slub.c revision 98072e4d977aabe6a39abb95951cd8bf2c2202d5
1/* 2 * SLUB: A slab allocator that limits cache line use instead of queuing 3 * objects in per cpu and per node lists. 4 * 5 * The allocator synchronizes using per slab locks and only 6 * uses a centralized lock to manage a pool of partial slabs. 7 * 8 * (C) 2007 SGI, Christoph Lameter 9 */ 10 11#include <linux/mm.h> 12#include <linux/swap.h> /* struct reclaim_state */ 13#include <linux/module.h> 14#include <linux/bit_spinlock.h> 15#include <linux/interrupt.h> 16#include <linux/bitops.h> 17#include <linux/slab.h> 18#include <linux/proc_fs.h> 19#include <linux/seq_file.h> 20#include <linux/kmemcheck.h> 21#include <linux/cpu.h> 22#include <linux/cpuset.h> 23#include <linux/mempolicy.h> 24#include <linux/ctype.h> 25#include <linux/debugobjects.h> 26#include <linux/kallsyms.h> 27#include <linux/memory.h> 28#include <linux/math64.h> 29#include <linux/fault-inject.h> 30 31#include <trace/events/kmem.h> 32 33/* 34 * Lock order: 35 * 1. slab_lock(page) 36 * 2. slab->list_lock 37 * 38 * The slab_lock protects operations on the object of a particular 39 * slab and its metadata in the page struct. If the slab lock 40 * has been taken then no allocations nor frees can be performed 41 * on the objects in the slab nor can the slab be added or removed 42 * from the partial or full lists since this would mean modifying 43 * the page_struct of the slab. 44 * 45 * The list_lock protects the partial and full list on each node and 46 * the partial slab counter. If taken then no new slabs may be added or 47 * removed from the lists nor make the number of partial slabs be modified. 48 * (Note that the total number of slabs is an atomic value that may be 49 * modified without taking the list lock). 50 * 51 * The list_lock is a centralized lock and thus we avoid taking it as 52 * much as possible. As long as SLUB does not have to handle partial 53 * slabs, operations can continue without any centralized lock. F.e. 54 * allocating a long series of objects that fill up slabs does not require 55 * the list lock. 56 * 57 * The lock order is sometimes inverted when we are trying to get a slab 58 * off a list. We take the list_lock and then look for a page on the list 59 * to use. While we do that objects in the slabs may be freed. We can 60 * only operate on the slab if we have also taken the slab_lock. So we use 61 * a slab_trylock() on the slab. If trylock was successful then no frees 62 * can occur anymore and we can use the slab for allocations etc. If the 63 * slab_trylock() does not succeed then frees are in progress in the slab and 64 * we must stay away from it for a while since we may cause a bouncing 65 * cacheline if we try to acquire the lock. So go onto the next slab. 66 * If all pages are busy then we may allocate a new slab instead of reusing 67 * a partial slab. A new slab has noone operating on it and thus there is 68 * no danger of cacheline contention. 69 * 70 * Interrupts are disabled during allocation and deallocation in order to 71 * make the slab allocator safe to use in the context of an irq. In addition 72 * interrupts are disabled to ensure that the processor does not change 73 * while handling per_cpu slabs, due to kernel preemption. 74 * 75 * SLUB assigns one slab for allocation to each processor. 76 * Allocations only occur from these slabs called cpu slabs. 77 * 78 * Slabs with free elements are kept on a partial list and during regular 79 * operations no list for full slabs is used. If an object in a full slab is 80 * freed then the slab will show up again on the partial lists. 81 * We track full slabs for debugging purposes though because otherwise we 82 * cannot scan all objects. 83 * 84 * Slabs are freed when they become empty. Teardown and setup is 85 * minimal so we rely on the page allocators per cpu caches for 86 * fast frees and allocs. 87 * 88 * Overloading of page flags that are otherwise used for LRU management. 89 * 90 * PageActive The slab is frozen and exempt from list processing. 91 * This means that the slab is dedicated to a purpose 92 * such as satisfying allocations for a specific 93 * processor. Objects may be freed in the slab while 94 * it is frozen but slab_free will then skip the usual 95 * list operations. It is up to the processor holding 96 * the slab to integrate the slab into the slab lists 97 * when the slab is no longer needed. 98 * 99 * One use of this flag is to mark slabs that are 100 * used for allocations. Then such a slab becomes a cpu 101 * slab. The cpu slab may be equipped with an additional 102 * freelist that allows lockless access to 103 * free objects in addition to the regular freelist 104 * that requires the slab lock. 105 * 106 * PageError Slab requires special handling due to debug 107 * options set. This moves slab handling out of 108 * the fast path and disables lockless freelists. 109 */ 110 111#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 112 SLAB_TRACE | SLAB_DEBUG_FREE) 113 114static inline int kmem_cache_debug(struct kmem_cache *s) 115{ 116#ifdef CONFIG_SLUB_DEBUG 117 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 118#else 119 return 0; 120#endif 121} 122 123/* 124 * Issues still to be resolved: 125 * 126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 127 * 128 * - Variable sizing of the per node arrays 129 */ 130 131/* Enable to test recovery from slab corruption on boot */ 132#undef SLUB_RESILIENCY_TEST 133 134/* 135 * Mininum number of partial slabs. These will be left on the partial 136 * lists even if they are empty. kmem_cache_shrink may reclaim them. 137 */ 138#define MIN_PARTIAL 5 139 140/* 141 * Maximum number of desirable partial slabs. 142 * The existence of more partial slabs makes kmem_cache_shrink 143 * sort the partial list by the number of objects in the. 144 */ 145#define MAX_PARTIAL 10 146 147#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 148 SLAB_POISON | SLAB_STORE_USER) 149 150/* 151 * Debugging flags that require metadata to be stored in the slab. These get 152 * disabled when slub_debug=O is used and a cache's min order increases with 153 * metadata. 154 */ 155#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 156 157/* 158 * Set of flags that will prevent slab merging 159 */ 160#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ 162 SLAB_FAILSLAB) 163 164#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 165 SLAB_CACHE_DMA | SLAB_NOTRACK) 166 167#define OO_SHIFT 16 168#define OO_MASK ((1 << OO_SHIFT) - 1) 169#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ 170 171/* Internal SLUB flags */ 172#define __OBJECT_POISON 0x80000000UL /* Poison object */ 173 174static int kmem_size = sizeof(struct kmem_cache); 175 176#ifdef CONFIG_SMP 177static struct notifier_block slab_notifier; 178#endif 179 180static enum { 181 DOWN, /* No slab functionality available */ 182 PARTIAL, /* Kmem_cache_node works */ 183 UP, /* Everything works but does not show up in sysfs */ 184 SYSFS /* Sysfs up */ 185} slab_state = DOWN; 186 187/* A list of all slab caches on the system */ 188static DECLARE_RWSEM(slub_lock); 189static LIST_HEAD(slab_caches); 190 191/* 192 * Tracking user of a slab. 193 */ 194struct track { 195 unsigned long addr; /* Called from address */ 196 int cpu; /* Was running on cpu */ 197 int pid; /* Pid context */ 198 unsigned long when; /* When did the operation occur */ 199}; 200 201enum track_item { TRACK_ALLOC, TRACK_FREE }; 202 203#ifdef CONFIG_SYSFS 204static int sysfs_slab_add(struct kmem_cache *); 205static int sysfs_slab_alias(struct kmem_cache *, const char *); 206static void sysfs_slab_remove(struct kmem_cache *); 207 208#else 209static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 210static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 211 { return 0; } 212static inline void sysfs_slab_remove(struct kmem_cache *s) 213{ 214 kfree(s->name); 215 kfree(s); 216} 217 218#endif 219 220static inline void stat(struct kmem_cache *s, enum stat_item si) 221{ 222#ifdef CONFIG_SLUB_STATS 223 __this_cpu_inc(s->cpu_slab->stat[si]); 224#endif 225} 226 227/******************************************************************** 228 * Core slab cache functions 229 *******************************************************************/ 230 231int slab_is_available(void) 232{ 233 return slab_state >= UP; 234} 235 236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 237{ 238 return s->node[node]; 239} 240 241/* Verify that a pointer has an address that is valid within a slab page */ 242static inline int check_valid_pointer(struct kmem_cache *s, 243 struct page *page, const void *object) 244{ 245 void *base; 246 247 if (!object) 248 return 1; 249 250 base = page_address(page); 251 if (object < base || object >= base + page->objects * s->size || 252 (object - base) % s->size) { 253 return 0; 254 } 255 256 return 1; 257} 258 259static inline void *get_freepointer(struct kmem_cache *s, void *object) 260{ 261 return *(void **)(object + s->offset); 262} 263 264static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 265{ 266 *(void **)(object + s->offset) = fp; 267} 268 269/* Loop over all objects in a slab */ 270#define for_each_object(__p, __s, __addr, __objects) \ 271 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 272 __p += (__s)->size) 273 274/* Scan freelist */ 275#define for_each_free_object(__p, __s, __free) \ 276 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 277 278/* Determine object index from a given position */ 279static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 280{ 281 return (p - addr) / s->size; 282} 283 284static inline struct kmem_cache_order_objects oo_make(int order, 285 unsigned long size) 286{ 287 struct kmem_cache_order_objects x = { 288 (order << OO_SHIFT) + (PAGE_SIZE << order) / size 289 }; 290 291 return x; 292} 293 294static inline int oo_order(struct kmem_cache_order_objects x) 295{ 296 return x.x >> OO_SHIFT; 297} 298 299static inline int oo_objects(struct kmem_cache_order_objects x) 300{ 301 return x.x & OO_MASK; 302} 303 304#ifdef CONFIG_SLUB_DEBUG 305/* 306 * Debug settings: 307 */ 308#ifdef CONFIG_SLUB_DEBUG_ON 309static int slub_debug = DEBUG_DEFAULT_FLAGS; 310#else 311static int slub_debug; 312#endif 313 314static char *slub_debug_slabs; 315static int disable_higher_order_debug; 316 317/* 318 * Object debugging 319 */ 320static void print_section(char *text, u8 *addr, unsigned int length) 321{ 322 int i, offset; 323 int newline = 1; 324 char ascii[17]; 325 326 ascii[16] = 0; 327 328 for (i = 0; i < length; i++) { 329 if (newline) { 330 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 331 newline = 0; 332 } 333 printk(KERN_CONT " %02x", addr[i]); 334 offset = i % 16; 335 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 336 if (offset == 15) { 337 printk(KERN_CONT " %s\n", ascii); 338 newline = 1; 339 } 340 } 341 if (!newline) { 342 i %= 16; 343 while (i < 16) { 344 printk(KERN_CONT " "); 345 ascii[i] = ' '; 346 i++; 347 } 348 printk(KERN_CONT " %s\n", ascii); 349 } 350} 351 352static struct track *get_track(struct kmem_cache *s, void *object, 353 enum track_item alloc) 354{ 355 struct track *p; 356 357 if (s->offset) 358 p = object + s->offset + sizeof(void *); 359 else 360 p = object + s->inuse; 361 362 return p + alloc; 363} 364 365static void set_track(struct kmem_cache *s, void *object, 366 enum track_item alloc, unsigned long addr) 367{ 368 struct track *p = get_track(s, object, alloc); 369 370 if (addr) { 371 p->addr = addr; 372 p->cpu = smp_processor_id(); 373 p->pid = current->pid; 374 p->when = jiffies; 375 } else 376 memset(p, 0, sizeof(struct track)); 377} 378 379static void init_tracking(struct kmem_cache *s, void *object) 380{ 381 if (!(s->flags & SLAB_STORE_USER)) 382 return; 383 384 set_track(s, object, TRACK_FREE, 0UL); 385 set_track(s, object, TRACK_ALLOC, 0UL); 386} 387 388static void print_track(const char *s, struct track *t) 389{ 390 if (!t->addr) 391 return; 392 393 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 394 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 395} 396 397static void print_tracking(struct kmem_cache *s, void *object) 398{ 399 if (!(s->flags & SLAB_STORE_USER)) 400 return; 401 402 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 403 print_track("Freed", get_track(s, object, TRACK_FREE)); 404} 405 406static void print_page_info(struct page *page) 407{ 408 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 409 page, page->objects, page->inuse, page->freelist, page->flags); 410 411} 412 413static void slab_bug(struct kmem_cache *s, char *fmt, ...) 414{ 415 va_list args; 416 char buf[100]; 417 418 va_start(args, fmt); 419 vsnprintf(buf, sizeof(buf), fmt, args); 420 va_end(args); 421 printk(KERN_ERR "========================================" 422 "=====================================\n"); 423 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 424 printk(KERN_ERR "----------------------------------------" 425 "-------------------------------------\n\n"); 426} 427 428static void slab_fix(struct kmem_cache *s, char *fmt, ...) 429{ 430 va_list args; 431 char buf[100]; 432 433 va_start(args, fmt); 434 vsnprintf(buf, sizeof(buf), fmt, args); 435 va_end(args); 436 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 437} 438 439static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 440{ 441 unsigned int off; /* Offset of last byte */ 442 u8 *addr = page_address(page); 443 444 print_tracking(s, p); 445 446 print_page_info(page); 447 448 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 449 p, p - addr, get_freepointer(s, p)); 450 451 if (p > addr + 16) 452 print_section("Bytes b4", p - 16, 16); 453 454 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 455 456 if (s->flags & SLAB_RED_ZONE) 457 print_section("Redzone", p + s->objsize, 458 s->inuse - s->objsize); 459 460 if (s->offset) 461 off = s->offset + sizeof(void *); 462 else 463 off = s->inuse; 464 465 if (s->flags & SLAB_STORE_USER) 466 off += 2 * sizeof(struct track); 467 468 if (off != s->size) 469 /* Beginning of the filler is the free pointer */ 470 print_section("Padding", p + off, s->size - off); 471 472 dump_stack(); 473} 474 475static void object_err(struct kmem_cache *s, struct page *page, 476 u8 *object, char *reason) 477{ 478 slab_bug(s, "%s", reason); 479 print_trailer(s, page, object); 480} 481 482static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 483{ 484 va_list args; 485 char buf[100]; 486 487 va_start(args, fmt); 488 vsnprintf(buf, sizeof(buf), fmt, args); 489 va_end(args); 490 slab_bug(s, "%s", buf); 491 print_page_info(page); 492 dump_stack(); 493} 494 495static void init_object(struct kmem_cache *s, void *object, u8 val) 496{ 497 u8 *p = object; 498 499 if (s->flags & __OBJECT_POISON) { 500 memset(p, POISON_FREE, s->objsize - 1); 501 p[s->objsize - 1] = POISON_END; 502 } 503 504 if (s->flags & SLAB_RED_ZONE) 505 memset(p + s->objsize, val, s->inuse - s->objsize); 506} 507 508static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 509{ 510 while (bytes) { 511 if (*start != (u8)value) 512 return start; 513 start++; 514 bytes--; 515 } 516 return NULL; 517} 518 519static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 520 void *from, void *to) 521{ 522 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 523 memset(from, data, to - from); 524} 525 526static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 527 u8 *object, char *what, 528 u8 *start, unsigned int value, unsigned int bytes) 529{ 530 u8 *fault; 531 u8 *end; 532 533 fault = check_bytes(start, value, bytes); 534 if (!fault) 535 return 1; 536 537 end = start + bytes; 538 while (end > fault && end[-1] == value) 539 end--; 540 541 slab_bug(s, "%s overwritten", what); 542 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 543 fault, end - 1, fault[0], value); 544 print_trailer(s, page, object); 545 546 restore_bytes(s, what, value, fault, end); 547 return 0; 548} 549 550/* 551 * Object layout: 552 * 553 * object address 554 * Bytes of the object to be managed. 