slab.c revision 897e679b17460b52752a038af29db356fe1bd759
1/* 2 * linux/mm/slab.c 3 * Written by Mark Hemment, 1996/97. 4 * (markhe@nextd.demon.co.uk) 5 * 6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 7 * 8 * Major cleanup, different bufctl logic, per-cpu arrays 9 * (c) 2000 Manfred Spraul 10 * 11 * Cleanup, make the head arrays unconditional, preparation for NUMA 12 * (c) 2002 Manfred Spraul 13 * 14 * An implementation of the Slab Allocator as described in outline in; 15 * UNIX Internals: The New Frontiers by Uresh Vahalia 16 * Pub: Prentice Hall ISBN 0-13-101908-2 17 * or with a little more detail in; 18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 19 * Jeff Bonwick (Sun Microsystems). 20 * Presented at: USENIX Summer 1994 Technical Conference 21 * 22 * The memory is organized in caches, one cache for each object type. 23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 24 * Each cache consists out of many slabs (they are small (usually one 25 * page long) and always contiguous), and each slab contains multiple 26 * initialized objects. 27 * 28 * This means, that your constructor is used only for newly allocated 29 * slabs and you must pass objects with the same intializations to 30 * kmem_cache_free. 31 * 32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 33 * normal). If you need a special memory type, then must create a new 34 * cache for that memory type. 35 * 36 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 37 * full slabs with 0 free objects 38 * partial slabs 39 * empty slabs with no allocated objects 40 * 41 * If partial slabs exist, then new allocations come from these slabs, 42 * otherwise from empty slabs or new slabs are allocated. 43 * 44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 46 * 47 * Each cache has a short per-cpu head array, most allocs 48 * and frees go into that array, and if that array overflows, then 1/2 49 * of the entries in the array are given back into the global cache. 50 * The head array is strictly LIFO and should improve the cache hit rates. 51 * On SMP, it additionally reduces the spinlock operations. 52 * 53 * The c_cpuarray may not be read with enabled local interrupts - 54 * it's changed with a smp_call_function(). 55 * 56 * SMP synchronization: 57 * constructors and destructors are called without any locking. 58 * Several members in struct kmem_cache and struct slab never change, they 59 * are accessed without any locking. 60 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 61 * and local interrupts are disabled so slab code is preempt-safe. 62 * The non-constant members are protected with a per-cache irq spinlock. 63 * 64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 65 * in 2000 - many ideas in the current implementation are derived from 66 * his patch. 67 * 68 * Further notes from the original documentation: 69 * 70 * 11 April '97. Started multi-threading - markhe 71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'. 72 * The sem is only needed when accessing/extending the cache-chain, which 73 * can never happen inside an interrupt (kmem_cache_create(), 74 * kmem_cache_shrink() and kmem_cache_reap()). 75 * 76 * At present, each engine can be growing a cache. This should be blocked. 77 * 78 * 15 March 2005. NUMA slab allocator. 79 * Shai Fultheim <shai@scalex86.org>. 80 * Shobhit Dayal <shobhit@calsoftinc.com> 81 * Alok N Kataria <alokk@calsoftinc.com> 82 * Christoph Lameter <christoph@lameter.com> 83 * 84 * Modified the slab allocator to be node aware on NUMA systems. 85 * Each node has its own list of partial, free and full slabs. 86 * All object allocations for a node occur from node specific slab lists. 87 */ 88 89#include <linux/slab.h> 90#include <linux/mm.h> 91#include <linux/poison.h> 92#include <linux/swap.h> 93#include <linux/cache.h> 94#include <linux/interrupt.h> 95#include <linux/init.h> 96#include <linux/compiler.h> 97#include <linux/cpuset.h> 98#include <linux/seq_file.h> 99#include <linux/notifier.h> 100#include <linux/kallsyms.h> 101#include <linux/cpu.h> 102#include <linux/sysctl.h> 103#include <linux/module.h> 104#include <linux/rcupdate.h> 105#include <linux/string.h> 106#include <linux/uaccess.h> 107#include <linux/nodemask.h> 108#include <linux/mempolicy.h> 109#include <linux/mutex.h> 110#include <linux/fault-inject.h> 111#include <linux/rtmutex.h> 112#include <linux/reciprocal_div.h> 113 114#include <asm/cacheflush.h> 115#include <asm/tlbflush.h> 116#include <asm/page.h> 117 118/* 119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. 120 * 0 for faster, smaller code (especially in the critical paths). 121 * 122 * STATS - 1 to collect stats for /proc/slabinfo. 123 * 0 for faster, smaller code (especially in the critical paths). 124 * 125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 126 */ 127 128#ifdef CONFIG_DEBUG_SLAB 129#define DEBUG 1 130#define STATS 1 131#define FORCED_DEBUG 1 132#else 133#define DEBUG 0 134#define STATS 0 135#define FORCED_DEBUG 0 136#endif 137 138/* Shouldn't this be in a header file somewhere? */ 139#define BYTES_PER_WORD sizeof(void *) 140#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) 141 142#ifndef cache_line_size 143#define cache_line_size() L1_CACHE_BYTES 144#endif 145 146#ifndef ARCH_KMALLOC_MINALIGN 147/* 148 * Enforce a minimum alignment for the kmalloc caches. 149 * Usually, the kmalloc caches are cache_line_size() aligned, except when 150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. 151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed 152 * alignment larger than the alignment of a 64-bit integer. 153 * ARCH_KMALLOC_MINALIGN allows that. 154 * Note that increasing this value may disable some debug features. 155 */ 156#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 157#endif 158 159#ifndef ARCH_SLAB_MINALIGN 160/* 161 * Enforce a minimum alignment for all caches. 162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD 163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. 164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables 165 * some debug features. 166 */ 167#define ARCH_SLAB_MINALIGN 0 168#endif 169 170#ifndef ARCH_KMALLOC_FLAGS 171#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 172#endif 173 174/* Legal flag mask for kmem_cache_create(). */ 175#if DEBUG 176# define CREATE_MASK (SLAB_RED_ZONE | \ 177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 178 SLAB_CACHE_DMA | \ 179 SLAB_STORE_USER | \ 180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 182#else 183# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 184 SLAB_CACHE_DMA | \ 185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 187#endif 188 189/* 190 * kmem_bufctl_t: 191 * 192 * Bufctl's are used for linking objs within a slab 193 * linked offsets. 194 * 195 * This implementation relies on "struct page" for locating the cache & 196 * slab an object belongs to. 197 * This allows the bufctl structure to be small (one int), but limits 198 * the number of objects a slab (not a cache) can contain when off-slab 199 * bufctls are used. The limit is the size of the largest general cache 200 * that does not use off-slab slabs. 201 * For 32bit archs with 4 kB pages, is this 56. 202 * This is not serious, as it is only for large objects, when it is unwise 203 * to have too many per slab. 204 * Note: This limit can be raised by introducing a general cache whose size 205 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 206 */ 207 208typedef unsigned int kmem_bufctl_t; 209#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 210#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 211#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 212#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 213 214/* 215 * struct slab 216 * 217 * Manages the objs in a slab. Placed either at the beginning of mem allocated 218 * for a slab, or allocated from an general cache. 219 * Slabs are chained into three list: fully used, partial, fully free slabs. 220 */ 221struct slab { 222 struct list_head list; 223 unsigned long colouroff; 224 void *s_mem; /* including colour offset */ 225 unsigned int inuse; /* num of objs active in slab */ 226 kmem_bufctl_t free; 227 unsigned short nodeid; 228}; 229 230/* 231 * struct slab_rcu 232 * 233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 234 * arrange for kmem_freepages to be called via RCU. This is useful if 235 * we need to approach a kernel structure obliquely, from its address 236 * obtained without the usual locking. We can lock the structure to 237 * stabilize it and check it's still at the given address, only if we 238 * can be sure that the memory has not been meanwhile reused for some 239 * other kind of object (which our subsystem's lock might corrupt). 240 * 241 * rcu_read_lock before reading the address, then rcu_read_unlock after 242 * taking the spinlock within the structure expected at that address. 243 * 244 * We assume struct slab_rcu can overlay struct slab when destroying. 245 */ 246struct slab_rcu { 247 struct rcu_head head; 248 struct kmem_cache *cachep; 249 void *addr; 250}; 251 252/* 253 * struct array_cache 254 * 255 * Purpose: 256 * - LIFO ordering, to hand out cache-warm objects from _alloc 257 * - reduce the number of linked list operations 258 * - reduce spinlock operations 259 * 260 * The limit is stored in the per-cpu structure to reduce the data cache 261 * footprint. 262 * 263 */ 264struct array_cache { 265 unsigned int avail; 266 unsigned int limit; 267 unsigned int batchcount; 268 unsigned int touched; 269 spinlock_t lock; 270 void *entry[0]; /* 271 * Must have this definition in here for the proper 272 * alignment of array_cache. Also simplifies accessing 273 * the entries. 274 * [0] is for gcc 2.95. It should really be []. 275 */ 276}; 277 278/* 279 * bootstrap: The caches do not work without cpuarrays anymore, but the 280 * cpuarrays are allocated from the generic caches... 281 */ 282#define BOOT_CPUCACHE_ENTRIES 1 283struct arraycache_init { 284 struct array_cache cache; 285 void *entries[BOOT_CPUCACHE_ENTRIES]; 286}; 287 288/* 289 * The slab lists for all objects. 290 */ 291struct kmem_list3 { 292 struct list_head slabs_partial; /* partial list first, better asm code */ 293 struct list_head slabs_full; 294 struct list_head slabs_free; 295 unsigned long free_objects; 296 unsigned int free_limit; 297 unsigned int colour_next; /* Per-node cache coloring */ 298 spinlock_t list_lock; 299 struct array_cache *shared; /* shared per node */ 300 struct array_cache **alien; /* on other nodes */ 301 unsigned long next_reap; /* updated without locking */ 302 int free_touched; /* updated without locking */ 303}; 304 305/* 306 * Need this for bootstrapping a per node allocator. 307 */ 308#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) 309struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 310#define CACHE_CACHE 0 311#define SIZE_AC 1 312#define SIZE_L3 (1 + MAX_NUMNODES) 313 314static int drain_freelist(struct kmem_cache *cache, 315 struct kmem_list3 *l3, int tofree); 316static void free_block(struct kmem_cache *cachep, void **objpp, int len, 317 int node); 318static int enable_cpucache(struct kmem_cache *cachep); 319static void cache_reap(struct work_struct *unused); 320 321/* 322 * This function must be completely optimized away if a constant is passed to 323 * it. Mostly the same as what is in linux/slab.h except it returns an index. 324 */ 325static __always_inline int index_of(const size_t size) 326{ 327 extern void __bad_size(void); 328 329 if (__builtin_constant_p(size)) { 330 int i = 0; 331 332#define CACHE(x) \ 333 if (size <=x) \ 334 return i; \ 335 else \ 336 i++; 337#include "linux/kmalloc_sizes.h" 338#undef CACHE 339 __bad_size(); 340 } else 341 __bad_size(); 342 return 0; 343} 344 345static int slab_early_init = 1; 346 347#define INDEX_AC index_of(sizeof(struct arraycache_init)) 348#define INDEX_L3 index_of(sizeof(struct kmem_list3)) 349 350static void kmem_list3_init(struct kmem_list3 *parent) 351{ 352 INIT_LIST_HEAD(&parent->slabs_full); 353 INIT_LIST_HEAD(&parent->slabs_partial); 354 INIT_LIST_HEAD(&parent->slabs_free); 355 parent->shared = NULL; 356 parent->alien = NULL; 357 parent->colour_next = 0; 358 spin_lock_init(&parent->list_lock); 359 parent->free_objects = 0; 360 parent->free_touched = 0; 361} 362 363#define MAKE_LIST(cachep, listp, slab, nodeid) \ 364 do { \ 365 INIT_LIST_HEAD(listp); \ 366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 367 } while (0) 368 369#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 370 do { \ 371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 374 } while (0) 375 376/* 377 * struct kmem_cache 378 * 379 * manages a cache. 380 */ 381 382struct kmem_cache { 383/* 1) per-cpu data, touched during every alloc/free */ 384 struct array_cache *array[NR_CPUS]; 385/* 2) Cache tunables. Protected by cache_chain_mutex */ 386 unsigned int batchcount; 387 unsigned int limit; 388 unsigned int shared; 389 390 unsigned int buffer_size; 391 u32 reciprocal_buffer_size; 392/* 3) touched by every alloc & free from the backend */ 393 394 unsigned int flags; /* constant flags */ 395 unsigned int num; /* # of objs per slab */ 396 397/* 4) cache_grow/shrink */ 398 /* order of pgs per slab (2^n) */ 399 unsigned int gfporder; 400 401 /* force GFP flags, e.g. GFP_DMA */ 402 gfp_t gfpflags; 403 404 size_t colour; /* cache colouring range */ 405 unsigned int colour_off; /* colour offset */ 406 struct kmem_cache *slabp_cache; 407 unsigned int slab_size; 408 unsigned int dflags; /* dynamic flags */ 409 410 /* constructor func */ 411 void (*ctor) (void *, struct kmem_cache *, unsigned long); 412 413/* 5) cache creation/removal */ 414 const char *name; 415 struct list_head next; 416 417/* 6) statistics */ 418#if STATS 419 unsigned long num_active; 420 unsigned long num_allocations; 421 unsigned long high_mark; 422 unsigned long grown; 423 unsigned long reaped; 424 unsigned long errors; 425 unsigned long max_freeable; 426 unsigned long node_allocs; 427 unsigned long node_frees; 428 unsigned long node_overflow; 429 atomic_t allochit; 430 atomic_t allocmiss; 431 atomic_t freehit; 432 atomic_t freemiss; 433#endif 434#if DEBUG 435 /* 436 * If debugging is enabled, then the allocator can add additional 437 * fields and/or padding to every object. buffer_size contains the total 438 * object size including these internal fields, the following two 439 * variables contain the offset to the user object and its size. 440 */ 441 int obj_offset; 442 int obj_size; 443#endif 444 /* 445 * We put nodelists[] at the end of kmem_cache, because we want to size 446 * this array to nr_node_ids slots instead of MAX_NUMNODES 447 * (see kmem_cache_init()) 448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache 449 * is statically defined, so we reserve the max number of nodes. 450 */ 451 struct kmem_list3 *nodelists[MAX_NUMNODES]; 452 /* 453 * Do not add fields after nodelists[] 454 */ 455}; 456 457#define CFLGS_OFF_SLAB (0x80000000UL) 458#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 459 460#define BATCHREFILL_LIMIT 16 461/* 462 * Optimization question: fewer reaps means less probability for unnessary 463 * cpucache drain/refill cycles. 464 * 465 * OTOH the cpuarrays can contain lots of objects, 466 * which could lock up otherwise freeable slabs. 