slub.c revision d18a90dd85f8243ed20cdadb6d8a37d595df456d
1/* 2 * SLUB: A slab allocator that limits cache line use instead of queuing 3 * objects in per cpu and per node lists. 4 * 5 * The allocator synchronizes using per slab locks and only 6 * uses a centralized lock to manage a pool of partial slabs. 7 * 8 * (C) 2007 SGI, Christoph Lameter 9 */ 10 11#include <linux/mm.h> 12#include <linux/swap.h> /* struct reclaim_state */ 13#include <linux/module.h> 14#include <linux/bit_spinlock.h> 15#include <linux/interrupt.h> 16#include <linux/bitops.h> 17#include <linux/slab.h> 18#include <linux/proc_fs.h> 19#include <linux/seq_file.h> 20#include <linux/kmemcheck.h> 21#include <linux/cpu.h> 22#include <linux/cpuset.h> 23#include <linux/mempolicy.h> 24#include <linux/ctype.h> 25#include <linux/debugobjects.h> 26#include <linux/kallsyms.h> 27#include <linux/memory.h> 28#include <linux/math64.h> 29#include <linux/fault-inject.h> 30 31#include <trace/events/kmem.h> 32 33/* 34 * Lock order: 35 * 1. slab_lock(page) 36 * 2. slab->list_lock 37 * 38 * The slab_lock protects operations on the object of a particular 39 * slab and its metadata in the page struct. If the slab lock 40 * has been taken then no allocations nor frees can be performed 41 * on the objects in the slab nor can the slab be added or removed 42 * from the partial or full lists since this would mean modifying 43 * the page_struct of the slab. 44 * 45 * The list_lock protects the partial and full list on each node and 46 * the partial slab counter. If taken then no new slabs may be added or 47 * removed from the lists nor make the number of partial slabs be modified. 48 * (Note that the total number of slabs is an atomic value that may be 49 * modified without taking the list lock). 50 * 51 * The list_lock is a centralized lock and thus we avoid taking it as 52 * much as possible. As long as SLUB does not have to handle partial 53 * slabs, operations can continue without any centralized lock. F.e. 54 * allocating a long series of objects that fill up slabs does not require 55 * the list lock. 56 * 57 * The lock order is sometimes inverted when we are trying to get a slab 58 * off a list. We take the list_lock and then look for a page on the list 59 * to use. While we do that objects in the slabs may be freed. We can 60 * only operate on the slab if we have also taken the slab_lock. So we use 61 * a slab_trylock() on the slab. If trylock was successful then no frees 62 * can occur anymore and we can use the slab for allocations etc. If the 63 * slab_trylock() does not succeed then frees are in progress in the slab and 64 * we must stay away from it for a while since we may cause a bouncing 65 * cacheline if we try to acquire the lock. So go onto the next slab. 66 * If all pages are busy then we may allocate a new slab instead of reusing 67 * a partial slab. A new slab has no one operating on it and thus there is 68 * no danger of cacheline contention. 69 * 70 * Interrupts are disabled during allocation and deallocation in order to 71 * make the slab allocator safe to use in the context of an irq. In addition 72 * interrupts are disabled to ensure that the processor does not change 73 * while handling per_cpu slabs, due to kernel preemption. 74 * 75 * SLUB assigns one slab for allocation to each processor. 76 * Allocations only occur from these slabs called cpu slabs. 77 * 78 * Slabs with free elements are kept on a partial list and during regular 79 * operations no list for full slabs is used. If an object in a full slab is 80 * freed then the slab will show up again on the partial lists. 81 * We track full slabs for debugging purposes though because otherwise we 82 * cannot scan all objects. 83 * 84 * Slabs are freed when they become empty. Teardown and setup is 85 * minimal so we rely on the page allocators per cpu caches for 86 * fast frees and allocs. 87 * 88 * Overloading of page flags that are otherwise used for LRU management. 89 * 90 * PageActive The slab is frozen and exempt from list processing. 91 * This means that the slab is dedicated to a purpose 92 * such as satisfying allocations for a specific 93 * processor. Objects may be freed in the slab while 94 * it is frozen but slab_free will then skip the usual 95 * list operations. It is up to the processor holding 96 * the slab to integrate the slab into the slab lists 97 * when the slab is no longer needed. 98 * 99 * One use of this flag is to mark slabs that are 100 * used for allocations. Then such a slab becomes a cpu 101 * slab. The cpu slab may be equipped with an additional 102 * freelist that allows lockless access to 103 * free objects in addition to the regular freelist 104 * that requires the slab lock. 105 * 106 * PageError Slab requires special handling due to debug 107 * options set. This moves slab handling out of 108 * the fast path and disables lockless freelists. 109 */ 110 111#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 112 SLAB_TRACE | SLAB_DEBUG_FREE) 113 114static inline int kmem_cache_debug(struct kmem_cache *s) 115{ 116#ifdef CONFIG_SLUB_DEBUG 117 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 118#else 119 return 0; 120#endif 121} 122 123/* 124 * Issues still to be resolved: 125 * 126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 127 * 128 * - Variable sizing of the per node arrays 129 */ 130 131/* Enable to test recovery from slab corruption on boot */ 132#undef SLUB_RESILIENCY_TEST 133 134/* 135 * Mininum number of partial slabs. These will be left on the partial 136 * lists even if they are empty. kmem_cache_shrink may reclaim them. 137 */ 138#define MIN_PARTIAL 5 139 140/* 141 * Maximum number of desirable partial slabs. 142 * The existence of more partial slabs makes kmem_cache_shrink 143 * sort the partial list by the number of objects in the. 144 */ 145#define MAX_PARTIAL 10 146 147#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 148 SLAB_POISON | SLAB_STORE_USER) 149 150/* 151 * Debugging flags that require metadata to be stored in the slab. These get 152 * disabled when slub_debug=O is used and a cache's min order increases with 153 * metadata. 154 */ 155#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 156 157/* 158 * Set of flags that will prevent slab merging 159 */ 160#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ 162 SLAB_FAILSLAB) 163 164#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 165 SLAB_CACHE_DMA | SLAB_NOTRACK) 166 167#define OO_SHIFT 16 168#define OO_MASK ((1 << OO_SHIFT) - 1) 169#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ 170 171/* Internal SLUB flags */ 172#define __OBJECT_POISON 0x80000000UL /* Poison object */ 173 174static int kmem_size = sizeof(struct kmem_cache); 175 176#ifdef CONFIG_SMP 177static struct notifier_block slab_notifier; 178#endif 179 180static enum { 181 DOWN, /* No slab functionality available */ 182 PARTIAL, /* Kmem_cache_node works */ 183 UP, /* Everything works but does not show up in sysfs */ 184 SYSFS /* Sysfs up */ 185} slab_state = DOWN; 186 187/* A list of all slab caches on the system */ 188static DECLARE_RWSEM(slub_lock); 189static LIST_HEAD(slab_caches); 190 191/* 192 * Tracking user of a slab. 193 */ 194#define TRACK_ADDRS_COUNT 16 195struct track { 196 unsigned long addr; /* Called from address */ 197#ifdef CONFIG_STACKTRACE 198 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 199#endif 200 int cpu; /* Was running on cpu */ 201 int pid; /* Pid context */ 202 unsigned long when; /* When did the operation occur */ 203}; 204 205enum track_item { TRACK_ALLOC, TRACK_FREE }; 206 207#ifdef CONFIG_SYSFS 208static int sysfs_slab_add(struct kmem_cache *); 209static int sysfs_slab_alias(struct kmem_cache *, const char *); 210static void sysfs_slab_remove(struct kmem_cache *); 211 212#else 213static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 214static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 215 { return 0; } 216static inline void sysfs_slab_remove(struct kmem_cache *s) 217{ 218 kfree(s->name); 219 kfree(s); 220} 221 222#endif 223 224static inline void stat(const struct kmem_cache *s, enum stat_item si) 225{ 226#ifdef CONFIG_SLUB_STATS 227 __this_cpu_inc(s->cpu_slab->stat[si]); 228#endif 229} 230 231/******************************************************************** 232 * Core slab cache functions 233 *******************************************************************/ 234 235int slab_is_available(void) 236{ 237 return slab_state >= UP; 238} 239 240static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 241{ 242 return s->node[node]; 243} 244 245/* Verify that a pointer has an address that is valid within a slab page */ 246static inline int check_valid_pointer(struct kmem_cache *s, 247 struct page *page, const void *object) 248{ 249 void *base; 250 251 if (!object) 252 return 1; 253 254 base = page_address(page); 255 if (object < base || object >= base + page->objects * s->size || 256 (object - base) % s->size) { 257 return 0; 258 } 259 260 return 1; 261} 262 263static inline void *get_freepointer(struct kmem_cache *s, void *object) 264{ 265 return *(void **)(object + s->offset); 266} 267 268static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 269{ 270 void *p; 271 272#ifdef CONFIG_DEBUG_PAGEALLOC 273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p)); 274#else 275 p = get_freepointer(s, object); 276#endif 277 return p; 278} 279 280static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 281{ 282 *(void **)(object + s->offset) = fp; 283} 284 285/* Loop over all objects in a slab */ 286#define for_each_object(__p, __s, __addr, __objects) \ 287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 288 __p += (__s)->size) 289 290/* Determine object index from a given position */ 291static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 292{ 293 return (p - addr) / s->size; 294} 295 296static inline size_t slab_ksize(const struct kmem_cache *s) 297{ 298#ifdef CONFIG_SLUB_DEBUG 299 /* 300 * Debugging requires use of the padding between object 301 * and whatever may come after it. 302 */ 303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 304 return s->objsize; 305 306#endif 307 /* 308 * If we have the need to store the freelist pointer 309 * back there or track user information then we can 310 * only use the space before that information. 311 */ 312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 313 return s->inuse; 314 /* 315 * Else we can use all the padding etc for the allocation 316 */ 317 return s->size; 318} 319 320static inline int order_objects(int order, unsigned long size, int reserved) 321{ 322 return ((PAGE_SIZE << order) - reserved) / size; 323} 324 325static inline struct kmem_cache_order_objects oo_make(int order, 326 unsigned long size, int reserved) 327{ 328 struct kmem_cache_order_objects x = { 329 (order << OO_SHIFT) + order_objects(order, size, reserved) 330 }; 331 332 return x; 333} 334 335static inline int oo_order(struct kmem_cache_order_objects x) 336{ 337 return x.x >> OO_SHIFT; 338} 339 340static inline int oo_objects(struct kmem_cache_order_objects x) 341{ 342 return x.x & OO_MASK; 343} 344 345#ifdef CONFIG_SLUB_DEBUG 346/* 347 * Determine a map of object in use on a page. 348 * 349 * Slab lock or node listlock must be held to guarantee that the page does 350 * not vanish from under us. 351 */ 352static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) 353{ 354 void *p; 355 void *addr = page_address(page); 356 357 for (p = page->freelist; p; p = get_freepointer(s, p)) 358 set_bit(slab_index(p, s, addr), map); 359} 360 361/* 362 * Debug settings: 363 */ 364#ifdef CONFIG_SLUB_DEBUG_ON 365static int slub_debug = DEBUG_DEFAULT_FLAGS; 366#else 367static int slub_debug; 368#endif 369 370static char *slub_debug_slabs; 371static int disable_higher_order_debug; 372 373/* 374 * Object debugging 375 */ 376static void print_section(char *text, u8 *addr, unsigned int length) 377{ 378 int i, offset; 379 int newline = 1; 380 char ascii[17]; 381 382 ascii[16] = 0; 383 384 for (i = 0; i < length; i++) { 385 if (newline) { 386 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 387 newline = 0; 388 } 389 printk(KERN_CONT " %02x", addr[i]); 390 offset = i % 16; 391 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 392 if (offset == 15) { 393 printk(KERN_CONT " %s\n", ascii); 394 newline = 1; 395 } 396 } 397 if (!newline) { 398 i %= 16; 399 while (i < 16) { 400 printk(KERN_CONT " "); 401 ascii[i] = ' '; 402 i++; 403 } 404 printk(KERN_CONT " %s\n", ascii); 405 } 406} 407 408static struct track *get_track(struct kmem_cache *s, void *object, 409 enum track_item alloc) 410{ 411 struct track *p; 412 413 if (s->offset) 414 p = object + s->offset + sizeof(void *); 415 else 416 p = object + s->inuse; 417 418 return p + alloc; 419} 420 421static void set_track(struct kmem_cache *s, void *object, 422 enum track_item alloc, unsigned long addr) 423{ 424 struct track *p = get_track(s, object, alloc); 425 426 if (addr) { 427#ifdef CONFIG_STACKTRACE 428 struct stack_trace trace; 429 int i; 430 431 trace.nr_entries = 0; 432 trace.max_entries = TRACK_ADDRS_COUNT; 433 trace.entries = p->addrs; 434 trace.skip = 3; 435 save_stack_trace(&trace); 436 437 /* See rant in lockdep.c */ 438 if (trace.nr_entries != 0 && 439 trace.entries[trace.nr_entries - 1] == ULONG_MAX) 440 trace.nr_entries--; 441 442 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) 443 p->addrs[i] = 0; 444#endif 445 p->addr = addr; 446 p->cpu = smp_processor_id(); 447 p->pid = current->pid; 448 p->when = jiffies; 449 } else 450 memset(p, 0, sizeof(struct track)); 451} 452 453static void init_tracking(struct kmem_cache *s, void *object) 454{ 455 if (!(s->flags & SLAB_STORE_USER)) 456 return; 457 458 set_track(s, object, TRACK_FREE, 0UL); 459 set_track(s, object, TRACK_ALLOC, 0UL); 460} 461 462static void print_track(const char *s, struct track *t) 463{ 464 if (!t->addr) 465 return; 466 467 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 468 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 469#ifdef CONFIG_STACKTRACE 470 { 471 int i; 472 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 473 if (t->addrs[i]) 474 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]); 475 else 476 break; 477 } 478#endif 479} 480 481static void print_tracking(struct kmem_cache *s, void *object) 482{ 483 if (!(s->flags & SLAB_STORE_USER)) 484 return; 485 486 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 487 print_track("Freed", get_track(s, object, TRACK_FREE)); 488} 489 490static void print_page_info(struct page *page) 491{ 492 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 493 page, page->objects, page->inuse, page->freelist, page->flags); 494 495} 496 497static void slab_bug(struct kmem_cache *s, char *fmt, ...) 498{ 499 va_list args; 500 char buf[100]; 501 502 va_start(args, fmt); 503 vsnprintf(buf, sizeof(buf), fmt, args); 504 va_end(args); 505 printk(KERN_ERR "========================================" 506 "=====================================\n"); 507 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 508 printk(KERN_ERR "----------------------------------------" 509 "-------------------------------------\n\n"); 510} 511 512static void slab_fix(struct kmem_cache *s, char *fmt, ...) 513{ 514 va_list args; 515 char buf[100]; 516 517 va_start(args, fmt); 518 vsnprintf(buf, sizeof(buf), fmt, args); 519 va_end(args); 520 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 521} 522 523static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 524{ 525 unsigned int off; /* Offset of last byte */ 526 u8 *addr = page_address(page); 527 528 print_tracking(s, p); 529 530 print_page_info(page); 531 532 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 533 p, p - addr, get_freepointer(s, p)); 534 535 if (p > addr + 16) 536 print_section("Bytes b4", p - 16, 16); 537 538 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 539 540 if (s->flags & SLAB_RED_ZONE) 541 print_section("Redzone", p + s->objsize, 542 s->inuse - s->objsize); 543 544 if (s->offset) 545 off = s->offset + sizeof(void *); 546 else 547 off = s->inuse; 548 549 if (s->flags & SLAB_STORE_USER) 550 off += 2 * sizeof(struct track); 551 552 if (off != s->size) 553 /* Beginning of the filler is the free pointer */ 554 print_section("Padding", p + off, s->size - off); 555 556 dump_stack(); 557} 558 559static void object_err(struct kmem_cache *s, struct page *page, 560 u8 *object, char *reason) 561{ 562 slab_bug(s, "%s", reason); 563 print_trailer(s, page, object); 564} 565 566static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 567{ 568 va_list args; 569 char buf[100]; 570 571 va_start(args, fmt); 572 vsnprintf(buf, sizeof(buf), fmt, args); 573 va_end(args); 574 slab_bug(s, "%s", buf); 575 print_page_info(page); 576 dump_stack(); 577} 578 579static void init_object(struct kmem_cache *s, void *object, u8 val) 580{ 581 u8 *p = object; 582 583 if (s->flags & __OBJECT_POISON) { 584 memset(p, POISON_FREE, s->objsize - 1); 585 p[s->objsize - 1] = POISON_END; 586 } 587 588 if (s->flags & SLAB_RED_ZONE) 589 memset(p + s->objsize, val, s->inuse - s->objsize); 590} 591 592static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 593{ 594 while (bytes) { 595 if (*start != (u8)value) 596 return start; 597 start++; 598 bytes--; 599 } 600 return NULL; 601} 602 603static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 604 void *from, void *to) 605{ 606 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 607 memset(from, data, to - from); 608} 609 610static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 611 u8 *object, char *what, 612 u8 *start, unsigned int value, unsigned int bytes) 613{ 614 u8 *fault; 615 u8 *end; 616 617 fault = check_bytes(start, value, bytes); 618 if (!fault) 619 return 1; 620 621 end = start + bytes; 622 while (end > fault && end[-1] == value) 623 end--; 624 625 slab_bug(s, "%s overwritten", what); 626 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 627 fault, end - 1, fault[0], value); 628 print_trailer(s, page, object); 629 630 restore_bytes(s, what, value, fault, end); 631 return 0; 632} 633 634/* 635 * Object layout: 636 * 637 * object address 638 * Bytes of the object to be managed. 