1/* Modified by Broadcom Corp. Portions Copyright (c) Broadcom Corp, 2012. */ 2/* 3 * SLUB: A slab allocator that limits cache line use instead of queuing 4 * objects in per cpu and per node lists. 5 * 6 * The allocator synchronizes using per slab locks and only 7 * uses a centralized lock to manage a pool of partial slabs. 8 * 9 * (C) 2007 SGI, Christoph Lameter 10 */ 11 12#include <linux/mm.h> 13#include <linux/swap.h> /* struct reclaim_state */ 14#include <linux/module.h> 15#include <linux/bit_spinlock.h> 16#include <linux/interrupt.h> 17#include <linux/bitops.h> 18#include <linux/slab.h> 19#include <linux/proc_fs.h> 20#include <linux/seq_file.h> 21#include <linux/kmemcheck.h> 22#include <linux/cpu.h> 23#include <linux/cpuset.h> 24#include <linux/mempolicy.h> 25#include <linux/ctype.h> 26#include <linux/debugobjects.h> 27#include <linux/kallsyms.h> 28#include <linux/memory.h> 29#include <linux/math64.h> 30#include <linux/fault-inject.h> 31 32#include <typedefs.h> 33#include <bcmdefs.h> 34 35/* 36 * Lock order: 37 * 1. slab_lock(page) 38 * 2. slab->list_lock 39 * 40 * The slab_lock protects operations on the object of a particular 41 * slab and its metadata in the page struct. If the slab lock 42 * has been taken then no allocations nor frees can be performed 43 * on the objects in the slab nor can the slab be added or removed 44 * from the partial or full lists since this would mean modifying 45 * the page_struct of the slab. 46 * 47 * The list_lock protects the partial and full list on each node and 48 * the partial slab counter. If taken then no new slabs may be added or 49 * removed from the lists nor make the number of partial slabs be modified. 50 * (Note that the total number of slabs is an atomic value that may be 51 * modified without taking the list lock). 52 * 53 * The list_lock is a centralized lock and thus we avoid taking it as 54 * much as possible. As long as SLUB does not have to handle partial 55 * slabs, operations can continue without any centralized lock. F.e. 56 * allocating a long series of objects that fill up slabs does not require 57 * the list lock. 58 * 59 * The lock order is sometimes inverted when we are trying to get a slab 60 * off a list. We take the list_lock and then look for a page on the list 61 * to use. While we do that objects in the slabs may be freed. We can 62 * only operate on the slab if we have also taken the slab_lock. So we use 63 * a slab_trylock() on the slab. If trylock was successful then no frees 64 * can occur anymore and we can use the slab for allocations etc. If the 65 * slab_trylock() does not succeed then frees are in progress in the slab and 66 * we must stay away from it for a while since we may cause a bouncing 67 * cacheline if we try to acquire the lock. So go onto the next slab. 68 * If all pages are busy then we may allocate a new slab instead of reusing 69 * a partial slab. A new slab has noone operating on it and thus there is 70 * no danger of cacheline contention. 71 * 72 * Interrupts are disabled during allocation and deallocation in order to 73 * make the slab allocator safe to use in the context of an irq. In addition 74 * interrupts are disabled to ensure that the processor does not change 75 * while handling per_cpu slabs, due to kernel preemption. 76 * 77 * SLUB assigns one slab for allocation to each processor. 78 * Allocations only occur from these slabs called cpu slabs. 79 * 80 * Slabs with free elements are kept on a partial list and during regular 81 * operations no list for full slabs is used. If an object in a full slab is 82 * freed then the slab will show up again on the partial lists. 83 * We track full slabs for debugging purposes though because otherwise we 84 * cannot scan all objects. 85 * 86 * Slabs are freed when they become empty. Teardown and setup is 87 * minimal so we rely on the page allocators per cpu caches for 88 * fast frees and allocs. 89 * 90 * Overloading of page flags that are otherwise used for LRU management. 91 * 92 * PageActive The slab is frozen and exempt from list processing. 93 * This means that the slab is dedicated to a purpose 94 * such as satisfying allocations for a specific 95 * processor. Objects may be freed in the slab while 96 * it is frozen but slab_free will then skip the usual 97 * list operations. It is up to the processor holding 98 * the slab to integrate the slab into the slab lists 99 * when the slab is no longer needed. 100 * 101 * One use of this flag is to mark slabs that are 102 * used for allocations. Then such a slab becomes a cpu 103 * slab. The cpu slab may be equipped with an additional 104 * freelist that allows lockless access to 105 * free objects in addition to the regular freelist 106 * that requires the slab lock. 107 * 108 * PageError Slab requires special handling due to debug 109 * options set. This moves slab handling out of 110 * the fast path and disables lockless freelists. 111 */ 112 113#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 114 SLAB_TRACE | SLAB_DEBUG_FREE) 115 116static inline int kmem_cache_debug(struct kmem_cache *s) 117{ 118#ifdef CONFIG_SLUB_DEBUG 119 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 120#else 121 return 0; 122#endif 123} 124 125/* 126 * Issues still to be resolved: 127 * 128 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 129 * 130 * - Variable sizing of the per node arrays 131 */ 132 133/* Enable to test recovery from slab corruption on boot */ 134#undef SLUB_RESILIENCY_TEST 135 136/* 137 * Mininum number of partial slabs. These will be left on the partial 138 * lists even if they are empty. kmem_cache_shrink may reclaim them. 139 */ 140#define MIN_PARTIAL 5 141 142/* 143 * Maximum number of desirable partial slabs. 144 * The existence of more partial slabs makes kmem_cache_shrink 145 * sort the partial list by the number of objects in the. 146 */ 147#define MAX_PARTIAL 10 148 149#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 150 SLAB_POISON | SLAB_STORE_USER) 151 152/* 153 * Debugging flags that require metadata to be stored in the slab. These get 154 * disabled when slub_debug=O is used and a cache's min order increases with 155 * metadata. 156 */ 157#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 158 159/* 160 * Set of flags that will prevent slab merging 161 */ 162#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 163 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ 164 SLAB_FAILSLAB) 165 166#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 167 SLAB_CACHE_DMA | SLAB_NOTRACK) 168 169#define OO_SHIFT 16 170#define OO_MASK ((1 << OO_SHIFT) - 1) 171#define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */ 172 173/* Internal SLUB flags */ 174#define __OBJECT_POISON 0x80000000UL /* Poison object */ 175#define __SYSFS_ADD_DEFERRED 0x40000000UL /* Not yet visible via sysfs */ 176 177static int kmem_size = sizeof(struct kmem_cache); 178 179#ifdef CONFIG_SMP 180static struct notifier_block slab_notifier; 181#endif 182 183static enum { 184 DOWN, /* No slab functionality available */ 185 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 186 UP, /* Everything works but does not show up in sysfs */ 187 SYSFS /* Sysfs up */ 188} slab_state = DOWN; 189 190/* A list of all slab caches on the system */ 191static DECLARE_RWSEM(slub_lock); 192static LIST_HEAD(slab_caches); 193 194/* 195 * Tracking user of a slab. 196 */ 197struct track { 198 unsigned long addr; /* Called from address */ 199 int cpu; /* Was running on cpu */ 200 int pid; /* Pid context */ 201 unsigned long when; /* When did the operation occur */ 202}; 203 204enum track_item { TRACK_ALLOC, TRACK_FREE }; 205 206#ifdef CONFIG_SLUB_DEBUG 207static int sysfs_slab_add(struct kmem_cache *); 208static int sysfs_slab_alias(struct kmem_cache *, const char *); 209static void sysfs_slab_remove(struct kmem_cache *); 210 211#else 212static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 213static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 214 { return 0; } 215static inline void sysfs_slab_remove(struct kmem_cache *s) 216{ 217 kfree(s); 218} 219 220#endif 221 222static inline void stat(struct kmem_cache *s, enum stat_item si) 223{ 224#ifdef CONFIG_SLUB_STATS 225 __this_cpu_inc(s->cpu_slab->stat[si]); 226#endif 227} 228 229/******************************************************************** 230 * Core slab cache functions 231 *******************************************************************/ 232 233int slab_is_available(void) 234{ 235 return slab_state >= UP; 236} 237 238static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 239{ 240#ifdef CONFIG_NUMA 241 return s->node[node]; 242#else 243 return &s->local_node; 244#endif 245} 246 247/* Verify that a pointer has an address that is valid within a slab page */ 248static inline int check_valid_pointer(struct kmem_cache *s, 249 struct page *page, const void *object) 250{ 251 void *base; 252 253 if (!object) 254 return 1; 255 256 base = page_address(page); 257 if (object < base || object >= base + page->objects * s->size || 258 (object - base) % s->size) { 259 return 0; 260 } 261 262 return 1; 263} 264 265static inline void *get_freepointer(struct kmem_cache *s, void *object) 266{ 267 return *(void **)(object + s->offset); 268} 269 270static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 271{ 272 *(void **)(object + s->offset) = fp; 273} 274 275/* Loop over all objects in a slab */ 276#define for_each_object(__p, __s, __addr, __objects) \ 277 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 278 __p += (__s)->size) 279 280/* Scan freelist */ 281#define for_each_free_object(__p, __s, __free) \ 282 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 283 284/* Determine object index from a given position */ 285static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 286{ 287 return (p - addr) / s->size; 288} 289 290static inline struct kmem_cache_order_objects oo_make(int order, 291 unsigned long size) 292{ 293 struct kmem_cache_order_objects x = { 294 (order << OO_SHIFT) + (PAGE_SIZE << order) / size 295 }; 296 297 return x; 298} 299 300static inline int oo_order(struct kmem_cache_order_objects x) 301{ 302 return x.x >> OO_SHIFT; 303} 304 305static inline int oo_objects(struct kmem_cache_order_objects x) 306{ 307 return x.x & OO_MASK; 308} 309 310#ifdef CONFIG_SLUB_DEBUG 311/* 312 * Debug settings: 313 */ 314#ifdef CONFIG_SLUB_DEBUG_ON 315static int slub_debug = DEBUG_DEFAULT_FLAGS; 316#else 317static int slub_debug; 318#endif 319 320static char *slub_debug_slabs; 321static int disable_higher_order_debug; 322 323/* 324 * Object debugging 325 */ 326static void print_section(char *text, u8 *addr, unsigned int length) 327{ 328 int i, offset; 329 int newline = 1; 330 char ascii[17]; 331 332 ascii[16] = 0; 333 334 for (i = 0; i < length; i++) { 335 if (newline) { 336 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 337 newline = 0; 338 } 339 printk(KERN_CONT " %02x", addr[i]); 340 offset = i % 16; 341 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 342 if (offset == 15) { 343 printk(KERN_CONT " %s\n", ascii); 344 newline = 1; 345 } 346 } 347 if (!newline) { 348 i %= 16; 349 while (i < 16) { 350 printk(KERN_CONT " "); 351 ascii[i] = ' '; 352 i++; 353 } 354 printk(KERN_CONT " %s\n", ascii); 355 } 356} 357 358static struct track *get_track(struct kmem_cache *s, void *object, 359 enum track_item alloc) 360{ 361 struct track *p; 362 363 if (s->offset) 364 p = object + s->offset + sizeof(void *); 365 else 366 p = object + s->inuse; 367 368 return p + alloc; 369} 370 371static void set_track(struct kmem_cache *s, void *object, 372 enum track_item alloc, unsigned long addr) 373{ 374 struct track *p = get_track(s, object, alloc); 375 376 if (addr) { 377 p->addr = addr; 378 p->cpu = smp_processor_id(); 379 p->pid = current->pid; 380 p->when = jiffies; 381 } else 382 memset(p, 0, sizeof(struct track)); 383} 384 385static void init_tracking(struct kmem_cache *s, void *object) 386{ 387 if (!(s->flags & SLAB_STORE_USER)) 388 return; 389 390 set_track(s, object, TRACK_FREE, 0UL); 391 set_track(s, object, TRACK_ALLOC, 0UL); 392} 393 394static void print_track(const char *s, struct track *t) 395{ 396 if (!t->addr) 397 return; 398 399 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 400 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 401} 402 403static void print_tracking(struct kmem_cache *s, void *object) 404{ 405 if (!(s->flags & SLAB_STORE_USER)) 406 return; 407 408 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 409 print_track("Freed", get_track(s, object, TRACK_FREE)); 410} 411 412static void print_page_info(struct page *page) 413{ 414 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 415 page, page->objects, page->inuse, page->freelist, page->flags); 416 417} 418 419static void slab_bug(struct kmem_cache *s, char *fmt, ...) 420{ 421 va_list args; 422 char buf[100]; 423 424 va_start(args, fmt); 425 vsnprintf(buf, sizeof(buf), fmt, args); 426 va_end(args); 427 printk(KERN_ERR "========================================" 428 "=====================================\n"); 429 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 430 printk(KERN_ERR "----------------------------------------" 431 "-------------------------------------\n\n"); 432} 433 434static void slab_fix(struct kmem_cache *s, char *fmt, ...) 435{ 436 va_list args; 437 char buf[100]; 438 439 va_start(args, fmt); 440 vsnprintf(buf, sizeof(buf), fmt, args); 441 va_end(args); 442 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 443} 444 445static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 446{ 447 unsigned int off; /* Offset of last byte */ 448 u8 *addr = page_address(page); 449 450 print_tracking(s, p); 451 452 print_page_info(page); 453 454 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 455 p, p - addr, get_freepointer(s, p)); 456 457 if (p > addr + 16) 458 print_section("Bytes b4", p - 16, 16); 459 460 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE)); 461 462 if (s->flags & SLAB_RED_ZONE) 463 print_section("Redzone", p + s->objsize, 464 s->inuse - s->objsize); 465 466 if (s->offset) 467 off = s->offset + sizeof(void *); 468 else 469 off = s->inuse; 470 471 if (s->flags & SLAB_STORE_USER) 472 off += 2 * sizeof(struct track); 473 474 if (off != s->size) 475 /* Beginning of the filler is the free pointer */ 476 print_section("Padding", p + off, s->size - off); 477 478 dump_stack(); 479} 480 481static void object_err(struct kmem_cache *s, struct page *page, 482 u8 *object, char *reason) 483{ 484 slab_bug(s, "%s", reason); 485 print_trailer(s, page, object); 486} 487 488static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 489{ 490 va_list args; 491 char buf[100]; 492 493 va_start(args, fmt); 494 vsnprintf(buf, sizeof(buf), fmt, args); 495 va_end(args); 496 slab_bug(s, "%s", buf); 497 print_page_info(page); 498 dump_stack(); 499} 500 501static void init_object(struct kmem_cache *s, void *object, int active) 502{ 503 u8 *p = object; 504 505 if (s->flags & __OBJECT_POISON) { 506 memset(p, POISON_FREE, s->objsize - 1); 507 p[s->objsize - 1] = POISON_END; 508 } 509 510 if (s->flags & SLAB_RED_ZONE) 511 memset(p + s->objsize, 512 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 513 s->inuse - s->objsize); 514} 515 516static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 517{ 518 while (bytes) { 519 if (*start != (u8)value) 520 return start; 521 start++; 522 bytes--; 523 } 524 return NULL; 525} 526 527static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 528 void *from, void *to) 529{ 530 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 531 memset(from, data, to - from); 532} 533 534static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 535 u8 *object, char *what, 536 u8 *start, unsigned int value, unsigned int bytes) 537{ 538 u8 *fault; 539 u8 *end; 540 541 fault = check_bytes(start, value, bytes); 542 if (!fault) 543 return 1; 544 545 end = start + bytes; 546 while (end > fault && end[-1] == value) 547 end--; 548 549 slab_bug(s, "%s overwritten", what); 550 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 551 fault, end - 1, fault[0], value); 552 print_trailer(s, page, object); 553 554 restore_bytes(s, what, value, fault, end); 555 return 0; 556} 557 558/* 559 * Object layout: 560 * 561 * object address 562 * Bytes of the object to be managed. 