467 */ 468#define REAPTIMEOUT_CPUC (2*HZ) 469#define REAPTIMEOUT_LIST3 (4*HZ) 470 471#if STATS 472#define STATS_INC_ACTIVE(x) ((x)->num_active++) 473#define STATS_DEC_ACTIVE(x) ((x)->num_active--) 474#define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 475#define STATS_INC_GROWN(x) ((x)->grown++) 476#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 477#define STATS_SET_HIGH(x) \ 478 do { \ 479 if ((x)->num_active > (x)->high_mark) \ 480 (x)->high_mark = (x)->num_active; \ 481 } while (0) 482#define STATS_INC_ERR(x) ((x)->errors++) 483#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 484#define STATS_INC_NODEFREES(x) ((x)->node_frees++) 485#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 486#define STATS_SET_FREEABLE(x, i) \ 487 do { \ 488 if ((x)->max_freeable < i) \ 489 (x)->max_freeable = i; \ 490 } while (0) 491#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 492#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 493#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 494#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 495#else 496#define STATS_INC_ACTIVE(x) do { } while (0) 497#define STATS_DEC_ACTIVE(x) do { } while (0) 498#define STATS_INC_ALLOCED(x) do { } while (0) 499#define STATS_INC_GROWN(x) do { } while (0) 500#define STATS_ADD_REAPED(x,y) do { } while (0) 501#define STATS_SET_HIGH(x) do { } while (0) 502#define STATS_INC_ERR(x) do { } while (0) 503#define STATS_INC_NODEALLOCS(x) do { } while (0) 504#define STATS_INC_NODEFREES(x) do { } while (0) 505#define STATS_INC_ACOVERFLOW(x) do { } while (0) 506#define STATS_SET_FREEABLE(x, i) do { } while (0) 507#define STATS_INC_ALLOCHIT(x) do { } while (0) 508#define STATS_INC_ALLOCMISS(x) do { } while (0) 509#define STATS_INC_FREEHIT(x) do { } while (0) 510#define STATS_INC_FREEMISS(x) do { } while (0) 511#endif 512 513#if DEBUG 514 515/* 516 * memory layout of objects: 517 * 0 : objp 518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 519 * the end of an object is aligned with the end of the real 520 * allocation. Catches writes behind the end of the allocation. 521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 522 * redzone word. 523 * cachep->obj_offset: The real object. 524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 526 * [BYTES_PER_WORD long] 527 */ 528static int obj_offset(struct kmem_cache *cachep) 529{ 530 return cachep->obj_offset; 531} 532 533static int obj_size(struct kmem_cache *cachep) 534{ 535 return cachep->obj_size; 536} 537 538static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 539{ 540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 541 return (unsigned long long*) (objp + obj_offset(cachep) - 542 sizeof(unsigned long long)); 543} 544 545static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 546{ 547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 548 if (cachep->flags & SLAB_STORE_USER) 549 return (unsigned long long *)(objp + cachep->buffer_size - 550 sizeof(unsigned long long) - 551 REDZONE_ALIGN); 552 return (unsigned long long *) (objp + cachep->buffer_size - 553 sizeof(unsigned long long)); 554} 555 556static void **dbg_userword(struct kmem_cache *cachep, void *objp) 557{ 558 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 560} 561 562#else 563 564#define obj_offset(x) 0 565#define obj_size(cachep) (cachep->buffer_size) 566#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 567#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 568#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 569 570#endif 571 572/* 573 * Do not go above this order unless 0 objects fit into the slab. 574 */ 575#define BREAK_GFP_ORDER_HI 1 576#define BREAK_GFP_ORDER_LO 0 577static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 578 579/* 580 * Functions for storing/retrieving the cachep and or slab from the page 581 * allocator. These are used to find the slab an obj belongs to. With kfree(), 582 * these are used to find the cache which an obj belongs to. 583 */ 584static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 585{ 586 page->lru.next = (struct list_head *)cache; 587} 588 589static inline struct kmem_cache *page_get_cache(struct page *page) 590{ 591 page = compound_head(page); 592 BUG_ON(!PageSlab(page)); 593 return (struct kmem_cache *)page->lru.next; 594} 595 596static inline void page_set_slab(struct page *page, struct slab *slab) 597{ 598 page->lru.prev = (struct list_head *)slab; 599} 600 601static inline struct slab *page_get_slab(struct page *page) 602{ 603 BUG_ON(!PageSlab(page)); 604 return (struct slab *)page->lru.prev; 605} 606 607static inline struct kmem_cache *virt_to_cache(const void *obj) 608{ 609 struct page *page = virt_to_head_page(obj); 610 return page_get_cache(page); 611} 612 613static inline struct slab *virt_to_slab(const void *obj) 614{ 615 struct page *page = virt_to_head_page(obj); 616 return page_get_slab(page); 617} 618 619static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 620 unsigned int idx) 621{ 622 return slab->s_mem + cache->buffer_size * idx; 623} 624 625/* 626 * We want to avoid an expensive divide : (offset / cache->buffer_size) 627 * Using the fact that buffer_size is a constant for a particular cache, 628 * we can replace (offset / cache->buffer_size) by 629 * reciprocal_divide(offset, cache->reciprocal_buffer_size) 630 */ 631static inline unsigned int obj_to_index(const struct kmem_cache *cache, 632 const struct slab *slab, void *obj) 633{ 634 u32 offset = (obj - slab->s_mem); 635 return reciprocal_divide(offset, cache->reciprocal_buffer_size); 636} 637 638/* 639 * These are the default caches for kmalloc. Custom caches can have other sizes. 640 */ 641struct cache_sizes malloc_sizes[] = { 642#define CACHE(x) { .cs_size = (x) }, 643#include <linux/kmalloc_sizes.h> 644 CACHE(ULONG_MAX) 645#undef CACHE 646}; 647EXPORT_SYMBOL(malloc_sizes); 648 649/* Must match cache_sizes above. Out of line to keep cache footprint low. */ 650struct cache_names { 651 char *name; 652 char *name_dma; 653}; 654 655static struct cache_names __initdata cache_names[] = { 656#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 657#include <linux/kmalloc_sizes.h> 658 {NULL,} 659#undef CACHE 660}; 661 662static struct arraycache_init initarray_cache __initdata = 663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 664static struct arraycache_init initarray_generic = 665 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 666 667/* internal cache of cache description objs */ 668static struct kmem_cache cache_cache = { 669 .batchcount = 1, 670 .limit = BOOT_CPUCACHE_ENTRIES, 671 .shared = 1, 672 .buffer_size = sizeof(struct kmem_cache), 673 .name = "kmem_cache", 674}; 675 676#define BAD_ALIEN_MAGIC 0x01020304ul 677 678#ifdef CONFIG_LOCKDEP 679 680/* 681 * Slab sometimes uses the kmalloc slabs to store the slab headers 682 * for other slabs "off slab". 683 * The locking for this is tricky in that it nests within the locks 684 * of all other slabs in a few places; to deal with this special 685 * locking we put on-slab caches into a separate lock-class. 686 * 687 * We set lock class for alien array caches which are up during init. 688 * The lock annotation will be lost if all cpus of a node goes down and 689 * then comes back up during hotplug 690 */ 691static struct lock_class_key on_slab_l3_key; 692static struct lock_class_key on_slab_alc_key; 693 694static inline void init_lock_keys(void) 695 696{ 697 int q; 698 struct cache_sizes *s = malloc_sizes; 699 700 while (s->cs_size != ULONG_MAX) { 701 for_each_node(q) { 702 struct array_cache **alc; 703 int r; 704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; 705 if (!l3 || OFF_SLAB(s->cs_cachep)) 706 continue; 707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key); 708 alc = l3->alien; 709 /* 710 * FIXME: This check for BAD_ALIEN_MAGIC 711 * should go away when common slab code is taught to 712 * work even without alien caches. 713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC 714 * for alloc_alien_cache, 715 */ 716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) 717 continue; 718 for_each_node(r) { 719 if (alc[r]) 720 lockdep_set_class(&alc[r]->lock, 721 &on_slab_alc_key); 722 } 723 } 724 s++; 725 } 726} 727#else 728static inline void init_lock_keys(void) 729{ 730} 731#endif 732 733/* 734 * 1. Guard access to the cache-chain. 735 * 2. Protect sanity of cpu_online_map against cpu hotplug events 736 */ 737static DEFINE_MUTEX(cache_chain_mutex); 738static struct list_head cache_chain; 739 740/* 741 * chicken and egg problem: delay the per-cpu array allocation 742 * until the general caches are up. 743 */ 744static enum { 745 NONE, 746 PARTIAL_AC, 747 PARTIAL_L3, 748 FULL 749} g_cpucache_up; 750 751/* 752 * used by boot code to determine if it can use slab based allocator 753 */ 754int slab_is_available(void) 755{ 756 return g_cpucache_up == FULL; 757} 758 759static DEFINE_PER_CPU(struct delayed_work, reap_work); 760 761static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 762{ 763 return cachep->array[smp_processor_id()]; 764} 765 766static inline struct kmem_cache *__find_general_cachep(size_t size, 767 gfp_t gfpflags) 768{ 769 struct cache_sizes *csizep = malloc_sizes; 770 771#if DEBUG 772 /* This happens if someone tries to call 773 * kmem_cache_create(), or __kmalloc(), before 774 * the generic caches are initialized. 775 */ 776 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 777#endif 778 while (size > csizep->cs_size) 779 csizep++; 780 781 /* 782 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 783 * has cs_{dma,}cachep==NULL. Thus no special case 784 * for large kmalloc calls required. 785 */ 786#ifdef CONFIG_ZONE_DMA 787 if (unlikely(gfpflags & GFP_DMA)) 788 return csizep->cs_dmacachep; 789#endif 790 return csizep->cs_cachep; 791} 792 793static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 794{ 795 return __find_general_cachep(size, gfpflags); 796} 797 798static size_t slab_mgmt_size(size_t nr_objs, size_t align) 799{ 800 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 801} 802 803/* 804 * Calculate the number of objects and left-over bytes for a given buffer size. 805 */ 806static void cache_estimate(unsigned long gfporder, size_t buffer_size, 807 size_t align, int flags, size_t *left_over, 808 unsigned int *num) 809{ 810 int nr_objs; 811 size_t mgmt_size; 812 size_t slab_size = PAGE_SIZE << gfporder; 813 814 /* 815 * The slab management structure can be either off the slab or 816 * on it. For the latter case, the memory allocated for a 817 * slab is used for: 818 * 819 * - The struct slab 820 * - One kmem_bufctl_t for each object 821 * - Padding to respect alignment of @align 822 * - @buffer_size bytes for each object 823 * 824 * If the slab management structure is off the slab, then the 825 * alignment will already be calculated into the size. Because 826 * the slabs are all pages aligned, the objects will be at the 827 * correct alignment when allocated. 828 */ 829 if (flags & CFLGS_OFF_SLAB) { 830 mgmt_size = 0; 831 nr_objs = slab_size / buffer_size; 832 833 if (nr_objs > SLAB_LIMIT) 834 nr_objs = SLAB_LIMIT; 835 } else { 836 /* 837 * Ignore padding for the initial guess. The padding 838 * is at most @align-1 bytes, and @buffer_size is at 839 * least @align. In the worst case, this result will 840 * be one greater than the number of objects that fit 841 * into the memory allocation when taking the padding 842 * into account. 843 */ 844 nr_objs = (slab_size - sizeof(struct slab)) / 845 (buffer_size + sizeof(kmem_bufctl_t)); 846 847 /* 848 * This calculated number will be either the right 849 * amount, or one greater than what we want. 850 */ 851 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 852 > slab_size) 853 nr_objs--; 854 855 if (nr_objs > SLAB_LIMIT) 856 nr_objs = SLAB_LIMIT; 857 858 mgmt_size = slab_mgmt_size(nr_objs, align); 859 } 860 *num = nr_objs; 861 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 862} 863 864#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) 865 866static void __slab_error(const char *function, struct kmem_cache *cachep, 867 char *msg) 868{ 869 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 870 function, cachep->name, msg); 871 dump_stack(); 872} 873 874/* 875 * By default on NUMA we use alien caches to stage the freeing of 876 * objects allocated from other nodes. This causes massive memory 877 * inefficiencies when using fake NUMA setup to split memory into a 878 * large number of small nodes, so it can be disabled on the command 879 * line 880 */ 881 882static int use_alien_caches __read_mostly = 1; 883static int __init noaliencache_setup(char *s) 884{ 885 use_alien_caches = 0; 886 return 1; 887} 888__setup("noaliencache", noaliencache_setup); 889 890#ifdef CONFIG_NUMA 891/* 892 * Special reaping functions for NUMA systems called from cache_reap(). 893 * These take care of doing round robin flushing of alien caches (containing 894 * objects freed on different nodes from which they were allocated) and the 895 * flushing of remote pcps by calling drain_node_pages. 896 */ 897static DEFINE_PER_CPU(unsigned long, reap_node); 898 899static void init_reap_node(int cpu) 900{ 901 int node; 902 903 node = next_node(cpu_to_node(cpu), node_online_map); 904 if (node == MAX_NUMNODES) 905 node = first_node(node_online_map); 906 907 per_cpu(reap_node, cpu) = node; 908} 909 910static void next_reap_node(void) 911{ 912 int node = __get_cpu_var(reap_node); 913 914 node = next_node(node, node_online_map); 915 if (unlikely(node >= MAX_NUMNODES)) 916 node = first_node(node_online_map); 917 __get_cpu_var(reap_node) = node; 918} 919 920#else 921#define init_reap_node(cpu) do { } while (0) 922#define next_reap_node(void) do { } while (0) 923#endif 924 925/* 926 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 927 * via the workqueue/eventd. 928 * Add the CPU number into the expiration time to minimize the possibility of 929 * the CPUs getting into lockstep and contending for the global cache chain 930 * lock. 931 */ 932static void __cpuinit start_cpu_timer(int cpu) 933{ 934 struct delayed_work *reap_work = &per_cpu(reap_work, cpu); 935 936 /* 937 * When this gets called from do_initcalls via cpucache_init(), 938 * init_workqueues() has already run, so keventd will be setup 939 * at that time. 940 */ 941 if (keventd_up() && reap_work->work.func == NULL) { 942 init_reap_node(cpu); 943 INIT_DELAYED_WORK(reap_work, cache_reap); 944 schedule_delayed_work_on(cpu, reap_work, 945 __round_jiffies_relative(HZ, cpu)); 946 } 947} 948 949static struct array_cache *alloc_arraycache(int node, int entries, 950 int batchcount) 951{ 952 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 953 struct array_cache *nc = NULL; 954 955 nc = kmalloc_node(memsize, GFP_KERNEL, node); 956 if (nc) { 957 nc->avail = 0; 958 nc->limit = entries; 959 nc->batchcount = batchcount; 960 nc->touched = 0; 961 spin_lock_init(&nc->lock); 962 } 963 return nc; 964} 965 966/* 967 * Transfer objects in one arraycache to another. 968 * Locking must be handled by the caller. 969 * 970 * Return the number of entries transferred. 971 */ 972static int transfer_objects(struct array_cache *to, 973 struct array_cache *from, unsigned int max) 974{ 975 /* Figure out how many entries to transfer */ 976 int nr = min(min(from->avail, max), to->limit - to->avail); 977 978 if (!nr) 979 return 0; 980 981 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 982 sizeof(void *) *nr); 983 984 from->avail -= nr; 985 to->avail += nr; 986 to->touched = 1; 987 return nr; 988} 989 990#ifndef CONFIG_NUMA 991 992#define drain_alien_cache(cachep, alien) do { } while (0) 993#define reap_alien(cachep, l3) do { } while (0) 994 995static inline struct array_cache **alloc_alien_cache(int node, int limit) 996{ 997 return (struct array_cache **)BAD_ALIEN_MAGIC; 998} 999 1000static inline void free_alien_cache(struct array_cache **ac_ptr) 1001{ 1002} 1003 1004static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1005{ 1006 return 0; 1007} 1008 1009static inline void *alternate_node_alloc(struct kmem_cache *cachep, 1010 gfp_t flags) 1011{ 1012 return NULL; 1013} 1014 1015static inline void *____cache_alloc_node(struct kmem_cache *cachep, 1016 gfp_t flags, int nodeid) 1017{ 1018 return NULL; 1019} 1020 1021#else /* CONFIG_NUMA */ 1022 1023static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 1024static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 1025 1026static struct array_cache **alloc_alien_cache(int node, int limit) 1027{ 1028 struct array_cache **ac_ptr; 1029 int memsize = sizeof(void *) * nr_node_ids; 1030 int i; 1031 1032 if (limit > 1) 1033 limit = 12; 1034 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); 1035 if (ac_ptr) { 1036 for_each_node(i) { 1037 if (i == node || !node_online(i)) { 1038 ac_ptr[i] = NULL; 1039 continue; 1040 } 1041 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); 1042 if (!ac_ptr[i]) { 1043 for (i--; i <= 0; i--) 1044 kfree(ac_ptr[i]); 1045 kfree(ac_ptr); 1046 return NULL; 1047 } 1048 } 1049 } 1050 return ac_ptr; 1051} 1052 1053static void free_alien_cache(struct array_cache **ac_ptr) 1054{ 1055 int i; 1056 1057 if (!ac_ptr) 1058 return; 1059 for_each_node(i) 1060 kfree(ac_ptr[i]); 1061 kfree(ac_ptr); 1062} 1063 1064static void __drain_alien_cache(struct kmem_cache *cachep, 1065 struct array_cache *ac, int node) 1066{ 1067 struct kmem_list3 *rl3 = cachep->nodelists[node]; 1068 1069 if (ac->avail) { 1070 spin_lock(&rl3->list_lock); 1071 /* 1072 * Stuff objects into the remote nodes shared array first. 