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 "slab.h" 20 #include <linux/proc_fs.h> 21 #include <linux/notifier.h> 22 #include <linux/seq_file.h> 23 #include <linux/kmemcheck.h> 24 #include <linux/cpu.h> 25 #include <linux/cpuset.h> 26 #include <linux/mempolicy.h> 27 #include <linux/ctype.h> 28 #include <linux/debugobjects.h> 29 #include <linux/kallsyms.h> 30 #include <linux/memory.h> 31 #include <linux/math64.h> 32 #include <linux/fault-inject.h> 33 #include <linux/stacktrace.h> 34 #include <linux/prefetch.h> 35 #include <linux/memcontrol.h> 36 37 #include <trace/events/kmem.h> 38 39 #include "internal.h" 40 41 /* 42 * Lock order: 43 * 1. slab_mutex (Global Mutex) 44 * 2. node->list_lock 45 * 3. slab_lock(page) (Only on some arches and for debugging) 46 * 47 * slab_mutex 48 * 49 * The role of the slab_mutex is to protect the list of all the slabs 50 * and to synchronize major metadata changes to slab cache structures. 51 * 52 * The slab_lock is only used for debugging and on arches that do not 53 * have the ability to do a cmpxchg_double. It only protects the second 54 * double word in the page struct. Meaning 55 * A. page->freelist -> List of object free in a page 56 * B. page->counters -> Counters of objects 57 * C. page->frozen -> frozen state 58 * 59 * If a slab is frozen then it is exempt from list management. It is not 60 * on any list. The processor that froze the slab is the one who can 61 * perform list operations on the page. Other processors may put objects 62 * onto the freelist but the processor that froze the slab is the only 63 * one that can retrieve the objects from the page's freelist. 64 * 65 * The list_lock protects the partial and full list on each node and 66 * the partial slab counter. If taken then no new slabs may be added or 67 * removed from the lists nor make the number of partial slabs be modified. 68 * (Note that the total number of slabs is an atomic value that may be 69 * modified without taking the list lock). 70 * 71 * The list_lock is a centralized lock and thus we avoid taking it as 72 * much as possible. As long as SLUB does not have to handle partial 73 * slabs, operations can continue without any centralized lock. F.e. 74 * allocating a long series of objects that fill up slabs does not require 75 * the list lock. 76 * Interrupts are disabled during allocation and deallocation in order to 77 * make the slab allocator safe to use in the context of an irq. In addition 78 * interrupts are disabled to ensure that the processor does not change 79 * while handling per_cpu slabs, due to kernel preemption. 80 * 81 * SLUB assigns one slab for allocation to each processor. 82 * Allocations only occur from these slabs called cpu slabs. 83 * 84 * Slabs with free elements are kept on a partial list and during regular 85 * operations no list for full slabs is used. If an object in a full slab is 86 * freed then the slab will show up again on the partial lists. 87 * We track full slabs for debugging purposes though because otherwise we 88 * cannot scan all objects. 89 * 90 * Slabs are freed when they become empty. Teardown and setup is 91 * minimal so we rely on the page allocators per cpu caches for 92 * fast frees and allocs. 93 * 94 * Overloading of page flags that are otherwise used for LRU management. 95 * 96 * PageActive The slab is frozen and exempt from list processing. 97 * This means that the slab is dedicated to a purpose 98 * such as satisfying allocations for a specific 99 * processor. Objects may be freed in the slab while 100 * it is frozen but slab_free will then skip the usual 101 * list operations. It is up to the processor holding 102 * the slab to integrate the slab into the slab lists 103 * when the slab is no longer needed. 104 * 105 * One use of this flag is to mark slabs that are 106 * used for allocations. Then such a slab becomes a cpu 107 * slab. The cpu slab may be equipped with an additional 108 * freelist that allows lockless access to 109 * free objects in addition to the regular freelist 110 * that requires the slab lock. 111 * 112 * PageError Slab requires special handling due to debug 113 * options set. This moves slab handling out of 114 * the fast path and disables lockless freelists. 115 */ 116 117 static inline int kmem_cache_debug(struct kmem_cache *s) 118 { 119 #ifdef CONFIG_SLUB_DEBUG 120 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 121 #else 122 return 0; 123 #endif 124 } 125 126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) 127 { 128 #ifdef CONFIG_SLUB_CPU_PARTIAL 129 return !kmem_cache_debug(s); 130 #else 131 return false; 132 #endif 133 } 134 135 /* 136 * Issues still to be resolved: 137 * 138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 139 * 140 * - Variable sizing of the per node arrays 141 */ 142 143 /* Enable to test recovery from slab corruption on boot */ 144 #undef SLUB_RESILIENCY_TEST 145 146 /* Enable to log cmpxchg failures */ 147 #undef SLUB_DEBUG_CMPXCHG 148 149 /* 150 * Mininum number of partial slabs. These will be left on the partial 151 * lists even if they are empty. kmem_cache_shrink may reclaim them. 152 */ 153 #define MIN_PARTIAL 5 154 155 /* 156 * Maximum number of desirable partial slabs. 157 * The existence of more partial slabs makes kmem_cache_shrink 158 * sort the partial list by the number of objects in use. 159 */ 160 #define MAX_PARTIAL 10 161 162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \ 163 SLAB_POISON | SLAB_STORE_USER) 164 165 /* 166 * Debugging flags that require metadata to be stored in the slab. These get 167 * disabled when slub_debug=O is used and a cache's min order increases with 168 * metadata. 169 */ 170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 171 172 /* 173 * Set of flags that will prevent slab merging 174 */ 175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \ 176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \ 177 SLAB_FAILSLAB) 178 179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \ 180 SLAB_CACHE_DMA | SLAB_NOTRACK) 181 182 #define OO_SHIFT 16 183 #define OO_MASK ((1 << OO_SHIFT) - 1) 184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ 185 186 /* Internal SLUB flags */ 187 #define __OBJECT_POISON 0x80000000UL /* Poison object */ 188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */ 189 190 #ifdef CONFIG_SMP 191 static struct notifier_block slab_notifier; 192 #endif 193 194 /* 195 * Tracking user of a slab. 196 */ 197 #define TRACK_ADDRS_COUNT 16 198 struct track { 199 unsigned long addr; /* Called from address */ 200 #ifdef CONFIG_STACKTRACE 201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 202 #endif 203 int cpu; /* Was running on cpu */ 204 int pid; /* Pid context */ 205 unsigned long when; /* When did the operation occur */ 206 }; 207 208 enum track_item { TRACK_ALLOC, TRACK_FREE }; 209 210 #ifdef CONFIG_SYSFS 211 static int sysfs_slab_add(struct kmem_cache *); 212 static int sysfs_slab_alias(struct kmem_cache *, const char *); 213 static void memcg_propagate_slab_attrs(struct kmem_cache *s); 214 #else 215 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 216 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 217 { return 0; } 218 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { } 219 #endif 220 221 static inline void stat(const struct kmem_cache *s, enum stat_item si) 222 { 223 #ifdef CONFIG_SLUB_STATS 224 /* 225 * The rmw is racy on a preemptible kernel but this is acceptable, so 226 * avoid this_cpu_add()'s irq-disable overhead. 227 */ 228 raw_cpu_inc(s->cpu_slab->stat[si]); 229 #endif 230 } 231 232 /******************************************************************** 233 * Core slab cache functions 234 *******************************************************************/ 235 236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) 237 { 238 return s->node[node]; 239 } 240 241 /* Verify that a pointer has an address that is valid within a slab page */ 242 static inline int check_valid_pointer(struct kmem_cache *s, 243 struct page *page, const void *object) 244 { 245 void *base; 246 247 if (!object) 248 return 1; 249 250 base = page_address(page); 251 if (object < base || object >= base + page->objects * s->size || 252 (object - base) % s->size) { 253 return 0; 254 } 255 256 return 1; 257 } 258 259 static inline void *get_freepointer(struct kmem_cache *s, void *object) 260 { 261 return *(void **)(object + s->offset); 262 } 263 264 static void prefetch_freepointer(const struct kmem_cache *s, void *object) 265 { 266 prefetch(object + s->offset); 267 } 268 269 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 270 { 271 void *p; 272 273 #ifdef CONFIG_DEBUG_PAGEALLOC 274 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p)); 275 #else 276 p = get_freepointer(s, object); 277 #endif 278 return p; 279 } 280 281 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 282 { 283 *(void **)(object + s->offset) = fp; 284 } 285 286 /* Loop over all objects in a slab */ 287 #define for_each_object(__p, __s, __addr, __objects) \ 288 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\ 289 __p += (__s)->size) 290 291 /* Determine object index from a given position */ 292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 293 { 294 return (p - addr) / s->size; 295 } 296 297 static inline size_t slab_ksize(const struct kmem_cache *s) 298 { 299 #ifdef CONFIG_SLUB_DEBUG 300 /* 301 * Debugging requires use of the padding between object 302 * and whatever may come after it. 303 */ 304 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON)) 305 return s->object_size; 306 307 #endif 308 /* 309 * If we have the need to store the freelist pointer 310 * back there or track user information then we can 311 * only use the space before that information. 