1 /* 2 * linux/mm/slab.c 3 * Written by Mark Hemment, 1996/97. 4 * (markhe@nextd.demon.co.uk) 5 * 6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 7 * 8 * Major cleanup, different bufctl logic, per-cpu arrays 9 * (c) 2000 Manfred Spraul 10 * 11 * Cleanup, make the head arrays unconditional, preparation for NUMA 12 * (c) 2002 Manfred Spraul 13 * 14 * An implementation of the Slab Allocator as described in outline in; 15 * UNIX Internals: The New Frontiers by Uresh Vahalia 16 * Pub: Prentice Hall ISBN 0-13-101908-2 17 * or with a little more detail in; 18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 19 * Jeff Bonwick (Sun Microsystems). 20 * Presented at: USENIX Summer 1994 Technical Conference 21 * 22 * The memory is organized in caches, one cache for each object type. 23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 24 * Each cache consists out of many slabs (they are small (usually one 25 * page long) and always contiguous), and each slab contains multiple 26 * initialized objects. 27 * 28 * This means, that your constructor is used only for newly allocated 29 * slabs and you must pass objects with the same initializations to 30 * kmem_cache_free. 31 * 32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 33 * normal). If you need a special memory type, then must create a new 34 * cache for that memory type. 35 * 36 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 37 * full slabs with 0 free objects 38 * partial slabs 39 * empty slabs with no allocated objects 40 * 41 * If partial slabs exist, then new allocations come from these slabs, 42 * otherwise from empty slabs or new slabs are allocated. 43 * 44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 46 * 47 * Each cache has a short per-cpu head array, most allocs 48 * and frees go into that array, and if that array overflows, then 1/2 49 * of the entries in the array are given back into the global cache. 50 * The head array is strictly LIFO and should improve the cache hit rates. 51 * On SMP, it additionally reduces the spinlock operations. 52 * 53 * The c_cpuarray may not be read with enabled local interrupts - 54 * it's changed with a smp_call_function(). 55 * 56 * SMP synchronization: 57 * constructors and destructors are called without any locking. 58 * Several members in struct kmem_cache and struct slab never change, they 59 * are accessed without any locking. 60 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 61 * and local interrupts are disabled so slab code is preempt-safe. 62 * The non-constant members are protected with a per-cache irq spinlock. 63 * 64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 65 * in 2000 - many ideas in the current implementation are derived from 66 * his patch. 67 * 68 * Further notes from the original documentation: 69 * 70 * 11 April '97. Started multi-threading - markhe 71 * The global cache-chain is protected by the mutex 'slab_mutex'. 72 * The sem is only needed when accessing/extending the cache-chain, which 73 * can never happen inside an interrupt (kmem_cache_create(), 74 * kmem_cache_shrink() and kmem_cache_reap()). 75 * 76 * At present, each engine can be growing a cache. This should be blocked. 77 * 78 * 15 March 2005. NUMA slab allocator. 79 * Shai Fultheim <shai@scalex86.org>. 80 * Shobhit Dayal <shobhit@calsoftinc.com> 81 * Alok N Kataria <alokk@calsoftinc.com> 82 * Christoph Lameter <christoph@lameter.com> 83 * 84 * Modified the slab allocator to be node aware on NUMA systems. 85 * Each node has its own list of partial, free and full slabs. 86 * All object allocations for a node occur from node specific slab lists. 87 */ 88 89 #include <linux/slab.h> 90 #include <linux/mm.h> 91 #include <linux/poison.h> 92 #include <linux/swap.h> 93 #include <linux/cache.h> 94 #include <linux/interrupt.h> 95 #include <linux/init.h> 96 #include <linux/compiler.h> 97 #include <linux/cpuset.h> 98 #include <linux/proc_fs.h> 99 #include <linux/seq_file.h> 100 #include <linux/notifier.h> 101 #include <linux/kallsyms.h> 102 #include <linux/cpu.h> 103 #include <linux/sysctl.h> 104 #include <linux/module.h> 105 #include <linux/rcupdate.h> 106 #include <linux/string.h> 107 #include <linux/uaccess.h> 108 #include <linux/nodemask.h> 109 #include <linux/kmemleak.h> 110 #include <linux/mempolicy.h> 111 #include <linux/mutex.h> 112 #include <linux/fault-inject.h> 113 #include <linux/rtmutex.h> 114 #include <linux/reciprocal_div.h> 115 #include <linux/debugobjects.h> 116 #include <linux/kmemcheck.h> 117 #include <linux/memory.h> 118 #include <linux/prefetch.h> 119 120 #include <net/sock.h> 121 122 #include <asm/cacheflush.h> 123 #include <asm/tlbflush.h> 124 #include <asm/page.h> 125 126 #include <trace/events/kmem.h> 127 128 #include "internal.h" 129 130 #include "slab.h" 131 132 /* 133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. 134 * 0 for faster, smaller code (especially in the critical paths). 135 * 136 * STATS - 1 to collect stats for /proc/slabinfo. 137 * 0 for faster, smaller code (especially in the critical paths). 138 * 139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 140 */ 141 142 #ifdef CONFIG_DEBUG_SLAB 143 #define DEBUG 1 144 #define STATS 1 145 #define FORCED_DEBUG 1 146 #else 147 #define DEBUG 0 148 #define STATS 0 149 #define FORCED_DEBUG 0 150 #endif 151 152 /* Shouldn't this be in a header file somewhere? */ 153 #define BYTES_PER_WORD sizeof(void *) 154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) 155 156 #ifndef ARCH_KMALLOC_FLAGS 157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 158 #endif 159 160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \ 161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0) 162 163 #if FREELIST_BYTE_INDEX 164 typedef unsigned char freelist_idx_t; 165 #else 166 typedef unsigned short freelist_idx_t; 167 #endif 168 169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1) 170 171 /* 172 * struct array_cache 173 * 174 * Purpose: 175 * - LIFO ordering, to hand out cache-warm objects from _alloc 176 * - reduce the number of linked list operations 177 * - reduce spinlock operations 178 * 179 * The limit is stored in the per-cpu structure to reduce the data cache 180 * footprint. 181 * 182 */ 183 struct array_cache { 184 unsigned int avail; 185 unsigned int limit; 186 unsigned int batchcount; 187 unsigned int touched; 188 void *entry[]; /* 189 * Must have this definition in here for the proper 190 * alignment of array_cache. Also simplifies accessing 191 * the entries. 192 */ 193 }; 194 195 struct alien_cache { 196 spinlock_t lock; 197 struct array_cache ac; 198 }; 199 200 /* 201 * Need this for bootstrapping a per node allocator. 202 */ 203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES) 204 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS]; 205 #define CACHE_CACHE 0 206 #define SIZE_NODE (MAX_NUMNODES) 207 208 static int drain_freelist(struct kmem_cache *cache, 209 struct kmem_cache_node *n, int tofree); 210 static void free_block(struct kmem_cache *cachep, void **objpp, int len, 211 int node, struct list_head *list); 212 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list); 213 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); 214 static void cache_reap(struct work_struct *unused); 215 216 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep, 217 void **list); 218 static inline void fixup_slab_list(struct kmem_cache *cachep, 219 struct kmem_cache_node *n, struct page *page, 220 void **list); 221 static int slab_early_init = 1; 222 223 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node)) 224 225 static void kmem_cache_node_init(struct kmem_cache_node *parent) 226 { 227 INIT_LIST_HEAD(&parent->slabs_full); 228 INIT_LIST_HEAD(&parent->slabs_partial); 229 INIT_LIST_HEAD(&parent->slabs_free); 230 parent->shared = NULL; 231 parent->alien = NULL; 232 parent->colour_next = 0; 233 spin_lock_init(&parent->list_lock); 234 parent->free_objects = 0; 235 parent->free_touched = 0; 236 } 237 238 #define MAKE_LIST(cachep, listp, slab, nodeid) \ 239 do { \ 240 INIT_LIST_HEAD(listp); \ 241 list_splice(&get_node(cachep, nodeid)->slab, listp); \ 242 } while (0) 243 244 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 245 do { \ 246 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 247 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 248 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 249 } while (0) 250 251 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL) 252 #define CFLGS_OFF_SLAB (0x80000000UL) 253 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB) 254 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 255 256 #define BATCHREFILL_LIMIT 16 257 /* 258 * Optimization question: fewer reaps means less probability for unnessary 259 * cpucache drain/refill cycles. 260 * 261 * OTOH the cpuarrays can contain lots of objects, 262 * which could lock up otherwise freeable slabs. 263 */ 264 #define REAPTIMEOUT_AC (2*HZ) 265 #define REAPTIMEOUT_NODE (4*HZ) 266 267 #if STATS 268 #define STATS_INC_ACTIVE(x) ((x)->num_active++) 269 #define STATS_DEC_ACTIVE(x) ((x)->num_active--) 270 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 271 #define STATS_INC_GROWN(x) ((x)->grown++) 272 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 273 #define STATS_SET_HIGH(x) \ 274 do { \ 275 if ((x)->num_active > (x)->high_mark) \ 276 (x)->high_mark = (x)->num_active; \ 277 } while (0) 278 #define STATS_INC_ERR(x) ((x)->errors++) 279 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 280 #define STATS_INC_NODEFREES(x) ((x)->node_frees++) 281 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 282 #define STATS_SET_FREEABLE(x, i) \ 283 do { \ 284 if ((x)->max_freeable < i) \ 285 (x)->max_freeable = i; \ 286 } while (0) 287 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 288 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 289 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 290 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 291 #else 292 #define STATS_INC_ACTIVE(x) do { } while (0) 293 #define STATS_DEC_ACTIVE(x) do { } while (0) 294 #define STATS_INC_ALLOCED(x) do { } while (0) 295 #define STATS_INC_GROWN(x) do { } while (0) 296 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) 297 #define STATS_SET_HIGH(x) do { } while (0) 298 #define STATS_INC_ERR(x) do { } while (0) 299 #define STATS_INC_NODEALLOCS(x) do { } while (0) 300 #define STATS_INC_NODEFREES(x) do { } while (0) 301 #define STATS_INC_ACOVERFLOW(x) do { } while (0) 302 #define STATS_SET_FREEABLE(x, i) do { } while (0) 303 #define STATS_INC_ALLOCHIT(x) do { } while (0) 304 #define STATS_INC_ALLOCMISS(x) do { } while (0) 305 #define STATS_INC_FREEHIT(x) do { } while (0) 306 #define STATS_INC_FREEMISS(x) do { } while (0) 307 #endif 308 309 #if DEBUG 310 311 /* 312 * memory layout of objects: 313 * 0 : objp 314 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 315 * the end of an object is aligned with the end of the real 316 * allocation. Catches writes behind the end of the allocation. 317 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 318 * redzone word. 319 * cachep->obj_offset: The real object. 320 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 321 * cachep->size - 1* BYTES_PER_WORD: last caller address 322 * [BYTES_PER_WORD long] 323 */ 324 static int obj_offset(struct kmem_cache *cachep) 325 { 326 return cachep->obj_offset; 327 } 328 329 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 330 { 331 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 332 return (unsigned long long*) (objp + obj_offset(cachep) - 333 sizeof(unsigned long long)); 334 } 335 336 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 337 { 338 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 339 if (cachep->flags & SLAB_STORE_USER) 340 return (unsigned long long *)(objp + cachep->size - 341 sizeof(unsigned long long) - 342 REDZONE_ALIGN); 343 return (unsigned long long *) (objp + cachep->size - 344 sizeof(unsigned long long)); 345 } 346 347 static void **dbg_userword(struct kmem_cache *cachep, void *objp) 348 { 349 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 350 return (void **)(objp + cachep->size - BYTES_PER_WORD); 351 } 352 353 #else 354 355 #define obj_offset(x) 0 356 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 357 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 358 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 359 360 #endif 361 362 #ifdef CONFIG_DEBUG_SLAB_LEAK 363 364 static inline bool is_store_user_clean(struct kmem_cache *cachep) 365 { 366 return atomic_read(&cachep->store_user_clean) == 1; 367 } 368 369 static inline void set_store_user_clean(struct kmem_cache *cachep) 370 { 371 atomic_set(&cachep->store_user_clean, 1); 372 } 373 374 static inline void set_store_user_dirty(struct kmem_cache *cachep) 375 { 376 if (is_store_user_clean(cachep)) 377 atomic_set(&cachep->store_user_clean, 0); 378 } 379 380 #else 381 static inline void set_store_user_dirty(struct kmem_cache *cachep) {} 382 383 #endif 384 385 /* 386 * Do not go above this order unless 0 objects fit into the slab or 387 * overridden on the command line. 388 */ 389 #define SLAB_MAX_ORDER_HI 1 390 #define SLAB_MAX_ORDER_LO 0 391 static int slab_max_order = SLAB_MAX_ORDER_LO; 392 static bool slab_max_order_set __initdata; 393 394 static inline struct kmem_cache *virt_to_cache(const void *obj) 395 { 396 struct page *page = virt_to_head_page(obj); 397 return page->slab_cache; 398 } 399 400 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page, 401 unsigned int idx) 402 { 403 return page->s_mem + cache->size * idx; 404 } 405 406 /* 407 * We want to avoid an expensive divide : (offset / cache->size) 408 * Using the fact that size is a constant for a particular cache, 409 * we can replace (offset / cache->size) by 410 * reciprocal_divide(offset, cache->reciprocal_buffer_size) 411 */ 412 static inline unsigned int obj_to_index(const struct kmem_cache *cache, 413 const struct page *page, void *obj) 414 { 415 u32 offset = (obj - page->s_mem); 416 return reciprocal_divide(offset, cache->reciprocal_buffer_size); 417 } 418 419 #define BOOT_CPUCACHE_ENTRIES 1 420 /* internal cache of cache description objs */ 421 static struct kmem_cache kmem_cache_boot = { 422 .batchcount = 1, 423 .limit = BOOT_CPUCACHE_ENTRIES, 424 .shared = 1, 425 .size = sizeof(struct kmem_cache), 426 .name = "kmem_cache", 427 }; 428 429 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); 430 431 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 432 { 433 return this_cpu_ptr(cachep->cpu_cache); 434 } 435 436 /* 437 * Calculate the number of objects and left-over bytes for a given buffer size. 438 */ 439 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size, 440 unsigned long flags, size_t *left_over) 441 { 442 unsigned int num; 443 size_t slab_size = PAGE_SIZE << gfporder; 444 445 /* 446 * The slab management structure can be either off the slab or 447 * on it. For the latter case, the memory allocated for a 448 * slab is used for: 449 * 450 * - @buffer_size bytes for each object 451 * - One freelist_idx_t for each object 452 * 453 * We don't need to consider alignment of freelist because 454 * freelist will be at the end of slab page. The objects will be 455 * at the correct alignment. 456 * 457 * If the slab management structure is off the slab, then the 458 * alignment will already be calculated into the size. Because 459 * the slabs are all pages aligned, the objects will be at the 460 * correct alignment when allocated. 