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 or atomic operatios 6 * and only uses a centralized lock to manage a pool of partial slabs. 7 * 8 * (C) 2007 SGI, Christoph Lameter 9 * (C) 2011 Linux Foundation, Christoph Lameter 10 */ 11 12 #include <linux/mm.h> 13 #include <linux/swap.h> /* struct reclaim_state */ 14 #include <linux/module.h> 15 #include <linux/bit_spinlock.h> 16 #include <linux/interrupt.h> 17 #include <linux/bitops.h> 18 #include <linux/slab.h> 19 #include <linux/proc_fs.h> 20 #include <linux/seq_file.h> 21 #include <linux/kmemcheck.h> 22 #include <linux/cpu.h> 23 #include <linux/cpuset.h> 24 #include <linux/mempolicy.h> 25 #include <linux/ctype.h> 26 #include <linux/debugobjects.h> 27 #include <linux/kallsyms.h> 28 #include <linux/memory.h> 29 #include <linux/math64.h> 30 #include <linux/fault-inject.h> 31 #include <linux/stacktrace.h> 32 33 #include <trace/events/kmem.h> 34 35 /* 36 * Lock order: 37 * 1. slub_lock (Global Semaphore) 38 * 2. node->list_lock 39 * 3. slab_lock(page) (Only on some arches and for debugging) 40 * 41 * slub_lock 42 * 43 * The role of the slub_lock is to protect the list of all the slabs 44 * and to synchronize major metadata changes to slab cache structures. 45 * 46 * The slab_lock is only used for debugging and on arches that do not 47 * have the ability to do a cmpxchg_double. It only protects the second 48 * double word in the page struct. Meaning 49 * A. page->freelist -> List of object free in a page 50 * B. page->counters -> Counters of objects 51 * C. page->frozen -> frozen state 52 * 53 * If a slab is frozen then it is exempt from list management. It is not 54 * on any list. The processor that froze the slab is the one who can 55 * perform list operations on the page. Other processors may put objects 56 * onto the freelist but the processor that froze the slab is the only 57 * one that can retrieve the objects from the page's freelist. 58 * 59 * The list_lock protects the partial and full list on each node and 60 * the partial slab counter. If taken then no new slabs may be added or 61 * removed from the lists nor make the number of partial slabs be modified. 62 * (Note that the total number of slabs is an atomic value that may be 63 * modified without taking the list lock). 64 * 65 * The list_lock is a centralized lock and thus we avoid taking it as 66 * much as possible. As long as SLUB does not have to handle partial 67 * slabs, operations can continue without any centralized lock. F.e. 68 * allocating a long series of objects that fill up slabs does not require 69 * the list lock. 70 * Interrupts are disabled during allocation and deallocation in order to 71 * make the slab allocator safe to use in the context of an irq. In addition 72 * interrupts are disabled to ensure that the processor does not change 73 * while handling per_cpu slabs, due to kernel preemption. 74 * 75 * SLUB assigns one slab for allocation to each processor. 76 * Allocations only occur from these slabs called cpu slabs. 77 * 78 * Slabs with free elements are kept on a partial list and during regular 79 * operations no list for full slabs is used. If an object in a full slab is 80 * freed then the slab will show up again on the partial lists. 81 * We track full slabs for debugging purposes though because otherwise we 82 * cannot scan all objects. 83 * 84 * Slabs are freed when they become empty. Teardown and setup is 85 * minimal so we rely on the page allocators per cpu caches for 86 * fast frees and allocs. 87 * 88 * Overloading of page flags that are otherwise used for LRU management. 89 * 90 * PageActive The slab is frozen and exempt from list processing. 91 * This means that the slab is dedicated to a purpose 92 * such as satisfying allocations for a specific 93 * processor. Objects may be freed in the slab while 94 * it is frozen but slab_free will then skip the usual 95 * list operations. It is up to the processor holding 96 * the slab to integrate the slab into the slab lists 97 * when the slab is no longer needed. 98 * 99 * One use of this flag is to mark slabs that are 100 * used for allocations. Then such a slab becomes a cpu 101 * slab. The cpu slab may be equipped with an additional 102 * freelist that allows lockless access to 103 * free objects in addition to the regular freelist 104 * that requires the slab lock. 105 * 106 * PageError Slab requires special handling due to debug 107 * options set. This moves slab handling out of 108 * the fast path and disables lockless freelists. 109 */ 110 111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 112 SLAB_TRACE | SLAB_DEBUG_FREE) 113 114 static inline int kmem_cache_debug(struct kmem_cache *s) 115 { 116 #ifdef CONFIG_SLUB_DEBUG 117 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 118 #else 119 return 0; 120 #endif 121 } 122 123 /* 124 * Issues still to be resolved: 125 * 126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 127 * 128 * - Variable sizing of the per node arrays 129 */ 130 131 /* Enable to test recovery from slab corruption on boot */ 132 #undef SLUB_RESILIENCY_TEST 133 134 /* Enable to log cmpxchg failures */ 135 #undef SLUB_DEBUG_CMPXCHG 136 137 /* 138 * Mininum number of partial slabs. These will be left on the partial 139 * lists even if they are empty. kmem_cache_shrink may reclaim them. 140 */ 141 #define MIN_PARTIAL 5 142 143 /* 144 * Maximum number of desirable partial slabs. 145 * The existence of more partial slabs makes kmem_cache_shrink 146 * sort the partial list by the number of objects in the. 147 */ 148 #define MAX_PARTIAL 10 149 150 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 151 SLAB_POISON | SLAB_STORE_USER) 152 153 /* 154 * Debugging flags that require metadata to be stored in the slab. These get 155 * disabled when slub_debug=O is used and a cache's min order increases with 156 * metadata. 157 */ 158 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 159 160 /* 161 * Set of flags that will prevent slab merging 162 */ 163 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ 165 SLAB_FAILSLAB) 166 167 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 168 SLAB_CACHE_DMA | SLAB_NOTRACK) 169 170 #define OO_SHIFT 16 171 #define OO_MASK ((1 << OO_SHIFT) - 1) 172 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ 173 174 /* Internal SLUB flags */ 175 #define __OBJECT_POISON 0x80000000UL /* Poison object */ 176 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */ 177 178 static int kmem_size = sizeof(struct kmem_cache); 179 180 #ifdef CONFIG_SMP 181 static struct notifier_block slab_notifier; 182 #endif 183 184 static enum { 185 DOWN, /* No slab functionality available */ 186 PARTIAL, /* Kmem_cache_node works */ 187 UP, /* Everything works but does not show up in sysfs */ 188 SYSFS /* Sysfs up */ 189 } slab_state = DOWN; 190 191 /* A list of all slab caches on the system */ 192 static DECLARE_RWSEM(slub_lock); 193 static LIST_HEAD(slab_caches); 194 195 /* 196 * Tracking user of a slab. 197 */ 198 #define TRACK_ADDRS_COUNT 16 199 struct track { 200 unsigned long addr; /* Called from address */ 201 #ifdef CONFIG_STACKTRACE 202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 203 #endif 204 int cpu; /* Was running on cpu */ 205 int pid; /* Pid context */ 206 unsigned long when; /* When did the operation occur */ 207 }; 208 209 enum track_item { TRACK_ALLOC, TRACK_FREE }; 210 211 #ifdef CONFIG_SYSFS 212 static int sysfs_slab_add(struct kmem_cache *); 213 static int sysfs_slab_alias(struct kmem_cache *, const char *); 214 static void sysfs_slab_remove(struct kmem_cache *); 215 216 #else 217 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 218 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 219 { return 0; } 220 static inline void sysfs_slab_remove(struct kmem_cache *s) 221 { 222 kfree(s->name); 223 kfree(s); 224 } 225 226 #endif 227 228 static inline void stat(const struct kmem_cache *s, enum stat_item si) 229 { 230 #ifdef CONFIG_SLUB_STATS 231 __this_cpu_inc(s->cpu_slab->stat[si]); 232 #endif 233 } 234 235 /******************************************************************** 236 * Core slab cache functions 237 *******************************************************************/ 238 239 int slab_is_available(void) 240 { 241 return slab_state >= UP; 242 } 243 244 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 245 { 246 return s->node[node]; 247 } 248 249 /* Verify that a pointer has an address that is valid within a slab page */ 250 static inline int check_valid_pointer(struct kmem_cache *s, 251 struct page *page, const void *object) 252 { 253 void *base; 254 255 if (!object) 256 return 1; 257 258 base = page_address(page); 259 if (object < base || object >= base + page->objects * s->size || 260 (object - base) % s->size) { 261 return 0; 262 } 263 264 return 1; 265 } 266 267 static inline void *get_freepointer(struct kmem_cache *s, void *object) 268 { 269 return *(void **)(object + s->offset); 270 } 271 272 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 273 { 274 void *p; 275 276 #ifdef CONFIG_DEBUG_PAGEALLOC 277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p)); 278 #else 279 p = get_freepointer(s, object); 280 #endif 281 return p; 282 } 283 284 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 285 { 286 *(void **)(object + s->offset) = fp; 287 } 288 289 /* Loop over all objects in a slab */ 290 #define for_each_object(__p, __s, __addr, __objects) \ 291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 292 __p += (__s)->size) 293 294 /* Determine object index from a given position */ 295 static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 296 { 297 return (p - addr) / s->size; 298 } 299 300 static inline size_t slab_ksize(const struct kmem_cache *s) 301 { 302 #ifdef CONFIG_SLUB_DEBUG 303 /* 304 * Debugging requires use of the padding between object 305 * and whatever may come after it. 306 */ 307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 308 return s->objsize; 309 310 #endif 311 /* 312 * If we have the need to store the freelist pointer 313 * back there or track user information then we can 314 * only use the space before that information. 315 */ 316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 317 return s->inuse; 318 /* 319 * Else we can use all the padding etc for the allocation 320 */ 321 return s->size; 322 } 323 324 static inline int order_objects(int order, unsigned long size, int reserved) 325 { 326 return ((PAGE_SIZE << order) - reserved) / size; 327 } 328 329 static inline struct kmem_cache_order_objects oo_make(int order, 330 unsigned long size, int reserved) 331 { 332 struct kmem_cache_order_objects x = { 333 (order << OO_SHIFT) + order_objects(order, size, reserved) 334 }; 335 336 return x; 337 } 338 339 static inline int oo_order(struct kmem_cache_order_objects x) 340 { 341 return x.x >> OO_SHIFT; 342 } 343 344 static inline int oo_objects(struct kmem_cache_order_objects x) 345 { 346 return x.x & OO_MASK; 347 } 348 349 /* 350 * Per slab locking using the pagelock 351 */ 352 static __always_inline void slab_lock(struct page *page) 353 { 354 bit_spin_lock(PG_locked, &page->flags); 355 } 356 357 static __always_inline void slab_unlock(struct page *page) 358 { 359 __bit_spin_unlock(PG_locked, &page->flags); 360 } 361 362 /* Interrupts must be disabled (for the fallback code to work right) */ 363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 364 void *freelist_old, unsigned long counters_old, 365 void *freelist_new, unsigned long counters_new, 366 const char *n) 367 { 368 VM_BUG_ON(!irqs_disabled()); 369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 371 if (s->flags & __CMPXCHG_DOUBLE) { 372 if (cmpxchg_double(&page->freelist, &page->counters, 373 freelist_old, counters_old, 374 freelist_new, counters_new)) 375 return 1; 376 } else 377 #endif 378 { 379 slab_lock(page); 380 if (page->freelist == freelist_old && page->counters == counters_old) { 381 page->freelist = freelist_new; 382 page->counters = counters_new; 383 slab_unlock(page); 384 return 1; 385 } 386 slab_unlock(page); 387 } 388 389 cpu_relax(); 390 stat(s, CMPXCHG_DOUBLE_FAIL); 391 392 #ifdef SLUB_DEBUG_CMPXCHG 393 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name); 394 #endif 395 396 return 0; 397 } 398 399 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 400 void *freelist_old, unsigned long counters_old, 401 void *freelist_new, unsigned long counters_new, 402 const char *n) 403 { 404 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 405 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 406 if (s->flags & __CMPXCHG_DOUBLE) { 407 if (cmpxchg_double(&page->freelist, &page->counters, 408 freelist_old, counters_old, 409 freelist_new, counters_new)) 410 return 1; 411 } else 412 #endif 413 { 414 unsigned long flags; 415 416 local_irq_save(flags); 417 slab_lock(page); 418 if (page->freelist == freelist_old && page->counters == counters_old) { 419 page->freelist = freelist_new; 420 page->counters = counters_new; 421 slab_unlock(page); 422 local_irq_restore(flags); 423 return 1; 424 } 425 slab_unlock(page); 426 local_irq_restore(flags); 427 } 428 429 cpu_relax(); 430 stat(s, CMPXCHG_DOUBLE_FAIL); 431 432 #ifdef SLUB_DEBUG_CMPXCHG 433 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name); 434 #endif 435 436 return 0; 437 } 438 439 #ifdef CONFIG_SLUB_DEBUG 440 /* 441 * Determine a map of object in use on a page. 442 * 443 * Node listlock must be held to guarantee that the page does 444 * not vanish from under us. 445 */ 446 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) 447 { 448 void *p; 449 void *addr = page_address(page); 450 451 for (p = page->freelist; p; p = get_freepointer(s, p)) 452 set_bit(slab_index(p, s, addr), map); 453 } 454 455 /* 456 * Debug settings: 457 */ 458 #ifdef CONFIG_SLUB_DEBUG_ON 459 static int slub_debug = DEBUG_DEFAULT_FLAGS; 460 #else 461 static int slub_debug; 462 #endif 463 464 static char *slub_debug_slabs; 465 static int disable_higher_order_debug; 466 467 /* 468 * Object debugging 469 */ 470 static void print_section(char *text, u8 *addr, unsigned int length) 471 { 472 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr, 473 length, 1); 474 } 475 476 static struct track *get_track(struct kmem_cache *s, void *object, 477 enum track_item alloc) 478 { 479 struct track *p; 480 481 if (s->offset) 482 p = object + s->offset + sizeof(void *); 483 else 484 p = object + s->inuse; 485 486 return p + alloc; 487 } 488 489 static void set_track(struct kmem_cache *s, void *object, 490 enum track_item alloc, unsigned long addr) 491 { 492 struct track *p = get_track(s, object, alloc); 493 494 if (addr) { 495 #ifdef CONFIG_STACKTRACE 496 struct stack_trace trace; 497 int i; 498 499 trace.nr_entries = 0; 500 trace.max_entries = TRACK_ADDRS_COUNT; 501 trace.entries = p->addrs; 502 trace.skip = 3; 503 save_stack_trace(&trace); 504 505 /* See rant in lockdep.c */ 506 if (trace.nr_entries != 0 && 507 trace.entries[trace.nr_entries - 1] == ULONG_MAX) 508 trace.nr_entries--; 509 510 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) 511 p->addrs[i] = 0; 512 #endif 513 p->addr = addr; 514 p->cpu = smp_processor_id(); 515 p->pid = current->pid; 516 p->when = jiffies; 517 } else 518 memset(p, 0, sizeof(struct track)); 519 } 520 521 static void init_tracking(struct kmem_cache *s, void *object) 522 { 523 if (!(s->flags & SLAB_STORE_USER)) 524 return; 525 526 set_track(s, object, TRACK_FREE, 0UL); 527 set_track(s, object, TRACK_ALLOC, 0UL); 528 } 529 530 static void print_track(const char *s, struct track *t) 531 { 532 if (!t->addr) 533 return; 534 535 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 536 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 537 #ifdef CONFIG_STACKTRACE 538 { 539 int i; 540 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 541 if (t->addrs[i]) 542 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]); 543 else 544 break; 545 } 546 #endif 547 } 548 549 static void print_tracking(struct kmem_cache *s, void *object) 550 { 551 if (!(s->flags & SLAB_STORE_USER)) 552 return; 553 554 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 555 print_track("Freed", get_track(s, object, TRACK_FREE)); 556 } 557 558 static void print_page_info(struct page *page) 559 { 560 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 561 page, page->objects, page->inuse, page->freelist, page->flags); 562 563 } 564 565 static void slab_bug(struct kmem_cache *s, char *fmt, ...) 566 { 567 va_list args; 568 char buf[100]; 569 570 va_start(args, fmt); 571 vsnprintf(buf, sizeof(buf), fmt, args); 572 va_end(args); 573 printk(KERN_ERR "========================================" 574 "=====================================\n"); 575 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf); 576 printk(KERN_ERR "----------------------------------------" 577 "-------------------------------------\n\n"); 578 } 579 580 static void slab_fix(struct kmem_cache *s, char *fmt, ...) 581 { 582 va_list args; 583 char buf[100]; 584 585 va_start(args, fmt); 586 vsnprintf(buf, sizeof(buf), fmt, args); 587 va_end(args); 588 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 589 } 590 591 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 592 { 593 unsigned int off; /* Offset of last byte */ 594 u8 *addr = page_address(page); 595 596 print_tracking(s, p); 597 598 print_page_info(page); 599 600 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 601 p, p - addr, get_freepointer(s, p)); 602 603 if (p > addr + 16) 604 print_section("Bytes b4 ", p - 16, 16); 605 606 print_section("Object ", p, min_t(unsigned long, s->objsize, 607 PAGE_SIZE)); 608 if (s->flags & SLAB_RED_ZONE) 609 print_section("Redzone ", p + s->objsize, 610 s->inuse - s->objsize); 611 612 if (s->offset) 613 off = s->offset + sizeof(void *); 614 else 615 off = s->inuse; 616 617 if (s->flags & SLAB_STORE_USER) 618 off += 2 * sizeof(struct track); 619 620 if (off != s->size) 621 /* Beginning of the filler is the free pointer */ 622 print_section("Padding ", p + off, s->size - off); 623 624 dump_stack(); 625 } 626 627 static void object_err(struct kmem_cache *s, struct page *page, 628 u8 *object, char *reason) 629 { 630 slab_bug(s, "%s", reason); 631 print_trailer(s, page, object); 632 } 633 634 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...) 