slub.c revision 294a80a8ed004b383ab214837e1c05ca4098a717
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 2 176 177/* 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 */ 207 208/* Not all arches define cache_line_size */ 209#ifndef cache_line_size 210#define cache_line_size() L1_CACHE_BYTES 211#endif 212 213static int kmem_size = sizeof(struct kmem_cache); 214 215#ifdef CONFIG_SMP 216static struct notifier_block slab_notifier; 217#endif 218 219static enum { 220 DOWN, /* No slab functionality available */ 221 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 222 UP, /* Everything works but does not show up in sysfs */ 223 SYSFS /* Sysfs up */ 224} slab_state = DOWN; 225 226/* A list of all slab caches on the system */ 227static DECLARE_RWSEM(slub_lock); 228static LIST_HEAD(slab_caches); 229 230/* 231 * Tracking user of a slab. 232 */ 233struct track { 234 void *addr; /* Called from address */ 235 int cpu; /* Was running on cpu */ 236 int pid; /* Pid context */ 237 unsigned long when; /* When did the operation occur */ 238}; 239 240enum track_item { TRACK_ALLOC, TRACK_FREE }; 241 242#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) 243static int sysfs_slab_add(struct kmem_cache *); 244static int sysfs_slab_alias(struct kmem_cache *, const char *); 245static void sysfs_slab_remove(struct kmem_cache *); 246#else 247static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 248static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 249 { return 0; } 250static inline void sysfs_slab_remove(struct kmem_cache *s) {} 251#endif 252 253/******************************************************************** 254 * Core slab cache functions 255 *******************************************************************/ 256 257int slab_is_available(void) 258{ 259 return slab_state >= UP; 260} 261 262static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 263{ 264#ifdef CONFIG_NUMA 265 return s->node[node]; 266#else 267 return &s->local_node; 268#endif 269} 270 271static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) 272{ 273#ifdef CONFIG_SMP 274 return s->cpu_slab[cpu]; 275#else 276 return &s->cpu_slab; 277#endif 278} 279 280static inline int check_valid_pointer(struct kmem_cache *s, 281 struct page *page, const void *object) 282{ 283 void *base; 284 285 if (!object) 286 return 1; 287 288 base = page_address(page); 289 if (object < base || object >= base + s->objects * s->size || 290 (object - base) % s->size) { 291 return 0; 292 } 293 294 return 1; 295} 296 297/* 298 * Slow version of get and set free pointer. 299 * 300 * This version requires touching the cache lines of kmem_cache which 301 * we avoid to do in the fast alloc free paths. There we obtain the offset 302 * from the page struct. 303 */ 304static inline void *get_freepointer(struct kmem_cache *s, void *object) 305{ 306 return *(void **)(object + s->offset); 307} 308 309static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 310{ 311 *(void **)(object + s->offset) = fp; 312} 313 314/* Loop over all objects in a slab */ 315#define for_each_object(__p, __s, __addr) \ 316 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\ 317 __p += (__s)->size) 318 319/* Scan freelist */ 320#define for_each_free_object(__p, __s, __free) \ 321 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 322 323/* Determine object index from a given position */ 324static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 325{ 326 return (p - addr) / s->size; 327} 328 329#ifdef CONFIG_SLUB_DEBUG 330/* 331 * Debug settings: 332 */ 333#ifdef CONFIG_SLUB_DEBUG_ON 334static int slub_debug = DEBUG_DEFAULT_FLAGS; 335#else 336static int slub_debug; 337#endif 338 339static char *slub_debug_slabs; 340 341/* 342 * Object debugging 343 */ 344static void print_section(char *text, u8 *addr, unsigned int length) 345{ 346 int i, offset; 347 int newline = 1; 348 char ascii[17]; 349 350 ascii[16] = 0; 351 352 for (i = 0; i < length; i++) { 353 if (newline) { 354 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 355 newline = 0; 356 } 357 printk(" %02x", addr[i]); 358 offset = i % 16; 359 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 360 if (offset == 15) { 361 printk(" %s\n",ascii); 362 newline = 1; 363 } 364 } 365 if (!newline) { 366 i %= 16; 367 while (i < 16) { 368 printk(" "); 369 ascii[i] = ' '; 370 i++; 371 } 372 printk(" %s\n", ascii); 373 } 374} 375 376static struct track *get_track(struct kmem_cache *s, void *object, 377 enum track_item alloc) 378{ 379 struct track *p; 380 381 if (s->offset) 382 p = object + s->offset + sizeof(void *); 383 else 384 p = object + s->inuse; 385 386 return p + alloc; 387} 388 389static void set_track(struct kmem_cache *s, void *object, 390 enum track_item alloc, void *addr) 391{ 392 struct track *p; 393 394 if (s->offset) 395 p = object + s->offset + sizeof(void *); 396 else 397 p = object + s->inuse; 398 399 p += alloc; 400 if (addr) { 401 p->addr = addr; 402 p->cpu = smp_processor_id(); 403 p->pid = current ? current->pid : -1; 404 p->when = jiffies; 405 } else 406 memset(p, 0, sizeof(struct track)); 407} 408 409static void init_tracking(struct kmem_cache *s, void *object) 410{ 411 if (!(s->flags & SLAB_STORE_USER)) 412 return; 413 414 set_track(s, object, TRACK_FREE, NULL); 415 set_track(s, object, TRACK_ALLOC, NULL); 416} 417 418static void print_track(const char *s, struct track *t) 419{ 420 if (!t->addr) 421 return; 422 423 printk(KERN_ERR "INFO: %s in ", s); 424 __print_symbol("%s", (unsigned long)t->addr); 425 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); 426} 427 428static void print_tracking(struct kmem_cache *s, void *object) 429{ 430 if (!(s->flags & SLAB_STORE_USER)) 431 return; 432 433 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 434 print_track("Freed", get_track(s, object, TRACK_FREE)); 435} 436 437static void print_page_info(struct page *page) 438{ 439 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n", 440 page, page->inuse, page->freelist, page->flags); 441 442} 443 444static void slab_bug(struct kmem_cache *s, char *fmt, ...) 445{ 446 va_list args; 447 char buf[100]; 448 449 va_start(args, fmt); 450 vsnprintf(buf, sizeof(buf), fmt, args); 451 va_end(args); 452 printk(KERN_ERR "========================================" 453 "=====================================\n"); 454 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 455 printk(KERN_ERR "----------------------------------------" 456 "-------------------------------------\n\n"); 457} 458 459static void slab_fix(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 "FIX %s: %s\n", s->name, buf); 468} 469 470static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 471{ 472 unsigned int off; /* Offset of last byte */ 473 u8 *addr = page_address(page); 474 475 print_tracking(s, p); 476 477 print_page_info(page); 478 479 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 480 p, p - addr, get_freepointer(s, p)); 481 482 if (p > addr + 16) 483 print_section("Bytes b4", p - 16, 16); 484 485 print_section("Object", p, min(s->objsize, 128)); 486 487 if (s->flags & SLAB_RED_ZONE) 488 print_section("Redzone", p + s->objsize, 489 s->inuse - s->objsize); 490 491 if (s->offset) 492 off = s->offset + sizeof(void *); 493 else 494 off = s->inuse; 495 496 if (s->flags & SLAB_STORE_USER) 497 off += 2 * sizeof(struct track); 498 499 if (off != s->size) 500 /* Beginning of the filler is the free pointer */ 501 print_section("Padding", p + off, s->size - off); 502 503 dump_stack(); 504} 505 506static void object_err(struct kmem_cache *s, struct page *page, 507 u8 *object, char *reason) 508{ 509 slab_bug(s, reason); 510 print_trailer(s, page, object); 511} 512 513static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 514{ 515 va_list args; 516 char buf[100]; 517 518 va_start(args, fmt); 519 vsnprintf(buf, sizeof(buf), fmt, args); 520 va_end(args); 521 slab_bug(s, fmt); 522 print_page_info(page); 523 dump_stack(); 524} 525 526static void init_object(struct kmem_cache *s, void *object, int active) 527{ 528 u8 *p = object; 529 530 if (s->flags & __OBJECT_POISON) { 531 memset(p, POISON_FREE, s->objsize - 1); 532 p[s->objsize -1] = POISON_END; 533 } 534 535 if (s->flags & SLAB_RED_ZONE) 536 memset(p + s->objsize, 537 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 538 s->inuse - s->objsize); 539} 540 541static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 542{ 543 while (bytes) { 544 if (*start != (u8)value) 545 return start; 546 start++; 547 bytes--; 548 } 549 return NULL; 550} 551 552static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 553 void *from, void *to) 554{ 555 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 556 memset(from, data, to - from); 557} 558 559static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 560 u8 *object, char *what, 561 u8* start, unsigned int value, unsigned int bytes) 562{ 563 u8 *fault; 564 u8 *end; 565 566 fault = check_bytes(start, value, bytes); 567 if (!fault) 568 return 1; 569 570 end = start + bytes; 571 while (end > fault && end[-1] == value) 572 end--; 573 574 slab_bug(s, "%s overwritten", what); 575 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 576 fault, end - 1, fault[0], value); 577 print_trailer(s, page, object); 578 579 restore_bytes(s, what, value, fault, end); 580 return 0; 581} 582 583/* 584 * Object layout: 585 * 586 * object address 587 * Bytes of the object to be managed. 588 * If the freepointer may overlay the object then the free 589 * pointer is the first word of the object. 590 * 591 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 592 * 0xa5 (POISON_END) 593 * 594 * object + s->objsize 595 * Padding to reach word boundary. This is also used for Redzoning. 596 * Padding is extended by another word if Redzoning is enabled and 597 * objsize == inuse. 598 * 599 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 600 * 0xcc (RED_ACTIVE) for objects in use. 601 * 602 * object + s->inuse 603 * Meta data starts here. 604 * 605 * A. Free pointer (if we cannot overwrite object on free) 606 * B. Tracking data for SLAB_STORE_USER 607 * C. Padding to reach required alignment boundary or at mininum 608 * one word if debuggin is on to be able to detect writes 609 * before the word boundary. 610 * 611 * Padding is done using 0x5a (POISON_INUSE) 612 * 613 * object + s->size 614 * Nothing is used beyond s->size. 615 * 616 * If slabcaches are merged then the objsize and inuse boundaries are mostly 617 * ignored. And therefore no slab options that rely on these boundaries 618 * may be used with merged slabcaches. 619 */ 620 621static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 622{ 623 unsigned long off = s->inuse; /* The end of info */ 624 625 if (s->offset) 626 /* Freepointer is placed after the object. */ 627 off += sizeof(void *); 628 629 if (s->flags & SLAB_STORE_USER) 630 /* We also have user information there */ 631 off += 2 * sizeof(struct track); 632 633 if (s->size == off) 634 return 1; 635 636 return check_bytes_and_report(s, page, p, "Object padding", 637 p + off, POISON_INUSE, s->size - off); 638} 639 640static int slab_pad_check(struct kmem_cache *s, struct page *page) 641{ 642 u8 *start; 643 u8 *fault; 644 u8 *end; 645 int length; 646 int remainder; 647 648 if (!(s->flags & SLAB_POISON)) 649 return 1; 650 651 start = page_address(page); 652 end = start + (PAGE_SIZE << s->order); 653 length = s->objects * s->size; 654 remainder = end - (start + length); 655 if (!remainder) 656 return 1; 657 658 fault = check_bytes(start + length, POISON_INUSE, remainder); 659 if (!fault) 660 return 1; 661 while (end > fault && end[-1] == POISON_INUSE) 662 end--; 663 664 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 665 print_section("Padding", start, length); 666 667 restore_bytes(s, "slab padding", POISON_INUSE, start, end); 668 return 0; 669} 670 671static int check_object(struct kmem_cache *s, struct page *page, 672 void *object, int active) 673{ 674 u8 *p = object; 675 u8 *endobject = object + s->objsize; 676 677 if (s->flags & SLAB_RED_ZONE) { 678 unsigned int red = 679 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 680 681 if (!check_bytes_and_report(s, page, object, "Redzone", 682 endobject, red, s->inuse - s->objsize)) 683 return 0; 684 } else { 685 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) 686 check_bytes_and_report(s, page, p, "Alignment padding", endobject, 687 POISON_INUSE, s->inuse - s->objsize); 688 } 689 690 if (s->flags & SLAB_POISON) { 691 if (!active && (s->flags & __OBJECT_POISON) && 692 (!check_bytes_and_report(s, page, p, "Poison", p, 693 POISON_FREE, s->objsize - 1) || 694 !check_bytes_and_report(s, page, p, "Poison", 695 p + s->objsize -1, POISON_END, 1))) 696 return 0; 697 /* 698 * check_pad_bytes cleans up on its own. 699 */ 700 check_pad_bytes(s, page, p); 701 } 702 703 if (!s->offset && active) 704 /* 705 * Object and freepointer overlap. Cannot check 706 * freepointer while object is allocated. 