slab.c revision 7ae439ce0c01d7db0c70d1542985969e95ef750d
1/* 2 * linux/mm/slab.c 3 * Written by Mark Hemment, 1996/97. 4 * (markhe@nextd.demon.co.uk) 5 * 6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 7 * 8 * Major cleanup, different bufctl logic, per-cpu arrays 9 * (c) 2000 Manfred Spraul 10 * 11 * Cleanup, make the head arrays unconditional, preparation for NUMA 12 * (c) 2002 Manfred Spraul 13 * 14 * An implementation of the Slab Allocator as described in outline in; 15 * UNIX Internals: The New Frontiers by Uresh Vahalia 16 * Pub: Prentice Hall ISBN 0-13-101908-2 17 * or with a little more detail in; 18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 19 * Jeff Bonwick (Sun Microsystems). 20 * Presented at: USENIX Summer 1994 Technical Conference 21 * 22 * The memory is organized in caches, one cache for each object type. 23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 24 * Each cache consists out of many slabs (they are small (usually one 25 * page long) and always contiguous), and each slab contains multiple 26 * initialized objects. 27 * 28 * This means, that your constructor is used only for newly allocated 29 * slabs and you must pass objects with the same intializations to 30 * kmem_cache_free. 31 * 32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 33 * normal). If you need a special memory type, then must create a new 34 * cache for that memory type. 35 * 36 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 37 * full slabs with 0 free objects 38 * partial slabs 39 * empty slabs with no allocated objects 40 * 41 * If partial slabs exist, then new allocations come from these slabs, 42 * otherwise from empty slabs or new slabs are allocated. 43 * 44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 46 * 47 * Each cache has a short per-cpu head array, most allocs 48 * and frees go into that array, and if that array overflows, then 1/2 49 * of the entries in the array are given back into the global cache. 50 * The head array is strictly LIFO and should improve the cache hit rates. 51 * On SMP, it additionally reduces the spinlock operations. 52 * 53 * The c_cpuarray may not be read with enabled local interrupts - 54 * it's changed with a smp_call_function(). 55 * 56 * SMP synchronization: 57 * constructors and destructors are called without any locking. 58 * Several members in struct kmem_cache and struct slab never change, they 59 * are accessed without any locking. 60 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 61 * and local interrupts are disabled so slab code is preempt-safe. 62 * The non-constant members are protected with a per-cache irq spinlock. 63 * 64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 65 * in 2000 - many ideas in the current implementation are derived from 66 * his patch. 67 * 68 * Further notes from the original documentation: 69 * 70 * 11 April '97. Started multi-threading - markhe 71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'. 72 * The sem is only needed when accessing/extending the cache-chain, which 73 * can never happen inside an interrupt (kmem_cache_create(), 74 * kmem_cache_shrink() and kmem_cache_reap()). 75 * 76 * At present, each engine can be growing a cache. This should be blocked. 77 * 78 * 15 March 2005. NUMA slab allocator. 79 * Shai Fultheim <shai@scalex86.org>. 80 * Shobhit Dayal <shobhit@calsoftinc.com> 81 * Alok N Kataria <alokk@calsoftinc.com> 82 * Christoph Lameter <christoph@lameter.com> 83 * 84 * Modified the slab allocator to be node aware on NUMA systems. 85 * Each node has its own list of partial, free and full slabs. 86 * All object allocations for a node occur from node specific slab lists. 87 */ 88 89#include <linux/slab.h> 90#include <linux/mm.h> 91#include <linux/poison.h> 92#include <linux/swap.h> 93#include <linux/cache.h> 94#include <linux/interrupt.h> 95#include <linux/init.h> 96#include <linux/compiler.h> 97#include <linux/cpuset.h> 98#include <linux/seq_file.h> 99#include <linux/notifier.h> 100#include <linux/kallsyms.h> 101#include <linux/cpu.h> 102#include <linux/sysctl.h> 103#include <linux/module.h> 104#include <linux/rcupdate.h> 105#include <linux/string.h> 106#include <linux/uaccess.h> 107#include <linux/nodemask.h> 108#include <linux/mempolicy.h> 109#include <linux/mutex.h> 110#include <linux/fault-inject.h> 111#include <linux/rtmutex.h> 112#include <linux/reciprocal_div.h> 113 114#include <asm/cacheflush.h> 115#include <asm/tlbflush.h> 116#include <asm/page.h> 117 118/* 119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. 120 * 0 for faster, smaller code (especially in the critical paths). 121 * 122 * STATS - 1 to collect stats for /proc/slabinfo. 123 * 0 for faster, smaller code (especially in the critical paths). 124 * 125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 126 */ 127 128#ifdef CONFIG_DEBUG_SLAB 129#define DEBUG 1 130#define STATS 1 131#define FORCED_DEBUG 1 132#else 133#define DEBUG 0 134#define STATS 0 135#define FORCED_DEBUG 0 136#endif 137 138/* Shouldn't this be in a header file somewhere? */ 139#define BYTES_PER_WORD sizeof(void *) 140 141#ifndef cache_line_size 142#define cache_line_size() L1_CACHE_BYTES 143#endif 144 145#ifndef ARCH_KMALLOC_MINALIGN 146/* 147 * Enforce a minimum alignment for the kmalloc caches. 148 * Usually, the kmalloc caches are cache_line_size() aligned, except when 149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. 150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed 151 * alignment larger than the alignment of a 64-bit integer. 152 * ARCH_KMALLOC_MINALIGN allows that. 153 * Note that increasing this value may disable some debug features. 154 */ 155#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 156#endif 157 158#ifndef ARCH_SLAB_MINALIGN 159/* 160 * Enforce a minimum alignment for all caches. 161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD 162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. 163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables 164 * some debug features. 165 */ 166#define ARCH_SLAB_MINALIGN 0 167#endif 168 169#ifndef ARCH_KMALLOC_FLAGS 170#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 171#endif 172 173/* Legal flag mask for kmem_cache_create(). */ 174#if DEBUG 175# define CREATE_MASK (SLAB_RED_ZONE | \ 176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 177 SLAB_CACHE_DMA | \ 178 SLAB_STORE_USER | \ 179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 181#else 182# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 183 SLAB_CACHE_DMA | \ 184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 186#endif 187 188/* 189 * kmem_bufctl_t: 190 * 191 * Bufctl's are used for linking objs within a slab 192 * linked offsets. 193 * 194 * This implementation relies on "struct page" for locating the cache & 195 * slab an object belongs to. 196 * This allows the bufctl structure to be small (one int), but limits 197 * the number of objects a slab (not a cache) can contain when off-slab 198 * bufctls are used. The limit is the size of the largest general cache 199 * that does not use off-slab slabs. 200 * For 32bit archs with 4 kB pages, is this 56. 201 * This is not serious, as it is only for large objects, when it is unwise 202 * to have too many per slab. 203 * Note: This limit can be raised by introducing a general cache whose size 204 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 205 */ 206 207typedef unsigned int kmem_bufctl_t; 208#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 209#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 210#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 211#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 212 213/* 214 * struct slab 215 * 216 * Manages the objs in a slab. Placed either at the beginning of mem allocated 217 * for a slab, or allocated from an general cache. 218 * Slabs are chained into three list: fully used, partial, fully free slabs. 219 */ 220struct slab { 221 struct list_head list; 222 unsigned long colouroff; 223 void *s_mem; /* including colour offset */ 224 unsigned int inuse; /* num of objs active in slab */ 225 kmem_bufctl_t free; 226 unsigned short nodeid; 227}; 228 229/* 230 * struct slab_rcu 231 * 232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 233 * arrange for kmem_freepages to be called via RCU. This is useful if 234 * we need to approach a kernel structure obliquely, from its address 235 * obtained without the usual locking. We can lock the structure to 236 * stabilize it and check it's still at the given address, only if we 237 * can be sure that the memory has not been meanwhile reused for some 238 * other kind of object (which our subsystem's lock might corrupt). 239 * 240 * rcu_read_lock before reading the address, then rcu_read_unlock after 241 * taking the spinlock within the structure expected at that address. 242 * 243 * We assume struct slab_rcu can overlay struct slab when destroying. 244 */ 245struct slab_rcu { 246 struct rcu_head head; 247 struct kmem_cache *cachep; 248 void *addr; 249}; 250 251/* 252 * struct array_cache 253 * 254 * Purpose: 255 * - LIFO ordering, to hand out cache-warm objects from _alloc 256 * - reduce the number of linked list operations 257 * - reduce spinlock operations 258 * 259 * The limit is stored in the per-cpu structure to reduce the data cache 260 * footprint. 261 * 262 */ 263struct array_cache { 264 unsigned int avail; 265 unsigned int limit; 266 unsigned int batchcount; 267 unsigned int touched; 268 spinlock_t lock; 269 void *entry[0]; /* 270 * Must have this definition in here for the proper 271 * alignment of array_cache. Also simplifies accessing 272 * the entries. 273 * [0] is for gcc 2.95. It should really be []. 274 */ 275}; 276 277/* 278 * bootstrap: The caches do not work without cpuarrays anymore, but the 279 * cpuarrays are allocated from the generic caches... 280 */ 281#define BOOT_CPUCACHE_ENTRIES 1 282struct arraycache_init { 283 struct array_cache cache; 284 void *entries[BOOT_CPUCACHE_ENTRIES]; 285}; 286 287/* 288 * The slab lists for all objects. 289 */ 290struct kmem_list3 { 291 struct list_head slabs_partial; /* partial list first, better asm code */ 292 struct list_head slabs_full; 293 struct list_head slabs_free; 294 unsigned long free_objects; 295 unsigned int free_limit; 296 unsigned int colour_next; /* Per-node cache coloring */ 297 spinlock_t list_lock; 298 struct array_cache *shared; /* shared per node */ 299 struct array_cache **alien; /* on other nodes */ 300 unsigned long next_reap; /* updated without locking */ 301 int free_touched; /* updated without locking */ 302}; 303 304/* 305 * Need this for bootstrapping a per node allocator. 306 */ 307#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) 308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 309#define CACHE_CACHE 0 310#define SIZE_AC 1 311#define SIZE_L3 (1 + MAX_NUMNODES) 312 313static int drain_freelist(struct kmem_cache *cache, 314 struct kmem_list3 *l3, int tofree); 315static void free_block(struct kmem_cache *cachep, void **objpp, int len, 316 int node); 317static int enable_cpucache(struct kmem_cache *cachep); 318static void cache_reap(struct work_struct *unused); 319 320/* 321 * This function must be completely optimized away if a constant is passed to 322 * it. Mostly the same as what is in linux/slab.h except it returns an index. 323 */ 324static __always_inline int index_of(const size_t size) 325{ 326 extern void __bad_size(void); 327 328 if (__builtin_constant_p(size)) { 329 int i = 0; 330 331#define CACHE(x) \ 332 if (size <=x) \ 333 return i; \ 334 else \ 335 i++; 336#include "linux/kmalloc_sizes.h" 337#undef CACHE 338 __bad_size(); 339 } else 340 __bad_size(); 341 return 0; 342} 343 344static int slab_early_init = 1; 345 346#define INDEX_AC index_of(sizeof(struct arraycache_init)) 347#define INDEX_L3 index_of(sizeof(struct kmem_list3)) 348 349static void kmem_list3_init(struct kmem_list3 *parent) 350{ 351 INIT_LIST_HEAD(&parent->slabs_full); 352 INIT_LIST_HEAD(&parent->slabs_partial); 353 INIT_LIST_HEAD(&parent->slabs_free); 354 parent->shared = NULL; 355 parent->alien = NULL; 356 parent->colour_next = 0; 357 spin_lock_init(&parent->list_lock); 358 parent->free_objects = 0; 359 parent->free_touched = 0; 360} 361 362#define MAKE_LIST(cachep, listp, slab, nodeid) \ 363 do { \ 364 INIT_LIST_HEAD(listp); \ 365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 366 } while (0) 367 368#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 369 do { \ 370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 373 } while (0) 374 375/* 376 * struct kmem_cache 377 * 378 * manages a cache. 379 */ 380 381struct kmem_cache { 382/* 1) per-cpu data, touched during every alloc/free */ 383 struct array_cache *array[NR_CPUS]; 384/* 2) Cache tunables. Protected by cache_chain_mutex */ 385 unsigned int batchcount; 386 unsigned int limit; 387 unsigned int shared; 388 389 unsigned int buffer_size; 390 u32 reciprocal_buffer_size; 391/* 3) touched by every alloc & free from the backend */ 392 393 unsigned int flags; /* constant flags */ 394 unsigned int num; /* # of objs per slab */ 395 396/* 4) cache_grow/shrink */ 397 /* order of pgs per slab (2^n) */ 398 unsigned int gfporder; 399 400 /* force GFP flags, e.g. GFP_DMA */ 401 gfp_t gfpflags; 402 403 size_t colour; /* cache colouring range */ 404 unsigned int colour_off; /* colour offset */ 405 struct kmem_cache *slabp_cache; 406 unsigned int slab_size; 407 unsigned int dflags; /* dynamic flags */ 408 409 /* constructor func */ 410 void (*ctor) (void *, struct kmem_cache *, unsigned long); 411 412 /* de-constructor func */ 413 void (*dtor) (void *, struct kmem_cache *, unsigned long); 414 415/* 5) cache creation/removal */ 416 const char *name; 417 struct list_head next; 418 419/* 6) statistics */ 420#if STATS 421 unsigned long num_active; 422 unsigned long num_allocations; 423 unsigned long high_mark; 424 unsigned long grown; 425 unsigned long reaped; 426 unsigned long errors; 427 unsigned long max_freeable; 428 unsigned long node_allocs; 429 unsigned long node_frees; 430 unsigned long node_overflow; 431 atomic_t allochit; 432 atomic_t allocmiss; 433 atomic_t freehit; 434 atomic_t freemiss; 435#endif 436#if DEBUG 437 /* 438 * If debugging is enabled, then the allocator can add additional 439 * fields and/or padding to every object. buffer_size contains the total 440 * object size including these internal fields, the following two 441 * variables contain the offset to the user object and its size. 442 */ 443 int obj_offset; 444 int obj_size; 445#endif 446 /* 447 * We put nodelists[] at the end of kmem_cache, because we want to size 448 * this array to nr_node_ids slots instead of MAX_NUMNODES 449 * (see kmem_cache_init()) 450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache 451 * is statically defined, so we reserve the max number of nodes. 452 */ 453 struct kmem_list3 *nodelists[MAX_NUMNODES]; 454 /* 455 * Do not add fields after nodelists[] 456 */ 457}; 458 459#define CFLGS_OFF_SLAB (0x80000000UL) 460#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 461 462#define BATCHREFILL_LIMIT 16 463/* 464 * Optimization question: fewer reaps means less probability for unnessary 465 * cpucache drain/refill cycles. 466 * 467 * OTOH the cpuarrays can contain lots of objects, 468 * which could lock up otherwise freeable slabs. 469 */ 470#define REAPTIMEOUT_CPUC (2*HZ) 471#define REAPTIMEOUT_LIST3 (4*HZ) 472 473#if STATS 474#define STATS_INC_ACTIVE(x) ((x)->num_active++) 475#define STATS_DEC_ACTIVE(x) ((x)->num_active--) 476#define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 477#define STATS_INC_GROWN(x) ((x)->grown++) 478#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 479#define STATS_SET_HIGH(x) \ 480 do { \ 481 if ((x)->num_active > (x)->high_mark) \ 482 (x)->high_mark = (x)->num_active; \ 483 } while (0) 484#define STATS_INC_ERR(x) ((x)->errors++) 485#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 486#define STATS_INC_NODEFREES(x) ((x)->node_frees++) 487#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 488#define STATS_SET_FREEABLE(x, i) \ 489 do { \ 490 if ((x)->max_freeable < i) \ 491 (x)->max_freeable = i; \ 492 } while (0) 493#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 494#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 495#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 496#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 497#else 498#define STATS_INC_ACTIVE(x) do { } while (0) 499#define STATS_DEC_ACTIVE(x) do { } while (0) 500#define STATS_INC_ALLOCED(x) do { } while (0) 501#define STATS_INC_GROWN(x) do { } while (0) 502#define STATS_ADD_REAPED(x,y) do { } while (0) 503#define STATS_SET_HIGH(x) do { } while (0) 504#define STATS_INC_ERR(x) do { } while (0) 505#define STATS_INC_NODEALLOCS(x) do { } while (0) 506#define STATS_INC_NODEFREES(x) do { } while (0) 507#define STATS_INC_ACOVERFLOW(x) do { } while (0) 508#define STATS_SET_FREEABLE(x, i) do { } while (0) 509#define STATS_INC_ALLOCHIT(x) do { } while (0) 510#define STATS_INC_ALLOCMISS(x) do { } while (0) 511#define STATS_INC_FREEHIT(x) do { } while (0) 512#define STATS_INC_FREEMISS(x) do { } while (0) 513#endif 514 515#if DEBUG 516 517/* 518 * memory layout of objects: 519 * 0 : objp 520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 521 * the end of an object is aligned with the end of the real 522 * allocation. Catches writes behind the end of the allocation. 