555 * If the freepointer may overlay the object then the free 556 * pointer is the first word of the object. 557 * 558 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 559 * 0xa5 (POISON_END) 560 * 561 * object + s->objsize 562 * Padding to reach word boundary. This is also used for Redzoning. 563 * Padding is extended by another word if Redzoning is enabled and 564 * objsize == inuse. 565 * 566 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 567 * 0xcc (RED_ACTIVE) for objects in use. 568 * 569 * object + s->inuse 570 * Meta data starts here. 571 * 572 * A. Free pointer (if we cannot overwrite object on free) 573 * B. Tracking data for SLAB_STORE_USER 574 * C. Padding to reach required alignment boundary or at mininum 575 * one word if debugging is on to be able to detect writes 576 * before the word boundary. 577 * 578 * Padding is done using 0x5a (POISON_INUSE) 579 * 580 * object + s->size 581 * Nothing is used beyond s->size. 582 * 583 * If slabcaches are merged then the objsize and inuse boundaries are mostly 584 * ignored. And therefore no slab options that rely on these boundaries 585 * may be used with merged slabcaches. 586 */ 587 588static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 589{ 590 unsigned long off = s->inuse; /* The end of info */ 591 592 if (s->offset) 593 /* Freepointer is placed after the object. */ 594 off += sizeof(void *); 595 596 if (s->flags & SLAB_STORE_USER) 597 /* We also have user information there */ 598 off += 2 * sizeof(struct track); 599 600 if (s->size == off) 601 return 1; 602 603 return check_bytes_and_report(s, page, p, "Object padding", 604 p + off, POISON_INUSE, s->size - off); 605} 606 607/* Check the pad bytes at the end of a slab page */ 608static int slab_pad_check(struct kmem_cache *s, struct page *page) 609{ 610 u8 *start; 611 u8 *fault; 612 u8 *end; 613 int length; 614 int remainder; 615 616 if (!(s->flags & SLAB_POISON)) 617 return 1; 618 619 start = page_address(page); 620 length = (PAGE_SIZE << compound_order(page)); 621 end = start + length; 622 remainder = length % s->size; 623 if (!remainder) 624 return 1; 625 626 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 627 if (!fault) 628 return 1; 629 while (end > fault && end[-1] == POISON_INUSE) 630 end--; 631 632 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 633 print_section("Padding", end - remainder, remainder); 634 635 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 636 return 0; 637} 638 639static int check_object(struct kmem_cache *s, struct page *page, 640 void *object, u8 val) 641{ 642 u8 *p = object; 643 u8 *endobject = object + s->objsize; 644 645 if (s->flags & SLAB_RED_ZONE) { 646 if (!check_bytes_and_report(s, page, object, "Redzone", 647 endobject, val, s->inuse - s->objsize)) 648 return 0; 649 } else { 650 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 651 check_bytes_and_report(s, page, p, "Alignment padding", 652 endobject, POISON_INUSE, s->inuse - s->objsize); 653 } 654 } 655 656 if (s->flags & SLAB_POISON) { 657 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 658 (!check_bytes_and_report(s, page, p, "Poison", p, 659 POISON_FREE, s->objsize - 1) || 660 !check_bytes_and_report(s, page, p, "Poison", 661 p + s->objsize - 1, POISON_END, 1))) 662 return 0; 663 /* 664 * check_pad_bytes cleans up on its own. 665 */ 666 check_pad_bytes(s, page, p); 667 } 668 669 if (!s->offset && val == SLUB_RED_ACTIVE) 670 /* 671 * Object and freepointer overlap. Cannot check 672 * freepointer while object is allocated. 673 */ 674 return 1; 675 676 /* Check free pointer validity */ 677 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 678 object_err(s, page, p, "Freepointer corrupt"); 679 /* 680 * No choice but to zap it and thus lose the remainder 681 * of the free objects in this slab. May cause 682 * another error because the object count is now wrong. 683 */ 684 set_freepointer(s, p, NULL); 685 return 0; 686 } 687 return 1; 688} 689 690static int check_slab(struct kmem_cache *s, struct page *page) 691{ 692 int maxobj; 693 694 VM_BUG_ON(!irqs_disabled()); 695 696 if (!PageSlab(page)) { 697 slab_err(s, page, "Not a valid slab page"); 698 return 0; 699 } 700 701 maxobj = (PAGE_SIZE << compound_order(page)) / s->size; 702 if (page->objects > maxobj) { 703 slab_err(s, page, "objects %u > max %u", 704 s->name, page->objects, maxobj); 705 return 0; 706 } 707 if (page->inuse > page->objects) { 708 slab_err(s, page, "inuse %u > max %u", 709 s->name, page->inuse, page->objects); 710 return 0; 711 } 712 /* Slab_pad_check fixes things up after itself */ 713 slab_pad_check(s, page); 714 return 1; 715} 716 717/* 718 * Determine if a certain object on a page is on the freelist. Must hold the 719 * slab lock to guarantee that the chains are in a consistent state. 720 */ 721static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 722{ 723 int nr = 0; 724 void *fp = page->freelist; 725 void *object = NULL; 726 unsigned long max_objects; 727 728 while (fp && nr <= page->objects) { 729 if (fp == search) 730 return 1; 731 if (!check_valid_pointer(s, page, fp)) { 732 if (object) { 733 object_err(s, page, object, 734 "Freechain corrupt"); 735 set_freepointer(s, object, NULL); 736 break; 737 } else { 738 slab_err(s, page, "Freepointer corrupt"); 739 page->freelist = NULL; 740 page->inuse = page->objects; 741 slab_fix(s, "Freelist cleared"); 742 return 0; 743 } 744 break; 745 } 746 object = fp; 747 fp = get_freepointer(s, object); 748 nr++; 749 } 750 751 max_objects = (PAGE_SIZE << compound_order(page)) / s->size; 752 if (max_objects > MAX_OBJS_PER_PAGE) 753 max_objects = MAX_OBJS_PER_PAGE; 754 755 if (page->objects != max_objects) { 756 slab_err(s, page, "Wrong number of objects. Found %d but " 757 "should be %d", page->objects, max_objects); 758 page->objects = max_objects; 759 slab_fix(s, "Number of objects adjusted."); 760 } 761 if (page->inuse != page->objects - nr) { 762 slab_err(s, page, "Wrong object count. Counter is %d but " 763 "counted were %d", page->inuse, page->objects - nr); 764 page->inuse = page->objects - nr; 765 slab_fix(s, "Object count adjusted."); 766 } 767 return search == NULL; 768} 769 770static void trace(struct kmem_cache *s, struct page *page, void *object, 771 int alloc) 772{ 773 if (s->flags & SLAB_TRACE) { 774 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 775 s->name, 776 alloc ? "alloc" : "free", 777 object, page->inuse, 778 page->freelist); 779 780 if (!alloc) 781 print_section("Object", (void *)object, s->objsize); 782 783 dump_stack(); 784 } 785} 786 787/* 788 * Hooks for other subsystems that check memory allocations. In a typical 789 * production configuration these hooks all should produce no code at all. 790 */ 791static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 792{ 793 flags &= gfp_allowed_mask; 794 lockdep_trace_alloc(flags); 795 might_sleep_if(flags & __GFP_WAIT); 796 797 return should_failslab(s->objsize, flags, s->flags); 798} 799 800static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object) 801{ 802 flags &= gfp_allowed_mask; 803 kmemcheck_slab_alloc(s, flags, object, s->objsize); 804 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags); 805} 806 807static inline void slab_free_hook(struct kmem_cache *s, void *x) 808{ 809 kmemleak_free_recursive(x, s->flags); 810} 811 812static inline void slab_free_hook_irq(struct kmem_cache *s, void *object) 813{ 814 kmemcheck_slab_free(s, object, s->objsize); 815 debug_check_no_locks_freed(object, s->objsize); 816 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 817 debug_check_no_obj_freed(object, s->objsize); 818} 819 820/* 821 * Tracking of fully allocated slabs for debugging purposes. 822 */ 823static void add_full(struct kmem_cache_node *n, struct page *page) 824{ 825 spin_lock(&n->list_lock); 826 list_add(&page->lru, &n->full); 827 spin_unlock(&n->list_lock); 828} 829 830static void remove_full(struct kmem_cache *s, struct page *page) 831{ 832 struct kmem_cache_node *n; 833 834 if (!(s->flags & SLAB_STORE_USER)) 835 return; 836 837 n = get_node(s, page_to_nid(page)); 838 839 spin_lock(&n->list_lock); 840 list_del(&page->lru); 841 spin_unlock(&n->list_lock); 842} 843 844/* Tracking of the number of slabs for debugging purposes */ 845static inline unsigned long slabs_node(struct kmem_cache *s, int node) 846{ 847 struct kmem_cache_node *n = get_node(s, node); 848 849 return atomic_long_read(&n->nr_slabs); 850} 851 852static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 853{ 854 return atomic_long_read(&n->nr_slabs); 855} 856 857static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 858{ 859 struct kmem_cache_node *n = get_node(s, node); 860 861 /* 862 * May be called early in order to allocate a slab for the 863 * kmem_cache_node structure. Solve the chicken-egg 864 * dilemma by deferring the increment of the count during 865 * bootstrap (see early_kmem_cache_node_alloc). 866 */ 867 if (n) { 868 atomic_long_inc(&n->nr_slabs); 869 atomic_long_add(objects, &n->total_objects); 870 } 871} 872static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 873{ 874 struct kmem_cache_node *n = get_node(s, node); 875 876 atomic_long_dec(&n->nr_slabs); 877 atomic_long_sub(objects, &n->total_objects); 878} 879 880/* Object debug checks for alloc/free paths */ 881static void setup_object_debug(struct kmem_cache *s, struct page *page, 882 void *object) 883{ 884 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 885 return; 886 887 init_object(s, object, SLUB_RED_INACTIVE); 888 init_tracking(s, object); 889} 890 891static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page, 892 void *object, unsigned long addr) 893{ 894 if (!check_slab(s, page)) 895 goto bad; 896 897 if (!on_freelist(s, page, object)) { 898 object_err(s, page, object, "Object already allocated"); 899 goto bad; 900 } 901 902 if (!check_valid_pointer(s, page, object)) { 903 object_err(s, page, object, "Freelist Pointer check fails"); 904 goto bad; 905 } 906 907 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 908 goto bad; 909 910 /* Success perform special debug activities for allocs */ 911 if (s->flags & SLAB_STORE_USER) 912 set_track(s, object, TRACK_ALLOC, addr); 913 trace(s, page, object, 1); 914 init_object(s, object, SLUB_RED_ACTIVE); 915 return 1; 916 917bad: 918 if (PageSlab(page)) { 919 /* 920 * If this is a slab page then lets do the best we can 921 * to avoid issues in the future. Marking all objects 922 * as used avoids touching the remaining objects. 923 */ 924 slab_fix(s, "Marking all objects used"); 925 page->inuse = page->objects; 926 page->freelist = NULL; 927 } 928 return 0; 929} 930 931static noinline int free_debug_processing(struct kmem_cache *s, 932 struct page *page, void *object, unsigned long addr) 933{ 934 if (!check_slab(s, page)) 935 goto fail; 936 937 if (!check_valid_pointer(s, page, object)) { 938 slab_err(s, page, "Invalid object pointer 0x%p", object); 939 goto fail; 940 } 941 942 if (on_freelist(s, page, object)) { 943 object_err(s, page, object, "Object already free"); 944 goto fail; 945 } 946 947 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 948 return 0; 949 950 if (unlikely(s != page->slab)) { 951 if (!PageSlab(page)) { 952 slab_err(s, page, "Attempt to free object(0x%p) " 953 "outside of slab", object); 954 } else if (!page->slab) { 955 printk(KERN_ERR 956 "SLUB <none>: no slab for object 0x%p.\n", 957 object); 958 dump_stack(); 959 } else 960 object_err(s, page, object, 961 "page slab pointer corrupt."); 962 goto fail; 963 } 964 965 /* Special debug activities for freeing objects */ 966 if (!PageSlubFrozen(page) && !page->freelist) 967 remove_full(s, page); 968 if (s->flags & SLAB_STORE_USER) 969 set_track(s, object, TRACK_FREE, addr); 970 trace(s, page, object, 0); 971 init_object(s, object, SLUB_RED_INACTIVE); 972 return 1; 973 974fail: 975 slab_fix(s, "Object at 0x%p not freed", object); 976 return 0; 977} 978 979static int __init setup_slub_debug(char *str) 980{ 981 slub_debug = DEBUG_DEFAULT_FLAGS; 982 if (*str++ != '=' || !*str) 983 /* 984 * No options specified. Switch on full debugging. 985 */ 986 goto out; 987 988 if (*str == ',') 989 /* 990 * No options but restriction on slabs. This means full 991 * debugging for slabs matching a pattern. 992 */ 993 goto check_slabs; 994 995 if (tolower(*str) == 'o') { 996 /* 997 * Avoid enabling debugging on caches if its minimum order 998 * would increase as a result. 999 */ 1000 disable_higher_order_debug = 1; 1001 goto out; 1002 } 1003 1004 slub_debug = 0; 1005 if (*str == '-') 1006 /* 1007 * Switch off all debugging measures. 1008 */ 1009 goto out; 1010 1011 /* 1012 * Determine which debug features should be switched on 1013 */ 1014 for (; *str && *str != ','; str++) { 1015 switch (tolower(*str)) { 1016 case 'f': 1017 slub_debug |= SLAB_DEBUG_FREE; 1018 break; 1019 case 'z': 1020 slub_debug |= SLAB_RED_ZONE; 1021 break; 1022 case 'p': 1023 slub_debug |= SLAB_POISON; 1024 break; 1025 case 'u': 1026 slub_debug |= SLAB_STORE_USER; 1027 break; 1028 case 't': 1029 slub_debug |= SLAB_TRACE; 1030 break; 1031 case 'a': 1032 slub_debug |= SLAB_FAILSLAB; 1033 break; 1034 default: 1035 printk(KERN_ERR "slub_debug option '%c' " 1036 "unknown. skipped\n", *str); 1037 } 1038 } 1039 1040check_slabs: 1041 if (*str == ',') 1042 slub_debug_slabs = str + 1; 1043out: 1044 return 1; 1045} 1046 1047__setup("slub_debug", setup_slub_debug); 1048 1049static unsigned long kmem_cache_flags(unsigned long objsize, 1050 unsigned long flags, const char *name, 1051 void (*ctor)(void *)) 1052{ 1053 /* 1054 * Enable debugging if selected on the kernel commandline. 1055 */ 1056 if (slub_debug && (!slub_debug_slabs || 1057 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) 1058 flags |= slub_debug; 1059 1060 return flags; 1061} 1062#else 1063static inline void setup_object_debug(struct kmem_cache *s, 1064 struct page *page, void *object) {} 1065 1066static inline int alloc_debug_processing(struct kmem_cache *s, 1067 struct page *page, void *object, unsigned long addr) { return 0; } 1068 1069static inline int free_debug_processing(struct kmem_cache *s, 1070 struct page *page, void *object, unsigned long addr) { return 0; } 1071 1072static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1073 { return 1; } 1074static inline int check_object(struct kmem_cache *s, struct page *page, 1075 void *object, u8 val) { return 1; } 1076static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1077static inline unsigned long kmem_cache_flags(unsigned long objsize, 1078 unsigned long flags, const char *name, 1079 void (*ctor)(void *)) 1080{ 1081 return flags; 1082} 1083#define slub_debug 0 1084 1085#define disable_higher_order_debug 0 1086 1087static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1088 { return 0; } 1089static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1090 { return 0; } 1091static inline void inc_slabs_node(struct kmem_cache *s, int node, 1092 int objects) {} 1093static inline void dec_slabs_node(struct kmem_cache *s, int node, 1094 int objects) {} 1095 1096static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 1097 { return 0; } 1098 1099static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, 1100 void *object) {} 1101 1102static inline void slab_free_hook(struct kmem_cache *s, void *x) {} 1103 1104static inline void slab_free_hook_irq(struct kmem_cache *s, 1105 void *object) {} 1106 1107#endif /* CONFIG_SLUB_DEBUG */ 1108 1109/* 1110 * Slab allocation and freeing 1111 */ 1112static inline struct page *alloc_slab_page(gfp_t flags, int node, 1113 struct kmem_cache_order_objects oo) 1114{ 1115 int order = oo_order(oo); 1116 1117 flags |= __GFP_NOTRACK; 1118 1119 if (node == NUMA_NO_NODE) 1120 return alloc_pages(flags, order); 1121 else 1122 return alloc_pages_exact_node(node, flags, order); 1123} 1124 1125static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1126{ 1127 struct page *page; 1128 struct kmem_cache_order_objects oo = s->oo; 1129 gfp_t alloc_gfp; 1130 1131 flags |= s->allocflags; 1132 1133 /* 1134 * Let the initial higher-order allocation fail under memory pressure 1135 * so we fall-back to the minimum order allocation. 