slub.c revision 231367fd9bccbb36309ab5bf5012e11a84231031
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/module.h> 13#include <linux/bit_spinlock.h> 14#include <linux/interrupt.h> 15#include <linux/bitops.h> 16#include <linux/slab.h> 17#include <linux/seq_file.h> 18#include <linux/cpu.h> 19#include <linux/cpuset.h> 20#include <linux/mempolicy.h> 21#include <linux/ctype.h> 22#include <linux/debugobjects.h> 23#include <linux/kallsyms.h> 24#include <linux/memory.h> 25#include <linux/math64.h> 26 27/* 28 * Lock order: 29 * 1. slab_lock(page) 30 * 2. slab->list_lock 31 * 32 * The slab_lock protects operations on the object of a particular 33 * slab and its metadata in the page struct. If the slab lock 34 * has been taken then no allocations nor frees can be performed 35 * on the objects in the slab nor can the slab be added or removed 36 * from the partial or full lists since this would mean modifying 37 * the page_struct of the slab. 38 * 39 * The list_lock protects the partial and full list on each node and 40 * the partial slab counter. If taken then no new slabs may be added or 41 * removed from the lists nor make the number of partial slabs be modified. 42 * (Note that the total number of slabs is an atomic value that may be 43 * modified without taking the list lock). 44 * 45 * The list_lock is a centralized lock and thus we avoid taking it as 46 * much as possible. As long as SLUB does not have to handle partial 47 * slabs, operations can continue without any centralized lock. F.e. 48 * allocating a long series of objects that fill up slabs does not require 49 * the list lock. 50 * 51 * The lock order is sometimes inverted when we are trying to get a slab 52 * off a list. We take the list_lock and then look for a page on the list 53 * to use. While we do that objects in the slabs may be freed. We can 54 * only operate on the slab if we have also taken the slab_lock. So we use 55 * a slab_trylock() on the slab. If trylock was successful then no frees 56 * can occur anymore and we can use the slab for allocations etc. If the 57 * slab_trylock() does not succeed then frees are in progress in the slab and 58 * we must stay away from it for a while since we may cause a bouncing 59 * cacheline if we try to acquire the lock. So go onto the next slab. 60 * If all pages are busy then we may allocate a new slab instead of reusing 61 * a partial slab. A new slab has noone operating on it and thus there is 62 * no danger of cacheline contention. 63 * 64 * Interrupts are disabled during allocation and deallocation in order to 65 * make the slab allocator safe to use in the context of an irq. In addition 66 * interrupts are disabled to ensure that the processor does not change 67 * while handling per_cpu slabs, due to kernel preemption. 68 * 69 * SLUB assigns one slab for allocation to each processor. 70 * Allocations only occur from these slabs called cpu slabs. 71 * 72 * Slabs with free elements are kept on a partial list and during regular 73 * operations no list for full slabs is used. If an object in a full slab is 74 * freed then the slab will show up again on the partial lists. 75 * We track full slabs for debugging purposes though because otherwise we 76 * cannot scan all objects. 77 * 78 * Slabs are freed when they become empty. Teardown and setup is 79 * minimal so we rely on the page allocators per cpu caches for 80 * fast frees and allocs. 81 * 82 * Overloading of page flags that are otherwise used for LRU management. 83 * 84 * PageActive The slab is frozen and exempt from list processing. 85 * This means that the slab is dedicated to a purpose 86 * such as satisfying allocations for a specific 87 * processor. Objects may be freed in the slab while 88 * it is frozen but slab_free will then skip the usual 89 * list operations. It is up to the processor holding 90 * the slab to integrate the slab into the slab lists 91 * when the slab is no longer needed. 92 * 93 * One use of this flag is to mark slabs that are 94 * used for allocations. Then such a slab becomes a cpu 95 * slab. The cpu slab may be equipped with an additional 96 * freelist that allows lockless access to 97 * free objects in addition to the regular freelist 98 * that requires the slab lock. 99 * 100 * PageError Slab requires special handling due to debug 101 * options set. This moves slab handling out of 102 * the fast path and disables lockless freelists. 103 */ 104 105#ifdef CONFIG_SLUB_DEBUG 106#define SLABDEBUG 1 107#else 108#define SLABDEBUG 0 109#endif 110 111/* 112 * Issues still to be resolved: 113 * 114 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 115 * 116 * - Variable sizing of the per node arrays 117 */ 118 119/* Enable to test recovery from slab corruption on boot */ 120#undef SLUB_RESILIENCY_TEST 121 122/* 123 * Mininum number of partial slabs. These will be left on the partial 124 * lists even if they are empty. kmem_cache_shrink may reclaim them. 125 */ 126#define MIN_PARTIAL 5 127 128/* 129 * Maximum number of desirable partial slabs. 130 * The existence of more partial slabs makes kmem_cache_shrink 131 * sort the partial list by the number of objects in the. 132 */ 133#define MAX_PARTIAL 10 134 135#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 136 SLAB_POISON | SLAB_STORE_USER) 137 138/* 139 * Set of flags that will prevent slab merging 140 */ 141#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 142 SLAB_TRACE | SLAB_DESTROY_BY_RCU) 143 144#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 145 SLAB_CACHE_DMA) 146 147#ifndef ARCH_KMALLOC_MINALIGN 148#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 149#endif 150 151#ifndef ARCH_SLAB_MINALIGN 152#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) 153#endif 154 155/* Internal SLUB flags */ 156#define __OBJECT_POISON 0x80000000 /* Poison object */ 157#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ 158 159static int kmem_size = sizeof(struct kmem_cache); 160 161#ifdef CONFIG_SMP 162static struct notifier_block slab_notifier; 163#endif 164 165static enum { 166 DOWN, /* No slab functionality available */ 167 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 168 UP, /* Everything works but does not show up in sysfs */ 169 SYSFS /* Sysfs up */ 170} slab_state = DOWN; 171 172/* A list of all slab caches on the system */ 173static DECLARE_RWSEM(slub_lock); 174static LIST_HEAD(slab_caches); 175 176/* 177 * Tracking user of a slab. 178 */ 179struct track { 180 void *addr; /* Called from address */ 181 int cpu; /* Was running on cpu */ 182 int pid; /* Pid context */ 183 unsigned long when; /* When did the operation occur */ 184}; 185 186enum track_item { TRACK_ALLOC, TRACK_FREE }; 187 188#ifdef CONFIG_SLUB_DEBUG 189static int sysfs_slab_add(struct kmem_cache *); 190static int sysfs_slab_alias(struct kmem_cache *, const char *); 191static void sysfs_slab_remove(struct kmem_cache *); 192 193#else 194static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 195static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 196 { return 0; } 197static inline void sysfs_slab_remove(struct kmem_cache *s) 198{ 199 kfree(s); 200} 201 202#endif 203 204static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) 205{ 206#ifdef CONFIG_SLUB_STATS 207 c->stat[si]++; 208#endif 209} 210 211/******************************************************************** 212 * Core slab cache functions 213 *******************************************************************/ 214 215int slab_is_available(void) 216{ 217 return slab_state >= UP; 218} 219 220static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 221{ 222#ifdef CONFIG_NUMA 223 return s->node[node]; 224#else 225 return &s->local_node; 226#endif 227} 228 229static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) 230{ 231#ifdef CONFIG_SMP 232 return s->cpu_slab[cpu]; 233#else 234 return &s->cpu_slab; 235#endif 236} 237 238/* Verify that a pointer has an address that is valid within a slab page */ 239static inline int check_valid_pointer(struct kmem_cache *s, 240 struct page *page, const void *object) 241{ 242 void *base; 243 244 if (!object) 245 return 1; 246 247 base = page_address(page); 248 if (object < base || object >= base + page->objects * s->size || 249 (object - base) % s->size) { 250 return 0; 251 } 252 253 return 1; 254} 255 256/* 257 * Slow version of get and set free pointer. 258 * 259 * This version requires touching the cache lines of kmem_cache which 260 * we avoid to do in the fast alloc free paths. There we obtain the offset 261 * from the page struct. 262 */ 263static inline void *get_freepointer(struct kmem_cache *s, void *object) 264{ 265 return *(void **)(object + s->offset); 266} 267 268static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 269{ 270 *(void **)(object + s->offset) = fp; 271} 272 273/* Loop over all objects in a slab */ 274#define for_each_object(__p, __s, __addr, __objects) \ 275 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 276 __p += (__s)->size) 277 278/* Scan freelist */ 279#define for_each_free_object(__p, __s, __free) \ 280 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 281 282/* Determine object index from a given position */ 283static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 284{ 285 return (p - addr) / s->size; 286} 287 288static inline struct kmem_cache_order_objects oo_make(int order, 289 unsigned long size) 290{ 291 struct kmem_cache_order_objects x = { 292 (order << 16) + (PAGE_SIZE << order) / size 293 }; 294 295 return x; 296} 297 298static inline int oo_order(struct kmem_cache_order_objects x) 299{ 300 return x.x >> 16; 301} 302 303static inline int oo_objects(struct kmem_cache_order_objects x) 304{ 305 return x.x & ((1 << 16) - 1); 306} 307 308#ifdef CONFIG_SLUB_DEBUG 309/* 310 * Debug settings: 311 */ 312#ifdef CONFIG_SLUB_DEBUG_ON 313static int slub_debug = DEBUG_DEFAULT_FLAGS; 314#else 315static int slub_debug; 316#endif 317 318static char *slub_debug_slabs; 319 320/* 321 * Object debugging 322 */ 323static void print_section(char *text, u8 *addr, unsigned int length) 324{ 325 int i, offset; 326 int newline = 1; 327 char ascii[17]; 328 329 ascii[16] = 0; 330 331 for (i = 0; i < length; i++) { 332 if (newline) { 333 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 334 newline = 0; 335 } 336 printk(KERN_CONT " %02x", addr[i]); 337 offset = i % 16; 338 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 339 if (offset == 15) { 340 printk(KERN_CONT " %s\n", ascii); 341 newline = 1; 342 } 343 } 344 if (!newline) { 345 i %= 16; 346 while (i < 16) { 347 printk(KERN_CONT " "); 348 ascii[i] = ' '; 349 i++; 350 } 351 printk(KERN_CONT " %s\n", ascii); 352 } 353} 354 355static struct track *get_track(struct kmem_cache *s, void *object, 356 enum track_item alloc) 357{ 358 struct track *p; 359 360 if (s->offset) 361 p = object + s->offset + sizeof(void *); 362 else 363 p = object + s->inuse; 364 365 return p + alloc; 366} 367 368static void set_track(struct kmem_cache *s, void *object, 369 enum track_item alloc, void *addr) 370{ 371 struct track *p; 372 373 if (s->offset) 374 p = object + s->offset + sizeof(void *); 375 else 376 p = object + s->inuse; 377 378 p += alloc; 379 if (addr) { 380 p->addr = addr; 381 p->cpu = smp_processor_id(); 382 p->pid = current->pid; 383 p->when = jiffies; 384 } else 385 memset(p, 0, sizeof(struct track)); 386} 387 388static void init_tracking(struct kmem_cache *s, void *object) 389{ 390 if (!(s->flags & SLAB_STORE_USER)) 391 return; 392 393 set_track(s, object, TRACK_FREE, NULL); 394 set_track(s, object, TRACK_ALLOC, NULL); 395} 396 397static void print_track(const char *s, struct track *t) 398{ 399 if (!t->addr) 400 return; 401 402 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 403 s, t->addr, jiffies - t->when, t->cpu, t->pid); 404} 405 406static void print_tracking(struct kmem_cache *s, void *object) 407{ 408 if (!(s->flags & SLAB_STORE_USER)) 409 return; 410 411 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 412 print_track("Freed", get_track(s, object, TRACK_FREE)); 413} 414 415static void print_page_info(struct page *page) 416{ 417 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 418 page, page->objects, page->inuse, page->freelist, page->flags); 419 420} 421 422static void slab_bug(struct kmem_cache *s, char *fmt, ...) 423{ 424 va_list args; 425 char buf[100]; 426 427 va_start(args, fmt); 428 vsnprintf(buf, sizeof(buf), fmt, args); 429 va_end(args); 430 printk(KERN_ERR "========================================" 431 "=====================================\n"); 432 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 433 printk(KERN_ERR "----------------------------------------" 434 "-------------------------------------\n\n"); 435} 436 437static void slab_fix(struct kmem_cache *s, char *fmt, ...) 438{ 439 va_list args; 440 char buf[100]; 441 442 va_start(args, fmt); 443 vsnprintf(buf, sizeof(buf), fmt, args); 444 va_end(args); 445 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 446} 447 448static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 449{ 450 unsigned int off; /* Offset of last byte */ 451 u8 *addr = page_address(page); 452 453 print_tracking(s, p); 454 455 print_page_info(page); 456 457 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 458 p, p - addr, get_freepointer(s, p)); 459 460 if (p > addr + 16) 461 print_section("Bytes b4", p - 16, 16); 462 463 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 464 465 if (s->flags & SLAB_RED_ZONE) 466 print_section("Redzone", p + s->objsize, 467 s->inuse - s->objsize); 468 469 if (s->offset) 470 off = s->offset + sizeof(void *); 471 else 472 off = s->inuse; 473 474 if (s->flags & SLAB_STORE_USER) 475 off += 2 * sizeof(struct track); 476 477 if (off != s->size) 478 /* Beginning of the filler is the free pointer */ 479 print_section("Padding", p + off, s->size - off); 480 481 dump_stack(); 482} 483 484static void object_err(struct kmem_cache *s, struct page *page, 485 u8 *object, char *reason) 486{ 487 slab_bug(s, "%s", reason); 488 print_trailer(s, page, object); 489} 490 491static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 492{ 493 va_list args; 494 char buf[100]; 495 496 va_start(args, fmt); 497 vsnprintf(buf, sizeof(buf), fmt, args); 498 va_end(args); 499 slab_bug(s, "%s", buf); 500 print_page_info(page); 501 dump_stack(); 502} 503 504static void init_object(struct kmem_cache *s, void *object, int active) 505{ 506 u8 *p = object; 507 508 if (s->flags & __OBJECT_POISON) { 509 memset(p, POISON_FREE, s->objsize - 1); 510 p[s->objsize - 1] = POISON_END; 511 } 512 513 if (s->flags & SLAB_RED_ZONE) 514 memset(p + s->objsize, 515 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 516 s->inuse - s->objsize); 517} 518 519static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 520{ 521 while (bytes) { 522 if (*start != (u8)value) 523 return start; 524 start++; 525 bytes--; 526 } 527 return NULL; 528} 529 530static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 531 void *from, void *to) 532{ 533 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 534 memset(from, data, to - from); 535} 536 537static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 538 u8 *object, char *what, 539 u8 *start, unsigned int value, unsigned int bytes) 540{ 541 u8 *fault; 542 u8 *end; 543 544 fault = check_bytes(start, value, bytes); 545 if (!fault) 546 return 1; 547 548 end = start + bytes; 549 while (end > fault && end[-1] == value) 550 end--; 551 552 slab_bug(s, "%s overwritten", what); 553 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 554 fault, end - 1, fault[0], value); 555 print_trailer(s, page, object); 556 557 restore_bytes(s, what, value, fault, end); 558 return 0; 559} 560 561/* 562 * Object layout: 563 * 564 * object address 565 * Bytes of the object to be managed. 