707 */ 708 return 1; 709 710 /* Check free pointer validity */ 711 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 712 object_err(s, page, p, "Freepointer corrupt"); 713 /* 714 * No choice but to zap it and thus loose the remainder 715 * of the free objects in this slab. May cause 716 * another error because the object count is now wrong. 717 */ 718 set_freepointer(s, p, NULL); 719 return 0; 720 } 721 return 1; 722} 723 724static int check_slab(struct kmem_cache *s, struct page *page) 725{ 726 VM_BUG_ON(!irqs_disabled()); 727 728 if (!PageSlab(page)) { 729 slab_err(s, page, "Not a valid slab page"); 730 return 0; 731 } 732 if (page->inuse > s->objects) { 733 slab_err(s, page, "inuse %u > max %u", 734 s->name, page->inuse, s->objects); 735 return 0; 736 } 737 /* Slab_pad_check fixes things up after itself */ 738 slab_pad_check(s, page); 739 return 1; 740} 741 742/* 743 * Determine if a certain object on a page is on the freelist. Must hold the 744 * slab lock to guarantee that the chains are in a consistent state. 745 */ 746static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 747{ 748 int nr = 0; 749 void *fp = page->freelist; 750 void *object = NULL; 751 752 while (fp && nr <= s->objects) { 753 if (fp == search) 754 return 1; 755 if (!check_valid_pointer(s, page, fp)) { 756 if (object) { 757 object_err(s, page, object, 758 "Freechain corrupt"); 759 set_freepointer(s, object, NULL); 760 break; 761 } else { 762 slab_err(s, page, "Freepointer corrupt"); 763 page->freelist = NULL; 764 page->inuse = s->objects; 765 slab_fix(s, "Freelist cleared"); 766 return 0; 767 } 768 break; 769 } 770 object = fp; 771 fp = get_freepointer(s, object); 772 nr++; 773 } 774 775 if (page->inuse != s->objects - nr) { 776 slab_err(s, page, "Wrong object count. Counter is %d but " 777 "counted were %d", page->inuse, s->objects - nr); 778 page->inuse = s->objects - nr; 779 slab_fix(s, "Object count adjusted."); 780 } 781 return search == NULL; 782} 783 784static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc) 785{ 786 if (s->flags & SLAB_TRACE) { 787 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 788 s->name, 789 alloc ? "alloc" : "free", 790 object, page->inuse, 791 page->freelist); 792 793 if (!alloc) 794 print_section("Object", (void *)object, s->objsize); 795 796 dump_stack(); 797 } 798} 799 800/* 801 * Tracking of fully allocated slabs for debugging purposes. 802 */ 803static void add_full(struct kmem_cache_node *n, struct page *page) 804{ 805 spin_lock(&n->list_lock); 806 list_add(&page->lru, &n->full); 807 spin_unlock(&n->list_lock); 808} 809 810static void remove_full(struct kmem_cache *s, struct page *page) 811{ 812 struct kmem_cache_node *n; 813 814 if (!(s->flags & SLAB_STORE_USER)) 815 return; 816 817 n = get_node(s, page_to_nid(page)); 818 819 spin_lock(&n->list_lock); 820 list_del(&page->lru); 821 spin_unlock(&n->list_lock); 822} 823 824static void setup_object_debug(struct kmem_cache *s, struct page *page, 825 void *object) 826{ 827 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 828 return; 829 830 init_object(s, object, 0); 831 init_tracking(s, object); 832} 833 834static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 835 void *object, void *addr) 836{ 837 if (!check_slab(s, page)) 838 goto bad; 839 840 if (object && !on_freelist(s, page, object)) { 841 object_err(s, page, object, "Object already allocated"); 842 goto bad; 843 } 844 845 if (!check_valid_pointer(s, page, object)) { 846 object_err(s, page, object, "Freelist Pointer check fails"); 847 goto bad; 848 } 849 850 if (object && !check_object(s, page, object, 0)) 851 goto bad; 852 853 /* Success perform special debug activities for allocs */ 854 if (s->flags & SLAB_STORE_USER) 855 set_track(s, object, TRACK_ALLOC, addr); 856 trace(s, page, object, 1); 857 init_object(s, object, 1); 858 return 1; 859 860bad: 861 if (PageSlab(page)) { 862 /* 863 * If this is a slab page then lets do the best we can 864 * to avoid issues in the future. Marking all objects 865 * as used avoids touching the remaining objects. 866 */ 867 slab_fix(s, "Marking all objects used"); 868 page->inuse = s->objects; 869 page->freelist = NULL; 870 } 871 return 0; 872} 873 874static int free_debug_processing(struct kmem_cache *s, struct page *page, 875 void *object, void *addr) 876{ 877 if (!check_slab(s, page)) 878 goto fail; 879 880 if (!check_valid_pointer(s, page, object)) { 881 slab_err(s, page, "Invalid object pointer 0x%p", object); 882 goto fail; 883 } 884 885 if (on_freelist(s, page, object)) { 886 object_err(s, page, object, "Object already free"); 887 goto fail; 888 } 889 890 if (!check_object(s, page, object, 1)) 891 return 0; 892 893 if (unlikely(s != page->slab)) { 894 if (!PageSlab(page)) 895 slab_err(s, page, "Attempt to free object(0x%p) " 896 "outside of slab", object); 897 else 898 if (!page->slab) { 899 printk(KERN_ERR 900 "SLUB <none>: no slab for object 0x%p.\n", 901 object); 902 dump_stack(); 903 } 904 else 905 object_err(s, page, object, 906 "page slab pointer corrupt."); 907 goto fail; 908 } 909 910 /* Special debug activities for freeing objects */ 911 if (!SlabFrozen(page) && !page->freelist) 912 remove_full(s, page); 913 if (s->flags & SLAB_STORE_USER) 914 set_track(s, object, TRACK_FREE, addr); 915 trace(s, page, object, 0); 916 init_object(s, object, 0); 917 return 1; 918 919fail: 920 slab_fix(s, "Object at 0x%p not freed", object); 921 return 0; 922} 923 924static int __init setup_slub_debug(char *str) 925{ 926 slub_debug = DEBUG_DEFAULT_FLAGS; 927 if (*str++ != '=' || !*str) 928 /* 929 * No options specified. Switch on full debugging. 930 */ 931 goto out; 932 933 if (*str == ',') 934 /* 935 * No options but restriction on slabs. This means full 936 * debugging for slabs matching a pattern. 937 */ 938 goto check_slabs; 939 940 slub_debug = 0; 941 if (*str == '-') 942 /* 943 * Switch off all debugging measures. 944 */ 945 goto out; 946 947 /* 948 * Determine which debug features should be switched on 949 */ 950 for ( ;*str && *str != ','; str++) { 951 switch (tolower(*str)) { 952 case 'f': 953 slub_debug |= SLAB_DEBUG_FREE; 954 break; 955 case 'z': 956 slub_debug |= SLAB_RED_ZONE; 957 break; 958 case 'p': 959 slub_debug |= SLAB_POISON; 960 break; 961 case 'u': 962 slub_debug |= SLAB_STORE_USER; 963 break; 964 case 't': 965 slub_debug |= SLAB_TRACE; 966 break; 967 default: 968 printk(KERN_ERR "slub_debug option '%c' " 969 "unknown. skipped\n",*str); 970 } 971 } 972 973check_slabs: 974 if (*str == ',') 975 slub_debug_slabs = str + 1; 976out: 977 return 1; 978} 979 980__setup("slub_debug", setup_slub_debug); 981 982static unsigned long kmem_cache_flags(unsigned long objsize, 983 unsigned long flags, const char *name, 984 void (*ctor)(struct kmem_cache *, void *)) 985{ 986 /* 987 * The page->offset field is only 16 bit wide. This is an offset 988 * in units of words from the beginning of an object. If the slab 989 * size is bigger then we cannot move the free pointer behind the 990 * object anymore. 991 * 992 * On 32 bit platforms the limit is 256k. On 64bit platforms 993 * the limit is 512k. 994 * 995 * Debugging or ctor may create a need to move the free 996 * pointer. Fail if this happens. 997 */ 998 if (objsize >= 65535 * sizeof(void *)) { 999 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON | 1000 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU)); 1001 BUG_ON(ctor); 1002 } else { 1003 /* 1004 * Enable debugging if selected on the kernel commandline. 1005 */ 1006 if (slub_debug && (!slub_debug_slabs || 1007 strncmp(slub_debug_slabs, name, 1008 strlen(slub_debug_slabs)) == 0)) 1009 flags |= slub_debug; 1010 } 1011 1012 return flags; 1013} 1014#else 1015static inline void setup_object_debug(struct kmem_cache *s, 1016 struct page *page, void *object) {} 1017 1018static inline int alloc_debug_processing(struct kmem_cache *s, 1019 struct page *page, void *object, void *addr) { return 0; } 1020 1021static inline int free_debug_processing(struct kmem_cache *s, 1022 struct page *page, void *object, void *addr) { return 0; } 1023 1024static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1025 { return 1; } 1026static inline int check_object(struct kmem_cache *s, struct page *page, 1027 void *object, int active) { return 1; } 1028static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1029static inline unsigned long kmem_cache_flags(unsigned long objsize, 1030 unsigned long flags, const char *name, 1031 void (*ctor)(struct kmem_cache *, void *)) 1032{ 1033 return flags; 1034} 1035#define slub_debug 0 1036#endif 1037/* 1038 * Slab allocation and freeing 1039 */ 1040static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1041{ 1042 struct page * page; 1043 int pages = 1 << s->order; 1044 1045 if (s->order) 1046 flags |= __GFP_COMP; 1047 1048 if (s->flags & SLAB_CACHE_DMA) 1049 flags |= SLUB_DMA; 1050 1051 if (s->flags & SLAB_RECLAIM_ACCOUNT) 1052 flags |= __GFP_RECLAIMABLE; 1053 1054 if (node == -1) 1055 page = alloc_pages(flags, s->order); 1056 else 1057 page = alloc_pages_node(node, flags, s->order); 1058 1059 if (!page) 1060 return NULL; 1061 1062 mod_zone_page_state(page_zone(page), 1063 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1064 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1065 pages); 1066 1067 return page; 1068} 1069 1070static void setup_object(struct kmem_cache *s, struct page *page, 1071 void *object) 1072{ 1073 setup_object_debug(s, page, object); 1074 if (unlikely(s->ctor)) 1075 s->ctor(s, object); 1076} 1077 1078static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1079{ 1080 struct page *page; 1081 struct kmem_cache_node *n; 1082 void *start; 1083 void *last; 1084 void *p; 1085 1086 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1087 1088 page = allocate_slab(s, 1089 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1090 if (!page) 1091 goto out; 1092 1093 n = get_node(s, page_to_nid(page)); 1094 if (n) 1095 atomic_long_inc(&n->nr_slabs); 1096 page->slab = s; 1097 page->flags |= 1 << PG_slab; 1098 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | 1099 SLAB_STORE_USER | SLAB_TRACE)) 1100 SetSlabDebug(page); 1101 1102 start = page_address(page); 1103 1104 if (unlikely(s->flags & SLAB_POISON)) 1105 memset(start, POISON_INUSE, PAGE_SIZE << s->order); 1106 1107 last = start; 1108 for_each_object(p, s, start) { 1109 setup_object(s, page, last); 1110 set_freepointer(s, last, p); 1111 last = p; 1112 } 1113 setup_object(s, page, last); 1114 set_freepointer(s, last, NULL); 1115 1116 page->freelist = start; 1117 page->inuse = 0; 1118out: 1119 return page; 1120} 1121 1122static void __free_slab(struct kmem_cache *s, struct page *page) 1123{ 1124 int pages = 1 << s->order; 1125 1126 if (unlikely(SlabDebug(page))) { 1127 void *p; 1128 1129 slab_pad_check(s, page); 1130 for_each_object(p, s, page_address(page)) 1131 check_object(s, page, p, 0); 1132 ClearSlabDebug(page); 1133 } 1134 1135 mod_zone_page_state(page_zone(page), 1136 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1137 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1138 - pages); 1139 1140 __free_pages(page, s->order); 1141} 1142 1143static void rcu_free_slab(struct rcu_head *h) 1144{ 1145 struct page *page; 1146 1147 page = container_of((struct list_head *)h, struct page, lru); 1148 __free_slab(page->slab, page); 1149} 1150 1151static void free_slab(struct kmem_cache *s, struct page *page) 1152{ 1153 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1154 /* 1155 * RCU free overloads the RCU head over the LRU 1156 */ 1157 struct rcu_head *head = (void *)&page->lru; 1158 1159 call_rcu(head, rcu_free_slab); 1160 } else 1161 __free_slab(s, page); 1162} 1163 1164static void discard_slab(struct kmem_cache *s, struct page *page) 1165{ 1166 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1167 1168 atomic_long_dec(&n->nr_slabs); 1169 reset_page_mapcount(page); 1170 __ClearPageSlab(page); 1171 free_slab(s, page); 1172} 1173 1174/* 1175 * Per slab locking using the pagelock 1176 */ 1177static __always_inline void slab_lock(struct page *page) 1178{ 1179 bit_spin_lock(PG_locked, &page->flags); 1180} 1181 1182static __always_inline void slab_unlock(struct page *page) 1183{ 1184 bit_spin_unlock(PG_locked, &page->flags); 1185} 1186 1187static __always_inline int slab_trylock(struct page *page) 1188{ 1189 int rc = 1; 1190 1191 rc = bit_spin_trylock(PG_locked, &page->flags); 1192 return rc; 1193} 1194 1195/* 1196 * Management of partially allocated slabs 1197 */ 1198static void add_partial_tail(struct kmem_cache_node *n, struct page *page) 1199{ 1200 spin_lock(&n->list_lock); 1201 n->nr_partial++; 1202 list_add_tail(&page->lru, &n->partial); 1203 spin_unlock(&n->list_lock); 1204} 1205 1206static void add_partial(struct kmem_cache_node *n, struct page *page) 1207{ 1208 spin_lock(&n->list_lock); 1209 n->nr_partial++; 1210 list_add(&page->lru, &n->partial); 1211 spin_unlock(&n->list_lock); 1212} 1213 1214static void remove_partial(struct kmem_cache *s, 1215 struct page *page) 1216{ 1217 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1218 1219 spin_lock(&n->list_lock); 1220 list_del(&page->lru); 1221 n->nr_partial--; 1222 spin_unlock(&n->list_lock); 1223} 1224 1225/* 1226 * Lock slab and remove from the partial list. 1227 * 1228 * Must hold list_lock. 1229 */ 1230static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page) 1231{ 1232 if (slab_trylock(page)) { 1233 list_del(&page->lru); 1234 n->nr_partial--; 1235 SetSlabFrozen(page); 1236 return 1; 1237 } 1238 return 0; 1239} 1240 1241/* 1242 * Try to allocate a partial slab from a specific node. 1243 */ 1244static struct page *get_partial_node(struct kmem_cache_node *n) 1245{ 1246 struct page *page; 1247 1248 /* 1249 * Racy check. If we mistakenly see no partial slabs then we 1250 * just allocate an empty slab. If we mistakenly try to get a 1251 * partial slab and there is none available then get_partials() 1252 * will return NULL. 1253 */ 1254 if (!n || !n->nr_partial) 1255 return NULL; 1256 1257 spin_lock(&n->list_lock); 1258 list_for_each_entry(page, &n->partial, lru) 1259 if (lock_and_freeze_slab(n, page)) 1260 goto out; 1261 page = NULL; 1262out: 1263 spin_unlock(&n->list_lock); 1264 return page; 1265} 1266 1267/* 1268 * Get a page from somewhere. Search in increasing NUMA distances. 