1 /* 2 * linux/mm/slab.c 3 * Written by Mark Hemment, 1996/97. 4 * (markhe@nextd.demon.co.uk) 5 * 6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 7 * 8 * Major cleanup, different bufctl logic, per-cpu arrays 9 * (c) 2000 Manfred Spraul 10 * 11 * Cleanup, make the head arrays unconditional, preparation for NUMA 12 * (c) 2002 Manfred Spraul 13 * 14 * An implementation of the Slab Allocator as described in outline in; 15 * UNIX Internals: The New Frontiers by Uresh Vahalia 16 * Pub: Prentice Hall ISBN 0-13-101908-2 17 * or with a little more detail in; 18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 19 * Jeff Bonwick (Sun Microsystems). 20 * Presented at: USENIX Summer 1994 Technical Conference 21 * 22 * The memory is organized in caches, one cache for each object type. 23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 24 * Each cache consists out of many slabs (they are small (usually one 25 * page long) and always contiguous), and each slab contains multiple 26 * initialized objects. 27 * 28 * This means, that your constructor is used only for newly allocated 29 * slabs and you must pass objects with the same intializations to 30 * kmem_cache_free. 31 * 32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 33 * normal). If you need a special memory type, then must create a new 34 * cache for that memory type. 35 * 36 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 37 * full slabs with 0 free objects 38 * partial slabs 39 * empty slabs with no allocated objects 40 * 41 * If partial slabs exist, then new allocations come from these slabs, 42 * otherwise from empty slabs or new slabs are allocated. 43 * 44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 46 * 47 * Each cache has a short per-cpu head array, most allocs 48 * and frees go into that array, and if that array overflows, then 1/2 49 * of the entries in the array are given back into the global cache. 50 * The head array is strictly LIFO and should improve the cache hit rates. 51 * On SMP, it additionally reduces the spinlock operations. 52 * 53 * The c_cpuarray may not be read with enabled local interrupts - 54 * it's changed with a smp_call_function(). 55 * 56 * SMP synchronization: 57 * constructors and destructors are called without any locking. 58 * Several members in struct kmem_cache and struct slab never change, they 59 * are accessed without any locking. 60 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 61 * and local interrupts are disabled so slab code is preempt-safe. 62 * The non-constant members are protected with a per-cache irq spinlock. 63 * 64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 65 * in 2000 - many ideas in the current implementation are derived from 66 * his patch. 67 * 68 * Further notes from the original documentation: 69 * 70 * 11 April '97. Started multi-threading - markhe 71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'. 72 * The sem is only needed when accessing/extending the cache-chain, which 73 * can never happen inside an interrupt (kmem_cache_create(), 74 * kmem_cache_shrink() and kmem_cache_reap()). 75 * 76 * At present, each engine can be growing a cache. This should be blocked. 77 * 78 * 15 March 2005. NUMA slab allocator. 79 * Shai Fultheim <shai@scalex86.org>. 80 * Shobhit Dayal <shobhit@calsoftinc.com> 81 * Alok N Kataria <alokk@calsoftinc.com> 82 * Christoph Lameter <christoph@lameter.com> 83 * 84 * Modified the slab allocator to be node aware on NUMA systems. 85 * Each node has its own list of partial, free and full slabs. 86 * All object allocations for a node occur from node specific slab lists. 87 */ 88 89 #include <linux/slab.h> 90 #include <linux/mm.h> 91 #include <linux/poison.h> 92 #include <linux/swap.h> 93 #include <linux/cache.h> 94 #include <linux/interrupt.h> 95 #include <linux/init.h> 96 #include <linux/compiler.h> 97 #include <linux/cpuset.h> 98 #include <linux/seq_file.h> 99 #include <linux/notifier.h> 100 #include <linux/kallsyms.h> 101 #include <linux/cpu.h> 102 #include <linux/sysctl.h> 103 #include <linux/module.h> 104 #include <linux/rcupdate.h> 105 #include <linux/string.h> 106 #include <linux/uaccess.h> 107 #include <linux/nodemask.h> 108 #include <linux/mempolicy.h> 109 #include <linux/mutex.h> 110 #include <linux/fault-inject.h> 111 #include <linux/rtmutex.h> 112 #include <linux/reciprocal_div.h> 113 114 #include <asm/cacheflush.h> 115 #include <asm/tlbflush.h> 116 #include <asm/page.h> 117 118 /* 119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. 120 * 0 for faster, smaller code (especially in the critical paths). 121 * 122 * STATS - 1 to collect stats for /proc/slabinfo. 123 * 0 for faster, smaller code (especially in the critical paths). 124 * 125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 126 */ 127 128 #ifdef CONFIG_DEBUG_SLAB 129 #define DEBUG 1 130 #define STATS 1 131 #define FORCED_DEBUG 1 132 #else 133 #define DEBUG 0 134 #define STATS 0 135 #define FORCED_DEBUG 0 136 #endif 137 138 /* Shouldn't this be in a header file somewhere? */ 139 #define BYTES_PER_WORD sizeof(void *) 140 141 #ifndef cache_line_size 142 #define cache_line_size() L1_CACHE_BYTES 143 #endif 144 145 #ifndef ARCH_KMALLOC_MINALIGN 146 /* 147 * Enforce a minimum alignment for the kmalloc caches. 148 * Usually, the kmalloc caches are cache_line_size() aligned, except when 149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. 150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed 151 * alignment larger than the alignment of a 64-bit integer. 152 * ARCH_KMALLOC_MINALIGN allows that. 153 * Note that increasing this value may disable some debug features. 154 */ 155 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 156 #endif 157 158 #ifndef ARCH_SLAB_MINALIGN 159 /* 160 * Enforce a minimum alignment for all caches. 161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD 162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. 163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables 164 * some debug features. 165 */ 166 #define ARCH_SLAB_MINALIGN 0 167 #endif 168 169 #ifndef ARCH_KMALLOC_FLAGS 170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 171 #endif 172 173 /* Legal flag mask for kmem_cache_create(). */ 174 #if DEBUG 175 # define CREATE_MASK (SLAB_RED_ZONE | \ 176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \ 177 SLAB_CACHE_DMA | \ 178 SLAB_STORE_USER | \ 179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 181 #else 182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ 183 SLAB_CACHE_DMA | \ 184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ 185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) 186 #endif 187 188 /* 189 * kmem_bufctl_t: 190 * 191 * Bufctl's are used for linking objs within a slab 192 * linked offsets. 193 * 194 * This implementation relies on "struct page" for locating the cache & 195 * slab an object belongs to. 196 * This allows the bufctl structure to be small (one int), but limits 197 * the number of objects a slab (not a cache) can contain when off-slab 198 * bufctls are used. The limit is the size of the largest general cache 199 * that does not use off-slab slabs. 200 * For 32bit archs with 4 kB pages, is this 56. 201 * This is not serious, as it is only for large objects, when it is unwise 202 * to have too many per slab. 203 * Note: This limit can be raised by introducing a general cache whose size 204 * is less than 512 (PAGE_SIZE<<3), but greater than 256. 205 */ 206 207 typedef unsigned int kmem_bufctl_t; 208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) 209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) 210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) 211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) 212 213 /* 214 * struct slab 215 * 216 * Manages the objs in a slab. Placed either at the beginning of mem allocated 217 * for a slab, or allocated from an general cache. 218 * Slabs are chained into three list: fully used, partial, fully free slabs. 219 */ 220 struct slab { 221 struct list_head list; 222 unsigned long colouroff; 223 void *s_mem; /* including colour offset */ 224 unsigned int inuse; /* num of objs active in slab */ 225 kmem_bufctl_t free; 226 unsigned short nodeid; 227 }; 228 229 /* 230 * struct slab_rcu 231 * 232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to 233 * arrange for kmem_freepages to be called via RCU. This is useful if 234 * we need to approach a kernel structure obliquely, from its address 235 * obtained without the usual locking. We can lock the structure to 236 * stabilize it and check it's still at the given address, only if we 237 * can be sure that the memory has not been meanwhile reused for some 238 * other kind of object (which our subsystem's lock might corrupt). 239 * 240 * rcu_read_lock before reading the address, then rcu_read_unlock after 241 * taking the spinlock within the structure expected at that address. 242 * 243 * We assume struct slab_rcu can overlay struct slab when destroying. 244 */ 245 struct slab_rcu { 246 struct rcu_head head; 247 struct kmem_cache *cachep; 248 void *addr; 249 }; 250 251 /* 252 * struct array_cache 253 * 254 * Purpose: 255 * - LIFO ordering, to hand out cache-warm objects from _alloc 256 * - reduce the number of linked list operations 257 * - reduce spinlock operations 258 * 259 * The limit is stored in the per-cpu structure to reduce the data cache 260 * footprint. 261 * 262 */ 263 struct array_cache { 264 unsigned int avail; 265 unsigned int limit; 266 unsigned int batchcount; 267 unsigned int touched; 268 spinlock_t lock; 269 void *entry[0]; /* 270 * Must have this definition in here for the proper 271 * alignment of array_cache. Also simplifies accessing 272 * the entries. 273 * [0] is for gcc 2.95. It should really be []. 274 */ 275 }; 276 277 /* 278 * bootstrap: The caches do not work without cpuarrays anymore, but the 279 * cpuarrays are allocated from the generic caches... 280 */ 281 #define BOOT_CPUCACHE_ENTRIES 1 282 struct arraycache_init { 283 struct array_cache cache; 284 void *entries[BOOT_CPUCACHE_ENTRIES]; 285 }; 286 287 /* 288 * The slab lists for all objects. 289 */ 290 struct kmem_list3 { 291 struct list_head slabs_partial; /* partial list first, better asm code */ 292 struct list_head slabs_full; 293 struct list_head slabs_free; 294 unsigned long free_objects; 295 unsigned int free_limit; 296 unsigned int colour_next; /* Per-node cache coloring */ 297 spinlock_t list_lock; 298 struct array_cache *shared; /* shared per node */ 299 struct array_cache **alien; /* on other nodes */ 300 unsigned long next_reap; /* updated without locking */ 301 int free_touched; /* updated without locking */ 302 }; 303 304 /* 305 * Need this for bootstrapping a per node allocator. 306 */ 307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) 308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; 309 #define CACHE_CACHE 0 310 #define SIZE_AC 1 311 #define SIZE_L3 (1 + MAX_NUMNODES) 312 313 static int drain_freelist(struct kmem_cache *cache, 314 struct kmem_list3 *l3, int tofree); 315 static void free_block(struct kmem_cache *cachep, void **objpp, int len, 316 int node); 317 static int enable_cpucache(struct kmem_cache *cachep); 318 static void cache_reap(struct work_struct *unused); 319 320 /* 321 * This function must be completely optimized away if a constant is passed to 322 * it. Mostly the same as what is in linux/slab.h except it returns an index. 323 */ 324 static __always_inline int index_of(const size_t size) 325 { 326 extern void __bad_size(void); 327 328 if (__builtin_constant_p(size)) { 329 int i = 0; 330 331 #define CACHE(x) \ 332 if (size <=x) \ 333 return i; \ 334 else \ 335 i++; 336 #include "linux/kmalloc_sizes.h" 337 #undef CACHE 338 __bad_size(); 339 } else 340 __bad_size(); 341 return 0; 342 } 343 344 static int slab_early_init = 1; 345 346 #define INDEX_AC index_of(sizeof(struct arraycache_init)) 347 #define INDEX_L3 index_of(sizeof(struct kmem_list3)) 348 349 static void kmem_list3_init(struct kmem_list3 *parent) 350 { 351 INIT_LIST_HEAD(&parent->slabs_full); 352 INIT_LIST_HEAD(&parent->slabs_partial); 353 INIT_LIST_HEAD(&parent->slabs_free); 354 parent->shared = NULL; 355 parent->alien = NULL; 356 parent->colour_next = 0; 357 spin_lock_init(&parent->list_lock); 358 parent->free_objects = 0; 359 parent->free_touched = 0; 360 } 361 362 #define MAKE_LIST(cachep, listp, slab, nodeid) \ 363 do { \ 364 INIT_LIST_HEAD(listp); \ 365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ 366 } while (0) 367 368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 369 do { \ 370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 373 } while (0) 374 375 /* 376 * struct kmem_cache 377 * 378 * manages a cache. 379 */ 380 381 struct kmem_cache { 382 /* 1) per-cpu data, touched during every alloc/free */ 383 struct array_cache *array[NR_CPUS]; 384 /* 2) Cache tunables. Protected by cache_chain_mutex */ 385 unsigned int batchcount; 386 unsigned int limit; 387 unsigned int shared; 388 389 unsigned int buffer_size; 390 u32 reciprocal_buffer_size; 391 /* 3) touched by every alloc & free from the backend */ 392 393 unsigned int flags; /* constant flags */ 394 unsigned int num; /* # of objs per slab */ 395 396 /* 4) cache_grow/shrink */ 397 /* order of pgs per slab (2^n) */ 398 unsigned int gfporder; 399 400 /* force GFP flags, e.g. GFP_DMA */ 401 gfp_t gfpflags; 402 403 size_t colour; /* cache colouring range */ 404 unsigned int colour_off; /* colour offset */ 405 struct kmem_cache *slabp_cache; 406 unsigned int slab_size; 407 unsigned int dflags; /* dynamic flags */ 408 409 /* constructor func */ 410 void (*ctor) (void *, struct kmem_cache *, unsigned long); 411 412 /* de-constructor func */ 413 void (*dtor) (void *, struct kmem_cache *, unsigned long); 414 415 /* 5) cache creation/removal */ 416 const char *name; 417 struct list_head next; 418 419 /* 6) statistics */ 420 #if STATS 421 unsigned long num_active; 422 unsigned long num_allocations; 423 unsigned long high_mark; 424 unsigned long grown; 425 unsigned long reaped; 426 unsigned long errors; 427 unsigned long max_freeable; 428 unsigned long node_allocs; 429 unsigned long node_frees; 430 unsigned long node_overflow; 431 atomic_t allochit; 432 atomic_t allocmiss; 433 atomic_t freehit; 434 atomic_t freemiss; 435 #endif 436 #if DEBUG 437 /* 438 * If debugging is enabled, then the allocator can add additional 439 * fields and/or padding to every object. buffer_size contains the total 440 * object size including these internal fields, the following two 441 * variables contain the offset to the user object and its size. 442 */ 443 int obj_offset; 444 int obj_size; 445 #endif 446 /* 447 * We put nodelists[] at the end of kmem_cache, because we want to size 448 * this array to nr_node_ids slots instead of MAX_NUMNODES 449 * (see kmem_cache_init()) 450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache 451 * is statically defined, so we reserve the max number of nodes. 452 */ 453 struct kmem_list3 *nodelists[MAX_NUMNODES]; 454 /* 455 * Do not add fields after nodelists[] 456 */ 457 }; 458 459 #define CFLGS_OFF_SLAB (0x80000000UL) 460 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 461 462 #define BATCHREFILL_LIMIT 16 463 /* 464 * Optimization question: fewer reaps means less probability for unnessary 465 * cpucache drain/refill cycles. 466 * 467 * OTOH the cpuarrays can contain lots of objects, 468 * which could lock up otherwise freeable slabs. 469 */ 470 #define REAPTIMEOUT_CPUC (2*HZ) 471 #define REAPTIMEOUT_LIST3 (4*HZ) 472 473 #if STATS 474 #define STATS_INC_ACTIVE(x) ((x)->num_active++) 475 #define STATS_DEC_ACTIVE(x) ((x)->num_active--) 476 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 477 #define STATS_INC_GROWN(x) ((x)->grown++) 478 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 479 #define STATS_SET_HIGH(x) \ 480 do { \ 481 if ((x)->num_active > (x)->high_mark) \ 482 (x)->high_mark = (x)->num_active; \ 483 } while (0) 484 #define STATS_INC_ERR(x) ((x)->errors++) 485 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 486 #define STATS_INC_NODEFREES(x) ((x)->node_frees++) 487 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 488 #define STATS_SET_FREEABLE(x, i) \ 489 do { \ 490 if ((x)->max_freeable < i) \ 491 (x)->max_freeable = i; \ 492 } while (0) 493 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 494 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 495 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 496 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 497 #else 498 #define STATS_INC_ACTIVE(x) do { } while (0) 499 #define STATS_DEC_ACTIVE(x) do { } while (0) 500 #define STATS_INC_ALLOCED(x) do { } while (0) 501 #define STATS_INC_GROWN(x) do { } while (0) 502 #define STATS_ADD_REAPED(x,y) do { } while (0) 503 #define STATS_SET_HIGH(x) do { } while (0) 504 #define STATS_INC_ERR(x) do { } while (0) 505 #define STATS_INC_NODEALLOCS(x) do { } while (0) 506 #define STATS_INC_NODEFREES(x) do { } while (0) 507 #define STATS_INC_ACOVERFLOW(x) do { } while (0) 508 #define STATS_SET_FREEABLE(x, i) do { } while (0) 509 #define STATS_INC_ALLOCHIT(x) do { } while (0) 510 #define STATS_INC_ALLOCMISS(x) do { } while (0) 511 #define STATS_INC_FREEHIT(x) do { } while (0) 512 #define STATS_INC_FREEMISS(x) do { } while (0) 513 #endif 514 515 #if DEBUG 516 517 /* 518 * memory layout of objects: 519 * 0 : objp 520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 521 * the end of an object is aligned with the end of the real 522 * allocation. Catches writes behind the end of the allocation. 