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 #include <kunit/test.h> 40 41 #include <linux/debugfs.h> 42 #include <trace/events/kmem.h> 43 44 #include "internal.h" 45 46 /* 47 * Lock order: 48 * 1. slab_mutex (Global Mutex) 49 * 2. node->list_lock 50 * 3. slab_lock(page) (Only on some arches and for debugging) 51 * 52 * slab_mutex 53 * 54 * The role of the slab_mutex is to protect the list of all the slabs 55 * and to synchronize major metadata changes to slab cache structures. 56 * 57 * The slab_lock is only used for debugging and on arches that do not 58 * have the ability to do a cmpxchg_double. It only protects: 59 * A. page->freelist -> List of object free in a page 60 * B. page->inuse -> Number of objects in use 61 * C. page->objects -> Number of objects in page 62 * D. page->frozen -> frozen state 63 * 64 * If a slab is frozen then it is exempt from list management. It is not 65 * on any list except per cpu partial list. The processor that froze the 66 * slab is the one who can perform list operations on the page. Other 67 * processors may put objects onto the freelist but the processor that 68 * froze the slab is the only one that can retrieve the objects from the 69 * page's freelist. 70 * 71 * The list_lock protects the partial and full list on each node and 72 * the partial slab counter. If taken then no new slabs may be added or 73 * removed from the lists nor make the number of partial slabs be modified. 74 * (Note that the total number of slabs is an atomic value that may be 75 * modified without taking the list lock). 76 * 77 * The list_lock is a centralized lock and thus we avoid taking it as 78 * much as possible. As long as SLUB does not have to handle partial 79 * slabs, operations can continue without any centralized lock. F.e. 80 * allocating a long series of objects that fill up slabs does not require 81 * the list lock. 82 * Interrupts are disabled during allocation and deallocation in order to 83 * make the slab allocator safe to use in the context of an irq. In addition 84 * interrupts are disabled to ensure that the processor does not change 85 * while handling per_cpu slabs, due to kernel preemption. 86 * 87 * SLUB assigns one slab for allocation to each processor. 88 * Allocations only occur from these slabs called cpu slabs. 89 * 90 * Slabs with free elements are kept on a partial list and during regular 91 * operations no list for full slabs is used. If an object in a full slab is 92 * freed then the slab will show up again on the partial lists. 93 * We track full slabs for debugging purposes though because otherwise we 94 * cannot scan all objects. 95 * 96 * Slabs are freed when they become empty. Teardown and setup is 97 * minimal so we rely on the page allocators per cpu caches for 98 * fast frees and allocs. 99 * 100 * page->frozen The slab is frozen and exempt from list processing. 101 * This means that the slab is dedicated to a purpose 102 * such as satisfying allocations for a specific 103 * processor. Objects may be freed in the slab while 104 * it is frozen but slab_free will then skip the usual 105 * list operations. It is up to the processor holding 106 * the slab to integrate the slab into the slab lists 107 * when the slab is no longer needed. 108 * 109 * One use of this flag is to mark slabs that are 110 * used for allocations. Then such a slab becomes a cpu 111 * slab. The cpu slab may be equipped with an additional 112 * freelist that allows lockless access to 113 * free objects in addition to the regular freelist 114 * that requires the slab lock. 115 * 116 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug 117 * options set. This moves slab handling out of 118 * the fast path and disables lockless freelists. 119 */ 120 121 #ifdef CONFIG_SLUB_DEBUG 122 #ifdef CONFIG_SLUB_DEBUG_ON 123 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled); 124 #else 125 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled); 126 #endif 127 #endif /* CONFIG_SLUB_DEBUG */ 128 129 static inline bool kmem_cache_debug(struct kmem_cache *s) 130 { 131 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS); 132 } 133 134 void *fixup_red_left(struct kmem_cache *s, void *p) 135 { 136 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) 137 p += s->red_left_pad; 138 139 return p; 140 } 141 142 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) 143 { 144 #ifdef CONFIG_SLUB_CPU_PARTIAL 145 return !kmem_cache_debug(s); 146 #else 147 return false; 148 #endif 149 } 150 151 /* 152 * Issues still to be resolved: 153 * 154 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 155 * 156 * - Variable sizing of the per node arrays 157 */ 158 159 /* Enable to log cmpxchg failures */ 160 #undef SLUB_DEBUG_CMPXCHG 161 162 /* 163 * Minimum number of partial slabs. These will be left on the partial 164 * lists even if they are empty. kmem_cache_shrink may reclaim them. 165 */ 166 #define MIN_PARTIAL 5 167 168 /* 169 * Maximum number of desirable partial slabs. 170 * The existence of more partial slabs makes kmem_cache_shrink 171 * sort the partial list by the number of objects in use. 172 */ 173 #define MAX_PARTIAL 10 174 175 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ 176 SLAB_POISON | SLAB_STORE_USER) 177 178 /* 179 * These debug flags cannot use CMPXCHG because there might be consistency 180 * issues when checking or reading debug information 181 */ 182 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ 183 SLAB_TRACE) 184 185 186 /* 187 * Debugging flags that require metadata to be stored in the slab. These get 188 * disabled when slub_debug=O is used and a cache's min order increases with 189 * metadata. 190 */ 191 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 192 193 #define OO_SHIFT 16 194 #define OO_MASK ((1 << OO_SHIFT) - 1) 195 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ 196 197 /* Internal SLUB flags */ 198 /* Poison object */ 199 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U) 200 /* Use cmpxchg_double */ 201 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U) 202 203 /* 204 * Tracking user of a slab. 205 */ 206 #define TRACK_ADDRS_COUNT 16 207 struct track { 208 unsigned long addr; /* Called from address */ 209 #ifdef CONFIG_STACKTRACE 210 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 211 #endif 212 int cpu; /* Was running on cpu */ 213 int pid; /* Pid context */ 214 unsigned long when; /* When did the operation occur */ 215 }; 216 217 enum track_item { TRACK_ALLOC, TRACK_FREE }; 218 219 #ifdef CONFIG_SYSFS 220 static int sysfs_slab_add(struct kmem_cache *); 221 static int sysfs_slab_alias(struct kmem_cache *, const char *); 222 #else 223 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 224 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 225 { return 0; } 226 #endif 227 228 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG) 229 static void debugfs_slab_add(struct kmem_cache *); 230 #else 231 static inline void debugfs_slab_add(struct kmem_cache *s) { } 232 #endif 233 234 static inline void stat(const struct kmem_cache *s, enum stat_item si) 235 { 236 #ifdef CONFIG_SLUB_STATS 237 /* 238 * The rmw is racy on a preemptible kernel but this is acceptable, so 239 * avoid this_cpu_add()'s irq-disable overhead. 240 */ 241 raw_cpu_inc(s->cpu_slab->stat[si]); 242 #endif 243 } 244 245 /* 246 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated. 247 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily 248 * differ during memory hotplug/hotremove operations. 249 * Protected by slab_mutex. 250 */ 251 static nodemask_t slab_nodes; 252 253 /******************************************************************** 254 * Core slab cache functions 255 *******************************************************************/ 256 257 /* 258 * Returns freelist pointer (ptr). With hardening, this is obfuscated 259 * with an XOR of the address where the pointer is held and a per-cache 260 * random number. 261 */ 262 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr, 263 unsigned long ptr_addr) 264 { 265 #ifdef CONFIG_SLAB_FREELIST_HARDENED 266 /* 267 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged. 268 * Normally, this doesn't cause any issues, as both set_freepointer() 269 * and get_freepointer() are called with a pointer with the same tag. 270 * However, there are some issues with CONFIG_SLUB_DEBUG code. For 271 * example, when __free_slub() iterates over objects in a cache, it 272 * passes untagged pointers to check_object(). check_object() in turns 273 * calls get_freepointer() with an untagged pointer, which causes the 274 * freepointer to be restored incorrectly. 275 */ 276 return (void *)((unsigned long)ptr ^ s->random ^ 277 swab((unsigned long)kasan_reset_tag((void *)ptr_addr))); 278 #else 279 return ptr; 280 #endif 281 } 282 283 /* Returns the freelist pointer recorded at location ptr_addr. */ 284 static inline void *freelist_dereference(const struct kmem_cache *s, 285 void *ptr_addr) 286 { 287 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr), 288 (unsigned long)ptr_addr); 289 } 290 291 static inline void *get_freepointer(struct kmem_cache *s, void *object) 292 { 293 object = kasan_reset_tag(object); 294 return freelist_dereference(s, object + s->offset); 295 } 296 297 static void prefetch_freepointer(const struct kmem_cache *s, void *object) 298 { 299 prefetch(object + s->offset); 300 } 301 302 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 303 { 304 unsigned long freepointer_addr; 305 void *p; 306 307 if (!debug_pagealloc_enabled_static()) 308 return get_freepointer(s, object); 309 310 object = kasan_reset_tag(object); 311 freepointer_addr = (unsigned long)object + s->offset; 312 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p)); 313 return freelist_ptr(s, p, freepointer_addr); 314 } 315 316 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 317 { 318 unsigned long freeptr_addr = (unsigned long)object + s->offset; 319 320 #ifdef CONFIG_SLAB_FREELIST_HARDENED 321 BUG_ON(object == fp); /* naive detection of double free or corruption */ 322 #endif 323 324 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr); 325 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr); 326 } 327 328 /* Loop over all objects in a slab */ 329 #define for_each_object(__p, __s, __addr, __objects) \ 330 for (__p = fixup_red_left(__s, __addr); \ 331 __p < (__addr) + (__objects) * (__s)->size; \ 332 __p += (__s)->size) 333 334 static inline unsigned int order_objects(unsigned int order, unsigned int size) 335 { 336 return ((unsigned int)PAGE_SIZE << order) / size; 337 } 338 339 static inline struct kmem_cache_order_objects oo_make(unsigned int order, 340 unsigned int size) 341 { 342 struct kmem_cache_order_objects x = { 343 (order << OO_SHIFT) + order_objects(order, size) 344 }; 345 346 return x; 347 } 348 349 static inline unsigned int oo_order(struct kmem_cache_order_objects x) 350 { 351 return x.x >> OO_SHIFT; 352 } 353 354 static inline unsigned int oo_objects(struct kmem_cache_order_objects x) 355 { 356 return x.x & OO_MASK; 357 } 358 359 /* 360 * Per slab locking using the pagelock 361 */ 362 static __always_inline void slab_lock(struct page *page) 363 { 364 VM_BUG_ON_PAGE(PageTail(page), page); 365 bit_spin_lock(PG_locked, &page->flags); 366 } 367 368 static __always_inline void slab_unlock(struct page *page) 369 { 370 VM_BUG_ON_PAGE(PageTail(page), page); 371 __bit_spin_unlock(PG_locked, &page->flags); 372 } 373 374 /* Interrupts must be disabled (for the fallback code to work right) */ 375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 376 void *freelist_old, unsigned long counters_old, 377 void *freelist_new, unsigned long counters_new, 378 const char *n) 379 { 380 VM_BUG_ON(!irqs_disabled()); 381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 383 if (s->flags & __CMPXCHG_DOUBLE) { 384 if (cmpxchg_double(&page->freelist, &page->counters, 385 freelist_old, counters_old, 386 freelist_new, counters_new)) 387 return true; 388 } else 389 #endif 390 { 391 slab_lock(page); 392 if (page->freelist == freelist_old && 393 page->counters == counters_old) { 394 page->freelist = freelist_new; 395 page->counters = counters_new; 396 slab_unlock(page); 397 return true; 398 } 399 slab_unlock(page); 400 } 401 402 cpu_relax(); 403 stat(s, CMPXCHG_DOUBLE_FAIL); 404 405 #ifdef SLUB_DEBUG_CMPXCHG 406 pr_info("%s %s: cmpxchg double redo ", n, s->name); 407 #endif 408 409 return false; 410 } 411 412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 413 void *freelist_old, unsigned long counters_old, 414 void *freelist_new, unsigned long counters_new, 415 const char *n) 416 { 417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 419 if (s->flags & __CMPXCHG_DOUBLE) { 420 if (cmpxchg_double(&page->freelist, &page->counters, 421 freelist_old, counters_old, 422 freelist_new, counters_new)) 423 return true; 424 } else 425 #endif 426 { 427 unsigned long flags; 428 429 local_irq_save(flags); 430 slab_lock(page); 431 if (page->freelist == freelist_old && 432 page->counters == counters_old) { 433 page->freelist = freelist_new; 434 page->counters = counters_new; 435 slab_unlock(page); 436 local_irq_restore(flags); 437 return true; 438 } 439 slab_unlock(page); 440 local_irq_restore(flags); 441 } 442 443 cpu_relax(); 444 stat(s, CMPXCHG_DOUBLE_FAIL); 445 446 #ifdef SLUB_DEBUG_CMPXCHG 447 pr_info("%s %s: cmpxchg double redo ", n, s->name); 448 #endif 449 450 return false; 451 } 452 453 #ifdef CONFIG_SLUB_DEBUG 454 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)]; 455 static DEFINE_SPINLOCK(object_map_lock); 456 457 #if IS_ENABLED(CONFIG_KUNIT) 458 static bool slab_add_kunit_errors(void) 459 { 460 struct kunit_resource *resource; 461 462 if (likely(!current->kunit_test)) 463 return false; 464 465 resource = kunit_find_named_resource(current->kunit_test, "slab_errors"); 466 if (!resource) 467 return false; 468 469 (*(int *)resource->data)++; 470 kunit_put_resource(resource); 471 return true; 472 } 473 #else 474 static inline bool slab_add_kunit_errors(void) { return false; } 475 #endif 476 477 /* 478 * Determine a map of object in use on a page. 479 * 480 * Node listlock must be held to guarantee that the page does 481 * not vanish from under us. 482 */ 483 static unsigned long *get_map(struct kmem_cache *s, struct page *page) 484 __acquires(&object_map_lock) 485 { 486 void *p; 487 void *addr = page_address(page); 488 489 VM_BUG_ON(!irqs_disabled()); 490 491 spin_lock(&object_map_lock); 492 493 bitmap_zero(object_map, page->objects); 494 495 for (p = page->freelist; p; p = get_freepointer(s, p)) 496 set_bit(__obj_to_index(s, addr, p), object_map); 497 498 return object_map; 499 } 500 501 static void put_map(unsigned long *map) __releases(&object_map_lock) 502 { 503 VM_BUG_ON(map != object_map); 504 spin_unlock(&object_map_lock); 505 } 506 507 static inline unsigned int size_from_object(struct kmem_cache *s) 508 { 509 if (s->flags & SLAB_RED_ZONE) 510 return s->size - s->red_left_pad; 511 512 return s->size; 513 } 514 515 static inline void *restore_red_left(struct kmem_cache *s, void *p) 516 { 517 if (s->flags & SLAB_RED_ZONE) 518 p -= s->red_left_pad; 519 520 return p; 521 } 522 523 /* 524 * Debug settings: 525 */ 526 #if defined(CONFIG_SLUB_DEBUG_ON) 527 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; 528 #else 529 static slab_flags_t slub_debug; 530 #endif 531 532 static char *slub_debug_string; 533 static int disable_higher_order_debug; 534 535 /* 536 * slub is about to manipulate internal object metadata. This memory lies 537 * outside the range of the allocated object, so accessing it would normally 538 * be reported by kasan as a bounds error. metadata_access_enable() is used 539 * to tell kasan that these accesses are OK. 540 */ 541 static inline void metadata_access_enable(void) 542 { 543 kasan_disable_current(); 544 } 545 546 static inline void metadata_access_disable(void) 547 { 548 kasan_enable_current(); 549 } 550 551 /* 552 * Object debugging 553 */ 554 555 /* Verify that a pointer has an address that is valid within a slab page */ 556 static inline int check_valid_pointer(struct kmem_cache *s, 557 struct page *page, void *object) 558 { 559 void *base; 560 561 if (!object) 562 return 1; 563 564 base = page_address(page); 565 object = kasan_reset_tag(object); 566 object = restore_red_left(s, object); 567 if (object < base || object >= base + page->objects * s->size || 568 (object - base) % s->size) { 569 return 0; 570 } 571 572 return 1; 573 } 574 575 static void print_section(char *level, char *text, u8 *addr, 576 unsigned int length) 577 { 578 metadata_access_enable(); 579 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 580 16, 1, kasan_reset_tag((void *)addr), length, 1); 581 metadata_access_disable(); 582 } 583 584 /* 585 * See comment in calculate_sizes(). 586 */ 587 static inline bool freeptr_outside_object(struct kmem_cache *s) 588 { 589 return s->offset >= s->inuse; 590 } 591 592 /* 593 * Return offset of the end of info block which is inuse + free pointer if 594 * not overlapping with object. 595 */ 596 static inline unsigned int get_info_end(struct kmem_cache *s) 597 { 598 if (freeptr_outside_object(s)) 599 return s->inuse + sizeof(void *); 600 else 601 return s->inuse; 602 } 603 604 static struct track *get_track(struct kmem_cache *s, void *object, 605 enum track_item alloc) 606 { 607 struct track *p; 608 609 p = object + get_info_end(s); 610 611 return kasan_reset_tag(p + alloc); 612 } 613 614 static void set_track(struct kmem_cache *s, void *object, 615 enum track_item alloc, unsigned long addr) 616 { 617 struct track *p = get_track(s, object, alloc); 618 619 if (addr) { 620 #ifdef CONFIG_STACKTRACE 621 unsigned int nr_entries; 622 623 metadata_access_enable(); 624 nr_entries = stack_trace_save(kasan_reset_tag(p->addrs), 625 TRACK_ADDRS_COUNT, 3); 626 metadata_access_disable(); 627 628 if (nr_entries < TRACK_ADDRS_COUNT) 629 p->addrs[nr_entries] = 0; 630 #endif 631 p->addr = addr; 632 p->cpu = smp_processor_id(); 633 p->pid = current->pid; 634 p->when = jiffies; 635 } else { 636 memset(p, 0, sizeof(struct track)); 637 } 638 } 639 640 static void init_tracking(struct kmem_cache *s, void *object) 641 { 642 if (!(s->flags & SLAB_STORE_USER)) 643 return; 644 645 set_track(s, object, TRACK_FREE, 0UL); 646 set_track(s, object, TRACK_ALLOC, 0UL); 647 } 648 649 static void print_track(const char *s, struct track *t, unsigned long pr_time) 650 { 651 if (!t->addr) 652 return; 653 654 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n", 655 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid); 656 #ifdef CONFIG_STACKTRACE 657 { 658 int i; 659 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 660 if (t->addrs[i]) 661 pr_err("\t%pS\n", (void *)t->addrs[i]); 662 else 663 break; 664 } 665 #endif 666 } 667 668 void print_tracking(struct kmem_cache *s, void *object) 669 { 670 unsigned long pr_time = jiffies; 671 if (!(s->flags & SLAB_STORE_USER)) 672 return; 673 674 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time); 675 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time); 676 } 677 678 static void print_page_info(struct page *page) 679 { 680 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n", 681 page, page->objects, page->inuse, page->freelist, 682 page->flags, &page->flags); 683 684 } 685 686 static void slab_bug(struct kmem_cache *s, char *fmt, ...) 687 { 688 struct va_format vaf; 689 va_list args; 690 691 va_start(args, fmt); 692 vaf.fmt = fmt; 693 vaf.va = &args; 694 pr_err("=============================================================================\n"); 695 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf); 696 pr_err("-----------------------------------------------------------------------------\n\n"); 697 va_end(args); 698 } 699 700 __printf(2, 3) 701 static void slab_fix(struct kmem_cache *s, char *fmt, ...) 