483 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3484 } 3485 3486 /* 3487 * Release an obj back to its cache. If the obj has a constructed state, it must 3488 * be in this state _before_ it is released. Called with disabled ints. 3489 */ 3490 static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3491 { 3492 struct array_cache *ac = cpu_cache_get(cachep); 3493 3494 check_irq_off(); 3495 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3496 3497 if (cache_free_alien(cachep, objp)) 3498 return; 3499 3500 if (likely(ac->avail < ac->limit)) { 3501 STATS_INC_FREEHIT(cachep); 3502 ac->entry[ac->avail++] = objp; 3503 return; 3504 } else { 3505 STATS_INC_FREEMISS(cachep); 3506 cache_flusharray(cachep, ac); 3507 ac->entry[ac->avail++] = objp; 3508 } 3509 } 3510 3511 /** 3512 * kmem_cache_alloc - Allocate an object 3513 * @cachep: The cache to allocate from. 3514 * @flags: See kmalloc(). 3515 * 3516 * Allocate an object from this cache. The flags are only relevant 3517 * if the cache has no available objects. 3518 */ 3519 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3520 { 3521 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3522 } 3523 EXPORT_SYMBOL(kmem_cache_alloc); 3524 3525 /** 3526 * kmem_cache_zalloc - Allocate an object. The memory is set to zero. 3527 * @cache: The cache to allocate from. 3528 * @flags: See kmalloc(). 3529 * 3530 * Allocate an object from this cache and set the allocated memory to zero. 3531 * The flags are only relevant if the cache has no available objects. 3532 */ 3533 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) 3534 { 3535 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); 3536 if (ret) 3537 memset(ret, 0, obj_size(cache)); 3538 return ret; 3539 } 3540 EXPORT_SYMBOL(kmem_cache_zalloc); 3541 3542 /** 3543 * kmem_ptr_validate - check if an untrusted pointer might 3544 * be a slab entry. 3545 * @cachep: the cache we're checking against 3546 * @ptr: pointer to validate 3547 * 3548 * This verifies that the untrusted pointer looks sane: 3549 * it is _not_ a guarantee that the pointer is actually 3550 * part of the slab cache in question, but it at least 3551 * validates that the pointer can be dereferenced and 3552 * looks half-way sane. 3553 * 3554 * Currently only used for dentry validation. 3555 */ 3556 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) 3557 { 3558 unsigned long addr = (unsigned long)ptr; 3559 unsigned long min_addr = PAGE_OFFSET; 3560 unsigned long align_mask = BYTES_PER_WORD - 1; 3561 unsigned long size = cachep->buffer_size; 3562 struct page *page; 3563 3564 if (unlikely(addr < min_addr)) 3565 goto out; 3566 if (unlikely(addr > (unsigned long)high_memory - size)) 3567 goto out; 3568 if (unlikely(addr & align_mask)) 3569 goto out; 3570 if (unlikely(!kern_addr_valid(addr))) 3571 goto out; 3572 if (unlikely(!kern_addr_valid(addr + size - 1))) 3573 goto out; 3574 page = virt_to_page(ptr); 3575 if (unlikely(!PageSlab(page))) 3576 goto out; 3577 if (unlikely(page_get_cache(page) != cachep)) 3578 goto out; 3579 return 1; 3580 out: 3581 return 0; 3582 } 3583 3584 #ifdef CONFIG_NUMA 3585 /** 3586 * kmem_cache_alloc_node - Allocate an object on the specified node 3587 * @cachep: The cache to allocate from. 3588 * @flags: See kmalloc(). 3589 * @nodeid: node number of the target node. 3590 * @caller: return address of caller, used for debug information 3591 * 3592 * Identical to kmem_cache_alloc but it will allocate memory on the given 3593 * node, which can improve the performance for cpu bound structures. 