523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 524 * redzone word. 525 * cachep->obj_offset: The real object. 526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address 528 * [BYTES_PER_WORD long] 529 */ 530 static int obj_offset(struct kmem_cache *cachep) 531 { 532 return cachep->obj_offset; 533 } 534 535 static int obj_size(struct kmem_cache *cachep) 536 { 537 return cachep->obj_size; 538 } 539 540 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 541 { 542 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 543 return (unsigned long long*) (objp + obj_offset(cachep) - 544 sizeof(unsigned long long)); 545 } 546 547 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 548 { 549 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 550 if (cachep->flags & SLAB_STORE_USER) 551 return (unsigned long long *)(objp + cachep->buffer_size - 552 sizeof(unsigned long long) - 553 BYTES_PER_WORD); 554 return (unsigned long long *) (objp + cachep->buffer_size - 555 sizeof(unsigned long long)); 556 } 557 558 static void **dbg_userword(struct kmem_cache *cachep, void *objp) 559 { 560 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 561 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); 562 } 563 564 #else 565 566 #define obj_offset(x) 0 567 #define obj_size(cachep) (cachep->buffer_size) 568 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 569 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 570 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 571 572 #endif 573 574 /* 575 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp 576 * order. 577 */ 578 #if defined(CONFIG_LARGE_ALLOCS) 579 #define MAX_OBJ_ORDER 13 /* up to 32Mb */ 580 #define MAX_GFP_ORDER 13 /* up to 32Mb */ 581 #elif defined(CONFIG_MMU) 582 #define MAX_OBJ_ORDER 5 /* 32 pages */ 583 #define MAX_GFP_ORDER 5 /* 32 pages */ 584 #else 585 #define MAX_OBJ_ORDER 8 /* up to 1Mb */ 586 #define MAX_GFP_ORDER 8 /* up to 1Mb */ 587 #endif 588 589 /* 590 * Do not go above this order unless 0 objects fit into the slab. 591 */ 592 #define BREAK_GFP_ORDER_HI 1 593 #define BREAK_GFP_ORDER_LO 0 594 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; 595 596 /* 597 * Functions for storing/retrieving the cachep and or slab from the page 598 * allocator. These are used to find the slab an obj belongs to. With kfree(), 599 * these are used to find the cache which an obj belongs to. 600 */ 601 static inline void page_set_cache(struct page *page, struct kmem_cache *cache) 602 { 603 page->lru.next = (struct list_head *)cache; 604 } 605 606 static inline struct kmem_cache *page_get_cache(struct page *page) 607 { 608 page = compound_head(page); 609 BUG_ON(!PageSlab(page)); 610 return (struct kmem_cache *)page->lru.next; 611 } 612 613 static inline void page_set_slab(struct page *page, struct slab *slab) 614 { 615 page->lru.prev = (struct list_head *)slab; 616 } 617 618 static inline struct slab *page_get_slab(struct page *page) 619 { 620 BUG_ON(!PageSlab(page)); 621 return (struct slab *)page->lru.prev; 622 } 623 624 static inline struct kmem_cache *virt_to_cache(const void *obj) 625 { 626 struct page *page = virt_to_head_page(obj); 627 return page_get_cache(page); 628 } 629 630 static inline struct slab *virt_to_slab(const void *obj) 631 { 632 struct page *page = virt_to_head_page(obj); 633 return page_get_slab(page); 634 } 635 636 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, 637 unsigned int idx) 638 { 639 return slab->s_mem + cache->buffer_size * idx; 640 } 641 642 /* 643 * We want to avoid an expensive divide : (offset / cache->buffer_size) 644 * Using the fact that buffer_size is a constant for a particular cache, 645 * we can replace (offset / cache->buffer_size) by 646 * reciprocal_divide(offset, cache->reciprocal_buffer_size) 647 */ 648 static inline unsigned int obj_to_index(const struct kmem_cache *cache, 649 const struct slab *slab, void *obj) 650 { 651 u32 offset = (obj - slab->s_mem); 652 return reciprocal_divide(offset, cache->reciprocal_buffer_size); 653 } 654 655 /* 656 * These are the default caches for kmalloc. Custom caches can have other sizes. 657 */ 658 struct cache_sizes malloc_sizes[] = { 659 #define CACHE(x) { .cs_size = (x) }, 660 #include <linux/kmalloc_sizes.h> 661 CACHE(ULONG_MAX) 662 #undef CACHE 663 }; 664 EXPORT_SYMBOL(malloc_sizes); 665 666 /* Must match cache_sizes above. Out of line to keep cache footprint low. */ 667 struct cache_names { 668 char *name; 669 char *name_dma; 670 }; 671 672 static struct cache_names __initdata cache_names[] = { 673 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, 674 #include <linux/kmalloc_sizes.h> 675 {NULL,} 676 #undef CACHE 677 }; 678 679 static struct arraycache_init initarray_cache __initdata = 680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 681 static struct arraycache_init initarray_generic = 682 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; 683 684 /* internal cache of cache description objs */ 685 static struct kmem_cache cache_cache = { 686 .batchcount = 1, 687 .limit = BOOT_CPUCACHE_ENTRIES, 688 .shared = 1, 689 .buffer_size = sizeof(struct kmem_cache), 690 .name = "kmem_cache", 691 }; 692 693 #define BAD_ALIEN_MAGIC 0x01020304ul 694 695 #ifdef CONFIG_LOCKDEP 696 697 /* 698 * Slab sometimes uses the kmalloc slabs to store the slab headers 699 * for other slabs "off slab". 700 * The locking for this is tricky in that it nests within the locks 701 * of all other slabs in a few places; to deal with this special 702 * locking we put on-slab caches into a separate lock-class. 703 * 704 * We set lock class for alien array caches which are up during init. 705 * The lock annotation will be lost if all cpus of a node goes down and 706 * then comes back up during hotplug 707 */ 708 static struct lock_class_key on_slab_l3_key; 709 static struct lock_class_key on_slab_alc_key; 710 711 static inline void init_lock_keys(void) 712 713 { 714 int q; 715 struct cache_sizes *s = malloc_sizes; 716 717 while (s->cs_size != ULONG_MAX) { 718 for_each_node(q) { 719 struct array_cache **alc; 720 int r; 721 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; 722 if (!l3 || OFF_SLAB(s->cs_cachep)) 723 continue; 724 lockdep_set_class(&l3->list_lock, &on_slab_l3_key); 725 alc = l3->alien; 726 /* 727 * FIXME: This check for BAD_ALIEN_MAGIC 728 * should go away when common slab code is taught to 729 * work even without alien caches. 730 * Currently, non NUMA code returns BAD_ALIEN_MAGIC 731 * for alloc_alien_cache, 732 */ 733 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) 734 continue; 735 for_each_node(r) { 736 if (alc[r]) 737 lockdep_set_class(&alc[r]->lock, 738 &on_slab_alc_key); 739 } 740 } 741 s++; 742 } 743 } 744 #else 745 static inline void init_lock_keys(void) 746 { 747 } 748 #endif 749 750 /* 751 * 1. Guard access to the cache-chain. 752 * 2. Protect sanity of cpu_online_map against cpu hotplug events 753 */ 754 static DEFINE_MUTEX(cache_chain_mutex); 755 static struct list_head cache_chain; 756 757 /* 758 * chicken and egg problem: delay the per-cpu array allocation 759 * until the general caches are up. 760 */ 761 static enum { 762 NONE, 763 PARTIAL_AC, 764 PARTIAL_L3, 765 FULL 766 } g_cpucache_up; 767 768 /* 769 * used by boot code to determine if it can use slab based allocator 770 */ 771 int slab_is_available(void) 772 { 773 return g_cpucache_up == FULL; 774 } 775 776 static DEFINE_PER_CPU(struct delayed_work, reap_work); 777 778 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 779 { 780 return cachep->array[smp_processor_id()]; 781 } 782 783 static inline struct kmem_cache *__find_general_cachep(size_t size, 784 gfp_t gfpflags) 785 { 786 struct cache_sizes *csizep = malloc_sizes; 787 788 #if DEBUG 789 /* This happens if someone tries to call 790 * kmem_cache_create(), or __kmalloc(), before 791 * the generic caches are initialized. 792 */ 793 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); 794 #endif 795 while (size > csizep->cs_size) 796 csizep++; 797 798 /* 799 * Really subtle: The last entry with cs->cs_size==ULONG_MAX 800 * has cs_{dma,}cachep==NULL. Thus no special case 801 * for large kmalloc calls required. 802 */ 803 #ifdef CONFIG_ZONE_DMA 804 if (unlikely(gfpflags & GFP_DMA)) 805 return csizep->cs_dmacachep; 806 #endif 807 return csizep->cs_cachep; 808 } 809 810 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) 811 { 812 return __find_general_cachep(size, gfpflags); 813 } 814 815 static size_t slab_mgmt_size(size_t nr_objs, size_t align) 816 { 817 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); 818 } 819 820 /* 821 * Calculate the number of objects and left-over bytes for a given buffer size. 822 */ 823 static void cache_estimate(unsigned long gfporder, size_t buffer_size, 824 size_t align, int flags, size_t *left_over, 825 unsigned int *num) 826 { 827 int nr_objs; 828 size_t mgmt_size; 829 size_t slab_size = PAGE_SIZE << gfporder; 830 831 /* 832 * The slab management structure can be either off the slab or 833 * on it. For the latter case, the memory allocated for a 834 * slab is used for: 835 * 836 * - The struct slab 837 * - One kmem_bufctl_t for each object 838 * - Padding to respect alignment of @align 839 * - @buffer_size bytes for each object 840 * 841 * If the slab management structure is off the slab, then the 842 * alignment will already be calculated into the size. Because 843 * the slabs are all pages aligned, the objects will be at the 844 * correct alignment when allocated. 845 */ 846 if (flags & CFLGS_OFF_SLAB) { 847 mgmt_size = 0; 848 nr_objs = slab_size / buffer_size; 849 850 if (nr_objs > SLAB_LIMIT) 851 nr_objs = SLAB_LIMIT; 852 } else { 853 /* 854 * Ignore padding for the initial guess. The padding 855 * is at most @align-1 bytes, and @buffer_size is at 856 * least @align. In the worst case, this result will 857 * be one greater than the number of objects that fit 858 * into the memory allocation when taking the padding 859 * into account. 860 */ 861 nr_objs = (slab_size - sizeof(struct slab)) / 862 (buffer_size + sizeof(kmem_bufctl_t)); 863 864 /* 865 * This calculated number will be either the right 866 * amount, or one greater than what we want. 867 */ 868 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size 869 > slab_size) 870 nr_objs--; 871 872 if (nr_objs > SLAB_LIMIT) 873 nr_objs = SLAB_LIMIT; 874 875 mgmt_size = slab_mgmt_size(nr_objs, align); 876 } 877 *num = nr_objs; 878 *left_over = slab_size - nr_objs*buffer_size - mgmt_size; 879 } 880 881 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) 882 883 static void __slab_error(const char *function, struct kmem_cache *cachep, 884 char *msg) 885 { 886 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", 887 function, cachep->name, msg); 888 dump_stack(); 889 } 890 891 /* 892 * By default on NUMA we use alien caches to stage the freeing of 893 * objects allocated from other nodes. This causes massive memory 894 * inefficiencies when using fake NUMA setup to split memory into a 895 * large number of small nodes, so it can be disabled on the command 896 * line 897 */ 898 899 static int use_alien_caches __read_mostly = 1; 900 static int __init noaliencache_setup(char *s) 901 { 902 use_alien_caches = 0; 903 return 1; 904 } 905 __setup("noaliencache", noaliencache_setup); 906 907 #ifdef CONFIG_NUMA 908 /* 909 * Special reaping functions for NUMA systems called from cache_reap(). 910 * These take care of doing round robin flushing of alien caches (containing 911 * objects freed on different nodes from which they were allocated) and the 912 * flushing of remote pcps by calling drain_node_pages. 913 */ 914 static DEFINE_PER_CPU(unsigned long, reap_node); 915 916 static void init_reap_node(int cpu) 917 { 918 int node; 919 920 node = next_node(cpu_to_node(cpu), node_online_map); 921 if (node == MAX_NUMNODES) 922 node = first_node(node_online_map); 923 924 per_cpu(reap_node, cpu) = node; 925 } 926 927 static void next_reap_node(void) 928 { 929 int node = __get_cpu_var(reap_node); 930 931 node = next_node(node, node_online_map); 932 if (unlikely(node >= MAX_NUMNODES)) 933 node = first_node(node_online_map); 934 __get_cpu_var(reap_node) = node; 935 } 936 937 #else 938 #define init_reap_node(cpu) do { } while (0) 939 #define next_reap_node(void) do { } while (0) 940 #endif 941 942 /* 943 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 944 * via the workqueue/eventd. 945 * Add the CPU number into the expiration time to minimize the possibility of 946 * the CPUs getting into lockstep and contending for the global cache chain 947 * lock. 948 */ 949 static void __devinit start_cpu_timer(int cpu) 950 { 951 struct delayed_work *reap_work = &per_cpu(reap_work, cpu); 952 953 /* 954 * When this gets called from do_initcalls via cpucache_init(), 955 * init_workqueues() has already run, so keventd will be setup 956 * at that time. 957 */ 958 if (keventd_up() && reap_work->work.func == NULL) { 959 init_reap_node(cpu); 960 INIT_DELAYED_WORK(reap_work, cache_reap); 961 schedule_delayed_work_on(cpu, reap_work, 962 __round_jiffies_relative(HZ, cpu)); 963 } 964 } 965 966 static struct array_cache *alloc_arraycache(int node, int entries, 967 int batchcount) 968 { 969 int memsize = sizeof(void *) * entries + sizeof(struct array_cache); 970 struct array_cache *nc = NULL; 971 972 nc = kmalloc_node(memsize, GFP_KERNEL, node); 973 if (nc) { 974 nc->avail = 0; 975 nc->limit = entries; 976 nc->batchcount = batchcount; 977 nc->touched = 0; 978 spin_lock_init(&nc->lock); 979 } 980 return nc; 981 } 982 983 /* 984 * Transfer objects in one arraycache to another. 985 * Locking must be handled by the caller. 986 * 987 * Return the number of entries transferred. 988 */ 989 static int transfer_objects(struct array_cache *to, 990 struct array_cache *from, unsigned int max) 991 { 992 /* Figure out how many entries to transfer */ 993 int nr = min(min(from->avail, max), to->limit - to->avail); 994 995 if (!nr) 996 return 0; 997 998 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 999 sizeof(void *) *nr); 1000 1001 from->avail -= nr; 1002 to->avail += nr; 1003 to->touched = 1; 1004 return nr; 1005 } 1006 1007 #ifndef CONFIG_NUMA 1008 1009 #define drain_alien_cache(cachep, alien) do { } while (0) 1010 #define reap_alien(cachep, l3) do { } while (0) 1011 1012 static inline struct array_cache **alloc_alien_cache(int node, int limit) 1013 { 1014 return (struct array_cache **)BAD_ALIEN_MAGIC; 1015 } 1016 1017 static inline void free_alien_cache(struct array_cache **ac_ptr) 1018 { 1019 } 1020 1021 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1022 { 1023 return 0; 1024 } 1025 1026 static inline void *alternate_node_alloc(struct kmem_cache *cachep, 1027 gfp_t flags) 1028 { 1029 return NULL; 1030 } 1031 1032 static inline void *____cache_alloc_node(struct kmem_cache *cachep, 1033 gfp_t flags, int nodeid) 1034 { 1035 return NULL; 1036 } 1037 1038 #else /* CONFIG_NUMA */ 1039 1040 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 1041 static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 1042 1043 static struct array_cache **alloc_alien_cache(int node, int limit) 1044 { 1045 struct array_cache **ac_ptr; 1046 int memsize = sizeof(void *) * nr_node_ids; 1047 int i; 1048 1049 if (limit > 1) 1050 limit = 12; 1051 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); 1052 if (ac_ptr) { 1053 for_each_node(i) { 1054 if (i == node || !node_online(i)) { 1055 ac_ptr[i] = NULL; 1056 continue; 1057 } 1058 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); 1059 if (!ac_ptr[i]) { 1060 for (i--; i <= 0; i--) 1061 kfree(ac_ptr[i]); 1062 kfree(ac_ptr); 1063 return NULL; 1064 } 1065 } 1066 } 1067 return ac_ptr; 1068 } 1069 1070 static void free_alien_cache(struct array_cache **ac_ptr) 1071 { 1072 int i; 1073 1074 if (!ac_ptr) 1075 return; 1076 for_each_node(i) 1077 kfree(ac_ptr[i]); 1078 kfree(ac_ptr); 1079 } 1080 1081 static void __drain_alien_cache(struct kmem_cache *cachep, 1082 struct array_cache *ac, int node) 1083 { 1084 struct kmem_list3 *rl3 = cachep->nodelists[node]; 1085 1086 if (ac->avail) { 1087 spin_lock(&rl3->list_lock); 1088 /* 1089 * Stuff objects into the remote nodes shared array first. 