1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * SLUB: A slab allocator that limits cache line use instead of queuing 4 * objects in per cpu and per node lists. 5 * 6 * The allocator synchronizes using per slab locks or atomic operations 7 * and only uses a centralized lock to manage a pool of partial slabs. 8 * 9 * (C) 2007 SGI, Christoph Lameter 10 * (C) 2011 Linux Foundation, Christoph Lameter 11 */ 12 13 #include <linux/mm.h> 14 #include <linux/swap.h> /* struct reclaim_state */ 15 #include <linux/module.h> 16 #include <linux/bit_spinlock.h> 17 #include <linux/interrupt.h> 18 #include <linux/swab.h> 19 #include <linux/bitops.h> 20 #include <linux/slab.h> 21 #include "slab.h" 22 #include <linux/proc_fs.h> 23 #include <linux/seq_file.h> 24 #include <linux/kasan.h> 25 #include <linux/cpu.h> 26 #include <linux/cpuset.h> 27 #include <linux/mempolicy.h> 28 #include <linux/ctype.h> 29 #include <linux/debugobjects.h> 30 #include <linux/kallsyms.h> 31 #include <linux/kfence.h> 32 #include <linux/memory.h> 33 #include <linux/math64.h> 34 #include <linux/fault-inject.h> 35 #include <linux/stacktrace.h> 36 #include <linux/prefetch.h> 37 #include <linux/memcontrol.h> 38 #include <linux/random.h> 39 40 #include <trace/events/kmem.h> 41 42 #include "internal.h" 43 44 /* 45 * Lock order: 46 * 1. slab_mutex (Global Mutex) 47 * 2. node->list_lock 48 * 3. slab_lock(page) (Only on some arches and for debugging) 49 * 50 * slab_mutex 51 * 52 * The role of the slab_mutex is to protect the list of all the slabs 53 * and to synchronize major metadata changes to slab cache structures. 54 * 55 * The slab_lock is only used for debugging and on arches that do not 56 * have the ability to do a cmpxchg_double. It only protects: 57 * A. page->freelist -> List of object free in a page 58 * B. page->inuse -> Number of objects in use 59 * C. page->objects -> Number of objects in page 60 * D. page->frozen -> frozen state 61 * 62 * If a slab is frozen then it is exempt from list management. It is not 63 * on any list except per cpu partial list. The processor that froze the 64 * slab is the one who can perform list operations on the page. Other 65 * processors may put objects onto the freelist but the processor that 66 * froze the slab is the only one that can retrieve the objects from the 67 * page's freelist. 68 * 69 * The list_lock protects the partial and full list on each node and 70 * the partial slab counter. If taken then no new slabs may be added or 71 * removed from the lists nor make the number of partial slabs be modified. 72 * (Note that the total number of slabs is an atomic value that may be 73 * modified without taking the list lock). 74 * 75 * The list_lock is a centralized lock and thus we avoid taking it as 76 * much as possible. As long as SLUB does not have to handle partial 77 * slabs, operations can continue without any centralized lock. F.e. 78 * allocating a long series of objects that fill up slabs does not require 79 * the list lock. 80 * Interrupts are disabled during allocation and deallocation in order to 81 * make the slab allocator safe to use in the context of an irq. In addition 82 * interrupts are disabled to ensure that the processor does not change 83 * while handling per_cpu slabs, due to kernel preemption. 84 * 85 * SLUB assigns one slab for allocation to each processor. 86 * Allocations only occur from these slabs called cpu slabs. 87 * 88 * Slabs with free elements are kept on a partial list and during regular 89 * operations no list for full slabs is used. If an object in a full slab is 90 * freed then the slab will show up again on the partial lists. 91 * We track full slabs for debugging purposes though because otherwise we 92 * cannot scan all objects. 93 * 94 * Slabs are freed when they become empty. Teardown and setup is 95 * minimal so we rely on the page allocators per cpu caches for 96 * fast frees and allocs. 97 * 98 * page->frozen The slab is frozen and exempt from list processing. 99 * This means that the slab is dedicated to a purpose 100 * such as satisfying allocations for a specific 101 * processor. Objects may be freed in the slab while 102 * it is frozen but slab_free will then skip the usual 103 * list operations. It is up to the processor holding 104 * the slab to integrate the slab into the slab lists 105 * when the slab is no longer needed. 106 * 107 * One use of this flag is to mark slabs that are 108 * used for allocations. Then such a slab becomes a cpu 109 * slab. The cpu slab may be equipped with an additional 110 * freelist that allows lockless access to 111 * free objects in addition to the regular freelist 112 * that requires the slab lock. 113 * 114 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug 115 * options set. This moves slab handling out of 116 * the fast path and disables lockless freelists. 117 */ 118 119 #ifdef CONFIG_SLUB_DEBUG 120 #ifdef CONFIG_SLUB_DEBUG_ON 121 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled); 122 #else 123 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled); 124 #endif 125 #endif 126 127 static inline bool kmem_cache_debug(struct kmem_cache *s) 128 { 129 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS); 130 } 131 132 void *fixup_red_left(struct kmem_cache *s, void *p) 133 { 134 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) 135 p += s->red_left_pad; 136 137 return p; 138 } 139 140 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) 141 { 142 #ifdef CONFIG_SLUB_CPU_PARTIAL 143 return !kmem_cache_debug(s); 144 #else 145 return false; 146 #endif 147 } 148 149 /* 150 * Issues still to be resolved: 151 * 152 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 153 * 154 * - Variable sizing of the per node arrays 155 */ 156 157 /* Enable to test recovery from slab corruption on boot */ 158 #undef SLUB_RESILIENCY_TEST 159 160 /* Enable to log cmpxchg failures */ 161 #undef SLUB_DEBUG_CMPXCHG 162 163 /* 164 * Minimum number of partial slabs. These will be left on the partial 165 * lists even if they are empty. kmem_cache_shrink may reclaim them. 166 */ 167 #define MIN_PARTIAL 5 168 169 /* 170 * Maximum number of desirable partial slabs. 171 * The existence of more partial slabs makes kmem_cache_shrink 172 * sort the partial list by the number of objects in use. 173 */ 174 #define MAX_PARTIAL 10 175 176 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ 177 SLAB_POISON | SLAB_STORE_USER) 178 179 /* 180 * These debug flags cannot use CMPXCHG because there might be consistency 181 * issues when checking or reading debug information 182 */ 183 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ 184 SLAB_TRACE) 185 186 187 /* 188 * Debugging flags that require metadata to be stored in the slab. These get 189 * disabled when slub_debug=O is used and a cache's min order increases with 190 * metadata. 191 */ 192 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 193 194 #define OO_SHIFT 16 195 #define OO_MASK ((1 << OO_SHIFT) - 1) 196 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ 197 198 /* Internal SLUB flags */ 199 /* Poison object */ 200 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U) 201 /* Use cmpxchg_double */ 202 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U) 203 204 /* 205 * Tracking user of a slab. 206 */ 207 #define TRACK_ADDRS_COUNT 16 208 struct track { 209 unsigned long addr; /* Called from address */ 210 #ifdef CONFIG_STACKTRACE 211 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 212 #endif 213 int cpu; /* Was running on cpu */ 214 int pid; /* Pid context */ 215 unsigned long when; /* When did the operation occur */ 216 }; 217 218 enum track_item { TRACK_ALLOC, TRACK_FREE }; 219 220 #ifdef CONFIG_SYSFS 221 static int sysfs_slab_add(struct kmem_cache *); 222 static int sysfs_slab_alias(struct kmem_cache *, const char *); 223 #else 224 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 225 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 226 { return 0; } 227 #endif 228 229 static inline void stat(const struct kmem_cache *s, enum stat_item si) 230 { 231 #ifdef CONFIG_SLUB_STATS 232 /* 233 * The rmw is racy on a preemptible kernel but this is acceptable, so 234 * avoid this_cpu_add()'s irq-disable overhead. 235 */ 236 raw_cpu_inc(s->cpu_slab->stat[si]); 237 #endif 238 } 239 240 /* 241 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated. 242 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily 243 * differ during memory hotplug/hotremove operations. 244 * Protected by slab_mutex. 245 */ 246 static nodemask_t slab_nodes; 247 248 /******************************************************************** 249 * Core slab cache functions 250 *******************************************************************/ 251 252 /* 253 * Returns freelist pointer (ptr). With hardening, this is obfuscated 254 * with an XOR of the address where the pointer is held and a per-cache 255 * random number. 256 */ 257 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr, 258 unsigned long ptr_addr) 259 { 260 #ifdef CONFIG_SLAB_FREELIST_HARDENED 261 /* 262 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged. 263 * Normally, this doesn't cause any issues, as both set_freepointer() 264 * and get_freepointer() are called with a pointer with the same tag. 265 * However, there are some issues with CONFIG_SLUB_DEBUG code. For 266 * example, when __free_slub() iterates over objects in a cache, it 267 * passes untagged pointers to check_object(). check_object() in turns 268 * calls get_freepointer() with an untagged pointer, which causes the 269 * freepointer to be restored incorrectly. 270 */ 271 return (void *)((unsigned long)ptr ^ s->random ^ 272 swab((unsigned long)kasan_reset_tag((void *)ptr_addr))); 273 #else 274 return ptr; 275 #endif 276 } 277 278 /* Returns the freelist pointer recorded at location ptr_addr. */ 279 static inline void *freelist_dereference(const struct kmem_cache *s, 280 void *ptr_addr) 281 { 282 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr), 283 (unsigned long)ptr_addr); 284 } 285 286 static inline void *get_freepointer(struct kmem_cache *s, void *object) 287 { 288 object = kasan_reset_tag(object); 289 return freelist_dereference(s, object + s->offset); 290 } 291 292 static void prefetch_freepointer(const struct kmem_cache *s, void *object) 293 { 294 prefetch(object + s->offset); 295 } 296 297 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 298 { 299 unsigned long freepointer_addr; 300 void *p; 301 302 if (!debug_pagealloc_enabled_static()) 303 return get_freepointer(s, object); 304 305 object = kasan_reset_tag(object); 306 freepointer_addr = (unsigned long)object + s->offset; 307 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p)); 308 return freelist_ptr(s, p, freepointer_addr); 309 } 310 311 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 312 { 313 unsigned long freeptr_addr = (unsigned long)object + s->offset; 314 315 #ifdef CONFIG_SLAB_FREELIST_HARDENED 316 BUG_ON(object == fp); /* naive detection of double free or corruption */ 317 #endif 318 319 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr); 320 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr); 321 } 322 323 /* Loop over all objects in a slab */ 324 #define for_each_object(__p, __s, __addr, __objects) \ 325 for (__p = fixup_red_left(__s, __addr); \ 326 __p < (__addr) + (__objects) * (__s)->size; \ 327 __p += (__s)->size) 328 329 static inline unsigned int order_objects(unsigned int order, unsigned int size) 330 { 331 return ((unsigned int)PAGE_SIZE << order) / size; 332 } 333 334 static inline struct kmem_cache_order_objects oo_make(unsigned int order, 335 unsigned int size) 336 { 337 struct kmem_cache_order_objects x = { 338 (order << OO_SHIFT) + order_objects(order, size) 339 }; 340 341 return x; 342 } 343 344 static inline unsigned int oo_order(struct kmem_cache_order_objects x) 345 { 346 return x.x >> OO_SHIFT; 347 } 348 349 static inline unsigned int oo_objects(struct kmem_cache_order_objects x) 350 { 351 return x.x & OO_MASK; 352 } 353 354 /* 355 * Per slab locking using the pagelock 356 */ 357 static __always_inline void slab_lock(struct page *page) 358 { 359 VM_BUG_ON_PAGE(PageTail(page), page); 360 bit_spin_lock(PG_locked, &page->flags); 361 } 362 363 static __always_inline void slab_unlock(struct page *page) 364 { 365 VM_BUG_ON_PAGE(PageTail(page), page); 366 __bit_spin_unlock(PG_locked, &page->flags); 367 } 368 369 /* Interrupts must be disabled (for the fallback code to work right) */ 370 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 371 void *freelist_old, unsigned long counters_old, 372 void *freelist_new, unsigned long counters_new, 373 const char *n) 374 { 375 VM_BUG_ON(!irqs_disabled()); 376 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 377 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 378 if (s->flags & __CMPXCHG_DOUBLE) { 379 if (cmpxchg_double(&page->freelist, &page->counters, 380 freelist_old, counters_old, 381 freelist_new, counters_new)) 382 return true; 383 } else 384 #endif 385 { 386 slab_lock(page); 387 if (page->freelist == freelist_old && 388 page->counters == counters_old) { 389 page->freelist = freelist_new; 390 page->counters = counters_new; 391 slab_unlock(page); 392 return true; 393 } 394 slab_unlock(page); 395 } 396 397 cpu_relax(); 398 stat(s, CMPXCHG_DOUBLE_FAIL); 399 400 #ifdef SLUB_DEBUG_CMPXCHG 401 pr_info("%s %s: cmpxchg double redo ", n, s->name); 402 #endif 403 404 return false; 405 } 406 407 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 408 void *freelist_old, unsigned long counters_old, 409 void *freelist_new, unsigned long counters_new, 410 const char *n) 411 { 412 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 413 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 414 if (s->flags & __CMPXCHG_DOUBLE) { 415 if (cmpxchg_double(&page->freelist, &page->counters, 416 freelist_old, counters_old, 417 freelist_new, counters_new)) 418 return true; 419 } else 420 #endif 421 { 422 unsigned long flags; 423 424 local_irq_save(flags); 425 slab_lock(page); 426 if (page->freelist == freelist_old && 427 page->counters == counters_old) { 428 page->freelist = freelist_new; 429 page->counters = counters_new; 430 slab_unlock(page); 431 local_irq_restore(flags); 432 return true; 433 } 434 slab_unlock(page); 435 local_irq_restore(flags); 436 } 437 438 cpu_relax(); 439 stat(s, CMPXCHG_DOUBLE_FAIL); 440 441 #ifdef SLUB_DEBUG_CMPXCHG 442 pr_info("%s %s: cmpxchg double redo ", n, s->name); 443 #endif 444 445 return false; 446 } 447 448 #ifdef CONFIG_SLUB_DEBUG 449 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)]; 450 static DEFINE_SPINLOCK(object_map_lock); 451 452 /* 453 * Determine a map of object in use on a page. 454 * 455 * Node listlock must be held to guarantee that the page does 456 * not vanish from under us. 457 */ 458 static unsigned long *get_map(struct kmem_cache *s, struct page *page) 459 __acquires(&object_map_lock) 460 { 461 void *p; 462 void *addr = page_address(page); 463 464 VM_BUG_ON(!irqs_disabled()); 465 466 spin_lock(&object_map_lock); 467 468 bitmap_zero(object_map, page->objects); 469 470 for (p = page->freelist; p; p = get_freepointer(s, p)) 471 set_bit(__obj_to_index(s, addr, p), object_map); 472 473 return object_map; 474 } 475 476 static void put_map(unsigned long *map) __releases(&object_map_lock) 477 { 478 VM_BUG_ON(map != object_map); 479 spin_unlock(&object_map_lock); 480 } 481 482 static inline unsigned int size_from_object(struct kmem_cache *s) 483 { 484 if (s->flags & SLAB_RED_ZONE) 485 return s->size - s->red_left_pad; 486 487 return s->size; 488 } 489 490 static inline void *restore_red_left(struct kmem_cache *s, void *p) 491 { 492 if (s->flags & SLAB_RED_ZONE) 493 p -= s->red_left_pad; 494 495 return p; 496 } 497 498 /* 499 * Debug settings: 500 */ 501 #if defined(CONFIG_SLUB_DEBUG_ON) 502 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; 503 #else 504 static slab_flags_t slub_debug; 505 #endif 506 507 static char *slub_debug_string; 508 static int disable_higher_order_debug; 509 510 /* 511 * slub is about to manipulate internal object metadata. This memory lies 512 * outside the range of the allocated object, so accessing it would normally 513 * be reported by kasan as a bounds error. metadata_access_enable() is used 514 * to tell kasan that these accesses are OK. 515 */ 516 static inline void metadata_access_enable(void) 517 { 518 kasan_disable_current(); 519 } 520 521 static inline void metadata_access_disable(void) 522 { 523 kasan_enable_current(); 524 } 525 526 /* 527 * Object debugging 528 */ 529 530 /* Verify that a pointer has an address that is valid within a slab page */ 531 static inline int check_valid_pointer(struct kmem_cache *s, 532 struct page *page, void *object) 533 { 534 void *base; 535 536 if (!object) 537 return 1; 538 539 base = page_address(page); 540 object = kasan_reset_tag(object); 541 object = restore_red_left(s, object); 542 if (object < base || object >= base + page->objects * s->size || 543 (object - base) % s->size) { 544 return 0; 545 } 546 547 return 1; 548 } 549 550 static void print_section(char *level, char *text, u8 *addr, 551 unsigned int length) 552 { 553 metadata_access_enable(); 554 print_hex_dump(level, kasan_reset_tag(text), DUMP_PREFIX_ADDRESS, 555 16, 1, addr, length, 1); 556 metadata_access_disable(); 557 } 558 559 /* 560 * See comment in calculate_sizes(). 561 */ 562 static inline bool freeptr_outside_object(struct kmem_cache *s) 563 { 564 return s->offset >= s->inuse; 565 } 566 567 /* 568 * Return offset of the end of info block which is inuse + free pointer if 569 * not overlapping with object. 570 */ 571 static inline unsigned int get_info_end(struct kmem_cache *s) 572 { 573 if (freeptr_outside_object(s)) 574 return s->inuse + sizeof(void *); 575 else 576 return s->inuse; 577 } 578 579 static struct track *get_track(struct kmem_cache *s, void *object, 580 enum track_item alloc) 581 { 582 struct track *p; 583 584 p = object + get_info_end(s); 585 586 return kasan_reset_tag(p + alloc); 587 } 588 589 static void set_track(struct kmem_cache *s, void *object, 590 enum track_item alloc, unsigned long addr) 591 { 592 struct track *p = get_track(s, object, alloc); 593 594 if (addr) { 595 #ifdef CONFIG_STACKTRACE 596 unsigned int nr_entries; 597 598 metadata_access_enable(); 599 nr_entries = stack_trace_save(kasan_reset_tag(p->addrs), 600 TRACK_ADDRS_COUNT, 3); 601 metadata_access_disable(); 602 603 if (nr_entries < TRACK_ADDRS_COUNT) 604 p->addrs[nr_entries] = 0; 605 #endif 606 p->addr = addr; 607 p->cpu = smp_processor_id(); 608 p->pid = current->pid; 609 p->when = jiffies; 610 } else { 611 memset(p, 0, sizeof(struct track)); 612 } 613 } 614 615 static void init_tracking(struct kmem_cache *s, void *object) 616 { 617 if (!(s->flags & SLAB_STORE_USER)) 618 return; 619 620 set_track(s, object, TRACK_FREE, 0UL); 621 set_track(s, object, TRACK_ALLOC, 0UL); 622 } 623 624 static void print_track(const char *s, struct track *t, unsigned long pr_time) 625 { 626 if (!t->addr) 627 return; 628 629 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n", 630 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid); 631 #ifdef CONFIG_STACKTRACE 632 { 633 int i; 634 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 635 if (t->addrs[i]) 636 pr_err("\t%pS\n", (void *)t->addrs[i]); 637 else 638 break; 639 } 640 #endif 641 } 642 643 void print_tracking(struct kmem_cache *s, void *object) 644 { 645 unsigned long pr_time = jiffies; 646 if (!(s->flags & SLAB_STORE_USER)) 647 return; 648 649 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time); 650 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time); 651 } 652 653 static void print_page_info(struct page *page) 654 { 655 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n", 656 page, page->objects, page->inuse, page->freelist, 657 page->flags, &page->flags); 658 659 } 660 661 static void slab_bug(struct kmem_cache *s, char *fmt, ...) 662 { 663 struct va_format vaf; 664 va_list args; 665 666 va_start(args, fmt); 667 vaf.fmt = fmt; 668 vaf.va = &args; 669 pr_err("=============================================================================\n"); 670 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf); 671 pr_err("-----------------------------------------------------------------------------\n\n"); 672 673 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 674 va_end(args); 675 } 676 677 static void slab_fix(struct kmem_cache *s, char *fmt, ...) 