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 initializations 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/proc_fs.h> 99 #include <linux/seq_file.h> 100 #include <linux/notifier.h> 101 #include <linux/kallsyms.h> 102 #include <linux/cpu.h> 103 #include <linux/sysctl.h> 104 #include <linux/module.h> 105 #include <linux/rcupdate.h> 106 #include <linux/string.h> 107 #include <linux/uaccess.h> 108 #include <linux/nodemask.h> 109 #include <linux/kmemleak.h> 110 #include <linux/mempolicy.h> 111 #include <linux/mutex.h> 112 #include <linux/fault-inject.h> 113 #include <linux/rtmutex.h> 114 #include <linux/reciprocal_div.h> 115 #include <linux/debugobjects.h> 116 #include <linux/kmemcheck.h> 117 #include <linux/memory.h> 118 119 #include <asm/cacheflush.h> 120 #include <asm/tlbflush.h> 121 #include <asm/page.h> 122 123 /* 124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. 125 * 0 for faster, smaller code (especially in the critical paths). 126 * 127 * STATS - 1 to collect stats for /proc/slabinfo. 128 * 0 for faster, smaller code (especially in the critical paths). 129 * 130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 131 */ 132 133 #ifdef CONFIG_DEBUG_SLAB 134 #define DEBUG 1 135 #define STATS 1 136 #define FORCED_DEBUG 1 137 #else 138 #define DEBUG 0 139 #define STATS 0 140 #define FORCED_DEBUG 0 141 #endif 142 143 /* Shouldn't this be in a header file somewhere? */ 144 #define BYTES_PER_WORD sizeof(void *) 145 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) 146 147 #ifndef ARCH_KMALLOC_FLAGS 148 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 149 #endif 150 151 /* Legal flag mask for kmem_cache_create(). */ 152 #if DEBUG 153 # define CREATE_MASK (SLAB_RED_ZONE | \ 154 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 155 SLAB_CACHE_DMA | \ 156 SLAB_STORE_USER | \ 157 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 158 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ 159 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) 160 #else 161 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 162 SLAB_CACHE_DMA | \ 163 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 164 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ 165 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) 166 #endif 167 168 /* 169 * kmem_bufctl_t: 170 * 171 * Bufctl's are used for linking objs within a slab 172 * linked offsets. 173 * 174 * This implementation relies on "struct page" for locating the cache & 175 * slab an object belongs to. 176 * This allows the bufctl structure to be small (one int), but limits 177 * the number of objects a slab (not a cache) can contain when off-slab 178 * bufctls are used. The limit is the size of the largest general cache 179 * that does not use off-slab slabs. 180 * For 32bit archs with 4 kB pages, is this 56. 181 * This is not serious, as it is only for large objects, when it is unwise 182 * to have too many per slab. 183 * Note: This limit can be raised by introducing a general cache whose size 184 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 185 */ 186 187 typedef unsigned int kmem_bufctl_t; 188 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 189 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 190 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 191 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 192 193 /* 194 * struct slab_rcu 195 * 196 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 197 * arrange for kmem_freepages to be called via RCU. This is useful if 198 * we need to approach a kernel structure obliquely, from its address 199 * obtained without the usual locking. We can lock the structure to 200 * stabilize it and check it's still at the given address, only if we 201 * can be sure that the memory has not been meanwhile reused for some 202 * other kind of object (which our subsystem's lock might corrupt). 203 * 204 * rcu_read_lock before reading the address, then rcu_read_unlock after 205 * taking the spinlock within the structure expected at that address. 206 */ 207 struct slab_rcu { 208 struct rcu_head head; 209 struct kmem_cache *cachep; 210 void *addr; 211 }; 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 union { 222 struct { 223 struct list_head list; 224 unsigned long colouroff; 225 void *s_mem; /* including colour offset */ 226 unsigned int inuse; /* num of objs active in slab */ 227 kmem_bufctl_t free; 228 unsigned short nodeid; 229 }; 230 struct slab_rcu __slab_cover_slab_rcu; 231 }; 232 }; 233 234 /* 235 * struct array_cache 236 * 237 * Purpose: 238 * - LIFO ordering, to hand out cache-warm objects from _alloc 239 * - reduce the number of linked list operations 240 * - reduce spinlock operations 241 * 242 * The limit is stored in the per-cpu structure to reduce the data cache 243 * footprint. 244 * 245 */ 246 struct array_cache { 247 unsigned int avail; 248 unsigned int limit; 249 unsigned int batchcount; 250 unsigned int touched; 251 spinlock_t lock; 252 void *entry[]; /* 253 * Must have this definition in here for the proper 254 * alignment of array_cache. Also simplifies accessing 255 * the entries. 256 */ 257 }; 258 259 /* 260 * bootstrap: The caches do not work without cpuarrays anymore, but the 261 * cpuarrays are allocated from the generic caches... 262 */ 263 #define BOOT_CPUCACHE_ENTRIES 1 264 struct arraycache_init { 265 struct array_cache cache; 266 void *entries[BOOT_CPUCACHE_ENTRIES]; 267 }; 268 269 /* 270 * The slab lists for all objects. 271 */ 272 struct kmem_list3 { 273 struct list_head slabs_partial; /* partial list first, better asm code */ 274 struct list_head slabs_full; 275 struct list_head slabs_free; 276 unsigned long free_objects; 277 unsigned int free_limit; 278 unsigned int colour_next; /* Per-node cache coloring */ 279 spinlock_t list_lock; 280 struct array_cache *shared; /* shared per node */ 281 struct array_cache **alien; /* on other nodes */ 282 unsigned long next_reap; /* updated without locking */ 283 int free_touched; /* updated without locking */ 284 }; 285 286 /* 287 * Need this for bootstrapping a per node allocator. 288 */ 289 #define NUM_INIT_LISTS (3 * MAX_NUMNODES) 290 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 291 #define CACHE_CACHE 0 292 #define SIZE_AC MAX_NUMNODES 293 #define SIZE_L3 (2 * MAX_NUMNODES) 294 295 static int drain_freelist(struct kmem_cache *cache, 296 struct kmem_list3 *l3, int tofree); 297 static void free_block(struct kmem_cache *cachep, void **objpp, int len, 298 int node); 299 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); 300 static void cache_reap(struct work_struct *unused); 301 302 /* 303 * This function must be completely optimized away if a constant is passed to 304 * it. Mostly the same as what is in linux/slab.h except it returns an index. 305 */ 306 static __always_inline int index_of(const size_t size) 307 { 308 extern void __bad_size(void); 309 310 if (__builtin_constant_p(size)) { 311 int i = 0; 312 313 #define CACHE(x) \ 314 if (size <=x) \ 315 return i; \ 316 else \ 317 i++; 318 #include <linux/kmalloc_sizes.h> 319 #undef CACHE 320 __bad_size(); 321 } else 322 __bad_size(); 323 return 0; 324 } 325 326 static int slab_early_init = 1; 327 328 #define INDEX_AC index_of(sizeof(struct arraycache_init)) 329 #define INDEX_L3 index_of(sizeof(struct kmem_list3)) 330 331 static void kmem_list3_init(struct kmem_list3 *parent) 332 { 333 INIT_LIST_HEAD(&parent->slabs_full); 334 INIT_LIST_HEAD(&parent->slabs_partial); 335 INIT_LIST_HEAD(&parent->slabs_free); 336 parent->shared = NULL; 337 parent->alien = NULL; 338 parent->colour_next = 0; 339 spin_lock_init(&parent->list_lock); 340 parent->free_objects = 0; 341 parent->free_touched = 0; 342 } 343 344 #define MAKE_LIST(cachep, listp, slab, nodeid) \ 345 do { \ 346 INIT_LIST_HEAD(listp); \ 347 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 348 } while (0) 349 350 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 351 do { \ 352 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 353 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 354 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 355 } while (0) 356 357 #define CFLGS_OFF_SLAB (0x80000000UL) 358 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 359 360 #define BATCHREFILL_LIMIT 16 361 /* 362 * Optimization question: fewer reaps means less probability for unnessary 363 * cpucache drain/refill cycles. 364 * 365 * OTOH the cpuarrays can contain lots of objects, 366 * which could lock up otherwise freeable slabs. 367 */ 368 #define REAPTIMEOUT_CPUC (2*HZ) 369 #define REAPTIMEOUT_LIST3 (4*HZ) 370 371 #if STATS 372 #define STATS_INC_ACTIVE(x) ((x)->num_active++) 373 #define STATS_DEC_ACTIVE(x) ((x)->num_active--) 374 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 375 #define STATS_INC_GROWN(x) ((x)->grown++) 376 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 377 #define STATS_SET_HIGH(x) \ 378 do { \ 379 if ((x)->num_active > (x)->high_mark) \ 380 (x)->high_mark = (x)->num_active; \ 381 } while (0) 382 #define STATS_INC_ERR(x) ((x)->errors++) 383 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 384 #define STATS_INC_NODEFREES(x) ((x)->node_frees++) 385 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 386 #define STATS_SET_FREEABLE(x, i) \ 387 do { \ 388 if ((x)->max_freeable < i) \ 389 (x)->max_freeable = i; \ 390 } while (0) 391 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 392 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 393 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 394 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 395 #else 396 #define STATS_INC_ACTIVE(x) do { } while (0) 397 #define STATS_DEC_ACTIVE(x) do { } while (0) 398 #define STATS_INC_ALLOCED(x) do { } while (0) 399 #define STATS_INC_GROWN(x) do { } while (0) 400 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) 401 #define STATS_SET_HIGH(x) do { } while (0) 402 #define STATS_INC_ERR(x) do { } while (0) 403 #define STATS_INC_NODEALLOCS(x) do { } while (0) 404 #define STATS_INC_NODEFREES(x) do { } while (0) 405 #define STATS_INC_ACOVERFLOW(x) do { } while (0) 406 #define STATS_SET_FREEABLE(x, i) do { } while (0) 407 #define STATS_INC_ALLOCHIT(x) do { } while (0) 408 #define STATS_INC_ALLOCMISS(x) do { } while (0) 409 #define STATS_INC_FREEHIT(x) do { } while (0) 410 #define STATS_INC_FREEMISS(x) do { } while (0) 411 #endif 412 413 #if DEBUG 414 415 /* 416 * memory layout of objects: 417 * 0 : objp 418 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 419 * the end of an object is aligned with the end of the real 420 * allocation. Catches writes behind the end of the allocation. 421 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 422 * redzone word. 423 * cachep->obj_offset: The real object. 424 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 425 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 426 * [BYTES_PER_WORD long] 427 */ 428 static int obj_offset(struct kmem_cache *cachep) 429 { 430 return cachep->obj_offset; 431 } 432 433 static int obj_size(struct kmem_cache *cachep) 434 { 435 return cachep->obj_size; 436 } 437 438 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 439 { 440 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 441 return (unsigned long long*) (objp + obj_offset(cachep) - 442 sizeof(unsigned long long)); 443 } 444 445 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 446 { 447 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 448 if (cachep->flags & SLAB_STORE_USER) 449 return (unsigned long long *)(objp + cachep->buffer_size - 450 sizeof(unsigned long long) - 451 REDZONE_ALIGN); 452 return (unsigned long long *) (objp + cachep->buffer_size - 453 sizeof(unsigned long long)); 454 } 455 456 static void **dbg_userword(struct kmem_cache *cachep, void *objp) 457 { 458 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 459 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 460 } 461 462 #else 463 464 #define obj_offset(x) 0 465 #define obj_size(cachep) (cachep->buffer_size) 466 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 467 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 468 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 469 470 #endif 471 472 #ifdef CONFIG_TRACING 473 size_t slab_buffer_size(struct kmem_cache *cachep) 474 { 475 return cachep->buffer_size; 476 } 477 EXPORT_SYMBOL(slab_buffer_size); 478 #endif 479 480 /* 481 * Do not go above this order unless 0 objects fit into the slab. 482 */ 483 #define BREAK_GFP_ORDER_HI 1 484 #define BREAK_GFP_ORDER_LO 0 485 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 486 487 /* 488 * Functions for storing/retrieving the cachep and or slab from the page 489 * allocator. These are used to find the slab an obj belongs to. With kfree(), 490 * these are used to find the cache which an obj belongs to. 491 */ 492 static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 493 { 494 page->lru.next = (struct list_head *)cache; 495 } 496 497 static inline struct kmem_cache *page_get_cache(struct page *page) 498 { 499 page = compound_head(page); 500 BUG_ON(!PageSlab(page)); 501 return (struct kmem_cache *)page->lru.next; 502 } 503 504 static inline void page_set_slab(struct page *page, struct slab *slab) 505 { 506 page->lru.prev = (struct list_head *)slab; 507 } 508 509 static inline struct slab *page_get_slab(struct page *page) 510 { 511 BUG_ON(!PageSlab(page)); 512 return (struct slab *)page->lru.prev; 513 } 514 515 static inline struct kmem_cache *virt_to_cache(const void *obj) 516 { 517 struct page *page = virt_to_head_page(obj); 518 return page_get_cache(page); 519 } 520 521 static inline struct slab *virt_to_slab(const void *obj) 522 { 523 struct page *page = virt_to_head_page(obj); 524 return page_get_slab(page); 525 } 526 527 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 528 unsigned int idx) 529 { 530 return slab->s_mem + cache->buffer_size * idx; 531 } 532 533 /* 534 * We want to avoid an expensive divide : (offset / cache->buffer_size) 535 * Using the fact that buffer_size is a constant for a particular cache, 536 * we can replace (offset / cache->buffer_size) by 537 * reciprocal_divide(offset, cache->reciprocal_buffer_size) 538 */ 539 static inline unsigned int obj_to_index(const struct kmem_cache *cache, 540 const struct slab *slab, void *obj) 541 { 542 u32 offset = (obj - slab->s_mem); 543 return reciprocal_divide(offset, cache->reciprocal_buffer_size); 544 } 545 546 /* 547 * These are the default caches for kmalloc. Custom caches can have other sizes. 548 */ 549 struct cache_sizes malloc_sizes[] = { 550 #define CACHE(x) { .cs_size = (x) }, 551 #include <linux/kmalloc_sizes.h> 552 CACHE(ULONG_MAX) 553 #undef CACHE 554 }; 555 EXPORT_SYMBOL(malloc_sizes); 556 557 /* Must match cache_sizes above. Out of line to keep cache footprint low. */ 558 struct cache_names { 559 char *name; 560 char *name_dma; 561 }; 562 563 static struct cache_names __initdata cache_names[] = { 564 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 565 #include <linux/kmalloc_sizes.h> 566 {NULL,} 567 #undef CACHE 568 }; 569 570 static struct arraycache_init initarray_cache __initdata = 571 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 572 static struct arraycache_init initarray_generic = 573 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 574 575 /* internal cache of cache description objs */ 576 static struct kmem_cache cache_cache = { 577 .batchcount = 1, 578 .limit = BOOT_CPUCACHE_ENTRIES, 579 .shared = 1, 580 .buffer_size = sizeof(struct kmem_cache), 581 .name = "kmem_cache", 582 }; 583 584 #define BAD_ALIEN_MAGIC 0x01020304ul 585 586 /* 587 * chicken and egg problem: delay the per-cpu array allocation 588 * until the general caches are up. 589 */ 590 static enum { 591 NONE, 592 PARTIAL_AC, 593 PARTIAL_L3, 594 EARLY, 595 FULL 596 } g_cpucache_up; 597 598 /* 599 * used by boot code to determine if it can use slab based allocator 600 */ 601 int slab_is_available(void) 602 { 603 return g_cpucache_up >= EARLY; 604 } 605 606 #ifdef CONFIG_LOCKDEP 607 608 /* 609 * Slab sometimes uses the kmalloc slabs to store the slab headers 610 * for other slabs "off slab". 