slub.c revision 0094de92a4f1da3a845ccc4ecb12ec0db8e48997
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/proc_fs.h> 18#include <linux/seq_file.h> 19#include <linux/cpu.h> 20#include <linux/cpuset.h> 21#include <linux/mempolicy.h> 22#include <linux/ctype.h> 23#include <linux/debugobjects.h> 24#include <linux/kallsyms.h> 25#include <linux/memory.h> 26#include <linux/math64.h> 27 28/* 29 * Lock order: 30 * 1. slab_lock(page) 31 * 2. slab->list_lock 32 * 33 * The slab_lock protects operations on the object of a particular 34 * slab and its metadata in the page struct. If the slab lock 35 * has been taken then no allocations nor frees can be performed 36 * on the objects in the slab nor can the slab be added or removed 37 * from the partial or full lists since this would mean modifying 38 * the page_struct of the slab. 39 * 40 * The list_lock protects the partial and full list on each node and 41 * the partial slab counter. If taken then no new slabs may be added or 42 * removed from the lists nor make the number of partial slabs be modified. 43 * (Note that the total number of slabs is an atomic value that may be 44 * modified without taking the list lock). 45 * 46 * The list_lock is a centralized lock and thus we avoid taking it as 47 * much as possible. As long as SLUB does not have to handle partial 48 * slabs, operations can continue without any centralized lock. F.e. 49 * allocating a long series of objects that fill up slabs does not require 50 * the list lock. 51 * 52 * The lock order is sometimes inverted when we are trying to get a slab 53 * off a list. We take the list_lock and then look for a page on the list 54 * to use. While we do that objects in the slabs may be freed. We can 55 * only operate on the slab if we have also taken the slab_lock. So we use 56 * a slab_trylock() on the slab. If trylock was successful then no frees 57 * can occur anymore and we can use the slab for allocations etc. If the 58 * slab_trylock() does not succeed then frees are in progress in the slab and 59 * we must stay away from it for a while since we may cause a bouncing 60 * cacheline if we try to acquire the lock. So go onto the next slab. 61 * If all pages are busy then we may allocate a new slab instead of reusing 62 * a partial slab. A new slab has noone operating on it and thus there is 63 * no danger of cacheline contention. 64 * 65 * Interrupts are disabled during allocation and deallocation in order to 66 * make the slab allocator safe to use in the context of an irq. In addition 67 * interrupts are disabled to ensure that the processor does not change 68 * while handling per_cpu slabs, due to kernel preemption. 69 * 70 * SLUB assigns one slab for allocation to each processor. 71 * Allocations only occur from these slabs called cpu slabs. 72 * 73 * Slabs with free elements are kept on a partial list and during regular 74 * operations no list for full slabs is used. If an object in a full slab is 75 * freed then the slab will show up again on the partial lists. 76 * We track full slabs for debugging purposes though because otherwise we 77 * cannot scan all objects. 78 * 79 * Slabs are freed when they become empty. Teardown and setup is 80 * minimal so we rely on the page allocators per cpu caches for 81 * fast frees and allocs. 82 * 83 * Overloading of page flags that are otherwise used for LRU management. 84 * 85 * PageActive The slab is frozen and exempt from list processing. 86 * This means that the slab is dedicated to a purpose 87 * such as satisfying allocations for a specific 88 * processor. Objects may be freed in the slab while 89 * it is frozen but slab_free will then skip the usual 90 * list operations. It is up to the processor holding 91 * the slab to integrate the slab into the slab lists 92 * when the slab is no longer needed. 93 * 94 * One use of this flag is to mark slabs that are 95 * used for allocations. Then such a slab becomes a cpu 96 * slab. The cpu slab may be equipped with an additional 97 * freelist that allows lockless access to 98 * free objects in addition to the regular freelist 99 * that requires the slab lock. 100 * 101 * PageError Slab requires special handling due to debug 102 * options set. This moves slab handling out of 103 * the fast path and disables lockless freelists. 104 */ 105 106#ifdef CONFIG_SLUB_DEBUG 107#define SLABDEBUG 1 108#else 109#define SLABDEBUG 0 110#endif 111 112/* 113 * Issues still to be resolved: 114 * 115 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 116 * 117 * - Variable sizing of the per node arrays 118 */ 119 120/* Enable to test recovery from slab corruption on boot */ 121#undef SLUB_RESILIENCY_TEST 122 123/* 124 * Mininum number of partial slabs. These will be left on the partial 125 * lists even if they are empty. kmem_cache_shrink may reclaim them. 126 */ 127#define MIN_PARTIAL 5 128 129/* 130 * Maximum number of desirable partial slabs. 131 * The existence of more partial slabs makes kmem_cache_shrink 132 * sort the partial list by the number of objects in the. 133 */ 134#define MAX_PARTIAL 10 135 136#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 137 SLAB_POISON | SLAB_STORE_USER) 138 139/* 140 * Set of flags that will prevent slab merging 141 */ 142#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 143 SLAB_TRACE | SLAB_DESTROY_BY_RCU) 144 145#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 146 SLAB_CACHE_DMA) 147 148#ifndef ARCH_KMALLOC_MINALIGN 149#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 150#endif 151 152#ifndef ARCH_SLAB_MINALIGN 153#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) 154#endif 155 156#define OO_SHIFT 16 157#define OO_MASK ((1 << OO_SHIFT) - 1) 158#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ 159 160/* Internal SLUB flags */ 161#define __OBJECT_POISON 0x80000000 /* Poison object */ 162#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ 163 164static int kmem_size = sizeof(struct kmem_cache); 165 166#ifdef CONFIG_SMP 167static struct notifier_block slab_notifier; 168#endif 169 170static enum { 171 DOWN, /* No slab functionality available */ 172 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 173 UP, /* Everything works but does not show up in sysfs */ 174 SYSFS /* Sysfs up */ 175} slab_state = DOWN; 176 177/* A list of all slab caches on the system */ 178static DECLARE_RWSEM(slub_lock); 179static LIST_HEAD(slab_caches); 180 181/* 182 * Tracking user of a slab. 183 */ 184struct track { 185 unsigned long addr; /* Called from address */ 186 int cpu; /* Was running on cpu */ 187 int pid; /* Pid context */ 188 unsigned long when; /* When did the operation occur */ 189}; 190 191enum track_item { TRACK_ALLOC, TRACK_FREE }; 192 193#ifdef CONFIG_SLUB_DEBUG 194static int sysfs_slab_add(struct kmem_cache *); 195static int sysfs_slab_alias(struct kmem_cache *, const char *); 196static void sysfs_slab_remove(struct kmem_cache *); 197 198#else 199static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 200static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 201 { return 0; } 202static inline void sysfs_slab_remove(struct kmem_cache *s) 203{ 204 kfree(s); 205} 206 207#endif 208 209static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) 210{ 211#ifdef CONFIG_SLUB_STATS 212 c->stat[si]++; 213#endif 214} 215 216/******************************************************************** 217 * Core slab cache functions 218 *******************************************************************/ 219 220int slab_is_available(void) 221{ 222 return slab_state >= UP; 223} 224 225static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 226{ 227#ifdef CONFIG_NUMA 228 return s->node[node]; 229#else 230 return &s->local_node; 231#endif 232} 233 234static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) 235{ 236#ifdef CONFIG_SMP 237 return s->cpu_slab[cpu]; 238#else 239 return &s->cpu_slab; 240#endif 241} 242 243/* Verify that a pointer has an address that is valid within a slab page */ 244static inline int check_valid_pointer(struct kmem_cache *s, 245 struct page *page, const void *object) 246{ 247 void *base; 248 249 if (!object) 250 return 1; 251 252 base = page_address(page); 253 if (object < base || object >= base + page->objects * s->size || 254 (object - base) % s->size) { 255 return 0; 256 } 257 258 return 1; 259} 260 261/* 262 * Slow version of get and set free pointer. 263 * 264 * This version requires touching the cache lines of kmem_cache which 265 * we avoid to do in the fast alloc free paths. There we obtain the offset 266 * from the page struct. 267 */ 268static inline void *get_freepointer(struct kmem_cache *s, void *object) 269{ 270 return *(void **)(object + s->offset); 271} 272 273static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 274{ 275 *(void **)(object + s->offset) = fp; 276} 277 278/* Loop over all objects in a slab */ 279#define for_each_object(__p, __s, __addr, __objects) \ 280 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 281 __p += (__s)->size) 282 283/* Scan freelist */ 284#define for_each_free_object(__p, __s, __free) \ 285 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 286 287/* Determine object index from a given position */ 288static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 289{ 290 return (p - addr) / s->size; 291} 292 293static inline struct kmem_cache_order_objects oo_make(int order, 294 unsigned long size) 295{ 296 struct kmem_cache_order_objects x = { 297 (order << OO_SHIFT) + (PAGE_SIZE << order) / size 298 }; 299 300 return x; 301} 302 303static inline int oo_order(struct kmem_cache_order_objects x) 304{ 305 return x.x >> OO_SHIFT; 306} 307 308static inline int oo_objects(struct kmem_cache_order_objects x) 309{ 310 return x.x & OO_MASK; 311} 312 313#ifdef CONFIG_SLUB_DEBUG 314/* 315 * Debug settings: 316 */ 317#ifdef CONFIG_SLUB_DEBUG_ON 318static int slub_debug = DEBUG_DEFAULT_FLAGS; 319#else 320static int slub_debug; 321#endif 322 323static char *slub_debug_slabs; 324 325/* 326 * Object debugging 327 */ 328static void print_section(char *text, u8 *addr, unsigned int length) 329{ 330 int i, offset; 331 int newline = 1; 332 char ascii[17]; 333 334 ascii[16] = 0; 335 336 for (i = 0; i < length; i++) { 337 if (newline) { 338 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 339 newline = 0; 340 } 341 printk(KERN_CONT " %02x", addr[i]); 342 offset = i % 16; 343 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 344 if (offset == 15) { 345 printk(KERN_CONT " %s\n", ascii); 346 newline = 1; 347 } 348 } 349 if (!newline) { 350 i %= 16; 351 while (i < 16) { 352 printk(KERN_CONT " "); 353 ascii[i] = ' '; 354 i++; 355 } 356 printk(KERN_CONT " %s\n", ascii); 357 } 358} 359 360static struct track *get_track(struct kmem_cache *s, void *object, 361 enum track_item alloc) 362{ 363 struct track *p; 364 365 if (s->offset) 366 p = object + s->offset + sizeof(void *); 367 else 368 p = object + s->inuse; 369 370 return p + alloc; 371} 372 373static void set_track(struct kmem_cache *s, void *object, 374 enum track_item alloc, unsigned long addr) 375{ 376 struct track *p; 377 378 if (s->offset) 379 p = object + s->offset + sizeof(void *); 380 else 381 p = object + s->inuse; 382 383 p += alloc; 384 if (addr) { 385 p->addr = addr; 386 p->cpu = smp_processor_id(); 387 p->pid = current->pid; 388 p->when = jiffies; 389 } else 390 memset(p, 0, sizeof(struct track)); 391} 392 393static void init_tracking(struct kmem_cache *s, void *object) 394{ 395 if (!(s->flags & SLAB_STORE_USER)) 396 return; 397 398 set_track(s, object, TRACK_FREE, 0UL); 399 set_track(s, object, TRACK_ALLOC, 0UL); 400} 401 402static void print_track(const char *s, struct track *t) 403{ 404 if (!t->addr) 405 return; 406 407 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 408 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 409} 410 411static void print_tracking(struct kmem_cache *s, void *object) 412{ 413 if (!(s->flags & SLAB_STORE_USER)) 414 return; 415 416 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 417 print_track("Freed", get_track(s, object, TRACK_FREE)); 418} 419 420static void print_page_info(struct page *page) 421{ 422 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 423 page, page->objects, page->inuse, page->freelist, page->flags); 424 425} 426 427static void slab_bug(struct kmem_cache *s, char *fmt, ...) 428{ 429 va_list args; 430 char buf[100]; 431 432 va_start(args, fmt); 433 vsnprintf(buf, sizeof(buf), fmt, args); 434 va_end(args); 435 printk(KERN_ERR "========================================" 436 "=====================================\n"); 437 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 438 printk(KERN_ERR "----------------------------------------" 439 "-------------------------------------\n\n"); 440} 441 442static void slab_fix(struct kmem_cache *s, char *fmt, ...) 443{ 444 va_list args; 445 char buf[100]; 446 447 va_start(args, fmt); 448 vsnprintf(buf, sizeof(buf), fmt, args); 449 va_end(args); 450 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 451} 452 453static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 454{ 455 unsigned int off; /* Offset of last byte */ 456 u8 *addr = page_address(page); 457 458 print_tracking(s, p); 459 460 print_page_info(page); 461 462 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 463 p, p - addr, get_freepointer(s, p)); 464 465 if (p > addr + 16) 466 print_section("Bytes b4", p - 16, 16); 467 468 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 469 470 if (s->flags & SLAB_RED_ZONE) 471 print_section("Redzone", p + s->objsize, 472 s->inuse - s->objsize); 473 474 if (s->offset) 475 off = s->offset + sizeof(void *); 476 else 477 off = s->inuse; 478 479 if (s->flags & SLAB_STORE_USER) 480 off += 2 * sizeof(struct track); 481 482 if (off != s->size) 483 /* Beginning of the filler is the free pointer */ 484 print_section("Padding", p + off, s->size - off); 485 486 dump_stack(); 487} 488 489static void object_err(struct kmem_cache *s, struct page *page, 490 u8 *object, char *reason) 491{ 492 slab_bug(s, "%s", reason); 493 print_trailer(s, page, object); 494} 495 496static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 497{ 498 va_list args; 499 char buf[100]; 500 501 va_start(args, fmt); 502 vsnprintf(buf, sizeof(buf), fmt, args); 503 va_end(args); 504 slab_bug(s, "%s", buf); 505 print_page_info(page); 506 dump_stack(); 507} 508 509static void init_object(struct kmem_cache *s, void *object, int active) 510{ 511 u8 *p = object; 512 513 if (s->flags & __OBJECT_POISON) { 514 memset(p, POISON_FREE, s->objsize - 1); 515 p[s->objsize - 1] = POISON_END; 516 } 517 518 if (s->flags & SLAB_RED_ZONE) 519 memset(p + s->objsize, 520 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 521 s->inuse - s->objsize); 522} 523 524static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 525{ 526 while (bytes) { 527 if (*start != (u8)value) 528 return start; 529 start++; 530 bytes--; 531 } 532 return NULL; 533} 534 535static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 536 void *from, void *to) 537{ 538 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 539 memset(from, data, to - from); 540} 541 542static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 543 u8 *object, char *what, 544 u8 *start, unsigned int value, unsigned int bytes) 545{ 546 u8 *fault; 547 u8 *end; 548 549 fault = check_bytes(start, value, bytes); 550 if (!fault) 551 return 1; 552 553 end = start + bytes; 554 while (end > fault && end[-1] == value) 555 end--; 556 557 slab_bug(s, "%s overwritten", what); 558 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 559 fault, end - 1, fault[0], value); 560 print_trailer(s, page, object); 561 562 restore_bytes(s, what, value, fault, end); 563 return 0; 564} 565 566/* 567 * Object layout: 568 * 569 * object address 570 * Bytes of the object to be managed. 