461 */ 462 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) { 463 num = slab_size / buffer_size; 464 *left_over = slab_size % buffer_size; 465 } else { 466 num = slab_size / (buffer_size + sizeof(freelist_idx_t)); 467 *left_over = slab_size % 468 (buffer_size + sizeof(freelist_idx_t)); 469 } 470 471 return num; 472 } 473 474 #if DEBUG 475 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) 476 477 static void __slab_error(const char *function, struct kmem_cache *cachep, 478 char *msg) 479 { 480 pr_err("slab error in %s(): cache `%s': %s\n", 481 function, cachep->name, msg); 482 dump_stack(); 483 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 484 } 485 #endif 486 487 /* 488 * By default on NUMA we use alien caches to stage the freeing of 489 * objects allocated from other nodes. This causes massive memory 490 * inefficiencies when using fake NUMA setup to split memory into a 491 * large number of small nodes, so it can be disabled on the command 492 * line 493 */ 494 495 static int use_alien_caches __read_mostly = 1; 496 static int __init noaliencache_setup(char *s) 497 { 498 use_alien_caches = 0; 499 return 1; 500 } 501 __setup("noaliencache", noaliencache_setup); 502 503 static int __init slab_max_order_setup(char *str) 504 { 505 get_option(&str, &slab_max_order); 506 slab_max_order = slab_max_order < 0 ? 0 : 507 min(slab_max_order, MAX_ORDER - 1); 508 slab_max_order_set = true; 509 510 return 1; 511 } 512 __setup("slab_max_order=", slab_max_order_setup); 513 514 #ifdef CONFIG_NUMA 515 /* 516 * Special reaping functions for NUMA systems called from cache_reap(). 517 * These take care of doing round robin flushing of alien caches (containing 518 * objects freed on different nodes from which they were allocated) and the 519 * flushing of remote pcps by calling drain_node_pages. 520 */ 521 static DEFINE_PER_CPU(unsigned long, slab_reap_node); 522 523 static void init_reap_node(int cpu) 524 { 525 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu), 526 node_online_map); 527 } 528 529 static void next_reap_node(void) 530 { 531 int node = __this_cpu_read(slab_reap_node); 532 533 node = next_node_in(node, node_online_map); 534 __this_cpu_write(slab_reap_node, node); 535 } 536 537 #else 538 #define init_reap_node(cpu) do { } while (0) 539 #define next_reap_node(void) do { } while (0) 540 #endif 541 542 /* 543 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 544 * via the workqueue/eventd. 545 * Add the CPU number into the expiration time to minimize the possibility of 546 * the CPUs getting into lockstep and contending for the global cache chain 547 * lock. 548 */ 549 static void start_cpu_timer(int cpu) 550 { 551 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); 552 553 /* 554 * When this gets called from do_initcalls via cpucache_init(), 555 * init_workqueues() has already run, so keventd will be setup 556 * at that time. 557 */ 558 if (keventd_up() && reap_work->work.func == NULL) { 559 init_reap_node(cpu); 560 INIT_DEFERRABLE_WORK(reap_work, cache_reap); 561 schedule_delayed_work_on(cpu, reap_work, 562 __round_jiffies_relative(HZ, cpu)); 563 } 564 } 565 566 static void init_arraycache(struct array_cache *ac, int limit, int batch) 567 { 568 /* 569 * The array_cache structures contain pointers to free object. 570 * However, when such objects are allocated or transferred to another 571 * cache the pointers are not cleared and they could be counted as 572 * valid references during a kmemleak scan. Therefore, kmemleak must 573 * not scan such objects. 574 */ 575 kmemleak_no_scan(ac); 576 if (ac) { 577 ac->avail = 0; 578 ac->limit = limit; 579 ac->batchcount = batch; 580 ac->touched = 0; 581 } 582 } 583 584 static struct array_cache *alloc_arraycache(int node, int entries, 585 int batchcount, gfp_t gfp) 586 { 587 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache); 588 struct array_cache *ac = NULL; 589 590 ac = kmalloc_node(memsize, gfp, node); 591 init_arraycache(ac, entries, batchcount); 592 return ac; 593 } 594 595 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep, 596 struct page *page, void *objp) 597 { 598 struct kmem_cache_node *n; 599 int page_node; 600 LIST_HEAD(list); 601 602 page_node = page_to_nid(page); 603 n = get_node(cachep, page_node); 604 605 spin_lock(&n->list_lock); 606 free_block(cachep, &objp, 1, page_node, &list); 607 spin_unlock(&n->list_lock); 608 609 slabs_destroy(cachep, &list); 610 } 611 612 /* 613 * Transfer objects in one arraycache to another. 614 * Locking must be handled by the caller. 615 * 616 * Return the number of entries transferred. 617 */ 618 static int transfer_objects(struct array_cache *to, 619 struct array_cache *from, unsigned int max) 620 { 621 /* Figure out how many entries to transfer */ 622 int nr = min3(from->avail, max, to->limit - to->avail); 623 624 if (!nr) 625 return 0; 626 627 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 628 sizeof(void *) *nr); 629 630 from->avail -= nr; 631 to->avail += nr; 632 return nr; 633 } 634 635 #ifndef CONFIG_NUMA 636 637 #define drain_alien_cache(cachep, alien) do { } while (0) 638 #define reap_alien(cachep, n) do { } while (0) 639 640 static inline struct alien_cache **alloc_alien_cache(int node, 641 int limit, gfp_t gfp) 642 { 643 return NULL; 644 } 645 646 static inline void free_alien_cache(struct alien_cache **ac_ptr) 647 { 648 } 649 650 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 651 { 652 return 0; 653 } 654 655 static inline void *alternate_node_alloc(struct kmem_cache *cachep, 656 gfp_t flags) 657 { 658 return NULL; 659 } 660 661 static inline void *____cache_alloc_node(struct kmem_cache *cachep, 662 gfp_t flags, int nodeid) 663 { 664 return NULL; 665 } 666 667 static inline gfp_t gfp_exact_node(gfp_t flags) 668 { 669 return flags & ~__GFP_NOFAIL; 670 } 671 672 #else /* CONFIG_NUMA */ 673 674 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 675 static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 676 677 static struct alien_cache *__alloc_alien_cache(int node, int entries, 678 int batch, gfp_t gfp) 679 { 680 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache); 681 struct alien_cache *alc = NULL; 682 683 alc = kmalloc_node(memsize, gfp, node); 684 init_arraycache(&alc->ac, entries, batch); 685 spin_lock_init(&alc->lock); 686 return alc; 687 } 688 689 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) 690 { 691 struct alien_cache **alc_ptr; 692 size_t memsize = sizeof(void *) * nr_node_ids; 693 int i; 694 695 if (limit > 1) 696 limit = 12; 697 alc_ptr = kzalloc_node(memsize, gfp, node); 698 if (!alc_ptr) 699 return NULL; 700 701 for_each_node(i) { 702 if (i == node || !node_online(i)) 703 continue; 704 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp); 705 if (!alc_ptr[i]) { 706 for (i--; i >= 0; i--) 707 kfree(alc_ptr[i]); 708 kfree(alc_ptr); 709 return NULL; 710 } 711 } 712 return alc_ptr; 713 } 714 715 static void free_alien_cache(struct alien_cache **alc_ptr) 716 { 717 int i; 718 719 if (!alc_ptr) 720 return; 721 for_each_node(i) 722 kfree(alc_ptr[i]); 723 kfree(alc_ptr); 724 } 725 726 static void __drain_alien_cache(struct kmem_cache *cachep, 727 struct array_cache *ac, int node, 728 struct list_head *list) 729 { 730 struct kmem_cache_node *n = get_node(cachep, node); 731 732 if (ac->avail) { 733 spin_lock(&n->list_lock); 734 /* 735 * Stuff objects into the remote nodes shared array first. 736 * That way we could avoid the overhead of putting the objects 737 * into the free lists and getting them back later. 738 */ 739 if (n->shared) 740 transfer_objects(n->shared, ac, ac->limit); 741 742 free_block(cachep, ac->entry, ac->avail, node, list); 743 ac->avail = 0; 744 spin_unlock(&n->list_lock); 745 } 746 } 747 748 /* 749 * Called from cache_reap() to regularly drain alien caches round robin. 750 */ 751 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n) 752 { 753 int node = __this_cpu_read(slab_reap_node); 754 755 if (n->alien) { 756 struct alien_cache *alc = n->alien[node]; 757 struct array_cache *ac; 758 759 if (alc) { 760 ac = &alc->ac; 761 if (ac->avail && spin_trylock_irq(&alc->lock)) { 762 LIST_HEAD(list); 763 764 __drain_alien_cache(cachep, ac, node, &list); 765 spin_unlock_irq(&alc->lock); 766 slabs_destroy(cachep, &list); 767 } 768 } 769 } 770 } 771 772 static void drain_alien_cache(struct kmem_cache *cachep, 773 struct alien_cache **alien) 774 { 775 int i = 0; 776 struct alien_cache *alc; 777 struct array_cache *ac; 778 unsigned long flags; 779 780 for_each_online_node(i) { 781 alc = alien[i]; 782 if (alc) { 783 LIST_HEAD(list); 784 785 ac = &alc->ac; 786 spin_lock_irqsave(&alc->lock, flags); 787 __drain_alien_cache(cachep, ac, i, &list); 788 spin_unlock_irqrestore(&alc->lock, flags); 789 slabs_destroy(cachep, &list); 790 } 791 } 792 } 793 794 static int __cache_free_alien(struct kmem_cache *cachep, void *objp, 795 int node, int page_node) 796 { 797 struct kmem_cache_node *n; 798 struct alien_cache *alien = NULL; 799 struct array_cache *ac; 800 LIST_HEAD(list); 801 802 n = get_node(cachep, node); 803 STATS_INC_NODEFREES(cachep); 804 if (n->alien && n->alien[page_node]) { 805 alien = n->alien[page_node]; 806 ac = &alien->ac; 807 spin_lock(&alien->lock); 808 if (unlikely(ac->avail == ac->limit)) { 809 STATS_INC_ACOVERFLOW(cachep); 810 __drain_alien_cache(cachep, ac, page_node, &list); 811 } 812 ac->entry[ac->avail++] = objp; 813 spin_unlock(&alien->lock); 814 slabs_destroy(cachep, &list); 815 } else { 816 n = get_node(cachep, page_node); 817 spin_lock(&n->list_lock); 818 free_block(cachep, &objp, 1, page_node, &list); 819 spin_unlock(&n->list_lock); 820 slabs_destroy(cachep, &list); 821 } 822 return 1; 823 } 824 825 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 826 { 827 int page_node = page_to_nid(virt_to_page(objp)); 828 int node = numa_mem_id(); 829 /* 830 * Make sure we are not freeing a object from another node to the array 831 * cache on this cpu. 832 */ 833 if (likely(node == page_node)) 834 return 0; 835 836 return __cache_free_alien(cachep, objp, node, page_node); 837 } 838 839 /* 840 * Construct gfp mask to allocate from a specific node but do not reclaim or 841 * warn about failures. 842 */ 843 static inline gfp_t gfp_exact_node(gfp_t flags) 844 { 845 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL); 846 } 847 #endif 848 849 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp) 850 { 851 struct kmem_cache_node *n; 852 853 /* 854 * Set up the kmem_cache_node for cpu before we can 855 * begin anything. Make sure some other cpu on this 856 * node has not already allocated this 857 */ 858 n = get_node(cachep, node); 859 if (n) { 860 spin_lock_irq(&n->list_lock); 861 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount + 862 cachep->num; 863 spin_unlock_irq(&n->list_lock); 864 865 return 0; 866 } 867 868 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node); 869 if (!n) 870 return -ENOMEM; 871 872 kmem_cache_node_init(n); 873 n->next_reap = jiffies + REAPTIMEOUT_NODE + 874 ((unsigned long)cachep) % REAPTIMEOUT_NODE; 875 876 n->free_limit = 877 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num; 878 879 /* 880 * The kmem_cache_nodes don't come and go as CPUs 881 * come and go. slab_mutex is sufficient 882 * protection here. 883 */ 884 cachep->node[node] = n; 885 886 return 0; 887 } 888 889 /* 890 * Allocates and initializes node for a node on each slab cache, used for 891 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node 892 * will be allocated off-node since memory is not yet online for the new node. 893 * When hotplugging memory or a cpu, existing node are not replaced if 894 * already in use. 895 * 896 * Must hold slab_mutex. 897 */ 898 static int init_cache_node_node(int node) 899 { 900 int ret; 901 struct kmem_cache *cachep; 902 903 list_for_each_entry(cachep, &slab_caches, list) { 904 ret = init_cache_node(cachep, node, GFP_KERNEL); 905 if (ret) 906 return ret; 907 } 908 909 return 0; 910 } 911 912 static int setup_kmem_cache_node(struct kmem_cache *cachep, 913 int node, gfp_t gfp, bool force_change) 914 { 915 int ret = -ENOMEM; 916 struct kmem_cache_node *n; 917 struct array_cache *old_shared = NULL; 918 struct array_cache *new_shared = NULL; 919 struct alien_cache **new_alien = NULL; 920 LIST_HEAD(list); 921 922 if (use_alien_caches) { 923 new_alien = alloc_alien_cache(node, cachep->limit, gfp); 924 if (!new_alien) 925 goto fail; 926 } 927 928 if (cachep->shared) { 929 new_shared = alloc_arraycache(node, 930 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp); 931 if (!new_shared) 932 goto fail; 933 } 934 935 ret = init_cache_node(cachep, node, gfp); 936 if (ret) 937 goto fail; 938 939 n = get_node(cachep, node); 940 spin_lock_irq(&n->list_lock); 941 if (n->shared && force_change) { 942 free_block(cachep, n->shared->entry, 943 n->shared->avail, node, &list); 944 n->shared->avail = 0; 945 } 946 947 if (!n->shared || force_change) { 948 old_shared = n->shared; 949 n->shared = new_shared; 950 new_shared = NULL; 951 } 952 953 if (!n->alien) { 954 n->alien = new_alien; 955 new_alien = NULL; 956 } 957 958 spin_unlock_irq(&n->list_lock); 959 slabs_destroy(cachep, &list); 960 961 /* 962 * To protect lockless access to n->shared during irq disabled context. 963 * If n->shared isn't NULL in irq disabled context, accessing to it is 964 * guaranteed to be valid until irq is re-enabled, because it will be 965 * freed after synchronize_sched(). 966 */ 967 if (force_change) 968 synchronize_sched(); 969 970 fail: 971 kfree(old_shared); 972 kfree(new_shared); 973 free_alien_cache(new_alien); 974 975 return ret; 976 } 977 978 static void cpuup_canceled(long cpu) 979 { 980 struct kmem_cache *cachep; 981 struct kmem_cache_node *n = NULL; 982 int node = cpu_to_mem(cpu); 983 const struct cpumask *mask = cpumask_of_node(node); 984 985 list_for_each_entry(cachep, &slab_caches, list) { 986 struct array_cache *nc; 987 struct array_cache *shared; 988 struct alien_cache **alien; 989 LIST_HEAD(list); 990 991 n = get_node(cachep, node); 992 if (!n) 993 continue; 994 995 spin_lock_irq(&n->list_lock); 996 997 /* Free limit for this kmem_cache_node */ 998 n->free_limit -= cachep->batchcount; 999 1000 /* cpu is dead; no one can alloc from it. */ 1001 nc = per_cpu_ptr(cachep->cpu_cache, cpu); 1002 if (nc) { 1003 free_block(cachep, nc->entry, nc->avail, node, &list); 1004 nc->avail = 0; 1005 } 1006 1007 if (!cpumask_empty(mask)) { 1008 spin_unlock_irq(&n->list_lock); 1009 goto free_slab; 1010 } 1011 1012 shared = n->shared; 1013 if (shared) { 1014 free_block(cachep, shared->entry, 1015 shared->avail, node, &list); 1016 n->shared = NULL; 1017 } 1018 1019 alien = n->alien; 1020 n->alien = NULL; 1021 1022 spin_unlock_irq(&n->list_lock); 1023 1024 kfree(shared); 1025 if (alien) { 1026 drain_alien_cache(cachep, alien); 1027 free_alien_cache(alien); 1028 } 1029 1030 free_slab: 1031 slabs_destroy(cachep, &list); 1032 } 1033 /* 1034 * In the previous loop, all the objects were freed to 1035 * the respective cache's slabs, now we can go ahead and 1036 * shrink each nodelist to its limit. 1037 */ 1038 list_for_each_entry(cachep, &slab_caches, list) { 1039 n = get_node(cachep, node); 1040 if (!n) 1041 continue; 1042 drain_freelist(cachep, n, INT_MAX); 1043 } 1044 } 1045 1046 static int cpuup_prepare(long cpu) 1047 { 1048 struct kmem_cache *cachep; 1049 int node = cpu_to_mem(cpu); 1050 int err; 1051 1052 /* 1053 * We need to do this right in the beginning since 1054 * alloc_arraycache's are going to use this list. 1055 * kmalloc_node allows us to add the slab to the right 1056 * kmem_cache_node and not this cpu's kmem_cache_node 1057 */ 1058 err = init_cache_node_node(node); 1059 if (err < 0) 1060 goto bad; 1061 1062 /* 1063 * Now we can go ahead with allocating the shared arrays and 1064 * array caches 1065 */ 1066 list_for_each_entry(cachep, &slab_caches, list) { 1067 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false); 1068 if (err) 1069 goto bad; 1070 } 1071 1072 return 0; 1073 bad: 1074 cpuup_canceled(cpu); 1075 return -ENOMEM; 1076 } 1077 1078 static int cpuup_callback(struct notifier_block *nfb, 1079 unsigned long action, void *hcpu) 1080 { 1081 long cpu = (long)hcpu; 1082 int err = 0; 1083 1084 switch (action) { 1085 case CPU_UP_PREPARE: 1086 case CPU_UP_PREPARE_FROZEN: 1087 mutex_lock(&slab_mutex); 1088 err = cpuup_prepare(cpu); 1089 mutex_unlock(&slab_mutex); 1090 break; 1091 case CPU_ONLINE: 1092 case CPU_ONLINE_FROZEN: 1093 start_cpu_timer(cpu); 1094 break; 1095 #ifdef CONFIG_HOTPLUG_CPU 1096 case CPU_DOWN_PREPARE: 1097 case CPU_DOWN_PREPARE_FROZEN: 1098 /* 1099 * Shutdown cache reaper. Note that the slab_mutex is 1100 * held so that if cache_reap() is invoked it cannot do 1101 * anything expensive but will only modify reap_work 1102 * and reschedule the timer. 