3594 * 3595 * Fallback to other node is possible if __GFP_THISNODE is not set. 3596 */ 3597 static __always_inline void * 3598 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3599 int nodeid, void *caller) 3600 { 3601 unsigned long save_flags; 3602 void *ptr = NULL; 3603 3604 cache_alloc_debugcheck_before(cachep, flags); 3605 local_irq_save(save_flags); 3606 3607 if (unlikely(nodeid == -1)) 3608 nodeid = numa_node_id(); 3609 3610 if (likely(cachep->nodelists[nodeid])) { 3611 if (nodeid == numa_node_id()) { 3612 /* 3613 * Use the locally cached objects if possible. 3614 * However ____cache_alloc does not allow fallback 3615 * to other nodes. It may fail while we still have 3616 * objects on other nodes available. 3617 */ 3618 ptr = ____cache_alloc(cachep, flags); 3619 } 3620 if (!ptr) { 3621 /* ___cache_alloc_node can fall back to other nodes */ 3622 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3623 } 3624 } else { 3625 /* Node not bootstrapped yet */ 3626 if (!(flags & __GFP_THISNODE)) 3627 ptr = fallback_alloc(cachep, flags); 3628 } 3629 3630 local_irq_restore(save_flags); 3631 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3632 3633 return ptr; 3634 } 3635 3636 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3637 { 3638 return __cache_alloc_node(cachep, flags, nodeid, 3639 __builtin_return_address(0)); 3640 } 3641 EXPORT_SYMBOL(kmem_cache_alloc_node); 3642 3643 static __always_inline void * 3644 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) 3645 { 3646 struct kmem_cache *cachep; 3647 3648 cachep = kmem_find_general_cachep(size, flags); 3649 if (unlikely(cachep == NULL)) 3650 return NULL; 3651 return kmem_cache_alloc_node(cachep, flags, node); 3652 } 3653 3654 #ifdef CONFIG_DEBUG_SLAB 3655 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3656 { 3657 return __do_kmalloc_node(size, flags, node, 3658 __builtin_return_address(0)); 3659 } 3660 EXPORT_SYMBOL(__kmalloc_node); 3661 3662 void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3663 int node, void *caller) 3664 { 3665 return __do_kmalloc_node(size, flags, node, caller); 3666 } 3667 EXPORT_SYMBOL(__kmalloc_node_track_caller); 3668 #else 3669 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3670 { 3671 return __do_kmalloc_node(size, flags, node, NULL); 3672 } 3673 EXPORT_SYMBOL(__kmalloc_node); 3674 #endif /* CONFIG_DEBUG_SLAB */ 3675 #endif /* CONFIG_NUMA */ 3676 3677 /** 3678 * __do_kmalloc - allocate memory 3679 * @size: how many bytes of memory are required. 3680 * @flags: the type of memory to allocate (see kmalloc). 3681 * @caller: function caller for debug tracking of the caller 3682 */ 3683 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3684 void *caller) 3685 { 3686 struct kmem_cache *cachep; 3687 3688 /* If you want to save a few bytes .text space: replace 3689 * __ with kmem_. 3690 * Then kmalloc uses the uninlined functions instead of the inline 3691 * functions. 3692 */ 3693 cachep = __find_general_cachep(size, flags); 3694 if (unlikely(cachep == NULL)) 3695 return NULL; 3696 return __cache_alloc(cachep, flags, caller); 3697 } 3698 3699 3700 #ifdef CONFIG_DEBUG_SLAB 3701 void *__kmalloc(size_t size, gfp_t flags) 3702 { 3703 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3704 } 3705 EXPORT_SYMBOL(__kmalloc); 3706 3707 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) 3708 { 3709 return __do_kmalloc(size, flags, caller); 3710 } 3711 EXPORT_SYMBOL(__kmalloc_track_caller); 3712 3713 #else 3714 void *__kmalloc(size_t size, gfp_t flags) 3715 { 3716 return __do_kmalloc(size, flags, NULL); 3717 } 3718 EXPORT_SYMBOL(__kmalloc); 3719 #endif 3720 3721 /** 3722 * kmem_cache_free - Deallocate an object 3723 * @cachep: The cache the allocation was from. 3724 * @objp: The previously allocated object. 