571 * If the freepointer may overlay the object then the free 572 * pointer is the first word of the object. 573 * 574 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 575 * 0xa5 (POISON_END) 576 * 577 * object + s->objsize 578 * Padding to reach word boundary. This is also used for Redzoning. 579 * Padding is extended by another word if Redzoning is enabled and 580 * objsize == inuse. 581 * 582 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 583 * 0xcc (RED_ACTIVE) for objects in use. 584 * 585 * object + s->inuse 586 * Meta data starts here. 587 * 588 * A. Free pointer (if we cannot overwrite object on free) 589 * B. Tracking data for SLAB_STORE_USER 590 * C. Padding to reach required alignment boundary or at mininum 591 * one word if debugging is on to be able to detect writes 592 * before the word boundary. 593 * 594 * Padding is done using 0x5a (POISON_INUSE) 595 * 596 * object + s->size 597 * Nothing is used beyond s->size. 598 * 599 * If slabcaches are merged then the objsize and inuse boundaries are mostly 600 * ignored. And therefore no slab options that rely on these boundaries 601 * may be used with merged slabcaches. 602 */ 603 604static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 605{ 606 unsigned long off = s->inuse; /* The end of info */ 607 608 if (s->offset) 609 /* Freepointer is placed after the object. */ 610 off += sizeof(void *); 611 612 if (s->flags & SLAB_STORE_USER) 613 /* We also have user information there */ 614 off += 2 * sizeof(struct track); 615 616 if (s->size == off) 617 return 1; 618 619 return check_bytes_and_report(s, page, p, "Object padding", 620 p + off, POISON_INUSE, s->size - off); 621} 622 623/* Check the pad bytes at the end of a slab page */ 624static int slab_pad_check(struct kmem_cache *s, struct page *page) 625{ 626 u8 *start; 627 u8 *fault; 628 u8 *end; 629 int length; 630 int remainder; 631 632 if (!(s->flags & SLAB_POISON)) 633 return 1; 634 635 start = page_address(page); 636 length = (PAGE_SIZE << compound_order(page)); 637 end = start + length; 638 remainder = length % s->size; 639 if (!remainder) 640 return 1; 641 642 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 643 if (!fault) 644 return 1; 645 while (end > fault && end[-1] == POISON_INUSE) 646 end--; 647 648 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 649 print_section("Padding", end - remainder, remainder); 650 651 restore_bytes(s, "slab padding", POISON_INUSE, start, end); 652 return 0; 653} 654 655static int check_object(struct kmem_cache *s, struct page *page, 656 void *object, int active) 657{ 658 u8 *p = object; 659 u8 *endobject = object + s->objsize; 660 661 if (s->flags & SLAB_RED_ZONE) { 662 unsigned int red = 663 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 664 665 if (!check_bytes_and_report(s, page, object, "Redzone", 666 endobject, red, s->inuse - s->objsize)) 667 return 0; 668 } else { 669 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 670 check_bytes_and_report(s, page, p, "Alignment padding", 671 endobject, POISON_INUSE, s->inuse - s->objsize); 672 } 673 } 674 675 if (s->flags & SLAB_POISON) { 676 if (!active && (s->flags & __OBJECT_POISON) && 677 (!check_bytes_and_report(s, page, p, "Poison", p, 678 POISON_FREE, s->objsize - 1) || 679 !check_bytes_and_report(s, page, p, "Poison", 680 p + s->objsize - 1, POISON_END, 1))) 681 return 0; 682 /* 683 * check_pad_bytes cleans up on its own. 684 */ 685 check_pad_bytes(s, page, p); 686 } 687 688 if (!s->offset && active) 689 /* 690 * Object and freepointer overlap. Cannot check 691 * freepointer while object is allocated. 692 */ 693 return 1; 694 695 /* Check free pointer validity */ 696 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 697 object_err(s, page, p, "Freepointer corrupt"); 698 /* 699 * No choice but to zap it and thus loose the remainder 700 * of the free objects in this slab. May cause 701 * another error because the object count is now wrong. 702 */ 703 set_freepointer(s, p, NULL); 704 return 0; 705 } 706 return 1; 707} 708 709static int check_slab(struct kmem_cache *s, struct page *page) 710{ 711 int maxobj; 712 713 VM_BUG_ON(!irqs_disabled()); 714 715 if (!PageSlab(page)) { 716 slab_err(s, page, "Not a valid slab page"); 717 return 0; 718 } 719 720 maxobj = (PAGE_SIZE << compound_order(page)) / s->size; 721 if (page->objects > maxobj) { 722 slab_err(s, page, "objects %u > max %u", 723 s->name, page->objects, maxobj); 724 return 0; 725 } 726 if (page->inuse > page->objects) { 727 slab_err(s, page, "inuse %u > max %u", 728 s->name, page->inuse, page->objects); 729 return 0; 730 } 731 /* Slab_pad_check fixes things up after itself */ 732 slab_pad_check(s, page); 733 return 1; 734} 735 736/* 737 * Determine if a certain object on a page is on the freelist. Must hold the 738 * slab lock to guarantee that the chains are in a consistent state. 739 */ 740static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 741{ 742 int nr = 0; 743 void *fp = page->freelist; 744 void *object = NULL; 745 unsigned long max_objects; 746 747 while (fp && nr <= page->objects) { 748 if (fp == search) 749 return 1; 750 if (!check_valid_pointer(s, page, fp)) { 751 if (object) { 752 object_err(s, page, object, 753 "Freechain corrupt"); 754 set_freepointer(s, object, NULL); 755 break; 756 } else { 757 slab_err(s, page, "Freepointer corrupt"); 758 page->freelist = NULL; 759 page->inuse = page->objects; 760 slab_fix(s, "Freelist cleared"); 761 return 0; 762 } 763 break; 764 } 765 object = fp; 766 fp = get_freepointer(s, object); 767 nr++; 768 } 769 770 max_objects = (PAGE_SIZE << compound_order(page)) / s->size; 771 if (max_objects > MAX_OBJS_PER_PAGE) 772 max_objects = MAX_OBJS_PER_PAGE; 773 774 if (page->objects != max_objects) { 775 slab_err(s, page, "Wrong number of objects. Found %d but " 776 "should be %d", page->objects, max_objects); 777 page->objects = max_objects; 778 slab_fix(s, "Number of objects adjusted."); 779 } 780 if (page->inuse != page->objects - nr) { 781 slab_err(s, page, "Wrong object count. Counter is %d but " 782 "counted were %d", page->inuse, page->objects - nr); 783 page->inuse = page->objects - nr; 784 slab_fix(s, "Object count adjusted."); 785 } 786 return search == NULL; 787} 788 789static void trace(struct kmem_cache *s, struct page *page, void *object, 790 int alloc) 791{ 792 if (s->flags & SLAB_TRACE) { 793 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 794 s->name, 795 alloc ? "alloc" : "free", 796 object, page->inuse, 797 page->freelist); 798 799 if (!alloc) 800 print_section("Object", (void *)object, s->objsize); 801 802 dump_stack(); 803 } 804} 805 806/* 807 * Tracking of fully allocated slabs for debugging purposes. 808 */ 809static void add_full(struct kmem_cache_node *n, struct page *page) 810{ 811 spin_lock(&n->list_lock); 812 list_add(&page->lru, &n->full); 813 spin_unlock(&n->list_lock); 814} 815 816static void remove_full(struct kmem_cache *s, struct page *page) 817{ 818 struct kmem_cache_node *n; 819 820 if (!(s->flags & SLAB_STORE_USER)) 821 return; 822 823 n = get_node(s, page_to_nid(page)); 824 825 spin_lock(&n->list_lock); 826 list_del(&page->lru); 827 spin_unlock(&n->list_lock); 828} 829 830/* Tracking of the number of slabs for debugging purposes */ 831static inline unsigned long slabs_node(struct kmem_cache *s, int node) 832{ 833 struct kmem_cache_node *n = get_node(s, node); 834 835 return atomic_long_read(&n->nr_slabs); 836} 837 838static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 839{ 840 struct kmem_cache_node *n = get_node(s, node); 841 842 /* 843 * May be called early in order to allocate a slab for the 844 * kmem_cache_node structure. Solve the chicken-egg 845 * dilemma by deferring the increment of the count during 846 * bootstrap (see early_kmem_cache_node_alloc). 847 */ 848 if (!NUMA_BUILD || n) { 849 atomic_long_inc(&n->nr_slabs); 850 atomic_long_add(objects, &n->total_objects); 851 } 852} 853static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 854{ 855 struct kmem_cache_node *n = get_node(s, node); 856 857 atomic_long_dec(&n->nr_slabs); 858 atomic_long_sub(objects, &n->total_objects); 859} 860 861/* Object debug checks for alloc/free paths */ 862static void setup_object_debug(struct kmem_cache *s, struct page *page, 863 void *object) 864{ 865 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 866 return; 867 868 init_object(s, object, 0); 869 init_tracking(s, object); 870} 871 872static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 873 void *object, unsigned long addr) 874{ 875 if (!check_slab(s, page)) 876 goto bad; 877 878 if (!on_freelist(s, page, object)) { 879 object_err(s, page, object, "Object already allocated"); 880 goto bad; 881 } 882 883 if (!check_valid_pointer(s, page, object)) { 884 object_err(s, page, object, "Freelist Pointer check fails"); 885 goto bad; 886 } 887 888 if (!check_object(s, page, object, 0)) 889 goto bad; 890 891 /* Success perform special debug activities for allocs */ 892 if (s->flags & SLAB_STORE_USER) 893 set_track(s, object, TRACK_ALLOC, addr); 894 trace(s, page, object, 1); 895 init_object(s, object, 1); 896 return 1; 897 898bad: 899 if (PageSlab(page)) { 900 /* 901 * If this is a slab page then lets do the best we can 902 * to avoid issues in the future. Marking all objects 903 * as used avoids touching the remaining objects. 904 */ 905 slab_fix(s, "Marking all objects used"); 906 page->inuse = page->objects; 907 page->freelist = NULL; 908 } 909 return 0; 910} 911 912static int free_debug_processing(struct kmem_cache *s, struct page *page, 913 void *object, unsigned long addr) 914{ 915 if (!check_slab(s, page)) 916 goto fail; 917 918 if (!check_valid_pointer(s, page, object)) { 919 slab_err(s, page, "Invalid object pointer 0x%p", object); 920 goto fail; 921 } 922 923 if (on_freelist(s, page, object)) { 924 object_err(s, page, object, "Object already free"); 925 goto fail; 926 } 927 928 if (!check_object(s, page, object, 1)) 929 return 0; 930 931 if (unlikely(s != page->slab)) { 932 if (!PageSlab(page)) { 933 slab_err(s, page, "Attempt to free object(0x%p) " 934 "outside of slab", object); 935 } else if (!page->slab) { 936 printk(KERN_ERR 937 "SLUB <none>: no slab for object 0x%p.\n", 938 object); 939 dump_stack(); 940 } else 941 object_err(s, page, object, 942 "page slab pointer corrupt."); 943 goto fail; 944 } 945 946 /* Special debug activities for freeing objects */ 947 if (!PageSlubFrozen(page) && !page->freelist) 948 remove_full(s, page); 949 if (s->flags & SLAB_STORE_USER) 950 set_track(s, object, TRACK_FREE, addr); 951 trace(s, page, object, 0); 952 init_object(s, object, 0); 953 return 1; 954 955fail: 956 slab_fix(s, "Object at 0x%p not freed", object); 957 return 0; 958} 959 960static int __init setup_slub_debug(char *str) 961{ 962 slub_debug = DEBUG_DEFAULT_FLAGS; 963 if (*str++ != '=' || !*str) 964 /* 965 * No options specified. Switch on full debugging. 966 */ 967 goto out; 968 969 if (*str == ',') 970 /* 971 * No options but restriction on slabs. This means full 972 * debugging for slabs matching a pattern. 973 */ 974 goto check_slabs; 975 976 slub_debug = 0; 977 if (*str == '-') 978 /* 979 * Switch off all debugging measures. 980 */ 981 goto out; 982 983 /* 984 * Determine which debug features should be switched on 985 */ 986 for (; *str && *str != ','; str++) { 987 switch (tolower(*str)) { 988 case 'f': 989 slub_debug |= SLAB_DEBUG_FREE; 990 break; 991 case 'z': 992 slub_debug |= SLAB_RED_ZONE; 993 break; 994 case 'p': 995 slub_debug |= SLAB_POISON; 996 break; 997 case 'u': 998 slub_debug |= SLAB_STORE_USER; 999 break; 1000 case 't': 1001 slub_debug |= SLAB_TRACE; 1002 break; 1003 default: 1004 printk(KERN_ERR "slub_debug option '%c' " 1005 "unknown. skipped\n", *str); 1006 } 1007 } 1008 1009check_slabs: 1010 if (*str == ',') 1011 slub_debug_slabs = str + 1; 1012out: 1013 return 1; 1014} 1015 1016__setup("slub_debug", setup_slub_debug); 1017 1018static unsigned long kmem_cache_flags(unsigned long objsize, 1019 unsigned long flags, const char *name, 1020 void (*ctor)(void *)) 1021{ 1022 /* 1023 * Enable debugging if selected on the kernel commandline. 1024 */ 1025 if (slub_debug && (!slub_debug_slabs || 1026 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0)) 1027 flags |= slub_debug; 1028 1029 return flags; 1030} 1031#else 1032static inline void setup_object_debug(struct kmem_cache *s, 1033 struct page *page, void *object) {} 1034 1035static inline int alloc_debug_processing(struct kmem_cache *s, 1036 struct page *page, void *object, unsigned long addr) { return 0; } 1037 1038static inline int free_debug_processing(struct kmem_cache *s, 1039 struct page *page, void *object, unsigned long addr) { return 0; } 1040 1041static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1042 { return 1; } 1043static inline int check_object(struct kmem_cache *s, struct page *page, 1044 void *object, int active) { return 1; } 1045static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1046static inline unsigned long kmem_cache_flags(unsigned long objsize, 1047 unsigned long flags, const char *name, 1048 void (*ctor)(void *)) 1049{ 1050 return flags; 1051} 1052#define slub_debug 0 1053 1054static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1055 { return 0; } 1056static inline void inc_slabs_node(struct kmem_cache *s, int node, 1057 int objects) {} 1058static inline void dec_slabs_node(struct kmem_cache *s, int node, 1059 int objects) {} 1060#endif 1061 1062/* 1063 * Slab allocation and freeing 1064 */ 1065static inline struct page *alloc_slab_page(gfp_t flags, int node, 1066 struct kmem_cache_order_objects oo) 1067{ 1068 int order = oo_order(oo); 1069 1070 if (node == -1) 1071 return alloc_pages(flags, order); 1072 else 1073 return alloc_pages_node(node, flags, order); 1074} 1075 1076static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1077{ 1078 struct page *page; 1079 struct kmem_cache_order_objects oo = s->oo; 1080 1081 flags |= s->allocflags; 1082 1083 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node, 1084 oo); 1085 if (unlikely(!page)) { 1086 oo = s->min; 1087 /* 1088 * Allocation may have failed due to fragmentation. 