611 * The locking for this is tricky in that it nests within the locks 612 * of all other slabs in a few places; to deal with this special 613 * locking we put on-slab caches into a separate lock-class. 614 * 615 * We set lock class for alien array caches which are up during init. 616 * The lock annotation will be lost if all cpus of a node goes down and 617 * then comes back up during hotplug 618 */ 619 static struct lock_class_key on_slab_l3_key; 620 static struct lock_class_key on_slab_alc_key; 621 622 static void init_node_lock_keys(int q) 623 { 624 struct cache_sizes *s = malloc_sizes; 625 626 if (g_cpucache_up != FULL) 627 return; 628 629 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { 630 struct array_cache **alc; 631 struct kmem_list3 *l3; 632 int r; 633 634 l3 = s->cs_cachep->nodelists[q]; 635 if (!l3 || OFF_SLAB(s->cs_cachep)) 636 continue; 637 lockdep_set_class(&l3->list_lock, &on_slab_l3_key); 638 alc = l3->alien; 639 /* 640 * FIXME: This check for BAD_ALIEN_MAGIC 641 * should go away when common slab code is taught to 642 * work even without alien caches. 643 * Currently, non NUMA code returns BAD_ALIEN_MAGIC 644 * for alloc_alien_cache, 645 */ 646 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) 647 continue; 648 for_each_node(r) { 649 if (alc[r]) 650 lockdep_set_class(&alc[r]->lock, 651 &on_slab_alc_key); 652 } 653 } 654 } 655 656 static inline void init_lock_keys(void) 657 { 658 int node; 659 660 for_each_node(node) 661 init_node_lock_keys(node); 662 } 663 #else 664 static void init_node_lock_keys(int q) 665 { 666 } 667 668 static inline void init_lock_keys(void) 669 { 670 } 671 #endif 672 673 /* 674 * Guard access to the cache-chain. 675 */ 676 static DEFINE_MUTEX(cache_chain_mutex); 677 static struct list_head cache_chain; 678 679 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); 680 681 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 682 { 683 return cachep->array[smp_processor_id()]; 684 } 685 686 static inline struct kmem_cache *__find_general_cachep(size_t size, 687 gfp_t gfpflags) 688 { 689 struct cache_sizes *csizep = malloc_sizes; 690 691 #if DEBUG 692 /* This happens if someone tries to call 693 * kmem_cache_create(), or __kmalloc(), before 694 * the generic caches are initialized. 695 */ 696 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 697 #endif 698 if (!size) 699 return ZERO_SIZE_PTR; 700 701 while (size > csizep->cs_size) 702 csizep++; 703 704 /* 705 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 706 * has cs_{dma,}cachep==NULL. Thus no special case 707 * for large kmalloc calls required. 708 */ 709 #ifdef CONFIG_ZONE_DMA 710 if (unlikely(gfpflags & GFP_DMA)) 711 return csizep->cs_dmacachep; 712 #endif 713 return csizep->cs_cachep; 714 } 715 716 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 717 { 718 return __find_general_cachep(size, gfpflags); 719 } 720 721 static size_t slab_mgmt_size(size_t nr_objs, size_t align) 722 { 723 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 724 } 725 726 /* 727 * Calculate the number of objects and left-over bytes for a given buffer size. 728 */ 729 static void cache_estimate(unsigned long gfporder, size_t buffer_size, 730 size_t align, int flags, size_t *left_over, 731 unsigned int *num) 732 { 733 int nr_objs; 734 size_t mgmt_size; 735 size_t slab_size = PAGE_SIZE << gfporder; 736 737 /* 738 * The slab management structure can be either off the slab or 739 * on it. For the latter case, the memory allocated for a 740 * slab is used for: 741 * 742 * - The struct slab 743 * - One kmem_bufctl_t for each object 744 * - Padding to respect alignment of @align 745 * - @buffer_size bytes for each object 746 * 747 * If the slab management structure is off the slab, then the 748 * alignment will already be calculated into the size. Because 749 * the slabs are all pages aligned, the objects will be at the 750 * correct alignment when allocated. 751 */ 752 if (flags & CFLGS_OFF_SLAB) { 753 mgmt_size = 0; 754 nr_objs = slab_size / buffer_size; 755 756 if (nr_objs > SLAB_LIMIT) 757 nr_objs = SLAB_LIMIT; 758 } else { 759 /* 760 * Ignore padding for the initial guess. The padding 761 * is at most @align-1 bytes, and @buffer_size is at 762 * least @align. In the worst case, this result will 763 * be one greater than the number of objects that fit 764 * into the memory allocation when taking the padding 765 * into account. 766 */ 767 nr_objs = (slab_size - sizeof(struct slab)) / 768 (buffer_size + sizeof(kmem_bufctl_t)); 769 770 /* 771 * This calculated number will be either the right 772 * amount, or one greater than what we want. 773 */ 774 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 775 > slab_size) 776 nr_objs--; 777 778 if (nr_objs > SLAB_LIMIT) 779 nr_objs = SLAB_LIMIT; 780 781 mgmt_size = slab_mgmt_size(nr_objs, align); 782 } 783 *num = nr_objs; 784 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 785 } 786 787 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) 788 789 static void __slab_error(const char *function, struct kmem_cache *cachep, 790 char *msg) 791 { 792 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 793 function, cachep->name, msg); 794 dump_stack(); 795 } 796 797 /* 798 * By default on NUMA we use alien caches to stage the freeing of 799 * objects allocated from other nodes. This causes massive memory 800 * inefficiencies when using fake NUMA setup to split memory into a 801 * large number of small nodes, so it can be disabled on the command 802 * line 803 */ 804 805 static int use_alien_caches __read_mostly = 1; 806 static int __init noaliencache_setup(char *s) 807 { 808 use_alien_caches = 0; 809 return 1; 810 } 811 __setup("noaliencache", noaliencache_setup); 812 813 #ifdef CONFIG_NUMA 814 /* 815 * Special reaping functions for NUMA systems called from cache_reap(). 816 * These take care of doing round robin flushing of alien caches (containing 817 * objects freed on different nodes from which they were allocated) and the 818 * flushing of remote pcps by calling drain_node_pages. 819 */ 820 static DEFINE_PER_CPU(unsigned long, slab_reap_node); 821 822 static void init_reap_node(int cpu) 823 { 824 int node; 825 826 node = next_node(cpu_to_mem(cpu), node_online_map); 827 if (node == MAX_NUMNODES) 828 node = first_node(node_online_map); 829 830 per_cpu(slab_reap_node, cpu) = node; 831 } 832 833 static void next_reap_node(void) 834 { 835 int node = __this_cpu_read(slab_reap_node); 836 837 node = next_node(node, node_online_map); 838 if (unlikely(node >= MAX_NUMNODES)) 839 node = first_node(node_online_map); 840 __this_cpu_write(slab_reap_node, node); 841 } 842 843 #else 844 #define init_reap_node(cpu) do { } while (0) 845 #define next_reap_node(void) do { } while (0) 846 #endif 847 848 /* 849 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 850 * via the workqueue/eventd. 851 * Add the CPU number into the expiration time to minimize the possibility of 852 * the CPUs getting into lockstep and contending for the global cache chain 853 * lock. 854 */ 855 static void __cpuinit start_cpu_timer(int cpu) 856 { 857 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); 858 859 /* 860 * When this gets called from do_initcalls via cpucache_init(), 861 * init_workqueues() has already run, so keventd will be setup 862 * at that time. 863 */ 864 if (keventd_up() && reap_work->work.func == NULL) { 865 init_reap_node(cpu); 866 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); 867 schedule_delayed_work_on(cpu, reap_work, 868 __round_jiffies_relative(HZ, cpu)); 869 } 870 } 871 872 static struct array_cache *alloc_arraycache(int node, int entries, 873 int batchcount, gfp_t gfp) 874 { 875 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 876 struct array_cache *nc = NULL; 877 878 nc = kmalloc_node(memsize, gfp, node); 879 /* 880 * The array_cache structures contain pointers to free object. 881 * However, when such objects are allocated or transfered to another 882 * cache the pointers are not cleared and they could be counted as 883 * valid references during a kmemleak scan. Therefore, kmemleak must 884 * not scan such objects. 885 */ 886 kmemleak_no_scan(nc); 887 if (nc) { 888 nc->avail = 0; 889 nc->limit = entries; 890 nc->batchcount = batchcount; 891 nc->touched = 0; 892 spin_lock_init(&nc->lock); 893 } 894 return nc; 895 } 896 897 /* 898 * Transfer objects in one arraycache to another. 899 * Locking must be handled by the caller. 900 * 901 * Return the number of entries transferred. 902 */ 903 static int transfer_objects(struct array_cache *to, 904 struct array_cache *from, unsigned int max) 905 { 906 /* Figure out how many entries to transfer */ 907 int nr = min3(from->avail, max, to->limit - to->avail); 908 909 if (!nr) 910 return 0; 911 912 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 913 sizeof(void *) *nr); 914 915 from->avail -= nr; 916 to->avail += nr; 917 return nr; 918 } 919 920 #ifndef CONFIG_NUMA 921 922 #define drain_alien_cache(cachep, alien) do { } while (0) 923 #define reap_alien(cachep, l3) do { } while (0) 924 925 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) 926 { 927 return (struct array_cache **)BAD_ALIEN_MAGIC; 928 } 929 930 static inline void free_alien_cache(struct array_cache **ac_ptr) 931 { 932 } 933 934 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 935 { 936 return 0; 937 } 938 939 static inline void *alternate_node_alloc(struct kmem_cache *cachep, 940 gfp_t flags) 941 { 942 return NULL; 943 } 944 945 static inline void *____cache_alloc_node(struct kmem_cache *cachep, 946 gfp_t flags, int nodeid) 947 { 948 return NULL; 949 } 950 951 #else /* CONFIG_NUMA */ 952 953 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 954 static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 955 956 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) 957 { 958 struct array_cache **ac_ptr; 959 int memsize = sizeof(void *) * nr_node_ids; 960 int i; 961 962 if (limit > 1) 963 limit = 12; 964 ac_ptr = kzalloc_node(memsize, gfp, node); 965 if (ac_ptr) { 966 for_each_node(i) { 967 if (i == node || !node_online(i)) 968 continue; 969 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); 970 if (!ac_ptr[i]) { 971 for (i--; i >= 0; i--) 972 kfree(ac_ptr[i]); 973 kfree(ac_ptr); 974 return NULL; 975 } 976 } 977 } 978 return ac_ptr; 979 } 980 981 static void free_alien_cache(struct array_cache **ac_ptr) 982 { 983 int i; 984 985 if (!ac_ptr) 986 return; 987 for_each_node(i) 988 kfree(ac_ptr[i]); 989 kfree(ac_ptr); 990 } 991 992 static void __drain_alien_cache(struct kmem_cache *cachep, 993 struct array_cache *ac, int node) 994 { 995 struct kmem_list3 *rl3 = cachep->nodelists[node]; 996 997 if (ac->avail) { 998 spin_lock(&rl3->list_lock); 999 /* 1000 * Stuff objects into the remote nodes shared array first. 1001 * That way we could avoid the overhead of putting the objects 1002 * into the free lists and getting them back later. 1003 */ 1004 if (rl3->shared) 1005 transfer_objects(rl3->shared, ac, ac->limit); 1006 1007 free_block(cachep, ac->entry, ac->avail, node); 1008 ac->avail = 0; 1009 spin_unlock(&rl3->list_lock); 1010 } 1011 } 1012 1013 /* 1014 * Called from cache_reap() to regularly drain alien caches round robin. 1015 */ 1016 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 1017 { 1018 int node = __this_cpu_read(slab_reap_node); 1019 1020 if (l3->alien) { 1021 struct array_cache *ac = l3->alien[node]; 1022 1023 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1024 __drain_alien_cache(cachep, ac, node); 1025 spin_unlock_irq(&ac->lock); 1026 } 1027 } 1028 } 1029 1030 static void drain_alien_cache(struct kmem_cache *cachep, 1031 struct array_cache **alien) 1032 { 1033 int i = 0; 1034 struct array_cache *ac; 1035 unsigned long flags; 1036 1037 for_each_online_node(i) { 1038 ac = alien[i]; 1039 if (ac) { 1040 spin_lock_irqsave(&ac->lock, flags); 1041 __drain_alien_cache(cachep, ac, i); 1042 spin_unlock_irqrestore(&ac->lock, flags); 1043 } 1044 } 1045 } 1046 1047 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1048 { 1049 struct slab *slabp = virt_to_slab(objp); 1050 int nodeid = slabp->nodeid; 1051 struct kmem_list3 *l3; 1052 struct array_cache *alien = NULL; 1053 int node; 1054 1055 node = numa_mem_id(); 1056 1057 /* 1058 * Make sure we are not freeing a object from another node to the array 1059 * cache on this cpu. 1060 */ 1061 if (likely(slabp->nodeid == node)) 1062 return 0; 1063 1064 l3 = cachep->nodelists[node]; 1065 STATS_INC_NODEFREES(cachep); 1066 if (l3->alien && l3->alien[nodeid]) { 1067 alien = l3->alien[nodeid]; 1068 spin_lock(&alien->lock); 1069 if (unlikely(alien->avail == alien->limit)) { 1070 STATS_INC_ACOVERFLOW(cachep); 1071 __drain_alien_cache(cachep, alien, nodeid); 1072 } 1073 alien->entry[alien->avail++] = objp; 1074 spin_unlock(&alien->lock); 1075 } else { 1076 spin_lock(&(cachep->nodelists[nodeid])->list_lock); 1077 free_block(cachep, &objp, 1, nodeid); 1078 spin_unlock(&(cachep->nodelists[nodeid])->list_lock); 1079 } 1080 return 1; 1081 } 1082 #endif 1083 1084 /* 1085 * Allocates and initializes nodelists for a node on each slab cache, used for 1086 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 1087 * will be allocated off-node since memory is not yet online for the new node. 1088 * When hotplugging memory or a cpu, existing nodelists are not replaced if 1089 * already in use. 1090 * 1091 * Must hold cache_chain_mutex. 1092 */ 1093 static int init_cache_nodelists_node(int node) 1094 { 1095 struct kmem_cache *cachep; 1096 struct kmem_list3 *l3; 1097 const int memsize = sizeof(struct kmem_list3); 1098 1099 list_for_each_entry(cachep, &cache_chain, next) { 1100 /* 1101 * Set up the size64 kmemlist for cpu before we can 1102 * begin anything. Make sure some other cpu on this 1103 * node has not already allocated this 1104 */ 1105 if (!cachep->nodelists[node]) { 1106 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1107 if (!l3) 1108 return -ENOMEM; 1109 kmem_list3_init(l3); 1110 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1111 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1112 1113 /* 1114 * The l3s don't come and go as CPUs come and 1115 * go. cache_chain_mutex is sufficient 1116 * protection here. 1117 */ 1118 cachep->nodelists[node] = l3; 1119 } 1120 1121 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1122 cachep->nodelists[node]->free_limit = 1123 (1 + nr_cpus_node(node)) * 1124 cachep->batchcount + cachep->num; 1125 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1126 } 1127 return 0; 1128 } 1129 1130 static void __cpuinit cpuup_canceled(long cpu) 1131 { 1132 struct kmem_cache *cachep; 1133 struct kmem_list3 *l3 = NULL; 1134 int node = cpu_to_mem(cpu); 1135 const struct cpumask *mask = cpumask_of_node(node); 1136 1137 list_for_each_entry(cachep, &cache_chain, next) { 1138 struct array_cache *nc; 1139 struct array_cache *shared; 1140 struct array_cache **alien; 1141 1142 /* cpu is dead; no one can alloc from it. */ 1143 nc = cachep->array[cpu]; 1144 cachep->array[cpu] = NULL; 1145 l3 = cachep->nodelists[node]; 1146 1147 if (!l3) 1148 goto free_array_cache; 1149 1150 spin_lock_irq(&l3->list_lock); 1151 1152 /* Free limit for this kmem_list3 */ 1153 l3->free_limit -= cachep->batchcount; 1154 if (nc) 1155 free_block(cachep, nc->entry, nc->avail, node); 1156 1157 if (!cpumask_empty(mask)) { 1158 spin_unlock_irq(&l3->list_lock); 1159 goto free_array_cache; 1160 } 1161 1162 shared = l3->shared; 1163 if (shared) { 1164 free_block(cachep, shared->entry, 1165 shared->avail, node); 1166 l3->shared = NULL; 1167 } 1168 1169 alien = l3->alien; 1170 l3->alien = NULL; 1171 1172 spin_unlock_irq(&l3->list_lock); 1173 1174 kfree(shared); 1175 if (alien) { 1176 drain_alien_cache(cachep, alien); 1177 free_alien_cache(alien); 1178 } 1179 free_array_cache: 1180 kfree(nc); 1181 } 1182 /* 1183 * In the previous loop, all the objects were freed to 1184 * the respective cache's slabs, now we can go ahead and 1185 * shrink each nodelist to its limit. 