635 { 636 va_list args; 637 char buf[100]; 638 639 va_start(args, fmt); 640 vsnprintf(buf, sizeof(buf), fmt, args); 641 va_end(args); 642 slab_bug(s, "%s", buf); 643 print_page_info(page); 644 dump_stack(); 645 } 646 647 static void init_object(struct kmem_cache *s, void *object, u8 val) 648 { 649 u8 *p = object; 650 651 if (s->flags & __OBJECT_POISON) { 652 memset(p, POISON_FREE, s->objsize - 1); 653 p[s->objsize - 1] = POISON_END; 654 } 655 656 if (s->flags & SLAB_RED_ZONE) 657 memset(p + s->objsize, val, s->inuse - s->objsize); 658 } 659 660 static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 661 void *from, void *to) 662 { 663 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 664 memset(from, data, to - from); 665 } 666 667 static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 668 u8 *object, char *what, 669 u8 *start, unsigned int value, unsigned int bytes) 670 { 671 u8 *fault; 672 u8 *end; 673 674 fault = memchr_inv(start, value, bytes); 675 if (!fault) 676 return 1; 677 678 end = start + bytes; 679 while (end > fault && end[-1] == value) 680 end--; 681 682 slab_bug(s, "%s overwritten", what); 683 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 684 fault, end - 1, fault[0], value); 685 print_trailer(s, page, object); 686 687 restore_bytes(s, what, value, fault, end); 688 return 0; 689 } 690 691 /* 692 * Object layout: 693 * 694 * object address 695 * Bytes of the object to be managed. 696 * If the freepointer may overlay the object then the free 697 * pointer is the first word of the object. 698 * 699 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 700 * 0xa5 (POISON_END) 701 * 702 * object + s->objsize 703 * Padding to reach word boundary. This is also used for Redzoning. 704 * Padding is extended by another word if Redzoning is enabled and 705 * objsize == inuse. 706 * 707 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 708 * 0xcc (RED_ACTIVE) for objects in use. 709 * 710 * object + s->inuse 711 * Meta data starts here. 712 * 713 * A. Free pointer (if we cannot overwrite object on free) 714 * B. Tracking data for SLAB_STORE_USER 715 * C. Padding to reach required alignment boundary or at mininum 716 * one word if debugging is on to be able to detect writes 717 * before the word boundary. 718 * 719 * Padding is done using 0x5a (POISON_INUSE) 720 * 721 * object + s->size 722 * Nothing is used beyond s->size. 723 * 724 * If slabcaches are merged then the objsize and inuse boundaries are mostly 725 * ignored. And therefore no slab options that rely on these boundaries 726 * may be used with merged slabcaches. 727 */ 728 729 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 730 { 731 unsigned long off = s->inuse; /* The end of info */ 732 733 if (s->offset) 734 /* Freepointer is placed after the object. */ 735 off += sizeof(void *); 736 737 if (s->flags & SLAB_STORE_USER) 738 /* We also have user information there */ 739 off += 2 * sizeof(struct track); 740 741 if (s->size == off) 742 return 1; 743 744 return check_bytes_and_report(s, page, p, "Object padding", 745 p + off, POISON_INUSE, s->size - off); 746 } 747 748 /* Check the pad bytes at the end of a slab page */ 749 static int slab_pad_check(struct kmem_cache *s, struct page *page) 750 { 751 u8 *start; 752 u8 *fault; 753 u8 *end; 754 int length; 755 int remainder; 756 757 if (!(s->flags & SLAB_POISON)) 758 return 1; 759 760 start = page_address(page); 761 length = (PAGE_SIZE << compound_order(page)) - s->reserved; 762 end = start + length; 763 remainder = length % s->size; 764 if (!remainder) 765 return 1; 766 767 fault = memchr_inv(end - remainder, POISON_INUSE, remainder); 768 if (!fault) 769 return 1; 770 while (end > fault && end[-1] == POISON_INUSE) 771 end--; 772 773 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 774 print_section("Padding ", end - remainder, remainder); 775 776 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 777 return 0; 778 } 779 780 static int check_object(struct kmem_cache *s, struct page *page, 781 void *object, u8 val) 782 { 783 u8 *p = object; 784 u8 *endobject = object + s->objsize; 785 786 if (s->flags & SLAB_RED_ZONE) { 787 if (!check_bytes_and_report(s, page, object, "Redzone", 788 endobject, val, s->inuse - s->objsize)) 789 return 0; 790 } else { 791 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) { 792 check_bytes_and_report(s, page, p, "Alignment padding", 793 endobject, POISON_INUSE, s->inuse - s->objsize); 794 } 795 } 796 797 if (s->flags & SLAB_POISON) { 798 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 799 (!check_bytes_and_report(s, page, p, "Poison", p, 800 POISON_FREE, s->objsize - 1) || 801 !check_bytes_and_report(s, page, p, "Poison", 802 p + s->objsize - 1, POISON_END, 1))) 803 return 0; 804 /* 805 * check_pad_bytes cleans up on its own. 806 */ 807 check_pad_bytes(s, page, p); 808 } 809 810 if (!s->offset && val == SLUB_RED_ACTIVE) 811 /* 812 * Object and freepointer overlap. Cannot check 813 * freepointer while object is allocated. 814 */ 815 return 1; 816 817 /* Check free pointer validity */ 818 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 819 object_err(s, page, p, "Freepointer corrupt"); 820 /* 821 * No choice but to zap it and thus lose the remainder 822 * of the free objects in this slab. May cause 823 * another error because the object count is now wrong. 824 */ 825 set_freepointer(s, p, NULL); 826 return 0; 827 } 828 return 1; 829 } 830 831 static int check_slab(struct kmem_cache *s, struct page *page) 832 { 833 int maxobj; 834 835 VM_BUG_ON(!irqs_disabled()); 836 837 if (!PageSlab(page)) { 838 slab_err(s, page, "Not a valid slab page"); 839 return 0; 840 } 841 842 maxobj = order_objects(compound_order(page), s->size, s->reserved); 843 if (page->objects > maxobj) { 844 slab_err(s, page, "objects %u > max %u", 845 s->name, page->objects, maxobj); 846 return 0; 847 } 848 if (page->inuse > page->objects) { 849 slab_err(s, page, "inuse %u > max %u", 850 s->name, page->inuse, page->objects); 851 return 0; 852 } 853 /* Slab_pad_check fixes things up after itself */ 854 slab_pad_check(s, page); 855 return 1; 856 } 857 858 /* 859 * Determine if a certain object on a page is on the freelist. Must hold the 860 * slab lock to guarantee that the chains are in a consistent state. 861 */ 862 static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 863 { 864 int nr = 0; 865 void *fp; 866 void *object = NULL; 867 unsigned long max_objects; 868 869 fp = page->freelist; 870 while (fp && nr <= page->objects) { 871 if (fp == search) 872 return 1; 873 if (!check_valid_pointer(s, page, fp)) { 874 if (object) { 875 object_err(s, page, object, 876 "Freechain corrupt"); 877 set_freepointer(s, object, NULL); 878 break; 879 } else { 880 slab_err(s, page, "Freepointer corrupt"); 881 page->freelist = NULL; 882 page->inuse = page->objects; 883 slab_fix(s, "Freelist cleared"); 884 return 0; 885 } 886 break; 887 } 888 object = fp; 889 fp = get_freepointer(s, object); 890 nr++; 891 } 892 893 max_objects = order_objects(compound_order(page), s->size, s->reserved); 894 if (max_objects > MAX_OBJS_PER_PAGE) 895 max_objects = MAX_OBJS_PER_PAGE; 896 897 if (page->objects != max_objects) { 898 slab_err(s, page, "Wrong number of objects. Found %d but " 899 "should be %d", page->objects, max_objects); 900 page->objects = max_objects; 901 slab_fix(s, "Number of objects adjusted."); 902 } 903 if (page->inuse != page->objects - nr) { 904 slab_err(s, page, "Wrong object count. Counter is %d but " 905 "counted were %d", page->inuse, page->objects - nr); 906 page->inuse = page->objects - nr; 907 slab_fix(s, "Object count adjusted."); 908 } 909 return search == NULL; 910 } 911 912 static void trace(struct kmem_cache *s, struct page *page, void *object, 913 int alloc) 914 { 915 if (s->flags & SLAB_TRACE) { 916 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 917 s->name, 918 alloc ? "alloc" : "free", 919 object, page->inuse, 920 page->freelist); 921 922 if (!alloc) 923 print_section("Object ", (void *)object, s->objsize); 924 925 dump_stack(); 926 } 927 } 928 929 /* 930 * Hooks for other subsystems that check memory allocations. In a typical 931 * production configuration these hooks all should produce no code at all. 932 */ 933 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 934 { 935 flags &= gfp_allowed_mask; 936 lockdep_trace_alloc(flags); 937 might_sleep_if(flags & __GFP_WAIT); 938 939 return should_failslab(s->objsize, flags, s->flags); 940 } 941 942 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object) 943 { 944 flags &= gfp_allowed_mask; 945 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s)); 946 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags); 947 } 948 949 static inline void slab_free_hook(struct kmem_cache *s, void *x) 950 { 951 kmemleak_free_recursive(x, s->flags); 952 953 /* 954 * Trouble is that we may no longer disable interupts in the fast path 955 * So in order to make the debug calls that expect irqs to be 956 * disabled we need to disable interrupts temporarily. 957 */ 958 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) 959 { 960 unsigned long flags; 961 962 local_irq_save(flags); 963 kmemcheck_slab_free(s, x, s->objsize); 964 debug_check_no_locks_freed(x, s->objsize); 965 local_irq_restore(flags); 966 } 967 #endif 968 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 969 debug_check_no_obj_freed(x, s->objsize); 970 } 971 972 /* 973 * Tracking of fully allocated slabs for debugging purposes. 974 * 975 * list_lock must be held. 976 */ 977 static void add_full(struct kmem_cache *s, 978 struct kmem_cache_node *n, struct page *page) 979 { 980 if (!(s->flags & SLAB_STORE_USER)) 981 return; 982 983 list_add(&page->lru, &n->full); 984 } 985 986 /* 987 * list_lock must be held. 988 */ 989 static void remove_full(struct kmem_cache *s, struct page *page) 990 { 991 if (!(s->flags & SLAB_STORE_USER)) 992 return; 993 994 list_del(&page->lru); 995 } 996 997 /* Tracking of the number of slabs for debugging purposes */ 998 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 999 { 1000 struct kmem_cache_node *n = get_node(s, node); 1001 1002 return atomic_long_read(&n->nr_slabs); 1003 } 1004 1005 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1006 { 1007 return atomic_long_read(&n->nr_slabs); 1008 } 1009 1010 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 1011 { 1012 struct kmem_cache_node *n = get_node(s, node); 1013 1014 /* 1015 * May be called early in order to allocate a slab for the 1016 * kmem_cache_node structure. Solve the chicken-egg 1017 * dilemma by deferring the increment of the count during 1018 * bootstrap (see early_kmem_cache_node_alloc). 1019 */ 1020 if (n) { 1021 atomic_long_inc(&n->nr_slabs); 1022 atomic_long_add(objects, &n->total_objects); 1023 } 1024 } 1025 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 1026 { 1027 struct kmem_cache_node *n = get_node(s, node); 1028 1029 atomic_long_dec(&n->nr_slabs); 1030 atomic_long_sub(objects, &n->total_objects); 1031 } 1032 1033 /* Object debug checks for alloc/free paths */ 1034 static void setup_object_debug(struct kmem_cache *s, struct page *page, 1035 void *object) 1036 { 1037 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 1038 return; 1039 1040 init_object(s, object, SLUB_RED_INACTIVE); 1041 init_tracking(s, object); 1042 } 1043 1044 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page, 1045 void *object, unsigned long addr) 1046 { 1047 if (!check_slab(s, page)) 1048 goto bad; 1049 1050 if (!check_valid_pointer(s, page, object)) { 1051 object_err(s, page, object, "Freelist Pointer check fails"); 1052 goto bad; 1053 } 1054 1055 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1056 goto bad; 1057 1058 /* Success perform special debug activities for allocs */ 1059 if (s->flags & SLAB_STORE_USER) 1060 set_track(s, object, TRACK_ALLOC, addr); 1061 trace(s, page, object, 1); 1062 init_object(s, object, SLUB_RED_ACTIVE); 1063 return 1; 1064 1065 bad: 1066 if (PageSlab(page)) { 1067 /* 1068 * If this is a slab page then lets do the best we can 1069 * to avoid issues in the future. Marking all objects 1070 * as used avoids touching the remaining objects. 1071 */ 1072 slab_fix(s, "Marking all objects used"); 1073 page->inuse = page->objects; 1074 page->freelist = NULL; 1075 } 1076 return 0; 1077 } 1078 1079 static noinline int free_debug_processing(struct kmem_cache *s, 1080 struct page *page, void *object, unsigned long addr) 1081 { 1082 unsigned long flags; 1083 int rc = 0; 1084 1085 local_irq_save(flags); 1086 slab_lock(page); 1087 1088 if (!check_slab(s, page)) 1089 goto fail; 1090 1091 if (!check_valid_pointer(s, page, object)) { 1092 slab_err(s, page, "Invalid object pointer 0x%p", object); 1093 goto fail; 1094 } 1095 1096 if (on_freelist(s, page, object)) { 1097 object_err(s, page, object, "Object already free"); 1098 goto fail; 1099 } 1100 1101 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1102 goto out; 1103 1104 if (unlikely(s != page->slab)) { 1105 if (!PageSlab(page)) { 1106 slab_err(s, page, "Attempt to free object(0x%p) " 1107 "outside of slab", object); 1108 } else if (!page->slab) { 1109 printk(KERN_ERR 1110 "SLUB <none>: no slab for object 0x%p.\n", 1111 object); 1112 dump_stack(); 1113 } else 1114 object_err(s, page, object, 1115 "page slab pointer corrupt."); 1116 goto fail; 1117 } 1118 1119 if (s->flags & SLAB_STORE_USER) 1120 set_track(s, object, TRACK_FREE, addr); 1121 trace(s, page, object, 0); 1122 init_object(s, object, SLUB_RED_INACTIVE); 1123 rc = 1; 1124 out: 1125 slab_unlock(page); 1126 local_irq_restore(flags); 1127 return rc; 1128 1129 fail: 1130 slab_fix(s, "Object at 0x%p not freed", object); 1131 goto out; 1132 } 1133 1134 static int __init setup_slub_debug(char *str) 1135 { 1136 slub_debug = DEBUG_DEFAULT_FLAGS; 1137 if (*str++ != '=' || !*str) 1138 /* 1139 * No options specified. Switch on full debugging. 1140 */ 1141 goto out; 1142 1143 if (*str == ',') 1144 /* 1145 * No options but restriction on slabs. This means full 1146 * debugging for slabs matching a pattern. 1147 */ 1148 goto check_slabs; 1149 1150 if (tolower(*str) == 'o') { 1151 /* 1152 * Avoid enabling debugging on caches if its minimum order 1153 * would increase as a result. 1154 */ 1155 disable_higher_order_debug = 1; 1156 goto out; 1157 } 1158 1159 slub_debug = 0; 1160 if (*str == '-') 1161 /* 1162 * Switch off all debugging measures. 1163 */ 1164 goto out; 1165 1166 /* 1167 * Determine which debug features should be switched on 1168 */ 1169 for (; *str && *str != ','; str++) { 1170 switch (tolower(*str)) { 1171 case 'f': 1172 slub_debug |= SLAB_DEBUG_FREE; 1173 break; 1174 case 'z': 1175 slub_debug |= SLAB_RED_ZONE; 1176 break; 1177 case 'p': 1178 slub_debug |= SLAB_POISON; 1179 break; 1180 case 'u': 1181 slub_debug |= SLAB_STORE_USER; 1182 break; 1183 case 't': 1184 slub_debug |= SLAB_TRACE; 1185 break; 1186 case 'a': 1187 slub_debug |= SLAB_FAILSLAB; 1188 break; 1189 default: 1190 printk(KERN_ERR "slub_debug option '%c' " 1191 "unknown. skipped\n", *str); 1192 } 1193 } 1194 1195 check_slabs: 1196 if (*str == ',') 1197 slub_debug_slabs = str + 1; 1198 out: 1199 return 1; 1200 } 1201 1202 __setup("slub_debug", setup_slub_debug); 1203 1204 static unsigned long kmem_cache_flags(unsigned long objsize, 1205 unsigned long flags, const char *name, 1206 void (*ctor)(void *)) 1207 { 1208 /* 1209 * Enable debugging if selected on the kernel commandline. 1210 */ 1211 if (slub_debug && (!slub_debug_slabs || 1212 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))) 1213 flags |= slub_debug; 1214 1215 return flags; 1216 } 1217 #else 1218 static inline void setup_object_debug(struct kmem_cache *s, 1219 struct page *page, void *object) {} 1220 1221 static inline int alloc_debug_processing(struct kmem_cache *s, 1222 struct page *page, void *object, unsigned long addr) { return 0; } 1223 1224 static inline int free_debug_processing(struct kmem_cache *s, 1225 struct page *page, void *object, unsigned long addr) { return 0; } 1226 1227 static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1228 { return 1; } 1229 static inline int check_object(struct kmem_cache *s, struct page *page, 1230 void *object, u8 val) { return 1; } 1231 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, 1232 struct page *page) {} 1233 static inline void remove_full(struct kmem_cache *s, struct page *page) {} 1234 static inline unsigned long kmem_cache_flags(unsigned long objsize, 1235 unsigned long flags, const char *name, 1236 void (*ctor)(void *)) 1237 { 1238 return flags; 1239 } 1240 #define slub_debug 0 1241 1242 #define disable_higher_order_debug 0 1243 1244 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1245 { return 0; } 1246 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1247 { return 0; } 1248 static inline void inc_slabs_node(struct kmem_cache *s, int node, 1249 int objects) {} 1250 static inline void dec_slabs_node(struct kmem_cache *s, int node, 1251 int objects) {} 1252 1253 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 1254 { return 0; } 1255 1256 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, 1257 void *object) {} 1258 1259 static inline void slab_free_hook(struct kmem_cache *s, void *x) {} 1260 1261 #endif /* CONFIG_SLUB_DEBUG */ 1262 1263 /* 1264 * Slab allocation and freeing 1265 */ 1266 static inline struct page *alloc_slab_page(gfp_t flags, int node, 1267 struct kmem_cache_order_objects oo) 1268 { 1269 int order = oo_order(oo); 1270 1271 flags |= __GFP_NOTRACK; 1272 1273 if (node == NUMA_NO_NODE) 1274 return alloc_pages(flags, order); 1275 else 1276 return alloc_pages_exact_node(node, flags, order); 1277 } 1278 1279 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1280 { 1281 struct page *page; 1282 struct kmem_cache_order_objects oo = s->oo; 1283 gfp_t alloc_gfp; 1284 1285 flags &= gfp_allowed_mask; 1286 1287 if (flags & __GFP_WAIT) 1288 local_irq_enable(); 1289 1290 flags |= s->allocflags; 1291 1292 /* 1293 * Let the initial higher-order allocation fail under memory pressure 1294 * so we fall-back to the minimum order allocation. 1295 */ 1296 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1297 1298 page = alloc_slab_page(alloc_gfp, node, oo); 1299 if (unlikely(!page)) { 1300 oo = s->min; 1301 /* 1302 * Allocation may have failed due to fragmentation. 