1090 * That way we could avoid the overhead of putting the objects 1091 * into the free lists and getting them back later. 1092 */ 1093 if (rl3->shared) 1094 transfer_objects(rl3->shared, ac, ac->limit); 1095 1096 free_block(cachep, ac->entry, ac->avail, node); 1097 ac->avail = 0; 1098 spin_unlock(&rl3->list_lock); 1099 } 1100 } 1101 1102 /* 1103 * Called from cache_reap() to regularly drain alien caches round robin. 1104 */ 1105 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) 1106 { 1107 int node = __get_cpu_var(reap_node); 1108 1109 if (l3->alien) { 1110 struct array_cache *ac = l3->alien[node]; 1111 1112 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { 1113 __drain_alien_cache(cachep, ac, node); 1114 spin_unlock_irq(&ac->lock); 1115 } 1116 } 1117 } 1118 1119 static void drain_alien_cache(struct kmem_cache *cachep, 1120 struct array_cache **alien) 1121 { 1122 int i = 0; 1123 struct array_cache *ac; 1124 unsigned long flags; 1125 1126 for_each_online_node(i) { 1127 ac = alien[i]; 1128 if (ac) { 1129 spin_lock_irqsave(&ac->lock, flags); 1130 __drain_alien_cache(cachep, ac, i); 1131 spin_unlock_irqrestore(&ac->lock, flags); 1132 } 1133 } 1134 } 1135 1136 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 1137 { 1138 struct slab *slabp = virt_to_slab(objp); 1139 int nodeid = slabp->nodeid; 1140 struct kmem_list3 *l3; 1141 struct array_cache *alien = NULL; 1142 int node; 1143 1144 node = numa_node_id(); 1145 1146 /* 1147 * Make sure we are not freeing a object from another node to the array 1148 * cache on this cpu. 1149 */ 1150 if (likely(slabp->nodeid == node)) 1151 return 0; 1152 1153 l3 = cachep->nodelists[node]; 1154 STATS_INC_NODEFREES(cachep); 1155 if (l3->alien && l3->alien[nodeid]) { 1156 alien = l3->alien[nodeid]; 1157 spin_lock(&alien->lock); 1158 if (unlikely(alien->avail == alien->limit)) { 1159 STATS_INC_ACOVERFLOW(cachep); 1160 __drain_alien_cache(cachep, alien, nodeid); 1161 } 1162 alien->entry[alien->avail++] = objp; 1163 spin_unlock(&alien->lock); 1164 } else { 1165 spin_lock(&(cachep->nodelists[nodeid])->list_lock); 1166 free_block(cachep, &objp, 1, nodeid); 1167 spin_unlock(&(cachep->nodelists[nodeid])->list_lock); 1168 } 1169 return 1; 1170 } 1171 #endif 1172 1173 static int __cpuinit cpuup_callback(struct notifier_block *nfb, 1174 unsigned long action, void *hcpu) 1175 { 1176 long cpu = (long)hcpu; 1177 struct kmem_cache *cachep; 1178 struct kmem_list3 *l3 = NULL; 1179 int node = cpu_to_node(cpu); 1180 int memsize = sizeof(struct kmem_list3); 1181 1182 switch (action) { 1183 case CPU_LOCK_ACQUIRE: 1184 mutex_lock(&cache_chain_mutex); 1185 break; 1186 case CPU_UP_PREPARE: 1187 case CPU_UP_PREPARE_FROZEN: 1188 /* 1189 * We need to do this right in the beginning since 1190 * alloc_arraycache's are going to use this list. 1191 * kmalloc_node allows us to add the slab to the right 1192 * kmem_list3 and not this cpu's kmem_list3 1193 */ 1194 1195 list_for_each_entry(cachep, &cache_chain, next) { 1196 /* 1197 * Set up the size64 kmemlist for cpu before we can 1198 * begin anything. Make sure some other cpu on this 1199 * node has not already allocated this 1200 */ 1201 if (!cachep->nodelists[node]) { 1202 l3 = kmalloc_node(memsize, GFP_KERNEL, node); 1203 if (!l3) 1204 goto bad; 1205 kmem_list3_init(l3); 1206 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 1207 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1208 1209 /* 1210 * The l3s don't come and go as CPUs come and 1211 * go. cache_chain_mutex is sufficient 1212 * protection here. 1213 */ 1214 cachep->nodelists[node] = l3; 1215 } 1216 1217 spin_lock_irq(&cachep->nodelists[node]->list_lock); 1218 cachep->nodelists[node]->free_limit = 1219 (1 + nr_cpus_node(node)) * 1220 cachep->batchcount + cachep->num; 1221 spin_unlock_irq(&cachep->nodelists[node]->list_lock); 1222 } 1223 1224 /* 1225 * Now we can go ahead with allocating the shared arrays and 1226 * array caches 1227 */ 1228 list_for_each_entry(cachep, &cache_chain, next) { 1229 struct array_cache *nc; 1230 struct array_cache *shared = NULL; 1231 struct array_cache **alien = NULL; 1232 1233 nc = alloc_arraycache(node, cachep->limit, 1234 cachep->batchcount); 1235 if (!nc) 1236 goto bad; 1237 if (cachep->shared) { 1238 shared = alloc_arraycache(node, 1239 cachep->shared * cachep->batchcount, 1240 0xbaadf00d); 1241 if (!shared) 1242 goto bad; 1243 } 1244 if (use_alien_caches) { 1245 alien = alloc_alien_cache(node, cachep->limit); 1246 if (!alien) 1247 goto bad; 1248 } 1249 cachep->array[cpu] = nc; 1250 l3 = cachep->nodelists[node]; 1251 BUG_ON(!l3); 1252 1253 spin_lock_irq(&l3->list_lock); 1254 if (!l3->shared) { 1255 /* 1256 * We are serialised from CPU_DEAD or 1257 * CPU_UP_CANCELLED by the cpucontrol lock 1258 */ 1259 l3->shared = shared; 1260 shared = NULL; 1261 } 1262 #ifdef CONFIG_NUMA 1263 if (!l3->alien) { 1264 l3->alien = alien; 1265 alien = NULL; 1266 } 1267 #endif 1268 spin_unlock_irq(&l3->list_lock); 1269 kfree(shared); 1270 free_alien_cache(alien); 1271 } 1272 break; 1273 case CPU_ONLINE: 1274 case CPU_ONLINE_FROZEN: 1275 start_cpu_timer(cpu); 1276 break; 1277 #ifdef CONFIG_HOTPLUG_CPU 1278 case CPU_DOWN_PREPARE: 1279 case CPU_DOWN_PREPARE_FROZEN: 1280 /* 1281 * Shutdown cache reaper. Note that the cache_chain_mutex is 1282 * held so that if cache_reap() is invoked it cannot do 1283 * anything expensive but will only modify reap_work 1284 * and reschedule the timer. 1285 */ 1286 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); 1287 /* Now the cache_reaper is guaranteed to be not running. */ 1288 per_cpu(reap_work, cpu).work.func = NULL; 1289 break; 1290 case CPU_DOWN_FAILED: 1291 case CPU_DOWN_FAILED_FROZEN: 1292 start_cpu_timer(cpu); 1293 break; 1294 case CPU_DEAD: 1295 case CPU_DEAD_FROZEN: 1296 /* 1297 * Even if all the cpus of a node are down, we don't free the 1298 * kmem_list3 of any cache. This to avoid a race between 1299 * cpu_down, and a kmalloc allocation from another cpu for 1300 * memory from the node of the cpu going down. The list3 1301 * structure is usually allocated from kmem_cache_create() and 1302 * gets destroyed at kmem_cache_destroy(). 1303 */ 1304 /* fall thru */ 1305 #endif 1306 case CPU_UP_CANCELED: 1307 case CPU_UP_CANCELED_FROZEN: 1308 list_for_each_entry(cachep, &cache_chain, next) { 1309 struct array_cache *nc; 1310 struct array_cache *shared; 1311 struct array_cache **alien; 1312 cpumask_t mask; 1313 1314 mask = node_to_cpumask(node); 1315 /* cpu is dead; no one can alloc from it. */ 1316 nc = cachep->array[cpu]; 1317 cachep->array[cpu] = NULL; 1318 l3 = cachep->nodelists[node]; 1319 1320 if (!l3) 1321 goto free_array_cache; 1322 1323 spin_lock_irq(&l3->list_lock); 1324 1325 /* Free limit for this kmem_list3 */ 1326 l3->free_limit -= cachep->batchcount; 1327 if (nc) 1328 free_block(cachep, nc->entry, nc->avail, node); 1329 1330 if (!cpus_empty(mask)) { 1331 spin_unlock_irq(&l3->list_lock); 1332 goto free_array_cache; 1333 } 1334 1335 shared = l3->shared; 1336 if (shared) { 1337 free_block(cachep, shared->entry, 1338 shared->avail, node); 1339 l3->shared = NULL; 1340 } 1341 1342 alien = l3->alien; 1343 l3->alien = NULL; 1344 1345 spin_unlock_irq(&l3->list_lock); 1346 1347 kfree(shared); 1348 if (alien) { 1349 drain_alien_cache(cachep, alien); 1350 free_alien_cache(alien); 1351 } 1352 free_array_cache: 1353 kfree(nc); 1354 } 1355 /* 1356 * In the previous loop, all the objects were freed to 1357 * the respective cache's slabs, now we can go ahead and 1358 * shrink each nodelist to its limit. 1359 */ 1360 list_for_each_entry(cachep, &cache_chain, next) { 1361 l3 = cachep->nodelists[node]; 1362 if (!l3) 1363 continue; 1364 drain_freelist(cachep, l3, l3->free_objects); 1365 } 1366 break; 1367 case CPU_LOCK_RELEASE: 1368 mutex_unlock(&cache_chain_mutex); 1369 break; 1370 } 1371 return NOTIFY_OK; 1372 bad: 1373 return NOTIFY_BAD; 1374 } 1375 1376 static struct notifier_block __cpuinitdata cpucache_notifier = { 1377 &cpuup_callback, NULL, 0 1378 }; 1379 1380 /* 1381 * swap the static kmem_list3 with kmalloced memory 1382 */ 1383 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, 1384 int nodeid) 1385 { 1386 struct kmem_list3 *ptr; 1387 1388 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); 1389 BUG_ON(!ptr); 1390 1391 local_irq_disable(); 1392 memcpy(ptr, list, sizeof(struct kmem_list3)); 1393 /* 1394 * Do not assume that spinlocks can be initialized via memcpy: 1395 */ 1396 spin_lock_init(&ptr->list_lock); 1397 1398 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1399 cachep->nodelists[nodeid] = ptr; 1400 local_irq_enable(); 1401 } 1402 1403 /* 1404 * Initialisation. Called after the page allocator have been initialised and 1405 * before smp_init(). 1406 */ 1407 void __init kmem_cache_init(void) 1408 { 1409 size_t left_over; 1410 struct cache_sizes *sizes; 1411 struct cache_names *names; 1412 int i; 1413 int order; 1414 int node; 1415 1416 if (num_possible_nodes() == 1) 1417 use_alien_caches = 0; 1418 1419 for (i = 0; i < NUM_INIT_LISTS; i++) { 1420 kmem_list3_init(&initkmem_list3[i]); 1421 if (i < MAX_NUMNODES) 1422 cache_cache.nodelists[i] = NULL; 1423 } 1424 1425 /* 1426 * Fragmentation resistance on low memory - only use bigger 1427 * page orders on machines with more than 32MB of memory. 1428 */ 1429 if (num_physpages > (32 << 20) >> PAGE_SHIFT) 1430 slab_break_gfp_order = BREAK_GFP_ORDER_HI; 1431 1432 /* Bootstrap is tricky, because several objects are allocated 1433 * from caches that do not exist yet: 1434 * 1) initialize the cache_cache cache: it contains the struct 1435 * kmem_cache structures of all caches, except cache_cache itself: 1436 * cache_cache is statically allocated. 1437 * Initially an __init data area is used for the head array and the 1438 * kmem_list3 structures, it's replaced with a kmalloc allocated 1439 * array at the end of the bootstrap. 1440 * 2) Create the first kmalloc cache. 1441 * The struct kmem_cache for the new cache is allocated normally. 1442 * An __init data area is used for the head array. 1443 * 3) Create the remaining kmalloc caches, with minimally sized 1444 * head arrays. 1445 * 4) Replace the __init data head arrays for cache_cache and the first 1446 * kmalloc cache with kmalloc allocated arrays. 1447 * 5) Replace the __init data for kmem_list3 for cache_cache and 1448 * the other cache's with kmalloc allocated memory. 1449 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1450 */ 1451 1452 node = numa_node_id(); 1453 1454 /* 1) create the cache_cache */ 1455 INIT_LIST_HEAD(&cache_chain); 1456 list_add(&cache_cache.next, &cache_chain); 1457 cache_cache.colour_off = cache_line_size(); 1458 cache_cache.array[smp_processor_id()] = &initarray_cache.cache; 1459 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; 1460 1461 /* 1462 * struct kmem_cache size depends on nr_node_ids, which 1463 * can be less than MAX_NUMNODES. 1464 */ 1465 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + 1466 nr_node_ids * sizeof(struct kmem_list3 *); 1467 #if DEBUG 1468 cache_cache.obj_size = cache_cache.buffer_size; 1469 #endif 1470 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, 1471 cache_line_size()); 1472 cache_cache.reciprocal_buffer_size = 1473 reciprocal_value(cache_cache.buffer_size); 1474 1475 for (order = 0; order < MAX_ORDER; order++) { 1476 cache_estimate(order, cache_cache.buffer_size, 1477 cache_line_size(), 0, &left_over, &cache_cache.num); 1478 if (cache_cache.num) 1479 break; 1480 } 1481 BUG_ON(!cache_cache.num); 1482 cache_cache.gfporder = order; 1483 cache_cache.colour = left_over / cache_cache.colour_off; 1484 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + 1485 sizeof(struct slab), cache_line_size()); 1486 1487 /* 2+3) create the kmalloc caches */ 1488 sizes = malloc_sizes; 1489 names = cache_names; 1490 1491 /* 1492 * Initialize the caches that provide memory for the array cache and the 1493 * kmem_list3 structures first. Without this, further allocations will 1494 * bug. 1495 */ 1496 1497 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, 1498 sizes[INDEX_AC].cs_size, 1499 ARCH_KMALLOC_MINALIGN, 1500 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1501 NULL, NULL); 1502 1503 if (INDEX_AC != INDEX_L3) { 1504 sizes[INDEX_L3].cs_cachep = 1505 kmem_cache_create(names[INDEX_L3].name, 1506 sizes[INDEX_L3].cs_size, 1507 ARCH_KMALLOC_MINALIGN, 1508 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1509 NULL, NULL); 1510 } 1511 1512 slab_early_init = 0; 1513 1514 while (sizes->cs_size != ULONG_MAX) { 1515 /* 1516 * For performance, all the general caches are L1 aligned. 1517 * This should be particularly beneficial on SMP boxes, as it 1518 * eliminates "false sharing". 1519 * Note for systems short on memory removing the alignment will 1520 * allow tighter packing of the smaller caches. 1521 */ 1522 if (!sizes->cs_cachep) { 1523 sizes->cs_cachep = kmem_cache_create(names->name, 1524 sizes->cs_size, 1525 ARCH_KMALLOC_MINALIGN, 1526 ARCH_KMALLOC_FLAGS|SLAB_PANIC, 1527 NULL, NULL); 1528 } 1529 #ifdef CONFIG_ZONE_DMA 1530 sizes->cs_dmacachep = kmem_cache_create( 1531 names->name_dma, 1532 sizes->cs_size, 1533 ARCH_KMALLOC_MINALIGN, 1534 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| 1535 SLAB_PANIC, 1536 NULL, NULL); 1537 #endif 1538 sizes++; 1539 names++; 1540 } 1541 /* 4) Replace the bootstrap head arrays */ 1542 { 1543 struct array_cache *ptr; 1544 1545 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1546 1547 local_irq_disable(); 1548 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); 1549 memcpy(ptr, cpu_cache_get(&cache_cache), 1550 sizeof(struct arraycache_init)); 1551 /* 1552 * Do not assume that spinlocks can be initialized via memcpy: 1553 */ 1554 spin_lock_init(&ptr->lock); 1555 1556 cache_cache.array[smp_processor_id()] = ptr; 1557 local_irq_enable(); 1558 1559 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 1560 1561 local_irq_disable(); 1562 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) 1563 != &initarray_generic.cache); 1564 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), 1565 sizeof(struct arraycache_init)); 1566 /* 1567 * Do not assume that spinlocks can be initialized via memcpy: 1568 */ 1569 spin_lock_init(&ptr->lock); 1570 1571 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = 1572 ptr; 1573 local_irq_enable(); 1574 } 1575 /* 5) Replace the bootstrap kmem_list3's */ 1576 { 1577 int nid; 1578 1579 /* Replace the static kmem_list3 structures for the boot cpu */ 1580 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); 1581 1582 for_each_online_node(nid) { 1583 init_list(malloc_sizes[INDEX_AC].cs_cachep, 1584 &initkmem_list3[SIZE_AC + nid], nid); 1585 1586 if (INDEX_AC != INDEX_L3) { 1587 init_list(malloc_sizes[INDEX_L3].cs_cachep, 1588 &initkmem_list3[SIZE_L3 + nid], nid); 1589 } 1590 } 1591 } 1592 1593 /* 6) resize the head arrays to their final sizes */ 1594 { 1595 struct kmem_cache *cachep; 1596 mutex_lock(&cache_chain_mutex); 1597 list_for_each_entry(cachep, &cache_chain, next) 1598 if (enable_cpucache(cachep)) 1599 BUG(); 1600 mutex_unlock(&cache_chain_mutex); 1601 } 1602 1603 /* Annotate slab for lockdep -- annotate the malloc caches */ 1604 init_lock_keys(); 1605 1606 1607 /* Done! */ 1608 g_cpucache_up = FULL; 1609 1610 /* 1611 * Register a cpu startup notifier callback that initializes 1612 * cpu_cache_get for all new cpus 1613 */ 1614 register_cpu_notifier(&cpucache_notifier); 1615 1616 /* 1617 * The reap timers are started later, with a module init call: That part 1618 * of the kernel is not yet operational. 1619 */ 1620 } 1621 1622 static int __init cpucache_init(void) 1623 { 1624 int cpu; 1625 1626 /* 1627 * Register the timers that return unneeded pages to the page allocator 1628 */ 1629 for_each_online_cpu(cpu) 1630 start_cpu_timer(cpu); 1631 return 0; 1632 } 1633 __initcall(cpucache_init); 1634 1635 /* 1636 * Interface to system's page allocator. No need to hold the cache-lock. 1637 * 1638 * If we requested dmaable memory, we will get it. Even if we 1639 * did not request dmaable memory, we might get it, but that 1640 * would be relatively rare and ignorable. 