slub.c revision dc1fb7f43636754a4d06f7bdb8ea3269a7d71d6d
1/* 2 * SLUB: A slab allocator that limits cache line use instead of queuing 3 * objects in per cpu and per node lists. 4 * 5 * The allocator synchronizes using per slab locks and only 6 * uses a centralized lock to manage a pool of partial slabs. 7 * 8 * (C) 2007 SGI, Christoph Lameter 9 */ 10 11#include <linux/mm.h> 12#include <linux/swap.h> /* struct reclaim_state */ 13#include <linux/module.h> 14#include <linux/bit_spinlock.h> 15#include <linux/interrupt.h> 16#include <linux/bitops.h> 17#include <linux/slab.h> 18#include <linux/proc_fs.h> 19#include <linux/seq_file.h> 20#include <linux/kmemcheck.h> 21#include <linux/cpu.h> 22#include <linux/cpuset.h> 23#include <linux/mempolicy.h> 24#include <linux/ctype.h> 25#include <linux/debugobjects.h> 26#include <linux/kallsyms.h> 27#include <linux/memory.h> 28#include <linux/math64.h> 29#include <linux/fault-inject.h> 30 31#include <trace/events/kmem.h> 32 33/* 34 * Lock order: 35 * 1. slab_lock(page) 36 * 2. slab->list_lock 37 * 38 * The slab_lock protects operations on the object of a particular 39 * slab and its metadata in the page struct. If the slab lock 40 * has been taken then no allocations nor frees can be performed 41 * on the objects in the slab nor can the slab be added or removed 42 * from the partial or full lists since this would mean modifying 43 * the page_struct of the slab. 44 * 45 * The list_lock protects the partial and full list on each node and 46 * the partial slab counter. If taken then no new slabs may be added or 47 * removed from the lists nor make the number of partial slabs be modified. 48 * (Note that the total number of slabs is an atomic value that may be 49 * modified without taking the list lock). 50 * 51 * The list_lock is a centralized lock and thus we avoid taking it as 52 * much as possible. As long as SLUB does not have to handle partial 53 * slabs, operations can continue without any centralized lock. F.e. 54 * allocating a long series of objects that fill up slabs does not require 55 * the list lock. 56 * 57 * The lock order is sometimes inverted when we are trying to get a slab 58 * off a list. We take the list_lock and then look for a page on the list 59 * to use. While we do that objects in the slabs may be freed. We can 60 * only operate on the slab if we have also taken the slab_lock. So we use 61 * a slab_trylock() on the slab. If trylock was successful then no frees 62 * can occur anymore and we can use the slab for allocations etc. If the 63 * slab_trylock() does not succeed then frees are in progress in the slab and 64 * we must stay away from it for a while since we may cause a bouncing 65 * cacheline if we try to acquire the lock. So go onto the next slab. 66 * If all pages are busy then we may allocate a new slab instead of reusing 67 * a partial slab. A new slab has noone operating on it and thus there is 68 * no danger of cacheline contention. 69 * 70 * Interrupts are disabled during allocation and deallocation in order to 71 * make the slab allocator safe to use in the context of an irq. In addition 72 * interrupts are disabled to ensure that the processor does not change 73 * while handling per_cpu slabs, due to kernel preemption. 74 * 75 * SLUB assigns one slab for allocation to each processor. 76 * Allocations only occur from these slabs called cpu slabs. 77 * 78 * Slabs with free elements are kept on a partial list and during regular 79 * operations no list for full slabs is used. If an object in a full slab is 80 * freed then the slab will show up again on the partial lists. 81 * We track full slabs for debugging purposes though because otherwise we 82 * cannot scan all objects. 83 * 84 * Slabs are freed when they become empty. Teardown and setup is 85 * minimal so we rely on the page allocators per cpu caches for 86 * fast frees and allocs. 87 * 88 * Overloading of page flags that are otherwise used for LRU management. 89 * 90 * PageActive The slab is frozen and exempt from list processing. 91 * This means that the slab is dedicated to a purpose 92 * such as satisfying allocations for a specific 93 * processor. Objects may be freed in the slab while 94 * it is frozen but slab_free will then skip the usual 95 * list operations. It is up to the processor holding 96 * the slab to integrate the slab into the slab lists 97 * when the slab is no longer needed. 98 * 99 * One use of this flag is to mark slabs that are 100 * used for allocations. Then such a slab becomes a cpu 101 * slab. The cpu slab may be equipped with an additional 102 * freelist that allows lockless access to 103 * free objects in addition to the regular freelist 104 * that requires the slab lock. 105 * 106 * PageError Slab requires special handling due to debug 107 * options set. This moves slab handling out of 108 * the fast path and disables lockless freelists. 109 */ 110 111#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 112 SLAB_TRACE | SLAB_DEBUG_FREE) 113 114static inline int kmem_cache_debug(struct kmem_cache *s) 115{ 116#ifdef CONFIG_SLUB_DEBUG 117 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 118#else 119 return 0; 120#endif 121} 122 123/* 124 * Issues still to be resolved: 125 * 126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 127 * 128 * - Variable sizing of the per node arrays 129 */ 130 131/* Enable to test recovery from slab corruption on boot */ 132#undef SLUB_RESILIENCY_TEST 133 134/* 135 * Mininum number of partial slabs. These will be left on the partial 136 * lists even if they are empty. kmem_cache_shrink may reclaim them. 137 */ 138#define MIN_PARTIAL 5 139 140/* 141 * Maximum number of desirable partial slabs. 142 * The existence of more partial slabs makes kmem_cache_shrink 143 * sort the partial list by the number of objects in the. 144 */ 145#define MAX_PARTIAL 10 146 147#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 148 SLAB_POISON | SLAB_STORE_USER) 149 150/* 151 * Debugging flags that require metadata to be stored in the slab. These get 152 * disabled when slub_debug=O is used and a cache's min order increases with 153 * metadata. 154 */ 155#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 156 157/* 158 * Set of flags that will prevent slab merging 159 */ 160#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ 162 SLAB_FAILSLAB) 163 164#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 165 SLAB_CACHE_DMA | SLAB_NOTRACK) 166 167#define OO_SHIFT 16 168#define OO_MASK ((1 << OO_SHIFT) - 1) 169#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ 170 171/* Internal SLUB flags */ 172#define __OBJECT_POISON 0x80000000UL /* Poison object */ 173 174static int kmem_size = sizeof(struct kmem_cache); 175 176#ifdef CONFIG_SMP 177static struct notifier_block slab_notifier; 178#endif 179 180static enum { 181 DOWN, /* No slab functionality available */ 182 PARTIAL, /* Kmem_cache_node works */ 183 UP, /* Everything works but does not show up in sysfs */ 184 SYSFS /* Sysfs up */ 185} slab_state = DOWN; 186 187/* A list of all slab caches on the system */ 188static DECLARE_RWSEM(slub_lock); 189static LIST_HEAD(slab_caches); 190 191/* 192 * Tracking user of a slab. 193 */ 194struct track { 195 unsigned long addr; /* Called from address */ 196 int cpu; /* Was running on cpu */ 197 int pid; /* Pid context */ 198 unsigned long when; /* When did the operation occur */ 199}; 200 201enum track_item { TRACK_ALLOC, TRACK_FREE }; 202 203#ifdef CONFIG_SYSFS 204static int sysfs_slab_add(struct kmem_cache *); 205static int sysfs_slab_alias(struct kmem_cache *, const char *); 206static void sysfs_slab_remove(struct kmem_cache *); 207 208#else 209static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 210static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 211 { return 0; } 212static inline void sysfs_slab_remove(struct kmem_cache *s) 213{ 214 kfree(s->name); 215 kfree(s); 216} 217 218#endif 219 220static inline void stat(const struct kmem_cache *s, enum stat_item si) 221{ 222#ifdef CONFIG_SLUB_STATS 223 __this_cpu_inc(s->cpu_slab->stat[si]); 224#endif 225} 226 227/******************************************************************** 228 * Core slab cache functions 229 *******************************************************************/ 230 231int slab_is_available(void) 232{ 233 return slab_state >= UP; 234} 235 236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 237{ 238 return s->node[node]; 239} 240 241/* Verify that a pointer has an address that is valid within a slab page */ 242static inline int check_valid_pointer(struct kmem_cache *s, 243 struct page *page, const void *object) 244{ 245 void *base; 246 247 if (!object) 248 return 1; 249 250 base = page_address(page); 251 if (object < base || object >= base + page->objects * s->size || 252 (object - base) % s->size) { 253 return 0; 254 } 255 256 return 1; 257} 258 259static inline void *get_freepointer(struct kmem_cache *s, void *object) 260{ 261 return *(void **)(object + s->offset); 262} 263 264static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 265{ 266 *(void **)(object + s->offset) = fp; 267} 268 269/* Loop over all objects in a slab */ 270#define for_each_object(__p, __s, __addr, __objects) \ 271 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 272 __p += (__s)->size) 273 274/* Determine object index from a given position */ 275static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 276{ 277 return (p - addr) / s->size; 278} 279 280static inline size_t slab_ksize(const struct kmem_cache *s) 281{ 282#ifdef CONFIG_SLUB_DEBUG 283 /* 284 * Debugging requires use of the padding between object 285 * and whatever may come after it. 286 */ 287 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 288 return s->objsize; 289 290#endif 291 /* 292 * If we have the need to store the freelist pointer 293 * back there or track user information then we can 294 * only use the space before that information. 295 */ 296 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 297 return s->inuse; 298 /* 299 * Else we can use all the padding etc for the allocation 300 */ 301 return s->size; 302} 303 304static inline int order_objects(int order, unsigned long size, int reserved) 305{ 306 return ((PAGE_SIZE << order) - reserved) / size; 307} 308 309static inline struct kmem_cache_order_objects oo_make(int order, 310 unsigned long size, int reserved) 311{ 312 struct kmem_cache_order_objects x = { 313 (order << OO_SHIFT) + order_objects(order, size, reserved) 314 }; 315 316 return x; 317} 318 319static inline int oo_order(struct kmem_cache_order_objects x) 320{ 321 return x.x >> OO_SHIFT; 322} 323 324static inline int oo_objects(struct kmem_cache_order_objects x) 325{ 326 return x.x & OO_MASK; 327} 328 329/* 330 * Determine a map of object in use on a page. 331 * 332 * Slab lock or node listlock must be held to guarantee that the page does 333 * not vanish from under us. 334 */ 335static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) 336{ 337 void *p; 338 void *addr = page_address(page); 339 340 for (p = page->freelist; p; p = get_freepointer(s, p)) 341 set_bit(slab_index(p, s, addr), map); 342} 343 344#ifdef CONFIG_SLUB_DEBUG 345/* 346 * Debug settings: 347 */ 348#ifdef CONFIG_SLUB_DEBUG_ON 349static int slub_debug = DEBUG_DEFAULT_FLAGS; 350#else 351static int slub_debug; 352#endif 353 354static char *slub_debug_slabs; 355static int disable_higher_order_debug; 356 357/* 358 * Object debugging 359 */ 360static void print_section(char *text, u8 *addr, unsigned int length) 361{ 362 int i, offset; 363 int newline = 1; 364 char ascii[17]; 365 366 ascii[16] = 0; 367 368 for (i = 0; i < length; i++) { 369 if (newline) { 370 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 371 newline = 0; 372 } 373 printk(KERN_CONT " %02x", addr[i]); 374 offset = i % 16; 375 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 376 if (offset == 15) { 377 printk(KERN_CONT " %s\n", ascii); 378 newline = 1; 379 } 380 } 381 if (!newline) { 382 i %= 16; 383 while (i < 16) { 384 printk(KERN_CONT " "); 385 ascii[i] = ' '; 386 i++; 387 } 388 printk(KERN_CONT " %s\n", ascii); 389 } 390} 391 392static struct track *get_track(struct kmem_cache *s, void *object, 393 enum track_item alloc) 394{ 395 struct track *p; 396 397 if (s->offset) 398 p = object + s->offset + sizeof(void *); 399 else 400 p = object + s->inuse; 401 402 return p + alloc; 403} 404 405static void set_track(struct kmem_cache *s, void *object, 406 enum track_item alloc, unsigned long addr) 407{ 408 struct track *p = get_track(s, object, alloc); 409 410 if (addr) { 411 p->addr = addr; 412 p->cpu = smp_processor_id(); 413 p->pid = current->pid; 414 p->when = jiffies; 415 } else 416 memset(p, 0, sizeof(struct track)); 417} 418 419static void init_tracking(struct kmem_cache *s, void *object) 420{ 421 if (!(s->flags & SLAB_STORE_USER)) 422 return; 423 424 set_track(s, object, TRACK_FREE, 0UL); 425 set_track(s, object, TRACK_ALLOC, 0UL); 426} 427 428static void print_track(const char *s, struct track *t) 429{ 430 if (!t->addr) 431 return; 432 433 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 434 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 435} 436 437static void print_tracking(struct kmem_cache *s, void *object) 438{ 439 if (!(s->flags & SLAB_STORE_USER)) 440 return; 441 442 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 443 print_track("Freed", get_track(s, object, TRACK_FREE)); 444} 445 446static void print_page_info(struct page *page) 447{ 448 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 449 page, page->objects, page->inuse, page->freelist, page->flags); 450 451} 452 453static void slab_bug(struct kmem_cache *s, char *fmt, ...) 454{ 455 va_list args; 456 char buf[100]; 457 458 va_start(args, fmt); 459 vsnprintf(buf, sizeof(buf), fmt, args); 460 va_end(args); 461 printk(KERN_ERR "========================================" 462 "=====================================\n"); 463 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 464 printk(KERN_ERR "----------------------------------------" 465 "-------------------------------------\n\n"); 466} 467 468static void slab_fix(struct kmem_cache *s, char *fmt, ...) 469{ 470 va_list args; 471 char buf[100]; 472 473 va_start(args, fmt); 474 vsnprintf(buf, sizeof(buf), fmt, args); 475 va_end(args); 476 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 477} 478 479static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 480{ 481 unsigned int off; /* Offset of last byte */ 482 u8 *addr = page_address(page); 483 484 print_tracking(s, p); 485 486 print_page_info(page); 487 488 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 489 p, p - addr, get_freepointer(s, p)); 490 491 if (p > addr + 16) 492 print_section("Bytes b4", p - 16, 16); 493 494 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 495 496 if (s->flags & SLAB_RED_ZONE) 497 print_section("Redzone", p + s->objsize, 498 s->inuse - s->objsize); 499 500 if (s->offset) 501 off = s->offset + sizeof(void *); 502 else 503 off = s->inuse; 504 505 if (s->flags & SLAB_STORE_USER) 506 off += 2 * sizeof(struct track); 507 508 if (off != s->size) 509 /* Beginning of the filler is the free pointer */ 510 print_section("Padding", p + off, s->size - off); 511 512 dump_stack(); 513} 514 515static void object_err(struct kmem_cache *s, struct page *page, 516 u8 *object, char *reason) 517{ 518 slab_bug(s, "%s", reason); 519 print_trailer(s, page, object); 520} 521 522static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 523{ 524 va_list args; 525 char buf[100]; 526 527 va_start(args, fmt); 528 vsnprintf(buf, sizeof(buf), fmt, args); 529 va_end(args); 530 slab_bug(s, "%s", buf); 531 print_page_info(page); 532 dump_stack(); 533} 534 535static void init_object(struct kmem_cache *s, void *object, u8 val) 536{ 537 u8 *p = object; 538 539 if (s->flags & __OBJECT_POISON) { 540 memset(p, POISON_FREE, s->objsize - 1); 541 p[s->objsize - 1] = POISON_END; 542 } 543 544 if (s->flags & SLAB_RED_ZONE) 545 memset(p + s->objsize, val, s->inuse - s->objsize); 546} 547 548static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 549{ 550 while (bytes) { 551 if (*start != (u8)value) 552 return start; 553 start++; 554 bytes--; 555 } 556 return NULL; 557} 558 559static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 560 void *from, void *to) 561{ 562 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 563 memset(from, data, to - from); 564} 565 566static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 567 u8 *object, char *what, 568 u8 *start, unsigned int value, unsigned int bytes) 569{ 570 u8 *fault; 571 u8 *end; 572 573 fault = check_bytes(start, value, bytes); 574 if (!fault) 575 return 1; 576 577 end = start + bytes; 578 while (end > fault && end[-1] == value) 579 end--; 580 581 slab_bug(s, "%s overwritten", what); 582 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 583 fault, end - 1, fault[0], value); 584 print_trailer(s, page, object); 585 586 restore_bytes(s, what, value, fault, end); 587 return 0; 588} 589 590/* 591 * Object layout: 592 * 593 * object address 594 * Bytes of the object to be managed. 