566 * If the freepointer may overlay the object then the free 567 * pointer is the first word of the object. 568 * 569 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 570 * 0xa5 (POISON_END) 571 * 572 * object + s->objsize 573 * Padding to reach word boundary. This is also used for Redzoning. 574 * Padding is extended by another word if Redzoning is enabled and 575 * objsize == inuse. 576 * 577 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 578 * 0xcc (RED_ACTIVE) for objects in use. 579 * 580 * object + s->inuse 581 * Meta data starts here. 582 * 583 * A. Free pointer (if we cannot overwrite object on free) 584 * B. Tracking data for SLAB_STORE_USER 585 * C. Padding to reach required alignment boundary or at mininum 586 * one word if debugging is on to be able to detect writes 587 * before the word boundary. 588 * 589 * Padding is done using 0x5a (POISON_INUSE) 590 * 591 * object + s->size 592 * Nothing is used beyond s->size. 593 * 594 * If slabcaches are merged then the objsize and inuse boundaries are mostly 595 * ignored. And therefore no slab options that rely on these boundaries 596 * may be used with merged slabcaches. 597 */ 598 599static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 600{ 601 unsigned long off = s->inuse; /* The end of info */ 602 603 if (s->offset) 604 /* Freepointer is placed after the object. */ 605 off += sizeof(void *); 606 607 if (s->flags & SLAB_STORE_USER) 608 /* We also have user information there */ 609 off += 2 * sizeof(struct track); 610 611 if (s->size == off) 612 return 1; 613 614 return check_bytes_and_report(s, page, p, "Object padding", 615 p + off, POISON_INUSE, s->size - off); 616} 617 618/* Check the pad bytes at the end of a slab page */ 619static int slab_pad_check(struct kmem_cache *s, struct page *page) 620{ 621 u8 *start; 622 u8 *fault; 623 u8 *end; 624 int length; 625 int remainder; 626 627 if (!(s->flags & SLAB_POISON)) 628 return 1; 629 630 start = page_address(page); 631 length = (PAGE_SIZE << compound_order(page)); 632 end = start + length; 633 remainder = length % s->size; 634 if (!remainder) 635 return 1; 636 637 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 638 if (!fault) 639 return 1; 640 while (end > fault && end[-1] == POISON_INUSE) 641 end--; 642 643 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 644 print_section("Padding", end - remainder, remainder); 645 646 restore_bytes(s, "slab padding", POISON_INUSE, start, end); 647 return 0; 648} 649 650static int check_object(struct kmem_cache *s, struct page *page, 651 void *object, int active) 652{ 653 u8 *p = object; 654 u8 *endobject = object + s->objsize; 655 656 if (s->flags & SLAB_RED_ZONE) { 657 unsigned int red = 658 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 659 660 if (!check_bytes_and_report(s, page, object, "Redzone", 661 endobject, red, s->inuse - s->objsize)) 662 return 0; 663 } else { 664 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 665 check_bytes_and_report(s, page, p, "Alignment padding", 666 endobject, POISON_INUSE, s->inuse - s->objsize); 667 } 668 } 669 670 if (s->flags & SLAB_POISON) { 671 if (!active && (s->flags & __OBJECT_POISON) && 672 (!check_bytes_and_report(s, page, p, "Poison", p, 673 POISON_FREE, s->objsize - 1) || 674 !check_bytes_and_report(s, page, p, "Poison", 675 p + s->objsize - 1, POISON_END, 1))) 676 return 0; 677 /* 678 * check_pad_bytes cleans up on its own. 679 */ 680 check_pad_bytes(s, page, p); 681 } 682 683 if (!s->offset && active) 684 /* 685 * Object and freepointer overlap. Cannot check 686 * freepointer while object is allocated. 687 */ 688 return 1; 689 690 /* Check free pointer validity */ 691 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 692 object_err(s, page, p, "Freepointer corrupt"); 693 /* 694 * No choice but to zap it and thus loose the remainder 695 * of the free objects in this slab. May cause 696 * another error because the object count is now wrong. 697 */ 698 set_freepointer(s, p, NULL); 699 return 0; 700 } 701 return 1; 702} 703 704static int check_slab(struct kmem_cache *s, struct page *page) 705{ 706 int maxobj; 707 708 VM_BUG_ON(!irqs_disabled()); 709 710 if (!PageSlab(page)) { 711 slab_err(s, page, "Not a valid slab page"); 712 return 0; 713 } 714 715 maxobj = (PAGE_SIZE << compound_order(page)) / s->size; 716 if (page->objects > maxobj) { 717 slab_err(s, page, "objects %u > max %u", 718 s->name, page->objects, maxobj); 719 return 0; 720 } 721 if (page->inuse > page->objects) { 722 slab_err(s, page, "inuse %u > max %u", 723 s->name, page->inuse, page->objects); 724 return 0; 725 } 726 /* Slab_pad_check fixes things up after itself */ 727 slab_pad_check(s, page); 728 return 1; 729} 730 731/* 732 * Determine if a certain object on a page is on the freelist. Must hold the 733 * slab lock to guarantee that the chains are in a consistent state. 734 */ 735static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 736{ 737 int nr = 0; 738 void *fp = page->freelist; 739 void *object = NULL; 740 unsigned long max_objects; 741 742 while (fp && nr <= page->objects) { 743 if (fp == search) 744 return 1; 745 if (!check_valid_pointer(s, page, fp)) { 746 if (object) { 747 object_err(s, page, object, 748 "Freechain corrupt"); 749 set_freepointer(s, object, NULL); 750 break; 751 } else { 752 slab_err(s, page, "Freepointer corrupt"); 753 page->freelist = NULL; 754 page->inuse = page->objects; 755 slab_fix(s, "Freelist cleared"); 756 return 0; 757 } 758 break; 759 } 760 object = fp; 761 fp = get_freepointer(s, object); 762 nr++; 763 } 764 765 max_objects = (PAGE_SIZE << compound_order(page)) / s->size; 766 if (max_objects > 65535) 767 max_objects = 65535; 768 769 if (page->objects != max_objects) { 770 slab_err(s, page, "Wrong number of objects. Found %d but " 771 "should be %d", page->objects, max_objects); 772 page->objects = max_objects; 773 slab_fix(s, "Number of objects adjusted."); 774 } 775 if (page->inuse != page->objects - nr) { 776 slab_err(s, page, "Wrong object count. Counter is %d but " 777 "counted were %d", page->inuse, page->objects - nr); 778 page->inuse = page->objects - nr; 779 slab_fix(s, "Object count adjusted."); 780 } 781 return search == NULL; 782} 783 784static void trace(struct kmem_cache *s, struct page *page, void *object, 785 int alloc) 786{ 787 if (s->flags & SLAB_TRACE) { 788 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 789 s->name, 790 alloc ? "alloc" : "free", 791 object, page->inuse, 792 page->freelist); 793 794 if (!alloc) 795 print_section("Object", (void *)object, s->objsize); 796 797 dump_stack(); 798 } 799} 800 801/* 802 * Tracking of fully allocated slabs for debugging purposes. 803 */ 804static void add_full(struct kmem_cache_node *n, struct page *page) 805{ 806 spin_lock(&n->list_lock); 807 list_add(&page->lru, &n->full); 808 spin_unlock(&n->list_lock); 809} 810 811static void remove_full(struct kmem_cache *s, struct page *page) 812{ 813 struct kmem_cache_node *n; 814 815 if (!(s->flags & SLAB_STORE_USER)) 816 return; 817 818 n = get_node(s, page_to_nid(page)); 819 820 spin_lock(&n->list_lock); 821 list_del(&page->lru); 822 spin_unlock(&n->list_lock); 823} 824 825/* Tracking of the number of slabs for debugging purposes */ 826static inline unsigned long slabs_node(struct kmem_cache *s, int node) 827{ 828 struct kmem_cache_node *n = get_node(s, node); 829 830 return atomic_long_read(&n->nr_slabs); 831} 832 833static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 834{ 835 struct kmem_cache_node *n = get_node(s, node); 836 837 /* 838 * May be called early in order to allocate a slab for the 839 * kmem_cache_node structure. Solve the chicken-egg 840 * dilemma by deferring the increment of the count during 841 * bootstrap (see early_kmem_cache_node_alloc). 842 */ 843 if (!NUMA_BUILD || n) { 844 atomic_long_inc(&n->nr_slabs); 845 atomic_long_add(objects, &n->total_objects); 846 } 847} 848static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 849{ 850 struct kmem_cache_node *n = get_node(s, node); 851 852 atomic_long_dec(&n->nr_slabs); 853 atomic_long_sub(objects, &n->total_objects); 854} 855 856/* Object debug checks for alloc/free paths */ 857static void setup_object_debug(struct kmem_cache *s, struct page *page, 858 void *object) 859{ 860 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 861 return; 862 863 init_object(s, object, 0); 864 init_tracking(s, object); 865} 866 867static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 868 void *object, void *addr) 869{ 870 if (!check_slab(s, page)) 871 goto bad; 872 873 if (!on_freelist(s, page, object)) { 874 object_err(s, page, object, "Object already allocated"); 875 goto bad; 876 } 877 878 if (!check_valid_pointer(s, page, object)) { 879 object_err(s, page, object, "Freelist Pointer check fails"); 880 goto bad; 881 } 882 883 if (!check_object(s, page, object, 0)) 884 goto bad; 885 886 /* Success perform special debug activities for allocs */ 887 if (s->flags & SLAB_STORE_USER) 888 set_track(s, object, TRACK_ALLOC, addr); 889 trace(s, page, object, 1); 890 init_object(s, object, 1); 891 return 1; 892 893bad: 894 if (PageSlab(page)) { 895 /* 896 * If this is a slab page then lets do the best we can 897 * to avoid issues in the future. Marking all objects 898 * as used avoids touching the remaining objects. 899 */ 900 slab_fix(s, "Marking all objects used"); 901 page->inuse = page->objects; 902 page->freelist = NULL; 903 } 904 return 0; 905} 906 907static int free_debug_processing(struct kmem_cache *s, struct page *page, 908 void *object, void *addr) 909{ 910 if (!check_slab(s, page)) 911 goto fail; 912 913 if (!check_valid_pointer(s, page, object)) { 914 slab_err(s, page, "Invalid object pointer 0x%p", object); 915 goto fail; 916 } 917 918 if (on_freelist(s, page, object)) { 919 object_err(s, page, object, "Object already free"); 920 goto fail; 921 } 922 923 if (!check_object(s, page, object, 1)) 924 return 0; 925 926 if (unlikely(s != page->slab)) { 927 if (!PageSlab(page)) { 928 slab_err(s, page, "Attempt to free object(0x%p) " 929 "outside of slab", object); 930 } else if (!page->slab) { 931 printk(KERN_ERR 932 "SLUB <none>: no slab for object 0x%p.\n", 933 object); 934 dump_stack(); 935 } else 936 object_err(s, page, object, 937 "page slab pointer corrupt."); 938 goto fail; 939 } 940 941 /* Special debug activities for freeing objects */ 942 if (!PageSlubFrozen(page) && !page->freelist) 943 remove_full(s, page); 944 if (s->flags & SLAB_STORE_USER) 945 set_track(s, object, TRACK_FREE, addr); 946 trace(s, page, object, 0); 947 init_object(s, object, 0); 948 return 1; 949 950fail: 951 slab_fix(s, "Object at 0x%p not freed", object); 952 return 0; 953} 954 955static int __init setup_slub_debug(char *str) 956{ 957 slub_debug = DEBUG_DEFAULT_FLAGS; 958 if (*str++ != '=' || !*str) 959 /* 960 * No options specified. Switch on full debugging. 961 */ 962 goto out; 963 964 if (*str == ',') 965 /* 966 * No options but restriction on slabs. This means full 967 * debugging for slabs matching a pattern. 968 */ 969 goto check_slabs; 970 971 slub_debug = 0; 972 if (*str == '-') 973 /* 974 * Switch off all debugging measures. 975 */ 976 goto out; 977 978 /* 979 * Determine which debug features should be switched on 980 */ 981 for (; *str && *str != ','; str++) { 982 switch (tolower(*str)) { 983 case 'f': 984 slub_debug |= SLAB_DEBUG_FREE; 985 break; 986 case 'z': 987 slub_debug |= SLAB_RED_ZONE; 988 break; 989 case 'p': 990 slub_debug |= SLAB_POISON; 991 break; 992 case 'u': 993 slub_debug |= SLAB_STORE_USER; 994 break; 995 case 't': 996 slub_debug |= SLAB_TRACE; 997 break; 998 default: 999 printk(KERN_ERR "slub_debug option '%c' " 1000 "unknown. skipped\n", *str); 1001 } 1002 } 1003 1004check_slabs: 1005 if (*str == ',') 1006 slub_debug_slabs = str + 1; 1007out: 1008 return 1; 1009} 1010 1011__setup("slub_debug", setup_slub_debug); 1012 1013static unsigned long kmem_cache_flags(unsigned long objsize, 1014 unsigned long flags, const char *name, 1015 void (*ctor)(void *)) 1016{ 1017 /* 1018 * Enable debugging if selected on the kernel commandline. 1019 */ 1020 if (slub_debug && (!slub_debug_slabs || 1021 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0)) 1022 flags |= slub_debug; 1023 1024 return flags; 1025} 1026#else 1027static inline void setup_object_debug(struct kmem_cache *s, 1028 struct page *page, void *object) {} 1029 1030static inline int alloc_debug_processing(struct kmem_cache *s, 1031 struct page *page, void *object, void *addr) { return 0; } 1032 1033static inline int free_debug_processing(struct kmem_cache *s, 1034 struct page *page, void *object, void *addr) { return 0; } 1035 1036static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1037 { return 1; } 1038static inline int check_object(struct kmem_cache *s, struct page *page, 1039 void *object, int active) { return 1; } 1040static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1041static inline unsigned long kmem_cache_flags(unsigned long objsize, 1042 unsigned long flags, const char *name, 1043 void (*ctor)(void *)) 1044{ 1045 return flags; 1046} 1047#define slub_debug 0 1048 1049static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1050 { return 0; } 1051static inline void inc_slabs_node(struct kmem_cache *s, int node, 1052 int objects) {} 1053static inline void dec_slabs_node(struct kmem_cache *s, int node, 1054 int objects) {} 1055#endif 1056 1057/* 1058 * Slab allocation and freeing 1059 */ 1060static inline struct page *alloc_slab_page(gfp_t flags, int node, 1061 struct kmem_cache_order_objects oo) 1062{ 1063 int order = oo_order(oo); 1064 1065 if (node == -1) 1066 return alloc_pages(flags, order); 1067 else 1068 return alloc_pages_node(node, flags, order); 1069} 1070 1071static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1072{ 1073 struct page *page; 1074 struct kmem_cache_order_objects oo = s->oo; 1075 1076 flags |= s->allocflags; 1077 1078 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node, 1079 oo); 1080 if (unlikely(!page)) { 1081 oo = s->min; 1082 /* 1083 * Allocation may have failed due to fragmentation. 