slab.c revision e5ac9c5aec7c4bc57fa93f2d37d760a22cb7bd33
1/* 2 * linux/mm/slab.c 3 * Written by Mark Hemment, 1996/97. 4 * (markhe@nextd.demon.co.uk) 5 * 6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 7 * 8 * Major cleanup, different bufctl logic, per-cpu arrays 9 * (c) 2000 Manfred Spraul 10 * 11 * Cleanup, make the head arrays unconditional, preparation for NUMA 12 * (c) 2002 Manfred Spraul 13 * 14 * An implementation of the Slab Allocator as described in outline in; 15 * UNIX Internals: The New Frontiers by Uresh Vahalia 16 * Pub: Prentice Hall ISBN 0-13-101908-2 17 * or with a little more detail in; 18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 19 * Jeff Bonwick (Sun Microsystems). 20 * Presented at: USENIX Summer 1994 Technical Conference 21 * 22 * The memory is organized in caches, one cache for each object type. 23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 24 * Each cache consists out of many slabs (they are small (usually one 25 * page long) and always contiguous), and each slab contains multiple 26 * initialized objects. 27 * 28 * This means, that your constructor is used only for newly allocated 29 * slabs and you must pass objects with the same intializations to 30 * kmem_cache_free. 31 * 32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 33 * normal). If you need a special memory type, then must create a new 34 * cache for that memory type. 35 * 36 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 37 * full slabs with 0 free objects 38 * partial slabs 39 * empty slabs with no allocated objects 40 * 41 * If partial slabs exist, then new allocations come from these slabs, 42 * otherwise from empty slabs or new slabs are allocated. 43 * 44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 46 * 47 * Each cache has a short per-cpu head array, most allocs 48 * and frees go into that array, and if that array overflows, then 1/2 49 * of the entries in the array are given back into the global cache. 50 * The head array is strictly LIFO and should improve the cache hit rates. 51 * On SMP, it additionally reduces the spinlock operations. 52 * 53 * The c_cpuarray may not be read with enabled local interrupts - 54 * it's changed with a smp_call_function(). 55 * 56 * SMP synchronization: 57 * constructors and destructors are called without any locking. 58 * Several members in struct kmem_cache and struct slab never change, they 59 * are accessed without any locking. 60 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 61 * and local interrupts are disabled so slab code is preempt-safe. 62 * The non-constant members are protected with a per-cache irq spinlock. 63 * 64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 65 * in 2000 - many ideas in the current implementation are derived from 66 * his patch. 67 * 68 * Further notes from the original documentation: 69 * 70 * 11 April '97. Started multi-threading - markhe 71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'. 72 * The sem is only needed when accessing/extending the cache-chain, which 73 * can never happen inside an interrupt (kmem_cache_create(), 74 * kmem_cache_shrink() and kmem_cache_reap()). 75 * 76 * At present, each engine can be growing a cache. This should be blocked. 77 * 78 * 15 March 2005. NUMA slab allocator. 79 * Shai Fultheim <shai@scalex86.org>. 80 * Shobhit Dayal <shobhit@calsoftinc.com> 81 * Alok N Kataria <alokk@calsoftinc.com> 82 * Christoph Lameter <christoph@lameter.com> 83 * 84 * Modified the slab allocator to be node aware on NUMA systems. 85 * Each node has its own list of partial, free and full slabs. 86 * All object allocations for a node occur from node specific slab lists. 87 */ 88 89#include <linux/config.h> 90#include <linux/slab.h> 91#include <linux/mm.h> 92#include <linux/poison.h> 93#include <linux/swap.h> 94#include <linux/cache.h> 95#include <linux/interrupt.h> 96#include <linux/init.h> 97#include <linux/compiler.h> 98#include <linux/cpuset.h> 99#include <linux/seq_file.h> 100#include <linux/notifier.h> 101#include <linux/kallsyms.h> 102#include <linux/cpu.h> 103#include <linux/sysctl.h> 104#include <linux/module.h> 105#include <linux/rcupdate.h> 106#include <linux/string.h> 107#include <linux/nodemask.h> 108#include <linux/mempolicy.h> 109#include <linux/mutex.h> 110#include <linux/rtmutex.h> 111 112#include <asm/uaccess.h> 113#include <asm/cacheflush.h> 114#include <asm/tlbflush.h> 115#include <asm/page.h> 116 117/* 118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, 119 * 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 BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. 152 * Note that this flag disables some debug features. 153 */ 154#define ARCH_KMALLOC_MINALIGN 0 155#endif 156 157#ifndef ARCH_SLAB_MINALIGN 158/* 159 * Enforce a minimum alignment for all caches. 160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD 161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. 162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables 163 * some debug features. 164 */ 165#define ARCH_SLAB_MINALIGN 0 166#endif 167 168#ifndef ARCH_KMALLOC_FLAGS 169#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 170#endif 171 172/* Legal flag mask for kmem_cache_create(). */ 173#if DEBUG 174# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ 175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 176 SLAB_CACHE_DMA | \ 177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ 178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 180#else 181# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ 183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 185#endif 186 187/* 188 * kmem_bufctl_t: 189 * 190 * Bufctl's are used for linking objs within a slab 191 * linked offsets. 192 * 193 * This implementation relies on "struct page" for locating the cache & 194 * slab an object belongs to. 195 * This allows the bufctl structure to be small (one int), but limits 196 * the number of objects a slab (not a cache) can contain when off-slab 197 * bufctls are used. The limit is the size of the largest general cache 198 * that does not use off-slab slabs. 199 * For 32bit archs with 4 kB pages, is this 56. 200 * This is not serious, as it is only for large objects, when it is unwise 201 * to have too many per slab. 202 * Note: This limit can be raised by introducing a general cache whose size 203 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 204 */ 205 206typedef unsigned int kmem_bufctl_t; 207#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 208#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 209#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 210#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 211 212/* 213 * struct slab 214 * 215 * Manages the objs in a slab. Placed either at the beginning of mem allocated 216 * for a slab, or allocated from an general cache. 217 * Slabs are chained into three list: fully used, partial, fully free slabs. 218 */ 219struct slab { 220 struct list_head list; 221 unsigned long colouroff; 222 void *s_mem; /* including colour offset */ 223 unsigned int inuse; /* num of objs active in slab */ 224 kmem_bufctl_t free; 225 unsigned short nodeid; 226}; 227 228/* 229 * struct slab_rcu 230 * 231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 232 * arrange for kmem_freepages to be called via RCU. This is useful if 233 * we need to approach a kernel structure obliquely, from its address 234 * obtained without the usual locking. We can lock the structure to 235 * stabilize it and check it's still at the given address, only if we 236 * can be sure that the memory has not been meanwhile reused for some 237 * other kind of object (which our subsystem's lock might corrupt). 238 * 239 * rcu_read_lock before reading the address, then rcu_read_unlock after 240 * taking the spinlock within the structure expected at that address. 241 * 242 * We assume struct slab_rcu can overlay struct slab when destroying. 243 */ 244struct slab_rcu { 245 struct rcu_head head; 246 struct kmem_cache *cachep; 247 void *addr; 248}; 249 250/* 251 * struct array_cache 252 * 253 * Purpose: 254 * - LIFO ordering, to hand out cache-warm objects from _alloc 255 * - reduce the number of linked list operations 256 * - reduce spinlock operations 257 * 258 * The limit is stored in the per-cpu structure to reduce the data cache 259 * footprint. 260 * 261 */ 262struct array_cache { 263 unsigned int avail; 264 unsigned int limit; 265 unsigned int batchcount; 266 unsigned int touched; 267 spinlock_t lock; 268 void *entry[0]; /* 269 * Must have this definition in here for the proper 270 * alignment of array_cache. Also simplifies accessing 271 * the entries. 272 * [0] is for gcc 2.95. It should really be []. 273 */ 274}; 275 276/* 277 * bootstrap: The caches do not work without cpuarrays anymore, but the 278 * cpuarrays are allocated from the generic caches... 279 */ 280#define BOOT_CPUCACHE_ENTRIES 1 281struct arraycache_init { 282 struct array_cache cache; 283 void *entries[BOOT_CPUCACHE_ENTRIES]; 284}; 285 286/* 287 * The slab lists for all objects. 288 */ 289struct kmem_list3 { 290 struct list_head slabs_partial; /* partial list first, better asm code */ 291 struct list_head slabs_full; 292 struct list_head slabs_free; 293 unsigned long free_objects; 294 unsigned int free_limit; 295 unsigned int colour_next; /* Per-node cache coloring */ 296 spinlock_t list_lock; 297 struct array_cache *shared; /* shared per node */ 298 struct array_cache **alien; /* on other nodes */ 299 unsigned long next_reap; /* updated without locking */ 300 int free_touched; /* updated without locking */ 301}; 302 303/* 304 * Need this for bootstrapping a per node allocator. 305 */ 306#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) 307struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 308#define CACHE_CACHE 0 309#define SIZE_AC 1 310#define SIZE_L3 (1 + MAX_NUMNODES) 311 312static int drain_freelist(struct kmem_cache *cache, 313 struct kmem_list3 *l3, int tofree); 314static void free_block(struct kmem_cache *cachep, void **objpp, int len, 315 int node); 316static void enable_cpucache(struct kmem_cache *cachep); 317static void cache_reap(void *unused); 318 319/* 320 * This function must be completely optimized away if a constant is passed to 321 * it. Mostly the same as what is in linux/slab.h except it returns an index. 322 */ 323static __always_inline int index_of(const size_t size) 324{ 325 extern void __bad_size(void); 326 327 if (__builtin_constant_p(size)) { 328 int i = 0; 329 330#define CACHE(x) \ 331 if (size <=x) \ 332 return i; \ 333 else \ 334 i++; 335#include "linux/kmalloc_sizes.h" 336#undef CACHE 337 __bad_size(); 338 } else 339 __bad_size(); 340 return 0; 341} 342 343static int slab_early_init = 1; 344 345#define INDEX_AC index_of(sizeof(struct arraycache_init)) 346#define INDEX_L3 index_of(sizeof(struct kmem_list3)) 347 348static void kmem_list3_init(struct kmem_list3 *parent) 349{ 350 INIT_LIST_HEAD(&parent->slabs_full); 351 INIT_LIST_HEAD(&parent->slabs_partial); 352 INIT_LIST_HEAD(&parent->slabs_free); 353 parent->shared = NULL; 354 parent->alien = NULL; 355 parent->colour_next = 0; 356 spin_lock_init(&parent->list_lock); 357 parent->free_objects = 0; 358 parent->free_touched = 0; 359} 360 361#define MAKE_LIST(cachep, listp, slab, nodeid) \ 362 do { \ 363 INIT_LIST_HEAD(listp); \ 364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 365 } while (0) 366 367#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 368 do { \ 369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 372 } while (0) 373 374/* 375 * struct kmem_cache 376 * 377 * manages a cache. 378 */ 379 380struct kmem_cache { 381/* 1) per-cpu data, touched during every alloc/free */ 382 struct array_cache *array[NR_CPUS]; 383/* 2) Cache tunables. Protected by cache_chain_mutex */ 384 unsigned int batchcount; 385 unsigned int limit; 386 unsigned int shared; 387 388 unsigned int buffer_size; 389/* 3) touched by every alloc & free from the backend */ 390 struct kmem_list3 *nodelists[MAX_NUMNODES]; 391 392 unsigned int flags; /* constant flags */ 393 unsigned int num; /* # of objs per slab */ 394 395/* 4) cache_grow/shrink */ 396 /* order of pgs per slab (2^n) */ 397 unsigned int gfporder; 398 399 /* force GFP flags, e.g. GFP_DMA */ 400 gfp_t gfpflags; 401 402 size_t colour; /* cache colouring range */ 403 unsigned int colour_off; /* colour offset */ 404 struct kmem_cache *slabp_cache; 405 unsigned int slab_size; 406 unsigned int dflags; /* dynamic flags */ 407 408 /* constructor func */ 409 void (*ctor) (void *, struct kmem_cache *, unsigned long); 410 411 /* de-constructor func */ 412 void (*dtor) (void *, struct kmem_cache *, unsigned long); 413 414/* 5) cache creation/removal */ 415 const char *name; 416 struct list_head next; 417 418/* 6) statistics */ 419#if STATS 420 unsigned long num_active; 421 unsigned long num_allocations; 422 unsigned long high_mark; 423 unsigned long grown; 424 unsigned long reaped; 425 unsigned long errors; 426 unsigned long max_freeable; 427 unsigned long node_allocs; 428 unsigned long node_frees; 429 unsigned long node_overflow; 430 atomic_t allochit; 431 atomic_t allocmiss; 432 atomic_t freehit; 433 atomic_t freemiss; 434#endif 435#if DEBUG 436 /* 437 * If debugging is enabled, then the allocator can add additional 438 * fields and/or padding to every object. buffer_size contains the total 439 * object size including these internal fields, the following two 440 * variables contain the offset to the user object and its size. 441 */ 442 int obj_offset; 443 int obj_size; 444#endif 445}; 446 447#define CFLGS_OFF_SLAB (0x80000000UL) 448#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 449 450#define BATCHREFILL_LIMIT 16 451/* 452 * Optimization question: fewer reaps means less probability for unnessary 453 * cpucache drain/refill cycles. 454 * 455 * OTOH the cpuarrays can contain lots of objects, 456 * which could lock up otherwise freeable slabs. 457 */ 458#define REAPTIMEOUT_CPUC (2*HZ) 459#define REAPTIMEOUT_LIST3 (4*HZ) 460 461#if STATS 462#define STATS_INC_ACTIVE(x) ((x)->num_active++) 463#define STATS_DEC_ACTIVE(x) ((x)->num_active--) 464#define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 465#define STATS_INC_GROWN(x) ((x)->grown++) 466#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 467#define STATS_SET_HIGH(x) \ 468 do { \ 469 if ((x)->num_active > (x)->high_mark) \ 470 (x)->high_mark = (x)->num_active; \ 471 } while (0) 472#define STATS_INC_ERR(x) ((x)->errors++) 473#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 474#define STATS_INC_NODEFREES(x) ((x)->node_frees++) 475#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 476#define STATS_SET_FREEABLE(x, i) \ 477 do { \ 478 if ((x)->max_freeable < i) \ 479 (x)->max_freeable = i; \ 480 } while (0) 481#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 482#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 483#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 484#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 485#else 486#define STATS_INC_ACTIVE(x) do { } while (0) 487#define STATS_DEC_ACTIVE(x) do { } while (0) 488#define STATS_INC_ALLOCED(x) do { } while (0) 489#define STATS_INC_GROWN(x) do { } while (0) 490#define STATS_ADD_REAPED(x,y) do { } while (0) 491#define STATS_SET_HIGH(x) do { } while (0) 492#define STATS_INC_ERR(x) do { } while (0) 493#define STATS_INC_NODEALLOCS(x) do { } while (0) 494#define STATS_INC_NODEFREES(x) do { } while (0) 495#define STATS_INC_ACOVERFLOW(x) do { } while (0) 496#define STATS_SET_FREEABLE(x, i) do { } while (0) 497#define STATS_INC_ALLOCHIT(x) do { } while (0) 498#define STATS_INC_ALLOCMISS(x) do { } while (0) 499#define STATS_INC_FREEHIT(x) do { } while (0) 500#define STATS_INC_FREEMISS(x) do { } while (0) 501#endif 502 503#if DEBUG 504 505/* 506 * memory layout of objects: 507 * 0 : objp 508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 509 * the end of an object is aligned with the end of the real 510 * allocation. Catches writes behind the end of the allocation. 