slub.c revision 49bd5221ce8fb55d12c04a3ffd375201c5bbfb7a
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 <clameter@sgi.com> 9 */ 10 11#include <linux/mm.h> 12#include <linux/module.h> 13#include <linux/bit_spinlock.h> 14#include <linux/interrupt.h> 15#include <linux/bitops.h> 16#include <linux/slab.h> 17#include <linux/seq_file.h> 18#include <linux/cpu.h> 19#include <linux/cpuset.h> 20#include <linux/mempolicy.h> 21#include <linux/ctype.h> 22#include <linux/kallsyms.h> 23#include <linux/memory.h> 24 25/* 26 * Lock order: 27 * 1. slab_lock(page) 28 * 2. slab->list_lock 29 * 30 * The slab_lock protects operations on the object of a particular 31 * slab and its metadata in the page struct. If the slab lock 32 * has been taken then no allocations nor frees can be performed 33 * on the objects in the slab nor can the slab be added or removed 34 * from the partial or full lists since this would mean modifying 35 * the page_struct of the slab. 36 * 37 * The list_lock protects the partial and full list on each node and 38 * the partial slab counter. If taken then no new slabs may be added or 39 * removed from the lists nor make the number of partial slabs be modified. 40 * (Note that the total number of slabs is an atomic value that may be 41 * modified without taking the list lock). 42 * 43 * The list_lock is a centralized lock and thus we avoid taking it as 44 * much as possible. As long as SLUB does not have to handle partial 45 * slabs, operations can continue without any centralized lock. F.e. 46 * allocating a long series of objects that fill up slabs does not require 47 * the list lock. 48 * 49 * The lock order is sometimes inverted when we are trying to get a slab 50 * off a list. We take the list_lock and then look for a page on the list 51 * to use. While we do that objects in the slabs may be freed. We can 52 * only operate on the slab if we have also taken the slab_lock. So we use 53 * a slab_trylock() on the slab. If trylock was successful then no frees 54 * can occur anymore and we can use the slab for allocations etc. If the 55 * slab_trylock() does not succeed then frees are in progress in the slab and 56 * we must stay away from it for a while since we may cause a bouncing 57 * cacheline if we try to acquire the lock. So go onto the next slab. 58 * If all pages are busy then we may allocate a new slab instead of reusing 59 * a partial slab. A new slab has noone operating on it and thus there is 60 * no danger of cacheline contention. 61 * 62 * Interrupts are disabled during allocation and deallocation in order to 63 * make the slab allocator safe to use in the context of an irq. In addition 64 * interrupts are disabled to ensure that the processor does not change 65 * while handling per_cpu slabs, due to kernel preemption. 66 * 67 * SLUB assigns one slab for allocation to each processor. 68 * Allocations only occur from these slabs called cpu slabs. 69 * 70 * Slabs with free elements are kept on a partial list and during regular 71 * operations no list for full slabs is used. If an object in a full slab is 72 * freed then the slab will show up again on the partial lists. 73 * We track full slabs for debugging purposes though because otherwise we 74 * cannot scan all objects. 75 * 76 * Slabs are freed when they become empty. Teardown and setup is 77 * minimal so we rely on the page allocators per cpu caches for 78 * fast frees and allocs. 79 * 80 * Overloading of page flags that are otherwise used for LRU management. 81 * 82 * PageActive The slab is frozen and exempt from list processing. 83 * This means that the slab is dedicated to a purpose 84 * such as satisfying allocations for a specific 85 * processor. Objects may be freed in the slab while 86 * it is frozen but slab_free will then skip the usual 87 * list operations. It is up to the processor holding 88 * the slab to integrate the slab into the slab lists 89 * when the slab is no longer needed. 90 * 91 * One use of this flag is to mark slabs that are 92 * used for allocations. Then such a slab becomes a cpu 93 * slab. The cpu slab may be equipped with an additional 94 * freelist that allows lockless access to 95 * free objects in addition to the regular freelist 96 * that requires the slab lock. 97 * 98 * PageError Slab requires special handling due to debug 99 * options set. This moves slab handling out of 100 * the fast path and disables lockless freelists. 101 */ 102 103#define FROZEN (1 << PG_active) 104 105#ifdef CONFIG_SLUB_DEBUG 106#define SLABDEBUG (1 << PG_error) 107#else 108#define SLABDEBUG 0 109#endif 110 111static inline int SlabFrozen(struct page *page) 112{ 113 return page->flags & FROZEN; 114} 115 116static inline void SetSlabFrozen(struct page *page) 117{ 118 page->flags |= FROZEN; 119} 120 121static inline void ClearSlabFrozen(struct page *page) 122{ 123 page->flags &= ~FROZEN; 124} 125 126static inline int SlabDebug(struct page *page) 127{ 128 return page->flags & SLABDEBUG; 129} 130 131static inline void SetSlabDebug(struct page *page) 132{ 133 page->flags |= SLABDEBUG; 134} 135 136static inline void ClearSlabDebug(struct page *page) 137{ 138 page->flags &= ~SLABDEBUG; 139} 140 141/* 142 * Issues still to be resolved: 143 * 144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 145 * 146 * - Variable sizing of the per node arrays 147 */ 148 149/* Enable to test recovery from slab corruption on boot */ 150#undef SLUB_RESILIENCY_TEST 151 152#if PAGE_SHIFT <= 12 153 154/* 155 * Small page size. Make sure that we do not fragment memory 156 */ 157#define DEFAULT_MAX_ORDER 1 158#define DEFAULT_MIN_OBJECTS 4 159 160#else 161 162/* 163 * Large page machines are customarily able to handle larger 164 * page orders. 165 */ 166#define DEFAULT_MAX_ORDER 2 167#define DEFAULT_MIN_OBJECTS 8 168 169#endif 170 171/* 172 * Mininum number of partial slabs. These will be left on the partial 173 * lists even if they are empty. kmem_cache_shrink may reclaim them. 174 */ 175#define MIN_PARTIAL 5 176 177/* 178 * Maximum number of desirable partial slabs. 179 * The existence of more partial slabs makes kmem_cache_shrink 180 * sort the partial list by the number of objects in the. 181 */ 182#define MAX_PARTIAL 10 183 184#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 185 SLAB_POISON | SLAB_STORE_USER) 186 187/* 188 * Set of flags that will prevent slab merging 189 */ 190#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 191 SLAB_TRACE | SLAB_DESTROY_BY_RCU) 192 193#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 194 SLAB_CACHE_DMA) 195 196#ifndef ARCH_KMALLOC_MINALIGN 197#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 198#endif 199 200#ifndef ARCH_SLAB_MINALIGN 201#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) 202#endif 203 204/* Internal SLUB flags */ 205#define __OBJECT_POISON 0x80000000 /* Poison object */ 206#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ 207#define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */ 208#define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */ 209 210/* Not all arches define cache_line_size */ 211#ifndef cache_line_size 212#define cache_line_size() L1_CACHE_BYTES 213#endif 214 215static int kmem_size = sizeof(struct kmem_cache); 216 217#ifdef CONFIG_SMP 218static struct notifier_block slab_notifier; 219#endif 220 221static enum { 222 DOWN, /* No slab functionality available */ 223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 224 UP, /* Everything works but does not show up in sysfs */ 225 SYSFS /* Sysfs up */ 226} slab_state = DOWN; 227 228/* A list of all slab caches on the system */ 229static DECLARE_RWSEM(slub_lock); 230static LIST_HEAD(slab_caches); 231 232/* 233 * Tracking user of a slab. 234 */ 235struct track { 236 void *addr; /* Called from address */ 237 int cpu; /* Was running on cpu */ 238 int pid; /* Pid context */ 239 unsigned long when; /* When did the operation occur */ 240}; 241 242enum track_item { TRACK_ALLOC, TRACK_FREE }; 243 244#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) 245static int sysfs_slab_add(struct kmem_cache *); 246static int sysfs_slab_alias(struct kmem_cache *, const char *); 247static void sysfs_slab_remove(struct kmem_cache *); 248 249#else 250static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 251static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 252 { return 0; } 253static inline void sysfs_slab_remove(struct kmem_cache *s) 254{ 255 kfree(s); 256} 257 258#endif 259 260static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) 261{ 262#ifdef CONFIG_SLUB_STATS 263 c->stat[si]++; 264#endif 265} 266 267/******************************************************************** 268 * Core slab cache functions 269 *******************************************************************/ 270 271int slab_is_available(void) 272{ 273 return slab_state >= UP; 274} 275 276static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 277{ 278#ifdef CONFIG_NUMA 279 return s->node[node]; 280#else 281 return &s->local_node; 282#endif 283} 284 285static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) 286{ 287#ifdef CONFIG_SMP 288 return s->cpu_slab[cpu]; 289#else 290 return &s->cpu_slab; 291#endif 292} 293 294/* Verify that a pointer has an address that is valid within a slab page */ 295static inline int check_valid_pointer(struct kmem_cache *s, 296 struct page *page, const void *object) 297{ 298 void *base; 299 300 if (!object) 301 return 1; 302 303 base = page_address(page); 304 if (object < base || object >= base + s->objects * s->size || 305 (object - base) % s->size) { 306 return 0; 307 } 308 309 return 1; 310} 311 312/* 313 * Slow version of get and set free pointer. 314 * 315 * This version requires touching the cache lines of kmem_cache which 316 * we avoid to do in the fast alloc free paths. There we obtain the offset 317 * from the page struct. 318 */ 319static inline void *get_freepointer(struct kmem_cache *s, void *object) 320{ 321 return *(void **)(object + s->offset); 322} 323 324static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 325{ 326 *(void **)(object + s->offset) = fp; 327} 328 329/* Loop over all objects in a slab */ 330#define for_each_object(__p, __s, __addr) \ 331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\ 332 __p += (__s)->size) 333 334/* Scan freelist */ 335#define for_each_free_object(__p, __s, __free) \ 336 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 337 338/* Determine object index from a given position */ 339static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 340{ 341 return (p - addr) / s->size; 342} 343 344#ifdef CONFIG_SLUB_DEBUG 345/* 346 * Debug settings: 347 */ 348#ifdef CONFIG_SLUB_DEBUG_ON 349static int slub_debug = DEBUG_DEFAULT_FLAGS; 350#else 351static int slub_debug; 352#endif 353 354static char *slub_debug_slabs; 355 356/* 357 * Object debugging 358 */ 359static void print_section(char *text, u8 *addr, unsigned int length) 360{ 361 int i, offset; 362 int newline = 1; 363 char ascii[17]; 364 365 ascii[16] = 0; 366 367 for (i = 0; i < length; i++) { 368 if (newline) { 369 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 370 newline = 0; 371 } 372 printk(KERN_CONT " %02x", addr[i]); 373 offset = i % 16; 374 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 375 if (offset == 15) { 376 printk(KERN_CONT " %s\n", ascii); 377 newline = 1; 378 } 379 } 380 if (!newline) { 381 i %= 16; 382 while (i < 16) { 383 printk(KERN_CONT " "); 384 ascii[i] = ' '; 385 i++; 386 } 387 printk(KERN_CONT " %s\n", ascii); 388 } 389} 390 391static struct track *get_track(struct kmem_cache *s, void *object, 392 enum track_item alloc) 393{ 394 struct track *p; 395 396 if (s->offset) 397 p = object + s->offset + sizeof(void *); 398 else 399 p = object + s->inuse; 400 401 return p + alloc; 402} 403 404static void set_track(struct kmem_cache *s, void *object, 405 enum track_item alloc, void *addr) 406{ 407 struct track *p; 408 409 if (s->offset) 410 p = object + s->offset + sizeof(void *); 411 else 412 p = object + s->inuse; 413 414 p += alloc; 415 if (addr) { 416 p->addr = addr; 417 p->cpu = smp_processor_id(); 418 p->pid = current ? current->pid : -1; 419 p->when = jiffies; 420 } else 421 memset(p, 0, sizeof(struct track)); 422} 423 424static void init_tracking(struct kmem_cache *s, void *object) 425{ 426 if (!(s->flags & SLAB_STORE_USER)) 427 return; 428 429 set_track(s, object, TRACK_FREE, NULL); 430 set_track(s, object, TRACK_ALLOC, NULL); 431} 432 433static void print_track(const char *s, struct track *t) 434{ 435 if (!t->addr) 436 return; 437 438 printk(KERN_ERR "INFO: %s in ", s); 439 __print_symbol("%s", (unsigned long)t->addr); 440 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); 441} 442 443static void print_tracking(struct kmem_cache *s, void *object) 444{ 445 if (!(s->flags & SLAB_STORE_USER)) 446 return; 447 448 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 449 print_track("Freed", get_track(s, object, TRACK_FREE)); 450} 451 452static void print_page_info(struct page *page) 453{ 454 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n", 455 page, page->inuse, page->freelist, page->flags); 456 457} 458 459static void slab_bug(struct kmem_cache *s, char *fmt, ...) 460{ 461 va_list args; 462 char buf[100]; 463 464 va_start(args, fmt); 465 vsnprintf(buf, sizeof(buf), fmt, args); 466 va_end(args); 467 printk(KERN_ERR "========================================" 468 "=====================================\n"); 469 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 470 printk(KERN_ERR "----------------------------------------" 471 "-------------------------------------\n\n"); 472} 473 474static void slab_fix(struct kmem_cache *s, char *fmt, ...) 