1073 * That way we could avoid the overhead of putting the objects 1074 * into the free lists and getting them back later. 1075 */ 1076 if (rl3->shared) 1077 transfer_objects(rl3->shared, ac, ac->limit); 1078 1079 free_block(cachep, ac->entry, ac->avail, node); 1080 ac->avail = 0; 1081 spin_unlock(&rl3->list_lock); 1082 } 1083} 1084 1085/* 1086 * Called from cache_reap() to regularly drain alien caches round robin. 1087 */ 1088static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 1089{ 1090 int node = __get_cpu_var(reap_node); 1091 1092 if (l3->alien) { 1093 struct array_cache *ac = l3->alien[node]; 1094 1095 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1096 __drain_alien_cache(cachep, ac, node); 1097 spin_unlock_irq(&ac->lock); 1098 } 1099 } 1100} 1101 1102static void drain_alien_cache(struct kmem_cache *cachep, 1103 struct array_cache **alien) 1104{ 1105 int i = 0; 1106 struct array_cache *ac; 1107 unsigned long flags; 1108 1109 for_each_online_node(i) { 1110 ac = alien[i]; 1111 if (ac) { 1112 spin_lock_irqsave(&ac->lock, flags); 1113 __drain_alien_cache(cachep, ac, i); 1114 spin_unlock_irqrestore(&ac->lock, flags); 1115 } 1116 } 1117} 1118 1119static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1120{ 1121 struct slab *slabp = virt_to_slab(objp); 1122 int nodeid = slabp->nodeid; 1123 struct kmem_list3 *l3; 1124 struct array_cache *alien = NULL; 1125 int node; 1126 1127 node = numa_node_id(); 1128 1129 /* 1130 * Make sure we are not freeing a object from another node to the array 1131 * cache on this cpu. 1132 */ 1133 if (likely(slabp->nodeid == node)) 1134 return 0; 1135 1136 l3 = cachep->nodelists[node]; 1137 STATS_INC_NODEFREES(cachep); 1138 if (l3->alien && l3->alien[nodeid]) { 1139 alien = l3->alien[nodeid]; 1140 spin_lock(&alien->lock); 1141 if (unlikely(alien->avail == alien->limit)) { 1142 STATS_INC_ACOVERFLOW(cachep); 1143 __drain_alien_cache(cachep, alien, nodeid); 1144 } 1145 alien->entry[alien->avail++] = objp; 1146 spin_unlock(&alien->lock); 1147 } else { 1148 spin_lock(&(cachep->nodelists[nodeid])->list_lock); 1149 free_block(cachep, &objp, 1, nodeid); 1150 spin_unlock(&(cachep->nodelists[nodeid])->list_lock); 1151 } 1152 return 1; 1153} 1154#endif 1155 1156static int __cpuinit cpuup_callback(struct notifier_block *nfb, 1157 unsigned long action, void *hcpu) 1158{ 1159 long cpu = (long)hcpu; 1160 struct kmem_cache *cachep; 1161 struct kmem_list3 *l3 = NULL; 1162 int node = cpu_to_node(cpu); 1163 int memsize = sizeof(struct kmem_list3); 1164 1165 switch (action) { 1166 case CPU_LOCK_ACQUIRE: 1167 mutex_lock(&cache_chain_mutex); 1168 break; 1169 case CPU_UP_PREPARE: 1170 case CPU_UP_PREPARE_FROZEN: 1171 /* 1172 * We need to do this right in the beginning since 1173 * alloc_arraycache's are going to use this list. 1174 * kmalloc_node allows us to add the slab to the right 1175 * kmem_list3 and not this cpu's kmem_list3 1176 */ 1177 1178 list_for_each_entry(cachep, &cache_chain, next) { 1179 /* 1180 * Set up the size64 kmemlist for cpu before we can 1181 * begin anything. Make sure some other cpu on this 1182 * node has not already allocated this 1183 */ 1184 if (!cachep->nodelists[node]) { 1185 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1186 if (!l3) 1187 goto bad; 1188 kmem_list3_init(l3); 1189 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1190 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1191 1192 /* 1193 * The l3s don't come and go as CPUs come and 1194 * go. cache_chain_mutex is sufficient 1195 * protection here. 1196 */ 1197 cachep->nodelists[node] = l3; 1198 } 1199 1200 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1201 cachep->nodelists[node]->free_limit = 1202 (1 + nr_cpus_node(node)) * 1203 cachep->batchcount + cachep->num; 1204 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1205 } 1206 1207 /* 1208 * Now we can go ahead with allocating the shared arrays and 1209 * array caches 1210 */ 1211 list_for_each_entry(cachep, &cache_chain, next) { 1212 struct array_cache *nc; 1213 struct array_cache *shared = NULL; 1214 struct array_cache **alien = NULL; 1215 1216 nc = alloc_arraycache(node, cachep->limit, 1217 cachep->batchcount); 1218 if (!nc) 1219 goto bad; 1220 if (cachep->shared) { 1221 shared = alloc_arraycache(node, 1222 cachep->shared * cachep->batchcount, 1223 0xbaadf00d); 1224 if (!shared) 1225 goto bad; 1226 } 1227 if (use_alien_caches) { 1228 alien = alloc_alien_cache(node, cachep->limit); 1229 if (!alien) 1230 goto bad; 1231 } 1232 cachep->array[cpu] = nc; 1233 l3 = cachep->nodelists[node]; 1234 BUG_ON(!l3); 1235 1236 spin_lock_irq(&l3->list_lock); 1237 if (!l3->shared) { 1238 /* 1239 * We are serialised from CPU_DEAD or 1240 * CPU_UP_CANCELLED by the cpucontrol lock 1241 */ 1242 l3->shared = shared; 1243 shared = NULL; 1244 } 1245#ifdef CONFIG_NUMA 1246 if (!l3->alien) { 1247 l3->alien = alien; 1248 alien = NULL; 1249 } 1250#endif 1251 spin_unlock_irq(&l3->list_lock); 1252 kfree(shared); 1253 free_alien_cache(alien); 1254 } 1255 break; 1256 case CPU_ONLINE: 1257 case CPU_ONLINE_FROZEN: 1258 start_cpu_timer(cpu); 1259 break; 1260#ifdef CONFIG_HOTPLUG_CPU 1261 case CPU_DOWN_PREPARE: 1262 case CPU_DOWN_PREPARE_FROZEN: 1263 /* 1264 * Shutdown cache reaper. Note that the cache_chain_mutex is 1265 * held so that if cache_reap() is invoked it cannot do 1266 * anything expensive but will only modify reap_work 1267 * and reschedule the timer. 1268 */ 1269 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); 1270 /* Now the cache_reaper is guaranteed to be not running. */ 1271 per_cpu(reap_work, cpu).work.func = NULL; 1272 break; 1273 case CPU_DOWN_FAILED: 1274 case CPU_DOWN_FAILED_FROZEN: 1275 start_cpu_timer(cpu); 1276 break; 1277 case CPU_DEAD: 1278 case CPU_DEAD_FROZEN: 1279 /* 1280 * Even if all the cpus of a node are down, we don't free the 1281 * kmem_list3 of any cache. This to avoid a race between 1282 * cpu_down, and a kmalloc allocation from another cpu for 1283 * memory from the node of the cpu going down. The list3 1284 * structure is usually allocated from kmem_cache_create() and 1285 * gets destroyed at kmem_cache_destroy(). 1286 */ 1287 /* fall thru */ 1288#endif 1289 case CPU_UP_CANCELED: 1290 case CPU_UP_CANCELED_FROZEN: 1291 list_for_each_entry(cachep, &cache_chain, next) { 1292 struct array_cache *nc; 1293 struct array_cache *shared; 1294 struct array_cache **alien; 1295 cpumask_t mask; 1296 1297 mask = node_to_cpumask(node); 1298 /* cpu is dead; no one can alloc from it. */ 1299 nc = cachep->array[cpu]; 1300 cachep->array[cpu] = NULL; 1301 l3 = cachep->nodelists[node]; 1302 1303 if (!l3) 1304 goto free_array_cache; 1305 1306 spin_lock_irq(&l3->list_lock); 1307 1308 /* Free limit for this kmem_list3 */ 1309 l3->free_limit -= cachep->batchcount; 1310 if (nc) 1311 free_block(cachep, nc->entry, nc->avail, node); 1312 1313 if (!cpus_empty(mask)) { 1314 spin_unlock_irq(&l3->list_lock); 1315 goto free_array_cache; 1316 } 1317 1318 shared = l3->shared; 1319 if (shared) { 1320 free_block(cachep, shared->entry, 1321 shared->avail, node); 1322 l3->shared = NULL; 1323 } 1324 1325 alien = l3->alien; 1326 l3->alien = NULL; 1327 1328 spin_unlock_irq(&l3->list_lock); 1329 1330 kfree(shared); 1331 if (alien) { 1332 drain_alien_cache(cachep, alien); 1333 free_alien_cache(alien); 1334 } 1335free_array_cache: 1336 kfree(nc); 1337 } 1338 /* 1339 * In the previous loop, all the objects were freed to 1340 * the respective cache's slabs, now we can go ahead and 1341 * shrink each nodelist to its limit. 1342 */ 1343 list_for_each_entry(cachep, &cache_chain, next) { 1344 l3 = cachep->nodelists[node]; 1345 if (!l3) 1346 continue; 1347 drain_freelist(cachep, l3, l3->free_objects); 1348 } 1349 break; 1350 case CPU_LOCK_RELEASE: 1351 mutex_unlock(&cache_chain_mutex); 1352 break; 1353 } 1354 return NOTIFY_OK; 1355bad: 1356 return NOTIFY_BAD; 1357} 1358 1359static struct notifier_block __cpuinitdata cpucache_notifier = { 1360 &cpuup_callback, NULL, 0 1361}; 1362 1363/* 1364 * swap the static kmem_list3 with kmalloced memory 1365 */ 1366static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1367 int nodeid) 1368{ 1369 struct kmem_list3 *ptr; 1370 1371 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); 1372 BUG_ON(!ptr); 1373 1374 local_irq_disable(); 1375 memcpy(ptr, list, sizeof(struct kmem_list3)); 1376 /* 1377 * Do not assume that spinlocks can be initialized via memcpy: 1378 */ 1379 spin_lock_init(&ptr->list_lock); 1380 1381 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1382 cachep->nodelists[nodeid] = ptr; 1383 local_irq_enable(); 1384} 1385 1386/* 1387 * Initialisation. Called after the page allocator have been initialised and 1388 * before smp_init(). 1389 */ 1390void __init kmem_cache_init(void) 1391{ 1392 size_t left_over; 1393 struct cache_sizes *sizes; 1394 struct cache_names *names; 1395 int i; 1396 int order; 1397 int node; 1398 1399 if (num_possible_nodes() == 1) 1400 use_alien_caches = 0; 1401 1402 for (i = 0; i < NUM_INIT_LISTS; i++) { 1403 kmem_list3_init(&initkmem_list3[i]); 1404 if (i < MAX_NUMNODES) 1405 cache_cache.nodelists[i] = NULL; 1406 } 1407 1408 /* 1409 * Fragmentation resistance on low memory - only use bigger 1410 * page orders on machines with more than 32MB of memory. 1411 */ 1412 if (num_physpages > (32 << 20) >> PAGE_SHIFT) 1413 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1414 1415 /* Bootstrap is tricky, because several objects are allocated 1416 * from caches that do not exist yet: 1417 * 1) initialize the cache_cache cache: it contains the struct 1418 * kmem_cache structures of all caches, except cache_cache itself: 1419 * cache_cache is statically allocated. 1420 * Initially an __init data area is used for the head array and the 1421 * kmem_list3 structures, it's replaced with a kmalloc allocated 1422 * array at the end of the bootstrap. 1423 * 2) Create the first kmalloc cache. 1424 * The struct kmem_cache for the new cache is allocated normally. 1425 * An __init data area is used for the head array. 1426 * 3) Create the remaining kmalloc caches, with minimally sized 1427 * head arrays. 1428 * 4) Replace the __init data head arrays for cache_cache and the first 1429 * kmalloc cache with kmalloc allocated arrays. 1430 * 5) Replace the __init data for kmem_list3 for cache_cache and 1431 * the other cache's with kmalloc allocated memory. 1432 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1433 */ 1434 1435 node = numa_node_id(); 1436 1437 /* 1) create the cache_cache */ 1438 INIT_LIST_HEAD(&cache_chain); 1439 list_add(&cache_cache.next, &cache_chain); 1440 cache_cache.colour_off = cache_line_size(); 1441 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1442 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; 1443 1444 /* 1445 * struct kmem_cache size depends on nr_node_ids, which 1446 * can be less than MAX_NUMNODES. 1447 */ 1448 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + 1449 nr_node_ids * sizeof(struct kmem_list3 *); 1450#if DEBUG 1451 cache_cache.obj_size = cache_cache.buffer_size; 1452#endif 1453 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1454 cache_line_size()); 1455 cache_cache.reciprocal_buffer_size = 1456 reciprocal_value(cache_cache.buffer_size); 1457 1458 for (order = 0; order < MAX_ORDER; order++) { 1459 cache_estimate(order, cache_cache.buffer_size, 1460 cache_line_size(), 0, &left_over, &cache_cache.num); 1461 if (cache_cache.num) 1462 break; 1463 } 1464 BUG_ON(!cache_cache.num); 1465 cache_cache.gfporder = order; 1466 cache_cache.colour = left_over / cache_cache.colour_off; 1467 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1468 sizeof(struct slab), cache_line_size()); 1469 1470 /* 2+3) create the kmalloc caches */ 1471 sizes = malloc_sizes; 1472 names = cache_names; 1473 1474 /* 1475 * Initialize the caches that provide memory for the array cache and the 1476 * kmem_list3 structures first. Without this, further allocations will 1477 * bug. 1478 */ 1479 1480 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1481 sizes[INDEX_AC].cs_size, 1482 ARCH_KMALLOC_MINALIGN, 1483 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1484 NULL, NULL); 1485 1486 if (INDEX_AC != INDEX_L3) { 1487 sizes[INDEX_L3].cs_cachep = 1488 kmem_cache_create(names[INDEX_L3].name, 1489 sizes[INDEX_L3].cs_size, 1490 ARCH_KMALLOC_MINALIGN, 1491 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1492 NULL, NULL); 1493 } 1494 1495 slab_early_init = 0; 1496 1497 while (sizes->cs_size != ULONG_MAX) { 1498 /* 1499 * For performance, all the general caches are L1 aligned. 1500 * This should be particularly beneficial on SMP boxes, as it 1501 * eliminates "false sharing". 1502 * Note for systems short on memory removing the alignment will 1503 * allow tighter packing of the smaller caches. 1504 */ 1505 if (!sizes->cs_cachep) { 1506 sizes->cs_cachep = kmem_cache_create(names->name, 1507 sizes->cs_size, 1508 ARCH_KMALLOC_MINALIGN, 1509 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1510 NULL, NULL); 1511 } 1512#ifdef CONFIG_ZONE_DMA 1513 sizes->cs_dmacachep = kmem_cache_create( 1514 names->name_dma, 1515 sizes->cs_size, 1516 ARCH_KMALLOC_MINALIGN, 1517 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1518 SLAB_PANIC, 1519 NULL, NULL); 1520#endif 1521 sizes++; 1522 names++; 1523 } 1524 /* 4) Replace the bootstrap head arrays */ 1525 { 1526 struct array_cache *ptr; 1527 1528 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1529 1530 local_irq_disable(); 1531 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1532 memcpy(ptr, cpu_cache_get(&cache_cache), 1533 sizeof(struct arraycache_init)); 1534 /* 1535 * Do not assume that spinlocks can be initialized via memcpy: 1536 */ 1537 spin_lock_init(&ptr->lock); 1538 1539 cache_cache.array[smp_processor_id()] = ptr; 1540 local_irq_enable(); 1541 1542 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1543 1544 local_irq_disable(); 1545 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1546 != &initarray_generic.cache); 1547 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1548 sizeof(struct arraycache_init)); 1549 /* 1550 * Do not assume that spinlocks can be initialized via memcpy: 1551 */ 1552 spin_lock_init(&ptr->lock); 1553 1554 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1555 ptr; 1556 local_irq_enable(); 1557 } 1558 /* 5) Replace the bootstrap kmem_list3's */ 1559 { 1560 int nid; 1561 1562 /* Replace the static kmem_list3 structures for the boot cpu */ 1563 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); 1564 1565 for_each_online_node(nid) { 1566 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1567 &initkmem_list3[SIZE_AC + nid], nid); 1568 1569 if (INDEX_AC != INDEX_L3) { 1570 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1571 &initkmem_list3[SIZE_L3 + nid], nid); 1572 } 1573 } 1574 } 1575 1576 /* 6) resize the head arrays to their final sizes */ 1577 { 1578 struct kmem_cache *cachep; 1579 mutex_lock(&cache_chain_mutex); 1580 list_for_each_entry(cachep, &cache_chain, next) 1581 if (enable_cpucache(cachep)) 1582 BUG(); 1583 mutex_unlock(&cache_chain_mutex); 1584 } 1585 1586 /* Annotate slab for lockdep -- annotate the malloc caches */ 1587 init_lock_keys(); 1588 1589 1590 /* Done! */ 1591 g_cpucache_up = FULL; 1592 1593 /* 1594 * Register a cpu startup notifier callback that initializes 1595 * cpu_cache_get for all new cpus 1596 */ 1597 register_cpu_notifier(&cpucache_notifier); 1598 1599 /* 1600 * The reap timers are started later, with a module init call: That part 1601 * of the kernel is not yet operational. 1602 */ 1603} 1604 1605static int __init cpucache_init(void) 1606{ 1607 int cpu; 1608 1609 /* 1610 * Register the timers that return unneeded pages to the page allocator 1611 */ 1612 for_each_online_cpu(cpu) 1613 start_cpu_timer(cpu); 1614 return 0; 1615} 1616__initcall(cpucache_init); 1617 1618/* 1619 * Interface to system's page allocator. No need to hold the cache-lock. 1620 * 1621 * If we requested dmaable memory, we will get it. Even if we 1622 * did not request dmaable memory, we might get it, but that 1623 * would be relatively rare and ignorable. 