595 * If the freepointer may overlay the object then the free 596 * pointer is the first word of the object. 597 * 598 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 599 * 0xa5 (POISON_END) 600 * 601 * object + s->objsize 602 * Padding to reach word boundary. This is also used for Redzoning. 603 * Padding is extended by another word if Redzoning is enabled and 604 * objsize == inuse. 605 * 606 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 607 * 0xcc (RED_ACTIVE) for objects in use. 608 * 609 * object + s->inuse 610 * Meta data starts here. 611 * 612 * A. Free pointer (if we cannot overwrite object on free) 613 * B. Tracking data for SLAB_STORE_USER 614 * C. Padding to reach required alignment boundary or at mininum 615 * one word if debugging is on to be able to detect writes 616 * before the word boundary. 617 * 618 * Padding is done using 0x5a (POISON_INUSE) 619 * 620 * object + s->size 621 * Nothing is used beyond s->size. 622 * 623 * If slabcaches are merged then the objsize and inuse boundaries are mostly 624 * ignored. And therefore no slab options that rely on these boundaries 625 * may be used with merged slabcaches. 626 */ 627 628static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 629{ 630 unsigned long off = s->inuse; /* The end of info */ 631 632 if (s->offset) 633 /* Freepointer is placed after the object. */ 634 off += sizeof(void *); 635 636 if (s->flags & SLAB_STORE_USER) 637 /* We also have user information there */ 638 off += 2 * sizeof(struct track); 639 640 if (s->size == off) 641 return 1; 642 643 return check_bytes_and_report(s, page, p, "Object padding", 644 p + off, POISON_INUSE, s->size - off); 645} 646 647/* Check the pad bytes at the end of a slab page */ 648static int slab_pad_check(struct kmem_cache *s, struct page *page) 649{ 650 u8 *start; 651 u8 *fault; 652 u8 *end; 653 int length; 654 int remainder; 655 656 if (!(s->flags & SLAB_POISON)) 657 return 1; 658 659 start = page_address(page); 660 length = (PAGE_SIZE << compound_order(page)) - s->reserved; 661 end = start + length; 662 remainder = length % s->size; 663 if (!remainder) 664 return 1; 665 666 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 667 if (!fault) 668 return 1; 669 while (end > fault && end[-1] == POISON_INUSE) 670 end--; 671 672 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 673 print_section("Padding", end - remainder, remainder); 674 675 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 676 return 0; 677} 678 679static int check_object(struct kmem_cache *s, struct page *page, 680 void *object, u8 val) 681{ 682 u8 *p = object; 683 u8 *endobject = object + s->objsize; 684 685 if (s->flags & SLAB_RED_ZONE) { 686 if (!check_bytes_and_report(s, page, object, "Redzone", 687 endobject, val, s->inuse - s->objsize)) 688 return 0; 689 } else { 690 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 691 check_bytes_and_report(s, page, p, "Alignment padding", 692 endobject, POISON_INUSE, s->inuse - s->objsize); 693 } 694 } 695 696 if (s->flags & SLAB_POISON) { 697 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 698 (!check_bytes_and_report(s, page, p, "Poison", p, 699 POISON_FREE, s->objsize - 1) || 700 !check_bytes_and_report(s, page, p, "Poison", 701 p + s->objsize - 1, POISON_END, 1))) 702 return 0; 703 /* 704 * check_pad_bytes cleans up on its own. 705 */ 706 check_pad_bytes(s, page, p); 707 } 708 709 if (!s->offset && val == SLUB_RED_ACTIVE) 710 /* 711 * Object and freepointer overlap. Cannot check 712 * freepointer while object is allocated. 713 */ 714 return 1; 715 716 /* Check free pointer validity */ 717 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 718 object_err(s, page, p, "Freepointer corrupt"); 719 /* 720 * No choice but to zap it and thus lose the remainder 721 * of the free objects in this slab. May cause 722 * another error because the object count is now wrong. 723 */ 724 set_freepointer(s, p, NULL); 725 return 0; 726 } 727 return 1; 728} 729 730static int check_slab(struct kmem_cache *s, struct page *page) 731{ 732 int maxobj; 733 734 VM_BUG_ON(!irqs_disabled()); 735 736 if (!PageSlab(page)) { 737 slab_err(s, page, "Not a valid slab page"); 738 return 0; 739 } 740 741 maxobj = order_objects(compound_order(page), s->size, s->reserved); 742 if (page->objects > maxobj) { 743 slab_err(s, page, "objects %u > max %u", 744 s->name, page->objects, maxobj); 745 return 0; 746 } 747 if (page->inuse > page->objects) { 748 slab_err(s, page, "inuse %u > max %u", 749 s->name, page->inuse, page->objects); 750 return 0; 751 } 752 /* Slab_pad_check fixes things up after itself */ 753 slab_pad_check(s, page); 754 return 1; 755} 756 757/* 758 * Determine if a certain object on a page is on the freelist. Must hold the 759 * slab lock to guarantee that the chains are in a consistent state. 760 */ 761static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 762{ 763 int nr = 0; 764 void *fp = page->freelist; 765 void *object = NULL; 766 unsigned long max_objects; 767 768 while (fp && nr <= page->objects) { 769 if (fp == search) 770 return 1; 771 if (!check_valid_pointer(s, page, fp)) { 772 if (object) { 773 object_err(s, page, object, 774 "Freechain corrupt"); 775 set_freepointer(s, object, NULL); 776 break; 777 } else { 778 slab_err(s, page, "Freepointer corrupt"); 779 page->freelist = NULL; 780 page->inuse = page->objects; 781 slab_fix(s, "Freelist cleared"); 782 return 0; 783 } 784 break; 785 } 786 object = fp; 787 fp = get_freepointer(s, object); 788 nr++; 789 } 790 791 max_objects = order_objects(compound_order(page), s->size, s->reserved); 792 if (max_objects > MAX_OBJS_PER_PAGE) 793 max_objects = MAX_OBJS_PER_PAGE; 794 795 if (page->objects != max_objects) { 796 slab_err(s, page, "Wrong number of objects. Found %d but " 797 "should be %d", page->objects, max_objects); 798 page->objects = max_objects; 799 slab_fix(s, "Number of objects adjusted."); 800 } 801 if (page->inuse != page->objects - nr) { 802 slab_err(s, page, "Wrong object count. Counter is %d but " 803 "counted were %d", page->inuse, page->objects - nr); 804 page->inuse = page->objects - nr; 805 slab_fix(s, "Object count adjusted."); 806 } 807 return search == NULL; 808} 809 810static void trace(struct kmem_cache *s, struct page *page, void *object, 811 int alloc) 812{ 813 if (s->flags & SLAB_TRACE) { 814 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 815 s->name, 816 alloc ? "alloc" : "free", 817 object, page->inuse, 818 page->freelist); 819 820 if (!alloc) 821 print_section("Object", (void *)object, s->objsize); 822 823 dump_stack(); 824 } 825} 826 827/* 828 * Hooks for other subsystems that check memory allocations. In a typical 829 * production configuration these hooks all should produce no code at all. 830 */ 831static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 832{ 833 flags &= gfp_allowed_mask; 834 lockdep_trace_alloc(flags); 835 might_sleep_if(flags & __GFP_WAIT); 836 837 return should_failslab(s->objsize, flags, s->flags); 838} 839 840static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object) 841{ 842 flags &= gfp_allowed_mask; 843 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s)); 844 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags); 845} 846 847static inline void slab_free_hook(struct kmem_cache *s, void *x) 848{ 849 kmemleak_free_recursive(x, s->flags); 850 851 /* 852 * Trouble is that we may no longer disable interupts in the fast path 853 * So in order to make the debug calls that expect irqs to be 854 * disabled we need to disable interrupts temporarily. 855 */ 856#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) 857 { 858 unsigned long flags; 859 860 local_irq_save(flags); 861 kmemcheck_slab_free(s, x, s->objsize); 862 debug_check_no_locks_freed(x, s->objsize); 863 local_irq_restore(flags); 864 } 865#endif 866 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 867 debug_check_no_obj_freed(x, s->objsize); 868} 869 870/* 871 * Tracking of fully allocated slabs for debugging purposes. 872 */ 873static void add_full(struct kmem_cache_node *n, struct page *page) 874{ 875 spin_lock(&n->list_lock); 876 list_add(&page->lru, &n->full); 877 spin_unlock(&n->list_lock); 878} 879 880static void remove_full(struct kmem_cache *s, struct page *page) 881{ 882 struct kmem_cache_node *n; 883 884 if (!(s->flags & SLAB_STORE_USER)) 885 return; 886 887 n = get_node(s, page_to_nid(page)); 888 889 spin_lock(&n->list_lock); 890 list_del(&page->lru); 891 spin_unlock(&n->list_lock); 892} 893 894/* Tracking of the number of slabs for debugging purposes */ 895static inline unsigned long slabs_node(struct kmem_cache *s, int node) 896{ 897 struct kmem_cache_node *n = get_node(s, node); 898 899 return atomic_long_read(&n->nr_slabs); 900} 901 902static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 903{ 904 return atomic_long_read(&n->nr_slabs); 905} 906 907static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 908{ 909 struct kmem_cache_node *n = get_node(s, node); 910 911 /* 912 * May be called early in order to allocate a slab for the 913 * kmem_cache_node structure. Solve the chicken-egg 914 * dilemma by deferring the increment of the count during 915 * bootstrap (see early_kmem_cache_node_alloc). 916 */ 917 if (n) { 918 atomic_long_inc(&n->nr_slabs); 919 atomic_long_add(objects, &n->total_objects); 920 } 921} 922static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 923{ 924 struct kmem_cache_node *n = get_node(s, node); 925 926 atomic_long_dec(&n->nr_slabs); 927 atomic_long_sub(objects, &n->total_objects); 928} 929 930/* Object debug checks for alloc/free paths */ 931static void setup_object_debug(struct kmem_cache *s, struct page *page, 932 void *object) 933{ 934 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 935 return; 936 937 init_object(s, object, SLUB_RED_INACTIVE); 938 init_tracking(s, object); 939} 940 941static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page, 942 void *object, unsigned long addr) 943{ 944 if (!check_slab(s, page)) 945 goto bad; 946 947 if (!on_freelist(s, page, object)) { 948 object_err(s, page, object, "Object already allocated"); 949 goto bad; 950 } 951 952 if (!check_valid_pointer(s, page, object)) { 953 object_err(s, page, object, "Freelist Pointer check fails"); 954 goto bad; 955 } 956 957 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 958 goto bad; 959 960 /* Success perform special debug activities for allocs */ 961 if (s->flags & SLAB_STORE_USER) 962 set_track(s, object, TRACK_ALLOC, addr); 963 trace(s, page, object, 1); 964 init_object(s, object, SLUB_RED_ACTIVE); 965 return 1; 966 967bad: 968 if (PageSlab(page)) { 969 /* 970 * If this is a slab page then lets do the best we can 971 * to avoid issues in the future. Marking all objects 972 * as used avoids touching the remaining objects. 973 */ 974 slab_fix(s, "Marking all objects used"); 975 page->inuse = page->objects; 976 page->freelist = NULL; 977 } 978 return 0; 979} 980 981static noinline int free_debug_processing(struct kmem_cache *s, 982 struct page *page, void *object, unsigned long addr) 983{ 984 if (!check_slab(s, page)) 985 goto fail; 986 987 if (!check_valid_pointer(s, page, object)) { 988 slab_err(s, page, "Invalid object pointer 0x%p", object); 989 goto fail; 990 } 991 992 if (on_freelist(s, page, object)) { 993 object_err(s, page, object, "Object already free"); 994 goto fail; 995 } 996 997 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 998 return 0; 999 1000 if (unlikely(s != page->slab)) { 1001 if (!PageSlab(page)) { 1002 slab_err(s, page, "Attempt to free object(0x%p) " 1003 "outside of slab", object); 1004 } else if (!page->slab) { 1005 printk(KERN_ERR 1006 "SLUB <none>: no slab for object 0x%p.\n", 1007 object); 1008 dump_stack(); 1009 } else 1010 object_err(s, page, object, 1011 "page slab pointer corrupt."); 1012 goto fail; 1013 } 1014 1015 /* Special debug activities for freeing objects */ 1016 if (!PageSlubFrozen(page) && !page->freelist) 1017 remove_full(s, page); 1018 if (s->flags & SLAB_STORE_USER) 1019 set_track(s, object, TRACK_FREE, addr); 1020 trace(s, page, object, 0); 1021 init_object(s, object, SLUB_RED_INACTIVE); 1022 return 1; 1023 1024fail: 1025 slab_fix(s, "Object at 0x%p not freed", object); 1026 return 0; 1027} 1028 1029static int __init setup_slub_debug(char *str) 1030{ 1031 slub_debug = DEBUG_DEFAULT_FLAGS; 1032 if (*str++ != '=' || !*str) 1033 /* 1034 * No options specified. Switch on full debugging. 1035 */ 1036 goto out; 1037 1038 if (*str == ',') 1039 /* 1040 * No options but restriction on slabs. This means full 1041 * debugging for slabs matching a pattern. 1042 */ 1043 goto check_slabs; 1044 1045 if (tolower(*str) == 'o') { 1046 /* 1047 * Avoid enabling debugging on caches if its minimum order 1048 * would increase as a result. 1049 */ 1050 disable_higher_order_debug = 1; 1051 goto out; 1052 } 1053 1054 slub_debug = 0; 1055 if (*str == '-') 1056 /* 1057 * Switch off all debugging measures. 1058 */ 1059 goto out; 1060 1061 /* 1062 * Determine which debug features should be switched on 1063 */ 1064 for (; *str && *str != ','; str++) { 1065 switch (tolower(*str)) { 1066 case 'f': 1067 slub_debug |= SLAB_DEBUG_FREE; 1068 break; 1069 case 'z': 1070 slub_debug |= SLAB_RED_ZONE; 1071 break; 1072 case 'p': 1073 slub_debug |= SLAB_POISON; 1074 break; 1075 case 'u': 1076 slub_debug |= SLAB_STORE_USER; 1077 break; 1078 case 't': 1079 slub_debug |= SLAB_TRACE; 1080 break; 1081 case 'a': 1082 slub_debug |= SLAB_FAILSLAB; 1083 break; 1084 default: 1085 printk(KERN_ERR "slub_debug option '%c' " 1086 "unknown. skipped\n", *str); 1087 } 1088 } 1089 1090check_slabs: 1091 if (*str == ',') 1092 slub_debug_slabs = str + 1; 1093out: 1094 return 1; 1095} 1096 1097__setup("slub_debug", setup_slub_debug); 1098 1099static unsigned long kmem_cache_flags(unsigned long objsize, 1100 unsigned long flags, const char *name, 1101 void (*ctor)(void *)) 1102{ 1103 /* 1104 * Enable debugging if selected on the kernel commandline. 1105 */ 1106 if (slub_debug && (!slub_debug_slabs || 1107 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) 1108 flags |= slub_debug; 1109 1110 return flags; 1111} 1112#else 1113static inline void setup_object_debug(struct kmem_cache *s, 1114 struct page *page, void *object) {} 1115 1116static inline int alloc_debug_processing(struct kmem_cache *s, 1117 struct page *page, void *object, unsigned long addr) { return 0; } 1118 1119static inline int free_debug_processing(struct kmem_cache *s, 1120 struct page *page, void *object, unsigned long addr) { return 0; } 1121 1122static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1123 { return 1; } 1124static inline int check_object(struct kmem_cache *s, struct page *page, 1125 void *object, u8 val) { return 1; } 1126static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1127static inline unsigned long kmem_cache_flags(unsigned long objsize, 1128 unsigned long flags, const char *name, 1129 void (*ctor)(void *)) 1130{ 1131 return flags; 1132} 1133#define slub_debug 0 1134 1135#define disable_higher_order_debug 0 1136 1137static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1138 { return 0; } 1139static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1140 { return 0; } 1141static inline void inc_slabs_node(struct kmem_cache *s, int node, 1142 int objects) {} 1143static inline void dec_slabs_node(struct kmem_cache *s, int node, 1144 int objects) {} 1145 1146static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 1147 { return 0; } 1148 1149static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, 1150 void *object) {} 1151 1152static inline void slab_free_hook(struct kmem_cache *s, void *x) {} 1153 1154#endif /* CONFIG_SLUB_DEBUG */ 1155 1156/* 1157 * Slab allocation and freeing 1158 */ 1159static inline struct page *alloc_slab_page(gfp_t flags, int node, 1160 struct kmem_cache_order_objects oo) 1161{ 1162 int order = oo_order(oo); 1163 1164 flags |= __GFP_NOTRACK; 1165 1166 if (node == NUMA_NO_NODE) 1167 return alloc_pages(flags, order); 1168 else 1169 return alloc_pages_exact_node(node, flags, order); 1170} 1171 1172static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1173{ 1174 struct page *page; 1175 struct kmem_cache_order_objects oo = s->oo; 1176 gfp_t alloc_gfp; 1177 1178 flags |= s->allocflags; 1179 1180 /* 1181 * Let the initial higher-order allocation fail under memory pressure 1182 * so we fall-back to the minimum order allocation. 