523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 524 * redzone word. 525 * cachep->obj_offset: The real object. 526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 528 * [BYTES_PER_WORD long] 529 */ 530static int obj_offset(struct kmem_cache *cachep) 531{ 532 return cachep->obj_offset; 533} 534 535static int obj_size(struct kmem_cache *cachep) 536{ 537 return cachep->obj_size; 538} 539 540static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 541{ 542 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 543 return (unsigned long long*) (objp + obj_offset(cachep) - 544 sizeof(unsigned long long)); 545} 546 547static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 548{ 549 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 550 if (cachep->flags & SLAB_STORE_USER) 551 return (unsigned long long *)(objp + cachep->buffer_size - 552 sizeof(unsigned long long) - 553 BYTES_PER_WORD); 554 return (unsigned long long *) (objp + cachep->buffer_size - 555 sizeof(unsigned long long)); 556} 557 558static void **dbg_userword(struct kmem_cache *cachep, void *objp) 559{ 560 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 561 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 562} 563 564#else 565 566#define obj_offset(x) 0 567#define obj_size(cachep) (cachep->buffer_size) 568#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 569#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 570#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 571 572#endif 573 574/* 575 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp 576 * order. 577 */ 578#if defined(CONFIG_LARGE_ALLOCS) 579#define MAX_OBJ_ORDER 13 /* up to 32Mb */ 580#define MAX_GFP_ORDER 13 /* up to 32Mb */ 581#elif defined(CONFIG_MMU) 582#define MAX_OBJ_ORDER 5 /* 32 pages */ 583#define MAX_GFP_ORDER 5 /* 32 pages */ 584#else 585#define MAX_OBJ_ORDER 8 /* up to 1Mb */ 586#define MAX_GFP_ORDER 8 /* up to 1Mb */ 587#endif 588 589/* 590 * Do not go above this order unless 0 objects fit into the slab. 591 */ 592#define BREAK_GFP_ORDER_HI 1 593#define BREAK_GFP_ORDER_LO 0 594static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 595 596/* 597 * Functions for storing/retrieving the cachep and or slab from the page 598 * allocator. These are used to find the slab an obj belongs to. With kfree(), 599 * these are used to find the cache which an obj belongs to. 600 */ 601static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 602{ 603 page->lru.next = (struct list_head *)cache; 604} 605 606static inline struct kmem_cache *page_get_cache(struct page *page) 607{ 608 page = compound_head(page); 609 BUG_ON(!PageSlab(page)); 610 return (struct kmem_cache *)page->lru.next; 611} 612 613static inline void page_set_slab(struct page *page, struct slab *slab) 614{ 615 page->lru.prev = (struct list_head *)slab; 616} 617 618static inline struct slab *page_get_slab(struct page *page) 619{ 620 BUG_ON(!PageSlab(page)); 621 return (struct slab *)page->lru.prev; 622} 623 624static inline struct kmem_cache *virt_to_cache(const void *obj) 625{ 626 struct page *page = virt_to_head_page(obj); 627 return page_get_cache(page); 628} 629 630static inline struct slab *virt_to_slab(const void *obj) 631{ 632 struct page *page = virt_to_head_page(obj); 633 return page_get_slab(page); 634} 635 636static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 637 unsigned int idx) 638{ 639 return slab->s_mem + cache->buffer_size * idx; 640} 641 642/* 643 * We want to avoid an expensive divide : (offset / cache->buffer_size) 644 * Using the fact that buffer_size is a constant for a particular cache, 645 * we can replace (offset / cache->buffer_size) by 646 * reciprocal_divide(offset, cache->reciprocal_buffer_size) 647 */ 648static inline unsigned int obj_to_index(const struct kmem_cache *cache, 649 const struct slab *slab, void *obj) 650{ 651 u32 offset = (obj - slab->s_mem); 652 return reciprocal_divide(offset, cache->reciprocal_buffer_size); 653} 654 655/* 656 * These are the default caches for kmalloc. Custom caches can have other sizes. 657 */ 658struct cache_sizes malloc_sizes[] = { 659#define CACHE(x) { .cs_size = (x) }, 660#include <linux/kmalloc_sizes.h> 661 CACHE(ULONG_MAX) 662#undef CACHE 663}; 664EXPORT_SYMBOL(malloc_sizes); 665 666/* Must match cache_sizes above. Out of line to keep cache footprint low. */ 667struct cache_names { 668 char *name; 669 char *name_dma; 670}; 671 672static struct cache_names __initdata cache_names[] = { 673#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 674#include <linux/kmalloc_sizes.h> 675 {NULL,} 676#undef CACHE 677}; 678 679static struct arraycache_init initarray_cache __initdata = 680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 681static struct arraycache_init initarray_generic = 682 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 683 684/* internal cache of cache description objs */ 685static struct kmem_cache cache_cache = { 686 .batchcount = 1, 687 .limit = BOOT_CPUCACHE_ENTRIES, 688 .shared = 1, 689 .buffer_size = sizeof(struct kmem_cache), 690 .name = "kmem_cache", 691}; 692 693#define BAD_ALIEN_MAGIC 0x01020304ul 694 695#ifdef CONFIG_LOCKDEP 696 697/* 698 * Slab sometimes uses the kmalloc slabs to store the slab headers 699 * for other slabs "off slab". 700 * The locking for this is tricky in that it nests within the locks 701 * of all other slabs in a few places; to deal with this special 702 * locking we put on-slab caches into a separate lock-class. 703 * 704 * We set lock class for alien array caches which are up during init. 705 * The lock annotation will be lost if all cpus of a node goes down and 706 * then comes back up during hotplug 707 */ 708static struct lock_class_key on_slab_l3_key; 709static struct lock_class_key on_slab_alc_key; 710 711static inline void init_lock_keys(void) 712 713{ 714 int q; 715 struct cache_sizes *s = malloc_sizes; 716 717 while (s->cs_size != ULONG_MAX) { 718 for_each_node(q) { 719 struct array_cache **alc; 720 int r; 721 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; 722 if (!l3 || OFF_SLAB(s->cs_cachep)) 723 continue; 724 lockdep_set_class(&l3->list_lock, &on_slab_l3_key); 725 alc = l3->alien; 726 /* 727 * FIXME: This check for BAD_ALIEN_MAGIC 728 * should go away when common slab code is taught to 729 * work even without alien caches. 730 * Currently, non NUMA code returns BAD_ALIEN_MAGIC 731 * for alloc_alien_cache, 732 */ 733 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) 734 continue; 735 for_each_node(r) { 736 if (alc[r]) 737 lockdep_set_class(&alc[r]->lock, 738 &on_slab_alc_key); 739 } 740 } 741 s++; 742 } 743} 744#else 745static inline void init_lock_keys(void) 746{ 747} 748#endif 749 750/* 751 * 1. Guard access to the cache-chain. 752 * 2. Protect sanity of cpu_online_map against cpu hotplug events 753 */ 754static DEFINE_MUTEX(cache_chain_mutex); 755static struct list_head cache_chain; 756 757/* 758 * chicken and egg problem: delay the per-cpu array allocation 759 * until the general caches are up. 760 */ 761static enum { 762 NONE, 763 PARTIAL_AC, 764 PARTIAL_L3, 765 FULL 766} g_cpucache_up; 767 768/* 769 * used by boot code to determine if it can use slab based allocator 770 */ 771int slab_is_available(void) 772{ 773 return g_cpucache_up == FULL; 774} 775 776static DEFINE_PER_CPU(struct delayed_work, reap_work); 777 778static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 779{ 780 return cachep->array[smp_processor_id()]; 781} 782 783static inline struct kmem_cache *__find_general_cachep(size_t size, 784 gfp_t gfpflags) 785{ 786 struct cache_sizes *csizep = malloc_sizes; 787 788#if DEBUG 789 /* This happens if someone tries to call 790 * kmem_cache_create(), or __kmalloc(), before 791 * the generic caches are initialized. 792 */ 793 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 794#endif 795 while (size > csizep->cs_size) 796 csizep++; 797 798 /* 799 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 800 * has cs_{dma,}cachep==NULL. Thus no special case 801 * for large kmalloc calls required. 802 */ 803#ifdef CONFIG_ZONE_DMA 804 if (unlikely(gfpflags & GFP_DMA)) 805 return csizep->cs_dmacachep; 806#endif 807 return csizep->cs_cachep; 808} 809 810static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 811{ 812 return __find_general_cachep(size, gfpflags); 813} 814 815static size_t slab_mgmt_size(size_t nr_objs, size_t align) 816{ 817 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 818} 819 820/* 821 * Calculate the number of objects and left-over bytes for a given buffer size. 822 */ 823static void cache_estimate(unsigned long gfporder, size_t buffer_size, 824 size_t align, int flags, size_t *left_over, 825 unsigned int *num) 826{ 827 int nr_objs; 828 size_t mgmt_size; 829 size_t slab_size = PAGE_SIZE << gfporder; 830 831 /* 832 * The slab management structure can be either off the slab or 833 * on it. For the latter case, the memory allocated for a 834 * slab is used for: 835 * 836 * - The struct slab 837 * - One kmem_bufctl_t for each object 838 * - Padding to respect alignment of @align 839 * - @buffer_size bytes for each object 840 * 841 * If the slab management structure is off the slab, then the 842 * alignment will already be calculated into the size. Because 843 * the slabs are all pages aligned, the objects will be at the 844 * correct alignment when allocated. 845 */ 846 if (flags & CFLGS_OFF_SLAB) { 847 mgmt_size = 0; 848 nr_objs = slab_size / buffer_size; 849 850 if (nr_objs > SLAB_LIMIT) 851 nr_objs = SLAB_LIMIT; 852 } else { 853 /* 854 * Ignore padding for the initial guess. The padding 855 * is at most @align-1 bytes, and @buffer_size is at 856 * least @align. In the worst case, this result will 857 * be one greater than the number of objects that fit 858 * into the memory allocation when taking the padding 859 * into account. 860 */ 861 nr_objs = (slab_size - sizeof(struct slab)) / 862 (buffer_size + sizeof(kmem_bufctl_t)); 863 864 /* 865 * This calculated number will be either the right 866 * amount, or one greater than what we want. 867 */ 868 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 869 > slab_size) 870 nr_objs--; 871 872 if (nr_objs > SLAB_LIMIT) 873 nr_objs = SLAB_LIMIT; 874 875 mgmt_size = slab_mgmt_size(nr_objs, align); 876 } 877 *num = nr_objs; 878 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 879} 880 881#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) 882 883static void __slab_error(const char *function, struct kmem_cache *cachep, 884 char *msg) 885{ 886 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 887 function, cachep->name, msg); 888 dump_stack(); 889} 890 891/* 892 * By default on NUMA we use alien caches to stage the freeing of 893 * objects allocated from other nodes. This causes massive memory 894 * inefficiencies when using fake NUMA setup to split memory into a 895 * large number of small nodes, so it can be disabled on the command 896 * line 897 */ 898 899static int use_alien_caches __read_mostly = 1; 900static int __init noaliencache_setup(char *s) 901{ 902 use_alien_caches = 0; 903 return 1; 904} 905__setup("noaliencache", noaliencache_setup); 906 907#ifdef CONFIG_NUMA 908/* 909 * Special reaping functions for NUMA systems called from cache_reap(). 910 * These take care of doing round robin flushing of alien caches (containing 911 * objects freed on different nodes from which they were allocated) and the 912 * flushing of remote pcps by calling drain_node_pages. 913 */ 914static DEFINE_PER_CPU(unsigned long, reap_node); 915 916static void init_reap_node(int cpu) 917{ 918 int node; 919 920 node = next_node(cpu_to_node(cpu), node_online_map); 921 if (node == MAX_NUMNODES) 922 node = first_node(node_online_map); 923 924 per_cpu(reap_node, cpu) = node; 925} 926 927static void next_reap_node(void) 928{ 929 int node = __get_cpu_var(reap_node); 930 931 /* 932 * Also drain per cpu pages on remote zones 933 */ 934 if (node != numa_node_id()) 935 drain_node_pages(node); 936 937 node = next_node(node, node_online_map); 938 if (unlikely(node >= MAX_NUMNODES)) 939 node = first_node(node_online_map); 940 __get_cpu_var(reap_node) = node; 941} 942 943#else 944#define init_reap_node(cpu) do { } while (0) 945#define next_reap_node(void) do { } while (0) 946#endif 947 948/* 949 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 950 * via the workqueue/eventd. 951 * Add the CPU number into the expiration time to minimize the possibility of 952 * the CPUs getting into lockstep and contending for the global cache chain 953 * lock. 954 */ 955static void __devinit start_cpu_timer(int cpu) 956{ 957 struct delayed_work *reap_work = &per_cpu(reap_work, cpu); 958 959 /* 960 * When this gets called from do_initcalls via cpucache_init(), 961 * init_workqueues() has already run, so keventd will be setup 962 * at that time. 963 */ 964 if (keventd_up() && reap_work->work.func == NULL) { 965 init_reap_node(cpu); 966 INIT_DELAYED_WORK(reap_work, cache_reap); 967 schedule_delayed_work_on(cpu, reap_work, 968 __round_jiffies_relative(HZ, cpu)); 969 } 970} 971 972static struct array_cache *alloc_arraycache(int node, int entries, 973 int batchcount) 974{ 975 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 976 struct array_cache *nc = NULL; 977 978 nc = kmalloc_node(memsize, GFP_KERNEL, node); 979 if (nc) { 980 nc->avail = 0; 981 nc->limit = entries; 982 nc->batchcount = batchcount; 983 nc->touched = 0; 984 spin_lock_init(&nc->lock); 985 } 986 return nc; 987} 988 989/* 990 * Transfer objects in one arraycache to another. 991 * Locking must be handled by the caller. 992 * 993 * Return the number of entries transferred. 994 */ 995static int transfer_objects(struct array_cache *to, 996 struct array_cache *from, unsigned int max) 997{ 998 /* Figure out how many entries to transfer */ 999 int nr = min(min(from->avail, max), to->limit - to->avail); 1000 1001 if (!nr) 1002 return 0; 1003 1004 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 1005 sizeof(void *) *nr); 1006 1007 from->avail -= nr; 1008 to->avail += nr; 1009 to->touched = 1; 1010 return nr; 1011} 1012 1013#ifndef CONFIG_NUMA 1014 1015#define drain_alien_cache(cachep, alien) do { } while (0) 1016#define reap_alien(cachep, l3) do { } while (0) 1017 1018static inline struct array_cache **alloc_alien_cache(int node, int limit) 1019{ 1020 return (struct array_cache **)BAD_ALIEN_MAGIC; 1021} 1022 1023static inline void free_alien_cache(struct array_cache **ac_ptr) 1024{ 1025} 1026 1027static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1028{ 1029 return 0; 1030} 1031 1032static inline void *alternate_node_alloc(struct kmem_cache *cachep, 1033 gfp_t flags) 1034{ 1035 return NULL; 1036} 1037 1038static inline void *____cache_alloc_node(struct kmem_cache *cachep, 1039 gfp_t flags, int nodeid) 1040{ 1041 return NULL; 1042} 1043 1044#else /* CONFIG_NUMA */ 1045 1046static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 1047static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 1048 1049static struct array_cache **alloc_alien_cache(int node, int limit) 1050{ 1051 struct array_cache **ac_ptr; 1052 int memsize = sizeof(void *) * nr_node_ids; 1053 int i; 1054 1055 if (limit > 1) 1056 limit = 12; 1057 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); 1058 if (ac_ptr) { 1059 for_each_node(i) { 1060 if (i == node || !node_online(i)) { 1061 ac_ptr[i] = NULL; 1062 continue; 1063 } 1064 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); 1065 if (!ac_ptr[i]) { 1066 for (i--; i <= 0; i--) 1067 kfree(ac_ptr[i]); 1068 kfree(ac_ptr); 1069 return NULL; 1070 } 1071 } 1072 } 1073 return ac_ptr; 1074} 1075 1076static void free_alien_cache(struct array_cache **ac_ptr) 1077{ 1078 int i; 1079 1080 if (!ac_ptr) 1081 return; 1082 for_each_node(i) 1083 kfree(ac_ptr[i]); 1084 kfree(ac_ptr); 1085} 1086 1087static void __drain_alien_cache(struct kmem_cache *cachep, 1088 struct array_cache *ac, int node) 1089{ 1090 struct kmem_list3 *rl3 = cachep->nodelists[node]; 1091 1092 if (ac->avail) { 1093 spin_lock(&rl3->list_lock); 1094 /* 1095 * Stuff objects into the remote nodes shared array first. 