563 * If the freepointer may overlay the object then the free 564 * pointer is the first word of the object. 565 * 566 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 567 * 0xa5 (POISON_END) 568 * 569 * object + s->objsize 570 * Padding to reach word boundary. This is also used for Redzoning. 571 * Padding is extended by another word if Redzoning is enabled and 572 * objsize == inuse. 573 * 574 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 575 * 0xcc (RED_ACTIVE) for objects in use. 576 * 577 * object + s->inuse 578 * Meta data starts here. 579 * 580 * A. Free pointer (if we cannot overwrite object on free) 581 * B. Tracking data for SLAB_STORE_USER 582 * C. Padding to reach required alignment boundary or at mininum 583 * one word if debugging is on to be able to detect writes 584 * before the word boundary. 585 * 586 * Padding is done using 0x5a (POISON_INUSE) 587 * 588 * object + s->size 589 * Nothing is used beyond s->size. 590 * 591 * If slabcaches are merged then the objsize and inuse boundaries are mostly 592 * ignored. And therefore no slab options that rely on these boundaries 593 * may be used with merged slabcaches. 594 */ 595 596static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 597{ 598 unsigned long off = s->inuse; /* The end of info */ 599 600 if (s->offset) 601 /* Freepointer is placed after the object. */ 602 off += sizeof(void *); 603 604 if (s->flags & SLAB_STORE_USER) 605 /* We also have user information there */ 606 off += 2 * sizeof(struct track); 607 608 if (s->size == off) 609 return 1; 610 611 return check_bytes_and_report(s, page, p, "Object padding", 612 p + off, POISON_INUSE, s->size - off); 613} 614 615/* Check the pad bytes at the end of a slab page */ 616static int slab_pad_check(struct kmem_cache *s, struct page *page) 617{ 618 u8 *start; 619 u8 *fault; 620 u8 *end; 621 int length; 622 int remainder; 623 624 if (!(s->flags & SLAB_POISON)) 625 return 1; 626 627 start = page_address(page); 628 length = (PAGE_SIZE << compound_order(page)); 629 end = start + length; 630 remainder = length % s->size; 631 if (!remainder) 632 return 1; 633 634 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 635 if (!fault) 636 return 1; 637 while (end > fault && end[-1] == POISON_INUSE) 638 end--; 639 640 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 641 print_section("Padding", end - remainder, remainder); 642 643 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 644 return 0; 645} 646 647static int check_object(struct kmem_cache *s, struct page *page, 648 void *object, int active) 649{ 650 u8 *p = object; 651 u8 *endobject = object + s->objsize; 652 653 if (s->flags & SLAB_RED_ZONE) { 654 unsigned int red = 655 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 656 657 if (!check_bytes_and_report(s, page, object, "Redzone", 658 endobject, red, s->inuse - s->objsize)) 659 return 0; 660 } else { 661 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 662 check_bytes_and_report(s, page, p, "Alignment padding", 663 endobject, POISON_INUSE, s->inuse - s->objsize); 664 } 665 } 666 667 if (s->flags & SLAB_POISON) { 668 if (!active && (s->flags & __OBJECT_POISON) && 669 (!check_bytes_and_report(s, page, p, "Poison", p, 670 POISON_FREE, s->objsize - 1) || 671 !check_bytes_and_report(s, page, p, "Poison", 672 p + s->objsize - 1, POISON_END, 1))) 673 return 0; 674 /* 675 * check_pad_bytes cleans up on its own. 676 */ 677 check_pad_bytes(s, page, p); 678 } 679 680 if (!s->offset && active) 681 /* 682 * Object and freepointer overlap. Cannot check 683 * freepointer while object is allocated. 684 */ 685 return 1; 686 687 /* Check free pointer validity */ 688 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 689 object_err(s, page, p, "Freepointer corrupt"); 690 /* 691 * No choice but to zap it and thus lose the remainder 692 * of the free objects in this slab. May cause 693 * another error because the object count is now wrong. 694 */ 695 set_freepointer(s, p, NULL); 696 return 0; 697 } 698 return 1; 699} 700 701static int check_slab(struct kmem_cache *s, struct page *page) 702{ 703 int maxobj; 704 705 VM_BUG_ON(!irqs_disabled()); 706 707 if (!PageSlab(page)) { 708 slab_err(s, page, "Not a valid slab page"); 709 return 0; 710 } 711 712 maxobj = (PAGE_SIZE << compound_order(page)) / s->size; 713 if (page->objects > maxobj) { 714 slab_err(s, page, "objects %u > max %u", 715 s->name, page->objects, maxobj); 716 return 0; 717 } 718 if (page->inuse > page->objects) { 719 slab_err(s, page, "inuse %u > max %u", 720 s->name, page->inuse, page->objects); 721 return 0; 722 } 723 /* Slab_pad_check fixes things up after itself */ 724 slab_pad_check(s, page); 725 return 1; 726} 727 728/* 729 * Determine if a certain object on a page is on the freelist. Must hold the 730 * slab lock to guarantee that the chains are in a consistent state. 731 */ 732static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 733{ 734 int nr = 0; 735 void *fp = page->freelist; 736 void *object = NULL; 737 unsigned long max_objects; 738 739 while (fp && nr <= page->objects) { 740 if (fp == search) 741 return 1; 742 if (!check_valid_pointer(s, page, fp)) { 743 if (object) { 744 object_err(s, page, object, 745 "Freechain corrupt"); 746 set_freepointer(s, object, NULL); 747 break; 748 } else { 749 slab_err(s, page, "Freepointer corrupt"); 750 page->freelist = NULL; 751 page->inuse = page->objects; 752 slab_fix(s, "Freelist cleared"); 753 return 0; 754 } 755 break; 756 } 757 object = fp; 758 fp = get_freepointer(s, object); 759 nr++; 760 } 761 762 max_objects = (PAGE_SIZE << compound_order(page)) / s->size; 763 if (max_objects > MAX_OBJS_PER_PAGE) 764 max_objects = MAX_OBJS_PER_PAGE; 765 766 if (page->objects != max_objects) { 767 slab_err(s, page, "Wrong number of objects. Found %d but " 768 "should be %d", page->objects, max_objects); 769 page->objects = max_objects; 770 slab_fix(s, "Number of objects adjusted."); 771 } 772 if (page->inuse != page->objects - nr) { 773 slab_err(s, page, "Wrong object count. Counter is %d but " 774 "counted were %d", page->inuse, page->objects - nr); 775 page->inuse = page->objects - nr; 776 slab_fix(s, "Object count adjusted."); 777 } 778 return search == NULL; 779} 780 781static void trace(struct kmem_cache *s, struct page *page, void *object, 782 int alloc) 783{ 784 if (s->flags & SLAB_TRACE) { 785 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 786 s->name, 787 alloc ? "alloc" : "free", 788 object, page->inuse, 789 page->freelist); 790 791 if (!alloc) 792 print_section("Object", (void *)object, s->objsize); 793 794 dump_stack(); 795 } 796} 797 798/* 799 * Tracking of fully allocated slabs for debugging purposes. 800 */ 801static void add_full(struct kmem_cache_node *n, struct page *page) 802{ 803 spin_lock(&n->list_lock); 804 list_add(&page->lru, &n->full); 805 spin_unlock(&n->list_lock); 806} 807 808static void remove_full(struct kmem_cache *s, struct page *page) 809{ 810 struct kmem_cache_node *n; 811 812 if (!(s->flags & SLAB_STORE_USER)) 813 return; 814 815 n = get_node(s, page_to_nid(page)); 816 817 spin_lock(&n->list_lock); 818 list_del(&page->lru); 819 spin_unlock(&n->list_lock); 820} 821 822/* Tracking of the number of slabs for debugging purposes */ 823static inline unsigned long slabs_node(struct kmem_cache *s, int node) 824{ 825 struct kmem_cache_node *n = get_node(s, node); 826 827 return atomic_long_read(&n->nr_slabs); 828} 829 830static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 831{ 832 return atomic_long_read(&n->nr_slabs); 833} 834 835static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 836{ 837 struct kmem_cache_node *n = get_node(s, node); 838 839 /* 840 * May be called early in order to allocate a slab for the 841 * kmem_cache_node structure. Solve the chicken-egg 842 * dilemma by deferring the increment of the count during 843 * bootstrap (see early_kmem_cache_node_alloc). 844 */ 845 if (!NUMA_BUILD || n) { 846 atomic_long_inc(&n->nr_slabs); 847 atomic_long_add(objects, &n->total_objects); 848 } 849} 850static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 851{ 852 struct kmem_cache_node *n = get_node(s, node); 853 854 atomic_long_dec(&n->nr_slabs); 855 atomic_long_sub(objects, &n->total_objects); 856} 857 858/* Object debug checks for alloc/free paths */ 859static void setup_object_debug(struct kmem_cache *s, struct page *page, 860 void *object) 861{ 862 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 863 return; 864 865 init_object(s, object, 0); 866 init_tracking(s, object); 867} 868 869static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 870 void *object, unsigned long addr) 871{ 872 if (!check_slab(s, page)) 873 goto bad; 874 875 if (!on_freelist(s, page, object)) { 876 object_err(s, page, object, "Object already allocated"); 877 goto bad; 878 } 879 880 if (!check_valid_pointer(s, page, object)) { 881 object_err(s, page, object, "Freelist Pointer check fails"); 882 goto bad; 883 } 884 885 if (!check_object(s, page, object, 0)) 886 goto bad; 887 888 /* Success perform special debug activities for allocs */ 889 if (s->flags & SLAB_STORE_USER) 890 set_track(s, object, TRACK_ALLOC, addr); 891 trace(s, page, object, 1); 892 init_object(s, object, 1); 893 return 1; 894 895bad: 896 if (PageSlab(page)) { 897 /* 898 * If this is a slab page then lets do the best we can 899 * to avoid issues in the future. Marking all objects 900 * as used avoids touching the remaining objects. 901 */ 902 slab_fix(s, "Marking all objects used"); 903 page->inuse = page->objects; 904 page->freelist = NULL; 905 } 906 return 0; 907} 908 909static int free_debug_processing(struct kmem_cache *s, struct page *page, 910 void *object, unsigned long addr) 911{ 912 if (!check_slab(s, page)) 913 goto fail; 914 915 if (!check_valid_pointer(s, page, object)) { 916 slab_err(s, page, "Invalid object pointer 0x%p", object); 917 goto fail; 918 } 919 920 if (on_freelist(s, page, object)) { 921 object_err(s, page, object, "Object already free"); 922 goto fail; 923 } 924 925 if (!check_object(s, page, object, 1)) 926 return 0; 927 928 if (unlikely(s != page->slab)) { 929 if (!PageSlab(page)) { 930 slab_err(s, page, "Attempt to free object(0x%p) " 931 "outside of slab", object); 932 } else if (!page->slab) { 933 printk(KERN_ERR 934 "SLUB <none>: no slab for object 0x%p.\n", 935 object); 936 dump_stack(); 937 } else 938 object_err(s, page, object, 939 "page slab pointer corrupt."); 940 goto fail; 941 } 942 943 /* Special debug activities for freeing objects */ 944 if (!PageSlubFrozen(page) && !page->freelist) 945 remove_full(s, page); 946 if (s->flags & SLAB_STORE_USER) 947 set_track(s, object, TRACK_FREE, addr); 948 trace(s, page, object, 0); 949 init_object(s, object, 0); 950 return 1; 951 952fail: 953 slab_fix(s, "Object at 0x%p not freed", object); 954 return 0; 955} 956 957static int __init setup_slub_debug(char *str) 958{ 959 slub_debug = DEBUG_DEFAULT_FLAGS; 960 if (*str++ != '=' || !*str) 961 /* 962 * No options specified. Switch on full debugging. 963 */ 964 goto out; 965 966 if (*str == ',') 967 /* 968 * No options but restriction on slabs. This means full 969 * debugging for slabs matching a pattern. 970 */ 971 goto check_slabs; 972 973 if (tolower(*str) == 'o') { 974 /* 975 * Avoid enabling debugging on caches if its minimum order 976 * would increase as a result. 977 */ 978 disable_higher_order_debug = 1; 979 goto out; 980 } 981 982 slub_debug = 0; 983 if (*str == '-') 984 /* 985 * Switch off all debugging measures. 986 */ 987 goto out; 988 989 /* 990 * Determine which debug features should be switched on 991 */ 992 for (; *str && *str != ','; str++) { 993 switch (tolower(*str)) { 994 case 'f': 995 slub_debug |= SLAB_DEBUG_FREE; 996 break; 997 case 'z': 998 slub_debug |= SLAB_RED_ZONE; 999 break; 1000 case 'p': 1001 slub_debug |= SLAB_POISON; 1002 break; 1003 case 'u': 1004 slub_debug |= SLAB_STORE_USER; 1005 break; 1006 case 't': 1007 slub_debug |= SLAB_TRACE; 1008 break; 1009 case 'a': 1010 slub_debug |= SLAB_FAILSLAB; 1011 break; 1012 default: 1013 printk(KERN_ERR "slub_debug option '%c' " 1014 "unknown. skipped\n", *str); 1015 } 1016 } 1017 1018check_slabs: 1019 if (*str == ',') 1020 slub_debug_slabs = str + 1; 1021out: 1022 return 1; 1023} 1024 1025__setup("slub_debug", setup_slub_debug); 1026 1027static unsigned long kmem_cache_flags(unsigned long objsize, 1028 unsigned long flags, const char *name, 1029 void (*ctor)(void *)) 1030{ 1031 /* 1032 * Enable debugging if selected on the kernel commandline. 1033 */ 1034 if (slub_debug && (!slub_debug_slabs || 1035 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) 1036 flags |= slub_debug; 1037 1038 return flags; 1039} 1040#else 1041static inline void setup_object_debug(struct kmem_cache *s, 1042 struct page *page, void *object) {} 1043 1044static inline int alloc_debug_processing(struct kmem_cache *s, 1045 struct page *page, void *object, unsigned long addr) { return 0; } 1046 1047static inline int free_debug_processing(struct kmem_cache *s, 1048 struct page *page, void *object, unsigned long addr) { return 0; } 1049 1050static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1051 { return 1; } 1052static inline int check_object(struct kmem_cache *s, struct page *page, 1053 void *object, int active) { return 1; } 1054static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1055static inline unsigned long kmem_cache_flags(unsigned long objsize, 1056 unsigned long flags, const char *name, 1057 void (*ctor)(void *)) 1058{ 1059 return flags; 1060} 1061#define slub_debug 0 1062 1063#define disable_higher_order_debug 0 1064 1065static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1066 { return 0; } 1067static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1068 { return 0; } 1069static inline void inc_slabs_node(struct kmem_cache *s, int node, 1070 int objects) {} 1071static inline void dec_slabs_node(struct kmem_cache *s, int node, 1072 int objects) {} 1073#endif 1074 1075/* 1076 * Slab allocation and freeing 1077 */ 1078static inline struct page *alloc_slab_page(gfp_t flags, int node, 1079 struct kmem_cache_order_objects oo) 1080{ 1081 int order = oo_order(oo); 1082 1083 flags |= __GFP_NOTRACK; 1084 1085 if (node == NUMA_NO_NODE) 1086 return alloc_pages(flags, order); 1087 else 1088 return alloc_pages_exact_node(node, flags, order); 1089} 1090 1091static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1092{ 1093 struct page *page; 1094 struct kmem_cache_order_objects oo = s->oo; 1095 gfp_t alloc_gfp; 1096 1097 flags |= s->allocflags; 1098 1099 /* 1100 * Let the initial higher-order allocation fail under memory pressure 1101 * so we fall-back to the minimum order allocation. 1102 */ 1103 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1104 1105 page = alloc_slab_page(alloc_gfp, node, oo); 1106 if (unlikely(!page)) { 1107 oo = s->min; 1108 /* 1109 * Allocation may have failed due to fragmentation. 