1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * linux/mm/slab.c 4 * Written by Mark Hemment, 1996/97. 5 * (markhe@nextd.demon.co.uk) 6 * 7 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli 8 * 9 * Major cleanup, different bufctl logic, per-cpu arrays 10 * (c) 2000 Manfred Spraul 11 * 12 * Cleanup, make the head arrays unconditional, preparation for NUMA 13 * (c) 2002 Manfred Spraul 14 * 15 * An implementation of the Slab Allocator as described in outline in; 16 * UNIX Internals: The New Frontiers by Uresh Vahalia 17 * Pub: Prentice Hall ISBN 0-13-101908-2 18 * or with a little more detail in; 19 * The Slab Allocator: An Object-Caching Kernel Memory Allocator 20 * Jeff Bonwick (Sun Microsystems). 21 * Presented at: USENIX Summer 1994 Technical Conference 22 * 23 * The memory is organized in caches, one cache for each object type. 24 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) 25 * Each cache consists out of many slabs (they are small (usually one 26 * page long) and always contiguous), and each slab contains multiple 27 * initialized objects. 28 * 29 * This means, that your constructor is used only for newly allocated 30 * slabs and you must pass objects with the same initializations to 31 * kmem_cache_free. 32 * 33 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, 34 * normal). If you need a special memory type, then must create a new 35 * cache for that memory type. 36 * 37 * In order to reduce fragmentation, the slabs are sorted in 3 groups: 38 * full slabs with 0 free objects 39 * partial slabs 40 * empty slabs with no allocated objects 41 * 42 * If partial slabs exist, then new allocations come from these slabs, 43 * otherwise from empty slabs or new slabs are allocated. 44 * 45 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache 46 * during kmem_cache_destroy(). The caller must prevent concurrent allocs. 47 * 48 * Each cache has a short per-cpu head array, most allocs 49 * and frees go into that array, and if that array overflows, then 1/2 50 * of the entries in the array are given back into the global cache. 51 * The head array is strictly LIFO and should improve the cache hit rates. 52 * On SMP, it additionally reduces the spinlock operations. 53 * 54 * The c_cpuarray may not be read with enabled local interrupts - 55 * it's changed with a smp_call_function(). 56 * 57 * SMP synchronization: 58 * constructors and destructors are called without any locking. 59 * Several members in struct kmem_cache and struct slab never change, they 60 * are accessed without any locking. 61 * The per-cpu arrays are never accessed from the wrong cpu, no locking, 62 * and local interrupts are disabled so slab code is preempt-safe. 63 * The non-constant members are protected with a per-cache irq spinlock. 64 * 65 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch 66 * in 2000 - many ideas in the current implementation are derived from 67 * his patch. 68 * 69 * Further notes from the original documentation: 70 * 71 * 11 April '97. Started multi-threading - markhe 72 * The global cache-chain is protected by the mutex 'slab_mutex'. 73 * The sem is only needed when accessing/extending the cache-chain, which 74 * can never happen inside an interrupt (kmem_cache_create(), 75 * kmem_cache_shrink() and kmem_cache_reap()). 76 * 77 * At present, each engine can be growing a cache. This should be blocked. 78 * 79 * 15 March 2005. NUMA slab allocator. 80 * Shai Fultheim <shai@scalex86.org>. 81 * Shobhit Dayal <shobhit@calsoftinc.com> 82 * Alok N Kataria <alokk@calsoftinc.com> 83 * Christoph Lameter <christoph@lameter.com> 84 * 85 * Modified the slab allocator to be node aware on NUMA systems. 86 * Each node has its own list of partial, free and full slabs. 87 * All object allocations for a node occur from node specific slab lists. 88 */ 89 90 #include <linux/slab.h> 91 #include <linux/mm.h> 92 #include <linux/poison.h> 93 #include <linux/swap.h> 94 #include <linux/cache.h> 95 #include <linux/interrupt.h> 96 #include <linux/init.h> 97 #include <linux/compiler.h> 98 #include <linux/cpuset.h> 99 #include <linux/proc_fs.h> 100 #include <linux/seq_file.h> 101 #include <linux/notifier.h> 102 #include <linux/kallsyms.h> 103 #include <linux/cpu.h> 104 #include <linux/sysctl.h> 105 #include <linux/module.h> 106 #include <linux/rcupdate.h> 107 #include <linux/string.h> 108 #include <linux/uaccess.h> 109 #include <linux/nodemask.h> 110 #include <linux/kmemleak.h> 111 #include <linux/mempolicy.h> 112 #include <linux/mutex.h> 113 #include <linux/fault-inject.h> 114 #include <linux/rtmutex.h> 115 #include <linux/reciprocal_div.h> 116 #include <linux/debugobjects.h> 117 #include <linux/memory.h> 118 #include <linux/prefetch.h> 119 #include <linux/sched/task_stack.h> 120 121 #include <net/sock.h> 122 123 #include <asm/cacheflush.h> 124 #include <asm/tlbflush.h> 125 #include <asm/page.h> 126 127 #include <trace/events/kmem.h> 128 129 #include "internal.h" 130 131 #include "slab.h" 132 133 /* 134 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. 135 * 0 for faster, smaller code (especially in the critical paths). 136 * 137 * STATS - 1 to collect stats for /proc/slabinfo. 138 * 0 for faster, smaller code (especially in the critical paths). 139 * 140 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) 141 */ 142 143 #ifdef CONFIG_DEBUG_SLAB 144 #define DEBUG 1 145 #define STATS 1 146 #define FORCED_DEBUG 1 147 #else 148 #define DEBUG 0 149 #define STATS 0 150 #define FORCED_DEBUG 0 151 #endif 152 153 /* Shouldn't this be in a header file somewhere? */ 154 #define BYTES_PER_WORD sizeof(void *) 155 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) 156 157 #ifndef ARCH_KMALLOC_FLAGS 158 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN 159 #endif 160 161 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \ 162 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0) 163 164 #if FREELIST_BYTE_INDEX 165 typedef unsigned char freelist_idx_t; 166 #else 167 typedef unsigned short freelist_idx_t; 168 #endif 169 170 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1) 171 172 /* 173 * struct array_cache 174 * 175 * Purpose: 176 * - LIFO ordering, to hand out cache-warm objects from _alloc 177 * - reduce the number of linked list operations 178 * - reduce spinlock operations 179 * 180 * The limit is stored in the per-cpu structure to reduce the data cache 181 * footprint. 182 * 183 */ 184 struct array_cache { 185 unsigned int avail; 186 unsigned int limit; 187 unsigned int batchcount; 188 unsigned int touched; 189 void *entry[]; /* 190 * Must have this definition in here for the proper 191 * alignment of array_cache. Also simplifies accessing 192 * the entries. 193 */ 194 }; 195 196 struct alien_cache { 197 spinlock_t lock; 198 struct array_cache ac; 199 }; 200 201 /* 202 * Need this for bootstrapping a per node allocator. 203 */ 204 #define NUM_INIT_LISTS (2 * MAX_NUMNODES) 205 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS]; 206 #define CACHE_CACHE 0 207 #define SIZE_NODE (MAX_NUMNODES) 208 209 static int drain_freelist(struct kmem_cache *cache, 210 struct kmem_cache_node *n, int tofree); 211 static void free_block(struct kmem_cache *cachep, void **objpp, int len, 212 int node, struct list_head *list); 213 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list); 214 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); 215 static void cache_reap(struct work_struct *unused); 216 217 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep, 218 void **list); 219 static inline void fixup_slab_list(struct kmem_cache *cachep, 220 struct kmem_cache_node *n, struct page *page, 221 void **list); 222 static int slab_early_init = 1; 223 224 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node)) 225 226 static void kmem_cache_node_init(struct kmem_cache_node *parent) 227 { 228 INIT_LIST_HEAD(&parent->slabs_full); 229 INIT_LIST_HEAD(&parent->slabs_partial); 230 INIT_LIST_HEAD(&parent->slabs_free); 231 parent->total_slabs = 0; 232 parent->free_slabs = 0; 233 parent->shared = NULL; 234 parent->alien = NULL; 235 parent->colour_next = 0; 236 spin_lock_init(&parent->list_lock); 237 parent->free_objects = 0; 238 parent->free_touched = 0; 239 } 240 241 #define MAKE_LIST(cachep, listp, slab, nodeid) \ 242 do { \ 243 INIT_LIST_HEAD(listp); \ 244 list_splice(&get_node(cachep, nodeid)->slab, listp); \ 245 } while (0) 246 247 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ 248 do { \ 249 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ 250 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ 251 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ 252 } while (0) 253 254 #define CFLGS_OBJFREELIST_SLAB ((slab_flags_t __force)0x40000000U) 255 #define CFLGS_OFF_SLAB ((slab_flags_t __force)0x80000000U) 256 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB) 257 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) 258 259 #define BATCHREFILL_LIMIT 16 260 /* 261 * Optimization question: fewer reaps means less probability for unnessary 262 * cpucache drain/refill cycles. 263 * 264 * OTOH the cpuarrays can contain lots of objects, 265 * which could lock up otherwise freeable slabs. 266 */ 267 #define REAPTIMEOUT_AC (2*HZ) 268 #define REAPTIMEOUT_NODE (4*HZ) 269 270 #if STATS 271 #define STATS_INC_ACTIVE(x) ((x)->num_active++) 272 #define STATS_DEC_ACTIVE(x) ((x)->num_active--) 273 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) 274 #define STATS_INC_GROWN(x) ((x)->grown++) 275 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) 276 #define STATS_SET_HIGH(x) \ 277 do { \ 278 if ((x)->num_active > (x)->high_mark) \ 279 (x)->high_mark = (x)->num_active; \ 280 } while (0) 281 #define STATS_INC_ERR(x) ((x)->errors++) 282 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) 283 #define STATS_INC_NODEFREES(x) ((x)->node_frees++) 284 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) 285 #define STATS_SET_FREEABLE(x, i) \ 286 do { \ 287 if ((x)->max_freeable < i) \ 288 (x)->max_freeable = i; \ 289 } while (0) 290 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) 291 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) 292 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) 293 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) 294 #else 295 #define STATS_INC_ACTIVE(x) do { } while (0) 296 #define STATS_DEC_ACTIVE(x) do { } while (0) 297 #define STATS_INC_ALLOCED(x) do { } while (0) 298 #define STATS_INC_GROWN(x) do { } while (0) 299 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) 300 #define STATS_SET_HIGH(x) do { } while (0) 301 #define STATS_INC_ERR(x) do { } while (0) 302 #define STATS_INC_NODEALLOCS(x) do { } while (0) 303 #define STATS_INC_NODEFREES(x) do { } while (0) 304 #define STATS_INC_ACOVERFLOW(x) do { } while (0) 305 #define STATS_SET_FREEABLE(x, i) do { } while (0) 306 #define STATS_INC_ALLOCHIT(x) do { } while (0) 307 #define STATS_INC_ALLOCMISS(x) do { } while (0) 308 #define STATS_INC_FREEHIT(x) do { } while (0) 309 #define STATS_INC_FREEMISS(x) do { } while (0) 310 #endif 311 312 #if DEBUG 313 314 /* 315 * memory layout of objects: 316 * 0 : objp 317 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that 318 * the end of an object is aligned with the end of the real 319 * allocation. Catches writes behind the end of the allocation. 320 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: 321 * redzone word. 322 * cachep->obj_offset: The real object. 323 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] 324 * cachep->size - 1* BYTES_PER_WORD: last caller address 325 * [BYTES_PER_WORD long] 326 */ 327 static int obj_offset(struct kmem_cache *cachep) 328 { 329 return cachep->obj_offset; 330 } 331 332 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) 333 { 334 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 335 return (unsigned long long*) (objp + obj_offset(cachep) - 336 sizeof(unsigned long long)); 337 } 338 339 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) 340 { 341 BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); 342 if (cachep->flags & SLAB_STORE_USER) 343 return (unsigned long long *)(objp + cachep->size - 344 sizeof(unsigned long long) - 345 REDZONE_ALIGN); 346 return (unsigned long long *) (objp + cachep->size - 347 sizeof(unsigned long long)); 348 } 349 350 static void **dbg_userword(struct kmem_cache *cachep, void *objp) 351 { 352 BUG_ON(!(cachep->flags & SLAB_STORE_USER)); 353 return (void **)(objp + cachep->size - BYTES_PER_WORD); 354 } 355 356 #else 357 358 #define obj_offset(x) 0 359 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 360 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) 361 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) 362 363 #endif 364 365 #ifdef CONFIG_DEBUG_SLAB_LEAK 366 367 static inline bool is_store_user_clean(struct kmem_cache *cachep) 368 { 369 return atomic_read(&cachep->store_user_clean) == 1; 370 } 371 372 static inline void set_store_user_clean(struct kmem_cache *cachep) 373 { 374 atomic_set(&cachep->store_user_clean, 1); 375 } 376 377 static inline void set_store_user_dirty(struct kmem_cache *cachep) 378 { 379 if (is_store_user_clean(cachep)) 380 atomic_set(&cachep->store_user_clean, 0); 381 } 382 383 #else 384 static inline void set_store_user_dirty(struct kmem_cache *cachep) {} 385 386 #endif 387 388 /* 389 * Do not go above this order unless 0 objects fit into the slab or 390 * overridden on the command line. 391 */ 392 #define SLAB_MAX_ORDER_HI 1 393 #define SLAB_MAX_ORDER_LO 0 394 static int slab_max_order = SLAB_MAX_ORDER_LO; 395 static bool slab_max_order_set __initdata; 396 397 static inline struct kmem_cache *virt_to_cache(const void *obj) 398 { 399 struct page *page = virt_to_head_page(obj); 400 return page->slab_cache; 401 } 402 403 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page, 404 unsigned int idx) 405 { 406 return page->s_mem + cache->size * idx; 407 } 408 409 #define BOOT_CPUCACHE_ENTRIES 1 410 /* internal cache of cache description objs */ 411 static struct kmem_cache kmem_cache_boot = { 412 .batchcount = 1, 413 .limit = BOOT_CPUCACHE_ENTRIES, 414 .shared = 1, 415 .size = sizeof(struct kmem_cache), 416 .name = "kmem_cache", 417 }; 418 419 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); 420 421 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) 422 { 423 return this_cpu_ptr(cachep->cpu_cache); 424 } 425 426 /* 427 * Calculate the number of objects and left-over bytes for a given buffer size. 428 */ 429 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size, 430 slab_flags_t flags, size_t *left_over) 431 { 432 unsigned int num; 433 size_t slab_size = PAGE_SIZE << gfporder; 434 435 /* 436 * The slab management structure can be either off the slab or 437 * on it. For the latter case, the memory allocated for a 438 * slab is used for: 439 * 440 * - @buffer_size bytes for each object 441 * - One freelist_idx_t for each object 442 * 443 * We don't need to consider alignment of freelist because 444 * freelist will be at the end of slab page. The objects will be 445 * at the correct alignment. 446 * 447 * If the slab management structure is off the slab, then the 448 * alignment will already be calculated into the size. Because 449 * the slabs are all pages aligned, the objects will be at the 450 * correct alignment when allocated. 