1089 * Try a lower order alloc if possible 1090 */ 1091 page = alloc_slab_page(flags, node, oo); 1092 if (!page) 1093 return NULL; 1094 1095 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK); 1096 } 1097 page->objects = oo_objects(oo); 1098 mod_zone_page_state(page_zone(page), 1099 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1100 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1101 1 << oo_order(oo)); 1102 1103 return page; 1104} 1105 1106static void setup_object(struct kmem_cache *s, struct page *page, 1107 void *object) 1108{ 1109 setup_object_debug(s, page, object); 1110 if (unlikely(s->ctor)) 1111 s->ctor(object); 1112} 1113 1114static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1115{ 1116 struct page *page; 1117 void *start; 1118 void *last; 1119 void *p; 1120 1121 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1122 1123 page = allocate_slab(s, 1124 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1125 if (!page) 1126 goto out; 1127 1128 inc_slabs_node(s, page_to_nid(page), page->objects); 1129 page->slab = s; 1130 page->flags |= 1 << PG_slab; 1131 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | 1132 SLAB_STORE_USER | SLAB_TRACE)) 1133 __SetPageSlubDebug(page); 1134 1135 start = page_address(page); 1136 1137 if (unlikely(s->flags & SLAB_POISON)) 1138 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1139 1140 last = start; 1141 for_each_object(p, s, start, page->objects) { 1142 setup_object(s, page, last); 1143 set_freepointer(s, last, p); 1144 last = p; 1145 } 1146 setup_object(s, page, last); 1147 set_freepointer(s, last, NULL); 1148 1149 page->freelist = start; 1150 page->inuse = 0; 1151out: 1152 return page; 1153} 1154 1155static void __free_slab(struct kmem_cache *s, struct page *page) 1156{ 1157 int order = compound_order(page); 1158 int pages = 1 << order; 1159 1160 if (unlikely(SLABDEBUG && PageSlubDebug(page))) { 1161 void *p; 1162 1163 slab_pad_check(s, page); 1164 for_each_object(p, s, page_address(page), 1165 page->objects) 1166 check_object(s, page, p, 0); 1167 __ClearPageSlubDebug(page); 1168 } 1169 1170 mod_zone_page_state(page_zone(page), 1171 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1172 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1173 -pages); 1174 1175 __ClearPageSlab(page); 1176 reset_page_mapcount(page); 1177 __free_pages(page, order); 1178} 1179 1180static void rcu_free_slab(struct rcu_head *h) 1181{ 1182 struct page *page; 1183 1184 page = container_of((struct list_head *)h, struct page, lru); 1185 __free_slab(page->slab, page); 1186} 1187 1188static void free_slab(struct kmem_cache *s, struct page *page) 1189{ 1190 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1191 /* 1192 * RCU free overloads the RCU head over the LRU 1193 */ 1194 struct rcu_head *head = (void *)&page->lru; 1195 1196 call_rcu(head, rcu_free_slab); 1197 } else 1198 __free_slab(s, page); 1199} 1200 1201static void discard_slab(struct kmem_cache *s, struct page *page) 1202{ 1203 dec_slabs_node(s, page_to_nid(page), page->objects); 1204 free_slab(s, page); 1205} 1206 1207/* 1208 * Per slab locking using the pagelock 1209 */ 1210static __always_inline void slab_lock(struct page *page) 1211{ 1212 bit_spin_lock(PG_locked, &page->flags); 1213} 1214 1215static __always_inline void slab_unlock(struct page *page) 1216{ 1217 __bit_spin_unlock(PG_locked, &page->flags); 1218} 1219 1220static __always_inline int slab_trylock(struct page *page) 1221{ 1222 int rc = 1; 1223 1224 rc = bit_spin_trylock(PG_locked, &page->flags); 1225 return rc; 1226} 1227 1228/* 1229 * Management of partially allocated slabs 1230 */ 1231static void add_partial(struct kmem_cache_node *n, 1232 struct page *page, int tail) 1233{ 1234 spin_lock(&n->list_lock); 1235 n->nr_partial++; 1236 if (tail) 1237 list_add_tail(&page->lru, &n->partial); 1238 else 1239 list_add(&page->lru, &n->partial); 1240 spin_unlock(&n->list_lock); 1241} 1242 1243static void remove_partial(struct kmem_cache *s, struct page *page) 1244{ 1245 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1246 1247 spin_lock(&n->list_lock); 1248 list_del(&page->lru); 1249 n->nr_partial--; 1250 spin_unlock(&n->list_lock); 1251} 1252 1253/* 1254 * Lock slab and remove from the partial list. 1255 * 1256 * Must hold list_lock. 1257 */ 1258static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1259 struct page *page) 1260{ 1261 if (slab_trylock(page)) { 1262 list_del(&page->lru); 1263 n->nr_partial--; 1264 __SetPageSlubFrozen(page); 1265 return 1; 1266 } 1267 return 0; 1268} 1269 1270/* 1271 * Try to allocate a partial slab from a specific node. 1272 */ 1273static struct page *get_partial_node(struct kmem_cache_node *n) 1274{ 1275 struct page *page; 1276 1277 /* 1278 * Racy check. If we mistakenly see no partial slabs then we 1279 * just allocate an empty slab. If we mistakenly try to get a 1280 * partial slab and there is none available then get_partials() 1281 * will return NULL. 1282 */ 1283 if (!n || !n->nr_partial) 1284 return NULL; 1285 1286 spin_lock(&n->list_lock); 1287 list_for_each_entry(page, &n->partial, lru) 1288 if (lock_and_freeze_slab(n, page)) 1289 goto out; 1290 page = NULL; 1291out: 1292 spin_unlock(&n->list_lock); 1293 return page; 1294} 1295 1296/* 1297 * Get a page from somewhere. Search in increasing NUMA distances. 1298 */ 1299static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1300{ 1301#ifdef CONFIG_NUMA 1302 struct zonelist *zonelist; 1303 struct zoneref *z; 1304 struct zone *zone; 1305 enum zone_type high_zoneidx = gfp_zone(flags); 1306 struct page *page; 1307 1308 /* 1309 * The defrag ratio allows a configuration of the tradeoffs between 1310 * inter node defragmentation and node local allocations. A lower 1311 * defrag_ratio increases the tendency to do local allocations 1312 * instead of attempting to obtain partial slabs from other nodes. 1313 * 1314 * If the defrag_ratio is set to 0 then kmalloc() always 1315 * returns node local objects. If the ratio is higher then kmalloc() 1316 * may return off node objects because partial slabs are obtained 1317 * from other nodes and filled up. 1318 * 1319 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1320 * defrag_ratio = 1000) then every (well almost) allocation will 1321 * first attempt to defrag slab caches on other nodes. This means 1322 * scanning over all nodes to look for partial slabs which may be 1323 * expensive if we do it every time we are trying to find a slab 1324 * with available objects. 1325 */ 1326 if (!s->remote_node_defrag_ratio || 1327 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1328 return NULL; 1329 1330 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1331 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1332 struct kmem_cache_node *n; 1333 1334 n = get_node(s, zone_to_nid(zone)); 1335 1336 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1337 n->nr_partial > n->min_partial) { 1338 page = get_partial_node(n); 1339 if (page) 1340 return page; 1341 } 1342 } 1343#endif 1344 return NULL; 1345} 1346 1347/* 1348 * Get a partial page, lock it and return it. 1349 */ 1350static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1351{ 1352 struct page *page; 1353 int searchnode = (node == -1) ? numa_node_id() : node; 1354 1355 page = get_partial_node(get_node(s, searchnode)); 1356 if (page || (flags & __GFP_THISNODE)) 1357 return page; 1358 1359 return get_any_partial(s, flags); 1360} 1361 1362/* 1363 * Move a page back to the lists. 1364 * 1365 * Must be called with the slab lock held. 1366 * 1367 * On exit the slab lock will have been dropped. 1368 */ 1369static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1370{ 1371 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1372 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id()); 1373 1374 __ClearPageSlubFrozen(page); 1375 if (page->inuse) { 1376 1377 if (page->freelist) { 1378 add_partial(n, page, tail); 1379 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1380 } else { 1381 stat(c, DEACTIVATE_FULL); 1382 if (SLABDEBUG && PageSlubDebug(page) && 1383 (s->flags & SLAB_STORE_USER)) 1384 add_full(n, page); 1385 } 1386 slab_unlock(page); 1387 } else { 1388 stat(c, DEACTIVATE_EMPTY); 1389 if (n->nr_partial < n->min_partial) { 1390 /* 1391 * Adding an empty slab to the partial slabs in order 1392 * to avoid page allocator overhead. This slab needs 1393 * to come after the other slabs with objects in 1394 * so that the others get filled first. That way the 1395 * size of the partial list stays small. 1396 * 1397 * kmem_cache_shrink can reclaim any empty slabs from 1398 * the partial list. 1399 */ 1400 add_partial(n, page, 1); 1401 slab_unlock(page); 1402 } else { 1403 slab_unlock(page); 1404 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB); 1405 discard_slab(s, page); 1406 } 1407 } 1408} 1409 1410/* 1411 * Remove the cpu slab 1412 */ 1413static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1414{ 1415 struct page *page = c->page; 1416 int tail = 1; 1417 1418 if (page->freelist) 1419 stat(c, DEACTIVATE_REMOTE_FREES); 1420 /* 1421 * Merge cpu freelist into slab freelist. Typically we get here 1422 * because both freelists are empty. So this is unlikely 1423 * to occur. 1424 */ 1425 while (unlikely(c->freelist)) { 1426 void **object; 1427 1428 tail = 0; /* Hot objects. Put the slab first */ 1429 1430 /* Retrieve object from cpu_freelist */ 1431 object = c->freelist; 1432 c->freelist = c->freelist[c->offset]; 1433 1434 /* And put onto the regular freelist */ 1435 object[c->offset] = page->freelist; 1436 page->freelist = object; 1437 page->inuse--; 1438 } 1439 c->page = NULL; 1440 unfreeze_slab(s, page, tail); 1441} 1442 1443static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1444{ 1445 stat(c, CPUSLAB_FLUSH); 1446 slab_lock(c->page); 1447 deactivate_slab(s, c); 1448} 1449 1450/* 1451 * Flush cpu slab. 1452 * 1453 * Called from IPI handler with interrupts disabled. 1454 */ 1455static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1456{ 1457 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1458 1459 if (likely(c && c->page)) 1460 flush_slab(s, c); 1461} 1462 1463static void flush_cpu_slab(void *d) 1464{ 1465 struct kmem_cache *s = d; 1466 1467 __flush_cpu_slab(s, smp_processor_id()); 1468} 1469 1470static void flush_all(struct kmem_cache *s) 1471{ 1472 on_each_cpu(flush_cpu_slab, s, 1); 1473} 1474 1475/* 1476 * Check if the objects in a per cpu structure fit numa 1477 * locality expectations. 1478 */ 1479static inline int node_match(struct kmem_cache_cpu *c, int node) 1480{ 1481#ifdef CONFIG_NUMA 1482 if (node != -1 && c->node != node) 1483 return 0; 1484#endif 1485 return 1; 1486} 1487 1488/* 1489 * Slow path. The lockless freelist is empty or we need to perform 1490 * debugging duties. 1491 * 1492 * Interrupts are disabled. 1493 * 1494 * Processing is still very fast if new objects have been freed to the 1495 * regular freelist. In that case we simply take over the regular freelist 1496 * as the lockless freelist and zap the regular freelist. 1497 * 1498 * If that is not working then we fall back to the partial lists. We take the 1499 * first element of the freelist as the object to allocate now and move the 1500 * rest of the freelist to the lockless freelist. 1501 * 1502 * And if we were unable to get a new slab from the partial slab lists then 1503 * we need to allocate a new slab. This is the slowest path since it involves 1504 * a call to the page allocator and the setup of a new slab. 1505 */ 1506static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 1507 unsigned long addr, struct kmem_cache_cpu *c) 1508{ 1509 void **object; 1510 struct page *new; 1511 1512 /* We handle __GFP_ZERO in the caller */ 1513 gfpflags &= ~__GFP_ZERO; 1514 1515 if (!c->page) 1516 goto new_slab; 1517 1518 slab_lock(c->page); 1519 if (unlikely(!node_match(c, node))) 1520 goto another_slab; 1521 1522 stat(c, ALLOC_REFILL); 1523 1524load_freelist: 1525 object = c->page->freelist; 1526 if (unlikely(!object)) 1527 goto another_slab; 1528 if (unlikely(SLABDEBUG && PageSlubDebug(c->page))) 1529 goto debug; 1530 1531 c->freelist = object[c->offset]; 1532 c->page->inuse = c->page->objects; 1533 c->page->freelist = NULL; 1534 c->node = page_to_nid(c->page); 1535unlock_out: 1536 slab_unlock(c->page); 1537 stat(c, ALLOC_SLOWPATH); 1538 return object; 1539 1540another_slab: 1541 deactivate_slab(s, c); 1542 1543new_slab: 1544 new = get_partial(s, gfpflags, node); 1545 if (new) { 1546 c->page = new; 1547 stat(c, ALLOC_FROM_PARTIAL); 1548 goto load_freelist; 1549 } 1550 1551 if (gfpflags & __GFP_WAIT) 1552 local_irq_enable(); 1553 1554 new = new_slab(s, gfpflags, node); 1555 1556 if (gfpflags & __GFP_WAIT) 1557 local_irq_disable(); 1558 1559 if (new) { 1560 c = get_cpu_slab(s, smp_processor_id()); 1561 stat(c, ALLOC_SLAB); 1562 if (c->page) 1563 flush_slab(s, c); 1564 slab_lock(new); 1565 __SetPageSlubFrozen(new); 1566 c->page = new; 1567 goto load_freelist; 1568 } 1569 return NULL; 1570debug: 1571 if (!alloc_debug_processing(s, c->page, object, addr)) 1572 goto another_slab; 1573 1574 c->page->inuse++; 1575 c->page->freelist = object[c->offset]; 1576 c->node = -1; 1577 goto unlock_out; 1578} 1579 1580/* 1581 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1582 * have the fastpath folded into their functions. So no function call 1583 * overhead for requests that can be satisfied on the fastpath. 1584 * 1585 * The fastpath works by first checking if the lockless freelist can be used. 1586 * If not then __slab_alloc is called for slow processing. 1587 * 1588 * Otherwise we can simply pick the next object from the lockless free list. 1589 */ 1590static __always_inline void *slab_alloc(struct kmem_cache *s, 1591 gfp_t gfpflags, int node, unsigned long addr) 1592{ 1593 void **object; 1594 struct kmem_cache_cpu *c; 1595 unsigned long flags; 1596 unsigned int objsize; 1597 1598 local_irq_save(flags); 1599 c = get_cpu_slab(s, smp_processor_id()); 1600 objsize = c->objsize; 1601 if (unlikely(!c->freelist || !node_match(c, node))) 1602 1603 object = __slab_alloc(s, gfpflags, node, addr, c); 1604 1605 else { 1606 object = c->freelist; 1607 c->freelist = object[c->offset]; 1608 stat(c, ALLOC_FASTPATH); 1609 } 1610 local_irq_restore(flags); 1611 1612 if (unlikely((gfpflags & __GFP_ZERO) && object)) 1613 memset(object, 0, objsize); 1614 1615 return object; 1616} 1617 1618void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1619{ 1620 return slab_alloc(s, gfpflags, -1, _RET_IP_); 1621} 1622EXPORT_SYMBOL(kmem_cache_alloc); 1623 1624#ifdef CONFIG_NUMA 1625void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1626{ 1627 return slab_alloc(s, gfpflags, node, _RET_IP_); 1628} 1629EXPORT_SYMBOL(kmem_cache_alloc_node); 1630#endif 1631 1632/* 1633 * Slow patch handling. This may still be called frequently since objects 1634 * have a longer lifetime than the cpu slabs in most processing loads. 1635 * 1636 * So we still attempt to reduce cache line usage. Just take the slab 1637 * lock and free the item. If there is no additional partial page 1638 * handling required then we can return immediately. 