3725 * 3726 * Free an object which was previously allocated from this 3727 * cache. 3728 */ 3729 void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3730 { 3731 unsigned long flags; 3732 3733 BUG_ON(virt_to_cache(objp) != cachep); 3734 3735 local_irq_save(flags); 3736 __cache_free(cachep, objp); 3737 local_irq_restore(flags); 3738 } 3739 EXPORT_SYMBOL(kmem_cache_free); 3740 3741 /** 3742 * kfree - free previously allocated memory 3743 * @objp: pointer returned by kmalloc. 3744 * 3745 * If @objp is NULL, no operation is performed. 3746 * 3747 * Don't free memory not originally allocated by kmalloc() 3748 * or you will run into trouble. 3749 */ 3750 void kfree(const void *objp) 3751 { 3752 struct kmem_cache *c; 3753 unsigned long flags; 3754 3755 if (unlikely(!objp)) 3756 return; 3757 local_irq_save(flags); 3758 kfree_debugcheck(objp); 3759 c = virt_to_cache(objp); 3760 debug_check_no_locks_freed(objp, obj_size(c)); 3761 __cache_free(c, (void *)objp); 3762 local_irq_restore(flags); 3763 } 3764 EXPORT_SYMBOL(kfree); 3765 3766 unsigned int kmem_cache_size(struct kmem_cache *cachep) 3767 { 3768 return obj_size(cachep); 3769 } 3770 EXPORT_SYMBOL(kmem_cache_size); 3771 3772 const char *kmem_cache_name(struct kmem_cache *cachep) 3773 { 3774 return cachep->name; 3775 } 3776 EXPORT_SYMBOL_GPL(kmem_cache_name); 3777 3778 /* 3779 * This initializes kmem_list3 or resizes varioius caches for all nodes. 3780 */ 3781 static int alloc_kmemlist(struct kmem_cache *cachep) 3782 { 3783 int node; 3784 struct kmem_list3 *l3; 3785 struct array_cache *new_shared; 3786 struct array_cache **new_alien = NULL; 3787 3788 for_each_online_node(node) { 3789 3790 if (use_alien_caches) { 3791 new_alien = alloc_alien_cache(node, cachep->limit); 3792 if (!new_alien) 3793 goto fail; 3794 } 3795 3796 new_shared = alloc_arraycache(node, 3797 cachep->shared*cachep->batchcount, 3798 0xbaadf00d); 3799 if (!new_shared) { 3800 free_alien_cache(new_alien); 3801 goto fail; 3802 } 3803 3804 l3 = cachep->nodelists[node]; 3805 if (l3) { 3806 struct array_cache *shared = l3->shared; 3807 3808 spin_lock_irq(&l3->list_lock); 3809 3810 if (shared) 3811 free_block(cachep, shared->entry, 3812 shared->avail, node); 3813 3814 l3->shared = new_shared; 3815 if (!l3->alien) { 3816 l3->alien = new_alien; 3817 new_alien = NULL; 3818 } 3819 l3->free_limit = (1 + nr_cpus_node(node)) * 3820 cachep->batchcount + cachep->num; 3821 spin_unlock_irq(&l3->list_lock); 3822 kfree(shared); 3823 free_alien_cache(new_alien); 3824 continue; 3825 } 3826 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); 3827 if (!l3) { 3828 free_alien_cache(new_alien); 3829 kfree(new_shared); 3830 goto fail; 3831 } 3832 3833 kmem_list3_init(l3); 3834 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3835 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3836 l3->shared = new_shared; 3837 l3->alien = new_alien; 3838 l3->free_limit = (1 + nr_cpus_node(node)) * 3839 cachep->batchcount + cachep->num; 3840 cachep->nodelists[node] = l3; 3841 } 3842 return 0; 3843 3844 fail: 3845 if (!cachep->next.next) { 3846 /* Cache is not active yet. Roll back