1110 * Try a lower order alloc if possible 1111 */ 1112 page = alloc_slab_page(flags, node, oo); 1113 if (!page) 1114 return NULL; 1115 1116 stat(s, ORDER_FALLBACK); 1117 } 1118 1119 if (kmemcheck_enabled 1120 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1121 int pages = 1 << oo_order(oo); 1122 1123 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); 1124 1125 /* 1126 * Objects from caches that have a constructor don't get 1127 * cleared when they're allocated, so we need to do it here. 1128 */ 1129 if (s->ctor) 1130 kmemcheck_mark_uninitialized_pages(page, pages); 1131 else 1132 kmemcheck_mark_unallocated_pages(page, pages); 1133 } 1134 1135 page->objects = oo_objects(oo); 1136 mod_zone_page_state(page_zone(page), 1137 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1138 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1139 1 << oo_order(oo)); 1140 1141 return page; 1142} 1143 1144static void setup_object(struct kmem_cache *s, struct page *page, 1145 void *object) 1146{ 1147 setup_object_debug(s, page, object); 1148 if (unlikely(s->ctor)) 1149 s->ctor(object); 1150} 1151 1152static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1153{ 1154 struct page *page; 1155 void *start; 1156 void *last; 1157 void *p; 1158 1159 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1160 1161 page = allocate_slab(s, 1162 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1163 if (!page) 1164 goto out; 1165 1166 inc_slabs_node(s, page_to_nid(page), page->objects); 1167 page->slab = s; 1168 page->flags |= 1 << PG_slab; 1169 1170 start = page_address(page); 1171 1172 if (unlikely(s->flags & SLAB_POISON)) 1173 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1174 1175 last = start; 1176 for_each_object(p, s, start, page->objects) { 1177 setup_object(s, page, last); 1178 set_freepointer(s, last, p); 1179 last = p; 1180 } 1181 setup_object(s, page, last); 1182 set_freepointer(s, last, NULL); 1183 1184 page->freelist = start; 1185 page->inuse = 0; 1186out: 1187 return page; 1188} 1189 1190static void __free_slab(struct kmem_cache *s, struct page *page) 1191{ 1192 int order = compound_order(page); 1193 int pages = 1 << order; 1194 1195 if (kmem_cache_debug(s)) { 1196 void *p; 1197 1198 slab_pad_check(s, page); 1199 for_each_object(p, s, page_address(page), 1200 page->objects) 1201 check_object(s, page, p, 0); 1202 } 1203 1204 kmemcheck_free_shadow(page, compound_order(page)); 1205 1206 mod_zone_page_state(page_zone(page), 1207 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1208 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1209 -pages); 1210 1211 __ClearPageSlab(page); 1212 reset_page_mapcount(page); 1213 if (current->reclaim_state) 1214 current->reclaim_state->reclaimed_slab += pages; 1215 __free_pages(page, order); 1216} 1217 1218static void rcu_free_slab(struct rcu_head *h) 1219{ 1220 struct page *page; 1221 1222 page = container_of((struct list_head *)h, struct page, lru); 1223 __free_slab(page->slab, page); 1224} 1225 1226static void free_slab(struct kmem_cache *s, struct page *page) 1227{ 1228 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1229 /* 1230 * RCU free overloads the RCU head over the LRU 1231 */ 1232 struct rcu_head *head = (void *)&page->lru; 1233 1234 call_rcu(head, rcu_free_slab); 1235 } else 1236 __free_slab(s, page); 1237} 1238 1239static void discard_slab(struct kmem_cache *s, struct page *page) 1240{ 1241 dec_slabs_node(s, page_to_nid(page), page->objects); 1242 free_slab(s, page); 1243} 1244 1245/* 1246 * Per slab locking using the pagelock 1247 */ 1248static __always_inline void slab_lock(struct page *page) 1249{ 1250 bit_spin_lock(PG_locked, &page->flags); 1251} 1252 1253static __always_inline void slab_unlock(struct page *page) 1254{ 1255 __bit_spin_unlock(PG_locked, &page->flags); 1256} 1257 1258static __always_inline int slab_trylock(struct page *page) 1259{ 1260 int rc = 1; 1261 1262 rc = bit_spin_trylock(PG_locked, &page->flags); 1263 return rc; 1264} 1265 1266/* 1267 * Management of partially allocated slabs 1268 */ 1269static void add_partial(struct kmem_cache_node *n, 1270 struct page *page, int tail) 1271{ 1272 spin_lock(&n->list_lock); 1273 n->nr_partial++; 1274 if (tail) 1275 list_add_tail(&page->lru, &n->partial); 1276 else 1277 list_add(&page->lru, &n->partial); 1278 spin_unlock(&n->list_lock); 1279} 1280 1281static void remove_partial(struct kmem_cache *s, struct page *page) 1282{ 1283 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1284 1285 spin_lock(&n->list_lock); 1286 list_del(&page->lru); 1287 n->nr_partial--; 1288 spin_unlock(&n->list_lock); 1289} 1290 1291/* 1292 * Lock slab and remove from the partial list. 1293 * 1294 * Must hold list_lock. 1295 */ 1296static inline int lock_and_freeze_slab(struct kmem_cache_node *n, 1297 struct page *page) 1298{ 1299 if (slab_trylock(page)) { 1300 list_del(&page->lru); 1301 n->nr_partial--; 1302 __SetPageSlubFrozen(page); 1303 return 1; 1304 } 1305 return 0; 1306} 1307 1308/* 1309 * Try to allocate a partial slab from a specific node. 1310 */ 1311static struct page *get_partial_node(struct kmem_cache_node *n) 1312{ 1313 struct page *page; 1314 1315 /* 1316 * Racy check. If we mistakenly see no partial slabs then we 1317 * just allocate an empty slab. If we mistakenly try to get a 1318 * partial slab and there is none available then get_partials() 1319 * will return NULL. 1320 */ 1321 if (!n || !n->nr_partial) 1322 return NULL; 1323 1324 spin_lock(&n->list_lock); 1325 list_for_each_entry(page, &n->partial, lru) 1326 if (lock_and_freeze_slab(n, page)) 1327 goto out; 1328 page = NULL; 1329out: 1330 spin_unlock(&n->list_lock); 1331 return page; 1332} 1333 1334/* 1335 * Get a page from somewhere. Search in increasing NUMA distances. 1336 */ 1337static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1338{ 1339#ifdef CONFIG_NUMA 1340 struct zonelist *zonelist; 1341 struct zoneref *z; 1342 struct zone *zone; 1343 enum zone_type high_zoneidx = gfp_zone(flags); 1344 struct page *page; 1345 1346 /* 1347 * The defrag ratio allows a configuration of the tradeoffs between 1348 * inter node defragmentation and node local allocations. A lower 1349 * defrag_ratio increases the tendency to do local allocations 1350 * instead of attempting to obtain partial slabs from other nodes. 1351 * 1352 * If the defrag_ratio is set to 0 then kmalloc() always 1353 * returns node local objects. If the ratio is higher then kmalloc() 1354 * may return off node objects because partial slabs are obtained 1355 * from other nodes and filled up. 1356 * 1357 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1358 * defrag_ratio = 1000) then every (well almost) allocation will 1359 * first attempt to defrag slab caches on other nodes. This means 1360 * scanning over all nodes to look for partial slabs which may be 1361 * expensive if we do it every time we are trying to find a slab 1362 * with available objects. 1363 */ 1364 if (!s->remote_node_defrag_ratio || 1365 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1366 return NULL; 1367 1368 get_mems_allowed(); 1369 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1370 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1371 struct kmem_cache_node *n; 1372 1373 n = get_node(s, zone_to_nid(zone)); 1374 1375 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1376 n->nr_partial > s->min_partial) { 1377 page = get_partial_node(n); 1378 if (page) { 1379 put_mems_allowed(); 1380 return page; 1381 } 1382 } 1383 } 1384 put_mems_allowed(); 1385#endif 1386 return NULL; 1387} 1388 1389/* 1390 * Get a partial page, lock it and return it. 1391 */ 1392static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1393{ 1394 struct page *page; 1395 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; 1396 1397 page = get_partial_node(get_node(s, searchnode)); 1398 if (page || node != -1) 1399 return page; 1400 1401 return get_any_partial(s, flags); 1402} 1403 1404/* 1405 * Move a page back to the lists. 1406 * 1407 * Must be called with the slab lock held. 1408 * 1409 * On exit the slab lock will have been dropped. 1410 */ 1411static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1412{ 1413 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1414 1415 __ClearPageSlubFrozen(page); 1416 if (page->inuse) { 1417 1418 if (page->freelist) { 1419 add_partial(n, page, tail); 1420 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1421 } else { 1422 stat(s, DEACTIVATE_FULL); 1423 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER)) 1424 add_full(n, page); 1425 } 1426 slab_unlock(page); 1427 } else { 1428 stat(s, DEACTIVATE_EMPTY); 1429 if (n->nr_partial < s->min_partial) { 1430 /* 1431 * Adding an empty slab to the partial slabs in order 1432 * to avoid page allocator overhead. This slab needs 1433 * to come after the other slabs with objects in 1434 * so that the others get filled first. That way the 1435 * size of the partial list stays small. 1436 * 1437 * kmem_cache_shrink can reclaim any empty slabs from 1438 * the partial list. 1439 */ 1440 add_partial(n, page, 1); 1441 slab_unlock(page); 1442 } else { 1443 slab_unlock(page); 1444 stat(s, FREE_SLAB); 1445 discard_slab(s, page); 1446 } 1447 } 1448} 1449 1450/* 1451 * Remove the cpu slab 1452 */ 1453static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1454{ 1455 struct page *page = c->page; 1456 int tail = 1; 1457 1458 if (page->freelist) 1459 stat(s, DEACTIVATE_REMOTE_FREES); 1460 /* 1461 * Merge cpu freelist into slab freelist. Typically we get here 1462 * because both freelists are empty. So this is unlikely 1463 * to occur. 1464 */ 1465 while (unlikely(c->freelist)) { 1466 void **object; 1467 1468 tail = 0; /* Hot objects. Put the slab first */ 1469 1470 /* Retrieve object from cpu_freelist */ 1471 object = c->freelist; 1472 c->freelist = get_freepointer(s, c->freelist); 1473 1474 /* And put onto the regular freelist */ 1475 set_freepointer(s, object, page->freelist); 1476 page->freelist = object; 1477 page->inuse--; 1478 } 1479 c->page = NULL; 1480 unfreeze_slab(s, page, tail); 1481} 1482 1483static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1484{ 1485 stat(s, CPUSLAB_FLUSH); 1486 slab_lock(c->page); 1487 deactivate_slab(s, c); 1488} 1489 1490/* 1491 * Flush cpu slab. 1492 * 1493 * Called from IPI handler with interrupts disabled. 1494 */ 1495static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1496{ 1497 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 1498 1499 if (likely(c && c->page)) 1500 flush_slab(s, c); 1501} 1502 1503static void flush_cpu_slab(void *d) 1504{ 1505 struct kmem_cache *s = d; 1506 1507 __flush_cpu_slab(s, smp_processor_id()); 1508} 1509 1510static void flush_all(struct kmem_cache *s) 1511{ 1512 on_each_cpu(flush_cpu_slab, s, 1); 1513} 1514 1515/* 1516 * Check if the objects in a per cpu structure fit numa 1517 * locality expectations. 1518 */ 1519static inline int node_match(struct kmem_cache_cpu *c, int node) 1520{ 1521#ifdef CONFIG_NUMA 1522 if (node != NUMA_NO_NODE && c->node != node) 1523 return 0; 1524#endif 1525 return 1; 1526} 1527 1528static int count_free(struct page *page) 1529{ 1530 return page->objects - page->inuse; 1531} 1532 1533static unsigned long count_partial(struct kmem_cache_node *n, 1534 int (*get_count)(struct page *)) 1535{ 1536 unsigned long flags; 1537 unsigned long x = 0; 1538 struct page *page; 1539 1540 spin_lock_irqsave(&n->list_lock, flags); 1541 list_for_each_entry(page, &n->partial, lru) 1542 x += get_count(page); 1543 spin_unlock_irqrestore(&n->list_lock, flags); 1544 return x; 1545} 1546 1547static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 1548{ 1549#ifdef CONFIG_SLUB_DEBUG 1550 return atomic_long_read(&n->total_objects); 1551#else 1552 return 0; 1553#endif 1554} 1555 1556static noinline void 1557slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 1558{ 1559 int node; 1560 1561 printk(KERN_WARNING 1562 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 1563 nid, gfpflags); 1564 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 1565 "default order: %d, min order: %d\n", s->name, s->objsize, 1566 s->size, oo_order(s->oo), oo_order(s->min)); 1567 1568 if (oo_order(s->min) > get_order(s->objsize)) 1569 printk(KERN_WARNING " %s debugging increased min order, use " 1570 "slub_debug=O to disable.\n", s->name); 1571 1572 for_each_online_node(node) { 1573 struct kmem_cache_node *n = get_node(s, node); 1574 unsigned long nr_slabs; 1575 unsigned long nr_objs; 1576 unsigned long nr_free; 1577 1578 if (!n) 1579 continue; 1580 1581 nr_free = count_partial(n, count_free); 1582 nr_slabs = node_nr_slabs(n); 1583 nr_objs = node_nr_objs(n); 1584 1585 printk(KERN_WARNING 1586 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 1587 node, nr_slabs, nr_objs, nr_free); 1588 } 1589} 1590 1591/* 1592 * Slow path. The lockless freelist is empty or we need to perform 1593 * debugging duties. 1594 * 1595 * Interrupts are disabled. 1596 * 1597 * Processing is still very fast if new objects have been freed to the 1598 * regular freelist. In that case we simply take over the regular freelist 1599 * as the lockless freelist and zap the regular freelist. 1600 * 1601 * If that is not working then we fall back to the partial lists. We take the 1602 * first element of the freelist as the object to allocate now and move the 1603 * rest of the freelist to the lockless freelist. 1604 * 1605 * And if we were unable to get a new slab from the partial slab lists then 1606 * we need to allocate a new slab. This is the slowest path since it involves 1607 * a call to the page allocator and the setup of a new slab. 1608 */ 1609static void * BCMFASTPATH_HOST __slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 1610 unsigned long addr, struct kmem_cache_cpu *c) 1611{ 1612 void **object; 1613 struct page *new; 1614 1615 /* We handle __GFP_ZERO in the caller */ 1616 gfpflags &= ~__GFP_ZERO; 1617 1618 if (!c->page) 1619 goto new_slab; 1620 1621 slab_lock(c->page); 1622 if (unlikely(!node_match(c, node))) 1623 goto another_slab; 1624 1625 stat(s, ALLOC_REFILL); 1626 1627load_freelist: 1628 object = c->page->freelist; 1629 if (unlikely(!object)) 1630 goto another_slab; 1631 if (kmem_cache_debug(s)) 1632 goto debug; 1633 1634 c->freelist = get_freepointer(s, object); 1635 c->page->inuse = c->page->objects; 1636 c->page->freelist = NULL; 1637 c->node = page_to_nid(c->page); 1638unlock_out: 1639 slab_unlock(c->page); 1640 stat(s, ALLOC_SLOWPATH); 1641 return object; 1642 1643another_slab: 1644 deactivate_slab(s, c); 1645 1646new_slab: 1647 new = get_partial(s, gfpflags, node); 1648 if (new) { 1649 c->page = new; 1650 stat(s, ALLOC_FROM_PARTIAL); 1651 goto load_freelist; 1652 } 1653 1654 if (gfpflags & __GFP_WAIT) 1655 local_irq_enable(); 1656 1657 new = new_slab(s, gfpflags, node); 1658 1659 if (gfpflags & __GFP_WAIT) 1660 local_irq_disable(); 1661 1662 if (new) { 1663 c = __this_cpu_ptr(s->cpu_slab); 1664 stat(s, ALLOC_SLAB); 1665 if (c->page) 1666 flush_slab(s, c); 1667 slab_lock(new); 1668 __SetPageSlubFrozen(new); 1669 c->page = new; 1670 goto load_freelist; 1671 } 1672 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 1673 slab_out_of_memory(s, gfpflags, node); 1674 return NULL; 1675debug: 1676 if (!alloc_debug_processing(s, c->page, object, addr)) 1677 goto another_slab; 1678 1679 c->page->inuse++; 1680 c->page->freelist = get_freepointer(s, object); 1681 c->node = -1; 1682 goto unlock_out; 1683} 1684 1685/* 1686 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1687 * have the fastpath folded into their functions. So no function call 1688 * overhead for requests that can be satisfied on the fastpath. 1689 * 1690 * The fastpath works by first checking if the lockless freelist can be used. 1691 * If not then __slab_alloc is called for slow processing. 