1269 */ 1270static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1271{ 1272#ifdef CONFIG_NUMA 1273 struct zonelist *zonelist; 1274 struct zone **z; 1275 struct page *page; 1276 1277 /* 1278 * The defrag ratio allows a configuration of the tradeoffs between 1279 * inter node defragmentation and node local allocations. A lower 1280 * defrag_ratio increases the tendency to do local allocations 1281 * instead of attempting to obtain partial slabs from other nodes. 1282 * 1283 * If the defrag_ratio is set to 0 then kmalloc() always 1284 * returns node local objects. If the ratio is higher then kmalloc() 1285 * may return off node objects because partial slabs are obtained 1286 * from other nodes and filled up. 1287 * 1288 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes 1289 * defrag_ratio = 1000) then every (well almost) allocation will 1290 * first attempt to defrag slab caches on other nodes. This means 1291 * scanning over all nodes to look for partial slabs which may be 1292 * expensive if we do it every time we are trying to find a slab 1293 * with available objects. 1294 */ 1295 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio) 1296 return NULL; 1297 1298 zonelist = &NODE_DATA(slab_node(current->mempolicy)) 1299 ->node_zonelists[gfp_zone(flags)]; 1300 for (z = zonelist->zones; *z; z++) { 1301 struct kmem_cache_node *n; 1302 1303 n = get_node(s, zone_to_nid(*z)); 1304 1305 if (n && cpuset_zone_allowed_hardwall(*z, flags) && 1306 n->nr_partial > MIN_PARTIAL) { 1307 page = get_partial_node(n); 1308 if (page) 1309 return page; 1310 } 1311 } 1312#endif 1313 return NULL; 1314} 1315 1316/* 1317 * Get a partial page, lock it and return it. 1318 */ 1319static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1320{ 1321 struct page *page; 1322 int searchnode = (node == -1) ? numa_node_id() : node; 1323 1324 page = get_partial_node(get_node(s, searchnode)); 1325 if (page || (flags & __GFP_THISNODE)) 1326 return page; 1327 1328 return get_any_partial(s, flags); 1329} 1330 1331/* 1332 * Move a page back to the lists. 1333 * 1334 * Must be called with the slab lock held. 1335 * 1336 * On exit the slab lock will have been dropped. 1337 */ 1338static void unfreeze_slab(struct kmem_cache *s, struct page *page) 1339{ 1340 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1341 1342 ClearSlabFrozen(page); 1343 if (page->inuse) { 1344 1345 if (page->freelist) 1346 add_partial(n, page); 1347 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER)) 1348 add_full(n, page); 1349 slab_unlock(page); 1350 1351 } else { 1352 if (n->nr_partial < MIN_PARTIAL) { 1353 /* 1354 * Adding an empty slab to the partial slabs in order 1355 * to avoid page allocator overhead. This slab needs 1356 * to come after the other slabs with objects in 1357 * order to fill them up. That way the size of the 1358 * partial list stays small. kmem_cache_shrink can 1359 * reclaim empty slabs from the partial list. 1360 */ 1361 add_partial_tail(n, page); 1362 slab_unlock(page); 1363 } else { 1364 slab_unlock(page); 1365 discard_slab(s, page); 1366 } 1367 } 1368} 1369 1370/* 1371 * Remove the cpu slab 1372 */ 1373static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1374{ 1375 struct page *page = c->page; 1376 /* 1377 * Merge cpu freelist into freelist. Typically we get here 1378 * because both freelists are empty. So this is unlikely 1379 * to occur. 1380 */ 1381 while (unlikely(c->freelist)) { 1382 void **object; 1383 1384 /* Retrieve object from cpu_freelist */ 1385 object = c->freelist; 1386 c->freelist = c->freelist[c->offset]; 1387 1388 /* And put onto the regular freelist */ 1389 object[c->offset] = page->freelist; 1390 page->freelist = object; 1391 page->inuse--; 1392 } 1393 c->page = NULL; 1394 unfreeze_slab(s, page); 1395} 1396 1397static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1398{ 1399 slab_lock(c->page); 1400 deactivate_slab(s, c); 1401} 1402 1403/* 1404 * Flush cpu slab. 1405 * Called from IPI handler with interrupts disabled. 1406 */ 1407static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1408{ 1409 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1410 1411 if (likely(c && c->page)) 1412 flush_slab(s, c); 1413} 1414 1415static void flush_cpu_slab(void *d) 1416{ 1417 struct kmem_cache *s = d; 1418 1419 __flush_cpu_slab(s, smp_processor_id()); 1420} 1421 1422static void flush_all(struct kmem_cache *s) 1423{ 1424#ifdef CONFIG_SMP 1425 on_each_cpu(flush_cpu_slab, s, 1, 1); 1426#else 1427 unsigned long flags; 1428 1429 local_irq_save(flags); 1430 flush_cpu_slab(s); 1431 local_irq_restore(flags); 1432#endif 1433} 1434 1435/* 1436 * Check if the objects in a per cpu structure fit numa 1437 * locality expectations. 1438 */ 1439static inline int node_match(struct kmem_cache_cpu *c, int node) 1440{ 1441#ifdef CONFIG_NUMA 1442 if (node != -1 && c->node != node) 1443 return 0; 1444#endif 1445 return 1; 1446} 1447 1448/* 1449 * Slow path. The lockless freelist is empty or we need to perform 1450 * debugging duties. 1451 * 1452 * Interrupts are disabled. 1453 * 1454 * Processing is still very fast if new objects have been freed to the 1455 * regular freelist. In that case we simply take over the regular freelist 1456 * as the lockless freelist and zap the regular freelist. 1457 * 1458 * If that is not working then we fall back to the partial lists. We take the 1459 * first element of the freelist as the object to allocate now and move the 1460 * rest of the freelist to the lockless freelist. 1461 * 1462 * And if we were unable to get a new slab from the partial slab lists then 1463 * we need to allocate a new slab. This is slowest path since we may sleep. 1464 */ 1465static void *__slab_alloc(struct kmem_cache *s, 1466 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c) 1467{ 1468 void **object; 1469 struct page *new; 1470 1471 if (!c->page) 1472 goto new_slab; 1473 1474 slab_lock(c->page); 1475 if (unlikely(!node_match(c, node))) 1476 goto another_slab; 1477load_freelist: 1478 object = c->page->freelist; 1479 if (unlikely(!object)) 1480 goto another_slab; 1481 if (unlikely(SlabDebug(c->page))) 1482 goto debug; 1483 1484 object = c->page->freelist; 1485 c->freelist = object[c->offset]; 1486 c->page->inuse = s->objects; 1487 c->page->freelist = NULL; 1488 c->node = page_to_nid(c->page); 1489 slab_unlock(c->page); 1490 return object; 1491 1492another_slab: 1493 deactivate_slab(s, c); 1494 1495new_slab: 1496 new = get_partial(s, gfpflags, node); 1497 if (new) { 1498 c->page = new; 1499 goto load_freelist; 1500 } 1501 1502 if (gfpflags & __GFP_WAIT) 1503 local_irq_enable(); 1504 1505 new = new_slab(s, gfpflags, node); 1506 1507 if (gfpflags & __GFP_WAIT) 1508 local_irq_disable(); 1509 1510 if (new) { 1511 c = get_cpu_slab(s, smp_processor_id()); 1512 if (c->page) 1513 flush_slab(s, c); 1514 slab_lock(new); 1515 SetSlabFrozen(new); 1516 c->page = new; 1517 goto load_freelist; 1518 } 1519 return NULL; 1520debug: 1521 object = c->page->freelist; 1522 if (!alloc_debug_processing(s, c->page, object, addr)) 1523 goto another_slab; 1524 1525 c->page->inuse++; 1526 c->page->freelist = object[c->offset]; 1527 c->node = -1; 1528 slab_unlock(c->page); 1529 return object; 1530} 1531 1532/* 1533 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1534 * have the fastpath folded into their functions. So no function call 1535 * overhead for requests that can be satisfied on the fastpath. 1536 * 1537 * The fastpath works by first checking if the lockless freelist can be used. 1538 * If not then __slab_alloc is called for slow processing. 1539 * 1540 * Otherwise we can simply pick the next object from the lockless free list. 1541 */ 1542static void __always_inline *slab_alloc(struct kmem_cache *s, 1543 gfp_t gfpflags, int node, void *addr) 1544{ 1545 void **object; 1546 unsigned long flags; 1547 struct kmem_cache_cpu *c; 1548 1549 local_irq_save(flags); 1550 c = get_cpu_slab(s, smp_processor_id()); 1551 if (unlikely(!c->freelist || !node_match(c, node))) 1552 1553 object = __slab_alloc(s, gfpflags, node, addr, c); 1554 1555 else { 1556 object = c->freelist; 1557 c->freelist = object[c->offset]; 1558 } 1559 local_irq_restore(flags); 1560 1561 if (unlikely((gfpflags & __GFP_ZERO) && object)) 1562 memset(object, 0, c->objsize); 1563 1564 return object; 1565} 1566 1567void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1568{ 1569 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); 1570} 1571EXPORT_SYMBOL(kmem_cache_alloc); 1572 1573#ifdef CONFIG_NUMA 1574void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1575{ 1576 return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); 1577} 1578EXPORT_SYMBOL(kmem_cache_alloc_node); 1579#endif 1580 1581/* 1582 * Slow patch handling. This may still be called frequently since objects 1583 * have a longer lifetime than the cpu slabs in most processing loads. 1584 * 1585 * So we still attempt to reduce cache line usage. Just take the slab 1586 * lock and free the item. If there is no additional partial page 1587 * handling required then we can return immediately. 1588 */ 1589static void __slab_free(struct kmem_cache *s, struct page *page, 1590 void *x, void *addr, unsigned int offset) 1591{ 1592 void *prior; 1593 void **object = (void *)x; 1594 1595 slab_lock(page); 1596 1597 if (unlikely(SlabDebug(page))) 1598 goto debug; 1599checks_ok: 1600 prior = object[offset] = page->freelist; 1601 page->freelist = object; 1602 page->inuse--; 1603 1604 if (unlikely(SlabFrozen(page))) 1605 goto out_unlock; 1606 1607 if (unlikely(!page->inuse)) 1608 goto slab_empty; 1609 1610 /* 1611 * Objects left in the slab. If it 1612 * was not on the partial list before 1613 * then add it. 1614 */ 1615 if (unlikely(!prior)) 1616 add_partial(get_node(s, page_to_nid(page)), page); 1617 1618out_unlock: 1619 slab_unlock(page); 1620 return; 1621 1622slab_empty: 1623 if (prior) 1624 /* 1625 * Slab still on the partial list. 1626 */ 1627 remove_partial(s, page); 1628 1629 slab_unlock(page); 1630 discard_slab(s, page); 1631 return; 1632 1633debug: 1634 if (!free_debug_processing(s, page, x, addr)) 1635 goto out_unlock; 1636 goto checks_ok; 1637} 1638 1639/* 1640 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1641 * can perform fastpath freeing without additional function calls. 1642 * 1643 * The fastpath is only possible if we are freeing to the current cpu slab 1644 * of this processor. This typically the case if we have just allocated 1645 * the item before. 1646 * 1647 * If fastpath is not possible then fall back to __slab_free where we deal 1648 * with all sorts of special processing. 