what we did */ 3847 node--; 3848 while (node >= 0) { 3849 if (cachep->nodelists[node]) { 3850 l3 = cachep->nodelists[node]; 3851 3852 kfree(l3->shared); 3853 free_alien_cache(l3->alien); 3854 kfree(l3); 3855 cachep->nodelists[node] = NULL; 3856 } 3857 node--; 3858 } 3859 } 3860 return -ENOMEM; 3861 } 3862 3863 struct ccupdate_struct { 3864 struct kmem_cache *cachep; 3865 struct array_cache *new[NR_CPUS]; 3866 }; 3867 3868 static void do_ccupdate_local(void *info) 3869 { 3870 struct ccupdate_struct *new = info; 3871 struct array_cache *old; 3872 3873 check_irq_off(); 3874 old = cpu_cache_get(new->cachep); 3875 3876 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3877 new->new[smp_processor_id()] = old; 3878 } 3879 3880 /* Always called with the cache_chain_mutex held */ 3881 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3882 int batchcount, int shared) 3883 { 3884 struct ccupdate_struct *new; 3885 int i; 3886 3887 new = kzalloc(sizeof(*new), GFP_KERNEL); 3888 if (!new) 3889 return -ENOMEM; 3890 3891 for_each_online_cpu(i) { 3892 new->new[i] = alloc_arraycache(cpu_to_node(i), limit, 3893 batchcount); 3894 if (!new->new[i]) { 3895 for (i--; i >= 0; i--) 3896 kfree(new->new[i]); 3897 kfree(new); 3898 return -ENOMEM; 3899 } 3900 } 3901 new->cachep = cachep; 3902 3903 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); 3904 3905 check_irq_on(); 3906 cachep->batchcount = batchcount; 3907 cachep->limit = limit; 3908 cachep->shared = shared; 3909 3910 for_each_online_cpu(i) { 3911 struct array_cache *ccold = new->new[i]; 3912 if (!ccold) 3913 continue; 3914 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3915 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 3916 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3917 kfree(ccold); 3918 } 3919 kfree(new); 3920 return alloc_kmemlist(cachep); 3921 } 3922 3923 /* Called with cache_chain_mutex held always */ 3924 static int enable_cpucache(struct kmem_cache *cachep) 3925 { 3926 int err; 3927 int limit, shared; 3928 3929 /* 3930 * The head array serves three purposes: 3931 * - create a LIFO ordering, i.e. return objects that are cache-warm 3932 * - reduce the number of spinlock operations. 3933 * - reduce the number of linked list operations on the slab and 3934 * bufctl chains: array operations are cheaper. 3935 * The numbers are guessed, we should auto-tune as described by 3936 * Bonwick. 3937 */ 3938 if (cachep->buffer_size > 131072) 3939 limit = 1; 3940 else if (cachep->buffer_size > PAGE_SIZE) 3941 limit = 8; 3942 else if (cachep->buffer_size > 1024) 3943 limit = 24; 3944 else if (cachep->buffer_size > 256) 3945 limit = 54; 3946 else 3947 limit = 120; 3948 3949 /* 3950 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 3951 * allocation behaviour: Most allocs on one cpu, most free operations 3952 * on another cpu. For these cases, an efficient object passing between 3953 * cpus is necessary. This is provided by a shared array. The array 3954 * replaces Bonwick's magazine layer. 3955 * On uniprocessor, it's functionally equivalent (but less efficient) 3956 * to a larger limit. Thus disabled by default. 