1136 */ 1137 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1138 1139 page = alloc_slab_page(alloc_gfp, node, oo); 1140 if (unlikely(!page)) { 1141 oo = s->min; 1142 /* 1143 * Allocation may have failed due to fragmentation. 1144 * Try a lower order alloc if possible 1145 */ 1146 page = alloc_slab_page(flags, node, oo); 1147 if (!page) 1148 return NULL; 1149 1150 stat(s, ORDER_FALLBACK); 1151 } 1152 1153 if (kmemcheck_enabled 1154 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1155 int pages = 1 << oo_order(oo); 1156 1157 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); 1158 1159 /* 1160 * Objects from caches that have a constructor don't get 1161 * cleared when they're allocated, so we need to do it here. 1162 */ 1163 if (s->ctor) 1164 kmemcheck_mark_uninitialized_pages(page, pages); 1165 else 1166 kmemcheck_mark_unallocated_pages(page, pages); 1167 } 1168 1169 page->objects = oo_objects(oo); 1170 mod_zone_page_state(page_zone(page), 1171 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1172 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1173 1 << oo_order(oo)); 1174 1175 return page; 1176} 1177 1178static void setup_object(struct kmem_cache *s, struct page *page, 1179 void *object) 1180{ 1181 setup_object_debug(s, page, object); 1182 if (unlikely(s->ctor)) 1183 s->ctor(object); 1184} 1185 1186static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1187{ 1188 struct page *page; 1189 void *start; 1190 void *last; 1191 void *p; 1192 1193 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1194 1195 page = allocate_slab(s, 1196 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1197 if (!page) 1198 goto out; 1199 1200 inc_slabs_node(s, page_to_nid(page), page->objects); 1201 page->slab = s; 1202 page->flags |= 1 << PG_slab; 1203 1204 start = page_address(page); 1205 1206 if (unlikely(s->flags & SLAB_POISON)) 1207 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1208 1209 last = start; 1210 for_each_object(p, s, start, page->objects) { 1211 setup_object(s, page, last); 1212 set_freepointer(s, last, p); 1213 last = p; 1214 } 1215 setup_object(s, page, last); 1216 set_freepointer(s, last, NULL); 1217 1218 page->freelist = start; 1219 page->inuse = 0; 1220out: 1221 return page; 1222} 1223 1224static void __free_slab(struct kmem_cache *s, struct page *page) 1225{ 1226 int order = compound_order(page); 1227 int pages = 1 << order; 1228 1229 if (kmem_cache_debug(s)) { 1230 void *p; 1231 1232 slab_pad_check(s, page); 1233 for_each_object(p, s, page_address(page), 1234 page->objects) 1235 check_object(s, page, p, SLUB_RED_INACTIVE); 1236 } 1237 1238 kmemcheck_free_shadow(page, compound_order(page)); 1239 1240 mod_zone_page_state(page_zone(page), 1241 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1242 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1243 -pages); 1244 1245 __ClearPageSlab(page); 1246 reset_page_mapcount(page); 1247 if (current->reclaim_state) 1248 current->reclaim_state->reclaimed_slab += pages; 1249 __free_pages(page, order); 1250} 1251 1252static void rcu_free_slab(struct rcu_head *h) 1253{ 1254 struct page *page; 1255 1256 page = container_of((struct list_head *)h, struct page, lru); 1257 __free_slab(page->slab, page); 1258} 1259 1260static void free_slab(struct kmem_cache *s, struct page *page) 1261{ 1262 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1263 /* 1264 * RCU free overloads the RCU head over the LRU 1265 */ 1266 struct rcu_head *head = (void *)&page->lru; 1267 1268 call_rcu(head, rcu_free_slab); 1269 } else 1270 __free_slab(s, page); 1271} 1272 1273static void discard_slab(struct kmem_cache *s, struct page *page) 1274{ 1275 dec_slabs_node(s, page_to_nid(page), page->objects); 1276 free_slab(s, page); 1277} 1278 1279/* 1280 * Per slab locking using the pagelock 1281 */ 1282static __always_inline void slab_lock(struct page *page) 1283{ 1284 bit_spin_lock(PG_locked, &page->flags); 1285} 1286 1287static __always_inline void slab_unlock(struct page *page) 1288{ 1289 __bit_spin_unlock(PG_locked, &page->flags); 1290} 1291 1292static __always_inline int slab_trylock(struct page *page) 1293{ 1294 int rc = 1; 1295 1296 rc = bit_spin_trylock(PG_locked, &page->flags); 1297 return rc; 1298} 1299 1300/* 1301 * Management of partially allocated slabs 1302 */ 1303static void add_partial(struct kmem_cache_node *n, 1304 struct page *page, int tail) 1305{ 1306 spin_lock(&n->list_lock); 1307 n->nr_partial++; 1308 if (tail) 1309 list_add_tail(&page->lru, &n->partial); 1310 else 1311 list_add(&page->lru, &n->partial); 1312 spin_unlock(&n->list_lock); 1313} 1314 1315static inline void __remove_partial(struct kmem_cache_node *n, 1316 struct page *page) 1317{ 1318 list_del(&page->lru); 1319 n->nr_partial--; 1320} 1321 1322static void remove_partial(struct kmem_cache *s, struct page *page) 1323{ 1324 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1325 1326 spin_lock(&n->list_lock); 1327 __remove_partial(n, page); 1328 spin_unlock(&n->list_lock); 1329} 1330 1331/* 1332 * Lock slab and remove from the partial list. 1333 * 1334 * Must hold list_lock. 1335 */ 1336static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1337 struct page *page) 1338{ 1339 if (slab_trylock(page)) { 1340 __remove_partial(n, page); 1341 __SetPageSlubFrozen(page); 1342 return 1; 1343 } 1344 return 0; 1345} 1346 1347/* 1348 * Try to allocate a partial slab from a specific node. 1349 */ 1350static struct page *get_partial_node(struct kmem_cache_node *n) 1351{ 1352 struct page *page; 1353 1354 /* 1355 * Racy check. If we mistakenly see no partial slabs then we 1356 * just allocate an empty slab. If we mistakenly try to get a 1357 * partial slab and there is none available then get_partials() 1358 * will return NULL. 1359 */ 1360 if (!n || !n->nr_partial) 1361 return NULL; 1362 1363 spin_lock(&n->list_lock); 1364 list_for_each_entry(page, &n->partial, lru) 1365 if (lock_and_freeze_slab(n, page)) 1366 goto out; 1367 page = NULL; 1368out: 1369 spin_unlock(&n->list_lock); 1370 return page; 1371} 1372 1373/* 1374 * Get a page from somewhere. Search in increasing NUMA distances. 1375 */ 1376static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1377{ 1378#ifdef CONFIG_NUMA 1379 struct zonelist *zonelist; 1380 struct zoneref *z; 1381 struct zone *zone; 1382 enum zone_type high_zoneidx = gfp_zone(flags); 1383 struct page *page; 1384 1385 /* 1386 * The defrag ratio allows a configuration of the tradeoffs between 1387 * inter node defragmentation and node local allocations. A lower 1388 * defrag_ratio increases the tendency to do local allocations 1389 * instead of attempting to obtain partial slabs from other nodes. 1390 * 1391 * If the defrag_ratio is set to 0 then kmalloc() always 1392 * returns node local objects. If the ratio is higher then kmalloc() 1393 * may return off node objects because partial slabs are obtained 1394 * from other nodes and filled up. 1395 * 1396 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1397 * defrag_ratio = 1000) then every (well almost) allocation will 1398 * first attempt to defrag slab caches on other nodes. This means 1399 * scanning over all nodes to look for partial slabs which may be 1400 * expensive if we do it every time we are trying to find a slab 1401 * with available objects. 1402 */ 1403 if (!s->remote_node_defrag_ratio || 1404 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1405 return NULL; 1406 1407 get_mems_allowed(); 1408 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1409 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1410 struct kmem_cache_node *n; 1411 1412 n = get_node(s, zone_to_nid(zone)); 1413 1414 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1415 n->nr_partial > s->min_partial) { 1416 page = get_partial_node(n); 1417 if (page) { 1418 put_mems_allowed(); 1419 return page; 1420 } 1421 } 1422 } 1423 put_mems_allowed(); 1424#endif 1425 return NULL; 1426} 1427 1428/* 1429 * Get a partial page, lock it and return it. 1430 */ 1431static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1432{ 1433 struct page *page; 1434 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; 1435 1436 page = get_partial_node(get_node(s, searchnode)); 1437 if (page || node != -1) 1438 return page; 1439 1440 return get_any_partial(s, flags); 1441} 1442 1443/* 1444 * Move a page back to the lists. 1445 * 1446 * Must be called with the slab lock held. 1447 * 1448 * On exit the slab lock will have been dropped. 1449 */ 1450static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1451 __releases(bitlock) 1452{ 1453 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1454 1455 __ClearPageSlubFrozen(page); 1456 if (page->inuse) { 1457 1458 if (page->freelist) { 1459 add_partial(n, page, tail); 1460 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1461 } else { 1462 stat(s, DEACTIVATE_FULL); 1463 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER)) 1464 add_full(n, page); 1465 } 1466 slab_unlock(page); 1467 } else { 1468 stat(s, DEACTIVATE_EMPTY); 1469 if (n->nr_partial < s->min_partial) { 1470 /* 1471 * Adding an empty slab to the partial slabs in order 1472 * to avoid page allocator overhead. This slab needs 1473 * to come after the other slabs with objects in 1474 * so that the others get filled first. That way the 1475 * size of the partial list stays small. 1476 * 1477 * kmem_cache_shrink can reclaim any empty slabs from 1478 * the partial list. 1479 */ 1480 add_partial(n, page, 1); 1481 slab_unlock(page); 1482 } else { 1483 slab_unlock(page); 1484 stat(s, FREE_SLAB); 1485 discard_slab(s, page); 1486 } 1487 } 1488} 1489 1490/* 1491 * Remove the cpu slab 1492 */ 1493static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1494 __releases(bitlock) 1495{ 1496 struct page *page = c->page; 1497 int tail = 1; 1498 1499 if (page->freelist) 1500 stat(s, DEACTIVATE_REMOTE_FREES); 1501 /* 1502 * Merge cpu freelist into slab freelist. Typically we get here 1503 * because both freelists are empty. So this is unlikely 1504 * to occur. 1505 */ 1506 while (unlikely(c->freelist)) { 1507 void **object; 1508 1509 tail = 0; /* Hot objects. Put the slab first */ 1510 1511 /* Retrieve object from cpu_freelist */ 1512 object = c->freelist; 1513 c->freelist = get_freepointer(s, c->freelist); 1514 1515 /* And put onto the regular freelist */ 1516 set_freepointer(s, object, page->freelist); 1517 page->freelist = object; 1518 page->inuse--; 1519 } 1520 c->page = NULL; 1521 unfreeze_slab(s, page, tail); 1522} 1523 1524static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1525{ 1526 stat(s, CPUSLAB_FLUSH); 1527 slab_lock(c->page); 1528 deactivate_slab(s, c); 1529} 1530 1531/* 1532 * Flush cpu slab. 1533 * 1534 * Called from IPI handler with interrupts disabled. 1535 */ 1536static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1537{ 1538 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 1539 1540 if (likely(c && c->page)) 1541 flush_slab(s, c); 1542} 1543 1544static void flush_cpu_slab(void *d) 1545{ 1546 struct kmem_cache *s = d; 1547 1548 __flush_cpu_slab(s, smp_processor_id()); 1549} 1550 1551static void flush_all(struct kmem_cache *s) 1552{ 1553 on_each_cpu(flush_cpu_slab, s, 1); 1554} 1555 1556/* 1557 * Check if the objects in a per cpu structure fit numa 1558 * locality expectations. 1559 */ 1560static inline int node_match(struct kmem_cache_cpu *c, int node) 1561{ 1562#ifdef CONFIG_NUMA 1563 if (node != NUMA_NO_NODE && c->node != node) 1564 return 0; 1565#endif 1566 return 1; 1567} 1568 1569static int count_free(struct page *page) 1570{ 1571 return page->objects - page->inuse; 1572} 1573 1574static unsigned long count_partial(struct kmem_cache_node *n, 1575 int (*get_count)(struct page *)) 1576{ 1577 unsigned long flags; 1578 unsigned long x = 0; 1579 struct page *page; 1580 1581 spin_lock_irqsave(&n->list_lock, flags); 1582 list_for_each_entry(page, &n->partial, lru) 1583 x += get_count(page); 1584 spin_unlock_irqrestore(&n->list_lock, flags); 1585 return x; 1586} 1587 1588static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 1589{ 1590#ifdef CONFIG_SLUB_DEBUG 1591 return atomic_long_read(&n->total_objects); 1592#else 1593 return 0; 1594#endif 1595} 1596 1597static noinline void 1598slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 1599{ 1600 int node; 1601 1602 printk(KERN_WARNING 1603 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 1604 nid, gfpflags); 1605 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 1606 "default order: %d, min order: %d\n", s->name, s->objsize, 1607 s->size, oo_order(s->oo), oo_order(s->min)); 1608 1609 if (oo_order(s->min) > get_order(s->objsize)) 1610 printk(KERN_WARNING " %s debugging increased min order, use " 1611 "slub_debug=O to disable.\n", s->name); 1612 1613 for_each_online_node(node) { 1614 struct kmem_cache_node *n = get_node(s, node); 1615 unsigned long nr_slabs; 1616 unsigned long nr_objs; 1617 unsigned long nr_free; 1618 1619 if (!n) 1620 continue; 1621 1622 nr_free = count_partial(n, count_free); 1623 nr_slabs = node_nr_slabs(n); 1624 nr_objs = node_nr_objs(n); 1625 1626 printk(KERN_WARNING 1627 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 1628 node, nr_slabs, nr_objs, nr_free); 1629 } 1630} 1631 1632/* 1633 * Slow path. The lockless freelist is empty or we need to perform 1634 * debugging duties. 1635 * 1636 * Interrupts are disabled. 1637 * 1638 * Processing is still very fast if new objects have been freed to the 1639 * regular freelist. In that case we simply take over the regular freelist 1640 * as the lockless freelist and zap the regular freelist. 1641 * 1642 * If that is not working then we fall back to the partial lists. We take the 1643 * first element of the freelist as the object to allocate now and move the 1644 * rest of the freelist to the lockless freelist. 1645 * 1646 * And if we were unable to get a new slab from the partial slab lists then 1647 * we need to allocate a new slab. This is the slowest path since it involves 1648 * a call to the page allocator and the setup of a new slab. 1649 */ 1650static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 1651 unsigned long addr, struct kmem_cache_cpu *c) 1652{ 1653 void **object; 1654 struct page *new; 1655 1656 /* We handle __GFP_ZERO in the caller */ 1657 gfpflags &= ~__GFP_ZERO; 1658 1659 if (!c->page) 1660 goto new_slab; 1661 1662 slab_lock(c->page); 1663 if (unlikely(!node_match(c, node))) 1664 goto another_slab; 1665 1666 stat(s, ALLOC_REFILL); 1667 1668load_freelist: 1669 object = c->page->freelist; 1670 if (unlikely(!object)) 1671 goto another_slab; 1672 if (kmem_cache_debug(s)) 1673 goto debug; 1674 1675 c->freelist = get_freepointer(s, object); 1676 c->page->inuse = c->page->objects; 1677 c->page->freelist = NULL; 1678 c->node = page_to_nid(c->page); 1679unlock_out: 1680 slab_unlock(c->page); 1681 stat(s, ALLOC_SLOWPATH); 1682 return object; 1683 1684another_slab: 1685 deactivate_slab(s, c); 1686 1687new_slab: 1688 new = get_partial(s, gfpflags, node); 1689 if (new) { 1690 c->page = new; 1691 stat(s, ALLOC_FROM_PARTIAL); 1692 goto load_freelist; 1693 } 1694 1695 gfpflags &= gfp_allowed_mask; 1696 if (gfpflags & __GFP_WAIT) 1697 local_irq_enable(); 1698 1699 new = new_slab(s, gfpflags, node); 1700 1701 if (gfpflags & __GFP_WAIT) 1702 local_irq_disable(); 1703 1704 if (new) { 1705 c = __this_cpu_ptr(s->cpu_slab); 1706 stat(s, ALLOC_SLAB); 1707 if (c->page) 1708 flush_slab(s, c); 1709 slab_lock(new); 1710 __SetPageSlubFrozen(new); 1711 c->page = new; 1712 goto load_freelist; 1713 } 1714 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 1715 slab_out_of_memory(s, gfpflags, node); 1716 return NULL; 1717debug: 1718 if (!alloc_debug_processing(s, c->page, object, addr)) 1719 goto another_slab; 1720 1721 c->page->inuse++; 1722 c->page->freelist = get_freepointer(s, object); 1723 c->node = NUMA_NO_NODE; 1724 goto unlock_out; 1725} 1726 1727/* 1728 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1729 * have the fastpath folded into their functions. So no function call 1730 * overhead for requests that can be satisfied on the fastpath. 