1639 */ 1640static void __slab_free(struct kmem_cache *s, struct page *page, 1641 void *x, unsigned long addr, unsigned int offset) 1642{ 1643 void *prior; 1644 void **object = (void *)x; 1645 struct kmem_cache_cpu *c; 1646 1647 c = get_cpu_slab(s, raw_smp_processor_id()); 1648 stat(c, FREE_SLOWPATH); 1649 slab_lock(page); 1650 1651 if (unlikely(SLABDEBUG && PageSlubDebug(page))) 1652 goto debug; 1653 1654checks_ok: 1655 prior = object[offset] = page->freelist; 1656 page->freelist = object; 1657 page->inuse--; 1658 1659 if (unlikely(PageSlubFrozen(page))) { 1660 stat(c, FREE_FROZEN); 1661 goto out_unlock; 1662 } 1663 1664 if (unlikely(!page->inuse)) 1665 goto slab_empty; 1666 1667 /* 1668 * Objects left in the slab. If it was not on the partial list before 1669 * then add it. 1670 */ 1671 if (unlikely(!prior)) { 1672 add_partial(get_node(s, page_to_nid(page)), page, 1); 1673 stat(c, FREE_ADD_PARTIAL); 1674 } 1675 1676out_unlock: 1677 slab_unlock(page); 1678 return; 1679 1680slab_empty: 1681 if (prior) { 1682 /* 1683 * Slab still on the partial list. 1684 */ 1685 remove_partial(s, page); 1686 stat(c, FREE_REMOVE_PARTIAL); 1687 } 1688 slab_unlock(page); 1689 stat(c, FREE_SLAB); 1690 discard_slab(s, page); 1691 return; 1692 1693debug: 1694 if (!free_debug_processing(s, page, x, addr)) 1695 goto out_unlock; 1696 goto checks_ok; 1697} 1698 1699/* 1700 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1701 * can perform fastpath freeing without additional function calls. 1702 * 1703 * The fastpath is only possible if we are freeing to the current cpu slab 1704 * of this processor. This typically the case if we have just allocated 1705 * the item before. 1706 * 1707 * If fastpath is not possible then fall back to __slab_free where we deal 1708 * with all sorts of special processing. 1709 */ 1710static __always_inline void slab_free(struct kmem_cache *s, 1711 struct page *page, void *x, unsigned long addr) 1712{ 1713 void **object = (void *)x; 1714 struct kmem_cache_cpu *c; 1715 unsigned long flags; 1716 1717 local_irq_save(flags); 1718 c = get_cpu_slab(s, smp_processor_id()); 1719 debug_check_no_locks_freed(object, c->objsize); 1720 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1721 debug_check_no_obj_freed(object, s->objsize); 1722 if (likely(page == c->page && c->node >= 0)) { 1723 object[c->offset] = c->freelist; 1724 c->freelist = object; 1725 stat(c, FREE_FASTPATH); 1726 } else 1727 __slab_free(s, page, x, addr, c->offset); 1728 1729 local_irq_restore(flags); 1730} 1731 1732void kmem_cache_free(struct kmem_cache *s, void *x) 1733{ 1734 struct page *page; 1735 1736 page = virt_to_head_page(x); 1737 1738 slab_free(s, page, x, _RET_IP_); 1739} 1740EXPORT_SYMBOL(kmem_cache_free); 1741 1742/* Figure out on which slab page the object resides */ 1743static struct page *get_object_page(const void *x) 1744{ 1745 struct page *page = virt_to_head_page(x); 1746 1747 if (!PageSlab(page)) 1748 return NULL; 1749 1750 return page; 1751} 1752 1753/* 1754 * Object placement in a slab is made very easy because we always start at 1755 * offset 0. If we tune the size of the object to the alignment then we can 1756 * get the required alignment by putting one properly sized object after 1757 * another. 1758 * 1759 * Notice that the allocation order determines the sizes of the per cpu 1760 * caches. Each processor has always one slab available for allocations. 1761 * Increasing the allocation order reduces the number of times that slabs 1762 * must be moved on and off the partial lists and is therefore a factor in 1763 * locking overhead. 1764 */ 1765 1766/* 1767 * Mininum / Maximum order of slab pages. This influences locking overhead 1768 * and slab fragmentation. A higher order reduces the number of partial slabs 1769 * and increases the number of allocations possible without having to 1770 * take the list_lock. 1771 */ 1772static int slub_min_order; 1773static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 1774static int slub_min_objects; 1775 1776/* 1777 * Merge control. If this is set then no merging of slab caches will occur. 1778 * (Could be removed. This was introduced to pacify the merge skeptics.) 1779 */ 1780static int slub_nomerge; 1781 1782/* 1783 * Calculate the order of allocation given an slab object size. 1784 * 1785 * The order of allocation has significant impact on performance and other 1786 * system components. Generally order 0 allocations should be preferred since 1787 * order 0 does not cause fragmentation in the page allocator. Larger objects 1788 * be problematic to put into order 0 slabs because there may be too much 1789 * unused space left. We go to a higher order if more than 1/16th of the slab 1790 * would be wasted. 1791 * 1792 * In order to reach satisfactory performance we must ensure that a minimum 1793 * number of objects is in one slab. Otherwise we may generate too much 1794 * activity on the partial lists which requires taking the list_lock. This is 1795 * less a concern for large slabs though which are rarely used. 1796 * 1797 * slub_max_order specifies the order where we begin to stop considering the 1798 * number of objects in a slab as critical. If we reach slub_max_order then 1799 * we try to keep the page order as low as possible. So we accept more waste 1800 * of space in favor of a small page order. 1801 * 1802 * Higher order allocations also allow the placement of more objects in a 1803 * slab and thereby reduce object handling overhead. If the user has 1804 * requested a higher mininum order then we start with that one instead of 1805 * the smallest order which will fit the object. 1806 */ 1807static inline int slab_order(int size, int min_objects, 1808 int max_order, int fract_leftover) 1809{ 1810 int order; 1811 int rem; 1812 int min_order = slub_min_order; 1813 1814 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE) 1815 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 1816 1817 for (order = max(min_order, 1818 fls(min_objects * size - 1) - PAGE_SHIFT); 1819 order <= max_order; order++) { 1820 1821 unsigned long slab_size = PAGE_SIZE << order; 1822 1823 if (slab_size < min_objects * size) 1824 continue; 1825 1826 rem = slab_size % size; 1827 1828 if (rem <= slab_size / fract_leftover) 1829 break; 1830 1831 } 1832 1833 return order; 1834} 1835 1836static inline int calculate_order(int size) 1837{ 1838 int order; 1839 int min_objects; 1840 int fraction; 1841 1842 /* 1843 * Attempt to find best configuration for a slab. This 1844 * works by first attempting to generate a layout with 1845 * the best configuration and backing off gradually. 1846 * 1847 * First we reduce the acceptable waste in a slab. Then 1848 * we reduce the minimum objects required in a slab. 1849 */ 1850 min_objects = slub_min_objects; 1851 if (!min_objects) 1852 min_objects = 4 * (fls(nr_cpu_ids) + 1); 1853 while (min_objects > 1) { 1854 fraction = 16; 1855 while (fraction >= 4) { 1856 order = slab_order(size, min_objects, 1857 slub_max_order, fraction); 1858 if (order <= slub_max_order) 1859 return order; 1860 fraction /= 2; 1861 } 1862 min_objects /= 2; 1863 } 1864 1865 /* 1866 * We were unable to place multiple objects in a slab. Now 1867 * lets see if we can place a single object there. 1868 */ 1869 order = slab_order(size, 1, slub_max_order, 1); 1870 if (order <= slub_max_order) 1871 return order; 1872 1873 /* 1874 * Doh this slab cannot be placed using slub_max_order. 1875 */ 1876 order = slab_order(size, 1, MAX_ORDER, 1); 1877 if (order <= MAX_ORDER) 1878 return order; 1879 return -ENOSYS; 1880} 1881 1882/* 1883 * Figure out what the alignment of the objects will be. 1884 */ 1885static unsigned long calculate_alignment(unsigned long flags, 1886 unsigned long align, unsigned long size) 1887{ 1888 /* 1889 * If the user wants hardware cache aligned objects then follow that 1890 * suggestion if the object is sufficiently large. 1891 * 1892 * The hardware cache alignment cannot override the specified 1893 * alignment though. If that is greater then use it. 1894 */ 1895 if (flags & SLAB_HWCACHE_ALIGN) { 1896 unsigned long ralign = cache_line_size(); 1897 while (size <= ralign / 2) 1898 ralign /= 2; 1899 align = max(align, ralign); 1900 } 1901 1902 if (align < ARCH_SLAB_MINALIGN) 1903 align = ARCH_SLAB_MINALIGN; 1904 1905 return ALIGN(align, sizeof(void *)); 1906} 1907 1908static void init_kmem_cache_cpu(struct kmem_cache *s, 1909 struct kmem_cache_cpu *c) 1910{ 1911 c->page = NULL; 1912 c->freelist = NULL; 1913 c->node = 0; 1914 c->offset = s->offset / sizeof(void *); 1915 c->objsize = s->objsize; 1916#ifdef CONFIG_SLUB_STATS 1917 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); 1918#endif 1919} 1920 1921static void 1922init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 1923{ 1924 n->nr_partial = 0; 1925 1926 /* 1927 * The larger the object size is, the more pages we want on the partial 1928 * list to avoid pounding the page allocator excessively. 1929 */ 1930 n->min_partial = ilog2(s->size); 1931 if (n->min_partial < MIN_PARTIAL) 1932 n->min_partial = MIN_PARTIAL; 1933 else if (n->min_partial > MAX_PARTIAL) 1934 n->min_partial = MAX_PARTIAL; 1935 1936 spin_lock_init(&n->list_lock); 1937 INIT_LIST_HEAD(&n->partial); 1938#ifdef CONFIG_SLUB_DEBUG 1939 atomic_long_set(&n->nr_slabs, 0); 1940 atomic_long_set(&n->total_objects, 0); 1941 INIT_LIST_HEAD(&n->full); 1942#endif 1943} 1944 1945#ifdef CONFIG_SMP 1946/* 1947 * Per cpu array for per cpu structures. 1948 * 1949 * The per cpu array places all kmem_cache_cpu structures from one processor 1950 * close together meaning that it becomes possible that multiple per cpu 1951 * structures are contained in one cacheline. This may be particularly 1952 * beneficial for the kmalloc caches. 1953 * 1954 * A desktop system typically has around 60-80 slabs. With 100 here we are 1955 * likely able to get per cpu structures for all caches from the array defined 1956 * here. We must be able to cover all kmalloc caches during bootstrap. 1957 * 1958 * If the per cpu array is exhausted then fall back to kmalloc 1959 * of individual cachelines. No sharing is possible then. 1960 */ 1961#define NR_KMEM_CACHE_CPU 100 1962 1963static DEFINE_PER_CPU(struct kmem_cache_cpu, 1964 kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; 1965 1966static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); 1967static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; 1968 1969static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, 1970 int cpu, gfp_t flags) 1971{ 1972 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); 1973 1974 if (c) 1975 per_cpu(kmem_cache_cpu_free, cpu) = 1976 (void *)c->freelist; 1977 else { 1978 /* Table overflow: So allocate ourselves */ 1979 c = kmalloc_node( 1980 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), 1981 flags, cpu_to_node(cpu)); 1982 if (!c) 1983 return NULL; 1984 } 1985 1986 init_kmem_cache_cpu(s, c); 1987 return c; 1988} 1989 1990static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) 1991{ 1992 if (c < per_cpu(kmem_cache_cpu, cpu) || 1993 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { 1994 kfree(c); 1995 return; 1996 } 1997 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); 1998 per_cpu(kmem_cache_cpu_free, cpu) = c; 1999} 2000 2001static void free_kmem_cache_cpus(struct kmem_cache *s) 2002{ 2003 int cpu; 2004 2005 for_each_online_cpu(cpu) { 2006 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 2007 2008 if (c) { 2009 s->cpu_slab[cpu] = NULL; 2010 free_kmem_cache_cpu(c, cpu); 2011 } 2012 } 2013} 2014 2015static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2016{ 2017 int cpu; 2018 2019 for_each_online_cpu(cpu) { 2020 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 2021 2022 if (c) 2023 continue; 2024 2025 c = alloc_kmem_cache_cpu(s, cpu, flags); 2026 if (!c) { 2027 free_kmem_cache_cpus(s); 2028 return 0; 2029 } 2030 s->cpu_slab[cpu] = c; 2031 } 2032 return 1; 2033} 2034 2035/* 2036 * Initialize the per cpu array. 2037 */ 2038static void init_alloc_cpu_cpu(int cpu) 2039{ 2040 int i; 2041 2042 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) 2043 return; 2044 2045 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) 2046 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); 2047 2048 cpu_set(cpu, kmem_cach_cpu_free_init_once); 2049} 2050 2051static void __init init_alloc_cpu(void) 2052{ 2053 int cpu; 2054 2055 for_each_online_cpu(cpu) 2056 init_alloc_cpu_cpu(cpu); 2057 } 2058 2059#else 2060static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} 2061static inline void init_alloc_cpu(void) {} 2062 2063static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2064{ 2065 init_kmem_cache_cpu(s, &s->cpu_slab); 2066 return 1; 2067} 2068#endif 2069 2070#ifdef CONFIG_NUMA 2071/* 2072 * No kmalloc_node yet so do it by hand. We know that this is the first 2073 * slab on the node for this slabcache. There are no concurrent accesses 2074 * possible. 2075 * 2076 * Note that this function only works on the kmalloc_node_cache 2077 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2078 * memory on a fresh node that has no slab structures yet. 2079 */ 2080static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node) 2081{ 2082 struct page *page; 2083 struct kmem_cache_node *n; 2084 unsigned long flags; 2085 2086 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2087 2088 page = new_slab(kmalloc_caches, gfpflags, node); 2089 2090 BUG_ON(!page); 2091 if (page_to_nid(page) != node) { 2092 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2093 "node %d\n", node); 2094 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2095 "in order to be able to continue\n"); 2096 } 2097 2098 n = page->freelist; 2099 BUG_ON(!n); 2100 page->freelist = get_freepointer(kmalloc_caches, n); 2101 page->inuse++; 2102 kmalloc_caches->node[node] = n; 2103#ifdef CONFIG_SLUB_DEBUG 2104 init_object(kmalloc_caches, n, 1); 2105 init_tracking(kmalloc_caches, n); 2106#endif 2107 init_kmem_cache_node(n, kmalloc_caches); 2108 inc_slabs_node(kmalloc_caches, node, page->objects); 2109 2110 /* 2111 * lockdep requires consistent irq usage for each lock 2112 * so even though there cannot be a race this early in 2113 * the boot sequence, we still disable irqs. 2114 */ 2115 local_irq_save(flags); 2116 add_partial(n, page, 0); 2117 local_irq_restore(flags); 2118} 2119 2120static void free_kmem_cache_nodes(struct kmem_cache *s) 2121{ 2122 int node; 2123 2124 for_each_node_state(node, N_NORMAL_MEMORY) { 2125 struct kmem_cache_node *n = s->node[node]; 2126 if (n && n != &s->local_node) 2127 kmem_cache_free(kmalloc_caches, n); 2128 s->node[node] = NULL; 2129 } 2130} 2131 2132static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2133{ 2134 int node; 2135 int local_node; 2136 2137 if (slab_state >= UP) 2138 local_node = page_to_nid(virt_to_page(s)); 2139 else 2140 local_node = 0; 2141 2142 for_each_node_state(node, N_NORMAL_MEMORY) { 2143 struct kmem_cache_node *n; 2144 2145 if (local_node == node) 2146 n = &s->local_node; 2147 else { 2148 if (slab_state == DOWN) { 2149 early_kmem_cache_node_alloc(gfpflags, node); 2150 continue; 2151 } 2152 n = kmem_cache_alloc_node(kmalloc_caches, 2153 gfpflags, node); 2154 2155 if (!n) { 2156 free_kmem_cache_nodes(s); 2157 return 0; 2158 } 2159 2160 } 2161 s->node[node] = n; 2162 init_kmem_cache_node(n, s); 2163 } 2164 return 1; 2165} 2166#else 2167static void free_kmem_cache_nodes(struct kmem_cache *s) 2168{ 2169} 2170 2171static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2172{ 2173 init_kmem_cache_node(&s->local_node, s); 2174 return 1; 2175} 2176#endif 2177 2178/* 2179 * calculate_sizes() determines the order and the distribution of data within 2180 * a slab object. 