511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 512 * redzone word. 513 * cachep->obj_offset: The real object. 514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 516 * [BYTES_PER_WORD long] 517 */ 518static int obj_offset(struct kmem_cache *cachep) 519{ 520 return cachep->obj_offset; 521} 522 523static int obj_size(struct kmem_cache *cachep) 524{ 525 return cachep->obj_size; 526} 527 528static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 529{ 530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); 532} 533 534static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 535{ 536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 537 if (cachep->flags & SLAB_STORE_USER) 538 return (unsigned long *)(objp + cachep->buffer_size - 539 2 * BYTES_PER_WORD); 540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); 541} 542 543static void **dbg_userword(struct kmem_cache *cachep, void *objp) 544{ 545 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 547} 548 549#else 550 551#define obj_offset(x) 0 552#define obj_size(cachep) (cachep->buffer_size) 553#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) 554#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) 555#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 556 557#endif 558 559/* 560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp 561 * order. 562 */ 563#if defined(CONFIG_LARGE_ALLOCS) 564#define MAX_OBJ_ORDER 13 /* up to 32Mb */ 565#define MAX_GFP_ORDER 13 /* up to 32Mb */ 566#elif defined(CONFIG_MMU) 567#define MAX_OBJ_ORDER 5 /* 32 pages */ 568#define MAX_GFP_ORDER 5 /* 32 pages */ 569#else 570#define MAX_OBJ_ORDER 8 /* up to 1Mb */ 571#define MAX_GFP_ORDER 8 /* up to 1Mb */ 572#endif 573 574/* 575 * Do not go above this order unless 0 objects fit into the slab. 576 */ 577#define BREAK_GFP_ORDER_HI 1 578#define BREAK_GFP_ORDER_LO 0 579static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 580 581/* 582 * Functions for storing/retrieving the cachep and or slab from the page 583 * allocator. These are used to find the slab an obj belongs to. With kfree(), 584 * these are used to find the cache which an obj belongs to. 585 */ 586static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 587{ 588 page->lru.next = (struct list_head *)cache; 589} 590 591static inline struct kmem_cache *page_get_cache(struct page *page) 592{ 593 if (unlikely(PageCompound(page))) 594 page = (struct page *)page_private(page); 595 BUG_ON(!PageSlab(page)); 596 return (struct kmem_cache *)page->lru.next; 597} 598 599static inline void page_set_slab(struct page *page, struct slab *slab) 600{ 601 page->lru.prev = (struct list_head *)slab; 602} 603 604static inline struct slab *page_get_slab(struct page *page) 605{ 606 if (unlikely(PageCompound(page))) 607 page = (struct page *)page_private(page); 608 BUG_ON(!PageSlab(page)); 609 return (struct slab *)page->lru.prev; 610} 611 612static inline struct kmem_cache *virt_to_cache(const void *obj) 613{ 614 struct page *page = virt_to_page(obj); 615 return page_get_cache(page); 616} 617 618static inline struct slab *virt_to_slab(const void *obj) 619{ 620 struct page *page = virt_to_page(obj); 621 return page_get_slab(page); 622} 623 624static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 625 unsigned int idx) 626{ 627 return slab->s_mem + cache->buffer_size * idx; 628} 629 630static inline unsigned int obj_to_index(struct kmem_cache *cache, 631 struct slab *slab, void *obj) 632{ 633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size; 634} 635 636/* 637 * These are the default caches for kmalloc. Custom caches can have other sizes. 638 */ 639struct cache_sizes malloc_sizes[] = { 640#define CACHE(x) { .cs_size = (x) }, 641#include <linux/kmalloc_sizes.h> 642 CACHE(ULONG_MAX) 643#undef CACHE 644}; 645EXPORT_SYMBOL(malloc_sizes); 646 647/* Must match cache_sizes above. Out of line to keep cache footprint low. */ 648struct cache_names { 649 char *name; 650 char *name_dma; 651}; 652 653static struct cache_names __initdata cache_names[] = { 654#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 655#include <linux/kmalloc_sizes.h> 656 {NULL,} 657#undef CACHE 658}; 659 660static struct arraycache_init initarray_cache __initdata = 661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 662static struct arraycache_init initarray_generic = 663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 664 665/* internal cache of cache description objs */ 666static struct kmem_cache cache_cache = { 667 .batchcount = 1, 668 .limit = BOOT_CPUCACHE_ENTRIES, 669 .shared = 1, 670 .buffer_size = sizeof(struct kmem_cache), 671 .name = "kmem_cache", 672#if DEBUG 673 .obj_size = sizeof(struct kmem_cache), 674#endif 675}; 676 677#ifdef CONFIG_LOCKDEP 678 679/* 680 * Slab sometimes uses the kmalloc slabs to store the slab headers 681 * for other slabs "off slab". 682 * The locking for this is tricky in that it nests within the locks 683 * of all other slabs in a few places; to deal with this special 684 * locking we put on-slab caches into a separate lock-class. 685 */ 686static struct lock_class_key on_slab_key; 687 688static inline void init_lock_keys(struct cache_sizes *s) 689{ 690 int q; 691 692 for (q = 0; q < MAX_NUMNODES; q++) { 693 if (!s->cs_cachep->nodelists[q] || OFF_SLAB(s->cs_cachep)) 694 continue; 695 lockdep_set_class(&s->cs_cachep->nodelists[q]->list_lock, 696 &on_slab_key); 697 } 698} 699 700#else 701static inline void init_lock_keys(struct cache_sizes *s) 702{ 703} 704#endif 705 706 707 708/* Guard access to the cache-chain. */ 709static DEFINE_MUTEX(cache_chain_mutex); 710static struct list_head cache_chain; 711 712/* 713 * vm_enough_memory() looks at this to determine how many slab-allocated pages 714 * are possibly freeable under pressure 715 * 716 * SLAB_RECLAIM_ACCOUNT turns this on per-slab 717 */ 718atomic_t slab_reclaim_pages; 719 720/* 721 * chicken and egg problem: delay the per-cpu array allocation 722 * until the general caches are up. 723 */ 724static enum { 725 NONE, 726 PARTIAL_AC, 727 PARTIAL_L3, 728 FULL 729} g_cpucache_up; 730 731/* 732 * used by boot code to determine if it can use slab based allocator 733 */ 734int slab_is_available(void) 735{ 736 return g_cpucache_up == FULL; 737} 738 739static DEFINE_PER_CPU(struct work_struct, reap_work); 740 741static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 742{ 743 return cachep->array[smp_processor_id()]; 744} 745 746static inline struct kmem_cache *__find_general_cachep(size_t size, 747 gfp_t gfpflags) 748{ 749 struct cache_sizes *csizep = malloc_sizes; 750 751#if DEBUG 752 /* This happens if someone tries to call 753 * kmem_cache_create(), or __kmalloc(), before 754 * the generic caches are initialized. 755 */ 756 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 757#endif 758 while (size > csizep->cs_size) 759 csizep++; 760 761 /* 762 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 763 * has cs_{dma,}cachep==NULL. Thus no special case 764 * for large kmalloc calls required. 765 */ 766 if (unlikely(gfpflags & GFP_DMA)) 767 return csizep->cs_dmacachep; 768 return csizep->cs_cachep; 769} 770 771static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 772{ 773 return __find_general_cachep(size, gfpflags); 774} 775 776static size_t slab_mgmt_size(size_t nr_objs, size_t align) 777{ 778 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 779} 780 781/* 782 * Calculate the number of objects and left-over bytes for a given buffer size. 783 */ 784static void cache_estimate(unsigned long gfporder, size_t buffer_size, 785 size_t align, int flags, size_t *left_over, 786 unsigned int *num) 787{ 788 int nr_objs; 789 size_t mgmt_size; 790 size_t slab_size = PAGE_SIZE << gfporder; 791 792 /* 793 * The slab management structure can be either off the slab or 794 * on it. For the latter case, the memory allocated for a 795 * slab is used for: 796 * 797 * - The struct slab 798 * - One kmem_bufctl_t for each object 799 * - Padding to respect alignment of @align 800 * - @buffer_size bytes for each object 801 * 802 * If the slab management structure is off the slab, then the 803 * alignment will already be calculated into the size. Because 804 * the slabs are all pages aligned, the objects will be at the 805 * correct alignment when allocated. 806 */ 807 if (flags & CFLGS_OFF_SLAB) { 808 mgmt_size = 0; 809 nr_objs = slab_size / buffer_size; 810 811 if (nr_objs > SLAB_LIMIT) 812 nr_objs = SLAB_LIMIT; 813 } else { 814 /* 815 * Ignore padding for the initial guess. The padding 816 * is at most @align-1 bytes, and @buffer_size is at 817 * least @align. In the worst case, this result will 818 * be one greater than the number of objects that fit 819 * into the memory allocation when taking the padding 820 * into account. 821 */ 822 nr_objs = (slab_size - sizeof(struct slab)) / 823 (buffer_size + sizeof(kmem_bufctl_t)); 824 825 /* 826 * This calculated number will be either the right 827 * amount, or one greater than what we want. 828 */ 829 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 830 > slab_size) 831 nr_objs--; 832 833 if (nr_objs > SLAB_LIMIT) 834 nr_objs = SLAB_LIMIT; 835 836 mgmt_size = slab_mgmt_size(nr_objs, align); 837 } 838 *num = nr_objs; 839 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 840} 841 842#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) 843 844static void __slab_error(const char *function, struct kmem_cache *cachep, 845 char *msg) 846{ 847 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 848 function, cachep->name, msg); 849 dump_stack(); 850} 851 852#ifdef CONFIG_NUMA 853/* 854 * Special reaping functions for NUMA systems called from cache_reap(). 855 * These take care of doing round robin flushing of alien caches (containing 856 * objects freed on different nodes from which they were allocated) and the 857 * flushing of remote pcps by calling drain_node_pages. 858 */ 859static DEFINE_PER_CPU(unsigned long, reap_node); 860 861static void init_reap_node(int cpu) 862{ 863 int node; 864 865 node = next_node(cpu_to_node(cpu), node_online_map); 866 if (node == MAX_NUMNODES) 867 node = first_node(node_online_map); 868 869 __get_cpu_var(reap_node) = node; 870} 871 872static void next_reap_node(void) 873{ 874 int node = __get_cpu_var(reap_node); 875 876 /* 877 * Also drain per cpu pages on remote zones 878 */ 879 if (node != numa_node_id()) 880 drain_node_pages(node); 881 882 node = next_node(node, node_online_map); 883 if (unlikely(node >= MAX_NUMNODES)) 884 node = first_node(node_online_map); 885 __get_cpu_var(reap_node) = node; 886} 887 888#else 889#define init_reap_node(cpu) do { } while (0) 890#define next_reap_node(void) do { } while (0) 891#endif 892 893/* 894 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 895 * via the workqueue/eventd. 896 * Add the CPU number into the expiration time to minimize the possibility of 897 * the CPUs getting into lockstep and contending for the global cache chain 898 * lock. 899 */ 900static void __devinit start_cpu_timer(int cpu) 901{ 902 struct work_struct *reap_work = &per_cpu(reap_work, cpu); 903 904 /* 905 * When this gets called from do_initcalls via cpucache_init(), 906 * init_workqueues() has already run, so keventd will be setup 907 * at that time. 908 */ 909 if (keventd_up() && reap_work->func == NULL) { 910 init_reap_node(cpu); 911 INIT_WORK(reap_work, cache_reap, NULL); 912 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); 913 } 914} 915 916static struct array_cache *alloc_arraycache(int node, int entries, 917 int batchcount) 918{ 919 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 920 struct array_cache *nc = NULL; 921 922 nc = kmalloc_node(memsize, GFP_KERNEL, node); 923 if (nc) { 924 nc->avail = 0; 925 nc->limit = entries; 926 nc->batchcount = batchcount; 927 nc->touched = 0; 928 spin_lock_init(&nc->lock); 929 } 930 return nc; 931} 932 933/* 934 * Transfer objects in one arraycache to another. 935 * Locking must be handled by the caller. 936 * 937 * Return the number of entries transferred. 938 */ 939static int transfer_objects(struct array_cache *to, 940 struct array_cache *from, unsigned int max) 941{ 942 /* Figure out how many entries to transfer */ 943 int nr = min(min(from->avail, max), to->limit - to->avail); 944 945 if (!nr) 946 return 0; 947 948 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 949 sizeof(void *) *nr); 950 951 from->avail -= nr; 952 to->avail += nr; 953 to->touched = 1; 954 return nr; 955} 956 957#ifdef CONFIG_NUMA 958static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); 959static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 960 961static struct array_cache **alloc_alien_cache(int node, int limit) 962{ 963 struct array_cache **ac_ptr; 964 int memsize = sizeof(void *) * MAX_NUMNODES; 965 int i; 966 967 if (limit > 1) 968 limit = 12; 969 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); 970 if (ac_ptr) { 971 for_each_node(i) { 972 if (i == node || !node_online(i)) { 973 ac_ptr[i] = NULL; 974 continue; 975 } 976 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); 977 if (!ac_ptr[i]) { 978 for (i--; i <= 0; i--) 979 kfree(ac_ptr[i]); 980 kfree(ac_ptr); 981 return NULL; 982 } 983 } 984 } 985 return ac_ptr; 986} 987 988static void free_alien_cache(struct array_cache **ac_ptr) 989{ 990 int i; 991 992 if (!ac_ptr) 993 return; 994 for_each_node(i) 995 kfree(ac_ptr[i]); 996 kfree(ac_ptr); 997} 998 999static void __drain_alien_cache(struct kmem_cache *cachep, 1000 struct array_cache *ac, int node) 1001{ 1002 struct kmem_list3 *rl3 = cachep->nodelists[node]; 1003 1004 if (ac->avail) { 1005 spin_lock(&rl3->list_lock); 1006 /* 1007 * Stuff objects into the remote nodes shared array first. 1008 * That way we could avoid the overhead of putting the objects 1009 * into the free lists and getting them back later. 1010 */ 1011 if (rl3->shared) 1012 transfer_objects(rl3->shared, ac, ac->limit); 1013 1014 free_block(cachep, ac->entry, ac->avail, node); 1015 ac->avail = 0; 1016 spin_unlock(&rl3->list_lock); 1017 } 1018} 1019 1020/* 1021 * Called from cache_reap() to regularly drain alien caches round robin. 