1649 */ 1650static void __always_inline slab_free(struct kmem_cache *s, 1651 struct page *page, void *x, void *addr) 1652{ 1653 void **object = (void *)x; 1654 unsigned long flags; 1655 struct kmem_cache_cpu *c; 1656 1657 local_irq_save(flags); 1658 debug_check_no_locks_freed(object, s->objsize); 1659 c = get_cpu_slab(s, smp_processor_id()); 1660 if (likely(page == c->page && c->node >= 0)) { 1661 object[c->offset] = c->freelist; 1662 c->freelist = object; 1663 } else 1664 __slab_free(s, page, x, addr, c->offset); 1665 1666 local_irq_restore(flags); 1667} 1668 1669void kmem_cache_free(struct kmem_cache *s, void *x) 1670{ 1671 struct page *page; 1672 1673 page = virt_to_head_page(x); 1674 1675 slab_free(s, page, x, __builtin_return_address(0)); 1676} 1677EXPORT_SYMBOL(kmem_cache_free); 1678 1679/* Figure out on which slab object the object resides */ 1680static struct page *get_object_page(const void *x) 1681{ 1682 struct page *page = virt_to_head_page(x); 1683 1684 if (!PageSlab(page)) 1685 return NULL; 1686 1687 return page; 1688} 1689 1690/* 1691 * Object placement in a slab is made very easy because we always start at 1692 * offset 0. If we tune the size of the object to the alignment then we can 1693 * get the required alignment by putting one properly sized object after 1694 * another. 1695 * 1696 * Notice that the allocation order determines the sizes of the per cpu 1697 * caches. Each processor has always one slab available for allocations. 1698 * Increasing the allocation order reduces the number of times that slabs 1699 * must be moved on and off the partial lists and is therefore a factor in 1700 * locking overhead. 1701 */ 1702 1703/* 1704 * Mininum / Maximum order of slab pages. This influences locking overhead 1705 * and slab fragmentation. A higher order reduces the number of partial slabs 1706 * and increases the number of allocations possible without having to 1707 * take the list_lock. 1708 */ 1709static int slub_min_order; 1710static int slub_max_order = DEFAULT_MAX_ORDER; 1711static int slub_min_objects = DEFAULT_MIN_OBJECTS; 1712 1713/* 1714 * Merge control. If this is set then no merging of slab caches will occur. 1715 * (Could be removed. This was introduced to pacify the merge skeptics.) 1716 */ 1717static int slub_nomerge; 1718 1719/* 1720 * Calculate the order of allocation given an slab object size. 1721 * 1722 * The order of allocation has significant impact on performance and other 1723 * system components. Generally order 0 allocations should be preferred since 1724 * order 0 does not cause fragmentation in the page allocator. Larger objects 1725 * be problematic to put into order 0 slabs because there may be too much 1726 * unused space left. We go to a higher order if more than 1/8th of the slab 1727 * would be wasted. 1728 * 1729 * In order to reach satisfactory performance we must ensure that a minimum 1730 * number of objects is in one slab. Otherwise we may generate too much 1731 * activity on the partial lists which requires taking the list_lock. This is 1732 * less a concern for large slabs though which are rarely used. 1733 * 1734 * slub_max_order specifies the order where we begin to stop considering the 1735 * number of objects in a slab as critical. If we reach slub_max_order then 1736 * we try to keep the page order as low as possible. So we accept more waste 1737 * of space in favor of a small page order. 1738 * 1739 * Higher order allocations also allow the placement of more objects in a 1740 * slab and thereby reduce object handling overhead. If the user has 1741 * requested a higher mininum order then we start with that one instead of 1742 * the smallest order which will fit the object. 1743 */ 1744static inline int slab_order(int size, int min_objects, 1745 int max_order, int fract_leftover) 1746{ 1747 int order; 1748 int rem; 1749 int min_order = slub_min_order; 1750 1751 for (order = max(min_order, 1752 fls(min_objects * size - 1) - PAGE_SHIFT); 1753 order <= max_order; order++) { 1754 1755 unsigned long slab_size = PAGE_SIZE << order; 1756 1757 if (slab_size < min_objects * size) 1758 continue; 1759 1760 rem = slab_size % size; 1761 1762 if (rem <= slab_size / fract_leftover) 1763 break; 1764 1765 } 1766 1767 return order; 1768} 1769 1770static inline int calculate_order(int size) 1771{ 1772 int order; 1773 int min_objects; 1774 int fraction; 1775 1776 /* 1777 * Attempt to find best configuration for a slab. This 1778 * works by first attempting to generate a layout with 1779 * the best configuration and backing off gradually. 1780 * 1781 * First we reduce the acceptable waste in a slab. Then 1782 * we reduce the minimum objects required in a slab. 1783 */ 1784 min_objects = slub_min_objects; 1785 while (min_objects > 1) { 1786 fraction = 8; 1787 while (fraction >= 4) { 1788 order = slab_order(size, min_objects, 1789 slub_max_order, fraction); 1790 if (order <= slub_max_order) 1791 return order; 1792 fraction /= 2; 1793 } 1794 min_objects /= 2; 1795 } 1796 1797 /* 1798 * We were unable to place multiple objects in a slab. Now 1799 * lets see if we can place a single object there. 1800 */ 1801 order = slab_order(size, 1, slub_max_order, 1); 1802 if (order <= slub_max_order) 1803 return order; 1804 1805 /* 1806 * Doh this slab cannot be placed using slub_max_order. 1807 */ 1808 order = slab_order(size, 1, MAX_ORDER, 1); 1809 if (order <= MAX_ORDER) 1810 return order; 1811 return -ENOSYS; 1812} 1813 1814/* 1815 * Figure out what the alignment of the objects will be. 1816 */ 1817static unsigned long calculate_alignment(unsigned long flags, 1818 unsigned long align, unsigned long size) 1819{ 1820 /* 1821 * If the user wants hardware cache aligned objects then 1822 * follow that suggestion if the object is sufficiently 1823 * large. 1824 * 1825 * The hardware cache alignment cannot override the 1826 * specified alignment though. If that is greater 1827 * then use it. 1828 */ 1829 if ((flags & SLAB_HWCACHE_ALIGN) && 1830 size > cache_line_size() / 2) 1831 return max_t(unsigned long, align, cache_line_size()); 1832 1833 if (align < ARCH_SLAB_MINALIGN) 1834 return ARCH_SLAB_MINALIGN; 1835 1836 return ALIGN(align, sizeof(void *)); 1837} 1838 1839static void init_kmem_cache_cpu(struct kmem_cache *s, 1840 struct kmem_cache_cpu *c) 1841{ 1842 c->page = NULL; 1843 c->freelist = NULL; 1844 c->node = 0; 1845 c->offset = s->offset / sizeof(void *); 1846 c->objsize = s->objsize; 1847} 1848 1849static void init_kmem_cache_node(struct kmem_cache_node *n) 1850{ 1851 n->nr_partial = 0; 1852 atomic_long_set(&n->nr_slabs, 0); 1853 spin_lock_init(&n->list_lock); 1854 INIT_LIST_HEAD(&n->partial); 1855#ifdef CONFIG_SLUB_DEBUG 1856 INIT_LIST_HEAD(&n->full); 1857#endif 1858} 1859 1860#ifdef CONFIG_SMP 1861/* 1862 * Per cpu array for per cpu structures. 1863 * 1864 * The per cpu array places all kmem_cache_cpu structures from one processor 1865 * close together meaning that it becomes possible that multiple per cpu 1866 * structures are contained in one cacheline. This may be particularly 1867 * beneficial for the kmalloc caches. 1868 * 1869 * A desktop system typically has around 60-80 slabs. With 100 here we are 1870 * likely able to get per cpu structures for all caches from the array defined 1871 * here. We must be able to cover all kmalloc caches during bootstrap. 1872 * 1873 * If the per cpu array is exhausted then fall back to kmalloc 1874 * of individual cachelines. No sharing is possible then. 1875 */ 1876#define NR_KMEM_CACHE_CPU 100 1877 1878static DEFINE_PER_CPU(struct kmem_cache_cpu, 1879 kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; 1880 1881static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); 1882static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; 1883 1884static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, 1885 int cpu, gfp_t flags) 1886{ 1887 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); 1888 1889 if (c) 1890 per_cpu(kmem_cache_cpu_free, cpu) = 1891 (void *)c->freelist; 1892 else { 1893 /* Table overflow: So allocate ourselves */ 1894 c = kmalloc_node( 1895 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), 1896 flags, cpu_to_node(cpu)); 1897 if (!c) 1898 return NULL; 1899 } 1900 1901 init_kmem_cache_cpu(s, c); 1902 return c; 1903} 1904 1905static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) 1906{ 1907 if (c < per_cpu(kmem_cache_cpu, cpu) || 1908 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { 1909 kfree(c); 1910 return; 1911 } 1912 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); 1913 per_cpu(kmem_cache_cpu_free, cpu) = c; 1914} 1915 1916static void free_kmem_cache_cpus(struct kmem_cache *s) 1917{ 1918 int cpu; 1919 1920 for_each_online_cpu(cpu) { 1921 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1922 1923 if (c) { 1924 s->cpu_slab[cpu] = NULL; 1925 free_kmem_cache_cpu(c, cpu); 1926 } 1927 } 1928} 1929 1930static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 1931{ 1932 int cpu; 1933 1934 for_each_online_cpu(cpu) { 1935 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1936 1937 if (c) 1938 continue; 1939 1940 c = alloc_kmem_cache_cpu(s, cpu, flags); 1941 if (!c) { 1942 free_kmem_cache_cpus(s); 1943 return 0; 1944 } 1945 s->cpu_slab[cpu] = c; 1946 } 1947 return 1; 1948} 1949 1950/* 1951 * Initialize the per cpu array. 1952 */ 1953static void init_alloc_cpu_cpu(int cpu) 1954{ 1955 int i; 1956 1957 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) 1958 return; 1959 1960 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) 1961 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); 1962 1963 cpu_set(cpu, kmem_cach_cpu_free_init_once); 1964} 1965 1966static void __init init_alloc_cpu(void) 1967{ 1968 int cpu; 1969 1970 for_each_online_cpu(cpu) 1971 init_alloc_cpu_cpu(cpu); 1972 } 1973 1974#else 1975static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} 1976static inline void init_alloc_cpu(void) {} 1977 1978static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 1979{ 1980 init_kmem_cache_cpu(s, &s->cpu_slab); 1981 return 1; 1982} 1983#endif 1984 1985#ifdef CONFIG_NUMA 1986/* 1987 * No kmalloc_node yet so do it by hand. We know that this is the first 1988 * slab on the node for this slabcache. There are no concurrent accesses 1989 * possible. 1990 * 1991 * Note that this function only works on the kmalloc_node_cache 1992 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 1993 * memory on a fresh node that has no slab structures yet. 1994 */ 1995static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags, 1996 int node) 1997{ 1998 struct page *page; 1999 struct kmem_cache_node *n; 2000 2001 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2002 2003 page = new_slab(kmalloc_caches, gfpflags, node); 2004 2005 BUG_ON(!page); 2006 if (page_to_nid(page) != node) { 2007 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2008 "node %d\n", node); 2009 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2010 "in order to be able to continue\n"); 2011 } 2012 2013 n = page->freelist; 2014 BUG_ON(!n); 2015 page->freelist = get_freepointer(kmalloc_caches, n); 2016 page->inuse++; 2017 kmalloc_caches->node[node] = n; 2018#ifdef CONFIG_SLUB_DEBUG 2019 init_object(kmalloc_caches, n, 1); 2020 init_tracking(kmalloc_caches, n); 2021#endif 2022 init_kmem_cache_node(n); 2023 atomic_long_inc(&n->nr_slabs); 2024 add_partial(n, page); 2025 return n; 2026} 2027 2028static void free_kmem_cache_nodes(struct kmem_cache *s) 2029{ 2030 int node; 2031 2032 for_each_node_state(node, N_NORMAL_MEMORY) { 2033 struct kmem_cache_node *n = s->node[node]; 2034 if (n && n != &s->local_node) 2035 kmem_cache_free(kmalloc_caches, n); 2036 s->node[node] = NULL; 2037 } 2038} 2039 2040static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2041{ 2042 int node; 2043 int local_node; 2044 2045 if (slab_state >= UP) 2046 local_node = page_to_nid(virt_to_page(s)); 2047 else 2048 local_node = 0; 2049 2050 for_each_node_state(node, N_NORMAL_MEMORY) { 2051 struct kmem_cache_node *n; 2052 2053 if (local_node == node) 2054 n = &s->local_node; 2055 else { 2056 if (slab_state == DOWN) { 2057 n = early_kmem_cache_node_alloc(gfpflags, 2058 node); 2059 continue; 2060 } 2061 n = kmem_cache_alloc_node(kmalloc_caches, 2062 gfpflags, node); 2063 2064 if (!n) { 2065 free_kmem_cache_nodes(s); 2066 return 0; 2067 } 2068 2069 } 2070 s->node[node] = n; 2071 init_kmem_cache_node(n); 2072 } 2073 return 1; 2074} 2075#else 2076static void free_kmem_cache_nodes(struct kmem_cache *s) 2077{ 2078} 2079 2080static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2081{ 2082 init_kmem_cache_node(&s->local_node); 2083 return 1; 2084} 2085#endif 2086 2087/* 2088 * calculate_sizes() determines the order and the distribution of data within 2089 * a slab object. 2090 */ 2091static int calculate_sizes(struct kmem_cache *s) 2092{ 2093 unsigned long flags = s->flags; 2094 unsigned long size = s->objsize; 2095 unsigned long align = s->align; 2096 2097 /* 2098 * Determine if we can poison the object itself. If the user of 2099 * the slab may touch the object after free or before allocation 2100 * then we should never poison the object itself. 2101 */ 2102 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2103 !s->ctor) 2104 s->flags |= __OBJECT_POISON; 2105 else 2106 s->flags &= ~__OBJECT_POISON; 2107 2108 /* 2109 * Round up object size to the next word boundary. We can only 2110 * place the free pointer at word boundaries and this determines 2111 * the possible location of the free pointer. 