312 */ 313 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER)) 314 return s->inuse; 315 /* 316 * Else we can use all the padding etc for the allocation 317 */ 318 return s->size; 319 } 320 321 static inline int order_objects(int order, unsigned long size, int reserved) 322 { 323 return ((PAGE_SIZE << order) - reserved) / size; 324 } 325 326 static inline struct kmem_cache_order_objects oo_make(int order, 327 unsigned long size, int reserved) 328 { 329 struct kmem_cache_order_objects x = { 330 (order << OO_SHIFT) + order_objects(order, size, reserved) 331 }; 332 333 return x; 334 } 335 336 static inline int oo_order(struct kmem_cache_order_objects x) 337 { 338 return x.x >> OO_SHIFT; 339 } 340 341 static inline int oo_objects(struct kmem_cache_order_objects x) 342 { 343 return x.x & OO_MASK; 344 } 345 346 /* 347 * Per slab locking using the pagelock 348 */ 349 static __always_inline void slab_lock(struct page *page) 350 { 351 bit_spin_lock(PG_locked, &page->flags); 352 } 353 354 static __always_inline void slab_unlock(struct page *page) 355 { 356 __bit_spin_unlock(PG_locked, &page->flags); 357 } 358 359 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new) 360 { 361 struct page tmp; 362 tmp.counters = counters_new; 363 /* 364 * page->counters can cover frozen/inuse/objects as well 365 * as page->_count. If we assign to ->counters directly 366 * we run the risk of losing updates to page->_count, so 367 * be careful and only assign to the fields we need. 368 */ 369 page->frozen = tmp.frozen; 370 page->inuse = tmp.inuse; 371 page->objects = tmp.objects; 372 } 373 374 /* Interrupts must be disabled (for the fallback code to work right) */ 375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 376 void *freelist_old, unsigned long counters_old, 377 void *freelist_new, unsigned long counters_new, 378 const char *n) 379 { 380 VM_BUG_ON(!irqs_disabled()); 381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 383 if (s->flags & __CMPXCHG_DOUBLE) { 384 if (cmpxchg_double(&page->freelist, &page->counters, 385 freelist_old, counters_old, 386 freelist_new, counters_new)) 387 return 1; 388 } else 389 #endif 390 { 391 slab_lock(page); 392 if (page->freelist == freelist_old && 393 page->counters == counters_old) { 394 page->freelist = freelist_new; 395 set_page_slub_counters(page, counters_new); 396 slab_unlock(page); 397 return 1; 398 } 399 slab_unlock(page); 400 } 401 402 cpu_relax(); 403 stat(s, CMPXCHG_DOUBLE_FAIL); 404 405 #ifdef SLUB_DEBUG_CMPXCHG 406 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name); 407 #endif 408 409 return 0; 410 } 411 412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 413 void *freelist_old, unsigned long counters_old, 414 void *freelist_new, unsigned long counters_new, 415 const char *n) 416 { 417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 419 if (s->flags & __CMPXCHG_DOUBLE) { 420 if (cmpxchg_double(&page->freelist, &page->counters, 421 freelist_old, counters_old, 422 freelist_new, counters_new)) 423 return 1; 424 } else 425 #endif 426 { 427 unsigned long flags; 428 429 local_irq_save(flags); 430 slab_lock(page); 431 if (page->freelist == freelist_old && 432 page->counters == counters_old) { 433 page->freelist = freelist_new; 434 set_page_slub_counters(page, counters_new); 435 slab_unlock(page); 436 local_irq_restore(flags); 437 return 1; 438 } 439 slab_unlock(page); 440 local_irq_restore(flags); 441 } 442 443 cpu_relax(); 444 stat(s, CMPXCHG_DOUBLE_FAIL); 445 446 #ifdef SLUB_DEBUG_CMPXCHG 447 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name); 448 #endif 449 450 return 0; 451 } 452 453 #ifdef CONFIG_SLUB_DEBUG 454 /* 455 * Determine a map of object in use on a page. 456 * 457 * Node listlock must be held to guarantee that the page does 458 * not vanish from under us. 459 */ 460 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) 461 { 462 void *p; 463 void *addr = page_address(page); 464 465 for (p = page->freelist; p; p = get_freepointer(s, p)) 466 set_bit(slab_index(p, s, addr), map); 467 } 468 469 /* 470 * Debug settings: 471 */ 472 #ifdef CONFIG_SLUB_DEBUG_ON 473 static int slub_debug = DEBUG_DEFAULT_FLAGS; 474 #else 475 static int slub_debug; 476 #endif 477 478 static char *slub_debug_slabs; 479 static int disable_higher_order_debug; 480 481 /* 482 * Object debugging 483 */ 484 static void print_section(char *text, u8 *addr, unsigned int length) 485 { 486 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr, 487 length, 1); 488 } 489 490 static struct track *get_track(struct kmem_cache *s, void *object, 491 enum track_item alloc) 492 { 493 struct track *p; 494 495 if (s->offset) 496 p = object + s->offset + sizeof(void *); 497 else 498 p = object + s->inuse; 499 500 return p + alloc; 501 } 502 503 static void set_track(struct kmem_cache *s, void *object, 504 enum track_item alloc, unsigned long addr) 505 { 506 struct track *p = get_track(s, object, alloc); 507 508 if (addr) { 509 #ifdef CONFIG_STACKTRACE 510 struct stack_trace trace; 511 int i; 512 513 trace.nr_entries = 0; 514 trace.max_entries = TRACK_ADDRS_COUNT; 515 trace.entries = p->addrs; 516 trace.skip = 3; 517 save_stack_trace(&trace); 518 519 /* See rant in lockdep.c */ 520 if (trace.nr_entries != 0 && 521 trace.entries[trace.nr_entries - 1] == ULONG_MAX) 522 trace.nr_entries--; 523 524 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) 525 p->addrs[i] = 0; 526 #endif 527 p->addr = addr; 528 p->cpu = smp_processor_id(); 529 p->pid = current->pid; 530 p->when = jiffies; 531 } else 532 memset(p, 0, sizeof(struct track)); 533 } 534 535 static void init_tracking(struct kmem_cache *s, void *object) 536 { 537 if (!(s->flags & SLAB_STORE_USER)) 538 return; 539 540 set_track(s, object, TRACK_FREE, 0UL); 541 set_track(s, object, TRACK_ALLOC, 0UL); 542 } 543 544 static void print_track(const char *s, struct track *t) 545 { 546 if (!t->addr) 547 return; 548 549 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 550 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 551 #ifdef CONFIG_STACKTRACE 552 { 553 int i; 554 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 555 if (t->addrs[i]) 556 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]); 557 else 558 break; 559 } 560 #endif 561 } 562 563 static void print_tracking(struct kmem_cache *s, void *object) 564 { 565 if (!(s->flags & SLAB_STORE_USER)) 566 return; 567 568 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 569 print_track("Freed", get_track(s, object, TRACK_FREE)); 570 } 571 572 static void print_page_info(struct page *page) 573 { 574 printk(KERN_ERR 575 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 576 page, page->objects, page->inuse, page->freelist, page->flags); 577 578 } 579 580 static void slab_bug(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 "========================================" 589 "=====================================\n"); 590 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf); 591 printk(KERN_ERR "----------------------------------------" 592 "-------------------------------------\n\n"); 593 594 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 595 } 596 597 static void slab_fix(struct kmem_cache *s, char *fmt, ...) 598 { 599 va_list args; 600 char buf[100]; 601 602 va_start(args, fmt); 603 vsnprintf(buf, sizeof(buf), fmt, args); 604 va_end(args); 605 printk(KERN_ERR "FIX %s: %s\n", s->name, buf); 606 } 607 608 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 609 { 610 unsigned int off; /* Offset of last byte */ 611 u8 *addr = page_address(page); 612 613 print_tracking(s, p); 614 615 print_page_info(page); 616 617 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 618 p, p - addr, get_freepointer(s, p)); 619 620 if (p > addr + 16) 621 print_section("Bytes b4 ", p - 16, 16); 622 623 print_section("Object ", p, min_t(unsigned long, s->object_size, 624 PAGE_SIZE)); 625 if (s->flags & SLAB_RED_ZONE) 626 print_section("Redzone ", p + s->object_size, 627 s->inuse - s->object_size); 628 629 if (s->offset) 630 off = s->offset + sizeof(void *); 631 else 632 off = s->inuse; 633 634 if (s->flags & SLAB_STORE_USER) 635 off += 2 * sizeof(struct track); 636 637 if (off != s->size) 638 /* Beginning of the filler is the free pointer */ 639 print_section("Padding ", p + off, s->size - off); 640 641 dump_stack(); 642 } 643 644 static void object_err(struct kmem_cache *s, struct page *page, 645 u8 *object, char *reason) 646 { 647 slab_bug(s, "%s", reason); 648 print_trailer(s, page, object); 649 } 650 651 static void slab_err(struct kmem_cache *s, struct page *page, 652 const char *fmt, ...) 