1022 */ 1023static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 1024{ 1025 int node = __get_cpu_var(reap_node); 1026 1027 if (l3->alien) { 1028 struct array_cache *ac = l3->alien[node]; 1029 1030 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1031 __drain_alien_cache(cachep, ac, node); 1032 spin_unlock_irq(&ac->lock); 1033 } 1034 } 1035} 1036 1037static void drain_alien_cache(struct kmem_cache *cachep, 1038 struct array_cache **alien) 1039{ 1040 int i = 0; 1041 struct array_cache *ac; 1042 unsigned long flags; 1043 1044 for_each_online_node(i) { 1045 ac = alien[i]; 1046 if (ac) { 1047 spin_lock_irqsave(&ac->lock, flags); 1048 __drain_alien_cache(cachep, ac, i); 1049 spin_unlock_irqrestore(&ac->lock, flags); 1050 } 1051 } 1052} 1053 1054static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1055{ 1056 struct slab *slabp = virt_to_slab(objp); 1057 int nodeid = slabp->nodeid; 1058 struct kmem_list3 *l3; 1059 struct array_cache *alien = NULL; 1060 1061 /* 1062 * Make sure we are not freeing a object from another node to the array 1063 * cache on this cpu. 1064 */ 1065 if (likely(slabp->nodeid == numa_node_id())) 1066 return 0; 1067 1068 l3 = cachep->nodelists[numa_node_id()]; 1069 STATS_INC_NODEFREES(cachep); 1070 if (l3->alien && l3->alien[nodeid]) { 1071 alien = l3->alien[nodeid]; 1072 spin_lock(&alien->lock); 1073 if (unlikely(alien->avail == alien->limit)) { 1074 STATS_INC_ACOVERFLOW(cachep); 1075 __drain_alien_cache(cachep, alien, nodeid); 1076 } 1077 alien->entry[alien->avail++] = objp; 1078 spin_unlock(&alien->lock); 1079 } else { 1080 spin_lock(&(cachep->nodelists[nodeid])->list_lock); 1081 free_block(cachep, &objp, 1, nodeid); 1082 spin_unlock(&(cachep->nodelists[nodeid])->list_lock); 1083 } 1084 return 1; 1085} 1086 1087#else 1088 1089#define drain_alien_cache(cachep, alien) do { } while (0) 1090#define reap_alien(cachep, l3) do { } while (0) 1091 1092static inline struct array_cache **alloc_alien_cache(int node, int limit) 1093{ 1094 return (struct array_cache **) 0x01020304ul; 1095} 1096 1097static inline void free_alien_cache(struct array_cache **ac_ptr) 1098{ 1099} 1100 1101static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1102{ 1103 return 0; 1104} 1105 1106#endif 1107 1108static int __cpuinit cpuup_callback(struct notifier_block *nfb, 1109 unsigned long action, void *hcpu) 1110{ 1111 long cpu = (long)hcpu; 1112 struct kmem_cache *cachep; 1113 struct kmem_list3 *l3 = NULL; 1114 int node = cpu_to_node(cpu); 1115 int memsize = sizeof(struct kmem_list3); 1116 1117 switch (action) { 1118 case CPU_UP_PREPARE: 1119 mutex_lock(&cache_chain_mutex); 1120 /* 1121 * We need to do this right in the beginning since 1122 * alloc_arraycache's are going to use this list. 1123 * kmalloc_node allows us to add the slab to the right 1124 * kmem_list3 and not this cpu's kmem_list3 1125 */ 1126 1127 list_for_each_entry(cachep, &cache_chain, next) { 1128 /* 1129 * Set up the size64 kmemlist for cpu before we can 1130 * begin anything. Make sure some other cpu on this 1131 * node has not already allocated this 1132 */ 1133 if (!cachep->nodelists[node]) { 1134 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1135 if (!l3) 1136 goto bad; 1137 kmem_list3_init(l3); 1138 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1139 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1140 1141 /* 1142 * The l3s don't come and go as CPUs come and 1143 * go. cache_chain_mutex is sufficient 1144 * protection here. 1145 */ 1146 cachep->nodelists[node] = l3; 1147 } 1148 1149 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1150 cachep->nodelists[node]->free_limit = 1151 (1 + nr_cpus_node(node)) * 1152 cachep->batchcount + cachep->num; 1153 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1154 } 1155 1156 /* 1157 * Now we can go ahead with allocating the shared arrays and 1158 * array caches 1159 */ 1160 list_for_each_entry(cachep, &cache_chain, next) { 1161 struct array_cache *nc; 1162 struct array_cache *shared; 1163 struct array_cache **alien; 1164 1165 nc = alloc_arraycache(node, cachep->limit, 1166 cachep->batchcount); 1167 if (!nc) 1168 goto bad; 1169 shared = alloc_arraycache(node, 1170 cachep->shared * cachep->batchcount, 1171 0xbaadf00d); 1172 if (!shared) 1173 goto bad; 1174 1175 alien = alloc_alien_cache(node, cachep->limit); 1176 if (!alien) 1177 goto bad; 1178 cachep->array[cpu] = nc; 1179 l3 = cachep->nodelists[node]; 1180 BUG_ON(!l3); 1181 1182 spin_lock_irq(&l3->list_lock); 1183 if (!l3->shared) { 1184 /* 1185 * We are serialised from CPU_DEAD or 1186 * CPU_UP_CANCELLED by the cpucontrol lock 1187 */ 1188 l3->shared = shared; 1189 shared = NULL; 1190 } 1191#ifdef CONFIG_NUMA 1192 if (!l3->alien) { 1193 l3->alien = alien; 1194 alien = NULL; 1195 } 1196#endif 1197 spin_unlock_irq(&l3->list_lock); 1198 kfree(shared); 1199 free_alien_cache(alien); 1200 } 1201 mutex_unlock(&cache_chain_mutex); 1202 break; 1203 case CPU_ONLINE: 1204 start_cpu_timer(cpu); 1205 break; 1206#ifdef CONFIG_HOTPLUG_CPU 1207 case CPU_DEAD: 1208 /* 1209 * Even if all the cpus of a node are down, we don't free the 1210 * kmem_list3 of any cache. This to avoid a race between 1211 * cpu_down, and a kmalloc allocation from another cpu for 1212 * memory from the node of the cpu going down. The list3 1213 * structure is usually allocated from kmem_cache_create() and 1214 * gets destroyed at kmem_cache_destroy(). 1215 */ 1216 /* fall thru */ 1217 case CPU_UP_CANCELED: 1218 mutex_lock(&cache_chain_mutex); 1219 list_for_each_entry(cachep, &cache_chain, next) { 1220 struct array_cache *nc; 1221 struct array_cache *shared; 1222 struct array_cache **alien; 1223 cpumask_t mask; 1224 1225 mask = node_to_cpumask(node); 1226 /* cpu is dead; no one can alloc from it. */ 1227 nc = cachep->array[cpu]; 1228 cachep->array[cpu] = NULL; 1229 l3 = cachep->nodelists[node]; 1230 1231 if (!l3) 1232 goto free_array_cache; 1233 1234 spin_lock_irq(&l3->list_lock); 1235 1236 /* Free limit for this kmem_list3 */ 1237 l3->free_limit -= cachep->batchcount; 1238 if (nc) 1239 free_block(cachep, nc->entry, nc->avail, node); 1240 1241 if (!cpus_empty(mask)) { 1242 spin_unlock_irq(&l3->list_lock); 1243 goto free_array_cache; 1244 } 1245 1246 shared = l3->shared; 1247 if (shared) { 1248 free_block(cachep, l3->shared->entry, 1249 l3->shared->avail, node); 1250 l3->shared = NULL; 1251 } 1252 1253 alien = l3->alien; 1254 l3->alien = NULL; 1255 1256 spin_unlock_irq(&l3->list_lock); 1257 1258 kfree(shared); 1259 if (alien) { 1260 drain_alien_cache(cachep, alien); 1261 free_alien_cache(alien); 1262 } 1263free_array_cache: 1264 kfree(nc); 1265 } 1266 /* 1267 * In the previous loop, all the objects were freed to 1268 * the respective cache's slabs, now we can go ahead and 1269 * shrink each nodelist to its limit. 1270 */ 1271 list_for_each_entry(cachep, &cache_chain, next) { 1272 l3 = cachep->nodelists[node]; 1273 if (!l3) 1274 continue; 1275 drain_freelist(cachep, l3, l3->free_objects); 1276 } 1277 mutex_unlock(&cache_chain_mutex); 1278 break; 1279#endif 1280 } 1281 return NOTIFY_OK; 1282bad: 1283 mutex_unlock(&cache_chain_mutex); 1284 return NOTIFY_BAD; 1285} 1286 1287static struct notifier_block __cpuinitdata cpucache_notifier = { 1288 &cpuup_callback, NULL, 0 1289}; 1290 1291/* 1292 * swap the static kmem_list3 with kmalloced memory 1293 */ 1294static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1295 int nodeid) 1296{ 1297 struct kmem_list3 *ptr; 1298 1299 BUG_ON(cachep->nodelists[nodeid] != list); 1300 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); 1301 BUG_ON(!ptr); 1302 1303 local_irq_disable(); 1304 memcpy(ptr, list, sizeof(struct kmem_list3)); 1305 /* 1306 * Do not assume that spinlocks can be initialized via memcpy: 1307 */ 1308 spin_lock_init(&ptr->list_lock); 1309 1310 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1311 cachep->nodelists[nodeid] = ptr; 1312 local_irq_enable(); 1313} 1314 1315/* 1316 * Initialisation. Called after the page allocator have been initialised and 1317 * before smp_init(). 1318 */ 1319void __init kmem_cache_init(void) 1320{ 1321 size_t left_over; 1322 struct cache_sizes *sizes; 1323 struct cache_names *names; 1324 int i; 1325 int order; 1326 1327 for (i = 0; i < NUM_INIT_LISTS; i++) { 1328 kmem_list3_init(&initkmem_list3[i]); 1329 if (i < MAX_NUMNODES) 1330 cache_cache.nodelists[i] = NULL; 1331 } 1332 1333 /* 1334 * Fragmentation resistance on low memory - only use bigger 1335 * page orders on machines with more than 32MB of memory. 1336 */ 1337 if (num_physpages > (32 << 20) >> PAGE_SHIFT) 1338 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1339 1340 /* Bootstrap is tricky, because several objects are allocated 1341 * from caches that do not exist yet: 1342 * 1) initialize the cache_cache cache: it contains the struct 1343 * kmem_cache structures of all caches, except cache_cache itself: 1344 * cache_cache is statically allocated. 1345 * Initially an __init data area is used for the head array and the 1346 * kmem_list3 structures, it's replaced with a kmalloc allocated 1347 * array at the end of the bootstrap. 1348 * 2) Create the first kmalloc cache. 1349 * The struct kmem_cache for the new cache is allocated normally. 1350 * An __init data area is used for the head array. 1351 * 3) Create the remaining kmalloc caches, with minimally sized 1352 * head arrays. 1353 * 4) Replace the __init data head arrays for cache_cache and the first 1354 * kmalloc cache with kmalloc allocated arrays. 1355 * 5) Replace the __init data for kmem_list3 for cache_cache and 1356 * the other cache's with kmalloc allocated memory. 1357 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1358 */ 1359 1360 /* 1) create the cache_cache */ 1361 INIT_LIST_HEAD(&cache_chain); 1362 list_add(&cache_cache.next, &cache_chain); 1363 cache_cache.colour_off = cache_line_size(); 1364 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1365 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; 1366 1367 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1368 cache_line_size()); 1369 1370 for (order = 0; order < MAX_ORDER; order++) { 1371 cache_estimate(order, cache_cache.buffer_size, 1372 cache_line_size(), 0, &left_over, &cache_cache.num); 1373 if (cache_cache.num) 1374 break; 1375 } 1376 BUG_ON(!cache_cache.num); 1377 cache_cache.gfporder = order; 1378 cache_cache.colour = left_over / cache_cache.colour_off; 1379 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1380 sizeof(struct slab), cache_line_size()); 1381 1382 /* 2+3) create the kmalloc caches */ 1383 sizes = malloc_sizes; 1384 names = cache_names; 1385 1386 /* 1387 * Initialize the caches that provide memory for the array cache and the 1388 * kmem_list3 structures first. Without this, further allocations will 1389 * bug. 1390 */ 1391 1392 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1393 sizes[INDEX_AC].cs_size, 1394 ARCH_KMALLOC_MINALIGN, 1395 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1396 NULL, NULL); 1397 1398 if (INDEX_AC != INDEX_L3) { 1399 sizes[INDEX_L3].cs_cachep = 1400 kmem_cache_create(names[INDEX_L3].name, 1401 sizes[INDEX_L3].cs_size, 1402 ARCH_KMALLOC_MINALIGN, 1403 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1404 NULL, NULL); 1405 } 1406 1407 slab_early_init = 0; 1408 1409 while (sizes->cs_size != ULONG_MAX) { 1410 /* 1411 * For performance, all the general caches are L1 aligned. 1412 * This should be particularly beneficial on SMP boxes, as it 1413 * eliminates "false sharing". 1414 * Note for systems short on memory removing the alignment will 1415 * allow tighter packing of the smaller caches. 1416 */ 1417 if (!sizes->cs_cachep) { 1418 sizes->cs_cachep = kmem_cache_create(names->name, 1419 sizes->cs_size, 1420 ARCH_KMALLOC_MINALIGN, 1421 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1422 NULL, NULL); 1423 } 1424 init_lock_keys(sizes); 1425 1426 sizes->cs_dmacachep = kmem_cache_create(names->name_dma, 1427 sizes->cs_size, 1428 ARCH_KMALLOC_MINALIGN, 1429 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1430 SLAB_PANIC, 1431 NULL, NULL); 1432 sizes++; 1433 names++; 1434 } 1435 /* 4) Replace the bootstrap head arrays */ 1436 { 1437 struct array_cache *ptr; 1438 1439 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1440 1441 local_irq_disable(); 1442 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1443 memcpy(ptr, cpu_cache_get(&cache_cache), 1444 sizeof(struct arraycache_init)); 1445 /* 1446 * Do not assume that spinlocks can be initialized via memcpy: 1447 */ 1448 spin_lock_init(&ptr->lock); 1449 1450 cache_cache.array[smp_processor_id()] = ptr; 1451 local_irq_enable(); 1452 1453 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1454 1455 local_irq_disable(); 1456 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1457 != &initarray_generic.cache); 1458 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1459 sizeof(struct arraycache_init)); 1460 /* 1461 * Do not assume that spinlocks can be initialized via memcpy: 1462 */ 1463 spin_lock_init(&ptr->lock); 1464 1465 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1466 ptr; 1467 local_irq_enable(); 1468 } 1469 /* 5) Replace the bootstrap kmem_list3's */ 1470 { 1471 int node; 1472 /* Replace the static kmem_list3 structures for the boot cpu */ 1473 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], 1474 numa_node_id()); 1475 1476 for_each_online_node(node) { 1477 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1478 &initkmem_list3[SIZE_AC + node], node); 1479 1480 if (INDEX_AC != INDEX_L3) { 1481 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1482 &initkmem_list3[SIZE_L3 + node], 1483 node); 1484 } 1485 } 1486 } 1487 1488 /* 6) resize the head arrays to their final sizes */ 1489 { 1490 struct kmem_cache *cachep; 1491 mutex_lock(&cache_chain_mutex); 1492 list_for_each_entry(cachep, &cache_chain, next) 1493 enable_cpucache(cachep); 1494 mutex_unlock(&cache_chain_mutex); 1495 } 1496 1497 /* Done! */ 1498 g_cpucache_up = FULL; 1499 1500 /* 1501 * Register a cpu startup notifier callback that initializes 1502 * cpu_cache_get for all new cpus 1503 */ 1504 register_cpu_notifier(&cpucache_notifier); 1505 1506 /* 1507 * The reap timers are started later, with a module init call: That part 1508 * of the kernel is not yet operational. 1509 */ 1510} 1511 1512static int __init cpucache_init(void) 1513{ 1514 int cpu; 1515 1516 /* 1517 * Register the timers that return unneeded pages to the page allocator 1518 */ 1519 for_each_online_cpu(cpu) 1520 start_cpu_timer(cpu); 1521 return 0; 1522} 1523__initcall(cpucache_init); 1524 1525/* 1526 * Interface to system's page allocator. No need to hold the cache-lock. 1527 * 1528 * If we requested dmaable memory, we will get it. Even if we 1529 * did not request dmaable memory, we might get it, but that 1530 * would be relatively rare and ignorable. 1531 */ 1532static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1533{ 1534 struct page *page; 1535 int nr_pages; 1536 int i; 1537 1538#ifndef CONFIG_MMU 1539 /* 1540 * Nommu uses slab's for process anonymous memory allocations, and thus 1541 * requires __GFP_COMP to properly refcount higher order allocations 1542 */ 1543 flags |= __GFP_COMP; 1544#endif 1545 flags |= cachep->gfpflags; 1546 1547 page = alloc_pages_node(nodeid, flags, cachep->gfporder); 1548 if (!page) 1549 return NULL; 1550 1551 nr_pages = (1 << cachep->gfporder); 1552 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1553 atomic_add(nr_pages, &slab_reclaim_pages); 1554 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages); 1555 for (i = 0; i < nr_pages; i++) 1556 __SetPageSlab(page + i); 1557 return page_address(page); 1558} 1559 1560/* 1561 * Interface to system's page release. 