1183 */ 1184 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1185 1186 page = alloc_slab_page(alloc_gfp, node, oo); 1187 if (unlikely(!page)) { 1188 oo = s->min; 1189 /* 1190 * Allocation may have failed due to fragmentation. 1191 * Try a lower order alloc if possible 1192 */ 1193 page = alloc_slab_page(flags, node, oo); 1194 if (!page) 1195 return NULL; 1196 1197 stat(s, ORDER_FALLBACK); 1198 } 1199 1200 if (kmemcheck_enabled 1201 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1202 int pages = 1 << oo_order(oo); 1203 1204 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); 1205 1206 /* 1207 * Objects from caches that have a constructor don't get 1208 * cleared when they're allocated, so we need to do it here. 1209 */ 1210 if (s->ctor) 1211 kmemcheck_mark_uninitialized_pages(page, pages); 1212 else 1213 kmemcheck_mark_unallocated_pages(page, pages); 1214 } 1215 1216 page->objects = oo_objects(oo); 1217 mod_zone_page_state(page_zone(page), 1218 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1219 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1220 1 << oo_order(oo)); 1221 1222 return page; 1223} 1224 1225static void setup_object(struct kmem_cache *s, struct page *page, 1226 void *object) 1227{ 1228 setup_object_debug(s, page, object); 1229 if (unlikely(s->ctor)) 1230 s->ctor(object); 1231} 1232 1233static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1234{ 1235 struct page *page; 1236 void *start; 1237 void *last; 1238 void *p; 1239 1240 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1241 1242 page = allocate_slab(s, 1243 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1244 if (!page) 1245 goto out; 1246 1247 inc_slabs_node(s, page_to_nid(page), page->objects); 1248 page->slab = s; 1249 page->flags |= 1 << PG_slab; 1250 1251 start = page_address(page); 1252 1253 if (unlikely(s->flags & SLAB_POISON)) 1254 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1255 1256 last = start; 1257 for_each_object(p, s, start, page->objects) { 1258 setup_object(s, page, last); 1259 set_freepointer(s, last, p); 1260 last = p; 1261 } 1262 setup_object(s, page, last); 1263 set_freepointer(s, last, NULL); 1264 1265 page->freelist = start; 1266 page->inuse = 0; 1267out: 1268 return page; 1269} 1270 1271static void __free_slab(struct kmem_cache *s, struct page *page) 1272{ 1273 int order = compound_order(page); 1274 int pages = 1 << order; 1275 1276 if (kmem_cache_debug(s)) { 1277 void *p; 1278 1279 slab_pad_check(s, page); 1280 for_each_object(p, s, page_address(page), 1281 page->objects) 1282 check_object(s, page, p, SLUB_RED_INACTIVE); 1283 } 1284 1285 kmemcheck_free_shadow(page, compound_order(page)); 1286 1287 mod_zone_page_state(page_zone(page), 1288 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1289 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1290 -pages); 1291 1292 __ClearPageSlab(page); 1293 reset_page_mapcount(page); 1294 if (current->reclaim_state) 1295 current->reclaim_state->reclaimed_slab += pages; 1296 __free_pages(page, order); 1297} 1298 1299#define need_reserve_slab_rcu \ 1300 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) 1301 1302static void rcu_free_slab(struct rcu_head *h) 1303{ 1304 struct page *page; 1305 1306 if (need_reserve_slab_rcu) 1307 page = virt_to_head_page(h); 1308 else 1309 page = container_of((struct list_head *)h, struct page, lru); 1310 1311 __free_slab(page->slab, page); 1312} 1313 1314static void free_slab(struct kmem_cache *s, struct page *page) 1315{ 1316 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1317 struct rcu_head *head; 1318 1319 if (need_reserve_slab_rcu) { 1320 int order = compound_order(page); 1321 int offset = (PAGE_SIZE << order) - s->reserved; 1322 1323 VM_BUG_ON(s->reserved != sizeof(*head)); 1324 head = page_address(page) + offset; 1325 } else { 1326 /* 1327 * RCU free overloads the RCU head over the LRU 1328 */ 1329 head = (void *)&page->lru; 1330 } 1331 1332 call_rcu(head, rcu_free_slab); 1333 } else 1334 __free_slab(s, page); 1335} 1336 1337static void discard_slab(struct kmem_cache *s, struct page *page) 1338{ 1339 dec_slabs_node(s, page_to_nid(page), page->objects); 1340 free_slab(s, page); 1341} 1342 1343/* 1344 * Per slab locking using the pagelock 1345 */ 1346static __always_inline void slab_lock(struct page *page) 1347{ 1348 bit_spin_lock(PG_locked, &page->flags); 1349} 1350 1351static __always_inline void slab_unlock(struct page *page) 1352{ 1353 __bit_spin_unlock(PG_locked, &page->flags); 1354} 1355 1356static __always_inline int slab_trylock(struct page *page) 1357{ 1358 int rc = 1; 1359 1360 rc = bit_spin_trylock(PG_locked, &page->flags); 1361 return rc; 1362} 1363 1364/* 1365 * Management of partially allocated slabs 1366 */ 1367static void add_partial(struct kmem_cache_node *n, 1368 struct page *page, int tail) 1369{ 1370 spin_lock(&n->list_lock); 1371 n->nr_partial++; 1372 if (tail) 1373 list_add_tail(&page->lru, &n->partial); 1374 else 1375 list_add(&page->lru, &n->partial); 1376 spin_unlock(&n->list_lock); 1377} 1378 1379static inline void __remove_partial(struct kmem_cache_node *n, 1380 struct page *page) 1381{ 1382 list_del(&page->lru); 1383 n->nr_partial--; 1384} 1385 1386static void remove_partial(struct kmem_cache *s, struct page *page) 1387{ 1388 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1389 1390 spin_lock(&n->list_lock); 1391 __remove_partial(n, page); 1392 spin_unlock(&n->list_lock); 1393} 1394 1395/* 1396 * Lock slab and remove from the partial list. 1397 * 1398 * Must hold list_lock. 1399 */ 1400static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1401 struct page *page) 1402{ 1403 if (slab_trylock(page)) { 1404 __remove_partial(n, page); 1405 __SetPageSlubFrozen(page); 1406 return 1; 1407 } 1408 return 0; 1409} 1410 1411/* 1412 * Try to allocate a partial slab from a specific node. 1413 */ 1414static struct page *get_partial_node(struct kmem_cache_node *n) 1415{ 1416 struct page *page; 1417 1418 /* 1419 * Racy check. If we mistakenly see no partial slabs then we 1420 * just allocate an empty slab. If we mistakenly try to get a 1421 * partial slab and there is none available then get_partials() 1422 * will return NULL. 1423 */ 1424 if (!n || !n->nr_partial) 1425 return NULL; 1426 1427 spin_lock(&n->list_lock); 1428 list_for_each_entry(page, &n->partial, lru) 1429 if (lock_and_freeze_slab(n, page)) 1430 goto out; 1431 page = NULL; 1432out: 1433 spin_unlock(&n->list_lock); 1434 return page; 1435} 1436 1437/* 1438 * Get a page from somewhere. Search in increasing NUMA distances. 1439 */ 1440static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1441{ 1442#ifdef CONFIG_NUMA 1443 struct zonelist *zonelist; 1444 struct zoneref *z; 1445 struct zone *zone; 1446 enum zone_type high_zoneidx = gfp_zone(flags); 1447 struct page *page; 1448 1449 /* 1450 * The defrag ratio allows a configuration of the tradeoffs between 1451 * inter node defragmentation and node local allocations. A lower 1452 * defrag_ratio increases the tendency to do local allocations 1453 * instead of attempting to obtain partial slabs from other nodes. 1454 * 1455 * If the defrag_ratio is set to 0 then kmalloc() always 1456 * returns node local objects. If the ratio is higher then kmalloc() 1457 * may return off node objects because partial slabs are obtained 1458 * from other nodes and filled up. 1459 * 1460 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1461 * defrag_ratio = 1000) then every (well almost) allocation will 1462 * first attempt to defrag slab caches on other nodes. This means 1463 * scanning over all nodes to look for partial slabs which may be 1464 * expensive if we do it every time we are trying to find a slab 1465 * with available objects. 1466 */ 1467 if (!s->remote_node_defrag_ratio || 1468 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1469 return NULL; 1470 1471 get_mems_allowed(); 1472 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1473 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1474 struct kmem_cache_node *n; 1475 1476 n = get_node(s, zone_to_nid(zone)); 1477 1478 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1479 n->nr_partial > s->min_partial) { 1480 page = get_partial_node(n); 1481 if (page) { 1482 put_mems_allowed(); 1483 return page; 1484 } 1485 } 1486 } 1487 put_mems_allowed(); 1488#endif 1489 return NULL; 1490} 1491 1492/* 1493 * Get a partial page, lock it and return it. 1494 */ 1495static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1496{ 1497 struct page *page; 1498 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; 1499 1500 page = get_partial_node(get_node(s, searchnode)); 1501 if (page || node != NUMA_NO_NODE) 1502 return page; 1503 1504 return get_any_partial(s, flags); 1505} 1506 1507/* 1508 * Move a page back to the lists. 1509 * 1510 * Must be called with the slab lock held. 1511 * 1512 * On exit the slab lock will have been dropped. 1513 */ 1514static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1515 __releases(bitlock) 1516{ 1517 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1518 1519 __ClearPageSlubFrozen(page); 1520 if (page->inuse) { 1521 1522 if (page->freelist) { 1523 add_partial(n, page, tail); 1524 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1525 } else { 1526 stat(s, DEACTIVATE_FULL); 1527 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER)) 1528 add_full(n, page); 1529 } 1530 slab_unlock(page); 1531 } else { 1532 stat(s, DEACTIVATE_EMPTY); 1533 if (n->nr_partial < s->min_partial) { 1534 /* 1535 * Adding an empty slab to the partial slabs in order 1536 * to avoid page allocator overhead. This slab needs 1537 * to come after the other slabs with objects in 1538 * so that the others get filled first. That way the 1539 * size of the partial list stays small. 1540 * 1541 * kmem_cache_shrink can reclaim any empty slabs from 1542 * the partial list. 1543 */ 1544 add_partial(n, page, 1); 1545 slab_unlock(page); 1546 } else { 1547 slab_unlock(page); 1548 stat(s, FREE_SLAB); 1549 discard_slab(s, page); 1550 } 1551 } 1552} 1553 1554#ifdef CONFIG_CMPXCHG_LOCAL 1555#ifdef CONFIG_PREEMPT 1556/* 1557 * Calculate the next globally unique transaction for disambiguiation 1558 * during cmpxchg. The transactions start with the cpu number and are then 1559 * incremented by CONFIG_NR_CPUS. 1560 */ 1561#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 1562#else 1563/* 1564 * No preemption supported therefore also no need to check for 1565 * different cpus. 1566 */ 1567#define TID_STEP 1 1568#endif 1569 1570static inline unsigned long next_tid(unsigned long tid) 1571{ 1572 return tid + TID_STEP; 1573} 1574 1575static inline unsigned int tid_to_cpu(unsigned long tid) 1576{ 1577 return tid % TID_STEP; 1578} 1579 1580static inline unsigned long tid_to_event(unsigned long tid) 1581{ 1582 return tid / TID_STEP; 1583} 1584 1585static inline unsigned int init_tid(int cpu) 1586{ 1587 return cpu; 1588} 1589 1590static inline void note_cmpxchg_failure(const char *n, 1591 const struct kmem_cache *s, unsigned long tid) 1592{ 1593#ifdef SLUB_DEBUG_CMPXCHG 1594 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 1595 1596 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); 1597 1598#ifdef CONFIG_PREEMPT 1599 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 1600 printk("due to cpu change %d -> %d\n", 1601 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 1602 else 1603#endif 1604 if (tid_to_event(tid) != tid_to_event(actual_tid)) 1605 printk("due to cpu running other code. Event %ld->%ld\n", 1606 tid_to_event(tid), tid_to_event(actual_tid)); 1607 else 1608 printk("for unknown reason: actual=%lx was=%lx target=%lx\n", 1609 actual_tid, tid, next_tid(tid)); 1610#endif 1611 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 1612} 1613 1614#endif 1615 1616void init_kmem_cache_cpus(struct kmem_cache *s) 1617{ 1618#ifdef CONFIG_CMPXCHG_LOCAL 1619 int cpu; 1620 1621 for_each_possible_cpu(cpu) 1622 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 1623#endif 1624 1625} 1626/* 1627 * Remove the cpu slab 1628 */ 1629static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1630 __releases(bitlock) 1631{ 1632 struct page *page = c->page; 1633 int tail = 1; 1634 1635 if (page->freelist) 1636 stat(s, DEACTIVATE_REMOTE_FREES); 1637 /* 1638 * Merge cpu freelist into slab freelist. Typically we get here 1639 * because both freelists are empty. So this is unlikely 1640 * to occur. 1641 */ 1642 while (unlikely(c->freelist)) { 1643 void **object; 1644 1645 tail = 0; /* Hot objects. Put the slab first */ 1646 1647 /* Retrieve object from cpu_freelist */ 1648 object = c->freelist; 1649 c->freelist = get_freepointer(s, c->freelist); 1650 1651 /* And put onto the regular freelist */ 1652 set_freepointer(s, object, page->freelist); 1653 page->freelist = object; 1654 page->inuse--; 1655 } 1656 c->page = NULL; 1657#ifdef CONFIG_CMPXCHG_LOCAL 1658 c->tid = next_tid(c->tid); 1659#endif 1660 unfreeze_slab(s, page, tail); 1661} 1662 1663static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1664{ 1665 stat(s, CPUSLAB_FLUSH); 1666 slab_lock(c->page); 1667 deactivate_slab(s, c); 1668} 1669 1670/* 1671 * Flush cpu slab. 1672 * 1673 * Called from IPI handler with interrupts disabled. 1674 */ 1675static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1676{ 1677 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 1678 1679 if (likely(c && c->page)) 1680 flush_slab(s, c); 1681} 1682 1683static void flush_cpu_slab(void *d) 1684{ 1685 struct kmem_cache *s = d; 1686 1687 __flush_cpu_slab(s, smp_processor_id()); 1688} 1689 1690static void flush_all(struct kmem_cache *s) 1691{ 1692 on_each_cpu(flush_cpu_slab, s, 1); 1693} 1694 1695/* 1696 * Check if the objects in a per cpu structure fit numa 1697 * locality expectations. 1698 */ 1699static inline int node_match(struct kmem_cache_cpu *c, int node) 1700{ 1701#ifdef CONFIG_NUMA 1702 if (node != NUMA_NO_NODE && c->node != node) 1703 return 0; 1704#endif 1705 return 1; 1706} 1707 1708static int count_free(struct page *page) 1709{ 1710 return page->objects - page->inuse; 1711} 1712 1713static unsigned long count_partial(struct kmem_cache_node *n, 1714 int (*get_count)(struct page *)) 1715{ 1716 unsigned long flags; 1717 unsigned long x = 0; 1718 struct page *page; 1719 1720 spin_lock_irqsave(&n->list_lock, flags); 1721 list_for_each_entry(page, &n->partial, lru) 1722 x += get_count(page); 1723 spin_unlock_irqrestore(&n->list_lock, flags); 1724 return x; 1725} 1726 1727static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 1728{ 1729#ifdef CONFIG_SLUB_DEBUG 1730 return atomic_long_read(&n->total_objects); 1731#else 1732 return 0; 1733#endif 1734} 1735 1736static noinline void 1737slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 1738{ 1739 int node; 1740 1741 printk(KERN_WARNING 1742 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 1743 nid, gfpflags); 1744 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 1745 "default order: %d, min order: %d\n", s->name, s->objsize, 1746 s->size, oo_order(s->oo), oo_order(s->min)); 1747 1748 if (oo_order(s->min) > get_order(s->objsize)) 1749 printk(KERN_WARNING " %s debugging increased min order, use " 1750 "slub_debug=O to disable.\n", s->name); 1751 1752 for_each_online_node(node) { 1753 struct kmem_cache_node *n = get_node(s, node); 1754 unsigned long nr_slabs; 1755 unsigned long nr_objs; 1756 unsigned long nr_free; 1757 1758 if (!n) 1759 continue; 1760 1761 nr_free = count_partial(n, count_free); 1762 nr_slabs = node_nr_slabs(n); 1763 nr_objs = node_nr_objs(n); 1764 1765 printk(KERN_WARNING 1766 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 1767 node, nr_slabs, nr_objs, nr_free); 1768 } 1769} 1770 1771/* 1772 * Slow path. The lockless freelist is empty or we need to perform 1773 * debugging duties. 1774 * 1775 * Interrupts are disabled. 1776 * 1777 * Processing is still very fast if new objects have been freed to the 1778 * regular freelist. In that case we simply take over the regular freelist 1779 * as the lockless freelist and zap the regular freelist. 1780 * 1781 * If that is not working then we fall back to the partial lists. We take the 1782 * first element of the freelist as the object to allocate now and move the 1783 * rest of the freelist to the lockless freelist. 