1084 * Try a lower order alloc if possible 1085 */ 1086 page = alloc_slab_page(flags, node, oo); 1087 if (!page) 1088 return NULL; 1089 1090 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK); 1091 } 1092 page->objects = oo_objects(oo); 1093 mod_zone_page_state(page_zone(page), 1094 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1095 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1096 1 << oo_order(oo)); 1097 1098 return page; 1099} 1100 1101static void setup_object(struct kmem_cache *s, struct page *page, 1102 void *object) 1103{ 1104 setup_object_debug(s, page, object); 1105 if (unlikely(s->ctor)) 1106 s->ctor(object); 1107} 1108 1109static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1110{ 1111 struct page *page; 1112 void *start; 1113 void *last; 1114 void *p; 1115 1116 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1117 1118 page = allocate_slab(s, 1119 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1120 if (!page) 1121 goto out; 1122 1123 inc_slabs_node(s, page_to_nid(page), page->objects); 1124 page->slab = s; 1125 page->flags |= 1 << PG_slab; 1126 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | 1127 SLAB_STORE_USER | SLAB_TRACE)) 1128 __SetPageSlubDebug(page); 1129 1130 start = page_address(page); 1131 1132 if (unlikely(s->flags & SLAB_POISON)) 1133 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1134 1135 last = start; 1136 for_each_object(p, s, start, page->objects) { 1137 setup_object(s, page, last); 1138 set_freepointer(s, last, p); 1139 last = p; 1140 } 1141 setup_object(s, page, last); 1142 set_freepointer(s, last, NULL); 1143 1144 page->freelist = start; 1145 page->inuse = 0; 1146out: 1147 return page; 1148} 1149 1150static void __free_slab(struct kmem_cache *s, struct page *page) 1151{ 1152 int order = compound_order(page); 1153 int pages = 1 << order; 1154 1155 if (unlikely(SLABDEBUG && PageSlubDebug(page))) { 1156 void *p; 1157 1158 slab_pad_check(s, page); 1159 for_each_object(p, s, page_address(page), 1160 page->objects) 1161 check_object(s, page, p, 0); 1162 __ClearPageSlubDebug(page); 1163 } 1164 1165 mod_zone_page_state(page_zone(page), 1166 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1167 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1168 -pages); 1169 1170 __ClearPageSlab(page); 1171 reset_page_mapcount(page); 1172 __free_pages(page, order); 1173} 1174 1175static void rcu_free_slab(struct rcu_head *h) 1176{ 1177 struct page *page; 1178 1179 page = container_of((struct list_head *)h, struct page, lru); 1180 __free_slab(page->slab, page); 1181} 1182 1183static void free_slab(struct kmem_cache *s, struct page *page) 1184{ 1185 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1186 /* 1187 * RCU free overloads the RCU head over the LRU 1188 */ 1189 struct rcu_head *head = (void *)&page->lru; 1190 1191 call_rcu(head, rcu_free_slab); 1192 } else 1193 __free_slab(s, page); 1194} 1195 1196static void discard_slab(struct kmem_cache *s, struct page *page) 1197{ 1198 dec_slabs_node(s, page_to_nid(page), page->objects); 1199 free_slab(s, page); 1200} 1201 1202/* 1203 * Per slab locking using the pagelock 1204 */ 1205static __always_inline void slab_lock(struct page *page) 1206{ 1207 bit_spin_lock(PG_locked, &page->flags); 1208} 1209 1210static __always_inline void slab_unlock(struct page *page) 1211{ 1212 __bit_spin_unlock(PG_locked, &page->flags); 1213} 1214 1215static __always_inline int slab_trylock(struct page *page) 1216{ 1217 int rc = 1; 1218 1219 rc = bit_spin_trylock(PG_locked, &page->flags); 1220 return rc; 1221} 1222 1223/* 1224 * Management of partially allocated slabs 1225 */ 1226static void add_partial(struct kmem_cache_node *n, 1227 struct page *page, int tail) 1228{ 1229 spin_lock(&n->list_lock); 1230 n->nr_partial++; 1231 if (tail) 1232 list_add_tail(&page->lru, &n->partial); 1233 else 1234 list_add(&page->lru, &n->partial); 1235 spin_unlock(&n->list_lock); 1236} 1237 1238static void remove_partial(struct kmem_cache *s, struct page *page) 1239{ 1240 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1241 1242 spin_lock(&n->list_lock); 1243 list_del(&page->lru); 1244 n->nr_partial--; 1245 spin_unlock(&n->list_lock); 1246} 1247 1248/* 1249 * Lock slab and remove from the partial list. 1250 * 1251 * Must hold list_lock. 1252 */ 1253static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1254 struct page *page) 1255{ 1256 if (slab_trylock(page)) { 1257 list_del(&page->lru); 1258 n->nr_partial--; 1259 __SetPageSlubFrozen(page); 1260 return 1; 1261 } 1262 return 0; 1263} 1264 1265/* 1266 * Try to allocate a partial slab from a specific node. 1267 */ 1268static struct page *get_partial_node(struct kmem_cache_node *n) 1269{ 1270 struct page *page; 1271 1272 /* 1273 * Racy check. If we mistakenly see no partial slabs then we 1274 * just allocate an empty slab. If we mistakenly try to get a 1275 * partial slab and there is none available then get_partials() 1276 * will return NULL. 1277 */ 1278 if (!n || !n->nr_partial) 1279 return NULL; 1280 1281 spin_lock(&n->list_lock); 1282 list_for_each_entry(page, &n->partial, lru) 1283 if (lock_and_freeze_slab(n, page)) 1284 goto out; 1285 page = NULL; 1286out: 1287 spin_unlock(&n->list_lock); 1288 return page; 1289} 1290 1291/* 1292 * Get a page from somewhere. Search in increasing NUMA distances. 1293 */ 1294static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1295{ 1296#ifdef CONFIG_NUMA 1297 struct zonelist *zonelist; 1298 struct zoneref *z; 1299 struct zone *zone; 1300 enum zone_type high_zoneidx = gfp_zone(flags); 1301 struct page *page; 1302 1303 /* 1304 * The defrag ratio allows a configuration of the tradeoffs between 1305 * inter node defragmentation and node local allocations. A lower 1306 * defrag_ratio increases the tendency to do local allocations 1307 * instead of attempting to obtain partial slabs from other nodes. 1308 * 1309 * If the defrag_ratio is set to 0 then kmalloc() always 1310 * returns node local objects. If the ratio is higher then kmalloc() 1311 * may return off node objects because partial slabs are obtained 1312 * from other nodes and filled up. 1313 * 1314 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1315 * defrag_ratio = 1000) then every (well almost) allocation will 1316 * first attempt to defrag slab caches on other nodes. This means 1317 * scanning over all nodes to look for partial slabs which may be 1318 * expensive if we do it every time we are trying to find a slab 1319 * with available objects. 1320 */ 1321 if (!s->remote_node_defrag_ratio || 1322 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1323 return NULL; 1324 1325 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1326 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1327 struct kmem_cache_node *n; 1328 1329 n = get_node(s, zone_to_nid(zone)); 1330 1331 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1332 n->nr_partial > MIN_PARTIAL) { 1333 page = get_partial_node(n); 1334 if (page) 1335 return page; 1336 } 1337 } 1338#endif 1339 return NULL; 1340} 1341 1342/* 1343 * Get a partial page, lock it and return it. 1344 */ 1345static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1346{ 1347 struct page *page; 1348 int searchnode = (node == -1) ? numa_node_id() : node; 1349 1350 page = get_partial_node(get_node(s, searchnode)); 1351 if (page || (flags & __GFP_THISNODE)) 1352 return page; 1353 1354 return get_any_partial(s, flags); 1355} 1356 1357/* 1358 * Move a page back to the lists. 1359 * 1360 * Must be called with the slab lock held. 1361 * 1362 * On exit the slab lock will have been dropped. 1363 */ 1364static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1365{ 1366 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1367 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id()); 1368 1369 __ClearPageSlubFrozen(page); 1370 if (page->inuse) { 1371 1372 if (page->freelist) { 1373 add_partial(n, page, tail); 1374 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1375 } else { 1376 stat(c, DEACTIVATE_FULL); 1377 if (SLABDEBUG && PageSlubDebug(page) && 1378 (s->flags & SLAB_STORE_USER)) 1379 add_full(n, page); 1380 } 1381 slab_unlock(page); 1382 } else { 1383 stat(c, DEACTIVATE_EMPTY); 1384 if (n->nr_partial < MIN_PARTIAL) { 1385 /* 1386 * Adding an empty slab to the partial slabs in order 1387 * to avoid page allocator overhead. This slab needs 1388 * to come after the other slabs with objects in 1389 * so that the others get filled first. That way the 1390 * size of the partial list stays small. 1391 * 1392 * kmem_cache_shrink can reclaim any empty slabs from 1393 * the partial list. 1394 */ 1395 add_partial(n, page, 1); 1396 slab_unlock(page); 1397 } else { 1398 slab_unlock(page); 1399 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB); 1400 discard_slab(s, page); 1401 } 1402 } 1403} 1404 1405/* 1406 * Remove the cpu slab 1407 */ 1408static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1409{ 1410 struct page *page = c->page; 1411 int tail = 1; 1412 1413 if (page->freelist) 1414 stat(c, DEACTIVATE_REMOTE_FREES); 1415 /* 1416 * Merge cpu freelist into slab freelist. Typically we get here 1417 * because both freelists are empty. So this is unlikely 1418 * to occur. 1419 */ 1420 while (unlikely(c->freelist)) { 1421 void **object; 1422 1423 tail = 0; /* Hot objects. Put the slab first */ 1424 1425 /* Retrieve object from cpu_freelist */ 1426 object = c->freelist; 1427 c->freelist = c->freelist[c->offset]; 1428 1429 /* And put onto the regular freelist */ 1430 object[c->offset] = page->freelist; 1431 page->freelist = object; 1432 page->inuse--; 1433 } 1434 c->page = NULL; 1435 unfreeze_slab(s, page, tail); 1436} 1437 1438static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1439{ 1440 stat(c, CPUSLAB_FLUSH); 1441 slab_lock(c->page); 1442 deactivate_slab(s, c); 1443} 1444 1445/* 1446 * Flush cpu slab. 1447 * 1448 * Called from IPI handler with interrupts disabled. 1449 */ 1450static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1451{ 1452 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1453 1454 if (likely(c && c->page)) 1455 flush_slab(s, c); 1456} 1457 1458static void flush_cpu_slab(void *d) 1459{ 1460 struct kmem_cache *s = d; 1461 1462 __flush_cpu_slab(s, smp_processor_id()); 1463} 1464 1465static void flush_all(struct kmem_cache *s) 1466{ 1467 on_each_cpu(flush_cpu_slab, s, 1); 1468} 1469 1470/* 1471 * Check if the objects in a per cpu structure fit numa 1472 * locality expectations. 1473 */ 1474static inline int node_match(struct kmem_cache_cpu *c, int node) 1475{ 1476#ifdef CONFIG_NUMA 1477 if (node != -1 && c->node != node) 1478 return 0; 1479#endif 1480 return 1; 1481} 1482 1483/* 1484 * Slow path. The lockless freelist is empty or we need to perform 1485 * debugging duties. 1486 * 1487 * Interrupts are disabled. 1488 * 1489 * Processing is still very fast if new objects have been freed to the 1490 * regular freelist. In that case we simply take over the regular freelist 1491 * as the lockless freelist and zap the regular freelist. 1492 * 1493 * If that is not working then we fall back to the partial lists. We take the 1494 * first element of the freelist as the object to allocate now and move the 1495 * rest of the freelist to the lockless freelist. 1496 * 1497 * And if we were unable to get a new slab from the partial slab lists then 1498 * we need to allocate a new slab. This is the slowest path since it involves 1499 * a call to the page allocator and the setup of a new slab. 1500 */ 1501static void *__slab_alloc(struct kmem_cache *s, 1502 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c) 1503{ 1504 void **object; 1505 struct page *new; 1506 1507 /* We handle __GFP_ZERO in the caller */ 1508 gfpflags &= ~__GFP_ZERO; 1509 1510 if (!c->page) 1511 goto new_slab; 1512 1513 slab_lock(c->page); 1514 if (unlikely(!node_match(c, node))) 1515 goto another_slab; 1516 1517 stat(c, ALLOC_REFILL); 1518 1519load_freelist: 1520 object = c->page->freelist; 1521 if (unlikely(!object)) 1522 goto another_slab; 1523 if (unlikely(SLABDEBUG && PageSlubDebug(c->page))) 1524 goto debug; 1525 1526 c->freelist = object[c->offset]; 1527 c->page->inuse = c->page->objects; 1528 c->page->freelist = NULL; 1529 c->node = page_to_nid(c->page); 1530unlock_out: 1531 slab_unlock(c->page); 1532 stat(c, ALLOC_SLOWPATH); 1533 return object; 1534 1535another_slab: 1536 deactivate_slab(s, c); 1537 1538new_slab: 1539 new = get_partial(s, gfpflags, node); 1540 if (new) { 1541 c->page = new; 1542 stat(c, ALLOC_FROM_PARTIAL); 1543 goto load_freelist; 1544 } 1545 1546 if (gfpflags & __GFP_WAIT) 1547 local_irq_enable(); 1548 1549 new = new_slab(s, gfpflags, node); 1550 1551 if (gfpflags & __GFP_WAIT) 1552 local_irq_disable(); 1553 1554 if (new) { 1555 c = get_cpu_slab(s, smp_processor_id()); 1556 stat(c, ALLOC_SLAB); 1557 if (c->page) 1558 flush_slab(s, c); 1559 slab_lock(new); 1560 __SetPageSlubFrozen(new); 1561 c->page = new; 1562 goto load_freelist; 1563 } 1564 return NULL; 1565debug: 1566 if (!alloc_debug_processing(s, c->page, object, addr)) 1567 goto another_slab; 1568 1569 c->page->inuse++; 1570 c->page->freelist = object[c->offset]; 1571 c->node = -1; 1572 goto unlock_out; 1573} 1574 1575/* 1576 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1577 * have the fastpath folded into their functions. So no function call 1578 * overhead for requests that can be satisfied on the fastpath. 1579 * 1580 * The fastpath works by first checking if the lockless freelist can be used. 1581 * If not then __slab_alloc is called for slow processing. 1582 * 1583 * Otherwise we can simply pick the next object from the lockless free list. 1584 */ 1585static __always_inline void *slab_alloc(struct kmem_cache *s, 1586 gfp_t gfpflags, int node, void *addr) 1587{ 1588 void **object; 1589 struct kmem_cache_cpu *c; 1590 unsigned long flags; 1591 unsigned int objsize; 1592 1593 local_irq_save(flags); 1594 c = get_cpu_slab(s, smp_processor_id()); 1595 objsize = c->objsize; 1596 if (unlikely(!c->freelist || !node_match(c, node))) 1597 1598 object = __slab_alloc(s, gfpflags, node, addr, c); 1599 1600 else { 1601 object = c->freelist; 1602 c->freelist = object[c->offset]; 1603 stat(c, ALLOC_FASTPATH); 1604 } 1605 local_irq_restore(flags); 1606 1607 if (unlikely((gfpflags & __GFP_ZERO) && object)) 1608 memset(object, 0, objsize); 1609 1610 return object; 1611} 1612 1613void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1614{ 1615 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); 1616} 1617EXPORT_SYMBOL(kmem_cache_alloc); 1618 1619#ifdef CONFIG_NUMA 1620void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1621{ 1622 return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); 1623} 1624EXPORT_SYMBOL(kmem_cache_alloc_node); 1625#endif 1626 1627/* 1628 * Slow patch handling. This may still be called frequently since objects 1629 * have a longer lifetime than the cpu slabs in most processing loads. 1630 * 1631 * So we still attempt to reduce cache line usage. Just take the slab 1632 * lock and free the item. If there is no additional partial page 1633 * handling required then we can return immediately. 