1692 * 1693 * Otherwise we can simply pick the next object from the lockless free list. 1694 */ 1695static __always_inline void *slab_alloc(struct kmem_cache *s, 1696 gfp_t gfpflags, int node, unsigned long addr) 1697{ 1698 void **object; 1699 struct kmem_cache_cpu *c; 1700 unsigned long flags; 1701 1702 gfpflags &= gfp_allowed_mask; 1703 1704 lockdep_trace_alloc(gfpflags); 1705 might_sleep_if(gfpflags & __GFP_WAIT); 1706 1707 if (should_failslab(s->objsize, gfpflags, s->flags)) 1708 return NULL; 1709 1710 local_irq_save(flags); 1711 c = __this_cpu_ptr(s->cpu_slab); 1712 object = c->freelist; 1713 if (unlikely(!object || !node_match(c, node))) 1714 1715 object = __slab_alloc(s, gfpflags, node, addr, c); 1716 1717 else { 1718 c->freelist = get_freepointer(s, object); 1719 stat(s, ALLOC_FASTPATH); 1720 } 1721 local_irq_restore(flags); 1722 1723 if (unlikely(gfpflags & __GFP_ZERO) && object) 1724 memset(object, 0, s->objsize); 1725 1726 kmemcheck_slab_alloc(s, gfpflags, object, s->objsize); 1727 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, gfpflags); 1728 1729 return object; 1730} 1731 1732void * BCMFASTPATH_HOST kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1733{ 1734 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 1735 1736 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); 1737 1738 return ret; 1739} 1740EXPORT_SYMBOL(kmem_cache_alloc); 1741 1742#ifdef CONFIG_TRACING 1743void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags) 1744{ 1745 return slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 1746} 1747EXPORT_SYMBOL(kmem_cache_alloc_notrace); 1748#endif 1749 1750#ifdef CONFIG_NUMA 1751void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1752{ 1753 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 1754 1755 trace_kmem_cache_alloc_node(_RET_IP_, ret, 1756 s->objsize, s->size, gfpflags, node); 1757 1758 return ret; 1759} 1760EXPORT_SYMBOL(kmem_cache_alloc_node); 1761#endif 1762 1763#ifdef CONFIG_TRACING 1764void *kmem_cache_alloc_node_notrace(struct kmem_cache *s, 1765 gfp_t gfpflags, 1766 int node) 1767{ 1768 return slab_alloc(s, gfpflags, node, _RET_IP_); 1769} 1770EXPORT_SYMBOL(kmem_cache_alloc_node_notrace); 1771#endif 1772 1773/* 1774 * Slow patch handling. This may still be called frequently since objects 1775 * have a longer lifetime than the cpu slabs in most processing loads. 1776 * 1777 * So we still attempt to reduce cache line usage. Just take the slab 1778 * lock and free the item. If there is no additional partial page 1779 * handling required then we can return immediately. 1780 */ 1781static void BCMFASTPATH_HOST __slab_free(struct kmem_cache *s, struct page *page, 1782 void *x, unsigned long addr) 1783{ 1784 void *prior; 1785 void **object = (void *)x; 1786 1787 stat(s, FREE_SLOWPATH); 1788 slab_lock(page); 1789 1790 if (kmem_cache_debug(s)) 1791 goto debug; 1792 1793checks_ok: 1794 prior = page->freelist; 1795 set_freepointer(s, object, prior); 1796 page->freelist = object; 1797 page->inuse--; 1798 1799 if (unlikely(PageSlubFrozen(page))) { 1800 stat(s, FREE_FROZEN); 1801 goto out_unlock; 1802 } 1803 1804 if (unlikely(!page->inuse)) 1805 goto slab_empty; 1806 1807 /* 1808 * Objects left in the slab. If it was not on the partial list before 1809 * then add it. 1810 */ 1811 if (unlikely(!prior)) { 1812 add_partial(get_node(s, page_to_nid(page)), page, 1); 1813 stat(s, FREE_ADD_PARTIAL); 1814 } 1815 1816out_unlock: 1817 slab_unlock(page); 1818 return; 1819 1820slab_empty: 1821 if (prior) { 1822 /* 1823 * Slab still on the partial list. 1824 */ 1825 remove_partial(s, page); 1826 stat(s, FREE_REMOVE_PARTIAL); 1827 } 1828 slab_unlock(page); 1829 stat(s, FREE_SLAB); 1830 discard_slab(s, page); 1831 return; 1832 1833debug: 1834 if (!free_debug_processing(s, page, x, addr)) 1835 goto out_unlock; 1836 goto checks_ok; 1837} 1838 1839/* 1840 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1841 * can perform fastpath freeing without additional function calls. 1842 * 1843 * The fastpath is only possible if we are freeing to the current cpu slab 1844 * of this processor. This typically the case if we have just allocated 1845 * the item before. 1846 * 1847 * If fastpath is not possible then fall back to __slab_free where we deal 1848 * with all sorts of special processing. 1849 */ 1850static __always_inline void slab_free(struct kmem_cache *s, 1851 struct page *page, void *x, unsigned long addr) 1852{ 1853 void **object = (void *)x; 1854 struct kmem_cache_cpu *c; 1855 unsigned long flags; 1856 1857 kmemleak_free_recursive(x, s->flags); 1858 local_irq_save(flags); 1859 c = __this_cpu_ptr(s->cpu_slab); 1860 kmemcheck_slab_free(s, object, s->objsize); 1861 debug_check_no_locks_freed(object, s->objsize); 1862 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1863 debug_check_no_obj_freed(object, s->objsize); 1864 if (likely(page == c->page && c->node >= 0)) { 1865 set_freepointer(s, object, c->freelist); 1866 c->freelist = object; 1867 stat(s, FREE_FASTPATH); 1868 } else 1869 __slab_free(s, page, x, addr); 1870 1871 local_irq_restore(flags); 1872} 1873 1874void BCMFASTPATH_HOST kmem_cache_free(struct kmem_cache *s, void *x) 1875{ 1876 struct page *page; 1877 1878 page = virt_to_head_page(x); 1879 1880 slab_free(s, page, x, _RET_IP_); 1881 1882 trace_kmem_cache_free(_RET_IP_, x); 1883} 1884EXPORT_SYMBOL(kmem_cache_free); 1885 1886/* Figure out on which slab page the object resides */ 1887static struct page *get_object_page(const void *x) 1888{ 1889 struct page *page = virt_to_head_page(x); 1890 1891 if (!PageSlab(page)) 1892 return NULL; 1893 1894 return page; 1895} 1896 1897/* 1898 * Object placement in a slab is made very easy because we always start at 1899 * offset 0. If we tune the size of the object to the alignment then we can 1900 * get the required alignment by putting one properly sized object after 1901 * another. 1902 * 1903 * Notice that the allocation order determines the sizes of the per cpu 1904 * caches. Each processor has always one slab available for allocations. 1905 * Increasing the allocation order reduces the number of times that slabs 1906 * must be moved on and off the partial lists and is therefore a factor in 1907 * locking overhead. 1908 */ 1909 1910/* 1911 * Mininum / Maximum order of slab pages. This influences locking overhead 1912 * and slab fragmentation. A higher order reduces the number of partial slabs 1913 * and increases the number of allocations possible without having to 1914 * take the list_lock. 1915 */ 1916static int slub_min_order; 1917static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 1918static int slub_min_objects; 1919 1920/* 1921 * Merge control. If this is set then no merging of slab caches will occur. 1922 * (Could be removed. This was introduced to pacify the merge skeptics.) 1923 */ 1924static int slub_nomerge; 1925 1926/* 1927 * Calculate the order of allocation given an slab object size. 1928 * 1929 * The order of allocation has significant impact on performance and other 1930 * system components. Generally order 0 allocations should be preferred since 1931 * order 0 does not cause fragmentation in the page allocator. Larger objects 1932 * be problematic to put into order 0 slabs because there may be too much 1933 * unused space left. We go to a higher order if more than 1/16th of the slab 1934 * would be wasted. 1935 * 1936 * In order to reach satisfactory performance we must ensure that a minimum 1937 * number of objects is in one slab. Otherwise we may generate too much 1938 * activity on the partial lists which requires taking the list_lock. This is 1939 * less a concern for large slabs though which are rarely used. 1940 * 1941 * slub_max_order specifies the order where we begin to stop considering the 1942 * number of objects in a slab as critical. If we reach slub_max_order then 1943 * we try to keep the page order as low as possible. So we accept more waste 1944 * of space in favor of a small page order. 1945 * 1946 * Higher order allocations also allow the placement of more objects in a 1947 * slab and thereby reduce object handling overhead. If the user has 1948 * requested a higher mininum order then we start with that one instead of 1949 * the smallest order which will fit the object. 1950 */ 1951static inline int slab_order(int size, int min_objects, 1952 int max_order, int fract_leftover) 1953{ 1954 int order; 1955 int rem; 1956 int min_order = slub_min_order; 1957 1958 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE) 1959 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 1960 1961 for (order = max(min_order, 1962 fls(min_objects * size - 1) - PAGE_SHIFT); 1963 order <= max_order; order++) { 1964 1965 unsigned long slab_size = PAGE_SIZE << order; 1966 1967 if (slab_size < min_objects * size) 1968 continue; 1969 1970 rem = slab_size % size; 1971 1972 if (rem <= slab_size / fract_leftover) 1973 break; 1974 1975 } 1976 1977 return order; 1978} 1979 1980static inline int calculate_order(int size) 1981{ 1982 int order; 1983 int min_objects; 1984 int fraction; 1985 int max_objects; 1986 1987 /* 1988 * Attempt to find best configuration for a slab. This 1989 * works by first attempting to generate a layout with 1990 * the best configuration and backing off gradually. 1991 * 1992 * First we reduce the acceptable waste in a slab. Then 1993 * we reduce the minimum objects required in a slab. 1994 */ 1995 min_objects = slub_min_objects; 1996 if (!min_objects) 1997 min_objects = 4 * (fls(nr_cpu_ids) + 1); 1998 max_objects = (PAGE_SIZE << slub_max_order)/size; 1999 min_objects = min(min_objects, max_objects); 2000 2001 while (min_objects > 1) { 2002 fraction = 16; 2003 while (fraction >= 4) { 2004 order = slab_order(size, min_objects, 2005 slub_max_order, fraction); 2006 if (order <= slub_max_order) 2007 return order; 2008 fraction /= 2; 2009 } 2010 min_objects--; 2011 } 2012 2013 /* 2014 * We were unable to place multiple objects in a slab. Now 2015 * lets see if we can place a single object there. 2016 */ 2017 order = slab_order(size, 1, slub_max_order, 1); 2018 if (order <= slub_max_order) 2019 return order; 2020 2021 /* 2022 * Doh this slab cannot be placed using slub_max_order. 2023 */ 2024 order = slab_order(size, 1, MAX_ORDER, 1); 2025 if (order < MAX_ORDER) 2026 return order; 2027 return -ENOSYS; 2028} 2029 2030/* 2031 * Figure out what the alignment of the objects will be. 2032 */ 2033static unsigned long calculate_alignment(unsigned long flags, 2034 unsigned long align, unsigned long size) 2035{ 2036 /* 2037 * If the user wants hardware cache aligned objects then follow that 2038 * suggestion if the object is sufficiently large. 2039 * 2040 * The hardware cache alignment cannot override the specified 2041 * alignment though. If that is greater then use it. 2042 */ 2043 if (flags & SLAB_HWCACHE_ALIGN) { 2044 unsigned long ralign = cache_line_size(); 2045 while (size <= ralign / 2) 2046 ralign /= 2; 2047 align = max(align, ralign); 2048 } 2049 2050 if (align < ARCH_SLAB_MINALIGN) 2051 align = ARCH_SLAB_MINALIGN; 2052 2053 return ALIGN(align, sizeof(void *)); 2054} 2055 2056static void 2057init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 2058{ 2059 n->nr_partial = 0; 2060 spin_lock_init(&n->list_lock); 2061 INIT_LIST_HEAD(&n->partial); 2062#ifdef CONFIG_SLUB_DEBUG 2063 atomic_long_set(&n->nr_slabs, 0); 2064 atomic_long_set(&n->total_objects, 0); 2065 INIT_LIST_HEAD(&n->full); 2066#endif 2067} 2068 2069static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[KMALLOC_CACHES]); 2070 2071static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2072{ 2073 if (s < kmalloc_caches + KMALLOC_CACHES && s >= kmalloc_caches) 2074 /* 2075 * Boot time creation of the kmalloc array. Use static per cpu data 2076 * since the per cpu allocator is not available yet. 2077 */ 2078 s->cpu_slab = kmalloc_percpu + (s - kmalloc_caches); 2079 else 2080 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu); 2081 2082 if (!s->cpu_slab) 2083 return 0; 2084 2085 return 1; 2086} 2087 2088#ifdef CONFIG_NUMA 2089/* 2090 * No kmalloc_node yet so do it by hand. We know that this is the first 2091 * slab on the node for this slabcache. There are no concurrent accesses 2092 * possible. 2093 * 2094 * Note that this function only works on the kmalloc_node_cache 2095 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2096 * memory on a fresh node that has no slab structures yet. 2097 */ 2098static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node) 2099{ 2100 struct page *page; 2101 struct kmem_cache_node *n; 2102 unsigned long flags; 2103 2104 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2105 2106 page = new_slab(kmalloc_caches, gfpflags, node); 2107 2108 BUG_ON(!page); 2109 if (page_to_nid(page) != node) { 2110 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2111 "node %d\n", node); 2112 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2113 "in order to be able to continue\n"); 2114 } 2115 2116 n = page->freelist; 2117 BUG_ON(!n); 2118 page->freelist = get_freepointer(kmalloc_caches, n); 2119 page->inuse++; 2120 kmalloc_caches->node[node] = n; 2121#ifdef CONFIG_SLUB_DEBUG 2122 init_object(kmalloc_caches, n, 1); 2123 init_tracking(kmalloc_caches, n); 2124#endif 2125 init_kmem_cache_node(n, kmalloc_caches); 2126 inc_slabs_node(kmalloc_caches, node, page->objects); 2127 2128 /* 2129 * lockdep requires consistent irq usage for each lock 2130 * so even though there cannot be a race this early in 2131 * the boot sequence, we still disable irqs. 2132 */ 2133 local_irq_save(flags); 2134 add_partial(n, page, 0); 2135 local_irq_restore(flags); 2136} 2137 2138static void free_kmem_cache_nodes(struct kmem_cache *s) 2139{ 2140 int node; 2141 2142 for_each_node_state(node, N_NORMAL_MEMORY) { 2143 struct kmem_cache_node *n = s->node[node]; 2144 if (n) 2145 kmem_cache_free(kmalloc_caches, n); 2146 s->node[node] = NULL; 2147 } 2148} 2149 2150static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2151{ 2152 int node; 2153 2154 for_each_node_state(node, N_NORMAL_MEMORY) { 2155 struct kmem_cache_node *n; 2156 2157 if (slab_state == DOWN) { 2158 early_kmem_cache_node_alloc(gfpflags, node); 2159 continue; 2160 } 2161 n = kmem_cache_alloc_node(kmalloc_caches, 2162 gfpflags, node); 2163 2164 if (!n) { 2165 free_kmem_cache_nodes(s); 2166 return 0; 2167 } 2168 2169 s->node[node] = n; 2170 init_kmem_cache_node(n, s); 2171 } 2172 return 1; 2173} 2174#else 2175static void free_kmem_cache_nodes(struct kmem_cache *s) 2176{ 2177} 2178 2179static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2180{ 2181 init_kmem_cache_node(&s->local_node, s); 2182 return 1; 2183} 2184#endif 2185 2186static void set_min_partial(struct kmem_cache *s, unsigned long min) 2187{ 2188 if (min < MIN_PARTIAL) 2189 min = MIN_PARTIAL; 2190 else if (min > MAX_PARTIAL) 2191 min = MAX_PARTIAL; 2192 s->min_partial = min; 2193} 2194 2195/* 2196 * calculate_sizes() determines the order and the distribution of data within 2197 * a slab object. 2198 */ 2199static int calculate_sizes(struct kmem_cache *s, int forced_order) 2200{ 2201 unsigned long flags = s->flags; 2202 unsigned long size = s->objsize; 2203 unsigned long align = s->align; 2204 int order; 2205 2206 /* 2207 * Round up object size to the next word boundary. We can only 2208 * place the free pointer at word boundaries and this determines 2209 * the possible location of the free pointer. 2210 */ 2211 size = ALIGN(size, sizeof(void *)); 2212 2213#ifdef CONFIG_SLUB_DEBUG 2214 /* 2215 * Determine if we can poison the object itself. If the user of 2216 * the slab may touch the object after free or before allocation 2217 * then we should never poison the object itself. 