1186 */ 1187 list_for_each_entry(cachep, &cache_chain, next) { 1188 l3 = cachep->nodelists[node]; 1189 if (!l3) 1190 continue; 1191 drain_freelist(cachep, l3, l3->free_objects); 1192 } 1193 } 1194 1195 static int __cpuinit cpuup_prepare(long cpu) 1196 { 1197 struct kmem_cache *cachep; 1198 struct kmem_list3 *l3 = NULL; 1199 int node = cpu_to_mem(cpu); 1200 int err; 1201 1202 /* 1203 * We need to do this right in the beginning since 1204 * alloc_arraycache's are going to use this list. 1205 * kmalloc_node allows us to add the slab to the right 1206 * kmem_list3 and not this cpu's kmem_list3 1207 */ 1208 err = init_cache_nodelists_node(node); 1209 if (err < 0) 1210 goto bad; 1211 1212 /* 1213 * Now we can go ahead with allocating the shared arrays and 1214 * array caches 1215 */ 1216 list_for_each_entry(cachep, &cache_chain, next) { 1217 struct array_cache *nc; 1218 struct array_cache *shared = NULL; 1219 struct array_cache **alien = NULL; 1220 1221 nc = alloc_arraycache(node, cachep->limit, 1222 cachep->batchcount, GFP_KERNEL); 1223 if (!nc) 1224 goto bad; 1225 if (cachep->shared) { 1226 shared = alloc_arraycache(node, 1227 cachep->shared * cachep->batchcount, 1228 0xbaadf00d, GFP_KERNEL); 1229 if (!shared) { 1230 kfree(nc); 1231 goto bad; 1232 } 1233 } 1234 if (use_alien_caches) { 1235 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); 1236 if (!alien) { 1237 kfree(shared); 1238 kfree(nc); 1239 goto bad; 1240 } 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 init_node_lock_keys(node); 1266 1267 return 0; 1268 bad: 1269 cpuup_canceled(cpu); 1270 return -ENOMEM; 1271 } 1272 1273 static int __cpuinit cpuup_callback(struct notifier_block *nfb, 1274 unsigned long action, void *hcpu) 1275 { 1276 long cpu = (long)hcpu; 1277 int err = 0; 1278 1279 switch (action) { 1280 case CPU_UP_PREPARE: 1281 case CPU_UP_PREPARE_FROZEN: 1282 mutex_lock(&cache_chain_mutex); 1283 err = cpuup_prepare(cpu); 1284 mutex_unlock(&cache_chain_mutex); 1285 break; 1286 case CPU_ONLINE: 1287 case CPU_ONLINE_FROZEN: 1288 start_cpu_timer(cpu); 1289 break; 1290 #ifdef CONFIG_HOTPLUG_CPU 1291 case CPU_DOWN_PREPARE: 1292 case CPU_DOWN_PREPARE_FROZEN: 1293 /* 1294 * Shutdown cache reaper. Note that the cache_chain_mutex is 1295 * held so that if cache_reap() is invoked it cannot do 1296 * anything expensive but will only modify reap_work 1297 * and reschedule the timer. 1298 */ 1299 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); 1300 /* Now the cache_reaper is guaranteed to be not running. */ 1301 per_cpu(slab_reap_work, cpu).work.func = NULL; 1302 break; 1303 case CPU_DOWN_FAILED: 1304 case CPU_DOWN_FAILED_FROZEN: 1305 start_cpu_timer(cpu); 1306 break; 1307 case CPU_DEAD: 1308 case CPU_DEAD_FROZEN: 1309 /* 1310 * Even if all the cpus of a node are down, we don't free the 1311 * kmem_list3 of any cache. This to avoid a race between 1312 * cpu_down, and a kmalloc allocation from another cpu for 1313 * memory from the node of the cpu going down. The list3 1314 * structure is usually allocated from kmem_cache_create() and 1315 * gets destroyed at kmem_cache_destroy(). 1316 */ 1317 /* fall through */ 1318 #endif 1319 case CPU_UP_CANCELED: 1320 case CPU_UP_CANCELED_FROZEN: 1321 mutex_lock(&cache_chain_mutex); 1322 cpuup_canceled(cpu); 1323 mutex_unlock(&cache_chain_mutex); 1324 break; 1325 } 1326 return notifier_from_errno(err); 1327 } 1328 1329 static struct notifier_block __cpuinitdata cpucache_notifier = { 1330 &cpuup_callback, NULL, 0 1331 }; 1332 1333 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 1334 /* 1335 * Drains freelist for a node on each slab cache, used for memory hot-remove. 1336 * Returns -EBUSY if all objects cannot be drained so that the node is not 1337 * removed. 1338 * 1339 * Must hold cache_chain_mutex. 1340 */ 1341 static int __meminit drain_cache_nodelists_node(int node) 1342 { 1343 struct kmem_cache *cachep; 1344 int ret = 0; 1345 1346 list_for_each_entry(cachep, &cache_chain, next) { 1347 struct kmem_list3 *l3; 1348 1349 l3 = cachep->nodelists[node]; 1350 if (!l3) 1351 continue; 1352 1353 drain_freelist(cachep, l3, l3->free_objects); 1354 1355 if (!list_empty(&l3->slabs_full) || 1356 !list_empty(&l3->slabs_partial)) { 1357 ret = -EBUSY; 1358 break; 1359 } 1360 } 1361 return ret; 1362 } 1363 1364 static int __meminit slab_memory_callback(struct notifier_block *self, 1365 unsigned long action, void *arg) 1366 { 1367 struct memory_notify *mnb = arg; 1368 int ret = 0; 1369 int nid; 1370 1371 nid = mnb->status_change_nid; 1372 if (nid < 0) 1373 goto out; 1374 1375 switch (action) { 1376 case MEM_GOING_ONLINE: 1377 mutex_lock(&cache_chain_mutex); 1378 ret = init_cache_nodelists_node(nid); 1379 mutex_unlock(&cache_chain_mutex); 1380 break; 1381 case MEM_GOING_OFFLINE: 1382 mutex_lock(&cache_chain_mutex); 1383 ret = drain_cache_nodelists_node(nid); 1384 mutex_unlock(&cache_chain_mutex); 1385 break; 1386 case MEM_ONLINE: 1387 case MEM_OFFLINE: 1388 case MEM_CANCEL_ONLINE: 1389 case MEM_CANCEL_OFFLINE: 1390 break; 1391 } 1392 out: 1393 return notifier_from_errno(ret); 1394 } 1395 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ 1396 1397 /* 1398 * swap the static kmem_list3 with kmalloced memory 1399 */ 1400 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1401 int nodeid) 1402 { 1403 struct kmem_list3 *ptr; 1404 1405 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); 1406 BUG_ON(!ptr); 1407 1408 memcpy(ptr, list, sizeof(struct kmem_list3)); 1409 /* 1410 * Do not assume that spinlocks can be initialized via memcpy: 1411 */ 1412 spin_lock_init(&ptr->list_lock); 1413 1414 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1415 cachep->nodelists[nodeid] = ptr; 1416 } 1417 1418 /* 1419 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1420 * size of kmem_list3. 1421 */ 1422 static void __init set_up_list3s(struct kmem_cache *cachep, int index) 1423 { 1424 int node; 1425 1426 for_each_online_node(node) { 1427 cachep->nodelists[node] = &initkmem_list3[index + node]; 1428 cachep->nodelists[node]->next_reap = jiffies + 1429 REAPTIMEOUT_LIST3 + 1430 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1431 } 1432 } 1433 1434 /* 1435 * Initialisation. Called after the page allocator have been initialised and 1436 * before smp_init(). 1437 */ 1438 void __init kmem_cache_init(void) 1439 { 1440 size_t left_over; 1441 struct cache_sizes *sizes; 1442 struct cache_names *names; 1443 int i; 1444 int order; 1445 int node; 1446 1447 if (num_possible_nodes() == 1) 1448 use_alien_caches = 0; 1449 1450 for (i = 0; i < NUM_INIT_LISTS; i++) { 1451 kmem_list3_init(&initkmem_list3[i]); 1452 if (i < MAX_NUMNODES) 1453 cache_cache.nodelists[i] = NULL; 1454 } 1455 set_up_list3s(&cache_cache, CACHE_CACHE); 1456 1457 /* 1458 * Fragmentation resistance on low memory - only use bigger 1459 * page orders on machines with more than 32MB of memory. 1460 */ 1461 if (totalram_pages > (32 << 20) >> PAGE_SHIFT) 1462 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1463 1464 /* Bootstrap is tricky, because several objects are allocated 1465 * from caches that do not exist yet: 1466 * 1) initialize the cache_cache cache: it contains the struct 1467 * kmem_cache structures of all caches, except cache_cache itself: 1468 * cache_cache is statically allocated. 1469 * Initially an __init data area is used for the head array and the 1470 * kmem_list3 structures, it's replaced with a kmalloc allocated 1471 * array at the end of the bootstrap. 1472 * 2) Create the first kmalloc cache. 1473 * The struct kmem_cache for the new cache is allocated normally. 1474 * An __init data area is used for the head array. 1475 * 3) Create the remaining kmalloc caches, with minimally sized 1476 * head arrays. 1477 * 4) Replace the __init data head arrays for cache_cache and the first 1478 * kmalloc cache with kmalloc allocated arrays. 1479 * 5) Replace the __init data for kmem_list3 for cache_cache and 1480 * the other cache's with kmalloc allocated memory. 1481 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1482 */ 1483 1484 node = numa_mem_id(); 1485 1486 /* 1) create the cache_cache */ 1487 INIT_LIST_HEAD(&cache_chain); 1488 list_add(&cache_cache.next, &cache_chain); 1489 cache_cache.colour_off = cache_line_size(); 1490 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1491 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; 1492 1493 /* 1494 * struct kmem_cache size depends on nr_node_ids, which 1495 * can be less than MAX_NUMNODES. 1496 */ 1497 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + 1498 nr_node_ids * sizeof(struct kmem_list3 *); 1499 #if DEBUG 1500 cache_cache.obj_size = cache_cache.buffer_size; 1501 #endif 1502 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1503 cache_line_size()); 1504 cache_cache.reciprocal_buffer_size = 1505 reciprocal_value(cache_cache.buffer_size); 1506 1507 for (order = 0; order < MAX_ORDER; order++) { 1508 cache_estimate(order, cache_cache.buffer_size, 1509 cache_line_size(), 0, &left_over, &cache_cache.num); 1510 if (cache_cache.num) 1511 break; 1512 } 1513 BUG_ON(!cache_cache.num); 1514 cache_cache.gfporder = order; 1515 cache_cache.colour = left_over / cache_cache.colour_off; 1516 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1517 sizeof(struct slab), cache_line_size()); 1518 1519 /* 2+3) create the kmalloc caches */ 1520 sizes = malloc_sizes; 1521 names = cache_names; 1522 1523 /* 1524 * Initialize the caches that provide memory for the array cache and the 1525 * kmem_list3 structures first. Without this, further allocations will 1526 * bug. 1527 */ 1528 1529 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1530 sizes[INDEX_AC].cs_size, 1531 ARCH_KMALLOC_MINALIGN, 1532 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1533 NULL); 1534 1535 if (INDEX_AC != INDEX_L3) { 1536 sizes[INDEX_L3].cs_cachep = 1537 kmem_cache_create(names[INDEX_L3].name, 1538 sizes[INDEX_L3].cs_size, 1539 ARCH_KMALLOC_MINALIGN, 1540 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1541 NULL); 1542 } 1543 1544 slab_early_init = 0; 1545 1546 while (sizes->cs_size != ULONG_MAX) { 1547 /* 1548 * For performance, all the general caches are L1 aligned. 1549 * This should be particularly beneficial on SMP boxes, as it 1550 * eliminates "false sharing". 1551 * Note for systems short on memory removing the alignment will 1552 * allow tighter packing of the smaller caches. 1553 */ 1554 if (!sizes->cs_cachep) { 1555 sizes->cs_cachep = kmem_cache_create(names->name, 1556 sizes->cs_size, 1557 ARCH_KMALLOC_MINALIGN, 1558 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1559 NULL); 1560 } 1561 #ifdef CONFIG_ZONE_DMA 1562 sizes->cs_dmacachep = kmem_cache_create( 1563 names->name_dma, 1564 sizes->cs_size, 1565 ARCH_KMALLOC_MINALIGN, 1566 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1567 SLAB_PANIC, 1568 NULL); 1569 #endif 1570 sizes++; 1571 names++; 1572 } 1573 /* 4) Replace the bootstrap head arrays */ 1574 { 1575 struct array_cache *ptr; 1576 1577 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); 1578 1579 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1580 memcpy(ptr, cpu_cache_get(&cache_cache), 1581 sizeof(struct arraycache_init)); 1582 /* 1583 * Do not assume that spinlocks can be initialized via memcpy: 1584 */ 1585 spin_lock_init(&ptr->lock); 1586 1587 cache_cache.array[smp_processor_id()] = ptr; 1588 1589 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); 1590 1591 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1592 != &initarray_generic.cache); 1593 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1594 sizeof(struct arraycache_init)); 1595 /* 1596 * Do not assume that spinlocks can be initialized via memcpy: 1597 */ 1598 spin_lock_init(&ptr->lock); 1599 1600 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1601 ptr; 1602 } 1603 /* 5) Replace the bootstrap kmem_list3's */ 1604 { 1605 int nid; 1606 1607 for_each_online_node(nid) { 1608 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); 1609 1610 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1611 &initkmem_list3[SIZE_AC + nid], nid); 1612 1613 if (INDEX_AC != INDEX_L3) { 1614 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1615 &initkmem_list3[SIZE_L3 + nid], nid); 1616 } 1617 } 1618 } 1619 1620 g_cpucache_up = EARLY; 1621 } 1622 1623 void __init kmem_cache_init_late(void) 1624 { 1625 struct kmem_cache *cachep; 1626 1627 /* 6) resize the head arrays to their final sizes */ 1628 mutex_lock(&cache_chain_mutex); 1629 list_for_each_entry(cachep, &cache_chain, next) 1630 if (enable_cpucache(cachep, GFP_NOWAIT)) 1631 BUG(); 1632 mutex_unlock(&cache_chain_mutex); 1633 1634 /* Done! */ 1635 g_cpucache_up = FULL; 1636 1637 /* Annotate slab for lockdep -- annotate the malloc caches */ 1638 init_lock_keys(); 1639 1640 /* 1641 * Register a cpu startup notifier callback that initializes 1642 * cpu_cache_get for all new cpus 1643 */ 1644 register_cpu_notifier(&cpucache_notifier); 1645 1646 #ifdef CONFIG_NUMA 1647 /* 1648 * Register a memory hotplug callback that initializes and frees 1649 * nodelists. 1650 */ 1651 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 1652 #endif 1653 1654 /* 1655 * The reap timers are started later, with a module init call: That part 1656 * of the kernel is not yet operational. 1657 */ 1658 } 1659 1660 static int __init cpucache_init(void) 1661 { 1662 int cpu; 1663 1664 /* 1665 * Register the timers that return unneeded pages to the page allocator 1666 */ 1667 for_each_online_cpu(cpu) 1668 start_cpu_timer(cpu); 1669 return 0; 1670 } 1671 __initcall(cpucache_init); 1672 1673 /* 1674 * Interface to system's page allocator. No need to hold the cache-lock. 1675 * 1676 * If we requested dmaable memory, we will get it. Even if we 1677 * did not request dmaable memory, we might get it, but that 1678 * would be relatively rare and ignorable. 1679 */ 1680 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1681 { 1682 struct page *page; 1683 int nr_pages; 1684 int i; 1685 1686 #ifndef CONFIG_MMU 1687 /* 1688 * Nommu uses slab's for process anonymous memory allocations, and thus 1689 * requires __GFP_COMP to properly refcount higher order allocations 1690 */ 1691 flags |= __GFP_COMP; 1692 #endif 1693 1694 flags |= cachep->gfpflags; 1695 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1696 flags |= __GFP_RECLAIMABLE; 1697 1698 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); 1699 if (!page) 1700 return NULL; 1701 1702 nr_pages = (1 << cachep->gfporder); 1703 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1704 add_zone_page_state(page_zone(page), 1705 NR_SLAB_RECLAIMABLE, nr_pages); 1706 else 1707 add_zone_page_state(page_zone(page), 1708 NR_SLAB_UNRECLAIMABLE, nr_pages); 1709 for (i = 0; i < nr_pages; i++) 1710 __SetPageSlab(page + i); 1711 1712 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { 1713 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); 1714 1715 if (cachep->ctor) 1716 kmemcheck_mark_uninitialized_pages(page, nr_pages); 1717 else 1718 kmemcheck_mark_unallocated_pages(page, nr_pages); 1719 } 1720 1721 return page_address(page); 1722 } 1723 1724 /* 1725 * Interface to system's page release. 