2112 */ 2113 size = ALIGN(size, sizeof(void *)); 2114 2115#ifdef CONFIG_SLUB_DEBUG 2116 /* 2117 * If we are Redzoning then check if there is some space between the 2118 * end of the object and the free pointer. If not then add an 2119 * additional word to have some bytes to store Redzone information. 2120 */ 2121 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2122 size += sizeof(void *); 2123#endif 2124 2125 /* 2126 * With that we have determined the number of bytes in actual use 2127 * by the object. This is the potential offset to the free pointer. 2128 */ 2129 s->inuse = size; 2130 2131 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2132 s->ctor)) { 2133 /* 2134 * Relocate free pointer after the object if it is not 2135 * permitted to overwrite the first word of the object on 2136 * kmem_cache_free. 2137 * 2138 * This is the case if we do RCU, have a constructor or 2139 * destructor or are poisoning the objects. 2140 */ 2141 s->offset = size; 2142 size += sizeof(void *); 2143 } 2144 2145#ifdef CONFIG_SLUB_DEBUG 2146 if (flags & SLAB_STORE_USER) 2147 /* 2148 * Need to store information about allocs and frees after 2149 * the object. 2150 */ 2151 size += 2 * sizeof(struct track); 2152 2153 if (flags & SLAB_RED_ZONE) 2154 /* 2155 * Add some empty padding so that we can catch 2156 * overwrites from earlier objects rather than let 2157 * tracking information or the free pointer be 2158 * corrupted if an user writes before the start 2159 * of the object. 2160 */ 2161 size += sizeof(void *); 2162#endif 2163 2164 /* 2165 * Determine the alignment based on various parameters that the 2166 * user specified and the dynamic determination of cache line size 2167 * on bootup. 2168 */ 2169 align = calculate_alignment(flags, align, s->objsize); 2170 2171 /* 2172 * SLUB stores one object immediately after another beginning from 2173 * offset 0. In order to align the objects we have to simply size 2174 * each object to conform to the alignment. 2175 */ 2176 size = ALIGN(size, align); 2177 s->size = size; 2178 2179 s->order = calculate_order(size); 2180 if (s->order < 0) 2181 return 0; 2182 2183 /* 2184 * Determine the number of objects per slab 2185 */ 2186 s->objects = (PAGE_SIZE << s->order) / size; 2187 2188 return !!s->objects; 2189 2190} 2191 2192static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2193 const char *name, size_t size, 2194 size_t align, unsigned long flags, 2195 void (*ctor)(struct kmem_cache *, void *)) 2196{ 2197 memset(s, 0, kmem_size); 2198 s->name = name; 2199 s->ctor = ctor; 2200 s->objsize = size; 2201 s->align = align; 2202 s->flags = kmem_cache_flags(size, flags, name, ctor); 2203 2204 if (!calculate_sizes(s)) 2205 goto error; 2206 2207 s->refcount = 1; 2208#ifdef CONFIG_NUMA 2209 s->defrag_ratio = 100; 2210#endif 2211 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2212 goto error; 2213 2214 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2215 return 1; 2216 free_kmem_cache_nodes(s); 2217error: 2218 if (flags & SLAB_PANIC) 2219 panic("Cannot create slab %s size=%lu realsize=%u " 2220 "order=%u offset=%u flags=%lx\n", 2221 s->name, (unsigned long)size, s->size, s->order, 2222 s->offset, flags); 2223 return 0; 2224} 2225 2226/* 2227 * Check if a given pointer is valid 2228 */ 2229int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2230{ 2231 struct page * page; 2232 2233 page = get_object_page(object); 2234 2235 if (!page || s != page->slab) 2236 /* No slab or wrong slab */ 2237 return 0; 2238 2239 if (!check_valid_pointer(s, page, object)) 2240 return 0; 2241 2242 /* 2243 * We could also check if the object is on the slabs freelist. 2244 * But this would be too expensive and it seems that the main 2245 * purpose of kmem_ptr_valid is to check if the object belongs 2246 * to a certain slab. 2247 */ 2248 return 1; 2249} 2250EXPORT_SYMBOL(kmem_ptr_validate); 2251 2252/* 2253 * Determine the size of a slab object 2254 */ 2255unsigned int kmem_cache_size(struct kmem_cache *s) 2256{ 2257 return s->objsize; 2258} 2259EXPORT_SYMBOL(kmem_cache_size); 2260 2261const char *kmem_cache_name(struct kmem_cache *s) 2262{ 2263 return s->name; 2264} 2265EXPORT_SYMBOL(kmem_cache_name); 2266 2267/* 2268 * Attempt to free all slabs on a node. Return the number of slabs we 2269 * were unable to free. 2270 */ 2271static int free_list(struct kmem_cache *s, struct kmem_cache_node *n, 2272 struct list_head *list) 2273{ 2274 int slabs_inuse = 0; 2275 unsigned long flags; 2276 struct page *page, *h; 2277 2278 spin_lock_irqsave(&n->list_lock, flags); 2279 list_for_each_entry_safe(page, h, list, lru) 2280 if (!page->inuse) { 2281 list_del(&page->lru); 2282 discard_slab(s, page); 2283 } else 2284 slabs_inuse++; 2285 spin_unlock_irqrestore(&n->list_lock, flags); 2286 return slabs_inuse; 2287} 2288 2289/* 2290 * Release all resources used by a slab cache. 2291 */ 2292static inline int kmem_cache_close(struct kmem_cache *s) 2293{ 2294 int node; 2295 2296 flush_all(s); 2297 2298 /* Attempt to free all objects */ 2299 free_kmem_cache_cpus(s); 2300 for_each_node_state(node, N_NORMAL_MEMORY) { 2301 struct kmem_cache_node *n = get_node(s, node); 2302 2303 n->nr_partial -= free_list(s, n, &n->partial); 2304 if (atomic_long_read(&n->nr_slabs)) 2305 return 1; 2306 } 2307 free_kmem_cache_nodes(s); 2308 return 0; 2309} 2310 2311/* 2312 * Close a cache and release the kmem_cache structure 2313 * (must be used for caches created using kmem_cache_create) 2314 */ 2315void kmem_cache_destroy(struct kmem_cache *s) 2316{ 2317 down_write(&slub_lock); 2318 s->refcount--; 2319 if (!s->refcount) { 2320 list_del(&s->list); 2321 up_write(&slub_lock); 2322 if (kmem_cache_close(s)) 2323 WARN_ON(1); 2324 sysfs_slab_remove(s); 2325 kfree(s); 2326 } else 2327 up_write(&slub_lock); 2328} 2329EXPORT_SYMBOL(kmem_cache_destroy); 2330 2331/******************************************************************** 2332 * Kmalloc subsystem 2333 *******************************************************************/ 2334 2335struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned; 2336EXPORT_SYMBOL(kmalloc_caches); 2337 2338#ifdef CONFIG_ZONE_DMA 2339static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT]; 2340#endif 2341 2342static int __init setup_slub_min_order(char *str) 2343{ 2344 get_option (&str, &slub_min_order); 2345 2346 return 1; 2347} 2348 2349__setup("slub_min_order=", setup_slub_min_order); 2350 2351static int __init setup_slub_max_order(char *str) 2352{ 2353 get_option (&str, &slub_max_order); 2354 2355 return 1; 2356} 2357 2358__setup("slub_max_order=", setup_slub_max_order); 2359 2360static int __init setup_slub_min_objects(char *str) 2361{ 2362 get_option (&str, &slub_min_objects); 2363 2364 return 1; 2365} 2366 2367__setup("slub_min_objects=", setup_slub_min_objects); 2368 2369static int __init setup_slub_nomerge(char *str) 2370{ 2371 slub_nomerge = 1; 2372 return 1; 2373} 2374 2375__setup("slub_nomerge", setup_slub_nomerge); 2376 2377static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2378 const char *name, int size, gfp_t gfp_flags) 2379{ 2380 unsigned int flags = 0; 2381 2382 if (gfp_flags & SLUB_DMA) 2383 flags = SLAB_CACHE_DMA; 2384 2385 down_write(&slub_lock); 2386 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2387 flags, NULL)) 2388 goto panic; 2389 2390 list_add(&s->list, &slab_caches); 2391 up_write(&slub_lock); 2392 if (sysfs_slab_add(s)) 2393 goto panic; 2394 return s; 2395 2396panic: 2397 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2398} 2399 2400#ifdef CONFIG_ZONE_DMA 2401 2402static void sysfs_add_func(struct work_struct *w) 2403{ 2404 struct kmem_cache *s; 2405 2406 down_write(&slub_lock); 2407 list_for_each_entry(s, &slab_caches, list) { 2408 if (s->flags & __SYSFS_ADD_DEFERRED) { 2409 s->flags &= ~__SYSFS_ADD_DEFERRED; 2410 sysfs_slab_add(s); 2411 } 2412 } 2413 up_write(&slub_lock); 2414} 2415 2416static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2417 2418static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2419{ 2420 struct kmem_cache *s; 2421 char *text; 2422 size_t realsize; 2423 2424 s = kmalloc_caches_dma[index]; 2425 if (s) 2426 return s; 2427 2428 /* Dynamically create dma cache */ 2429 if (flags & __GFP_WAIT) 2430 down_write(&slub_lock); 2431 else { 2432 if (!down_write_trylock(&slub_lock)) 2433 goto out; 2434 } 2435 2436 if (kmalloc_caches_dma[index]) 2437 goto unlock_out; 2438 2439 realsize = kmalloc_caches[index].objsize; 2440 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize), 2441 s = kmalloc(kmem_size, flags & ~SLUB_DMA); 2442 2443 if (!s || !text || !kmem_cache_open(s, flags, text, 2444 realsize, ARCH_KMALLOC_MINALIGN, 2445 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { 2446 kfree(s); 2447 kfree(text); 2448 goto unlock_out; 2449 } 2450 2451 list_add(&s->list, &slab_caches); 2452 kmalloc_caches_dma[index] = s; 2453 2454 schedule_work(&sysfs_add_work); 2455 2456unlock_out: 2457 up_write(&slub_lock); 2458out: 2459 return kmalloc_caches_dma[index]; 2460} 2461#endif 2462 2463/* 2464 * Conversion table for small slabs sizes / 8 to the index in the 2465 * kmalloc array. This is necessary for slabs < 192 since we have non power 2466 * of two cache sizes there. The size of larger slabs can be determined using 2467 * fls. 2468 */ 2469static s8 size_index[24] = { 2470 3, /* 8 */ 2471 4, /* 16 */ 2472 5, /* 24 */ 2473 5, /* 32 */ 2474 6, /* 40 */ 2475 6, /* 48 */ 2476 6, /* 56 */ 2477 6, /* 64 */ 2478 1, /* 72 */ 2479 1, /* 80 */ 2480 1, /* 88 */ 2481 1, /* 96 */ 2482 7, /* 104 */ 2483 7, /* 112 */ 2484 7, /* 120 */ 2485 7, /* 128 */ 2486 2, /* 136 */ 2487 2, /* 144 */ 2488 2, /* 152 */ 2489 2, /* 160 */ 2490 2, /* 168 */ 2491 2, /* 176 */ 2492 2, /* 184 */ 2493 2 /* 192 */ 2494}; 2495 2496static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2497{ 2498 int index; 2499 2500 if (size <= 192) { 2501 if (!size) 2502 return ZERO_SIZE_PTR; 2503 2504 index = size_index[(size - 1) / 8]; 2505 } else 2506 index = fls(size - 1); 2507 2508#ifdef CONFIG_ZONE_DMA 2509 if (unlikely((flags & SLUB_DMA))) 2510 return dma_kmalloc_cache(index, flags); 2511 2512#endif 2513 return &kmalloc_caches[index]; 2514} 2515 2516void *__kmalloc(size_t size, gfp_t flags) 2517{ 2518 struct kmem_cache *s; 2519 2520 if (unlikely(size > PAGE_SIZE / 2)) 2521 return (void *)__get_free_pages(flags | __GFP_COMP, 2522 get_order(size)); 2523 2524 s = get_slab(size, flags); 2525 2526 if (unlikely(ZERO_OR_NULL_PTR(s))) 2527 return s; 2528 2529 return slab_alloc(s, flags, -1, __builtin_return_address(0)); 2530} 2531EXPORT_SYMBOL(__kmalloc); 2532 2533#ifdef CONFIG_NUMA 2534void *__kmalloc_node(size_t size, gfp_t flags, int node) 2535{ 2536 struct kmem_cache *s; 2537 2538 if (unlikely(size > PAGE_SIZE / 2)) 2539 return (void *)__get_free_pages(flags | __GFP_COMP, 2540 get_order(size)); 2541 2542 s = get_slab(size, flags); 2543 2544 if (unlikely(ZERO_OR_NULL_PTR(s))) 2545 return s; 2546 2547 return slab_alloc(s, flags, node, __builtin_return_address(0)); 2548} 2549EXPORT_SYMBOL(__kmalloc_node); 2550#endif 2551 2552size_t ksize(const void *object) 2553{ 2554 struct page *page; 2555 struct kmem_cache *s; 2556 2557 BUG_ON(!object); 2558 if (unlikely(object == ZERO_SIZE_PTR)) 2559 return 0; 2560 2561 page = virt_to_head_page(object); 2562 BUG_ON(!page); 2563 2564 if (unlikely(!PageSlab(page))) 2565 return PAGE_SIZE << compound_order(page); 2566 2567 s = page->slab; 2568 BUG_ON(!s); 2569 2570 /* 2571 * Debugging requires use of the padding between object 2572 * and whatever may come after it. 2573 */ 2574 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2575 return s->objsize; 2576 2577 /* 2578 * If we have the need to store the freelist pointer 2579 * back there or track user information then we can 2580 * only use the space before that information. 2581 */ 2582 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2583 return s->inuse; 2584 2585 /* 2586 * Else we can use all the padding etc for the allocation 2587 */ 2588 return s->size; 2589} 2590EXPORT_SYMBOL(ksize); 2591 2592void kfree(const void *x) 2593{ 2594 struct page *page; 2595 2596 if (unlikely(ZERO_OR_NULL_PTR(x))) 2597 return; 2598 2599 page = virt_to_head_page(x); 2600 if (unlikely(!PageSlab(page))) { 2601 put_page(page); 2602 return; 2603 } 2604 slab_free(page->slab, page, (void *)x, __builtin_return_address(0)); 2605} 2606EXPORT_SYMBOL(kfree); 2607 2608/* 2609 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2610 * the remaining slabs by the number of items in use. The slabs with the 2611 * most items in use come first. New allocations will then fill those up 2612 * and thus they can be removed from the partial lists. 