678 { 679 struct va_format vaf; 680 va_list args; 681 682 va_start(args, fmt); 683 vaf.fmt = fmt; 684 vaf.va = &args; 685 pr_err("FIX %s: %pV\n", s->name, &vaf); 686 va_end(args); 687 } 688 689 static bool freelist_corrupted(struct kmem_cache *s, struct page *page, 690 void **freelist, void *nextfree) 691 { 692 if ((s->flags & SLAB_CONSISTENCY_CHECKS) && 693 !check_valid_pointer(s, page, nextfree) && freelist) { 694 object_err(s, page, *freelist, "Freechain corrupt"); 695 *freelist = NULL; 696 slab_fix(s, "Isolate corrupted freechain"); 697 return true; 698 } 699 700 return false; 701 } 702 703 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 704 { 705 unsigned int off; /* Offset of last byte */ 706 u8 *addr = page_address(page); 707 708 print_tracking(s, p); 709 710 print_page_info(page); 711 712 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n", 713 p, p - addr, get_freepointer(s, p)); 714 715 if (s->flags & SLAB_RED_ZONE) 716 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad, 717 s->red_left_pad); 718 else if (p > addr + 16) 719 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16); 720 721 print_section(KERN_ERR, "Object ", p, 722 min_t(unsigned int, s->object_size, PAGE_SIZE)); 723 if (s->flags & SLAB_RED_ZONE) 724 print_section(KERN_ERR, "Redzone ", p + s->object_size, 725 s->inuse - s->object_size); 726 727 off = get_info_end(s); 728 729 if (s->flags & SLAB_STORE_USER) 730 off += 2 * sizeof(struct track); 731 732 off += kasan_metadata_size(s); 733 734 if (off != size_from_object(s)) 735 /* Beginning of the filler is the free pointer */ 736 print_section(KERN_ERR, "Padding ", p + off, 737 size_from_object(s) - off); 738 739 dump_stack(); 740 } 741 742 void object_err(struct kmem_cache *s, struct page *page, 743 u8 *object, char *reason) 744 { 745 slab_bug(s, "%s", reason); 746 print_trailer(s, page, object); 747 } 748 749 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page, 750 const char *fmt, ...) 751 { 752 va_list args; 753 char buf[100]; 754 755 va_start(args, fmt); 756 vsnprintf(buf, sizeof(buf), fmt, args); 757 va_end(args); 758 slab_bug(s, "%s", buf); 759 print_page_info(page); 760 dump_stack(); 761 } 762 763 static void init_object(struct kmem_cache *s, void *object, u8 val) 764 { 765 u8 *p = kasan_reset_tag(object); 766 767 if (s->flags & SLAB_RED_ZONE) 768 memset(p - s->red_left_pad, val, s->red_left_pad); 769 770 if (s->flags & __OBJECT_POISON) { 771 memset(p, POISON_FREE, s->object_size - 1); 772 p[s->object_size - 1] = POISON_END; 773 } 774 775 if (s->flags & SLAB_RED_ZONE) 776 memset(p + s->object_size, val, s->inuse - s->object_size); 777 } 778 779 static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 780 void *from, void *to) 781 { 782 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 783 memset(from, data, to - from); 784 } 785 786 static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 787 u8 *object, char *what, 788 u8 *start, unsigned int value, unsigned int bytes) 789 { 790 u8 *fault; 791 u8 *end; 792 u8 *addr = page_address(page); 793 794 metadata_access_enable(); 795 fault = memchr_inv(kasan_reset_tag(start), value, bytes); 796 metadata_access_disable(); 797 if (!fault) 798 return 1; 799 800 end = start + bytes; 801 while (end > fault && end[-1] == value) 802 end--; 803 804 slab_bug(s, "%s overwritten", what); 805 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n", 806 fault, end - 1, fault - addr, 807 fault[0], value); 808 print_trailer(s, page, object); 809 810 restore_bytes(s, what, value, fault, end); 811 return 0; 812 } 813 814 /* 815 * Object layout: 816 * 817 * object address 818 * Bytes of the object to be managed. 819 * If the freepointer may overlay the object then the free 820 * pointer is at the middle of the object. 821 * 822 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 823 * 0xa5 (POISON_END) 824 * 825 * object + s->object_size 826 * Padding to reach word boundary. This is also used for Redzoning. 827 * Padding is extended by another word if Redzoning is enabled and 828 * object_size == inuse. 829 * 830 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 831 * 0xcc (RED_ACTIVE) for objects in use. 832 * 833 * object + s->inuse 834 * Meta data starts here. 835 * 836 * A. Free pointer (if we cannot overwrite object on free) 837 * B. Tracking data for SLAB_STORE_USER 838 * C. Padding to reach required alignment boundary or at minimum 839 * one word if debugging is on to be able to detect writes 840 * before the word boundary. 841 * 842 * Padding is done using 0x5a (POISON_INUSE) 843 * 844 * object + s->size 845 * Nothing is used beyond s->size. 846 * 847 * If slabcaches are merged then the object_size and inuse boundaries are mostly 848 * ignored. And therefore no slab options that rely on these boundaries 849 * may be used with merged slabcaches. 850 */ 851 852 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 853 { 854 unsigned long off = get_info_end(s); /* The end of info */ 855 856 if (s->flags & SLAB_STORE_USER) 857 /* We also have user information there */ 858 off += 2 * sizeof(struct track); 859 860 off += kasan_metadata_size(s); 861 862 if (size_from_object(s) == off) 863 return 1; 864 865 return check_bytes_and_report(s, page, p, "Object padding", 866 p + off, POISON_INUSE, size_from_object(s) - off); 867 } 868 869 /* Check the pad bytes at the end of a slab page */ 870 static int slab_pad_check(struct kmem_cache *s, struct page *page) 871 { 872 u8 *start; 873 u8 *fault; 874 u8 *end; 875 u8 *pad; 876 int length; 877 int remainder; 878 879 if (!(s->flags & SLAB_POISON)) 880 return 1; 881 882 start = page_address(page); 883 length = page_size(page); 884 end = start + length; 885 remainder = length % s->size; 886 if (!remainder) 887 return 1; 888 889 pad = end - remainder; 890 metadata_access_enable(); 891 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder); 892 metadata_access_disable(); 893 if (!fault) 894 return 1; 895 while (end > fault && end[-1] == POISON_INUSE) 896 end--; 897 898 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu", 899 fault, end - 1, fault - start); 900 print_section(KERN_ERR, "Padding ", pad, remainder); 901 902 restore_bytes(s, "slab padding", POISON_INUSE, fault, end); 903 return 0; 904 } 905 906 static int check_object(struct kmem_cache *s, struct page *page, 907 void *object, u8 val) 908 { 909 u8 *p = object; 910 u8 *endobject = object + s->object_size; 911 912 if (s->flags & SLAB_RED_ZONE) { 913 if (!check_bytes_and_report(s, page, object, "Left Redzone", 914 object - s->red_left_pad, val, s->red_left_pad)) 915 return 0; 916 917 if (!check_bytes_and_report(s, page, object, "Right Redzone", 918 endobject, val, s->inuse - s->object_size)) 919 return 0; 920 } else { 921 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { 922 check_bytes_and_report(s, page, p, "Alignment padding", 923 endobject, POISON_INUSE, 924 s->inuse - s->object_size); 925 } 926 } 927 928 if (s->flags & SLAB_POISON) { 929 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 930 (!check_bytes_and_report(s, page, p, "Poison", p, 931 POISON_FREE, s->object_size - 1) || 932 !check_bytes_and_report(s, page, p, "End Poison", 933 p + s->object_size - 1, POISON_END, 1))) 934 return 0; 935 /* 936 * check_pad_bytes cleans up on its own. 937 */ 938 check_pad_bytes(s, page, p); 939 } 940 941 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE) 942 /* 943 * Object and freepointer overlap. Cannot check 944 * freepointer while object is allocated. 945 */ 946 return 1; 947 948 /* Check free pointer validity */ 949 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 950 object_err(s, page, p, "Freepointer corrupt"); 951 /* 952 * No choice but to zap it and thus lose the remainder 953 * of the free objects in this slab. May cause 954 * another error because the object count is now wrong. 955 */ 956 set_freepointer(s, p, NULL); 957 return 0; 958 } 959 return 1; 960 } 961 962 static int check_slab(struct kmem_cache *s, struct page *page) 963 { 964 int maxobj; 965 966 VM_BUG_ON(!irqs_disabled()); 967 968 if (!PageSlab(page)) { 969 slab_err(s, page, "Not a valid slab page"); 970 return 0; 971 } 972 973 maxobj = order_objects(compound_order(page), s->size); 974 if (page->objects > maxobj) { 975 slab_err(s, page, "objects %u > max %u", 976 page->objects, maxobj); 977 return 0; 978 } 979 if (page->inuse > page->objects) { 980 slab_err(s, page, "inuse %u > max %u", 981 page->inuse, page->objects); 982 return 0; 983 } 984 /* Slab_pad_check fixes things up after itself */ 985 slab_pad_check(s, page); 986 return 1; 987 } 988 989 /* 990 * Determine if a certain object on a page is on the freelist. Must hold the 991 * slab lock to guarantee that the chains are in a consistent state. 992 */ 993 static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 994 { 995 int nr = 0; 996 void *fp; 997 void *object = NULL; 998 int max_objects; 999 1000 fp = page->freelist; 1001 while (fp && nr <= page->objects) { 1002 if (fp == search) 1003 return 1; 1004 if (!check_valid_pointer(s, page, fp)) { 1005 if (object) { 1006 object_err(s, page, object, 1007 "Freechain corrupt"); 1008 set_freepointer(s, object, NULL); 1009 } else { 1010 slab_err(s, page, "Freepointer corrupt"); 1011 page->freelist = NULL; 1012 page->inuse = page->objects; 1013 slab_fix(s, "Freelist cleared"); 1014 return 0; 1015 } 1016 break; 1017 } 1018 object = fp; 1019 fp = get_freepointer(s, object); 1020 nr++; 1021 } 1022 1023 max_objects = order_objects(compound_order(page), s->size); 1024 if (max_objects > MAX_OBJS_PER_PAGE) 1025 max_objects = MAX_OBJS_PER_PAGE; 1026 1027 if (page->objects != max_objects) { 1028 slab_err(s, page, "Wrong number of objects. Found %d but should be %d", 1029 page->objects, max_objects); 1030 page->objects = max_objects; 1031 slab_fix(s, "Number of objects adjusted."); 1032 } 1033 if (page->inuse != page->objects - nr) { 1034 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d", 1035 page->inuse, page->objects - nr); 1036 page->inuse = page->objects - nr; 1037 slab_fix(s, "Object count adjusted."); 1038 } 1039 return search == NULL; 1040 } 1041 1042 static void trace(struct kmem_cache *s, struct page *page, void *object, 1043 int alloc) 1044 { 1045 if (s->flags & SLAB_TRACE) { 1046 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 1047 s->name, 1048 alloc ? "alloc" : "free", 1049 object, page->inuse, 1050 page->freelist); 1051 1052 if (!alloc) 1053 print_section(KERN_INFO, "Object ", (void *)object, 1054 s->object_size); 1055 1056 dump_stack(); 1057 } 1058 } 1059 1060 /* 1061 * Tracking of fully allocated slabs for debugging purposes. 1062 */ 1063 static void add_full(struct kmem_cache *s, 1064 struct kmem_cache_node *n, struct page *page) 1065 { 1066 if (!(s->flags & SLAB_STORE_USER)) 1067 return; 1068 1069 lockdep_assert_held(&n->list_lock); 1070 list_add(&page->slab_list, &n->full); 1071 } 1072 1073 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page) 1074 { 1075 if (!(s->flags & SLAB_STORE_USER)) 1076 return; 1077 1078 lockdep_assert_held(&n->list_lock); 1079 list_del(&page->slab_list); 1080 } 1081 1082 /* Tracking of the number of slabs for debugging purposes */ 1083 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1084 { 1085 struct kmem_cache_node *n = get_node(s, node); 1086 1087 return atomic_long_read(&n->nr_slabs); 1088 } 1089 1090 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1091 { 1092 return atomic_long_read(&n->nr_slabs); 1093 } 1094 1095 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 1096 { 1097 struct kmem_cache_node *n = get_node(s, node); 1098 1099 /* 1100 * May be called early in order to allocate a slab for the 1101 * kmem_cache_node structure. Solve the chicken-egg 1102 * dilemma by deferring the increment of the count during 1103 * bootstrap (see early_kmem_cache_node_alloc). 1104 */ 1105 if (likely(n)) { 1106 atomic_long_inc(&n->nr_slabs); 1107 atomic_long_add(objects, &n->total_objects); 1108 } 1109 } 1110 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 1111 { 1112 struct kmem_cache_node *n = get_node(s, node); 1113 1114 atomic_long_dec(&n->nr_slabs); 1115 atomic_long_sub(objects, &n->total_objects); 1116 } 1117 1118 /* Object debug checks for alloc/free paths */ 1119 static void setup_object_debug(struct kmem_cache *s, struct page *page, 1120 void *object) 1121 { 1122 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)) 1123 return; 1124 1125 init_object(s, object, SLUB_RED_INACTIVE); 1126 init_tracking(s, object); 1127 } 1128 1129 static 1130 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) 1131 { 1132 if (!kmem_cache_debug_flags(s, SLAB_POISON)) 1133 return; 1134 1135 metadata_access_enable(); 1136 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page)); 1137 metadata_access_disable(); 1138 } 1139 1140 static inline int alloc_consistency_checks(struct kmem_cache *s, 1141 struct page *page, void *object) 1142 { 1143 if (!check_slab(s, page)) 1144 return 0; 1145 1146 if (!check_valid_pointer(s, page, object)) { 1147 object_err(s, page, object, "Freelist Pointer check fails"); 1148 return 0; 1149 } 1150 1151 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1152 return 0; 1153 1154 return 1; 1155 } 1156 1157 static noinline int alloc_debug_processing(struct kmem_cache *s, 1158 struct page *page, 1159 void *object, unsigned long addr) 1160 { 1161 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1162 if (!alloc_consistency_checks(s, page, object)) 1163 goto bad; 1164 } 1165 1166 /* Success perform special debug activities for allocs */ 1167 if (s->flags & SLAB_STORE_USER) 1168 set_track(s, object, TRACK_ALLOC, addr); 1169 trace(s, page, object, 1); 1170 init_object(s, object, SLUB_RED_ACTIVE); 1171 return 1; 1172 1173 bad: 1174 if (PageSlab(page)) { 1175 /* 1176 * If this is a slab page then lets do the best we can 1177 * to avoid issues in the future. Marking all objects 1178 * as used avoids touching the remaining objects. 1179 */ 1180 slab_fix(s, "Marking all objects used"); 1181 page->inuse = page->objects; 1182 page->freelist = NULL; 1183 } 1184 return 0; 1185 } 1186 1187 static inline int free_consistency_checks(struct kmem_cache *s, 1188 struct page *page, void *object, unsigned long addr) 1189 { 1190 if (!check_valid_pointer(s, page, object)) { 1191 slab_err(s, page, "Invalid object pointer 0x%p", object); 1192 return 0; 1193 } 1194 1195 if (on_freelist(s, page, object)) { 1196 object_err(s, page, object, "Object already free"); 1197 return 0; 1198 } 1199 1200 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1201 return 0; 1202 1203 if (unlikely(s != page->slab_cache)) { 1204 if (!PageSlab(page)) { 1205 slab_err(s, page, "Attempt to free object(0x%p) outside of slab", 1206 object); 1207 } else if (!page->slab_cache) { 1208 pr_err("SLUB <none>: no slab for object 0x%p.\n", 1209 object); 1210 dump_stack(); 1211 } else 1212 object_err(s, page, object, 1213 "page slab pointer corrupt."); 1214 return 0; 1215 } 1216 return 1; 1217 } 1218 1219 /* Supports checking bulk free of a constructed freelist */ 1220 static noinline int free_debug_processing( 1221 struct kmem_cache *s, struct page *page, 1222 void *head, void *tail, int bulk_cnt, 1223 unsigned long addr) 1224 { 1225 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1226 void *object = head; 1227 int cnt = 0; 1228 unsigned long flags; 1229 int ret = 0; 1230 1231 spin_lock_irqsave(&n->list_lock, flags); 1232 slab_lock(page); 1233 1234 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1235 if (!check_slab(s, page)) 1236 goto out; 1237 } 1238 1239 next_object: 1240 cnt++; 1241 1242 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1243 if (!free_consistency_checks(s, page, object, addr)) 1244 goto out; 1245 } 1246 1247 if (s->flags & SLAB_STORE_USER) 1248 set_track(s, object, TRACK_FREE, addr); 1249 trace(s, page, object, 0); 1250 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ 1251 init_object(s, object, SLUB_RED_INACTIVE); 1252 1253 /* Reached end of constructed freelist yet? */ 1254 if (object != tail) { 1255 object = get_freepointer(s, object); 1256 goto next_object; 1257 } 1258 ret = 1; 1259 1260 out: 1261 if (cnt != bulk_cnt) 1262 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n", 1263 bulk_cnt, cnt); 1264 1265 slab_unlock(page); 1266 spin_unlock_irqrestore(&n->list_lock, flags); 1267 if (!ret) 1268 slab_fix(s, "Object at 0x%p not freed", object); 1269 return ret; 1270 } 1271 1272 /* 1273 * Parse a block of slub_debug options. Blocks are delimited by ';' 1274 * 1275 * @str: start of block 1276 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified 1277 * @slabs: return start of list of slabs, or NULL when there's no list 1278 * @init: assume this is initial parsing and not per-kmem-create parsing 1279 * 1280 * returns the start of next block if there's any, or NULL 1281 */ 1282 static char * 1283 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init) 1284 { 1285 bool higher_order_disable = false; 1286 1287 /* Skip any completely empty blocks */ 1288 while (*str && *str == ';') 1289 str++; 1290 1291 if (*str == ',') { 1292 /* 1293 * No options but restriction on slabs. This means full 1294 * debugging for slabs matching a pattern. 1295 */ 1296 *flags = DEBUG_DEFAULT_FLAGS; 1297 goto check_slabs; 1298 } 1299 *flags = 0; 1300 1301 /* Determine which debug features should be switched on */ 1302 for (; *str && *str != ',' && *str != ';'; str++) { 1303 switch (tolower(*str)) { 1304 case '-': 1305 *flags = 0; 1306 break; 1307 case 'f': 1308 *flags |= SLAB_CONSISTENCY_CHECKS; 1309 break; 1310 case 'z': 1311 *flags |= SLAB_RED_ZONE; 1312 break; 1313 case 'p': 1314 *flags |= SLAB_POISON; 1315 break; 1316 case 'u': 1317 *flags |= SLAB_STORE_USER; 1318 break; 1319 case 't': 1320 *flags |= SLAB_TRACE; 1321 break; 1322 case 'a': 1323 *flags |= SLAB_FAILSLAB; 1324 break; 1325 case 'o': 1326 /* 1327 * Avoid enabling debugging on caches if its minimum 1328 * order would increase as a result. 1329 */ 1330 higher_order_disable = true; 1331 break; 1332 default: 1333 if (init) 1334 pr_err("slub_debug option '%c' unknown. skipped\n", *str); 1335 } 1336 } 1337 check_slabs: 1338 if (*str == ',') 1339 *slabs = ++str; 1340 else 1341 *slabs = NULL; 1342 1343 /* Skip over the slab list */ 1344 while (*str && *str != ';') 1345 str++; 1346 1347 /* Skip any completely empty blocks */ 1348 while (*str && *str == ';') 1349 str++; 1350 1351 if (init && higher_order_disable) 1352 disable_higher_order_debug = 1; 1353 1354 if (*str) 1355 return str; 1356 else 1357 return NULL; 1358 } 1359 1360 static int __init setup_slub_debug(char *str) 1361 { 1362 slab_flags_t flags; 1363 char *saved_str; 1364 char *slab_list; 1365 bool global_slub_debug_changed = false; 1366 bool slab_list_specified = false; 1367 1368 slub_debug = DEBUG_DEFAULT_FLAGS; 1369 if (*str++ != '=' || !*str) 1370 /* 1371 * No options specified. Switch on full debugging. 1372 */ 1373 goto out; 1374 1375 saved_str = str; 1376 while (str) { 1377 str = parse_slub_debug_flags(str, &flags, &slab_list, true); 1378 1379 if (!slab_list) { 1380 slub_debug = flags; 1381 global_slub_debug_changed = true; 1382 } else { 1383 slab_list_specified = true; 1384 } 1385 } 1386 1387 /* 1388 * For backwards compatibility, a single list of flags with list of 1389 * slabs means debugging is only enabled for those slabs, so the global 1390 * slub_debug should be 0. We can extended that to multiple lists as 1391 * long as there is no option specifying flags without a slab list. 1392 */ 1393 if (slab_list_specified) { 1394 if (!global_slub_debug_changed) 1395 slub_debug = 0; 1396 slub_debug_string = saved_str; 1397 } 1398 out: 1399 if (slub_debug != 0 || slub_debug_string) 1400 static_branch_enable(&slub_debug_enabled); 1401 if ((static_branch_unlikely(&init_on_alloc) || 1402 static_branch_unlikely(&init_on_free)) && 1403 (slub_debug & SLAB_POISON)) 1404 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n"); 1405 return 1; 1406 } 1407 1408 __setup("slub_debug", setup_slub_debug); 1409 1410 /* 1411 * kmem_cache_flags - apply debugging options to the cache 1412 * @object_size: the size of an object without meta data 1413 * @flags: flags to set 1414 * @name: name of the cache 1415 * 1416 * Debug option(s) are applied to @flags. In addition to the debug 1417 * option(s), if a slab name (or multiple) is specified i.e. 1418 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ... 1419 * then only the select slabs will receive the debug option(s). 