653 { 654 va_list args; 655 char buf[100]; 656 657 va_start(args, fmt); 658 vsnprintf(buf, sizeof(buf), fmt, args); 659 va_end(args); 660 slab_bug(s, "%s", buf); 661 print_page_info(page); 662 dump_stack(); 663 } 664 665 static void init_object(struct kmem_cache *s, void *object, u8 val) 666 { 667 u8 *p = object; 668 669 if (s->flags & __OBJECT_POISON) { 670 memset(p, POISON_FREE, s->object_size - 1); 671 p[s->object_size - 1] = POISON_END; 672 } 673 674 if (s->flags & SLAB_RED_ZONE) 675 memset(p + s->object_size, val, s->inuse - s->object_size); 676 } 677 678 static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 679 void *from, void *to) 680 { 681 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 682 memset(from, data, to - from); 683 } 684 685 static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 686 u8 *object, char *what, 687 u8 *start, unsigned int value, unsigned int bytes) 688 { 689 u8 *fault; 690 u8 *end; 691 692 fault = memchr_inv(start, value, bytes); 693 if (!fault) 694 return 1; 695 696 end = start + bytes; 697 while (end > fault && end[-1] == value) 698 end--; 699 700 slab_bug(s, "%s overwritten", what); 701 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 702 fault, end - 1, fault[0], value); 703 print_trailer(s, page, object); 704 705 restore_bytes(s, what, value, fault, end); 706 return 0; 707 } 708 709 /* 710 * Object layout: 711 * 712 * object address 713 * Bytes of the object to be managed. 714 * If the freepointer may overlay the object then the free 715 * pointer is the first word of the object. 716 * 717 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 718 * 0xa5 (POISON_END) 719 * 720 * object + s->object_size 721 * Padding to reach word boundary. This is also used for Redzoning. 722 * Padding is extended by another word if Redzoning is enabled and 723 * object_size == inuse. 724 * 725 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 726 * 0xcc (RED_ACTIVE) for objects in use. 727 * 728 * object + s->inuse 729 * Meta data starts here. 730 * 731 * A. Free pointer (if we cannot overwrite object on free) 732 * B. Tracking data for SLAB_STORE_USER 733 * C. Padding to reach required alignment boundary or at mininum 734 * one word if debugging is on to be able to detect writes 735 * before the word boundary. 736 * 737 * Padding is done using 0x5a (POISON_INUSE) 738 * 739 * object + s->size 740 * Nothing is used beyond s->size. 741 * 742 * If slabcaches are merged then the object_size and inuse boundaries are mostly 743 * ignored. And therefore no slab options that rely on these boundaries 744 * may be used with merged slabcaches. 745 */ 746 747 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 748 { 749 unsigned long off = s->inuse; /* The end of info */ 750 751 if (s->offset) 752 /* Freepointer is placed after the object. */ 753 off += sizeof(void *); 754 755 if (s->flags & SLAB_STORE_USER) 756 /* We also have user information there */ 757 off += 2 * sizeof(struct track); 758 759 if (s->size == off) 760 return 1; 761 762 return check_bytes_and_report(s, page, p, "Object padding", 763 p + off, POISON_INUSE, s->size - off); 764 } 765 766 /* Check the pad bytes at the end of a slab page */ 767 static int slab_pad_check(struct kmem_cache *s, struct page *page) 768 { 769 u8 *start; 770 u8 *fault; 771 u8 *end; 772 int length; 773 int remainder; 774 775 if (!(s->flags & SLAB_POISON)) 776 return 1; 777 778 start = page_address(page); 779 length = (PAGE_SIZE << compound_order(page)) - s->reserved; 780 end = start + length; 781 remainder = length % s->size; 782 if (!remainder) 783 return 1; 784 785 fault = memchr_inv(end - remainder, POISON_INUSE, remainder); 786 if (!fault) 787 return 1; 788 while (end > fault && end[-1] == POISON_INUSE) 789 end--; 790 791 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 792 print_section("Padding ", end - remainder, remainder); 793 794 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 795 return 0; 796 } 797 798 static int check_object(struct kmem_cache *s, struct page *page, 799 void *object, u8 val) 800 { 801 u8 *p = object; 802 u8 *endobject = object + s->object_size; 803 804 if (s->flags & SLAB_RED_ZONE) { 805 if (!check_bytes_and_report(s, page, object, "Redzone", 806 endobject, val, s->inuse - s->object_size)) 807 return 0; 808 } else { 809 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { 810 check_bytes_and_report(s, page, p, "Alignment padding", 811 endobject, POISON_INUSE, 812 s->inuse - s->object_size); 813 } 814 } 815 816 if (s->flags & SLAB_POISON) { 817 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 818 (!check_bytes_and_report(s, page, p, "Poison", p, 819 POISON_FREE, s->object_size - 1) || 820 !check_bytes_and_report(s, page, p, "Poison", 821 p + s->object_size - 1, POISON_END, 1))) 822 return 0; 823 /* 824 * check_pad_bytes cleans up on its own. 825 */ 826 check_pad_bytes(s, page, p); 827 } 828 829 if (!s->offset && val == SLUB_RED_ACTIVE) 830 /* 831 * Object and freepointer overlap. Cannot check 832 * freepointer while object is allocated. 833 */ 834 return 1; 835 836 /* Check free pointer validity */ 837 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 838 object_err(s, page, p, "Freepointer corrupt"); 839 /* 840 * No choice but to zap it and thus lose the remainder 841 * of the free objects in this slab. May cause 842 * another error because the object count is now wrong. 843 */ 844 set_freepointer(s, p, NULL); 845 return 0; 846 } 847 return 1; 848 } 849 850 static int check_slab(struct kmem_cache *s, struct page *page) 851 { 852 int maxobj; 853 854 VM_BUG_ON(!irqs_disabled()); 855 856 if (!PageSlab(page)) { 857 slab_err(s, page, "Not a valid slab page"); 858 return 0; 859 } 860 861 maxobj = order_objects(compound_order(page), s->size, s->reserved); 862 if (page->objects > maxobj) { 863 slab_err(s, page, "objects %u > max %u", 864 s->name, page->objects, maxobj); 865 return 0; 866 } 867 if (page->inuse > page->objects) { 868 slab_err(s, page, "inuse %u > max %u", 869 s->name, page->inuse, page->objects); 870 return 0; 871 } 872 /* Slab_pad_check fixes things up after itself */ 873 slab_pad_check(s, page); 874 return 1; 875 } 876 877 /* 878 * Determine if a certain object on a page is on the freelist. Must hold the 879 * slab lock to guarantee that the chains are in a consistent state. 880 */ 881 static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 882 { 883 int nr = 0; 884 void *fp; 885 void *object = NULL; 886 unsigned long max_objects; 887 888 fp = page->freelist; 889 while (fp && nr <= page->objects) { 890 if (fp == search) 891 return 1; 892 if (!check_valid_pointer(s, page, fp)) { 893 if (object) { 894 object_err(s, page, object, 895 "Freechain corrupt"); 896 set_freepointer(s, object, NULL); 897 } else { 898 slab_err(s, page, "Freepointer corrupt"); 899 page->freelist = NULL; 900 page->inuse = page->objects; 901 slab_fix(s, "Freelist cleared"); 902 return 0; 903 } 904 break; 905 } 906 object = fp; 907 fp = get_freepointer(s, object); 908 nr++; 909 } 910 911 max_objects = order_objects(compound_order(page), s->size, s->reserved); 912 if (max_objects > MAX_OBJS_PER_PAGE) 913 max_objects = MAX_OBJS_PER_PAGE; 914 915 if (page->objects != max_objects) { 916 slab_err(s, page, "Wrong number of objects. Found %d but " 917 "should be %d", page->objects, max_objects); 918 page->objects = max_objects; 919 slab_fix(s, "Number of objects adjusted."); 920 } 921 if (page->inuse != page->objects - nr) { 922 slab_err(s, page, "Wrong object count. Counter is %d but " 923 "counted were %d", page->inuse, page->objects - nr); 924 page->inuse = page->objects - nr; 925 slab_fix(s, "Object count adjusted."); 926 } 927 return search == NULL; 928 } 929 930 static void trace(struct kmem_cache *s, struct page *page, void *object, 931 int alloc) 932 { 933 if (s->flags & SLAB_TRACE) { 934 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 935 s->name, 936 alloc ? "alloc" : "free", 937 object, page->inuse, 938 page->freelist); 939 940 if (!alloc) 941 print_section("Object ", (void *)object, 942 s->object_size); 943 944 dump_stack(); 945 } 946 } 947 948 /* 949 * Hooks for other subsystems that check memory allocations. In a typical 950 * production configuration these hooks all should produce no code at all. 951 */ 952 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) 953 { 954 kmemleak_alloc(ptr, size, 1, flags); 955 } 956 957 static inline void kfree_hook(const void *x) 958 { 959 kmemleak_free(x); 960 } 961 962 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 963 { 964 flags &= gfp_allowed_mask; 965 lockdep_trace_alloc(flags); 966 might_sleep_if(flags & __GFP_WAIT); 967 968 return should_failslab(s->object_size, flags, s->flags); 969 } 970 971 static inline void slab_post_alloc_hook(struct kmem_cache *s, 972 gfp_t flags, void *object) 973 { 974 flags &= gfp_allowed_mask; 975 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s)); 976 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags); 977 } 978 979 static inline void slab_free_hook(struct kmem_cache *s, void *x) 980 { 981 kmemleak_free_recursive(x, s->flags); 982 983 /* 984 * Trouble is that we may no longer disable interrupts in the fast path 985 * So in order to make the debug calls that expect irqs to be 986 * disabled we need to disable interrupts temporarily. 