475{ 476 va_list args; 477 char buf[100]; 478 479 va_start(args, fmt); 480 vsnprintf(buf, sizeof(buf), fmt, args); 481 va_end(args); 482 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 483} 484 485static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 486{ 487 unsigned int off; /* Offset of last byte */ 488 u8 *addr = page_address(page); 489 490 print_tracking(s, p); 491 492 print_page_info(page); 493 494 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 495 p, p - addr, get_freepointer(s, p)); 496 497 if (p > addr + 16) 498 print_section("Bytes b4", p - 16, 16); 499 500 print_section("Object", p, min(s->objsize, 128)); 501 502 if (s->flags & SLAB_RED_ZONE) 503 print_section("Redzone", p + s->objsize, 504 s->inuse - s->objsize); 505 506 if (s->offset) 507 off = s->offset + sizeof(void *); 508 else 509 off = s->inuse; 510 511 if (s->flags & SLAB_STORE_USER) 512 off += 2 * sizeof(struct track); 513 514 if (off != s->size) 515 /* Beginning of the filler is the free pointer */ 516 print_section("Padding", p + off, s->size - off); 517 518 dump_stack(); 519} 520 521static void object_err(struct kmem_cache *s, struct page *page, 522 u8 *object, char *reason) 523{ 524 slab_bug(s, reason); 525 print_trailer(s, page, object); 526} 527 528static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 529{ 530 va_list args; 531 char buf[100]; 532 533 va_start(args, fmt); 534 vsnprintf(buf, sizeof(buf), fmt, args); 535 va_end(args); 536 slab_bug(s, fmt); 537 print_page_info(page); 538 dump_stack(); 539} 540 541static void init_object(struct kmem_cache *s, void *object, int active) 542{ 543 u8 *p = object; 544 545 if (s->flags & __OBJECT_POISON) { 546 memset(p, POISON_FREE, s->objsize - 1); 547 p[s->objsize - 1] = POISON_END; 548 } 549 550 if (s->flags & SLAB_RED_ZONE) 551 memset(p + s->objsize, 552 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 553 s->inuse - s->objsize); 554} 555 556static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 557{ 558 while (bytes) { 559 if (*start != (u8)value) 560 return start; 561 start++; 562 bytes--; 563 } 564 return NULL; 565} 566 567static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 568 void *from, void *to) 569{ 570 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 571 memset(from, data, to - from); 572} 573 574static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 575 u8 *object, char *what, 576 u8 *start, unsigned int value, unsigned int bytes) 577{ 578 u8 *fault; 579 u8 *end; 580 581 fault = check_bytes(start, value, bytes); 582 if (!fault) 583 return 1; 584 585 end = start + bytes; 586 while (end > fault && end[-1] == value) 587 end--; 588 589 slab_bug(s, "%s overwritten", what); 590 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 591 fault, end - 1, fault[0], value); 592 print_trailer(s, page, object); 593 594 restore_bytes(s, what, value, fault, end); 595 return 0; 596} 597 598/* 599 * Object layout: 600 * 601 * object address 602 * Bytes of the object to be managed. 603 * If the freepointer may overlay the object then the free 604 * pointer is the first word of the object. 605 * 606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 607 * 0xa5 (POISON_END) 608 * 609 * object + s->objsize 610 * Padding to reach word boundary. This is also used for Redzoning. 611 * Padding is extended by another word if Redzoning is enabled and 612 * objsize == inuse. 613 * 614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 615 * 0xcc (RED_ACTIVE) for objects in use. 616 * 617 * object + s->inuse 618 * Meta data starts here. 619 * 620 * A. Free pointer (if we cannot overwrite object on free) 621 * B. Tracking data for SLAB_STORE_USER 622 * C. Padding to reach required alignment boundary or at mininum 623 * one word if debugging is on to be able to detect writes 624 * before the word boundary. 625 * 626 * Padding is done using 0x5a (POISON_INUSE) 627 * 628 * object + s->size 629 * Nothing is used beyond s->size. 630 * 631 * If slabcaches are merged then the objsize and inuse boundaries are mostly 632 * ignored. And therefore no slab options that rely on these boundaries 633 * may be used with merged slabcaches. 634 */ 635 636static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 637{ 638 unsigned long off = s->inuse; /* The end of info */ 639 640 if (s->offset) 641 /* Freepointer is placed after the object. */ 642 off += sizeof(void *); 643 644 if (s->flags & SLAB_STORE_USER) 645 /* We also have user information there */ 646 off += 2 * sizeof(struct track); 647 648 if (s->size == off) 649 return 1; 650 651 return check_bytes_and_report(s, page, p, "Object padding", 652 p + off, POISON_INUSE, s->size - off); 653} 654 655static int slab_pad_check(struct kmem_cache *s, struct page *page) 656{ 657 u8 *start; 658 u8 *fault; 659 u8 *end; 660 int length; 661 int remainder; 662 663 if (!(s->flags & SLAB_POISON)) 664 return 1; 665 666 start = page_address(page); 667 end = start + (PAGE_SIZE << s->order); 668 length = s->objects * s->size; 669 remainder = end - (start + length); 670 if (!remainder) 671 return 1; 672 673 fault = check_bytes(start + length, POISON_INUSE, remainder); 674 if (!fault) 675 return 1; 676 while (end > fault && end[-1] == POISON_INUSE) 677 end--; 678 679 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 680 print_section("Padding", start, length); 681 682 restore_bytes(s, "slab padding", POISON_INUSE, start, end); 683 return 0; 684} 685 686static int check_object(struct kmem_cache *s, struct page *page, 687 void *object, int active) 688{ 689 u8 *p = object; 690 u8 *endobject = object + s->objsize; 691 692 if (s->flags & SLAB_RED_ZONE) { 693 unsigned int red = 694 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 695 696 if (!check_bytes_and_report(s, page, object, "Redzone", 697 endobject, red, s->inuse - s->objsize)) 698 return 0; 699 } else { 700 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 701 check_bytes_and_report(s, page, p, "Alignment padding", 702 endobject, POISON_INUSE, s->inuse - s->objsize); 703 } 704 } 705 706 if (s->flags & SLAB_POISON) { 707 if (!active && (s->flags & __OBJECT_POISON) && 708 (!check_bytes_and_report(s, page, p, "Poison", p, 709 POISON_FREE, s->objsize - 1) || 710 !check_bytes_and_report(s, page, p, "Poison", 711 p + s->objsize - 1, POISON_END, 1))) 712 return 0; 713 /* 714 * check_pad_bytes cleans up on its own. 715 */ 716 check_pad_bytes(s, page, p); 717 } 718 719 if (!s->offset && active) 720 /* 721 * Object and freepointer overlap. Cannot check 722 * freepointer while object is allocated. 723 */ 724 return 1; 725 726 /* Check free pointer validity */ 727 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 728 object_err(s, page, p, "Freepointer corrupt"); 729 /* 730 * No choice but to zap it and thus loose the remainder 731 * of the free objects in this slab. May cause 732 * another error because the object count is now wrong. 733 */ 734 set_freepointer(s, p, NULL); 735 return 0; 736 } 737 return 1; 738} 739 740static int check_slab(struct kmem_cache *s, struct page *page) 741{ 742 VM_BUG_ON(!irqs_disabled()); 743 744 if (!PageSlab(page)) { 745 slab_err(s, page, "Not a valid slab page"); 746 return 0; 747 } 748 if (page->inuse > s->objects) { 749 slab_err(s, page, "inuse %u > max %u", 750 s->name, page->inuse, s->objects); 751 return 0; 752 } 753 /* Slab_pad_check fixes things up after itself */ 754 slab_pad_check(s, page); 755 return 1; 756} 757 758/* 759 * Determine if a certain object on a page is on the freelist. Must hold the 760 * slab lock to guarantee that the chains are in a consistent state. 761 */ 762static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 763{ 764 int nr = 0; 765 void *fp = page->freelist; 766 void *object = NULL; 767 768 while (fp && nr <= s->objects) { 769 if (fp == search) 770 return 1; 771 if (!check_valid_pointer(s, page, fp)) { 772 if (object) { 773 object_err(s, page, object, 774 "Freechain corrupt"); 775 set_freepointer(s, object, NULL); 776 break; 777 } else { 778 slab_err(s, page, "Freepointer corrupt"); 779 page->freelist = NULL; 780 page->inuse = s->objects; 781 slab_fix(s, "Freelist cleared"); 782 return 0; 783 } 784 break; 785 } 786 object = fp; 787 fp = get_freepointer(s, object); 788 nr++; 789 } 790 791 if (page->inuse != s->objects - nr) { 792 slab_err(s, page, "Wrong object count. Counter is %d but " 793 "counted were %d", page->inuse, s->objects - nr); 794 page->inuse = s->objects - nr; 795 slab_fix(s, "Object count adjusted."); 796 } 797 return search == NULL; 798} 799 800static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc) 801{ 802 if (s->flags & SLAB_TRACE) { 803 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 804 s->name, 805 alloc ? "alloc" : "free", 806 object, page->inuse, 807 page->freelist); 808 809 if (!alloc) 810 print_section("Object", (void *)object, s->objsize); 811 812 dump_stack(); 813 } 814} 815 816/* 817 * Tracking of fully allocated slabs for debugging purposes. 818 */ 819static void add_full(struct kmem_cache_node *n, struct page *page) 820{ 821 spin_lock(&n->list_lock); 822 list_add(&page->lru, &n->full); 823 spin_unlock(&n->list_lock); 824} 825 826static void remove_full(struct kmem_cache *s, struct page *page) 827{ 828 struct kmem_cache_node *n; 829 830 if (!(s->flags & SLAB_STORE_USER)) 831 return; 832 833 n = get_node(s, page_to_nid(page)); 834 835 spin_lock(&n->list_lock); 836 list_del(&page->lru); 837 spin_unlock(&n->list_lock); 838} 839 840static void setup_object_debug(struct kmem_cache *s, struct page *page, 841 void *object) 842{ 843 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 844 return; 845 846 init_object(s, object, 0); 847 init_tracking(s, object); 848} 849 850static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 851 void *object, void *addr) 852{ 853 if (!check_slab(s, page)) 854 goto bad; 855 856 if (!on_freelist(s, page, object)) { 857 object_err(s, page, object, "Object already allocated"); 858 goto bad; 859 } 860 861 if (!check_valid_pointer(s, page, object)) { 862 object_err(s, page, object, "Freelist Pointer check fails"); 863 goto bad; 864 } 865 866 if (!check_object(s, page, object, 0)) 867 goto bad; 868 869 /* Success perform special debug activities for allocs */ 870 if (s->flags & SLAB_STORE_USER) 871 set_track(s, object, TRACK_ALLOC, addr); 872 trace(s, page, object, 1); 873 init_object(s, object, 1); 874 return 1; 875 876bad: 877 if (PageSlab(page)) { 878 /* 879 * If this is a slab page then lets do the best we can 880 * to avoid issues in the future. Marking all objects 881 * as used avoids touching the remaining objects. 882 */ 883 slab_fix(s, "Marking all objects used"); 884 page->inuse = s->objects; 885 page->freelist = NULL; 886 } 887 return 0; 888} 889 890static int free_debug_processing(struct kmem_cache *s, struct page *page, 891 void *object, void *addr) 892{ 893 if (!check_slab(s, page)) 894 goto fail; 895 896 if (!check_valid_pointer(s, page, object)) { 897 slab_err(s, page, "Invalid object pointer 0x%p", object); 898 goto fail; 899 } 900 901 if (on_freelist(s, page, object)) { 902 object_err(s, page, object, "Object already free"); 903 goto fail; 904 } 905 906 if (!check_object(s, page, object, 1)) 907 return 0; 908 909 if (unlikely(s != page->slab)) { 910 if (!PageSlab(page)) { 911 slab_err(s, page, "Attempt to free object(0x%p) " 912 "outside of slab", object); 913 } else if (!page->slab) { 914 printk(KERN_ERR 915 "SLUB <none>: no slab for object 0x%p.\n", 916 object); 917 dump_stack(); 918 } else 919 object_err(s, page, object, 920 "page slab pointer corrupt."); 921 goto fail; 922 } 923 924 /* Special debug activities for freeing objects */ 925 if (!SlabFrozen(page) && !page->freelist) 926 remove_full(s, page); 927 if (s->flags & SLAB_STORE_USER) 928 set_track(s, object, TRACK_FREE, addr); 929 trace(s, page, object, 0); 930 init_object(s, object, 0); 931 return 1; 932 933fail: 934 slab_fix(s, "Object at 0x%p not freed", object); 935 return 0; 936} 937 938static int __init setup_slub_debug(char *str) 939{ 940 slub_debug = DEBUG_DEFAULT_FLAGS; 941 if (*str++ != '=' || !*str) 942 /* 943 * No options specified. Switch on full debugging. 944 */ 945 goto out; 946 947 if (*str == ',') 948 /* 949 * No options but restriction on slabs. This means full 950 * debugging for slabs matching a pattern. 951 */ 952 goto check_slabs; 953 954 slub_debug = 0; 955 if (*str == '-') 956 /* 957 * Switch off all debugging measures. 958 */ 959 goto out; 960 961 /* 962 * Determine which debug features should be switched on 963 */ 964 for (; *str && *str != ','; str++) { 965 switch (tolower(*str)) { 966 case 'f': 967 slub_debug |= SLAB_DEBUG_FREE; 968 break; 969 case 'z': 970 slub_debug |= SLAB_RED_ZONE; 971 break; 972 case 'p': 973 slub_debug |= SLAB_POISON; 974 break; 975 case 'u': 976 slub_debug |= SLAB_STORE_USER; 977 break; 978 case 't': 979 slub_debug |= SLAB_TRACE; 980 break; 981 default: 982 printk(KERN_ERR "slub_debug option '%c' " 983 "unknown. skipped\n", *str); 984 } 985 } 986 987check_slabs: 988 if (*str == ',') 989 slub_debug_slabs = str + 1; 990out: 991 return 1; 992} 993 994__setup("slub_debug", setup_slub_debug); 995 996static unsigned long kmem_cache_flags(unsigned long objsize, 997 unsigned long flags, const char *name, 998 void (*ctor)(struct kmem_cache *, void *)) 999{ 1000 /* 1001 * Enable debugging if selected on the kernel commandline. 