702 { 703 struct va_format vaf; 704 va_list args; 705 706 if (slab_add_kunit_errors()) 707 return; 708 709 va_start(args, fmt); 710 vaf.fmt = fmt; 711 vaf.va = &args; 712 pr_err("FIX %s: %pV\n", s->name, &vaf); 713 va_end(args); 714 } 715 716 static bool freelist_corrupted(struct kmem_cache *s, struct page *page, 717 void **freelist, void *nextfree) 718 { 719 if ((s->flags & SLAB_CONSISTENCY_CHECKS) && 720 !check_valid_pointer(s, page, nextfree) && freelist) { 721 object_err(s, page, *freelist, "Freechain corrupt"); 722 *freelist = NULL; 723 slab_fix(s, "Isolate corrupted freechain"); 724 return true; 725 } 726 727 return false; 728 } 729 730 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 731 { 732 unsigned int off; /* Offset of last byte */ 733 u8 *addr = page_address(page); 734 735 print_tracking(s, p); 736 737 print_page_info(page); 738 739 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n", 740 p, p - addr, get_freepointer(s, p)); 741 742 if (s->flags & SLAB_RED_ZONE) 743 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad, 744 s->red_left_pad); 745 else if (p > addr + 16) 746 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16); 747 748 print_section(KERN_ERR, "Object ", p, 749 min_t(unsigned int, s->object_size, PAGE_SIZE)); 750 if (s->flags & SLAB_RED_ZONE) 751 print_section(KERN_ERR, "Redzone ", p + s->object_size, 752 s->inuse - s->object_size); 753 754 off = get_info_end(s); 755 756 if (s->flags & SLAB_STORE_USER) 757 off += 2 * sizeof(struct track); 758 759 off += kasan_metadata_size(s); 760 761 if (off != size_from_object(s)) 762 /* Beginning of the filler is the free pointer */ 763 print_section(KERN_ERR, "Padding ", p + off, 764 size_from_object(s) - off); 765 766 dump_stack(); 767 } 768 769 void object_err(struct kmem_cache *s, struct page *page, 770 u8 *object, char *reason) 771 { 772 if (slab_add_kunit_errors()) 773 return; 774 775 slab_bug(s, "%s", reason); 776 print_trailer(s, page, object); 777 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 778 } 779 780 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page, 781 const char *fmt, ...) 782 { 783 va_list args; 784 char buf[100]; 785 786 if (slab_add_kunit_errors()) 787 return; 788 789 va_start(args, fmt); 790 vsnprintf(buf, sizeof(buf), fmt, args); 791 va_end(args); 792 slab_bug(s, "%s", buf); 793 print_page_info(page); 794 dump_stack(); 795 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 796 } 797 798 static void init_object(struct kmem_cache *s, void *object, u8 val) 799 { 800 u8 *p = kasan_reset_tag(object); 801 802 if (s->flags & SLAB_RED_ZONE) 803 memset(p - s->red_left_pad, val, s->red_left_pad); 804 805 if (s->flags & __OBJECT_POISON) { 806 memset(p, POISON_FREE, s->object_size - 1); 807 p[s->object_size - 1] = POISON_END; 808 } 809 810 if (s->flags & SLAB_RED_ZONE) 811 memset(p + s->object_size, val, s->inuse - s->object_size); 812 } 813 814 static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 815 void *from, void *to) 816 { 817 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data); 818 memset(from, data, to - from); 819 } 820 821 static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 822 u8 *object, char *what, 823 u8 *start, unsigned int value, unsigned int bytes) 824 { 825 u8 *fault; 826 u8 *end; 827 u8 *addr = page_address(page); 828 829 metadata_access_enable(); 830 fault = memchr_inv(kasan_reset_tag(start), value, bytes); 831 metadata_access_disable(); 832 if (!fault) 833 return 1; 834 835 end = start + bytes; 836 while (end > fault && end[-1] == value) 837 end--; 838 839 if (slab_add_kunit_errors()) 840 goto skip_bug_print; 841 842 slab_bug(s, "%s overwritten", what); 843 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n", 844 fault, end - 1, fault - addr, 845 fault[0], value); 846 print_trailer(s, page, object); 847 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 848 849 skip_bug_print: 850 restore_bytes(s, what, value, fault, end); 851 return 0; 852 } 853 854 /* 855 * Object layout: 856 * 857 * object address 858 * Bytes of the object to be managed. 859 * If the freepointer may overlay the object then the free 860 * pointer is at the middle of the object. 861 * 862 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 863 * 0xa5 (POISON_END) 864 * 865 * object + s->object_size 866 * Padding to reach word boundary. This is also used for Redzoning. 867 * Padding is extended by another word if Redzoning is enabled and 868 * object_size == inuse. 869 * 870 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 871 * 0xcc (RED_ACTIVE) for objects in use. 872 * 873 * object + s->inuse 874 * Meta data starts here. 875 * 876 * A. Free pointer (if we cannot overwrite object on free) 877 * B. Tracking data for SLAB_STORE_USER 878 * C. Padding to reach required alignment boundary or at minimum 879 * one word if debugging is on to be able to detect writes 880 * before the word boundary. 881 * 882 * Padding is done using 0x5a (POISON_INUSE) 883 * 884 * object + s->size 885 * Nothing is used beyond s->size. 886 * 887 * If slabcaches are merged then the object_size and inuse boundaries are mostly 888 * ignored. And therefore no slab options that rely on these boundaries 889 * may be used with merged slabcaches. 890 */ 891 892 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 893 { 894 unsigned long off = get_info_end(s); /* The end of info */ 895 896 if (s->flags & SLAB_STORE_USER) 897 /* We also have user information there */ 898 off += 2 * sizeof(struct track); 899 900 off += kasan_metadata_size(s); 901 902 if (size_from_object(s) == off) 903 return 1; 904 905 return check_bytes_and_report(s, page, p, "Object padding", 906 p + off, POISON_INUSE, size_from_object(s) - off); 907 } 908 909 /* Check the pad bytes at the end of a slab page */ 910 static int slab_pad_check(struct kmem_cache *s, struct page *page) 911 { 912 u8 *start; 913 u8 *fault; 914 u8 *end; 915 u8 *pad; 916 int length; 917 int remainder; 918 919 if (!(s->flags & SLAB_POISON)) 920 return 1; 921 922 start = page_address(page); 923 length = page_size(page); 924 end = start + length; 925 remainder = length % s->size; 926 if (!remainder) 927 return 1; 928 929 pad = end - remainder; 930 metadata_access_enable(); 931 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder); 932 metadata_access_disable(); 933 if (!fault) 934 return 1; 935 while (end > fault && end[-1] == POISON_INUSE) 936 end--; 937 938 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu", 939 fault, end - 1, fault - start); 940 print_section(KERN_ERR, "Padding ", pad, remainder); 941 942 restore_bytes(s, "slab padding", POISON_INUSE, fault, end); 943 return 0; 944 } 945 946 static int check_object(struct kmem_cache *s, struct page *page, 947 void *object, u8 val) 948 { 949 u8 *p = object; 950 u8 *endobject = object + s->object_size; 951 952 if (s->flags & SLAB_RED_ZONE) { 953 if (!check_bytes_and_report(s, page, object, "Left Redzone", 954 object - s->red_left_pad, val, s->red_left_pad)) 955 return 0; 956 957 if (!check_bytes_and_report(s, page, object, "Right Redzone", 958 endobject, val, s->inuse - s->object_size)) 959 return 0; 960 } else { 961 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { 962 check_bytes_and_report(s, page, p, "Alignment padding", 963 endobject, POISON_INUSE, 964 s->inuse - s->object_size); 965 } 966 } 967 968 if (s->flags & SLAB_POISON) { 969 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 970 (!check_bytes_and_report(s, page, p, "Poison", p, 971 POISON_FREE, s->object_size - 1) || 972 !check_bytes_and_report(s, page, p, "End Poison", 973 p + s->object_size - 1, POISON_END, 1))) 974 return 0; 975 /* 976 * check_pad_bytes cleans up on its own. 977 */ 978 check_pad_bytes(s, page, p); 979 } 980 981 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE) 982 /* 983 * Object and freepointer overlap. Cannot check 984 * freepointer while object is allocated. 985 */ 986 return 1; 987 988 /* Check free pointer validity */ 989 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 990 object_err(s, page, p, "Freepointer corrupt"); 991 /* 992 * No choice but to zap it and thus lose the remainder 993 * of the free objects in this slab. May cause 994 * another error because the object count is now wrong. 995 */ 996 set_freepointer(s, p, NULL); 997 return 0; 998 } 999 return 1; 1000 } 1001 1002 static int check_slab(struct kmem_cache *s, struct page *page) 1003 { 1004 int maxobj; 1005 1006 VM_BUG_ON(!irqs_disabled()); 1007 1008 if (!PageSlab(page)) { 1009 slab_err(s, page, "Not a valid slab page"); 1010 return 0; 1011 } 1012 1013 maxobj = order_objects(compound_order(page), s->size); 1014 if (page->objects > maxobj) { 1015 slab_err(s, page, "objects %u > max %u", 1016 page->objects, maxobj); 1017 return 0; 1018 } 1019 if (page->inuse > page->objects) { 1020 slab_err(s, page, "inuse %u > max %u", 1021 page->inuse, page->objects); 1022 return 0; 1023 } 1024 /* Slab_pad_check fixes things up after itself */ 1025 slab_pad_check(s, page); 1026 return 1; 1027 } 1028 1029 /* 1030 * Determine if a certain object on a page is on the freelist. Must hold the 1031 * slab lock to guarantee that the chains are in a consistent state. 1032 */ 1033 static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 1034 { 1035 int nr = 0; 1036 void *fp; 1037 void *object = NULL; 1038 int max_objects; 1039 1040 fp = page->freelist; 1041 while (fp && nr <= page->objects) { 1042 if (fp == search) 1043 return 1; 1044 if (!check_valid_pointer(s, page, fp)) { 1045 if (object) { 1046 object_err(s, page, object, 1047 "Freechain corrupt"); 1048 set_freepointer(s, object, NULL); 1049 } else { 1050 slab_err(s, page, "Freepointer corrupt"); 1051 page->freelist = NULL; 1052 page->inuse = page->objects; 1053 slab_fix(s, "Freelist cleared"); 1054 return 0; 1055 } 1056 break; 1057 } 1058 object = fp; 1059 fp = get_freepointer(s, object); 1060 nr++; 1061 } 1062 1063 max_objects = order_objects(compound_order(page), s->size); 1064 if (max_objects > MAX_OBJS_PER_PAGE) 1065 max_objects = MAX_OBJS_PER_PAGE; 1066 1067 if (page->objects != max_objects) { 1068 slab_err(s, page, "Wrong number of objects. Found %d but should be %d", 1069 page->objects, max_objects); 1070 page->objects = max_objects; 1071 slab_fix(s, "Number of objects adjusted"); 1072 } 1073 if (page->inuse != page->objects - nr) { 1074 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d", 1075 page->inuse, page->objects - nr); 1076 page->inuse = page->objects - nr; 1077 slab_fix(s, "Object count adjusted"); 1078 } 1079 return search == NULL; 1080 } 1081 1082 static void trace(struct kmem_cache *s, struct page *page, void *object, 1083 int alloc) 1084 { 1085 if (s->flags & SLAB_TRACE) { 1086 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 1087 s->name, 1088 alloc ? "alloc" : "free", 1089 object, page->inuse, 1090 page->freelist); 1091 1092 if (!alloc) 1093 print_section(KERN_INFO, "Object ", (void *)object, 1094 s->object_size); 1095 1096 dump_stack(); 1097 } 1098 } 1099 1100 /* 1101 * Tracking of fully allocated slabs for debugging purposes. 1102 */ 1103 static void add_full(struct kmem_cache *s, 1104 struct kmem_cache_node *n, struct page *page) 1105 { 1106 if (!(s->flags & SLAB_STORE_USER)) 1107 return; 1108 1109 lockdep_assert_held(&n->list_lock); 1110 list_add(&page->slab_list, &n->full); 1111 } 1112 1113 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page) 1114 { 1115 if (!(s->flags & SLAB_STORE_USER)) 1116 return; 1117 1118 lockdep_assert_held(&n->list_lock); 1119 list_del(&page->slab_list); 1120 } 1121 1122 /* Tracking of the number of slabs for debugging purposes */ 1123 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1124 { 1125 struct kmem_cache_node *n = get_node(s, node); 1126 1127 return atomic_long_read(&n->nr_slabs); 1128 } 1129 1130 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1131 { 1132 return atomic_long_read(&n->nr_slabs); 1133 } 1134 1135 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 1136 { 1137 struct kmem_cache_node *n = get_node(s, node); 1138 1139 /* 1140 * May be called early in order to allocate a slab for the 1141 * kmem_cache_node structure. Solve the chicken-egg 1142 * dilemma by deferring the increment of the count during 1143 * bootstrap (see early_kmem_cache_node_alloc). 1144 */ 1145 if (likely(n)) { 1146 atomic_long_inc(&n->nr_slabs); 1147 atomic_long_add(objects, &n->total_objects); 1148 } 1149 } 1150 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 1151 { 1152 struct kmem_cache_node *n = get_node(s, node); 1153 1154 atomic_long_dec(&n->nr_slabs); 1155 atomic_long_sub(objects, &n->total_objects); 1156 } 1157 1158 /* Object debug checks for alloc/free paths */ 1159 static void setup_object_debug(struct kmem_cache *s, struct page *page, 1160 void *object) 1161 { 1162 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)) 1163 return; 1164 1165 init_object(s, object, SLUB_RED_INACTIVE); 1166 init_tracking(s, object); 1167 } 1168 1169 static 1170 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) 1171 { 1172 if (!kmem_cache_debug_flags(s, SLAB_POISON)) 1173 return; 1174 1175 metadata_access_enable(); 1176 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page)); 1177 metadata_access_disable(); 1178 } 1179 1180 static inline int alloc_consistency_checks(struct kmem_cache *s, 1181 struct page *page, void *object) 1182 { 1183 if (!check_slab(s, page)) 1184 return 0; 1185 1186 if (!check_valid_pointer(s, page, object)) { 1187 object_err(s, page, object, "Freelist Pointer check fails"); 1188 return 0; 1189 } 1190 1191 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1192 return 0; 1193 1194 return 1; 1195 } 1196 1197 static noinline int alloc_debug_processing(struct kmem_cache *s, 1198 struct page *page, 1199 void *object, unsigned long addr) 1200 { 1201 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1202 if (!alloc_consistency_checks(s, page, object)) 1203 goto bad; 1204 } 1205 1206 /* Success perform special debug activities for allocs */ 1207 if (s->flags & SLAB_STORE_USER) 1208 set_track(s, object, TRACK_ALLOC, addr); 1209 trace(s, page, object, 1); 1210 init_object(s, object, SLUB_RED_ACTIVE); 1211 return 1; 1212 1213 bad: 1214 if (PageSlab(page)) { 1215 /* 1216 * If this is a slab page then lets do the best we can 1217 * to avoid issues in the future. Marking all objects 1218 * as used avoids touching the remaining objects. 1219 */ 1220 slab_fix(s, "Marking all objects used"); 1221 page->inuse = page->objects; 1222 page->freelist = NULL; 1223 } 1224 return 0; 1225 } 1226 1227 static inline int free_consistency_checks(struct kmem_cache *s, 1228 struct page *page, void *object, unsigned long addr) 1229 { 1230 if (!check_valid_pointer(s, page, object)) { 1231 slab_err(s, page, "Invalid object pointer 0x%p", object); 1232 return 0; 1233 } 1234 1235 if (on_freelist(s, page, object)) { 1236 object_err(s, page, object, "Object already free"); 1237 return 0; 1238 } 1239 1240 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1241 return 0; 1242 1243 if (unlikely(s != page->slab_cache)) { 1244 if (!PageSlab(page)) { 1245 slab_err(s, page, "Attempt to free object(0x%p) outside of slab", 1246 object); 1247 } else if (!page->slab_cache) { 1248 pr_err("SLUB <none>: no slab for object 0x%p.\n", 1249 object); 1250 dump_stack(); 1251 } else 1252 object_err(s, page, object, 1253 "page slab pointer corrupt."); 1254 return 0; 1255 } 1256 return 1; 1257 } 1258 1259 /* Supports checking bulk free of a constructed freelist */ 1260 static noinline int free_debug_processing( 1261 struct kmem_cache *s, struct page *page, 1262 void *head, void *tail, int bulk_cnt, 1263 unsigned long addr) 1264 { 1265 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1266 void *object = head; 1267 int cnt = 0; 1268 unsigned long flags; 1269 int ret = 0; 1270 1271 spin_lock_irqsave(&n->list_lock, flags); 1272 slab_lock(page); 1273 1274 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1275 if (!check_slab(s, page)) 1276 goto out; 1277 } 1278 1279 next_object: 1280 cnt++; 1281 1282 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1283 if (!free_consistency_checks(s, page, object, addr)) 1284 goto out; 1285 } 1286 1287 if (s->flags & SLAB_STORE_USER) 1288 set_track(s, object, TRACK_FREE, addr); 1289 trace(s, page, object, 0); 1290 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ 1291 init_object(s, object, SLUB_RED_INACTIVE); 1292 1293 /* Reached end of constructed freelist yet? */ 1294 if (object != tail) { 1295 object = get_freepointer(s, object); 1296 goto next_object; 1297 } 1298 ret = 1; 1299 1300 out: 1301 if (cnt != bulk_cnt) 1302 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n", 1303 bulk_cnt, cnt); 1304 1305 slab_unlock(page); 1306 spin_unlock_irqrestore(&n->list_lock, flags); 1307 if (!ret) 1308 slab_fix(s, "Object at 0x%p not freed", object); 1309 return ret; 1310 } 1311 1312 /* 1313 * Parse a block of slub_debug options. Blocks are delimited by ';' 1314 * 1315 * @str: start of block 1316 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified 1317 * @slabs: return start of list of slabs, or NULL when there's no list 1318 * @init: assume this is initial parsing and not per-kmem-create parsing 1319 * 1320 * returns the start of next block if there's any, or NULL 1321 */ 1322 static char * 1323 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init) 1324 { 1325 bool higher_order_disable = false; 1326 1327 /* Skip any completely empty blocks */ 1328 while (*str && *str == ';') 1329 str++; 1330 1331 if (*str == ',') { 1332 /* 1333 * No options but restriction on slabs. This means full 1334 * debugging for slabs matching a pattern. 1335 */ 1336 *flags = DEBUG_DEFAULT_FLAGS; 1337 goto check_slabs; 1338 } 1339 *flags = 0; 1340 1341 /* Determine which debug features should be switched on */ 1342 for (; *str && *str != ',' && *str != ';'; str++) { 1343 switch (tolower(*str)) { 1344 case '-': 1345 *flags = 0; 1346 break; 1347 case 'f': 1348 *flags |= SLAB_CONSISTENCY_CHECKS; 1349 break; 1350 case 'z': 1351 *flags |= SLAB_RED_ZONE; 1352 break; 1353 case 'p': 1354 *flags |= SLAB_POISON; 1355 break; 1356 case 'u': 1357 *flags |= SLAB_STORE_USER; 1358 break; 1359 case 't': 1360 *flags |= SLAB_TRACE; 1361 break; 1362 case 'a': 1363 *flags |= SLAB_FAILSLAB; 1364 break; 1365 case 'o': 1366 /* 1367 * Avoid enabling debugging on caches if its minimum 1368 * order would increase as a result. 1369 */ 1370 higher_order_disable = true; 1371 break; 1372 default: 1373 if (init) 1374 pr_err("slub_debug option '%c' unknown. skipped\n", *str); 1375 } 1376 } 1377 check_slabs: 1378 if (*str == ',') 1379 *slabs = ++str; 1380 else 1381 *slabs = NULL; 1382 1383 /* Skip over the slab list */ 1384 while (*str && *str != ';') 1385 str++; 1386 1387 /* Skip any completely empty blocks */ 1388 while (*str && *str == ';') 1389 str++; 1390 1391 if (init && higher_order_disable) 1392 disable_higher_order_debug = 1; 1393 1394 if (*str) 1395 return str; 1396 else 1397 return NULL; 1398 } 1399 1400 static int __init setup_slub_debug(char *str) 1401 { 1402 slab_flags_t flags; 1403 slab_flags_t global_flags; 1404 char *saved_str; 1405 char *slab_list; 1406 bool global_slub_debug_changed = false; 1407 bool slab_list_specified = false; 1408 1409 global_flags = DEBUG_DEFAULT_FLAGS; 1410 if (*str++ != '=' || !*str) 1411 /* 1412 * No options specified. Switch on full debugging. 1413 */ 1414 goto out; 1415 1416 saved_str = str; 1417 while (str) { 1418 str = parse_slub_debug_flags(str, &flags, &slab_list, true); 1419 1420 if (!slab_list) { 1421 global_flags = flags; 1422 global_slub_debug_changed = true; 1423 } else { 1424 slab_list_specified = true; 1425 } 1426 } 1427 1428 /* 1429 * For backwards compatibility, a single list of flags with list of 1430 * slabs means debugging is only changed for those slabs, so the global 1431 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending 1432 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as 1433 * long as there is no option specifying flags without a slab list. 1434 */ 1435 if (slab_list_specified) { 1436 if (!global_slub_debug_changed) 1437 global_flags = slub_debug; 1438 slub_debug_string = saved_str; 1439 } 1440 out: 1441 slub_debug = global_flags; 1442 if (slub_debug != 0 || slub_debug_string) 1443 static_branch_enable(&slub_debug_enabled); 1444 else 1445 static_branch_disable(&slub_debug_enabled); 1446 if ((static_branch_unlikely(&init_on_alloc) || 1447 static_branch_unlikely(&init_on_free)) && 1448 (slub_debug & SLAB_POISON)) 1449 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n"); 1450 return 1; 1451 } 1452 1453 __setup("slub_debug", setup_slub_debug); 1454 1455 /* 1456 * kmem_cache_flags - apply debugging options to the cache 1457 * @object_size: the size of an object without meta data 1458 * @flags: flags to set 1459 * @name: name of the cache 1460 * 1461 * Debug option(s) are applied to @flags. In addition to the debug 1462 * option(s), if a slab name (or multiple) is specified i.e. 1463 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ... 1464 * then only the select slabs will receive the debug option(s). 1465 */ 1466 slab_flags_t kmem_cache_flags(unsigned int object_size, 1467 slab_flags_t flags, const char *name) 1468 { 1469 char *iter; 1470 size_t len; 1471 char *next_block; 1472 slab_flags_t block_flags; 1473 slab_flags_t slub_debug_local = slub_debug; 1474 1475 /* 1476 * If the slab cache is for debugging (e.g. kmemleak) then 1477 * don't store user (stack trace) information by default, 1478 * but let the user enable it via the command line below. 