1726 */ 1727 static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1728 { 1729 unsigned long i = (1 << cachep->gfporder); 1730 struct page *page = virt_to_page(addr); 1731 const unsigned long nr_freed = i; 1732 1733 kmemcheck_free_shadow(page, cachep->gfporder); 1734 1735 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1736 sub_zone_page_state(page_zone(page), 1737 NR_SLAB_RECLAIMABLE, nr_freed); 1738 else 1739 sub_zone_page_state(page_zone(page), 1740 NR_SLAB_UNRECLAIMABLE, nr_freed); 1741 while (i--) { 1742 BUG_ON(!PageSlab(page)); 1743 __ClearPageSlab(page); 1744 page++; 1745 } 1746 if (current->reclaim_state) 1747 current->reclaim_state->reclaimed_slab += nr_freed; 1748 free_pages((unsigned long)addr, cachep->gfporder); 1749 } 1750 1751 static void kmem_rcu_free(struct rcu_head *head) 1752 { 1753 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1754 struct kmem_cache *cachep = slab_rcu->cachep; 1755 1756 kmem_freepages(cachep, slab_rcu->addr); 1757 if (OFF_SLAB(cachep)) 1758 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1759 } 1760 1761 #if DEBUG 1762 1763 #ifdef CONFIG_DEBUG_PAGEALLOC 1764 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1765 unsigned long caller) 1766 { 1767 int size = obj_size(cachep); 1768 1769 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1770 1771 if (size < 5 * sizeof(unsigned long)) 1772 return; 1773 1774 *addr++ = 0x12345678; 1775 *addr++ = caller; 1776 *addr++ = smp_processor_id(); 1777 size -= 3 * sizeof(unsigned long); 1778 { 1779 unsigned long *sptr = &caller; 1780 unsigned long svalue; 1781 1782 while (!kstack_end(sptr)) { 1783 svalue = *sptr++; 1784 if (kernel_text_address(svalue)) { 1785 *addr++ = svalue; 1786 size -= sizeof(unsigned long); 1787 if (size <= sizeof(unsigned long)) 1788 break; 1789 } 1790 } 1791 1792 } 1793 *addr++ = 0x87654321; 1794 } 1795 #endif 1796 1797 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1798 { 1799 int size = obj_size(cachep); 1800 addr = &((char *)addr)[obj_offset(cachep)]; 1801 1802 memset(addr, val, size); 1803 *(unsigned char *)(addr + size - 1) = POISON_END; 1804 } 1805 1806 static void dump_line(char *data, int offset, int limit) 1807 { 1808 int i; 1809 unsigned char error = 0; 1810 int bad_count = 0; 1811 1812 printk(KERN_ERR "%03x:", offset); 1813 for (i = 0; i < limit; i++) { 1814 if (data[offset + i] != POISON_FREE) { 1815 error = data[offset + i]; 1816 bad_count++; 1817 } 1818 printk(" %02x", (unsigned char)data[offset + i]); 1819 } 1820 printk("\n"); 1821 1822 if (bad_count == 1) { 1823 error ^= POISON_FREE; 1824 if (!(error & (error - 1))) { 1825 printk(KERN_ERR "Single bit error detected. Probably " 1826 "bad RAM.\n"); 1827 #ifdef CONFIG_X86 1828 printk(KERN_ERR "Run memtest86+ or a similar memory " 1829 "test tool.\n"); 1830 #else 1831 printk(KERN_ERR "Run a memory test tool.\n"); 1832 #endif 1833 } 1834 } 1835 } 1836 #endif 1837 1838 #if DEBUG 1839 1840 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1841 { 1842 int i, size; 1843 char *realobj; 1844 1845 if (cachep->flags & SLAB_RED_ZONE) { 1846 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", 1847 *dbg_redzone1(cachep, objp), 1848 *dbg_redzone2(cachep, objp)); 1849 } 1850 1851 if (cachep->flags & SLAB_STORE_USER) { 1852 printk(KERN_ERR "Last user: [<%p>]", 1853 *dbg_userword(cachep, objp)); 1854 print_symbol("(%s)", 1855 (unsigned long)*dbg_userword(cachep, objp)); 1856 printk("\n"); 1857 } 1858 realobj = (char *)objp + obj_offset(cachep); 1859 size = obj_size(cachep); 1860 for (i = 0; i < size && lines; i += 16, lines--) { 1861 int limit; 1862 limit = 16; 1863 if (i + limit > size) 1864 limit = size - i; 1865 dump_line(realobj, i, limit); 1866 } 1867 } 1868 1869 static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1870 { 1871 char *realobj; 1872 int size, i; 1873 int lines = 0; 1874 1875 realobj = (char *)objp + obj_offset(cachep); 1876 size = obj_size(cachep); 1877 1878 for (i = 0; i < size; i++) { 1879 char exp = POISON_FREE; 1880 if (i == size - 1) 1881 exp = POISON_END; 1882 if (realobj[i] != exp) { 1883 int limit; 1884 /* Mismatch ! */ 1885 /* Print header */ 1886 if (lines == 0) { 1887 printk(KERN_ERR 1888 "Slab corruption: %s start=%p, len=%d\n", 1889 cachep->name, realobj, size); 1890 print_objinfo(cachep, objp, 0); 1891 } 1892 /* Hexdump the affected line */ 1893 i = (i / 16) * 16; 1894 limit = 16; 1895 if (i + limit > size) 1896 limit = size - i; 1897 dump_line(realobj, i, limit); 1898 i += 16; 1899 lines++; 1900 /* Limit to 5 lines */ 1901 if (lines > 5) 1902 break; 1903 } 1904 } 1905 if (lines != 0) { 1906 /* Print some data about the neighboring objects, if they 1907 * exist: 1908 */ 1909 struct slab *slabp = virt_to_slab(objp); 1910 unsigned int objnr; 1911 1912 objnr = obj_to_index(cachep, slabp, objp); 1913 if (objnr) { 1914 objp = index_to_obj(cachep, slabp, objnr - 1); 1915 realobj = (char *)objp + obj_offset(cachep); 1916 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1917 realobj, size); 1918 print_objinfo(cachep, objp, 2); 1919 } 1920 if (objnr + 1 < cachep->num) { 1921 objp = index_to_obj(cachep, slabp, objnr + 1); 1922 realobj = (char *)objp + obj_offset(cachep); 1923 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1924 realobj, size); 1925 print_objinfo(cachep, objp, 2); 1926 } 1927 } 1928 } 1929 #endif 1930 1931 #if DEBUG 1932 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) 1933 { 1934 int i; 1935 for (i = 0; i < cachep->num; i++) { 1936 void *objp = index_to_obj(cachep, slabp, i); 1937 1938 if (cachep->flags & SLAB_POISON) { 1939 #ifdef CONFIG_DEBUG_PAGEALLOC 1940 if (cachep->buffer_size % PAGE_SIZE == 0 && 1941 OFF_SLAB(cachep)) 1942 kernel_map_pages(virt_to_page(objp), 1943 cachep->buffer_size / PAGE_SIZE, 1); 1944 else 1945 check_poison_obj(cachep, objp); 1946 #else 1947 check_poison_obj(cachep, objp); 1948 #endif 1949 } 1950 if (cachep->flags & SLAB_RED_ZONE) { 1951 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1952 slab_error(cachep, "start of a freed object " 1953 "was overwritten"); 1954 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1955 slab_error(cachep, "end of a freed object " 1956 "was overwritten"); 1957 } 1958 } 1959 } 1960 #else 1961 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) 1962 { 1963 } 1964 #endif 1965 1966 /** 1967 * slab_destroy - destroy and release all objects in a slab 1968 * @cachep: cache pointer being destroyed 1969 * @slabp: slab pointer being destroyed 1970 * 1971 * Destroy all the objs in a slab, and release the mem back to the system. 1972 * Before calling the slab must have been unlinked from the cache. The 1973 * cache-lock is not held/needed. 1974 */ 1975 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1976 { 1977 void *addr = slabp->s_mem - slabp->colouroff; 1978 1979 slab_destroy_debugcheck(cachep, slabp); 1980 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1981 struct slab_rcu *slab_rcu; 1982 1983 slab_rcu = (struct slab_rcu *)slabp; 1984 slab_rcu->cachep = cachep; 1985 slab_rcu->addr = addr; 1986 call_rcu(&slab_rcu->head, kmem_rcu_free); 1987 } else { 1988 kmem_freepages(cachep, addr); 1989 if (OFF_SLAB(cachep)) 1990 kmem_cache_free(cachep->slabp_cache, slabp); 1991 } 1992 } 1993 1994 static void __kmem_cache_destroy(struct kmem_cache *cachep) 1995 { 1996 int i; 1997 struct kmem_list3 *l3; 1998 1999 for_each_online_cpu(i) 2000 kfree(cachep->array[i]); 2001 2002 /* NUMA: free the list3 structures */ 2003 for_each_online_node(i) { 2004 l3 = cachep->nodelists[i]; 2005 if (l3) { 2006 kfree(l3->shared); 2007 free_alien_cache(l3->alien); 2008 kfree(l3); 2009 } 2010 } 2011 kmem_cache_free(&cache_cache, cachep); 2012 } 2013 2014 2015 /** 2016 * calculate_slab_order - calculate size (page order) of slabs 2017 * @cachep: pointer to the cache that is being created 2018 * @size: size of objects to be created in this cache. 2019 * @align: required alignment for the objects. 2020 * @flags: slab allocation flags 2021 * 2022 * Also calculates the number of objects per slab. 2023 * 2024 * This could be made much more intelligent. For now, try to avoid using 2025 * high order pages for slabs. When the gfp() functions are more friendly 2026 * towards high-order requests, this should be changed. 2027 */ 2028 static size_t calculate_slab_order(struct kmem_cache *cachep, 2029 size_t size, size_t align, unsigned long flags) 2030 { 2031 unsigned long offslab_limit; 2032 size_t left_over = 0; 2033 int gfporder; 2034 2035 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { 2036 unsigned int num; 2037 size_t remainder; 2038 2039 cache_estimate(gfporder, size, align, flags, &remainder, &num); 2040 if (!num) 2041 continue; 2042 2043 if (flags & CFLGS_OFF_SLAB) { 2044 /* 2045 * Max number of objs-per-slab for caches which 2046 * use off-slab slabs. Needed to avoid a possible 2047 * looping condition in cache_grow(). 2048 */ 2049 offslab_limit = size - sizeof(struct slab); 2050 offslab_limit /= sizeof(kmem_bufctl_t); 2051 2052 if (num > offslab_limit) 2053 break; 2054 } 2055 2056 /* Found something acceptable - save it away */ 2057 cachep->num = num; 2058 cachep->gfporder = gfporder; 2059 left_over = remainder; 2060 2061 /* 2062 * A VFS-reclaimable slab tends to have most allocations 2063 * as GFP_NOFS and we really don't want to have to be allocating 2064 * higher-order pages when we are unable to shrink dcache. 2065 */ 2066 if (flags & SLAB_RECLAIM_ACCOUNT) 2067 break; 2068 2069 /* 2070 * Large number of objects is good, but very large slabs are 2071 * currently bad for the gfp()s. 2072 */ 2073 if (gfporder >= slab_break_gfp_order) 2074 break; 2075 2076 /* 2077 * Acceptable internal fragmentation? 2078 */ 2079 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 2080 break; 2081 } 2082 return left_over; 2083 } 2084 2085 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) 2086 { 2087 if (g_cpucache_up == FULL) 2088 return enable_cpucache(cachep, gfp); 2089 2090 if (g_cpucache_up == NONE) { 2091 /* 2092 * Note: the first kmem_cache_create must create the cache 2093 * that's used by kmalloc(24), otherwise the creation of 2094 * further caches will BUG(). 2095 */ 2096 cachep->array[smp_processor_id()] = &initarray_generic.cache; 2097 2098 /* 2099 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 2100 * the first cache, then we need to set up all its list3s, 2101 * otherwise the creation of further caches will BUG(). 2102 */ 2103 set_up_list3s(cachep, SIZE_AC); 2104 if (INDEX_AC == INDEX_L3) 2105 g_cpucache_up = PARTIAL_L3; 2106 else 2107 g_cpucache_up = PARTIAL_AC; 2108 } else { 2109 cachep->array[smp_processor_id()] = 2110 kmalloc(sizeof(struct arraycache_init), gfp); 2111 2112 if (g_cpucache_up == PARTIAL_AC) { 2113 set_up_list3s(cachep, SIZE_L3); 2114 g_cpucache_up = PARTIAL_L3; 2115 } else { 2116 int node; 2117 for_each_online_node(node) { 2118 cachep->nodelists[node] = 2119 kmalloc_node(sizeof(struct kmem_list3), 2120 gfp, node); 2121 BUG_ON(!cachep->nodelists[node]); 2122 kmem_list3_init(cachep->nodelists[node]); 2123 } 2124 } 2125 } 2126 cachep->nodelists[numa_mem_id()]->next_reap = 2127 jiffies + REAPTIMEOUT_LIST3 + 2128 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 2129 2130 cpu_cache_get(cachep)->avail = 0; 2131 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 2132 cpu_cache_get(cachep)->batchcount = 1; 2133 cpu_cache_get(cachep)->touched = 0; 2134 cachep->batchcount = 1; 2135 cachep->limit = BOOT_CPUCACHE_ENTRIES; 2136 return 0; 2137 } 2138 2139 /** 2140 * kmem_cache_create - Create a cache. 2141 * @name: A string which is used in /proc/slabinfo to identify this cache. 2142 * @size: The size of objects to be created in this cache. 2143 * @align: The required alignment for the objects. 2144 * @flags: SLAB flags 2145 * @ctor: A constructor for the objects. 2146 * 2147 * Returns a ptr to the cache on success, NULL on failure. 2148 * Cannot be called within a int, but can be interrupted. 2149 * The @ctor is run when new pages are allocated by the cache. 2150 * 2151 * @name must be valid until the cache is destroyed. This implies that 2152 * the module calling this has to destroy the cache before getting unloaded. 2153 * 2154 * The flags are 2155 * 2156 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 2157 * to catch references to uninitialised memory. 2158 * 2159 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 2160 * for buffer overruns. 2161 * 2162 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 2163 * cacheline. This can be beneficial if you're counting cycles as closely 2164 * as davem. 2165 */ 2166 struct kmem_cache * 2167 kmem_cache_create (const char *name, size_t size, size_t align, 2168 unsigned long flags, void (*ctor)(void *)) 2169 { 2170 size_t left_over, slab_size, ralign; 2171 struct kmem_cache *cachep = NULL, *pc; 2172 gfp_t gfp; 2173 2174 /* 2175 * Sanity checks... these are all serious usage bugs. 2176 */ 2177 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 2178 size > KMALLOC_MAX_SIZE) { 2179 printk(KERN_ERR "%s: Early error in slab %s\n", __func__, 2180 name); 2181 BUG(); 2182 } 2183 2184 /* 2185 * We use cache_chain_mutex to ensure a consistent view of 2186 * cpu_online_mask as well. Please see cpuup_callback 2187 */ 2188 if (slab_is_available()) { 2189 get_online_cpus(); 2190 mutex_lock(&cache_chain_mutex); 2191 } 2192 2193 list_for_each_entry(pc, &cache_chain, next) { 2194 char tmp; 2195 int res; 2196 2197 /* 2198 * This happens when the module gets unloaded and doesn't 2199 * destroy its slab cache and no-one else reuses the vmalloc 2200 * area of the module. Print a warning. 2201 */ 2202 res = probe_kernel_address(pc->name, tmp); 2203 if (res) { 2204 printk(KERN_ERR 2205 "SLAB: cache with size %d has lost its name\n", 2206 pc->buffer_size); 2207 continue; 2208 } 2209 2210 if (!strcmp(pc->name, name)) { 2211 printk(KERN_ERR 2212 "kmem_cache_create: duplicate cache %s\n", name); 2213 dump_stack(); 2214 goto oops; 2215 } 2216 } 2217 2218 #if DEBUG 2219 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 2220 #if FORCED_DEBUG 2221 /* 2222 * Enable redzoning and last user accounting, except for caches with 2223 * large objects, if the increased size would increase the object size 2224 * above the next power of two: caches with object sizes just above a 2225 * power of two have a significant amount of internal fragmentation. 2226 */ 2227 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + 2228 2 * sizeof(unsigned long long))) 2229 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2230 if (!(flags & SLAB_DESTROY_BY_RCU)) 2231 flags |= SLAB_POISON; 2232 #endif 2233 if (flags & SLAB_DESTROY_BY_RCU) 2234 BUG_ON(flags & SLAB_POISON); 2235 #endif 2236 /* 2237 * Always checks flags, a caller might be expecting debug support which 2238 * isn't available. 2239 */ 2240 BUG_ON(flags & ~CREATE_MASK); 2241 2242 /* 2243 * Check that size is in terms of words. This is needed to avoid 2244 * unaligned accesses for some archs when redzoning is used, and makes 2245 * sure any on-slab bufctl's are also correctly aligned. 2246 */ 2247 if (size & (BYTES_PER_WORD - 1)) { 2248 size += (BYTES_PER_WORD - 1); 2249 size &= ~(BYTES_PER_WORD - 1); 2250 } 2251 2252 /* calculate the final buffer alignment: */ 2253 2254 /* 1) arch recommendation: can be overridden for debug */ 2255 if (flags & SLAB_HWCACHE_ALIGN) { 2256 /* 2257 * Default alignment: as specified by the arch code. Except if 2258 * an object is really small, then squeeze multiple objects into 2259 * one cacheline. 2260 */ 2261 ralign = cache_line_size(); 2262 while (size <= ralign / 2) 2263 ralign /= 2; 2264 } else { 2265 ralign = BYTES_PER_WORD; 2266 } 2267 2268 /* 2269 * Redzoning and user store require word alignment or possibly larger. 2270 * Note this will be overridden by architecture or caller mandated 2271 * alignment if either is greater than BYTES_PER_WORD. 2272 */ 2273 if (flags & SLAB_STORE_USER) 2274 ralign = BYTES_PER_WORD; 2275 2276 if (flags & SLAB_RED_ZONE) { 2277 ralign = REDZONE_ALIGN; 2278 /* If redzoning, ensure that the second redzone is suitably 2279 * aligned, by adjusting the object size accordingly. */ 2280 size += REDZONE_ALIGN - 1; 2281 size &= ~(REDZONE_ALIGN - 1); 2282 } 2283 2284 /* 2) arch mandated alignment */ 2285 if (ralign < ARCH_SLAB_MINALIGN) { 2286 ralign = ARCH_SLAB_MINALIGN; 2287 } 2288 /* 3) caller mandated alignment */ 2289 if (ralign < align) { 2290 ralign = align; 2291 } 2292 /* disable debug if necessary */ 2293 if (ralign > __alignof__(unsigned long long)) 2294 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2295 /* 2296 * 4) Store it. 2297 */ 2298 align = ralign; 2299 2300 if (slab_is_available()) 2301 gfp = GFP_KERNEL; 2302 else 2303 gfp = GFP_NOWAIT; 2304 2305 /* Get cache's description obj. */ 2306 cachep = kmem_cache_zalloc(&cache_cache, gfp); 2307 if (!cachep) 2308 goto oops; 2309 2310 #if DEBUG 2311 cachep->obj_size = size; 2312 2313 /* 2314 * Both debugging options require word-alignment which is calculated 2315 * into align above. 