1103 */ 1104 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); 1105 /* Now the cache_reaper is guaranteed to be not running. */ 1106 per_cpu(slab_reap_work, cpu).work.func = NULL; 1107 break; 1108 case CPU_DOWN_FAILED: 1109 case CPU_DOWN_FAILED_FROZEN: 1110 start_cpu_timer(cpu); 1111 break; 1112 case CPU_DEAD: 1113 case CPU_DEAD_FROZEN: 1114 /* 1115 * Even if all the cpus of a node are down, we don't free the 1116 * kmem_cache_node of any cache. This to avoid a race between 1117 * cpu_down, and a kmalloc allocation from another cpu for 1118 * memory from the node of the cpu going down. The node 1119 * structure is usually allocated from kmem_cache_create() and 1120 * gets destroyed at kmem_cache_destroy(). 1121 */ 1122 /* fall through */ 1123 #endif 1124 case CPU_UP_CANCELED: 1125 case CPU_UP_CANCELED_FROZEN: 1126 mutex_lock(&slab_mutex); 1127 cpuup_canceled(cpu); 1128 mutex_unlock(&slab_mutex); 1129 break; 1130 } 1131 return notifier_from_errno(err); 1132 } 1133 1134 static struct notifier_block cpucache_notifier = { 1135 &cpuup_callback, NULL, 0 1136 }; 1137 1138 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 1139 /* 1140 * Drains freelist for a node on each slab cache, used for memory hot-remove. 1141 * Returns -EBUSY if all objects cannot be drained so that the node is not 1142 * removed. 1143 * 1144 * Must hold slab_mutex. 1145 */ 1146 static int __meminit drain_cache_node_node(int node) 1147 { 1148 struct kmem_cache *cachep; 1149 int ret = 0; 1150 1151 list_for_each_entry(cachep, &slab_caches, list) { 1152 struct kmem_cache_node *n; 1153 1154 n = get_node(cachep, node); 1155 if (!n) 1156 continue; 1157 1158 drain_freelist(cachep, n, INT_MAX); 1159 1160 if (!list_empty(&n->slabs_full) || 1161 !list_empty(&n->slabs_partial)) { 1162 ret = -EBUSY; 1163 break; 1164 } 1165 } 1166 return ret; 1167 } 1168 1169 static int __meminit slab_memory_callback(struct notifier_block *self, 1170 unsigned long action, void *arg) 1171 { 1172 struct memory_notify *mnb = arg; 1173 int ret = 0; 1174 int nid; 1175 1176 nid = mnb->status_change_nid; 1177 if (nid < 0) 1178 goto out; 1179 1180 switch (action) { 1181 case MEM_GOING_ONLINE: 1182 mutex_lock(&slab_mutex); 1183 ret = init_cache_node_node(nid); 1184 mutex_unlock(&slab_mutex); 1185 break; 1186 case MEM_GOING_OFFLINE: 1187 mutex_lock(&slab_mutex); 1188 ret = drain_cache_node_node(nid); 1189 mutex_unlock(&slab_mutex); 1190 break; 1191 case MEM_ONLINE: 1192 case MEM_OFFLINE: 1193 case MEM_CANCEL_ONLINE: 1194 case MEM_CANCEL_OFFLINE: 1195 break; 1196 } 1197 out: 1198 return notifier_from_errno(ret); 1199 } 1200 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ 1201 1202 /* 1203 * swap the static kmem_cache_node with kmalloced memory 1204 */ 1205 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list, 1206 int nodeid) 1207 { 1208 struct kmem_cache_node *ptr; 1209 1210 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid); 1211 BUG_ON(!ptr); 1212 1213 memcpy(ptr, list, sizeof(struct kmem_cache_node)); 1214 /* 1215 * Do not assume that spinlocks can be initialized via memcpy: 1216 */ 1217 spin_lock_init(&ptr->list_lock); 1218 1219 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1220 cachep->node[nodeid] = ptr; 1221 } 1222 1223 /* 1224 * For setting up all the kmem_cache_node for cache whose buffer_size is same as 1225 * size of kmem_cache_node. 1226 */ 1227 static void __init set_up_node(struct kmem_cache *cachep, int index) 1228 { 1229 int node; 1230 1231 for_each_online_node(node) { 1232 cachep->node[node] = &init_kmem_cache_node[index + node]; 1233 cachep->node[node]->next_reap = jiffies + 1234 REAPTIMEOUT_NODE + 1235 ((unsigned long)cachep) % REAPTIMEOUT_NODE; 1236 } 1237 } 1238 1239 /* 1240 * Initialisation. Called after the page allocator have been initialised and 1241 * before smp_init(). 1242 */ 1243 void __init kmem_cache_init(void) 1244 { 1245 int i; 1246 1247 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) < 1248 sizeof(struct rcu_head)); 1249 kmem_cache = &kmem_cache_boot; 1250 1251 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1) 1252 use_alien_caches = 0; 1253 1254 for (i = 0; i < NUM_INIT_LISTS; i++) 1255 kmem_cache_node_init(&init_kmem_cache_node[i]); 1256 1257 /* 1258 * Fragmentation resistance on low memory - only use bigger 1259 * page orders on machines with more than 32MB of memory if 1260 * not overridden on the command line. 1261 */ 1262 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) 1263 slab_max_order = SLAB_MAX_ORDER_HI; 1264 1265 /* Bootstrap is tricky, because several objects are allocated 1266 * from caches that do not exist yet: 1267 * 1) initialize the kmem_cache cache: it contains the struct 1268 * kmem_cache structures of all caches, except kmem_cache itself: 1269 * kmem_cache is statically allocated. 1270 * Initially an __init data area is used for the head array and the 1271 * kmem_cache_node structures, it's replaced with a kmalloc allocated 1272 * array at the end of the bootstrap. 1273 * 2) Create the first kmalloc cache. 1274 * The struct kmem_cache for the new cache is allocated normally. 1275 * An __init data area is used for the head array. 1276 * 3) Create the remaining kmalloc caches, with minimally sized 1277 * head arrays. 1278 * 4) Replace the __init data head arrays for kmem_cache and the first 1279 * kmalloc cache with kmalloc allocated arrays. 1280 * 5) Replace the __init data for kmem_cache_node for kmem_cache and 1281 * the other cache's with kmalloc allocated memory. 1282 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1283 */ 1284 1285 /* 1) create the kmem_cache */ 1286 1287 /* 1288 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids 1289 */ 1290 create_boot_cache(kmem_cache, "kmem_cache", 1291 offsetof(struct kmem_cache, node) + 1292 nr_node_ids * sizeof(struct kmem_cache_node *), 1293 SLAB_HWCACHE_ALIGN); 1294 list_add(&kmem_cache->list, &slab_caches); 1295 slab_state = PARTIAL; 1296 1297 /* 1298 * Initialize the caches that provide memory for the kmem_cache_node 1299 * structures first. Without this, further allocations will bug. 1300 */ 1301 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node", 1302 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS); 1303 slab_state = PARTIAL_NODE; 1304 setup_kmalloc_cache_index_table(); 1305 1306 slab_early_init = 0; 1307 1308 /* 5) Replace the bootstrap kmem_cache_node */ 1309 { 1310 int nid; 1311 1312 for_each_online_node(nid) { 1313 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid); 1314 1315 init_list(kmalloc_caches[INDEX_NODE], 1316 &init_kmem_cache_node[SIZE_NODE + nid], nid); 1317 } 1318 } 1319 1320 create_kmalloc_caches(ARCH_KMALLOC_FLAGS); 1321 } 1322 1323 void __init kmem_cache_init_late(void) 1324 { 1325 struct kmem_cache *cachep; 1326 1327 slab_state = UP; 1328 1329 /* 6) resize the head arrays to their final sizes */ 1330 mutex_lock(&slab_mutex); 1331 list_for_each_entry(cachep, &slab_caches, list) 1332 if (enable_cpucache(cachep, GFP_NOWAIT)) 1333 BUG(); 1334 mutex_unlock(&slab_mutex); 1335 1336 /* Done! */ 1337 slab_state = FULL; 1338 1339 /* 1340 * Register a cpu startup notifier callback that initializes 1341 * cpu_cache_get for all new cpus 1342 */ 1343 register_cpu_notifier(&cpucache_notifier); 1344 1345 #ifdef CONFIG_NUMA 1346 /* 1347 * Register a memory hotplug callback that initializes and frees 1348 * node. 1349 */ 1350 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 1351 #endif 1352 1353 /* 1354 * The reap timers are started later, with a module init call: That part 1355 * of the kernel is not yet operational. 1356 */ 1357 } 1358 1359 static int __init cpucache_init(void) 1360 { 1361 int cpu; 1362 1363 /* 1364 * Register the timers that return unneeded pages to the page allocator 1365 */ 1366 for_each_online_cpu(cpu) 1367 start_cpu_timer(cpu); 1368 1369 /* Done! */ 1370 slab_state = FULL; 1371 return 0; 1372 } 1373 __initcall(cpucache_init); 1374 1375 static noinline void 1376 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) 1377 { 1378 #if DEBUG 1379 struct kmem_cache_node *n; 1380 struct page *page; 1381 unsigned long flags; 1382 int node; 1383 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL, 1384 DEFAULT_RATELIMIT_BURST); 1385 1386 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs)) 1387 return; 1388 1389 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", 1390 nodeid, gfpflags, &gfpflags); 1391 pr_warn(" cache: %s, object size: %d, order: %d\n", 1392 cachep->name, cachep->size, cachep->gfporder); 1393 1394 for_each_kmem_cache_node(cachep, node, n) { 1395 unsigned long active_objs = 0, num_objs = 0, free_objects = 0; 1396 unsigned long active_slabs = 0, num_slabs = 0; 1397 1398 spin_lock_irqsave(&n->list_lock, flags); 1399 list_for_each_entry(page, &n->slabs_full, lru) { 1400 active_objs += cachep->num; 1401 active_slabs++; 1402 } 1403 list_for_each_entry(page, &n->slabs_partial, lru) { 1404 active_objs += page->active; 1405 active_slabs++; 1406 } 1407 list_for_each_entry(page, &n->slabs_free, lru) 1408 num_slabs++; 1409 1410 free_objects += n->free_objects; 1411 spin_unlock_irqrestore(&n->list_lock, flags); 1412 1413 num_slabs += active_slabs; 1414 num_objs = num_slabs * cachep->num; 1415 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", 1416 node, active_slabs, num_slabs, active_objs, num_objs, 1417 free_objects); 1418 } 1419 #endif 1420 } 1421 1422 /* 1423 * Interface to system's page allocator. No need to hold the 1424 * kmem_cache_node ->list_lock. 1425 * 1426 * If we requested dmaable memory, we will get it. Even if we 1427 * did not request dmaable memory, we might get it, but that 1428 * would be relatively rare and ignorable. 1429 */ 1430 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, 1431 int nodeid) 1432 { 1433 struct page *page; 1434 int nr_pages; 1435 1436 flags |= cachep->allocflags; 1437 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1438 flags |= __GFP_RECLAIMABLE; 1439 1440 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); 1441 if (!page) { 1442 slab_out_of_memory(cachep, flags, nodeid); 1443 return NULL; 1444 } 1445 1446 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) { 1447 __free_pages(page, cachep->gfporder); 1448 return NULL; 1449 } 1450 1451 nr_pages = (1 << cachep->gfporder); 1452 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1453 add_zone_page_state(page_zone(page), 1454 NR_SLAB_RECLAIMABLE, nr_pages); 1455 else 1456 add_zone_page_state(page_zone(page), 1457 NR_SLAB_UNRECLAIMABLE, nr_pages); 1458 1459 __SetPageSlab(page); 1460 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ 1461 if (sk_memalloc_socks() && page_is_pfmemalloc(page)) 1462 SetPageSlabPfmemalloc(page); 1463 1464 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { 1465 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); 1466 1467 if (cachep->ctor) 1468 kmemcheck_mark_uninitialized_pages(page, nr_pages); 1469 else 1470 kmemcheck_mark_unallocated_pages(page, nr_pages); 1471 } 1472 1473 return page; 1474 } 1475 1476 /* 1477 * Interface to system's page release. 1478 */ 1479 static void kmem_freepages(struct kmem_cache *cachep, struct page *page) 1480 { 1481 int order = cachep->gfporder; 1482 unsigned long nr_freed = (1 << order); 1483 1484 kmemcheck_free_shadow(page, order); 1485 1486 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1487 sub_zone_page_state(page_zone(page), 1488 NR_SLAB_RECLAIMABLE, nr_freed); 1489 else 1490 sub_zone_page_state(page_zone(page), 1491 NR_SLAB_UNRECLAIMABLE, nr_freed); 1492 1493 BUG_ON(!PageSlab(page)); 1494 __ClearPageSlabPfmemalloc(page); 1495 __ClearPageSlab(page); 1496 page_mapcount_reset(page); 1497 page->mapping = NULL; 1498 1499 if (current->reclaim_state) 1500 current->reclaim_state->reclaimed_slab += nr_freed; 1501 memcg_uncharge_slab(page, order, cachep); 1502 __free_pages(page, order); 1503 } 1504 1505 static void kmem_rcu_free(struct rcu_head *head) 1506 { 1507 struct kmem_cache *cachep; 1508 struct page *page; 1509 1510 page = container_of(head, struct page, rcu_head); 1511 cachep = page->slab_cache; 1512 1513 kmem_freepages(cachep, page); 1514 } 1515 1516 #if DEBUG 1517 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep) 1518 { 1519 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) && 1520 (cachep->size % PAGE_SIZE) == 0) 1521 return true; 1522 1523 return false; 1524 } 1525 1526 #ifdef CONFIG_DEBUG_PAGEALLOC 1527 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1528 unsigned long caller) 1529 { 1530 int size = cachep->object_size; 1531 1532 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1533 1534 if (size < 5 * sizeof(unsigned long)) 1535 return; 1536 1537 *addr++ = 0x12345678; 1538 *addr++ = caller; 1539 *addr++ = smp_processor_id(); 1540 size -= 3 * sizeof(unsigned long); 1541 { 1542 unsigned long *sptr = &caller; 1543 unsigned long svalue; 1544 1545 while (!kstack_end(sptr)) { 1546 svalue = *sptr++; 1547 if (kernel_text_address(svalue)) { 1548 *addr++ = svalue; 1549 size -= sizeof(unsigned long); 1550 if (size <= sizeof(unsigned long)) 1551 break; 1552 } 1553 } 1554 1555 } 1556 *addr++ = 0x87654321; 1557 } 1558 1559 static void slab_kernel_map(struct kmem_cache *cachep, void *objp, 1560 int map, unsigned long caller) 1561 { 1562 if (!is_debug_pagealloc_cache(cachep)) 1563 return; 1564 1565 if (caller) 1566 store_stackinfo(cachep, objp, caller); 1567 1568 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map); 1569 } 1570 1571 #else 1572 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp, 1573 int map, unsigned long caller) {} 1574 1575 #endif 1576 1577 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1578 { 1579 int size = cachep->object_size; 1580 addr = &((char *)addr)[obj_offset(cachep)]; 1581 1582 memset(addr, val, size); 1583 *(unsigned char *)(addr + size - 1) = POISON_END; 1584 } 1585 1586 static void dump_line(char *data, int offset, int limit) 1587 { 1588 int i; 1589 unsigned char error = 0; 1590 int bad_count = 0; 1591 1592 pr_err("%03x: ", offset); 1593 for (i = 0; i < limit; i++) { 1594 if (data[offset + i] != POISON_FREE) { 1595 error = data[offset + i]; 1596 bad_count++; 1597 } 1598 } 1599 print_hex_dump(KERN_CONT, "", 0, 16, 1, 1600 &data[offset], limit, 1); 1601 1602 if (bad_count == 1) { 1603 error ^= POISON_FREE; 1604 if (!(error & (error - 1))) { 1605 pr_err("Single bit error detected. Probably bad RAM.\n"); 1606 #ifdef CONFIG_X86 1607 pr_err("Run memtest86+ or a similar memory test tool.\n"); 1608 #else 1609 pr_err("Run a memory test tool.