slub.c revision 756dee75872a2a764b478e18076360b8a4ec9045
1/* 2 * SLUB: A slab allocator that limits cache line use instead of queuing 3 * objects in per cpu and per node lists. 4 * 5 * The allocator synchronizes using per slab locks and only 6 * uses a centralized lock to manage a pool of partial slabs. 7 * 8 * (C) 2007 SGI, Christoph Lameter 9 */ 10 11#include <linux/mm.h> 12#include <linux/swap.h> /* struct reclaim_state */ 13#include <linux/module.h> 14#include <linux/bit_spinlock.h> 15#include <linux/interrupt.h> 16#include <linux/bitops.h> 17#include <linux/slab.h> 18#include <linux/proc_fs.h> 19#include <linux/seq_file.h> 20#include <linux/kmemtrace.h> 21#include <linux/kmemcheck.h> 22#include <linux/cpu.h> 23#include <linux/cpuset.h> 24#include <linux/mempolicy.h> 25#include <linux/ctype.h> 26#include <linux/debugobjects.h> 27#include <linux/kallsyms.h> 28#include <linux/memory.h> 29#include <linux/math64.h> 30#include <linux/fault-inject.h> 31 32/* 33 * Lock order: 34 * 1. slab_lock(page) 35 * 2. slab->list_lock 36 * 37 * The slab_lock protects operations on the object of a particular 38 * slab and its metadata in the page struct. If the slab lock 39 * has been taken then no allocations nor frees can be performed 40 * on the objects in the slab nor can the slab be added or removed 41 * from the partial or full lists since this would mean modifying 42 * the page_struct of the slab. 43 * 44 * The list_lock protects the partial and full list on each node and 45 * the partial slab counter. If taken then no new slabs may be added or 46 * removed from the lists nor make the number of partial slabs be modified. 47 * (Note that the total number of slabs is an atomic value that may be 48 * modified without taking the list lock). 49 * 50 * The list_lock is a centralized lock and thus we avoid taking it as 51 * much as possible. As long as SLUB does not have to handle partial 52 * slabs, operations can continue without any centralized lock. F.e. 53 * allocating a long series of objects that fill up slabs does not require 54 * the list lock. 55 * 56 * The lock order is sometimes inverted when we are trying to get a slab 57 * off a list. We take the list_lock and then look for a page on the list 58 * to use. While we do that objects in the slabs may be freed. We can 59 * only operate on the slab if we have also taken the slab_lock. So we use 60 * a slab_trylock() on the slab. If trylock was successful then no frees 61 * can occur anymore and we can use the slab for allocations etc. If the 62 * slab_trylock() does not succeed then frees are in progress in the slab and 63 * we must stay away from it for a while since we may cause a bouncing 64 * cacheline if we try to acquire the lock. So go onto the next slab. 65 * If all pages are busy then we may allocate a new slab instead of reusing 66 * a partial slab. A new slab has noone operating on it and thus there is 67 * no danger of cacheline contention. 68 * 69 * Interrupts are disabled during allocation and deallocation in order to 70 * make the slab allocator safe to use in the context of an irq. In addition 71 * interrupts are disabled to ensure that the processor does not change 72 * while handling per_cpu slabs, due to kernel preemption. 73 * 74 * SLUB assigns one slab for allocation to each processor. 75 * Allocations only occur from these slabs called cpu slabs. 76 * 77 * Slabs with free elements are kept on a partial list and during regular 78 * operations no list for full slabs is used. If an object in a full slab is 79 * freed then the slab will show up again on the partial lists. 80 * We track full slabs for debugging purposes though because otherwise we 81 * cannot scan all objects. 82 * 83 * Slabs are freed when they become empty. Teardown and setup is 84 * minimal so we rely on the page allocators per cpu caches for 85 * fast frees and allocs. 86 * 87 * Overloading of page flags that are otherwise used for LRU management. 88 * 89 * PageActive The slab is frozen and exempt from list processing. 90 * This means that the slab is dedicated to a purpose 91 * such as satisfying allocations for a specific 92 * processor. Objects may be freed in the slab while 93 * it is frozen but slab_free will then skip the usual 94 * list operations. It is up to the processor holding 95 * the slab to integrate the slab into the slab lists 96 * when the slab is no longer needed. 97 * 98 * One use of this flag is to mark slabs that are 99 * used for allocations. Then such a slab becomes a cpu 100 * slab. The cpu slab may be equipped with an additional 101 * freelist that allows lockless access to 102 * free objects in addition to the regular freelist 103 * that requires the slab lock. 104 * 105 * PageError Slab requires special handling due to debug 106 * options set. This moves slab handling out of 107 * the fast path and disables lockless freelists. 108 */ 109 110#ifdef CONFIG_SLUB_DEBUG 111#define SLABDEBUG 1 112#else 113#define SLABDEBUG 0 114#endif 115 116/* 117 * Issues still to be resolved: 118 * 119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 120 * 121 * - Variable sizing of the per node arrays 122 */ 123 124/* Enable to test recovery from slab corruption on boot */ 125#undef SLUB_RESILIENCY_TEST 126 127/* 128 * Mininum number of partial slabs. These will be left on the partial 129 * lists even if they are empty. kmem_cache_shrink may reclaim them. 130 */ 131#define MIN_PARTIAL 5 132 133/* 134 * Maximum number of desirable partial slabs. 135 * The existence of more partial slabs makes kmem_cache_shrink 136 * sort the partial list by the number of objects in the. 137 */ 138#define MAX_PARTIAL 10 139 140#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 141 SLAB_POISON | SLAB_STORE_USER) 142 143/* 144 * Debugging flags that require metadata to be stored in the slab. These get 145 * disabled when slub_debug=O is used and a cache's min order increases with 146 * metadata. 147 */ 148#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 149 150/* 151 * Set of flags that will prevent slab merging 152 */ 153#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 154 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE) 155 156#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 157 SLAB_CACHE_DMA | SLAB_NOTRACK) 158 159#ifndef ARCH_KMALLOC_MINALIGN 160#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 161#endif 162 163#ifndef ARCH_SLAB_MINALIGN 164#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) 165#endif 166 167#define OO_SHIFT 16 168#define OO_MASK ((1 << OO_SHIFT) - 1) 169#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ 170 171/* Internal SLUB flags */ 172#define __OBJECT_POISON 0x80000000 /* Poison object */ 173#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ 174 175static int kmem_size = sizeof(struct kmem_cache); 176 177#ifdef CONFIG_SMP 178static struct notifier_block slab_notifier; 179#endif 180 181static enum { 182 DOWN, /* No slab functionality available */ 183 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 184 UP, /* Everything works but does not show up in sysfs */ 185 SYSFS /* Sysfs up */ 186} slab_state = DOWN; 187 188/* A list of all slab caches on the system */ 189static DECLARE_RWSEM(slub_lock); 190static LIST_HEAD(slab_caches); 191 192/* 193 * Tracking user of a slab. 194 */ 195struct track { 196 unsigned long addr; /* Called from address */ 197 int cpu; /* Was running on cpu */ 198 int pid; /* Pid context */ 199 unsigned long when; /* When did the operation occur */ 200}; 201 202enum track_item { TRACK_ALLOC, TRACK_FREE }; 203 204#ifdef CONFIG_SLUB_DEBUG 205static int sysfs_slab_add(struct kmem_cache *); 206static int sysfs_slab_alias(struct kmem_cache *, const char *); 207static void sysfs_slab_remove(struct kmem_cache *); 208 209#else 210static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 211static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 212 { return 0; } 213static inline void sysfs_slab_remove(struct kmem_cache *s) 214{ 215 kfree(s); 216} 217 218#endif 219 220static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) 221{ 222#ifdef CONFIG_SLUB_STATS 223 c->stat[si]++; 224#endif 225} 226 227/******************************************************************** 228 * Core slab cache functions 229 *******************************************************************/ 230 231int slab_is_available(void) 232{ 233 return slab_state >= UP; 234} 235 236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 237{ 238#ifdef CONFIG_NUMA 239 return s->node[node]; 240#else 241 return &s->local_node; 242#endif 243} 244 245/* Verify that a pointer has an address that is valid within a slab page */ 246static inline int check_valid_pointer(struct kmem_cache *s, 247 struct page *page, const void *object) 248{ 249 void *base; 250 251 if (!object) 252 return 1; 253 254 base = page_address(page); 255 if (object < base || object >= base + page->objects * s->size || 256 (object - base) % s->size) { 257 return 0; 258 } 259 260 return 1; 261} 262 263/* 264 * Slow version of get and set free pointer. 265 * 266 * This version requires touching the cache lines of kmem_cache which 267 * we avoid to do in the fast alloc free paths. There we obtain the offset 268 * from the page struct. 269 */ 270static inline void *get_freepointer(struct kmem_cache *s, void *object) 271{ 272 return *(void **)(object + s->offset); 273} 274 275static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 276{ 277 *(void **)(object + s->offset) = fp; 278} 279 280/* Loop over all objects in a slab */ 281#define for_each_object(__p, __s, __addr, __objects) \ 282 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 283 __p += (__s)->size) 284 285/* Scan freelist */ 286#define for_each_free_object(__p, __s, __free) \ 287 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 288 289/* Determine object index from a given position */ 290static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 291{ 292 return (p - addr) / s->size; 293} 294 295static inline struct kmem_cache_order_objects oo_make(int order, 296 unsigned long size) 297{ 298 struct kmem_cache_order_objects x = { 299 (order << OO_SHIFT) + (PAGE_SIZE << order) / size 300 }; 301 302 return x; 303} 304 305static inline int oo_order(struct kmem_cache_order_objects x) 306{ 307 return x.x >> OO_SHIFT; 308} 309 310static inline int oo_objects(struct kmem_cache_order_objects x) 311{ 312 return x.x & OO_MASK; 313} 314 315#ifdef CONFIG_SLUB_DEBUG 316/* 317 * Debug settings: 318 */ 319#ifdef CONFIG_SLUB_DEBUG_ON 320static int slub_debug = DEBUG_DEFAULT_FLAGS; 321#else 322static int slub_debug; 323#endif 324 325static char *slub_debug_slabs; 326static int disable_higher_order_debug; 327 328/* 329 * Object debugging 330 */ 331static void print_section(char *text, u8 *addr, unsigned int length) 332{ 333 int i, offset; 334 int newline = 1; 335 char ascii[17]; 336 337 ascii[16] = 0; 338 339 for (i = 0; i < length; i++) { 340 if (newline) { 341 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 342 newline = 0; 343 } 344 printk(KERN_CONT " %02x", addr[i]); 345 offset = i % 16; 346 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 347 if (offset == 15) { 348 printk(KERN_CONT " %s\n", ascii); 349 newline = 1; 350 } 351 } 352 if (!newline) { 353 i %= 16; 354 while (i < 16) { 355 printk(KERN_CONT " "); 356 ascii[i] = ' '; 357 i++; 358 } 359 printk(KERN_CONT " %s\n", ascii); 360 } 361} 362 363static struct track *get_track(struct kmem_cache *s, void *object, 364 enum track_item alloc) 365{ 366 struct track *p; 367 368 if (s->offset) 369 p = object + s->offset + sizeof(void *); 370 else 371 p = object + s->inuse; 372 373 return p + alloc; 374} 375 376static void set_track(struct kmem_cache *s, void *object, 377 enum track_item alloc, unsigned long addr) 378{ 379 struct track *p = get_track(s, object, alloc); 380 381 if (addr) { 382 p->addr = addr; 383 p->cpu = smp_processor_id(); 384 p->pid = current->pid; 385 p->when = jiffies; 386 } else 387 memset(p, 0, sizeof(struct track)); 388} 389 390static void init_tracking(struct kmem_cache *s, void *object) 391{ 392 if (!(s->flags & SLAB_STORE_USER)) 393 return; 394 395 set_track(s, object, TRACK_FREE, 0UL); 396 set_track(s, object, TRACK_ALLOC, 0UL); 397} 398 399static void print_track(const char *s, struct track *t) 400{ 401 if (!t->addr) 402 return; 403 404 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 405 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 406} 407 408static void print_tracking(struct kmem_cache *s, void *object) 409{ 410 if (!(s->flags & SLAB_STORE_USER)) 411 return; 412 413 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 414 print_track("Freed", get_track(s, object, TRACK_FREE)); 415} 416 417static void print_page_info(struct page *page) 418{ 419 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 420 page, page->objects, page->inuse, page->freelist, page->flags); 421 422} 423 424static void slab_bug(struct kmem_cache *s, char *fmt, ...) 425{ 426 va_list args; 427 char buf[100]; 428 429 va_start(args, fmt); 430 vsnprintf(buf, sizeof(buf), fmt, args); 431 va_end(args); 432 printk(KERN_ERR "========================================" 433 "=====================================\n"); 434 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 435 printk(KERN_ERR "----------------------------------------" 436 "-------------------------------------\n\n"); 437} 438 439static void slab_fix(struct kmem_cache *s, char *fmt, ...) 440{ 441 va_list args; 442 char buf[100]; 443 444 va_start(args, fmt); 445 vsnprintf(buf, sizeof(buf), fmt, args); 446 va_end(args); 447 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 448} 449 450static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 451{ 452 unsigned int off; /* Offset of last byte */ 453 u8 *addr = page_address(page); 454 455 print_tracking(s, p); 456 457 print_page_info(page); 458 459 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 460 p, p - addr, get_freepointer(s, p)); 461 462 if (p > addr + 16) 463 print_section("Bytes b4", p - 16, 16); 464 465 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 466 467 if (s->flags & SLAB_RED_ZONE) 468 print_section("Redzone", p + s->objsize, 469 s->inuse - s->objsize); 470 471 if (s->offset) 472 off = s->offset + sizeof(void *); 473 else 474 off = s->inuse; 475 476 if (s->flags & SLAB_STORE_USER) 477 off += 2 * sizeof(struct track); 478 479 if (off != s->size) 480 /* Beginning of the filler is the free pointer */ 481 print_section("Padding", p + off, s->size - off); 482 483 dump_stack(); 484} 485 486static void object_err(struct kmem_cache *s, struct page *page, 487 u8 *object, char *reason) 488{ 489 slab_bug(s, "%s", reason); 490 print_trailer(s, page, object); 491} 492 493static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 494{ 495 va_list args; 496 char buf[100]; 497 498 va_start(args, fmt); 499 vsnprintf(buf, sizeof(buf), fmt, args); 500 va_end(args); 501 slab_bug(s, "%s", buf); 502 print_page_info(page); 503 dump_stack(); 504} 505 506static void init_object(struct kmem_cache *s, void *object, int active) 507{ 508 u8 *p = object; 509 510 if (s->flags & __OBJECT_POISON) { 511 memset(p, POISON_FREE, s->objsize - 1); 512 p[s->objsize - 1] = POISON_END; 513 } 514 515 if (s->flags & SLAB_RED_ZONE) 516 memset(p + s->objsize, 517 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 518 s->inuse - s->objsize); 519} 520 521static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 522{ 523 while (bytes) { 524 if (*start != (u8)value) 525 return start; 526 start++; 527 bytes--; 528 } 529 return NULL; 530} 531 532static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 533 void *from, void *to) 534{ 535 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 536 memset(from, data, to - from); 537} 538 539static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 540 u8 *object, char *what, 541 u8 *start, unsigned int value, unsigned int bytes) 542{ 543 u8 *fault; 544 u8 *end; 545 546 fault = check_bytes(start, value, bytes); 547 if (!fault) 548 return 1; 549 550 end = start + bytes; 551 while (end > fault && end[-1] == value) 552 end--; 553 554 slab_bug(s, "%s overwritten", what); 555 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 556 fault, end - 1, fault[0], value); 557 print_trailer(s, page, object); 558 559 restore_bytes(s, what, value, fault, end); 560 return 0; 561} 562 563/* 564 * Object layout: 565 * 566 * object address 567 * Bytes of the object to be managed. 