1002 */ 1003 if (slub_debug && (!slub_debug_slabs || 1004 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0)) 1005 flags |= slub_debug; 1006 1007 return flags; 1008} 1009#else 1010static inline void setup_object_debug(struct kmem_cache *s, 1011 struct page *page, void *object) {} 1012 1013static inline int alloc_debug_processing(struct kmem_cache *s, 1014 struct page *page, void *object, void *addr) { return 0; } 1015 1016static inline int free_debug_processing(struct kmem_cache *s, 1017 struct page *page, void *object, void *addr) { return 0; } 1018 1019static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1020 { return 1; } 1021static inline int check_object(struct kmem_cache *s, struct page *page, 1022 void *object, int active) { return 1; } 1023static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1024static inline unsigned long kmem_cache_flags(unsigned long objsize, 1025 unsigned long flags, const char *name, 1026 void (*ctor)(struct kmem_cache *, void *)) 1027{ 1028 return flags; 1029} 1030#define slub_debug 0 1031#endif 1032/* 1033 * Slab allocation and freeing 1034 */ 1035static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1036{ 1037 struct page *page; 1038 int pages = 1 << s->order; 1039 1040 flags |= s->allocflags; 1041 1042 if (node == -1) 1043 page = alloc_pages(flags, s->order); 1044 else 1045 page = alloc_pages_node(node, flags, s->order); 1046 1047 if (!page) 1048 return NULL; 1049 1050 mod_zone_page_state(page_zone(page), 1051 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1052 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1053 pages); 1054 1055 return page; 1056} 1057 1058static void setup_object(struct kmem_cache *s, struct page *page, 1059 void *object) 1060{ 1061 setup_object_debug(s, page, object); 1062 if (unlikely(s->ctor)) 1063 s->ctor(s, object); 1064} 1065 1066static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1067{ 1068 struct page *page; 1069 struct kmem_cache_node *n; 1070 void *start; 1071 void *last; 1072 void *p; 1073 1074 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1075 1076 page = allocate_slab(s, 1077 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1078 if (!page) 1079 goto out; 1080 1081 n = get_node(s, page_to_nid(page)); 1082 if (n) 1083 atomic_long_inc(&n->nr_slabs); 1084 page->slab = s; 1085 page->flags |= 1 << PG_slab; 1086 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | 1087 SLAB_STORE_USER | SLAB_TRACE)) 1088 SetSlabDebug(page); 1089 1090 start = page_address(page); 1091 1092 if (unlikely(s->flags & SLAB_POISON)) 1093 memset(start, POISON_INUSE, PAGE_SIZE << s->order); 1094 1095 last = start; 1096 for_each_object(p, s, start) { 1097 setup_object(s, page, last); 1098 set_freepointer(s, last, p); 1099 last = p; 1100 } 1101 setup_object(s, page, last); 1102 set_freepointer(s, last, NULL); 1103 1104 page->freelist = start; 1105 page->inuse = 0; 1106out: 1107 return page; 1108} 1109 1110static void __free_slab(struct kmem_cache *s, struct page *page) 1111{ 1112 int pages = 1 << s->order; 1113 1114 if (unlikely(SlabDebug(page))) { 1115 void *p; 1116 1117 slab_pad_check(s, page); 1118 for_each_object(p, s, page_address(page)) 1119 check_object(s, page, p, 0); 1120 ClearSlabDebug(page); 1121 } 1122 1123 mod_zone_page_state(page_zone(page), 1124 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1125 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1126 -pages); 1127 1128 __ClearPageSlab(page); 1129 reset_page_mapcount(page); 1130 __free_pages(page, s->order); 1131} 1132 1133static void rcu_free_slab(struct rcu_head *h) 1134{ 1135 struct page *page; 1136 1137 page = container_of((struct list_head *)h, struct page, lru); 1138 __free_slab(page->slab, page); 1139} 1140 1141static void free_slab(struct kmem_cache *s, struct page *page) 1142{ 1143 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1144 /* 1145 * RCU free overloads the RCU head over the LRU 1146 */ 1147 struct rcu_head *head = (void *)&page->lru; 1148 1149 call_rcu(head, rcu_free_slab); 1150 } else 1151 __free_slab(s, page); 1152} 1153 1154static void discard_slab(struct kmem_cache *s, struct page *page) 1155{ 1156 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1157 1158 atomic_long_dec(&n->nr_slabs); 1159 free_slab(s, page); 1160} 1161 1162/* 1163 * Per slab locking using the pagelock 1164 */ 1165static __always_inline void slab_lock(struct page *page) 1166{ 1167 bit_spin_lock(PG_locked, &page->flags); 1168} 1169 1170static __always_inline void slab_unlock(struct page *page) 1171{ 1172 __bit_spin_unlock(PG_locked, &page->flags); 1173} 1174 1175static __always_inline int slab_trylock(struct page *page) 1176{ 1177 int rc = 1; 1178 1179 rc = bit_spin_trylock(PG_locked, &page->flags); 1180 return rc; 1181} 1182 1183/* 1184 * Management of partially allocated slabs 1185 */ 1186static void add_partial(struct kmem_cache_node *n, 1187 struct page *page, int tail) 1188{ 1189 spin_lock(&n->list_lock); 1190 n->nr_partial++; 1191 if (tail) 1192 list_add_tail(&page->lru, &n->partial); 1193 else 1194 list_add(&page->lru, &n->partial); 1195 spin_unlock(&n->list_lock); 1196} 1197 1198static void remove_partial(struct kmem_cache *s, 1199 struct page *page) 1200{ 1201 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1202 1203 spin_lock(&n->list_lock); 1204 list_del(&page->lru); 1205 n->nr_partial--; 1206 spin_unlock(&n->list_lock); 1207} 1208 1209/* 1210 * Lock slab and remove from the partial list. 1211 * 1212 * Must hold list_lock. 1213 */ 1214static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page) 1215{ 1216 if (slab_trylock(page)) { 1217 list_del(&page->lru); 1218 n->nr_partial--; 1219 SetSlabFrozen(page); 1220 return 1; 1221 } 1222 return 0; 1223} 1224 1225/* 1226 * Try to allocate a partial slab from a specific node. 1227 */ 1228static struct page *get_partial_node(struct kmem_cache_node *n) 1229{ 1230 struct page *page; 1231 1232 /* 1233 * Racy check. If we mistakenly see no partial slabs then we 1234 * just allocate an empty slab. If we mistakenly try to get a 1235 * partial slab and there is none available then get_partials() 1236 * will return NULL. 1237 */ 1238 if (!n || !n->nr_partial) 1239 return NULL; 1240 1241 spin_lock(&n->list_lock); 1242 list_for_each_entry(page, &n->partial, lru) 1243 if (lock_and_freeze_slab(n, page)) 1244 goto out; 1245 page = NULL; 1246out: 1247 spin_unlock(&n->list_lock); 1248 return page; 1249} 1250 1251/* 1252 * Get a page from somewhere. Search in increasing NUMA distances. 1253 */ 1254static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1255{ 1256#ifdef CONFIG_NUMA 1257 struct zonelist *zonelist; 1258 struct zone **z; 1259 struct page *page; 1260 1261 /* 1262 * The defrag ratio allows a configuration of the tradeoffs between 1263 * inter node defragmentation and node local allocations. A lower 1264 * defrag_ratio increases the tendency to do local allocations 1265 * instead of attempting to obtain partial slabs from other nodes. 1266 * 1267 * If the defrag_ratio is set to 0 then kmalloc() always 1268 * returns node local objects. If the ratio is higher then kmalloc() 1269 * may return off node objects because partial slabs are obtained 1270 * from other nodes and filled up. 1271 * 1272 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1273 * defrag_ratio = 1000) then every (well almost) allocation will 1274 * first attempt to defrag slab caches on other nodes. This means 1275 * scanning over all nodes to look for partial slabs which may be 1276 * expensive if we do it every time we are trying to find a slab 1277 * with available objects. 1278 */ 1279 if (!s->remote_node_defrag_ratio || 1280 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1281 return NULL; 1282 1283 zonelist = &NODE_DATA( 1284 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)]; 1285 for (z = zonelist->zones; *z; z++) { 1286 struct kmem_cache_node *n; 1287 1288 n = get_node(s, zone_to_nid(*z)); 1289 1290 if (n && cpuset_zone_allowed_hardwall(*z, flags) && 1291 n->nr_partial > MIN_PARTIAL) { 1292 page = get_partial_node(n); 1293 if (page) 1294 return page; 1295 } 1296 } 1297#endif 1298 return NULL; 1299} 1300 1301/* 1302 * Get a partial page, lock it and return it. 1303 */ 1304static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1305{ 1306 struct page *page; 1307 int searchnode = (node == -1) ? numa_node_id() : node; 1308 1309 page = get_partial_node(get_node(s, searchnode)); 1310 if (page || (flags & __GFP_THISNODE)) 1311 return page; 1312 1313 return get_any_partial(s, flags); 1314} 1315 1316/* 1317 * Move a page back to the lists. 1318 * 1319 * Must be called with the slab lock held. 1320 * 1321 * On exit the slab lock will have been dropped. 1322 */ 1323static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1324{ 1325 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1326 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id()); 1327 1328 ClearSlabFrozen(page); 1329 if (page->inuse) { 1330 1331 if (page->freelist) { 1332 add_partial(n, page, tail); 1333 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1334 } else { 1335 stat(c, DEACTIVATE_FULL); 1336 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER)) 1337 add_full(n, page); 1338 } 1339 slab_unlock(page); 1340 } else { 1341 stat(c, DEACTIVATE_EMPTY); 1342 if (n->nr_partial < MIN_PARTIAL) { 1343 /* 1344 * Adding an empty slab to the partial slabs in order 1345 * to avoid page allocator overhead. This slab needs 1346 * to come after the other slabs with objects in 1347 * so that the others get filled first. That way the 1348 * size of the partial list stays small. 1349 * 1350 * kmem_cache_shrink can reclaim any empty slabs from the 1351 * partial list. 1352 */ 1353 add_partial(n, page, 1); 1354 slab_unlock(page); 1355 } else { 1356 slab_unlock(page); 1357 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB); 1358 discard_slab(s, page); 1359 } 1360 } 1361} 1362 1363/* 1364 * Remove the cpu slab 1365 */ 1366static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1367{ 1368 struct page *page = c->page; 1369 int tail = 1; 1370 1371 if (page->freelist) 1372 stat(c, DEACTIVATE_REMOTE_FREES); 1373 /* 1374 * Merge cpu freelist into slab freelist. Typically we get here 1375 * because both freelists are empty. So this is unlikely 1376 * to occur. 1377 */ 1378 while (unlikely(c->freelist)) { 1379 void **object; 1380 1381 tail = 0; /* Hot objects. Put the slab first */ 1382 1383 /* Retrieve object from cpu_freelist */ 1384 object = c->freelist; 1385 c->freelist = c->freelist[c->offset]; 1386 1387 /* And put onto the regular freelist */ 1388 object[c->offset] = page->freelist; 1389 page->freelist = object; 1390 page->inuse--; 1391 } 1392 c->page = NULL; 1393 unfreeze_slab(s, page, tail); 1394} 1395 1396static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1397{ 1398 stat(c, CPUSLAB_FLUSH); 1399 slab_lock(c->page); 1400 deactivate_slab(s, c); 1401} 1402 1403/* 1404 * Flush cpu slab. 1405 * 1406 * Called from IPI handler with interrupts disabled. 1407 */ 1408static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1409{ 1410 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1411 1412 if (likely(c && c->page)) 1413 flush_slab(s, c); 1414} 1415 1416static void flush_cpu_slab(void *d) 1417{ 1418 struct kmem_cache *s = d; 1419 1420 __flush_cpu_slab(s, smp_processor_id()); 1421} 1422 1423static void flush_all(struct kmem_cache *s) 1424{ 1425#ifdef CONFIG_SMP 1426 on_each_cpu(flush_cpu_slab, s, 1, 1); 1427#else 1428 unsigned long flags; 1429 1430 local_irq_save(flags); 1431 flush_cpu_slab(s); 1432 local_irq_restore(flags); 1433#endif 1434} 1435 1436/* 1437 * Check if the objects in a per cpu structure fit numa 1438 * locality expectations. 1439 */ 1440static inline int node_match(struct kmem_cache_cpu *c, int node) 1441{ 1442#ifdef CONFIG_NUMA 1443 if (node != -1 && c->node != node) 1444 return 0; 1445#endif 1446 return 1; 1447} 1448 1449/* 1450 * Slow path. The lockless freelist is empty or we need to perform 1451 * debugging duties. 1452 * 1453 * Interrupts are disabled. 1454 * 1455 * Processing is still very fast if new objects have been freed to the 1456 * regular freelist. In that case we simply take over the regular freelist 1457 * as the lockless freelist and zap the regular freelist. 1458 * 1459 * If that is not working then we fall back to the partial lists. We take the 1460 * first element of the freelist as the object to allocate now and move the 1461 * rest of the freelist to the lockless freelist. 1462 * 1463 * And if we were unable to get a new slab from the partial slab lists then 1464 * we need to allocate a new slab. This is the slowest path since it involves 1465 * a call to the page allocator and the setup of a new slab. 1466 */ 1467static void *__slab_alloc(struct kmem_cache *s, 1468 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c) 1469{ 1470 void **object; 1471 struct page *new; 1472 1473 /* We handle __GFP_ZERO in the caller */ 1474 gfpflags &= ~__GFP_ZERO; 1475 1476 if (!c->page) 1477 goto new_slab; 1478 1479 slab_lock(c->page); 1480 if (unlikely(!node_match(c, node))) 1481 goto another_slab; 1482 1483 stat(c, ALLOC_REFILL); 1484 1485load_freelist: 1486 object = c->page->freelist; 1487 if (unlikely(!object)) 1488 goto another_slab; 1489 if (unlikely(SlabDebug(c->page))) 1490 goto debug; 1491 1492 c->freelist = object[c->offset]; 1493 c->page->inuse = s->objects; 1494 c->page->freelist = NULL; 1495 c->node = page_to_nid(c->page); 1496unlock_out: 1497 slab_unlock(c->page); 1498 stat(c, ALLOC_SLOWPATH); 1499 return object; 1500 1501another_slab: 1502 deactivate_slab(s, c); 1503 1504new_slab: 1505 new = get_partial(s, gfpflags, node); 1506 if (new) { 1507 c->page = new; 1508 stat(c, ALLOC_FROM_PARTIAL); 1509 goto load_freelist; 1510 } 1511 1512 if (gfpflags & __GFP_WAIT) 1513 local_irq_enable(); 1514 1515 new = new_slab(s, gfpflags, node); 1516 1517 if (gfpflags & __GFP_WAIT) 1518 local_irq_disable(); 1519 1520 if (new) { 1521 c = get_cpu_slab(s, smp_processor_id()); 1522 stat(c, ALLOC_SLAB); 1523 if (c->page) 1524 flush_slab(s, c); 1525 slab_lock(new); 1526 SetSlabFrozen(new); 1527 c->page = new; 1528 goto load_freelist; 1529 } 1530 1531 /* 1532 * No memory available. 1533 * 1534 * If the slab uses higher order allocs but the object is 1535 * smaller than a page size then we can fallback in emergencies 1536 * to the page allocator via kmalloc_large. The page allocator may 1537 * have failed to obtain a higher order page and we can try to 1538 * allocate a single page if the object fits into a single page. 1539 * That is only possible if certain conditions are met that are being 1540 * checked when a slab is created. 1541 */ 1542 if (!(gfpflags & __GFP_NORETRY) && 1543 (s->flags & __PAGE_ALLOC_FALLBACK)) { 1544 if (gfpflags & __GFP_WAIT) 1545 local_irq_enable(); 1546 object = kmalloc_large(s->objsize, gfpflags); 1547 if (gfpflags & __GFP_WAIT) 1548 local_irq_disable(); 1549 return object; 1550 } 1551 return NULL; 1552debug: 1553 if (!alloc_debug_processing(s, c->page, object, addr)) 1554 goto another_slab; 1555 1556 c->page->inuse++; 1557 c->page->freelist = object[c->offset]; 1558 c->node = -1; 1559 goto unlock_out; 1560} 1561 1562/* 1563 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1564 * have the fastpath folded into their functions. So no function call 1565 * overhead for requests that can be satisfied on the fastpath. 1566 * 1567 * The fastpath works by first checking if the lockless freelist can be used. 1568 * If not then __slab_alloc is called for slow processing. 1569 * 1570 * Otherwise we can simply pick the next object from the lockless free list. 1571 */ 1572static __always_inline void *slab_alloc(struct kmem_cache *s, 1573 gfp_t gfpflags, int node, void *addr) 1574{ 1575 void **object; 1576 struct kmem_cache_cpu *c; 1577 unsigned long flags; 1578 1579 local_irq_save(flags); 1580 c = get_cpu_slab(s, smp_processor_id()); 1581 if (unlikely(!c->freelist || !node_match(c, node))) 1582 1583 object = __slab_alloc(s, gfpflags, node, addr, c); 1584 1585 else { 1586 object = c->freelist; 1587 c->freelist = object[c->offset]; 1588 stat(c, ALLOC_FASTPATH); 1589 } 1590 local_irq_restore(flags); 1591 1592 if (unlikely((gfpflags & __GFP_ZERO) && object)) 1593 memset(object, 0, c->objsize); 1594 1595 return object; 1596} 1597 1598void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1599{ 1600 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); 1601} 1602EXPORT_SYMBOL(kmem_cache_alloc); 1603 1604#ifdef CONFIG_NUMA 1605void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1606{ 1607 return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); 1608} 1609EXPORT_SYMBOL(kmem_cache_alloc_node); 1610#endif 1611 1612/* 1613 * Slow patch handling. This may still be called frequently since objects 1614 * have a longer lifetime than the cpu slabs in most processing loads. 