1420 */ 1421 slab_flags_t kmem_cache_flags(unsigned int object_size, 1422 slab_flags_t flags, const char *name) 1423 { 1424 char *iter; 1425 size_t len; 1426 char *next_block; 1427 slab_flags_t block_flags; 1428 slab_flags_t slub_debug_local = slub_debug; 1429 1430 /* 1431 * If the slab cache is for debugging (e.g. kmemleak) then 1432 * don't store user (stack trace) information by default, 1433 * but let the user enable it via the command line below. 1434 */ 1435 if (flags & SLAB_NOLEAKTRACE) 1436 slub_debug_local &= ~SLAB_STORE_USER; 1437 1438 len = strlen(name); 1439 next_block = slub_debug_string; 1440 /* Go through all blocks of debug options, see if any matches our slab's name */ 1441 while (next_block) { 1442 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false); 1443 if (!iter) 1444 continue; 1445 /* Found a block that has a slab list, search it */ 1446 while (*iter) { 1447 char *end, *glob; 1448 size_t cmplen; 1449 1450 end = strchrnul(iter, ','); 1451 if (next_block && next_block < end) 1452 end = next_block - 1; 1453 1454 glob = strnchr(iter, end - iter, '*'); 1455 if (glob) 1456 cmplen = glob - iter; 1457 else 1458 cmplen = max_t(size_t, len, (end - iter)); 1459 1460 if (!strncmp(name, iter, cmplen)) { 1461 flags |= block_flags; 1462 return flags; 1463 } 1464 1465 if (!*end || *end == ';') 1466 break; 1467 iter = end + 1; 1468 } 1469 } 1470 1471 return flags | slub_debug_local; 1472 } 1473 #else /* !CONFIG_SLUB_DEBUG */ 1474 static inline void setup_object_debug(struct kmem_cache *s, 1475 struct page *page, void *object) {} 1476 static inline 1477 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {} 1478 1479 static inline int alloc_debug_processing(struct kmem_cache *s, 1480 struct page *page, void *object, unsigned long addr) { return 0; } 1481 1482 static inline int free_debug_processing( 1483 struct kmem_cache *s, struct page *page, 1484 void *head, void *tail, int bulk_cnt, 1485 unsigned long addr) { return 0; } 1486 1487 static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1488 { return 1; } 1489 static inline int check_object(struct kmem_cache *s, struct page *page, 1490 void *object, u8 val) { return 1; } 1491 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, 1492 struct page *page) {} 1493 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, 1494 struct page *page) {} 1495 slab_flags_t kmem_cache_flags(unsigned int object_size, 1496 slab_flags_t flags, const char *name) 1497 { 1498 return flags; 1499 } 1500 #define slub_debug 0 1501 1502 #define disable_higher_order_debug 0 1503 1504 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1505 { return 0; } 1506 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1507 { return 0; } 1508 static inline void inc_slabs_node(struct kmem_cache *s, int node, 1509 int objects) {} 1510 static inline void dec_slabs_node(struct kmem_cache *s, int node, 1511 int objects) {} 1512 1513 static bool freelist_corrupted(struct kmem_cache *s, struct page *page, 1514 void **freelist, void *nextfree) 1515 { 1516 return false; 1517 } 1518 #endif /* CONFIG_SLUB_DEBUG */ 1519 1520 /* 1521 * Hooks for other subsystems that check memory allocations. In a typical 1522 * production configuration these hooks all should produce no code at all. 1523 */ 1524 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) 1525 { 1526 ptr = kasan_kmalloc_large(ptr, size, flags); 1527 /* As ptr might get tagged, call kmemleak hook after KASAN. */ 1528 kmemleak_alloc(ptr, size, 1, flags); 1529 return ptr; 1530 } 1531 1532 static __always_inline void kfree_hook(void *x) 1533 { 1534 kmemleak_free(x); 1535 kasan_kfree_large(x); 1536 } 1537 1538 static __always_inline bool slab_free_hook(struct kmem_cache *s, 1539 void *x, bool init) 1540 { 1541 kmemleak_free_recursive(x, s->flags); 1542 1543 /* 1544 * Trouble is that we may no longer disable interrupts in the fast path 1545 * So in order to make the debug calls that expect irqs to be 1546 * disabled we need to disable interrupts temporarily. 1547 */ 1548 #ifdef CONFIG_LOCKDEP 1549 { 1550 unsigned long flags; 1551 1552 local_irq_save(flags); 1553 debug_check_no_locks_freed(x, s->object_size); 1554 local_irq_restore(flags); 1555 } 1556 #endif 1557 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1558 debug_check_no_obj_freed(x, s->object_size); 1559 1560 /* Use KCSAN to help debug racy use-after-free. */ 1561 if (!(s->flags & SLAB_TYPESAFE_BY_RCU)) 1562 __kcsan_check_access(x, s->object_size, 1563 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT); 1564 1565 /* 1566 * As memory initialization might be integrated into KASAN, 1567 * kasan_slab_free and initialization memset's must be 1568 * kept together to avoid discrepancies in behavior. 1569 * 1570 * The initialization memset's clear the object and the metadata, 1571 * but don't touch the SLAB redzone. 1572 */ 1573 if (init) { 1574 int rsize; 1575 1576 if (!kasan_has_integrated_init()) 1577 memset(kasan_reset_tag(x), 0, s->object_size); 1578 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0; 1579 memset((char *)kasan_reset_tag(x) + s->inuse, 0, 1580 s->size - s->inuse - rsize); 1581 } 1582 /* KASAN might put x into memory quarantine, delaying its reuse. */ 1583 return kasan_slab_free(s, x, init); 1584 } 1585 1586 static inline bool slab_free_freelist_hook(struct kmem_cache *s, 1587 void **head, void **tail) 1588 { 1589 1590 void *object; 1591 void *next = *head; 1592 void *old_tail = *tail ? *tail : *head; 1593 1594 if (is_kfence_address(next)) { 1595 slab_free_hook(s, next, false); 1596 return true; 1597 } 1598 1599 /* Head and tail of the reconstructed freelist */ 1600 *head = NULL; 1601 *tail = NULL; 1602 1603 do { 1604 object = next; 1605 next = get_freepointer(s, object); 1606 1607 /* If object's reuse doesn't have to be delayed */ 1608 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) { 1609 /* Move object to the new freelist */ 1610 set_freepointer(s, object, *head); 1611 *head = object; 1612 if (!*tail) 1613 *tail = object; 1614 } 1615 } while (object != old_tail); 1616 1617 if (*head == *tail) 1618 *tail = NULL; 1619 1620 return *head != NULL; 1621 } 1622 1623 static void *setup_object(struct kmem_cache *s, struct page *page, 1624 void *object) 1625 { 1626 setup_object_debug(s, page, object); 1627 object = kasan_init_slab_obj(s, object); 1628 if (unlikely(s->ctor)) { 1629 kasan_unpoison_object_data(s, object); 1630 s->ctor(object); 1631 kasan_poison_object_data(s, object); 1632 } 1633 return object; 1634 } 1635 1636 /* 1637 * Slab allocation and freeing 1638 */ 1639 static inline struct page *alloc_slab_page(struct kmem_cache *s, 1640 gfp_t flags, int node, struct kmem_cache_order_objects oo) 1641 { 1642 struct page *page; 1643 unsigned int order = oo_order(oo); 1644 1645 if (node == NUMA_NO_NODE) 1646 page = alloc_pages(flags, order); 1647 else 1648 page = __alloc_pages_node(node, flags, order); 1649 1650 return page; 1651 } 1652 1653 #ifdef CONFIG_SLAB_FREELIST_RANDOM 1654 /* Pre-initialize the random sequence cache */ 1655 static int init_cache_random_seq(struct kmem_cache *s) 1656 { 1657 unsigned int count = oo_objects(s->oo); 1658 int err; 1659 1660 /* Bailout if already initialised */ 1661 if (s->random_seq) 1662 return 0; 1663 1664 err = cache_random_seq_create(s, count, GFP_KERNEL); 1665 if (err) { 1666 pr_err("SLUB: Unable to initialize free list for %s\n", 1667 s->name); 1668 return err; 1669 } 1670 1671 /* Transform to an offset on the set of pages */ 1672 if (s->random_seq) { 1673 unsigned int i; 1674 1675 for (i = 0; i < count; i++) 1676 s->random_seq[i] *= s->size; 1677 } 1678 return 0; 1679 } 1680 1681 /* Initialize each random sequence freelist per cache */ 1682 static void __init init_freelist_randomization(void) 1683 { 1684 struct kmem_cache *s; 1685 1686 mutex_lock(&slab_mutex); 1687 1688 list_for_each_entry(s, &slab_caches, list) 1689 init_cache_random_seq(s); 1690 1691 mutex_unlock(&slab_mutex); 1692 } 1693 1694 /* Get the next entry on the pre-computed freelist randomized */ 1695 static void *next_freelist_entry(struct kmem_cache *s, struct page *page, 1696 unsigned long *pos, void *start, 1697 unsigned long page_limit, 1698 unsigned long freelist_count) 1699 { 1700 unsigned int idx; 1701 1702 /* 1703 * If the target page allocation failed, the number of objects on the 1704 * page might be smaller than the usual size defined by the cache. 1705 */ 1706 do { 1707 idx = s->random_seq[*pos]; 1708 *pos += 1; 1709 if (*pos >= freelist_count) 1710 *pos = 0; 1711 } while (unlikely(idx >= page_limit)); 1712 1713 return (char *)start + idx; 1714 } 1715 1716 /* Shuffle the single linked freelist based on a random pre-computed sequence */ 1717 static bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1718 { 1719 void *start; 1720 void *cur; 1721 void *next; 1722 unsigned long idx, pos, page_limit, freelist_count; 1723 1724 if (page->objects < 2 || !s->random_seq) 1725 return false; 1726 1727 freelist_count = oo_objects(s->oo); 1728 pos = get_random_int() % freelist_count; 1729 1730 page_limit = page->objects * s->size; 1731 start = fixup_red_left(s, page_address(page)); 1732 1733 /* First entry is used as the base of the freelist */ 1734 cur = next_freelist_entry(s, page, &pos, start, page_limit, 1735 freelist_count); 1736 cur = setup_object(s, page, cur); 1737 page->freelist = cur; 1738 1739 for (idx = 1; idx < page->objects; idx++) { 1740 next = next_freelist_entry(s, page, &pos, start, page_limit, 1741 freelist_count); 1742 next = setup_object(s, page, next); 1743 set_freepointer(s, cur, next); 1744 cur = next; 1745 } 1746 set_freepointer(s, cur, NULL); 1747 1748 return true; 1749 } 1750 #else 1751 static inline int init_cache_random_seq(struct kmem_cache *s) 1752 { 1753 return 0; 1754 } 1755 static inline void init_freelist_randomization(void) { } 1756 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1757 { 1758 return false; 1759 } 1760 #endif /* CONFIG_SLAB_FREELIST_RANDOM */ 1761 1762 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1763 { 1764 struct page *page; 1765 struct kmem_cache_order_objects oo = s->oo; 1766 gfp_t alloc_gfp; 1767 void *start, *p, *next; 1768 int idx; 1769 bool shuffle; 1770 1771 flags &= gfp_allowed_mask; 1772 1773 if (gfpflags_allow_blocking(flags)) 1774 local_irq_enable(); 1775 1776 flags |= s->allocflags; 1777 1778 /* 1779 * Let the initial higher-order allocation fail under memory pressure 1780 * so we fall-back to the minimum order allocation. 1781 */ 1782 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1783 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min)) 1784 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL); 1785 1786 page = alloc_slab_page(s, alloc_gfp, node, oo); 1787 if (unlikely(!page)) { 1788 oo = s->min; 1789 alloc_gfp = flags; 1790 /* 1791 * Allocation may have failed due to fragmentation. 1792 * Try a lower order alloc if possible 1793 */ 1794 page = alloc_slab_page(s, alloc_gfp, node, oo); 1795 if (unlikely(!page)) 1796 goto out; 1797 stat(s, ORDER_FALLBACK); 1798 } 1799 1800 page->objects = oo_objects(oo); 1801 1802 account_slab_page(page, oo_order(oo), s, flags); 1803 1804 page->slab_cache = s; 1805 __SetPageSlab(page); 1806 if (page_is_pfmemalloc(page)) 1807 SetPageSlabPfmemalloc(page); 1808 1809 kasan_poison_slab(page); 1810 1811 start = page_address(page); 1812 1813 setup_page_debug(s, page, start); 1814 1815 shuffle = shuffle_freelist(s, page); 1816 1817 if (!shuffle) { 1818 start = fixup_red_left(s, start); 1819 start = setup_object(s, page, start); 1820 page->freelist = start; 1821 for (idx = 0, p = start; idx < page->objects - 1; idx++) { 1822 next = p + s->size; 1823 next = setup_object(s, page, next); 1824 set_freepointer(s, p, next); 1825 p = next; 1826 } 1827 set_freepointer(s, p, NULL); 1828 } 1829 1830 page->inuse = page->objects; 1831 page->frozen = 1; 1832 1833 out: 1834 if (gfpflags_allow_blocking(flags)) 1835 local_irq_disable(); 1836 if (!page) 1837 return NULL; 1838 1839 inc_slabs_node(s, page_to_nid(page), page->objects); 1840 1841 return page; 1842 } 1843 1844 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1845 { 1846 if (unlikely(flags & GFP_SLAB_BUG_MASK)) 1847 flags = kmalloc_fix_flags(flags); 1848 1849 return allocate_slab(s, 1850 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1851 } 1852 1853 static void __free_slab(struct kmem_cache *s, struct page *page) 1854 { 1855 int order = compound_order(page); 1856 int pages = 1 << order; 1857 1858 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) { 1859 void *p; 1860 1861 slab_pad_check(s, page); 1862 for_each_object(p, s, page_address(page), 1863 page->objects) 1864 check_object(s, page, p, SLUB_RED_INACTIVE); 1865 } 1866 1867 __ClearPageSlabPfmemalloc(page); 1868 __ClearPageSlab(page); 1869 /* In union with page->mapping where page allocator expects NULL */ 1870 page->slab_cache = NULL; 1871 if (current->reclaim_state) 1872 current->reclaim_state->reclaimed_slab += pages; 1873 unaccount_slab_page(page, order, s); 1874 __free_pages(page, order); 1875 } 1876 1877 static void rcu_free_slab(struct rcu_head *h) 1878 { 1879 struct page *page = container_of(h, struct page, rcu_head); 1880 1881 __free_slab(page->slab_cache, page); 1882 } 1883 1884 static void free_slab(struct kmem_cache *s, struct page *page) 1885 { 1886 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) { 1887 call_rcu(&page->rcu_head, rcu_free_slab); 1888 } else 1889 __free_slab(s, page); 1890 } 1891 1892 static void discard_slab(struct kmem_cache *s, struct page *page) 1893 { 1894 dec_slabs_node(s, page_to_nid(page), page->objects); 1895 free_slab(s, page); 1896 } 1897 1898 /* 1899 * Management of partially allocated slabs. 1900 */ 1901 static inline void 1902 __add_partial(struct kmem_cache_node *n, struct page *page, int tail) 1903 { 1904 n->nr_partial++; 1905 if (tail == DEACTIVATE_TO_TAIL) 1906 list_add_tail(&page->slab_list, &n->partial); 1907 else 1908 list_add(&page->slab_list, &n->partial); 1909 } 1910 1911 static inline void add_partial(struct kmem_cache_node *n, 1912 struct page *page, int tail) 1913 { 1914 lockdep_assert_held(&n->list_lock); 1915 __add_partial(n, page, tail); 1916 } 1917 1918 static inline void remove_partial(struct kmem_cache_node *n, 1919 struct page *page) 1920 { 1921 lockdep_assert_held(&n->list_lock); 1922 list_del(&page->slab_list); 1923 n->nr_partial--; 1924 } 1925 1926 /* 1927 * Remove slab from the partial list, freeze it and 1928 * return the pointer to the freelist. 1929 * 1930 * Returns a list of objects or NULL if it fails. 1931 */ 1932 static inline void *acquire_slab(struct kmem_cache *s, 1933 struct kmem_cache_node *n, struct page *page, 1934 int mode, int *objects) 1935 { 1936 void *freelist; 1937 unsigned long counters; 1938 struct page new; 1939 1940 lockdep_assert_held(&n->list_lock); 1941 1942 /* 1943 * Zap the freelist and set the frozen bit. 1944 * The old freelist is the list of objects for the 1945 * per cpu allocation list. 1946 */ 1947 freelist = page->freelist; 1948 counters = page->counters; 1949 new.counters = counters; 1950 *objects = new.objects - new.inuse; 1951 if (mode) { 1952 new.inuse = page->objects; 1953 new.freelist = NULL; 1954 } else { 1955 new.freelist = freelist; 1956 } 1957 1958 VM_BUG_ON(new.frozen); 1959 new.frozen = 1; 1960 1961 if (!__cmpxchg_double_slab(s, page, 1962 freelist, counters, 1963 new.freelist, new.counters, 1964 "acquire_slab")) 1965 return NULL; 1966 1967 remove_partial(n, page); 1968 WARN_ON(!freelist); 1969 return freelist; 1970 } 1971 1972 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); 1973 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags); 1974 1975 /* 1976 * Try to allocate a partial slab from a specific node. 1977 */ 1978 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, 1979 struct kmem_cache_cpu *c, gfp_t flags) 1980 { 1981 struct page *page, *page2; 1982 void *object = NULL; 1983 unsigned int available = 0; 1984 int objects; 1985 1986 /* 1987 * Racy check. If we mistakenly see no partial slabs then we 1988 * just allocate an empty slab. If we mistakenly try to get a 1989 * partial slab and there is none available then get_partial() 1990 * will return NULL. 1991 */ 1992 if (!n || !n->nr_partial) 1993 return NULL; 1994 1995 spin_lock(&n->list_lock); 1996 list_for_each_entry_safe(page, page2, &n->partial, slab_list) { 1997 void *t; 1998 1999 if (!pfmemalloc_match(page, flags)) 2000 continue; 2001 2002 t = acquire_slab(s, n, page, object == NULL, &objects); 2003 if (!t) 2004 break; 2005 2006 available += objects; 2007 if (!object) { 2008 c->page = page; 2009 stat(s, ALLOC_FROM_PARTIAL); 2010 object = t; 2011 } else { 2012 put_cpu_partial(s, page, 0); 2013 stat(s, CPU_PARTIAL_NODE); 2014 } 2015 if (!kmem_cache_has_cpu_partial(s) 2016 || available > slub_cpu_partial(s) / 2) 2017 break; 2018 2019 } 2020 spin_unlock(&n->list_lock); 2021 return object; 2022 } 2023 2024 /* 2025 * Get a page from somewhere. Search in increasing NUMA distances. 2026 */ 2027 static void *get_any_partial(struct kmem_cache *s, gfp_t flags, 2028 struct kmem_cache_cpu *c) 2029 { 2030 #ifdef CONFIG_NUMA 2031 struct zonelist *zonelist; 2032 struct zoneref *z; 2033 struct zone *zone; 2034 enum zone_type highest_zoneidx = gfp_zone(flags); 2035 void *object; 2036 unsigned int cpuset_mems_cookie; 2037 2038 /* 2039 * The defrag ratio allows a configuration of the tradeoffs between 2040 * inter node defragmentation and node local allocations. A lower 2041 * defrag_ratio increases the tendency to do local allocations 2042 * instead of attempting to obtain partial slabs from other nodes. 2043 * 2044 * If the defrag_ratio is set to 0 then kmalloc() always 2045 * returns node local objects. If the ratio is higher then kmalloc() 2046 * may return off node objects because partial slabs are obtained 2047 * from other nodes and filled up. 2048 * 2049 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 2050 * (which makes defrag_ratio = 1000) then every (well almost) 2051 * allocation will first attempt to defrag slab caches on other nodes. 2052 * This means scanning over all nodes to look for partial slabs which 2053 * may be expensive if we do it every time we are trying to find a slab 2054 * with available objects. 2055 */ 2056 if (!s->remote_node_defrag_ratio || 2057 get_cycles() % 1024 > s->remote_node_defrag_ratio) 2058 return NULL; 2059 2060 do { 2061 cpuset_mems_cookie = read_mems_allowed_begin(); 2062 zonelist = node_zonelist(mempolicy_slab_node(), flags); 2063 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) { 2064 struct kmem_cache_node *n; 2065 2066 n = get_node(s, zone_to_nid(zone)); 2067 2068 if (n && cpuset_zone_allowed(zone, flags) && 2069 n->nr_partial > s->min_partial) { 2070 object = get_partial_node(s, n, c, flags); 2071 if (object) { 2072 /* 2073 * Don't check read_mems_allowed_retry() 2074 * here - if mems_allowed was updated in 2075 * parallel, that was a harmless race 2076 * between allocation and the cpuset 2077 * update 2078 */ 2079 return object; 2080 } 2081 } 2082 } 2083 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 2084 #endif /* CONFIG_NUMA */ 2085 return NULL; 2086 } 2087 2088 /* 2089 * Get a partial page, lock it and return it. 2090 */ 2091 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, 2092 struct kmem_cache_cpu *c) 2093 { 2094 void *object; 2095 int searchnode = node; 2096 2097 if (node == NUMA_NO_NODE) 2098 searchnode = numa_mem_id(); 2099 2100 object = get_partial_node(s, get_node(s, searchnode), c, flags); 2101 if (object || node != NUMA_NO_NODE) 2102 return object; 2103 2104 return get_any_partial(s, flags, c); 2105 } 2106 2107 #ifdef CONFIG_PREEMPTION 2108 /* 2109 * Calculate the next globally unique transaction for disambiguation 2110 * during cmpxchg. The transactions start with the cpu number and are then 2111 * incremented by CONFIG_NR_CPUS. 2112 */ 2113 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 2114 #else 2115 /* 2116 * No preemption supported therefore also no need to check for 2117 * different cpus. 2118 */ 2119 #define TID_STEP 1 2120 #endif 2121 2122 static inline unsigned long next_tid(unsigned long tid) 2123 { 2124 return tid + TID_STEP; 2125 } 2126 2127 #ifdef SLUB_DEBUG_CMPXCHG 2128 static inline unsigned int tid_to_cpu(unsigned long tid) 2129 { 2130 return tid % TID_STEP; 2131 } 2132 2133 static inline unsigned long tid_to_event(unsigned long tid) 2134 { 2135 return tid / TID_STEP; 2136 } 2137 #endif 2138 2139 static inline unsigned int init_tid(int cpu) 2140 { 2141 return cpu; 2142 } 2143 2144 static inline void note_cmpxchg_failure(const char *n, 2145 const struct kmem_cache *s, unsigned long tid) 2146 { 2147 #ifdef SLUB_DEBUG_CMPXCHG 2148 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 2149 2150 pr_info("%s %s: cmpxchg redo ", n, s->name); 2151 2152 #ifdef CONFIG_PREEMPTION 2153 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 2154 pr_warn("due to cpu change %d -> %d\n", 2155 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 2156 else 2157 #endif 2158 if (tid_to_event(tid) != tid_to_event(actual_tid)) 2159 pr_warn("due to cpu running other code. Event %ld->%ld\n", 2160 tid_to_event(tid), tid_to_event(actual_tid)); 2161 else 2162 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n", 2163 actual_tid, tid, next_tid(tid)); 2164 #endif 2165 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 2166 } 2167 2168 static void init_kmem_cache_cpus(struct kmem_cache *s) 2169 { 2170 int cpu; 2171 2172 for_each_possible_cpu(cpu) 2173 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 2174 } 2175 2176 /* 2177 * Remove the cpu slab 2178 */ 2179 static void deactivate_slab(struct kmem_cache *s, struct page *page, 2180 void *freelist, struct kmem_cache_cpu *c) 2181 { 2182 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; 2183 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 2184 int lock = 0, free_delta = 0; 2185 enum slab_modes l = M_NONE, m = M_NONE; 2186 void *nextfree, *freelist_iter, *freelist_tail; 2187 int tail = DEACTIVATE_TO_HEAD; 2188 struct page new; 2189 struct page old; 2190 2191 if (page->freelist) { 2192 stat(s, DEACTIVATE_REMOTE_FREES); 2193 tail = DEACTIVATE_TO_TAIL; 2194 } 2195 2196 /* 2197 * Stage one: Count the objects on cpu's freelist as free_delta and 2198 * remember the last object in freelist_tail for later splicing. 