2316 */ 2317 if (flags & SLAB_RED_ZONE) { 2318 /* add space for red zone words */ 2319 cachep->obj_offset += sizeof(unsigned long long); 2320 size += 2 * sizeof(unsigned long long); 2321 } 2322 if (flags & SLAB_STORE_USER) { 2323 /* user store requires one word storage behind the end of 2324 * the real object. But if the second red zone needs to be 2325 * aligned to 64 bits, we must allow that much space. 2326 */ 2327 if (flags & SLAB_RED_ZONE) 2328 size += REDZONE_ALIGN; 2329 else 2330 size += BYTES_PER_WORD; 2331 } 2332 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2333 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2334 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { 2335 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); 2336 size = PAGE_SIZE; 2337 } 2338 #endif 2339 #endif 2340 2341 /* 2342 * Determine if the slab management is 'on' or 'off' slab. 2343 * (bootstrapping cannot cope with offslab caches so don't do 2344 * it too early on. Always use on-slab management when 2345 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) 2346 */ 2347 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && 2348 !(flags & SLAB_NOLEAKTRACE)) 2349 /* 2350 * Size is large, assume best to place the slab management obj 2351 * off-slab (should allow better packing of objs). 2352 */ 2353 flags |= CFLGS_OFF_SLAB; 2354 2355 size = ALIGN(size, align); 2356 2357 left_over = calculate_slab_order(cachep, size, align, flags); 2358 2359 if (!cachep->num) { 2360 printk(KERN_ERR 2361 "kmem_cache_create: couldn't create cache %s.\n", name); 2362 kmem_cache_free(&cache_cache, cachep); 2363 cachep = NULL; 2364 goto oops; 2365 } 2366 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2367 + sizeof(struct slab), align); 2368 2369 /* 2370 * If the slab has been placed off-slab, and we have enough space then 2371 * move it on-slab. This is at the expense of any extra colouring. 2372 */ 2373 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2374 flags &= ~CFLGS_OFF_SLAB; 2375 left_over -= slab_size; 2376 } 2377 2378 if (flags & CFLGS_OFF_SLAB) { 2379 /* really off slab. No need for manual alignment */ 2380 slab_size = 2381 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2382 2383 #ifdef CONFIG_PAGE_POISONING 2384 /* If we're going to use the generic kernel_map_pages() 2385 * poisoning, then it's going to smash the contents of 2386 * the redzone and userword anyhow, so switch them off. 2387 */ 2388 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) 2389 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2390 #endif 2391 } 2392 2393 cachep->colour_off = cache_line_size(); 2394 /* Offset must be a multiple of the alignment. */ 2395 if (cachep->colour_off < align) 2396 cachep->colour_off = align; 2397 cachep->colour = left_over / cachep->colour_off; 2398 cachep->slab_size = slab_size; 2399 cachep->flags = flags; 2400 cachep->gfpflags = 0; 2401 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) 2402 cachep->gfpflags |= GFP_DMA; 2403 cachep->buffer_size = size; 2404 cachep->reciprocal_buffer_size = reciprocal_value(size); 2405 2406 if (flags & CFLGS_OFF_SLAB) { 2407 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2408 /* 2409 * This is a possibility for one of the malloc_sizes caches. 2410 * But since we go off slab only for object size greater than 2411 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, 2412 * this should not happen at all. 2413 * But leave a BUG_ON for some lucky dude. 2414 */ 2415 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); 2416 } 2417 cachep->ctor = ctor; 2418 cachep->name = name; 2419 2420 if (setup_cpu_cache(cachep, gfp)) { 2421 __kmem_cache_destroy(cachep); 2422 cachep = NULL; 2423 goto oops; 2424 } 2425 2426 /* cache setup completed, link it into the list */ 2427 list_add(&cachep->next, &cache_chain); 2428 oops: 2429 if (!cachep && (flags & SLAB_PANIC)) 2430 panic("kmem_cache_create(): failed to create slab `%s'\n", 2431 name); 2432 if (slab_is_available()) { 2433 mutex_unlock(&cache_chain_mutex); 2434 put_online_cpus(); 2435 } 2436 return cachep; 2437 } 2438 EXPORT_SYMBOL(kmem_cache_create); 2439 2440 #if DEBUG 2441 static void check_irq_off(void) 2442 { 2443 BUG_ON(!irqs_disabled()); 2444 } 2445 2446 static void check_irq_on(void) 2447 { 2448 BUG_ON(irqs_disabled()); 2449 } 2450 2451 static void check_spinlock_acquired(struct kmem_cache *cachep) 2452 { 2453 #ifdef CONFIG_SMP 2454 check_irq_off(); 2455 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); 2456 #endif 2457 } 2458 2459 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2460 { 2461 #ifdef CONFIG_SMP 2462 check_irq_off(); 2463 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2464 #endif 2465 } 2466 2467 #else 2468 #define check_irq_off() do { } while(0) 2469 #define check_irq_on() do { } while(0) 2470 #define check_spinlock_acquired(x) do { } while(0) 2471 #define check_spinlock_acquired_node(x, y) do { } while(0) 2472 #endif 2473 2474 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2475 struct array_cache *ac, 2476 int force, int node); 2477 2478 static void do_drain(void *arg) 2479 { 2480 struct kmem_cache *cachep = arg; 2481 struct array_cache *ac; 2482 int node = numa_mem_id(); 2483 2484 check_irq_off(); 2485 ac = cpu_cache_get(cachep); 2486 spin_lock(&cachep->nodelists[node]->list_lock); 2487 free_block(cachep, ac->entry, ac->avail, node); 2488 spin_unlock(&cachep->nodelists[node]->list_lock); 2489 ac->avail = 0; 2490 } 2491 2492 static void drain_cpu_caches(struct kmem_cache *cachep) 2493 { 2494 struct kmem_list3 *l3; 2495 int node; 2496 2497 on_each_cpu(do_drain, cachep, 1); 2498 check_irq_on(); 2499 for_each_online_node(node) { 2500 l3 = cachep->nodelists[node]; 2501 if (l3 && l3->alien) 2502 drain_alien_cache(cachep, l3->alien); 2503 } 2504 2505 for_each_online_node(node) { 2506 l3 = cachep->nodelists[node]; 2507 if (l3) 2508 drain_array(cachep, l3, l3->shared, 1, node); 2509 } 2510 } 2511 2512 /* 2513 * Remove slabs from the list of free slabs. 2514 * Specify the number of slabs to drain in tofree. 2515 * 2516 * Returns the actual number of slabs released. 2517 */ 2518 static int drain_freelist(struct kmem_cache *cache, 2519 struct kmem_list3 *l3, int tofree) 2520 { 2521 struct list_head *p; 2522 int nr_freed; 2523 struct slab *slabp; 2524 2525 nr_freed = 0; 2526 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { 2527 2528 spin_lock_irq(&l3->list_lock); 2529 p = l3->slabs_free.prev; 2530 if (p == &l3->slabs_free) { 2531 spin_unlock_irq(&l3->list_lock); 2532 goto out; 2533 } 2534 2535 slabp = list_entry(p, struct slab, list); 2536 #if DEBUG 2537 BUG_ON(slabp->inuse); 2538 #endif 2539 list_del(&slabp->list); 2540 /* 2541 * Safe to drop the lock. The slab is no longer linked 2542 * to the cache. 2543 */ 2544 l3->free_objects -= cache->num; 2545 spin_unlock_irq(&l3->list_lock); 2546 slab_destroy(cache, slabp); 2547 nr_freed++; 2548 } 2549 out: 2550 return nr_freed; 2551 } 2552 2553 /* Called with cache_chain_mutex held to protect against cpu hotplug */ 2554 static int __cache_shrink(struct kmem_cache *cachep) 2555 { 2556 int ret = 0, i = 0; 2557 struct kmem_list3 *l3; 2558 2559 drain_cpu_caches(cachep); 2560 2561 check_irq_on(); 2562 for_each_online_node(i) { 2563 l3 = cachep->nodelists[i]; 2564 if (!l3) 2565 continue; 2566 2567 drain_freelist(cachep, l3, l3->free_objects); 2568 2569 ret += !list_empty(&l3->slabs_full) || 2570 !list_empty(&l3->slabs_partial); 2571 } 2572 return (ret ? 1 : 0); 2573 } 2574 2575 /** 2576 * kmem_cache_shrink - Shrink a cache. 2577 * @cachep: The cache to shrink. 2578 * 2579 * Releases as many slabs as possible for a cache. 2580 * To help debugging, a zero exit status indicates all slabs were released. 2581 */ 2582 int kmem_cache_shrink(struct kmem_cache *cachep) 2583 { 2584 int ret; 2585 BUG_ON(!cachep || in_interrupt()); 2586 2587 get_online_cpus(); 2588 mutex_lock(&cache_chain_mutex); 2589 ret = __cache_shrink(cachep); 2590 mutex_unlock(&cache_chain_mutex); 2591 put_online_cpus(); 2592 return ret; 2593 } 2594 EXPORT_SYMBOL(kmem_cache_shrink); 2595 2596 /** 2597 * kmem_cache_destroy - delete a cache 2598 * @cachep: the cache to destroy 2599 * 2600 * Remove a &struct kmem_cache object from the slab cache. 2601 * 2602 * It is expected this function will be called by a module when it is 2603 * unloaded. This will remove the cache completely, and avoid a duplicate 2604 * cache being allocated each time a module is loaded and unloaded, if the 2605 * module doesn't have persistent in-kernel storage across loads and unloads. 2606 * 2607 * The cache must be empty before calling this function. 2608 * 2609 * The caller must guarantee that noone will allocate memory from the cache 2610 * during the kmem_cache_destroy(). 2611 */ 2612 void kmem_cache_destroy(struct kmem_cache *cachep) 2613 { 2614 BUG_ON(!cachep || in_interrupt()); 2615 2616 /* Find the cache in the chain of caches. */ 2617 get_online_cpus(); 2618 mutex_lock(&cache_chain_mutex); 2619 /* 2620 * the chain is never empty, cache_cache is never destroyed 2621 */ 2622 list_del(&cachep->next); 2623 if (__cache_shrink(cachep)) { 2624 slab_error(cachep, "Can't free all objects"); 2625 list_add(&cachep->next, &cache_chain); 2626 mutex_unlock(&cache_chain_mutex); 2627 put_online_cpus(); 2628 return; 2629 } 2630 2631 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2632 rcu_barrier(); 2633 2634 __kmem_cache_destroy(cachep); 2635 mutex_unlock(&cache_chain_mutex); 2636 put_online_cpus(); 2637 } 2638 EXPORT_SYMBOL(kmem_cache_destroy); 2639 2640 /* 2641 * Get the memory for a slab management obj. 2642 * For a slab cache when the slab descriptor is off-slab, slab descriptors 2643 * always come from malloc_sizes caches. The slab descriptor cannot 2644 * come from the same cache which is getting created because, 2645 * when we are searching for an appropriate cache for these 2646 * descriptors in kmem_cache_create, we search through the malloc_sizes array. 2647 * If we are creating a malloc_sizes cache here it would not be visible to 2648 * kmem_find_general_cachep till the initialization is complete. 2649 * Hence we cannot have slabp_cache same as the original cache. 2650 */ 2651 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2652 int colour_off, gfp_t local_flags, 2653 int nodeid) 2654 { 2655 struct slab *slabp; 2656 2657 if (OFF_SLAB(cachep)) { 2658 /* Slab management obj is off-slab. */ 2659 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2660 local_flags, nodeid); 2661 /* 2662 * If the first object in the slab is leaked (it's allocated 2663 * but no one has a reference to it), we want to make sure 2664 * kmemleak does not treat the ->s_mem pointer as a reference 2665 * to the object. Otherwise we will not report the leak. 2666 */ 2667 kmemleak_scan_area(&slabp->list, sizeof(struct list_head), 2668 local_flags); 2669 if (!slabp) 2670 return NULL; 2671 } else { 2672 slabp = objp + colour_off; 2673 colour_off += cachep->slab_size; 2674 } 2675 slabp->inuse = 0; 2676 slabp->colouroff = colour_off; 2677 slabp->s_mem = objp + colour_off; 2678 slabp->nodeid = nodeid; 2679 slabp->free = 0; 2680 return slabp; 2681 } 2682 2683 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2684 { 2685 return (kmem_bufctl_t *) (slabp + 1); 2686 } 2687 2688 static void cache_init_objs(struct kmem_cache *cachep, 2689 struct slab *slabp) 2690 { 2691 int i; 2692 2693 for (i = 0; i < cachep->num; i++) { 2694 void *objp = index_to_obj(cachep, slabp, i); 2695 #if DEBUG 2696 /* need to poison the objs? */ 2697 if (cachep->flags & SLAB_POISON) 2698 poison_obj(cachep, objp, POISON_FREE); 2699 if (cachep->flags & SLAB_STORE_USER) 2700 *dbg_userword(cachep, objp) = NULL; 2701 2702 if (cachep->flags & SLAB_RED_ZONE) { 2703 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2704 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2705 } 2706 /* 2707 * Constructors are not allowed to allocate memory from the same 2708 * cache which they are a constructor for. Otherwise, deadlock. 2709 * They must also be threaded. 2710 */ 2711 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2712 cachep->ctor(objp + obj_offset(cachep)); 2713 2714 if (cachep->flags & SLAB_RED_ZONE) { 2715 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2716 slab_error(cachep, "constructor overwrote the" 2717 " end of an object"); 2718 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2719 slab_error(cachep, "constructor overwrote the" 2720 " start of an object"); 2721 } 2722 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2723 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2724 kernel_map_pages(virt_to_page(objp), 2725 cachep->buffer_size / PAGE_SIZE, 0); 2726 #else 2727 if (cachep->ctor) 2728 cachep->ctor(objp); 2729 #endif 2730 slab_bufctl(slabp)[i] = i + 1; 2731 } 2732 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2733 } 2734 2735 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2736 { 2737 if (CONFIG_ZONE_DMA_FLAG) { 2738 if (flags & GFP_DMA) 2739 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2740 else 2741 BUG_ON(cachep->gfpflags & GFP_DMA); 2742 } 2743 } 2744 2745 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2746 int nodeid) 2747 { 2748 void *objp = index_to_obj(cachep, slabp, slabp->free); 2749 kmem_bufctl_t next; 2750 2751 slabp->inuse++; 2752 next = slab_bufctl(slabp)[slabp->free]; 2753 #if DEBUG 2754 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2755 WARN_ON(slabp->nodeid != nodeid); 2756 #endif 2757 slabp->free = next; 2758 2759 return objp; 2760 } 2761 2762 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2763 void *objp, int nodeid) 2764 { 2765 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2766 2767 #if DEBUG 2768 /* Verify that the slab belongs to the intended node */ 2769 WARN_ON(slabp->nodeid != nodeid); 2770 2771 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2772 printk(KERN_ERR "slab: double free detected in cache " 2773 "'%s', objp %p\n", cachep->name, objp); 2774 BUG(); 2775 } 2776 #endif 2777 slab_bufctl(slabp)[objnr] = slabp->free; 2778 slabp->free = objnr; 2779 slabp->inuse--; 2780 } 2781 2782 /* 2783 * Map pages beginning at addr to the given cache and slab. This is required 2784 * for the slab allocator to be able to lookup the cache and slab of a 2785 * virtual address for kfree, ksize, and slab debugging. 2786 */ 2787 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, 2788 void *addr) 2789 { 2790 int nr_pages; 2791 struct page *page; 2792 2793 page = virt_to_page(addr); 2794 2795 nr_pages = 1; 2796 if (likely(!PageCompound(page))) 2797 nr_pages <<= cache->gfporder; 2798 2799 do { 2800 page_set_cache(page, cache); 2801 page_set_slab(page, slab); 2802 page++; 2803 } while (--nr_pages); 2804 } 2805 2806 /* 2807 * Grow (by 1) the number of slabs within a cache. This is called by 2808 * kmem_cache_alloc() when there are no active objs left in a cache. 2809 */ 2810 static int cache_grow(struct kmem_cache *cachep, 2811 gfp_t flags, int nodeid, void *objp) 2812 { 2813 struct slab *slabp; 2814 size_t offset; 2815 gfp_t local_flags; 2816 struct kmem_list3 *l3; 2817 2818 /* 2819 * Be lazy and only check for valid flags here, keeping it out of the 2820 * critical path in kmem_cache_alloc(). 2821 */ 2822 BUG_ON(flags & GFP_SLAB_BUG_MASK); 2823 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); 2824 2825 /* Take the l3 list lock to change the colour_next on this node */ 2826 check_irq_off(); 2827 l3 = cachep->nodelists[nodeid]; 2828 spin_lock(&l3->list_lock); 2829 2830 /* Get colour for the slab, and cal the next value. */ 2831 offset = l3->colour_next; 2832 l3->colour_next++; 2833 if (l3->colour_next >= cachep->colour) 2834 l3->colour_next = 0; 2835 spin_unlock(&l3->list_lock); 2836 2837 offset *= cachep->colour_off; 2838 2839 if (local_flags & __GFP_WAIT) 2840 local_irq_enable(); 2841 2842 /* 2843 * The test for missing atomic flag is performed here, rather than 2844 * the more obvious place, simply to reduce the critical path length 2845 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2846 * will eventually be caught here (where it matters). 