1784 * 1785 * And if we were unable to get a new slab from the partial slab lists then 1786 * we need to allocate a new slab. This is the slowest path since it involves 1787 * a call to the page allocator and the setup of a new slab. 1788 */ 1789static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 1790 unsigned long addr, struct kmem_cache_cpu *c) 1791{ 1792 void **object; 1793 struct page *page; 1794#ifdef CONFIG_CMPXCHG_LOCAL 1795 unsigned long flags; 1796 1797 local_irq_save(flags); 1798#ifdef CONFIG_PREEMPT 1799 /* 1800 * We may have been preempted and rescheduled on a different 1801 * cpu before disabling interrupts. Need to reload cpu area 1802 * pointer. 1803 */ 1804 c = this_cpu_ptr(s->cpu_slab); 1805#endif 1806#endif 1807 1808 /* We handle __GFP_ZERO in the caller */ 1809 gfpflags &= ~__GFP_ZERO; 1810 1811 page = c->page; 1812 if (!page) 1813 goto new_slab; 1814 1815 slab_lock(page); 1816 if (unlikely(!node_match(c, node))) 1817 goto another_slab; 1818 1819 stat(s, ALLOC_REFILL); 1820 1821load_freelist: 1822 object = page->freelist; 1823 if (unlikely(!object)) 1824 goto another_slab; 1825 if (kmem_cache_debug(s)) 1826 goto debug; 1827 1828 c->freelist = get_freepointer(s, object); 1829 page->inuse = page->objects; 1830 page->freelist = NULL; 1831 1832unlock_out: 1833 slab_unlock(page); 1834#ifdef CONFIG_CMPXCHG_LOCAL 1835 c->tid = next_tid(c->tid); 1836 local_irq_restore(flags); 1837#endif 1838 stat(s, ALLOC_SLOWPATH); 1839 return object; 1840 1841another_slab: 1842 deactivate_slab(s, c); 1843 1844new_slab: 1845 page = get_partial(s, gfpflags, node); 1846 if (page) { 1847 stat(s, ALLOC_FROM_PARTIAL); 1848load_from_page: 1849 c->node = page_to_nid(page); 1850 c->page = page; 1851 goto load_freelist; 1852 } 1853 1854 gfpflags &= gfp_allowed_mask; 1855 if (gfpflags & __GFP_WAIT) 1856 local_irq_enable(); 1857 1858 page = new_slab(s, gfpflags, node); 1859 1860 if (gfpflags & __GFP_WAIT) 1861 local_irq_disable(); 1862 1863 if (page) { 1864 c = __this_cpu_ptr(s->cpu_slab); 1865 stat(s, ALLOC_SLAB); 1866 if (c->page) 1867 flush_slab(s, c); 1868 1869 slab_lock(page); 1870 __SetPageSlubFrozen(page); 1871 1872 goto load_from_page; 1873 } 1874 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 1875 slab_out_of_memory(s, gfpflags, node); 1876#ifdef CONFIG_CMPXCHG_LOCAL 1877 local_irq_restore(flags); 1878#endif 1879 return NULL; 1880debug: 1881 if (!alloc_debug_processing(s, page, object, addr)) 1882 goto another_slab; 1883 1884 page->inuse++; 1885 page->freelist = get_freepointer(s, object); 1886 c->node = NUMA_NO_NODE; 1887 goto unlock_out; 1888} 1889 1890/* 1891 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1892 * have the fastpath folded into their functions. So no function call 1893 * overhead for requests that can be satisfied on the fastpath. 1894 * 1895 * The fastpath works by first checking if the lockless freelist can be used. 1896 * If not then __slab_alloc is called for slow processing. 1897 * 1898 * Otherwise we can simply pick the next object from the lockless free list. 1899 */ 1900static __always_inline void *slab_alloc(struct kmem_cache *s, 1901 gfp_t gfpflags, int node, unsigned long addr) 1902{ 1903 void **object; 1904 struct kmem_cache_cpu *c; 1905#ifdef CONFIG_CMPXCHG_LOCAL 1906 unsigned long tid; 1907#else 1908 unsigned long flags; 1909#endif 1910 1911 if (slab_pre_alloc_hook(s, gfpflags)) 1912 return NULL; 1913 1914#ifndef CONFIG_CMPXCHG_LOCAL 1915 local_irq_save(flags); 1916#else 1917redo: 1918#endif 1919 1920 /* 1921 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 1922 * enabled. We may switch back and forth between cpus while 1923 * reading from one cpu area. That does not matter as long 1924 * as we end up on the original cpu again when doing the cmpxchg. 1925 */ 1926 c = __this_cpu_ptr(s->cpu_slab); 1927 1928#ifdef CONFIG_CMPXCHG_LOCAL 1929 /* 1930 * The transaction ids are globally unique per cpu and per operation on 1931 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 1932 * occurs on the right processor and that there was no operation on the 1933 * linked list in between. 1934 */ 1935 tid = c->tid; 1936 barrier(); 1937#endif 1938 1939 object = c->freelist; 1940 if (unlikely(!object || !node_match(c, node))) 1941 1942 object = __slab_alloc(s, gfpflags, node, addr, c); 1943 1944 else { 1945#ifdef CONFIG_CMPXCHG_LOCAL 1946 /* 1947 * The cmpxchg will only match if there was no additonal 1948 * operation and if we are on the right processor. 1949 * 1950 * The cmpxchg does the following atomically (without lock semantics!) 1951 * 1. Relocate first pointer to the current per cpu area. 1952 * 2. Verify that tid and freelist have not been changed 1953 * 3. If they were not changed replace tid and freelist 1954 * 1955 * Since this is without lock semantics the protection is only against 1956 * code executing on this cpu *not* from access by other cpus. 1957 */ 1958 if (unlikely(!this_cpu_cmpxchg_double( 1959 s->cpu_slab->freelist, s->cpu_slab->tid, 1960 object, tid, 1961 get_freepointer(s, object), next_tid(tid)))) { 1962 1963 note_cmpxchg_failure("slab_alloc", s, tid); 1964 goto redo; 1965 } 1966#else 1967 c->freelist = get_freepointer(s, object); 1968#endif 1969 stat(s, ALLOC_FASTPATH); 1970 } 1971 1972#ifndef CONFIG_CMPXCHG_LOCAL 1973 local_irq_restore(flags); 1974#endif 1975 1976 if (unlikely(gfpflags & __GFP_ZERO) && object) 1977 memset(object, 0, s->objsize); 1978 1979 slab_post_alloc_hook(s, gfpflags, object); 1980 1981 return object; 1982} 1983 1984void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1985{ 1986 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 1987 1988 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); 1989 1990 return ret; 1991} 1992EXPORT_SYMBOL(kmem_cache_alloc); 1993 1994#ifdef CONFIG_TRACING 1995void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 1996{ 1997 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 1998 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 1999 return ret; 2000} 2001EXPORT_SYMBOL(kmem_cache_alloc_trace); 2002 2003void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) 2004{ 2005 void *ret = kmalloc_order(size, flags, order); 2006 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); 2007 return ret; 2008} 2009EXPORT_SYMBOL(kmalloc_order_trace); 2010#endif 2011 2012#ifdef CONFIG_NUMA 2013void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2014{ 2015 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 2016 2017 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2018 s->objsize, s->size, gfpflags, node); 2019 2020 return ret; 2021} 2022EXPORT_SYMBOL(kmem_cache_alloc_node); 2023 2024#ifdef CONFIG_TRACING 2025void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2026 gfp_t gfpflags, 2027 int node, size_t size) 2028{ 2029 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 2030 2031 trace_kmalloc_node(_RET_IP_, ret, 2032 size, s->size, gfpflags, node); 2033 return ret; 2034} 2035EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2036#endif 2037#endif 2038 2039/* 2040 * Slow patch handling. This may still be called frequently since objects 2041 * have a longer lifetime than the cpu slabs in most processing loads. 2042 * 2043 * So we still attempt to reduce cache line usage. Just take the slab 2044 * lock and free the item. If there is no additional partial page 2045 * handling required then we can return immediately. 2046 */ 2047static void __slab_free(struct kmem_cache *s, struct page *page, 2048 void *x, unsigned long addr) 2049{ 2050 void *prior; 2051 void **object = (void *)x; 2052#ifdef CONFIG_CMPXCHG_LOCAL 2053 unsigned long flags; 2054 2055 local_irq_save(flags); 2056#endif 2057 slab_lock(page); 2058 stat(s, FREE_SLOWPATH); 2059 2060 if (kmem_cache_debug(s)) 2061 goto debug; 2062 2063checks_ok: 2064 prior = page->freelist; 2065 set_freepointer(s, object, prior); 2066 page->freelist = object; 2067 page->inuse--; 2068 2069 if (unlikely(PageSlubFrozen(page))) { 2070 stat(s, FREE_FROZEN); 2071 goto out_unlock; 2072 } 2073 2074 if (unlikely(!page->inuse)) 2075 goto slab_empty; 2076 2077 /* 2078 * Objects left in the slab. If it was not on the partial list before 2079 * then add it. 2080 */ 2081 if (unlikely(!prior)) { 2082 add_partial(get_node(s, page_to_nid(page)), page, 1); 2083 stat(s, FREE_ADD_PARTIAL); 2084 } 2085 2086out_unlock: 2087 slab_unlock(page); 2088#ifdef CONFIG_CMPXCHG_LOCAL 2089 local_irq_restore(flags); 2090#endif 2091 return; 2092 2093slab_empty: 2094 if (prior) { 2095 /* 2096 * Slab still on the partial list. 2097 */ 2098 remove_partial(s, page); 2099 stat(s, FREE_REMOVE_PARTIAL); 2100 } 2101 slab_unlock(page); 2102#ifdef CONFIG_CMPXCHG_LOCAL 2103 local_irq_restore(flags); 2104#endif 2105 stat(s, FREE_SLAB); 2106 discard_slab(s, page); 2107 return; 2108 2109debug: 2110 if (!free_debug_processing(s, page, x, addr)) 2111 goto out_unlock; 2112 goto checks_ok; 2113} 2114 2115/* 2116 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 2117 * can perform fastpath freeing without additional function calls. 2118 * 2119 * The fastpath is only possible if we are freeing to the current cpu slab 2120 * of this processor. This typically the case if we have just allocated 2121 * the item before. 2122 * 2123 * If fastpath is not possible then fall back to __slab_free where we deal 2124 * with all sorts of special processing. 2125 */ 2126static __always_inline void slab_free(struct kmem_cache *s, 2127 struct page *page, void *x, unsigned long addr) 2128{ 2129 void **object = (void *)x; 2130 struct kmem_cache_cpu *c; 2131#ifdef CONFIG_CMPXCHG_LOCAL 2132 unsigned long tid; 2133#else 2134 unsigned long flags; 2135#endif 2136 2137 slab_free_hook(s, x); 2138 2139#ifndef CONFIG_CMPXCHG_LOCAL 2140 local_irq_save(flags); 2141 2142#else 2143redo: 2144#endif 2145 2146 /* 2147 * Determine the currently cpus per cpu slab. 2148 * The cpu may change afterward. However that does not matter since 2149 * data is retrieved via this pointer. If we are on the same cpu 2150 * during the cmpxchg then the free will succedd. 2151 */ 2152 c = __this_cpu_ptr(s->cpu_slab); 2153 2154#ifdef CONFIG_CMPXCHG_LOCAL 2155 tid = c->tid; 2156 barrier(); 2157#endif 2158 2159 if (likely(page == c->page && c->node != NUMA_NO_NODE)) { 2160 set_freepointer(s, object, c->freelist); 2161 2162#ifdef CONFIG_CMPXCHG_LOCAL 2163 if (unlikely(!this_cpu_cmpxchg_double( 2164 s->cpu_slab->freelist, s->cpu_slab->tid, 2165 c->freelist, tid, 2166 object, next_tid(tid)))) { 2167 2168 note_cmpxchg_failure("slab_free", s, tid); 2169 goto redo; 2170 } 2171#else 2172 c->freelist = object; 2173#endif 2174 stat(s, FREE_FASTPATH); 2175 } else 2176 __slab_free(s, page, x, addr); 2177 2178#ifndef CONFIG_CMPXCHG_LOCAL 2179 local_irq_restore(flags); 2180#endif 2181} 2182 2183void kmem_cache_free(struct kmem_cache *s, void *x) 2184{ 2185 struct page *page; 2186 2187 page = virt_to_head_page(x); 2188 2189 slab_free(s, page, x, _RET_IP_); 2190 2191 trace_kmem_cache_free(_RET_IP_, x); 2192} 2193EXPORT_SYMBOL(kmem_cache_free); 2194 2195/* 2196 * Object placement in a slab is made very easy because we always start at 2197 * offset 0. If we tune the size of the object to the alignment then we can 2198 * get the required alignment by putting one properly sized object after 2199 * another. 2200 * 2201 * Notice that the allocation order determines the sizes of the per cpu 2202 * caches. Each processor has always one slab available for allocations. 2203 * Increasing the allocation order reduces the number of times that slabs 2204 * must be moved on and off the partial lists and is therefore a factor in 2205 * locking overhead. 2206 */ 2207 2208/* 2209 * Mininum / Maximum order of slab pages. This influences locking overhead 2210 * and slab fragmentation. A higher order reduces the number of partial slabs 2211 * and increases the number of allocations possible without having to 2212 * take the list_lock. 2213 */ 2214static int slub_min_order; 2215static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 2216static int slub_min_objects; 2217 2218/* 2219 * Merge control. If this is set then no merging of slab caches will occur. 2220 * (Could be removed. This was introduced to pacify the merge skeptics.) 2221 */ 2222static int slub_nomerge; 2223 2224/* 2225 * Calculate the order of allocation given an slab object size. 2226 * 2227 * The order of allocation has significant impact on performance and other 2228 * system components. Generally order 0 allocations should be preferred since 2229 * order 0 does not cause fragmentation in the page allocator. Larger objects 2230 * be problematic to put into order 0 slabs because there may be too much 2231 * unused space left. We go to a higher order if more than 1/16th of the slab 2232 * would be wasted. 2233 * 2234 * In order to reach satisfactory performance we must ensure that a minimum 2235 * number of objects is in one slab. Otherwise we may generate too much 2236 * activity on the partial lists which requires taking the list_lock. This is 2237 * less a concern for large slabs though which are rarely used. 2238 * 2239 * slub_max_order specifies the order where we begin to stop considering the 2240 * number of objects in a slab as critical. If we reach slub_max_order then 2241 * we try to keep the page order as low as possible. So we accept more waste 2242 * of space in favor of a small page order. 2243 * 2244 * Higher order allocations also allow the placement of more objects in a 2245 * slab and thereby reduce object handling overhead. If the user has 2246 * requested a higher mininum order then we start with that one instead of 2247 * the smallest order which will fit the object. 2248 */ 2249static inline int slab_order(int size, int min_objects, 2250 int max_order, int fract_leftover, int reserved) 2251{ 2252 int order; 2253 int rem; 2254 int min_order = slub_min_order; 2255 2256 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) 2257 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 2258 2259 for (order = max(min_order, 2260 fls(min_objects * size - 1) - PAGE_SHIFT); 2261 order <= max_order; order++) { 2262 2263 unsigned long slab_size = PAGE_SIZE << order; 2264 2265 if (slab_size < min_objects * size + reserved) 2266 continue; 2267 2268 rem = (slab_size - reserved) % size; 2269 2270 if (rem <= slab_size / fract_leftover) 2271 break; 2272 2273 } 2274 2275 return order; 2276} 2277 2278static inline int calculate_order(int size, int reserved) 2279{ 2280 int order; 2281 int min_objects; 2282 int fraction; 2283 int max_objects; 2284 2285 /* 2286 * Attempt to find best configuration for a slab. This 2287 * works by first attempting to generate a layout with 2288 * the best configuration and backing off gradually. 2289 * 2290 * First we reduce the acceptable waste in a slab. Then 2291 * we reduce the minimum objects required in a slab. 2292 */ 2293 min_objects = slub_min_objects; 2294 if (!min_objects) 2295 min_objects = 4 * (fls(nr_cpu_ids) + 1); 2296 max_objects = order_objects(slub_max_order, size, reserved); 2297 min_objects = min(min_objects, max_objects); 2298 2299 while (min_objects > 1) { 2300 fraction = 16; 2301 while (fraction >= 4) { 2302 order = slab_order(size, min_objects, 2303 slub_max_order, fraction, reserved); 2304 if (order <= slub_max_order) 2305 return order; 2306 fraction /= 2; 2307 } 2308 min_objects--; 2309 } 2310 2311 /* 2312 * We were unable to place multiple objects in a slab. Now 2313 * lets see if we can place a single object there. 2314 */ 2315 order = slab_order(size, 1, slub_max_order, 1, reserved); 2316 if (order <= slub_max_order) 2317 return order; 2318 2319 /* 2320 * Doh this slab cannot be placed using slub_max_order. 2321 */ 2322 order = slab_order(size, 1, MAX_ORDER, 1, reserved); 2323 if (order < MAX_ORDER) 2324 return order; 2325 return -ENOSYS; 2326} 2327 2328/* 2329 * Figure out what the alignment of the objects will be. 2330 */ 2331static unsigned long calculate_alignment(unsigned long flags, 2332 unsigned long align, unsigned long size) 2333{ 2334 /* 2335 * If the user wants hardware cache aligned objects then follow that 2336 * suggestion if the object is sufficiently large. 