987 */ 988 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) 989 { 990 unsigned long flags; 991 992 local_irq_save(flags); 993 kmemcheck_slab_free(s, x, s->object_size); 994 debug_check_no_locks_freed(x, s->object_size); 995 local_irq_restore(flags); 996 } 997 #endif 998 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 999 debug_check_no_obj_freed(x, s->object_size); 1000 } 1001 1002 /* 1003 * Tracking of fully allocated slabs for debugging purposes. 1004 */ 1005 static void add_full(struct kmem_cache *s, 1006 struct kmem_cache_node *n, struct page *page) 1007 { 1008 if (!(s->flags & SLAB_STORE_USER)) 1009 return; 1010 1011 lockdep_assert_held(&n->list_lock); 1012 list_add(&page->lru, &n->full); 1013 } 1014 1015 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page) 1016 { 1017 if (!(s->flags & SLAB_STORE_USER)) 1018 return; 1019 1020 lockdep_assert_held(&n->list_lock); 1021 list_del(&page->lru); 1022 } 1023 1024 /* Tracking of the number of slabs for debugging purposes */ 1025 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1026 { 1027 struct kmem_cache_node *n = get_node(s, node); 1028 1029 return atomic_long_read(&n->nr_slabs); 1030 } 1031 1032 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1033 { 1034 return atomic_long_read(&n->nr_slabs); 1035 } 1036 1037 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 1038 { 1039 struct kmem_cache_node *n = get_node(s, node); 1040 1041 /* 1042 * May be called early in order to allocate a slab for the 1043 * kmem_cache_node structure. Solve the chicken-egg 1044 * dilemma by deferring the increment of the count during 1045 * bootstrap (see early_kmem_cache_node_alloc). 1046 */ 1047 if (likely(n)) { 1048 atomic_long_inc(&n->nr_slabs); 1049 atomic_long_add(objects, &n->total_objects); 1050 } 1051 } 1052 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 1053 { 1054 struct kmem_cache_node *n = get_node(s, node); 1055 1056 atomic_long_dec(&n->nr_slabs); 1057 atomic_long_sub(objects, &n->total_objects); 1058 } 1059 1060 /* Object debug checks for alloc/free paths */ 1061 static void setup_object_debug(struct kmem_cache *s, struct page *page, 1062 void *object) 1063 { 1064 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 1065 return; 1066 1067 init_object(s, object, SLUB_RED_INACTIVE); 1068 init_tracking(s, object); 1069 } 1070 1071 static noinline int alloc_debug_processing(struct kmem_cache *s, 1072 struct page *page, 1073 void *object, unsigned long addr) 1074 { 1075 if (!check_slab(s, page)) 1076 goto bad; 1077 1078 if (!check_valid_pointer(s, page, object)) { 1079 object_err(s, page, object, "Freelist Pointer check fails"); 1080 goto bad; 1081 } 1082 1083 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1084 goto bad; 1085 1086 /* Success perform special debug activities for allocs */ 1087 if (s->flags & SLAB_STORE_USER) 1088 set_track(s, object, TRACK_ALLOC, addr); 1089 trace(s, page, object, 1); 1090 init_object(s, object, SLUB_RED_ACTIVE); 1091 return 1; 1092 1093 bad: 1094 if (PageSlab(page)) { 1095 /* 1096 * If this is a slab page then lets do the best we can 1097 * to avoid issues in the future. Marking all objects 1098 * as used avoids touching the remaining objects. 1099 */ 1100 slab_fix(s, "Marking all objects used"); 1101 page->inuse = page->objects; 1102 page->freelist = NULL; 1103 } 1104 return 0; 1105 } 1106 1107 static noinline struct kmem_cache_node *free_debug_processing( 1108 struct kmem_cache *s, struct page *page, void *object, 1109 unsigned long addr, unsigned long *flags) 1110 { 1111 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1112 1113 spin_lock_irqsave(&n->list_lock, *flags); 1114 slab_lock(page); 1115 1116 if (!check_slab(s, page)) 1117 goto fail; 1118 1119 if (!check_valid_pointer(s, page, object)) { 1120 slab_err(s, page, "Invalid object pointer 0x%p", object); 1121 goto fail; 1122 } 1123 1124 if (on_freelist(s, page, object)) { 1125 object_err(s, page, object, "Object already free"); 1126 goto fail; 1127 } 1128 1129 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1130 goto out; 1131 1132 if (unlikely(s != page->slab_cache)) { 1133 if (!PageSlab(page)) { 1134 slab_err(s, page, "Attempt to free object(0x%p) " 1135 "outside of slab", object); 1136 } else if (!page->slab_cache) { 1137 printk(KERN_ERR 1138 "SLUB <none>: no slab for object 0x%p.\n", 1139 object); 1140 dump_stack(); 1141 } else 1142 object_err(s, page, object, 1143 "page slab pointer corrupt."); 1144 goto fail; 1145 } 1146 1147 if (s->flags & SLAB_STORE_USER) 1148 set_track(s, object, TRACK_FREE, addr); 1149 trace(s, page, object, 0); 1150 init_object(s, object, SLUB_RED_INACTIVE); 1151 out: 1152 slab_unlock(page); 1153 /* 1154 * Keep node_lock to preserve integrity 1155 * until the object is actually freed 1156 */ 1157 return n; 1158 1159 fail: 1160 slab_unlock(page); 1161 spin_unlock_irqrestore(&n->list_lock, *flags); 1162 slab_fix(s, "Object at 0x%p not freed", object); 1163 return NULL; 1164 } 1165 1166 static int __init setup_slub_debug(char *str) 1167 { 1168 slub_debug = DEBUG_DEFAULT_FLAGS; 1169 if (*str++ != '=' || !*str) 1170 /* 1171 * No options specified. Switch on full debugging. 1172 */ 1173 goto out; 1174 1175 if (*str == ',') 1176 /* 1177 * No options but restriction on slabs. This means full 1178 * debugging for slabs matching a pattern. 1179 */ 1180 goto check_slabs; 1181 1182 if (tolower(*str) == 'o') { 1183 /* 1184 * Avoid enabling debugging on caches if its minimum order 1185 * would increase as a result. 1186 */ 1187 disable_higher_order_debug = 1; 1188 goto out; 1189 } 1190 1191 slub_debug = 0; 1192 if (*str == '-') 1193 /* 1194 * Switch off all debugging measures. 1195 */ 1196 goto out; 1197 1198 /* 1199 * Determine which debug features should be switched on 1200 */ 1201 for (; *str && *str != ','; str++) { 1202 switch (tolower(*str)) { 1203 case 'f': 1204 slub_debug |= SLAB_DEBUG_FREE; 1205 break; 1206 case 'z': 1207 slub_debug |= SLAB_RED_ZONE; 1208 break; 1209 case 'p': 1210 slub_debug |= SLAB_POISON; 1211 break; 1212 case 'u': 1213 slub_debug |= SLAB_STORE_USER; 1214 break; 1215 case 't': 1216 slub_debug |= SLAB_TRACE; 1217 break; 1218 case 'a': 1219 slub_debug |= SLAB_FAILSLAB; 1220 break; 1221 default: 1222 printk(KERN_ERR "slub_debug option '%c' " 1223 "unknown. skipped\n", *str); 1224 } 1225 } 1226 1227 check_slabs: 1228 if (*str == ',') 1229 slub_debug_slabs = str + 1; 1230 out: 1231 return 1; 1232 } 1233 1234 __setup("slub_debug", setup_slub_debug); 1235 1236 static unsigned long kmem_cache_flags(unsigned long object_size, 1237 unsigned long flags, const char *name, 1238 void (*ctor)(void *)) 1239 { 1240 /* 1241 * Enable debugging if selected on the kernel commandline. 