2218 */ 2219 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2220 !s->ctor) 2221 s->flags |= __OBJECT_POISON; 2222 else 2223 s->flags &= ~__OBJECT_POISON; 2224 2225 2226 /* 2227 * If we are Redzoning then check if there is some space between the 2228 * end of the object and the free pointer. If not then add an 2229 * additional word to have some bytes to store Redzone information. 2230 */ 2231 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2232 size += sizeof(void *); 2233#endif 2234 2235 /* 2236 * With that we have determined the number of bytes in actual use 2237 * by the object. This is the potential offset to the free pointer. 2238 */ 2239 s->inuse = size; 2240 2241 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2242 s->ctor)) { 2243 /* 2244 * Relocate free pointer after the object if it is not 2245 * permitted to overwrite the first word of the object on 2246 * kmem_cache_free. 2247 * 2248 * This is the case if we do RCU, have a constructor or 2249 * destructor or are poisoning the objects. 2250 */ 2251 s->offset = size; 2252 size += sizeof(void *); 2253 } 2254 2255#ifdef CONFIG_SLUB_DEBUG 2256 if (flags & SLAB_STORE_USER) 2257 /* 2258 * Need to store information about allocs and frees after 2259 * the object. 2260 */ 2261 size += 2 * sizeof(struct track); 2262 2263 if (flags & SLAB_RED_ZONE) 2264 /* 2265 * Add some empty padding so that we can catch 2266 * overwrites from earlier objects rather than let 2267 * tracking information or the free pointer be 2268 * corrupted if a user writes before the start 2269 * of the object. 2270 */ 2271 size += sizeof(void *); 2272#endif 2273 2274 /* 2275 * Determine the alignment based on various parameters that the 2276 * user specified and the dynamic determination of cache line size 2277 * on bootup. 2278 */ 2279 align = calculate_alignment(flags, align, s->objsize); 2280 s->align = align; 2281 2282 /* 2283 * SLUB stores one object immediately after another beginning from 2284 * offset 0. In order to align the objects we have to simply size 2285 * each object to conform to the alignment. 2286 */ 2287 size = ALIGN(size, align); 2288 s->size = size; 2289 if (forced_order >= 0) 2290 order = forced_order; 2291 else 2292 order = calculate_order(size); 2293 2294 if (order < 0) 2295 return 0; 2296 2297 s->allocflags = 0; 2298 if (order) 2299 s->allocflags |= __GFP_COMP; 2300 2301 if (s->flags & SLAB_CACHE_DMA) 2302 s->allocflags |= SLUB_DMA; 2303 2304 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2305 s->allocflags |= __GFP_RECLAIMABLE; 2306 2307 /* 2308 * Determine the number of objects per slab 2309 */ 2310 s->oo = oo_make(order, size); 2311 s->min = oo_make(get_order(size), size); 2312 if (oo_objects(s->oo) > oo_objects(s->max)) 2313 s->max = s->oo; 2314 2315 return !!oo_objects(s->oo); 2316 2317} 2318 2319static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2320 const char *name, size_t size, 2321 size_t align, unsigned long flags, 2322 void (*ctor)(void *)) 2323{ 2324 memset(s, 0, kmem_size); 2325 s->name = name; 2326 s->ctor = ctor; 2327 s->objsize = size; 2328 s->align = align; 2329 s->flags = kmem_cache_flags(size, flags, name, ctor); 2330 2331 if (!calculate_sizes(s, -1)) 2332 goto error; 2333 if (disable_higher_order_debug) { 2334 /* 2335 * Disable debugging flags that store metadata if the min slab 2336 * order increased. 2337 */ 2338 if (get_order(s->size) > get_order(s->objsize)) { 2339 s->flags &= ~DEBUG_METADATA_FLAGS; 2340 s->offset = 0; 2341 if (!calculate_sizes(s, -1)) 2342 goto error; 2343 } 2344 } 2345 2346 /* 2347 * The larger the object size is, the more pages we want on the partial 2348 * list to avoid pounding the page allocator excessively. 2349 */ 2350 set_min_partial(s, ilog2(s->size)); 2351 s->refcount = 1; 2352#ifdef CONFIG_NUMA 2353 s->remote_node_defrag_ratio = 1000; 2354#endif 2355 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2356 goto error; 2357 2358 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2359 return 1; 2360 2361 free_kmem_cache_nodes(s); 2362error: 2363 if (flags & SLAB_PANIC) 2364 panic("Cannot create slab %s size=%lu realsize=%u " 2365 "order=%u offset=%u flags=%lx\n", 2366 s->name, (unsigned long)size, s->size, oo_order(s->oo), 2367 s->offset, flags); 2368 return 0; 2369} 2370 2371/* 2372 * Check if a given pointer is valid 2373 */ 2374int kmem_ptr_validate(struct kmem_cache *s, const void *object) 2375{ 2376 struct page *page; 2377 2378 if (!kern_ptr_validate(object, s->size)) 2379 return 0; 2380 2381 page = get_object_page(object); 2382 2383 if (!page || s != page->slab) 2384 /* No slab or wrong slab */ 2385 return 0; 2386 2387 if (!check_valid_pointer(s, page, object)) 2388 return 0; 2389 2390 /* 2391 * We could also check if the object is on the slabs freelist. 2392 * But this would be too expensive and it seems that the main 2393 * purpose of kmem_ptr_valid() is to check if the object belongs 2394 * to a certain slab. 2395 */ 2396 return 1; 2397} 2398EXPORT_SYMBOL(kmem_ptr_validate); 2399 2400/* 2401 * Determine the size of a slab object 2402 */ 2403unsigned int kmem_cache_size(struct kmem_cache *s) 2404{ 2405 return s->objsize; 2406} 2407EXPORT_SYMBOL(kmem_cache_size); 2408 2409const char *kmem_cache_name(struct kmem_cache *s) 2410{ 2411 return s->name; 2412} 2413EXPORT_SYMBOL(kmem_cache_name); 2414 2415static void list_slab_objects(struct kmem_cache *s, struct page *page, 2416 const char *text) 2417{ 2418#ifdef CONFIG_SLUB_DEBUG 2419 void *addr = page_address(page); 2420 void *p; 2421 long *map = kzalloc(BITS_TO_LONGS(page->objects) * sizeof(long), 2422 GFP_ATOMIC); 2423 2424 if (!map) 2425 return; 2426 slab_err(s, page, "%s", text); 2427 slab_lock(page); 2428 for_each_free_object(p, s, page->freelist) 2429 set_bit(slab_index(p, s, addr), map); 2430 2431 for_each_object(p, s, addr, page->objects) { 2432 2433 if (!test_bit(slab_index(p, s, addr), map)) { 2434 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2435 p, p - addr); 2436 print_tracking(s, p); 2437 } 2438 } 2439 slab_unlock(page); 2440 kfree(map); 2441#endif 2442} 2443 2444/* 2445 * Attempt to free all partial slabs on a node. 2446 */ 2447static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2448{ 2449 unsigned long flags; 2450 struct page *page, *h; 2451 2452 spin_lock_irqsave(&n->list_lock, flags); 2453 list_for_each_entry_safe(page, h, &n->partial, lru) { 2454 if (!page->inuse) { 2455 list_del(&page->lru); 2456 discard_slab(s, page); 2457 n->nr_partial--; 2458 } else { 2459 list_slab_objects(s, page, 2460 "Objects remaining on kmem_cache_close()"); 2461 } 2462 } 2463 spin_unlock_irqrestore(&n->list_lock, flags); 2464} 2465 2466/* 2467 * Release all resources used by a slab cache. 2468 */ 2469static inline int kmem_cache_close(struct kmem_cache *s) 2470{ 2471 int node; 2472 2473 flush_all(s); 2474 free_percpu(s->cpu_slab); 2475 /* Attempt to free all objects */ 2476 for_each_node_state(node, N_NORMAL_MEMORY) { 2477 struct kmem_cache_node *n = get_node(s, node); 2478 2479 free_partial(s, n); 2480 if (n->nr_partial || slabs_node(s, node)) 2481 return 1; 2482 } 2483 free_kmem_cache_nodes(s); 2484 return 0; 2485} 2486 2487/* 2488 * Close a cache and release the kmem_cache structure 2489 * (must be used for caches created using kmem_cache_create) 2490 */ 2491void kmem_cache_destroy(struct kmem_cache *s) 2492{ 2493 down_write(&slub_lock); 2494 s->refcount--; 2495 if (!s->refcount) { 2496 list_del(&s->list); 2497 if (kmem_cache_close(s)) { 2498 printk(KERN_ERR "SLUB %s: %s called for cache that " 2499 "still has objects.\n", s->name, __func__); 2500 dump_stack(); 2501 } 2502 if (s->flags & SLAB_DESTROY_BY_RCU) 2503 rcu_barrier(); 2504 sysfs_slab_remove(s); 2505 } 2506 up_write(&slub_lock); 2507} 2508EXPORT_SYMBOL(kmem_cache_destroy); 2509 2510/******************************************************************** 2511 * Kmalloc subsystem 2512 *******************************************************************/ 2513 2514struct kmem_cache kmalloc_caches[KMALLOC_CACHES] __cacheline_aligned; 2515EXPORT_SYMBOL(kmalloc_caches); 2516 2517static int __init setup_slub_min_order(char *str) 2518{ 2519 get_option(&str, &slub_min_order); 2520 2521 return 1; 2522} 2523 2524__setup("slub_min_order=", setup_slub_min_order); 2525 2526static int __init setup_slub_max_order(char *str) 2527{ 2528 get_option(&str, &slub_max_order); 2529 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 2530 2531 return 1; 2532} 2533 2534__setup("slub_max_order=", setup_slub_max_order); 2535 2536static int __init setup_slub_min_objects(char *str) 2537{ 2538 get_option(&str, &slub_min_objects); 2539 2540 return 1; 2541} 2542 2543__setup("slub_min_objects=", setup_slub_min_objects); 2544 2545static int __init setup_slub_nomerge(char *str) 2546{ 2547 slub_nomerge = 1; 2548 return 1; 2549} 2550 2551__setup("slub_nomerge", setup_slub_nomerge); 2552 2553static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2554 const char *name, int size, gfp_t gfp_flags) 2555{ 2556 unsigned int flags = 0; 2557 2558 if (gfp_flags & SLUB_DMA) 2559 flags = SLAB_CACHE_DMA; 2560 2561 /* 2562 * This function is called with IRQs disabled during early-boot on 2563 * single CPU so there's no need to take slub_lock here. 2564 */ 2565 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2566 flags, NULL)) 2567 goto panic; 2568 2569 list_add(&s->list, &slab_caches); 2570 2571 if (sysfs_slab_add(s)) 2572 goto panic; 2573 return s; 2574 2575panic: 2576 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2577} 2578 2579#ifdef CONFIG_ZONE_DMA 2580static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT]; 2581 2582static void sysfs_add_func(struct work_struct *w) 2583{ 2584 struct kmem_cache *s; 2585 2586 down_write(&slub_lock); 2587 list_for_each_entry(s, &slab_caches, list) { 2588 if (s->flags & __SYSFS_ADD_DEFERRED) { 2589 s->flags &= ~__SYSFS_ADD_DEFERRED; 2590 sysfs_slab_add(s); 2591 } 2592 } 2593 up_write(&slub_lock); 2594} 2595 2596static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2597 2598static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2599{ 2600 struct kmem_cache *s; 2601 char *text; 2602 size_t realsize; 2603 unsigned long slabflags; 2604 int i; 2605 2606 s = kmalloc_caches_dma[index]; 2607 if (s) 2608 return s; 2609 2610 /* Dynamically create dma cache */ 2611 if (flags & __GFP_WAIT) 2612 down_write(&slub_lock); 2613 else { 2614 if (!down_write_trylock(&slub_lock)) 2615 goto out; 2616 } 2617 2618 if (kmalloc_caches_dma[index]) 2619 goto unlock_out; 2620 2621 realsize = kmalloc_caches[index].objsize; 2622 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", 2623 (unsigned int)realsize); 2624 2625 s = NULL; 2626 for (i = 0; i < KMALLOC_CACHES; i++) 2627 if (!kmalloc_caches[i].size) 2628 break; 2629 2630 BUG_ON(i >= KMALLOC_CACHES); 2631 s = kmalloc_caches + i; 2632 2633 /* 2634 * Must defer sysfs creation to a workqueue because we don't know 2635 * what context we are called from. Before sysfs comes up, we don't 2636 * need to do anything because our sysfs initcall will start by 2637 * adding all existing slabs to sysfs. 2638 */ 2639 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK; 2640 if (slab_state >= SYSFS) 2641 slabflags |= __SYSFS_ADD_DEFERRED; 2642 2643 if (!text || !kmem_cache_open(s, flags, text, 2644 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) { 2645 s->size = 0; 2646 kfree(text); 2647 goto unlock_out; 2648 } 2649 2650 list_add(&s->list, &slab_caches); 2651 kmalloc_caches_dma[index] = s; 2652 2653 if (slab_state >= SYSFS) 2654 schedule_work(&sysfs_add_work); 2655 2656unlock_out: 2657 up_write(&slub_lock); 2658out: 2659 return kmalloc_caches_dma[index]; 2660} 2661#endif 2662 2663/* 2664 * Conversion table for small slabs sizes / 8 to the index in the 2665 * kmalloc array. This is necessary for slabs < 192 since we have non power 2666 * of two cache sizes there. The size of larger slabs can be determined using 2667 * fls. 2668 */ 2669static s8 size_index[24] = { 2670 3, /* 8 */ 2671 4, /* 16 */ 2672 5, /* 24 */ 2673 5, /* 32 */ 2674 6, /* 40 */ 2675 6, /* 48 */ 2676 6, /* 56 */ 2677 6, /* 64 */ 2678 1, /* 72 */ 2679 1, /* 80 */ 2680 1, /* 88 */ 2681 1, /* 96 */ 2682 7, /* 104 */ 2683 7, /* 112 */ 2684 7, /* 120 */ 2685 7, /* 128 */ 2686 2, /* 136 */ 2687 2, /* 144 */ 2688 2, /* 152 */ 2689 2, /* 160 */ 2690 2, /* 168 */ 2691 2, /* 176 */ 2692 2, /* 184 */ 2693 2 /* 192 */ 2694}; 2695 2696static inline int size_index_elem(size_t bytes) 2697{ 2698 return (bytes - 1) / 8; 2699} 2700 2701static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2702{ 2703 int index; 2704 2705 if (size <= 192) { 2706 if (!size) 2707 return ZERO_SIZE_PTR; 2708 2709 index = size_index[size_index_elem(size)]; 2710 } else 2711 index = fls(size - 1); 2712 2713#ifdef CONFIG_ZONE_DMA 2714 if (unlikely((flags & SLUB_DMA))) 2715 return dma_kmalloc_cache(index, flags); 2716 2717#endif 2718 return &kmalloc_caches[index]; 2719} 2720 2721void *__kmalloc(size_t size, gfp_t flags) 2722{ 2723 struct kmem_cache *s; 2724 void *ret; 2725 2726 if (unlikely(size > SLUB_MAX_SIZE)) 2727 return kmalloc_large(size, flags); 2728 2729 s = get_slab(size, flags); 2730 2731 if (unlikely(ZERO_OR_NULL_PTR(s))) 2732 return s; 2733 2734 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); 2735 2736 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 2737 2738 return ret; 2739} 2740EXPORT_SYMBOL(__kmalloc); 2741 2742static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2743{ 2744 struct page *page; 2745 void *ptr = NULL; 2746 2747 flags |= __GFP_COMP | __GFP_NOTRACK; 2748 page = alloc_pages_node(node, flags, get_order(size)); 2749 if (page) 2750 ptr = page_address(page); 2751 2752 kmemleak_alloc(ptr, size, 1, flags); 2753 return ptr; 2754} 2755 2756#ifdef CONFIG_NUMA 2757void *__kmalloc_node(size_t size, gfp_t flags, int node) 2758{ 2759 struct kmem_cache *s; 2760 void *ret; 2761 2762 if (unlikely(size > SLUB_MAX_SIZE)) { 2763 ret = kmalloc_large_node(size, flags, node); 2764 2765 trace_kmalloc_node(_RET_IP_, ret, 2766 size, PAGE_SIZE << get_order(size), 2767 flags, node); 2768 2769 return ret; 2770 } 2771 2772 s = get_slab(size, flags); 2773 2774 if (unlikely(ZERO_OR_NULL_PTR(s))) 2775 return s; 2776 2777 ret = slab_alloc(s, flags, node, _RET_IP_); 2778 2779 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 2780 2781 return ret; 2782} 2783EXPORT_SYMBOL(__kmalloc_node); 2784#endif 2785 2786size_t ksize(const void *object) 2787{ 2788 struct page *page; 2789 struct kmem_cache *s; 2790 2791 if (unlikely(object == ZERO_SIZE_PTR)) 2792 return 0; 2793 2794 page = virt_to_head_page(object); 2795 2796 if (unlikely(!PageSlab(page))) { 2797 WARN_ON(!PageCompound(page)); 2798 return PAGE_SIZE << compound_order(page); 2799 } 2800 s = page->slab; 2801 2802#ifdef CONFIG_SLUB_DEBUG 2803 /* 2804 * Debugging requires use of the padding between object 2805 * and whatever may come after it. 2806 */ 2807 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2808 return s->objsize; 2809 2810#endif 2811 /* 2812 * If we have the need to store the freelist pointer 2813 * back there or track user information then we can 2814 * only use the space before that information. 2815 */ 2816 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2817 return s->inuse; 2818 /* 2819 * Else we can use all the padding etc for the allocation 2820 */ 2821 return s->size; 2822} 2823EXPORT_SYMBOL(ksize); 2824 2825void BCMFASTPATH_HOST kfree(const void *x) 2826{ 2827 struct page *page; 2828 void *object = (void *)x; 2829 2830 trace_kfree(_RET_IP_, x); 2831 2832 if (unlikely(ZERO_OR_NULL_PTR(x))) 2833 return; 2834 2835 page = virt_to_head_page(x); 2836 if (unlikely(!PageSlab(page))) { 2837 BUG_ON(!PageCompound(page)); 2838 kmemleak_free(x); 2839 put_page(page); 2840 return; 2841 } 2842 slab_free(page->slab, page, object, _RET_IP_); 2843} 2844EXPORT_SYMBOL(kfree); 2845 2846/* 2847 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2848 * the remaining slabs by the number of items in use. The slabs with the 2849 * most items in use come first. New allocations will then fill those up 2850 * and thus they can be removed from the partial lists. 