1731 * 1732 * The fastpath works by first checking if the lockless freelist can be used. 1733 * If not then __slab_alloc is called for slow processing. 1734 * 1735 * Otherwise we can simply pick the next object from the lockless free list. 1736 */ 1737static __always_inline void *slab_alloc(struct kmem_cache *s, 1738 gfp_t gfpflags, int node, unsigned long addr) 1739{ 1740 void **object; 1741 struct kmem_cache_cpu *c; 1742 unsigned long flags; 1743 1744 if (slab_pre_alloc_hook(s, gfpflags)) 1745 return NULL; 1746 1747 local_irq_save(flags); 1748 c = __this_cpu_ptr(s->cpu_slab); 1749 object = c->freelist; 1750 if (unlikely(!object || !node_match(c, node))) 1751 1752 object = __slab_alloc(s, gfpflags, node, addr, c); 1753 1754 else { 1755 c->freelist = get_freepointer(s, object); 1756 stat(s, ALLOC_FASTPATH); 1757 } 1758 local_irq_restore(flags); 1759 1760 if (unlikely(gfpflags & __GFP_ZERO) && object) 1761 memset(object, 0, s->objsize); 1762 1763 slab_post_alloc_hook(s, gfpflags, object); 1764 1765 return object; 1766} 1767 1768void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1769{ 1770 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 1771 1772 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); 1773 1774 return ret; 1775} 1776EXPORT_SYMBOL(kmem_cache_alloc); 1777 1778#ifdef CONFIG_TRACING 1779void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 1780{ 1781 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 1782 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 1783 return ret; 1784} 1785EXPORT_SYMBOL(kmem_cache_alloc_trace); 1786 1787void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) 1788{ 1789 void *ret = kmalloc_order(size, flags, order); 1790 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); 1791 return ret; 1792} 1793EXPORT_SYMBOL(kmalloc_order_trace); 1794#endif 1795 1796#ifdef CONFIG_NUMA 1797void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1798{ 1799 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 1800 1801 trace_kmem_cache_alloc_node(_RET_IP_, ret, 1802 s->objsize, s->size, gfpflags, node); 1803 1804 return ret; 1805} 1806EXPORT_SYMBOL(kmem_cache_alloc_node); 1807 1808#ifdef CONFIG_TRACING 1809void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 1810 gfp_t gfpflags, 1811 int node, size_t size) 1812{ 1813 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 1814 1815 trace_kmalloc_node(_RET_IP_, ret, 1816 size, s->size, gfpflags, node); 1817 return ret; 1818} 1819EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 1820#endif 1821#endif 1822 1823/* 1824 * Slow patch handling. This may still be called frequently since objects 1825 * have a longer lifetime than the cpu slabs in most processing loads. 1826 * 1827 * So we still attempt to reduce cache line usage. Just take the slab 1828 * lock and free the item. If there is no additional partial page 1829 * handling required then we can return immediately. 1830 */ 1831static void __slab_free(struct kmem_cache *s, struct page *page, 1832 void *x, unsigned long addr) 1833{ 1834 void *prior; 1835 void **object = (void *)x; 1836 1837 stat(s, FREE_SLOWPATH); 1838 slab_lock(page); 1839 1840 if (kmem_cache_debug(s)) 1841 goto debug; 1842 1843checks_ok: 1844 prior = page->freelist; 1845 set_freepointer(s, object, prior); 1846 page->freelist = object; 1847 page->inuse--; 1848 1849 if (unlikely(PageSlubFrozen(page))) { 1850 stat(s, FREE_FROZEN); 1851 goto out_unlock; 1852 } 1853 1854 if (unlikely(!page->inuse)) 1855 goto slab_empty; 1856 1857 /* 1858 * Objects left in the slab. If it was not on the partial list before 1859 * then add it. 1860 */ 1861 if (unlikely(!prior)) { 1862 add_partial(get_node(s, page_to_nid(page)), page, 1); 1863 stat(s, FREE_ADD_PARTIAL); 1864 } 1865 1866out_unlock: 1867 slab_unlock(page); 1868 return; 1869 1870slab_empty: 1871 if (prior) { 1872 /* 1873 * Slab still on the partial list. 1874 */ 1875 remove_partial(s, page); 1876 stat(s, FREE_REMOVE_PARTIAL); 1877 } 1878 slab_unlock(page); 1879 stat(s, FREE_SLAB); 1880 discard_slab(s, page); 1881 return; 1882 1883debug: 1884 if (!free_debug_processing(s, page, x, addr)) 1885 goto out_unlock; 1886 goto checks_ok; 1887} 1888 1889/* 1890 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1891 * can perform fastpath freeing without additional function calls. 1892 * 1893 * The fastpath is only possible if we are freeing to the current cpu slab 1894 * of this processor. This typically the case if we have just allocated 1895 * the item before. 1896 * 1897 * If fastpath is not possible then fall back to __slab_free where we deal 1898 * with all sorts of special processing. 1899 */ 1900static __always_inline void slab_free(struct kmem_cache *s, 1901 struct page *page, void *x, unsigned long addr) 1902{ 1903 void **object = (void *)x; 1904 struct kmem_cache_cpu *c; 1905 unsigned long flags; 1906 1907 slab_free_hook(s, x); 1908 1909 local_irq_save(flags); 1910 c = __this_cpu_ptr(s->cpu_slab); 1911 1912 slab_free_hook_irq(s, x); 1913 1914 if (likely(page == c->page && c->node != NUMA_NO_NODE)) { 1915 set_freepointer(s, object, c->freelist); 1916 c->freelist = object; 1917 stat(s, FREE_FASTPATH); 1918 } else 1919 __slab_free(s, page, x, addr); 1920 1921 local_irq_restore(flags); 1922} 1923 1924void kmem_cache_free(struct kmem_cache *s, void *x) 1925{ 1926 struct page *page; 1927 1928 page = virt_to_head_page(x); 1929 1930 slab_free(s, page, x, _RET_IP_); 1931 1932 trace_kmem_cache_free(_RET_IP_, x); 1933} 1934EXPORT_SYMBOL(kmem_cache_free); 1935 1936/* Figure out on which slab page the object resides */ 1937static struct page *get_object_page(const void *x) 1938{ 1939 struct page *page = virt_to_head_page(x); 1940 1941 if (!PageSlab(page)) 1942 return NULL; 1943 1944 return page; 1945} 1946 1947/* 1948 * Object placement in a slab is made very easy because we always start at 1949 * offset 0. If we tune the size of the object to the alignment then we can 1950 * get the required alignment by putting one properly sized object after 1951 * another. 1952 * 1953 * Notice that the allocation order determines the sizes of the per cpu 1954 * caches. Each processor has always one slab available for allocations. 1955 * Increasing the allocation order reduces the number of times that slabs 1956 * must be moved on and off the partial lists and is therefore a factor in 1957 * locking overhead. 1958 */ 1959 1960/* 1961 * Mininum / Maximum order of slab pages. This influences locking overhead 1962 * and slab fragmentation. A higher order reduces the number of partial slabs 1963 * and increases the number of allocations possible without having to 1964 * take the list_lock. 1965 */ 1966static int slub_min_order; 1967static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 1968static int slub_min_objects; 1969 1970/* 1971 * Merge control. If this is set then no merging of slab caches will occur. 1972 * (Could be removed. This was introduced to pacify the merge skeptics.) 1973 */ 1974static int slub_nomerge; 1975 1976/* 1977 * Calculate the order of allocation given an slab object size. 1978 * 1979 * The order of allocation has significant impact on performance and other 1980 * system components. Generally order 0 allocations should be preferred since 1981 * order 0 does not cause fragmentation in the page allocator. Larger objects 1982 * be problematic to put into order 0 slabs because there may be too much 1983 * unused space left. We go to a higher order if more than 1/16th of the slab 1984 * would be wasted. 1985 * 1986 * In order to reach satisfactory performance we must ensure that a minimum 1987 * number of objects is in one slab. Otherwise we may generate too much 1988 * activity on the partial lists which requires taking the list_lock. This is 1989 * less a concern for large slabs though which are rarely used. 1990 * 1991 * slub_max_order specifies the order where we begin to stop considering the 1992 * number of objects in a slab as critical. If we reach slub_max_order then 1993 * we try to keep the page order as low as possible. So we accept more waste 1994 * of space in favor of a small page order. 1995 * 1996 * Higher order allocations also allow the placement of more objects in a 1997 * slab and thereby reduce object handling overhead. If the user has 1998 * requested a higher mininum order then we start with that one instead of 1999 * the smallest order which will fit the object. 2000 */ 2001static inline int slab_order(int size, int min_objects, 2002 int max_order, int fract_leftover) 2003{ 2004 int order; 2005 int rem; 2006 int min_order = slub_min_order; 2007 2008 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE) 2009 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 2010 2011 for (order = max(min_order, 2012 fls(min_objects * size - 1) - PAGE_SHIFT); 2013 order <= max_order; order++) { 2014 2015 unsigned long slab_size = PAGE_SIZE << order; 2016 2017 if (slab_size < min_objects * size) 2018 continue; 2019 2020 rem = slab_size % size; 2021 2022 if (rem <= slab_size / fract_leftover) 2023 break; 2024 2025 } 2026 2027 return order; 2028} 2029 2030static inline int calculate_order(int size) 2031{ 2032 int order; 2033 int min_objects; 2034 int fraction; 2035 int max_objects; 2036 2037 /* 2038 * Attempt to find best configuration for a slab. This 2039 * works by first attempting to generate a layout with 2040 * the best configuration and backing off gradually. 2041 * 2042 * First we reduce the acceptable waste in a slab. Then 2043 * we reduce the minimum objects required in a slab. 2044 */ 2045 min_objects = slub_min_objects; 2046 if (!min_objects) 2047 min_objects = 4 * (fls(nr_cpu_ids) + 1); 2048 max_objects = (PAGE_SIZE << slub_max_order)/size; 2049 min_objects = min(min_objects, max_objects); 2050 2051 while (min_objects > 1) { 2052 fraction = 16; 2053 while (fraction >= 4) { 2054 order = slab_order(size, min_objects, 2055 slub_max_order, fraction); 2056 if (order <= slub_max_order) 2057 return order; 2058 fraction /= 2; 2059 } 2060 min_objects--; 2061 } 2062 2063 /* 2064 * We were unable to place multiple objects in a slab. Now 2065 * lets see if we can place a single object there. 2066 */ 2067 order = slab_order(size, 1, slub_max_order, 1); 2068 if (order <= slub_max_order) 2069 return order; 2070 2071 /* 2072 * Doh this slab cannot be placed using slub_max_order. 2073 */ 2074 order = slab_order(size, 1, MAX_ORDER, 1); 2075 if (order < MAX_ORDER) 2076 return order; 2077 return -ENOSYS; 2078} 2079 2080/* 2081 * Figure out what the alignment of the objects will be. 2082 */ 2083static unsigned long calculate_alignment(unsigned long flags, 2084 unsigned long align, unsigned long size) 2085{ 2086 /* 2087 * If the user wants hardware cache aligned objects then follow that 2088 * suggestion if the object is sufficiently large. 2089 * 2090 * The hardware cache alignment cannot override the specified 2091 * alignment though. If that is greater then use it. 2092 */ 2093 if (flags & SLAB_HWCACHE_ALIGN) { 2094 unsigned long ralign = cache_line_size(); 2095 while (size <= ralign / 2) 2096 ralign /= 2; 2097 align = max(align, ralign); 2098 } 2099 2100 if (align < ARCH_SLAB_MINALIGN) 2101 align = ARCH_SLAB_MINALIGN; 2102 2103 return ALIGN(align, sizeof(void *)); 2104} 2105 2106static void 2107init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 2108{ 2109 n->nr_partial = 0; 2110 spin_lock_init(&n->list_lock); 2111 INIT_LIST_HEAD(&n->partial); 2112#ifdef CONFIG_SLUB_DEBUG 2113 atomic_long_set(&n->nr_slabs, 0); 2114 atomic_long_set(&n->total_objects, 0); 2115 INIT_LIST_HEAD(&n->full); 2116#endif 2117} 2118 2119static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 2120{ 2121 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 2122 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); 2123 2124 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu); 2125 2126 return s->cpu_slab != NULL; 2127} 2128 2129static struct kmem_cache *kmem_cache_node; 2130 2131/* 2132 * No kmalloc_node yet so do it by hand. We know that this is the first 2133 * slab on the node for this slabcache. There are no concurrent accesses 2134 * possible. 2135 * 2136 * Note that this function only works on the kmalloc_node_cache 2137 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2138 * memory on a fresh node that has no slab structures yet. 2139 */ 2140static void early_kmem_cache_node_alloc(int node) 2141{ 2142 struct page *page; 2143 struct kmem_cache_node *n; 2144 unsigned long flags; 2145 2146 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 2147 2148 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 2149 2150 BUG_ON(!page); 2151 if (page_to_nid(page) != node) { 2152 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2153 "node %d\n", node); 2154 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2155 "in order to be able to continue\n"); 2156 } 2157 2158 n = page->freelist; 2159 BUG_ON(!n); 2160 page->freelist = get_freepointer(kmem_cache_node, n); 2161 page->inuse++; 2162 kmem_cache_node->node[node] = n; 2163#ifdef CONFIG_SLUB_DEBUG 2164 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 2165 init_tracking(kmem_cache_node, n); 2166#endif 2167 init_kmem_cache_node(n, kmem_cache_node); 2168 inc_slabs_node(kmem_cache_node, node, page->objects); 2169 2170 /* 2171 * lockdep requires consistent irq usage for each lock 2172 * so even though there cannot be a race this early in 2173 * the boot sequence, we still disable irqs. 2174 */ 2175 local_irq_save(flags); 2176 add_partial(n, page, 0); 2177 local_irq_restore(flags); 2178} 2179 2180static void free_kmem_cache_nodes(struct kmem_cache *s) 2181{ 2182 int node; 2183 2184 for_each_node_state(node, N_NORMAL_MEMORY) { 2185 struct kmem_cache_node *n = s->node[node]; 2186 2187 if (n) 2188 kmem_cache_free(kmem_cache_node, n); 2189 2190 s->node[node] = NULL; 2191 } 2192} 2193 2194static int init_kmem_cache_nodes(struct kmem_cache *s) 2195{ 2196 int node; 2197 2198 for_each_node_state(node, N_NORMAL_MEMORY) { 2199 struct kmem_cache_node *n; 2200 2201 if (slab_state == DOWN) { 2202 early_kmem_cache_node_alloc(node); 2203 continue; 2204 } 2205 n = kmem_cache_alloc_node(kmem_cache_node, 2206 GFP_KERNEL, node); 2207 2208 if (!n) { 2209 free_kmem_cache_nodes(s); 2210 return 0; 2211 } 2212 2213 s->node[node] = n; 2214 init_kmem_cache_node(n, s); 2215 } 2216 return 1; 2217} 2218 2219static void set_min_partial(struct kmem_cache *s, unsigned long min) 2220{ 2221 if (min < MIN_PARTIAL) 2222 min = MIN_PARTIAL; 2223 else if (min > MAX_PARTIAL) 2224 min = MAX_PARTIAL; 2225 s->min_partial = min; 2226} 2227 2228/* 2229 * calculate_sizes() determines the order and the distribution of data within 2230 * a slab object. 2231 */ 2232static int calculate_sizes(struct kmem_cache *s, int forced_order) 2233{ 2234 unsigned long flags = s->flags; 2235 unsigned long size = s->objsize; 2236 unsigned long align = s->align; 2237 int order; 2238 2239 /* 2240 * Round up object size to the next word boundary. We can only 2241 * place the free pointer at word boundaries and this determines 2242 * the possible location of the free pointer. 2243 */ 2244 size = ALIGN(size, sizeof(void *)); 2245 2246#ifdef CONFIG_SLUB_DEBUG 2247 /* 2248 * Determine if we can poison the object itself. If the user of 2249 * the slab may touch the object after free or before allocation 2250 * then we should never poison the object itself. 