\n"); 1610 #endif 1611 } 1612 } 1613 } 1614 #endif 1615 1616 #if DEBUG 1617 1618 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1619 { 1620 int i, size; 1621 char *realobj; 1622 1623 if (cachep->flags & SLAB_RED_ZONE) { 1624 pr_err("Redzone: 0x%llx/0x%llx\n", 1625 *dbg_redzone1(cachep, objp), 1626 *dbg_redzone2(cachep, objp)); 1627 } 1628 1629 if (cachep->flags & SLAB_STORE_USER) { 1630 pr_err("Last user: [<%p>](%pSR)\n", 1631 *dbg_userword(cachep, objp), 1632 *dbg_userword(cachep, objp)); 1633 } 1634 realobj = (char *)objp + obj_offset(cachep); 1635 size = cachep->object_size; 1636 for (i = 0; i < size && lines; i += 16, lines--) { 1637 int limit; 1638 limit = 16; 1639 if (i + limit > size) 1640 limit = size - i; 1641 dump_line(realobj, i, limit); 1642 } 1643 } 1644 1645 static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1646 { 1647 char *realobj; 1648 int size, i; 1649 int lines = 0; 1650 1651 if (is_debug_pagealloc_cache(cachep)) 1652 return; 1653 1654 realobj = (char *)objp + obj_offset(cachep); 1655 size = cachep->object_size; 1656 1657 for (i = 0; i < size; i++) { 1658 char exp = POISON_FREE; 1659 if (i == size - 1) 1660 exp = POISON_END; 1661 if (realobj[i] != exp) { 1662 int limit; 1663 /* Mismatch ! */ 1664 /* Print header */ 1665 if (lines == 0) { 1666 pr_err("Slab corruption (%s): %s start=%p, len=%d\n", 1667 print_tainted(), cachep->name, 1668 realobj, size); 1669 print_objinfo(cachep, objp, 0); 1670 } 1671 /* Hexdump the affected line */ 1672 i = (i / 16) * 16; 1673 limit = 16; 1674 if (i + limit > size) 1675 limit = size - i; 1676 dump_line(realobj, i, limit); 1677 i += 16; 1678 lines++; 1679 /* Limit to 5 lines */ 1680 if (lines > 5) 1681 break; 1682 } 1683 } 1684 if (lines != 0) { 1685 /* Print some data about the neighboring objects, if they 1686 * exist: 1687 */ 1688 struct page *page = virt_to_head_page(objp); 1689 unsigned int objnr; 1690 1691 objnr = obj_to_index(cachep, page, objp); 1692 if (objnr) { 1693 objp = index_to_obj(cachep, page, objnr - 1); 1694 realobj = (char *)objp + obj_offset(cachep); 1695 pr_err("Prev obj: start=%p, len=%d\n", realobj, size); 1696 print_objinfo(cachep, objp, 2); 1697 } 1698 if (objnr + 1 < cachep->num) { 1699 objp = index_to_obj(cachep, page, objnr + 1); 1700 realobj = (char *)objp + obj_offset(cachep); 1701 pr_err("Next obj: start=%p, len=%d\n", realobj, size); 1702 print_objinfo(cachep, objp, 2); 1703 } 1704 } 1705 } 1706 #endif 1707 1708 #if DEBUG 1709 static void slab_destroy_debugcheck(struct kmem_cache *cachep, 1710 struct page *page) 1711 { 1712 int i; 1713 1714 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) { 1715 poison_obj(cachep, page->freelist - obj_offset(cachep), 1716 POISON_FREE); 1717 } 1718 1719 for (i = 0; i < cachep->num; i++) { 1720 void *objp = index_to_obj(cachep, page, i); 1721 1722 if (cachep->flags & SLAB_POISON) { 1723 check_poison_obj(cachep, objp); 1724 slab_kernel_map(cachep, objp, 1, 0); 1725 } 1726 if (cachep->flags & SLAB_RED_ZONE) { 1727 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1728 slab_error(cachep, "start of a freed object was overwritten"); 1729 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1730 slab_error(cachep, "end of a freed object was overwritten"); 1731 } 1732 } 1733 } 1734 #else 1735 static void slab_destroy_debugcheck(struct kmem_cache *cachep, 1736 struct page *page) 1737 { 1738 } 1739 #endif 1740 1741 /** 1742 * slab_destroy - destroy and release all objects in a slab 1743 * @cachep: cache pointer being destroyed 1744 * @page: page pointer being destroyed 1745 * 1746 * Destroy all the objs in a slab page, and release the mem back to the system. 1747 * Before calling the slab page must have been unlinked from the cache. The 1748 * kmem_cache_node ->list_lock is not held/needed. 1749 */ 1750 static void slab_destroy(struct kmem_cache *cachep, struct page *page) 1751 { 1752 void *freelist; 1753 1754 freelist = page->freelist; 1755 slab_destroy_debugcheck(cachep, page); 1756 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) 1757 call_rcu(&page->rcu_head, kmem_rcu_free); 1758 else 1759 kmem_freepages(cachep, page); 1760 1761 /* 1762 * From now on, we don't use freelist 1763 * although actual page can be freed in rcu context 1764 */ 1765 if (OFF_SLAB(cachep)) 1766 kmem_cache_free(cachep->freelist_cache, freelist); 1767 } 1768 1769 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list) 1770 { 1771 struct page *page, *n; 1772 1773 list_for_each_entry_safe(page, n, list, lru) { 1774 list_del(&page->lru); 1775 slab_destroy(cachep, page); 1776 } 1777 } 1778 1779 /** 1780 * calculate_slab_order - calculate size (page order) of slabs 1781 * @cachep: pointer to the cache that is being created 1782 * @size: size of objects to be created in this cache. 1783 * @flags: slab allocation flags 1784 * 1785 * Also calculates the number of objects per slab. 1786 * 1787 * This could be made much more intelligent. For now, try to avoid using 1788 * high order pages for slabs. When the gfp() functions are more friendly 1789 * towards high-order requests, this should be changed. 1790 */ 1791 static size_t calculate_slab_order(struct kmem_cache *cachep, 1792 size_t size, unsigned long flags) 1793 { 1794 size_t left_over = 0; 1795 int gfporder; 1796 1797 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { 1798 unsigned int num; 1799 size_t remainder; 1800 1801 num = cache_estimate(gfporder, size, flags, &remainder); 1802 if (!num) 1803 continue; 1804 1805 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */ 1806 if (num > SLAB_OBJ_MAX_NUM) 1807 break; 1808 1809 if (flags & CFLGS_OFF_SLAB) { 1810 struct kmem_cache *freelist_cache; 1811 size_t freelist_size; 1812 1813 freelist_size = num * sizeof(freelist_idx_t); 1814 freelist_cache = kmalloc_slab(freelist_size, 0u); 1815 if (!freelist_cache) 1816 continue; 1817 1818 /* 1819 * Needed to avoid possible looping condition 1820 * in cache_grow_begin() 1821 */ 1822 if (OFF_SLAB(freelist_cache)) 1823 continue; 1824 1825 /* check if off slab has enough benefit */ 1826 if (freelist_cache->size > cachep->size / 2) 1827 continue; 1828 } 1829 1830 /* Found something acceptable - save it away */ 1831 cachep->num = num; 1832 cachep->gfporder = gfporder; 1833 left_over = remainder; 1834 1835 /* 1836 * A VFS-reclaimable slab tends to have most allocations 1837 * as GFP_NOFS and we really don't want to have to be allocating 1838 * higher-order pages when we are unable to shrink dcache. 1839 */ 1840 if (flags & SLAB_RECLAIM_ACCOUNT) 1841 break; 1842 1843 /* 1844 * Large number of objects is good, but very large slabs are 1845 * currently bad for the gfp()s. 1846 */ 1847 if (gfporder >= slab_max_order) 1848 break; 1849 1850 /* 1851 * Acceptable internal fragmentation? 1852 */ 1853 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 1854 break; 1855 } 1856 return left_over; 1857 } 1858 1859 static struct array_cache __percpu *alloc_kmem_cache_cpus( 1860 struct kmem_cache *cachep, int entries, int batchcount) 1861 { 1862 int cpu; 1863 size_t size; 1864 struct array_cache __percpu *cpu_cache; 1865 1866 size = sizeof(void *) * entries + sizeof(struct array_cache); 1867 cpu_cache = __alloc_percpu(size, sizeof(void *)); 1868 1869 if (!cpu_cache) 1870 return NULL; 1871 1872 for_each_possible_cpu(cpu) { 1873 init_arraycache(per_cpu_ptr(cpu_cache, cpu), 1874 entries, batchcount); 1875 } 1876 1877 return cpu_cache; 1878 } 1879 1880 static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) 1881 { 1882 if (slab_state >= FULL) 1883 return enable_cpucache(cachep, gfp); 1884 1885 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1); 1886 if (!cachep->cpu_cache) 1887 return 1; 1888 1889 if (slab_state == DOWN) { 1890 /* Creation of first cache (kmem_cache). */ 1891 set_up_node(kmem_cache, CACHE_CACHE); 1892 } else if (slab_state == PARTIAL) { 1893 /* For kmem_cache_node */ 1894 set_up_node(cachep, SIZE_NODE); 1895 } else { 1896 int node; 1897 1898 for_each_online_node(node) { 1899 cachep->node[node] = kmalloc_node( 1900 sizeof(struct kmem_cache_node), gfp, node); 1901 BUG_ON(!cachep->node[node]); 1902 kmem_cache_node_init(cachep->node[node]); 1903 } 1904 } 1905 1906 cachep->node[numa_mem_id()]->next_reap = 1907 jiffies + REAPTIMEOUT_NODE + 1908 ((unsigned long)cachep) % REAPTIMEOUT_NODE; 1909 1910 cpu_cache_get(cachep)->avail = 0; 1911 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 1912 cpu_cache_get(cachep)->batchcount = 1; 1913 cpu_cache_get(cachep)->touched = 0; 1914 cachep->batchcount = 1; 1915 cachep->limit = BOOT_CPUCACHE_ENTRIES; 1916 return 0; 1917 } 1918 1919 unsigned long kmem_cache_flags(unsigned long object_size, 1920 unsigned long flags, const char *name, 1921 void (*ctor)(void *)) 1922 { 1923 return flags; 1924 } 1925 1926 struct kmem_cache * 1927 __kmem_cache_alias(const char *name, size_t size, size_t align, 1928 unsigned long flags, void (*ctor)(void *)) 1929 { 1930 struct kmem_cache *cachep; 1931 1932 cachep = find_mergeable(size, align, flags, name, ctor); 1933 if (cachep) { 1934 cachep->refcount++; 1935 1936 /* 1937 * Adjust the object sizes so that we clear 1938 * the complete object on kzalloc. 1939 */ 1940 cachep->object_size = max_t(int, cachep->object_size, size); 1941 } 1942 return cachep; 1943 } 1944 1945 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep, 1946 size_t size, unsigned long flags) 1947 { 1948 size_t left; 1949 1950 cachep->num = 0; 1951 1952 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU) 1953 return false; 1954 1955 left = calculate_slab_order(cachep, size, 1956 flags | CFLGS_OBJFREELIST_SLAB); 1957 if (!cachep->num) 1958 return false; 1959 1960 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size) 1961 return false; 1962 1963 cachep->colour = left / cachep->colour_off; 1964 1965 return true; 1966 } 1967 1968 static bool set_off_slab_cache(struct kmem_cache *cachep, 1969 size_t size, unsigned long flags) 1970 { 1971 size_t left; 1972 1973 cachep->num = 0; 1974 1975 /* 1976 * Always use on-slab management when SLAB_NOLEAKTRACE 1977 * to avoid recursive calls into kmemleak. 1978 */ 1979 if (flags & SLAB_NOLEAKTRACE) 1980 return false; 1981 1982 /* 1983 * Size is large, assume best to place the slab management obj 1984 * off-slab (should allow better packing of objs). 1985 */ 1986 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB); 1987 if (!cachep->num) 1988 return false; 1989 1990 /* 1991 * If the slab has been placed off-slab, and we have enough space then 1992 * move it on-slab. This is at the expense of any extra colouring. 1993 */ 1994 if (left >= cachep->num * sizeof(freelist_idx_t)) 1995 return false; 1996 1997 cachep->colour = left / cachep->colour_off; 1998 1999 return true; 2000 } 2001 2002 static bool set_on_slab_cache(struct kmem_cache *cachep, 2003 size_t size, unsigned long flags) 2004 { 2005 size_t left; 2006 2007 cachep->num = 0; 2008 2009 left = calculate_slab_order(cachep, size, flags); 2010 if (!cachep->num) 2011 return false; 2012 2013 cachep->colour = left / cachep->colour_off; 2014 2015 return true; 2016 } 2017 2018 /** 2019 * __kmem_cache_create - Create a cache. 2020 * @cachep: cache management descriptor 2021 * @flags: SLAB flags 2022 * 2023 * Returns a ptr to the cache on success, NULL on failure. 2024 * Cannot be called within a int, but can be interrupted. 2025 * The @ctor is run when new pages are allocated by the cache. 2026 * 2027 * The flags are 2028 * 2029 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 2030 * to catch references to uninitialised memory. 2031 * 2032 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 2033 * for buffer overruns. 2034 * 2035 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 2036 * cacheline. This can be beneficial if you're counting cycles as closely 2037 * as davem. 2038 */ 2039 int 2040 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) 2041 { 2042 size_t ralign = BYTES_PER_WORD; 2043 gfp_t gfp; 2044 int err; 2045 size_t size = cachep->size; 2046 2047 #if DEBUG 2048 #if FORCED_DEBUG 2049 /* 2050 * Enable redzoning and last user accounting, except for caches with 2051 * large objects, if the increased size would increase the object size 2052 * above the next power of two: caches with object sizes just above a 2053 * power of two have a significant amount of internal fragmentation. 2054 */ 2055 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + 2056 2 * sizeof(unsigned long long))) 2057 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2058 if (!(flags & SLAB_DESTROY_BY_RCU)) 2059 flags |= SLAB_POISON; 2060 #endif 2061 #endif 2062 2063 /* 2064 * Check that size is in terms of words. This is needed to avoid 2065 * unaligned accesses for some archs when redzoning is used, and makes 2066 * sure any on-slab bufctl's are also correctly aligned. 2067 */ 2068 if (size & (BYTES_PER_WORD - 1)) { 2069 size += (BYTES_PER_WORD - 1); 2070 size &= ~(BYTES_PER_WORD - 1); 2071 } 2072 2073 if (flags & SLAB_RED_ZONE) { 2074 ralign = REDZONE_ALIGN; 2075 /* If redzoning, ensure that the second redzone is suitably 2076 * aligned, by adjusting the object size accordingly. */ 2077 size += REDZONE_ALIGN - 1; 2078 size &= ~(REDZONE_ALIGN - 1); 2079 } 2080 2081 /* 3) caller mandated alignment */ 2082 if (ralign < cachep->align) { 2083 ralign = cachep->align; 2084 } 2085 /* disable debug if necessary */ 2086 if (ralign > __alignof__(unsigned long long)) 2087 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2088 /* 2089 * 4) Store it. 2090 */ 2091 cachep->align = ralign; 2092 cachep->colour_off = cache_line_size(); 2093 /* Offset must be a multiple of the alignment. */ 2094 if (cachep->colour_off < cachep->align) 2095 cachep->colour_off = cachep->align; 2096 2097 if (slab_is_available()) 2098 gfp = GFP_KERNEL; 2099 else 2100 gfp = GFP_NOWAIT; 2101 2102 #if DEBUG 2103 2104 /* 2105 * Both debugging options require word-alignment which is calculated 2106 * into align above. 2107 */ 2108 if (flags & SLAB_RED_ZONE) { 2109 /* add space for red zone words */ 2110 cachep->obj_offset += sizeof(unsigned long long); 2111 size += 2 * sizeof(unsigned long long); 2112 } 2113 if (flags & SLAB_STORE_USER) { 2114 /* user store requires one word storage behind the end of 2115 * the real object. But if the second red zone needs to be 2116 * aligned to 64 bits, we must allow that much space. 2117 */ 2118 if (flags & SLAB_RED_ZONE) 2119 size += REDZONE_ALIGN; 2120 else 2121 size += BYTES_PER_WORD; 2122 } 2123 #endif 2124 2125 kasan_cache_create(cachep, &size, &flags); 2126 2127 size = ALIGN(size, cachep->align); 2128 /* 2129 * We should restrict the number of objects in a slab to implement 2130 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition. 2131 */ 2132 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE) 2133 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align); 2134 2135 #if DEBUG 2136 /* 2137 * To activate debug pagealloc, off-slab management is necessary 2138 * requirement. In early phase of initialization, small sized slab 2139 * doesn't get initialized so it would not be possible. So, we need 2140 * to check size >= 256. It guarantees that all necessary small 2141 * sized slab is initialized in current slab initialization sequence. 