1479 */ 1480 if (flags & SLAB_NOLEAKTRACE) 1481 slub_debug_local &= ~SLAB_STORE_USER; 1482 1483 len = strlen(name); 1484 next_block = slub_debug_string; 1485 /* Go through all blocks of debug options, see if any matches our slab's name */ 1486 while (next_block) { 1487 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false); 1488 if (!iter) 1489 continue; 1490 /* Found a block that has a slab list, search it */ 1491 while (*iter) { 1492 char *end, *glob; 1493 size_t cmplen; 1494 1495 end = strchrnul(iter, ','); 1496 if (next_block && next_block < end) 1497 end = next_block - 1; 1498 1499 glob = strnchr(iter, end - iter, '*'); 1500 if (glob) 1501 cmplen = glob - iter; 1502 else 1503 cmplen = max_t(size_t, len, (end - iter)); 1504 1505 if (!strncmp(name, iter, cmplen)) { 1506 flags |= block_flags; 1507 return flags; 1508 } 1509 1510 if (!*end || *end == ';') 1511 break; 1512 iter = end + 1; 1513 } 1514 } 1515 1516 return flags | slub_debug_local; 1517 } 1518 #else /* !CONFIG_SLUB_DEBUG */ 1519 static inline void setup_object_debug(struct kmem_cache *s, 1520 struct page *page, void *object) {} 1521 static inline 1522 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {} 1523 1524 static inline int alloc_debug_processing(struct kmem_cache *s, 1525 struct page *page, void *object, unsigned long addr) { return 0; } 1526 1527 static inline int free_debug_processing( 1528 struct kmem_cache *s, struct page *page, 1529 void *head, void *tail, int bulk_cnt, 1530 unsigned long addr) { return 0; } 1531 1532 static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1533 { return 1; } 1534 static inline int check_object(struct kmem_cache *s, struct page *page, 1535 void *object, u8 val) { return 1; } 1536 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, 1537 struct page *page) {} 1538 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, 1539 struct page *page) {} 1540 slab_flags_t kmem_cache_flags(unsigned int object_size, 1541 slab_flags_t flags, const char *name) 1542 { 1543 return flags; 1544 } 1545 #define slub_debug 0 1546 1547 #define disable_higher_order_debug 0 1548 1549 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1550 { return 0; } 1551 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1552 { return 0; } 1553 static inline void inc_slabs_node(struct kmem_cache *s, int node, 1554 int objects) {} 1555 static inline void dec_slabs_node(struct kmem_cache *s, int node, 1556 int objects) {} 1557 1558 static bool freelist_corrupted(struct kmem_cache *s, struct page *page, 1559 void **freelist, void *nextfree) 1560 { 1561 return false; 1562 } 1563 #endif /* CONFIG_SLUB_DEBUG */ 1564 1565 /* 1566 * Hooks for other subsystems that check memory allocations. In a typical 1567 * production configuration these hooks all should produce no code at all. 1568 */ 1569 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) 1570 { 1571 ptr = kasan_kmalloc_large(ptr, size, flags); 1572 /* As ptr might get tagged, call kmemleak hook after KASAN. */ 1573 kmemleak_alloc(ptr, size, 1, flags); 1574 return ptr; 1575 } 1576 1577 static __always_inline void kfree_hook(void *x) 1578 { 1579 kmemleak_free(x); 1580 kasan_kfree_large(x); 1581 } 1582 1583 static __always_inline bool slab_free_hook(struct kmem_cache *s, 1584 void *x, bool init) 1585 { 1586 kmemleak_free_recursive(x, s->flags); 1587 1588 /* 1589 * Trouble is that we may no longer disable interrupts in the fast path 1590 * So in order to make the debug calls that expect irqs to be 1591 * disabled we need to disable interrupts temporarily. 1592 */ 1593 #ifdef CONFIG_LOCKDEP 1594 { 1595 unsigned long flags; 1596 1597 local_irq_save(flags); 1598 debug_check_no_locks_freed(x, s->object_size); 1599 local_irq_restore(flags); 1600 } 1601 #endif 1602 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1603 debug_check_no_obj_freed(x, s->object_size); 1604 1605 /* Use KCSAN to help debug racy use-after-free. */ 1606 if (!(s->flags & SLAB_TYPESAFE_BY_RCU)) 1607 __kcsan_check_access(x, s->object_size, 1608 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT); 1609 1610 /* 1611 * As memory initialization might be integrated into KASAN, 1612 * kasan_slab_free and initialization memset's must be 1613 * kept together to avoid discrepancies in behavior. 1614 * 1615 * The initialization memset's clear the object and the metadata, 1616 * but don't touch the SLAB redzone. 1617 */ 1618 if (init) { 1619 int rsize; 1620 1621 if (!kasan_has_integrated_init()) 1622 memset(kasan_reset_tag(x), 0, s->object_size); 1623 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0; 1624 memset((char *)kasan_reset_tag(x) + s->inuse, 0, 1625 s->size - s->inuse - rsize); 1626 } 1627 /* KASAN might put x into memory quarantine, delaying its reuse. */ 1628 return kasan_slab_free(s, x, init); 1629 } 1630 1631 static inline bool slab_free_freelist_hook(struct kmem_cache *s, 1632 void **head, void **tail) 1633 { 1634 1635 void *object; 1636 void *next = *head; 1637 void *old_tail = *tail ? *tail : *head; 1638 1639 if (is_kfence_address(next)) { 1640 slab_free_hook(s, next, false); 1641 return true; 1642 } 1643 1644 /* Head and tail of the reconstructed freelist */ 1645 *head = NULL; 1646 *tail = NULL; 1647 1648 do { 1649 object = next; 1650 next = get_freepointer(s, object); 1651 1652 /* If object's reuse doesn't have to be delayed */ 1653 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) { 1654 /* Move object to the new freelist */ 1655 set_freepointer(s, object, *head); 1656 *head = object; 1657 if (!*tail) 1658 *tail = object; 1659 } 1660 } while (object != old_tail); 1661 1662 if (*head == *tail) 1663 *tail = NULL; 1664 1665 return *head != NULL; 1666 } 1667 1668 static void *setup_object(struct kmem_cache *s, struct page *page, 1669 void *object) 1670 { 1671 setup_object_debug(s, page, object); 1672 object = kasan_init_slab_obj(s, object); 1673 if (unlikely(s->ctor)) { 1674 kasan_unpoison_object_data(s, object); 1675 s->ctor(object); 1676 kasan_poison_object_data(s, object); 1677 } 1678 return object; 1679 } 1680 1681 /* 1682 * Slab allocation and freeing 1683 */ 1684 static inline struct page *alloc_slab_page(struct kmem_cache *s, 1685 gfp_t flags, int node, struct kmem_cache_order_objects oo) 1686 { 1687 struct page *page; 1688 unsigned int order = oo_order(oo); 1689 1690 if (node == NUMA_NO_NODE) 1691 page = alloc_pages(flags, order); 1692 else 1693 page = __alloc_pages_node(node, flags, order); 1694 1695 return page; 1696 } 1697 1698 #ifdef CONFIG_SLAB_FREELIST_RANDOM 1699 /* Pre-initialize the random sequence cache */ 1700 static int init_cache_random_seq(struct kmem_cache *s) 1701 { 1702 unsigned int count = oo_objects(s->oo); 1703 int err; 1704 1705 /* Bailout if already initialised */ 1706 if (s->random_seq) 1707 return 0; 1708 1709 err = cache_random_seq_create(s, count, GFP_KERNEL); 1710 if (err) { 1711 pr_err("SLUB: Unable to initialize free list for %s\n", 1712 s->name); 1713 return err; 1714 } 1715 1716 /* Transform to an offset on the set of pages */ 1717 if (s->random_seq) { 1718 unsigned int i; 1719 1720 for (i = 0; i < count; i++) 1721 s->random_seq[i] *= s->size; 1722 } 1723 return 0; 1724 } 1725 1726 /* Initialize each random sequence freelist per cache */ 1727 static void __init init_freelist_randomization(void) 1728 { 1729 struct kmem_cache *s; 1730 1731 mutex_lock(&slab_mutex); 1732 1733 list_for_each_entry(s, &slab_caches, list) 1734 init_cache_random_seq(s); 1735 1736 mutex_unlock(&slab_mutex); 1737 } 1738 1739 /* Get the next entry on the pre-computed freelist randomized */ 1740 static void *next_freelist_entry(struct kmem_cache *s, struct page *page, 1741 unsigned long *pos, void *start, 1742 unsigned long page_limit, 1743 unsigned long freelist_count) 1744 { 1745 unsigned int idx; 1746 1747 /* 1748 * If the target page allocation failed, the number of objects on the 1749 * page might be smaller than the usual size defined by the cache. 1750 */ 1751 do { 1752 idx = s->random_seq[*pos]; 1753 *pos += 1; 1754 if (*pos >= freelist_count) 1755 *pos = 0; 1756 } while (unlikely(idx >= page_limit)); 1757 1758 return (char *)start + idx; 1759 } 1760 1761 /* Shuffle the single linked freelist based on a random pre-computed sequence */ 1762 static bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1763 { 1764 void *start; 1765 void *cur; 1766 void *next; 1767 unsigned long idx, pos, page_limit, freelist_count; 1768 1769 if (page->objects < 2 || !s->random_seq) 1770 return false; 1771 1772 freelist_count = oo_objects(s->oo); 1773 pos = get_random_int() % freelist_count; 1774 1775 page_limit = page->objects * s->size; 1776 start = fixup_red_left(s, page_address(page)); 1777 1778 /* First entry is used as the base of the freelist */ 1779 cur = next_freelist_entry(s, page, &pos, start, page_limit, 1780 freelist_count); 1781 cur = setup_object(s, page, cur); 1782 page->freelist = cur; 1783 1784 for (idx = 1; idx < page->objects; idx++) { 1785 next = next_freelist_entry(s, page, &pos, start, page_limit, 1786 freelist_count); 1787 next = setup_object(s, page, next); 1788 set_freepointer(s, cur, next); 1789 cur = next; 1790 } 1791 set_freepointer(s, cur, NULL); 1792 1793 return true; 1794 } 1795 #else 1796 static inline int init_cache_random_seq(struct kmem_cache *s) 1797 { 1798 return 0; 1799 } 1800 static inline void init_freelist_randomization(void) { } 1801 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1802 { 1803 return false; 1804 } 1805 #endif /* CONFIG_SLAB_FREELIST_RANDOM */ 1806 1807 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1808 { 1809 struct page *page; 1810 struct kmem_cache_order_objects oo = s->oo; 1811 gfp_t alloc_gfp; 1812 void *start, *p, *next; 1813 int idx; 1814 bool shuffle; 1815 1816 flags &= gfp_allowed_mask; 1817 1818 if (gfpflags_allow_blocking(flags)) 1819 local_irq_enable(); 1820 1821 flags |= s->allocflags; 1822 1823 /* 1824 * Let the initial higher-order allocation fail under memory pressure 1825 * so we fall-back to the minimum order allocation. 1826 */ 1827 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1828 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min)) 1829 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL); 1830 1831 page = alloc_slab_page(s, alloc_gfp, node, oo); 1832 if (unlikely(!page)) { 1833 oo = s->min; 1834 alloc_gfp = flags; 1835 /* 1836 * Allocation may have failed due to fragmentation. 1837 * Try a lower order alloc if possible 1838 */ 1839 page = alloc_slab_page(s, alloc_gfp, node, oo); 1840 if (unlikely(!page)) 1841 goto out; 1842 stat(s, ORDER_FALLBACK); 1843 } 1844 1845 page->objects = oo_objects(oo); 1846 1847 account_slab_page(page, oo_order(oo), s, flags); 1848 1849 page->slab_cache = s; 1850 __SetPageSlab(page); 1851 if (page_is_pfmemalloc(page)) 1852 SetPageSlabPfmemalloc(page); 1853 1854 kasan_poison_slab(page); 1855 1856 start = page_address(page); 1857 1858 setup_page_debug(s, page, start); 1859 1860 shuffle = shuffle_freelist(s, page); 1861 1862 if (!shuffle) { 1863 start = fixup_red_left(s, start); 1864 start = setup_object(s, page, start); 1865 page->freelist = start; 1866 for (idx = 0, p = start; idx < page->objects - 1; idx++) { 1867 next = p + s->size; 1868 next = setup_object(s, page, next); 1869 set_freepointer(s, p, next); 1870 p = next; 1871 } 1872 set_freepointer(s, p, NULL); 1873 } 1874 1875 page->inuse = page->objects; 1876 page->frozen = 1; 1877 1878 out: 1879 if (gfpflags_allow_blocking(flags)) 1880 local_irq_disable(); 1881 if (!page) 1882 return NULL; 1883 1884 inc_slabs_node(s, page_to_nid(page), page->objects); 1885 1886 return page; 1887 } 1888 1889 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1890 { 1891 if (unlikely(flags & GFP_SLAB_BUG_MASK)) 1892 flags = kmalloc_fix_flags(flags); 1893 1894 return allocate_slab(s, 1895 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1896 } 1897 1898 static void __free_slab(struct kmem_cache *s, struct page *page) 1899 { 1900 int order = compound_order(page); 1901 int pages = 1 << order; 1902 1903 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) { 1904 void *p; 1905 1906 slab_pad_check(s, page); 1907 for_each_object(p, s, page_address(page), 1908 page->objects) 1909 check_object(s, page, p, SLUB_RED_INACTIVE); 1910 } 1911 1912 __ClearPageSlabPfmemalloc(page); 1913 __ClearPageSlab(page); 1914 /* In union with page->mapping where page allocator expects NULL */ 1915 page->slab_cache = NULL; 1916 if (current->reclaim_state) 1917 current->reclaim_state->reclaimed_slab += pages; 1918 unaccount_slab_page(page, order, s); 1919 __free_pages(page, order); 1920 } 1921 1922 static void rcu_free_slab(struct rcu_head *h) 1923 { 1924 struct page *page = container_of(h, struct page, rcu_head); 1925 1926 __free_slab(page->slab_cache, page); 1927 } 1928 1929 static void free_slab(struct kmem_cache *s, struct page *page) 1930 { 1931 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) { 1932 call_rcu(&page->rcu_head, rcu_free_slab); 1933 } else 1934 __free_slab(s, page); 1935 } 1936 1937 static void discard_slab(struct kmem_cache *s, struct page *page) 1938 { 1939 dec_slabs_node(s, page_to_nid(page), page->objects); 1940 free_slab(s, page); 1941 } 1942 1943 /* 1944 * Management of partially allocated slabs. 1945 */ 1946 static inline void 1947 __add_partial(struct kmem_cache_node *n, struct page *page, int tail) 1948 { 1949 n->nr_partial++; 1950 if (tail == DEACTIVATE_TO_TAIL) 1951 list_add_tail(&page->slab_list, &n->partial); 1952 else 1953 list_add(&page->slab_list, &n->partial); 1954 } 1955 1956 static inline void add_partial(struct kmem_cache_node *n, 1957 struct page *page, int tail) 1958 { 1959 lockdep_assert_held(&n->list_lock); 1960 __add_partial(n, page, tail); 1961 } 1962 1963 static inline void remove_partial(struct kmem_cache_node *n, 1964 struct page *page) 1965 { 1966 lockdep_assert_held(&n->list_lock); 1967 list_del(&page->slab_list); 1968 n->nr_partial--; 1969 } 1970 1971 /* 1972 * Remove slab from the partial list, freeze it and 1973 * return the pointer to the freelist. 1974 * 1975 * Returns a list of objects or NULL if it fails. 1976 */ 1977 static inline void *acquire_slab(struct kmem_cache *s, 1978 struct kmem_cache_node *n, struct page *page, 1979 int mode, int *objects) 1980 { 1981 void *freelist; 1982 unsigned long counters; 1983 struct page new; 1984 1985 lockdep_assert_held(&n->list_lock); 1986 1987 /* 1988 * Zap the freelist and set the frozen bit. 1989 * The old freelist is the list of objects for the 1990 * per cpu allocation list. 1991 */ 1992 freelist = page->freelist; 1993 counters = page->counters; 1994 new.counters = counters; 1995 *objects = new.objects - new.inuse; 1996 if (mode) { 1997 new.inuse = page->objects; 1998 new.freelist = NULL; 1999 } else { 2000 new.freelist = freelist; 2001 } 2002 2003 VM_BUG_ON(new.frozen); 2004 new.frozen = 1; 2005 2006 if (!__cmpxchg_double_slab(s, page, 2007 freelist, counters, 2008 new.freelist, new.counters, 2009 "acquire_slab")) 2010 return NULL; 2011 2012 remove_partial(n, page); 2013 WARN_ON(!freelist); 2014 return freelist; 2015 } 2016 2017 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); 2018 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags); 2019 2020 /* 2021 * Try to allocate a partial slab from a specific node. 2022 */ 2023 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, 2024 struct kmem_cache_cpu *c, gfp_t flags) 2025 { 2026 struct page *page, *page2; 2027 void *object = NULL; 2028 unsigned int available = 0; 2029 int objects; 2030 2031 /* 2032 * Racy check. If we mistakenly see no partial slabs then we 2033 * just allocate an empty slab. If we mistakenly try to get a 2034 * partial slab and there is none available then get_partial() 2035 * will return NULL. 2036 */ 2037 if (!n || !n->nr_partial) 2038 return NULL; 2039 2040 spin_lock(&n->list_lock); 2041 list_for_each_entry_safe(page, page2, &n->partial, slab_list) { 2042 void *t; 2043 2044 if (!pfmemalloc_match(page, flags)) 2045 continue; 2046 2047 t = acquire_slab(s, n, page, object == NULL, &objects); 2048 if (!t) 2049 break; 2050 2051 available += objects; 2052 if (!object) { 2053 c->page = page; 2054 stat(s, ALLOC_FROM_PARTIAL); 2055 object = t; 2056 } else { 2057 put_cpu_partial(s, page, 0); 2058 stat(s, CPU_PARTIAL_NODE); 2059 } 2060 if (!kmem_cache_has_cpu_partial(s) 2061 || available > slub_cpu_partial(s) / 2) 2062 break; 2063 2064 } 2065 spin_unlock(&n->list_lock); 2066 return object; 2067 } 2068 2069 /* 2070 * Get a page from somewhere. Search in increasing NUMA distances. 2071 */ 2072 static void *get_any_partial(struct kmem_cache *s, gfp_t flags, 2073 struct kmem_cache_cpu *c) 2074 { 2075 #ifdef CONFIG_NUMA 2076 struct zonelist *zonelist; 2077 struct zoneref *z; 2078 struct zone *zone; 2079 enum zone_type highest_zoneidx = gfp_zone(flags); 2080 void *object; 2081 unsigned int cpuset_mems_cookie; 2082 2083 /* 2084 * The defrag ratio allows a configuration of the tradeoffs between 2085 * inter node defragmentation and node local allocations. A lower 2086 * defrag_ratio increases the tendency to do local allocations 2087 * instead of attempting to obtain partial slabs from other nodes. 2088 * 2089 * If the defrag_ratio is set to 0 then kmalloc() always 2090 * returns node local objects. If the ratio is higher then kmalloc() 2091 * may return off node objects because partial slabs are obtained 2092 * from other nodes and filled up. 2093 * 2094 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 2095 * (which makes defrag_ratio = 1000) then every (well almost) 2096 * allocation will first attempt to defrag slab caches on other nodes. 2097 * This means scanning over all nodes to look for partial slabs which 2098 * may be expensive if we do it every time we are trying to find a slab 2099 * with available objects. 2100 */ 2101 if (!s->remote_node_defrag_ratio || 2102 get_cycles() % 1024 > s->remote_node_defrag_ratio) 2103 return NULL; 2104 2105 do { 2106 cpuset_mems_cookie = read_mems_allowed_begin(); 2107 zonelist = node_zonelist(mempolicy_slab_node(), flags); 2108 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) { 2109 struct kmem_cache_node *n; 2110 2111 n = get_node(s, zone_to_nid(zone)); 2112 2113 if (n && cpuset_zone_allowed(zone, flags) && 2114 n->nr_partial > s->min_partial) { 2115 object = get_partial_node(s, n, c, flags); 2116 if (object) { 2117 /* 2118 * Don't check read_mems_allowed_retry() 2119 * here - if mems_allowed was updated in 2120 * parallel, that was a harmless race 2121 * between allocation and the cpuset 2122 * update 2123 */ 2124 return object; 2125 } 2126 } 2127 } 2128 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 2129 #endif /* CONFIG_NUMA */ 2130 return NULL; 2131 } 2132 2133 /* 2134 * Get a partial page, lock it and return it. 2135 */ 2136 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, 2137 struct kmem_cache_cpu *c) 2138 { 2139 void *object; 2140 int searchnode = node; 2141 2142 if (node == NUMA_NO_NODE) 2143 searchnode = numa_mem_id(); 2144 2145 object = get_partial_node(s, get_node(s, searchnode), c, flags); 2146 if (object || node != NUMA_NO_NODE) 2147 return object; 2148 2149 return get_any_partial(s, flags, c); 2150 } 2151 2152 #ifdef CONFIG_PREEMPTION 2153 /* 2154 * Calculate the next globally unique transaction for disambiguation 2155 * during cmpxchg. The transactions start with the cpu number and are then 2156 * incremented by CONFIG_NR_CPUS. 2157 */ 2158 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 2159 #else 2160 /* 2161 * No preemption supported therefore also no need to check for 2162 * different cpus. 2163 */ 2164 #define TID_STEP 1 2165 #endif 2166 2167 static inline unsigned long next_tid(unsigned long tid) 2168 { 2169 return tid + TID_STEP; 2170 } 2171 2172 #ifdef SLUB_DEBUG_CMPXCHG 2173 static inline unsigned int tid_to_cpu(unsigned long tid) 2174 { 2175 return tid % TID_STEP; 2176 } 2177 2178 static inline unsigned long tid_to_event(unsigned long tid) 2179 { 2180 return tid / TID_STEP; 2181 } 2182 #endif 2183 2184 static inline unsigned int init_tid(int cpu) 2185 { 2186 return cpu; 2187 } 2188 2189 static inline void note_cmpxchg_failure(const char *n, 2190 const struct kmem_cache *s, unsigned long tid) 2191 { 2192 #ifdef SLUB_DEBUG_CMPXCHG 2193 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 2194 2195 pr_info("%s %s: cmpxchg redo ", n, s->name); 2196 2197 #ifdef CONFIG_PREEMPTION 2198 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 2199 pr_warn("due to cpu change %d -> %d\n", 2200 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 2201 else 2202 #endif 2203 if (tid_to_event(tid) != tid_to_event(actual_tid)) 2204 pr_warn("due to cpu running other code. Event %ld->%ld\n", 2205 tid_to_event(tid), tid_to_event(actual_tid)); 2206 else 2207 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n", 2208 actual_tid, tid, next_tid(tid)); 2209 #endif 2210 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 2211 } 2212 2213 static void init_kmem_cache_cpus(struct kmem_cache *s) 2214 { 2215 int cpu; 2216 2217 for_each_possible_cpu(cpu) 2218 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 2219 } 2220 2221 /* 2222 * Remove the cpu slab 2223 */ 2224 static void deactivate_slab(struct kmem_cache *s, struct page *page, 2225 void *freelist, struct kmem_cache_cpu *c) 2226 { 2227 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; 2228 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 2229 int lock = 0, free_delta = 0; 2230 enum slab_modes l = M_NONE, m = M_NONE; 2231 void *nextfree, *freelist_iter, *freelist_tail; 2232 int tail = DEACTIVATE_TO_HEAD; 2233 struct page new; 2234 struct page old; 2235 2236 if (page->freelist) { 2237 stat(s, DEACTIVATE_REMOTE_FREES); 2238 tail = DEACTIVATE_TO_TAIL; 2239 } 2240 2241 /* 2242 * Stage one: Count the objects on cpu's freelist as free_delta and 2243 * remember the last object in freelist_tail for later splicing. 