2251 */ 2252 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2253 !s->ctor) 2254 s->flags |= __OBJECT_POISON; 2255 else 2256 s->flags &= ~__OBJECT_POISON; 2257 2258 2259 /* 2260 * If we are Redzoning then check if there is some space between the 2261 * end of the object and the free pointer. If not then add an 2262 * additional word to have some bytes to store Redzone information. 2263 */ 2264 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2265 size += sizeof(void *); 2266#endif 2267 2268 /* 2269 * With that we have determined the number of bytes in actual use 2270 * by the object. This is the potential offset to the free pointer. 2271 */ 2272 s->inuse = size; 2273 2274 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2275 s->ctor)) { 2276 /* 2277 * Relocate free pointer after the object if it is not 2278 * permitted to overwrite the first word of the object on 2279 * kmem_cache_free. 2280 * 2281 * This is the case if we do RCU, have a constructor or 2282 * destructor or are poisoning the objects. 2283 */ 2284 s->offset = size; 2285 size += sizeof(void *); 2286 } 2287 2288#ifdef CONFIG_SLUB_DEBUG 2289 if (flags & SLAB_STORE_USER) 2290 /* 2291 * Need to store information about allocs and frees after 2292 * the object. 2293 */ 2294 size += 2 * sizeof(struct track); 2295 2296 if (flags & SLAB_RED_ZONE) 2297 /* 2298 * Add some empty padding so that we can catch 2299 * overwrites from earlier objects rather than let 2300 * tracking information or the free pointer be 2301 * corrupted if a user writes before the start 2302 * of the object. 2303 */ 2304 size += sizeof(void *); 2305#endif 2306 2307 /* 2308 * Determine the alignment based on various parameters that the 2309 * user specified and the dynamic determination of cache line size 2310 * on bootup. 2311 */ 2312 align = calculate_alignment(flags, align, s->objsize); 2313 s->align = align; 2314 2315 /* 2316 * SLUB stores one object immediately after another beginning from 2317 * offset 0. In order to align the objects we have to simply size 2318 * each object to conform to the alignment. 2319 */ 2320 size = ALIGN(size, align); 2321 s->size = size; 2322 if (forced_order >= 0) 2323 order = forced_order; 2324 else 2325 order = calculate_order(size); 2326 2327 if (order < 0) 2328 return 0; 2329 2330 s->allocflags = 0; 2331 if (order) 2332 s->allocflags |= __GFP_COMP; 2333 2334 if (s->flags & SLAB_CACHE_DMA) 2335 s->allocflags |= SLUB_DMA; 2336 2337 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2338 s->allocflags |= __GFP_RECLAIMABLE; 2339 2340 /* 2341 * Determine the number of objects per slab 2342 */ 2343 s->oo = oo_make(order, size); 2344 s->min = oo_make(get_order(size), size); 2345 if (oo_objects(s->oo) > oo_objects(s->max)) 2346 s->max = s->oo; 2347 2348 return !!oo_objects(s->oo); 2349 2350} 2351 2352static int kmem_cache_open(struct kmem_cache *s, 2353 const char *name, size_t size, 2354 size_t align, unsigned long flags, 2355 void (*ctor)(void *)) 2356{ 2357 memset(s, 0, kmem_size); 2358 s->name = name; 2359 s->ctor = ctor; 2360 s->objsize = size; 2361 s->align = align; 2362 s->flags = kmem_cache_flags(size, flags, name, ctor); 2363 2364 if (!calculate_sizes(s, -1)) 2365 goto error; 2366 if (disable_higher_order_debug) { 2367 /* 2368 * Disable debugging flags that store metadata if the min slab 2369 * order increased. 2370 */ 2371 if (get_order(s->size) > get_order(s->objsize)) { 2372 s->flags &= ~DEBUG_METADATA_FLAGS; 2373 s->offset = 0; 2374 if (!calculate_sizes(s, -1)) 2375 goto error; 2376 } 2377 } 2378 2379 /* 2380 * The larger the object size is, the more pages we want on the partial 2381 * list to avoid pounding the page allocator excessively. 2382 */ 2383 set_min_partial(s, ilog2(s->size)); 2384 s->refcount = 1; 2385#ifdef CONFIG_NUMA 2386 s->remote_node_defrag_ratio = 1000; 2387#endif 2388 if (!init_kmem_cache_nodes(s)) 2389 goto error; 2390 2391 if (alloc_kmem_cache_cpus(s)) 2392 return 1; 2393 2394 free_kmem_cache_nodes(s); 2395error: 2396 if (flags & SLAB_PANIC) 2397 panic("Cannot create slab %s size=%lu realsize=%u " 2398 "order=%u offset=%u flags=%lx\n", 2399 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2400 s->offset, flags); 2401 return 0; 2402} 2403 2404/* 2405 * Check if a given pointer is valid 2406 */ 2407int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2408{ 2409 struct page *page; 2410 2411 if (!kern_ptr_validate(object, s->size)) 2412 return 0; 2413 2414 page = get_object_page(object); 2415 2416 if (!page || s != page->slab) 2417 /* No slab or wrong slab */ 2418 return 0; 2419 2420 if (!check_valid_pointer(s, page, object)) 2421 return 0; 2422 2423 /* 2424 * We could also check if the object is on the slabs freelist. 2425 * But this would be too expensive and it seems that the main 2426 * purpose of kmem_ptr_valid() is to check if the object belongs 2427 * to a certain slab. 2428 */ 2429 return 1; 2430} 2431EXPORT_SYMBOL(kmem_ptr_validate); 2432 2433/* 2434 * Determine the size of a slab object 2435 */ 2436unsigned int kmem_cache_size(struct kmem_cache *s) 2437{ 2438 return s->objsize; 2439} 2440EXPORT_SYMBOL(kmem_cache_size); 2441 2442const char *kmem_cache_name(struct kmem_cache *s) 2443{ 2444 return s->name; 2445} 2446EXPORT_SYMBOL(kmem_cache_name); 2447 2448static void list_slab_objects(struct kmem_cache *s, struct page *page, 2449 const char *text) 2450{ 2451#ifdef CONFIG_SLUB_DEBUG 2452 void *addr = page_address(page); 2453 void *p; 2454 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * 2455 sizeof(long), GFP_ATOMIC); 2456 if (!map) 2457 return; 2458 slab_err(s, page, "%s", text); 2459 slab_lock(page); 2460 for_each_free_object(p, s, page->freelist) 2461 set_bit(slab_index(p, s, addr), map); 2462 2463 for_each_object(p, s, addr, page->objects) { 2464 2465 if (!test_bit(slab_index(p, s, addr), map)) { 2466 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2467 p, p - addr); 2468 print_tracking(s, p); 2469 } 2470 } 2471 slab_unlock(page); 2472 kfree(map); 2473#endif 2474} 2475 2476/* 2477 * Attempt to free all partial slabs on a node. 2478 */ 2479static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2480{ 2481 unsigned long flags; 2482 struct page *page, *h; 2483 2484 spin_lock_irqsave(&n->list_lock, flags); 2485 list_for_each_entry_safe(page, h, &n->partial, lru) { 2486 if (!page->inuse) { 2487 __remove_partial(n, page); 2488 discard_slab(s, page); 2489 } else { 2490 list_slab_objects(s, page, 2491 "Objects remaining on kmem_cache_close()"); 2492 } 2493 } 2494 spin_unlock_irqrestore(&n->list_lock, flags); 2495} 2496 2497/* 2498 * Release all resources used by a slab cache. 2499 */ 2500static inline int kmem_cache_close(struct kmem_cache *s) 2501{ 2502 int node; 2503 2504 flush_all(s); 2505 free_percpu(s->cpu_slab); 2506 /* Attempt to free all objects */ 2507 for_each_node_state(node, N_NORMAL_MEMORY) { 2508 struct kmem_cache_node *n = get_node(s, node); 2509 2510 free_partial(s, n); 2511 if (n->nr_partial || slabs_node(s, node)) 2512 return 1; 2513 } 2514 free_kmem_cache_nodes(s); 2515 return 0; 2516} 2517 2518/* 2519 * Close a cache and release the kmem_cache structure 2520 * (must be used for caches created using kmem_cache_create) 2521 */ 2522void kmem_cache_destroy(struct kmem_cache *s) 2523{ 2524 down_write(&slub_lock); 2525 s->refcount--; 2526 if (!s->refcount) { 2527 list_del(&s->list); 2528 if (kmem_cache_close(s)) { 2529 printk(KERN_ERR "SLUB %s: %s called for cache that " 2530 "still has objects.\n", s->name, __func__); 2531 dump_stack(); 2532 } 2533 if (s->flags & SLAB_DESTROY_BY_RCU) 2534 rcu_barrier(); 2535 sysfs_slab_remove(s); 2536 } 2537 up_write(&slub_lock); 2538} 2539EXPORT_SYMBOL(kmem_cache_destroy); 2540 2541/******************************************************************** 2542 * Kmalloc subsystem 2543 *******************************************************************/ 2544 2545struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; 2546EXPORT_SYMBOL(kmalloc_caches); 2547 2548static struct kmem_cache *kmem_cache; 2549 2550#ifdef CONFIG_ZONE_DMA 2551static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; 2552#endif 2553 2554static int __init setup_slub_min_order(char *str) 2555{ 2556 get_option(&str, &slub_min_order); 2557 2558 return 1; 2559} 2560 2561__setup("slub_min_order=", setup_slub_min_order); 2562 2563static int __init setup_slub_max_order(char *str) 2564{ 2565 get_option(&str, &slub_max_order); 2566 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 2567 2568 return 1; 2569} 2570 2571__setup("slub_max_order=", setup_slub_max_order); 2572 2573static int __init setup_slub_min_objects(char *str) 2574{ 2575 get_option(&str, &slub_min_objects); 2576 2577 return 1; 2578} 2579 2580__setup("slub_min_objects=", setup_slub_min_objects); 2581 2582static int __init setup_slub_nomerge(char *str) 2583{ 2584 slub_nomerge = 1; 2585 return 1; 2586} 2587 2588__setup("slub_nomerge", setup_slub_nomerge); 2589 2590static struct kmem_cache *__init create_kmalloc_cache(const char *name, 2591 int size, unsigned int flags) 2592{ 2593 struct kmem_cache *s; 2594 2595 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 2596 2597 /* 2598 * This function is called with IRQs disabled during early-boot on 2599 * single CPU so there's no need to take slub_lock here. 2600 */ 2601 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, 2602 flags, NULL)) 2603 goto panic; 2604 2605 list_add(&s->list, &slab_caches); 2606 return s; 2607 2608panic: 2609 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2610 return NULL; 2611} 2612 2613/* 2614 * Conversion table for small slabs sizes / 8 to the index in the 2615 * kmalloc array. This is necessary for slabs < 192 since we have non power 2616 * of two cache sizes there. The size of larger slabs can be determined using 2617 * fls. 2618 */ 2619static s8 size_index[24] = { 2620 3, /* 8 */ 2621 4, /* 16 */ 2622 5, /* 24 */ 2623 5, /* 32 */ 2624 6, /* 40 */ 2625 6, /* 48 */ 2626 6, /* 56 */ 2627 6, /* 64 */ 2628 1, /* 72 */ 2629 1, /* 80 */ 2630 1, /* 88 */ 2631 1, /* 96 */ 2632 7, /* 104 */ 2633 7, /* 112 */ 2634 7, /* 120 */ 2635 7, /* 128 */ 2636 2, /* 136 */ 2637 2, /* 144 */ 2638 2, /* 152 */ 2639 2, /* 160 */ 2640 2, /* 168 */ 2641 2, /* 176 */ 2642 2, /* 184 */ 2643 2 /* 192 */ 2644}; 2645 2646static inline int size_index_elem(size_t bytes) 2647{ 2648 return (bytes - 1) / 8; 2649} 2650 2651static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2652{ 2653 int index; 2654 2655 if (size <= 192) { 2656 if (!size) 2657 return ZERO_SIZE_PTR; 2658 2659 index = size_index[size_index_elem(size)]; 2660 } else 2661 index = fls(size - 1); 2662 2663#ifdef CONFIG_ZONE_DMA 2664 if (unlikely((flags & SLUB_DMA))) 2665 return kmalloc_dma_caches[index]; 2666 2667#endif 2668 return kmalloc_caches[index]; 2669} 2670 2671void *__kmalloc(size_t size, gfp_t flags) 2672{ 2673 struct kmem_cache *s; 2674 void *ret; 2675 2676 if (unlikely(size > SLUB_MAX_SIZE)) 2677 return kmalloc_large(size, flags); 2678 2679 s = get_slab(size, flags); 2680 2681 if (unlikely(ZERO_OR_NULL_PTR(s))) 2682 return s; 2683 2684 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); 2685 2686 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 2687 2688 return ret; 2689} 2690EXPORT_SYMBOL(__kmalloc); 2691 2692#ifdef CONFIG_NUMA 2693static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2694{ 2695 struct page *page; 2696 void *ptr = NULL; 2697 2698 flags |= __GFP_COMP | __GFP_NOTRACK; 2699 page = alloc_pages_node(node, flags, get_order(size)); 2700 if (page) 2701 ptr = page_address(page); 2702 2703 kmemleak_alloc(ptr, size, 1, flags); 2704 return ptr; 2705} 2706 2707void *__kmalloc_node(size_t size, gfp_t flags, int node) 2708{ 2709 struct kmem_cache *s; 2710 void *ret; 2711 2712 if (unlikely(size > SLUB_MAX_SIZE)) { 2713 ret = kmalloc_large_node(size, flags, node); 2714 2715 trace_kmalloc_node(_RET_IP_, ret, 2716 size, PAGE_SIZE << get_order(size), 2717 flags, node); 2718 2719 return ret; 2720 } 2721 2722 s = get_slab(size, flags); 2723 2724 if (unlikely(ZERO_OR_NULL_PTR(s))) 2725 return s; 2726 2727 ret = slab_alloc(s, flags, node, _RET_IP_); 2728 2729 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 2730 2731 return ret; 2732} 2733EXPORT_SYMBOL(__kmalloc_node); 2734#endif 2735 2736size_t ksize(const void *object) 2737{ 2738 struct page *page; 2739 struct kmem_cache *s; 2740 2741 if (unlikely(object == ZERO_SIZE_PTR)) 2742 return 0; 2743 2744 page = virt_to_head_page(object); 2745 2746 if (unlikely(!PageSlab(page))) { 2747 WARN_ON(!PageCompound(page)); 2748 return PAGE_SIZE << compound_order(page); 2749 } 2750 s = page->slab; 2751 2752#ifdef CONFIG_SLUB_DEBUG 2753 /* 2754 * Debugging requires use of the padding between object 2755 * and whatever may come after it. 2756 */ 2757 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2758 return s->objsize; 2759 2760#endif 2761 /* 2762 * If we have the need to store the freelist pointer 2763 * back there or track user information then we can 2764 * only use the space before that information. 2765 */ 2766 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2767 return s->inuse; 2768 /* 2769 * Else we can use all the padding etc for the allocation 2770 */ 2771 return s->size; 2772} 2773EXPORT_SYMBOL(ksize); 2774 2775void kfree(const void *x) 2776{ 2777 struct page *page; 2778 void *object = (void *)x; 2779 2780 trace_kfree(_RET_IP_, x); 2781 2782 if (unlikely(ZERO_OR_NULL_PTR(x))) 2783 return; 2784 2785 page = virt_to_head_page(x); 2786 if (unlikely(!PageSlab(page))) { 2787 BUG_ON(!PageCompound(page)); 2788 kmemleak_free(x); 2789 put_page(page); 2790 return; 2791 } 2792 slab_free(page->slab, page, object, _RET_IP_); 2793} 2794EXPORT_SYMBOL(kfree); 2795 2796/* 2797 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2798 * the remaining slabs by the number of items in use. The slabs with the 2799 * most items in use come first. New allocations will then fill those up 2800 * and thus they can be removed from the partial lists. 2801 * 2802 * The slabs with the least items are placed last. This results in them 2803 * being allocated from last increasing the chance that the last objects 2804 * are freed in them. 2805 */ 2806int kmem_cache_shrink(struct kmem_cache *s) 2807{ 2808 int node; 2809 int i; 2810 struct kmem_cache_node *n; 2811 struct page *page; 2812 struct page *t; 2813 int objects = oo_objects(s->max); 2814 struct list_head *slabs_by_inuse = 2815 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 2816 unsigned long flags; 2817 2818 if (!slabs_by_inuse) 2819 return -ENOMEM; 2820 2821 flush_all(s); 2822 for_each_node_state(node, N_NORMAL_MEMORY) { 2823 n = get_node(s, node); 2824 2825 if (!n->nr_partial) 2826 continue; 2827 2828 for (i = 0; i < objects; i++) 2829 INIT_LIST_HEAD(slabs_by_inuse + i); 2830 2831 spin_lock_irqsave(&n->list_lock, flags); 2832 2833 /* 2834 * Build lists indexed by the items in use in each slab. 2835 * 2836 * Note that concurrent frees may occur while we hold the 2837 * list_lock. page->inuse here is the upper limit. 