1303 * Try a lower order alloc if possible 1304 */ 1305 page = alloc_slab_page(flags, node, oo); 1306 1307 if (page) 1308 stat(s, ORDER_FALLBACK); 1309 } 1310 1311 if (flags & __GFP_WAIT) 1312 local_irq_disable(); 1313 1314 if (!page) 1315 return NULL; 1316 1317 if (kmemcheck_enabled 1318 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1319 int pages = 1 << oo_order(oo); 1320 1321 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node); 1322 1323 /* 1324 * Objects from caches that have a constructor don't get 1325 * cleared when they're allocated, so we need to do it here. 1326 */ 1327 if (s->ctor) 1328 kmemcheck_mark_uninitialized_pages(page, pages); 1329 else 1330 kmemcheck_mark_unallocated_pages(page, pages); 1331 } 1332 1333 page->objects = oo_objects(oo); 1334 mod_zone_page_state(page_zone(page), 1335 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1336 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1337 1 << oo_order(oo)); 1338 1339 return page; 1340 } 1341 1342 static void setup_object(struct kmem_cache *s, struct page *page, 1343 void *object) 1344 { 1345 setup_object_debug(s, page, object); 1346 if (unlikely(s->ctor)) 1347 s->ctor(object); 1348 } 1349 1350 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1351 { 1352 struct page *page; 1353 void *start; 1354 void *last; 1355 void *p; 1356 1357 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1358 1359 page = allocate_slab(s, 1360 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1361 if (!page) 1362 goto out; 1363 1364 inc_slabs_node(s, page_to_nid(page), page->objects); 1365 page->slab = s; 1366 page->flags |= 1 << PG_slab; 1367 1368 start = page_address(page); 1369 1370 if (unlikely(s->flags & SLAB_POISON)) 1371 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page)); 1372 1373 last = start; 1374 for_each_object(p, s, start, page->objects) { 1375 setup_object(s, page, last); 1376 set_freepointer(s, last, p); 1377 last = p; 1378 } 1379 setup_object(s, page, last); 1380 set_freepointer(s, last, NULL); 1381 1382 page->freelist = start; 1383 page->inuse = page->objects; 1384 page->frozen = 1; 1385 out: 1386 return page; 1387 } 1388 1389 static void __free_slab(struct kmem_cache *s, struct page *page) 1390 { 1391 int order = compound_order(page); 1392 int pages = 1 << order; 1393 1394 if (kmem_cache_debug(s)) { 1395 void *p; 1396 1397 slab_pad_check(s, page); 1398 for_each_object(p, s, page_address(page), 1399 page->objects) 1400 check_object(s, page, p, SLUB_RED_INACTIVE); 1401 } 1402 1403 kmemcheck_free_shadow(page, compound_order(page)); 1404 1405 mod_zone_page_state(page_zone(page), 1406 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1407 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1408 -pages); 1409 1410 __ClearPageSlab(page); 1411 reset_page_mapcount(page); 1412 if (current->reclaim_state) 1413 current->reclaim_state->reclaimed_slab += pages; 1414 __free_pages(page, order); 1415 } 1416 1417 #define need_reserve_slab_rcu \ 1418 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) 1419 1420 static void rcu_free_slab(struct rcu_head *h) 1421 { 1422 struct page *page; 1423 1424 if (need_reserve_slab_rcu) 1425 page = virt_to_head_page(h); 1426 else 1427 page = container_of((struct list_head *)h, struct page, lru); 1428 1429 __free_slab(page->slab, page); 1430 } 1431 1432 static void free_slab(struct kmem_cache *s, struct page *page) 1433 { 1434 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1435 struct rcu_head *head; 1436 1437 if (need_reserve_slab_rcu) { 1438 int order = compound_order(page); 1439 int offset = (PAGE_SIZE << order) - s->reserved; 1440 1441 VM_BUG_ON(s->reserved != sizeof(*head)); 1442 head = page_address(page) + offset; 1443 } else { 1444 /* 1445 * RCU free overloads the RCU head over the LRU 1446 */ 1447 head = (void *)&page->lru; 1448 } 1449 1450 call_rcu(head, rcu_free_slab); 1451 } else 1452 __free_slab(s, page); 1453 } 1454 1455 static void discard_slab(struct kmem_cache *s, struct page *page) 1456 { 1457 dec_slabs_node(s, page_to_nid(page), page->objects); 1458 free_slab(s, page); 1459 } 1460 1461 /* 1462 * Management of partially allocated slabs. 1463 * 1464 * list_lock must be held. 1465 */ 1466 static inline void add_partial(struct kmem_cache_node *n, 1467 struct page *page, int tail) 1468 { 1469 n->nr_partial++; 1470 if (tail == DEACTIVATE_TO_TAIL) 1471 list_add_tail(&page->lru, &n->partial); 1472 else 1473 list_add(&page->lru, &n->partial); 1474 } 1475 1476 /* 1477 * list_lock must be held. 1478 */ 1479 static inline void remove_partial(struct kmem_cache_node *n, 1480 struct page *page) 1481 { 1482 list_del(&page->lru); 1483 n->nr_partial--; 1484 } 1485 1486 /* 1487 * Lock slab, remove from the partial list and put the object into the 1488 * per cpu freelist. 1489 * 1490 * Returns a list of objects or NULL if it fails. 1491 * 1492 * Must hold list_lock. 1493 */ 1494 static inline void *acquire_slab(struct kmem_cache *s, 1495 struct kmem_cache_node *n, struct page *page, 1496 int mode) 1497 { 1498 void *freelist; 1499 unsigned long counters; 1500 struct page new; 1501 1502 /* 1503 * Zap the freelist and set the frozen bit. 1504 * The old freelist is the list of objects for the 1505 * per cpu allocation list. 1506 */ 1507 do { 1508 freelist = page->freelist; 1509 counters = page->counters; 1510 new.counters = counters; 1511 if (mode) 1512 new.inuse = page->objects; 1513 1514 VM_BUG_ON(new.frozen); 1515 new.frozen = 1; 1516 1517 } while (!__cmpxchg_double_slab(s, page, 1518 freelist, counters, 1519 NULL, new.counters, 1520 "lock and freeze")); 1521 1522 remove_partial(n, page); 1523 return freelist; 1524 } 1525 1526 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); 1527 1528 /* 1529 * Try to allocate a partial slab from a specific node. 1530 */ 1531 static void *get_partial_node(struct kmem_cache *s, 1532 struct kmem_cache_node *n, struct kmem_cache_cpu *c) 1533 { 1534 struct page *page, *page2; 1535 void *object = NULL; 1536 1537 /* 1538 * Racy check. If we mistakenly see no partial slabs then we 1539 * just allocate an empty slab. If we mistakenly try to get a 1540 * partial slab and there is none available then get_partials() 1541 * will return NULL. 1542 */ 1543 if (!n || !n->nr_partial) 1544 return NULL; 1545 1546 spin_lock(&n->list_lock); 1547 list_for_each_entry_safe(page, page2, &n->partial, lru) { 1548 void *t = acquire_slab(s, n, page, object == NULL); 1549 int available; 1550 1551 if (!t) 1552 break; 1553 1554 if (!object) { 1555 c->page = page; 1556 c->node = page_to_nid(page); 1557 stat(s, ALLOC_FROM_PARTIAL); 1558 object = t; 1559 available = page->objects - page->inuse; 1560 } else { 1561 page->freelist = t; 1562 available = put_cpu_partial(s, page, 0); 1563 } 1564 if (kmem_cache_debug(s) || available > s->cpu_partial / 2) 1565 break; 1566 1567 } 1568 spin_unlock(&n->list_lock); 1569 return object; 1570 } 1571 1572 /* 1573 * Get a page from somewhere. Search in increasing NUMA distances. 1574 */ 1575 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags, 1576 struct kmem_cache_cpu *c) 1577 { 1578 #ifdef CONFIG_NUMA 1579 struct zonelist *zonelist; 1580 struct zoneref *z; 1581 struct zone *zone; 1582 enum zone_type high_zoneidx = gfp_zone(flags); 1583 void *object; 1584 unsigned int cpuset_mems_cookie; 1585 1586 /* 1587 * The defrag ratio allows a configuration of the tradeoffs between 1588 * inter node defragmentation and node local allocations. A lower 1589 * defrag_ratio increases the tendency to do local allocations 1590 * instead of attempting to obtain partial slabs from other nodes. 1591 * 1592 * If the defrag_ratio is set to 0 then kmalloc() always 1593 * returns node local objects. If the ratio is higher then kmalloc() 1594 * may return off node objects because partial slabs are obtained 1595 * from other nodes and filled up. 1596 * 1597 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1598 * defrag_ratio = 1000) then every (well almost) allocation will 1599 * first attempt to defrag slab caches on other nodes. This means 1600 * scanning over all nodes to look for partial slabs which may be 1601 * expensive if we do it every time we are trying to find a slab 1602 * with available objects. 1603 */ 1604 if (!s->remote_node_defrag_ratio || 1605 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1606 return NULL; 1607 1608 do { 1609 cpuset_mems_cookie = get_mems_allowed(); 1610 zonelist = node_zonelist(slab_node(current->mempolicy), flags); 1611 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1612 struct kmem_cache_node *n; 1613 1614 n = get_node(s, zone_to_nid(zone)); 1615 1616 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1617 n->nr_partial > s->min_partial) { 1618 object = get_partial_node(s, n, c); 1619 if (object) { 1620 /* 1621 * Return the object even if 1622 * put_mems_allowed indicated that 1623 * the cpuset mems_allowed was 1624 * updated in parallel. It's a 1625 * harmless race between the alloc 1626 * and the cpuset update. 1627 */ 1628 put_mems_allowed(cpuset_mems_cookie); 1629 return object; 1630 } 1631 } 1632 } 1633 } while (!put_mems_allowed(cpuset_mems_cookie)); 1634 #endif 1635 return NULL; 1636 } 1637 1638 /* 1639 * Get a partial page, lock it and return it. 1640 */ 1641 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, 1642 struct kmem_cache_cpu *c) 1643 { 1644 void *object; 1645 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; 1646 1647 object = get_partial_node(s, get_node(s, searchnode), c); 1648 if (object || node != NUMA_NO_NODE) 1649 return object; 1650 1651 return get_any_partial(s, flags, c); 1652 } 1653 1654 #ifdef CONFIG_PREEMPT 1655 /* 1656 * Calculate the next globally unique transaction for disambiguiation 1657 * during cmpxchg. The transactions start with the cpu number and are then 1658 * incremented by CONFIG_NR_CPUS. 1659 */ 1660 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 1661 #else 1662 /* 1663 * No preemption supported therefore also no need to check for 1664 * different cpus. 1665 */ 1666 #define TID_STEP 1 1667 #endif 1668 1669 static inline unsigned long next_tid(unsigned long tid) 1670 { 1671 return tid + TID_STEP; 1672 } 1673 1674 static inline unsigned int tid_to_cpu(unsigned long tid) 1675 { 1676 return tid % TID_STEP; 1677 } 1678 1679 static inline unsigned long tid_to_event(unsigned long tid) 1680 { 1681 return tid / TID_STEP; 1682 } 1683 1684 static inline unsigned int init_tid(int cpu) 1685 { 1686 return cpu; 1687 } 1688 1689 static inline void note_cmpxchg_failure(const char *n, 1690 const struct kmem_cache *s, unsigned long tid) 1691 { 1692 #ifdef SLUB_DEBUG_CMPXCHG 1693 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 1694 1695 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); 1696 1697 #ifdef CONFIG_PREEMPT 1698 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 1699 printk("due to cpu change %d -> %d\n", 1700 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 1701 else 1702 #endif 1703 if (tid_to_event(tid) != tid_to_event(actual_tid)) 1704 printk("due to cpu running other code. Event %ld->%ld\n", 1705 tid_to_event(tid), tid_to_event(actual_tid)); 1706 else 1707 printk("for unknown reason: actual=%lx was=%lx target=%lx\n", 1708 actual_tid, tid, next_tid(tid)); 1709 #endif 1710 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 1711 } 1712 1713 void init_kmem_cache_cpus(struct kmem_cache *s) 1714 { 1715 int cpu; 1716 1717 for_each_possible_cpu(cpu) 1718 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 1719 } 1720 1721 /* 1722 * Remove the cpu slab 1723 */ 1724 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 1725 { 1726 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; 1727 struct page *page = c->page; 1728 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1729 int lock = 0; 1730 enum slab_modes l = M_NONE, m = M_NONE; 1731 void *freelist; 1732 void *nextfree; 1733 int tail = DEACTIVATE_TO_HEAD; 1734 struct page new; 1735 struct page old; 1736 1737 if (page->freelist) { 1738 stat(s, DEACTIVATE_REMOTE_FREES); 1739 tail = DEACTIVATE_TO_TAIL; 1740 } 1741 1742 c->tid = next_tid(c->tid); 1743 c->page = NULL; 1744 freelist = c->freelist; 1745 c->freelist = NULL; 1746 1747 /* 1748 * Stage one: Free all available per cpu objects back 1749 * to the page freelist while it is still frozen. Leave the 1750 * last one. 1751 * 1752 * There is no need to take the list->lock because the page 1753 * is still frozen. 1754 */ 1755 while (freelist && (nextfree = get_freepointer(s, freelist))) { 1756 void *prior; 1757 unsigned long counters; 1758 1759 do { 1760 prior = page->freelist; 1761 counters = page->counters; 1762 set_freepointer(s, freelist, prior); 1763 new.counters = counters; 1764 new.inuse--; 1765 VM_BUG_ON(!new.frozen); 1766 1767 } while (!__cmpxchg_double_slab(s, page, 1768 prior, counters, 1769 freelist, new.counters, 1770 "drain percpu freelist")); 1771 1772 freelist = nextfree; 1773 } 1774 1775 /* 1776 * Stage two: Ensure that the page is unfrozen while the 1777 * list presence reflects the actual number of objects 1778 * during unfreeze. 1779 * 1780 * We setup the list membership and then perform a cmpxchg 1781 * with the count. If there is a mismatch then the page 1782 * is not unfrozen but the page is on the wrong list. 1783 * 1784 * Then we restart the process which may have to remove 1785 * the page from the list that we just put it on again 1786 * because the number of objects in the slab may have 1787 * changed. 1788 */ 1789 redo: 1790 1791 old.freelist = page->freelist; 1792 old.counters = page->counters; 1793 VM_BUG_ON(!old.frozen); 1794 1795 /* Determine target state of the slab */ 1796 new.counters = old.counters; 1797 if (freelist) { 1798 new.inuse--; 1799 set_freepointer(s, freelist, old.freelist); 1800 new.freelist = freelist; 1801 } else 1802 new.freelist = old.freelist; 1803 1804 new.frozen = 0; 1805 1806 if (!new.inuse && n->nr_partial > s->min_partial) 1807 m = M_FREE; 1808 else if (new.freelist) { 1809 m = M_PARTIAL; 1810 if (!lock) { 1811 lock = 1; 1812 /* 1813 * Taking the spinlock removes the possiblity 1814 * that acquire_slab() will see a slab page that 1815 * is frozen 1816 */ 1817 spin_lock(&n->list_lock); 1818 } 1819 } else { 1820 m = M_FULL; 1821 if (kmem_cache_debug(s) && !lock) { 1822 lock = 1; 1823 /* 1824 * This also ensures that the scanning of full 1825 * slabs from diagnostic functions will not see 1826 * any frozen slabs. 1827 */ 1828 spin_lock(&n->list_lock); 1829 } 1830 } 1831 1832 if (l != m) { 1833 1834 if (l == M_PARTIAL) 1835 1836 remove_partial(n, page); 1837 1838 else if (l == M_FULL) 1839 1840 remove_full(s, page); 1841 1842 if (m == M_PARTIAL) { 1843 1844 add_partial(n, page, tail); 1845 stat(s, tail); 1846 1847 } else if (m == M_FULL) { 1848 1849 stat(s, DEACTIVATE_FULL); 1850 add_full(s, n, page); 1851 1852 } 1853 } 1854 1855 l = m; 1856 if (!__cmpxchg_double_slab(s, page, 1857 old.freelist, old.counters, 1858 new.freelist, new.counters, 1859 "unfreezing slab")) 1860 goto redo; 1861 1862 if (lock) 1863 spin_unlock(&n->list_lock); 1864 1865 if (m == M_FREE) { 1866 stat(s, DEACTIVATE_EMPTY); 1867 discard_slab(s, page); 1868 stat(s, FREE_SLAB); 1869 } 1870 } 1871 1872 /* Unfreeze all the cpu partial slabs */ 1873 static void unfreeze_partials(struct kmem_cache *s) 1874 { 1875 struct kmem_cache_node *n = NULL; 1876 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab); 1877 struct page *page, *discard_page = NULL; 1878 1879 while ((page = c->partial)) { 1880 enum slab_modes { M_PARTIAL, M_FREE }; 1881 enum slab_modes l, m; 1882 struct page new; 1883 struct page old; 1884 1885 c->partial = page->next; 1886 l = M_FREE; 1887 1888 do { 1889 1890 old.freelist = page->freelist; 1891 old.counters = page->counters; 1892 VM_BUG_ON(!old.frozen); 1893 1894 new.counters = old.counters; 1895 new.freelist = old.freelist; 1896 1897 new.frozen = 0; 1898 1899 if (!new.inuse && (!n || n->nr_partial > s->min_partial)) 1900 m = M_FREE; 1901 else { 1902 struct kmem_cache_node *n2 = get_node(s, 1903 page_to_nid(page)); 1904 1905 m = M_PARTIAL; 1906 if (n != n2) { 1907 if (n) 1908 spin_unlock(&n->list_lock); 1909 1910 n = n2; 1911 spin_lock(&n->list_lock); 1912 } 1913 } 1914 1915 if (l != m) { 1916 if (l == M_PARTIAL) { 1917 remove_partial(n, page); 1918 stat(s, FREE_REMOVE_PARTIAL); 1919 } else { 1920 add_partial(n, page, 1921 DEACTIVATE_TO_TAIL); 1922 stat(s, FREE_ADD_PARTIAL); 1923 } 1924 1925 l = m; 1926 } 1927 1928 } while (!cmpxchg_double_slab(s, page, 1929 old.freelist, old.counters, 1930 new.freelist, new.counters, 1931 "unfreezing slab")); 1932 1933 if (m == M_FREE) { 1934 page->next = discard_page; 1935 discard_page = page; 1936 } 1937 } 1938 1939 if (n) 1940 spin_unlock(&n->list_lock); 1941 1942 while (discard_page) { 1943 page = discard_page; 1944 discard_page = discard_page->next; 1945 1946 stat(s, DEACTIVATE_EMPTY); 1947 discard_slab(s, page); 1948 stat(s, FREE_SLAB); 1949 } 1950 } 1951 1952 /* 1953 * Put a page that was just frozen (in __slab_free) into a partial page 1954 * slot if available. This is done without interrupts disabled and without 1955 * preemption disabled. The cmpxchg is racy and may put the partial page 1956 * onto a random cpus partial slot. 1957 * 1958 * If we did not find a slot then simply move all the partials to the 1959 * per node partial list. 1960 */ 1961 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) 1962 { 1963 struct page *oldpage; 1964 int pages; 1965 int pobjects; 1966 1967 do { 1968 pages = 0; 1969 pobjects = 0; 1970 oldpage = this_cpu_read(s->cpu_slab->partial); 1971 1972 if (oldpage) { 1973 pobjects = oldpage->pobjects; 1974 pages = oldpage->pages; 1975 if (drain && pobjects > s->cpu_partial) { 1976 unsigned long flags; 1977 /* 1978 * partial array is full. Move the existing 1979 * set to the per node partial list. 1980 */ 1981 local_irq_save(flags); 1982 unfreeze_partials(s); 1983 local_irq_restore(flags); 1984 pobjects = 0; 1985 pages = 0; 1986 } 1987 } 1988 1989 pages++; 1990 pobjects += page->objects - page->inuse; 1991 1992 page->pages = pages; 1993 page->pobjects = pobjects; 1994 page->next = oldpage; 1995 1996 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage); 1997 stat(s, CPU_PARTIAL_FREE); 1998 return pobjects; 1999 } 2000 2001 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 2002 { 2003 stat(s, CPUSLAB_FLUSH); 2004 deactivate_slab(s, c); 2005 } 2006 2007 /* 2008 * Flush cpu slab. 2009 * 2010 * Called from IPI handler with interrupts disabled. 2011 */ 2012 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 2013 { 2014 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2015 2016 if (likely(c)) { 2017 if (c->page) 2018 flush_slab(s, c); 2019 2020 unfreeze_partials(s); 2021 } 2022 } 2023 2024 static void flush_cpu_slab(void *d) 2025 { 2026 struct kmem_cache *s = d; 2027 2028 __flush_cpu_slab(s, smp_processor_id()); 2029 } 2030 2031 static void flush_all(struct kmem_cache *s) 2032 { 2033 on_each_cpu(flush_cpu_slab, s, 1); 2034 } 2035 2036 /* 2037 * Check if the objects in a per cpu structure fit numa 2038 * locality expectations. 