2142 */ 2143 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) && 2144 size >= 256 && cachep->object_size > cache_line_size()) { 2145 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) { 2146 size_t tmp_size = ALIGN(size, PAGE_SIZE); 2147 2148 if (set_off_slab_cache(cachep, tmp_size, flags)) { 2149 flags |= CFLGS_OFF_SLAB; 2150 cachep->obj_offset += tmp_size - size; 2151 size = tmp_size; 2152 goto done; 2153 } 2154 } 2155 } 2156 #endif 2157 2158 if (set_objfreelist_slab_cache(cachep, size, flags)) { 2159 flags |= CFLGS_OBJFREELIST_SLAB; 2160 goto done; 2161 } 2162 2163 if (set_off_slab_cache(cachep, size, flags)) { 2164 flags |= CFLGS_OFF_SLAB; 2165 goto done; 2166 } 2167 2168 if (set_on_slab_cache(cachep, size, flags)) 2169 goto done; 2170 2171 return -E2BIG; 2172 2173 done: 2174 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t); 2175 cachep->flags = flags; 2176 cachep->allocflags = __GFP_COMP; 2177 if (flags & SLAB_CACHE_DMA) 2178 cachep->allocflags |= GFP_DMA; 2179 cachep->size = size; 2180 cachep->reciprocal_buffer_size = reciprocal_value(size); 2181 2182 #if DEBUG 2183 /* 2184 * If we're going to use the generic kernel_map_pages() 2185 * poisoning, then it's going to smash the contents of 2186 * the redzone and userword anyhow, so switch them off. 2187 */ 2188 if (IS_ENABLED(CONFIG_PAGE_POISONING) && 2189 (cachep->flags & SLAB_POISON) && 2190 is_debug_pagealloc_cache(cachep)) 2191 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2192 #endif 2193 2194 if (OFF_SLAB(cachep)) { 2195 cachep->freelist_cache = 2196 kmalloc_slab(cachep->freelist_size, 0u); 2197 } 2198 2199 err = setup_cpu_cache(cachep, gfp); 2200 if (err) { 2201 __kmem_cache_release(cachep); 2202 return err; 2203 } 2204 2205 return 0; 2206 } 2207 2208 #if DEBUG 2209 static void check_irq_off(void) 2210 { 2211 BUG_ON(!irqs_disabled()); 2212 } 2213 2214 static void check_irq_on(void) 2215 { 2216 BUG_ON(irqs_disabled()); 2217 } 2218 2219 static void check_mutex_acquired(void) 2220 { 2221 BUG_ON(!mutex_is_locked(&slab_mutex)); 2222 } 2223 2224 static void check_spinlock_acquired(struct kmem_cache *cachep) 2225 { 2226 #ifdef CONFIG_SMP 2227 check_irq_off(); 2228 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock); 2229 #endif 2230 } 2231 2232 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2233 { 2234 #ifdef CONFIG_SMP 2235 check_irq_off(); 2236 assert_spin_locked(&get_node(cachep, node)->list_lock); 2237 #endif 2238 } 2239 2240 #else 2241 #define check_irq_off() do { } while(0) 2242 #define check_irq_on() do { } while(0) 2243 #define check_mutex_acquired() do { } while(0) 2244 #define check_spinlock_acquired(x) do { } while(0) 2245 #define check_spinlock_acquired_node(x, y) do { } while(0) 2246 #endif 2247 2248 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, 2249 int node, bool free_all, struct list_head *list) 2250 { 2251 int tofree; 2252 2253 if (!ac || !ac->avail) 2254 return; 2255 2256 tofree = free_all ? ac->avail : (ac->limit + 4) / 5; 2257 if (tofree > ac->avail) 2258 tofree = (ac->avail + 1) / 2; 2259 2260 free_block(cachep, ac->entry, tofree, node, list); 2261 ac->avail -= tofree; 2262 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail); 2263 } 2264 2265 static void do_drain(void *arg) 2266 { 2267 struct kmem_cache *cachep = arg; 2268 struct array_cache *ac; 2269 int node = numa_mem_id(); 2270 struct kmem_cache_node *n; 2271 LIST_HEAD(list); 2272 2273 check_irq_off(); 2274 ac = cpu_cache_get(cachep); 2275 n = get_node(cachep, node); 2276 spin_lock(&n->list_lock); 2277 free_block(cachep, ac->entry, ac->avail, node, &list); 2278 spin_unlock(&n->list_lock); 2279 slabs_destroy(cachep, &list); 2280 ac->avail = 0; 2281 } 2282 2283 static void drain_cpu_caches(struct kmem_cache *cachep) 2284 { 2285 struct kmem_cache_node *n; 2286 int node; 2287 LIST_HEAD(list); 2288 2289 on_each_cpu(do_drain, cachep, 1); 2290 check_irq_on(); 2291 for_each_kmem_cache_node(cachep, node, n) 2292 if (n->alien) 2293 drain_alien_cache(cachep, n->alien); 2294 2295 for_each_kmem_cache_node(cachep, node, n) { 2296 spin_lock_irq(&n->list_lock); 2297 drain_array_locked(cachep, n->shared, node, true, &list); 2298 spin_unlock_irq(&n->list_lock); 2299 2300 slabs_destroy(cachep, &list); 2301 } 2302 } 2303 2304 /* 2305 * Remove slabs from the list of free slabs. 2306 * Specify the number of slabs to drain in tofree. 2307 * 2308 * Returns the actual number of slabs released. 2309 */ 2310 static int drain_freelist(struct kmem_cache *cache, 2311 struct kmem_cache_node *n, int tofree) 2312 { 2313 struct list_head *p; 2314 int nr_freed; 2315 struct page *page; 2316 2317 nr_freed = 0; 2318 while (nr_freed < tofree && !list_empty(&n->slabs_free)) { 2319 2320 spin_lock_irq(&n->list_lock); 2321 p = n->slabs_free.prev; 2322 if (p == &n->slabs_free) { 2323 spin_unlock_irq(&n->list_lock); 2324 goto out; 2325 } 2326 2327 page = list_entry(p, struct page, lru); 2328 list_del(&page->lru); 2329 /* 2330 * Safe to drop the lock. The slab is no longer linked 2331 * to the cache. 2332 */ 2333 n->free_objects -= cache->num; 2334 spin_unlock_irq(&n->list_lock); 2335 slab_destroy(cache, page); 2336 nr_freed++; 2337 } 2338 out: 2339 return nr_freed; 2340 } 2341 2342 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate) 2343 { 2344 int ret = 0; 2345 int node; 2346 struct kmem_cache_node *n; 2347 2348 drain_cpu_caches(cachep); 2349 2350 check_irq_on(); 2351 for_each_kmem_cache_node(cachep, node, n) { 2352 drain_freelist(cachep, n, INT_MAX); 2353 2354 ret += !list_empty(&n->slabs_full) || 2355 !list_empty(&n->slabs_partial); 2356 } 2357 return (ret ? 1 : 0); 2358 } 2359 2360 int __kmem_cache_shutdown(struct kmem_cache *cachep) 2361 { 2362 return __kmem_cache_shrink(cachep, false); 2363 } 2364 2365 void __kmem_cache_release(struct kmem_cache *cachep) 2366 { 2367 int i; 2368 struct kmem_cache_node *n; 2369 2370 cache_random_seq_destroy(cachep); 2371 2372 free_percpu(cachep->cpu_cache); 2373 2374 /* NUMA: free the node structures */ 2375 for_each_kmem_cache_node(cachep, i, n) { 2376 kfree(n->shared); 2377 free_alien_cache(n->alien); 2378 kfree(n); 2379 cachep->node[i] = NULL; 2380 } 2381 } 2382 2383 /* 2384 * Get the memory for a slab management obj. 2385 * 2386 * For a slab cache when the slab descriptor is off-slab, the 2387 * slab descriptor can't come from the same cache which is being created, 2388 * Because if it is the case, that means we defer the creation of 2389 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point. 2390 * And we eventually call down to __kmem_cache_create(), which 2391 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one. 2392 * This is a "chicken-and-egg" problem. 2393 * 2394 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches, 2395 * which are all initialized during kmem_cache_init(). 2396 */ 2397 static void *alloc_slabmgmt(struct kmem_cache *cachep, 2398 struct page *page, int colour_off, 2399 gfp_t local_flags, int nodeid) 2400 { 2401 void *freelist; 2402 void *addr = page_address(page); 2403 2404 page->s_mem = addr + colour_off; 2405 page->active = 0; 2406 2407 if (OBJFREELIST_SLAB(cachep)) 2408 freelist = NULL; 2409 else if (OFF_SLAB(cachep)) { 2410 /* Slab management obj is off-slab. */ 2411 freelist = kmem_cache_alloc_node(cachep->freelist_cache, 2412 local_flags, nodeid); 2413 if (!freelist) 2414 return NULL; 2415 } else { 2416 /* We will use last bytes at the slab for freelist */ 2417 freelist = addr + (PAGE_SIZE << cachep->gfporder) - 2418 cachep->freelist_size; 2419 } 2420 2421 return freelist; 2422 } 2423 2424 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx) 2425 { 2426 return ((freelist_idx_t *)page->freelist)[idx]; 2427 } 2428 2429 static inline void set_free_obj(struct page *page, 2430 unsigned int idx, freelist_idx_t val) 2431 { 2432 ((freelist_idx_t *)(page->freelist))[idx] = val; 2433 } 2434 2435 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page) 2436 { 2437 #if DEBUG 2438 int i; 2439 2440 for (i = 0; i < cachep->num; i++) { 2441 void *objp = index_to_obj(cachep, page, i); 2442 2443 if (cachep->flags & SLAB_STORE_USER) 2444 *dbg_userword(cachep, objp) = NULL; 2445 2446 if (cachep->flags & SLAB_RED_ZONE) { 2447 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2448 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2449 } 2450 /* 2451 * Constructors are not allowed to allocate memory from the same 2452 * cache which they are a constructor for. Otherwise, deadlock. 2453 * They must also be threaded. 2454 */ 2455 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) { 2456 kasan_unpoison_object_data(cachep, 2457 objp + obj_offset(cachep)); 2458 cachep->ctor(objp + obj_offset(cachep)); 2459 kasan_poison_object_data( 2460 cachep, objp + obj_offset(cachep)); 2461 } 2462 2463 if (cachep->flags & SLAB_RED_ZONE) { 2464 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2465 slab_error(cachep, "constructor overwrote the end of an object"); 2466 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2467 slab_error(cachep, "constructor overwrote the start of an object"); 2468 } 2469 /* need to poison the objs? */ 2470 if (cachep->flags & SLAB_POISON) { 2471 poison_obj(cachep, objp, POISON_FREE); 2472 slab_kernel_map(cachep, objp, 0, 0); 2473 } 2474 } 2475 #endif 2476 } 2477 2478 #ifdef CONFIG_SLAB_FREELIST_RANDOM 2479 /* Hold information during a freelist initialization */ 2480 union freelist_init_state { 2481 struct { 2482 unsigned int pos; 2483 unsigned int *list; 2484 unsigned int count; 2485 unsigned int rand; 2486 }; 2487 struct rnd_state rnd_state; 2488 }; 2489 2490 /* 2491 * Initialize the state based on the randomization methode available. 2492 * return true if the pre-computed list is available, false otherwize. 2493 */ 2494 static bool freelist_state_initialize(union freelist_init_state *state, 2495 struct kmem_cache *cachep, 2496 unsigned int count) 2497 { 2498 bool ret; 2499 unsigned int rand; 2500 2501 /* Use best entropy available to define a random shift */ 2502 rand = get_random_int(); 2503 2504 /* Use a random state if the pre-computed list is not available */ 2505 if (!cachep->random_seq) { 2506 prandom_seed_state(&state->rnd_state, rand); 2507 ret = false; 2508 } else { 2509 state->list = cachep->random_seq; 2510 state->count = count; 2511 state->pos = 0; 2512 state->rand = rand; 2513 ret = true; 2514 } 2515 return ret; 2516 } 2517 2518 /* Get the next entry on the list and randomize it using a random shift */ 2519 static freelist_idx_t next_random_slot(union freelist_init_state *state) 2520 { 2521 return (state->list[state->pos++] + state->rand) % state->count; 2522 } 2523 2524 /* Swap two freelist entries */ 2525 static void swap_free_obj(struct page *page, unsigned int a, unsigned int b) 2526 { 2527 swap(((freelist_idx_t *)page->freelist)[a], 2528 ((freelist_idx_t *)page->freelist)[b]); 2529 } 2530 2531 /* 2532 * Shuffle the freelist initialization state based on pre-computed lists. 2533 * return true if the list was successfully shuffled, false otherwise. 2534 */ 2535 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page) 2536 { 2537 unsigned int objfreelist = 0, i, rand, count = cachep->num; 2538 union freelist_init_state state; 2539 bool precomputed; 2540 2541 if (count < 2) 2542 return false; 2543 2544 precomputed = freelist_state_initialize(&state, cachep, count); 2545 2546 /* Take a random entry as the objfreelist */ 2547 if (OBJFREELIST_SLAB(cachep)) { 2548 if (!precomputed) 2549 objfreelist = count - 1; 2550 else 2551 objfreelist = next_random_slot(&state); 2552 page->freelist = index_to_obj(cachep, page, objfreelist) + 2553 obj_offset(cachep); 2554 count--; 2555 } 2556 2557 /* 2558 * On early boot, generate the list dynamically. 2559 * Later use a pre-computed list for speed. 2560 */ 2561 if (!precomputed) { 2562 for (i = 0; i < count; i++) 2563 set_free_obj(page, i, i); 2564 2565 /* Fisher-Yates shuffle */ 2566 for (i = count - 1; i > 0; i--) { 2567 rand = prandom_u32_state(&state.rnd_state); 2568 rand %= (i + 1); 2569 swap_free_obj(page, i, rand); 2570 } 2571 } else { 2572 for (i = 0; i < count; i++) 2573 set_free_obj(page, i, next_random_slot(&state)); 2574 } 2575 2576 if (OBJFREELIST_SLAB(cachep)) 2577 set_free_obj(page, cachep->num - 1, objfreelist); 2578 2579 return true; 2580 } 2581 #else 2582 static inline bool shuffle_freelist(struct kmem_cache *cachep, 2583 struct page *page) 2584 { 2585 return false; 2586 } 2587 #endif /* CONFIG_SLAB_FREELIST_RANDOM */ 2588 2589 static void cache_init_objs(struct kmem_cache *cachep, 2590 struct page *page) 2591 { 2592 int i; 2593 void *objp; 2594 bool shuffled; 2595 2596 cache_init_objs_debug(cachep, page); 2597 2598 /* Try to randomize the freelist if enabled */ 2599 shuffled = shuffle_freelist(cachep, page); 2600 2601 if (!shuffled && OBJFREELIST_SLAB(cachep)) { 2602 page->freelist = index_to_obj(cachep, page, cachep->num - 1) + 2603 obj_offset(cachep); 2604 } 2605 2606 for (i = 0; i < cachep->num; i++) { 2607 objp = index_to_obj(cachep, page, i); 2608 kasan_init_slab_obj(cachep, objp); 2609 2610 /* constructor could break poison info */ 2611 if (DEBUG == 0 && cachep->ctor) { 2612 kasan_unpoison_object_data(cachep, objp); 2613 cachep->ctor(objp); 2614 kasan_poison_object_data(cachep, objp); 2615 } 2616 2617 if (!shuffled) 2618 set_free_obj(page, i, i); 2619 } 2620 } 2621 2622 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page) 2623 { 2624 void *objp; 2625 2626 objp = index_to_obj(cachep, page, get_free_obj(page, page->active)); 2627 page->active++; 2628 2629 #if DEBUG 2630 if (cachep->flags & SLAB_STORE_USER) 2631 set_store_user_dirty(cachep); 2632 #endif 2633 2634 return objp; 2635 } 2636 2637 static void slab_put_obj(struct kmem_cache *cachep, 2638 struct page *page, void *objp) 2639 { 2640 unsigned int objnr = obj_to_index(cachep, page, objp); 2641 #if DEBUG 2642 unsigned int i; 2643 2644 /* Verify double free bug */ 2645 for (i = page->active; i < cachep->num; i++) { 2646 if (get_free_obj(page, i) == objnr) { 2647 pr_err("slab: double free detected in cache '%s', objp %p\n", 2648 cachep->name, objp); 2649 BUG(); 2650 } 2651 } 2652 #endif 2653 page->active--; 2654 if (!page->freelist) 2655 page->freelist = objp + obj_offset(cachep); 2656 2657 set_free_obj(page, page->active, objnr); 2658 } 2659 2660 /* 2661 * Map pages beginning at addr to the given cache and slab. This is required 2662 * for the slab allocator to be able to lookup the cache and slab of a 2663 * virtual address for kfree, ksize, and slab debugging. 2664 */ 2665 static void slab_map_pages(struct kmem_cache *cache, struct page *page, 2666 void *freelist) 2667 { 2668 page->slab_cache = cache; 2669 page->freelist = freelist; 2670 } 2671 2672 /* 2673 * Grow (by 1) the number of slabs within a cache. This is called by 2674 * kmem_cache_alloc() when there are no active objs left in a cache. 2675 */ 2676 static struct page *cache_grow_begin(struct kmem_cache *cachep, 2677 gfp_t flags, int nodeid) 2678 { 2679 void *freelist; 2680 size_t offset; 2681 gfp_t local_flags; 2682 int page_node; 2683 struct kmem_cache_node *n; 2684 struct page *page; 2685 2686 /* 2687 * Be lazy and only check for valid flags here, keeping it out of the 2688 * critical path in kmem_cache_alloc(). 2689 */ 2690 if (unlikely(flags & GFP_SLAB_BUG_MASK)) { 2691 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK; 2692 flags &= ~GFP_SLAB_BUG_MASK; 2693 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n", 2694 invalid_mask, &invalid_mask, flags, &flags); 2695 dump_stack(); 2696 } 2697 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); 2698 2699 check_irq_off(); 2700 if (gfpflags_allow_blocking(local_flags)) 2701 local_irq_enable(); 2702 2703 /* 2704 * Get mem for the objs. Attempt to allocate a physical page from 2705 * 'nodeid'. 2706 */ 2707 page = kmem_getpages(cachep, local_flags, nodeid); 2708 if (!page) 2709 goto failed; 2710 2711 page_node = page_to_nid(page); 2712 n = get_node(cachep, page_node); 2713 2714 /* Get colour for the slab, and cal the next value. */ 2715 n->colour_next++; 2716 if (n->colour_next >= cachep->colour) 2717 n->colour_next = 0; 2718 2719 offset = n->colour_next; 2720 if (offset >= cachep->colour) 2721 offset = 0; 2722 2723 offset *= cachep->colour_off; 2724 2725 /* Get slab management. */ 2726 freelist = alloc_slabmgmt(cachep, page, offset, 2727 local_flags & ~GFP_CONSTRAINT_MASK, page_node); 2728 if (OFF_SLAB(cachep) && !freelist) 2729 goto opps1; 2730 2731 slab_map_pages(cachep, page, freelist); 2732 2733 kasan_poison_slab(page); 2734 cache_init_objs(cachep, page); 2735 2736 if (gfpflags_allow_blocking(local_flags)) 2737 local_irq_disable(); 2738 2739 return page; 2740 2741 opps1: 2742 kmem_freepages(cachep, page); 2743 failed: 2744 if (gfpflags_allow_blocking(local_flags)) 2745 local_irq_disable(); 2746 return NULL; 2747 } 2748 2749 static void cache_grow_end(struct kmem_cache *cachep, struct page *page) 2750 { 2751 struct kmem_cache_node *n; 2752 void *list = NULL; 2753 2754 check_irq_off(); 2755 2756 if (!page) 2757 return; 2758 2759 INIT_LIST_HEAD(&page->lru); 2760 n = get_node(cachep, page_to_nid(page)); 2761 2762 spin_lock(&n->list_lock); 2763 if (!page->active) 2764 list_add_tail(&page->lru, &(n->slabs_free)); 2765 else 2766 fixup_slab_list(cachep, n, page, &list); 2767 STATS_INC_GROWN(cachep); 2768 n->free_objects += cachep->num - page->active; 2769 spin_unlock(&n->list_lock); 2770 2771 fixup_objfreelist_debug(cachep, &list); 2772 } 2773 2774 #if DEBUG 2775 2776 /* 2777 * Perform extra freeing checks: 2778 * - detect bad pointers. 