slub.c revision 319d1e240683d37924ea8977c91730c3393fd453
1/* 2 * SLUB: A slab allocator that limits cache line use instead of queuing 3 * objects in per cpu and per node lists. 4 * 5 * The allocator synchronizes using per slab locks and only 6 * uses a centralized lock to manage a pool of partial slabs. 7 * 8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com> 9 */ 10 11#include <linux/mm.h> 12#include <linux/module.h> 13#include <linux/bit_spinlock.h> 14#include <linux/interrupt.h> 15#include <linux/bitops.h> 16#include <linux/slab.h> 17#include <linux/seq_file.h> 18#include <linux/cpu.h> 19#include <linux/cpuset.h> 20#include <linux/mempolicy.h> 21#include <linux/ctype.h> 22#include <linux/kallsyms.h> 23#include <linux/memory.h> 24 25/* 26 * Lock order: 27 * 1. slab_lock(page) 28 * 2. slab->list_lock 29 * 30 * The slab_lock protects operations on the object of a particular 31 * slab and its metadata in the page struct. If the slab lock 32 * has been taken then no allocations nor frees can be performed 33 * on the objects in the slab nor can the slab be added or removed 34 * from the partial or full lists since this would mean modifying 35 * the page_struct of the slab. 36 * 37 * The list_lock protects the partial and full list on each node and 38 * the partial slab counter. If taken then no new slabs may be added or 39 * removed from the lists nor make the number of partial slabs be modified. 40 * (Note that the total number of slabs is an atomic value that may be 41 * modified without taking the list lock). 42 * 43 * The list_lock is a centralized lock and thus we avoid taking it as 44 * much as possible. As long as SLUB does not have to handle partial 45 * slabs, operations can continue without any centralized lock. F.e. 46 * allocating a long series of objects that fill up slabs does not require 47 * the list lock. 48 * 49 * The lock order is sometimes inverted when we are trying to get a slab 50 * off a list. We take the list_lock and then look for a page on the list 51 * to use. While we do that objects in the slabs may be freed. We can 52 * only operate on the slab if we have also taken the slab_lock. So we use 53 * a slab_trylock() on the slab. If trylock was successful then no frees 54 * can occur anymore and we can use the slab for allocations etc. If the 55 * slab_trylock() does not succeed then frees are in progress in the slab and 56 * we must stay away from it for a while since we may cause a bouncing 57 * cacheline if we try to acquire the lock. So go onto the next slab. 58 * If all pages are busy then we may allocate a new slab instead of reusing 59 * a partial slab. A new slab has noone operating on it and thus there is 60 * no danger of cacheline contention. 61 * 62 * Interrupts are disabled during allocation and deallocation in order to 63 * make the slab allocator safe to use in the context of an irq. In addition 64 * interrupts are disabled to ensure that the processor does not change 65 * while handling per_cpu slabs, due to kernel preemption. 66 * 67 * SLUB assigns one slab for allocation to each processor. 68 * Allocations only occur from these slabs called cpu slabs. 69 * 70 * Slabs with free elements are kept on a partial list and during regular 71 * operations no list for full slabs is used. If an object in a full slab is 72 * freed then the slab will show up again on the partial lists. 73 * We track full slabs for debugging purposes though because otherwise we 74 * cannot scan all objects. 75 * 76 * Slabs are freed when they become empty. Teardown and setup is 77 * minimal so we rely on the page allocators per cpu caches for 78 * fast frees and allocs. 79 * 80 * Overloading of page flags that are otherwise used for LRU management. 81 * 82 * PageActive The slab is frozen and exempt from list processing. 83 * This means that the slab is dedicated to a purpose 84 * such as satisfying allocations for a specific 85 * processor. Objects may be freed in the slab while 86 * it is frozen but slab_free will then skip the usual 87 * list operations. It is up to the processor holding 88 * the slab to integrate the slab into the slab lists 89 * when the slab is no longer needed. 90 * 91 * One use of this flag is to mark slabs that are 92 * used for allocations. Then such a slab becomes a cpu 93 * slab. The cpu slab may be equipped with an additional 94 * freelist that allows lockless access to 95 * free objects in addition to the regular freelist 96 * that requires the slab lock. 97 * 98 * PageError Slab requires special handling due to debug 99 * options set. This moves slab handling out of 100 * the fast path and disables lockless freelists. 101 */ 102 103#define FROZEN (1 << PG_active) 104 105#ifdef CONFIG_SLUB_DEBUG 106#define SLABDEBUG (1 << PG_error) 107#else 108#define SLABDEBUG 0 109#endif 110 111static inline int SlabFrozen(struct page *page) 112{ 113 return page->flags & FROZEN; 114} 115 116static inline void SetSlabFrozen(struct page *page) 117{ 118 page->flags |= FROZEN; 119} 120 121static inline void ClearSlabFrozen(struct page *page) 122{ 123 page->flags &= ~FROZEN; 124} 125 126static inline int SlabDebug(struct page *page) 127{ 128 return page->flags & SLABDEBUG; 129} 130 131static inline void SetSlabDebug(struct page *page) 132{ 133 page->flags |= SLABDEBUG; 134} 135 136static inline void ClearSlabDebug(struct page *page) 137{ 138 page->flags &= ~SLABDEBUG; 139} 140 141/* 142 * Issues still to be resolved: 143 * 144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 145 * 146 * - Variable sizing of the per node arrays 147 */ 148 149/* Enable to test recovery from slab corruption on boot */ 150#undef SLUB_RESILIENCY_TEST 151 152#if PAGE_SHIFT <= 12 153 154/* 155 * Small page size. Make sure that we do not fragment memory 156 */ 157#define DEFAULT_MAX_ORDER 1 158#define DEFAULT_MIN_OBJECTS 4 159 160#else 161 162/* 163 * Large page machines are customarily able to handle larger 164 * page orders. 165 */ 166#define DEFAULT_MAX_ORDER 2 167#define DEFAULT_MIN_OBJECTS 8 168 169#endif 170 171/* 172 * Mininum number of partial slabs. These will be left on the partial 173 * lists even if they are empty. kmem_cache_shrink may reclaim them. 174 */ 175#define MIN_PARTIAL 5 176 177/* 178 * Maximum number of desirable partial slabs. 179 * The existence of more partial slabs makes kmem_cache_shrink 180 * sort the partial list by the number of objects in the. 181 */ 182#define MAX_PARTIAL 10 183 184#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 185 SLAB_POISON | SLAB_STORE_USER) 186 187/* 188 * Set of flags that will prevent slab merging 189 */ 190#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 191 SLAB_TRACE | SLAB_DESTROY_BY_RCU) 192 193#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 194 SLAB_CACHE_DMA) 195 196#ifndef ARCH_KMALLOC_MINALIGN 197#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) 198#endif 199 200#ifndef ARCH_SLAB_MINALIGN 201#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long) 202#endif 203 204/* Internal SLUB flags */ 205#define __OBJECT_POISON 0x80000000 /* Poison object */ 206#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */ 207 208/* Not all arches define cache_line_size */ 209#ifndef cache_line_size 210#define cache_line_size() L1_CACHE_BYTES 211#endif 212 213static int kmem_size = sizeof(struct kmem_cache); 214 215#ifdef CONFIG_SMP 216static struct notifier_block slab_notifier; 217#endif 218 219static enum { 220 DOWN, /* No slab functionality available */ 221 PARTIAL, /* kmem_cache_open() works but kmalloc does not */ 222 UP, /* Everything works but does not show up in sysfs */ 223 SYSFS /* Sysfs up */ 224} slab_state = DOWN; 225 226/* A list of all slab caches on the system */ 227static DECLARE_RWSEM(slub_lock); 228static LIST_HEAD(slab_caches); 229 230/* 231 * Tracking user of a slab. 232 */ 233struct track { 234 void *addr; /* Called from address */ 235 int cpu; /* Was running on cpu */ 236 int pid; /* Pid context */ 237 unsigned long when; /* When did the operation occur */ 238}; 239 240enum track_item { TRACK_ALLOC, TRACK_FREE }; 241 242#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) 243static int sysfs_slab_add(struct kmem_cache *); 244static int sysfs_slab_alias(struct kmem_cache *, const char *); 245static void sysfs_slab_remove(struct kmem_cache *); 246 247#else 248static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 249static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 250 { return 0; } 251static inline void sysfs_slab_remove(struct kmem_cache *s) 252{ 253 kfree(s); 254} 255 256#endif 257 258static inline void stat(struct kmem_cache_cpu *c, enum stat_item si) 259{ 260#ifdef CONFIG_SLUB_STATS 261 c->stat[si]++; 262#endif 263} 264 265/******************************************************************** 266 * Core slab cache functions 267 *******************************************************************/ 268 269int slab_is_available(void) 270{ 271 return slab_state >= UP; 272} 273 274static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 275{ 276#ifdef CONFIG_NUMA 277 return s->node[node]; 278#else 279 return &s->local_node; 280#endif 281} 282 283static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu) 284{ 285#ifdef CONFIG_SMP 286 return s->cpu_slab[cpu]; 287#else 288 return &s->cpu_slab; 289#endif 290} 291 292/* Verify that a pointer has an address that is valid within a slab page */ 293static inline int check_valid_pointer(struct kmem_cache *s, 294 struct page *page, const void *object) 295{ 296 void *base; 297 298 if (!object) 299 return 1; 300 301 base = page_address(page); 302 if (object < base || object >= base + page->objects * s->size || 303 (object - base) % s->size) { 304 return 0; 305 } 306 307 return 1; 308} 309 310/* 311 * Slow version of get and set free pointer. 312 * 313 * This version requires touching the cache lines of kmem_cache which 314 * we avoid to do in the fast alloc free paths. There we obtain the offset 315 * from the page struct. 316 */ 317static inline void *get_freepointer(struct kmem_cache *s, void *object) 318{ 319 return *(void **)(object + s->offset); 320} 321 322static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 323{ 324 *(void **)(object + s->offset) = fp; 325} 326 327/* Loop over all objects in a slab */ 328#define for_each_object(__p, __s, __addr, __objects) \ 329 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 330 __p += (__s)->size) 331 332/* Scan freelist */ 333#define for_each_free_object(__p, __s, __free) \ 334 for (__p = (__free); __p; __p = get_freepointer((__s), __p)) 335 336/* Determine object index from a given position */ 337static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 338{ 339 return (p - addr) / s->size; 340} 341 342static inline struct kmem_cache_order_objects oo_make(int order, 343 unsigned long size) 344{ 345 struct kmem_cache_order_objects x = { 346 (order << 16) + (PAGE_SIZE << order) / size 347 }; 348 349 return x; 350} 351 352static inline int oo_order(struct kmem_cache_order_objects x) 353{ 354 return x.x >> 16; 355} 356 357static inline int oo_objects(struct kmem_cache_order_objects x) 358{ 359 return x.x & ((1 << 16) - 1); 360} 361 362#ifdef CONFIG_SLUB_DEBUG 363/* 364 * Debug settings: 365 */ 366#ifdef CONFIG_SLUB_DEBUG_ON 367static int slub_debug = DEBUG_DEFAULT_FLAGS; 368#else 369static int slub_debug; 370#endif 371 372static char *slub_debug_slabs; 373 374/* 375 * Object debugging 376 */ 377static void print_section(char *text, u8 *addr, unsigned int length) 378{ 379 int i, offset; 380 int newline = 1; 381 char ascii[17]; 382 383 ascii[16] = 0; 384 385 for (i = 0; i < length; i++) { 386 if (newline) { 387 printk(KERN_ERR "%8s 0x%p: ", text, addr + i); 388 newline = 0; 389 } 390 printk(KERN_CONT " %02x", addr[i]); 391 offset = i % 16; 392 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.'; 393 if (offset == 15) { 394 printk(KERN_CONT " %s\n", ascii); 395 newline = 1; 396 } 397 } 398 if (!newline) { 399 i %= 16; 400 while (i < 16) { 401 printk(KERN_CONT " "); 402 ascii[i] = ' '; 403 i++; 404 } 405 printk(KERN_CONT " %s\n", ascii); 406 } 407} 408 409static struct track *get_track(struct kmem_cache *s, void *object, 410 enum track_item alloc) 411{ 412 struct track *p; 413 414 if (s->offset) 415 p = object + s->offset + sizeof(void *); 416 else 417 p = object + s->inuse; 418 419 return p + alloc; 420} 421 422static void set_track(struct kmem_cache *s, void *object, 423 enum track_item alloc, void *addr) 424{ 425 struct track *p; 426 427 if (s->offset) 428 p = object + s->offset + sizeof(void *); 429 else 430 p = object + s->inuse; 431 432 p += alloc; 433 if (addr) { 434 p->addr = addr; 435 p->cpu = smp_processor_id(); 436 p->pid = current ? current->pid : -1; 437 p->when = jiffies; 438 } else 439 memset(p, 0, sizeof(struct track)); 440} 441 442static void init_tracking(struct kmem_cache *s, void *object) 443{ 444 if (!(s->flags & SLAB_STORE_USER)) 445 return; 446 447 set_track(s, object, TRACK_FREE, NULL); 448 set_track(s, object, TRACK_ALLOC, NULL); 449} 450 451static void print_track(const char *s, struct track *t) 452{ 453 if (!t->addr) 454 return; 455 456 printk(KERN_ERR "INFO: %s in ", s); 457 __print_symbol("%s", (unsigned long)t->addr); 458 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid); 459} 460 461static void print_tracking(struct kmem_cache *s, void *object) 462{ 463 if (!(s->flags & SLAB_STORE_USER)) 464 return; 465 466 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 467 print_track("Freed", get_track(s, object, TRACK_FREE)); 468} 469 470static void print_page_info(struct page *page) 471{ 472 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 473 page, page->objects, page->inuse, page->freelist, page->flags); 474 475} 476 477static void slab_bug(struct kmem_cache *s, char *fmt, ...) 478{ 479 va_list args; 480 char buf[100]; 481 482 va_start(args, fmt); 483 vsnprintf(buf, sizeof(buf), fmt, args); 484 va_end(args); 485 printk(KERN_ERR "========================================" 486 "=====================================\n"); 487 printk(KERN_ERR "BUG %s: %s\n", s->name, buf); 488 printk(KERN_ERR "----------------------------------------" 489 "-------------------------------------\n\n"); 490} 491 492static void slab_fix(struct kmem_cache *s, char *fmt, ...) 