451 */ 452 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) { 453 num = slab_size / buffer_size; 454 *left_over = slab_size % buffer_size; 455 } else { 456 num = slab_size / (buffer_size + sizeof(freelist_idx_t)); 457 *left_over = slab_size % 458 (buffer_size + sizeof(freelist_idx_t)); 459 } 460 461 return num; 462 } 463 464 #if DEBUG 465 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) 466 467 static void __slab_error(const char *function, struct kmem_cache *cachep, 468 char *msg) 469 { 470 pr_err("slab error in %s(): cache `%s': %s\n", 471 function, cachep->name, msg); 472 dump_stack(); 473 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 474 } 475 #endif 476 477 /* 478 * By default on NUMA we use alien caches to stage the freeing of 479 * objects allocated from other nodes. This causes massive memory 480 * inefficiencies when using fake NUMA setup to split memory into a 481 * large number of small nodes, so it can be disabled on the command 482 * line 483 */ 484 485 static int use_alien_caches __read_mostly = 1; 486 static int __init noaliencache_setup(char *s) 487 { 488 use_alien_caches = 0; 489 return 1; 490 } 491 __setup("noaliencache", noaliencache_setup); 492 493 static int __init slab_max_order_setup(char *str) 494 { 495 get_option(&str, &slab_max_order); 496 slab_max_order = slab_max_order < 0 ? 0 : 497 min(slab_max_order, MAX_ORDER - 1); 498 slab_max_order_set = true; 499 500 return 1; 501 } 502 __setup("slab_max_order=", slab_max_order_setup); 503 504 #ifdef CONFIG_NUMA 505 /* 506 * Special reaping functions for NUMA systems called from cache_reap(). 507 * These take care of doing round robin flushing of alien caches (containing 508 * objects freed on different nodes from which they were allocated) and the 509 * flushing of remote pcps by calling drain_node_pages. 510 */ 511 static DEFINE_PER_CPU(unsigned long, slab_reap_node); 512 513 static void init_reap_node(int cpu) 514 { 515 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu), 516 node_online_map); 517 } 518 519 static void next_reap_node(void) 520 { 521 int node = __this_cpu_read(slab_reap_node); 522 523 node = next_node_in(node, node_online_map); 524 __this_cpu_write(slab_reap_node, node); 525 } 526 527 #else 528 #define init_reap_node(cpu) do { } while (0) 529 #define next_reap_node(void) do { } while (0) 530 #endif 531 532 /* 533 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz 534 * via the workqueue/eventd. 535 * Add the CPU number into the expiration time to minimize the possibility of 536 * the CPUs getting into lockstep and contending for the global cache chain 537 * lock. 538 */ 539 static void start_cpu_timer(int cpu) 540 { 541 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); 542 543 if (reap_work->work.func == NULL) { 544 init_reap_node(cpu); 545 INIT_DEFERRABLE_WORK(reap_work, cache_reap); 546 schedule_delayed_work_on(cpu, reap_work, 547 __round_jiffies_relative(HZ, cpu)); 548 } 549 } 550 551 static void init_arraycache(struct array_cache *ac, int limit, int batch) 552 { 553 if (ac) { 554 ac->avail = 0; 555 ac->limit = limit; 556 ac->batchcount = batch; 557 ac->touched = 0; 558 } 559 } 560 561 static struct array_cache *alloc_arraycache(int node, int entries, 562 int batchcount, gfp_t gfp) 563 { 564 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache); 565 struct array_cache *ac = NULL; 566 567 ac = kmalloc_node(memsize, gfp, node); 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 init_arraycache(ac, entries, batchcount); 577 return ac; 578 } 579 580 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep, 581 struct page *page, void *objp) 582 { 583 struct kmem_cache_node *n; 584 int page_node; 585 LIST_HEAD(list); 586 587 page_node = page_to_nid(page); 588 n = get_node(cachep, page_node); 589 590 spin_lock(&n->list_lock); 591 free_block(cachep, &objp, 1, page_node, &list); 592 spin_unlock(&n->list_lock); 593 594 slabs_destroy(cachep, &list); 595 } 596 597 /* 598 * Transfer objects in one arraycache to another. 599 * Locking must be handled by the caller. 600 * 601 * Return the number of entries transferred. 602 */ 603 static int transfer_objects(struct array_cache *to, 604 struct array_cache *from, unsigned int max) 605 { 606 /* Figure out how many entries to transfer */ 607 int nr = min3(from->avail, max, to->limit - to->avail); 608 609 if (!nr) 610 return 0; 611 612 memcpy(to->entry + to->avail, from->entry + from->avail -nr, 613 sizeof(void *) *nr); 614 615 from->avail -= nr; 616 to->avail += nr; 617 return nr; 618 } 619 620 #ifndef CONFIG_NUMA 621 622 #define drain_alien_cache(cachep, alien) do { } while (0) 623 #define reap_alien(cachep, n) do { } while (0) 624 625 static inline struct alien_cache **alloc_alien_cache(int node, 626 int limit, gfp_t gfp) 627 { 628 return NULL; 629 } 630 631 static inline void free_alien_cache(struct alien_cache **ac_ptr) 632 { 633 } 634 635 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 636 { 637 return 0; 638 } 639 640 static inline void *alternate_node_alloc(struct kmem_cache *cachep, 641 gfp_t flags) 642 { 643 return NULL; 644 } 645 646 static inline void *____cache_alloc_node(struct kmem_cache *cachep, 647 gfp_t flags, int nodeid) 648 { 649 return NULL; 650 } 651 652 static inline gfp_t gfp_exact_node(gfp_t flags) 653 { 654 return flags & ~__GFP_NOFAIL; 655 } 656 657 #else /* CONFIG_NUMA */ 658 659 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); 660 static void *alternate_node_alloc(struct kmem_cache *, gfp_t); 661 662 static struct alien_cache *__alloc_alien_cache(int node, int entries, 663 int batch, gfp_t gfp) 664 { 665 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache); 666 struct alien_cache *alc = NULL; 667 668 alc = kmalloc_node(memsize, gfp, node); 669 if (alc) { 670 kmemleak_no_scan(alc); 671 init_arraycache(&alc->ac, entries, batch); 672 spin_lock_init(&alc->lock); 673 } 674 return alc; 675 } 676 677 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) 678 { 679 struct alien_cache **alc_ptr; 680 int i; 681 682 if (limit > 1) 683 limit = 12; 684 alc_ptr = kcalloc_node(nr_node_ids, sizeof(void *), gfp, node); 685 if (!alc_ptr) 686 return NULL; 687 688 for_each_node(i) { 689 if (i == node || !node_online(i)) 690 continue; 691 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp); 692 if (!alc_ptr[i]) { 693 for (i--; i >= 0; i--) 694 kfree(alc_ptr[i]); 695 kfree(alc_ptr); 696 return NULL; 697 } 698 } 699 return alc_ptr; 700 } 701 702 static void free_alien_cache(struct alien_cache **alc_ptr) 703 { 704 int i; 705 706 if (!alc_ptr) 707 return; 708 for_each_node(i) 709 kfree(alc_ptr[i]); 710 kfree(alc_ptr); 711 } 712 713 static void __drain_alien_cache(struct kmem_cache *cachep, 714 struct array_cache *ac, int node, 715 struct list_head *list) 716 { 717 struct kmem_cache_node *n = get_node(cachep, node); 718 719 if (ac->avail) { 720 spin_lock(&n->list_lock); 721 /* 722 * Stuff objects into the remote nodes shared array first. 723 * That way we could avoid the overhead of putting the objects 724 * into the free lists and getting them back later. 725 */ 726 if (n->shared) 727 transfer_objects(n->shared, ac, ac->limit); 728 729 free_block(cachep, ac->entry, ac->avail, node, list); 730 ac->avail = 0; 731 spin_unlock(&n->list_lock); 732 } 733 } 734 735 /* 736 * Called from cache_reap() to regularly drain alien caches round robin. 737 */ 738 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n) 739 { 740 int node = __this_cpu_read(slab_reap_node); 741 742 if (n->alien) { 743 struct alien_cache *alc = n->alien[node]; 744 struct array_cache *ac; 745 746 if (alc) { 747 ac = &alc->ac; 748 if (ac->avail && spin_trylock_irq(&alc->lock)) { 749 LIST_HEAD(list); 750 751 __drain_alien_cache(cachep, ac, node, &list); 752 spin_unlock_irq(&alc->lock); 753 slabs_destroy(cachep, &list); 754 } 755 } 756 } 757 } 758 759 static void drain_alien_cache(struct kmem_cache *cachep, 760 struct alien_cache **alien) 761 { 762 int i = 0; 763 struct alien_cache *alc; 764 struct array_cache *ac; 765 unsigned long flags; 766 767 for_each_online_node(i) { 768 alc = alien[i]; 769 if (alc) { 770 LIST_HEAD(list); 771 772 ac = &alc->ac; 773 spin_lock_irqsave(&alc->lock, flags); 774 __drain_alien_cache(cachep, ac, i, &list); 775 spin_unlock_irqrestore(&alc->lock, flags); 776 slabs_destroy(cachep, &list); 777 } 778 } 779 } 780 781 static int __cache_free_alien(struct kmem_cache *cachep, void *objp, 782 int node, int page_node) 783 { 784 struct kmem_cache_node *n; 785 struct alien_cache *alien = NULL; 786 struct array_cache *ac; 787 LIST_HEAD(list); 788 789 n = get_node(cachep, node); 790 STATS_INC_NODEFREES(cachep); 791 if (n->alien && n->alien[page_node]) { 792 alien = n->alien[page_node]; 793 ac = &alien->ac; 794 spin_lock(&alien->lock); 795 if (unlikely(ac->avail == ac->limit)) { 796 STATS_INC_ACOVERFLOW(cachep); 797 __drain_alien_cache(cachep, ac, page_node, &list); 798 } 799 ac->entry[ac->avail++] = objp; 800 spin_unlock(&alien->lock); 801 slabs_destroy(cachep, &list); 802 } else { 803 n = get_node(cachep, page_node); 804 spin_lock(&n->list_lock); 805 free_block(cachep, &objp, 1, page_node, &list); 806 spin_unlock(&n->list_lock); 807 slabs_destroy(cachep, &list); 808 } 809 return 1; 810 } 811 812 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) 813 { 814 int page_node = page_to_nid(virt_to_page(objp)); 815 int node = numa_mem_id(); 816 /* 817 * Make sure we are not freeing a object from another node to the array 818 * cache on this cpu. 819 */ 820 if (likely(node == page_node)) 821 return 0; 822 823 return __cache_free_alien(cachep, objp, node, page_node); 824 } 825 826 /* 827 * Construct gfp mask to allocate from a specific node but do not reclaim or 828 * warn about failures. 829 */ 830 static inline gfp_t gfp_exact_node(gfp_t flags) 831 { 832 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL); 833 } 834 #endif 835 836 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp) 837 { 838 struct kmem_cache_node *n; 839 840 /* 841 * Set up the kmem_cache_node for cpu before we can 842 * begin anything. Make sure some other cpu on this 843 * node has not already allocated this 844 */ 845 n = get_node(cachep, node); 846 if (n) { 847 spin_lock_irq(&n->list_lock); 848 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount + 849 cachep->num; 850 spin_unlock_irq(&n->list_lock); 851 852 return 0; 853 } 854 855 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node); 856 if (!n) 857 return -ENOMEM; 858 859 kmem_cache_node_init(n); 860 n->next_reap = jiffies + REAPTIMEOUT_NODE + 861 ((unsigned long)cachep) % REAPTIMEOUT_NODE; 862 863 n->free_limit = 864 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num; 865 866 /* 867 * The kmem_cache_nodes don't come and go as CPUs 868 * come and go. slab_mutex is sufficient 869 * protection here. 870 */ 871 cachep->node[node] = n; 872 873 return 0; 874 } 875 876 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP) 877 /* 878 * Allocates and initializes node for a node on each slab cache, used for 879 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node 880 * will be allocated off-node since memory is not yet online for the new node. 881 * When hotplugging memory or a cpu, existing node are not replaced if 882 * already in use. 883 * 884 * Must hold slab_mutex. 885 */ 886 static int init_cache_node_node(int node) 887 { 888 int ret; 889 struct kmem_cache *cachep; 890 891 list_for_each_entry(cachep, &slab_caches, list) { 892 ret = init_cache_node(cachep, node, GFP_KERNEL); 893 if (ret) 894 return ret; 895 } 896 897 return 0; 898 } 899 #endif 900 901 static int setup_kmem_cache_node(struct kmem_cache *cachep, 902 int node, gfp_t gfp, bool force_change) 903 { 904 int ret = -ENOMEM; 905 struct kmem_cache_node *n; 906 struct array_cache *old_shared = NULL; 907 struct array_cache *new_shared = NULL; 908 struct alien_cache **new_alien = NULL; 909 LIST_HEAD(list); 910 911 if (use_alien_caches) { 912 new_alien = alloc_alien_cache(node, cachep->limit, gfp); 913 if (!new_alien) 914 goto fail; 915 } 916 917 if (cachep->shared) { 918 new_shared = alloc_arraycache(node, 919 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp); 920 if (!new_shared) 921 goto fail; 922 } 923 924 ret = init_cache_node(cachep, node, gfp); 925 if (ret) 926 goto fail; 927 928 n = get_node(cachep, node); 929 spin_lock_irq(&n->list_lock); 930 if (n->shared && force_change) { 931 free_block(cachep, n->shared->entry, 932 n->shared->avail, node, &list); 933 n->shared->avail = 0; 934 } 935 936 if (!n->shared || force_change) { 937 old_shared = n->shared; 938 n->shared = new_shared; 939 new_shared = NULL; 940 } 941 942 if (!n->alien) { 943 n->alien = new_alien; 944 new_alien = NULL; 945 } 946 947 spin_unlock_irq(&n->list_lock); 948 slabs_destroy(cachep, &list); 949 950 /* 951 * To protect lockless access to n->shared during irq disabled context. 952 * If n->shared isn't NULL in irq disabled context, accessing to it is 953 * guaranteed to be valid until irq is re-enabled, because it will be 954 * freed after synchronize_rcu(). 955 */ 956 if (old_shared && force_change) 957 synchronize_rcu(); 958 959 fail: 960 kfree(old_shared); 961 kfree(new_shared); 962 free_alien_cache(new_alien); 963 964 return ret; 965 } 966 967 #ifdef CONFIG_SMP 968 969 static void cpuup_canceled(long cpu) 970 { 971 struct kmem_cache *cachep; 972 struct kmem_cache_node *n = NULL; 973 int node = cpu_to_mem(cpu); 974 const struct cpumask *mask = cpumask_of_node(node); 975 976 list_for_each_entry(cachep, &slab_caches, list) { 977 struct array_cache *nc; 978 struct array_cache *shared; 979 struct alien_cache **alien; 980 LIST_HEAD(list); 981 982 n = get_node(cachep, node); 983 if (!n) 984 continue; 985 986 spin_lock_irq(&n->list_lock); 987 988 /* Free limit for this kmem_cache_node */ 989 n->free_limit -= cachep->batchcount; 990 991 /* cpu is dead; no one can alloc from it. */ 992 nc = per_cpu_ptr(cachep->cpu_cache, cpu); 993 if (nc) { 994 free_block(cachep, nc->entry, nc->avail, node, &list); 995 nc->avail = 0; 996 } 997 998 if (!cpumask_empty(mask)) { 999 spin_unlock_irq(&n->list_lock); 1000 goto free_slab; 1001 } 1002 1003 shared = n->shared; 1004 if (shared) { 1005 free_block(cachep, shared->entry, 1006 shared->avail, node, &list); 1007 n->shared = NULL; 1008 } 1009 1010 alien = n->alien; 1011 n->alien = NULL; 1012 1013 spin_unlock_irq(&n->list_lock); 1014 1015 kfree(shared); 1016 if (alien) { 1017 drain_alien_cache(cachep, alien); 1018 free_alien_cache(alien); 1019 } 1020 1021 free_slab: 1022 slabs_destroy(cachep, &list); 1023 } 1024 /* 1025 * In the previous loop, all the objects were freed to 1026 * the respective cache's slabs, now we can go ahead and 1027 * shrink each nodelist to its limit. 1028 */ 1029 list_for_each_entry(cachep, &slab_caches, list) { 1030 n = get_node(cachep, node); 1031 if (!n) 1032 continue; 1033 drain_freelist(cachep, n, INT_MAX); 1034 } 1035 } 1036 1037 static int cpuup_prepare(long cpu) 1038 { 1039 struct kmem_cache *cachep; 1040 int node = cpu_to_mem(cpu); 1041 int err; 1042 1043 /* 1044 * We need to do this right in the beginning since 1045 * alloc_arraycache's are going to use this list. 1046 * kmalloc_node allows us to add the slab to the right 1047 * kmem_cache_node and not this cpu's kmem_cache_node 1048 */ 1049 err = init_cache_node_node(node); 1050 if (err < 0) 1051 goto bad; 1052 1053 /* 1054 * Now we can go ahead with allocating the shared arrays and 1055 * array caches 1056 */ 1057 list_for_each_entry(cachep, &slab_caches, list) { 1058 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false); 1059 if (err) 1060 goto bad; 1061 } 1062 1063 return 0; 1064 bad: 1065 cpuup_canceled(cpu); 1066 return -ENOMEM; 1067 } 1068 1069 int slab_prepare_cpu(unsigned int cpu) 1070 { 1071 int err; 1072 1073 mutex_lock(&slab_mutex); 1074 err = cpuup_prepare(cpu); 1075 mutex_unlock(&slab_mutex); 1076 return err; 1077 } 1078 1079 /* 1080 * This is called for a failed online attempt and for a successful 1081 * offline. 1082 * 1083 * Even if all the cpus of a node are down, we don't free the 1084 * kmem_list3 of any cache. This to avoid a race between cpu_down, and 1085 * a kmalloc allocation from another cpu for memory from the node of 1086 * the cpu going down. The list3 structure is usually allocated from 1087 * kmem_cache_create() and gets destroyed at kmem_cache_destroy(). 