2039 */ 2040 static inline int node_match(struct kmem_cache_cpu *c, int node) 2041 { 2042 #ifdef CONFIG_NUMA 2043 if (node != NUMA_NO_NODE && c->node != node) 2044 return 0; 2045 #endif 2046 return 1; 2047 } 2048 2049 static int count_free(struct page *page) 2050 { 2051 return page->objects - page->inuse; 2052 } 2053 2054 static unsigned long count_partial(struct kmem_cache_node *n, 2055 int (*get_count)(struct page *)) 2056 { 2057 unsigned long flags; 2058 unsigned long x = 0; 2059 struct page *page; 2060 2061 spin_lock_irqsave(&n->list_lock, flags); 2062 list_for_each_entry(page, &n->partial, lru) 2063 x += get_count(page); 2064 spin_unlock_irqrestore(&n->list_lock, flags); 2065 return x; 2066 } 2067 2068 static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 2069 { 2070 #ifdef CONFIG_SLUB_DEBUG 2071 return atomic_long_read(&n->total_objects); 2072 #else 2073 return 0; 2074 #endif 2075 } 2076 2077 static noinline void 2078 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 2079 { 2080 int node; 2081 2082 printk(KERN_WARNING 2083 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 2084 nid, gfpflags); 2085 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 2086 "default order: %d, min order: %d\n", s->name, s->objsize, 2087 s->size, oo_order(s->oo), oo_order(s->min)); 2088 2089 if (oo_order(s->min) > get_order(s->objsize)) 2090 printk(KERN_WARNING " %s debugging increased min order, use " 2091 "slub_debug=O to disable.\n", s->name); 2092 2093 for_each_online_node(node) { 2094 struct kmem_cache_node *n = get_node(s, node); 2095 unsigned long nr_slabs; 2096 unsigned long nr_objs; 2097 unsigned long nr_free; 2098 2099 if (!n) 2100 continue; 2101 2102 nr_free = count_partial(n, count_free); 2103 nr_slabs = node_nr_slabs(n); 2104 nr_objs = node_nr_objs(n); 2105 2106 printk(KERN_WARNING 2107 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 2108 node, nr_slabs, nr_objs, nr_free); 2109 } 2110 } 2111 2112 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, 2113 int node, struct kmem_cache_cpu **pc) 2114 { 2115 void *object; 2116 struct kmem_cache_cpu *c; 2117 struct page *page = new_slab(s, flags, node); 2118 2119 if (page) { 2120 c = __this_cpu_ptr(s->cpu_slab); 2121 if (c->page) 2122 flush_slab(s, c); 2123 2124 /* 2125 * No other reference to the page yet so we can 2126 * muck around with it freely without cmpxchg 2127 */ 2128 object = page->freelist; 2129 page->freelist = NULL; 2130 2131 stat(s, ALLOC_SLAB); 2132 c->node = page_to_nid(page); 2133 c->page = page; 2134 *pc = c; 2135 } else 2136 object = NULL; 2137 2138 return object; 2139 } 2140 2141 /* 2142 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist 2143 * or deactivate the page. 2144 * 2145 * The page is still frozen if the return value is not NULL. 2146 * 2147 * If this function returns NULL then the page has been unfrozen. 2148 */ 2149 static inline void *get_freelist(struct kmem_cache *s, struct page *page) 2150 { 2151 struct page new; 2152 unsigned long counters; 2153 void *freelist; 2154 2155 do { 2156 freelist = page->freelist; 2157 counters = page->counters; 2158 new.counters = counters; 2159 VM_BUG_ON(!new.frozen); 2160 2161 new.inuse = page->objects; 2162 new.frozen = freelist != NULL; 2163 2164 } while (!cmpxchg_double_slab(s, page, 2165 freelist, counters, 2166 NULL, new.counters, 2167 "get_freelist")); 2168 2169 return freelist; 2170 } 2171 2172 /* 2173 * Slow path. The lockless freelist is empty or we need to perform 2174 * debugging duties. 2175 * 2176 * Processing is still very fast if new objects have been freed to the 2177 * regular freelist. In that case we simply take over the regular freelist 2178 * as the lockless freelist and zap the regular freelist. 2179 * 2180 * If that is not working then we fall back to the partial lists. We take the 2181 * first element of the freelist as the object to allocate now and move the 2182 * rest of the freelist to the lockless freelist. 2183 * 2184 * And if we were unable to get a new slab from the partial slab lists then 2185 * we need to allocate a new slab. This is the slowest path since it involves 2186 * a call to the page allocator and the setup of a new slab. 2187 */ 2188 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2189 unsigned long addr, struct kmem_cache_cpu *c) 2190 { 2191 void **object; 2192 unsigned long flags; 2193 2194 local_irq_save(flags); 2195 #ifdef CONFIG_PREEMPT 2196 /* 2197 * We may have been preempted and rescheduled on a different 2198 * cpu before disabling interrupts. Need to reload cpu area 2199 * pointer. 2200 */ 2201 c = this_cpu_ptr(s->cpu_slab); 2202 #endif 2203 2204 if (!c->page) 2205 goto new_slab; 2206 redo: 2207 if (unlikely(!node_match(c, node))) { 2208 stat(s, ALLOC_NODE_MISMATCH); 2209 deactivate_slab(s, c); 2210 goto new_slab; 2211 } 2212 2213 /* must check again c->freelist in case of cpu migration or IRQ */ 2214 object = c->freelist; 2215 if (object) 2216 goto load_freelist; 2217 2218 stat(s, ALLOC_SLOWPATH); 2219 2220 object = get_freelist(s, c->page); 2221 2222 if (!object) { 2223 c->page = NULL; 2224 stat(s, DEACTIVATE_BYPASS); 2225 goto new_slab; 2226 } 2227 2228 stat(s, ALLOC_REFILL); 2229 2230 load_freelist: 2231 c->freelist = get_freepointer(s, object); 2232 c->tid = next_tid(c->tid); 2233 local_irq_restore(flags); 2234 return object; 2235 2236 new_slab: 2237 2238 if (c->partial) { 2239 c->page = c->partial; 2240 c->partial = c->page->next; 2241 c->node = page_to_nid(c->page); 2242 stat(s, CPU_PARTIAL_ALLOC); 2243 c->freelist = NULL; 2244 goto redo; 2245 } 2246 2247 /* Then do expensive stuff like retrieving pages from the partial lists */ 2248 object = get_partial(s, gfpflags, node, c); 2249 2250 if (unlikely(!object)) { 2251 2252 object = new_slab_objects(s, gfpflags, node, &c); 2253 2254 if (unlikely(!object)) { 2255 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 2256 slab_out_of_memory(s, gfpflags, node); 2257 2258 local_irq_restore(flags); 2259 return NULL; 2260 } 2261 } 2262 2263 if (likely(!kmem_cache_debug(s))) 2264 goto load_freelist; 2265 2266 /* Only entered in the debug case */ 2267 if (!alloc_debug_processing(s, c->page, object, addr)) 2268 goto new_slab; /* Slab failed checks. Next slab needed */ 2269 2270 c->freelist = get_freepointer(s, object); 2271 deactivate_slab(s, c); 2272 c->node = NUMA_NO_NODE; 2273 local_irq_restore(flags); 2274 return object; 2275 } 2276 2277 /* 2278 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 2279 * have the fastpath folded into their functions. So no function call 2280 * overhead for requests that can be satisfied on the fastpath. 2281 * 2282 * The fastpath works by first checking if the lockless freelist can be used. 2283 * If not then __slab_alloc is called for slow processing. 2284 * 2285 * Otherwise we can simply pick the next object from the lockless free list. 2286 */ 2287 static __always_inline void *slab_alloc(struct kmem_cache *s, 2288 gfp_t gfpflags, int node, unsigned long addr) 2289 { 2290 void **object; 2291 struct kmem_cache_cpu *c; 2292 unsigned long tid; 2293 2294 if (slab_pre_alloc_hook(s, gfpflags)) 2295 return NULL; 2296 2297 redo: 2298 2299 /* 2300 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 2301 * enabled. We may switch back and forth between cpus while 2302 * reading from one cpu area. That does not matter as long 2303 * as we end up on the original cpu again when doing the cmpxchg. 2304 */ 2305 c = __this_cpu_ptr(s->cpu_slab); 2306 2307 /* 2308 * The transaction ids are globally unique per cpu and per operation on 2309 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 2310 * occurs on the right processor and that there was no operation on the 2311 * linked list in between. 2312 */ 2313 tid = c->tid; 2314 barrier(); 2315 2316 object = c->freelist; 2317 if (unlikely(!object || !node_match(c, node))) 2318 2319 object = __slab_alloc(s, gfpflags, node, addr, c); 2320 2321 else { 2322 /* 2323 * The cmpxchg will only match if there was no additional 2324 * operation and if we are on the right processor. 2325 * 2326 * The cmpxchg does the following atomically (without lock semantics!) 2327 * 1. Relocate first pointer to the current per cpu area. 2328 * 2. Verify that tid and freelist have not been changed 2329 * 3. If they were not changed replace tid and freelist 2330 * 2331 * Since this is without lock semantics the protection is only against 2332 * code executing on this cpu *not* from access by other cpus. 2333 */ 2334 if (unlikely(!this_cpu_cmpxchg_double( 2335 s->cpu_slab->freelist, s->cpu_slab->tid, 2336 object, tid, 2337 get_freepointer_safe(s, object), next_tid(tid)))) { 2338 2339 note_cmpxchg_failure("slab_alloc", s, tid); 2340 goto redo; 2341 } 2342 stat(s, ALLOC_FASTPATH); 2343 } 2344 2345 if (unlikely(gfpflags & __GFP_ZERO) && object) 2346 memset(object, 0, s->objsize); 2347 2348 slab_post_alloc_hook(s, gfpflags, object); 2349 2350 return object; 2351 } 2352 2353 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 2354 { 2355 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 2356 2357 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags); 2358 2359 return ret; 2360 } 2361 EXPORT_SYMBOL(kmem_cache_alloc); 2362 2363 #ifdef CONFIG_TRACING 2364 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2365 { 2366 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_); 2367 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2368 return ret; 2369 } 2370 EXPORT_SYMBOL(kmem_cache_alloc_trace); 2371 2372 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order) 2373 { 2374 void *ret = kmalloc_order(size, flags, order); 2375 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags); 2376 return ret; 2377 } 2378 EXPORT_SYMBOL(kmalloc_order_trace); 2379 #endif 2380 2381 #ifdef CONFIG_NUMA 2382 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2383 { 2384 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 2385 2386 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2387 s->objsize, s->size, gfpflags, node); 2388 2389 return ret; 2390 } 2391 EXPORT_SYMBOL(kmem_cache_alloc_node); 2392 2393 #ifdef CONFIG_TRACING 2394 void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2395 gfp_t gfpflags, 2396 int node, size_t size) 2397 { 2398 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_); 2399 2400 trace_kmalloc_node(_RET_IP_, ret, 2401 size, s->size, gfpflags, node); 2402 return ret; 2403 } 2404 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2405 #endif 2406 #endif 2407 2408 /* 2409 * Slow patch handling. This may still be called frequently since objects 2410 * have a longer lifetime than the cpu slabs in most processing loads. 2411 * 2412 * So we still attempt to reduce cache line usage. Just take the slab 2413 * lock and free the item. If there is no additional partial page 2414 * handling required then we can return immediately. 2415 */ 2416 static void __slab_free(struct kmem_cache *s, struct page *page, 2417 void *x, unsigned long addr) 2418 { 2419 void *prior; 2420 void **object = (void *)x; 2421 int was_frozen; 2422 int inuse; 2423 struct page new; 2424 unsigned long counters; 2425 struct kmem_cache_node *n = NULL; 2426 unsigned long uninitialized_var(flags); 2427 2428 stat(s, FREE_SLOWPATH); 2429 2430 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr)) 2431 return; 2432 2433 do { 2434 prior = page->freelist; 2435 counters = page->counters; 2436 set_freepointer(s, object, prior); 2437 new.counters = counters; 2438 was_frozen = new.frozen; 2439 new.inuse--; 2440 if ((!new.inuse || !prior) && !was_frozen && !n) { 2441 2442 if (!kmem_cache_debug(s) && !prior) 2443 2444 /* 2445 * Slab was on no list before and will be partially empty 2446 * We can defer the list move and instead freeze it. 2447 */ 2448 new.frozen = 1; 2449 2450 else { /* Needs to be taken off a list */ 2451 2452 n = get_node(s, page_to_nid(page)); 2453 /* 2454 * Speculatively acquire the list_lock. 2455 * If the cmpxchg does not succeed then we may 2456 * drop the list_lock without any processing. 2457 * 2458 * Otherwise the list_lock will synchronize with 2459 * other processors updating the list of slabs. 2460 */ 2461 spin_lock_irqsave(&n->list_lock, flags); 2462 2463 } 2464 } 2465 inuse = new.inuse; 2466 2467 } while (!cmpxchg_double_slab(s, page, 2468 prior, counters, 2469 object, new.counters, 2470 "__slab_free")); 2471 2472 if (likely(!n)) { 2473 2474 /* 2475 * If we just froze the page then put it onto the 2476 * per cpu partial list. 2477 */ 2478 if (new.frozen && !was_frozen) 2479 put_cpu_partial(s, page, 1); 2480 2481 /* 2482 * The list lock was not taken therefore no list 2483 * activity can be necessary. 2484 */ 2485 if (was_frozen) 2486 stat(s, FREE_FROZEN); 2487 return; 2488 } 2489 2490 /* 2491 * was_frozen may have been set after we acquired the list_lock in 2492 * an earlier loop. So we need to check it here again. 2493 */ 2494 if (was_frozen) 2495 stat(s, FREE_FROZEN); 2496 else { 2497 if (unlikely(!inuse && n->nr_partial > s->min_partial)) 2498 goto slab_empty; 2499 2500 /* 2501 * Objects left in the slab. If it was not on the partial list before 2502 * then add it. 2503 */ 2504 if (unlikely(!prior)) { 2505 remove_full(s, page); 2506 add_partial(n, page, DEACTIVATE_TO_TAIL); 2507 stat(s, FREE_ADD_PARTIAL); 2508 } 2509 } 2510 spin_unlock_irqrestore(&n->list_lock, flags); 2511 return; 2512 2513 slab_empty: 2514 if (prior) { 2515 /* 2516 * Slab on the partial list. 2517 */ 2518 remove_partial(n, page); 2519 stat(s, FREE_REMOVE_PARTIAL); 2520 } else 2521 /* Slab must be on the full list */ 2522 remove_full(s, page); 2523 2524 spin_unlock_irqrestore(&n->list_lock, flags); 2525 stat(s, FREE_SLAB); 2526 discard_slab(s, page); 2527 } 2528 2529 /* 2530 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 2531 * can perform fastpath freeing without additional function calls. 2532 * 2533 * The fastpath is only possible if we are freeing to the current cpu slab 2534 * of this processor. This typically the case if we have just allocated 2535 * the item before. 2536 * 2537 * If fastpath is not possible then fall back to __slab_free where we deal 2538 * with all sorts of special processing. 2539 */ 2540 static __always_inline void slab_free(struct kmem_cache *s, 2541 struct page *page, void *x, unsigned long addr) 2542 { 2543 void **object = (void *)x; 2544 struct kmem_cache_cpu *c; 2545 unsigned long tid; 2546 2547 slab_free_hook(s, x); 2548 2549 redo: 2550 /* 2551 * Determine the currently cpus per cpu slab. 2552 * The cpu may change afterward. However that does not matter since 2553 * data is retrieved via this pointer. If we are on the same cpu 2554 * during the cmpxchg then the free will succedd. 2555 */ 2556 c = __this_cpu_ptr(s->cpu_slab); 2557 2558 tid = c->tid; 2559 barrier(); 2560 2561 if (likely(page == c->page)) { 2562 set_freepointer(s, object, c->freelist); 2563 2564 if (unlikely(!this_cpu_cmpxchg_double( 2565 s->cpu_slab->freelist, s->cpu_slab->tid, 2566 c->freelist, tid, 2567 object, next_tid(tid)))) { 2568 2569 note_cmpxchg_failure("slab_free", s, tid); 2570 goto redo; 2571 } 2572 stat(s, FREE_FASTPATH); 2573 } else 2574 __slab_free(s, page, x, addr); 2575 2576 } 2577 2578 void kmem_cache_free(struct kmem_cache *s, void *x) 2579 { 2580 struct page *page; 2581 2582 page = virt_to_head_page(x); 2583 2584 slab_free(s, page, x, _RET_IP_); 2585 2586 trace_kmem_cache_free(_RET_IP_, x); 2587 } 2588 EXPORT_SYMBOL(kmem_cache_free); 2589 2590 /* 2591 * Object placement in a slab is made very easy because we always start at 2592 * offset 0. If we tune the size of the object to the alignment then we can 2593 * get the required alignment by putting one properly sized object after 2594 * another. 2595 * 2596 * Notice that the allocation order determines the sizes of the per cpu 2597 * caches. Each processor has always one slab available for allocations. 2598 * Increasing the allocation order reduces the number of times that slabs 2599 * must be moved on and off the partial lists and is therefore a factor in 2600 * locking overhead. 2601 */ 2602 2603 /* 2604 * Mininum / Maximum order of slab pages. This influences locking overhead 2605 * and slab fragmentation. A higher order reduces the number of partial slabs 2606 * and increases the number of allocations possible without having to 2607 * take the list_lock. 2608 */ 2609 static int slub_min_order; 2610 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 2611 static int slub_min_objects; 2612 2613 /* 2614 * Merge control. If this is set then no merging of slab caches will occur. 2615 * (Could be removed. This was introduced to pacify the merge skeptics.) 2616 */ 2617 static int slub_nomerge; 2618 2619 /* 2620 * Calculate the order of allocation given an slab object size. 2621 * 2622 * The order of allocation has significant impact on performance and other 2623 * system components. Generally order 0 allocations should be preferred since 2624 * order 0 does not cause fragmentation in the page allocator. Larger objects 2625 * be problematic to put into order 0 slabs because there may be too much 2626 * unused space left. We go to a higher order if more than 1/16th of the slab 2627 * would be wasted. 2628 * 2629 * In order to reach satisfactory performance we must ensure that a minimum 2630 * number of objects is in one slab. Otherwise we may generate too much 2631 * activity on the partial lists which requires taking the list_lock. This is 2632 * less a concern for large slabs though which are rarely used. 2633 * 2634 * slub_max_order specifies the order where we begin to stop considering the 2635 * number of objects in a slab as critical. If we reach slub_max_order then 2636 * we try to keep the page order as low as possible. So we accept more waste 2637 * of space in favor of a small page order. 2638 * 2639 * Higher order allocations also allow the placement of more objects in a 2640 * slab and thereby reduce object handling overhead. If the user has 2641 * requested a higher mininum order then we start with that one instead of 2642 * the smallest order which will fit the object. 