1615 * 1616 * So we still attempt to reduce cache line usage. Just take the slab 1617 * lock and free the item. If there is no additional partial page 1618 * handling required then we can return immediately. 1619 */ 1620static void __slab_free(struct kmem_cache *s, struct page *page, 1621 void *x, void *addr, unsigned int offset) 1622{ 1623 void *prior; 1624 void **object = (void *)x; 1625 struct kmem_cache_cpu *c; 1626 1627 c = get_cpu_slab(s, raw_smp_processor_id()); 1628 stat(c, FREE_SLOWPATH); 1629 slab_lock(page); 1630 1631 if (unlikely(SlabDebug(page))) 1632 goto debug; 1633 1634checks_ok: 1635 prior = object[offset] = page->freelist; 1636 page->freelist = object; 1637 page->inuse--; 1638 1639 if (unlikely(SlabFrozen(page))) { 1640 stat(c, FREE_FROZEN); 1641 goto out_unlock; 1642 } 1643 1644 if (unlikely(!page->inuse)) 1645 goto slab_empty; 1646 1647 /* 1648 * Objects left in the slab. If it was not on the partial list before 1649 * then add it. 1650 */ 1651 if (unlikely(!prior)) { 1652 add_partial(get_node(s, page_to_nid(page)), page, 1); 1653 stat(c, FREE_ADD_PARTIAL); 1654 } 1655 1656out_unlock: 1657 slab_unlock(page); 1658 return; 1659 1660slab_empty: 1661 if (prior) { 1662 /* 1663 * Slab still on the partial list. 1664 */ 1665 remove_partial(s, page); 1666 stat(c, FREE_REMOVE_PARTIAL); 1667 } 1668 slab_unlock(page); 1669 stat(c, FREE_SLAB); 1670 discard_slab(s, page); 1671 return; 1672 1673debug: 1674 if (!free_debug_processing(s, page, x, addr)) 1675 goto out_unlock; 1676 goto checks_ok; 1677} 1678 1679/* 1680 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1681 * can perform fastpath freeing without additional function calls. 1682 * 1683 * The fastpath is only possible if we are freeing to the current cpu slab 1684 * of this processor. This typically the case if we have just allocated 1685 * the item before. 1686 * 1687 * If fastpath is not possible then fall back to __slab_free where we deal 1688 * with all sorts of special processing. 1689 */ 1690static __always_inline void slab_free(struct kmem_cache *s, 1691 struct page *page, void *x, void *addr) 1692{ 1693 void **object = (void *)x; 1694 struct kmem_cache_cpu *c; 1695 unsigned long flags; 1696 1697 local_irq_save(flags); 1698 c = get_cpu_slab(s, smp_processor_id()); 1699 debug_check_no_locks_freed(object, c->objsize); 1700 if (likely(page == c->page && c->node >= 0)) { 1701 object[c->offset] = c->freelist; 1702 c->freelist = object; 1703 stat(c, FREE_FASTPATH); 1704 } else 1705 __slab_free(s, page, x, addr, c->offset); 1706 1707 local_irq_restore(flags); 1708} 1709 1710void kmem_cache_free(struct kmem_cache *s, void *x) 1711{ 1712 struct page *page; 1713 1714 page = virt_to_head_page(x); 1715 1716 slab_free(s, page, x, __builtin_return_address(0)); 1717} 1718EXPORT_SYMBOL(kmem_cache_free); 1719 1720/* Figure out on which slab object the object resides */ 1721static struct page *get_object_page(const void *x) 1722{ 1723 struct page *page = virt_to_head_page(x); 1724 1725 if (!PageSlab(page)) 1726 return NULL; 1727 1728 return page; 1729} 1730 1731/* 1732 * Object placement in a slab is made very easy because we always start at 1733 * offset 0. If we tune the size of the object to the alignment then we can 1734 * get the required alignment by putting one properly sized object after 1735 * another. 1736 * 1737 * Notice that the allocation order determines the sizes of the per cpu 1738 * caches. Each processor has always one slab available for allocations. 1739 * Increasing the allocation order reduces the number of times that slabs 1740 * must be moved on and off the partial lists and is therefore a factor in 1741 * locking overhead. 1742 */ 1743 1744/* 1745 * Mininum / Maximum order of slab pages. This influences locking overhead 1746 * and slab fragmentation. A higher order reduces the number of partial slabs 1747 * and increases the number of allocations possible without having to 1748 * take the list_lock. 1749 */ 1750static int slub_min_order; 1751static int slub_max_order = DEFAULT_MAX_ORDER; 1752static int slub_min_objects = DEFAULT_MIN_OBJECTS; 1753 1754/* 1755 * Merge control. If this is set then no merging of slab caches will occur. 1756 * (Could be removed. This was introduced to pacify the merge skeptics.) 1757 */ 1758static int slub_nomerge; 1759 1760/* 1761 * Calculate the order of allocation given an slab object size. 1762 * 1763 * The order of allocation has significant impact on performance and other 1764 * system components. Generally order 0 allocations should be preferred since 1765 * order 0 does not cause fragmentation in the page allocator. Larger objects 1766 * be problematic to put into order 0 slabs because there may be too much 1767 * unused space left. We go to a higher order if more than 1/8th of the slab 1768 * would be wasted. 1769 * 1770 * In order to reach satisfactory performance we must ensure that a minimum 1771 * number of objects is in one slab. Otherwise we may generate too much 1772 * activity on the partial lists which requires taking the list_lock. This is 1773 * less a concern for large slabs though which are rarely used. 1774 * 1775 * slub_max_order specifies the order where we begin to stop considering the 1776 * number of objects in a slab as critical. If we reach slub_max_order then 1777 * we try to keep the page order as low as possible. So we accept more waste 1778 * of space in favor of a small page order. 1779 * 1780 * Higher order allocations also allow the placement of more objects in a 1781 * slab and thereby reduce object handling overhead. If the user has 1782 * requested a higher mininum order then we start with that one instead of 1783 * the smallest order which will fit the object. 1784 */ 1785static inline int slab_order(int size, int min_objects, 1786 int max_order, int fract_leftover) 1787{ 1788 int order; 1789 int rem; 1790 int min_order = slub_min_order; 1791 1792 for (order = max(min_order, 1793 fls(min_objects * size - 1) - PAGE_SHIFT); 1794 order <= max_order; order++) { 1795 1796 unsigned long slab_size = PAGE_SIZE << order; 1797 1798 if (slab_size < min_objects * size) 1799 continue; 1800 1801 rem = slab_size % size; 1802 1803 if (rem <= slab_size / fract_leftover) 1804 break; 1805 1806 } 1807 1808 return order; 1809} 1810 1811static inline int calculate_order(int size) 1812{ 1813 int order; 1814 int min_objects; 1815 int fraction; 1816 1817 /* 1818 * Attempt to find best configuration for a slab. This 1819 * works by first attempting to generate a layout with 1820 * the best configuration and backing off gradually. 1821 * 1822 * First we reduce the acceptable waste in a slab. Then 1823 * we reduce the minimum objects required in a slab. 1824 */ 1825 min_objects = slub_min_objects; 1826 while (min_objects > 1) { 1827 fraction = 8; 1828 while (fraction >= 4) { 1829 order = slab_order(size, min_objects, 1830 slub_max_order, fraction); 1831 if (order <= slub_max_order) 1832 return order; 1833 fraction /= 2; 1834 } 1835 min_objects /= 2; 1836 } 1837 1838 /* 1839 * We were unable to place multiple objects in a slab. Now 1840 * lets see if we can place a single object there. 1841 */ 1842 order = slab_order(size, 1, slub_max_order, 1); 1843 if (order <= slub_max_order) 1844 return order; 1845 1846 /* 1847 * Doh this slab cannot be placed using slub_max_order. 1848 */ 1849 order = slab_order(size, 1, MAX_ORDER, 1); 1850 if (order <= MAX_ORDER) 1851 return order; 1852 return -ENOSYS; 1853} 1854 1855/* 1856 * Figure out what the alignment of the objects will be. 1857 */ 1858static unsigned long calculate_alignment(unsigned long flags, 1859 unsigned long align, unsigned long size) 1860{ 1861 /* 1862 * If the user wants hardware cache aligned objects then follow that 1863 * suggestion if the object is sufficiently large. 1864 * 1865 * The hardware cache alignment cannot override the specified 1866 * alignment though. If that is greater then use it. 1867 */ 1868 if (flags & SLAB_HWCACHE_ALIGN) { 1869 unsigned long ralign = cache_line_size(); 1870 while (size <= ralign / 2) 1871 ralign /= 2; 1872 align = max(align, ralign); 1873 } 1874 1875 if (align < ARCH_SLAB_MINALIGN) 1876 align = ARCH_SLAB_MINALIGN; 1877 1878 return ALIGN(align, sizeof(void *)); 1879} 1880 1881static void init_kmem_cache_cpu(struct kmem_cache *s, 1882 struct kmem_cache_cpu *c) 1883{ 1884 c->page = NULL; 1885 c->freelist = NULL; 1886 c->node = 0; 1887 c->offset = s->offset / sizeof(void *); 1888 c->objsize = s->objsize; 1889#ifdef CONFIG_SLUB_STATS 1890 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); 1891#endif 1892} 1893 1894static void init_kmem_cache_node(struct kmem_cache_node *n) 1895{ 1896 n->nr_partial = 0; 1897 atomic_long_set(&n->nr_slabs, 0); 1898 spin_lock_init(&n->list_lock); 1899 INIT_LIST_HEAD(&n->partial); 1900#ifdef CONFIG_SLUB_DEBUG 1901 INIT_LIST_HEAD(&n->full); 1902#endif 1903} 1904 1905#ifdef CONFIG_SMP 1906/* 1907 * Per cpu array for per cpu structures. 1908 * 1909 * The per cpu array places all kmem_cache_cpu structures from one processor 1910 * close together meaning that it becomes possible that multiple per cpu 1911 * structures are contained in one cacheline. This may be particularly 1912 * beneficial for the kmalloc caches. 1913 * 1914 * A desktop system typically has around 60-80 slabs. With 100 here we are 1915 * likely able to get per cpu structures for all caches from the array defined 1916 * here. We must be able to cover all kmalloc caches during bootstrap. 1917 * 1918 * If the per cpu array is exhausted then fall back to kmalloc 1919 * of individual cachelines. No sharing is possible then. 1920 */ 1921#define NR_KMEM_CACHE_CPU 100 1922 1923static DEFINE_PER_CPU(struct kmem_cache_cpu, 1924 kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; 1925 1926static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); 1927static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; 1928 1929static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, 1930 int cpu, gfp_t flags) 1931{ 1932 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); 1933 1934 if (c) 1935 per_cpu(kmem_cache_cpu_free, cpu) = 1936 (void *)c->freelist; 1937 else { 1938 /* Table overflow: So allocate ourselves */ 1939 c = kmalloc_node( 1940 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), 1941 flags, cpu_to_node(cpu)); 1942 if (!c) 1943 return NULL; 1944 } 1945 1946 init_kmem_cache_cpu(s, c); 1947 return c; 1948} 1949 1950static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) 1951{ 1952 if (c < per_cpu(kmem_cache_cpu, cpu) || 1953 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { 1954 kfree(c); 1955 return; 1956 } 1957 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); 1958 per_cpu(kmem_cache_cpu_free, cpu) = c; 1959} 1960 1961static void free_kmem_cache_cpus(struct kmem_cache *s) 1962{ 1963 int cpu; 1964 1965 for_each_online_cpu(cpu) { 1966 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1967 1968 if (c) { 1969 s->cpu_slab[cpu] = NULL; 1970 free_kmem_cache_cpu(c, cpu); 1971 } 1972 } 1973} 1974 1975static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 1976{ 1977 int cpu; 1978 1979 for_each_online_cpu(cpu) { 1980 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1981 1982 if (c) 1983 continue; 1984 1985 c = alloc_kmem_cache_cpu(s, cpu, flags); 1986 if (!c) { 1987 free_kmem_cache_cpus(s); 1988 return 0; 1989 } 1990 s->cpu_slab[cpu] = c; 1991 } 1992 return 1; 1993} 1994 1995/* 1996 * Initialize the per cpu array. 1997 */ 1998static void init_alloc_cpu_cpu(int cpu) 1999{ 2000 int i; 2001 2002 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) 2003 return; 2004 2005 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) 2006 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); 2007 2008 cpu_set(cpu, kmem_cach_cpu_free_init_once); 2009} 2010 2011static void __init init_alloc_cpu(void) 2012{ 2013 int cpu; 2014 2015 for_each_online_cpu(cpu) 2016 init_alloc_cpu_cpu(cpu); 2017 } 2018 2019#else 2020static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} 2021static inline void init_alloc_cpu(void) {} 2022 2023static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2024{ 2025 init_kmem_cache_cpu(s, &s->cpu_slab); 2026 return 1; 2027} 2028#endif 2029 2030#ifdef CONFIG_NUMA 2031/* 2032 * No kmalloc_node yet so do it by hand. We know that this is the first 2033 * slab on the node for this slabcache. There are no concurrent accesses 2034 * possible. 2035 * 2036 * Note that this function only works on the kmalloc_node_cache 2037 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2038 * memory on a fresh node that has no slab structures yet. 2039 */ 2040static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags, 2041 int node) 2042{ 2043 struct page *page; 2044 struct kmem_cache_node *n; 2045 unsigned long flags; 2046 2047 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2048 2049 page = new_slab(kmalloc_caches, gfpflags, node); 2050 2051 BUG_ON(!page); 2052 if (page_to_nid(page) != node) { 2053 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2054 "node %d\n", node); 2055 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2056 "in order to be able to continue\n"); 2057 } 2058 2059 n = page->freelist; 2060 BUG_ON(!n); 2061 page->freelist = get_freepointer(kmalloc_caches, n); 2062 page->inuse++; 2063 kmalloc_caches->node[node] = n; 2064#ifdef CONFIG_SLUB_DEBUG 2065 init_object(kmalloc_caches, n, 1); 2066 init_tracking(kmalloc_caches, n); 2067#endif 2068 init_kmem_cache_node(n); 2069 atomic_long_inc(&n->nr_slabs); 2070 2071 /* 2072 * lockdep requires consistent irq usage for each lock 2073 * so even though there cannot be a race this early in 2074 * the boot sequence, we still disable irqs. 2075 */ 2076 local_irq_save(flags); 2077 add_partial(n, page, 0); 2078 local_irq_restore(flags); 2079 return n; 2080} 2081 2082static void free_kmem_cache_nodes(struct kmem_cache *s) 2083{ 2084 int node; 2085 2086 for_each_node_state(node, N_NORMAL_MEMORY) { 2087 struct kmem_cache_node *n = s->node[node]; 2088 if (n && n != &s->local_node) 2089 kmem_cache_free(kmalloc_caches, n); 2090 s->node[node] = NULL; 2091 } 2092} 2093 2094static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2095{ 2096 int node; 2097 int local_node; 2098 2099 if (slab_state >= UP) 2100 local_node = page_to_nid(virt_to_page(s)); 2101 else 2102 local_node = 0; 2103 2104 for_each_node_state(node, N_NORMAL_MEMORY) { 2105 struct kmem_cache_node *n; 2106 2107 if (local_node == node) 2108 n = &s->local_node; 2109 else { 2110 if (slab_state == DOWN) { 2111 n = early_kmem_cache_node_alloc(gfpflags, 2112 node); 2113 continue; 2114 } 2115 n = kmem_cache_alloc_node(kmalloc_caches, 2116 gfpflags, node); 2117 2118 if (!n) { 2119 free_kmem_cache_nodes(s); 2120 return 0; 2121 } 2122 2123 } 2124 s->node[node] = n; 2125 init_kmem_cache_node(n); 2126 } 2127 return 1; 2128} 2129#else 2130static void free_kmem_cache_nodes(struct kmem_cache *s) 2131{ 2132} 2133 2134static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2135{ 2136 init_kmem_cache_node(&s->local_node); 2137 return 1; 2138} 2139#endif 2140 2141/* 2142 * calculate_sizes() determines the order and the distribution of data within 2143 * a slab object. 