1088 */ 1089 int slab_dead_cpu(unsigned int cpu) 1090 { 1091 mutex_lock(&slab_mutex); 1092 cpuup_canceled(cpu); 1093 mutex_unlock(&slab_mutex); 1094 return 0; 1095 } 1096 #endif 1097 1098 static int slab_online_cpu(unsigned int cpu) 1099 { 1100 start_cpu_timer(cpu); 1101 return 0; 1102 } 1103 1104 static int slab_offline_cpu(unsigned int cpu) 1105 { 1106 /* 1107 * Shutdown cache reaper. Note that the slab_mutex is held so 1108 * that if cache_reap() is invoked it cannot do anything 1109 * expensive but will only modify reap_work and reschedule the 1110 * timer. 1111 */ 1112 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); 1113 /* Now the cache_reaper is guaranteed to be not running. */ 1114 per_cpu(slab_reap_work, cpu).work.func = NULL; 1115 return 0; 1116 } 1117 1118 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 1119 /* 1120 * Drains freelist for a node on each slab cache, used for memory hot-remove. 1121 * Returns -EBUSY if all objects cannot be drained so that the node is not 1122 * removed. 1123 * 1124 * Must hold slab_mutex. 1125 */ 1126 static int __meminit drain_cache_node_node(int node) 1127 { 1128 struct kmem_cache *cachep; 1129 int ret = 0; 1130 1131 list_for_each_entry(cachep, &slab_caches, list) { 1132 struct kmem_cache_node *n; 1133 1134 n = get_node(cachep, node); 1135 if (!n) 1136 continue; 1137 1138 drain_freelist(cachep, n, INT_MAX); 1139 1140 if (!list_empty(&n->slabs_full) || 1141 !list_empty(&n->slabs_partial)) { 1142 ret = -EBUSY; 1143 break; 1144 } 1145 } 1146 return ret; 1147 } 1148 1149 static int __meminit slab_memory_callback(struct notifier_block *self, 1150 unsigned long action, void *arg) 1151 { 1152 struct memory_notify *mnb = arg; 1153 int ret = 0; 1154 int nid; 1155 1156 nid = mnb->status_change_nid; 1157 if (nid < 0) 1158 goto out; 1159 1160 switch (action) { 1161 case MEM_GOING_ONLINE: 1162 mutex_lock(&slab_mutex); 1163 ret = init_cache_node_node(nid); 1164 mutex_unlock(&slab_mutex); 1165 break; 1166 case MEM_GOING_OFFLINE: 1167 mutex_lock(&slab_mutex); 1168 ret = drain_cache_node_node(nid); 1169 mutex_unlock(&slab_mutex); 1170 break; 1171 case MEM_ONLINE: 1172 case MEM_OFFLINE: 1173 case MEM_CANCEL_ONLINE: 1174 case MEM_CANCEL_OFFLINE: 1175 break; 1176 } 1177 out: 1178 return notifier_from_errno(ret); 1179 } 1180 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ 1181 1182 /* 1183 * swap the static kmem_cache_node with kmalloced memory 1184 */ 1185 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list, 1186 int nodeid) 1187 { 1188 struct kmem_cache_node *ptr; 1189 1190 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid); 1191 BUG_ON(!ptr); 1192 1193 memcpy(ptr, list, sizeof(struct kmem_cache_node)); 1194 /* 1195 * Do not assume that spinlocks can be initialized via memcpy: 1196 */ 1197 spin_lock_init(&ptr->list_lock); 1198 1199 MAKE_ALL_LISTS(cachep, ptr, nodeid); 1200 cachep->node[nodeid] = ptr; 1201 } 1202 1203 /* 1204 * For setting up all the kmem_cache_node for cache whose buffer_size is same as 1205 * size of kmem_cache_node. 1206 */ 1207 static void __init set_up_node(struct kmem_cache *cachep, int index) 1208 { 1209 int node; 1210 1211 for_each_online_node(node) { 1212 cachep->node[node] = &init_kmem_cache_node[index + node]; 1213 cachep->node[node]->next_reap = jiffies + 1214 REAPTIMEOUT_NODE + 1215 ((unsigned long)cachep) % REAPTIMEOUT_NODE; 1216 } 1217 } 1218 1219 /* 1220 * Initialisation. Called after the page allocator have been initialised and 1221 * before smp_init(). 1222 */ 1223 void __init kmem_cache_init(void) 1224 { 1225 int i; 1226 1227 kmem_cache = &kmem_cache_boot; 1228 1229 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1) 1230 use_alien_caches = 0; 1231 1232 for (i = 0; i < NUM_INIT_LISTS; i++) 1233 kmem_cache_node_init(&init_kmem_cache_node[i]); 1234 1235 /* 1236 * Fragmentation resistance on low memory - only use bigger 1237 * page orders on machines with more than 32MB of memory if 1238 * not overridden on the command line. 1239 */ 1240 if (!slab_max_order_set && totalram_pages() > (32 << 20) >> PAGE_SHIFT) 1241 slab_max_order = SLAB_MAX_ORDER_HI; 1242 1243 /* Bootstrap is tricky, because several objects are allocated 1244 * from caches that do not exist yet: 1245 * 1) initialize the kmem_cache cache: it contains the struct 1246 * kmem_cache structures of all caches, except kmem_cache itself: 1247 * kmem_cache is statically allocated. 1248 * Initially an __init data area is used for the head array and the 1249 * kmem_cache_node structures, it's replaced with a kmalloc allocated 1250 * array at the end of the bootstrap. 1251 * 2) Create the first kmalloc cache. 1252 * The struct kmem_cache for the new cache is allocated normally. 1253 * An __init data area is used for the head array. 1254 * 3) Create the remaining kmalloc caches, with minimally sized 1255 * head arrays. 1256 * 4) Replace the __init data head arrays for kmem_cache and the first 1257 * kmalloc cache with kmalloc allocated arrays. 1258 * 5) Replace the __init data for kmem_cache_node for kmem_cache and 1259 * the other cache's with kmalloc allocated memory. 1260 * 6) Resize the head arrays of the kmalloc caches to their final sizes. 1261 */ 1262 1263 /* 1) create the kmem_cache */ 1264 1265 /* 1266 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids 1267 */ 1268 create_boot_cache(kmem_cache, "kmem_cache", 1269 offsetof(struct kmem_cache, node) + 1270 nr_node_ids * sizeof(struct kmem_cache_node *), 1271 SLAB_HWCACHE_ALIGN, 0, 0); 1272 list_add(&kmem_cache->list, &slab_caches); 1273 memcg_link_cache(kmem_cache); 1274 slab_state = PARTIAL; 1275 1276 /* 1277 * Initialize the caches that provide memory for the kmem_cache_node 1278 * structures first. Without this, further allocations will bug. 1279 */ 1280 kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE] = create_kmalloc_cache( 1281 kmalloc_info[INDEX_NODE].name, 1282 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS, 1283 0, kmalloc_size(INDEX_NODE)); 1284 slab_state = PARTIAL_NODE; 1285 setup_kmalloc_cache_index_table(); 1286 1287 slab_early_init = 0; 1288 1289 /* 5) Replace the bootstrap kmem_cache_node */ 1290 { 1291 int nid; 1292 1293 for_each_online_node(nid) { 1294 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid); 1295 1296 init_list(kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE], 1297 &init_kmem_cache_node[SIZE_NODE + nid], nid); 1298 } 1299 } 1300 1301 create_kmalloc_caches(ARCH_KMALLOC_FLAGS); 1302 } 1303 1304 void __init kmem_cache_init_late(void) 1305 { 1306 struct kmem_cache *cachep; 1307 1308 /* 6) resize the head arrays to their final sizes */ 1309 mutex_lock(&slab_mutex); 1310 list_for_each_entry(cachep, &slab_caches, list) 1311 if (enable_cpucache(cachep, GFP_NOWAIT)) 1312 BUG(); 1313 mutex_unlock(&slab_mutex); 1314 1315 /* Done! */ 1316 slab_state = FULL; 1317 1318 #ifdef CONFIG_NUMA 1319 /* 1320 * Register a memory hotplug callback that initializes and frees 1321 * node. 1322 */ 1323 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 1324 #endif 1325 1326 /* 1327 * The reap timers are started later, with a module init call: That part 1328 * of the kernel is not yet operational. 1329 */ 1330 } 1331 1332 static int __init cpucache_init(void) 1333 { 1334 int ret; 1335 1336 /* 1337 * Register the timers that return unneeded pages to the page allocator 1338 */ 1339 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online", 1340 slab_online_cpu, slab_offline_cpu); 1341 WARN_ON(ret < 0); 1342 1343 return 0; 1344 } 1345 __initcall(cpucache_init); 1346 1347 static noinline void 1348 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) 1349 { 1350 #if DEBUG 1351 struct kmem_cache_node *n; 1352 unsigned long flags; 1353 int node; 1354 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL, 1355 DEFAULT_RATELIMIT_BURST); 1356 1357 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs)) 1358 return; 1359 1360 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", 1361 nodeid, gfpflags, &gfpflags); 1362 pr_warn(" cache: %s, object size: %d, order: %d\n", 1363 cachep->name, cachep->size, cachep->gfporder); 1364 1365 for_each_kmem_cache_node(cachep, node, n) { 1366 unsigned long total_slabs, free_slabs, free_objs; 1367 1368 spin_lock_irqsave(&n->list_lock, flags); 1369 total_slabs = n->total_slabs; 1370 free_slabs = n->free_slabs; 1371 free_objs = n->free_objects; 1372 spin_unlock_irqrestore(&n->list_lock, flags); 1373 1374 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n", 1375 node, total_slabs - free_slabs, total_slabs, 1376 (total_slabs * cachep->num) - free_objs, 1377 total_slabs * cachep->num); 1378 } 1379 #endif 1380 } 1381 1382 /* 1383 * Interface to system's page allocator. No need to hold the 1384 * kmem_cache_node ->list_lock. 1385 * 1386 * If we requested dmaable memory, we will get it. Even if we 1387 * did not request dmaable memory, we might get it, but that 1388 * would be relatively rare and ignorable. 1389 */ 1390 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, 1391 int nodeid) 1392 { 1393 struct page *page; 1394 int nr_pages; 1395 1396 flags |= cachep->allocflags; 1397 1398 page = __alloc_pages_node(nodeid, flags, cachep->gfporder); 1399 if (!page) { 1400 slab_out_of_memory(cachep, flags, nodeid); 1401 return NULL; 1402 } 1403 1404 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) { 1405 __free_pages(page, cachep->gfporder); 1406 return NULL; 1407 } 1408 1409 nr_pages = (1 << cachep->gfporder); 1410 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1411 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, nr_pages); 1412 else 1413 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, nr_pages); 1414 1415 __SetPageSlab(page); 1416 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ 1417 if (sk_memalloc_socks() && page_is_pfmemalloc(page)) 1418 SetPageSlabPfmemalloc(page); 1419 1420 return page; 1421 } 1422 1423 /* 1424 * Interface to system's page release. 1425 */ 1426 static void kmem_freepages(struct kmem_cache *cachep, struct page *page) 1427 { 1428 int order = cachep->gfporder; 1429 unsigned long nr_freed = (1 << order); 1430 1431 if (cachep->flags & SLAB_RECLAIM_ACCOUNT) 1432 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, -nr_freed); 1433 else 1434 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, -nr_freed); 1435 1436 BUG_ON(!PageSlab(page)); 1437 __ClearPageSlabPfmemalloc(page); 1438 __ClearPageSlab(page); 1439 page_mapcount_reset(page); 1440 page->mapping = NULL; 1441 1442 if (current->reclaim_state) 1443 current->reclaim_state->reclaimed_slab += nr_freed; 1444 memcg_uncharge_slab(page, order, cachep); 1445 __free_pages(page, order); 1446 } 1447 1448 static void kmem_rcu_free(struct rcu_head *head) 1449 { 1450 struct kmem_cache *cachep; 1451 struct page *page; 1452 1453 page = container_of(head, struct page, rcu_head); 1454 cachep = page->slab_cache; 1455 1456 kmem_freepages(cachep, page); 1457 } 1458 1459 #if DEBUG 1460 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep) 1461 { 1462 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) && 1463 (cachep->size % PAGE_SIZE) == 0) 1464 return true; 1465 1466 return false; 1467 } 1468 1469 #ifdef CONFIG_DEBUG_PAGEALLOC 1470 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, 1471 unsigned long caller) 1472 { 1473 int size = cachep->object_size; 1474 1475 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; 1476 1477 if (size < 5 * sizeof(unsigned long)) 1478 return; 1479 1480 *addr++ = 0x12345678; 1481 *addr++ = caller; 1482 *addr++ = smp_processor_id(); 1483 size -= 3 * sizeof(unsigned long); 1484 { 1485 unsigned long *sptr = &caller; 1486 unsigned long svalue; 1487 1488 while (!kstack_end(sptr)) { 1489 svalue = *sptr++; 1490 if (kernel_text_address(svalue)) { 1491 *addr++ = svalue; 1492 size -= sizeof(unsigned long); 1493 if (size <= sizeof(unsigned long)) 1494 break; 1495 } 1496 } 1497 1498 } 1499 *addr++ = 0x87654321; 1500 } 1501 1502 static void slab_kernel_map(struct kmem_cache *cachep, void *objp, 1503 int map, unsigned long caller) 1504 { 1505 if (!is_debug_pagealloc_cache(cachep)) 1506 return; 1507 1508 if (caller) 1509 store_stackinfo(cachep, objp, caller); 1510 1511 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map); 1512 } 1513 1514 #else 1515 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp, 1516 int map, unsigned long caller) {} 1517 1518 #endif 1519 1520 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) 1521 { 1522 int size = cachep->object_size; 1523 addr = &((char *)addr)[obj_offset(cachep)]; 1524 1525 memset(addr, val, size); 1526 *(unsigned char *)(addr + size - 1) = POISON_END; 1527 } 1528 1529 static void dump_line(char *data, int offset, int limit) 1530 { 1531 int i; 1532 unsigned char error = 0; 1533 int bad_count = 0; 1534 1535 pr_err("%03x: ", offset); 1536 for (i = 0; i < limit; i++) { 1537 if (data[offset + i] != POISON_FREE) { 1538 error = data[offset + i]; 1539 bad_count++; 1540 } 1541 } 1542 print_hex_dump(KERN_CONT, "", 0, 16, 1, 1543 &data[offset], limit, 1); 1544 1545 if (bad_count == 1) { 1546 error ^= POISON_FREE; 1547 if (!(error & (error - 1))) { 1548 pr_err("Single bit error detected. Probably bad RAM.\n"); 1549 #ifdef CONFIG_X86 1550 pr_err("Run memtest86+ or a similar memory test tool.\n"); 1551 #else 1552 pr_err("Run a memory test tool.\n"); 1553 #endif 1554 } 1555 } 1556 } 1557 #endif 1558 1559 #if DEBUG 1560 1561 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) 1562 { 1563 int i, size; 1564 char *realobj; 1565 1566 if (cachep->flags & SLAB_RED_ZONE) { 1567 pr_err("Redzone: 0x%llx/0x%llx\n", 1568 *dbg_redzone1(cachep, objp), 1569 *dbg_redzone2(cachep, objp)); 1570 } 1571 1572 if (cachep->flags & SLAB_STORE_USER) 1573 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp)); 1574 realobj = (char *)objp + obj_offset(cachep); 1575 size = cachep->object_size; 1576 for (i = 0; i < size && lines; i += 16, lines--) { 1577 int limit; 1578 limit = 16; 1579 if (i + limit > size) 1580 limit = size - i; 1581 dump_line(realobj, i, limit); 1582 } 1583 } 1584 1585 static void check_poison_obj(struct kmem_cache *cachep, void *objp) 1586 { 1587 char *realobj; 1588 int size, i; 1589 int lines = 0; 1590 1591 if (is_debug_pagealloc_cache(cachep)) 1592 return; 1593 1594 realobj = (char *)objp + obj_offset(cachep); 1595 size = cachep->object_size; 1596 1597 for (i = 0; i < size; i++) { 1598 char exp = POISON_FREE; 1599 if (i == size - 1) 1600 exp = POISON_END; 1601 if (realobj[i] != exp) { 1602 int limit; 1603 /* Mismatch ! */ 1604 /* Print header */ 1605 if (lines == 0) { 1606 pr_err("Slab corruption (%s): %s start=%px, len=%d\n", 1607 print_tainted(), cachep->name, 1608 realobj, size); 1609 print_objinfo(cachep, objp, 0); 1610 } 1611 /* Hexdump the affected line */ 1612 i = (i / 16) * 16; 1613 limit = 16; 1614 if (i + limit > size) 1615 limit = size - i; 1616 dump_line(realobj, i, limit); 1617 i += 16; 1618 lines++; 1619 /* Limit to 5 lines */ 1620 if (lines > 5) 1621 break; 1622 } 1623 } 1624 if (lines != 0) { 1625 /* Print some data about the neighboring objects, if they 1626 * exist: 1627 */ 1628 struct page *page = virt_to_head_page(objp); 1629 unsigned int objnr; 1630 1631 objnr = obj_to_index(cachep, page, objp); 1632 if (objnr) { 1633 objp = index_to_obj(cachep, page, objnr - 1); 1634 realobj = (char *)objp + obj_offset(cachep); 1635 pr_err("Prev obj: start=%px, len=%d\n", realobj, size); 1636 print_objinfo(cachep, objp, 2); 1637 } 1638 if (objnr + 1 < cachep->num) { 1639 objp = index_to_obj(cachep, page, objnr + 1); 1640 realobj = (char *)objp + obj_offset(cachep); 1641 pr_err("Next obj: start=%px, len=%d\n", realobj, size); 1642 print_objinfo(cachep, objp, 2); 1643 } 1644 } 1645 } 1646 #endif 1647 1648 #if DEBUG 1649 static void slab_destroy_debugcheck(struct kmem_cache *cachep, 1650 struct page *page) 1651 { 1652 int i; 1653 1654 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) { 1655 poison_obj(cachep, page->freelist - obj_offset(cachep), 1656 POISON_FREE); 1657 } 1658 1659 for (i = 0; i < cachep->num; i++) { 1660 void *objp = index_to_obj(cachep, page, i); 1661 1662 if (cachep->flags & SLAB_POISON) { 1663 check_poison_obj(cachep, objp); 1664 slab_kernel_map(cachep, objp, 1, 0); 1665 } 1666 if (cachep->flags & SLAB_RED_ZONE) { 1667 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 1668 slab_error(cachep, "start of a freed object was overwritten"); 1669 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 1670 slab_error(cachep, "end of a freed object was overwritten"); 1671 } 1672 } 1673 } 1674 #else 1675 static void slab_destroy_debugcheck(struct kmem_cache *cachep, 1676 struct page *page) 1677 { 1678 } 1679 #endif 1680 1681 /** 1682 * slab_destroy - destroy and release all objects in a slab 1683 * @cachep: cache pointer being destroyed 1684 * @page: page pointer being destroyed 1685 * 1686 * Destroy all the objs in a slab page, and release the mem back to the system. 