2199 */ 2200 freelist_tail = NULL; 2201 freelist_iter = freelist; 2202 while (freelist_iter) { 2203 nextfree = get_freepointer(s, freelist_iter); 2204 2205 /* 2206 * If 'nextfree' is invalid, it is possible that the object at 2207 * 'freelist_iter' is already corrupted. So isolate all objects 2208 * starting at 'freelist_iter' by skipping them. 2209 */ 2210 if (freelist_corrupted(s, page, &freelist_iter, nextfree)) 2211 break; 2212 2213 freelist_tail = freelist_iter; 2214 free_delta++; 2215 2216 freelist_iter = nextfree; 2217 } 2218 2219 /* 2220 * Stage two: Unfreeze the page while splicing the per-cpu 2221 * freelist to the head of page's freelist. 2222 * 2223 * Ensure that the page is unfrozen while the list presence 2224 * reflects the actual number of objects during unfreeze. 2225 * 2226 * We setup the list membership and then perform a cmpxchg 2227 * with the count. If there is a mismatch then the page 2228 * is not unfrozen but the page is on the wrong list. 2229 * 2230 * Then we restart the process which may have to remove 2231 * the page from the list that we just put it on again 2232 * because the number of objects in the slab may have 2233 * changed. 2234 */ 2235 redo: 2236 2237 old.freelist = READ_ONCE(page->freelist); 2238 old.counters = READ_ONCE(page->counters); 2239 VM_BUG_ON(!old.frozen); 2240 2241 /* Determine target state of the slab */ 2242 new.counters = old.counters; 2243 if (freelist_tail) { 2244 new.inuse -= free_delta; 2245 set_freepointer(s, freelist_tail, old.freelist); 2246 new.freelist = freelist; 2247 } else 2248 new.freelist = old.freelist; 2249 2250 new.frozen = 0; 2251 2252 if (!new.inuse && n->nr_partial >= s->min_partial) 2253 m = M_FREE; 2254 else if (new.freelist) { 2255 m = M_PARTIAL; 2256 if (!lock) { 2257 lock = 1; 2258 /* 2259 * Taking the spinlock removes the possibility 2260 * that acquire_slab() will see a slab page that 2261 * is frozen 2262 */ 2263 spin_lock(&n->list_lock); 2264 } 2265 } else { 2266 m = M_FULL; 2267 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) { 2268 lock = 1; 2269 /* 2270 * This also ensures that the scanning of full 2271 * slabs from diagnostic functions will not see 2272 * any frozen slabs. 2273 */ 2274 spin_lock(&n->list_lock); 2275 } 2276 } 2277 2278 if (l != m) { 2279 if (l == M_PARTIAL) 2280 remove_partial(n, page); 2281 else if (l == M_FULL) 2282 remove_full(s, n, page); 2283 2284 if (m == M_PARTIAL) 2285 add_partial(n, page, tail); 2286 else if (m == M_FULL) 2287 add_full(s, n, page); 2288 } 2289 2290 l = m; 2291 if (!__cmpxchg_double_slab(s, page, 2292 old.freelist, old.counters, 2293 new.freelist, new.counters, 2294 "unfreezing slab")) 2295 goto redo; 2296 2297 if (lock) 2298 spin_unlock(&n->list_lock); 2299 2300 if (m == M_PARTIAL) 2301 stat(s, tail); 2302 else if (m == M_FULL) 2303 stat(s, DEACTIVATE_FULL); 2304 else if (m == M_FREE) { 2305 stat(s, DEACTIVATE_EMPTY); 2306 discard_slab(s, page); 2307 stat(s, FREE_SLAB); 2308 } 2309 2310 c->page = NULL; 2311 c->freelist = NULL; 2312 } 2313 2314 /* 2315 * Unfreeze all the cpu partial slabs. 2316 * 2317 * This function must be called with interrupts disabled 2318 * for the cpu using c (or some other guarantee must be there 2319 * to guarantee no concurrent accesses). 2320 */ 2321 static void unfreeze_partials(struct kmem_cache *s, 2322 struct kmem_cache_cpu *c) 2323 { 2324 #ifdef CONFIG_SLUB_CPU_PARTIAL 2325 struct kmem_cache_node *n = NULL, *n2 = NULL; 2326 struct page *page, *discard_page = NULL; 2327 2328 while ((page = slub_percpu_partial(c))) { 2329 struct page new; 2330 struct page old; 2331 2332 slub_set_percpu_partial(c, page); 2333 2334 n2 = get_node(s, page_to_nid(page)); 2335 if (n != n2) { 2336 if (n) 2337 spin_unlock(&n->list_lock); 2338 2339 n = n2; 2340 spin_lock(&n->list_lock); 2341 } 2342 2343 do { 2344 2345 old.freelist = page->freelist; 2346 old.counters = page->counters; 2347 VM_BUG_ON(!old.frozen); 2348 2349 new.counters = old.counters; 2350 new.freelist = old.freelist; 2351 2352 new.frozen = 0; 2353 2354 } while (!__cmpxchg_double_slab(s, page, 2355 old.freelist, old.counters, 2356 new.freelist, new.counters, 2357 "unfreezing slab")); 2358 2359 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { 2360 page->next = discard_page; 2361 discard_page = page; 2362 } else { 2363 add_partial(n, page, DEACTIVATE_TO_TAIL); 2364 stat(s, FREE_ADD_PARTIAL); 2365 } 2366 } 2367 2368 if (n) 2369 spin_unlock(&n->list_lock); 2370 2371 while (discard_page) { 2372 page = discard_page; 2373 discard_page = discard_page->next; 2374 2375 stat(s, DEACTIVATE_EMPTY); 2376 discard_slab(s, page); 2377 stat(s, FREE_SLAB); 2378 } 2379 #endif /* CONFIG_SLUB_CPU_PARTIAL */ 2380 } 2381 2382 /* 2383 * Put a page that was just frozen (in __slab_free|get_partial_node) into a 2384 * partial page slot if available. 2385 * 2386 * If we did not find a slot then simply move all the partials to the 2387 * per node partial list. 2388 */ 2389 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) 2390 { 2391 #ifdef CONFIG_SLUB_CPU_PARTIAL 2392 struct page *oldpage; 2393 int pages; 2394 int pobjects; 2395 2396 preempt_disable(); 2397 do { 2398 pages = 0; 2399 pobjects = 0; 2400 oldpage = this_cpu_read(s->cpu_slab->partial); 2401 2402 if (oldpage) { 2403 pobjects = oldpage->pobjects; 2404 pages = oldpage->pages; 2405 if (drain && pobjects > slub_cpu_partial(s)) { 2406 unsigned long flags; 2407 /* 2408 * partial array is full. Move the existing 2409 * set to the per node partial list. 2410 */ 2411 local_irq_save(flags); 2412 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2413 local_irq_restore(flags); 2414 oldpage = NULL; 2415 pobjects = 0; 2416 pages = 0; 2417 stat(s, CPU_PARTIAL_DRAIN); 2418 } 2419 } 2420 2421 pages++; 2422 pobjects += page->objects - page->inuse; 2423 2424 page->pages = pages; 2425 page->pobjects = pobjects; 2426 page->next = oldpage; 2427 2428 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) 2429 != oldpage); 2430 if (unlikely(!slub_cpu_partial(s))) { 2431 unsigned long flags; 2432 2433 local_irq_save(flags); 2434 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2435 local_irq_restore(flags); 2436 } 2437 preempt_enable(); 2438 #endif /* CONFIG_SLUB_CPU_PARTIAL */ 2439 } 2440 2441 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 2442 { 2443 stat(s, CPUSLAB_FLUSH); 2444 deactivate_slab(s, c->page, c->freelist, c); 2445 2446 c->tid = next_tid(c->tid); 2447 } 2448 2449 /* 2450 * Flush cpu slab. 2451 * 2452 * Called from IPI handler with interrupts disabled. 2453 */ 2454 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 2455 { 2456 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2457 2458 if (c->page) 2459 flush_slab(s, c); 2460 2461 unfreeze_partials(s, c); 2462 } 2463 2464 static void flush_cpu_slab(void *d) 2465 { 2466 struct kmem_cache *s = d; 2467 2468 __flush_cpu_slab(s, smp_processor_id()); 2469 } 2470 2471 static bool has_cpu_slab(int cpu, void *info) 2472 { 2473 struct kmem_cache *s = info; 2474 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2475 2476 return c->page || slub_percpu_partial(c); 2477 } 2478 2479 static void flush_all(struct kmem_cache *s) 2480 { 2481 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1); 2482 } 2483 2484 /* 2485 * Use the cpu notifier to insure that the cpu slabs are flushed when 2486 * necessary. 2487 */ 2488 static int slub_cpu_dead(unsigned int cpu) 2489 { 2490 struct kmem_cache *s; 2491 unsigned long flags; 2492 2493 mutex_lock(&slab_mutex); 2494 list_for_each_entry(s, &slab_caches, list) { 2495 local_irq_save(flags); 2496 __flush_cpu_slab(s, cpu); 2497 local_irq_restore(flags); 2498 } 2499 mutex_unlock(&slab_mutex); 2500 return 0; 2501 } 2502 2503 /* 2504 * Check if the objects in a per cpu structure fit numa 2505 * locality expectations. 2506 */ 2507 static inline int node_match(struct page *page, int node) 2508 { 2509 #ifdef CONFIG_NUMA 2510 if (node != NUMA_NO_NODE && page_to_nid(page) != node) 2511 return 0; 2512 #endif 2513 return 1; 2514 } 2515 2516 #ifdef CONFIG_SLUB_DEBUG 2517 static int count_free(struct page *page) 2518 { 2519 return page->objects - page->inuse; 2520 } 2521 2522 static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 2523 { 2524 return atomic_long_read(&n->total_objects); 2525 } 2526 #endif /* CONFIG_SLUB_DEBUG */ 2527 2528 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS) 2529 static unsigned long count_partial(struct kmem_cache_node *n, 2530 int (*get_count)(struct page *)) 2531 { 2532 unsigned long flags; 2533 unsigned long x = 0; 2534 struct page *page; 2535 2536 spin_lock_irqsave(&n->list_lock, flags); 2537 list_for_each_entry(page, &n->partial, slab_list) 2538 x += get_count(page); 2539 spin_unlock_irqrestore(&n->list_lock, flags); 2540 return x; 2541 } 2542 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */ 2543 2544 static noinline void 2545 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 2546 { 2547 #ifdef CONFIG_SLUB_DEBUG 2548 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, 2549 DEFAULT_RATELIMIT_BURST); 2550 int node; 2551 struct kmem_cache_node *n; 2552 2553 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) 2554 return; 2555 2556 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", 2557 nid, gfpflags, &gfpflags); 2558 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n", 2559 s->name, s->object_size, s->size, oo_order(s->oo), 2560 oo_order(s->min)); 2561 2562 if (oo_order(s->min) > get_order(s->object_size)) 2563 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n", 2564 s->name); 2565 2566 for_each_kmem_cache_node(s, node, n) { 2567 unsigned long nr_slabs; 2568 unsigned long nr_objs; 2569 unsigned long nr_free; 2570 2571 nr_free = count_partial(n, count_free); 2572 nr_slabs = node_nr_slabs(n); 2573 nr_objs = node_nr_objs(n); 2574 2575 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n", 2576 node, nr_slabs, nr_objs, nr_free); 2577 } 2578 #endif 2579 } 2580 2581 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, 2582 int node, struct kmem_cache_cpu **pc) 2583 { 2584 void *freelist; 2585 struct kmem_cache_cpu *c = *pc; 2586 struct page *page; 2587 2588 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO)); 2589 2590 freelist = get_partial(s, flags, node, c); 2591 2592 if (freelist) 2593 return freelist; 2594 2595 page = new_slab(s, flags, node); 2596 if (page) { 2597 c = raw_cpu_ptr(s->cpu_slab); 2598 if (c->page) 2599 flush_slab(s, c); 2600 2601 /* 2602 * No other reference to the page yet so we can 2603 * muck around with it freely without cmpxchg 2604 */ 2605 freelist = page->freelist; 2606 page->freelist = NULL; 2607 2608 stat(s, ALLOC_SLAB); 2609 c->page = page; 2610 *pc = c; 2611 } 2612 2613 return freelist; 2614 } 2615 2616 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags) 2617 { 2618 if (unlikely(PageSlabPfmemalloc(page))) 2619 return gfp_pfmemalloc_allowed(gfpflags); 2620 2621 return true; 2622 } 2623 2624 /* 2625 * Check the page->freelist of a page and either transfer the freelist to the 2626 * per cpu freelist or deactivate the page. 2627 * 2628 * The page is still frozen if the return value is not NULL. 2629 * 2630 * If this function returns NULL then the page has been unfrozen. 2631 * 2632 * This function must be called with interrupt disabled. 2633 */ 2634 static inline void *get_freelist(struct kmem_cache *s, struct page *page) 2635 { 2636 struct page new; 2637 unsigned long counters; 2638 void *freelist; 2639 2640 do { 2641 freelist = page->freelist; 2642 counters = page->counters; 2643 2644 new.counters = counters; 2645 VM_BUG_ON(!new.frozen); 2646 2647 new.inuse = page->objects; 2648 new.frozen = freelist != NULL; 2649 2650 } while (!__cmpxchg_double_slab(s, page, 2651 freelist, counters, 2652 NULL, new.counters, 2653 "get_freelist")); 2654 2655 return freelist; 2656 } 2657 2658 /* 2659 * Slow path. The lockless freelist is empty or we need to perform 2660 * debugging duties. 2661 * 2662 * Processing is still very fast if new objects have been freed to the 2663 * regular freelist. In that case we simply take over the regular freelist 2664 * as the lockless freelist and zap the regular freelist. 2665 * 2666 * If that is not working then we fall back to the partial lists. We take the 2667 * first element of the freelist as the object to allocate now and move the 2668 * rest of the freelist to the lockless freelist. 2669 * 2670 * And if we were unable to get a new slab from the partial slab lists then 2671 * we need to allocate a new slab. This is the slowest path since it involves 2672 * a call to the page allocator and the setup of a new slab. 2673 * 2674 * Version of __slab_alloc to use when we know that interrupts are 2675 * already disabled (which is the case for bulk allocation). 2676 */ 2677 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2678 unsigned long addr, struct kmem_cache_cpu *c) 2679 { 2680 void *freelist; 2681 struct page *page; 2682 2683 stat(s, ALLOC_SLOWPATH); 2684 2685 page = c->page; 2686 if (!page) { 2687 /* 2688 * if the node is not online or has no normal memory, just 2689 * ignore the node constraint 2690 */ 2691 if (unlikely(node != NUMA_NO_NODE && 2692 !node_isset(node, slab_nodes))) 2693 node = NUMA_NO_NODE; 2694 goto new_slab; 2695 } 2696 redo: 2697 2698 if (unlikely(!node_match(page, node))) { 2699 /* 2700 * same as above but node_match() being false already 2701 * implies node != NUMA_NO_NODE 2702 */ 2703 if (!node_isset(node, slab_nodes)) { 2704 node = NUMA_NO_NODE; 2705 goto redo; 2706 } else { 2707 stat(s, ALLOC_NODE_MISMATCH); 2708 deactivate_slab(s, page, c->freelist, c); 2709 goto new_slab; 2710 } 2711 } 2712 2713 /* 2714 * By rights, we should be searching for a slab page that was 2715 * PFMEMALLOC but right now, we are losing the pfmemalloc 2716 * information when the page leaves the per-cpu allocator 2717 */ 2718 if (unlikely(!pfmemalloc_match(page, gfpflags))) { 2719 deactivate_slab(s, page, c->freelist, c); 2720 goto new_slab; 2721 } 2722 2723 /* must check again c->freelist in case of cpu migration or IRQ */ 2724 freelist = c->freelist; 2725 if (freelist) 2726 goto load_freelist; 2727 2728 freelist = get_freelist(s, page); 2729 2730 if (!freelist) { 2731 c->page = NULL; 2732 stat(s, DEACTIVATE_BYPASS); 2733 goto new_slab; 2734 } 2735 2736 stat(s, ALLOC_REFILL); 2737 2738 load_freelist: 2739 /* 2740 * freelist is pointing to the list of objects to be used. 2741 * page is pointing to the page from which the objects are obtained. 2742 * That page must be frozen for per cpu allocations to work. 2743 */ 2744 VM_BUG_ON(!c->page->frozen); 2745 c->freelist = get_freepointer(s, freelist); 2746 c->tid = next_tid(c->tid); 2747 return freelist; 2748 2749 new_slab: 2750 2751 if (slub_percpu_partial(c)) { 2752 page = c->page = slub_percpu_partial(c); 2753 slub_set_percpu_partial(c, page); 2754 stat(s, CPU_PARTIAL_ALLOC); 2755 goto redo; 2756 } 2757 2758 freelist = new_slab_objects(s, gfpflags, node, &c); 2759 2760 if (unlikely(!freelist)) { 2761 slab_out_of_memory(s, gfpflags, node); 2762 return NULL; 2763 } 2764 2765 page = c->page; 2766 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags))) 2767 goto load_freelist; 2768 2769 /* Only entered in the debug case */ 2770 if (kmem_cache_debug(s) && 2771 !alloc_debug_processing(s, page, freelist, addr)) 2772 goto new_slab; /* Slab failed checks. Next slab needed */ 2773 2774 deactivate_slab(s, page, get_freepointer(s, freelist), c); 2775 return freelist; 2776 } 2777 2778 /* 2779 * Another one that disabled interrupt and compensates for possible 2780 * cpu changes by refetching the per cpu area pointer. 2781 */ 2782 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2783 unsigned long addr, struct kmem_cache_cpu *c) 2784 { 2785 void *p; 2786 unsigned long flags; 2787 2788 local_irq_save(flags); 2789 #ifdef CONFIG_PREEMPTION 2790 /* 2791 * We may have been preempted and rescheduled on a different 2792 * cpu before disabling interrupts. Need to reload cpu area 2793 * pointer. 2794 */ 2795 c = this_cpu_ptr(s->cpu_slab); 2796 #endif 2797 2798 p = ___slab_alloc(s, gfpflags, node, addr, c); 2799 local_irq_restore(flags); 2800 return p; 2801 } 2802 2803 /* 2804 * If the object has been wiped upon free, make sure it's fully initialized by 2805 * zeroing out freelist pointer. 2806 */ 2807 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s, 2808 void *obj) 2809 { 2810 if (unlikely(slab_want_init_on_free(s)) && obj) 2811 memset((void *)((char *)kasan_reset_tag(obj) + s->offset), 2812 0, sizeof(void *)); 2813 } 2814 2815 /* 2816 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 2817 * have the fastpath folded into their functions. So no function call 2818 * overhead for requests that can be satisfied on the fastpath. 2819 * 2820 * The fastpath works by first checking if the lockless freelist can be used. 2821 * If not then __slab_alloc is called for slow processing. 2822 * 2823 * Otherwise we can simply pick the next object from the lockless free list. 2824 */ 2825 static __always_inline void *slab_alloc_node(struct kmem_cache *s, 2826 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) 2827 { 2828 void *object; 2829 struct kmem_cache_cpu *c; 2830 struct page *page; 2831 unsigned long tid; 2832 struct obj_cgroup *objcg = NULL; 2833 bool init = false; 2834 2835 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags); 2836 if (!s) 2837 return NULL; 2838 2839 object = kfence_alloc(s, orig_size, gfpflags); 2840 if (unlikely(object)) 2841 goto out; 2842 2843 redo: 2844 /* 2845 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 2846 * enabled. We may switch back and forth between cpus while 2847 * reading from one cpu area. That does not matter as long 2848 * as we end up on the original cpu again when doing the cmpxchg. 2849 * 2850 * We should guarantee that tid and kmem_cache are retrieved on 2851 * the same cpu. It could be different if CONFIG_PREEMPTION so we need 2852 * to check if it is matched or not. 2853 */ 2854 do { 2855 tid = this_cpu_read(s->cpu_slab->tid); 2856 c = raw_cpu_ptr(s->cpu_slab); 2857 } while (IS_ENABLED(CONFIG_PREEMPTION) && 2858 unlikely(tid != READ_ONCE(c->tid))); 2859 2860 /* 2861 * Irqless object alloc/free algorithm used here depends on sequence 2862 * of fetching cpu_slab's data. tid should be fetched before anything 2863 * on c to guarantee that object and page associated with previous tid 2864 * won't be used with current tid. If we fetch tid first, object and 2865 * page could be one associated with next tid and our alloc/free 2866 * request will be failed. In this case, we will retry. So, no problem. 2867 */ 2868 barrier(); 2869 2870 /* 2871 * The transaction ids are globally unique per cpu and per operation on 2872 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 2873 * occurs on the right processor and that there was no operation on the 2874 * linked list in between. 