2838 */ 2839 list_for_each_entry_safe(page, t, &n->partial, lru) { 2840 if (!page->inuse && slab_trylock(page)) { 2841 /* 2842 * Must hold slab lock here because slab_free 2843 * may have freed the last object and be 2844 * waiting to release the slab. 2845 */ 2846 __remove_partial(n, page); 2847 slab_unlock(page); 2848 discard_slab(s, page); 2849 } else { 2850 list_move(&page->lru, 2851 slabs_by_inuse + page->inuse); 2852 } 2853 } 2854 2855 /* 2856 * Rebuild the partial list with the slabs filled up most 2857 * first and the least used slabs at the end. 2858 */ 2859 for (i = objects - 1; i >= 0; i--) 2860 list_splice(slabs_by_inuse + i, n->partial.prev); 2861 2862 spin_unlock_irqrestore(&n->list_lock, flags); 2863 } 2864 2865 kfree(slabs_by_inuse); 2866 return 0; 2867} 2868EXPORT_SYMBOL(kmem_cache_shrink); 2869 2870#if defined(CONFIG_MEMORY_HOTPLUG) 2871static int slab_mem_going_offline_callback(void *arg) 2872{ 2873 struct kmem_cache *s; 2874 2875 down_read(&slub_lock); 2876 list_for_each_entry(s, &slab_caches, list) 2877 kmem_cache_shrink(s); 2878 up_read(&slub_lock); 2879 2880 return 0; 2881} 2882 2883static void slab_mem_offline_callback(void *arg) 2884{ 2885 struct kmem_cache_node *n; 2886 struct kmem_cache *s; 2887 struct memory_notify *marg = arg; 2888 int offline_node; 2889 2890 offline_node = marg->status_change_nid; 2891 2892 /* 2893 * If the node still has available memory. we need kmem_cache_node 2894 * for it yet. 2895 */ 2896 if (offline_node < 0) 2897 return; 2898 2899 down_read(&slub_lock); 2900 list_for_each_entry(s, &slab_caches, list) { 2901 n = get_node(s, offline_node); 2902 if (n) { 2903 /* 2904 * if n->nr_slabs > 0, slabs still exist on the node 2905 * that is going down. We were unable to free them, 2906 * and offline_pages() function shouldn't call this 2907 * callback. So, we must fail. 2908 */ 2909 BUG_ON(slabs_node(s, offline_node)); 2910 2911 s->node[offline_node] = NULL; 2912 kmem_cache_free(kmem_cache_node, n); 2913 } 2914 } 2915 up_read(&slub_lock); 2916} 2917 2918static int slab_mem_going_online_callback(void *arg) 2919{ 2920 struct kmem_cache_node *n; 2921 struct kmem_cache *s; 2922 struct memory_notify *marg = arg; 2923 int nid = marg->status_change_nid; 2924 int ret = 0; 2925 2926 /* 2927 * If the node's memory is already available, then kmem_cache_node is 2928 * already created. Nothing to do. 2929 */ 2930 if (nid < 0) 2931 return 0; 2932 2933 /* 2934 * We are bringing a node online. No memory is available yet. We must 2935 * allocate a kmem_cache_node structure in order to bring the node 2936 * online. 2937 */ 2938 down_read(&slub_lock); 2939 list_for_each_entry(s, &slab_caches, list) { 2940 /* 2941 * XXX: kmem_cache_alloc_node will fallback to other nodes 2942 * since memory is not yet available from the node that 2943 * is brought up. 2944 */ 2945 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 2946 if (!n) { 2947 ret = -ENOMEM; 2948 goto out; 2949 } 2950 init_kmem_cache_node(n, s); 2951 s->node[nid] = n; 2952 } 2953out: 2954 up_read(&slub_lock); 2955 return ret; 2956} 2957 2958static int slab_memory_callback(struct notifier_block *self, 2959 unsigned long action, void *arg) 2960{ 2961 int ret = 0; 2962 2963 switch (action) { 2964 case MEM_GOING_ONLINE: 2965 ret = slab_mem_going_online_callback(arg); 2966 break; 2967 case MEM_GOING_OFFLINE: 2968 ret = slab_mem_going_offline_callback(arg); 2969 break; 2970 case MEM_OFFLINE: 2971 case MEM_CANCEL_ONLINE: 2972 slab_mem_offline_callback(arg); 2973 break; 2974 case MEM_ONLINE: 2975 case MEM_CANCEL_OFFLINE: 2976 break; 2977 } 2978 if (ret) 2979 ret = notifier_from_errno(ret); 2980 else 2981 ret = NOTIFY_OK; 2982 return ret; 2983} 2984 2985#endif /* CONFIG_MEMORY_HOTPLUG */ 2986 2987/******************************************************************** 2988 * Basic setup of slabs 2989 *******************************************************************/ 2990 2991/* 2992 * Used for early kmem_cache structures that were allocated using 2993 * the page allocator 2994 */ 2995 2996static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) 2997{ 2998 int node; 2999 3000 list_add(&s->list, &slab_caches); 3001 s->refcount = -1; 3002 3003 for_each_node_state(node, N_NORMAL_MEMORY) { 3004 struct kmem_cache_node *n = get_node(s, node); 3005 struct page *p; 3006 3007 if (n) { 3008 list_for_each_entry(p, &n->partial, lru) 3009 p->slab = s; 3010 3011#ifdef CONFIG_SLAB_DEBUG 3012 list_for_each_entry(p, &n->full, lru) 3013 p->slab = s; 3014#endif 3015 } 3016 } 3017} 3018 3019void __init kmem_cache_init(void) 3020{ 3021 int i; 3022 int caches = 0; 3023 struct kmem_cache *temp_kmem_cache; 3024 int order; 3025 struct kmem_cache *temp_kmem_cache_node; 3026 unsigned long kmalloc_size; 3027 3028 kmem_size = offsetof(struct kmem_cache, node) + 3029 nr_node_ids * sizeof(struct kmem_cache_node *); 3030 3031 /* Allocate two kmem_caches from the page allocator */ 3032 kmalloc_size = ALIGN(kmem_size, cache_line_size()); 3033 order = get_order(2 * kmalloc_size); 3034 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); 3035 3036 /* 3037 * Must first have the slab cache available for the allocations of the 3038 * struct kmem_cache_node's. There is special bootstrap code in 3039 * kmem_cache_open for slab_state == DOWN. 3040 */ 3041 kmem_cache_node = (void *)kmem_cache + kmalloc_size; 3042 3043 kmem_cache_open(kmem_cache_node, "kmem_cache_node", 3044 sizeof(struct kmem_cache_node), 3045 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3046 3047 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 3048 3049 /* Able to allocate the per node structures */ 3050 slab_state = PARTIAL; 3051 3052 temp_kmem_cache = kmem_cache; 3053 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, 3054 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3055 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3056 memcpy(kmem_cache, temp_kmem_cache, kmem_size); 3057 3058 /* 3059 * Allocate kmem_cache_node properly from the kmem_cache slab. 3060 * kmem_cache_node is separately allocated so no need to 3061 * update any list pointers. 3062 */ 3063 temp_kmem_cache_node = kmem_cache_node; 3064 3065 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3066 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); 3067 3068 kmem_cache_bootstrap_fixup(kmem_cache_node); 3069 3070 caches++; 3071 kmem_cache_bootstrap_fixup(kmem_cache); 3072 caches++; 3073 /* Free temporary boot structure */ 3074 free_pages((unsigned long)temp_kmem_cache, order); 3075 3076 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 3077 3078 /* 3079 * Patch up the size_index table if we have strange large alignment 3080 * requirements for the kmalloc array. This is only the case for 3081 * MIPS it seems. The standard arches will not generate any code here. 3082 * 3083 * Largest permitted alignment is 256 bytes due to the way we 3084 * handle the index determination for the smaller caches. 3085 * 3086 * Make sure that nothing crazy happens if someone starts tinkering 3087 * around with ARCH_KMALLOC_MINALIGN 3088 */ 3089 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3090 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3091 3092 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { 3093 int elem = size_index_elem(i); 3094 if (elem >= ARRAY_SIZE(size_index)) 3095 break; 3096 size_index[elem] = KMALLOC_SHIFT_LOW; 3097 } 3098 3099 if (KMALLOC_MIN_SIZE == 64) { 3100 /* 3101 * The 96 byte size cache is not used if the alignment 3102 * is 64 byte. 3103 */ 3104 for (i = 64 + 8; i <= 96; i += 8) 3105 size_index[size_index_elem(i)] = 7; 3106 } else if (KMALLOC_MIN_SIZE == 128) { 3107 /* 3108 * The 192 byte sized cache is not used if the alignment 3109 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3110 * instead. 3111 */ 3112 for (i = 128 + 8; i <= 192; i += 8) 3113 size_index[size_index_elem(i)] = 8; 3114 } 3115 3116 /* Caches that are not of the two-to-the-power-of size */ 3117 if (KMALLOC_MIN_SIZE <= 32) { 3118 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); 3119 caches++; 3120 } 3121 3122 if (KMALLOC_MIN_SIZE <= 64) { 3123 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); 3124 caches++; 3125 } 3126 3127 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3128 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); 3129 caches++; 3130 } 3131 3132 slab_state = UP; 3133 3134 /* Provide the correct kmalloc names now that the caches are up */ 3135 if (KMALLOC_MIN_SIZE <= 32) { 3136 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); 3137 BUG_ON(!kmalloc_caches[1]->name); 3138 } 3139 3140 if (KMALLOC_MIN_SIZE <= 64) { 3141 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); 3142 BUG_ON(!kmalloc_caches[2]->name); 3143 } 3144 3145 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3146 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); 3147 3148 BUG_ON(!s); 3149 kmalloc_caches[i]->name = s; 3150 } 3151 3152#ifdef CONFIG_SMP 3153 register_cpu_notifier(&slab_notifier); 3154#endif 3155 3156#ifdef CONFIG_ZONE_DMA 3157 for (i = 0; i < SLUB_PAGE_SHIFT; i++) { 3158 struct kmem_cache *s = kmalloc_caches[i]; 3159 3160 if (s && s->size) { 3161 char *name = kasprintf(GFP_NOWAIT, 3162 "dma-kmalloc-%d", s->objsize); 3163 3164 BUG_ON(!name); 3165 kmalloc_dma_caches[i] = create_kmalloc_cache(name, 3166 s->objsize, SLAB_CACHE_DMA); 3167 } 3168 } 3169#endif 3170 printk(KERN_INFO 3171 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3172 " CPUs=%d, Nodes=%d\n", 3173 caches, cache_line_size(), 3174 slub_min_order, slub_max_order, slub_min_objects, 3175 nr_cpu_ids, nr_node_ids); 3176} 3177 3178void __init kmem_cache_init_late(void) 3179{ 3180} 3181 3182/* 3183 * Find a mergeable slab cache 3184 */ 3185static int slab_unmergeable(struct kmem_cache *s) 3186{ 3187 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3188 return 1; 3189 3190 if (s->ctor) 3191 return 1; 3192 3193 /* 3194 * We may have set a slab to be unmergeable during bootstrap. 3195 */ 3196 if (s->refcount < 0) 3197 return 1; 3198 3199 return 0; 3200} 3201 3202static struct kmem_cache *find_mergeable(size_t size, 3203 size_t align, unsigned long flags, const char *name, 3204 void (*ctor)(void *)) 3205{ 3206 struct kmem_cache *s; 3207 3208 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3209 return NULL; 3210 3211 if (ctor) 3212 return NULL; 3213 3214 size = ALIGN(size, sizeof(void *)); 3215 align = calculate_alignment(flags, align, size); 3216 size = ALIGN(size, align); 3217 flags = kmem_cache_flags(size, flags, name, NULL); 3218 3219 list_for_each_entry(s, &slab_caches, list) { 3220 if (slab_unmergeable(s)) 3221 continue; 3222 3223 if (size > s->size) 3224 continue; 3225 3226 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3227 continue; 3228 /* 3229 * Check if alignment is compatible. 3230 * Courtesy of Adrian Drzewiecki 3231 */ 3232 if ((s->size & ~(align - 1)) != s->size) 3233 continue; 3234 3235 if (s->size - size >= sizeof(void *)) 3236 continue; 3237 3238 return s; 3239 } 3240 return NULL; 3241} 3242 3243struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3244 size_t align, unsigned long flags, void (*ctor)(void *)) 3245{ 3246 struct kmem_cache *s; 3247 char *n; 3248 3249 if (WARN_ON(!name)) 3250 return NULL; 3251 3252 down_write(&slub_lock); 3253 s = find_mergeable(size, align, flags, name, ctor); 3254 if (s) { 3255 s->refcount++; 3256 /* 3257 * Adjust the object sizes so that we clear 3258 * the complete object on kzalloc. 3259 */ 3260 s->objsize = max(s->objsize, (int)size); 3261 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3262 3263 if (sysfs_slab_alias(s, name)) { 3264 s->refcount--; 3265 goto err; 3266 } 3267 up_write(&slub_lock); 3268 return s; 3269 } 3270 3271 n = kstrdup(name, GFP_KERNEL); 3272 if (!n) 3273 goto err; 3274 3275 s = kmalloc(kmem_size, GFP_KERNEL); 3276 if (s) { 3277 if (kmem_cache_open(s, n, 3278 size, align, flags, ctor)) { 3279 list_add(&s->list, &slab_caches); 3280 if (sysfs_slab_add(s)) { 3281 list_del(&s->list); 3282 kfree(n); 3283 kfree(s); 3284 goto err; 3285 } 3286 up_write(&slub_lock); 3287 return s; 3288 } 3289 kfree(n); 3290 kfree(s); 3291 } 3292err: 3293 up_write(&slub_lock); 3294 3295 if (flags & SLAB_PANIC) 3296 panic("Cannot create slabcache %s\n", name); 3297 else 3298 s = NULL; 3299 return s; 3300} 3301EXPORT_SYMBOL(kmem_cache_create); 3302 3303#ifdef CONFIG_SMP 3304/* 3305 * Use the cpu notifier to insure that the cpu slabs are flushed when 3306 * necessary. 3307 */ 3308static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3309 unsigned long action, void *hcpu) 3310{ 3311 long cpu = (long)hcpu; 3312 struct kmem_cache *s; 3313 unsigned long flags; 3314 3315 switch (action) { 3316 case CPU_UP_CANCELED: 3317 case CPU_UP_CANCELED_FROZEN: 3318 case CPU_DEAD: 3319 case CPU_DEAD_FROZEN: 3320 down_read(&slub_lock); 3321 list_for_each_entry(s, &slab_caches, list) { 3322 local_irq_save(flags); 3323 __flush_cpu_slab(s, cpu); 3324 local_irq_restore(flags); 3325 } 3326 up_read(&slub_lock); 3327 break; 3328 default: 3329 break; 3330 } 3331 return NOTIFY_OK; 3332} 3333 3334static struct notifier_block __cpuinitdata slab_notifier = { 3335 .notifier_call = slab_cpuup_callback 3336}; 3337 3338#endif 3339 3340void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3341{ 3342 struct kmem_cache *s; 3343 void *ret; 3344 3345 if (unlikely(size > SLUB_MAX_SIZE)) 3346 return kmalloc_large(size, gfpflags); 3347 3348 s = get_slab(size, gfpflags); 3349 3350 if (unlikely(ZERO_OR_NULL_PTR(s))) 3351 return s; 3352 3353 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); 3354 3355 /* Honor the call site pointer we recieved. */ 3356 trace_kmalloc(caller, ret, size, s->size, gfpflags); 3357 3358 return ret; 3359} 3360 3361#ifdef CONFIG_NUMA 3362void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3363 int node, unsigned long caller) 3364{ 3365 struct kmem_cache *s; 3366 void *ret; 3367 3368 if (unlikely(size > SLUB_MAX_SIZE)) { 3369 ret = kmalloc_large_node(size, gfpflags, node); 3370 3371 trace_kmalloc_node(caller, ret, 3372 size, PAGE_SIZE << get_order(size), 3373 gfpflags, node); 3374 3375 return ret; 3376 } 3377 3378 s = get_slab(size, gfpflags); 3379 3380 if (unlikely(ZERO_OR_NULL_PTR(s))) 3381 return s; 3382 3383 ret = slab_alloc(s, gfpflags, node, caller); 3384 3385 /* Honor the call site pointer we recieved. */ 3386 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 3387 3388 return ret; 3389} 3390#endif 3391 3392#ifdef CONFIG_SYSFS 3393static int count_inuse(struct page *page) 3394{ 3395 return page->inuse; 3396} 3397 3398static int count_total(struct page *page) 3399{ 3400 return page->objects; 3401} 3402#endif 3403 3404#ifdef CONFIG_SLUB_DEBUG 3405static int validate_slab(struct kmem_cache *s, struct page *page, 3406 unsigned long *map) 3407{ 3408 void *p; 3409 void *addr = page_address(page); 3410 3411 if (!check_slab(s, page) || 3412 !on_freelist(s, page, NULL)) 3413 return 0; 3414 3415 /* Now we know that a valid freelist exists */ 3416 bitmap_zero(map, page->objects); 3417 3418 for_each_free_object(p, s, page->freelist) { 3419 set_bit(slab_index(p, s, addr), map); 3420 if (!check_object(s, page, p, 0)) 3421 return 0; 3422 } 3423 3424 for_each_object(p, s, addr, page->objects) 3425 if (!test_bit(slab_index(p, s, addr), map)) 3426 if (!check_object(s, page, p, 1)) 3427 return 0; 3428 return 1; 3429} 3430 3431static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3432 unsigned long *map) 3433{ 3434 if (slab_trylock(page)) { 3435 validate_slab(s, page, map); 3436 slab_unlock(page); 3437 } else 3438 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3439 s->name, page); 3440} 3441 3442static int validate_slab_node(struct kmem_cache *s, 3443 struct kmem_cache_node *n, unsigned long *map) 3444{ 3445 unsigned long count = 0; 3446 struct page *page; 3447 unsigned long flags; 3448 3449 spin_lock_irqsave(&n->list_lock, flags); 3450 3451 list_for_each_entry(page, &n->partial, lru) { 3452 validate_slab_slab(s, page, map); 3453 count++; 3454 } 3455 if (count != n->nr_partial) 3456 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3457 "counter=%ld\n", s->name, count, n->nr_partial); 3458 3459 if (!(s->flags & SLAB_STORE_USER)) 3460 goto out; 3461 3462 list_for_each_entry(page, &n->full, lru) { 3463 validate_slab_slab(s, page, map); 3464 count++; 3465 } 3466 if (count != atomic_long_read(&n->nr_slabs)) 3467 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3468 "counter=%ld\n", s->name, count, 3469 atomic_long_read(&n->nr_slabs)); 3470 3471out: 3472 spin_unlock_irqrestore(&n->list_lock, flags); 3473 return count; 3474} 3475 3476static long validate_slab_cache(struct kmem_cache *s) 3477{ 3478 int node; 3479 unsigned long count = 0; 3480 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3481 sizeof(unsigned long), GFP_KERNEL); 3482 3483 if (!map) 3484 return -ENOMEM; 3485 3486 flush_all(s); 3487 for_each_node_state(node, N_NORMAL_MEMORY) { 3488 struct kmem_cache_node *n = get_node(s, node); 3489 3490 count += validate_slab_node(s, n, map); 3491 } 3492 kfree(map); 3493 return count; 3494} 3495/* 3496 * Generate lists of code addresses where slabcache objects are allocated 3497 * and freed. 