1634 */ 1635static void __slab_free(struct kmem_cache *s, struct page *page, 1636 void *x, void *addr, unsigned int offset) 1637{ 1638 void *prior; 1639 void **object = (void *)x; 1640 struct kmem_cache_cpu *c; 1641 1642 c = get_cpu_slab(s, raw_smp_processor_id()); 1643 stat(c, FREE_SLOWPATH); 1644 slab_lock(page); 1645 1646 if (unlikely(SLABDEBUG && PageSlubDebug(page))) 1647 goto debug; 1648 1649checks_ok: 1650 prior = object[offset] = page->freelist; 1651 page->freelist = object; 1652 page->inuse--; 1653 1654 if (unlikely(PageSlubFrozen(page))) { 1655 stat(c, FREE_FROZEN); 1656 goto out_unlock; 1657 } 1658 1659 if (unlikely(!page->inuse)) 1660 goto slab_empty; 1661 1662 /* 1663 * Objects left in the slab. If it was not on the partial list before 1664 * then add it. 1665 */ 1666 if (unlikely(!prior)) { 1667 add_partial(get_node(s, page_to_nid(page)), page, 1); 1668 stat(c, FREE_ADD_PARTIAL); 1669 } 1670 1671out_unlock: 1672 slab_unlock(page); 1673 return; 1674 1675slab_empty: 1676 if (prior) { 1677 /* 1678 * Slab still on the partial list. 1679 */ 1680 remove_partial(s, page); 1681 stat(c, FREE_REMOVE_PARTIAL); 1682 } 1683 slab_unlock(page); 1684 stat(c, FREE_SLAB); 1685 discard_slab(s, page); 1686 return; 1687 1688debug: 1689 if (!free_debug_processing(s, page, x, addr)) 1690 goto out_unlock; 1691 goto checks_ok; 1692} 1693 1694/* 1695 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1696 * can perform fastpath freeing without additional function calls. 1697 * 1698 * The fastpath is only possible if we are freeing to the current cpu slab 1699 * of this processor. This typically the case if we have just allocated 1700 * the item before. 1701 * 1702 * If fastpath is not possible then fall back to __slab_free where we deal 1703 * with all sorts of special processing. 1704 */ 1705static __always_inline void slab_free(struct kmem_cache *s, 1706 struct page *page, void *x, void *addr) 1707{ 1708 void **object = (void *)x; 1709 struct kmem_cache_cpu *c; 1710 unsigned long flags; 1711 1712 local_irq_save(flags); 1713 c = get_cpu_slab(s, smp_processor_id()); 1714 debug_check_no_locks_freed(object, c->objsize); 1715 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1716 debug_check_no_obj_freed(object, s->objsize); 1717 if (likely(page == c->page && c->node >= 0)) { 1718 object[c->offset] = c->freelist; 1719 c->freelist = object; 1720 stat(c, FREE_FASTPATH); 1721 } else 1722 __slab_free(s, page, x, addr, c->offset); 1723 1724 local_irq_restore(flags); 1725} 1726 1727void kmem_cache_free(struct kmem_cache *s, void *x) 1728{ 1729 struct page *page; 1730 1731 page = virt_to_head_page(x); 1732 1733 slab_free(s, page, x, __builtin_return_address(0)); 1734} 1735EXPORT_SYMBOL(kmem_cache_free); 1736 1737/* Figure out on which slab object the object resides */ 1738static struct page *get_object_page(const void *x) 1739{ 1740 struct page *page = virt_to_head_page(x); 1741 1742 if (!PageSlab(page)) 1743 return NULL; 1744 1745 return page; 1746} 1747 1748/* 1749 * Object placement in a slab is made very easy because we always start at 1750 * offset 0. If we tune the size of the object to the alignment then we can 1751 * get the required alignment by putting one properly sized object after 1752 * another. 1753 * 1754 * Notice that the allocation order determines the sizes of the per cpu 1755 * caches. Each processor has always one slab available for allocations. 1756 * Increasing the allocation order reduces the number of times that slabs 1757 * must be moved on and off the partial lists and is therefore a factor in 1758 * locking overhead. 1759 */ 1760 1761/* 1762 * Mininum / Maximum order of slab pages. This influences locking overhead 1763 * and slab fragmentation. A higher order reduces the number of partial slabs 1764 * and increases the number of allocations possible without having to 1765 * take the list_lock. 1766 */ 1767static int slub_min_order; 1768static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 1769static int slub_min_objects; 1770 1771/* 1772 * Merge control. If this is set then no merging of slab caches will occur. 1773 * (Could be removed. This was introduced to pacify the merge skeptics.) 1774 */ 1775static int slub_nomerge; 1776 1777/* 1778 * Calculate the order of allocation given an slab object size. 1779 * 1780 * The order of allocation has significant impact on performance and other 1781 * system components. Generally order 0 allocations should be preferred since 1782 * order 0 does not cause fragmentation in the page allocator. Larger objects 1783 * be problematic to put into order 0 slabs because there may be too much 1784 * unused space left. We go to a higher order if more than 1/16th of the slab 1785 * would be wasted. 1786 * 1787 * In order to reach satisfactory performance we must ensure that a minimum 1788 * number of objects is in one slab. Otherwise we may generate too much 1789 * activity on the partial lists which requires taking the list_lock. This is 1790 * less a concern for large slabs though which are rarely used. 1791 * 1792 * slub_max_order specifies the order where we begin to stop considering the 1793 * number of objects in a slab as critical. If we reach slub_max_order then 1794 * we try to keep the page order as low as possible. So we accept more waste 1795 * of space in favor of a small page order. 1796 * 1797 * Higher order allocations also allow the placement of more objects in a 1798 * slab and thereby reduce object handling overhead. If the user has 1799 * requested a higher mininum order then we start with that one instead of 1800 * the smallest order which will fit the object. 1801 */ 1802static inline int slab_order(int size, int min_objects, 1803 int max_order, int fract_leftover) 1804{ 1805 int order; 1806 int rem; 1807 int min_order = slub_min_order; 1808 1809 if ((PAGE_SIZE << min_order) / size > 65535) 1810 return get_order(size * 65535) - 1; 1811 1812 for (order = max(min_order, 1813 fls(min_objects * size - 1) - PAGE_SHIFT); 1814 order <= max_order; order++) { 1815 1816 unsigned long slab_size = PAGE_SIZE << order; 1817 1818 if (slab_size < min_objects * size) 1819 continue; 1820 1821 rem = slab_size % size; 1822 1823 if (rem <= slab_size / fract_leftover) 1824 break; 1825 1826 } 1827 1828 return order; 1829} 1830 1831static inline int calculate_order(int size) 1832{ 1833 int order; 1834 int min_objects; 1835 int fraction; 1836 1837 /* 1838 * Attempt to find best configuration for a slab. This 1839 * works by first attempting to generate a layout with 1840 * the best configuration and backing off gradually. 1841 * 1842 * First we reduce the acceptable waste in a slab. Then 1843 * we reduce the minimum objects required in a slab. 1844 */ 1845 min_objects = slub_min_objects; 1846 if (!min_objects) 1847 min_objects = 4 * (fls(nr_cpu_ids) + 1); 1848 while (min_objects > 1) { 1849 fraction = 16; 1850 while (fraction >= 4) { 1851 order = slab_order(size, min_objects, 1852 slub_max_order, fraction); 1853 if (order <= slub_max_order) 1854 return order; 1855 fraction /= 2; 1856 } 1857 min_objects /= 2; 1858 } 1859 1860 /* 1861 * We were unable to place multiple objects in a slab. Now 1862 * lets see if we can place a single object there. 1863 */ 1864 order = slab_order(size, 1, slub_max_order, 1); 1865 if (order <= slub_max_order) 1866 return order; 1867 1868 /* 1869 * Doh this slab cannot be placed using slub_max_order. 1870 */ 1871 order = slab_order(size, 1, MAX_ORDER, 1); 1872 if (order <= MAX_ORDER) 1873 return order; 1874 return -ENOSYS; 1875} 1876 1877/* 1878 * Figure out what the alignment of the objects will be. 1879 */ 1880static unsigned long calculate_alignment(unsigned long flags, 1881 unsigned long align, unsigned long size) 1882{ 1883 /* 1884 * If the user wants hardware cache aligned objects then follow that 1885 * suggestion if the object is sufficiently large. 1886 * 1887 * The hardware cache alignment cannot override the specified 1888 * alignment though. If that is greater then use it. 1889 */ 1890 if (flags & SLAB_HWCACHE_ALIGN) { 1891 unsigned long ralign = cache_line_size(); 1892 while (size <= ralign / 2) 1893 ralign /= 2; 1894 align = max(align, ralign); 1895 } 1896 1897 if (align < ARCH_SLAB_MINALIGN) 1898 align = ARCH_SLAB_MINALIGN; 1899 1900 return ALIGN(align, sizeof(void *)); 1901} 1902 1903static void init_kmem_cache_cpu(struct kmem_cache *s, 1904 struct kmem_cache_cpu *c) 1905{ 1906 c->page = NULL; 1907 c->freelist = NULL; 1908 c->node = 0; 1909 c->offset = s->offset / sizeof(void *); 1910 c->objsize = s->objsize; 1911#ifdef CONFIG_SLUB_STATS 1912 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); 1913#endif 1914} 1915 1916static void init_kmem_cache_node(struct kmem_cache_node *n) 1917{ 1918 n->nr_partial = 0; 1919 spin_lock_init(&n->list_lock); 1920 INIT_LIST_HEAD(&n->partial); 1921#ifdef CONFIG_SLUB_DEBUG 1922 atomic_long_set(&n->nr_slabs, 0); 1923 INIT_LIST_HEAD(&n->full); 1924#endif 1925} 1926 1927#ifdef CONFIG_SMP 1928/* 1929 * Per cpu array for per cpu structures. 1930 * 1931 * The per cpu array places all kmem_cache_cpu structures from one processor 1932 * close together meaning that it becomes possible that multiple per cpu 1933 * structures are contained in one cacheline. This may be particularly 1934 * beneficial for the kmalloc caches. 1935 * 1936 * A desktop system typically has around 60-80 slabs. With 100 here we are 1937 * likely able to get per cpu structures for all caches from the array defined 1938 * here. We must be able to cover all kmalloc caches during bootstrap. 1939 * 1940 * If the per cpu array is exhausted then fall back to kmalloc 1941 * of individual cachelines. No sharing is possible then. 1942 */ 1943#define NR_KMEM_CACHE_CPU 100 1944 1945static DEFINE_PER_CPU(struct kmem_cache_cpu, 1946 kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; 1947 1948static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); 1949static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; 1950 1951static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, 1952 int cpu, gfp_t flags) 1953{ 1954 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); 1955 1956 if (c) 1957 per_cpu(kmem_cache_cpu_free, cpu) = 1958 (void *)c->freelist; 1959 else { 1960 /* Table overflow: So allocate ourselves */ 1961 c = kmalloc_node( 1962 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), 1963 flags, cpu_to_node(cpu)); 1964 if (!c) 1965 return NULL; 1966 } 1967 1968 init_kmem_cache_cpu(s, c); 1969 return c; 1970} 1971 1972static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) 1973{ 1974 if (c < per_cpu(kmem_cache_cpu, cpu) || 1975 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { 1976 kfree(c); 1977 return; 1978 } 1979 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); 1980 per_cpu(kmem_cache_cpu_free, cpu) = c; 1981} 1982 1983static void free_kmem_cache_cpus(struct kmem_cache *s) 1984{ 1985 int cpu; 1986 1987 for_each_online_cpu(cpu) { 1988 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1989 1990 if (c) { 1991 s->cpu_slab[cpu] = NULL; 1992 free_kmem_cache_cpu(c, cpu); 1993 } 1994 } 1995} 1996 1997static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 1998{ 1999 int cpu; 2000 2001 for_each_online_cpu(cpu) { 2002 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 2003 2004 if (c) 2005 continue; 2006 2007 c = alloc_kmem_cache_cpu(s, cpu, flags); 2008 if (!c) { 2009 free_kmem_cache_cpus(s); 2010 return 0; 2011 } 2012 s->cpu_slab[cpu] = c; 2013 } 2014 return 1; 2015} 2016 2017/* 2018 * Initialize the per cpu array. 2019 */ 2020static void init_alloc_cpu_cpu(int cpu) 2021{ 2022 int i; 2023 2024 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) 2025 return; 2026 2027 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) 2028 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); 2029 2030 cpu_set(cpu, kmem_cach_cpu_free_init_once); 2031} 2032 2033static void __init init_alloc_cpu(void) 2034{ 2035 int cpu; 2036 2037 for_each_online_cpu(cpu) 2038 init_alloc_cpu_cpu(cpu); 2039 } 2040 2041#else 2042static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} 2043static inline void init_alloc_cpu(void) {} 2044 2045static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2046{ 2047 init_kmem_cache_cpu(s, &s->cpu_slab); 2048 return 1; 2049} 2050#endif 2051 2052#ifdef CONFIG_NUMA 2053/* 2054 * No kmalloc_node yet so do it by hand. We know that this is the first 2055 * slab on the node for this slabcache. There are no concurrent accesses 2056 * possible. 2057 * 2058 * Note that this function only works on the kmalloc_node_cache 2059 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2060 * memory on a fresh node that has no slab structures yet. 2061 */ 2062static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags, 2063 int node) 2064{ 2065 struct page *page; 2066 struct kmem_cache_node *n; 2067 unsigned long flags; 2068 2069 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2070 2071 page = new_slab(kmalloc_caches, gfpflags, node); 2072 2073 BUG_ON(!page); 2074 if (page_to_nid(page) != node) { 2075 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2076 "node %d\n", node); 2077 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2078 "in order to be able to continue\n"); 2079 } 2080 2081 n = page->freelist; 2082 BUG_ON(!n); 2083 page->freelist = get_freepointer(kmalloc_caches, n); 2084 page->inuse++; 2085 kmalloc_caches->node[node] = n; 2086#ifdef CONFIG_SLUB_DEBUG 2087 init_object(kmalloc_caches, n, 1); 2088 init_tracking(kmalloc_caches, n); 2089#endif 2090 init_kmem_cache_node(n); 2091 inc_slabs_node(kmalloc_caches, node, page->objects); 2092 2093 /* 2094 * lockdep requires consistent irq usage for each lock 2095 * so even though there cannot be a race this early in 2096 * the boot sequence, we still disable irqs. 2097 */ 2098 local_irq_save(flags); 2099 add_partial(n, page, 0); 2100 local_irq_restore(flags); 2101 return n; 2102} 2103 2104static void free_kmem_cache_nodes(struct kmem_cache *s) 2105{ 2106 int node; 2107 2108 for_each_node_state(node, N_NORMAL_MEMORY) { 2109 struct kmem_cache_node *n = s->node[node]; 2110 if (n && n != &s->local_node) 2111 kmem_cache_free(kmalloc_caches, n); 2112 s->node[node] = NULL; 2113 } 2114} 2115 2116static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2117{ 2118 int node; 2119 int local_node; 2120 2121 if (slab_state >= UP) 2122 local_node = page_to_nid(virt_to_page(s)); 2123 else 2124 local_node = 0; 2125 2126 for_each_node_state(node, N_NORMAL_MEMORY) { 2127 struct kmem_cache_node *n; 2128 2129 if (local_node == node) 2130 n = &s->local_node; 2131 else { 2132 if (slab_state == DOWN) { 2133 n = early_kmem_cache_node_alloc(gfpflags, 2134 node); 2135 continue; 2136 } 2137 n = kmem_cache_alloc_node(kmalloc_caches, 2138 gfpflags, node); 2139 2140 if (!n) { 2141 free_kmem_cache_nodes(s); 2142 return 0; 2143 } 2144 2145 } 2146 s->node[node] = n; 2147 init_kmem_cache_node(n); 2148 } 2149 return 1; 2150} 2151#else 2152static void free_kmem_cache_nodes(struct kmem_cache *s) 2153{ 2154} 2155 2156static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2157{ 2158 init_kmem_cache_node(&s->local_node); 2159 return 1; 2160} 2161#endif 2162 2163/* 2164 * calculate_sizes() determines the order and the distribution of data within 2165 * a slab object. 