639 * If the freepointer may overlay the object then the free 640 * pointer is the first word of the object. 641 * 642 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 643 * 0xa5 (POISON_END) 644 * 645 * object + s->objsize 646 * Padding to reach word boundary. This is also used for Redzoning. 647 * Padding is extended by another word if Redzoning is enabled and 648 * objsize == inuse. 649 * 650 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 651 * 0xcc (RED_ACTIVE) for objects in use. 652 * 653 * object + s->inuse 654 * Meta data starts here. 655 * 656 * A. Free pointer (if we cannot overwrite object on free) 657 * B. Tracking data for SLAB_STORE_USER 658 * C. Padding to reach required alignment boundary or at mininum 659 * one word if debugging is on to be able to detect writes 660 * before the word boundary. 661 * 662 * Padding is done using 0x5a (POISON_INUSE) 663 * 664 * object + s->size 665 * Nothing is used beyond s->size. 666 * 667 * If slabcaches are merged then the objsize and inuse boundaries are mostly 668 * ignored. And therefore no slab options that rely on these boundaries 669 * may be used with merged slabcaches. 670 */ 671 672static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 673{ 674 unsigned long off = s->inuse; /* The end of info */ 675 676 if (s->offset) 677 /* Freepointer is placed after the object. */ 678 off += sizeof(void *); 679 680 if (s->flags & SLAB_STORE_USER) 681 /* We also have user information there */ 682 off += 2 * sizeof(struct track); 683 684 if (s->size == off) 685 return 1; 686 687 return check_bytes_and_report(s, page, p, "Object padding", 688 p + off, POISON_INUSE, s->size - off); 689} 690 691/* Check the pad bytes at the end of a slab page */ 692static int slab_pad_check(struct kmem_cache *s, struct page *page) 693{ 694 u8 *start; 695 u8 *fault; 696 u8 *end; 697 int length; 698 int remainder; 699 700 if (!(s->flags & SLAB_POISON)) 701 return 1; 702 703 start = page_address(page); 704 length = (PAGE_SIZE << compound_order(page)) - s->reserved; 705 end = start + length; 706 remainder = length % s->size; 707 if (!remainder) 708 return 1; 709 710 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 711 if (!fault) 712 return 1; 713 while (end > fault && end[-1] == POISON_INUSE) 714 end--; 715 716 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 717 print_section("Padding", end - remainder, remainder); 718 719 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 720 return 0; 721} 722 723static int check_object(struct kmem_cache *s, struct page *page, 724 void *object, u8 val) 725{ 726 u8 *p = object; 727 u8 *endobject = object + s->objsize; 728 729 if (s->flags & SLAB_RED_ZONE) { 730 if (!check_bytes_and_report(s, page, object, "Redzone", 731 endobject, val, s->inuse - s->objsize)) 732 return 0; 733 } else { 734 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 735 check_bytes_and_report(s, page, p, "Alignment padding", 736 endobject, POISON_INUSE, s->inuse - s->objsize); 737 } 738 } 739 740 if (s->flags & SLAB_POISON) { 741 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 742 (!check_bytes_and_report(s, page, p, "Poison", p, 743 POISON_FREE, s->objsize - 1) || 744 !check_bytes_and_report(s, page, p, "Poison", 745 p + s->objsize - 1, POISON_END, 1))) 746 return 0; 747 /* 748 * check_pad_bytes cleans up on its own. 749 */ 750 check_pad_bytes(s, page, p); 751 } 752 753 if (!s->offset && val == SLUB_RED_ACTIVE) 754 /* 755 * Object and freepointer overlap. Cannot check 756 * freepointer while object is allocated. 757 */ 758 return 1; 759 760 /* Check free pointer validity */ 761 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 762 object_err(s, page, p, "Freepointer corrupt"); 763 /* 764 * No choice but to zap it and thus lose the remainder 765 * of the free objects in this slab. May cause 766 * another error because the object count is now wrong. 767 */ 768 set_freepointer(s, p, NULL); 769 return 0; 770 } 771 return 1; 772} 773 774static int check_slab(struct kmem_cache *s, struct page *page) 775{ 776 int maxobj; 777 778 VM_BUG_ON(!irqs_disabled()); 779 780 if (!PageSlab(page)) { 781 slab_err(s, page, "Not a valid slab page"); 782 return 0; 783 } 784 785 maxobj = order_objects(compound_order(page), s->size, s->reserved); 786 if (page->objects > maxobj) { 787 slab_err(s, page, "objects %u > max %u", 788 s->name, page->objects, maxobj); 789 return 0; 790 } 791 if (page->inuse > page->objects) { 792 slab_err(s, page, "inuse %u > max %u", 793 s->name, page->inuse, page->objects); 794 return 0; 795 } 796 /* Slab_pad_check fixes things up after itself */ 797 slab_pad_check(s, page); 798 return 1; 799} 800 801/* 802 * Determine if a certain object on a page is on the freelist. Must hold the 803 * slab lock to guarantee that the chains are in a consistent state. 804 */ 805static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 806{ 807 int nr = 0; 808 void *fp = page->freelist; 809 void *object = NULL; 810 unsigned long max_objects; 811 812 while (fp && nr <= page->objects) { 813 if (fp == search) 814 return 1; 815 if (!check_valid_pointer(s, page, fp)) { 816 if (object) { 817 object_err(s, page, object, 818 "Freechain corrupt"); 819 set_freepointer(s, object, NULL); 820 break; 821 } else { 822 slab_err(s, page, "Freepointer corrupt"); 823 page->freelist = NULL; 824 page->inuse = page->objects; 825 slab_fix(s, "Freelist cleared"); 826 return 0; 827 } 828 break; 829 } 830 object = fp; 831 fp = get_freepointer(s, object); 832 nr++; 833 } 834 835 max_objects = order_objects(compound_order(page), s->size, s->reserved); 836 if (max_objects > MAX_OBJS_PER_PAGE) 837 max_objects = MAX_OBJS_PER_PAGE; 838 839 if (page->objects != max_objects) { 840 slab_err(s, page, "Wrong number of objects. Found %d but " 841 "should be %d", page->objects, max_objects); 842 page->objects = max_objects; 843 slab_fix(s, "Number of objects adjusted."); 844 } 845 if (page->inuse != page->objects - nr) { 846 slab_err(s, page, "Wrong object count. Counter is %d but " 847 "counted were %d", page->inuse, page->objects - nr); 848 page->inuse = page->objects - nr; 849 slab_fix(s, "Object count adjusted."); 850 } 851 return search == NULL; 852} 853 854static void trace(struct kmem_cache *s, struct page *page, void *object, 855 int alloc) 856{ 857 if (s->flags & SLAB_TRACE) { 858 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 859 s->name, 860 alloc ? "alloc" : "free", 861 object, page->inuse, 862 page->freelist); 863 864 if (!alloc) 865 print_section("Object", (void *)object, s->objsize); 866 867 dump_stack(); 868 } 869} 870 871/* 872 * Hooks for other subsystems that check memory allocations. In a typical 873 * production configuration these hooks all should produce no code at all. 874 */ 875static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 876{ 877 flags &= gfp_allowed_mask; 878 lockdep_trace_alloc(flags); 879 might_sleep_if(flags & __GFP_WAIT); 880 881 return should_failslab(s->objsize, flags, s->flags); 882} 883 884static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object) 885{ 886 flags &= gfp_allowed_mask; 887 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s)); 888 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags); 889} 890 891static inline void slab_free_hook(struct kmem_cache *s, void *x) 892{ 893 kmemleak_free_recursive(x, s->flags); 894 895 /* 896 * Trouble is that we may no longer disable interupts in the fast path 897 * So in order to make the debug calls that expect irqs to be 898 * disabled we need to disable interrupts temporarily. 899 */ 900#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) 901 { 902 unsigned long flags; 903 904 local_irq_save(flags); 905 kmemcheck_slab_free(s, x, s->objsize); 906 debug_check_no_locks_freed(x, s->objsize); 907 local_irq_restore(flags); 908 } 909#endif 910 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 911 debug_check_no_obj_freed(x, s->objsize); 912} 913 914/* 915 * Tracking of fully allocated slabs for debugging purposes. 916 */ 917static void add_full(struct kmem_cache_node *n, struct page *page) 918{ 919 spin_lock(&n->list_lock); 920 list_add(&page->lru, &n->full); 921 spin_unlock(&n->list_lock); 922} 923 924static void remove_full(struct kmem_cache *s, struct page *page) 925{ 926 struct kmem_cache_node *n; 927 928 if (!(s->flags & SLAB_STORE_USER)) 929 return; 930 931 n = get_node(s, page_to_nid(page)); 932 933 spin_lock(&n->list_lock); 934 list_del(&page->lru); 935 spin_unlock(&n->list_lock); 936} 937 938/* Tracking of the number of slabs for debugging purposes */ 939static inline unsigned long slabs_node(struct kmem_cache *s, int node) 940{ 941 struct kmem_cache_node *n = get_node(s, node); 942 943 return atomic_long_read(&n->nr_slabs); 944} 945 946static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 947{ 948 return atomic_long_read(&n->nr_slabs); 949} 950 951static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 952{ 953 struct kmem_cache_node *n = get_node(s, node); 954 955 /* 956 * May be called early in order to allocate a slab for the 957 * kmem_cache_node structure. Solve the chicken-egg 958 * dilemma by deferring the increment of the count during 959 * bootstrap (see early_kmem_cache_node_alloc). 960 */ 961 if (n) { 962 atomic_long_inc(&n->nr_slabs); 963 atomic_long_add(objects, &n->total_objects); 964 } 965} 966static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 967{ 968 struct kmem_cache_node *n = get_node(s, node); 969 970 atomic_long_dec(&n->nr_slabs); 971 atomic_long_sub(objects, &n->total_objects); 972} 973 974/* Object debug checks for alloc/free paths */ 975static void setup_object_debug(struct kmem_cache *s, struct page *page, 976 void *object) 977{ 978 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 979 return; 980 981 init_object(s, object, SLUB_RED_INACTIVE); 982 init_tracking(s, object); 983} 984 985static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page, 986 void *object, unsigned long addr) 987{ 988 if (!check_slab(s, page)) 989 goto bad; 990 991 if (!on_freelist(s, page, object)) { 992 object_err(s, page, object, "Object already allocated"); 993 goto bad; 994 } 995 996 if (!check_valid_pointer(s, page, object)) { 997 object_err(s, page, object, "Freelist Pointer check fails"); 998 goto bad; 999 } 1000 1001 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1002 goto bad; 1003 1004 /* Success perform special debug activities for allocs */ 1005 if (s->flags & SLAB_STORE_USER) 1006 set_track(s, object, TRACK_ALLOC, addr); 1007 trace(s, page, object, 1); 1008 init_object(s, object, SLUB_RED_ACTIVE); 1009 return 1; 1010 1011bad: 1012 if (PageSlab(page)) { 1013 /* 1014 * If this is a slab page then lets do the best we can 1015 * to avoid issues in the future. Marking all objects 1016 * as used avoids touching the remaining objects. 1017 */ 1018 slab_fix(s, "Marking all objects used"); 1019 page->inuse = page->objects; 1020 page->freelist = NULL; 1021 } 1022 return 0; 1023} 1024 1025static noinline int free_debug_processing(struct kmem_cache *s, 1026 struct page *page, void *object, unsigned long addr) 1027{ 1028 if (!check_slab(s, page)) 1029 goto fail; 1030 1031 if (!check_valid_pointer(s, page, object)) { 1032 slab_err(s, page, "Invalid object pointer 0x%p", object); 1033 goto fail; 1034 } 1035 1036 if (on_freelist(s, page, object)) { 1037 object_err(s, page, object, "Object already free"); 1038 goto fail; 1039 } 1040 1041 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1042 return 0; 1043 1044 if (unlikely(s != page->slab)) { 1045 if (!PageSlab(page)) { 1046 slab_err(s, page, "Attempt to free object(0x%p) " 1047 "outside of slab", object); 1048 } else if (!page->slab) { 1049 printk(KERN_ERR 1050 "SLUB <none>: no slab for object 0x%p.\n", 1051 object); 1052 dump_stack(); 1053 } else 1054 object_err(s, page, object, 1055 "page slab pointer corrupt."); 1056 goto fail; 1057 } 1058 1059 /* Special debug activities for freeing objects */ 1060 if (!PageSlubFrozen(page) && !page->freelist) 1061 remove_full(s, page); 1062 if (s->flags & SLAB_STORE_USER) 1063 set_track(s, object, TRACK_FREE, addr); 1064 trace(s, page, object, 0); 1065 init_object(s, object, SLUB_RED_INACTIVE); 1066 return 1; 1067 1068fail: 1069 slab_fix(s, "Object at 0x%p not freed", object); 1070 return 0; 1071} 1072 1073static int __init setup_slub_debug(char *str) 1074{ 1075 slub_debug = DEBUG_DEFAULT_FLAGS; 1076 if (*str++ != '=' || !*str) 1077 /* 1078 * No options specified. Switch on full debugging. 1079 */ 1080 goto out; 1081 1082 if (*str == ',') 1083 /* 1084 * No options but restriction on slabs. This means full 1085 * debugging for slabs matching a pattern. 1086 */ 1087 goto check_slabs; 1088 1089 if (tolower(*str) == 'o') { 1090 /* 1091 * Avoid enabling debugging on caches if its minimum order 1092 * would increase as a result. 1093 */ 1094 disable_higher_order_debug = 1; 1095 goto out; 1096 } 1097 1098 slub_debug = 0; 1099 if (*str == '-') 1100 /* 1101 * Switch off all debugging measures. 1102 */ 1103 goto out; 1104 1105 /* 1106 * Determine which debug features should be switched on 1107 */ 1108 for (; *str && *str != ','; str++) { 1109 switch (tolower(*str)) { 1110 case 'f': 1111 slub_debug |= SLAB_DEBUG_FREE; 1112 break; 1113 case 'z': 1114 slub_debug |= SLAB_RED_ZONE; 1115 break; 1116 case 'p': 1117 slub_debug |= SLAB_POISON; 1118 break; 1119 case 'u': 1120 slub_debug |= SLAB_STORE_USER; 1121 break; 1122 case 't': 1123 slub_debug |= SLAB_TRACE; 1124 break; 1125 case 'a': 1126 slub_debug |= SLAB_FAILSLAB; 1127 break; 1128 default: 1129 printk(KERN_ERR "slub_debug option '%c' " 1130 "unknown. skipped\n", *str); 1131 } 1132 } 1133 1134check_slabs: 1135 if (*str == ',') 1136 slub_debug_slabs = str + 1; 1137out: 1138 return 1; 1139} 1140 1141__setup("slub_debug", setup_slub_debug); 1142 1143static unsigned long kmem_cache_flags(unsigned long objsize, 1144 unsigned long flags, const char *name, 1145 void (*ctor)(void *)) 1146{ 1147 /* 1148 * Enable debugging if selected on the kernel commandline. 1149 */ 1150 if (slub_debug && (!slub_debug_slabs || 1151 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) 1152 flags |= slub_debug; 1153 1154 return flags; 1155} 1156#else 1157static inline void setup_object_debug(struct kmem_cache *s, 1158 struct page *page, void *object) {} 1159 1160static inline int alloc_debug_processing(struct kmem_cache *s, 1161 struct page *page, void *object, unsigned long addr) { return 0; } 1162 1163static inline int free_debug_processing(struct kmem_cache *s, 1164 struct page *page, void *object, unsigned long addr) { return 0; } 1165 1166static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1167 { return 1; } 1168static inline int check_object(struct kmem_cache *s, struct page *page, 1169 void *object, u8 val) { return 1; } 1170static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1171static inline unsigned long kmem_cache_flags(unsigned long objsize, 1172 unsigned long flags, const char *name, 1173 void (*ctor)(void *)) 1174{ 1175 return flags; 1176} 1177#define slub_debug 0 1178 1179#define disable_higher_order_debug 0 1180 1181static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1182 { return 0; } 1183static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1184 { return 0; } 1185static inline void inc_slabs_node(struct kmem_cache *s, int node, 1186 int objects) {} 1187static inline void dec_slabs_node(struct kmem_cache *s, int node, 1188 int objects) {} 1189 1190static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 1191 { return 0; } 1192 1193static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, 1194 void *object) {} 1195 1196static inline void slab_free_hook(struct kmem_cache *s, void *x) {} 1197 1198#endif /* CONFIG_SLUB_DEBUG */ 1199 1200/* 1201 * Slab allocation and freeing 1202 */ 1203static inline struct page *alloc_slab_page(gfp_t flags, int node, 1204 struct kmem_cache_order_objects oo) 1205{ 1206 int order = oo_order(oo); 1207 1208 flags |= __GFP_NOTRACK; 1209 1210 if (node == NUMA_NO_NODE) 1211 return alloc_pages(flags, order); 1212 else 1213 return alloc_pages_exact_node(node, flags, order); 1214} 1215 1216static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1217{ 1218 struct page *page; 1219 struct kmem_cache_order_objects oo = s->oo; 1220 gfp_t alloc_gfp; 1221 1222 flags |= s->allocflags; 1223 1224 /* 1225 * Let the initial higher-order allocation fail under memory pressure 1226 * so we fall-back to the minimum order allocation. 