568 * If the freepointer may overlay the object then the free 569 * pointer is the first word of the object. 570 * 571 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 572 * 0xa5 (POISON_END) 573 * 574 * object + s->objsize 575 * Padding to reach word boundary. This is also used for Redzoning. 576 * Padding is extended by another word if Redzoning is enabled and 577 * objsize == inuse. 578 * 579 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 580 * 0xcc (RED_ACTIVE) for objects in use. 581 * 582 * object + s->inuse 583 * Meta data starts here. 584 * 585 * A. Free pointer (if we cannot overwrite object on free) 586 * B. Tracking data for SLAB_STORE_USER 587 * C. Padding to reach required alignment boundary or at mininum 588 * one word if debugging is on to be able to detect writes 589 * before the word boundary. 590 * 591 * Padding is done using 0x5a (POISON_INUSE) 592 * 593 * object + s->size 594 * Nothing is used beyond s->size. 595 * 596 * If slabcaches are merged then the objsize and inuse boundaries are mostly 597 * ignored. And therefore no slab options that rely on these boundaries 598 * may be used with merged slabcaches. 599 */ 600 601static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 602{ 603 unsigned long off = s->inuse; /* The end of info */ 604 605 if (s->offset) 606 /* Freepointer is placed after the object. */ 607 off += sizeof(void *); 608 609 if (s->flags & SLAB_STORE_USER) 610 /* We also have user information there */ 611 off += 2 * sizeof(struct track); 612 613 if (s->size == off) 614 return 1; 615 616 return check_bytes_and_report(s, page, p, "Object padding", 617 p + off, POISON_INUSE, s->size - off); 618} 619 620/* Check the pad bytes at the end of a slab page */ 621static int slab_pad_check(struct kmem_cache *s, struct page *page) 622{ 623 u8 *start; 624 u8 *fault; 625 u8 *end; 626 int length; 627 int remainder; 628 629 if (!(s->flags & SLAB_POISON)) 630 return 1; 631 632 start = page_address(page); 633 length = (PAGE_SIZE << compound_order(page)); 634 end = start + length; 635 remainder = length % s->size; 636 if (!remainder) 637 return 1; 638 639 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 640 if (!fault) 641 return 1; 642 while (end > fault && end[-1] == POISON_INUSE) 643 end--; 644 645 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 646 print_section("Padding", end - remainder, remainder); 647 648 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 649 return 0; 650} 651 652static int check_object(struct kmem_cache *s, struct page *page, 653 void *object, int active) 654{ 655 u8 *p = object; 656 u8 *endobject = object + s->objsize; 657 658 if (s->flags & SLAB_RED_ZONE) { 659 unsigned int red = 660 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 661 662 if (!check_bytes_and_report(s, page, object, "Redzone", 663 endobject, red, s->inuse - s->objsize)) 664 return 0; 665 } else { 666 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 667 check_bytes_and_report(s, page, p, "Alignment padding", 668 endobject, POISON_INUSE, s->inuse - s->objsize); 669 } 670 } 671 672 if (s->flags & SLAB_POISON) { 673 if (!active && (s->flags & __OBJECT_POISON) && 674 (!check_bytes_and_report(s, page, p, "Poison", p, 675 POISON_FREE, s->objsize - 1) || 676 !check_bytes_and_report(s, page, p, "Poison", 677 p + s->objsize - 1, POISON_END, 1))) 678 return 0; 679 /* 680 * check_pad_bytes cleans up on its own. 681 */ 682 check_pad_bytes(s, page, p); 683 } 684 685 if (!s->offset && active) 686 /* 687 * Object and freepointer overlap. Cannot check 688 * freepointer while object is allocated. 689 */ 690 return 1; 691 692 /* Check free pointer validity */ 693 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 694 object_err(s, page, p, "Freepointer corrupt"); 695 /* 696 * No choice but to zap it and thus lose the remainder 697 * of the free objects in this slab. May cause 698 * another error because the object count is now wrong. 699 */ 700 set_freepointer(s, p, NULL); 701 return 0; 702 } 703 return 1; 704} 705 706static int check_slab(struct kmem_cache *s, struct page *page) 707{ 708 int maxobj; 709 710 VM_BUG_ON(!irqs_disabled()); 711 712 if (!PageSlab(page)) { 713 slab_err(s, page, "Not a valid slab page"); 714 return 0; 715 } 716 717 maxobj = (PAGE_SIZE << compound_order(page)) / s->size; 718 if (page->objects > maxobj) { 719 slab_err(s, page, "objects %u > max %u", 720 s->name, page->objects, maxobj); 721 return 0; 722 } 723 if (page->inuse > page->objects) { 724 slab_err(s, page, "inuse %u > max %u", 725 s->name, page->inuse, page->objects); 726 return 0; 727 } 728 /* Slab_pad_check fixes things up after itself */ 729 slab_pad_check(s, page); 730 return 1; 731} 732 733/* 734 * Determine if a certain object on a page is on the freelist. Must hold the 735 * slab lock to guarantee that the chains are in a consistent state. 736 */ 737static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 738{ 739 int nr = 0; 740 void *fp = page->freelist; 741 void *object = NULL; 742 unsigned long max_objects; 743 744 while (fp && nr <= page->objects) { 745 if (fp == search) 746 return 1; 747 if (!check_valid_pointer(s, page, fp)) { 748 if (object) { 749 object_err(s, page, object, 750 "Freechain corrupt"); 751 set_freepointer(s, object, NULL); 752 break; 753 } else { 754 slab_err(s, page, "Freepointer corrupt"); 755 page->freelist = NULL; 756 page->inuse = page->objects; 757 slab_fix(s, "Freelist cleared"); 758 return 0; 759 } 760 break; 761 } 762 object = fp; 763 fp = get_freepointer(s, object); 764 nr++; 765 } 766 767 max_objects = (PAGE_SIZE << compound_order(page)) / s->size; 768 if (max_objects > MAX_OBJS_PER_PAGE) 769 max_objects = MAX_OBJS_PER_PAGE; 770 771 if (page->objects != max_objects) { 772 slab_err(s, page, "Wrong number of objects. Found %d but " 773 "should be %d", page->objects, max_objects); 774 page->objects = max_objects; 775 slab_fix(s, "Number of objects adjusted."); 776 } 777 if (page->inuse != page->objects - nr) { 778 slab_err(s, page, "Wrong object count. Counter is %d but " 779 "counted were %d", page->inuse, page->objects - nr); 780 page->inuse = page->objects - nr; 781 slab_fix(s, "Object count adjusted."); 782 } 783 return search == NULL; 784} 785 786static void trace(struct kmem_cache *s, struct page *page, void *object, 787 int alloc) 788{ 789 if (s->flags & SLAB_TRACE) { 790 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 791 s->name, 792 alloc ? "alloc" : "free", 793 object, page->inuse, 794 page->freelist); 795 796 if (!alloc) 797 print_section("Object", (void *)object, s->objsize); 798 799 dump_stack(); 800 } 801} 802 803/* 804 * Tracking of fully allocated slabs for debugging purposes. 805 */ 806static void add_full(struct kmem_cache_node *n, struct page *page) 807{ 808 spin_lock(&n->list_lock); 809 list_add(&page->lru, &n->full); 810 spin_unlock(&n->list_lock); 811} 812 813static void remove_full(struct kmem_cache *s, struct page *page) 814{ 815 struct kmem_cache_node *n; 816 817 if (!(s->flags & SLAB_STORE_USER)) 818 return; 819 820 n = get_node(s, page_to_nid(page)); 821 822 spin_lock(&n->list_lock); 823 list_del(&page->lru); 824 spin_unlock(&n->list_lock); 825} 826 827/* Tracking of the number of slabs for debugging purposes */ 828static inline unsigned long slabs_node(struct kmem_cache *s, int node) 829{ 830 struct kmem_cache_node *n = get_node(s, node); 831 832 return atomic_long_read(&n->nr_slabs); 833} 834 835static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 836{ 837 return atomic_long_read(&n->nr_slabs); 838} 839 840static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 841{ 842 struct kmem_cache_node *n = get_node(s, node); 843 844 /* 845 * May be called early in order to allocate a slab for the 846 * kmem_cache_node structure. Solve the chicken-egg 847 * dilemma by deferring the increment of the count during 848 * bootstrap (see early_kmem_cache_node_alloc). 849 */ 850 if (!NUMA_BUILD || n) { 851 atomic_long_inc(&n->nr_slabs); 852 atomic_long_add(objects, &n->total_objects); 853 } 854} 855static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 856{ 857 struct kmem_cache_node *n = get_node(s, node); 858 859 atomic_long_dec(&n->nr_slabs); 860 atomic_long_sub(objects, &n->total_objects); 861} 862 863/* Object debug checks for alloc/free paths */ 864static void setup_object_debug(struct kmem_cache *s, struct page *page, 865 void *object) 866{ 867 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 868 return; 869 870 init_object(s, object, 0); 871 init_tracking(s, object); 872} 873 874static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 875 void *object, unsigned long addr) 876{ 877 if (!check_slab(s, page)) 878 goto bad; 879 880 if (!on_freelist(s, page, object)) { 881 object_err(s, page, object, "Object already allocated"); 882 goto bad; 883 } 884 885 if (!check_valid_pointer(s, page, object)) { 886 object_err(s, page, object, "Freelist Pointer check fails"); 887 goto bad; 888 } 889 890 if (!check_object(s, page, object, 0)) 891 goto bad; 892 893 /* Success perform special debug activities for allocs */ 894 if (s->flags & SLAB_STORE_USER) 895 set_track(s, object, TRACK_ALLOC, addr); 896 trace(s, page, object, 1); 897 init_object(s, object, 1); 898 return 1; 899 900bad: 901 if (PageSlab(page)) { 902 /* 903 * If this is a slab page then lets do the best we can 904 * to avoid issues in the future. Marking all objects 905 * as used avoids touching the remaining objects. 906 */ 907 slab_fix(s, "Marking all objects used"); 908 page->inuse = page->objects; 909 page->freelist = NULL; 910 } 911 return 0; 912} 913 914static int free_debug_processing(struct kmem_cache *s, struct page *page, 915 void *object, unsigned long addr) 916{ 917 if (!check_slab(s, page)) 918 goto fail; 919 920 if (!check_valid_pointer(s, page, object)) { 921 slab_err(s, page, "Invalid object pointer 0x%p", object); 922 goto fail; 923 } 924 925 if (on_freelist(s, page, object)) { 926 object_err(s, page, object, "Object already free"); 927 goto fail; 928 } 929 930 if (!check_object(s, page, object, 1)) 931 return 0; 932 933 if (unlikely(s != page->slab)) { 934 if (!PageSlab(page)) { 935 slab_err(s, page, "Attempt to free object(0x%p) " 936 "outside of slab", object); 937 } else if (!page->slab) { 938 printk(KERN_ERR 939 "SLUB <none>: no slab for object 0x%p.\n", 940 object); 941 dump_stack(); 942 } else 943 object_err(s, page, object, 944 "page slab pointer corrupt."); 945 goto fail; 946 } 947 948 /* Special debug activities for freeing objects */ 949 if (!PageSlubFrozen(page) && !page->freelist) 950 remove_full(s, page); 951 if (s->flags & SLAB_STORE_USER) 952 set_track(s, object, TRACK_FREE, addr); 953 trace(s, page, object, 0); 954 init_object(s, object, 0); 955 return 1; 956 957fail: 958 slab_fix(s, "Object at 0x%p not freed", object); 959 return 0; 960} 961 962static int __init setup_slub_debug(char *str) 963{ 964 slub_debug = DEBUG_DEFAULT_FLAGS; 965 if (*str++ != '=' || !*str) 966 /* 967 * No options specified. Switch on full debugging. 968 */ 969 goto out; 970 971 if (*str == ',') 972 /* 973 * No options but restriction on slabs. This means full 974 * debugging for slabs matching a pattern. 975 */ 976 goto check_slabs; 977 978 if (tolower(*str) == 'o') { 979 /* 980 * Avoid enabling debugging on caches if its minimum order 981 * would increase as a result. 982 */ 983 disable_higher_order_debug = 1; 984 goto out; 985 } 986 987 slub_debug = 0; 988 if (*str == '-') 989 /* 990 * Switch off all debugging measures. 991 */ 992 goto out; 993 994 /* 995 * Determine which debug features should be switched on 996 */ 997 for (; *str && *str != ','; str++) { 998 switch (tolower(*str)) { 999 case 'f': 1000 slub_debug |= SLAB_DEBUG_FREE; 1001 break; 1002 case 'z': 1003 slub_debug |= SLAB_RED_ZONE; 1004 break; 1005 case 'p': 1006 slub_debug |= SLAB_POISON; 1007 break; 1008 case 'u': 1009 slub_debug |= SLAB_STORE_USER; 1010 break; 1011 case 't': 1012 slub_debug |= SLAB_TRACE; 1013 break; 1014 default: 1015 printk(KERN_ERR "slub_debug option '%c' " 1016 "unknown. skipped\n", *str); 1017 } 1018 } 1019 1020check_slabs: 1021 if (*str == ',') 1022 slub_debug_slabs = str + 1; 1023out: 1024 return 1; 1025} 1026 1027__setup("slub_debug", setup_slub_debug); 1028 1029static unsigned long kmem_cache_flags(unsigned long objsize, 1030 unsigned long flags, const char *name, 1031 void (*ctor)(void *)) 1032{ 1033 /* 1034 * Enable debugging if selected on the kernel commandline. 1035 */ 1036 if (slub_debug && (!slub_debug_slabs || 1037 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) 1038 flags |= slub_debug; 1039 1040 return flags; 1041} 1042#else 1043static inline void setup_object_debug(struct kmem_cache *s, 1044 struct page *page, void *object) {} 1045 1046static inline int alloc_debug_processing(struct kmem_cache *s, 1047 struct page *page, void *object, unsigned long addr) { return 0; } 1048 1049static inline int free_debug_processing(struct kmem_cache *s, 1050 struct page *page, void *object, unsigned long addr) { return 0; } 1051 1052static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1053 { return 1; } 1054static inline int check_object(struct kmem_cache *s, struct page *page, 1055 void *object, int active) { return 1; } 1056static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1057static inline unsigned long kmem_cache_flags(unsigned long objsize, 1058 unsigned long flags, const char *name, 1059 void (*ctor)(void *)) 1060{ 1061 return flags; 1062} 1063#define slub_debug 0 1064 1065#define disable_higher_order_debug 0 1066 1067static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1068 { return 0; } 1069static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1070 { return 0; } 1071static inline void inc_slabs_node(struct kmem_cache *s, int node, 1072 int objects) {} 1073static inline void dec_slabs_node(struct kmem_cache *s, int node, 1074 int objects) {} 1075#endif 1076 1077/* 1078 * Slab allocation and freeing 1079 */ 1080static inline struct page *alloc_slab_page(gfp_t flags, int node, 1081 struct kmem_cache_order_objects oo) 1082{ 1083 int order = oo_order(oo); 1084 1085 flags |= __GFP_NOTRACK; 1086 1087 if (node == -1) 1088 return alloc_pages(flags, order); 1089 else 1090 return alloc_pages_node(node, flags, order); 1091} 1092 1093static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1094{ 1095 struct page *page; 1096 struct kmem_cache_order_objects oo = s->oo; 1097 gfp_t alloc_gfp; 1098 1099 flags |= s->allocflags; 1100 1101 /* 1102 * Let the initial higher-order allocation fail under memory pressure 1103 * so we fall-back to the minimum order allocation. 1104 */ 1105 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1106 1107 page = alloc_slab_page(alloc_gfp, node, oo); 1108 if (unlikely(!page)) { 1109 oo = s->min; 1110 /* 1111 * Allocation may have failed due to fragmentation. 