2144 */ 2145static int calculate_sizes(struct kmem_cache *s) 2146{ 2147 unsigned long flags = s->flags; 2148 unsigned long size = s->objsize; 2149 unsigned long align = s->align; 2150 2151 /* 2152 * Round up object size to the next word boundary. We can only 2153 * place the free pointer at word boundaries and this determines 2154 * the possible location of the free pointer. 2155 */ 2156 size = ALIGN(size, sizeof(void *)); 2157 2158#ifdef CONFIG_SLUB_DEBUG 2159 /* 2160 * Determine if we can poison the object itself. If the user of 2161 * the slab may touch the object after free or before allocation 2162 * then we should never poison the object itself. 2163 */ 2164 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2165 !s->ctor) 2166 s->flags |= __OBJECT_POISON; 2167 else 2168 s->flags &= ~__OBJECT_POISON; 2169 2170 2171 /* 2172 * If we are Redzoning then check if there is some space between the 2173 * end of the object and the free pointer. If not then add an 2174 * additional word to have some bytes to store Redzone information. 2175 */ 2176 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2177 size += sizeof(void *); 2178#endif 2179 2180 /* 2181 * With that we have determined the number of bytes in actual use 2182 * by the object. This is the potential offset to the free pointer. 2183 */ 2184 s->inuse = size; 2185 2186 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2187 s->ctor)) { 2188 /* 2189 * Relocate free pointer after the object if it is not 2190 * permitted to overwrite the first word of the object on 2191 * kmem_cache_free. 2192 * 2193 * This is the case if we do RCU, have a constructor or 2194 * destructor or are poisoning the objects. 2195 */ 2196 s->offset = size; 2197 size += sizeof(void *); 2198 } 2199 2200#ifdef CONFIG_SLUB_DEBUG 2201 if (flags & SLAB_STORE_USER) 2202 /* 2203 * Need to store information about allocs and frees after 2204 * the object. 2205 */ 2206 size += 2 * sizeof(struct track); 2207 2208 if (flags & SLAB_RED_ZONE) 2209 /* 2210 * Add some empty padding so that we can catch 2211 * overwrites from earlier objects rather than let 2212 * tracking information or the free pointer be 2213 * corrupted if an user writes before the start 2214 * of the object. 2215 */ 2216 size += sizeof(void *); 2217#endif 2218 2219 /* 2220 * Determine the alignment based on various parameters that the 2221 * user specified and the dynamic determination of cache line size 2222 * on bootup. 2223 */ 2224 align = calculate_alignment(flags, align, s->objsize); 2225 2226 /* 2227 * SLUB stores one object immediately after another beginning from 2228 * offset 0. In order to align the objects we have to simply size 2229 * each object to conform to the alignment. 2230 */ 2231 size = ALIGN(size, align); 2232 s->size = size; 2233 2234 if ((flags & __KMALLOC_CACHE) && 2235 PAGE_SIZE / size < slub_min_objects) { 2236 /* 2237 * Kmalloc cache that would not have enough objects in 2238 * an order 0 page. Kmalloc slabs can fallback to 2239 * page allocator order 0 allocs so take a reasonably large 2240 * order that will allows us a good number of objects. 2241 */ 2242 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER); 2243 s->flags |= __PAGE_ALLOC_FALLBACK; 2244 s->allocflags |= __GFP_NOWARN; 2245 } else 2246 s->order = calculate_order(size); 2247 2248 if (s->order < 0) 2249 return 0; 2250 2251 s->allocflags = 0; 2252 if (s->order) 2253 s->allocflags |= __GFP_COMP; 2254 2255 if (s->flags & SLAB_CACHE_DMA) 2256 s->allocflags |= SLUB_DMA; 2257 2258 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2259 s->allocflags |= __GFP_RECLAIMABLE; 2260 2261 /* 2262 * Determine the number of objects per slab 2263 */ 2264 s->objects = (PAGE_SIZE << s->order) / size; 2265 2266 return !!s->objects; 2267 2268} 2269 2270static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2271 const char *name, size_t size, 2272 size_t align, unsigned long flags, 2273 void (*ctor)(struct kmem_cache *, void *)) 2274{ 2275 memset(s, 0, kmem_size); 2276 s->name = name; 2277 s->ctor = ctor; 2278 s->objsize = size; 2279 s->align = align; 2280 s->flags = kmem_cache_flags(size, flags, name, ctor); 2281 2282 if (!calculate_sizes(s)) 2283 goto error; 2284 2285 s->refcount = 1; 2286#ifdef CONFIG_NUMA 2287 s->remote_node_defrag_ratio = 100; 2288#endif 2289 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2290 goto error; 2291 2292 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2293 return 1; 2294 free_kmem_cache_nodes(s); 2295error: 2296 if (flags & SLAB_PANIC) 2297 panic("Cannot create slab %s size=%lu realsize=%u " 2298 "order=%u offset=%u flags=%lx\n", 2299 s->name, (unsigned long)size, s->size, s->order, 2300 s->offset, flags); 2301 return 0; 2302} 2303 2304/* 2305 * Check if a given pointer is valid 2306 */ 2307int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2308{ 2309 struct page *page; 2310 2311 page = get_object_page(object); 2312 2313 if (!page || s != page->slab) 2314 /* No slab or wrong slab */ 2315 return 0; 2316 2317 if (!check_valid_pointer(s, page, object)) 2318 return 0; 2319 2320 /* 2321 * We could also check if the object is on the slabs freelist. 2322 * But this would be too expensive and it seems that the main 2323 * purpose of kmem_ptr_valid() is to check if the object belongs 2324 * to a certain slab. 2325 */ 2326 return 1; 2327} 2328EXPORT_SYMBOL(kmem_ptr_validate); 2329 2330/* 2331 * Determine the size of a slab object 2332 */ 2333unsigned int kmem_cache_size(struct kmem_cache *s) 2334{ 2335 return s->objsize; 2336} 2337EXPORT_SYMBOL(kmem_cache_size); 2338 2339const char *kmem_cache_name(struct kmem_cache *s) 2340{ 2341 return s->name; 2342} 2343EXPORT_SYMBOL(kmem_cache_name); 2344 2345/* 2346 * Attempt to free all slabs on a node. Return the number of slabs we 2347 * were unable to free. 2348 */ 2349static int free_list(struct kmem_cache *s, struct kmem_cache_node *n, 2350 struct list_head *list) 2351{ 2352 int slabs_inuse = 0; 2353 unsigned long flags; 2354 struct page *page, *h; 2355 2356 spin_lock_irqsave(&n->list_lock, flags); 2357 list_for_each_entry_safe(page, h, list, lru) 2358 if (!page->inuse) { 2359 list_del(&page->lru); 2360 discard_slab(s, page); 2361 } else 2362 slabs_inuse++; 2363 spin_unlock_irqrestore(&n->list_lock, flags); 2364 return slabs_inuse; 2365} 2366 2367/* 2368 * Release all resources used by a slab cache. 2369 */ 2370static inline int kmem_cache_close(struct kmem_cache *s) 2371{ 2372 int node; 2373 2374 flush_all(s); 2375 2376 /* Attempt to free all objects */ 2377 free_kmem_cache_cpus(s); 2378 for_each_node_state(node, N_NORMAL_MEMORY) { 2379 struct kmem_cache_node *n = get_node(s, node); 2380 2381 n->nr_partial -= free_list(s, n, &n->partial); 2382 if (atomic_long_read(&n->nr_slabs)) 2383 return 1; 2384 } 2385 free_kmem_cache_nodes(s); 2386 return 0; 2387} 2388 2389/* 2390 * Close a cache and release the kmem_cache structure 2391 * (must be used for caches created using kmem_cache_create) 2392 */ 2393void kmem_cache_destroy(struct kmem_cache *s) 2394{ 2395 down_write(&slub_lock); 2396 s->refcount--; 2397 if (!s->refcount) { 2398 list_del(&s->list); 2399 up_write(&slub_lock); 2400 if (kmem_cache_close(s)) 2401 WARN_ON(1); 2402 sysfs_slab_remove(s); 2403 } else 2404 up_write(&slub_lock); 2405} 2406EXPORT_SYMBOL(kmem_cache_destroy); 2407 2408/******************************************************************** 2409 * Kmalloc subsystem 2410 *******************************************************************/ 2411 2412struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; 2413EXPORT_SYMBOL(kmalloc_caches); 2414 2415static int __init setup_slub_min_order(char *str) 2416{ 2417 get_option(&str, &slub_min_order); 2418 2419 return 1; 2420} 2421 2422__setup("slub_min_order=", setup_slub_min_order); 2423 2424static int __init setup_slub_max_order(char *str) 2425{ 2426 get_option(&str, &slub_max_order); 2427 2428 return 1; 2429} 2430 2431__setup("slub_max_order=", setup_slub_max_order); 2432 2433static int __init setup_slub_min_objects(char *str) 2434{ 2435 get_option(&str, &slub_min_objects); 2436 2437 return 1; 2438} 2439 2440__setup("slub_min_objects=", setup_slub_min_objects); 2441 2442static int __init setup_slub_nomerge(char *str) 2443{ 2444 slub_nomerge = 1; 2445 return 1; 2446} 2447 2448__setup("slub_nomerge", setup_slub_nomerge); 2449 2450static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2451 const char *name, int size, gfp_t gfp_flags) 2452{ 2453 unsigned int flags = 0; 2454 2455 if (gfp_flags & SLUB_DMA) 2456 flags = SLAB_CACHE_DMA; 2457 2458 down_write(&slub_lock); 2459 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2460 flags | __KMALLOC_CACHE, NULL)) 2461 goto panic; 2462 2463 list_add(&s->list, &slab_caches); 2464 up_write(&slub_lock); 2465 if (sysfs_slab_add(s)) 2466 goto panic; 2467 return s; 2468 2469panic: 2470 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2471} 2472 2473#ifdef CONFIG_ZONE_DMA 2474static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; 2475 2476static void sysfs_add_func(struct work_struct *w) 2477{ 2478 struct kmem_cache *s; 2479 2480 down_write(&slub_lock); 2481 list_for_each_entry(s, &slab_caches, list) { 2482 if (s->flags & __SYSFS_ADD_DEFERRED) { 2483 s->flags &= ~__SYSFS_ADD_DEFERRED; 2484 sysfs_slab_add(s); 2485 } 2486 } 2487 up_write(&slub_lock); 2488} 2489 2490static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2491 2492static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2493{ 2494 struct kmem_cache *s; 2495 char *text; 2496 size_t realsize; 2497 2498 s = kmalloc_caches_dma[index]; 2499 if (s) 2500 return s; 2501 2502 /* Dynamically create dma cache */ 2503 if (flags & __GFP_WAIT) 2504 down_write(&slub_lock); 2505 else { 2506 if (!down_write_trylock(&slub_lock)) 2507 goto out; 2508 } 2509 2510 if (kmalloc_caches_dma[index]) 2511 goto unlock_out; 2512 2513 realsize = kmalloc_caches[index].objsize; 2514 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", 2515 (unsigned int)realsize); 2516 s = kmalloc(kmem_size, flags & ~SLUB_DMA); 2517 2518 if (!s || !text || !kmem_cache_open(s, flags, text, 2519 realsize, ARCH_KMALLOC_MINALIGN, 2520 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { 2521 kfree(s); 2522 kfree(text); 2523 goto unlock_out; 2524 } 2525 2526 list_add(&s->list, &slab_caches); 2527 kmalloc_caches_dma[index] = s; 2528 2529 schedule_work(&sysfs_add_work); 2530 2531unlock_out: 2532 up_write(&slub_lock); 2533out: 2534 return kmalloc_caches_dma[index]; 2535} 2536#endif 2537 2538/* 2539 * Conversion table for small slabs sizes / 8 to the index in the 2540 * kmalloc array. This is necessary for slabs < 192 since we have non power 2541 * of two cache sizes there. The size of larger slabs can be determined using 2542 * fls. 2543 */ 2544static s8 size_index[24] = { 2545 3, /* 8 */ 2546 4, /* 16 */ 2547 5, /* 24 */ 2548 5, /* 32 */ 2549 6, /* 40 */ 2550 6, /* 48 */ 2551 6, /* 56 */ 2552 6, /* 64 */ 2553 1, /* 72 */ 2554 1, /* 80 */ 2555 1, /* 88 */ 2556 1, /* 96 */ 2557 7, /* 104 */ 2558 7, /* 112 */ 2559 7, /* 120 */ 2560 7, /* 128 */ 2561 2, /* 136 */ 2562 2, /* 144 */ 2563 2, /* 152 */ 2564 2, /* 160 */ 2565 2, /* 168 */ 2566 2, /* 176 */ 2567 2, /* 184 */ 2568 2 /* 192 */ 2569}; 2570 2571static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2572{ 2573 int index; 2574 2575 if (size <= 192) { 2576 if (!size) 2577 return ZERO_SIZE_PTR; 2578 2579 index = size_index[(size - 1) / 8]; 2580 } else 2581 index = fls(size - 1); 2582 2583#ifdef CONFIG_ZONE_DMA 2584 if (unlikely((flags & SLUB_DMA))) 2585 return dma_kmalloc_cache(index, flags); 2586 2587#endif 2588 return &kmalloc_caches[index]; 2589} 2590 2591void *__kmalloc(size_t size, gfp_t flags) 2592{ 2593 struct kmem_cache *s; 2594 2595 if (unlikely(size > PAGE_SIZE)) 2596 return kmalloc_large(size, flags); 2597 2598 s = get_slab(size, flags); 2599 2600 if (unlikely(ZERO_OR_NULL_PTR(s))) 2601 return s; 2602 2603 return slab_alloc(s, flags, -1, __builtin_return_address(0)); 2604} 2605EXPORT_SYMBOL(__kmalloc); 2606 2607static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2608{ 2609 struct page *page = alloc_pages_node(node, flags | __GFP_COMP, 2610 get_order(size)); 2611 2612 if (page) 2613 return page_address(page); 2614 else 2615 return NULL; 2616} 2617 2618#ifdef CONFIG_NUMA 2619void *__kmalloc_node(size_t size, gfp_t flags, int node) 2620{ 2621 struct kmem_cache *s; 2622 2623 if (unlikely(size > PAGE_SIZE)) 2624 return kmalloc_large_node(size, flags, node); 2625 2626 s = get_slab(size, flags); 2627 2628 if (unlikely(ZERO_OR_NULL_PTR(s))) 2629 return s; 2630 2631 return slab_alloc(s, flags, node, __builtin_return_address(0)); 2632} 2633EXPORT_SYMBOL(__kmalloc_node); 2634#endif 2635 2636size_t ksize(const void *object) 2637{ 2638 struct page *page; 2639 struct kmem_cache *s; 2640 2641 if (unlikely(object == ZERO_SIZE_PTR)) 2642 return 0; 2643 2644 page = virt_to_head_page(object); 2645 2646 if (unlikely(!PageSlab(page))) 2647 return PAGE_SIZE << compound_order(page); 2648 2649 s = page->slab; 2650 2651#ifdef CONFIG_SLUB_DEBUG 2652 /* 2653 * Debugging requires use of the padding between object 2654 * and whatever may come after it. 2655 */ 2656 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2657 return s->objsize; 2658 2659#endif 2660 /* 2661 * If we have the need to store the freelist pointer 2662 * back there or track user information then we can 2663 * only use the space before that information. 