1687 * Before calling the slab page must have been unlinked from the cache. The 1688 * kmem_cache_node ->list_lock is not held/needed. 1689 */ 1690 static void slab_destroy(struct kmem_cache *cachep, struct page *page) 1691 { 1692 void *freelist; 1693 1694 freelist = page->freelist; 1695 slab_destroy_debugcheck(cachep, page); 1696 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU)) 1697 call_rcu(&page->rcu_head, kmem_rcu_free); 1698 else 1699 kmem_freepages(cachep, page); 1700 1701 /* 1702 * From now on, we don't use freelist 1703 * although actual page can be freed in rcu context 1704 */ 1705 if (OFF_SLAB(cachep)) 1706 kmem_cache_free(cachep->freelist_cache, freelist); 1707 } 1708 1709 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list) 1710 { 1711 struct page *page, *n; 1712 1713 list_for_each_entry_safe(page, n, list, lru) { 1714 list_del(&page->lru); 1715 slab_destroy(cachep, page); 1716 } 1717 } 1718 1719 /** 1720 * calculate_slab_order - calculate size (page order) of slabs 1721 * @cachep: pointer to the cache that is being created 1722 * @size: size of objects to be created in this cache. 1723 * @flags: slab allocation flags 1724 * 1725 * Also calculates the number of objects per slab. 1726 * 1727 * This could be made much more intelligent. For now, try to avoid using 1728 * high order pages for slabs. When the gfp() functions are more friendly 1729 * towards high-order requests, this should be changed. 1730 * 1731 * Return: number of left-over bytes in a slab 1732 */ 1733 static size_t calculate_slab_order(struct kmem_cache *cachep, 1734 size_t size, slab_flags_t flags) 1735 { 1736 size_t left_over = 0; 1737 int gfporder; 1738 1739 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { 1740 unsigned int num; 1741 size_t remainder; 1742 1743 num = cache_estimate(gfporder, size, flags, &remainder); 1744 if (!num) 1745 continue; 1746 1747 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */ 1748 if (num > SLAB_OBJ_MAX_NUM) 1749 break; 1750 1751 if (flags & CFLGS_OFF_SLAB) { 1752 struct kmem_cache *freelist_cache; 1753 size_t freelist_size; 1754 1755 freelist_size = num * sizeof(freelist_idx_t); 1756 freelist_cache = kmalloc_slab(freelist_size, 0u); 1757 if (!freelist_cache) 1758 continue; 1759 1760 /* 1761 * Needed to avoid possible looping condition 1762 * in cache_grow_begin() 1763 */ 1764 if (OFF_SLAB(freelist_cache)) 1765 continue; 1766 1767 /* check if off slab has enough benefit */ 1768 if (freelist_cache->size > cachep->size / 2) 1769 continue; 1770 } 1771 1772 /* Found something acceptable - save it away */ 1773 cachep->num = num; 1774 cachep->gfporder = gfporder; 1775 left_over = remainder; 1776 1777 /* 1778 * A VFS-reclaimable slab tends to have most allocations 1779 * as GFP_NOFS and we really don't want to have to be allocating 1780 * higher-order pages when we are unable to shrink dcache. 1781 */ 1782 if (flags & SLAB_RECLAIM_ACCOUNT) 1783 break; 1784 1785 /* 1786 * Large number of objects is good, but very large slabs are 1787 * currently bad for the gfp()s. 1788 */ 1789 if (gfporder >= slab_max_order) 1790 break; 1791 1792 /* 1793 * Acceptable internal fragmentation? 1794 */ 1795 if (left_over * 8 <= (PAGE_SIZE << gfporder)) 1796 break; 1797 } 1798 return left_over; 1799 } 1800 1801 static struct array_cache __percpu *alloc_kmem_cache_cpus( 1802 struct kmem_cache *cachep, int entries, int batchcount) 1803 { 1804 int cpu; 1805 size_t size; 1806 struct array_cache __percpu *cpu_cache; 1807 1808 size = sizeof(void *) * entries + sizeof(struct array_cache); 1809 cpu_cache = __alloc_percpu(size, sizeof(void *)); 1810 1811 if (!cpu_cache) 1812 return NULL; 1813 1814 for_each_possible_cpu(cpu) { 1815 init_arraycache(per_cpu_ptr(cpu_cache, cpu), 1816 entries, batchcount); 1817 } 1818 1819 return cpu_cache; 1820 } 1821 1822 static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) 1823 { 1824 if (slab_state >= FULL) 1825 return enable_cpucache(cachep, gfp); 1826 1827 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1); 1828 if (!cachep->cpu_cache) 1829 return 1; 1830 1831 if (slab_state == DOWN) { 1832 /* Creation of first cache (kmem_cache). */ 1833 set_up_node(kmem_cache, CACHE_CACHE); 1834 } else if (slab_state == PARTIAL) { 1835 /* For kmem_cache_node */ 1836 set_up_node(cachep, SIZE_NODE); 1837 } else { 1838 int node; 1839 1840 for_each_online_node(node) { 1841 cachep->node[node] = kmalloc_node( 1842 sizeof(struct kmem_cache_node), gfp, node); 1843 BUG_ON(!cachep->node[node]); 1844 kmem_cache_node_init(cachep->node[node]); 1845 } 1846 } 1847 1848 cachep->node[numa_mem_id()]->next_reap = 1849 jiffies + REAPTIMEOUT_NODE + 1850 ((unsigned long)cachep) % REAPTIMEOUT_NODE; 1851 1852 cpu_cache_get(cachep)->avail = 0; 1853 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; 1854 cpu_cache_get(cachep)->batchcount = 1; 1855 cpu_cache_get(cachep)->touched = 0; 1856 cachep->batchcount = 1; 1857 cachep->limit = BOOT_CPUCACHE_ENTRIES; 1858 return 0; 1859 } 1860 1861 slab_flags_t kmem_cache_flags(unsigned int object_size, 1862 slab_flags_t flags, const char *name, 1863 void (*ctor)(void *)) 1864 { 1865 return flags; 1866 } 1867 1868 struct kmem_cache * 1869 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, 1870 slab_flags_t flags, void (*ctor)(void *)) 1871 { 1872 struct kmem_cache *cachep; 1873 1874 cachep = find_mergeable(size, align, flags, name, ctor); 1875 if (cachep) { 1876 cachep->refcount++; 1877 1878 /* 1879 * Adjust the object sizes so that we clear 1880 * the complete object on kzalloc. 1881 */ 1882 cachep->object_size = max_t(int, cachep->object_size, size); 1883 } 1884 return cachep; 1885 } 1886 1887 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep, 1888 size_t size, slab_flags_t flags) 1889 { 1890 size_t left; 1891 1892 cachep->num = 0; 1893 1894 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU) 1895 return false; 1896 1897 left = calculate_slab_order(cachep, size, 1898 flags | CFLGS_OBJFREELIST_SLAB); 1899 if (!cachep->num) 1900 return false; 1901 1902 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size) 1903 return false; 1904 1905 cachep->colour = left / cachep->colour_off; 1906 1907 return true; 1908 } 1909 1910 static bool set_off_slab_cache(struct kmem_cache *cachep, 1911 size_t size, slab_flags_t flags) 1912 { 1913 size_t left; 1914 1915 cachep->num = 0; 1916 1917 /* 1918 * Always use on-slab management when SLAB_NOLEAKTRACE 1919 * to avoid recursive calls into kmemleak. 1920 */ 1921 if (flags & SLAB_NOLEAKTRACE) 1922 return false; 1923 1924 /* 1925 * Size is large, assume best to place the slab management obj 1926 * off-slab (should allow better packing of objs). 1927 */ 1928 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB); 1929 if (!cachep->num) 1930 return false; 1931 1932 /* 1933 * If the slab has been placed off-slab, and we have enough space then 1934 * move it on-slab. This is at the expense of any extra colouring. 1935 */ 1936 if (left >= cachep->num * sizeof(freelist_idx_t)) 1937 return false; 1938 1939 cachep->colour = left / cachep->colour_off; 1940 1941 return true; 1942 } 1943 1944 static bool set_on_slab_cache(struct kmem_cache *cachep, 1945 size_t size, slab_flags_t flags) 1946 { 1947 size_t left; 1948 1949 cachep->num = 0; 1950 1951 left = calculate_slab_order(cachep, size, flags); 1952 if (!cachep->num) 1953 return false; 1954 1955 cachep->colour = left / cachep->colour_off; 1956 1957 return true; 1958 } 1959 1960 /** 1961 * __kmem_cache_create - Create a cache. 1962 * @cachep: cache management descriptor 1963 * @flags: SLAB flags 1964 * 1965 * Returns a ptr to the cache on success, NULL on failure. 1966 * Cannot be called within a int, but can be interrupted. 1967 * The @ctor is run when new pages are allocated by the cache. 1968 * 1969 * The flags are 1970 * 1971 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) 1972 * to catch references to uninitialised memory. 1973 * 1974 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check 1975 * for buffer overruns. 1976 * 1977 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware 1978 * cacheline. This can be beneficial if you're counting cycles as closely 1979 * as davem. 1980 * 1981 * Return: a pointer to the created cache or %NULL in case of error 1982 */ 1983 int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags) 1984 { 1985 size_t ralign = BYTES_PER_WORD; 1986 gfp_t gfp; 1987 int err; 1988 unsigned int size = cachep->size; 1989 1990 #if DEBUG 1991 #if FORCED_DEBUG 1992 /* 1993 * Enable redzoning and last user accounting, except for caches with 1994 * large objects, if the increased size would increase the object size 1995 * above the next power of two: caches with object sizes just above a 1996 * power of two have a significant amount of internal fragmentation. 1997 */ 1998 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + 1999 2 * sizeof(unsigned long long))) 2000 flags |= SLAB_RED_ZONE | SLAB_STORE_USER; 2001 if (!(flags & SLAB_TYPESAFE_BY_RCU)) 2002 flags |= SLAB_POISON; 2003 #endif 2004 #endif 2005 2006 /* 2007 * Check that size is in terms of words. This is needed to avoid 2008 * unaligned accesses for some archs when redzoning is used, and makes 2009 * sure any on-slab bufctl's are also correctly aligned. 2010 */ 2011 size = ALIGN(size, BYTES_PER_WORD); 2012 2013 if (flags & SLAB_RED_ZONE) { 2014 ralign = REDZONE_ALIGN; 2015 /* If redzoning, ensure that the second redzone is suitably 2016 * aligned, by adjusting the object size accordingly. */ 2017 size = ALIGN(size, REDZONE_ALIGN); 2018 } 2019 2020 /* 3) caller mandated alignment */ 2021 if (ralign < cachep->align) { 2022 ralign = cachep->align; 2023 } 2024 /* disable debug if necessary */ 2025 if (ralign > __alignof__(unsigned long long)) 2026 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2027 /* 2028 * 4) Store it. 2029 */ 2030 cachep->align = ralign; 2031 cachep->colour_off = cache_line_size(); 2032 /* Offset must be a multiple of the alignment. */ 2033 if (cachep->colour_off < cachep->align) 2034 cachep->colour_off = cachep->align; 2035 2036 if (slab_is_available()) 2037 gfp = GFP_KERNEL; 2038 else 2039 gfp = GFP_NOWAIT; 2040 2041 #if DEBUG 2042 2043 /* 2044 * Both debugging options require word-alignment which is calculated 2045 * into align above. 2046 */ 2047 if (flags & SLAB_RED_ZONE) { 2048 /* add space for red zone words */ 2049 cachep->obj_offset += sizeof(unsigned long long); 2050 size += 2 * sizeof(unsigned long long); 2051 } 2052 if (flags & SLAB_STORE_USER) { 2053 /* user store requires one word storage behind the end of 2054 * the real object. But if the second red zone needs to be 2055 * aligned to 64 bits, we must allow that much space. 2056 */ 2057 if (flags & SLAB_RED_ZONE) 2058 size += REDZONE_ALIGN; 2059 else 2060 size += BYTES_PER_WORD; 2061 } 2062 #endif 2063 2064 kasan_cache_create(cachep, &size, &flags); 2065 2066 size = ALIGN(size, cachep->align); 2067 /* 2068 * We should restrict the number of objects in a slab to implement 2069 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition. 2070 */ 2071 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE) 2072 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align); 2073 2074 #if DEBUG 2075 /* 2076 * To activate debug pagealloc, off-slab management is necessary 2077 * requirement. In early phase of initialization, small sized slab 2078 * doesn't get initialized so it would not be possible. So, we need 2079 * to check size >= 256. It guarantees that all necessary small 2080 * sized slab is initialized in current slab initialization sequence. 2081 */ 2082 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) && 2083 size >= 256 && cachep->object_size > cache_line_size()) { 2084 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) { 2085 size_t tmp_size = ALIGN(size, PAGE_SIZE); 2086 2087 if (set_off_slab_cache(cachep, tmp_size, flags)) { 2088 flags |= CFLGS_OFF_SLAB; 2089 cachep->obj_offset += tmp_size - size; 2090 size = tmp_size; 2091 goto done; 2092 } 2093 } 2094 } 2095 #endif 2096 2097 if (set_objfreelist_slab_cache(cachep, size, flags)) { 2098 flags |= CFLGS_OBJFREELIST_SLAB; 2099 goto done; 2100 } 2101 2102 if (set_off_slab_cache(cachep, size, flags)) { 2103 flags |= CFLGS_OFF_SLAB; 2104 goto done; 2105 } 2106 2107 if (set_on_slab_cache(cachep, size, flags)) 2108 goto done; 2109 2110 return -E2BIG; 2111 2112 done: 2113 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t); 2114 cachep->flags = flags; 2115 cachep->allocflags = __GFP_COMP; 2116 if (flags & SLAB_CACHE_DMA) 2117 cachep->allocflags |= GFP_DMA; 2118 if (flags & SLAB_RECLAIM_ACCOUNT) 2119 cachep->allocflags |= __GFP_RECLAIMABLE; 2120 cachep->size = size; 2121 cachep->reciprocal_buffer_size = reciprocal_value(size); 2122 2123 #if DEBUG 2124 /* 2125 * If we're going to use the generic kernel_map_pages() 2126 * poisoning, then it's going to smash the contents of 2127 * the redzone and userword anyhow, so switch them off. 2128 */ 2129 if (IS_ENABLED(CONFIG_PAGE_POISONING) && 2130 (cachep->flags & SLAB_POISON) && 2131 is_debug_pagealloc_cache(cachep)) 2132 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); 2133 #endif 2134 2135 if (OFF_SLAB(cachep)) { 2136 cachep->freelist_cache = 2137 kmalloc_slab(cachep->freelist_size, 0u); 2138 } 2139 2140 err = setup_cpu_cache(cachep, gfp); 2141 if (err) { 2142 __kmem_cache_release(cachep); 2143 return err; 2144 } 2145 2146 return 0; 2147 } 2148 2149 #if DEBUG 2150 static void check_irq_off(void) 2151 { 2152 BUG_ON(!irqs_disabled()); 2153 } 2154 2155 static void check_irq_on(void) 2156 { 2157 BUG_ON(irqs_disabled()); 2158 } 2159 2160 static void check_mutex_acquired(void) 2161 { 2162 BUG_ON(!mutex_is_locked(&slab_mutex)); 2163 } 2164 2165 static void check_spinlock_acquired(struct kmem_cache *cachep) 2166 { 2167 #ifdef CONFIG_SMP 2168 check_irq_off(); 2169 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock); 2170 #endif 2171 } 2172 2173 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) 2174 { 2175 #ifdef CONFIG_SMP 2176 check_irq_off(); 2177 assert_spin_locked(&get_node(cachep, node)->list_lock); 2178 #endif 2179 } 2180 2181 #else 2182 #define check_irq_off() do { } while(0) 2183 #define check_irq_on() do { } while(0) 2184 #define check_mutex_acquired() do { } while(0) 2185 #define check_spinlock_acquired(x) do { } while(0) 2186 #define check_spinlock_acquired_node(x, y) do { } while(0) 2187 #endif 2188 2189 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, 2190 int node, bool free_all, struct list_head *list) 2191 { 2192 int tofree; 2193 2194 if (!ac || !ac->avail) 2195 return; 2196 2197 tofree = free_all ? ac->avail : (ac->limit + 4) / 5; 2198 if (tofree > ac->avail) 2199 tofree = (ac->avail + 1) / 2; 2200 2201 free_block(cachep, ac->entry, tofree, node, list); 2202 ac->avail -= tofree; 2203 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail); 2204 } 2205 2206 static void do_drain(void *arg) 2207 { 2208 struct kmem_cache *cachep = arg; 2209 struct array_cache *ac; 2210 int node = numa_mem_id(); 2211 struct kmem_cache_node *n; 2212 LIST_HEAD(list); 2213 2214 check_irq_off(); 2215 ac = cpu_cache_get(cachep); 2216 n = get_node(cachep, node); 2217 spin_lock(&n->list_lock); 2218 free_block(cachep, ac->entry, ac->avail, node, &list); 2219 spin_unlock(&n->list_lock); 2220 slabs_destroy(cachep, &list); 2221 ac->avail = 0; 2222 } 2223 2224 static void drain_cpu_caches(struct kmem_cache *cachep) 2225 { 2226 struct kmem_cache_node *n; 2227 int node; 2228 LIST_HEAD(list); 2229 2230 on_each_cpu(do_drain, cachep, 1); 2231 check_irq_on(); 2232 for_each_kmem_cache_node(cachep, node, n) 2233 if (n->alien) 2234 drain_alien_cache(cachep, n->alien); 2235 2236 for_each_kmem_cache_node(cachep, node, n) { 2237 spin_lock_irq(&n->list_lock); 2238 drain_array_locked(cachep, n->shared, node, true, &list); 2239 spin_unlock_irq(&n->list_lock); 2240 2241 slabs_destroy(cachep, &list); 2242 } 2243 } 2244 2245 /* 2246 * Remove slabs from the list of free slabs. 