1112 * Try a lower order alloc if possible 1113 */ 1114 page = alloc_slab_page(flags, node, oo); 1115 if (!page) 1116 return NULL; 1117 1118 stat(this_cpu_ptr(s->cpu_slab), ORDER_FALLBACK); 1119 } 1120 1121 if (kmemcheck_enabled 1122 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1123 int pages = 1 << oo_order(oo); 1124 1125 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); 1126 1127 /* 1128 * Objects from caches that have a constructor don't get 1129 * cleared when they're allocated, so we need to do it here. 1130 */ 1131 if (s->ctor) 1132 kmemcheck_mark_uninitialized_pages(page, pages); 1133 else 1134 kmemcheck_mark_unallocated_pages(page, pages); 1135 } 1136 1137 page->objects = oo_objects(oo); 1138 mod_zone_page_state(page_zone(page), 1139 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1140 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1141 1 << oo_order(oo)); 1142 1143 return page; 1144} 1145 1146static void setup_object(struct kmem_cache *s, struct page *page, 1147 void *object) 1148{ 1149 setup_object_debug(s, page, object); 1150 if (unlikely(s->ctor)) 1151 s->ctor(object); 1152} 1153 1154static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1155{ 1156 struct page *page; 1157 void *start; 1158 void *last; 1159 void *p; 1160 1161 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1162 1163 page = allocate_slab(s, 1164 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1165 if (!page) 1166 goto out; 1167 1168 inc_slabs_node(s, page_to_nid(page), page->objects); 1169 page->slab = s; 1170 page->flags |= 1 << PG_slab; 1171 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | 1172 SLAB_STORE_USER | SLAB_TRACE)) 1173 __SetPageSlubDebug(page); 1174 1175 start = page_address(page); 1176 1177 if (unlikely(s->flags & SLAB_POISON)) 1178 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1179 1180 last = start; 1181 for_each_object(p, s, start, page->objects) { 1182 setup_object(s, page, last); 1183 set_freepointer(s, last, p); 1184 last = p; 1185 } 1186 setup_object(s, page, last); 1187 set_freepointer(s, last, NULL); 1188 1189 page->freelist = start; 1190 page->inuse = 0; 1191out: 1192 return page; 1193} 1194 1195static void __free_slab(struct kmem_cache *s, struct page *page) 1196{ 1197 int order = compound_order(page); 1198 int pages = 1 << order; 1199 1200 if (unlikely(SLABDEBUG && PageSlubDebug(page))) { 1201 void *p; 1202 1203 slab_pad_check(s, page); 1204 for_each_object(p, s, page_address(page), 1205 page->objects) 1206 check_object(s, page, p, 0); 1207 __ClearPageSlubDebug(page); 1208 } 1209 1210 kmemcheck_free_shadow(page, compound_order(page)); 1211 1212 mod_zone_page_state(page_zone(page), 1213 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1214 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1215 -pages); 1216 1217 __ClearPageSlab(page); 1218 reset_page_mapcount(page); 1219 if (current->reclaim_state) 1220 current->reclaim_state->reclaimed_slab += pages; 1221 __free_pages(page, order); 1222} 1223 1224static void rcu_free_slab(struct rcu_head *h) 1225{ 1226 struct page *page; 1227 1228 page = container_of((struct list_head *)h, struct page, lru); 1229 __free_slab(page->slab, page); 1230} 1231 1232static void free_slab(struct kmem_cache *s, struct page *page) 1233{ 1234 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1235 /* 1236 * RCU free overloads the RCU head over the LRU 1237 */ 1238 struct rcu_head *head = (void *)&page->lru; 1239 1240 call_rcu(head, rcu_free_slab); 1241 } else 1242 __free_slab(s, page); 1243} 1244 1245static void discard_slab(struct kmem_cache *s, struct page *page) 1246{ 1247 dec_slabs_node(s, page_to_nid(page), page->objects); 1248 free_slab(s, page); 1249} 1250 1251/* 1252 * Per slab locking using the pagelock 1253 */ 1254static __always_inline void slab_lock(struct page *page) 1255{ 1256 bit_spin_lock(PG_locked, &page->flags); 1257} 1258 1259static __always_inline void slab_unlock(struct page *page) 1260{ 1261 __bit_spin_unlock(PG_locked, &page->flags); 1262} 1263 1264static __always_inline int slab_trylock(struct page *page) 1265{ 1266 int rc = 1; 1267 1268 rc = bit_spin_trylock(PG_locked, &page->flags); 1269 return rc; 1270} 1271 1272/* 1273 * Management of partially allocated slabs 1274 */ 1275static void add_partial(struct kmem_cache_node *n, 1276 struct page *page, int tail) 1277{ 1278 spin_lock(&n->list_lock); 1279 n->nr_partial++; 1280 if (tail) 1281 list_add_tail(&page->lru, &n->partial); 1282 else 1283 list_add(&page->lru, &n->partial); 1284 spin_unlock(&n->list_lock); 1285} 1286 1287static void remove_partial(struct kmem_cache *s, struct page *page) 1288{ 1289 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1290 1291 spin_lock(&n->list_lock); 1292 list_del(&page->lru); 1293 n->nr_partial--; 1294 spin_unlock(&n->list_lock); 1295} 1296 1297/* 1298 * Lock slab and remove from the partial list. 1299 * 1300 * Must hold list_lock. 1301 */ 1302static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1303 struct page *page) 1304{ 1305 if (slab_trylock(page)) { 1306 list_del(&page->lru); 1307 n->nr_partial--; 1308 __SetPageSlubFrozen(page); 1309 return 1; 1310 } 1311 return 0; 1312} 1313 1314/* 1315 * Try to allocate a partial slab from a specific node. 1316 */ 1317static struct page *get_partial_node(struct kmem_cache_node *n) 1318{ 1319 struct page *page; 1320 1321 /* 1322 * Racy check. If we mistakenly see no partial slabs then we 1323 * just allocate an empty slab. If we mistakenly try to get a 1324 * partial slab and there is none available then get_partials() 1325 * will return NULL. 1326 */ 1327 if (!n || !n->nr_partial) 1328 return NULL; 1329 1330 spin_lock(&n->list_lock); 1331 list_for_each_entry(page, &n->partial, lru) 1332 if (lock_and_freeze_slab(n, page)) 1333 goto out; 1334 page = NULL; 1335out: 1336 spin_unlock(&n->list_lock); 1337 return page; 1338} 1339 1340/* 1341 * Get a page from somewhere. Search in increasing NUMA distances. 1342 */ 1343static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1344{ 1345#ifdef CONFIG_NUMA 1346 struct zonelist *zonelist; 1347 struct zoneref *z; 1348 struct zone *zone; 1349 enum zone_type high_zoneidx = gfp_zone(flags); 1350 struct page *page; 1351 1352 /* 1353 * The defrag ratio allows a configuration of the tradeoffs between 1354 * inter node defragmentation and node local allocations. A lower 1355 * defrag_ratio increases the tendency to do local allocations 1356 * instead of attempting to obtain partial slabs from other nodes. 1357 * 1358 * If the defrag_ratio is set to 0 then kmalloc() always 1359 * returns node local objects. If the ratio is higher then kmalloc() 1360 * may return off node objects because partial slabs are obtained 1361 * from other nodes and filled up. 1362 * 1363 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1364 * defrag_ratio = 1000) then every (well almost) allocation will 1365 * first attempt to defrag slab caches on other nodes. This means 1366 * scanning over all nodes to look for partial slabs which may be 1367 * expensive if we do it every time we are trying to find a slab 1368 * with available objects. 1369 */ 1370 if (!s->remote_node_defrag_ratio || 1371 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1372 return NULL; 1373 1374 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1375 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1376 struct kmem_cache_node *n; 1377 1378 n = get_node(s, zone_to_nid(zone)); 1379 1380 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1381 n->nr_partial > s->min_partial) { 1382 page = get_partial_node(n); 1383 if (page) 1384 return page; 1385 } 1386 } 1387#endif 1388 return NULL; 1389} 1390 1391/* 1392 * Get a partial page, lock it and return it. 1393 */ 1394static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1395{ 1396 struct page *page; 1397 int searchnode = (node == -1) ? numa_node_id() : node; 1398 1399 page = get_partial_node(get_node(s, searchnode)); 1400 if (page || (flags & __GFP_THISNODE)) 1401 return page; 1402 1403 return get_any_partial(s, flags); 1404} 1405 1406/* 1407 * Move a page back to the lists. 1408 * 1409 * Must be called with the slab lock held. 1410 * 1411 * On exit the slab lock will have been dropped. 1412 */ 1413static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1414{ 1415 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1416 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab); 1417 1418 __ClearPageSlubFrozen(page); 1419 if (page->inuse) { 1420 1421 if (page->freelist) { 1422 add_partial(n, page, tail); 1423 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1424 } else { 1425 stat(c, DEACTIVATE_FULL); 1426 if (SLABDEBUG && PageSlubDebug(page) && 1427 (s->flags & SLAB_STORE_USER)) 1428 add_full(n, page); 1429 } 1430 slab_unlock(page); 1431 } else { 1432 stat(c, DEACTIVATE_EMPTY); 1433 if (n->nr_partial < s->min_partial) { 1434 /* 1435 * Adding an empty slab to the partial slabs in order 1436 * to avoid page allocator overhead. This slab needs 1437 * to come after the other slabs with objects in 1438 * so that the others get filled first. That way the 1439 * size of the partial list stays small. 1440 * 1441 * kmem_cache_shrink can reclaim any empty slabs from 1442 * the partial list. 1443 */ 1444 add_partial(n, page, 1); 1445 slab_unlock(page); 1446 } else { 1447 slab_unlock(page); 1448 stat(__this_cpu_ptr(s->cpu_slab), FREE_SLAB); 1449 discard_slab(s, page); 1450 } 1451 } 1452} 1453 1454/* 1455 * Remove the cpu slab 1456 */ 1457static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1458{ 1459 struct page *page = c->page; 1460 int tail = 1; 1461 1462 if (page->freelist) 1463 stat(c, DEACTIVATE_REMOTE_FREES); 1464 /* 1465 * Merge cpu freelist into slab freelist. Typically we get here 1466 * because both freelists are empty. So this is unlikely 1467 * to occur. 1468 */ 1469 while (unlikely(c->freelist)) { 1470 void **object; 1471 1472 tail = 0; /* Hot objects. Put the slab first */ 1473 1474 /* Retrieve object from cpu_freelist */ 1475 object = c->freelist; 1476 c->freelist = c->freelist[c->offset]; 1477 1478 /* And put onto the regular freelist */ 1479 object[c->offset] = page->freelist; 1480 page->freelist = object; 1481 page->inuse--; 1482 } 1483 c->page = NULL; 1484 unfreeze_slab(s, page, tail); 1485} 1486 1487static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1488{ 1489 stat(c, CPUSLAB_FLUSH); 1490 slab_lock(c->page); 1491 deactivate_slab(s, c); 1492} 1493 1494/* 1495 * Flush cpu slab. 1496 * 1497 * Called from IPI handler with interrupts disabled. 1498 */ 1499static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1500{ 1501 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 1502 1503 if (likely(c && c->page)) 1504 flush_slab(s, c); 1505} 1506 1507static void flush_cpu_slab(void *d) 1508{ 1509 struct kmem_cache *s = d; 1510 1511 __flush_cpu_slab(s, smp_processor_id()); 1512} 1513 1514static void flush_all(struct kmem_cache *s) 1515{ 1516 on_each_cpu(flush_cpu_slab, s, 1); 1517} 1518 1519/* 1520 * Check if the objects in a per cpu structure fit numa 1521 * locality expectations. 1522 */ 1523static inline int node_match(struct kmem_cache_cpu *c, int node) 1524{ 1525#ifdef CONFIG_NUMA 1526 if (node != -1 && c->node != node) 1527 return 0; 1528#endif 1529 return 1; 1530} 1531 1532static int count_free(struct page *page) 1533{ 1534 return page->objects - page->inuse; 1535} 1536 1537static unsigned long count_partial(struct kmem_cache_node *n, 1538 int (*get_count)(struct page *)) 1539{ 1540 unsigned long flags; 1541 unsigned long x = 0; 1542 struct page *page; 1543 1544 spin_lock_irqsave(&n->list_lock, flags); 1545 list_for_each_entry(page, &n->partial, lru) 1546 x += get_count(page); 1547 spin_unlock_irqrestore(&n->list_lock, flags); 1548 return x; 1549} 1550 1551static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 1552{ 1553#ifdef CONFIG_SLUB_DEBUG 1554 return atomic_long_read(&n->total_objects); 1555#else 1556 return 0; 1557#endif 1558} 1559 1560static noinline void 1561slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 1562{ 1563 int node; 1564 1565 printk(KERN_WARNING 1566 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 1567 nid, gfpflags); 1568 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 1569 "default order: %d, min order: %d\n", s->name, s->objsize, 1570 s->size, oo_order(s->oo), oo_order(s->min)); 1571 1572 if (oo_order(s->min) > get_order(s->objsize)) 1573 printk(KERN_WARNING " %s debugging increased min order, use " 1574 "slub_debug=O to disable.\n", s->name); 1575 1576 for_each_online_node(node) { 1577 struct kmem_cache_node *n = get_node(s, node); 1578 unsigned long nr_slabs; 1579 unsigned long nr_objs; 1580 unsigned long nr_free; 1581 1582 if (!n) 1583 continue; 1584 1585 nr_free = count_partial(n, count_free); 1586 nr_slabs = node_nr_slabs(n); 1587 nr_objs = node_nr_objs(n); 1588 1589 printk(KERN_WARNING 1590 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 1591 node, nr_slabs, nr_objs, nr_free); 1592 } 1593} 1594 1595/* 1596 * Slow path. The lockless freelist is empty or we need to perform 1597 * debugging duties. 1598 * 1599 * Interrupts are disabled. 1600 * 1601 * Processing is still very fast if new objects have been freed to the 1602 * regular freelist. In that case we simply take over the regular freelist 1603 * as the lockless freelist and zap the regular freelist. 1604 * 1605 * If that is not working then we fall back to the partial lists. We take the 1606 * first element of the freelist as the object to allocate now and move the 1607 * rest of the freelist to the lockless freelist. 1608 * 1609 * And if we were unable to get a new slab from the partial slab lists then 1610 * we need to allocate a new slab. This is the slowest path since it involves 1611 * a call to the page allocator and the setup of a new slab. 1612 */ 1613static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 1614 unsigned long addr, struct kmem_cache_cpu *c) 1615{ 1616 void **object; 1617 struct page *new; 1618 1619 /* We handle __GFP_ZERO in the caller */ 1620 gfpflags &= ~__GFP_ZERO; 1621 1622 if (!c->page) 1623 goto new_slab; 1624 1625 slab_lock(c->page); 1626 if (unlikely(!node_match(c, node))) 1627 goto another_slab; 1628 1629 stat(c, ALLOC_REFILL); 1630 1631load_freelist: 1632 object = c->page->freelist; 1633 if (unlikely(!object)) 1634 goto another_slab; 1635 if (unlikely(SLABDEBUG && PageSlubDebug(c->page))) 1636 goto debug; 1637 1638 c->freelist = object[c->offset]; 1639 c->page->inuse = c->page->objects; 1640 c->page->freelist = NULL; 1641 c->node = page_to_nid(c->page); 1642unlock_out: 1643 slab_unlock(c->page); 1644 stat(c, ALLOC_SLOWPATH); 1645 return object; 1646 1647another_slab: 1648 deactivate_slab(s, c); 1649 1650new_slab: 1651 new = get_partial(s, gfpflags, node); 1652 if (new) { 1653 c->page = new; 1654 stat(c, ALLOC_FROM_PARTIAL); 1655 goto load_freelist; 1656 } 1657 1658 if (gfpflags & __GFP_WAIT) 1659 local_irq_enable(); 1660 1661 new = new_slab(s, gfpflags, node); 1662 1663 if (gfpflags & __GFP_WAIT) 1664 local_irq_disable(); 1665 1666 if (new) { 1667 c = __this_cpu_ptr(s->cpu_slab); 1668 stat(c, ALLOC_SLAB); 1669 if (c->page) 1670 flush_slab(s, c); 1671 slab_lock(new); 1672 __SetPageSlubFrozen(new); 1673 c->page = new; 1674 goto load_freelist; 1675 } 1676 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 1677 slab_out_of_memory(s, gfpflags, node); 1678 return NULL; 1679debug: 1680 if (!alloc_debug_processing(s, c->page, object, addr)) 1681 goto another_slab; 1682 1683 c->page->inuse++; 1684 c->page->freelist = object[c->offset]; 1685 c->node = -1; 1686 goto unlock_out; 1687} 1688 1689/* 1690 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1691 * have the fastpath folded into their functions. So no function call 1692 * overhead for requests that can be satisfied on the fastpath. 1693 * 1694 * The fastpath works by first checking if the lockless freelist can be used. 1695 * If not then __slab_alloc is called for slow processing. 