2181 */ 2182static int calculate_sizes(struct kmem_cache *s, int forced_order) 2183{ 2184 unsigned long flags = s->flags; 2185 unsigned long size = s->objsize; 2186 unsigned long align = s->align; 2187 int order; 2188 2189 /* 2190 * Round up object size to the next word boundary. We can only 2191 * place the free pointer at word boundaries and this determines 2192 * the possible location of the free pointer. 2193 */ 2194 size = ALIGN(size, sizeof(void *)); 2195 2196#ifdef CONFIG_SLUB_DEBUG 2197 /* 2198 * Determine if we can poison the object itself. If the user of 2199 * the slab may touch the object after free or before allocation 2200 * then we should never poison the object itself. 2201 */ 2202 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2203 !s->ctor) 2204 s->flags |= __OBJECT_POISON; 2205 else 2206 s->flags &= ~__OBJECT_POISON; 2207 2208 2209 /* 2210 * If we are Redzoning then check if there is some space between the 2211 * end of the object and the free pointer. If not then add an 2212 * additional word to have some bytes to store Redzone information. 2213 */ 2214 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2215 size += sizeof(void *); 2216#endif 2217 2218 /* 2219 * With that we have determined the number of bytes in actual use 2220 * by the object. This is the potential offset to the free pointer. 2221 */ 2222 s->inuse = size; 2223 2224 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2225 s->ctor)) { 2226 /* 2227 * Relocate free pointer after the object if it is not 2228 * permitted to overwrite the first word of the object on 2229 * kmem_cache_free. 2230 * 2231 * This is the case if we do RCU, have a constructor or 2232 * destructor or are poisoning the objects. 2233 */ 2234 s->offset = size; 2235 size += sizeof(void *); 2236 } 2237 2238#ifdef CONFIG_SLUB_DEBUG 2239 if (flags & SLAB_STORE_USER) 2240 /* 2241 * Need to store information about allocs and frees after 2242 * the object. 2243 */ 2244 size += 2 * sizeof(struct track); 2245 2246 if (flags & SLAB_RED_ZONE) 2247 /* 2248 * Add some empty padding so that we can catch 2249 * overwrites from earlier objects rather than let 2250 * tracking information or the free pointer be 2251 * corrupted if an user writes before the start 2252 * of the object. 2253 */ 2254 size += sizeof(void *); 2255#endif 2256 2257 /* 2258 * Determine the alignment based on various parameters that the 2259 * user specified and the dynamic determination of cache line size 2260 * on bootup. 2261 */ 2262 align = calculate_alignment(flags, align, s->objsize); 2263 2264 /* 2265 * SLUB stores one object immediately after another beginning from 2266 * offset 0. In order to align the objects we have to simply size 2267 * each object to conform to the alignment. 2268 */ 2269 size = ALIGN(size, align); 2270 s->size = size; 2271 if (forced_order >= 0) 2272 order = forced_order; 2273 else 2274 order = calculate_order(size); 2275 2276 if (order < 0) 2277 return 0; 2278 2279 s->allocflags = 0; 2280 if (order) 2281 s->allocflags |= __GFP_COMP; 2282 2283 if (s->flags & SLAB_CACHE_DMA) 2284 s->allocflags |= SLUB_DMA; 2285 2286 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2287 s->allocflags |= __GFP_RECLAIMABLE; 2288 2289 /* 2290 * Determine the number of objects per slab 2291 */ 2292 s->oo = oo_make(order, size); 2293 s->min = oo_make(get_order(size), size); 2294 if (oo_objects(s->oo) > oo_objects(s->max)) 2295 s->max = s->oo; 2296 2297 return !!oo_objects(s->oo); 2298 2299} 2300 2301static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2302 const char *name, size_t size, 2303 size_t align, unsigned long flags, 2304 void (*ctor)(void *)) 2305{ 2306 memset(s, 0, kmem_size); 2307 s->name = name; 2308 s->ctor = ctor; 2309 s->objsize = size; 2310 s->align = align; 2311 s->flags = kmem_cache_flags(size, flags, name, ctor); 2312 2313 if (!calculate_sizes(s, -1)) 2314 goto error; 2315 2316 s->refcount = 1; 2317#ifdef CONFIG_NUMA 2318 s->remote_node_defrag_ratio = 1000; 2319#endif 2320 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2321 goto error; 2322 2323 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2324 return 1; 2325 free_kmem_cache_nodes(s); 2326error: 2327 if (flags & SLAB_PANIC) 2328 panic("Cannot create slab %s size=%lu realsize=%u " 2329 "order=%u offset=%u flags=%lx\n", 2330 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2331 s->offset, flags); 2332 return 0; 2333} 2334 2335/* 2336 * Check if a given pointer is valid 2337 */ 2338int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2339{ 2340 struct page *page; 2341 2342 page = get_object_page(object); 2343 2344 if (!page || s != page->slab) 2345 /* No slab or wrong slab */ 2346 return 0; 2347 2348 if (!check_valid_pointer(s, page, object)) 2349 return 0; 2350 2351 /* 2352 * We could also check if the object is on the slabs freelist. 2353 * But this would be too expensive and it seems that the main 2354 * purpose of kmem_ptr_valid() is to check if the object belongs 2355 * to a certain slab. 2356 */ 2357 return 1; 2358} 2359EXPORT_SYMBOL(kmem_ptr_validate); 2360 2361/* 2362 * Determine the size of a slab object 2363 */ 2364unsigned int kmem_cache_size(struct kmem_cache *s) 2365{ 2366 return s->objsize; 2367} 2368EXPORT_SYMBOL(kmem_cache_size); 2369 2370const char *kmem_cache_name(struct kmem_cache *s) 2371{ 2372 return s->name; 2373} 2374EXPORT_SYMBOL(kmem_cache_name); 2375 2376static void list_slab_objects(struct kmem_cache *s, struct page *page, 2377 const char *text) 2378{ 2379#ifdef CONFIG_SLUB_DEBUG 2380 void *addr = page_address(page); 2381 void *p; 2382 DECLARE_BITMAP(map, page->objects); 2383 2384 bitmap_zero(map, page->objects); 2385 slab_err(s, page, "%s", text); 2386 slab_lock(page); 2387 for_each_free_object(p, s, page->freelist) 2388 set_bit(slab_index(p, s, addr), map); 2389 2390 for_each_object(p, s, addr, page->objects) { 2391 2392 if (!test_bit(slab_index(p, s, addr), map)) { 2393 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2394 p, p - addr); 2395 print_tracking(s, p); 2396 } 2397 } 2398 slab_unlock(page); 2399#endif 2400} 2401 2402/* 2403 * Attempt to free all partial slabs on a node. 2404 */ 2405static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2406{ 2407 unsigned long flags; 2408 struct page *page, *h; 2409 2410 spin_lock_irqsave(&n->list_lock, flags); 2411 list_for_each_entry_safe(page, h, &n->partial, lru) { 2412 if (!page->inuse) { 2413 list_del(&page->lru); 2414 discard_slab(s, page); 2415 n->nr_partial--; 2416 } else { 2417 list_slab_objects(s, page, 2418 "Objects remaining on kmem_cache_close()"); 2419 } 2420 } 2421 spin_unlock_irqrestore(&n->list_lock, flags); 2422} 2423 2424/* 2425 * Release all resources used by a slab cache. 2426 */ 2427static inline int kmem_cache_close(struct kmem_cache *s) 2428{ 2429 int node; 2430 2431 flush_all(s); 2432 2433 /* Attempt to free all objects */ 2434 free_kmem_cache_cpus(s); 2435 for_each_node_state(node, N_NORMAL_MEMORY) { 2436 struct kmem_cache_node *n = get_node(s, node); 2437 2438 free_partial(s, n); 2439 if (n->nr_partial || slabs_node(s, node)) 2440 return 1; 2441 } 2442 free_kmem_cache_nodes(s); 2443 return 0; 2444} 2445 2446/* 2447 * Close a cache and release the kmem_cache structure 2448 * (must be used for caches created using kmem_cache_create) 2449 */ 2450void kmem_cache_destroy(struct kmem_cache *s) 2451{ 2452 down_write(&slub_lock); 2453 s->refcount--; 2454 if (!s->refcount) { 2455 list_del(&s->list); 2456 up_write(&slub_lock); 2457 if (kmem_cache_close(s)) { 2458 printk(KERN_ERR "SLUB %s: %s called for cache that " 2459 "still has objects.\n", s->name, __func__); 2460 dump_stack(); 2461 } 2462 sysfs_slab_remove(s); 2463 } else 2464 up_write(&slub_lock); 2465} 2466EXPORT_SYMBOL(kmem_cache_destroy); 2467 2468/******************************************************************** 2469 * Kmalloc subsystem 2470 *******************************************************************/ 2471 2472struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; 2473EXPORT_SYMBOL(kmalloc_caches); 2474 2475static int __init setup_slub_min_order(char *str) 2476{ 2477 get_option(&str, &slub_min_order); 2478 2479 return 1; 2480} 2481 2482__setup("slub_min_order=", setup_slub_min_order); 2483 2484static int __init setup_slub_max_order(char *str) 2485{ 2486 get_option(&str, &slub_max_order); 2487 2488 return 1; 2489} 2490 2491__setup("slub_max_order=", setup_slub_max_order); 2492 2493static int __init setup_slub_min_objects(char *str) 2494{ 2495 get_option(&str, &slub_min_objects); 2496 2497 return 1; 2498} 2499 2500__setup("slub_min_objects=", setup_slub_min_objects); 2501 2502static int __init setup_slub_nomerge(char *str) 2503{ 2504 slub_nomerge = 1; 2505 return 1; 2506} 2507 2508__setup("slub_nomerge", setup_slub_nomerge); 2509 2510static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2511 const char *name, int size, gfp_t gfp_flags) 2512{ 2513 unsigned int flags = 0; 2514 2515 if (gfp_flags & SLUB_DMA) 2516 flags = SLAB_CACHE_DMA; 2517 2518 down_write(&slub_lock); 2519 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2520 flags, NULL)) 2521 goto panic; 2522 2523 list_add(&s->list, &slab_caches); 2524 up_write(&slub_lock); 2525 if (sysfs_slab_add(s)) 2526 goto panic; 2527 return s; 2528 2529panic: 2530 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2531} 2532 2533#ifdef CONFIG_ZONE_DMA 2534static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; 2535 2536static void sysfs_add_func(struct work_struct *w) 2537{ 2538 struct kmem_cache *s; 2539 2540 down_write(&slub_lock); 2541 list_for_each_entry(s, &slab_caches, list) { 2542 if (s->flags & __SYSFS_ADD_DEFERRED) { 2543 s->flags &= ~__SYSFS_ADD_DEFERRED; 2544 sysfs_slab_add(s); 2545 } 2546 } 2547 up_write(&slub_lock); 2548} 2549 2550static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2551 2552static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2553{ 2554 struct kmem_cache *s; 2555 char *text; 2556 size_t realsize; 2557 2558 s = kmalloc_caches_dma[index]; 2559 if (s) 2560 return s; 2561 2562 /* Dynamically create dma cache */ 2563 if (flags & __GFP_WAIT) 2564 down_write(&slub_lock); 2565 else { 2566 if (!down_write_trylock(&slub_lock)) 2567 goto out; 2568 } 2569 2570 if (kmalloc_caches_dma[index]) 2571 goto unlock_out; 2572 2573 realsize = kmalloc_caches[index].objsize; 2574 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", 2575 (unsigned int)realsize); 2576 s = kmalloc(kmem_size, flags & ~SLUB_DMA); 2577 2578 if (!s || !text || !kmem_cache_open(s, flags, text, 2579 realsize, ARCH_KMALLOC_MINALIGN, 2580 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { 2581 kfree(s); 2582 kfree(text); 2583 goto unlock_out; 2584 } 2585 2586 list_add(&s->list, &slab_caches); 2587 kmalloc_caches_dma[index] = s; 2588 2589 schedule_work(&sysfs_add_work); 2590 2591unlock_out: 2592 up_write(&slub_lock); 2593out: 2594 return kmalloc_caches_dma[index]; 2595} 2596#endif 2597 2598/* 2599 * Conversion table for small slabs sizes / 8 to the index in the 2600 * kmalloc array. This is necessary for slabs < 192 since we have non power 2601 * of two cache sizes there. The size of larger slabs can be determined using 2602 * fls. 2603 */ 2604static s8 size_index[24] = { 2605 3, /* 8 */ 2606 4, /* 16 */ 2607 5, /* 24 */ 2608 5, /* 32 */ 2609 6, /* 40 */ 2610 6, /* 48 */ 2611 6, /* 56 */ 2612 6, /* 64 */ 2613 1, /* 72 */ 2614 1, /* 80 */ 2615 1, /* 88 */ 2616 1, /* 96 */ 2617 7, /* 104 */ 2618 7, /* 112 */ 2619 7, /* 120 */ 2620 7, /* 128 */ 2621 2, /* 136 */ 2622 2, /* 144 */ 2623 2, /* 152 */ 2624 2, /* 160 */ 2625 2, /* 168 */ 2626 2, /* 176 */ 2627 2, /* 184 */ 2628 2 /* 192 */ 2629}; 2630 2631static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2632{ 2633 int index; 2634 2635 if (size <= 192) { 2636 if (!size) 2637 return ZERO_SIZE_PTR; 2638 2639 index = size_index[(size - 1) / 8]; 2640 } else 2641 index = fls(size - 1); 2642 2643#ifdef CONFIG_ZONE_DMA 2644 if (unlikely((flags & SLUB_DMA))) 2645 return dma_kmalloc_cache(index, flags); 2646 2647#endif 2648 return &kmalloc_caches[index]; 2649} 2650 2651void *__kmalloc(size_t size, gfp_t flags) 2652{ 2653 struct kmem_cache *s; 2654 2655 if (unlikely(size > PAGE_SIZE)) 2656 return kmalloc_large(size, flags); 2657 2658 s = get_slab(size, flags); 2659 2660 if (unlikely(ZERO_OR_NULL_PTR(s))) 2661 return s; 2662 2663 return slab_alloc(s, flags, -1, _RET_IP_); 2664} 2665EXPORT_SYMBOL(__kmalloc); 2666 2667static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2668{ 2669 struct page *page = alloc_pages_node(node, flags | __GFP_COMP, 2670 get_order(size)); 2671 2672 if (page) 2673 return page_address(page); 2674 else 2675 return NULL; 2676} 2677 2678#ifdef CONFIG_NUMA 2679void *__kmalloc_node(size_t size, gfp_t flags, int node) 2680{ 2681 struct kmem_cache *s; 2682 2683 if (unlikely(size > PAGE_SIZE)) 2684 return kmalloc_large_node(size, flags, node); 2685 2686 s = get_slab(size, flags); 2687 2688 if (unlikely(ZERO_OR_NULL_PTR(s))) 2689 return s; 2690 2691 return slab_alloc(s, flags, node, _RET_IP_); 2692} 2693EXPORT_SYMBOL(__kmalloc_node); 2694#endif 2695 2696size_t ksize(const void *object) 2697{ 2698 struct page *page; 2699 struct kmem_cache *s; 2700 2701 if (unlikely(object == ZERO_SIZE_PTR)) 2702 return 0; 2703 2704 page = virt_to_head_page(object); 2705 2706 if (unlikely(!PageSlab(page))) { 2707 WARN_ON(!PageCompound(page)); 2708 return PAGE_SIZE << compound_order(page); 2709 } 2710 s = page->slab; 2711 2712#ifdef CONFIG_SLUB_DEBUG 2713 /* 2714 * Debugging requires use of the padding between object 2715 * and whatever may come after it. 2716 */ 2717 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2718 return s->objsize; 2719 2720#endif 2721 /* 2722 * If we have the need to store the freelist pointer 2723 * back there or track user information then we can 2724 * only use the space before that information. 2725 */ 2726 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2727 return s->inuse; 2728 /* 2729 * Else we can use all the padding etc for the allocation 2730 */ 2731 return s->size; 2732} 2733 2734void kfree(const void *x) 2735{ 2736 struct page *page; 2737 void *object = (void *)x; 2738 2739 if (unlikely(ZERO_OR_NULL_PTR(x))) 2740 return; 2741 2742 page = virt_to_head_page(x); 2743 if (unlikely(!PageSlab(page))) { 2744 BUG_ON(!PageCompound(page)); 2745 put_page(page); 2746 return; 2747 } 2748 slab_free(page->slab, page, object, _RET_IP_); 2749} 2750EXPORT_SYMBOL(kfree); 2751 2752/* 2753 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2754 * the remaining slabs by the number of items in use. The slabs with the 2755 * most items in use come first. New allocations will then fill those up 2756 * and thus they can be removed from the partial lists. 