2875 */ 2876 2877 object = c->freelist; 2878 page = c->page; 2879 if (unlikely(!object || !page || !node_match(page, node))) { 2880 object = __slab_alloc(s, gfpflags, node, addr, c); 2881 } else { 2882 void *next_object = get_freepointer_safe(s, object); 2883 2884 /* 2885 * The cmpxchg will only match if there was no additional 2886 * operation and if we are on the right processor. 2887 * 2888 * The cmpxchg does the following atomically (without lock 2889 * semantics!) 2890 * 1. Relocate first pointer to the current per cpu area. 2891 * 2. Verify that tid and freelist have not been changed 2892 * 3. If they were not changed replace tid and freelist 2893 * 2894 * Since this is without lock semantics the protection is only 2895 * against code executing on this cpu *not* from access by 2896 * other cpus. 2897 */ 2898 if (unlikely(!this_cpu_cmpxchg_double( 2899 s->cpu_slab->freelist, s->cpu_slab->tid, 2900 object, tid, 2901 next_object, next_tid(tid)))) { 2902 2903 note_cmpxchg_failure("slab_alloc", s, tid); 2904 goto redo; 2905 } 2906 prefetch_freepointer(s, next_object); 2907 stat(s, ALLOC_FASTPATH); 2908 } 2909 2910 maybe_wipe_obj_freeptr(s, object); 2911 init = slab_want_init_on_alloc(gfpflags, s); 2912 2913 out: 2914 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init); 2915 2916 return object; 2917 } 2918 2919 static __always_inline void *slab_alloc(struct kmem_cache *s, 2920 gfp_t gfpflags, unsigned long addr, size_t orig_size) 2921 { 2922 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size); 2923 } 2924 2925 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 2926 { 2927 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size); 2928 2929 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, 2930 s->size, gfpflags); 2931 2932 return ret; 2933 } 2934 EXPORT_SYMBOL(kmem_cache_alloc); 2935 2936 #ifdef CONFIG_TRACING 2937 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2938 { 2939 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size); 2940 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2941 ret = kasan_kmalloc(s, ret, size, gfpflags); 2942 return ret; 2943 } 2944 EXPORT_SYMBOL(kmem_cache_alloc_trace); 2945 #endif 2946 2947 #ifdef CONFIG_NUMA 2948 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2949 { 2950 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size); 2951 2952 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2953 s->object_size, s->size, gfpflags, node); 2954 2955 return ret; 2956 } 2957 EXPORT_SYMBOL(kmem_cache_alloc_node); 2958 2959 #ifdef CONFIG_TRACING 2960 void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2961 gfp_t gfpflags, 2962 int node, size_t size) 2963 { 2964 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size); 2965 2966 trace_kmalloc_node(_RET_IP_, ret, 2967 size, s->size, gfpflags, node); 2968 2969 ret = kasan_kmalloc(s, ret, size, gfpflags); 2970 return ret; 2971 } 2972 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2973 #endif 2974 #endif /* CONFIG_NUMA */ 2975 2976 /* 2977 * Slow path handling. This may still be called frequently since objects 2978 * have a longer lifetime than the cpu slabs in most processing loads. 2979 * 2980 * So we still attempt to reduce cache line usage. Just take the slab 2981 * lock and free the item. If there is no additional partial page 2982 * handling required then we can return immediately. 2983 */ 2984 static void __slab_free(struct kmem_cache *s, struct page *page, 2985 void *head, void *tail, int cnt, 2986 unsigned long addr) 2987 2988 { 2989 void *prior; 2990 int was_frozen; 2991 struct page new; 2992 unsigned long counters; 2993 struct kmem_cache_node *n = NULL; 2994 unsigned long flags; 2995 2996 stat(s, FREE_SLOWPATH); 2997 2998 if (kfence_free(head)) 2999 return; 3000 3001 if (kmem_cache_debug(s) && 3002 !free_debug_processing(s, page, head, tail, cnt, addr)) 3003 return; 3004 3005 do { 3006 if (unlikely(n)) { 3007 spin_unlock_irqrestore(&n->list_lock, flags); 3008 n = NULL; 3009 } 3010 prior = page->freelist; 3011 counters = page->counters; 3012 set_freepointer(s, tail, prior); 3013 new.counters = counters; 3014 was_frozen = new.frozen; 3015 new.inuse -= cnt; 3016 if ((!new.inuse || !prior) && !was_frozen) { 3017 3018 if (kmem_cache_has_cpu_partial(s) && !prior) { 3019 3020 /* 3021 * Slab was on no list before and will be 3022 * partially empty 3023 * We can defer the list move and instead 3024 * freeze it. 3025 */ 3026 new.frozen = 1; 3027 3028 } else { /* Needs to be taken off a list */ 3029 3030 n = get_node(s, page_to_nid(page)); 3031 /* 3032 * Speculatively acquire the list_lock. 3033 * If the cmpxchg does not succeed then we may 3034 * drop the list_lock without any processing. 3035 * 3036 * Otherwise the list_lock will synchronize with 3037 * other processors updating the list of slabs. 3038 */ 3039 spin_lock_irqsave(&n->list_lock, flags); 3040 3041 } 3042 } 3043 3044 } while (!cmpxchg_double_slab(s, page, 3045 prior, counters, 3046 head, new.counters, 3047 "__slab_free")); 3048 3049 if (likely(!n)) { 3050 3051 if (likely(was_frozen)) { 3052 /* 3053 * The list lock was not taken therefore no list 3054 * activity can be necessary. 3055 */ 3056 stat(s, FREE_FROZEN); 3057 } else if (new.frozen) { 3058 /* 3059 * If we just froze the page then put it onto the 3060 * per cpu partial list. 3061 */ 3062 put_cpu_partial(s, page, 1); 3063 stat(s, CPU_PARTIAL_FREE); 3064 } 3065 3066 return; 3067 } 3068 3069 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) 3070 goto slab_empty; 3071 3072 /* 3073 * Objects left in the slab. If it was not on the partial list before 3074 * then add it. 3075 */ 3076 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { 3077 remove_full(s, n, page); 3078 add_partial(n, page, DEACTIVATE_TO_TAIL); 3079 stat(s, FREE_ADD_PARTIAL); 3080 } 3081 spin_unlock_irqrestore(&n->list_lock, flags); 3082 return; 3083 3084 slab_empty: 3085 if (prior) { 3086 /* 3087 * Slab on the partial list. 3088 */ 3089 remove_partial(n, page); 3090 stat(s, FREE_REMOVE_PARTIAL); 3091 } else { 3092 /* Slab must be on the full list */ 3093 remove_full(s, n, page); 3094 } 3095 3096 spin_unlock_irqrestore(&n->list_lock, flags); 3097 stat(s, FREE_SLAB); 3098 discard_slab(s, page); 3099 } 3100 3101 /* 3102 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 3103 * can perform fastpath freeing without additional function calls. 3104 * 3105 * The fastpath is only possible if we are freeing to the current cpu slab 3106 * of this processor. This typically the case if we have just allocated 3107 * the item before. 3108 * 3109 * If fastpath is not possible then fall back to __slab_free where we deal 3110 * with all sorts of special processing. 3111 * 3112 * Bulk free of a freelist with several objects (all pointing to the 3113 * same page) possible by specifying head and tail ptr, plus objects 3114 * count (cnt). Bulk free indicated by tail pointer being set. 3115 */ 3116 static __always_inline void do_slab_free(struct kmem_cache *s, 3117 struct page *page, void *head, void *tail, 3118 int cnt, unsigned long addr) 3119 { 3120 void *tail_obj = tail ? : head; 3121 struct kmem_cache_cpu *c; 3122 unsigned long tid; 3123 3124 memcg_slab_free_hook(s, &head, 1); 3125 redo: 3126 /* 3127 * Determine the currently cpus per cpu slab. 3128 * The cpu may change afterward. However that does not matter since 3129 * data is retrieved via this pointer. If we are on the same cpu 3130 * during the cmpxchg then the free will succeed. 3131 */ 3132 do { 3133 tid = this_cpu_read(s->cpu_slab->tid); 3134 c = raw_cpu_ptr(s->cpu_slab); 3135 } while (IS_ENABLED(CONFIG_PREEMPTION) && 3136 unlikely(tid != READ_ONCE(c->tid))); 3137 3138 /* Same with comment on barrier() in slab_alloc_node() */ 3139 barrier(); 3140 3141 if (likely(page == c->page)) { 3142 void **freelist = READ_ONCE(c->freelist); 3143 3144 set_freepointer(s, tail_obj, freelist); 3145 3146 if (unlikely(!this_cpu_cmpxchg_double( 3147 s->cpu_slab->freelist, s->cpu_slab->tid, 3148 freelist, tid, 3149 head, next_tid(tid)))) { 3150 3151 note_cmpxchg_failure("slab_free", s, tid); 3152 goto redo; 3153 } 3154 stat(s, FREE_FASTPATH); 3155 } else 3156 __slab_free(s, page, head, tail_obj, cnt, addr); 3157 3158 } 3159 3160 static __always_inline void slab_free(struct kmem_cache *s, struct page *page, 3161 void *head, void *tail, int cnt, 3162 unsigned long addr) 3163 { 3164 /* 3165 * With KASAN enabled slab_free_freelist_hook modifies the freelist 3166 * to remove objects, whose reuse must be delayed. 3167 */ 3168 if (slab_free_freelist_hook(s, &head, &tail)) 3169 do_slab_free(s, page, head, tail, cnt, addr); 3170 } 3171 3172 #ifdef CONFIG_KASAN_GENERIC 3173 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) 3174 { 3175 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr); 3176 } 3177 #endif 3178 3179 void kmem_cache_free(struct kmem_cache *s, void *x) 3180 { 3181 s = cache_from_obj(s, x); 3182 if (!s) 3183 return; 3184 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_); 3185 trace_kmem_cache_free(_RET_IP_, x, s->name); 3186 } 3187 EXPORT_SYMBOL(kmem_cache_free); 3188 3189 struct detached_freelist { 3190 struct page *page; 3191 void *tail; 3192 void *freelist; 3193 int cnt; 3194 struct kmem_cache *s; 3195 }; 3196 3197 /* 3198 * This function progressively scans the array with free objects (with 3199 * a limited look ahead) and extract objects belonging to the same 3200 * page. It builds a detached freelist directly within the given 3201 * page/objects. This can happen without any need for 3202 * synchronization, because the objects are owned by running process. 3203 * The freelist is build up as a single linked list in the objects. 3204 * The idea is, that this detached freelist can then be bulk 3205 * transferred to the real freelist(s), but only requiring a single 3206 * synchronization primitive. Look ahead in the array is limited due 3207 * to performance reasons. 3208 */ 3209 static inline 3210 int build_detached_freelist(struct kmem_cache *s, size_t size, 3211 void **p, struct detached_freelist *df) 3212 { 3213 size_t first_skipped_index = 0; 3214 int lookahead = 3; 3215 void *object; 3216 struct page *page; 3217 3218 /* Always re-init detached_freelist */ 3219 df->page = NULL; 3220 3221 do { 3222 object = p[--size]; 3223 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */ 3224 } while (!object && size); 3225 3226 if (!object) 3227 return 0; 3228 3229 page = virt_to_head_page(object); 3230 if (!s) { 3231 /* Handle kalloc'ed objects */ 3232 if (unlikely(!PageSlab(page))) { 3233 BUG_ON(!PageCompound(page)); 3234 kfree_hook(object); 3235 __free_pages(page, compound_order(page)); 3236 p[size] = NULL; /* mark object processed */ 3237 return size; 3238 } 3239 /* Derive kmem_cache from object */ 3240 df->s = page->slab_cache; 3241 } else { 3242 df->s = cache_from_obj(s, object); /* Support for memcg */ 3243 } 3244 3245 if (is_kfence_address(object)) { 3246 slab_free_hook(df->s, object, false); 3247 __kfence_free(object); 3248 p[size] = NULL; /* mark object processed */ 3249 return size; 3250 } 3251 3252 /* Start new detached freelist */ 3253 df->page = page; 3254 set_freepointer(df->s, object, NULL); 3255 df->tail = object; 3256 df->freelist = object; 3257 p[size] = NULL; /* mark object processed */ 3258 df->cnt = 1; 3259 3260 while (size) { 3261 object = p[--size]; 3262 if (!object) 3263 continue; /* Skip processed objects */ 3264 3265 /* df->page is always set at this point */ 3266 if (df->page == virt_to_head_page(object)) { 3267 /* Opportunity build freelist */ 3268 set_freepointer(df->s, object, df->freelist); 3269 df->freelist = object; 3270 df->cnt++; 3271 p[size] = NULL; /* mark object processed */ 3272 3273 continue; 3274 } 3275 3276 /* Limit look ahead search */ 3277 if (!--lookahead) 3278 break; 3279 3280 if (!first_skipped_index) 3281 first_skipped_index = size + 1; 3282 } 3283 3284 return first_skipped_index; 3285 } 3286 3287 /* Note that interrupts must be enabled when calling this function. */ 3288 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) 3289 { 3290 if (WARN_ON(!size)) 3291 return; 3292 3293 memcg_slab_free_hook(s, p, size); 3294 do { 3295 struct detached_freelist df; 3296 3297 size = build_detached_freelist(s, size, p, &df); 3298 if (!df.page) 3299 continue; 3300 3301 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_); 3302 } while (likely(size)); 3303 } 3304 EXPORT_SYMBOL(kmem_cache_free_bulk); 3305 3306 /* Note that interrupts must be enabled when calling this function. */ 3307 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, 3308 void **p) 3309 { 3310 struct kmem_cache_cpu *c; 3311 int i; 3312 struct obj_cgroup *objcg = NULL; 3313 3314 /* memcg and kmem_cache debug support */ 3315 s = slab_pre_alloc_hook(s, &objcg, size, flags); 3316 if (unlikely(!s)) 3317 return false; 3318 /* 3319 * Drain objects in the per cpu slab, while disabling local 3320 * IRQs, which protects against PREEMPT and interrupts 3321 * handlers invoking normal fastpath. 3322 */ 3323 local_irq_disable(); 3324 c = this_cpu_ptr(s->cpu_slab); 3325 3326 for (i = 0; i < size; i++) { 3327 void *object = kfence_alloc(s, s->object_size, flags); 3328 3329 if (unlikely(object)) { 3330 p[i] = object; 3331 continue; 3332 } 3333 3334 object = c->freelist; 3335 if (unlikely(!object)) { 3336 /* 3337 * We may have removed an object from c->freelist using 3338 * the fastpath in the previous iteration; in that case, 3339 * c->tid has not been bumped yet. 3340 * Since ___slab_alloc() may reenable interrupts while 3341 * allocating memory, we should bump c->tid now. 3342 */ 3343 c->tid = next_tid(c->tid); 3344 3345 /* 3346 * Invoking slow path likely have side-effect 3347 * of re-populating per CPU c->freelist 3348 */ 3349 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, 3350 _RET_IP_, c); 3351 if (unlikely(!p[i])) 3352 goto error; 3353 3354 c = this_cpu_ptr(s->cpu_slab); 3355 maybe_wipe_obj_freeptr(s, p[i]); 3356 3357 continue; /* goto for-loop */ 3358 } 3359 c->freelist = get_freepointer(s, object); 3360 p[i] = object; 3361 maybe_wipe_obj_freeptr(s, p[i]); 3362 } 3363 c->tid = next_tid(c->tid); 3364 local_irq_enable(); 3365 3366 /* 3367 * memcg and kmem_cache debug support and memory initialization. 3368 * Done outside of the IRQ disabled fastpath loop. 3369 */ 3370 slab_post_alloc_hook(s, objcg, flags, size, p, 3371 slab_want_init_on_alloc(flags, s)); 3372 return i; 3373 error: 3374 local_irq_enable(); 3375 slab_post_alloc_hook(s, objcg, flags, i, p, false); 3376 __kmem_cache_free_bulk(s, i, p); 3377 return 0; 3378 } 3379 EXPORT_SYMBOL(kmem_cache_alloc_bulk); 3380 3381 3382 /* 3383 * Object placement in a slab is made very easy because we always start at 3384 * offset 0. If we tune the size of the object to the alignment then we can 3385 * get the required alignment by putting one properly sized object after 3386 * another. 3387 * 3388 * Notice that the allocation order determines the sizes of the per cpu 3389 * caches. Each processor has always one slab available for allocations. 3390 * Increasing the allocation order reduces the number of times that slabs 3391 * must be moved on and off the partial lists and is therefore a factor in 3392 * locking overhead. 3393 */ 3394 3395 /* 3396 * Minimum / Maximum order of slab pages. This influences locking overhead 3397 * and slab fragmentation. A higher order reduces the number of partial slabs 3398 * and increases the number of allocations possible without having to 3399 * take the list_lock. 3400 */ 3401 static unsigned int slub_min_order; 3402 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 3403 static unsigned int slub_min_objects; 3404 3405 /* 3406 * Calculate the order of allocation given an slab object size. 3407 * 3408 * The order of allocation has significant impact on performance and other 3409 * system components. Generally order 0 allocations should be preferred since 3410 * order 0 does not cause fragmentation in the page allocator. Larger objects 3411 * be problematic to put into order 0 slabs because there may be too much 3412 * unused space left. We go to a higher order if more than 1/16th of the slab 3413 * would be wasted. 3414 * 3415 * In order to reach satisfactory performance we must ensure that a minimum 3416 * number of objects is in one slab. Otherwise we may generate too much 3417 * activity on the partial lists which requires taking the list_lock. This is 3418 * less a concern for large slabs though which are rarely used. 3419 * 3420 * slub_max_order specifies the order where we begin to stop considering the 3421 * number of objects in a slab as critical. If we reach slub_max_order then 3422 * we try to keep the page order as low as possible. So we accept more waste 3423 * of space in favor of a small page order. 3424 * 3425 * Higher order allocations also allow the placement of more objects in a 3426 * slab and thereby reduce object handling overhead. If the user has 3427 * requested a higher minimum order then we start with that one instead of 3428 * the smallest order which will fit the object. 3429 */ 3430 static inline unsigned int slab_order(unsigned int size, 3431 unsigned int min_objects, unsigned int max_order, 3432 unsigned int fract_leftover) 3433 { 3434 unsigned int min_order = slub_min_order; 3435 unsigned int order; 3436 3437 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE) 3438 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 3439 3440 for (order = max(min_order, (unsigned int)get_order(min_objects * size)); 3441 order <= max_order; order++) { 3442 3443 unsigned int slab_size = (unsigned int)PAGE_SIZE << order; 3444 unsigned int rem; 3445 3446 rem = slab_size % size; 3447 3448 if (rem <= slab_size / fract_leftover) 3449 break; 3450 } 3451 3452 return order; 3453 } 3454 3455 static inline int calculate_order(unsigned int size) 3456 { 3457 unsigned int order; 3458 unsigned int min_objects; 3459 unsigned int max_objects; 3460 unsigned int nr_cpus; 3461 3462 /* 3463 * Attempt to find best configuration for a slab. This 3464 * works by first attempting to generate a layout with 3465 * the best configuration and backing off gradually. 3466 * 3467 * First we increase the acceptable waste in a slab. Then 3468 * we reduce the minimum objects required in a slab. 3469 */ 3470 min_objects = slub_min_objects; 3471 if (!min_objects) { 3472 /* 3473 * Some architectures will only update present cpus when 3474 * onlining them, so don't trust the number if it's just 1. But 3475 * we also don't want to use nr_cpu_ids always, as on some other 3476 * architectures, there can be many possible cpus, but never 3477 * onlined. Here we compromise between trying to avoid too high 3478 * order on systems that appear larger than they are, and too 3479 * low order on systems that appear smaller than they are. 3480 */ 3481 nr_cpus = num_present_cpus(); 3482 if (nr_cpus <= 1) 3483 nr_cpus = nr_cpu_ids; 3484 min_objects = 4 * (fls(nr_cpus) + 1); 3485 } 3486 max_objects = order_objects(slub_max_order, size); 3487 min_objects = min(min_objects, max_objects); 3488 3489 while (min_objects > 1) { 3490 unsigned int fraction; 3491 3492 fraction = 16; 3493 while (fraction >= 4) { 3494 order = slab_order(size, min_objects, 3495 slub_max_order, fraction); 3496 if (order <= slub_max_order) 3497 return order; 3498 fraction /= 2; 3499 } 3500 min_objects--; 3501 } 3502 3503 /* 3504 * We were unable to place multiple objects in a slab. Now 3505 * lets see if we can place a single object there. 3506 */ 3507 order = slab_order(size, 1, slub_max_order, 1); 3508 if (order <= slub_max_order) 3509 return order; 3510 3511 /* 3512 * Doh this slab cannot be placed using slub_max_order. 3513 */ 3514 order = slab_order(size, 1, MAX_ORDER, 1); 3515 if (order < MAX_ORDER) 3516 return order; 3517 return -ENOSYS; 3518 } 3519 3520 static void 3521 init_kmem_cache_node(struct kmem_cache_node *n) 3522 { 3523 n->nr_partial = 0; 3524 spin_lock_init(&n->list_lock); 3525 INIT_LIST_HEAD(&n->partial); 3526 #ifdef CONFIG_SLUB_DEBUG 3527 atomic_long_set(&n->nr_slabs, 0); 3528 atomic_long_set(&n->total_objects, 0); 3529 INIT_LIST_HEAD(&n->full); 3530 #endif 3531 } 3532 3533 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 3534 { 3535 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 3536 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu)); 3537 3538 /* 3539 * Must align to double word boundary for the double cmpxchg 3540 * instructions to work; see __pcpu_double_call_return_bool(). 