2247 * Specify the number of slabs to drain in tofree. 2248 * 2249 * Returns the actual number of slabs released. 2250 */ 2251 static int drain_freelist(struct kmem_cache *cache, 2252 struct kmem_cache_node *n, int tofree) 2253 { 2254 struct list_head *p; 2255 int nr_freed; 2256 struct page *page; 2257 2258 nr_freed = 0; 2259 while (nr_freed < tofree && !list_empty(&n->slabs_free)) { 2260 2261 spin_lock_irq(&n->list_lock); 2262 p = n->slabs_free.prev; 2263 if (p == &n->slabs_free) { 2264 spin_unlock_irq(&n->list_lock); 2265 goto out; 2266 } 2267 2268 page = list_entry(p, struct page, lru); 2269 list_del(&page->lru); 2270 n->free_slabs--; 2271 n->total_slabs--; 2272 /* 2273 * Safe to drop the lock. The slab is no longer linked 2274 * to the cache. 2275 */ 2276 n->free_objects -= cache->num; 2277 spin_unlock_irq(&n->list_lock); 2278 slab_destroy(cache, page); 2279 nr_freed++; 2280 } 2281 out: 2282 return nr_freed; 2283 } 2284 2285 bool __kmem_cache_empty(struct kmem_cache *s) 2286 { 2287 int node; 2288 struct kmem_cache_node *n; 2289 2290 for_each_kmem_cache_node(s, node, n) 2291 if (!list_empty(&n->slabs_full) || 2292 !list_empty(&n->slabs_partial)) 2293 return false; 2294 return true; 2295 } 2296 2297 int __kmem_cache_shrink(struct kmem_cache *cachep) 2298 { 2299 int ret = 0; 2300 int node; 2301 struct kmem_cache_node *n; 2302 2303 drain_cpu_caches(cachep); 2304 2305 check_irq_on(); 2306 for_each_kmem_cache_node(cachep, node, n) { 2307 drain_freelist(cachep, n, INT_MAX); 2308 2309 ret += !list_empty(&n->slabs_full) || 2310 !list_empty(&n->slabs_partial); 2311 } 2312 return (ret ? 1 : 0); 2313 } 2314 2315 #ifdef CONFIG_MEMCG 2316 void __kmemcg_cache_deactivate(struct kmem_cache *cachep) 2317 { 2318 __kmem_cache_shrink(cachep); 2319 } 2320 #endif 2321 2322 int __kmem_cache_shutdown(struct kmem_cache *cachep) 2323 { 2324 return __kmem_cache_shrink(cachep); 2325 } 2326 2327 void __kmem_cache_release(struct kmem_cache *cachep) 2328 { 2329 int i; 2330 struct kmem_cache_node *n; 2331 2332 cache_random_seq_destroy(cachep); 2333 2334 free_percpu(cachep->cpu_cache); 2335 2336 /* NUMA: free the node structures */ 2337 for_each_kmem_cache_node(cachep, i, n) { 2338 kfree(n->shared); 2339 free_alien_cache(n->alien); 2340 kfree(n); 2341 cachep->node[i] = NULL; 2342 } 2343 } 2344 2345 /* 2346 * Get the memory for a slab management obj. 2347 * 2348 * For a slab cache when the slab descriptor is off-slab, the 2349 * slab descriptor can't come from the same cache which is being created, 2350 * Because if it is the case, that means we defer the creation of 2351 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point. 2352 * And we eventually call down to __kmem_cache_create(), which 2353 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one. 2354 * This is a "chicken-and-egg" problem. 2355 * 2356 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches, 2357 * which are all initialized during kmem_cache_init(). 2358 */ 2359 static void *alloc_slabmgmt(struct kmem_cache *cachep, 2360 struct page *page, int colour_off, 2361 gfp_t local_flags, int nodeid) 2362 { 2363 void *freelist; 2364 void *addr = page_address(page); 2365 2366 page->s_mem = addr + colour_off; 2367 page->active = 0; 2368 2369 if (OBJFREELIST_SLAB(cachep)) 2370 freelist = NULL; 2371 else if (OFF_SLAB(cachep)) { 2372 /* Slab management obj is off-slab. */ 2373 freelist = kmem_cache_alloc_node(cachep->freelist_cache, 2374 local_flags, nodeid); 2375 freelist = kasan_reset_tag(freelist); 2376 if (!freelist) 2377 return NULL; 2378 } else { 2379 /* We will use last bytes at the slab for freelist */ 2380 freelist = addr + (PAGE_SIZE << cachep->gfporder) - 2381 cachep->freelist_size; 2382 } 2383 2384 return freelist; 2385 } 2386 2387 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx) 2388 { 2389 return ((freelist_idx_t *)page->freelist)[idx]; 2390 } 2391 2392 static inline void set_free_obj(struct page *page, 2393 unsigned int idx, freelist_idx_t val) 2394 { 2395 ((freelist_idx_t *)(page->freelist))[idx] = val; 2396 } 2397 2398 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page) 2399 { 2400 #if DEBUG 2401 int i; 2402 2403 for (i = 0; i < cachep->num; i++) { 2404 void *objp = index_to_obj(cachep, page, i); 2405 2406 if (cachep->flags & SLAB_STORE_USER) 2407 *dbg_userword(cachep, objp) = NULL; 2408 2409 if (cachep->flags & SLAB_RED_ZONE) { 2410 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2411 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2412 } 2413 /* 2414 * Constructors are not allowed to allocate memory from the same 2415 * cache which they are a constructor for. Otherwise, deadlock. 2416 * They must also be threaded. 2417 */ 2418 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) { 2419 kasan_unpoison_object_data(cachep, 2420 objp + obj_offset(cachep)); 2421 cachep->ctor(objp + obj_offset(cachep)); 2422 kasan_poison_object_data( 2423 cachep, objp + obj_offset(cachep)); 2424 } 2425 2426 if (cachep->flags & SLAB_RED_ZONE) { 2427 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) 2428 slab_error(cachep, "constructor overwrote the end of an object"); 2429 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) 2430 slab_error(cachep, "constructor overwrote the start of an object"); 2431 } 2432 /* need to poison the objs? */ 2433 if (cachep->flags & SLAB_POISON) { 2434 poison_obj(cachep, objp, POISON_FREE); 2435 slab_kernel_map(cachep, objp, 0, 0); 2436 } 2437 } 2438 #endif 2439 } 2440 2441 #ifdef CONFIG_SLAB_FREELIST_RANDOM 2442 /* Hold information during a freelist initialization */ 2443 union freelist_init_state { 2444 struct { 2445 unsigned int pos; 2446 unsigned int *list; 2447 unsigned int count; 2448 }; 2449 struct rnd_state rnd_state; 2450 }; 2451 2452 /* 2453 * Initialize the state based on the randomization methode available. 2454 * return true if the pre-computed list is available, false otherwize. 2455 */ 2456 static bool freelist_state_initialize(union freelist_init_state *state, 2457 struct kmem_cache *cachep, 2458 unsigned int count) 2459 { 2460 bool ret; 2461 unsigned int rand; 2462 2463 /* Use best entropy available to define a random shift */ 2464 rand = get_random_int(); 2465 2466 /* Use a random state if the pre-computed list is not available */ 2467 if (!cachep->random_seq) { 2468 prandom_seed_state(&state->rnd_state, rand); 2469 ret = false; 2470 } else { 2471 state->list = cachep->random_seq; 2472 state->count = count; 2473 state->pos = rand % count; 2474 ret = true; 2475 } 2476 return ret; 2477 } 2478 2479 /* Get the next entry on the list and randomize it using a random shift */ 2480 static freelist_idx_t next_random_slot(union freelist_init_state *state) 2481 { 2482 if (state->pos >= state->count) 2483 state->pos = 0; 2484 return state->list[state->pos++]; 2485 } 2486 2487 /* Swap two freelist entries */ 2488 static void swap_free_obj(struct page *page, unsigned int a, unsigned int b) 2489 { 2490 swap(((freelist_idx_t *)page->freelist)[a], 2491 ((freelist_idx_t *)page->freelist)[b]); 2492 } 2493 2494 /* 2495 * Shuffle the freelist initialization state based on pre-computed lists. 2496 * return true if the list was successfully shuffled, false otherwise. 2497 */ 2498 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page) 2499 { 2500 unsigned int objfreelist = 0, i, rand, count = cachep->num; 2501 union freelist_init_state state; 2502 bool precomputed; 2503 2504 if (count < 2) 2505 return false; 2506 2507 precomputed = freelist_state_initialize(&state, cachep, count); 2508 2509 /* Take a random entry as the objfreelist */ 2510 if (OBJFREELIST_SLAB(cachep)) { 2511 if (!precomputed) 2512 objfreelist = count - 1; 2513 else 2514 objfreelist = next_random_slot(&state); 2515 page->freelist = index_to_obj(cachep, page, objfreelist) + 2516 obj_offset(cachep); 2517 count--; 2518 } 2519 2520 /* 2521 * On early boot, generate the list dynamically. 2522 * Later use a pre-computed list for speed. 2523 */ 2524 if (!precomputed) { 2525 for (i = 0; i < count; i++) 2526 set_free_obj(page, i, i); 2527 2528 /* Fisher-Yates shuffle */ 2529 for (i = count - 1; i > 0; i--) { 2530 rand = prandom_u32_state(&state.rnd_state); 2531 rand %= (i + 1); 2532 swap_free_obj(page, i, rand); 2533 } 2534 } else { 2535 for (i = 0; i < count; i++) 2536 set_free_obj(page, i, next_random_slot(&state)); 2537 } 2538 2539 if (OBJFREELIST_SLAB(cachep)) 2540 set_free_obj(page, cachep->num - 1, objfreelist); 2541 2542 return true; 2543 } 2544 #else 2545 static inline bool shuffle_freelist(struct kmem_cache *cachep, 2546 struct page *page) 2547 { 2548 return false; 2549 } 2550 #endif /* CONFIG_SLAB_FREELIST_RANDOM */ 2551 2552 static void cache_init_objs(struct kmem_cache *cachep, 2553 struct page *page) 2554 { 2555 int i; 2556 void *objp; 2557 bool shuffled; 2558 2559 cache_init_objs_debug(cachep, page); 2560 2561 /* Try to randomize the freelist if enabled */ 2562 shuffled = shuffle_freelist(cachep, page); 2563 2564 if (!shuffled && OBJFREELIST_SLAB(cachep)) { 2565 page->freelist = index_to_obj(cachep, page, cachep->num - 1) + 2566 obj_offset(cachep); 2567 } 2568 2569 for (i = 0; i < cachep->num; i++) { 2570 objp = index_to_obj(cachep, page, i); 2571 objp = kasan_init_slab_obj(cachep, objp); 2572 2573 /* constructor could break poison info */ 2574 if (DEBUG == 0 && cachep->ctor) { 2575 kasan_unpoison_object_data(cachep, objp); 2576 cachep->ctor(objp); 2577 kasan_poison_object_data(cachep, objp); 2578 } 2579 2580 if (!shuffled) 2581 set_free_obj(page, i, i); 2582 } 2583 } 2584 2585 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page) 2586 { 2587 void *objp; 2588 2589 objp = index_to_obj(cachep, page, get_free_obj(page, page->active)); 2590 page->active++; 2591 2592 #if DEBUG 2593 if (cachep->flags & SLAB_STORE_USER) 2594 set_store_user_dirty(cachep); 2595 #endif 2596 2597 return objp; 2598 } 2599 2600 static void slab_put_obj(struct kmem_cache *cachep, 2601 struct page *page, void *objp) 2602 { 2603 unsigned int objnr = obj_to_index(cachep, page, objp); 2604 #if DEBUG 2605 unsigned int i; 2606 2607 /* Verify double free bug */ 2608 for (i = page->active; i < cachep->num; i++) { 2609 if (get_free_obj(page, i) == objnr) { 2610 pr_err("slab: double free detected in cache '%s', objp %px\n", 2611 cachep->name, objp); 2612 BUG(); 2613 } 2614 } 2615 #endif 2616 page->active--; 2617 if (!page->freelist) 2618 page->freelist = objp + obj_offset(cachep); 2619 2620 set_free_obj(page, page->active, objnr); 2621 } 2622 2623 /* 2624 * Map pages beginning at addr to the given cache and slab. This is required 2625 * for the slab allocator to be able to lookup the cache and slab of a 2626 * virtual address for kfree, ksize, and slab debugging. 2627 */ 2628 static void slab_map_pages(struct kmem_cache *cache, struct page *page, 2629 void *freelist) 2630 { 2631 page->slab_cache = cache; 2632 page->freelist = freelist; 2633 } 2634 2635 /* 2636 * Grow (by 1) the number of slabs within a cache. This is called by 2637 * kmem_cache_alloc() when there are no active objs left in a cache. 2638 */ 2639 static struct page *cache_grow_begin(struct kmem_cache *cachep, 2640 gfp_t flags, int nodeid) 2641 { 2642 void *freelist; 2643 size_t offset; 2644 gfp_t local_flags; 2645 int page_node; 2646 struct kmem_cache_node *n; 2647 struct page *page; 2648 2649 /* 2650 * Be lazy and only check for valid flags here, keeping it out of the 2651 * critical path in kmem_cache_alloc(). 2652 */ 2653 if (unlikely(flags & GFP_SLAB_BUG_MASK)) { 2654 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK; 2655 flags &= ~GFP_SLAB_BUG_MASK; 2656 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n", 2657 invalid_mask, &invalid_mask, flags, &flags); 2658 dump_stack(); 2659 } 2660 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO)); 2661 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); 2662 2663 check_irq_off(); 2664 if (gfpflags_allow_blocking(local_flags)) 2665 local_irq_enable(); 2666 2667 /* 2668 * Get mem for the objs. Attempt to allocate a physical page from 2669 * 'nodeid'. 2670 */ 2671 page = kmem_getpages(cachep, local_flags, nodeid); 2672 if (!page) 2673 goto failed; 2674 2675 page_node = page_to_nid(page); 2676 n = get_node(cachep, page_node); 2677 2678 /* Get colour for the slab, and cal the next value. */ 2679 n->colour_next++; 2680 if (n->colour_next >= cachep->colour) 2681 n->colour_next = 0; 2682 2683 offset = n->colour_next; 2684 if (offset >= cachep->colour) 2685 offset = 0; 2686 2687 offset *= cachep->colour_off; 2688 2689 /* 2690 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so 2691 * page_address() in the latter returns a non-tagged pointer, 2692 * as it should be for slab pages. 2693 */ 2694 kasan_poison_slab(page); 2695 2696 /* Get slab management. */ 2697 freelist = alloc_slabmgmt(cachep, page, offset, 2698 local_flags & ~GFP_CONSTRAINT_MASK, page_node); 2699 if (OFF_SLAB(cachep) && !freelist) 2700 goto opps1; 2701 2702 slab_map_pages(cachep, page, freelist); 2703 2704 cache_init_objs(cachep, page); 2705 2706 if (gfpflags_allow_blocking(local_flags)) 2707 local_irq_disable(); 2708 2709 return page; 2710 2711 opps1: 2712 kmem_freepages(cachep, page); 2713 failed: 2714 if (gfpflags_allow_blocking(local_flags)) 2715 local_irq_disable(); 2716 return NULL; 2717 } 2718 2719 static void cache_grow_end(struct kmem_cache *cachep, struct page *page) 2720 { 2721 struct kmem_cache_node *n; 2722 void *list = NULL; 2723 2724 check_irq_off(); 2725 2726 if (!page) 2727 return; 2728 2729 INIT_LIST_HEAD(&page->lru); 2730 n = get_node(cachep, page_to_nid(page)); 2731 2732 spin_lock(&n->list_lock); 2733 n->total_slabs++; 2734 if (!page->active) { 2735 list_add_tail(&page->lru, &(n->slabs_free)); 2736 n->free_slabs++; 2737 } else 2738 fixup_slab_list(cachep, n, page, &list); 2739 2740 STATS_INC_GROWN(cachep); 2741 n->free_objects += cachep->num - page->active; 2742 spin_unlock(&n->list_lock); 2743 2744 fixup_objfreelist_debug(cachep, &list); 2745 } 2746 2747 #if DEBUG 2748 2749 /* 2750 * Perform extra freeing checks: 2751 * - detect bad pointers. 