2757 * 2758 * The slabs with the least items are placed last. This results in them 2759 * being allocated from last increasing the chance that the last objects 2760 * are freed in them. 2761 */ 2762int kmem_cache_shrink(struct kmem_cache *s) 2763{ 2764 int node; 2765 int i; 2766 struct kmem_cache_node *n; 2767 struct page *page; 2768 struct page *t; 2769 int objects = oo_objects(s->max); 2770 struct list_head *slabs_by_inuse = 2771 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 2772 unsigned long flags; 2773 2774 if (!slabs_by_inuse) 2775 return -ENOMEM; 2776 2777 flush_all(s); 2778 for_each_node_state(node, N_NORMAL_MEMORY) { 2779 n = get_node(s, node); 2780 2781 if (!n->nr_partial) 2782 continue; 2783 2784 for (i = 0; i < objects; i++) 2785 INIT_LIST_HEAD(slabs_by_inuse + i); 2786 2787 spin_lock_irqsave(&n->list_lock, flags); 2788 2789 /* 2790 * Build lists indexed by the items in use in each slab. 2791 * 2792 * Note that concurrent frees may occur while we hold the 2793 * list_lock. page->inuse here is the upper limit. 2794 */ 2795 list_for_each_entry_safe(page, t, &n->partial, lru) { 2796 if (!page->inuse && slab_trylock(page)) { 2797 /* 2798 * Must hold slab lock here because slab_free 2799 * may have freed the last object and be 2800 * waiting to release the slab. 2801 */ 2802 list_del(&page->lru); 2803 n->nr_partial--; 2804 slab_unlock(page); 2805 discard_slab(s, page); 2806 } else { 2807 list_move(&page->lru, 2808 slabs_by_inuse + page->inuse); 2809 } 2810 } 2811 2812 /* 2813 * Rebuild the partial list with the slabs filled up most 2814 * first and the least used slabs at the end. 2815 */ 2816 for (i = objects - 1; i >= 0; i--) 2817 list_splice(slabs_by_inuse + i, n->partial.prev); 2818 2819 spin_unlock_irqrestore(&n->list_lock, flags); 2820 } 2821 2822 kfree(slabs_by_inuse); 2823 return 0; 2824} 2825EXPORT_SYMBOL(kmem_cache_shrink); 2826 2827#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2828static int slab_mem_going_offline_callback(void *arg) 2829{ 2830 struct kmem_cache *s; 2831 2832 down_read(&slub_lock); 2833 list_for_each_entry(s, &slab_caches, list) 2834 kmem_cache_shrink(s); 2835 up_read(&slub_lock); 2836 2837 return 0; 2838} 2839 2840static void slab_mem_offline_callback(void *arg) 2841{ 2842 struct kmem_cache_node *n; 2843 struct kmem_cache *s; 2844 struct memory_notify *marg = arg; 2845 int offline_node; 2846 2847 offline_node = marg->status_change_nid; 2848 2849 /* 2850 * If the node still has available memory. we need kmem_cache_node 2851 * for it yet. 2852 */ 2853 if (offline_node < 0) 2854 return; 2855 2856 down_read(&slub_lock); 2857 list_for_each_entry(s, &slab_caches, list) { 2858 n = get_node(s, offline_node); 2859 if (n) { 2860 /* 2861 * if n->nr_slabs > 0, slabs still exist on the node 2862 * that is going down. We were unable to free them, 2863 * and offline_pages() function shoudn't call this 2864 * callback. So, we must fail. 2865 */ 2866 BUG_ON(slabs_node(s, offline_node)); 2867 2868 s->node[offline_node] = NULL; 2869 kmem_cache_free(kmalloc_caches, n); 2870 } 2871 } 2872 up_read(&slub_lock); 2873} 2874 2875static int slab_mem_going_online_callback(void *arg) 2876{ 2877 struct kmem_cache_node *n; 2878 struct kmem_cache *s; 2879 struct memory_notify *marg = arg; 2880 int nid = marg->status_change_nid; 2881 int ret = 0; 2882 2883 /* 2884 * If the node's memory is already available, then kmem_cache_node is 2885 * already created. Nothing to do. 2886 */ 2887 if (nid < 0) 2888 return 0; 2889 2890 /* 2891 * We are bringing a node online. No memory is available yet. We must 2892 * allocate a kmem_cache_node structure in order to bring the node 2893 * online. 2894 */ 2895 down_read(&slub_lock); 2896 list_for_each_entry(s, &slab_caches, list) { 2897 /* 2898 * XXX: kmem_cache_alloc_node will fallback to other nodes 2899 * since memory is not yet available from the node that 2900 * is brought up. 2901 */ 2902 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 2903 if (!n) { 2904 ret = -ENOMEM; 2905 goto out; 2906 } 2907 init_kmem_cache_node(n, s); 2908 s->node[nid] = n; 2909 } 2910out: 2911 up_read(&slub_lock); 2912 return ret; 2913} 2914 2915static int slab_memory_callback(struct notifier_block *self, 2916 unsigned long action, void *arg) 2917{ 2918 int ret = 0; 2919 2920 switch (action) { 2921 case MEM_GOING_ONLINE: 2922 ret = slab_mem_going_online_callback(arg); 2923 break; 2924 case MEM_GOING_OFFLINE: 2925 ret = slab_mem_going_offline_callback(arg); 2926 break; 2927 case MEM_OFFLINE: 2928 case MEM_CANCEL_ONLINE: 2929 slab_mem_offline_callback(arg); 2930 break; 2931 case MEM_ONLINE: 2932 case MEM_CANCEL_OFFLINE: 2933 break; 2934 } 2935 2936 ret = notifier_from_errno(ret); 2937 return ret; 2938} 2939 2940#endif /* CONFIG_MEMORY_HOTPLUG */ 2941 2942/******************************************************************** 2943 * Basic setup of slabs 2944 *******************************************************************/ 2945 2946void __init kmem_cache_init(void) 2947{ 2948 int i; 2949 int caches = 0; 2950 2951 init_alloc_cpu(); 2952 2953#ifdef CONFIG_NUMA 2954 /* 2955 * Must first have the slab cache available for the allocations of the 2956 * struct kmem_cache_node's. There is special bootstrap code in 2957 * kmem_cache_open for slab_state == DOWN. 2958 */ 2959 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 2960 sizeof(struct kmem_cache_node), GFP_KERNEL); 2961 kmalloc_caches[0].refcount = -1; 2962 caches++; 2963 2964 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 2965#endif 2966 2967 /* Able to allocate the per node structures */ 2968 slab_state = PARTIAL; 2969 2970 /* Caches that are not of the two-to-the-power-of size */ 2971 if (KMALLOC_MIN_SIZE <= 64) { 2972 create_kmalloc_cache(&kmalloc_caches[1], 2973 "kmalloc-96", 96, GFP_KERNEL); 2974 caches++; 2975 create_kmalloc_cache(&kmalloc_caches[2], 2976 "kmalloc-192", 192, GFP_KERNEL); 2977 caches++; 2978 } 2979 2980 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { 2981 create_kmalloc_cache(&kmalloc_caches[i], 2982 "kmalloc", 1 << i, GFP_KERNEL); 2983 caches++; 2984 } 2985 2986 2987 /* 2988 * Patch up the size_index table if we have strange large alignment 2989 * requirements for the kmalloc array. This is only the case for 2990 * MIPS it seems. The standard arches will not generate any code here. 2991 * 2992 * Largest permitted alignment is 256 bytes due to the way we 2993 * handle the index determination for the smaller caches. 2994 * 2995 * Make sure that nothing crazy happens if someone starts tinkering 2996 * around with ARCH_KMALLOC_MINALIGN 2997 */ 2998 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 2999 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3000 3001 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) 3002 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; 3003 3004 if (KMALLOC_MIN_SIZE == 128) { 3005 /* 3006 * The 192 byte sized cache is not used if the alignment 3007 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3008 * instead. 3009 */ 3010 for (i = 128 + 8; i <= 192; i += 8) 3011 size_index[(i - 1) / 8] = 8; 3012 } 3013 3014 slab_state = UP; 3015 3016 /* Provide the correct kmalloc names now that the caches are up */ 3017 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) 3018 kmalloc_caches[i]. name = 3019 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); 3020 3021#ifdef CONFIG_SMP 3022 register_cpu_notifier(&slab_notifier); 3023 kmem_size = offsetof(struct kmem_cache, cpu_slab) + 3024 nr_cpu_ids * sizeof(struct kmem_cache_cpu *); 3025#else 3026 kmem_size = sizeof(struct kmem_cache); 3027#endif 3028 3029 printk(KERN_INFO 3030 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3031 " CPUs=%d, Nodes=%d\n", 3032 caches, cache_line_size(), 3033 slub_min_order, slub_max_order, slub_min_objects, 3034 nr_cpu_ids, nr_node_ids); 3035} 3036 3037/* 3038 * Find a mergeable slab cache 3039 */ 3040static int slab_unmergeable(struct kmem_cache *s) 3041{ 3042 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3043 return 1; 3044 3045 if (s->ctor) 3046 return 1; 3047 3048 /* 3049 * We may have set a slab to be unmergeable during bootstrap. 3050 */ 3051 if (s->refcount < 0) 3052 return 1; 3053 3054 return 0; 3055} 3056 3057static struct kmem_cache *find_mergeable(size_t size, 3058 size_t align, unsigned long flags, const char *name, 3059 void (*ctor)(void *)) 3060{ 3061 struct kmem_cache *s; 3062 3063 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3064 return NULL; 3065 3066 if (ctor) 3067 return NULL; 3068 3069 size = ALIGN(size, sizeof(void *)); 3070 align = calculate_alignment(flags, align, size); 3071 size = ALIGN(size, align); 3072 flags = kmem_cache_flags(size, flags, name, NULL); 3073 3074 list_for_each_entry(s, &slab_caches, list) { 3075 if (slab_unmergeable(s)) 3076 continue; 3077 3078 if (size > s->size) 3079 continue; 3080 3081 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3082 continue; 3083 /* 3084 * Check if alignment is compatible. 3085 * Courtesy of Adrian Drzewiecki 3086 */ 3087 if ((s->size & ~(align - 1)) != s->size) 3088 continue; 3089 3090 if (s->size - size >= sizeof(void *)) 3091 continue; 3092 3093 return s; 3094 } 3095 return NULL; 3096} 3097 3098struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3099 size_t align, unsigned long flags, void (*ctor)(void *)) 3100{ 3101 struct kmem_cache *s; 3102 3103 down_write(&slub_lock); 3104 s = find_mergeable(size, align, flags, name, ctor); 3105 if (s) { 3106 int cpu; 3107 3108 s->refcount++; 3109 /* 3110 * Adjust the object sizes so that we clear 3111 * the complete object on kzalloc. 3112 */ 3113 s->objsize = max(s->objsize, (int)size); 3114 3115 /* 3116 * And then we need to update the object size in the 3117 * per cpu structures 3118 */ 3119 for_each_online_cpu(cpu) 3120 get_cpu_slab(s, cpu)->objsize = s->objsize; 3121 3122 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3123 up_write(&slub_lock); 3124 3125 if (sysfs_slab_alias(s, name)) 3126 goto err; 3127 return s; 3128 } 3129 3130 s = kmalloc(kmem_size, GFP_KERNEL); 3131 if (s) { 3132 if (kmem_cache_open(s, GFP_KERNEL, name, 3133 size, align, flags, ctor)) { 3134 list_add(&s->list, &slab_caches); 3135 up_write(&slub_lock); 3136 if (sysfs_slab_add(s)) 3137 goto err; 3138 return s; 3139 } 3140 kfree(s); 3141 } 3142 up_write(&slub_lock); 3143 3144err: 3145 if (flags & SLAB_PANIC) 3146 panic("Cannot create slabcache %s\n", name); 3147 else 3148 s = NULL; 3149 return s; 3150} 3151EXPORT_SYMBOL(kmem_cache_create); 3152 3153#ifdef CONFIG_SMP 3154/* 3155 * Use the cpu notifier to insure that the cpu slabs are flushed when 3156 * necessary. 3157 */ 3158static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3159 unsigned long action, void *hcpu) 3160{ 3161 long cpu = (long)hcpu; 3162 struct kmem_cache *s; 3163 unsigned long flags; 3164 3165 switch (action) { 3166 case CPU_UP_PREPARE: 3167 case CPU_UP_PREPARE_FROZEN: 3168 init_alloc_cpu_cpu(cpu); 3169 down_read(&slub_lock); 3170 list_for_each_entry(s, &slab_caches, list) 3171 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, 3172 GFP_KERNEL); 3173 up_read(&slub_lock); 3174 break; 3175 3176 case CPU_UP_CANCELED: 3177 case CPU_UP_CANCELED_FROZEN: 3178 case CPU_DEAD: 3179 case CPU_DEAD_FROZEN: 3180 down_read(&slub_lock); 3181 list_for_each_entry(s, &slab_caches, list) { 3182 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3183 3184 local_irq_save(flags); 3185 __flush_cpu_slab(s, cpu); 3186 local_irq_restore(flags); 3187 free_kmem_cache_cpu(c, cpu); 3188 s->cpu_slab[cpu] = NULL; 3189 } 3190 up_read(&slub_lock); 3191 break; 3192 default: 3193 break; 3194 } 3195 return NOTIFY_OK; 3196} 3197 3198static struct notifier_block __cpuinitdata slab_notifier = { 3199 .notifier_call = slab_cpuup_callback 3200}; 3201 3202#endif 3203 3204void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3205{ 3206 struct kmem_cache *s; 3207 3208 if (unlikely(size > PAGE_SIZE)) 3209 return kmalloc_large(size, gfpflags); 3210 3211 s = get_slab(size, gfpflags); 3212 3213 if (unlikely(ZERO_OR_NULL_PTR(s))) 3214 return s; 3215 3216 return slab_alloc(s, gfpflags, -1, caller); 3217} 3218 3219void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3220 int node, unsigned long caller) 3221{ 3222 struct kmem_cache *s; 3223 3224 if (unlikely(size > PAGE_SIZE)) 3225 return kmalloc_large_node(size, gfpflags, node); 3226 3227 s = get_slab(size, gfpflags); 3228 3229 if (unlikely(ZERO_OR_NULL_PTR(s))) 3230 return s; 3231 3232 return slab_alloc(s, gfpflags, node, caller); 3233} 3234 3235#ifdef CONFIG_SLUB_DEBUG 3236static unsigned long count_partial(struct kmem_cache_node *n, 3237 int (*get_count)(struct page *)) 3238{ 3239 unsigned long flags; 3240 unsigned long x = 0; 3241 struct page *page; 3242 3243 spin_lock_irqsave(&n->list_lock, flags); 3244 list_for_each_entry(page, &n->partial, lru) 3245 x += get_count(page); 3246 spin_unlock_irqrestore(&n->list_lock, flags); 3247 return x; 3248} 3249 3250static int count_inuse(struct page *page) 3251{ 3252 return page->inuse; 3253} 3254 3255static int count_total(struct page *page) 3256{ 3257 return page->objects; 3258} 3259 3260static int count_free(struct page *page) 3261{ 3262 return page->objects - page->inuse; 3263} 3264 3265static int validate_slab(struct kmem_cache *s, struct page *page, 3266 unsigned long *map) 3267{ 3268 void *p; 3269 void *addr = page_address(page); 3270 3271 if (!check_slab(s, page) || 3272 !on_freelist(s, page, NULL)) 3273 return 0; 3274 3275 /* Now we know that a valid freelist exists */ 3276 bitmap_zero(map, page->objects); 3277 3278 for_each_free_object(p, s, page->freelist) { 3279 set_bit(slab_index(p, s, addr), map); 3280 if (!check_object(s, page, p, 0)) 3281 return 0; 3282 } 3283 3284 for_each_object(p, s, addr, page->objects) 3285 if (!test_bit(slab_index(p, s, addr), map)) 3286 if (!check_object(s, page, p, 1)) 3287 return 0; 3288 return 1; 3289} 3290 3291static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3292 unsigned long *map) 3293{ 3294 if (slab_trylock(page)) { 3295 validate_slab(s, page, map); 3296 slab_unlock(page); 3297 } else 3298 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3299 s->name, page); 3300 3301 if (s->flags & DEBUG_DEFAULT_FLAGS) { 3302 if (!PageSlubDebug(page)) 3303 printk(KERN_ERR "SLUB %s: SlubDebug not set " 3304 "on slab 0x%p\n", s->name, page); 3305 } else { 3306 if (PageSlubDebug(page)) 3307 printk(KERN_ERR "SLUB %s: SlubDebug set on " 3308 "slab 0x%p\n", s->name, page); 3309 } 3310} 3311 3312static int validate_slab_node(struct kmem_cache *s, 3313 struct kmem_cache_node *n, unsigned long *map) 3314{ 3315 unsigned long count = 0; 3316 struct page *page; 3317 unsigned long flags; 3318 3319 spin_lock_irqsave(&n->list_lock, flags); 3320 3321 list_for_each_entry(page, &n->partial, lru) { 3322 validate_slab_slab(s, page, map); 3323 count++; 3324 } 3325 if (count != n->nr_partial) 3326 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3327 "counter=%ld\n", s->name, count, n->nr_partial); 3328 3329 if (!