1242 */ 1243 if (slub_debug && (!slub_debug_slabs || (name && 1244 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))) 1245 flags |= slub_debug; 1246 1247 return flags; 1248 } 1249 #else 1250 static inline void setup_object_debug(struct kmem_cache *s, 1251 struct page *page, void *object) {} 1252 1253 static inline int alloc_debug_processing(struct kmem_cache *s, 1254 struct page *page, void *object, unsigned long addr) { return 0; } 1255 1256 static inline struct kmem_cache_node *free_debug_processing( 1257 struct kmem_cache *s, struct page *page, void *object, 1258 unsigned long addr, unsigned long *flags) { return NULL; } 1259 1260 static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1261 { return 1; } 1262 static inline int check_object(struct kmem_cache *s, struct page *page, 1263 void *object, u8 val) { return 1; } 1264 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, 1265 struct page *page) {} 1266 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, 1267 struct page *page) {} 1268 static inline unsigned long kmem_cache_flags(unsigned long object_size, 1269 unsigned long flags, const char *name, 1270 void (*ctor)(void *)) 1271 { 1272 return flags; 1273 } 1274 #define slub_debug 0 1275 1276 #define disable_higher_order_debug 0 1277 1278 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1279 { return 0; } 1280 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1281 { return 0; } 1282 static inline void inc_slabs_node(struct kmem_cache *s, int node, 1283 int objects) {} 1284 static inline void dec_slabs_node(struct kmem_cache *s, int node, 1285 int objects) {} 1286 1287 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) 1288 { 1289 kmemleak_alloc(ptr, size, 1, flags); 1290 } 1291 1292 static inline void kfree_hook(const void *x) 1293 { 1294 kmemleak_free(x); 1295 } 1296 1297 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags) 1298 { return 0; } 1299 1300 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, 1301 void *object) 1302 { 1303 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, 1304 flags & gfp_allowed_mask); 1305 } 1306 1307 static inline void slab_free_hook(struct kmem_cache *s, void *x) 1308 { 1309 kmemleak_free_recursive(x, s->flags); 1310 } 1311 1312 #endif /* CONFIG_SLUB_DEBUG */ 1313 1314 /* 1315 * Slab allocation and freeing 1316 */ 1317 static inline struct page *alloc_slab_page(gfp_t flags, int node, 1318 struct kmem_cache_order_objects oo) 1319 { 1320 int order = oo_order(oo); 1321 1322 flags |= __GFP_NOTRACK; 1323 1324 if (node == NUMA_NO_NODE) 1325 return alloc_pages(flags, order); 1326 else 1327 return alloc_pages_exact_node(node, flags, order); 1328 } 1329 1330 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1331 { 1332 struct page *page; 1333 struct kmem_cache_order_objects oo = s->oo; 1334 gfp_t alloc_gfp; 1335 1336 flags &= gfp_allowed_mask; 1337 1338 if (flags & __GFP_WAIT) 1339 local_irq_enable(); 1340 1341 flags |= s->allocflags; 1342 1343 /* 1344 * Let the initial higher-order allocation fail under memory pressure 1345 * so we fall-back to the minimum order allocation. 1346 */ 1347 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1348 1349 page = alloc_slab_page(alloc_gfp, node, oo); 1350 if (unlikely(!page)) { 1351 oo = s->min; 1352 alloc_gfp = flags; 1353 /* 1354 * Allocation may have failed due to fragmentation. 1355 * Try a lower order alloc if possible 1356 */ 1357 page = alloc_slab_page(alloc_gfp, node, oo); 1358 1359 if (page) 1360 stat(s, ORDER_FALLBACK); 1361 } 1362 1363 if (kmemcheck_enabled && page 1364 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1365 int pages = 1 << oo_order(oo); 1366 1367 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node); 1368 1369 /* 1370 * Objects from caches that have a constructor don't get 1371 * cleared when they're allocated, so we need to do it here. 1372 */ 1373 if (s->ctor) 1374 kmemcheck_mark_uninitialized_pages(page, pages); 1375 else 1376 kmemcheck_mark_unallocated_pages(page, pages); 1377 } 1378 1379 if (flags & __GFP_WAIT) 1380 local_irq_disable(); 1381 if (!page) 1382 return NULL; 1383 1384 page->objects = oo_objects(oo); 1385 mod_zone_page_state(page_zone(page), 1386 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1387 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1388 1 << oo_order(oo)); 1389 1390 return page; 1391 } 1392 1393 static void setup_object(struct kmem_cache *s, struct page *page, 1394 void *object) 1395 { 1396 setup_object_debug(s, page, object); 1397 if (unlikely(s->ctor)) 1398 s->ctor(object); 1399 } 1400 1401 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1402 { 1403 struct page *page; 1404 void *start; 1405 void *last; 1406 void *p; 1407 int order; 1408 1409 BUG_ON(flags & GFP_SLAB_BUG_MASK); 1410 1411 page = allocate_slab(s, 1412 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1413 if (!page) 1414 goto out; 1415 1416 order = compound_order(page); 1417 inc_slabs_node(s, page_to_nid(page), page->objects); 1418 memcg_bind_pages(s, order); 1419 page->slab_cache = s; 1420 __SetPageSlab(page); 1421 if (page->pfmemalloc) 1422 SetPageSlabPfmemalloc(page); 1423 1424 start = page_address(page); 1425 1426 if (unlikely(s->flags & SLAB_POISON)) 1427 memset(start, POISON_INUSE, PAGE_SIZE << order); 1428 1429 last = start; 1430 for_each_object(p, s, start, page->objects) { 1431 setup_object(s, page, last); 1432 set_freepointer(s, last, p); 1433 last = p; 1434 } 1435 setup_object(s, page, last); 1436 set_freepointer(s, last, NULL); 1437 1438 page->freelist = start; 1439 page->inuse = page->objects; 1440 page->frozen = 1; 1441 out: 1442 return page; 1443 } 1444 1445 static void __free_slab(struct kmem_cache *s, struct page *page) 1446 { 1447 int order = compound_order(page); 1448 int pages = 1 << order; 1449 1450 if (kmem_cache_debug(s)) { 1451 void *p; 1452 1453 slab_pad_check(s, page); 1454 for_each_object(p, s, page_address(page), 1455 page->objects) 1456 check_object(s, page, p, SLUB_RED_INACTIVE); 1457 } 1458 1459 kmemcheck_free_shadow(page, compound_order(page)); 1460 1461 mod_zone_page_state(page_zone(page), 1462 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1463 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1464 -pages); 1465 1466 __ClearPageSlabPfmemalloc(page); 1467 __ClearPageSlab(page); 1468 1469 memcg_release_pages(s, order); 1470 page_mapcount_reset(page); 1471 if (current->reclaim_state) 1472 current->reclaim_state->reclaimed_slab += pages; 1473 __free_memcg_kmem_pages(page, order); 1474 } 1475 1476 #define need_reserve_slab_rcu \ 1477 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) 1478 1479 static void rcu_free_slab(struct rcu_head *h) 1480 { 1481 struct page *page; 1482 1483 if (need_reserve_slab_rcu) 1484 page = virt_to_head_page(h); 1485 else 1486 page = container_of((struct list_head *)h, struct page, lru); 1487 1488 __free_slab(page->slab_cache, page); 1489 } 1490 1491 static void free_slab(struct kmem_cache *s, struct page *page) 1492 { 1493 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1494 struct rcu_head *head; 1495 1496 if (need_reserve_slab_rcu) { 1497 int order = compound_order(page); 1498 int offset = (PAGE_SIZE << order) - s->reserved; 1499 1500 VM_BUG_ON(s->reserved != sizeof(*head)); 1501 head = page_address(page) + offset; 1502 } else { 1503 /* 1504 * RCU free overloads the RCU head over the LRU 1505 */ 1506 head = (void *)&page->lru; 1507 } 1508 1509 call_rcu(head, rcu_free_slab); 1510 } else 1511 __free_slab(s, page); 1512 } 1513 1514 static void discard_slab(struct kmem_cache *s, struct page *page) 1515 { 1516 dec_slabs_node(s, page_to_nid(page), page->objects); 1517 free_slab(s, page); 1518 } 1519 1520 /* 1521 * Management of partially allocated slabs. 1522 */ 1523 static inline void 1524 __add_partial(struct kmem_cache_node *n, struct page *page, int tail) 1525 { 1526 n->nr_partial++; 1527 if (tail == DEACTIVATE_TO_TAIL) 1528 list_add_tail(&page->lru, &n->partial); 1529 else 1530 list_add(&page->lru, &n->partial); 1531 } 1532 1533 static inline void add_partial(struct kmem_cache_node *n, 1534 struct page *page, int tail) 1535 { 1536 lockdep_assert_held(&n->list_lock); 1537 __add_partial(n, page, tail); 1538 } 1539 1540 static inline void 1541 __remove_partial(struct kmem_cache_node *n, struct page *page) 1542 { 1543 list_del(&page->lru); 1544 n->nr_partial--; 1545 } 1546 1547 static inline void remove_partial(struct kmem_cache_node *n, 1548 struct page *page) 1549 { 1550 lockdep_assert_held(&n->list_lock); 1551 __remove_partial(n, page); 1552 } 1553 1554 /* 1555 * Remove slab from the partial list, freeze it and 1556 * return the pointer to the freelist. 1557 * 1558 * Returns a list of objects or NULL if it fails. 1559 */ 1560 static inline void *acquire_slab(struct kmem_cache *s, 1561 struct kmem_cache_node *n, struct page *page, 1562 int mode, int *objects) 1563 { 1564 void *freelist; 1565 unsigned long counters; 1566 struct page new; 1567 1568 lockdep_assert_held(&n->list_lock); 1569 1570 /* 1571 * Zap the freelist and set the frozen bit. 1572 * The old freelist is the list of objects for the 1573 * per cpu allocation list. 1574 */ 1575 freelist = page->freelist; 1576 counters = page->counters; 1577 new.counters = counters; 1578 *objects = new.objects - new.inuse; 1579 if (mode) { 1580 new.inuse = page->objects; 1581 new.freelist = NULL; 1582 } else { 1583 new.freelist = freelist; 1584 } 1585 1586 VM_BUG_ON(new.frozen); 1587 new.frozen = 1; 1588 1589 if (!