1096 * That way we could avoid the overhead of putting the objects 1097 * into the free lists and getting them back later. 1098 */ 1099 if (rl3->shared) 1100 transfer_objects(rl3->shared, ac, ac->limit); 1101 1102 free_block(cachep, ac->entry, ac->avail, node); 1103 ac->avail = 0; 1104 spin_unlock(&rl3->list_lock); 1105 } 1106} 1107 1108/* 1109 * Called from cache_reap() to regularly drain alien caches round robin. 1110 */ 1111static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 1112{ 1113 int node = __get_cpu_var(reap_node); 1114 1115 if (l3->alien) { 1116 struct array_cache *ac = l3->alien[node]; 1117 1118 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1119 __drain_alien_cache(cachep, ac, node); 1120 spin_unlock_irq(&ac->lock); 1121 } 1122 } 1123} 1124 1125static void drain_alien_cache(struct kmem_cache *cachep, 1126 struct array_cache **alien) 1127{ 1128 int i = 0; 1129 struct array_cache *ac; 1130 unsigned long flags; 1131 1132 for_each_online_node(i) { 1133 ac = alien[i]; 1134 if (ac) { 1135 spin_lock_irqsave(&ac->lock, flags); 1136 __drain_alien_cache(cachep, ac, i); 1137 spin_unlock_irqrestore(&ac->lock, flags); 1138 } 1139 } 1140} 1141 1142static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1143{ 1144 struct slab *slabp = virt_to_slab(objp); 1145 int nodeid = slabp->nodeid; 1146 struct kmem_list3 *l3; 1147 struct array_cache *alien = NULL; 1148 int node; 1149 1150 node = numa_node_id(); 1151 1152 /* 1153 * Make sure we are not freeing a object from another node to the array 1154 * cache on this cpu. 1155 */ 1156 if (likely(slabp->nodeid == node)) 1157 return 0; 1158 1159 l3 = cachep->nodelists[node]; 1160 STATS_INC_NODEFREES(cachep); 1161 if (l3->alien && l3->alien[nodeid]) { 1162 alien = l3->alien[nodeid]; 1163 spin_lock(&alien->lock); 1164 if (unlikely(alien->avail == alien->limit)) { 1165 STATS_INC_ACOVERFLOW(cachep); 1166 __drain_alien_cache(cachep, alien, nodeid); 1167 } 1168 alien->entry[alien->avail++] = objp; 1169 spin_unlock(&alien->lock); 1170 } else { 1171 spin_lock(&(cachep->nodelists[nodeid])->list_lock); 1172 free_block(cachep, &objp, 1, nodeid); 1173 spin_unlock(&(cachep->nodelists[nodeid])->list_lock); 1174 } 1175 return 1; 1176} 1177#endif 1178 1179static int __cpuinit cpuup_callback(struct notifier_block *nfb, 1180 unsigned long action, void *hcpu) 1181{ 1182 long cpu = (long)hcpu; 1183 struct kmem_cache *cachep; 1184 struct kmem_list3 *l3 = NULL; 1185 int node = cpu_to_node(cpu); 1186 int memsize = sizeof(struct kmem_list3); 1187 1188 switch (action) { 1189 case CPU_UP_PREPARE: 1190 mutex_lock(&cache_chain_mutex); 1191 /* 1192 * We need to do this right in the beginning since 1193 * alloc_arraycache's are going to use this list. 1194 * kmalloc_node allows us to add the slab to the right 1195 * kmem_list3 and not this cpu's kmem_list3 1196 */ 1197 1198 list_for_each_entry(cachep, &cache_chain, next) { 1199 /* 1200 * Set up the size64 kmemlist for cpu before we can 1201 * begin anything. Make sure some other cpu on this 1202 * node has not already allocated this 1203 */ 1204 if (!cachep->nodelists[node]) { 1205 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1206 if (!l3) 1207 goto bad; 1208 kmem_list3_init(l3); 1209 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1210 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1211 1212 /* 1213 * The l3s don't come and go as CPUs come and 1214 * go. cache_chain_mutex is sufficient 1215 * protection here. 1216 */ 1217 cachep->nodelists[node] = l3; 1218 } 1219 1220 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1221 cachep->nodelists[node]->free_limit = 1222 (1 + nr_cpus_node(node)) * 1223 cachep->batchcount + cachep->num; 1224 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1225 } 1226 1227 /* 1228 * Now we can go ahead with allocating the shared arrays and 1229 * array caches 1230 */ 1231 list_for_each_entry(cachep, &cache_chain, next) { 1232 struct array_cache *nc; 1233 struct array_cache *shared = NULL; 1234 struct array_cache **alien = NULL; 1235 1236 nc = alloc_arraycache(node, cachep->limit, 1237 cachep->batchcount); 1238 if (!nc) 1239 goto bad; 1240 if (cachep->shared) { 1241 shared = alloc_arraycache(node, 1242 cachep->shared * cachep->batchcount, 1243 0xbaadf00d); 1244 if (!shared) 1245 goto bad; 1246 } 1247 if (use_alien_caches) { 1248 alien = alloc_alien_cache(node, cachep->limit); 1249 if (!alien) 1250 goto bad; 1251 } 1252 cachep->array[cpu] = nc; 1253 l3 = cachep->nodelists[node]; 1254 BUG_ON(!l3); 1255 1256 spin_lock_irq(&l3->list_lock); 1257 if (!l3->shared) { 1258 /* 1259 * We are serialised from CPU_DEAD or 1260 * CPU_UP_CANCELLED by the cpucontrol lock 1261 */ 1262 l3->shared = shared; 1263 shared = NULL; 1264 } 1265#ifdef CONFIG_NUMA 1266 if (!l3->alien) { 1267 l3->alien = alien; 1268 alien = NULL; 1269 } 1270#endif 1271 spin_unlock_irq(&l3->list_lock); 1272 kfree(shared); 1273 free_alien_cache(alien); 1274 } 1275 break; 1276 case CPU_ONLINE: 1277 mutex_unlock(&cache_chain_mutex); 1278 start_cpu_timer(cpu); 1279 break; 1280#ifdef CONFIG_HOTPLUG_CPU 1281 case CPU_DOWN_PREPARE: 1282 mutex_lock(&cache_chain_mutex); 1283 break; 1284 case CPU_DOWN_FAILED: 1285 mutex_unlock(&cache_chain_mutex); 1286 break; 1287 case CPU_DEAD: 1288 /* 1289 * Even if all the cpus of a node are down, we don't free the 1290 * kmem_list3 of any cache. This to avoid a race between 1291 * cpu_down, and a kmalloc allocation from another cpu for 1292 * memory from the node of the cpu going down. The list3 1293 * structure is usually allocated from kmem_cache_create() and 1294 * gets destroyed at kmem_cache_destroy(). 1295 */ 1296 /* fall thru */ 1297#endif 1298 case CPU_UP_CANCELED: 1299 list_for_each_entry(cachep, &cache_chain, next) { 1300 struct array_cache *nc; 1301 struct array_cache *shared; 1302 struct array_cache **alien; 1303 cpumask_t mask; 1304 1305 mask = node_to_cpumask(node); 1306 /* cpu is dead; no one can alloc from it. */ 1307 nc = cachep->array[cpu]; 1308 cachep->array[cpu] = NULL; 1309 l3 = cachep->nodelists[node]; 1310 1311 if (!l3) 1312 goto free_array_cache; 1313 1314 spin_lock_irq(&l3->list_lock); 1315 1316 /* Free limit for this kmem_list3 */ 1317 l3->free_limit -= cachep->batchcount; 1318 if (nc) 1319 free_block(cachep, nc->entry, nc->avail, node); 1320 1321 if (!cpus_empty(mask)) { 1322 spin_unlock_irq(&l3->list_lock); 1323 goto free_array_cache; 1324 } 1325 1326 shared = l3->shared; 1327 if (shared) { 1328 free_block(cachep, shared->entry, 1329 shared->avail, node); 1330 l3->shared = NULL; 1331 } 1332 1333 alien = l3->alien; 1334 l3->alien = NULL; 1335 1336 spin_unlock_irq(&l3->list_lock); 1337 1338 kfree(shared); 1339 if (alien) { 1340 drain_alien_cache(cachep, alien); 1341 free_alien_cache(alien); 1342 } 1343free_array_cache: 1344 kfree(nc); 1345 } 1346 /* 1347 * In the previous loop, all the objects were freed to 1348 * the respective cache's slabs, now we can go ahead and 1349 * shrink each nodelist to its limit. 1350 */ 1351 list_for_each_entry(cachep, &cache_chain, next) { 1352 l3 = cachep->nodelists[node]; 1353 if (!l3) 1354 continue; 1355 drain_freelist(cachep, l3, l3->free_objects); 1356 } 1357 mutex_unlock(&cache_chain_mutex); 1358 break; 1359 } 1360 return NOTIFY_OK; 1361bad: 1362 return NOTIFY_BAD; 1363} 1364 1365static struct notifier_block __cpuinitdata cpucache_notifier = { 1366 &cpuup_callback, NULL, 0 1367}; 1368 1369/* 1370 * swap the static kmem_list3 with kmalloced memory 1371 */ 1372static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1373 int nodeid) 1374{ 1375 struct kmem_list3 *ptr; 1376 1377 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); 1378 BUG_ON(!ptr); 1379 1380 local_irq_disable(); 1381 memcpy(ptr, list, sizeof(struct kmem_list3)); 1382 /* 1383 * Do not assume that spinlocks can be initialized via memcpy: 1384 */ 1385 spin_lock_init(&ptr->list_lock); 1386 1387 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1388 cachep->nodelists[nodeid] = ptr; 1389 local_irq_enable(); 1390} 1391 1392/* 1393 * Initialisation. Called after the page allocator have been initialised and 1394 * before smp_init(). 1395 */ 1396void __init kmem_cache_init(void) 1397{ 1398 size_t left_over; 1399 struct cache_sizes *sizes; 1400 struct cache_names *names; 1401 int i; 1402 int order; 1403 int node; 1404 1405 if (num_possible_nodes() == 1) 1406 use_alien_caches = 0; 1407 1408 for (i = 0; i < NUM_INIT_LISTS; i++) { 1409 kmem_list3_init(&initkmem_list3[i]); 1410 if (i < MAX_NUMNODES) 1411 cache_cache.nodelists[i] = NULL; 1412 } 1413 1414 /* 1415 * Fragmentation resistance on low memory - only use bigger 1416 * page orders on machines with more than 32MB of memory. 1417 */ 1418 if (num_physpages > (32 << 20) >> PAGE_SHIFT) 1419 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1420 1421 /* Bootstrap is tricky, because several objects are allocated 1422 * from caches that do not exist yet: 1423 * 1) initialize the cache_cache cache: it contains the struct 1424 * kmem_cache structures of all caches, except cache_cache itself: 1425 * cache_cache is statically allocated. 1426 * Initially an __init data area is used for the head array and the 1427 * kmem_list3 structures, it's replaced with a kmalloc allocated 1428 * array at the end of the bootstrap. 1429 * 2) Create the first kmalloc cache. 1430 * The struct kmem_cache for the new cache is allocated normally. 1431 * An __init data area is used for the head array. 1432 * 3) Create the remaining kmalloc caches, with minimally sized 1433 * head arrays. 1434 * 4) Replace the __init data head arrays for cache_cache and the first 1435 * kmalloc cache with kmalloc allocated arrays. 1436 * 5) Replace the __init data for kmem_list3 for cache_cache and 1437 * the other cache's with kmalloc allocated memory. 1438 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1439 */ 1440 1441 node = numa_node_id(); 1442 1443 /* 1) create the cache_cache */ 1444 INIT_LIST_HEAD(&cache_chain); 1445 list_add(&cache_cache.next, &cache_chain); 1446 cache_cache.colour_off = cache_line_size(); 1447 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1448 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; 1449 1450 /* 1451 * struct kmem_cache size depends on nr_node_ids, which 1452 * can be less than MAX_NUMNODES. 1453 */ 1454 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + 1455 nr_node_ids * sizeof(struct kmem_list3 *); 1456#if DEBUG 1457 cache_cache.obj_size = cache_cache.buffer_size; 1458#endif 1459 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1460 cache_line_size()); 1461 cache_cache.reciprocal_buffer_size = 1462 reciprocal_value(cache_cache.buffer_size); 1463 1464 for (order = 0; order < MAX_ORDER; order++) { 1465 cache_estimate(order, cache_cache.buffer_size, 1466 cache_line_size(), 0, &left_over, &cache_cache.num); 1467 if (cache_cache.num) 1468 break; 1469 } 1470 BUG_ON(!cache_cache.num); 1471 cache_cache.gfporder = order; 1472 cache_cache.colour = left_over / cache_cache.colour_off; 1473 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1474 sizeof(struct slab), cache_line_size()); 1475 1476 /* 2+3) create the kmalloc caches */ 1477 sizes = malloc_sizes; 1478 names = cache_names; 1479 1480 /* 1481 * Initialize the caches that provide memory for the array cache and the 1482 * kmem_list3 structures first. Without this, further allocations will 1483 * bug. 1484 */ 1485 1486 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1487 sizes[INDEX_AC].cs_size, 1488 ARCH_KMALLOC_MINALIGN, 1489 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1490 NULL, NULL); 1491 1492 if (INDEX_AC != INDEX_L3) { 1493 sizes[INDEX_L3].cs_cachep = 1494 kmem_cache_create(names[INDEX_L3].name, 1495 sizes[INDEX_L3].cs_size, 1496 ARCH_KMALLOC_MINALIGN, 1497 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1498 NULL, NULL); 1499 } 1500 1501 slab_early_init = 0; 1502 1503 while (sizes->cs_size != ULONG_MAX) { 1504 /* 1505 * For performance, all the general caches are L1 aligned. 1506 * This should be particularly beneficial on SMP boxes, as it 1507 * eliminates "false sharing". 1508 * Note for systems short on memory removing the alignment will 1509 * allow tighter packing of the smaller caches. 1510 */ 1511 if (!sizes->cs_cachep) { 1512 sizes->cs_cachep = kmem_cache_create(names->name, 1513 sizes->cs_size, 1514 ARCH_KMALLOC_MINALIGN, 1515 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1516 NULL, NULL); 1517 } 1518#ifdef CONFIG_ZONE_DMA 1519 sizes->cs_dmacachep = kmem_cache_create( 1520 names->name_dma, 1521 sizes->cs_size, 1522 ARCH_KMALLOC_MINALIGN, 1523 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1524 SLAB_PANIC, 1525 NULL, NULL); 1526#endif 1527 sizes++; 1528 names++; 1529 } 1530 /* 4) Replace the bootstrap head arrays */ 1531 { 1532 struct array_cache *ptr; 1533 1534 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1535 1536 local_irq_disable(); 1537 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1538 memcpy(ptr, cpu_cache_get(&cache_cache), 1539 sizeof(struct arraycache_init)); 1540 /* 1541 * Do not assume that spinlocks can be initialized via memcpy: 1542 */ 1543 spin_lock_init(&ptr->lock); 1544 1545 cache_cache.array[smp_processor_id()] = ptr; 1546 local_irq_enable(); 1547 1548 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1549 1550 local_irq_disable(); 1551 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1552 != &initarray_generic.cache); 1553 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1554 sizeof(struct arraycache_init)); 1555 /* 1556 * Do not assume that spinlocks can be initialized via memcpy: 1557 */ 1558 spin_lock_init(&ptr->lock); 1559 1560 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1561 ptr; 1562 local_irq_enable(); 1563 } 1564 /* 5) Replace the bootstrap kmem_list3's */ 1565 { 1566 int nid; 1567 1568 /* Replace the static kmem_list3 structures for the boot cpu */ 1569 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); 1570 1571 for_each_online_node(nid) { 1572 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1573 &initkmem_list3[SIZE_AC + nid], nid); 1574 1575 if (INDEX_AC != INDEX_L3) { 1576 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1577 &initkmem_list3[SIZE_L3 + nid], nid); 1578 } 1579 } 1580 } 1581 1582 /* 6) resize the head arrays to their final sizes */ 1583 { 1584 struct kmem_cache *cachep; 1585 mutex_lock(&cache_chain_mutex); 1586 list_for_each_entry(cachep, &cache_chain, next) 1587 if (enable_cpucache(cachep)) 1588 BUG(); 1589 mutex_unlock(&cache_chain_mutex); 1590 } 1591 1592 /* Annotate slab for lockdep -- annotate the malloc caches */ 1593 init_lock_keys(); 1594 1595 1596 /* Done! */ 1597 g_cpucache_up = FULL; 1598 1599 /* 1600 * Register a cpu startup notifier callback that initializes 1601 * cpu_cache_get for all new cpus 1602 */ 1603 register_cpu_notifier(&cpucache_notifier); 1604 1605 /* 1606 * The reap timers are started later, with a module init call: That part 1607 * of the kernel is not yet operational. 1608 */ 1609} 1610 1611static int __init cpucache_init(void) 1612{ 1613 int cpu; 1614 1615 /* 1616 * Register the timers that return unneeded pages to the page allocator 1617 */ 1618 for_each_online_cpu(cpu) 1619 start_cpu_timer(cpu); 1620 return 0; 1621} 1622__initcall(cpucache_init); 1623 1624/* 1625 * Interface to system's page allocator. No need to hold the cache-lock. 1626 * 1627 * If we requested dmaable memory, we will get it. Even if we 1628 * did not request dmaable memory, we might get it, but that 1629 * would be relatively rare and ignorable. 