1562 */ 1563static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1564{ 1565 unsigned long i = (1 << cachep->gfporder); 1566 struct page *page = virt_to_page(addr); 1567 const unsigned long nr_freed = i; 1568 1569 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed); 1570 while (i--) { 1571 BUG_ON(!PageSlab(page)); 1572 __ClearPageSlab(page); 1573 page++; 1574 } 1575 if (current->reclaim_state) 1576 current->reclaim_state->reclaimed_slab += nr_freed; 1577 free_pages((unsigned long)addr, cachep->gfporder); 1578 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1579 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); 1580} 1581 1582static void kmem_rcu_free(struct rcu_head *head) 1583{ 1584 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1585 struct kmem_cache *cachep = slab_rcu->cachep; 1586 1587 kmem_freepages(cachep, slab_rcu->addr); 1588 if (OFF_SLAB(cachep)) 1589 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1590} 1591 1592#if DEBUG 1593 1594#ifdef CONFIG_DEBUG_PAGEALLOC 1595static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1596 unsigned long caller) 1597{ 1598 int size = obj_size(cachep); 1599 1600 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1601 1602 if (size < 5 * sizeof(unsigned long)) 1603 return; 1604 1605 *addr++ = 0x12345678; 1606 *addr++ = caller; 1607 *addr++ = smp_processor_id(); 1608 size -= 3 * sizeof(unsigned long); 1609 { 1610 unsigned long *sptr = &caller; 1611 unsigned long svalue; 1612 1613 while (!kstack_end(sptr)) { 1614 svalue = *sptr++; 1615 if (kernel_text_address(svalue)) { 1616 *addr++ = svalue; 1617 size -= sizeof(unsigned long); 1618 if (size <= sizeof(unsigned long)) 1619 break; 1620 } 1621 } 1622 1623 } 1624 *addr++ = 0x87654321; 1625} 1626#endif 1627 1628static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1629{ 1630 int size = obj_size(cachep); 1631 addr = &((char *)addr)[obj_offset(cachep)]; 1632 1633 memset(addr, val, size); 1634 *(unsigned char *)(addr + size - 1) = POISON_END; 1635} 1636 1637static void dump_line(char *data, int offset, int limit) 1638{ 1639 int i; 1640 printk(KERN_ERR "%03x:", offset); 1641 for (i = 0; i < limit; i++) 1642 printk(" %02x", (unsigned char)data[offset + i]); 1643 printk("\n"); 1644} 1645#endif 1646 1647#if DEBUG 1648 1649static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1650{ 1651 int i, size; 1652 char *realobj; 1653 1654 if (cachep->flags & SLAB_RED_ZONE) { 1655 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", 1656 *dbg_redzone1(cachep, objp), 1657 *dbg_redzone2(cachep, objp)); 1658 } 1659 1660 if (cachep->flags & SLAB_STORE_USER) { 1661 printk(KERN_ERR "Last user: [<%p>]", 1662 *dbg_userword(cachep, objp)); 1663 print_symbol("(%s)", 1664 (unsigned long)*dbg_userword(cachep, objp)); 1665 printk("\n"); 1666 } 1667 realobj = (char *)objp + obj_offset(cachep); 1668 size = obj_size(cachep); 1669 for (i = 0; i < size && lines; i += 16, lines--) { 1670 int limit; 1671 limit = 16; 1672 if (i + limit > size) 1673 limit = size - i; 1674 dump_line(realobj, i, limit); 1675 } 1676} 1677 1678static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1679{ 1680 char *realobj; 1681 int size, i; 1682 int lines = 0; 1683 1684 realobj = (char *)objp + obj_offset(cachep); 1685 size = obj_size(cachep); 1686 1687 for (i = 0; i < size; i++) { 1688 char exp = POISON_FREE; 1689 if (i == size - 1) 1690 exp = POISON_END; 1691 if (realobj[i] != exp) { 1692 int limit; 1693 /* Mismatch ! */ 1694 /* Print header */ 1695 if (lines == 0) { 1696 printk(KERN_ERR 1697 "Slab corruption: start=%p, len=%d\n", 1698 realobj, size); 1699 print_objinfo(cachep, objp, 0); 1700 } 1701 /* Hexdump the affected line */ 1702 i = (i / 16) * 16; 1703 limit = 16; 1704 if (i + limit > size) 1705 limit = size - i; 1706 dump_line(realobj, i, limit); 1707 i += 16; 1708 lines++; 1709 /* Limit to 5 lines */ 1710 if (lines > 5) 1711 break; 1712 } 1713 } 1714 if (lines != 0) { 1715 /* Print some data about the neighboring objects, if they 1716 * exist: 1717 */ 1718 struct slab *slabp = virt_to_slab(objp); 1719 unsigned int objnr; 1720 1721 objnr = obj_to_index(cachep, slabp, objp); 1722 if (objnr) { 1723 objp = index_to_obj(cachep, slabp, objnr - 1); 1724 realobj = (char *)objp + obj_offset(cachep); 1725 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1726 realobj, size); 1727 print_objinfo(cachep, objp, 2); 1728 } 1729 if (objnr + 1 < cachep->num) { 1730 objp = index_to_obj(cachep, slabp, objnr + 1); 1731 realobj = (char *)objp + obj_offset(cachep); 1732 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1733 realobj, size); 1734 print_objinfo(cachep, objp, 2); 1735 } 1736 } 1737} 1738#endif 1739 1740#if DEBUG 1741/** 1742 * slab_destroy_objs - destroy a slab and its objects 1743 * @cachep: cache pointer being destroyed 1744 * @slabp: slab pointer being destroyed 1745 * 1746 * Call the registered destructor for each object in a slab that is being 1747 * destroyed. 1748 */ 1749static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1750{ 1751 int i; 1752 for (i = 0; i < cachep->num; i++) { 1753 void *objp = index_to_obj(cachep, slabp, i); 1754 1755 if (cachep->flags & SLAB_POISON) { 1756#ifdef CONFIG_DEBUG_PAGEALLOC 1757 if (cachep->buffer_size % PAGE_SIZE == 0 && 1758 OFF_SLAB(cachep)) 1759 kernel_map_pages(virt_to_page(objp), 1760 cachep->buffer_size / PAGE_SIZE, 1); 1761 else 1762 check_poison_obj(cachep, objp); 1763#else 1764 check_poison_obj(cachep, objp); 1765#endif 1766 } 1767 if (cachep->flags & SLAB_RED_ZONE) { 1768 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1769 slab_error(cachep, "start of a freed object " 1770 "was overwritten"); 1771 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1772 slab_error(cachep, "end of a freed object " 1773 "was overwritten"); 1774 } 1775 if (cachep->dtor && !(cachep->flags & SLAB_POISON)) 1776 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); 1777 } 1778} 1779#else 1780static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1781{ 1782 if (cachep->dtor) { 1783 int i; 1784 for (i = 0; i < cachep->num; i++) { 1785 void *objp = index_to_obj(cachep, slabp, i); 1786 (cachep->dtor) (objp, cachep, 0); 1787 } 1788 } 1789} 1790#endif 1791 1792/** 1793 * slab_destroy - destroy and release all objects in a slab 1794 * @cachep: cache pointer being destroyed 1795 * @slabp: slab pointer being destroyed 1796 * 1797 * Destroy all the objs in a slab, and release the mem back to the system. 1798 * Before calling the slab must have been unlinked from the cache. The 1799 * cache-lock is not held/needed. 1800 */ 1801static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1802{ 1803 void *addr = slabp->s_mem - slabp->colouroff; 1804 1805 slab_destroy_objs(cachep, slabp); 1806 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1807 struct slab_rcu *slab_rcu; 1808 1809 slab_rcu = (struct slab_rcu *)slabp; 1810 slab_rcu->cachep = cachep; 1811 slab_rcu->addr = addr; 1812 call_rcu(&slab_rcu->head, kmem_rcu_free); 1813 } else { 1814 kmem_freepages(cachep, addr); 1815 if (OFF_SLAB(cachep)) 1816 kmem_cache_free(cachep->slabp_cache, slabp); 1817 } 1818} 1819 1820/* 1821 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1822 * size of kmem_list3. 1823 */ 1824static void set_up_list3s(struct kmem_cache *cachep, int index) 1825{ 1826 int node; 1827 1828 for_each_online_node(node) { 1829 cachep->nodelists[node] = &initkmem_list3[index + node]; 1830 cachep->nodelists[node]->next_reap = jiffies + 1831 REAPTIMEOUT_LIST3 + 1832 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1833 } 1834} 1835 1836/** 1837 * calculate_slab_order - calculate size (page order) of slabs 1838 * @cachep: pointer to the cache that is being created 1839 * @size: size of objects to be created in this cache. 1840 * @align: required alignment for the objects. 1841 * @flags: slab allocation flags 1842 * 1843 * Also calculates the number of objects per slab. 1844 * 1845 * This could be made much more intelligent. For now, try to avoid using 1846 * high order pages for slabs. When the gfp() functions are more friendly 1847 * towards high-order requests, this should be changed. 1848 */ 1849static size_t calculate_slab_order(struct kmem_cache *cachep, 1850 size_t size, size_t align, unsigned long flags) 1851{ 1852 unsigned long offslab_limit; 1853 size_t left_over = 0; 1854 int gfporder; 1855 1856 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { 1857 unsigned int num; 1858 size_t remainder; 1859 1860 cache_estimate(gfporder, size, align, flags, &remainder, &num); 1861 if (!num) 1862 continue; 1863 1864 if (flags & CFLGS_OFF_SLAB) { 1865 /* 1866 * Max number of objs-per-slab for caches which 1867 * use off-slab slabs. Needed to avoid a possible 1868 * looping condition in cache_grow(). 1869 */ 1870 offslab_limit = size - sizeof(struct slab); 1871 offslab_limit /= sizeof(kmem_bufctl_t); 1872 1873 if (num > offslab_limit) 1874 break; 1875 } 1876 1877 /* Found something acceptable - save it away */ 1878 cachep->num = num; 1879 cachep->gfporder = gfporder; 1880 left_over = remainder; 1881 1882 /* 1883 * A VFS-reclaimable slab tends to have most allocations 1884 * as GFP_NOFS and we really don't want to have to be allocating 1885 * higher-order pages when we are unable to shrink dcache. 1886 */ 1887 if (flags & SLAB_RECLAIM_ACCOUNT) 1888 break; 1889 1890 /* 1891 * Large number of objects is good, but very large slabs are 1892 * currently bad for the gfp()s. 1893 */ 1894 if (gfporder >= slab_break_gfp_order) 1895 break; 1896 1897 /* 1898 * Acceptable internal fragmentation? 1899 */ 1900 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 1901 break; 1902 } 1903 return left_over; 1904} 1905 1906static void setup_cpu_cache(struct kmem_cache *cachep) 1907{ 1908 if (g_cpucache_up == FULL) { 1909 enable_cpucache(cachep); 1910 return; 1911 } 1912 if (g_cpucache_up == NONE) { 1913 /* 1914 * Note: the first kmem_cache_create must create the cache 1915 * that's used by kmalloc(24), otherwise the creation of 1916 * further caches will BUG(). 1917 */ 1918 cachep->array[smp_processor_id()] = &initarray_generic.cache; 1919 1920 /* 1921 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 1922 * the first cache, then we need to set up all its list3s, 1923 * otherwise the creation of further caches will BUG(). 1924 */ 1925 set_up_list3s(cachep, SIZE_AC); 1926 if (INDEX_AC == INDEX_L3) 1927 g_cpucache_up = PARTIAL_L3; 1928 else 1929 g_cpucache_up = PARTIAL_AC; 1930 } else { 1931 cachep->array[smp_processor_id()] = 1932 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1933 1934 if (g_cpucache_up == PARTIAL_AC) { 1935 set_up_list3s(cachep, SIZE_L3); 1936 g_cpucache_up = PARTIAL_L3; 1937 } else { 1938 int node; 1939 for_each_online_node(node) { 1940 cachep->nodelists[node] = 1941 kmalloc_node(sizeof(struct kmem_list3), 1942 GFP_KERNEL, node); 1943 BUG_ON(!cachep->nodelists[node]); 1944 kmem_list3_init(cachep->nodelists[node]); 1945 } 1946 } 1947 } 1948 cachep->nodelists[numa_node_id()]->next_reap = 1949 jiffies + REAPTIMEOUT_LIST3 + 1950 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1951 1952 cpu_cache_get(cachep)->avail = 0; 1953 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 1954 cpu_cache_get(cachep)->batchcount = 1; 1955 cpu_cache_get(cachep)->touched = 0; 1956 cachep->batchcount = 1; 1957 cachep->limit = BOOT_CPUCACHE_ENTRIES; 1958} 1959 1960/** 1961 * kmem_cache_create - Create a cache. 1962 * @name: A string which is used in /proc/slabinfo to identify this cache. 1963 * @size: The size of objects to be created in this cache. 1964 * @align: The required alignment for the objects. 1965 * @flags: SLAB flags 1966 * @ctor: A constructor for the objects. 1967 * @dtor: A destructor for the objects. 1968 * 1969 * Returns a ptr to the cache on success, NULL on failure. 1970 * Cannot be called within a int, but can be interrupted. 1971 * The @ctor is run when new pages are allocated by the cache 1972 * and the @dtor is run before the pages are handed back. 1973 * 1974 * @name must be valid until the cache is destroyed. This implies that 1975 * the module calling this has to destroy the cache before getting unloaded. 1976 * 1977 * The flags are 1978 * 1979 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 1980 * to catch references to uninitialised memory. 1981 * 1982 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 1983 * for buffer overruns. 1984 * 1985 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 1986 * cacheline. This can be beneficial if you're counting cycles as closely 1987 * as davem. 1988 */ 1989struct kmem_cache * 1990kmem_cache_create (const char *name, size_t size, size_t align, 1991 unsigned long flags, 1992 void (*ctor)(void*, struct kmem_cache *, unsigned long), 1993 void (*dtor)(void*, struct kmem_cache *, unsigned long)) 1994{ 1995 size_t left_over, slab_size, ralign; 1996 struct kmem_cache *cachep = NULL, *pc; 1997 1998 /* 1999 * Sanity checks... these are all serious usage bugs. 2000 */ 2001 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 2002 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { 2003 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, 2004 name); 2005 BUG(); 2006 } 2007 2008 /* 2009 * Prevent CPUs from coming and going. 2010 * lock_cpu_hotplug() nests outside cache_chain_mutex 2011 */ 2012 lock_cpu_hotplug(); 2013 2014 mutex_lock(&cache_chain_mutex); 2015 2016 list_for_each_entry(pc, &cache_chain, next) { 2017 mm_segment_t old_fs = get_fs(); 2018 char tmp; 2019 int res; 2020 2021 /* 2022 * This happens when the module gets unloaded and doesn't 2023 * destroy its slab cache and no-one else reuses the vmalloc 2024 * area of the module. Print a warning. 2025 */ 2026 set_fs(KERNEL_DS); 2027 res = __get_user(tmp, pc->name); 2028 set_fs(old_fs); 2029 if (res) { 2030 printk("SLAB: cache with size %d has lost its name\n", 2031 pc->buffer_size); 2032 continue; 2033 } 2034 2035 if (!strcmp(pc->name, name)) { 2036 printk("kmem_cache_create: duplicate cache %s\n", name); 2037 dump_stack(); 2038 goto oops; 2039 } 2040 } 2041 2042#if DEBUG 2043 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 2044 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { 2045 /* No constructor, but inital state check requested */ 2046 printk(KERN_ERR "%s: No con, but init state check " 2047 "requested - %s\n", __FUNCTION__, name); 2048 flags &= ~SLAB_DEBUG_INITIAL; 2049 } 2050#if FORCED_DEBUG 2051 /* 2052 * Enable redzoning and last user accounting, except for caches with 2053 * large objects, if the increased size would increase the object size 2054 * above the next power of two: caches with object sizes just above a 2055 * power of two have a significant amount of internal fragmentation. 2056 */ 2057 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) 2058 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2059 if (!(flags & SLAB_DESTROY_BY_RCU)) 2060 flags |= SLAB_POISON; 2061#endif 2062 if (flags & SLAB_DESTROY_BY_RCU) 2063 BUG_ON(flags & SLAB_POISON); 2064#endif 2065 if (flags & SLAB_DESTROY_BY_RCU) 2066 BUG_ON(dtor); 2067 2068 /* 2069 * Always checks flags, a caller might be expecting debug support which 2070 * isn't available. 2071 */ 2072 BUG_ON(flags & ~CREATE_MASK); 2073 2074 /* 2075 * Check that size is in terms of words. This is needed to avoid 2076 * unaligned accesses for some archs when redzoning is used, and makes 2077 * sure any on-slab bufctl's are also correctly aligned. 2078 */ 2079 if (size & (BYTES_PER_WORD - 1)) { 2080 size += (BYTES_PER_WORD - 1); 2081 size &= ~(BYTES_PER_WORD - 1); 2082 } 2083 2084 /* calculate the final buffer alignment: */ 2085 2086 /* 1) arch recommendation: can be overridden for debug */ 2087 if (flags & SLAB_HWCACHE_ALIGN) { 2088 /* 2089 * Default alignment: as specified by the arch code. Except if 2090 * an object is really small, then squeeze multiple objects into 2091 * one cacheline. 