2244 */ 2245 freelist_tail = NULL; 2246 freelist_iter = freelist; 2247 while (freelist_iter) { 2248 nextfree = get_freepointer(s, freelist_iter); 2249 2250 /* 2251 * If 'nextfree' is invalid, it is possible that the object at 2252 * 'freelist_iter' is already corrupted. So isolate all objects 2253 * starting at 'freelist_iter' by skipping them. 2254 */ 2255 if (freelist_corrupted(s, page, &freelist_iter, nextfree)) 2256 break; 2257 2258 freelist_tail = freelist_iter; 2259 free_delta++; 2260 2261 freelist_iter = nextfree; 2262 } 2263 2264 /* 2265 * Stage two: Unfreeze the page while splicing the per-cpu 2266 * freelist to the head of page's freelist. 2267 * 2268 * Ensure that the page is unfrozen while the list presence 2269 * reflects the actual number of objects during unfreeze. 2270 * 2271 * We setup the list membership and then perform a cmpxchg 2272 * with the count. If there is a mismatch then the page 2273 * is not unfrozen but the page is on the wrong list. 2274 * 2275 * Then we restart the process which may have to remove 2276 * the page from the list that we just put it on again 2277 * because the number of objects in the slab may have 2278 * changed. 2279 */ 2280 redo: 2281 2282 old.freelist = READ_ONCE(page->freelist); 2283 old.counters = READ_ONCE(page->counters); 2284 VM_BUG_ON(!old.frozen); 2285 2286 /* Determine target state of the slab */ 2287 new.counters = old.counters; 2288 if (freelist_tail) { 2289 new.inuse -= free_delta; 2290 set_freepointer(s, freelist_tail, old.freelist); 2291 new.freelist = freelist; 2292 } else 2293 new.freelist = old.freelist; 2294 2295 new.frozen = 0; 2296 2297 if (!new.inuse && n->nr_partial >= s->min_partial) 2298 m = M_FREE; 2299 else if (new.freelist) { 2300 m = M_PARTIAL; 2301 if (!lock) { 2302 lock = 1; 2303 /* 2304 * Taking the spinlock removes the possibility 2305 * that acquire_slab() will see a slab page that 2306 * is frozen 2307 */ 2308 spin_lock(&n->list_lock); 2309 } 2310 } else { 2311 m = M_FULL; 2312 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) { 2313 lock = 1; 2314 /* 2315 * This also ensures that the scanning of full 2316 * slabs from diagnostic functions will not see 2317 * any frozen slabs. 2318 */ 2319 spin_lock(&n->list_lock); 2320 } 2321 } 2322 2323 if (l != m) { 2324 if (l == M_PARTIAL) 2325 remove_partial(n, page); 2326 else if (l == M_FULL) 2327 remove_full(s, n, page); 2328 2329 if (m == M_PARTIAL) 2330 add_partial(n, page, tail); 2331 else if (m == M_FULL) 2332 add_full(s, n, page); 2333 } 2334 2335 l = m; 2336 if (!__cmpxchg_double_slab(s, page, 2337 old.freelist, old.counters, 2338 new.freelist, new.counters, 2339 "unfreezing slab")) 2340 goto redo; 2341 2342 if (lock) 2343 spin_unlock(&n->list_lock); 2344 2345 if (m == M_PARTIAL) 2346 stat(s, tail); 2347 else if (m == M_FULL) 2348 stat(s, DEACTIVATE_FULL); 2349 else if (m == M_FREE) { 2350 stat(s, DEACTIVATE_EMPTY); 2351 discard_slab(s, page); 2352 stat(s, FREE_SLAB); 2353 } 2354 2355 c->page = NULL; 2356 c->freelist = NULL; 2357 } 2358 2359 /* 2360 * Unfreeze all the cpu partial slabs. 2361 * 2362 * This function must be called with interrupts disabled 2363 * for the cpu using c (or some other guarantee must be there 2364 * to guarantee no concurrent accesses). 2365 */ 2366 static void unfreeze_partials(struct kmem_cache *s, 2367 struct kmem_cache_cpu *c) 2368 { 2369 #ifdef CONFIG_SLUB_CPU_PARTIAL 2370 struct kmem_cache_node *n = NULL, *n2 = NULL; 2371 struct page *page, *discard_page = NULL; 2372 2373 while ((page = slub_percpu_partial(c))) { 2374 struct page new; 2375 struct page old; 2376 2377 slub_set_percpu_partial(c, page); 2378 2379 n2 = get_node(s, page_to_nid(page)); 2380 if (n != n2) { 2381 if (n) 2382 spin_unlock(&n->list_lock); 2383 2384 n = n2; 2385 spin_lock(&n->list_lock); 2386 } 2387 2388 do { 2389 2390 old.freelist = page->freelist; 2391 old.counters = page->counters; 2392 VM_BUG_ON(!old.frozen); 2393 2394 new.counters = old.counters; 2395 new.freelist = old.freelist; 2396 2397 new.frozen = 0; 2398 2399 } while (!__cmpxchg_double_slab(s, page, 2400 old.freelist, old.counters, 2401 new.freelist, new.counters, 2402 "unfreezing slab")); 2403 2404 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { 2405 page->next = discard_page; 2406 discard_page = page; 2407 } else { 2408 add_partial(n, page, DEACTIVATE_TO_TAIL); 2409 stat(s, FREE_ADD_PARTIAL); 2410 } 2411 } 2412 2413 if (n) 2414 spin_unlock(&n->list_lock); 2415 2416 while (discard_page) { 2417 page = discard_page; 2418 discard_page = discard_page->next; 2419 2420 stat(s, DEACTIVATE_EMPTY); 2421 discard_slab(s, page); 2422 stat(s, FREE_SLAB); 2423 } 2424 #endif /* CONFIG_SLUB_CPU_PARTIAL */ 2425 } 2426 2427 /* 2428 * Put a page that was just frozen (in __slab_free|get_partial_node) into a 2429 * partial page slot if available. 2430 * 2431 * If we did not find a slot then simply move all the partials to the 2432 * per node partial list. 2433 */ 2434 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) 2435 { 2436 #ifdef CONFIG_SLUB_CPU_PARTIAL 2437 struct page *oldpage; 2438 int pages; 2439 int pobjects; 2440 2441 preempt_disable(); 2442 do { 2443 pages = 0; 2444 pobjects = 0; 2445 oldpage = this_cpu_read(s->cpu_slab->partial); 2446 2447 if (oldpage) { 2448 pobjects = oldpage->pobjects; 2449 pages = oldpage->pages; 2450 if (drain && pobjects > slub_cpu_partial(s)) { 2451 unsigned long flags; 2452 /* 2453 * partial array is full. Move the existing 2454 * set to the per node partial list. 2455 */ 2456 local_irq_save(flags); 2457 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2458 local_irq_restore(flags); 2459 oldpage = NULL; 2460 pobjects = 0; 2461 pages = 0; 2462 stat(s, CPU_PARTIAL_DRAIN); 2463 } 2464 } 2465 2466 pages++; 2467 pobjects += page->objects - page->inuse; 2468 2469 page->pages = pages; 2470 page->pobjects = pobjects; 2471 page->next = oldpage; 2472 2473 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) 2474 != oldpage); 2475 if (unlikely(!slub_cpu_partial(s))) { 2476 unsigned long flags; 2477 2478 local_irq_save(flags); 2479 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2480 local_irq_restore(flags); 2481 } 2482 preempt_enable(); 2483 #endif /* CONFIG_SLUB_CPU_PARTIAL */ 2484 } 2485 2486 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 2487 { 2488 stat(s, CPUSLAB_FLUSH); 2489 deactivate_slab(s, c->page, c->freelist, c); 2490 2491 c->tid = next_tid(c->tid); 2492 } 2493 2494 /* 2495 * Flush cpu slab. 2496 * 2497 * Called from IPI handler with interrupts disabled. 2498 */ 2499 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 2500 { 2501 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2502 2503 if (c->page) 2504 flush_slab(s, c); 2505 2506 unfreeze_partials(s, c); 2507 } 2508 2509 static void flush_cpu_slab(void *d) 2510 { 2511 struct kmem_cache *s = d; 2512 2513 __flush_cpu_slab(s, smp_processor_id()); 2514 } 2515 2516 static bool has_cpu_slab(int cpu, void *info) 2517 { 2518 struct kmem_cache *s = info; 2519 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2520 2521 return c->page || slub_percpu_partial(c); 2522 } 2523 2524 static void flush_all(struct kmem_cache *s) 2525 { 2526 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1); 2527 } 2528 2529 /* 2530 * Use the cpu notifier to insure that the cpu slabs are flushed when 2531 * necessary. 2532 */ 2533 static int slub_cpu_dead(unsigned int cpu) 2534 { 2535 struct kmem_cache *s; 2536 unsigned long flags; 2537 2538 mutex_lock(&slab_mutex); 2539 list_for_each_entry(s, &slab_caches, list) { 2540 local_irq_save(flags); 2541 __flush_cpu_slab(s, cpu); 2542 local_irq_restore(flags); 2543 } 2544 mutex_unlock(&slab_mutex); 2545 return 0; 2546 } 2547 2548 /* 2549 * Check if the objects in a per cpu structure fit numa 2550 * locality expectations. 2551 */ 2552 static inline int node_match(struct page *page, int node) 2553 { 2554 #ifdef CONFIG_NUMA 2555 if (node != NUMA_NO_NODE && page_to_nid(page) != node) 2556 return 0; 2557 #endif 2558 return 1; 2559 } 2560 2561 #ifdef CONFIG_SLUB_DEBUG 2562 static int count_free(struct page *page) 2563 { 2564 return page->objects - page->inuse; 2565 } 2566 2567 static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 2568 { 2569 return atomic_long_read(&n->total_objects); 2570 } 2571 #endif /* CONFIG_SLUB_DEBUG */ 2572 2573 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS) 2574 static unsigned long count_partial(struct kmem_cache_node *n, 2575 int (*get_count)(struct page *)) 2576 { 2577 unsigned long flags; 2578 unsigned long x = 0; 2579 struct page *page; 2580 2581 spin_lock_irqsave(&n->list_lock, flags); 2582 list_for_each_entry(page, &n->partial, slab_list) 2583 x += get_count(page); 2584 spin_unlock_irqrestore(&n->list_lock, flags); 2585 return x; 2586 } 2587 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */ 2588 2589 static noinline void 2590 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 2591 { 2592 #ifdef CONFIG_SLUB_DEBUG 2593 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, 2594 DEFAULT_RATELIMIT_BURST); 2595 int node; 2596 struct kmem_cache_node *n; 2597 2598 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) 2599 return; 2600 2601 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", 2602 nid, gfpflags, &gfpflags); 2603 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n", 2604 s->name, s->object_size, s->size, oo_order(s->oo), 2605 oo_order(s->min)); 2606 2607 if (oo_order(s->min) > get_order(s->object_size)) 2608 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n", 2609 s->name); 2610 2611 for_each_kmem_cache_node(s, node, n) { 2612 unsigned long nr_slabs; 2613 unsigned long nr_objs; 2614 unsigned long nr_free; 2615 2616 nr_free = count_partial(n, count_free); 2617 nr_slabs = node_nr_slabs(n); 2618 nr_objs = node_nr_objs(n); 2619 2620 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n", 2621 node, nr_slabs, nr_objs, nr_free); 2622 } 2623 #endif 2624 } 2625 2626 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, 2627 int node, struct kmem_cache_cpu **pc) 2628 { 2629 void *freelist; 2630 struct kmem_cache_cpu *c = *pc; 2631 struct page *page; 2632 2633 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO)); 2634 2635 freelist = get_partial(s, flags, node, c); 2636 2637 if (freelist) 2638 return freelist; 2639 2640 page = new_slab(s, flags, node); 2641 if (page) { 2642 c = raw_cpu_ptr(s->cpu_slab); 2643 if (c->page) 2644 flush_slab(s, c); 2645 2646 /* 2647 * No other reference to the page yet so we can 2648 * muck around with it freely without cmpxchg 2649 */ 2650 freelist = page->freelist; 2651 page->freelist = NULL; 2652 2653 stat(s, ALLOC_SLAB); 2654 c->page = page; 2655 *pc = c; 2656 } 2657 2658 return freelist; 2659 } 2660 2661 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags) 2662 { 2663 if (unlikely(PageSlabPfmemalloc(page))) 2664 return gfp_pfmemalloc_allowed(gfpflags); 2665 2666 return true; 2667 } 2668 2669 /* 2670 * Check the page->freelist of a page and either transfer the freelist to the 2671 * per cpu freelist or deactivate the page. 2672 * 2673 * The page is still frozen if the return value is not NULL. 2674 * 2675 * If this function returns NULL then the page has been unfrozen. 2676 * 2677 * This function must be called with interrupt disabled. 2678 */ 2679 static inline void *get_freelist(struct kmem_cache *s, struct page *page) 2680 { 2681 struct page new; 2682 unsigned long counters; 2683 void *freelist; 2684 2685 do { 2686 freelist = page->freelist; 2687 counters = page->counters; 2688 2689 new.counters = counters; 2690 VM_BUG_ON(!new.frozen); 2691 2692 new.inuse = page->objects; 2693 new.frozen = freelist != NULL; 2694 2695 } while (!__cmpxchg_double_slab(s, page, 2696 freelist, counters, 2697 NULL, new.counters, 2698 "get_freelist")); 2699 2700 return freelist; 2701 } 2702 2703 /* 2704 * Slow path. The lockless freelist is empty or we need to perform 2705 * debugging duties. 2706 * 2707 * Processing is still very fast if new objects have been freed to the 2708 * regular freelist. In that case we simply take over the regular freelist 2709 * as the lockless freelist and zap the regular freelist. 2710 * 2711 * If that is not working then we fall back to the partial lists. We take the 2712 * first element of the freelist as the object to allocate now and move the 2713 * rest of the freelist to the lockless freelist. 2714 * 2715 * And if we were unable to get a new slab from the partial slab lists then 2716 * we need to allocate a new slab. This is the slowest path since it involves 2717 * a call to the page allocator and the setup of a new slab. 2718 * 2719 * Version of __slab_alloc to use when we know that interrupts are 2720 * already disabled (which is the case for bulk allocation). 2721 */ 2722 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2723 unsigned long addr, struct kmem_cache_cpu *c) 2724 { 2725 void *freelist; 2726 struct page *page; 2727 2728 stat(s, ALLOC_SLOWPATH); 2729 2730 page = c->page; 2731 if (!page) { 2732 /* 2733 * if the node is not online or has no normal memory, just 2734 * ignore the node constraint 2735 */ 2736 if (unlikely(node != NUMA_NO_NODE && 2737 !node_isset(node, slab_nodes))) 2738 node = NUMA_NO_NODE; 2739 goto new_slab; 2740 } 2741 redo: 2742 2743 if (unlikely(!node_match(page, node))) { 2744 /* 2745 * same as above but node_match() being false already 2746 * implies node != NUMA_NO_NODE 2747 */ 2748 if (!node_isset(node, slab_nodes)) { 2749 node = NUMA_NO_NODE; 2750 goto redo; 2751 } else { 2752 stat(s, ALLOC_NODE_MISMATCH); 2753 deactivate_slab(s, page, c->freelist, c); 2754 goto new_slab; 2755 } 2756 } 2757 2758 /* 2759 * By rights, we should be searching for a slab page that was 2760 * PFMEMALLOC but right now, we are losing the pfmemalloc 2761 * information when the page leaves the per-cpu allocator 2762 */ 2763 if (unlikely(!pfmemalloc_match(page, gfpflags))) { 2764 deactivate_slab(s, page, c->freelist, c); 2765 goto new_slab; 2766 } 2767 2768 /* must check again c->freelist in case of cpu migration or IRQ */ 2769 freelist = c->freelist; 2770 if (freelist) 2771 goto load_freelist; 2772 2773 freelist = get_freelist(s, page); 2774 2775 if (!freelist) { 2776 c->page = NULL; 2777 stat(s, DEACTIVATE_BYPASS); 2778 goto new_slab; 2779 } 2780 2781 stat(s, ALLOC_REFILL); 2782 2783 load_freelist: 2784 /* 2785 * freelist is pointing to the list of objects to be used. 2786 * page is pointing to the page from which the objects are obtained. 2787 * That page must be frozen for per cpu allocations to work. 2788 */ 2789 VM_BUG_ON(!c->page->frozen); 2790 c->freelist = get_freepointer(s, freelist); 2791 c->tid = next_tid(c->tid); 2792 return freelist; 2793 2794 new_slab: 2795 2796 if (slub_percpu_partial(c)) { 2797 page = c->page = slub_percpu_partial(c); 2798 slub_set_percpu_partial(c, page); 2799 stat(s, CPU_PARTIAL_ALLOC); 2800 goto redo; 2801 } 2802 2803 freelist = new_slab_objects(s, gfpflags, node, &c); 2804 2805 if (unlikely(!freelist)) { 2806 slab_out_of_memory(s, gfpflags, node); 2807 return NULL; 2808 } 2809 2810 page = c->page; 2811 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags))) 2812 goto load_freelist; 2813 2814 /* Only entered in the debug case */ 2815 if (kmem_cache_debug(s) && 2816 !alloc_debug_processing(s, page, freelist, addr)) 2817 goto new_slab; /* Slab failed checks. Next slab needed */ 2818 2819 deactivate_slab(s, page, get_freepointer(s, freelist), c); 2820 return freelist; 2821 } 2822 2823 /* 2824 * Another one that disabled interrupt and compensates for possible 2825 * cpu changes by refetching the per cpu area pointer. 2826 */ 2827 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2828 unsigned long addr, struct kmem_cache_cpu *c) 2829 { 2830 void *p; 2831 unsigned long flags; 2832 2833 local_irq_save(flags); 2834 #ifdef CONFIG_PREEMPTION 2835 /* 2836 * We may have been preempted and rescheduled on a different 2837 * cpu before disabling interrupts. Need to reload cpu area 2838 * pointer. 2839 */ 2840 c = this_cpu_ptr(s->cpu_slab); 2841 #endif 2842 2843 p = ___slab_alloc(s, gfpflags, node, addr, c); 2844 local_irq_restore(flags); 2845 return p; 2846 } 2847 2848 /* 2849 * If the object has been wiped upon free, make sure it's fully initialized by 2850 * zeroing out freelist pointer. 2851 */ 2852 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s, 2853 void *obj) 2854 { 2855 if (unlikely(slab_want_init_on_free(s)) && obj) 2856 memset((void *)((char *)kasan_reset_tag(obj) + s->offset), 2857 0, sizeof(void *)); 2858 } 2859 2860 /* 2861 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 2862 * have the fastpath folded into their functions. So no function call 2863 * overhead for requests that can be satisfied on the fastpath. 2864 * 2865 * The fastpath works by first checking if the lockless freelist can be used. 2866 * If not then __slab_alloc is called for slow processing. 2867 * 2868 * Otherwise we can simply pick the next object from the lockless free list. 2869 */ 2870 static __always_inline void *slab_alloc_node(struct kmem_cache *s, 2871 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) 2872 { 2873 void *object; 2874 struct kmem_cache_cpu *c; 2875 struct page *page; 2876 unsigned long tid; 2877 struct obj_cgroup *objcg = NULL; 2878 bool init = false; 2879 2880 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags); 2881 if (!s) 2882 return NULL; 2883 2884 object = kfence_alloc(s, orig_size, gfpflags); 2885 if (unlikely(object)) 2886 goto out; 2887 2888 redo: 2889 /* 2890 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 2891 * enabled. We may switch back and forth between cpus while 2892 * reading from one cpu area. That does not matter as long 2893 * as we end up on the original cpu again when doing the cmpxchg. 2894 * 2895 * We should guarantee that tid and kmem_cache are retrieved on 2896 * the same cpu. It could be different if CONFIG_PREEMPTION so we need 2897 * to check if it is matched or not. 2898 */ 2899 do { 2900 tid = this_cpu_read(s->cpu_slab->tid); 2901 c = raw_cpu_ptr(s->cpu_slab); 2902 } while (IS_ENABLED(CONFIG_PREEMPTION) && 2903 unlikely(tid != READ_ONCE(c->tid))); 2904 2905 /* 2906 * Irqless object alloc/free algorithm used here depends on sequence 2907 * of fetching cpu_slab's data. tid should be fetched before anything 2908 * on c to guarantee that object and page associated with previous tid 2909 * won't be used with current tid. If we fetch tid first, object and 2910 * page could be one associated with next tid and our alloc/free 2911 * request will be failed. In this case, we will retry. So, no problem. 2912 */ 2913 barrier(); 2914 2915 /* 2916 * The transaction ids are globally unique per cpu and per operation on 2917 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 2918 * occurs on the right processor and that there was no operation on the 2919 * linked list in between. 2920 */ 2921 2922 object = c->freelist; 2923 page = c->page; 2924 if (unlikely(!object || !page || !node_match(page, node))) { 2925 object = __slab_alloc(s, gfpflags, node, addr, c); 2926 } else { 2927 void *next_object = get_freepointer_safe(s, object); 2928 2929 /* 2930 * The cmpxchg will only match if there was no additional 2931 * operation and if we are on the right processor. 