3498 */ 3499 3500struct location { 3501 unsigned long count; 3502 unsigned long addr; 3503 long long sum_time; 3504 long min_time; 3505 long max_time; 3506 long min_pid; 3507 long max_pid; 3508 DECLARE_BITMAP(cpus, NR_CPUS); 3509 nodemask_t nodes; 3510}; 3511 3512struct loc_track { 3513 unsigned long max; 3514 unsigned long count; 3515 struct location *loc; 3516}; 3517 3518static void free_loc_track(struct loc_track *t) 3519{ 3520 if (t->max) 3521 free_pages((unsigned long)t->loc, 3522 get_order(sizeof(struct location) * t->max)); 3523} 3524 3525static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3526{ 3527 struct location *l; 3528 int order; 3529 3530 order = get_order(sizeof(struct location) * max); 3531 3532 l = (void *)__get_free_pages(flags, order); 3533 if (!l) 3534 return 0; 3535 3536 if (t->count) { 3537 memcpy(l, t->loc, sizeof(struct location) * t->count); 3538 free_loc_track(t); 3539 } 3540 t->max = max; 3541 t->loc = l; 3542 return 1; 3543} 3544 3545static int add_location(struct loc_track *t, struct kmem_cache *s, 3546 const struct track *track) 3547{ 3548 long start, end, pos; 3549 struct location *l; 3550 unsigned long caddr; 3551 unsigned long age = jiffies - track->when; 3552 3553 start = -1; 3554 end = t->count; 3555 3556 for ( ; ; ) { 3557 pos = start + (end - start + 1) / 2; 3558 3559 /* 3560 * There is nothing at "end". If we end up there 3561 * we need to add something to before end. 3562 */ 3563 if (pos == end) 3564 break; 3565 3566 caddr = t->loc[pos].addr; 3567 if (track->addr == caddr) { 3568 3569 l = &t->loc[pos]; 3570 l->count++; 3571 if (track->when) { 3572 l->sum_time += age; 3573 if (age < l->min_time) 3574 l->min_time = age; 3575 if (age > l->max_time) 3576 l->max_time = age; 3577 3578 if (track->pid < l->min_pid) 3579 l->min_pid = track->pid; 3580 if (track->pid > l->max_pid) 3581 l->max_pid = track->pid; 3582 3583 cpumask_set_cpu(track->cpu, 3584 to_cpumask(l->cpus)); 3585 } 3586 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3587 return 1; 3588 } 3589 3590 if (track->addr < caddr) 3591 end = pos; 3592 else 3593 start = pos; 3594 } 3595 3596 /* 3597 * Not found. Insert new tracking element. 3598 */ 3599 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3600 return 0; 3601 3602 l = t->loc + pos; 3603 if (pos < t->count) 3604 memmove(l + 1, l, 3605 (t->count - pos) * sizeof(struct location)); 3606 t->count++; 3607 l->count = 1; 3608 l->addr = track->addr; 3609 l->sum_time = age; 3610 l->min_time = age; 3611 l->max_time = age; 3612 l->min_pid = track->pid; 3613 l->max_pid = track->pid; 3614 cpumask_clear(to_cpumask(l->cpus)); 3615 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 3616 nodes_clear(l->nodes); 3617 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3618 return 1; 3619} 3620 3621static void process_slab(struct loc_track *t, struct kmem_cache *s, 3622 struct page *page, enum track_item alloc, 3623 unsigned long *map) 3624{ 3625 void *addr = page_address(page); 3626 void *p; 3627 3628 bitmap_zero(map, page->objects); 3629 for_each_free_object(p, s, page->freelist) 3630 set_bit(slab_index(p, s, addr), map); 3631 3632 for_each_object(p, s, addr, page->objects) 3633 if (!test_bit(slab_index(p, s, addr), map)) 3634 add_location(t, s, get_track(s, p, alloc)); 3635} 3636 3637static int list_locations(struct kmem_cache *s, char *buf, 3638 enum track_item alloc) 3639{ 3640 int len = 0; 3641 unsigned long i; 3642 struct loc_track t = { 0, 0, NULL }; 3643 int node; 3644 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3645 sizeof(unsigned long), GFP_KERNEL); 3646 3647 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3648 GFP_TEMPORARY)) { 3649 kfree(map); 3650 return sprintf(buf, "Out of memory\n"); 3651 } 3652 /* Push back cpu slabs */ 3653 flush_all(s); 3654 3655 for_each_node_state(node, N_NORMAL_MEMORY) { 3656 struct kmem_cache_node *n = get_node(s, node); 3657 unsigned long flags; 3658 struct page *page; 3659 3660 if (!atomic_long_read(&n->nr_slabs)) 3661 continue; 3662 3663 spin_lock_irqsave(&n->list_lock, flags); 3664 list_for_each_entry(page, &n->partial, lru) 3665 process_slab(&t, s, page, alloc, map); 3666 list_for_each_entry(page, &n->full, lru) 3667 process_slab(&t, s, page, alloc, map); 3668 spin_unlock_irqrestore(&n->list_lock, flags); 3669 } 3670 3671 for (i = 0; i < t.count; i++) { 3672 struct location *l = &t.loc[i]; 3673 3674 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 3675 break; 3676 len += sprintf(buf + len, "%7ld ", l->count); 3677 3678 if (l->addr) 3679 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3680 else 3681 len += sprintf(buf + len, "<not-available>"); 3682 3683 if (l->sum_time != l->min_time) { 3684 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3685 l->min_time, 3686 (long)div_u64(l->sum_time, l->count), 3687 l->max_time); 3688 } else 3689 len += sprintf(buf + len, " age=%ld", 3690 l->min_time); 3691 3692 if (l->min_pid != l->max_pid) 3693 len += sprintf(buf + len, " pid=%ld-%ld", 3694 l->min_pid, l->max_pid); 3695 else 3696 len += sprintf(buf + len, " pid=%ld", 3697 l->min_pid); 3698 3699 if (num_online_cpus() > 1 && 3700 !cpumask_empty(to_cpumask(l->cpus)) && 3701 len < PAGE_SIZE - 60) { 3702 len += sprintf(buf + len, " cpus="); 3703 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3704 to_cpumask(l->cpus)); 3705 } 3706 3707 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 3708 len < PAGE_SIZE - 60) { 3709 len += sprintf(buf + len, " nodes="); 3710 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3711 l->nodes); 3712 } 3713 3714 len += sprintf(buf + len, "\n"); 3715 } 3716 3717 free_loc_track(&t); 3718 kfree(map); 3719 if (!t.count) 3720 len += sprintf(buf, "No data\n"); 3721 return len; 3722} 3723#endif 3724 3725#ifdef SLUB_RESILIENCY_TEST 3726static void resiliency_test(void) 3727{ 3728 u8 *p; 3729 3730 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); 3731 3732 printk(KERN_ERR "SLUB resiliency testing\n"); 3733 printk(KERN_ERR "-----------------------\n"); 3734 printk(KERN_ERR "A. Corruption after allocation\n"); 3735 3736 p = kzalloc(16, GFP_KERNEL); 3737 p[16] = 0x12; 3738 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3739 " 0x12->0x%p\n\n", p + 16); 3740 3741 validate_slab_cache(kmalloc_caches[4]); 3742 3743 /* Hmmm... The next two are dangerous */ 3744 p = kzalloc(32, GFP_KERNEL); 3745 p[32 + sizeof(void *)] = 0x34; 3746 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3747 " 0x34 -> -0x%p\n", p); 3748 printk(KERN_ERR 3749 "If allocated object is overwritten then not detectable\n\n"); 3750 3751 validate_slab_cache(kmalloc_caches[5]); 3752 p = kzalloc(64, GFP_KERNEL); 3753 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3754 *p = 0x56; 3755 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3756 p); 3757 printk(KERN_ERR 3758 "If allocated object is overwritten then not detectable\n\n"); 3759 validate_slab_cache(kmalloc_caches[6]); 3760 3761 printk(KERN_ERR "\nB. Corruption after free\n"); 3762 p = kzalloc(128, GFP_KERNEL); 3763 kfree(p); 3764 *p = 0x78; 3765 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3766 validate_slab_cache(kmalloc_caches[7]); 3767 3768 p = kzalloc(256, GFP_KERNEL); 3769 kfree(p); 3770 p[50] = 0x9a; 3771 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3772 p); 3773 validate_slab_cache(kmalloc_caches[8]); 3774 3775 p = kzalloc(512, GFP_KERNEL); 3776 kfree(p); 3777 p[512] = 0xab; 3778 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3779 validate_slab_cache(kmalloc_caches[9]); 3780} 3781#else 3782#ifdef CONFIG_SYSFS 3783static void resiliency_test(void) {}; 3784#endif 3785#endif 3786 3787#ifdef CONFIG_SYSFS 3788enum slab_stat_type { 3789 SL_ALL, /* All slabs */ 3790 SL_PARTIAL, /* Only partially allocated slabs */ 3791 SL_CPU, /* Only slabs used for cpu caches */ 3792 SL_OBJECTS, /* Determine allocated objects not slabs */ 3793 SL_TOTAL /* Determine object capacity not slabs */ 3794}; 3795 3796#define SO_ALL (1 << SL_ALL) 3797#define SO_PARTIAL (1 << SL_PARTIAL) 3798#define SO_CPU (1 << SL_CPU) 3799#define SO_OBJECTS (1 << SL_OBJECTS) 3800#define SO_TOTAL (1 << SL_TOTAL) 3801 3802static ssize_t show_slab_objects(struct kmem_cache *s, 3803 char *buf, unsigned long flags) 3804{ 3805 unsigned long total = 0; 3806 int node; 3807 int x; 3808 unsigned long *nodes; 3809 unsigned long *per_cpu; 3810 3811 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3812 if (!nodes) 3813 return -ENOMEM; 3814 per_cpu = nodes + nr_node_ids; 3815 3816 if (flags & SO_CPU) { 3817 int cpu; 3818 3819 for_each_possible_cpu(cpu) { 3820 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 3821 3822 if (!c || c->node < 0) 3823 continue; 3824 3825 if (c->page) { 3826 if (flags & SO_TOTAL) 3827 x = c->page->objects; 3828 else if (flags & SO_OBJECTS) 3829 x = c->page->inuse; 3830 else 3831 x = 1; 3832 3833 total += x; 3834 nodes[c->node] += x; 3835 } 3836 per_cpu[c->node]++; 3837 } 3838 } 3839 3840 down_read(&slub_lock); 3841#ifdef CONFIG_SLUB_DEBUG 3842 if (flags & SO_ALL) { 3843 for_each_node_state(node, N_NORMAL_MEMORY) { 3844 struct kmem_cache_node *n = get_node(s, node); 3845 3846 if (flags & SO_TOTAL) 3847 x = atomic_long_read(&n->total_objects); 3848 else if (flags & SO_OBJECTS) 3849 x = atomic_long_read(&n->total_objects) - 3850 count_partial(n, count_free); 3851 3852 else 3853 x = atomic_long_read(&n->nr_slabs); 3854 total += x; 3855 nodes[node] += x; 3856 } 3857 3858 } else 3859#endif 3860 if (flags & SO_PARTIAL) { 3861 for_each_node_state(node, N_NORMAL_MEMORY) { 3862 struct kmem_cache_node *n = get_node(s, node); 3863 3864 if (flags & SO_TOTAL) 3865 x = count_partial(n, count_total); 3866 else if (flags & SO_OBJECTS) 3867 x = count_partial(n, count_inuse); 3868 else 3869 x = n->nr_partial; 3870 total += x; 3871 nodes[node] += x; 3872 } 3873 } 3874 x = sprintf(buf, "%lu", total); 3875#ifdef CONFIG_NUMA 3876 for_each_node_state(node, N_NORMAL_MEMORY) 3877 if (nodes[node]) 3878 x += sprintf(buf + x, " N%d=%lu", 3879 node, nodes[node]); 3880#endif 3881 up_read(&slub_lock); 3882 kfree(nodes); 3883 return x + sprintf(buf + x, "\n"); 3884} 3885 3886#ifdef CONFIG_SLUB_DEBUG 3887static int any_slab_objects(struct kmem_cache *s) 3888{ 3889 int node; 3890 3891 for_each_online_node(node) { 3892 struct kmem_cache_node *n = get_node(s, node); 3893 3894 if (!n) 3895 continue; 3896 3897 if (atomic_long_read(&n->total_objects)) 3898 return 1; 3899 } 3900 return 0; 3901} 3902#endif 3903 3904#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3905#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3906 3907struct slab_attribute { 3908 struct attribute attr; 3909 ssize_t (*show)(struct kmem_cache *s, char *buf); 3910 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3911}; 3912 3913#define SLAB_ATTR_RO(_name) \ 3914 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3915 3916#define SLAB_ATTR(_name) \ 3917 static struct slab_attribute _name##_attr = \ 3918 __ATTR(_name, 0644, _name##_show, _name##_store) 3919 3920static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3921{ 3922 return sprintf(buf, "%d\n", s->size); 3923} 3924SLAB_ATTR_RO(slab_size); 3925 3926static ssize_t align_show(struct kmem_cache *s, char *buf) 3927{ 3928 return sprintf(buf, "%d\n", s->align); 3929} 3930SLAB_ATTR_RO(align); 3931 3932static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3933{ 3934 return sprintf(buf, "%d\n", s->objsize); 3935} 3936SLAB_ATTR_RO(object_size); 3937 3938static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3939{ 3940 return sprintf(buf, "%d\n", oo_objects(s->oo)); 3941} 3942SLAB_ATTR_RO(objs_per_slab); 3943 3944static ssize_t order_store(struct kmem_cache *s, 3945 const char *buf, size_t length) 3946{ 3947 unsigned long order; 3948 int err; 3949 3950 err = strict_strtoul(buf, 10, &order); 3951 if (err) 3952 return err; 3953 3954 if (order > slub_max_order || order < slub_min_order) 3955 return -EINVAL; 3956 3957 calculate_sizes(s, order); 3958 return length; 3959} 3960 3961static ssize_t order_show(struct kmem_cache *s, char *buf) 3962{ 3963 return sprintf(buf, "%d\n", oo_order(s->oo)); 3964} 3965SLAB_ATTR(order); 3966 3967static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 3968{ 3969 return sprintf(buf, "%lu\n", s->min_partial); 3970} 3971 3972static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 3973 size_t length) 3974{ 3975 unsigned long min; 3976 int err; 3977 3978 err = strict_strtoul(buf, 10, &min); 3979 if (err) 3980 return err; 3981 3982 set_min_partial(s, min); 3983 return length; 3984} 3985SLAB_ATTR(min_partial); 3986 3987static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3988{ 3989 if (s->ctor) { 3990 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3991 3992 return n + sprintf(buf + n, "\n"); 3993 } 3994 return 0; 3995} 3996SLAB_ATTR_RO(ctor); 3997 3998static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3999{ 4000 return sprintf(buf, "%d\n", s->refcount - 1); 4001} 4002SLAB_ATTR_RO(aliases); 4003 4004static ssize_t partial_show(struct kmem_cache *s, char *buf) 4005{ 4006 return show_slab_objects(s, buf, SO_PARTIAL); 4007} 4008SLAB_ATTR_RO(partial); 4009 4010static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4011{ 4012 return show_slab_objects(s, buf, SO_CPU); 4013} 4014SLAB_ATTR_RO(cpu_slabs); 4015 4016static ssize_t objects_show(struct kmem_cache *s, char *buf) 4017{ 4018 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4019} 4020SLAB_ATTR_RO(objects); 4021 4022static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4023{ 4024 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4025} 4026SLAB_ATTR_RO(objects_partial); 4027 4028static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4029{ 4030 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4031} 4032 4033static ssize_t reclaim_account_store(struct kmem_cache *s, 4034 const char *buf, size_t length) 4035{ 4036 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4037 if (buf[0] == '1') 4038 s->flags |= SLAB_RECLAIM_ACCOUNT; 4039 return length; 4040} 4041SLAB_ATTR(reclaim_account); 4042 4043static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4044{ 4045 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4046} 4047SLAB_ATTR_RO(hwcache_align); 4048 4049#ifdef CONFIG_ZONE_DMA 4050static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4051{ 4052 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4053} 4054SLAB_ATTR_RO(cache_dma); 4055#endif 4056 4057static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4058{ 4059 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4060} 4061SLAB_ATTR_RO(destroy_by_rcu); 4062 4063#ifdef CONFIG_SLUB_DEBUG 4064static ssize_t slabs_show(struct kmem_cache *s, char *buf) 4065{ 4066 return show_slab_objects(s, buf, SO_ALL); 4067} 4068SLAB_ATTR_RO(slabs); 4069 4070static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 4071{ 4072 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 4073} 4074SLAB_ATTR_RO(total_objects); 4075 4076static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 4077{ 4078 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 4079} 4080 4081static ssize_t sanity_checks_store(struct kmem_cache *s, 4082 const char *buf, size_t length) 4083{ 4084 s->flags &= ~SLAB_DEBUG_FREE; 4085 if (buf[0] == '1') 4086 s->flags |= SLAB_DEBUG_FREE; 4087 return length; 4088} 4089SLAB_ATTR(sanity_checks); 4090 4091static ssize_t trace_show(struct kmem_cache *s, char *buf) 4092{ 4093 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4094} 4095 4096static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4097 size_t length) 4098{ 4099 s->flags &= ~SLAB_TRACE; 4100 if (buf[0] == '1') 4101 s->flags |= SLAB_TRACE; 4102 return length; 4103} 4104SLAB_ATTR(trace); 4105 4106static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4107{ 4108 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4109} 4110 4111static ssize_t red_zone_store(struct kmem_cache *s, 4112 const char *buf, size_t length) 4113{ 4114 if (any_slab_objects(s)) 4115 return -EBUSY; 4116 4117 s->flags &= ~SLAB_RED_ZONE; 4118 if (buf[0] == '1') 4119 s->flags |= SLAB_RED_ZONE; 4120 calculate_sizes(s, -1); 4121 return length; 4122} 4123SLAB_ATTR(red_zone); 4124 4125static ssize_t poison_show(struct kmem_cache *s, char *buf) 4126{ 4127 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4128} 4129 4130static ssize_t poison_store(struct kmem_cache *s, 4131 const char *buf, size_t length) 4132{ 4133 if (any_slab_objects(s)) 4134 return -EBUSY; 4135 4136 s->flags &= ~SLAB_POISON; 4137 if (buf[0] == '1') 4138 s->flags |= SLAB_POISON; 4139 calculate_sizes(s, -1); 4140 return length; 4141} 4142SLAB_ATTR(poison); 4143 4144static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4145{ 4146 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4147} 4148 4149static ssize_t store_user_store(struct kmem_cache *s, 4150 const char *buf, size_t length) 4151{ 4152 if (any_slab_objects(s)) 4153 return -EBUSY; 4154 4155 s->flags &= ~SLAB_STORE_USER; 4156 if (buf[0] == '1') 4157 s->flags |= SLAB_STORE_USER; 4158 calculate_sizes(s, -1); 4159 return length; 4160} 4161SLAB_ATTR(store_user); 4162 4163static ssize_t validate_show(struct kmem_cache *s, char *buf) 4164{ 4165 return 0; 4166} 4167 4168static ssize_t validate_store(struct kmem_cache *s, 4169 const char *buf, size_t length) 4170{ 4171 int ret = -EINVAL; 4172 4173 if (buf[0] == '1') { 4174 ret = validate_slab_cache(s); 4175 if (ret >= 0) 4176 ret = length; 4177 } 4178 return ret; 4179} 4180SLAB_ATTR(validate); 4181 4182static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4183{ 4184 if (!