1696 * 1697 * Otherwise we can simply pick the next object from the lockless free list. 1698 */ 1699static __always_inline void *slab_alloc(struct kmem_cache *s, 1700 gfp_t gfpflags, int node, unsigned long addr) 1701{ 1702 void **object; 1703 struct kmem_cache_cpu *c; 1704 unsigned long flags; 1705 unsigned long objsize; 1706 1707 gfpflags &= gfp_allowed_mask; 1708 1709 lockdep_trace_alloc(gfpflags); 1710 might_sleep_if(gfpflags & __GFP_WAIT); 1711 1712 if (should_failslab(s->objsize, gfpflags)) 1713 return NULL; 1714 1715 local_irq_save(flags); 1716 c = __this_cpu_ptr(s->cpu_slab); 1717 object = c->freelist; 1718 objsize = c->objsize; 1719 if (unlikely(!object || !node_match(c, node))) 1720 1721 object = __slab_alloc(s, gfpflags, node, addr, c); 1722 1723 else { 1724 c->freelist = object[c->offset]; 1725 stat(c, ALLOC_FASTPATH); 1726 } 1727 local_irq_restore(flags); 1728 1729 if (unlikely(gfpflags & __GFP_ZERO) && object) 1730 memset(object, 0, objsize); 1731 1732 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize); 1733 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags); 1734 1735 return object; 1736} 1737 1738void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1739{ 1740 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_); 1741 1742 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); 1743 1744 return ret; 1745} 1746EXPORT_SYMBOL(kmem_cache_alloc); 1747 1748#ifdef CONFIG_TRACING 1749void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags) 1750{ 1751 return slab_alloc(s, gfpflags, -1, _RET_IP_); 1752} 1753EXPORT_SYMBOL(kmem_cache_alloc_notrace); 1754#endif 1755 1756#ifdef CONFIG_NUMA 1757void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1758{ 1759 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 1760 1761 trace_kmem_cache_alloc_node(_RET_IP_, ret, 1762 s->objsize, s->size, gfpflags, node); 1763 1764 return ret; 1765} 1766EXPORT_SYMBOL(kmem_cache_alloc_node); 1767#endif 1768 1769#ifdef CONFIG_TRACING 1770void *kmem_cache_alloc_node_notrace(struct kmem_cache *s, 1771 gfp_t gfpflags, 1772 int node) 1773{ 1774 return slab_alloc(s, gfpflags, node, _RET_IP_); 1775} 1776EXPORT_SYMBOL(kmem_cache_alloc_node_notrace); 1777#endif 1778 1779/* 1780 * Slow patch handling. This may still be called frequently since objects 1781 * have a longer lifetime than the cpu slabs in most processing loads. 1782 * 1783 * So we still attempt to reduce cache line usage. Just take the slab 1784 * lock and free the item. If there is no additional partial page 1785 * handling required then we can return immediately. 1786 */ 1787static void __slab_free(struct kmem_cache *s, struct page *page, 1788 void *x, unsigned long addr, unsigned int offset) 1789{ 1790 void *prior; 1791 void **object = (void *)x; 1792 struct kmem_cache_cpu *c; 1793 1794 c = __this_cpu_ptr(s->cpu_slab); 1795 stat(c, FREE_SLOWPATH); 1796 slab_lock(page); 1797 1798 if (unlikely(SLABDEBUG && PageSlubDebug(page))) 1799 goto debug; 1800 1801checks_ok: 1802 prior = object[offset] = page->freelist; 1803 page->freelist = object; 1804 page->inuse--; 1805 1806 if (unlikely(PageSlubFrozen(page))) { 1807 stat(c, FREE_FROZEN); 1808 goto out_unlock; 1809 } 1810 1811 if (unlikely(!page->inuse)) 1812 goto slab_empty; 1813 1814 /* 1815 * Objects left in the slab. If it was not on the partial list before 1816 * then add it. 1817 */ 1818 if (unlikely(!prior)) { 1819 add_partial(get_node(s, page_to_nid(page)), page, 1); 1820 stat(c, FREE_ADD_PARTIAL); 1821 } 1822 1823out_unlock: 1824 slab_unlock(page); 1825 return; 1826 1827slab_empty: 1828 if (prior) { 1829 /* 1830 * Slab still on the partial list. 1831 */ 1832 remove_partial(s, page); 1833 stat(c, FREE_REMOVE_PARTIAL); 1834 } 1835 slab_unlock(page); 1836 stat(c, FREE_SLAB); 1837 discard_slab(s, page); 1838 return; 1839 1840debug: 1841 if (!free_debug_processing(s, page, x, addr)) 1842 goto out_unlock; 1843 goto checks_ok; 1844} 1845 1846/* 1847 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1848 * can perform fastpath freeing without additional function calls. 1849 * 1850 * The fastpath is only possible if we are freeing to the current cpu slab 1851 * of this processor. This typically the case if we have just allocated 1852 * the item before. 1853 * 1854 * If fastpath is not possible then fall back to __slab_free where we deal 1855 * with all sorts of special processing. 1856 */ 1857static __always_inline void slab_free(struct kmem_cache *s, 1858 struct page *page, void *x, unsigned long addr) 1859{ 1860 void **object = (void *)x; 1861 struct kmem_cache_cpu *c; 1862 unsigned long flags; 1863 1864 kmemleak_free_recursive(x, s->flags); 1865 local_irq_save(flags); 1866 c = __this_cpu_ptr(s->cpu_slab); 1867 kmemcheck_slab_free(s, object, c->objsize); 1868 debug_check_no_locks_freed(object, c->objsize); 1869 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1870 debug_check_no_obj_freed(object, c->objsize); 1871 if (likely(page == c->page && c->node >= 0)) { 1872 object[c->offset] = c->freelist; 1873 c->freelist = object; 1874 stat(c, FREE_FASTPATH); 1875 } else 1876 __slab_free(s, page, x, addr, c->offset); 1877 1878 local_irq_restore(flags); 1879} 1880 1881void kmem_cache_free(struct kmem_cache *s, void *x) 1882{ 1883 struct page *page; 1884 1885 page = virt_to_head_page(x); 1886 1887 slab_free(s, page, x, _RET_IP_); 1888 1889 trace_kmem_cache_free(_RET_IP_, x); 1890} 1891EXPORT_SYMBOL(kmem_cache_free); 1892 1893/* Figure out on which slab page the object resides */ 1894static struct page *get_object_page(const void *x) 1895{ 1896 struct page *page = virt_to_head_page(x); 1897 1898 if (!PageSlab(page)) 1899 return NULL; 1900 1901 return page; 1902} 1903 1904/* 1905 * Object placement in a slab is made very easy because we always start at 1906 * offset 0. If we tune the size of the object to the alignment then we can 1907 * get the required alignment by putting one properly sized object after 1908 * another. 1909 * 1910 * Notice that the allocation order determines the sizes of the per cpu 1911 * caches. Each processor has always one slab available for allocations. 1912 * Increasing the allocation order reduces the number of times that slabs 1913 * must be moved on and off the partial lists and is therefore a factor in 1914 * locking overhead. 1915 */ 1916 1917/* 1918 * Mininum / Maximum order of slab pages. This influences locking overhead 1919 * and slab fragmentation. A higher order reduces the number of partial slabs 1920 * and increases the number of allocations possible without having to 1921 * take the list_lock. 1922 */ 1923static int slub_min_order; 1924static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 1925static int slub_min_objects; 1926 1927/* 1928 * Merge control. If this is set then no merging of slab caches will occur. 1929 * (Could be removed. This was introduced to pacify the merge skeptics.) 1930 */ 1931static int slub_nomerge; 1932 1933/* 1934 * Calculate the order of allocation given an slab object size. 1935 * 1936 * The order of allocation has significant impact on performance and other 1937 * system components. Generally order 0 allocations should be preferred since 1938 * order 0 does not cause fragmentation in the page allocator. Larger objects 1939 * be problematic to put into order 0 slabs because there may be too much 1940 * unused space left. We go to a higher order if more than 1/16th of the slab 1941 * would be wasted. 1942 * 1943 * In order to reach satisfactory performance we must ensure that a minimum 1944 * number of objects is in one slab. Otherwise we may generate too much 1945 * activity on the partial lists which requires taking the list_lock. This is 1946 * less a concern for large slabs though which are rarely used. 1947 * 1948 * slub_max_order specifies the order where we begin to stop considering the 1949 * number of objects in a slab as critical. If we reach slub_max_order then 1950 * we try to keep the page order as low as possible. So we accept more waste 1951 * of space in favor of a small page order. 1952 * 1953 * Higher order allocations also allow the placement of more objects in a 1954 * slab and thereby reduce object handling overhead. If the user has 1955 * requested a higher mininum order then we start with that one instead of 1956 * the smallest order which will fit the object. 1957 */ 1958static inline int slab_order(int size, int min_objects, 1959 int max_order, int fract_leftover) 1960{ 1961 int order; 1962 int rem; 1963 int min_order = slub_min_order; 1964 1965 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE) 1966 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 1967 1968 for (order = max(min_order, 1969 fls(min_objects * size - 1) - PAGE_SHIFT); 1970 order <= max_order; order++) { 1971 1972 unsigned long slab_size = PAGE_SIZE << order; 1973 1974 if (slab_size < min_objects * size) 1975 continue; 1976 1977 rem = slab_size % size; 1978 1979 if (rem <= slab_size / fract_leftover) 1980 break; 1981 1982 } 1983 1984 return order; 1985} 1986 1987static inline int calculate_order(int size) 1988{ 1989 int order; 1990 int min_objects; 1991 int fraction; 1992 int max_objects; 1993 1994 /* 1995 * Attempt to find best configuration for a slab. This 1996 * works by first attempting to generate a layout with 1997 * the best configuration and backing off gradually. 1998 * 1999 * First we reduce the acceptable waste in a slab. Then 2000 * we reduce the minimum objects required in a slab. 2001 */ 2002 min_objects = slub_min_objects; 2003 if (!min_objects) 2004 min_objects = 4 * (fls(nr_cpu_ids) + 1); 2005 max_objects = (PAGE_SIZE << slub_max_order)/size; 2006 min_objects = min(min_objects, max_objects); 2007 2008 while (min_objects > 1) { 2009 fraction = 16; 2010 while (fraction >= 4) { 2011 order = slab_order(size, min_objects, 2012 slub_max_order, fraction); 2013 if (order <= slub_max_order) 2014 return order; 2015 fraction /= 2; 2016 } 2017 min_objects--; 2018 } 2019 2020 /* 2021 * We were unable to place multiple objects in a slab. Now 2022 * lets see if we can place a single object there. 2023 */ 2024 order = slab_order(size, 1, slub_max_order, 1); 2025 if (order <= slub_max_order) 2026 return order; 2027 2028 /* 2029 * Doh this slab cannot be placed using slub_max_order. 2030 */ 2031 order = slab_order(size, 1, MAX_ORDER, 1); 2032 if (order < MAX_ORDER) 2033 return order; 2034 return -ENOSYS; 2035} 2036 2037/* 2038 * Figure out what the alignment of the objects will be. 2039 */ 2040static unsigned long calculate_alignment(unsigned long flags, 2041 unsigned long align, unsigned long size) 2042{ 2043 /* 2044 * If the user wants hardware cache aligned objects then follow that 2045 * suggestion if the object is sufficiently large. 2046 * 2047 * The hardware cache alignment cannot override the specified 2048 * alignment though. If that is greater then use it. 2049 */ 2050 if (flags & SLAB_HWCACHE_ALIGN) { 2051 unsigned long ralign = cache_line_size(); 2052 while (size <= ralign / 2) 2053 ralign /= 2; 2054 align = max(align, ralign); 2055 } 2056 2057 if (align < ARCH_SLAB_MINALIGN) 2058 align = ARCH_SLAB_MINALIGN; 2059 2060 return ALIGN(align, sizeof(void *)); 2061} 2062 2063static void init_kmem_cache_cpu(struct kmem_cache *s, 2064 struct kmem_cache_cpu *c) 2065{ 2066 c->page = NULL; 2067 c->freelist = NULL; 2068 c->node = 0; 2069 c->offset = s->offset / sizeof(void *); 2070 c->objsize = s->objsize; 2071#ifdef CONFIG_SLUB_STATS 2072 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); 2073#endif 2074} 2075 2076static void 2077init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 2078{ 2079 n->nr_partial = 0; 2080 spin_lock_init(&n->list_lock); 2081 INIT_LIST_HEAD(&n->partial); 2082#ifdef CONFIG_SLUB_DEBUG 2083 atomic_long_set(&n->nr_slabs, 0); 2084 atomic_long_set(&n->total_objects, 0); 2085 INIT_LIST_HEAD(&n->full); 2086#endif 2087} 2088 2089static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[SLUB_PAGE_SHIFT]); 2090 2091static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2092{ 2093 int cpu; 2094 2095 if (s < kmalloc_caches + KMALLOC_CACHES && s >= kmalloc_caches) 2096 /* 2097 * Boot time creation of the kmalloc array. Use static per cpu data 2098 * since the per cpu allocator is not available yet. 2099 */ 2100 s->cpu_slab = per_cpu_var(kmalloc_percpu) + (s - kmalloc_caches); 2101 else 2102 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu); 2103 2104 if (!s->cpu_slab) 2105 return 0; 2106 2107 for_each_possible_cpu(cpu) 2108 init_kmem_cache_cpu(s, per_cpu_ptr(s->cpu_slab, cpu)); 2109 return 1; 2110} 2111 2112#ifdef CONFIG_NUMA 2113/* 2114 * No kmalloc_node yet so do it by hand. We know that this is the first 2115 * slab on the node for this slabcache. There are no concurrent accesses 2116 * possible. 2117 * 2118 * Note that this function only works on the kmalloc_node_cache 2119 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2120 * memory on a fresh node that has no slab structures yet. 2121 */ 2122static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node) 2123{ 2124 struct page *page; 2125 struct kmem_cache_node *n; 2126 unsigned long flags; 2127 2128 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2129 2130 page = new_slab(kmalloc_caches, gfpflags, node); 2131 2132 BUG_ON(!page); 2133 if (page_to_nid(page) != node) { 2134 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2135 "node %d\n", node); 2136 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2137 "in order to be able to continue\n"); 2138 } 2139 2140 n = page->freelist; 2141 BUG_ON(!n); 2142 page->freelist = get_freepointer(kmalloc_caches, n); 2143 page->inuse++; 2144 kmalloc_caches->node[node] = n; 2145#ifdef CONFIG_SLUB_DEBUG 2146 init_object(kmalloc_caches, n, 1); 2147 init_tracking(kmalloc_caches, n); 2148#endif 2149 init_kmem_cache_node(n, kmalloc_caches); 2150 inc_slabs_node(kmalloc_caches, node, page->objects); 2151 2152 /* 2153 * lockdep requires consistent irq usage for each lock 2154 * so even though there cannot be a race this early in 2155 * the boot sequence, we still disable irqs. 2156 */ 2157 local_irq_save(flags); 2158 add_partial(n, page, 0); 2159 local_irq_restore(flags); 2160} 2161 2162static void free_kmem_cache_nodes(struct kmem_cache *s) 2163{ 2164 int node; 2165 2166 for_each_node_state(node, N_NORMAL_MEMORY) { 2167 struct kmem_cache_node *n = s->node[node]; 2168 if (n && n != &s->local_node) 2169 kmem_cache_free(kmalloc_caches, n); 2170 s->node[node] = NULL; 2171 } 2172} 2173 2174static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2175{ 2176 int node; 2177 int local_node; 2178 2179 if (slab_state >= UP) 2180 local_node = page_to_nid(virt_to_page(s)); 2181 else 2182 local_node = 0; 2183 2184 for_each_node_state(node, N_NORMAL_MEMORY) { 2185 struct kmem_cache_node *n; 2186 2187 if (local_node == node) 2188 n = &s->local_node; 2189 else { 2190 if (slab_state == DOWN) { 2191 early_kmem_cache_node_alloc(gfpflags, node); 2192 continue; 2193 } 2194 n = kmem_cache_alloc_node(kmalloc_caches, 2195 gfpflags, node); 2196 2197 if (!n) { 2198 free_kmem_cache_nodes(s); 2199 return 0; 2200 } 2201 2202 } 2203 s->node[node] = n; 2204 init_kmem_cache_node(n, s); 2205 } 2206 return 1; 2207} 2208#else 2209static void free_kmem_cache_nodes(struct kmem_cache *s) 2210{ 2211} 2212 2213static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2214{ 2215 init_kmem_cache_node(&s->local_node, s); 2216 return 1; 2217} 2218#endif 2219 2220static void set_min_partial(struct kmem_cache *s, unsigned long min) 2221{ 2222 if (min < MIN_PARTIAL) 2223 min = MIN_PARTIAL; 2224 else if (min > MAX_PARTIAL) 2225 min = MAX_PARTIAL; 2226 s->min_partial = min; 2227} 2228 2229/* 2230 * calculate_sizes() determines the order and the distribution of data within 2231 * a slab object. 