493{ 494 va_list args; 495 char buf[100]; 496 497 va_start(args, fmt); 498 vsnprintf(buf, sizeof(buf), fmt, args); 499 va_end(args); 500 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 501} 502 503static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 504{ 505 unsigned int off; /* Offset of last byte */ 506 u8 *addr = page_address(page); 507 508 print_tracking(s, p); 509 510 print_page_info(page); 511 512 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 513 p, p - addr, get_freepointer(s, p)); 514 515 if (p > addr + 16) 516 print_section("Bytes b4", p - 16, 16); 517 518 print_section("Object", p, min(s->objsize, 128)); 519 520 if (s->flags & SLAB_RED_ZONE) 521 print_section("Redzone", p + s->objsize, 522 s->inuse - s->objsize); 523 524 if (s->offset) 525 off = s->offset + sizeof(void *); 526 else 527 off = s->inuse; 528 529 if (s->flags & SLAB_STORE_USER) 530 off += 2 * sizeof(struct track); 531 532 if (off != s->size) 533 /* Beginning of the filler is the free pointer */ 534 print_section("Padding", p + off, s->size - off); 535 536 dump_stack(); 537} 538 539static void object_err(struct kmem_cache *s, struct page *page, 540 u8 *object, char *reason) 541{ 542 slab_bug(s, "%s", reason); 543 print_trailer(s, page, object); 544} 545 546static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 547{ 548 va_list args; 549 char buf[100]; 550 551 va_start(args, fmt); 552 vsnprintf(buf, sizeof(buf), fmt, args); 553 va_end(args); 554 slab_bug(s, "%s", buf); 555 print_page_info(page); 556 dump_stack(); 557} 558 559static void init_object(struct kmem_cache *s, void *object, int active) 560{ 561 u8 *p = object; 562 563 if (s->flags & __OBJECT_POISON) { 564 memset(p, POISON_FREE, s->objsize - 1); 565 p[s->objsize - 1] = POISON_END; 566 } 567 568 if (s->flags & SLAB_RED_ZONE) 569 memset(p + s->objsize, 570 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE, 571 s->inuse - s->objsize); 572} 573 574static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes) 575{ 576 while (bytes) { 577 if (*start != (u8)value) 578 return start; 579 start++; 580 bytes--; 581 } 582 return NULL; 583} 584 585static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 586 void *from, void *to) 587{ 588 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 589 memset(from, data, to - from); 590} 591 592static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 593 u8 *object, char *what, 594 u8 *start, unsigned int value, unsigned int bytes) 595{ 596 u8 *fault; 597 u8 *end; 598 599 fault = check_bytes(start, value, bytes); 600 if (!fault) 601 return 1; 602 603 end = start + bytes; 604 while (end > fault && end[-1] == value) 605 end--; 606 607 slab_bug(s, "%s overwritten", what); 608 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 609 fault, end - 1, fault[0], value); 610 print_trailer(s, page, object); 611 612 restore_bytes(s, what, value, fault, end); 613 return 0; 614} 615 616/* 617 * Object layout: 618 * 619 * object address 620 * Bytes of the object to be managed. 621 * If the freepointer may overlay the object then the free 622 * pointer is the first word of the object. 623 * 624 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 625 * 0xa5 (POISON_END) 626 * 627 * object + s->objsize 628 * Padding to reach word boundary. This is also used for Redzoning. 629 * Padding is extended by another word if Redzoning is enabled and 630 * objsize == inuse. 631 * 632 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 633 * 0xcc (RED_ACTIVE) for objects in use. 634 * 635 * object + s->inuse 636 * Meta data starts here. 637 * 638 * A. Free pointer (if we cannot overwrite object on free) 639 * B. Tracking data for SLAB_STORE_USER 640 * C. Padding to reach required alignment boundary or at mininum 641 * one word if debugging is on to be able to detect writes 642 * before the word boundary. 643 * 644 * Padding is done using 0x5a (POISON_INUSE) 645 * 646 * object + s->size 647 * Nothing is used beyond s->size. 648 * 649 * If slabcaches are merged then the objsize and inuse boundaries are mostly 650 * ignored. And therefore no slab options that rely on these boundaries 651 * may be used with merged slabcaches. 652 */ 653 654static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 655{ 656 unsigned long off = s->inuse; /* The end of info */ 657 658 if (s->offset) 659 /* Freepointer is placed after the object. */ 660 off += sizeof(void *); 661 662 if (s->flags & SLAB_STORE_USER) 663 /* We also have user information there */ 664 off += 2 * sizeof(struct track); 665 666 if (s->size == off) 667 return 1; 668 669 return check_bytes_and_report(s, page, p, "Object padding", 670 p + off, POISON_INUSE, s->size - off); 671} 672 673/* Check the pad bytes at the end of a slab page */ 674static int slab_pad_check(struct kmem_cache *s, struct page *page) 675{ 676 u8 *start; 677 u8 *fault; 678 u8 *end; 679 int length; 680 int remainder; 681 682 if (!(s->flags & SLAB_POISON)) 683 return 1; 684 685 start = page_address(page); 686 length = (PAGE_SIZE << compound_order(page)); 687 end = start + length; 688 remainder = length % s->size; 689 if (!remainder) 690 return 1; 691 692 fault = check_bytes(end - remainder, POISON_INUSE, remainder); 693 if (!fault) 694 return 1; 695 while (end > fault && end[-1] == POISON_INUSE) 696 end--; 697 698 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 699 print_section("Padding", end - remainder, remainder); 700 701 restore_bytes(s, "slab padding", POISON_INUSE, start, end); 702 return 0; 703} 704 705static int check_object(struct kmem_cache *s, struct page *page, 706 void *object, int active) 707{ 708 u8 *p = object; 709 u8 *endobject = object + s->objsize; 710 711 if (s->flags & SLAB_RED_ZONE) { 712 unsigned int red = 713 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE; 714 715 if (!check_bytes_and_report(s, page, object, "Redzone", 716 endobject, red, s->inuse - s->objsize)) 717 return 0; 718 } else { 719 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 720 check_bytes_and_report(s, page, p, "Alignment padding", 721 endobject, POISON_INUSE, s->inuse - s->objsize); 722 } 723 } 724 725 if (s->flags & SLAB_POISON) { 726 if (!active && (s->flags & __OBJECT_POISON) && 727 (!check_bytes_and_report(s, page, p, "Poison", p, 728 POISON_FREE, s->objsize - 1) || 729 !check_bytes_and_report(s, page, p, "Poison", 730 p + s->objsize - 1, POISON_END, 1))) 731 return 0; 732 /* 733 * check_pad_bytes cleans up on its own. 734 */ 735 check_pad_bytes(s, page, p); 736 } 737 738 if (!s->offset && active) 739 /* 740 * Object and freepointer overlap. Cannot check 741 * freepointer while object is allocated. 742 */ 743 return 1; 744 745 /* Check free pointer validity */ 746 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 747 object_err(s, page, p, "Freepointer corrupt"); 748 /* 749 * No choice but to zap it and thus loose the remainder 750 * of the free objects in this slab. May cause 751 * another error because the object count is now wrong. 752 */ 753 set_freepointer(s, p, NULL); 754 return 0; 755 } 756 return 1; 757} 758 759static int check_slab(struct kmem_cache *s, struct page *page) 760{ 761 int maxobj; 762 763 VM_BUG_ON(!irqs_disabled()); 764 765 if (!PageSlab(page)) { 766 slab_err(s, page, "Not a valid slab page"); 767 return 0; 768 } 769 770 maxobj = (PAGE_SIZE << compound_order(page)) / s->size; 771 if (page->objects > maxobj) { 772 slab_err(s, page, "objects %u > max %u", 773 s->name, page->objects, maxobj); 774 return 0; 775 } 776 if (page->inuse > page->objects) { 777 slab_err(s, page, "inuse %u > max %u", 778 s->name, page->inuse, page->objects); 779 return 0; 780 } 781 /* Slab_pad_check fixes things up after itself */ 782 slab_pad_check(s, page); 783 return 1; 784} 785 786/* 787 * Determine if a certain object on a page is on the freelist. Must hold the 788 * slab lock to guarantee that the chains are in a consistent state. 789 */ 790static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 791{ 792 int nr = 0; 793 void *fp = page->freelist; 794 void *object = NULL; 795 unsigned long max_objects; 796 797 while (fp && nr <= page->objects) { 798 if (fp == search) 799 return 1; 800 if (!check_valid_pointer(s, page, fp)) { 801 if (object) { 802 object_err(s, page, object, 803 "Freechain corrupt"); 804 set_freepointer(s, object, NULL); 805 break; 806 } else { 807 slab_err(s, page, "Freepointer corrupt"); 808 page->freelist = NULL; 809 page->inuse = page->objects; 810 slab_fix(s, "Freelist cleared"); 811 return 0; 812 } 813 break; 814 } 815 object = fp; 816 fp = get_freepointer(s, object); 817 nr++; 818 } 819 820 max_objects = (PAGE_SIZE << compound_order(page)) / s->size; 821 if (max_objects > 65535) 822 max_objects = 65535; 823 824 if (page->objects != max_objects) { 825 slab_err(s, page, "Wrong number of objects. Found %d but " 826 "should be %d", page->objects, max_objects); 827 page->objects = max_objects; 828 slab_fix(s, "Number of objects adjusted."); 829 } 830 if (page->inuse != page->objects - nr) { 831 slab_err(s, page, "Wrong object count. Counter is %d but " 832 "counted were %d", page->inuse, page->objects - nr); 833 page->inuse = page->objects - nr; 834 slab_fix(s, "Object count adjusted."); 835 } 836 return search == NULL; 837} 838 839static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc) 840{ 841 if (s->flags & SLAB_TRACE) { 842 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 843 s->name, 844 alloc ? "alloc" : "free", 845 object, page->inuse, 846 page->freelist); 847 848 if (!alloc) 849 print_section("Object", (void *)object, s->objsize); 850 851 dump_stack(); 852 } 853} 854 855/* 856 * Tracking of fully allocated slabs for debugging purposes. 857 */ 858static void add_full(struct kmem_cache_node *n, struct page *page) 859{ 860 spin_lock(&n->list_lock); 861 list_add(&page->lru, &n->full); 862 spin_unlock(&n->list_lock); 863} 864 865static void remove_full(struct kmem_cache *s, struct page *page) 866{ 867 struct kmem_cache_node *n; 868 869 if (!(s->flags & SLAB_STORE_USER)) 870 return; 871 872 n = get_node(s, page_to_nid(page)); 873 874 spin_lock(&n->list_lock); 875 list_del(&page->lru); 876 spin_unlock(&n->list_lock); 877} 878 879/* Tracking of the number of slabs for debugging purposes */ 880static inline unsigned long slabs_node(struct kmem_cache *s, int node) 881{ 882 struct kmem_cache_node *n = get_node(s, node); 883 884 return atomic_long_read(&n->nr_slabs); 885} 886 887static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 888{ 889 struct kmem_cache_node *n = get_node(s, node); 890 891 /* 892 * May be called early in order to allocate a slab for the 893 * kmem_cache_node structure. Solve the chicken-egg 894 * dilemma by deferring the increment of the count during 895 * bootstrap (see early_kmem_cache_node_alloc). 896 */ 897 if (!NUMA_BUILD || n) { 898 atomic_long_inc(&n->nr_slabs); 899 atomic_long_add(objects, &n->total_objects); 900 } 901} 902static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 903{ 904 struct kmem_cache_node *n = get_node(s, node); 905 906 atomic_long_dec(&n->nr_slabs); 907 atomic_long_sub(objects, &n->total_objects); 908} 909 910/* Object debug checks for alloc/free paths */ 911static void setup_object_debug(struct kmem_cache *s, struct page *page, 912 void *object) 913{ 914 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 915 return; 916 917 init_object(s, object, 0); 918 init_tracking(s, object); 919} 920 921static int alloc_debug_processing(struct kmem_cache *s, struct page *page, 922 void *object, void *addr) 923{ 924 if (!check_slab(s, page)) 925 goto bad; 926 927 if (!on_freelist(s, page, object)) { 928 object_err(s, page, object, "Object already allocated"); 929 goto bad; 930 } 931 932 if (!check_valid_pointer(s, page, object)) { 933 object_err(s, page, object, "Freelist Pointer check fails"); 934 goto bad; 935 } 936 937 if (!check_object(s, page, object, 0)) 938 goto bad; 939 940 /* Success perform special debug activities for allocs */ 941 if (s->flags & SLAB_STORE_USER) 942 set_track(s, object, TRACK_ALLOC, addr); 943 trace(s, page, object, 1); 944 init_object(s, object, 1); 945 return 1; 946 947bad: 948 if (PageSlab(page)) { 949 /* 950 * If this is a slab page then lets do the best we can 951 * to avoid issues in the future. Marking all objects 952 * as used avoids touching the remaining objects. 953 */ 954 slab_fix(s, "Marking all objects used"); 955 page->inuse = page->objects; 956 page->freelist = NULL; 957 } 958 return 0; 959} 960 961static int free_debug_processing(struct kmem_cache *s, struct page *page, 962 void *object, void *addr) 963{ 964 if (!check_slab(s, page)) 965 goto fail; 966 967 if (!check_valid_pointer(s, page, object)) { 968 slab_err(s, page, "Invalid object pointer 0x%p", object); 969 goto fail; 970 } 971 972 if (on_freelist(s, page, object)) { 973 object_err(s, page, object, "Object already free"); 974 goto fail; 975 } 976 977 if (!check_object(s, page, object, 1)) 978 return 0; 979 980 if (unlikely(s != page->slab)) { 981 if (!PageSlab(page)) { 982 slab_err(s, page, "Attempt to free object(0x%p) " 983 "outside of slab", object); 984 } else if (!page->slab) { 985 printk(KERN_ERR 986 "SLUB <none>: no slab for object 0x%p.\n", 987 object); 988 dump_stack(); 989 } else 990 object_err(s, page, object, 991 "page slab pointer corrupt."); 992 goto fail; 993 } 994 995 /* Special debug activities for freeing objects */ 996 if (!SlabFrozen(page) && !page->freelist) 997 remove_full(s, page); 998 if (s->flags & SLAB_STORE_USER) 999 set_track(s, object, TRACK_FREE, addr); 1000 trace(s, page, object, 0); 1001 init_object(s, object, 0); 1002 return 1; 1003 1004fail: 1005 slab_fix(s, "Object at 0x%p not freed", object); 1006 return 0; 1007} 1008 1009static int __init setup_slub_debug(char *str) 1010{ 1011 slub_debug = DEBUG_DEFAULT_FLAGS; 1012 if (*str++ != '=' || !*str) 1013 /* 1014 * No options specified. Switch on full debugging. 1015 */ 1016 goto out; 1017 1018 if (*str == ',') 1019 /* 1020 * No options but restriction on slabs. This means full 1021 * debugging for slabs matching a pattern. 1022 */ 1023 goto check_slabs; 1024 1025 slub_debug = 0; 1026 if (*str == '-') 1027 /* 1028 * Switch off all debugging measures. 1029 */ 1030 goto out; 1031 1032 /* 1033 * Determine which debug features should be switched on 1034 */ 1035 for (; *str && *str != ','; str++) { 1036 switch (tolower(*str)) { 1037 case 'f': 1038 slub_debug |= SLAB_DEBUG_FREE; 1039 break; 1040 case 'z': 1041 slub_debug |= SLAB_RED_ZONE; 1042 break; 1043 case 'p': 1044 slub_debug |= SLAB_POISON; 1045 break; 1046 case 'u': 1047 slub_debug |= SLAB_STORE_USER; 1048 break; 1049 case 't': 1050 slub_debug |= SLAB_TRACE; 1051 break; 1052 default: 1053 printk(KERN_ERR "slub_debug option '%c' " 1054 "unknown. skipped\n", *str); 1055 } 1056 } 1057 1058check_slabs: 1059 if (*str == ',') 1060 slub_debug_slabs = str + 1; 1061out: 1062 return 1; 1063} 1064 1065__setup("slub_debug", setup_slub_debug); 1066 1067static unsigned long kmem_cache_flags(unsigned long objsize, 1068 unsigned long flags, const char *name, 1069 void (*ctor)(struct kmem_cache *, void *)) 1070{ 1071 /* 1072 * Enable debugging if selected on the kernel commandline. 