2847 */ 2848 kmem_flagcheck(cachep, flags); 2849 2850 /* 2851 * Get mem for the objs. Attempt to allocate a physical page from 2852 * 'nodeid'. 2853 */ 2854 if (!objp) 2855 objp = kmem_getpages(cachep, local_flags, nodeid); 2856 if (!objp) 2857 goto failed; 2858 2859 /* Get slab management. */ 2860 slabp = alloc_slabmgmt(cachep, objp, offset, 2861 local_flags & ~GFP_CONSTRAINT_MASK, nodeid); 2862 if (!slabp) 2863 goto opps1; 2864 2865 slab_map_pages(cachep, slabp, objp); 2866 2867 cache_init_objs(cachep, slabp); 2868 2869 if (local_flags & __GFP_WAIT) 2870 local_irq_disable(); 2871 check_irq_off(); 2872 spin_lock(&l3->list_lock); 2873 2874 /* Make slab active. */ 2875 list_add_tail(&slabp->list, &(l3->slabs_free)); 2876 STATS_INC_GROWN(cachep); 2877 l3->free_objects += cachep->num; 2878 spin_unlock(&l3->list_lock); 2879 return 1; 2880 opps1: 2881 kmem_freepages(cachep, objp); 2882 failed: 2883 if (local_flags & __GFP_WAIT) 2884 local_irq_disable(); 2885 return 0; 2886 } 2887 2888 #if DEBUG 2889 2890 /* 2891 * Perform extra freeing checks: 2892 * - detect bad pointers. 2893 * - POISON/RED_ZONE checking 2894 */ 2895 static void kfree_debugcheck(const void *objp) 2896 { 2897 if (!virt_addr_valid(objp)) { 2898 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2899 (unsigned long)objp); 2900 BUG(); 2901 } 2902 } 2903 2904 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2905 { 2906 unsigned long long redzone1, redzone2; 2907 2908 redzone1 = *dbg_redzone1(cache, obj); 2909 redzone2 = *dbg_redzone2(cache, obj); 2910 2911 /* 2912 * Redzone is ok. 2913 */ 2914 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2915 return; 2916 2917 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2918 slab_error(cache, "double free detected"); 2919 else 2920 slab_error(cache, "memory outside object was overwritten"); 2921 2922 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", 2923 obj, redzone1, redzone2); 2924 } 2925 2926 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2927 void *caller) 2928 { 2929 struct page *page; 2930 unsigned int objnr; 2931 struct slab *slabp; 2932 2933 BUG_ON(virt_to_cache(objp) != cachep); 2934 2935 objp -= obj_offset(cachep); 2936 kfree_debugcheck(objp); 2937 page = virt_to_head_page(objp); 2938 2939 slabp = page_get_slab(page); 2940 2941 if (cachep->flags & SLAB_RED_ZONE) { 2942 verify_redzone_free(cachep, objp); 2943 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2944 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2945 } 2946 if (cachep->flags & SLAB_STORE_USER) 2947 *dbg_userword(cachep, objp) = caller; 2948 2949 objnr = obj_to_index(cachep, slabp, objp); 2950 2951 BUG_ON(objnr >= cachep->num); 2952 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2953 2954 #ifdef CONFIG_DEBUG_SLAB_LEAK 2955 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2956 #endif 2957 if (cachep->flags & SLAB_POISON) { 2958 #ifdef CONFIG_DEBUG_PAGEALLOC 2959 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2960 store_stackinfo(cachep, objp, (unsigned long)caller); 2961 kernel_map_pages(virt_to_page(objp), 2962 cachep->buffer_size / PAGE_SIZE, 0); 2963 } else { 2964 poison_obj(cachep, objp, POISON_FREE); 2965 } 2966 #else 2967 poison_obj(cachep, objp, POISON_FREE); 2968 #endif 2969 } 2970 return objp; 2971 } 2972 2973 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2974 { 2975 kmem_bufctl_t i; 2976 int entries = 0; 2977 2978 /* Check slab's freelist to see if this obj is there. */ 2979 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2980 entries++; 2981 if (entries > cachep->num || i >= cachep->num) 2982 goto bad; 2983 } 2984 if (entries != cachep->num - slabp->inuse) { 2985 bad: 2986 printk(KERN_ERR "slab: Internal list corruption detected in " 2987 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2988 cachep->name, cachep->num, slabp, slabp->inuse); 2989 for (i = 0; 2990 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2991 i++) { 2992 if (i % 16 == 0) 2993 printk("\n%03x:", i); 2994 printk(" %02x", ((unsigned char *)slabp)[i]); 2995 } 2996 printk("\n"); 2997 BUG(); 2998 } 2999 } 3000 #else 3001 #define kfree_debugcheck(x) do { } while(0) 3002 #define cache_free_debugcheck(x,objp,z) (objp) 3003 #define check_slabp(x,y) do { } while(0) 3004 #endif 3005 3006 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 3007 { 3008 int batchcount; 3009 struct kmem_list3 *l3; 3010 struct array_cache *ac; 3011 int node; 3012 3013 retry: 3014 check_irq_off(); 3015 node = numa_mem_id(); 3016 ac = cpu_cache_get(cachep); 3017 batchcount = ac->batchcount; 3018 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 3019 /* 3020 * If there was little recent activity on this cache, then 3021 * perform only a partial refill. Otherwise we could generate 3022 * refill bouncing. 3023 */ 3024 batchcount = BATCHREFILL_LIMIT; 3025 } 3026 l3 = cachep->nodelists[node]; 3027 3028 BUG_ON(ac->avail > 0 || !l3); 3029 spin_lock(&l3->list_lock); 3030 3031 /* See if we can refill from the shared array */ 3032 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { 3033 l3->shared->touched = 1; 3034 goto alloc_done; 3035 } 3036 3037 while (batchcount > 0) { 3038 struct list_head *entry; 3039 struct slab *slabp; 3040 /* Get slab alloc is to come from. */ 3041 entry = l3->slabs_partial.next; 3042 if (entry == &l3->slabs_partial) { 3043 l3->free_touched = 1; 3044 entry = l3->slabs_free.next; 3045 if (entry == &l3->slabs_free) 3046 goto must_grow; 3047 } 3048 3049 slabp = list_entry(entry, struct slab, list); 3050 check_slabp(cachep, slabp); 3051 check_spinlock_acquired(cachep); 3052 3053 /* 3054 * The slab was either on partial or free list so 3055 * there must be at least one object available for 3056 * allocation. 3057 */ 3058 BUG_ON(slabp->inuse >= cachep->num); 3059 3060 while (slabp->inuse < cachep->num && batchcount--) { 3061 STATS_INC_ALLOCED(cachep); 3062 STATS_INC_ACTIVE(cachep); 3063 STATS_SET_HIGH(cachep); 3064 3065 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 3066 node); 3067 } 3068 check_slabp(cachep, slabp); 3069 3070 /* move slabp to correct slabp list: */ 3071 list_del(&slabp->list); 3072 if (slabp->free == BUFCTL_END) 3073 list_add(&slabp->list, &l3->slabs_full); 3074 else 3075 list_add(&slabp->list, &l3->slabs_partial); 3076 } 3077 3078 must_grow: 3079 l3->free_objects -= ac->avail; 3080 alloc_done: 3081 spin_unlock(&l3->list_lock); 3082 3083 if (unlikely(!ac->avail)) { 3084 int x; 3085 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); 3086 3087 /* cache_grow can reenable interrupts, then ac could change. */ 3088 ac = cpu_cache_get(cachep); 3089 if (!x && ac->avail == 0) /* no objects in sight? abort */ 3090 return NULL; 3091 3092 if (!ac->avail) /* objects refilled by interrupt? */ 3093 goto retry; 3094 } 3095 ac->touched = 1; 3096 return ac->entry[--ac->avail]; 3097 } 3098 3099 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3100 gfp_t flags) 3101 { 3102 might_sleep_if(flags & __GFP_WAIT); 3103 #if DEBUG 3104 kmem_flagcheck(cachep, flags); 3105 #endif 3106 } 3107 3108 #if DEBUG 3109 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3110 gfp_t flags, void *objp, void *caller) 3111 { 3112 if (!objp) 3113 return objp; 3114 if (cachep->flags & SLAB_POISON) { 3115 #ifdef CONFIG_DEBUG_PAGEALLOC 3116 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 3117 kernel_map_pages(virt_to_page(objp), 3118 cachep->buffer_size / PAGE_SIZE, 1); 3119 else 3120 check_poison_obj(cachep, objp); 3121 #else 3122 check_poison_obj(cachep, objp); 3123 #endif 3124 poison_obj(cachep, objp, POISON_INUSE); 3125 } 3126 if (cachep->flags & SLAB_STORE_USER) 3127 *dbg_userword(cachep, objp) = caller; 3128 3129 if (cachep->flags & SLAB_RED_ZONE) { 3130 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3131 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3132 slab_error(cachep, "double free, or memory outside" 3133 " object was overwritten"); 3134 printk(KERN_ERR 3135 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", 3136 objp, *dbg_redzone1(cachep, objp), 3137 *dbg_redzone2(cachep, objp)); 3138 } 3139 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3140 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3141 } 3142 #ifdef CONFIG_DEBUG_SLAB_LEAK 3143 { 3144 struct slab *slabp; 3145 unsigned objnr; 3146 3147 slabp = page_get_slab(virt_to_head_page(objp)); 3148 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 3149 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 3150 } 3151 #endif 3152 objp += obj_offset(cachep); 3153 if (cachep->ctor && cachep->flags & SLAB_POISON) 3154 cachep->ctor(objp); 3155 #if ARCH_SLAB_MINALIGN 3156 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { 3157 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3158 objp, ARCH_SLAB_MINALIGN); 3159 } 3160 #endif 3161 return objp; 3162 } 3163 #else 3164 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3165 #endif 3166 3167 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) 3168 { 3169 if (cachep == &cache_cache) 3170 return false; 3171 3172 return should_failslab(obj_size(cachep), flags, cachep->flags); 3173 } 3174 3175 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3176 { 3177 void *objp; 3178 struct array_cache *ac; 3179 3180 check_irq_off(); 3181 3182 ac = cpu_cache_get(cachep); 3183 if (likely(ac->avail)) { 3184 STATS_INC_ALLOCHIT(cachep); 3185 ac->touched = 1; 3186 objp = ac->entry[--ac->avail]; 3187 } else { 3188 STATS_INC_ALLOCMISS(cachep); 3189 objp = cache_alloc_refill(cachep, flags); 3190 /* 3191 * the 'ac' may be updated by cache_alloc_refill(), 3192 * and kmemleak_erase() requires its correct value. 3193 */ 3194 ac = cpu_cache_get(cachep); 3195 } 3196 /* 3197 * To avoid a false negative, if an object that is in one of the 3198 * per-CPU caches is leaked, we need to make sure kmemleak doesn't 3199 * treat the array pointers as a reference to the object. 3200 */ 3201 if (objp) 3202 kmemleak_erase(&ac->entry[ac->avail]); 3203 return objp; 3204 } 3205 3206 #ifdef CONFIG_NUMA 3207 /* 3208 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 3209 * 3210 * If we are in_interrupt, then process context, including cpusets and 3211 * mempolicy, may not apply and should not be used for allocation policy. 3212 */ 3213 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3214 { 3215 int nid_alloc, nid_here; 3216 3217 if (in_interrupt() || (flags & __GFP_THISNODE)) 3218 return NULL; 3219 nid_alloc = nid_here = numa_mem_id(); 3220 get_mems_allowed(); 3221 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3222 nid_alloc = cpuset_slab_spread_node(); 3223 else if (current->mempolicy) 3224 nid_alloc = slab_node(current->mempolicy); 3225 put_mems_allowed(); 3226 if (nid_alloc != nid_here) 3227 return ____cache_alloc_node(cachep, flags, nid_alloc); 3228 return NULL; 3229 } 3230 3231 /* 3232 * Fallback function if there was no memory available and no objects on a 3233 * certain node and fall back is permitted. First we scan all the 3234 * available nodelists for available objects. If that fails then we 3235 * perform an allocation without specifying a node. This allows the page 3236 * allocator to do its reclaim / fallback magic. We then insert the 3237 * slab into the proper nodelist and then allocate from it. 3238 */ 3239 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3240 { 3241 struct zonelist *zonelist; 3242 gfp_t local_flags; 3243 struct zoneref *z; 3244 struct zone *zone; 3245 enum zone_type high_zoneidx = gfp_zone(flags); 3246 void *obj = NULL; 3247 int nid; 3248 3249 if (flags & __GFP_THISNODE) 3250 return NULL; 3251 3252 get_mems_allowed(); 3253 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 3254 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); 3255 3256 retry: 3257 /* 3258 * Look through allowed nodes for objects available 3259 * from existing per node queues. 3260 */ 3261 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 3262 nid = zone_to_nid(zone); 3263 3264 if (cpuset_zone_allowed_hardwall(zone, flags) && 3265 cache->nodelists[nid] && 3266 cache->nodelists[nid]->free_objects) { 3267 obj = ____cache_alloc_node(cache, 3268 flags | GFP_THISNODE, nid); 3269 if (obj) 3270 break; 3271 } 3272 } 3273 3274 if (!obj) { 3275 /* 3276 * This allocation will be performed within the constraints 3277 * of the current cpuset / memory policy requirements. 3278 * We may trigger various forms of reclaim on the allowed 3279 * set and go into memory reserves if necessary. 3280 */ 3281 if (local_flags & __GFP_WAIT) 3282 local_irq_enable(); 3283 kmem_flagcheck(cache, flags); 3284 obj = kmem_getpages(cache, local_flags, numa_mem_id()); 3285 if (local_flags & __GFP_WAIT) 3286 local_irq_disable(); 3287 if (obj) { 3288 /* 3289 * Insert into the appropriate per node queues 3290 */ 3291 nid = page_to_nid(virt_to_page(obj)); 3292 if (cache_grow(cache, flags, nid, obj)) { 3293 obj = ____cache_alloc_node(cache, 3294 flags | GFP_THISNODE, nid); 3295 if (!obj) 3296 /* 3297 * Another processor may allocate the 3298 * objects in the slab since we are 3299 * not holding any locks. 3300 */ 3301 goto retry; 3302 } else { 3303 /* cache_grow already freed obj */ 3304 obj = NULL; 3305 } 3306 } 3307 } 3308 put_mems_allowed(); 3309 return obj; 3310 } 3311 3312 /* 3313 * A interface to enable slab creation on nodeid 3314 */ 3315 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3316 int nodeid) 3317 { 3318 struct list_head *entry; 3319 struct slab *slabp; 3320 struct kmem_list3 *l3; 3321 void *obj; 3322 int x; 3323 3324 l3 = cachep->nodelists[nodeid]; 3325 BUG_ON(!l3); 3326 3327 retry: 3328 check_irq_off(); 3329 spin_lock(&l3->list_lock); 3330 entry = l3->slabs_partial.next; 3331 if (entry == &l3->slabs_partial) { 3332 l3->free_touched = 1; 3333 entry = l3->slabs_free.next; 3334 if (entry == &l3->slabs_free) 3335 goto must_grow; 3336 } 3337 3338 slabp = list_entry(entry, struct slab, list); 3339 check_spinlock_acquired_node(cachep, nodeid); 3340 check_slabp(cachep, slabp); 3341 3342 STATS_INC_NODEALLOCS(cachep); 3343 STATS_INC_ACTIVE(cachep); 3344 STATS_SET_HIGH(cachep); 3345 3346 BUG_ON(slabp->inuse == cachep->num); 3347 3348 obj = slab_get_obj(cachep, slabp, nodeid); 3349 check_slabp(cachep, slabp); 3350 l3->free_objects--; 3351 /* move slabp to correct slabp list: */ 3352 list_del(&slabp->list); 3353 3354 if (slabp->free == BUFCTL_END) 3355 list_add(&slabp->list, &l3->slabs_full); 3356 else 3357 list_add(&slabp->list, &l3->slabs_partial); 3358 3359 spin_unlock(&l3->list_lock); 3360 goto done; 3361 3362 must_grow: 3363 spin_unlock(&l3->list_lock); 3364 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); 3365 if (x) 3366 goto retry; 3367 3368 return fallback_alloc(cachep, flags); 3369 3370 done: 3371 return obj; 3372 } 3373 3374 /** 3375 * kmem_cache_alloc_node - Allocate an object on the specified node 3376 * @cachep: The cache to allocate from. 3377 * @flags: See kmalloc(). 3378 * @nodeid: node number of the target node. 3379 * @caller: return address of caller, used for debug information 3380 * 3381 * Identical to kmem_cache_alloc but it will allocate memory on the given 3382 * node, which can improve the performance for cpu bound structures. 3383 * 3384 * Fallback to other node is possible if __GFP_THISNODE is not set. 3385 */ 3386 static __always_inline void * 3387 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, 3388 void *caller) 3389 { 3390 unsigned long save_flags; 3391 void *ptr; 3392 int slab_node = numa_mem_id(); 3393 3394 flags &= gfp_allowed_mask; 3395 3396 lockdep_trace_alloc(flags); 3397 3398 if (slab_should_failslab(cachep, flags)) 3399 return NULL; 3400 3401 cache_alloc_debugcheck_before(cachep, flags); 3402 local_irq_save(save_flags); 3403 3404 if (nodeid == -1) 3405 nodeid = slab_node; 3406 3407 if (unlikely(!cachep->nodelists[nodeid])) { 3408 /* Node not bootstrapped yet */ 3409 ptr = fallback_alloc(cachep, flags); 3410 goto out; 3411 } 3412 3413 if (nodeid == slab_node) { 3414 /* 3415 * Use the locally cached objects if possible. 