1641 */ 1642 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) 1643 { 1644 struct page *page; 1645 int nr_pages; 1646 int i; 1647 1648 #ifndef CONFIG_MMU 1649 /* 1650 * Nommu uses slab's for process anonymous memory allocations, and thus 1651 * requires __GFP_COMP to properly refcount higher order allocations 1652 */ 1653 flags |= __GFP_COMP; 1654 #endif 1655 1656 flags |= cachep->gfpflags; 1657 1658 page = alloc_pages_node(nodeid, flags, cachep->gfporder); 1659 if (!page) 1660 return NULL; 1661 1662 nr_pages = (1 << cachep->gfporder); 1663 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1664 add_zone_page_state(page_zone(page), 1665 NR_SLAB_RECLAIMABLE, nr_pages); 1666 else 1667 add_zone_page_state(page_zone(page), 1668 NR_SLAB_UNRECLAIMABLE, nr_pages); 1669 for (i = 0; i < nr_pages; i++) 1670 __SetPageSlab(page + i); 1671 return page_address(page); 1672 } 1673 1674 /* 1675 * Interface to system's page release. 1676 */ 1677 static void kmem_freepages(struct kmem_cache *cachep, void *addr) 1678 { 1679 unsigned long i = (1 << cachep->gfporder); 1680 struct page *page = virt_to_page(addr); 1681 const unsigned long nr_freed = i; 1682 1683 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1684 sub_zone_page_state(page_zone(page), 1685 NR_SLAB_RECLAIMABLE, nr_freed); 1686 else 1687 sub_zone_page_state(page_zone(page), 1688 NR_SLAB_UNRECLAIMABLE, nr_freed); 1689 while (i--) { 1690 BUG_ON(!PageSlab(page)); 1691 __ClearPageSlab(page); 1692 page++; 1693 } 1694 if (current->reclaim_state) 1695 current->reclaim_state->reclaimed_slab += nr_freed; 1696 free_pages((unsigned long)addr, cachep->gfporder); 1697 } 1698 1699 static void kmem_rcu_free(struct rcu_head *head) 1700 { 1701 struct slab_rcu *slab_rcu = (struct slab_rcu *)head; 1702 struct kmem_cache *cachep = slab_rcu->cachep; 1703 1704 kmem_freepages(cachep, slab_rcu->addr); 1705 if (OFF_SLAB(cachep)) 1706 kmem_cache_free(cachep->slabp_cache, slab_rcu); 1707 } 1708 1709 #if DEBUG 1710 1711 #ifdef CONFIG_DEBUG_PAGEALLOC 1712 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1713 unsigned long caller) 1714 { 1715 int size = obj_size(cachep); 1716 1717 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1718 1719 if (size < 5 * sizeof(unsigned long)) 1720 return; 1721 1722 *addr++ = 0x12345678; 1723 *addr++ = caller; 1724 *addr++ = smp_processor_id(); 1725 size -= 3 * sizeof(unsigned long); 1726 { 1727 unsigned long *sptr = &caller; 1728 unsigned long svalue; 1729 1730 while (!kstack_end(sptr)) { 1731 svalue = *sptr++; 1732 if (kernel_text_address(svalue)) { 1733 *addr++ = svalue; 1734 size -= sizeof(unsigned long); 1735 if (size <= sizeof(unsigned long)) 1736 break; 1737 } 1738 } 1739 1740 } 1741 *addr++ = 0x87654321; 1742 } 1743 #endif 1744 1745 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1746 { 1747 int size = obj_size(cachep); 1748 addr = &((char *)addr)[obj_offset(cachep)]; 1749 1750 memset(addr, val, size); 1751 *(unsigned char *)(addr + size - 1) = POISON_END; 1752 } 1753 1754 static void dump_line(char *data, int offset, int limit) 1755 { 1756 int i; 1757 unsigned char error = 0; 1758 int bad_count = 0; 1759 1760 printk(KERN_ERR "%03x:", offset); 1761 for (i = 0; i < limit; i++) { 1762 if (data[offset + i] != POISON_FREE) { 1763 error = data[offset + i]; 1764 bad_count++; 1765 } 1766 printk(" %02x", (unsigned char)data[offset + i]); 1767 } 1768 printk("\n"); 1769 1770 if (bad_count == 1) { 1771 error ^= POISON_FREE; 1772 if (!(error & (error - 1))) { 1773 printk(KERN_ERR "Single bit error detected. Probably " 1774 "bad RAM.\n"); 1775 #ifdef CONFIG_X86 1776 printk(KERN_ERR "Run memtest86+ or a similar memory " 1777 "test tool.\n"); 1778 #else 1779 printk(KERN_ERR "Run a memory test tool.\n"); 1780 #endif 1781 } 1782 } 1783 } 1784 #endif 1785 1786 #if DEBUG 1787 1788 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1789 { 1790 int i, size; 1791 char *realobj; 1792 1793 if (cachep->flags & SLAB_RED_ZONE) { 1794 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", 1795 *dbg_redzone1(cachep, objp), 1796 *dbg_redzone2(cachep, objp)); 1797 } 1798 1799 if (cachep->flags & SLAB_STORE_USER) { 1800 printk(KERN_ERR "Last user: [<%p>]", 1801 *dbg_userword(cachep, objp)); 1802 print_symbol("(%s)", 1803 (unsigned long)*dbg_userword(cachep, objp)); 1804 printk("\n"); 1805 } 1806 realobj = (char *)objp + obj_offset(cachep); 1807 size = obj_size(cachep); 1808 for (i = 0; i < size && lines; i += 16, lines--) { 1809 int limit; 1810 limit = 16; 1811 if (i + limit > size) 1812 limit = size - i; 1813 dump_line(realobj, i, limit); 1814 } 1815 } 1816 1817 static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1818 { 1819 char *realobj; 1820 int size, i; 1821 int lines = 0; 1822 1823 realobj = (char *)objp + obj_offset(cachep); 1824 size = obj_size(cachep); 1825 1826 for (i = 0; i < size; i++) { 1827 char exp = POISON_FREE; 1828 if (i == size - 1) 1829 exp = POISON_END; 1830 if (realobj[i] != exp) { 1831 int limit; 1832 /* Mismatch ! */ 1833 /* Print header */ 1834 if (lines == 0) { 1835 printk(KERN_ERR 1836 "Slab corruption: %s start=%p, len=%d\n", 1837 cachep->name, realobj, size); 1838 print_objinfo(cachep, objp, 0); 1839 } 1840 /* Hexdump the affected line */ 1841 i = (i / 16) * 16; 1842 limit = 16; 1843 if (i + limit > size) 1844 limit = size - i; 1845 dump_line(realobj, i, limit); 1846 i += 16; 1847 lines++; 1848 /* Limit to 5 lines */ 1849 if (lines > 5) 1850 break; 1851 } 1852 } 1853 if (lines != 0) { 1854 /* Print some data about the neighboring objects, if they 1855 * exist: 1856 */ 1857 struct slab *slabp = virt_to_slab(objp); 1858 unsigned int objnr; 1859 1860 objnr = obj_to_index(cachep, slabp, objp); 1861 if (objnr) { 1862 objp = index_to_obj(cachep, slabp, objnr - 1); 1863 realobj = (char *)objp + obj_offset(cachep); 1864 printk(KERN_ERR "Prev obj: start=%p, len=%d\n", 1865 realobj, size); 1866 print_objinfo(cachep, objp, 2); 1867 } 1868 if (objnr + 1 < cachep->num) { 1869 objp = index_to_obj(cachep, slabp, objnr + 1); 1870 realobj = (char *)objp + obj_offset(cachep); 1871 printk(KERN_ERR "Next obj: start=%p, len=%d\n", 1872 realobj, size); 1873 print_objinfo(cachep, objp, 2); 1874 } 1875 } 1876 } 1877 #endif 1878 1879 #if DEBUG 1880 /** 1881 * slab_destroy_objs - destroy a slab and its objects 1882 * @cachep: cache pointer being destroyed 1883 * @slabp: slab pointer being destroyed 1884 * 1885 * Call the registered destructor for each object in a slab that is being 1886 * destroyed. 1887 */ 1888 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1889 { 1890 int i; 1891 for (i = 0; i < cachep->num; i++) { 1892 void *objp = index_to_obj(cachep, slabp, i); 1893 1894 if (cachep->flags & SLAB_POISON) { 1895 #ifdef CONFIG_DEBUG_PAGEALLOC 1896 if (cachep->buffer_size % PAGE_SIZE == 0 && 1897 OFF_SLAB(cachep)) 1898 kernel_map_pages(virt_to_page(objp), 1899 cachep->buffer_size / PAGE_SIZE, 1); 1900 else 1901 check_poison_obj(cachep, objp); 1902 #else 1903 check_poison_obj(cachep, objp); 1904 #endif 1905 } 1906 if (cachep->flags & SLAB_RED_ZONE) { 1907 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1908 slab_error(cachep, "start of a freed object " 1909 "was overwritten"); 1910 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1911 slab_error(cachep, "end of a freed object " 1912 "was overwritten"); 1913 } 1914 if (cachep->dtor && !(cachep->flags & SLAB_POISON)) 1915 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); 1916 } 1917 } 1918 #else 1919 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) 1920 { 1921 if (cachep->dtor) { 1922 int i; 1923 for (i = 0; i < cachep->num; i++) { 1924 void *objp = index_to_obj(cachep, slabp, i); 1925 (cachep->dtor) (objp, cachep, 0); 1926 } 1927 } 1928 } 1929 #endif 1930 1931 /** 1932 * slab_destroy - destroy and release all objects in a slab 1933 * @cachep: cache pointer being destroyed 1934 * @slabp: slab pointer being destroyed 1935 * 1936 * Destroy all the objs in a slab, and release the mem back to the system. 1937 * Before calling the slab must have been unlinked from the cache. The 1938 * cache-lock is not held/needed. 1939 */ 1940 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) 1941 { 1942 void *addr = slabp->s_mem - slabp->colouroff; 1943 1944 slab_destroy_objs(cachep, slabp); 1945 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { 1946 struct slab_rcu *slab_rcu; 1947 1948 slab_rcu = (struct slab_rcu *)slabp; 1949 slab_rcu->cachep = cachep; 1950 slab_rcu->addr = addr; 1951 call_rcu(&slab_rcu->head, kmem_rcu_free); 1952 } else { 1953 kmem_freepages(cachep, addr); 1954 if (OFF_SLAB(cachep)) 1955 kmem_cache_free(cachep->slabp_cache, slabp); 1956 } 1957 } 1958 1959 /* 1960 * For setting up all the kmem_list3s for cache whose buffer_size is same as 1961 * size of kmem_list3. 1962 */ 1963 static void __init set_up_list3s(struct kmem_cache *cachep, int index) 1964 { 1965 int node; 1966 1967 for_each_online_node(node) { 1968 cachep->nodelists[node] = &initkmem_list3[index + node]; 1969 cachep->nodelists[node]->next_reap = jiffies + 1970 REAPTIMEOUT_LIST3 + 1971 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 1972 } 1973 } 1974 1975 static void __kmem_cache_destroy(struct kmem_cache *cachep) 1976 { 1977 int i; 1978 struct kmem_list3 *l3; 1979 1980 for_each_online_cpu(i) 1981 kfree(cachep->array[i]); 1982 1983 /* NUMA: free the list3 structures */ 1984 for_each_online_node(i) { 1985 l3 = cachep->nodelists[i]; 1986 if (l3) { 1987 kfree(l3->shared); 1988 free_alien_cache(l3->alien); 1989 kfree(l3); 1990 } 1991 } 1992 kmem_cache_free(&cache_cache, cachep); 1993 } 1994 1995 1996 /** 1997 * calculate_slab_order - calculate size (page order) of slabs 1998 * @cachep: pointer to the cache that is being created 1999 * @size: size of objects to be created in this cache. 2000 * @align: required alignment for the objects. 2001 * @flags: slab allocation flags 2002 * 2003 * Also calculates the number of objects per slab. 2004 * 2005 * This could be made much more intelligent. For now, try to avoid using 2006 * high order pages for slabs. When the gfp() functions are more friendly 2007 * towards high-order requests, this should be changed. 2008 */ 2009 static size_t calculate_slab_order(struct kmem_cache *cachep, 2010 size_t size, size_t align, unsigned long flags) 2011 { 2012 unsigned long offslab_limit; 2013 size_t left_over = 0; 2014 int gfporder; 2015 2016 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { 2017 unsigned int num; 2018 size_t remainder; 2019 2020 cache_estimate(gfporder, size, align, flags, &remainder, &num); 2021 if (!num) 2022 continue; 2023 2024 if (flags & CFLGS_OFF_SLAB) { 2025 /* 2026 * Max number of objs-per-slab for caches which 2027 * use off-slab slabs. Needed to avoid a possible 2028 * looping condition in cache_grow(). 2029 */ 2030 offslab_limit = size - sizeof(struct slab); 2031 offslab_limit /= sizeof(kmem_bufctl_t); 2032 2033 if (num > offslab_limit) 2034 break; 2035 } 2036 2037 /* Found something acceptable - save it away */ 2038 cachep->num = num; 2039 cachep->gfporder = gfporder; 2040 left_over = remainder; 2041 2042 /* 2043 * A VFS-reclaimable slab tends to have most allocations 2044 * as GFP_NOFS and we really don't want to have to be allocating 2045 * higher-order pages when we are unable to shrink dcache. 2046 */ 2047 if (flags & SLAB_RECLAIM_ACCOUNT) 2048 break; 2049 2050 /* 2051 * Large number of objects is good, but very large slabs are 2052 * currently bad for the gfp()s. 2053 */ 2054 if (gfporder >= slab_break_gfp_order) 2055 break; 2056 2057 /* 2058 * Acceptable internal fragmentation? 2059 */ 2060 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 2061 break; 2062 } 2063 return left_over; 2064 } 2065 2066 static int setup_cpu_cache(struct kmem_cache *cachep) 2067 { 2068 if (g_cpucache_up == FULL) 2069 return enable_cpucache(cachep); 2070 2071 if (g_cpucache_up == NONE) { 2072 /* 2073 * Note: the first kmem_cache_create must create the cache 2074 * that's used by kmalloc(24), otherwise the creation of 2075 * further caches will BUG(). 2076 */ 2077 cachep->array[smp_processor_id()] = &initarray_generic.cache; 2078 2079 /* 2080 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is 2081 * the first cache, then we need to set up all its list3s, 2082 * otherwise the creation of further caches will BUG(). 2083 */ 2084 set_up_list3s(cachep, SIZE_AC); 2085 if (INDEX_AC == INDEX_L3) 2086 g_cpucache_up = PARTIAL_L3; 2087 else 2088 g_cpucache_up = PARTIAL_AC; 2089 } else { 2090 cachep->array[smp_processor_id()] = 2091 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); 2092 2093 if (g_cpucache_up == PARTIAL_AC) { 2094 set_up_list3s(cachep, SIZE_L3); 2095 g_cpucache_up = PARTIAL_L3; 2096 } else { 2097 int node; 2098 for_each_online_node(node) { 2099 cachep->nodelists[node] = 2100 kmalloc_node(sizeof(struct kmem_list3), 2101 GFP_KERNEL, node); 2102 BUG_ON(!cachep->nodelists[node]); 2103 kmem_list3_init(cachep->nodelists[node]); 2104 } 2105 } 2106 } 2107 cachep->nodelists[numa_node_id()]->next_reap = 2108 jiffies + REAPTIMEOUT_LIST3 + 2109 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 2110 2111 cpu_cache_get(cachep)->avail = 0; 2112 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 2113 cpu_cache_get(cachep)->batchcount = 1; 2114 cpu_cache_get(cachep)->touched = 0; 2115 cachep->batchcount = 1; 2116 cachep->limit = BOOT_CPUCACHE_ENTRIES; 2117 return 0; 2118 } 2119 2120 /** 2121 * kmem_cache_create - Create a cache. 2122 * @name: A string which is used in /proc/slabinfo to identify this cache. 2123 * @size: The size of objects to be created in this cache. 2124 * @align: The required alignment for the objects. 2125 * @flags: SLAB flags 2126 * @ctor: A constructor for the objects. 2127 * @dtor: A destructor for the objects. 2128 * 2129 * Returns a ptr to the cache on success, NULL on failure. 2130 * Cannot be called within a int, but can be interrupted. 2131 * The @ctor is run when new pages are allocated by the cache 2132 * and the @dtor is run before the pages are handed back. 2133 * 2134 * @name must be valid until the cache is destroyed. This implies that 2135 * the module calling this has to destroy the cache before getting unloaded. 2136 * 2137 * The flags are 2138 * 2139 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 2140 * to catch references to uninitialised memory. 2141 * 2142 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 2143 * for buffer overruns. 2144 * 2145 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 2146 * cacheline. This can be beneficial if you're counting cycles as closely 2147 * as davem. 2148 */ 2149 struct kmem_cache * 2150 kmem_cache_create (const char *name, size_t size, size_t align, 2151 unsigned long flags, 2152 void (*ctor)(void*, struct kmem_cache *, unsigned long), 2153 void (*dtor)(void*, struct kmem_cache *, unsigned long)) 2154 { 2155 size_t left_over, slab_size, ralign; 2156 struct kmem_cache *cachep = NULL, *pc; 2157 2158 /* 2159 * Sanity checks... these are all serious usage bugs. 2160 */ 2161 if (!name || in_interrupt() || (size < BYTES_PER_WORD) || 2162 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { 2163 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, 2164 name); 2165 BUG(); 2166 } 2167 2168 /* 2169 * We use cache_chain_mutex to ensure a consistent view of 2170 * cpu_online_map as well. Please see cpuup_callback 2171 */ 2172 mutex_lock(&cache_chain_mutex); 2173 2174 list_for_each_entry(pc, &cache_chain, next) { 2175 char tmp; 2176 int res; 2177 2178 /* 2179 * This happens when the module gets unloaded and doesn't 2180 * destroy its slab cache and no-one else reuses the vmalloc 2181 * area of the module. Print a warning. 2182 */ 2183 res = probe_kernel_address(pc->name, tmp); 2184 if (res) { 2185 printk(KERN_ERR 2186 "SLAB: cache with size %d has lost its name\n", 2187 pc->buffer_size); 2188 continue; 2189 } 2190 2191 if (!strcmp(pc->name, name)) { 2192 printk(KERN_ERR 2193 "kmem_cache_create: duplicate cache %s\n", name); 2194 dump_stack(); 2195 goto oops; 2196 } 2197 } 2198 2199 #if DEBUG 2200 WARN_ON(strchr(name, ' ')); /* It confuses parsers */ 2201 #if FORCED_DEBUG 2202 /* 2203 * Enable redzoning and last user accounting, except for caches with 2204 * large objects, if the increased size would increase the object size 2205 * above the next power of two: caches with object sizes just above a 2206 * power of two have a significant amount of internal fragmentation. 