1073 */ 1074 if (slub_debug && (!slub_debug_slabs || 1075 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0)) 1076 flags |= slub_debug; 1077 1078 return flags; 1079} 1080#else 1081static inline void setup_object_debug(struct kmem_cache *s, 1082 struct page *page, void *object) {} 1083 1084static inline int alloc_debug_processing(struct kmem_cache *s, 1085 struct page *page, void *object, void *addr) { return 0; } 1086 1087static inline int free_debug_processing(struct kmem_cache *s, 1088 struct page *page, void *object, void *addr) { return 0; } 1089 1090static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1091 { return 1; } 1092static inline int check_object(struct kmem_cache *s, struct page *page, 1093 void *object, int active) { return 1; } 1094static inline void add_full(struct kmem_cache_node *n, struct page *page) {} 1095static inline unsigned long kmem_cache_flags(unsigned long objsize, 1096 unsigned long flags, const char *name, 1097 void (*ctor)(struct kmem_cache *, void *)) 1098{ 1099 return flags; 1100} 1101#define slub_debug 0 1102 1103static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1104 { return 0; } 1105static inline void inc_slabs_node(struct kmem_cache *s, int node, 1106 int objects) {} 1107static inline void dec_slabs_node(struct kmem_cache *s, int node, 1108 int objects) {} 1109#endif 1110 1111/* 1112 * Slab allocation and freeing 1113 */ 1114static inline struct page *alloc_slab_page(gfp_t flags, int node, 1115 struct kmem_cache_order_objects oo) 1116{ 1117 int order = oo_order(oo); 1118 1119 if (node == -1) 1120 return alloc_pages(flags, order); 1121 else 1122 return alloc_pages_node(node, flags, order); 1123} 1124 1125static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1126{ 1127 struct page *page; 1128 struct kmem_cache_order_objects oo = s->oo; 1129 1130 flags |= s->allocflags; 1131 1132 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node, 1133 oo); 1134 if (unlikely(!page)) { 1135 oo = s->min; 1136 /* 1137 * Allocation may have failed due to fragmentation. 1138 * Try a lower order alloc if possible 1139 */ 1140 page = alloc_slab_page(flags, node, oo); 1141 if (!page) 1142 return NULL; 1143 1144 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK); 1145 } 1146 page->objects = oo_objects(oo); 1147 mod_zone_page_state(page_zone(page), 1148 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1149 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1150 1 << oo_order(oo)); 1151 1152 return page; 1153} 1154 1155static void setup_object(struct kmem_cache *s, struct page *page, 1156 void *object) 1157{ 1158 setup_object_debug(s, page, object); 1159 if (unlikely(s->ctor)) 1160 s->ctor(s, object); 1161} 1162 1163static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1164{ 1165 struct page *page; 1166 void *start; 1167 void *last; 1168 void *p; 1169 1170 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1171 1172 page = allocate_slab(s, 1173 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1174 if (!page) 1175 goto out; 1176 1177 inc_slabs_node(s, page_to_nid(page), page->objects); 1178 page->slab = s; 1179 page->flags |= 1 << PG_slab; 1180 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON | 1181 SLAB_STORE_USER | SLAB_TRACE)) 1182 SetSlabDebug(page); 1183 1184 start = page_address(page); 1185 1186 if (unlikely(s->flags & SLAB_POISON)) 1187 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1188 1189 last = start; 1190 for_each_object(p, s, start, page->objects) { 1191 setup_object(s, page, last); 1192 set_freepointer(s, last, p); 1193 last = p; 1194 } 1195 setup_object(s, page, last); 1196 set_freepointer(s, last, NULL); 1197 1198 page->freelist = start; 1199 page->inuse = 0; 1200out: 1201 return page; 1202} 1203 1204static void __free_slab(struct kmem_cache *s, struct page *page) 1205{ 1206 int order = compound_order(page); 1207 int pages = 1 << order; 1208 1209 if (unlikely(SlabDebug(page))) { 1210 void *p; 1211 1212 slab_pad_check(s, page); 1213 for_each_object(p, s, page_address(page), 1214 page->objects) 1215 check_object(s, page, p, 0); 1216 ClearSlabDebug(page); 1217 } 1218 1219 mod_zone_page_state(page_zone(page), 1220 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1221 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1222 -pages); 1223 1224 __ClearPageSlab(page); 1225 reset_page_mapcount(page); 1226 __free_pages(page, order); 1227} 1228 1229static void rcu_free_slab(struct rcu_head *h) 1230{ 1231 struct page *page; 1232 1233 page = container_of((struct list_head *)h, struct page, lru); 1234 __free_slab(page->slab, page); 1235} 1236 1237static void free_slab(struct kmem_cache *s, struct page *page) 1238{ 1239 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1240 /* 1241 * RCU free overloads the RCU head over the LRU 1242 */ 1243 struct rcu_head *head = (void *)&page->lru; 1244 1245 call_rcu(head, rcu_free_slab); 1246 } else 1247 __free_slab(s, page); 1248} 1249 1250static void discard_slab(struct kmem_cache *s, struct page *page) 1251{ 1252 dec_slabs_node(s, page_to_nid(page), page->objects); 1253 free_slab(s, page); 1254} 1255 1256/* 1257 * Per slab locking using the pagelock 1258 */ 1259static __always_inline void slab_lock(struct page *page) 1260{ 1261 bit_spin_lock(PG_locked, &page->flags); 1262} 1263 1264static __always_inline void slab_unlock(struct page *page) 1265{ 1266 __bit_spin_unlock(PG_locked, &page->flags); 1267} 1268 1269static __always_inline int slab_trylock(struct page *page) 1270{ 1271 int rc = 1; 1272 1273 rc = bit_spin_trylock(PG_locked, &page->flags); 1274 return rc; 1275} 1276 1277/* 1278 * Management of partially allocated slabs 1279 */ 1280static void add_partial(struct kmem_cache_node *n, 1281 struct page *page, int tail) 1282{ 1283 spin_lock(&n->list_lock); 1284 n->nr_partial++; 1285 if (tail) 1286 list_add_tail(&page->lru, &n->partial); 1287 else 1288 list_add(&page->lru, &n->partial); 1289 spin_unlock(&n->list_lock); 1290} 1291 1292static void remove_partial(struct kmem_cache *s, 1293 struct page *page) 1294{ 1295 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1296 1297 spin_lock(&n->list_lock); 1298 list_del(&page->lru); 1299 n->nr_partial--; 1300 spin_unlock(&n->list_lock); 1301} 1302 1303/* 1304 * Lock slab and remove from the partial list. 1305 * 1306 * Must hold list_lock. 1307 */ 1308static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page) 1309{ 1310 if (slab_trylock(page)) { 1311 list_del(&page->lru); 1312 n->nr_partial--; 1313 SetSlabFrozen(page); 1314 return 1; 1315 } 1316 return 0; 1317} 1318 1319/* 1320 * Try to allocate a partial slab from a specific node. 1321 */ 1322static struct page *get_partial_node(struct kmem_cache_node *n) 1323{ 1324 struct page *page; 1325 1326 /* 1327 * Racy check. If we mistakenly see no partial slabs then we 1328 * just allocate an empty slab. If we mistakenly try to get a 1329 * partial slab and there is none available then get_partials() 1330 * will return NULL. 1331 */ 1332 if (!n || !n->nr_partial) 1333 return NULL; 1334 1335 spin_lock(&n->list_lock); 1336 list_for_each_entry(page, &n->partial, lru) 1337 if (lock_and_freeze_slab(n, page)) 1338 goto out; 1339 page = NULL; 1340out: 1341 spin_unlock(&n->list_lock); 1342 return page; 1343} 1344 1345/* 1346 * Get a page from somewhere. Search in increasing NUMA distances. 1347 */ 1348static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags) 1349{ 1350#ifdef CONFIG_NUMA 1351 struct zonelist *zonelist; 1352 struct zone **z; 1353 struct page *page; 1354 1355 /* 1356 * The defrag ratio allows a configuration of the tradeoffs between 1357 * inter node defragmentation and node local allocations. A lower 1358 * defrag_ratio increases the tendency to do local allocations 1359 * instead of attempting to obtain partial slabs from other nodes. 1360 * 1361 * If the defrag_ratio is set to 0 then kmalloc() always 1362 * returns node local objects. If the ratio is higher then kmalloc() 1363 * may return off node objects because partial slabs are obtained 1364 * from other nodes and filled up. 1365 * 1366 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1367 * defrag_ratio = 1000) then every (well almost) allocation will 1368 * first attempt to defrag slab caches on other nodes. This means 1369 * scanning over all nodes to look for partial slabs which may be 1370 * expensive if we do it every time we are trying to find a slab 1371 * with available objects. 1372 */ 1373 if (!s->remote_node_defrag_ratio || 1374 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1375 return NULL; 1376 1377 zonelist = &NODE_DATA( 1378 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)]; 1379 for (z = zonelist->zones; *z; z++) { 1380 struct kmem_cache_node *n; 1381 1382 n = get_node(s, zone_to_nid(*z)); 1383 1384 if (n && cpuset_zone_allowed_hardwall(*z, flags) && 1385 n->nr_partial > MIN_PARTIAL) { 1386 page = get_partial_node(n); 1387 if (page) 1388 return page; 1389 } 1390 } 1391#endif 1392 return NULL; 1393} 1394 1395/* 1396 * Get a partial page, lock it and return it. 1397 */ 1398static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node) 1399{ 1400 struct page *page; 1401 int searchnode = (node == -1) ? numa_node_id() : node; 1402 1403 page = get_partial_node(get_node(s, searchnode)); 1404 if (page || (flags & __GFP_THISNODE)) 1405 return page; 1406 1407 return get_any_partial(s, flags); 1408} 1409 1410/* 1411 * Move a page back to the lists. 1412 * 1413 * Must be called with the slab lock held. 1414 * 1415 * On exit the slab lock will have been dropped. 1416 */ 1417static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail) 1418{ 1419 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1420 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id()); 1421 1422 ClearSlabFrozen(page); 1423 if (page->inuse) { 1424 1425 if (page->freelist) { 1426 add_partial(n, page, tail); 1427 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD); 1428 } else { 1429 stat(c, DEACTIVATE_FULL); 1430 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER)) 1431 add_full(n, page); 1432 } 1433 slab_unlock(page); 1434 } else { 1435 stat(c, DEACTIVATE_EMPTY); 1436 if (n->nr_partial < MIN_PARTIAL) { 1437 /* 1438 * Adding an empty slab to the partial slabs in order 1439 * to avoid page allocator overhead. This slab needs 1440 * to come after the other slabs with objects in 1441 * so that the others get filled first. That way the 1442 * size of the partial list stays small. 1443 * 1444 * kmem_cache_shrink can reclaim any empty slabs from the 1445 * partial list. 1446 */ 1447 add_partial(n, page, 1); 1448 slab_unlock(page); 1449 } else { 1450 slab_unlock(page); 1451 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB); 1452 discard_slab(s, page); 1453 } 1454 } 1455} 1456 1457/* 1458 * Remove the cpu slab 1459 */ 1460static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1461{ 1462 struct page *page = c->page; 1463 int tail = 1; 1464 1465 if (page->freelist) 1466 stat(c, DEACTIVATE_REMOTE_FREES); 1467 /* 1468 * Merge cpu freelist into slab freelist. Typically we get here 1469 * because both freelists are empty. So this is unlikely 1470 * to occur. 1471 */ 1472 while (unlikely(c->freelist)) { 1473 void **object; 1474 1475 tail = 0; /* Hot objects. Put the slab first */ 1476 1477 /* Retrieve object from cpu_freelist */ 1478 object = c->freelist; 1479 c->freelist = c->freelist[c->offset]; 1480 1481 /* And put onto the regular freelist */ 1482 object[c->offset] = page->freelist; 1483 page->freelist = object; 1484 page->inuse--; 1485 } 1486 c->page = NULL; 1487 unfreeze_slab(s, page, tail); 1488} 1489 1490static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1491{ 1492 stat(c, CPUSLAB_FLUSH); 1493 slab_lock(c->page); 1494 deactivate_slab(s, c); 1495} 1496 1497/* 1498 * Flush cpu slab. 1499 * 1500 * Called from IPI handler with interrupts disabled. 1501 */ 1502static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 1503{ 1504 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 1505 1506 if (likely(c && c->page)) 1507 flush_slab(s, c); 1508} 1509 1510static void flush_cpu_slab(void *d) 1511{ 1512 struct kmem_cache *s = d; 1513 1514 __flush_cpu_slab(s, smp_processor_id()); 1515} 1516 1517static void flush_all(struct kmem_cache *s) 1518{ 1519#ifdef CONFIG_SMP 1520 on_each_cpu(flush_cpu_slab, s, 1, 1); 1521#else 1522 unsigned long flags; 1523 1524 local_irq_save(flags); 1525 flush_cpu_slab(s); 1526 local_irq_restore(flags); 1527#endif 1528} 1529 1530/* 1531 * Check if the objects in a per cpu structure fit numa 1532 * locality expectations. 1533 */ 1534static inline int node_match(struct kmem_cache_cpu *c, int node) 1535{ 1536#ifdef CONFIG_NUMA 1537 if (node != -1 && c->node != node) 1538 return 0; 1539#endif 1540 return 1; 1541} 1542 1543/* 1544 * Slow path. The lockless freelist is empty or we need to perform 1545 * debugging duties. 1546 * 1547 * Interrupts are disabled. 1548 * 1549 * Processing is still very fast if new objects have been freed to the 1550 * regular freelist. In that case we simply take over the regular freelist 1551 * as the lockless freelist and zap the regular freelist. 1552 * 1553 * If that is not working then we fall back to the partial lists. We take the 1554 * first element of the freelist as the object to allocate now and move the 1555 * rest of the freelist to the lockless freelist. 1556 * 1557 * And if we were unable to get a new slab from the partial slab lists then 1558 * we need to allocate a new slab. This is the slowest path since it involves 1559 * a call to the page allocator and the setup of a new slab. 1560 */ 1561static void *__slab_alloc(struct kmem_cache *s, 1562 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c) 1563{ 1564 void **object; 1565 struct page *new; 1566 1567 /* We handle __GFP_ZERO in the caller */ 1568 gfpflags &= ~__GFP_ZERO; 1569 1570 if (!c->page) 1571 goto new_slab; 1572 1573 slab_lock(c->page); 1574 if (unlikely(!node_match(c, node))) 1575 goto another_slab; 1576 1577 stat(c, ALLOC_REFILL); 1578 1579load_freelist: 1580 object = c->page->freelist; 1581 if (unlikely(!object)) 1582 goto another_slab; 1583 if (unlikely(SlabDebug(c->page))) 1584 goto debug; 1585 1586 c->freelist = object[c->offset]; 1587 c->page->inuse = c->page->objects; 1588 c->page->freelist = NULL; 1589 c->node = page_to_nid(c->page); 1590unlock_out: 1591 slab_unlock(c->page); 1592 stat(c, ALLOC_SLOWPATH); 1593 return object; 1594 1595another_slab: 1596 deactivate_slab(s, c); 1597 1598new_slab: 1599 new = get_partial(s, gfpflags, node); 1600 if (new) { 1601 c->page = new; 1602 stat(c, ALLOC_FROM_PARTIAL); 1603 goto load_freelist; 1604 } 1605 1606 if (gfpflags & __GFP_WAIT) 1607 local_irq_enable(); 1608 1609 new = new_slab(s, gfpflags, node); 1610 1611 if (gfpflags & __GFP_WAIT) 1612 local_irq_disable(); 1613 1614 if (new) { 1615 c = get_cpu_slab(s, smp_processor_id()); 1616 stat(c, ALLOC_SLAB); 1617 if (c->page) 1618 flush_slab(s, c); 1619 slab_lock(new); 1620 SetSlabFrozen(new); 1621 c->page = new; 1622 goto load_freelist; 1623 } 1624 return NULL; 1625debug: 1626 if (!alloc_debug_processing(s, c->page, object, addr)) 1627 goto another_slab; 1628 1629 c->page->inuse++; 1630 c->page->freelist = object[c->offset]; 1631 c->node = -1; 1632 goto unlock_out; 1633} 1634 1635/* 1636 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 1637 * have the fastpath folded into their functions. So no function call 1638 * overhead for requests that can be satisfied on the fastpath. 1639 * 1640 * The fastpath works by first checking if the lockless freelist can be used. 1641 * If not then __slab_alloc is called for slow processing. 