2092 */ 2093 ralign = cache_line_size(); 2094 while (size <= ralign / 2) 2095 ralign /= 2; 2096 } else { 2097 ralign = BYTES_PER_WORD; 2098 } 2099 2100 /* 2101 * Redzoning and user store require word alignment. Note this will be 2102 * overridden by architecture or caller mandated alignment if either 2103 * is greater than BYTES_PER_WORD. 2104 */ 2105 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) 2106 ralign = BYTES_PER_WORD; 2107 2108 /* 2) arch mandated alignment: disables debug if necessary */ 2109 if (ralign < ARCH_SLAB_MINALIGN) { 2110 ralign = ARCH_SLAB_MINALIGN; 2111 if (ralign > BYTES_PER_WORD) 2112 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2113 } 2114 /* 3) caller mandated alignment: disables debug if necessary */ 2115 if (ralign < align) { 2116 ralign = align; 2117 if (ralign > BYTES_PER_WORD) 2118 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2119 } 2120 /* 2121 * 4) Store it. 2122 */ 2123 align = ralign; 2124 2125 /* Get cache's description obj. */ 2126 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); 2127 if (!cachep) 2128 goto oops; 2129 2130#if DEBUG 2131 cachep->obj_size = size; 2132 2133 /* 2134 * Both debugging options require word-alignment which is calculated 2135 * into align above. 2136 */ 2137 if (flags & SLAB_RED_ZONE) { 2138 /* add space for red zone words */ 2139 cachep->obj_offset += BYTES_PER_WORD; 2140 size += 2 * BYTES_PER_WORD; 2141 } 2142 if (flags & SLAB_STORE_USER) { 2143 /* user store requires one word storage behind the end of 2144 * the real object. 2145 */ 2146 size += BYTES_PER_WORD; 2147 } 2148#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2149 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2150 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { 2151 cachep->obj_offset += PAGE_SIZE - size; 2152 size = PAGE_SIZE; 2153 } 2154#endif 2155#endif 2156 2157 /* 2158 * Determine if the slab management is 'on' or 'off' slab. 2159 * (bootstrapping cannot cope with offslab caches so don't do 2160 * it too early on.) 2161 */ 2162 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) 2163 /* 2164 * Size is large, assume best to place the slab management obj 2165 * off-slab (should allow better packing of objs). 2166 */ 2167 flags |= CFLGS_OFF_SLAB; 2168 2169 size = ALIGN(size, align); 2170 2171 left_over = calculate_slab_order(cachep, size, align, flags); 2172 2173 if (!cachep->num) { 2174 printk("kmem_cache_create: couldn't create cache %s.\n", name); 2175 kmem_cache_free(&cache_cache, cachep); 2176 cachep = NULL; 2177 goto oops; 2178 } 2179 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2180 + sizeof(struct slab), align); 2181 2182 /* 2183 * If the slab has been placed off-slab, and we have enough space then 2184 * move it on-slab. This is at the expense of any extra colouring. 2185 */ 2186 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2187 flags &= ~CFLGS_OFF_SLAB; 2188 left_over -= slab_size; 2189 } 2190 2191 if (flags & CFLGS_OFF_SLAB) { 2192 /* really off slab. No need for manual alignment */ 2193 slab_size = 2194 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2195 } 2196 2197 cachep->colour_off = cache_line_size(); 2198 /* Offset must be a multiple of the alignment. */ 2199 if (cachep->colour_off < align) 2200 cachep->colour_off = align; 2201 cachep->colour = left_over / cachep->colour_off; 2202 cachep->slab_size = slab_size; 2203 cachep->flags = flags; 2204 cachep->gfpflags = 0; 2205 if (flags & SLAB_CACHE_DMA) 2206 cachep->gfpflags |= GFP_DMA; 2207 cachep->buffer_size = size; 2208 2209 if (flags & CFLGS_OFF_SLAB) { 2210 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2211 /* 2212 * This is a possibility for one of the malloc_sizes caches. 2213 * But since we go off slab only for object size greater than 2214 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, 2215 * this should not happen at all. 2216 * But leave a BUG_ON for some lucky dude. 2217 */ 2218 BUG_ON(!cachep->slabp_cache); 2219 } 2220 cachep->ctor = ctor; 2221 cachep->dtor = dtor; 2222 cachep->name = name; 2223 2224 2225 setup_cpu_cache(cachep); 2226 2227 /* cache setup completed, link it into the list */ 2228 list_add(&cachep->next, &cache_chain); 2229oops: 2230 if (!cachep && (flags & SLAB_PANIC)) 2231 panic("kmem_cache_create(): failed to create slab `%s'\n", 2232 name); 2233 mutex_unlock(&cache_chain_mutex); 2234 unlock_cpu_hotplug(); 2235 return cachep; 2236} 2237EXPORT_SYMBOL(kmem_cache_create); 2238 2239#if DEBUG 2240static void check_irq_off(void) 2241{ 2242 BUG_ON(!irqs_disabled()); 2243} 2244 2245static void check_irq_on(void) 2246{ 2247 BUG_ON(irqs_disabled()); 2248} 2249 2250static void check_spinlock_acquired(struct kmem_cache *cachep) 2251{ 2252#ifdef CONFIG_SMP 2253 check_irq_off(); 2254 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); 2255#endif 2256} 2257 2258static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2259{ 2260#ifdef CONFIG_SMP 2261 check_irq_off(); 2262 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2263#endif 2264} 2265 2266#else 2267#define check_irq_off() do { } while(0) 2268#define check_irq_on() do { } while(0) 2269#define check_spinlock_acquired(x) do { } while(0) 2270#define check_spinlock_acquired_node(x, y) do { } while(0) 2271#endif 2272 2273static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2274 struct array_cache *ac, 2275 int force, int node); 2276 2277static void do_drain(void *arg) 2278{ 2279 struct kmem_cache *cachep = arg; 2280 struct array_cache *ac; 2281 int node = numa_node_id(); 2282 2283 check_irq_off(); 2284 ac = cpu_cache_get(cachep); 2285 spin_lock(&cachep->nodelists[node]->list_lock); 2286 free_block(cachep, ac->entry, ac->avail, node); 2287 spin_unlock(&cachep->nodelists[node]->list_lock); 2288 ac->avail = 0; 2289} 2290 2291static void drain_cpu_caches(struct kmem_cache *cachep) 2292{ 2293 struct kmem_list3 *l3; 2294 int node; 2295 2296 on_each_cpu(do_drain, cachep, 1, 1); 2297 check_irq_on(); 2298 for_each_online_node(node) { 2299 l3 = cachep->nodelists[node]; 2300 if (l3 && l3->alien) 2301 drain_alien_cache(cachep, l3->alien); 2302 } 2303 2304 for_each_online_node(node) { 2305 l3 = cachep->nodelists[node]; 2306 if (l3) 2307 drain_array(cachep, l3, l3->shared, 1, node); 2308 } 2309} 2310 2311/* 2312 * Remove slabs from the list of free slabs. 2313 * Specify the number of slabs to drain in tofree. 2314 * 2315 * Returns the actual number of slabs released. 2316 */ 2317static int drain_freelist(struct kmem_cache *cache, 2318 struct kmem_list3 *l3, int tofree) 2319{ 2320 struct list_head *p; 2321 int nr_freed; 2322 struct slab *slabp; 2323 2324 nr_freed = 0; 2325 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { 2326 2327 spin_lock_irq(&l3->list_lock); 2328 p = l3->slabs_free.prev; 2329 if (p == &l3->slabs_free) { 2330 spin_unlock_irq(&l3->list_lock); 2331 goto out; 2332 } 2333 2334 slabp = list_entry(p, struct slab, list); 2335#if DEBUG 2336 BUG_ON(slabp->inuse); 2337#endif 2338 list_del(&slabp->list); 2339 /* 2340 * Safe to drop the lock. The slab is no longer linked 2341 * to the cache. 2342 */ 2343 l3->free_objects -= cache->num; 2344 spin_unlock_irq(&l3->list_lock); 2345 slab_destroy(cache, slabp); 2346 nr_freed++; 2347 } 2348out: 2349 return nr_freed; 2350} 2351 2352static int __cache_shrink(struct kmem_cache *cachep) 2353{ 2354 int ret = 0, i = 0; 2355 struct kmem_list3 *l3; 2356 2357 drain_cpu_caches(cachep); 2358 2359 check_irq_on(); 2360 for_each_online_node(i) { 2361 l3 = cachep->nodelists[i]; 2362 if (!l3) 2363 continue; 2364 2365 drain_freelist(cachep, l3, l3->free_objects); 2366 2367 ret += !list_empty(&l3->slabs_full) || 2368 !list_empty(&l3->slabs_partial); 2369 } 2370 return (ret ? 1 : 0); 2371} 2372 2373/** 2374 * kmem_cache_shrink - Shrink a cache. 2375 * @cachep: The cache to shrink. 2376 * 2377 * Releases as many slabs as possible for a cache. 2378 * To help debugging, a zero exit status indicates all slabs were released. 2379 */ 2380int kmem_cache_shrink(struct kmem_cache *cachep) 2381{ 2382 BUG_ON(!cachep || in_interrupt()); 2383 2384 return __cache_shrink(cachep); 2385} 2386EXPORT_SYMBOL(kmem_cache_shrink); 2387 2388/** 2389 * kmem_cache_destroy - delete a cache 2390 * @cachep: the cache to destroy 2391 * 2392 * Remove a struct kmem_cache object from the slab cache. 2393 * Returns 0 on success. 2394 * 2395 * It is expected this function will be called by a module when it is 2396 * unloaded. This will remove the cache completely, and avoid a duplicate 2397 * cache being allocated each time a module is loaded and unloaded, if the 2398 * module doesn't have persistent in-kernel storage across loads and unloads. 2399 * 2400 * The cache must be empty before calling this function. 2401 * 2402 * The caller must guarantee that noone will allocate memory from the cache 2403 * during the kmem_cache_destroy(). 2404 */ 2405int kmem_cache_destroy(struct kmem_cache *cachep) 2406{ 2407 int i; 2408 struct kmem_list3 *l3; 2409 2410 BUG_ON(!cachep || in_interrupt()); 2411 2412 /* Don't let CPUs to come and go */ 2413 lock_cpu_hotplug(); 2414 2415 /* Find the cache in the chain of caches. */ 2416 mutex_lock(&cache_chain_mutex); 2417 /* 2418 * the chain is never empty, cache_cache is never destroyed 2419 */ 2420 list_del(&cachep->next); 2421 mutex_unlock(&cache_chain_mutex); 2422 2423 if (__cache_shrink(cachep)) { 2424 slab_error(cachep, "Can't free all objects"); 2425 mutex_lock(&cache_chain_mutex); 2426 list_add(&cachep->next, &cache_chain); 2427 mutex_unlock(&cache_chain_mutex); 2428 unlock_cpu_hotplug(); 2429 return 1; 2430 } 2431 2432 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2433 synchronize_rcu(); 2434 2435 for_each_online_cpu(i) 2436 kfree(cachep->array[i]); 2437 2438 /* NUMA: free the list3 structures */ 2439 for_each_online_node(i) { 2440 l3 = cachep->nodelists[i]; 2441 if (l3) { 2442 kfree(l3->shared); 2443 free_alien_cache(l3->alien); 2444 kfree(l3); 2445 } 2446 } 2447 kmem_cache_free(&cache_cache, cachep); 2448 unlock_cpu_hotplug(); 2449 return 0; 2450} 2451EXPORT_SYMBOL(kmem_cache_destroy); 2452 2453/* 2454 * Get the memory for a slab management obj. 2455 * For a slab cache when the slab descriptor is off-slab, slab descriptors 2456 * always come from malloc_sizes caches. The slab descriptor cannot 2457 * come from the same cache which is getting created because, 2458 * when we are searching for an appropriate cache for these 2459 * descriptors in kmem_cache_create, we search through the malloc_sizes array. 2460 * If we are creating a malloc_sizes cache here it would not be visible to 2461 * kmem_find_general_cachep till the initialization is complete. 2462 * Hence we cannot have slabp_cache same as the original cache. 2463 */ 2464static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2465 int colour_off, gfp_t local_flags, 2466 int nodeid) 2467{ 2468 struct slab *slabp; 2469 2470 if (OFF_SLAB(cachep)) { 2471 /* Slab management obj is off-slab. */ 2472 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2473 local_flags, nodeid); 2474 if (!slabp) 2475 return NULL; 2476 } else { 2477 slabp = objp + colour_off; 2478 colour_off += cachep->slab_size; 2479 } 2480 slabp->inuse = 0; 2481 slabp->colouroff = colour_off; 2482 slabp->s_mem = objp + colour_off; 2483 slabp->nodeid = nodeid; 2484 return slabp; 2485} 2486 2487static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2488{ 2489 return (kmem_bufctl_t *) (slabp + 1); 2490} 2491 2492static void cache_init_objs(struct kmem_cache *cachep, 2493 struct slab *slabp, unsigned long ctor_flags) 2494{ 2495 int i; 2496 2497 for (i = 0; i < cachep->num; i++) { 2498 void *objp = index_to_obj(cachep, slabp, i); 2499#if DEBUG 2500 /* need to poison the objs? */ 2501 if (cachep->flags & SLAB_POISON) 2502 poison_obj(cachep, objp, POISON_FREE); 2503 if (cachep->flags & SLAB_STORE_USER) 2504 *dbg_userword(cachep, objp) = NULL; 2505 2506 if (cachep->flags & SLAB_RED_ZONE) { 2507 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2508 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2509 } 2510 /* 2511 * Constructors are not allowed to allocate memory from the same 2512 * cache which they are a constructor for. Otherwise, deadlock. 2513 * They must also be threaded. 2514 */ 2515 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2516 cachep->ctor(objp + obj_offset(cachep), cachep, 2517 ctor_flags); 2518 2519 if (cachep->flags & SLAB_RED_ZONE) { 2520 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2521 slab_error(cachep, "constructor overwrote the" 2522 " end of an object"); 2523 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2524 slab_error(cachep, "constructor overwrote the" 2525 " start of an object"); 2526 } 2527 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2528 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2529 kernel_map_pages(virt_to_page(objp), 2530 cachep->buffer_size / PAGE_SIZE, 0); 2531#else 2532 if (cachep->ctor) 2533 cachep->ctor(objp, cachep, ctor_flags); 2534#endif 2535 slab_bufctl(slabp)[i] = i + 1; 2536 } 2537 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2538 slabp->free = 0; 2539} 2540 2541static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2542{ 2543 if (flags & SLAB_DMA) 2544 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2545 else 2546 BUG_ON(cachep->gfpflags & GFP_DMA); 2547} 2548 2549static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2550 int nodeid) 2551{ 2552 void *objp = index_to_obj(cachep, slabp, slabp->free); 2553 kmem_bufctl_t next; 2554 2555 slabp->inuse++; 2556 next = slab_bufctl(slabp)[slabp->free]; 2557#if DEBUG 2558 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2559 WARN_ON(slabp->nodeid != nodeid); 2560#endif 2561 slabp->free = next; 2562 2563 return objp; 2564} 2565 2566static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2567 void *objp, int nodeid) 2568{ 2569 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2570 2571#if DEBUG 2572 /* Verify that the slab belongs to the intended node */ 2573 WARN_ON(slabp->nodeid != nodeid); 2574 2575 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2576 printk(KERN_ERR "slab: double free detected in cache " 2577 "'%s', objp %p\n", cachep->name, objp); 2578 BUG(); 2579 } 2580#endif 2581 slab_bufctl(slabp)[objnr] = slabp->free; 2582 slabp->free = objnr; 2583 slabp->inuse--; 2584} 2585 2586/* 2587 * Map pages beginning at addr to the given cache and slab. This is required 2588 * for the slab allocator to be able to lookup the cache and slab of a 2589 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. 2590 */ 2591static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, 2592 void *addr) 2593{ 2594 int nr_pages; 2595 struct page *page; 2596 2597 page = virt_to_page(addr); 2598 2599 nr_pages = 1; 2600 if (likely(!PageCompound(page))) 2601 nr_pages <<= cache->gfporder; 2602 2603 do { 2604 page_set_cache(page, cache); 2605 page_set_slab(page, slab); 2606 page++; 2607 } while (--nr_pages); 2608} 2609 2610/* 2611 * Grow (by 1) the number of slabs within a cache. This is called by 2612 * kmem_cache_alloc() when there are no active objs left in a cache. 2613 */ 2614static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) 2615{ 2616 struct slab *slabp; 2617 void *objp; 2618 size_t offset; 2619 gfp_t local_flags; 2620 unsigned long ctor_flags; 2621 struct kmem_list3 *l3; 2622 2623 /* 2624 * Be lazy and only check for valid flags here, keeping it out of the 2625 * critical path in kmem_cache_alloc(). 2626 */ 2627 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); 2628 if (flags & SLAB_NO_GROW) 2629 return 0; 2630 2631 ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2632 local_flags = (flags & SLAB_LEVEL_MASK); 2633 if (!