2337 * 2338 * The hardware cache alignment cannot override the specified 2339 * alignment though. If that is greater then use it. 2340 */ 2341 if (flags & SLAB_HWCACHE_ALIGN) { 2342 unsigned long ralign = cache_line_size(); 2343 while (size <= ralign / 2) 2344 ralign /= 2; 2345 align = max(align, ralign); 2346 } 2347 2348 if (align < ARCH_SLAB_MINALIGN) 2349 align = ARCH_SLAB_MINALIGN; 2350 2351 return ALIGN(align, sizeof(void *)); 2352} 2353 2354static void 2355init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 2356{ 2357 n->nr_partial = 0; 2358 spin_lock_init(&n->list_lock); 2359 INIT_LIST_HEAD(&n->partial); 2360#ifdef CONFIG_SLUB_DEBUG 2361 atomic_long_set(&n->nr_slabs, 0); 2362 atomic_long_set(&n->total_objects, 0); 2363 INIT_LIST_HEAD(&n->full); 2364#endif 2365} 2366 2367static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 2368{ 2369 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 2370 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); 2371 2372#ifdef CONFIG_CMPXCHG_LOCAL 2373 /* 2374 * Must align to double word boundary for the double cmpxchg instructions 2375 * to work. 2376 */ 2377 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *)); 2378#else 2379 /* Regular alignment is sufficient */ 2380 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu); 2381#endif 2382 2383 if (!s->cpu_slab) 2384 return 0; 2385 2386 init_kmem_cache_cpus(s); 2387 2388 return 1; 2389} 2390 2391static struct kmem_cache *kmem_cache_node; 2392 2393/* 2394 * No kmalloc_node yet so do it by hand. We know that this is the first 2395 * slab on the node for this slabcache. There are no concurrent accesses 2396 * possible. 2397 * 2398 * Note that this function only works on the kmalloc_node_cache 2399 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2400 * memory on a fresh node that has no slab structures yet. 2401 */ 2402static void early_kmem_cache_node_alloc(int node) 2403{ 2404 struct page *page; 2405 struct kmem_cache_node *n; 2406 unsigned long flags; 2407 2408 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 2409 2410 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 2411 2412 BUG_ON(!page); 2413 if (page_to_nid(page) != node) { 2414 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2415 "node %d\n", node); 2416 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2417 "in order to be able to continue\n"); 2418 } 2419 2420 n = page->freelist; 2421 BUG_ON(!n); 2422 page->freelist = get_freepointer(kmem_cache_node, n); 2423 page->inuse++; 2424 kmem_cache_node->node[node] = n; 2425#ifdef CONFIG_SLUB_DEBUG 2426 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 2427 init_tracking(kmem_cache_node, n); 2428#endif 2429 init_kmem_cache_node(n, kmem_cache_node); 2430 inc_slabs_node(kmem_cache_node, node, page->objects); 2431 2432 /* 2433 * lockdep requires consistent irq usage for each lock 2434 * so even though there cannot be a race this early in 2435 * the boot sequence, we still disable irqs. 2436 */ 2437 local_irq_save(flags); 2438 add_partial(n, page, 0); 2439 local_irq_restore(flags); 2440} 2441 2442static void free_kmem_cache_nodes(struct kmem_cache *s) 2443{ 2444 int node; 2445 2446 for_each_node_state(node, N_NORMAL_MEMORY) { 2447 struct kmem_cache_node *n = s->node[node]; 2448 2449 if (n) 2450 kmem_cache_free(kmem_cache_node, n); 2451 2452 s->node[node] = NULL; 2453 } 2454} 2455 2456static int init_kmem_cache_nodes(struct kmem_cache *s) 2457{ 2458 int node; 2459 2460 for_each_node_state(node, N_NORMAL_MEMORY) { 2461 struct kmem_cache_node *n; 2462 2463 if (slab_state == DOWN) { 2464 early_kmem_cache_node_alloc(node); 2465 continue; 2466 } 2467 n = kmem_cache_alloc_node(kmem_cache_node, 2468 GFP_KERNEL, node); 2469 2470 if (!n) { 2471 free_kmem_cache_nodes(s); 2472 return 0; 2473 } 2474 2475 s->node[node] = n; 2476 init_kmem_cache_node(n, s); 2477 } 2478 return 1; 2479} 2480 2481static void set_min_partial(struct kmem_cache *s, unsigned long min) 2482{ 2483 if (min < MIN_PARTIAL) 2484 min = MIN_PARTIAL; 2485 else if (min > MAX_PARTIAL) 2486 min = MAX_PARTIAL; 2487 s->min_partial = min; 2488} 2489 2490/* 2491 * calculate_sizes() determines the order and the distribution of data within 2492 * a slab object. 2493 */ 2494static int calculate_sizes(struct kmem_cache *s, int forced_order) 2495{ 2496 unsigned long flags = s->flags; 2497 unsigned long size = s->objsize; 2498 unsigned long align = s->align; 2499 int order; 2500 2501 /* 2502 * Round up object size to the next word boundary. We can only 2503 * place the free pointer at word boundaries and this determines 2504 * the possible location of the free pointer. 2505 */ 2506 size = ALIGN(size, sizeof(void *)); 2507 2508#ifdef CONFIG_SLUB_DEBUG 2509 /* 2510 * Determine if we can poison the object itself. If the user of 2511 * the slab may touch the object after free or before allocation 2512 * then we should never poison the object itself. 2513 */ 2514 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2515 !s->ctor) 2516 s->flags |= __OBJECT_POISON; 2517 else 2518 s->flags &= ~__OBJECT_POISON; 2519 2520 2521 /* 2522 * If we are Redzoning then check if there is some space between the 2523 * end of the object and the free pointer. If not then add an 2524 * additional word to have some bytes to store Redzone information. 2525 */ 2526 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2527 size += sizeof(void *); 2528#endif 2529 2530 /* 2531 * With that we have determined the number of bytes in actual use 2532 * by the object. This is the potential offset to the free pointer. 2533 */ 2534 s->inuse = size; 2535 2536 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2537 s->ctor)) { 2538 /* 2539 * Relocate free pointer after the object if it is not 2540 * permitted to overwrite the first word of the object on 2541 * kmem_cache_free. 2542 * 2543 * This is the case if we do RCU, have a constructor or 2544 * destructor or are poisoning the objects. 2545 */ 2546 s->offset = size; 2547 size += sizeof(void *); 2548 } 2549 2550#ifdef CONFIG_SLUB_DEBUG 2551 if (flags & SLAB_STORE_USER) 2552 /* 2553 * Need to store information about allocs and frees after 2554 * the object. 2555 */ 2556 size += 2 * sizeof(struct track); 2557 2558 if (flags & SLAB_RED_ZONE) 2559 /* 2560 * Add some empty padding so that we can catch 2561 * overwrites from earlier objects rather than let 2562 * tracking information or the free pointer be 2563 * corrupted if a user writes before the start 2564 * of the object. 2565 */ 2566 size += sizeof(void *); 2567#endif 2568 2569 /* 2570 * Determine the alignment based on various parameters that the 2571 * user specified and the dynamic determination of cache line size 2572 * on bootup. 2573 */ 2574 align = calculate_alignment(flags, align, s->objsize); 2575 s->align = align; 2576 2577 /* 2578 * SLUB stores one object immediately after another beginning from 2579 * offset 0. In order to align the objects we have to simply size 2580 * each object to conform to the alignment. 2581 */ 2582 size = ALIGN(size, align); 2583 s->size = size; 2584 if (forced_order >= 0) 2585 order = forced_order; 2586 else 2587 order = calculate_order(size, s->reserved); 2588 2589 if (order < 0) 2590 return 0; 2591 2592 s->allocflags = 0; 2593 if (order) 2594 s->allocflags |= __GFP_COMP; 2595 2596 if (s->flags & SLAB_CACHE_DMA) 2597 s->allocflags |= SLUB_DMA; 2598 2599 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2600 s->allocflags |= __GFP_RECLAIMABLE; 2601 2602 /* 2603 * Determine the number of objects per slab 2604 */ 2605 s->oo = oo_make(order, size, s->reserved); 2606 s->min = oo_make(get_order(size), size, s->reserved); 2607 if (oo_objects(s->oo) > oo_objects(s->max)) 2608 s->max = s->oo; 2609 2610 return !!oo_objects(s->oo); 2611 2612} 2613 2614static int kmem_cache_open(struct kmem_cache *s, 2615 const char *name, size_t size, 2616 size_t align, unsigned long flags, 2617 void (*ctor)(void *)) 2618{ 2619 memset(s, 0, kmem_size); 2620 s->name = name; 2621 s->ctor = ctor; 2622 s->objsize = size; 2623 s->align = align; 2624 s->flags = kmem_cache_flags(size, flags, name, ctor); 2625 s->reserved = 0; 2626 2627 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) 2628 s->reserved = sizeof(struct rcu_head); 2629 2630 if (!calculate_sizes(s, -1)) 2631 goto error; 2632 if (disable_higher_order_debug) { 2633 /* 2634 * Disable debugging flags that store metadata if the min slab 2635 * order increased. 2636 */ 2637 if (get_order(s->size) > get_order(s->objsize)) { 2638 s->flags &= ~DEBUG_METADATA_FLAGS; 2639 s->offset = 0; 2640 if (!calculate_sizes(s, -1)) 2641 goto error; 2642 } 2643 } 2644 2645 /* 2646 * The larger the object size is, the more pages we want on the partial 2647 * list to avoid pounding the page allocator excessively. 2648 */ 2649 set_min_partial(s, ilog2(s->size)); 2650 s->refcount = 1; 2651#ifdef CONFIG_NUMA 2652 s->remote_node_defrag_ratio = 1000; 2653#endif 2654 if (!init_kmem_cache_nodes(s)) 2655 goto error; 2656 2657 if (alloc_kmem_cache_cpus(s)) 2658 return 1; 2659 2660 free_kmem_cache_nodes(s); 2661error: 2662 if (flags & SLAB_PANIC) 2663 panic("Cannot create slab %s size=%lu realsize=%u " 2664 "order=%u offset=%u flags=%lx\n", 2665 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2666 s->offset, flags); 2667 return 0; 2668} 2669 2670/* 2671 * Determine the size of a slab object 2672 */ 2673unsigned int kmem_cache_size(struct kmem_cache *s) 2674{ 2675 return s->objsize; 2676} 2677EXPORT_SYMBOL(kmem_cache_size); 2678 2679static void list_slab_objects(struct kmem_cache *s, struct page *page, 2680 const char *text) 2681{ 2682#ifdef CONFIG_SLUB_DEBUG 2683 void *addr = page_address(page); 2684 void *p; 2685 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * 2686 sizeof(long), GFP_ATOMIC); 2687 if (!map) 2688 return; 2689 slab_err(s, page, "%s", text); 2690 slab_lock(page); 2691 2692 get_map(s, page, map); 2693 for_each_object(p, s, addr, page->objects) { 2694 2695 if (!test_bit(slab_index(p, s, addr), map)) { 2696 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2697 p, p - addr); 2698 print_tracking(s, p); 2699 } 2700 } 2701 slab_unlock(page); 2702 kfree(map); 2703#endif 2704} 2705 2706/* 2707 * Attempt to free all partial slabs on a node. 2708 */ 2709static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2710{ 2711 unsigned long flags; 2712 struct page *page, *h; 2713 2714 spin_lock_irqsave(&n->list_lock, flags); 2715 list_for_each_entry_safe(page, h, &n->partial, lru) { 2716 if (!page->inuse) { 2717 __remove_partial(n, page); 2718 discard_slab(s, page); 2719 } else { 2720 list_slab_objects(s, page, 2721 "Objects remaining on kmem_cache_close()"); 2722 } 2723 } 2724 spin_unlock_irqrestore(&n->list_lock, flags); 2725} 2726 2727/* 2728 * Release all resources used by a slab cache. 2729 */ 2730static inline int kmem_cache_close(struct kmem_cache *s) 2731{ 2732 int node; 2733 2734 flush_all(s); 2735 free_percpu(s->cpu_slab); 2736 /* Attempt to free all objects */ 2737 for_each_node_state(node, N_NORMAL_MEMORY) { 2738 struct kmem_cache_node *n = get_node(s, node); 2739 2740 free_partial(s, n); 2741 if (n->nr_partial || slabs_node(s, node)) 2742 return 1; 2743 } 2744 free_kmem_cache_nodes(s); 2745 return 0; 2746} 2747 2748/* 2749 * Close a cache and release the kmem_cache structure 2750 * (must be used for caches created using kmem_cache_create) 2751 */ 2752void kmem_cache_destroy(struct kmem_cache *s) 2753{ 2754 down_write(&slub_lock); 2755 s->refcount--; 2756 if (!s->refcount) { 2757 list_del(&s->list); 2758 if (kmem_cache_close(s)) { 2759 printk(KERN_ERR "SLUB %s: %s called for cache that " 2760 "still has objects.\n", s->name, __func__); 2761 dump_stack(); 2762 } 2763 if (s->flags & SLAB_DESTROY_BY_RCU) 2764 rcu_barrier(); 2765 sysfs_slab_remove(s); 2766 } 2767 up_write(&slub_lock); 2768} 2769EXPORT_SYMBOL(kmem_cache_destroy); 2770 2771/******************************************************************** 2772 * Kmalloc subsystem 2773 *******************************************************************/ 2774 2775struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; 2776EXPORT_SYMBOL(kmalloc_caches); 2777 2778static struct kmem_cache *kmem_cache; 2779 2780#ifdef CONFIG_ZONE_DMA 2781static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; 2782#endif 2783 2784static int __init setup_slub_min_order(char *str) 2785{ 2786 get_option(&str, &slub_min_order); 2787 2788 return 1; 2789} 2790 2791__setup("slub_min_order=", setup_slub_min_order); 2792 2793static int __init setup_slub_max_order(char *str) 2794{ 2795 get_option(&str, &slub_max_order); 2796 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 2797 2798 return 1; 2799} 2800 2801__setup("slub_max_order=", setup_slub_max_order); 2802 2803static int __init setup_slub_min_objects(char *str) 2804{ 2805 get_option(&str, &slub_min_objects); 2806 2807 return 1; 2808} 2809 2810__setup("slub_min_objects=", setup_slub_min_objects); 2811 2812static int __init setup_slub_nomerge(char *str) 2813{ 2814 slub_nomerge = 1; 2815 return 1; 2816} 2817 2818__setup("slub_nomerge", setup_slub_nomerge); 2819 2820static struct kmem_cache *__init create_kmalloc_cache(const char *name, 2821 int size, unsigned int flags) 2822{ 2823 struct kmem_cache *s; 2824 2825 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 2826 2827 /* 2828 * This function is called with IRQs disabled during early-boot on 2829 * single CPU so there's no need to take slub_lock here. 2830 */ 2831 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, 2832 flags, NULL)) 2833 goto panic; 2834 2835 list_add(&s->list, &slab_caches); 2836 return s; 2837 2838panic: 2839 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2840 return NULL; 2841} 2842 2843/* 2844 * Conversion table for small slabs sizes / 8 to the index in the 2845 * kmalloc array. This is necessary for slabs < 192 since we have non power 2846 * of two cache sizes there. The size of larger slabs can be determined using 2847 * fls. 2848 */ 2849static s8 size_index[24] = { 2850 3, /* 8 */ 2851 4, /* 16 */ 2852 5, /* 24 */ 2853 5, /* 32 */ 2854 6, /* 40 */ 2855 6, /* 48 */ 2856 6, /* 56 */ 2857 6, /* 64 */ 2858 1, /* 72 */ 2859 1, /* 80 */ 2860 1, /* 88 */ 2861 1, /* 96 */ 2862 7, /* 104 */ 2863 7, /* 112 */ 2864 7, /* 120 */ 2865 7, /* 128 */ 2866 2, /* 136 */ 2867 2, /* 144 */ 2868 2, /* 152 */ 2869 2, /* 160 */ 2870 2, /* 168 */ 2871 2, /* 176 */ 2872 2, /* 184 */ 2873 2 /* 192 */ 2874}; 2875 2876static inline int size_index_elem(size_t bytes) 2877{ 2878 return (bytes - 1) / 8; 2879} 2880 2881static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2882{ 2883 int index; 2884 2885 if (size <= 192) { 2886 if (!size) 2887 return ZERO_SIZE_PTR; 2888 2889 index = size_index[size_index_elem(size)]; 2890 } else 2891 index = fls(size - 1); 2892 2893#ifdef CONFIG_ZONE_DMA 2894 if (unlikely((flags & SLUB_DMA))) 2895 return kmalloc_dma_caches[index]; 2896 2897#endif 2898 return kmalloc_caches[index]; 2899} 2900 2901void *__kmalloc(size_t size, gfp_t flags) 2902{ 2903 struct kmem_cache *s; 2904 void *ret; 2905 2906 if (unlikely(size > SLUB_MAX_SIZE)) 2907 return kmalloc_large(size, flags); 2908 2909 s = get_slab(size, flags); 2910 2911 if (unlikely(ZERO_OR_NULL_PTR(s))) 2912 return s; 2913 2914 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); 2915 2916 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 2917 2918 return ret; 2919} 2920EXPORT_SYMBOL(__kmalloc); 2921 2922#ifdef CONFIG_NUMA 2923static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2924{ 2925 struct page *page; 2926 void *ptr = NULL; 2927 2928 flags |= __GFP_COMP | __GFP_NOTRACK; 2929 page = alloc_pages_node(node, flags, get_order(size)); 2930 if (page) 2931 ptr = page_address(page); 2932 2933 kmemleak_alloc(ptr, size, 1, flags); 2934 return ptr; 2935} 2936 2937void *__kmalloc_node(size_t size, gfp_t flags, int node) 2938{ 2939 struct kmem_cache *s; 2940 void *ret; 2941 2942 if (unlikely(size > SLUB_MAX_SIZE)) { 2943 ret = kmalloc_large_node(size, flags, node); 2944 2945 trace_kmalloc_node(_RET_IP_, ret, 2946 size, PAGE_SIZE << get_order(size), 2947 flags, node); 2948 2949 return ret; 2950 } 2951 2952 s = get_slab(size, flags); 2953 2954 if (unlikely(ZERO_OR_NULL_PTR(s))) 2955 return s; 2956 2957 ret = slab_alloc(s, flags, node, _RET_IP_); 2958 2959 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 2960 2961 return ret; 2962} 2963EXPORT_SYMBOL(__kmalloc_node); 2964#endif 2965 2966size_t ksize(const void *object) 2967{ 2968 struct page *page; 2969 2970 if (unlikely(object == ZERO_SIZE_PTR)) 2971 return 0; 2972 2973 page = virt_to_head_page(object); 2974 2975 if (unlikely(!PageSlab(page))) { 2976 WARN_ON(!PageCompound(page)); 2977 return PAGE_SIZE << compound_order(page); 2978 } 2979 2980 return slab_ksize(page->slab); 2981} 2982EXPORT_SYMBOL(ksize); 2983 2984void kfree(const void *x) 2985{ 2986 struct page *page; 2987 void *object = (void *)x; 2988 2989 trace_kfree(_RET_IP_, x); 2990 2991 if (unlikely(ZERO_OR_NULL_PTR(x))) 2992 return; 2993 2994 page = virt_to_head_page(x); 2995 if (unlikely(!PageSlab(page))) { 2996 BUG_ON(!PageCompound(page)); 2997 kmemleak_free(x); 2998 put_page(page); 2999 return; 3000 } 3001 slab_free(page->slab, page, object, _RET_IP_); 3002} 3003EXPORT_SYMBOL(kfree); 3004 3005/* 3006 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 3007 * the remaining slabs by the number of items in use. The slabs with the 3008 * most items in use come first. New allocations will then fill those up 3009 * and thus they can be removed from the partial lists. 