1227 */ 1228 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1229 1230 page = alloc_slab_page(alloc_gfp, node, oo); 1231 if (unlikely(!page)) { 1232 oo = s->min; 1233 /* 1234 * Allocation may have failed due to fragmentation. 1235 * Try a lower order alloc if possible 1236 */ 1237 page = alloc_slab_page(flags, node, oo); 1238 if (!page) 1239 return NULL; 1240 1241 stat(s, ORDER_FALLBACK); 1242 } 1243 1244 if (kmemcheck_enabled 1245 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1246 int pages = 1 << oo_order(oo); 1247 1248 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); 1249 1250 /* 1251 * Objects from caches that have a constructor don't get 1252 * cleared when they're allocated, so we need to do it here. 1253 */ 1254 if (s->ctor) 1255 kmemcheck_mark_uninitialized_pages(page, pages); 1256 else 1257 kmemcheck_mark_unallocated_pages(page, pages); 1258 } 1259 1260 page->objects = oo_objects(oo); 1261 mod_zone_page_state(page_zone(page), 1262 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1263 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1264 1 << oo_order(oo)); 1265 1266 return page; 1267} 1268 1269static void setup_object(struct kmem_cache *s, struct page *page, 1270 void *object) 1271{ 1272 setup_object_debug(s, page, object); 1273 if (unlikely(s->ctor)) 1274 s->ctor(object); 1275} 1276 1277static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1278{ 1279 struct page *page; 1280 void *start; 1281 void *last; 1282 void *p; 1283 1284 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1285 1286 page = allocate_slab(s, 1287 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1288 if (!page) 1289 goto out; 1290 1291 inc_slabs_node(s, page_to_nid(page), page->objects); 1292 page->slab = s; 1293 page->flags |= 1 << PG_slab; 1294 1295 start = page_address(page); 1296 1297 if (unlikely(s->flags & SLAB_POISON)) 1298 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1299 1300 last = start; 1301 for_each_object(p, s, start, page->objects) { 1302 setup_object(s, page, last); 1303 set_freepointer(s, last, p); 1304 last = p; 1305 } 1306 setup_object(s, page, last); 1307 set_freepointer(s, last, NULL); 1308 1309 page->freelist = start; 1310 page->inuse = 0; 1311out: 1312 return page; 1313} 1314 1315static void __free_slab(struct kmem_cache *s, struct page *page) 1316{ 1317 int order = compound_order(page); 1318 int pages = 1 << order; 1319 1320 if (kmem_cache_debug(s)) { 1321 void *p; 1322 1323 slab_pad_check(s, page); 1324 for_each_object(p, s, page_address(page), 1325 page->objects) 1326 check_object(s, page, p, SLUB_RED_INACTIVE); 1327 } 1328 1329 kmemcheck_free_shadow(page, compound_order(page)); 1330 1331 mod_zone_page_state(page_zone(page), 1332 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1333 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1334 -pages); 1335 1336 __ClearPageSlab(page); 1337 reset_page_mapcount(page); 1338 if (current->reclaim_state) 1339 current->reclaim_state->reclaimed_slab += pages; 1340 __free_pages(page, order); 1341} 1342 1343#define need_reserve_slab_rcu \ 1344 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) 1345 1346static void rcu_free_slab(struct rcu_head *h) 1347{ 1348 struct page *page; 1349 1350 if (need_reserve_slab_rcu) 1351 page = virt_to_head_page(h); 1352 else 1353 page = container_of((struct list_head *)h, struct page, lru); 1354 1355 __free_slab(page->slab, page); 1356} 1357 1358static void free_slab(struct kmem_cache *s, struct page *page) 1359{ 1360 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1361 struct rcu_head *head; 1362 1363 if (need_reserve_slab_rcu) { 1364 int order = compound_order(page); 1365 int offset = (PAGE_SIZE << order) - s->reserved; 1366 1367 VM_BUG_ON(s->reserved != sizeof(*head)); 1368 head = page_address(page) + offset; 1369 } else { 1370 /* 1371 * RCU free overloads the RCU head over the LRU 1372 */ 1373 head = (void *)&page->lru; 1374 } 1375 1376 call_rcu(head, rcu_free_slab); 1377 } else 1378 __free_slab(s, page); 1379} 1380 1381static void discard_slab(struct kmem_cache *s, struct page *page) 1382{ 1383 dec_slabs_node(s, page_to_nid(page), page->objects); 1384 free_slab(s, page); 1385} 1386 1387/* 1388 * Per slab locking using the pagelock 1389 */ 1390static __always_inline void slab_lock(struct page *page) 1391{ 1392 bit_spin_lock(PG_locked, &page->flags); 1393} 1394 1395static __always_inline void slab_unlock(struct page *page) 1396{ 1397 __bit_spin_unlock(PG_locked, &page->flags); 1398} 1399 1400static __always_inline int slab_trylock(struct page *page) 1401{ 1402 int rc = 1; 1403 1404 rc = bit_spin_trylock(PG_locked, &page->flags); 1405 return rc; 1406} 1407 1408/* 1409 * Management of partially allocated slabs 1410 */ 1411static void add_partial(struct kmem_cache_node *n, 1412 struct page *page, int tail) 1413{ 1414 spin_lock(&n->list_lock); 1415 n->nr_partial++; 1416 if (tail) 1417 list_add_tail(&page->lru, &n->partial); 1418 else 1419 list_add(&page->lru, &n->partial); 1420 spin_unlock(&n->list_lock); 1421} 1422 1423static inline void __remove_partial(struct kmem_cache_node *n, 1424 struct page *page) 1425{ 1426 list_del(&page->lru); 1427 n->nr_partial--; 1428} 1429 1430static void remove_partial(struct kmem_cache *s, struct page *page) 1431{ 1432 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1433 1434 spin_lock(&n->list_lock); 1435 __remove_partial(n, page); 1436 spin_unlock(&n->list_lock); 1437} 1438 1439/* 1440 * Lock slab and remove from the partial list. 1441 * 1442 * Must hold list_lock. 1443 */ 1444static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1445 struct page *page) 1446{ 1447 if (slab_trylock(page)) { 1448 __remove_partial(n, page); 1449 __SetPageSlubFrozen(page); 1450 return 1; 1451 } 1452 return 0; 1453} 1454 1455/* 1456 * Try to allocate a partial slab from a specific node. 1457 */ 1458static struct page *get_partial_node(struct kmem_cache_node *n) 1459{ 1460 struct page *page; 1461 1462 /* 1463 * Racy check. If we mistakenly see no partial slabs then we 1464 * just allocate an empty slab. If we mistakenly try to get a 1465 * partial slab and there is none available then get_partials() 1466 * will return NULL. 1467 */ 1468 if (!n || !n->nr_partial) 1469 return NULL; 1470 1471 spin_lock(&n->list_lock); 1472 list_for_each_entry(page, &n->partial, lru) 1473 if (lock_and_freeze_slab(n, page)) 1474 goto out; 1475 page = NULL; 1476out: 1477 spin_unlock(&n->list_lock); 1478 return page; 1479} 1480 1481/* 1482 * Get a page from somewhere. Search in increasing NUMA distances. 1483 */ 1484static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1485{ 1486#ifdef CONFIG_NUMA 1487 struct zonelist *zonelist; 1488 struct zoneref *z; 1489 struct zone *zone; 1490 enum zone_type high_zoneidx = gfp_zone(flags); 1491 struct page *page; 1492 1493 /* 1494 * The defrag ratio allows a configuration of the tradeoffs between 1495 * inter node defragmentation and node local allocations. A lower 1496 * defrag_ratio increases the tendency to do local allocations 1497 * instead of attempting to obtain partial slabs from other nodes. 1498 * 1499 * If the defrag_ratio is set to 0 then kmalloc() always 1500 * returns node local objects. If the ratio is higher then kmalloc() 1501 * may return off node objects because partial slabs are obtained 1502 * from other nodes and filled up. 1503 * 1504 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1505 * defrag_ratio = 1000) then every (well almost) allocation will 1506 * first attempt to defrag slab caches on other nodes. This means 1507 * scanning over all nodes to look for partial slabs which may be 1508 * expensive if we do it every time we are trying to find a slab 1509 * with available objects. 1510 */ 1511 if (!s->remote_node_defrag_ratio || 1512 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1513 return NULL; 1514 1515 get_mems_allowed(); 1516 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1517 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1518 struct kmem_cache_node *n; 1519 1520 n = get_node(s, zone_to_nid(zone)); 1521 1522 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1523 n->nr_partial > s->min_partial) { 1524 page = get_partial_node(n); 1525 if (page) { 1526 put_mems_allowed(); 1527 return page; 1528 } 1529 } 1530 } 1531 put_mems_allowed(); 1532#endif 1533 return NULL; 1534} 1535 1536/* 1537 * Get a partial page, lock it and return it. 1538 */ 1539static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1540{ 1541 struct page *page; 1542 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; 1543 1544 page = get_partial_node(get_node(s, searchnode)); 1545 if (page || node != NUMA_NO_NODE) 1546 return page; 1547 1548 return get_any_partial(s, flags); 1549} 1550 1551/* 1552 * Move a page back to the lists. 1553 * 1554 * Must be called with the slab lock held. 1555 * 1556 * On exit the slab lock will have been dropped. 1557 */ 1558static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1559 __releases(bitlock) 1560{ 1561 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1562 1563 __ClearPageSlubFrozen(page); 1564 if (page->inuse) { 1565 1566 if (page->freelist) { 1567 add_partial(n, page, tail); 1568 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1569 } else { 1570 stat(s, DEACTIVATE_FULL); 1571 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER)) 1572 add_full(n, page); 1573 } 1574 slab_unlock(page); 1575 } else { 1576 stat(s, DEACTIVATE_EMPTY); 1577 if (n->nr_partial < s->min_partial) { 1578 /* 1579 * Adding an empty slab to the partial slabs in order 1580 * to avoid page allocator overhead. This slab needs 1581 * to come after the other slabs with objects in 1582 * so that the others get filled first. That way the 1583 * size of the partial list stays small. 1584 * 1585 * kmem_cache_shrink can reclaim any empty slabs from 1586 * the partial list. 1587 */ 1588 add_partial(n, page, 1); 1589 slab_unlock(page); 1590 } else { 1591 slab_unlock(page); 1592 stat(s, FREE_SLAB); 1593 discard_slab(s, page); 1594 } 1595 } 1596} 1597 1598#ifdef CONFIG_PREEMPT 1599/* 1600 * Calculate the next globally unique transaction for disambiguiation 1601 * during cmpxchg. The transactions start with the cpu number and are then 1602 * incremented by CONFIG_NR_CPUS. 1603 */ 1604#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 1605#else 1606/* 1607 * No preemption supported therefore also no need to check for 1608 * different cpus. 1609 */ 1610#define TID_STEP 1 1611#endif 1612 1613static inline unsigned long next_tid(unsigned long tid) 1614{ 1615 return tid + TID_STEP; 1616} 1617 1618static inline unsigned int tid_to_cpu(unsigned long tid) 1619{ 1620 return tid % TID_STEP; 1621} 1622 1623static inline unsigned long tid_to_event(unsigned long tid) 1624{ 1625 return tid / TID_STEP; 1626} 1627 1628static inline unsigned int init_tid(int cpu) 1629{ 1630 return cpu; 1631} 1632 1633static inline void note_cmpxchg_failure(const char *n, 1634 const struct kmem_cache *s, unsigned long tid) 1635{ 1636#ifdef SLUB_DEBUG_CMPXCHG 1637 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 1638 1639 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); 1640 1641#ifdef CONFIG_PREEMPT 1642 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 1643 printk("due to cpu change %d -> %d\n", 1644 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 1645 else 1646#endif 1647 if (tid_to_event(tid) != tid_to_event(actual_tid)) 1648 printk("due to cpu running other code. Event %ld->%ld\n", 1649 tid_to_event(tid), tid_to_event(actual_tid)); 1650 else 1651 printk("for unknown reason: actual=%lx was=%lx target=%lx\n", 1652 actual_tid, tid, next_tid(tid)); 1653#endif 1654 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 1655} 1656 1657void init_kmem_cache_cpus(struct kmem_cache *s) 1658{ 1659 int cpu; 1660 1661 for_each_possible_cpu(cpu) 1662 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 1663} 1664/* 1665 * Remove the cpu slab 1666 */ 1667static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1668 __releases(bitlock) 1669{ 1670 struct page *page = c->page; 1671 int tail = 1; 1672 1673 if (page->freelist) 1674 stat(s, DEACTIVATE_REMOTE_FREES); 1675 /* 1676 * Merge cpu freelist into slab freelist. Typically we get here 1677 * because both freelists are empty. So this is unlikely 1678 * to occur. 1679 */ 1680 while (unlikely(c->freelist)) { 1681 void **object; 1682 1683 tail = 0; /* Hot objects. Put the slab first */ 1684 1685 /* Retrieve object from cpu_freelist */ 1686 object = c->freelist; 1687 c->freelist = get_freepointer(s, c->freelist); 1688 1689 /* And put onto the regular freelist */ 1690 set_freepointer(s, object, page->freelist); 1691 page->freelist = object; 1692 page->inuse--; 1693 } 1694 c->page = NULL; 1695 c->tid = next_tid(c->tid); 1696 unfreeze_slab(s, page, tail); 1697} 1698 1699static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1700{ 1701 stat(s, CPUSLAB_FLUSH); 1702 slab_lock(c->page); 1703 deactivate_slab(s, c); 1704} 1705 1706/* 1707 * Flush cpu slab. 1708 * 1709 * Called from IPI handler with interrupts disabled. 1710 */ 1711static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1712{ 1713 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 1714 1715 if (likely(c && c->page)) 1716 flush_slab(s, c); 1717} 1718 1719static void flush_cpu_slab(void *d) 1720{ 1721 struct kmem_cache *s = d; 1722 1723 __flush_cpu_slab(s, smp_processor_id()); 1724} 1725 1726static void flush_all(struct kmem_cache *s) 1727{ 1728 on_each_cpu(flush_cpu_slab, s, 1); 1729} 1730 1731/* 1732 * Check if the objects in a per cpu structure fit numa 1733 * locality expectations. 1734 */ 1735static inline int node_match(struct kmem_cache_cpu *c, int node) 1736{ 1737#ifdef CONFIG_NUMA 1738 if (node != NUMA_NO_NODE && c->node != node) 1739 return 0; 1740#endif 1741 return 1; 1742} 1743 1744static int count_free(struct page *page) 1745{ 1746 return page->objects - page->inuse; 1747} 1748 1749static unsigned long count_partial(struct kmem_cache_node *n, 1750 int (*get_count)(struct page *)) 1751{ 1752 unsigned long flags; 1753 unsigned long x = 0; 1754 struct page *page; 1755 1756 spin_lock_irqsave(&n->list_lock, flags); 1757 list_for_each_entry(page, &n->partial, lru) 1758 x += get_count(page); 1759 spin_unlock_irqrestore(&n->list_lock, flags); 1760 return x; 1761} 1762 1763static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 1764{ 1765#ifdef CONFIG_SLUB_DEBUG 1766 return atomic_long_read(&n->total_objects); 1767#else 1768 return 0; 1769#endif 1770} 1771 1772static noinline void 1773slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 1774{ 1775 int node; 1776 1777 printk(KERN_WARNING 1778 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 1779 nid, gfpflags); 1780 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 1781 "default order: %d, min order: %d\n", s->name, s->objsize, 1782 s->size, oo_order(s->oo), oo_order(s->min)); 1783 1784 if (oo_order(s->min) > get_order(s->objsize)) 1785 printk(KERN_WARNING " %s debugging increased min order, use " 1786 "slub_debug=O to disable.\n", s->name); 1787 1788 for_each_online_node(node) { 1789 struct kmem_cache_node *n = get_node(s, node); 1790 unsigned long nr_slabs; 1791 unsigned long nr_objs; 1792 unsigned long nr_free; 1793 1794 if (!n) 1795 continue; 1796 1797 nr_free = count_partial(n, count_free); 1798 nr_slabs = node_nr_slabs(n); 1799 nr_objs = node_nr_objs(n); 1800 1801 printk(KERN_WARNING 1802 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 1803 node, nr_slabs, nr_objs, nr_free); 1804 } 1805} 1806 1807/* 1808 * Slow path. The lockless freelist is empty or we need to perform 1809 * debugging duties. 