1642 * 1643 * Otherwise we can simply pick the next object from the lockless free list. 1644 */ 1645static __always_inline void *slab_alloc(struct kmem_cache *s, 1646 gfp_t gfpflags, int node, void *addr) 1647{ 1648 void **object; 1649 struct kmem_cache_cpu *c; 1650 unsigned long flags; 1651 1652 local_irq_save(flags); 1653 c = get_cpu_slab(s, smp_processor_id()); 1654 if (unlikely(!c->freelist || !node_match(c, node))) 1655 1656 object = __slab_alloc(s, gfpflags, node, addr, c); 1657 1658 else { 1659 object = c->freelist; 1660 c->freelist = object[c->offset]; 1661 stat(c, ALLOC_FASTPATH); 1662 } 1663 local_irq_restore(flags); 1664 1665 if (unlikely((gfpflags & __GFP_ZERO) && object)) 1666 memset(object, 0, c->objsize); 1667 1668 return object; 1669} 1670 1671void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 1672{ 1673 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0)); 1674} 1675EXPORT_SYMBOL(kmem_cache_alloc); 1676 1677#ifdef CONFIG_NUMA 1678void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 1679{ 1680 return slab_alloc(s, gfpflags, node, __builtin_return_address(0)); 1681} 1682EXPORT_SYMBOL(kmem_cache_alloc_node); 1683#endif 1684 1685/* 1686 * Slow patch handling. This may still be called frequently since objects 1687 * have a longer lifetime than the cpu slabs in most processing loads. 1688 * 1689 * So we still attempt to reduce cache line usage. Just take the slab 1690 * lock and free the item. If there is no additional partial page 1691 * handling required then we can return immediately. 1692 */ 1693static void __slab_free(struct kmem_cache *s, struct page *page, 1694 void *x, void *addr, unsigned int offset) 1695{ 1696 void *prior; 1697 void **object = (void *)x; 1698 struct kmem_cache_cpu *c; 1699 1700 c = get_cpu_slab(s, raw_smp_processor_id()); 1701 stat(c, FREE_SLOWPATH); 1702 slab_lock(page); 1703 1704 if (unlikely(SlabDebug(page))) 1705 goto debug; 1706 1707checks_ok: 1708 prior = object[offset] = page->freelist; 1709 page->freelist = object; 1710 page->inuse--; 1711 1712 if (unlikely(SlabFrozen(page))) { 1713 stat(c, FREE_FROZEN); 1714 goto out_unlock; 1715 } 1716 1717 if (unlikely(!page->inuse)) 1718 goto slab_empty; 1719 1720 /* 1721 * Objects left in the slab. If it was not on the partial list before 1722 * then add it. 1723 */ 1724 if (unlikely(!prior)) { 1725 add_partial(get_node(s, page_to_nid(page)), page, 1); 1726 stat(c, FREE_ADD_PARTIAL); 1727 } 1728 1729out_unlock: 1730 slab_unlock(page); 1731 return; 1732 1733slab_empty: 1734 if (prior) { 1735 /* 1736 * Slab still on the partial list. 1737 */ 1738 remove_partial(s, page); 1739 stat(c, FREE_REMOVE_PARTIAL); 1740 } 1741 slab_unlock(page); 1742 stat(c, FREE_SLAB); 1743 discard_slab(s, page); 1744 return; 1745 1746debug: 1747 if (!free_debug_processing(s, page, x, addr)) 1748 goto out_unlock; 1749 goto checks_ok; 1750} 1751 1752/* 1753 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 1754 * can perform fastpath freeing without additional function calls. 1755 * 1756 * The fastpath is only possible if we are freeing to the current cpu slab 1757 * of this processor. This typically the case if we have just allocated 1758 * the item before. 1759 * 1760 * If fastpath is not possible then fall back to __slab_free where we deal 1761 * with all sorts of special processing. 1762 */ 1763static __always_inline void slab_free(struct kmem_cache *s, 1764 struct page *page, void *x, void *addr) 1765{ 1766 void **object = (void *)x; 1767 struct kmem_cache_cpu *c; 1768 unsigned long flags; 1769 1770 local_irq_save(flags); 1771 c = get_cpu_slab(s, smp_processor_id()); 1772 debug_check_no_locks_freed(object, c->objsize); 1773 if (likely(page == c->page && c->node >= 0)) { 1774 object[c->offset] = c->freelist; 1775 c->freelist = object; 1776 stat(c, FREE_FASTPATH); 1777 } else 1778 __slab_free(s, page, x, addr, c->offset); 1779 1780 local_irq_restore(flags); 1781} 1782 1783void kmem_cache_free(struct kmem_cache *s, void *x) 1784{ 1785 struct page *page; 1786 1787 page = virt_to_head_page(x); 1788 1789 slab_free(s, page, x, __builtin_return_address(0)); 1790} 1791EXPORT_SYMBOL(kmem_cache_free); 1792 1793/* Figure out on which slab object the object resides */ 1794static struct page *get_object_page(const void *x) 1795{ 1796 struct page *page = virt_to_head_page(x); 1797 1798 if (!PageSlab(page)) 1799 return NULL; 1800 1801 return page; 1802} 1803 1804/* 1805 * Object placement in a slab is made very easy because we always start at 1806 * offset 0. If we tune the size of the object to the alignment then we can 1807 * get the required alignment by putting one properly sized object after 1808 * another. 1809 * 1810 * Notice that the allocation order determines the sizes of the per cpu 1811 * caches. Each processor has always one slab available for allocations. 1812 * Increasing the allocation order reduces the number of times that slabs 1813 * must be moved on and off the partial lists and is therefore a factor in 1814 * locking overhead. 1815 */ 1816 1817/* 1818 * Mininum / Maximum order of slab pages. This influences locking overhead 1819 * and slab fragmentation. A higher order reduces the number of partial slabs 1820 * and increases the number of allocations possible without having to 1821 * take the list_lock. 1822 */ 1823static int slub_min_order; 1824static int slub_max_order = DEFAULT_MAX_ORDER; 1825static int slub_min_objects = DEFAULT_MIN_OBJECTS; 1826 1827/* 1828 * Merge control. If this is set then no merging of slab caches will occur. 1829 * (Could be removed. This was introduced to pacify the merge skeptics.) 1830 */ 1831static int slub_nomerge; 1832 1833/* 1834 * Calculate the order of allocation given an slab object size. 1835 * 1836 * The order of allocation has significant impact on performance and other 1837 * system components. Generally order 0 allocations should be preferred since 1838 * order 0 does not cause fragmentation in the page allocator. Larger objects 1839 * be problematic to put into order 0 slabs because there may be too much 1840 * unused space left. We go to a higher order if more than 1/8th of the slab 1841 * would be wasted. 1842 * 1843 * In order to reach satisfactory performance we must ensure that a minimum 1844 * number of objects is in one slab. Otherwise we may generate too much 1845 * activity on the partial lists which requires taking the list_lock. This is 1846 * less a concern for large slabs though which are rarely used. 1847 * 1848 * slub_max_order specifies the order where we begin to stop considering the 1849 * number of objects in a slab as critical. If we reach slub_max_order then 1850 * we try to keep the page order as low as possible. So we accept more waste 1851 * of space in favor of a small page order. 1852 * 1853 * Higher order allocations also allow the placement of more objects in a 1854 * slab and thereby reduce object handling overhead. If the user has 1855 * requested a higher mininum order then we start with that one instead of 1856 * the smallest order which will fit the object. 1857 */ 1858static inline int slab_order(int size, int min_objects, 1859 int max_order, int fract_leftover) 1860{ 1861 int order; 1862 int rem; 1863 int min_order = slub_min_order; 1864 1865 if ((PAGE_SIZE << min_order) / size > 65535) 1866 return get_order(size * 65535) - 1; 1867 1868 for (order = max(min_order, 1869 fls(min_objects * size - 1) - PAGE_SHIFT); 1870 order <= max_order; order++) { 1871 1872 unsigned long slab_size = PAGE_SIZE << order; 1873 1874 if (slab_size < min_objects * size) 1875 continue; 1876 1877 rem = slab_size % size; 1878 1879 if (rem <= slab_size / fract_leftover) 1880 break; 1881 1882 } 1883 1884 return order; 1885} 1886 1887static inline int calculate_order(int size) 1888{ 1889 int order; 1890 int min_objects; 1891 int fraction; 1892 1893 /* 1894 * Attempt to find best configuration for a slab. This 1895 * works by first attempting to generate a layout with 1896 * the best configuration and backing off gradually. 1897 * 1898 * First we reduce the acceptable waste in a slab. Then 1899 * we reduce the minimum objects required in a slab. 1900 */ 1901 min_objects = slub_min_objects; 1902 while (min_objects > 1) { 1903 fraction = 8; 1904 while (fraction >= 4) { 1905 order = slab_order(size, min_objects, 1906 slub_max_order, fraction); 1907 if (order <= slub_max_order) 1908 return order; 1909 fraction /= 2; 1910 } 1911 min_objects /= 2; 1912 } 1913 1914 /* 1915 * We were unable to place multiple objects in a slab. Now 1916 * lets see if we can place a single object there. 1917 */ 1918 order = slab_order(size, 1, slub_max_order, 1); 1919 if (order <= slub_max_order) 1920 return order; 1921 1922 /* 1923 * Doh this slab cannot be placed using slub_max_order. 1924 */ 1925 order = slab_order(size, 1, MAX_ORDER, 1); 1926 if (order <= MAX_ORDER) 1927 return order; 1928 return -ENOSYS; 1929} 1930 1931/* 1932 * Figure out what the alignment of the objects will be. 1933 */ 1934static unsigned long calculate_alignment(unsigned long flags, 1935 unsigned long align, unsigned long size) 1936{ 1937 /* 1938 * If the user wants hardware cache aligned objects then follow that 1939 * suggestion if the object is sufficiently large. 1940 * 1941 * The hardware cache alignment cannot override the specified 1942 * alignment though. If that is greater then use it. 1943 */ 1944 if (flags & SLAB_HWCACHE_ALIGN) { 1945 unsigned long ralign = cache_line_size(); 1946 while (size <= ralign / 2) 1947 ralign /= 2; 1948 align = max(align, ralign); 1949 } 1950 1951 if (align < ARCH_SLAB_MINALIGN) 1952 align = ARCH_SLAB_MINALIGN; 1953 1954 return ALIGN(align, sizeof(void *)); 1955} 1956 1957static void init_kmem_cache_cpu(struct kmem_cache *s, 1958 struct kmem_cache_cpu *c) 1959{ 1960 c->page = NULL; 1961 c->freelist = NULL; 1962 c->node = 0; 1963 c->offset = s->offset / sizeof(void *); 1964 c->objsize = s->objsize; 1965#ifdef CONFIG_SLUB_STATS 1966 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned)); 1967#endif 1968} 1969 1970static void init_kmem_cache_node(struct kmem_cache_node *n) 1971{ 1972 n->nr_partial = 0; 1973 spin_lock_init(&n->list_lock); 1974 INIT_LIST_HEAD(&n->partial); 1975#ifdef CONFIG_SLUB_DEBUG 1976 atomic_long_set(&n->nr_slabs, 0); 1977 INIT_LIST_HEAD(&n->full); 1978#endif 1979} 1980 1981#ifdef CONFIG_SMP 1982/* 1983 * Per cpu array for per cpu structures. 1984 * 1985 * The per cpu array places all kmem_cache_cpu structures from one processor 1986 * close together meaning that it becomes possible that multiple per cpu 1987 * structures are contained in one cacheline. This may be particularly 1988 * beneficial for the kmalloc caches. 1989 * 1990 * A desktop system typically has around 60-80 slabs. With 100 here we are 1991 * likely able to get per cpu structures for all caches from the array defined 1992 * here. We must be able to cover all kmalloc caches during bootstrap. 1993 * 1994 * If the per cpu array is exhausted then fall back to kmalloc 1995 * of individual cachelines. No sharing is possible then. 1996 */ 1997#define NR_KMEM_CACHE_CPU 100 1998 1999static DEFINE_PER_CPU(struct kmem_cache_cpu, 2000 kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; 2001 2002static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); 2003static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE; 2004 2005static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s, 2006 int cpu, gfp_t flags) 2007{ 2008 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu); 2009 2010 if (c) 2011 per_cpu(kmem_cache_cpu_free, cpu) = 2012 (void *)c->freelist; 2013 else { 2014 /* Table overflow: So allocate ourselves */ 2015 c = kmalloc_node( 2016 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()), 2017 flags, cpu_to_node(cpu)); 2018 if (!c) 2019 return NULL; 2020 } 2021 2022 init_kmem_cache_cpu(s, c); 2023 return c; 2024} 2025 2026static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu) 2027{ 2028 if (c < per_cpu(kmem_cache_cpu, cpu) || 2029 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) { 2030 kfree(c); 2031 return; 2032 } 2033 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu); 2034 per_cpu(kmem_cache_cpu_free, cpu) = c; 2035} 2036 2037static void free_kmem_cache_cpus(struct kmem_cache *s) 2038{ 2039 int cpu; 2040 2041 for_each_online_cpu(cpu) { 2042 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 2043 2044 if (c) { 2045 s->cpu_slab[cpu] = NULL; 2046 free_kmem_cache_cpu(c, cpu); 2047 } 2048 } 2049} 2050 2051static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2052{ 2053 int cpu; 2054 2055 for_each_online_cpu(cpu) { 2056 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 2057 2058 if (c) 2059 continue; 2060 2061 c = alloc_kmem_cache_cpu(s, cpu, flags); 2062 if (!c) { 2063 free_kmem_cache_cpus(s); 2064 return 0; 2065 } 2066 s->cpu_slab[cpu] = c; 2067 } 2068 return 1; 2069} 2070 2071/* 2072 * Initialize the per cpu array. 2073 */ 2074static void init_alloc_cpu_cpu(int cpu) 2075{ 2076 int i; 2077 2078 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once)) 2079 return; 2080 2081 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--) 2082 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu); 2083 2084 cpu_set(cpu, kmem_cach_cpu_free_init_once); 2085} 2086 2087static void __init init_alloc_cpu(void) 2088{ 2089 int cpu; 2090 2091 for_each_online_cpu(cpu) 2092 init_alloc_cpu_cpu(cpu); 2093 } 2094 2095#else 2096static inline void free_kmem_cache_cpus(struct kmem_cache *s) {} 2097static inline void init_alloc_cpu(void) {} 2098 2099static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags) 2100{ 2101 init_kmem_cache_cpu(s, &s->cpu_slab); 2102 return 1; 2103} 2104#endif 2105 2106#ifdef CONFIG_NUMA 2107/* 2108 * No kmalloc_node yet so do it by hand. We know that this is the first 2109 * slab on the node for this slabcache. There are no concurrent accesses 2110 * possible. 2111 * 2112 * Note that this function only works on the kmalloc_node_cache 2113 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2114 * memory on a fresh node that has no slab structures yet. 2115 */ 2116static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags, 2117 int node) 2118{ 2119 struct page *page; 2120 struct kmem_cache_node *n; 2121 unsigned long flags; 2122 2123 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node)); 2124 2125 page = new_slab(kmalloc_caches, gfpflags, node); 2126 2127 BUG_ON(!page); 2128 if (page_to_nid(page) != node) { 2129 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2130 "node %d\n", node); 2131 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2132 "in order to be able to continue\n"); 2133 } 2134 2135 n = page->freelist; 2136 BUG_ON(!n); 2137 page->freelist = get_freepointer(kmalloc_caches, n); 2138 page->inuse++; 2139 kmalloc_caches->node[node] = n; 2140#ifdef CONFIG_SLUB_DEBUG 2141 init_object(kmalloc_caches, n, 1); 2142 init_tracking(kmalloc_caches, n); 2143#endif 2144 init_kmem_cache_node(n); 2145 inc_slabs_node(kmalloc_caches, node, page->objects); 2146 2147 /* 2148 * lockdep requires consistent irq usage for each lock 2149 * so even though there cannot be a race this early in 2150 * the boot sequence, we still disable irqs. 2151 */ 2152 local_irq_save(flags); 2153 add_partial(n, page, 0); 2154 local_irq_restore(flags); 2155 return n; 2156} 2157 2158static void free_kmem_cache_nodes(struct kmem_cache *s) 2159{ 2160 int node; 2161 2162 for_each_node_state(node, N_NORMAL_MEMORY) { 2163 struct kmem_cache_node *n = s->node[node]; 2164 if (n && n != &s->local_node) 2165 kmem_cache_free(kmalloc_caches, n); 2166 s->node[node] = NULL; 2167 } 2168} 2169 2170static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2171{ 2172 int node; 2173 int local_node; 2174 2175 if (slab_state >= UP) 2176 local_node = page_to_nid(virt_to_page(s)); 2177 else 2178 local_node = 0; 2179 2180 for_each_node_state(node, N_NORMAL_MEMORY) { 2181 struct kmem_cache_node *n; 2182 2183 if (local_node == node) 2184 n = &s->local_node; 2185 else { 2186 if (slab_state == DOWN) { 2187 n = early_kmem_cache_node_alloc(gfpflags, 2188 node); 2189 continue; 2190 } 2191 n = kmem_cache_alloc_node(kmalloc_caches, 2192 gfpflags, node); 2193 2194 if (!n) { 2195 free_kmem_cache_nodes(s); 2196 return 0; 2197 } 2198 2199 } 2200 s->node[node] = n; 2201 init_kmem_cache_node(n); 2202 } 2203 return 1; 2204} 2205#else 2206static void free_kmem_cache_nodes(struct kmem_cache *s) 2207{ 2208} 2209 2210static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags) 2211{ 2212 init_kmem_cache_node(&s->local_node); 2213 return 1; 2214} 2215#endif 2216 2217/* 2218 * calculate_sizes() determines the order and the distribution of data within 2219 * a slab object. 