2643 */ 2644 static inline int slab_order(int size, int min_objects, 2645 int max_order, int fract_leftover, int reserved) 2646 { 2647 int order; 2648 int rem; 2649 int min_order = slub_min_order; 2650 2651 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) 2652 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 2653 2654 for (order = max(min_order, 2655 fls(min_objects * size - 1) - PAGE_SHIFT); 2656 order <= max_order; order++) { 2657 2658 unsigned long slab_size = PAGE_SIZE << order; 2659 2660 if (slab_size < min_objects * size + reserved) 2661 continue; 2662 2663 rem = (slab_size - reserved) % size; 2664 2665 if (rem <= slab_size / fract_leftover) 2666 break; 2667 2668 } 2669 2670 return order; 2671 } 2672 2673 static inline int calculate_order(int size, int reserved) 2674 { 2675 int order; 2676 int min_objects; 2677 int fraction; 2678 int max_objects; 2679 2680 /* 2681 * Attempt to find best configuration for a slab. This 2682 * works by first attempting to generate a layout with 2683 * the best configuration and backing off gradually. 2684 * 2685 * First we reduce the acceptable waste in a slab. Then 2686 * we reduce the minimum objects required in a slab. 2687 */ 2688 min_objects = slub_min_objects; 2689 if (!min_objects) 2690 min_objects = 4 * (fls(nr_cpu_ids) + 1); 2691 max_objects = order_objects(slub_max_order, size, reserved); 2692 min_objects = min(min_objects, max_objects); 2693 2694 while (min_objects > 1) { 2695 fraction = 16; 2696 while (fraction >= 4) { 2697 order = slab_order(size, min_objects, 2698 slub_max_order, fraction, reserved); 2699 if (order <= slub_max_order) 2700 return order; 2701 fraction /= 2; 2702 } 2703 min_objects--; 2704 } 2705 2706 /* 2707 * We were unable to place multiple objects in a slab. Now 2708 * lets see if we can place a single object there. 2709 */ 2710 order = slab_order(size, 1, slub_max_order, 1, reserved); 2711 if (order <= slub_max_order) 2712 return order; 2713 2714 /* 2715 * Doh this slab cannot be placed using slub_max_order. 2716 */ 2717 order = slab_order(size, 1, MAX_ORDER, 1, reserved); 2718 if (order < MAX_ORDER) 2719 return order; 2720 return -ENOSYS; 2721 } 2722 2723 /* 2724 * Figure out what the alignment of the objects will be. 2725 */ 2726 static unsigned long calculate_alignment(unsigned long flags, 2727 unsigned long align, unsigned long size) 2728 { 2729 /* 2730 * If the user wants hardware cache aligned objects then follow that 2731 * suggestion if the object is sufficiently large. 2732 * 2733 * The hardware cache alignment cannot override the specified 2734 * alignment though. If that is greater then use it. 2735 */ 2736 if (flags & SLAB_HWCACHE_ALIGN) { 2737 unsigned long ralign = cache_line_size(); 2738 while (size <= ralign / 2) 2739 ralign /= 2; 2740 align = max(align, ralign); 2741 } 2742 2743 if (align < ARCH_SLAB_MINALIGN) 2744 align = ARCH_SLAB_MINALIGN; 2745 2746 return ALIGN(align, sizeof(void *)); 2747 } 2748 2749 static void 2750 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s) 2751 { 2752 n->nr_partial = 0; 2753 spin_lock_init(&n->list_lock); 2754 INIT_LIST_HEAD(&n->partial); 2755 #ifdef CONFIG_SLUB_DEBUG 2756 atomic_long_set(&n->nr_slabs, 0); 2757 atomic_long_set(&n->total_objects, 0); 2758 INIT_LIST_HEAD(&n->full); 2759 #endif 2760 } 2761 2762 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 2763 { 2764 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 2765 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu)); 2766 2767 /* 2768 * Must align to double word boundary for the double cmpxchg 2769 * instructions to work; see __pcpu_double_call_return_bool(). 2770 */ 2771 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2772 2 * sizeof(void *)); 2773 2774 if (!s->cpu_slab) 2775 return 0; 2776 2777 init_kmem_cache_cpus(s); 2778 2779 return 1; 2780 } 2781 2782 static struct kmem_cache *kmem_cache_node; 2783 2784 /* 2785 * No kmalloc_node yet so do it by hand. We know that this is the first 2786 * slab on the node for this slabcache. There are no concurrent accesses 2787 * possible. 2788 * 2789 * Note that this function only works on the kmalloc_node_cache 2790 * when allocating for the kmalloc_node_cache. This is used for bootstrapping 2791 * memory on a fresh node that has no slab structures yet. 2792 */ 2793 static void early_kmem_cache_node_alloc(int node) 2794 { 2795 struct page *page; 2796 struct kmem_cache_node *n; 2797 2798 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 2799 2800 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 2801 2802 BUG_ON(!page); 2803 if (page_to_nid(page) != node) { 2804 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2805 "node %d\n", node); 2806 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2807 "in order to be able to continue\n"); 2808 } 2809 2810 n = page->freelist; 2811 BUG_ON(!n); 2812 page->freelist = get_freepointer(kmem_cache_node, n); 2813 page->inuse = 1; 2814 page->frozen = 0; 2815 kmem_cache_node->node[node] = n; 2816 #ifdef CONFIG_SLUB_DEBUG 2817 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 2818 init_tracking(kmem_cache_node, n); 2819 #endif 2820 init_kmem_cache_node(n, kmem_cache_node); 2821 inc_slabs_node(kmem_cache_node, node, page->objects); 2822 2823 add_partial(n, page, DEACTIVATE_TO_HEAD); 2824 } 2825 2826 static void free_kmem_cache_nodes(struct kmem_cache *s) 2827 { 2828 int node; 2829 2830 for_each_node_state(node, N_NORMAL_MEMORY) { 2831 struct kmem_cache_node *n = s->node[node]; 2832 2833 if (n) 2834 kmem_cache_free(kmem_cache_node, n); 2835 2836 s->node[node] = NULL; 2837 } 2838 } 2839 2840 static int init_kmem_cache_nodes(struct kmem_cache *s) 2841 { 2842 int node; 2843 2844 for_each_node_state(node, N_NORMAL_MEMORY) { 2845 struct kmem_cache_node *n; 2846 2847 if (slab_state == DOWN) { 2848 early_kmem_cache_node_alloc(node); 2849 continue; 2850 } 2851 n = kmem_cache_alloc_node(kmem_cache_node, 2852 GFP_KERNEL, node); 2853 2854 if (!n) { 2855 free_kmem_cache_nodes(s); 2856 return 0; 2857 } 2858 2859 s->node[node] = n; 2860 init_kmem_cache_node(n, s); 2861 } 2862 return 1; 2863 } 2864 2865 static void set_min_partial(struct kmem_cache *s, unsigned long min) 2866 { 2867 if (min < MIN_PARTIAL) 2868 min = MIN_PARTIAL; 2869 else if (min > MAX_PARTIAL) 2870 min = MAX_PARTIAL; 2871 s->min_partial = min; 2872 } 2873 2874 /* 2875 * calculate_sizes() determines the order and the distribution of data within 2876 * a slab object. 2877 */ 2878 static int calculate_sizes(struct kmem_cache *s, int forced_order) 2879 { 2880 unsigned long flags = s->flags; 2881 unsigned long size = s->objsize; 2882 unsigned long align = s->align; 2883 int order; 2884 2885 /* 2886 * Round up object size to the next word boundary. We can only 2887 * place the free pointer at word boundaries and this determines 2888 * the possible location of the free pointer. 2889 */ 2890 size = ALIGN(size, sizeof(void *)); 2891 2892 #ifdef CONFIG_SLUB_DEBUG 2893 /* 2894 * Determine if we can poison the object itself. If the user of 2895 * the slab may touch the object after free or before allocation 2896 * then we should never poison the object itself. 2897 */ 2898 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2899 !s->ctor) 2900 s->flags |= __OBJECT_POISON; 2901 else 2902 s->flags &= ~__OBJECT_POISON; 2903 2904 2905 /* 2906 * If we are Redzoning then check if there is some space between the 2907 * end of the object and the free pointer. If not then add an 2908 * additional word to have some bytes to store Redzone information. 2909 */ 2910 if ((flags & SLAB_RED_ZONE) && size == s->objsize) 2911 size += sizeof(void *); 2912 #endif 2913 2914 /* 2915 * With that we have determined the number of bytes in actual use 2916 * by the object. This is the potential offset to the free pointer. 2917 */ 2918 s->inuse = size; 2919 2920 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 2921 s->ctor)) { 2922 /* 2923 * Relocate free pointer after the object if it is not 2924 * permitted to overwrite the first word of the object on 2925 * kmem_cache_free. 2926 * 2927 * This is the case if we do RCU, have a constructor or 2928 * destructor or are poisoning the objects. 2929 */ 2930 s->offset = size; 2931 size += sizeof(void *); 2932 } 2933 2934 #ifdef CONFIG_SLUB_DEBUG 2935 if (flags & SLAB_STORE_USER) 2936 /* 2937 * Need to store information about allocs and frees after 2938 * the object. 2939 */ 2940 size += 2 * sizeof(struct track); 2941 2942 if (flags & SLAB_RED_ZONE) 2943 /* 2944 * Add some empty padding so that we can catch 2945 * overwrites from earlier objects rather than let 2946 * tracking information or the free pointer be 2947 * corrupted if a user writes before the start 2948 * of the object. 2949 */ 2950 size += sizeof(void *); 2951 #endif 2952 2953 /* 2954 * Determine the alignment based on various parameters that the 2955 * user specified and the dynamic determination of cache line size 2956 * on bootup. 2957 */ 2958 align = calculate_alignment(flags, align, s->objsize); 2959 s->align = align; 2960 2961 /* 2962 * SLUB stores one object immediately after another beginning from 2963 * offset 0. In order to align the objects we have to simply size 2964 * each object to conform to the alignment. 2965 */ 2966 size = ALIGN(size, align); 2967 s->size = size; 2968 if (forced_order >= 0) 2969 order = forced_order; 2970 else 2971 order = calculate_order(size, s->reserved); 2972 2973 if (order < 0) 2974 return 0; 2975 2976 s->allocflags = 0; 2977 if (order) 2978 s->allocflags |= __GFP_COMP; 2979 2980 if (s->flags & SLAB_CACHE_DMA) 2981 s->allocflags |= SLUB_DMA; 2982 2983 if (s->flags & SLAB_RECLAIM_ACCOUNT) 2984 s->allocflags |= __GFP_RECLAIMABLE; 2985 2986 /* 2987 * Determine the number of objects per slab 2988 */ 2989 s->oo = oo_make(order, size, s->reserved); 2990 s->min = oo_make(get_order(size), size, s->reserved); 2991 if (oo_objects(s->oo) > oo_objects(s->max)) 2992 s->max = s->oo; 2993 2994 return !!oo_objects(s->oo); 2995 2996 } 2997 2998 static int kmem_cache_open(struct kmem_cache *s, 2999 const char *name, size_t size, 3000 size_t align, unsigned long flags, 3001 void (*ctor)(void *)) 3002 { 3003 memset(s, 0, kmem_size); 3004 s->name = name; 3005 s->ctor = ctor; 3006 s->objsize = size; 3007 s->align = align; 3008 s->flags = kmem_cache_flags(size, flags, name, ctor); 3009 s->reserved = 0; 3010 3011 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) 3012 s->reserved = sizeof(struct rcu_head); 3013 3014 if (!calculate_sizes(s, -1)) 3015 goto error; 3016 if (disable_higher_order_debug) { 3017 /* 3018 * Disable debugging flags that store metadata if the min slab 3019 * order increased. 3020 */ 3021 if (get_order(s->size) > get_order(s->objsize)) { 3022 s->flags &= ~DEBUG_METADATA_FLAGS; 3023 s->offset = 0; 3024 if (!calculate_sizes(s, -1)) 3025 goto error; 3026 } 3027 } 3028 3029 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 3030 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 3031 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0) 3032 /* Enable fast mode */ 3033 s->flags |= __CMPXCHG_DOUBLE; 3034 #endif 3035 3036 /* 3037 * The larger the object size is, the more pages we want on the partial 3038 * list to avoid pounding the page allocator excessively. 3039 */ 3040 set_min_partial(s, ilog2(s->size) / 2); 3041 3042 /* 3043 * cpu_partial determined the maximum number of objects kept in the 3044 * per cpu partial lists of a processor. 3045 * 3046 * Per cpu partial lists mainly contain slabs that just have one 3047 * object freed. If they are used for allocation then they can be 3048 * filled up again with minimal effort. The slab will never hit the 3049 * per node partial lists and therefore no locking will be required. 3050 * 3051 * This setting also determines 3052 * 3053 * A) The number of objects from per cpu partial slabs dumped to the 3054 * per node list when we reach the limit. 3055 * B) The number of objects in cpu partial slabs to extract from the 3056 * per node list when we run out of per cpu objects. We only fetch 50% 3057 * to keep some capacity around for frees. 3058 */ 3059 if (kmem_cache_debug(s)) 3060 s->cpu_partial = 0; 3061 else if (s->size >= PAGE_SIZE) 3062 s->cpu_partial = 2; 3063 else if (s->size >= 1024) 3064 s->cpu_partial = 6; 3065 else if (s->size >= 256) 3066 s->cpu_partial = 13; 3067 else 3068 s->cpu_partial = 30; 3069 3070 s->refcount = 1; 3071 #ifdef CONFIG_NUMA 3072 s->remote_node_defrag_ratio = 1000; 3073 #endif 3074 if (!init_kmem_cache_nodes(s)) 3075 goto error; 3076 3077 if (alloc_kmem_cache_cpus(s)) 3078 return 1; 3079 3080 free_kmem_cache_nodes(s); 3081 error: 3082 if (flags & SLAB_PANIC) 3083 panic("Cannot create slab %s size=%lu realsize=%u " 3084 "order=%u offset=%u flags=%lx\n", 3085 s->name, (unsigned long)size, s->size, oo_order(s->oo), 3086 s->offset, flags); 3087 return 0; 3088 } 3089 3090 /* 3091 * Determine the size of a slab object 3092 */ 3093 unsigned int kmem_cache_size(struct kmem_cache *s) 3094 { 3095 return s->objsize; 3096 } 3097 EXPORT_SYMBOL(kmem_cache_size); 3098 3099 static void list_slab_objects(struct kmem_cache *s, struct page *page, 3100 const char *text) 3101 { 3102 #ifdef CONFIG_SLUB_DEBUG 3103 void *addr = page_address(page); 3104 void *p; 3105 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * 3106 sizeof(long), GFP_ATOMIC); 3107 if (!map) 3108 return; 3109 slab_err(s, page, "%s", text); 3110 slab_lock(page); 3111 3112 get_map(s, page, map); 3113 for_each_object(p, s, addr, page->objects) { 3114 3115 if (!test_bit(slab_index(p, s, addr), map)) { 3116 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 3117 p, p - addr); 3118 print_tracking(s, p); 3119 } 3120 } 3121 slab_unlock(page); 3122 kfree(map); 3123 #endif 3124 } 3125 3126 /* 3127 * Attempt to free all partial slabs on a node. 3128 * This is called from kmem_cache_close(). We must be the last thread 3129 * using the cache and therefore we do not need to lock anymore. 3130 */ 3131 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 3132 { 3133 struct page *page, *h; 3134 3135 list_for_each_entry_safe(page, h, &n->partial, lru) { 3136 if (!page->inuse) { 3137 remove_partial(n, page); 3138 discard_slab(s, page); 3139 } else { 3140 list_slab_objects(s, page, 3141 "Objects remaining on kmem_cache_close()"); 3142 } 3143 } 3144 } 3145 3146 /* 3147 * Release all resources used by a slab cache. 3148 */ 3149 static inline int kmem_cache_close(struct kmem_cache *s) 3150 { 3151 int node; 3152 3153 flush_all(s); 3154 free_percpu(s->cpu_slab); 3155 /* Attempt to free all objects */ 3156 for_each_node_state(node, N_NORMAL_MEMORY) { 3157 struct kmem_cache_node *n = get_node(s, node); 3158 3159 free_partial(s, n); 3160 if (n->nr_partial || slabs_node(s, node)) 3161 return 1; 3162 } 3163 free_kmem_cache_nodes(s); 3164 return 0; 3165 } 3166 3167 /* 3168 * Close a cache and release the kmem_cache structure 3169 * (must be used for caches created using kmem_cache_create) 3170 */ 3171 void kmem_cache_destroy(struct kmem_cache *s) 3172 { 3173 down_write(&slub_lock); 3174 s->refcount--; 3175 if (!s->refcount) { 3176 list_del(&s->list); 3177 up_write(&slub_lock); 3178 if (kmem_cache_close(s)) { 3179 printk(KERN_ERR "SLUB %s: %s called for cache that " 3180 "still has objects.\n", s->name, __func__); 3181 dump_stack(); 3182 } 3183 if (s->flags & SLAB_DESTROY_BY_RCU) 3184 rcu_barrier(); 3185 sysfs_slab_remove(s); 3186 } else 3187 up_write(&slub_lock); 3188 } 3189 EXPORT_SYMBOL(kmem_cache_destroy); 3190 3191 /******************************************************************** 3192 * Kmalloc subsystem 3193 *******************************************************************/ 3194 3195 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT]; 3196 EXPORT_SYMBOL(kmalloc_caches); 3197 3198 static struct kmem_cache *kmem_cache; 3199 3200 #ifdef CONFIG_ZONE_DMA 3201 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT]; 3202 #endif 3203 3204 static int __init setup_slub_min_order(char *str) 3205 { 3206 get_option(&str, &slub_min_order); 3207 3208 return 1; 3209 } 3210 3211 __setup("slub_min_order=", setup_slub_min_order); 3212 3213 static int __init setup_slub_max_order(char *str) 3214 { 3215 get_option(&str, &slub_max_order); 3216 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 3217 3218 return 1; 3219 } 3220 3221 __setup("slub_max_order=", setup_slub_max_order); 3222 3223 static int __init setup_slub_min_objects(char *str) 3224 { 3225 get_option(&str, &slub_min_objects); 3226 3227 return 1; 3228 } 3229 3230 __setup("slub_min_objects=", setup_slub_min_objects); 3231 3232 static int __init setup_slub_nomerge(char *str) 3233 { 3234 slub_nomerge = 1; 3235 return 1; 3236 } 3237 3238 __setup("slub_nomerge", setup_slub_nomerge); 3239 3240 static struct kmem_cache *__init create_kmalloc_cache(const char *name, 3241 int size, unsigned int flags) 3242 { 3243 struct kmem_cache *s; 3244 3245 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3246 3247 /* 3248 * This function is called with IRQs disabled during early-boot on 3249 * single CPU so there's no need to take slub_lock here. 3250 */ 3251 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN, 3252 flags, NULL)) 3253 goto panic; 3254 3255 list_add(&s->list, &slab_caches); 3256 return s; 3257 3258 panic: 3259 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size); 3260 return NULL; 3261 } 3262 3263 /* 3264 * Conversion table for small slabs sizes / 8 to the index in the 3265 * kmalloc array. This is necessary for slabs < 192 since we have non power 3266 * of two cache sizes there. The size of larger slabs can be determined using 3267 * fls. 3268 */ 3269 static s8 size_index[24] = { 3270 3, /* 8 */ 3271 4, /* 16 */ 3272 5, /* 24 */ 3273 5, /* 32 */ 3274 6, /* 40 */ 3275 6, /* 48 */ 3276 6, /* 56 */ 3277 6, /* 64 */ 3278 1, /* 72 */ 3279 1, /* 80 */ 3280 1, /* 88 */ 3281 1, /* 96 */ 3282 7, /* 104 */ 3283 7, /* 112 */ 3284 7, /* 120 */ 3285 7, /* 128 */ 3286 2, /* 136 */ 3287 2, /* 144 */ 3288 2, /* 152 */ 3289 2, /* 160 */ 3290 2, /* 168 */ 3291 2, /* 176 */ 3292 2, /* 184 */ 3293 2 /* 192 */ 3294 }; 3295 3296 static inline int size_index_elem(size_t bytes) 3297 { 3298 return (bytes - 1) / 8; 3299 } 3300 3301 static struct kmem_cache *get_slab(size_t size, gfp_t flags) 3302 { 3303 int index; 3304 3305 if (size <= 192) { 3306 if (!size) 3307 return ZERO_SIZE_PTR; 3308 3309 index = size_index[size_index_elem(size)]; 3310 } else 3311 index = fls(size - 1); 3312 3313 #ifdef CONFIG_ZONE_DMA 3314 if (unlikely((flags & SLUB_DMA))) 3315 return kmalloc_dma_caches[index]; 3316 3317 #endif 3318 return kmalloc_caches[index]; 3319 } 3320 3321 void *__kmalloc(size_t size, gfp_t flags) 3322 { 3323 struct kmem_cache *s; 3324 void *ret; 3325 3326 if (unlikely(size > SLUB_MAX_SIZE)) 3327 return kmalloc_large(size, flags); 3328 3329 s = get_slab(size, flags); 3330 3331 if (unlikely(ZERO_OR_NULL_PTR(s))) 3332 return s; 3333 3334 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_); 3335 3336 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 