2932 * 2933 * The cmpxchg does the following atomically (without lock 2934 * semantics!) 2935 * 1. Relocate first pointer to the current per cpu area. 2936 * 2. Verify that tid and freelist have not been changed 2937 * 3. If they were not changed replace tid and freelist 2938 * 2939 * Since this is without lock semantics the protection is only 2940 * against code executing on this cpu *not* from access by 2941 * other cpus. 2942 */ 2943 if (unlikely(!this_cpu_cmpxchg_double( 2944 s->cpu_slab->freelist, s->cpu_slab->tid, 2945 object, tid, 2946 next_object, next_tid(tid)))) { 2947 2948 note_cmpxchg_failure("slab_alloc", s, tid); 2949 goto redo; 2950 } 2951 prefetch_freepointer(s, next_object); 2952 stat(s, ALLOC_FASTPATH); 2953 } 2954 2955 maybe_wipe_obj_freeptr(s, object); 2956 init = slab_want_init_on_alloc(gfpflags, s); 2957 2958 out: 2959 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init); 2960 2961 return object; 2962 } 2963 2964 static __always_inline void *slab_alloc(struct kmem_cache *s, 2965 gfp_t gfpflags, unsigned long addr, size_t orig_size) 2966 { 2967 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size); 2968 } 2969 2970 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 2971 { 2972 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size); 2973 2974 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, 2975 s->size, gfpflags); 2976 2977 return ret; 2978 } 2979 EXPORT_SYMBOL(kmem_cache_alloc); 2980 2981 #ifdef CONFIG_TRACING 2982 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2983 { 2984 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size); 2985 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2986 ret = kasan_kmalloc(s, ret, size, gfpflags); 2987 return ret; 2988 } 2989 EXPORT_SYMBOL(kmem_cache_alloc_trace); 2990 #endif 2991 2992 #ifdef CONFIG_NUMA 2993 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2994 { 2995 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size); 2996 2997 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2998 s->object_size, s->size, gfpflags, node); 2999 3000 return ret; 3001 } 3002 EXPORT_SYMBOL(kmem_cache_alloc_node); 3003 3004 #ifdef CONFIG_TRACING 3005 void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 3006 gfp_t gfpflags, 3007 int node, size_t size) 3008 { 3009 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size); 3010 3011 trace_kmalloc_node(_RET_IP_, ret, 3012 size, s->size, gfpflags, node); 3013 3014 ret = kasan_kmalloc(s, ret, size, gfpflags); 3015 return ret; 3016 } 3017 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 3018 #endif 3019 #endif /* CONFIG_NUMA */ 3020 3021 /* 3022 * Slow path handling. This may still be called frequently since objects 3023 * have a longer lifetime than the cpu slabs in most processing loads. 3024 * 3025 * So we still attempt to reduce cache line usage. Just take the slab 3026 * lock and free the item. If there is no additional partial page 3027 * handling required then we can return immediately. 3028 */ 3029 static void __slab_free(struct kmem_cache *s, struct page *page, 3030 void *head, void *tail, int cnt, 3031 unsigned long addr) 3032 3033 { 3034 void *prior; 3035 int was_frozen; 3036 struct page new; 3037 unsigned long counters; 3038 struct kmem_cache_node *n = NULL; 3039 unsigned long flags; 3040 3041 stat(s, FREE_SLOWPATH); 3042 3043 if (kfence_free(head)) 3044 return; 3045 3046 if (kmem_cache_debug(s) && 3047 !free_debug_processing(s, page, head, tail, cnt, addr)) 3048 return; 3049 3050 do { 3051 if (unlikely(n)) { 3052 spin_unlock_irqrestore(&n->list_lock, flags); 3053 n = NULL; 3054 } 3055 prior = page->freelist; 3056 counters = page->counters; 3057 set_freepointer(s, tail, prior); 3058 new.counters = counters; 3059 was_frozen = new.frozen; 3060 new.inuse -= cnt; 3061 if ((!new.inuse || !prior) && !was_frozen) { 3062 3063 if (kmem_cache_has_cpu_partial(s) && !prior) { 3064 3065 /* 3066 * Slab was on no list before and will be 3067 * partially empty 3068 * We can defer the list move and instead 3069 * freeze it. 3070 */ 3071 new.frozen = 1; 3072 3073 } else { /* Needs to be taken off a list */ 3074 3075 n = get_node(s, page_to_nid(page)); 3076 /* 3077 * Speculatively acquire the list_lock. 3078 * If the cmpxchg does not succeed then we may 3079 * drop the list_lock without any processing. 3080 * 3081 * Otherwise the list_lock will synchronize with 3082 * other processors updating the list of slabs. 3083 */ 3084 spin_lock_irqsave(&n->list_lock, flags); 3085 3086 } 3087 } 3088 3089 } while (!cmpxchg_double_slab(s, page, 3090 prior, counters, 3091 head, new.counters, 3092 "__slab_free")); 3093 3094 if (likely(!n)) { 3095 3096 if (likely(was_frozen)) { 3097 /* 3098 * The list lock was not taken therefore no list 3099 * activity can be necessary. 3100 */ 3101 stat(s, FREE_FROZEN); 3102 } else if (new.frozen) { 3103 /* 3104 * If we just froze the page then put it onto the 3105 * per cpu partial list. 3106 */ 3107 put_cpu_partial(s, page, 1); 3108 stat(s, CPU_PARTIAL_FREE); 3109 } 3110 3111 return; 3112 } 3113 3114 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) 3115 goto slab_empty; 3116 3117 /* 3118 * Objects left in the slab. If it was not on the partial list before 3119 * then add it. 3120 */ 3121 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { 3122 remove_full(s, n, page); 3123 add_partial(n, page, DEACTIVATE_TO_TAIL); 3124 stat(s, FREE_ADD_PARTIAL); 3125 } 3126 spin_unlock_irqrestore(&n->list_lock, flags); 3127 return; 3128 3129 slab_empty: 3130 if (prior) { 3131 /* 3132 * Slab on the partial list. 3133 */ 3134 remove_partial(n, page); 3135 stat(s, FREE_REMOVE_PARTIAL); 3136 } else { 3137 /* Slab must be on the full list */ 3138 remove_full(s, n, page); 3139 } 3140 3141 spin_unlock_irqrestore(&n->list_lock, flags); 3142 stat(s, FREE_SLAB); 3143 discard_slab(s, page); 3144 } 3145 3146 /* 3147 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 3148 * can perform fastpath freeing without additional function calls. 3149 * 3150 * The fastpath is only possible if we are freeing to the current cpu slab 3151 * of this processor. This typically the case if we have just allocated 3152 * the item before. 3153 * 3154 * If fastpath is not possible then fall back to __slab_free where we deal 3155 * with all sorts of special processing. 3156 * 3157 * Bulk free of a freelist with several objects (all pointing to the 3158 * same page) possible by specifying head and tail ptr, plus objects 3159 * count (cnt). Bulk free indicated by tail pointer being set. 3160 */ 3161 static __always_inline void do_slab_free(struct kmem_cache *s, 3162 struct page *page, void *head, void *tail, 3163 int cnt, unsigned long addr) 3164 { 3165 void *tail_obj = tail ? : head; 3166 struct kmem_cache_cpu *c; 3167 unsigned long tid; 3168 3169 memcg_slab_free_hook(s, &head, 1); 3170 redo: 3171 /* 3172 * Determine the currently cpus per cpu slab. 3173 * The cpu may change afterward. However that does not matter since 3174 * data is retrieved via this pointer. If we are on the same cpu 3175 * during the cmpxchg then the free will succeed. 3176 */ 3177 do { 3178 tid = this_cpu_read(s->cpu_slab->tid); 3179 c = raw_cpu_ptr(s->cpu_slab); 3180 } while (IS_ENABLED(CONFIG_PREEMPTION) && 3181 unlikely(tid != READ_ONCE(c->tid))); 3182 3183 /* Same with comment on barrier() in slab_alloc_node() */ 3184 barrier(); 3185 3186 if (likely(page == c->page)) { 3187 void **freelist = READ_ONCE(c->freelist); 3188 3189 set_freepointer(s, tail_obj, freelist); 3190 3191 if (unlikely(!this_cpu_cmpxchg_double( 3192 s->cpu_slab->freelist, s->cpu_slab->tid, 3193 freelist, tid, 3194 head, next_tid(tid)))) { 3195 3196 note_cmpxchg_failure("slab_free", s, tid); 3197 goto redo; 3198 } 3199 stat(s, FREE_FASTPATH); 3200 } else 3201 __slab_free(s, page, head, tail_obj, cnt, addr); 3202 3203 } 3204 3205 static __always_inline void slab_free(struct kmem_cache *s, struct page *page, 3206 void *head, void *tail, int cnt, 3207 unsigned long addr) 3208 { 3209 /* 3210 * With KASAN enabled slab_free_freelist_hook modifies the freelist 3211 * to remove objects, whose reuse must be delayed. 3212 */ 3213 if (slab_free_freelist_hook(s, &head, &tail)) 3214 do_slab_free(s, page, head, tail, cnt, addr); 3215 } 3216 3217 #ifdef CONFIG_KASAN_GENERIC 3218 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) 3219 { 3220 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr); 3221 } 3222 #endif 3223 3224 void kmem_cache_free(struct kmem_cache *s, void *x) 3225 { 3226 s = cache_from_obj(s, x); 3227 if (!s) 3228 return; 3229 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_); 3230 trace_kmem_cache_free(_RET_IP_, x, s->name); 3231 } 3232 EXPORT_SYMBOL(kmem_cache_free); 3233 3234 struct detached_freelist { 3235 struct page *page; 3236 void *tail; 3237 void *freelist; 3238 int cnt; 3239 struct kmem_cache *s; 3240 }; 3241 3242 static inline void free_nonslab_page(struct page *page, void *object) 3243 { 3244 unsigned int order = compound_order(page); 3245 3246 VM_BUG_ON_PAGE(!PageCompound(page), page); 3247 kfree_hook(object); 3248 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, -(PAGE_SIZE << order)); 3249 __free_pages(page, order); 3250 } 3251 3252 /* 3253 * This function progressively scans the array with free objects (with 3254 * a limited look ahead) and extract objects belonging to the same 3255 * page. It builds a detached freelist directly within the given 3256 * page/objects. This can happen without any need for 3257 * synchronization, because the objects are owned by running process. 3258 * The freelist is build up as a single linked list in the objects. 3259 * The idea is, that this detached freelist can then be bulk 3260 * transferred to the real freelist(s), but only requiring a single 3261 * synchronization primitive. Look ahead in the array is limited due 3262 * to performance reasons. 3263 */ 3264 static inline 3265 int build_detached_freelist(struct kmem_cache *s, size_t size, 3266 void **p, struct detached_freelist *df) 3267 { 3268 size_t first_skipped_index = 0; 3269 int lookahead = 3; 3270 void *object; 3271 struct page *page; 3272 3273 /* Always re-init detached_freelist */ 3274 df->page = NULL; 3275 3276 do { 3277 object = p[--size]; 3278 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */ 3279 } while (!object && size); 3280 3281 if (!object) 3282 return 0; 3283 3284 page = virt_to_head_page(object); 3285 if (!s) { 3286 /* Handle kalloc'ed objects */ 3287 if (unlikely(!PageSlab(page))) { 3288 free_nonslab_page(page, object); 3289 p[size] = NULL; /* mark object processed */ 3290 return size; 3291 } 3292 /* Derive kmem_cache from object */ 3293 df->s = page->slab_cache; 3294 } else { 3295 df->s = cache_from_obj(s, object); /* Support for memcg */ 3296 } 3297 3298 if (is_kfence_address(object)) { 3299 slab_free_hook(df->s, object, false); 3300 __kfence_free(object); 3301 p[size] = NULL; /* mark object processed */ 3302 return size; 3303 } 3304 3305 /* Start new detached freelist */ 3306 df->page = page; 3307 set_freepointer(df->s, object, NULL); 3308 df->tail = object; 3309 df->freelist = object; 3310 p[size] = NULL; /* mark object processed */ 3311 df->cnt = 1; 3312 3313 while (size) { 3314 object = p[--size]; 3315 if (!object) 3316 continue; /* Skip processed objects */ 3317 3318 /* df->page is always set at this point */ 3319 if (df->page == virt_to_head_page(object)) { 3320 /* Opportunity build freelist */ 3321 set_freepointer(df->s, object, df->freelist); 3322 df->freelist = object; 3323 df->cnt++; 3324 p[size] = NULL; /* mark object processed */ 3325 3326 continue; 3327 } 3328 3329 /* Limit look ahead search */ 3330 if (!--lookahead) 3331 break; 3332 3333 if (!first_skipped_index) 3334 first_skipped_index = size + 1; 3335 } 3336 3337 return first_skipped_index; 3338 } 3339 3340 /* Note that interrupts must be enabled when calling this function. */ 3341 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) 3342 { 3343 if (WARN_ON(!size)) 3344 return; 3345 3346 memcg_slab_free_hook(s, p, size); 3347 do { 3348 struct detached_freelist df; 3349 3350 size = build_detached_freelist(s, size, p, &df); 3351 if (!df.page) 3352 continue; 3353 3354 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_); 3355 } while (likely(size)); 3356 } 3357 EXPORT_SYMBOL(kmem_cache_free_bulk); 3358 3359 /* Note that interrupts must be enabled when calling this function. */ 3360 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, 3361 void **p) 3362 { 3363 struct kmem_cache_cpu *c; 3364 int i; 3365 struct obj_cgroup *objcg = NULL; 3366 3367 /* memcg and kmem_cache debug support */ 3368 s = slab_pre_alloc_hook(s, &objcg, size, flags); 3369 if (unlikely(!s)) 3370 return false; 3371 /* 3372 * Drain objects in the per cpu slab, while disabling local 3373 * IRQs, which protects against PREEMPT and interrupts 3374 * handlers invoking normal fastpath. 3375 */ 3376 local_irq_disable(); 3377 c = this_cpu_ptr(s->cpu_slab); 3378 3379 for (i = 0; i < size; i++) { 3380 void *object = kfence_alloc(s, s->object_size, flags); 3381 3382 if (unlikely(object)) { 3383 p[i] = object; 3384 continue; 3385 } 3386 3387 object = c->freelist; 3388 if (unlikely(!object)) { 3389 /* 3390 * We may have removed an object from c->freelist using 3391 * the fastpath in the previous iteration; in that case, 3392 * c->tid has not been bumped yet. 3393 * Since ___slab_alloc() may reenable interrupts while 3394 * allocating memory, we should bump c->tid now. 3395 */ 3396 c->tid = next_tid(c->tid); 3397 3398 /* 3399 * Invoking slow path likely have side-effect 3400 * of re-populating per CPU c->freelist 3401 */ 3402 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, 3403 _RET_IP_, c); 3404 if (unlikely(!p[i])) 3405 goto error; 3406 3407 c = this_cpu_ptr(s->cpu_slab); 3408 maybe_wipe_obj_freeptr(s, p[i]); 3409 3410 continue; /* goto for-loop */ 3411 } 3412 c->freelist = get_freepointer(s, object); 3413 p[i] = object; 3414 maybe_wipe_obj_freeptr(s, p[i]); 3415 } 3416 c->tid = next_tid(c->tid); 3417 local_irq_enable(); 3418 3419 /* 3420 * memcg and kmem_cache debug support and memory initialization. 3421 * Done outside of the IRQ disabled fastpath loop. 3422 */ 3423 slab_post_alloc_hook(s, objcg, flags, size, p, 3424 slab_want_init_on_alloc(flags, s)); 3425 return i; 3426 error: 3427 local_irq_enable(); 3428 slab_post_alloc_hook(s, objcg, flags, i, p, false); 3429 __kmem_cache_free_bulk(s, i, p); 3430 return 0; 3431 } 3432 EXPORT_SYMBOL(kmem_cache_alloc_bulk); 3433 3434 3435 /* 3436 * Object placement in a slab is made very easy because we always start at 3437 * offset 0. If we tune the size of the object to the alignment then we can 3438 * get the required alignment by putting one properly sized object after 3439 * another. 3440 * 3441 * Notice that the allocation order determines the sizes of the per cpu 3442 * caches. Each processor has always one slab available for allocations. 3443 * Increasing the allocation order reduces the number of times that slabs 3444 * must be moved on and off the partial lists and is therefore a factor in 3445 * locking overhead. 3446 */ 3447 3448 /* 3449 * Minimum / Maximum order of slab pages. This influences locking overhead 3450 * and slab fragmentation. A higher order reduces the number of partial slabs 3451 * and increases the number of allocations possible without having to 3452 * take the list_lock. 3453 */ 3454 static unsigned int slub_min_order; 3455 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 3456 static unsigned int slub_min_objects; 3457 3458 /* 3459 * Calculate the order of allocation given an slab object size. 3460 * 3461 * The order of allocation has significant impact on performance and other 3462 * system components. Generally order 0 allocations should be preferred since 3463 * order 0 does not cause fragmentation in the page allocator. Larger objects 3464 * be problematic to put into order 0 slabs because there may be too much 3465 * unused space left. We go to a higher order if more than 1/16th of the slab 3466 * would be wasted. 3467 * 3468 * In order to reach satisfactory performance we must ensure that a minimum 3469 * number of objects is in one slab. Otherwise we may generate too much 3470 * activity on the partial lists which requires taking the list_lock. This is 3471 * less a concern for large slabs though which are rarely used. 3472 * 3473 * slub_max_order specifies the order where we begin to stop considering the 3474 * number of objects in a slab as critical. If we reach slub_max_order then 3475 * we try to keep the page order as low as possible. So we accept more waste 3476 * of space in favor of a small page order. 3477 * 3478 * Higher order allocations also allow the placement of more objects in a 3479 * slab and thereby reduce object handling overhead. If the user has 3480 * requested a higher minimum order then we start with that one instead of 3481 * the smallest order which will fit the object. 3482 */ 3483 static inline unsigned int slab_order(unsigned int size, 3484 unsigned int min_objects, unsigned int max_order, 3485 unsigned int fract_leftover) 3486 { 3487 unsigned int min_order = slub_min_order; 3488 unsigned int order; 3489 3490 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE) 3491 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 3492 3493 for (order = max(min_order, (unsigned int)get_order(min_objects * size)); 3494 order <= max_order; order++) { 3495 3496 unsigned int slab_size = (unsigned int)PAGE_SIZE << order; 3497 unsigned int rem; 3498 3499 rem = slab_size % size; 3500 3501 if (rem <= slab_size / fract_leftover) 3502 break; 3503 } 3504 3505 return order; 3506 } 3507 3508 static inline int calculate_order(unsigned int size) 3509 { 3510 unsigned int order; 3511 unsigned int min_objects; 3512 unsigned int max_objects; 3513 unsigned int nr_cpus; 3514 3515 /* 3516 * Attempt to find best configuration for a slab. This 3517 * works by first attempting to generate a layout with 3518 * the best configuration and backing off gradually. 3519 * 3520 * First we increase the acceptable waste in a slab. Then 3521 * we reduce the minimum objects required in a slab. 3522 */ 3523 min_objects = slub_min_objects; 3524 if (!min_objects) { 3525 /* 3526 * Some architectures will only update present cpus when 3527 * onlining them, so don't trust the number if it's just 1. But 3528 * we also don't want to use nr_cpu_ids always, as on some other 3529 * architectures, there can be many possible cpus, but never 3530 * onlined. Here we compromise between trying to avoid too high 3531 * order on systems that appear larger than they are, and too 3532 * low order on systems that appear smaller than they are. 3533 */ 3534 nr_cpus = num_present_cpus(); 3535 if (nr_cpus <= 1) 3536 nr_cpus = nr_cpu_ids; 3537 min_objects = 4 * (fls(nr_cpus) + 1); 3538 } 3539 max_objects = order_objects(slub_max_order, size); 3540 min_objects = min(min_objects, max_objects); 3541 3542 while (min_objects > 1) { 3543 unsigned int fraction; 3544 3545 fraction = 16; 3546 while (fraction >= 4) { 3547 order = slab_order(size, min_objects, 3548 slub_max_order, fraction); 3549 if (order <= slub_max_order) 3550 return order; 3551 fraction /= 2; 3552 } 3553 min_objects--; 3554 } 3555 3556 /* 3557 * We were unable to place multiple objects in a slab. Now 3558 * lets see if we can place a single object there. 3559 */ 3560 order = slab_order(size, 1, slub_max_order, 1); 3561 if (order <= slub_max_order) 3562 return order; 3563 3564 /* 3565 * Doh this slab cannot be placed using slub_max_order. 3566 */ 3567 order = slab_order(size, 1, MAX_ORDER, 1); 3568 if (order < MAX_ORDER) 3569 return order; 3570 return -ENOSYS; 3571 } 3572 3573 static void 3574 init_kmem_cache_node(struct kmem_cache_node *n) 3575 { 3576 n->nr_partial = 0; 3577 spin_lock_init(&n->list_lock); 3578 INIT_LIST_HEAD(&n->partial); 3579 #ifdef CONFIG_SLUB_DEBUG 3580 atomic_long_set(&n->nr_slabs, 0); 3581 atomic_long_set(&n->total_objects, 0); 3582 INIT_LIST_HEAD(&n->full); 3583 #endif 3584 } 3585 3586 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 3587 { 3588 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 3589 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu)); 3590 3591 /* 3592 * Must align to double word boundary for the double cmpxchg 3593 * instructions to work; see __pcpu_double_call_return_bool(). 3594 */ 3595 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 3596 2 * sizeof(void *)); 3597 3598 if (!s->cpu_slab) 3599 return 0; 3600 3601 init_kmem_cache_cpus(s); 3602 3603 return 1; 3604 } 3605 3606 static struct kmem_cache *kmem_cache_node; 3607 3608 /* 3609 * No kmalloc_node yet so do it by hand. We know that this is the first 3610 * slab on the node for this slabcache. There are no concurrent accesses 3611 * possible. 