2613 * 2614 * The slabs with the least items are placed last. This results in them 2615 * being allocated from last increasing the chance that the last objects 2616 * are freed in them. 2617 */ 2618int kmem_cache_shrink(struct kmem_cache *s) 2619{ 2620 int node; 2621 int i; 2622 struct kmem_cache_node *n; 2623 struct page *page; 2624 struct page *t; 2625 struct list_head *slabs_by_inuse = 2626 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL); 2627 unsigned long flags; 2628 2629 if (!slabs_by_inuse) 2630 return -ENOMEM; 2631 2632 flush_all(s); 2633 for_each_node_state(node, N_NORMAL_MEMORY) { 2634 n = get_node(s, node); 2635 2636 if (!n->nr_partial) 2637 continue; 2638 2639 for (i = 0; i < s->objects; i++) 2640 INIT_LIST_HEAD(slabs_by_inuse + i); 2641 2642 spin_lock_irqsave(&n->list_lock, flags); 2643 2644 /* 2645 * Build lists indexed by the items in use in each slab. 2646 * 2647 * Note that concurrent frees may occur while we hold the 2648 * list_lock. page->inuse here is the upper limit. 2649 */ 2650 list_for_each_entry_safe(page, t, &n->partial, lru) { 2651 if (!page->inuse && slab_trylock(page)) { 2652 /* 2653 * Must hold slab lock here because slab_free 2654 * may have freed the last object and be 2655 * waiting to release the slab. 2656 */ 2657 list_del(&page->lru); 2658 n->nr_partial--; 2659 slab_unlock(page); 2660 discard_slab(s, page); 2661 } else { 2662 list_move(&page->lru, 2663 slabs_by_inuse + page->inuse); 2664 } 2665 } 2666 2667 /* 2668 * Rebuild the partial list with the slabs filled up most 2669 * first and the least used slabs at the end. 2670 */ 2671 for (i = s->objects - 1; i >= 0; i--) 2672 list_splice(slabs_by_inuse + i, n->partial.prev); 2673 2674 spin_unlock_irqrestore(&n->list_lock, flags); 2675 } 2676 2677 kfree(slabs_by_inuse); 2678 return 0; 2679} 2680EXPORT_SYMBOL(kmem_cache_shrink); 2681 2682#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2683static int slab_mem_going_offline_callback(void *arg) 2684{ 2685 struct kmem_cache *s; 2686 2687 down_read(&slub_lock); 2688 list_for_each_entry(s, &slab_caches, list) 2689 kmem_cache_shrink(s); 2690 up_read(&slub_lock); 2691 2692 return 0; 2693} 2694 2695static void slab_mem_offline_callback(void *arg) 2696{ 2697 struct kmem_cache_node *n; 2698 struct kmem_cache *s; 2699 struct memory_notify *marg = arg; 2700 int offline_node; 2701 2702 offline_node = marg->status_change_nid; 2703 2704 /* 2705 * If the node still has available memory. we need kmem_cache_node 2706 * for it yet. 2707 */ 2708 if (offline_node < 0) 2709 return; 2710 2711 down_read(&slub_lock); 2712 list_for_each_entry(s, &slab_caches, list) { 2713 n = get_node(s, offline_node); 2714 if (n) { 2715 /* 2716 * if n->nr_slabs > 0, slabs still exist on the node 2717 * that is going down. We were unable to free them, 2718 * and offline_pages() function shoudn't call this 2719 * callback. So, we must fail. 2720 */ 2721 BUG_ON(atomic_long_read(&n->nr_slabs)); 2722 2723 s->node[offline_node] = NULL; 2724 kmem_cache_free(kmalloc_caches, n); 2725 } 2726 } 2727 up_read(&slub_lock); 2728} 2729 2730static int slab_mem_going_online_callback(void *arg) 2731{ 2732 struct kmem_cache_node *n; 2733 struct kmem_cache *s; 2734 struct memory_notify *marg = arg; 2735 int nid = marg->status_change_nid; 2736 int ret = 0; 2737 2738 /* 2739 * If the node's memory is already available, then kmem_cache_node is 2740 * already created. Nothing to do. 2741 */ 2742 if (nid < 0) 2743 return 0; 2744 2745 /* 2746 * We are bringing a node online. No memory is availabe yet. We must 2747 * allocate a kmem_cache_node structure in order to bring the node 2748 * online. 2749 */ 2750 down_read(&slub_lock); 2751 list_for_each_entry(s, &slab_caches, list) { 2752 /* 2753 * XXX: kmem_cache_alloc_node will fallback to other nodes 2754 * since memory is not yet available from the node that 2755 * is brought up. 2756 */ 2757 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 2758 if (!n) { 2759 ret = -ENOMEM; 2760 goto out; 2761 } 2762 init_kmem_cache_node(n); 2763 s->node[nid] = n; 2764 } 2765out: 2766 up_read(&slub_lock); 2767 return ret; 2768} 2769 2770static int slab_memory_callback(struct notifier_block *self, 2771 unsigned long action, void *arg) 2772{ 2773 int ret = 0; 2774 2775 switch (action) { 2776 case MEM_GOING_ONLINE: 2777 ret = slab_mem_going_online_callback(arg); 2778 break; 2779 case MEM_GOING_OFFLINE: 2780 ret = slab_mem_going_offline_callback(arg); 2781 break; 2782 case MEM_OFFLINE: 2783 case MEM_CANCEL_ONLINE: 2784 slab_mem_offline_callback(arg); 2785 break; 2786 case MEM_ONLINE: 2787 case MEM_CANCEL_OFFLINE: 2788 break; 2789 } 2790 2791 ret = notifier_from_errno(ret); 2792 return ret; 2793} 2794 2795#endif /* CONFIG_MEMORY_HOTPLUG */ 2796 2797/******************************************************************** 2798 * Basic setup of slabs 2799 *******************************************************************/ 2800 2801void __init kmem_cache_init(void) 2802{ 2803 int i; 2804 int caches = 0; 2805 2806 init_alloc_cpu(); 2807 2808#ifdef CONFIG_NUMA 2809 /* 2810 * Must first have the slab cache available for the allocations of the 2811 * struct kmem_cache_node's. There is special bootstrap code in 2812 * kmem_cache_open for slab_state == DOWN. 2813 */ 2814 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 2815 sizeof(struct kmem_cache_node), GFP_KERNEL); 2816 kmalloc_caches[0].refcount = -1; 2817 caches++; 2818 2819 hotplug_memory_notifier(slab_memory_callback, 1); 2820#endif 2821 2822 /* Able to allocate the per node structures */ 2823 slab_state = PARTIAL; 2824 2825 /* Caches that are not of the two-to-the-power-of size */ 2826 if (KMALLOC_MIN_SIZE <= 64) { 2827 create_kmalloc_cache(&kmalloc_caches[1], 2828 "kmalloc-96", 96, GFP_KERNEL); 2829 caches++; 2830 } 2831 if (KMALLOC_MIN_SIZE <= 128) { 2832 create_kmalloc_cache(&kmalloc_caches[2], 2833 "kmalloc-192", 192, GFP_KERNEL); 2834 caches++; 2835 } 2836 2837 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) { 2838 create_kmalloc_cache(&kmalloc_caches[i], 2839 "kmalloc", 1 << i, GFP_KERNEL); 2840 caches++; 2841 } 2842 2843 2844 /* 2845 * Patch up the size_index table if we have strange large alignment 2846 * requirements for the kmalloc array. This is only the case for 2847 * mips it seems. The standard arches will not generate any code here. 2848 * 2849 * Largest permitted alignment is 256 bytes due to the way we 2850 * handle the index determination for the smaller caches. 2851 * 2852 * Make sure that nothing crazy happens if someone starts tinkering 2853 * around with ARCH_KMALLOC_MINALIGN 2854 */ 2855 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 2856 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 2857 2858 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) 2859 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; 2860 2861 slab_state = UP; 2862 2863 /* Provide the correct kmalloc names now that the caches are up */ 2864 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) 2865 kmalloc_caches[i]. name = 2866 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); 2867 2868#ifdef CONFIG_SMP 2869 register_cpu_notifier(&slab_notifier); 2870 kmem_size = offsetof(struct kmem_cache, cpu_slab) + 2871 nr_cpu_ids * sizeof(struct kmem_cache_cpu *); 2872#else 2873 kmem_size = sizeof(struct kmem_cache); 2874#endif 2875 2876 2877 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 2878 " CPUs=%d, Nodes=%d\n", 2879 caches, cache_line_size(), 2880 slub_min_order, slub_max_order, slub_min_objects, 2881 nr_cpu_ids, nr_node_ids); 2882} 2883 2884/* 2885 * Find a mergeable slab cache 2886 */ 2887static int slab_unmergeable(struct kmem_cache *s) 2888{ 2889 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 2890 return 1; 2891 2892 if (s->ctor) 2893 return 1; 2894 2895 /* 2896 * We may have set a slab to be unmergeable during bootstrap. 2897 */ 2898 if (s->refcount < 0) 2899 return 1; 2900 2901 return 0; 2902} 2903 2904static struct kmem_cache *find_mergeable(size_t size, 2905 size_t align, unsigned long flags, const char *name, 2906 void (*ctor)(struct kmem_cache *, void *)) 2907{ 2908 struct kmem_cache *s; 2909 2910 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 2911 return NULL; 2912 2913 if (ctor) 2914 return NULL; 2915 2916 size = ALIGN(size, sizeof(void *)); 2917 align = calculate_alignment(flags, align, size); 2918 size = ALIGN(size, align); 2919 flags = kmem_cache_flags(size, flags, name, NULL); 2920 2921 list_for_each_entry(s, &slab_caches, list) { 2922 if (slab_unmergeable(s)) 2923 continue; 2924 2925 if (size > s->size) 2926 continue; 2927 2928 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 2929 continue; 2930 /* 2931 * Check if alignment is compatible. 2932 * Courtesy of Adrian Drzewiecki 2933 */ 2934 if ((s->size & ~(align -1)) != s->size) 2935 continue; 2936 2937 if (s->size - size >= sizeof(void *)) 2938 continue; 2939 2940 return s; 2941 } 2942 return NULL; 2943} 2944 2945struct kmem_cache *kmem_cache_create(const char *name, size_t size, 2946 size_t align, unsigned long flags, 2947 void (*ctor)(struct kmem_cache *, void *)) 2948{ 2949 struct kmem_cache *s; 2950 2951 down_write(&slub_lock); 2952 s = find_mergeable(size, align, flags, name, ctor); 2953 if (s) { 2954 int cpu; 2955 2956 s->refcount++; 2957 /* 2958 * Adjust the object sizes so that we clear 2959 * the complete object on kzalloc. 2960 */ 2961 s->objsize = max(s->objsize, (int)size); 2962 2963 /* 2964 * And then we need to update the object size in the 2965 * per cpu structures 2966 */ 2967 for_each_online_cpu(cpu) 2968 get_cpu_slab(s, cpu)->objsize = s->objsize; 2969 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 2970 up_write(&slub_lock); 2971 if (sysfs_slab_alias(s, name)) 2972 goto err; 2973 return s; 2974 } 2975 s = kmalloc(kmem_size, GFP_KERNEL); 2976 if (s) { 2977 if (kmem_cache_open(s, GFP_KERNEL, name, 2978 size, align, flags, ctor)) { 2979 list_add(&s->list, &slab_caches); 2980 up_write(&slub_lock); 2981 if (sysfs_slab_add(s)) 2982 goto err; 2983 return s; 2984 } 2985 kfree(s); 2986 } 2987 up_write(&slub_lock); 2988 2989err: 2990 if (flags & SLAB_PANIC) 2991 panic("Cannot create slabcache %s\n", name); 2992 else 2993 s = NULL; 2994 return s; 2995} 2996EXPORT_SYMBOL(kmem_cache_create); 2997 2998#ifdef CONFIG_SMP 2999/* 3000 * Use the cpu notifier to insure that the cpu slabs are flushed when 3001 * necessary. 3002 */ 3003static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3004 unsigned long action, void *hcpu) 3005{ 3006 long cpu = (long)hcpu; 3007 struct kmem_cache *s; 3008 unsigned long flags; 3009 3010 switch (action) { 3011 case CPU_UP_PREPARE: 3012 case CPU_UP_PREPARE_FROZEN: 3013 init_alloc_cpu_cpu(cpu); 3014 down_read(&slub_lock); 3015 list_for_each_entry(s, &slab_caches, list) 3016 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, 3017 GFP_KERNEL); 3018 up_read(&slub_lock); 3019 break; 3020 3021 case CPU_UP_CANCELED: 3022 case CPU_UP_CANCELED_FROZEN: 3023 case CPU_DEAD: 3024 case CPU_DEAD_FROZEN: 3025 down_read(&slub_lock); 3026 list_for_each_entry(s, &slab_caches, list) { 3027 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3028 3029 local_irq_save(flags); 3030 __flush_cpu_slab(s, cpu); 3031 local_irq_restore(flags); 3032 free_kmem_cache_cpu(c, cpu); 3033 s->cpu_slab[cpu] = NULL; 3034 } 3035 up_read(&slub_lock); 3036 break; 3037 default: 3038 break; 3039 } 3040 return NOTIFY_OK; 3041} 3042 3043static struct notifier_block __cpuinitdata slab_notifier = 3044 { &slab_cpuup_callback, NULL, 0 }; 3045 3046#endif 3047 3048void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) 3049{ 3050 struct kmem_cache *s; 3051 3052 if (unlikely(size > PAGE_SIZE / 2)) 3053 return (void *)__get_free_pages(gfpflags | __GFP_COMP, 3054 get_order(size)); 3055 s = get_slab(size, gfpflags); 3056 3057 if (unlikely(ZERO_OR_NULL_PTR(s))) 3058 return s; 3059 3060 return slab_alloc(s, gfpflags, -1, caller); 3061} 3062 3063void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3064 int node, void *caller) 3065{ 3066 struct kmem_cache *s; 3067 3068 if (unlikely(size > PAGE_SIZE / 2)) 3069 return (void *)__get_free_pages(gfpflags | __GFP_COMP, 3070 get_order(size)); 3071 s = get_slab(size, gfpflags); 3072 3073 if (unlikely(ZERO_OR_NULL_PTR(s))) 3074 return s; 3075 3076 return slab_alloc(s, gfpflags, node, caller); 3077} 3078 3079#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) 3080static int validate_slab(struct kmem_cache *s, struct page *page, 3081 unsigned long *map) 3082{ 3083 void *p; 3084 void *addr = page_address(page); 3085 3086 if (!check_slab(s, page) || 3087 !on_freelist(s, page, NULL)) 3088 return 0; 3089 3090 /* Now we know that a valid freelist exists */ 3091 bitmap_zero(map, s->objects); 3092 3093 for_each_free_object(p, s, page->freelist) { 3094 set_bit(slab_index(p, s, addr), map); 3095 if (!check_object(s, page, p, 0)) 3096 return 0; 3097 } 3098 3099 for_each_object(p, s, addr) 3100 if (!test_bit(slab_index(p, s, addr), map)) 3101 if (!check_object(s, page, p, 1)) 3102 return 0; 3103 return 1; 3104} 3105 3106static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3107 unsigned long *map) 3108{ 3109 if (slab_trylock(page)) { 3110 validate_slab(s, page, map); 3111 slab_unlock(page); 3112 } else 3113 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3114 s->name, page); 3115 3116 if (s->flags & DEBUG_DEFAULT_FLAGS) { 3117 if (!SlabDebug(page)) 3118 printk(KERN_ERR "SLUB %s: SlabDebug not set " 3119 "on slab 0x%p\n", s->name, page); 3120 } else { 3121 if (SlabDebug(page)) 3122 printk(KERN_ERR "SLUB %s: SlabDebug set on " 3123 "slab 0x%p\n", s->name, page); 3124 } 3125} 3126 3127static int validate_slab_node(struct kmem_cache *s, 3128 struct kmem_cache_node *n, unsigned long *map) 3129{ 3130 unsigned long count = 0; 3131 struct page *page; 3132 unsigned long flags; 3133 3134 spin_lock_irqsave(&n->list_lock, flags); 3135 3136 list_for_each_entry(page, &n->partial, lru) { 3137 validate_slab_slab(s, page, map); 3138 count++; 3139 } 3140 if (count != n->nr_partial) 3141 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3142 "counter=%ld\n", s->name, count, n->nr_partial); 3143 3144 if (!