1624 */ 1625static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1626{ 1627 struct page *page; 1628 int nr_pages; 1629 int i; 1630 1631#ifndef CONFIG_MMU 1632 /* 1633 * Nommu uses slab's for process anonymous memory allocations, and thus 1634 * requires __GFP_COMP to properly refcount higher order allocations 1635 */ 1636 flags |= __GFP_COMP; 1637#endif 1638 1639 flags |= cachep->gfpflags; 1640 1641 page = alloc_pages_node(nodeid, flags, cachep->gfporder); 1642 if (!page) 1643 return NULL; 1644 1645 nr_pages = (1 << cachep->gfporder); 1646 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1647 add_zone_page_state(page_zone(page), 1648 NR_SLAB_RECLAIMABLE, nr_pages); 1649 else 1650 add_zone_page_state(page_zone(page), 1651 NR_SLAB_UNRECLAIMABLE, nr_pages); 1652 for (i = 0; i < nr_pages; i++) 1653 __SetPageSlab(page + i); 1654 return page_address(page); 1655} 1656 1657/* 1658 * Interface to system's page release. 1659 */ 1660static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1661{ 1662 unsigned long i = (1 << cachep->gfporder); 1663 struct page *page = virt_to_page(addr); 1664 const unsigned long nr_freed = i; 1665 1666 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1667 sub_zone_page_state(page_zone(page), 1668 NR_SLAB_RECLAIMABLE, nr_freed); 1669 else 1670 sub_zone_page_state(page_zone(page), 1671 NR_SLAB_UNRECLAIMABLE, nr_freed); 1672 while (i--) { 1673 BUG_ON(!PageSlab(page)); 1674 __ClearPageSlab(page); 1675 page++; 1676 } 1677 if (current->reclaim_state) 1678 current->reclaim_state->reclaimed_slab += nr_freed; 1679 free_pages((unsigned long)addr, cachep->gfporder); 1680} 1681 1682static void kmem_rcu_free(struct rcu_head *head) 1683{ 1684 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1685 struct kmem_cache *cachep = slab_rcu->cachep; 1686 1687 kmem_freepages(cachep, slab_rcu->addr); 1688 if (OFF_SLAB(cachep)) 1689 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1690} 1691 1692#if DEBUG 1693 1694#ifdef CONFIG_DEBUG_PAGEALLOC 1695static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1696 unsigned long caller) 1697{ 1698 int size = obj_size(cachep); 1699 1700 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1701 1702 if (size < 5 * sizeof(unsigned long)) 1703 return; 1704 1705 *addr++ = 0x12345678; 1706 *addr++ = caller; 1707 *addr++ = smp_processor_id(); 1708 size -= 3 * sizeof(unsigned long); 1709 { 1710 unsigned long *sptr = &caller; 1711 unsigned long svalue; 1712 1713 while (!kstack_end(sptr)) { 1714 svalue = *sptr++; 1715 if (kernel_text_address(svalue)) { 1716 *addr++ = svalue; 1717 size -= sizeof(unsigned long); 1718 if (size <= sizeof(unsigned long)) 1719 break; 1720 } 1721 } 1722 1723 } 1724 *addr++ = 0x87654321; 1725} 1726#endif 1727 1728static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1729{ 1730 int size = obj_size(cachep); 1731 addr = &((char *)addr)[obj_offset(cachep)]; 1732 1733 memset(addr, val, size); 1734 *(unsigned char *)(addr + size - 1) = POISON_END; 1735} 1736 1737static void dump_line(char *data, int offset, int limit) 1738{ 1739 int i; 1740 unsigned char error = 0; 1741 int bad_count = 0; 1742 1743 printk(KERN_ERR "%03x:", offset); 1744 for (i = 0; i < limit; i++) { 1745 if (data[offset + i] != POISON_FREE) { 1746 error = data[offset + i]; 1747 bad_count++; 1748 } 1749 printk(" %02x", (unsigned char)data[offset + i]); 1750 } 1751 printk("\n"); 1752 1753 if (bad_count == 1) { 1754 error ^= POISON_FREE; 1755 if (!(error & (error - 1))) { 1756 printk(KERN_ERR "Single bit error detected. Probably " 1757 "bad RAM.\n"); 1758#ifdef CONFIG_X86 1759 printk(KERN_ERR "Run memtest86+ or a similar memory " 1760 "test tool.\n"); 1761#else 1762 printk(KERN_ERR "Run a memory test tool.\n"); 1763#endif 1764 } 1765 } 1766} 1767#endif 1768 1769#if DEBUG 1770 1771static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1772{ 1773 int i, size; 1774 char *realobj; 1775 1776 if (cachep->flags & SLAB_RED_ZONE) { 1777 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", 1778 *dbg_redzone1(cachep, objp), 1779 *dbg_redzone2(cachep, objp)); 1780 } 1781 1782 if (cachep->flags & SLAB_STORE_USER) { 1783 printk(KERN_ERR "Last user: [<%p>]", 1784 *dbg_userword(cachep, objp)); 1785 print_symbol("(%s)", 1786 (unsigned long)*dbg_userword(cachep, objp)); 1787 printk("\n"); 1788 } 1789 realobj = (char *)objp + obj_offset(cachep); 1790 size = obj_size(cachep); 1791 for (i = 0; i < size && lines; i += 16, lines--) { 1792 int limit; 1793 limit = 16; 1794 if (i + limit > size) 1795 limit = size - i; 1796 dump_line(realobj, i, limit); 1797 } 1798} 1799 1800static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1801{ 1802 char *realobj; 1803 int size, i; 1804 int lines = 0; 1805 1806 realobj = (char *)objp + obj_offset(cachep); 1807 size = obj_size(cachep); 1808 1809 for (i = 0; i < size; i++) { 1810 char exp = POISON_FREE; 1811 if (i == size - 1) 1812 exp = POISON_END; 1813 if (realobj[i] != exp) { 1814 int limit; 1815 /* Mismatch ! */ 1816 /* Print header */ 1817 if (lines == 0) { 1818 printk(KERN_ERR 1819 "Slab corruption: %s start=%p, len=%d\n", 1820 cachep->name, realobj, size); 1821 print_objinfo(cachep, objp, 0); 1822 } 1823 /* Hexdump the affected line */ 1824 i = (i / 16) * 16; 1825 limit = 16; 1826 if (i + limit > size) 1827 limit = size - i; 1828 dump_line(realobj, i, limit); 1829 i += 16; 1830 lines++; 1831 /* Limit to 5 lines */ 1832 if (lines > 5) 1833 break; 1834 } 1835 } 1836 if (lines != 0) { 1837 /* Print some data about the neighboring objects, if they 1838 * exist: 1839 */ 1840 struct slab *slabp = virt_to_slab(objp); 1841 unsigned int objnr; 1842 1843 objnr = obj_to_index(cachep, slabp, objp); 1844 if (objnr) { 1845 objp = index_to_obj(cachep, slabp, objnr - 1); 1846 realobj = (char *)objp + obj_offset(cachep); 1847 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1848 realobj, size); 1849 print_objinfo(cachep, objp, 2); 1850 } 1851 if (objnr + 1 < cachep->num) { 1852 objp = index_to_obj(cachep, slabp, objnr + 1); 1853 realobj = (char *)objp + obj_offset(cachep); 1854 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1855 realobj, size); 1856 print_objinfo(cachep, objp, 2); 1857 } 1858 } 1859} 1860#endif 1861 1862#if DEBUG 1863/** 1864 * slab_destroy_objs - destroy a slab and its objects 1865 * @cachep: cache pointer being destroyed 1866 * @slabp: slab pointer being destroyed 1867 * 1868 * Call the registered destructor for each object in a slab that is being 1869 * destroyed. 1870 */ 1871static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1872{ 1873 int i; 1874 for (i = 0; i < cachep->num; i++) { 1875 void *objp = index_to_obj(cachep, slabp, i); 1876 1877 if (cachep->flags & SLAB_POISON) { 1878#ifdef CONFIG_DEBUG_PAGEALLOC 1879 if (cachep->buffer_size % PAGE_SIZE == 0 && 1880 OFF_SLAB(cachep)) 1881 kernel_map_pages(virt_to_page(objp), 1882 cachep->buffer_size / PAGE_SIZE, 1); 1883 else 1884 check_poison_obj(cachep, objp); 1885#else 1886 check_poison_obj(cachep, objp); 1887#endif 1888 } 1889 if (cachep->flags & SLAB_RED_ZONE) { 1890 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1891 slab_error(cachep, "start of a freed object " 1892 "was overwritten"); 1893 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1894 slab_error(cachep, "end of a freed object " 1895 "was overwritten"); 1896 } 1897 } 1898} 1899#else 1900static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1901{ 1902} 1903#endif 1904 1905/** 1906 * slab_destroy - destroy and release all objects in a slab 1907 * @cachep: cache pointer being destroyed 1908 * @slabp: slab pointer being destroyed 1909 * 1910 * Destroy all the objs in a slab, and release the mem back to the system. 1911 * Before calling the slab must have been unlinked from the cache. The 1912 * cache-lock is not held/needed. 1913 */ 1914static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1915{ 1916 void *addr = slabp->s_mem - slabp->colouroff; 1917 1918 slab_destroy_objs(cachep, slabp); 1919 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1920 struct slab_rcu *slab_rcu; 1921 1922 slab_rcu = (struct slab_rcu *)slabp; 1923 slab_rcu->cachep = cachep; 1924 slab_rcu->addr = addr; 1925 call_rcu(&slab_rcu->head, kmem_rcu_free); 1926 } else { 1927 kmem_freepages(cachep, addr); 1928 if (OFF_SLAB(cachep)) 1929 kmem_cache_free(cachep->slabp_cache, slabp); 1930 } 1931} 1932 1933/* 1934 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1935 * size of kmem_list3. 1936 */ 1937static void __init set_up_list3s(struct kmem_cache *cachep, int index) 1938{ 1939 int node; 1940 1941 for_each_online_node(node) { 1942 cachep->nodelists[node] = &initkmem_list3[index + node]; 1943 cachep->nodelists[node]->next_reap = jiffies + 1944 REAPTIMEOUT_LIST3 + 1945 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1946 } 1947} 1948 1949static void __kmem_cache_destroy(struct kmem_cache *cachep) 1950{ 1951 int i; 1952 struct kmem_list3 *l3; 1953 1954 for_each_online_cpu(i) 1955 kfree(cachep->array[i]); 1956 1957 /* NUMA: free the list3 structures */ 1958 for_each_online_node(i) { 1959 l3 = cachep->nodelists[i]; 1960 if (l3) { 1961 kfree(l3->shared); 1962 free_alien_cache(l3->alien); 1963 kfree(l3); 1964 } 1965 } 1966 kmem_cache_free(&cache_cache, cachep); 1967} 1968 1969 1970/** 1971 * calculate_slab_order - calculate size (page order) of slabs 1972 * @cachep: pointer to the cache that is being created 1973 * @size: size of objects to be created in this cache. 1974 * @align: required alignment for the objects. 1975 * @flags: slab allocation flags 1976 * 1977 * Also calculates the number of objects per slab. 1978 * 1979 * This could be made much more intelligent. For now, try to avoid using 1980 * high order pages for slabs. When the gfp() functions are more friendly 1981 * towards high-order requests, this should be changed. 1982 */ 1983static size_t calculate_slab_order(struct kmem_cache *cachep, 1984 size_t size, size_t align, unsigned long flags) 1985{ 1986 unsigned long offslab_limit; 1987 size_t left_over = 0; 1988 int gfporder; 1989 1990 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { 1991 unsigned int num; 1992 size_t remainder; 1993 1994 cache_estimate(gfporder, size, align, flags, &remainder, &num); 1995 if (!num) 1996 continue; 1997 1998 if (flags & CFLGS_OFF_SLAB) { 1999 /* 2000 * Max number of objs-per-slab for caches which 2001 * use off-slab slabs. Needed to avoid a possible 2002 * looping condition in cache_grow(). 2003 */ 2004 offslab_limit = size - sizeof(struct slab); 2005 offslab_limit /= sizeof(kmem_bufctl_t); 2006 2007 if (num > offslab_limit) 2008 break; 2009 } 2010 2011 /* Found something acceptable - save it away */ 2012 cachep->num = num; 2013 cachep->gfporder = gfporder; 2014 left_over = remainder; 2015 2016 /* 2017 * A VFS-reclaimable slab tends to have most allocations 2018 * as GFP_NOFS and we really don't want to have to be allocating 2019 * higher-order pages when we are unable to shrink dcache. 2020 */ 2021 if (flags & SLAB_RECLAIM_ACCOUNT) 2022 break; 2023 2024 /* 2025 * Large number of objects is good, but very large slabs are 2026 * currently bad for the gfp()s. 2027 */ 2028 if (gfporder >= slab_break_gfp_order) 2029 break; 2030 2031 /* 2032 * Acceptable internal fragmentation? 2033 */ 2034 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 2035 break; 2036 } 2037 return left_over; 2038} 2039 2040static int __init_refok setup_cpu_cache(struct kmem_cache *cachep) 2041{ 2042 if (g_cpucache_up == FULL) 2043 return enable_cpucache(cachep); 2044 2045 if (g_cpucache_up == NONE) { 2046 /* 2047 * Note: the first kmem_cache_create must create the cache 2048 * that's used by kmalloc(24), otherwise the creation of 2049 * further caches will BUG(). 2050 */ 2051 cachep->array[smp_processor_id()] = &initarray_generic.cache; 2052 2053 /* 2054 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 2055 * the first cache, then we need to set up all its list3s, 2056 * otherwise the creation of further caches will BUG(). 2057 */ 2058 set_up_list3s(cachep, SIZE_AC); 2059 if (INDEX_AC == INDEX_L3) 2060 g_cpucache_up = PARTIAL_L3; 2061 else 2062 g_cpucache_up = PARTIAL_AC; 2063 } else { 2064 cachep->array[smp_processor_id()] = 2065 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 2066 2067 if (g_cpucache_up == PARTIAL_AC) { 2068 set_up_list3s(cachep, SIZE_L3); 2069 g_cpucache_up = PARTIAL_L3; 2070 } else { 2071 int node; 2072 for_each_online_node(node) { 2073 cachep->nodelists[node] = 2074 kmalloc_node(sizeof(struct kmem_list3), 2075 GFP_KERNEL, node); 2076 BUG_ON(!cachep->nodelists[node]); 2077 kmem_list3_init(cachep->nodelists[node]); 2078 } 2079 } 2080 } 2081 cachep->nodelists[numa_node_id()]->next_reap = 2082 jiffies + REAPTIMEOUT_LIST3 + 2083 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 2084 2085 cpu_cache_get(cachep)->avail = 0; 2086 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 2087 cpu_cache_get(cachep)->batchcount = 1; 2088 cpu_cache_get(cachep)->touched = 0; 2089 cachep->batchcount = 1; 2090 cachep->limit = BOOT_CPUCACHE_ENTRIES; 2091 return 0; 2092} 2093 2094/** 2095 * kmem_cache_create - Create a cache. 2096 * @name: A string which is used in /proc/slabinfo to identify this cache. 2097 * @size: The size of objects to be created in this cache. 2098 * @align: The required alignment for the objects. 2099 * @flags: SLAB flags 2100 * @ctor: A constructor for the objects. 2101 * @dtor: A destructor for the objects (not implemented anymore). 2102 * 2103 * Returns a ptr to the cache on success, NULL on failure. 2104 * Cannot be called within a int, but can be interrupted. 2105 * The @ctor is run when new pages are allocated by the cache 2106 * and the @dtor is run before the pages are handed back. 2107 * 2108 * @name must be valid until the cache is destroyed. This implies that 2109 * the module calling this has to destroy the cache before getting unloaded. 2110 * 2111 * The flags are 2112 * 2113 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 2114 * to catch references to uninitialised memory. 2115 * 2116 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 2117 * for buffer overruns. 2118 * 2119 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 2120 * cacheline. This can be beneficial if you're counting cycles as closely 2121 * as davem. 2122 */ 2123struct kmem_cache * 2124kmem_cache_create (const char *name, size_t size, size_t align, 2125 unsigned long flags, 2126 void (*ctor)(void*, struct kmem_cache *, unsigned long), 2127 void (*dtor)(void*, struct kmem_cache *, unsigned long)) 2128{ 2129 size_t left_over, slab_size, ralign; 2130 struct kmem_cache *cachep = NULL, *pc; 2131 2132 /* 2133 * Sanity checks... these are all serious usage bugs. 2134 */ 2135 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 2136 size > KMALLOC_MAX_SIZE || dtor) { 2137 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, 2138 name); 2139 BUG(); 2140 } 2141 2142 /* 2143 * We use cache_chain_mutex to ensure a consistent view of 2144 * cpu_online_map as well. Please see cpuup_callback 2145 */ 2146 mutex_lock(&cache_chain_mutex); 2147 2148 list_for_each_entry(pc, &cache_chain, next) { 2149 char tmp; 2150 int res; 2151 2152 /* 2153 * This happens when the module gets unloaded and doesn't 2154 * destroy its slab cache and no-one else reuses the vmalloc 2155 * area of the module. Print a warning. 2156 */ 2157 res = probe_kernel_address(pc->name, tmp); 2158 if (res) { 2159 printk(KERN_ERR 2160 "SLAB: cache with size %d has lost its name\n", 2161 pc->buffer_size); 2162 continue; 2163 } 2164 2165 if (!strcmp(pc->name, name)) { 2166 printk(KERN_ERR 2167 "kmem_cache_create: duplicate cache %s\n", name); 2168 dump_stack(); 2169 goto oops; 2170 } 2171 } 2172 2173#if DEBUG 2174 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 2175#if FORCED_DEBUG 2176 /* 2177 * Enable redzoning and last user accounting, except for caches with 2178 * large objects, if the increased size would increase the object size 2179 * above the next power of two: caches with object sizes just above a 2180 * power of two have a significant amount of internal fragmentation. 2181 */ 2182 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + 2183 2 * sizeof(unsigned long long))) 2184 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2185 if (!(flags & SLAB_DESTROY_BY_RCU)) 2186 flags |= SLAB_POISON; 2187#endif 2188 if (flags & SLAB_DESTROY_BY_RCU) 2189 BUG_ON(flags & SLAB_POISON); 2190#endif 2191 /* 2192 * Always checks flags, a caller might be expecting debug support which 2193 * isn't available. 