3957 */ 3958 shared = 0; 3959 #ifdef CONFIG_SMP 3960 if (cachep->buffer_size <= PAGE_SIZE) 3961 shared = 8; 3962 #endif 3963 3964 #if DEBUG 3965 /* 3966 * With debugging enabled, large batchcount lead to excessively long 3967 * periods with disabled local interrupts. Limit the batchcount 3968 */ 3969 if (limit > 32) 3970 limit = 32; 3971 #endif 3972 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); 3973 if (err) 3974 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 3975 cachep->name, -err); 3976 return err; 3977 } 3978 3979 /* 3980 * Drain an array if it contains any elements taking the l3 lock only if 3981 * necessary. Note that the l3 listlock also protects the array_cache 3982 * if drain_array() is used on the shared array. 3983 */ 3984 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 3985 struct array_cache *ac, int force, int node) 3986 { 3987 int tofree; 3988 3989 if (!ac || !ac->avail) 3990 return; 3991 if (ac->touched && !force) { 3992 ac->touched = 0; 3993 } else { 3994 spin_lock_irq(&l3->list_lock); 3995 if (ac->avail) { 3996 tofree = force ? ac->avail : (ac->limit + 4) / 5; 3997 if (tofree > ac->avail) 3998 tofree = (ac->avail + 1) / 2; 3999 free_block(cachep, ac->entry, tofree, node); 4000 ac->avail -= tofree; 4001 memmove(ac->entry, &(ac->entry[tofree]), 4002 sizeof(void *) * ac->avail); 4003 } 4004 spin_unlock_irq(&l3->list_lock); 4005 } 4006 } 4007 4008 /** 4009 * cache_reap - Reclaim memory from caches. 4010 * @unused: unused parameter 4011 * 4012 * Called from workqueue/eventd every few seconds. 4013 * Purpose: 4014 * - clear the per-cpu caches for this CPU. 4015 * - return freeable pages to the main free memory pool. 4016 * 4017 * If we cannot acquire the cache chain mutex then just give up - we'll try 4018 * again on the next iteration. 4019 */ 4020 static void cache_reap(struct work_struct *unused) 4021 { 4022 struct kmem_cache *searchp; 4023 struct kmem_list3 *l3; 4024 int node = numa_node_id(); 4025 4026 if (!mutex_trylock(&cache_chain_mutex)) { 4027 /* Give up. Setup the next iteration. */ 4028 schedule_delayed_work(&__get_cpu_var(reap_work), 4029 round_jiffies_relative(REAPTIMEOUT_CPUC)); 4030 return; 4031 } 4032 4033 list_for_each_entry(searchp, &cache_chain, next) { 4034 check_irq_on(); 4035 4036 /* 4037 * We only take the l3 lock if absolutely necessary and we 4038 * have established with reasonable certainty that 4039 * we can do some work if the lock was obtained. 4040 */ 4041 l3 = searchp->nodelists[node]; 4042 4043 reap_alien(searchp, l3); 4044 4045 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 4046 4047 /* 4048 * These are racy checks but it does not matter 4049 * if we skip one check or scan twice. 4050 */ 4051 if (time_after(l3->next_reap, jiffies)) 4052 goto next; 4053 4054 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 4055 4056 drain_array(searchp, l3, l3->shared, 0, node); 4057 4058 if (l3->free_touched) 4059 l3->free_touched = 0; 4060 else { 4061 int freed; 4062 4063 freed = drain_freelist(searchp, l3, (l3->free_limit + 4064 5 * searchp->num - 1) / (5 * searchp->num)); 4065 STATS_ADD_REAPED(searchp, freed); 4066 } 4067 next: 4068 cond_resched(); 4069 } 4070 check_irq_on(); 4071 mutex_unlock(&cache_chain_mutex); 4072 next_reap_node(); 4073 refresh_cpu_vm_stats(smp_processor_id()); 4074 /* Set up the next iteration */ 4075 schedule_delayed_work(&__get_cpu_var(reap_work), 4076 round_jiffies_relative(REAPTIMEOUT_CPUC)); 4077 } 4078 4079 #ifdef CONFIG_PROC_FS 4080 4081 static void print_slabinfo_header(struct seq_file *m) 4082 { 4083 /* 4084 * Output format version, so at least we can change it 4085 * without _too_ many complaints. 