2220 */ 2221static int calculate_sizes(struct kmem_cache *s) 2222{ 2223 unsigned long flags = s->flags; 2224 unsigned long size = s->objsize; 2225 unsigned long align = s->align; 2226 int order; 2227 2228 /* 2229 * Round up object size to the next word boundary. We can only 2230 * place the free pointer at word boundaries and this determines 2231 * the possible location of the free pointer. 2232 */ 2233 size = ALIGN(size, sizeof(void *)); 2234 2235#ifdef CONFIG_SLUB_DEBUG 2236 /* 2237 * Determine if we can poison the object itself. If the user of 2238 * the slab may touch the object after free or before allocation 2239 * then we should never poison the object itself. 2240 */ 2241 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2242 !s->ctor) 2243 s->flags |= __OBJECT_POISON; 2244 else 2245 s->flags &= ~__OBJECT_POISON; 2246 2247 2248 /* 2249 * If we are Redzoning then check if there is some space between the 2250 * end of the object and the free pointer. If not then add an 2251 * additional word to have some bytes to store Redzone information. 2252 */ 2253 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2254 size += sizeof(void *); 2255#endif 2256 2257 /* 2258 * With that we have determined the number of bytes in actual use 2259 * by the object. This is the potential offset to the free pointer. 2260 */ 2261 s->inuse = size; 2262 2263 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2264 s->ctor)) { 2265 /* 2266 * Relocate free pointer after the object if it is not 2267 * permitted to overwrite the first word of the object on 2268 * kmem_cache_free. 2269 * 2270 * This is the case if we do RCU, have a constructor or 2271 * destructor or are poisoning the objects. 2272 */ 2273 s->offset = size; 2274 size += sizeof(void *); 2275 } 2276 2277#ifdef CONFIG_SLUB_DEBUG 2278 if (flags & SLAB_STORE_USER) 2279 /* 2280 * Need to store information about allocs and frees after 2281 * the object. 2282 */ 2283 size += 2 * sizeof(struct track); 2284 2285 if (flags & SLAB_RED_ZONE) 2286 /* 2287 * Add some empty padding so that we can catch 2288 * overwrites from earlier objects rather than let 2289 * tracking information or the free pointer be 2290 * corrupted if an user writes before the start 2291 * of the object. 2292 */ 2293 size += sizeof(void *); 2294#endif 2295 2296 /* 2297 * Determine the alignment based on various parameters that the 2298 * user specified and the dynamic determination of cache line size 2299 * on bootup. 2300 */ 2301 align = calculate_alignment(flags, align, s->objsize); 2302 2303 /* 2304 * SLUB stores one object immediately after another beginning from 2305 * offset 0. In order to align the objects we have to simply size 2306 * each object to conform to the alignment. 2307 */ 2308 size = ALIGN(size, align); 2309 s->size = size; 2310 order = calculate_order(size); 2311 2312 if (order < 0) 2313 return 0; 2314 2315 s->allocflags = 0; 2316 if (order) 2317 s->allocflags |= __GFP_COMP; 2318 2319 if (s->flags & SLAB_CACHE_DMA) 2320 s->allocflags |= SLUB_DMA; 2321 2322 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2323 s->allocflags |= __GFP_RECLAIMABLE; 2324 2325 /* 2326 * Determine the number of objects per slab 2327 */ 2328 s->oo = oo_make(order, size); 2329 s->min = oo_make(get_order(size), size); 2330 if (oo_objects(s->oo) > oo_objects(s->max)) 2331 s->max = s->oo; 2332 2333 return !!oo_objects(s->oo); 2334 2335} 2336 2337static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags, 2338 const char *name, size_t size, 2339 size_t align, unsigned long flags, 2340 void (*ctor)(struct kmem_cache *, void *)) 2341{ 2342 memset(s, 0, kmem_size); 2343 s->name = name; 2344 s->ctor = ctor; 2345 s->objsize = size; 2346 s->align = align; 2347 s->flags = kmem_cache_flags(size, flags, name, ctor); 2348 2349 if (!calculate_sizes(s)) 2350 goto error; 2351 2352 s->refcount = 1; 2353#ifdef CONFIG_NUMA 2354 s->remote_node_defrag_ratio = 100; 2355#endif 2356 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA)) 2357 goto error; 2358 2359 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA)) 2360 return 1; 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 page = get_object_page(object); 2379 2380 if (!page || s != page->slab) 2381 /* No slab or wrong slab */ 2382 return 0; 2383 2384 if (!check_valid_pointer(s, page, object)) 2385 return 0; 2386 2387 /* 2388 * We could also check if the object is on the slabs freelist. 2389 * But this would be too expensive and it seems that the main 2390 * purpose of kmem_ptr_valid() is to check if the object belongs 2391 * to a certain slab. 2392 */ 2393 return 1; 2394} 2395EXPORT_SYMBOL(kmem_ptr_validate); 2396 2397/* 2398 * Determine the size of a slab object 2399 */ 2400unsigned int kmem_cache_size(struct kmem_cache *s) 2401{ 2402 return s->objsize; 2403} 2404EXPORT_SYMBOL(kmem_cache_size); 2405 2406const char *kmem_cache_name(struct kmem_cache *s) 2407{ 2408 return s->name; 2409} 2410EXPORT_SYMBOL(kmem_cache_name); 2411 2412static void list_slab_objects(struct kmem_cache *s, struct page *page, 2413 const char *text) 2414{ 2415#ifdef CONFIG_SLUB_DEBUG 2416 void *addr = page_address(page); 2417 void *p; 2418 DECLARE_BITMAP(map, page->objects); 2419 2420 bitmap_zero(map, page->objects); 2421 slab_err(s, page, "%s", text); 2422 slab_lock(page); 2423 for_each_free_object(p, s, page->freelist) 2424 set_bit(slab_index(p, s, addr), map); 2425 2426 for_each_object(p, s, addr, page->objects) { 2427 2428 if (!test_bit(slab_index(p, s, addr), map)) { 2429 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 2430 p, p - addr); 2431 print_tracking(s, p); 2432 } 2433 } 2434 slab_unlock(page); 2435#endif 2436} 2437 2438/* 2439 * Attempt to free all partial slabs on a node. 2440 */ 2441static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 2442{ 2443 unsigned long flags; 2444 struct page *page, *h; 2445 2446 spin_lock_irqsave(&n->list_lock, flags); 2447 list_for_each_entry_safe(page, h, &n->partial, lru) { 2448 if (!page->inuse) { 2449 list_del(&page->lru); 2450 discard_slab(s, page); 2451 n->nr_partial--; 2452 } else { 2453 list_slab_objects(s, page, 2454 "Objects remaining on kmem_cache_close()"); 2455 } 2456 } 2457 spin_unlock_irqrestore(&n->list_lock, flags); 2458} 2459 2460/* 2461 * Release all resources used by a slab cache. 2462 */ 2463static inline int kmem_cache_close(struct kmem_cache *s) 2464{ 2465 int node; 2466 2467 flush_all(s); 2468 2469 /* Attempt to free all objects */ 2470 free_kmem_cache_cpus(s); 2471 for_each_node_state(node, N_NORMAL_MEMORY) { 2472 struct kmem_cache_node *n = get_node(s, node); 2473 2474 free_partial(s, n); 2475 if (n->nr_partial || slabs_node(s, node)) 2476 return 1; 2477 } 2478 free_kmem_cache_nodes(s); 2479 return 0; 2480} 2481 2482/* 2483 * Close a cache and release the kmem_cache structure 2484 * (must be used for caches created using kmem_cache_create) 2485 */ 2486void kmem_cache_destroy(struct kmem_cache *s) 2487{ 2488 down_write(&slub_lock); 2489 s->refcount--; 2490 if (!s->refcount) { 2491 list_del(&s->list); 2492 up_write(&slub_lock); 2493 if (kmem_cache_close(s)) { 2494 printk(KERN_ERR "SLUB %s: %s called for cache that " 2495 "still has objects.\n", s->name, __func__); 2496 dump_stack(); 2497 } 2498 sysfs_slab_remove(s); 2499 } else 2500 up_write(&slub_lock); 2501} 2502EXPORT_SYMBOL(kmem_cache_destroy); 2503 2504/******************************************************************** 2505 * Kmalloc subsystem 2506 *******************************************************************/ 2507 2508struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned; 2509EXPORT_SYMBOL(kmalloc_caches); 2510 2511static int __init setup_slub_min_order(char *str) 2512{ 2513 get_option(&str, &slub_min_order); 2514 2515 return 1; 2516} 2517 2518__setup("slub_min_order=", setup_slub_min_order); 2519 2520static int __init setup_slub_max_order(char *str) 2521{ 2522 get_option(&str, &slub_max_order); 2523 2524 return 1; 2525} 2526 2527__setup("slub_max_order=", setup_slub_max_order); 2528 2529static int __init setup_slub_min_objects(char *str) 2530{ 2531 get_option(&str, &slub_min_objects); 2532 2533 return 1; 2534} 2535 2536__setup("slub_min_objects=", setup_slub_min_objects); 2537 2538static int __init setup_slub_nomerge(char *str) 2539{ 2540 slub_nomerge = 1; 2541 return 1; 2542} 2543 2544__setup("slub_nomerge", setup_slub_nomerge); 2545 2546static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s, 2547 const char *name, int size, gfp_t gfp_flags) 2548{ 2549 unsigned int flags = 0; 2550 2551 if (gfp_flags & SLUB_DMA) 2552 flags = SLAB_CACHE_DMA; 2553 2554 down_write(&slub_lock); 2555 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN, 2556 flags, NULL)) 2557 goto panic; 2558 2559 list_add(&s->list, &slab_caches); 2560 up_write(&slub_lock); 2561 if (sysfs_slab_add(s)) 2562 goto panic; 2563 return s; 2564 2565panic: 2566 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 2567} 2568 2569#ifdef CONFIG_ZONE_DMA 2570static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1]; 2571 2572static void sysfs_add_func(struct work_struct *w) 2573{ 2574 struct kmem_cache *s; 2575 2576 down_write(&slub_lock); 2577 list_for_each_entry(s, &slab_caches, list) { 2578 if (s->flags & __SYSFS_ADD_DEFERRED) { 2579 s->flags &= ~__SYSFS_ADD_DEFERRED; 2580 sysfs_slab_add(s); 2581 } 2582 } 2583 up_write(&slub_lock); 2584} 2585 2586static DECLARE_WORK(sysfs_add_work, sysfs_add_func); 2587 2588static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags) 2589{ 2590 struct kmem_cache *s; 2591 char *text; 2592 size_t realsize; 2593 2594 s = kmalloc_caches_dma[index]; 2595 if (s) 2596 return s; 2597 2598 /* Dynamically create dma cache */ 2599 if (flags & __GFP_WAIT) 2600 down_write(&slub_lock); 2601 else { 2602 if (!down_write_trylock(&slub_lock)) 2603 goto out; 2604 } 2605 2606 if (kmalloc_caches_dma[index]) 2607 goto unlock_out; 2608 2609 realsize = kmalloc_caches[index].objsize; 2610 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", 2611 (unsigned int)realsize); 2612 s = kmalloc(kmem_size, flags & ~SLUB_DMA); 2613 2614 if (!s || !text || !kmem_cache_open(s, flags, text, 2615 realsize, ARCH_KMALLOC_MINALIGN, 2616 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) { 2617 kfree(s); 2618 kfree(text); 2619 goto unlock_out; 2620 } 2621 2622 list_add(&s->list, &slab_caches); 2623 kmalloc_caches_dma[index] = s; 2624 2625 schedule_work(&sysfs_add_work); 2626 2627unlock_out: 2628 up_write(&slub_lock); 2629out: 2630 return kmalloc_caches_dma[index]; 2631} 2632#endif 2633 2634/* 2635 * Conversion table for small slabs sizes / 8 to the index in the 2636 * kmalloc array. This is necessary for slabs < 192 since we have non power 2637 * of two cache sizes there. The size of larger slabs can be determined using 2638 * fls. 2639 */ 2640static s8 size_index[24] = { 2641 3, /* 8 */ 2642 4, /* 16 */ 2643 5, /* 24 */ 2644 5, /* 32 */ 2645 6, /* 40 */ 2646 6, /* 48 */ 2647 6, /* 56 */ 2648 6, /* 64 */ 2649 1, /* 72 */ 2650 1, /* 80 */ 2651 1, /* 88 */ 2652 1, /* 96 */ 2653 7, /* 104 */ 2654 7, /* 112 */ 2655 7, /* 120 */ 2656 7, /* 128 */ 2657 2, /* 136 */ 2658 2, /* 144 */ 2659 2, /* 152 */ 2660 2, /* 160 */ 2661 2, /* 168 */ 2662 2, /* 176 */ 2663 2, /* 184 */ 2664 2 /* 192 */ 2665}; 2666 2667static struct kmem_cache *get_slab(size_t size, gfp_t flags) 2668{ 2669 int index; 2670 2671 if (size <= 192) { 2672 if (!size) 2673 return ZERO_SIZE_PTR; 2674 2675 index = size_index[(size - 1) / 8]; 2676 } else 2677 index = fls(size - 1); 2678 2679#ifdef CONFIG_ZONE_DMA 2680 if (unlikely((flags & SLUB_DMA))) 2681 return dma_kmalloc_cache(index, flags); 2682 2683#endif 2684 return &kmalloc_caches[index]; 2685} 2686 2687void *__kmalloc(size_t size, gfp_t flags) 2688{ 2689 struct kmem_cache *s; 2690 2691 if (unlikely(size > PAGE_SIZE)) 2692 return kmalloc_large(size, flags); 2693 2694 s = get_slab(size, flags); 2695 2696 if (unlikely(ZERO_OR_NULL_PTR(s))) 2697 return s; 2698 2699 return slab_alloc(s, flags, -1, __builtin_return_address(0)); 2700} 2701EXPORT_SYMBOL(__kmalloc); 2702 2703static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 2704{ 2705 struct page *page = alloc_pages_node(node, flags | __GFP_COMP, 2706 get_order(size)); 2707 2708 if (page) 2709 return page_address(page); 2710 else 2711 return NULL; 2712} 2713 2714#ifdef CONFIG_NUMA 2715void *__kmalloc_node(size_t size, gfp_t flags, int node) 2716{ 2717 struct kmem_cache *s; 2718 2719 if (unlikely(size > PAGE_SIZE)) 2720 return kmalloc_large_node(size, flags, node); 2721 2722 s = get_slab(size, flags); 2723 2724 if (unlikely(ZERO_OR_NULL_PTR(s))) 2725 return s; 2726 2727 return slab_alloc(s, flags, node, __builtin_return_address(0)); 2728} 2729EXPORT_SYMBOL(__kmalloc_node); 2730#endif 2731 2732size_t ksize(const void *object) 2733{ 2734 struct page *page; 2735 struct kmem_cache *s; 2736 2737 if (unlikely(object == ZERO_SIZE_PTR)) 2738 return 0; 2739 2740 page = virt_to_head_page(object); 2741 2742 if (unlikely(!PageSlab(page))) 2743 return PAGE_SIZE << compound_order(page); 2744 2745 s = page->slab; 2746 2747#ifdef CONFIG_SLUB_DEBUG 2748 /* 2749 * Debugging requires use of the padding between object 2750 * and whatever may come after it. 2751 */ 2752 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 2753 return s->objsize; 2754 2755#endif 2756 /* 2757 * If we have the need to store the freelist pointer 2758 * back there or track user information then we can 2759 * only use the space before that information. 