(s->flags & SLAB_STORE_USER)) 4185 return -ENOSYS; 4186 return list_locations(s, buf, TRACK_ALLOC); 4187} 4188SLAB_ATTR_RO(alloc_calls); 4189 4190static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4191{ 4192 if (!(s->flags & SLAB_STORE_USER)) 4193 return -ENOSYS; 4194 return list_locations(s, buf, TRACK_FREE); 4195} 4196SLAB_ATTR_RO(free_calls); 4197#endif /* CONFIG_SLUB_DEBUG */ 4198 4199#ifdef CONFIG_FAILSLAB 4200static ssize_t failslab_show(struct kmem_cache *s, char *buf) 4201{ 4202 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 4203} 4204 4205static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 4206 size_t length) 4207{ 4208 s->flags &= ~SLAB_FAILSLAB; 4209 if (buf[0] == '1') 4210 s->flags |= SLAB_FAILSLAB; 4211 return length; 4212} 4213SLAB_ATTR(failslab); 4214#endif 4215 4216static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4217{ 4218 return 0; 4219} 4220 4221static ssize_t shrink_store(struct kmem_cache *s, 4222 const char *buf, size_t length) 4223{ 4224 if (buf[0] == '1') { 4225 int rc = kmem_cache_shrink(s); 4226 4227 if (rc) 4228 return rc; 4229 } else 4230 return -EINVAL; 4231 return length; 4232} 4233SLAB_ATTR(shrink); 4234 4235#ifdef CONFIG_NUMA 4236static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4237{ 4238 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4239} 4240 4241static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4242 const char *buf, size_t length) 4243{ 4244 unsigned long ratio; 4245 int err; 4246 4247 err = strict_strtoul(buf, 10, &ratio); 4248 if (err) 4249 return err; 4250 4251 if (ratio <= 100) 4252 s->remote_node_defrag_ratio = ratio * 10; 4253 4254 return length; 4255} 4256SLAB_ATTR(remote_node_defrag_ratio); 4257#endif 4258 4259#ifdef CONFIG_SLUB_STATS 4260static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4261{ 4262 unsigned long sum = 0; 4263 int cpu; 4264 int len; 4265 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4266 4267 if (!data) 4268 return -ENOMEM; 4269 4270 for_each_online_cpu(cpu) { 4271 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 4272 4273 data[cpu] = x; 4274 sum += x; 4275 } 4276 4277 len = sprintf(buf, "%lu", sum); 4278 4279#ifdef CONFIG_SMP 4280 for_each_online_cpu(cpu) { 4281 if (data[cpu] && len < PAGE_SIZE - 20) 4282 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4283 } 4284#endif 4285 kfree(data); 4286 return len + sprintf(buf + len, "\n"); 4287} 4288 4289static void clear_stat(struct kmem_cache *s, enum stat_item si) 4290{ 4291 int cpu; 4292 4293 for_each_online_cpu(cpu) 4294 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 4295} 4296 4297#define STAT_ATTR(si, text) \ 4298static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4299{ \ 4300 return show_stat(s, buf, si); \ 4301} \ 4302static ssize_t text##_store(struct kmem_cache *s, \ 4303 const char *buf, size_t length) \ 4304{ \ 4305 if (buf[0] != '0') \ 4306 return -EINVAL; \ 4307 clear_stat(s, si); \ 4308 return length; \ 4309} \ 4310SLAB_ATTR(text); \ 4311 4312STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4313STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4314STAT_ATTR(FREE_FASTPATH, free_fastpath); 4315STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4316STAT_ATTR(FREE_FROZEN, free_frozen); 4317STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4318STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4319STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4320STAT_ATTR(ALLOC_SLAB, alloc_slab); 4321STAT_ATTR(ALLOC_REFILL, alloc_refill); 4322STAT_ATTR(FREE_SLAB, free_slab); 4323STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4324STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4325STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4326STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4327STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4328STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4329STAT_ATTR(ORDER_FALLBACK, order_fallback); 4330#endif 4331 4332static struct attribute *slab_attrs[] = { 4333 &slab_size_attr.attr, 4334 &object_size_attr.attr, 4335 &objs_per_slab_attr.attr, 4336 &order_attr.attr, 4337 &min_partial_attr.attr, 4338 &objects_attr.attr, 4339 &objects_partial_attr.attr, 4340 &partial_attr.attr, 4341 &cpu_slabs_attr.attr, 4342 &ctor_attr.attr, 4343 &aliases_attr.attr, 4344 &align_attr.attr, 4345 &hwcache_align_attr.attr, 4346 &reclaim_account_attr.attr, 4347 &destroy_by_rcu_attr.attr, 4348 &shrink_attr.attr, 4349#ifdef CONFIG_SLUB_DEBUG 4350 &total_objects_attr.attr, 4351 &slabs_attr.attr, 4352 &sanity_checks_attr.attr, 4353 &trace_attr.attr, 4354 &red_zone_attr.attr, 4355 &poison_attr.attr, 4356 &store_user_attr.attr, 4357 &validate_attr.attr, 4358 &alloc_calls_attr.attr, 4359 &free_calls_attr.attr, 4360#endif 4361#ifdef CONFIG_ZONE_DMA 4362 &cache_dma_attr.attr, 4363#endif 4364#ifdef CONFIG_NUMA 4365 &remote_node_defrag_ratio_attr.attr, 4366#endif 4367#ifdef CONFIG_SLUB_STATS 4368 &alloc_fastpath_attr.attr, 4369 &alloc_slowpath_attr.attr, 4370 &free_fastpath_attr.attr, 4371 &free_slowpath_attr.attr, 4372 &free_frozen_attr.attr, 4373 &free_add_partial_attr.attr, 4374 &free_remove_partial_attr.attr, 4375 &alloc_from_partial_attr.attr, 4376 &alloc_slab_attr.attr, 4377 &alloc_refill_attr.attr, 4378 &free_slab_attr.attr, 4379 &cpuslab_flush_attr.attr, 4380 &deactivate_full_attr.attr, 4381 &deactivate_empty_attr.attr, 4382 &deactivate_to_head_attr.attr, 4383 &deactivate_to_tail_attr.attr, 4384 &deactivate_remote_frees_attr.attr, 4385 &order_fallback_attr.attr, 4386#endif 4387#ifdef CONFIG_FAILSLAB 4388 &failslab_attr.attr, 4389#endif 4390 4391 NULL 4392}; 4393 4394static struct attribute_group slab_attr_group = { 4395 .attrs = slab_attrs, 4396}; 4397 4398static ssize_t slab_attr_show(struct kobject *kobj, 4399 struct attribute *attr, 4400 char *buf) 4401{ 4402 struct slab_attribute *attribute; 4403 struct kmem_cache *s; 4404 int err; 4405 4406 attribute = to_slab_attr(attr); 4407 s = to_slab(kobj); 4408 4409 if (!attribute->show) 4410 return -EIO; 4411 4412 err = attribute->show(s, buf); 4413 4414 return err; 4415} 4416 4417static ssize_t slab_attr_store(struct kobject *kobj, 4418 struct attribute *attr, 4419 const char *buf, size_t len) 4420{ 4421 struct slab_attribute *attribute; 4422 struct kmem_cache *s; 4423 int err; 4424 4425 attribute = to_slab_attr(attr); 4426 s = to_slab(kobj); 4427 4428 if (!attribute->store) 4429 return -EIO; 4430 4431 err = attribute->store(s, buf, len); 4432 4433 return err; 4434} 4435 4436static void kmem_cache_release(struct kobject *kobj) 4437{ 4438 struct kmem_cache *s = to_slab(kobj); 4439 4440 kfree(s->name); 4441 kfree(s); 4442} 4443 4444static const struct sysfs_ops slab_sysfs_ops = { 4445 .show = slab_attr_show, 4446 .store = slab_attr_store, 4447}; 4448 4449static struct kobj_type slab_ktype = { 4450 .sysfs_ops = &slab_sysfs_ops, 4451 .release = kmem_cache_release 4452}; 4453 4454static int uevent_filter(struct kset *kset, struct kobject *kobj) 4455{ 4456 struct kobj_type *ktype = get_ktype(kobj); 4457 4458 if (ktype == &slab_ktype) 4459 return 1; 4460 return 0; 4461} 4462 4463static const struct kset_uevent_ops slab_uevent_ops = { 4464 .filter = uevent_filter, 4465}; 4466 4467static struct kset *slab_kset; 4468 4469#define ID_STR_LENGTH 64 4470 4471/* Create a unique string id for a slab cache: 4472 * 4473 * Format :[flags-]size 4474 */ 4475static char *create_unique_id(struct kmem_cache *s) 4476{ 4477 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4478 char *p = name; 4479 4480 BUG_ON(!name); 4481 4482 *p++ = ':'; 4483 /* 4484 * First flags affecting slabcache operations. We will only 4485 * get here for aliasable slabs so we do not need to support 4486 * too many flags. The flags here must cover all flags that 4487 * are matched during merging to guarantee that the id is 4488 * unique. 4489 */ 4490 if (s->flags & SLAB_CACHE_DMA) 4491 *p++ = 'd'; 4492 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4493 *p++ = 'a'; 4494 if (s->flags & SLAB_DEBUG_FREE) 4495 *p++ = 'F'; 4496 if (!(s->flags & SLAB_NOTRACK)) 4497 *p++ = 't'; 4498 if (p != name + 1) 4499 *p++ = '-'; 4500 p += sprintf(p, "%07d", s->size); 4501 BUG_ON(p > name + ID_STR_LENGTH - 1); 4502 return name; 4503} 4504 4505static int sysfs_slab_add(struct kmem_cache *s) 4506{ 4507 int err; 4508 const char *name; 4509 int unmergeable; 4510 4511 if (slab_state < SYSFS) 4512 /* Defer until later */ 4513 return 0; 4514 4515 unmergeable = slab_unmergeable(s); 4516 if (unmergeable) { 4517 /* 4518 * Slabcache can never be merged so we can use the name proper. 4519 * This is typically the case for debug situations. In that 4520 * case we can catch duplicate names easily. 4521 */ 4522 sysfs_remove_link(&slab_kset->kobj, s->name); 4523 name = s->name; 4524 } else { 4525 /* 4526 * Create a unique name for the slab as a target 4527 * for the symlinks. 4528 */ 4529 name = create_unique_id(s); 4530 } 4531 4532 s->kobj.kset = slab_kset; 4533 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4534 if (err) { 4535 kobject_put(&s->kobj); 4536 return err; 4537 } 4538 4539 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4540 if (err) { 4541 kobject_del(&s->kobj); 4542 kobject_put(&s->kobj); 4543 return err; 4544 } 4545 kobject_uevent(&s->kobj, KOBJ_ADD); 4546 if (!unmergeable) { 4547 /* Setup first alias */ 4548 sysfs_slab_alias(s, s->name); 4549 kfree(name); 4550 } 4551 return 0; 4552} 4553 4554static void sysfs_slab_remove(struct kmem_cache *s) 4555{ 4556 if (slab_state < SYSFS) 4557 /* 4558 * Sysfs has not been setup yet so no need to remove the 4559 * cache from sysfs. 4560 */ 4561 return; 4562 4563 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4564 kobject_del(&s->kobj); 4565 kobject_put(&s->kobj); 4566} 4567 4568/* 4569 * Need to buffer aliases during bootup until sysfs becomes 4570 * available lest we lose that information. 4571 */ 4572struct saved_alias { 4573 struct kmem_cache *s; 4574 const char *name; 4575 struct saved_alias *next; 4576}; 4577 4578static struct saved_alias *alias_list; 4579 4580static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4581{ 4582 struct saved_alias *al; 4583 4584 if (slab_state == SYSFS) { 4585 /* 4586 * If we have a leftover link then remove it. 4587 */ 4588 sysfs_remove_link(&slab_kset->kobj, name); 4589 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4590 } 4591 4592 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4593 if (!al) 4594 return -ENOMEM; 4595 4596 al->s = s; 4597 al->name = name; 4598 al->next = alias_list; 4599 alias_list = al; 4600 return 0; 4601} 4602 4603static int __init slab_sysfs_init(void) 4604{ 4605 struct kmem_cache *s; 4606 int err; 4607 4608 down_write(&slub_lock); 4609 4610 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4611 if (!slab_kset) { 4612 up_write(&slub_lock); 4613 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4614 return -ENOSYS; 4615 } 4616 4617 slab_state = SYSFS; 4618 4619 list_for_each_entry(s, &slab_caches, list) { 4620 err = sysfs_slab_add(s); 4621 if (err) 4622 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4623 " to sysfs\n", s->name); 4624 } 4625 4626 while (alias_list) { 4627 struct saved_alias *al = alias_list; 4628 4629 alias_list = alias_list->next; 4630 err = sysfs_slab_alias(al->s, al->name); 4631 if (err) 4632 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4633 " %s to sysfs\n", s->name); 4634 kfree(al); 4635 } 4636 4637 up_write(&slub_lock); 4638 resiliency_test(); 4639 return 0; 4640} 4641 4642__initcall(slab_sysfs_init); 4643#endif /* CONFIG_SYSFS */ 4644 4645/* 4646 * The /proc/slabinfo ABI 4647 */ 4648#ifdef CONFIG_SLABINFO 4649static void print_slabinfo_header(struct seq_file *m) 4650{ 4651 seq_puts(m, "slabinfo - version: 2.1\n"); 4652 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4653 "<objperslab> <pagesperslab>"); 4654 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4655 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4656 seq_putc(m, '\n'); 4657} 4658 4659static void *s_start(struct seq_file *m, loff_t *pos) 4660{ 4661 loff_t n = *pos; 4662 4663 down_read(&slub_lock); 4664 if (!n) 4665 print_slabinfo_header(m); 4666 4667 return seq_list_start(&slab_caches, *pos); 4668} 4669 4670static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4671{ 4672 return seq_list_next(p, &slab_caches, pos); 4673} 4674 4675static void s_stop(struct seq_file *m, void *p) 4676{ 4677 up_read(&slub_lock); 4678} 4679 4680static int s_show(struct seq_file *m, void *p) 4681{ 4682 unsigned long nr_partials = 0; 4683 unsigned long nr_slabs = 0; 4684 unsigned long nr_inuse = 0; 4685 unsigned long nr_objs = 0; 4686 unsigned long nr_free = 0; 4687 struct kmem_cache *s; 4688 int node; 4689 4690 s = list_entry(p, struct kmem_cache, list); 4691 4692 for_each_online_node(node) { 4693 struct kmem_cache_node *n = get_node(s, node); 4694 4695 if (!n) 4696 continue; 4697 4698 nr_partials += n->nr_partial; 4699 nr_slabs += atomic_long_read(&n->nr_slabs); 4700 nr_objs += atomic_long_read(&n->total_objects); 4701 nr_free += count_partial(n, count_free); 4702 } 4703 4704 nr_inuse = nr_objs - nr_free; 4705 4706 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4707 nr_objs, s->size, oo_objects(s->oo), 4708 (1 << oo_order(s->oo))); 4709 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4710 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4711 0UL); 4712 seq_putc(m, '\n'); 4713 return 0; 4714} 4715 4716static const struct seq_operations slabinfo_op = { 4717 .start = s_start, 4718 .next = s_next, 4719 .stop = s_stop, 4720 .show = s_show, 4721}; 4722 4723static int slabinfo_open(struct inode *inode, struct file *file) 4724{ 4725 return seq_open(file, &slabinfo_op); 4726} 4727 4728static const struct file_operations proc_slabinfo_operations = { 4729 .open = slabinfo_open, 4730 .read = seq_read, 4731 .llseek = seq_lseek, 4732 .release = seq_release, 4733}; 4734 4735static int __init slab_proc_init(void) 4736{ 4737 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations); 4738 return 0; 4739} 4740module_init(slab_proc_init); 4741#endif /* CONFIG_SLABINFO */ 4742