2194 */ 2195 BUG_ON(flags & ~CREATE_MASK); 2196 2197 /* 2198 * Check that size is in terms of words. This is needed to avoid 2199 * unaligned accesses for some archs when redzoning is used, and makes 2200 * sure any on-slab bufctl's are also correctly aligned. 2201 */ 2202 if (size & (BYTES_PER_WORD - 1)) { 2203 size += (BYTES_PER_WORD - 1); 2204 size &= ~(BYTES_PER_WORD - 1); 2205 } 2206 2207 /* calculate the final buffer alignment: */ 2208 2209 /* 1) arch recommendation: can be overridden for debug */ 2210 if (flags & SLAB_HWCACHE_ALIGN) { 2211 /* 2212 * Default alignment: as specified by the arch code. Except if 2213 * an object is really small, then squeeze multiple objects into 2214 * one cacheline. 2215 */ 2216 ralign = cache_line_size(); 2217 while (size <= ralign / 2) 2218 ralign /= 2; 2219 } else { 2220 ralign = BYTES_PER_WORD; 2221 } 2222 2223 /* 2224 * Redzoning and user store require word alignment or possibly larger. 2225 * Note this will be overridden by architecture or caller mandated 2226 * alignment if either is greater than BYTES_PER_WORD. 2227 */ 2228 if (flags & SLAB_STORE_USER) 2229 ralign = BYTES_PER_WORD; 2230 2231 if (flags & SLAB_RED_ZONE) { 2232 ralign = REDZONE_ALIGN; 2233 /* If redzoning, ensure that the second redzone is suitably 2234 * aligned, by adjusting the object size accordingly. */ 2235 size += REDZONE_ALIGN - 1; 2236 size &= ~(REDZONE_ALIGN - 1); 2237 } 2238 2239 /* 2) arch mandated alignment */ 2240 if (ralign < ARCH_SLAB_MINALIGN) { 2241 ralign = ARCH_SLAB_MINALIGN; 2242 } 2243 /* 3) caller mandated alignment */ 2244 if (ralign < align) { 2245 ralign = align; 2246 } 2247 /* disable debug if necessary */ 2248 if (ralign > __alignof__(unsigned long long)) 2249 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2250 /* 2251 * 4) Store it. 2252 */ 2253 align = ralign; 2254 2255 /* Get cache's description obj. */ 2256 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); 2257 if (!cachep) 2258 goto oops; 2259 2260#if DEBUG 2261 cachep->obj_size = size; 2262 2263 /* 2264 * Both debugging options require word-alignment which is calculated 2265 * into align above. 2266 */ 2267 if (flags & SLAB_RED_ZONE) { 2268 /* add space for red zone words */ 2269 cachep->obj_offset += sizeof(unsigned long long); 2270 size += 2 * sizeof(unsigned long long); 2271 } 2272 if (flags & SLAB_STORE_USER) { 2273 /* user store requires one word storage behind the end of 2274 * the real object. But if the second red zone needs to be 2275 * aligned to 64 bits, we must allow that much space. 2276 */ 2277 if (flags & SLAB_RED_ZONE) 2278 size += REDZONE_ALIGN; 2279 else 2280 size += BYTES_PER_WORD; 2281 } 2282#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2283 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2284 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { 2285 cachep->obj_offset += PAGE_SIZE - size; 2286 size = PAGE_SIZE; 2287 } 2288#endif 2289#endif 2290 2291 /* 2292 * Determine if the slab management is 'on' or 'off' slab. 2293 * (bootstrapping cannot cope with offslab caches so don't do 2294 * it too early on.) 2295 */ 2296 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) 2297 /* 2298 * Size is large, assume best to place the slab management obj 2299 * off-slab (should allow better packing of objs). 2300 */ 2301 flags |= CFLGS_OFF_SLAB; 2302 2303 size = ALIGN(size, align); 2304 2305 left_over = calculate_slab_order(cachep, size, align, flags); 2306 2307 if (!cachep->num) { 2308 printk(KERN_ERR 2309 "kmem_cache_create: couldn't create cache %s.\n", name); 2310 kmem_cache_free(&cache_cache, cachep); 2311 cachep = NULL; 2312 goto oops; 2313 } 2314 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2315 + sizeof(struct slab), align); 2316 2317 /* 2318 * If the slab has been placed off-slab, and we have enough space then 2319 * move it on-slab. This is at the expense of any extra colouring. 2320 */ 2321 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2322 flags &= ~CFLGS_OFF_SLAB; 2323 left_over -= slab_size; 2324 } 2325 2326 if (flags & CFLGS_OFF_SLAB) { 2327 /* really off slab. No need for manual alignment */ 2328 slab_size = 2329 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2330 } 2331 2332 cachep->colour_off = cache_line_size(); 2333 /* Offset must be a multiple of the alignment. */ 2334 if (cachep->colour_off < align) 2335 cachep->colour_off = align; 2336 cachep->colour = left_over / cachep->colour_off; 2337 cachep->slab_size = slab_size; 2338 cachep->flags = flags; 2339 cachep->gfpflags = 0; 2340 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) 2341 cachep->gfpflags |= GFP_DMA; 2342 cachep->buffer_size = size; 2343 cachep->reciprocal_buffer_size = reciprocal_value(size); 2344 2345 if (flags & CFLGS_OFF_SLAB) { 2346 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2347 /* 2348 * This is a possibility for one of the malloc_sizes caches. 2349 * But since we go off slab only for object size greater than 2350 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, 2351 * this should not happen at all. 2352 * But leave a BUG_ON for some lucky dude. 2353 */ 2354 BUG_ON(!cachep->slabp_cache); 2355 } 2356 cachep->ctor = ctor; 2357 cachep->name = name; 2358 2359 if (setup_cpu_cache(cachep)) { 2360 __kmem_cache_destroy(cachep); 2361 cachep = NULL; 2362 goto oops; 2363 } 2364 2365 /* cache setup completed, link it into the list */ 2366 list_add(&cachep->next, &cache_chain); 2367oops: 2368 if (!cachep && (flags & SLAB_PANIC)) 2369 panic("kmem_cache_create(): failed to create slab `%s'\n", 2370 name); 2371 mutex_unlock(&cache_chain_mutex); 2372 return cachep; 2373} 2374EXPORT_SYMBOL(kmem_cache_create); 2375 2376#if DEBUG 2377static void check_irq_off(void) 2378{ 2379 BUG_ON(!irqs_disabled()); 2380} 2381 2382static void check_irq_on(void) 2383{ 2384 BUG_ON(irqs_disabled()); 2385} 2386 2387static void check_spinlock_acquired(struct kmem_cache *cachep) 2388{ 2389#ifdef CONFIG_SMP 2390 check_irq_off(); 2391 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); 2392#endif 2393} 2394 2395static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2396{ 2397#ifdef CONFIG_SMP 2398 check_irq_off(); 2399 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2400#endif 2401} 2402 2403#else 2404#define check_irq_off() do { } while(0) 2405#define check_irq_on() do { } while(0) 2406#define check_spinlock_acquired(x) do { } while(0) 2407#define check_spinlock_acquired_node(x, y) do { } while(0) 2408#endif 2409 2410static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2411 struct array_cache *ac, 2412 int force, int node); 2413 2414static void do_drain(void *arg) 2415{ 2416 struct kmem_cache *cachep = arg; 2417 struct array_cache *ac; 2418 int node = numa_node_id(); 2419 2420 check_irq_off(); 2421 ac = cpu_cache_get(cachep); 2422 spin_lock(&cachep->nodelists[node]->list_lock); 2423 free_block(cachep, ac->entry, ac->avail, node); 2424 spin_unlock(&cachep->nodelists[node]->list_lock); 2425 ac->avail = 0; 2426} 2427 2428static void drain_cpu_caches(struct kmem_cache *cachep) 2429{ 2430 struct kmem_list3 *l3; 2431 int node; 2432 2433 on_each_cpu(do_drain, cachep, 1, 1); 2434 check_irq_on(); 2435 for_each_online_node(node) { 2436 l3 = cachep->nodelists[node]; 2437 if (l3 && l3->alien) 2438 drain_alien_cache(cachep, l3->alien); 2439 } 2440 2441 for_each_online_node(node) { 2442 l3 = cachep->nodelists[node]; 2443 if (l3) 2444 drain_array(cachep, l3, l3->shared, 1, node); 2445 } 2446} 2447 2448/* 2449 * Remove slabs from the list of free slabs. 2450 * Specify the number of slabs to drain in tofree. 2451 * 2452 * Returns the actual number of slabs released. 2453 */ 2454static int drain_freelist(struct kmem_cache *cache, 2455 struct kmem_list3 *l3, int tofree) 2456{ 2457 struct list_head *p; 2458 int nr_freed; 2459 struct slab *slabp; 2460 2461 nr_freed = 0; 2462 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { 2463 2464 spin_lock_irq(&l3->list_lock); 2465 p = l3->slabs_free.prev; 2466 if (p == &l3->slabs_free) { 2467 spin_unlock_irq(&l3->list_lock); 2468 goto out; 2469 } 2470 2471 slabp = list_entry(p, struct slab, list); 2472#if DEBUG 2473 BUG_ON(slabp->inuse); 2474#endif 2475 list_del(&slabp->list); 2476 /* 2477 * Safe to drop the lock. The slab is no longer linked 2478 * to the cache. 2479 */ 2480 l3->free_objects -= cache->num; 2481 spin_unlock_irq(&l3->list_lock); 2482 slab_destroy(cache, slabp); 2483 nr_freed++; 2484 } 2485out: 2486 return nr_freed; 2487} 2488 2489/* Called with cache_chain_mutex held to protect against cpu hotplug */ 2490static int __cache_shrink(struct kmem_cache *cachep) 2491{ 2492 int ret = 0, i = 0; 2493 struct kmem_list3 *l3; 2494 2495 drain_cpu_caches(cachep); 2496 2497 check_irq_on(); 2498 for_each_online_node(i) { 2499 l3 = cachep->nodelists[i]; 2500 if (!l3) 2501 continue; 2502 2503 drain_freelist(cachep, l3, l3->free_objects); 2504 2505 ret += !list_empty(&l3->slabs_full) || 2506 !list_empty(&l3->slabs_partial); 2507 } 2508 return (ret ? 1 : 0); 2509} 2510 2511/** 2512 * kmem_cache_shrink - Shrink a cache. 2513 * @cachep: The cache to shrink. 2514 * 2515 * Releases as many slabs as possible for a cache. 2516 * To help debugging, a zero exit status indicates all slabs were released. 2517 */ 2518int kmem_cache_shrink(struct kmem_cache *cachep) 2519{ 2520 int ret; 2521 BUG_ON(!cachep || in_interrupt()); 2522 2523 mutex_lock(&cache_chain_mutex); 2524 ret = __cache_shrink(cachep); 2525 mutex_unlock(&cache_chain_mutex); 2526 return ret; 2527} 2528EXPORT_SYMBOL(kmem_cache_shrink); 2529 2530/** 2531 * kmem_cache_destroy - delete a cache 2532 * @cachep: the cache to destroy 2533 * 2534 * Remove a &struct kmem_cache object from the slab cache. 2535 * 2536 * It is expected this function will be called by a module when it is 2537 * unloaded. This will remove the cache completely, and avoid a duplicate 2538 * cache being allocated each time a module is loaded and unloaded, if the 2539 * module doesn't have persistent in-kernel storage across loads and unloads. 2540 * 2541 * The cache must be empty before calling this function. 2542 * 2543 * The caller must guarantee that noone will allocate memory from the cache 2544 * during the kmem_cache_destroy(). 2545 */ 2546void kmem_cache_destroy(struct kmem_cache *cachep) 2547{ 2548 BUG_ON(!cachep || in_interrupt()); 2549 2550 /* Find the cache in the chain of caches. */ 2551 mutex_lock(&cache_chain_mutex); 2552 /* 2553 * the chain is never empty, cache_cache is never destroyed 2554 */ 2555 list_del(&cachep->next); 2556 if (__cache_shrink(cachep)) { 2557 slab_error(cachep, "Can't free all objects"); 2558 list_add(&cachep->next, &cache_chain); 2559 mutex_unlock(&cache_chain_mutex); 2560 return; 2561 } 2562 2563 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2564 synchronize_rcu(); 2565 2566 __kmem_cache_destroy(cachep); 2567 mutex_unlock(&cache_chain_mutex); 2568} 2569EXPORT_SYMBOL(kmem_cache_destroy); 2570 2571/* 2572 * Get the memory for a slab management obj. 2573 * For a slab cache when the slab descriptor is off-slab, slab descriptors 2574 * always come from malloc_sizes caches. The slab descriptor cannot 2575 * come from the same cache which is getting created because, 2576 * when we are searching for an appropriate cache for these 2577 * descriptors in kmem_cache_create, we search through the malloc_sizes array. 2578 * If we are creating a malloc_sizes cache here it would not be visible to 2579 * kmem_find_general_cachep till the initialization is complete. 2580 * Hence we cannot have slabp_cache same as the original cache. 2581 */ 2582static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2583 int colour_off, gfp_t local_flags, 2584 int nodeid) 2585{ 2586 struct slab *slabp; 2587 2588 if (OFF_SLAB(cachep)) { 2589 /* Slab management obj is off-slab. */ 2590 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2591 local_flags & ~GFP_THISNODE, nodeid); 2592 if (!slabp) 2593 return NULL; 2594 } else { 2595 slabp = objp + colour_off; 2596 colour_off += cachep->slab_size; 2597 } 2598 slabp->inuse = 0; 2599 slabp->colouroff = colour_off; 2600 slabp->s_mem = objp + colour_off; 2601 slabp->nodeid = nodeid; 2602 return slabp; 2603} 2604 2605static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2606{ 2607 return (kmem_bufctl_t *) (slabp + 1); 2608} 2609 2610static void cache_init_objs(struct kmem_cache *cachep, 2611 struct slab *slabp) 2612{ 2613 int i; 2614 2615 for (i = 0; i < cachep->num; i++) { 2616 void *objp = index_to_obj(cachep, slabp, i); 2617#if DEBUG 2618 /* need to poison the objs? */ 2619 if (cachep->flags & SLAB_POISON) 2620 poison_obj(cachep, objp, POISON_FREE); 2621 if (cachep->flags & SLAB_STORE_USER) 2622 *dbg_userword(cachep, objp) = NULL; 2623 2624 if (cachep->flags & SLAB_RED_ZONE) { 2625 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2626 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2627 } 2628 /* 2629 * Constructors are not allowed to allocate memory from the same 2630 * cache which they are a constructor for. Otherwise, deadlock. 2631 * They must also be threaded. 2632 */ 2633 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2634 cachep->ctor(objp + obj_offset(cachep), cachep, 2635 0); 2636 2637 if (cachep->flags & SLAB_RED_ZONE) { 2638 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2639 slab_error(cachep, "constructor overwrote the" 2640 " end of an object"); 2641 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2642 slab_error(cachep, "constructor overwrote the" 2643 " start of an object"); 2644 } 2645 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2646 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2647 kernel_map_pages(virt_to_page(objp), 2648 cachep->buffer_size / PAGE_SIZE, 0); 2649#else 2650 if (cachep->ctor) 2651 cachep->ctor(objp, cachep, 0); 2652#endif 2653 slab_bufctl(slabp)[i] = i + 1; 2654 } 2655 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2656 slabp->free = 0; 2657} 2658 2659static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2660{ 2661 if (CONFIG_ZONE_DMA_FLAG) { 2662 if (flags & GFP_DMA) 2663 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2664 else 2665 BUG_ON(cachep->gfpflags & GFP_DMA); 2666 } 2667} 2668 2669static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2670 int nodeid) 2671{ 2672 void *objp = index_to_obj(cachep, slabp, slabp->free); 2673 kmem_bufctl_t next; 2674 2675 slabp->inuse++; 2676 next = slab_bufctl(slabp)[slabp->free]; 2677#if DEBUG 2678 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2679 WARN_ON(slabp->nodeid != nodeid); 2680#endif 2681 slabp->free = next; 2682 2683 return objp; 2684} 2685 2686static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2687 void *objp, int nodeid) 2688{ 2689 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2690 2691#if DEBUG 2692 /* Verify that the slab belongs to the intended node */ 2693 WARN_ON(slabp->nodeid != nodeid); 2694 2695 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2696 printk(KERN_ERR "slab: double free detected in cache " 2697 "'%s', objp %p\n", cachep->name, objp); 2698 BUG(); 2699 } 2700#endif 2701 slab_bufctl(slabp)[objnr] = slabp->free; 2702 slabp->free = objnr; 2703 slabp->inuse--; 2704} 2705 2706/* 2707 * Map pages beginning at addr to the given cache and slab. This is required 2708 * for the slab allocator to be able to lookup the cache and slab of a 2709 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. 2710 */ 2711static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, 2712 void *addr) 2713{ 2714 int nr_pages; 2715 struct page *page; 2716 2717 page = virt_to_page(addr); 2718 2719 nr_pages = 1; 2720 if (likely(!PageCompound(page))) 2721 nr_pages <<= cache->gfporder; 2722 2723 do { 2724 page_set_cache(page, cache); 2725 page_set_slab(page, slab); 2726 page++; 2727 } while (--nr_pages); 2728} 2729 2730/* 2731 * Grow (by 1) the number of slabs within a cache. This is called by 2732 * kmem_cache_alloc() when there are no active objs left in a cache. 2733 */ 2734static int cache_grow(struct kmem_cache *cachep, 2735 gfp_t flags, int nodeid, void *objp) 2736{ 2737 struct slab *slabp; 2738 size_t offset; 2739 gfp_t local_flags; 2740 struct kmem_list3 *l3; 2741 2742 /* 2743 * Be lazy and only check for valid flags here, keeping it out of the 2744 * critical path in kmem_cache_alloc(). 