3337 3338 return ret; 3339 } 3340 EXPORT_SYMBOL(__kmalloc); 3341 3342 #ifdef CONFIG_NUMA 3343 static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 3344 { 3345 struct page *page; 3346 void *ptr = NULL; 3347 3348 flags |= __GFP_COMP | __GFP_NOTRACK; 3349 page = alloc_pages_node(node, flags, get_order(size)); 3350 if (page) 3351 ptr = page_address(page); 3352 3353 kmemleak_alloc(ptr, size, 1, flags); 3354 return ptr; 3355 } 3356 3357 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3358 { 3359 struct kmem_cache *s; 3360 void *ret; 3361 3362 if (unlikely(size > SLUB_MAX_SIZE)) { 3363 ret = kmalloc_large_node(size, flags, node); 3364 3365 trace_kmalloc_node(_RET_IP_, ret, 3366 size, PAGE_SIZE << get_order(size), 3367 flags, node); 3368 3369 return ret; 3370 } 3371 3372 s = get_slab(size, flags); 3373 3374 if (unlikely(ZERO_OR_NULL_PTR(s))) 3375 return s; 3376 3377 ret = slab_alloc(s, flags, node, _RET_IP_); 3378 3379 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 3380 3381 return ret; 3382 } 3383 EXPORT_SYMBOL(__kmalloc_node); 3384 #endif 3385 3386 size_t ksize(const void *object) 3387 { 3388 struct page *page; 3389 3390 if (unlikely(object == ZERO_SIZE_PTR)) 3391 return 0; 3392 3393 page = virt_to_head_page(object); 3394 3395 if (unlikely(!PageSlab(page))) { 3396 WARN_ON(!PageCompound(page)); 3397 return PAGE_SIZE << compound_order(page); 3398 } 3399 3400 return slab_ksize(page->slab); 3401 } 3402 EXPORT_SYMBOL(ksize); 3403 3404 #ifdef CONFIG_SLUB_DEBUG 3405 bool verify_mem_not_deleted(const void *x) 3406 { 3407 struct page *page; 3408 void *object = (void *)x; 3409 unsigned long flags; 3410 bool rv; 3411 3412 if (unlikely(ZERO_OR_NULL_PTR(x))) 3413 return false; 3414 3415 local_irq_save(flags); 3416 3417 page = virt_to_head_page(x); 3418 if (unlikely(!PageSlab(page))) { 3419 /* maybe it was from stack? */ 3420 rv = true; 3421 goto out_unlock; 3422 } 3423 3424 slab_lock(page); 3425 if (on_freelist(page->slab, page, object)) { 3426 object_err(page->slab, page, object, "Object is on free-list"); 3427 rv = false; 3428 } else { 3429 rv = true; 3430 } 3431 slab_unlock(page); 3432 3433 out_unlock: 3434 local_irq_restore(flags); 3435 return rv; 3436 } 3437 EXPORT_SYMBOL(verify_mem_not_deleted); 3438 #endif 3439 3440 void kfree(const void *x) 3441 { 3442 struct page *page; 3443 void *object = (void *)x; 3444 3445 trace_kfree(_RET_IP_, x); 3446 3447 if (unlikely(ZERO_OR_NULL_PTR(x))) 3448 return; 3449 3450 page = virt_to_head_page(x); 3451 if (unlikely(!PageSlab(page))) { 3452 BUG_ON(!PageCompound(page)); 3453 kmemleak_free(x); 3454 put_page(page); 3455 return; 3456 } 3457 slab_free(page->slab, page, object, _RET_IP_); 3458 } 3459 EXPORT_SYMBOL(kfree); 3460 3461 /* 3462 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 3463 * the remaining slabs by the number of items in use. The slabs with the 3464 * most items in use come first. New allocations will then fill those up 3465 * and thus they can be removed from the partial lists. 3466 * 3467 * The slabs with the least items are placed last. This results in them 3468 * being allocated from last increasing the chance that the last objects 3469 * are freed in them. 3470 */ 3471 int kmem_cache_shrink(struct kmem_cache *s) 3472 { 3473 int node; 3474 int i; 3475 struct kmem_cache_node *n; 3476 struct page *page; 3477 struct page *t; 3478 int objects = oo_objects(s->max); 3479 struct list_head *slabs_by_inuse = 3480 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 3481 unsigned long flags; 3482 3483 if (!slabs_by_inuse) 3484 return -ENOMEM; 3485 3486 flush_all(s); 3487 for_each_node_state(node, N_NORMAL_MEMORY) { 3488 n = get_node(s, node); 3489 3490 if (!n->nr_partial) 3491 continue; 3492 3493 for (i = 0; i < objects; i++) 3494 INIT_LIST_HEAD(slabs_by_inuse + i); 3495 3496 spin_lock_irqsave(&n->list_lock, flags); 3497 3498 /* 3499 * Build lists indexed by the items in use in each slab. 3500 * 3501 * Note that concurrent frees may occur while we hold the 3502 * list_lock. page->inuse here is the upper limit. 3503 */ 3504 list_for_each_entry_safe(page, t, &n->partial, lru) { 3505 list_move(&page->lru, slabs_by_inuse + page->inuse); 3506 if (!page->inuse) 3507 n->nr_partial--; 3508 } 3509 3510 /* 3511 * Rebuild the partial list with the slabs filled up most 3512 * first and the least used slabs at the end. 3513 */ 3514 for (i = objects - 1; i > 0; i--) 3515 list_splice(slabs_by_inuse + i, n->partial.prev); 3516 3517 spin_unlock_irqrestore(&n->list_lock, flags); 3518 3519 /* Release empty slabs */ 3520 list_for_each_entry_safe(page, t, slabs_by_inuse, lru) 3521 discard_slab(s, page); 3522 } 3523 3524 kfree(slabs_by_inuse); 3525 return 0; 3526 } 3527 EXPORT_SYMBOL(kmem_cache_shrink); 3528 3529 #if defined(CONFIG_MEMORY_HOTPLUG) 3530 static int slab_mem_going_offline_callback(void *arg) 3531 { 3532 struct kmem_cache *s; 3533 3534 down_read(&slub_lock); 3535 list_for_each_entry(s, &slab_caches, list) 3536 kmem_cache_shrink(s); 3537 up_read(&slub_lock); 3538 3539 return 0; 3540 } 3541 3542 static void slab_mem_offline_callback(void *arg) 3543 { 3544 struct kmem_cache_node *n; 3545 struct kmem_cache *s; 3546 struct memory_notify *marg = arg; 3547 int offline_node; 3548 3549 offline_node = marg->status_change_nid; 3550 3551 /* 3552 * If the node still has available memory. we need kmem_cache_node 3553 * for it yet. 3554 */ 3555 if (offline_node < 0) 3556 return; 3557 3558 down_read(&slub_lock); 3559 list_for_each_entry(s, &slab_caches, list) { 3560 n = get_node(s, offline_node); 3561 if (n) { 3562 /* 3563 * if n->nr_slabs > 0, slabs still exist on the node 3564 * that is going down. We were unable to free them, 3565 * and offline_pages() function shouldn't call this 3566 * callback. So, we must fail. 3567 */ 3568 BUG_ON(slabs_node(s, offline_node)); 3569 3570 s->node[offline_node] = NULL; 3571 kmem_cache_free(kmem_cache_node, n); 3572 } 3573 } 3574 up_read(&slub_lock); 3575 } 3576 3577 static int slab_mem_going_online_callback(void *arg) 3578 { 3579 struct kmem_cache_node *n; 3580 struct kmem_cache *s; 3581 struct memory_notify *marg = arg; 3582 int nid = marg->status_change_nid; 3583 int ret = 0; 3584 3585 /* 3586 * If the node's memory is already available, then kmem_cache_node is 3587 * already created. Nothing to do. 3588 */ 3589 if (nid < 0) 3590 return 0; 3591 3592 /* 3593 * We are bringing a node online. No memory is available yet. We must 3594 * allocate a kmem_cache_node structure in order to bring the node 3595 * online. 3596 */ 3597 down_read(&slub_lock); 3598 list_for_each_entry(s, &slab_caches, list) { 3599 /* 3600 * XXX: kmem_cache_alloc_node will fallback to other nodes 3601 * since memory is not yet available from the node that 3602 * is brought up. 3603 */ 3604 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 3605 if (!n) { 3606 ret = -ENOMEM; 3607 goto out; 3608 } 3609 init_kmem_cache_node(n, s); 3610 s->node[nid] = n; 3611 } 3612 out: 3613 up_read(&slub_lock); 3614 return ret; 3615 } 3616 3617 static int slab_memory_callback(struct notifier_block *self, 3618 unsigned long action, void *arg) 3619 { 3620 int ret = 0; 3621 3622 switch (action) { 3623 case MEM_GOING_ONLINE: 3624 ret = slab_mem_going_online_callback(arg); 3625 break; 3626 case MEM_GOING_OFFLINE: 3627 ret = slab_mem_going_offline_callback(arg); 3628 break; 3629 case MEM_OFFLINE: 3630 case MEM_CANCEL_ONLINE: 3631 slab_mem_offline_callback(arg); 3632 break; 3633 case MEM_ONLINE: 3634 case MEM_CANCEL_OFFLINE: 3635 break; 3636 } 3637 if (ret) 3638 ret = notifier_from_errno(ret); 3639 else 3640 ret = NOTIFY_OK; 3641 return ret; 3642 } 3643 3644 #endif /* CONFIG_MEMORY_HOTPLUG */ 3645 3646 /******************************************************************** 3647 * Basic setup of slabs 3648 *******************************************************************/ 3649 3650 /* 3651 * Used for early kmem_cache structures that were allocated using 3652 * the page allocator 3653 */ 3654 3655 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s) 3656 { 3657 int node; 3658 3659 list_add(&s->list, &slab_caches); 3660 s->refcount = -1; 3661 3662 for_each_node_state(node, N_NORMAL_MEMORY) { 3663 struct kmem_cache_node *n = get_node(s, node); 3664 struct page *p; 3665 3666 if (n) { 3667 list_for_each_entry(p, &n->partial, lru) 3668 p->slab = s; 3669 3670 #ifdef CONFIG_SLUB_DEBUG 3671 list_for_each_entry(p, &n->full, lru) 3672 p->slab = s; 3673 #endif 3674 } 3675 } 3676 } 3677 3678 void __init kmem_cache_init(void) 3679 { 3680 int i; 3681 int caches = 0; 3682 struct kmem_cache *temp_kmem_cache; 3683 int order; 3684 struct kmem_cache *temp_kmem_cache_node; 3685 unsigned long kmalloc_size; 3686 3687 if (debug_guardpage_minorder()) 3688 slub_max_order = 0; 3689 3690 kmem_size = offsetof(struct kmem_cache, node) + 3691 nr_node_ids * sizeof(struct kmem_cache_node *); 3692 3693 /* Allocate two kmem_caches from the page allocator */ 3694 kmalloc_size = ALIGN(kmem_size, cache_line_size()); 3695 order = get_order(2 * kmalloc_size); 3696 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order); 3697 3698 /* 3699 * Must first have the slab cache available for the allocations of the 3700 * struct kmem_cache_node's. There is special bootstrap code in 3701 * kmem_cache_open for slab_state == DOWN. 3702 */ 3703 kmem_cache_node = (void *)kmem_cache + kmalloc_size; 3704 3705 kmem_cache_open(kmem_cache_node, "kmem_cache_node", 3706 sizeof(struct kmem_cache_node), 3707 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3708 3709 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); 3710 3711 /* Able to allocate the per node structures */ 3712 slab_state = PARTIAL; 3713 3714 temp_kmem_cache = kmem_cache; 3715 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size, 3716 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL); 3717 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3718 memcpy(kmem_cache, temp_kmem_cache, kmem_size); 3719 3720 /* 3721 * Allocate kmem_cache_node properly from the kmem_cache slab. 3722 * kmem_cache_node is separately allocated so no need to 3723 * update any list pointers. 3724 */ 3725 temp_kmem_cache_node = kmem_cache_node; 3726 3727 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT); 3728 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size); 3729 3730 kmem_cache_bootstrap_fixup(kmem_cache_node); 3731 3732 caches++; 3733 kmem_cache_bootstrap_fixup(kmem_cache); 3734 caches++; 3735 /* Free temporary boot structure */ 3736 free_pages((unsigned long)temp_kmem_cache, order); 3737 3738 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 3739 3740 /* 3741 * Patch up the size_index table if we have strange large alignment 3742 * requirements for the kmalloc array. This is only the case for 3743 * MIPS it seems. The standard arches will not generate any code here. 3744 * 3745 * Largest permitted alignment is 256 bytes due to the way we 3746 * handle the index determination for the smaller caches. 3747 * 3748 * Make sure that nothing crazy happens if someone starts tinkering 3749 * around with ARCH_KMALLOC_MINALIGN 3750 */ 3751 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 || 3752 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1))); 3753 3754 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) { 3755 int elem = size_index_elem(i); 3756 if (elem >= ARRAY_SIZE(size_index)) 3757 break; 3758 size_index[elem] = KMALLOC_SHIFT_LOW; 3759 } 3760 3761 if (KMALLOC_MIN_SIZE == 64) { 3762 /* 3763 * The 96 byte size cache is not used if the alignment 3764 * is 64 byte. 3765 */ 3766 for (i = 64 + 8; i <= 96; i += 8) 3767 size_index[size_index_elem(i)] = 7; 3768 } else if (KMALLOC_MIN_SIZE == 128) { 3769 /* 3770 * The 192 byte sized cache is not used if the alignment 3771 * is 128 byte. Redirect kmalloc to use the 256 byte cache 3772 * instead. 3773 */ 3774 for (i = 128 + 8; i <= 192; i += 8) 3775 size_index[size_index_elem(i)] = 8; 3776 } 3777 3778 /* Caches that are not of the two-to-the-power-of size */ 3779 if (KMALLOC_MIN_SIZE <= 32) { 3780 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0); 3781 caches++; 3782 } 3783 3784 if (KMALLOC_MIN_SIZE <= 64) { 3785 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0); 3786 caches++; 3787 } 3788 3789 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3790 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0); 3791 caches++; 3792 } 3793 3794 slab_state = UP; 3795 3796 /* Provide the correct kmalloc names now that the caches are up */ 3797 if (KMALLOC_MIN_SIZE <= 32) { 3798 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT); 3799 BUG_ON(!kmalloc_caches[1]->name); 3800 } 3801 3802 if (KMALLOC_MIN_SIZE <= 64) { 3803 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT); 3804 BUG_ON(!kmalloc_caches[2]->name); 3805 } 3806 3807 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) { 3808 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i); 3809 3810 BUG_ON(!s); 3811 kmalloc_caches[i]->name = s; 3812 } 3813 3814 #ifdef CONFIG_SMP 3815 register_cpu_notifier(&slab_notifier); 3816 #endif 3817 3818 #ifdef CONFIG_ZONE_DMA 3819 for (i = 0; i < SLUB_PAGE_SHIFT; i++) { 3820 struct kmem_cache *s = kmalloc_caches[i]; 3821 3822 if (s && s->size) { 3823 char *name = kasprintf(GFP_NOWAIT, 3824 "dma-kmalloc-%d", s->objsize); 3825 3826 BUG_ON(!name); 3827 kmalloc_dma_caches[i] = create_kmalloc_cache(name, 3828 s->objsize, SLAB_CACHE_DMA); 3829 } 3830 } 3831 #endif 3832 printk(KERN_INFO 3833 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d," 3834 " CPUs=%d, Nodes=%d\n", 3835 caches, cache_line_size(), 3836 slub_min_order, slub_max_order, slub_min_objects, 3837 nr_cpu_ids, nr_node_ids); 3838 } 3839 3840 void __init kmem_cache_init_late(void) 3841 { 3842 } 3843 3844 /* 3845 * Find a mergeable slab cache 3846 */ 3847 static int slab_unmergeable(struct kmem_cache *s) 3848 { 3849 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3850 return 1; 3851 3852 if (s->ctor) 3853 return 1; 3854 3855 /* 3856 * We may have set a slab to be unmergeable during bootstrap. 3857 */ 3858 if (s->refcount < 0) 3859 return 1; 3860 3861 return 0; 3862 } 3863 3864 static struct kmem_cache *find_mergeable(size_t size, 3865 size_t align, unsigned long flags, const char *name, 3866 void (*ctor)(void *)) 3867 { 3868 struct kmem_cache *s; 3869 3870 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3871 return NULL; 3872 3873 if (ctor) 3874 return NULL; 3875 3876 size = ALIGN(size, sizeof(void *)); 3877 align = calculate_alignment(flags, align, size); 3878 size = ALIGN(size, align); 3879 flags = kmem_cache_flags(size, flags, name, NULL); 3880 3881 list_for_each_entry(s, &slab_caches, list) { 3882 if (slab_unmergeable(s)) 3883 continue; 3884 3885 if (size > s->size) 3886 continue; 3887 3888 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3889 continue; 3890 /* 3891 * Check if alignment is compatible. 3892 * Courtesy of Adrian Drzewiecki 3893 */ 3894 if ((s->size & ~(align - 1)) != s->size) 3895 continue; 3896 3897 if (s->size - size >= sizeof(void *)) 3898 continue; 3899 3900 return s; 3901 } 3902 return NULL; 3903 } 3904 3905 struct kmem_cache *kmem_cache_create(const char *name, size_t size, 3906 size_t align, unsigned long flags, void (*ctor)(void *)) 3907 { 3908 struct kmem_cache *s; 3909 char *n; 3910 3911 if (WARN_ON(!name)) 3912 return NULL; 3913 3914 down_write(&slub_lock); 3915 s = find_mergeable(size, align, flags, name, ctor); 3916 if (s) { 3917 s->refcount++; 3918 /* 3919 * Adjust the object sizes so that we clear 3920 * the complete object on kzalloc. 3921 */ 3922 s->objsize = max(s->objsize, (int)size); 3923 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3924 3925 if (sysfs_slab_alias(s, name)) { 3926 s->refcount--; 3927 goto err; 3928 } 3929 up_write(&slub_lock); 3930 return s; 3931 } 3932 3933 n = kstrdup(name, GFP_KERNEL); 3934 if (!n) 3935 goto err; 3936 3937 s = kmalloc(kmem_size, GFP_KERNEL); 3938 if (s) { 3939 if (kmem_cache_open(s, n, 3940 size, align, flags, ctor)) { 3941 list_add(&s->list, &slab_caches); 3942 if (sysfs_slab_add(s)) { 3943 list_del(&s->list); 3944 kfree(n); 3945 kfree(s); 3946 goto err; 3947 } 3948 up_write(&slub_lock); 3949 return s; 3950 } 3951 kfree(n); 3952 kfree(s); 3953 } 3954 err: 3955 up_write(&slub_lock); 3956 3957 if (flags & SLAB_PANIC) 3958 panic("Cannot create slabcache %s\n", name); 3959 else 3960 s = NULL; 3961 return s; 3962 } 3963 EXPORT_SYMBOL(kmem_cache_create); 3964 3965 #ifdef CONFIG_SMP 3966 /* 3967 * Use the cpu notifier to insure that the cpu slabs are flushed when 3968 * necessary. 3969 */ 3970 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb, 3971 unsigned long action, void *hcpu) 3972 { 3973 long cpu = (long)hcpu; 3974 struct kmem_cache *s; 3975 unsigned long flags; 3976 3977 switch (action) { 3978 case CPU_UP_CANCELED: 3979 case CPU_UP_CANCELED_FROZEN: 3980 case CPU_DEAD: 3981 case CPU_DEAD_FROZEN: 3982 down_read(&slub_lock); 3983 list_for_each_entry(s, &slab_caches, list) { 3984 local_irq_save(flags); 3985 __flush_cpu_slab(s, cpu); 3986 local_irq_restore(flags); 3987 } 3988 up_read(&slub_lock); 3989 break; 3990 default: 3991 break; 3992 } 3993 return NOTIFY_OK; 3994 } 3995 3996 static struct notifier_block __cpuinitdata slab_notifier = { 3997 .notifier_call = slab_cpuup_callback 3998 }; 3999 4000 #endif 4001 4002 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 4003 { 4004 struct kmem_cache *s; 4005 void *ret; 4006 4007 if (unlikely(size > SLUB_MAX_SIZE)) 4008 return kmalloc_large(size, gfpflags); 4009 4010 s = get_slab(size, gfpflags); 4011 4012 if (unlikely(ZERO_OR_NULL_PTR(s))) 4013 return s; 4014 4015 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller); 4016 4017 /* Honor the call site pointer we received. */ 4018 trace_kmalloc(caller, ret, size, s->size, gfpflags); 4019 4020 return ret; 4021 } 4022 4023 #ifdef CONFIG_NUMA 4024 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 4025 int node, unsigned long caller) 4026 { 4027 struct kmem_cache *s; 4028 void *ret; 4029 4030 if (unlikely(size > SLUB_MAX_SIZE)) { 4031 ret = kmalloc_large_node(size, gfpflags, node); 4032 4033 trace_kmalloc_node(caller, ret, 4034 size, PAGE_SIZE << get_order(size), 4035 gfpflags, node); 4036 4037 return ret; 4038 } 4039 4040 s = get_slab(size, gfpflags); 4041 4042 if (unlikely(ZERO_OR_NULL_PTR(s))) 4043 return s; 4044 4045 ret = slab_alloc(s, gfpflags, node, caller); 4046 4047 /* Honor the call site pointer we received. */ 4048 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 4049 4050 return ret; 4051 } 4052 #endif 4053 4054 #ifdef CONFIG_SYSFS 4055 static int count_inuse(struct page *page) 4056 { 4057 return page->inuse; 4058 } 4059 4060 static int count_total(struct page *page) 4061 { 4062 return page->objects; 4063 } 4064 #endif 4065 4066 #ifdef CONFIG_SLUB_DEBUG 4067 static int validate_slab(struct kmem_cache *s, struct page *page, 4068 unsigned long *map) 4069 { 4070 void *p; 4071 void *addr = page_address(page); 4072 4073 if (!check_slab(s, page) || 4074 !on_freelist(s, page, NULL)) 4075 return 0; 4076 4077 /* Now we know that a valid freelist exists */ 4078 bitmap_zero(map, page->objects); 4079 4080 get_map(s, page, map); 4081 for_each_object(p, s, addr, page->objects) { 4082 if (test_bit(slab_index(p, s, addr), map)) 4083 if (!check_object(s, page, p, SLUB_RED_INACTIVE)) 4084 return 0; 4085 } 4086 4087 for_each_object(p, s, addr, page->objects) 4088 if (!test_bit(slab_index(p, s, addr), map)) 4089 if (!check_object(s, page, p, SLUB_RED_ACTIVE)) 4090 return 0; 4091 return 1; 4092 } 4093 4094 static void validate_slab_slab(struct kmem_cache *s, struct page *page, 4095 unsigned long *map) 4096 { 4097 slab_lock(page); 4098 validate_slab(s, page, map); 4099 slab_unlock(page); 4100 } 4101 4102 static int validate_slab_node(struct kmem_cache *s, 4103 struct kmem_cache_node *n, unsigned long *map) 4104 { 4105 unsigned long count = 0; 4106 struct page *page; 4107 unsigned long flags; 4108 4109 spin_lock_irqsave(&n->list_lock, flags); 4110 4111 list_for_each_entry(page, &n->partial, lru) { 4112 validate_slab_slab(s, page, map); 4113 count++; 4114 } 4115 if (count != n->nr_partial) 4116 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 4117 "counter=%ld\n", s->name, count, n->nr_partial); 4118 4119 if (!