2166 */ 2167static int calculate_sizes(struct kmem_cache *s, int forced_order) 2168{ 2169 unsigned long flags = s->flags; 2170 unsigned long size = s->objsize; 2171 unsigned long align = s->align; 2172 int order; 2173 2174 /* 2175 * Round up object size to the next word boundary. We can only 2176 * place the free pointer at word boundaries and this determines 2177 * the possible location of the free pointer. 2178 */ 2179 size = ALIGN(size, sizeof(void *)); 2180 2181#ifdef CONFIG_SLUB_DEBUG 2182 /* 2183 * Determine if we can poison the object itself. If the user of 2184 * the slab may touch the object after free or before allocation 2185 * then we should never poison the object itself. 2186 */ 2187 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2188 !s->ctor) 2189 s->flags |= __OBJECT_POISON; 2190 else 2191 s->flags &= ~__OBJECT_POISON; 2192 2193 2194 /* 2195 * If we are Redzoning then check if there is some space between the 2196 * end of the object and the free pointer. If not then add an 2197 * additional word to have some bytes to store Redzone information. 2198 */ 2199 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2200 size += sizeof(void *); 2201#endif 2202 2203 /* 2204 * With that we have determined the number of bytes in actual use 2205 * by the object. This is the potential offset to the free pointer. 2206 */ 2207 s->inuse = size; 2208 2209 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2210 s->ctor)) { 2211 /* 2212 * Relocate free pointer after the object if it is not 2213 * permitted to overwrite the first word of the object on 2214 * kmem_cache_free. 2215 * 2216 * This is the case if we do RCU, have a constructor or 2217 * destructor or are poisoning the objects. 2218 */ 2219 s->offset = size; 2220 size += sizeof(void *); 2221 } 2222 2223#ifdef CONFIG_SLUB_DEBUG 2224 if (flags & SLAB_STORE_USER) 2225 /* 2226 * Need to store information about allocs and frees after 2227 * the object. 2228 */ 2229 size += 2 * sizeof(struct track); 2230 2231 if (flags & SLAB_RED_ZONE) 2232 /* 2233 * Add some empty padding so that we can catch 2234 * overwrites from earlier objects rather than let 2235 * tracking information or the free pointer be 2236 * corrupted if an user writes before the start 2237 * of the object. 2238 */ 2239 size += sizeof(void *); 2240#endif 2241 2242 /* 2243 * Determine the alignment based on various parameters that the 2244 * user specified and the dynamic determination of cache line size 2245 * on bootup. 2246 */ 2247 align = calculate_alignment(flags, align, s->objsize); 2248 2249 /* 2250 * SLUB stores one object immediately after another beginning from 2251 * offset 0. In order to align the objects we have to simply size 2252 * each object to conform to the alignment. 2253 */ 2254 size = ALIGN(size, align); 2255 s->size = size; 2256 if (forced_order >= 0) 2257 order = forced_order; 2258 else 2259 order = calculate_order(size); 2260 2261 if (order < 0) 2262 return 0; 2263 2264 s->allocflags = 0; 2265 if (order) 2266 s->allocflags |= __GFP_COMP; 2267 2268 if (s->flags & SLAB_CACHE_DMA) 2269 s->allocflags |= SLUB_DMA; 2270 2271 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2272 s->allocflags |= __GFP_RECLAIMABLE; 2273 2274 /* 2275 * Determine the number of objects per slab 2276 */ 2277 s->oo = oo_make(order, size); 2278 s->min = oo_make(get_order(size), size); 2279 if (oo_objects(s->oo) > oo_objects(s->max)) 2280 s->max = s->oo; 2281 2282 return !!oo_objects(s->oo); 2283 2284} 2285 2286static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2287 const char *name, size_t size, 2288 size_t align, unsigned long flags, 2289 void (*ctor)(void *)) 2290{ 2291 memset(s, 0, kmem_size); 2292 s->name = name; 2293 s->ctor = ctor; 2294 s->objsize = size; 2295 s->align = align; 2296 s->flags = kmem_cache_flags(size, flags, name, ctor); 2297 2298 if (!calculate_sizes(s, -1)) 2299 goto error; 2300 2301 s->refcount = 1; 2302#ifdef CONFIG_NUMA 2303 s->remote_node_defrag_ratio = 100; 2304#endif 2305 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2306 goto error; 2307 2308 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2309 return 1; 2310 free_kmem_cache_nodes(s); 2311error: 2312 if (flags & SLAB_PANIC) 2313 panic("Cannot create slab %s size=%lu realsize=%u " 2314 "order=%u offset=%u flags=%lx\n", 2315 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2316 s->offset, flags); 2317 return 0; 2318} 2319 2320/* 2321 * Check if a given pointer is valid 2322 */ 2323int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2324{ 2325 struct page *page; 2326 2327 page = get_object_page(object); 2328 2329 if (!page || s != page->slab) 2330 /* No slab or wrong slab */ 2331 return 0; 2332 2333 if (!check_valid_pointer(s, page, object)) 2334 return 0; 2335 2336 /* 2337 * We could also check if the object is on the slabs freelist. 2338 * But this would be too expensive and it seems that the main 2339 * purpose of kmem_ptr_valid() is to check if the object belongs 2340 * to a certain slab. 2341 */ 2342 return 1; 2343} 2344EXPORT_SYMBOL(kmem_ptr_validate); 2345 2346/* 2347 * Determine the size of a slab object 2348 */ 2349unsigned int kmem_cache_size(struct kmem_cache *s) 2350{ 2351 return s->objsize; 2352} 2353EXPORT_SYMBOL(kmem_cache_size); 2354 2355const char *kmem_cache_name(struct kmem_cache *s) 2356{ 2357 return s->name; 2358} 2359EXPORT_SYMBOL(kmem_cache_name); 2360 2361static void list_slab_objects(struct kmem_cache *s, struct page *page, 2362 const char *text) 2363{ 2364#ifdef CONFIG_SLUB_DEBUG 2365 void *addr = page_address(page); 2366 void *p; 2367 DECLARE_BITMAP(map, page->objects); 2368 2369 bitmap_zero(map, page->objects); 2370 slab_err(s, page, "%s", text); 2371 slab_lock(page); 2372 for_each_free_object(p, s, page->freelist) 2373 set_bit(slab_index(p, s, addr), map); 2374 2375 for_each_object(p, s, addr, page->objects) { 2376 2377 if (!test_bit(slab_index(p, s, addr), map)) { 2378 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2379 p, p - addr); 2380 print_tracking(s, p); 2381 } 2382 } 2383 slab_unlock(page); 2384#endif 2385} 2386 2387/* 2388 * Attempt to free all partial slabs on a node. 2389 */ 2390static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2391{ 2392 unsigned long flags; 2393 struct page *page, *h; 2394 2395 spin_lock_irqsave(&n->list_lock, flags); 2396 list_for_each_entry_safe(page, h, &n->partial, lru) { 2397 if (!page->inuse) { 2398 list_del(&page->lru); 2399 discard_slab(s, page); 2400 n->nr_partial--; 2401 } else { 2402 list_slab_objects(s, page, 2403 "Objects remaining on kmem_cache_close()"); 2404 } 2405 } 2406 spin_unlock_irqrestore(&n->list_lock, flags); 2407} 2408 2409/* 2410 * Release all resources used by a slab cache. 2411 */ 2412static inline int kmem_cache_close(struct kmem_cache *s) 2413{ 2414 int node; 2415 2416 flush_all(s); 2417 2418 /* Attempt to free all objects */ 2419 free_kmem_cache_cpus(s); 2420 for_each_node_state(node, N_NORMAL_MEMORY) { 2421 struct kmem_cache_node *n = get_node(s, node); 2422 2423 free_partial(s, n); 2424 if (n->nr_partial || slabs_node(s, node)) 2425 return 1; 2426 } 2427 free_kmem_cache_nodes(s); 2428 return 0; 2429} 2430 2431/* 2432 * Close a cache and release the kmem_cache structure 2433 * (must be used for caches created using kmem_cache_create) 2434 */ 2435void kmem_cache_destroy(struct kmem_cache *s) 2436{ 2437 down_write(&slub_lock); 2438 s->refcount--; 2439 if (!s->refcount) { 2440 list_del(&s->list); 2441 up_write(&slub_lock); 2442 if (kmem_cache_close(s)) { 2443 printk(KERN_ERR "SLUB %s: %s called for cache that " 2444 "still has objects.\n", s->name, __func__); 2445 dump_stack(); 2446 } 2447 sysfs_slab_remove(s); 2448 } else 2449 up_write(&slub_lock); 2450} 2451EXPORT_SYMBOL(kmem_cache_destroy); 2452 2453/******************************************************************** 2454 * Kmalloc subsystem 2455 *******************************************************************/ 2456 2457struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; 2458EXPORT_SYMBOL(kmalloc_caches); 2459 2460static int __init setup_slub_min_order(char *str) 2461{ 2462 get_option(&str, &slub_min_order); 2463 2464 return 1; 2465} 2466 2467__setup("slub_min_order=", setup_slub_min_order); 2468 2469static int __init setup_slub_max_order(char *str) 2470{ 2471 get_option(&str, &slub_max_order); 2472 2473 return 1; 2474} 2475 2476__setup("slub_max_order=", setup_slub_max_order); 2477 2478static int __init setup_slub_min_objects(char *str) 2479{ 2480 get_option(&str, &slub_min_objects); 2481 2482 return 1; 2483} 2484 2485__setup("slub_min_objects=", setup_slub_min_objects); 2486 2487static int __init setup_slub_nomerge(char *str) 2488{ 2489 slub_nomerge = 1; 2490 return 1; 2491} 2492 2493__setup("slub_nomerge", setup_slub_nomerge); 2494 2495static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2496 const char *name, int size, gfp_t gfp_flags) 2497{ 2498 unsigned int flags = 0; 2499 2500 if (gfp_flags & SLUB_DMA) 2501 flags = SLAB_CACHE_DMA; 2502 2503 down_write(&slub_lock); 2504 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2505 flags, NULL)) 2506 goto panic; 2507 2508 list_add(&s->list, &slab_caches); 2509 up_write(&slub_lock); 2510 if (sysfs_slab_add(s)) 2511 goto panic; 2512 return s; 2513 2514panic: 2515 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2516} 2517 2518#ifdef CONFIG_ZONE_DMA 2519static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; 2520 2521static void sysfs_add_func(struct work_struct *w) 2522{ 2523 struct kmem_cache *s; 2524 2525 down_write(&slub_lock); 2526 list_for_each_entry(s, &slab_caches, list) { 2527 if (s->flags & __SYSFS_ADD_DEFERRED) { 2528 s->flags &= ~__SYSFS_ADD_DEFERRED; 2529 sysfs_slab_add(s); 2530 } 2531 } 2532 up_write(&slub_lock); 2533} 2534 2535static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2536 2537static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2538{ 2539 struct kmem_cache *s; 2540 char *text; 2541 size_t realsize; 2542 2543 s = kmalloc_caches_dma[index]; 2544 if (s) 2545 return s; 2546 2547 /* Dynamically create dma cache */ 2548 if (flags & __GFP_WAIT) 2549 down_write(&slub_lock); 2550 else { 2551 if (!down_write_trylock(&slub_lock)) 2552 goto out; 2553 } 2554 2555 if (kmalloc_caches_dma[index]) 2556 goto unlock_out; 2557 2558 realsize = kmalloc_caches[index].objsize; 2559 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", 2560 (unsigned int)realsize); 2561 s = kmalloc(kmem_size, flags & ~SLUB_DMA); 2562 2563 if (!s || !text || !kmem_cache_open(s, flags, text, 2564 realsize, ARCH_KMALLOC_MINALIGN, 2565 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { 2566 kfree(s); 2567 kfree(text); 2568 goto unlock_out; 2569 } 2570 2571 list_add(&s->list, &slab_caches); 2572 kmalloc_caches_dma[index] = s; 2573 2574 schedule_work(&sysfs_add_work); 2575 2576unlock_out: 2577 up_write(&slub_lock); 2578out: 2579 return kmalloc_caches_dma[index]; 2580} 2581#endif 2582 2583/* 2584 * Conversion table for small slabs sizes / 8 to the index in the 2585 * kmalloc array. This is necessary for slabs < 192 since we have non power 2586 * of two cache sizes there. The size of larger slabs can be determined using 2587 * fls. 2588 */ 2589static s8 size_index[24] = { 2590 3, /* 8 */ 2591 4, /* 16 */ 2592 5, /* 24 */ 2593 5, /* 32 */ 2594 6, /* 40 */ 2595 6, /* 48 */ 2596 6, /* 56 */ 2597 6, /* 64 */ 2598 1, /* 72 */ 2599 1, /* 80 */ 2600 1, /* 88 */ 2601 1, /* 96 */ 2602 7, /* 104 */ 2603 7, /* 112 */ 2604 7, /* 120 */ 2605 7, /* 128 */ 2606 2, /* 136 */ 2607 2, /* 144 */ 2608 2, /* 152 */ 2609 2, /* 160 */ 2610 2, /* 168 */ 2611 2, /* 176 */ 2612 2, /* 184 */ 2613 2 /* 192 */ 2614}; 2615 2616static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2617{ 2618 int index; 2619 2620 if (size <= 192) { 2621 if (!size) 2622 return ZERO_SIZE_PTR; 2623 2624 index = size_index[(size - 1) / 8]; 2625 } else 2626 index = fls(size - 1); 2627 2628#ifdef CONFIG_ZONE_DMA 2629 if (unlikely((flags & SLUB_DMA))) 2630 return dma_kmalloc_cache(index, flags); 2631 2632#endif 2633 return &kmalloc_caches[index]; 2634} 2635 2636void *__kmalloc(size_t size, gfp_t flags) 2637{ 2638 struct kmem_cache *s; 2639 2640 if (unlikely(size > PAGE_SIZE)) 2641 return kmalloc_large(size, flags); 2642 2643 s = get_slab(size, flags); 2644 2645 if (unlikely(ZERO_OR_NULL_PTR(s))) 2646 return s; 2647 2648 return slab_alloc(s, flags, -1, __builtin_return_address(0)); 2649} 2650EXPORT_SYMBOL(__kmalloc); 2651 2652static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2653{ 2654 struct page *page = alloc_pages_node(node, flags | __GFP_COMP, 2655 get_order(size)); 2656 2657 if (page) 2658 return page_address(page); 2659 else 2660 return NULL; 2661} 2662 2663#ifdef CONFIG_NUMA 2664void *__kmalloc_node(size_t size, gfp_t flags, int node) 2665{ 2666 struct kmem_cache *s; 2667 2668 if (unlikely(size > PAGE_SIZE)) 2669 return kmalloc_large_node(size, flags, node); 2670 2671 s = get_slab(size, flags); 2672 2673 if (unlikely(ZERO_OR_NULL_PTR(s))) 2674 return s; 2675 2676 return slab_alloc(s, flags, node, __builtin_return_address(0)); 2677} 2678EXPORT_SYMBOL(__kmalloc_node); 2679#endif 2680 2681size_t ksize(const void *object) 2682{ 2683 struct page *page; 2684 struct kmem_cache *s; 2685 2686 if (unlikely(object == ZERO_SIZE_PTR)) 2687 return 0; 2688 2689 page = virt_to_head_page(object); 2690 2691 if (unlikely(!PageSlab(page))) { 2692 WARN_ON(!PageCompound(page)); 2693 return PAGE_SIZE << compound_order(page); 2694 } 2695 s = page->slab; 2696 2697#ifdef CONFIG_SLUB_DEBUG 2698 /* 2699 * Debugging requires use of the padding between object 2700 * and whatever may come after it. 2701 */ 2702 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2703 return s->objsize; 2704 2705#endif 2706 /* 2707 * If we have the need to store the freelist pointer 2708 * back there or track user information then we can 2709 * only use the space before that information. 2710 */ 2711 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2712 return s->inuse; 2713 /* 2714 * Else we can use all the padding etc for the allocation 2715 */ 2716 return s->size; 2717} 2718 2719void kfree(const void *x) 2720{ 2721 struct page *page; 2722 void *object = (void *)x; 2723 2724 if (unlikely(ZERO_OR_NULL_PTR(x))) 2725 return; 2726 2727 page = virt_to_head_page(x); 2728 if (unlikely(!PageSlab(page))) { 2729 BUG_ON(!PageCompound(page)); 2730 put_page(page); 2731 return; 2732 } 2733 slab_free(page->slab, page, object, __builtin_return_address(0)); 2734} 2735EXPORT_SYMBOL(kfree); 2736 2737/* 2738 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2739 * the remaining slabs by the number of items in use. The slabs with the 2740 * most items in use come first. New allocations will then fill those up 2741 * and thus they can be removed from the partial lists. 