1810 * 1811 * Interrupts are disabled. 1812 * 1813 * Processing is still very fast if new objects have been freed to the 1814 * regular freelist. In that case we simply take over the regular freelist 1815 * as the lockless freelist and zap the regular freelist. 1816 * 1817 * If that is not working then we fall back to the partial lists. We take the 1818 * first element of the freelist as the object to allocate now and move the 1819 * rest of the freelist to the lockless freelist. 1820 * 1821 * And if we were unable to get a new slab from the partial slab lists then 1822 * we need to allocate a new slab. This is the slowest path since it involves 1823 * a call to the page allocator and the setup of a new slab. 1824 */ 1825static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 1826 unsigned long addr, struct kmem_cache_cpu *c) 1827{ 1828 void **object; 1829 struct page *page; 1830 unsigned long flags; 1831 1832 local_irq_save(flags); 1833#ifdef CONFIG_PREEMPT 1834 /* 1835 * We may have been preempted and rescheduled on a different 1836 * cpu before disabling interrupts. Need to reload cpu area 1837 * pointer. 1838 */ 1839 c = this_cpu_ptr(s->cpu_slab); 1840#endif 1841 1842 /* We handle __GFP_ZERO in the caller */ 1843 gfpflags &= ~__GFP_ZERO; 1844 1845 page = c->page; 1846 if (!page) 1847 goto new_slab; 1848 1849 slab_lock(page); 1850 if (unlikely(!node_match(c, node))) 1851 goto another_slab; 1852 1853 stat(s, ALLOC_REFILL); 1854 1855load_freelist: 1856 object = page->freelist; 1857 if (unlikely(!object)) 1858 goto another_slab; 1859 if (kmem_cache_debug(s)) 1860 goto debug; 1861 1862 c->freelist = get_freepointer(s, object); 1863 page->inuse = page->objects; 1864 page->freelist = NULL; 1865 1866 slab_unlock(page); 1867 c->tid = next_tid(c->tid); 1868 local_irq_restore(flags); 1869 stat(s, ALLOC_SLOWPATH); 1870 return object; 1871 1872another_slab: 1873 deactivate_slab(s, c); 1874 1875new_slab: 1876 page = get_partial(s, gfpflags, node); 1877 if (page) { 1878 stat(s, ALLOC_FROM_PARTIAL); 1879 c->node = page_to_nid(page); 1880 c->page = page; 1881 goto load_freelist; 1882 } 1883 1884 gfpflags &= gfp_allowed_mask; 1885 if (gfpflags & __GFP_WAIT) 1886 local_irq_enable(); 1887 1888 page = new_slab(s, gfpflags, node); 1889 1890 if (gfpflags & __GFP_WAIT) 1891 local_irq_disable(); 1892 1893 if (page) { 1894 c = __this_cpu_ptr(s->cpu_slab); 1895 stat(s, ALLOC_SLAB); 1896 if (c->page) 1897 flush_slab(s, c); 1898 1899 slab_lock(page); 1900 __SetPageSlubFrozen(page); 1901 c->node = page_to_nid(page); 1902 c->page = page; 1903 goto load_freelist; 1904 } 1905 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 1906 slab_out_of_memory(s, gfpflags, node); 1907 local_irq_restore(flags); 1908 return NULL; 1909debug: 1910 if (!alloc_debug_processing(s, page, object, addr)) 1911 goto another_slab; 1912 1913 page->inuse++; 1914 page->freelist = get_freepointer(s, object); 1915 deactivate_slab(s, c); 1916 c->page = NULL; 1917 c->node = NUMA_NO_NODE; 1918 local_irq_restore(flags); 1919 return object; 1920} 1921 1922/* 1923 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1924 * have the fastpath folded into their functions. So no function call 1925 * overhead for requests that can be satisfied on the fastpath. 1926 * 1927 * The fastpath works by first checking if the lockless freelist can be used. 1928 * If not then __slab_alloc is called for slow processing. 1929 * 1930 * Otherwise we can simply pick the next object from the lockless free list. 1931 */ 1932static __always_inline void *slab_alloc(struct kmem_cache *s, 1933 gfp_t gfpflags, int node, unsigned long addr) 1934{ 1935 void **object; 1936 struct kmem_cache_cpu *c; 1937 unsigned long tid; 1938 1939 if (slab_pre_alloc_hook(s, gfpflags)) 1940 return NULL; 1941 1942redo: 1943 1944 /* 1945 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 1946 * enabled. We may switch back and forth between cpus while 1947 * reading from one cpu area. That does not matter as long 1948 * as we end up on the original cpu again when doing the cmpxchg. 1949 */ 1950 c = __this_cpu_ptr(s->cpu_slab); 1951 1952 /* 1953 * The transaction ids are globally unique per cpu and per operation on 1954 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 1955 * occurs on the right processor and that there was no operation on the 1956 * linked list in between. 1957 */ 1958 tid = c->tid; 1959 barrier(); 1960 1961 object = c->freelist; 1962 if (unlikely(!object || !node_match(c, node))) 1963 1964 object = __slab_alloc(s, gfpflags, node, addr, c); 1965 1966 else { 1967 /* 1968 * The cmpxchg will only match if there was no additional 1969 * operation and if we are on the right processor. 1970 * 1971 * The cmpxchg does the following atomically (without lock semantics!) 1972 * 1. Relocate first pointer to the current per cpu area. 1973 * 2. Verify that tid and freelist have not been changed 1974 * 3. If they were not changed replace tid and freelist 1975 * 1976 * Since this is without lock semantics the protection is only against 1977 * code executing on this cpu *not* from access by other cpus. 1978 */ 1979 if (unlikely(!irqsafe_cpu_cmpxchg_double( 1980 s->cpu_slab->freelist, s->cpu_slab->tid, 1981 object, tid, 1982 get_freepointer_safe(s, object), next_tid(tid)))) { 1983 1984 note_cmpxchg_failure("slab_alloc", s, tid); 1985 goto redo; 1986 } 1987 stat(s, ALLOC_FASTPATH); 1988 } 1989 1990 if (unlikely(gfpflags & __GFP_ZERO) && object) 1991 memset(object, 0, s->objsize); 1992 1993 slab_post_alloc_hook(s, gfpflags, object); 1994 1995 return object; 1996} 1997 1998void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1999{ 2000 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 2001 2002 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); 2003 2004 return ret; 2005} 2006EXPORT_SYMBOL(kmem_cache_alloc); 2007 2008#ifdef CONFIG_TRACING 2009void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2010{ 2011 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 2012 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2013 return ret; 2014} 2015EXPORT_SYMBOL(kmem_cache_alloc_trace); 2016 2017void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) 2018{ 2019 void *ret = kmalloc_order(size, flags, order); 2020 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); 2021 return ret; 2022} 2023EXPORT_SYMBOL(kmalloc_order_trace); 2024#endif 2025 2026#ifdef CONFIG_NUMA 2027void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2028{ 2029 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 2030 2031 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2032 s->objsize, s->size, gfpflags, node); 2033 2034 return ret; 2035} 2036EXPORT_SYMBOL(kmem_cache_alloc_node); 2037 2038#ifdef CONFIG_TRACING 2039void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2040 gfp_t gfpflags, 2041 int node, size_t size) 2042{ 2043 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 2044 2045 trace_kmalloc_node(_RET_IP_, ret, 2046 size, s->size, gfpflags, node); 2047 return ret; 2048} 2049EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2050#endif 2051#endif 2052 2053/* 2054 * Slow patch handling. This may still be called frequently since objects 2055 * have a longer lifetime than the cpu slabs in most processing loads. 2056 * 2057 * So we still attempt to reduce cache line usage. Just take the slab 2058 * lock and free the item. If there is no additional partial page 2059 * handling required then we can return immediately. 2060 */ 2061static void __slab_free(struct kmem_cache *s, struct page *page, 2062 void *x, unsigned long addr) 2063{ 2064 void *prior; 2065 void **object = (void *)x; 2066 unsigned long flags; 2067 2068 local_irq_save(flags); 2069 slab_lock(page); 2070 stat(s, FREE_SLOWPATH); 2071 2072 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr)) 2073 goto out_unlock; 2074 2075 prior = page->freelist; 2076 set_freepointer(s, object, prior); 2077 page->freelist = object; 2078 page->inuse--; 2079 2080 if (unlikely(PageSlubFrozen(page))) { 2081 stat(s, FREE_FROZEN); 2082 goto out_unlock; 2083 } 2084 2085 if (unlikely(!page->inuse)) 2086 goto slab_empty; 2087 2088 /* 2089 * Objects left in the slab. If it was not on the partial list before 2090 * then add it. 2091 */ 2092 if (unlikely(!prior)) { 2093 add_partial(get_node(s, page_to_nid(page)), page, 1); 2094 stat(s, FREE_ADD_PARTIAL); 2095 } 2096 2097out_unlock: 2098 slab_unlock(page); 2099 local_irq_restore(flags); 2100 return; 2101 2102slab_empty: 2103 if (prior) { 2104 /* 2105 * Slab still on the partial list. 2106 */ 2107 remove_partial(s, page); 2108 stat(s, FREE_REMOVE_PARTIAL); 2109 } 2110 slab_unlock(page); 2111 local_irq_restore(flags); 2112 stat(s, FREE_SLAB); 2113 discard_slab(s, page); 2114} 2115 2116/* 2117 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 2118 * can perform fastpath freeing without additional function calls. 2119 * 2120 * The fastpath is only possible if we are freeing to the current cpu slab 2121 * of this processor. This typically the case if we have just allocated 2122 * the item before. 2123 * 2124 * If fastpath is not possible then fall back to __slab_free where we deal 2125 * with all sorts of special processing. 2126 */ 2127static __always_inline void slab_free(struct kmem_cache *s, 2128 struct page *page, void *x, unsigned long addr) 2129{ 2130 void **object = (void *)x; 2131 struct kmem_cache_cpu *c; 2132 unsigned long tid; 2133 2134 slab_free_hook(s, x); 2135 2136redo: 2137 2138 /* 2139 * Determine the currently cpus per cpu slab. 2140 * The cpu may change afterward. However that does not matter since 2141 * data is retrieved via this pointer. If we are on the same cpu 2142 * during the cmpxchg then the free will succedd. 2143 */ 2144 c = __this_cpu_ptr(s->cpu_slab); 2145 2146 tid = c->tid; 2147 barrier(); 2148 2149 if (likely(page == c->page)) { 2150 set_freepointer(s, object, c->freelist); 2151 2152 if (unlikely(!irqsafe_cpu_cmpxchg_double( 2153 s->cpu_slab->freelist, s->cpu_slab->tid, 2154 c->freelist, tid, 2155 object, next_tid(tid)))) { 2156 2157 note_cmpxchg_failure("slab_free", s, tid); 2158 goto redo; 2159 } 2160 stat(s, FREE_FASTPATH); 2161 } else 2162 __slab_free(s, page, x, addr); 2163 2164} 2165 2166void kmem_cache_free(struct kmem_cache *s, void *x) 2167{ 2168 struct page *page; 2169 2170 page = virt_to_head_page(x); 2171 2172 slab_free(s, page, x, _RET_IP_); 2173 2174 trace_kmem_cache_free(_RET_IP_, x); 2175} 2176EXPORT_SYMBOL(kmem_cache_free); 2177 2178/* 2179 * Object placement in a slab is made very easy because we always start at 2180 * offset 0. If we tune the size of the object to the alignment then we can 2181 * get the required alignment by putting one properly sized object after 2182 * another. 2183 * 2184 * Notice that the allocation order determines the sizes of the per cpu 2185 * caches. Each processor has always one slab available for allocations. 2186 * Increasing the allocation order reduces the number of times that slabs 2187 * must be moved on and off the partial lists and is therefore a factor in 2188 * locking overhead. 2189 */ 2190 2191/* 2192 * Mininum / Maximum order of slab pages. This influences locking overhead 2193 * and slab fragmentation. A higher order reduces the number of partial slabs 2194 * and increases the number of allocations possible without having to 2195 * take the list_lock. 2196 */ 2197static int slub_min_order; 2198static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 2199static int slub_min_objects; 2200 2201/* 2202 * Merge control. If this is set then no merging of slab caches will occur. 2203 * (Could be removed. This was introduced to pacify the merge skeptics.) 2204 */ 2205static int slub_nomerge; 2206 2207/* 2208 * Calculate the order of allocation given an slab object size. 2209 * 2210 * The order of allocation has significant impact on performance and other 2211 * system components. Generally order 0 allocations should be preferred since 2212 * order 0 does not cause fragmentation in the page allocator. Larger objects 2213 * be problematic to put into order 0 slabs because there may be too much 2214 * unused space left. We go to a higher order if more than 1/16th of the slab 2215 * would be wasted. 2216 * 2217 * In order to reach satisfactory performance we must ensure that a minimum 2218 * number of objects is in one slab. Otherwise we may generate too much 2219 * activity on the partial lists which requires taking the list_lock. This is 2220 * less a concern for large slabs though which are rarely used. 2221 * 2222 * slub_max_order specifies the order where we begin to stop considering the 2223 * number of objects in a slab as critical. If we reach slub_max_order then 2224 * we try to keep the page order as low as possible. So we accept more waste 2225 * of space in favor of a small page order. 2226 * 2227 * Higher order allocations also allow the placement of more objects in a 2228 * slab and thereby reduce object handling overhead. If the user has 2229 * requested a higher mininum order then we start with that one instead of 2230 * the smallest order which will fit the object. 2231 */ 2232static inline int slab_order(int size, int min_objects, 2233 int max_order, int fract_leftover, int reserved) 2234{ 2235 int order; 2236 int rem; 2237 int min_order = slub_min_order; 2238 2239 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) 2240 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 2241 2242 for (order = max(min_order, 2243 fls(min_objects * size - 1) - PAGE_SHIFT); 2244 order <= max_order; order++) { 2245 2246 unsigned long slab_size = PAGE_SIZE << order; 2247 2248 if (slab_size < min_objects * size + reserved) 2249 continue; 2250 2251 rem = (slab_size - reserved) % size; 2252 2253 if (rem <= slab_size / fract_leftover) 2254 break; 2255 2256 } 2257 2258 return order; 2259} 2260 2261static inline int calculate_order(int size, int reserved) 2262{ 2263 int order; 2264 int min_objects; 2265 int fraction; 2266 int max_objects; 2267 2268 /* 2269 * Attempt to find best configuration for a slab. This 2270 * works by first attempting to generate a layout with 2271 * the best configuration and backing off gradually. 2272 * 2273 * First we reduce the acceptable waste in a slab. Then 2274 * we reduce the minimum objects required in a slab. 2275 */ 2276 min_objects = slub_min_objects; 2277 if (!min_objects) 2278 min_objects = 4 * (fls(nr_cpu_ids) + 1); 2279 max_objects = order_objects(slub_max_order, size, reserved); 2280 min_objects = min(min_objects, max_objects); 2281 2282 while (min_objects > 1) { 2283 fraction = 16; 2284 while (fraction >= 4) { 2285 order = slab_order(size, min_objects, 2286 slub_max_order, fraction, reserved); 2287 if (order <= slub_max_order) 2288 return order; 2289 fraction /= 2; 2290 } 2291 min_objects--; 2292 } 2293 2294 /* 2295 * We were unable to place multiple objects in a slab. Now 2296 * lets see if we can place a single object there. 2297 */ 2298 order = slab_order(size, 1, slub_max_order, 1, reserved); 2299 if (order <= slub_max_order) 2300 return order; 2301 2302 /* 2303 * Doh this slab cannot be placed using slub_max_order. 2304 */ 2305 order = slab_order(size, 1, MAX_ORDER, 1, reserved); 2306 if (order < MAX_ORDER) 2307 return order; 2308 return -ENOSYS; 2309} 2310 2311/* 2312 * Figure out what the alignment of the objects will be. 2313 */ 2314static unsigned long calculate_alignment(unsigned long flags, 2315 unsigned long align, unsigned long size) 2316{ 2317 /* 2318 * If the user wants hardware cache aligned objects then follow that 2319 * suggestion if the object is sufficiently large. 2320 * 2321 * The hardware cache alignment cannot override the specified 2322 * alignment though. If that is greater then use it. 2323 */ 2324 if (flags & SLAB_HWCACHE_ALIGN) { 2325 unsigned long ralign = cache_line_size(); 2326 while (size <= ralign / 2) 2327 ralign /= 2; 2328 align = max(align, ralign); 2329 } 2330 2331 if (align < ARCH_SLAB_MINALIGN) 2332 align = ARCH_SLAB_MINALIGN; 2333 2334 return ALIGN(align, sizeof(void *)); 2335} 2336 2337static void 2338init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 2339{ 2340 n->nr_partial = 0; 2341 spin_lock_init(&n->list_lock); 2342 INIT_LIST_HEAD(&n->partial); 2343#ifdef CONFIG_SLUB_DEBUG 2344 atomic_long_set(&n->nr_slabs, 0); 2345 atomic_long_set(&n->total_objects, 0); 2346 INIT_LIST_HEAD(&n->full); 2347#endif 2348} 2349 2350static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 2351{ 2352 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 2353 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); 2354 2355#ifdef CONFIG_CMPXCHG_LOCAL 2356 /* 2357 * Must align to double word boundary for the double cmpxchg instructions 2358 * to work. 2359 */ 2360 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *)); 2361#else 2362 /* Regular alignment is sufficient */ 2363 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu); 2364#endif 2365 2366 if (!s->cpu_slab) 2367 return 0; 2368 2369 init_kmem_cache_cpus(s); 2370 2371 return 1; 2372} 2373 2374static struct kmem_cache *kmem_cache_node; 2375 2376/* 2377 * No kmalloc_node yet so do it by hand. We know that this is the first 2378 * slab on the node for this slabcache. There are no concurrent accesses 2379 * possible. 