2851 * 2852 * The slabs with the least items are placed last. This results in them 2853 * being allocated from last increasing the chance that the last objects 2854 * are freed in them. 2855 */ 2856int kmem_cache_shrink(struct kmem_cache *s) 2857{ 2858 int node; 2859 int i; 2860 struct kmem_cache_node *n; 2861 struct page *page; 2862 struct page *t; 2863 int objects = oo_objects(s->max); 2864 struct list_head *slabs_by_inuse = 2865 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 2866 unsigned long flags; 2867 2868 if (!slabs_by_inuse) 2869 return -ENOMEM; 2870 2871 flush_all(s); 2872 for_each_node_state(node, N_NORMAL_MEMORY) { 2873 n = get_node(s, node); 2874 2875 if (!n->nr_partial) 2876 continue; 2877 2878 for (i = 0; i < objects; i++) 2879 INIT_LIST_HEAD(slabs_by_inuse + i); 2880 2881 spin_lock_irqsave(&n->list_lock, flags); 2882 2883 /* 2884 * Build lists indexed by the items in use in each slab. 2885 * 2886 * Note that concurrent frees may occur while we hold the 2887 * list_lock. page->inuse here is the upper limit. 2888 */ 2889 list_for_each_entry_safe(page, t, &n->partial, lru) { 2890 if (!page->inuse && slab_trylock(page)) { 2891 /* 2892 * Must hold slab lock here because slab_free 2893 * may have freed the last object and be 2894 * waiting to release the slab. 2895 */ 2896 list_del(&page->lru); 2897 n->nr_partial--; 2898 slab_unlock(page); 2899 discard_slab(s, page); 2900 } else { 2901 list_move(&page->lru, 2902 slabs_by_inuse + page->inuse); 2903 } 2904 } 2905 2906 /* 2907 * Rebuild the partial list with the slabs filled up most 2908 * first and the least used slabs at the end. 2909 */ 2910 for (i = objects - 1; i >= 0; i--) 2911 list_splice(slabs_by_inuse + i, n->partial.prev); 2912 2913 spin_unlock_irqrestore(&n->list_lock, flags); 2914 } 2915 2916 kfree(slabs_by_inuse); 2917 return 0; 2918} 2919EXPORT_SYMBOL(kmem_cache_shrink); 2920 2921#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2922static int slab_mem_going_offline_callback(void *arg) 2923{ 2924 struct kmem_cache *s; 2925 2926 down_read(&slub_lock); 2927 list_for_each_entry(s, &slab_caches, list) 2928 kmem_cache_shrink(s); 2929 up_read(&slub_lock); 2930 2931 return 0; 2932} 2933 2934static void slab_mem_offline_callback(void *arg) 2935{ 2936 struct kmem_cache_node *n; 2937 struct kmem_cache *s; 2938 struct memory_notify *marg = arg; 2939 int offline_node; 2940 2941 offline_node = marg->status_change_nid; 2942 2943 /* 2944 * If the node still has available memory. we need kmem_cache_node 2945 * for it yet. 2946 */ 2947 if (offline_node < 0) 2948 return; 2949 2950 down_read(&slub_lock); 2951 list_for_each_entry(s, &slab_caches, list) { 2952 n = get_node(s, offline_node); 2953 if (n) { 2954 /* 2955 * if n->nr_slabs > 0, slabs still exist on the node 2956 * that is going down. We were unable to free them, 2957 * and offline_pages() function shouldn't call this 2958 * callback. So, we must fail. 2959 */ 2960 BUG_ON(slabs_node(s, offline_node)); 2961 2962 s->node[offline_node] = NULL; 2963 kmem_cache_free(kmalloc_caches, n); 2964 } 2965 } 2966 up_read(&slub_lock); 2967} 2968 2969static int slab_mem_going_online_callback(void *arg) 2970{ 2971 struct kmem_cache_node *n; 2972 struct kmem_cache *s; 2973 struct memory_notify *marg = arg; 2974 int nid = marg->status_change_nid; 2975 int ret = 0; 2976 2977 /* 2978 * If the node's memory is already available, then kmem_cache_node is 2979 * already created. Nothing to do. 2980 */ 2981 if (nid < 0) 2982 return 0; 2983 2984 /* 2985 * We are bringing a node online. No memory is available yet. We must 2986 * allocate a kmem_cache_node structure in order to bring the node 2987 * online. 2988 */ 2989 down_read(&slub_lock); 2990 list_for_each_entry(s, &slab_caches, list) { 2991 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 2992 if (!n) { 2993 ret = -ENOMEM; 2994 goto out; 2995 } 2996 init_kmem_cache_node(n, s); 2997 s->node[nid] = n; 2998 } 2999out: 3000 up_read(&slub_lock); 3001 return ret; 3002} 3003 3004static int slab_memory_callback(struct notifier_block *self, 3005 unsigned long action, void *arg) 3006{ 3007 int ret = 0; 3008 3009 switch (action) { 3010 case MEM_GOING_ONLINE: 3011 ret = slab_mem_going_online_callback(arg); 3012 break; 3013 case MEM_GOING_OFFLINE: 3014 ret = slab_mem_going_offline_callback(arg); 3015 break; 3016 case MEM_OFFLINE: 3017 case MEM_CANCEL_ONLINE: 3018 slab_mem_offline_callback(arg); 3019 break; 3020 case MEM_ONLINE: 3021 case MEM_CANCEL_OFFLINE: 3022 break; 3023 } 3024 if (ret) 3025 ret = notifier_from_errno(ret); 3026 else 3027 ret = NOTIFY_OK; 3028 return ret; 3029} 3030 3031#endif /* CONFIG_MEMORY_HOTPLUG */ 3032 3033/******************************************************************** 3034 * Basic setup of slabs 3035 *******************************************************************/ 3036 3037void __init kmem_cache_init(void) 3038{ 3039 int i; 3040 int caches = 0; 3041 3042#ifdef CONFIG_NUMA 3043 /* 3044 * Must first have the slab cache available for the allocations of the 3045 * struct kmem_cache_node's. There is special bootstrap code in 3046 * kmem_cache_open for slab_state == DOWN. 3047 */ 3048 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 3049 sizeof(struct kmem_cache_node), GFP_NOWAIT); 3050 kmalloc_caches[0].refcount = -1; 3051 caches++; 3052 3053 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 3054#endif 3055 3056 /* Able to allocate the per node structures */ 3057 slab_state = PARTIAL; 3058 3059 /* Caches that are not of the two-to-the-power-of size */ 3060 if (KMALLOC_MIN_SIZE <= 32) { 3061 create_kmalloc_cache(&kmalloc_caches[1], 3062 "kmalloc-96", 96, GFP_NOWAIT); 3063 caches++; 3064 } 3065 if (KMALLOC_MIN_SIZE <= 64) { 3066 create_kmalloc_cache(&kmalloc_caches[2], 3067 "kmalloc-192", 192, GFP_NOWAIT); 3068 caches++; 3069 } 3070 3071 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3072 create_kmalloc_cache(&kmalloc_caches[i], 3073 "kmalloc", 1 << i, GFP_NOWAIT); 3074 caches++; 3075 } 3076 3077 3078 /* 3079 * Patch up the size_index table if we have strange large alignment 3080 * requirements for the kmalloc array. This is only the case for 3081 * MIPS it seems. The standard arches will not generate any code here. 3082 * 3083 * Largest permitted alignment is 256 bytes due to the way we 3084 * handle the index determination for the smaller caches. 3085 * 3086 * Make sure that nothing crazy happens if someone starts tinkering 3087 * around with ARCH_KMALLOC_MINALIGN 3088 */ 3089 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3090 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3091 3092 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { 3093 int elem = size_index_elem(i); 3094 if (elem >= ARRAY_SIZE(size_index)) 3095 break; 3096 size_index[elem] = KMALLOC_SHIFT_LOW; 3097 } 3098 3099 if (KMALLOC_MIN_SIZE == 64) { 3100 /* 3101 * The 96 byte size cache is not used if the alignment 3102 * is 64 byte. 3103 */ 3104 for (i = 64 + 8; i <= 96; i += 8) 3105 size_index[size_index_elem(i)] = 7; 3106 } else if (KMALLOC_MIN_SIZE == 128) { 3107 /* 3108 * The 192 byte sized cache is not used if the alignment 3109 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3110 * instead. 3111 */ 3112 for (i = 128 + 8; i <= 192; i += 8) 3113 size_index[size_index_elem(i)] = 8; 3114 } 3115 3116 slab_state = UP; 3117 3118 /* Provide the correct kmalloc names now that the caches are up */ 3119 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3120 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); 3121 3122 BUG_ON(!s); 3123 kmalloc_caches[i].name = s; 3124 } 3125 3126#ifdef CONFIG_SMP 3127 register_cpu_notifier(&slab_notifier); 3128#endif 3129#ifdef CONFIG_NUMA 3130 kmem_size = offsetof(struct kmem_cache, node) + 3131 nr_node_ids * sizeof(struct kmem_cache_node *); 3132#else 3133 kmem_size = sizeof(struct kmem_cache); 3134#endif 3135 3136 printk(KERN_INFO 3137 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3138 " CPUs=%d, Nodes=%d\n", 3139 caches, cache_line_size(), 3140 slub_min_order, slub_max_order, slub_min_objects, 3141 nr_cpu_ids, nr_node_ids); 3142} 3143 3144void __init kmem_cache_init_late(void) 3145{ 3146} 3147 3148/* 3149 * Find a mergeable slab cache 3150 */ 3151static int slab_unmergeable(struct kmem_cache *s) 3152{ 3153 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3154 return 1; 3155 3156 if (s->ctor) 3157 return 1; 3158 3159 /* 3160 * We may have set a slab to be unmergeable during bootstrap. 3161 */ 3162 if (s->refcount < 0) 3163 return 1; 3164 3165 return 0; 3166} 3167 3168static struct kmem_cache *find_mergeable(size_t size, 3169 size_t align, unsigned long flags, const char *name, 3170 void (*ctor)(void *)) 3171{ 3172 struct kmem_cache *s; 3173 3174 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3175 return NULL; 3176 3177 if (ctor) 3178 return NULL; 3179 3180 size = ALIGN(size, sizeof(void *)); 3181 align = calculate_alignment(flags, align, size); 3182 size = ALIGN(size, align); 3183 flags = kmem_cache_flags(size, flags, name, NULL); 3184 3185 list_for_each_entry(s, &slab_caches, list) { 3186 if (slab_unmergeable(s)) 3187 continue; 3188 3189 if (size > s->size) 3190 continue; 3191 3192 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3193 continue; 3194 /* 3195 * Check if alignment is compatible. 3196 * Courtesy of Adrian Drzewiecki 3197 */ 3198 if ((s->size & ~(align - 1)) != s->size) 3199 continue; 3200 3201 if (s->size - size >= sizeof(void *)) 3202 continue; 3203 3204 return s; 3205 } 3206 return NULL; 3207} 3208 3209struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3210 size_t align, unsigned long flags, void (*ctor)(void *)) 3211{ 3212 struct kmem_cache *s; 3213 3214 if (WARN_ON(!name)) 3215 return NULL; 3216 3217 down_write(&slub_lock); 3218 s = find_mergeable(size, align, flags, name, ctor); 3219 if (s) { 3220 s->refcount++; 3221 /* 3222 * Adjust the object sizes so that we clear 3223 * the complete object on kzalloc. 3224 */ 3225 s->objsize = max(s->objsize, (int)size); 3226 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3227 3228 if (sysfs_slab_alias(s, name)) { 3229 s->refcount--; 3230 goto err; 3231 } 3232 up_write(&slub_lock); 3233 return s; 3234 } 3235 3236 s = kmalloc(kmem_size, GFP_KERNEL); 3237 if (s) { 3238 if (kmem_cache_open(s, GFP_KERNEL, name, 3239 size, align, flags, ctor)) { 3240 list_add(&s->list, &slab_caches); 3241 if (sysfs_slab_add(s)) { 3242 list_del(&s->list); 3243 kfree(s); 3244 goto err; 3245 } 3246 up_write(&slub_lock); 3247 return s; 3248 } 3249 kfree(s); 3250 } 3251 up_write(&slub_lock); 3252 3253err: 3254 if (flags & SLAB_PANIC) 3255 panic("Cannot create slabcache %s\n", name); 3256 else 3257 s = NULL; 3258 return s; 3259} 3260EXPORT_SYMBOL(kmem_cache_create); 3261 3262#ifdef CONFIG_SMP 3263/* 3264 * Use the cpu notifier to insure that the cpu slabs are flushed when 3265 * necessary. 3266 */ 3267static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3268 unsigned long action, void *hcpu) 3269{ 3270 long cpu = (long)hcpu; 3271 struct kmem_cache *s; 3272 unsigned long flags; 3273 3274 switch (action) { 3275 case CPU_UP_CANCELED: 3276 case CPU_UP_CANCELED_FROZEN: 3277 case CPU_DEAD: 3278 case CPU_DEAD_FROZEN: 3279 down_read(&slub_lock); 3280 list_for_each_entry(s, &slab_caches, list) { 3281 local_irq_save(flags); 3282 __flush_cpu_slab(s, cpu); 3283 local_irq_restore(flags); 3284 } 3285 up_read(&slub_lock); 3286 break; 3287 default: 3288 break; 3289 } 3290 return NOTIFY_OK; 3291} 3292 3293static struct notifier_block __cpuinitdata slab_notifier = { 3294 .notifier_call = slab_cpuup_callback 3295}; 3296 3297#endif 3298 3299void * BCMFASTPATH_HOST __kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3300{ 3301 struct kmem_cache *s; 3302 void *ret; 3303 3304 if (unlikely(size > SLUB_MAX_SIZE)) 3305 return kmalloc_large(size, gfpflags); 3306 3307 s = get_slab(size, gfpflags); 3308 3309 if (unlikely(ZERO_OR_NULL_PTR(s))) 3310 return s; 3311 3312 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); 3313 3314 /* Honor the call site pointer we recieved. */ 3315 trace_kmalloc(caller, ret, size, s->size, gfpflags); 3316 3317 return ret; 3318} 3319 3320void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3321 int node, unsigned long caller) 3322{ 3323 struct kmem_cache *s; 3324 void *ret; 3325 3326 if (unlikely(size > SLUB_MAX_SIZE)) { 3327 ret = kmalloc_large_node(size, gfpflags, node); 3328 3329 trace_kmalloc_node(caller, ret, 3330 size, PAGE_SIZE << get_order(size), 3331 gfpflags, node); 3332 3333 return ret; 3334 } 3335 3336 s = get_slab(size, gfpflags); 3337 3338 if (unlikely(ZERO_OR_NULL_PTR(s))) 3339 return s; 3340 3341 ret = slab_alloc(s, gfpflags, node, caller); 3342 3343 /* Honor the call site pointer we recieved. */ 3344 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 3345 3346 return ret; 3347} 3348 3349#ifdef CONFIG_SLUB_DEBUG 3350static int count_inuse(struct page *page) 3351{ 3352 return page->inuse; 3353} 3354 3355static int count_total(struct page *page) 3356{ 3357 return page->objects; 3358} 3359 3360static int validate_slab(struct kmem_cache *s, struct page *page, 3361 unsigned long *map) 3362{ 3363 void *p; 3364 void *addr = page_address(page); 3365 3366 if (!check_slab(s, page) || 3367 !on_freelist(s, page, NULL)) 3368 return 0; 3369 3370 /* Now we know that a valid freelist exists */ 3371 bitmap_zero(map, page->objects); 3372 3373 for_each_free_object(p, s, page->freelist) { 3374 set_bit(slab_index(p, s, addr), map); 3375 if (!check_object(s, page, p, 0)) 3376 return 0; 3377 } 3378 3379 for_each_object(p, s, addr, page->objects) 3380 if (!test_bit(slab_index(p, s, addr), map)) 3381 if (!check_object(s, page, p, 1)) 3382 return 0; 3383 return 1; 3384} 3385 3386static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3387 unsigned long *map) 3388{ 3389 if (slab_trylock(page)) { 3390 validate_slab(s, page, map); 3391 slab_unlock(page); 3392 } else 3393 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3394 s->name, page); 3395} 3396 3397static int validate_slab_node(struct kmem_cache *s, 3398 struct kmem_cache_node *n, unsigned long *map) 3399{ 3400 unsigned long count = 0; 3401 struct page *page; 3402 unsigned long flags; 3403 3404 spin_lock_irqsave(&n->list_lock, flags); 3405 3406 list_for_each_entry(page, &n->partial, lru) { 3407 validate_slab_slab(s, page, map); 3408 count++; 3409 } 3410 if (count != n->nr_partial) 3411 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3412 "counter=%ld\n", s->name, count, n->nr_partial); 3413 3414 if (!