2742 * 2743 * The slabs with the least items are placed last. This results in them 2744 * being allocated from last increasing the chance that the last objects 2745 * are freed in them. 2746 */ 2747int kmem_cache_shrink(struct kmem_cache *s) 2748{ 2749 int node; 2750 int i; 2751 struct kmem_cache_node *n; 2752 struct page *page; 2753 struct page *t; 2754 int objects = oo_objects(s->max); 2755 struct list_head *slabs_by_inuse = 2756 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 2757 unsigned long flags; 2758 2759 if (!slabs_by_inuse) 2760 return -ENOMEM; 2761 2762 flush_all(s); 2763 for_each_node_state(node, N_NORMAL_MEMORY) { 2764 n = get_node(s, node); 2765 2766 if (!n->nr_partial) 2767 continue; 2768 2769 for (i = 0; i < objects; i++) 2770 INIT_LIST_HEAD(slabs_by_inuse + i); 2771 2772 spin_lock_irqsave(&n->list_lock, flags); 2773 2774 /* 2775 * Build lists indexed by the items in use in each slab. 2776 * 2777 * Note that concurrent frees may occur while we hold the 2778 * list_lock. page->inuse here is the upper limit. 2779 */ 2780 list_for_each_entry_safe(page, t, &n->partial, lru) { 2781 if (!page->inuse && slab_trylock(page)) { 2782 /* 2783 * Must hold slab lock here because slab_free 2784 * may have freed the last object and be 2785 * waiting to release the slab. 2786 */ 2787 list_del(&page->lru); 2788 n->nr_partial--; 2789 slab_unlock(page); 2790 discard_slab(s, page); 2791 } else { 2792 list_move(&page->lru, 2793 slabs_by_inuse + page->inuse); 2794 } 2795 } 2796 2797 /* 2798 * Rebuild the partial list with the slabs filled up most 2799 * first and the least used slabs at the end. 2800 */ 2801 for (i = objects - 1; i >= 0; i--) 2802 list_splice(slabs_by_inuse + i, n->partial.prev); 2803 2804 spin_unlock_irqrestore(&n->list_lock, flags); 2805 } 2806 2807 kfree(slabs_by_inuse); 2808 return 0; 2809} 2810EXPORT_SYMBOL(kmem_cache_shrink); 2811 2812#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2813static int slab_mem_going_offline_callback(void *arg) 2814{ 2815 struct kmem_cache *s; 2816 2817 down_read(&slub_lock); 2818 list_for_each_entry(s, &slab_caches, list) 2819 kmem_cache_shrink(s); 2820 up_read(&slub_lock); 2821 2822 return 0; 2823} 2824 2825static void slab_mem_offline_callback(void *arg) 2826{ 2827 struct kmem_cache_node *n; 2828 struct kmem_cache *s; 2829 struct memory_notify *marg = arg; 2830 int offline_node; 2831 2832 offline_node = marg->status_change_nid; 2833 2834 /* 2835 * If the node still has available memory. we need kmem_cache_node 2836 * for it yet. 2837 */ 2838 if (offline_node < 0) 2839 return; 2840 2841 down_read(&slub_lock); 2842 list_for_each_entry(s, &slab_caches, list) { 2843 n = get_node(s, offline_node); 2844 if (n) { 2845 /* 2846 * if n->nr_slabs > 0, slabs still exist on the node 2847 * that is going down. We were unable to free them, 2848 * and offline_pages() function shoudn't call this 2849 * callback. So, we must fail. 2850 */ 2851 BUG_ON(slabs_node(s, offline_node)); 2852 2853 s->node[offline_node] = NULL; 2854 kmem_cache_free(kmalloc_caches, n); 2855 } 2856 } 2857 up_read(&slub_lock); 2858} 2859 2860static int slab_mem_going_online_callback(void *arg) 2861{ 2862 struct kmem_cache_node *n; 2863 struct kmem_cache *s; 2864 struct memory_notify *marg = arg; 2865 int nid = marg->status_change_nid; 2866 int ret = 0; 2867 2868 /* 2869 * If the node's memory is already available, then kmem_cache_node is 2870 * already created. Nothing to do. 2871 */ 2872 if (nid < 0) 2873 return 0; 2874 2875 /* 2876 * We are bringing a node online. No memory is available yet. We must 2877 * allocate a kmem_cache_node structure in order to bring the node 2878 * online. 2879 */ 2880 down_read(&slub_lock); 2881 list_for_each_entry(s, &slab_caches, list) { 2882 /* 2883 * XXX: kmem_cache_alloc_node will fallback to other nodes 2884 * since memory is not yet available from the node that 2885 * is brought up. 2886 */ 2887 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 2888 if (!n) { 2889 ret = -ENOMEM; 2890 goto out; 2891 } 2892 init_kmem_cache_node(n); 2893 s->node[nid] = n; 2894 } 2895out: 2896 up_read(&slub_lock); 2897 return ret; 2898} 2899 2900static int slab_memory_callback(struct notifier_block *self, 2901 unsigned long action, void *arg) 2902{ 2903 int ret = 0; 2904 2905 switch (action) { 2906 case MEM_GOING_ONLINE: 2907 ret = slab_mem_going_online_callback(arg); 2908 break; 2909 case MEM_GOING_OFFLINE: 2910 ret = slab_mem_going_offline_callback(arg); 2911 break; 2912 case MEM_OFFLINE: 2913 case MEM_CANCEL_ONLINE: 2914 slab_mem_offline_callback(arg); 2915 break; 2916 case MEM_ONLINE: 2917 case MEM_CANCEL_OFFLINE: 2918 break; 2919 } 2920 2921 ret = notifier_from_errno(ret); 2922 return ret; 2923} 2924 2925#endif /* CONFIG_MEMORY_HOTPLUG */ 2926 2927/******************************************************************** 2928 * Basic setup of slabs 2929 *******************************************************************/ 2930 2931void __init kmem_cache_init(void) 2932{ 2933 int i; 2934 int caches = 0; 2935 2936 init_alloc_cpu(); 2937 2938#ifdef CONFIG_NUMA 2939 /* 2940 * Must first have the slab cache available for the allocations of the 2941 * struct kmem_cache_node's. There is special bootstrap code in 2942 * kmem_cache_open for slab_state == DOWN. 2943 */ 2944 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 2945 sizeof(struct kmem_cache_node), GFP_KERNEL); 2946 kmalloc_caches[0].refcount = -1; 2947 caches++; 2948 2949 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 2950#endif 2951 2952 /* Able to allocate the per node structures */ 2953 slab_state = PARTIAL; 2954 2955 /* Caches that are not of the two-to-the-power-of size */ 2956 if (KMALLOC_MIN_SIZE <= 64) { 2957 create_kmalloc_cache(&kmalloc_caches[1], 2958 "kmalloc-96", 96, GFP_KERNEL); 2959 caches++; 2960 create_kmalloc_cache(&kmalloc_caches[2], 2961 "kmalloc-192", 192, GFP_KERNEL); 2962 caches++; 2963 } 2964 2965 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { 2966 create_kmalloc_cache(&kmalloc_caches[i], 2967 "kmalloc", 1 << i, GFP_KERNEL); 2968 caches++; 2969 } 2970 2971 2972 /* 2973 * Patch up the size_index table if we have strange large alignment 2974 * requirements for the kmalloc array. This is only the case for 2975 * MIPS it seems. The standard arches will not generate any code here. 2976 * 2977 * Largest permitted alignment is 256 bytes due to the way we 2978 * handle the index determination for the smaller caches. 2979 * 2980 * Make sure that nothing crazy happens if someone starts tinkering 2981 * around with ARCH_KMALLOC_MINALIGN 2982 */ 2983 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 2984 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 2985 2986 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) 2987 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; 2988 2989 if (KMALLOC_MIN_SIZE == 128) { 2990 /* 2991 * The 192 byte sized cache is not used if the alignment 2992 * is 128 byte. Redirect kmalloc to use the 256 byte cache 2993 * instead. 2994 */ 2995 for (i = 128 + 8; i <= 192; i += 8) 2996 size_index[(i - 1) / 8] = 8; 2997 } 2998 2999 slab_state = UP; 3000 3001 /* Provide the correct kmalloc names now that the caches are up */ 3002 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) 3003 kmalloc_caches[i]. name = 3004 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); 3005 3006#ifdef CONFIG_SMP 3007 register_cpu_notifier(&slab_notifier); 3008 kmem_size = offsetof(struct kmem_cache, cpu_slab) + 3009 nr_cpu_ids * sizeof(struct kmem_cache_cpu *); 3010#else 3011 kmem_size = sizeof(struct kmem_cache); 3012#endif 3013 3014 printk(KERN_INFO 3015 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3016 " CPUs=%d, Nodes=%d\n", 3017 caches, cache_line_size(), 3018 slub_min_order, slub_max_order, slub_min_objects, 3019 nr_cpu_ids, nr_node_ids); 3020} 3021 3022/* 3023 * Find a mergeable slab cache 3024 */ 3025static int slab_unmergeable(struct kmem_cache *s) 3026{ 3027 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3028 return 1; 3029 3030 if (s->ctor) 3031 return 1; 3032 3033 /* 3034 * We may have set a slab to be unmergeable during bootstrap. 3035 */ 3036 if (s->refcount < 0) 3037 return 1; 3038 3039 return 0; 3040} 3041 3042static struct kmem_cache *find_mergeable(size_t size, 3043 size_t align, unsigned long flags, const char *name, 3044 void (*ctor)(void *)) 3045{ 3046 struct kmem_cache *s; 3047 3048 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3049 return NULL; 3050 3051 if (ctor) 3052 return NULL; 3053 3054 size = ALIGN(size, sizeof(void *)); 3055 align = calculate_alignment(flags, align, size); 3056 size = ALIGN(size, align); 3057 flags = kmem_cache_flags(size, flags, name, NULL); 3058 3059 list_for_each_entry(s, &slab_caches, list) { 3060 if (slab_unmergeable(s)) 3061 continue; 3062 3063 if (size > s->size) 3064 continue; 3065 3066 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3067 continue; 3068 /* 3069 * Check if alignment is compatible. 3070 * Courtesy of Adrian Drzewiecki 3071 */ 3072 if ((s->size & ~(align - 1)) != s->size) 3073 continue; 3074 3075 if (s->size - size >= sizeof(void *)) 3076 continue; 3077 3078 return s; 3079 } 3080 return NULL; 3081} 3082 3083struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3084 size_t align, unsigned long flags, void (*ctor)(void *)) 3085{ 3086 struct kmem_cache *s; 3087 3088 down_write(&slub_lock); 3089 s = find_mergeable(size, align, flags, name, ctor); 3090 if (s) { 3091 int cpu; 3092 3093 s->refcount++; 3094 /* 3095 * Adjust the object sizes so that we clear 3096 * the complete object on kzalloc. 3097 */ 3098 s->objsize = max(s->objsize, (int)size); 3099 3100 /* 3101 * And then we need to update the object size in the 3102 * per cpu structures 3103 */ 3104 for_each_online_cpu(cpu) 3105 get_cpu_slab(s, cpu)->objsize = s->objsize; 3106 3107 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3108 up_write(&slub_lock); 3109 3110 if (sysfs_slab_alias(s, name)) 3111 goto err; 3112 return s; 3113 } 3114 3115 s = kmalloc(kmem_size, GFP_KERNEL); 3116 if (s) { 3117 if (kmem_cache_open(s, GFP_KERNEL, name, 3118 size, align, flags, ctor)) { 3119 list_add(&s->list, &slab_caches); 3120 up_write(&slub_lock); 3121 if (sysfs_slab_add(s)) 3122 goto err; 3123 return s; 3124 } 3125 kfree(s); 3126 } 3127 up_write(&slub_lock); 3128 3129err: 3130 if (flags & SLAB_PANIC) 3131 panic("Cannot create slabcache %s\n", name); 3132 else 3133 s = NULL; 3134 return s; 3135} 3136EXPORT_SYMBOL(kmem_cache_create); 3137 3138#ifdef CONFIG_SMP 3139/* 3140 * Use the cpu notifier to insure that the cpu slabs are flushed when 3141 * necessary. 3142 */ 3143static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3144 unsigned long action, void *hcpu) 3145{ 3146 long cpu = (long)hcpu; 3147 struct kmem_cache *s; 3148 unsigned long flags; 3149 3150 switch (action) { 3151 case CPU_UP_PREPARE: 3152 case CPU_UP_PREPARE_FROZEN: 3153 init_alloc_cpu_cpu(cpu); 3154 down_read(&slub_lock); 3155 list_for_each_entry(s, &slab_caches, list) 3156 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, 3157 GFP_KERNEL); 3158 up_read(&slub_lock); 3159 break; 3160 3161 case CPU_UP_CANCELED: 3162 case CPU_UP_CANCELED_FROZEN: 3163 case CPU_DEAD: 3164 case CPU_DEAD_FROZEN: 3165 down_read(&slub_lock); 3166 list_for_each_entry(s, &slab_caches, list) { 3167 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3168 3169 local_irq_save(flags); 3170 __flush_cpu_slab(s, cpu); 3171 local_irq_restore(flags); 3172 free_kmem_cache_cpu(c, cpu); 3173 s->cpu_slab[cpu] = NULL; 3174 } 3175 up_read(&slub_lock); 3176 break; 3177 default: 3178 break; 3179 } 3180 return NOTIFY_OK; 3181} 3182 3183static struct notifier_block __cpuinitdata slab_notifier = { 3184 .notifier_call = slab_cpuup_callback 3185}; 3186 3187#endif 3188 3189void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) 3190{ 3191 struct kmem_cache *s; 3192 3193 if (unlikely(size > PAGE_SIZE)) 3194 return kmalloc_large(size, gfpflags); 3195 3196 s = get_slab(size, gfpflags); 3197 3198 if (unlikely(ZERO_OR_NULL_PTR(s))) 3199 return s; 3200 3201 return slab_alloc(s, gfpflags, -1, caller); 3202} 3203 3204void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3205 int node, void *caller) 3206{ 3207 struct kmem_cache *s; 3208 3209 if (unlikely(size > PAGE_SIZE)) 3210 return kmalloc_large_node(size, gfpflags, node); 3211 3212 s = get_slab(size, gfpflags); 3213 3214 if (unlikely(ZERO_OR_NULL_PTR(s))) 3215 return s; 3216 3217 return slab_alloc(s, gfpflags, node, caller); 3218} 3219 3220#ifdef CONFIG_SLUB_DEBUG 3221static unsigned long count_partial(struct kmem_cache_node *n, 3222 int (*get_count)(struct page *)) 3223{ 3224 unsigned long flags; 3225 unsigned long x = 0; 3226 struct page *page; 3227 3228 spin_lock_irqsave(&n->list_lock, flags); 3229 list_for_each_entry(page, &n->partial, lru) 3230 x += get_count(page); 3231 spin_unlock_irqrestore(&n->list_lock, flags); 3232 return x; 3233} 3234 3235static int count_inuse(struct page *page) 3236{ 3237 return page->inuse; 3238} 3239 3240static int count_total(struct page *page) 3241{ 3242 return page->objects; 3243} 3244 3245static int count_free(struct page *page) 3246{ 3247 return page->objects - page->inuse; 3248} 3249 3250static int validate_slab(struct kmem_cache *s, struct page *page, 3251 unsigned long *map) 3252{ 3253 void *p; 3254 void *addr = page_address(page); 3255 3256 if (!check_slab(s, page) || 3257 !on_freelist(s, page, NULL)) 3258 return 0; 3259 3260 /* Now we know that a valid freelist exists */ 3261 bitmap_zero(map, page->objects); 3262 3263 for_each_free_object(p, s, page->freelist) { 3264 set_bit(slab_index(p, s, addr), map); 3265 if (!check_object(s, page, p, 0)) 3266 return 0; 3267 } 3268 3269 for_each_object(p, s, addr, page->objects) 3270 if (!test_bit(slab_index(p, s, addr), map)) 3271 if (!check_object(s, page, p, 1)) 3272 return 0; 3273 return 1; 3274} 3275 3276static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3277 unsigned long *map) 3278{ 3279 if (slab_trylock(page)) { 3280 validate_slab(s, page, map); 3281 slab_unlock(page); 3282 } else 3283 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3284 s->name, page); 3285 3286 if (s->flags & DEBUG_DEFAULT_FLAGS) { 3287 if (!PageSlubDebug(page)) 3288 printk(KERN_ERR "SLUB %s: SlubDebug not set " 3289 "on slab 0x%p\n", s->name, page); 3290 } else { 3291 if (PageSlubDebug(page)) 3292 printk(KERN_ERR "SLUB %s: SlubDebug set on " 3293 "slab 0x%p\n", s->name, page); 3294 } 3295} 3296 3297static int validate_slab_node(struct kmem_cache *s, 3298 struct kmem_cache_node *n, unsigned long *map) 3299{ 3300 unsigned long count = 0; 3301 struct page *page; 3302 unsigned long flags; 3303 3304 spin_lock_irqsave(&n->list_lock, flags); 3305 3306 list_for_each_entry(page, &n->partial, lru) { 3307 validate_slab_slab(s, page, map); 3308 count++; 3309 } 3310 if (count != n->nr_partial) 3311 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3312 "counter=%ld\n", s->name, count, n->nr_partial); 3313 3314 if (!