2380 * 2381 * Note that this function only works on the kmalloc_node_cache 2382 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2383 * memory on a fresh node that has no slab structures yet. 2384 */ 2385static void early_kmem_cache_node_alloc(int node) 2386{ 2387 struct page *page; 2388 struct kmem_cache_node *n; 2389 unsigned long flags; 2390 2391 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 2392 2393 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 2394 2395 BUG_ON(!page); 2396 if (page_to_nid(page) != node) { 2397 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2398 "node %d\n", node); 2399 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2400 "in order to be able to continue\n"); 2401 } 2402 2403 n = page->freelist; 2404 BUG_ON(!n); 2405 page->freelist = get_freepointer(kmem_cache_node, n); 2406 page->inuse++; 2407 kmem_cache_node->node[node] = n; 2408#ifdef CONFIG_SLUB_DEBUG 2409 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 2410 init_tracking(kmem_cache_node, n); 2411#endif 2412 init_kmem_cache_node(n, kmem_cache_node); 2413 inc_slabs_node(kmem_cache_node, node, page->objects); 2414 2415 /* 2416 * lockdep requires consistent irq usage for each lock 2417 * so even though there cannot be a race this early in 2418 * the boot sequence, we still disable irqs. 2419 */ 2420 local_irq_save(flags); 2421 add_partial(n, page, 0); 2422 local_irq_restore(flags); 2423} 2424 2425static void free_kmem_cache_nodes(struct kmem_cache *s) 2426{ 2427 int node; 2428 2429 for_each_node_state(node, N_NORMAL_MEMORY) { 2430 struct kmem_cache_node *n = s->node[node]; 2431 2432 if (n) 2433 kmem_cache_free(kmem_cache_node, n); 2434 2435 s->node[node] = NULL; 2436 } 2437} 2438 2439static int init_kmem_cache_nodes(struct kmem_cache *s) 2440{ 2441 int node; 2442 2443 for_each_node_state(node, N_NORMAL_MEMORY) { 2444 struct kmem_cache_node *n; 2445 2446 if (slab_state == DOWN) { 2447 early_kmem_cache_node_alloc(node); 2448 continue; 2449 } 2450 n = kmem_cache_alloc_node(kmem_cache_node, 2451 GFP_KERNEL, node); 2452 2453 if (!n) { 2454 free_kmem_cache_nodes(s); 2455 return 0; 2456 } 2457 2458 s->node[node] = n; 2459 init_kmem_cache_node(n, s); 2460 } 2461 return 1; 2462} 2463 2464static void set_min_partial(struct kmem_cache *s, unsigned long min) 2465{ 2466 if (min < MIN_PARTIAL) 2467 min = MIN_PARTIAL; 2468 else if (min > MAX_PARTIAL) 2469 min = MAX_PARTIAL; 2470 s->min_partial = min; 2471} 2472 2473/* 2474 * calculate_sizes() determines the order and the distribution of data within 2475 * a slab object. 2476 */ 2477static int calculate_sizes(struct kmem_cache *s, int forced_order) 2478{ 2479 unsigned long flags = s->flags; 2480 unsigned long size = s->objsize; 2481 unsigned long align = s->align; 2482 int order; 2483 2484 /* 2485 * Round up object size to the next word boundary. We can only 2486 * place the free pointer at word boundaries and this determines 2487 * the possible location of the free pointer. 2488 */ 2489 size = ALIGN(size, sizeof(void *)); 2490 2491#ifdef CONFIG_SLUB_DEBUG 2492 /* 2493 * Determine if we can poison the object itself. If the user of 2494 * the slab may touch the object after free or before allocation 2495 * then we should never poison the object itself. 2496 */ 2497 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2498 !s->ctor) 2499 s->flags |= __OBJECT_POISON; 2500 else 2501 s->flags &= ~__OBJECT_POISON; 2502 2503 2504 /* 2505 * If we are Redzoning then check if there is some space between the 2506 * end of the object and the free pointer. If not then add an 2507 * additional word to have some bytes to store Redzone information. 2508 */ 2509 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2510 size += sizeof(void *); 2511#endif 2512 2513 /* 2514 * With that we have determined the number of bytes in actual use 2515 * by the object. This is the potential offset to the free pointer. 2516 */ 2517 s->inuse = size; 2518 2519 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2520 s->ctor)) { 2521 /* 2522 * Relocate free pointer after the object if it is not 2523 * permitted to overwrite the first word of the object on 2524 * kmem_cache_free. 2525 * 2526 * This is the case if we do RCU, have a constructor or 2527 * destructor or are poisoning the objects. 2528 */ 2529 s->offset = size; 2530 size += sizeof(void *); 2531 } 2532 2533#ifdef CONFIG_SLUB_DEBUG 2534 if (flags & SLAB_STORE_USER) 2535 /* 2536 * Need to store information about allocs and frees after 2537 * the object. 2538 */ 2539 size += 2 * sizeof(struct track); 2540 2541 if (flags & SLAB_RED_ZONE) 2542 /* 2543 * Add some empty padding so that we can catch 2544 * overwrites from earlier objects rather than let 2545 * tracking information or the free pointer be 2546 * corrupted if a user writes before the start 2547 * of the object. 2548 */ 2549 size += sizeof(void *); 2550#endif 2551 2552 /* 2553 * Determine the alignment based on various parameters that the 2554 * user specified and the dynamic determination of cache line size 2555 * on bootup. 2556 */ 2557 align = calculate_alignment(flags, align, s->objsize); 2558 s->align = align; 2559 2560 /* 2561 * SLUB stores one object immediately after another beginning from 2562 * offset 0. In order to align the objects we have to simply size 2563 * each object to conform to the alignment. 2564 */ 2565 size = ALIGN(size, align); 2566 s->size = size; 2567 if (forced_order >= 0) 2568 order = forced_order; 2569 else 2570 order = calculate_order(size, s->reserved); 2571 2572 if (order < 0) 2573 return 0; 2574 2575 s->allocflags = 0; 2576 if (order) 2577 s->allocflags |= __GFP_COMP; 2578 2579 if (s->flags & SLAB_CACHE_DMA) 2580 s->allocflags |= SLUB_DMA; 2581 2582 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2583 s->allocflags |= __GFP_RECLAIMABLE; 2584 2585 /* 2586 * Determine the number of objects per slab 2587 */ 2588 s->oo = oo_make(order, size, s->reserved); 2589 s->min = oo_make(get_order(size), size, s->reserved); 2590 if (oo_objects(s->oo) > oo_objects(s->max)) 2591 s->max = s->oo; 2592 2593 return !!oo_objects(s->oo); 2594 2595} 2596 2597static int kmem_cache_open(struct kmem_cache *s, 2598 const char *name, size_t size, 2599 size_t align, unsigned long flags, 2600 void (*ctor)(void *)) 2601{ 2602 memset(s, 0, kmem_size); 2603 s->name = name; 2604 s->ctor = ctor; 2605 s->objsize = size; 2606 s->align = align; 2607 s->flags = kmem_cache_flags(size, flags, name, ctor); 2608 s->reserved = 0; 2609 2610 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) 2611 s->reserved = sizeof(struct rcu_head); 2612 2613 if (!calculate_sizes(s, -1)) 2614 goto error; 2615 if (disable_higher_order_debug) { 2616 /* 2617 * Disable debugging flags that store metadata if the min slab 2618 * order increased. 2619 */ 2620 if (get_order(s->size) > get_order(s->objsize)) { 2621 s->flags &= ~DEBUG_METADATA_FLAGS; 2622 s->offset = 0; 2623 if (!calculate_sizes(s, -1)) 2624 goto error; 2625 } 2626 } 2627 2628 /* 2629 * The larger the object size is, the more pages we want on the partial 2630 * list to avoid pounding the page allocator excessively. 2631 */ 2632 set_min_partial(s, ilog2(s->size)); 2633 s->refcount = 1; 2634#ifdef CONFIG_NUMA 2635 s->remote_node_defrag_ratio = 1000; 2636#endif 2637 if (!init_kmem_cache_nodes(s)) 2638 goto error; 2639 2640 if (alloc_kmem_cache_cpus(s)) 2641 return 1; 2642 2643 free_kmem_cache_nodes(s); 2644error: 2645 if (flags & SLAB_PANIC) 2646 panic("Cannot create slab %s size=%lu realsize=%u " 2647 "order=%u offset=%u flags=%lx\n", 2648 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2649 s->offset, flags); 2650 return 0; 2651} 2652 2653/* 2654 * Determine the size of a slab object 2655 */ 2656unsigned int kmem_cache_size(struct kmem_cache *s) 2657{ 2658 return s->objsize; 2659} 2660EXPORT_SYMBOL(kmem_cache_size); 2661 2662static void list_slab_objects(struct kmem_cache *s, struct page *page, 2663 const char *text) 2664{ 2665#ifdef CONFIG_SLUB_DEBUG 2666 void *addr = page_address(page); 2667 void *p; 2668 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * 2669 sizeof(long), GFP_ATOMIC); 2670 if (!map) 2671 return; 2672 slab_err(s, page, "%s", text); 2673 slab_lock(page); 2674 2675 get_map(s, page, map); 2676 for_each_object(p, s, addr, page->objects) { 2677 2678 if (!test_bit(slab_index(p, s, addr), map)) { 2679 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2680 p, p - addr); 2681 print_tracking(s, p); 2682 } 2683 } 2684 slab_unlock(page); 2685 kfree(map); 2686#endif 2687} 2688 2689/* 2690 * Attempt to free all partial slabs on a node. 2691 */ 2692static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2693{ 2694 unsigned long flags; 2695 struct page *page, *h; 2696 2697 spin_lock_irqsave(&n->list_lock, flags); 2698 list_for_each_entry_safe(page, h, &n->partial, lru) { 2699 if (!page->inuse) { 2700 __remove_partial(n, page); 2701 discard_slab(s, page); 2702 } else { 2703 list_slab_objects(s, page, 2704 "Objects remaining on kmem_cache_close()"); 2705 } 2706 } 2707 spin_unlock_irqrestore(&n->list_lock, flags); 2708} 2709 2710/* 2711 * Release all resources used by a slab cache. 2712 */ 2713static inline int kmem_cache_close(struct kmem_cache *s) 2714{ 2715 int node; 2716 2717 flush_all(s); 2718 free_percpu(s->cpu_slab); 2719 /* Attempt to free all objects */ 2720 for_each_node_state(node, N_NORMAL_MEMORY) { 2721 struct kmem_cache_node *n = get_node(s, node); 2722 2723 free_partial(s, n); 2724 if (n->nr_partial || slabs_node(s, node)) 2725 return 1; 2726 } 2727 free_kmem_cache_nodes(s); 2728 return 0; 2729} 2730 2731/* 2732 * Close a cache and release the kmem_cache structure 2733 * (must be used for caches created using kmem_cache_create) 2734 */ 2735void kmem_cache_destroy(struct kmem_cache *s) 2736{ 2737 down_write(&slub_lock); 2738 s->refcount--; 2739 if (!s->refcount) { 2740 list_del(&s->list); 2741 if (kmem_cache_close(s)) { 2742 printk(KERN_ERR "SLUB %s: %s called for cache that " 2743 "still has objects.\n", s->name, __func__); 2744 dump_stack(); 2745 } 2746 if (s->flags & SLAB_DESTROY_BY_RCU) 2747 rcu_barrier(); 2748 sysfs_slab_remove(s); 2749 } 2750 up_write(&slub_lock); 2751} 2752EXPORT_SYMBOL(kmem_cache_destroy); 2753 2754/******************************************************************** 2755 * Kmalloc subsystem 2756 *******************************************************************/ 2757 2758struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; 2759EXPORT_SYMBOL(kmalloc_caches); 2760 2761static struct kmem_cache *kmem_cache; 2762 2763#ifdef CONFIG_ZONE_DMA 2764static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; 2765#endif 2766 2767static int __init setup_slub_min_order(char *str) 2768{ 2769 get_option(&str, &slub_min_order); 2770 2771 return 1; 2772} 2773 2774__setup("slub_min_order=", setup_slub_min_order); 2775 2776static int __init setup_slub_max_order(char *str) 2777{ 2778 get_option(&str, &slub_max_order); 2779 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 2780 2781 return 1; 2782} 2783 2784__setup("slub_max_order=", setup_slub_max_order); 2785 2786static int __init setup_slub_min_objects(char *str) 2787{ 2788 get_option(&str, &slub_min_objects); 2789 2790 return 1; 2791} 2792 2793__setup("slub_min_objects=", setup_slub_min_objects); 2794 2795static int __init setup_slub_nomerge(char *str) 2796{ 2797 slub_nomerge = 1; 2798 return 1; 2799} 2800 2801__setup("slub_nomerge", setup_slub_nomerge); 2802 2803static struct kmem_cache *__init create_kmalloc_cache(const char *name, 2804 int size, unsigned int flags) 2805{ 2806 struct kmem_cache *s; 2807 2808 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 2809 2810 /* 2811 * This function is called with IRQs disabled during early-boot on 2812 * single CPU so there's no need to take slub_lock here. 2813 */ 2814 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, 2815 flags, NULL)) 2816 goto panic; 2817 2818 list_add(&s->list, &slab_caches); 2819 return s; 2820 2821panic: 2822 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2823 return NULL; 2824} 2825 2826/* 2827 * Conversion table for small slabs sizes / 8 to the index in the 2828 * kmalloc array. This is necessary for slabs < 192 since we have non power 2829 * of two cache sizes there. The size of larger slabs can be determined using 2830 * fls. 2831 */ 2832static s8 size_index[24] = { 2833 3, /* 8 */ 2834 4, /* 16 */ 2835 5, /* 24 */ 2836 5, /* 32 */ 2837 6, /* 40 */ 2838 6, /* 48 */ 2839 6, /* 56 */ 2840 6, /* 64 */ 2841 1, /* 72 */ 2842 1, /* 80 */ 2843 1, /* 88 */ 2844 1, /* 96 */ 2845 7, /* 104 */ 2846 7, /* 112 */ 2847 7, /* 120 */ 2848 7, /* 128 */ 2849 2, /* 136 */ 2850 2, /* 144 */ 2851 2, /* 152 */ 2852 2, /* 160 */ 2853 2, /* 168 */ 2854 2, /* 176 */ 2855 2, /* 184 */ 2856 2 /* 192 */ 2857}; 2858 2859static inline int size_index_elem(size_t bytes) 2860{ 2861 return (bytes - 1) / 8; 2862} 2863 2864static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2865{ 2866 int index; 2867 2868 if (size <= 192) { 2869 if (!size) 2870 return ZERO_SIZE_PTR; 2871 2872 index = size_index[size_index_elem(size)]; 2873 } else 2874 index = fls(size - 1); 2875 2876#ifdef CONFIG_ZONE_DMA 2877 if (unlikely((flags & SLUB_DMA))) 2878 return kmalloc_dma_caches[index]; 2879 2880#endif 2881 return kmalloc_caches[index]; 2882} 2883 2884void *__kmalloc(size_t size, gfp_t flags) 2885{ 2886 struct kmem_cache *s; 2887 void *ret; 2888 2889 if (unlikely(size > SLUB_MAX_SIZE)) 2890 return kmalloc_large(size, flags); 2891 2892 s = get_slab(size, flags); 2893 2894 if (unlikely(ZERO_OR_NULL_PTR(s))) 2895 return s; 2896 2897 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); 2898 2899 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 2900 2901 return ret; 2902} 2903EXPORT_SYMBOL(__kmalloc); 2904 2905#ifdef CONFIG_NUMA 2906static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2907{ 2908 struct page *page; 2909 void *ptr = NULL; 2910 2911 flags |= __GFP_COMP | __GFP_NOTRACK; 2912 page = alloc_pages_node(node, flags, get_order(size)); 2913 if (page) 2914 ptr = page_address(page); 2915 2916 kmemleak_alloc(ptr, size, 1, flags); 2917 return ptr; 2918} 2919 2920void *__kmalloc_node(size_t size, gfp_t flags, int node) 2921{ 2922 struct kmem_cache *s; 2923 void *ret; 2924 2925 if (unlikely(size > SLUB_MAX_SIZE)) { 2926 ret = kmalloc_large_node(size, flags, node); 2927 2928 trace_kmalloc_node(_RET_IP_, ret, 2929 size, PAGE_SIZE << get_order(size), 2930 flags, node); 2931 2932 return ret; 2933 } 2934 2935 s = get_slab(size, flags); 2936 2937 if (unlikely(ZERO_OR_NULL_PTR(s))) 2938 return s; 2939 2940 ret = slab_alloc(s, flags, node, _RET_IP_); 2941 2942 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 2943 2944 return ret; 2945} 2946EXPORT_SYMBOL(__kmalloc_node); 2947#endif 2948 2949size_t ksize(const void *object) 2950{ 2951 struct page *page; 2952 2953 if (unlikely(object == ZERO_SIZE_PTR)) 2954 return 0; 2955 2956 page = virt_to_head_page(object); 2957 2958 if (unlikely(!PageSlab(page))) { 2959 WARN_ON(!PageCompound(page)); 2960 return PAGE_SIZE << compound_order(page); 2961 } 2962 2963 return slab_ksize(page->slab); 2964} 2965EXPORT_SYMBOL(ksize); 2966 2967#ifdef CONFIG_SLUB_DEBUG 2968bool verify_mem_not_deleted(const void *x) 2969{ 2970 struct page *page; 2971 void *object = (void *)x; 2972 unsigned long flags; 2973 bool rv; 2974 2975 if (unlikely(ZERO_OR_NULL_PTR(x))) 2976 return false; 2977 2978 local_irq_save(flags); 2979 2980 page = virt_to_head_page(x); 2981 if (unlikely(!PageSlab(page))) { 2982 /* maybe it was from stack? */ 2983 rv = true; 2984 goto out_unlock; 2985 } 2986 2987 slab_lock(page); 2988 if (on_freelist(page->slab, page, object)) { 2989 object_err(page->slab, page, object, "Object is on free-list"); 2990 rv = false; 2991 } else { 2992 rv = true; 2993 } 2994 slab_unlock(page); 2995 2996out_unlock: 2997 local_irq_restore(flags); 2998 return rv; 2999} 3000EXPORT_SYMBOL(verify_mem_not_deleted); 3001#endif 3002 3003void kfree(const void *x) 3004{ 3005 struct page *page; 3006 void *object = (void *)x; 3007 3008 trace_kfree(_RET_IP_, x); 3009 3010 if (unlikely(ZERO_OR_NULL_PTR(x))) 3011 return; 3012 3013 page = virt_to_head_page(x); 3014 if (unlikely(!PageSlab(page))) { 3015 BUG_ON(!PageCompound(page)); 3016 kmemleak_free(x); 3017 put_page(page); 3018 return; 3019 } 3020 slab_free(page->slab, page, object, _RET_IP_); 3021} 3022EXPORT_SYMBOL(kfree); 3023 3024/* 3025 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 3026 * the remaining slabs by the number of items in use. The slabs with the 3027 * most items in use come first. New allocations will then fill those up 3028 * and thus they can be removed from the partial lists. 