4086 */ 4087 #if STATS 4088 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 4089 #else 4090 seq_puts(m, "slabinfo - version: 2.1\n"); 4091 #endif 4092 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4093 "<objperslab> <pagesperslab>"); 4094 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4095 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4096 #if STATS 4097 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 4098 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 4099 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 4100 #endif 4101 seq_putc(m, '\n'); 4102 } 4103 4104 static void *s_start(struct seq_file *m, loff_t *pos) 4105 { 4106 loff_t n = *pos; 4107 struct list_head *p; 4108 4109 mutex_lock(&cache_chain_mutex); 4110 if (!n) 4111 print_slabinfo_header(m); 4112 p = cache_chain.next; 4113 while (n--) { 4114 p = p->next; 4115 if (p == &cache_chain) 4116 return NULL; 4117 } 4118 return list_entry(p, struct kmem_cache, next); 4119 } 4120 4121 static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4122 { 4123 struct kmem_cache *cachep = p; 4124 ++*pos; 4125 return cachep->next.next == &cache_chain ? 4126 NULL : list_entry(cachep->next.next, struct kmem_cache, next); 4127 } 4128 4129 static void s_stop(struct seq_file *m, void *p) 4130 { 4131 mutex_unlock(&cache_chain_mutex); 4132 } 4133 4134 static int s_show(struct seq_file *m, void *p) 4135 { 4136 struct kmem_cache *cachep = p; 4137 struct slab *slabp; 4138 unsigned long active_objs; 4139 unsigned long num_objs; 4140 unsigned long active_slabs = 0; 4141 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 4142 const char *name; 4143 char *error = NULL; 4144 int node; 4145 struct kmem_list3 *l3; 4146 4147 active_objs = 0; 4148 num_slabs = 0; 4149 for_each_online_node(node) { 4150 l3 = cachep->nodelists[node]; 4151 if (!l3) 4152 continue; 4153 4154 check_irq_on(); 4155 spin_lock_irq(&l3->list_lock); 4156 4157 list_for_each_entry(slabp, &l3->slabs_full, list) { 4158 if (slabp->inuse != cachep->num && !error) 4159 error = "slabs_full accounting error"; 4160 active_objs += cachep->num; 4161 active_slabs++; 4162 } 4163 list_for_each_entry(slabp, &l3->slabs_partial, list) { 4164 if (slabp->inuse == cachep->num && !error) 4165 error = "slabs_partial inuse accounting error"; 4166 if (!slabp->inuse && !error) 4167 error = "slabs_partial/inuse accounting error"; 4168 active_objs += slabp->inuse; 4169 active_slabs++; 4170 } 4171 list_for_each_entry(slabp, &l3->slabs_free, list) { 4172 if (slabp->inuse && !error) 4173 error = "slabs_free/inuse accounting error"; 4174 num_slabs++; 4175 } 4176 free_objects += l3->free_objects; 4177 if (l3->shared) 4178 shared_avail += l3->shared->avail; 4179 4180 spin_unlock_irq(&l3->list_lock); 4181 } 4182 num_slabs += active_slabs; 4183 num_objs = num_slabs * cachep->num; 4184 if (num_objs - active_objs != free_objects && !error) 4185 error = "free_objects accounting error"; 4186 4187 name = cachep->name; 4188 if (error) 4189 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 4190 4191 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 4192 name, active_objs, num_objs, cachep->buffer_size, 4193 cachep->num, (1 << cachep->gfporder)); 4194 seq_printf(m, " : tunables %4u %4u %4u", 4195 cachep->limit, cachep->batchcount, cachep->shared); 4196 seq_printf(m, " : slabdata %6lu %6lu %6lu", 4197 active_slabs, num_slabs, shared_avail); 4198 #if STATS 4199 { /* list3 stats */ 4200 unsigned long high = cachep->high_mark; 4201 unsigned long allocs = cachep->num_allocations; 4202 unsigned long grown = cachep->grown; 4203 unsigned long reaped = cachep->reaped; 4204 unsigned long errors = cachep->errors; 4205 unsigned long max_freeable = cachep->max_freeable; 4206 unsigned long node_allocs = cachep->node_allocs; 4207 unsigned long node_frees = cachep->node_frees; 4208 unsigned long overflows = cachep->node_overflow; 4209 4210 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ 4211 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, 4212 reaped, errors, max_freeable, node_allocs, 4213 node_frees, overflows); 4214 } 4215 /* cpu stats */ 4216 { 4217 unsigned long allochit = atomic_read(&cachep->allochit); 4218 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4219 unsigned long freehit = atomic_read(&cachep->freehit); 4220 unsigned long freemiss = atomic_read(&cachep->freemiss); 4221 4222 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4223 allochit, allocmiss, freehit, freemiss); 4224 } 4225 #endif 4226 seq_putc(m, '\n'); 4227 return 0; 4228 } 4229 4230 /* 4231 * slabinfo_op - iterator that generates /proc/slabinfo 4232 * 4233 * Output layout: 4234 * cache-name 4235 * num-active-objs 4236 * total-objs 4237 * object size 4238 * num-active-slabs 4239 * total-slabs 4240 * num-pages-per-slab 4241 * + further values on SMP and with statistics enabled 4242 */ 4243 4244 const struct seq_operations slabinfo_op = { 4245 .start = s_start, 4246 .next = s_next, 4247 .stop = s_stop, 4248 .show = s_show, 4249 }; 4250 4251 #define MAX_SLABINFO_WRITE 128 4252 /** 4253 * slabinfo_write - Tuning for the slab allocator 4254 * @file: unused 4255 * @buffer: user buffer 4256 * @count: data length 4257 * @ppos: unused 4258 */ 4259 ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4260 size_t count, loff_t *ppos) 4261 { 4262 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4263 int limit, batchcount, shared, res; 4264 struct kmem_cache *cachep; 4265 4266 if (count > MAX_SLABINFO_WRITE) 4267 return -EINVAL; 4268 if (copy_from_user(&kbuf, buffer, count)) 4269 return -EFAULT; 4270 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4271 4272 tmp = strchr(kbuf, ' '); 4273 if (!tmp) 4274 return -EINVAL; 4275 *tmp = '\0'; 4276 tmp++; 4277 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4278 return -EINVAL; 4279 4280 /* Find the cache in the chain of caches. */ 4281 mutex_lock(&cache_chain_mutex); 4282 res = -EINVAL; 4283 list_for_each_entry(cachep, &cache_chain, next) { 4284 if (!strcmp(cachep->name, kbuf)) { 4285 if (limit < 1 || batchcount < 1 || 4286 batchcount > limit || shared < 0) { 4287 res = 0; 4288 } else { 4289 res = do_tune_cpucache(cachep, limit, 4290 batchcount, shared); 4291 } 4292 break; 4293 } 4294 } 4295 mutex_unlock(&cache_chain_mutex); 4296 if (res >= 0) 4297 res = count; 4298 return res; 4299 } 4300 4301 #ifdef CONFIG_DEBUG_SLAB_LEAK 4302 4303 static void *leaks_start(struct seq_file *m, loff_t *pos) 4304 { 4305 loff_t n = *pos; 4306 struct list_head *p; 4307 4308 mutex_lock(&cache_chain_mutex); 4309 p = cache_chain.next; 4310 while (n--) { 4311 p = p->next; 4312 if (p == &cache_chain) 4313 return NULL; 4314 } 4315 return list_entry(p, struct kmem_cache, next); 4316 } 4317 4318 static inline int add_caller(unsigned long *n, unsigned long v) 4319 { 4320 unsigned long *p; 4321 int l; 4322 if (!v) 4323 return 1; 4324 l = n[1]; 4325 p = n + 2; 4326 while (l) { 4327 int i = l/2; 4328 unsigned long *q = p + 2 * i; 4329 if (*q == v) { 4330 q[1]++; 4331 return 1; 4332 } 4333 if (*q > v) { 4334 l = i; 4335 } else { 4336 p = q + 2; 4337 l -= i + 1; 4338 } 4339 } 4340 if (++n[1] == n[0]) 4341 return 