2752 * - POISON/RED_ZONE checking 2753 */ 2754 static void kfree_debugcheck(const void *objp) 2755 { 2756 if (!virt_addr_valid(objp)) { 2757 pr_err("kfree_debugcheck: out of range ptr %lxh\n", 2758 (unsigned long)objp); 2759 BUG(); 2760 } 2761 } 2762 2763 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) 2764 { 2765 unsigned long long redzone1, redzone2; 2766 2767 redzone1 = *dbg_redzone1(cache, obj); 2768 redzone2 = *dbg_redzone2(cache, obj); 2769 2770 /* 2771 * Redzone is ok. 2772 */ 2773 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) 2774 return; 2775 2776 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) 2777 slab_error(cache, "double free detected"); 2778 else 2779 slab_error(cache, "memory outside object was overwritten"); 2780 2781 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n", 2782 obj, redzone1, redzone2); 2783 } 2784 2785 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, 2786 unsigned long caller) 2787 { 2788 unsigned int objnr; 2789 struct page *page; 2790 2791 BUG_ON(virt_to_cache(objp) != cachep); 2792 2793 objp -= obj_offset(cachep); 2794 kfree_debugcheck(objp); 2795 page = virt_to_head_page(objp); 2796 2797 if (cachep->flags & SLAB_RED_ZONE) { 2798 verify_redzone_free(cachep, objp); 2799 *dbg_redzone1(cachep, objp) = RED_INACTIVE; 2800 *dbg_redzone2(cachep, objp) = RED_INACTIVE; 2801 } 2802 if (cachep->flags & SLAB_STORE_USER) { 2803 set_store_user_dirty(cachep); 2804 *dbg_userword(cachep, objp) = (void *)caller; 2805 } 2806 2807 objnr = obj_to_index(cachep, page, objp); 2808 2809 BUG_ON(objnr >= cachep->num); 2810 BUG_ON(objp != index_to_obj(cachep, page, objnr)); 2811 2812 if (cachep->flags & SLAB_POISON) { 2813 poison_obj(cachep, objp, POISON_FREE); 2814 slab_kernel_map(cachep, objp, 0, caller); 2815 } 2816 return objp; 2817 } 2818 2819 #else 2820 #define kfree_debugcheck(x) do { } while(0) 2821 #define cache_free_debugcheck(x,objp,z) (objp) 2822 #endif 2823 2824 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep, 2825 void **list) 2826 { 2827 #if DEBUG 2828 void *next = *list; 2829 void *objp; 2830 2831 while (next) { 2832 objp = next - obj_offset(cachep); 2833 next = *(void **)next; 2834 poison_obj(cachep, objp, POISON_FREE); 2835 } 2836 #endif 2837 } 2838 2839 static inline void fixup_slab_list(struct kmem_cache *cachep, 2840 struct kmem_cache_node *n, struct page *page, 2841 void **list) 2842 { 2843 /* move slabp to correct slabp list: */ 2844 list_del(&page->lru); 2845 if (page->active == cachep->num) { 2846 list_add(&page->lru, &n->slabs_full); 2847 if (OBJFREELIST_SLAB(cachep)) { 2848 #if DEBUG 2849 /* Poisoning will be done without holding the lock */ 2850 if (cachep->flags & SLAB_POISON) { 2851 void **objp = page->freelist; 2852 2853 *objp = *list; 2854 *list = objp; 2855 } 2856 #endif 2857 page->freelist = NULL; 2858 } 2859 } else 2860 list_add(&page->lru, &n->slabs_partial); 2861 } 2862 2863 /* Try to find non-pfmemalloc slab if needed */ 2864 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n, 2865 struct page *page, bool pfmemalloc) 2866 { 2867 if (!page) 2868 return NULL; 2869 2870 if (pfmemalloc) 2871 return page; 2872 2873 if (!PageSlabPfmemalloc(page)) 2874 return page; 2875 2876 /* No need to keep pfmemalloc slab if we have enough free objects */ 2877 if (n->free_objects > n->free_limit) { 2878 ClearPageSlabPfmemalloc(page); 2879 return page; 2880 } 2881 2882 /* Move pfmemalloc slab to the end of list to speed up next search */ 2883 list_del(&page->lru); 2884 if (!page->active) { 2885 list_add_tail(&page->lru, &n->slabs_free); 2886 n->free_slabs++; 2887 } else 2888 list_add_tail(&page->lru, &n->slabs_partial); 2889 2890 list_for_each_entry(page, &n->slabs_partial, lru) { 2891 if (!PageSlabPfmemalloc(page)) 2892 return page; 2893 } 2894 2895 n->free_touched = 1; 2896 list_for_each_entry(page, &n->slabs_free, lru) { 2897 if (!PageSlabPfmemalloc(page)) { 2898 n->free_slabs--; 2899 return page; 2900 } 2901 } 2902 2903 return NULL; 2904 } 2905 2906 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc) 2907 { 2908 struct page *page; 2909 2910 assert_spin_locked(&n->list_lock); 2911 page = list_first_entry_or_null(&n->slabs_partial, struct page, lru); 2912 if (!page) { 2913 n->free_touched = 1; 2914 page = list_first_entry_or_null(&n->slabs_free, struct page, 2915 lru); 2916 if (page) 2917 n->free_slabs--; 2918 } 2919 2920 if (sk_memalloc_socks()) 2921 page = get_valid_first_slab(n, page, pfmemalloc); 2922 2923 return page; 2924 } 2925 2926 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep, 2927 struct kmem_cache_node *n, gfp_t flags) 2928 { 2929 struct page *page; 2930 void *obj; 2931 void *list = NULL; 2932 2933 if (!gfp_pfmemalloc_allowed(flags)) 2934 return NULL; 2935 2936 spin_lock(&n->list_lock); 2937 page = get_first_slab(n, true); 2938 if (!page) { 2939 spin_unlock(&n->list_lock); 2940 return NULL; 2941 } 2942 2943 obj = slab_get_obj(cachep, page); 2944 n->free_objects--; 2945 2946 fixup_slab_list(cachep, n, page, &list); 2947 2948 spin_unlock(&n->list_lock); 2949 fixup_objfreelist_debug(cachep, &list); 2950 2951 return obj; 2952 } 2953 2954 /* 2955 * Slab list should be fixed up by fixup_slab_list() for existing slab 2956 * or cache_grow_end() for new slab 2957 */ 2958 static __always_inline int alloc_block(struct kmem_cache *cachep, 2959 struct array_cache *ac, struct page *page, int batchcount) 2960 { 2961 /* 2962 * There must be at least one object available for 2963 * allocation. 2964 */ 2965 BUG_ON(page->active >= cachep->num); 2966 2967 while (page->active < cachep->num && batchcount--) { 2968 STATS_INC_ALLOCED(cachep); 2969 STATS_INC_ACTIVE(cachep); 2970 STATS_SET_HIGH(cachep); 2971 2972 ac->entry[ac->avail++] = slab_get_obj(cachep, page); 2973 } 2974 2975 return batchcount; 2976 } 2977 2978 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) 2979 { 2980 int batchcount; 2981 struct kmem_cache_node *n; 2982 struct array_cache *ac, *shared; 2983 int node; 2984 void *list = NULL; 2985 struct page *page; 2986 2987 check_irq_off(); 2988 node = numa_mem_id(); 2989 2990 ac = cpu_cache_get(cachep); 2991 batchcount = ac->batchcount; 2992 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { 2993 /* 2994 * If there was little recent activity on this cache, then 2995 * perform only a partial refill. Otherwise we could generate 2996 * refill bouncing. 2997 */ 2998 batchcount = BATCHREFILL_LIMIT; 2999 } 3000 n = get_node(cachep, node); 3001 3002 BUG_ON(ac->avail > 0 || !n); 3003 shared = READ_ONCE(n->shared); 3004 if (!n->free_objects && (!shared || !shared->avail)) 3005 goto direct_grow; 3006 3007 spin_lock(&n->list_lock); 3008 shared = READ_ONCE(n->shared); 3009 3010 /* See if we can refill from the shared array */ 3011 if (shared && transfer_objects(ac, shared, batchcount)) { 3012 shared->touched = 1; 3013 goto alloc_done; 3014 } 3015 3016 while (batchcount > 0) { 3017 /* Get slab alloc is to come from. */ 3018 page = get_first_slab(n, false); 3019 if (!page) 3020 goto must_grow; 3021 3022 check_spinlock_acquired(cachep); 3023 3024 batchcount = alloc_block(cachep, ac, page, batchcount); 3025 fixup_slab_list(cachep, n, page, &list); 3026 } 3027 3028 must_grow: 3029 n->free_objects -= ac->avail; 3030 alloc_done: 3031 spin_unlock(&n->list_lock); 3032 fixup_objfreelist_debug(cachep, &list); 3033 3034 direct_grow: 3035 if (unlikely(!ac->avail)) { 3036 /* Check if we can use obj in pfmemalloc slab */ 3037 if (sk_memalloc_socks()) { 3038 void *obj = cache_alloc_pfmemalloc(cachep, n, flags); 3039 3040 if (obj) 3041 return obj; 3042 } 3043 3044 page = cache_grow_begin(cachep, gfp_exact_node(flags), node); 3045 3046 /* 3047 * cache_grow_begin() can reenable interrupts, 3048 * then ac could change. 3049 */ 3050 ac = cpu_cache_get(cachep); 3051 if (!ac->avail && page) 3052 alloc_block(cachep, ac, page, batchcount); 3053 cache_grow_end(cachep, page); 3054 3055 if (!ac->avail) 3056 return NULL; 3057 } 3058 ac->touched = 1; 3059 3060 return ac->entry[--ac->avail]; 3061 } 3062 3063 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, 3064 gfp_t flags) 3065 { 3066 might_sleep_if(gfpflags_allow_blocking(flags)); 3067 } 3068 3069 #if DEBUG 3070 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, 3071 gfp_t flags, void *objp, unsigned long caller) 3072 { 3073 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO)); 3074 if (!objp) 3075 return objp; 3076 if (cachep->flags & SLAB_POISON) { 3077 check_poison_obj(cachep, objp); 3078 slab_kernel_map(cachep, objp, 1, 0); 3079 poison_obj(cachep, objp, POISON_INUSE); 3080 } 3081 if (cachep->flags & SLAB_STORE_USER) 3082 *dbg_userword(cachep, objp) = (void *)caller; 3083 3084 if (cachep->flags & SLAB_RED_ZONE) { 3085 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || 3086 *dbg_redzone2(cachep, objp) != RED_INACTIVE) { 3087 slab_error(cachep, "double free, or memory outside object was overwritten"); 3088 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n", 3089 objp, *dbg_redzone1(cachep, objp), 3090 *dbg_redzone2(cachep, objp)); 3091 } 3092 *dbg_redzone1(cachep, objp) = RED_ACTIVE; 3093 *dbg_redzone2(cachep, objp) = RED_ACTIVE; 3094 } 3095 3096 objp += obj_offset(cachep); 3097 if (cachep->ctor && cachep->flags & SLAB_POISON) 3098 cachep->ctor(objp); 3099 if (ARCH_SLAB_MINALIGN && 3100 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { 3101 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n", 3102 objp, (int)ARCH_SLAB_MINALIGN); 3103 } 3104 return objp; 3105 } 3106 #else 3107 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) 3108 #endif 3109 3110 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3111 { 3112 void *objp; 3113 struct array_cache *ac; 3114 3115 check_irq_off(); 3116 3117 ac = cpu_cache_get(cachep); 3118 if (likely(ac->avail)) { 3119 ac->touched = 1; 3120 objp = ac->entry[--ac->avail]; 3121 3122 STATS_INC_ALLOCHIT(cachep); 3123 goto out; 3124 } 3125 3126 STATS_INC_ALLOCMISS(cachep); 3127 objp = cache_alloc_refill(cachep, flags); 3128 /* 3129 * the 'ac' may be updated by cache_alloc_refill(), 3130 * and kmemleak_erase() requires its correct value. 3131 */ 3132 ac = cpu_cache_get(cachep); 3133 3134 out: 3135 /* 3136 * To avoid a false negative, if an object that is in one of the 3137 * per-CPU caches is leaked, we need to make sure kmemleak doesn't 3138 * treat the array pointers as a reference to the object. 3139 */ 3140 if (objp) 3141 kmemleak_erase(&ac->entry[ac->avail]); 3142 return objp; 3143 } 3144 3145 #ifdef CONFIG_NUMA 3146 /* 3147 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set. 3148 * 3149 * If we are in_interrupt, then process context, including cpusets and 3150 * mempolicy, may not apply and should not be used for allocation policy. 3151 */ 3152 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) 3153 { 3154 int nid_alloc, nid_here; 3155 3156 if (in_interrupt() || (flags & __GFP_THISNODE)) 3157 return NULL; 3158 nid_alloc = nid_here = numa_mem_id(); 3159 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) 3160 nid_alloc = cpuset_slab_spread_node(); 3161 else if (current->mempolicy) 3162 nid_alloc = mempolicy_slab_node(); 3163 if (nid_alloc != nid_here) 3164 return ____cache_alloc_node(cachep, flags, nid_alloc); 3165 return NULL; 3166 } 3167 3168 /* 3169 * Fallback function if there was no memory available and no objects on a 3170 * certain node and fall back is permitted. First we scan all the 3171 * available node for available objects. If that fails then we 3172 * perform an allocation without specifying a node. This allows the page 3173 * allocator to do its reclaim / fallback magic. We then insert the 3174 * slab into the proper nodelist and then allocate from it. 3175 */ 3176 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) 3177 { 3178 struct zonelist *zonelist; 3179 struct zoneref *z; 3180 struct zone *zone; 3181 enum zone_type high_zoneidx = gfp_zone(flags); 3182 void *obj = NULL; 3183 struct page *page; 3184 int nid; 3185 unsigned int cpuset_mems_cookie; 3186 3187 if (flags & __GFP_THISNODE) 3188 return NULL; 3189 3190 retry_cpuset: 3191 cpuset_mems_cookie = read_mems_allowed_begin(); 3192 zonelist = node_zonelist(mempolicy_slab_node(), flags); 3193 3194 retry: 3195 /* 3196 * Look through allowed nodes for objects available 3197 * from existing per node queues. 3198 */ 3199 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 3200 nid = zone_to_nid(zone); 3201 3202 if (cpuset_zone_allowed(zone, flags) && 3203 get_node(cache, nid) && 3204 get_node(cache, nid)->free_objects) { 3205 obj = ____cache_alloc_node(cache, 3206 gfp_exact_node(flags), nid); 3207 if (obj) 3208 break; 3209 } 3210 } 3211 3212 if (!obj) { 3213 /* 3214 * This allocation will be performed within the constraints 3215 * of the current cpuset / memory policy requirements. 3216 * We may trigger various forms of reclaim on the allowed 3217 * set and go into memory reserves if necessary. 3218 */ 3219 page = cache_grow_begin(cache, flags, numa_mem_id()); 3220 cache_grow_end(cache, page); 3221 if (page) { 3222 nid = page_to_nid(page); 3223 obj = ____cache_alloc_node(cache, 3224 gfp_exact_node(flags), nid); 3225 3226 /* 3227 * Another processor may allocate the objects in 3228 * the slab since we are not holding any locks. 3229 */ 3230 if (!obj) 3231 goto retry; 3232 } 3233 } 3234 3235 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie))) 3236 goto retry_cpuset; 3237 return obj; 3238 } 3239 3240 /* 3241 * A interface to enable slab creation on nodeid 3242 */ 3243 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, 3244 int nodeid) 3245 { 3246 struct page *page; 3247 struct kmem_cache_node *n; 3248 void *obj = NULL; 3249 void *list = NULL; 3250 3251 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES); 3252 n = get_node(cachep, nodeid); 3253 BUG_ON(!n); 3254 3255 check_irq_off(); 3256 spin_lock(&n->list_lock); 3257 page = get_first_slab(n, false); 3258 if (!page) 3259 goto must_grow; 3260 3261 check_spinlock_acquired_node(cachep, nodeid); 3262 3263 STATS_INC_NODEALLOCS(cachep); 3264 STATS_INC_ACTIVE(cachep); 3265 STATS_SET_HIGH(cachep); 3266 3267 BUG_ON(page->active == cachep->num); 3268 3269 obj = slab_get_obj(cachep, page); 3270 n->free_objects--; 3271 3272 fixup_slab_list(cachep, n, page, &list); 3273 3274 spin_unlock(&n->list_lock); 3275 fixup_objfreelist_debug(cachep, &list); 3276 return obj; 3277 3278 must_grow: 3279 spin_unlock(&n->list_lock); 3280 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid); 3281 if (page) { 3282 /* This slab isn't counted yet so don't update free_objects */ 3283 obj = slab_get_obj(cachep, page); 3284 } 3285 cache_grow_end(cachep, page); 3286 3287 return obj ? obj : fallback_alloc(cachep, flags); 3288 } 3289 3290 static __always_inline void * 3291 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, 3292 unsigned long caller) 3293 { 3294 unsigned long save_flags; 3295 void *ptr; 3296 int slab_node = numa_mem_id(); 3297 3298 flags &= gfp_allowed_mask; 3299 cachep = slab_pre_alloc_hook(cachep, flags); 3300 if (unlikely(!cachep)) 3301 return NULL; 3302 3303 cache_alloc_debugcheck_before(cachep, flags); 3304 local_irq_save(save_flags); 3305 3306 if (nodeid == NUMA_NO_NODE) 3307 nodeid = slab_node; 3308 3309 if (unlikely(!get_node(cachep, nodeid))) { 3310 /* Node not bootstrapped yet */ 3311 ptr = fallback_alloc(cachep, flags); 3312 goto out; 3313 } 3314 3315 if (nodeid == slab_node) { 3316 /* 3317 * Use the locally cached objects if possible. 3318 * However ____cache_alloc does not allow fallback 3319 * to other nodes. It may fail while we still have 3320 * objects on other nodes available. 