2760 */ 2761 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 2762 return s->inuse; 2763 /* 2764 * Else we can use all the padding etc for the allocation 2765 */ 2766 return s->size; 2767} 2768EXPORT_SYMBOL(ksize); 2769 2770void kfree(const void *x) 2771{ 2772 struct page *page; 2773 void *object = (void *)x; 2774 2775 if (unlikely(ZERO_OR_NULL_PTR(x))) 2776 return; 2777 2778 page = virt_to_head_page(x); 2779 if (unlikely(!PageSlab(page))) { 2780 put_page(page); 2781 return; 2782 } 2783 slab_free(page->slab, page, object, __builtin_return_address(0)); 2784} 2785EXPORT_SYMBOL(kfree); 2786 2787/* 2788 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 2789 * the remaining slabs by the number of items in use. The slabs with the 2790 * most items in use come first. New allocations will then fill those up 2791 * and thus they can be removed from the partial lists. 2792 * 2793 * The slabs with the least items are placed last. This results in them 2794 * being allocated from last increasing the chance that the last objects 2795 * are freed in them. 2796 */ 2797int kmem_cache_shrink(struct kmem_cache *s) 2798{ 2799 int node; 2800 int i; 2801 struct kmem_cache_node *n; 2802 struct page *page; 2803 struct page *t; 2804 int objects = oo_objects(s->max); 2805 struct list_head *slabs_by_inuse = 2806 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 2807 unsigned long flags; 2808 2809 if (!slabs_by_inuse) 2810 return -ENOMEM; 2811 2812 flush_all(s); 2813 for_each_node_state(node, N_NORMAL_MEMORY) { 2814 n = get_node(s, node); 2815 2816 if (!n->nr_partial) 2817 continue; 2818 2819 for (i = 0; i < objects; i++) 2820 INIT_LIST_HEAD(slabs_by_inuse + i); 2821 2822 spin_lock_irqsave(&n->list_lock, flags); 2823 2824 /* 2825 * Build lists indexed by the items in use in each slab. 2826 * 2827 * Note that concurrent frees may occur while we hold the 2828 * list_lock. page->inuse here is the upper limit. 2829 */ 2830 list_for_each_entry_safe(page, t, &n->partial, lru) { 2831 if (!page->inuse && slab_trylock(page)) { 2832 /* 2833 * Must hold slab lock here because slab_free 2834 * may have freed the last object and be 2835 * waiting to release the slab. 2836 */ 2837 list_del(&page->lru); 2838 n->nr_partial--; 2839 slab_unlock(page); 2840 discard_slab(s, page); 2841 } else { 2842 list_move(&page->lru, 2843 slabs_by_inuse + page->inuse); 2844 } 2845 } 2846 2847 /* 2848 * Rebuild the partial list with the slabs filled up most 2849 * first and the least used slabs at the end. 2850 */ 2851 for (i = objects - 1; i >= 0; i--) 2852 list_splice(slabs_by_inuse + i, n->partial.prev); 2853 2854 spin_unlock_irqrestore(&n->list_lock, flags); 2855 } 2856 2857 kfree(slabs_by_inuse); 2858 return 0; 2859} 2860EXPORT_SYMBOL(kmem_cache_shrink); 2861 2862#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) 2863static int slab_mem_going_offline_callback(void *arg) 2864{ 2865 struct kmem_cache *s; 2866 2867 down_read(&slub_lock); 2868 list_for_each_entry(s, &slab_caches, list) 2869 kmem_cache_shrink(s); 2870 up_read(&slub_lock); 2871 2872 return 0; 2873} 2874 2875static void slab_mem_offline_callback(void *arg) 2876{ 2877 struct kmem_cache_node *n; 2878 struct kmem_cache *s; 2879 struct memory_notify *marg = arg; 2880 int offline_node; 2881 2882 offline_node = marg->status_change_nid; 2883 2884 /* 2885 * If the node still has available memory. we need kmem_cache_node 2886 * for it yet. 2887 */ 2888 if (offline_node < 0) 2889 return; 2890 2891 down_read(&slub_lock); 2892 list_for_each_entry(s, &slab_caches, list) { 2893 n = get_node(s, offline_node); 2894 if (n) { 2895 /* 2896 * if n->nr_slabs > 0, slabs still exist on the node 2897 * that is going down. We were unable to free them, 2898 * and offline_pages() function shoudn't call this 2899 * callback. So, we must fail. 2900 */ 2901 BUG_ON(slabs_node(s, offline_node)); 2902 2903 s->node[offline_node] = NULL; 2904 kmem_cache_free(kmalloc_caches, n); 2905 } 2906 } 2907 up_read(&slub_lock); 2908} 2909 2910static int slab_mem_going_online_callback(void *arg) 2911{ 2912 struct kmem_cache_node *n; 2913 struct kmem_cache *s; 2914 struct memory_notify *marg = arg; 2915 int nid = marg->status_change_nid; 2916 int ret = 0; 2917 2918 /* 2919 * If the node's memory is already available, then kmem_cache_node is 2920 * already created. Nothing to do. 2921 */ 2922 if (nid < 0) 2923 return 0; 2924 2925 /* 2926 * We are bringing a node online. No memory is availabe yet. We must 2927 * allocate a kmem_cache_node structure in order to bring the node 2928 * online. 2929 */ 2930 down_read(&slub_lock); 2931 list_for_each_entry(s, &slab_caches, list) { 2932 /* 2933 * XXX: kmem_cache_alloc_node will fallback to other nodes 2934 * since memory is not yet available from the node that 2935 * is brought up. 2936 */ 2937 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL); 2938 if (!n) { 2939 ret = -ENOMEM; 2940 goto out; 2941 } 2942 init_kmem_cache_node(n); 2943 s->node[nid] = n; 2944 } 2945out: 2946 up_read(&slub_lock); 2947 return ret; 2948} 2949 2950static int slab_memory_callback(struct notifier_block *self, 2951 unsigned long action, void *arg) 2952{ 2953 int ret = 0; 2954 2955 switch (action) { 2956 case MEM_GOING_ONLINE: 2957 ret = slab_mem_going_online_callback(arg); 2958 break; 2959 case MEM_GOING_OFFLINE: 2960 ret = slab_mem_going_offline_callback(arg); 2961 break; 2962 case MEM_OFFLINE: 2963 case MEM_CANCEL_ONLINE: 2964 slab_mem_offline_callback(arg); 2965 break; 2966 case MEM_ONLINE: 2967 case MEM_CANCEL_OFFLINE: 2968 break; 2969 } 2970 2971 ret = notifier_from_errno(ret); 2972 return ret; 2973} 2974 2975#endif /* CONFIG_MEMORY_HOTPLUG */ 2976 2977/******************************************************************** 2978 * Basic setup of slabs 2979 *******************************************************************/ 2980 2981void __init kmem_cache_init(void) 2982{ 2983 int i; 2984 int caches = 0; 2985 2986 init_alloc_cpu(); 2987 2988#ifdef CONFIG_NUMA 2989 /* 2990 * Must first have the slab cache available for the allocations of the 2991 * struct kmem_cache_node's. There is special bootstrap code in 2992 * kmem_cache_open for slab_state == DOWN. 2993 */ 2994 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node", 2995 sizeof(struct kmem_cache_node), GFP_KERNEL); 2996 kmalloc_caches[0].refcount = -1; 2997 caches++; 2998 2999 hotplug_memory_notifier(slab_memory_callback, 1); 3000#endif 3001 3002 /* Able to allocate the per node structures */ 3003 slab_state = PARTIAL; 3004 3005 /* Caches that are not of the two-to-the-power-of size */ 3006 if (KMALLOC_MIN_SIZE <= 64) { 3007 create_kmalloc_cache(&kmalloc_caches[1], 3008 "kmalloc-96", 96, GFP_KERNEL); 3009 caches++; 3010 } 3011 if (KMALLOC_MIN_SIZE <= 128) { 3012 create_kmalloc_cache(&kmalloc_caches[2], 3013 "kmalloc-192", 192, GFP_KERNEL); 3014 caches++; 3015 } 3016 3017 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) { 3018 create_kmalloc_cache(&kmalloc_caches[i], 3019 "kmalloc", 1 << i, GFP_KERNEL); 3020 caches++; 3021 } 3022 3023 3024 /* 3025 * Patch up the size_index table if we have strange large alignment 3026 * requirements for the kmalloc array. This is only the case for 3027 * MIPS it seems. The standard arches will not generate any code here. 3028 * 3029 * Largest permitted alignment is 256 bytes due to the way we 3030 * handle the index determination for the smaller caches. 3031 * 3032 * Make sure that nothing crazy happens if someone starts tinkering 3033 * around with ARCH_KMALLOC_MINALIGN 3034 */ 3035 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3036 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3037 3038 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) 3039 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW; 3040 3041 slab_state = UP; 3042 3043 /* Provide the correct kmalloc names now that the caches are up */ 3044 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) 3045 kmalloc_caches[i]. name = 3046 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i); 3047 3048#ifdef CONFIG_SMP 3049 register_cpu_notifier(&slab_notifier); 3050 kmem_size = offsetof(struct kmem_cache, cpu_slab) + 3051 nr_cpu_ids * sizeof(struct kmem_cache_cpu *); 3052#else 3053 kmem_size = sizeof(struct kmem_cache); 3054#endif 3055 3056 printk(KERN_INFO 3057 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3058 " CPUs=%d, Nodes=%d\n", 3059 caches, cache_line_size(), 3060 slub_min_order, slub_max_order, slub_min_objects, 3061 nr_cpu_ids, nr_node_ids); 3062} 3063 3064/* 3065 * Find a mergeable slab cache 3066 */ 3067static int slab_unmergeable(struct kmem_cache *s) 3068{ 3069 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3070 return 1; 3071 3072 if (s->ctor) 3073 return 1; 3074 3075 /* 3076 * We may have set a slab to be unmergeable during bootstrap. 3077 */ 3078 if (s->refcount < 0) 3079 return 1; 3080 3081 return 0; 3082} 3083 3084static struct kmem_cache *find_mergeable(size_t size, 3085 size_t align, unsigned long flags, const char *name, 3086 void (*ctor)(struct kmem_cache *, void *)) 3087{ 3088 struct kmem_cache *s; 3089 3090 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3091 return NULL; 3092 3093 if (ctor) 3094 return NULL; 3095 3096 size = ALIGN(size, sizeof(void *)); 3097 align = calculate_alignment(flags, align, size); 3098 size = ALIGN(size, align); 3099 flags = kmem_cache_flags(size, flags, name, NULL); 3100 3101 list_for_each_entry(s, &slab_caches, list) { 3102 if (slab_unmergeable(s)) 3103 continue; 3104 3105 if (size > s->size) 3106 continue; 3107 3108 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3109 continue; 3110 /* 3111 * Check if alignment is compatible. 3112 * Courtesy of Adrian Drzewiecki 3113 */ 3114 if ((s->size & ~(align - 1)) != s->size) 3115 continue; 3116 3117 if (s->size - size >= sizeof(void *)) 3118 continue; 3119 3120 return s; 3121 } 3122 return NULL; 3123} 3124 3125struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3126 size_t align, unsigned long flags, 3127 void (*ctor)(struct kmem_cache *, void *)) 3128{ 3129 struct kmem_cache *s; 3130 3131 down_write(&slub_lock); 3132 s = find_mergeable(size, align, flags, name, ctor); 3133 if (s) { 3134 int cpu; 3135 3136 s->refcount++; 3137 /* 3138 * Adjust the object sizes so that we clear 3139 * the complete object on kzalloc. 3140 */ 3141 s->objsize = max(s->objsize, (int)size); 3142 3143 /* 3144 * And then we need to update the object size in the 3145 * per cpu structures 3146 */ 3147 for_each_online_cpu(cpu) 3148 get_cpu_slab(s, cpu)->objsize = s->objsize; 3149 3150 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3151 up_write(&slub_lock); 3152 3153 if (sysfs_slab_alias(s, name)) 3154 goto err; 3155 return s; 3156 } 3157 3158 s = kmalloc(kmem_size, GFP_KERNEL); 3159 if (s) { 3160 if (kmem_cache_open(s, GFP_KERNEL, name, 3161 size, align, flags, ctor)) { 3162 list_add(&s->list, &slab_caches); 3163 up_write(&slub_lock); 3164 if (sysfs_slab_add(s)) 3165 goto err; 3166 return s; 3167 } 3168 kfree(s); 3169 } 3170 up_write(&slub_lock); 3171 3172err: 3173 if (flags & SLAB_PANIC) 3174 panic("Cannot create slabcache %s\n", name); 3175 else 3176 s = NULL; 3177 return s; 3178} 3179EXPORT_SYMBOL(kmem_cache_create); 3180 3181#ifdef CONFIG_SMP 3182/* 3183 * Use the cpu notifier to insure that the cpu slabs are flushed when 3184 * necessary. 3185 */ 3186static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3187 unsigned long action, void *hcpu) 3188{ 3189 long cpu = (long)hcpu; 3190 struct kmem_cache *s; 3191 unsigned long flags; 3192 3193 switch (action) { 3194 case CPU_UP_PREPARE: 3195 case CPU_UP_PREPARE_FROZEN: 3196 init_alloc_cpu_cpu(cpu); 3197 down_read(&slub_lock); 3198 list_for_each_entry(s, &slab_caches, list) 3199 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu, 3200 GFP_KERNEL); 3201 up_read(&slub_lock); 3202 break; 3203 3204 case CPU_UP_CANCELED: 3205 case CPU_UP_CANCELED_FROZEN: 3206 case CPU_DEAD: 3207 case CPU_DEAD_FROZEN: 3208 down_read(&slub_lock); 3209 list_for_each_entry(s, &slab_caches, list) { 3210 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3211 3212 local_irq_save(flags); 3213 __flush_cpu_slab(s, cpu); 3214 local_irq_restore(flags); 3215 free_kmem_cache_cpu(c, cpu); 3216 s->cpu_slab[cpu] = NULL; 3217 } 3218 up_read(&slub_lock); 3219 break; 3220 default: 3221 break; 3222 } 3223 return NOTIFY_OK; 3224} 3225 3226static struct notifier_block __cpuinitdata slab_notifier = { 3227 .notifier_call = slab_cpuup_callback 3228}; 3229 3230#endif 3231 3232void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller) 3233{ 3234 struct kmem_cache *s; 3235 3236 if (unlikely(size > PAGE_SIZE)) 3237 return kmalloc_large(size, gfpflags); 3238 3239 s = get_slab(size, gfpflags); 3240 3241 if (unlikely(ZERO_OR_NULL_PTR(s))) 3242 return s; 3243 3244 return slab_alloc(s, gfpflags, -1, caller); 3245} 3246 3247void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3248 int node, void *caller) 3249{ 3250 struct kmem_cache *s; 3251 3252 if (unlikely(size > PAGE_SIZE)) 3253 return kmalloc_large_node(size, gfpflags, node); 3254 3255 s = get_slab(size, gfpflags); 3256 3257 if (unlikely(ZERO_OR_NULL_PTR(s))) 3258 return s; 3259 3260 return slab_alloc(s, gfpflags, node, caller); 3261} 3262 3263#if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO) 3264static unsigned long count_partial(struct kmem_cache_node *n, 3265 int (*get_count)(struct page *)) 3266{ 3267 unsigned long flags; 3268 unsigned long x = 0; 3269 struct page *page; 3270 3271 spin_lock_irqsave(&n->list_lock, flags); 3272 list_for_each_entry(page, &n->partial, lru) 3273 x += get_count(page); 3274 spin_unlock_irqrestore(&n->list_lock, flags); 3275 return x; 3276} 3277 3278static int count_inuse(struct page *page) 3279{ 3280 return page->inuse; 3281} 3282 3283static int count_total(struct page *page) 3284{ 3285 return page->objects; 3286} 3287 3288static int count_free(struct page *page) 3289{ 3290 return page->objects - page->inuse; 3291} 3292#endif 3293 3294#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG) 3295static int validate_slab(struct kmem_cache *s, struct page *page, 3296 unsigned long *map) 3297{ 3298 void *p; 3299 void *addr = page_address(page); 3300 3301 if (!check_slab(s, page) || 3302 !on_freelist(s, page, NULL)) 3303 return 0; 3304 3305 /* Now we know that a valid freelist exists */ 3306 bitmap_zero(map, page->objects); 3307 3308 for_each_free_object(p, s, page->freelist) { 3309 set_bit(slab_index(p, s, addr), map); 3310 if (!check_object(s, page, p, 0)) 3311 return 0; 3312 } 3313 3314 for_each_object(p, s, addr, page->objects) 3315 if (!test_bit(slab_index(p, s, addr), map)) 3316 if (!check_object(s, page, p, 1)) 3317 return 0; 3318 return 1; 3319} 3320 3321static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3322 unsigned long *map) 3323{ 3324 if (slab_trylock(page)) { 3325 validate_slab(s, page, map); 3326 slab_unlock(page); 3327 } else 3328 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n", 3329 s->name, page); 3330 3331 if (s->flags & DEBUG_DEFAULT_FLAGS) { 3332 if (!SlabDebug(page)) 3333 printk(KERN_ERR "SLUB %s: SlabDebug not set " 3334 "on slab 0x%p\n", s->name, page); 3335 } else { 3336 if (SlabDebug(page)) 3337 printk(KERN_ERR "SLUB %s: SlabDebug set on " 3338 "slab 0x%p\n", s->name, page); 3339 } 3340} 3341 3342static int validate_slab_node(struct kmem_cache *s, 3343 struct kmem_cache_node *n, unsigned long *map) 3344{ 3345 unsigned long count = 0; 3346 struct page *page; 3347 unsigned long flags; 3348 3349 spin_lock_irqsave(&n->list_lock, flags); 3350 3351 list_for_each_entry(page, &n->partial, lru) { 3352 validate_slab_slab(s, page, map); 3353 count++; 3354 } 3355 if (count != n->nr_partial) 3356 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3357 "counter=%ld\n", s->name, count, n->nr_partial); 3358 3359 if (!