(s->flags & SLAB_STORE_USER)) 3145 goto out; 3146 3147 list_for_each_entry(page, &n->full, lru) { 3148 validate_slab_slab(s, page, map); 3149 count++; 3150 } 3151 if (count != atomic_long_read(&n->nr_slabs)) 3152 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3153 "counter=%ld\n", s->name, count, 3154 atomic_long_read(&n->nr_slabs)); 3155 3156out: 3157 spin_unlock_irqrestore(&n->list_lock, flags); 3158 return count; 3159} 3160 3161static long validate_slab_cache(struct kmem_cache *s) 3162{ 3163 int node; 3164 unsigned long count = 0; 3165 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) * 3166 sizeof(unsigned long), GFP_KERNEL); 3167 3168 if (!map) 3169 return -ENOMEM; 3170 3171 flush_all(s); 3172 for_each_node_state(node, N_NORMAL_MEMORY) { 3173 struct kmem_cache_node *n = get_node(s, node); 3174 3175 count += validate_slab_node(s, n, map); 3176 } 3177 kfree(map); 3178 return count; 3179} 3180 3181#ifdef SLUB_RESILIENCY_TEST 3182static void resiliency_test(void) 3183{ 3184 u8 *p; 3185 3186 printk(KERN_ERR "SLUB resiliency testing\n"); 3187 printk(KERN_ERR "-----------------------\n"); 3188 printk(KERN_ERR "A. Corruption after allocation\n"); 3189 3190 p = kzalloc(16, GFP_KERNEL); 3191 p[16] = 0x12; 3192 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3193 " 0x12->0x%p\n\n", p + 16); 3194 3195 validate_slab_cache(kmalloc_caches + 4); 3196 3197 /* Hmmm... The next two are dangerous */ 3198 p = kzalloc(32, GFP_KERNEL); 3199 p[32 + sizeof(void *)] = 0x34; 3200 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3201 " 0x34 -> -0x%p\n", p); 3202 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); 3203 3204 validate_slab_cache(kmalloc_caches + 5); 3205 p = kzalloc(64, GFP_KERNEL); 3206 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3207 *p = 0x56; 3208 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3209 p); 3210 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n"); 3211 validate_slab_cache(kmalloc_caches + 6); 3212 3213 printk(KERN_ERR "\nB. Corruption after free\n"); 3214 p = kzalloc(128, GFP_KERNEL); 3215 kfree(p); 3216 *p = 0x78; 3217 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3218 validate_slab_cache(kmalloc_caches + 7); 3219 3220 p = kzalloc(256, GFP_KERNEL); 3221 kfree(p); 3222 p[50] = 0x9a; 3223 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); 3224 validate_slab_cache(kmalloc_caches + 8); 3225 3226 p = kzalloc(512, GFP_KERNEL); 3227 kfree(p); 3228 p[512] = 0xab; 3229 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3230 validate_slab_cache(kmalloc_caches + 9); 3231} 3232#else 3233static void resiliency_test(void) {}; 3234#endif 3235 3236/* 3237 * Generate lists of code addresses where slabcache objects are allocated 3238 * and freed. 3239 */ 3240 3241struct location { 3242 unsigned long count; 3243 void *addr; 3244 long long sum_time; 3245 long min_time; 3246 long max_time; 3247 long min_pid; 3248 long max_pid; 3249 cpumask_t cpus; 3250 nodemask_t nodes; 3251}; 3252 3253struct loc_track { 3254 unsigned long max; 3255 unsigned long count; 3256 struct location *loc; 3257}; 3258 3259static void free_loc_track(struct loc_track *t) 3260{ 3261 if (t->max) 3262 free_pages((unsigned long)t->loc, 3263 get_order(sizeof(struct location) * t->max)); 3264} 3265 3266static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3267{ 3268 struct location *l; 3269 int order; 3270 3271 order = get_order(sizeof(struct location) * max); 3272 3273 l = (void *)__get_free_pages(flags, order); 3274 if (!l) 3275 return 0; 3276 3277 if (t->count) { 3278 memcpy(l, t->loc, sizeof(struct location) * t->count); 3279 free_loc_track(t); 3280 } 3281 t->max = max; 3282 t->loc = l; 3283 return 1; 3284} 3285 3286static int add_location(struct loc_track *t, struct kmem_cache *s, 3287 const struct track *track) 3288{ 3289 long start, end, pos; 3290 struct location *l; 3291 void *caddr; 3292 unsigned long age = jiffies - track->when; 3293 3294 start = -1; 3295 end = t->count; 3296 3297 for ( ; ; ) { 3298 pos = start + (end - start + 1) / 2; 3299 3300 /* 3301 * There is nothing at "end". If we end up there 3302 * we need to add something to before end. 3303 */ 3304 if (pos == end) 3305 break; 3306 3307 caddr = t->loc[pos].addr; 3308 if (track->addr == caddr) { 3309 3310 l = &t->loc[pos]; 3311 l->count++; 3312 if (track->when) { 3313 l->sum_time += age; 3314 if (age < l->min_time) 3315 l->min_time = age; 3316 if (age > l->max_time) 3317 l->max_time = age; 3318 3319 if (track->pid < l->min_pid) 3320 l->min_pid = track->pid; 3321 if (track->pid > l->max_pid) 3322 l->max_pid = track->pid; 3323 3324 cpu_set(track->cpu, l->cpus); 3325 } 3326 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3327 return 1; 3328 } 3329 3330 if (track->addr < caddr) 3331 end = pos; 3332 else 3333 start = pos; 3334 } 3335 3336 /* 3337 * Not found. Insert new tracking element. 3338 */ 3339 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3340 return 0; 3341 3342 l = t->loc + pos; 3343 if (pos < t->count) 3344 memmove(l + 1, l, 3345 (t->count - pos) * sizeof(struct location)); 3346 t->count++; 3347 l->count = 1; 3348 l->addr = track->addr; 3349 l->sum_time = age; 3350 l->min_time = age; 3351 l->max_time = age; 3352 l->min_pid = track->pid; 3353 l->max_pid = track->pid; 3354 cpus_clear(l->cpus); 3355 cpu_set(track->cpu, l->cpus); 3356 nodes_clear(l->nodes); 3357 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3358 return 1; 3359} 3360 3361static void process_slab(struct loc_track *t, struct kmem_cache *s, 3362 struct page *page, enum track_item alloc) 3363{ 3364 void *addr = page_address(page); 3365 DECLARE_BITMAP(map, s->objects); 3366 void *p; 3367 3368 bitmap_zero(map, s->objects); 3369 for_each_free_object(p, s, page->freelist) 3370 set_bit(slab_index(p, s, addr), map); 3371 3372 for_each_object(p, s, addr) 3373 if (!test_bit(slab_index(p, s, addr), map)) 3374 add_location(t, s, get_track(s, p, alloc)); 3375} 3376 3377static int list_locations(struct kmem_cache *s, char *buf, 3378 enum track_item alloc) 3379{ 3380 int n = 0; 3381 unsigned long i; 3382 struct loc_track t = { 0, 0, NULL }; 3383 int node; 3384 3385 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3386 GFP_TEMPORARY)) 3387 return sprintf(buf, "Out of memory\n"); 3388 3389 /* Push back cpu slabs */ 3390 flush_all(s); 3391 3392 for_each_node_state(node, N_NORMAL_MEMORY) { 3393 struct kmem_cache_node *n = get_node(s, node); 3394 unsigned long flags; 3395 struct page *page; 3396 3397 if (!atomic_long_read(&n->nr_slabs)) 3398 continue; 3399 3400 spin_lock_irqsave(&n->list_lock, flags); 3401 list_for_each_entry(page, &n->partial, lru) 3402 process_slab(&t, s, page, alloc); 3403 list_for_each_entry(page, &n->full, lru) 3404 process_slab(&t, s, page, alloc); 3405 spin_unlock_irqrestore(&n->list_lock, flags); 3406 } 3407 3408 for (i = 0; i < t.count; i++) { 3409 struct location *l = &t.loc[i]; 3410 3411 if (n > PAGE_SIZE - 100) 3412 break; 3413 n += sprintf(buf + n, "%7ld ", l->count); 3414 3415 if (l->addr) 3416 n += sprint_symbol(buf + n, (unsigned long)l->addr); 3417 else 3418 n += sprintf(buf + n, "<not-available>"); 3419 3420 if (l->sum_time != l->min_time) { 3421 unsigned long remainder; 3422 3423 n += sprintf(buf + n, " age=%ld/%ld/%ld", 3424 l->min_time, 3425 div_long_long_rem(l->sum_time, l->count, &remainder), 3426 l->max_time); 3427 } else 3428 n += sprintf(buf + n, " age=%ld", 3429 l->min_time); 3430 3431 if (l->min_pid != l->max_pid) 3432 n += sprintf(buf + n, " pid=%ld-%ld", 3433 l->min_pid, l->max_pid); 3434 else 3435 n += sprintf(buf + n, " pid=%ld", 3436 l->min_pid); 3437 3438 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && 3439 n < PAGE_SIZE - 60) { 3440 n += sprintf(buf + n, " cpus="); 3441 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50, 3442 l->cpus); 3443 } 3444 3445 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && 3446 n < PAGE_SIZE - 60) { 3447 n += sprintf(buf + n, " nodes="); 3448 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50, 3449 l->nodes); 3450 } 3451 3452 n += sprintf(buf + n, "\n"); 3453 } 3454 3455 free_loc_track(&t); 3456 if (!t.count) 3457 n += sprintf(buf, "No data\n"); 3458 return n; 3459} 3460 3461static unsigned long count_partial(struct kmem_cache_node *n) 3462{ 3463 unsigned long flags; 3464 unsigned long x = 0; 3465 struct page *page; 3466 3467 spin_lock_irqsave(&n->list_lock, flags); 3468 list_for_each_entry(page, &n->partial, lru) 3469 x += page->inuse; 3470 spin_unlock_irqrestore(&n->list_lock, flags); 3471 return x; 3472} 3473 3474enum slab_stat_type { 3475 SL_FULL, 3476 SL_PARTIAL, 3477 SL_CPU, 3478 SL_OBJECTS 3479}; 3480 3481#define SO_FULL (1 << SL_FULL) 3482#define SO_PARTIAL (1 << SL_PARTIAL) 3483#define SO_CPU (1 << SL_CPU) 3484#define SO_OBJECTS (1 << SL_OBJECTS) 3485 3486static unsigned long slab_objects(struct kmem_cache *s, 3487 char *buf, unsigned long flags) 3488{ 3489 unsigned long total = 0; 3490 int cpu; 3491 int node; 3492 int x; 3493 unsigned long *nodes; 3494 unsigned long *per_cpu; 3495 3496 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3497 per_cpu = nodes + nr_node_ids; 3498 3499 for_each_possible_cpu(cpu) { 3500 struct page *page; 3501 int node; 3502 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3503 3504 if (!c) 3505 continue; 3506 3507 page = c->page; 3508 node = c->node; 3509 if (node < 0) 3510 continue; 3511 if (page) { 3512 if (flags & SO_CPU) { 3513 int x = 0; 3514 3515 if (flags & SO_OBJECTS) 3516 x = page->inuse; 3517 else 3518 x = 1; 3519 total += x; 3520 nodes[node] += x; 3521 } 3522 per_cpu[node]++; 3523 } 3524 } 3525 3526 for_each_node_state(node, N_NORMAL_MEMORY) { 3527 struct kmem_cache_node *n = get_node(s, node); 3528 3529 if (flags & SO_PARTIAL) { 3530 if (flags & SO_OBJECTS) 3531 x = count_partial(n); 3532 else 3533 x = n->nr_partial; 3534 total += x; 3535 nodes[node] += x; 3536 } 3537 3538 if (flags & SO_FULL) { 3539 int full_slabs = atomic_long_read(&n->nr_slabs) 3540 - per_cpu[node] 3541 - n->nr_partial; 3542 3543 if (flags & SO_OBJECTS) 3544 x = full_slabs * s->objects; 3545 else 3546 x = full_slabs; 3547 total += x; 3548 nodes[node] += x; 3549 } 3550 } 3551 3552 x = sprintf(buf, "%lu", total); 3553#ifdef CONFIG_NUMA 3554 for_each_node_state(node, N_NORMAL_MEMORY) 3555 if (nodes[node]) 3556 x += sprintf(buf + x, " N%d=%lu", 3557 node, nodes[node]); 3558#endif 3559 kfree(nodes); 3560 return x + sprintf(buf + x, "\n"); 3561} 3562 3563static int any_slab_objects(struct kmem_cache *s) 3564{ 3565 int node; 3566 int cpu; 3567 3568 for_each_possible_cpu(cpu) { 3569 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3570 3571 if (c && c->page) 3572 return 1; 3573 } 3574 3575 for_each_online_node(node) { 3576 struct kmem_cache_node *n = get_node(s, node); 3577 3578 if (!n) 3579 continue; 3580 3581 if (n->nr_partial || atomic_long_read(&n->nr_slabs)) 3582 return 1; 3583 } 3584 return 0; 3585} 3586 3587#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3588#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3589 3590struct slab_attribute { 3591 struct attribute attr; 3592 ssize_t (*show)(struct kmem_cache *s, char *buf); 3593 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3594}; 3595 3596#define SLAB_ATTR_RO(_name) \ 3597 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3598 3599#define SLAB_ATTR(_name) \ 3600 static struct slab_attribute _name##_attr = \ 3601 __ATTR(_name, 0644, _name##_show, _name##_store) 