2664 */ 2665 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2666 return s->inuse; 2667 /* 2668 * Else we can use all the padding etc for the allocation 2669 */ 2670 return s->size; 2671} 2672EXPORT_SYMBOL(ksize); 2673 2674void kfree(const void *x) 2675{ 2676 struct page *page; 2677 void *object = (void *)x; 2678 2679 if (unlikely(ZERO_OR_NULL_PTR(x))) 2680 return; 2681 2682 page = virt_to_head_page(x); 2683 if (unlikely(!PageSlab(page))) { 2684 put_page(page); 2685 return; 2686 } 2687 slab_free(page->slab, page, object, __builtin_return_address(0)); 2688} 2689EXPORT_SYMBOL(kfree); 2690 2691/* 2692 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2693 * the remaining slabs by the number of items in use. The slabs with the 2694 * most items in use come first. New allocations will then fill those up 2695 * and thus they can be removed from the partial lists. 2696 * 2697 * The slabs with the least items are placed last. This results in them 2698 * being allocated from last increasing the chance that the last objects 2699 * are freed in them. 2700 */ 2701int kmem_cache_shrink(struct kmem_cache *s) 2702{ 2703 int node; 2704 int i; 2705 struct kmem_cache_node *n; 2706 struct page *page; 2707 struct page *t; 2708 struct list_head *slabs_by_inuse = 2709 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL); 2710 unsigned long flags; 2711 2712 if (!slabs_by_inuse) 2713 return -ENOMEM; 2714 2715 flush_all(s); 2716 for_each_node_state(node, N_NORMAL_MEMORY) { 2717 n = get_node(s, node); 2718 2719 if (!n->nr_partial) 2720 continue; 2721 2722 for (i = 0; i < s->objects; i++) 2723 INIT_LIST_HEAD(slabs_by_inuse + i); 2724 2725 spin_lock_irqsave(&n->list_lock, flags); 2726 2727 /* 2728 * Build lists indexed by the items in use in each slab. 2729 * 2730 * Note that concurrent frees may occur while we hold the 2731 * list_lock. page->inuse here is the upper limit. 2732 */ 2733 list_for_each_entry_safe(page, t, &n->partial, lru) { 2734 if (!page->inuse && slab_trylock(page)) { 2735 /* 2736 * Must hold slab lock here because slab_free 2737 * may have freed the last object and be 2738 * waiting to release the slab. 2739 */ 2740 list_del(&page->lru); 2741 n->nr_partial--; 2742 slab_unlock(page); 2743 discard_slab(s, page); 2744 } else { 2745 list_move(&page->lru, 2746 slabs_by_inuse + page->inuse); 2747 } 2748 } 2749 2750 /* 2751 * Rebuild the partial list with the slabs filled up most 2752 * first and the least used slabs at the end. 2753 */ 2754 for (i = s->objects - 1; i >= 0; i--) 2755 list_splice(slabs_by_inuse + i, n->partial.prev); 2756 2757 spin_unlock_irqrestore(&n->list_lock, flags); 2758 } 2759 2760 kfree(slabs_by_inuse); 2761 return 0; 2762} 2763EXPORT_SYMBOL(kmem_cache_shrink); 2764 2765#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2766static int slab_mem_going_offline_callback(void *arg) 2767{ 2768 struct kmem_cache *s; 2769 2770 down_read(&slub_lock); 2771 list_for_each_entry(s, &slab_caches, list) 2772 kmem_cache_shrink(s); 2773 up_read(&slub_lock); 2774 2775 return 0; 2776} 2777 2778static void slab_mem_offline_callback(void *arg) 2779{ 2780 struct kmem_cache_node *n; 2781 struct kmem_cache *s; 2782 struct memory_notify *marg = arg; 2783 int offline_node; 2784 2785 offline_node = marg->status_change_nid; 2786 2787 /* 2788 * If the node still has available memory. we need kmem_cache_node 2789 * for it yet. 2790 */ 2791 if (offline_node < 0) 2792 return; 2793 2794 down_read(&slub_lock); 2795 list_for_each_entry(s, &slab_caches, list) { 2796 n = get_node(s, offline_node); 2797 if (n) { 2798 /* 2799 * if n->nr_slabs > 0, slabs still exist on the node 2800 * that is going down. We were unable to free them, 2801 * and offline_pages() function shoudn't call this 2802 * callback. So, we must fail. 2803 */ 2804 BUG_ON(atomic_long_read(&n->nr_slabs)); 2805 2806 s->node[offline_node] = NULL; 2807 kmem_cache_free(kmalloc_caches, n); 2808 } 2809 } 2810 up_read(&slub_lock); 2811} 2812 2813static int slab_mem_going_online_callback(void *arg) 2814{ 2815 struct kmem_cache_node *n; 2816 struct kmem_cache *s; 2817 struct memory_notify *marg = arg; 2818 int nid = marg->status_change_nid; 2819 int ret = 0; 2820 2821 /* 2822 * If the node's memory is already available, then kmem_cache_node is 2823 * already created. Nothing to do. 2824 */ 2825 if (nid < 0) 2826 return 0; 2827 2828 /* 2829 * We are bringing a node online. No memory is availabe yet. We must 2830 * allocate a kmem_cache_node structure in order to bring the node 2831 * online. 2832 */ 2833 down_read(&slub_lock); 2834 list_for_each_entry(s, &slab_caches, list) { 2835 /* 2836 * XXX: kmem_cache_alloc_node will fallback to other nodes 2837 * since memory is not yet available from the node that 2838 * is brought up. 2839 */ 2840 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 2841 if (!n) { 2842 ret = -ENOMEM; 2843 goto out; 2844 } 2845 init_kmem_cache_node(n); 2846 s->node[nid] = n; 2847 } 2848out: 2849 up_read(&slub_lock); 2850 return ret; 2851} 2852 2853static int slab_memory_callback(struct notifier_block *self, 2854 unsigned long action, void *arg) 2855{ 2856 int ret = 0; 2857 2858 switch (action) { 2859 case MEM_GOING_ONLINE: 2860 ret = slab_mem_going_online_callback(arg); 2861 break; 2862 case MEM_GOING_OFFLINE: 2863 ret = slab_mem_going_offline_callback(arg); 2864 break; 2865 case MEM_OFFLINE: 2866 case MEM_CANCEL_ONLINE: 2867 slab_mem_offline_callback(arg); 2868 break; 2869 case MEM_ONLINE: 2870 case MEM_CANCEL_OFFLINE: 2871 break; 2872 } 2873 2874 ret = notifier_from_errno(ret); 2875 return ret; 2876} 2877 2878#endif /* CONFIG_MEMORY_HOTPLUG */ 2879 2880/******************************************************************** 2881 * Basic setup of slabs 2882 *******************************************************************/ 2883 2884void __init kmem_cache_init(void) 2885{ 2886 int i; 2887 int caches = 0; 2888 2889 init_alloc_cpu(); 2890 2891#ifdef CONFIG_NUMA 2892 /* 2893 * Must first have the slab cache available for the allocations of the 2894 * struct kmem_cache_node's. There is special bootstrap code in 2895 * kmem_cache_open for slab_state == DOWN. 2896 */ 2897 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 2898 sizeof(struct kmem_cache_node), GFP_KERNEL); 2899 kmalloc_caches[0].refcount = -1; 2900 caches++; 2901 2902 hotplug_memory_notifier(slab_memory_callback, 1); 2903#endif 2904 2905 /* Able to allocate the per node structures */ 2906 slab_state = PARTIAL; 2907 2908 /* Caches that are not of the two-to-the-power-of size */ 2909 if (KMALLOC_MIN_SIZE <= 64) { 2910 create_kmalloc_cache(&kmalloc_caches[1], 2911 "kmalloc-96", 96, GFP_KERNEL); 2912 caches++; 2913 } 2914 if (KMALLOC_MIN_SIZE <= 128) { 2915 create_kmalloc_cache(&kmalloc_caches[2], 2916 "kmalloc-192", 192, GFP_KERNEL); 2917 caches++; 2918 } 2919 2920 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { 2921 create_kmalloc_cache(&kmalloc_caches[i], 2922 "kmalloc", 1 << i, GFP_KERNEL); 2923 caches++; 2924 } 2925 2926 2927 /* 2928 * Patch up the size_index table if we have strange large alignment 2929 * requirements for the kmalloc array. This is only the case for 2930 * MIPS it seems. The standard arches will not generate any code here. 2931 * 2932 * Largest permitted alignment is 256 bytes due to the way we 2933 * handle the index determination for the smaller caches. 2934 * 2935 * Make sure that nothing crazy happens if someone starts tinkering 2936 * around with ARCH_KMALLOC_MINALIGN 2937 */ 2938 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 2939 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 2940 2941 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) 2942 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; 2943 2944 slab_state = UP; 2945 2946 /* Provide the correct kmalloc names now that the caches are up */ 2947 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) 2948 kmalloc_caches[i]. name = 2949 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); 2950 2951#ifdef CONFIG_SMP 2952 register_cpu_notifier(&slab_notifier); 2953 kmem_size = offsetof(struct kmem_cache, cpu_slab) + 2954 nr_cpu_ids * sizeof(struct kmem_cache_cpu *); 2955#else 2956 kmem_size = sizeof(struct kmem_cache); 2957#endif 2958 2959 printk(KERN_INFO 2960 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 2961 " CPUs=%d, Nodes=%d\n", 2962 caches, cache_line_size(), 2963 slub_min_order, slub_max_order, slub_min_objects, 2964 nr_cpu_ids, nr_node_ids); 2965} 2966 2967/* 2968 * Find a mergeable slab cache 2969 */ 2970static int slab_unmergeable(struct kmem_cache *s) 2971{ 2972 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 2973 return 1; 2974 2975 if ((s->flags & __PAGE_ALLOC_FALLBACK)) 2976 return 1; 2977 2978 if (s->ctor) 2979 return 1; 2980 2981 /* 2982 * We may have set a slab to be unmergeable during bootstrap. 2983 */ 2984 if (s->refcount < 0) 2985 return 1; 2986 2987 return 0; 2988} 2989 2990static struct kmem_cache *find_mergeable(size_t size, 2991 size_t align, unsigned long flags, const char *name, 2992 void (*ctor)(struct kmem_cache *, void *)) 2993{ 2994 struct kmem_cache *s; 2995 2996 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 2997 return NULL; 2998 2999 if (ctor) 3000 return NULL; 3001 3002 size = ALIGN(size, sizeof(void *)); 3003 align = calculate_alignment(flags, align, size); 3004 size = ALIGN(size, align); 3005 flags = kmem_cache_flags(size, flags, name, NULL); 3006 3007 list_for_each_entry(s, &slab_caches, list) { 3008 if (slab_unmergeable(s)) 3009 continue; 3010 3011 if (size > s->size) 3012 continue; 3013 3014 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3015 continue; 3016 /* 3017 * Check if alignment is compatible. 3018 * Courtesy of Adrian Drzewiecki 3019 */ 3020 if ((s->size & ~(align - 1)) != s->size) 3021 continue; 3022 3023 if (s->size - size >= sizeof(void *)) 3024 continue; 3025 3026 return s; 3027 } 3028 return NULL; 3029} 3030 3031struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3032 size_t align, unsigned long flags, 3033 void (*ctor)(struct kmem_cache *, void *)) 3034{ 3035 struct kmem_cache *s; 3036 3037 down_write(&slub_lock); 3038 s = find_mergeable(size, align, flags, name, ctor); 3039 if (s) { 3040 int cpu; 3041 3042 s->refcount++; 3043 /* 3044 * Adjust the object sizes so that we clear 3045 * the complete object on kzalloc. 3046 */ 3047 s->objsize = max(s->objsize, (int)size); 3048 3049 /* 3050 * And then we need to update the object size in the 3051 * per cpu structures 3052 */ 3053 for_each_online_cpu(cpu) 3054 get_cpu_slab(s, cpu)->objsize = s->objsize; 3055 3056 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3057 up_write(&slub_lock); 3058 3059 if (sysfs_slab_alias(s, name)) 3060 goto err; 3061 return s; 3062 } 3063 3064 s = kmalloc(kmem_size, GFP_KERNEL); 3065 if (s) { 3066 if (kmem_cache_open(s, GFP_KERNEL, name, 3067 size, align, flags, ctor)) { 3068 list_add(&s->list, &slab_caches); 3069 up_write(&slub_lock); 3070 if (sysfs_slab_add(s)) 3071 goto err; 3072 return s; 3073 } 3074 kfree(s); 3075 } 3076 up_write(&slub_lock); 3077 3078err: 3079 if (flags & SLAB_PANIC) 3080 panic("Cannot create slabcache %s\n", name); 3081 else 3082 s = NULL; 3083 return s; 3084} 3085EXPORT_SYMBOL(kmem_cache_create); 3086 3087#ifdef CONFIG_SMP 3088/* 3089 * Use the cpu notifier to insure that the cpu slabs are flushed when 3090 * necessary. 3091 */ 3092static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3093 unsigned long action, void *hcpu) 3094{ 3095 long cpu = (long)hcpu; 3096 struct kmem_cache *s; 3097 unsigned long flags; 3098 3099 switch (action) { 3100 case CPU_UP_PREPARE: 3101 case CPU_UP_PREPARE_FROZEN: 3102 init_alloc_cpu_cpu(cpu); 3103 down_read(&slub_lock); 3104 list_for_each_entry(s, &slab_caches, list) 3105 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, 3106 GFP_KERNEL); 3107 up_read(&slub_lock); 3108 break; 3109 3110 case CPU_UP_CANCELED: 3111 case CPU_UP_CANCELED_FROZEN: 3112 case CPU_DEAD: 3113 case CPU_DEAD_FROZEN: 3114 down_read(&slub_lock); 3115 list_for_each_entry(s, &slab_caches, list) { 3116 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3117 3118 local_irq_save(flags); 3119 __flush_cpu_slab(s, cpu); 3120 local_irq_restore(flags); 3121 free_kmem_cache_cpu(c, cpu); 3122 s->cpu_slab[cpu] = NULL; 3123 } 3124 up_read(&slub_lock); 3125 break; 3126 default: 3127 break; 3128 } 3129 return NOTIFY_OK; 3130} 3131 3132static struct notifier_block __cpuinitdata slab_notifier = { 3133 .notifier_call = slab_cpuup_callback 3134}; 3135 3136#endif 3137 3138void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) 3139{ 3140 struct kmem_cache *s; 3141 3142 if (unlikely(size > PAGE_SIZE)) 3143 return kmalloc_large(size, gfpflags); 3144 3145 s = get_slab(size, gfpflags); 3146 3147 if (unlikely(ZERO_OR_NULL_PTR(s))) 3148 return s; 3149 3150 return slab_alloc(s, gfpflags, -1, caller); 3151} 3152 3153void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3154 int node, void *caller) 3155{ 3156 struct kmem_cache *s; 3157 3158 if (unlikely(size > PAGE_SIZE)) 3159 return kmalloc_large_node(size, gfpflags, node); 3160 3161 s = get_slab(size, gfpflags); 3162 3163 if (unlikely(ZERO_OR_NULL_PTR(s))) 3164 return s; 3165 3166 return slab_alloc(s, gfpflags, node, caller); 3167} 3168 3169#if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO) 3170static unsigned long count_partial(struct kmem_cache_node *n) 3171{ 3172 unsigned long flags; 3173 unsigned long x = 0; 3174 struct page *page; 3175 3176 spin_lock_irqsave(&n->list_lock, flags); 3177 list_for_each_entry(page, &n->partial, lru) 3178 x += page->inuse; 3179 spin_unlock_irqrestore(&n->list_lock, flags); 3180 return x; 3181} 3182#endif 3183 3184#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) 3185static int validate_slab(struct kmem_cache *s, struct page *page, 3186 unsigned long *map) 3187{ 3188 void *p; 3189 void *addr = page_address(page); 3190 3191 if (!check_slab(s, page) || 3192 !on_freelist(s, page, NULL)) 3193 return 0; 3194 3195 /* Now we know that a valid freelist exists */ 3196 bitmap_zero(map, s->objects); 3197 3198 for_each_free_object(p, s, page->freelist) { 3199 set_bit(slab_index(p, s, addr), map); 3200 if (!check_object(s, page, p, 0)) 3201 return 0; 3202 } 3203 3204 for_each_object(p, s, addr) 3205 if (!test_bit(slab_index(p, s, addr), map)) 3206 if (!check_object(s, page, p, 1)) 3207 return 0; 3208 return 1; 3209} 3210 3211static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3212 unsigned long *map) 3213{ 3214 if (slab_trylock(page)) { 3215 validate_slab(s, page, map); 3216 slab_unlock(page); 3217 } else 3218 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3219 s->name, page); 3220 3221 if (s->flags & DEBUG_DEFAULT_FLAGS) { 3222 if (!SlabDebug(page)) 3223 printk(KERN_ERR "SLUB %s: SlabDebug not set " 3224 "on slab 0x%p\n", s->name, page); 3225 } else { 3226 if (SlabDebug(page)) 3227 printk(KERN_ERR "SLUB %s: SlabDebug set on " 3228 "slab 0x%p\n", s->name, page); 3229 } 3230} 3231 3232static int validate_slab_node(struct kmem_cache *s, 3233 struct kmem_cache_node *n, unsigned long *map) 3234{ 3235 unsigned long count = 0; 3236 struct page *page; 3237 unsigned long flags; 3238 3239 spin_lock_irqsave(&n->list_lock, flags); 3240 3241 list_for_each_entry(page, &n->partial, lru) { 3242 validate_slab_slab(s, page, map); 3243 count++; 3244 } 3245 if (count != n->nr_partial) 3246 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3247 "counter=%ld\n", s->name, count, n->nr_partial); 3248 3249 if (!