(s->flags & SLAB_STORE_USER)) 3360 goto out; 3361 3362 list_for_each_entry(page, &n->full, lru) { 3363 validate_slab_slab(s, page, map); 3364 count++; 3365 } 3366 if (count != atomic_long_read(&n->nr_slabs)) 3367 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3368 "counter=%ld\n", s->name, count, 3369 atomic_long_read(&n->nr_slabs)); 3370 3371out: 3372 spin_unlock_irqrestore(&n->list_lock, flags); 3373 return count; 3374} 3375 3376static long validate_slab_cache(struct kmem_cache *s) 3377{ 3378 int node; 3379 unsigned long count = 0; 3380 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3381 sizeof(unsigned long), GFP_KERNEL); 3382 3383 if (!map) 3384 return -ENOMEM; 3385 3386 flush_all(s); 3387 for_each_node_state(node, N_NORMAL_MEMORY) { 3388 struct kmem_cache_node *n = get_node(s, node); 3389 3390 count += validate_slab_node(s, n, map); 3391 } 3392 kfree(map); 3393 return count; 3394} 3395 3396#ifdef SLUB_RESILIENCY_TEST 3397static void resiliency_test(void) 3398{ 3399 u8 *p; 3400 3401 printk(KERN_ERR "SLUB resiliency testing\n"); 3402 printk(KERN_ERR "-----------------------\n"); 3403 printk(KERN_ERR "A. Corruption after allocation\n"); 3404 3405 p = kzalloc(16, GFP_KERNEL); 3406 p[16] = 0x12; 3407 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 3408 " 0x12->0x%p\n\n", p + 16); 3409 3410 validate_slab_cache(kmalloc_caches + 4); 3411 3412 /* Hmmm... The next two are dangerous */ 3413 p = kzalloc(32, GFP_KERNEL); 3414 p[32 + sizeof(void *)] = 0x34; 3415 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 3416 " 0x34 -> -0x%p\n", p); 3417 printk(KERN_ERR 3418 "If allocated object is overwritten then not detectable\n\n"); 3419 3420 validate_slab_cache(kmalloc_caches + 5); 3421 p = kzalloc(64, GFP_KERNEL); 3422 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 3423 *p = 0x56; 3424 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 3425 p); 3426 printk(KERN_ERR 3427 "If allocated object is overwritten then not detectable\n\n"); 3428 validate_slab_cache(kmalloc_caches + 6); 3429 3430 printk(KERN_ERR "\nB. Corruption after free\n"); 3431 p = kzalloc(128, GFP_KERNEL); 3432 kfree(p); 3433 *p = 0x78; 3434 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 3435 validate_slab_cache(kmalloc_caches + 7); 3436 3437 p = kzalloc(256, GFP_KERNEL); 3438 kfree(p); 3439 p[50] = 0x9a; 3440 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 3441 p); 3442 validate_slab_cache(kmalloc_caches + 8); 3443 3444 p = kzalloc(512, GFP_KERNEL); 3445 kfree(p); 3446 p[512] = 0xab; 3447 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 3448 validate_slab_cache(kmalloc_caches + 9); 3449} 3450#else 3451static void resiliency_test(void) {}; 3452#endif 3453 3454/* 3455 * Generate lists of code addresses where slabcache objects are allocated 3456 * and freed. 3457 */ 3458 3459struct location { 3460 unsigned long count; 3461 void *addr; 3462 long long sum_time; 3463 long min_time; 3464 long max_time; 3465 long min_pid; 3466 long max_pid; 3467 cpumask_t cpus; 3468 nodemask_t nodes; 3469}; 3470 3471struct loc_track { 3472 unsigned long max; 3473 unsigned long count; 3474 struct location *loc; 3475}; 3476 3477static void free_loc_track(struct loc_track *t) 3478{ 3479 if (t->max) 3480 free_pages((unsigned long)t->loc, 3481 get_order(sizeof(struct location) * t->max)); 3482} 3483 3484static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 3485{ 3486 struct location *l; 3487 int order; 3488 3489 order = get_order(sizeof(struct location) * max); 3490 3491 l = (void *)__get_free_pages(flags, order); 3492 if (!l) 3493 return 0; 3494 3495 if (t->count) { 3496 memcpy(l, t->loc, sizeof(struct location) * t->count); 3497 free_loc_track(t); 3498 } 3499 t->max = max; 3500 t->loc = l; 3501 return 1; 3502} 3503 3504static int add_location(struct loc_track *t, struct kmem_cache *s, 3505 const struct track *track) 3506{ 3507 long start, end, pos; 3508 struct location *l; 3509 void *caddr; 3510 unsigned long age = jiffies - track->when; 3511 3512 start = -1; 3513 end = t->count; 3514 3515 for ( ; ; ) { 3516 pos = start + (end - start + 1) / 2; 3517 3518 /* 3519 * There is nothing at "end". If we end up there 3520 * we need to add something to before end. 3521 */ 3522 if (pos == end) 3523 break; 3524 3525 caddr = t->loc[pos].addr; 3526 if (track->addr == caddr) { 3527 3528 l = &t->loc[pos]; 3529 l->count++; 3530 if (track->when) { 3531 l->sum_time += age; 3532 if (age < l->min_time) 3533 l->min_time = age; 3534 if (age > l->max_time) 3535 l->max_time = age; 3536 3537 if (track->pid < l->min_pid) 3538 l->min_pid = track->pid; 3539 if (track->pid > l->max_pid) 3540 l->max_pid = track->pid; 3541 3542 cpu_set(track->cpu, l->cpus); 3543 } 3544 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3545 return 1; 3546 } 3547 3548 if (track->addr < caddr) 3549 end = pos; 3550 else 3551 start = pos; 3552 } 3553 3554 /* 3555 * Not found. Insert new tracking element. 3556 */ 3557 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 3558 return 0; 3559 3560 l = t->loc + pos; 3561 if (pos < t->count) 3562 memmove(l + 1, l, 3563 (t->count - pos) * sizeof(struct location)); 3564 t->count++; 3565 l->count = 1; 3566 l->addr = track->addr; 3567 l->sum_time = age; 3568 l->min_time = age; 3569 l->max_time = age; 3570 l->min_pid = track->pid; 3571 l->max_pid = track->pid; 3572 cpus_clear(l->cpus); 3573 cpu_set(track->cpu, l->cpus); 3574 nodes_clear(l->nodes); 3575 node_set(page_to_nid(virt_to_page(track)), l->nodes); 3576 return 1; 3577} 3578 3579static void process_slab(struct loc_track *t, struct kmem_cache *s, 3580 struct page *page, enum track_item alloc) 3581{ 3582 void *addr = page_address(page); 3583 DECLARE_BITMAP(map, page->objects); 3584 void *p; 3585 3586 bitmap_zero(map, page->objects); 3587 for_each_free_object(p, s, page->freelist) 3588 set_bit(slab_index(p, s, addr), map); 3589 3590 for_each_object(p, s, addr, page->objects) 3591 if (!test_bit(slab_index(p, s, addr), map)) 3592 add_location(t, s, get_track(s, p, alloc)); 3593} 3594 3595static int list_locations(struct kmem_cache *s, char *buf, 3596 enum track_item alloc) 3597{ 3598 int len = 0; 3599 unsigned long i; 3600 struct loc_track t = { 0, 0, NULL }; 3601 int node; 3602 3603 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 3604 GFP_TEMPORARY)) 3605 return sprintf(buf, "Out of memory\n"); 3606 3607 /* Push back cpu slabs */ 3608 flush_all(s); 3609 3610 for_each_node_state(node, N_NORMAL_MEMORY) { 3611 struct kmem_cache_node *n = get_node(s, node); 3612 unsigned long flags; 3613 struct page *page; 3614 3615 if (!atomic_long_read(&n->nr_slabs)) 3616 continue; 3617 3618 spin_lock_irqsave(&n->list_lock, flags); 3619 list_for_each_entry(page, &n->partial, lru) 3620 process_slab(&t, s, page, alloc); 3621 list_for_each_entry(page, &n->full, lru) 3622 process_slab(&t, s, page, alloc); 3623 spin_unlock_irqrestore(&n->list_lock, flags); 3624 } 3625 3626 for (i = 0; i < t.count; i++) { 3627 struct location *l = &t.loc[i]; 3628 3629 if (len > PAGE_SIZE - 100) 3630 break; 3631 len += sprintf(buf + len, "%7ld ", l->count); 3632 3633 if (l->addr) 3634 len += sprint_symbol(buf + len, (unsigned long)l->addr); 3635 else 3636 len += sprintf(buf + len, "<not-available>"); 3637 3638 if (l->sum_time != l->min_time) { 3639 unsigned long remainder; 3640 3641 len += sprintf(buf + len, " age=%ld/%ld/%ld", 3642 l->min_time, 3643 div_long_long_rem(l->sum_time, l->count, &remainder), 3644 l->max_time); 3645 } else 3646 len += sprintf(buf + len, " age=%ld", 3647 l->min_time); 3648 3649 if (l->min_pid != l->max_pid) 3650 len += sprintf(buf + len, " pid=%ld-%ld", 3651 l->min_pid, l->max_pid); 3652 else 3653 len += sprintf(buf + len, " pid=%ld", 3654 l->min_pid); 3655 3656 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) && 3657 len < PAGE_SIZE - 60) { 3658 len += sprintf(buf + len, " cpus="); 3659 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3660 l->cpus); 3661 } 3662 3663 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) && 3664 len < PAGE_SIZE - 60) { 3665 len += sprintf(buf + len, " nodes="); 3666 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 3667 l->nodes); 3668 } 3669 3670 len += sprintf(buf + len, "\n"); 3671 } 3672 3673 free_loc_track(&t); 3674 if (!t.count) 3675 len += sprintf(buf, "No data\n"); 3676 return len; 3677} 3678 3679enum slab_stat_type { 3680 SL_ALL, /* All slabs */ 3681 SL_PARTIAL, /* Only partially allocated slabs */ 3682 SL_CPU, /* Only slabs used for cpu caches */ 3683 SL_OBJECTS, /* Determine allocated objects not slabs */ 3684 SL_TOTAL /* Determine object capacity not slabs */ 3685}; 3686 3687#define SO_ALL (1 << SL_ALL) 3688#define SO_PARTIAL (1 << SL_PARTIAL) 3689#define SO_CPU (1 << SL_CPU) 3690#define SO_OBJECTS (1 << SL_OBJECTS) 3691#define SO_TOTAL (1 << SL_TOTAL) 3692 3693static ssize_t show_slab_objects(struct kmem_cache *s, 3694 char *buf, unsigned long flags) 3695{ 3696 unsigned long total = 0; 3697 int node; 3698 int x; 3699 unsigned long *nodes; 3700 unsigned long *per_cpu; 3701 3702 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 3703 if (!nodes) 3704 return -ENOMEM; 3705 per_cpu = nodes + nr_node_ids; 3706 3707 if (flags & SO_CPU) { 3708 int cpu; 3709 3710 for_each_possible_cpu(cpu) { 3711 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3712 3713 if (!c || c->node < 0) 3714 continue; 3715 3716 if (c->page) { 3717 if (flags & SO_TOTAL) 3718 x = c->page->objects; 3719 else if (flags & SO_OBJECTS) 3720 x = c->page->inuse; 3721 else 3722 x = 1; 3723 3724 total += x; 3725 nodes[c->node] += x; 3726 } 3727 per_cpu[c->node]++; 3728 } 3729 } 3730 3731 if (flags & SO_ALL) { 3732 for_each_node_state(node, N_NORMAL_MEMORY) { 3733 struct kmem_cache_node *n = get_node(s, node); 3734 3735 if (flags & SO_TOTAL) 3736 x = atomic_long_read(&n->total_objects); 3737 else if (flags & SO_OBJECTS) 3738 x = atomic_long_read(&n->total_objects) - 3739 count_partial(n, count_free); 3740 3741 else 3742 x = atomic_long_read(&n->nr_slabs); 3743 total += x; 3744 nodes[node] += x; 3745 } 3746 3747 } else if (flags & SO_PARTIAL) { 3748 for_each_node_state(node, N_NORMAL_MEMORY) { 3749 struct kmem_cache_node *n = get_node(s, node); 3750 3751 if (flags & SO_TOTAL) 3752 x = count_partial(n, count_total); 3753 else if (flags & SO_OBJECTS) 3754 x = count_partial(n, count_inuse); 3755 else 3756 x = n->nr_partial; 3757 total += x; 3758 nodes[node] += x; 3759 } 3760 } 3761 x = sprintf(buf, "%lu", total); 3762#ifdef CONFIG_NUMA 3763 for_each_node_state(node, N_NORMAL_MEMORY) 3764 if (nodes[node]) 3765 x += sprintf(buf + x, " N%d=%lu", 3766 node, nodes[node]); 3767#endif 3768 kfree(nodes); 3769 return x + sprintf(buf + x, "\n"); 3770} 3771 3772static int any_slab_objects(struct kmem_cache *s) 3773{ 3774 int node; 3775 int cpu; 3776 3777 for_each_possible_cpu(cpu) { 3778 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu); 3779 3780 if (c && c->page) 3781 return 1; 3782 } 3783 3784 for_each_online_node(node) { 3785 struct kmem_cache_node *n = get_node(s, node); 3786 3787 if (!n) 3788 continue; 3789 3790 if (n->nr_partial || atomic_long_read(&n->nr_slabs)) 3791 return 1; 3792 } 3793 return 0; 3794} 3795 3796#define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 3797#define to_slab(n) container_of(n, struct kmem_cache, kobj); 3798 3799struct slab_attribute { 3800 struct attribute attr; 3801 ssize_t (*show)(struct kmem_cache *s, char *buf); 3802 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 3803}; 3804 3805#define SLAB_ATTR_RO(_name) \ 3806 static struct slab_attribute _name##_attr = __ATTR_RO(_name) 3807 3808#define SLAB_ATTR(_name) \ 3809 static struct slab_attribute _name##_attr = \ 3810 __ATTR(_name, 0644, _name##_show, _name##_store) 3811 3812static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 3813{ 3814 return sprintf(buf, "%d\n", s->size); 3815} 3816SLAB_ATTR_RO(slab_size); 3817 3818static ssize_t align_show(struct kmem_cache *s, char *buf) 3819{ 3820 return sprintf(buf, "%d\n", s->align); 3821} 3822SLAB_ATTR_RO(align); 3823 3824static ssize_t object_size_show(struct kmem_cache *s, char *buf) 3825{ 3826 return sprintf(buf, "%d\n", s->objsize); 3827} 3828SLAB_ATTR_RO(object_size); 3829 3830static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 3831{ 3832 return sprintf(buf, "%d\n", oo_objects(s->oo)); 3833} 3834SLAB_ATTR_RO(objs_per_slab); 3835 3836static ssize_t order_show(struct kmem_cache *s, char *buf) 3837{ 3838 return sprintf(buf, "%d\n", oo_order(s->oo)); 3839} 3840SLAB_ATTR_RO(order); 3841 3842static ssize_t ctor_show(struct kmem_cache *s, char *buf) 3843{ 3844 if (s->ctor) { 3845 int n = sprint_symbol(buf, (unsigned long)s->ctor); 3846 3847 return n + sprintf(buf + n, "\n"); 3848 } 3849 return 0; 3850} 3851SLAB_ATTR_RO(ctor); 3852 3853static ssize_t aliases_show(struct kmem_cache *s, char *buf) 3854{ 3855 return sprintf(buf, "%d\n", s->refcount - 1); 3856} 3857SLAB_ATTR_RO(aliases); 3858 3859static ssize_t slabs_show(struct kmem_cache *s, char *buf) 3860{ 3861 return show_slab_objects(s, buf, SO_ALL); 3862} 3863SLAB_ATTR_RO(slabs); 3864 3865static ssize_t partial_show(struct kmem_cache *s, char *buf) 3866{ 3867 return show_slab_objects(s, buf, SO_PARTIAL); 3868} 3869SLAB_ATTR_RO(partial); 3870 3871static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 3872{ 3873 return show_slab_objects(s, buf, SO_CPU); 3874} 3875SLAB_ATTR_RO(cpu_slabs); 3876 3877static ssize_t objects_show(struct kmem_cache *s, char *buf) 3878{ 3879 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 3880} 3881SLAB_ATTR_RO(objects); 3882 3883static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 3884{ 3885 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 3886} 3887SLAB_ATTR_RO(objects_partial); 3888 3889static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 3890{ 3891 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 3892} 3893SLAB_ATTR_RO(total_objects); 3894 3895static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 3896{ 3897 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 3898} 3899 3900static ssize_t sanity_checks_store(struct kmem_cache *s, 3901 const char *buf, size_t length) 3902{ 3903 s->flags &= ~SLAB_DEBUG_FREE; 3904 if (buf[0] == '1') 3905 s->flags |= SLAB_DEBUG_FREE; 3906 return length; 3907} 3908SLAB_ATTR(sanity_checks); 3909 3910static ssize_t trace_show(struct kmem_cache *s, char *buf) 3911{ 3912 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 3913} 3914 3915static ssize_t trace_store(struct kmem_cache *s, const char *buf, 3916 size_t length) 3917{ 3918 s->flags &= ~SLAB_TRACE; 3919 if (buf[0] == '1') 3920 s->flags |= SLAB_TRACE; 3921 return length; 3922} 3923SLAB_ATTR(trace); 3924 