2232 */ 2233static int calculate_sizes(struct kmem_cache *s, int forced_order) 2234{ 2235 unsigned long flags = s->flags; 2236 unsigned long size = s->objsize; 2237 unsigned long align = s->align; 2238 int order; 2239 2240 /* 2241 * Round up object size to the next word boundary. We can only 2242 * place the free pointer at word boundaries and this determines 2243 * the possible location of the free pointer. 2244 */ 2245 size = ALIGN(size, sizeof(void *)); 2246 2247#ifdef CONFIG_SLUB_DEBUG 2248 /* 2249 * Determine if we can poison the object itself. If the user of 2250 * the slab may touch the object after free or before allocation 2251 * then we should never poison the object itself. 2252 */ 2253 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2254 !s->ctor) 2255 s->flags |= __OBJECT_POISON; 2256 else 2257 s->flags &= ~__OBJECT_POISON; 2258 2259 2260 /* 2261 * If we are Redzoning then check if there is some space between the 2262 * end of the object and the free pointer. If not then add an 2263 * additional word to have some bytes to store Redzone information. 2264 */ 2265 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2266 size += sizeof(void *); 2267#endif 2268 2269 /* 2270 * With that we have determined the number of bytes in actual use 2271 * by the object. This is the potential offset to the free pointer. 2272 */ 2273 s->inuse = size; 2274 2275 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2276 s->ctor)) { 2277 /* 2278 * Relocate free pointer after the object if it is not 2279 * permitted to overwrite the first word of the object on 2280 * kmem_cache_free. 2281 * 2282 * This is the case if we do RCU, have a constructor or 2283 * destructor or are poisoning the objects. 2284 */ 2285 s->offset = size; 2286 size += sizeof(void *); 2287 } 2288 2289#ifdef CONFIG_SLUB_DEBUG 2290 if (flags & SLAB_STORE_USER) 2291 /* 2292 * Need to store information about allocs and frees after 2293 * the object. 2294 */ 2295 size += 2 * sizeof(struct track); 2296 2297 if (flags & SLAB_RED_ZONE) 2298 /* 2299 * Add some empty padding so that we can catch 2300 * overwrites from earlier objects rather than let 2301 * tracking information or the free pointer be 2302 * corrupted if a user writes before the start 2303 * of the object. 2304 */ 2305 size += sizeof(void *); 2306#endif 2307 2308 /* 2309 * Determine the alignment based on various parameters that the 2310 * user specified and the dynamic determination of cache line size 2311 * on bootup. 2312 */ 2313 align = calculate_alignment(flags, align, s->objsize); 2314 s->align = align; 2315 2316 /* 2317 * SLUB stores one object immediately after another beginning from 2318 * offset 0. In order to align the objects we have to simply size 2319 * each object to conform to the alignment. 2320 */ 2321 size = ALIGN(size, align); 2322 s->size = size; 2323 if (forced_order >= 0) 2324 order = forced_order; 2325 else 2326 order = calculate_order(size); 2327 2328 if (order < 0) 2329 return 0; 2330 2331 s->allocflags = 0; 2332 if (order) 2333 s->allocflags |= __GFP_COMP; 2334 2335 if (s->flags & SLAB_CACHE_DMA) 2336 s->allocflags |= SLUB_DMA; 2337 2338 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2339 s->allocflags |= __GFP_RECLAIMABLE; 2340 2341 /* 2342 * Determine the number of objects per slab 2343 */ 2344 s->oo = oo_make(order, size); 2345 s->min = oo_make(get_order(size), size); 2346 if (oo_objects(s->oo) > oo_objects(s->max)) 2347 s->max = s->oo; 2348 2349 return !!oo_objects(s->oo); 2350 2351} 2352 2353static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2354 const char *name, size_t size, 2355 size_t align, unsigned long flags, 2356 void (*ctor)(void *)) 2357{ 2358 memset(s, 0, kmem_size); 2359 s->name = name; 2360 s->ctor = ctor; 2361 s->objsize = size; 2362 s->align = align; 2363 s->flags = kmem_cache_flags(size, flags, name, ctor); 2364 2365 if (!calculate_sizes(s, -1)) 2366 goto error; 2367 if (disable_higher_order_debug) { 2368 /* 2369 * Disable debugging flags that store metadata if the min slab 2370 * order increased. 2371 */ 2372 if (get_order(s->size) > get_order(s->objsize)) { 2373 s->flags &= ~DEBUG_METADATA_FLAGS; 2374 s->offset = 0; 2375 if (!calculate_sizes(s, -1)) 2376 goto error; 2377 } 2378 } 2379 2380 /* 2381 * The larger the object size is, the more pages we want on the partial 2382 * list to avoid pounding the page allocator excessively. 2383 */ 2384 set_min_partial(s, ilog2(s->size)); 2385 s->refcount = 1; 2386#ifdef CONFIG_NUMA 2387 s->remote_node_defrag_ratio = 1000; 2388#endif 2389 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2390 goto error; 2391 2392 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2393 return 1; 2394 free_kmem_cache_nodes(s); 2395error: 2396 if (flags & SLAB_PANIC) 2397 panic("Cannot create slab %s size=%lu realsize=%u " 2398 "order=%u offset=%u flags=%lx\n", 2399 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2400 s->offset, flags); 2401 return 0; 2402} 2403 2404/* 2405 * Check if a given pointer is valid 2406 */ 2407int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2408{ 2409 struct page *page; 2410 2411 page = get_object_page(object); 2412 2413 if (!page || s != page->slab) 2414 /* No slab or wrong slab */ 2415 return 0; 2416 2417 if (!check_valid_pointer(s, page, object)) 2418 return 0; 2419 2420 /* 2421 * We could also check if the object is on the slabs freelist. 2422 * But this would be too expensive and it seems that the main 2423 * purpose of kmem_ptr_valid() is to check if the object belongs 2424 * to a certain slab. 2425 */ 2426 return 1; 2427} 2428EXPORT_SYMBOL(kmem_ptr_validate); 2429 2430/* 2431 * Determine the size of a slab object 2432 */ 2433unsigned int kmem_cache_size(struct kmem_cache *s) 2434{ 2435 return s->objsize; 2436} 2437EXPORT_SYMBOL(kmem_cache_size); 2438 2439const char *kmem_cache_name(struct kmem_cache *s) 2440{ 2441 return s->name; 2442} 2443EXPORT_SYMBOL(kmem_cache_name); 2444 2445static void list_slab_objects(struct kmem_cache *s, struct page *page, 2446 const char *text) 2447{ 2448#ifdef CONFIG_SLUB_DEBUG 2449 void *addr = page_address(page); 2450 void *p; 2451 DECLARE_BITMAP(map, page->objects); 2452 2453 bitmap_zero(map, page->objects); 2454 slab_err(s, page, "%s", text); 2455 slab_lock(page); 2456 for_each_free_object(p, s, page->freelist) 2457 set_bit(slab_index(p, s, addr), map); 2458 2459 for_each_object(p, s, addr, page->objects) { 2460 2461 if (!test_bit(slab_index(p, s, addr), map)) { 2462 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2463 p, p - addr); 2464 print_tracking(s, p); 2465 } 2466 } 2467 slab_unlock(page); 2468#endif 2469} 2470 2471/* 2472 * Attempt to free all partial slabs on a node. 2473 */ 2474static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2475{ 2476 unsigned long flags; 2477 struct page *page, *h; 2478 2479 spin_lock_irqsave(&n->list_lock, flags); 2480 list_for_each_entry_safe(page, h, &n->partial, lru) { 2481 if (!page->inuse) { 2482 list_del(&page->lru); 2483 discard_slab(s, page); 2484 n->nr_partial--; 2485 } else { 2486 list_slab_objects(s, page, 2487 "Objects remaining on kmem_cache_close()"); 2488 } 2489 } 2490 spin_unlock_irqrestore(&n->list_lock, flags); 2491} 2492 2493/* 2494 * Release all resources used by a slab cache. 2495 */ 2496static inline int kmem_cache_close(struct kmem_cache *s) 2497{ 2498 int node; 2499 2500 flush_all(s); 2501 free_percpu(s->cpu_slab); 2502 /* Attempt to free all objects */ 2503 for_each_node_state(node, N_NORMAL_MEMORY) { 2504 struct kmem_cache_node *n = get_node(s, node); 2505 2506 free_partial(s, n); 2507 if (n->nr_partial || slabs_node(s, node)) 2508 return 1; 2509 } 2510 free_kmem_cache_nodes(s); 2511 return 0; 2512} 2513 2514/* 2515 * Close a cache and release the kmem_cache structure 2516 * (must be used for caches created using kmem_cache_create) 2517 */ 2518void kmem_cache_destroy(struct kmem_cache *s) 2519{ 2520 down_write(&slub_lock); 2521 s->refcount--; 2522 if (!s->refcount) { 2523 list_del(&s->list); 2524 up_write(&slub_lock); 2525 if (kmem_cache_close(s)) { 2526 printk(KERN_ERR "SLUB %s: %s called for cache that " 2527 "still has objects.\n", s->name, __func__); 2528 dump_stack(); 2529 } 2530 if (s->flags & SLAB_DESTROY_BY_RCU) 2531 rcu_barrier(); 2532 sysfs_slab_remove(s); 2533 } else 2534 up_write(&slub_lock); 2535} 2536EXPORT_SYMBOL(kmem_cache_destroy); 2537 2538/******************************************************************** 2539 * Kmalloc subsystem 2540 *******************************************************************/ 2541 2542struct kmem_cache kmalloc_caches[KMALLOC_CACHES] __cacheline_aligned; 2543EXPORT_SYMBOL(kmalloc_caches); 2544 2545static int __init setup_slub_min_order(char *str) 2546{ 2547 get_option(&str, &slub_min_order); 2548 2549 return 1; 2550} 2551 2552__setup("slub_min_order=", setup_slub_min_order); 2553 2554static int __init setup_slub_max_order(char *str) 2555{ 2556 get_option(&str, &slub_max_order); 2557 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 2558 2559 return 1; 2560} 2561 2562__setup("slub_max_order=", setup_slub_max_order); 2563 2564static int __init setup_slub_min_objects(char *str) 2565{ 2566 get_option(&str, &slub_min_objects); 2567 2568 return 1; 2569} 2570 2571__setup("slub_min_objects=", setup_slub_min_objects); 2572 2573static int __init setup_slub_nomerge(char *str) 2574{ 2575 slub_nomerge = 1; 2576 return 1; 2577} 2578 2579__setup("slub_nomerge", setup_slub_nomerge); 2580 2581static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2582 const char *name, int size, gfp_t gfp_flags) 2583{ 2584 unsigned int flags = 0; 2585 2586 if (gfp_flags & SLUB_DMA) 2587 flags = SLAB_CACHE_DMA; 2588 2589 /* 2590 * This function is called with IRQs disabled during early-boot on 2591 * single CPU so there's no need to take slub_lock here. 2592 */ 2593 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2594 flags, NULL)) 2595 goto panic; 2596 2597 list_add(&s->list, &slab_caches); 2598 2599 if (sysfs_slab_add(s)) 2600 goto panic; 2601 return s; 2602 2603panic: 2604 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2605} 2606 2607#ifdef CONFIG_ZONE_DMA 2608static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT]; 2609 2610static void sysfs_add_func(struct work_struct *w) 2611{ 2612 struct kmem_cache *s; 2613 2614 down_write(&slub_lock); 2615 list_for_each_entry(s, &slab_caches, list) { 2616 if (s->flags & __SYSFS_ADD_DEFERRED) { 2617 s->flags &= ~__SYSFS_ADD_DEFERRED; 2618 sysfs_slab_add(s); 2619 } 2620 } 2621 up_write(&slub_lock); 2622} 2623 2624static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2625 2626static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2627{ 2628 struct kmem_cache *s; 2629 char *text; 2630 size_t realsize; 2631 unsigned long slabflags; 2632 int i; 2633 2634 s = kmalloc_caches_dma[index]; 2635 if (s) 2636 return s; 2637 2638 /* Dynamically create dma cache */ 2639 if (flags & __GFP_WAIT) 2640 down_write(&slub_lock); 2641 else { 2642 if (!down_write_trylock(&slub_lock)) 2643 goto out; 2644 } 2645 2646 if (kmalloc_caches_dma[index]) 2647 goto unlock_out; 2648 2649 realsize = kmalloc_caches[index].objsize; 2650 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", 2651 (unsigned int)realsize); 2652 2653 s = NULL; 2654 for (i = 0; i < KMALLOC_CACHES; i++) 2655 if (!kmalloc_caches[i].size) 2656 break; 2657 2658 BUG_ON(i >= KMALLOC_CACHES); 2659 s = kmalloc_caches + i; 2660 2661 /* 2662 * Must defer sysfs creation to a workqueue because we don't know 2663 * what context we are called from. Before sysfs comes up, we don't 2664 * need to do anything because our sysfs initcall will start by 2665 * adding all existing slabs to sysfs. 2666 */ 2667 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK; 2668 if (slab_state >= SYSFS) 2669 slabflags |= __SYSFS_ADD_DEFERRED; 2670 2671 if (!s || !text || !kmem_cache_open(s, flags, text, 2672 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) { 2673 s->size = 0; 2674 kfree(text); 2675 goto unlock_out; 2676 } 2677 2678 list_add(&s->list, &slab_caches); 2679 kmalloc_caches_dma[index] = s; 2680 2681 if (slab_state >= SYSFS) 2682 schedule_work(&sysfs_add_work); 2683 2684unlock_out: 2685 up_write(&slub_lock); 2686out: 2687 return kmalloc_caches_dma[index]; 2688} 2689#endif 2690 2691/* 2692 * Conversion table for small slabs sizes / 8 to the index in the 2693 * kmalloc array. This is necessary for slabs < 192 since we have non power 2694 * of two cache sizes there. The size of larger slabs can be determined using 2695 * fls. 2696 */ 2697static s8 size_index[24] = { 2698 3, /* 8 */ 2699 4, /* 16 */ 2700 5, /* 24 */ 2701 5, /* 32 */ 2702 6, /* 40 */ 2703 6, /* 48 */ 2704 6, /* 56 */ 2705 6, /* 64 */ 2706 1, /* 72 */ 2707 1, /* 80 */ 2708 1, /* 88 */ 2709 1, /* 96 */ 2710 7, /* 104 */ 2711 7, /* 112 */ 2712 7, /* 120 */ 2713 7, /* 128 */ 2714 2, /* 136 */ 2715 2, /* 144 */ 2716 2, /* 152 */ 2717 2, /* 160 */ 2718 2, /* 168 */ 2719 2, /* 176 */ 2720 2, /* 184 */ 2721 2 /* 192 */ 2722}; 2723 2724static inline int size_index_elem(size_t bytes) 2725{ 2726 return (bytes - 1) / 8; 2727} 2728 2729static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2730{ 2731 int index; 2732 2733 if (size <= 192) { 2734 if (!size) 2735 return ZERO_SIZE_PTR; 2736 2737 index = size_index[size_index_elem(size)]; 2738 } else 2739 index = fls(size - 1); 2740 2741#ifdef CONFIG_ZONE_DMA 2742 if (unlikely((flags & SLUB_DMA))) 2743 return dma_kmalloc_cache(index, flags); 2744 2745#endif 2746 return &kmalloc_caches[index]; 2747} 2748 2749void *__kmalloc(size_t size, gfp_t flags) 2750{ 2751 struct kmem_cache *s; 2752 void *ret; 2753 2754 if (unlikely(size > SLUB_MAX_SIZE)) 2755 return kmalloc_large(size, flags); 2756 2757 s = get_slab(size, flags); 2758 2759 if (unlikely(ZERO_OR_NULL_PTR(s))) 2760 return s; 2761 2762 ret = slab_alloc(s, flags, -1, _RET_IP_); 2763 2764 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 2765 2766 return ret; 2767} 2768EXPORT_SYMBOL(__kmalloc); 2769 2770static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2771{ 2772 struct page *page; 2773 void *ptr = NULL; 2774 2775 flags |= __GFP_COMP | __GFP_NOTRACK; 2776 page = alloc_pages_node(node, flags, get_order(size)); 2777 if (page) 2778 ptr = page_address(page); 2779 2780 kmemleak_alloc(ptr, size, 1, flags); 2781 return ptr; 2782} 2783 2784#ifdef CONFIG_NUMA 2785void *__kmalloc_node(size_t size, gfp_t flags, int node) 2786{ 2787 struct kmem_cache *s; 2788 void *ret; 2789 2790 if (unlikely(size > SLUB_MAX_SIZE)) { 2791 ret = kmalloc_large_node(size, flags, node); 2792 2793 trace_kmalloc_node(_RET_IP_, ret, 2794 size, PAGE_SIZE << get_order(size), 2795 flags, node); 2796 2797 return ret; 2798 } 2799 2800 s = get_slab(size, flags); 2801 2802 if (unlikely(ZERO_OR_NULL_PTR(s))) 2803 return s; 2804 2805 ret = slab_alloc(s, flags, node, _RET_IP_); 2806 2807 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 2808 2809 return ret; 2810} 2811EXPORT_SYMBOL(__kmalloc_node); 2812#endif 2813 2814size_t ksize(const void *object) 2815{ 2816 struct page *page; 2817 struct kmem_cache *s; 2818 2819 if (unlikely(object == ZERO_SIZE_PTR)) 2820 return 0; 2821 2822 page = virt_to_head_page(object); 2823 2824 if (unlikely(!PageSlab(page))) { 2825 WARN_ON(!PageCompound(page)); 2826 return PAGE_SIZE << compound_order(page); 2827 } 2828 s = page->slab; 2829 2830#ifdef CONFIG_SLUB_DEBUG 2831 /* 2832 * Debugging requires use of the padding between object 2833 * and whatever may come after it. 2834 */ 2835 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2836 return s->objsize; 2837 2838#endif 2839 /* 2840 * If we have the need to store the freelist pointer 2841 * back there or track user information then we can 2842 * only use the space before that information. 