1630 */ 1631static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1632{ 1633 struct page *page; 1634 int nr_pages; 1635 int i; 1636 1637#ifndef CONFIG_MMU 1638 /* 1639 * Nommu uses slab's for process anonymous memory allocations, and thus 1640 * requires __GFP_COMP to properly refcount higher order allocations 1641 */ 1642 flags |= __GFP_COMP; 1643#endif 1644 1645 flags |= cachep->gfpflags; 1646 1647 page = alloc_pages_node(nodeid, flags, cachep->gfporder); 1648 if (!page) 1649 return NULL; 1650 1651 nr_pages = (1 << cachep->gfporder); 1652 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1653 add_zone_page_state(page_zone(page), 1654 NR_SLAB_RECLAIMABLE, nr_pages); 1655 else 1656 add_zone_page_state(page_zone(page), 1657 NR_SLAB_UNRECLAIMABLE, nr_pages); 1658 for (i = 0; i < nr_pages; i++) 1659 __SetPageSlab(page + i); 1660 return page_address(page); 1661} 1662 1663/* 1664 * Interface to system's page release. 1665 */ 1666static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1667{ 1668 unsigned long i = (1 << cachep->gfporder); 1669 struct page *page = virt_to_page(addr); 1670 const unsigned long nr_freed = i; 1671 1672 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1673 sub_zone_page_state(page_zone(page), 1674 NR_SLAB_RECLAIMABLE, nr_freed); 1675 else 1676 sub_zone_page_state(page_zone(page), 1677 NR_SLAB_UNRECLAIMABLE, nr_freed); 1678 while (i--) { 1679 BUG_ON(!PageSlab(page)); 1680 __ClearPageSlab(page); 1681 page++; 1682 } 1683 if (current->reclaim_state) 1684 current->reclaim_state->reclaimed_slab += nr_freed; 1685 free_pages((unsigned long)addr, cachep->gfporder); 1686} 1687 1688static void kmem_rcu_free(struct rcu_head *head) 1689{ 1690 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1691 struct kmem_cache *cachep = slab_rcu->cachep; 1692 1693 kmem_freepages(cachep, slab_rcu->addr); 1694 if (OFF_SLAB(cachep)) 1695 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1696} 1697 1698#if DEBUG 1699 1700#ifdef CONFIG_DEBUG_PAGEALLOC 1701static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1702 unsigned long caller) 1703{ 1704 int size = obj_size(cachep); 1705 1706 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1707 1708 if (size < 5 * sizeof(unsigned long)) 1709 return; 1710 1711 *addr++ = 0x12345678; 1712 *addr++ = caller; 1713 *addr++ = smp_processor_id(); 1714 size -= 3 * sizeof(unsigned long); 1715 { 1716 unsigned long *sptr = &caller; 1717 unsigned long svalue; 1718 1719 while (!kstack_end(sptr)) { 1720 svalue = *sptr++; 1721 if (kernel_text_address(svalue)) { 1722 *addr++ = svalue; 1723 size -= sizeof(unsigned long); 1724 if (size <= sizeof(unsigned long)) 1725 break; 1726 } 1727 } 1728 1729 } 1730 *addr++ = 0x87654321; 1731} 1732#endif 1733 1734static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1735{ 1736 int size = obj_size(cachep); 1737 addr = &((char *)addr)[obj_offset(cachep)]; 1738 1739 memset(addr, val, size); 1740 *(unsigned char *)(addr + size - 1) = POISON_END; 1741} 1742 1743static void dump_line(char *data, int offset, int limit) 1744{ 1745 int i; 1746 unsigned char error = 0; 1747 int bad_count = 0; 1748 1749 printk(KERN_ERR "%03x:", offset); 1750 for (i = 0; i < limit; i++) { 1751 if (data[offset + i] != POISON_FREE) { 1752 error = data[offset + i]; 1753 bad_count++; 1754 } 1755 printk(" %02x", (unsigned char)data[offset + i]); 1756 } 1757 printk("\n"); 1758 1759 if (bad_count == 1) { 1760 error ^= POISON_FREE; 1761 if (!(error & (error - 1))) { 1762 printk(KERN_ERR "Single bit error detected. Probably " 1763 "bad RAM.\n"); 1764#ifdef CONFIG_X86 1765 printk(KERN_ERR "Run memtest86+ or a similar memory " 1766 "test tool.\n"); 1767#else 1768 printk(KERN_ERR "Run a memory test tool.\n"); 1769#endif 1770 } 1771 } 1772} 1773#endif 1774 1775#if DEBUG 1776 1777static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1778{ 1779 int i, size; 1780 char *realobj; 1781 1782 if (cachep->flags & SLAB_RED_ZONE) { 1783 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", 1784 *dbg_redzone1(cachep, objp), 1785 *dbg_redzone2(cachep, objp)); 1786 } 1787 1788 if (cachep->flags & SLAB_STORE_USER) { 1789 printk(KERN_ERR "Last user: [<%p>]", 1790 *dbg_userword(cachep, objp)); 1791 print_symbol("(%s)", 1792 (unsigned long)*dbg_userword(cachep, objp)); 1793 printk("\n"); 1794 } 1795 realobj = (char *)objp + obj_offset(cachep); 1796 size = obj_size(cachep); 1797 for (i = 0; i < size && lines; i += 16, lines--) { 1798 int limit; 1799 limit = 16; 1800 if (i + limit > size) 1801 limit = size - i; 1802 dump_line(realobj, i, limit); 1803 } 1804} 1805 1806static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1807{ 1808 char *realobj; 1809 int size, i; 1810 int lines = 0; 1811 1812 realobj = (char *)objp + obj_offset(cachep); 1813 size = obj_size(cachep); 1814 1815 for (i = 0; i < size; i++) { 1816 char exp = POISON_FREE; 1817 if (i == size - 1) 1818 exp = POISON_END; 1819 if (realobj[i] != exp) { 1820 int limit; 1821 /* Mismatch ! */ 1822 /* Print header */ 1823 if (lines == 0) { 1824 printk(KERN_ERR 1825 "Slab corruption: %s start=%p, len=%d\n", 1826 cachep->name, realobj, size); 1827 print_objinfo(cachep, objp, 0); 1828 } 1829 /* Hexdump the affected line */ 1830 i = (i / 16) * 16; 1831 limit = 16; 1832 if (i + limit > size) 1833 limit = size - i; 1834 dump_line(realobj, i, limit); 1835 i += 16; 1836 lines++; 1837 /* Limit to 5 lines */ 1838 if (lines > 5) 1839 break; 1840 } 1841 } 1842 if (lines != 0) { 1843 /* Print some data about the neighboring objects, if they 1844 * exist: 1845 */ 1846 struct slab *slabp = virt_to_slab(objp); 1847 unsigned int objnr; 1848 1849 objnr = obj_to_index(cachep, slabp, objp); 1850 if (objnr) { 1851 objp = index_to_obj(cachep, slabp, objnr - 1); 1852 realobj = (char *)objp + obj_offset(cachep); 1853 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1854 realobj, size); 1855 print_objinfo(cachep, objp, 2); 1856 } 1857 if (objnr + 1 < cachep->num) { 1858 objp = index_to_obj(cachep, slabp, objnr + 1); 1859 realobj = (char *)objp + obj_offset(cachep); 1860 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1861 realobj, size); 1862 print_objinfo(cachep, objp, 2); 1863 } 1864 } 1865} 1866#endif 1867 1868#if DEBUG 1869/** 1870 * slab_destroy_objs - destroy a slab and its objects 1871 * @cachep: cache pointer being destroyed 1872 * @slabp: slab pointer being destroyed 1873 * 1874 * Call the registered destructor for each object in a slab that is being 1875 * destroyed. 1876 */ 1877static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1878{ 1879 int i; 1880 for (i = 0; i < cachep->num; i++) { 1881 void *objp = index_to_obj(cachep, slabp, i); 1882 1883 if (cachep->flags & SLAB_POISON) { 1884#ifdef CONFIG_DEBUG_PAGEALLOC 1885 if (cachep->buffer_size % PAGE_SIZE == 0 && 1886 OFF_SLAB(cachep)) 1887 kernel_map_pages(virt_to_page(objp), 1888 cachep->buffer_size / PAGE_SIZE, 1); 1889 else 1890 check_poison_obj(cachep, objp); 1891#else 1892 check_poison_obj(cachep, objp); 1893#endif 1894 } 1895 if (cachep->flags & SLAB_RED_ZONE) { 1896 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1897 slab_error(cachep, "start of a freed object " 1898 "was overwritten"); 1899 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1900 slab_error(cachep, "end of a freed object " 1901 "was overwritten"); 1902 } 1903 if (cachep->dtor && !(cachep->flags & SLAB_POISON)) 1904 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); 1905 } 1906} 1907#else 1908static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1909{ 1910 if (cachep->dtor) { 1911 int i; 1912 for (i = 0; i < cachep->num; i++) { 1913 void *objp = index_to_obj(cachep, slabp, i); 1914 (cachep->dtor) (objp, cachep, 0); 1915 } 1916 } 1917} 1918#endif 1919 1920/** 1921 * slab_destroy - destroy and release all objects in a slab 1922 * @cachep: cache pointer being destroyed 1923 * @slabp: slab pointer being destroyed 1924 * 1925 * Destroy all the objs in a slab, and release the mem back to the system. 1926 * Before calling the slab must have been unlinked from the cache. The 1927 * cache-lock is not held/needed. 1928 */ 1929static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1930{ 1931 void *addr = slabp->s_mem - slabp->colouroff; 1932 1933 slab_destroy_objs(cachep, slabp); 1934 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1935 struct slab_rcu *slab_rcu; 1936 1937 slab_rcu = (struct slab_rcu *)slabp; 1938 slab_rcu->cachep = cachep; 1939 slab_rcu->addr = addr; 1940 call_rcu(&slab_rcu->head, kmem_rcu_free); 1941 } else { 1942 kmem_freepages(cachep, addr); 1943 if (OFF_SLAB(cachep)) 1944 kmem_cache_free(cachep->slabp_cache, slabp); 1945 } 1946} 1947 1948/* 1949 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1950 * size of kmem_list3. 1951 */ 1952static void __init set_up_list3s(struct kmem_cache *cachep, int index) 1953{ 1954 int node; 1955 1956 for_each_online_node(node) { 1957 cachep->nodelists[node] = &initkmem_list3[index + node]; 1958 cachep->nodelists[node]->next_reap = jiffies + 1959 REAPTIMEOUT_LIST3 + 1960 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1961 } 1962} 1963 1964static void __kmem_cache_destroy(struct kmem_cache *cachep) 1965{ 1966 int i; 1967 struct kmem_list3 *l3; 1968 1969 for_each_online_cpu(i) 1970 kfree(cachep->array[i]); 1971 1972 /* NUMA: free the list3 structures */ 1973 for_each_online_node(i) { 1974 l3 = cachep->nodelists[i]; 1975 if (l3) { 1976 kfree(l3->shared); 1977 free_alien_cache(l3->alien); 1978 kfree(l3); 1979 } 1980 } 1981 kmem_cache_free(&cache_cache, cachep); 1982} 1983 1984 1985/** 1986 * calculate_slab_order - calculate size (page order) of slabs 1987 * @cachep: pointer to the cache that is being created 1988 * @size: size of objects to be created in this cache. 1989 * @align: required alignment for the objects. 1990 * @flags: slab allocation flags 1991 * 1992 * Also calculates the number of objects per slab. 1993 * 1994 * This could be made much more intelligent. For now, try to avoid using 1995 * high order pages for slabs. When the gfp() functions are more friendly 1996 * towards high-order requests, this should be changed. 1997 */ 1998static size_t calculate_slab_order(struct kmem_cache *cachep, 1999 size_t size, size_t align, unsigned long flags) 2000{ 2001 unsigned long offslab_limit; 2002 size_t left_over = 0; 2003 int gfporder; 2004 2005 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { 2006 unsigned int num; 2007 size_t remainder; 2008 2009 cache_estimate(gfporder, size, align, flags, &remainder, &num); 2010 if (!num) 2011 continue; 2012 2013 if (flags & CFLGS_OFF_SLAB) { 2014 /* 2015 * Max number of objs-per-slab for caches which 2016 * use off-slab slabs. Needed to avoid a possible 2017 * looping condition in cache_grow(). 2018 */ 2019 offslab_limit = size - sizeof(struct slab); 2020 offslab_limit /= sizeof(kmem_bufctl_t); 2021 2022 if (num > offslab_limit) 2023 break; 2024 } 2025 2026 /* Found something acceptable - save it away */ 2027 cachep->num = num; 2028 cachep->gfporder = gfporder; 2029 left_over = remainder; 2030 2031 /* 2032 * A VFS-reclaimable slab tends to have most allocations 2033 * as GFP_NOFS and we really don't want to have to be allocating 2034 * higher-order pages when we are unable to shrink dcache. 2035 */ 2036 if (flags & SLAB_RECLAIM_ACCOUNT) 2037 break; 2038 2039 /* 2040 * Large number of objects is good, but very large slabs are 2041 * currently bad for the gfp()s. 2042 */ 2043 if (gfporder >= slab_break_gfp_order) 2044 break; 2045 2046 /* 2047 * Acceptable internal fragmentation? 2048 */ 2049 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 2050 break; 2051 } 2052 return left_over; 2053} 2054 2055static int setup_cpu_cache(struct kmem_cache *cachep) 2056{ 2057 if (g_cpucache_up == FULL) 2058 return enable_cpucache(cachep); 2059 2060 if (g_cpucache_up == NONE) { 2061 /* 2062 * Note: the first kmem_cache_create must create the cache 2063 * that's used by kmalloc(24), otherwise the creation of 2064 * further caches will BUG(). 2065 */ 2066 cachep->array[smp_processor_id()] = &initarray_generic.cache; 2067 2068 /* 2069 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 2070 * the first cache, then we need to set up all its list3s, 2071 * otherwise the creation of further caches will BUG(). 2072 */ 2073 set_up_list3s(cachep, SIZE_AC); 2074 if (INDEX_AC == INDEX_L3) 2075 g_cpucache_up = PARTIAL_L3; 2076 else 2077 g_cpucache_up = PARTIAL_AC; 2078 } else { 2079 cachep->array[smp_processor_id()] = 2080 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 2081 2082 if (g_cpucache_up == PARTIAL_AC) { 2083 set_up_list3s(cachep, SIZE_L3); 2084 g_cpucache_up = PARTIAL_L3; 2085 } else { 2086 int node; 2087 for_each_online_node(node) { 2088 cachep->nodelists[node] = 2089 kmalloc_node(sizeof(struct kmem_list3), 2090 GFP_KERNEL, node); 2091 BUG_ON(!cachep->nodelists[node]); 2092 kmem_list3_init(cachep->nodelists[node]); 2093 } 2094 } 2095 } 2096 cachep->nodelists[numa_node_id()]->next_reap = 2097 jiffies + REAPTIMEOUT_LIST3 + 2098 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 2099 2100 cpu_cache_get(cachep)->avail = 0; 2101 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 2102 cpu_cache_get(cachep)->batchcount = 1; 2103 cpu_cache_get(cachep)->touched = 0; 2104 cachep->batchcount = 1; 2105 cachep->limit = BOOT_CPUCACHE_ENTRIES; 2106 return 0; 2107} 2108 2109/** 2110 * kmem_cache_create - Create a cache. 2111 * @name: A string which is used in /proc/slabinfo to identify this cache. 2112 * @size: The size of objects to be created in this cache. 2113 * @align: The required alignment for the objects. 2114 * @flags: SLAB flags 2115 * @ctor: A constructor for the objects. 2116 * @dtor: A destructor for the objects. 2117 * 2118 * Returns a ptr to the cache on success, NULL on failure. 2119 * Cannot be called within a int, but can be interrupted. 2120 * The @ctor is run when new pages are allocated by the cache 2121 * and the @dtor is run before the pages are handed back. 2122 * 2123 * @name must be valid until the cache is destroyed. This implies that 2124 * the module calling this has to destroy the cache before getting unloaded. 2125 * 2126 * The flags are 2127 * 2128 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 2129 * to catch references to uninitialised memory. 2130 * 2131 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 2132 * for buffer overruns. 2133 * 2134 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 2135 * cacheline. This can be beneficial if you're counting cycles as closely 2136 * as davem. 2137 */ 2138struct kmem_cache * 2139kmem_cache_create (const char *name, size_t size, size_t align, 2140 unsigned long flags, 2141 void (*ctor)(void*, struct kmem_cache *, unsigned long), 2142 void (*dtor)(void*, struct kmem_cache *, unsigned long)) 2143{ 2144 size_t left_over, slab_size, ralign; 2145 struct kmem_cache *cachep = NULL, *pc; 2146 2147 /* 2148 * Sanity checks... these are all serious usage bugs. 2149 */ 2150 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 2151 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { 2152 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, 2153 name); 2154 BUG(); 2155 } 2156 2157 /* 2158 * We use cache_chain_mutex to ensure a consistent view of 2159 * cpu_online_map as well. Please see cpuup_callback 2160 */ 2161 mutex_lock(&cache_chain_mutex); 2162 2163 list_for_each_entry(pc, &cache_chain, next) { 2164 char tmp; 2165 int res; 2166 2167 /* 2168 * This happens when the module gets unloaded and doesn't 2169 * destroy its slab cache and no-one else reuses the vmalloc 2170 * area of the module. Print a warning. 2171 */ 2172 res = probe_kernel_address(pc->name, tmp); 2173 if (res) { 2174 printk(KERN_ERR 2175 "SLAB: cache with size %d has lost its name\n", 2176 pc->buffer_size); 2177 continue; 2178 } 2179 2180 if (!strcmp(pc->name, name)) { 2181 printk(KERN_ERR 2182 "kmem_cache_create: duplicate cache %s\n", name); 2183 dump_stack(); 2184 goto oops; 2185 } 2186 } 2187 2188#if DEBUG 2189 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 2190#if FORCED_DEBUG 2191 /* 2192 * Enable redzoning and last user accounting, except for caches with 2193 * large objects, if the increased size would increase the object size 2194 * above the next power of two: caches with object sizes just above a 2195 * power of two have a significant amount of internal fragmentation. 2196 */ 2197 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) 2198 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2199 if (!