3541 */ 3542 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 3543 2 * sizeof(void *)); 3544 3545 if (!s->cpu_slab) 3546 return 0; 3547 3548 init_kmem_cache_cpus(s); 3549 3550 return 1; 3551 } 3552 3553 static struct kmem_cache *kmem_cache_node; 3554 3555 /* 3556 * No kmalloc_node yet so do it by hand. We know that this is the first 3557 * slab on the node for this slabcache. There are no concurrent accesses 3558 * possible. 3559 * 3560 * Note that this function only works on the kmem_cache_node 3561 * when allocating for the kmem_cache_node. This is used for bootstrapping 3562 * memory on a fresh node that has no slab structures yet. 3563 */ 3564 static void early_kmem_cache_node_alloc(int node) 3565 { 3566 struct page *page; 3567 struct kmem_cache_node *n; 3568 3569 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 3570 3571 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 3572 3573 BUG_ON(!page); 3574 if (page_to_nid(page) != node) { 3575 pr_err("SLUB: Unable to allocate memory from node %d\n", node); 3576 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n"); 3577 } 3578 3579 n = page->freelist; 3580 BUG_ON(!n); 3581 #ifdef CONFIG_SLUB_DEBUG 3582 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 3583 init_tracking(kmem_cache_node, n); 3584 #endif 3585 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false); 3586 page->freelist = get_freepointer(kmem_cache_node, n); 3587 page->inuse = 1; 3588 page->frozen = 0; 3589 kmem_cache_node->node[node] = n; 3590 init_kmem_cache_node(n); 3591 inc_slabs_node(kmem_cache_node, node, page->objects); 3592 3593 /* 3594 * No locks need to be taken here as it has just been 3595 * initialized and there is no concurrent access. 3596 */ 3597 __add_partial(n, page, DEACTIVATE_TO_HEAD); 3598 } 3599 3600 static void free_kmem_cache_nodes(struct kmem_cache *s) 3601 { 3602 int node; 3603 struct kmem_cache_node *n; 3604 3605 for_each_kmem_cache_node(s, node, n) { 3606 s->node[node] = NULL; 3607 kmem_cache_free(kmem_cache_node, n); 3608 } 3609 } 3610 3611 void __kmem_cache_release(struct kmem_cache *s) 3612 { 3613 cache_random_seq_destroy(s); 3614 free_percpu(s->cpu_slab); 3615 free_kmem_cache_nodes(s); 3616 } 3617 3618 static int init_kmem_cache_nodes(struct kmem_cache *s) 3619 { 3620 int node; 3621 3622 for_each_node_mask(node, slab_nodes) { 3623 struct kmem_cache_node *n; 3624 3625 if (slab_state == DOWN) { 3626 early_kmem_cache_node_alloc(node); 3627 continue; 3628 } 3629 n = kmem_cache_alloc_node(kmem_cache_node, 3630 GFP_KERNEL, node); 3631 3632 if (!n) { 3633 free_kmem_cache_nodes(s); 3634 return 0; 3635 } 3636 3637 init_kmem_cache_node(n); 3638 s->node[node] = n; 3639 } 3640 return 1; 3641 } 3642 3643 static void set_min_partial(struct kmem_cache *s, unsigned long min) 3644 { 3645 if (min < MIN_PARTIAL) 3646 min = MIN_PARTIAL; 3647 else if (min > MAX_PARTIAL) 3648 min = MAX_PARTIAL; 3649 s->min_partial = min; 3650 } 3651 3652 static void set_cpu_partial(struct kmem_cache *s) 3653 { 3654 #ifdef CONFIG_SLUB_CPU_PARTIAL 3655 /* 3656 * cpu_partial determined the maximum number of objects kept in the 3657 * per cpu partial lists of a processor. 3658 * 3659 * Per cpu partial lists mainly contain slabs that just have one 3660 * object freed. If they are used for allocation then they can be 3661 * filled up again with minimal effort. The slab will never hit the 3662 * per node partial lists and therefore no locking will be required. 3663 * 3664 * This setting also determines 3665 * 3666 * A) The number of objects from per cpu partial slabs dumped to the 3667 * per node list when we reach the limit. 3668 * B) The number of objects in cpu partial slabs to extract from the 3669 * per node list when we run out of per cpu objects. We only fetch 3670 * 50% to keep some capacity around for frees. 3671 */ 3672 if (!kmem_cache_has_cpu_partial(s)) 3673 slub_set_cpu_partial(s, 0); 3674 else if (s->size >= PAGE_SIZE) 3675 slub_set_cpu_partial(s, 2); 3676 else if (s->size >= 1024) 3677 slub_set_cpu_partial(s, 6); 3678 else if (s->size >= 256) 3679 slub_set_cpu_partial(s, 13); 3680 else 3681 slub_set_cpu_partial(s, 30); 3682 #endif 3683 } 3684 3685 /* 3686 * calculate_sizes() determines the order and the distribution of data within 3687 * a slab object. 3688 */ 3689 static int calculate_sizes(struct kmem_cache *s, int forced_order) 3690 { 3691 slab_flags_t flags = s->flags; 3692 unsigned int size = s->object_size; 3693 unsigned int order; 3694 3695 /* 3696 * Round up object size to the next word boundary. We can only 3697 * place the free pointer at word boundaries and this determines 3698 * the possible location of the free pointer. 3699 */ 3700 size = ALIGN(size, sizeof(void *)); 3701 3702 #ifdef CONFIG_SLUB_DEBUG 3703 /* 3704 * Determine if we can poison the object itself. If the user of 3705 * the slab may touch the object after free or before allocation 3706 * then we should never poison the object itself. 3707 */ 3708 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && 3709 !s->ctor) 3710 s->flags |= __OBJECT_POISON; 3711 else 3712 s->flags &= ~__OBJECT_POISON; 3713 3714 3715 /* 3716 * If we are Redzoning then check if there is some space between the 3717 * end of the object and the free pointer. If not then add an 3718 * additional word to have some bytes to store Redzone information. 3719 */ 3720 if ((flags & SLAB_RED_ZONE) && size == s->object_size) 3721 size += sizeof(void *); 3722 #endif 3723 3724 /* 3725 * With that we have determined the number of bytes in actual use 3726 * by the object and redzoning. 3727 */ 3728 s->inuse = size; 3729 3730 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || 3731 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) || 3732 s->ctor) { 3733 /* 3734 * Relocate free pointer after the object if it is not 3735 * permitted to overwrite the first word of the object on 3736 * kmem_cache_free. 3737 * 3738 * This is the case if we do RCU, have a constructor or 3739 * destructor, are poisoning the objects, or are 3740 * redzoning an object smaller than sizeof(void *). 3741 * 3742 * The assumption that s->offset >= s->inuse means free 3743 * pointer is outside of the object is used in the 3744 * freeptr_outside_object() function. If that is no 3745 * longer true, the function needs to be modified. 3746 */ 3747 s->offset = size; 3748 size += sizeof(void *); 3749 } else { 3750 /* 3751 * Store freelist pointer near middle of object to keep 3752 * it away from the edges of the object to avoid small 3753 * sized over/underflows from neighboring allocations. 3754 */ 3755 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *)); 3756 } 3757 3758 #ifdef CONFIG_SLUB_DEBUG 3759 if (flags & SLAB_STORE_USER) 3760 /* 3761 * Need to store information about allocs and frees after 3762 * the object. 3763 */ 3764 size += 2 * sizeof(struct track); 3765 #endif 3766 3767 kasan_cache_create(s, &size, &s->flags); 3768 #ifdef CONFIG_SLUB_DEBUG 3769 if (flags & SLAB_RED_ZONE) { 3770 /* 3771 * Add some empty padding so that we can catch 3772 * overwrites from earlier objects rather than let 3773 * tracking information or the free pointer be 3774 * corrupted if a user writes before the start 3775 * of the object. 3776 */ 3777 size += sizeof(void *); 3778 3779 s->red_left_pad = sizeof(void *); 3780 s->red_left_pad = ALIGN(s->red_left_pad, s->align); 3781 size += s->red_left_pad; 3782 } 3783 #endif 3784 3785 /* 3786 * SLUB stores one object immediately after another beginning from 3787 * offset 0. In order to align the objects we have to simply size 3788 * each object to conform to the alignment. 3789 */ 3790 size = ALIGN(size, s->align); 3791 s->size = size; 3792 s->reciprocal_size = reciprocal_value(size); 3793 if (forced_order >= 0) 3794 order = forced_order; 3795 else 3796 order = calculate_order(size); 3797 3798 if ((int)order < 0) 3799 return 0; 3800 3801 s->allocflags = 0; 3802 if (order) 3803 s->allocflags |= __GFP_COMP; 3804 3805 if (s->flags & SLAB_CACHE_DMA) 3806 s->allocflags |= GFP_DMA; 3807 3808 if (s->flags & SLAB_CACHE_DMA32) 3809 s->allocflags |= GFP_DMA32; 3810 3811 if (s->flags & SLAB_RECLAIM_ACCOUNT) 3812 s->allocflags |= __GFP_RECLAIMABLE; 3813 3814 /* 3815 * Determine the number of objects per slab 3816 */ 3817 s->oo = oo_make(order, size); 3818 s->min = oo_make(get_order(size), size); 3819 if (oo_objects(s->oo) > oo_objects(s->max)) 3820 s->max = s->oo; 3821 3822 return !!oo_objects(s->oo); 3823 } 3824 3825 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) 3826 { 3827 s->flags = kmem_cache_flags(s->size, flags, s->name); 3828 #ifdef CONFIG_SLAB_FREELIST_HARDENED 3829 s->random = get_random_long(); 3830 #endif 3831 3832 if (!calculate_sizes(s, -1)) 3833 goto error; 3834 if (disable_higher_order_debug) { 3835 /* 3836 * Disable debugging flags that store metadata if the min slab 3837 * order increased. 3838 */ 3839 if (get_order(s->size) > get_order(s->object_size)) { 3840 s->flags &= ~DEBUG_METADATA_FLAGS; 3841 s->offset = 0; 3842 if (!calculate_sizes(s, -1)) 3843 goto error; 3844 } 3845 } 3846 3847 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 3848 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 3849 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0) 3850 /* Enable fast mode */ 3851 s->flags |= __CMPXCHG_DOUBLE; 3852 #endif 3853 3854 /* 3855 * The larger the object size is, the more pages we want on the partial 3856 * list to avoid pounding the page allocator excessively. 3857 */ 3858 set_min_partial(s, ilog2(s->size) / 2); 3859 3860 set_cpu_partial(s); 3861 3862 #ifdef CONFIG_NUMA 3863 s->remote_node_defrag_ratio = 1000; 3864 #endif 3865 3866 /* Initialize the pre-computed randomized freelist if slab is up */ 3867 if (slab_state >= UP) { 3868 if (init_cache_random_seq(s)) 3869 goto error; 3870 } 3871 3872 if (!init_kmem_cache_nodes(s)) 3873 goto error; 3874 3875 if (alloc_kmem_cache_cpus(s)) 3876 return 0; 3877 3878 free_kmem_cache_nodes(s); 3879 error: 3880 return -EINVAL; 3881 } 3882 3883 static void list_slab_objects(struct kmem_cache *s, struct page *page, 3884 const char *text) 3885 { 3886 #ifdef CONFIG_SLUB_DEBUG 3887 void *addr = page_address(page); 3888 unsigned long *map; 3889 void *p; 3890 3891 slab_err(s, page, text, s->name); 3892 slab_lock(page); 3893 3894 map = get_map(s, page); 3895 for_each_object(p, s, addr, page->objects) { 3896 3897 if (!test_bit(__obj_to_index(s, addr, p), map)) { 3898 pr_err("Object 0x%p @offset=%tu\n", p, p - addr); 3899 print_tracking(s, p); 3900 } 3901 } 3902 put_map(map); 3903 slab_unlock(page); 3904 #endif 3905 } 3906 3907 /* 3908 * Attempt to free all partial slabs on a node. 3909 * This is called from __kmem_cache_shutdown(). We must take list_lock 3910 * because sysfs file might still access partial list after the shutdowning. 3911 */ 3912 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 3913 { 3914 LIST_HEAD(discard); 3915 struct page *page, *h; 3916 3917 BUG_ON(irqs_disabled()); 3918 spin_lock_irq(&n->list_lock); 3919 list_for_each_entry_safe(page, h, &n->partial, slab_list) { 3920 if (!page->inuse) { 3921 remove_partial(n, page); 3922 list_add(&page->slab_list, &discard); 3923 } else { 3924 list_slab_objects(s, page, 3925 "Objects remaining in %s on __kmem_cache_shutdown()"); 3926 } 3927 } 3928 spin_unlock_irq(&n->list_lock); 3929 3930 list_for_each_entry_safe(page, h, &discard, slab_list) 3931 discard_slab(s, page); 3932 } 3933 3934 bool __kmem_cache_empty(struct kmem_cache *s) 3935 { 3936 int node; 3937 struct kmem_cache_node *n; 3938 3939 for_each_kmem_cache_node(s, node, n) 3940 if (n->nr_partial || slabs_node(s, node)) 3941 return false; 3942 return true; 3943 } 3944 3945 /* 3946 * Release all resources used by a slab cache. 3947 */ 3948 int __kmem_cache_shutdown(struct kmem_cache *s) 3949 { 3950 int node; 3951 struct kmem_cache_node *n; 3952 3953 flush_all(s); 3954 /* Attempt to free all objects */ 3955 for_each_kmem_cache_node(s, node, n) { 3956 free_partial(s, n); 3957 if (n->nr_partial || slabs_node(s, node)) 3958 return 1; 3959 } 3960 return 0; 3961 } 3962 3963 #ifdef CONFIG_PRINTK 3964 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page) 3965 { 3966 void *base; 3967 int __maybe_unused i; 3968 unsigned int objnr; 3969 void *objp; 3970 void *objp0; 3971 struct kmem_cache *s = page->slab_cache; 3972 struct track __maybe_unused *trackp; 3973 3974 kpp->kp_ptr = object; 3975 kpp->kp_page = page; 3976 kpp->kp_slab_cache = s; 3977 base = page_address(page); 3978 objp0 = kasan_reset_tag(object); 3979 #ifdef CONFIG_SLUB_DEBUG 3980 objp = restore_red_left(s, objp0); 3981 #else 3982 objp = objp0; 3983 #endif 3984 objnr = obj_to_index(s, page, objp); 3985 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp); 3986 objp = base + s->size * objnr; 3987 kpp->kp_objp = objp; 3988 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) || 3989 !(s->flags & SLAB_STORE_USER)) 3990 return; 3991 #ifdef CONFIG_SLUB_DEBUG 3992 trackp = get_track(s, objp, TRACK_ALLOC); 3993 kpp->kp_ret = (void *)trackp->addr; 3994 #ifdef CONFIG_STACKTRACE 3995 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) { 3996 kpp->kp_stack[i] = (void *)trackp->addrs[i]; 3997 if (!kpp->kp_stack[i]) 3998 break; 3999 } 4000 #endif 4001 #endif 4002 } 4003 #endif 4004 4005 /******************************************************************** 4006 * Kmalloc subsystem 4007 *******************************************************************/ 4008 4009 static int __init setup_slub_min_order(char *str) 4010 { 4011 get_option(&str, (int *)&slub_min_order); 4012 4013 return 1; 4014 } 4015 4016 __setup("slub_min_order=", setup_slub_min_order); 4017 4018 static int __init setup_slub_max_order(char *str) 4019 { 4020 get_option(&str, (int *)&slub_max_order); 4021 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1); 4022 4023 return 1; 4024 } 4025 4026 __setup("slub_max_order=", setup_slub_max_order); 4027 4028 static int __init setup_slub_min_objects(char *str) 4029 { 4030 get_option(&str, (int *)&slub_min_objects); 4031 4032 return 1; 4033 } 4034 4035 __setup("slub_min_objects=", setup_slub_min_objects); 4036 4037 void *__kmalloc(size_t size, gfp_t flags) 4038 { 4039 struct kmem_cache *s; 4040 void *ret; 4041 4042 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 4043 return kmalloc_large(size, flags); 4044 4045 s = kmalloc_slab(size, flags); 4046 4047 if (unlikely(ZERO_OR_NULL_PTR(s))) 4048 return s; 4049 4050 ret = slab_alloc(s, flags, _RET_IP_, size); 4051 4052 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 4053 4054 ret = kasan_kmalloc(s, ret, size, flags); 4055 4056 return ret; 4057 } 4058 EXPORT_SYMBOL(__kmalloc); 4059 4060 #ifdef CONFIG_NUMA 4061 static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 4062 { 4063 struct page *page; 4064 void *ptr = NULL; 4065 unsigned int order = get_order(size); 4066 4067 flags |= __GFP_COMP; 4068 page = alloc_pages_node(node, flags, order); 4069 if (page) { 4070 ptr = page_address(page); 4071 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, 4072 PAGE_SIZE << order); 4073 } 4074 4075 return kmalloc_large_node_hook(ptr, size, flags); 4076 } 4077 4078 void *__kmalloc_node(size_t size, gfp_t flags, int node) 4079 { 4080 struct kmem_cache *s; 4081 void *ret; 4082 4083 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 4084 ret = kmalloc_large_node(size, flags, node); 4085 4086 trace_kmalloc_node(_RET_IP_, ret, 4087 size, PAGE_SIZE << get_order(size), 4088 flags, node); 4089 4090 return ret; 4091 } 4092 4093 s = kmalloc_slab(size, flags); 4094 4095 if (unlikely(ZERO_OR_NULL_PTR(s))) 4096 return s; 4097 4098 ret = slab_alloc_node(s, flags, node, _RET_IP_, size); 4099 4100 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 4101 4102 ret = kasan_kmalloc(s, ret, size, flags); 4103 4104 return ret; 4105 } 4106 EXPORT_SYMBOL(__kmalloc_node); 4107 #endif /* CONFIG_NUMA */ 4108 4109 #ifdef CONFIG_HARDENED_USERCOPY 4110 /* 4111 * Rejects incorrectly sized objects and objects that are to be copied 4112 * to/from userspace but do not fall entirely within the containing slab 4113 * cache's usercopy region. 4114 * 4115 * Returns NULL if check passes, otherwise const char * to name of cache 4116 * to indicate an error. 4117 */ 4118 void __check_heap_object(const void *ptr, unsigned long n, struct page *page, 4119 bool to_user) 4120 { 4121 struct kmem_cache *s; 4122 unsigned int offset; 4123 size_t object_size; 4124 bool is_kfence = is_kfence_address(ptr); 4125 4126 ptr = kasan_reset_tag(ptr); 4127 4128 /* Find object and usable object size. */ 4129 s = page->slab_cache; 4130 4131 /* Reject impossible pointers. */ 4132 if (ptr < page_address(page)) 4133 usercopy_abort("SLUB object not in SLUB page?!", NULL, 4134 to_user, 0, n); 4135 4136 /* Find offset within object. */ 4137 if (is_kfence) 4138 offset = ptr - kfence_object_start(ptr); 4139 else 4140 offset = (ptr - page_address(page)) % s->size; 4141 4142 /* Adjust for redzone and reject if within the redzone. */ 4143 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) { 4144 if (offset < s->red_left_pad) 4145 usercopy_abort("SLUB object in left red zone", 4146 s->name, to_user, offset, n); 4147 offset -= s->red_left_pad; 4148 } 4149 4150 /* Allow address range falling entirely within usercopy region. */ 4151 if (offset >= s->useroffset && 4152 offset - s->useroffset <= s->usersize && 4153 n <= s->useroffset - offset + s->usersize) 4154 return; 4155 4156 /* 4157 * If the copy is still within the allocated object, produce 4158 * a warning instead of rejecting the copy. This is intended 4159 * to be a temporary method to find any missing usercopy 4160 * whitelists. 4161 */ 4162 object_size = slab_ksize(s); 4163 if (usercopy_fallback && 4164 offset <= object_size && n <= object_size - offset) { 4165 usercopy_warn("SLUB object", s->name, to_user, offset, n); 4166 return; 4167 } 4168 4169 usercopy_abort("SLUB object", s->name, to_user, offset, n); 4170 } 4171 #endif /* CONFIG_HARDENED_USERCOPY */ 4172 4173 size_t __ksize(const void *object) 4174 { 4175 struct page *page; 4176 4177 if (unlikely(object == ZERO_SIZE_PTR)) 4178 return 0; 4179 4180 page = virt_to_head_page(object); 4181 4182 if (unlikely(!PageSlab(page))) { 4183 WARN_ON(!PageCompound(page)); 4184 return page_size(page); 4185 } 4186 4187 return slab_ksize(page->slab_cache); 4188 } 4189 EXPORT_SYMBOL(__ksize); 4190 4191 void kfree(const void *x) 4192 { 4193 struct page *page; 4194 void *object = (void *)x; 4195 4196 trace_kfree(_RET_IP_, x); 4197 4198 if (unlikely(ZERO_OR_NULL_PTR(x))) 4199 return; 4200 4201 page = virt_to_head_page(x); 4202 if (unlikely(!PageSlab(page))) { 4203 unsigned int order = compound_order(page); 4204 4205 BUG_ON(!PageCompound(page)); 4206 kfree_hook(object); 4207 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, 4208 -(PAGE_SIZE << order)); 4209 __free_pages(page, order); 4210 return; 4211 } 4212 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_); 4213 } 4214 EXPORT_SYMBOL(kfree); 4215 4216 #define SHRINK_PROMOTE_MAX 32 4217 4218 /* 4219 * kmem_cache_shrink discards empty slabs and promotes the slabs filled 4220 * up most to the head of the partial lists. New allocations will then 4221 * fill those up and thus they can be removed from the partial lists. 4222 * 4223 * The slabs with the least items are placed last. This results in them 4224 * being allocated from last increasing the chance that the last objects 4225 * are freed in them. 4226 */ 4227 int __kmem_cache_shrink(struct kmem_cache *s) 4228 { 4229 int node; 4230 int i; 4231 struct kmem_cache_node *n; 4232 struct page *page; 4233 struct page *t; 4234 struct list_head discard; 4235 struct list_head promote[SHRINK_PROMOTE_MAX]; 4236 unsigned long flags; 4237 int ret = 0; 4238 4239 flush_all(s); 4240 for_each_kmem_cache_node(s, node, n) { 4241 INIT_LIST_HEAD(&discard); 4242 for (i = 0; i < SHRINK_PROMOTE_MAX; i++) 4243 INIT_LIST_HEAD(promote + i); 4244 4245 spin_lock_irqsave(&n->list_lock, flags); 4246 4247 /* 4248 * Build lists of slabs to discard or promote. 4249 * 4250 * Note that concurrent frees may occur while we hold the 4251 * list_lock. page->inuse here is the upper limit. 