2843 */ 2844 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2845 return s->inuse; 2846 /* 2847 * Else we can use all the padding etc for the allocation 2848 */ 2849 return s->size; 2850} 2851EXPORT_SYMBOL(ksize); 2852 2853void kfree(const void *x) 2854{ 2855 struct page *page; 2856 void *object = (void *)x; 2857 2858 trace_kfree(_RET_IP_, x); 2859 2860 if (unlikely(ZERO_OR_NULL_PTR(x))) 2861 return; 2862 2863 page = virt_to_head_page(x); 2864 if (unlikely(!PageSlab(page))) { 2865 BUG_ON(!PageCompound(page)); 2866 kmemleak_free(x); 2867 put_page(page); 2868 return; 2869 } 2870 slab_free(page->slab, page, object, _RET_IP_); 2871} 2872EXPORT_SYMBOL(kfree); 2873 2874/* 2875 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2876 * the remaining slabs by the number of items in use. The slabs with the 2877 * most items in use come first. New allocations will then fill those up 2878 * and thus they can be removed from the partial lists. 2879 * 2880 * The slabs with the least items are placed last. This results in them 2881 * being allocated from last increasing the chance that the last objects 2882 * are freed in them. 2883 */ 2884int kmem_cache_shrink(struct kmem_cache *s) 2885{ 2886 int node; 2887 int i; 2888 struct kmem_cache_node *n; 2889 struct page *page; 2890 struct page *t; 2891 int objects = oo_objects(s->max); 2892 struct list_head *slabs_by_inuse = 2893 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 2894 unsigned long flags; 2895 2896 if (!slabs_by_inuse) 2897 return -ENOMEM; 2898 2899 flush_all(s); 2900 for_each_node_state(node, N_NORMAL_MEMORY) { 2901 n = get_node(s, node); 2902 2903 if (!n->nr_partial) 2904 continue; 2905 2906 for (i = 0; i < objects; i++) 2907 INIT_LIST_HEAD(slabs_by_inuse + i); 2908 2909 spin_lock_irqsave(&n->list_lock, flags); 2910 2911 /* 2912 * Build lists indexed by the items in use in each slab. 2913 * 2914 * Note that concurrent frees may occur while we hold the 2915 * list_lock. page->inuse here is the upper limit. 2916 */ 2917 list_for_each_entry_safe(page, t, &n->partial, lru) { 2918 if (!page->inuse && slab_trylock(page)) { 2919 /* 2920 * Must hold slab lock here because slab_free 2921 * may have freed the last object and be 2922 * waiting to release the slab. 2923 */ 2924 list_del(&page->lru); 2925 n->nr_partial--; 2926 slab_unlock(page); 2927 discard_slab(s, page); 2928 } else { 2929 list_move(&page->lru, 2930 slabs_by_inuse + page->inuse); 2931 } 2932 } 2933 2934 /* 2935 * Rebuild the partial list with the slabs filled up most 2936 * first and the least used slabs at the end. 2937 */ 2938 for (i = objects - 1; i >= 0; i--) 2939 list_splice(slabs_by_inuse + i, n->partial.prev); 2940 2941 spin_unlock_irqrestore(&n->list_lock, flags); 2942 } 2943 2944 kfree(slabs_by_inuse); 2945 return 0; 2946} 2947EXPORT_SYMBOL(kmem_cache_shrink); 2948 2949#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2950static int slab_mem_going_offline_callback(void *arg) 2951{ 2952 struct kmem_cache *s; 2953 2954 down_read(&slub_lock); 2955 list_for_each_entry(s, &slab_caches, list) 2956 kmem_cache_shrink(s); 2957 up_read(&slub_lock); 2958 2959 return 0; 2960} 2961 2962static void slab_mem_offline_callback(void *arg) 2963{ 2964 struct kmem_cache_node *n; 2965 struct kmem_cache *s; 2966 struct memory_notify *marg = arg; 2967 int offline_node; 2968 2969 offline_node = marg->status_change_nid; 2970 2971 /* 2972 * If the node still has available memory. we need kmem_cache_node 2973 * for it yet. 2974 */ 2975 if (offline_node < 0) 2976 return; 2977 2978 down_read(&slub_lock); 2979 list_for_each_entry(s, &slab_caches, list) { 2980 n = get_node(s, offline_node); 2981 if (n) { 2982 /* 2983 * if n->nr_slabs > 0, slabs still exist on the node 2984 * that is going down. We were unable to free them, 2985 * and offline_pages() function shoudn't call this 2986 * callback. So, we must fail. 2987 */ 2988 BUG_ON(slabs_node(s, offline_node)); 2989 2990 s->node[offline_node] = NULL; 2991 kmem_cache_free(kmalloc_caches, n); 2992 } 2993 } 2994 up_read(&slub_lock); 2995} 2996 2997static int slab_mem_going_online_callback(void *arg) 2998{ 2999 struct kmem_cache_node *n; 3000 struct kmem_cache *s; 3001 struct memory_notify *marg = arg; 3002 int nid = marg->status_change_nid; 3003 int ret = 0; 3004 3005 /* 3006 * If the node's memory is already available, then kmem_cache_node is 3007 * already created. Nothing to do. 3008 */ 3009 if (nid < 0) 3010 return 0; 3011 3012 /* 3013 * We are bringing a node online. No memory is available yet. We must 3014 * allocate a kmem_cache_node structure in order to bring the node 3015 * online. 3016 */ 3017 down_read(&slub_lock); 3018 list_for_each_entry(s, &slab_caches, list) { 3019 /* 3020 * XXX: kmem_cache_alloc_node will fallback to other nodes 3021 * since memory is not yet available from the node that 3022 * is brought up. 3023 */ 3024 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 3025 if (!n) { 3026 ret = -ENOMEM; 3027 goto out; 3028 } 3029 init_kmem_cache_node(n, s); 3030 s->node[nid] = n; 3031 } 3032out: 3033 up_read(&slub_lock); 3034 return ret; 3035} 3036 3037static int slab_memory_callback(struct notifier_block *self, 3038 unsigned long action, void *arg) 3039{ 3040 int ret = 0; 3041 3042 switch (action) { 3043 case MEM_GOING_ONLINE: 3044 ret = slab_mem_going_online_callback(arg); 3045 break; 3046 case MEM_GOING_OFFLINE: 3047 ret = slab_mem_going_offline_callback(arg); 3048 break; 3049 case MEM_OFFLINE: 3050 case MEM_CANCEL_ONLINE: 3051 slab_mem_offline_callback(arg); 3052 break; 3053 case MEM_ONLINE: 3054 case MEM_CANCEL_OFFLINE: 3055 break; 3056 } 3057 if (ret) 3058 ret = notifier_from_errno(ret); 3059 else 3060 ret = NOTIFY_OK; 3061 return ret; 3062} 3063 3064#endif /* CONFIG_MEMORY_HOTPLUG */ 3065 3066/******************************************************************** 3067 * Basic setup of slabs 3068 *******************************************************************/ 3069 3070void __init kmem_cache_init(void) 3071{ 3072 int i; 3073 int caches = 0; 3074 3075#ifdef CONFIG_NUMA 3076 /* 3077 * Must first have the slab cache available for the allocations of the 3078 * struct kmem_cache_node's. There is special bootstrap code in 3079 * kmem_cache_open for slab_state == DOWN. 3080 */ 3081 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 3082 sizeof(struct kmem_cache_node), GFP_NOWAIT); 3083 kmalloc_caches[0].refcount = -1; 3084 caches++; 3085 3086 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 3087#endif 3088 3089 /* Able to allocate the per node structures */ 3090 slab_state = PARTIAL; 3091 3092 /* Caches that are not of the two-to-the-power-of size */ 3093 if (KMALLOC_MIN_SIZE <= 32) { 3094 create_kmalloc_cache(&kmalloc_caches[1], 3095 "kmalloc-96", 96, GFP_NOWAIT); 3096 caches++; 3097 } 3098 if (KMALLOC_MIN_SIZE <= 64) { 3099 create_kmalloc_cache(&kmalloc_caches[2], 3100 "kmalloc-192", 192, GFP_NOWAIT); 3101 caches++; 3102 } 3103 3104 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3105 create_kmalloc_cache(&kmalloc_caches[i], 3106 "kmalloc", 1 << i, GFP_NOWAIT); 3107 caches++; 3108 } 3109 3110 3111 /* 3112 * Patch up the size_index table if we have strange large alignment 3113 * requirements for the kmalloc array. This is only the case for 3114 * MIPS it seems. The standard arches will not generate any code here. 3115 * 3116 * Largest permitted alignment is 256 bytes due to the way we 3117 * handle the index determination for the smaller caches. 3118 * 3119 * Make sure that nothing crazy happens if someone starts tinkering 3120 * around with ARCH_KMALLOC_MINALIGN 3121 */ 3122 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3123 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3124 3125 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { 3126 int elem = size_index_elem(i); 3127 if (elem >= ARRAY_SIZE(size_index)) 3128 break; 3129 size_index[elem] = KMALLOC_SHIFT_LOW; 3130 } 3131 3132 if (KMALLOC_MIN_SIZE == 64) { 3133 /* 3134 * The 96 byte size cache is not used if the alignment 3135 * is 64 byte. 3136 */ 3137 for (i = 64 + 8; i <= 96; i += 8) 3138 size_index[size_index_elem(i)] = 7; 3139 } else if (KMALLOC_MIN_SIZE == 128) { 3140 /* 3141 * The 192 byte sized cache is not used if the alignment 3142 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3143 * instead. 3144 */ 3145 for (i = 128 + 8; i <= 192; i += 8) 3146 size_index[size_index_elem(i)] = 8; 3147 } 3148 3149 slab_state = UP; 3150 3151 /* Provide the correct kmalloc names now that the caches are up */ 3152 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) 3153 kmalloc_caches[i]. name = 3154 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); 3155 3156#ifdef CONFIG_SMP 3157 register_cpu_notifier(&slab_notifier); 3158#endif 3159#ifdef CONFIG_NUMA 3160 kmem_size = offsetof(struct kmem_cache, node) + 3161 nr_node_ids * sizeof(struct kmem_cache_node *); 3162#else 3163 kmem_size = sizeof(struct kmem_cache); 3164#endif 3165 3166 printk(KERN_INFO 3167 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3168 " CPUs=%d, Nodes=%d\n", 3169 caches, cache_line_size(), 3170 slub_min_order, slub_max_order, slub_min_objects, 3171 nr_cpu_ids, nr_node_ids); 3172} 3173 3174void __init kmem_cache_init_late(void) 3175{ 3176} 3177 3178/* 3179 * Find a mergeable slab cache 3180 */ 3181static int slab_unmergeable(struct kmem_cache *s) 3182{ 3183 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3184 return 1; 3185 3186 if (s->ctor) 3187 return 1; 3188 3189 /* 3190 * We may have set a slab to be unmergeable during bootstrap. 3191 */ 3192 if (s->refcount < 0) 3193 return 1; 3194 3195 return 0; 3196} 3197 3198static struct kmem_cache *find_mergeable(size_t size, 3199 size_t align, unsigned long flags, const char *name, 3200 void (*ctor)(void *)) 3201{ 3202 struct kmem_cache *s; 3203 3204 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3205 return NULL; 3206 3207 if (ctor) 3208 return NULL; 3209 3210 size = ALIGN(size, sizeof(void *)); 3211 align = calculate_alignment(flags, align, size); 3212 size = ALIGN(size, align); 3213 flags = kmem_cache_flags(size, flags, name, NULL); 3214 3215 list_for_each_entry(s, &slab_caches, list) { 3216 if (slab_unmergeable(s)) 3217 continue; 3218 3219 if (size > s->size) 3220 continue; 3221 3222 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3223 continue; 3224 /* 3225 * Check if alignment is compatible. 3226 * Courtesy of Adrian Drzewiecki 3227 */ 3228 if ((s->size & ~(align - 1)) != s->size) 3229 continue; 3230 3231 if (s->size - size >= sizeof(void *)) 3232 continue; 3233 3234 return s; 3235 } 3236 return NULL; 3237} 3238 3239struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3240 size_t align, unsigned long flags, void (*ctor)(void *)) 3241{ 3242 struct kmem_cache *s; 3243 3244 if (WARN_ON(!name)) 3245 return NULL; 3246 3247 down_write(&slub_lock); 3248 s = find_mergeable(size, align, flags, name, ctor); 3249 if (s) { 3250 int cpu; 3251 3252 s->refcount++; 3253 /* 3254 * Adjust the object sizes so that we clear 3255 * the complete object on kzalloc. 3256 */ 3257 s->objsize = max(s->objsize, (int)size); 3258 3259 /* 3260 * And then we need to update the object size in the 3261 * per cpu structures 3262 */ 3263 for_each_online_cpu(cpu) 3264 per_cpu_ptr(s->cpu_slab, cpu)->objsize = s->objsize; 3265 3266 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3267 up_write(&slub_lock); 3268 3269 if (sysfs_slab_alias(s, name)) { 3270 down_write(&slub_lock); 3271 s->refcount--; 3272 up_write(&slub_lock); 3273 goto err; 3274 } 3275 return s; 3276 } 3277 3278 s = kmalloc(kmem_size, GFP_KERNEL); 3279 if (s) { 3280 if (kmem_cache_open(s, GFP_KERNEL, name, 3281 size, align, flags, ctor)) { 3282 list_add(&s->list, &slab_caches); 3283 up_write(&slub_lock); 3284 if (sysfs_slab_add(s)) { 3285 down_write(&slub_lock); 3286 list_del(&s->list); 3287 up_write(&slub_lock); 3288 kfree(s); 3289 goto err; 3290 } 3291 return s; 3292 } 3293 kfree(s); 3294 } 3295 up_write(&slub_lock); 3296 3297err: 3298 if (flags & SLAB_PANIC) 3299 panic("Cannot create slabcache %s\n", name); 3300 else 3301 s = NULL; 3302 return s; 3303} 3304EXPORT_SYMBOL(kmem_cache_create); 3305 3306#ifdef CONFIG_SMP 3307/* 3308 * Use the cpu notifier to insure that the cpu slabs are flushed when 3309 * necessary. 3310 */ 3311static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3312 unsigned long action, void *hcpu) 3313{ 3314 long cpu = (long)hcpu; 3315 struct kmem_cache *s; 3316 unsigned long flags; 3317 3318 switch (action) { 3319 case CPU_UP_PREPARE: 3320 case CPU_UP_PREPARE_FROZEN: 3321 down_read(&slub_lock); 3322 list_for_each_entry(s, &slab_caches, list) 3323 init_kmem_cache_cpu(s, per_cpu_ptr(s->cpu_slab, cpu)); 3324 up_read(&slub_lock); 3325 break; 3326 3327 case CPU_UP_CANCELED: 3328 case CPU_UP_CANCELED_FROZEN: 3329 case CPU_DEAD: 3330 case CPU_DEAD_FROZEN: 3331 down_read(&slub_lock); 3332 list_for_each_entry(s, &slab_caches, list) { 3333 local_irq_save(flags); 3334 __flush_cpu_slab(s, cpu); 3335 local_irq_restore(flags); 3336 } 3337 up_read(&slub_lock); 3338 break; 3339 default: 3340 break; 3341 } 3342 return NOTIFY_OK; 3343} 3344 3345static struct notifier_block __cpuinitdata slab_notifier = { 3346 .notifier_call = slab_cpuup_callback 3347}; 3348 3349#endif 3350 3351void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3352{ 3353 struct kmem_cache *s; 3354 void *ret; 3355 3356 if (unlikely(size > SLUB_MAX_SIZE)) 3357 return kmalloc_large(size, gfpflags); 3358 3359 s = get_slab(size, gfpflags); 3360 3361 if (unlikely(ZERO_OR_NULL_PTR(s))) 3362 return s; 3363 3364 ret = slab_alloc(s, gfpflags, -1, caller); 3365 3366 /* Honor the call site pointer we recieved. */ 3367 trace_kmalloc(caller, ret, size, s->size, gfpflags); 3368 3369 return ret; 3370} 3371 3372void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3373 int node, unsigned long caller) 3374{ 3375 struct kmem_cache *s; 3376 void *ret; 3377 3378 if (unlikely(size > SLUB_MAX_SIZE)) 3379 return kmalloc_large_node(size, gfpflags, node); 3380 3381 s = get_slab(size, gfpflags); 3382 3383 if (unlikely(ZERO_OR_NULL_PTR(s))) 3384 return s; 3385 3386 ret = slab_alloc(s, gfpflags, node, caller); 3387 3388 /* Honor the call site pointer we recieved. */ 3389 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 3390 3391 return ret; 3392} 3393 3394#ifdef CONFIG_SLUB_DEBUG 3395static int count_inuse(struct page *page) 3396{ 3397 return page->inuse; 3398} 3399 3400static int count_total(struct page *page) 3401{ 3402 return page->objects; 3403} 3404 3405static int validate_slab(struct kmem_cache *s, struct page *page, 3406 unsigned long *map) 3407{ 3408 void *p; 3409 void *addr = page_address(page); 3410 3411 if (!check_slab(s, page) || 3412 !on_freelist(s, page, NULL)) 3413 return 0; 3414 3415 /* Now we know that a valid freelist exists */ 3416 bitmap_zero(map, page->objects); 3417 3418 for_each_free_object(p, s, page->freelist) { 3419 set_bit(slab_index(p, s, addr), map); 3420 if (!check_object(s, page, p, 0)) 3421 return 0; 3422 } 3423 3424 for_each_object(p, s, addr, page->objects) 3425 if (!test_bit(slab_index(p, s, addr), map)) 3426 if (!check_object(s, page, p, 1)) 3427 return 0; 3428 return 1; 3429} 3430 3431static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3432 unsigned long *map) 3433{ 3434 if (slab_trylock(page)) { 3435 validate_slab(s, page, map); 3436 slab_unlock(page); 3437 } else 3438 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3439 s->name, page); 3440 3441 if (s->flags & DEBUG_DEFAULT_FLAGS) { 3442 if (!PageSlubDebug(page)) 3443 printk(KERN_ERR "SLUB %s: SlubDebug not set " 3444 "on slab 0x%p\n", s->name, page); 3445 } else { 3446 if (PageSlubDebug(page)) 3447 printk(KERN_ERR "SLUB %s: SlubDebug set on " 3448 "slab 0x%p\n", s->name, page); 3449 } 3450} 3451 3452static int validate_slab_node(struct kmem_cache *s, 3453 struct kmem_cache_node *n, unsigned long *map) 3454{ 3455 unsigned long count = 0; 3456 struct page *page; 3457 unsigned long flags; 3458 3459 spin_lock_irqsave(&n->list_lock, flags); 3460 3461 list_for_each_entry(page, &n->partial, lru) { 3462 validate_slab_slab(s, page, map); 3463 count++; 3464 } 3465 if (count != n->nr_partial) 3466 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3467 "counter=%ld\n", s->name, count, n->nr_partial); 3468 3469 if (!