0; 4342 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4343 p[0] = v; 4344 p[1] = 1; 4345 return 1; 4346 } 4347 4348 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4349 { 4350 void *p; 4351 int i; 4352 if (n[0] == n[1]) 4353 return; 4354 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4355 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4356 continue; 4357 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4358 return; 4359 } 4360 } 4361 4362 static void show_symbol(struct seq_file *m, unsigned long address) 4363 { 4364 #ifdef CONFIG_KALLSYMS 4365 char *modname; 4366 const char *name; 4367 unsigned long offset, size; 4368 char namebuf[KSYM_NAME_LEN+1]; 4369 4370 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); 4371 4372 if (name) { 4373 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4374 if (modname) 4375 seq_printf(m, " [%s]", modname); 4376 return; 4377 } 4378 #endif 4379 seq_printf(m, "%p", (void *)address); 4380 } 4381 4382 static int leaks_show(struct seq_file *m, void *p) 4383 { 4384 struct kmem_cache *cachep = p; 4385 struct slab *slabp; 4386 struct kmem_list3 *l3; 4387 const char *name; 4388 unsigned long *n = m->private; 4389 int node; 4390 int i; 4391 4392 if (!(cachep->flags & SLAB_STORE_USER)) 4393 return 0; 4394 if (!(cachep->flags & SLAB_RED_ZONE)) 4395 return 0; 4396 4397 /* OK, we can do it */ 4398 4399 n[1] = 0; 4400 4401 for_each_online_node(node) { 4402 l3 = cachep->nodelists[node]; 4403 if (!l3) 4404 continue; 4405 4406 check_irq_on(); 4407 spin_lock_irq(&l3->list_lock); 4408 4409 list_for_each_entry(slabp, &l3->slabs_full, list) 4410 handle_slab(n, cachep, slabp); 4411 list_for_each_entry(slabp, &l3->slabs_partial, list) 4412 handle_slab(n, cachep, slabp); 4413 spin_unlock_irq(&l3->list_lock); 4414 } 4415 name = cachep->name; 4416 if (n[0] == n[1]) { 4417 /* Increase the buffer size */ 4418 mutex_unlock(&cache_chain_mutex); 4419 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4420 if (!m->private) { 4421 /* Too bad, we are really out */ 4422 m->private = n; 4423 mutex_lock(&cache_chain_mutex); 4424 return -ENOMEM; 4425 } 4426 *(unsigned long *)m->private = n[0] * 2; 4427 kfree(n); 4428 mutex_lock(&cache_chain_mutex); 4429 /* Now make sure this entry will be retried */ 4430 m->count = m->size; 4431 return 0; 4432 } 4433 for (i = 0; i < n[1]; i++) { 4434 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4435 show_symbol(m, n[2*i+2]); 4436 seq_putc(m, '\n'); 4437 } 4438 4439 return 0; 4440 } 4441 4442 const struct seq_operations slabstats_op = { 4443 .start = leaks_start, 4444 .next = s_next, 4445 .stop = s_stop, 4446 .show = leaks_show, 4447 }; 4448 #endif 4449 #endif 4450 4451 /** 4452 * ksize - get the actual amount of memory allocated for a given object 4453 * @objp: Pointer to the object 4454 * 4455 * kmalloc may internally round up allocations and return more memory 4456 * than requested. ksize() can be used to determine the actual amount of 4457 * memory allocated. The caller may use this additional memory, even though 4458 * a smaller amount of memory was initially specified with the kmalloc call. 4459 * The caller must guarantee that objp points to a valid object previously 4460 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4461 * must not be freed during the duration of the call. 4462 */ 4463 unsigned int ksize(const void *objp) 4464 { 4465 if (unlikely(objp == NULL)) 4466 return 0; 4467 4468 return obj_size(virt_to_cache(objp)); 4469 } 4470