3321 */ 3322 ptr = ____cache_alloc(cachep, flags); 3323 if (ptr) 3324 goto out; 3325 } 3326 /* ___cache_alloc_node can fall back to other nodes */ 3327 ptr = ____cache_alloc_node(cachep, flags, nodeid); 3328 out: 3329 local_irq_restore(save_flags); 3330 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); 3331 3332 if (unlikely(flags & __GFP_ZERO) && ptr) 3333 memset(ptr, 0, cachep->object_size); 3334 3335 slab_post_alloc_hook(cachep, flags, 1, &ptr); 3336 return ptr; 3337 } 3338 3339 static __always_inline void * 3340 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) 3341 { 3342 void *objp; 3343 3344 if (current->mempolicy || cpuset_do_slab_mem_spread()) { 3345 objp = alternate_node_alloc(cache, flags); 3346 if (objp) 3347 goto out; 3348 } 3349 objp = ____cache_alloc(cache, flags); 3350 3351 /* 3352 * We may just have run out of memory on the local node. 3353 * ____cache_alloc_node() knows how to locate memory on other nodes 3354 */ 3355 if (!objp) 3356 objp = ____cache_alloc_node(cache, flags, numa_mem_id()); 3357 3358 out: 3359 return objp; 3360 } 3361 #else 3362 3363 static __always_inline void * 3364 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3365 { 3366 return ____cache_alloc(cachep, flags); 3367 } 3368 3369 #endif /* CONFIG_NUMA */ 3370 3371 static __always_inline void * 3372 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) 3373 { 3374 unsigned long save_flags; 3375 void *objp; 3376 3377 flags &= gfp_allowed_mask; 3378 cachep = slab_pre_alloc_hook(cachep, flags); 3379 if (unlikely(!cachep)) 3380 return NULL; 3381 3382 cache_alloc_debugcheck_before(cachep, flags); 3383 local_irq_save(save_flags); 3384 objp = __do_cache_alloc(cachep, flags); 3385 local_irq_restore(save_flags); 3386 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); 3387 prefetchw(objp); 3388 3389 if (unlikely(flags & __GFP_ZERO) && objp) 3390 memset(objp, 0, cachep->object_size); 3391 3392 slab_post_alloc_hook(cachep, flags, 1, &objp); 3393 return objp; 3394 } 3395 3396 /* 3397 * Caller needs to acquire correct kmem_cache_node's list_lock 3398 * @list: List of detached free slabs should be freed by caller 3399 */ 3400 static void free_block(struct kmem_cache *cachep, void **objpp, 3401 int nr_objects, int node, struct list_head *list) 3402 { 3403 int i; 3404 struct kmem_cache_node *n = get_node(cachep, node); 3405 struct page *page; 3406 3407 n->free_objects += nr_objects; 3408 3409 for (i = 0; i < nr_objects; i++) { 3410 void *objp; 3411 struct page *page; 3412 3413 objp = objpp[i]; 3414 3415 page = virt_to_head_page(objp); 3416 list_del(&page->lru); 3417 check_spinlock_acquired_node(cachep, node); 3418 slab_put_obj(cachep, page, objp); 3419 STATS_DEC_ACTIVE(cachep); 3420 3421 /* fixup slab chains */ 3422 if (page->active == 0) { 3423 list_add(&page->lru, &n->slabs_free); 3424 n->free_slabs++; 3425 } else { 3426 /* Unconditionally move a slab to the end of the 3427 * partial list on free - maximum time for the 3428 * other objects to be freed, too. 3429 */ 3430 list_add_tail(&page->lru, &n->slabs_partial); 3431 } 3432 } 3433 3434 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) { 3435 n->free_objects -= cachep->num; 3436 3437 page = list_last_entry(&n->slabs_free, struct page, lru); 3438 list_move(&page->lru, list); 3439 n->free_slabs--; 3440 n->total_slabs--; 3441 } 3442 } 3443 3444 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) 3445 { 3446 int batchcount; 3447 struct kmem_cache_node *n; 3448 int node = numa_mem_id(); 3449 LIST_HEAD(list); 3450 3451 batchcount = ac->batchcount; 3452 3453 check_irq_off(); 3454 n = get_node(cachep, node); 3455 spin_lock(&n->list_lock); 3456 if (n->shared) { 3457 struct array_cache *shared_array = n->shared; 3458 int max = shared_array->limit - shared_array->avail; 3459 if (max) { 3460 if (batchcount > max) 3461 batchcount = max; 3462 memcpy(&(shared_array->entry[shared_array->avail]), 3463 ac->entry, sizeof(void *) * batchcount); 3464 shared_array->avail += batchcount; 3465 goto free_done; 3466 } 3467 } 3468 3469 free_block(cachep, ac->entry, batchcount, node, &list); 3470 free_done: 3471 #if STATS 3472 { 3473 int i = 0; 3474 struct page *page; 3475 3476 list_for_each_entry(page, &n->slabs_free, lru) { 3477 BUG_ON(page->active); 3478 3479 i++; 3480 } 3481 STATS_SET_FREEABLE(cachep, i); 3482 } 3483 #endif 3484 spin_unlock(&n->list_lock); 3485 slabs_destroy(cachep, &list); 3486 ac->avail -= batchcount; 3487 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); 3488 } 3489 3490 /* 3491 * Release an obj back to its cache. If the obj has a constructed state, it must 3492 * be in this state _before_ it is released. Called with disabled ints. 3493 */ 3494 static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp, 3495 unsigned long caller) 3496 { 3497 /* Put the object into the quarantine, don't touch it for now. */ 3498 if (kasan_slab_free(cachep, objp, _RET_IP_)) 3499 return; 3500 3501 ___cache_free(cachep, objp, caller); 3502 } 3503 3504 void ___cache_free(struct kmem_cache *cachep, void *objp, 3505 unsigned long caller) 3506 { 3507 struct array_cache *ac = cpu_cache_get(cachep); 3508 3509 check_irq_off(); 3510 kmemleak_free_recursive(objp, cachep->flags); 3511 objp = cache_free_debugcheck(cachep, objp, caller); 3512 3513 /* 3514 * Skip calling cache_free_alien() when the platform is not numa. 3515 * This will avoid cache misses that happen while accessing slabp (which 3516 * is per page memory reference) to get nodeid. Instead use a global 3517 * variable to skip the call, which is mostly likely to be present in 3518 * the cache. 3519 */ 3520 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) 3521 return; 3522 3523 if (ac->avail < ac->limit) { 3524 STATS_INC_FREEHIT(cachep); 3525 } else { 3526 STATS_INC_FREEMISS(cachep); 3527 cache_flusharray(cachep, ac); 3528 } 3529 3530 if (sk_memalloc_socks()) { 3531 struct page *page = virt_to_head_page(objp); 3532 3533 if (unlikely(PageSlabPfmemalloc(page))) { 3534 cache_free_pfmemalloc(cachep, page, objp); 3535 return; 3536 } 3537 } 3538 3539 ac->entry[ac->avail++] = objp; 3540 } 3541 3542 /** 3543 * kmem_cache_alloc - Allocate an object 3544 * @cachep: The cache to allocate from. 3545 * @flags: See kmalloc(). 3546 * 3547 * Allocate an object from this cache. The flags are only relevant 3548 * if the cache has no available objects. 3549 * 3550 * Return: pointer to the new object or %NULL in case of error 3551 */ 3552 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) 3553 { 3554 void *ret = slab_alloc(cachep, flags, _RET_IP_); 3555 3556 trace_kmem_cache_alloc(_RET_IP_, ret, 3557 cachep->object_size, cachep->size, flags); 3558 3559 return ret; 3560 } 3561 EXPORT_SYMBOL(kmem_cache_alloc); 3562 3563 static __always_inline void 3564 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags, 3565 size_t size, void **p, unsigned long caller) 3566 { 3567 size_t i; 3568 3569 for (i = 0; i < size; i++) 3570 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller); 3571 } 3572 3573 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, 3574 void **p) 3575 { 3576 size_t i; 3577 3578 s = slab_pre_alloc_hook(s, flags); 3579 if (!s) 3580 return 0; 3581 3582 cache_alloc_debugcheck_before(s, flags); 3583 3584 local_irq_disable(); 3585 for (i = 0; i < size; i++) { 3586 void *objp = __do_cache_alloc(s, flags); 3587 3588 if (unlikely(!objp)) 3589 goto error; 3590 p[i] = objp; 3591 } 3592 local_irq_enable(); 3593 3594 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_); 3595 3596 /* Clear memory outside IRQ disabled section */ 3597 if (unlikely(flags & __GFP_ZERO)) 3598 for (i = 0; i < size; i++) 3599 memset(p[i], 0, s->object_size); 3600 3601 slab_post_alloc_hook(s, flags, size, p); 3602 /* FIXME: Trace call missing. Christoph would like a bulk variant */ 3603 return size; 3604 error: 3605 local_irq_enable(); 3606 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_); 3607 slab_post_alloc_hook(s, flags, i, p); 3608 __kmem_cache_free_bulk(s, i, p); 3609 return 0; 3610 } 3611 EXPORT_SYMBOL(kmem_cache_alloc_bulk); 3612 3613 #ifdef CONFIG_TRACING 3614 void * 3615 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) 3616 { 3617 void *ret; 3618 3619 ret = slab_alloc(cachep, flags, _RET_IP_); 3620 3621 ret = kasan_kmalloc(cachep, ret, size, flags); 3622 trace_kmalloc(_RET_IP_, ret, 3623 size, cachep->size, flags); 3624 return ret; 3625 } 3626 EXPORT_SYMBOL(kmem_cache_alloc_trace); 3627 #endif 3628 3629 #ifdef CONFIG_NUMA 3630 /** 3631 * kmem_cache_alloc_node - Allocate an object on the specified node 3632 * @cachep: The cache to allocate from. 3633 * @flags: See kmalloc(). 3634 * @nodeid: node number of the target node. 3635 * 3636 * Identical to kmem_cache_alloc but it will allocate memory on the given 3637 * node, which can improve the performance for cpu bound structures. 3638 * 3639 * Fallback to other node is possible if __GFP_THISNODE is not set. 3640 * 3641 * Return: pointer to the new object or %NULL in case of error 3642 */ 3643 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) 3644 { 3645 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); 3646 3647 trace_kmem_cache_alloc_node(_RET_IP_, ret, 3648 cachep->object_size, cachep->size, 3649 flags, nodeid); 3650 3651 return ret; 3652 } 3653 EXPORT_SYMBOL(kmem_cache_alloc_node); 3654 3655 #ifdef CONFIG_TRACING 3656 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, 3657 gfp_t flags, 3658 int nodeid, 3659 size_t size) 3660 { 3661 void *ret; 3662 3663 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); 3664 3665 ret = kasan_kmalloc(cachep, ret, size, flags); 3666 trace_kmalloc_node(_RET_IP_, ret, 3667 size, cachep->size, 3668 flags, nodeid); 3669 return ret; 3670 } 3671 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 3672 #endif 3673 3674 static __always_inline void * 3675 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) 3676 { 3677 struct kmem_cache *cachep; 3678 void *ret; 3679 3680 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 3681 return NULL; 3682 cachep = kmalloc_slab(size, flags); 3683 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3684 return cachep; 3685 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size); 3686 ret = kasan_kmalloc(cachep, ret, size, flags); 3687 3688 return ret; 3689 } 3690 3691 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3692 { 3693 return __do_kmalloc_node(size, flags, node, _RET_IP_); 3694 } 3695 EXPORT_SYMBOL(__kmalloc_node); 3696 3697 void *__kmalloc_node_track_caller(size_t size, gfp_t flags, 3698 int node, unsigned long caller) 3699 { 3700 return __do_kmalloc_node(size, flags, node, caller); 3701 } 3702 EXPORT_SYMBOL(__kmalloc_node_track_caller); 3703 #endif /* CONFIG_NUMA */ 3704 3705 /** 3706 * __do_kmalloc - allocate memory 3707 * @size: how many bytes of memory are required. 3708 * @flags: the type of memory to allocate (see kmalloc). 3709 * @caller: function caller for debug tracking of the caller 3710 * 3711 * Return: pointer to the allocated memory or %NULL in case of error 3712 */ 3713 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, 3714 unsigned long caller) 3715 { 3716 struct kmem_cache *cachep; 3717 void *ret; 3718 3719 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 3720 return NULL; 3721 cachep = kmalloc_slab(size, flags); 3722 if (unlikely(ZERO_OR_NULL_PTR(cachep))) 3723 return cachep; 3724 ret = slab_alloc(cachep, flags, caller); 3725 3726 ret = kasan_kmalloc(cachep, ret, size, flags); 3727 trace_kmalloc(caller, ret, 3728 size, cachep->size, flags); 3729 3730 return ret; 3731 } 3732 3733 void *__kmalloc(size_t size, gfp_t flags) 3734 { 3735 return __do_kmalloc(size, flags, _RET_IP_); 3736 } 3737 EXPORT_SYMBOL(__kmalloc); 3738 3739 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) 3740 { 3741 return __do_kmalloc(size, flags, caller); 3742 } 3743 EXPORT_SYMBOL(__kmalloc_track_caller); 3744 3745 /** 3746 * kmem_cache_free - Deallocate an object 3747 * @cachep: The cache the allocation was from. 3748 * @objp: The previously allocated object. 3749 * 3750 * Free an object which was previously allocated from this 3751 * cache. 3752 */ 3753 void kmem_cache_free(struct kmem_cache *cachep, void *objp) 3754 { 3755 unsigned long flags; 3756 cachep = cache_from_obj(cachep, objp); 3757 if (!cachep) 3758 return; 3759 3760 local_irq_save(flags); 3761 debug_check_no_locks_freed(objp, cachep->object_size); 3762 if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) 3763 debug_check_no_obj_freed(objp, cachep->object_size); 3764 __cache_free(cachep, objp, _RET_IP_); 3765 local_irq_restore(flags); 3766 3767 trace_kmem_cache_free(_RET_IP_, objp); 3768 } 3769 EXPORT_SYMBOL(kmem_cache_free); 3770 3771 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p) 3772 { 3773 struct kmem_cache *s; 3774 size_t i; 3775 3776 local_irq_disable(); 3777 for (i = 0; i < size; i++) { 3778 void *objp = p[i]; 3779 3780 if (!orig_s) /* called via kfree_bulk */ 3781 s = virt_to_cache(objp); 3782 else 3783 s = cache_from_obj(orig_s, objp); 3784 3785 debug_check_no_locks_freed(objp, s->object_size); 3786 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 3787 debug_check_no_obj_freed(objp, s->object_size); 3788 3789 __cache_free(s, objp, _RET_IP_); 3790 } 3791 local_irq_enable(); 3792 3793 /* FIXME: add tracing */ 3794 } 3795 EXPORT_SYMBOL(kmem_cache_free_bulk); 3796 3797 /** 3798 * kfree - free previously allocated memory 3799 * @objp: pointer returned by kmalloc. 3800 * 3801 * If @objp is NULL, no operation is performed. 3802 * 3803 * Don't free memory not originally allocated by kmalloc() 3804 * or you will run into trouble. 3805 */ 3806 void kfree(const void *objp) 3807 { 3808 struct kmem_cache *c; 3809 unsigned long flags; 3810 3811 trace_kfree(_RET_IP_, objp); 3812 3813 if (unlikely(ZERO_OR_NULL_PTR(objp))) 3814 return; 3815 local_irq_save(flags); 3816 kfree_debugcheck(objp); 3817 c = virt_to_cache(objp); 3818 debug_check_no_locks_freed(objp, c->object_size); 3819 3820 debug_check_no_obj_freed(objp, c->object_size); 3821 __cache_free(c, (void *)objp, _RET_IP_); 3822 local_irq_restore(flags); 3823 } 3824 EXPORT_SYMBOL(kfree); 3825 3826 /* 3827 * This initializes kmem_cache_node or resizes various caches for all nodes. 3828 */ 3829 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp) 3830 { 3831 int ret; 3832 int node; 3833 struct kmem_cache_node *n; 3834 3835 for_each_online_node(node) { 3836 ret = setup_kmem_cache_node(cachep, node, gfp, true); 3837 if (ret) 3838 goto fail; 3839 3840 } 3841 3842 return 0; 3843 3844 fail: 3845 if (!cachep->list.next) { 3846 /* Cache is not active yet. Roll back what we did */ 3847 node--; 3848 while (node >= 0) { 3849 n = get_node(cachep, node); 3850 if (n) { 3851 kfree(n->shared); 3852 free_alien_cache(n->alien); 3853 kfree(n); 3854 cachep->node[node] = NULL; 3855 } 3856 node--; 3857 } 3858 } 3859 return -ENOMEM; 3860 } 3861 3862 /* Always called with the slab_mutex held */ 3863 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit, 3864 int batchcount, int shared, gfp_t gfp) 3865 { 3866 struct array_cache __percpu *cpu_cache, *prev; 3867 int cpu; 3868 3869 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount); 3870 if (!cpu_cache) 3871 return -ENOMEM; 3872 3873 prev = cachep->cpu_cache; 3874 cachep->cpu_cache = cpu_cache; 3875 /* 3876 * Without a previous cpu_cache there's no need to synchronize remote 3877 * cpus, so skip the IPIs. 3878 */ 3879 if (prev) 3880 kick_all_cpus_sync(); 3881 3882 check_irq_on(); 3883 cachep->batchcount = batchcount; 3884 cachep->limit = limit; 3885 cachep->shared = shared; 3886 3887 if (!prev) 3888 goto setup_node; 3889 3890 for_each_online_cpu(cpu) { 3891 LIST_HEAD(list); 3892 int node; 3893 struct kmem_cache_node *n; 3894 struct array_cache *ac = per_cpu_ptr(prev, cpu); 3895 3896 node = cpu_to_mem(cpu); 3897 n = get_node(cachep, node); 3898 spin_lock_irq(&n->list_lock); 3899 free_block(cachep, ac->entry, ac->avail, node, &list); 3900 spin_unlock_irq(&n->list_lock); 3901 slabs_destroy(cachep, &list); 3902 } 3903 free_percpu(prev); 3904 3905 setup_node: 3906 return setup_kmem_cache_nodes(cachep, gfp); 3907 } 3908 3909 static int do_tune_cpucache(struct kmem_cache *cachep, int limit, 3910 int batchcount, int shared, gfp_t gfp) 3911 { 3912 int ret; 3913 struct kmem_cache *c; 3914 3915 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp); 3916 3917 if (slab_state < FULL) 3918 return ret; 3919 3920 if ((ret < 0) || !is_root_cache(cachep)) 3921 return ret; 3922 3923 lockdep_assert_held(&slab_mutex); 3924 for_each_memcg_cache(c, cachep) { 3925 /* return value determined by the root cache only */ 3926 __do_tune_cpucache(c, limit, batchcount, shared, gfp); 3927 } 3928 3929 return ret; 3930 } 3931 3932 /* Called with slab_mutex held always */ 3933 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) 3934 { 3935 int err; 3936 int limit = 0; 3937 int shared = 0; 3938 int batchcount = 0; 3939 3940 err = cache_random_seq_create(cachep, cachep->num, gfp); 3941 if (err) 3942 goto end; 3943 3944 if (!is_root_cache(cachep)) { 3945 struct kmem_cache *root = memcg_root_cache(cachep); 3946 limit = root->limit; 3947 shared = root->shared; 3948 batchcount = root->batchcount; 3949 } 3950 3951 if (limit && shared && batchcount) 3952 goto skip_setup; 3953 /* 3954 * The head array serves three purposes: 3955 * - create a LIFO ordering, i.e. return objects that are cache-warm 3956 * - reduce the number of spinlock operations. 