(s->flags & SLAB_STORE_USER)) 3470 goto out; 3471 3472 list_for_each_entry(page, &n->full, lru) { 3473 validate_slab_slab(s, page, map); 3474 count++; 3475 } 3476 if (count != atomic_long_read(&n->nr_slabs)) 3477 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3478 "counter=%ld\n", s->name, count, 3479 atomic_long_read(&n->nr_slabs)); 3480 3481out: 3482 spin_unlock_irqrestore(&n->list_lock, flags); 3483 return count; 3484} 3485 3486static long validate_slab_cache(struct kmem_cache *s) 3487{ 3488 int node; 3489 unsigned long count = 0; 3490 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3491 sizeof(unsigned long), GFP_KERNEL); 3492 3493 if (!map) 3494 return -ENOMEM; 3495 3496 flush_all(s); 3497 for_each_node_state(node, N_NORMAL_MEMORY) { 3498 struct kmem_cache_node *n = get_node(s, node); 3499 3500 count += validate_slab_node(s, n, map); 3501 } 3502 kfree(map); 3503 return count; 3504} 3505 3506#ifdef SLUB_RESILIENCY_TEST 3507static void resiliency_test(void) 3508{ 3509 u8 *p; 3510 3511 printk(KERN_ERR "SLUB resiliency testing\n"); 3512 printk(KERN_ERR "-----------------------\n"); 3513 printk(KERN_ERR "A. Corruption after allocation\n"); 3514 3515 p = kzalloc(16, GFP_KERNEL); 3516 p[16] = 0x12; 3517 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3518 " 0x12->0x%p\n\n", p + 16); 3519 3520 validate_slab_cache(kmalloc_caches + 4); 3521 3522 /* Hmmm... The next two are dangerous */ 3523 p = kzalloc(32, GFP_KERNEL); 3524 p[32 + sizeof(void *)] = 0x34; 3525 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3526 " 0x34 -> -0x%p\n", p); 3527 printk(KERN_ERR 3528 "If allocated object is overwritten then not detectable\n\n"); 3529 3530 validate_slab_cache(kmalloc_caches + 5); 3531 p = kzalloc(64, GFP_KERNEL); 3532 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3533 *p = 0x56; 3534 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3535 p); 3536 printk(KERN_ERR 3537 "If allocated object is overwritten then not detectable\n\n"); 3538 validate_slab_cache(kmalloc_caches + 6); 3539 3540 printk(KERN_ERR "\nB. Corruption after free\n"); 3541 p = kzalloc(128, GFP_KERNEL); 3542 kfree(p); 3543 *p = 0x78; 3544 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3545 validate_slab_cache(kmalloc_caches + 7); 3546 3547 p = kzalloc(256, GFP_KERNEL); 3548 kfree(p); 3549 p[50] = 0x9a; 3550 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3551 p); 3552 validate_slab_cache(kmalloc_caches + 8); 3553 3554 p = kzalloc(512, GFP_KERNEL); 3555 kfree(p); 3556 p[512] = 0xab; 3557 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3558 validate_slab_cache(kmalloc_caches + 9); 3559} 3560#else 3561static void resiliency_test(void) {}; 3562#endif 3563 3564/* 3565 * Generate lists of code addresses where slabcache objects are allocated 3566 * and freed. 3567 */ 3568 3569struct location { 3570 unsigned long count; 3571 unsigned long addr; 3572 long long sum_time; 3573 long min_time; 3574 long max_time; 3575 long min_pid; 3576 long max_pid; 3577 DECLARE_BITMAP(cpus, NR_CPUS); 3578 nodemask_t nodes; 3579}; 3580 3581struct loc_track { 3582 unsigned long max; 3583 unsigned long count; 3584 struct location *loc; 3585}; 3586 3587static void free_loc_track(struct loc_track *t) 3588{ 3589 if (t->max) 3590 free_pages((unsigned long)t->loc, 3591 get_order(sizeof(struct location) * t->max)); 3592} 3593 3594static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3595{ 3596 struct location *l; 3597 int order; 3598 3599 order = get_order(sizeof(struct location) * max); 3600 3601 l = (void *)__get_free_pages(flags, order); 3602 if (!l) 3603 return 0; 3604 3605 if (t->count) { 3606 memcpy(l, t->loc, sizeof(struct location) * t->count); 3607 free_loc_track(t); 3608 } 3609 t->max = max; 3610 t->loc = l; 3611 return 1; 3612} 3613 3614static int add_location(struct loc_track *t, struct kmem_cache *s, 3615 const struct track *track) 3616{ 3617 long start, end, pos; 3618 struct location *l; 3619 unsigned long caddr; 3620 unsigned long age = jiffies - track->when; 3621 3622 start = -1; 3623 end = t->count; 3624 3625 for ( ; ; ) { 3626 pos = start + (end - start + 1) / 2; 3627 3628 /* 3629 * There is nothing at "end". If we end up there 3630 * we need to add something to before end. 3631 */ 3632 if (pos == end) 3633 break; 3634 3635 caddr = t->loc[pos].addr; 3636 if (track->addr == caddr) { 3637 3638 l = &t->loc[pos]; 3639 l->count++; 3640 if (track->when) { 3641 l->sum_time += age; 3642 if (age < l->min_time) 3643 l->min_time = age; 3644 if (age > l->max_time) 3645 l->max_time = age; 3646 3647 if (track->pid < l->min_pid) 3648 l->min_pid = track->pid; 3649 if (track->pid > l->max_pid) 3650 l->max_pid = track->pid; 3651 3652 cpumask_set_cpu(track->cpu, 3653 to_cpumask(l->cpus)); 3654 } 3655 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3656 return 1; 3657 } 3658 3659 if (track->addr < caddr) 3660 end = pos; 3661 else 3662 start = pos; 3663 } 3664 3665 /* 3666 * Not found. Insert new tracking element. 3667 */ 3668 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3669 return 0; 3670 3671 l = t->loc + pos; 3672 if (pos < t->count) 3673 memmove(l + 1, l, 3674 (t->count - pos) * sizeof(struct location)); 3675 t->count++; 3676 l->count = 1; 3677 l->addr = track->addr; 3678 l->sum_time = age; 3679 l->min_time = age; 3680 l->max_time = age; 3681 l->min_pid = track->pid; 3682 l->max_pid = track->pid; 3683 cpumask_clear(to_cpumask(l->cpus)); 3684 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 3685 nodes_clear(l->nodes); 3686 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3687 return 1; 3688} 3689 3690static void process_slab(struct loc_track *t, struct kmem_cache *s, 3691 struct page *page, enum track_item alloc) 3692{ 3693 void *addr = page_address(page); 3694 DECLARE_BITMAP(map, page->objects); 3695 void *p; 3696 3697 bitmap_zero(map, page->objects); 3698 for_each_free_object(p, s, page->freelist) 3699 set_bit(slab_index(p, s, addr), map); 3700 3701 for_each_object(p, s, addr, page->objects) 3702 if (!test_bit(slab_index(p, s, addr), map)) 3703 add_location(t, s, get_track(s, p, alloc)); 3704} 3705 3706static int list_locations(struct kmem_cache *s, char *buf, 3707 enum track_item alloc) 3708{ 3709 int len = 0; 3710 unsigned long i; 3711 struct loc_track t = { 0, 0, NULL }; 3712 int node; 3713 3714 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3715 GFP_TEMPORARY)) 3716 return sprintf(buf, "Out of memory\n"); 3717 3718 /* Push back cpu slabs */ 3719 flush_all(s); 3720 3721 for_each_node_state(node, N_NORMAL_MEMORY) { 3722 struct kmem_cache_node *n = get_node(s, node); 3723 unsigned long flags; 3724 struct page *page; 3725 3726 if (!atomic_long_read(&n->nr_slabs)) 3727 continue; 3728 3729 spin_lock_irqsave(&n->list_lock, flags); 3730 list_for_each_entry(page, &n->partial, lru) 3731 process_slab(&t, s, page, alloc); 3732 list_for_each_entry(page, &n->full, lru) 3733 process_slab(&t, s, page, alloc); 3734 spin_unlock_irqrestore(&n->list_lock, flags); 3735 } 3736 3737 for (i = 0; i < t.count; i++) { 3738 struct location *l = &t.loc[i]; 3739 3740 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 3741 break; 3742 len += sprintf(buf + len, "%7ld ", l->count); 3743 3744 if (l->addr) 3745 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3746 else 3747 len += sprintf(buf + len, "<not-available>"); 3748 3749 if (l->sum_time != l->min_time) { 3750 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3751 l->min_time, 3752 (long)div_u64(l->sum_time, l->count), 3753 l->max_time); 3754 } else 3755 len += sprintf(buf + len, " age=%ld", 3756 l->min_time); 3757 3758 if (l->min_pid != l->max_pid) 3759 len += sprintf(buf + len, " pid=%ld-%ld", 3760 l->min_pid, l->max_pid); 3761 else 3762 len += sprintf(buf + len, " pid=%ld", 3763 l->min_pid); 3764 3765 if (num_online_cpus() > 1 && 3766 !cpumask_empty(to_cpumask(l->cpus)) && 3767 len < PAGE_SIZE - 60) { 3768 len += sprintf(buf + len, " cpus="); 3769 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3770 to_cpumask(l->cpus)); 3771 } 3772 3773 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 3774 len < PAGE_SIZE - 60) { 3775 len += sprintf(buf + len, " nodes="); 3776 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3777 l->nodes); 3778 } 3779 3780 len += sprintf(buf + len, "\n"); 3781 } 3782 3783 free_loc_track(&t); 3784 if (!t.count) 3785 len += sprintf(buf, "No data\n"); 3786 return len; 3787} 3788 3789enum slab_stat_type { 3790 SL_ALL, /* All slabs */ 3791 SL_PARTIAL, /* Only partially allocated slabs */ 3792 SL_CPU, /* Only slabs used for cpu caches */ 3793 SL_OBJECTS, /* Determine allocated objects not slabs */ 3794 SL_TOTAL /* Determine object capacity not slabs */ 3795}; 3796 3797#define SO_ALL (1 << SL_ALL) 3798#define SO_PARTIAL (1 << SL_PARTIAL) 3799#define SO_CPU (1 << SL_CPU) 3800#define SO_OBJECTS (1 << SL_OBJECTS) 3801#define SO_TOTAL (1 << SL_TOTAL) 3802 3803static ssize_t show_slab_objects(struct kmem_cache *s, 3804 char *buf, unsigned long flags) 3805{ 3806 unsigned long total = 0; 3807 int node; 3808 int x; 3809 unsigned long *nodes; 3810 unsigned long *per_cpu; 3811 3812 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3813 if (!nodes) 3814 return -ENOMEM; 3815 per_cpu = nodes + nr_node_ids; 3816 3817 if (flags & SO_CPU) { 3818 int cpu; 3819 3820 for_each_possible_cpu(cpu) { 3821 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 3822 3823 if (!c || c->node < 0) 3824 continue; 3825 3826 if (c->page) { 3827 if (flags & SO_TOTAL) 3828 x = c->page->objects; 3829 else if (flags & SO_OBJECTS) 3830 x = c->page->inuse; 3831 else 3832 x = 1; 3833 3834 total += x; 3835 nodes[c->node] += x; 3836 } 3837 per_cpu[c->node]++; 3838 } 3839 } 3840 3841 if (flags & SO_ALL) { 3842 for_each_node_state(node, N_NORMAL_MEMORY) { 3843 struct kmem_cache_node *n = get_node(s, node); 3844 3845 if (flags & SO_TOTAL) 3846 x = atomic_long_read(&n->total_objects); 3847 else if (flags & SO_OBJECTS) 3848 x = atomic_long_read(&n->total_objects) - 3849 count_partial(n, count_free); 3850 3851 else 3852 x = atomic_long_read(&n->nr_slabs); 3853 total += x; 3854 nodes[node] += x; 3855 } 3856 3857 } else if (flags & SO_PARTIAL) { 3858 for_each_node_state(node, N_NORMAL_MEMORY) { 3859 struct kmem_cache_node *n = get_node(s, node); 3860 3861 if (flags & SO_TOTAL) 3862 x = count_partial(n, count_total); 3863 else if (flags & SO_OBJECTS) 3864 x = count_partial(n, count_inuse); 3865 else 3866 x = n->nr_partial; 3867 total += x; 3868 nodes[node] += x; 3869 } 3870 } 3871 x = sprintf(buf, "%lu", total); 3872#ifdef CONFIG_NUMA 3873 for_each_node_state(node, N_NORMAL_MEMORY) 3874 if (nodes[node]) 3875 x += sprintf(buf + x, " N%d=%lu", 3876 node, nodes[node]); 3877#endif 3878 kfree(nodes); 3879 return x + sprintf(buf + x, "\n"); 3880} 3881 3882static int any_slab_objects(struct kmem_cache *s) 3883{ 3884 int node; 3885 3886 for_each_online_node(node) { 3887 struct kmem_cache_node *n = get_node(s, node); 3888 3889 if (!n) 3890 continue; 3891 3892 if (atomic_long_read(&n->total_objects)) 3893 return 1; 3894 } 3895 return 0; 3896} 3897 3898#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3899#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3900 3901struct slab_attribute { 3902 struct attribute attr; 3903 ssize_t (*show)(struct kmem_cache *s, char *buf); 3904 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3905}; 3906 3907#define SLAB_ATTR_RO(_name) \ 3908 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3909 3910#define SLAB_ATTR(_name) \ 3911 static struct slab_attribute _name##_attr = \ 3912 __ATTR(_name, 0644, _name##_show, _name##_store) 3913 3914static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3915{ 3916 return sprintf(buf, "%d\n", s->size); 3917} 3918SLAB_ATTR_RO(slab_size); 3919 3920static ssize_t align_show(struct kmem_cache *s, char *buf) 3921{ 3922 return sprintf(buf, "%d\n", s->align); 3923} 3924SLAB_ATTR_RO(align); 3925 3926static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3927{ 3928 return sprintf(buf, "%d\n", s->objsize); 3929} 3930SLAB_ATTR_RO(object_size); 3931 3932static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3933{ 3934 return sprintf(buf, "%d\n", oo_objects(s->oo)); 3935} 3936SLAB_ATTR_RO(objs_per_slab); 3937 3938static ssize_t order_store(struct kmem_cache *s, 3939 const char *buf, size_t length) 3940{ 3941 unsigned long order; 3942 int err; 3943 3944 err = strict_strtoul(buf, 10, &order); 3945 if (err) 3946 return err; 3947 3948 if (order > slub_max_order || order < slub_min_order) 3949 return -EINVAL; 3950 3951 calculate_sizes(s, order); 3952 return length; 3953} 3954 3955static ssize_t order_show(struct kmem_cache *s, char *buf) 3956{ 3957 return sprintf(buf, "%d\n", oo_order(s->oo)); 3958} 3959SLAB_ATTR(order); 3960 3961static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 3962{ 3963 return sprintf(buf, "%lu\n", s->min_partial); 3964} 3965 3966static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 3967 size_t length) 3968{ 3969 unsigned long min; 3970 int err; 3971 3972 err = strict_strtoul(buf, 10, &min); 3973 if (err) 3974 return err; 3975 3976 set_min_partial(s, min); 3977 return length; 3978} 3979SLAB_ATTR(min_partial); 3980 3981static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3982{ 3983 if (s->ctor) { 3984 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3985 3986 return n + sprintf(buf + n, "\n"); 3987 } 3988 return 0; 3989} 3990SLAB_ATTR_RO(ctor); 3991 3992static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3993{ 3994 return sprintf(buf, "%d\n", s->refcount - 1); 3995} 3996SLAB_ATTR_RO(aliases); 3997 3998static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3999{ 4000 return show_slab_objects(s, buf, SO_ALL); 4001} 4002SLAB_ATTR_RO(slabs); 4003 4004static ssize_t partial_show(struct kmem_cache *s, char *buf) 4005{ 4006 return show_slab_objects(s, buf, SO_PARTIAL); 4007} 4008SLAB_ATTR_RO(partial); 4009 4010static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4011{ 4012 return show_slab_objects(s, buf, SO_CPU); 4013} 4014SLAB_ATTR_RO(cpu_slabs); 4015 4016static ssize_t objects_show(struct kmem_cache *s, char *buf) 4017{ 4018 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4019} 4020SLAB_ATTR_RO(objects); 4021 4022static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4023{ 4024 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4025} 4026SLAB_ATTR_RO(objects_partial); 4027 4028static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 4029{ 4030 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 4031} 4032SLAB_ATTR_RO(total_objects); 4033 4034static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 4035{ 4036 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 4037} 4038 4039static ssize_t sanity_checks_store(struct kmem_cache *s, 4040 const char *buf, size_t length) 4041{ 4042 s->flags &= ~SLAB_DEBUG_FREE; 4043 if (buf[0] == '1') 4044 s->flags |= SLAB_DEBUG_FREE; 4045 return length; 4046} 4047SLAB_ATTR(sanity_checks); 4048 4049static ssize_t trace_show(struct kmem_cache *s, char *buf) 4050{ 4051 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4052} 4053 4054static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4055 size_t length) 4056{ 4057 s->flags &= ~SLAB_TRACE; 4058 if (buf[0] == '1') 4059 s->flags |= SLAB_TRACE; 4060 return length; 4061} 4062SLAB_ATTR(trace); 4063 4064static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4065{ 4066 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4067} 4068 4069static ssize_t reclaim_account_store(struct kmem_cache *s, 4070 const char *buf, size_t length) 4071{ 4072 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4073 