(local_flags & __GFP_WAIT)) 2634 /* 2635 * Not allowed to sleep. Need to tell a constructor about 2636 * this - it might need to know... 2637 */ 2638 ctor_flags |= SLAB_CTOR_ATOMIC; 2639 2640 /* Take the l3 list lock to change the colour_next on this node */ 2641 check_irq_off(); 2642 l3 = cachep->nodelists[nodeid]; 2643 spin_lock(&l3->list_lock); 2644 2645 /* Get colour for the slab, and cal the next value. */ 2646 offset = l3->colour_next; 2647 l3->colour_next++; 2648 if (l3->colour_next >= cachep->colour) 2649 l3->colour_next = 0; 2650 spin_unlock(&l3->list_lock); 2651 2652 offset *= cachep->colour_off; 2653 2654 if (local_flags & __GFP_WAIT) 2655 local_irq_enable(); 2656 2657 /* 2658 * The test for missing atomic flag is performed here, rather than 2659 * the more obvious place, simply to reduce the critical path length 2660 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2661 * will eventually be caught here (where it matters). 2662 */ 2663 kmem_flagcheck(cachep, flags); 2664 2665 /* 2666 * Get mem for the objs. Attempt to allocate a physical page from 2667 * 'nodeid'. 2668 */ 2669 objp = kmem_getpages(cachep, flags, nodeid); 2670 if (!objp) 2671 goto failed; 2672 2673 /* Get slab management. */ 2674 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); 2675 if (!slabp) 2676 goto opps1; 2677 2678 slabp->nodeid = nodeid; 2679 slab_map_pages(cachep, slabp, objp); 2680 2681 cache_init_objs(cachep, slabp, ctor_flags); 2682 2683 if (local_flags & __GFP_WAIT) 2684 local_irq_disable(); 2685 check_irq_off(); 2686 spin_lock(&l3->list_lock); 2687 2688 /* Make slab active. */ 2689 list_add_tail(&slabp->list, &(l3->slabs_free)); 2690 STATS_INC_GROWN(cachep); 2691 l3->free_objects += cachep->num; 2692 spin_unlock(&l3->list_lock); 2693 return 1; 2694opps1: 2695 kmem_freepages(cachep, objp); 2696failed: 2697 if (local_flags & __GFP_WAIT) 2698 local_irq_disable(); 2699 return 0; 2700} 2701 2702#if DEBUG 2703 2704/* 2705 * Perform extra freeing checks: 2706 * - detect bad pointers. 2707 * - POISON/RED_ZONE checking 2708 * - destructor calls, for caches with POISON+dtor 2709 */ 2710static void kfree_debugcheck(const void *objp) 2711{ 2712 struct page *page; 2713 2714 if (!virt_addr_valid(objp)) { 2715 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2716 (unsigned long)objp); 2717 BUG(); 2718 } 2719 page = virt_to_page(objp); 2720 if (!PageSlab(page)) { 2721 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", 2722 (unsigned long)objp); 2723 BUG(); 2724 } 2725} 2726 2727static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2728{ 2729 unsigned long redzone1, redzone2; 2730 2731 redzone1 = *dbg_redzone1(cache, obj); 2732 redzone2 = *dbg_redzone2(cache, obj); 2733 2734 /* 2735 * Redzone is ok. 2736 */ 2737 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2738 return; 2739 2740 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2741 slab_error(cache, "double free detected"); 2742 else 2743 slab_error(cache, "memory outside object was overwritten"); 2744 2745 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n", 2746 obj, redzone1, redzone2); 2747} 2748 2749static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2750 void *caller) 2751{ 2752 struct page *page; 2753 unsigned int objnr; 2754 struct slab *slabp; 2755 2756 objp -= obj_offset(cachep); 2757 kfree_debugcheck(objp); 2758 page = virt_to_page(objp); 2759 2760 slabp = page_get_slab(page); 2761 2762 if (cachep->flags & SLAB_RED_ZONE) { 2763 verify_redzone_free(cachep, objp); 2764 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2765 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2766 } 2767 if (cachep->flags & SLAB_STORE_USER) 2768 *dbg_userword(cachep, objp) = caller; 2769 2770 objnr = obj_to_index(cachep, slabp, objp); 2771 2772 BUG_ON(objnr >= cachep->num); 2773 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2774 2775 if (cachep->flags & SLAB_DEBUG_INITIAL) { 2776 /* 2777 * Need to call the slab's constructor so the caller can 2778 * perform a verify of its state (debugging). Called without 2779 * the cache-lock held. 2780 */ 2781 cachep->ctor(objp + obj_offset(cachep), 2782 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); 2783 } 2784 if (cachep->flags & SLAB_POISON && cachep->dtor) { 2785 /* we want to cache poison the object, 2786 * call the destruction callback 2787 */ 2788 cachep->dtor(objp + obj_offset(cachep), cachep, 0); 2789 } 2790#ifdef CONFIG_DEBUG_SLAB_LEAK 2791 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2792#endif 2793 if (cachep->flags & SLAB_POISON) { 2794#ifdef CONFIG_DEBUG_PAGEALLOC 2795 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2796 store_stackinfo(cachep, objp, (unsigned long)caller); 2797 kernel_map_pages(virt_to_page(objp), 2798 cachep->buffer_size / PAGE_SIZE, 0); 2799 } else { 2800 poison_obj(cachep, objp, POISON_FREE); 2801 } 2802#else 2803 poison_obj(cachep, objp, POISON_FREE); 2804#endif 2805 } 2806 return objp; 2807} 2808 2809static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2810{ 2811 kmem_bufctl_t i; 2812 int entries = 0; 2813 2814 /* Check slab's freelist to see if this obj is there. */ 2815 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2816 entries++; 2817 if (entries > cachep->num || i >= cachep->num) 2818 goto bad; 2819 } 2820 if (entries != cachep->num - slabp->inuse) { 2821bad: 2822 printk(KERN_ERR "slab: Internal list corruption detected in " 2823 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2824 cachep->name, cachep->num, slabp, slabp->inuse); 2825 for (i = 0; 2826 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2827 i++) { 2828 if (i % 16 == 0) 2829 printk("\n%03x:", i); 2830 printk(" %02x", ((unsigned char *)slabp)[i]); 2831 } 2832 printk("\n"); 2833 BUG(); 2834 } 2835} 2836#else 2837#define kfree_debugcheck(x) do { } while(0) 2838#define cache_free_debugcheck(x,objp,z) (objp) 2839#define check_slabp(x,y) do { } while(0) 2840#endif 2841 2842static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2843{ 2844 int batchcount; 2845 struct kmem_list3 *l3; 2846 struct array_cache *ac; 2847 2848 check_irq_off(); 2849 ac = cpu_cache_get(cachep); 2850retry: 2851 batchcount = ac->batchcount; 2852 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2853 /* 2854 * If there was little recent activity on this cache, then 2855 * perform only a partial refill. Otherwise we could generate 2856 * refill bouncing. 2857 */ 2858 batchcount = BATCHREFILL_LIMIT; 2859 } 2860 l3 = cachep->nodelists[numa_node_id()]; 2861 2862 BUG_ON(ac->avail > 0 || !l3); 2863 spin_lock(&l3->list_lock); 2864 2865 /* See if we can refill from the shared array */ 2866 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) 2867 goto alloc_done; 2868 2869 while (batchcount > 0) { 2870 struct list_head *entry; 2871 struct slab *slabp; 2872 /* Get slab alloc is to come from. */ 2873 entry = l3->slabs_partial.next; 2874 if (entry == &l3->slabs_partial) { 2875 l3->free_touched = 1; 2876 entry = l3->slabs_free.next; 2877 if (entry == &l3->slabs_free) 2878 goto must_grow; 2879 } 2880 2881 slabp = list_entry(entry, struct slab, list); 2882 check_slabp(cachep, slabp); 2883 check_spinlock_acquired(cachep); 2884 while (slabp->inuse < cachep->num && batchcount--) { 2885 STATS_INC_ALLOCED(cachep); 2886 STATS_INC_ACTIVE(cachep); 2887 STATS_SET_HIGH(cachep); 2888 2889 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 2890 numa_node_id()); 2891 } 2892 check_slabp(cachep, slabp); 2893 2894 /* move slabp to correct slabp list: */ 2895 list_del(&slabp->list); 2896 if (slabp->free == BUFCTL_END) 2897 list_add(&slabp->list, &l3->slabs_full); 2898 else 2899 list_add(&slabp->list, &l3->slabs_partial); 2900 } 2901 2902must_grow: 2903 l3->free_objects -= ac->avail; 2904alloc_done: 2905 spin_unlock(&l3->list_lock); 2906 2907 if (unlikely(!ac->avail)) { 2908 int x; 2909 x = cache_grow(cachep, flags, numa_node_id()); 2910 2911 /* cache_grow can reenable interrupts, then ac could change. */ 2912 ac = cpu_cache_get(cachep); 2913 if (!x && ac->avail == 0) /* no objects in sight? abort */ 2914 return NULL; 2915 2916 if (!ac->avail) /* objects refilled by interrupt? */ 2917 goto retry; 2918 } 2919 ac->touched = 1; 2920 return ac->entry[--ac->avail]; 2921} 2922 2923static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 2924 gfp_t flags) 2925{ 2926 might_sleep_if(flags & __GFP_WAIT); 2927#if DEBUG 2928 kmem_flagcheck(cachep, flags); 2929#endif 2930} 2931 2932#if DEBUG 2933static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 2934 gfp_t flags, void *objp, void *caller) 2935{ 2936 if (!objp) 2937 return objp; 2938 if (cachep->flags & SLAB_POISON) { 2939#ifdef CONFIG_DEBUG_PAGEALLOC 2940 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 2941 kernel_map_pages(virt_to_page(objp), 2942 cachep->buffer_size / PAGE_SIZE, 1); 2943 else 2944 check_poison_obj(cachep, objp); 2945#else 2946 check_poison_obj(cachep, objp); 2947#endif 2948 poison_obj(cachep, objp, POISON_INUSE); 2949 } 2950 if (cachep->flags & SLAB_STORE_USER) 2951 *dbg_userword(cachep, objp) = caller; 2952 2953 if (cachep->flags & SLAB_RED_ZONE) { 2954 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 2955 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 2956 slab_error(cachep, "double free, or memory outside" 2957 " object was overwritten"); 2958 printk(KERN_ERR 2959 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", 2960 objp, *dbg_redzone1(cachep, objp), 2961 *dbg_redzone2(cachep, objp)); 2962 } 2963 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 2964 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 2965 } 2966#ifdef CONFIG_DEBUG_SLAB_LEAK 2967 { 2968 struct slab *slabp; 2969 unsigned objnr; 2970 2971 slabp = page_get_slab(virt_to_page(objp)); 2972 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 2973 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 2974 } 2975#endif 2976 objp += obj_offset(cachep); 2977 if (cachep->ctor && cachep->flags & SLAB_POISON) { 2978 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2979 2980 if (!(flags & __GFP_WAIT)) 2981 ctor_flags |= SLAB_CTOR_ATOMIC; 2982 2983 cachep->ctor(objp, cachep, ctor_flags); 2984 } 2985 return objp; 2986} 2987#else 2988#define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 2989#endif 2990 2991static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 2992{ 2993 void *objp; 2994 struct array_cache *ac; 2995 2996#ifdef CONFIG_NUMA 2997 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { 2998 objp = alternate_node_alloc(cachep, flags); 2999 if (objp != NULL) 3000 return objp; 3001 } 3002#endif 3003 3004 check_irq_off(); 3005 ac = cpu_cache_get(cachep); 3006 if (likely(ac->avail)) { 3007 STATS_INC_ALLOCHIT(cachep); 3008 ac->touched = 1; 3009 objp = ac->entry[--ac->avail]; 3010 } else { 3011 STATS_INC_ALLOCMISS(cachep); 3012 objp = cache_alloc_refill(cachep, flags); 3013 } 3014 return objp; 3015} 3016 3017static __always_inline void *__cache_alloc(struct kmem_cache *cachep, 3018 gfp_t flags, void *caller) 3019{ 3020 unsigned long save_flags; 3021 void *objp; 3022 3023 cache_alloc_debugcheck_before(cachep, flags); 3024 3025 local_irq_save(save_flags); 3026 objp = ____cache_alloc(cachep, flags); 3027 local_irq_restore(save_flags); 3028 objp = cache_alloc_debugcheck_after(cachep, flags, objp, 3029 caller); 3030 prefetchw(objp); 3031 return objp; 3032} 3033 3034#ifdef CONFIG_NUMA 3035/* 3036 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 3037 * 3038 * If we are in_interrupt, then process context, including cpusets and 3039 * mempolicy, may not apply and should not be used for allocation policy. 3040 */ 3041static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3042{ 3043 int nid_alloc, nid_here; 3044 3045 if (in_interrupt()) 3046 return NULL; 3047 nid_alloc = nid_here = numa_node_id(); 3048 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3049 nid_alloc = cpuset_mem_spread_node(); 3050 else if (current->mempolicy) 3051 nid_alloc = slab_node(current->mempolicy); 3052 if (nid_alloc != nid_here) 3053 return __cache_alloc_node(cachep, flags, nid_alloc); 3054 return NULL; 3055} 3056 3057/* 3058 * A interface to enable slab creation on nodeid 3059 */ 3060static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3061 int nodeid) 3062{ 3063 struct list_head *entry; 3064 struct slab *slabp; 3065 struct kmem_list3 *l3; 3066 void *obj; 3067 int x; 3068 3069 l3 = cachep->nodelists[nodeid]; 3070 BUG_ON(!l3); 3071 3072retry: 3073 check_irq_off(); 3074 spin_lock(&l3->list_lock); 3075 entry = l3->slabs_partial.next; 3076 if (entry == &l3->slabs_partial) { 3077 l3->free_touched = 1; 3078 entry = l3->slabs_free.next; 3079 if (entry == &l3->slabs_free) 3080 goto must_grow; 3081 } 3082 3083 slabp = list_entry(entry, struct slab, list); 3084 check_spinlock_acquired_node(cachep, nodeid); 3085 check_slabp(cachep, slabp); 3086 3087 STATS_INC_NODEALLOCS(cachep); 3088 STATS_INC_ACTIVE(cachep); 3089 STATS_SET_HIGH(cachep); 3090 3091 BUG_ON(slabp->inuse == cachep->num); 3092 3093 obj = slab_get_obj(cachep, slabp, nodeid); 3094 check_slabp(cachep, slabp); 3095 l3->free_objects--; 3096 /* move slabp to correct slabp list: */ 3097 list_del(&slabp->list); 3098 3099 if (slabp->free == BUFCTL_END) 3100 list_add(&slabp->list, &l3->slabs_full); 3101 else 3102 list_add(&slabp->list, &l3->slabs_partial); 3103 3104 spin_unlock(&l3->list_lock); 3105 goto done; 3106 3107must_grow: 3108 spin_unlock(&l3->list_lock); 3109 x = cache_grow(cachep, flags, nodeid); 3110 3111 if (!x) 3112 return NULL; 3113 3114 goto retry; 3115done: 3116 return obj; 3117} 3118#endif 3119 3120/* 3121 * Caller needs to acquire correct kmem_list's list_lock 3122 */ 3123static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 3124 int node) 3125{ 3126 int i; 3127 struct kmem_list3 *l3; 3128 3129 for (i = 0; i < nr_objects; i++) { 3130 void *objp = objpp[i]; 3131 struct slab *slabp; 3132 3133 slabp = virt_to_slab(objp); 3134 l3 = cachep->nodelists[node]; 3135 list_del(&slabp->list); 3136 check_spinlock_acquired_node(cachep, node); 3137 check_slabp(cachep, slabp); 3138 slab_put_obj(cachep, slabp, objp, node); 3139 STATS_DEC_ACTIVE(cachep); 3140 l3->free_objects++; 3141 check_slabp(cachep, slabp); 3142 3143 /* fixup slab chains */ 3144 if (slabp->inuse == 0) { 3145 if (l3->free_objects > l3->free_limit) { 3146 l3->free_objects -= cachep->num; 3147 /* No need to drop any previously held 3148 * lock here, even if we have a off-slab slab 3149 * descriptor it is guaranteed to come from 3150 * a different cache, refer to comments before 3151 * alloc_slabmgmt. 