3612 * 3613 * Note that this function only works on the kmem_cache_node 3614 * when allocating for the kmem_cache_node. This is used for bootstrapping 3615 * memory on a fresh node that has no slab structures yet. 3616 */ 3617 static void early_kmem_cache_node_alloc(int node) 3618 { 3619 struct page *page; 3620 struct kmem_cache_node *n; 3621 3622 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 3623 3624 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 3625 3626 BUG_ON(!page); 3627 if (page_to_nid(page) != node) { 3628 pr_err("SLUB: Unable to allocate memory from node %d\n", node); 3629 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n"); 3630 } 3631 3632 n = page->freelist; 3633 BUG_ON(!n); 3634 #ifdef CONFIG_SLUB_DEBUG 3635 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 3636 init_tracking(kmem_cache_node, n); 3637 #endif 3638 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false); 3639 page->freelist = get_freepointer(kmem_cache_node, n); 3640 page->inuse = 1; 3641 page->frozen = 0; 3642 kmem_cache_node->node[node] = n; 3643 init_kmem_cache_node(n); 3644 inc_slabs_node(kmem_cache_node, node, page->objects); 3645 3646 /* 3647 * No locks need to be taken here as it has just been 3648 * initialized and there is no concurrent access. 3649 */ 3650 __add_partial(n, page, DEACTIVATE_TO_HEAD); 3651 } 3652 3653 static void free_kmem_cache_nodes(struct kmem_cache *s) 3654 { 3655 int node; 3656 struct kmem_cache_node *n; 3657 3658 for_each_kmem_cache_node(s, node, n) { 3659 s->node[node] = NULL; 3660 kmem_cache_free(kmem_cache_node, n); 3661 } 3662 } 3663 3664 void __kmem_cache_release(struct kmem_cache *s) 3665 { 3666 cache_random_seq_destroy(s); 3667 free_percpu(s->cpu_slab); 3668 free_kmem_cache_nodes(s); 3669 } 3670 3671 static int init_kmem_cache_nodes(struct kmem_cache *s) 3672 { 3673 int node; 3674 3675 for_each_node_mask(node, slab_nodes) { 3676 struct kmem_cache_node *n; 3677 3678 if (slab_state == DOWN) { 3679 early_kmem_cache_node_alloc(node); 3680 continue; 3681 } 3682 n = kmem_cache_alloc_node(kmem_cache_node, 3683 GFP_KERNEL, node); 3684 3685 if (!n) { 3686 free_kmem_cache_nodes(s); 3687 return 0; 3688 } 3689 3690 init_kmem_cache_node(n); 3691 s->node[node] = n; 3692 } 3693 return 1; 3694 } 3695 3696 static void set_min_partial(struct kmem_cache *s, unsigned long min) 3697 { 3698 if (min < MIN_PARTIAL) 3699 min = MIN_PARTIAL; 3700 else if (min > MAX_PARTIAL) 3701 min = MAX_PARTIAL; 3702 s->min_partial = min; 3703 } 3704 3705 static void set_cpu_partial(struct kmem_cache *s) 3706 { 3707 #ifdef CONFIG_SLUB_CPU_PARTIAL 3708 /* 3709 * cpu_partial determined the maximum number of objects kept in the 3710 * per cpu partial lists of a processor. 3711 * 3712 * Per cpu partial lists mainly contain slabs that just have one 3713 * object freed. If they are used for allocation then they can be 3714 * filled up again with minimal effort. The slab will never hit the 3715 * per node partial lists and therefore no locking will be required. 3716 * 3717 * This setting also determines 3718 * 3719 * A) The number of objects from per cpu partial slabs dumped to the 3720 * per node list when we reach the limit. 3721 * B) The number of objects in cpu partial slabs to extract from the 3722 * per node list when we run out of per cpu objects. We only fetch 3723 * 50% to keep some capacity around for frees. 3724 */ 3725 if (!kmem_cache_has_cpu_partial(s)) 3726 slub_set_cpu_partial(s, 0); 3727 else if (s->size >= PAGE_SIZE) 3728 slub_set_cpu_partial(s, 2); 3729 else if (s->size >= 1024) 3730 slub_set_cpu_partial(s, 6); 3731 else if (s->size >= 256) 3732 slub_set_cpu_partial(s, 13); 3733 else 3734 slub_set_cpu_partial(s, 30); 3735 #endif 3736 } 3737 3738 /* 3739 * calculate_sizes() determines the order and the distribution of data within 3740 * a slab object. 3741 */ 3742 static int calculate_sizes(struct kmem_cache *s, int forced_order) 3743 { 3744 slab_flags_t flags = s->flags; 3745 unsigned int size = s->object_size; 3746 unsigned int order; 3747 3748 /* 3749 * Round up object size to the next word boundary. We can only 3750 * place the free pointer at word boundaries and this determines 3751 * the possible location of the free pointer. 3752 */ 3753 size = ALIGN(size, sizeof(void *)); 3754 3755 #ifdef CONFIG_SLUB_DEBUG 3756 /* 3757 * Determine if we can poison the object itself. If the user of 3758 * the slab may touch the object after free or before allocation 3759 * then we should never poison the object itself. 3760 */ 3761 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && 3762 !s->ctor) 3763 s->flags |= __OBJECT_POISON; 3764 else 3765 s->flags &= ~__OBJECT_POISON; 3766 3767 3768 /* 3769 * If we are Redzoning then check if there is some space between the 3770 * end of the object and the free pointer. If not then add an 3771 * additional word to have some bytes to store Redzone information. 3772 */ 3773 if ((flags & SLAB_RED_ZONE) && size == s->object_size) 3774 size += sizeof(void *); 3775 #endif 3776 3777 /* 3778 * With that we have determined the number of bytes in actual use 3779 * by the object and redzoning. 3780 */ 3781 s->inuse = size; 3782 3783 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || 3784 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) || 3785 s->ctor) { 3786 /* 3787 * Relocate free pointer after the object if it is not 3788 * permitted to overwrite the first word of the object on 3789 * kmem_cache_free. 3790 * 3791 * This is the case if we do RCU, have a constructor or 3792 * destructor, are poisoning the objects, or are 3793 * redzoning an object smaller than sizeof(void *). 3794 * 3795 * The assumption that s->offset >= s->inuse means free 3796 * pointer is outside of the object is used in the 3797 * freeptr_outside_object() function. If that is no 3798 * longer true, the function needs to be modified. 3799 */ 3800 s->offset = size; 3801 size += sizeof(void *); 3802 } else { 3803 /* 3804 * Store freelist pointer near middle of object to keep 3805 * it away from the edges of the object to avoid small 3806 * sized over/underflows from neighboring allocations. 3807 */ 3808 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *)); 3809 } 3810 3811 #ifdef CONFIG_SLUB_DEBUG 3812 if (flags & SLAB_STORE_USER) 3813 /* 3814 * Need to store information about allocs and frees after 3815 * the object. 3816 */ 3817 size += 2 * sizeof(struct track); 3818 #endif 3819 3820 kasan_cache_create(s, &size, &s->flags); 3821 #ifdef CONFIG_SLUB_DEBUG 3822 if (flags & SLAB_RED_ZONE) { 3823 /* 3824 * Add some empty padding so that we can catch 3825 * overwrites from earlier objects rather than let 3826 * tracking information or the free pointer be 3827 * corrupted if a user writes before the start 3828 * of the object. 3829 */ 3830 size += sizeof(void *); 3831 3832 s->red_left_pad = sizeof(void *); 3833 s->red_left_pad = ALIGN(s->red_left_pad, s->align); 3834 size += s->red_left_pad; 3835 } 3836 #endif 3837 3838 /* 3839 * SLUB stores one object immediately after another beginning from 3840 * offset 0. In order to align the objects we have to simply size 3841 * each object to conform to the alignment. 3842 */ 3843 size = ALIGN(size, s->align); 3844 s->size = size; 3845 s->reciprocal_size = reciprocal_value(size); 3846 if (forced_order >= 0) 3847 order = forced_order; 3848 else 3849 order = calculate_order(size); 3850 3851 if ((int)order < 0) 3852 return 0; 3853 3854 s->allocflags = 0; 3855 if (order) 3856 s->allocflags |= __GFP_COMP; 3857 3858 if (s->flags & SLAB_CACHE_DMA) 3859 s->allocflags |= GFP_DMA; 3860 3861 if (s->flags & SLAB_CACHE_DMA32) 3862 s->allocflags |= GFP_DMA32; 3863 3864 if (s->flags & SLAB_RECLAIM_ACCOUNT) 3865 s->allocflags |= __GFP_RECLAIMABLE; 3866 3867 /* 3868 * Determine the number of objects per slab 3869 */ 3870 s->oo = oo_make(order, size); 3871 s->min = oo_make(get_order(size), size); 3872 if (oo_objects(s->oo) > oo_objects(s->max)) 3873 s->max = s->oo; 3874 3875 return !!oo_objects(s->oo); 3876 } 3877 3878 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) 3879 { 3880 s->flags = kmem_cache_flags(s->size, flags, s->name); 3881 #ifdef CONFIG_SLAB_FREELIST_HARDENED 3882 s->random = get_random_long(); 3883 #endif 3884 3885 if (!calculate_sizes(s, -1)) 3886 goto error; 3887 if (disable_higher_order_debug) { 3888 /* 3889 * Disable debugging flags that store metadata if the min slab 3890 * order increased. 3891 */ 3892 if (get_order(s->size) > get_order(s->object_size)) { 3893 s->flags &= ~DEBUG_METADATA_FLAGS; 3894 s->offset = 0; 3895 if (!calculate_sizes(s, -1)) 3896 goto error; 3897 } 3898 } 3899 3900 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 3901 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 3902 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0) 3903 /* Enable fast mode */ 3904 s->flags |= __CMPXCHG_DOUBLE; 3905 #endif 3906 3907 /* 3908 * The larger the object size is, the more pages we want on the partial 3909 * list to avoid pounding the page allocator excessively. 3910 */ 3911 set_min_partial(s, ilog2(s->size) / 2); 3912 3913 set_cpu_partial(s); 3914 3915 #ifdef CONFIG_NUMA 3916 s->remote_node_defrag_ratio = 1000; 3917 #endif 3918 3919 /* Initialize the pre-computed randomized freelist if slab is up */ 3920 if (slab_state >= UP) { 3921 if (init_cache_random_seq(s)) 3922 goto error; 3923 } 3924 3925 if (!init_kmem_cache_nodes(s)) 3926 goto error; 3927 3928 if (alloc_kmem_cache_cpus(s)) 3929 return 0; 3930 3931 free_kmem_cache_nodes(s); 3932 error: 3933 return -EINVAL; 3934 } 3935 3936 static void list_slab_objects(struct kmem_cache *s, struct page *page, 3937 const char *text) 3938 { 3939 #ifdef CONFIG_SLUB_DEBUG 3940 void *addr = page_address(page); 3941 unsigned long *map; 3942 void *p; 3943 3944 slab_err(s, page, text, s->name); 3945 slab_lock(page); 3946 3947 map = get_map(s, page); 3948 for_each_object(p, s, addr, page->objects) { 3949 3950 if (!test_bit(__obj_to_index(s, addr, p), map)) { 3951 pr_err("Object 0x%p @offset=%tu\n", p, p - addr); 3952 print_tracking(s, p); 3953 } 3954 } 3955 put_map(map); 3956 slab_unlock(page); 3957 #endif 3958 } 3959 3960 /* 3961 * Attempt to free all partial slabs on a node. 3962 * This is called from __kmem_cache_shutdown(). We must take list_lock 3963 * because sysfs file might still access partial list after the shutdowning. 3964 */ 3965 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 3966 { 3967 LIST_HEAD(discard); 3968 struct page *page, *h; 3969 3970 BUG_ON(irqs_disabled()); 3971 spin_lock_irq(&n->list_lock); 3972 list_for_each_entry_safe(page, h, &n->partial, slab_list) { 3973 if (!page->inuse) { 3974 remove_partial(n, page); 3975 list_add(&page->slab_list, &discard); 3976 } else { 3977 list_slab_objects(s, page, 3978 "Objects remaining in %s on __kmem_cache_shutdown()"); 3979 } 3980 } 3981 spin_unlock_irq(&n->list_lock); 3982 3983 list_for_each_entry_safe(page, h, &discard, slab_list) 3984 discard_slab(s, page); 3985 } 3986 3987 bool __kmem_cache_empty(struct kmem_cache *s) 3988 { 3989 int node; 3990 struct kmem_cache_node *n; 3991 3992 for_each_kmem_cache_node(s, node, n) 3993 if (n->nr_partial || slabs_node(s, node)) 3994 return false; 3995 return true; 3996 } 3997 3998 /* 3999 * Release all resources used by a slab cache. 4000 */ 4001 int __kmem_cache_shutdown(struct kmem_cache *s) 4002 { 4003 int node; 4004 struct kmem_cache_node *n; 4005 4006 flush_all(s); 4007 /* Attempt to free all objects */ 4008 for_each_kmem_cache_node(s, node, n) { 4009 free_partial(s, n); 4010 if (n->nr_partial || slabs_node(s, node)) 4011 return 1; 4012 } 4013 return 0; 4014 } 4015 4016 #ifdef CONFIG_PRINTK 4017 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page) 4018 { 4019 void *base; 4020 int __maybe_unused i; 4021 unsigned int objnr; 4022 void *objp; 4023 void *objp0; 4024 struct kmem_cache *s = page->slab_cache; 4025 struct track __maybe_unused *trackp; 4026 4027 kpp->kp_ptr = object; 4028 kpp->kp_page = page; 4029 kpp->kp_slab_cache = s; 4030 base = page_address(page); 4031 objp0 = kasan_reset_tag(object); 4032 #ifdef CONFIG_SLUB_DEBUG 4033 objp = restore_red_left(s, objp0); 4034 #else 4035 objp = objp0; 4036 #endif 4037 objnr = obj_to_index(s, page, objp); 4038 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp); 4039 objp = base + s->size * objnr; 4040 kpp->kp_objp = objp; 4041 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) || 4042 !(s->flags & SLAB_STORE_USER)) 4043 return; 4044 #ifdef CONFIG_SLUB_DEBUG 4045 objp = fixup_red_left(s, objp); 4046 trackp = get_track(s, objp, TRACK_ALLOC); 4047 kpp->kp_ret = (void *)trackp->addr; 4048 #ifdef CONFIG_STACKTRACE 4049 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) { 4050 kpp->kp_stack[i] = (void *)trackp->addrs[i]; 4051 if (!kpp->kp_stack[i]) 4052 break; 4053 } 4054 4055 trackp = get_track(s, objp, TRACK_FREE); 4056 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) { 4057 kpp->kp_free_stack[i] = (void *)trackp->addrs[i]; 4058 if (!kpp->kp_free_stack[i]) 4059 break; 4060 } 4061 #endif 4062 #endif 4063 } 4064 #endif 4065 4066 /******************************************************************** 4067 * Kmalloc subsystem 4068 *******************************************************************/ 4069 4070 static int __init setup_slub_min_order(char *str) 4071 { 4072 get_option(&str, (int *)&slub_min_order); 4073 4074 return 1; 4075 } 4076 4077 __setup("slub_min_order=", setup_slub_min_order); 4078 4079 static int __init setup_slub_max_order(char *str) 4080 { 4081 get_option(&str, (int *)&slub_max_order); 4082 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1); 4083 4084 return 1; 4085 } 4086 4087 __setup("slub_max_order=", setup_slub_max_order); 4088 4089 static int __init setup_slub_min_objects(char *str) 4090 { 4091 get_option(&str, (int *)&slub_min_objects); 4092 4093 return 1; 4094 } 4095 4096 __setup("slub_min_objects=", setup_slub_min_objects); 4097 4098 void *__kmalloc(size_t size, gfp_t flags) 4099 { 4100 struct kmem_cache *s; 4101 void *ret; 4102 4103 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 4104 return kmalloc_large(size, flags); 4105 4106 s = kmalloc_slab(size, flags); 4107 4108 if (unlikely(ZERO_OR_NULL_PTR(s))) 4109 return s; 4110 4111 ret = slab_alloc(s, flags, _RET_IP_, size); 4112 4113 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 4114 4115 ret = kasan_kmalloc(s, ret, size, flags); 4116 4117 return ret; 4118 } 4119 EXPORT_SYMBOL(__kmalloc); 4120 4121 #ifdef CONFIG_NUMA 4122 static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 4123 { 4124 struct page *page; 4125 void *ptr = NULL; 4126 unsigned int order = get_order(size); 4127 4128 flags |= __GFP_COMP; 4129 page = alloc_pages_node(node, flags, order); 4130 if (page) { 4131 ptr = page_address(page); 4132 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, 4133 PAGE_SIZE << order); 4134 } 4135 4136 return kmalloc_large_node_hook(ptr, size, flags); 4137 } 4138 4139 void *__kmalloc_node(size_t size, gfp_t flags, int node) 4140 { 4141 struct kmem_cache *s; 4142 void *ret; 4143 4144 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 4145 ret = kmalloc_large_node(size, flags, node); 4146 4147 trace_kmalloc_node(_RET_IP_, ret, 4148 size, PAGE_SIZE << get_order(size), 4149 flags, node); 4150 4151 return ret; 4152 } 4153 4154 s = kmalloc_slab(size, flags); 4155 4156 if (unlikely(ZERO_OR_NULL_PTR(s))) 4157 return s; 4158 4159 ret = slab_alloc_node(s, flags, node, _RET_IP_, size); 4160 4161 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 4162 4163 ret = kasan_kmalloc(s, ret, size, flags); 4164 4165 return ret; 4166 } 4167 EXPORT_SYMBOL(__kmalloc_node); 4168 #endif /* CONFIG_NUMA */ 4169 4170 #ifdef CONFIG_HARDENED_USERCOPY 4171 /* 4172 * Rejects incorrectly sized objects and objects that are to be copied 4173 * to/from userspace but do not fall entirely within the containing slab 4174 * cache's usercopy region. 4175 * 4176 * Returns NULL if check passes, otherwise const char * to name of cache 4177 * to indicate an error. 4178 */ 4179 void __check_heap_object(const void *ptr, unsigned long n, struct page *page, 4180 bool to_user) 4181 { 4182 struct kmem_cache *s; 4183 unsigned int offset; 4184 size_t object_size; 4185 bool is_kfence = is_kfence_address(ptr); 4186 4187 ptr = kasan_reset_tag(ptr); 4188 4189 /* Find object and usable object size. */ 4190 s = page->slab_cache; 4191 4192 /* Reject impossible pointers. */ 4193 if (ptr < page_address(page)) 4194 usercopy_abort("SLUB object not in SLUB page?!", NULL, 4195 to_user, 0, n); 4196 4197 /* Find offset within object. */ 4198 if (is_kfence) 4199 offset = ptr - kfence_object_start(ptr); 4200 else 4201 offset = (ptr - page_address(page)) % s->size; 4202 4203 /* Adjust for redzone and reject if within the redzone. */ 4204 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) { 4205 if (offset < s->red_left_pad) 4206 usercopy_abort("SLUB object in left red zone", 4207 s->name, to_user, offset, n); 4208 offset -= s->red_left_pad; 4209 } 4210 4211 /* Allow address range falling entirely within usercopy region. */ 4212 if (offset >= s->useroffset && 4213 offset - s->useroffset <= s->usersize && 4214 n <= s->useroffset - offset + s->usersize) 4215 return; 4216 4217 /* 4218 * If the copy is still within the allocated object, produce 4219 * a warning instead of rejecting the copy. This is intended 4220 * to be a temporary method to find any missing usercopy 4221 * whitelists. 4222 */ 4223 object_size = slab_ksize(s); 4224 if (usercopy_fallback && 4225 offset <= object_size && n <= object_size - offset) { 4226 usercopy_warn("SLUB object", s->name, to_user, offset, n); 4227 return; 4228 } 4229 4230 usercopy_abort("SLUB object", s->name, to_user, offset, n); 4231 } 4232 #endif /* CONFIG_HARDENED_USERCOPY */ 4233 4234 size_t __ksize(const void *object) 4235 { 4236 struct page *page; 4237 4238 if (unlikely(object == ZERO_SIZE_PTR)) 4239 return 0; 4240 4241 page = virt_to_head_page(object); 4242 4243 if (unlikely(!PageSlab(page))) { 4244 WARN_ON(!PageCompound(page)); 4245 return page_size(page); 4246 } 4247 4248 return slab_ksize(page->slab_cache); 4249 } 4250 EXPORT_SYMBOL(__ksize); 4251 4252 void kfree(const void *x) 4253 { 4254 struct page *page; 4255 void *object = (void *)x; 4256 4257 trace_kfree(_RET_IP_, x); 4258 4259 if (unlikely(ZERO_OR_NULL_PTR(x))) 4260 return; 4261 4262 page = virt_to_head_page(x); 4263 if (unlikely(!PageSlab(page))) { 4264 free_nonslab_page(page, object); 4265 return; 4266 } 4267 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_); 4268 } 4269 EXPORT_SYMBOL(kfree); 4270 4271 #define SHRINK_PROMOTE_MAX 32 4272 4273 /* 4274 * kmem_cache_shrink discards empty slabs and promotes the slabs filled 4275 * up most to the head of the partial lists. New allocations will then 4276 * fill those up and thus they can be removed from the partial lists. 4277 * 4278 * The slabs with the least items are placed last. This results in them 4279 * being allocated from last increasing the chance that the last objects 4280 * are freed in them. 4281 */ 4282 int __kmem_cache_shrink(struct kmem_cache *s) 4283 { 4284 int node; 4285 int i; 4286 struct kmem_cache_node *n; 4287 struct page *page; 4288 struct page *t; 4289 struct list_head discard; 4290 struct list_head promote[SHRINK_PROMOTE_MAX]; 4291 unsigned long flags; 4292 int ret = 0; 4293 4294 flush_all(s); 4295 for_each_kmem_cache_node(s, node, n) { 4296 INIT_LIST_HEAD(&discard); 4297 for (i = 0; i < SHRINK_PROMOTE_MAX; i++) 4298 INIT_LIST_HEAD(promote + i); 4299 4300 spin_lock_irqsave(&n->list_lock, flags); 4301 4302 /* 4303 * Build lists of slabs to discard or promote. 4304 * 4305 * Note that concurrent frees may occur while we hold the 4306 * list_lock. page->inuse here is the upper limit. 