if (buf[0] == '1') 4074 s->flags |= SLAB_RECLAIM_ACCOUNT; 4075 return length; 4076} 4077SLAB_ATTR(reclaim_account); 4078 4079static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4080{ 4081 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4082} 4083SLAB_ATTR_RO(hwcache_align); 4084 4085#ifdef CONFIG_ZONE_DMA 4086static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4087{ 4088 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4089} 4090SLAB_ATTR_RO(cache_dma); 4091#endif 4092 4093static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4094{ 4095 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4096} 4097SLAB_ATTR_RO(destroy_by_rcu); 4098 4099static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4100{ 4101 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4102} 4103 4104static ssize_t red_zone_store(struct kmem_cache *s, 4105 const char *buf, size_t length) 4106{ 4107 if (any_slab_objects(s)) 4108 return -EBUSY; 4109 4110 s->flags &= ~SLAB_RED_ZONE; 4111 if (buf[0] == '1') 4112 s->flags |= SLAB_RED_ZONE; 4113 calculate_sizes(s, -1); 4114 return length; 4115} 4116SLAB_ATTR(red_zone); 4117 4118static ssize_t poison_show(struct kmem_cache *s, char *buf) 4119{ 4120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4121} 4122 4123static ssize_t poison_store(struct kmem_cache *s, 4124 const char *buf, size_t length) 4125{ 4126 if (any_slab_objects(s)) 4127 return -EBUSY; 4128 4129 s->flags &= ~SLAB_POISON; 4130 if (buf[0] == '1') 4131 s->flags |= SLAB_POISON; 4132 calculate_sizes(s, -1); 4133 return length; 4134} 4135SLAB_ATTR(poison); 4136 4137static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4138{ 4139 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4140} 4141 4142static ssize_t store_user_store(struct kmem_cache *s, 4143 const char *buf, size_t length) 4144{ 4145 if (any_slab_objects(s)) 4146 return -EBUSY; 4147 4148 s->flags &= ~SLAB_STORE_USER; 4149 if (buf[0] == '1') 4150 s->flags |= SLAB_STORE_USER; 4151 calculate_sizes(s, -1); 4152 return length; 4153} 4154SLAB_ATTR(store_user); 4155 4156static ssize_t validate_show(struct kmem_cache *s, char *buf) 4157{ 4158 return 0; 4159} 4160 4161static ssize_t validate_store(struct kmem_cache *s, 4162 const char *buf, size_t length) 4163{ 4164 int ret = -EINVAL; 4165 4166 if (buf[0] == '1') { 4167 ret = validate_slab_cache(s); 4168 if (ret >= 0) 4169 ret = length; 4170 } 4171 return ret; 4172} 4173SLAB_ATTR(validate); 4174 4175static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4176{ 4177 return 0; 4178} 4179 4180static ssize_t shrink_store(struct kmem_cache *s, 4181 const char *buf, size_t length) 4182{ 4183 if (buf[0] == '1') { 4184 int rc = kmem_cache_shrink(s); 4185 4186 if (rc) 4187 return rc; 4188 } else 4189 return -EINVAL; 4190 return length; 4191} 4192SLAB_ATTR(shrink); 4193 4194static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4195{ 4196 if (!(s->flags & SLAB_STORE_USER)) 4197 return -ENOSYS; 4198 return list_locations(s, buf, TRACK_ALLOC); 4199} 4200SLAB_ATTR_RO(alloc_calls); 4201 4202static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4203{ 4204 if (!(s->flags & SLAB_STORE_USER)) 4205 return -ENOSYS; 4206 return list_locations(s, buf, TRACK_FREE); 4207} 4208SLAB_ATTR_RO(free_calls); 4209 4210#ifdef CONFIG_NUMA 4211static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4212{ 4213 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4214} 4215 4216static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4217 const char *buf, size_t length) 4218{ 4219 unsigned long ratio; 4220 int err; 4221 4222 err = strict_strtoul(buf, 10, &ratio); 4223 if (err) 4224 return err; 4225 4226 if (ratio <= 100) 4227 s->remote_node_defrag_ratio = ratio * 10; 4228 4229 return length; 4230} 4231SLAB_ATTR(remote_node_defrag_ratio); 4232#endif 4233 4234#ifdef CONFIG_SLUB_STATS 4235static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4236{ 4237 unsigned long sum = 0; 4238 int cpu; 4239 int len; 4240 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4241 4242 if (!data) 4243 return -ENOMEM; 4244 4245 for_each_online_cpu(cpu) { 4246 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 4247 4248 data[cpu] = x; 4249 sum += x; 4250 } 4251 4252 len = sprintf(buf, "%lu", sum); 4253 4254#ifdef CONFIG_SMP 4255 for_each_online_cpu(cpu) { 4256 if (data[cpu] && len < PAGE_SIZE - 20) 4257 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4258 } 4259#endif 4260 kfree(data); 4261 return len + sprintf(buf + len, "\n"); 4262} 4263 4264static void clear_stat(struct kmem_cache *s, enum stat_item si) 4265{ 4266 int cpu; 4267 4268 for_each_online_cpu(cpu) 4269 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 4270} 4271 4272#define STAT_ATTR(si, text) \ 4273static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4274{ \ 4275 return show_stat(s, buf, si); \ 4276} \ 4277static ssize_t text##_store(struct kmem_cache *s, \ 4278 const char *buf, size_t length) \ 4279{ \ 4280 if (buf[0] != '0') \ 4281 return -EINVAL; \ 4282 clear_stat(s, si); \ 4283 return length; \ 4284} \ 4285SLAB_ATTR(text); \ 4286 4287STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4288STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4289STAT_ATTR(FREE_FASTPATH, free_fastpath); 4290STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4291STAT_ATTR(FREE_FROZEN, free_frozen); 4292STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4293STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4294STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4295STAT_ATTR(ALLOC_SLAB, alloc_slab); 4296STAT_ATTR(ALLOC_REFILL, alloc_refill); 4297STAT_ATTR(FREE_SLAB, free_slab); 4298STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4299STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4300STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4301STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4302STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4303STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4304STAT_ATTR(ORDER_FALLBACK, order_fallback); 4305#endif 4306 4307static struct attribute *slab_attrs[] = { 4308 &slab_size_attr.attr, 4309 &object_size_attr.attr, 4310 &objs_per_slab_attr.attr, 4311 &order_attr.attr, 4312 &min_partial_attr.attr, 4313 &objects_attr.attr, 4314 &objects_partial_attr.attr, 4315 &total_objects_attr.attr, 4316 &slabs_attr.attr, 4317 &partial_attr.attr, 4318 &cpu_slabs_attr.attr, 4319 &ctor_attr.attr, 4320 &aliases_attr.attr, 4321 &align_attr.attr, 4322 &sanity_checks_attr.attr, 4323 &trace_attr.attr, 4324 &hwcache_align_attr.attr, 4325 &reclaim_account_attr.attr, 4326 &destroy_by_rcu_attr.attr, 4327 &red_zone_attr.attr, 4328 &poison_attr.attr, 4329 &store_user_attr.attr, 4330 &validate_attr.attr, 4331 &shrink_attr.attr, 4332 &alloc_calls_attr.attr, 4333 &free_calls_attr.attr, 4334#ifdef CONFIG_ZONE_DMA 4335 &cache_dma_attr.attr, 4336#endif 4337#ifdef CONFIG_NUMA 4338 &remote_node_defrag_ratio_attr.attr, 4339#endif 4340#ifdef CONFIG_SLUB_STATS 4341 &alloc_fastpath_attr.attr, 4342 &alloc_slowpath_attr.attr, 4343 &free_fastpath_attr.attr, 4344 &free_slowpath_attr.attr, 4345 &free_frozen_attr.attr, 4346 &free_add_partial_attr.attr, 4347 &free_remove_partial_attr.attr, 4348 &alloc_from_partial_attr.attr, 4349 &alloc_slab_attr.attr, 4350 &alloc_refill_attr.attr, 4351 &free_slab_attr.attr, 4352 &cpuslab_flush_attr.attr, 4353 &deactivate_full_attr.attr, 4354 &deactivate_empty_attr.attr, 4355 &deactivate_to_head_attr.attr, 4356 &deactivate_to_tail_attr.attr, 4357 &deactivate_remote_frees_attr.attr, 4358 &order_fallback_attr.attr, 4359#endif 4360 NULL 4361}; 4362 4363static struct attribute_group slab_attr_group = { 4364 .attrs = slab_attrs, 4365}; 4366 4367static ssize_t slab_attr_show(struct kobject *kobj, 4368 struct attribute *attr, 4369 char *buf) 4370{ 4371 struct slab_attribute *attribute; 4372 struct kmem_cache *s; 4373 int err; 4374 4375 attribute = to_slab_attr(attr); 4376 s = to_slab(kobj); 4377 4378 if (!attribute->show) 4379 return -EIO; 4380 4381 err = attribute->show(s, buf); 4382 4383 return err; 4384} 4385 4386static ssize_t slab_attr_store(struct kobject *kobj, 4387 struct attribute *attr, 4388 const char *buf, size_t len) 4389{ 4390 struct slab_attribute *attribute; 4391 struct kmem_cache *s; 4392 int err; 4393 4394 attribute = to_slab_attr(attr); 4395 s = to_slab(kobj); 4396 4397 if (!attribute->store) 4398 return -EIO; 4399 4400 err = attribute->store(s, buf, len); 4401 4402 return err; 4403} 4404 4405static void kmem_cache_release(struct kobject *kobj) 4406{ 4407 struct kmem_cache *s = to_slab(kobj); 4408 4409 kfree(s); 4410} 4411 4412static struct sysfs_ops slab_sysfs_ops = { 4413 .show = slab_attr_show, 4414 .store = slab_attr_store, 4415}; 4416 4417static struct kobj_type slab_ktype = { 4418 .sysfs_ops = &slab_sysfs_ops, 4419 .release = kmem_cache_release 4420}; 4421 4422static int uevent_filter(struct kset *kset, struct kobject *kobj) 4423{ 4424 struct kobj_type *ktype = get_ktype(kobj); 4425 4426 if (ktype == &slab_ktype) 4427 return 1; 4428 return 0; 4429} 4430 4431static struct kset_uevent_ops slab_uevent_ops = { 4432 .filter = uevent_filter, 4433}; 4434 4435static struct kset *slab_kset; 4436 4437#define ID_STR_LENGTH 64 4438 4439/* Create a unique string id for a slab cache: 4440 * 4441 * Format :[flags-]size 4442 */ 4443static char *create_unique_id(struct kmem_cache *s) 4444{ 4445 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4446 char *p = name; 4447 4448 BUG_ON(!name); 4449 4450 *p++ = ':'; 4451 /* 4452 * First flags affecting slabcache operations. We will only 4453 * get here for aliasable slabs so we do not need to support 4454 * too many flags. The flags here must cover all flags that 4455 * are matched during merging to guarantee that the id is 4456 * unique. 4457 */ 4458 if (s->flags & SLAB_CACHE_DMA) 4459 *p++ = 'd'; 4460 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4461 *p++ = 'a'; 4462 if (s->flags & SLAB_DEBUG_FREE) 4463 *p++ = 'F'; 4464 if (!(s->flags & SLAB_NOTRACK)) 4465 *p++ = 't'; 4466 if (p != name + 1) 4467 *p++ = '-'; 4468 p += sprintf(p, "%07d", s->size); 4469 BUG_ON(p > name + ID_STR_LENGTH - 1); 4470 return name; 4471} 4472 4473static int sysfs_slab_add(struct kmem_cache *s) 4474{ 4475 int err; 4476 const char *name; 4477 int unmergeable; 4478 4479 if (slab_state < SYSFS) 4480 /* Defer until later */ 4481 return 0; 4482 4483 unmergeable = slab_unmergeable(s); 4484 if (unmergeable) { 4485 /* 4486 * Slabcache can never be merged so we can use the name proper. 4487 * This is typically the case for debug situations. In that 4488 * case we can catch duplicate names easily. 4489 */ 4490 sysfs_remove_link(&slab_kset->kobj, s->name); 4491 name = s->name; 4492 } else { 4493 /* 4494 * Create a unique name for the slab as a target 4495 * for the symlinks. 4496 */ 4497 name = create_unique_id(s); 4498 } 4499 4500 s->kobj.kset = slab_kset; 4501 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4502 if (err) { 4503 kobject_put(&s->kobj); 4504 return err; 4505 } 4506 4507 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4508 if (err) { 4509 kobject_del(&s->kobj); 4510 kobject_put(&s->kobj); 4511 return err; 4512 } 4513 kobject_uevent(&s->kobj, KOBJ_ADD); 4514 if (!unmergeable) { 4515 /* Setup first alias */ 4516 sysfs_slab_alias(s, s->name); 4517 kfree(name); 4518 } 4519 return 0; 4520} 4521 4522static void sysfs_slab_remove(struct kmem_cache *s) 4523{ 4524 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4525 kobject_del(&s->kobj); 4526 kobject_put(&s->kobj); 4527} 4528 4529/* 4530 * Need to buffer aliases during bootup until sysfs becomes 4531 * available lest we lose that information. 4532 */ 4533struct saved_alias { 4534 struct kmem_cache *s; 4535 const char *name; 4536 struct saved_alias *next; 4537}; 4538 4539static struct saved_alias *alias_list; 4540 4541static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4542{ 4543 struct saved_alias *al; 4544 4545 if (slab_state == SYSFS) { 4546 /* 4547 * If we have a leftover link then remove it. 4548 */ 4549 sysfs_remove_link(&slab_kset->kobj, name); 4550 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4551 } 4552 4553 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4554 if (!al) 4555 return -ENOMEM; 4556 4557 al->s = s; 4558 al->name = name; 4559 al->next = alias_list; 4560 alias_list = al; 4561 return 0; 4562} 4563 4564static int __init slab_sysfs_init(void) 4565{ 4566 struct kmem_cache *s; 4567 int err; 4568 4569 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4570 if (!slab_kset) { 4571 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4572 return -ENOSYS; 4573 } 4574 4575 slab_state = SYSFS; 4576 4577 list_for_each_entry(s, &slab_caches, list) { 4578 err = sysfs_slab_add(s); 4579 if (err) 4580 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4581 " to sysfs\n", s->name); 4582 } 4583 4584 while (alias_list) { 4585 struct saved_alias *al = alias_list; 4586 4587 alias_list = alias_list->next; 4588 err = sysfs_slab_alias(al->s, al->name); 4589 if (err) 4590 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4591 " %s to sysfs\n", s->name); 4592 kfree(al); 4593 } 4594 4595 resiliency_test(); 4596 return 0; 4597} 4598 4599__initcall(slab_sysfs_init); 4600#endif 4601 4602/* 4603 * The /proc/slabinfo ABI 4604 */ 4605#ifdef CONFIG_SLABINFO 4606static void print_slabinfo_header(struct seq_file *m) 4607{ 4608 seq_puts(m, "slabinfo - version: 2.1\n"); 4609 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4610 "<objperslab> <pagesperslab>"); 4611 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4612 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4613 seq_putc(m, '\n'); 4614} 4615 4616static void *s_start(struct seq_file *m, loff_t *pos) 4617{ 4618 loff_t n = *pos; 4619 4620 down_read(&slub_lock); 4621 if (!n) 4622 print_slabinfo_header(m); 4623 4624 return seq_list_start(&slab_caches, *pos); 4625} 4626 4627static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4628{ 4629 return seq_list_next(p, &slab_caches, pos); 4630} 4631 4632static void s_stop(struct seq_file *m, void *p) 4633{ 4634 up_read(&slub_lock); 4635} 4636 4637static int s_show(struct seq_file *m, void *p) 4638{ 4639 unsigned long nr_partials = 0; 4640 unsigned long nr_slabs = 0; 4641 unsigned long nr_inuse = 0; 4642 unsigned long nr_objs = 0; 4643 unsigned long nr_free = 0; 4644 struct kmem_cache *s; 4645 int node; 4646 4647 s = list_entry(p, struct kmem_cache, list); 4648 4649 for_each_online_node(node) { 4650 struct kmem_cache_node *n = get_node(s, node); 4651 4652 if (!n) 4653 continue; 4654 4655 nr_partials += n->nr_partial; 4656 nr_slabs += atomic_long_read(&n->nr_slabs); 4657 nr_objs += atomic_long_read(&n->total_objects); 4658 nr_free += count_partial(n, count_free); 4659 } 4660 4661 nr_inuse = nr_objs - nr_free; 4662 4663 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4664 nr_objs, s->size, oo_objects(s->oo), 4665 (1 << oo_order(s->oo))); 4666 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4667 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4668 0UL); 4669 seq_putc(m, '\n'); 4670 return 0; 4671} 4672 4673static const struct seq_operations slabinfo_op = { 4674 .start = s_start, 4675 .next = s_next, 4676 .stop = s_stop, 4677 .show = s_show, 4678}; 4679 4680static int slabinfo_open(struct inode *inode, struct file *file) 4681{ 4682 return seq_open(file, &slabinfo_op); 4683} 4684 4685static const struct file_operations proc_slabinfo_operations = { 4686 .open = slabinfo_open, 4687 .read = seq_read, 4688 .llseek = seq_lseek, 4689 .release = seq_release, 4690}; 4691 4692static int __init slab_proc_init(void) 4693{ 4694 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations); 4695 return 0; 4696} 4697module_init(slab_proc_init); 4698#endif /* CONFIG_SLABINFO */ 4699