3010 * 3011 * The slabs with the least items are placed last. This results in them 3012 * being allocated from last increasing the chance that the last objects 3013 * are freed in them. 3014 */ 3015int kmem_cache_shrink(struct kmem_cache *s) 3016{ 3017 int node; 3018 int i; 3019 struct kmem_cache_node *n; 3020 struct page *page; 3021 struct page *t; 3022 int objects = oo_objects(s->max); 3023 struct list_head *slabs_by_inuse = 3024 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 3025 unsigned long flags; 3026 3027 if (!slabs_by_inuse) 3028 return -ENOMEM; 3029 3030 flush_all(s); 3031 for_each_node_state(node, N_NORMAL_MEMORY) { 3032 n = get_node(s, node); 3033 3034 if (!n->nr_partial) 3035 continue; 3036 3037 for (i = 0; i < objects; i++) 3038 INIT_LIST_HEAD(slabs_by_inuse + i); 3039 3040 spin_lock_irqsave(&n->list_lock, flags); 3041 3042 /* 3043 * Build lists indexed by the items in use in each slab. 3044 * 3045 * Note that concurrent frees may occur while we hold the 3046 * list_lock. page->inuse here is the upper limit. 3047 */ 3048 list_for_each_entry_safe(page, t, &n->partial, lru) { 3049 if (!page->inuse && slab_trylock(page)) { 3050 /* 3051 * Must hold slab lock here because slab_free 3052 * may have freed the last object and be 3053 * waiting to release the slab. 3054 */ 3055 __remove_partial(n, page); 3056 slab_unlock(page); 3057 discard_slab(s, page); 3058 } else { 3059 list_move(&page->lru, 3060 slabs_by_inuse + page->inuse); 3061 } 3062 } 3063 3064 /* 3065 * Rebuild the partial list with the slabs filled up most 3066 * first and the least used slabs at the end. 3067 */ 3068 for (i = objects - 1; i >= 0; i--) 3069 list_splice(slabs_by_inuse + i, n->partial.prev); 3070 3071 spin_unlock_irqrestore(&n->list_lock, flags); 3072 } 3073 3074 kfree(slabs_by_inuse); 3075 return 0; 3076} 3077EXPORT_SYMBOL(kmem_cache_shrink); 3078 3079#if defined(CONFIG_MEMORY_HOTPLUG) 3080static int slab_mem_going_offline_callback(void *arg) 3081{ 3082 struct kmem_cache *s; 3083 3084 down_read(&slub_lock); 3085 list_for_each_entry(s, &slab_caches, list) 3086 kmem_cache_shrink(s); 3087 up_read(&slub_lock); 3088 3089 return 0; 3090} 3091 3092static void slab_mem_offline_callback(void *arg) 3093{ 3094 struct kmem_cache_node *n; 3095 struct kmem_cache *s; 3096 struct memory_notify *marg = arg; 3097 int offline_node; 3098 3099 offline_node = marg->status_change_nid; 3100 3101 /* 3102 * If the node still has available memory. we need kmem_cache_node 3103 * for it yet. 3104 */ 3105 if (offline_node < 0) 3106 return; 3107 3108 down_read(&slub_lock); 3109 list_for_each_entry(s, &slab_caches, list) { 3110 n = get_node(s, offline_node); 3111 if (n) { 3112 /* 3113 * if n->nr_slabs > 0, slabs still exist on the node 3114 * that is going down. We were unable to free them, 3115 * and offline_pages() function shouldn't call this 3116 * callback. So, we must fail. 3117 */ 3118 BUG_ON(slabs_node(s, offline_node)); 3119 3120 s->node[offline_node] = NULL; 3121 kmem_cache_free(kmem_cache_node, n); 3122 } 3123 } 3124 up_read(&slub_lock); 3125} 3126 3127static int slab_mem_going_online_callback(void *arg) 3128{ 3129 struct kmem_cache_node *n; 3130 struct kmem_cache *s; 3131 struct memory_notify *marg = arg; 3132 int nid = marg->status_change_nid; 3133 int ret = 0; 3134 3135 /* 3136 * If the node's memory is already available, then kmem_cache_node is 3137 * already created. Nothing to do. 3138 */ 3139 if (nid < 0) 3140 return 0; 3141 3142 /* 3143 * We are bringing a node online. No memory is available yet. We must 3144 * allocate a kmem_cache_node structure in order to bring the node 3145 * online. 3146 */ 3147 down_read(&slub_lock); 3148 list_for_each_entry(s, &slab_caches, list) { 3149 /* 3150 * XXX: kmem_cache_alloc_node will fallback to other nodes 3151 * since memory is not yet available from the node that 3152 * is brought up. 3153 */ 3154 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 3155 if (!n) { 3156 ret = -ENOMEM; 3157 goto out; 3158 } 3159 init_kmem_cache_node(n, s); 3160 s->node[nid] = n; 3161 } 3162out: 3163 up_read(&slub_lock); 3164 return ret; 3165} 3166 3167static int slab_memory_callback(struct notifier_block *self, 3168 unsigned long action, void *arg) 3169{ 3170 int ret = 0; 3171 3172 switch (action) { 3173 case MEM_GOING_ONLINE: 3174 ret = slab_mem_going_online_callback(arg); 3175 break; 3176 case MEM_GOING_OFFLINE: 3177 ret = slab_mem_going_offline_callback(arg); 3178 break; 3179 case MEM_OFFLINE: 3180 case MEM_CANCEL_ONLINE: 3181 slab_mem_offline_callback(arg); 3182 break; 3183 case MEM_ONLINE: 3184 case MEM_CANCEL_OFFLINE: 3185 break; 3186 } 3187 if (ret) 3188 ret = notifier_from_errno(ret); 3189 else 3190 ret = NOTIFY_OK; 3191 return ret; 3192} 3193 3194#endif /* CONFIG_MEMORY_HOTPLUG */ 3195 3196/******************************************************************** 3197 * Basic setup of slabs 3198 *******************************************************************/ 3199 3200/* 3201 * Used for early kmem_cache structures that were allocated using 3202 * the page allocator 3203 */ 3204 3205static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) 3206{ 3207 int node; 3208 3209 list_add(&s->list, &slab_caches); 3210 s->refcount = -1; 3211 3212 for_each_node_state(node, N_NORMAL_MEMORY) { 3213 struct kmem_cache_node *n = get_node(s, node); 3214 struct page *p; 3215 3216 if (n) { 3217 list_for_each_entry(p, &n->partial, lru) 3218 p->slab = s; 3219 3220#ifdef CONFIG_SLUB_DEBUG 3221 list_for_each_entry(p, &n->full, lru) 3222 p->slab = s; 3223#endif 3224 } 3225 } 3226} 3227 3228void __init kmem_cache_init(void) 3229{ 3230 int i; 3231 int caches = 0; 3232 struct kmem_cache *temp_kmem_cache; 3233 int order; 3234 struct kmem_cache *temp_kmem_cache_node; 3235 unsigned long kmalloc_size; 3236 3237 kmem_size = offsetof(struct kmem_cache, node) + 3238 nr_node_ids * sizeof(struct kmem_cache_node *); 3239 3240 /* Allocate two kmem_caches from the page allocator */ 3241 kmalloc_size = ALIGN(kmem_size, cache_line_size()); 3242 order = get_order(2 * kmalloc_size); 3243 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); 3244 3245 /* 3246 * Must first have the slab cache available for the allocations of the 3247 * struct kmem_cache_node's. There is special bootstrap code in 3248 * kmem_cache_open for slab_state == DOWN. 3249 */ 3250 kmem_cache_node = (void *)kmem_cache + kmalloc_size; 3251 3252 kmem_cache_open(kmem_cache_node, "kmem_cache_node", 3253 sizeof(struct kmem_cache_node), 3254 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3255 3256 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 3257 3258 /* Able to allocate the per node structures */ 3259 slab_state = PARTIAL; 3260 3261 temp_kmem_cache = kmem_cache; 3262 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, 3263 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3264 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3265 memcpy(kmem_cache, temp_kmem_cache, kmem_size); 3266 3267 /* 3268 * Allocate kmem_cache_node properly from the kmem_cache slab. 3269 * kmem_cache_node is separately allocated so no need to 3270 * update any list pointers. 3271 */ 3272 temp_kmem_cache_node = kmem_cache_node; 3273 3274 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3275 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); 3276 3277 kmem_cache_bootstrap_fixup(kmem_cache_node); 3278 3279 caches++; 3280 kmem_cache_bootstrap_fixup(kmem_cache); 3281 caches++; 3282 /* Free temporary boot structure */ 3283 free_pages((unsigned long)temp_kmem_cache, order); 3284 3285 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 3286 3287 /* 3288 * Patch up the size_index table if we have strange large alignment 3289 * requirements for the kmalloc array. This is only the case for 3290 * MIPS it seems. The standard arches will not generate any code here. 3291 * 3292 * Largest permitted alignment is 256 bytes due to the way we 3293 * handle the index determination for the smaller caches. 3294 * 3295 * Make sure that nothing crazy happens if someone starts tinkering 3296 * around with ARCH_KMALLOC_MINALIGN 3297 */ 3298 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3299 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3300 3301 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { 3302 int elem = size_index_elem(i); 3303 if (elem >= ARRAY_SIZE(size_index)) 3304 break; 3305 size_index[elem] = KMALLOC_SHIFT_LOW; 3306 } 3307 3308 if (KMALLOC_MIN_SIZE == 64) { 3309 /* 3310 * The 96 byte size cache is not used if the alignment 3311 * is 64 byte. 3312 */ 3313 for (i = 64 + 8; i <= 96; i += 8) 3314 size_index[size_index_elem(i)] = 7; 3315 } else if (KMALLOC_MIN_SIZE == 128) { 3316 /* 3317 * The 192 byte sized cache is not used if the alignment 3318 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3319 * instead. 3320 */ 3321 for (i = 128 + 8; i <= 192; i += 8) 3322 size_index[size_index_elem(i)] = 8; 3323 } 3324 3325 /* Caches that are not of the two-to-the-power-of size */ 3326 if (KMALLOC_MIN_SIZE <= 32) { 3327 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); 3328 caches++; 3329 } 3330 3331 if (KMALLOC_MIN_SIZE <= 64) { 3332 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); 3333 caches++; 3334 } 3335 3336 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3337 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); 3338 caches++; 3339 } 3340 3341 slab_state = UP; 3342 3343 /* Provide the correct kmalloc names now that the caches are up */ 3344 if (KMALLOC_MIN_SIZE <= 32) { 3345 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); 3346 BUG_ON(!kmalloc_caches[1]->name); 3347 } 3348 3349 if (KMALLOC_MIN_SIZE <= 64) { 3350 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); 3351 BUG_ON(!kmalloc_caches[2]->name); 3352 } 3353 3354 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3355 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); 3356 3357 BUG_ON(!s); 3358 kmalloc_caches[i]->name = s; 3359 } 3360 3361#ifdef CONFIG_SMP 3362 register_cpu_notifier(&slab_notifier); 3363#endif 3364 3365#ifdef CONFIG_ZONE_DMA 3366 for (i = 0; i < SLUB_PAGE_SHIFT; i++) { 3367 struct kmem_cache *s = kmalloc_caches[i]; 3368 3369 if (s && s->size) { 3370 char *name = kasprintf(GFP_NOWAIT, 3371 "dma-kmalloc-%d", s->objsize); 3372 3373 BUG_ON(!name); 3374 kmalloc_dma_caches[i] = create_kmalloc_cache(name, 3375 s->objsize, SLAB_CACHE_DMA); 3376 } 3377 } 3378#endif 3379 printk(KERN_INFO 3380 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3381 " CPUs=%d, Nodes=%d\n", 3382 caches, cache_line_size(), 3383 slub_min_order, slub_max_order, slub_min_objects, 3384 nr_cpu_ids, nr_node_ids); 3385} 3386 3387void __init kmem_cache_init_late(void) 3388{ 3389} 3390 3391/* 3392 * Find a mergeable slab cache 3393 */ 3394static int slab_unmergeable(struct kmem_cache *s) 3395{ 3396 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3397 return 1; 3398 3399 if (s->ctor) 3400 return 1; 3401 3402 /* 3403 * We may have set a slab to be unmergeable during bootstrap. 3404 */ 3405 if (s->refcount < 0) 3406 return 1; 3407 3408 return 0; 3409} 3410 3411static struct kmem_cache *find_mergeable(size_t size, 3412 size_t align, unsigned long flags, const char *name, 3413 void (*ctor)(void *)) 3414{ 3415 struct kmem_cache *s; 3416 3417 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3418 return NULL; 3419 3420 if (ctor) 3421 return NULL; 3422 3423 size = ALIGN(size, sizeof(void *)); 3424 align = calculate_alignment(flags, align, size); 3425 size = ALIGN(size, align); 3426 flags = kmem_cache_flags(size, flags, name, NULL); 3427 3428 list_for_each_entry(s, &slab_caches, list) { 3429 if (slab_unmergeable(s)) 3430 continue; 3431 3432 if (size > s->size) 3433 continue; 3434 3435 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3436 continue; 3437 /* 3438 * Check if alignment is compatible. 3439 * Courtesy of Adrian Drzewiecki 3440 */ 3441 if ((s->size & ~(align - 1)) != s->size) 3442 continue; 3443 3444 if (s->size - size >= sizeof(void *)) 3445 continue; 3446 3447 return s; 3448 } 3449 return NULL; 3450} 3451 3452struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3453 size_t align, unsigned long flags, void (*ctor)(void *)) 3454{ 3455 struct kmem_cache *s; 3456 char *n; 3457 3458 if (WARN_ON(!name)) 3459 return NULL; 3460 3461 down_write(&slub_lock); 3462 s = find_mergeable(size, align, flags, name, ctor); 3463 if (s) { 3464 s->refcount++; 3465 /* 3466 * Adjust the object sizes so that we clear 3467 * the complete object on kzalloc. 3468 */ 3469 s->objsize = max(s->objsize, (int)size); 3470 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3471 3472 if (sysfs_slab_alias(s, name)) { 3473 s->refcount--; 3474 goto err; 3475 } 3476 up_write(&slub_lock); 3477 return s; 3478 } 3479 3480 n = kstrdup(name, GFP_KERNEL); 3481 if (!n) 3482 goto err; 3483 3484 s = kmalloc(kmem_size, GFP_KERNEL); 3485 if (s) { 3486 if (kmem_cache_open(s, n, 3487 size, align, flags, ctor)) { 3488 list_add(&s->list, &slab_caches); 3489 if (sysfs_slab_add(s)) { 3490 list_del(&s->list); 3491 kfree(n); 3492 kfree(s); 3493 goto err; 3494 } 3495 up_write(&slub_lock); 3496 return s; 3497 } 3498 kfree(n); 3499 kfree(s); 3500 } 3501err: 3502 up_write(&slub_lock); 3503 3504 if (flags & SLAB_PANIC) 3505 panic("Cannot create slabcache %s\n", name); 3506 else 3507 s = NULL; 3508 return s; 3509} 3510EXPORT_SYMBOL(kmem_cache_create); 3511 3512#ifdef CONFIG_SMP 3513/* 3514 * Use the cpu notifier to insure that the cpu slabs are flushed when 3515 * necessary. 3516 */ 3517static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3518 unsigned long action, void *hcpu) 3519{ 3520 long cpu = (long)hcpu; 3521 struct kmem_cache *s; 3522 unsigned long flags; 3523 3524 switch (action) { 3525 case CPU_UP_CANCELED: 3526 case CPU_UP_CANCELED_FROZEN: 3527 case CPU_DEAD: 3528 case CPU_DEAD_FROZEN: 3529 down_read(&slub_lock); 3530 list_for_each_entry(s, &slab_caches, list) { 3531 local_irq_save(flags); 3532 __flush_cpu_slab(s, cpu); 3533 local_irq_restore(flags); 3534 } 3535 up_read(&slub_lock); 3536 break; 3537 default: 3538 break; 3539 } 3540 return NOTIFY_OK; 3541} 3542 3543static struct notifier_block __cpuinitdata slab_notifier = { 3544 .notifier_call = slab_cpuup_callback 3545}; 3546 3547#endif 3548 3549void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3550{ 3551 struct kmem_cache *s; 3552 void *ret; 3553 3554 if (unlikely(size > SLUB_MAX_SIZE)) 3555 return kmalloc_large(size, gfpflags); 3556 3557 s = get_slab(size, gfpflags); 3558 3559 if (unlikely(ZERO_OR_NULL_PTR(s))) 3560 return s; 3561 3562 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); 3563 3564 /* Honor the call site pointer we recieved. */ 3565 trace_kmalloc(caller, ret, size, s->size, gfpflags); 3566 3567 return ret; 3568} 3569 3570#ifdef CONFIG_NUMA 3571void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3572 int node, unsigned long caller) 3573{ 3574 struct kmem_cache *s; 3575 void *ret; 3576 3577 if (unlikely(size > SLUB_MAX_SIZE)) { 3578 ret = kmalloc_large_node(size, gfpflags, node); 3579 3580 trace_kmalloc_node(caller, ret, 3581 size, PAGE_SIZE << get_order(size), 3582 gfpflags, node); 3583 3584 return ret; 3585 } 3586 3587 s = get_slab(size, gfpflags); 3588 3589 if (unlikely(ZERO_OR_NULL_PTR(s))) 3590 return s; 3591 3592 ret = slab_alloc(s, gfpflags, node, caller); 3593 3594 /* Honor the call site pointer we recieved. */ 3595 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 3596 3597 return ret; 3598} 3599#endif 3600 3601#ifdef CONFIG_SYSFS 3602static int count_inuse(struct page *page) 3603{ 3604 return page->inuse; 3605} 3606 3607static int count_total(struct page *page) 3608{ 3609 return page->objects; 3610} 3611#endif 3612 3613#ifdef CONFIG_SLUB_DEBUG 3614static int validate_slab(struct kmem_cache *s, struct page *page, 3615 unsigned long *map) 3616{ 3617 void *p; 3618 void *addr = page_address(page); 3619 3620 if (!check_slab(s, page) || 3621 !on_freelist(s, page, NULL)) 3622 return 0; 3623 3624 /* Now we know that a valid freelist exists */ 3625 bitmap_zero(map, page->objects); 3626 3627 get_map(s, page, map); 3628 for_each_object(p, s, addr, page->objects) { 3629 if (test_bit(slab_index(p, s, addr), map)) 3630 if (!check_object(s, page, p, SLUB_RED_INACTIVE)) 3631 return 0; 3632 } 3633 3634 for_each_object(p, s, addr, page->objects) 3635 if (!test_bit(slab_index(p, s, addr), map)) 3636 if (!check_object(s, page, p, SLUB_RED_ACTIVE)) 3637 return 0; 3638 return 1; 3639} 3640 3641static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3642 unsigned long *map) 3643{ 3644 if (slab_trylock(page)) { 3645 validate_slab(s, page, map); 3646 slab_unlock(page); 3647 } else 3648 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3649 s->name, page); 3650} 3651 3652static int validate_slab_node(struct kmem_cache *s, 3653 struct kmem_cache_node *n, unsigned long *map) 3654{ 3655 unsigned long count = 0; 3656 struct page *page; 3657 unsigned long flags; 3658 3659 spin_lock_irqsave(&n->list_lock, flags); 3660 3661 list_for_each_entry(page, &n->partial, lru) { 3662 validate_slab_slab(s, page, map); 3663 count++; 3664 } 3665 if (count != n->nr_partial) 3666 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3667 "counter=%ld\n", s->name, count, n->nr_partial); 3668 3669 if (!