3029 * 3030 * The slabs with the least items are placed last. This results in them 3031 * being allocated from last increasing the chance that the last objects 3032 * are freed in them. 3033 */ 3034int kmem_cache_shrink(struct kmem_cache *s) 3035{ 3036 int node; 3037 int i; 3038 struct kmem_cache_node *n; 3039 struct page *page; 3040 struct page *t; 3041 int objects = oo_objects(s->max); 3042 struct list_head *slabs_by_inuse = 3043 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 3044 unsigned long flags; 3045 3046 if (!slabs_by_inuse) 3047 return -ENOMEM; 3048 3049 flush_all(s); 3050 for_each_node_state(node, N_NORMAL_MEMORY) { 3051 n = get_node(s, node); 3052 3053 if (!n->nr_partial) 3054 continue; 3055 3056 for (i = 0; i < objects; i++) 3057 INIT_LIST_HEAD(slabs_by_inuse + i); 3058 3059 spin_lock_irqsave(&n->list_lock, flags); 3060 3061 /* 3062 * Build lists indexed by the items in use in each slab. 3063 * 3064 * Note that concurrent frees may occur while we hold the 3065 * list_lock. page->inuse here is the upper limit. 3066 */ 3067 list_for_each_entry_safe(page, t, &n->partial, lru) { 3068 if (!page->inuse && slab_trylock(page)) { 3069 /* 3070 * Must hold slab lock here because slab_free 3071 * may have freed the last object and be 3072 * waiting to release the slab. 3073 */ 3074 __remove_partial(n, page); 3075 slab_unlock(page); 3076 discard_slab(s, page); 3077 } else { 3078 list_move(&page->lru, 3079 slabs_by_inuse + page->inuse); 3080 } 3081 } 3082 3083 /* 3084 * Rebuild the partial list with the slabs filled up most 3085 * first and the least used slabs at the end. 3086 */ 3087 for (i = objects - 1; i >= 0; i--) 3088 list_splice(slabs_by_inuse + i, n->partial.prev); 3089 3090 spin_unlock_irqrestore(&n->list_lock, flags); 3091 } 3092 3093 kfree(slabs_by_inuse); 3094 return 0; 3095} 3096EXPORT_SYMBOL(kmem_cache_shrink); 3097 3098#if defined(CONFIG_MEMORY_HOTPLUG) 3099static int slab_mem_going_offline_callback(void *arg) 3100{ 3101 struct kmem_cache *s; 3102 3103 down_read(&slub_lock); 3104 list_for_each_entry(s, &slab_caches, list) 3105 kmem_cache_shrink(s); 3106 up_read(&slub_lock); 3107 3108 return 0; 3109} 3110 3111static void slab_mem_offline_callback(void *arg) 3112{ 3113 struct kmem_cache_node *n; 3114 struct kmem_cache *s; 3115 struct memory_notify *marg = arg; 3116 int offline_node; 3117 3118 offline_node = marg->status_change_nid; 3119 3120 /* 3121 * If the node still has available memory. we need kmem_cache_node 3122 * for it yet. 3123 */ 3124 if (offline_node < 0) 3125 return; 3126 3127 down_read(&slub_lock); 3128 list_for_each_entry(s, &slab_caches, list) { 3129 n = get_node(s, offline_node); 3130 if (n) { 3131 /* 3132 * if n->nr_slabs > 0, slabs still exist on the node 3133 * that is going down. We were unable to free them, 3134 * and offline_pages() function shouldn't call this 3135 * callback. So, we must fail. 3136 */ 3137 BUG_ON(slabs_node(s, offline_node)); 3138 3139 s->node[offline_node] = NULL; 3140 kmem_cache_free(kmem_cache_node, n); 3141 } 3142 } 3143 up_read(&slub_lock); 3144} 3145 3146static int slab_mem_going_online_callback(void *arg) 3147{ 3148 struct kmem_cache_node *n; 3149 struct kmem_cache *s; 3150 struct memory_notify *marg = arg; 3151 int nid = marg->status_change_nid; 3152 int ret = 0; 3153 3154 /* 3155 * If the node's memory is already available, then kmem_cache_node is 3156 * already created. Nothing to do. 3157 */ 3158 if (nid < 0) 3159 return 0; 3160 3161 /* 3162 * We are bringing a node online. No memory is available yet. We must 3163 * allocate a kmem_cache_node structure in order to bring the node 3164 * online. 3165 */ 3166 down_read(&slub_lock); 3167 list_for_each_entry(s, &slab_caches, list) { 3168 /* 3169 * XXX: kmem_cache_alloc_node will fallback to other nodes 3170 * since memory is not yet available from the node that 3171 * is brought up. 3172 */ 3173 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 3174 if (!n) { 3175 ret = -ENOMEM; 3176 goto out; 3177 } 3178 init_kmem_cache_node(n, s); 3179 s->node[nid] = n; 3180 } 3181out: 3182 up_read(&slub_lock); 3183 return ret; 3184} 3185 3186static int slab_memory_callback(struct notifier_block *self, 3187 unsigned long action, void *arg) 3188{ 3189 int ret = 0; 3190 3191 switch (action) { 3192 case MEM_GOING_ONLINE: 3193 ret = slab_mem_going_online_callback(arg); 3194 break; 3195 case MEM_GOING_OFFLINE: 3196 ret = slab_mem_going_offline_callback(arg); 3197 break; 3198 case MEM_OFFLINE: 3199 case MEM_CANCEL_ONLINE: 3200 slab_mem_offline_callback(arg); 3201 break; 3202 case MEM_ONLINE: 3203 case MEM_CANCEL_OFFLINE: 3204 break; 3205 } 3206 if (ret) 3207 ret = notifier_from_errno(ret); 3208 else 3209 ret = NOTIFY_OK; 3210 return ret; 3211} 3212 3213#endif /* CONFIG_MEMORY_HOTPLUG */ 3214 3215/******************************************************************** 3216 * Basic setup of slabs 3217 *******************************************************************/ 3218 3219/* 3220 * Used for early kmem_cache structures that were allocated using 3221 * the page allocator 3222 */ 3223 3224static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) 3225{ 3226 int node; 3227 3228 list_add(&s->list, &slab_caches); 3229 s->refcount = -1; 3230 3231 for_each_node_state(node, N_NORMAL_MEMORY) { 3232 struct kmem_cache_node *n = get_node(s, node); 3233 struct page *p; 3234 3235 if (n) { 3236 list_for_each_entry(p, &n->partial, lru) 3237 p->slab = s; 3238 3239#ifdef CONFIG_SLUB_DEBUG 3240 list_for_each_entry(p, &n->full, lru) 3241 p->slab = s; 3242#endif 3243 } 3244 } 3245} 3246 3247void __init kmem_cache_init(void) 3248{ 3249 int i; 3250 int caches = 0; 3251 struct kmem_cache *temp_kmem_cache; 3252 int order; 3253 struct kmem_cache *temp_kmem_cache_node; 3254 unsigned long kmalloc_size; 3255 3256 kmem_size = offsetof(struct kmem_cache, node) + 3257 nr_node_ids * sizeof(struct kmem_cache_node *); 3258 3259 /* Allocate two kmem_caches from the page allocator */ 3260 kmalloc_size = ALIGN(kmem_size, cache_line_size()); 3261 order = get_order(2 * kmalloc_size); 3262 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); 3263 3264 /* 3265 * Must first have the slab cache available for the allocations of the 3266 * struct kmem_cache_node's. There is special bootstrap code in 3267 * kmem_cache_open for slab_state == DOWN. 3268 */ 3269 kmem_cache_node = (void *)kmem_cache + kmalloc_size; 3270 3271 kmem_cache_open(kmem_cache_node, "kmem_cache_node", 3272 sizeof(struct kmem_cache_node), 3273 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3274 3275 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 3276 3277 /* Able to allocate the per node structures */ 3278 slab_state = PARTIAL; 3279 3280 temp_kmem_cache = kmem_cache; 3281 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, 3282 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3283 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3284 memcpy(kmem_cache, temp_kmem_cache, kmem_size); 3285 3286 /* 3287 * Allocate kmem_cache_node properly from the kmem_cache slab. 3288 * kmem_cache_node is separately allocated so no need to 3289 * update any list pointers. 3290 */ 3291 temp_kmem_cache_node = kmem_cache_node; 3292 3293 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3294 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); 3295 3296 kmem_cache_bootstrap_fixup(kmem_cache_node); 3297 3298 caches++; 3299 kmem_cache_bootstrap_fixup(kmem_cache); 3300 caches++; 3301 /* Free temporary boot structure */ 3302 free_pages((unsigned long)temp_kmem_cache, order); 3303 3304 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 3305 3306 /* 3307 * Patch up the size_index table if we have strange large alignment 3308 * requirements for the kmalloc array. This is only the case for 3309 * MIPS it seems. The standard arches will not generate any code here. 3310 * 3311 * Largest permitted alignment is 256 bytes due to the way we 3312 * handle the index determination for the smaller caches. 3313 * 3314 * Make sure that nothing crazy happens if someone starts tinkering 3315 * around with ARCH_KMALLOC_MINALIGN 3316 */ 3317 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3318 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3319 3320 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { 3321 int elem = size_index_elem(i); 3322 if (elem >= ARRAY_SIZE(size_index)) 3323 break; 3324 size_index[elem] = KMALLOC_SHIFT_LOW; 3325 } 3326 3327 if (KMALLOC_MIN_SIZE == 64) { 3328 /* 3329 * The 96 byte size cache is not used if the alignment 3330 * is 64 byte. 3331 */ 3332 for (i = 64 + 8; i <= 96; i += 8) 3333 size_index[size_index_elem(i)] = 7; 3334 } else if (KMALLOC_MIN_SIZE == 128) { 3335 /* 3336 * The 192 byte sized cache is not used if the alignment 3337 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3338 * instead. 3339 */ 3340 for (i = 128 + 8; i <= 192; i += 8) 3341 size_index[size_index_elem(i)] = 8; 3342 } 3343 3344 /* Caches that are not of the two-to-the-power-of size */ 3345 if (KMALLOC_MIN_SIZE <= 32) { 3346 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); 3347 caches++; 3348 } 3349 3350 if (KMALLOC_MIN_SIZE <= 64) { 3351 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); 3352 caches++; 3353 } 3354 3355 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3356 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); 3357 caches++; 3358 } 3359 3360 slab_state = UP; 3361 3362 /* Provide the correct kmalloc names now that the caches are up */ 3363 if (KMALLOC_MIN_SIZE <= 32) { 3364 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); 3365 BUG_ON(!kmalloc_caches[1]->name); 3366 } 3367 3368 if (KMALLOC_MIN_SIZE <= 64) { 3369 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); 3370 BUG_ON(!kmalloc_caches[2]->name); 3371 } 3372 3373 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3374 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); 3375 3376 BUG_ON(!s); 3377 kmalloc_caches[i]->name = s; 3378 } 3379 3380#ifdef CONFIG_SMP 3381 register_cpu_notifier(&slab_notifier); 3382#endif 3383 3384#ifdef CONFIG_ZONE_DMA 3385 for (i = 0; i < SLUB_PAGE_SHIFT; i++) { 3386 struct kmem_cache *s = kmalloc_caches[i]; 3387 3388 if (s && s->size) { 3389 char *name = kasprintf(GFP_NOWAIT, 3390 "dma-kmalloc-%d", s->objsize); 3391 3392 BUG_ON(!name); 3393 kmalloc_dma_caches[i] = create_kmalloc_cache(name, 3394 s->objsize, SLAB_CACHE_DMA); 3395 } 3396 } 3397#endif 3398 printk(KERN_INFO 3399 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3400 " CPUs=%d, Nodes=%d\n", 3401 caches, cache_line_size(), 3402 slub_min_order, slub_max_order, slub_min_objects, 3403 nr_cpu_ids, nr_node_ids); 3404} 3405 3406void __init kmem_cache_init_late(void) 3407{ 3408} 3409 3410/* 3411 * Find a mergeable slab cache 3412 */ 3413static int slab_unmergeable(struct kmem_cache *s) 3414{ 3415 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3416 return 1; 3417 3418 if (s->ctor) 3419 return 1; 3420 3421 /* 3422 * We may have set a slab to be unmergeable during bootstrap. 3423 */ 3424 if (s->refcount < 0) 3425 return 1; 3426 3427 return 0; 3428} 3429 3430static struct kmem_cache *find_mergeable(size_t size, 3431 size_t align, unsigned long flags, const char *name, 3432 void (*ctor)(void *)) 3433{ 3434 struct kmem_cache *s; 3435 3436 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3437 return NULL; 3438 3439 if (ctor) 3440 return NULL; 3441 3442 size = ALIGN(size, sizeof(void *)); 3443 align = calculate_alignment(flags, align, size); 3444 size = ALIGN(size, align); 3445 flags = kmem_cache_flags(size, flags, name, NULL); 3446 3447 list_for_each_entry(s, &slab_caches, list) { 3448 if (slab_unmergeable(s)) 3449 continue; 3450 3451 if (size > s->size) 3452 continue; 3453 3454 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3455 continue; 3456 /* 3457 * Check if alignment is compatible. 3458 * Courtesy of Adrian Drzewiecki 3459 */ 3460 if ((s->size & ~(align - 1)) != s->size) 3461 continue; 3462 3463 if (s->size - size >= sizeof(void *)) 3464 continue; 3465 3466 return s; 3467 } 3468 return NULL; 3469} 3470 3471struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3472 size_t align, unsigned long flags, void (*ctor)(void *)) 3473{ 3474 struct kmem_cache *s; 3475 char *n; 3476 3477 if (WARN_ON(!name)) 3478 return NULL; 3479 3480 down_write(&slub_lock); 3481 s = find_mergeable(size, align, flags, name, ctor); 3482 if (s) { 3483 s->refcount++; 3484 /* 3485 * Adjust the object sizes so that we clear 3486 * the complete object on kzalloc. 3487 */ 3488 s->objsize = max(s->objsize, (int)size); 3489 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3490 3491 if (sysfs_slab_alias(s, name)) { 3492 s->refcount--; 3493 goto err; 3494 } 3495 up_write(&slub_lock); 3496 return s; 3497 } 3498 3499 n = kstrdup(name, GFP_KERNEL); 3500 if (!n) 3501 goto err; 3502 3503 s = kmalloc(kmem_size, GFP_KERNEL); 3504 if (s) { 3505 if (kmem_cache_open(s, n, 3506 size, align, flags, ctor)) { 3507 list_add(&s->list, &slab_caches); 3508 if (sysfs_slab_add(s)) { 3509 list_del(&s->list); 3510 kfree(n); 3511 kfree(s); 3512 goto err; 3513 } 3514 up_write(&slub_lock); 3515 return s; 3516 } 3517 kfree(n); 3518 kfree(s); 3519 } 3520err: 3521 up_write(&slub_lock); 3522 3523 if (flags & SLAB_PANIC) 3524 panic("Cannot create slabcache %s\n", name); 3525 else 3526 s = NULL; 3527 return s; 3528} 3529EXPORT_SYMBOL(kmem_cache_create); 3530 3531#ifdef CONFIG_SMP 3532/* 3533 * Use the cpu notifier to insure that the cpu slabs are flushed when 3534 * necessary. 3535 */ 3536static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3537 unsigned long action, void *hcpu) 3538{ 3539 long cpu = (long)hcpu; 3540 struct kmem_cache *s; 3541 unsigned long flags; 3542 3543 switch (action) { 3544 case CPU_UP_CANCELED: 3545 case CPU_UP_CANCELED_FROZEN: 3546 case CPU_DEAD: 3547 case CPU_DEAD_FROZEN: 3548 down_read(&slub_lock); 3549 list_for_each_entry(s, &slab_caches, list) { 3550 local_irq_save(flags); 3551 __flush_cpu_slab(s, cpu); 3552 local_irq_restore(flags); 3553 } 3554 up_read(&slub_lock); 3555 break; 3556 default: 3557 break; 3558 } 3559 return NOTIFY_OK; 3560} 3561 3562static struct notifier_block __cpuinitdata slab_notifier = { 3563 .notifier_call = slab_cpuup_callback 3564}; 3565 3566#endif 3567 3568void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3569{ 3570 struct kmem_cache *s; 3571 void *ret; 3572 3573 if (unlikely(size > SLUB_MAX_SIZE)) 3574 return kmalloc_large(size, gfpflags); 3575 3576 s = get_slab(size, gfpflags); 3577 3578 if (unlikely(ZERO_OR_NULL_PTR(s))) 3579 return s; 3580 3581 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); 3582 3583 /* Honor the call site pointer we received. */ 3584 trace_kmalloc(caller, ret, size, s->size, gfpflags); 3585 3586 return ret; 3587} 3588 3589#ifdef CONFIG_NUMA 3590void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3591 int node, unsigned long caller) 3592{ 3593 struct kmem_cache *s; 3594 void *ret; 3595 3596 if (unlikely(size > SLUB_MAX_SIZE)) { 3597 ret = kmalloc_large_node(size, gfpflags, node); 3598 3599 trace_kmalloc_node(caller, ret, 3600 size, PAGE_SIZE << get_order(size), 3601 gfpflags, node); 3602 3603 return ret; 3604 } 3605 3606 s = get_slab(size, gfpflags); 3607 3608 if (unlikely(ZERO_OR_NULL_PTR(s))) 3609 return s; 3610 3611 ret = slab_alloc(s, gfpflags, node, caller); 3612 3613 /* Honor the call site pointer we received. */ 3614 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 3615 3616 return ret; 3617} 3618#endif 3619 3620#ifdef CONFIG_SYSFS 3621static int count_inuse(struct page *page) 3622{ 3623 return page->inuse; 3624} 3625 3626static int count_total(struct page *page) 3627{ 3628 return page->objects; 3629} 3630#endif 3631 3632#ifdef CONFIG_SLUB_DEBUG 3633static int validate_slab(struct kmem_cache *s, struct page *page, 3634 unsigned long *map) 3635{ 3636 void *p; 3637 void *addr = page_address(page); 3638 3639 if (!check_slab(s, page) || 3640 !on_freelist(s, page, NULL)) 3641 return 0; 3642 3643 /* Now we know that a valid freelist exists */ 3644 bitmap_zero(map, page->objects); 3645 3646 get_map(s, page, map); 3647 for_each_object(p, s, addr, page->objects) { 3648 if (test_bit(slab_index(p, s, addr), map)) 3649 if (!check_object(s, page, p, SLUB_RED_INACTIVE)) 3650 return 0; 3651 } 3652 3653 for_each_object(p, s, addr, page->objects) 3654 if (!test_bit(slab_index(p, s, addr), map)) 3655 if (!check_object(s, page, p, SLUB_RED_ACTIVE)) 3656 return 0; 3657 return 1; 3658} 3659 3660static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3661 unsigned long *map) 3662{ 3663 if (slab_trylock(page)) { 3664 validate_slab(s, page, map); 3665 slab_unlock(page); 3666 } else 3667 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3668 s->name, page); 3669} 3670 3671static int validate_slab_node(struct kmem_cache *s, 3672 struct kmem_cache_node *n, unsigned long *map) 3673{ 3674 unsigned long count = 0; 3675 struct page *page; 3676 unsigned long flags; 3677 3678 spin_lock_irqsave(&n->list_lock, flags); 3679 3680 list_for_each_entry(page, &n->partial, lru) { 3681 validate_slab_slab(s, page, map); 3682 count++; 3683 } 3684 if (count != n->nr_partial) 3685 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3686 "counter=%ld\n", s->name, count, n->nr_partial); 3687 3688 if (!