4252 */ 4253 list_for_each_entry_safe(page, t, &n->partial, slab_list) { 4254 int free = page->objects - page->inuse; 4255 4256 /* Do not reread page->inuse */ 4257 barrier(); 4258 4259 /* We do not keep full slabs on the list */ 4260 BUG_ON(free <= 0); 4261 4262 if (free == page->objects) { 4263 list_move(&page->slab_list, &discard); 4264 n->nr_partial--; 4265 } else if (free <= SHRINK_PROMOTE_MAX) 4266 list_move(&page->slab_list, promote + free - 1); 4267 } 4268 4269 /* 4270 * Promote the slabs filled up most to the head of the 4271 * partial list. 4272 */ 4273 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) 4274 list_splice(promote + i, &n->partial); 4275 4276 spin_unlock_irqrestore(&n->list_lock, flags); 4277 4278 /* Release empty slabs */ 4279 list_for_each_entry_safe(page, t, &discard, slab_list) 4280 discard_slab(s, page); 4281 4282 if (slabs_node(s, node)) 4283 ret = 1; 4284 } 4285 4286 return ret; 4287 } 4288 4289 static int slab_mem_going_offline_callback(void *arg) 4290 { 4291 struct kmem_cache *s; 4292 4293 mutex_lock(&slab_mutex); 4294 list_for_each_entry(s, &slab_caches, list) 4295 __kmem_cache_shrink(s); 4296 mutex_unlock(&slab_mutex); 4297 4298 return 0; 4299 } 4300 4301 static void slab_mem_offline_callback(void *arg) 4302 { 4303 struct memory_notify *marg = arg; 4304 int offline_node; 4305 4306 offline_node = marg->status_change_nid_normal; 4307 4308 /* 4309 * If the node still has available memory. we need kmem_cache_node 4310 * for it yet. 4311 */ 4312 if (offline_node < 0) 4313 return; 4314 4315 mutex_lock(&slab_mutex); 4316 node_clear(offline_node, slab_nodes); 4317 /* 4318 * We no longer free kmem_cache_node structures here, as it would be 4319 * racy with all get_node() users, and infeasible to protect them with 4320 * slab_mutex. 4321 */ 4322 mutex_unlock(&slab_mutex); 4323 } 4324 4325 static int slab_mem_going_online_callback(void *arg) 4326 { 4327 struct kmem_cache_node *n; 4328 struct kmem_cache *s; 4329 struct memory_notify *marg = arg; 4330 int nid = marg->status_change_nid_normal; 4331 int ret = 0; 4332 4333 /* 4334 * If the node's memory is already available, then kmem_cache_node is 4335 * already created. Nothing to do. 4336 */ 4337 if (nid < 0) 4338 return 0; 4339 4340 /* 4341 * We are bringing a node online. No memory is available yet. We must 4342 * allocate a kmem_cache_node structure in order to bring the node 4343 * online. 4344 */ 4345 mutex_lock(&slab_mutex); 4346 list_for_each_entry(s, &slab_caches, list) { 4347 /* 4348 * The structure may already exist if the node was previously 4349 * onlined and offlined. 4350 */ 4351 if (get_node(s, nid)) 4352 continue; 4353 /* 4354 * XXX: kmem_cache_alloc_node will fallback to other nodes 4355 * since memory is not yet available from the node that 4356 * is brought up. 4357 */ 4358 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 4359 if (!n) { 4360 ret = -ENOMEM; 4361 goto out; 4362 } 4363 init_kmem_cache_node(n); 4364 s->node[nid] = n; 4365 } 4366 /* 4367 * Any cache created after this point will also have kmem_cache_node 4368 * initialized for the new node. 4369 */ 4370 node_set(nid, slab_nodes); 4371 out: 4372 mutex_unlock(&slab_mutex); 4373 return ret; 4374 } 4375 4376 static int slab_memory_callback(struct notifier_block *self, 4377 unsigned long action, void *arg) 4378 { 4379 int ret = 0; 4380 4381 switch (action) { 4382 case MEM_GOING_ONLINE: 4383 ret = slab_mem_going_online_callback(arg); 4384 break; 4385 case MEM_GOING_OFFLINE: 4386 ret = slab_mem_going_offline_callback(arg); 4387 break; 4388 case MEM_OFFLINE: 4389 case MEM_CANCEL_ONLINE: 4390 slab_mem_offline_callback(arg); 4391 break; 4392 case MEM_ONLINE: 4393 case MEM_CANCEL_OFFLINE: 4394 break; 4395 } 4396 if (ret) 4397 ret = notifier_from_errno(ret); 4398 else 4399 ret = NOTIFY_OK; 4400 return ret; 4401 } 4402 4403 static struct notifier_block slab_memory_callback_nb = { 4404 .notifier_call = slab_memory_callback, 4405 .priority = SLAB_CALLBACK_PRI, 4406 }; 4407 4408 /******************************************************************** 4409 * Basic setup of slabs 4410 *******************************************************************/ 4411 4412 /* 4413 * Used for early kmem_cache structures that were allocated using 4414 * the page allocator. Allocate them properly then fix up the pointers 4415 * that may be pointing to the wrong kmem_cache structure. 4416 */ 4417 4418 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) 4419 { 4420 int node; 4421 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); 4422 struct kmem_cache_node *n; 4423 4424 memcpy(s, static_cache, kmem_cache->object_size); 4425 4426 /* 4427 * This runs very early, and only the boot processor is supposed to be 4428 * up. Even if it weren't true, IRQs are not up so we couldn't fire 4429 * IPIs around. 4430 */ 4431 __flush_cpu_slab(s, smp_processor_id()); 4432 for_each_kmem_cache_node(s, node, n) { 4433 struct page *p; 4434 4435 list_for_each_entry(p, &n->partial, slab_list) 4436 p->slab_cache = s; 4437 4438 #ifdef CONFIG_SLUB_DEBUG 4439 list_for_each_entry(p, &n->full, slab_list) 4440 p->slab_cache = s; 4441 #endif 4442 } 4443 list_add(&s->list, &slab_caches); 4444 return s; 4445 } 4446 4447 void __init kmem_cache_init(void) 4448 { 4449 static __initdata struct kmem_cache boot_kmem_cache, 4450 boot_kmem_cache_node; 4451 int node; 4452 4453 if (debug_guardpage_minorder()) 4454 slub_max_order = 0; 4455 4456 kmem_cache_node = &boot_kmem_cache_node; 4457 kmem_cache = &boot_kmem_cache; 4458 4459 /* 4460 * Initialize the nodemask for which we will allocate per node 4461 * structures. Here we don't need taking slab_mutex yet. 4462 */ 4463 for_each_node_state(node, N_NORMAL_MEMORY) 4464 node_set(node, slab_nodes); 4465 4466 create_boot_cache(kmem_cache_node, "kmem_cache_node", 4467 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0); 4468 4469 register_hotmemory_notifier(&slab_memory_callback_nb); 4470 4471 /* Able to allocate the per node structures */ 4472 slab_state = PARTIAL; 4473 4474 create_boot_cache(kmem_cache, "kmem_cache", 4475 offsetof(struct kmem_cache, node) + 4476 nr_node_ids * sizeof(struct kmem_cache_node *), 4477 SLAB_HWCACHE_ALIGN, 0, 0); 4478 4479 kmem_cache = bootstrap(&boot_kmem_cache); 4480 kmem_cache_node = bootstrap(&boot_kmem_cache_node); 4481 4482 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 4483 setup_kmalloc_cache_index_table(); 4484 create_kmalloc_caches(0); 4485 4486 /* Setup random freelists for each cache */ 4487 init_freelist_randomization(); 4488 4489 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL, 4490 slub_cpu_dead); 4491 4492 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n", 4493 cache_line_size(), 4494 slub_min_order, slub_max_order, slub_min_objects, 4495 nr_cpu_ids, nr_node_ids); 4496 } 4497 4498 void __init kmem_cache_init_late(void) 4499 { 4500 } 4501 4502 struct kmem_cache * 4503 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, 4504 slab_flags_t flags, void (*ctor)(void *)) 4505 { 4506 struct kmem_cache *s; 4507 4508 s = find_mergeable(size, align, flags, name, ctor); 4509 if (s) { 4510 s->refcount++; 4511 4512 /* 4513 * Adjust the object sizes so that we clear 4514 * the complete object on kzalloc. 4515 */ 4516 s->object_size = max(s->object_size, size); 4517 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *))); 4518 4519 if (sysfs_slab_alias(s, name)) { 4520 s->refcount--; 4521 s = NULL; 4522 } 4523 } 4524 4525 return s; 4526 } 4527 4528 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) 4529 { 4530 int err; 4531 4532 err = kmem_cache_open(s, flags); 4533 if (err) 4534 return err; 4535 4536 /* Mutex is not taken during early boot */ 4537 if (slab_state <= UP) 4538 return 0; 4539 4540 err = sysfs_slab_add(s); 4541 if (err) 4542 __kmem_cache_release(s); 4543 4544 return err; 4545 } 4546 4547 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 4548 { 4549 struct kmem_cache *s; 4550 void *ret; 4551 4552 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 4553 return kmalloc_large(size, gfpflags); 4554 4555 s = kmalloc_slab(size, gfpflags); 4556 4557 if (unlikely(ZERO_OR_NULL_PTR(s))) 4558 return s; 4559 4560 ret = slab_alloc(s, gfpflags, caller, size); 4561 4562 /* Honor the call site pointer we received. */ 4563 trace_kmalloc(caller, ret, size, s->size, gfpflags); 4564 4565 return ret; 4566 } 4567 EXPORT_SYMBOL(__kmalloc_track_caller); 4568 4569 #ifdef CONFIG_NUMA 4570 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 4571 int node, unsigned long caller) 4572 { 4573 struct kmem_cache *s; 4574 void *ret; 4575 4576 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 4577 ret = kmalloc_large_node(size, gfpflags, node); 4578 4579 trace_kmalloc_node(caller, ret, 4580 size, PAGE_SIZE << get_order(size), 4581 gfpflags, node); 4582 4583 return ret; 4584 } 4585 4586 s = kmalloc_slab(size, gfpflags); 4587 4588 if (unlikely(ZERO_OR_NULL_PTR(s))) 4589 return s; 4590 4591 ret = slab_alloc_node(s, gfpflags, node, caller, size); 4592 4593 /* Honor the call site pointer we received. */ 4594 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 4595 4596 return ret; 4597 } 4598 EXPORT_SYMBOL(__kmalloc_node_track_caller); 4599 #endif 4600 4601 #ifdef CONFIG_SYSFS 4602 static int count_inuse(struct page *page) 4603 { 4604 return page->inuse; 4605 } 4606 4607 static int count_total(struct page *page) 4608 { 4609 return page->objects; 4610 } 4611 #endif 4612 4613 #ifdef CONFIG_SLUB_DEBUG 4614 static void validate_slab(struct kmem_cache *s, struct page *page) 4615 { 4616 void *p; 4617 void *addr = page_address(page); 4618 unsigned long *map; 4619 4620 slab_lock(page); 4621 4622 if (!check_slab(s, page) || !on_freelist(s, page, NULL)) 4623 goto unlock; 4624 4625 /* Now we know that a valid freelist exists */ 4626 map = get_map(s, page); 4627 for_each_object(p, s, addr, page->objects) { 4628 u8 val = test_bit(__obj_to_index(s, addr, p), map) ? 4629 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE; 4630 4631 if (!check_object(s, page, p, val)) 4632 break; 4633 } 4634 put_map(map); 4635 unlock: 4636 slab_unlock(page); 4637 } 4638 4639 static int validate_slab_node(struct kmem_cache *s, 4640 struct kmem_cache_node *n) 4641 { 4642 unsigned long count = 0; 4643 struct page *page; 4644 unsigned long flags; 4645 4646 spin_lock_irqsave(&n->list_lock, flags); 4647 4648 list_for_each_entry(page, &n->partial, slab_list) { 4649 validate_slab(s, page); 4650 count++; 4651 } 4652 if (count != n->nr_partial) 4653 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n", 4654 s->name, count, n->nr_partial); 4655 4656 if (!(s->flags & SLAB_STORE_USER)) 4657 goto out; 4658 4659 list_for_each_entry(page, &n->full, slab_list) { 4660 validate_slab(s, page); 4661 count++; 4662 } 4663 if (count != atomic_long_read(&n->nr_slabs)) 4664 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n", 4665 s->name, count, atomic_long_read(&n->nr_slabs)); 4666 4667 out: 4668 spin_unlock_irqrestore(&n->list_lock, flags); 4669 return count; 4670 } 4671 4672 static long validate_slab_cache(struct kmem_cache *s) 4673 { 4674 int node; 4675 unsigned long count = 0; 4676 struct kmem_cache_node *n; 4677 4678 flush_all(s); 4679 for_each_kmem_cache_node(s, node, n) 4680 count += validate_slab_node(s, n); 4681 4682 return count; 4683 } 4684 /* 4685 * Generate lists of code addresses where slabcache objects are allocated 4686 * and freed. 4687 */ 4688 4689 struct location { 4690 unsigned long count; 4691 unsigned long addr; 4692 long long sum_time; 4693 long min_time; 4694 long max_time; 4695 long min_pid; 4696 long max_pid; 4697 DECLARE_BITMAP(cpus, NR_CPUS); 4698 nodemask_t nodes; 4699 }; 4700 4701 struct loc_track { 4702 unsigned long max; 4703 unsigned long count; 4704 struct location *loc; 4705 }; 4706 4707 static void free_loc_track(struct loc_track *t) 4708 { 4709 if (t->max) 4710 free_pages((unsigned long)t->loc, 4711 get_order(sizeof(struct location) * t->max)); 4712 } 4713 4714 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 4715 { 4716 struct location *l; 4717 int order; 4718 4719 order = get_order(sizeof(struct location) * max); 4720 4721 l = (void *)__get_free_pages(flags, order); 4722 if (!l) 4723 return 0; 4724 4725 if (t->count) { 4726 memcpy(l, t->loc, sizeof(struct location) * t->count); 4727 free_loc_track(t); 4728 } 4729 t->max = max; 4730 t->loc = l; 4731 return 1; 4732 } 4733 4734 static int add_location(struct loc_track *t, struct kmem_cache *s, 4735 const struct track *track) 4736 { 4737 long start, end, pos; 4738 struct location *l; 4739 unsigned long caddr; 4740 unsigned long age = jiffies - track->when; 4741 4742 start = -1; 4743 end = t->count; 4744 4745 for ( ; ; ) { 4746 pos = start + (end - start + 1) / 2; 4747 4748 /* 4749 * There is nothing at "end". If we end up there 4750 * we need to add something to before end. 4751 */ 4752 if (pos == end) 4753 break; 4754 4755 caddr = t->loc[pos].addr; 4756 if (track->addr == caddr) { 4757 4758 l = &t->loc[pos]; 4759 l->count++; 4760 if (track->when) { 4761 l->sum_time += age; 4762 if (age < l->min_time) 4763 l->min_time = age; 4764 if (age > l->max_time) 4765 l->max_time = age; 4766 4767 if (track->pid < l->min_pid) 4768 l->min_pid = track->pid; 4769 if (track->pid > l->max_pid) 4770 l->max_pid = track->pid; 4771 4772 cpumask_set_cpu(track->cpu, 4773 to_cpumask(l->cpus)); 4774 } 4775 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4776 return 1; 4777 } 4778 4779 if (track->addr < caddr) 4780 end = pos; 4781 else 4782 start = pos; 4783 } 4784 4785 /* 4786 * Not found. Insert new tracking element. 4787 */ 4788 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 4789 return 0; 4790 4791 l = t->loc + pos; 4792 if (pos < t->count) 4793 memmove(l + 1, l, 4794 (t->count - pos) * sizeof(struct location)); 4795 t->count++; 4796 l->count = 1; 4797 l->addr = track->addr; 4798 l->sum_time = age; 4799 l->min_time = age; 4800 l->max_time = age; 4801 l->min_pid = track->pid; 4802 l->max_pid = track->pid; 4803 cpumask_clear(to_cpumask(l->cpus)); 4804 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 4805 nodes_clear(l->nodes); 4806 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4807 return 1; 4808 } 4809 4810 static void process_slab(struct loc_track *t, struct kmem_cache *s, 4811 struct page *page, enum track_item alloc) 4812 { 4813 void *addr = page_address(page); 4814 void *p; 4815 unsigned long *map; 4816 4817 map = get_map(s, page); 4818 for_each_object(p, s, addr, page->objects) 4819 if (!test_bit(__obj_to_index(s, addr, p), map)) 4820 add_location(t, s, get_track(s, p, alloc)); 4821 put_map(map); 4822 } 4823 4824 static int list_locations(struct kmem_cache *s, char *buf, 4825 enum track_item alloc) 4826 { 4827 int len = 0; 4828 unsigned long i; 4829 struct loc_track t = { 0, 0, NULL }; 4830 int node; 4831 struct kmem_cache_node *n; 4832 4833 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 4834 GFP_KERNEL)) { 4835 return sysfs_emit(buf, "Out of memory\n"); 4836 } 4837 /* Push back cpu slabs */ 4838 flush_all(s); 4839 4840 for_each_kmem_cache_node(s, node, n) { 4841 unsigned long flags; 4842 struct page *page; 4843 4844 if (!atomic_long_read(&n->nr_slabs)) 4845 continue; 4846 4847 spin_lock_irqsave(&n->list_lock, flags); 4848 list_for_each_entry(page, &n->partial, slab_list) 4849 process_slab(&t, s, page, alloc); 4850 list_for_each_entry(page, &n->full, slab_list) 4851 process_slab(&t, s, page, alloc); 4852 spin_unlock_irqrestore(&n->list_lock, flags); 4853 } 4854 4855 for (i = 0; i < t.count; i++) { 4856 struct location *l = &t.loc[i]; 4857 4858 len += sysfs_emit_at(buf, len, "%7ld ", l->count); 4859 4860 if (l->addr) 4861 len += sysfs_emit_at(buf, len, "%pS", (void *)l->addr); 4862 else 4863 len += sysfs_emit_at(buf, len, "<not-available>"); 4864 4865 if (l->sum_time != l->min_time) 4866 len += sysfs_emit_at(buf, len, " age=%ld/%ld/%ld", 4867 l->min_time, 4868 (long)div_u64(l->sum_time, 4869 l->count), 4870 l->max_time); 4871 else 4872 len += sysfs_emit_at(buf, len, " age=%ld", l->min_time); 4873 4874 if (l->min_pid != l->max_pid) 4875 len += sysfs_emit_at(buf, len, " pid=%ld-%ld", 4876 l->min_pid, l->max_pid); 4877 else 4878 len += sysfs_emit_at(buf, len, " pid=%ld", 4879 l->min_pid); 4880 4881 if (num_online_cpus() > 1 && 4882 !cpumask_empty(to_cpumask(l->cpus))) 4883 len += sysfs_emit_at(buf, len, " cpus=%*pbl", 4884 cpumask_pr_args(to_cpumask(l->cpus))); 4885 4886 if (nr_online_nodes > 1 && !nodes_empty(l->nodes)) 4887 len += sysfs_emit_at(buf, len, " nodes=%*pbl", 4888 nodemask_pr_args(&l->nodes)); 4889 4890 len += sysfs_emit_at(buf, len, "\n"); 4891 } 4892 4893 free_loc_track(&t); 4894 if (!t.count) 4895 len += sysfs_emit_at(buf, len, "No data\n"); 4896 4897 return len; 4898 } 4899 #endif /* CONFIG_SLUB_DEBUG */ 4900 4901 #ifdef SLUB_RESILIENCY_TEST 4902 static void __init resiliency_test(void) 4903 { 4904 u8 *p; 4905 int type = KMALLOC_NORMAL; 4906 4907 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10); 4908 4909 pr_err("SLUB resiliency testing\n"); 4910 pr_err("-----------------------\n"); 4911 pr_err("A. Corruption after allocation\n"); 4912 4913 p = kzalloc(16, GFP_KERNEL); 4914 p[16] = 0x12; 4915 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n", 4916 p + 16); 4917 4918 validate_slab_cache(kmalloc_caches[type][4]); 4919 4920 /* Hmmm... The next two are dangerous */ 4921 p = kzalloc(32, GFP_KERNEL); 4922 p[32 + sizeof(void *)] = 0x34; 4923 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n", 4924 p); 4925 pr_err("If allocated object is overwritten then not detectable\n\n"); 4926 4927 validate_slab_cache(kmalloc_caches[type][5]); 4928 p = kzalloc(64, GFP_KERNEL); 4929 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 4930 *p = 0x56; 4931 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 4932 p); 4933 pr_err("If allocated object is overwritten then not detectable\n\n"); 4934 validate_slab_cache(kmalloc_caches[type][6]); 4935 4936 pr_err("\nB. Corruption after free\n"); 4937 p = kzalloc(128, GFP_KERNEL); 4938 kfree(p); 4939 *p = 0x78; 4940 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 4941 validate_slab_cache(kmalloc_caches[type][7]); 4942 4943 p = kzalloc(256, GFP_KERNEL); 4944 kfree(p); 4945 p[50] = 0x9a; 4946 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); 4947 validate_slab_cache(kmalloc_caches[type][8]); 4948 4949 p = kzalloc(512, GFP_KERNEL); 4950 kfree(p); 4951 p[512] = 0xab; 4952 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 4953 validate_slab_cache(kmalloc_caches[type][9]); 4954 } 4955 #else 4956 #ifdef CONFIG_SYSFS 4957 static void resiliency_test(void) {}; 4958 #endif 4959 #endif /* SLUB_RESILIENCY_TEST */ 4960 4961 #ifdef CONFIG_SYSFS 4962 enum slab_stat_type { 4963 SL_ALL, /* All slabs */ 4964 SL_PARTIAL, /* Only partially allocated slabs */ 4965 SL_CPU, /* Only slabs used for cpu caches */ 4966 SL_OBJECTS, /* Determine allocated objects not slabs */ 4967 SL_TOTAL /* Determine object capacity not slabs */ 4968 }; 4969 4970 #define SO_ALL (1 << SL_ALL) 4971 #define SO_PARTIAL (1 << SL_PARTIAL) 4972 #define SO_CPU (1 << SL_CPU) 4973 #define SO_OBJECTS (1 << SL_OBJECTS) 4974 #define SO_TOTAL (1 << SL_TOTAL) 4975 4976 static ssize_t show_slab_objects(struct kmem_cache *s, 4977 char *buf, unsigned long flags) 4978 { 4979 unsigned long total = 0; 4980 int node; 4981 int x; 4982 unsigned long *nodes; 4983 int len = 0; 4984 4985 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL); 4986 if (!nodes) 4987 return -ENOMEM; 4988 4989 if (flags & SO_CPU) { 4990 int cpu; 4991 4992 for_each_possible_cpu(cpu) { 4993 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, 4994 cpu); 4995 int node; 4996 struct page *page; 4997 4998 page = READ_ONCE(c->page); 4999 if (!page) 5000 continue; 5001 5002 node = page_to_nid(page); 5003 if (flags & SO_TOTAL) 5004 x = page->objects; 5005 else if (flags & SO_OBJECTS) 5006 x = page->inuse; 5007 else 5008 x = 1; 5009 5010 total += x; 5011 nodes[node] += x; 5012 5013 page = slub_percpu_partial_read_once(c); 5014 if (page) { 5015 node = page_to_nid(page); 5016 if (flags & SO_TOTAL) 5017 WARN_ON_ONCE(1); 5018 else if (flags & SO_OBJECTS) 5019 WARN_ON_ONCE(1); 5020 else 5021 x = page->pages; 5022 total += x; 5023 nodes[node] += x; 5024 } 5025 } 5026 } 5027 5028 /* 5029 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex" 5030 * already held which will conflict with an existing lock order: 5031 * 5032 * mem_hotplug_lock->slab_mutex->kernfs_mutex 5033 * 5034 * We don't really need mem_hotplug_lock (to hold off 5035 * slab_mem_going_offline_callback) here because slab's memory hot 5036 * unplug code doesn't destroy the kmem_cache->node[] data. 5037 */ 5038 5039 #ifdef CONFIG_SLUB_DEBUG 5040 if (flags & SO_ALL) { 5041 struct kmem_cache_node *n; 5042 5043 for_each_kmem_cache_node(s, node, n) { 5044 5045 if (flags & SO_TOTAL) 5046 x = atomic_long_read(&n->total_objects); 5047 else if (flags & SO_OBJECTS) 5048 x = atomic_long_read(&n->total_objects) - 5049 count_partial(n, count_free); 5050 else 5051 x = atomic_long_read(&n->nr_slabs); 5052 total += x; 5053 nodes[node] += x; 5054 } 5055 5056 } else 5057 #endif 5058 if (flags & SO_PARTIAL) { 5059 struct kmem_cache_node *n; 5060 5061 