(s->flags & SLAB_STORE_USER)) 3315 goto out; 3316 3317 list_for_each_entry(page, &n->full, lru) { 3318 validate_slab_slab(s, page, map); 3319 count++; 3320 } 3321 if (count != atomic_long_read(&n->nr_slabs)) 3322 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3323 "counter=%ld\n", s->name, count, 3324 atomic_long_read(&n->nr_slabs)); 3325 3326out: 3327 spin_unlock_irqrestore(&n->list_lock, flags); 3328 return count; 3329} 3330 3331static long validate_slab_cache(struct kmem_cache *s) 3332{ 3333 int node; 3334 unsigned long count = 0; 3335 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3336 sizeof(unsigned long), GFP_KERNEL); 3337 3338 if (!map) 3339 return -ENOMEM; 3340 3341 flush_all(s); 3342 for_each_node_state(node, N_NORMAL_MEMORY) { 3343 struct kmem_cache_node *n = get_node(s, node); 3344 3345 count += validate_slab_node(s, n, map); 3346 } 3347 kfree(map); 3348 return count; 3349} 3350 3351#ifdef SLUB_RESILIENCY_TEST 3352static void resiliency_test(void) 3353{ 3354 u8 *p; 3355 3356 printk(KERN_ERR "SLUB resiliency testing\n"); 3357 printk(KERN_ERR "-----------------------\n"); 3358 printk(KERN_ERR "A. Corruption after allocation\n"); 3359 3360 p = kzalloc(16, GFP_KERNEL); 3361 p[16] = 0x12; 3362 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3363 " 0x12->0x%p\n\n", p + 16); 3364 3365 validate_slab_cache(kmalloc_caches + 4); 3366 3367 /* Hmmm... The next two are dangerous */ 3368 p = kzalloc(32, GFP_KERNEL); 3369 p[32 + sizeof(void *)] = 0x34; 3370 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3371 " 0x34 -> -0x%p\n", p); 3372 printk(KERN_ERR 3373 "If allocated object is overwritten then not detectable\n\n"); 3374 3375 validate_slab_cache(kmalloc_caches + 5); 3376 p = kzalloc(64, GFP_KERNEL); 3377 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3378 *p = 0x56; 3379 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3380 p); 3381 printk(KERN_ERR 3382 "If allocated object is overwritten then not detectable\n\n"); 3383 validate_slab_cache(kmalloc_caches + 6); 3384 3385 printk(KERN_ERR "\nB. Corruption after free\n"); 3386 p = kzalloc(128, GFP_KERNEL); 3387 kfree(p); 3388 *p = 0x78; 3389 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3390 validate_slab_cache(kmalloc_caches + 7); 3391 3392 p = kzalloc(256, GFP_KERNEL); 3393 kfree(p); 3394 p[50] = 0x9a; 3395 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3396 p); 3397 validate_slab_cache(kmalloc_caches + 8); 3398 3399 p = kzalloc(512, GFP_KERNEL); 3400 kfree(p); 3401 p[512] = 0xab; 3402 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3403 validate_slab_cache(kmalloc_caches + 9); 3404} 3405#else 3406static void resiliency_test(void) {}; 3407#endif 3408 3409/* 3410 * Generate lists of code addresses where slabcache objects are allocated 3411 * and freed. 3412 */ 3413 3414struct location { 3415 unsigned long count; 3416 void *addr; 3417 long long sum_time; 3418 long min_time; 3419 long max_time; 3420 long min_pid; 3421 long max_pid; 3422 cpumask_t cpus; 3423 nodemask_t nodes; 3424}; 3425 3426struct loc_track { 3427 unsigned long max; 3428 unsigned long count; 3429 struct location *loc; 3430}; 3431 3432static void free_loc_track(struct loc_track *t) 3433{ 3434 if (t->max) 3435 free_pages((unsigned long)t->loc, 3436 get_order(sizeof(struct location) * t->max)); 3437} 3438 3439static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3440{ 3441 struct location *l; 3442 int order; 3443 3444 order = get_order(sizeof(struct location) * max); 3445 3446 l = (void *)__get_free_pages(flags, order); 3447 if (!l) 3448 return 0; 3449 3450 if (t->count) { 3451 memcpy(l, t->loc, sizeof(struct location) * t->count); 3452 free_loc_track(t); 3453 } 3454 t->max = max; 3455 t->loc = l; 3456 return 1; 3457} 3458 3459static int add_location(struct loc_track *t, struct kmem_cache *s, 3460 const struct track *track) 3461{ 3462 long start, end, pos; 3463 struct location *l; 3464 void *caddr; 3465 unsigned long age = jiffies - track->when; 3466 3467 start = -1; 3468 end = t->count; 3469 3470 for ( ; ; ) { 3471 pos = start + (end - start + 1) / 2; 3472 3473 /* 3474 * There is nothing at "end". If we end up there 3475 * we need to add something to before end. 3476 */ 3477 if (pos == end) 3478 break; 3479 3480 caddr = t->loc[pos].addr; 3481 if (track->addr == caddr) { 3482 3483 l = &t->loc[pos]; 3484 l->count++; 3485 if (track->when) { 3486 l->sum_time += age; 3487 if (age < l->min_time) 3488 l->min_time = age; 3489 if (age > l->max_time) 3490 l->max_time = age; 3491 3492 if (track->pid < l->min_pid) 3493 l->min_pid = track->pid; 3494 if (track->pid > l->max_pid) 3495 l->max_pid = track->pid; 3496 3497 cpu_set(track->cpu, l->cpus); 3498 } 3499 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3500 return 1; 3501 } 3502 3503 if (track->addr < caddr) 3504 end = pos; 3505 else 3506 start = pos; 3507 } 3508 3509 /* 3510 * Not found. Insert new tracking element. 3511 */ 3512 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3513 return 0; 3514 3515 l = t->loc + pos; 3516 if (pos < t->count) 3517 memmove(l + 1, l, 3518 (t->count - pos) * sizeof(struct location)); 3519 t->count++; 3520 l->count = 1; 3521 l->addr = track->addr; 3522 l->sum_time = age; 3523 l->min_time = age; 3524 l->max_time = age; 3525 l->min_pid = track->pid; 3526 l->max_pid = track->pid; 3527 cpus_clear(l->cpus); 3528 cpu_set(track->cpu, l->cpus); 3529 nodes_clear(l->nodes); 3530 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3531 return 1; 3532} 3533 3534static void process_slab(struct loc_track *t, struct kmem_cache *s, 3535 struct page *page, enum track_item alloc) 3536{ 3537 void *addr = page_address(page); 3538 DECLARE_BITMAP(map, page->objects); 3539 void *p; 3540 3541 bitmap_zero(map, page->objects); 3542 for_each_free_object(p, s, page->freelist) 3543 set_bit(slab_index(p, s, addr), map); 3544 3545 for_each_object(p, s, addr, page->objects) 3546 if (!test_bit(slab_index(p, s, addr), map)) 3547 add_location(t, s, get_track(s, p, alloc)); 3548} 3549 3550static int list_locations(struct kmem_cache *s, char *buf, 3551 enum track_item alloc) 3552{ 3553 int len = 0; 3554 unsigned long i; 3555 struct loc_track t = { 0, 0, NULL }; 3556 int node; 3557 3558 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3559 GFP_TEMPORARY)) 3560 return sprintf(buf, "Out of memory\n"); 3561 3562 /* Push back cpu slabs */ 3563 flush_all(s); 3564 3565 for_each_node_state(node, N_NORMAL_MEMORY) { 3566 struct kmem_cache_node *n = get_node(s, node); 3567 unsigned long flags; 3568 struct page *page; 3569 3570 if (!atomic_long_read(&n->nr_slabs)) 3571 continue; 3572 3573 spin_lock_irqsave(&n->list_lock, flags); 3574 list_for_each_entry(page, &n->partial, lru) 3575 process_slab(&t, s, page, alloc); 3576 list_for_each_entry(page, &n->full, lru) 3577 process_slab(&t, s, page, alloc); 3578 spin_unlock_irqrestore(&n->list_lock, flags); 3579 } 3580 3581 for (i = 0; i < t.count; i++) { 3582 struct location *l = &t.loc[i]; 3583 3584 if (len > PAGE_SIZE - 100) 3585 break; 3586 len += sprintf(buf + len, "%7ld ", l->count); 3587 3588 if (l->addr) 3589 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3590 else 3591 len += sprintf(buf + len, "<not-available>"); 3592 3593 if (l->sum_time != l->min_time) { 3594 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3595 l->min_time, 3596 (long)div_u64(l->sum_time, l->count), 3597 l->max_time); 3598 } else 3599 len += sprintf(buf + len, " age=%ld", 3600 l->min_time); 3601 3602 if (l->min_pid != l->max_pid) 3603 len += sprintf(buf + len, " pid=%ld-%ld", 3604 l->min_pid, l->max_pid); 3605 else 3606 len += sprintf(buf + len, " pid=%ld", 3607 l->min_pid); 3608 3609 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && 3610 len < PAGE_SIZE - 60) { 3611 len += sprintf(buf + len, " cpus="); 3612 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3613 l->cpus); 3614 } 3615 3616 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && 3617 len < PAGE_SIZE - 60) { 3618 len += sprintf(buf + len, " nodes="); 3619 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3620 l->nodes); 3621 } 3622 3623 len += sprintf(buf + len, "\n"); 3624 } 3625 3626 free_loc_track(&t); 3627 if (!t.count) 3628 len += sprintf(buf, "No data\n"); 3629 return len; 3630} 3631 3632enum slab_stat_type { 3633 SL_ALL, /* All slabs */ 3634 SL_PARTIAL, /* Only partially allocated slabs */ 3635 SL_CPU, /* Only slabs used for cpu caches */ 3636 SL_OBJECTS, /* Determine allocated objects not slabs */ 3637 SL_TOTAL /* Determine object capacity not slabs */ 3638}; 3639 3640#define SO_ALL (1 << SL_ALL) 3641#define SO_PARTIAL (1 << SL_PARTIAL) 3642#define SO_CPU (1 << SL_CPU) 3643#define SO_OBJECTS (1 << SL_OBJECTS) 3644#define SO_TOTAL (1 << SL_TOTAL) 3645 3646static ssize_t show_slab_objects(struct kmem_cache *s, 3647 char *buf, unsigned long flags) 3648{ 3649 unsigned long total = 0; 3650 int node; 3651 int x; 3652 unsigned long *nodes; 3653 unsigned long *per_cpu; 3654 3655 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3656 if (!nodes) 3657 return -ENOMEM; 3658 per_cpu = nodes + nr_node_ids; 3659 3660 if (flags & SO_CPU) { 3661 int cpu; 3662 3663 for_each_possible_cpu(cpu) { 3664 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3665 3666 if (!c || c->node < 0) 3667 continue; 3668 3669 if (c->page) { 3670 if (flags & SO_TOTAL) 3671 x = c->page->objects; 3672 else if (flags & SO_OBJECTS) 3673 x = c->page->inuse; 3674 else 3675 x = 1; 3676 3677 total += x; 3678 nodes[c->node] += x; 3679 } 3680 per_cpu[c->node]++; 3681 } 3682 } 3683 3684 if (flags & SO_ALL) { 3685 for_each_node_state(node, N_NORMAL_MEMORY) { 3686 struct kmem_cache_node *n = get_node(s, node); 3687 3688 if (flags & SO_TOTAL) 3689 x = atomic_long_read(&n->total_objects); 3690 else if (flags & SO_OBJECTS) 3691 x = atomic_long_read(&n->total_objects) - 3692 count_partial(n, count_free); 3693 3694 else 3695 x = atomic_long_read(&n->nr_slabs); 3696 total += x; 3697 nodes[node] += x; 3698 } 3699 3700 } else if (flags & SO_PARTIAL) { 3701 for_each_node_state(node, N_NORMAL_MEMORY) { 3702 struct kmem_cache_node *n = get_node(s, node); 3703 3704 if (flags & SO_TOTAL) 3705 x = count_partial(n, count_total); 3706 else if (flags & SO_OBJECTS) 3707 x = count_partial(n, count_inuse); 3708 else 3709 x = n->nr_partial; 3710 total += x; 3711 nodes[node] += x; 3712 } 3713 } 3714 x = sprintf(buf, "%lu", total); 3715#ifdef CONFIG_NUMA 3716 for_each_node_state(node, N_NORMAL_MEMORY) 3717 if (nodes[node]) 3718 x += sprintf(buf + x, " N%d=%lu", 3719 node, nodes[node]); 3720#endif 3721 kfree(nodes); 3722 return x + sprintf(buf + x, "\n"); 3723} 3724 3725static int any_slab_objects(struct kmem_cache *s) 3726{ 3727 int node; 3728 3729 for_each_online_node(node) { 3730 struct kmem_cache_node *n = get_node(s, node); 3731 3732 if (!n) 3733 continue; 3734 3735 if (atomic_long_read(&n->total_objects)) 3736 return 1; 3737 } 3738 return 0; 3739} 3740 3741#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3742#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3743 3744struct slab_attribute { 3745 struct attribute attr; 3746 ssize_t (*show)(struct kmem_cache *s, char *buf); 3747 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3748}; 3749 3750#define SLAB_ATTR_RO(_name) \ 3751 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3752 3753#define SLAB_ATTR(_name) \ 3754 static struct slab_attribute _name##_attr = \ 3755 __ATTR(_name, 0644, _name##_show, _name##_store) 3756 3757static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3758{ 3759 return sprintf(buf, "%d\n", s->size); 3760} 3761SLAB_ATTR_RO(slab_size); 3762 3763static ssize_t align_show(struct kmem_cache *s, char *buf) 3764{ 3765 return sprintf(buf, "%d\n", s->align); 3766} 3767SLAB_ATTR_RO(align); 3768 3769static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3770{ 3771 return sprintf(buf, "%d\n", s->objsize); 3772} 3773SLAB_ATTR_RO(object_size); 3774 3775static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3776{ 3777 return sprintf(buf, "%d\n", oo_objects(s->oo)); 3778} 3779SLAB_ATTR_RO(objs_per_slab); 3780 3781static ssize_t order_store(struct kmem_cache *s, 3782 const char *buf, size_t length) 3783{ 3784 unsigned long order; 3785 int err; 3786 3787 err = strict_strtoul(buf, 10, &order); 3788 if (err) 3789 return err; 3790 3791 if (order > slub_max_order || order < slub_min_order) 3792 return -EINVAL; 3793 3794 calculate_sizes(s, order); 3795 return length; 3796} 3797 3798static ssize_t order_show(struct kmem_cache *s, char *buf) 3799{ 3800 return sprintf(buf, "%d\n", oo_order(s->oo)); 3801} 3802SLAB_ATTR(order); 3803 3804static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3805{ 3806 if (s->ctor) { 3807 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3808 3809 return n + sprintf(buf + n, "\n"); 3810 } 3811 return 0; 3812} 3813SLAB_ATTR_RO(ctor); 3814 3815static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3816{ 3817 return sprintf(buf, "%d\n", s->refcount - 1); 3818} 3819SLAB_ATTR_RO(aliases); 3820 3821static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3822{ 3823 return show_slab_objects(s, buf, SO_ALL); 3824} 3825SLAB_ATTR_RO(slabs); 3826 3827static ssize_t partial_show(struct kmem_cache *s, char *buf) 3828{ 3829 return show_slab_objects(s, buf, SO_PARTIAL); 3830} 3831SLAB_ATTR_RO(partial); 3832 3833static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 3834{ 3835 return show_slab_objects(s, buf, SO_CPU); 3836} 3837SLAB_ATTR_RO(cpu_slabs); 3838 3839static ssize_t objects_show(struct kmem_cache *s, char *buf) 3840{ 3841 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 3842} 3843SLAB_ATTR_RO(objects); 3844 3845static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 3846{ 3847 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 3848} 3849SLAB_ATTR_RO(objects_partial); 3850 3851static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 3852{ 3853 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 3854} 3855SLAB_ATTR_RO(total_objects); 3856 3857static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 3858{ 3859 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 3860} 3861 3862static ssize_t sanity_checks_store(struct kmem_cache *s, 3863 const char *buf, size_t length) 3864{ 3865 s->flags &= ~SLAB_DEBUG_FREE; 3866 if (buf[0] == '1') 3867 s->flags |= SLAB_DEBUG_FREE; 3868 return length; 3869} 3870SLAB_ATTR(sanity_checks); 3871 3872static ssize_t trace_show(struct kmem_cache *s, char *buf) 3873{ 3874 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 3875} 3876 3877static ssize_t trace_store(struct kmem_cache *s, const char *buf, 3878 size_t length) 3879{ 3880 s->flags &= ~SLAB_TRACE; 3881 if (buf[0] == '1') 3882 s->flags |= SLAB_TRACE; 3883 return length; 3884} 3885SLAB_ATTR(trace); 3886 3887static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 3888{ 3889 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 3890} 3891 3892static ssize_t reclaim_account_store(struct kmem_cache *s, 3893 const char *buf, size_t length) 3894{ 3895 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 3896 if (buf[0] == '1') 3897 s->flags |= SLAB_RECLAIM_ACCOUNT; 3898 return length; 3899} 3900SLAB_ATTR(reclaim_account); 3901 3902static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 3903{ 3904 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 3905} 3906SLAB_ATTR_RO(hwcache_align); 3907 3908#ifdef CONFIG_ZONE_DMA 3909static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 3910{ 3911 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 3912} 3913SLAB_ATTR_RO(cache_dma); 3914#endif 3915 3916static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 3917{ 3918 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 3919} 3920SLAB_ATTR_RO(destroy_by_rcu); 3921 3922static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 3923{ 3924 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 3925} 3926 3927static ssize_t red_zone_store(struct kmem_cache *s, 3928 const char *buf, size_t length) 3929{ 3930 if (any_slab_objects(s)) 3931 return -EBUSY; 3932 3933 s->flags &= ~SLAB_RED_ZONE; 3934 if (buf[0] == '1') 3935 s->flags |= SLAB_RED_ZONE; 3936 calculate_sizes(s, -1); 3937 return length; 3938} 3939SLAB_ATTR(red_zone); 3940 3941static ssize_t poison_show(struct kmem_cache *s, char *buf) 3942{ 3943 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 3944} 3945 3946static ssize_t poison_store(struct kmem_cache *s, 3947 const char *buf, size_t length) 3948{ 3949 if (any_slab_objects(s)) 3950 return -EBUSY; 3951 3952 s->flags &= ~SLAB_POISON; 3953 if (buf[0] == '1') 3954 s->flags |= SLAB_POISON; 3955 calculate_sizes(s, -1); 3956 return length; 3957} 3958SLAB_ATTR(poison); 3959 3960static ssize_t store_user_show(struct kmem_cache *s, char *buf) 3961{ 3962 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 3963} 3964 3965static ssize_t store_user_store(struct kmem_cache *s, 3966 const char *buf, size_t length) 3967{ 3968 if (any_slab_objects(s)) 3969 return -EBUSY; 3970 3971 s->flags &= ~SLAB_STORE_USER; 3972 if (buf[0] == '1') 3973 s->flags |= SLAB_STORE_USER; 3974 calculate_sizes(s, -1); 3975 return length; 3976} 3977SLAB_ATTR(store_user); 3978 3979static ssize_t validate_show(struct kmem_cache *s, char *buf) 3980{ 3981 return 0; 3982} 3983 3984static ssize_t validate_store(struct kmem_cache *s, 3985 const char *buf, size_t length) 3986{ 3987 int ret = -EINVAL; 3988 3989 if (buf[0] == '1') { 3990 ret = validate_slab_cache(s); 3991 if (ret >= 0) 3992 ret = length; 3993 } 3994 return ret; 3995} 3996SLAB_ATTR(validate); 3997 3998static ssize_t shrink_show(struct kmem_cache *s, char *buf) 3999{ 4000 return 0; 4001} 4002 4003static ssize_t shrink_store(struct kmem_cache *s, 4004 const char *buf, size_t length) 4005{ 4006 if (buf[0] == '1') { 4007 int rc = kmem_cache_shrink(s); 4008 4009 if (rc) 4010 return rc; 4011 } else 4012 return -EINVAL; 4013 return length; 4014} 4015SLAB_ATTR(shrink); 4016 4017static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4018{ 4019 if (!(s->flags & SLAB_STORE_USER)) 4020 return -ENOSYS; 4021 return list_locations(s, buf, TRACK_ALLOC); 4022} 4023SLAB_ATTR_RO(alloc_calls); 4024 4025static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4026{ 4027 if (!(s->flags & SLAB_STORE_USER)) 4028 return -ENOSYS; 4029 return list_locations(s, buf, TRACK_FREE); 4030} 4031SLAB_ATTR_RO(free_calls); 4032 4033#ifdef CONFIG_NUMA 4034static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4035{ 4036 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4037} 4038 4039static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4040 const char *buf, size_t length) 4041{ 4042 unsigned long ratio; 4043 int err; 4044 4045 err = strict_strtoul(buf, 10, &ratio); 4046 if (err) 4047 return err; 4048 4049 if (ratio < 100) 4050 s->remote_node_defrag_ratio = ratio * 10; 4051 4052 return length; 4053} 4054SLAB_ATTR(remote_node_defrag_ratio); 4055#endif 4056 4057#ifdef CONFIG_SLUB_STATS 4058static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4059{ 4060 unsigned long sum = 0; 4061 int cpu; 4062 int len; 4063 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4064 4065 if (!data) 4066 return -ENOMEM; 4067 4068 for_each_online_cpu(cpu) { 4069 unsigned x = get_cpu_slab(s, cpu)->stat[si]; 4070 4071 data[cpu] = x; 4072 sum += x; 4073 } 4074 4075 len = sprintf(buf, "%lu", sum); 4076 4077#ifdef CONFIG_SMP 4078 for_each_online_cpu(cpu) { 4079 if (data[cpu] && len < PAGE_SIZE - 20) 4080 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4081 } 4082#endif 4083 kfree(data); 4084 return len + sprintf(buf + len, "\n"); 4085} 4086 4087#define STAT_ATTR(si, text) \ 4088static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4089{ \ 4090 return show_stat(s, buf, si); \ 4091} \ 4092SLAB_ATTR_RO(text); \ 4093 4094STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4095STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4096STAT_ATTR(FREE_FASTPATH, free_fastpath); 4097STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4098STAT_ATTR(FREE_FROZEN, free_frozen); 4099STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4100STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4101STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4102STAT_ATTR(ALLOC_SLAB, alloc_slab); 4103STAT_ATTR(ALLOC_REFILL, alloc_refill); 4104STAT_ATTR(FREE_SLAB, free_slab); 4105STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4106STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4107STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4108STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4109STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4110STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4111STAT_ATTR(ORDER_FALLBACK, order_fallback); 4112#endif 4113 4114static struct attribute *slab_attrs[] = { 4115 &slab_size_attr.attr, 4116 &object_size_attr.attr, 4117 &objs_per_slab_attr.attr, 4118 &order_attr.attr, 4119 &objects_attr.attr, 4120 &objects_partial_attr.attr, 4121 &total_objects_attr.attr, 4122 &slabs_attr.attr, 4123 &partial_attr.attr, 4124 &cpu_slabs_attr.attr, 4125 &ctor_attr.attr, 4126 &aliases_attr.attr, 4127 &align_attr.attr, 4128 &sanity_checks_attr.attr, 4129 &trace_attr.attr, 4130 &hwcache_align_attr.attr, 4131 &reclaim_account_attr.attr, 4132 &destroy_by_rcu_attr.attr, 4133 &red_zone_attr.attr, 4134 &poison_attr.attr, 4135 &store_user_attr.attr, 4136 &validate_attr.attr, 4137 &shrink_attr.attr, 4138 &alloc_calls_attr.attr, 4139 &free_calls_attr.attr, 4140#ifdef CONFIG_ZONE_DMA 4141 &cache_dma_attr.attr, 4142#endif 4143#ifdef CONFIG_NUMA 4144 &remote_node_defrag_ratio_attr.attr, 4145#endif 4146#ifdef CONFIG_SLUB_STATS 4147 &alloc_fastpath_attr.attr, 4148 &alloc_slowpath_attr.attr, 4149 &free_fastpath_attr.attr, 4150 &free_slowpath_attr.attr, 4151 &free_frozen_attr.attr, 4152 &free_add_partial_attr.attr, 4153 &free_remove_partial_attr.attr, 4154 &alloc_from_partial_attr.attr, 4155 &alloc_slab_attr.attr, 4156 &alloc_refill_attr.attr, 4157 &free_slab_attr.attr, 4158 &cpuslab_flush_attr.attr, 4159 &deactivate_full_attr.attr, 4160 &deactivate_empty_attr.attr, 4161 &deactivate_to_head_attr.attr, 4162 &deactivate_to_tail_attr.attr, 4163 &deactivate_remote_frees_attr.attr, 4164 &order_fallback_attr.attr, 4165#endif 4166 NULL 4167}; 4168 4169static struct attribute_group slab_attr_group = { 4170 .attrs = slab_attrs, 4171}; 4172 4173static ssize_t slab_attr_show(struct kobject *kobj, 4174 struct attribute *attr, 4175 char *buf) 4176{ 4177 struct slab_attribute *attribute; 4178 struct kmem_cache *s; 4179 int err; 4180 4181 attribute = to_slab_attr(attr); 4182 s = to_slab(kobj); 4183 4184 if (!attribute->show) 4185 return -EIO; 4186 4187 err = attribute->show(s, buf); 4188 4189 return err; 4190} 4191 4192static ssize_t slab_attr_store(struct kobject *kobj, 4193 struct attribute *attr, 4194 const char *buf, size_t len) 4195{ 4196 struct slab_attribute *attribute; 4197 struct kmem_cache *s; 4198 int err; 4199 4200 attribute = to_slab_attr(attr); 4201 s = to_slab(kobj); 4202 4203 if (!attribute->store) 4204 return -EIO; 4205 4206 err = attribute->store(s, buf, len); 4207 4208 return err; 4209} 4210 4211static void kmem_cache_release(struct kobject *kobj) 4212{ 4213 struct kmem_cache *s = to_slab(kobj); 4214 4215 kfree(s); 4216} 4217 4218static struct sysfs_ops slab_sysfs_ops = { 4219 .show = slab_attr_show, 4220 .store = slab_attr_store, 4221}; 4222 4223static struct kobj_type slab_ktype = { 4224 .sysfs_ops = &slab_sysfs_ops, 4225 .release = kmem_cache_release 4226}; 4227 4228static int uevent_filter(struct kset *kset, struct kobject *kobj) 4229{ 4230 struct kobj_type *ktype = get_ktype(kobj); 4231 4232 if (ktype == &slab_ktype) 4233 return 1; 4234 return 0; 4235} 4236 4237static struct kset_uevent_ops slab_uevent_ops = { 4238 .filter = uevent_filter, 4239}; 4240 4241static struct kset *slab_kset; 4242 4243#define ID_STR_LENGTH 64 4244 4245/* Create a unique string id for a slab cache: 4246 * 4247 * Format :[flags-]size 4248 */ 4249static char *create_unique_id(struct kmem_cache *s) 4250{ 4251 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4252 char *p = name; 4253 4254 BUG_ON(!name); 4255 4256 *p++ = ':'; 4257 /* 4258 * First flags affecting slabcache operations. We will only 4259 * get here for aliasable slabs so we do not need to support 4260 * too many flags. The flags here must cover all flags that 4261 * are matched during merging to guarantee that the id is 4262 * unique. 4263 */ 4264 if (s->flags & SLAB_CACHE_DMA) 4265 *p++ = 'd'; 4266 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4267 *p++ = 'a'; 4268 if (s->flags & SLAB_DEBUG_FREE) 4269 *p++ = 'F'; 4270 if (p != name + 1) 4271 *p++ = '-'; 4272 p += sprintf(p, "%07d", s->size); 4273 BUG_ON(p > name + ID_STR_LENGTH - 1); 4274 return name; 4275} 4276 4277static int sysfs_slab_add(struct kmem_cache *s) 4278{ 4279 int err; 4280 const char *name; 4281 int unmergeable; 4282 4283 if (slab_state < SYSFS) 4284 /* Defer until later */ 4285 return 0; 4286 4287 unmergeable = slab_unmergeable(s); 4288 if (unmergeable) { 4289 /* 4290 * Slabcache can never be merged so we can use the name proper. 4291 * This is typically the case for debug situations. In that 4292 * case we can catch duplicate names easily. 4293 */ 4294 sysfs_remove_link(&slab_kset->kobj, s->name); 4295 name = s->name; 4296 } else { 4297 /* 4298 * Create a unique name for the slab as a target 4299 * for the symlinks. 4300 */ 4301 name = create_unique_id(s); 4302 } 4303 4304 s->kobj.kset = slab_kset; 4305 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4306 if (err) { 4307 kobject_put(&s->kobj); 4308 return err; 4309 } 4310 4311 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4312 if (err) 4313 return err; 4314 kobject_uevent(&s->kobj, KOBJ_ADD); 4315 if (!unmergeable) { 4316 /* Setup first alias */ 4317 sysfs_slab_alias(s, s->name); 4318 kfree(name); 4319 } 4320 return 0; 4321} 4322 4323static void sysfs_slab_remove(struct kmem_cache *s) 4324{ 4325 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4326 kobject_del(&s->kobj); 4327 kobject_put(&s->kobj); 4328} 4329 4330/* 4331 * Need to buffer aliases during bootup until sysfs becomes 4332 * available lest we loose that information. 4333 */ 4334struct saved_alias { 4335 struct kmem_cache *s; 4336 const char *name; 4337 struct saved_alias *next; 4338}; 4339 4340static struct saved_alias *alias_list; 4341 4342static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4343{ 4344 struct saved_alias *al; 4345 4346 if (slab_state == SYSFS) { 4347 /* 4348 * If we have a leftover link then remove it. 4349 */ 4350 sysfs_remove_link(&slab_kset->kobj, name); 4351 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4352 } 4353 4354 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4355 if (!al) 4356 return -ENOMEM; 4357 4358 al->s = s; 4359 al->name = name; 4360 al->next = alias_list; 4361 alias_list = al; 4362 return 0; 4363} 4364 4365static int __init slab_sysfs_init(void) 4366{ 4367 struct kmem_cache *s; 4368 int err; 4369 4370 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4371 if (!slab_kset) { 4372 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4373 return -ENOSYS; 4374 } 4375 4376 slab_state = SYSFS; 4377 4378 list_for_each_entry(s, &slab_caches, list) { 4379 err = sysfs_slab_add(s); 4380 if (err) 4381 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4382 " to sysfs\n", s->name); 4383 } 4384 4385 while (alias_list) { 4386 struct saved_alias *al = alias_list; 4387 4388 alias_list = alias_list->next; 4389 err = sysfs_slab_alias(al->s, al->name); 4390 if (err) 4391 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4392 " %s to sysfs\n", s->name); 4393 kfree(al); 4394 } 4395 4396 resiliency_test(); 4397 return 0; 4398} 4399 4400__initcall(slab_sysfs_init); 4401#endif 4402 4403/* 4404 * The /proc/slabinfo ABI 4405 */ 4406#ifdef CONFIG_SLABINFO 4407 4408ssize_t slabinfo_write(struct file *file, const char __user *buffer, 4409 size_t count, loff_t *ppos) 4410{ 4411 return -EINVAL; 4412} 4413 4414 4415static void print_slabinfo_header(struct seq_file *m) 4416{ 4417 seq_puts(m, "slabinfo - version: 2.1\n"); 4418 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4419 "<objperslab> <pagesperslab>"); 4420 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4421 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4422 seq_putc(m, '\n'); 4423} 4424 4425static void *s_start(struct seq_file *m, loff_t *pos) 4426{ 4427 loff_t n = *pos; 4428 4429 down_read(&slub_lock); 4430 if (!n) 4431 print_slabinfo_header(m); 4432 4433 return seq_list_start(&slab_caches, *pos); 4434} 4435 4436static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4437{ 4438 return seq_list_next(p, &slab_caches, pos); 4439} 4440 4441static void s_stop(struct seq_file *m, void *p) 4442{ 4443 up_read(&slub_lock); 4444} 4445 4446static int s_show(struct seq_file *m, void *p) 4447{ 4448 unsigned long nr_partials = 0; 4449 unsigned long nr_slabs = 0; 4450 unsigned long nr_inuse = 0; 4451 unsigned long nr_objs = 0; 4452 unsigned long nr_free = 0; 4453 struct kmem_cache *s; 4454 int node; 4455 4456 s = list_entry(p, struct kmem_cache, list); 4457 4458 for_each_online_node(node) { 4459 struct kmem_cache_node *n = get_node(s, node); 4460 4461 if (!n) 4462 continue; 4463 4464 nr_partials += n->nr_partial; 4465 nr_slabs += atomic_long_read(&n->nr_slabs); 4466 nr_objs += atomic_long_read(&n->total_objects); 4467 nr_free += count_partial(n, count_free); 4468 } 4469 4470 nr_inuse = nr_objs - nr_free; 4471 4472 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4473 nr_objs, s->size, oo_objects(s->oo), 4474 (1 << oo_order(s->oo))); 4475 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4476 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4477 0UL); 4478 seq_putc(m, '\n'); 4479 return 0; 4480} 4481 4482const struct seq_operations slabinfo_op = { 4483 .start = s_start, 4484 .next = s_next, 4485 .stop = s_stop, 4486 .show = s_show, 4487}; 4488 4489#endif /* CONFIG_SLABINFO */ 4490