(s->flags & SLAB_STORE_USER)) 3689 goto out; 3690 3691 list_for_each_entry(page, &n->full, lru) { 3692 validate_slab_slab(s, page, map); 3693 count++; 3694 } 3695 if (count != atomic_long_read(&n->nr_slabs)) 3696 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3697 "counter=%ld\n", s->name, count, 3698 atomic_long_read(&n->nr_slabs)); 3699 3700out: 3701 spin_unlock_irqrestore(&n->list_lock, flags); 3702 return count; 3703} 3704 3705static long validate_slab_cache(struct kmem_cache *s) 3706{ 3707 int node; 3708 unsigned long count = 0; 3709 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3710 sizeof(unsigned long), GFP_KERNEL); 3711 3712 if (!map) 3713 return -ENOMEM; 3714 3715 flush_all(s); 3716 for_each_node_state(node, N_NORMAL_MEMORY) { 3717 struct kmem_cache_node *n = get_node(s, node); 3718 3719 count += validate_slab_node(s, n, map); 3720 } 3721 kfree(map); 3722 return count; 3723} 3724/* 3725 * Generate lists of code addresses where slabcache objects are allocated 3726 * and freed. 3727 */ 3728 3729struct location { 3730 unsigned long count; 3731 unsigned long addr; 3732 long long sum_time; 3733 long min_time; 3734 long max_time; 3735 long min_pid; 3736 long max_pid; 3737 DECLARE_BITMAP(cpus, NR_CPUS); 3738 nodemask_t nodes; 3739}; 3740 3741struct loc_track { 3742 unsigned long max; 3743 unsigned long count; 3744 struct location *loc; 3745}; 3746 3747static void free_loc_track(struct loc_track *t) 3748{ 3749 if (t->max) 3750 free_pages((unsigned long)t->loc, 3751 get_order(sizeof(struct location) * t->max)); 3752} 3753 3754static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3755{ 3756 struct location *l; 3757 int order; 3758 3759 order = get_order(sizeof(struct location) * max); 3760 3761 l = (void *)__get_free_pages(flags, order); 3762 if (!l) 3763 return 0; 3764 3765 if (t->count) { 3766 memcpy(l, t->loc, sizeof(struct location) * t->count); 3767 free_loc_track(t); 3768 } 3769 t->max = max; 3770 t->loc = l; 3771 return 1; 3772} 3773 3774static int add_location(struct loc_track *t, struct kmem_cache *s, 3775 const struct track *track) 3776{ 3777 long start, end, pos; 3778 struct location *l; 3779 unsigned long caddr; 3780 unsigned long age = jiffies - track->when; 3781 3782 start = -1; 3783 end = t->count; 3784 3785 for ( ; ; ) { 3786 pos = start + (end - start + 1) / 2; 3787 3788 /* 3789 * There is nothing at "end". If we end up there 3790 * we need to add something to before end. 3791 */ 3792 if (pos == end) 3793 break; 3794 3795 caddr = t->loc[pos].addr; 3796 if (track->addr == caddr) { 3797 3798 l = &t->loc[pos]; 3799 l->count++; 3800 if (track->when) { 3801 l->sum_time += age; 3802 if (age < l->min_time) 3803 l->min_time = age; 3804 if (age > l->max_time) 3805 l->max_time = age; 3806 3807 if (track->pid < l->min_pid) 3808 l->min_pid = track->pid; 3809 if (track->pid > l->max_pid) 3810 l->max_pid = track->pid; 3811 3812 cpumask_set_cpu(track->cpu, 3813 to_cpumask(l->cpus)); 3814 } 3815 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3816 return 1; 3817 } 3818 3819 if (track->addr < caddr) 3820 end = pos; 3821 else 3822 start = pos; 3823 } 3824 3825 /* 3826 * Not found. Insert new tracking element. 3827 */ 3828 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3829 return 0; 3830 3831 l = t->loc + pos; 3832 if (pos < t->count) 3833 memmove(l + 1, l, 3834 (t->count - pos) * sizeof(struct location)); 3835 t->count++; 3836 l->count = 1; 3837 l->addr = track->addr; 3838 l->sum_time = age; 3839 l->min_time = age; 3840 l->max_time = age; 3841 l->min_pid = track->pid; 3842 l->max_pid = track->pid; 3843 cpumask_clear(to_cpumask(l->cpus)); 3844 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 3845 nodes_clear(l->nodes); 3846 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3847 return 1; 3848} 3849 3850static void process_slab(struct loc_track *t, struct kmem_cache *s, 3851 struct page *page, enum track_item alloc, 3852 unsigned long *map) 3853{ 3854 void *addr = page_address(page); 3855 void *p; 3856 3857 bitmap_zero(map, page->objects); 3858 get_map(s, page, map); 3859 3860 for_each_object(p, s, addr, page->objects) 3861 if (!test_bit(slab_index(p, s, addr), map)) 3862 add_location(t, s, get_track(s, p, alloc)); 3863} 3864 3865static int list_locations(struct kmem_cache *s, char *buf, 3866 enum track_item alloc) 3867{ 3868 int len = 0; 3869 unsigned long i; 3870 struct loc_track t = { 0, 0, NULL }; 3871 int node; 3872 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3873 sizeof(unsigned long), GFP_KERNEL); 3874 3875 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3876 GFP_TEMPORARY)) { 3877 kfree(map); 3878 return sprintf(buf, "Out of memory\n"); 3879 } 3880 /* Push back cpu slabs */ 3881 flush_all(s); 3882 3883 for_each_node_state(node, N_NORMAL_MEMORY) { 3884 struct kmem_cache_node *n = get_node(s, node); 3885 unsigned long flags; 3886 struct page *page; 3887 3888 if (!atomic_long_read(&n->nr_slabs)) 3889 continue; 3890 3891 spin_lock_irqsave(&n->list_lock, flags); 3892 list_for_each_entry(page, &n->partial, lru) 3893 process_slab(&t, s, page, alloc, map); 3894 list_for_each_entry(page, &n->full, lru) 3895 process_slab(&t, s, page, alloc, map); 3896 spin_unlock_irqrestore(&n->list_lock, flags); 3897 } 3898 3899 for (i = 0; i < t.count; i++) { 3900 struct location *l = &t.loc[i]; 3901 3902 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 3903 break; 3904 len += sprintf(buf + len, "%7ld ", l->count); 3905 3906 if (l->addr) 3907 len += sprintf(buf + len, "%pS", (void *)l->addr); 3908 else 3909 len += sprintf(buf + len, "<not-available>"); 3910 3911 if (l->sum_time != l->min_time) { 3912 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3913 l->min_time, 3914 (long)div_u64(l->sum_time, l->count), 3915 l->max_time); 3916 } else 3917 len += sprintf(buf + len, " age=%ld", 3918 l->min_time); 3919 3920 if (l->min_pid != l->max_pid) 3921 len += sprintf(buf + len, " pid=%ld-%ld", 3922 l->min_pid, l->max_pid); 3923 else 3924 len += sprintf(buf + len, " pid=%ld", 3925 l->min_pid); 3926 3927 if (num_online_cpus() > 1 && 3928 !cpumask_empty(to_cpumask(l->cpus)) && 3929 len < PAGE_SIZE - 60) { 3930 len += sprintf(buf + len, " cpus="); 3931 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3932 to_cpumask(l->cpus)); 3933 } 3934 3935 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 3936 len < PAGE_SIZE - 60) { 3937 len += sprintf(buf + len, " nodes="); 3938 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3939 l->nodes); 3940 } 3941 3942 len += sprintf(buf + len, "\n"); 3943 } 3944 3945 free_loc_track(&t); 3946 kfree(map); 3947 if (!t.count) 3948 len += sprintf(buf, "No data\n"); 3949 return len; 3950} 3951#endif 3952 3953#ifdef SLUB_RESILIENCY_TEST 3954static void resiliency_test(void) 3955{ 3956 u8 *p; 3957 3958 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); 3959 3960 printk(KERN_ERR "SLUB resiliency testing\n"); 3961 printk(KERN_ERR "-----------------------\n"); 3962 printk(KERN_ERR "A. Corruption after allocation\n"); 3963 3964 p = kzalloc(16, GFP_KERNEL); 3965 p[16] = 0x12; 3966 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3967 " 0x12->0x%p\n\n", p + 16); 3968 3969 validate_slab_cache(kmalloc_caches[4]); 3970 3971 /* Hmmm... The next two are dangerous */ 3972 p = kzalloc(32, GFP_KERNEL); 3973 p[32 + sizeof(void *)] = 0x34; 3974 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3975 " 0x34 -> -0x%p\n", p); 3976 printk(KERN_ERR 3977 "If allocated object is overwritten then not detectable\n\n"); 3978 3979 validate_slab_cache(kmalloc_caches[5]); 3980 p = kzalloc(64, GFP_KERNEL); 3981 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3982 *p = 0x56; 3983 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3984 p); 3985 printk(KERN_ERR 3986 "If allocated object is overwritten then not detectable\n\n"); 3987 validate_slab_cache(kmalloc_caches[6]); 3988 3989 printk(KERN_ERR "\nB. Corruption after free\n"); 3990 p = kzalloc(128, GFP_KERNEL); 3991 kfree(p); 3992 *p = 0x78; 3993 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3994 validate_slab_cache(kmalloc_caches[7]); 3995 3996 p = kzalloc(256, GFP_KERNEL); 3997 kfree(p); 3998 p[50] = 0x9a; 3999 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 4000 p); 4001 validate_slab_cache(kmalloc_caches[8]); 4002 4003 p = kzalloc(512, GFP_KERNEL); 4004 kfree(p); 4005 p[512] = 0xab; 4006 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 4007 validate_slab_cache(kmalloc_caches[9]); 4008} 4009#else 4010#ifdef CONFIG_SYSFS 4011static void resiliency_test(void) {}; 4012#endif 4013#endif 4014 4015#ifdef CONFIG_SYSFS 4016enum slab_stat_type { 4017 SL_ALL, /* All slabs */ 4018 SL_PARTIAL, /* Only partially allocated slabs */ 4019 SL_CPU, /* Only slabs used for cpu caches */ 4020 SL_OBJECTS, /* Determine allocated objects not slabs */ 4021 SL_TOTAL /* Determine object capacity not slabs */ 4022}; 4023 4024#define SO_ALL (1 << SL_ALL) 4025#define SO_PARTIAL (1 << SL_PARTIAL) 4026#define SO_CPU (1 << SL_CPU) 4027#define SO_OBJECTS (1 << SL_OBJECTS) 4028#define SO_TOTAL (1 << SL_TOTAL) 4029 4030static ssize_t show_slab_objects(struct kmem_cache *s, 4031 char *buf, unsigned long flags) 4032{ 4033 unsigned long total = 0; 4034 int node; 4035 int x; 4036 unsigned long *nodes; 4037 unsigned long *per_cpu; 4038 4039 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 4040 if (!nodes) 4041 return -ENOMEM; 4042 per_cpu = nodes + nr_node_ids; 4043 4044 if (flags & SO_CPU) { 4045 int cpu; 4046 4047 for_each_possible_cpu(cpu) { 4048 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 4049 4050 if (!c || c->node < 0) 4051 continue; 4052 4053 if (c->page) { 4054 if (flags & SO_TOTAL) 4055 x = c->page->objects; 4056 else if (flags & SO_OBJECTS) 4057 x = c->page->inuse; 4058 else 4059 x = 1; 4060 4061 total += x; 4062 nodes[c->node] += x; 4063 } 4064 per_cpu[c->node]++; 4065 } 4066 } 4067 4068 lock_memory_hotplug(); 4069#ifdef CONFIG_SLUB_DEBUG 4070 if (flags & SO_ALL) { 4071 for_each_node_state(node, N_NORMAL_MEMORY) { 4072 struct kmem_cache_node *n = get_node(s, node); 4073 4074 if (flags & SO_TOTAL) 4075 x = atomic_long_read(&n->total_objects); 4076 else if (flags & SO_OBJECTS) 4077 x = atomic_long_read(&n->total_objects) - 4078 count_partial(n, count_free); 4079 4080 else 4081 x = atomic_long_read(&n->nr_slabs); 4082 total += x; 4083 nodes[node] += x; 4084 } 4085 4086 } else 4087#endif 4088 if (flags & SO_PARTIAL) { 4089 for_each_node_state(node, N_NORMAL_MEMORY) { 4090 struct kmem_cache_node *n = get_node(s, node); 4091 4092 if (flags & SO_TOTAL) 4093 x = count_partial(n, count_total); 4094 else if (flags & SO_OBJECTS) 4095 x = count_partial(n, count_inuse); 4096 else 4097 x = n->nr_partial; 4098 total += x; 4099 nodes[node] += x; 4100 } 4101 } 4102 x = sprintf(buf, "%lu", total); 4103#ifdef CONFIG_NUMA 4104 for_each_node_state(node, N_NORMAL_MEMORY) 4105 if (nodes[node]) 4106 x += sprintf(buf + x, " N%d=%lu", 4107 node, nodes[node]); 4108#endif 4109 unlock_memory_hotplug(); 4110 kfree(nodes); 4111 return x + sprintf(buf + x, "\n"); 4112} 4113 4114#ifdef CONFIG_SLUB_DEBUG 4115static int any_slab_objects(struct kmem_cache *s) 4116{ 4117 int node; 4118 4119 for_each_online_node(node) { 4120 struct kmem_cache_node *n = get_node(s, node); 4121 4122 if (!n) 4123 continue; 4124 4125 if (atomic_long_read(&n->total_objects)) 4126 return 1; 4127 } 4128 return 0; 4129} 4130#endif 4131 4132#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 4133#define to_slab(n) container_of(n, struct kmem_cache, kobj); 4134 4135struct slab_attribute { 4136 struct attribute attr; 4137 ssize_t (*show)(struct kmem_cache *s, char *buf); 4138 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 4139}; 4140 4141#define SLAB_ATTR_RO(_name) \ 4142 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 4143 4144#define SLAB_ATTR(_name) \ 4145 static struct slab_attribute _name##_attr = \ 4146 __ATTR(_name, 0644, _name##_show, _name##_store) 4147 4148static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 4149{ 4150 return sprintf(buf, "%d\n", s->size); 4151} 4152SLAB_ATTR_RO(slab_size); 4153 4154static ssize_t align_show(struct kmem_cache *s, char *buf) 4155{ 4156 return sprintf(buf, "%d\n", s->align); 4157} 4158SLAB_ATTR_RO(align); 4159 4160static ssize_t object_size_show(struct kmem_cache *s, char *buf) 4161{ 4162 return sprintf(buf, "%d\n", s->objsize); 4163} 4164SLAB_ATTR_RO(object_size); 4165 4166static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 4167{ 4168 return sprintf(buf, "%d\n", oo_objects(s->oo)); 4169} 4170SLAB_ATTR_RO(objs_per_slab); 4171 4172static ssize_t order_store(struct kmem_cache *s, 4173 const char *buf, size_t length) 4174{ 4175 unsigned long order; 4176 int err; 4177 4178 err = strict_strtoul(buf, 10, &order); 4179 if (err) 4180 return err; 4181 4182 if (order > slub_max_order || order < slub_min_order) 4183 return -EINVAL; 4184 4185 calculate_sizes(s, order); 4186 return length; 4187} 4188 4189static ssize_t order_show(struct kmem_cache *s, char *buf) 4190{ 4191 return sprintf(buf, "%d\n", oo_order(s->oo)); 4192} 4193SLAB_ATTR(order); 4194 4195static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 4196{ 4197 return sprintf(buf, "%lu\n", s->min_partial); 4198} 4199 4200static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 4201 size_t length) 4202{ 4203 unsigned long min; 4204 int err; 4205 4206 err = strict_strtoul(buf, 10, &min); 4207 if (err) 4208 return err; 4209 4210 set_min_partial(s, min); 4211 return length; 4212} 4213SLAB_ATTR(min_partial); 4214 4215static ssize_t ctor_show(struct kmem_cache *s, char *buf) 4216{ 4217 if (!s->ctor) 4218 return 0; 4219 return sprintf(buf, "%pS\n", s->ctor); 4220} 4221SLAB_ATTR_RO(ctor); 4222 4223static ssize_t aliases_show(struct kmem_cache *s, char *buf) 4224{ 4225 return sprintf(buf, "%d\n", s->refcount - 1); 4226} 4227SLAB_ATTR_RO(aliases); 4228 4229static ssize_t partial_show(struct kmem_cache *s, char *buf) 4230{ 4231 return show_slab_objects(s, buf, SO_PARTIAL); 4232} 4233SLAB_ATTR_RO(partial); 4234 4235static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4236{ 4237 return show_slab_objects(s, buf, SO_CPU); 4238} 4239SLAB_ATTR_RO(cpu_slabs); 4240 4241static ssize_t objects_show(struct kmem_cache *s, char *buf) 4242{ 4243 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4244} 4245SLAB_ATTR_RO(objects); 4246 4247static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4248{ 4249 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4250} 4251SLAB_ATTR_RO(objects_partial); 4252 4253static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4254{ 4255 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4256} 4257 4258static ssize_t reclaim_account_store(struct kmem_cache *s, 4259 const char *buf, size_t length) 4260{ 4261 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4262 if (buf[0] == '1') 4263 s->flags |= SLAB_RECLAIM_ACCOUNT; 4264 return length; 4265} 4266SLAB_ATTR(reclaim_account); 4267 4268static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4269{ 4270 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4271} 4272SLAB_ATTR_RO(hwcache_align); 4273 4274#ifdef CONFIG_ZONE_DMA 4275static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4276{ 4277 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4278} 4279SLAB_ATTR_RO(cache_dma); 4280#endif 4281 4282static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4283{ 4284 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4285} 4286SLAB_ATTR_RO(destroy_by_rcu); 4287 4288static ssize_t reserved_show(struct kmem_cache *s, char *buf) 4289{ 4290 return sprintf(buf, "%d\n", s->reserved); 4291} 4292SLAB_ATTR_RO(reserved); 4293 4294#ifdef CONFIG_SLUB_DEBUG 4295static ssize_t slabs_show(struct kmem_cache *s, char *buf) 4296{ 4297 return show_slab_objects(s, buf, SO_ALL); 4298} 4299SLAB_ATTR_RO(slabs); 4300 4301static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 4302{ 4303 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 4304} 