3152 */ 3153 slab_destroy(cachep, slabp); 3154 } else { 3155 list_add(&slabp->list, &l3->slabs_free); 3156 } 3157 } else { 3158 /* Unconditionally move a slab to the end of the 3159 * partial list on free - maximum time for the 3160 * other objects to be freed, too. 3161 */ 3162 list_add_tail(&slabp->list, &l3->slabs_partial); 3163 } 3164 } 3165} 3166 3167static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3168{ 3169 int batchcount; 3170 struct kmem_list3 *l3; 3171 int node = numa_node_id(); 3172 3173 batchcount = ac->batchcount; 3174#if DEBUG 3175 BUG_ON(!batchcount || batchcount > ac->avail); 3176#endif 3177 check_irq_off(); 3178 l3 = cachep->nodelists[node]; 3179 spin_lock(&l3->list_lock); 3180 if (l3->shared) { 3181 struct array_cache *shared_array = l3->shared; 3182 int max = shared_array->limit - shared_array->avail; 3183 if (max) { 3184 if (batchcount > max) 3185 batchcount = max; 3186 memcpy(&(shared_array->entry[shared_array->avail]), 3187 ac->entry, sizeof(void *) * batchcount); 3188 shared_array->avail += batchcount; 3189 goto free_done; 3190 } 3191 } 3192 3193 free_block(cachep, ac->entry, batchcount, node); 3194free_done: 3195#if STATS 3196 { 3197 int i = 0; 3198 struct list_head *p; 3199 3200 p = l3->slabs_free.next; 3201 while (p != &(l3->slabs_free)) { 3202 struct slab *slabp; 3203 3204 slabp = list_entry(p, struct slab, list); 3205 BUG_ON(slabp->inuse); 3206 3207 i++; 3208 p = p->next; 3209 } 3210 STATS_SET_FREEABLE(cachep, i); 3211 } 3212#endif 3213 spin_unlock(&l3->list_lock); 3214 ac->avail -= batchcount; 3215 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3216} 3217 3218/* 3219 * Release an obj back to its cache. If the obj has a constructed state, it must 3220 * be in this state _before_ it is released. Called with disabled ints. 3221 */ 3222static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3223{ 3224 struct array_cache *ac = cpu_cache_get(cachep); 3225 3226 check_irq_off(); 3227 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3228 3229 if (cache_free_alien(cachep, objp)) 3230 return; 3231 3232 if (likely(ac->avail < ac->limit)) { 3233 STATS_INC_FREEHIT(cachep); 3234 ac->entry[ac->avail++] = objp; 3235 return; 3236 } else { 3237 STATS_INC_FREEMISS(cachep); 3238 cache_flusharray(cachep, ac); 3239 ac->entry[ac->avail++] = objp; 3240 } 3241} 3242 3243/** 3244 * kmem_cache_alloc - Allocate an object 3245 * @cachep: The cache to allocate from. 3246 * @flags: See kmalloc(). 3247 * 3248 * Allocate an object from this cache. The flags are only relevant 3249 * if the cache has no available objects. 3250 */ 3251void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3252{ 3253 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3254} 3255EXPORT_SYMBOL(kmem_cache_alloc); 3256 3257/** 3258 * kmem_cache_zalloc - Allocate an object. The memory is set to zero. 3259 * @cache: The cache to allocate from. 3260 * @flags: See kmalloc(). 3261 * 3262 * Allocate an object from this cache and set the allocated memory to zero. 3263 * The flags are only relevant if the cache has no available objects. 3264 */ 3265void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) 3266{ 3267 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); 3268 if (ret) 3269 memset(ret, 0, obj_size(cache)); 3270 return ret; 3271} 3272EXPORT_SYMBOL(kmem_cache_zalloc); 3273 3274/** 3275 * kmem_ptr_validate - check if an untrusted pointer might 3276 * be a slab entry. 3277 * @cachep: the cache we're checking against 3278 * @ptr: pointer to validate 3279 * 3280 * This verifies that the untrusted pointer looks sane: 3281 * it is _not_ a guarantee that the pointer is actually 3282 * part of the slab cache in question, but it at least 3283 * validates that the pointer can be dereferenced and 3284 * looks half-way sane. 3285 * 3286 * Currently only used for dentry validation. 3287 */ 3288int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) 3289{ 3290 unsigned long addr = (unsigned long)ptr; 3291 unsigned long min_addr = PAGE_OFFSET; 3292 unsigned long align_mask = BYTES_PER_WORD - 1; 3293 unsigned long size = cachep->buffer_size; 3294 struct page *page; 3295 3296 if (unlikely(addr < min_addr)) 3297 goto out; 3298 if (unlikely(addr > (unsigned long)high_memory - size)) 3299 goto out; 3300 if (unlikely(addr & align_mask)) 3301 goto out; 3302 if (unlikely(!kern_addr_valid(addr))) 3303 goto out; 3304 if (unlikely(!kern_addr_valid(addr + size - 1))) 3305 goto out; 3306 page = virt_to_page(ptr); 3307 if (unlikely(!PageSlab(page))) 3308 goto out; 3309 if (unlikely(page_get_cache(page) != cachep)) 3310 goto out; 3311 return 1; 3312out: 3313 return 0; 3314} 3315 3316#ifdef CONFIG_NUMA 3317/** 3318 * kmem_cache_alloc_node - Allocate an object on the specified node 3319 * @cachep: The cache to allocate from. 3320 * @flags: See kmalloc(). 3321 * @nodeid: node number of the target node. 3322 * 3323 * Identical to kmem_cache_alloc, except that this function is slow 3324 * and can sleep. And it will allocate memory on the given node, which 3325 * can improve the performance for cpu bound structures. 3326 * New and improved: it will now make sure that the object gets 3327 * put on the correct node list so that there is no false sharing. 3328 */ 3329void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3330{ 3331 unsigned long save_flags; 3332 void *ptr; 3333 3334 cache_alloc_debugcheck_before(cachep, flags); 3335 local_irq_save(save_flags); 3336 3337 if (nodeid == -1 || nodeid == numa_node_id() || 3338 !cachep->nodelists[nodeid]) 3339 ptr = ____cache_alloc(cachep, flags); 3340 else 3341 ptr = __cache_alloc_node(cachep, flags, nodeid); 3342 local_irq_restore(save_flags); 3343 3344 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, 3345 __builtin_return_address(0)); 3346 3347 return ptr; 3348} 3349EXPORT_SYMBOL(kmem_cache_alloc_node); 3350 3351void *kmalloc_node(size_t size, gfp_t flags, int node) 3352{ 3353 struct kmem_cache *cachep; 3354 3355 cachep = kmem_find_general_cachep(size, flags); 3356 if (unlikely(cachep == NULL)) 3357 return NULL; 3358 return kmem_cache_alloc_node(cachep, flags, node); 3359} 3360EXPORT_SYMBOL(kmalloc_node); 3361#endif 3362 3363/** 3364 * __do_kmalloc - allocate memory 3365 * @size: how many bytes of memory are required. 3366 * @flags: the type of memory to allocate (see kmalloc). 3367 * @caller: function caller for debug tracking of the caller 3368 */ 3369static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3370 void *caller) 3371{ 3372 struct kmem_cache *cachep; 3373 3374 /* If you want to save a few bytes .text space: replace 3375 * __ with kmem_. 3376 * Then kmalloc uses the uninlined functions instead of the inline 3377 * functions. 3378 */ 3379 cachep = __find_general_cachep(size, flags); 3380 if (unlikely(cachep == NULL)) 3381 return NULL; 3382 return __cache_alloc(cachep, flags, caller); 3383} 3384 3385 3386void *__kmalloc(size_t size, gfp_t flags) 3387{ 3388#ifndef CONFIG_DEBUG_SLAB 3389 return __do_kmalloc(size, flags, NULL); 3390#else 3391 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3392#endif 3393} 3394EXPORT_SYMBOL(__kmalloc); 3395 3396#ifdef CONFIG_DEBUG_SLAB 3397void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) 3398{ 3399 return __do_kmalloc(size, flags, caller); 3400} 3401EXPORT_SYMBOL(__kmalloc_track_caller); 3402#endif 3403 3404#ifdef CONFIG_SMP 3405/** 3406 * percpu_depopulate - depopulate per-cpu data for given cpu 3407 * @__pdata: per-cpu data to depopulate 3408 * @cpu: depopulate per-cpu data for this cpu 3409 * 3410 * Depopulating per-cpu data for a cpu going offline would be a typical 3411 * use case. You need to register a cpu hotplug handler for that purpose. 3412 */ 3413void percpu_depopulate(void *__pdata, int cpu) 3414{ 3415 struct percpu_data *pdata = __percpu_disguise(__pdata); 3416 if (pdata->ptrs[cpu]) { 3417 kfree(pdata->ptrs[cpu]); 3418 pdata->ptrs[cpu] = NULL; 3419 } 3420} 3421EXPORT_SYMBOL_GPL(percpu_depopulate); 3422 3423/** 3424 * percpu_depopulate_mask - depopulate per-cpu data for some cpu's 3425 * @__pdata: per-cpu data to depopulate 3426 * @mask: depopulate per-cpu data for cpu's selected through mask bits 3427 */ 3428void __percpu_depopulate_mask(void *__pdata, cpumask_t *mask) 3429{ 3430 int cpu; 3431 for_each_cpu_mask(cpu, *mask) 3432 percpu_depopulate(__pdata, cpu); 3433} 3434EXPORT_SYMBOL_GPL(__percpu_depopulate_mask); 3435 3436/** 3437 * percpu_populate - populate per-cpu data for given cpu 3438 * @__pdata: per-cpu data to populate further 3439 * @size: size of per-cpu object 3440 * @gfp: may sleep or not etc. 3441 * @cpu: populate per-data for this cpu 3442 * 3443 * Populating per-cpu data for a cpu coming online would be a typical 3444 * use case. You need to register a cpu hotplug handler for that purpose. 3445 * Per-cpu object is populated with zeroed buffer. 3446 */ 3447void *percpu_populate(void *__pdata, size_t size, gfp_t gfp, int cpu) 3448{ 3449 struct percpu_data *pdata = __percpu_disguise(__pdata); 3450 int node = cpu_to_node(cpu); 3451 3452 BUG_ON(pdata->ptrs[cpu]); 3453 if (node_online(node)) { 3454 /* FIXME: kzalloc_node(size, gfp, node) */ 3455 pdata->ptrs[cpu] = kmalloc_node(size, gfp, node); 3456 if (pdata->ptrs[cpu]) 3457 memset(pdata->ptrs[cpu], 0, size); 3458 } else 3459 pdata->ptrs[cpu] = kzalloc(size, gfp); 3460 return pdata->ptrs[cpu]; 3461} 3462EXPORT_SYMBOL_GPL(percpu_populate); 3463 3464/** 3465 * percpu_populate_mask - populate per-cpu data for more cpu's 3466 * @__pdata: per-cpu data to populate further 3467 * @size: size of per-cpu object 3468 * @gfp: may sleep or not etc. 3469 * @mask: populate per-cpu data for cpu's selected through mask bits 3470 * 3471 * Per-cpu objects are populated with zeroed buffers. 3472 */ 3473int __percpu_populate_mask(void *__pdata, size_t size, gfp_t gfp, 3474 cpumask_t *mask) 3475{ 3476 cpumask_t populated = CPU_MASK_NONE; 3477 int cpu; 3478 3479 for_each_cpu_mask(cpu, *mask) 3480 if (unlikely(!percpu_populate(__pdata, size, gfp, cpu))) { 3481 __percpu_depopulate_mask(__pdata, &populated); 3482 return -ENOMEM; 3483 } else 3484 cpu_set(cpu, populated); 3485 return 0; 3486} 3487EXPORT_SYMBOL_GPL(__percpu_populate_mask); 3488 3489/** 3490 * percpu_alloc_mask - initial setup of per-cpu data 3491 * @size: size of per-cpu object 3492 * @gfp: may sleep or not etc. 3493 * @mask: populate per-data for cpu's selected through mask bits 3494 * 3495 * Populating per-cpu data for all online cpu's would be a typical use case, 3496 * which is simplified by the percpu_alloc() wrapper. 3497 * Per-cpu objects are populated with zeroed buffers. 3498 */ 3499void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask) 3500{ 3501 void *pdata = kzalloc(sizeof(struct percpu_data), gfp); 3502 void *__pdata = __percpu_disguise(pdata); 3503 3504 if (unlikely(!pdata)) 3505 return NULL; 3506 if (likely(!__percpu_populate_mask(__pdata, size, gfp, mask))) 3507 return __pdata; 3508 kfree(pdata); 3509 return NULL; 3510} 3511EXPORT_SYMBOL_GPL(__percpu_alloc_mask); 3512 3513/** 3514 * percpu_free - final cleanup of per-cpu data 3515 * @__pdata: object to clean up 3516 * 3517 * We simply clean up any per-cpu object left. No need for the client to 3518 * track and specify through a bis mask which per-cpu objects are to free. 3519 */ 3520void percpu_free(void *__pdata) 3521{ 3522 __percpu_depopulate_mask(__pdata, &cpu_possible_map); 3523 kfree(__percpu_disguise(__pdata)); 3524} 3525EXPORT_SYMBOL_GPL(percpu_free); 3526#endif /* CONFIG_SMP */ 3527 3528/** 3529 * kmem_cache_free - Deallocate an object 3530 * @cachep: The cache the allocation was from. 3531 * @objp: The previously allocated object. 3532 * 3533 * Free an object which was previously allocated from this 3534 * cache. 3535 */ 3536void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3537{ 3538 unsigned long flags; 3539 3540 BUG_ON(virt_to_cache(objp) != cachep); 3541 3542 local_irq_save(flags); 3543 __cache_free(cachep, objp); 3544 local_irq_restore(flags); 3545} 3546EXPORT_SYMBOL(kmem_cache_free); 3547 3548/** 3549 * kfree - free previously allocated memory 3550 * @objp: pointer returned by kmalloc. 3551 * 3552 * If @objp is NULL, no operation is performed. 3553 * 3554 * Don't free memory not originally allocated by kmalloc() 3555 * or you will run into trouble. 3556 */ 3557void kfree(const void *objp) 3558{ 3559 struct kmem_cache *c; 3560 unsigned long flags; 3561 3562 if (unlikely(!objp)) 3563 return; 3564 local_irq_save(flags); 3565 kfree_debugcheck(objp); 3566 c = virt_to_cache(objp); 3567 debug_check_no_locks_freed(objp, obj_size(c)); 3568 __cache_free(c, (void *)objp); 3569 local_irq_restore(flags); 3570} 3571EXPORT_SYMBOL(kfree); 3572 3573unsigned int kmem_cache_size(struct kmem_cache *cachep) 3574{ 3575 return obj_size(cachep); 3576} 3577EXPORT_SYMBOL(kmem_cache_size); 3578 3579const char *kmem_cache_name(struct kmem_cache *cachep) 3580{ 3581 return cachep->name; 3582} 3583EXPORT_SYMBOL_GPL(kmem_cache_name); 3584 3585/* 3586 * This initializes kmem_list3 or resizes varioius caches for all nodes. 3587 */ 3588static int alloc_kmemlist(struct kmem_cache *cachep) 3589{ 3590 int node; 3591 struct kmem_list3 *l3; 3592 struct array_cache *new_shared; 3593 struct array_cache **new_alien; 3594 3595 for_each_online_node(node) { 3596 3597 new_alien = alloc_alien_cache(node, cachep->limit); 3598 if (!new_alien) 3599 goto fail; 3600 3601 new_shared = alloc_arraycache(node, 3602 cachep->shared*cachep->batchcount, 3603 0xbaadf00d); 3604 if (!new_shared) { 3605 free_alien_cache(new_alien); 3606 goto fail; 3607 } 3608 3609 l3 = cachep->nodelists[node]; 3610 if (l3) { 3611 struct array_cache *shared = l3->shared; 3612 3613 spin_lock_irq(&l3->list_lock); 3614 3615 if (shared) 3616 free_block(cachep, shared->entry, 3617 shared->avail, node); 3618 3619 l3->shared = new_shared; 3620 if (!l3->alien) { 3621 l3->alien = new_alien; 3622 new_alien = NULL; 3623 } 3624 l3->free_limit = (1 + nr_cpus_node(node)) * 3625 cachep->batchcount + cachep->num; 3626 spin_unlock_irq(&l3->list_lock); 3627 kfree(shared); 3628 free_alien_cache(new_alien); 3629 continue; 3630 } 3631 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); 3632 if (!l3) { 3633 free_alien_cache(new_alien); 3634 kfree(new_shared); 3635 goto fail; 3636 } 3637 3638 kmem_list3_init(l3); 3639 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3640 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3641 l3->shared = new_shared; 3642 l3->alien = new_alien; 3643 l3->free_limit = (1 + nr_cpus_node(node)) * 3644 cachep->batchcount + cachep->num; 3645 cachep->nodelists[node] = l3; 3646 } 3647 return 0; 3648 3649fail: 3650 if (!cachep->next.next) { 3651 /* Cache is not active yet. Roll back what we did */ 3652 node--; 3653 while (node >= 0) { 3654 if (cachep->nodelists[node]) { 3655 l3 = cachep->nodelists[node]; 3656 3657 kfree(l3->shared); 3658 free_alien_cache(l3->alien); 3659 kfree(l3); 3660 cachep->nodelists[node] = NULL; 3661 } 3662 node--; 3663 } 3664 } 3665 return -ENOMEM; 3666} 3667 3668struct ccupdate_struct { 3669 struct kmem_cache *cachep; 3670 struct array_cache *new[NR_CPUS]; 3671}; 3672 3673static void do_ccupdate_local(void *info) 3674{ 3675 struct ccupdate_struct *new = info; 3676 struct array_cache *old; 3677 3678 check_irq_off(); 3679 old = cpu_cache_get(new->cachep); 3680 3681 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3682 new->new[smp_processor_id()] = old; 3683} 3684 3685/* Always called with the cache_chain_mutex held */ 3686static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3687 int batchcount, int shared) 3688{ 3689 struct ccupdate_struct new; 3690 int i, err; 3691 3692 memset(&new.new, 0, sizeof(new.new)); 3693 for_each_online_cpu(i) { 3694 new.new[i] = alloc_arraycache(cpu_to_node(i), limit, 3695 batchcount); 3696 if (!new.new[i]) { 3697 for (i--; i >= 0; i--) 3698 kfree(new.new[i]); 3699 return -ENOMEM; 3700 } 3701 } 3702 new.cachep = cachep; 3703 3704 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1); 3705 3706 check_irq_on(); 3707 cachep->batchcount = batchcount; 3708 cachep->limit = limit; 3709 cachep->shared = shared; 3710 3711 for_each_online_cpu(i) { 3712 struct array_cache *ccold = new.new[i]; 3713 if (!ccold) 3714 continue; 3715 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3716 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 3717 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3718 kfree(ccold); 3719 } 3720 3721 err = alloc_kmemlist(cachep); 3722 if (err) { 3723 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", 3724 cachep->name, -err); 3725 BUG(); 3726 } 3727 return 0; 3728} 3729 3730/* Called with cache_chain_mutex held always */ 3731static void enable_cpucache(struct kmem_cache *cachep) 3732{ 3733 int err; 3734 int limit, shared; 3735 3736 /* 3737 * The head array serves three purposes: 3738 * - create a LIFO ordering, i.e. return objects that are cache-warm 3739 * - reduce the number of spinlock operations. 