2745 */ 2746 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK)); 2747 2748 local_flags = (flags & GFP_LEVEL_MASK); 2749 /* Take the l3 list lock to change the colour_next on this node */ 2750 check_irq_off(); 2751 l3 = cachep->nodelists[nodeid]; 2752 spin_lock(&l3->list_lock); 2753 2754 /* Get colour for the slab, and cal the next value. */ 2755 offset = l3->colour_next; 2756 l3->colour_next++; 2757 if (l3->colour_next >= cachep->colour) 2758 l3->colour_next = 0; 2759 spin_unlock(&l3->list_lock); 2760 2761 offset *= cachep->colour_off; 2762 2763 if (local_flags & __GFP_WAIT) 2764 local_irq_enable(); 2765 2766 /* 2767 * The test for missing atomic flag is performed here, rather than 2768 * the more obvious place, simply to reduce the critical path length 2769 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2770 * will eventually be caught here (where it matters). 2771 */ 2772 kmem_flagcheck(cachep, flags); 2773 2774 /* 2775 * Get mem for the objs. Attempt to allocate a physical page from 2776 * 'nodeid'. 2777 */ 2778 if (!objp) 2779 objp = kmem_getpages(cachep, flags, nodeid); 2780 if (!objp) 2781 goto failed; 2782 2783 /* Get slab management. */ 2784 slabp = alloc_slabmgmt(cachep, objp, offset, 2785 local_flags & ~GFP_THISNODE, nodeid); 2786 if (!slabp) 2787 goto opps1; 2788 2789 slabp->nodeid = nodeid; 2790 slab_map_pages(cachep, slabp, objp); 2791 2792 cache_init_objs(cachep, slabp); 2793 2794 if (local_flags & __GFP_WAIT) 2795 local_irq_disable(); 2796 check_irq_off(); 2797 spin_lock(&l3->list_lock); 2798 2799 /* Make slab active. */ 2800 list_add_tail(&slabp->list, &(l3->slabs_free)); 2801 STATS_INC_GROWN(cachep); 2802 l3->free_objects += cachep->num; 2803 spin_unlock(&l3->list_lock); 2804 return 1; 2805opps1: 2806 kmem_freepages(cachep, objp); 2807failed: 2808 if (local_flags & __GFP_WAIT) 2809 local_irq_disable(); 2810 return 0; 2811} 2812 2813#if DEBUG 2814 2815/* 2816 * Perform extra freeing checks: 2817 * - detect bad pointers. 2818 * - POISON/RED_ZONE checking 2819 */ 2820static void kfree_debugcheck(const void *objp) 2821{ 2822 if (!virt_addr_valid(objp)) { 2823 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2824 (unsigned long)objp); 2825 BUG(); 2826 } 2827} 2828 2829static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2830{ 2831 unsigned long long redzone1, redzone2; 2832 2833 redzone1 = *dbg_redzone1(cache, obj); 2834 redzone2 = *dbg_redzone2(cache, obj); 2835 2836 /* 2837 * Redzone is ok. 2838 */ 2839 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2840 return; 2841 2842 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2843 slab_error(cache, "double free detected"); 2844 else 2845 slab_error(cache, "memory outside object was overwritten"); 2846 2847 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", 2848 obj, redzone1, redzone2); 2849} 2850 2851static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2852 void *caller) 2853{ 2854 struct page *page; 2855 unsigned int objnr; 2856 struct slab *slabp; 2857 2858 objp -= obj_offset(cachep); 2859 kfree_debugcheck(objp); 2860 page = virt_to_head_page(objp); 2861 2862 slabp = page_get_slab(page); 2863 2864 if (cachep->flags & SLAB_RED_ZONE) { 2865 verify_redzone_free(cachep, objp); 2866 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2867 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2868 } 2869 if (cachep->flags & SLAB_STORE_USER) 2870 *dbg_userword(cachep, objp) = caller; 2871 2872 objnr = obj_to_index(cachep, slabp, objp); 2873 2874 BUG_ON(objnr >= cachep->num); 2875 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2876 2877#ifdef CONFIG_DEBUG_SLAB_LEAK 2878 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2879#endif 2880 if (cachep->flags & SLAB_POISON) { 2881#ifdef CONFIG_DEBUG_PAGEALLOC 2882 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2883 store_stackinfo(cachep, objp, (unsigned long)caller); 2884 kernel_map_pages(virt_to_page(objp), 2885 cachep->buffer_size / PAGE_SIZE, 0); 2886 } else { 2887 poison_obj(cachep, objp, POISON_FREE); 2888 } 2889#else 2890 poison_obj(cachep, objp, POISON_FREE); 2891#endif 2892 } 2893 return objp; 2894} 2895 2896static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2897{ 2898 kmem_bufctl_t i; 2899 int entries = 0; 2900 2901 /* Check slab's freelist to see if this obj is there. */ 2902 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2903 entries++; 2904 if (entries > cachep->num || i >= cachep->num) 2905 goto bad; 2906 } 2907 if (entries != cachep->num - slabp->inuse) { 2908bad: 2909 printk(KERN_ERR "slab: Internal list corruption detected in " 2910 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2911 cachep->name, cachep->num, slabp, slabp->inuse); 2912 for (i = 0; 2913 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2914 i++) { 2915 if (i % 16 == 0) 2916 printk("\n%03x:", i); 2917 printk(" %02x", ((unsigned char *)slabp)[i]); 2918 } 2919 printk("\n"); 2920 BUG(); 2921 } 2922} 2923#else 2924#define kfree_debugcheck(x) do { } while(0) 2925#define cache_free_debugcheck(x,objp,z) (objp) 2926#define check_slabp(x,y) do { } while(0) 2927#endif 2928 2929static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2930{ 2931 int batchcount; 2932 struct kmem_list3 *l3; 2933 struct array_cache *ac; 2934 int node; 2935 2936 node = numa_node_id(); 2937 2938 check_irq_off(); 2939 ac = cpu_cache_get(cachep); 2940retry: 2941 batchcount = ac->batchcount; 2942 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2943 /* 2944 * If there was little recent activity on this cache, then 2945 * perform only a partial refill. Otherwise we could generate 2946 * refill bouncing. 2947 */ 2948 batchcount = BATCHREFILL_LIMIT; 2949 } 2950 l3 = cachep->nodelists[node]; 2951 2952 BUG_ON(ac->avail > 0 || !l3); 2953 spin_lock(&l3->list_lock); 2954 2955 /* See if we can refill from the shared array */ 2956 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) 2957 goto alloc_done; 2958 2959 while (batchcount > 0) { 2960 struct list_head *entry; 2961 struct slab *slabp; 2962 /* Get slab alloc is to come from. */ 2963 entry = l3->slabs_partial.next; 2964 if (entry == &l3->slabs_partial) { 2965 l3->free_touched = 1; 2966 entry = l3->slabs_free.next; 2967 if (entry == &l3->slabs_free) 2968 goto must_grow; 2969 } 2970 2971 slabp = list_entry(entry, struct slab, list); 2972 check_slabp(cachep, slabp); 2973 check_spinlock_acquired(cachep); 2974 2975 /* 2976 * The slab was either on partial or free list so 2977 * there must be at least one object available for 2978 * allocation. 2979 */ 2980 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num); 2981 2982 while (slabp->inuse < cachep->num && batchcount--) { 2983 STATS_INC_ALLOCED(cachep); 2984 STATS_INC_ACTIVE(cachep); 2985 STATS_SET_HIGH(cachep); 2986 2987 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 2988 node); 2989 } 2990 check_slabp(cachep, slabp); 2991 2992 /* move slabp to correct slabp list: */ 2993 list_del(&slabp->list); 2994 if (slabp->free == BUFCTL_END) 2995 list_add(&slabp->list, &l3->slabs_full); 2996 else 2997 list_add(&slabp->list, &l3->slabs_partial); 2998 } 2999 3000must_grow: 3001 l3->free_objects -= ac->avail; 3002alloc_done: 3003 spin_unlock(&l3->list_lock); 3004 3005 if (unlikely(!ac->avail)) { 3006 int x; 3007 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); 3008 3009 /* cache_grow can reenable interrupts, then ac could change. */ 3010 ac = cpu_cache_get(cachep); 3011 if (!x && ac->avail == 0) /* no objects in sight? abort */ 3012 return NULL; 3013 3014 if (!ac->avail) /* objects refilled by interrupt? */ 3015 goto retry; 3016 } 3017 ac->touched = 1; 3018 return ac->entry[--ac->avail]; 3019} 3020 3021static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3022 gfp_t flags) 3023{ 3024 might_sleep_if(flags & __GFP_WAIT); 3025#if DEBUG 3026 kmem_flagcheck(cachep, flags); 3027#endif 3028} 3029 3030#if DEBUG 3031static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3032 gfp_t flags, void *objp, void *caller) 3033{ 3034 if (!objp) 3035 return objp; 3036 if (cachep->flags & SLAB_POISON) { 3037#ifdef CONFIG_DEBUG_PAGEALLOC 3038 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 3039 kernel_map_pages(virt_to_page(objp), 3040 cachep->buffer_size / PAGE_SIZE, 1); 3041 else 3042 check_poison_obj(cachep, objp); 3043#else 3044 check_poison_obj(cachep, objp); 3045#endif 3046 poison_obj(cachep, objp, POISON_INUSE); 3047 } 3048 if (cachep->flags & SLAB_STORE_USER) 3049 *dbg_userword(cachep, objp) = caller; 3050 3051 if (cachep->flags & SLAB_RED_ZONE) { 3052 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3053 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3054 slab_error(cachep, "double free, or memory outside" 3055 " object was overwritten"); 3056 printk(KERN_ERR 3057 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", 3058 objp, *dbg_redzone1(cachep, objp), 3059 *dbg_redzone2(cachep, objp)); 3060 } 3061 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3062 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3063 } 3064#ifdef CONFIG_DEBUG_SLAB_LEAK 3065 { 3066 struct slab *slabp; 3067 unsigned objnr; 3068 3069 slabp = page_get_slab(virt_to_head_page(objp)); 3070 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 3071 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 3072 } 3073#endif 3074 objp += obj_offset(cachep); 3075 if (cachep->ctor && cachep->flags & SLAB_POISON) 3076 cachep->ctor(objp, cachep, 0); 3077#if ARCH_SLAB_MINALIGN 3078 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { 3079 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3080 objp, ARCH_SLAB_MINALIGN); 3081 } 3082#endif 3083 return objp; 3084} 3085#else 3086#define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3087#endif 3088 3089#ifdef CONFIG_FAILSLAB 3090 3091static struct failslab_attr { 3092 3093 struct fault_attr attr; 3094 3095 u32 ignore_gfp_wait; 3096#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3097 struct dentry *ignore_gfp_wait_file; 3098#endif 3099 3100} failslab = { 3101 .attr = FAULT_ATTR_INITIALIZER, 3102 .ignore_gfp_wait = 1, 3103}; 3104 3105static int __init setup_failslab(char *str) 3106{ 3107 return setup_fault_attr(&failslab.attr, str); 3108} 3109__setup("failslab=", setup_failslab); 3110 3111static int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3112{ 3113 if (cachep == &cache_cache) 3114 return 0; 3115 if (flags & __GFP_NOFAIL) 3116 return 0; 3117 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) 3118 return 0; 3119 3120 return should_fail(&failslab.attr, obj_size(cachep)); 3121} 3122 3123#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3124 3125static int __init failslab_debugfs(void) 3126{ 3127 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 3128 struct dentry *dir; 3129 int err; 3130 3131 err = init_fault_attr_dentries(&failslab.attr, "failslab"); 3132 if (err) 3133 return err; 3134 dir = failslab.attr.dentries.dir; 3135 3136 failslab.ignore_gfp_wait_file = 3137 debugfs_create_bool("ignore-gfp-wait", mode, dir, 3138 &failslab.ignore_gfp_wait); 3139 3140 if (!failslab.ignore_gfp_wait_file) { 3141 err = -ENOMEM; 3142 debugfs_remove(failslab.ignore_gfp_wait_file); 3143 cleanup_fault_attr_dentries(&failslab.attr); 3144 } 3145 3146 return err; 3147} 3148 3149late_initcall(failslab_debugfs); 3150 3151#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 3152 3153#else /* CONFIG_FAILSLAB */ 3154 3155static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3156{ 3157 return 0; 3158} 3159 3160#endif /* CONFIG_FAILSLAB */ 3161 3162static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3163{ 3164 void *objp; 3165 struct array_cache *ac; 3166 3167 check_irq_off(); 3168 3169 ac = cpu_cache_get(cachep); 3170 if (likely(ac->avail)) { 3171 STATS_INC_ALLOCHIT(cachep); 3172 ac->touched = 1; 3173 objp = ac->entry[--ac->avail]; 3174 } else { 3175 STATS_INC_ALLOCMISS(cachep); 3176 objp = cache_alloc_refill(cachep, flags); 3177 } 3178 return objp; 3179} 3180 3181#ifdef CONFIG_NUMA 3182/* 3183 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 3184 * 3185 * If we are in_interrupt, then process context, including cpusets and 3186 * mempolicy, may not apply and should not be used for allocation policy. 3187 */ 3188static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3189{ 3190 int nid_alloc, nid_here; 3191 3192 if (in_interrupt() || (flags & __GFP_THISNODE)) 3193 return NULL; 3194 nid_alloc = nid_here = numa_node_id(); 3195 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3196 nid_alloc = cpuset_mem_spread_node(); 3197 else if (current->mempolicy) 3198 nid_alloc = slab_node(current->mempolicy); 3199 if (nid_alloc != nid_here) 3200 return ____cache_alloc_node(cachep, flags, nid_alloc); 3201 return NULL; 3202} 3203 3204/* 3205 * Fallback function if there was no memory available and no objects on a 3206 * certain node and fall back is permitted. First we scan all the 3207 * available nodelists for available objects. If that fails then we 3208 * perform an allocation without specifying a node. This allows the page 3209 * allocator to do its reclaim / fallback magic. We then insert the 3210 * slab into the proper nodelist and then allocate from it. 3211 */ 3212static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3213{ 3214 struct zonelist *zonelist; 3215 gfp_t local_flags; 3216 struct zone **z; 3217 void *obj = NULL; 3218 int nid; 3219 3220 if (flags & __GFP_THISNODE) 3221 return NULL; 3222 3223 zonelist = &NODE_DATA(slab_node(current->mempolicy)) 3224 ->node_zonelists[gfp_zone(flags)]; 3225 local_flags = (flags & GFP_LEVEL_MASK); 3226 3227retry: 3228 /* 3229 * Look through allowed nodes for objects available 3230 * from existing per node queues. 3231 */ 3232 for (z = zonelist->zones; *z && !obj; z++) { 3233 nid = zone_to_nid(*z); 3234 3235 if (cpuset_zone_allowed_hardwall(*z, flags) && 3236 cache->nodelists[nid] && 3237 cache->nodelists[nid]->free_objects) 3238 obj = ____cache_alloc_node(cache, 3239 flags | GFP_THISNODE, nid); 3240 } 3241 3242 if (!obj) { 3243 /* 3244 * This allocation will be performed within the constraints 3245 * of the current cpuset / memory policy requirements. 3246 * We may trigger various forms of reclaim on the allowed 3247 * set and go into memory reserves if necessary. 3248 */ 3249 if (local_flags & __GFP_WAIT) 3250 local_irq_enable(); 3251 kmem_flagcheck(cache, flags); 3252 obj = kmem_getpages(cache, flags, -1); 3253 if (local_flags & __GFP_WAIT) 3254 local_irq_disable(); 3255 if (obj) { 3256 /* 3257 * Insert into the appropriate per node queues 3258 */ 3259 nid = page_to_nid(virt_to_page(obj)); 3260 if (cache_grow(cache, flags, nid, obj)) { 3261 obj = ____cache_alloc_node(cache, 3262 flags | GFP_THISNODE, nid); 3263 if (!obj) 3264 /* 3265 * Another processor may allocate the 3266 * objects in the slab since we are 3267 * not holding any locks. 3268 */ 3269 goto retry; 3270 } else { 3271 /* cache_grow already freed obj */ 3272 obj = NULL; 3273 } 3274 } 3275 } 3276 return obj; 3277} 3278 3279/* 3280 * A interface to enable slab creation on nodeid 3281 */ 3282static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3283 int nodeid) 3284{ 3285 struct list_head *entry; 3286 struct slab *slabp; 3287 struct kmem_list3 *l3; 3288 void *obj; 3289 int x; 3290 3291 l3 = cachep->nodelists[nodeid]; 3292 BUG_ON(!l3); 3293 3294retry: 3295 check_irq_off(); 3296 spin_lock(&l3->list_lock); 3297 entry = l3->slabs_partial.next; 3298 if (entry == &l3->slabs_partial) { 3299 l3->free_touched = 1; 3300 entry = l3->slabs_free.next; 3301 if (entry == &l3->slabs_free) 3302 goto must_grow; 3303 } 3304 3305 slabp = list_entry(entry, struct slab, list); 3306 check_spinlock_acquired_node(cachep, nodeid); 3307 check_slabp(cachep, slabp); 3308 3309 STATS_INC_NODEALLOCS(cachep); 3310 STATS_INC_ACTIVE(cachep); 3311 STATS_SET_HIGH(cachep); 3312 3313 BUG_ON(slabp->inuse == cachep->num); 3314 3315 obj = slab_get_obj(cachep, slabp, nodeid); 3316 check_slabp(cachep, slabp); 3317 l3->free_objects--; 3318 /* move slabp to correct slabp list: */ 3319 list_del(&slabp->list); 3320 3321 if (slabp->free == BUFCTL_END) 3322 list_add(&slabp->list, &l3->slabs_full); 3323 else 3324 list_add(&slabp->list, &l3->slabs_partial); 3325 3326 spin_unlock(&l3->list_lock); 3327 goto done; 3328 3329must_grow: 3330 spin_unlock(&l3->list_lock); 3331 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); 3332 if (x) 3333 goto retry; 3334 3335 return fallback_alloc(cachep, flags); 3336 3337done: 3338 return obj; 3339} 3340 3341/** 3342 * kmem_cache_alloc_node - Allocate an object on the specified node 3343 * @cachep: The cache to allocate from. 3344 * @flags: See kmalloc(). 