(s->flags & SLAB_STORE_USER)) 3415 goto out; 3416 3417 list_for_each_entry(page, &n->full, lru) { 3418 validate_slab_slab(s, page, map); 3419 count++; 3420 } 3421 if (count != atomic_long_read(&n->nr_slabs)) 3422 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3423 "counter=%ld\n", s->name, count, 3424 atomic_long_read(&n->nr_slabs)); 3425 3426out: 3427 spin_unlock_irqrestore(&n->list_lock, flags); 3428 return count; 3429} 3430 3431static long validate_slab_cache(struct kmem_cache *s) 3432{ 3433 int node; 3434 unsigned long count = 0; 3435 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3436 sizeof(unsigned long), GFP_KERNEL); 3437 3438 if (!map) 3439 return -ENOMEM; 3440 3441 flush_all(s); 3442 for_each_node_state(node, N_NORMAL_MEMORY) { 3443 struct kmem_cache_node *n = get_node(s, node); 3444 3445 count += validate_slab_node(s, n, map); 3446 } 3447 kfree(map); 3448 return count; 3449} 3450 3451#ifdef SLUB_RESILIENCY_TEST 3452static void resiliency_test(void) 3453{ 3454 u8 *p; 3455 3456 printk(KERN_ERR "SLUB resiliency testing\n"); 3457 printk(KERN_ERR "-----------------------\n"); 3458 printk(KERN_ERR "A. Corruption after allocation\n"); 3459 3460 p = kzalloc(16, GFP_KERNEL); 3461 p[16] = 0x12; 3462 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3463 " 0x12->0x%p\n\n", p + 16); 3464 3465 validate_slab_cache(kmalloc_caches + 4); 3466 3467 /* Hmmm... The next two are dangerous */ 3468 p = kzalloc(32, GFP_KERNEL); 3469 p[32 + sizeof(void *)] = 0x34; 3470 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3471 " 0x34 -> -0x%p\n", p); 3472 printk(KERN_ERR 3473 "If allocated object is overwritten then not detectable\n\n"); 3474 3475 validate_slab_cache(kmalloc_caches + 5); 3476 p = kzalloc(64, GFP_KERNEL); 3477 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3478 *p = 0x56; 3479 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3480 p); 3481 printk(KERN_ERR 3482 "If allocated object is overwritten then not detectable\n\n"); 3483 validate_slab_cache(kmalloc_caches + 6); 3484 3485 printk(KERN_ERR "\nB. Corruption after free\n"); 3486 p = kzalloc(128, GFP_KERNEL); 3487 kfree(p); 3488 *p = 0x78; 3489 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3490 validate_slab_cache(kmalloc_caches + 7); 3491 3492 p = kzalloc(256, GFP_KERNEL); 3493 kfree(p); 3494 p[50] = 0x9a; 3495 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3496 p); 3497 validate_slab_cache(kmalloc_caches + 8); 3498 3499 p = kzalloc(512, GFP_KERNEL); 3500 kfree(p); 3501 p[512] = 0xab; 3502 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3503 validate_slab_cache(kmalloc_caches + 9); 3504} 3505#else 3506static void resiliency_test(void) {}; 3507#endif 3508 3509/* 3510 * Generate lists of code addresses where slabcache objects are allocated 3511 * and freed. 3512 */ 3513 3514struct location { 3515 unsigned long count; 3516 unsigned long addr; 3517 long long sum_time; 3518 long min_time; 3519 long max_time; 3520 long min_pid; 3521 long max_pid; 3522 DECLARE_BITMAP(cpus, NR_CPUS); 3523 nodemask_t nodes; 3524}; 3525 3526struct loc_track { 3527 unsigned long max; 3528 unsigned long count; 3529 struct location *loc; 3530}; 3531 3532static void free_loc_track(struct loc_track *t) 3533{ 3534 if (t->max) 3535 free_pages((unsigned long)t->loc, 3536 get_order(sizeof(struct location) * t->max)); 3537} 3538 3539static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3540{ 3541 struct location *l; 3542 int order; 3543 3544 order = get_order(sizeof(struct location) * max); 3545 3546 l = (void *)__get_free_pages(flags, order); 3547 if (!l) 3548 return 0; 3549 3550 if (t->count) { 3551 memcpy(l, t->loc, sizeof(struct location) * t->count); 3552 free_loc_track(t); 3553 } 3554 t->max = max; 3555 t->loc = l; 3556 return 1; 3557} 3558 3559static int add_location(struct loc_track *t, struct kmem_cache *s, 3560 const struct track *track) 3561{ 3562 long start, end, pos; 3563 struct location *l; 3564 unsigned long caddr; 3565 unsigned long age = jiffies - track->when; 3566 3567 start = -1; 3568 end = t->count; 3569 3570 for ( ; ; ) { 3571 pos = start + (end - start + 1) / 2; 3572 3573 /* 3574 * There is nothing at "end". If we end up there 3575 * we need to add something to before end. 3576 */ 3577 if (pos == end) 3578 break; 3579 3580 caddr = t->loc[pos].addr; 3581 if (track->addr == caddr) { 3582 3583 l = &t->loc[pos]; 3584 l->count++; 3585 if (track->when) { 3586 l->sum_time += age; 3587 if (age < l->min_time) 3588 l->min_time = age; 3589 if (age > l->max_time) 3590 l->max_time = age; 3591 3592 if (track->pid < l->min_pid) 3593 l->min_pid = track->pid; 3594 if (track->pid > l->max_pid) 3595 l->max_pid = track->pid; 3596 3597 cpumask_set_cpu(track->cpu, 3598 to_cpumask(l->cpus)); 3599 } 3600 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3601 return 1; 3602 } 3603 3604 if (track->addr < caddr) 3605 end = pos; 3606 else 3607 start = pos; 3608 } 3609 3610 /* 3611 * Not found. Insert new tracking element. 3612 */ 3613 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3614 return 0; 3615 3616 l = t->loc + pos; 3617 if (pos < t->count) 3618 memmove(l + 1, l, 3619 (t->count - pos) * sizeof(struct location)); 3620 t->count++; 3621 l->count = 1; 3622 l->addr = track->addr; 3623 l->sum_time = age; 3624 l->min_time = age; 3625 l->max_time = age; 3626 l->min_pid = track->pid; 3627 l->max_pid = track->pid; 3628 cpumask_clear(to_cpumask(l->cpus)); 3629 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 3630 nodes_clear(l->nodes); 3631 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3632 return 1; 3633} 3634 3635static void process_slab(struct loc_track *t, struct kmem_cache *s, 3636 struct page *page, enum track_item alloc, 3637 long *map) 3638{ 3639 void *addr = page_address(page); 3640 void *p; 3641 3642 bitmap_zero(map, page->objects); 3643 for_each_free_object(p, s, page->freelist) 3644 set_bit(slab_index(p, s, addr), map); 3645 3646 for_each_object(p, s, addr, page->objects) 3647 if (!test_bit(slab_index(p, s, addr), map)) 3648 add_location(t, s, get_track(s, p, alloc)); 3649} 3650 3651static int list_locations(struct kmem_cache *s, char *buf, 3652 enum track_item alloc) 3653{ 3654 int len = 0; 3655 unsigned long i; 3656 struct loc_track t = { 0, 0, NULL }; 3657 int node; 3658 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3659 sizeof(unsigned long), GFP_KERNEL); 3660 3661 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3662 GFP_TEMPORARY)) { 3663 kfree(map); 3664 return sprintf(buf, "Out of memory\n"); 3665 } 3666 /* Push back cpu slabs */ 3667 flush_all(s); 3668 3669 for_each_node_state(node, N_NORMAL_MEMORY) { 3670 struct kmem_cache_node *n = get_node(s, node); 3671 unsigned long flags; 3672 struct page *page; 3673 3674 if (!atomic_long_read(&n->nr_slabs)) 3675 continue; 3676 3677 spin_lock_irqsave(&n->list_lock, flags); 3678 list_for_each_entry(page, &n->partial, lru) 3679 process_slab(&t, s, page, alloc, map); 3680 list_for_each_entry(page, &n->full, lru) 3681 process_slab(&t, s, page, alloc, map); 3682 spin_unlock_irqrestore(&n->list_lock, flags); 3683 } 3684 3685 for (i = 0; i < t.count; i++) { 3686 struct location *l = &t.loc[i]; 3687 3688 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 3689 break; 3690 len += sprintf(buf + len, "%7ld ", l->count); 3691 3692 if (l->addr) 3693 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3694 else 3695 len += sprintf(buf + len, "<not-available>"); 3696 3697 if (l->sum_time != l->min_time) { 3698 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3699 l->min_time, 3700 (long)div_u64(l->sum_time, l->count), 3701 l->max_time); 3702 } else 3703 len += sprintf(buf + len, " age=%ld", 3704 l->min_time); 3705 3706 if (l->min_pid != l->max_pid) 3707 len += sprintf(buf + len, " pid=%ld-%ld", 3708 l->min_pid, l->max_pid); 3709 else 3710 len += sprintf(buf + len, " pid=%ld", 3711 l->min_pid); 3712 3713 if (num_online_cpus() > 1 && 3714 !cpumask_empty(to_cpumask(l->cpus)) && 3715 len < PAGE_SIZE - 60) { 3716 len += sprintf(buf + len, " cpus="); 3717 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3718 to_cpumask(l->cpus)); 3719 } 3720 3721 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 3722 len < PAGE_SIZE - 60) { 3723 len += sprintf(buf + len, " nodes="); 3724 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3725 l->nodes); 3726 } 3727 3728 len += sprintf(buf + len, "\n"); 3729 } 3730 3731 free_loc_track(&t); 3732 kfree(map); 3733 if (!t.count) 3734 len += sprintf(buf, "No data\n"); 3735 return len; 3736} 3737 3738enum slab_stat_type { 3739 SL_ALL, /* All slabs */ 3740 SL_PARTIAL, /* Only partially allocated slabs */ 3741 SL_CPU, /* Only slabs used for cpu caches */ 3742 SL_OBJECTS, /* Determine allocated objects not slabs */ 3743 SL_TOTAL /* Determine object capacity not slabs */ 3744}; 3745 3746#define SO_ALL (1 << SL_ALL) 3747#define SO_PARTIAL (1 << SL_PARTIAL) 3748#define SO_CPU (1 << SL_CPU) 3749#define SO_OBJECTS (1 << SL_OBJECTS) 3750#define SO_TOTAL (1 << SL_TOTAL) 3751 3752static ssize_t show_slab_objects(struct kmem_cache *s, 3753 char *buf, unsigned long flags) 3754{ 3755 unsigned long total = 0; 3756 int node; 3757 int x; 3758 unsigned long *nodes; 3759 unsigned long *per_cpu; 3760 3761 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3762 if (!nodes) 3763 return -ENOMEM; 3764 per_cpu = nodes + nr_node_ids; 3765 3766 if (flags & SO_CPU) { 3767 int cpu; 3768 3769 for_each_possible_cpu(cpu) { 3770 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 3771 3772 if (!c || c->node < 0) 3773 continue; 3774 3775 if (c->page) { 3776 if (flags & SO_TOTAL) 3777 x = c->page->objects; 3778 else if (flags & SO_OBJECTS) 3779 x = c->page->inuse; 3780 else 3781 x = 1; 3782 3783 total += x; 3784 nodes[c->node] += x; 3785 } 3786 per_cpu[c->node]++; 3787 } 3788 } 3789 3790 if (flags & SO_ALL) { 3791 for_each_node_state(node, N_NORMAL_MEMORY) { 3792 struct kmem_cache_node *n = get_node(s, node); 3793 3794 if (flags & SO_TOTAL) 3795 x = atomic_long_read(&n->total_objects); 3796 else if (flags & SO_OBJECTS) 3797 x = atomic_long_read(&n->total_objects) - 3798 count_partial(n, count_free); 3799 3800 else 3801 x = atomic_long_read(&n->nr_slabs); 3802 total += x; 3803 nodes[node] += x; 3804 } 3805 3806 } else if (flags & SO_PARTIAL) { 3807 for_each_node_state(node, N_NORMAL_MEMORY) { 3808 struct kmem_cache_node *n = get_node(s, node); 3809 3810 if (flags & SO_TOTAL) 3811 x = count_partial(n, count_total); 3812 else if (flags & SO_OBJECTS) 3813 x = count_partial(n, count_inuse); 3814 else 3815 x = n->nr_partial; 3816 total += x; 3817 nodes[node] += x; 3818 } 3819 } 3820 x = sprintf(buf, "%lu", total); 3821#ifdef CONFIG_NUMA 3822 for_each_node_state(node, N_NORMAL_MEMORY) 3823 if (nodes[node]) 3824 x += sprintf(buf + x, " N%d=%lu", 3825 node, nodes[node]); 3826#endif 3827 kfree(nodes); 3828 return x + sprintf(buf + x, "\n"); 3829} 3830 3831static int any_slab_objects(struct kmem_cache *s) 3832{ 3833 int node; 3834 3835 for_each_online_node(node) { 3836 struct kmem_cache_node *n = get_node(s, node); 3837 3838 if (!n) 3839 continue; 3840 3841 if (atomic_long_read(&n->total_objects)) 3842 return 1; 3843 } 3844 return 0; 3845} 3846 3847#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3848#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3849 3850struct slab_attribute { 3851 struct attribute attr; 3852 ssize_t (*show)(struct kmem_cache *s, char *buf); 3853 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3854}; 3855 3856#define SLAB_ATTR_RO(_name) \ 3857 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3858 3859#define SLAB_ATTR(_name) \ 3860 static struct slab_attribute _name##_attr = \ 3861 __ATTR(_name, 0644, _name##_show, _name##_store) 3862 3863static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3864{ 3865 return sprintf(buf, "%d\n", s->size); 3866} 3867SLAB_ATTR_RO(slab_size); 3868 3869static ssize_t align_show(struct kmem_cache *s, char *buf) 3870{ 3871 return sprintf(buf, "%d\n", s->align); 3872} 3873SLAB_ATTR_RO(align); 3874 3875static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3876{ 3877 return sprintf(buf, "%d\n", s->objsize); 3878} 3879SLAB_ATTR_RO(object_size); 3880 3881static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3882{ 3883 return sprintf(buf, "%d\n", oo_objects(s->oo)); 3884} 3885SLAB_ATTR_RO(objs_per_slab); 3886 3887static ssize_t order_store(struct kmem_cache *s, 3888 const char *buf, size_t length) 3889{ 3890 unsigned long order; 3891 int err; 3892 3893 err = strict_strtoul(buf, 10, &order); 3894 if (err) 3895 return err; 3896 3897 if (order > slub_max_order || order < slub_min_order) 3898 return -EINVAL; 3899 3900 calculate_sizes(s, order); 3901 return length; 3902} 3903 3904static ssize_t order_show(struct kmem_cache *s, char *buf) 3905{ 3906 return sprintf(buf, "%d\n", oo_order(s->oo)); 3907} 3908SLAB_ATTR(order); 3909 3910static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 3911{ 3912 return sprintf(buf, "%lu\n", s->min_partial); 3913} 3914 3915static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 3916 size_t length) 3917{ 3918 unsigned long min; 3919 int err; 3920 3921 err = strict_strtoul(buf, 10, &min); 3922 if (err) 3923 return err; 3924 3925 set_min_partial(s, min); 3926 return length; 3927} 3928SLAB_ATTR(min_partial); 3929 3930static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3931{ 3932 if (s->ctor) { 3933 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3934 3935 return n + sprintf(buf + n, "\n"); 3936 } 3937 return 0; 3938} 3939SLAB_ATTR_RO(ctor); 3940 3941static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3942{ 3943 return sprintf(buf, "%d\n", s->refcount - 1); 3944} 3945SLAB_ATTR_RO(aliases); 3946 3947static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3948{ 3949 return show_slab_objects(s, buf, SO_ALL); 3950} 3951SLAB_ATTR_RO(slabs); 3952 3953static ssize_t partial_show(struct kmem_cache *s, char *buf) 3954{ 3955 return show_slab_objects(s, buf, SO_PARTIAL); 3956} 3957SLAB_ATTR_RO(partial); 3958 3959static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 3960{ 3961 return show_slab_objects(s, buf, SO_CPU); 3962} 3963SLAB_ATTR_RO(cpu_slabs); 3964 3965static ssize_t objects_show(struct kmem_cache *s, char *buf) 3966{ 3967 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 3968} 3969SLAB_ATTR_RO(objects); 3970 3971static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 3972{ 3973 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 3974} 3975SLAB_ATTR_RO(objects_partial); 3976 3977static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 3978{ 3979 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 3980} 3981SLAB_ATTR_RO(total_objects); 3982 3983static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 3984{ 3985 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 3986} 3987 3988static ssize_t sanity_checks_store(struct kmem_cache *s, 3989 const char *buf, size_t length) 3990{ 3991 s->flags &= ~SLAB_DEBUG_FREE; 3992 if (buf[0] == '1') 3993 s->flags |= SLAB_DEBUG_FREE; 3994 return length; 3995} 3996SLAB_ATTR(sanity_checks); 3997 3998static ssize_t trace_show(struct kmem_cache *s, char *buf) 3999{ 4000 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4001} 4002 4003static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4004 size_t length) 4005{ 4006 s->flags &= ~SLAB_TRACE; 4007 if (buf[0] == '1') 4008 s->flags |= SLAB_TRACE; 4009 return length; 4010} 4011SLAB_ATTR(trace); 4012 4013#ifdef CONFIG_FAILSLAB 4014static ssize_t failslab_show(struct kmem_cache *s, char *buf) 4015{ 4016 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 4017} 4018 4019static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 4020 size_t length) 4021{ 4022 s->flags &= ~SLAB_FAILSLAB; 4023 if (buf[0] == '1') 4024 s->flags |= SLAB_FAILSLAB; 4025 return length; 4026} 4027SLAB_ATTR(failslab); 4028#endif 4029 4030static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4031{ 4032 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4033} 4034 4035static ssize_t reclaim_account_store(struct kmem_cache *s, 4036 const char *buf, size_t length) 4037{ 4038 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4039 if (buf[0] == '1') 4040 s->flags |= SLAB_RECLAIM_ACCOUNT; 4041 return length; 4042} 4043SLAB_ATTR(reclaim_account); 4044 4045static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4046{ 4047 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4048} 4049SLAB_ATTR_RO(hwcache_align); 4050 4051#ifdef CONFIG_ZONE_DMA 4052static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4053{ 4054 