3416 * However ____cache_alloc does not allow fallback 3417 * to other nodes. It may fail while we still have 3418 * objects on other nodes available. 3419 */ 3420 ptr = ____cache_alloc(cachep, flags); 3421 if (ptr) 3422 goto out; 3423 } 3424 /* ___cache_alloc_node can fall back to other nodes */ 3425 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3426 out: 3427 local_irq_restore(save_flags); 3428 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3429 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags, 3430 flags); 3431 3432 if (likely(ptr)) 3433 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep)); 3434 3435 if (unlikely((flags & __GFP_ZERO) && ptr)) 3436 memset(ptr, 0, obj_size(cachep)); 3437 3438 return ptr; 3439 } 3440 3441 static __always_inline void * 3442 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) 3443 { 3444 void *objp; 3445 3446 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { 3447 objp = alternate_node_alloc(cache, flags); 3448 if (objp) 3449 goto out; 3450 } 3451 objp = ____cache_alloc(cache, flags); 3452 3453 /* 3454 * We may just have run out of memory on the local node. 3455 * ____cache_alloc_node() knows how to locate memory on other nodes 3456 */ 3457 if (!objp) 3458 objp = ____cache_alloc_node(cache, flags, numa_mem_id()); 3459 3460 out: 3461 return objp; 3462 } 3463 #else 3464 3465 static __always_inline void * 3466 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3467 { 3468 return ____cache_alloc(cachep, flags); 3469 } 3470 3471 #endif /* CONFIG_NUMA */ 3472 3473 static __always_inline void * 3474 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) 3475 { 3476 unsigned long save_flags; 3477 void *objp; 3478 3479 flags &= gfp_allowed_mask; 3480 3481 lockdep_trace_alloc(flags); 3482 3483 if (slab_should_failslab(cachep, flags)) 3484 return NULL; 3485 3486 cache_alloc_debugcheck_before(cachep, flags); 3487 local_irq_save(save_flags); 3488 objp = __do_cache_alloc(cachep, flags); 3489 local_irq_restore(save_flags); 3490 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); 3491 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags, 3492 flags); 3493 prefetchw(objp); 3494 3495 if (likely(objp)) 3496 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep)); 3497 3498 if (unlikely((flags & __GFP_ZERO) && objp)) 3499 memset(objp, 0, obj_size(cachep)); 3500 3501 return objp; 3502 } 3503 3504 /* 3505 * Caller needs to acquire correct kmem_list's list_lock 3506 */ 3507 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 3508 int node) 3509 { 3510 int i; 3511 struct kmem_list3 *l3; 3512 3513 for (i = 0; i < nr_objects; i++) { 3514 void *objp = objpp[i]; 3515 struct slab *slabp; 3516 3517 slabp = virt_to_slab(objp); 3518 l3 = cachep->nodelists[node]; 3519 list_del(&slabp->list); 3520 check_spinlock_acquired_node(cachep, node); 3521 check_slabp(cachep, slabp); 3522 slab_put_obj(cachep, slabp, objp, node); 3523 STATS_DEC_ACTIVE(cachep); 3524 l3->free_objects++; 3525 check_slabp(cachep, slabp); 3526 3527 /* fixup slab chains */ 3528 if (slabp->inuse == 0) { 3529 if (l3->free_objects > l3->free_limit) { 3530 l3->free_objects -= cachep->num; 3531 /* No need to drop any previously held 3532 * lock here, even if we have a off-slab slab 3533 * descriptor it is guaranteed to come from 3534 * a different cache, refer to comments before 3535 * alloc_slabmgmt. 3536 */ 3537 slab_destroy(cachep, slabp); 3538 } else { 3539 list_add(&slabp->list, &l3->slabs_free); 3540 } 3541 } else { 3542 /* Unconditionally move a slab to the end of the 3543 * partial list on free - maximum time for the 3544 * other objects to be freed, too. 3545 */ 3546 list_add_tail(&slabp->list, &l3->slabs_partial); 3547 } 3548 } 3549 } 3550 3551 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3552 { 3553 int batchcount; 3554 struct kmem_list3 *l3; 3555 int node = numa_mem_id(); 3556 3557 batchcount = ac->batchcount; 3558 #if DEBUG 3559 BUG_ON(!batchcount || batchcount > ac->avail); 3560 #endif 3561 check_irq_off(); 3562 l3 = cachep->nodelists[node]; 3563 spin_lock(&l3->list_lock); 3564 if (l3->shared) { 3565 struct array_cache *shared_array = l3->shared; 3566 int max = shared_array->limit - shared_array->avail; 3567 if (max) { 3568 if (batchcount > max) 3569 batchcount = max; 3570 memcpy(&(shared_array->entry[shared_array->avail]), 3571 ac->entry, sizeof(void *) * batchcount); 3572 shared_array->avail += batchcount; 3573 goto free_done; 3574 } 3575 } 3576 3577 free_block(cachep, ac->entry, batchcount, node); 3578 free_done: 3579 #if STATS 3580 { 3581 int i = 0; 3582 struct list_head *p; 3583 3584 p = l3->slabs_free.next; 3585 while (p != &(l3->slabs_free)) { 3586 struct slab *slabp; 3587 3588 slabp = list_entry(p, struct slab, list); 3589 BUG_ON(slabp->inuse); 3590 3591 i++; 3592 p = p->next; 3593 } 3594 STATS_SET_FREEABLE(cachep, i); 3595 } 3596 #endif 3597 spin_unlock(&l3->list_lock); 3598 ac->avail -= batchcount; 3599 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3600 } 3601 3602 /* 3603 * Release an obj back to its cache. If the obj has a constructed state, it must 3604 * be in this state _before_ it is released. Called with disabled ints. 3605 */ 3606 static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3607 { 3608 struct array_cache *ac = cpu_cache_get(cachep); 3609 3610 check_irq_off(); 3611 kmemleak_free_recursive(objp, cachep->flags); 3612 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3613 3614 kmemcheck_slab_free(cachep, objp, obj_size(cachep)); 3615 3616 /* 3617 * Skip calling cache_free_alien() when the platform is not numa. 3618 * This will avoid cache misses that happen while accessing slabp (which 3619 * is per page memory reference) to get nodeid. Instead use a global 3620 * variable to skip the call, which is mostly likely to be present in 3621 * the cache. 3622 */ 3623 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) 3624 return; 3625 3626 if (likely(ac->avail < ac->limit)) { 3627 STATS_INC_FREEHIT(cachep); 3628 ac->entry[ac->avail++] = objp; 3629 return; 3630 } else { 3631 STATS_INC_FREEMISS(cachep); 3632 cache_flusharray(cachep, ac); 3633 ac->entry[ac->avail++] = objp; 3634 } 3635 } 3636 3637 /** 3638 * kmem_cache_alloc - Allocate an object 3639 * @cachep: The cache to allocate from. 3640 * @flags: See kmalloc(). 3641 * 3642 * Allocate an object from this cache. The flags are only relevant 3643 * if the cache has no available objects. 3644 */ 3645 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3646 { 3647 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); 3648 3649 trace_kmem_cache_alloc(_RET_IP_, ret, 3650 obj_size(cachep), cachep->buffer_size, flags); 3651 3652 return ret; 3653 } 3654 EXPORT_SYMBOL(kmem_cache_alloc); 3655 3656 #ifdef CONFIG_TRACING 3657 void * 3658 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) 3659 { 3660 void *ret; 3661 3662 ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); 3663 3664 trace_kmalloc(_RET_IP_, ret, 3665 size, slab_buffer_size(cachep), flags); 3666 return ret; 3667 } 3668 EXPORT_SYMBOL(kmem_cache_alloc_trace); 3669 #endif 3670 3671 #ifdef CONFIG_NUMA 3672 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3673 { 3674 void *ret = __cache_alloc_node(cachep, flags, nodeid, 3675 __builtin_return_address(0)); 3676 3677 trace_kmem_cache_alloc_node(_RET_IP_, ret, 3678 obj_size(cachep), cachep->buffer_size, 3679 flags, nodeid); 3680 3681 return ret; 3682 } 3683 EXPORT_SYMBOL(kmem_cache_alloc_node); 3684 3685 #ifdef CONFIG_TRACING 3686 void *kmem_cache_alloc_node_trace(size_t size, 3687 struct kmem_cache *cachep, 3688 gfp_t flags, 3689 int nodeid) 3690 { 3691 void *ret; 3692 3693 ret = __cache_alloc_node(cachep, flags, nodeid, 3694 __builtin_return_address(0)); 3695 trace_kmalloc_node(_RET_IP_, ret, 3696 size, slab_buffer_size(cachep), 3697 flags, nodeid); 3698 return ret; 3699 } 3700 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 3701 #endif 3702 3703 static __always_inline void * 3704 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) 3705 { 3706 struct kmem_cache *cachep; 3707 3708 cachep = kmem_find_general_cachep(size, flags); 3709 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3710 return cachep; 3711 return kmem_cache_alloc_node_trace(size, cachep, flags, node); 3712 } 3713 3714 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) 3715 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3716 { 3717 return __do_kmalloc_node(size, flags, node, 3718 __builtin_return_address(0)); 3719 } 3720 EXPORT_SYMBOL(__kmalloc_node); 3721 3722 void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3723 int node, unsigned long caller) 3724 { 3725 return __do_kmalloc_node(size, flags, node, (void *)caller); 3726 } 3727 EXPORT_SYMBOL(__kmalloc_node_track_caller); 3728 #else 3729 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3730 { 3731 return __do_kmalloc_node(size, flags, node, NULL); 3732 } 3733 EXPORT_SYMBOL(__kmalloc_node); 3734 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ 3735 #endif /* CONFIG_NUMA */ 3736 3737 /** 3738 * __do_kmalloc - allocate memory 3739 * @size: how many bytes of memory are required. 3740 * @flags: the type of memory to allocate (see kmalloc). 3741 * @caller: function caller for debug tracking of the caller 3742 */ 3743 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3744 void *caller) 3745 { 3746 struct kmem_cache *cachep; 3747 void *ret; 3748 3749 /* If you want to save a few bytes .text space: replace 3750 * __ with kmem_. 3751 * Then kmalloc uses the uninlined functions instead of the inline 3752 * functions. 3753 */ 3754 cachep = __find_general_cachep(size, flags); 3755 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3756 return cachep; 3757 ret = __cache_alloc(cachep, flags, caller); 3758 3759 trace_kmalloc((unsigned long) caller, ret, 3760 size, cachep->buffer_size, flags); 3761 3762 return ret; 3763 } 3764 3765 3766 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) 3767 void *__kmalloc(size_t size, gfp_t flags) 3768 { 3769 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3770 } 3771 EXPORT_SYMBOL(__kmalloc); 3772 3773 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) 3774 { 3775 return __do_kmalloc(size, flags, (void *)caller); 3776 } 3777 EXPORT_SYMBOL(__kmalloc_track_caller); 3778 3779 #else 3780 void *__kmalloc(size_t size, gfp_t flags) 3781 { 3782 return __do_kmalloc(size, flags, NULL); 3783 } 3784 EXPORT_SYMBOL(__kmalloc); 3785 #endif 3786 3787 /** 3788 * kmem_cache_free - Deallocate an object 3789 * @cachep: The cache the allocation was from. 3790 * @objp: The previously allocated object. 3791 * 3792 * Free an object which was previously allocated from this 3793 * cache. 3794 */ 3795 void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3796 { 3797 unsigned long flags; 3798 3799 local_irq_save(flags); 3800 debug_check_no_locks_freed(objp, obj_size(cachep)); 3801 if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) 3802 debug_check_no_obj_freed(objp, obj_size(cachep)); 3803 __cache_free(cachep, objp); 3804 local_irq_restore(flags); 3805 3806 trace_kmem_cache_free(_RET_IP_, objp); 3807 } 3808 EXPORT_SYMBOL(kmem_cache_free); 3809 3810 /** 3811 * kfree - free previously allocated memory 3812 * @objp: pointer returned by kmalloc. 3813 * 3814 * If @objp is NULL, no operation is performed. 3815 * 3816 * Don't free memory not originally allocated by kmalloc() 3817 * or you will run into trouble. 3818 */ 3819 void kfree(const void *objp) 3820 { 3821 struct kmem_cache *c; 3822 unsigned long flags; 3823 3824 trace_kfree(_RET_IP_, objp); 3825 3826 if (unlikely(ZERO_OR_NULL_PTR(objp))) 3827 return; 3828 local_irq_save(flags); 3829 kfree_debugcheck(objp); 3830 c = virt_to_cache(objp); 3831 debug_check_no_locks_freed(objp, obj_size(c)); 3832 debug_check_no_obj_freed(objp, obj_size(c)); 3833 __cache_free(c, (void *)objp); 3834 local_irq_restore(flags); 3835 } 3836 EXPORT_SYMBOL(kfree); 3837 3838 unsigned int kmem_cache_size(struct kmem_cache *cachep) 3839 { 3840 return obj_size(cachep); 3841 } 3842 EXPORT_SYMBOL(kmem_cache_size); 3843 3844 /* 3845 * This initializes kmem_list3 or resizes various caches for all nodes. 3846 */ 3847 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) 3848 { 3849 int node; 3850 struct kmem_list3 *l3; 3851 struct array_cache *new_shared; 3852 struct array_cache **new_alien = NULL; 3853 3854 for_each_online_node(node) { 3855 3856 if (use_alien_caches) { 3857 new_alien = alloc_alien_cache(node, cachep->limit, gfp); 3858 if (!new_alien) 3859 goto fail; 3860 } 3861 3862 new_shared = NULL; 3863 if (cachep->shared) { 3864 new_shared = alloc_arraycache(node, 3865 cachep->shared*cachep->batchcount, 3866 0xbaadf00d, gfp); 3867 if (!new_shared) { 3868 free_alien_cache(new_alien); 3869 goto fail; 3870 } 3871 } 3872 3873 l3 = cachep->nodelists[node]; 3874 if (l3) { 3875 struct array_cache *shared = l3->shared; 3876 3877 spin_lock_irq(&l3->list_lock); 3878 3879 if (shared) 3880 free_block(cachep, shared->entry, 3881 shared->avail, node); 3882 3883 l3->shared = new_shared; 3884 if (!l3->alien) { 3885 l3->alien = new_alien; 3886 new_alien = NULL; 3887 } 3888 l3->free_limit = (1 + nr_cpus_node(node)) * 3889 cachep->batchcount + cachep->num; 3890 spin_unlock_irq(&l3->list_lock); 3891 kfree(shared); 3892 free_alien_cache(new_alien); 3893 continue; 3894 } 3895 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); 3896 if (!l3) { 3897 free_alien_cache(new_alien); 3898 kfree(new_shared); 3899 goto fail; 3900 } 3901 3902 kmem_list3_init(l3); 3903 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3904 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3905 l3->shared = new_shared; 3906 l3->alien = new_alien; 3907 l3->free_limit = (1 + nr_cpus_node(node)) * 3908 cachep->batchcount + cachep->num; 3909 cachep->nodelists[node] = l3; 3910 } 3911 return 0; 3912 3913 fail: 3914 if (!cachep->next.next) { 3915 /* Cache is not active yet. Roll back what we did */ 3916 node--; 3917 while (node >= 0) { 3918 if (cachep->nodelists[node]) { 3919 l3 = cachep->nodelists[node]; 3920 3921 kfree(l3->shared); 3922 free_alien_cache(l3->alien); 3923 kfree(l3); 3924 cachep->nodelists[node] = NULL; 3925 } 3926 node--; 3927 } 3928 } 3929 return -ENOMEM; 3930 } 3931 3932 struct ccupdate_struct { 3933 struct kmem_cache *cachep; 3934 struct array_cache *new[NR_CPUS]; 3935 }; 3936 3937 static void do_ccupdate_local(void *info) 3938 { 3939 struct ccupdate_struct *new = info; 3940 struct array_cache *old; 3941 3942 check_irq_off(); 3943 old = cpu_cache_get(new->cachep); 3944 3945 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3946 new->new[smp_processor_id()] = old; 3947 } 3948 3949 /* Always called with the cache_chain_mutex held */ 3950 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3951 int batchcount, int shared, gfp_t gfp) 3952 { 3953 struct ccupdate_struct *new; 3954 int i; 3955 3956 new = kzalloc(sizeof(*new), gfp); 3957 if (!new) 3958 return -ENOMEM; 3959 3960 for_each_online_cpu(i) { 3961 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, 3962 batchcount, gfp); 3963 if (!new->new[i]) { 3964 for (i--; i >= 0; i--) 3965 kfree(new->new[i]); 3966 kfree(new); 3967 return -ENOMEM; 3968 } 3969 } 3970 new->cachep = cachep; 3971 3972 on_each_cpu(do_ccupdate_local, (void *)new, 1); 3973 3974 check_irq_on(); 3975 cachep->batchcount = batchcount; 3976 cachep->limit = limit; 3977 cachep->shared = shared; 3978 3979 for_each_online_cpu(i) { 3980 struct array_cache *ccold = new->new[i]; 3981 if (!ccold) 3982 continue; 3983 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); 3984 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); 3985 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); 3986 kfree(ccold); 3987 } 3988 kfree(new); 3989 return alloc_kmemlist(cachep, gfp); 3990 } 3991 3992 /* Called with cache_chain_mutex held always */ 3993 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) 3994 { 3995 int err; 3996 int limit, shared; 3997 3998 /* 3999 * The head array serves three purposes: 4000 * - create a LIFO ordering, i.e. return objects that are cache-warm 4001 * - reduce the number of spinlock operations. 