(flags & SLAB_DESTROY_BY_RCU)) 2200 flags |= SLAB_POISON; 2201#endif 2202 if (flags & SLAB_DESTROY_BY_RCU) 2203 BUG_ON(flags & SLAB_POISON); 2204#endif 2205 if (flags & SLAB_DESTROY_BY_RCU) 2206 BUG_ON(dtor); 2207 2208 /* 2209 * Always checks flags, a caller might be expecting debug support which 2210 * isn't available. 2211 */ 2212 BUG_ON(flags & ~CREATE_MASK); 2213 2214 /* 2215 * Check that size is in terms of words. This is needed to avoid 2216 * unaligned accesses for some archs when redzoning is used, and makes 2217 * sure any on-slab bufctl's are also correctly aligned. 2218 */ 2219 if (size & (BYTES_PER_WORD - 1)) { 2220 size += (BYTES_PER_WORD - 1); 2221 size &= ~(BYTES_PER_WORD - 1); 2222 } 2223 2224 /* calculate the final buffer alignment: */ 2225 2226 /* 1) arch recommendation: can be overridden for debug */ 2227 if (flags & SLAB_HWCACHE_ALIGN) { 2228 /* 2229 * Default alignment: as specified by the arch code. Except if 2230 * an object is really small, then squeeze multiple objects into 2231 * one cacheline. 2232 */ 2233 ralign = cache_line_size(); 2234 while (size <= ralign / 2) 2235 ralign /= 2; 2236 } else { 2237 ralign = BYTES_PER_WORD; 2238 } 2239 2240 /* 2241 * Redzoning and user store require word alignment. Note this will be 2242 * overridden by architecture or caller mandated alignment if either 2243 * is greater than BYTES_PER_WORD. 2244 */ 2245 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) 2246 ralign = __alignof__(unsigned long long); 2247 2248 /* 2) arch mandated alignment */ 2249 if (ralign < ARCH_SLAB_MINALIGN) { 2250 ralign = ARCH_SLAB_MINALIGN; 2251 } 2252 /* 3) caller mandated alignment */ 2253 if (ralign < align) { 2254 ralign = align; 2255 } 2256 /* disable debug if necessary */ 2257 if (ralign > __alignof__(unsigned long long)) 2258 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2259 /* 2260 * 4) Store it. 2261 */ 2262 align = ralign; 2263 2264 /* Get cache's description obj. */ 2265 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); 2266 if (!cachep) 2267 goto oops; 2268 2269#if DEBUG 2270 cachep->obj_size = size; 2271 2272 /* 2273 * Both debugging options require word-alignment which is calculated 2274 * into align above. 2275 */ 2276 if (flags & SLAB_RED_ZONE) { 2277 /* add space for red zone words */ 2278 cachep->obj_offset += sizeof(unsigned long long); 2279 size += 2 * sizeof(unsigned long long); 2280 } 2281 if (flags & SLAB_STORE_USER) { 2282 /* user store requires one word storage behind the end of 2283 * the real object. 2284 */ 2285 size += BYTES_PER_WORD; 2286 } 2287#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2288 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2289 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { 2290 cachep->obj_offset += PAGE_SIZE - size; 2291 size = PAGE_SIZE; 2292 } 2293#endif 2294#endif 2295 2296 /* 2297 * Determine if the slab management is 'on' or 'off' slab. 2298 * (bootstrapping cannot cope with offslab caches so don't do 2299 * it too early on.) 2300 */ 2301 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) 2302 /* 2303 * Size is large, assume best to place the slab management obj 2304 * off-slab (should allow better packing of objs). 2305 */ 2306 flags |= CFLGS_OFF_SLAB; 2307 2308 size = ALIGN(size, align); 2309 2310 left_over = calculate_slab_order(cachep, size, align, flags); 2311 2312 if (!cachep->num) { 2313 printk(KERN_ERR 2314 "kmem_cache_create: couldn't create cache %s.\n", name); 2315 kmem_cache_free(&cache_cache, cachep); 2316 cachep = NULL; 2317 goto oops; 2318 } 2319 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2320 + sizeof(struct slab), align); 2321 2322 /* 2323 * If the slab has been placed off-slab, and we have enough space then 2324 * move it on-slab. This is at the expense of any extra colouring. 2325 */ 2326 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2327 flags &= ~CFLGS_OFF_SLAB; 2328 left_over -= slab_size; 2329 } 2330 2331 if (flags & CFLGS_OFF_SLAB) { 2332 /* really off slab. No need for manual alignment */ 2333 slab_size = 2334 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2335 } 2336 2337 cachep->colour_off = cache_line_size(); 2338 /* Offset must be a multiple of the alignment. */ 2339 if (cachep->colour_off < align) 2340 cachep->colour_off = align; 2341 cachep->colour = left_over / cachep->colour_off; 2342 cachep->slab_size = slab_size; 2343 cachep->flags = flags; 2344 cachep->gfpflags = 0; 2345 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) 2346 cachep->gfpflags |= GFP_DMA; 2347 cachep->buffer_size = size; 2348 cachep->reciprocal_buffer_size = reciprocal_value(size); 2349 2350 if (flags & CFLGS_OFF_SLAB) { 2351 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2352 /* 2353 * This is a possibility for one of the malloc_sizes caches. 2354 * But since we go off slab only for object size greater than 2355 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, 2356 * this should not happen at all. 2357 * But leave a BUG_ON for some lucky dude. 2358 */ 2359 BUG_ON(!cachep->slabp_cache); 2360 } 2361 cachep->ctor = ctor; 2362 cachep->dtor = dtor; 2363 cachep->name = name; 2364 2365 if (setup_cpu_cache(cachep)) { 2366 __kmem_cache_destroy(cachep); 2367 cachep = NULL; 2368 goto oops; 2369 } 2370 2371 /* cache setup completed, link it into the list */ 2372 list_add(&cachep->next, &cache_chain); 2373oops: 2374 if (!cachep && (flags & SLAB_PANIC)) 2375 panic("kmem_cache_create(): failed to create slab `%s'\n", 2376 name); 2377 mutex_unlock(&cache_chain_mutex); 2378 return cachep; 2379} 2380EXPORT_SYMBOL(kmem_cache_create); 2381 2382#if DEBUG 2383static void check_irq_off(void) 2384{ 2385 BUG_ON(!irqs_disabled()); 2386} 2387 2388static void check_irq_on(void) 2389{ 2390 BUG_ON(irqs_disabled()); 2391} 2392 2393static void check_spinlock_acquired(struct kmem_cache *cachep) 2394{ 2395#ifdef CONFIG_SMP 2396 check_irq_off(); 2397 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); 2398#endif 2399} 2400 2401static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2402{ 2403#ifdef CONFIG_SMP 2404 check_irq_off(); 2405 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2406#endif 2407} 2408 2409#else 2410#define check_irq_off() do { } while(0) 2411#define check_irq_on() do { } while(0) 2412#define check_spinlock_acquired(x) do { } while(0) 2413#define check_spinlock_acquired_node(x, y) do { } while(0) 2414#endif 2415 2416static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2417 struct array_cache *ac, 2418 int force, int node); 2419 2420static void do_drain(void *arg) 2421{ 2422 struct kmem_cache *cachep = arg; 2423 struct array_cache *ac; 2424 int node = numa_node_id(); 2425 2426 check_irq_off(); 2427 ac = cpu_cache_get(cachep); 2428 spin_lock(&cachep->nodelists[node]->list_lock); 2429 free_block(cachep, ac->entry, ac->avail, node); 2430 spin_unlock(&cachep->nodelists[node]->list_lock); 2431 ac->avail = 0; 2432} 2433 2434static void drain_cpu_caches(struct kmem_cache *cachep) 2435{ 2436 struct kmem_list3 *l3; 2437 int node; 2438 2439 on_each_cpu(do_drain, cachep, 1, 1); 2440 check_irq_on(); 2441 for_each_online_node(node) { 2442 l3 = cachep->nodelists[node]; 2443 if (l3 && l3->alien) 2444 drain_alien_cache(cachep, l3->alien); 2445 } 2446 2447 for_each_online_node(node) { 2448 l3 = cachep->nodelists[node]; 2449 if (l3) 2450 drain_array(cachep, l3, l3->shared, 1, node); 2451 } 2452} 2453 2454/* 2455 * Remove slabs from the list of free slabs. 2456 * Specify the number of slabs to drain in tofree. 2457 * 2458 * Returns the actual number of slabs released. 2459 */ 2460static int drain_freelist(struct kmem_cache *cache, 2461 struct kmem_list3 *l3, int tofree) 2462{ 2463 struct list_head *p; 2464 int nr_freed; 2465 struct slab *slabp; 2466 2467 nr_freed = 0; 2468 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { 2469 2470 spin_lock_irq(&l3->list_lock); 2471 p = l3->slabs_free.prev; 2472 if (p == &l3->slabs_free) { 2473 spin_unlock_irq(&l3->list_lock); 2474 goto out; 2475 } 2476 2477 slabp = list_entry(p, struct slab, list); 2478#if DEBUG 2479 BUG_ON(slabp->inuse); 2480#endif 2481 list_del(&slabp->list); 2482 /* 2483 * Safe to drop the lock. The slab is no longer linked 2484 * to the cache. 2485 */ 2486 l3->free_objects -= cache->num; 2487 spin_unlock_irq(&l3->list_lock); 2488 slab_destroy(cache, slabp); 2489 nr_freed++; 2490 } 2491out: 2492 return nr_freed; 2493} 2494 2495/* Called with cache_chain_mutex held to protect against cpu hotplug */ 2496static int __cache_shrink(struct kmem_cache *cachep) 2497{ 2498 int ret = 0, i = 0; 2499 struct kmem_list3 *l3; 2500 2501 drain_cpu_caches(cachep); 2502 2503 check_irq_on(); 2504 for_each_online_node(i) { 2505 l3 = cachep->nodelists[i]; 2506 if (!l3) 2507 continue; 2508 2509 drain_freelist(cachep, l3, l3->free_objects); 2510 2511 ret += !list_empty(&l3->slabs_full) || 2512 !list_empty(&l3->slabs_partial); 2513 } 2514 return (ret ? 1 : 0); 2515} 2516 2517/** 2518 * kmem_cache_shrink - Shrink a cache. 2519 * @cachep: The cache to shrink. 2520 * 2521 * Releases as many slabs as possible for a cache. 2522 * To help debugging, a zero exit status indicates all slabs were released. 2523 */ 2524int kmem_cache_shrink(struct kmem_cache *cachep) 2525{ 2526 int ret; 2527 BUG_ON(!cachep || in_interrupt()); 2528 2529 mutex_lock(&cache_chain_mutex); 2530 ret = __cache_shrink(cachep); 2531 mutex_unlock(&cache_chain_mutex); 2532 return ret; 2533} 2534EXPORT_SYMBOL(kmem_cache_shrink); 2535 2536/** 2537 * kmem_cache_destroy - delete a cache 2538 * @cachep: the cache to destroy 2539 * 2540 * Remove a &struct kmem_cache object from the slab cache. 2541 * 2542 * It is expected this function will be called by a module when it is 2543 * unloaded. This will remove the cache completely, and avoid a duplicate 2544 * cache being allocated each time a module is loaded and unloaded, if the 2545 * module doesn't have persistent in-kernel storage across loads and unloads. 2546 * 2547 * The cache must be empty before calling this function. 2548 * 2549 * The caller must guarantee that noone will allocate memory from the cache 2550 * during the kmem_cache_destroy(). 2551 */ 2552void kmem_cache_destroy(struct kmem_cache *cachep) 2553{ 2554 BUG_ON(!cachep || in_interrupt()); 2555 2556 /* Find the cache in the chain of caches. */ 2557 mutex_lock(&cache_chain_mutex); 2558 /* 2559 * the chain is never empty, cache_cache is never destroyed 2560 */ 2561 list_del(&cachep->next); 2562 if (__cache_shrink(cachep)) { 2563 slab_error(cachep, "Can't free all objects"); 2564 list_add(&cachep->next, &cache_chain); 2565 mutex_unlock(&cache_chain_mutex); 2566 return; 2567 } 2568 2569 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2570 synchronize_rcu(); 2571 2572 __kmem_cache_destroy(cachep); 2573 mutex_unlock(&cache_chain_mutex); 2574} 2575EXPORT_SYMBOL(kmem_cache_destroy); 2576 2577/* 2578 * Get the memory for a slab management obj. 2579 * For a slab cache when the slab descriptor is off-slab, slab descriptors 2580 * always come from malloc_sizes caches. The slab descriptor cannot 2581 * come from the same cache which is getting created because, 2582 * when we are searching for an appropriate cache for these 2583 * descriptors in kmem_cache_create, we search through the malloc_sizes array. 2584 * If we are creating a malloc_sizes cache here it would not be visible to 2585 * kmem_find_general_cachep till the initialization is complete. 2586 * Hence we cannot have slabp_cache same as the original cache. 2587 */ 2588static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2589 int colour_off, gfp_t local_flags, 2590 int nodeid) 2591{ 2592 struct slab *slabp; 2593 2594 if (OFF_SLAB(cachep)) { 2595 /* Slab management obj is off-slab. */ 2596 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2597 local_flags & ~GFP_THISNODE, nodeid); 2598 if (!slabp) 2599 return NULL; 2600 } else { 2601 slabp = objp + colour_off; 2602 colour_off += cachep->slab_size; 2603 } 2604 slabp->inuse = 0; 2605 slabp->colouroff = colour_off; 2606 slabp->s_mem = objp + colour_off; 2607 slabp->nodeid = nodeid; 2608 return slabp; 2609} 2610 2611static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2612{ 2613 return (kmem_bufctl_t *) (slabp + 1); 2614} 2615 2616static void cache_init_objs(struct kmem_cache *cachep, 2617 struct slab *slabp, unsigned long ctor_flags) 2618{ 2619 int i; 2620 2621 for (i = 0; i < cachep->num; i++) { 2622 void *objp = index_to_obj(cachep, slabp, i); 2623#if DEBUG 2624 /* need to poison the objs? */ 2625 if (cachep->flags & SLAB_POISON) 2626 poison_obj(cachep, objp, POISON_FREE); 2627 if (cachep->flags & SLAB_STORE_USER) 2628 *dbg_userword(cachep, objp) = NULL; 2629 2630 if (cachep->flags & SLAB_RED_ZONE) { 2631 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2632 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2633 } 2634 /* 2635 * Constructors are not allowed to allocate memory from the same 2636 * cache which they are a constructor for. Otherwise, deadlock. 2637 * They must also be threaded. 2638 */ 2639 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2640 cachep->ctor(objp + obj_offset(cachep), cachep, 2641 ctor_flags); 2642 2643 if (cachep->flags & SLAB_RED_ZONE) { 2644 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2645 slab_error(cachep, "constructor overwrote the" 2646 " end of an object"); 2647 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2648 slab_error(cachep, "constructor overwrote the" 2649 " start of an object"); 2650 } 2651 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2652 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2653 kernel_map_pages(virt_to_page(objp), 2654 cachep->buffer_size / PAGE_SIZE, 0); 2655#else 2656 if (cachep->ctor) 2657 cachep->ctor(objp, cachep, ctor_flags); 2658#endif 2659 slab_bufctl(slabp)[i] = i + 1; 2660 } 2661 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2662 slabp->free = 0; 2663} 2664 2665static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2666{ 2667 if (CONFIG_ZONE_DMA_FLAG) { 2668 if (flags & GFP_DMA) 2669 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2670 else 2671 BUG_ON(cachep->gfpflags & GFP_DMA); 2672 } 2673} 2674 2675static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2676 int nodeid) 2677{ 2678 void *objp = index_to_obj(cachep, slabp, slabp->free); 2679 kmem_bufctl_t next; 2680 2681 slabp->inuse++; 2682 next = slab_bufctl(slabp)[slabp->free]; 2683#if DEBUG 2684 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2685 WARN_ON(slabp->nodeid != nodeid); 2686#endif 2687 slabp->free = next; 2688 2689 return objp; 2690} 2691 2692static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2693 void *objp, int nodeid) 2694{ 2695 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2696 2697#if DEBUG 2698 /* Verify that the slab belongs to the intended node */ 2699 WARN_ON(slabp->nodeid != nodeid); 2700 2701 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2702 printk(KERN_ERR "slab: double free detected in cache " 2703 "'%s', objp %p\n", cachep->name, objp); 2704 BUG(); 2705 } 2706#endif 2707 slab_bufctl(slabp)[objnr] = slabp->free; 2708 slabp->free = objnr; 2709 slabp->inuse--; 2710} 2711 2712/* 2713 * Map pages beginning at addr to the given cache and slab. This is required 2714 * for the slab allocator to be able to lookup the cache and slab of a 2715 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. 2716 */ 2717static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, 2718 void *addr) 2719{ 2720 int nr_pages; 2721 struct page *page; 2722 2723 page = virt_to_page(addr); 2724 2725 nr_pages = 1; 2726 if (likely(!PageCompound(page))) 2727 nr_pages <<= cache->gfporder; 2728 2729 do { 2730 page_set_cache(page, cache); 2731 page_set_slab(page, slab); 2732 page++; 2733 } while (--nr_pages); 2734} 2735 2736/* 2737 * Grow (by 1) the number of slabs within a cache. This is called by 2738 * kmem_cache_alloc() when there are no active objs left in a cache. 2739 */ 2740static int cache_grow(struct kmem_cache *cachep, 2741 gfp_t flags, int nodeid, void *objp) 2742{ 2743 struct slab *slabp; 2744 size_t offset; 2745 gfp_t local_flags; 2746 unsigned long ctor_flags; 2747 struct kmem_list3 *l3; 2748 2749 /* 2750 * Be lazy and only check for valid flags here, keeping it out of the 2751 * critical path in kmem_cache_alloc(). 