2207 */ 2208 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) 2209 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2210 if (!(flags & SLAB_DESTROY_BY_RCU)) 2211 flags |= SLAB_POISON; 2212 #endif 2213 if (flags & SLAB_DESTROY_BY_RCU) 2214 BUG_ON(flags & SLAB_POISON); 2215 #endif 2216 if (flags & SLAB_DESTROY_BY_RCU) 2217 BUG_ON(dtor); 2218 2219 /* 2220 * Always checks flags, a caller might be expecting debug support which 2221 * isn't available. 2222 */ 2223 BUG_ON(flags & ~CREATE_MASK); 2224 2225 /* 2226 * Check that size is in terms of words. This is needed to avoid 2227 * unaligned accesses for some archs when redzoning is used, and makes 2228 * sure any on-slab bufctl's are also correctly aligned. 2229 */ 2230 if (size & (BYTES_PER_WORD - 1)) { 2231 size += (BYTES_PER_WORD - 1); 2232 size &= ~(BYTES_PER_WORD - 1); 2233 } 2234 2235 /* calculate the final buffer alignment: */ 2236 2237 /* 1) arch recommendation: can be overridden for debug */ 2238 if (flags & SLAB_HWCACHE_ALIGN) { 2239 /* 2240 * Default alignment: as specified by the arch code. Except if 2241 * an object is really small, then squeeze multiple objects into 2242 * one cacheline. 2243 */ 2244 ralign = cache_line_size(); 2245 while (size <= ralign / 2) 2246 ralign /= 2; 2247 } else { 2248 ralign = BYTES_PER_WORD; 2249 } 2250 2251 /* 2252 * Redzoning and user store require word alignment. Note this will be 2253 * overridden by architecture or caller mandated alignment if either 2254 * is greater than BYTES_PER_WORD. 2255 */ 2256 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) 2257 ralign = __alignof__(unsigned long long); 2258 2259 /* 2) arch mandated alignment */ 2260 if (ralign < ARCH_SLAB_MINALIGN) { 2261 ralign = ARCH_SLAB_MINALIGN; 2262 } 2263 /* 3) caller mandated alignment */ 2264 if (ralign < align) { 2265 ralign = align; 2266 } 2267 /* disable debug if necessary */ 2268 if (ralign > __alignof__(unsigned long long)) 2269 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2270 /* 2271 * 4) Store it. 2272 */ 2273 align = ralign; 2274 2275 /* Get cache's description obj. */ 2276 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); 2277 if (!cachep) 2278 goto oops; 2279 2280 #if DEBUG 2281 cachep->obj_size = size; 2282 2283 /* 2284 * Both debugging options require word-alignment which is calculated 2285 * into align above. 2286 */ 2287 if (flags & SLAB_RED_ZONE) { 2288 /* add space for red zone words */ 2289 cachep->obj_offset += sizeof(unsigned long long); 2290 size += 2 * sizeof(unsigned long long); 2291 } 2292 if (flags & SLAB_STORE_USER) { 2293 /* user store requires one word storage behind the end of 2294 * the real object. 2295 */ 2296 size += BYTES_PER_WORD; 2297 } 2298 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) 2299 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size 2300 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { 2301 cachep->obj_offset += PAGE_SIZE - size; 2302 size = PAGE_SIZE; 2303 } 2304 #endif 2305 #endif 2306 2307 /* 2308 * Determine if the slab management is 'on' or 'off' slab. 2309 * (bootstrapping cannot cope with offslab caches so don't do 2310 * it too early on.) 2311 */ 2312 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) 2313 /* 2314 * Size is large, assume best to place the slab management obj 2315 * off-slab (should allow better packing of objs). 2316 */ 2317 flags |= CFLGS_OFF_SLAB; 2318 2319 size = ALIGN(size, align); 2320 2321 left_over = calculate_slab_order(cachep, size, align, flags); 2322 2323 if (!cachep->num) { 2324 printk(KERN_ERR 2325 "kmem_cache_create: couldn't create cache %s.\n", name); 2326 kmem_cache_free(&cache_cache, cachep); 2327 cachep = NULL; 2328 goto oops; 2329 } 2330 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) 2331 + sizeof(struct slab), align); 2332 2333 /* 2334 * If the slab has been placed off-slab, and we have enough space then 2335 * move it on-slab. This is at the expense of any extra colouring. 2336 */ 2337 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { 2338 flags &= ~CFLGS_OFF_SLAB; 2339 left_over -= slab_size; 2340 } 2341 2342 if (flags & CFLGS_OFF_SLAB) { 2343 /* really off slab. No need for manual alignment */ 2344 slab_size = 2345 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); 2346 } 2347 2348 cachep->colour_off = cache_line_size(); 2349 /* Offset must be a multiple of the alignment. */ 2350 if (cachep->colour_off < align) 2351 cachep->colour_off = align; 2352 cachep->colour = left_over / cachep->colour_off; 2353 cachep->slab_size = slab_size; 2354 cachep->flags = flags; 2355 cachep->gfpflags = 0; 2356 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) 2357 cachep->gfpflags |= GFP_DMA; 2358 cachep->buffer_size = size; 2359 cachep->reciprocal_buffer_size = reciprocal_value(size); 2360 2361 if (flags & CFLGS_OFF_SLAB) { 2362 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); 2363 /* 2364 * This is a possibility for one of the malloc_sizes caches. 2365 * But since we go off slab only for object size greater than 2366 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, 2367 * this should not happen at all. 2368 * But leave a BUG_ON for some lucky dude. 2369 */ 2370 BUG_ON(!cachep->slabp_cache); 2371 } 2372 cachep->ctor = ctor; 2373 cachep->dtor = dtor; 2374 cachep->name = name; 2375 2376 if (setup_cpu_cache(cachep)) { 2377 __kmem_cache_destroy(cachep); 2378 cachep = NULL; 2379 goto oops; 2380 } 2381 2382 /* cache setup completed, link it into the list */ 2383 list_add(&cachep->next, &cache_chain); 2384 oops: 2385 if (!cachep && (flags & SLAB_PANIC)) 2386 panic("kmem_cache_create(): failed to create slab `%s'\n", 2387 name); 2388 mutex_unlock(&cache_chain_mutex); 2389 return cachep; 2390 } 2391 EXPORT_SYMBOL(kmem_cache_create); 2392 2393 #if DEBUG 2394 static void check_irq_off(void) 2395 { 2396 BUG_ON(!irqs_disabled()); 2397 } 2398 2399 static void check_irq_on(void) 2400 { 2401 BUG_ON(irqs_disabled()); 2402 } 2403 2404 static void check_spinlock_acquired(struct kmem_cache *cachep) 2405 { 2406 #ifdef CONFIG_SMP 2407 check_irq_off(); 2408 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); 2409 #endif 2410 } 2411 2412 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2413 { 2414 #ifdef CONFIG_SMP 2415 check_irq_off(); 2416 assert_spin_locked(&cachep->nodelists[node]->list_lock); 2417 #endif 2418 } 2419 2420 #else 2421 #define check_irq_off() do { } while(0) 2422 #define check_irq_on() do { } while(0) 2423 #define check_spinlock_acquired(x) do { } while(0) 2424 #define check_spinlock_acquired_node(x, y) do { } while(0) 2425 #endif 2426 2427 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 2428 struct array_cache *ac, 2429 int force, int node); 2430 2431 static void do_drain(void *arg) 2432 { 2433 struct kmem_cache *cachep = arg; 2434 struct array_cache *ac; 2435 int node = numa_node_id(); 2436 2437 check_irq_off(); 2438 ac = cpu_cache_get(cachep); 2439 spin_lock(&cachep->nodelists[node]->list_lock); 2440 free_block(cachep, ac->entry, ac->avail, node); 2441 spin_unlock(&cachep->nodelists[node]->list_lock); 2442 ac->avail = 0; 2443 } 2444 2445 static void drain_cpu_caches(struct kmem_cache *cachep) 2446 { 2447 struct kmem_list3 *l3; 2448 int node; 2449 2450 on_each_cpu(do_drain, cachep, 1, 1); 2451 check_irq_on(); 2452 for_each_online_node(node) { 2453 l3 = cachep->nodelists[node]; 2454 if (l3 && l3->alien) 2455 drain_alien_cache(cachep, l3->alien); 2456 } 2457 2458 for_each_online_node(node) { 2459 l3 = cachep->nodelists[node]; 2460 if (l3) 2461 drain_array(cachep, l3, l3->shared, 1, node); 2462 } 2463 } 2464 2465 /* 2466 * Remove slabs from the list of free slabs. 2467 * Specify the number of slabs to drain in tofree. 2468 * 2469 * Returns the actual number of slabs released. 2470 */ 2471 static int drain_freelist(struct kmem_cache *cache, 2472 struct kmem_list3 *l3, int tofree) 2473 { 2474 struct list_head *p; 2475 int nr_freed; 2476 struct slab *slabp; 2477 2478 nr_freed = 0; 2479 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { 2480 2481 spin_lock_irq(&l3->list_lock); 2482 p = l3->slabs_free.prev; 2483 if (p == &l3->slabs_free) { 2484 spin_unlock_irq(&l3->list_lock); 2485 goto out; 2486 } 2487 2488 slabp = list_entry(p, struct slab, list); 2489 #if DEBUG 2490 BUG_ON(slabp->inuse); 2491 #endif 2492 list_del(&slabp->list); 2493 /* 2494 * Safe to drop the lock. The slab is no longer linked 2495 * to the cache. 2496 */ 2497 l3->free_objects -= cache->num; 2498 spin_unlock_irq(&l3->list_lock); 2499 slab_destroy(cache, slabp); 2500 nr_freed++; 2501 } 2502 out: 2503 return nr_freed; 2504 } 2505 2506 /* Called with cache_chain_mutex held to protect against cpu hotplug */ 2507 static int __cache_shrink(struct kmem_cache *cachep) 2508 { 2509 int ret = 0, i = 0; 2510 struct kmem_list3 *l3; 2511 2512 drain_cpu_caches(cachep); 2513 2514 check_irq_on(); 2515 for_each_online_node(i) { 2516 l3 = cachep->nodelists[i]; 2517 if (!l3) 2518 continue; 2519 2520 drain_freelist(cachep, l3, l3->free_objects); 2521 2522 ret += !list_empty(&l3->slabs_full) || 2523 !list_empty(&l3->slabs_partial); 2524 } 2525 return (ret ? 1 : 0); 2526 } 2527 2528 /** 2529 * kmem_cache_shrink - Shrink a cache. 2530 * @cachep: The cache to shrink. 2531 * 2532 * Releases as many slabs as possible for a cache. 2533 * To help debugging, a zero exit status indicates all slabs were released. 2534 */ 2535 int kmem_cache_shrink(struct kmem_cache *cachep) 2536 { 2537 int ret; 2538 BUG_ON(!cachep || in_interrupt()); 2539 2540 mutex_lock(&cache_chain_mutex); 2541 ret = __cache_shrink(cachep); 2542 mutex_unlock(&cache_chain_mutex); 2543 return ret; 2544 } 2545 EXPORT_SYMBOL(kmem_cache_shrink); 2546 2547 /** 2548 * kmem_cache_destroy - delete a cache 2549 * @cachep: the cache to destroy 2550 * 2551 * Remove a &struct kmem_cache object from the slab cache. 2552 * 2553 * It is expected this function will be called by a module when it is 2554 * unloaded. This will remove the cache completely, and avoid a duplicate 2555 * cache being allocated each time a module is loaded and unloaded, if the 2556 * module doesn't have persistent in-kernel storage across loads and unloads. 2557 * 2558 * The cache must be empty before calling this function. 2559 * 2560 * The caller must guarantee that noone will allocate memory from the cache 2561 * during the kmem_cache_destroy(). 2562 */ 2563 void kmem_cache_destroy(struct kmem_cache *cachep) 2564 { 2565 BUG_ON(!cachep || in_interrupt()); 2566 2567 /* Find the cache in the chain of caches. */ 2568 mutex_lock(&cache_chain_mutex); 2569 /* 2570 * the chain is never empty, cache_cache is never destroyed 2571 */ 2572 list_del(&cachep->next); 2573 if (__cache_shrink(cachep)) { 2574 slab_error(cachep, "Can't free all objects"); 2575 list_add(&cachep->next, &cache_chain); 2576 mutex_unlock(&cache_chain_mutex); 2577 return; 2578 } 2579 2580 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 2581 synchronize_rcu(); 2582 2583 __kmem_cache_destroy(cachep); 2584 mutex_unlock(&cache_chain_mutex); 2585 } 2586 EXPORT_SYMBOL(kmem_cache_destroy); 2587 2588 /* 2589 * Get the memory for a slab management obj. 2590 * For a slab cache when the slab descriptor is off-slab, slab descriptors 2591 * always come from malloc_sizes caches. The slab descriptor cannot 2592 * come from the same cache which is getting created because, 2593 * when we are searching for an appropriate cache for these 2594 * descriptors in kmem_cache_create, we search through the malloc_sizes array. 2595 * If we are creating a malloc_sizes cache here it would not be visible to 2596 * kmem_find_general_cachep till the initialization is complete. 2597 * Hence we cannot have slabp_cache same as the original cache. 2598 */ 2599 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, 2600 int colour_off, gfp_t local_flags, 2601 int nodeid) 2602 { 2603 struct slab *slabp; 2604 2605 if (OFF_SLAB(cachep)) { 2606 /* Slab management obj is off-slab. */ 2607 slabp = kmem_cache_alloc_node(cachep->slabp_cache, 2608 local_flags & ~GFP_THISNODE, nodeid); 2609 if (!slabp) 2610 return NULL; 2611 } else { 2612 slabp = objp + colour_off; 2613 colour_off += cachep->slab_size; 2614 } 2615 slabp->inuse = 0; 2616 slabp->colouroff = colour_off; 2617 slabp->s_mem = objp + colour_off; 2618 slabp->nodeid = nodeid; 2619 return slabp; 2620 } 2621 2622 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) 2623 { 2624 return (kmem_bufctl_t *) (slabp + 1); 2625 } 2626 2627 static void cache_init_objs(struct kmem_cache *cachep, 2628 struct slab *slabp, unsigned long ctor_flags) 2629 { 2630 int i; 2631 2632 for (i = 0; i < cachep->num; i++) { 2633 void *objp = index_to_obj(cachep, slabp, i); 2634 #if DEBUG 2635 /* need to poison the objs? */ 2636 if (cachep->flags & SLAB_POISON) 2637 poison_obj(cachep, objp, POISON_FREE); 2638 if (cachep->flags & SLAB_STORE_USER) 2639 *dbg_userword(cachep, objp) = NULL; 2640 2641 if (cachep->flags & SLAB_RED_ZONE) { 2642 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2643 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2644 } 2645 /* 2646 * Constructors are not allowed to allocate memory from the same 2647 * cache which they are a constructor for. Otherwise, deadlock. 2648 * They must also be threaded. 2649 */ 2650 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) 2651 cachep->ctor(objp + obj_offset(cachep), cachep, 2652 ctor_flags); 2653 2654 if (cachep->flags & SLAB_RED_ZONE) { 2655 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2656 slab_error(cachep, "constructor overwrote the" 2657 " end of an object"); 2658 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2659 slab_error(cachep, "constructor overwrote the" 2660 " start of an object"); 2661 } 2662 if ((cachep->buffer_size % PAGE_SIZE) == 0 && 2663 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) 2664 kernel_map_pages(virt_to_page(objp), 2665 cachep->buffer_size / PAGE_SIZE, 0); 2666 #else 2667 if (cachep->ctor) 2668 cachep->ctor(objp, cachep, ctor_flags); 2669 #endif 2670 slab_bufctl(slabp)[i] = i + 1; 2671 } 2672 slab_bufctl(slabp)[i - 1] = BUFCTL_END; 2673 slabp->free = 0; 2674 } 2675 2676 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) 2677 { 2678 if (CONFIG_ZONE_DMA_FLAG) { 2679 if (flags & GFP_DMA) 2680 BUG_ON(!(cachep->gfpflags & GFP_DMA)); 2681 else 2682 BUG_ON(cachep->gfpflags & GFP_DMA); 2683 } 2684 } 2685 2686 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, 2687 int nodeid) 2688 { 2689 void *objp = index_to_obj(cachep, slabp, slabp->free); 2690 kmem_bufctl_t next; 2691 2692 slabp->inuse++; 2693 next = slab_bufctl(slabp)[slabp->free]; 2694 #if DEBUG 2695 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; 2696 WARN_ON(slabp->nodeid != nodeid); 2697 #endif 2698 slabp->free = next; 2699 2700 return objp; 2701 } 2702 2703 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, 2704 void *objp, int nodeid) 2705 { 2706 unsigned int objnr = obj_to_index(cachep, slabp, objp); 2707 2708 #if DEBUG 2709 /* Verify that the slab belongs to the intended node */ 2710 WARN_ON(slabp->nodeid != nodeid); 2711 2712 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { 2713 printk(KERN_ERR "slab: double free detected in cache " 2714 "'%s', objp %p\n", cachep->name, objp); 2715 BUG(); 2716 } 2717 #endif 2718 slab_bufctl(slabp)[objnr] = slabp->free; 2719 slabp->free = objnr; 2720 slabp->inuse--; 2721 } 2722 2723 /* 2724 * Map pages beginning at addr to the given cache and slab. This is required 2725 * for the slab allocator to be able to lookup the cache and slab of a 2726 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. 2727 */ 2728 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, 2729 void *addr) 2730 { 2731 int nr_pages; 2732 struct page *page; 2733 2734 page = virt_to_page(addr); 2735 2736 nr_pages = 1; 2737 if (likely(!PageCompound(page))) 2738 nr_pages <<= cache->gfporder; 2739 2740 do { 2741 page_set_cache(page, cache); 2742 page_set_slab(page, slab); 2743 page++; 2744 } while (--nr_pages); 2745 } 2746 2747 /* 2748 * Grow (by 1) the number of slabs within a cache. This is called by 2749 * kmem_cache_alloc() when there are no active objs left in a cache. 2750 */ 2751 static int cache_grow(struct kmem_cache *cachep, 2752 gfp_t flags, int nodeid, void *objp) 2753 { 2754 struct slab *slabp; 2755 size_t offset; 2756 gfp_t local_flags; 2757 unsigned long ctor_flags; 2758 struct kmem_list3 *l3; 2759 2760 /* 2761 * Be lazy and only check for valid flags here, keeping it out of the 2762 * critical path in kmem_cache_alloc(). 