3957 * - reduce the number of linked list operations on the slab and 3958 * bufctl chains: array operations are cheaper. 3959 * The numbers are guessed, we should auto-tune as described by 3960 * Bonwick. 3961 */ 3962 if (cachep->size > 131072) 3963 limit = 1; 3964 else if (cachep->size > PAGE_SIZE) 3965 limit = 8; 3966 else if (cachep->size > 1024) 3967 limit = 24; 3968 else if (cachep->size > 256) 3969 limit = 54; 3970 else 3971 limit = 120; 3972 3973 /* 3974 * CPU bound tasks (e.g. network routing) can exhibit cpu bound 3975 * allocation behaviour: Most allocs on one cpu, most free operations 3976 * on another cpu. For these cases, an efficient object passing between 3977 * cpus is necessary. This is provided by a shared array. The array 3978 * replaces Bonwick's magazine layer. 3979 * On uniprocessor, it's functionally equivalent (but less efficient) 3980 * to a larger limit. Thus disabled by default. 3981 */ 3982 shared = 0; 3983 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) 3984 shared = 8; 3985 3986 #if DEBUG 3987 /* 3988 * With debugging enabled, large batchcount lead to excessively long 3989 * periods with disabled local interrupts. Limit the batchcount 3990 */ 3991 if (limit > 32) 3992 limit = 32; 3993 #endif 3994 batchcount = (limit + 1) / 2; 3995 skip_setup: 3996 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp); 3997 end: 3998 if (err) 3999 pr_err("enable_cpucache failed for %s, error %d\n", 4000 cachep->name, -err); 4001 return err; 4002 } 4003 4004 /* 4005 * Drain an array if it contains any elements taking the node lock only if 4006 * necessary. Note that the node listlock also protects the array_cache 4007 * if drain_array() is used on the shared array. 4008 */ 4009 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n, 4010 struct array_cache *ac, int node) 4011 { 4012 LIST_HEAD(list); 4013 4014 /* ac from n->shared can be freed if we don't hold the slab_mutex. */ 4015 check_mutex_acquired(); 4016 4017 if (!ac || !ac->avail) 4018 return; 4019 4020 if (ac->touched) { 4021 ac->touched = 0; 4022 return; 4023 } 4024 4025 spin_lock_irq(&n->list_lock); 4026 drain_array_locked(cachep, ac, node, false, &list); 4027 spin_unlock_irq(&n->list_lock); 4028 4029 slabs_destroy(cachep, &list); 4030 } 4031 4032 /** 4033 * cache_reap - Reclaim memory from caches. 4034 * @w: work descriptor 4035 * 4036 * Called from workqueue/eventd every few seconds. 4037 * Purpose: 4038 * - clear the per-cpu caches for this CPU. 4039 * - return freeable pages to the main free memory pool. 4040 * 4041 * If we cannot acquire the cache chain mutex then just give up - we'll try 4042 * again on the next iteration. 4043 */ 4044 static void cache_reap(struct work_struct *w) 4045 { 4046 struct kmem_cache *searchp; 4047 struct kmem_cache_node *n; 4048 int node = numa_mem_id(); 4049 struct delayed_work *work = to_delayed_work(w); 4050 4051 if (!mutex_trylock(&slab_mutex)) 4052 /* Give up. Setup the next iteration. */ 4053 goto out; 4054 4055 list_for_each_entry(searchp, &slab_caches, list) { 4056 check_irq_on(); 4057 4058 /* 4059 * We only take the node lock if absolutely necessary and we 4060 * have established with reasonable certainty that 4061 * we can do some work if the lock was obtained. 4062 */ 4063 n = get_node(searchp, node); 4064 4065 reap_alien(searchp, n); 4066 4067 drain_array(searchp, n, cpu_cache_get(searchp), node); 4068 4069 /* 4070 * These are racy checks but it does not matter 4071 * if we skip one check or scan twice. 4072 */ 4073 if (time_after(n->next_reap, jiffies)) 4074 goto next; 4075 4076 n->next_reap = jiffies + REAPTIMEOUT_NODE; 4077 4078 drain_array(searchp, n, n->shared, node); 4079 4080 if (n->free_touched) 4081 n->free_touched = 0; 4082 else { 4083 int freed; 4084 4085 freed = drain_freelist(searchp, n, (n->free_limit + 4086 5 * searchp->num - 1) / (5 * searchp->num)); 4087 STATS_ADD_REAPED(searchp, freed); 4088 } 4089 next: 4090 cond_resched(); 4091 } 4092 check_irq_on(); 4093 mutex_unlock(&slab_mutex); 4094 next_reap_node(); 4095 out: 4096 /* Set up the next iteration */ 4097 schedule_delayed_work_on(smp_processor_id(), work, 4098 round_jiffies_relative(REAPTIMEOUT_AC)); 4099 } 4100 4101 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) 4102 { 4103 unsigned long active_objs, num_objs, active_slabs; 4104 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0; 4105 unsigned long free_slabs = 0; 4106 int node; 4107 struct kmem_cache_node *n; 4108 4109 for_each_kmem_cache_node(cachep, node, n) { 4110 check_irq_on(); 4111 spin_lock_irq(&n->list_lock); 4112 4113 total_slabs += n->total_slabs; 4114 free_slabs += n->free_slabs; 4115 free_objs += n->free_objects; 4116 4117 if (n->shared) 4118 shared_avail += n->shared->avail; 4119 4120 spin_unlock_irq(&n->list_lock); 4121 } 4122 num_objs = total_slabs * cachep->num; 4123 active_slabs = total_slabs - free_slabs; 4124 active_objs = num_objs - free_objs; 4125 4126 sinfo->active_objs = active_objs; 4127 sinfo->num_objs = num_objs; 4128 sinfo->active_slabs = active_slabs; 4129 sinfo->num_slabs = total_slabs; 4130 sinfo->shared_avail = shared_avail; 4131 sinfo->limit = cachep->limit; 4132 sinfo->batchcount = cachep->batchcount; 4133 sinfo->shared = cachep->shared; 4134 sinfo->objects_per_slab = cachep->num; 4135 sinfo->cache_order = cachep->gfporder; 4136 } 4137 4138 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) 4139 { 4140 #if STATS 4141 { /* node stats */ 4142 unsigned long high = cachep->high_mark; 4143 unsigned long allocs = cachep->num_allocations; 4144 unsigned long grown = cachep->grown; 4145 unsigned long reaped = cachep->reaped; 4146 unsigned long errors = cachep->errors; 4147 unsigned long max_freeable = cachep->max_freeable; 4148 unsigned long node_allocs = cachep->node_allocs; 4149 unsigned long node_frees = cachep->node_frees; 4150 unsigned long overflows = cachep->node_overflow; 4151 4152 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu", 4153 allocs, high, grown, 4154 reaped, errors, max_freeable, node_allocs, 4155 node_frees, overflows); 4156 } 4157 /* cpu stats */ 4158 { 4159 unsigned long allochit = atomic_read(&cachep->allochit); 4160 unsigned long allocmiss = atomic_read(&cachep->allocmiss); 4161 unsigned long freehit = atomic_read(&cachep->freehit); 4162 unsigned long freemiss = atomic_read(&cachep->freemiss); 4163 4164 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", 4165 allochit, allocmiss, freehit, freemiss); 4166 } 4167 #endif 4168 } 4169 4170 #define MAX_SLABINFO_WRITE 128 4171 /** 4172 * slabinfo_write - Tuning for the slab allocator 4173 * @file: unused 4174 * @buffer: user buffer 4175 * @count: data length 4176 * @ppos: unused 4177 * 4178 * Return: %0 on success, negative error code otherwise. 4179 */ 4180 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 4181 size_t count, loff_t *ppos) 4182 { 4183 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; 4184 int limit, batchcount, shared, res; 4185 struct kmem_cache *cachep; 4186 4187 if (count > MAX_SLABINFO_WRITE) 4188 return -EINVAL; 4189 if (copy_from_user(&kbuf, buffer, count)) 4190 return -EFAULT; 4191 kbuf[MAX_SLABINFO_WRITE] = '\0'; 4192 4193 tmp = strchr(kbuf, ' '); 4194 if (!tmp) 4195 return -EINVAL; 4196 *tmp = '\0'; 4197 tmp++; 4198 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) 4199 return -EINVAL; 4200 4201 /* Find the cache in the chain of caches. */ 4202 mutex_lock(&slab_mutex); 4203 res = -EINVAL; 4204 list_for_each_entry(cachep, &slab_caches, list) { 4205 if (!strcmp(cachep->name, kbuf)) { 4206 if (limit < 1 || batchcount < 1 || 4207 batchcount > limit || shared < 0) { 4208 res = 0; 4209 } else { 4210 res = do_tune_cpucache(cachep, limit, 4211 batchcount, shared, 4212 GFP_KERNEL); 4213 } 4214 break; 4215 } 4216 } 4217 mutex_unlock(&slab_mutex); 4218 if (res >= 0) 4219 res = count; 4220 return res; 4221 } 4222 4223 #ifdef CONFIG_DEBUG_SLAB_LEAK 4224 4225 static inline int add_caller(unsigned long *n, unsigned long v) 4226 { 4227 unsigned long *p; 4228 int l; 4229 if (!v) 4230 return 1; 4231 l = n[1]; 4232 p = n + 2; 4233 while (l) { 4234 int i = l/2; 4235 unsigned long *q = p + 2 * i; 4236 if (*q == v) { 4237 q[1]++; 4238 return 1; 4239 } 4240 if (*q > v) { 4241 l = i; 4242 } else { 4243 p = q + 2; 4244 l -= i + 1; 4245 } 4246 } 4247 if (++n[1] == n[0]) 4248 return 0; 4249 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); 4250 p[0] = v; 4251 p[1] = 1; 4252 return 1; 4253 } 4254 4255 static void handle_slab(unsigned long *n, struct kmem_cache *c, 4256 struct page *page) 4257 { 4258 void *p; 4259 int i, j; 4260 unsigned long v; 4261 4262 if (n[0] == n[1]) 4263 return; 4264 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) { 4265 bool active = true; 4266 4267 for (j = page->active; j < c->num; j++) { 4268 if (get_free_obj(page, j) == i) { 4269 active = false; 4270 break; 4271 } 4272 } 4273 4274 if (!active) 4275 continue; 4276 4277 /* 4278 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table 4279 * mapping is established when actual object allocation and 4280 * we could mistakenly access the unmapped object in the cpu 4281 * cache. 4282 */ 4283 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v))) 4284 continue; 4285 4286 if (!add_caller(n, v)) 4287 return; 4288 } 4289 } 4290 4291 static void show_symbol(struct seq_file *m, unsigned long address) 4292 { 4293 #ifdef CONFIG_KALLSYMS 4294 unsigned long offset, size; 4295 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; 4296 4297 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { 4298 seq_printf(m, "%s+%#lx/%#lx", name, offset, size); 4299 if (modname[0]) 4300 seq_printf(m, " [%s]", modname); 4301 return; 4302 } 4303 #endif 4304 seq_printf(m, "%px", (void *)address); 4305 } 4306 4307 static int leaks_show(struct seq_file *m, void *p) 4308 { 4309 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); 4310 struct page *page; 4311 struct kmem_cache_node *n; 4312 const char *name; 4313 unsigned long *x = m->private; 4314 int node; 4315 int i; 4316 4317 if (!(cachep->flags & SLAB_STORE_USER)) 4318 return 0; 4319 if (!(cachep->flags & SLAB_RED_ZONE)) 4320 return 0; 4321 4322 /* 4323 * Set store_user_clean and start to grab stored user information 4324 * for all objects on this cache. If some alloc/free requests comes 4325 * during the processing, information would be wrong so restart 4326 * whole processing. 4327 */ 4328 do { 4329 set_store_user_clean(cachep); 4330 drain_cpu_caches(cachep); 4331 4332 x[1] = 0; 4333 4334 for_each_kmem_cache_node(cachep, node, n) { 4335 4336 check_irq_on(); 4337 spin_lock_irq(&n->list_lock); 4338 4339 list_for_each_entry(page, &n->slabs_full, lru) 4340 handle_slab(x, cachep, page); 4341 list_for_each_entry(page, &n->slabs_partial, lru) 4342 handle_slab(x, cachep, page); 4343 spin_unlock_irq(&n->list_lock); 4344 } 4345 } while (!is_store_user_clean(cachep)); 4346 4347 name = cachep->name; 4348 if (x[0] == x[1]) { 4349 /* Increase the buffer size */ 4350 mutex_unlock(&slab_mutex); 4351 m->private = kcalloc(x[0] * 4, sizeof(unsigned long), 4352 GFP_KERNEL); 4353 if (!m->private) { 4354 /* Too bad, we are really out */ 4355 m->private = x; 4356 mutex_lock(&slab_mutex); 4357 return -ENOMEM; 4358 } 4359 *(unsigned long *)m->private = x[0] * 2; 4360 kfree(x); 4361 mutex_lock(&slab_mutex); 4362 /* Now make sure this entry will be retried */ 4363 m->count = m->size; 4364 return 0; 4365 } 4366 for (i = 0; i < x[1]; i++) { 4367 seq_printf(m, "%s: %lu ", name, x[2*i+3]); 4368 show_symbol(m, x[2*i+2]); 4369 seq_putc(m, '\n'); 4370 } 4371 4372 return 0; 4373 } 4374 4375 static const struct seq_operations slabstats_op = { 4376 .start = slab_start, 4377 .next = slab_next, 4378 .stop = slab_stop, 4379 .show = leaks_show, 4380 }; 4381 4382 static int slabstats_open(struct inode *inode, struct file *file) 4383 { 4384 unsigned long *n; 4385 4386 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE); 4387 if (!n) 4388 return -ENOMEM; 4389 4390 *n = PAGE_SIZE / (2 * sizeof(unsigned long)); 4391 4392 return 0; 4393 } 4394 4395 static const struct file_operations proc_slabstats_operations = { 4396 .open = slabstats_open, 4397 .read = seq_read, 4398 .llseek = seq_lseek, 4399 .release = seq_release_private, 4400 }; 4401 #endif 4402 4403 static int __init slab_proc_init(void) 4404 { 4405 #ifdef CONFIG_DEBUG_SLAB_LEAK 4406 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); 4407 #endif 4408 return 0; 4409 } 4410 module_init(slab_proc_init); 4411 4412 #ifdef CONFIG_HARDENED_USERCOPY 4413 /* 4414 * Rejects incorrectly sized objects and objects that are to be copied 4415 * to/from userspace but do not fall entirely within the containing slab 4416 * cache's usercopy region. 4417 * 4418 * Returns NULL if check passes, otherwise const char * to name of cache 4419 * to indicate an error. 4420 */ 4421 void __check_heap_object(const void *ptr, unsigned long n, struct page *page, 4422 bool to_user) 4423 { 4424 struct kmem_cache *cachep; 4425 unsigned int objnr; 4426 unsigned long offset; 4427 4428 ptr = kasan_reset_tag(ptr); 4429 4430 /* Find and validate object. */ 4431 cachep = page->slab_cache; 4432 objnr = obj_to_index(cachep, page, (void *)ptr); 4433 BUG_ON(objnr >= cachep->num); 4434 4435 /* Find offset within object. */ 4436 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep); 4437 4438 /* Allow address range falling entirely within usercopy region. */ 4439 if (offset >= cachep->useroffset && 4440 offset - cachep->useroffset <= cachep->usersize && 4441 n <= cachep->useroffset - offset + cachep->usersize) 4442 return; 4443 4444 /* 4445 * If the copy is still within the allocated object, produce 4446 * a warning instead of rejecting the copy. This is intended 4447 * to be a temporary method to find any missing usercopy 4448 * whitelists. 4449 */ 4450 if (usercopy_fallback && 4451 offset <= cachep->object_size && 4452 n <= cachep->object_size - offset) { 4453 usercopy_warn("SLAB object", cachep->name, to_user, offset, n); 4454 return; 4455 } 4456 4457 usercopy_abort("SLAB object", cachep->name, to_user, offset, n); 4458 } 4459 #endif /* CONFIG_HARDENED_USERCOPY */ 4460 4461 /** 4462 * ksize - get the actual amount of memory allocated for a given object 4463 * @objp: Pointer to the object 4464 * 4465 * kmalloc may internally round up allocations and return more memory 4466 * than requested. ksize() can be used to determine the actual amount of 4467 * memory allocated. The caller may use this additional memory, even though 4468 * a smaller amount of memory was initially specified with the kmalloc call. 4469 * The caller must guarantee that objp points to a valid object previously 4470 * allocated with either kmalloc() or kmem_cache_alloc(). The object 4471 * must not be freed during the duration of the call. 4472 * 4473 * Return: size of the actual memory used by @objp in bytes 4474 */ 4475 size_t ksize(const void *objp) 4476 { 4477 size_t size; 4478 4479 BUG_ON(!objp); 4480 if (unlikely(objp == ZERO_SIZE_PTR)) 4481 return 0; 4482 4483 size = virt_to_cache(objp)->object_size; 4484 /* We assume that ksize callers could use the whole allocated area, 4485 * so we need to unpoison this area. 4486 */ 4487 kasan_unpoison_shadow(objp, size); 4488 4489 return size; 4490 } 4491 EXPORT_SYMBOL(ksize); 4492