for_each_kmem_cache_node(s, node, n) { 5062 if (flags & SO_TOTAL) 5063 x = count_partial(n, count_total); 5064 else if (flags & SO_OBJECTS) 5065 x = count_partial(n, count_inuse); 5066 else 5067 x = n->nr_partial; 5068 total += x; 5069 nodes[node] += x; 5070 } 5071 } 5072 5073 len += sysfs_emit_at(buf, len, "%lu", total); 5074 #ifdef CONFIG_NUMA 5075 for (node = 0; node < nr_node_ids; node++) { 5076 if (nodes[node]) 5077 len += sysfs_emit_at(buf, len, " N%d=%lu", 5078 node, nodes[node]); 5079 } 5080 #endif 5081 len += sysfs_emit_at(buf, len, "\n"); 5082 kfree(nodes); 5083 5084 return len; 5085 } 5086 5087 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 5088 #define to_slab(n) container_of(n, struct kmem_cache, kobj) 5089 5090 struct slab_attribute { 5091 struct attribute attr; 5092 ssize_t (*show)(struct kmem_cache *s, char *buf); 5093 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 5094 }; 5095 5096 #define SLAB_ATTR_RO(_name) \ 5097 static struct slab_attribute _name##_attr = \ 5098 __ATTR(_name, 0400, _name##_show, NULL) 5099 5100 #define SLAB_ATTR(_name) \ 5101 static struct slab_attribute _name##_attr = \ 5102 __ATTR(_name, 0600, _name##_show, _name##_store) 5103 5104 static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 5105 { 5106 return sysfs_emit(buf, "%u\n", s->size); 5107 } 5108 SLAB_ATTR_RO(slab_size); 5109 5110 static ssize_t align_show(struct kmem_cache *s, char *buf) 5111 { 5112 return sysfs_emit(buf, "%u\n", s->align); 5113 } 5114 SLAB_ATTR_RO(align); 5115 5116 static ssize_t object_size_show(struct kmem_cache *s, char *buf) 5117 { 5118 return sysfs_emit(buf, "%u\n", s->object_size); 5119 } 5120 SLAB_ATTR_RO(object_size); 5121 5122 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 5123 { 5124 return sysfs_emit(buf, "%u\n", oo_objects(s->oo)); 5125 } 5126 SLAB_ATTR_RO(objs_per_slab); 5127 5128 static ssize_t order_show(struct kmem_cache *s, char *buf) 5129 { 5130 return sysfs_emit(buf, "%u\n", oo_order(s->oo)); 5131 } 5132 SLAB_ATTR_RO(order); 5133 5134 static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 5135 { 5136 return sysfs_emit(buf, "%lu\n", s->min_partial); 5137 } 5138 5139 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 5140 size_t length) 5141 { 5142 unsigned long min; 5143 int err; 5144 5145 err = kstrtoul(buf, 10, &min); 5146 if (err) 5147 return err; 5148 5149 set_min_partial(s, min); 5150 return length; 5151 } 5152 SLAB_ATTR(min_partial); 5153 5154 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) 5155 { 5156 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s)); 5157 } 5158 5159 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, 5160 size_t length) 5161 { 5162 unsigned int objects; 5163 int err; 5164 5165 err = kstrtouint(buf, 10, &objects); 5166 if (err) 5167 return err; 5168 if (objects && !kmem_cache_has_cpu_partial(s)) 5169 return -EINVAL; 5170 5171 slub_set_cpu_partial(s, objects); 5172 flush_all(s); 5173 return length; 5174 } 5175 SLAB_ATTR(cpu_partial); 5176 5177 static ssize_t ctor_show(struct kmem_cache *s, char *buf) 5178 { 5179 if (!s->ctor) 5180 return 0; 5181 return sysfs_emit(buf, "%pS\n", s->ctor); 5182 } 5183 SLAB_ATTR_RO(ctor); 5184 5185 static ssize_t aliases_show(struct kmem_cache *s, char *buf) 5186 { 5187 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1); 5188 } 5189 SLAB_ATTR_RO(aliases); 5190 5191 static ssize_t partial_show(struct kmem_cache *s, char *buf) 5192 { 5193 return show_slab_objects(s, buf, SO_PARTIAL); 5194 } 5195 SLAB_ATTR_RO(partial); 5196 5197 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 5198 { 5199 return show_slab_objects(s, buf, SO_CPU); 5200 } 5201 SLAB_ATTR_RO(cpu_slabs); 5202 5203 static ssize_t objects_show(struct kmem_cache *s, char *buf) 5204 { 5205 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 5206 } 5207 SLAB_ATTR_RO(objects); 5208 5209 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 5210 { 5211 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 5212 } 5213 SLAB_ATTR_RO(objects_partial); 5214 5215 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) 5216 { 5217 int objects = 0; 5218 int pages = 0; 5219 int cpu; 5220 int len = 0; 5221 5222 for_each_online_cpu(cpu) { 5223 struct page *page; 5224 5225 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); 5226 5227 if (page) { 5228 pages += page->pages; 5229 objects += page->pobjects; 5230 } 5231 } 5232 5233 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages); 5234 5235 #ifdef CONFIG_SMP 5236 for_each_online_cpu(cpu) { 5237 struct page *page; 5238 5239 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); 5240 if (page) 5241 len += sysfs_emit_at(buf, len, " C%d=%d(%d)", 5242 cpu, page->pobjects, page->pages); 5243 } 5244 #endif 5245 len += sysfs_emit_at(buf, len, "\n"); 5246 5247 return len; 5248 } 5249 SLAB_ATTR_RO(slabs_cpu_partial); 5250 5251 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 5252 { 5253 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 5254 } 5255 SLAB_ATTR_RO(reclaim_account); 5256 5257 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 5258 { 5259 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 5260 } 5261 SLAB_ATTR_RO(hwcache_align); 5262 5263 #ifdef CONFIG_ZONE_DMA 5264 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 5265 { 5266 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 5267 } 5268 SLAB_ATTR_RO(cache_dma); 5269 #endif 5270 5271 static ssize_t usersize_show(struct kmem_cache *s, char *buf) 5272 { 5273 return sysfs_emit(buf, "%u\n", s->usersize); 5274 } 5275 SLAB_ATTR_RO(usersize); 5276 5277 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 5278 { 5279 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU)); 5280 } 5281 SLAB_ATTR_RO(destroy_by_rcu); 5282 5283 #ifdef CONFIG_SLUB_DEBUG 5284 static ssize_t slabs_show(struct kmem_cache *s, char *buf) 5285 { 5286 return show_slab_objects(s, buf, SO_ALL); 5287 } 5288 SLAB_ATTR_RO(slabs); 5289 5290 static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 5291 { 5292 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 5293 } 5294 SLAB_ATTR_RO(total_objects); 5295 5296 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 5297 { 5298 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS)); 5299 } 5300 SLAB_ATTR_RO(sanity_checks); 5301 5302 static ssize_t trace_show(struct kmem_cache *s, char *buf) 5303 { 5304 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 5305 } 5306 SLAB_ATTR_RO(trace); 5307 5308 static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 5309 { 5310 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 5311 } 5312 5313 SLAB_ATTR_RO(red_zone); 5314 5315 static ssize_t poison_show(struct kmem_cache *s, char *buf) 5316 { 5317 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON)); 5318 } 5319 5320 SLAB_ATTR_RO(poison); 5321 5322 static ssize_t store_user_show(struct kmem_cache *s, char *buf) 5323 { 5324 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 5325 } 5326 5327 SLAB_ATTR_RO(store_user); 5328 5329 static ssize_t validate_show(struct kmem_cache *s, char *buf) 5330 { 5331 return 0; 5332 } 5333 5334 static ssize_t validate_store(struct kmem_cache *s, 5335 const char *buf, size_t length) 5336 { 5337 int ret = -EINVAL; 5338 5339 if (buf[0] == '1') { 5340 ret = validate_slab_cache(s); 5341 if (ret >= 0) 5342 ret = length; 5343 } 5344 return ret; 5345 } 5346 SLAB_ATTR(validate); 5347 5348 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 5349 { 5350 if (!(s->flags & SLAB_STORE_USER)) 5351 return -ENOSYS; 5352 return list_locations(s, buf, TRACK_ALLOC); 5353 } 5354 SLAB_ATTR_RO(alloc_calls); 5355 5356 static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 5357 { 5358 if (!(s->flags & SLAB_STORE_USER)) 5359 return -ENOSYS; 5360 return list_locations(s, buf, TRACK_FREE); 5361 } 5362 SLAB_ATTR_RO(free_calls); 5363 #endif /* CONFIG_SLUB_DEBUG */ 5364 5365 #ifdef CONFIG_FAILSLAB 5366 static ssize_t failslab_show(struct kmem_cache *s, char *buf) 5367 { 5368 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 5369 } 5370 SLAB_ATTR_RO(failslab); 5371 #endif 5372 5373 static ssize_t shrink_show(struct kmem_cache *s, char *buf) 5374 { 5375 return 0; 5376 } 5377 5378 static ssize_t shrink_store(struct kmem_cache *s, 5379 const char *buf, size_t length) 5380 { 5381 if (buf[0] == '1') 5382 kmem_cache_shrink(s); 5383 else 5384 return -EINVAL; 5385 return length; 5386 } 5387 SLAB_ATTR(shrink); 5388 5389 #ifdef CONFIG_NUMA 5390 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 5391 { 5392 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10); 5393 } 5394 5395 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 5396 const char *buf, size_t length) 5397 { 5398 unsigned int ratio; 5399 int err; 5400 5401 err = kstrtouint(buf, 10, &ratio); 5402 if (err) 5403 return err; 5404 if (ratio > 100) 5405 return -ERANGE; 5406 5407 s->remote_node_defrag_ratio = ratio * 10; 5408 5409 return length; 5410 } 5411 SLAB_ATTR(remote_node_defrag_ratio); 5412 #endif 5413 5414 #ifdef CONFIG_SLUB_STATS 5415 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 5416 { 5417 unsigned long sum = 0; 5418 int cpu; 5419 int len = 0; 5420 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL); 5421 5422 if (!data) 5423 return -ENOMEM; 5424 5425 for_each_online_cpu(cpu) { 5426 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 5427 5428 data[cpu] = x; 5429 sum += x; 5430 } 5431 5432 len += sysfs_emit_at(buf, len, "%lu", sum); 5433 5434 #ifdef CONFIG_SMP 5435 for_each_online_cpu(cpu) { 5436 if (data[cpu]) 5437 len += sysfs_emit_at(buf, len, " C%d=%u", 5438 cpu, data[cpu]); 5439 } 5440 #endif 5441 kfree(data); 5442 len += sysfs_emit_at(buf, len, "\n"); 5443 5444 return len; 5445 } 5446 5447 static void clear_stat(struct kmem_cache *s, enum stat_item si) 5448 { 5449 int cpu; 5450 5451 for_each_online_cpu(cpu) 5452 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 5453 } 5454 5455 #define STAT_ATTR(si, text) \ 5456 static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 5457 { \ 5458 return show_stat(s, buf, si); \ 5459 } \ 5460 static ssize_t text##_store(struct kmem_cache *s, \ 5461 const char *buf, size_t length) \ 5462 { \ 5463 if (buf[0] != '0') \ 5464 return -EINVAL; \ 5465 clear_stat(s, si); \ 5466 return length; \ 5467 } \ 5468 SLAB_ATTR(text); \ 5469 5470 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 5471 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 5472 STAT_ATTR(FREE_FASTPATH, free_fastpath); 5473 STAT_ATTR(FREE_SLOWPATH, free_slowpath); 5474 STAT_ATTR(FREE_FROZEN, free_frozen); 5475 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 5476 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 5477 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 5478 STAT_ATTR(ALLOC_SLAB, alloc_slab); 5479 STAT_ATTR(ALLOC_REFILL, alloc_refill); 5480 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); 5481 STAT_ATTR(FREE_SLAB, free_slab); 5482 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 5483 STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 5484 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 5485 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 5486 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 5487 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 5488 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); 5489 STAT_ATTR(ORDER_FALLBACK, order_fallback); 5490 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); 5491 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); 5492 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); 5493 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); 5494 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); 5495 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); 5496 #endif /* CONFIG_SLUB_STATS */ 5497 5498 static struct attribute *slab_attrs[] = { 5499 &slab_size_attr.attr, 5500 &object_size_attr.attr, 5501 &objs_per_slab_attr.attr, 5502 &order_attr.attr, 5503 &min_partial_attr.attr, 5504 &cpu_partial_attr.attr, 5505 &objects_attr.attr, 5506 &objects_partial_attr.attr, 5507 &partial_attr.attr, 5508 &cpu_slabs_attr.attr, 5509 &ctor_attr.attr, 5510 &aliases_attr.attr, 5511 &align_attr.attr, 5512 &hwcache_align_attr.attr, 5513 &reclaim_account_attr.attr, 5514 &destroy_by_rcu_attr.attr, 5515 &shrink_attr.attr, 5516 &slabs_cpu_partial_attr.attr, 5517 #ifdef CONFIG_SLUB_DEBUG 5518 &total_objects_attr.attr, 5519 &slabs_attr.attr, 5520 &sanity_checks_attr.attr, 5521 &trace_attr.attr, 5522 &red_zone_attr.attr, 5523 &poison_attr.attr, 5524 &store_user_attr.attr, 5525 &validate_attr.attr, 5526 &alloc_calls_attr.attr, 5527 &free_calls_attr.attr, 5528 #endif 5529 #ifdef CONFIG_ZONE_DMA 5530 &cache_dma_attr.attr, 5531 #endif 5532 #ifdef CONFIG_NUMA 5533 &remote_node_defrag_ratio_attr.attr, 5534 #endif 5535 #ifdef CONFIG_SLUB_STATS 5536 &alloc_fastpath_attr.attr, 5537 &alloc_slowpath_attr.attr, 5538 &free_fastpath_attr.attr, 5539 &free_slowpath_attr.attr, 5540 &free_frozen_attr.attr, 5541 &free_add_partial_attr.attr, 5542 &free_remove_partial_attr.attr, 5543 &alloc_from_partial_attr.attr, 5544 &alloc_slab_attr.attr, 5545 &alloc_refill_attr.attr, 5546 &alloc_node_mismatch_attr.attr, 5547 &free_slab_attr.attr, 5548 &cpuslab_flush_attr.attr, 5549 &deactivate_full_attr.attr, 5550 &deactivate_empty_attr.attr, 5551 &deactivate_to_head_attr.attr, 5552 &deactivate_to_tail_attr.attr, 5553 &deactivate_remote_frees_attr.attr, 5554 &deactivate_bypass_attr.attr, 5555 &order_fallback_attr.attr, 5556 &cmpxchg_double_fail_attr.attr, 5557 &cmpxchg_double_cpu_fail_attr.attr, 5558 &cpu_partial_alloc_attr.attr, 5559 &cpu_partial_free_attr.attr, 5560 &cpu_partial_node_attr.attr, 5561 &cpu_partial_drain_attr.attr, 5562 #endif 5563 #ifdef CONFIG_FAILSLAB 5564 &failslab_attr.attr, 5565 #endif 5566 &usersize_attr.attr, 5567 5568 NULL 5569 }; 5570 5571 static const struct attribute_group slab_attr_group = { 5572 .attrs = slab_attrs, 5573 }; 5574 5575 static ssize_t slab_attr_show(struct kobject *kobj, 5576 struct attribute *attr, 5577 char *buf) 5578 { 5579 struct slab_attribute *attribute; 5580 struct kmem_cache *s; 5581 int err; 5582 5583 attribute = to_slab_attr(attr); 5584 s = to_slab(kobj); 5585 5586 if (!attribute->show) 5587 return -EIO; 5588 5589 err = attribute->show(s, buf); 5590 5591 return err; 5592 } 5593 5594 static ssize_t slab_attr_store(struct kobject *kobj, 5595 struct attribute *attr, 5596 const char *buf, size_t len) 5597 { 5598 struct slab_attribute *attribute; 5599 struct kmem_cache *s; 5600 int err; 5601 5602 attribute = to_slab_attr(attr); 5603 s = to_slab(kobj); 5604 5605 if (!attribute->store) 5606 return -EIO; 5607 5608 err = attribute->store(s, buf, len); 5609 return err; 5610 } 5611 5612 static void kmem_cache_release(struct kobject *k) 5613 { 5614 slab_kmem_cache_release(to_slab(k)); 5615 } 5616 5617 static const struct sysfs_ops slab_sysfs_ops = { 5618 .show = slab_attr_show, 5619 .store = slab_attr_store, 5620 }; 5621 5622 static struct kobj_type slab_ktype = { 5623 .sysfs_ops = &slab_sysfs_ops, 5624 .release = kmem_cache_release, 5625 }; 5626 5627 static struct kset *slab_kset; 5628 5629 static inline struct kset *cache_kset(struct kmem_cache *s) 5630 { 5631 return slab_kset; 5632 } 5633 5634 #define ID_STR_LENGTH 64 5635 5636 /* Create a unique string id for a slab cache: 5637 * 5638 * Format :[flags-]size 5639 */ 5640 static char *create_unique_id(struct kmem_cache *s) 5641 { 5642 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 5643 char *p = name; 5644 5645 BUG_ON(!name); 5646 5647 *p++ = ':'; 5648 /* 5649 * First flags affecting slabcache operations. We will only 5650 * get here for aliasable slabs so we do not need to support 5651 * too many flags. The flags here must cover all flags that 5652 * are matched during merging to guarantee that the id is 5653 * unique. 5654 */ 5655 if (s->flags & SLAB_CACHE_DMA) 5656 *p++ = 'd'; 5657 if (s->flags & SLAB_CACHE_DMA32) 5658 *p++ = 'D'; 5659 if (s->flags & SLAB_RECLAIM_ACCOUNT) 5660 *p++ = 'a'; 5661 if (s->flags & SLAB_CONSISTENCY_CHECKS) 5662 *p++ = 'F'; 5663 if (s->flags & SLAB_ACCOUNT) 5664 *p++ = 'A'; 5665 if (p != name + 1) 5666 *p++ = '-'; 5667 p += sprintf(p, "%07u", s->size); 5668 5669 BUG_ON(p > name + ID_STR_LENGTH - 1); 5670 return name; 5671 } 5672 5673 static int sysfs_slab_add(struct kmem_cache *s) 5674 { 5675 int err; 5676 const char *name; 5677 struct kset *kset = cache_kset(s); 5678 int unmergeable = slab_unmergeable(s); 5679 5680 if (!kset) { 5681 kobject_init(&s->kobj, &slab_ktype); 5682 return 0; 5683 } 5684 5685 if (!unmergeable && disable_higher_order_debug && 5686 (slub_debug & DEBUG_METADATA_FLAGS)) 5687 unmergeable = 1; 5688 5689 if (unmergeable) { 5690 /* 5691 * Slabcache can never be merged so we can use the name proper. 5692 * This is typically the case for debug situations. In that 5693 * case we can catch duplicate names easily. 5694 */ 5695 sysfs_remove_link(&slab_kset->kobj, s->name); 5696 name = s->name; 5697 } else { 5698 /* 5699 * Create a unique name for the slab as a target 5700 * for the symlinks. 5701 */ 5702 name = create_unique_id(s); 5703 } 5704 5705 s->kobj.kset = kset; 5706 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); 5707 if (err) 5708 goto out; 5709 5710 err = sysfs_create_group(&s->kobj, &slab_attr_group); 5711 if (err) 5712 goto out_del_kobj; 5713 5714 if (!unmergeable) { 5715 /* Setup first alias */ 5716 sysfs_slab_alias(s, s->name); 5717 } 5718 out: 5719 if (!unmergeable) 5720 kfree(name); 5721 return err; 5722 out_del_kobj: 5723 kobject_del(&s->kobj); 5724 goto out; 5725 } 5726 5727 void sysfs_slab_unlink(struct kmem_cache *s) 5728 { 5729 if (slab_state >= FULL) 5730 kobject_del(&s->kobj); 5731 } 5732 5733 void sysfs_slab_release(struct kmem_cache *s) 5734 { 5735 if (slab_state >= FULL) 5736 kobject_put(&s->kobj); 5737 } 5738 5739 /* 5740 * Need to buffer aliases during bootup until sysfs becomes 5741 * available lest we lose that information. 5742 */ 5743 struct saved_alias { 5744 struct kmem_cache *s; 5745 const char *name; 5746 struct saved_alias *next; 5747 }; 5748 5749 static struct saved_alias *alias_list; 5750 5751 static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 5752 { 5753 struct saved_alias *al; 5754 5755 if (slab_state == FULL) { 5756 /* 5757 * If we have a leftover link then remove it. 5758 */ 5759 sysfs_remove_link(&slab_kset->kobj, name); 5760 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 5761 } 5762 5763 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 5764 if (!al) 5765 return -ENOMEM; 5766 5767 al->s = s; 5768 al->name = name; 5769 al->next = alias_list; 5770 alias_list = al; 5771 return 0; 5772 } 5773 5774 static int __init slab_sysfs_init(void) 5775 { 5776 struct kmem_cache *s; 5777 int err; 5778 5779 mutex_lock(&slab_mutex); 5780 5781 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj); 5782 if (!slab_kset) { 5783 mutex_unlock(&slab_mutex); 5784 pr_err("Cannot register slab subsystem.\n"); 5785 return -ENOSYS; 5786 } 5787 5788 slab_state = FULL; 5789 5790 list_for_each_entry(s, &slab_caches, list) { 5791 err = sysfs_slab_add(s); 5792 if (err) 5793 pr_err("SLUB: Unable to add boot slab %s to sysfs\n", 5794 s->name); 5795 } 5796 5797 while (alias_list) { 5798 struct saved_alias *al = alias_list; 5799 5800 alias_list = alias_list->next; 5801 err = sysfs_slab_alias(al->s, al->name); 5802 if (err) 5803 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n", 5804 al->name); 5805 kfree(al); 5806 } 5807 5808 mutex_unlock(&slab_mutex); 5809 resiliency_test(); 5810 return 0; 5811 } 5812 5813 __initcall(slab_sysfs_init); 5814 #endif /* CONFIG_SYSFS */ 5815 5816 /* 5817 * The /proc/slabinfo ABI 5818 */ 5819 #ifdef CONFIG_SLUB_DEBUG 5820 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) 5821 { 5822 unsigned long nr_slabs = 0; 5823 unsigned long nr_objs = 0; 5824 unsigned long nr_free = 0; 5825 int node; 5826 struct kmem_cache_node *n; 5827 5828 for_each_kmem_cache_node(s, node, n) { 5829 nr_slabs += node_nr_slabs(n); 5830 nr_objs += node_nr_objs(n); 5831 nr_free += count_partial(n, count_free); 5832 } 5833 5834 sinfo->active_objs = nr_objs - nr_free; 5835 sinfo->num_objs = nr_objs; 5836 sinfo->active_slabs = nr_slabs; 5837 sinfo->num_slabs = nr_slabs; 5838 sinfo->objects_per_slab = oo_objects(s->oo); 5839 sinfo->cache_order = oo_order(s->oo); 5840 } 5841 5842 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) 5843 { 5844 } 5845 5846 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 5847 size_t count, loff_t *ppos) 5848 { 5849 return -EIO; 5850 } 5851 #endif /* CONFIG_SLUB_DEBUG */ 5852