2763 */ 2764 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK)); 2765 2766 ctor_flags = SLAB_CTOR_CONSTRUCTOR; 2767 local_flags = (flags & GFP_LEVEL_MASK); 2768 /* Take the l3 list lock to change the colour_next on this node */ 2769 check_irq_off(); 2770 l3 = cachep->nodelists[nodeid]; 2771 spin_lock(&l3->list_lock); 2772 2773 /* Get colour for the slab, and cal the next value. */ 2774 offset = l3->colour_next; 2775 l3->colour_next++; 2776 if (l3->colour_next >= cachep->colour) 2777 l3->colour_next = 0; 2778 spin_unlock(&l3->list_lock); 2779 2780 offset *= cachep->colour_off; 2781 2782 if (local_flags & __GFP_WAIT) 2783 local_irq_enable(); 2784 2785 /* 2786 * The test for missing atomic flag is performed here, rather than 2787 * the more obvious place, simply to reduce the critical path length 2788 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they 2789 * will eventually be caught here (where it matters). 2790 */ 2791 kmem_flagcheck(cachep, flags); 2792 2793 /* 2794 * Get mem for the objs. Attempt to allocate a physical page from 2795 * 'nodeid'. 2796 */ 2797 if (!objp) 2798 objp = kmem_getpages(cachep, flags, nodeid); 2799 if (!objp) 2800 goto failed; 2801 2802 /* Get slab management. */ 2803 slabp = alloc_slabmgmt(cachep, objp, offset, 2804 local_flags & ~GFP_THISNODE, nodeid); 2805 if (!slabp) 2806 goto opps1; 2807 2808 slabp->nodeid = nodeid; 2809 slab_map_pages(cachep, slabp, objp); 2810 2811 cache_init_objs(cachep, slabp, ctor_flags); 2812 2813 if (local_flags & __GFP_WAIT) 2814 local_irq_disable(); 2815 check_irq_off(); 2816 spin_lock(&l3->list_lock); 2817 2818 /* Make slab active. */ 2819 list_add_tail(&slabp->list, &(l3->slabs_free)); 2820 STATS_INC_GROWN(cachep); 2821 l3->free_objects += cachep->num; 2822 spin_unlock(&l3->list_lock); 2823 return 1; 2824 opps1: 2825 kmem_freepages(cachep, objp); 2826 failed: 2827 if (local_flags & __GFP_WAIT) 2828 local_irq_disable(); 2829 return 0; 2830 } 2831 2832 #if DEBUG 2833 2834 /* 2835 * Perform extra freeing checks: 2836 * - detect bad pointers. 2837 * - POISON/RED_ZONE checking 2838 * - destructor calls, for caches with POISON+dtor 2839 */ 2840 static void kfree_debugcheck(const void *objp) 2841 { 2842 if (!virt_addr_valid(objp)) { 2843 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", 2844 (unsigned long)objp); 2845 BUG(); 2846 } 2847 } 2848 2849 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2850 { 2851 unsigned long long redzone1, redzone2; 2852 2853 redzone1 = *dbg_redzone1(cache, obj); 2854 redzone2 = *dbg_redzone2(cache, obj); 2855 2856 /* 2857 * Redzone is ok. 2858 */ 2859 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2860 return; 2861 2862 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2863 slab_error(cache, "double free detected"); 2864 else 2865 slab_error(cache, "memory outside object was overwritten"); 2866 2867 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", 2868 obj, redzone1, redzone2); 2869 } 2870 2871 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2872 void *caller) 2873 { 2874 struct page *page; 2875 unsigned int objnr; 2876 struct slab *slabp; 2877 2878 objp -= obj_offset(cachep); 2879 kfree_debugcheck(objp); 2880 page = virt_to_head_page(objp); 2881 2882 slabp = page_get_slab(page); 2883 2884 if (cachep->flags & SLAB_RED_ZONE) { 2885 verify_redzone_free(cachep, objp); 2886 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2887 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2888 } 2889 if (cachep->flags & SLAB_STORE_USER) 2890 *dbg_userword(cachep, objp) = caller; 2891 2892 objnr = obj_to_index(cachep, slabp, objp); 2893 2894 BUG_ON(objnr >= cachep->num); 2895 BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); 2896 2897 if (cachep->flags & SLAB_POISON && cachep->dtor) { 2898 /* we want to cache poison the object, 2899 * call the destruction callback 2900 */ 2901 cachep->dtor(objp + obj_offset(cachep), cachep, 0); 2902 } 2903 #ifdef CONFIG_DEBUG_SLAB_LEAK 2904 slab_bufctl(slabp)[objnr] = BUFCTL_FREE; 2905 #endif 2906 if (cachep->flags & SLAB_POISON) { 2907 #ifdef CONFIG_DEBUG_PAGEALLOC 2908 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { 2909 store_stackinfo(cachep, objp, (unsigned long)caller); 2910 kernel_map_pages(virt_to_page(objp), 2911 cachep->buffer_size / PAGE_SIZE, 0); 2912 } else { 2913 poison_obj(cachep, objp, POISON_FREE); 2914 } 2915 #else 2916 poison_obj(cachep, objp, POISON_FREE); 2917 #endif 2918 } 2919 return objp; 2920 } 2921 2922 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) 2923 { 2924 kmem_bufctl_t i; 2925 int entries = 0; 2926 2927 /* Check slab's freelist to see if this obj is there. */ 2928 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { 2929 entries++; 2930 if (entries > cachep->num || i >= cachep->num) 2931 goto bad; 2932 } 2933 if (entries != cachep->num - slabp->inuse) { 2934 bad: 2935 printk(KERN_ERR "slab: Internal list corruption detected in " 2936 "cache '%s'(%d), slabp %p(%d). Hexdump:\n", 2937 cachep->name, cachep->num, slabp, slabp->inuse); 2938 for (i = 0; 2939 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); 2940 i++) { 2941 if (i % 16 == 0) 2942 printk("\n%03x:", i); 2943 printk(" %02x", ((unsigned char *)slabp)[i]); 2944 } 2945 printk("\n"); 2946 BUG(); 2947 } 2948 } 2949 #else 2950 #define kfree_debugcheck(x) do { } while(0) 2951 #define cache_free_debugcheck(x,objp,z) (objp) 2952 #define check_slabp(x,y) do { } while(0) 2953 #endif 2954 2955 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2956 { 2957 int batchcount; 2958 struct kmem_list3 *l3; 2959 struct array_cache *ac; 2960 int node; 2961 2962 node = numa_node_id(); 2963 2964 check_irq_off(); 2965 ac = cpu_cache_get(cachep); 2966 retry: 2967 batchcount = ac->batchcount; 2968 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2969 /* 2970 * If there was little recent activity on this cache, then 2971 * perform only a partial refill. Otherwise we could generate 2972 * refill bouncing. 2973 */ 2974 batchcount = BATCHREFILL_LIMIT; 2975 } 2976 l3 = cachep->nodelists[node]; 2977 2978 BUG_ON(ac->avail > 0 || !l3); 2979 spin_lock(&l3->list_lock); 2980 2981 /* See if we can refill from the shared array */ 2982 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) 2983 goto alloc_done; 2984 2985 while (batchcount > 0) { 2986 struct list_head *entry; 2987 struct slab *slabp; 2988 /* Get slab alloc is to come from. */ 2989 entry = l3->slabs_partial.next; 2990 if (entry == &l3->slabs_partial) { 2991 l3->free_touched = 1; 2992 entry = l3->slabs_free.next; 2993 if (entry == &l3->slabs_free) 2994 goto must_grow; 2995 } 2996 2997 slabp = list_entry(entry, struct slab, list); 2998 check_slabp(cachep, slabp); 2999 check_spinlock_acquired(cachep); 3000 3001 /* 3002 * The slab was either on partial or free list so 3003 * there must be at least one object available for 3004 * allocation. 3005 */ 3006 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num); 3007 3008 while (slabp->inuse < cachep->num && batchcount--) { 3009 STATS_INC_ALLOCED(cachep); 3010 STATS_INC_ACTIVE(cachep); 3011 STATS_SET_HIGH(cachep); 3012 3013 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, 3014 node); 3015 } 3016 check_slabp(cachep, slabp); 3017 3018 /* move slabp to correct slabp list: */ 3019 list_del(&slabp->list); 3020 if (slabp->free == BUFCTL_END) 3021 list_add(&slabp->list, &l3->slabs_full); 3022 else 3023 list_add(&slabp->list, &l3->slabs_partial); 3024 } 3025 3026 must_grow: 3027 l3->free_objects -= ac->avail; 3028 alloc_done: 3029 spin_unlock(&l3->list_lock); 3030 3031 if (unlikely(!ac->avail)) { 3032 int x; 3033 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); 3034 3035 /* cache_grow can reenable interrupts, then ac could change. */ 3036 ac = cpu_cache_get(cachep); 3037 if (!x && ac->avail == 0) /* no objects in sight? abort */ 3038 return NULL; 3039 3040 if (!ac->avail) /* objects refilled by interrupt? */ 3041 goto retry; 3042 } 3043 ac->touched = 1; 3044 return ac->entry[--ac->avail]; 3045 } 3046 3047 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3048 gfp_t flags) 3049 { 3050 might_sleep_if(flags & __GFP_WAIT); 3051 #if DEBUG 3052 kmem_flagcheck(cachep, flags); 3053 #endif 3054 } 3055 3056 #if DEBUG 3057 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3058 gfp_t flags, void *objp, void *caller) 3059 { 3060 if (!objp) 3061 return objp; 3062 if (cachep->flags & SLAB_POISON) { 3063 #ifdef CONFIG_DEBUG_PAGEALLOC 3064 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) 3065 kernel_map_pages(virt_to_page(objp), 3066 cachep->buffer_size / PAGE_SIZE, 1); 3067 else 3068 check_poison_obj(cachep, objp); 3069 #else 3070 check_poison_obj(cachep, objp); 3071 #endif 3072 poison_obj(cachep, objp, POISON_INUSE); 3073 } 3074 if (cachep->flags & SLAB_STORE_USER) 3075 *dbg_userword(cachep, objp) = caller; 3076 3077 if (cachep->flags & SLAB_RED_ZONE) { 3078 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3079 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3080 slab_error(cachep, "double free, or memory outside" 3081 " object was overwritten"); 3082 printk(KERN_ERR 3083 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", 3084 objp, *dbg_redzone1(cachep, objp), 3085 *dbg_redzone2(cachep, objp)); 3086 } 3087 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3088 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3089 } 3090 #ifdef CONFIG_DEBUG_SLAB_LEAK 3091 { 3092 struct slab *slabp; 3093 unsigned objnr; 3094 3095 slabp = page_get_slab(virt_to_head_page(objp)); 3096 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; 3097 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; 3098 } 3099 #endif 3100 objp += obj_offset(cachep); 3101 if (cachep->ctor && cachep->flags & SLAB_POISON) 3102 cachep->ctor(objp, cachep, SLAB_CTOR_CONSTRUCTOR); 3103 #if ARCH_SLAB_MINALIGN 3104 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { 3105 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3106 objp, ARCH_SLAB_MINALIGN); 3107 } 3108 #endif 3109 return objp; 3110 } 3111 #else 3112 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3113 #endif 3114 3115 #ifdef CONFIG_FAILSLAB 3116 3117 static struct failslab_attr { 3118 3119 struct fault_attr attr; 3120 3121 u32 ignore_gfp_wait; 3122 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3123 struct dentry *ignore_gfp_wait_file; 3124 #endif 3125 3126 } failslab = { 3127 .attr = FAULT_ATTR_INITIALIZER, 3128 .ignore_gfp_wait = 1, 3129 }; 3130 3131 static int __init setup_failslab(char *str) 3132 { 3133 return setup_fault_attr(&failslab.attr, str); 3134 } 3135 __setup("failslab=", setup_failslab); 3136 3137 static int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3138 { 3139 if (cachep == &cache_cache) 3140 return 0; 3141 if (flags & __GFP_NOFAIL) 3142 return 0; 3143 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) 3144 return 0; 3145 3146 return should_fail(&failslab.attr, obj_size(cachep)); 3147 } 3148 3149 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 3150 3151 static int __init failslab_debugfs(void) 3152 { 3153 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 3154 struct dentry *dir; 3155 int err; 3156 3157 err = init_fault_attr_dentries(&failslab.attr, "failslab"); 3158 if (err) 3159 return err; 3160 dir = failslab.attr.dentries.dir; 3161 3162 failslab.ignore_gfp_wait_file = 3163 debugfs_create_bool("ignore-gfp-wait", mode, dir, 3164 &failslab.ignore_gfp_wait); 3165 3166 if (!failslab.ignore_gfp_wait_file) { 3167 err = -ENOMEM; 3168 debugfs_remove(failslab.ignore_gfp_wait_file); 3169 cleanup_fault_attr_dentries(&failslab.attr); 3170 } 3171 3172 return err; 3173 } 3174 3175 late_initcall(failslab_debugfs); 3176 3177 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 3178 3179 #else /* CONFIG_FAILSLAB */ 3180 3181 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) 3182 { 3183 return 0; 3184 } 3185 3186 #endif /* CONFIG_FAILSLAB */ 3187 3188 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3189 { 3190 void *objp; 3191 struct array_cache *ac; 3192 3193 check_irq_off(); 3194 3195 ac = cpu_cache_get(cachep); 3196 if (likely(ac->avail)) { 3197 STATS_INC_ALLOCHIT(cachep); 3198 ac->touched = 1; 3199 objp = ac->entry[--ac->avail]; 3200 } else { 3201 STATS_INC_ALLOCMISS(cachep); 3202 objp = cache_alloc_refill(cachep, flags); 3203 } 3204 return objp; 3205 } 3206 3207 #ifdef CONFIG_NUMA 3208 /* 3209 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. 3210 * 3211 * If we are in_interrupt, then process context, including cpusets and 3212 * mempolicy, may not apply and should not be used for allocation policy. 3213 */ 3214 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3215 { 3216 int nid_alloc, nid_here; 3217 3218 if (in_interrupt() || (flags & __GFP_THISNODE)) 3219 return NULL; 3220 nid_alloc = nid_here = numa_node_id(); 3221 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3222 nid_alloc = cpuset_mem_spread_node(); 3223 else if (current->mempolicy) 3224 nid_alloc = slab_node(current->mempolicy); 3225 if (nid_alloc != nid_here) 3226 return ____cache_alloc_node(cachep, flags, nid_alloc); 3227 return NULL; 3228 } 3229 3230 /* 3231 * Fallback function if there was no memory available and no objects on a 3232 * certain node and fall back is permitted. First we scan all the 3233 * available nodelists for available objects. If that fails then we 3234 * perform an allocation without specifying a node. This allows the page 3235 * allocator to do its reclaim / fallback magic. We then insert the 3236 * slab into the proper nodelist and then allocate from it. 3237 */ 3238 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3239 { 3240 struct zonelist *zonelist; 3241 gfp_t local_flags; 3242 struct zone **z; 3243 void *obj = NULL; 3244 int nid; 3245 3246 if (flags & __GFP_THISNODE) 3247 return NULL; 3248 3249 zonelist = &NODE_DATA(slab_node(current->mempolicy)) 3250 ->node_zonelists[gfp_zone(flags)]; 3251 local_flags = (flags & GFP_LEVEL_MASK); 3252 3253 retry: 3254 /* 3255 * Look through allowed nodes for objects available 3256 * from existing per node queues. 3257 */ 3258 for (z = zonelist->zones; *z && !obj; z++) { 3259 nid = zone_to_nid(*z); 3260 3261 if (cpuset_zone_allowed_hardwall(*z, flags) && 3262 cache->nodelists[nid] && 3263 cache->nodelists[nid]->free_objects) 3264 obj = ____cache_alloc_node(cache, 3265 flags | GFP_THISNODE, nid); 3266 } 3267 3268 if (!obj) { 3269 /* 3270 * This allocation will be performed within the constraints 3271 * of the current cpuset / memory policy requirements. 3272 * We may trigger various forms of reclaim on the allowed 3273 * set and go into memory reserves if necessary. 3274 */ 3275 if (local_flags & __GFP_WAIT) 3276 local_irq_enable(); 3277 kmem_flagcheck(cache, flags); 3278 obj = kmem_getpages(cache, flags, -1); 3279 if (local_flags & __GFP_WAIT) 3280 local_irq_disable(); 3281 if (obj) { 3282 /* 3283 * Insert into the appropriate per node queues 3284 */ 3285 nid = page_to_nid(virt_to_page(obj)); 3286 if (cache_grow(cache, flags, nid, obj)) { 3287 obj = ____cache_alloc_node(cache, 3288 flags | GFP_THISNODE, nid); 3289 if (!obj) 3290 /* 3291 * Another processor may allocate the 3292 * objects in the slab since we are 3293 * not holding any locks. 3294 */ 3295 goto retry; 3296 } else { 3297 /* cache_grow already freed obj */ 3298 obj = NULL; 3299 } 3300 } 3301 } 3302 return obj; 3303 } 3304 3305 /* 3306 * A interface to enable slab creation on nodeid 3307 */ 3308 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3309 int nodeid) 3310 { 3311 struct list_head *entry; 3312 struct slab *slabp; 3313 struct kmem_list3 *l3; 3314 void *obj; 3315 int x; 3316 3317 l3 = cachep->nodelists[nodeid]; 3318 BUG_ON(!l3); 3319 3320 retry: 3321 check_irq_off(); 3322 spin_lock(&l3->list_lock); 3323 entry = l3->slabs_partial.next; 3324 if (entry == &l3->slabs_partial) { 3325 l3->free_touched = 1; 3326 entry = l3->slabs_free.next; 3327 if (entry == &l3->slabs_free) 3328 goto must_grow; 3329 } 3330 3331 slabp = list_entry(entry, struct slab, list); 3332 check_spinlock_acquired_node(cachep, nodeid); 3333 check_slabp(cachep, slabp); 3334 3335 STATS_INC_NODEALLOCS(cachep); 3336 STATS_INC_ACTIVE(cachep); 3337 STATS_SET_HIGH(cachep); 3338 3339 BUG_ON(slabp->inuse == cachep->num); 3340 3341 obj = slab_get_obj(cachep, slabp, nodeid); 3342 check_slabp(cachep, slabp); 3343 l3->free_objects--; 3344 /* move slabp to correct slabp list: */ 3345 list_del(&slabp->list); 3346 3347 if (slabp->free == BUFCTL_END) 3348 list_add(&slabp->list, &l3->slabs_full); 3349 else 3350 list_add(&slabp->list, &l3->slabs_partial); 3351 3352 spin_unlock(&l3->list_lock); 3353 goto done; 3354 3355 must_grow: 3356 spin_unlock(&l3->list_lock); 3357 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); 3358 if (x) 3359 goto retry; 3360 3361 return fallback_alloc(cachep, flags); 3362 3363 done: 3364 return obj; 3365 } 3366 3367 /** 3368 * kmem_cache_alloc_node - Allocate an object on the specified node 3369 * @cachep: The cache to allocate from. 3370 * @flags: See kmalloc(). 