3345 * @nodeid: node number of the target node. 3346 * @caller: return address of caller, used for debug information 3347 * 3348 * Identical to kmem_cache_alloc but it will allocate memory on the given 3349 * node, which can improve the performance for cpu bound structures. 3350 * 3351 * Fallback to other node is possible if __GFP_THISNODE is not set. 3352 */ 3353static __always_inline void * 3354__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, 3355 void *caller) 3356{ 3357 unsigned long save_flags; 3358 void *ptr; 3359 3360 if (should_failslab(cachep, flags)) 3361 return NULL; 3362 3363 cache_alloc_debugcheck_before(cachep, flags); 3364 local_irq_save(save_flags); 3365 3366 if (unlikely(nodeid == -1)) 3367 nodeid = numa_node_id(); 3368 3369 if (unlikely(!cachep->nodelists[nodeid])) { 3370 /* Node not bootstrapped yet */ 3371 ptr = fallback_alloc(cachep, flags); 3372 goto out; 3373 } 3374 3375 if (nodeid == numa_node_id()) { 3376 /* 3377 * Use the locally cached objects if possible. 3378 * However ____cache_alloc does not allow fallback 3379 * to other nodes. It may fail while we still have 3380 * objects on other nodes available. 3381 */ 3382 ptr = ____cache_alloc(cachep, flags); 3383 if (ptr) 3384 goto out; 3385 } 3386 /* ___cache_alloc_node can fall back to other nodes */ 3387 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3388 out: 3389 local_irq_restore(save_flags); 3390 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3391 3392 return ptr; 3393} 3394 3395static __always_inline void * 3396__do_cache_alloc(struct kmem_cache *cache, gfp_t flags) 3397{ 3398 void *objp; 3399 3400 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { 3401 objp = alternate_node_alloc(cache, flags); 3402 if (objp) 3403 goto out; 3404 } 3405 objp = ____cache_alloc(cache, flags); 3406 3407 /* 3408 * We may just have run out of memory on the local node. 3409 * ____cache_alloc_node() knows how to locate memory on other nodes 3410 */ 3411 if (!objp) 3412 objp = ____cache_alloc_node(cache, flags, numa_node_id()); 3413 3414 out: 3415 return objp; 3416} 3417#else 3418 3419static __always_inline void * 3420__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3421{ 3422 return ____cache_alloc(cachep, flags); 3423} 3424 3425#endif /* CONFIG_NUMA */ 3426 3427static __always_inline void * 3428__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) 3429{ 3430 unsigned long save_flags; 3431 void *objp; 3432 3433 if (should_failslab(cachep, flags)) 3434 return NULL; 3435 3436 cache_alloc_debugcheck_before(cachep, flags); 3437 local_irq_save(save_flags); 3438 objp = __do_cache_alloc(cachep, flags); 3439 local_irq_restore(save_flags); 3440 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); 3441 prefetchw(objp); 3442 3443 return objp; 3444} 3445 3446/* 3447 * Caller needs to acquire correct kmem_list's list_lock 3448 */ 3449static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 3450 int node) 3451{ 3452 int i; 3453 struct kmem_list3 *l3; 3454 3455 for (i = 0; i < nr_objects; i++) { 3456 void *objp = objpp[i]; 3457 struct slab *slabp; 3458 3459 slabp = virt_to_slab(objp); 3460 l3 = cachep->nodelists[node]; 3461 list_del(&slabp->list); 3462 check_spinlock_acquired_node(cachep, node); 3463 check_slabp(cachep, slabp); 3464 slab_put_obj(cachep, slabp, objp, node); 3465 STATS_DEC_ACTIVE(cachep); 3466 l3->free_objects++; 3467 check_slabp(cachep, slabp); 3468 3469 /* fixup slab chains */ 3470 if (slabp->inuse == 0) { 3471 if (l3->free_objects > l3->free_limit) { 3472 l3->free_objects -= cachep->num; 3473 /* No need to drop any previously held 3474 * lock here, even if we have a off-slab slab 3475 * descriptor it is guaranteed to come from 3476 * a different cache, refer to comments before 3477 * alloc_slabmgmt. 3478 */ 3479 slab_destroy(cachep, slabp); 3480 } else { 3481 list_add(&slabp->list, &l3->slabs_free); 3482 } 3483 } else { 3484 /* Unconditionally move a slab to the end of the 3485 * partial list on free - maximum time for the 3486 * other objects to be freed, too. 3487 */ 3488 list_add_tail(&slabp->list, &l3->slabs_partial); 3489 } 3490 } 3491} 3492 3493static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3494{ 3495 int batchcount; 3496 struct kmem_list3 *l3; 3497 int node = numa_node_id(); 3498 3499 batchcount = ac->batchcount; 3500#if DEBUG 3501 BUG_ON(!batchcount || batchcount > ac->avail); 3502#endif 3503 check_irq_off(); 3504 l3 = cachep->nodelists[node]; 3505 spin_lock(&l3->list_lock); 3506 if (l3->shared) { 3507 struct array_cache *shared_array = l3->shared; 3508 int max = shared_array->limit - shared_array->avail; 3509 if (max) { 3510 if (batchcount > max) 3511 batchcount = max; 3512 memcpy(&(shared_array->entry[shared_array->avail]), 3513 ac->entry, sizeof(void *) * batchcount); 3514 shared_array->avail += batchcount; 3515 goto free_done; 3516 } 3517 } 3518 3519 free_block(cachep, ac->entry, batchcount, node); 3520free_done: 3521#if STATS 3522 { 3523 int i = 0; 3524 struct list_head *p; 3525 3526 p = l3->slabs_free.next; 3527 while (p != &(l3->slabs_free)) { 3528 struct slab *slabp; 3529 3530 slabp = list_entry(p, struct slab, list); 3531 BUG_ON(slabp->inuse); 3532 3533 i++; 3534 p = p->next; 3535 } 3536 STATS_SET_FREEABLE(cachep, i); 3537 } 3538#endif 3539 spin_unlock(&l3->list_lock); 3540 ac->avail -= batchcount; 3541 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3542} 3543 3544/* 3545 * Release an obj back to its cache. If the obj has a constructed state, it must 3546 * be in this state _before_ it is released. Called with disabled ints. 3547 */ 3548static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3549{ 3550 struct array_cache *ac = cpu_cache_get(cachep); 3551 3552 check_irq_off(); 3553 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3554 3555 if (cache_free_alien(cachep, objp)) 3556 return; 3557 3558 if (likely(ac->avail < ac->limit)) { 3559 STATS_INC_FREEHIT(cachep); 3560 ac->entry[ac->avail++] = objp; 3561 return; 3562 } else { 3563 STATS_INC_FREEMISS(cachep); 3564 cache_flusharray(cachep, ac); 3565 ac->entry[ac->avail++] = objp; 3566 } 3567} 3568 3569/** 3570 * kmem_cache_alloc - Allocate an object 3571 * @cachep: The cache to allocate from. 3572 * @flags: See kmalloc(). 3573 * 3574 * Allocate an object from this cache. The flags are only relevant 3575 * if the cache has no available objects. 3576 */ 3577void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3578{ 3579 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3580} 3581EXPORT_SYMBOL(kmem_cache_alloc); 3582 3583/** 3584 * kmem_cache_zalloc - Allocate an object. The memory is set to zero. 3585 * @cache: The cache to allocate from. 3586 * @flags: See kmalloc(). 3587 * 3588 * Allocate an object from this cache and set the allocated memory to zero. 3589 * The flags are only relevant if the cache has no available objects. 3590 */ 3591void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) 3592{ 3593 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); 3594 if (ret) 3595 memset(ret, 0, obj_size(cache)); 3596 return ret; 3597} 3598EXPORT_SYMBOL(kmem_cache_zalloc); 3599 3600/** 3601 * kmem_ptr_validate - check if an untrusted pointer might 3602 * be a slab entry. 3603 * @cachep: the cache we're checking against 3604 * @ptr: pointer to validate 3605 * 3606 * This verifies that the untrusted pointer looks sane: 3607 * it is _not_ a guarantee that the pointer is actually 3608 * part of the slab cache in question, but it at least 3609 * validates that the pointer can be dereferenced and 3610 * looks half-way sane. 3611 * 3612 * Currently only used for dentry validation. 3613 */ 3614int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) 3615{ 3616 unsigned long addr = (unsigned long)ptr; 3617 unsigned long min_addr = PAGE_OFFSET; 3618 unsigned long align_mask = BYTES_PER_WORD - 1; 3619 unsigned long size = cachep->buffer_size; 3620 struct page *page; 3621 3622 if (unlikely(addr < min_addr)) 3623 goto out; 3624 if (unlikely(addr > (unsigned long)high_memory - size)) 3625 goto out; 3626 if (unlikely(addr & align_mask)) 3627 goto out; 3628 if (unlikely(!kern_addr_valid(addr))) 3629 goto out; 3630 if (unlikely(!kern_addr_valid(addr + size - 1))) 3631 goto out; 3632 page = virt_to_page(ptr); 3633 if (unlikely(!PageSlab(page))) 3634 goto out; 3635 if (unlikely(page_get_cache(page) != cachep)) 3636 goto out; 3637 return 1; 3638out: 3639 return 0; 3640} 3641 3642#ifdef CONFIG_NUMA 3643void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3644{ 3645 return __cache_alloc_node(cachep, flags, nodeid, 3646 __builtin_return_address(0)); 3647} 3648EXPORT_SYMBOL(kmem_cache_alloc_node); 3649 3650static __always_inline void * 3651__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) 3652{ 3653 struct kmem_cache *cachep; 3654 3655 cachep = kmem_find_general_cachep(size, flags); 3656 if (unlikely(cachep == NULL)) 3657 return NULL; 3658 return kmem_cache_alloc_node(cachep, flags, node); 3659} 3660 3661#ifdef CONFIG_DEBUG_SLAB 3662void *__kmalloc_node(size_t size, gfp_t flags, int node) 3663{ 3664 return __do_kmalloc_node(size, flags, node, 3665 __builtin_return_address(0)); 3666} 3667EXPORT_SYMBOL(__kmalloc_node); 3668 3669void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3670 int node, void *caller) 3671{ 3672 return __do_kmalloc_node(size, flags, node, caller); 3673} 3674EXPORT_SYMBOL(__kmalloc_node_track_caller); 3675#else 3676void *__kmalloc_node(size_t size, gfp_t flags, int node) 3677{ 3678 return __do_kmalloc_node(size, flags, node, NULL); 3679} 3680EXPORT_SYMBOL(__kmalloc_node); 3681#endif /* CONFIG_DEBUG_SLAB */ 3682#endif /* CONFIG_NUMA */ 3683 3684/** 3685 * __do_kmalloc - allocate memory 3686 * @size: how many bytes of memory are required. 3687 * @flags: the type of memory to allocate (see kmalloc). 3688 * @caller: function caller for debug tracking of the caller 3689 */ 3690static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3691 void *caller) 3692{ 3693 struct kmem_cache *cachep; 3694 3695 /* If you want to save a few bytes .text space: replace 3696 * __ with kmem_. 3697 * Then kmalloc uses the uninlined functions instead of the inline 3698 * functions. 3699 */ 3700 cachep = __find_general_cachep(size, flags); 3701 if (unlikely(cachep == NULL)) 3702 return NULL; 3703 return __cache_alloc(cachep, flags, caller); 3704} 3705 3706 3707#ifdef CONFIG_DEBUG_SLAB 3708void *__kmalloc(size_t size, gfp_t flags) 3709{ 3710 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3711} 3712EXPORT_SYMBOL(__kmalloc); 3713 3714void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) 3715{ 3716 return __do_kmalloc(size, flags, caller); 3717} 3718EXPORT_SYMBOL(__kmalloc_track_caller); 3719 3720#else 3721void *__kmalloc(size_t size, gfp_t flags) 3722{ 3723 return __do_kmalloc(size, flags, NULL); 3724} 3725EXPORT_SYMBOL(__kmalloc); 3726#endif 3727 3728/** 3729 * krealloc - reallocate memory. The contents will remain unchanged. 3730 * @p: object to reallocate memory for. 3731 * @new_size: how many bytes of memory are required. 3732 * @flags: the type of memory to allocate. 3733 * 3734 * The contents of the object pointed to are preserved up to the 3735 * lesser of the new and old sizes. If @p is %NULL, krealloc() 3736 * behaves exactly like kmalloc(). If @size is 0 and @p is not a 3737 * %NULL pointer, the object pointed to is freed. 3738 */ 3739void *krealloc(const void *p, size_t new_size, gfp_t flags) 3740{ 3741 struct kmem_cache *cache, *new_cache; 3742 void *ret; 3743 3744 if (unlikely(!p)) 3745 return kmalloc_track_caller(new_size, flags); 3746 3747 if (unlikely(!new_size)) { 3748 kfree(p); 3749 return NULL; 3750 } 3751 3752 cache = virt_to_cache(p); 3753 new_cache = __find_general_cachep(new_size, flags); 3754 3755 /* 3756 * If new size fits in the current cache, bail out. 3757 */ 3758 if (likely(cache == new_cache)) 3759 return (void *)p; 3760 3761 /* 3762 * We are on the slow-path here so do not use __cache_alloc 3763 * because it bloats kernel text. 3764 */ 3765 ret = kmalloc_track_caller(new_size, flags); 3766 if (ret) { 3767 memcpy(ret, p, min(new_size, ksize(p))); 3768 kfree(p); 3769 } 3770 return ret; 3771} 3772EXPORT_SYMBOL(krealloc); 3773 3774/** 3775 * kmem_cache_free - Deallocate an object 3776 * @cachep: The cache the allocation was from. 3777 * @objp: The previously allocated object. 3778 * 3779 * Free an object which was previously allocated from this 3780 * cache. 3781 */ 3782void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3783{ 3784 unsigned long flags; 3785 3786 BUG_ON(virt_to_cache(objp) != cachep); 3787 3788 local_irq_save(flags); 3789 debug_check_no_locks_freed(objp, obj_size(cachep)); 3790 __cache_free(cachep, objp); 3791 local_irq_restore(flags); 3792} 3793EXPORT_SYMBOL(kmem_cache_free); 3794 3795/** 3796 * kfree - free previously allocated memory 3797 * @objp: pointer returned by kmalloc. 3798 * 3799 * If @objp is NULL, no operation is performed. 3800 * 3801 * Don't free memory not originally allocated by kmalloc() 3802 * or you will run into trouble. 3803 */ 3804void kfree(const void *objp) 3805{ 3806 struct kmem_cache *c; 3807 unsigned long flags; 3808 3809 if (unlikely(!objp)) 3810 return; 3811 local_irq_save(flags); 3812 kfree_debugcheck(objp); 3813 c = virt_to_cache(objp); 3814 debug_check_no_locks_freed(objp, obj_size(c)); 3815 __cache_free(c, (void *)objp); 3816 local_irq_restore(flags); 3817} 3818EXPORT_SYMBOL(kfree); 3819 3820unsigned int kmem_cache_size(struct kmem_cache *cachep) 3821{ 3822 return obj_size(cachep); 3823} 3824EXPORT_SYMBOL(kmem_cache_size); 3825 3826const char *kmem_cache_name(struct kmem_cache *cachep) 3827{ 3828 return cachep->name; 3829} 3830EXPORT_SYMBOL_GPL(kmem_cache_name); 3831 3832/* 3833 * This initializes kmem_list3 or resizes varioius caches for all nodes. 3834 */ 3835static int alloc_kmemlist(struct kmem_cache *cachep) 3836{ 3837 int node; 3838 struct kmem_list3 *l3; 3839 struct array_cache *new_shared; 3840 struct array_cache **new_alien = NULL; 3841 3842 for_each_online_node(node) { 3843 3844 if (use_alien_caches) { 3845 new_alien = alloc_alien_cache(node, cachep->limit); 3846 if (!new_alien) 3847 goto fail; 3848 } 3849 3850 new_shared = NULL; 3851 if (cachep->shared) { 3852 new_shared = alloc_arraycache(node, 3853 cachep->shared*cachep->batchcount, 3854 0xbaadf00d); 3855 if (!new_shared) { 3856 free_alien_cache(new_alien); 3857 goto fail; 3858 } 3859 } 3860 3861 l3 = cachep->nodelists[node]; 3862 if (l3) { 3863 struct array_cache *shared = l3->shared; 3864 3865 spin_lock_irq(&l3->list_lock); 3866 3867 if (shared) 3868 free_block(cachep, shared->entry, 3869 shared->avail, node); 3870 3871 l3->shared = new_shared; 3872 if (!l3->alien) { 3873 l3->alien = new_alien; 3874 new_alien = NULL; 3875 } 3876 l3->free_limit = (1 + nr_cpus_node(node)) * 3877 cachep->batchcount + cachep->num; 3878 spin_unlock_irq(&l3->list_lock); 3879 kfree(shared); 3880 free_alien_cache(new_alien); 3881 continue; 3882 } 3883 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); 3884 if (!l3) { 3885 free_alien_cache(new_alien); 3886 kfree(new_shared); 3887 goto fail; 3888 } 3889 3890 kmem_list3_init(l3); 3891 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3892 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3893 l3->shared = new_shared; 3894 l3->alien = new_alien; 3895 l3->free_limit = (1 + nr_cpus_node(node)) * 3896 cachep->batchcount + cachep->num; 3897 cachep->nodelists[node] = l3; 3898 } 3899 return 0; 3900 3901fail: 3902 if (!cachep->next.next) { 3903 /* Cache is not active yet. Roll back what we did */ 3904 node--; 3905 while (node >= 0) { 3906 if (cachep->nodelists[node]) { 3907 l3 = cachep->nodelists[node]; 3908 3909 kfree(l3->shared); 3910 free_alien_cache(l3->alien); 3911 kfree(l3); 3912 cachep->nodelists[node] = NULL; 3913 } 3914 node--; 3915 } 3916 } 3917 return -ENOMEM; 3918} 3919 3920struct ccupdate_struct { 3921 struct kmem_cache *cachep; 3922 struct array_cache *new[NR_CPUS]; 3923}; 3924 3925static void do_ccupdate_local(void *info) 3926{ 3927 struct ccupdate_struct *new = info; 3928 struct array_cache *old; 3929 3930 check_irq_off(); 3931 old = cpu_cache_get(new->cachep); 3932 3933 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3934 new->new[smp_processor_id()] = old; 3935} 3936 3937/* Always called with the cache_chain_mutex held */ 3938static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3939 int batchcount, int shared) 3940{ 3941 struct ccupdate_struct *new; 3942 int i; 3943 3944 new = kzalloc(sizeof(*new), GFP_KERNEL); 3945 if (!new) 3946 return -ENOMEM; 3947 3948 for_each_online_cpu(i) { 3949 new->new[i] = alloc_arraycache(cpu_to_node(i), limit, 3950 batchcount); 3951 if (!new->new[i]) { 3952 for (i--; i >= 0; i--) 3953 kfree(new->new[i]); 3954 kfree(new); 3955 return -ENOMEM; 3956 } 3957 } 3958 new->cachep = cachep; 3959 3960 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); 3961 3962 check_irq_on(); 3963 cachep->batchcount = batchcount; 3964 cachep->limit = limit; 3965 cachep->shared = shared; 3966 3967 for_each_online_cpu(i) { 3968 struct array_cache *ccold = new->new[i]; 3969 if (!ccold) 3970 continue; 3971 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3972 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 3973 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3974 kfree(ccold); 3975 } 3976 kfree(new); 3977 return alloc_kmemlist(cachep); 3978} 3979 3980/* Called with cache_chain_mutex held always */ 3981static int enable_cpucache(struct kmem_cache *cachep) 3982{ 3983 int err; 3984 int limit, shared; 3985 3986 /* 3987 * The head array serves three purposes: 3988 * - create a LIFO ordering, i.e. return objects that are cache-warm 3989 * - reduce the number of spinlock operations. 