3602 3603static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3604{ 3605 return sprintf(buf, "%d\n", s->size); 3606} 3607SLAB_ATTR_RO(slab_size); 3608 3609static ssize_t align_show(struct kmem_cache *s, char *buf) 3610{ 3611 return sprintf(buf, "%d\n", s->align); 3612} 3613SLAB_ATTR_RO(align); 3614 3615static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3616{ 3617 return sprintf(buf, "%d\n", s->objsize); 3618} 3619SLAB_ATTR_RO(object_size); 3620 3621static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3622{ 3623 return sprintf(buf, "%d\n", s->objects); 3624} 3625SLAB_ATTR_RO(objs_per_slab); 3626 3627static ssize_t order_show(struct kmem_cache *s, char *buf) 3628{ 3629 return sprintf(buf, "%d\n", s->order); 3630} 3631SLAB_ATTR_RO(order); 3632 3633static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3634{ 3635 if (s->ctor) { 3636 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3637 3638 return n + sprintf(buf + n, "\n"); 3639 } 3640 return 0; 3641} 3642SLAB_ATTR_RO(ctor); 3643 3644static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3645{ 3646 return sprintf(buf, "%d\n", s->refcount - 1); 3647} 3648SLAB_ATTR_RO(aliases); 3649 3650static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3651{ 3652 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU); 3653} 3654SLAB_ATTR_RO(slabs); 3655 3656static ssize_t partial_show(struct kmem_cache *s, char *buf) 3657{ 3658 return slab_objects(s, buf, SO_PARTIAL); 3659} 3660SLAB_ATTR_RO(partial); 3661 3662static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 3663{ 3664 return slab_objects(s, buf, SO_CPU); 3665} 3666SLAB_ATTR_RO(cpu_slabs); 3667 3668static ssize_t objects_show(struct kmem_cache *s, char *buf) 3669{ 3670 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS); 3671} 3672SLAB_ATTR_RO(objects); 3673 3674static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 3675{ 3676 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 3677} 3678 3679static ssize_t sanity_checks_store(struct kmem_cache *s, 3680 const char *buf, size_t length) 3681{ 3682 s->flags &= ~SLAB_DEBUG_FREE; 3683 if (buf[0] == '1') 3684 s->flags |= SLAB_DEBUG_FREE; 3685 return length; 3686} 3687SLAB_ATTR(sanity_checks); 3688 3689static ssize_t trace_show(struct kmem_cache *s, char *buf) 3690{ 3691 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 3692} 3693 3694static ssize_t trace_store(struct kmem_cache *s, const char *buf, 3695 size_t length) 3696{ 3697 s->flags &= ~SLAB_TRACE; 3698 if (buf[0] == '1') 3699 s->flags |= SLAB_TRACE; 3700 return length; 3701} 3702SLAB_ATTR(trace); 3703 3704static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 3705{ 3706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 3707} 3708 3709static ssize_t reclaim_account_store(struct kmem_cache *s, 3710 const char *buf, size_t length) 3711{ 3712 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 3713 if (buf[0] == '1') 3714 s->flags |= SLAB_RECLAIM_ACCOUNT; 3715 return length; 3716} 3717SLAB_ATTR(reclaim_account); 3718 3719static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 3720{ 3721 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 3722} 3723SLAB_ATTR_RO(hwcache_align); 3724 3725#ifdef CONFIG_ZONE_DMA 3726static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 3727{ 3728 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 3729} 3730SLAB_ATTR_RO(cache_dma); 3731#endif 3732 3733static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 3734{ 3735 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 3736} 3737SLAB_ATTR_RO(destroy_by_rcu); 3738 3739static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 3740{ 3741 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 3742} 3743 3744static ssize_t red_zone_store(struct kmem_cache *s, 3745 const char *buf, size_t length) 3746{ 3747 if (any_slab_objects(s)) 3748 return -EBUSY; 3749 3750 s->flags &= ~SLAB_RED_ZONE; 3751 if (buf[0] == '1') 3752 s->flags |= SLAB_RED_ZONE; 3753 calculate_sizes(s); 3754 return length; 3755} 3756SLAB_ATTR(red_zone); 3757 3758static ssize_t poison_show(struct kmem_cache *s, char *buf) 3759{ 3760 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 3761} 3762 3763static ssize_t poison_store(struct kmem_cache *s, 3764 const char *buf, size_t length) 3765{ 3766 if (any_slab_objects(s)) 3767 return -EBUSY; 3768 3769 s->flags &= ~SLAB_POISON; 3770 if (buf[0] == '1') 3771 s->flags |= SLAB_POISON; 3772 calculate_sizes(s); 3773 return length; 3774} 3775SLAB_ATTR(poison); 3776 3777static ssize_t store_user_show(struct kmem_cache *s, char *buf) 3778{ 3779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 3780} 3781 3782static ssize_t store_user_store(struct kmem_cache *s, 3783 const char *buf, size_t length) 3784{ 3785 if (any_slab_objects(s)) 3786 return -EBUSY; 3787 3788 s->flags &= ~SLAB_STORE_USER; 3789 if (buf[0] == '1') 3790 s->flags |= SLAB_STORE_USER; 3791 calculate_sizes(s); 3792 return length; 3793} 3794SLAB_ATTR(store_user); 3795 3796static ssize_t validate_show(struct kmem_cache *s, char *buf) 3797{ 3798 return 0; 3799} 3800 3801static ssize_t validate_store(struct kmem_cache *s, 3802 const char *buf, size_t length) 3803{ 3804 int ret = -EINVAL; 3805 3806 if (buf[0] == '1') { 3807 ret = validate_slab_cache(s); 3808 if (ret >= 0) 3809 ret = length; 3810 } 3811 return ret; 3812} 3813SLAB_ATTR(validate); 3814 3815static ssize_t shrink_show(struct kmem_cache *s, char *buf) 3816{ 3817 return 0; 3818} 3819 3820static ssize_t shrink_store(struct kmem_cache *s, 3821 const char *buf, size_t length) 3822{ 3823 if (buf[0] == '1') { 3824 int rc = kmem_cache_shrink(s); 3825 3826 if (rc) 3827 return rc; 3828 } else 3829 return -EINVAL; 3830 return length; 3831} 3832SLAB_ATTR(shrink); 3833 3834static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 3835{ 3836 if (!(s->flags & SLAB_STORE_USER)) 3837 return -ENOSYS; 3838 return list_locations(s, buf, TRACK_ALLOC); 3839} 3840SLAB_ATTR_RO(alloc_calls); 3841 3842static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 3843{ 3844 if (!(s->flags & SLAB_STORE_USER)) 3845 return -ENOSYS; 3846 return list_locations(s, buf, TRACK_FREE); 3847} 3848SLAB_ATTR_RO(free_calls); 3849 3850#ifdef CONFIG_NUMA 3851static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf) 3852{ 3853 return sprintf(buf, "%d\n", s->defrag_ratio / 10); 3854} 3855 3856static ssize_t defrag_ratio_store(struct kmem_cache *s, 3857 const char *buf, size_t length) 3858{ 3859 int n = simple_strtoul(buf, NULL, 10); 3860 3861 if (n < 100) 3862 s->defrag_ratio = n * 10; 3863 return length; 3864} 3865SLAB_ATTR(defrag_ratio); 3866#endif 3867 3868static struct attribute * slab_attrs[] = { 3869 &slab_size_attr.attr, 3870 &object_size_attr.attr, 3871 &objs_per_slab_attr.attr, 3872 &order_attr.attr, 3873 &objects_attr.attr, 3874 &slabs_attr.attr, 3875 &partial_attr.attr, 3876 &cpu_slabs_attr.attr, 3877 &ctor_attr.attr, 3878 &aliases_attr.attr, 3879 &align_attr.attr, 3880 &sanity_checks_attr.attr, 3881 &trace_attr.attr, 3882 &hwcache_align_attr.attr, 3883 &reclaim_account_attr.attr, 3884 &destroy_by_rcu_attr.attr, 3885 &red_zone_attr.attr, 3886 &poison_attr.attr, 3887 &store_user_attr.attr, 3888 &validate_attr.attr, 3889 &shrink_attr.attr, 3890 &alloc_calls_attr.attr, 3891 &free_calls_attr.attr, 3892#ifdef CONFIG_ZONE_DMA 3893 &cache_dma_attr.attr, 3894#endif 3895#ifdef CONFIG_NUMA 3896 &defrag_ratio_attr.attr, 3897#endif 3898 NULL 3899}; 3900 3901static struct attribute_group slab_attr_group = { 3902 .attrs = slab_attrs, 3903}; 3904 3905static ssize_t slab_attr_show(struct kobject *kobj, 3906 struct attribute *attr, 3907 char *buf) 3908{ 3909 struct slab_attribute *attribute; 3910 struct kmem_cache *s; 3911 int err; 3912 3913 attribute = to_slab_attr(attr); 3914 s = to_slab(kobj); 3915 3916 if (!attribute->show) 3917 return -EIO; 3918 3919 err = attribute->show(s, buf); 3920 3921 return err; 3922} 3923 3924static ssize_t slab_attr_store(struct kobject *kobj, 3925 struct attribute *attr, 3926 const char *buf, size_t len) 3927{ 3928 struct slab_attribute *attribute; 3929 struct kmem_cache *s; 3930 int err; 3931 3932 attribute = to_slab_attr(attr); 3933 s = to_slab(kobj); 3934 3935 if (!attribute->store) 3936 return -EIO; 3937 3938 err = attribute->store(s, buf, len); 3939 3940 return err; 3941} 3942 3943static struct sysfs_ops slab_sysfs_ops = { 3944 .show = slab_attr_show, 3945 .store = slab_attr_store, 3946}; 3947 3948static struct kobj_type slab_ktype = { 3949 .sysfs_ops = &slab_sysfs_ops, 3950}; 3951 3952static int uevent_filter(struct kset *kset, struct kobject *kobj) 3953{ 3954 struct kobj_type *ktype = get_ktype(kobj); 3955 3956 if (ktype == &slab_ktype) 3957 return 1; 3958 return 0; 3959} 3960 3961static struct kset_uevent_ops slab_uevent_ops = { 3962 .filter = uevent_filter, 3963}; 3964 3965static decl_subsys(slab, &slab_ktype, &slab_uevent_ops); 3966 3967#define ID_STR_LENGTH 64 3968 3969/* Create a unique string id for a slab cache: 3970 * format 3971 * :[flags-]size:[memory address of kmemcache] 3972 */ 3973static char *create_unique_id(struct kmem_cache *s) 3974{ 3975 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 3976 char *p = name; 3977 3978 BUG_ON(!name); 3979 3980 *p++ = ':'; 3981 /* 3982 * First flags affecting slabcache operations. We will only 3983 * get here for aliasable slabs so we do not need to support 3984 * too many flags. The flags here must cover all flags that 3985 * are matched during merging to guarantee that the id is 3986 * unique. 3987 */ 3988 if (s->flags & SLAB_CACHE_DMA) 3989 *p++ = 'd'; 3990 if (s->flags & SLAB_RECLAIM_ACCOUNT) 3991 *p++ = 'a'; 3992 if (s->flags & SLAB_DEBUG_FREE) 3993 *p++ = 'F'; 3994 if (p != name + 1) 3995 *p++ = '-'; 3996 p += sprintf(p, "%07d", s->size); 3997 BUG_ON(p > name + ID_STR_LENGTH - 1); 3998 return name; 3999} 4000 4001static int sysfs_slab_add(struct kmem_cache *s) 4002{ 4003 int err; 4004 const char *name; 4005 int unmergeable; 4006 4007 if (slab_state < SYSFS) 4008 /* Defer until later */ 4009 return 0; 4010 4011 unmergeable = slab_unmergeable(s); 4012 if (unmergeable) { 4013 /* 4014 * Slabcache can never be merged so we can use the name proper. 4015 * This is typically the case for debug situations. In that 4016 * case we can catch duplicate names easily. 4017 */ 4018 sysfs_remove_link(&slab_subsys.kobj, s->name); 4019 name = s->name; 4020 } else { 4021 /* 4022 * Create a unique name for the slab as a target 4023 * for the symlinks. 4024 */ 4025 name = create_unique_id(s); 4026 } 4027 4028 kobj_set_kset_s(s, slab_subsys); 4029 kobject_set_name(&s->kobj, name); 4030 kobject_init(&s->kobj); 4031 err = kobject_add(&s->kobj); 4032 if (err) 4033 return err; 4034 4035 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4036 if (err) 4037 return err; 4038 kobject_uevent(&s->kobj, KOBJ_ADD); 4039 if (!unmergeable) { 4040 /* Setup first alias */ 4041 sysfs_slab_alias(s, s->name); 4042 kfree(name); 4043 } 4044 return 0; 4045} 4046 4047static void sysfs_slab_remove(struct kmem_cache *s) 4048{ 4049 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4050 kobject_del(&s->kobj); 4051} 4052 4053/* 4054 * Need to buffer aliases during bootup until sysfs becomes 4055 * available lest we loose that information. 4056 */ 4057struct saved_alias { 4058 struct kmem_cache *s; 4059 const char *name; 4060 struct saved_alias *next; 4061}; 4062 4063static struct saved_alias *alias_list; 4064 4065static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4066{ 4067 struct saved_alias *al; 4068 4069 if (slab_state == SYSFS) { 4070 /* 4071 * If we have a leftover link then remove it. 4072 */ 4073 sysfs_remove_link(&slab_subsys.kobj, name); 4074 return sysfs_create_link(&slab_subsys.kobj, 4075 &s->kobj, name); 4076 } 4077 4078 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4079 if (!al) 4080 return -ENOMEM; 4081 4082 al->s = s; 4083 al->name = name; 4084 al->next = alias_list; 4085 alias_list = al; 4086 return 0; 4087} 4088 4089static int __init slab_sysfs_init(void) 4090{ 4091 struct kmem_cache *s; 4092 int err; 4093 4094 err = subsystem_register(&slab_subsys); 4095 if (err) { 4096 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4097 return -ENOSYS; 4098 } 4099 4100 slab_state = SYSFS; 4101 4102 list_for_each_entry(s, &slab_caches, list) { 4103 err = sysfs_slab_add(s); 4104 if (err) 4105 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4106 " to sysfs\n", s->name); 4107 } 4108 4109 while (alias_list) { 4110 struct saved_alias *al = alias_list; 4111 4112 alias_list = alias_list->next; 4113 err = sysfs_slab_alias(al->s, al->name); 4114 if (err) 4115 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4116 " %s to sysfs\n", s->name); 4117 kfree(al); 4118 } 4119 4120 resiliency_test(); 4121 return 0; 4122} 4123 4124__initcall(slab_sysfs_init); 4125#endif 4126