3371 * @nodeid: node number of the target node. 3372 * @caller: return address of caller, used for debug information 3373 * 3374 * Identical to kmem_cache_alloc but it will allocate memory on the given 3375 * node, which can improve the performance for cpu bound structures. 3376 * 3377 * Fallback to other node is possible if __GFP_THISNODE is not set. 3378 */ 3379 static __always_inline void * 3380 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, 3381 void *caller) 3382 { 3383 unsigned long save_flags; 3384 void *ptr; 3385 3386 if (should_failslab(cachep, flags)) 3387 return NULL; 3388 3389 cache_alloc_debugcheck_before(cachep, flags); 3390 local_irq_save(save_flags); 3391 3392 if (unlikely(nodeid == -1)) 3393 nodeid = numa_node_id(); 3394 3395 if (unlikely(!cachep->nodelists[nodeid])) { 3396 /* Node not bootstrapped yet */ 3397 ptr = fallback_alloc(cachep, flags); 3398 goto out; 3399 } 3400 3401 if (nodeid == numa_node_id()) { 3402 /* 3403 * Use the locally cached objects if possible. 3404 * However ____cache_alloc does not allow fallback 3405 * to other nodes. It may fail while we still have 3406 * objects on other nodes available. 3407 */ 3408 ptr = ____cache_alloc(cachep, flags); 3409 if (ptr) 3410 goto out; 3411 } 3412 /* ___cache_alloc_node can fall back to other nodes */ 3413 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3414 out: 3415 local_irq_restore(save_flags); 3416 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3417 3418 return ptr; 3419 } 3420 3421 static __always_inline void * 3422 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) 3423 { 3424 void *objp; 3425 3426 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { 3427 objp = alternate_node_alloc(cache, flags); 3428 if (objp) 3429 goto out; 3430 } 3431 objp = ____cache_alloc(cache, flags); 3432 3433 /* 3434 * We may just have run out of memory on the local node. 3435 * ____cache_alloc_node() knows how to locate memory on other nodes 3436 */ 3437 if (!objp) 3438 objp = ____cache_alloc_node(cache, flags, numa_node_id()); 3439 3440 out: 3441 return objp; 3442 } 3443 #else 3444 3445 static __always_inline void * 3446 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3447 { 3448 return ____cache_alloc(cachep, flags); 3449 } 3450 3451 #endif /* CONFIG_NUMA */ 3452 3453 static __always_inline void * 3454 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) 3455 { 3456 unsigned long save_flags; 3457 void *objp; 3458 3459 if (should_failslab(cachep, flags)) 3460 return NULL; 3461 3462 cache_alloc_debugcheck_before(cachep, flags); 3463 local_irq_save(save_flags); 3464 objp = __do_cache_alloc(cachep, flags); 3465 local_irq_restore(save_flags); 3466 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); 3467 prefetchw(objp); 3468 3469 return objp; 3470 } 3471 3472 /* 3473 * Caller needs to acquire correct kmem_list's list_lock 3474 */ 3475 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, 3476 int node) 3477 { 3478 int i; 3479 struct kmem_list3 *l3; 3480 3481 for (i = 0; i < nr_objects; i++) { 3482 void *objp = objpp[i]; 3483 struct slab *slabp; 3484 3485 slabp = virt_to_slab(objp); 3486 l3 = cachep->nodelists[node]; 3487 list_del(&slabp->list); 3488 check_spinlock_acquired_node(cachep, node); 3489 check_slabp(cachep, slabp); 3490 slab_put_obj(cachep, slabp, objp, node); 3491 STATS_DEC_ACTIVE(cachep); 3492 l3->free_objects++; 3493 check_slabp(cachep, slabp); 3494 3495 /* fixup slab chains */ 3496 if (slabp->inuse == 0) { 3497 if (l3->free_objects > l3->free_limit) { 3498 l3->free_objects -= cachep->num; 3499 /* No need to drop any previously held 3500 * lock here, even if we have a off-slab slab 3501 * descriptor it is guaranteed to come from 3502 * a different cache, refer to comments before 3503 * alloc_slabmgmt. 3504 */ 3505 slab_destroy(cachep, slabp); 3506 } else { 3507 list_add(&slabp->list, &l3->slabs_free); 3508 } 3509 } else { 3510 /* Unconditionally move a slab to the end of the 3511 * partial list on free - maximum time for the 3512 * other objects to be freed, too. 3513 */ 3514 list_add_tail(&slabp->list, &l3->slabs_partial); 3515 } 3516 } 3517 } 3518 3519 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3520 { 3521 int batchcount; 3522 struct kmem_list3 *l3; 3523 int node = numa_node_id(); 3524 3525 batchcount = ac->batchcount; 3526 #if DEBUG 3527 BUG_ON(!batchcount || batchcount > ac->avail); 3528 #endif 3529 check_irq_off(); 3530 l3 = cachep->nodelists[node]; 3531 spin_lock(&l3->list_lock); 3532 if (l3->shared) { 3533 struct array_cache *shared_array = l3->shared; 3534 int max = shared_array->limit - shared_array->avail; 3535 if (max) { 3536 if (batchcount > max) 3537 batchcount = max; 3538 memcpy(&(shared_array->entry[shared_array->avail]), 3539 ac->entry, sizeof(void *) * batchcount); 3540 shared_array->avail += batchcount; 3541 goto free_done; 3542 } 3543 } 3544 3545 free_block(cachep, ac->entry, batchcount, node); 3546 free_done: 3547 #if STATS 3548 { 3549 int i = 0; 3550 struct list_head *p; 3551 3552 p = l3->slabs_free.next; 3553 while (p != &(l3->slabs_free)) { 3554 struct slab *slabp; 3555 3556 slabp = list_entry(p, struct slab, list); 3557 BUG_ON(slabp->inuse); 3558 3559 i++; 3560 p = p->next; 3561 } 3562 STATS_SET_FREEABLE(cachep, i); 3563 } 3564 #endif 3565 spin_unlock(&l3->list_lock); 3566 ac->avail -= batchcount; 3567 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3568 } 3569 3570 /* 3571 * Release an obj back to its cache. If the obj has a constructed state, it must 3572 * be in this state _before_ it is released. Called with disabled ints. 3573 */ 3574 static inline void __cache_free(struct kmem_cache *cachep, void *objp) 3575 { 3576 struct array_cache *ac = cpu_cache_get(cachep); 3577 3578 check_irq_off(); 3579 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); 3580 3581 if (use_alien_caches && cache_free_alien(cachep, objp)) 3582 return; 3583 3584 if (likely(ac->avail < ac->limit)) { 3585 STATS_INC_FREEHIT(cachep); 3586 ac->entry[ac->avail++] = objp; 3587 return; 3588 } else { 3589 STATS_INC_FREEMISS(cachep); 3590 cache_flusharray(cachep, ac); 3591 ac->entry[ac->avail++] = objp; 3592 } 3593 } 3594 3595 /** 3596 * kmem_cache_alloc - Allocate an object 3597 * @cachep: The cache to allocate from. 3598 * @flags: See kmalloc(). 3599 * 3600 * Allocate an object from this cache. The flags are only relevant 3601 * if the cache has no available objects. 3602 */ 3603 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3604 { 3605 return __cache_alloc(cachep, flags, __builtin_return_address(0)); 3606 } 3607 EXPORT_SYMBOL(kmem_cache_alloc); 3608 3609 /** 3610 * kmem_cache_zalloc - Allocate an object. The memory is set to zero. 3611 * @cache: The cache to allocate from. 3612 * @flags: See kmalloc(). 3613 * 3614 * Allocate an object from this cache and set the allocated memory to zero. 3615 * The flags are only relevant if the cache has no available objects. 3616 */ 3617 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) 3618 { 3619 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); 3620 if (ret) 3621 memset(ret, 0, obj_size(cache)); 3622 return ret; 3623 } 3624 EXPORT_SYMBOL(kmem_cache_zalloc); 3625 3626 /** 3627 * kmem_ptr_validate - check if an untrusted pointer might 3628 * be a slab entry. 3629 * @cachep: the cache we're checking against 3630 * @ptr: pointer to validate 3631 * 3632 * This verifies that the untrusted pointer looks sane: 3633 * it is _not_ a guarantee that the pointer is actually 3634 * part of the slab cache in question, but it at least 3635 * validates that the pointer can be dereferenced and 3636 * looks half-way sane. 3637 * 3638 * Currently only used for dentry validation. 3639 */ 3640 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) 3641 { 3642 unsigned long addr = (unsigned long)ptr; 3643 unsigned long min_addr = PAGE_OFFSET; 3644 unsigned long align_mask = BYTES_PER_WORD - 1; 3645 unsigned long size = cachep->buffer_size; 3646 struct page *page; 3647 3648 if (unlikely(addr < min_addr)) 3649 goto out; 3650 if (unlikely(addr > (unsigned long)high_memory - size)) 3651 goto out; 3652 if (unlikely(addr & align_mask)) 3653 goto out; 3654 if (unlikely(!kern_addr_valid(addr))) 3655 goto out; 3656 if (unlikely(!kern_addr_valid(addr + size - 1))) 3657 goto out; 3658 page = virt_to_page(ptr); 3659 if (unlikely(!PageSlab(page))) 3660 goto out; 3661 if (unlikely(page_get_cache(page) != cachep)) 3662 goto out; 3663 return 1; 3664 out: 3665 return 0; 3666 } 3667 3668 #ifdef CONFIG_NUMA 3669 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3670 { 3671 return __cache_alloc_node(cachep, flags, nodeid, 3672 __builtin_return_address(0)); 3673 } 3674 EXPORT_SYMBOL(kmem_cache_alloc_node); 3675 3676 static __always_inline void * 3677 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) 3678 { 3679 struct kmem_cache *cachep; 3680 3681 cachep = kmem_find_general_cachep(size, flags); 3682 if (unlikely(cachep == NULL)) 3683 return NULL; 3684 return kmem_cache_alloc_node(cachep, flags, node); 3685 } 3686 3687 #ifdef CONFIG_DEBUG_SLAB 3688 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3689 { 3690 return __do_kmalloc_node(size, flags, node, 3691 __builtin_return_address(0)); 3692 } 3693 EXPORT_SYMBOL(__kmalloc_node); 3694 3695 void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3696 int node, void *caller) 3697 { 3698 return __do_kmalloc_node(size, flags, node, caller); 3699 } 3700 EXPORT_SYMBOL(__kmalloc_node_track_caller); 3701 #else 3702 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3703 { 3704 return __do_kmalloc_node(size, flags, node, NULL); 3705 } 3706 EXPORT_SYMBOL(__kmalloc_node); 3707 #endif /* CONFIG_DEBUG_SLAB */ 3708 #endif /* CONFIG_NUMA */ 3709 3710 /** 3711 * __do_kmalloc - allocate memory 3712 * @size: how many bytes of memory are required. 3713 * @flags: the type of memory to allocate (see kmalloc). 3714 * @caller: function caller for debug tracking of the caller 3715 */ 3716 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3717 void *caller) 3718 { 3719 struct kmem_cache *cachep; 3720 3721 /* If you want to save a few bytes .text space: replace 3722 * __ with kmem_. 3723 * Then kmalloc uses the uninlined functions instead of the inline 3724 * functions. 3725 */ 3726 cachep = __find_general_cachep(size, flags); 3727 if (unlikely(cachep == NULL)) 3728 return NULL; 3729 return __cache_alloc(cachep, flags, caller); 3730 } 3731 3732 3733 #ifdef CONFIG_DEBUG_SLAB 3734 void *__kmalloc(size_t size, gfp_t flags) 3735 { 3736 return __do_kmalloc(size, flags, __builtin_return_address(0)); 3737 } 3738 EXPORT_SYMBOL(__kmalloc); 3739 3740 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) 3741 { 3742 return __do_kmalloc(size, flags, caller); 3743 } 3744 EXPORT_SYMBOL(__kmalloc_track_caller); 3745 3746 #else 3747 void *__kmalloc(size_t size, gfp_t flags) 3748 { 3749 return __do_kmalloc(size, flags, NULL); 3750 } 3751 EXPORT_SYMBOL(__kmalloc); 3752 #endif 3753 3754 /** 3755 * krealloc - reallocate memory. The contents will remain unchanged. 3756 * @p: object to reallocate memory for. 3757 * @new_size: how many bytes of memory are required. 3758 * @flags: the type of memory to allocate. 3759 * 3760 * The contents of the object pointed to are preserved up to the 3761 * lesser of the new and old sizes. If @p is %NULL, krealloc() 3762 * behaves exactly like kmalloc(). If @size is 0 and @p is not a 3763 * %NULL pointer, the object pointed to is freed. 3764 */ 3765 void *krealloc(const void *p, size_t new_size, gfp_t flags) 3766 { 3767 struct kmem_cache *cache, *new_cache; 3768 void *ret; 3769 3770 if (unlikely(!p)) 3771 return kmalloc_track_caller(new_size, flags); 3772 3773 if (unlikely(!new_size)) { 3774 kfree(p); 3775 return NULL; 3776 } 3777 3778 cache = virt_to_cache(p); 3779 new_cache = __find_general_cachep(new_size, flags); 3780 3781 /* 3782 * If new size fits in the current cache, bail out. 3783 */ 3784 if (likely(cache == new_cache)) 3785 return (void *)p; 3786 3787 /* 3788 * We are on the slow-path here so do not use __cache_alloc 3789 * because it bloats kernel text. 3790 */ 3791 ret = kmalloc_track_caller(new_size, flags); 3792 if (ret) { 3793 memcpy(ret, p, min(new_size, ksize(p))); 3794 kfree(p); 3795 } 3796 return ret; 3797 } 3798 EXPORT_SYMBOL(krealloc); 3799 3800 /** 3801 * kmem_cache_free - Deallocate an object 3802 * @cachep: The cache the allocation was from. 3803 * @objp: The previously allocated object. 3804 * 3805 * Free an object which was previously allocated from this 3806 * cache. 3807 */ 3808 void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3809 { 3810 unsigned long flags; 3811 3812 BUG_ON(virt_to_cache(objp) != cachep); 3813 3814 local_irq_save(flags); 3815 debug_check_no_locks_freed(objp, obj_size(cachep)); 3816 __cache_free(cachep, objp); 3817 local_irq_restore(flags); 3818 } 3819 EXPORT_SYMBOL(kmem_cache_free); 3820 3821 /** 3822 * kfree - free previously allocated memory 3823 * @objp: pointer returned by kmalloc. 3824 * 3825 * If @objp is NULL, no operation is performed. 3826 * 3827 * Don't free memory not originally allocated by kmalloc() 3828 * or you will run into trouble. 3829 */ 3830 void kfree(const void *objp) 3831 { 3832 struct kmem_cache *c; 3833 unsigned long flags; 3834 3835 if (unlikely(!objp)) 3836 return; 3837 local_irq_save(flags); 3838 kfree_debugcheck(objp); 3839 c = virt_to_cache(objp); 3840 debug_check_no_locks_freed(objp, obj_size(c)); 3841 __cache_free(c, (void *)objp); 3842 local_irq_restore(flags); 3843 } 3844 EXPORT_SYMBOL(kfree); 3845 3846 unsigned int kmem_cache_size(struct kmem_cache *cachep) 3847 { 3848 return obj_size(cachep); 3849 } 3850 EXPORT_SYMBOL(kmem_cache_size); 3851 3852 const char *kmem_cache_name(struct kmem_cache *cachep) 3853 { 3854 return cachep->name; 3855 } 3856 EXPORT_SYMBOL_GPL(kmem_cache_name); 3857 3858 /* 3859 * This initializes kmem_list3 or resizes varioius caches for all nodes. 3860 */ 3861 static int alloc_kmemlist(struct kmem_cache *cachep) 3862 { 3863 int node; 3864 struct kmem_list3 *l3; 3865 struct array_cache *new_shared; 3866 struct array_cache **new_alien = NULL; 3867 3868 for_each_online_node(node) { 3869 3870 if (use_alien_caches) { 3871 new_alien = alloc_alien_cache(node, cachep->limit); 3872 if (!new_alien) 3873 goto fail; 3874 } 3875 3876 new_shared = NULL; 3877 if (cachep->shared) { 3878 new_shared = alloc_arraycache(node, 3879 cachep->shared*cachep->batchcount, 3880 0xbaadf00d); 3881 if (!new_shared) { 3882 free_alien_cache(new_alien); 3883 goto fail; 3884 } 3885 } 3886 3887 l3 = cachep->nodelists[node]; 3888 if (l3) { 3889 struct array_cache *shared = l3->shared; 3890 3891 spin_lock_irq(&l3->list_lock); 3892 3893 if (shared) 3894 free_block(cachep, shared->entry, 3895 shared->avail, node); 3896 3897 l3->shared = new_shared; 3898 if (!l3->alien) { 3899 l3->alien = new_alien; 3900 new_alien = NULL; 3901 } 3902 l3->free_limit = (1 + nr_cpus_node(node)) * 3903 cachep->batchcount + cachep->num; 3904 spin_unlock_irq(&l3->list_lock); 3905 kfree(shared); 3906 free_alien_cache(new_alien); 3907 continue; 3908 } 3909 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); 3910 if (!l3) { 3911 free_alien_cache(new_alien); 3912 kfree(new_shared); 3913 goto fail; 3914 } 3915 3916 kmem_list3_init(l3); 3917 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + 3918 ((unsigned long)cachep) % REAPTIMEOUT_LIST3; 3919 l3->shared = new_shared; 3920 l3->alien = new_alien; 3921 l3->free_limit = (1 + nr_cpus_node(node)) * 3922 cachep->batchcount + cachep->num; 3923 cachep->nodelists[node] = l3; 3924 } 3925 return 0; 3926 3927 fail: 3928 if (!cachep->next.next) { 3929 /* Cache is not active yet. Roll back what we did */ 3930 node--; 3931 while (node >= 0) { 3932 if (cachep->nodelists[node]) { 3933 l3 = cachep->nodelists[node]; 3934 3935 kfree(l3->shared); 3936 free_alien_cache(l3->alien); 3937 kfree(l3); 3938 cachep->nodelists[node] = NULL; 3939 } 3940 node--; 3941 } 3942 } 3943 return -ENOMEM; 3944 } 3945 3946 struct ccupdate_struct { 3947 struct kmem_cache *cachep; 3948 struct array_cache *new[NR_CPUS]; 3949 }; 3950 3951 static void do_ccupdate_local(void *info) 3952 { 3953 struct ccupdate_struct *new = info; 3954 struct array_cache *old; 3955 3956 check_irq_off(); 3957 old = cpu_cache_get(new->cachep); 3958 3959 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; 3960 new->new[smp_processor_id()] = old; 3961 } 3962 3963 /* Always called with the cache_chain_mutex held */ 3964 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3965 int batchcount, int shared) 3966 { 3967 struct ccupdate_struct *new; 3968 int i; 3969 3970 new = kzalloc(sizeof(*new), GFP_KERNEL); 3971 if (!new) 3972 return -ENOMEM; 3973 3974 for_each_online_cpu(i) { 3975 new->new[i] = alloc_arraycache(cpu_to_node(i), limit, 3976 batchcount); 3977 if (!new->new[i]) { 3978 for (i--; i >= 0; i--) 3979 kfree(new->new[i]); 3980 kfree(new); 3981 return -ENOMEM; 3982 } 3983 } 3984 new->cachep = cachep; 3985 3986 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); 3987 3988 check_irq_on(); 3989 cachep->batchcount = batchcount; 3990 cachep->limit = limit; 3991 cachep->shared = shared; 3992 3993 for_each_online_cpu(i) { 3994 struct array_cache *ccold = new->new[i]; 3995 if (!ccold) 3996 continue; 3997 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 3998 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); 3999 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); 4000 kfree(ccold); 4001 } 4002 kfree(new); 4003 return alloc_kmemlist(cachep); 4004 } 4005 4006 /* Called with cache_chain_mutex held always */ 4007 static int enable_cpucache(struct kmem_cache *cachep) 4008 { 4009 int err; 4010 int limit, shared; 4011 4012 /* 4013 * The head array serves three purposes: 4014 * - create a LIFO ordering, i.e. return objects that are cache-warm 4015 * - reduce the number of spinlock operations. 