4305SLAB_ATTR_RO(total_objects); 4306 4307static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 4308{ 4309 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 4310} 4311 4312static ssize_t sanity_checks_store(struct kmem_cache *s, 4313 const char *buf, size_t length) 4314{ 4315 s->flags &= ~SLAB_DEBUG_FREE; 4316 if (buf[0] == '1') 4317 s->flags |= SLAB_DEBUG_FREE; 4318 return length; 4319} 4320SLAB_ATTR(sanity_checks); 4321 4322static ssize_t trace_show(struct kmem_cache *s, char *buf) 4323{ 4324 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4325} 4326 4327static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4328 size_t length) 4329{ 4330 s->flags &= ~SLAB_TRACE; 4331 if (buf[0] == '1') 4332 s->flags |= SLAB_TRACE; 4333 return length; 4334} 4335SLAB_ATTR(trace); 4336 4337static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4338{ 4339 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4340} 4341 4342static ssize_t red_zone_store(struct kmem_cache *s, 4343 const char *buf, size_t length) 4344{ 4345 if (any_slab_objects(s)) 4346 return -EBUSY; 4347 4348 s->flags &= ~SLAB_RED_ZONE; 4349 if (buf[0] == '1') 4350 s->flags |= SLAB_RED_ZONE; 4351 calculate_sizes(s, -1); 4352 return length; 4353} 4354SLAB_ATTR(red_zone); 4355 4356static ssize_t poison_show(struct kmem_cache *s, char *buf) 4357{ 4358 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4359} 4360 4361static ssize_t poison_store(struct kmem_cache *s, 4362 const char *buf, size_t length) 4363{ 4364 if (any_slab_objects(s)) 4365 return -EBUSY; 4366 4367 s->flags &= ~SLAB_POISON; 4368 if (buf[0] == '1') 4369 s->flags |= SLAB_POISON; 4370 calculate_sizes(s, -1); 4371 return length; 4372} 4373SLAB_ATTR(poison); 4374 4375static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4376{ 4377 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4378} 4379 4380static ssize_t store_user_store(struct kmem_cache *s, 4381 const char *buf, size_t length) 4382{ 4383 if (any_slab_objects(s)) 4384 return -EBUSY; 4385 4386 s->flags &= ~SLAB_STORE_USER; 4387 if (buf[0] == '1') 4388 s->flags |= SLAB_STORE_USER; 4389 calculate_sizes(s, -1); 4390 return length; 4391} 4392SLAB_ATTR(store_user); 4393 4394static ssize_t validate_show(struct kmem_cache *s, char *buf) 4395{ 4396 return 0; 4397} 4398 4399static ssize_t validate_store(struct kmem_cache *s, 4400 const char *buf, size_t length) 4401{ 4402 int ret = -EINVAL; 4403 4404 if (buf[0] == '1') { 4405 ret = validate_slab_cache(s); 4406 if (ret >= 0) 4407 ret = length; 4408 } 4409 return ret; 4410} 4411SLAB_ATTR(validate); 4412 4413static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4414{ 4415 if (!(s->flags & SLAB_STORE_USER)) 4416 return -ENOSYS; 4417 return list_locations(s, buf, TRACK_ALLOC); 4418} 4419SLAB_ATTR_RO(alloc_calls); 4420 4421static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4422{ 4423 if (!(s->flags & SLAB_STORE_USER)) 4424 return -ENOSYS; 4425 return list_locations(s, buf, TRACK_FREE); 4426} 4427SLAB_ATTR_RO(free_calls); 4428#endif /* CONFIG_SLUB_DEBUG */ 4429 4430#ifdef CONFIG_FAILSLAB 4431static ssize_t failslab_show(struct kmem_cache *s, char *buf) 4432{ 4433 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 4434} 4435 4436static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 4437 size_t length) 4438{ 4439 s->flags &= ~SLAB_FAILSLAB; 4440 if (buf[0] == '1') 4441 s->flags |= SLAB_FAILSLAB; 4442 return length; 4443} 4444SLAB_ATTR(failslab); 4445#endif 4446 4447static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4448{ 4449 return 0; 4450} 4451 4452static ssize_t shrink_store(struct kmem_cache *s, 4453 const char *buf, size_t length) 4454{ 4455 if (buf[0] == '1') { 4456 int rc = kmem_cache_shrink(s); 4457 4458 if (rc) 4459 return rc; 4460 } else 4461 return -EINVAL; 4462 return length; 4463} 4464SLAB_ATTR(shrink); 4465 4466#ifdef CONFIG_NUMA 4467static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4468{ 4469 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4470} 4471 4472static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4473 const char *buf, size_t length) 4474{ 4475 unsigned long ratio; 4476 int err; 4477 4478 err = strict_strtoul(buf, 10, &ratio); 4479 if (err) 4480 return err; 4481 4482 if (ratio <= 100) 4483 s->remote_node_defrag_ratio = ratio * 10; 4484 4485 return length; 4486} 4487SLAB_ATTR(remote_node_defrag_ratio); 4488#endif 4489 4490#ifdef CONFIG_SLUB_STATS 4491static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4492{ 4493 unsigned long sum = 0; 4494 int cpu; 4495 int len; 4496 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4497 4498 if (!data) 4499 return -ENOMEM; 4500 4501 for_each_online_cpu(cpu) { 4502 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 4503 4504 data[cpu] = x; 4505 sum += x; 4506 } 4507 4508 len = sprintf(buf, "%lu", sum); 4509 4510#ifdef CONFIG_SMP 4511 for_each_online_cpu(cpu) { 4512 if (data[cpu] && len < PAGE_SIZE - 20) 4513 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4514 } 4515#endif 4516 kfree(data); 4517 return len + sprintf(buf + len, "\n"); 4518} 4519 4520static void clear_stat(struct kmem_cache *s, enum stat_item si) 4521{ 4522 int cpu; 4523 4524 for_each_online_cpu(cpu) 4525 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 4526} 4527 4528#define STAT_ATTR(si, text) \ 4529static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4530{ \ 4531 return show_stat(s, buf, si); \ 4532} \ 4533static ssize_t text##_store(struct kmem_cache *s, \ 4534 const char *buf, size_t length) \ 4535{ \ 4536 if (buf[0] != '0') \ 4537 return -EINVAL; \ 4538 clear_stat(s, si); \ 4539 return length; \ 4540} \ 4541SLAB_ATTR(text); \ 4542 4543STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4544STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4545STAT_ATTR(FREE_FASTPATH, free_fastpath); 4546STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4547STAT_ATTR(FREE_FROZEN, free_frozen); 4548STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4549STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4550STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4551STAT_ATTR(ALLOC_SLAB, alloc_slab); 4552STAT_ATTR(ALLOC_REFILL, alloc_refill); 4553STAT_ATTR(FREE_SLAB, free_slab); 4554STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4555STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4556STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4557STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4558STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4559STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4560STAT_ATTR(ORDER_FALLBACK, order_fallback); 4561#endif 4562 4563static struct attribute *slab_attrs[] = { 4564 &slab_size_attr.attr, 4565 &object_size_attr.attr, 4566 &objs_per_slab_attr.attr, 4567 &order_attr.attr, 4568 &min_partial_attr.attr, 4569 &objects_attr.attr, 4570 &objects_partial_attr.attr, 4571 &partial_attr.attr, 4572 &cpu_slabs_attr.attr, 4573 &ctor_attr.attr, 4574 &aliases_attr.attr, 4575 &align_attr.attr, 4576 &hwcache_align_attr.attr, 4577 &reclaim_account_attr.attr, 4578 &destroy_by_rcu_attr.attr, 4579 &shrink_attr.attr, 4580 &reserved_attr.attr, 4581#ifdef CONFIG_SLUB_DEBUG 4582 &total_objects_attr.attr, 4583 &slabs_attr.attr, 4584 &sanity_checks_attr.attr, 4585 &trace_attr.attr, 4586 &red_zone_attr.attr, 4587 &poison_attr.attr, 4588 &store_user_attr.attr, 4589 &validate_attr.attr, 4590 &alloc_calls_attr.attr, 4591 &free_calls_attr.attr, 4592#endif 4593#ifdef CONFIG_ZONE_DMA 4594 &cache_dma_attr.attr, 4595#endif 4596#ifdef CONFIG_NUMA 4597 &remote_node_defrag_ratio_attr.attr, 4598#endif 4599#ifdef CONFIG_SLUB_STATS 4600 &alloc_fastpath_attr.attr, 4601 &alloc_slowpath_attr.attr, 4602 &free_fastpath_attr.attr, 4603 &free_slowpath_attr.attr, 4604 &free_frozen_attr.attr, 4605 &free_add_partial_attr.attr, 4606 &free_remove_partial_attr.attr, 4607 &alloc_from_partial_attr.attr, 4608 &alloc_slab_attr.attr, 4609 &alloc_refill_attr.attr, 4610 &free_slab_attr.attr, 4611 &cpuslab_flush_attr.attr, 4612 &deactivate_full_attr.attr, 4613 &deactivate_empty_attr.attr, 4614 &deactivate_to_head_attr.attr, 4615 &deactivate_to_tail_attr.attr, 4616 &deactivate_remote_frees_attr.attr, 4617 &order_fallback_attr.attr, 4618#endif 4619#ifdef CONFIG_FAILSLAB 4620 &failslab_attr.attr, 4621#endif 4622 4623 NULL 4624}; 4625 4626static struct attribute_group slab_attr_group = { 4627 .attrs = slab_attrs, 4628}; 4629 4630static ssize_t slab_attr_show(struct kobject *kobj, 4631 struct attribute *attr, 4632 char *buf) 4633{ 4634 struct slab_attribute *attribute; 4635 struct kmem_cache *s; 4636 int err; 4637 4638 attribute = to_slab_attr(attr); 4639 s = to_slab(kobj); 4640 4641 if (!attribute->show) 4642 return -EIO; 4643 4644 err = attribute->show(s, buf); 4645 4646 return err; 4647} 4648 4649static ssize_t slab_attr_store(struct kobject *kobj, 4650 struct attribute *attr, 4651 const char *buf, size_t len) 4652{ 4653 struct slab_attribute *attribute; 4654 struct kmem_cache *s; 4655 int err; 4656 4657 attribute = to_slab_attr(attr); 4658 s = to_slab(kobj); 4659 4660 if (!attribute->store) 4661 return -EIO; 4662 4663 err = attribute->store(s, buf, len); 4664 4665 return err; 4666} 4667 4668static void kmem_cache_release(struct kobject *kobj) 4669{ 4670 struct kmem_cache *s = to_slab(kobj); 4671 4672 kfree(s->name); 4673 kfree(s); 4674} 4675 4676static const struct sysfs_ops slab_sysfs_ops = { 4677 .show = slab_attr_show, 4678 .store = slab_attr_store, 4679}; 4680 4681static struct kobj_type slab_ktype = { 4682 .sysfs_ops = &slab_sysfs_ops, 4683 .release = kmem_cache_release 4684}; 4685 4686static int uevent_filter(struct kset *kset, struct kobject *kobj) 4687{ 4688 struct kobj_type *ktype = get_ktype(kobj); 4689 4690 if (ktype == &slab_ktype) 4691 return 1; 4692 return 0; 4693} 4694 4695static const struct kset_uevent_ops slab_uevent_ops = { 4696 .filter = uevent_filter, 4697}; 4698 4699static struct kset *slab_kset; 4700 4701#define ID_STR_LENGTH 64 4702 4703/* Create a unique string id for a slab cache: 4704 * 4705 * Format :[flags-]size 4706 */ 4707static char *create_unique_id(struct kmem_cache *s) 4708{ 4709 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4710 char *p = name; 4711 4712 BUG_ON(!name); 4713 4714 *p++ = ':'; 4715 /* 4716 * First flags affecting slabcache operations. We will only 4717 * get here for aliasable slabs so we do not need to support 4718 * too many flags. The flags here must cover all flags that 4719 * are matched during merging to guarantee that the id is 4720 * unique. 4721 */ 4722 if (s->flags & SLAB_CACHE_DMA) 4723 *p++ = 'd'; 4724 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4725 *p++ = 'a'; 4726 if (s->flags & SLAB_DEBUG_FREE) 4727 *p++ = 'F'; 4728 if (!(s->flags & SLAB_NOTRACK)) 4729 *p++ = 't'; 4730 if (p != name + 1) 4731 *p++ = '-'; 4732 p += sprintf(p, "%07d", s->size); 4733 BUG_ON(p > name + ID_STR_LENGTH - 1); 4734 return name; 4735} 4736 4737static int sysfs_slab_add(struct kmem_cache *s) 4738{ 4739 int err; 4740 const char *name; 4741 int unmergeable; 4742 4743 if (slab_state < SYSFS) 4744 /* Defer until later */ 4745 return 0; 4746 4747 unmergeable = slab_unmergeable(s); 4748 if (unmergeable) { 4749 /* 4750 * Slabcache can never be merged so we can use the name proper. 4751 * This is typically the case for debug situations. In that 4752 * case we can catch duplicate names easily. 4753 */ 4754 sysfs_remove_link(&slab_kset->kobj, s->name); 4755 name = s->name; 4756 } else { 4757 /* 4758 * Create a unique name for the slab as a target 4759 * for the symlinks. 4760 */ 4761 name = create_unique_id(s); 4762 } 4763 4764 s->kobj.kset = slab_kset; 4765 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4766 if (err) { 4767 kobject_put(&s->kobj); 4768 return err; 4769 } 4770 4771 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4772 if (err) { 4773 kobject_del(&s->kobj); 4774 kobject_put(&s->kobj); 4775 return err; 4776 } 4777 kobject_uevent(&s->kobj, KOBJ_ADD); 4778 if (!unmergeable) { 4779 /* Setup first alias */ 4780 sysfs_slab_alias(s, s->name); 4781 kfree(name); 4782 } 4783 return 0; 4784} 4785 4786static void sysfs_slab_remove(struct kmem_cache *s) 4787{ 4788 if (slab_state < SYSFS) 4789 /* 4790 * Sysfs has not been setup yet so no need to remove the 4791 * cache from sysfs. 4792 */ 4793 return; 4794 4795 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4796 kobject_del(&s->kobj); 4797 kobject_put(&s->kobj); 4798} 4799 4800/* 4801 * Need to buffer aliases during bootup until sysfs becomes 4802 * available lest we lose that information. 4803 */ 4804struct saved_alias { 4805 struct kmem_cache *s; 4806 const char *name; 4807 struct saved_alias *next; 4808}; 4809 4810static struct saved_alias *alias_list; 4811 4812static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4813{ 4814 struct saved_alias *al; 4815 4816 if (slab_state == SYSFS) { 4817 /* 4818 * If we have a leftover link then remove it. 4819 */ 4820 sysfs_remove_link(&slab_kset->kobj, name); 4821 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4822 } 4823 4824 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4825 if (!al) 4826 return -ENOMEM; 4827 4828 al->s = s; 4829 al->name = name; 4830 al->next = alias_list; 4831 alias_list = al; 4832 return 0; 4833} 4834 4835static int __init slab_sysfs_init(void) 4836{ 4837 struct kmem_cache *s; 4838 int err; 4839 4840 down_write(&slub_lock); 4841 4842 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4843 if (!slab_kset) { 4844 up_write(&slub_lock); 4845 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4846 return -ENOSYS; 4847 } 4848 4849 slab_state = SYSFS; 4850 4851 list_for_each_entry(s, &slab_caches, list) { 4852 err = sysfs_slab_add(s); 4853 if (err) 4854 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4855 " to sysfs\n", s->name); 4856 } 4857 4858 while (alias_list) { 4859 struct saved_alias *al = alias_list; 4860 4861 alias_list = alias_list->next; 4862 err = sysfs_slab_alias(al->s, al->name); 4863 if (err) 4864 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4865 " %s to sysfs\n", s->name); 4866 kfree(al); 4867 } 4868 4869 up_write(&slub_lock); 4870 resiliency_test(); 4871 return 0; 4872} 4873 4874__initcall(slab_sysfs_init); 4875#endif /* CONFIG_SYSFS */ 4876 4877/* 4878 * The /proc/slabinfo ABI 4879 */ 4880#ifdef CONFIG_SLABINFO 4881static void print_slabinfo_header(struct seq_file *m) 4882{ 4883 seq_puts(m, "slabinfo - version: 2.1\n"); 4884 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4885 "<objperslab> <pagesperslab>"); 4886 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4887 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4888 seq_putc(m, '\n'); 4889} 4890 4891static void *s_start(struct seq_file *m, loff_t *pos) 4892{ 4893 loff_t n = *pos; 4894 4895 down_read(&slub_lock); 4896 if (!n) 4897 print_slabinfo_header(m); 4898 4899 return seq_list_start(&slab_caches, *pos); 4900} 4901 4902static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4903{ 4904 return seq_list_next(p, &slab_caches, pos); 4905} 4906 4907static void s_stop(struct seq_file *m, void *p) 4908{ 4909 up_read(&slub_lock); 4910} 4911 4912static int s_show(struct seq_file *m, void *p) 4913{ 4914 unsigned long nr_partials = 0; 4915 unsigned long nr_slabs = 0; 4916 unsigned long nr_inuse = 0; 4917 unsigned long nr_objs = 0; 4918 unsigned long nr_free = 0; 4919 struct kmem_cache *s; 4920 int node; 4921 4922 s = list_entry(p, struct kmem_cache, list); 4923 4924 for_each_online_node(node) { 4925 struct kmem_cache_node *n = get_node(s, node); 4926 4927 if (!n) 4928 continue; 4929 4930 nr_partials += n->nr_partial; 4931 nr_slabs += atomic_long_read(&n->nr_slabs); 4932 nr_objs += atomic_long_read(&n->total_objects); 4933 nr_free += count_partial(n, count_free); 4934 } 4935 4936 nr_inuse = nr_objs - nr_free; 4937 4938 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4939 nr_objs, s->size, oo_objects(s->oo), 4940 (1 << oo_order(s->oo))); 4941 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4942 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4943 0UL); 4944 seq_putc(m, '\n'); 4945 return 0; 4946} 4947 4948static const struct seq_operations slabinfo_op = { 4949 .start = s_start, 4950 .next = s_next, 4951 .stop = s_stop, 4952 .show = s_show, 4953}; 4954 4955static int slabinfo_open(struct inode *inode, struct file *file) 4956{ 4957 return seq_open(file, &slabinfo_op); 4958} 4959 4960static const struct file_operations proc_slabinfo_operations = { 4961 .open = slabinfo_open, 4962 .read = seq_read, 4963 .llseek = seq_lseek, 4964 .release = seq_release, 4965}; 4966 4967static int __init slab_proc_init(void) 4968{ 4969 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations); 4970 return 0; 4971} 4972module_init(slab_proc_init); 4973#endif /* CONFIG_SLABINFO */ 4974