(s->flags & SLAB_STORE_USER)) 3670 goto out; 3671 3672 list_for_each_entry(page, &n->full, lru) { 3673 validate_slab_slab(s, page, map); 3674 count++; 3675 } 3676 if (count != atomic_long_read(&n->nr_slabs)) 3677 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3678 "counter=%ld\n", s->name, count, 3679 atomic_long_read(&n->nr_slabs)); 3680 3681out: 3682 spin_unlock_irqrestore(&n->list_lock, flags); 3683 return count; 3684} 3685 3686static long validate_slab_cache(struct kmem_cache *s) 3687{ 3688 int node; 3689 unsigned long count = 0; 3690 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3691 sizeof(unsigned long), GFP_KERNEL); 3692 3693 if (!map) 3694 return -ENOMEM; 3695 3696 flush_all(s); 3697 for_each_node_state(node, N_NORMAL_MEMORY) { 3698 struct kmem_cache_node *n = get_node(s, node); 3699 3700 count += validate_slab_node(s, n, map); 3701 } 3702 kfree(map); 3703 return count; 3704} 3705/* 3706 * Generate lists of code addresses where slabcache objects are allocated 3707 * and freed. 3708 */ 3709 3710struct location { 3711 unsigned long count; 3712 unsigned long addr; 3713 long long sum_time; 3714 long min_time; 3715 long max_time; 3716 long min_pid; 3717 long max_pid; 3718 DECLARE_BITMAP(cpus, NR_CPUS); 3719 nodemask_t nodes; 3720}; 3721 3722struct loc_track { 3723 unsigned long max; 3724 unsigned long count; 3725 struct location *loc; 3726}; 3727 3728static void free_loc_track(struct loc_track *t) 3729{ 3730 if (t->max) 3731 free_pages((unsigned long)t->loc, 3732 get_order(sizeof(struct location) * t->max)); 3733} 3734 3735static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3736{ 3737 struct location *l; 3738 int order; 3739 3740 order = get_order(sizeof(struct location) * max); 3741 3742 l = (void *)__get_free_pages(flags, order); 3743 if (!l) 3744 return 0; 3745 3746 if (t->count) { 3747 memcpy(l, t->loc, sizeof(struct location) * t->count); 3748 free_loc_track(t); 3749 } 3750 t->max = max; 3751 t->loc = l; 3752 return 1; 3753} 3754 3755static int add_location(struct loc_track *t, struct kmem_cache *s, 3756 const struct track *track) 3757{ 3758 long start, end, pos; 3759 struct location *l; 3760 unsigned long caddr; 3761 unsigned long age = jiffies - track->when; 3762 3763 start = -1; 3764 end = t->count; 3765 3766 for ( ; ; ) { 3767 pos = start + (end - start + 1) / 2; 3768 3769 /* 3770 * There is nothing at "end". If we end up there 3771 * we need to add something to before end. 3772 */ 3773 if (pos == end) 3774 break; 3775 3776 caddr = t->loc[pos].addr; 3777 if (track->addr == caddr) { 3778 3779 l = &t->loc[pos]; 3780 l->count++; 3781 if (track->when) { 3782 l->sum_time += age; 3783 if (age < l->min_time) 3784 l->min_time = age; 3785 if (age > l->max_time) 3786 l->max_time = age; 3787 3788 if (track->pid < l->min_pid) 3789 l->min_pid = track->pid; 3790 if (track->pid > l->max_pid) 3791 l->max_pid = track->pid; 3792 3793 cpumask_set_cpu(track->cpu, 3794 to_cpumask(l->cpus)); 3795 } 3796 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3797 return 1; 3798 } 3799 3800 if (track->addr < caddr) 3801 end = pos; 3802 else 3803 start = pos; 3804 } 3805 3806 /* 3807 * Not found. Insert new tracking element. 3808 */ 3809 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3810 return 0; 3811 3812 l = t->loc + pos; 3813 if (pos < t->count) 3814 memmove(l + 1, l, 3815 (t->count - pos) * sizeof(struct location)); 3816 t->count++; 3817 l->count = 1; 3818 l->addr = track->addr; 3819 l->sum_time = age; 3820 l->min_time = age; 3821 l->max_time = age; 3822 l->min_pid = track->pid; 3823 l->max_pid = track->pid; 3824 cpumask_clear(to_cpumask(l->cpus)); 3825 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 3826 nodes_clear(l->nodes); 3827 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3828 return 1; 3829} 3830 3831static void process_slab(struct loc_track *t, struct kmem_cache *s, 3832 struct page *page, enum track_item alloc, 3833 unsigned long *map) 3834{ 3835 void *addr = page_address(page); 3836 void *p; 3837 3838 bitmap_zero(map, page->objects); 3839 get_map(s, page, map); 3840 3841 for_each_object(p, s, addr, page->objects) 3842 if (!test_bit(slab_index(p, s, addr), map)) 3843 add_location(t, s, get_track(s, p, alloc)); 3844} 3845 3846static int list_locations(struct kmem_cache *s, char *buf, 3847 enum track_item alloc) 3848{ 3849 int len = 0; 3850 unsigned long i; 3851 struct loc_track t = { 0, 0, NULL }; 3852 int node; 3853 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3854 sizeof(unsigned long), GFP_KERNEL); 3855 3856 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3857 GFP_TEMPORARY)) { 3858 kfree(map); 3859 return sprintf(buf, "Out of memory\n"); 3860 } 3861 /* Push back cpu slabs */ 3862 flush_all(s); 3863 3864 for_each_node_state(node, N_NORMAL_MEMORY) { 3865 struct kmem_cache_node *n = get_node(s, node); 3866 unsigned long flags; 3867 struct page *page; 3868 3869 if (!atomic_long_read(&n->nr_slabs)) 3870 continue; 3871 3872 spin_lock_irqsave(&n->list_lock, flags); 3873 list_for_each_entry(page, &n->partial, lru) 3874 process_slab(&t, s, page, alloc, map); 3875 list_for_each_entry(page, &n->full, lru) 3876 process_slab(&t, s, page, alloc, map); 3877 spin_unlock_irqrestore(&n->list_lock, flags); 3878 } 3879 3880 for (i = 0; i < t.count; i++) { 3881 struct location *l = &t.loc[i]; 3882 3883 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 3884 break; 3885 len += sprintf(buf + len, "%7ld ", l->count); 3886 3887 if (l->addr) 3888 len += sprintf(buf + len, "%pS", (void *)l->addr); 3889 else 3890 len += sprintf(buf + len, "<not-available>"); 3891 3892 if (l->sum_time != l->min_time) { 3893 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3894 l->min_time, 3895 (long)div_u64(l->sum_time, l->count), 3896 l->max_time); 3897 } else 3898 len += sprintf(buf + len, " age=%ld", 3899 l->min_time); 3900 3901 if (l->min_pid != l->max_pid) 3902 len += sprintf(buf + len, " pid=%ld-%ld", 3903 l->min_pid, l->max_pid); 3904 else 3905 len += sprintf(buf + len, " pid=%ld", 3906 l->min_pid); 3907 3908 if (num_online_cpus() > 1 && 3909 !cpumask_empty(to_cpumask(l->cpus)) && 3910 len < PAGE_SIZE - 60) { 3911 len += sprintf(buf + len, " cpus="); 3912 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3913 to_cpumask(l->cpus)); 3914 } 3915 3916 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 3917 len < PAGE_SIZE - 60) { 3918 len += sprintf(buf + len, " nodes="); 3919 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3920 l->nodes); 3921 } 3922 3923 len += sprintf(buf + len, "\n"); 3924 } 3925 3926 free_loc_track(&t); 3927 kfree(map); 3928 if (!t.count) 3929 len += sprintf(buf, "No data\n"); 3930 return len; 3931} 3932#endif 3933 3934#ifdef SLUB_RESILIENCY_TEST 3935static void resiliency_test(void) 3936{ 3937 u8 *p; 3938 3939 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); 3940 3941 printk(KERN_ERR "SLUB resiliency testing\n"); 3942 printk(KERN_ERR "-----------------------\n"); 3943 printk(KERN_ERR "A. Corruption after allocation\n"); 3944 3945 p = kzalloc(16, GFP_KERNEL); 3946 p[16] = 0x12; 3947 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3948 " 0x12->0x%p\n\n", p + 16); 3949 3950 validate_slab_cache(kmalloc_caches[4]); 3951 3952 /* Hmmm... The next two are dangerous */ 3953 p = kzalloc(32, GFP_KERNEL); 3954 p[32 + sizeof(void *)] = 0x34; 3955 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3956 " 0x34 -> -0x%p\n", p); 3957 printk(KERN_ERR 3958 "If allocated object is overwritten then not detectable\n\n"); 3959 3960 validate_slab_cache(kmalloc_caches[5]); 3961 p = kzalloc(64, GFP_KERNEL); 3962 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3963 *p = 0x56; 3964 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3965 p); 3966 printk(KERN_ERR 3967 "If allocated object is overwritten then not detectable\n\n"); 3968 validate_slab_cache(kmalloc_caches[6]); 3969 3970 printk(KERN_ERR "\nB. Corruption after free\n"); 3971 p = kzalloc(128, GFP_KERNEL); 3972 kfree(p); 3973 *p = 0x78; 3974 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3975 validate_slab_cache(kmalloc_caches[7]); 3976 3977 p = kzalloc(256, GFP_KERNEL); 3978 kfree(p); 3979 p[50] = 0x9a; 3980 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3981 p); 3982 validate_slab_cache(kmalloc_caches[8]); 3983 3984 p = kzalloc(512, GFP_KERNEL); 3985 kfree(p); 3986 p[512] = 0xab; 3987 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3988 validate_slab_cache(kmalloc_caches[9]); 3989} 3990#else 3991#ifdef CONFIG_SYSFS 3992static void resiliency_test(void) {}; 3993#endif 3994#endif 3995 3996#ifdef CONFIG_SYSFS 3997enum slab_stat_type { 3998 SL_ALL, /* All slabs */ 3999 SL_PARTIAL, /* Only partially allocated slabs */ 4000 SL_CPU, /* Only slabs used for cpu caches */ 4001 SL_OBJECTS, /* Determine allocated objects not slabs */ 4002 SL_TOTAL /* Determine object capacity not slabs */ 4003}; 4004 4005#define SO_ALL (1 << SL_ALL) 4006#define SO_PARTIAL (1 << SL_PARTIAL) 4007#define SO_CPU (1 << SL_CPU) 4008#define SO_OBJECTS (1 << SL_OBJECTS) 4009#define SO_TOTAL (1 << SL_TOTAL) 4010 4011static ssize_t show_slab_objects(struct kmem_cache *s, 4012 char *buf, unsigned long flags) 4013{ 4014 unsigned long total = 0; 4015 int node; 4016 int x; 4017 unsigned long *nodes; 4018 unsigned long *per_cpu; 4019 4020 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 4021 if (!nodes) 4022 return -ENOMEM; 4023 per_cpu = nodes + nr_node_ids; 4024 4025 if (flags & SO_CPU) { 4026 int cpu; 4027 4028 for_each_possible_cpu(cpu) { 4029 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 4030 4031 if (!c || c->node < 0) 4032 continue; 4033 4034 if (c->page) { 4035 if (flags & SO_TOTAL) 4036 x = c->page->objects; 4037 else if (flags & SO_OBJECTS) 4038 x = c->page->inuse; 4039 else 4040 x = 1; 4041 4042 total += x; 4043 nodes[c->node] += x; 4044 } 4045 per_cpu[c->node]++; 4046 } 4047 } 4048 4049 lock_memory_hotplug(); 4050#ifdef CONFIG_SLUB_DEBUG 4051 if (flags & SO_ALL) { 4052 for_each_node_state(node, N_NORMAL_MEMORY) { 4053 struct kmem_cache_node *n = get_node(s, node); 4054 4055 if (flags & SO_TOTAL) 4056 x = atomic_long_read(&n->total_objects); 4057 else if (flags & SO_OBJECTS) 4058 x = atomic_long_read(&n->total_objects) - 4059 count_partial(n, count_free); 4060 4061 else 4062 x = atomic_long_read(&n->nr_slabs); 4063 total += x; 4064 nodes[node] += x; 4065 } 4066 4067 } else 4068#endif 4069 if (flags & SO_PARTIAL) { 4070 for_each_node_state(node, N_NORMAL_MEMORY) { 4071 struct kmem_cache_node *n = get_node(s, node); 4072 4073 if (flags & SO_TOTAL) 4074 x = count_partial(n, count_total); 4075 else if (flags & SO_OBJECTS) 4076 x = count_partial(n, count_inuse); 4077 else 4078 x = n->nr_partial; 4079 total += x; 4080 nodes[node] += x; 4081 } 4082 } 4083 x = sprintf(buf, "%lu", total); 4084#ifdef CONFIG_NUMA 4085 for_each_node_state(node, N_NORMAL_MEMORY) 4086 if (nodes[node]) 4087 x += sprintf(buf + x, " N%d=%lu", 4088 node, nodes[node]); 4089#endif 4090 unlock_memory_hotplug(); 4091 kfree(nodes); 4092 return x + sprintf(buf + x, "\n"); 4093} 4094 4095#ifdef CONFIG_SLUB_DEBUG 4096static int any_slab_objects(struct kmem_cache *s) 4097{ 4098 int node; 4099 4100 for_each_online_node(node) { 4101 struct kmem_cache_node *n = get_node(s, node); 4102 4103 if (!n) 4104 continue; 4105 4106 if (atomic_long_read(&n->total_objects)) 4107 return 1; 4108 } 4109 return 0; 4110} 4111#endif 4112 4113#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 4114#define to_slab(n) container_of(n, struct kmem_cache, kobj); 4115 4116struct slab_attribute { 4117 struct attribute attr; 4118 ssize_t (*show)(struct kmem_cache *s, char *buf); 4119 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 4120}; 4121 4122#define SLAB_ATTR_RO(_name) \ 4123 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 4124 4125#define SLAB_ATTR(_name) \ 4126 static struct slab_attribute _name##_attr = \ 4127 __ATTR(_name, 0644, _name##_show, _name##_store) 4128 4129static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 4130{ 4131 return sprintf(buf, "%d\n", s->size); 4132} 4133SLAB_ATTR_RO(slab_size); 4134 4135static ssize_t align_show(struct kmem_cache *s, char *buf) 4136{ 4137 return sprintf(buf, "%d\n", s->align); 4138} 4139SLAB_ATTR_RO(align); 4140 4141static ssize_t object_size_show(struct kmem_cache *s, char *buf) 4142{ 4143 return sprintf(buf, "%d\n", s->objsize); 4144} 4145SLAB_ATTR_RO(object_size); 4146 4147static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 4148{ 4149 return sprintf(buf, "%d\n", oo_objects(s->oo)); 4150} 4151SLAB_ATTR_RO(objs_per_slab); 4152 4153static ssize_t order_store(struct kmem_cache *s, 4154 const char *buf, size_t length) 4155{ 4156 unsigned long order; 4157 int err; 4158 4159 err = strict_strtoul(buf, 10, &order); 4160 if (err) 4161 return err; 4162 4163 if (order > slub_max_order || order < slub_min_order) 4164 return -EINVAL; 4165 4166 calculate_sizes(s, order); 4167 return length; 4168} 4169 4170static ssize_t order_show(struct kmem_cache *s, char *buf) 4171{ 4172 return sprintf(buf, "%d\n", oo_order(s->oo)); 4173} 4174SLAB_ATTR(order); 4175 4176static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 4177{ 4178 return sprintf(buf, "%lu\n", s->min_partial); 4179} 4180 4181static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 4182 size_t length) 4183{ 4184 unsigned long min; 4185 int err; 4186 4187 err = strict_strtoul(buf, 10, &min); 4188 if (err) 4189 return err; 4190 4191 set_min_partial(s, min); 4192 return length; 4193} 4194SLAB_ATTR(min_partial); 4195 4196static ssize_t ctor_show(struct kmem_cache *s, char *buf) 4197{ 4198 if (!s->ctor) 4199 return 0; 4200 return sprintf(buf, "%pS\n", s->ctor); 4201} 4202SLAB_ATTR_RO(ctor); 4203 4204static ssize_t aliases_show(struct kmem_cache *s, char *buf) 4205{ 4206 return sprintf(buf, "%d\n", s->refcount - 1); 4207} 4208SLAB_ATTR_RO(aliases); 4209 4210static ssize_t partial_show(struct kmem_cache *s, char *buf) 4211{ 4212 return show_slab_objects(s, buf, SO_PARTIAL); 4213} 4214SLAB_ATTR_RO(partial); 4215 4216static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4217{ 4218 return show_slab_objects(s, buf, SO_CPU); 4219} 4220SLAB_ATTR_RO(cpu_slabs); 4221 4222static ssize_t objects_show(struct kmem_cache *s, char *buf) 4223{ 4224 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4225} 4226SLAB_ATTR_RO(objects); 4227 4228static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4229{ 4230 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4231} 4232SLAB_ATTR_RO(objects_partial); 4233 4234static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4235{ 4236 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4237} 4238 4239static ssize_t reclaim_account_store(struct kmem_cache *s, 4240 const char *buf, size_t length) 4241{ 4242 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4243 if (buf[0] == '1') 4244 s->flags |= SLAB_RECLAIM_ACCOUNT; 4245 return length; 4246} 4247SLAB_ATTR(reclaim_account); 4248 4249static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4250{ 4251 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4252} 4253SLAB_ATTR_RO(hwcache_align); 4254 4255#ifdef CONFIG_ZONE_DMA 4256static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4257{ 4258 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4259} 4260SLAB_ATTR_RO(cache_dma); 4261#endif 4262 4263static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4264{ 4265 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4266} 4267SLAB_ATTR_RO(destroy_by_rcu); 4268 4269static ssize_t reserved_show(struct kmem_cache *s, char *buf) 4270{ 4271 return sprintf(buf, "%d\n", s->reserved); 4272} 4273SLAB_ATTR_RO(reserved); 4274 4275#ifdef CONFIG_SLUB_DEBUG 4276static ssize_t slabs_show(struct kmem_cache *s, char *buf) 4277{ 4278 return show_slab_objects(s, buf, SO_ALL); 4279} 4280SLAB_ATTR_RO(slabs); 4281 4282static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 4283{ 4284 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 4285} 4286SLAB_ATTR_RO(total_objects); 4287 4288static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 4289{ 