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