(s->flags & SLAB_STORE_USER)) 3250 goto out; 3251 3252 list_for_each_entry(page, &n->full, lru) { 3253 validate_slab_slab(s, page, map); 3254 count++; 3255 } 3256 if (count != atomic_long_read(&n->nr_slabs)) 3257 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3258 "counter=%ld\n", s->name, count, 3259 atomic_long_read(&n->nr_slabs)); 3260 3261out: 3262 spin_unlock_irqrestore(&n->list_lock, flags); 3263 return count; 3264} 3265 3266static long validate_slab_cache(struct kmem_cache *s) 3267{ 3268 int node; 3269 unsigned long count = 0; 3270 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) * 3271 sizeof(unsigned long), GFP_KERNEL); 3272 3273 if (!map) 3274 return -ENOMEM; 3275 3276 flush_all(s); 3277 for_each_node_state(node, N_NORMAL_MEMORY) { 3278 struct kmem_cache_node *n = get_node(s, node); 3279 3280 count += validate_slab_node(s, n, map); 3281 } 3282 kfree(map); 3283 return count; 3284} 3285 3286#ifdef SLUB_RESILIENCY_TEST 3287static void resiliency_test(void) 3288{ 3289 u8 *p; 3290 3291 printk(KERN_ERR "SLUB resiliency testing\n"); 3292 printk(KERN_ERR "-----------------------\n"); 3293 printk(KERN_ERR "A. Corruption after allocation\n"); 3294 3295 p = kzalloc(16, GFP_KERNEL); 3296 p[16] = 0x12; 3297 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3298 " 0x12->0x%p\n\n", p + 16); 3299 3300 validate_slab_cache(kmalloc_caches + 4); 3301 3302 /* Hmmm... The next two are dangerous */ 3303 p = kzalloc(32, GFP_KERNEL); 3304 p[32 + sizeof(void *)] = 0x34; 3305 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3306 " 0x34 -> -0x%p\n", p); 3307 printk(KERN_ERR 3308 "If allocated object is overwritten then not detectable\n\n"); 3309 3310 validate_slab_cache(kmalloc_caches + 5); 3311 p = kzalloc(64, GFP_KERNEL); 3312 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3313 *p = 0x56; 3314 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3315 p); 3316 printk(KERN_ERR 3317 "If allocated object is overwritten then not detectable\n\n"); 3318 validate_slab_cache(kmalloc_caches + 6); 3319 3320 printk(KERN_ERR "\nB. Corruption after free\n"); 3321 p = kzalloc(128, GFP_KERNEL); 3322 kfree(p); 3323 *p = 0x78; 3324 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3325 validate_slab_cache(kmalloc_caches + 7); 3326 3327 p = kzalloc(256, GFP_KERNEL); 3328 kfree(p); 3329 p[50] = 0x9a; 3330 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3331 p); 3332 validate_slab_cache(kmalloc_caches + 8); 3333 3334 p = kzalloc(512, GFP_KERNEL); 3335 kfree(p); 3336 p[512] = 0xab; 3337 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3338 validate_slab_cache(kmalloc_caches + 9); 3339} 3340#else 3341static void resiliency_test(void) {}; 3342#endif 3343 3344/* 3345 * Generate lists of code addresses where slabcache objects are allocated 3346 * and freed. 3347 */ 3348 3349struct location { 3350 unsigned long count; 3351 void *addr; 3352 long long sum_time; 3353 long min_time; 3354 long max_time; 3355 long min_pid; 3356 long max_pid; 3357 cpumask_t cpus; 3358 nodemask_t nodes; 3359}; 3360 3361struct loc_track { 3362 unsigned long max; 3363 unsigned long count; 3364 struct location *loc; 3365}; 3366 3367static void free_loc_track(struct loc_track *t) 3368{ 3369 if (t->max) 3370 free_pages((unsigned long)t->loc, 3371 get_order(sizeof(struct location) * t->max)); 3372} 3373 3374static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3375{ 3376 struct location *l; 3377 int order; 3378 3379 order = get_order(sizeof(struct location) * max); 3380 3381 l = (void *)__get_free_pages(flags, order); 3382 if (!l) 3383 return 0; 3384 3385 if (t->count) { 3386 memcpy(l, t->loc, sizeof(struct location) * t->count); 3387 free_loc_track(t); 3388 } 3389 t->max = max; 3390 t->loc = l; 3391 return 1; 3392} 3393 3394static int add_location(struct loc_track *t, struct kmem_cache *s, 3395 const struct track *track) 3396{ 3397 long start, end, pos; 3398 struct location *l; 3399 void *caddr; 3400 unsigned long age = jiffies - track->when; 3401 3402 start = -1; 3403 end = t->count; 3404 3405 for ( ; ; ) { 3406 pos = start + (end - start + 1) / 2; 3407 3408 /* 3409 * There is nothing at "end". If we end up there 3410 * we need to add something to before end. 3411 */ 3412 if (pos == end) 3413 break; 3414 3415 caddr = t->loc[pos].addr; 3416 if (track->addr == caddr) { 3417 3418 l = &t->loc[pos]; 3419 l->count++; 3420 if (track->when) { 3421 l->sum_time += age; 3422 if (age < l->min_time) 3423 l->min_time = age; 3424 if (age > l->max_time) 3425 l->max_time = age; 3426 3427 if (track->pid < l->min_pid) 3428 l->min_pid = track->pid; 3429 if (track->pid > l->max_pid) 3430 l->max_pid = track->pid; 3431 3432 cpu_set(track->cpu, l->cpus); 3433 } 3434 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3435 return 1; 3436 } 3437 3438 if (track->addr < caddr) 3439 end = pos; 3440 else 3441 start = pos; 3442 } 3443 3444 /* 3445 * Not found. Insert new tracking element. 3446 */ 3447 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3448 return 0; 3449 3450 l = t->loc + pos; 3451 if (pos < t->count) 3452 memmove(l + 1, l, 3453 (t->count - pos) * sizeof(struct location)); 3454 t->count++; 3455 l->count = 1; 3456 l->addr = track->addr; 3457 l->sum_time = age; 3458 l->min_time = age; 3459 l->max_time = age; 3460 l->min_pid = track->pid; 3461 l->max_pid = track->pid; 3462 cpus_clear(l->cpus); 3463 cpu_set(track->cpu, l->cpus); 3464 nodes_clear(l->nodes); 3465 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3466 return 1; 3467} 3468 3469static void process_slab(struct loc_track *t, struct kmem_cache *s, 3470 struct page *page, enum track_item alloc) 3471{ 3472 void *addr = page_address(page); 3473 DECLARE_BITMAP(map, s->objects); 3474 void *p; 3475 3476 bitmap_zero(map, s->objects); 3477 for_each_free_object(p, s, page->freelist) 3478 set_bit(slab_index(p, s, addr), map); 3479 3480 for_each_object(p, s, addr) 3481 if (!test_bit(slab_index(p, s, addr), map)) 3482 add_location(t, s, get_track(s, p, alloc)); 3483} 3484 3485static int list_locations(struct kmem_cache *s, char *buf, 3486 enum track_item alloc) 3487{ 3488 int len = 0; 3489 unsigned long i; 3490 struct loc_track t = { 0, 0, NULL }; 3491 int node; 3492 3493 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3494 GFP_TEMPORARY)) 3495 return sprintf(buf, "Out of memory\n"); 3496 3497 /* Push back cpu slabs */ 3498 flush_all(s); 3499 3500 for_each_node_state(node, N_NORMAL_MEMORY) { 3501 struct kmem_cache_node *n = get_node(s, node); 3502 unsigned long flags; 3503 struct page *page; 3504 3505 if (!atomic_long_read(&n->nr_slabs)) 3506 continue; 3507 3508 spin_lock_irqsave(&n->list_lock, flags); 3509 list_for_each_entry(page, &n->partial, lru) 3510 process_slab(&t, s, page, alloc); 3511 list_for_each_entry(page, &n->full, lru) 3512 process_slab(&t, s, page, alloc); 3513 spin_unlock_irqrestore(&n->list_lock, flags); 3514 } 3515 3516 for (i = 0; i < t.count; i++) { 3517 struct location *l = &t.loc[i]; 3518 3519 if (len > PAGE_SIZE - 100) 3520 break; 3521 len += sprintf(buf + len, "%7ld ", l->count); 3522 3523 if (l->addr) 3524 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3525 else 3526 len += sprintf(buf + len, "<not-available>"); 3527 3528 if (l->sum_time != l->min_time) { 3529 unsigned long remainder; 3530 3531 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3532 l->min_time, 3533 div_long_long_rem(l->sum_time, l->count, &remainder), 3534 l->max_time); 3535 } else 3536 len += sprintf(buf + len, " age=%ld", 3537 l->min_time); 3538 3539 if (l->min_pid != l->max_pid) 3540 len += sprintf(buf + len, " pid=%ld-%ld", 3541 l->min_pid, l->max_pid); 3542 else 3543 len += sprintf(buf + len, " pid=%ld", 3544 l->min_pid); 3545 3546 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && 3547 len < PAGE_SIZE - 60) { 3548 len += sprintf(buf + len, " cpus="); 3549 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3550 l->cpus); 3551 } 3552 3553 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && 3554 len < PAGE_SIZE - 60) { 3555 len += sprintf(buf + len, " nodes="); 3556 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3557 l->nodes); 3558 } 3559 3560 len += sprintf(buf + len, "\n"); 3561 } 3562 3563 free_loc_track(&t); 3564 if (!t.count) 3565 len += sprintf(buf, "No data\n"); 3566 return len; 3567} 3568 3569enum slab_stat_type { 3570 SL_FULL, 3571 SL_PARTIAL, 3572 SL_CPU, 3573 SL_OBJECTS 3574}; 3575 3576#define SO_FULL (1 << SL_FULL) 3577#define SO_PARTIAL (1 << SL_PARTIAL) 3578#define SO_CPU (1 << SL_CPU) 3579#define SO_OBJECTS (1 << SL_OBJECTS) 3580 3581static ssize_t show_slab_objects(struct kmem_cache *s, 3582 char *buf, unsigned long flags) 3583{ 3584 unsigned long total = 0; 3585 int cpu; 3586 int node; 3587 int x; 3588 unsigned long *nodes; 3589 unsigned long *per_cpu; 3590 3591 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3592 if (!nodes) 3593 return -ENOMEM; 3594 per_cpu = nodes + nr_node_ids; 3595 3596 for_each_possible_cpu(cpu) { 3597 struct page *page; 3598 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3599 3600 if (!c) 3601 continue; 3602 3603 page = c->page; 3604 node = c->node; 3605 if (node < 0) 3606 continue; 3607 if (page) { 3608 if (flags & SO_CPU) { 3609 if (flags & SO_OBJECTS) 3610 x = page->inuse; 3611 else 3612 x = 1; 3613 total += x; 3614 nodes[node] += x; 3615 } 3616 per_cpu[node]++; 3617 } 3618 } 3619 3620 for_each_node_state(node, N_NORMAL_MEMORY) { 3621 struct kmem_cache_node *n = get_node(s, node); 3622 3623 if (flags & SO_PARTIAL) { 3624 if (flags & SO_OBJECTS) 3625 x = count_partial(n); 3626 else 3627 x = n->nr_partial; 3628 total += x; 3629 nodes[node] += x; 3630 } 3631 3632 if (flags & SO_FULL) { 3633 int full_slabs = atomic_long_read(&n->nr_slabs) 3634 - per_cpu[node] 3635 - n->nr_partial; 3636 3637 if (flags & SO_OBJECTS) 3638 x = full_slabs * s->objects; 3639 else 3640 x = full_slabs; 3641 total += x; 3642 nodes[node] += x; 3643 } 3644 } 3645 3646 x = sprintf(buf, "%lu", total); 3647#ifdef CONFIG_NUMA 3648 for_each_node_state(node, N_NORMAL_MEMORY) 3649 if (nodes[node]) 3650 x += sprintf(buf + x, " N%d=%lu", 3651 node, nodes[node]); 3652#endif 3653 kfree(nodes); 3654 return x + sprintf(buf + x, "\n"); 3655} 3656 3657static int any_slab_objects(struct kmem_cache *s) 3658{ 3659 int node; 3660 int cpu; 3661 3662 for_each_possible_cpu(cpu) { 3663 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3664 3665 if (c && c->page) 3666 return 1; 3667 } 3668 3669 for_each_online_node(node) { 3670 struct kmem_cache_node *n = get_node(s, node); 3671 3672 if (!n) 3673 continue; 3674 3675 if (n->nr_partial || atomic_long_read(&n->nr_slabs)) 3676 return 1; 3677 } 3678 return 0; 3679} 3680 3681#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3682#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3683 3684struct slab_attribute { 3685 struct attribute attr; 3686 ssize_t (*show)(struct kmem_cache *s, char *buf); 3687 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3688}; 3689 3690#define SLAB_ATTR_RO(_name) \ 3691 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3692 3693#define SLAB_ATTR(_name) \ 3694 static struct slab_attribute _name##_attr = \ 3695 __ATTR(_name, 0644, _name##_show, _name##_store) 3696 3697static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3698{ 3699 return sprintf(buf, "%d\n", s->size); 3700} 3701SLAB_ATTR_RO(slab_size); 3702 3703static ssize_t align_show(struct kmem_cache *s, char *buf) 3704{ 3705 return sprintf(buf, "%d\n", s->align); 3706} 3707SLAB_ATTR_RO(align); 3708 3709static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3710{ 3711 return sprintf(buf, "%d\n", s->objsize); 3712} 3713SLAB_ATTR_RO(object_size); 3714 3715static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3716{ 3717 return sprintf(buf, "%d\n", s->objects); 3718} 3719SLAB_ATTR_RO(objs_per_slab); 3720 3721static ssize_t order_show(struct kmem_cache *s, char *buf) 3722{ 3723 return sprintf(buf, "%d\n", s->order); 3724} 3725SLAB_ATTR_RO(order); 3726 3727static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3728{ 3729 if (s->ctor) { 3730 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3731 3732 return n + sprintf(buf + n, "\n"); 3733 } 3734 return 0; 3735} 3736SLAB_ATTR_RO(ctor); 3737 3738static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3739{ 3740 return sprintf(buf, "%d\n", s->refcount - 1); 3741} 3742SLAB_ATTR_RO(aliases); 3743 3744static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3745{ 3746 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU); 3747} 3748SLAB_ATTR_RO(slabs); 3749 3750static ssize_t partial_show(struct kmem_cache *s, char *buf) 3751{ 3752 return show_slab_objects(s, buf, SO_PARTIAL); 3753} 3754SLAB_ATTR_RO(partial); 3755 3756static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 3757{ 3758 return show_slab_objects(s, buf, SO_CPU); 3759} 3760SLAB_ATTR_RO(cpu_slabs); 3761 3762static ssize_t objects_show(struct kmem_cache *s, char *buf) 3763{ 3764 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS); 3765} 3766SLAB_ATTR_RO(objects); 3767 3768static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 3769{ 3770 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 3771} 3772 3773static ssize_t sanity_checks_store(struct kmem_cache *s, 3774 const char *buf, size_t length) 3775{ 3776 s->flags &= ~SLAB_DEBUG_FREE; 3777 if (buf[0] == '1') 3778 s->flags |= SLAB_DEBUG_FREE; 3779 return length; 3780} 3781SLAB_ATTR(sanity_checks); 3782 3783static ssize_t trace_show(struct kmem_cache *s, char *buf) 3784{ 3785 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 3786} 3787 3788static ssize_t trace_store(struct kmem_cache *s, const char *buf, 3789 size_t length) 3790{ 3791 s->flags &= ~SLAB_TRACE; 3792 if (buf[0] == '1') 3793 s->flags |= SLAB_TRACE; 3794 return length; 3795} 3796SLAB_ATTR(trace); 3797 3798static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 3799{ 3800 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 3801} 3802 3803static ssize_t reclaim_account_store(struct kmem_cache *s, 3804 const char *buf, size_t length) 3805{ 3806 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 3807 if (buf[0] == '1') 3808 