2779 * - POISON/RED_ZONE checking 2780 */ 2781 static void kfree_debugcheck(const void *objp) 2782 { 2783 if (!virt_addr_valid(objp)) { 2784 pr_err("kfree_debugcheck: out of range ptr %lxh\n", 2785 (unsigned long)objp); 2786 BUG(); 2787 } 2788 } 2789 2790 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2791 { 2792 unsigned long long redzone1, redzone2; 2793 2794 redzone1 = *dbg_redzone1(cache, obj); 2795 redzone2 = *dbg_redzone2(cache, obj); 2796 2797 /* 2798 * Redzone is ok. 2799 */ 2800 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2801 return; 2802 2803 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2804 slab_error(cache, "double free detected"); 2805 else 2806 slab_error(cache, "memory outside object was overwritten"); 2807 2808 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n", 2809 obj, redzone1, redzone2); 2810 } 2811 2812 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2813 unsigned long caller) 2814 { 2815 unsigned int objnr; 2816 struct page *page; 2817 2818 BUG_ON(virt_to_cache(objp) != cachep); 2819 2820 objp -= obj_offset(cachep); 2821 kfree_debugcheck(objp); 2822 page = virt_to_head_page(objp); 2823 2824 if (cachep->flags & SLAB_RED_ZONE) { 2825 verify_redzone_free(cachep, objp); 2826 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2827 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2828 } 2829 if (cachep->flags & SLAB_STORE_USER) { 2830 set_store_user_dirty(cachep); 2831 *dbg_userword(cachep, objp) = (void *)caller; 2832 } 2833 2834 objnr = obj_to_index(cachep, page, objp); 2835 2836 BUG_ON(objnr >= cachep->num); 2837 BUG_ON(objp != index_to_obj(cachep, page, objnr)); 2838 2839 if (cachep->flags & SLAB_POISON) { 2840 poison_obj(cachep, objp, POISON_FREE); 2841 slab_kernel_map(cachep, objp, 0, caller); 2842 } 2843 return objp; 2844 } 2845 2846 #else 2847 #define kfree_debugcheck(x) do { } while(0) 2848 #define cache_free_debugcheck(x,objp,z) (objp) 2849 #endif 2850 2851 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep, 2852 void **list) 2853 { 2854 #if DEBUG 2855 void *next = *list; 2856 void *objp; 2857 2858 while (next) { 2859 objp = next - obj_offset(cachep); 2860 next = *(void **)next; 2861 poison_obj(cachep, objp, POISON_FREE); 2862 } 2863 #endif 2864 } 2865 2866 static inline void fixup_slab_list(struct kmem_cache *cachep, 2867 struct kmem_cache_node *n, struct page *page, 2868 void **list) 2869 { 2870 /* move slabp to correct slabp list: */ 2871 list_del(&page->lru); 2872 if (page->active == cachep->num) { 2873 list_add(&page->lru, &n->slabs_full); 2874 if (OBJFREELIST_SLAB(cachep)) { 2875 #if DEBUG 2876 /* Poisoning will be done without holding the lock */ 2877 if (cachep->flags & SLAB_POISON) { 2878 void **objp = page->freelist; 2879 2880 *objp = *list; 2881 *list = objp; 2882 } 2883 #endif 2884 page->freelist = NULL; 2885 } 2886 } else 2887 list_add(&page->lru, &n->slabs_partial); 2888 } 2889 2890 /* Try to find non-pfmemalloc slab if needed */ 2891 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n, 2892 struct page *page, bool pfmemalloc) 2893 { 2894 if (!page) 2895 return NULL; 2896 2897 if (pfmemalloc) 2898 return page; 2899 2900 if (!PageSlabPfmemalloc(page)) 2901 return page; 2902 2903 /* No need to keep pfmemalloc slab if we have enough free objects */ 2904 if (n->free_objects > n->free_limit) { 2905 ClearPageSlabPfmemalloc(page); 2906 return page; 2907 } 2908 2909 /* Move pfmemalloc slab to the end of list to speed up next search */ 2910 list_del(&page->lru); 2911 if (!page->active) 2912 list_add_tail(&page->lru, &n->slabs_free); 2913 else 2914 list_add_tail(&page->lru, &n->slabs_partial); 2915 2916 list_for_each_entry(page, &n->slabs_partial, lru) { 2917 if (!PageSlabPfmemalloc(page)) 2918 return page; 2919 } 2920 2921 list_for_each_entry(page, &n->slabs_free, lru) { 2922 if (!PageSlabPfmemalloc(page)) 2923 return page; 2924 } 2925 2926 return NULL; 2927 } 2928 2929 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc) 2930 { 2931 struct page *page; 2932 2933 page = list_first_entry_or_null(&n->slabs_partial, 2934 struct page, lru); 2935 if (!page) { 2936 n->free_touched = 1; 2937 page = list_first_entry_or_null(&n->slabs_free, 2938 struct page, lru); 2939 } 2940 2941 if (sk_memalloc_socks()) 2942 return get_valid_first_slab(n, page, pfmemalloc); 2943 2944 return page; 2945 } 2946 2947 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep, 2948 struct kmem_cache_node *n, gfp_t flags) 2949 { 2950 struct page *page; 2951 void *obj; 2952 void *list = NULL; 2953 2954 if (!gfp_pfmemalloc_allowed(flags)) 2955 return NULL; 2956 2957 spin_lock(&n->list_lock); 2958 page = get_first_slab(n, true); 2959 if (!page) { 2960 spin_unlock(&n->list_lock); 2961 return NULL; 2962 } 2963 2964 obj = slab_get_obj(cachep, page); 2965 n->free_objects--; 2966 2967 fixup_slab_list(cachep, n, page, &list); 2968 2969 spin_unlock(&n->list_lock); 2970 fixup_objfreelist_debug(cachep, &list); 2971 2972 return obj; 2973 } 2974 2975 /* 2976 * Slab list should be fixed up by fixup_slab_list() for existing slab 2977 * or cache_grow_end() for new slab 2978 */ 2979 static __always_inline int alloc_block(struct kmem_cache *cachep, 2980 struct array_cache *ac, struct page *page, int batchcount) 2981 { 2982 /* 2983 * There must be at least one object available for 2984 * allocation. 2985 */ 2986 BUG_ON(page->active >= cachep->num); 2987 2988 while (page->active < cachep->num && batchcount--) { 2989 STATS_INC_ALLOCED(cachep); 2990 STATS_INC_ACTIVE(cachep); 2991 STATS_SET_HIGH(cachep); 2992 2993 ac->entry[ac->avail++] = slab_get_obj(cachep, page); 2994 } 2995 2996 return batchcount; 2997 } 2998 2999 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 3000 { 3001 int batchcount; 3002 struct kmem_cache_node *n; 3003 struct array_cache *ac, *shared; 3004 int node; 3005 void *list = NULL; 3006 struct page *page; 3007 3008 check_irq_off(); 3009 node = numa_mem_id(); 3010 3011 ac = cpu_cache_get(cachep); 3012 batchcount = ac->batchcount; 3013 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 3014 /* 3015 * If there was little recent activity on this cache, then 3016 * perform only a partial refill. Otherwise we could generate 3017 * refill bouncing. 3018 */ 3019 batchcount = BATCHREFILL_LIMIT; 3020 } 3021 n = get_node(cachep, node); 3022 3023 BUG_ON(ac->avail > 0 || !n); 3024 shared = READ_ONCE(n->shared); 3025 if (!n->free_objects && (!shared || !shared->avail)) 3026 goto direct_grow; 3027 3028 spin_lock(&n->list_lock); 3029 shared = READ_ONCE(n->shared); 3030 3031 /* See if we can refill from the shared array */ 3032 if (shared && transfer_objects(ac, shared, batchcount)) { 3033 shared->touched = 1; 3034 goto alloc_done; 3035 } 3036 3037 while (batchcount > 0) { 3038 /* Get slab alloc is to come from. */ 3039 page = get_first_slab(n, false); 3040 if (!page) 3041 goto must_grow; 3042 3043 check_spinlock_acquired(cachep); 3044 3045 batchcount = alloc_block(cachep, ac, page, batchcount); 3046 fixup_slab_list(cachep, n, page, &list); 3047 } 3048 3049 must_grow: 3050 n->free_objects -= ac->avail; 3051 alloc_done: 3052 spin_unlock(&n->list_lock); 3053 fixup_objfreelist_debug(cachep, &list); 3054 3055 direct_grow: 3056 if (unlikely(!ac->avail)) { 3057 /* Check if we can use obj in pfmemalloc slab */ 3058 if (sk_memalloc_socks()) { 3059 void *obj = cache_alloc_pfmemalloc(cachep, n, flags); 3060 3061 if (obj) 3062 return obj; 3063 } 3064 3065 page = cache_grow_begin(cachep, gfp_exact_node(flags), node); 3066 3067 /* 3068 * cache_grow_begin() can reenable interrupts, 3069 * then ac could change. 3070 */ 3071 ac = cpu_cache_get(cachep); 3072 if (!ac->avail && page) 3073 alloc_block(cachep, ac, page, batchcount); 3074 cache_grow_end(cachep, page); 3075 3076 if (!ac->avail) 3077 return NULL; 3078 } 3079 ac->touched = 1; 3080 3081 return ac->entry[--ac->avail]; 3082 } 3083 3084 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3085 gfp_t flags) 3086 { 3087 might_sleep_if(gfpflags_allow_blocking(flags)); 3088 } 3089 3090 #if DEBUG 3091 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3092 gfp_t flags, void *objp, unsigned long caller) 3093 { 3094 if (!objp) 3095 return objp; 3096 if (cachep->flags & SLAB_POISON) { 3097 check_poison_obj(cachep, objp); 3098 slab_kernel_map(cachep, objp, 1, 0); 3099 poison_obj(cachep, objp, POISON_INUSE); 3100 } 3101 if (cachep->flags & SLAB_STORE_USER) 3102 *dbg_userword(cachep, objp) = (void *)caller; 3103 3104 if (cachep->flags & SLAB_RED_ZONE) { 3105 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3106 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3107 slab_error(cachep, "double free, or memory outside object was overwritten"); 3108 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n", 3109 objp, *dbg_redzone1(cachep, objp), 3110 *dbg_redzone2(cachep, objp)); 3111 } 3112 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3113 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3114 } 3115 3116 objp += obj_offset(cachep); 3117 if (cachep->ctor && cachep->flags & SLAB_POISON) 3118 cachep->ctor(objp); 3119 if (ARCH_SLAB_MINALIGN && 3120 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { 3121 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3122 objp, (int)ARCH_SLAB_MINALIGN); 3123 } 3124 return objp; 3125 } 3126 #else 3127 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3128 #endif 3129 3130 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3131 { 3132 void *objp; 3133 struct array_cache *ac; 3134 3135 check_irq_off(); 3136 3137 ac = cpu_cache_get(cachep); 3138 if (likely(ac->avail)) { 3139 ac->touched = 1; 3140 objp = ac->entry[--ac->avail]; 3141 3142 STATS_INC_ALLOCHIT(cachep); 3143 goto out; 3144 } 3145 3146 STATS_INC_ALLOCMISS(cachep); 3147 objp = cache_alloc_refill(cachep, flags); 3148 /* 3149 * the 'ac' may be updated by cache_alloc_refill(), 3150 * and kmemleak_erase() requires its correct value. 3151 */ 3152 ac = cpu_cache_get(cachep); 3153 3154 out: 3155 /* 3156 * To avoid a false negative, if an object that is in one of the 3157 * per-CPU caches is leaked, we need to make sure kmemleak doesn't 3158 * treat the array pointers as a reference to the object. 3159 */ 3160 if (objp) 3161 kmemleak_erase(&ac->entry[ac->avail]); 3162 return objp; 3163 } 3164 3165 #ifdef CONFIG_NUMA 3166 /* 3167 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set. 3168 * 3169 * If we are in_interrupt, then process context, including cpusets and 3170 * mempolicy, may not apply and should not be used for allocation policy. 3171 */ 3172 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3173 { 3174 int nid_alloc, nid_here; 3175 3176 if (in_interrupt() || (flags & __GFP_THISNODE)) 3177 return NULL; 3178 nid_alloc = nid_here = numa_mem_id(); 3179 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3180 nid_alloc = cpuset_slab_spread_node(); 3181 else if (current->mempolicy) 3182 nid_alloc = mempolicy_slab_node(); 3183 if (nid_alloc != nid_here) 3184 return ____cache_alloc_node(cachep, flags, nid_alloc); 3185 return NULL; 3186 } 3187 3188 /* 3189 * Fallback function if there was no memory available and no objects on a 3190 * certain node and fall back is permitted. First we scan all the 3191 * available node for available objects. If that fails then we 3192 * perform an allocation without specifying a node. This allows the page 3193 * allocator to do its reclaim / fallback magic. We then insert the 3194 * slab into the proper nodelist and then allocate from it. 3195 */ 3196 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3197 { 3198 struct zonelist *zonelist; 3199 struct zoneref *z; 3200 struct zone *zone; 3201 enum zone_type high_zoneidx = gfp_zone(flags); 3202 void *obj = NULL; 3203 struct page *page; 3204 int nid; 3205 unsigned int cpuset_mems_cookie; 3206 3207 if (flags & __GFP_THISNODE) 3208 return NULL; 3209 3210 retry_cpuset: 3211 cpuset_mems_cookie = read_mems_allowed_begin(); 3212 zonelist = node_zonelist(mempolicy_slab_node(), flags); 3213 3214 retry: 3215 /* 3216 * Look through allowed nodes for objects available 3217 * from existing per node queues. 3218 */ 3219 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 3220 nid = zone_to_nid(zone); 3221 3222 if (cpuset_zone_allowed(zone, flags) && 3223 get_node(cache, nid) && 3224 get_node(cache, nid)->free_objects) { 3225 obj = ____cache_alloc_node(cache, 3226 gfp_exact_node(flags), nid); 3227 if (obj) 3228 break; 3229 } 3230 } 3231 3232 if (!obj) { 3233 /* 3234 * This allocation will be performed within the constraints 3235 * of the current cpuset / memory policy requirements. 3236 * We may trigger various forms of reclaim on the allowed 3237 * set and go into memory reserves if necessary. 3238 */ 3239 page = cache_grow_begin(cache, flags, numa_mem_id()); 3240 cache_grow_end(cache, page); 3241 if (page) { 3242 nid = page_to_nid(page); 3243 obj = ____cache_alloc_node(cache, 3244 gfp_exact_node(flags), nid); 3245 3246 /* 3247 * Another processor may allocate the objects in 3248 * the slab since we are not holding any locks. 3249 */ 3250 if (!obj) 3251 goto retry; 3252 } 3253 } 3254 3255 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie))) 3256 goto retry_cpuset; 3257 return obj; 3258 } 3259 3260 /* 3261 * A interface to enable slab creation on nodeid 3262 */ 3263 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3264 int nodeid) 3265 { 3266 struct page *page; 3267 struct kmem_cache_node *n; 3268 void *obj = NULL; 3269 void *list = NULL; 3270 3271 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES); 3272 n = get_node(cachep, nodeid); 3273 BUG_ON(!n); 3274 3275 check_irq_off(); 3276 spin_lock(&n->list_lock); 3277 page = get_first_slab(n, false); 3278 if (!page) 3279 goto must_grow; 3280 3281 check_spinlock_acquired_node(cachep, nodeid); 3282 3283 STATS_INC_NODEALLOCS(cachep); 3284 STATS_INC_ACTIVE(cachep); 3285 STATS_SET_HIGH(cachep); 3286 3287 BUG_ON(page->active == cachep->num); 3288 3289 obj = slab_get_obj(cachep, page); 3290 n->free_objects--; 3291 3292 fixup_slab_list(cachep, n, page, &list); 3293 3294 spin_unlock(&n->list_lock); 3295 fixup_objfreelist_debug(cachep, &list); 3296 return obj; 3297 3298 must_grow: 3299 spin_unlock(&n->list_lock); 3300 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid); 3301 if (page) { 3302 /* This slab isn't counted yet so don't update free_objects */ 3303 obj = slab_get_obj(cachep, page); 3304 } 3305 cache_grow_end(cachep, page); 3306 3307 return obj ? obj : fallback_alloc(cachep, flags); 3308 } 3309 3310 static __always_inline void * 3311 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, 3312 unsigned long caller) 3313 { 3314 unsigned long save_flags; 3315 void *ptr; 3316 int slab_node = numa_mem_id(); 3317 3318 flags &= gfp_allowed_mask; 3319 cachep = slab_pre_alloc_hook(cachep, flags); 3320 if (unlikely(!cachep)) 3321 return NULL; 3322 3323 cache_alloc_debugcheck_before(cachep, flags); 3324 local_irq_save(save_flags); 3325 3326 if (nodeid == NUMA_NO_NODE) 3327 nodeid = slab_node; 3328 3329 if (unlikely(!get_node(cachep, nodeid))) { 3330 /* Node not bootstrapped yet */ 3331 ptr = fallback_alloc(cachep, flags); 3332 goto out; 3333 } 3334 3335 if (nodeid == slab_node) { 3336 /* 3337 * Use the locally cached objects if possible. 3338 * However ____cache_alloc does not allow fallback 3339 * to other nodes. It may fail while we still have 3340 * objects on other nodes available. 