4290 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 4291} 4292 4293static ssize_t sanity_checks_store(struct kmem_cache *s, 4294 const char *buf, size_t length) 4295{ 4296 s->flags &= ~SLAB_DEBUG_FREE; 4297 if (buf[0] == '1') 4298 s->flags |= SLAB_DEBUG_FREE; 4299 return length; 4300} 4301SLAB_ATTR(sanity_checks); 4302 4303static ssize_t trace_show(struct kmem_cache *s, char *buf) 4304{ 4305 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4306} 4307 4308static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4309 size_t length) 4310{ 4311 s->flags &= ~SLAB_TRACE; 4312 if (buf[0] == '1') 4313 s->flags |= SLAB_TRACE; 4314 return length; 4315} 4316SLAB_ATTR(trace); 4317 4318static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4319{ 4320 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4321} 4322 4323static ssize_t red_zone_store(struct kmem_cache *s, 4324 const char *buf, size_t length) 4325{ 4326 if (any_slab_objects(s)) 4327 return -EBUSY; 4328 4329 s->flags &= ~SLAB_RED_ZONE; 4330 if (buf[0] == '1') 4331 s->flags |= SLAB_RED_ZONE; 4332 calculate_sizes(s, -1); 4333 return length; 4334} 4335SLAB_ATTR(red_zone); 4336 4337static ssize_t poison_show(struct kmem_cache *s, char *buf) 4338{ 4339 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4340} 4341 4342static ssize_t poison_store(struct kmem_cache *s, 4343 const char *buf, size_t length) 4344{ 4345 if (any_slab_objects(s)) 4346 return -EBUSY; 4347 4348 s->flags &= ~SLAB_POISON; 4349 if (buf[0] == '1') 4350 s->flags |= SLAB_POISON; 4351 calculate_sizes(s, -1); 4352 return length; 4353} 4354SLAB_ATTR(poison); 4355 4356static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4357{ 4358 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4359} 4360 4361static ssize_t store_user_store(struct kmem_cache *s, 4362 const char *buf, size_t length) 4363{ 4364 if (any_slab_objects(s)) 4365 return -EBUSY; 4366 4367 s->flags &= ~SLAB_STORE_USER; 4368 if (buf[0] == '1') 4369 s->flags |= SLAB_STORE_USER; 4370 calculate_sizes(s, -1); 4371 return length; 4372} 4373SLAB_ATTR(store_user); 4374 4375static ssize_t validate_show(struct kmem_cache *s, char *buf) 4376{ 4377 return 0; 4378} 4379 4380static ssize_t validate_store(struct kmem_cache *s, 4381 const char *buf, size_t length) 4382{ 4383 int ret = -EINVAL; 4384 4385 if (buf[0] == '1') { 4386 ret = validate_slab_cache(s); 4387 if (ret >= 0) 4388 ret = length; 4389 } 4390 return ret; 4391} 4392SLAB_ATTR(validate); 4393 4394static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4395{ 4396 if (!(s->flags & SLAB_STORE_USER)) 4397 return -ENOSYS; 4398 return list_locations(s, buf, TRACK_ALLOC); 4399} 4400SLAB_ATTR_RO(alloc_calls); 4401 4402static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4403{ 4404 if (!(s->flags & SLAB_STORE_USER)) 4405 return -ENOSYS; 4406 return list_locations(s, buf, TRACK_FREE); 4407} 4408SLAB_ATTR_RO(free_calls); 4409#endif /* CONFIG_SLUB_DEBUG */ 4410 4411#ifdef CONFIG_FAILSLAB 4412static ssize_t failslab_show(struct kmem_cache *s, char *buf) 4413{ 4414 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 4415} 4416 4417static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 4418 size_t length) 4419{ 4420 s->flags &= ~SLAB_FAILSLAB; 4421 if (buf[0] == '1') 4422 s->flags |= SLAB_FAILSLAB; 4423 return length; 4424} 4425SLAB_ATTR(failslab); 4426#endif 4427 4428static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4429{ 4430 return 0; 4431} 4432 4433static ssize_t shrink_store(struct kmem_cache *s, 4434 const char *buf, size_t length) 4435{ 4436 if (buf[0] == '1') { 4437 int rc = kmem_cache_shrink(s); 4438 4439 if (rc) 4440 return rc; 4441 } else 4442 return -EINVAL; 4443 return length; 4444} 4445SLAB_ATTR(shrink); 4446 4447#ifdef CONFIG_NUMA 4448static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4449{ 4450 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4451} 4452 4453static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4454 const char *buf, size_t length) 4455{ 4456 unsigned long ratio; 4457 int err; 4458 4459 err = strict_strtoul(buf, 10, &ratio); 4460 if (err) 4461 return err; 4462 4463 if (ratio <= 100) 4464 s->remote_node_defrag_ratio = ratio * 10; 4465 4466 return length; 4467} 4468SLAB_ATTR(remote_node_defrag_ratio); 4469#endif 4470 4471#ifdef CONFIG_SLUB_STATS 4472static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4473{ 4474 unsigned long sum = 0; 4475 int cpu; 4476 int len; 4477 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4478 4479 if (!data) 4480 return -ENOMEM; 4481 4482 for_each_online_cpu(cpu) { 4483 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 4484 4485 data[cpu] = x; 4486 sum += x; 4487 } 4488 4489 len = sprintf(buf, "%lu", sum); 4490 4491#ifdef CONFIG_SMP 4492 for_each_online_cpu(cpu) { 4493 if (data[cpu] && len < PAGE_SIZE - 20) 4494 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4495 } 4496#endif 4497 kfree(data); 4498 return len + sprintf(buf + len, "\n"); 4499} 4500 4501static void clear_stat(struct kmem_cache *s, enum stat_item si) 4502{ 4503 int cpu; 4504 4505 for_each_online_cpu(cpu) 4506 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 4507} 4508 4509#define STAT_ATTR(si, text) \ 4510static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4511{ \ 4512 return show_stat(s, buf, si); \ 4513} \ 4514static ssize_t text##_store(struct kmem_cache *s, \ 4515 const char *buf, size_t length) \ 4516{ \ 4517 if (buf[0] != '0') \ 4518 return -EINVAL; \ 4519 clear_stat(s, si); \ 4520 return length; \ 4521} \ 4522SLAB_ATTR(text); \ 4523 4524STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4525STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4526STAT_ATTR(FREE_FASTPATH, free_fastpath); 4527STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4528STAT_ATTR(FREE_FROZEN, free_frozen); 4529STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4530STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4531STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4532STAT_ATTR(ALLOC_SLAB, alloc_slab); 4533STAT_ATTR(ALLOC_REFILL, alloc_refill); 4534STAT_ATTR(FREE_SLAB, free_slab); 4535STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4536STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4537STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4538STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4539STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4540STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4541STAT_ATTR(ORDER_FALLBACK, order_fallback); 4542#endif 4543 4544static struct attribute *slab_attrs[] = { 4545 &slab_size_attr.attr, 4546 &object_size_attr.attr, 4547 &objs_per_slab_attr.attr, 4548 &order_attr.attr, 4549 &min_partial_attr.attr, 4550 &objects_attr.attr, 4551 &objects_partial_attr.attr, 4552 &partial_attr.attr, 4553 &cpu_slabs_attr.attr, 4554 &ctor_attr.attr, 4555 &aliases_attr.attr, 4556 &align_attr.attr, 4557 &hwcache_align_attr.attr, 4558 &reclaim_account_attr.attr, 4559 &destroy_by_rcu_attr.attr, 4560 &shrink_attr.attr, 4561 &reserved_attr.attr, 4562#ifdef CONFIG_SLUB_DEBUG 4563 &total_objects_attr.attr, 4564 &slabs_attr.attr, 4565 &sanity_checks_attr.attr, 4566 &trace_attr.attr, 4567 &red_zone_attr.attr, 4568 &poison_attr.attr, 4569 &store_user_attr.attr, 4570 &validate_attr.attr, 4571 &alloc_calls_attr.attr, 4572 &free_calls_attr.attr, 4573#endif 4574#ifdef CONFIG_ZONE_DMA 4575 &cache_dma_attr.attr, 4576#endif 4577#ifdef CONFIG_NUMA 4578 &remote_node_defrag_ratio_attr.attr, 4579#endif 4580#ifdef CONFIG_SLUB_STATS 4581 &alloc_fastpath_attr.attr, 4582 &alloc_slowpath_attr.attr, 4583 &free_fastpath_attr.attr, 4584 &free_slowpath_attr.attr, 4585 &free_frozen_attr.attr, 4586 &free_add_partial_attr.attr, 4587 &free_remove_partial_attr.attr, 4588 &alloc_from_partial_attr.attr, 4589 &alloc_slab_attr.attr, 4590 &alloc_refill_attr.attr, 4591 &free_slab_attr.attr, 4592 &cpuslab_flush_attr.attr, 4593 &deactivate_full_attr.attr, 4594 &deactivate_empty_attr.attr, 4595 &deactivate_to_head_attr.attr, 4596 &deactivate_to_tail_attr.attr, 4597 &deactivate_remote_frees_attr.attr, 4598 &order_fallback_attr.attr, 4599#endif 4600#ifdef CONFIG_FAILSLAB 4601 &failslab_attr.attr, 4602#endif 4603 4604 NULL 4605}; 4606 4607static struct attribute_group slab_attr_group = { 4608 .attrs = slab_attrs, 4609}; 4610 4611static ssize_t slab_attr_show(struct kobject *kobj, 4612 struct attribute *attr, 4613 char *buf) 4614{ 4615 struct slab_attribute *attribute; 4616 struct kmem_cache *s; 4617 int err; 4618 4619 attribute = to_slab_attr(attr); 4620 s = to_slab(kobj); 4621 4622 if (!attribute->show) 4623 return -EIO; 4624 4625 err = attribute->show(s, buf); 4626 4627 return err; 4628} 4629 4630static ssize_t slab_attr_store(struct kobject *kobj, 4631 struct attribute *attr, 4632 const char *buf, size_t len) 4633{ 4634 struct slab_attribute *attribute; 4635 struct kmem_cache *s; 4636 int err; 4637 4638 attribute = to_slab_attr(attr); 4639 s = to_slab(kobj); 4640 4641 if (!attribute->store) 4642 return -EIO; 4643 4644 err = attribute->store(s, buf, len); 4645 4646 return err; 4647} 4648 4649static void kmem_cache_release(struct kobject *kobj) 4650{ 4651 struct kmem_cache *s = to_slab(kobj); 4652 4653 kfree(s->name); 4654 kfree(s); 4655} 4656 4657static const struct sysfs_ops slab_sysfs_ops = { 4658 .show = slab_attr_show, 4659 .store = slab_attr_store, 4660}; 4661 4662static struct kobj_type slab_ktype = { 4663 .sysfs_ops = &slab_sysfs_ops, 4664 .release = kmem_cache_release 4665}; 4666 4667static int uevent_filter(struct kset *kset, struct kobject *kobj) 4668{ 4669 struct kobj_type *ktype = get_ktype(kobj); 4670 4671 if (ktype == &slab_ktype) 4672 return 1; 4673 return 0; 4674} 4675 4676static const struct kset_uevent_ops slab_uevent_ops = { 4677 .filter = uevent_filter, 4678}; 4679 4680static struct kset *slab_kset; 4681 4682#define ID_STR_LENGTH 64 4683 4684/* Create a unique string id for a slab cache: 4685 * 4686 * Format :[flags-]size 4687 */ 4688static char *create_unique_id(struct kmem_cache *s) 4689{ 4690 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4691 char *p = name; 4692 4693 BUG_ON(!name); 4694 4695 *p++ = ':'; 4696 /* 4697 * First flags affecting slabcache operations. We will only 4698 * get here for aliasable slabs so we do not need to support 4699 * too many flags. The flags here must cover all flags that 4700 * are matched during merging to guarantee that the id is 4701 * unique. 4702 */ 4703 if (s->flags & SLAB_CACHE_DMA) 4704 *p++ = 'd'; 4705 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4706 *p++ = 'a'; 4707 if (s->flags & SLAB_DEBUG_FREE) 4708 *p++ = 'F'; 4709 if (!(s->flags & SLAB_NOTRACK)) 4710 *p++ = 't'; 4711 if (p != name + 1) 4712 *p++ = '-'; 4713 p += sprintf(p, "%07d", s->size); 4714 BUG_ON(p > name + ID_STR_LENGTH - 1); 4715 return name; 4716} 4717 4718static int sysfs_slab_add(struct kmem_cache *s) 4719{ 4720 int err; 4721 const char *name; 4722 int unmergeable; 4723 4724 if (slab_state < SYSFS) 4725 /* Defer until later */ 4726 return 0; 4727 4728 unmergeable = slab_unmergeable(s); 4729 if (unmergeable) { 4730 /* 4731 * Slabcache can never be merged so we can use the name proper. 4732 * This is typically the case for debug situations. In that 4733 * case we can catch duplicate names easily. 4734 */ 4735 sysfs_remove_link(&slab_kset->kobj, s->name); 4736 name = s->name; 4737 } else { 4738 /* 4739 * Create a unique name for the slab as a target 4740 * for the symlinks. 4741 */ 4742 name = create_unique_id(s); 4743 } 4744 4745 s->kobj.kset = slab_kset; 4746 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4747 if (err) { 4748 kobject_put(&s->kobj); 4749 return err; 4750 } 4751 4752 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4753 if (err) { 4754 kobject_del(&s->kobj); 4755 kobject_put(&s->kobj); 4756 return err; 4757 } 4758 kobject_uevent(&s->kobj, KOBJ_ADD); 4759 if (!unmergeable) { 4760 /* Setup first alias */ 4761 sysfs_slab_alias(s, s->name); 4762 kfree(name); 4763 } 4764 return 0; 4765} 4766 4767static void sysfs_slab_remove(struct kmem_cache *s) 4768{ 4769 if (slab_state < SYSFS) 4770 /* 4771 * Sysfs has not been setup yet so no need to remove the 4772 * cache from sysfs. 4773 */ 4774 return; 4775 4776 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4777 kobject_del(&s->kobj); 4778 kobject_put(&s->kobj); 4779} 4780 4781/* 4782 * Need to buffer aliases during bootup until sysfs becomes 4783 * available lest we lose that information. 4784 */ 4785struct saved_alias { 4786 struct kmem_cache *s; 4787 const char *name; 4788 struct saved_alias *next; 4789}; 4790 4791static struct saved_alias *alias_list; 4792 4793static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4794{ 4795 struct saved_alias *al; 4796 4797 if (slab_state == SYSFS) { 4798 /* 4799 * If we have a leftover link then remove it. 4800 */ 4801 sysfs_remove_link(&slab_kset->kobj, name); 4802 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4803 } 4804 4805 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4806 if (!al) 4807 return -ENOMEM; 4808 4809 al->s = s; 4810 al->name = name; 4811 al->next = alias_list; 4812 alias_list = al; 4813 return 0; 4814} 4815 4816static int __init slab_sysfs_init(void) 4817{ 4818 struct kmem_cache *s; 4819 int err; 4820 4821 down_write(&slub_lock); 4822 4823 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4824 if (!slab_kset) { 4825 up_write(&slub_lock); 4826 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4827 return -ENOSYS; 4828 } 4829 4830 slab_state = SYSFS; 4831 4832 list_for_each_entry(s, &slab_caches, list) { 4833 err = sysfs_slab_add(s); 4834 if (err) 4835 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4836 " to sysfs\n", s->name); 4837 } 4838 4839 while (alias_list) { 4840 struct saved_alias *al = alias_list; 4841 4842 alias_list = alias_list->next; 4843 err = sysfs_slab_alias(al->s, al->name); 4844 if (err) 4845 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4846 " %s to sysfs\n", s->name); 4847 kfree(al); 4848 } 4849 4850 up_write(&slub_lock); 4851 resiliency_test(); 4852 return 0; 4853} 4854 4855__initcall(slab_sysfs_init); 4856#endif /* CONFIG_SYSFS */ 4857 4858/* 4859 * The /proc/slabinfo ABI 4860 */ 4861#ifdef CONFIG_SLABINFO 4862static void print_slabinfo_header(struct seq_file *m) 4863{ 4864 seq_puts(m, "slabinfo - version: 2.1\n"); 4865 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4866 "<objperslab> <pagesperslab>"); 4867 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4868 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4869 seq_putc(m, '\n'); 4870} 4871 4872static void *s_start(struct seq_file *m, loff_t *pos) 4873{ 4874 loff_t n = *pos; 4875 4876 down_read(&slub_lock); 4877 if (!n) 4878 print_slabinfo_header(m); 4879 4880 return seq_list_start(&slab_caches, *pos); 4881} 4882 4883static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4884{ 4885 return seq_list_next(p, &slab_caches, pos); 4886} 4887 4888static void s_stop(struct seq_file *m, void *p) 4889{ 4890 up_read(&slub_lock); 4891} 4892 4893static int s_show(struct seq_file *m, void *p) 4894{ 4895 unsigned long nr_partials = 0; 4896 unsigned long nr_slabs = 0; 4897 unsigned long nr_inuse = 0; 4898 unsigned long nr_objs = 0; 4899 unsigned long nr_free = 0; 4900 struct kmem_cache *s; 4901 int node; 4902 4903 s = list_entry(p, struct kmem_cache, list); 4904 4905 for_each_online_node(node) { 4906 struct kmem_cache_node *n = get_node(s, node); 4907 4908 if (!n) 4909 continue; 4910 4911 nr_partials += n->nr_partial; 4912 nr_slabs += atomic_long_read(&n->nr_slabs); 4913 nr_objs += atomic_long_read(&n->total_objects); 4914 nr_free += count_partial(n, count_free); 4915 } 4916 4917 nr_inuse = nr_objs - nr_free; 4918 4919 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4920 nr_objs, s->size, oo_objects(s->oo), 4921 (1 << oo_order(s->oo))); 4922 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4923 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4924 0UL); 4925 seq_putc(m, '\n'); 4926 return 0; 4927} 4928 4929static const struct seq_operations slabinfo_op = { 4930 .start = s_start, 4931 .next = s_next, 4932 .stop = s_stop, 4933 .show = s_show, 4934}; 4935 4936static int slabinfo_open(struct inode *inode, struct file *file) 4937{ 4938 return seq_open(file, &slabinfo_op); 4939} 4940 4941static const struct file_operations proc_slabinfo_operations = { 4942 .open = slabinfo_open, 4943 .read = seq_read, 4944 .llseek = seq_lseek, 4945 .release = seq_release, 4946}; 4947 4948static int __init slab_proc_init(void) 4949{ 4950 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations); 4951 return 0; 4952} 4953module_init(slab_proc_init); 4954#endif /* CONFIG_SLABINFO */ 4955