(s->flags & SLAB_STORE_USER)) 4120 goto out; 4121 4122 list_for_each_entry(page, &n->full, lru) { 4123 validate_slab_slab(s, page, map); 4124 count++; 4125 } 4126 if (count != atomic_long_read(&n->nr_slabs)) 4127 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 4128 "counter=%ld\n", s->name, count, 4129 atomic_long_read(&n->nr_slabs)); 4130 4131 out: 4132 spin_unlock_irqrestore(&n->list_lock, flags); 4133 return count; 4134 } 4135 4136 static long validate_slab_cache(struct kmem_cache *s) 4137 { 4138 int node; 4139 unsigned long count = 0; 4140 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 4141 sizeof(unsigned long), GFP_KERNEL); 4142 4143 if (!map) 4144 return -ENOMEM; 4145 4146 flush_all(s); 4147 for_each_node_state(node, N_NORMAL_MEMORY) { 4148 struct kmem_cache_node *n = get_node(s, node); 4149 4150 count += validate_slab_node(s, n, map); 4151 } 4152 kfree(map); 4153 return count; 4154 } 4155 /* 4156 * Generate lists of code addresses where slabcache objects are allocated 4157 * and freed. 4158 */ 4159 4160 struct location { 4161 unsigned long count; 4162 unsigned long addr; 4163 long long sum_time; 4164 long min_time; 4165 long max_time; 4166 long min_pid; 4167 long max_pid; 4168 DECLARE_BITMAP(cpus, NR_CPUS); 4169 nodemask_t nodes; 4170 }; 4171 4172 struct loc_track { 4173 unsigned long max; 4174 unsigned long count; 4175 struct location *loc; 4176 }; 4177 4178 static void free_loc_track(struct loc_track *t) 4179 { 4180 if (t->max) 4181 free_pages((unsigned long)t->loc, 4182 get_order(sizeof(struct location) * t->max)); 4183 } 4184 4185 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 4186 { 4187 struct location *l; 4188 int order; 4189 4190 order = get_order(sizeof(struct location) * max); 4191 4192 l = (void *)__get_free_pages(flags, order); 4193 if (!l) 4194 return 0; 4195 4196 if (t->count) { 4197 memcpy(l, t->loc, sizeof(struct location) * t->count); 4198 free_loc_track(t); 4199 } 4200 t->max = max; 4201 t->loc = l; 4202 return 1; 4203 } 4204 4205 static int add_location(struct loc_track *t, struct kmem_cache *s, 4206 const struct track *track) 4207 { 4208 long start, end, pos; 4209 struct location *l; 4210 unsigned long caddr; 4211 unsigned long age = jiffies - track->when; 4212 4213 start = -1; 4214 end = t->count; 4215 4216 for ( ; ; ) { 4217 pos = start + (end - start + 1) / 2; 4218 4219 /* 4220 * There is nothing at "end". If we end up there 4221 * we need to add something to before end. 4222 */ 4223 if (pos == end) 4224 break; 4225 4226 caddr = t->loc[pos].addr; 4227 if (track->addr == caddr) { 4228 4229 l = &t->loc[pos]; 4230 l->count++; 4231 if (track->when) { 4232 l->sum_time += age; 4233 if (age < l->min_time) 4234 l->min_time = age; 4235 if (age > l->max_time) 4236 l->max_time = age; 4237 4238 if (track->pid < l->min_pid) 4239 l->min_pid = track->pid; 4240 if (track->pid > l->max_pid) 4241 l->max_pid = track->pid; 4242 4243 cpumask_set_cpu(track->cpu, 4244 to_cpumask(l->cpus)); 4245 } 4246 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4247 return 1; 4248 } 4249 4250 if (track->addr < caddr) 4251 end = pos; 4252 else 4253 start = pos; 4254 } 4255 4256 /* 4257 * Not found. Insert new tracking element. 4258 */ 4259 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 4260 return 0; 4261 4262 l = t->loc + pos; 4263 if (pos < t->count) 4264 memmove(l + 1, l, 4265 (t->count - pos) * sizeof(struct location)); 4266 t->count++; 4267 l->count = 1; 4268 l->addr = track->addr; 4269 l->sum_time = age; 4270 l->min_time = age; 4271 l->max_time = age; 4272 l->min_pid = track->pid; 4273 l->max_pid = track->pid; 4274 cpumask_clear(to_cpumask(l->cpus)); 4275 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 4276 nodes_clear(l->nodes); 4277 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4278 return 1; 4279 } 4280 4281 static void process_slab(struct loc_track *t, struct kmem_cache *s, 4282 struct page *page, enum track_item alloc, 4283 unsigned long *map) 4284 { 4285 void *addr = page_address(page); 4286 void *p; 4287 4288 bitmap_zero(map, page->objects); 4289 get_map(s, page, map); 4290 4291 for_each_object(p, s, addr, page->objects) 4292 if (!test_bit(slab_index(p, s, addr), map)) 4293 add_location(t, s, get_track(s, p, alloc)); 4294 } 4295 4296 static int list_locations(struct kmem_cache *s, char *buf, 4297 enum track_item alloc) 4298 { 4299 int len = 0; 4300 unsigned long i; 4301 struct loc_track t = { 0, 0, NULL }; 4302 int node; 4303 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 4304 sizeof(unsigned long), GFP_KERNEL); 4305 4306 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 4307 GFP_TEMPORARY)) { 4308 kfree(map); 4309 return sprintf(buf, "Out of memory\n"); 4310 } 4311 /* Push back cpu slabs */ 4312 flush_all(s); 4313 4314 for_each_node_state(node, N_NORMAL_MEMORY) { 4315 struct kmem_cache_node *n = get_node(s, node); 4316 unsigned long flags; 4317 struct page *page; 4318 4319 if (!atomic_long_read(&n->nr_slabs)) 4320 continue; 4321 4322 spin_lock_irqsave(&n->list_lock, flags); 4323 list_for_each_entry(page, &n->partial, lru) 4324 process_slab(&t, s, page, alloc, map); 4325 list_for_each_entry(page, &n->full, lru) 4326 process_slab(&t, s, page, alloc, map); 4327 spin_unlock_irqrestore(&n->list_lock, flags); 4328 } 4329 4330 for (i = 0; i < t.count; i++) { 4331 struct location *l = &t.loc[i]; 4332 4333 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 4334 break; 4335 len += sprintf(buf + len, "%7ld ", l->count); 4336 4337 if (l->addr) 4338 len += sprintf(buf + len, "%pS", (void *)l->addr); 4339 else 4340 len += sprintf(buf + len, "<not-available>"); 4341 4342 if (l->sum_time != l->min_time) { 4343 len += sprintf(buf + len, " age=%ld/%ld/%ld", 4344 l->min_time, 4345 (long)div_u64(l->sum_time, l->count), 4346 l->max_time); 4347 } else 4348 len += sprintf(buf + len, " age=%ld", 4349 l->min_time); 4350 4351 if (l->min_pid != l->max_pid) 4352 len += sprintf(buf + len, " pid=%ld-%ld", 4353 l->min_pid, l->max_pid); 4354 else 4355 len += sprintf(buf + len, " pid=%ld", 4356 l->min_pid); 4357 4358 if (num_online_cpus() > 1 && 4359 !cpumask_empty(to_cpumask(l->cpus)) && 4360 len < PAGE_SIZE - 60) { 4361 len += sprintf(buf + len, " cpus="); 4362 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50, 4363 to_cpumask(l->cpus)); 4364 } 4365 4366 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 4367 len < PAGE_SIZE - 60) { 4368 len += sprintf(buf + len, " nodes="); 4369 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50, 4370 l->nodes); 4371 } 4372 4373 len += sprintf(buf + len, "\n"); 4374 } 4375 4376 free_loc_track(&t); 4377 kfree(map); 4378 if (!t.count) 4379 len += sprintf(buf, "No data\n"); 4380 return len; 4381 } 4382 #endif 4383 4384 #ifdef SLUB_RESILIENCY_TEST 4385 static void resiliency_test(void) 4386 { 4387 u8 *p; 4388 4389 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10); 4390 4391 printk(KERN_ERR "SLUB resiliency testing\n"); 4392 printk(KERN_ERR "-----------------------\n"); 4393 printk(KERN_ERR "A. Corruption after allocation\n"); 4394 4395 p = kzalloc(16, GFP_KERNEL); 4396 p[16] = 0x12; 4397 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 4398 " 0x12->0x%p\n\n", p + 16); 4399 4400 validate_slab_cache(kmalloc_caches[4]); 4401 4402 /* Hmmm... The next two are dangerous */ 4403 p = kzalloc(32, GFP_KERNEL); 4404 p[32 + sizeof(void *)] = 0x34; 4405 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 4406 " 0x34 -> -0x%p\n", p); 4407 printk(KERN_ERR 4408 "If allocated object is overwritten then not detectable\n\n"); 4409 4410 validate_slab_cache(kmalloc_caches[5]); 4411 p = kzalloc(64, GFP_KERNEL); 4412 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 4413 *p = 0x56; 4414 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 4415 p); 4416 printk(KERN_ERR 4417 "If allocated object is overwritten then not detectable\n\n"); 4418 validate_slab_cache(kmalloc_caches[6]); 4419 4420 printk(KERN_ERR "\nB. Corruption after free\n"); 4421 p = kzalloc(128, GFP_KERNEL); 4422 kfree(p); 4423 *p = 0x78; 4424 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 4425 validate_slab_cache(kmalloc_caches[7]); 4426 4427 p = kzalloc(256, GFP_KERNEL); 4428 kfree(p); 4429 p[50] = 0x9a; 4430 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 4431 p); 4432 validate_slab_cache(kmalloc_caches[8]); 4433 4434 p = kzalloc(512, GFP_KERNEL); 4435 kfree(p); 4436 p[512] = 0xab; 4437 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 4438 validate_slab_cache(kmalloc_caches[9]); 4439 } 4440 #else 4441 #ifdef CONFIG_SYSFS 4442 static void resiliency_test(void) {}; 4443 #endif 4444 #endif 4445 4446 #ifdef CONFIG_SYSFS 4447 enum slab_stat_type { 4448 SL_ALL, /* All slabs */ 4449 SL_PARTIAL, /* Only partially allocated slabs */ 4450 SL_CPU, /* Only slabs used for cpu caches */ 4451 SL_OBJECTS, /* Determine allocated objects not slabs */ 4452 SL_TOTAL /* Determine object capacity not slabs */ 4453 }; 4454 4455 #define SO_ALL (1 << SL_ALL) 4456 #define SO_PARTIAL (1 << SL_PARTIAL) 4457 #define SO_CPU (1 << SL_CPU) 4458 #define SO_OBJECTS (1 << SL_OBJECTS) 4459 #define SO_TOTAL (1 << SL_TOTAL) 4460 4461 static ssize_t show_slab_objects(struct kmem_cache *s, 4462 char *buf, unsigned long flags) 4463 { 4464 unsigned long total = 0; 4465 int node; 4466 int x; 4467 unsigned long *nodes; 4468 unsigned long *per_cpu; 4469 4470 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 4471 if (!nodes) 4472 return -ENOMEM; 4473 per_cpu = nodes + nr_node_ids; 4474 4475 if (flags & SO_CPU) { 4476 int cpu; 4477 4478 for_each_possible_cpu(cpu) { 4479 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 4480 int node = ACCESS_ONCE(c->node); 4481 struct page *page; 4482 4483 if (node < 0) 4484 continue; 4485 page = ACCESS_ONCE(c->page); 4486 if (page) { 4487 if (flags & SO_TOTAL) 4488 x = page->objects; 4489 else if (flags & SO_OBJECTS) 4490 x = page->inuse; 4491 else 4492 x = 1; 4493 4494 total += x; 4495 nodes[node] += x; 4496 } 4497 page = c->partial; 4498 4499 if (page) { 4500 x = page->pobjects; 4501 total += x; 4502 nodes[node] += x; 4503 } 4504 per_cpu[node]++; 4505 } 4506 } 4507 4508 lock_memory_hotplug(); 4509 #ifdef CONFIG_SLUB_DEBUG 4510 if (flags & SO_ALL) { 4511 for_each_node_state(node, N_NORMAL_MEMORY) { 4512 struct kmem_cache_node *n = get_node(s, node); 4513 4514 if (flags & SO_TOTAL) 4515 x = atomic_long_read(&n->total_objects); 4516 else if (flags & SO_OBJECTS) 4517 x = atomic_long_read(&n->total_objects) - 4518 count_partial(n, count_free); 4519 4520 else 4521 x = atomic_long_read(&n->nr_slabs); 4522 total += x; 4523 nodes[node] += x; 4524 } 4525 4526 } else 4527 #endif 4528 if (flags & SO_PARTIAL) { 4529 for_each_node_state(node, N_NORMAL_MEMORY) { 4530 struct kmem_cache_node *n = get_node(s, node); 4531 4532 if (flags & SO_TOTAL) 4533 x = count_partial(n, count_total); 4534 else if (flags & SO_OBJECTS) 4535 x = count_partial(n, count_inuse); 4536 else 4537 x = n->nr_partial; 4538 total += x; 4539 nodes[node] += x; 4540 } 4541 } 4542 x = sprintf(buf, "%lu", total); 4543 #ifdef CONFIG_NUMA 4544 for_each_node_state(node, N_NORMAL_MEMORY) 4545 if (nodes[node]) 4546 x += sprintf(buf + x, " N%d=%lu", 4547 node, nodes[node]); 4548 #endif 4549 unlock_memory_hotplug(); 4550 kfree(nodes); 4551 return x + sprintf(buf + x, "\n"); 4552 } 4553 4554 #ifdef CONFIG_SLUB_DEBUG 4555 static int any_slab_objects(struct kmem_cache *s) 4556 { 4557 int node; 4558 4559 for_each_online_node(node) { 4560 struct kmem_cache_node *n = get_node(s, node); 4561 4562 if (!n) 4563 continue; 4564 4565 if (atomic_long_read(&n->total_objects)) 4566 return 1; 4567 } 4568 return 0; 4569 } 4570 #endif 4571 4572 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 4573 #define to_slab(n) container_of(n, struct kmem_cache, kobj) 4574 4575 struct slab_attribute { 4576 struct attribute attr; 4577 ssize_t (*show)(struct kmem_cache *s, char *buf); 4578 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 4579 }; 4580 4581 #define SLAB_ATTR_RO(_name) \ 4582 static struct slab_attribute _name##_attr = \ 4583 __ATTR(_name, 0400, _name##_show, NULL) 4584 4585 #define SLAB_ATTR(_name) \ 4586 static struct slab_attribute _name##_attr = \ 4587 __ATTR(_name, 0600, _name##_show, _name##_store) 4588 4589 static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 4590 { 4591 return sprintf(buf, "%d\n", s->size); 4592 } 4593 SLAB_ATTR_RO(slab_size); 4594 4595 static ssize_t align_show(struct kmem_cache *s, char *buf) 4596 { 4597 return sprintf(buf, "%d\n", s->align); 4598 } 4599 SLAB_ATTR_RO(align); 4600 4601 static ssize_t object_size_show(struct kmem_cache *s, char *buf) 4602 { 4603 return sprintf(buf, "%d\n", s->objsize); 4604 } 4605 SLAB_ATTR_RO(object_size); 4606 4607 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 4608 { 4609 return sprintf(buf, "%d\n", oo_objects(s->oo)); 4610 } 4611 SLAB_ATTR_RO(objs_per_slab); 4612 4613 static ssize_t order_store(struct kmem_cache *s, 4614 const char *buf, size_t length) 4615 { 4616 unsigned long order; 4617 int err; 4618 4619 err = strict_strtoul(buf, 10, &order); 4620 if (err) 4621 return err; 4622 4623 if (order > slub_max_order || order < slub_min_order) 4624 return -EINVAL; 4625 4626 calculate_sizes(s, order); 4627 return length; 4628 } 4629 4630 static ssize_t order_show(struct kmem_cache *s, char *buf) 4631 { 4632 return sprintf(buf, "%d\n", oo_order(s->oo)); 4633 } 4634 SLAB_ATTR(order); 4635 4636 static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 4637 { 4638 return sprintf(buf, "%lu\n", s->min_partial); 4639 } 4640 4641 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 4642 size_t length) 4643 { 4644 unsigned long min; 4645 int err; 4646 4647 err = strict_strtoul(buf, 10, &min); 4648 if (err) 4649 return err; 4650 4651 set_min_partial(s, min); 4652 return length; 4653 } 4654 SLAB_ATTR(min_partial); 4655 4656 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) 4657 { 4658 return sprintf(buf, "%u\n", s->cpu_partial); 4659 } 4660 4661 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, 4662 size_t length) 4663 { 4664 unsigned long objects; 4665 int err; 4666 4667 err = strict_strtoul(buf, 10, &objects); 4668 if (err) 4669 return err; 4670 if (objects && kmem_cache_debug(s)) 4671 return -EINVAL; 4672 4673 s->cpu_partial = objects; 4674 flush_all(s); 4675 return length; 4676 } 4677 SLAB_ATTR(cpu_partial); 4678 4679 static ssize_t ctor_show(struct kmem_cache *s, char *buf) 4680 { 4681 if (!s->ctor) 4682 return 0; 4683 return sprintf(buf, "%pS\n", s->ctor); 4684 } 4685 SLAB_ATTR_RO(ctor); 4686 4687 static ssize_t aliases_show(struct kmem_cache *s, char *buf) 4688 { 4689 return sprintf(buf, "%d\n", s->refcount - 1); 4690 } 4691 SLAB_ATTR_RO(aliases); 4692 4693 static ssize_t partial_show(struct kmem_cache *s, char *buf) 4694 { 4695 return show_slab_objects(s, buf, SO_PARTIAL); 4696 } 4697 SLAB_ATTR_RO(partial); 4698 4699 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4700 { 4701 return show_slab_objects(s, buf, SO_CPU); 4702 } 4703 SLAB_ATTR_RO(cpu_slabs); 4704 4705 static ssize_t objects_show(struct kmem_cache *s, char *buf) 4706 { 4707 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4708 } 4709 SLAB_ATTR_RO(objects); 4710 4711 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4712 { 4713 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4714 } 4715 SLAB_ATTR_RO(objects_partial); 4716 4717 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) 4718 { 4719 int objects = 0; 4720 int pages = 0; 4721 int cpu; 4722 int len; 4723 4724 for_each_online_cpu(cpu) { 4725 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial; 4726 4727 if (page) { 4728 pages += page->pages; 4729 objects += page->pobjects; 4730 } 4731 } 4732 4733 len = sprintf(buf, "%d(%d)", objects, pages); 4734 4735 #ifdef CONFIG_SMP 4736 for_each_online_cpu(cpu) { 4737 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial; 4738 4739 if (page && len < PAGE_SIZE - 20) 4740 len += sprintf(buf + len, " C%d=%d(%d)", cpu, 4741 page->pobjects, page->pages); 4742 } 4743 #endif 4744 return len + sprintf(buf + len, "\n"); 4745 } 4746 SLAB_ATTR_RO(slabs_cpu_partial); 4747 4748 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4749 { 4750 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4751 } 4752 4753 static ssize_t reclaim_account_store(struct kmem_cache *s, 4754 const char *buf, size_t length) 4755 { 4756 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4757 if (buf[0] == '1') 4758 s->flags |= SLAB_RECLAIM_ACCOUNT; 4759 return length; 4760 } 4761 SLAB_ATTR(reclaim_account); 4762 4763 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4764 { 4765 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4766 } 4767 SLAB_ATTR_RO(hwcache_align); 4768 4769 #ifdef CONFIG_ZONE_DMA 4770 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4771 { 4772 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4773 } 4774 SLAB_ATTR_RO(cache_dma); 4775 #endif 4776 4777 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4778 { 4779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4780 } 4781 SLAB_ATTR_RO(destroy_by_rcu); 4782 4783 static ssize_t reserved_show(struct kmem_cache *s, char *buf) 4784 { 4785 return sprintf(buf, "%d\n", s->reserved); 4786 } 4787 SLAB_ATTR_RO(reserved); 4788 4789 #ifdef CONFIG_SLUB_DEBUG 4790 static ssize_t slabs_show(struct kmem_cache *s, char *buf) 4791 { 4792 return show_slab_objects(s, buf, SO_ALL); 4793 } 4794 SLAB_ATTR_RO(slabs); 4795 4796 static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 4797 { 4798 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 4799 } 4800 SLAB_ATTR_RO(total_objects); 4801 4802 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 4803 { 4804 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 4805 } 4806 4807 static ssize_t sanity_checks_store(struct kmem_cache *s, 4808 const char *buf, size_t length) 4809 { 4810 s->flags &= ~SLAB_DEBUG_FREE; 4811 if (buf[0] == '1') { 4812 s->flags &= ~__CMPXCHG_DOUBLE; 4813 s->flags |= SLAB_DEBUG_FREE; 4814 } 4815 return length; 4816 } 4817 SLAB_ATTR(sanity_checks); 4818 4819 static ssize_t trace_show(struct kmem_cache *s, char *buf) 4820 { 4821 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4822 } 4823 4824 static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4825 size_t length) 4826 { 4827 s->flags &= ~SLAB_TRACE; 4828 if (buf[0] == '1') { 4829 s->flags &= ~__CMPXCHG_DOUBLE; 4830 s->flags |= SLAB_TRACE; 4831 } 4832 return length; 4833 } 4834 SLAB_ATTR(trace); 4835 4836 static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4837 { 4838 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4839 } 4840 4841 static ssize_t red_zone_store(struct kmem_cache *s, 4842 const char *buf, size_t length) 4843 { 4844 if (any_slab_objects(s)) 4845 return -EBUSY; 4846 4847 s->flags &= ~SLAB_RED_ZONE; 4848 if (buf[0] == '1') { 4849 s->flags &= ~__CMPXCHG_DOUBLE; 4850 s->flags |= SLAB_RED_ZONE; 4851 } 4852 calculate_sizes(s, -1); 4853 return length; 4854 } 4855 SLAB_ATTR(red_zone); 4856 4857 static ssize_t poison_show(struct kmem_cache *s, char *buf) 4858 { 4859 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4860 } 4861 4862 static ssize_t poison_store(struct kmem_cache *s, 4863 const char *buf, size_t length) 4864 { 4865 if (any_slab_objects(s)) 4866 return -EBUSY; 4867 4868 s->flags &= ~SLAB_POISON; 4869 if (buf[0] == '1') { 4870 s->flags &= ~__CMPXCHG_DOUBLE; 4871 s->flags |= SLAB_POISON; 4872 } 4873 calculate_sizes(s, -1); 4874 return length; 4875 } 4876 SLAB_ATTR(poison); 4877 4878 static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4879 { 4880 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4881 } 4882 4883 static ssize_t store_user_store(struct kmem_cache *s, 4884 const char *buf, size_t length) 4885 { 4886 if (any_slab_objects(s)) 4887 return -EBUSY; 4888 4889 s->flags &= ~SLAB_STORE_USER; 4890 if (buf[0] == '1') { 4891 s->flags &= ~__CMPXCHG_DOUBLE; 4892 s->flags |= SLAB_STORE_USER; 4893 } 4894 calculate_sizes(s, -1); 4895 return length; 4896 } 4897 SLAB_ATTR(store_user); 4898 4899 static ssize_t validate_show(struct kmem_cache *s, char *buf) 4900 { 4901 return 0; 4902 } 4903 4904 static ssize_t validate_store(struct kmem_cache *s, 4905 const char *buf, size_t length) 4906 { 4907 int ret = -EINVAL; 4908 4909 if (buf[0] == '1') { 4910 ret = validate_slab_cache(s); 4911 if (ret >= 0) 4912 ret = length; 4913 } 4914 return ret; 4915 } 4916 SLAB_ATTR(validate); 4917 4918 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4919 { 4920 if (!(s->flags & SLAB_STORE_USER)) 4921 return -ENOSYS; 4922 return list_locations(s, buf, TRACK_ALLOC); 4923 } 4924 SLAB_ATTR_RO(alloc_calls); 4925 4926 static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4927 { 4928 if (!(s->flags & SLAB_STORE_USER)) 4929 return -ENOSYS; 4930 return list_locations(s, buf, TRACK_FREE); 4931 } 4932 SLAB_ATTR_RO(free_calls); 4933 #endif /* CONFIG_SLUB_DEBUG */ 4934 4935 #ifdef CONFIG_FAILSLAB 4936 static ssize_t failslab_show(struct kmem_cache *s, char *buf) 4937 { 4938 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 4939 } 4940 4941 static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 4942 size_t length) 4943 { 4944 s->flags &= ~SLAB_FAILSLAB; 4945 if (buf[0] == '1') 4946 s->flags |= SLAB_FAILSLAB; 4947 return length; 4948 } 4949 SLAB_ATTR(failslab); 4950 #endif 4951 4952 static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4953 { 4954 return 0; 4955 } 4956 4957 static ssize_t shrink_store(struct kmem_cache *s, 4958 const char *buf, size_t length) 4959 { 4960 if (buf[0] == '1') { 4961 int rc = kmem_cache_shrink(s); 4962 4963 if (rc) 4964 return rc; 4965 } else 4966 return -EINVAL; 4967 return length; 4968 } 4969 SLAB_ATTR(shrink); 4970 4971 #ifdef CONFIG_NUMA 4972 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4973 { 4974 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4975 } 4976 4977 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4978 const char *buf, size_t length) 4979 { 4980 unsigned long ratio; 4981 int err; 4982 4983 err = strict_strtoul(buf, 10, &ratio); 4984 if (err) 4985 return err; 4986 4987 if (ratio <= 100) 4988 s->remote_node_defrag_ratio = ratio * 10; 4989 4990 return length; 4991 } 4992 SLAB_ATTR(remote_node_defrag_ratio); 4993 #endif 4994 4995 #ifdef CONFIG_SLUB_STATS 4996 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4997 { 4998 unsigned long sum = 0; 4999 int cpu; 5000 int len; 5001 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 5002 5003 if (!data) 5004 return -ENOMEM; 5005 5006 for_each_online_cpu(cpu) { 5007 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 5008 5009 data[cpu] = x; 5010 sum += x; 5011 } 5012 5013 len = sprintf(buf, "%lu", sum); 5014 5015 #ifdef CONFIG_SMP 5016 for_each_online_cpu(cpu) { 5017 if (data[cpu] && len < PAGE_SIZE - 20) 5018 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 5019 } 5020 #endif 5021 kfree(data); 5022 return len + sprintf(buf + len, "\n"); 5023 } 5024 5025 static void clear_stat(struct kmem_cache *s, enum stat_item si) 5026 { 5027 int cpu; 5028 5029 for_each_online_cpu(cpu) 5030 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 5031 } 5032 5033 #define STAT_ATTR(si, text) \ 5034 static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 5035 { \ 5036 return show_stat(s, buf, si); \ 5037 } \ 5038 static ssize_t text##_store(struct kmem_cache *s, \ 5039 const char *buf, size_t length) \ 5040 { \ 5041 if (buf[0] != '0') \ 5042 return -EINVAL; \ 5043 clear_stat(s, si); \ 5044 return length; \ 5045 } \ 5046 SLAB_ATTR(text); \ 5047 5048 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 5049 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 5050 STAT_ATTR(FREE_FASTPATH, free_fastpath); 5051 STAT_ATTR(FREE_SLOWPATH, free_slowpath); 5052 STAT_ATTR(FREE_FROZEN, free_frozen); 5053 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 5054 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 5055 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 5056 STAT_ATTR(ALLOC_SLAB, alloc_slab); 5057 STAT_ATTR(ALLOC_REFILL, alloc_refill); 5058 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); 5059 STAT_ATTR(FREE_SLAB, free_slab); 5060 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 5061 STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 5062 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 5063 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 5064 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 5065 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 5066 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); 5067 STAT_ATTR(ORDER_FALLBACK, order_fallback); 5068 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); 5069 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); 5070 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); 5071 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); 5072 #endif 5073 5074 static struct attribute *slab_attrs[] = { 5075 &slab_size_attr.attr, 5076 &object_size_attr.attr, 5077 &objs_per_slab_attr.attr, 5078 &order_attr.attr, 5079 &min_partial_attr.attr, 5080 &cpu_partial_attr.attr, 5081 &objects_attr.attr, 5082 &objects_partial_attr.attr, 5083 &partial_attr.attr, 5084 &cpu_slabs_attr.attr, 5085 &ctor_attr.attr, 5086 &aliases_attr.attr, 5087 &align_attr.attr, 5088 &hwcache_align_attr.attr, 5089 &reclaim_account_attr.attr, 5090 &destroy_by_rcu_attr.attr, 5091 &shrink_attr.attr, 5092 &reserved_attr.attr, 5093 &slabs_cpu_partial_attr.attr, 5094 #ifdef CONFIG_SLUB_DEBUG 5095 &total_objects_attr.attr, 5096 &slabs_attr.attr, 5097 &sanity_checks_attr.attr, 5098 &trace_attr.attr, 5099 &red_zone_attr.attr, 5100 &poison_attr.attr, 5101 &store_user_attr.attr, 5102 &validate_attr.attr, 5103 &alloc_calls_attr.attr, 5104 &free_calls_attr.attr, 5105 #endif 5106 #ifdef CONFIG_ZONE_DMA 5107 &cache_dma_attr.attr, 5108 #endif 5109 #ifdef CONFIG_NUMA 5110 &remote_node_defrag_ratio_attr.attr, 5111 #endif 5112 #ifdef CONFIG_SLUB_STATS 5113 &alloc_fastpath_attr.attr, 5114 &alloc_slowpath_attr.attr, 5115 &free_fastpath_attr.attr, 5116 &free_slowpath_attr.attr, 5117 &free_frozen_attr.attr, 5118 &free_add_partial_attr.attr, 5119 &free_remove_partial_attr.attr, 5120 &alloc_from_partial_attr.attr, 5121 &alloc_slab_attr.attr, 5122 &alloc_refill_attr.attr, 5123 &alloc_node_mismatch_attr.attr, 5124 &free_slab_attr.attr, 5125 &cpuslab_flush_attr.attr, 5126 &deactivate_full_attr.attr, 5127 &deactivate_empty_attr.attr, 5128 &deactivate_to_head_attr.attr, 5129 &deactivate_to_tail_attr.attr, 5130 &deactivate_remote_frees_attr.attr, 5131 &deactivate_bypass_attr.attr, 5132 &order_fallback_attr.attr, 5133 &cmpxchg_double_fail_attr.attr, 5134 &cmpxchg_double_cpu_fail_attr.attr, 5135 &cpu_partial_alloc_attr.attr, 5136 &cpu_partial_free_attr.attr, 5137 #endif 5138 #ifdef CONFIG_FAILSLAB 5139 &failslab_attr.attr, 5140 #endif 5141 5142 NULL 5143 }; 5144 5145 static struct attribute_group slab_attr_group = { 5146 .attrs = slab_attrs, 5147 }; 5148 5149 static ssize_t slab_attr_show(struct kobject *kobj, 5150 struct attribute *attr, 5151 char *buf) 5152 { 5153 struct slab_attribute *attribute; 5154 struct kmem_cache *s; 5155 int err; 5156 5157 attribute = to_slab_attr(attr); 5158 s = to_slab(kobj); 5159 5160 if (!attribute->show) 5161 return -EIO; 5162 5163 err = attribute->show(s, buf); 5164 5165 return err; 5166 } 5167 5168 static ssize_t slab_attr_store(struct kobject *kobj, 5169 struct attribute *attr, 5170 const char *buf, size_t len) 5171 { 5172 struct slab_attribute *attribute; 5173 struct kmem_cache *s; 5174 int err; 5175 5176 attribute = to_slab_attr(attr); 5177 s = to_slab(kobj); 5178 5179 if (!attribute->store) 5180 return -EIO; 5181 5182 err = attribute->store(s, buf, len); 5183 5184 return err; 5185 } 5186 5187 static void kmem_cache_release(struct kobject *kobj) 5188 { 5189 struct kmem_cache *s = to_slab(kobj); 5190 5191 kfree(s->name); 5192 kfree(s); 5193 } 5194 5195 static const struct sysfs_ops slab_sysfs_ops = { 5196 .show = slab_attr_show, 5197 .store = slab_attr_store, 5198 }; 5199 5200 static struct kobj_type slab_ktype = { 5201 .sysfs_ops = &slab_sysfs_ops, 5202 .release = kmem_cache_release 5203 }; 5204 5205 static int uevent_filter(struct kset *kset, struct kobject *kobj) 5206 { 5207 struct kobj_type *ktype = get_ktype(kobj); 5208 5209 if (ktype == &slab_ktype) 5210 return 1; 5211 return 0; 5212 } 5213 5214 static const struct kset_uevent_ops slab_uevent_ops = { 5215 .filter = uevent_filter, 5216 }; 5217 5218 static struct kset *slab_kset; 5219 5220 #define ID_STR_LENGTH 64 5221 5222 /* Create a unique string id for a slab cache: 5223 * 5224 * Format :[flags-]size 5225 */ 5226 static char *create_unique_id(struct kmem_cache *s) 5227 { 5228 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 5229 char *p = name; 5230 5231 BUG_ON(!name); 5232 5233 *p++ = ':'; 5234 /* 5235 * First flags affecting slabcache operations. We will only 5236 * get here for aliasable slabs so we do not need to support 5237 * too many flags. The flags here must cover all flags that 5238 * are matched during merging to guarantee that the id is 5239 * unique. 5240 */ 5241 if (s->flags & SLAB_CACHE_DMA) 5242 *p++ = 'd'; 5243 if (s->flags & SLAB_RECLAIM_ACCOUNT) 5244 *p++ = 'a'; 5245 if (s->flags & SLAB_DEBUG_FREE) 5246 *p++ = 'F'; 5247 if (!(s->flags & SLAB_NOTRACK)) 5248 *p++ = 't'; 5249 if (p != name + 1) 5250 *p++ = '-'; 5251 p += sprintf(p, "%07d", s->size); 5252 BUG_ON(p > name + ID_STR_LENGTH - 1); 5253 return name; 5254 } 5255 5256 static int sysfs_slab_add(struct kmem_cache *s) 5257 { 5258 int err; 5259 const char *name; 5260 int unmergeable; 5261 5262 if (slab_state < SYSFS) 5263 /* Defer until later */ 5264 return 0; 5265 5266 unmergeable = slab_unmergeable(s); 5267 if (unmergeable) { 5268 /* 5269 * Slabcache can never be merged so we can use the name proper. 5270 * This is typically the case for debug situations. In that 5271 * case we can catch duplicate names easily. 5272 */ 5273 sysfs_remove_link(&slab_kset->kobj, s->name); 5274 name = s->name; 5275 } else { 5276 /* 5277 * Create a unique name for the slab as a target 5278 * for the symlinks. 5279 */ 5280 name = create_unique_id(s); 5281 } 5282 5283 s->kobj.kset = slab_kset; 5284 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name); 5285 if (err) { 5286 kobject_put(&s->kobj); 5287 return err; 5288 } 5289 5290 err = sysfs_create_group(&s->kobj, &slab_attr_group); 5291 if (err) { 5292 kobject_del(&s->kobj); 5293 kobject_put(&s->kobj); 5294 return err; 5295 } 5296 kobject_uevent(&s->kobj, KOBJ_ADD); 5297 if (!unmergeable) { 5298 /* Setup first alias */ 5299 sysfs_slab_alias(s, s->name); 5300 kfree(name); 5301 } 5302 return 0; 5303 } 5304 5305 static void sysfs_slab_remove(struct kmem_cache *s) 5306 { 5307 if (slab_state < SYSFS) 5308 /* 5309 * Sysfs has not been setup yet so no need to remove the 5310 * cache from sysfs. 5311 */ 5312 return; 5313 5314 kobject_uevent(&s->kobj, KOBJ_REMOVE); 5315 kobject_del(&s->kobj); 5316 kobject_put(&s->kobj); 5317 } 5318 5319 /* 5320 * Need to buffer aliases during bootup until sysfs becomes 5321 * available lest we lose that information. 5322 */ 5323 struct saved_alias { 5324 struct kmem_cache *s; 5325 const char *name; 5326 struct saved_alias *next; 5327 }; 5328 5329 static struct saved_alias *alias_list; 5330 5331 static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 5332 { 5333 struct saved_alias *al; 5334 5335 if (slab_state == SYSFS) { 5336 /* 5337 * If we have a leftover link then remove it. 5338 */ 5339 sysfs_remove_link(&slab_kset->kobj, name); 5340 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 5341 } 5342 5343 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 5344 if (!al) 5345 return -ENOMEM; 5346 5347 al->s = s; 5348 al->name = name; 5349 al->next = alias_list; 5350 alias_list = al; 5351 return 0; 5352 } 5353 5354 static int __init slab_sysfs_init(void) 5355 { 5356 struct kmem_cache *s; 5357 int err; 5358 5359 down_write(&slub_lock); 5360 5361 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 5362 if (!slab_kset) { 5363 up_write(&slub_lock); 5364 printk(KERN_ERR "Cannot register slab subsystem.\n"); 5365 return -ENOSYS; 5366 } 5367 5368 slab_state = SYSFS; 5369 5370 list_for_each_entry(s, &slab_caches, list) { 5371 err = sysfs_slab_add(s); 5372 if (err) 5373 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 5374 " to sysfs\n", s->name); 5375 } 5376 5377 while (alias_list) { 5378 struct saved_alias *al = alias_list; 5379 5380 alias_list = alias_list->next; 5381 err = sysfs_slab_alias(al->s, al->name); 5382 if (err) 5383 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 5384 " %s to sysfs\n", s->name); 5385 kfree(al); 5386 } 5387 5388 up_write(&slub_lock); 5389 resiliency_test(); 5390 return 0; 5391 } 5392 5393 __initcall(slab_sysfs_init); 5394 #endif /* CONFIG_SYSFS */ 5395 5396 /* 5397 * The /proc/slabinfo ABI 5398 */ 5399 #ifdef CONFIG_SLABINFO 5400 static void print_slabinfo_header(struct seq_file *m) 5401 { 5402 seq_puts(m, "slabinfo - version: 2.1\n"); 5403 seq_puts(m, "# name <active_objs> <num_objs> <objsize> " 5404 "<objperslab> <pagesperslab>"); 5405 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); 5406 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); 5407 seq_putc(m, '\n'); 5408 } 5409 5410 static void *s_start(struct seq_file *m, loff_t *pos) 5411 { 5412 loff_t n = *pos; 5413 5414 down_read(&slub_lock); 5415 if (!n) 5416 print_slabinfo_header(m); 5417 5418 return seq_list_start(&slab_caches, *pos); 5419 } 5420 5421 static void *s_next(struct seq_file *m, void *p, loff_t *pos) 5422 { 5423 return seq_list_next(p, &slab_caches, pos); 5424 } 5425 5426 static void s_stop(struct seq_file *m, void *p) 5427 { 5428 up_read(&slub_lock); 5429 } 5430 5431 static int s_show(struct seq_file *m, void *p) 5432 { 5433 unsigned long nr_partials = 0; 5434 unsigned long nr_slabs = 0; 5435 unsigned long nr_inuse = 0; 5436 unsigned long nr_objs = 0; 5437 unsigned long nr_free = 0; 5438 struct kmem_cache *s; 5439 int node; 5440 5441 s = list_entry(p, struct kmem_cache, list); 5442 5443 for_each_online_node(node) { 5444 struct kmem_cache_node *n = get_node(s, node); 5445 5446 if (!n) 5447 continue; 5448 5449 nr_partials += n->nr_partial; 5450 nr_slabs += atomic_long_read(&n->nr_slabs); 5451 nr_objs += atomic_long_read(&n->total_objects); 5452 nr_free += count_partial(n, count_free); 5453 } 5454 5455 nr_inuse = nr_objs - nr_free; 5456 5457 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse, 5458 nr_objs, s->size, oo_objects(s->oo), 5459 (1 << oo_order(s->oo))); 5460 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0); 5461 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs, 5462 0UL); 5463 seq_putc(m, '\n'); 5464 return 0; 5465 } 5466 5467 static const struct seq_operations slabinfo_op = { 5468 .start = s_start, 5469 .next = s_next, 5470 .stop = s_stop, 5471 .show = s_show, 5472 }; 5473 5474 static int slabinfo_open(struct inode *inode, struct file *file) 5475 { 5476 return seq_open(file, &slabinfo_op); 5477 } 5478 5479 static const struct file_operations proc_slabinfo_operations = { 5480 .open = slabinfo_open, 5481 .read = seq_read, 5482 .llseek = seq_lseek, 5483 .release = seq_release, 5484 }; 5485 5486 static int __init slab_proc_init(void) 5487 { 5488 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations); 5489 return 0; 5490 } 5491 module_init(slab_proc_init); 5492 #endif /* CONFIG_SLABINFO */ 5493