2752 */ 2753 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK)); 2754 2755 ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2756 local_flags = (flags & GFP_LEVEL_MASK); 2757 /* Take the l3 list lock to change the colour_next on this node */ 2758 check_irq_off(); 2759 l3 = cachep->nodelists[nodeid]; 2760 spin_lock(&l3->list_lock); 2761 2762 /* Get colour for the slab, and cal the next value. */ 2763 offset = l3->colour_next; 2764 l3->colour_next++; 2765 if (l3->colour_next >= cachep->colour) 2766 l3->colour_next = 0; 2767 spin_unlock(&l3->list_lock); 2768 2769 offset *= cachep->colour_off; 2770 2771 if (local_flags & __GFP_WAIT) 2772 local_irq_enable(); 2773 2774 /* 2775 * The test for missing atomic flag is performed here, rather than 2776 * the more obvious place, simply to reduce the critical path length 2777 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2778 * will eventually be caught here (where it matters). 2779 */ 2780 kmem_flagcheck(cachep, flags); 2781 2782 /* 2783 * Get mem for the objs. Attempt to allocate a physical page from 2784 * 'nodeid'. 2785 */ 2786 if (!objp) 2787 objp = kmem_getpages(cachep, flags, nodeid); 2788 if (!objp) 2789 goto failed; 2790 2791 /* Get slab management. */ 2792 slabp = alloc_slabmgmt(cachep, objp, offset, 2793 local_flags & ~GFP_THISNODE, nodeid); 2794 if (!slabp) 2795 goto opps1; 2796 2797 slabp->nodeid = nodeid; 2798 slab_map_pages(cachep, slabp, objp); 2799 2800 cache_init_objs(cachep, slabp, ctor_flags); 2801 2802 if (local_flags & __GFP_WAIT) 2803 local_irq_disable(); 2804 check_irq_off(); 2805 spin_lock(&l3->list_lock); 2806 2807 /* Make slab active. */ 2808 list_add_tail(&slabp->list, &(l3->slabs_free)); 2809 STATS_INC_GROWN(cachep); 2810 l3->free_objects += cachep->num; 2811 spin_unlock(&l3->list_lock); 2812 return 1; 2813opps1: 2814 kmem_freepages(cachep, objp); 2815failed: 2816 if (local_flags & __GFP_WAIT) 2817 local_irq_disable(); 2818 return 0; 2819} 2820 2821#if DEBUG 2822 2823/* 2824 * Perform extra freeing checks: 2825 * - detect bad pointers. 2826 * - POISON/RED_ZONE checking 2827 * - destructor calls, for caches with POISON+dtor 2828 */ 2829static void kfree_debugcheck(const void *objp) 2830{ 2831 if (!virt_addr_valid(objp)) { 2832 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2833 (unsigned long)objp); 2834 BUG(); 2835 } 2836} 2837 2838static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2839{ 2840 unsigned long long redzone1, redzone2; 2841 2842 redzone1 = *dbg_redzone1(cache, obj); 2843 redzone2 = *dbg_redzone2(cache, obj); 2844 2845 /* 2846 * Redzone is ok. 2847 */ 2848 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2849 return; 2850 2851 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2852 slab_error(cache, "double free detected"); 2853 else 2854 slab_error(cache, "memory outside object was overwritten"); 2855 2856 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", 2857 obj, redzone1, redzone2); 2858} 2859 2860static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2861 void *caller) 2862{ 2863 struct page *page; 2864 unsigned int objnr; 2865 struct slab *slabp; 2866 2867 objp -= obj_offset(cachep); 2868 kfree_debugcheck(objp); 2869 page = virt_to_head_page(objp); 2870 2871 slabp = page_get_slab(page); 2872 2873 if (cachep->flags & SLAB_RED_ZONE) { 2874 verify_redzone_free(cachep, objp); 2875 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2876 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2877 } 2878 if (cachep->flags & SLAB_STORE_USER) 2879 *dbg_userword(cachep, objp) = caller; 2880 2881 objnr = obj_to_index(cachep, slabp, objp); 2882 2883 BUG_ON(objnr >= cachep->num); 2884 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2885 2886 if (cachep->flags & SLAB_POISON && cachep->dtor) { 2887 /* we want to cache poison the object, 2888 * call the destruction callback 2889 */ 2890 cachep->dtor(objp + obj_offset(cachep), cachep, 0); 2891 } 2892#ifdef CONFIG_DEBUG_SLAB_LEAK 2893 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2894#endif 2895 if (cachep->flags & SLAB_POISON) { 2896#ifdef CONFIG_DEBUG_PAGEALLOC 2897 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2898 store_stackinfo(cachep, objp, (unsigned long)caller); 2899 kernel_map_pages(virt_to_page(objp), 2900 cachep->buffer_size / PAGE_SIZE, 0); 2901 } else { 2902 poison_obj(cachep, objp, POISON_FREE); 2903 } 2904#else 2905 poison_obj(cachep, objp, POISON_FREE); 2906#endif 2907 } 2908 return objp; 2909} 2910 2911static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2912{ 2913 kmem_bufctl_t i; 2914 int entries = 0; 2915 2916 /* Check slab's freelist to see if this obj is there. */ 2917 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2918 entries++; 2919 if (entries > cachep->num || i >= cachep->num) 2920 goto bad; 2921 } 2922 if (entries != cachep->num - slabp->inuse) { 2923bad: 2924 printk(KERN_ERR "slab: Internal list corruption detected in " 2925 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2926 cachep->name, cachep->num, slabp, slabp->inuse); 2927 for (i = 0; 2928 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2929 i++) { 2930 if (i % 16 == 0) 2931 printk("\n%03x:", i); 2932 printk(" %02x", ((unsigned char *)slabp)[i]); 2933 } 2934 printk("\n"); 2935 BUG(); 2936 } 2937} 2938#else 2939#define kfree_debugcheck(x) do { } while(0) 2940#define cache_free_debugcheck(x,objp,z) (objp) 2941#define check_slabp(x,y) do { } while(0) 2942#endif 2943 2944static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2945{ 2946 int batchcount; 2947 struct kmem_list3 *l3; 2948 struct array_cache *ac; 2949 int node; 2950 2951 node = numa_node_id(); 2952 2953 check_irq_off(); 2954 ac = cpu_cache_get(cachep); 2955retry: 2956 batchcount = ac->batchcount; 2957 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2958 /* 2959 * If there was little recent activity on this cache, then 2960 * perform only a partial refill. Otherwise we could generate 2961 * refill bouncing. 2962 */ 2963 batchcount = BATCHREFILL_LIMIT; 2964 } 2965 l3 = cachep->nodelists[node]; 2966 2967 BUG_ON(ac->avail > 0 || !l3); 2968 spin_lock(&l3->list_lock); 2969 2970 /* See if we can refill from the shared array */ 2971 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) 2972 goto alloc_done; 2973 2974 while (batchcount > 0) { 2975 struct list_head *entry; 2976 struct slab *slabp; 2977 /* Get slab alloc is to come from. */ 2978 entry = l3->slabs_partial.next; 2979 if (entry == &l3->slabs_partial) { 2980 l3->free_touched = 1; 2981 entry = l3->slabs_free.next; 2982 if (entry == &l3->slabs_free) 2983 goto must_grow; 2984 } 2985 2986 slabp = list_entry(entry, struct slab, list); 2987 check_slabp(cachep, slabp); 2988 check_spinlock_acquired(cachep); 2989 2990 /* 2991 * The slab was either on partial or free list so 2992 * there must be at least one object available for 2993 * allocation. 2994 */ 2995 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num); 2996 2997 while (slabp->inuse < cachep->num && batchcount--) { 2998 STATS_INC_ALLOCED(cachep); 2999 STATS_INC_ACTIVE(cachep); 3000 STATS_SET_HIGH(cachep); 3001 3002 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 3003 node); 3004 } 3005 check_slabp(cachep, slabp); 3006 3007 /* move slabp to correct slabp list: */ 3008 list_del(&slabp->list); 3009 if (slabp->free == BUFCTL_END) 3010 list_add(&slabp->list, &l3->slabs_full); 3011 else 3012 list_add(&slabp->list, &l3->slabs_partial); 3013 } 3014 3015must_grow: 3016 l3->free_objects -= ac->avail; 3017alloc_done: 3018 spin_unlock(&l3->list_lock); 3019 3020 if (unlikely(!ac->avail)) { 3021 int x; 3022 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); 3023 3024 /* cache_grow can reenable interrupts, then ac could change. */ 3025 ac = cpu_cache_get(cachep); 3026 if (!x && ac->avail == 0) /* no objects in sight? abort */ 3027 return NULL; 3028 3029 if (!ac->avail) /* objects refilled by interrupt? */ 3030 goto retry; 3031 } 3032 ac->touched = 1; 3033 return ac->entry[--ac->avail]; 3034} 3035 3036static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3037 gfp_t flags) 3038{ 3039 might_sleep_if(flags & __GFP_WAIT); 3040#if DEBUG 3041 kmem_flagcheck(cachep, flags); 3042#endif 3043} 3044 3045#if DEBUG 3046static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3047 gfp_t flags, void *objp, void *caller) 3048{ 3049 if (!objp) 3050 return objp; 3051 if (cachep->flags & SLAB_POISON) { 3052#ifdef CONFIG_DEBUG_PAGEALLOC 3053 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 3054 kernel_map_pages(virt_to_page(objp), 3055 cachep->buffer_size / PAGE_SIZE, 1); 3056 else 3057 check_poison_obj(cachep, objp); 3058#else 3059 check_poison_obj(cachep, objp); 3060#endif 3061 poison_obj(cachep, objp, POISON_INUSE); 3062 } 3063 if (cachep->flags & SLAB_STORE_USER) 3064 *dbg_userword(cachep, objp) = caller; 3065 3066 if (cachep->flags & SLAB_RED_ZONE) { 3067 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3068 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3069 slab_error(cachep, "double free, or memory outside" 3070 " object was overwritten"); 3071 printk(KERN_ERR 3072 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", 3073 objp, *dbg_redzone1(cachep, objp), 3074 *dbg_redzone2(cachep, objp)); 3075 } 3076 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3077 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3078 } 3079#ifdef CONFIG_DEBUG_SLAB_LEAK 3080 { 3081 struct slab *slabp; 3082 unsigned objnr; 3083 3084 slabp = page_get_slab(virt_to_head_page(objp)); 3085 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 3086 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 3087 } 3088#endif 3089 objp += obj_offset(cachep); 3090 if (cachep->ctor && cachep->flags & SLAB_POISON) 3091 cachep->ctor(objp, cachep, SLAB_CTOR_CONSTRUCTOR); 3092#if ARCH_SLAB_MINALIGN 3093 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { 3094 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3095 objp, ARCH_SLAB_MINALIGN); 3096 } 3097#endif 3098 return objp; 3099} 3100#else 3101#define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3102#endif 3103 3104#ifdef CONFIG_FAILSLAB 3105 3106static struct failslab_attr { 3107 3108 struct fault_attr attr; 3109 3110 u32 ignore_gfp_wait; 3111#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3112 struct dentry *ignore_gfp_wait_file; 3113#endif 3114 3115} failslab = { 3116 .attr = FAULT_ATTR_INITIALIZER, 3117 .ignore_gfp_wait = 1, 3118}; 3119 3120static int __init setup_failslab(char *str) 3121{ 3122 return setup_fault_attr(&failslab.attr, str); 3123} 3124__setup("failslab=", setup_failslab); 3125 3126static int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3127{ 3128 if (cachep == &cache_cache) 3129 return 0; 3130 if (flags & __GFP_NOFAIL) 3131 return 0; 3132 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) 3133 return 0; 3134 3135 return should_fail(&failslab.attr, obj_size(cachep)); 3136} 3137 3138#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3139 3140static int __init failslab_debugfs(void) 3141{ 3142 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 3143 struct dentry *dir; 3144 int err; 3145 3146 err = init_fault_attr_dentries(&failslab.attr, "failslab"); 3147 if (err) 3148 return err; 3149 dir = failslab.attr.dentries.dir; 3150 3151 failslab.ignore_gfp_wait_file = 3152 debugfs_create_bool("ignore-gfp-wait", mode, dir, 3153 &failslab.ignore_gfp_wait); 3154 3155 if (!failslab.ignore_gfp_wait_file) { 3156 err = -ENOMEM; 3157 debugfs_remove(failslab.ignore_gfp_wait_file); 3158 cleanup_fault_attr_dentries(&failslab.attr); 3159 } 3160 3161 return err; 3162} 3163 3164late_initcall(failslab_debugfs); 3165 3166#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 3167 3168#else /* CONFIG_FAILSLAB */ 3169 3170static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3171{ 3172 return 0; 3173} 3174 3175#endif /* CONFIG_FAILSLAB */ 3176 3177static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3178{ 3179 void *objp; 3180 struct array_cache *ac; 3181 3182 check_irq_off(); 3183 3184 ac = cpu_cache_get(cachep); 3185 if (likely(ac->avail)) { 3186 STATS_INC_ALLOCHIT(cachep); 3187 ac->touched = 1; 3188 objp = ac->entry[--ac->avail]; 3189 } else { 3190 STATS_INC_ALLOCMISS(cachep); 3191 objp = cache_alloc_refill(cachep, flags); 3192 } 3193 return objp; 3194} 3195 3196#ifdef CONFIG_NUMA 3197/* 3198 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 3199 * 3200 * If we are in_interrupt, then process context, including cpusets and 3201 * mempolicy, may not apply and should not be used for allocation policy. 3202 */ 3203static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3204{ 3205 int nid_alloc, nid_here; 3206 3207 if (in_interrupt() || (flags & __GFP_THISNODE)) 3208 return NULL; 3209 nid_alloc = nid_here = numa_node_id(); 3210 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3211 nid_alloc = cpuset_mem_spread_node(); 3212 else if (current->mempolicy) 3213 nid_alloc = slab_node(current->mempolicy); 3214 if (nid_alloc != nid_here) 3215 return ____cache_alloc_node(cachep, flags, nid_alloc); 3216 return NULL; 3217} 3218 3219/* 3220 * Fallback function if there was no memory available and no objects on a 3221 * certain node and fall back is permitted. First we scan all the 3222 * available nodelists for available objects. If that fails then we 3223 * perform an allocation without specifying a node. This allows the page 3224 * allocator to do its reclaim / fallback magic. We then insert the 3225 * slab into the proper nodelist and then allocate from it. 3226 */ 3227static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3228{ 3229 struct zonelist *zonelist; 3230 gfp_t local_flags; 3231 struct zone **z; 3232 void *obj = NULL; 3233 int nid; 3234 3235 if (flags & __GFP_THISNODE) 3236 return NULL; 3237 3238 zonelist = &NODE_DATA(slab_node(current->mempolicy)) 3239 ->node_zonelists[gfp_zone(flags)]; 3240 local_flags = (flags & GFP_LEVEL_MASK); 3241 3242retry: 3243 /* 3244 * Look through allowed nodes for objects available 3245 * from existing per node queues. 3246 */ 3247 for (z = zonelist->zones; *z && !obj; z++) { 3248 nid = zone_to_nid(*z); 3249 3250 if (cpuset_zone_allowed_hardwall(*z, flags) && 3251 cache->nodelists[nid] && 3252 cache->nodelists[nid]->free_objects) 3253 obj = ____cache_alloc_node(cache, 3254 flags | GFP_THISNODE, nid); 3255 } 3256 3257 if (!obj) { 3258 /* 3259 * This allocation will be performed within the constraints 3260 * of the current cpuset / memory policy requirements. 3261 * We may trigger various forms of reclaim on the allowed 3262 * set and go into memory reserves if necessary. 3263 */ 3264 if (local_flags & __GFP_WAIT) 3265 local_irq_enable(); 3266 kmem_flagcheck(cache, flags); 3267 obj = kmem_getpages(cache, flags, -1); 3268 if (local_flags & __GFP_WAIT) 3269 local_irq_disable(); 3270 if (obj) { 3271 /* 3272 * Insert into the appropriate per node queues 3273 */ 3274 nid = page_to_nid(virt_to_page(obj)); 3275 if (cache_grow(cache, flags, nid, obj)) { 3276 obj = ____cache_alloc_node(cache, 3277 flags | GFP_THISNODE, nid); 3278 if (!obj) 3279 /* 3280 * Another processor may allocate the 3281 * objects in the slab since we are 3282 * not holding any locks. 3283 */ 3284 goto retry; 3285 } else { 3286 /* cache_grow already freed obj */ 3287 obj = NULL; 3288 } 3289 } 3290 } 3291 return obj; 3292} 3293 3294/* 3295 * A interface to enable slab creation on nodeid 3296 */ 3297static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3298 int nodeid) 3299{ 3300 struct list_head *entry; 3301 struct slab *slabp; 3302 struct kmem_list3 *l3; 3303 void *obj; 3304 int x; 3305 3306 l3 = cachep->nodelists[nodeid]; 3307 BUG_ON(!l3); 3308 3309retry: 3310 check_irq_off(); 3311 spin_lock(&l3->list_lock); 3312 entry = l3->slabs_partial.next; 3313 if (entry == &l3->slabs_partial) { 3314 l3->free_touched = 1; 3315 entry = l3->slabs_free.next; 3316 if (entry == &l3->slabs_free) 3317 goto must_grow; 3318 } 3319 3320 slabp = list_entry(entry, struct slab, list); 3321 check_spinlock_acquired_node(cachep, nodeid); 3322 check_slabp(cachep, slabp); 3323 3324 STATS_INC_NODEALLOCS(cachep); 3325 STATS_INC_ACTIVE(cachep); 3326 STATS_SET_HIGH(cachep); 3327 3328 BUG_ON(slabp->inuse == cachep->num); 3329 3330 obj = slab_get_obj(cachep, slabp, nodeid); 3331 check_slabp(cachep, slabp); 3332 l3->free_objects--; 3333 /* move slabp to correct slabp list: */ 3334 list_del(&slabp->list); 3335 3336 if (slabp->free == BUFCTL_END) 3337 list_add(&slabp->list, &l3->slabs_full); 3338 else 3339 list_add(&slabp->list, &l3->slabs_partial); 3340 3341 spin_unlock(&l3->list_lock); 3342 goto done; 3343 3344must_grow: 3345 spin_unlock(&l3->list_lock); 3346 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); 3347 if (x) 3348 goto retry; 3349 3350 return fallback_alloc(cachep, flags); 3351 3352done: 3353 return obj; 3354} 3355 3356/** 3357 * kmem_cache_alloc_node - Allocate an object on the specified node 3358 * @cachep: The cache to allocate from. 3359 * @flags: See kmalloc(). 