__cmpxchg_double_slab(s, page, 1590 freelist, counters, 1591 new.freelist, new.counters, 1592 "acquire_slab")) 1593 return NULL; 1594 1595 remove_partial(n, page); 1596 WARN_ON(!freelist); 1597 return freelist; 1598 } 1599 1600 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); 1601 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags); 1602 1603 /* 1604 * Try to allocate a partial slab from a specific node. 1605 */ 1606 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, 1607 struct kmem_cache_cpu *c, gfp_t flags) 1608 { 1609 struct page *page, *page2; 1610 void *object = NULL; 1611 int available = 0; 1612 int objects; 1613 1614 /* 1615 * Racy check. If we mistakenly see no partial slabs then we 1616 * just allocate an empty slab. If we mistakenly try to get a 1617 * partial slab and there is none available then get_partials() 1618 * will return NULL. 1619 */ 1620 if (!n || !n->nr_partial) 1621 return NULL; 1622 1623 spin_lock(&n->list_lock); 1624 list_for_each_entry_safe(page, page2, &n->partial, lru) { 1625 void *t; 1626 1627 if (!pfmemalloc_match(page, flags)) 1628 continue; 1629 1630 t = acquire_slab(s, n, page, object == NULL, &objects); 1631 if (!t) 1632 break; 1633 1634 available += objects; 1635 if (!object) { 1636 c->page = page; 1637 stat(s, ALLOC_FROM_PARTIAL); 1638 object = t; 1639 } else { 1640 put_cpu_partial(s, page, 0); 1641 stat(s, CPU_PARTIAL_NODE); 1642 } 1643 if (!kmem_cache_has_cpu_partial(s) 1644 || available > s->cpu_partial / 2) 1645 break; 1646 1647 } 1648 spin_unlock(&n->list_lock); 1649 return object; 1650 } 1651 1652 /* 1653 * Get a page from somewhere. Search in increasing NUMA distances. 1654 */ 1655 static void *get_any_partial(struct kmem_cache *s, gfp_t flags, 1656 struct kmem_cache_cpu *c) 1657 { 1658 #ifdef CONFIG_NUMA 1659 struct zonelist *zonelist; 1660 struct zoneref *z; 1661 struct zone *zone; 1662 enum zone_type high_zoneidx = gfp_zone(flags); 1663 void *object; 1664 unsigned int cpuset_mems_cookie; 1665 1666 /* 1667 * The defrag ratio allows a configuration of the tradeoffs between 1668 * inter node defragmentation and node local allocations. A lower 1669 * defrag_ratio increases the tendency to do local allocations 1670 * instead of attempting to obtain partial slabs from other nodes. 1671 * 1672 * If the defrag_ratio is set to 0 then kmalloc() always 1673 * returns node local objects. If the ratio is higher then kmalloc() 1674 * may return off node objects because partial slabs are obtained 1675 * from other nodes and filled up. 1676 * 1677 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes 1678 * defrag_ratio = 1000) then every (well almost) allocation will 1679 * first attempt to defrag slab caches on other nodes. This means 1680 * scanning over all nodes to look for partial slabs which may be 1681 * expensive if we do it every time we are trying to find a slab 1682 * with available objects. 1683 */ 1684 if (!s->remote_node_defrag_ratio || 1685 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1686 return NULL; 1687 1688 do { 1689 cpuset_mems_cookie = read_mems_allowed_begin(); 1690 zonelist = node_zonelist(mempolicy_slab_node(), flags); 1691 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1692 struct kmem_cache_node *n; 1693 1694 n = get_node(s, zone_to_nid(zone)); 1695 1696 if (n && cpuset_zone_allowed_hardwall(zone, flags) && 1697 n->nr_partial > s->min_partial) { 1698 object = get_partial_node(s, n, c, flags); 1699 if (object) { 1700 /* 1701 * Don't check read_mems_allowed_retry() 1702 * here - if mems_allowed was updated in 1703 * parallel, that was a harmless race 1704 * between allocation and the cpuset 1705 * update 1706 */ 1707 return object; 1708 } 1709 } 1710 } 1711 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1712 #endif 1713 return NULL; 1714 } 1715 1716 /* 1717 * Get a partial page, lock it and return it. 1718 */ 1719 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, 1720 struct kmem_cache_cpu *c) 1721 { 1722 void *object; 1723 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node; 1724 1725 object = get_partial_node(s, get_node(s, searchnode), c, flags); 1726 if (object || node != NUMA_NO_NODE) 1727 return object; 1728 1729 return get_any_partial(s, flags, c); 1730 } 1731 1732 #ifdef CONFIG_PREEMPT 1733 /* 1734 * Calculate the next globally unique transaction for disambiguiation 1735 * during cmpxchg. The transactions start with the cpu number and are then 1736 * incremented by CONFIG_NR_CPUS. 1737 */ 1738 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 1739 #else 1740 /* 1741 * No preemption supported therefore also no need to check for 1742 * different cpus. 1743 */ 1744 #define TID_STEP 1 1745 #endif 1746 1747 static inline unsigned long next_tid(unsigned long tid) 1748 { 1749 return tid + TID_STEP; 1750 } 1751 1752 static inline unsigned int tid_to_cpu(unsigned long tid) 1753 { 1754 return tid % TID_STEP; 1755 } 1756 1757 static inline unsigned long tid_to_event(unsigned long tid) 1758 { 1759 return tid / TID_STEP; 1760 } 1761 1762 static inline unsigned int init_tid(int cpu) 1763 { 1764 return cpu; 1765 } 1766 1767 static inline void note_cmpxchg_failure(const char *n, 1768 const struct kmem_cache *s, unsigned long tid) 1769 { 1770 #ifdef SLUB_DEBUG_CMPXCHG 1771 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 1772 1773 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name); 1774 1775 #ifdef CONFIG_PREEMPT 1776 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 1777 printk("due to cpu change %d -> %d\n", 1778 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 1779 else 1780 #endif 1781 if (tid_to_event(tid) != tid_to_event(actual_tid)) 1782 printk("due to cpu running other code. Event %ld->%ld\n", 1783 tid_to_event(tid), tid_to_event(actual_tid)); 1784 else 1785 printk("for unknown reason: actual=%lx was=%lx target=%lx\n", 1786 actual_tid, tid, next_tid(tid)); 1787 #endif 1788 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 1789 } 1790 1791 static void init_kmem_cache_cpus(struct kmem_cache *s) 1792 { 1793 int cpu; 1794 1795 for_each_possible_cpu(cpu) 1796 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 1797 } 1798 1799 /* 1800 * Remove the cpu slab 1801 */ 1802 static void deactivate_slab(struct kmem_cache *s, struct page *page, 1803 void *freelist) 1804 { 1805 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; 1806 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1807 int lock = 0; 1808 enum slab_modes l = M_NONE, m = M_NONE; 1809 void *nextfree; 1810 int tail = DEACTIVATE_TO_HEAD; 1811 struct page new; 1812 struct page old; 1813 1814 if (page->freelist) { 1815 stat(s, DEACTIVATE_REMOTE_FREES); 1816 tail = DEACTIVATE_TO_TAIL; 1817 } 1818 1819 /* 1820 * Stage one: Free all available per cpu objects back 1821 * to the page freelist while it is still frozen. Leave the 1822 * last one. 1823 * 1824 * There is no need to take the list->lock because the page 1825 * is still frozen. 1826 */ 1827 while (freelist && (nextfree = get_freepointer(s, freelist))) { 1828 void *prior; 1829 unsigned long counters; 1830 1831 do { 1832 prior = page->freelist; 1833 counters = page->counters; 1834 set_freepointer(s, freelist, prior); 1835 new.counters = counters; 1836 new.inuse--; 1837 VM_BUG_ON(!new.frozen); 1838 1839 } while (!__cmpxchg_double_slab(s, page, 1840 prior, counters, 1841 freelist, new.counters, 1842 "drain percpu freelist")); 1843 1844 freelist = nextfree; 1845 } 1846 1847 /* 1848 * Stage two: Ensure that the page is unfrozen while the 1849 * list presence reflects the actual number of objects 1850 * during unfreeze. 1851 * 1852 * We setup the list membership and then perform a cmpxchg 1853 * with the count. If there is a mismatch then the page 1854 * is not unfrozen but the page is on the wrong list. 1855 * 1856 * Then we restart the process which may have to remove 1857 * the page from the list that we just put it on again 1858 * because the number of objects in the slab may have 1859 * changed. 1860 */ 1861 redo: 1862 1863 old.freelist = page->freelist; 1864 old.counters = page->counters; 1865 VM_BUG_ON(!old.frozen); 1866 1867 /* Determine target state of the slab */ 1868 new.counters = old.counters; 1869 if (freelist) { 1870 new.inuse--; 1871 set_freepointer(s, freelist, old.freelist); 1872 new.freelist = freelist; 1873 } else 1874 new.freelist = old.freelist; 1875 1876 new.frozen = 0; 1877 1878 if (!new.inuse && n->nr_partial > s->min_partial) 1879 m = M_FREE; 1880 else if (new.freelist) { 1881 m = M_PARTIAL; 1882 if (!lock) { 1883 lock = 1; 1884 /* 1885 * Taking the spinlock removes the possiblity 1886 * that acquire_slab() will see a slab page that 1887 * is frozen 1888 */ 1889 spin_lock(&n->list_lock); 1890 } 1891 } else { 1892 m = M_FULL; 1893 if (kmem_cache_debug(s) && !lock) { 1894 lock = 1; 1895 /* 1896 * This also ensures that the scanning of full 1897 * slabs from diagnostic functions will not see 1898 * any frozen slabs. 1899 */ 1900 spin_lock(&n->list_lock); 1901 } 1902 } 1903 1904 if (l != m) { 1905 1906 if (l == M_PARTIAL) 1907 1908 remove_partial(n, page); 1909 1910 else if (l == M_FULL) 1911 1912 remove_full(s, n, page); 1913 1914 if (m == M_PARTIAL) { 1915 1916 add_partial(n, page, tail); 1917 stat(s, tail); 1918 1919 } else if (m == M_FULL) { 1920 1921 stat(s, DEACTIVATE_FULL); 1922 add_full(s, n, page); 1923 1924 } 1925 } 1926 1927 l = m; 1928 if (!