3990 * - reduce the number of linked list operations on the slab and 3991 * bufctl chains: array operations are cheaper. 3992 * The numbers are guessed, we should auto-tune as described by 3993 * Bonwick. 3994 */ 3995 if (cachep->buffer_size > 131072) 3996 limit = 1; 3997 else if (cachep->buffer_size > PAGE_SIZE) 3998 limit = 8; 3999 else if (cachep->buffer_size > 1024) 4000 limit = 24; 4001 else if (cachep->buffer_size > 256) 4002 limit = 54; 4003 else 4004 limit = 120; 4005 4006 /* 4007 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 4008 * allocation behaviour: Most allocs on one cpu, most free operations 4009 * on another cpu. For these cases, an efficient object passing between 4010 * cpus is necessary. This is provided by a shared array. The array 4011 * replaces Bonwick's magazine layer. 4012 * On uniprocessor, it's functionally equivalent (but less efficient) 4013 * to a larger limit. Thus disabled by default. 4014 */ 4015 shared = 0; 4016 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) 4017 shared = 8; 4018 4019#if DEBUG 4020 /* 4021 * With debugging enabled, large batchcount lead to excessively long 4022 * periods with disabled local interrupts. Limit the batchcount 4023 */ 4024 if (limit > 32) 4025 limit = 32; 4026#endif 4027 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); 4028 if (err) 4029 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 4030 cachep->name, -err); 4031 return err; 4032} 4033 4034/* 4035 * Drain an array if it contains any elements taking the l3 lock only if 4036 * necessary. Note that the l3 listlock also protects the array_cache 4037 * if drain_array() is used on the shared array. 4038 */ 4039void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 4040 struct array_cache *ac, int force, int node) 4041{ 4042 int tofree; 4043 4044 if (!ac || !ac->avail) 4045 return; 4046 if (ac->touched && !force) { 4047 ac->touched = 0; 4048 } else { 4049 spin_lock_irq(&l3->list_lock); 4050 if (ac->avail) { 4051 tofree = force ? ac->avail : (ac->limit + 4) / 5; 4052 if (tofree > ac->avail) 4053 tofree = (ac->avail + 1) / 2; 4054 free_block(cachep, ac->entry, tofree, node); 4055 ac->avail -= tofree; 4056 memmove(ac->entry, &(ac->entry[tofree]), 4057 sizeof(void *) * ac->avail); 4058 } 4059 spin_unlock_irq(&l3->list_lock); 4060 } 4061} 4062 4063/** 4064 * cache_reap - Reclaim memory from caches. 4065 * @w: work descriptor 4066 * 4067 * Called from workqueue/eventd every few seconds. 4068 * Purpose: 4069 * - clear the per-cpu caches for this CPU. 4070 * - return freeable pages to the main free memory pool. 4071 * 4072 * If we cannot acquire the cache chain mutex then just give up - we'll try 4073 * again on the next iteration. 4074 */ 4075static void cache_reap(struct work_struct *w) 4076{ 4077 struct kmem_cache *searchp; 4078 struct kmem_list3 *l3; 4079 int node = numa_node_id(); 4080 struct delayed_work *work = 4081 container_of(w, struct delayed_work, work); 4082 4083 if (!mutex_trylock(&cache_chain_mutex)) 4084 /* Give up. Setup the next iteration. */ 4085 goto out; 4086 4087 list_for_each_entry(searchp, &cache_chain, next) { 4088 check_irq_on(); 4089 4090 /* 4091 * We only take the l3 lock if absolutely necessary and we 4092 * have established with reasonable certainty that 4093 * we can do some work if the lock was obtained. 4094 */ 4095 l3 = searchp->nodelists[node]; 4096 4097 reap_alien(searchp, l3); 4098 4099 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 4100 4101 /* 4102 * These are racy checks but it does not matter 4103 * if we skip one check or scan twice. 4104 */ 4105 if (time_after(l3->next_reap, jiffies)) 4106 goto next; 4107 4108 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 4109 4110 drain_array(searchp, l3, l3->shared, 0, node); 4111 4112 if (l3->free_touched) 4113 l3->free_touched = 0; 4114 else { 4115 int freed; 4116 4117 freed = drain_freelist(searchp, l3, (l3->free_limit + 4118 5 * searchp->num - 1) / (5 * searchp->num)); 4119 STATS_ADD_REAPED(searchp, freed); 4120 } 4121next: 4122 cond_resched(); 4123 } 4124 check_irq_on(); 4125 mutex_unlock(&cache_chain_mutex); 4126 next_reap_node(); 4127out: 4128 /* Set up the next iteration */ 4129 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); 4130} 4131 4132#ifdef CONFIG_PROC_FS 4133 4134static void print_slabinfo_header(struct seq_file *m) 4135{ 4136 /* 4137 * Output format version, so at least we can change it 4138 * without _too_ many complaints. 4139 */ 4140#if STATS 4141 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 4142#else 4143 seq_puts(m, "slabinfo - version: 2.1\n"); 4144#endif 4145 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4146 "<objperslab> <pagesperslab>"); 4147 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4148 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4149#if STATS 4150 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 4151 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 4152 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 4153#endif 4154 seq_putc(m, '\n'); 4155} 4156 4157static void *s_start(struct seq_file *m, loff_t *pos) 4158{ 4159 loff_t n = *pos; 4160 4161 mutex_lock(&cache_chain_mutex); 4162 if (!n) 4163 print_slabinfo_header(m); 4164 4165 return seq_list_start(&cache_chain, *pos); 4166} 4167 4168static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4169{ 4170 return seq_list_next(p, &cache_chain, pos); 4171} 4172 4173static void s_stop(struct seq_file *m, void *p) 4174{ 4175 mutex_unlock(&cache_chain_mutex); 4176} 4177 4178static int s_show(struct seq_file *m, void *p) 4179{ 4180 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); 4181 struct slab *slabp; 4182 unsigned long active_objs; 4183 unsigned long num_objs; 4184 unsigned long active_slabs = 0; 4185 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 4186 const char *name; 4187 char *error = NULL; 4188 int node; 4189 struct kmem_list3 *l3; 4190 4191 active_objs = 0; 4192 num_slabs = 0; 4193 for_each_online_node(node) { 4194 l3 = cachep->nodelists[node]; 4195 if (!l3) 4196 continue; 4197 4198 check_irq_on(); 4199 spin_lock_irq(&l3->list_lock); 4200 4201 list_for_each_entry(slabp, &l3->slabs_full, list) { 4202 if (slabp->inuse != cachep->num && !error) 4203 error = "slabs_full accounting error"; 4204 active_objs += cachep->num; 4205 active_slabs++; 4206 } 4207 list_for_each_entry(slabp, &l3->slabs_partial, list) { 4208 if (slabp->inuse == cachep->num && !error) 4209 error = "slabs_partial inuse accounting error"; 4210 if (!slabp->inuse && !error) 4211 error = "slabs_partial/inuse accounting error"; 4212 active_objs += slabp->inuse; 4213 active_slabs++; 4214 } 4215 list_for_each_entry(slabp, &l3->slabs_free, list) { 4216 if (slabp->inuse && !error) 4217 error = "slabs_free/inuse accounting error"; 4218 num_slabs++; 4219 } 4220 free_objects += l3->free_objects; 4221 if (l3->shared) 4222 shared_avail += l3->shared->avail; 4223 4224 spin_unlock_irq(&l3->list_lock); 4225 } 4226 num_slabs += active_slabs; 4227 num_objs = num_slabs * cachep->num; 4228 if (num_objs - active_objs != free_objects && !error) 4229 error = "free_objects accounting error"; 4230 4231 name = cachep->name; 4232 if (error) 4233 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 4234 4235 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 4236 name, active_objs, num_objs, cachep->buffer_size, 4237 cachep->num, (1 << cachep->gfporder)); 4238 seq_printf(m, " : tunables %4u %4u %4u", 4239 cachep->limit, cachep->batchcount, cachep->shared); 4240 seq_printf(m, " : slabdata %6lu %6lu %6lu", 4241 active_slabs, num_slabs, shared_avail); 4242#if STATS 4243 { /* list3 stats */ 4244 unsigned long high = cachep->high_mark; 4245 unsigned long allocs = cachep->num_allocations; 4246 unsigned long grown = cachep->grown; 4247 unsigned long reaped = cachep->reaped; 4248 unsigned long errors = cachep->errors; 4249 unsigned long max_freeable = cachep->max_freeable; 4250 unsigned long node_allocs = cachep->node_allocs; 4251 unsigned long node_frees = cachep->node_frees; 4252 unsigned long overflows = cachep->node_overflow; 4253 4254 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ 4255 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, 4256 reaped, errors, max_freeable, node_allocs, 4257 node_frees, overflows); 4258 } 4259 /* cpu stats */ 4260 { 4261 unsigned long allochit = atomic_read(&cachep->allochit); 4262 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4263 unsigned long freehit = atomic_read(&cachep->freehit); 4264 unsigned long freemiss = atomic_read(&cachep->freemiss); 4265 4266 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4267 allochit, allocmiss, freehit, freemiss); 4268 } 4269#endif 4270 seq_putc(m, '\n'); 4271 return 0; 4272} 4273 4274/* 4275 * slabinfo_op - iterator that generates /proc/slabinfo 4276 * 4277 * Output layout: 4278 * cache-name 4279 * num-active-objs 4280 * total-objs 4281 * object size 4282 * num-active-slabs 4283 * total-slabs 4284 * num-pages-per-slab 4285 * + further values on SMP and with statistics enabled 4286 */ 4287 4288const struct seq_operations slabinfo_op = { 4289 .start = s_start, 4290 .next = s_next, 4291 .stop = s_stop, 4292 .show = s_show, 4293}; 4294 4295#define MAX_SLABINFO_WRITE 128 4296/** 4297 * slabinfo_write - Tuning for the slab allocator 4298 * @file: unused 4299 * @buffer: user buffer 4300 * @count: data length 4301 * @ppos: unused 4302 */ 4303ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4304 size_t count, loff_t *ppos) 4305{ 4306 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4307 int limit, batchcount, shared, res; 4308 struct kmem_cache *cachep; 4309 4310 if (count > MAX_SLABINFO_WRITE) 4311 return -EINVAL; 4312 if (copy_from_user(&kbuf, buffer, count)) 4313 return -EFAULT; 4314 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4315 4316 tmp = strchr(kbuf, ' '); 4317 if (!tmp) 4318 return -EINVAL; 4319 *tmp = '\0'; 4320 tmp++; 4321 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4322 return -EINVAL; 4323 4324 /* Find the cache in the chain of caches. */ 4325 mutex_lock(&cache_chain_mutex); 4326 res = -EINVAL; 4327 list_for_each_entry(cachep, &cache_chain, next) { 4328 if (!strcmp(cachep->name, kbuf)) { 4329 if (limit < 1 || batchcount < 1 || 4330 batchcount > limit || shared < 0) { 4331 res = 0; 4332 } else { 4333 res = do_tune_cpucache(cachep, limit, 4334 batchcount, shared); 4335 } 4336 break; 4337 } 4338 } 4339 mutex_unlock(&cache_chain_mutex); 4340 if (res >= 0) 4341 res = count; 4342 return res; 4343} 4344 4345#ifdef CONFIG_DEBUG_SLAB_LEAK 4346 4347static void *leaks_start(struct seq_file *m, loff_t *pos) 4348{ 4349 mutex_lock(&cache_chain_mutex); 4350 return seq_list_start(&cache_chain, *pos); 4351} 4352 4353static inline int add_caller(unsigned long *n, unsigned long v) 4354{ 4355 unsigned long *p; 4356 int l; 4357 if (!v) 4358 return 1; 4359 l = n[1]; 4360 p = n + 2; 4361 while (l) { 4362 int i = l/2; 4363 unsigned long *q = p + 2 * i; 4364 if (*q == v) { 4365 q[1]++; 4366 return 1; 4367 } 4368 if (*q > v) { 4369 l = i; 4370 } else { 4371 p = q + 2; 4372 l -= i + 1; 4373 } 4374 } 4375 if (++n[1] == n[0]) 4376 return 0; 4377 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4378 p[0] = v; 4379 p[1] = 1; 4380 return 1; 4381} 4382 4383static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4384{ 4385 void *p; 4386 int i; 4387 if (n[0] == n[1]) 4388 return; 4389 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4390 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4391 continue; 4392 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4393 return; 4394 } 4395} 4396 4397static void show_symbol(struct seq_file *m, unsigned long address) 4398{ 4399#ifdef CONFIG_KALLSYMS 4400 unsigned long offset, size; 4401 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1]; 4402 4403 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { 4404 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4405 if (modname[0]) 4406 seq_printf(m, " [%s]", modname); 4407 return; 4408 } 4409#endif 4410 seq_printf(m, "%p", (void *)address); 4411} 4412 4413static int leaks_show(struct seq_file *m, void *p) 4414{ 4415 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); 4416 struct slab *slabp; 4417 struct kmem_list3 *l3; 4418 const char *name; 4419 unsigned long *n = m->private; 4420 int node; 4421 int i; 4422 4423 if (!(cachep->flags & SLAB_STORE_USER)) 4424 return 0; 4425 if (!(cachep->flags & SLAB_RED_ZONE)) 4426 return 0; 4427 4428 /* OK, we can do it */ 4429 4430 n[1] = 0; 4431 4432 for_each_online_node(node) { 4433 l3 = cachep->nodelists[node]; 4434 if (!l3) 4435 continue; 4436 4437 check_irq_on(); 4438 spin_lock_irq(&l3->list_lock); 4439 4440 list_for_each_entry(slabp, &l3->slabs_full, list) 4441 handle_slab(n, cachep, slabp); 4442 list_for_each_entry(slabp, &l3->slabs_partial, list) 4443 handle_slab(n, cachep, slabp); 4444 spin_unlock_irq(&l3->list_lock); 4445 } 4446 name = cachep->name; 4447 if (n[0] == n[1]) { 4448 /* Increase the buffer size */ 4449 mutex_unlock(&cache_chain_mutex); 4450 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4451 if (!m->private) { 4452 /* Too bad, we are really out */ 4453 m->private = n; 4454 mutex_lock(&cache_chain_mutex); 4455 return -ENOMEM; 4456 } 4457 *(unsigned long *)m->private = n[0] * 2; 4458 kfree(n); 4459 mutex_lock(&cache_chain_mutex); 4460 /* Now make sure this entry will be retried */ 4461 m->count = m->size; 4462 return 0; 4463 } 4464 for (i = 0; i < n[1]; i++) { 4465 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4466 show_symbol(m, n[2*i+2]); 4467 seq_putc(m, '\n'); 4468 } 4469 4470 return 0; 4471} 4472 4473const struct seq_operations slabstats_op = { 4474 .start = leaks_start, 4475 .next = s_next, 4476 .stop = s_stop, 4477 .show = leaks_show, 4478}; 4479#endif 4480#endif 4481 4482/** 4483 * ksize - get the actual amount of memory allocated for a given object 4484 * @objp: Pointer to the object 4485 * 4486 * kmalloc may internally round up allocations and return more memory 4487 * than requested. ksize() can be used to determine the actual amount of 4488 * memory allocated. The caller may use this additional memory, even though 4489 * a smaller amount of memory was initially specified with the kmalloc call. 4490 * The caller must guarantee that objp points to a valid object previously 4491 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4492 * must not be freed during the duration of the call. 4493 */ 4494size_t ksize(const void *objp) 4495{ 4496 if (unlikely(objp == NULL)) 4497 return 0; 4498 4499 return obj_size(virt_to_cache(objp)); 4500} 4501