3925static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 3926{ 3927 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 3928} 3929 3930static ssize_t reclaim_account_store(struct kmem_cache *s, 3931 const char *buf, size_t length) 3932{ 3933 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 3934 if (buf[0] == '1') 3935 s->flags |= SLAB_RECLAIM_ACCOUNT; 3936 return length; 3937} 3938SLAB_ATTR(reclaim_account); 3939 3940static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 3941{ 3942 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 3943} 3944SLAB_ATTR_RO(hwcache_align); 3945 3946#ifdef CONFIG_ZONE_DMA 3947static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 3948{ 3949 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 3950} 3951SLAB_ATTR_RO(cache_dma); 3952#endif 3953 3954static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 3955{ 3956 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 3957} 3958SLAB_ATTR_RO(destroy_by_rcu); 3959 3960static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 3961{ 3962 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 3963} 3964 3965static ssize_t red_zone_store(struct kmem_cache *s, 3966 const char *buf, size_t length) 3967{ 3968 if (any_slab_objects(s)) 3969 return -EBUSY; 3970 3971 s->flags &= ~SLAB_RED_ZONE; 3972 if (buf[0] == '1') 3973 s->flags |= SLAB_RED_ZONE; 3974 calculate_sizes(s); 3975 return length; 3976} 3977SLAB_ATTR(red_zone); 3978 3979static ssize_t poison_show(struct kmem_cache *s, char *buf) 3980{ 3981 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 3982} 3983 3984static ssize_t poison_store(struct kmem_cache *s, 3985 const char *buf, size_t length) 3986{ 3987 if (any_slab_objects(s)) 3988 return -EBUSY; 3989 3990 s->flags &= ~SLAB_POISON; 3991 if (buf[0] == '1') 3992 s->flags |= SLAB_POISON; 3993 calculate_sizes(s); 3994 return length; 3995} 3996SLAB_ATTR(poison); 3997 3998static ssize_t store_user_show(struct kmem_cache *s, char *buf) 3999{ 4000 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4001} 4002 4003static ssize_t store_user_store(struct kmem_cache *s, 4004 const char *buf, size_t length) 4005{ 4006 if (any_slab_objects(s)) 4007 return -EBUSY; 4008 4009 s->flags &= ~SLAB_STORE_USER; 4010 if (buf[0] == '1') 4011 s->flags |= SLAB_STORE_USER; 4012 calculate_sizes(s); 4013 return length; 4014} 4015SLAB_ATTR(store_user); 4016 4017static ssize_t validate_show(struct kmem_cache *s, char *buf) 4018{ 4019 return 0; 4020} 4021 4022static ssize_t validate_store(struct kmem_cache *s, 4023 const char *buf, size_t length) 4024{ 4025 int ret = -EINVAL; 4026 4027 if (buf[0] == '1') { 4028 ret = validate_slab_cache(s); 4029 if (ret >= 0) 4030 ret = length; 4031 } 4032 return ret; 4033} 4034SLAB_ATTR(validate); 4035 4036static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4037{ 4038 return 0; 4039} 4040 4041static ssize_t shrink_store(struct kmem_cache *s, 4042 const char *buf, size_t length) 4043{ 4044 if (buf[0] == '1') { 4045 int rc = kmem_cache_shrink(s); 4046 4047 if (rc) 4048 return rc; 4049 } else 4050 return -EINVAL; 4051 return length; 4052} 4053SLAB_ATTR(shrink); 4054 4055static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4056{ 4057 if (!(s->flags & SLAB_STORE_USER)) 4058 return -ENOSYS; 4059 return list_locations(s, buf, TRACK_ALLOC); 4060} 4061SLAB_ATTR_RO(alloc_calls); 4062 4063static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4064{ 4065 if (!(s->flags & SLAB_STORE_USER)) 4066 return -ENOSYS; 4067 return list_locations(s, buf, TRACK_FREE); 4068} 4069SLAB_ATTR_RO(free_calls); 4070 4071#ifdef CONFIG_NUMA 4072static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4073{ 4074 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4075} 4076 4077static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4078 const char *buf, size_t length) 4079{ 4080 int n = simple_strtoul(buf, NULL, 10); 4081 4082 if (n < 100) 4083 s->remote_node_defrag_ratio = n * 10; 4084 return length; 4085} 4086SLAB_ATTR(remote_node_defrag_ratio); 4087#endif 4088 4089#ifdef CONFIG_SLUB_STATS 4090static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4091{ 4092 unsigned long sum = 0; 4093 int cpu; 4094 int len; 4095 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4096 4097 if (!data) 4098 return -ENOMEM; 4099 4100 for_each_online_cpu(cpu) { 4101 unsigned x = get_cpu_slab(s, cpu)->stat[si]; 4102 4103 data[cpu] = x; 4104 sum += x; 4105 } 4106 4107 len = sprintf(buf, "%lu", sum); 4108 4109#ifdef CONFIG_SMP 4110 for_each_online_cpu(cpu) { 4111 if (data[cpu] && len < PAGE_SIZE - 20) 4112 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4113 } 4114#endif 4115 kfree(data); 4116 return len + sprintf(buf + len, "\n"); 4117} 4118 4119#define STAT_ATTR(si, text) \ 4120static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4121{ \ 4122 return show_stat(s, buf, si); \ 4123} \ 4124SLAB_ATTR_RO(text); \ 4125 4126STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4127STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4128STAT_ATTR(FREE_FASTPATH, free_fastpath); 4129STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4130STAT_ATTR(FREE_FROZEN, free_frozen); 4131STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4132STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4133STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4134STAT_ATTR(ALLOC_SLAB, alloc_slab); 4135STAT_ATTR(ALLOC_REFILL, alloc_refill); 4136STAT_ATTR(FREE_SLAB, free_slab); 4137STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4138STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4139STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4140STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4141STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4142STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4143STAT_ATTR(ORDER_FALLBACK, order_fallback); 4144#endif 4145 4146static struct attribute *slab_attrs[] = { 4147 &slab_size_attr.attr, 4148 &object_size_attr.attr, 4149 &objs_per_slab_attr.attr, 4150 &order_attr.attr, 4151 &objects_attr.attr, 4152 &objects_partial_attr.attr, 4153 &total_objects_attr.attr, 4154 &slabs_attr.attr, 4155 &partial_attr.attr, 4156 &cpu_slabs_attr.attr, 4157 &ctor_attr.attr, 4158 &aliases_attr.attr, 4159 &align_attr.attr, 4160 &sanity_checks_attr.attr, 4161 &trace_attr.attr, 4162 &hwcache_align_attr.attr, 4163 &reclaim_account_attr.attr, 4164 &destroy_by_rcu_attr.attr, 4165 &red_zone_attr.attr, 4166 &poison_attr.attr, 4167 &store_user_attr.attr, 4168 &validate_attr.attr, 4169 &shrink_attr.attr, 4170 &alloc_calls_attr.attr, 4171 &free_calls_attr.attr, 4172#ifdef CONFIG_ZONE_DMA 4173 &cache_dma_attr.attr, 4174#endif 4175#ifdef CONFIG_NUMA 4176 &remote_node_defrag_ratio_attr.attr, 4177#endif 4178#ifdef CONFIG_SLUB_STATS 4179 &alloc_fastpath_attr.attr, 4180 &alloc_slowpath_attr.attr, 4181 &free_fastpath_attr.attr, 4182 &free_slowpath_attr.attr, 4183 &free_frozen_attr.attr, 4184 &free_add_partial_attr.attr, 4185 &free_remove_partial_attr.attr, 4186 &alloc_from_partial_attr.attr, 4187 &alloc_slab_attr.attr, 4188 &alloc_refill_attr.attr, 4189 &free_slab_attr.attr, 4190 &cpuslab_flush_attr.attr, 4191 &deactivate_full_attr.attr, 4192 &deactivate_empty_attr.attr, 4193 &deactivate_to_head_attr.attr, 4194 &deactivate_to_tail_attr.attr, 4195 &deactivate_remote_frees_attr.attr, 4196 &order_fallback_attr.attr, 4197#endif 4198 NULL 4199}; 4200 4201static struct attribute_group slab_attr_group = { 4202 .attrs = slab_attrs, 4203}; 4204 4205static ssize_t slab_attr_show(struct kobject *kobj, 4206 struct attribute *attr, 4207 char *buf) 4208{ 4209 struct slab_attribute *attribute; 4210 struct kmem_cache *s; 4211 int err; 4212 4213 attribute = to_slab_attr(attr); 4214 s = to_slab(kobj); 4215 4216 if (!attribute->show) 4217 return -EIO; 4218 4219 err = attribute->show(s, buf); 4220 4221 return err; 4222} 4223 4224static ssize_t slab_attr_store(struct kobject *kobj, 4225 struct attribute *attr, 4226 const char *buf, size_t len) 4227{ 4228 struct slab_attribute *attribute; 4229 struct kmem_cache *s; 4230 int err; 4231 4232 attribute = to_slab_attr(attr); 4233 s = to_slab(kobj); 4234 4235 if (!attribute->store) 4236 return -EIO; 4237 4238 err = attribute->store(s, buf, len); 4239 4240 return err; 4241} 4242 4243static void kmem_cache_release(struct kobject *kobj) 4244{ 4245 struct kmem_cache *s = to_slab(kobj); 4246 4247 kfree(s); 4248} 4249 4250static struct sysfs_ops slab_sysfs_ops = { 4251 .show = slab_attr_show, 4252 .store = slab_attr_store, 4253}; 4254 4255static struct kobj_type slab_ktype = { 4256 .sysfs_ops = &slab_sysfs_ops, 4257 .release = kmem_cache_release 4258}; 4259 4260static int uevent_filter(struct kset *kset, struct kobject *kobj) 4261{ 4262 struct kobj_type *ktype = get_ktype(kobj); 4263 4264 if (ktype == &slab_ktype) 4265 return 1; 4266 return 0; 4267} 4268 4269static struct kset_uevent_ops slab_uevent_ops = { 4270 .filter = uevent_filter, 4271}; 4272 4273static struct kset *slab_kset; 4274 4275#define ID_STR_LENGTH 64 4276 4277/* Create a unique string id for a slab cache: 4278 * 4279 * Format :[flags-]size 4280 */ 4281static char *create_unique_id(struct kmem_cache *s) 4282{ 4283 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 4284 char *p = name; 4285 4286 BUG_ON(!name); 4287 4288 *p++ = ':'; 4289 /* 4290 * First flags affecting slabcache operations. We will only 4291 * get here for aliasable slabs so we do not need to support 4292 * too many flags. The flags here must cover all flags that 4293 * are matched during merging to guarantee that the id is 4294 * unique. 4295 */ 4296 if (s->flags & SLAB_CACHE_DMA) 4297 *p++ = 'd'; 4298 if (s->flags & SLAB_RECLAIM_ACCOUNT) 4299 *p++ = 'a'; 4300 if (s->flags & SLAB_DEBUG_FREE) 4301 *p++ = 'F'; 4302 if (p != name + 1) 4303 *p++ = '-'; 4304 p += sprintf(p, "%07d", s->size); 4305 BUG_ON(p > name + ID_STR_LENGTH - 1); 4306 return name; 4307} 4308 4309static int sysfs_slab_add(struct kmem_cache *s) 4310{ 4311 int err; 4312 const char *name; 4313 int unmergeable; 4314 4315 if (slab_state < SYSFS) 4316 /* Defer until later */ 4317 return 0; 4318 4319 unmergeable = slab_unmergeable(s); 4320 if (unmergeable) { 4321 /* 4322 * Slabcache can never be merged so we can use the name proper. 4323 * This is typically the case for debug situations. In that 4324 * case we can catch duplicate names easily. 4325 */ 4326 sysfs_remove_link(&slab_kset->kobj, s->name); 4327 name = s->name; 4328 } else { 4329 /* 4330 * Create a unique name for the slab as a target 4331 * for the symlinks. 4332 */ 4333 name = create_unique_id(s); 4334 } 4335 4336 s->kobj.kset = slab_kset; 4337 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 4338 if (err) { 4339 kobject_put(&s->kobj); 4340 return err; 4341 } 4342 4343 err = sysfs_create_group(&s->kobj, &slab_attr_group); 4344 if (err) 4345 return err; 4346 kobject_uevent(&s->kobj, KOBJ_ADD); 4347 if (!unmergeable) { 4348 /* Setup first alias */ 4349 sysfs_slab_alias(s, s->name); 4350 kfree(name); 4351 } 4352 return 0; 4353} 4354 4355static void sysfs_slab_remove(struct kmem_cache *s) 4356{ 4357 kobject_uevent(&s->kobj, KOBJ_REMOVE); 4358 kobject_del(&s->kobj); 4359 kobject_put(&s->kobj); 4360} 4361 4362/* 4363 * Need to buffer aliases during bootup until sysfs becomes 4364 * available lest we loose that information. 4365 */ 4366struct saved_alias { 4367 struct kmem_cache *s; 4368 const char *name; 4369 struct saved_alias *next; 4370}; 4371 4372static struct saved_alias *alias_list; 4373 4374static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 4375{ 4376 struct saved_alias *al; 4377 4378 if (slab_state == SYSFS) { 4379 /* 4380 * If we have a leftover link then remove it. 4381 */ 4382 sysfs_remove_link(&slab_kset->kobj, name); 4383 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 4384 } 4385 4386 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 4387 if (!al) 4388 return -ENOMEM; 4389 4390 al->s = s; 4391 al->name = name; 4392 al->next = alias_list; 4393 alias_list = al; 4394 return 0; 4395} 4396 4397static int __init slab_sysfs_init(void) 4398{ 4399 struct kmem_cache *s; 4400 int err; 4401 4402 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 4403 if (!slab_kset) { 4404 printk(KERN_ERR "Cannot register slab subsystem.\n"); 4405 return -ENOSYS; 4406 } 4407 4408 slab_state = SYSFS; 4409 4410 list_for_each_entry(s, &slab_caches, list) { 4411 err = sysfs_slab_add(s); 4412 if (err) 4413 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 4414 " to sysfs\n", s->name); 4415 } 4416 4417 while (alias_list) { 4418 struct saved_alias *al = alias_list; 4419 4420 alias_list = alias_list->next; 4421 err = sysfs_slab_alias(al->s, al->name); 4422 if (err) 4423 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 4424 " %s to sysfs\n", s->name); 4425 kfree(al); 4426 } 4427 4428 resiliency_test(); 4429 return 0; 4430} 4431 4432__initcall(slab_sysfs_init); 4433#endif 4434 4435/* 4436 * The /proc/slabinfo ABI 4437 */ 4438#ifdef CONFIG_SLABINFO 4439 4440ssize_t slabinfo_write(struct file *file, const char __user * buffer, 4441 size_t count, loff_t *ppos) 4442{ 4443 return -EINVAL; 4444} 4445 4446 4447static void print_slabinfo_header(struct seq_file *m) 4448{ 4449 seq_puts(m, "slabinfo - version: 2.1\n"); 4450 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 4451 "<objperslab> <pagesperslab>"); 4452 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 4453 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 4454 seq_putc(m, '\n'); 4455} 4456 4457static void *s_start(struct seq_file *m, loff_t *pos) 4458{ 4459 loff_t n = *pos; 4460 4461 down_read(&slub_lock); 4462 if (!n) 4463 print_slabinfo_header(m); 4464 4465 return seq_list_start(&slab_caches, *pos); 4466} 4467 4468static void *s_next(struct seq_file *m, void *p, loff_t *pos) 4469{ 4470 return seq_list_next(p, &slab_caches, pos); 4471} 4472 4473static void s_stop(struct seq_file *m, void *p) 4474{ 4475 up_read(&slub_lock); 4476} 4477 4478static int s_show(struct seq_file *m, void *p) 4479{ 4480 unsigned long nr_partials = 0; 4481 unsigned long nr_slabs = 0; 4482 unsigned long nr_inuse = 0; 4483 unsigned long nr_objs = 0; 4484 unsigned long nr_free = 0; 4485 struct kmem_cache *s; 4486 int node; 4487 4488 s = list_entry(p, struct kmem_cache, list); 4489 4490 for_each_online_node(node) { 4491 struct kmem_cache_node *n = get_node(s, node); 4492 4493 if (!n) 4494 continue; 4495 4496 nr_partials += n->nr_partial; 4497 nr_slabs += atomic_long_read(&n->nr_slabs); 4498 nr_objs += atomic_long_read(&n->total_objects); 4499 nr_free += count_partial(n, count_free); 4500 } 4501 4502 nr_inuse = nr_objs - nr_free; 4503 4504 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 4505 nr_objs, s->size, oo_objects(s->oo), 4506 (1 << oo_order(s->oo))); 4507 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 4508 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 4509 0UL); 4510 seq_putc(m, '\n'); 4511 return 0; 4512} 4513 4514const struct seq_operations slabinfo_op = { 4515 .start = s_start, 4516 .next = s_next, 4517 .stop = s_stop, 4518 .show = s_show, 4519}; 4520 4521#endif /* CONFIG_SLABINFO */ 4522