__cmpxchg_double_slab(s, page, 1929 old.freelist, old.counters, 1930 new.freelist, new.counters, 1931 "unfreezing slab")) 1932 goto redo; 1933 1934 if (lock) 1935 spin_unlock(&n->list_lock); 1936 1937 if (m == M_FREE) { 1938 stat(s, DEACTIVATE_EMPTY); 1939 discard_slab(s, page); 1940 stat(s, FREE_SLAB); 1941 } 1942 } 1943 1944 /* 1945 * Unfreeze all the cpu partial slabs. 1946 * 1947 * This function must be called with interrupts disabled 1948 * for the cpu using c (or some other guarantee must be there 1949 * to guarantee no concurrent accesses). 1950 */ 1951 static void unfreeze_partials(struct kmem_cache *s, 1952 struct kmem_cache_cpu *c) 1953 { 1954 #ifdef CONFIG_SLUB_CPU_PARTIAL 1955 struct kmem_cache_node *n = NULL, *n2 = NULL; 1956 struct page *page, *discard_page = NULL; 1957 1958 while ((page = c->partial)) { 1959 struct page new; 1960 struct page old; 1961 1962 c->partial = page->next; 1963 1964 n2 = get_node(s, page_to_nid(page)); 1965 if (n != n2) { 1966 if (n) 1967 spin_unlock(&n->list_lock); 1968 1969 n = n2; 1970 spin_lock(&n->list_lock); 1971 } 1972 1973 do { 1974 1975 old.freelist = page->freelist; 1976 old.counters = page->counters; 1977 VM_BUG_ON(!old.frozen); 1978 1979 new.counters = old.counters; 1980 new.freelist = old.freelist; 1981 1982 new.frozen = 0; 1983 1984 } while (!__cmpxchg_double_slab(s, page, 1985 old.freelist, old.counters, 1986 new.freelist, new.counters, 1987 "unfreezing slab")); 1988 1989 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) { 1990 page->next = discard_page; 1991 discard_page = page; 1992 } else { 1993 add_partial(n, page, DEACTIVATE_TO_TAIL); 1994 stat(s, FREE_ADD_PARTIAL); 1995 } 1996 } 1997 1998 if (n) 1999 spin_unlock(&n->list_lock); 2000 2001 while (discard_page) { 2002 page = discard_page; 2003 discard_page = discard_page->next; 2004 2005 stat(s, DEACTIVATE_EMPTY); 2006 discard_slab(s, page); 2007 stat(s, FREE_SLAB); 2008 } 2009 #endif 2010 } 2011 2012 /* 2013 * Put a page that was just frozen (in __slab_free) into a partial page 2014 * slot if available. This is done without interrupts disabled and without 2015 * preemption disabled. The cmpxchg is racy and may put the partial page 2016 * onto a random cpus partial slot. 2017 * 2018 * If we did not find a slot then simply move all the partials to the 2019 * per node partial list. 2020 */ 2021 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) 2022 { 2023 #ifdef CONFIG_SLUB_CPU_PARTIAL 2024 struct page *oldpage; 2025 int pages; 2026 int pobjects; 2027 2028 do { 2029 pages = 0; 2030 pobjects = 0; 2031 oldpage = this_cpu_read(s->cpu_slab->partial); 2032 2033 if (oldpage) { 2034 pobjects = oldpage->pobjects; 2035 pages = oldpage->pages; 2036 if (drain && pobjects > s->cpu_partial) { 2037 unsigned long flags; 2038 /* 2039 * partial array is full. Move the existing 2040 * set to the per node partial list. 2041 */ 2042 local_irq_save(flags); 2043 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2044 local_irq_restore(flags); 2045 oldpage = NULL; 2046 pobjects = 0; 2047 pages = 0; 2048 stat(s, CPU_PARTIAL_DRAIN); 2049 } 2050 } 2051 2052 pages++; 2053 pobjects += page->objects - page->inuse; 2054 2055 page->pages = pages; 2056 page->pobjects = pobjects; 2057 page->next = oldpage; 2058 2059 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) 2060 != oldpage); 2061 #endif 2062 } 2063 2064 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 2065 { 2066 stat(s, CPUSLAB_FLUSH); 2067 deactivate_slab(s, c->page, c->freelist); 2068 2069 c->tid = next_tid(c->tid); 2070 c->page = NULL; 2071 c->freelist = NULL; 2072 } 2073 2074 /* 2075 * Flush cpu slab. 2076 * 2077 * Called from IPI handler with interrupts disabled. 2078 */ 2079 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 2080 { 2081 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2082 2083 if (likely(c)) { 2084 if (c->page) 2085 flush_slab(s, c); 2086 2087 unfreeze_partials(s, c); 2088 } 2089 } 2090 2091 static void flush_cpu_slab(void *d) 2092 { 2093 struct kmem_cache *s = d; 2094 2095 __flush_cpu_slab(s, smp_processor_id()); 2096 } 2097 2098 static bool has_cpu_slab(int cpu, void *info) 2099 { 2100 struct kmem_cache *s = info; 2101 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2102 2103 return c->page || c->partial; 2104 } 2105 2106 static void flush_all(struct kmem_cache *s) 2107 { 2108 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC); 2109 } 2110 2111 /* 2112 * Check if the objects in a per cpu structure fit numa 2113 * locality expectations. 2114 */ 2115 static inline int node_match(struct page *page, int node) 2116 { 2117 #ifdef CONFIG_NUMA 2118 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node)) 2119 return 0; 2120 #endif 2121 return 1; 2122 } 2123 2124 static int count_free(struct page *page) 2125 { 2126 return page->objects - page->inuse; 2127 } 2128 2129 static unsigned long count_partial(struct kmem_cache_node *n, 2130 int (*get_count)(struct page *)) 2131 { 2132 unsigned long flags; 2133 unsigned long x = 0; 2134 struct page *page; 2135 2136 spin_lock_irqsave(&n->list_lock, flags); 2137 list_for_each_entry(page, &n->partial, lru) 2138 x += get_count(page); 2139 spin_unlock_irqrestore(&n->list_lock, flags); 2140 return x; 2141 } 2142 2143 static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 2144 { 2145 #ifdef CONFIG_SLUB_DEBUG 2146 return atomic_long_read(&n->total_objects); 2147 #else 2148 return 0; 2149 #endif 2150 } 2151 2152 static noinline void 2153 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 2154 { 2155 int node; 2156 2157 printk(KERN_WARNING 2158 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n", 2159 nid, gfpflags); 2160 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, " 2161 "default order: %d, min order: %d\n", s->name, s->object_size, 2162 s->size, oo_order(s->oo), oo_order(s->min)); 2163 2164 if (oo_order(s->min) > get_order(s->object_size)) 2165 printk(KERN_WARNING " %s debugging increased min order, use " 2166 "slub_debug=O to disable.\n", s->name); 2167 2168 for_each_online_node(node) { 2169 struct kmem_cache_node *n = get_node(s, node); 2170 unsigned long nr_slabs; 2171 unsigned long nr_objs; 2172 unsigned long nr_free; 2173 2174 if (!n) 2175 continue; 2176 2177 nr_free = count_partial(n, count_free); 2178 nr_slabs = node_nr_slabs(n); 2179 nr_objs = node_nr_objs(n); 2180 2181 printk(KERN_WARNING 2182 " node %d: slabs: %ld, objs: %ld, free: %ld\n", 2183 node, nr_slabs, nr_objs, nr_free); 2184 } 2185 } 2186 2187 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, 2188 int node, struct kmem_cache_cpu **pc) 2189 { 2190 void *freelist; 2191 struct kmem_cache_cpu *c = *pc; 2192 struct page *page; 2193 2194 freelist = get_partial(s, flags, node, c); 2195 2196 if (freelist) 2197 return freelist; 2198 2199 page = new_slab(s, flags, node); 2200 if (page) { 2201 c = __this_cpu_ptr(s->cpu_slab); 2202 if (c->page) 2203 flush_slab(s, c); 2204 2205 /* 2206 * No other reference to the page yet so we can 2207 * muck around with it freely without cmpxchg 2208 */ 2209 freelist = page->freelist; 2210 page->freelist = NULL; 2211 2212 stat(s, ALLOC_SLAB); 2213 c->page = page; 2214 *pc = c; 2215 } else 2216 freelist = NULL; 2217 2218 return freelist; 2219 } 2220 2221 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags) 2222 { 2223 if (unlikely(PageSlabPfmemalloc(page))) 2224 return gfp_pfmemalloc_allowed(gfpflags); 2225 2226 return true; 2227 } 2228 2229 /* 2230 * Check the page->freelist of a page and either transfer the freelist to the 2231 * per cpu freelist or deactivate the page. 2232 * 2233 * The page is still frozen if the return value is not NULL. 2234 * 2235 * If this function returns NULL then the page has been unfrozen. 2236 * 2237 * This function must be called with interrupt disabled. 2238 */ 2239 static inline void *get_freelist(struct kmem_cache *s, struct page *page) 2240 { 2241 struct page new; 2242 unsigned long counters; 2243 void *freelist; 2244 2245 do { 2246 freelist = page->freelist; 2247 counters = page->counters; 2248 2249 new.counters = counters; 2250 VM_BUG_ON(!new.frozen); 2251 2252 new.inuse = page->objects; 2253 new.frozen = freelist != NULL; 2254 2255 } while (!__cmpxchg_double_slab(s, page, 2256 freelist, counters, 2257 NULL, new.counters, 2258 "get_freelist")); 2259 2260 return freelist; 2261 } 2262 2263 /* 2264 * Slow path. The lockless freelist is empty or we need to perform 2265 * debugging duties. 2266 * 2267 * Processing is still very fast if new objects have been freed to the 2268 * regular freelist. In that case we simply take over the regular freelist 2269 * as the lockless freelist and zap the regular freelist. 