s->flags |= SLAB_RECLAIM_ACCOUNT; 3809 return length; 3810} 3811SLAB_ATTR(reclaim_account); 3812 3813static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 3814{ 3815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 3816} 3817SLAB_ATTR_RO(hwcache_align); 3818 3819#ifdef CONFIG_ZONE_DMA 3820static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 3821{ 3822 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 3823} 3824SLAB_ATTR_RO(cache_dma); 3825#endif 3826 3827static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 3828{ 3829 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 3830} 3831SLAB_ATTR_RO(destroy_by_rcu); 3832 3833static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 3834{ 3835 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 3836} 3837 3838static ssize_t red_zone_store(struct kmem_cache *s, 3839 const char *buf, size_t length) 3840{ 3841 if (any_slab_objects(s)) 3842 return -EBUSY; 3843 3844 s->flags &= ~SLAB_RED_ZONE; 3845 if (buf[0] == '1') 3846 s->flags |= SLAB_RED_ZONE; 3847 calculate_sizes(s); 3848 return length; 3849} 3850SLAB_ATTR(red_zone); 3851 3852static ssize_t poison_show(struct kmem_cache *s, char *buf) 3853{ 3854 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 3855} 3856 3857static ssize_t poison_store(struct kmem_cache *s, 3858 const char *buf, size_t length) 3859{ 3860 if (any_slab_objects(s)) 3861 return -EBUSY; 3862 3863 s->flags &= ~SLAB_POISON; 3864 if (buf[0] == '1') 3865 s->flags |= SLAB_POISON; 3866 calculate_sizes(s); 3867 return length; 3868} 3869SLAB_ATTR(poison); 3870 3871static ssize_t store_user_show(struct kmem_cache *s, char *buf) 3872{ 3873 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 3874} 3875 3876static ssize_t store_user_store(struct kmem_cache *s, 3877 const char *buf, size_t length) 3878{ 3879 if (any_slab_objects(s)) 3880 return -EBUSY; 3881 3882 s->flags &= ~SLAB_STORE_USER; 3883 if (buf[0] == '1') 3884 s->flags |= SLAB_STORE_USER; 3885 calculate_sizes(s); 3886 return length; 3887} 3888SLAB_ATTR(store_user); 3889 3890static ssize_t validate_show(struct kmem_cache *s, char *buf) 3891{ 3892 return 0; 3893} 3894 3895static ssize_t validate_store(struct kmem_cache *s, 3896 const char *buf, size_t length) 3897{ 3898 int ret = -EINVAL; 3899 3900 if (buf[0] == '1') { 3901 ret = validate_slab_cache(s); 3902 if (ret >= 0) 3903 ret = length; 3904 } 3905 return ret; 3906} 3907SLAB_ATTR(validate); 3908 3909static ssize_t shrink_show(struct kmem_cache *s, char *buf) 3910{ 3911 return 0; 3912} 3913 3914static ssize_t shrink_store(struct kmem_cache *s, 3915 const char *buf, size_t length) 3916{ 3917 if (buf[0] == '1') { 3918 int rc = kmem_cache_shrink(s); 3919 3920 if (rc) 3921 return rc; 3922 } else 3923 return -EINVAL; 3924 return length; 3925} 3926SLAB_ATTR(shrink); 3927 3928static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 3929{ 3930 if (!(s->flags & SLAB_STORE_USER)) 3931 return -ENOSYS; 3932 return list_locations(s, buf, TRACK_ALLOC); 3933} 3934SLAB_ATTR_RO(alloc_calls); 3935 3936static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 3937{ 3938 if (!(s->flags & SLAB_STORE_USER)) 3939 return -ENOSYS; 3940 return list_locations(s, buf, TRACK_FREE); 3941} 3942SLAB_ATTR_RO(free_calls); 3943 3944#ifdef CONFIG_NUMA 3945static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 3946{ 3947 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 3948} 3949 3950static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 3951 const char *buf, size_t length) 3952{ 3953 int n = simple_strtoul(buf, NULL, 10); 3954 3955 if (n < 100) 3956 s->remote_node_defrag_ratio = n * 10; 3957 return length; 3958} 3959SLAB_ATTR(remote_node_defrag_ratio); 3960#endif 3961 3962#ifdef CONFIG_SLUB_STATS 3963static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 3964{ 3965 unsigned long sum = 0; 3966 int cpu; 3967 int len; 3968 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 3969 3970 if (!data) 3971 return -ENOMEM; 3972 3973 for_each_online_cpu(cpu) { 3974 unsigned x = get_cpu_slab(s, cpu)->stat[si]; 3975 3976 data[cpu] = x; 3977 sum += x; 3978 } 3979 3980 len = sprintf(buf, "%lu", sum); 3981 3982#ifdef CONFIG_SMP 3983 for_each_online_cpu(cpu) { 3984 if (data[cpu] && len < PAGE_SIZE - 20) 3985 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 3986 } 3987#endif 3988 kfree(data); 3989 return len + sprintf(buf + len, "\n"); 3990} 3991 3992#define STAT_ATTR(si, text) \ 3993static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 3994{ \ 3995 return show_stat(s, buf, si); \ 3996} \ 3997SLAB_ATTR_RO(text); \ 3998 3999STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4000STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4001STAT_ATTR(FREE_FASTPATH, free_fastpath); 4002STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4003STAT_ATTR(FREE_FROZEN, free_frozen); 4004STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4005STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4006STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4007STAT_ATTR(ALLOC_SLAB, alloc_slab); 4008STAT_ATTR(ALLOC_REFILL, alloc_refill); 4009STAT_ATTR(FREE_SLAB, free_slab); 4010STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4011STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4012STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4013STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4014STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4015STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4016 4017#endif 4018 4019static struct attribute *slab_attrs[] = { 4020 &slab_size_attr.attr, 4021 &object_size_attr.attr, 4022 &objs_per_slab_attr.attr, 4023 &order_attr.attr, 4024 &objects_attr.attr, 4025 &slabs_attr.attr, 4026 &partial_attr.attr, 4027 &cpu_slabs_attr.attr, 4028 &ctor_attr.attr, 4029 &aliases_attr.attr, 4030 &align_attr.attr, 4031 &sanity_checks_attr.attr, 4032 &trace_attr.attr, 4033 &hwcache_align_attr.attr, 4034 &reclaim_account_attr.attr, 4035 &destroy_by_rcu_attr.attr, 4036 &red_zone_attr.attr, 4037 &poison_attr.attr, 4038 &store_user_attr.attr, 4039 &validate_attr.attr, 4040 &shrink_attr.attr, 4041 &alloc_calls_attr.attr, 4042 &free_calls_attr.attr, 4043#ifdef CONFIG_ZONE_DMA 4044 &cache_dma_attr.attr, 4045#endif 4046#ifdef CONFIG_NUMA 4047 &remote_node_defrag_ratio_attr.attr, 4048#endif 4049#ifdef CONFIG_SLUB_STATS 4050 &alloc_fastpath_attr.attr, 4051 &alloc_slowpath_attr.attr, 4052 &free_fastpath_attr.attr, 4053 &free_slowpath_attr.attr, 4054 &free_frozen_attr.attr, 4055 &free_add_partial_attr.attr, 4056 &free_remove_partial_attr.attr, 4057 &alloc_from_partial_attr.attr, 4058 &alloc_slab_attr.attr, 4059 &alloc_refill_attr.attr, 4060 &free_slab_attr.attr, 4061 &cpuslab_flush_attr.attr, 4062 &deactivate_full_attr.attr, 4063 &deactivate_empty_attr.attr, 4064 &deactivate_to_head_attr.attr, 4065 &deactivate_to_tail_attr.attr, 4066 &deactivate_remote_frees_attr.attr, 4067#endif 4068 NULL 4069}; 4070 4071static struct attribute_group slab_attr_group = { 4072 .attrs = slab_attrs, 4073}; 4074 4075static ssize_t slab_attr_show(struct kobject *kobj, 4076 struct attribute *attr, 4077 char *buf) 4078{ 4079 struct slab_attribute *attribute; 4080 struct kmem_cache *s; 4081 int err; 4082 4083 attribute = to_slab_attr(attr); 4084 s = to_slab(kobj); 4085 4086 if (!attribute->show) 4087 return -EIO; 4088 4089 err = attribute->show(s, buf); 4090 4091 return err; 4092} 4093 4094static ssize_t slab_attr_store(struct kobject *kobj, 4095 struct attribute *attr, 4096 const char *buf, size_t len) 4097{ 4098 struct slab_attribute *attribute; 4099 struct kmem_cache *s; 4100 int err; 4101 4102 attribute = to_slab_attr(attr); 4103 s = to_slab(kobj); 4104 4105 if (!attribute->store) 4106 return -EIO; 4107 4108 err = attribute->store(s, buf, len); 4109 4110 return err; 4111} 4112 4113static void kmem_cache_release(struct kobject *kobj) 4114{ 4115 struct kmem_cache *s = to_slab(kobj); 4116 4117 kfree(s); 4118} 4119 4120static struct sysfs_ops slab_sysfs_ops = { 4121 .show = slab_attr_show, 4122 .store = slab_attr_store, 4123}; 4124 4125static struct kobj_type slab_ktype = { 4126 .sysfs_ops = &slab_sysfs_ops, 4127 .release = kmem_cache_release 4128}; 4129 4130static int uevent_filter(struct kset *kset, struct kobject *kobj) 4131{ 4132 struct kobj_type *ktype = get_ktype(kobj); 4133 4134 if (ktype == &slab_ktype) 4135 return 1; 4136 return 0; 4137} 4138 4139static struct kset_uevent_ops slab_uevent_ops = { 4140 .filter = uevent_filter, 4141}; 4142 4143static struct kset *slab_kset; 4144 4145#define ID_STR_LENGTH 64 4146 4147/* Create a unique string id for a slab cache: 4148 * 4149 * Format :[flags-]size 4150 */ 4151static char *create_unique_id(struct kmem_cache *s) 4152{ 4153 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4154 char *p = name; 4155 4156 BUG_ON(!name); 4157 4158 *p++ = ':'; 4159 /* 4160 * First flags affecting slabcache operations. We will only 4161 * get here for aliasable slabs so we do not need to support 4162 * too many flags. The flags here must cover all flags that 4163 * are matched during merging to guarantee that the id is 4164 * unique. 4165 */ 4166 if (s->flags & SLAB_CACHE_DMA) 4167 *p++ = 'd'; 4168 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4169 *p++ = 'a'; 4170 if (s->flags & SLAB_DEBUG_FREE) 4171 *p++ = 'F'; 4172 if (p != name + 1) 4173 *p++ = '-'; 4174 p += sprintf(p, "%07d", s->size); 4175 BUG_ON(p > name + ID_STR_LENGTH - 1); 4176 return name; 4177} 4178 4179static int sysfs_slab_add(struct kmem_cache *s) 4180{ 4181 int err; 4182 const char *name; 4183 int unmergeable; 4184 4185 if (slab_state < SYSFS) 4186 /* Defer until later */ 4187 return 0; 4188 4189 unmergeable = slab_unmergeable(s); 4190 if (unmergeable) { 4191 /* 4192 * Slabcache can never be merged so we can use the name proper. 4193 * This is typically the case for debug situations. In that 4194 * case we can catch duplicate names easily. 4195 */ 4196 sysfs_remove_link(&slab_kset->kobj, s->name); 4197 name = s->name; 4198 } else { 4199 /* 4200 * Create a unique name for the slab as a target 4201 * for the symlinks. 4202 */ 4203 name = create_unique_id(s); 4204 } 4205 4206 s->kobj.kset = slab_kset; 4207 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4208 if (err) { 4209 kobject_put(&s->kobj); 4210 return err; 4211 } 4212 4213 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4214 if (err) 4215 return err; 4216 kobject_uevent(&s->kobj, KOBJ_ADD); 4217 if (!unmergeable) { 4218 /* Setup first alias */ 4219 sysfs_slab_alias(s, s->name); 4220 kfree(name); 4221 } 4222 return 0; 4223} 4224 4225static void sysfs_slab_remove(struct kmem_cache *s) 4226{ 4227 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4228 kobject_del(&s->kobj); 4229 kobject_put(&s->kobj); 4230} 4231 4232/* 4233 * Need to buffer aliases during bootup until sysfs becomes 4234 * available lest we loose that information. 4235 */ 4236struct saved_alias { 4237 struct kmem_cache *s; 4238 const char *name; 4239 struct saved_alias *next; 4240}; 4241 4242static struct saved_alias *alias_list; 4243 4244static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4245{ 4246 struct saved_alias *al; 4247 4248 if (slab_state == SYSFS) { 4249 /* 4250 * If we have a leftover link then remove it. 4251 */ 4252 sysfs_remove_link(&slab_kset->kobj, name); 4253 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4254 } 4255 4256 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4257 if (!al) 4258 return -ENOMEM; 4259 4260 al->s = s; 4261 al->name = name; 4262 al->next = alias_list; 4263 alias_list = al; 4264 return 0; 4265} 4266 4267static int __init slab_sysfs_init(void) 4268{ 4269 struct kmem_cache *s; 4270 int err; 4271 4272 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4273 if (!slab_kset) { 4274 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4275 return -ENOSYS; 4276 } 4277 4278 slab_state = SYSFS; 4279 4280 list_for_each_entry(s, &slab_caches, list) { 4281 err = sysfs_slab_add(s); 4282 if (err) 4283 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4284 " to sysfs\n", s->name); 4285 } 4286 4287 while (alias_list) { 4288 struct saved_alias *al = alias_list; 4289 4290 alias_list = alias_list->next; 4291 err = sysfs_slab_alias(al->s, al->name); 4292 if (err) 4293 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4294 " %s to sysfs\n", s->name); 4295 kfree(al); 4296 } 4297 4298 resiliency_test(); 4299 return 0; 4300} 4301 4302__initcall(slab_sysfs_init); 4303#endif 4304 4305/* 4306 * The /proc/slabinfo ABI 4307 */ 4308#ifdef CONFIG_SLABINFO 4309 4310ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4311 size_t count, loff_t *ppos) 4312{ 4313 return -EINVAL; 4314} 4315 4316 4317static void print_slabinfo_header(struct seq_file *m) 4318{ 4319 seq_puts(m, "slabinfo - version: 2.1\n"); 4320 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4321 "<objperslab> <pagesperslab>"); 4322 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4323 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4324 seq_putc(m, '\n'); 4325} 4326 4327static void *s_start(struct seq_file *m, loff_t *pos) 4328{ 4329 loff_t n = *pos; 4330 4331 down_read(&slub_lock); 4332 if (!n) 4333 print_slabinfo_header(m); 4334 4335 return seq_list_start(&slab_caches, *pos); 4336} 4337 4338static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4339{ 4340 return seq_list_next(p, &slab_caches, pos); 4341} 4342 4343static void s_stop(struct seq_file *m, void *p) 4344{ 4345 up_read(&slub_lock); 4346} 4347 4348static int s_show(struct seq_file *m, void *p) 4349{ 4350 unsigned long nr_partials = 0; 4351 unsigned long nr_slabs = 0; 4352 unsigned long nr_inuse = 0; 4353 unsigned long nr_objs; 4354 struct kmem_cache *s; 4355 int node; 4356 4357 s = list_entry(p, struct kmem_cache, list); 4358 4359 for_each_online_node(node) { 4360 struct kmem_cache_node *n = get_node(s, node); 4361 4362 if (!n) 4363 continue; 4364 4365 nr_partials += n->nr_partial; 4366 nr_slabs += atomic_long_read(&n->nr_slabs); 4367 nr_inuse += count_partial(n); 4368 } 4369 4370 nr_objs = nr_slabs * s->objects; 4371 nr_inuse += (nr_slabs - nr_partials) * s->objects; 4372 4373 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4374 nr_objs, s->size, s->objects, (1 << s->order)); 4375 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4376 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4377 0UL); 4378 seq_putc(m, '\n'); 4379 return 0; 4380} 4381 4382const struct seq_operations slabinfo_op = { 4383 .start = s_start, 4384 .next = s_next, 4385 .stop = s_stop, 4386 .show = s_show, 4387}; 4388 4389#endif /* CONFIG_SLABINFO */ 4390