4016 * - reduce the number of linked list operations on the slab and 4017 * bufctl chains: array operations are cheaper. 4018 * The numbers are guessed, we should auto-tune as described by 4019 * Bonwick. 4020 */ 4021 if (cachep->buffer_size > 131072) 4022 limit = 1; 4023 else if (cachep->buffer_size > PAGE_SIZE) 4024 limit = 8; 4025 else if (cachep->buffer_size > 1024) 4026 limit = 24; 4027 else if (cachep->buffer_size > 256) 4028 limit = 54; 4029 else 4030 limit = 120; 4031 4032 /* 4033 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 4034 * allocation behaviour: Most allocs on one cpu, most free operations 4035 * on another cpu. For these cases, an efficient object passing between 4036 * cpus is necessary. This is provided by a shared array. The array 4037 * replaces Bonwick's magazine layer. 4038 * On uniprocessor, it's functionally equivalent (but less efficient) 4039 * to a larger limit. Thus disabled by default. 4040 */ 4041 shared = 0; 4042 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) 4043 shared = 8; 4044 4045 #if DEBUG 4046 /* 4047 * With debugging enabled, large batchcount lead to excessively long 4048 * periods with disabled local interrupts. Limit the batchcount 4049 */ 4050 if (limit > 32) 4051 limit = 32; 4052 #endif 4053 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); 4054 if (err) 4055 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", 4056 cachep->name, -err); 4057 return err; 4058 } 4059 4060 /* 4061 * Drain an array if it contains any elements taking the l3 lock only if 4062 * necessary. Note that the l3 listlock also protects the array_cache 4063 * if drain_array() is used on the shared array. 4064 */ 4065 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, 4066 struct array_cache *ac, int force, int node) 4067 { 4068 int tofree; 4069 4070 if (!ac || !ac->avail) 4071 return; 4072 if (ac->touched && !force) { 4073 ac->touched = 0; 4074 } else { 4075 spin_lock_irq(&l3->list_lock); 4076 if (ac->avail) { 4077 tofree = force ? ac->avail : (ac->limit + 4) / 5; 4078 if (tofree > ac->avail) 4079 tofree = (ac->avail + 1) / 2; 4080 free_block(cachep, ac->entry, tofree, node); 4081 ac->avail -= tofree; 4082 memmove(ac->entry, &(ac->entry[tofree]), 4083 sizeof(void *) * ac->avail); 4084 } 4085 spin_unlock_irq(&l3->list_lock); 4086 } 4087 } 4088 4089 /** 4090 * cache_reap - Reclaim memory from caches. 4091 * @w: work descriptor 4092 * 4093 * Called from workqueue/eventd every few seconds. 4094 * Purpose: 4095 * - clear the per-cpu caches for this CPU. 4096 * - return freeable pages to the main free memory pool. 4097 * 4098 * If we cannot acquire the cache chain mutex then just give up - we'll try 4099 * again on the next iteration. 4100 */ 4101 static void cache_reap(struct work_struct *w) 4102 { 4103 struct kmem_cache *searchp; 4104 struct kmem_list3 *l3; 4105 int node = numa_node_id(); 4106 struct delayed_work *work = 4107 container_of(w, struct delayed_work, work); 4108 4109 if (!mutex_trylock(&cache_chain_mutex)) 4110 /* Give up. Setup the next iteration. */ 4111 goto out; 4112 4113 list_for_each_entry(searchp, &cache_chain, next) { 4114 check_irq_on(); 4115 4116 /* 4117 * We only take the l3 lock if absolutely necessary and we 4118 * have established with reasonable certainty that 4119 * we can do some work if the lock was obtained. 4120 */ 4121 l3 = searchp->nodelists[node]; 4122 4123 reap_alien(searchp, l3); 4124 4125 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); 4126 4127 /* 4128 * These are racy checks but it does not matter 4129 * if we skip one check or scan twice. 4130 */ 4131 if (time_after(l3->next_reap, jiffies)) 4132 goto next; 4133 4134 l3->next_reap = jiffies + REAPTIMEOUT_LIST3; 4135 4136 drain_array(searchp, l3, l3->shared, 0, node); 4137 4138 if (l3->free_touched) 4139 l3->free_touched = 0; 4140 else { 4141 int freed; 4142 4143 freed = drain_freelist(searchp, l3, (l3->free_limit + 4144 5 * searchp->num - 1) / (5 * searchp->num)); 4145 STATS_ADD_REAPED(searchp, freed); 4146 } 4147 next: 4148 cond_resched(); 4149 } 4150 check_irq_on(); 4151 mutex_unlock(&cache_chain_mutex); 4152 next_reap_node(); 4153 out: 4154 /* Set up the next iteration */ 4155 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); 4156 } 4157 4158 #ifdef CONFIG_PROC_FS 4159 4160 static void print_slabinfo_header(struct seq_file *m) 4161 { 4162 /* 4163 * Output format version, so at least we can change it 4164 * without _too_ many complaints. 4165 */ 4166 #if STATS 4167 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); 4168 #else 4169 seq_puts(m, "slabinfo - version: 2.1\n"); 4170 #endif 4171 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4172 "<objperslab> <pagesperslab>"); 4173 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4174 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4175 #if STATS 4176 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " 4177 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); 4178 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); 4179 #endif 4180 seq_putc(m, '\n'); 4181 } 4182 4183 static void *s_start(struct seq_file *m, loff_t *pos) 4184 { 4185 loff_t n = *pos; 4186 struct list_head *p; 4187 4188 mutex_lock(&cache_chain_mutex); 4189 if (!n) 4190 print_slabinfo_header(m); 4191 p = cache_chain.next; 4192 while (n--) { 4193 p = p->next; 4194 if (p == &cache_chain) 4195 return NULL; 4196 } 4197 return list_entry(p, struct kmem_cache, next); 4198 } 4199 4200 static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4201 { 4202 struct kmem_cache *cachep = p; 4203 ++*pos; 4204 return cachep->next.next == &cache_chain ? 4205 NULL : list_entry(cachep->next.next, struct kmem_cache, next); 4206 } 4207 4208 static void s_stop(struct seq_file *m, void *p) 4209 { 4210 mutex_unlock(&cache_chain_mutex); 4211 } 4212 4213 static int s_show(struct seq_file *m, void *p) 4214 { 4215 struct kmem_cache *cachep = p; 4216 struct slab *slabp; 4217 unsigned long active_objs; 4218 unsigned long num_objs; 4219 unsigned long active_slabs = 0; 4220 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 4221 const char *name; 4222 char *error = NULL; 4223 int node; 4224 struct kmem_list3 *l3; 4225 4226 active_objs = 0; 4227 num_slabs = 0; 4228 for_each_online_node(node) { 4229 l3 = cachep->nodelists[node]; 4230 if (!l3) 4231 continue; 4232 4233 check_irq_on(); 4234 spin_lock_irq(&l3->list_lock); 4235 4236 list_for_each_entry(slabp, &l3->slabs_full, list) { 4237 if (slabp->inuse != cachep->num && !error) 4238 error = "slabs_full accounting error"; 4239 active_objs += cachep->num; 4240 active_slabs++; 4241 } 4242 list_for_each_entry(slabp, &l3->slabs_partial, list) { 4243 if (slabp->inuse == cachep->num && !error) 4244 error = "slabs_partial inuse accounting error"; 4245 if (!slabp->inuse && !error) 4246 error = "slabs_partial/inuse accounting error"; 4247 active_objs += slabp->inuse; 4248 active_slabs++; 4249 } 4250 list_for_each_entry(slabp, &l3->slabs_free, list) { 4251 if (slabp->inuse && !error) 4252 error = "slabs_free/inuse accounting error"; 4253 num_slabs++; 4254 } 4255 free_objects += l3->free_objects; 4256 if (l3->shared) 4257 shared_avail += l3->shared->avail; 4258 4259 spin_unlock_irq(&l3->list_lock); 4260 } 4261 num_slabs += active_slabs; 4262 num_objs = num_slabs * cachep->num; 4263 if (num_objs - active_objs != free_objects && !error) 4264 error = "free_objects accounting error"; 4265 4266 name = cachep->name; 4267 if (error) 4268 printk(KERN_ERR "slab: cache %s error: %s\n", name, error); 4269 4270 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", 4271 name, active_objs, num_objs, cachep->buffer_size, 4272 cachep->num, (1 << cachep->gfporder)); 4273 seq_printf(m, " : tunables %4u %4u %4u", 4274 cachep->limit, cachep->batchcount, cachep->shared); 4275 seq_printf(m, " : slabdata %6lu %6lu %6lu", 4276 active_slabs, num_slabs, shared_avail); 4277 #if STATS 4278 { /* list3 stats */ 4279 unsigned long high = cachep->high_mark; 4280 unsigned long allocs = cachep->num_allocations; 4281 unsigned long grown = cachep->grown; 4282 unsigned long reaped = cachep->reaped; 4283 unsigned long errors = cachep->errors; 4284 unsigned long max_freeable = cachep->max_freeable; 4285 unsigned long node_allocs = cachep->node_allocs; 4286 unsigned long node_frees = cachep->node_frees; 4287 unsigned long overflows = cachep->node_overflow; 4288 4289 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ 4290 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, 4291 reaped, errors, max_freeable, node_allocs, 4292 node_frees, overflows); 4293 } 4294 /* cpu stats */ 4295 { 4296 unsigned long allochit = atomic_read(&cachep->allochit); 4297 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4298 unsigned long freehit = atomic_read(&cachep->freehit); 4299 unsigned long freemiss = atomic_read(&cachep->freemiss); 4300 4301 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4302 allochit, allocmiss, freehit, freemiss); 4303 } 4304 #endif 4305 seq_putc(m, '\n'); 4306 return 0; 4307 } 4308 4309 /* 4310 * slabinfo_op - iterator that generates /proc/slabinfo 4311 * 4312 * Output layout: 4313 * cache-name 4314 * num-active-objs 4315 * total-objs 4316 * object size 4317 * num-active-slabs 4318 * total-slabs 4319 * num-pages-per-slab 4320 * + further values on SMP and with statistics enabled 4321 */ 4322 4323 const struct seq_operations slabinfo_op = { 4324 .start = s_start, 4325 .next = s_next, 4326 .stop = s_stop, 4327 .show = s_show, 4328 }; 4329 4330 #define MAX_SLABINFO_WRITE 128 4331 /** 4332 * slabinfo_write - Tuning for the slab allocator 4333 * @file: unused 4334 * @buffer: user buffer 4335 * @count: data length 4336 * @ppos: unused 4337 */ 4338 ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4339 size_t count, loff_t *ppos) 4340 { 4341 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4342 int limit, batchcount, shared, res; 4343 struct kmem_cache *cachep; 4344 4345 if (count > MAX_SLABINFO_WRITE) 4346 return -EINVAL; 4347 if (copy_from_user(&kbuf, buffer, count)) 4348 return -EFAULT; 4349 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4350 4351 tmp = strchr(kbuf, ' '); 4352 if (!tmp) 4353 return -EINVAL; 4354 *tmp = '\0'; 4355 tmp++; 4356 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4357 return -EINVAL; 4358 4359 /* Find the cache in the chain of caches. */ 4360 mutex_lock(&cache_chain_mutex); 4361 res = -EINVAL; 4362 list_for_each_entry(cachep, &cache_chain, next) { 4363 if (!strcmp(cachep->name, kbuf)) { 4364 if (limit < 1 || batchcount < 1 || 4365 batchcount > limit || shared < 0) { 4366 res = 0; 4367 } else { 4368 res = do_tune_cpucache(cachep, limit, 4369 batchcount, shared); 4370 } 4371 break; 4372 } 4373 } 4374 mutex_unlock(&cache_chain_mutex); 4375 if (res >= 0) 4376 res = count; 4377 return res; 4378 } 4379 4380 #ifdef CONFIG_DEBUG_SLAB_LEAK 4381 4382 static void *leaks_start(struct seq_file *m, loff_t *pos) 4383 { 4384 loff_t n = *pos; 4385 struct list_head *p; 4386 4387 mutex_lock(&cache_chain_mutex); 4388 p = cache_chain.next; 4389 while (n--) { 4390 p = p->next; 4391 if (p == &cache_chain) 4392 return NULL; 4393 } 4394 return list_entry(p, struct kmem_cache, next); 4395 } 4396 4397 static inline int add_caller(unsigned long *n, unsigned long v) 4398 { 4399 unsigned long *p; 4400 int l; 4401 if (!v) 4402 return 1; 4403 l = n[1]; 4404 p = n + 2; 4405 while (l) { 4406 int i = l/2; 4407 unsigned long *q = p + 2 * i; 4408 if (*q == v) { 4409 q[1]++; 4410 return 1; 4411 } 4412 if (*q > v) { 4413 l = i; 4414 } else { 4415 p = q + 2; 4416 l -= i + 1; 4417 } 4418 } 4419 if (++n[1] == n[0]) 4420 return 0; 4421 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4422 p[0] = v; 4423 p[1] = 1; 4424 return 1; 4425 } 4426 4427 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) 4428 { 4429 void *p; 4430 int i; 4431 if (n[0] == n[1]) 4432 return; 4433 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { 4434 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) 4435 continue; 4436 if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) 4437 return; 4438 } 4439 } 4440 4441 static void show_symbol(struct seq_file *m, unsigned long address) 4442 { 4443 #ifdef CONFIG_KALLSYMS 4444 unsigned long offset, size; 4445 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1]; 4446 4447 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { 4448 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4449 if (modname[0]) 4450 seq_printf(m, " [%s]", modname); 4451 return; 4452 } 4453 #endif 4454 seq_printf(m, "%p", (void *)address); 4455 } 4456 4457 static int leaks_show(struct seq_file *m, void *p) 4458 { 4459 struct kmem_cache *cachep = p; 4460 struct slab *slabp; 4461 struct kmem_list3 *l3; 4462 const char *name; 4463 unsigned long *n = m->private; 4464 int node; 4465 int i; 4466 4467 if (!(cachep->flags & SLAB_STORE_USER)) 4468 return 0; 4469 if (!(cachep->flags & SLAB_RED_ZONE)) 4470 return 0; 4471 4472 /* OK, we can do it */ 4473 4474 n[1] = 0; 4475 4476 for_each_online_node(node) { 4477 l3 = cachep->nodelists[node]; 4478 if (!l3) 4479 continue; 4480 4481 check_irq_on(); 4482 spin_lock_irq(&l3->list_lock); 4483 4484 list_for_each_entry(slabp, &l3->slabs_full, list) 4485 handle_slab(n, cachep, slabp); 4486 list_for_each_entry(slabp, &l3->slabs_partial, list) 4487 handle_slab(n, cachep, slabp); 4488 spin_unlock_irq(&l3->list_lock); 4489 } 4490 name = cachep->name; 4491 if (n[0] == n[1]) { 4492 /* Increase the buffer size */ 4493 mutex_unlock(&cache_chain_mutex); 4494 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4495 if (!m->private) { 4496 /* Too bad, we are really out */ 4497 m->private = n; 4498 mutex_lock(&cache_chain_mutex); 4499 return -ENOMEM; 4500 } 4501 *(unsigned long *)m->private = n[0] * 2; 4502 kfree(n); 4503 mutex_lock(&cache_chain_mutex); 4504 /* Now make sure this entry will be retried */ 4505 m->count = m->size; 4506 return 0; 4507 } 4508 for (i = 0; i < n[1]; i++) { 4509 seq_printf(m, "%s: %lu ", name, n[2*i+3]); 4510 show_symbol(m, n[2*i+2]); 4511 seq_putc(m, '\n'); 4512 } 4513 4514 return 0; 4515 } 4516 4517 const struct seq_operations slabstats_op = { 4518 .start = leaks_start, 4519 .next = s_next, 4520 .stop = s_stop, 4521 .show = leaks_show, 4522 }; 4523 #endif 4524 #endif 4525 4526 /** 4527 * ksize - get the actual amount of memory allocated for a given object 4528 * @objp: Pointer to the object 4529 * 4530 * kmalloc may internally round up allocations and return more memory 4531 * than requested. ksize() can be used to determine the actual amount of 4532 * memory allocated. The caller may use this additional memory, even though 4533 * a smaller amount of memory was initially specified with the kmalloc call. 4534 * The caller must guarantee that objp points to a valid object previously 4535 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4536 * must not be freed during the duration of the call. 4537 */ 4538 size_t ksize(const void *objp) 4539 { 4540 if (unlikely(objp == NULL)) 4541 return 0; 4542 4543 return obj_size(virt_to_cache(objp)); 4544 } 4545