3360 * @nodeid: node number of the target node. 3361 * @caller: return address of caller, used for debug information 3362 * 3363 * Identical to kmem_cache_alloc but it will allocate memory on the given 3364 * node, which can improve the performance for cpu bound structures. 3365 * 3366 * Fallback to other node is possible if __GFP_THISNODE is not set. 3367 */ 3368static __always_inline void * 3369__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, 3370 void *caller) 3371{ 3372 unsigned long save_flags; 3373 void *ptr; 3374 3375 if (should_failslab(cachep, flags)) 3376 return NULL; 3377 3378 cache_alloc_debugcheck_before(cachep, flags); 3379 local_irq_save(save_flags); 3380 3381 if (unlikely(nodeid == -1)) 3382 nodeid = numa_node_id(); 3383 3384 if (unlikely(!cachep->nodelists[nodeid])) { 3385 /* Node not bootstrapped yet */ 3386 ptr = fallback_alloc(cachep, flags); 3387 goto out; 3388 } 3389 3390 if (nodeid == numa_node_id()) { 3391 /* 3392 * Use the locally cached objects if possible. 3393 * However ____cache_alloc does not allow fallback 3394 * to other nodes. It may fail while we still have 3395 * objects on other nodes available. 3396 */ 3397 ptr = ____cache_alloc(cachep, flags); 3398 if (ptr) 3399 goto out; 3400 } 3401 /* ___cache_alloc_node can fall back to other nodes */ 3402 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3403 out: 3404 local_irq_restore(save_flags); 3405 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3406 3407 return ptr; 3408} 3409 3410static __always_inline void * 3411__do_cache_alloc(struct kmem_cache *cache, gfp_t flags) 3412{ 3413 void *objp; 3414 3415 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { 3416 objp = alternate_node_alloc(cache, flags); 3417 if (objp) 3418 goto out; 3419 } 3420 objp = ____cache_alloc(cache, flags); 3421 3422 /* 3423 * We may just have run out of memory on the local node. 3424 * ____cache_alloc_node() knows how to locate memory on other nodes 3425 */ 3426 if (!objp) 3427 objp = ____cache_alloc_node(cache, flags, numa_node_id()); 3428 3429 out: 3430 return objp; 3431} 3432#else 3433 3434static __always_inline void * 3435__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3436{ 3437 return ____cache_alloc(cachep, flags); 3438} 3439 3440#endif /* CONFIG_NUMA */ 3441 3442static __always_inline void * 3443__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) 3444{ 3445 unsigned long save_flags; 3446 void *objp; 3447 3448 if (should_failslab(cachep, flags)) 3449 return NULL; 3450 3451 cache_alloc_debugcheck_before(cachep, flags); 3452 local_irq_save(save_flags); 3453 objp = __do_cache_alloc(cachep, flags); 3454 local_irq_restore(save_flags); 3455 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); 3456 prefetchw(objp); 3457 3458 return objp; 3459} 3460 3461/* 3462 * Caller needs to acquire correct kmem_list's list_lock 3463 */ 3464static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 3465 int node) 3466{ 3467 int i; 3468 struct kmem_list3 *l3; 3469 3470 for (i = 0; i < nr_objects; i++) { 3471 void *objp = objpp[i]; 3472 struct slab *slabp; 3473 3474 slabp = virt_to_slab(objp); 3475 l3 = cachep->nodelists[node]; 3476 list_del(&slabp->list); 3477 check_spinlock_acquired_node(cachep, node); 3478 check_slabp(cachep, slabp); 3479 slab_put_obj(cachep, slabp, objp, node); 3480 STATS_DEC_ACTIVE(cachep); 3481 l3->free_objects++; 3482 check_slabp(cachep, slabp); 3483 3484 /* fixup slab chains */ 3485 if (slabp->inuse == 0) { 3486 if (l3->free_objects > l3->free_limit) { 3487 l3->free_objects -= cachep->num; 3488 /* No need to drop any previously held 3489 * lock here, even if we have a off-slab slab 3490 * descriptor it is guaranteed to come from 3491 * a different cache, refer to comments before 3492 * alloc_slabmgmt. 3493 */ 3494 slab_destroy(cachep, slabp); 3495 } else { 3496 list_add(&slabp->list, &l3->slabs_free); 3497 } 3498 } else { 3499 /* Unconditionally move a slab to the end of the 3500 * partial list on free - maximum time for the 3501 * other objects to be freed, too. 3502 */ 3503 list_add_tail(&slabp->list, &l3->slabs_partial); 3504 } 3505 } 3506} 3507 3508static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3509{ 3510 int batchcount; 3511 struct kmem_list3 *l3; 3512 int node = numa_node_id(); 3513 3514 batchcount = ac->batchcount; 3515#if DEBUG 3516 BUG_ON(!batchcount || batchcount > ac->avail); 3517#endif 3518 check_irq_off(); 3519 l3 = cachep->nodelists[node]; 3520 spin_lock(&l3->list_lock); 3521 if (l3->shared) { 3522 struct array_cache *shared_array = l3->shared; 3523 int max = shared_array->limit - shared_array->avail; 3524 if (max) { 3525 if (batchcount > max) 3526 batchcount = max; 3527 memcpy(&(shared_array->entry[shared_array->avail]), 3528 ac->entry, sizeof(void *) * batchcount); 3529 shared_array->avail += batchcount; 3530 goto free_done; 3531 } 3532 } 3533 3534 free_block(cachep, ac->entry, batchcount, node); 3535free_done: 3536#if STATS 3537 { 3538 int i = 0; 3539 struct list_head *p; 3540 3541 p = l3->slabs_free.next; 3542 while (p != &(l3->slabs_free)) { 3543 struct slab *slabp; 3544 3545 slabp = list_entry(p, struct slab, list); 3546 BUG_ON(slabp->inuse); 3547 3548 i++; 3549 p = p->next; 3550 } 3551 STATS_SET_FREEABLE(cachep, i); 3552 } 3553#endif 3554 spin_unlock(&l3->list_lock); 3555 ac->avail -= batchcount; 3556 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3557} 3558 3559/* 3560 * Release an obj back to its cache. If the obj has a constructed state, it must 3561 * be in this state _before_ it is released. Called with disabled ints. 3562 */ 3563static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3564{ 3565 struct array_cache *ac = cpu_cache_get(cachep); 3566 3567 check_irq_off(); 3568 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3569 3570 if (use_alien_caches && cache_free_alien(cachep, objp)) 3571 return; 3572 3573 if (likely(ac->avail < ac->limit)) { 3574 STATS_INC_FREEHIT(cachep); 3575 ac->entry[ac->avail++] = objp; 3576 return; 3577 } else { 3578 STATS_INC_FREEMISS(cachep); 3579 cache_flusharray(cachep, ac); 3580 ac->entry[ac->avail++] = objp; 3581 } 3582} 3583 3584/** 3585 * kmem_cache_alloc - Allocate an object 3586 * @cachep: The cache to allocate from. 3587 * @flags: See kmalloc(). 3588 * 3589 * Allocate an object from this cache. The flags are only relevant 3590 * if the cache has no available objects. 3591 */ 3592void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3593{ 3594 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3595} 3596EXPORT_SYMBOL(kmem_cache_alloc); 3597 3598/** 3599 * kmem_cache_zalloc - Allocate an object. The memory is set to zero. 3600 * @cache: The cache to allocate from. 3601 * @flags: See kmalloc(). 3602 * 3603 * Allocate an object from this cache and set the allocated memory to zero. 3604 * The flags are only relevant if the cache has no available objects. 3605 */ 3606void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) 3607{ 3608 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); 3609 if (ret) 3610 memset(ret, 0, obj_size(cache)); 3611 return ret; 3612} 3613EXPORT_SYMBOL(kmem_cache_zalloc); 3614 3615/** 3616 * kmem_ptr_validate - check if an untrusted pointer might 3617 * be a slab entry. 3618 * @cachep: the cache we're checking against 3619 * @ptr: pointer to validate 3620 * 3621 * This verifies that the untrusted pointer looks sane: 3622 * it is _not_ a guarantee that the pointer is actually 3623 * part of the slab cache in question, but it at least 3624 * validates that the pointer can be dereferenced and 3625 * looks half-way sane. 3626 * 3627 * Currently only used for dentry validation. 3628 */ 3629int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) 3630{ 3631 unsigned long addr = (unsigned long)ptr; 3632 unsigned long min_addr = PAGE_OFFSET; 3633 unsigned long align_mask = BYTES_PER_WORD - 1; 3634 unsigned long size = cachep->buffer_size; 3635 struct page *page; 3636 3637 if (unlikely(addr < min_addr)) 3638 goto out; 3639 if (unlikely(addr > (unsigned long)high_memory - size)) 3640 goto out; 3641 if (unlikely(addr & align_mask)) 3642 goto out; 3643 if (unlikely(!kern_addr_valid(addr))) 3644 goto out; 3645 if (unlikely(!kern_addr_valid(addr + size - 1))) 3646 goto out; 3647 page = virt_to_page(ptr); 3648 if (unlikely(!PageSlab(page))) 3649 goto out; 3650 if (unlikely(page_get_cache(page) != cachep)) 3651 goto out; 3652 return 1; 3653out: 3654 return 0; 3655} 3656 3657#ifdef CONFIG_NUMA 3658void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3659{ 3660 return __cache_alloc_node(cachep, flags, nodeid, 3661 __builtin_return_address(0)); 3662} 3663EXPORT_SYMBOL(kmem_cache_alloc_node); 3664 3665static __always_inline void * 3666__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) 3667{ 3668 struct kmem_cache *cachep; 3669 3670 cachep = kmem_find_general_cachep(size, flags); 3671 if (unlikely(cachep == NULL)) 3672 return NULL; 3673 return kmem_cache_alloc_node(cachep, flags, node); 3674} 3675 3676#ifdef CONFIG_DEBUG_SLAB 3677void *__kmalloc_node(size_t size, gfp_t flags, int node) 3678{ 3679 return __do_kmalloc_node(size, flags, node, 3680 __builtin_return_address(0)); 3681} 3682EXPORT_SYMBOL(__kmalloc_node); 3683 3684void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3685 int node, void *caller) 3686{ 3687 return __do_kmalloc_node(size, flags, node, caller); 3688} 3689EXPORT_SYMBOL(__kmalloc_node_track_caller); 3690#else 3691void *__kmalloc_node(size_t size, gfp_t flags, int node) 3692{ 3693 return __do_kmalloc_node(size, flags, node, NULL); 3694} 3695EXPORT_SYMBOL(__kmalloc_node); 3696#endif /* CONFIG_DEBUG_SLAB */ 3697#endif /* CONFIG_NUMA */ 3698 3699/** 3700 * __do_kmalloc - allocate memory 3701 * @size: how many bytes of memory are required. 3702 * @flags: the type of memory to allocate (see kmalloc). 3703 * @caller: function caller for debug tracking of the caller 3704 */ 3705static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3706 void *caller) 3707{ 3708 struct kmem_cache *cachep; 3709 3710 /* If you want to save a few bytes .text space: replace 3711 * __ with kmem_. 3712 * Then kmalloc uses the uninlined functions instead of the inline 3713 * functions. 3714 */ 3715 cachep = __find_general_cachep(size, flags); 3716 if (unlikely(cachep == NULL)) 3717 return NULL; 3718 return __cache_alloc(cachep, flags, caller); 3719} 3720 3721 3722#ifdef CONFIG_DEBUG_SLAB 3723void *__kmalloc(size_t size, gfp_t flags) 3724{ 3725 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3726} 3727EXPORT_SYMBOL(__kmalloc); 3728 3729void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) 3730{ 3731 return __do_kmalloc(size, flags, caller); 3732} 3733EXPORT_SYMBOL(__kmalloc_track_caller); 3734 3735#else 3736void *__kmalloc(size_t size, gfp_t flags) 3737{ 3738 return __do_kmalloc(size, flags, NULL); 3739} 3740EXPORT_SYMBOL(__kmalloc); 3741#endif 3742 3743/** 3744 * krealloc - reallocate memory. The contents will remain unchanged. 3745 * @p: object to reallocate memory for. 3746 * @new_size: how many bytes of memory are required. 3747 * @flags: the type of memory to allocate. 3748 * 3749 * The contents of the object pointed to are preserved up to the 3750 * lesser of the new and old sizes. If @p is %NULL, krealloc() 3751 * behaves exactly like kmalloc(). If @size is 0 and @p is not a 3752 * %NULL pointer, the object pointed to is freed. 3753 */ 3754void *krealloc(const void *p, size_t new_size, gfp_t flags) 3755{ 3756 struct kmem_cache *cache, *new_cache; 3757 void *ret; 3758 3759 if (unlikely(!p)) 3760 return kmalloc_track_caller(new_size, flags); 3761 3762 if (unlikely(!new_size)) { 3763 kfree(p); 3764 return NULL; 3765 } 3766 3767 cache = virt_to_cache(p); 3768 new_cache = __find_general_cachep(new_size, flags); 3769 3770 /* 3771 * If new size fits in the current cache, bail out. 3772 */ 3773 if (likely(cache == new_cache)) 3774 return (void *)p; 3775 3776 /* 3777 * We are on the slow-path here so do not use __cache_alloc 3778 * because it bloats kernel text. 3779 */ 3780 ret = kmalloc_track_caller(new_size, flags); 3781 if (ret) { 3782 memcpy(ret, p, min(new_size, ksize(p))); 3783 kfree(p); 3784 } 3785 return ret; 3786} 3787EXPORT_SYMBOL(krealloc); 3788 3789/** 3790 * kmem_cache_free - Deallocate an object 3791 * @cachep: The cache the allocation was from. 3792 * @objp: The previously allocated object. 3793 * 3794 * Free an object which was previously allocated from this 3795 * cache. 3796 */ 3797void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3798{ 3799 unsigned long flags; 3800 3801 BUG_ON(virt_to_cache(objp) != cachep); 3802 3803 local_irq_save(flags); 3804 debug_check_no_locks_freed(objp, obj_size(cachep)); 3805 __cache_free(cachep, objp); 3806 local_irq_restore(flags); 3807} 3808EXPORT_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 */ 3819void kfree(const void *objp) 3820{ 3821 struct kmem_cache *c; 3822 unsigned long flags; 3823 3824 if (unlikely(!objp)) 3825 return; 3826 local_irq_save(flags); 3827 kfree_debugcheck(objp); 3828 c = virt_to_cache(objp); 3829 debug_check_no_locks_freed(objp, obj_size(c)); 3830 __cache_free(c, (void *)objp); 3831 local_irq_restore(flags); 3832} 3833EXPORT_SYMBOL(kfree); 3834 3835unsigned int kmem_cache_size(struct kmem_cache *cachep) 3836{ 3837 return obj_size(cachep); 3838} 3839EXPORT_SYMBOL(kmem_cache_size); 3840 3841const char *kmem_cache_name(struct kmem_cache *cachep) 3842{ 3843 return cachep->name; 3844} 3845EXPORT_SYMBOL_GPL(kmem_cache_name); 3846 3847/* 3848 * This initializes kmem_list3 or resizes varioius caches for all nodes. 3849 */ 3850static int alloc_kmemlist(struct kmem_cache *cachep) 3851{ 3852 int node; 3853 struct kmem_list3 *l3; 3854 struct array_cache *new_shared; 3855 struct array_cache **new_alien = NULL; 3856 3857 for_each_online_node(node) { 3858 3859 if (use_alien_caches) { 3860 new_alien = alloc_alien_cache(node, cachep->limit); 3861 if (!new_alien) 3862 goto fail; 3863 } 3864 3865 new_shared = NULL; 3866 if (cachep->shared) { 3867 new_shared = alloc_arraycache(node, 3868 cachep->shared*cachep->batchcount, 3869 0xbaadf00d); 3870 if (!new_shared) { 3871 free_alien_cache(new_alien); 3872 goto fail; 3873 } 3874 } 3875 3876 l3 = cachep->nodelists[node]; 3877 if (l3) { 3878 struct array_cache *shared = l3->shared; 3879 3880 spin_lock_irq(&l3->list_lock); 3881 3882 if (shared) 3883 free_block(cachep, shared->entry, 3884 shared->avail, node); 3885 3886 l3->shared = new_shared; 3887 if (!l3->alien) { 3888 l3->alien = new_alien; 3889 new_alien = NULL; 3890 } 3891 l3->free_limit = (1 + nr_cpus_node(node)) * 3892 cachep->batchcount + cachep->num; 3893 spin_unlock_irq(&l3->list_lock); 3894 kfree(shared); 3895 free_alien_cache(new_alien); 3896 continue; 3897 } 3898 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); 3899 if (!l3) { 3900 free_alien_cache(new_alien); 3901 kfree(new_shared); 3902 goto fail; 3903 } 3904 3905 kmem_list3_init(l3); 3906 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3907 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3908 l3->shared = new_shared; 3909 l3->alien = new_alien; 3910 l3->free_limit = (1 + nr_cpus_node(node)) * 3911 cachep->batchcount + cachep->num; 3912 cachep->nodelists[node] = l3; 3913 } 3914 return 0; 3915 3916fail: 3917 if (!cachep->next.next) { 3918 /* Cache is not active yet. Roll back what we did */ 3919 node--; 3920 while (node >= 0) { 3921 if (cachep->nodelists[node]) { 3922 l3 = cachep->nodelists[node]; 3923 3924 kfree(l3->shared); 3925 free_alien_cache(l3->alien); 3926 kfree(l3); 3927 cachep->nodelists[node] = NULL; 3928 } 3929 node--; 3930 } 3931 } 3932 return -ENOMEM; 3933} 3934 3935struct ccupdate_struct { 3936 struct kmem_cache *cachep; 3937 struct array_cache *new[NR_CPUS]; 3938}; 3939 3940static void do_ccupdate_local(void *info) 3941{ 3942 struct ccupdate_struct *new = info; 3943 struct array_cache *old; 3944 3945 check_irq_off(); 3946 old = cpu_cache_get(new->cachep); 3947 3948 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3949 new->new[smp_processor_id()] = old; 3950} 3951 3952/* Always called with the cache_chain_mutex held */ 3953static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3954 int batchcount, int shared) 3955{ 3956 struct ccupdate_struct *new; 3957 int i; 3958 3959 new = kzalloc(sizeof(*new), GFP_KERNEL); 3960 if (!new) 3961 return -ENOMEM; 3962 3963 for_each_online_cpu(i) { 3964 new->new[i] = alloc_arraycache(cpu_to_node(i), limit, 3965 batchcount); 3966 if (!new->new[i]) { 3967 for (i--; i >= 0; i--) 3968 kfree(new->new[i]); 3969 kfree(new); 3970 return -ENOMEM; 3971 } 3972 } 3973 new->cachep = cachep; 3974 3975 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); 3976 3977 check_irq_on(); 3978 cachep->batchcount = batchcount; 3979 cachep->limit = limit; 3980 cachep->shared = shared; 3981 3982 for_each_online_cpu(i) { 3983 struct array_cache *ccold = new->new[i]; 3984 if (!ccold) 3985 continue; 3986 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3987 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 3988 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3989 kfree(ccold); 3990 } 3991 kfree(new); 3992 return alloc_kmemlist(cachep); 3993} 3994 3995/* Called with cache_chain_mutex held always */ 3996static int enable_cpucache(struct kmem_cache *cachep) 3997{ 3998 int err; 3999 int limit, shared; 4000 4001 /* 4002 * The head array serves three purposes: 4003 * - create a LIFO ordering, i.e. return objects that are cache-warm 4004 * - reduce the number of spinlock operations. 