4307 */ 4308 list_for_each_entry_safe(page, t, &n->partial, slab_list) { 4309 int free = page->objects - page->inuse; 4310 4311 /* Do not reread page->inuse */ 4312 barrier(); 4313 4314 /* We do not keep full slabs on the list */ 4315 BUG_ON(free <= 0); 4316 4317 if (free == page->objects) { 4318 list_move(&page->slab_list, &discard); 4319 n->nr_partial--; 4320 } else if (free <= SHRINK_PROMOTE_MAX) 4321 list_move(&page->slab_list, promote + free - 1); 4322 } 4323 4324 /* 4325 * Promote the slabs filled up most to the head of the 4326 * partial list. 4327 */ 4328 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) 4329 list_splice(promote + i, &n->partial); 4330 4331 spin_unlock_irqrestore(&n->list_lock, flags); 4332 4333 /* Release empty slabs */ 4334 list_for_each_entry_safe(page, t, &discard, slab_list) 4335 discard_slab(s, page); 4336 4337 if (slabs_node(s, node)) 4338 ret = 1; 4339 } 4340 4341 return ret; 4342 } 4343 4344 static int slab_mem_going_offline_callback(void *arg) 4345 { 4346 struct kmem_cache *s; 4347 4348 mutex_lock(&slab_mutex); 4349 list_for_each_entry(s, &slab_caches, list) 4350 __kmem_cache_shrink(s); 4351 mutex_unlock(&slab_mutex); 4352 4353 return 0; 4354 } 4355 4356 static void slab_mem_offline_callback(void *arg) 4357 { 4358 struct memory_notify *marg = arg; 4359 int offline_node; 4360 4361 offline_node = marg->status_change_nid_normal; 4362 4363 /* 4364 * If the node still has available memory. we need kmem_cache_node 4365 * for it yet. 4366 */ 4367 if (offline_node < 0) 4368 return; 4369 4370 mutex_lock(&slab_mutex); 4371 node_clear(offline_node, slab_nodes); 4372 /* 4373 * We no longer free kmem_cache_node structures here, as it would be 4374 * racy with all get_node() users, and infeasible to protect them with 4375 * slab_mutex. 4376 */ 4377 mutex_unlock(&slab_mutex); 4378 } 4379 4380 static int slab_mem_going_online_callback(void *arg) 4381 { 4382 struct kmem_cache_node *n; 4383 struct kmem_cache *s; 4384 struct memory_notify *marg = arg; 4385 int nid = marg->status_change_nid_normal; 4386 int ret = 0; 4387 4388 /* 4389 * If the node's memory is already available, then kmem_cache_node is 4390 * already created. Nothing to do. 4391 */ 4392 if (nid < 0) 4393 return 0; 4394 4395 /* 4396 * We are bringing a node online. No memory is available yet. We must 4397 * allocate a kmem_cache_node structure in order to bring the node 4398 * online. 4399 */ 4400 mutex_lock(&slab_mutex); 4401 list_for_each_entry(s, &slab_caches, list) { 4402 /* 4403 * The structure may already exist if the node was previously 4404 * onlined and offlined. 4405 */ 4406 if (get_node(s, nid)) 4407 continue; 4408 /* 4409 * XXX: kmem_cache_alloc_node will fallback to other nodes 4410 * since memory is not yet available from the node that 4411 * is brought up. 4412 */ 4413 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 4414 if (!n) { 4415 ret = -ENOMEM; 4416 goto out; 4417 } 4418 init_kmem_cache_node(n); 4419 s->node[nid] = n; 4420 } 4421 /* 4422 * Any cache created after this point will also have kmem_cache_node 4423 * initialized for the new node. 4424 */ 4425 node_set(nid, slab_nodes); 4426 out: 4427 mutex_unlock(&slab_mutex); 4428 return ret; 4429 } 4430 4431 static int slab_memory_callback(struct notifier_block *self, 4432 unsigned long action, void *arg) 4433 { 4434 int ret = 0; 4435 4436 switch (action) { 4437 case MEM_GOING_ONLINE: 4438 ret = slab_mem_going_online_callback(arg); 4439 break; 4440 case MEM_GOING_OFFLINE: 4441 ret = slab_mem_going_offline_callback(arg); 4442 break; 4443 case MEM_OFFLINE: 4444 case MEM_CANCEL_ONLINE: 4445 slab_mem_offline_callback(arg); 4446 break; 4447 case MEM_ONLINE: 4448 case MEM_CANCEL_OFFLINE: 4449 break; 4450 } 4451 if (ret) 4452 ret = notifier_from_errno(ret); 4453 else 4454 ret = NOTIFY_OK; 4455 return ret; 4456 } 4457 4458 static struct notifier_block slab_memory_callback_nb = { 4459 .notifier_call = slab_memory_callback, 4460 .priority = SLAB_CALLBACK_PRI, 4461 }; 4462 4463 /******************************************************************** 4464 * Basic setup of slabs 4465 *******************************************************************/ 4466 4467 /* 4468 * Used for early kmem_cache structures that were allocated using 4469 * the page allocator. Allocate them properly then fix up the pointers 4470 * that may be pointing to the wrong kmem_cache structure. 4471 */ 4472 4473 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) 4474 { 4475 int node; 4476 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); 4477 struct kmem_cache_node *n; 4478 4479 memcpy(s, static_cache, kmem_cache->object_size); 4480 4481 /* 4482 * This runs very early, and only the boot processor is supposed to be 4483 * up. Even if it weren't true, IRQs are not up so we couldn't fire 4484 * IPIs around. 4485 */ 4486 __flush_cpu_slab(s, smp_processor_id()); 4487 for_each_kmem_cache_node(s, node, n) { 4488 struct page *p; 4489 4490 list_for_each_entry(p, &n->partial, slab_list) 4491 p->slab_cache = s; 4492 4493 #ifdef CONFIG_SLUB_DEBUG 4494 list_for_each_entry(p, &n->full, slab_list) 4495 p->slab_cache = s; 4496 #endif 4497 } 4498 list_add(&s->list, &slab_caches); 4499 return s; 4500 } 4501 4502 void __init kmem_cache_init(void) 4503 { 4504 static __initdata struct kmem_cache boot_kmem_cache, 4505 boot_kmem_cache_node; 4506 int node; 4507 4508 if (debug_guardpage_minorder()) 4509 slub_max_order = 0; 4510 4511 /* Print slub debugging pointers without hashing */ 4512 if (__slub_debug_enabled()) 4513 no_hash_pointers_enable(NULL); 4514 4515 kmem_cache_node = &boot_kmem_cache_node; 4516 kmem_cache = &boot_kmem_cache; 4517 4518 /* 4519 * Initialize the nodemask for which we will allocate per node 4520 * structures. Here we don't need taking slab_mutex yet. 4521 */ 4522 for_each_node_state(node, N_NORMAL_MEMORY) 4523 node_set(node, slab_nodes); 4524 4525 create_boot_cache(kmem_cache_node, "kmem_cache_node", 4526 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0); 4527 4528 register_hotmemory_notifier(&slab_memory_callback_nb); 4529 4530 /* Able to allocate the per node structures */ 4531 slab_state = PARTIAL; 4532 4533 create_boot_cache(kmem_cache, "kmem_cache", 4534 offsetof(struct kmem_cache, node) + 4535 nr_node_ids * sizeof(struct kmem_cache_node *), 4536 SLAB_HWCACHE_ALIGN, 0, 0); 4537 4538 kmem_cache = bootstrap(&boot_kmem_cache); 4539 kmem_cache_node = bootstrap(&boot_kmem_cache_node); 4540 4541 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 4542 setup_kmalloc_cache_index_table(); 4543 create_kmalloc_caches(0); 4544 4545 /* Setup random freelists for each cache */ 4546 init_freelist_randomization(); 4547 4548 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL, 4549 slub_cpu_dead); 4550 4551 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n", 4552 cache_line_size(), 4553 slub_min_order, slub_max_order, slub_min_objects, 4554 nr_cpu_ids, nr_node_ids); 4555 } 4556 4557 void __init kmem_cache_init_late(void) 4558 { 4559 } 4560 4561 struct kmem_cache * 4562 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, 4563 slab_flags_t flags, void (*ctor)(void *)) 4564 { 4565 struct kmem_cache *s; 4566 4567 s = find_mergeable(size, align, flags, name, ctor); 4568 if (s) { 4569 s->refcount++; 4570 4571 /* 4572 * Adjust the object sizes so that we clear 4573 * the complete object on kzalloc. 4574 */ 4575 s->object_size = max(s->object_size, size); 4576 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *))); 4577 4578 if (sysfs_slab_alias(s, name)) { 4579 s->refcount--; 4580 s = NULL; 4581 } 4582 } 4583 4584 return s; 4585 } 4586 4587 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) 4588 { 4589 int err; 4590 4591 err = kmem_cache_open(s, flags); 4592 if (err) 4593 return err; 4594 4595 /* Mutex is not taken during early boot */ 4596 if (slab_state <= UP) 4597 return 0; 4598 4599 err = sysfs_slab_add(s); 4600 if (err) 4601 __kmem_cache_release(s); 4602 4603 if (s->flags & SLAB_STORE_USER) 4604 debugfs_slab_add(s); 4605 4606 return err; 4607 } 4608 4609 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 4610 { 4611 struct kmem_cache *s; 4612 void *ret; 4613 4614 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 4615 return kmalloc_large(size, gfpflags); 4616 4617 s = kmalloc_slab(size, gfpflags); 4618 4619 if (unlikely(ZERO_OR_NULL_PTR(s))) 4620 return s; 4621 4622 ret = slab_alloc(s, gfpflags, caller, size); 4623 4624 /* Honor the call site pointer we received. */ 4625 trace_kmalloc(caller, ret, size, s->size, gfpflags); 4626 4627 return ret; 4628 } 4629 EXPORT_SYMBOL(__kmalloc_track_caller); 4630 4631 #ifdef CONFIG_NUMA 4632 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 4633 int node, unsigned long caller) 4634 { 4635 struct kmem_cache *s; 4636 void *ret; 4637 4638 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 4639 ret = kmalloc_large_node(size, gfpflags, node); 4640 4641 trace_kmalloc_node(caller, ret, 4642 size, PAGE_SIZE << get_order(size), 4643 gfpflags, node); 4644 4645 return ret; 4646 } 4647 4648 s = kmalloc_slab(size, gfpflags); 4649 4650 if (unlikely(ZERO_OR_NULL_PTR(s))) 4651 return s; 4652 4653 ret = slab_alloc_node(s, gfpflags, node, caller, size); 4654 4655 /* Honor the call site pointer we received. */ 4656 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 4657 4658 return ret; 4659 } 4660 EXPORT_SYMBOL(__kmalloc_node_track_caller); 4661 #endif 4662 4663 #ifdef CONFIG_SYSFS 4664 static int count_inuse(struct page *page) 4665 { 4666 return page->inuse; 4667 } 4668 4669 static int count_total(struct page *page) 4670 { 4671 return page->objects; 4672 } 4673 #endif 4674 4675 #ifdef CONFIG_SLUB_DEBUG 4676 static void validate_slab(struct kmem_cache *s, struct page *page) 4677 { 4678 void *p; 4679 void *addr = page_address(page); 4680 unsigned long *map; 4681 4682 slab_lock(page); 4683 4684 if (!check_slab(s, page) || !on_freelist(s, page, NULL)) 4685 goto unlock; 4686 4687 /* Now we know that a valid freelist exists */ 4688 map = get_map(s, page); 4689 for_each_object(p, s, addr, page->objects) { 4690 u8 val = test_bit(__obj_to_index(s, addr, p), map) ? 4691 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE; 4692 4693 if (!check_object(s, page, p, val)) 4694 break; 4695 } 4696 put_map(map); 4697 unlock: 4698 slab_unlock(page); 4699 } 4700 4701 static int validate_slab_node(struct kmem_cache *s, 4702 struct kmem_cache_node *n) 4703 { 4704 unsigned long count = 0; 4705 struct page *page; 4706 unsigned long flags; 4707 4708 spin_lock_irqsave(&n->list_lock, flags); 4709 4710 list_for_each_entry(page, &n->partial, slab_list) { 4711 validate_slab(s, page); 4712 count++; 4713 } 4714 if (count != n->nr_partial) { 4715 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n", 4716 s->name, count, n->nr_partial); 4717 slab_add_kunit_errors(); 4718 } 4719 4720 if (!(s->flags & SLAB_STORE_USER)) 4721 goto out; 4722 4723 list_for_each_entry(page, &n->full, slab_list) { 4724 validate_slab(s, page); 4725 count++; 4726 } 4727 if (count != atomic_long_read(&n->nr_slabs)) { 4728 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n", 4729 s->name, count, atomic_long_read(&n->nr_slabs)); 4730 slab_add_kunit_errors(); 4731 } 4732 4733 out: 4734 spin_unlock_irqrestore(&n->list_lock, flags); 4735 return count; 4736 } 4737 4738 long validate_slab_cache(struct kmem_cache *s) 4739 { 4740 int node; 4741 unsigned long count = 0; 4742 struct kmem_cache_node *n; 4743 4744 flush_all(s); 4745 for_each_kmem_cache_node(s, node, n) 4746 count += validate_slab_node(s, n); 4747 4748 return count; 4749 } 4750 EXPORT_SYMBOL(validate_slab_cache); 4751 4752 #ifdef CONFIG_DEBUG_FS 4753 /* 4754 * Generate lists of code addresses where slabcache objects are allocated 4755 * and freed. 4756 */ 4757 4758 struct location { 4759 unsigned long count; 4760 unsigned long addr; 4761 long long sum_time; 4762 long min_time; 4763 long max_time; 4764 long min_pid; 4765 long max_pid; 4766 DECLARE_BITMAP(cpus, NR_CPUS); 4767 nodemask_t nodes; 4768 }; 4769 4770 struct loc_track { 4771 unsigned long max; 4772 unsigned long count; 4773 struct location *loc; 4774 }; 4775 4776 static struct dentry *slab_debugfs_root; 4777 4778 static void free_loc_track(struct loc_track *t) 4779 { 4780 if (t->max) 4781 free_pages((unsigned long)t->loc, 4782 get_order(sizeof(struct location) * t->max)); 4783 } 4784 4785 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 4786 { 4787 struct location *l; 4788 int order; 4789 4790 order = get_order(sizeof(struct location) * max); 4791 4792 l = (void *)__get_free_pages(flags, order); 4793 if (!l) 4794 return 0; 4795 4796 if (t->count) { 4797 memcpy(l, t->loc, sizeof(struct location) * t->count); 4798 free_loc_track(t); 4799 } 4800 t->max = max; 4801 t->loc = l; 4802 return 1; 4803 } 4804 4805 static int add_location(struct loc_track *t, struct kmem_cache *s, 4806 const struct track *track) 4807 { 4808 long start, end, pos; 4809 struct location *l; 4810 unsigned long caddr; 4811 unsigned long age = jiffies - track->when; 4812 4813 start = -1; 4814 end = t->count; 4815 4816 for ( ; ; ) { 4817 pos = start + (end - start + 1) / 2; 4818 4819 /* 4820 * There is nothing at "end". If we end up there 4821 * we need to add something to before end. 4822 */ 4823 if (pos == end) 4824 break; 4825 4826 caddr = t->loc[pos].addr; 4827 if (track->addr == caddr) { 4828 4829 l = &t->loc[pos]; 4830 l->count++; 4831 if (track->when) { 4832 l->sum_time += age; 4833 if (age < l->min_time) 4834 l->min_time = age; 4835 if (age > l->max_time) 4836 l->max_time = age; 4837 4838 if (track->pid < l->min_pid) 4839 l->min_pid = track->pid; 4840 if (track->pid > l->max_pid) 4841 l->max_pid = track->pid; 4842 4843 cpumask_set_cpu(track->cpu, 4844 to_cpumask(l->cpus)); 4845 } 4846 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4847 return 1; 4848 } 4849 4850 if (track->addr < caddr) 4851 end = pos; 4852 else 4853 start = pos; 4854 } 4855 4856 /* 4857 * Not found. Insert new tracking element. 4858 */ 4859 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 4860 return 0; 4861 4862 l = t->loc + pos; 4863 if (pos < t->count) 4864 memmove(l + 1, l, 4865 (t->count - pos) * sizeof(struct location)); 4866 t->count++; 4867 l->count = 1; 4868 l->addr = track->addr; 4869 l->sum_time = age; 4870 l->min_time = age; 4871 l->max_time = age; 4872 l->min_pid = track->pid; 4873 l->max_pid = track->pid; 4874 cpumask_clear(to_cpumask(l->cpus)); 4875 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 4876 nodes_clear(l->nodes); 4877 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4878 return 1; 4879 } 4880 4881 static void process_slab(struct loc_track *t, struct kmem_cache *s, 4882 struct page *page, enum track_item alloc) 4883 { 4884 void *addr = page_address(page); 4885 void *p; 4886 unsigned long *map; 4887 4888 map = get_map(s, page); 4889 for_each_object(p, s, addr, page->objects) 4890 if (!test_bit(__obj_to_index(s, addr, p), map)) 4891 add_location(t, s, get_track(s, p, alloc)); 4892 put_map(map); 4893 } 4894 #endif /* CONFIG_DEBUG_FS */ 4895 #endif /* CONFIG_SLUB_DEBUG */ 4896 4897 #ifdef CONFIG_SYSFS 4898 enum slab_stat_type { 4899 SL_ALL, /* All slabs */ 4900 SL_PARTIAL, /* Only partially allocated slabs */ 4901 SL_CPU, /* Only slabs used for cpu caches */ 4902 SL_OBJECTS, /* Determine allocated objects not slabs */ 4903 SL_TOTAL /* Determine object capacity not slabs */ 4904 }; 4905 4906 #define SO_ALL (1 << SL_ALL) 4907 #define SO_PARTIAL (1 << SL_PARTIAL) 4908 #define SO_CPU (1 << SL_CPU) 4909 #define SO_OBJECTS (1 << SL_OBJECTS) 4910 #define SO_TOTAL (1 << SL_TOTAL) 4911 4912 static ssize_t show_slab_objects(struct kmem_cache *s, 4913 char *buf, unsigned long flags) 4914 { 4915 unsigned long total = 0; 4916 int node; 4917 int x; 4918 unsigned long *nodes; 4919 int len = 0; 4920 4921 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL); 4922 if (!nodes) 4923 return -ENOMEM; 4924 4925 if (flags & SO_CPU) { 4926 int cpu; 4927 4928 for_each_possible_cpu(cpu) { 4929 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, 4930 cpu); 4931 int node; 4932 struct page *page; 4933 4934 page = READ_ONCE(c->page); 4935 if (!page) 4936 continue; 4937 4938 node = page_to_nid(page); 4939 if (flags & SO_TOTAL) 4940 x = page->objects; 4941 else if (flags & SO_OBJECTS) 4942 x = page->inuse; 4943 else 4944 x = 1; 4945 4946 total += x; 4947 nodes[node] += x; 4948 4949 page = slub_percpu_partial_read_once(c); 4950 if (page) { 4951 node = page_to_nid(page); 4952 if (flags & SO_TOTAL) 4953 WARN_ON_ONCE(1); 4954 else if (flags & SO_OBJECTS) 4955 WARN_ON_ONCE(1); 4956 else 4957 x = page->pages; 4958 total += x; 4959 nodes[node] += x; 4960 } 4961 } 4962 } 4963 4964 /* 4965 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex" 4966 * already held which will conflict with an existing lock order: 4967 * 4968 * mem_hotplug_lock->slab_mutex->kernfs_mutex 4969 * 4970 * We don't really need mem_hotplug_lock (to hold off 4971 * slab_mem_going_offline_callback) here because slab's memory hot 4972 * unplug code doesn't destroy the kmem_cache->node[] data. 4973 */ 4974 4975 #ifdef CONFIG_SLUB_DEBUG 4976 if (flags & SO_ALL) { 4977 struct kmem_cache_node *n; 4978 4979 for_each_kmem_cache_node(s, node, n) { 4980 4981 if (flags & SO_TOTAL) 4982 x = atomic_long_read(&n->total_objects); 4983 else if (flags & SO_OBJECTS) 4984 x = atomic_long_read(&n->total_objects) - 4985 count_partial(n, count_free); 4986 else 4987 x = atomic_long_read(&n->nr_slabs); 4988 total += x; 4989 nodes[node] += x; 4990 } 4991 4992 } else 4993 #endif 4994 if (flags & SO_PARTIAL) { 4995 struct kmem_cache_node *n; 4996 4997 for_each_kmem_cache_node(s, node, n) { 4998 if (flags & SO_TOTAL) 4999 x = count_partial(n, count_total); 5000 else if (flags & SO_OBJECTS) 5001 x = count_partial(n, count_inuse); 5002 else 5003 x = n->nr_partial; 5004 total += x; 5005 nodes[node] += x; 5006 } 5007 } 5008 5009 len += sysfs_emit_at(buf, len, "%lu", total); 5010 #ifdef CONFIG_NUMA 5011 for (node = 0; node < nr_node_ids; node++) { 5012 if (nodes[node]) 5013 len += sysfs_emit_at(buf, len, " N%d=%lu", 5014 node, nodes[node]); 5015 } 5016 #endif 5017 len += sysfs_emit_at(buf, len, "\n"); 5018 kfree(nodes); 5019 5020 return len; 5021 } 5022 5023 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 5024 #define to_slab(n) container_of(n, struct kmem_cache, kobj) 5025 5026 struct slab_attribute { 5027 struct attribute attr; 5028 ssize_t (*show)(struct kmem_cache *s, char *buf); 5029 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 5030 }; 5031 5032 #define SLAB_ATTR_RO(_name) \ 5033 static struct slab_attribute _name##_attr = \ 5034 __ATTR(_name, 0400, _name##_show, NULL) 5035 5036 #define SLAB_ATTR(_name) \ 5037 static struct slab_attribute _name##_attr = \ 5038 __ATTR(_name, 0600, _name##_show, _name##_store) 5039 5040 static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 5041 { 5042 return sysfs_emit(buf, "%u\n", s->size); 5043 } 5044 SLAB_ATTR_RO(slab_size); 5045 5046 static ssize_t align_show(struct kmem_cache *s, char *buf) 5047 { 5048 return sysfs_emit(buf, "%u\n", s->align); 5049 } 5050 SLAB_ATTR_RO(align); 5051 5052 static ssize_t object_size_show(struct kmem_cache *s, char *buf) 5053 { 5054 return sysfs_emit(buf, "%u\n", s->object_size); 5055 } 5056 SLAB_ATTR_RO(object_size); 5057 5058 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 5059 { 5060 return sysfs_emit(buf, "%u\n", oo_objects(s->oo)); 5061 } 5062 SLAB_ATTR_RO(objs_per_slab); 5063 5064 static ssize_t order_show(struct kmem_cache *s, char *buf) 5065 { 5066 return sysfs_emit(buf, "%u\n", oo_order(s->oo)); 5067 } 5068 SLAB_ATTR_RO(order); 5069 5070 static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 5071 { 5072 return sysfs_emit(buf, "%lu\n", s->min_partial); 5073 } 5074 5075 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 5076 size_t length) 5077 { 5078 unsigned long min; 5079 int err; 5080 5081 err = kstrtoul(buf, 10, &min); 5082 if (err) 5083 return err; 5084 5085 set_min_partial(s, min); 5086 return length; 5087 } 5088 SLAB_ATTR(min_partial); 5089 5090 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) 5091 { 5092 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s)); 5093 } 5094 5095 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, 5096 size_t length) 5097 { 5098 unsigned int objects; 5099 int err; 5100 5101 err = kstrtouint(buf, 10, &objects); 5102 if (err) 5103 return err; 5104 if (objects && !kmem_cache_has_cpu_partial(s)) 5105 return -EINVAL; 5106 5107 slub_set_cpu_partial(s, objects); 5108 flush_all(s); 5109 return length; 5110 } 5111 SLAB_ATTR(cpu_partial); 5112 5113 static ssize_t ctor_show(struct kmem_cache *s, char *buf) 5114 { 5115 if (!s->ctor) 5116 return 0; 5117 return sysfs_emit(buf, "%pS\n", s->ctor); 5118 } 5119 SLAB_ATTR_RO(ctor); 5120 5121 static ssize_t aliases_show(struct kmem_cache *s, char *buf) 5122 { 5123 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1); 5124 } 5125 SLAB_ATTR_RO(aliases); 5126 5127 static ssize_t partial_show(struct kmem_cache *s, char *buf) 5128 { 5129 return show_slab_objects(s, buf, SO_PARTIAL); 5130 } 5131 SLAB_ATTR_RO(partial); 5132 5133 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 5134 { 5135 return show_slab_objects(s, buf, SO_CPU); 5136 } 5137 SLAB_ATTR_RO(cpu_slabs); 5138 5139 static ssize_t objects_show(struct kmem_cache *s, char *buf) 5140 { 5141 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 5142 } 5143 SLAB_ATTR_RO(objects); 5144 5145 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 5146 { 5147 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 5148 } 5149 SLAB_ATTR_RO(objects_partial); 5150 5151 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) 5152 { 5153 int objects = 0; 5154 int pages = 0; 5155 int cpu; 5156 int len = 0; 5157 5158 for_each_online_cpu(cpu) { 5159 struct page *page; 5160 5161 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); 5162 5163 if (page) { 5164 pages += page->pages; 5165 objects += page->pobjects; 5166 } 5167 } 5168 5169 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages); 5170 5171 #ifdef CONFIG_SMP 5172 for_each_online_cpu(cpu) { 5173 struct page *page; 5174 5175 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); 5176 if (page) 5177 len += sysfs_emit_at(buf, len, " C%d=%d(%d)", 5178 cpu, page->pobjects, page->pages); 5179 } 5180 #endif 5181 len += sysfs_emit_at(buf, len, "\n"); 5182 5183 return len; 5184 } 5185 SLAB_ATTR_RO(slabs_cpu_partial); 5186 5187 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 5188 { 5189 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 5190 } 5191 SLAB_ATTR_RO(reclaim_account); 5192 5193 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 5194 { 5195 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 5196 } 5197 SLAB_ATTR_RO(hwcache_align); 5198 5199 #ifdef CONFIG_ZONE_DMA 5200 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 5201 { 5202 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 5203 } 5204 SLAB_ATTR_RO(cache_dma); 5205 #endif 5206 5207 static ssize_t usersize_show(struct kmem_cache *s, char *buf) 5208 { 5209 return sysfs_emit(buf, "%u\n", s->usersize); 5210 } 5211 SLAB_ATTR_RO(usersize); 5212 5213 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 5214 { 5215 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU)); 5216 } 5217 SLAB_ATTR_RO(destroy_by_rcu); 5218 5219 #ifdef CONFIG_SLUB_DEBUG 5220 static ssize_t slabs_show(struct kmem_cache *s, char *buf) 5221 { 5222 return show_slab_objects(s, buf, SO_ALL); 5223 } 5224 SLAB_ATTR_RO(slabs); 5225 5226 static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 5227 { 5228 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 5229 } 5230 SLAB_ATTR_RO(total_objects); 5231 5232 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 5233 { 5234 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS)); 5235 } 5236 SLAB_ATTR_RO(sanity_checks); 5237 5238 static ssize_t trace_show(struct kmem_cache *s, char *buf) 5239 { 5240 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 5241 } 5242 SLAB_ATTR_RO(trace); 5243 5244 static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 5245 { 5246 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 5247 } 5248 5249 SLAB_ATTR_RO(red_zone); 5250 5251 static ssize_t poison_show(struct kmem_cache *s, char *buf) 5252 { 5253 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON)); 5254 } 5255 5256 SLAB_ATTR_RO(poison); 5257 5258 static ssize_t store_user_show(struct kmem_cache *s, char *buf) 5259 { 5260 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 5261 } 5262 5263 SLAB_ATTR_RO(store_user); 5264 5265 static ssize_t validate_show(struct kmem_cache *s, char *buf) 5266 { 5267 return 0; 5268 } 5269 5270 static ssize_t validate_store(struct kmem_cache *s, 5271 const char *buf, size_t length) 5272 { 5273 int ret = -EINVAL; 5274 5275 if (buf[0] == '1') { 5276 ret = validate_slab_cache(s); 5277 if (ret >= 0) 5278 ret = length; 5279 } 5280 return ret; 5281 } 5282 SLAB_ATTR(validate); 5283 5284 #endif /* CONFIG_SLUB_DEBUG */ 5285 5286 #ifdef CONFIG_FAILSLAB 5287 static ssize_t failslab_show(struct kmem_cache *s, char *buf) 5288 { 5289 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 5290 } 5291 SLAB_ATTR_RO(failslab); 5292 #endif 5293 5294 static ssize_t shrink_show(struct kmem_cache *s, char *buf) 5295 { 5296 return 0; 5297 } 5298 5299 static ssize_t shrink_store(struct kmem_cache *s, 5300 const char *buf, size_t length) 5301 { 5302 if (buf[0] == '1') 5303 kmem_cache_shrink(s); 5304 else 5305 return -EINVAL; 5306 return length; 5307 } 5308 SLAB_ATTR(shrink); 5309 5310 #ifdef CONFIG_NUMA 5311 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 5312 { 5313 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10); 5314 } 5315 5316 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 5317 const char *buf, size_t length) 5318 { 5319 unsigned int ratio; 5320 int err; 5321 5322 err = kstrtouint(buf, 10, &ratio); 5323 if (err) 5324 return err; 5325 if (ratio > 100) 5326 return -ERANGE; 5327 5328 s->remote_node_defrag_ratio = ratio * 10; 5329 5330 return length; 5331 } 5332 SLAB_ATTR(remote_node_defrag_ratio); 5333 #endif 5334 5335 #ifdef CONFIG_SLUB_STATS 5336 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 5337 { 5338 unsigned long sum = 0; 5339 int cpu; 5340 int len = 0; 5341 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL); 5342 5343 if (!data) 5344 return -ENOMEM; 5345 5346 for_each_online_cpu(cpu) { 5347 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 5348 5349 data[cpu] = x; 5350 sum += x; 5351 } 5352 5353 len += sysfs_emit_at(buf, len, "%lu", sum); 5354 5355 #ifdef CONFIG_SMP 5356 for_each_online_cpu(cpu) { 5357 if (data[cpu]) 5358 len += sysfs_emit_at(buf, len, " C%d=%u", 5359 cpu, data[cpu]); 5360 } 5361 #endif 5362 kfree(data); 5363 len += sysfs_emit_at(buf, len, "\n"); 5364 5365 return len; 5366 } 5367 5368 static void clear_stat(struct kmem_cache *s, enum stat_item si) 5369 { 5370 int cpu; 5371 5372 for_each_online_cpu(cpu) 5373 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 5374 } 5375 5376 #define STAT_ATTR(si, text) \ 5377 static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 5378 { \ 5379 return show_stat(s, buf, si); \ 5380 } \ 5381 static ssize_t text##_store(struct kmem_cache *s, \ 5382 const char *buf, size_t length) \ 5383 { \ 5384 if (buf[0] != '0') \ 5385 return -EINVAL; \ 5386 clear_stat(s, si); \ 5387 return length; \ 5388 } \ 5389 SLAB_ATTR(text); \ 5390 5391 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 5392 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 5393 STAT_ATTR(FREE_FASTPATH, free_fastpath); 5394 STAT_ATTR(FREE_SLOWPATH, free_slowpath); 5395 STAT_ATTR(FREE_FROZEN, free_frozen); 5396 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 5397 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 5398 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 5399 STAT_ATTR(ALLOC_SLAB, alloc_slab); 5400 STAT_ATTR(ALLOC_REFILL, alloc_refill); 5401 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); 5402 STAT_ATTR(FREE_SLAB, free_slab); 5403 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 5404 STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 5405 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 5406 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 5407 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 5408 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 5409 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); 5410 STAT_ATTR(ORDER_FALLBACK, order_fallback); 5411 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); 5412 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); 5413 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); 5414 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); 5415 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); 5416 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); 5417 #endif /* CONFIG_SLUB_STATS */ 5418 5419 static struct attribute *slab_attrs[] = { 5420 &slab_size_attr.attr, 5421 &object_size_attr.attr, 5422 &objs_per_slab_attr.attr, 5423 &order_attr.attr, 5424 &min_partial_attr.attr, 5425 &cpu_partial_attr.attr, 5426 &objects_attr.attr, 5427 &objects_partial_attr.attr, 5428 &partial_attr.attr, 5429 &cpu_slabs_attr.attr, 5430 &ctor_attr.attr, 5431 &aliases_attr.attr, 5432 &align_attr.attr, 5433 &hwcache_align_attr.attr, 5434 &reclaim_account_attr.attr, 5435 &destroy_by_rcu_attr.attr, 5436 &shrink_attr.attr, 5437 &slabs_cpu_partial_attr.attr, 5438 #ifdef CONFIG_SLUB_DEBUG 5439 &total_objects_attr.attr, 5440 &slabs_attr.attr, 5441 &sanity_checks_attr.attr, 5442 &trace_attr.attr, 5443 &red_zone_attr.attr, 5444 &poison_attr.attr, 5445 &store_user_attr.attr, 5446 &validate_attr.attr, 5447 #endif 5448 #ifdef CONFIG_ZONE_DMA 5449 &cache_dma_attr.attr, 5450 #endif 5451 #ifdef CONFIG_NUMA 5452 &remote_node_defrag_ratio_attr.attr, 5453 #endif 5454 #ifdef CONFIG_SLUB_STATS 5455 &alloc_fastpath_attr.attr, 5456 &alloc_slowpath_attr.attr, 5457 &free_fastpath_attr.attr, 5458 &free_slowpath_attr.attr, 5459 &free_frozen_attr.attr, 5460 &free_add_partial_attr.attr, 5461 &free_remove_partial_attr.attr, 5462 &alloc_from_partial_attr.attr, 5463 &alloc_slab_attr.attr, 5464 &alloc_refill_attr.attr, 5465 &alloc_node_mismatch_attr.attr, 5466 &free_slab_attr.attr, 5467 &cpuslab_flush_attr.attr, 5468 &deactivate_full_attr.attr, 5469 &deactivate_empty_attr.attr, 5470 &deactivate_to_head_attr.attr, 5471 &deactivate_to_tail_attr.attr, 5472 &deactivate_remote_frees_attr.attr, 5473 &deactivate_bypass_attr.attr, 5474 &order_fallback_attr.attr, 5475 &cmpxchg_double_fail_attr.attr, 5476 &cmpxchg_double_cpu_fail_attr.attr, 5477 &cpu_partial_alloc_attr.attr, 5478 &cpu_partial_free_attr.attr, 5479 &cpu_partial_node_attr.attr, 5480 &cpu_partial_drain_attr.attr, 5481 #endif 5482 #ifdef CONFIG_FAILSLAB 5483 &failslab_attr.attr, 5484 #endif 5485 &usersize_attr.attr, 5486 5487 NULL 5488 }; 5489 5490 static const struct attribute_group slab_attr_group = { 5491 .attrs = slab_attrs, 5492 }; 5493 5494 static ssize_t slab_attr_show(struct kobject *kobj, 5495 struct attribute *attr, 5496 char *buf) 5497 { 5498 struct slab_attribute *attribute; 5499 struct kmem_cache *s; 5500 int err; 5501 5502 attribute = to_slab_attr(attr); 5503 s = to_slab(kobj); 5504 5505 if (!attribute->show) 5506 return -EIO; 5507 5508 err = attribute->show(s, buf); 5509 5510 return err; 5511 } 5512 5513 static ssize_t slab_attr_store(struct kobject *kobj, 5514 struct attribute *attr, 5515 const char *buf, size_t len) 5516 { 5517 struct slab_attribute *attribute; 5518 struct kmem_cache *s; 5519 int err; 5520 5521 attribute = to_slab_attr(attr); 5522 s = to_slab(kobj); 5523 5524 if (!attribute->store) 5525 return -EIO; 5526 5527 err = attribute->store(s, buf, len); 5528 return err; 5529 } 5530 5531 static void kmem_cache_release(struct kobject *k) 5532 { 5533 slab_kmem_cache_release(to_slab(k)); 5534 } 5535 5536 static const struct sysfs_ops slab_sysfs_ops = { 5537 .show = slab_attr_show, 5538 .store = slab_attr_store, 5539 }; 5540 5541 static struct kobj_type slab_ktype = { 5542 .sysfs_ops = &slab_sysfs_ops, 5543 .release = kmem_cache_release, 5544 }; 5545 5546 static struct kset *slab_kset; 5547 5548 static inline struct kset *cache_kset(struct kmem_cache *s) 5549 { 5550 return slab_kset; 5551 } 5552 5553 #define ID_STR_LENGTH 64 5554 5555 /* Create a unique string id for a slab cache: 5556 * 5557 * Format :[flags-]size 5558 */ 5559 static char *create_unique_id(struct kmem_cache *s) 5560 { 5561 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 5562 char *p = name; 5563 5564 BUG_ON(!name); 5565 5566 *p++ = ':'; 5567 /* 5568 * First flags affecting slabcache operations. We will only 5569 * get here for aliasable slabs so we do not need to support 5570 * too many flags. The flags here must cover all flags that 5571 * are matched during merging to guarantee that the id is 5572 * unique. 5573 */ 5574 if (s->flags & SLAB_CACHE_DMA) 5575 *p++ = 'd'; 5576 if (s->flags & SLAB_CACHE_DMA32) 5577 *p++ = 'D'; 5578 if (s->flags & SLAB_RECLAIM_ACCOUNT) 5579 *p++ = 'a'; 5580 if (s->flags & SLAB_CONSISTENCY_CHECKS) 5581 *p++ = 'F'; 5582 if (s->flags & SLAB_ACCOUNT) 5583 *p++ = 'A'; 5584 if (p != name + 1) 5585 *p++ = '-'; 5586 p += sprintf(p, "%07u", s->size); 5587 5588 BUG_ON(p > name + ID_STR_LENGTH - 1); 5589 return name; 5590 } 5591 5592 static int sysfs_slab_add(struct kmem_cache *s) 5593 { 5594 int err; 5595 const char *name; 5596 struct kset *kset = cache_kset(s); 5597 int unmergeable = slab_unmergeable(s); 5598 5599 if (!kset) { 5600 kobject_init(&s->kobj, &slab_ktype); 5601 return 0; 5602 } 5603 5604 if (!unmergeable && disable_higher_order_debug && 5605 (slub_debug & DEBUG_METADATA_FLAGS)) 5606 unmergeable = 1; 5607 5608 if (unmergeable) { 5609 /* 5610 * Slabcache can never be merged so we can use the name proper. 5611 * This is typically the case for debug situations. In that 5612 * case we can catch duplicate names easily. 5613 */ 5614 sysfs_remove_link(&slab_kset->kobj, s->name); 5615 name = s->name; 5616 } else { 5617 /* 5618 * Create a unique name for the slab as a target 5619 * for the symlinks. 5620 */ 5621 name = create_unique_id(s); 5622 } 5623 5624 s->kobj.kset = kset; 5625 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); 5626 if (err) 5627 goto out; 5628 5629 err = sysfs_create_group(&s->kobj, &slab_attr_group); 5630 if (err) 5631 goto out_del_kobj; 5632 5633 if (!unmergeable) { 5634 /* Setup first alias */ 5635 sysfs_slab_alias(s, s->name); 5636 } 5637 out: 5638 if (!unmergeable) 5639 kfree(name); 5640 return err; 5641 out_del_kobj: 5642 kobject_del(&s->kobj); 5643 goto out; 5644 } 5645 5646 void sysfs_slab_unlink(struct kmem_cache *s) 5647 { 5648 if (slab_state >= FULL) 5649 kobject_del(&s->kobj); 5650 } 5651 5652 void sysfs_slab_release(struct kmem_cache *s) 5653 { 5654 if (slab_state >= FULL) 5655 kobject_put(&s->kobj); 5656 } 5657 5658 /* 5659 * Need to buffer aliases during bootup until sysfs becomes 5660 * available lest we lose that information. 5661 */ 5662 struct saved_alias { 5663 struct kmem_cache *s; 5664 const char *name; 5665 struct saved_alias *next; 5666 }; 5667 5668 static struct saved_alias *alias_list; 5669 5670 static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 5671 { 5672 struct saved_alias *al; 5673 5674 if (slab_state == FULL) { 5675 /* 5676 * If we have a leftover link then remove it. 5677 */ 5678 sysfs_remove_link(&slab_kset->kobj, name); 5679 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 5680 } 5681 5682 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 5683 if (!al) 5684 return -ENOMEM; 5685 5686 al->s = s; 5687 al->name = name; 5688 al->next = alias_list; 5689 alias_list = al; 5690 return 0; 5691 } 5692 5693 static int __init slab_sysfs_init(void) 5694 { 5695 struct kmem_cache *s; 5696 int err; 5697 5698 mutex_lock(&slab_mutex); 5699 5700 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj); 5701 if (!slab_kset) { 5702 mutex_unlock(&slab_mutex); 5703 pr_err("Cannot register slab subsystem.\n"); 5704 return -ENOSYS; 5705 } 5706 5707 slab_state = FULL; 5708 5709 list_for_each_entry(s, &slab_caches, list) { 5710 err = sysfs_slab_add(s); 5711 if (err) 5712 pr_err("SLUB: Unable to add boot slab %s to sysfs\n", 5713 s->name); 5714 } 5715 5716 while (alias_list) { 5717 struct saved_alias *al = alias_list; 5718 5719 alias_list = alias_list->next; 5720 err = sysfs_slab_alias(al->s, al->name); 5721 if (err) 5722 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n", 5723 al->name); 5724 kfree(al); 5725 } 5726 5727 mutex_unlock(&slab_mutex); 5728 return 0; 5729 } 5730 5731 __initcall(slab_sysfs_init); 5732 #endif /* CONFIG_SYSFS */ 5733 5734 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS) 5735 static int slab_debugfs_show(struct seq_file *seq, void *v) 5736 { 5737 5738 struct location *l; 5739 unsigned int idx = *(unsigned int *)v; 5740 struct loc_track *t = seq->private; 5741 5742 if (idx < t->count) { 5743 l = &t->loc[idx]; 5744 5745 seq_printf(seq, "%7ld ", l->count); 5746 5747 if (l->addr) 5748 seq_printf(seq, "%pS", (void *)l->addr); 5749 else 5750 seq_puts(seq, "<not-available>"); 5751 5752 if (l->sum_time != l->min_time) { 5753 seq_printf(seq, " age=%ld/%llu/%ld", 5754 l->min_time, div_u64(l->sum_time, l->count), 5755 l->max_time); 5756 } else 5757 seq_printf(seq, " age=%ld", l->min_time); 5758 5759 if (l->min_pid != l->max_pid) 5760 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid); 5761 else 5762 seq_printf(seq, " pid=%ld", 5763 l->min_pid); 5764 5765 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus))) 5766 seq_printf(seq, " cpus=%*pbl", 5767 cpumask_pr_args(to_cpumask(l->cpus))); 5768 5769 if (nr_online_nodes > 1 && !nodes_empty(l->nodes)) 5770 seq_printf(seq, " nodes=%*pbl", 5771 nodemask_pr_args(&l->nodes)); 5772 5773 seq_puts(seq, "\n"); 5774 } 5775 5776 if (!idx && !t->count) 5777 seq_puts(seq, "No data\n"); 5778 5779 return 0; 5780 } 5781 5782 static void slab_debugfs_stop(struct seq_file *seq, void *v) 5783 { 5784 } 5785 5786 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos) 5787 { 5788 struct loc_track *t = seq->private; 5789 5790 v = ppos; 5791 ++*ppos; 5792 if (*ppos <= t->count) 5793 return v; 5794 5795 return NULL; 5796 } 5797 5798 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos) 5799 { 5800 return ppos; 5801 } 5802 5803 static const struct seq_operations slab_debugfs_sops = { 5804 .start = slab_debugfs_start, 5805 .next = slab_debugfs_next, 5806 .stop = slab_debugfs_stop, 5807 .show = slab_debugfs_show, 5808 }; 5809 5810 static int slab_debug_trace_open(struct inode *inode, struct file *filep) 5811 { 5812 5813 struct kmem_cache_node *n; 5814 enum track_item alloc; 5815 int node; 5816 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops, 5817 sizeof(struct loc_track)); 5818 struct kmem_cache *s = file_inode(filep)->i_private; 5819 5820 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0) 5821 alloc = TRACK_ALLOC; 5822 else 5823 alloc = TRACK_FREE; 5824 5825 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) 5826 return -ENOMEM; 5827 5828 /* Push back cpu slabs */ 5829 flush_all(s); 5830 5831 for_each_kmem_cache_node(s, node, n) { 5832 unsigned long flags; 5833 struct page *page; 5834 5835 if (!atomic_long_read(&n->nr_slabs)) 5836 continue; 5837 5838 spin_lock_irqsave(&n->list_lock, flags); 5839 list_for_each_entry(page, &n->partial, slab_list) 5840 process_slab(t, s, page, alloc); 5841 list_for_each_entry(page, &n->full, slab_list) 5842 process_slab(t, s, page, alloc); 5843 spin_unlock_irqrestore(&n->list_lock, flags); 5844 } 5845 5846 return 0; 5847 } 5848 5849 static int slab_debug_trace_release(struct inode *inode, struct file *file) 5850 { 5851 struct seq_file *seq = file->private_data; 5852 struct loc_track *t = seq->private; 5853 5854 free_loc_track(t); 5855 return seq_release_private(inode, file); 5856 } 5857 5858 static const struct file_operations slab_debugfs_fops = { 5859 .open = slab_debug_trace_open, 5860 .read = seq_read, 5861 .llseek = seq_lseek, 5862 .release = slab_debug_trace_release, 5863 }; 5864 5865 static void debugfs_slab_add(struct kmem_cache *s) 5866 { 5867 struct dentry *slab_cache_dir; 5868 5869 if (unlikely(!slab_debugfs_root)) 5870 return; 5871 5872 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root); 5873 5874 debugfs_create_file("alloc_traces", 0400, 5875 slab_cache_dir, s, &slab_debugfs_fops); 5876 5877 debugfs_create_file("free_traces", 0400, 5878 slab_cache_dir, s, &slab_debugfs_fops); 5879 } 5880 5881 void debugfs_slab_release(struct kmem_cache *s) 5882 { 5883 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root)); 5884 } 5885 5886 static int __init slab_debugfs_init(void) 5887 { 5888 struct kmem_cache *s; 5889 5890 slab_debugfs_root = debugfs_create_dir("slab", NULL); 5891 5892 list_for_each_entry(s, &slab_caches, list) 5893 if (s->flags & SLAB_STORE_USER) 5894 debugfs_slab_add(s); 5895 5896 return 0; 5897 5898 } 5899 __initcall(slab_debugfs_init); 5900 #endif 5901 /* 5902 * The /proc/slabinfo ABI 5903 */ 5904 #ifdef CONFIG_SLUB_DEBUG 5905 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) 5906 { 5907 unsigned long nr_slabs = 0; 5908 unsigned long nr_objs = 0; 5909 unsigned long nr_free = 0; 5910 int node; 5911 struct kmem_cache_node *n; 5912 5913 for_each_kmem_cache_node(s, node, n) { 5914 nr_slabs += node_nr_slabs(n); 5915 nr_objs += node_nr_objs(n); 5916 nr_free += count_partial(n, count_free); 5917 } 5918 5919 sinfo->active_objs = nr_objs - nr_free; 5920 sinfo->num_objs = nr_objs; 5921 sinfo->active_slabs = nr_slabs; 5922 sinfo->num_slabs = nr_slabs; 5923 sinfo->objects_per_slab = oo_objects(s->oo); 5924 sinfo->cache_order = oo_order(s->oo); 5925 } 5926 5927 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) 5928 { 5929 } 5930 5931 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 5932 size_t count, loff_t *ppos) 5933 { 5934 return -EIO; 5935 } 5936 #endif /* CONFIG_SLUB_DEBUG */ 5937