4002 * - reduce the number of linked list operations on the slab and 4003 * bufctl chains: array operations are cheaper. 4004 * The numbers are guessed, we should auto-tune as described by 4005 * Bonwick. 4006 */ 4007 if (cachep->buffer_size > 131072) 4008 limit = 1; 4009 else if (cachep->buffer_size > PAGE_SIZE) 4010 limit = 8; 4011 else if (cachep->buffer_size > 1024) 4012 limit = 24; 4013 else if (cachep->buffer_size > 256) 4014 limit = 54; 4015 else 4016 limit = 120; 4017 4018 /* 4019 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 4020 * allocation behaviour: Most allocs on one cpu, most free operations 4021 * on another cpu. For these cases, an efficient object passing between 4022 * cpus is necessary. This is provided by a shared array. The array 4023 * replaces Bonwick's magazine layer. 4024 * On uniprocessor, it's functionally equivalent (but less efficient) 4025 * to a larger limit. Thus disabled by default. 4026 */ 4027 shared = 0; 4028 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) 4029 shared = 8; 4030 4031 #if DEBUG 4032 /* 4033 * With debugging enabled, large batchcount lead to excessively long 4034 * periods with disabled local interrupts. Limit the batchcount 4035 */ 4036 if (limit > 32) 4037 limit = 32; 4038 #endif 4039 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); 4040 if (err) 4041 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 4042 cachep->name, -err); 4043 return err; 4044 } 4045 4046 /* 4047 * Drain an array if it contains any elements taking the l3 lock only if 4048 * necessary. Note that the l3 listlock also protects the array_cache 4049 * if drain_array() is used on the shared array. 4050 */ 4051 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 4052 struct array_cache *ac, int force, int node) 4053 { 4054 int tofree; 4055 4056 if (!ac || !ac->avail) 4057 return; 4058 if (ac->touched && !force) { 4059 ac->touched = 0; 4060 } else { 4061 spin_lock_irq(&l3->list_lock); 4062 if (ac->avail) { 4063 tofree = force ? ac->avail : (ac->limit + 4) / 5; 4064 if (tofree > ac->avail) 4065 tofree = (ac->avail + 1) / 2; 4066 free_block(cachep, ac->entry, tofree, node); 4067 ac->avail -= tofree; 4068 memmove(ac->entry, &(ac->entry[tofree]), 4069 sizeof(void *) * ac->avail); 4070 } 4071 spin_unlock_irq(&l3->list_lock); 4072 } 4073 } 4074 4075 /** 4076 * cache_reap - Reclaim memory from caches. 4077 * @w: work descriptor 4078 * 4079 * Called from workqueue/eventd every few seconds. 4080 * Purpose: 4081 * - clear the per-cpu caches for this CPU. 4082 * - return freeable pages to the main free memory pool. 4083 * 4084 * If we cannot acquire the cache chain mutex then just give up - we'll try 4085 * again on the next iteration. 4086 */ 4087 static void cache_reap(struct work_struct *w) 4088 { 4089 struct kmem_cache *searchp; 4090 struct kmem_list3 *l3; 4091 int node = numa_mem_id(); 4092 struct delayed_work *work = to_delayed_work(w); 4093 4094 if (!mutex_trylock(&cache_chain_mutex)) 4095 /* Give up. Setup the next iteration. */ 4096 goto out; 4097 4098 list_for_each_entry(searchp, &cache_chain, next) { 4099 check_irq_on(); 4100 4101 /* 4102 * We only take the l3 lock if absolutely necessary and we 4103 * have established with reasonable certainty that 4104 * we can do some work if the lock was obtained. 4105 */ 4106 l3 = searchp->nodelists[node]; 4107 4108 reap_alien(searchp, l3); 4109 4110 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 4111 4112 /* 4113 * These are racy checks but it does not matter 4114 * if we skip one check or scan twice. 4115 */ 4116 if (time_after(l3->next_reap, jiffies)) 4117 goto next; 4118 4119 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 4120 4121 drain_array(searchp, l3, l3->shared, 0, node); 4122 4123 if (l3->free_touched) 4124 l3->free_touched = 0; 4125 else { 4126 int freed; 4127 4128 freed = drain_freelist(searchp, l3, (l3->free_limit + 4129 5 * searchp->num - 1) / (5 * searchp->num)); 4130 STATS_ADD_REAPED(searchp, freed); 4131 } 4132 next: 4133 cond_resched(); 4134 } 4135 check_irq_on(); 4136 mutex_unlock(&cache_chain_mutex); 4137 next_reap_node(); 4138 out: 4139 /* Set up the next iteration */ 4140 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); 4141 } 4142 4143 #ifdef CONFIG_SLABINFO 4144 4145 static void print_slabinfo_header(struct seq_file *m) 4146 { 4147 /* 4148 * Output format version, so at least we can change it 4149 * without _too_ many complaints. 4150 */ 4151 #if STATS 4152 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 4153 #else 4154 seq_puts(m, "slabinfo - version: 2.1\n"); 4155 #endif 4156 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4157 "<objperslab> <pagesperslab>"); 4158 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4159 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4160 #if STATS 4161 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 4162 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 4163 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 4164 #endif 4165 seq_putc(m, '\n'); 4166 } 4167 4168 static void *s_start(struct seq_file *m, loff_t *pos) 4169 { 4170 loff_t n = *pos; 4171 4172 mutex_lock(&cache_chain_mutex); 4173 if (!n) 4174 print_slabinfo_header(m); 4175 4176 return seq_list_start(&cache_chain, *pos); 4177 } 4178 4179 static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4180 { 4181 return seq_list_next(p, &cache_chain, pos); 4182 } 4183 4184 static void s_stop(struct seq_file *m, void *p) 4185 { 4186 mutex_unlock(&cache_chain_mutex); 4187 } 4188 4189 static int s_show(struct seq_file *m, void *p) 4190 { 4191 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); 4192 struct slab *slabp; 4193 unsigned long active_objs; 4194 unsigned long num_objs; 4195 unsigned long active_slabs = 0; 4196 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 4197 const char *name; 4198 char *error = NULL; 4199 int node; 4200 struct kmem_list3 *l3; 4201 4202 active_objs = 0; 4203 num_slabs = 0; 4204 for_each_online_node(node) { 4205 l3 = cachep->nodelists[node]; 4206 if (!l3) 4207 continue; 4208 4209 check_irq_on(); 4210 spin_lock_irq(&l3->list_lock); 4211 4212 list_for_each_entry(slabp, &l3->slabs_full, list) { 4213 if (slabp->inuse != cachep->num && !error) 4214 error = "slabs_full accounting error"; 4215 active_objs += cachep->num; 4216 active_slabs++; 4217 } 4218 list_for_each_entry(slabp, &l3->slabs_partial, list) { 4219 if (slabp->inuse == cachep->num && !error) 4220 error = "slabs_partial inuse accounting error"; 4221 if (!slabp->inuse && !error) 4222 error = "slabs_partial/inuse accounting error"; 4223 active_objs += slabp->inuse; 4224 active_slabs++; 4225 } 4226 list_for_each_entry(slabp, &l3->slabs_free, list) { 4227 if (slabp->inuse && !error) 4228 error = "slabs_free/inuse accounting error"; 4229 num_slabs++; 4230 } 4231 free_objects += l3->free_objects; 4232 if (l3->shared) 4233 shared_avail += l3->shared->avail; 4234 4235 spin_unlock_irq(&l3->list_lock); 4236 } 4237 num_slabs += active_slabs; 4238 num_objs = num_slabs * cachep->num; 4239 if (num_objs - active_objs != free_objects && !error) 4240 error = "free_objects accounting error"; 4241 4242 name = cachep->name; 4243 if (error) 4244 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 4245 4246 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 4247 name, active_objs, num_objs, cachep->buffer_size, 4248 cachep->num, (1 << cachep->gfporder)); 4249 seq_printf(m, " : tunables %4u %4u %4u", 4250 cachep->limit, cachep->batchcount, cachep->shared); 4251 seq_printf(m, " : slabdata %6lu %6lu %6lu", 4252 active_slabs, num_slabs, shared_avail); 4253 #if STATS 4254 { /* list3 stats */ 4255 unsigned long high = cachep->high_mark; 4256 unsigned long allocs = cachep->num_allocations; 4257 unsigned long grown = cachep->grown; 4258 unsigned long reaped = cachep->reaped; 4259 unsigned long errors = cachep->errors; 4260 unsigned long max_freeable = cachep->max_freeable; 4261 unsigned long node_allocs = cachep->node_allocs; 4262 unsigned long node_frees = cachep->node_frees; 4263 unsigned long overflows = cachep->node_overflow; 4264 4265 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " 4266 "%4lu %4lu %4lu %4lu %4lu", 4267 allocs, high, grown, 4268 reaped, errors, max_freeable, node_allocs, 4269 node_frees, overflows); 4270 } 4271 /* cpu stats */ 4272 { 4273 unsigned long allochit = atomic_read(&cachep->allochit); 4274 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4275 unsigned long freehit = atomic_read(&cachep->freehit); 4276 unsigned long freemiss = atomic_read(&cachep->freemiss); 4277 4278 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4279 allochit, allocmiss, freehit, freemiss); 4280 } 4281 #endif 4282 seq_putc(m, '\n'); 4283 return 0; 4284 } 4285 4286 /* 4287 * slabinfo_op - iterator that generates /proc/slabinfo 4288 * 4289 * Output layout: 4290 * cache-name 4291 * num-active-objs 4292 * total-objs 4293 * object size 4294 * num-active-slabs 4295 * total-slabs 4296 * num-pages-per-slab 4297 * + further values on SMP and with statistics enabled 4298 */ 4299 4300 static const struct seq_operations slabinfo_op = { 4301 .start = s_start, 4302 .next = s_next, 4303 .stop = s_stop, 4304 .show = s_show, 4305 }; 4306 4307 #define MAX_SLABINFO_WRITE 128 4308 /** 4309 * slabinfo_write - Tuning for the slab allocator 4310 * @file: unused 4311 * @buffer: user buffer 4312 * @count: data length 4313 * @ppos: unused 4314 */ 4315 static ssize_t slabinfo_write(struct file *file, const char __user *buffer, 4316 size_t count, loff_t *ppos) 4317 { 4318 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4319 int limit, batchcount, shared, res; 4320 struct kmem_cache *cachep; 4321 4322 if (count > MAX_SLABINFO_WRITE) 4323 return -EINVAL; 4324 if (copy_from_user(&kbuf, buffer, count)) 4325 return -EFAULT; 4326 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4327 4328 tmp = strchr(kbuf, ' '); 4329 if (!tmp) 4330 return -EINVAL; 4331 *tmp = '\0'; 4332 tmp++; 4333 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4334 return -EINVAL; 4335 4336 /* Find the cache in the chain of caches. */ 4337 mutex_lock(&cache_chain_mutex); 4338 res = -EINVAL; 4339 list_for_each_entry(cachep, &cache_chain, next) { 4340 if (!strcmp(cachep->name, kbuf)) { 4341 if (limit < 1 || batchcount < 1 || 4342 batchcount > limit || shared < 0) { 4343 res = 0; 4344 } else { 4345 res = do_tune_cpucache(cachep, limit, 4346 batchcount, shared, 4347 GFP_KERNEL); 4348 } 4349 break; 4350 } 4351 } 4352 mutex_unlock(&cache_chain_mutex); 4353 if (res >= 0) 4354 res = count; 4355 return res; 4356 } 4357 4358 static int slabinfo_open(struct inode *inode, struct file *file) 4359 { 4360 return seq_open(file, &slabinfo_op); 4361 } 4362 4363 static const struct file_operations proc_slabinfo_operations = { 4364 .open = slabinfo_open, 4365 .read = seq_read, 4366 .write = slabinfo_write, 4367 .llseek = seq_lseek, 4368 .release = seq_release, 4369 }; 4370 4371 #ifdef CONFIG_DEBUG_SLAB_LEAK 4372 4373 static void *leaks_start(struct seq_file *m, loff_t *pos) 4374 { 4375 mutex_lock(&cache_chain_mutex); 4376 return seq_list_start(&cache_chain, *pos); 4377 } 4378 4379 static inline int add_caller(unsigned long *n, unsigned long v) 4380 { 4381 unsigned long *p; 4382 int l; 4383 if (!v) 4384 return 1; 4385 l = n[1]; 4386 p = n + 2; 4387 while (l) { 4388 int i = l/2; 4389 unsigned long *q = p + 2 * i; 4390 if (*q == v) { 4391 q[1]++; 4392 return 1; 4393 } 4394 if (*q > v) { 4395 l = i; 4396 } else { 4397 p = q + 2; 4398 l -= i + 1; 4399 } 4400 } 4401 if (++n[1] == n[0]) 4402 return 0; 4403 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4404 p[0] = v; 4405 p[1] = 1; 4406 return 1; 4407 } 4408 4409 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4410 { 4411 void *p; 4412 int i; 4413 if (n[0] == n[1]) 4414 return; 4415 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4416 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4417 continue; 4418 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4419 return; 4420 } 4421 } 4422 4423 static void show_symbol(struct seq_file *m, unsigned long address) 4424 { 4425 #ifdef CONFIG_KALLSYMS 4426 unsigned long offset, size; 4427 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; 4428 4429 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { 4430 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4431 if (modname[0]) 4432 seq_printf(m, " [%s]", modname); 4433 return; 4434 } 4435 #endif 4436 seq_printf(m, "%p", (void *)address); 4437 } 4438 4439 static int leaks_show(struct seq_file *m, void *p) 4440 { 4441 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); 4442 struct slab *slabp; 4443 struct kmem_list3 *l3; 4444 const char *name; 4445 unsigned long *n = m->private; 4446 int node; 4447 int i; 4448 4449 if (!(cachep->flags & SLAB_STORE_USER)) 4450 return 0; 4451 if (!(cachep->flags & SLAB_RED_ZONE)) 4452 return 0; 4453 4454 /* OK, we can do it */ 4455 4456 n[1] = 0; 4457 4458 for_each_online_node(node) { 4459 l3 = cachep->nodelists[node]; 4460 if (!l3) 4461 continue; 4462 4463 check_irq_on(); 4464 spin_lock_irq(&l3->list_lock); 4465 4466 list_for_each_entry(slabp, &l3->slabs_full, list) 4467 handle_slab(n, cachep, slabp); 4468 list_for_each_entry(slabp, &l3->slabs_partial, list) 4469 handle_slab(n, cachep, slabp); 4470 spin_unlock_irq(&l3->list_lock); 4471 } 4472 name = cachep->name; 4473 if (n[0] == n[1]) { 4474 /* Increase the buffer size */ 4475 mutex_unlock(&cache_chain_mutex); 4476 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4477 if (!m->private) { 4478 /* Too bad, we are really out */ 4479 m->private = n; 4480 mutex_lock(&cache_chain_mutex); 4481 return -ENOMEM; 4482 } 4483 *(unsigned long *)m->private = n[0] * 2; 4484 kfree(n); 4485 mutex_lock(&cache_chain_mutex); 4486 /* Now make sure this entry will be retried */ 4487 m->count = m->size; 4488 return 0; 4489 } 4490 for (i = 0; i < n[1]; i++) { 4491 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4492 show_symbol(m, n[2*i+2]); 4493 seq_putc(m, '\n'); 4494 } 4495 4496 return 0; 4497 } 4498 4499 static const struct seq_operations slabstats_op = { 4500 .start = leaks_start, 4501 .next = s_next, 4502 .stop = s_stop, 4503 .show = leaks_show, 4504 }; 4505 4506 static int slabstats_open(struct inode *inode, struct file *file) 4507 { 4508 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); 4509 int ret = -ENOMEM; 4510 if (n) { 4511 ret = seq_open(file, &slabstats_op); 4512 if (!ret) { 4513 struct seq_file *m = file->private_data; 4514 *n = PAGE_SIZE / (2 * sizeof(unsigned long)); 4515 m->private = n; 4516 n = NULL; 4517 } 4518 kfree(n); 4519 } 4520 return ret; 4521 } 4522 4523 static const struct file_operations proc_slabstats_operations = { 4524 .open = slabstats_open, 4525 .read = seq_read, 4526 .llseek = seq_lseek, 4527 .release = seq_release_private, 4528 }; 4529 #endif 4530 4531 static int __init slab_proc_init(void) 4532 { 4533 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); 4534 #ifdef CONFIG_DEBUG_SLAB_LEAK 4535 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); 4536 #endif 4537 return 0; 4538 } 4539 module_init(slab_proc_init); 4540 #endif 4541 4542 /** 4543 * ksize - get the actual amount of memory allocated for a given object 4544 * @objp: Pointer to the object 4545 * 4546 * kmalloc may internally round up allocations and return more memory 4547 * than requested. ksize() can be used to determine the actual amount of 4548 * memory allocated. The caller may use this additional memory, even though 4549 * a smaller amount of memory was initially specified with the kmalloc call. 4550 * The caller must guarantee that objp points to a valid object previously 4551 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4552 * must not be freed during the duration of the call. 4553 */ 4554 size_t ksize(const void *objp) 4555 { 4556 BUG_ON(!objp); 4557 if (unlikely(objp == ZERO_SIZE_PTR)) 4558 return 0; 4559 4560 return obj_size(virt_to_cache(objp)); 4561 } 4562 EXPORT_SYMBOL(ksize); 4563