3341 */ 3342 ptr = ____cache_alloc(cachep, flags); 3343 if (ptr) 3344 goto out; 3345 } 3346 /* ___cache_alloc_node can fall back to other nodes */ 3347 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3348 out: 3349 local_irq_restore(save_flags); 3350 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3351 3352 if (unlikely(flags & __GFP_ZERO) && ptr) 3353 memset(ptr, 0, cachep->object_size); 3354 3355 slab_post_alloc_hook(cachep, flags, 1, &ptr); 3356 return ptr; 3357 } 3358 3359 static __always_inline void * 3360 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) 3361 { 3362 void *objp; 3363 3364 if (current->mempolicy || cpuset_do_slab_mem_spread()) { 3365 objp = alternate_node_alloc(cache, flags); 3366 if (objp) 3367 goto out; 3368 } 3369 objp = ____cache_alloc(cache, flags); 3370 3371 /* 3372 * We may just have run out of memory on the local node. 3373 * ____cache_alloc_node() knows how to locate memory on other nodes 3374 */ 3375 if (!objp) 3376 objp = ____cache_alloc_node(cache, flags, numa_mem_id()); 3377 3378 out: 3379 return objp; 3380 } 3381 #else 3382 3383 static __always_inline void * 3384 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3385 { 3386 return ____cache_alloc(cachep, flags); 3387 } 3388 3389 #endif /* CONFIG_NUMA */ 3390 3391 static __always_inline void * 3392 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) 3393 { 3394 unsigned long save_flags; 3395 void *objp; 3396 3397 flags &= gfp_allowed_mask; 3398 cachep = slab_pre_alloc_hook(cachep, flags); 3399 if (unlikely(!cachep)) 3400 return NULL; 3401 3402 cache_alloc_debugcheck_before(cachep, flags); 3403 local_irq_save(save_flags); 3404 objp = __do_cache_alloc(cachep, flags); 3405 local_irq_restore(save_flags); 3406 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); 3407 prefetchw(objp); 3408 3409 if (unlikely(flags & __GFP_ZERO) && objp) 3410 memset(objp, 0, cachep->object_size); 3411 3412 slab_post_alloc_hook(cachep, flags, 1, &objp); 3413 return objp; 3414 } 3415 3416 /* 3417 * Caller needs to acquire correct kmem_cache_node's list_lock 3418 * @list: List of detached free slabs should be freed by caller 3419 */ 3420 static void free_block(struct kmem_cache *cachep, void **objpp, 3421 int nr_objects, int node, struct list_head *list) 3422 { 3423 int i; 3424 struct kmem_cache_node *n = get_node(cachep, node); 3425 struct page *page; 3426 3427 n->free_objects += nr_objects; 3428 3429 for (i = 0; i < nr_objects; i++) { 3430 void *objp; 3431 struct page *page; 3432 3433 objp = objpp[i]; 3434 3435 page = virt_to_head_page(objp); 3436 list_del(&page->lru); 3437 check_spinlock_acquired_node(cachep, node); 3438 slab_put_obj(cachep, page, objp); 3439 STATS_DEC_ACTIVE(cachep); 3440 3441 /* fixup slab chains */ 3442 if (page->active == 0) 3443 list_add(&page->lru, &n->slabs_free); 3444 else { 3445 /* Unconditionally move a slab to the end of the 3446 * partial list on free - maximum time for the 3447 * other objects to be freed, too. 3448 */ 3449 list_add_tail(&page->lru, &n->slabs_partial); 3450 } 3451 } 3452 3453 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) { 3454 n->free_objects -= cachep->num; 3455 3456 page = list_last_entry(&n->slabs_free, struct page, lru); 3457 list_move(&page->lru, list); 3458 } 3459 } 3460 3461 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3462 { 3463 int batchcount; 3464 struct kmem_cache_node *n; 3465 int node = numa_mem_id(); 3466 LIST_HEAD(list); 3467 3468 batchcount = ac->batchcount; 3469 3470 check_irq_off(); 3471 n = get_node(cachep, node); 3472 spin_lock(&n->list_lock); 3473 if (n->shared) { 3474 struct array_cache *shared_array = n->shared; 3475 int max = shared_array->limit - shared_array->avail; 3476 if (max) { 3477 if (batchcount > max) 3478 batchcount = max; 3479 memcpy(&(shared_array->entry[shared_array->avail]), 3480 ac->entry, sizeof(void *) * batchcount); 3481 shared_array->avail += batchcount; 3482 goto free_done; 3483 } 3484 } 3485 3486 free_block(cachep, ac->entry, batchcount, node, &list); 3487 free_done: 3488 #if STATS 3489 { 3490 int i = 0; 3491 struct page *page; 3492 3493 list_for_each_entry(page, &n->slabs_free, lru) { 3494 BUG_ON(page->active); 3495 3496 i++; 3497 } 3498 STATS_SET_FREEABLE(cachep, i); 3499 } 3500 #endif 3501 spin_unlock(&n->list_lock); 3502 slabs_destroy(cachep, &list); 3503 ac->avail -= batchcount; 3504 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3505 } 3506 3507 /* 3508 * Release an obj back to its cache. If the obj has a constructed state, it must 3509 * be in this state _before_ it is released. Called with disabled ints. 3510 */ 3511 static inline void __cache_free(struct kmem_cache *cachep, void *objp, 3512 unsigned long caller) 3513 { 3514 /* Put the object into the quarantine, don't touch it for now. */ 3515 if (kasan_slab_free(cachep, objp)) 3516 return; 3517 3518 ___cache_free(cachep, objp, caller); 3519 } 3520 3521 void ___cache_free(struct kmem_cache *cachep, void *objp, 3522 unsigned long caller) 3523 { 3524 struct array_cache *ac = cpu_cache_get(cachep); 3525 3526 check_irq_off(); 3527 kmemleak_free_recursive(objp, cachep->flags); 3528 objp = cache_free_debugcheck(cachep, objp, caller); 3529 3530 kmemcheck_slab_free(cachep, objp, cachep->object_size); 3531 3532 /* 3533 * Skip calling cache_free_alien() when the platform is not numa. 3534 * This will avoid cache misses that happen while accessing slabp (which 3535 * is per page memory reference) to get nodeid. Instead use a global 3536 * variable to skip the call, which is mostly likely to be present in 3537 * the cache. 3538 */ 3539 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) 3540 return; 3541 3542 if (ac->avail < ac->limit) { 3543 STATS_INC_FREEHIT(cachep); 3544 } else { 3545 STATS_INC_FREEMISS(cachep); 3546 cache_flusharray(cachep, ac); 3547 } 3548 3549 if (sk_memalloc_socks()) { 3550 struct page *page = virt_to_head_page(objp); 3551 3552 if (unlikely(PageSlabPfmemalloc(page))) { 3553 cache_free_pfmemalloc(cachep, page, objp); 3554 return; 3555 } 3556 } 3557 3558 ac->entry[ac->avail++] = objp; 3559 } 3560 3561 /** 3562 * kmem_cache_alloc - Allocate an object 3563 * @cachep: The cache to allocate from. 3564 * @flags: See kmalloc(). 3565 * 3566 * Allocate an object from this cache. The flags are only relevant 3567 * if the cache has no available objects. 3568 */ 3569 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3570 { 3571 void *ret = slab_alloc(cachep, flags, _RET_IP_); 3572 3573 kasan_slab_alloc(cachep, ret, flags); 3574 trace_kmem_cache_alloc(_RET_IP_, ret, 3575 cachep->object_size, cachep->size, flags); 3576 3577 return ret; 3578 } 3579 EXPORT_SYMBOL(kmem_cache_alloc); 3580 3581 static __always_inline void 3582 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags, 3583 size_t size, void **p, unsigned long caller) 3584 { 3585 size_t i; 3586 3587 for (i = 0; i < size; i++) 3588 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller); 3589 } 3590 3591 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, 3592 void **p) 3593 { 3594 size_t i; 3595 3596 s = slab_pre_alloc_hook(s, flags); 3597 if (!s) 3598 return 0; 3599 3600 cache_alloc_debugcheck_before(s, flags); 3601 3602 local_irq_disable(); 3603 for (i = 0; i < size; i++) { 3604 void *objp = __do_cache_alloc(s, flags); 3605 3606 if (unlikely(!objp)) 3607 goto error; 3608 p[i] = objp; 3609 } 3610 local_irq_enable(); 3611 3612 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_); 3613 3614 /* Clear memory outside IRQ disabled section */ 3615 if (unlikely(flags & __GFP_ZERO)) 3616 for (i = 0; i < size; i++) 3617 memset(p[i], 0, s->object_size); 3618 3619 slab_post_alloc_hook(s, flags, size, p); 3620 /* FIXME: Trace call missing. Christoph would like a bulk variant */ 3621 return size; 3622 error: 3623 local_irq_enable(); 3624 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_); 3625 slab_post_alloc_hook(s, flags, i, p); 3626 __kmem_cache_free_bulk(s, i, p); 3627 return 0; 3628 } 3629 EXPORT_SYMBOL(kmem_cache_alloc_bulk); 3630 3631 #ifdef CONFIG_TRACING 3632 void * 3633 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) 3634 { 3635 void *ret; 3636 3637 ret = slab_alloc(cachep, flags, _RET_IP_); 3638 3639 kasan_kmalloc(cachep, ret, size, flags); 3640 trace_kmalloc(_RET_IP_, ret, 3641 size, cachep->size, flags); 3642 return ret; 3643 } 3644 EXPORT_SYMBOL(kmem_cache_alloc_trace); 3645 #endif 3646 3647 #ifdef CONFIG_NUMA 3648 /** 3649 * kmem_cache_alloc_node - Allocate an object on the specified node 3650 * @cachep: The cache to allocate from. 3651 * @flags: See kmalloc(). 3652 * @nodeid: node number of the target node. 3653 * 3654 * Identical to kmem_cache_alloc but it will allocate memory on the given 3655 * node, which can improve the performance for cpu bound structures. 3656 * 3657 * Fallback to other node is possible if __GFP_THISNODE is not set. 3658 */ 3659 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3660 { 3661 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); 3662 3663 kasan_slab_alloc(cachep, ret, flags); 3664 trace_kmem_cache_alloc_node(_RET_IP_, ret, 3665 cachep->object_size, cachep->size, 3666 flags, nodeid); 3667 3668 return ret; 3669 } 3670 EXPORT_SYMBOL(kmem_cache_alloc_node); 3671 3672 #ifdef CONFIG_TRACING 3673 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, 3674 gfp_t flags, 3675 int nodeid, 3676 size_t size) 3677 { 3678 void *ret; 3679 3680 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); 3681 3682 kasan_kmalloc(cachep, ret, size, flags); 3683 trace_kmalloc_node(_RET_IP_, ret, 3684 size, cachep->size, 3685 flags, nodeid); 3686 return ret; 3687 } 3688 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 3689 #endif 3690 3691 static __always_inline void * 3692 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) 3693 { 3694 struct kmem_cache *cachep; 3695 void *ret; 3696 3697 cachep = kmalloc_slab(size, flags); 3698 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3699 return cachep; 3700 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size); 3701 kasan_kmalloc(cachep, ret, size, flags); 3702 3703 return ret; 3704 } 3705 3706 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3707 { 3708 return __do_kmalloc_node(size, flags, node, _RET_IP_); 3709 } 3710 EXPORT_SYMBOL(__kmalloc_node); 3711 3712 void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3713 int node, unsigned long caller) 3714 { 3715 return __do_kmalloc_node(size, flags, node, caller); 3716 } 3717 EXPORT_SYMBOL(__kmalloc_node_track_caller); 3718 #endif /* CONFIG_NUMA */ 3719 3720 /** 3721 * __do_kmalloc - allocate memory 3722 * @size: how many bytes of memory are required. 3723 * @flags: the type of memory to allocate (see kmalloc). 3724 * @caller: function caller for debug tracking of the caller 3725 */ 3726 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3727 unsigned long caller) 3728 { 3729 struct kmem_cache *cachep; 3730 void *ret; 3731 3732 cachep = kmalloc_slab(size, flags); 3733 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3734 return cachep; 3735 ret = slab_alloc(cachep, flags, caller); 3736 3737 kasan_kmalloc(cachep, ret, size, flags); 3738 trace_kmalloc(caller, ret, 3739 size, cachep->size, flags); 3740 3741 return ret; 3742 } 3743 3744 void *__kmalloc(size_t size, gfp_t flags) 3745 { 3746 return __do_kmalloc(size, flags, _RET_IP_); 3747 } 3748 EXPORT_SYMBOL(__kmalloc); 3749 3750 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) 3751 { 3752 return __do_kmalloc(size, flags, caller); 3753 } 3754 EXPORT_SYMBOL(__kmalloc_track_caller); 3755 3756 /** 3757 * kmem_cache_free - Deallocate an object 3758 * @cachep: The cache the allocation was from. 3759 * @objp: The previously allocated object. 3760 * 3761 * Free an object which was previously allocated from this 3762 * cache. 3763 */ 3764 void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3765 { 3766 unsigned long flags; 3767 cachep = cache_from_obj(cachep, objp); 3768 if (!cachep) 3769 return; 3770 3771 local_irq_save(flags); 3772 debug_check_no_locks_freed(objp, cachep->object_size); 3773 if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) 3774 debug_check_no_obj_freed(objp, cachep->object_size); 3775 __cache_free(cachep, objp, _RET_IP_); 3776 local_irq_restore(flags); 3777 3778 trace_kmem_cache_free(_RET_IP_, objp); 3779 } 3780 EXPORT_SYMBOL(kmem_cache_free); 3781 3782 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p) 3783 { 3784 struct kmem_cache *s; 3785 size_t i; 3786 3787 local_irq_disable(); 3788 for (i = 0; i < size; i++) { 3789 void *objp = p[i]; 3790 3791 if (!orig_s) /* called via kfree_bulk */ 3792 s = virt_to_cache(objp); 3793 else 3794 s = cache_from_obj(orig_s, objp); 3795 3796 debug_check_no_locks_freed(objp, s->object_size); 3797 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 3798 debug_check_no_obj_freed(objp, s->object_size); 3799 3800 __cache_free(s, objp, _RET_IP_); 3801 } 3802 local_irq_enable(); 3803 3804 /* FIXME: add tracing */ 3805 } 3806 EXPORT_SYMBOL(kmem_cache_free_bulk); 3807 3808 /** 3809 * kfree - free previously allocated memory 3810 * @objp: pointer returned by kmalloc. 3811 * 3812 * If @objp is NULL, no operation is performed. 3813 * 3814 * Don't free memory not originally allocated by kmalloc() 3815 * or you will run into trouble. 3816 */ 3817 void kfree(const void *objp) 3818 { 3819 struct kmem_cache *c; 3820 unsigned long flags; 3821 3822 trace_kfree(_RET_IP_, objp); 3823 3824 if (unlikely(ZERO_OR_NULL_PTR(objp))) 3825 return; 3826 local_irq_save(flags); 3827 kfree_debugcheck(objp); 3828 c = virt_to_cache(objp); 3829 debug_check_no_locks_freed(objp, c->object_size); 3830 3831 debug_check_no_obj_freed(objp, c->object_size); 3832 __cache_free(c, (void *)objp, _RET_IP_); 3833 local_irq_restore(flags); 3834 } 3835 EXPORT_SYMBOL(kfree); 3836 3837 /* 3838 * This initializes kmem_cache_node or resizes various caches for all nodes. 3839 */ 3840 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp) 3841 { 3842 int ret; 3843 int node; 3844 struct kmem_cache_node *n; 3845 3846 for_each_online_node(node) { 3847 ret = setup_kmem_cache_node(cachep, node, gfp, true); 3848 if (ret) 3849 goto fail; 3850 3851 } 3852 3853 return 0; 3854 3855 fail: 3856 if (!cachep->list.next) { 3857 /* Cache is not active yet. Roll back what we did */ 3858 node--; 3859 while (node >= 0) { 3860 n = get_node(cachep, node); 3861 if (n) { 3862 kfree(n->shared); 3863 free_alien_cache(n->alien); 3864 kfree(n); 3865 cachep->node[node] = NULL; 3866 } 3867 node--; 3868 } 3869 } 3870 return -ENOMEM; 3871 } 3872 3873 /* Always called with the slab_mutex held */ 3874 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit, 3875 int batchcount, int shared, gfp_t gfp) 3876 { 3877 struct array_cache __percpu *cpu_cache, *prev; 3878 int cpu; 3879 3880 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount); 3881 if (!cpu_cache) 3882 return -ENOMEM; 3883 3884 prev = cachep->cpu_cache; 3885 cachep->cpu_cache = cpu_cache; 3886 kick_all_cpus_sync(); 3887 3888 check_irq_on(); 3889 cachep->batchcount = batchcount; 3890 cachep->limit = limit; 3891 cachep->shared = shared; 3892 3893 if (!prev) 3894 goto setup_node; 3895 3896 for_each_online_cpu(cpu) { 3897 LIST_HEAD(list); 3898 int node; 3899 struct kmem_cache_node *n; 3900 struct array_cache *ac = per_cpu_ptr(prev, cpu); 3901 3902 node = cpu_to_mem(cpu); 3903 n = get_node(cachep, node); 3904 spin_lock_irq(&n->list_lock); 3905 free_block(cachep, ac->entry, ac->avail, node, &list); 3906 spin_unlock_irq(&n->list_lock); 3907 slabs_destroy(cachep, &list); 3908 } 3909 free_percpu(prev); 3910 3911 setup_node: 3912 return setup_kmem_cache_nodes(cachep, gfp); 3913 } 3914 3915 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3916 int batchcount, int shared, gfp_t gfp) 3917 { 3918 int ret; 3919 struct kmem_cache *c; 3920 3921 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp); 3922 3923 if (slab_state < FULL) 3924 return ret; 3925 3926 if ((ret < 0) || !is_root_cache(cachep)) 3927 return ret; 3928 3929 lockdep_assert_held(&slab_mutex); 3930 for_each_memcg_cache(c, cachep) { 3931 /* return value determined by the root cache only */ 3932 __do_tune_cpucache(c, limit, batchcount, shared, gfp); 3933 } 3934 3935 return ret; 3936 } 3937 3938 /* Called with slab_mutex held always */ 3939 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) 3940 { 3941 int err; 3942 int limit = 0; 3943 int shared = 0; 3944 int batchcount = 0; 3945 3946 err = cache_random_seq_create(cachep, cachep->num, gfp); 3947 if (err) 3948 goto end; 3949 3950 if (!is_root_cache(cachep)) { 3951 struct kmem_cache *root = memcg_root_cache(cachep); 3952 limit = root->limit; 3953 shared = root->shared; 3954 batchcount = root->batchcount; 3955 } 3956 3957 if (limit && shared && batchcount) 3958 goto skip_setup; 3959 /* 3960 * The head array serves three purposes: 3961 * - create a LIFO ordering, i.e. return objects that are cache-warm 3962 * - reduce the number of spinlock operations. 