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4055} 4056SLAB_ATTR_RO(cache_dma); 4057#endif 4058 4059static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4060{ 4061 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4062} 4063SLAB_ATTR_RO(destroy_by_rcu); 4064 4065static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4066{ 4067 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4068} 4069 4070static ssize_t red_zone_store(struct kmem_cache *s, 4071 const char *buf, size_t length) 4072{ 4073 if (any_slab_objects(s)) 4074 return -EBUSY; 4075 4076 s->flags &= ~SLAB_RED_ZONE; 4077 if (buf[0] == '1') 4078 s->flags |= SLAB_RED_ZONE; 4079 calculate_sizes(s, -1); 4080 return length; 4081} 4082SLAB_ATTR(red_zone); 4083 4084static ssize_t poison_show(struct kmem_cache *s, char *buf) 4085{ 4086 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4087} 4088 4089static ssize_t poison_store(struct kmem_cache *s, 4090 const char *buf, size_t length) 4091{ 4092 if (any_slab_objects(s)) 4093 return -EBUSY; 4094 4095 s->flags &= ~SLAB_POISON; 4096 if (buf[0] == '1') 4097 s->flags |= SLAB_POISON; 4098 calculate_sizes(s, -1); 4099 return length; 4100} 4101SLAB_ATTR(poison); 4102 4103static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4104{ 4105 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4106} 4107 4108static ssize_t store_user_store(struct kmem_cache *s, 4109 const char *buf, size_t length) 4110{ 4111 if (any_slab_objects(s)) 4112 return -EBUSY; 4113 4114 s->flags &= ~SLAB_STORE_USER; 4115 if (buf[0] == '1') 4116 s->flags |= SLAB_STORE_USER; 4117 calculate_sizes(s, -1); 4118 return length; 4119} 4120SLAB_ATTR(store_user); 4121 4122static ssize_t validate_show(struct kmem_cache *s, char *buf) 4123{ 4124 return 0; 4125} 4126 4127static ssize_t validate_store(struct kmem_cache *s, 4128 const char *buf, size_t length) 4129{ 4130 int ret = -EINVAL; 4131 4132 if (buf[0] == '1') { 4133 ret = validate_slab_cache(s); 4134 if (ret >= 0) 4135 ret = length; 4136 } 4137 return ret; 4138} 4139SLAB_ATTR(validate); 4140 4141static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4142{ 4143 return 0; 4144} 4145 4146static ssize_t shrink_store(struct kmem_cache *s, 4147 const char *buf, size_t length) 4148{ 4149 if (buf[0] == '1') { 4150 int rc = kmem_cache_shrink(s); 4151 4152 if (rc) 4153 return rc; 4154 } else 4155 return -EINVAL; 4156 return length; 4157} 4158SLAB_ATTR(shrink); 4159 4160static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4161{ 4162 if (!(s->flags & SLAB_STORE_USER)) 4163 return -ENOSYS; 4164 return list_locations(s, buf, TRACK_ALLOC); 4165} 4166SLAB_ATTR_RO(alloc_calls); 4167 4168static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4169{ 4170 if (!(s->flags & SLAB_STORE_USER)) 4171 return -ENOSYS; 4172 return list_locations(s, buf, TRACK_FREE); 4173} 4174SLAB_ATTR_RO(free_calls); 4175 4176#ifdef CONFIG_NUMA 4177static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4178{ 4179 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4180} 4181 4182static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4183 const char *buf, size_t length) 4184{ 4185 unsigned long ratio; 4186 int err; 4187 4188 err = strict_strtoul(buf, 10, &ratio); 4189 if (err) 4190 return err; 4191 4192 if (ratio <= 100) 4193 s->remote_node_defrag_ratio = ratio * 10; 4194 4195 return length; 4196} 4197SLAB_ATTR(remote_node_defrag_ratio); 4198#endif 4199 4200#ifdef CONFIG_SLUB_STATS 4201static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4202{ 4203 unsigned long sum = 0; 4204 int cpu; 4205 int len; 4206 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4207 4208 if (!data) 4209 return -ENOMEM; 4210 4211 for_each_online_cpu(cpu) { 4212 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 4213 4214 data[cpu] = x; 4215 sum += x; 4216 } 4217 4218 len = sprintf(buf, "%lu", sum); 4219 4220#ifdef CONFIG_SMP 4221 for_each_online_cpu(cpu) { 4222 if (data[cpu] && len < PAGE_SIZE - 20) 4223 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4224 } 4225#endif 4226 kfree(data); 4227 return len + sprintf(buf + len, "\n"); 4228} 4229 4230static void clear_stat(struct kmem_cache *s, enum stat_item si) 4231{ 4232 int cpu; 4233 4234 for_each_online_cpu(cpu) 4235 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 4236} 4237 4238#define STAT_ATTR(si, text) \ 4239static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4240{ \ 4241 return show_stat(s, buf, si); \ 4242} \ 4243static ssize_t text##_store(struct kmem_cache *s, \ 4244 const char *buf, size_t length) \ 4245{ \ 4246 if (buf[0] != '0') \ 4247 return -EINVAL; \ 4248 clear_stat(s, si); \ 4249 return length; \ 4250} \ 4251SLAB_ATTR(text); \ 4252 4253STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4254STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4255STAT_ATTR(FREE_FASTPATH, free_fastpath); 4256STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4257STAT_ATTR(FREE_FROZEN, free_frozen); 4258STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4259STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4260STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4261STAT_ATTR(ALLOC_SLAB, alloc_slab); 4262STAT_ATTR(ALLOC_REFILL, alloc_refill); 4263STAT_ATTR(FREE_SLAB, free_slab); 4264STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4265STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4266STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4267STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4268STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4269STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4270STAT_ATTR(ORDER_FALLBACK, order_fallback); 4271#endif 4272 4273static struct attribute *slab_attrs[] = { 4274 &slab_size_attr.attr, 4275 &object_size_attr.attr, 4276 &objs_per_slab_attr.attr, 4277 &order_attr.attr, 4278 &min_partial_attr.attr, 4279 &objects_attr.attr, 4280 &objects_partial_attr.attr, 4281 &total_objects_attr.attr, 4282 &slabs_attr.attr, 4283 &partial_attr.attr, 4284 &cpu_slabs_attr.attr, 4285 &ctor_attr.attr, 4286 &aliases_attr.attr, 4287 &align_attr.attr, 4288 &sanity_checks_attr.attr, 4289 &trace_attr.attr, 4290 &hwcache_align_attr.attr, 4291 &reclaim_account_attr.attr, 4292 &destroy_by_rcu_attr.attr, 4293 &red_zone_attr.attr, 4294 &poison_attr.attr, 4295 &store_user_attr.attr, 4296 &validate_attr.attr, 4297 &shrink_attr.attr, 4298 &alloc_calls_attr.attr, 4299 &free_calls_attr.attr, 4300#ifdef CONFIG_ZONE_DMA 4301 &cache_dma_attr.attr, 4302#endif 4303#ifdef CONFIG_NUMA 4304 &remote_node_defrag_ratio_attr.attr, 4305#endif 4306#ifdef CONFIG_SLUB_STATS 4307 &alloc_fastpath_attr.attr, 4308 &alloc_slowpath_attr.attr, 4309 &free_fastpath_attr.attr, 4310 &free_slowpath_attr.attr, 4311 &free_frozen_attr.attr, 4312 &free_add_partial_attr.attr, 4313 &free_remove_partial_attr.attr, 4314 &alloc_from_partial_attr.attr, 4315 &alloc_slab_attr.attr, 4316 &alloc_refill_attr.attr, 4317 &free_slab_attr.attr, 4318 &cpuslab_flush_attr.attr, 4319 &deactivate_full_attr.attr, 4320 &deactivate_empty_attr.attr, 4321 &deactivate_to_head_attr.attr, 4322 &deactivate_to_tail_attr.attr, 4323 &deactivate_remote_frees_attr.attr, 4324 &order_fallback_attr.attr, 4325#endif 4326#ifdef CONFIG_FAILSLAB 4327 &failslab_attr.attr, 4328#endif 4329 4330 NULL 4331}; 4332 4333static struct attribute_group slab_attr_group = { 4334 .attrs = slab_attrs, 4335}; 4336 4337static ssize_t slab_attr_show(struct kobject *kobj, 4338 struct attribute *attr, 4339 char *buf) 4340{ 4341 struct slab_attribute *attribute; 4342 struct kmem_cache *s; 4343 int err; 4344 4345 attribute = to_slab_attr(attr); 4346 s = to_slab(kobj); 4347 4348 if (!attribute->show) 4349 return -EIO; 4350 4351 err = attribute->show(s, buf); 4352 4353 return err; 4354} 4355 4356static ssize_t slab_attr_store(struct kobject *kobj, 4357 struct attribute *attr, 4358 const char *buf, size_t len) 4359{ 4360 struct slab_attribute *attribute; 4361 struct kmem_cache *s; 4362 int err; 4363 4364 attribute = to_slab_attr(attr); 4365 s = to_slab(kobj); 4366 4367 if (!attribute->store) 4368 return -EIO; 4369 4370 err = attribute->store(s, buf, len); 4371 4372 return err; 4373} 4374 4375static void kmem_cache_release(struct kobject *kobj) 4376{ 4377 struct kmem_cache *s = to_slab(kobj); 4378 4379 kfree(s); 4380} 4381 4382static const struct sysfs_ops slab_sysfs_ops = { 4383 .show = slab_attr_show, 4384 .store = slab_attr_store, 4385}; 4386 4387static struct kobj_type slab_ktype = { 4388 .sysfs_ops = &slab_sysfs_ops, 4389 .release = kmem_cache_release 4390}; 4391 4392static int uevent_filter(struct kset *kset, struct kobject *kobj) 4393{ 4394 struct kobj_type *ktype = get_ktype(kobj); 4395 4396 if (ktype == &slab_ktype) 4397 return 1; 4398 return 0; 4399} 4400 4401static const struct kset_uevent_ops slab_uevent_ops = { 4402 .filter = uevent_filter, 4403}; 4404 4405static struct kset *slab_kset; 4406 4407#define ID_STR_LENGTH 64 4408 4409/* Create a unique string id for a slab cache: 4410 * 4411 * Format :[flags-]size 4412 */ 4413static char *create_unique_id(struct kmem_cache *s) 4414{ 4415 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4416 char *p = name; 4417 4418 BUG_ON(!name); 4419 4420 *p++ = ':'; 4421 /* 4422 * First flags affecting slabcache operations. We will only 4423 * get here for aliasable slabs so we do not need to support 4424 * too many flags. The flags here must cover all flags that 4425 * are matched during merging to guarantee that the id is 4426 * unique. 4427 */ 4428 if (s->flags & SLAB_CACHE_DMA) 4429 *p++ = 'd'; 4430 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4431 *p++ = 'a'; 4432 if (s->flags & SLAB_DEBUG_FREE) 4433 *p++ = 'F'; 4434 if (!(s->flags & SLAB_NOTRACK)) 4435 *p++ = 't'; 4436 if (p != name + 1) 4437 *p++ = '-'; 4438 p += sprintf(p, "%07d", s->size); 4439 BUG_ON(p > name + ID_STR_LENGTH - 1); 4440 return name; 4441} 4442 4443static int sysfs_slab_add(struct kmem_cache *s) 4444{ 4445 int err; 4446 const char *name; 4447 int unmergeable; 4448 4449 if (slab_state < SYSFS) 4450 /* Defer until later */ 4451 return 0; 4452 4453 unmergeable = slab_unmergeable(s); 4454 if (unmergeable) { 4455 /* 4456 * Slabcache can never be merged so we can use the name proper. 4457 * This is typically the case for debug situations. In that 4458 * case we can catch duplicate names easily. 4459 */ 4460 sysfs_remove_link(&slab_kset->kobj, s->name); 4461 name = s->name; 4462 } else { 4463 /* 4464 * Create a unique name for the slab as a target 4465 * for the symlinks. 4466 */ 4467 name = create_unique_id(s); 4468 } 4469 4470 s->kobj.kset = slab_kset; 4471 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4472 if (err) { 4473 kobject_put(&s->kobj); 4474 return err; 4475 } 4476 4477 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4478 if (err) { 4479 kobject_del(&s->kobj); 4480 kobject_put(&s->kobj); 4481 return err; 4482 } 4483 kobject_uevent(&s->kobj, KOBJ_ADD); 4484 if (!unmergeable) { 4485 /* Setup first alias */ 4486 sysfs_slab_alias(s, s->name); 4487 kfree(name); 4488 } 4489 return 0; 4490} 4491 4492static void sysfs_slab_remove(struct kmem_cache *s) 4493{ 4494 if (slab_state < SYSFS) 4495 /* 4496 * Sysfs has not been setup yet so no need to remove the 4497 * cache from sysfs. 4498 */ 4499 return; 4500 4501 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4502 kobject_del(&s->kobj); 4503 kobject_put(&s->kobj); 4504} 4505 4506/* 4507 * Need to buffer aliases during bootup until sysfs becomes 4508 * available lest we lose that information. 4509 */ 4510struct saved_alias { 4511 struct kmem_cache *s; 4512 const char *name; 4513 struct saved_alias *next; 4514}; 4515 4516static struct saved_alias *alias_list; 4517 4518static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4519{ 4520 struct saved_alias *al; 4521 4522 if (slab_state == SYSFS) { 4523 /* 4524 * If we have a leftover link then remove it. 4525 */ 4526 sysfs_remove_link(&slab_kset->kobj, name); 4527 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4528 } 4529 4530 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4531 if (!al) 4532 return -ENOMEM; 4533 4534 al->s = s; 4535 al->name = name; 4536 al->next = alias_list; 4537 alias_list = al; 4538 return 0; 4539} 4540 4541static int __init slab_sysfs_init(void) 4542{ 4543 struct kmem_cache *s; 4544 int err; 4545 4546 down_write(&slub_lock); 4547 4548 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4549 if (!slab_kset) { 4550 up_write(&slub_lock); 4551 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4552 return -ENOSYS; 4553 } 4554 4555 slab_state = SYSFS; 4556 4557 list_for_each_entry(s, &slab_caches, list) { 4558 err = sysfs_slab_add(s); 4559 if (err) 4560 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4561 " to sysfs\n", s->name); 4562 } 4563 4564 while (alias_list) { 4565 struct saved_alias *al = alias_list; 4566 4567 alias_list = alias_list->next; 4568 err = sysfs_slab_alias(al->s, al->name); 4569 if (err) 4570 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4571 " %s to sysfs\n", s->name); 4572 kfree(al); 4573 } 4574 4575 up_write(&slub_lock); 4576 resiliency_test(); 4577 return 0; 4578} 4579 4580__initcall(slab_sysfs_init); 4581#endif 4582 4583/* 4584 * The /proc/slabinfo ABI 4585 */ 4586#ifdef CONFIG_SLABINFO 4587static void print_slabinfo_header(struct seq_file *m) 4588{ 4589 seq_puts(m, "slabinfo - version: 2.1\n"); 4590 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4591 "<objperslab> <pagesperslab>"); 4592 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4593 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4594 seq_putc(m, '\n'); 4595} 4596 4597static void *s_start(struct seq_file *m, loff_t *pos) 4598{ 4599 loff_t n = *pos; 4600 4601 down_read(&slub_lock); 4602 if (!n) 4603 print_slabinfo_header(m); 4604 4605 return seq_list_start(&slab_caches, *pos); 4606} 4607 4608static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4609{ 4610 return seq_list_next(p, &slab_caches, pos); 4611} 4612 4613static void s_stop(struct seq_file *m, void *p) 4614{ 4615 up_read(&slub_lock); 4616} 4617 4618static int s_show(struct seq_file *m, void *p) 4619{ 4620 unsigned long nr_partials = 0; 4621 unsigned long nr_slabs = 0; 4622 unsigned long nr_inuse = 0; 4623 unsigned long nr_objs = 0; 4624 unsigned long nr_free = 0; 4625 struct kmem_cache *s; 4626 int node; 4627 4628 s = list_entry(p, struct kmem_cache, list); 4629 4630 for_each_online_node(node) { 4631 struct kmem_cache_node *n = get_node(s, node); 4632 4633 if (!n) 4634 continue; 4635 4636 nr_partials += n->nr_partial; 4637 nr_slabs += atomic_long_read(&n->nr_slabs); 4638 nr_objs += atomic_long_read(&n->total_objects); 4639 nr_free += count_partial(n, count_free); 4640 } 4641 4642 nr_inuse = nr_objs - nr_free; 4643 4644 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4645 nr_objs, s->size, oo_objects(s->oo), 4646 (1 << oo_order(s->oo))); 4647 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4648 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4649 0UL); 4650 seq_putc(m, '\n'); 4651 return 0; 4652} 4653 4654static const struct seq_operations slabinfo_op = { 4655 .start = s_start, 4656 .next = s_next, 4657 .stop = s_stop, 4658 .show = s_show, 4659}; 4660 4661static int slabinfo_open(struct inode *inode, struct file *file) 4662{ 4663 return seq_open(file, &slabinfo_op); 4664} 4665 4666static const struct file_operations proc_slabinfo_operations = { 4667 .open = slabinfo_open, 4668 .read = seq_read, 4669 .llseek = seq_lseek, 4670 .release = seq_release, 4671}; 4672 4673static int __init slab_proc_init(void) 4674{ 4675 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations); 4676 return 0; 4677} 4678module_init(slab_proc_init); 4679#endif /* CONFIG_SLABINFO */ 4680