(s->flags & SLAB_STORE_USER)) 3330 goto out; 3331 3332 list_for_each_entry(page, &n->full, lru) { 3333 validate_slab_slab(s, page, map); 3334 count++; 3335 } 3336 if (count != atomic_long_read(&n->nr_slabs)) 3337 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3338 "counter=%ld\n", s->name, count, 3339 atomic_long_read(&n->nr_slabs)); 3340 3341out: 3342 spin_unlock_irqrestore(&n->list_lock, flags); 3343 return count; 3344} 3345 3346static long validate_slab_cache(struct kmem_cache *s) 3347{ 3348 int node; 3349 unsigned long count = 0; 3350 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3351 sizeof(unsigned long), GFP_KERNEL); 3352 3353 if (!map) 3354 return -ENOMEM; 3355 3356 flush_all(s); 3357 for_each_node_state(node, N_NORMAL_MEMORY) { 3358 struct kmem_cache_node *n = get_node(s, node); 3359 3360 count += validate_slab_node(s, n, map); 3361 } 3362 kfree(map); 3363 return count; 3364} 3365 3366#ifdef SLUB_RESILIENCY_TEST 3367static void resiliency_test(void) 3368{ 3369 u8 *p; 3370 3371 printk(KERN_ERR "SLUB resiliency testing\n"); 3372 printk(KERN_ERR "-----------------------\n"); 3373 printk(KERN_ERR "A. Corruption after allocation\n"); 3374 3375 p = kzalloc(16, GFP_KERNEL); 3376 p[16] = 0x12; 3377 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3378 " 0x12->0x%p\n\n", p + 16); 3379 3380 validate_slab_cache(kmalloc_caches + 4); 3381 3382 /* Hmmm... The next two are dangerous */ 3383 p = kzalloc(32, GFP_KERNEL); 3384 p[32 + sizeof(void *)] = 0x34; 3385 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3386 " 0x34 -> -0x%p\n", p); 3387 printk(KERN_ERR 3388 "If allocated object is overwritten then not detectable\n\n"); 3389 3390 validate_slab_cache(kmalloc_caches + 5); 3391 p = kzalloc(64, GFP_KERNEL); 3392 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3393 *p = 0x56; 3394 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3395 p); 3396 printk(KERN_ERR 3397 "If allocated object is overwritten then not detectable\n\n"); 3398 validate_slab_cache(kmalloc_caches + 6); 3399 3400 printk(KERN_ERR "\nB. Corruption after free\n"); 3401 p = kzalloc(128, GFP_KERNEL); 3402 kfree(p); 3403 *p = 0x78; 3404 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3405 validate_slab_cache(kmalloc_caches + 7); 3406 3407 p = kzalloc(256, GFP_KERNEL); 3408 kfree(p); 3409 p[50] = 0x9a; 3410 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3411 p); 3412 validate_slab_cache(kmalloc_caches + 8); 3413 3414 p = kzalloc(512, GFP_KERNEL); 3415 kfree(p); 3416 p[512] = 0xab; 3417 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3418 validate_slab_cache(kmalloc_caches + 9); 3419} 3420#else 3421static void resiliency_test(void) {}; 3422#endif 3423 3424/* 3425 * Generate lists of code addresses where slabcache objects are allocated 3426 * and freed. 3427 */ 3428 3429struct location { 3430 unsigned long count; 3431 unsigned long addr; 3432 long long sum_time; 3433 long min_time; 3434 long max_time; 3435 long min_pid; 3436 long max_pid; 3437 cpumask_t cpus; 3438 nodemask_t nodes; 3439}; 3440 3441struct loc_track { 3442 unsigned long max; 3443 unsigned long count; 3444 struct location *loc; 3445}; 3446 3447static void free_loc_track(struct loc_track *t) 3448{ 3449 if (t->max) 3450 free_pages((unsigned long)t->loc, 3451 get_order(sizeof(struct location) * t->max)); 3452} 3453 3454static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3455{ 3456 struct location *l; 3457 int order; 3458 3459 order = get_order(sizeof(struct location) * max); 3460 3461 l = (void *)__get_free_pages(flags, order); 3462 if (!l) 3463 return 0; 3464 3465 if (t->count) { 3466 memcpy(l, t->loc, sizeof(struct location) * t->count); 3467 free_loc_track(t); 3468 } 3469 t->max = max; 3470 t->loc = l; 3471 return 1; 3472} 3473 3474static int add_location(struct loc_track *t, struct kmem_cache *s, 3475 const struct track *track) 3476{ 3477 long start, end, pos; 3478 struct location *l; 3479 unsigned long caddr; 3480 unsigned long age = jiffies - track->when; 3481 3482 start = -1; 3483 end = t->count; 3484 3485 for ( ; ; ) { 3486 pos = start + (end - start + 1) / 2; 3487 3488 /* 3489 * There is nothing at "end". If we end up there 3490 * we need to add something to before end. 3491 */ 3492 if (pos == end) 3493 break; 3494 3495 caddr = t->loc[pos].addr; 3496 if (track->addr == caddr) { 3497 3498 l = &t->loc[pos]; 3499 l->count++; 3500 if (track->when) { 3501 l->sum_time += age; 3502 if (age < l->min_time) 3503 l->min_time = age; 3504 if (age > l->max_time) 3505 l->max_time = age; 3506 3507 if (track->pid < l->min_pid) 3508 l->min_pid = track->pid; 3509 if (track->pid > l->max_pid) 3510 l->max_pid = track->pid; 3511 3512 cpu_set(track->cpu, l->cpus); 3513 } 3514 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3515 return 1; 3516 } 3517 3518 if (track->addr < caddr) 3519 end = pos; 3520 else 3521 start = pos; 3522 } 3523 3524 /* 3525 * Not found. Insert new tracking element. 3526 */ 3527 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3528 return 0; 3529 3530 l = t->loc + pos; 3531 if (pos < t->count) 3532 memmove(l + 1, l, 3533 (t->count - pos) * sizeof(struct location)); 3534 t->count++; 3535 l->count = 1; 3536 l->addr = track->addr; 3537 l->sum_time = age; 3538 l->min_time = age; 3539 l->max_time = age; 3540 l->min_pid = track->pid; 3541 l->max_pid = track->pid; 3542 cpus_clear(l->cpus); 3543 cpu_set(track->cpu, l->cpus); 3544 nodes_clear(l->nodes); 3545 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3546 return 1; 3547} 3548 3549static void process_slab(struct loc_track *t, struct kmem_cache *s, 3550 struct page *page, enum track_item alloc) 3551{ 3552 void *addr = page_address(page); 3553 DECLARE_BITMAP(map, page->objects); 3554 void *p; 3555 3556 bitmap_zero(map, page->objects); 3557 for_each_free_object(p, s, page->freelist) 3558 set_bit(slab_index(p, s, addr), map); 3559 3560 for_each_object(p, s, addr, page->objects) 3561 if (!test_bit(slab_index(p, s, addr), map)) 3562 add_location(t, s, get_track(s, p, alloc)); 3563} 3564 3565static int list_locations(struct kmem_cache *s, char *buf, 3566 enum track_item alloc) 3567{ 3568 int len = 0; 3569 unsigned long i; 3570 struct loc_track t = { 0, 0, NULL }; 3571 int node; 3572 3573 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3574 GFP_TEMPORARY)) 3575 return sprintf(buf, "Out of memory\n"); 3576 3577 /* Push back cpu slabs */ 3578 flush_all(s); 3579 3580 for_each_node_state(node, N_NORMAL_MEMORY) { 3581 struct kmem_cache_node *n = get_node(s, node); 3582 unsigned long flags; 3583 struct page *page; 3584 3585 if (!atomic_long_read(&n->nr_slabs)) 3586 continue; 3587 3588 spin_lock_irqsave(&n->list_lock, flags); 3589 list_for_each_entry(page, &n->partial, lru) 3590 process_slab(&t, s, page, alloc); 3591 list_for_each_entry(page, &n->full, lru) 3592 process_slab(&t, s, page, alloc); 3593 spin_unlock_irqrestore(&n->list_lock, flags); 3594 } 3595 3596 for (i = 0; i < t.count; i++) { 3597 struct location *l = &t.loc[i]; 3598 3599 if (len > PAGE_SIZE - 100) 3600 break; 3601 len += sprintf(buf + len, "%7ld ", l->count); 3602 3603 if (l->addr) 3604 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3605 else 3606 len += sprintf(buf + len, "<not-available>"); 3607 3608 if (l->sum_time != l->min_time) { 3609 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3610 l->min_time, 3611 (long)div_u64(l->sum_time, l->count), 3612 l->max_time); 3613 } else 3614 len += sprintf(buf + len, " age=%ld", 3615 l->min_time); 3616 3617 if (l->min_pid != l->max_pid) 3618 len += sprintf(buf + len, " pid=%ld-%ld", 3619 l->min_pid, l->max_pid); 3620 else 3621 len += sprintf(buf + len, " pid=%ld", 3622 l->min_pid); 3623 3624 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && 3625 len < PAGE_SIZE - 60) { 3626 len += sprintf(buf + len, " cpus="); 3627 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3628 l->cpus); 3629 } 3630 3631 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && 3632 len < PAGE_SIZE - 60) { 3633 len += sprintf(buf + len, " nodes="); 3634 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3635 l->nodes); 3636 } 3637 3638 len += sprintf(buf + len, "\n"); 3639 } 3640 3641 free_loc_track(&t); 3642 if (!t.count) 3643 len += sprintf(buf, "No data\n"); 3644 return len; 3645} 3646 3647enum slab_stat_type { 3648 SL_ALL, /* All slabs */ 3649 SL_PARTIAL, /* Only partially allocated slabs */ 3650 SL_CPU, /* Only slabs used for cpu caches */ 3651 SL_OBJECTS, /* Determine allocated objects not slabs */ 3652 SL_TOTAL /* Determine object capacity not slabs */ 3653}; 3654 3655#define SO_ALL (1 << SL_ALL) 3656#define SO_PARTIAL (1 << SL_PARTIAL) 3657#define SO_CPU (1 << SL_CPU) 3658#define SO_OBJECTS (1 << SL_OBJECTS) 3659#define SO_TOTAL (1 << SL_TOTAL) 3660 3661static ssize_t show_slab_objects(struct kmem_cache *s, 3662 char *buf, unsigned long flags) 3663{ 3664 unsigned long total = 0; 3665 int node; 3666 int x; 3667 unsigned long *nodes; 3668 unsigned long *per_cpu; 3669 3670 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3671 if (!nodes) 3672 return -ENOMEM; 3673 per_cpu = nodes + nr_node_ids; 3674 3675 if (flags & SO_CPU) { 3676 int cpu; 3677 3678 for_each_possible_cpu(cpu) { 3679 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3680 3681 if (!c || c->node < 0) 3682 continue; 3683 3684 if (c->page) { 3685 if (flags & SO_TOTAL) 3686 x = c->page->objects; 3687 else if (flags & SO_OBJECTS) 3688 x = c->page->inuse; 3689 else 3690 x = 1; 3691 3692 total += x; 3693 nodes[c->node] += x; 3694 } 3695 per_cpu[c->node]++; 3696 } 3697 } 3698 3699 if (flags & SO_ALL) { 3700 for_each_node_state(node, N_NORMAL_MEMORY) { 3701 struct kmem_cache_node *n = get_node(s, node); 3702 3703 if (flags & SO_TOTAL) 3704 x = atomic_long_read(&n->total_objects); 3705 else if (flags & SO_OBJECTS) 3706 x = atomic_long_read(&n->total_objects) - 3707 count_partial(n, count_free); 3708 3709 else 3710 x = atomic_long_read(&n->nr_slabs); 3711 total += x; 3712 nodes[node] += x; 3713 } 3714 3715 } else if (flags & SO_PARTIAL) { 3716 for_each_node_state(node, N_NORMAL_MEMORY) { 3717 struct kmem_cache_node *n = get_node(s, node); 3718 3719 if (flags & SO_TOTAL) 3720 x = count_partial(n, count_total); 3721 else if (flags & SO_OBJECTS) 3722 x = count_partial(n, count_inuse); 3723 else 3724 x = n->nr_partial; 3725 total += x; 3726 nodes[node] += x; 3727 } 3728 } 3729 x = sprintf(buf, "%lu", total); 3730#ifdef CONFIG_NUMA 3731 for_each_node_state(node, N_NORMAL_MEMORY) 3732 if (nodes[node]) 3733 x += sprintf(buf + x, " N%d=%lu", 3734 node, nodes[node]); 3735#endif 3736 kfree(nodes); 3737 return x + sprintf(buf + x, "\n"); 3738} 3739 3740static int any_slab_objects(struct kmem_cache *s) 3741{ 3742 int node; 3743 3744 for_each_online_node(node) { 3745 struct kmem_cache_node *n = get_node(s, node); 3746 3747 if (!n) 3748 continue; 3749 3750 if (atomic_long_read(&n->total_objects)) 3751 return 1; 3752 } 3753 return 0; 3754} 3755 3756#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3757#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3758 3759struct slab_attribute { 3760 struct attribute attr; 3761 ssize_t (*show)(struct kmem_cache *s, char *buf); 3762 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3763}; 3764 3765#define SLAB_ATTR_RO(_name) \ 3766 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3767 3768#define SLAB_ATTR(_name) \ 3769 static struct slab_attribute _name##_attr = \ 3770 __ATTR(_name, 0644, _name##_show, _name##_store) 3771 3772static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3773{ 3774 return sprintf(buf, "%d\n", s->size); 3775} 3776SLAB_ATTR_RO(slab_size); 3777 3778static ssize_t align_show(struct kmem_cache *s, char *buf) 3779{ 3780 return sprintf(buf, "%d\n", s->align); 3781} 3782SLAB_ATTR_RO(align); 3783 3784static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3785{ 3786 return sprintf(buf, "%d\n", s->objsize); 3787} 3788SLAB_ATTR_RO(object_size); 3789 3790static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3791{ 3792 return sprintf(buf, "%d\n", oo_objects(s->oo)); 3793} 3794SLAB_ATTR_RO(objs_per_slab); 3795 3796static ssize_t order_store(struct kmem_cache *s, 3797 const char *buf, size_t length) 3798{ 3799 unsigned long order; 3800 int err; 3801 3802 err = strict_strtoul(buf, 10, &order); 3803 if (err) 3804 return err; 3805 3806 if (order > slub_max_order || order < slub_min_order) 3807 return -EINVAL; 3808 3809 calculate_sizes(s, order); 3810 return length; 3811} 3812 3813static ssize_t order_show(struct kmem_cache *s, char *buf) 3814{ 3815 return sprintf(buf, "%d\n", oo_order(s->oo)); 3816} 3817SLAB_ATTR(order); 3818 3819static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3820{ 3821 if (s->ctor) { 3822 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3823 3824 return n + sprintf(buf + n, "\n"); 3825 } 3826 return 0; 3827} 3828SLAB_ATTR_RO(ctor); 3829 3830static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3831{ 3832 return sprintf(buf, "%d\n", s->refcount - 1); 3833} 3834SLAB_ATTR_RO(aliases); 3835 3836static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3837{ 3838 return show_slab_objects(s, buf, SO_ALL); 3839} 3840SLAB_ATTR_RO(slabs); 3841 3842static ssize_t partial_show(struct kmem_cache *s, char *buf) 3843{ 3844 return show_slab_objects(s, buf, SO_PARTIAL); 3845} 3846SLAB_ATTR_RO(partial); 3847 3848static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 3849{ 3850 return show_slab_objects(s, buf, SO_CPU); 3851} 3852SLAB_ATTR_RO(cpu_slabs); 3853 3854static ssize_t objects_show(struct kmem_cache *s, char *buf) 3855{ 3856 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 3857} 3858SLAB_ATTR_RO(objects); 3859 3860static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 3861{ 3862 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 3863} 3864SLAB_ATTR_RO(objects_partial); 3865 3866static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 3867{ 3868 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 3869} 3870SLAB_ATTR_RO(total_objects); 3871 3872static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 3873{ 3874 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 3875} 3876 3877static ssize_t sanity_checks_store(struct kmem_cache *s, 3878 const char *buf, size_t length) 3879{ 3880 s->flags &= ~SLAB_DEBUG_FREE; 3881 if (buf[0] == '1') 3882 s->flags |= SLAB_DEBUG_FREE; 3883 return length; 3884} 3885SLAB_ATTR(sanity_checks); 3886 3887static ssize_t trace_show(struct kmem_cache *s, char *buf) 3888{ 3889 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 3890} 3891 3892static ssize_t trace_store(struct kmem_cache *s, const char *buf, 3893 size_t length) 3894{ 3895 s->flags &= ~SLAB_TRACE; 3896 if (buf[0] == '1') 3897 s->flags |= SLAB_TRACE; 3898 return length; 3899} 3900SLAB_ATTR(trace); 3901 3902static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 3903{ 3904 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 3905} 3906 3907static ssize_t reclaim_account_store(struct kmem_cache *s, 3908 const char *buf, size_t length) 3909{ 3910 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 3911 if (buf[0] == '1') 3912 s->flags |= SLAB_RECLAIM_ACCOUNT; 3913 return length; 3914} 3915SLAB_ATTR(reclaim_account); 3916 