4005 * - reduce the number of linked list operations on the slab and 4006 * bufctl chains: array operations are cheaper. 4007 * The numbers are guessed, we should auto-tune as described by 4008 * Bonwick. 4009 */ 4010 if (cachep->buffer_size > 131072) 4011 limit = 1; 4012 else if (cachep->buffer_size > PAGE_SIZE) 4013 limit = 8; 4014 else if (cachep->buffer_size > 1024) 4015 limit = 24; 4016 else if (cachep->buffer_size > 256) 4017 limit = 54; 4018 else 4019 limit = 120; 4020 4021 /* 4022 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 4023 * allocation behaviour: Most allocs on one cpu, most free operations 4024 * on another cpu. For these cases, an efficient object passing between 4025 * cpus is necessary. This is provided by a shared array. The array 4026 * replaces Bonwick's magazine layer. 4027 * On uniprocessor, it's functionally equivalent (but less efficient) 4028 * to a larger limit. Thus disabled by default. 4029 */ 4030 shared = 0; 4031 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) 4032 shared = 8; 4033 4034#if DEBUG 4035 /* 4036 * With debugging enabled, large batchcount lead to excessively long 4037 * periods with disabled local interrupts. Limit the batchcount 4038 */ 4039 if (limit > 32) 4040 limit = 32; 4041#endif 4042 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); 4043 if (err) 4044 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 4045 cachep->name, -err); 4046 return err; 4047} 4048 4049/* 4050 * Drain an array if it contains any elements taking the l3 lock only if 4051 * necessary. Note that the l3 listlock also protects the array_cache 4052 * if drain_array() is used on the shared array. 4053 */ 4054void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 4055 struct array_cache *ac, int force, int node) 4056{ 4057 int tofree; 4058 4059 if (!ac || !ac->avail) 4060 return; 4061 if (ac->touched && !force) { 4062 ac->touched = 0; 4063 } else { 4064 spin_lock_irq(&l3->list_lock); 4065 if (ac->avail) { 4066 tofree = force ? ac->avail : (ac->limit + 4) / 5; 4067 if (tofree > ac->avail) 4068 tofree = (ac->avail + 1) / 2; 4069 free_block(cachep, ac->entry, tofree, node); 4070 ac->avail -= tofree; 4071 memmove(ac->entry, &(ac->entry[tofree]), 4072 sizeof(void *) * ac->avail); 4073 } 4074 spin_unlock_irq(&l3->list_lock); 4075 } 4076} 4077 4078/** 4079 * cache_reap - Reclaim memory from caches. 4080 * @w: work descriptor 4081 * 4082 * Called from workqueue/eventd every few seconds. 4083 * Purpose: 4084 * - clear the per-cpu caches for this CPU. 4085 * - return freeable pages to the main free memory pool. 4086 * 4087 * If we cannot acquire the cache chain mutex then just give up - we'll try 4088 * again on the next iteration. 4089 */ 4090static void cache_reap(struct work_struct *w) 4091{ 4092 struct kmem_cache *searchp; 4093 struct kmem_list3 *l3; 4094 int node = numa_node_id(); 4095 struct delayed_work *work = 4096 container_of(w, struct delayed_work, work); 4097 4098 if (!mutex_trylock(&cache_chain_mutex)) 4099 /* Give up. Setup the next iteration. */ 4100 goto out; 4101 4102 list_for_each_entry(searchp, &cache_chain, next) { 4103 check_irq_on(); 4104 4105 /* 4106 * We only take the l3 lock if absolutely necessary and we 4107 * have established with reasonable certainty that 4108 * we can do some work if the lock was obtained. 4109 */ 4110 l3 = searchp->nodelists[node]; 4111 4112 reap_alien(searchp, l3); 4113 4114 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 4115 4116 /* 4117 * These are racy checks but it does not matter 4118 * if we skip one check or scan twice. 4119 */ 4120 if (time_after(l3->next_reap, jiffies)) 4121 goto next; 4122 4123 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 4124 4125 drain_array(searchp, l3, l3->shared, 0, node); 4126 4127 if (l3->free_touched) 4128 l3->free_touched = 0; 4129 else { 4130 int freed; 4131 4132 freed = drain_freelist(searchp, l3, (l3->free_limit + 4133 5 * searchp->num - 1) / (5 * searchp->num)); 4134 STATS_ADD_REAPED(searchp, freed); 4135 } 4136next: 4137 cond_resched(); 4138 } 4139 check_irq_on(); 4140 mutex_unlock(&cache_chain_mutex); 4141 next_reap_node(); 4142 refresh_cpu_vm_stats(smp_processor_id()); 4143out: 4144 /* Set up the next iteration */ 4145 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); 4146} 4147 4148#ifdef CONFIG_PROC_FS 4149 4150static void print_slabinfo_header(struct seq_file *m) 4151{ 4152 /* 4153 * Output format version, so at least we can change it 4154 * without _too_ many complaints. 4155 */ 4156#if STATS 4157 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 4158#else 4159 seq_puts(m, "slabinfo - version: 2.1\n"); 4160#endif 4161 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4162 "<objperslab> <pagesperslab>"); 4163 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4164 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4165#if STATS 4166 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 4167 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 4168 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 4169#endif 4170 seq_putc(m, '\n'); 4171} 4172 4173static void *s_start(struct seq_file *m, loff_t *pos) 4174{ 4175 loff_t n = *pos; 4176 struct list_head *p; 4177 4178 mutex_lock(&cache_chain_mutex); 4179 if (!n) 4180 print_slabinfo_header(m); 4181 p = cache_chain.next; 4182 while (n--) { 4183 p = p->next; 4184 if (p == &cache_chain) 4185 return NULL; 4186 } 4187 return list_entry(p, struct kmem_cache, next); 4188} 4189 4190static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4191{ 4192 struct kmem_cache *cachep = p; 4193 ++*pos; 4194 return cachep->next.next == &cache_chain ? 4195 NULL : list_entry(cachep->next.next, struct kmem_cache, next); 4196} 4197 4198static void s_stop(struct seq_file *m, void *p) 4199{ 4200 mutex_unlock(&cache_chain_mutex); 4201} 4202 4203static int s_show(struct seq_file *m, void *p) 4204{ 4205 struct kmem_cache *cachep = p; 4206 struct slab *slabp; 4207 unsigned long active_objs; 4208 unsigned long num_objs; 4209 unsigned long active_slabs = 0; 4210 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 4211 const char *name; 4212 char *error = NULL; 4213 int node; 4214 struct kmem_list3 *l3; 4215 4216 active_objs = 0; 4217 num_slabs = 0; 4218 for_each_online_node(node) { 4219 l3 = cachep->nodelists[node]; 4220 if (!l3) 4221 continue; 4222 4223 check_irq_on(); 4224 spin_lock_irq(&l3->list_lock); 4225 4226 list_for_each_entry(slabp, &l3->slabs_full, list) { 4227 if (slabp->inuse != cachep->num && !error) 4228 error = "slabs_full accounting error"; 4229 active_objs += cachep->num; 4230 active_slabs++; 4231 } 4232 list_for_each_entry(slabp, &l3->slabs_partial, list) { 4233 if (slabp->inuse == cachep->num && !error) 4234 error = "slabs_partial inuse accounting error"; 4235 if (!slabp->inuse && !error) 4236 error = "slabs_partial/inuse accounting error"; 4237 active_objs += slabp->inuse; 4238 active_slabs++; 4239 } 4240 list_for_each_entry(slabp, &l3->slabs_free, list) { 4241 if (slabp->inuse && !error) 4242 error = "slabs_free/inuse accounting error"; 4243 num_slabs++; 4244 } 4245 free_objects += l3->free_objects; 4246 if (l3->shared) 4247 shared_avail += l3->shared->avail; 4248 4249 spin_unlock_irq(&l3->list_lock); 4250 } 4251 num_slabs += active_slabs; 4252 num_objs = num_slabs * cachep->num; 4253 if (num_objs - active_objs != free_objects && !error) 4254 error = "free_objects accounting error"; 4255 4256 name = cachep->name; 4257 if (error) 4258 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 4259 4260 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 4261 name, active_objs, num_objs, cachep->buffer_size, 4262 cachep->num, (1 << cachep->gfporder)); 4263 seq_printf(m, " : tunables %4u %4u %4u", 4264 cachep->limit, cachep->batchcount, cachep->shared); 4265 seq_printf(m, " : slabdata %6lu %6lu %6lu", 4266 active_slabs, num_slabs, shared_avail); 4267#if STATS 4268 { /* list3 stats */ 4269 unsigned long high = cachep->high_mark; 4270 unsigned long allocs = cachep->num_allocations; 4271 unsigned long grown = cachep->grown; 4272 unsigned long reaped = cachep->reaped; 4273 unsigned long errors = cachep->errors; 4274 unsigned long max_freeable = cachep->max_freeable; 4275 unsigned long node_allocs = cachep->node_allocs; 4276 unsigned long node_frees = cachep->node_frees; 4277 unsigned long overflows = cachep->node_overflow; 4278 4279 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ 4280 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, 4281 reaped, errors, max_freeable, node_allocs, 4282 node_frees, overflows); 4283 } 4284 /* cpu stats */ 4285 { 4286 unsigned long allochit = atomic_read(&cachep->allochit); 4287 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4288 unsigned long freehit = atomic_read(&cachep->freehit); 4289 unsigned long freemiss = atomic_read(&cachep->freemiss); 4290 4291 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4292 allochit, allocmiss, freehit, freemiss); 4293 } 4294#endif 4295 seq_putc(m, '\n'); 4296 return 0; 4297} 4298 4299/* 4300 * slabinfo_op - iterator that generates /proc/slabinfo 4301 * 4302 * Output layout: 4303 * cache-name 4304 * num-active-objs 4305 * total-objs 4306 * object size 4307 * num-active-slabs 4308 * total-slabs 4309 * num-pages-per-slab 4310 * + further values on SMP and with statistics enabled 4311 */ 4312 4313const struct seq_operations slabinfo_op = { 4314 .start = s_start, 4315 .next = s_next, 4316 .stop = s_stop, 4317 .show = s_show, 4318}; 4319 4320#define MAX_SLABINFO_WRITE 128 4321/** 4322 * slabinfo_write - Tuning for the slab allocator 4323 * @file: unused 4324 * @buffer: user buffer 4325 * @count: data length 4326 * @ppos: unused 4327 */ 4328ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4329 size_t count, loff_t *ppos) 4330{ 4331 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4332 int limit, batchcount, shared, res; 4333 struct kmem_cache *cachep; 4334 4335 if (count > MAX_SLABINFO_WRITE) 4336 return -EINVAL; 4337 if (copy_from_user(&kbuf, buffer, count)) 4338 return -EFAULT; 4339 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4340 4341 tmp = strchr(kbuf, ' '); 4342 if (!tmp) 4343 return -EINVAL; 4344 *tmp = '\0'; 4345 tmp++; 4346 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4347 return -EINVAL; 4348 4349 /* Find the cache in the chain of caches. */ 4350 mutex_lock(&cache_chain_mutex); 4351 res = -EINVAL; 4352 list_for_each_entry(cachep, &cache_chain, next) { 4353 if (!strcmp(cachep->name, kbuf)) { 4354 if (limit < 1 || batchcount < 1 || 4355 batchcount > limit || shared < 0) { 4356 res = 0; 4357 } else { 4358 res = do_tune_cpucache(cachep, limit, 4359 batchcount, shared); 4360 } 4361 break; 4362 } 4363 } 4364 mutex_unlock(&cache_chain_mutex); 4365 if (res >= 0) 4366 res = count; 4367 return res; 4368} 4369 4370#ifdef CONFIG_DEBUG_SLAB_LEAK 4371 4372static void *leaks_start(struct seq_file *m, loff_t *pos) 4373{ 4374 loff_t n = *pos; 4375 struct list_head *p; 4376 4377 mutex_lock(&cache_chain_mutex); 4378 p = cache_chain.next; 4379 while (n--) { 4380 p = p->next; 4381 if (p == &cache_chain) 4382 return NULL; 4383 } 4384 return list_entry(p, struct kmem_cache, next); 4385} 4386 4387static inline int add_caller(unsigned long *n, unsigned long v) 4388{ 4389 unsigned long *p; 4390 int l; 4391 if (!v) 4392 return 1; 4393 l = n[1]; 4394 p = n + 2; 4395 while (l) { 4396 int i = l/2; 4397 unsigned long *q = p + 2 * i; 4398 if (*q == v) { 4399 q[1]++; 4400 return 1; 4401 } 4402 if (*q > v) { 4403 l = i; 4404 } else { 4405 p = q + 2; 4406 l -= i + 1; 4407 } 4408 } 4409 if (++n[1] == n[0]) 4410 return 0; 4411 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4412 p[0] = v; 4413 p[1] = 1; 4414 return 1; 4415} 4416 4417static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4418{ 4419 void *p; 4420 int i; 4421 if (n[0] == n[1]) 4422 return; 4423 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4424 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4425 continue; 4426 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4427 return; 4428 } 4429} 4430 4431static void show_symbol(struct seq_file *m, unsigned long address) 4432{ 4433#ifdef CONFIG_KALLSYMS 4434 unsigned long offset, size; 4435 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1]; 4436 4437 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { 4438 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4439 if (modname[0]) 4440 seq_printf(m, " [%s]", modname); 4441 return; 4442 } 4443#endif 4444 seq_printf(m, "%p", (void *)address); 4445} 4446 4447static int leaks_show(struct seq_file *m, void *p) 4448{ 4449 struct kmem_cache *cachep = p; 4450 struct slab *slabp; 4451 struct kmem_list3 *l3; 4452 const char *name; 4453 unsigned long *n = m->private; 4454 int node; 4455 int i; 4456 4457 if (!(cachep->flags & SLAB_STORE_USER)) 4458 return 0; 4459 if (!(cachep->flags & SLAB_RED_ZONE)) 4460 return 0; 4461 4462 /* OK, we can do it */ 4463 4464 n[1] = 0; 4465 4466 for_each_online_node(node) { 4467 l3 = cachep->nodelists[node]; 4468 if (!l3) 4469 continue; 4470 4471 check_irq_on(); 4472 spin_lock_irq(&l3->list_lock); 4473 4474 list_for_each_entry(slabp, &l3->slabs_full, list) 4475 handle_slab(n, cachep, slabp); 4476 list_for_each_entry(slabp, &l3->slabs_partial, list) 4477 handle_slab(n, cachep, slabp); 4478 spin_unlock_irq(&l3->list_lock); 4479 } 4480 name = cachep->name; 4481 if (n[0] == n[1]) { 4482 /* Increase the buffer size */ 4483 mutex_unlock(&cache_chain_mutex); 4484 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4485 if (!m->private) { 4486 /* Too bad, we are really out */ 4487 m->private = n; 4488 mutex_lock(&cache_chain_mutex); 4489 return -ENOMEM; 4490 } 4491 *(unsigned long *)m->private = n[0] * 2; 4492 kfree(n); 4493 mutex_lock(&cache_chain_mutex); 4494 /* Now make sure this entry will be retried */ 4495 m->count = m->size; 4496 return 0; 4497 } 4498 for (i = 0; i < n[1]; i++) { 4499 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4500 show_symbol(m, n[2*i+2]); 4501 seq_putc(m, '\n'); 4502 } 4503 4504 return 0; 4505} 4506 4507const struct seq_operations slabstats_op = { 4508 .start = leaks_start, 4509 .next = s_next, 4510 .stop = s_stop, 4511 .show = leaks_show, 4512}; 4513#endif 4514#endif 4515 4516/** 4517 * ksize - get the actual amount of memory allocated for a given object 4518 * @objp: Pointer to the object 4519 * 4520 * kmalloc may internally round up allocations and return more memory 4521 * than requested. ksize() can be used to determine the actual amount of 4522 * memory allocated. The caller may use this additional memory, even though 4523 * a smaller amount of memory was initially specified with the kmalloc call. 4524 * The caller must guarantee that objp points to a valid object previously 4525 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4526 * must not be freed during the duration of the call. 4527 */ 4528size_t ksize(const void *objp) 4529{ 4530 if (unlikely(objp == NULL)) 4531 return 0; 4532 4533 return obj_size(virt_to_cache(objp)); 4534} 4535