3740 * - reduce the number of linked list operations on the slab and 3741 * bufctl chains: array operations are cheaper. 3742 * The numbers are guessed, we should auto-tune as described by 3743 * Bonwick. 3744 */ 3745 if (cachep->buffer_size > 131072) 3746 limit = 1; 3747 else if (cachep->buffer_size > PAGE_SIZE) 3748 limit = 8; 3749 else if (cachep->buffer_size > 1024) 3750 limit = 24; 3751 else if (cachep->buffer_size > 256) 3752 limit = 54; 3753 else 3754 limit = 120; 3755 3756 /* 3757 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 3758 * allocation behaviour: Most allocs on one cpu, most free operations 3759 * on another cpu. For these cases, an efficient object passing between 3760 * cpus is necessary. This is provided by a shared array. The array 3761 * replaces Bonwick's magazine layer. 3762 * On uniprocessor, it's functionally equivalent (but less efficient) 3763 * to a larger limit. Thus disabled by default. 3764 */ 3765 shared = 0; 3766#ifdef CONFIG_SMP 3767 if (cachep->buffer_size <= PAGE_SIZE) 3768 shared = 8; 3769#endif 3770 3771#if DEBUG 3772 /* 3773 * With debugging enabled, large batchcount lead to excessively long 3774 * periods with disabled local interrupts. Limit the batchcount 3775 */ 3776 if (limit > 32) 3777 limit = 32; 3778#endif 3779 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); 3780 if (err) 3781 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 3782 cachep->name, -err); 3783} 3784 3785/* 3786 * Drain an array if it contains any elements taking the l3 lock only if 3787 * necessary. Note that the l3 listlock also protects the array_cache 3788 * if drain_array() is used on the shared array. 3789 */ 3790void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 3791 struct array_cache *ac, int force, int node) 3792{ 3793 int tofree; 3794 3795 if (!ac || !ac->avail) 3796 return; 3797 if (ac->touched && !force) { 3798 ac->touched = 0; 3799 } else { 3800 spin_lock_irq(&l3->list_lock); 3801 if (ac->avail) { 3802 tofree = force ? ac->avail : (ac->limit + 4) / 5; 3803 if (tofree > ac->avail) 3804 tofree = (ac->avail + 1) / 2; 3805 free_block(cachep, ac->entry, tofree, node); 3806 ac->avail -= tofree; 3807 memmove(ac->entry, &(ac->entry[tofree]), 3808 sizeof(void *) * ac->avail); 3809 } 3810 spin_unlock_irq(&l3->list_lock); 3811 } 3812} 3813 3814/** 3815 * cache_reap - Reclaim memory from caches. 3816 * @unused: unused parameter 3817 * 3818 * Called from workqueue/eventd every few seconds. 3819 * Purpose: 3820 * - clear the per-cpu caches for this CPU. 3821 * - return freeable pages to the main free memory pool. 3822 * 3823 * If we cannot acquire the cache chain mutex then just give up - we'll try 3824 * again on the next iteration. 3825 */ 3826static void cache_reap(void *unused) 3827{ 3828 struct kmem_cache *searchp; 3829 struct kmem_list3 *l3; 3830 int node = numa_node_id(); 3831 3832 if (!mutex_trylock(&cache_chain_mutex)) { 3833 /* Give up. Setup the next iteration. */ 3834 schedule_delayed_work(&__get_cpu_var(reap_work), 3835 REAPTIMEOUT_CPUC); 3836 return; 3837 } 3838 3839 list_for_each_entry(searchp, &cache_chain, next) { 3840 check_irq_on(); 3841 3842 /* 3843 * We only take the l3 lock if absolutely necessary and we 3844 * have established with reasonable certainty that 3845 * we can do some work if the lock was obtained. 3846 */ 3847 l3 = searchp->nodelists[node]; 3848 3849 reap_alien(searchp, l3); 3850 3851 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 3852 3853 /* 3854 * These are racy checks but it does not matter 3855 * if we skip one check or scan twice. 3856 */ 3857 if (time_after(l3->next_reap, jiffies)) 3858 goto next; 3859 3860 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 3861 3862 drain_array(searchp, l3, l3->shared, 0, node); 3863 3864 if (l3->free_touched) 3865 l3->free_touched = 0; 3866 else { 3867 int freed; 3868 3869 freed = drain_freelist(searchp, l3, (l3->free_limit + 3870 5 * searchp->num - 1) / (5 * searchp->num)); 3871 STATS_ADD_REAPED(searchp, freed); 3872 } 3873next: 3874 cond_resched(); 3875 } 3876 check_irq_on(); 3877 mutex_unlock(&cache_chain_mutex); 3878 next_reap_node(); 3879 refresh_cpu_vm_stats(smp_processor_id()); 3880 /* Set up the next iteration */ 3881 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); 3882} 3883 3884#ifdef CONFIG_PROC_FS 3885 3886static void print_slabinfo_header(struct seq_file *m) 3887{ 3888 /* 3889 * Output format version, so at least we can change it 3890 * without _too_ many complaints. 3891 */ 3892#if STATS 3893 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 3894#else 3895 seq_puts(m, "slabinfo - version: 2.1\n"); 3896#endif 3897 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 3898 "<objperslab> <pagesperslab>"); 3899 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 3900 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 3901#if STATS 3902 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 3903 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 3904 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 3905#endif 3906 seq_putc(m, '\n'); 3907} 3908 3909static void *s_start(struct seq_file *m, loff_t *pos) 3910{ 3911 loff_t n = *pos; 3912 struct list_head *p; 3913 3914 mutex_lock(&cache_chain_mutex); 3915 if (!n) 3916 print_slabinfo_header(m); 3917 p = cache_chain.next; 3918 while (n--) { 3919 p = p->next; 3920 if (p == &cache_chain) 3921 return NULL; 3922 } 3923 return list_entry(p, struct kmem_cache, next); 3924} 3925 3926static void *s_next(struct seq_file *m, void *p, loff_t *pos) 3927{ 3928 struct kmem_cache *cachep = p; 3929 ++*pos; 3930 return cachep->next.next == &cache_chain ? 3931 NULL : list_entry(cachep->next.next, struct kmem_cache, next); 3932} 3933 3934static void s_stop(struct seq_file *m, void *p) 3935{ 3936 mutex_unlock(&cache_chain_mutex); 3937} 3938 3939static int s_show(struct seq_file *m, void *p) 3940{ 3941 struct kmem_cache *cachep = p; 3942 struct slab *slabp; 3943 unsigned long active_objs; 3944 unsigned long num_objs; 3945 unsigned long active_slabs = 0; 3946 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 3947 const char *name; 3948 char *error = NULL; 3949 int node; 3950 struct kmem_list3 *l3; 3951 3952 active_objs = 0; 3953 num_slabs = 0; 3954 for_each_online_node(node) { 3955 l3 = cachep->nodelists[node]; 3956 if (!l3) 3957 continue; 3958 3959 check_irq_on(); 3960 spin_lock_irq(&l3->list_lock); 3961 3962 list_for_each_entry(slabp, &l3->slabs_full, list) { 3963 if (slabp->inuse != cachep->num && !error) 3964 error = "slabs_full accounting error"; 3965 active_objs += cachep->num; 3966 active_slabs++; 3967 } 3968 list_for_each_entry(slabp, &l3->slabs_partial, list) { 3969 if (slabp->inuse == cachep->num && !error) 3970 error = "slabs_partial inuse accounting error"; 3971 if (!slabp->inuse && !error) 3972 error = "slabs_partial/inuse accounting error"; 3973 active_objs += slabp->inuse; 3974 active_slabs++; 3975 } 3976 list_for_each_entry(slabp, &l3->slabs_free, list) { 3977 if (slabp->inuse && !error) 3978 error = "slabs_free/inuse accounting error"; 3979 num_slabs++; 3980 } 3981 free_objects += l3->free_objects; 3982 if (l3->shared) 3983 shared_avail += l3->shared->avail; 3984 3985 spin_unlock_irq(&l3->list_lock); 3986 } 3987 num_slabs += active_slabs; 3988 num_objs = num_slabs * cachep->num; 3989 if (num_objs - active_objs != free_objects && !error) 3990 error = "free_objects accounting error"; 3991 3992 name = cachep->name; 3993 if (error) 3994 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 3995 3996 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 3997 name, active_objs, num_objs, cachep->buffer_size, 3998 cachep->num, (1 << cachep->gfporder)); 3999 seq_printf(m, " : tunables %4u %4u %4u", 4000 cachep->limit, cachep->batchcount, cachep->shared); 4001 seq_printf(m, " : slabdata %6lu %6lu %6lu", 4002 active_slabs, num_slabs, shared_avail); 4003#if STATS 4004 { /* list3 stats */ 4005 unsigned long high = cachep->high_mark; 4006 unsigned long allocs = cachep->num_allocations; 4007 unsigned long grown = cachep->grown; 4008 unsigned long reaped = cachep->reaped; 4009 unsigned long errors = cachep->errors; 4010 unsigned long max_freeable = cachep->max_freeable; 4011 unsigned long node_allocs = cachep->node_allocs; 4012 unsigned long node_frees = cachep->node_frees; 4013 unsigned long overflows = cachep->node_overflow; 4014 4015 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ 4016 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, 4017 reaped, errors, max_freeable, node_allocs, 4018 node_frees, overflows); 4019 } 4020 /* cpu stats */ 4021 { 4022 unsigned long allochit = atomic_read(&cachep->allochit); 4023 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4024 unsigned long freehit = atomic_read(&cachep->freehit); 4025 unsigned long freemiss = atomic_read(&cachep->freemiss); 4026 4027 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4028 allochit, allocmiss, freehit, freemiss); 4029 } 4030#endif 4031 seq_putc(m, '\n'); 4032 return 0; 4033} 4034 4035/* 4036 * slabinfo_op - iterator that generates /proc/slabinfo 4037 * 4038 * Output layout: 4039 * cache-name 4040 * num-active-objs 4041 * total-objs 4042 * object size 4043 * num-active-slabs 4044 * total-slabs 4045 * num-pages-per-slab 4046 * + further values on SMP and with statistics enabled 4047 */ 4048 4049struct seq_operations slabinfo_op = { 4050 .start = s_start, 4051 .next = s_next, 4052 .stop = s_stop, 4053 .show = s_show, 4054}; 4055 4056#define MAX_SLABINFO_WRITE 128 4057/** 4058 * slabinfo_write - Tuning for the slab allocator 4059 * @file: unused 4060 * @buffer: user buffer 4061 * @count: data length 4062 * @ppos: unused 4063 */ 4064ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4065 size_t count, loff_t *ppos) 4066{ 4067 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4068 int limit, batchcount, shared, res; 4069 struct kmem_cache *cachep; 4070 4071 if (count > MAX_SLABINFO_WRITE) 4072 return -EINVAL; 4073 if (copy_from_user(&kbuf, buffer, count)) 4074 return -EFAULT; 4075 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4076 4077 tmp = strchr(kbuf, ' '); 4078 if (!tmp) 4079 return -EINVAL; 4080 *tmp = '\0'; 4081 tmp++; 4082 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4083 return -EINVAL; 4084 4085 /* Find the cache in the chain of caches. */ 4086 mutex_lock(&cache_chain_mutex); 4087 res = -EINVAL; 4088 list_for_each_entry(cachep, &cache_chain, next) { 4089 if (!strcmp(cachep->name, kbuf)) { 4090 if (limit < 1 || batchcount < 1 || 4091 batchcount > limit || shared < 0) { 4092 res = 0; 4093 } else { 4094 res = do_tune_cpucache(cachep, limit, 4095 batchcount, shared); 4096 } 4097 break; 4098 } 4099 } 4100 mutex_unlock(&cache_chain_mutex); 4101 if (res >= 0) 4102 res = count; 4103 return res; 4104} 4105 4106#ifdef CONFIG_DEBUG_SLAB_LEAK 4107 4108static void *leaks_start(struct seq_file *m, loff_t *pos) 4109{ 4110 loff_t n = *pos; 4111 struct list_head *p; 4112 4113 mutex_lock(&cache_chain_mutex); 4114 p = cache_chain.next; 4115 while (n--) { 4116 p = p->next; 4117 if (p == &cache_chain) 4118 return NULL; 4119 } 4120 return list_entry(p, struct kmem_cache, next); 4121} 4122 4123static inline int add_caller(unsigned long *n, unsigned long v) 4124{ 4125 unsigned long *p; 4126 int l; 4127 if (!v) 4128 return 1; 4129 l = n[1]; 4130 p = n + 2; 4131 while (l) { 4132 int i = l/2; 4133 unsigned long *q = p + 2 * i; 4134 if (*q == v) { 4135 q[1]++; 4136 return 1; 4137 } 4138 if (*q > v) { 4139 l = i; 4140 } else { 4141 p = q + 2; 4142 l -= i + 1; 4143 } 4144 } 4145 if (++n[1] == n[0]) 4146 return 0; 4147 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4148 p[0] = v; 4149 p[1] = 1; 4150 return 1; 4151} 4152 4153static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4154{ 4155 void *p; 4156 int i; 4157 if (n[0] == n[1]) 4158 return; 4159 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4160 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4161 continue; 4162 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4163 return; 4164 } 4165} 4166 4167static void show_symbol(struct seq_file *m, unsigned long address) 4168{ 4169#ifdef CONFIG_KALLSYMS 4170 char *modname; 4171 const char *name; 4172 unsigned long offset, size; 4173 char namebuf[KSYM_NAME_LEN+1]; 4174 4175 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); 4176 4177 if (name) { 4178 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4179 if (modname) 4180 seq_printf(m, " [%s]", modname); 4181 return; 4182 } 4183#endif 4184 seq_printf(m, "%p", (void *)address); 4185} 4186 4187static int leaks_show(struct seq_file *m, void *p) 4188{ 4189 struct kmem_cache *cachep = p; 4190 struct slab *slabp; 4191 struct kmem_list3 *l3; 4192 const char *name; 4193 unsigned long *n = m->private; 4194 int node; 4195 int i; 4196 4197 if (!(cachep->flags & SLAB_STORE_USER)) 4198 return 0; 4199 if (!(cachep->flags & SLAB_RED_ZONE)) 4200 return 0; 4201 4202 /* OK, we can do it */ 4203 4204 n[1] = 0; 4205 4206 for_each_online_node(node) { 4207 l3 = cachep->nodelists[node]; 4208 if (!l3) 4209 continue; 4210 4211 check_irq_on(); 4212 spin_lock_irq(&l3->list_lock); 4213 4214 list_for_each_entry(slabp, &l3->slabs_full, list) 4215 handle_slab(n, cachep, slabp); 4216 list_for_each_entry(slabp, &l3->slabs_partial, list) 4217 handle_slab(n, cachep, slabp); 4218 spin_unlock_irq(&l3->list_lock); 4219 } 4220 name = cachep->name; 4221 if (n[0] == n[1]) { 4222 /* Increase the buffer size */ 4223 mutex_unlock(&cache_chain_mutex); 4224 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4225 if (!m->private) { 4226 /* Too bad, we are really out */ 4227 m->private = n; 4228 mutex_lock(&cache_chain_mutex); 4229 return -ENOMEM; 4230 } 4231 *(unsigned long *)m->private = n[0] * 2; 4232 kfree(n); 4233 mutex_lock(&cache_chain_mutex); 4234 /* Now make sure this entry will be retried */ 4235 m->count = m->size; 4236 return 0; 4237 } 4238 for (i = 0; i < n[1]; i++) { 4239 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4240 show_symbol(m, n[2*i+2]); 4241 seq_putc(m, '\n'); 4242 } 4243 return 0; 4244} 4245 4246struct seq_operations slabstats_op = { 4247 .start = leaks_start, 4248 .next = s_next, 4249 .stop = s_stop, 4250 .show = leaks_show, 4251}; 4252#endif 4253#endif 4254 4255/** 4256 * ksize - get the actual amount of memory allocated for a given object 4257 * @objp: Pointer to the object 4258 * 4259 * kmalloc may internally round up allocations and return more memory 4260 * than requested. ksize() can be used to determine the actual amount of 4261 * memory allocated. The caller may use this additional memory, even though 4262 * a smaller amount of memory was initially specified with the kmalloc call. 4263 * The caller must guarantee that objp points to a valid object previously 4264 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4265 * must not be freed during the duration of the call. 4266 */ 4267unsigned int ksize(const void *objp) 4268{ 4269 if (unlikely(objp == NULL)) 4270 return 0; 4271 4272 return obj_size(virt_to_cache(objp)); 4273} 4274