3963 * - reduce the number of linked list operations on the slab and 3964 * bufctl chains: array operations are cheaper. 3965 * The numbers are guessed, we should auto-tune as described by 3966 * Bonwick. 3967 */ 3968 if (cachep->size > 131072) 3969 limit = 1; 3970 else if (cachep->size > PAGE_SIZE) 3971 limit = 8; 3972 else if (cachep->size > 1024) 3973 limit = 24; 3974 else if (cachep->size > 256) 3975 limit = 54; 3976 else 3977 limit = 120; 3978 3979 /* 3980 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 3981 * allocation behaviour: Most allocs on one cpu, most free operations 3982 * on another cpu. For these cases, an efficient object passing between 3983 * cpus is necessary. This is provided by a shared array. The array 3984 * replaces Bonwick's magazine layer. 3985 * On uniprocessor, it's functionally equivalent (but less efficient) 3986 * to a larger limit. Thus disabled by default. 3987 */ 3988 shared = 0; 3989 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) 3990 shared = 8; 3991 3992 #if DEBUG 3993 /* 3994 * With debugging enabled, large batchcount lead to excessively long 3995 * periods with disabled local interrupts. Limit the batchcount 3996 */ 3997 if (limit > 32) 3998 limit = 32; 3999 #endif 4000 batchcount = (limit + 1) / 2; 4001 skip_setup: 4002 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp); 4003 end: 4004 if (err) 4005 pr_err("enable_cpucache failed for %s, error %d\n", 4006 cachep->name, -err); 4007 return err; 4008 } 4009 4010 /* 4011 * Drain an array if it contains any elements taking the node lock only if 4012 * necessary. Note that the node listlock also protects the array_cache 4013 * if drain_array() is used on the shared array. 4014 */ 4015 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n, 4016 struct array_cache *ac, int node) 4017 { 4018 LIST_HEAD(list); 4019 4020 /* ac from n->shared can be freed if we don't hold the slab_mutex. */ 4021 check_mutex_acquired(); 4022 4023 if (!ac || !ac->avail) 4024 return; 4025 4026 if (ac->touched) { 4027 ac->touched = 0; 4028 return; 4029 } 4030 4031 spin_lock_irq(&n->list_lock); 4032 drain_array_locked(cachep, ac, node, false, &list); 4033 spin_unlock_irq(&n->list_lock); 4034 4035 slabs_destroy(cachep, &list); 4036 } 4037 4038 /** 4039 * cache_reap - Reclaim memory from caches. 4040 * @w: work descriptor 4041 * 4042 * Called from workqueue/eventd every few seconds. 4043 * Purpose: 4044 * - clear the per-cpu caches for this CPU. 4045 * - return freeable pages to the main free memory pool. 4046 * 4047 * If we cannot acquire the cache chain mutex then just give up - we'll try 4048 * again on the next iteration. 4049 */ 4050 static void cache_reap(struct work_struct *w) 4051 { 4052 struct kmem_cache *searchp; 4053 struct kmem_cache_node *n; 4054 int node = numa_mem_id(); 4055 struct delayed_work *work = to_delayed_work(w); 4056 4057 if (!mutex_trylock(&slab_mutex)) 4058 /* Give up. Setup the next iteration. */ 4059 goto out; 4060 4061 list_for_each_entry(searchp, &slab_caches, list) { 4062 check_irq_on(); 4063 4064 /* 4065 * We only take the node lock if absolutely necessary and we 4066 * have established with reasonable certainty that 4067 * we can do some work if the lock was obtained. 4068 */ 4069 n = get_node(searchp, node); 4070 4071 reap_alien(searchp, n); 4072 4073 drain_array(searchp, n, cpu_cache_get(searchp), node); 4074 4075 /* 4076 * These are racy checks but it does not matter 4077 * if we skip one check or scan twice. 4078 */ 4079 if (time_after(n->next_reap, jiffies)) 4080 goto next; 4081 4082 n->next_reap = jiffies + REAPTIMEOUT_NODE; 4083 4084 drain_array(searchp, n, n->shared, node); 4085 4086 if (n->free_touched) 4087 n->free_touched = 0; 4088 else { 4089 int freed; 4090 4091 freed = drain_freelist(searchp, n, (n->free_limit + 4092 5 * searchp->num - 1) / (5 * searchp->num)); 4093 STATS_ADD_REAPED(searchp, freed); 4094 } 4095 next: 4096 cond_resched(); 4097 } 4098 check_irq_on(); 4099 mutex_unlock(&slab_mutex); 4100 next_reap_node(); 4101 out: 4102 /* Set up the next iteration */ 4103 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC)); 4104 } 4105 4106 #ifdef CONFIG_SLABINFO 4107 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) 4108 { 4109 struct page *page; 4110 unsigned long active_objs; 4111 unsigned long num_objs; 4112 unsigned long active_slabs = 0; 4113 unsigned long num_slabs, free_objects = 0, shared_avail = 0; 4114 const char *name; 4115 char *error = NULL; 4116 int node; 4117 struct kmem_cache_node *n; 4118 4119 active_objs = 0; 4120 num_slabs = 0; 4121 for_each_kmem_cache_node(cachep, node, n) { 4122 4123 check_irq_on(); 4124 spin_lock_irq(&n->list_lock); 4125 4126 list_for_each_entry(page, &n->slabs_full, lru) { 4127 if (page->active != cachep->num && !error) 4128 error = "slabs_full accounting error"; 4129 active_objs += cachep->num; 4130 active_slabs++; 4131 } 4132 list_for_each_entry(page, &n->slabs_partial, lru) { 4133 if (page->active == cachep->num && !error) 4134 error = "slabs_partial accounting error"; 4135 if (!page->active && !error) 4136 error = "slabs_partial accounting error"; 4137 active_objs += page->active; 4138 active_slabs++; 4139 } 4140 list_for_each_entry(page, &n->slabs_free, lru) { 4141 if (page->active && !error) 4142 error = "slabs_free accounting error"; 4143 num_slabs++; 4144 } 4145 free_objects += n->free_objects; 4146 if (n->shared) 4147 shared_avail += n->shared->avail; 4148 4149 spin_unlock_irq(&n->list_lock); 4150 } 4151 num_slabs += active_slabs; 4152 num_objs = num_slabs * cachep->num; 4153 if (num_objs - active_objs != free_objects && !error) 4154 error = "free_objects accounting error"; 4155 4156 name = cachep->name; 4157 if (error) 4158 pr_err("slab: cache %s error: %s\n", name, error); 4159 4160 sinfo->active_objs = active_objs; 4161 sinfo->num_objs = num_objs; 4162 sinfo->active_slabs = active_slabs; 4163 sinfo->num_slabs = num_slabs; 4164 sinfo->shared_avail = shared_avail; 4165 sinfo->limit = cachep->limit; 4166 sinfo->batchcount = cachep->batchcount; 4167 sinfo->shared = cachep->shared; 4168 sinfo->objects_per_slab = cachep->num; 4169 sinfo->cache_order = cachep->gfporder; 4170 } 4171 4172 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) 4173 { 4174 #if STATS 4175 { /* node stats */ 4176 unsigned long high = cachep->high_mark; 4177 unsigned long allocs = cachep->num_allocations; 4178 unsigned long grown = cachep->grown; 4179 unsigned long reaped = cachep->reaped; 4180 unsigned long errors = cachep->errors; 4181 unsigned long max_freeable = cachep->max_freeable; 4182 unsigned long node_allocs = cachep->node_allocs; 4183 unsigned long node_frees = cachep->node_frees; 4184 unsigned long overflows = cachep->node_overflow; 4185 4186 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu", 4187 allocs, high, grown, 4188 reaped, errors, max_freeable, node_allocs, 4189 node_frees, overflows); 4190 } 4191 /* cpu stats */ 4192 { 4193 unsigned long allochit = atomic_read(&cachep->allochit); 4194 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4195 unsigned long freehit = atomic_read(&cachep->freehit); 4196 unsigned long freemiss = atomic_read(&cachep->freemiss); 4197 4198 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4199 allochit, allocmiss, freehit, freemiss); 4200 } 4201 #endif 4202 } 4203 4204 #define MAX_SLABINFO_WRITE 128 4205 /** 4206 * slabinfo_write - Tuning for the slab allocator 4207 * @file: unused 4208 * @buffer: user buffer 4209 * @count: data length 4210 * @ppos: unused 4211 */ 4212 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 4213 size_t count, loff_t *ppos) 4214 { 4215 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4216 int limit, batchcount, shared, res; 4217 struct kmem_cache *cachep; 4218 4219 if (count > MAX_SLABINFO_WRITE) 4220 return -EINVAL; 4221 if (copy_from_user(&kbuf, buffer, count)) 4222 return -EFAULT; 4223 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4224 4225 tmp = strchr(kbuf, ' '); 4226 if (!tmp) 4227 return -EINVAL; 4228 *tmp = '\0'; 4229 tmp++; 4230 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4231 return -EINVAL; 4232 4233 /* Find the cache in the chain of caches. */ 4234 mutex_lock(&slab_mutex); 4235 res = -EINVAL; 4236 list_for_each_entry(cachep, &slab_caches, list) { 4237 if (!strcmp(cachep->name, kbuf)) { 4238 if (limit < 1 || batchcount < 1 || 4239 batchcount > limit || shared < 0) { 4240 res = 0; 4241 } else { 4242 res = do_tune_cpucache(cachep, limit, 4243 batchcount, shared, 4244 GFP_KERNEL); 4245 } 4246 break; 4247 } 4248 } 4249 mutex_unlock(&slab_mutex); 4250 if (res >= 0) 4251 res = count; 4252 return res; 4253 } 4254 4255 #ifdef CONFIG_DEBUG_SLAB_LEAK 4256 4257 static inline int add_caller(unsigned long *n, unsigned long v) 4258 { 4259 unsigned long *p; 4260 int l; 4261 if (!v) 4262 return 1; 4263 l = n[1]; 4264 p = n + 2; 4265 while (l) { 4266 int i = l/2; 4267 unsigned long *q = p + 2 * i; 4268 if (*q == v) { 4269 q[1]++; 4270 return 1; 4271 } 4272 if (*q > v) { 4273 l = i; 4274 } else { 4275 p = q + 2; 4276 l -= i + 1; 4277 } 4278 } 4279 if (++n[1] == n[0]) 4280 return 0; 4281 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4282 p[0] = v; 4283 p[1] = 1; 4284 return 1; 4285 } 4286 4287 static void handle_slab(unsigned long *n, struct kmem_cache *c, 4288 struct page *page) 4289 { 4290 void *p; 4291 int i, j; 4292 unsigned long v; 4293 4294 if (n[0] == n[1]) 4295 return; 4296 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) { 4297 bool active = true; 4298 4299 for (j = page->active; j < c->num; j++) { 4300 if (get_free_obj(page, j) == i) { 4301 active = false; 4302 break; 4303 } 4304 } 4305 4306 if (!active) 4307 continue; 4308 4309 /* 4310 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table 4311 * mapping is established when actual object allocation and 4312 * we could mistakenly access the unmapped object in the cpu 4313 * cache. 4314 */ 4315 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v))) 4316 continue; 4317 4318 if (!add_caller(n, v)) 4319 return; 4320 } 4321 } 4322 4323 static void show_symbol(struct seq_file *m, unsigned long address) 4324 { 4325 #ifdef CONFIG_KALLSYMS 4326 unsigned long offset, size; 4327 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; 4328 4329 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { 4330 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4331 if (modname[0]) 4332 seq_printf(m, " [%s]", modname); 4333 return; 4334 } 4335 #endif 4336 seq_printf(m, "%p", (void *)address); 4337 } 4338 4339 static int leaks_show(struct seq_file *m, void *p) 4340 { 4341 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); 4342 struct page *page; 4343 struct kmem_cache_node *n; 4344 const char *name; 4345 unsigned long *x = m->private; 4346 int node; 4347 int i; 4348 4349 if (!(cachep->flags & SLAB_STORE_USER)) 4350 return 0; 4351 if (!(cachep->flags & SLAB_RED_ZONE)) 4352 return 0; 4353 4354 /* 4355 * Set store_user_clean and start to grab stored user information 4356 * for all objects on this cache. If some alloc/free requests comes 4357 * during the processing, information would be wrong so restart 4358 * whole processing. 4359 */ 4360 do { 4361 set_store_user_clean(cachep); 4362 drain_cpu_caches(cachep); 4363 4364 x[1] = 0; 4365 4366 for_each_kmem_cache_node(cachep, node, n) { 4367 4368 check_irq_on(); 4369 spin_lock_irq(&n->list_lock); 4370 4371 list_for_each_entry(page, &n->slabs_full, lru) 4372 handle_slab(x, cachep, page); 4373 list_for_each_entry(page, &n->slabs_partial, lru) 4374 handle_slab(x, cachep, page); 4375 spin_unlock_irq(&n->list_lock); 4376 } 4377 } while (!is_store_user_clean(cachep)); 4378 4379 name = cachep->name; 4380 if (x[0] == x[1]) { 4381 /* Increase the buffer size */ 4382 mutex_unlock(&slab_mutex); 4383 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL); 4384 if (!m->private) { 4385 /* Too bad, we are really out */ 4386 m->private = x; 4387 mutex_lock(&slab_mutex); 4388 return -ENOMEM; 4389 } 4390 *(unsigned long *)m->private = x[0] * 2; 4391 kfree(x); 4392 mutex_lock(&slab_mutex); 4393 /* Now make sure this entry will be retried */ 4394 m->count = m->size; 4395 return 0; 4396 } 4397 for (i = 0; i < x[1]; i++) { 4398 seq_printf(m, "%s: %lu ", name, x[2*i+3]); 4399 show_symbol(m, x[2*i+2]); 4400 seq_putc(m, '\n'); 4401 } 4402 4403 return 0; 4404 } 4405 4406 static const struct seq_operations slabstats_op = { 4407 .start = slab_start, 4408 .next = slab_next, 4409 .stop = slab_stop, 4410 .show = leaks_show, 4411 }; 4412 4413 static int slabstats_open(struct inode *inode, struct file *file) 4414 { 4415 unsigned long *n; 4416 4417 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE); 4418 if (!n) 4419 return -ENOMEM; 4420 4421 *n = PAGE_SIZE / (2 * sizeof(unsigned long)); 4422 4423 return 0; 4424 } 4425 4426 static const struct file_operations proc_slabstats_operations = { 4427 .open = slabstats_open, 4428 .read = seq_read, 4429 .llseek = seq_lseek, 4430 .release = seq_release_private, 4431 }; 4432 #endif 4433 4434 static int __init slab_proc_init(void) 4435 { 4436 #ifdef CONFIG_DEBUG_SLAB_LEAK 4437 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); 4438 #endif 4439 return 0; 4440 } 4441 module_init(slab_proc_init); 4442 #endif 4443 4444 #ifdef CONFIG_HARDENED_USERCOPY 4445 /* 4446 * Rejects objects that are incorrectly sized. 4447 * 4448 * Returns NULL if check passes, otherwise const char * to name of cache 4449 * to indicate an error. 4450 */ 4451 const char *__check_heap_object(const void *ptr, unsigned long n, 4452 struct page *page) 4453 { 4454 struct kmem_cache *cachep; 4455 unsigned int objnr; 4456 unsigned long offset; 4457 4458 /* Find and validate object. */ 4459 cachep = page->slab_cache; 4460 objnr = obj_to_index(cachep, page, (void *)ptr); 4461 BUG_ON(objnr >= cachep->num); 4462 4463 /* Find offset within object. */ 4464 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep); 4465 4466 /* Allow address range falling entirely within object size. */ 4467 if (offset <= cachep->object_size && n <= cachep->object_size - offset) 4468 return NULL; 4469 4470 return cachep->name; 4471 } 4472 #endif /* CONFIG_HARDENED_USERCOPY */ 4473 4474 /** 4475 * ksize - get the actual amount of memory allocated for a given object 4476 * @objp: Pointer to the object 4477 * 4478 * kmalloc may internally round up allocations and return more memory 4479 * than requested. ksize() can be used to determine the actual amount of 4480 * memory allocated. The caller may use this additional memory, even though 4481 * a smaller amount of memory was initially specified with the kmalloc call. 4482 * The caller must guarantee that objp points to a valid object previously 4483 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4484 * must not be freed during the duration of the call. 4485 */ 4486 size_t ksize(const void *objp) 4487 { 4488 size_t size; 4489 4490 BUG_ON(!objp); 4491 if (unlikely(objp == ZERO_SIZE_PTR)) 4492 return 0; 4493 4494 size = virt_to_cache(objp)->object_size; 4495 /* We assume that ksize callers could use the whole allocated area, 4496 * so we need to unpoison this area. 4497 */ 4498 kasan_unpoison_shadow(objp, size); 4499 4500 return size; 4501 } 4502 EXPORT_SYMBOL(ksize); 4503