3917static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 3918{ 3919 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 3920} 3921SLAB_ATTR_RO(hwcache_align); 3922 3923#ifdef CONFIG_ZONE_DMA 3924static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 3925{ 3926 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 3927} 3928SLAB_ATTR_RO(cache_dma); 3929#endif 3930 3931static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 3932{ 3933 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 3934} 3935SLAB_ATTR_RO(destroy_by_rcu); 3936 3937static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 3938{ 3939 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 3940} 3941 3942static ssize_t red_zone_store(struct kmem_cache *s, 3943 const char *buf, size_t length) 3944{ 3945 if (any_slab_objects(s)) 3946 return -EBUSY; 3947 3948 s->flags &= ~SLAB_RED_ZONE; 3949 if (buf[0] == '1') 3950 s->flags |= SLAB_RED_ZONE; 3951 calculate_sizes(s, -1); 3952 return length; 3953} 3954SLAB_ATTR(red_zone); 3955 3956static ssize_t poison_show(struct kmem_cache *s, char *buf) 3957{ 3958 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 3959} 3960 3961static ssize_t poison_store(struct kmem_cache *s, 3962 const char *buf, size_t length) 3963{ 3964 if (any_slab_objects(s)) 3965 return -EBUSY; 3966 3967 s->flags &= ~SLAB_POISON; 3968 if (buf[0] == '1') 3969 s->flags |= SLAB_POISON; 3970 calculate_sizes(s, -1); 3971 return length; 3972} 3973SLAB_ATTR(poison); 3974 3975static ssize_t store_user_show(struct kmem_cache *s, char *buf) 3976{ 3977 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 3978} 3979 3980static ssize_t store_user_store(struct kmem_cache *s, 3981 const char *buf, size_t length) 3982{ 3983 if (any_slab_objects(s)) 3984 return -EBUSY; 3985 3986 s->flags &= ~SLAB_STORE_USER; 3987 if (buf[0] == '1') 3988 s->flags |= SLAB_STORE_USER; 3989 calculate_sizes(s, -1); 3990 return length; 3991} 3992SLAB_ATTR(store_user); 3993 3994static ssize_t validate_show(struct kmem_cache *s, char *buf) 3995{ 3996 return 0; 3997} 3998 3999static ssize_t validate_store(struct kmem_cache *s, 4000 const char *buf, size_t length) 4001{ 4002 int ret = -EINVAL; 4003 4004 if (buf[0] == '1') { 4005 ret = validate_slab_cache(s); 4006 if (ret >= 0) 4007 ret = length; 4008 } 4009 return ret; 4010} 4011SLAB_ATTR(validate); 4012 4013static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4014{ 4015 return 0; 4016} 4017 4018static ssize_t shrink_store(struct kmem_cache *s, 4019 const char *buf, size_t length) 4020{ 4021 if (buf[0] == '1') { 4022 int rc = kmem_cache_shrink(s); 4023 4024 if (rc) 4025 return rc; 4026 } else 4027 return -EINVAL; 4028 return length; 4029} 4030SLAB_ATTR(shrink); 4031 4032static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4033{ 4034 if (!(s->flags & SLAB_STORE_USER)) 4035 return -ENOSYS; 4036 return list_locations(s, buf, TRACK_ALLOC); 4037} 4038SLAB_ATTR_RO(alloc_calls); 4039 4040static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4041{ 4042 if (!(s->flags & SLAB_STORE_USER)) 4043 return -ENOSYS; 4044 return list_locations(s, buf, TRACK_FREE); 4045} 4046SLAB_ATTR_RO(free_calls); 4047 4048#ifdef CONFIG_NUMA 4049static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4050{ 4051 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4052} 4053 4054static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4055 const char *buf, size_t length) 4056{ 4057 unsigned long ratio; 4058 int err; 4059 4060 err = strict_strtoul(buf, 10, &ratio); 4061 if (err) 4062 return err; 4063 4064 if (ratio <= 100) 4065 s->remote_node_defrag_ratio = ratio * 10; 4066 4067 return length; 4068} 4069SLAB_ATTR(remote_node_defrag_ratio); 4070#endif 4071 4072#ifdef CONFIG_SLUB_STATS 4073static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4074{ 4075 unsigned long sum = 0; 4076 int cpu; 4077 int len; 4078 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4079 4080 if (!data) 4081 return -ENOMEM; 4082 4083 for_each_online_cpu(cpu) { 4084 unsigned x = get_cpu_slab(s, cpu)->stat[si]; 4085 4086 data[cpu] = x; 4087 sum += x; 4088 } 4089 4090 len = sprintf(buf, "%lu", sum); 4091 4092#ifdef CONFIG_SMP 4093 for_each_online_cpu(cpu) { 4094 if (data[cpu] && len < PAGE_SIZE - 20) 4095 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4096 } 4097#endif 4098 kfree(data); 4099 return len + sprintf(buf + len, "\n"); 4100} 4101 4102#define STAT_ATTR(si, text) \ 4103static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4104{ \ 4105 return show_stat(s, buf, si); \ 4106} \ 4107SLAB_ATTR_RO(text); \ 4108 4109STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4110STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4111STAT_ATTR(FREE_FASTPATH, free_fastpath); 4112STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4113STAT_ATTR(FREE_FROZEN, free_frozen); 4114STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4115STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4116STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4117STAT_ATTR(ALLOC_SLAB, alloc_slab); 4118STAT_ATTR(ALLOC_REFILL, alloc_refill); 4119STAT_ATTR(FREE_SLAB, free_slab); 4120STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4121STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4122STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4123STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4124STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4125STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4126STAT_ATTR(ORDER_FALLBACK, order_fallback); 4127#endif 4128 4129static struct attribute *slab_attrs[] = { 4130 &slab_size_attr.attr, 4131 &object_size_attr.attr, 4132 &objs_per_slab_attr.attr, 4133 &order_attr.attr, 4134 &objects_attr.attr, 4135 &objects_partial_attr.attr, 4136 &total_objects_attr.attr, 4137 &slabs_attr.attr, 4138 &partial_attr.attr, 4139 &cpu_slabs_attr.attr, 4140 &ctor_attr.attr, 4141 &aliases_attr.attr, 4142 &align_attr.attr, 4143 &sanity_checks_attr.attr, 4144 &trace_attr.attr, 4145 &hwcache_align_attr.attr, 4146 &reclaim_account_attr.attr, 4147 &destroy_by_rcu_attr.attr, 4148 &red_zone_attr.attr, 4149 &poison_attr.attr, 4150 &store_user_attr.attr, 4151 &validate_attr.attr, 4152 &shrink_attr.attr, 4153 &alloc_calls_attr.attr, 4154 &free_calls_attr.attr, 4155#ifdef CONFIG_ZONE_DMA 4156 &cache_dma_attr.attr, 4157#endif 4158#ifdef CONFIG_NUMA 4159 &remote_node_defrag_ratio_attr.attr, 4160#endif 4161#ifdef CONFIG_SLUB_STATS 4162 &alloc_fastpath_attr.attr, 4163 &alloc_slowpath_attr.attr, 4164 &free_fastpath_attr.attr, 4165 &free_slowpath_attr.attr, 4166 &free_frozen_attr.attr, 4167 &free_add_partial_attr.attr, 4168 &free_remove_partial_attr.attr, 4169 &alloc_from_partial_attr.attr, 4170 &alloc_slab_attr.attr, 4171 &alloc_refill_attr.attr, 4172 &free_slab_attr.attr, 4173 &cpuslab_flush_attr.attr, 4174 &deactivate_full_attr.attr, 4175 &deactivate_empty_attr.attr, 4176 &deactivate_to_head_attr.attr, 4177 &deactivate_to_tail_attr.attr, 4178 &deactivate_remote_frees_attr.attr, 4179 &order_fallback_attr.attr, 4180#endif 4181 NULL 4182}; 4183 4184static struct attribute_group slab_attr_group = { 4185 .attrs = slab_attrs, 4186}; 4187 4188static ssize_t slab_attr_show(struct kobject *kobj, 4189 struct attribute *attr, 4190 char *buf) 4191{ 4192 struct slab_attribute *attribute; 4193 struct kmem_cache *s; 4194 int err; 4195 4196 attribute = to_slab_attr(attr); 4197 s = to_slab(kobj); 4198 4199 if (!attribute->show) 4200 return -EIO; 4201 4202 err = attribute->show(s, buf); 4203 4204 return err; 4205} 4206 4207static ssize_t slab_attr_store(struct kobject *kobj, 4208 struct attribute *attr, 4209 const char *buf, size_t len) 4210{ 4211 struct slab_attribute *attribute; 4212 struct kmem_cache *s; 4213 int err; 4214 4215 attribute = to_slab_attr(attr); 4216 s = to_slab(kobj); 4217 4218 if (!attribute->store) 4219 return -EIO; 4220 4221 err = attribute->store(s, buf, len); 4222 4223 return err; 4224} 4225 4226static void kmem_cache_release(struct kobject *kobj) 4227{ 4228 struct kmem_cache *s = to_slab(kobj); 4229 4230 kfree(s); 4231} 4232 4233static struct sysfs_ops slab_sysfs_ops = { 4234 .show = slab_attr_show, 4235 .store = slab_attr_store, 4236}; 4237 4238static struct kobj_type slab_ktype = { 4239 .sysfs_ops = &slab_sysfs_ops, 4240 .release = kmem_cache_release 4241}; 4242 4243static int uevent_filter(struct kset *kset, struct kobject *kobj) 4244{ 4245 struct kobj_type *ktype = get_ktype(kobj); 4246 4247 if (ktype == &slab_ktype) 4248 return 1; 4249 return 0; 4250} 4251 4252static struct kset_uevent_ops slab_uevent_ops = { 4253 .filter = uevent_filter, 4254}; 4255 4256static struct kset *slab_kset; 4257 4258#define ID_STR_LENGTH 64 4259 4260/* Create a unique string id for a slab cache: 4261 * 4262 * Format :[flags-]size 4263 */ 4264static char *create_unique_id(struct kmem_cache *s) 4265{ 4266 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4267 char *p = name; 4268 4269 BUG_ON(!name); 4270 4271 *p++ = ':'; 4272 /* 4273 * First flags affecting slabcache operations. We will only 4274 * get here for aliasable slabs so we do not need to support 4275 * too many flags. The flags here must cover all flags that 4276 * are matched during merging to guarantee that the id is 4277 * unique. 4278 */ 4279 if (s->flags & SLAB_CACHE_DMA) 4280 *p++ = 'd'; 4281 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4282 *p++ = 'a'; 4283 if (s->flags & SLAB_DEBUG_FREE) 4284 *p++ = 'F'; 4285 if (p != name + 1) 4286 *p++ = '-'; 4287 p += sprintf(p, "%07d", s->size); 4288 BUG_ON(p > name + ID_STR_LENGTH - 1); 4289 return name; 4290} 4291 4292static int sysfs_slab_add(struct kmem_cache *s) 4293{ 4294 int err; 4295 const char *name; 4296 int unmergeable; 4297 4298 if (slab_state < SYSFS) 4299 /* Defer until later */ 4300 return 0; 4301 4302 unmergeable = slab_unmergeable(s); 4303 if (unmergeable) { 4304 /* 4305 * Slabcache can never be merged so we can use the name proper. 4306 * This is typically the case for debug situations. In that 4307 * case we can catch duplicate names easily. 4308 */ 4309 sysfs_remove_link(&slab_kset->kobj, s->name); 4310 name = s->name; 4311 } else { 4312 /* 4313 * Create a unique name for the slab as a target 4314 * for the symlinks. 4315 */ 4316 name = create_unique_id(s); 4317 } 4318 4319 s->kobj.kset = slab_kset; 4320 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4321 if (err) { 4322 kobject_put(&s->kobj); 4323 return err; 4324 } 4325 4326 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4327 if (err) 4328 return err; 4329 kobject_uevent(&s->kobj, KOBJ_ADD); 4330 if (!unmergeable) { 4331 /* Setup first alias */ 4332 sysfs_slab_alias(s, s->name); 4333 kfree(name); 4334 } 4335 return 0; 4336} 4337 4338static void sysfs_slab_remove(struct kmem_cache *s) 4339{ 4340 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4341 kobject_del(&s->kobj); 4342 kobject_put(&s->kobj); 4343} 4344 4345/* 4346 * Need to buffer aliases during bootup until sysfs becomes 4347 * available lest we loose that information. 4348 */ 4349struct saved_alias { 4350 struct kmem_cache *s; 4351 const char *name; 4352 struct saved_alias *next; 4353}; 4354 4355static struct saved_alias *alias_list; 4356 4357static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4358{ 4359 struct saved_alias *al; 4360 4361 if (slab_state == SYSFS) { 4362 /* 4363 * If we have a leftover link then remove it. 4364 */ 4365 sysfs_remove_link(&slab_kset->kobj, name); 4366 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4367 } 4368 4369 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4370 if (!al) 4371 return -ENOMEM; 4372 4373 al->s = s; 4374 al->name = name; 4375 al->next = alias_list; 4376 alias_list = al; 4377 return 0; 4378} 4379 4380static int __init slab_sysfs_init(void) 4381{ 4382 struct kmem_cache *s; 4383 int err; 4384 4385 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4386 if (!slab_kset) { 4387 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4388 return -ENOSYS; 4389 } 4390 4391 slab_state = SYSFS; 4392 4393 list_for_each_entry(s, &slab_caches, list) { 4394 err = sysfs_slab_add(s); 4395 if (err) 4396 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4397 " to sysfs\n", s->name); 4398 } 4399 4400 while (alias_list) { 4401 struct saved_alias *al = alias_list; 4402 4403 alias_list = alias_list->next; 4404 err = sysfs_slab_alias(al->s, al->name); 4405 if (err) 4406 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4407 " %s to sysfs\n", s->name); 4408 kfree(al); 4409 } 4410 4411 resiliency_test(); 4412 return 0; 4413} 4414 4415__initcall(slab_sysfs_init); 4416#endif 4417 4418/* 4419 * The /proc/slabinfo ABI 4420 */ 4421#ifdef CONFIG_SLABINFO 4422static void print_slabinfo_header(struct seq_file *m) 4423{ 4424 seq_puts(m, "slabinfo - version: 2.1\n"); 4425 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4426 "<objperslab> <pagesperslab>"); 4427 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4428 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4429 seq_putc(m, '\n'); 4430} 4431 4432static void *s_start(struct seq_file *m, loff_t *pos) 4433{ 4434 loff_t n = *pos; 4435 4436 down_read(&slub_lock); 4437 if (!n) 4438 print_slabinfo_header(m); 4439 4440 return seq_list_start(&slab_caches, *pos); 4441} 4442 4443static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4444{ 4445 return seq_list_next(p, &slab_caches, pos); 4446} 4447 4448static void s_stop(struct seq_file *m, void *p) 4449{ 4450 up_read(&slub_lock); 4451} 4452 4453static int s_show(struct seq_file *m, void *p) 4454{ 4455 unsigned long nr_partials = 0; 4456 unsigned long nr_slabs = 0; 4457 unsigned long nr_inuse = 0; 4458 unsigned long nr_objs = 0; 4459 unsigned long nr_free = 0; 4460 struct kmem_cache *s; 4461 int node; 4462 4463 s = list_entry(p, struct kmem_cache, list); 4464 4465 for_each_online_node(node) { 4466 struct kmem_cache_node *n = get_node(s, node); 4467 4468 if (!n) 4469 continue; 4470 4471 nr_partials += n->nr_partial; 4472 nr_slabs += atomic_long_read(&n->nr_slabs); 4473 nr_objs += atomic_long_read(&n->total_objects); 4474 nr_free += count_partial(n, count_free); 4475 } 4476 4477 nr_inuse = nr_objs - nr_free; 4478 4479 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4480 nr_objs, s->size, oo_objects(s->oo), 4481 (1 << oo_order(s->oo))); 4482 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4483 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4484 0UL); 4485 seq_putc(m, '\n'); 4486 return 0; 4487} 4488 4489static const struct seq_operations slabinfo_op = { 4490 .start = s_start, 4491 .next = s_next, 4492 .stop = s_stop, 4493 .show = s_show, 4494}; 4495 4496static int slabinfo_open(struct inode *inode, struct file *file) 4497{ 4498 return seq_open(file, &slabinfo_op); 4499} 4500 4501static const struct file_operations proc_slabinfo_operations = { 4502 .open = slabinfo_open, 4503 .read = seq_read, 4504 .llseek = seq_lseek, 4505 .release = seq_release, 4506}; 4507 4508static int __init slab_proc_init(void) 4509{ 4510 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); 4511 return 0; 4512} 4513module_init(slab_proc_init); 4514#endif /* CONFIG_SLABINFO */ 4515