2270 * 2271 * If that is not working then we fall back to the partial lists. We take the 2272 * first element of the freelist as the object to allocate now and move the 2273 * rest of the freelist to the lockless freelist. 2274 * 2275 * And if we were unable to get a new slab from the partial slab lists then 2276 * we need to allocate a new slab. This is the slowest path since it involves 2277 * a call to the page allocator and the setup of a new slab. 2278 */ 2279 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2280 unsigned long addr, struct kmem_cache_cpu *c) 2281 { 2282 void *freelist; 2283 struct page *page; 2284 unsigned long flags; 2285 2286 local_irq_save(flags); 2287 #ifdef CONFIG_PREEMPT 2288 /* 2289 * We may have been preempted and rescheduled on a different 2290 * cpu before disabling interrupts. Need to reload cpu area 2291 * pointer. 2292 */ 2293 c = this_cpu_ptr(s->cpu_slab); 2294 #endif 2295 2296 page = c->page; 2297 if (!page) 2298 goto new_slab; 2299 redo: 2300 2301 if (unlikely(!node_match(page, node))) { 2302 stat(s, ALLOC_NODE_MISMATCH); 2303 deactivate_slab(s, page, c->freelist); 2304 c->page = NULL; 2305 c->freelist = NULL; 2306 goto new_slab; 2307 } 2308 2309 /* 2310 * By rights, we should be searching for a slab page that was 2311 * PFMEMALLOC but right now, we are losing the pfmemalloc 2312 * information when the page leaves the per-cpu allocator 2313 */ 2314 if (unlikely(!pfmemalloc_match(page, gfpflags))) { 2315 deactivate_slab(s, page, c->freelist); 2316 c->page = NULL; 2317 c->freelist = NULL; 2318 goto new_slab; 2319 } 2320 2321 /* must check again c->freelist in case of cpu migration or IRQ */ 2322 freelist = c->freelist; 2323 if (freelist) 2324 goto load_freelist; 2325 2326 stat(s, ALLOC_SLOWPATH); 2327 2328 freelist = get_freelist(s, page); 2329 2330 if (!freelist) { 2331 c->page = NULL; 2332 stat(s, DEACTIVATE_BYPASS); 2333 goto new_slab; 2334 } 2335 2336 stat(s, ALLOC_REFILL); 2337 2338 load_freelist: 2339 /* 2340 * freelist is pointing to the list of objects to be used. 2341 * page is pointing to the page from which the objects are obtained. 2342 * That page must be frozen for per cpu allocations to work. 2343 */ 2344 VM_BUG_ON(!c->page->frozen); 2345 c->freelist = get_freepointer(s, freelist); 2346 c->tid = next_tid(c->tid); 2347 local_irq_restore(flags); 2348 return freelist; 2349 2350 new_slab: 2351 2352 if (c->partial) { 2353 page = c->page = c->partial; 2354 c->partial = page->next; 2355 stat(s, CPU_PARTIAL_ALLOC); 2356 c->freelist = NULL; 2357 goto redo; 2358 } 2359 2360 freelist = new_slab_objects(s, gfpflags, node, &c); 2361 2362 if (unlikely(!freelist)) { 2363 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit()) 2364 slab_out_of_memory(s, gfpflags, node); 2365 2366 local_irq_restore(flags); 2367 return NULL; 2368 } 2369 2370 page = c->page; 2371 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags))) 2372 goto load_freelist; 2373 2374 /* Only entered in the debug case */ 2375 if (kmem_cache_debug(s) && 2376 !alloc_debug_processing(s, page, freelist, addr)) 2377 goto new_slab; /* Slab failed checks. Next slab needed */ 2378 2379 deactivate_slab(s, page, get_freepointer(s, freelist)); 2380 c->page = NULL; 2381 c->freelist = NULL; 2382 local_irq_restore(flags); 2383 return freelist; 2384 } 2385 2386 /* 2387 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 2388 * have the fastpath folded into their functions. So no function call 2389 * overhead for requests that can be satisfied on the fastpath. 2390 * 2391 * The fastpath works by first checking if the lockless freelist can be used. 2392 * If not then __slab_alloc is called for slow processing. 2393 * 2394 * Otherwise we can simply pick the next object from the lockless free list. 2395 */ 2396 static __always_inline void *slab_alloc_node(struct kmem_cache *s, 2397 gfp_t gfpflags, int node, unsigned long addr) 2398 { 2399 void **object; 2400 struct kmem_cache_cpu *c; 2401 struct page *page; 2402 unsigned long tid; 2403 2404 if (slab_pre_alloc_hook(s, gfpflags)) 2405 return NULL; 2406 2407 s = memcg_kmem_get_cache(s, gfpflags); 2408 redo: 2409 /* 2410 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 2411 * enabled. We may switch back and forth between cpus while 2412 * reading from one cpu area. That does not matter as long 2413 * as we end up on the original cpu again when doing the cmpxchg. 2414 * 2415 * Preemption is disabled for the retrieval of the tid because that 2416 * must occur from the current processor. We cannot allow rescheduling 2417 * on a different processor between the determination of the pointer 2418 * and the retrieval of the tid. 2419 */ 2420 preempt_disable(); 2421 c = __this_cpu_ptr(s->cpu_slab); 2422 2423 /* 2424 * The transaction ids are globally unique per cpu and per operation on 2425 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 2426 * occurs on the right processor and that there was no operation on the 2427 * linked list in between. 2428 */ 2429 tid = c->tid; 2430 preempt_enable(); 2431 2432 object = c->freelist; 2433 page = c->page; 2434 if (unlikely(!object || !node_match(page, node))) 2435 object = __slab_alloc(s, gfpflags, node, addr, c); 2436 2437 else { 2438 void *next_object = get_freepointer_safe(s, object); 2439 2440 /* 2441 * The cmpxchg will only match if there was no additional 2442 * operation and if we are on the right processor. 2443 * 2444 * The cmpxchg does the following atomically (without lock 2445 * semantics!) 2446 * 1. Relocate first pointer to the current per cpu area. 2447 * 2. Verify that tid and freelist have not been changed 2448 * 3. If they were not changed replace tid and freelist 2449 * 2450 * Since this is without lock semantics the protection is only 2451 * against code executing on this cpu *not* from access by 2452 * other cpus. 2453 */ 2454 if (unlikely(!this_cpu_cmpxchg_double( 2455 s->cpu_slab->freelist, s->cpu_slab->tid, 2456 object, tid, 2457 next_object, next_tid(tid)))) { 2458 2459 note_cmpxchg_failure("slab_alloc", s, tid); 2460 goto redo; 2461 } 2462 prefetch_freepointer(s, next_object); 2463 stat(s, ALLOC_FASTPATH); 2464 } 2465 2466 if (unlikely(gfpflags & __GFP_ZERO) && object) 2467 memset(object, 0, s->object_size); 2468 2469 slab_post_alloc_hook(s, gfpflags, object); 2470 2471 return object; 2472 } 2473 2474 static __always_inline void *slab_alloc(struct kmem_cache *s, 2475 gfp_t gfpflags, unsigned long addr) 2476 { 2477 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr); 2478 } 2479 2480 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 2481 { 2482 void *ret = slab_alloc(s, gfpflags, _RET_IP_); 2483 2484 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, 2485 s->size, gfpflags); 2486 2487 return ret; 2488 } 2489 EXPORT_SYMBOL(kmem_cache_alloc); 2490 2491 #ifdef CONFIG_TRACING 2492 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2493 { 2494 void *ret = slab_alloc(s, gfpflags, _RET_IP_); 2495 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2496 return ret; 2497 } 2498 EXPORT_SYMBOL(kmem_cache_alloc_trace); 2499 #endif 2500 2501 #ifdef CONFIG_NUMA 2502 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2503 { 2504 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); 2505 2506 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2507 s->object_size, s->size, gfpflags, node); 2508 2509 return ret; 2510 } 2511 EXPORT_SYMBOL(kmem_cache_alloc_node); 2512 2513 #ifdef CONFIG_TRACING 2514 void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2515 gfp_t gfpflags, 2516 int node, size_t size) 2517 { 2518 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); 2519 2520 trace_kmalloc_node(_RET_IP_, ret, 2521 size, s->size, gfpflags, node); 2522 return ret; 2523 } 2524 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2525 #endif 2526 #endif 2527 2528 /* 2529 * Slow patch handling. This may still be called frequently since objects 2530 * have a longer lifetime than the cpu slabs in most processing loads. 2531 * 2532 * So we still attempt to reduce cache line usage. Just take the slab 2533 * lock and free the item. If there is no additional partial page 2534 * handling required then we can return immediately. 2535 */ 2536 static void __slab_free(struct kmem_cache *s, struct page *page, 2537 void *x, unsigned long addr) 2538 { 2539 void *prior; 2540 void **object = (void *)x; 2541 int was_frozen; 2542 struct page new; 2543 unsigned long counters; 2544 struct kmem_cache_node *n = NULL; 2545 unsigned long uninitialized_var(flags); 2546 2547 stat(s, FREE_SLOWPATH); 2548 2549 if (kmem_cache_debug(s) && 2550 !(n = free_debug_processing(s, page, x, addr, &flags))) 2551 return; 2552 2553 do { 2554 if (unlikely(n)) { 2555 spin_unlock_irqrestore(&n->list_lock, flags); 2556 n = NULL; 2557 } 2558 prior = page->freelist; 2559 counters = page->counters; 2560 set_freepointer(s, object, prior); 2561 new.counters = counters; 2562 was_frozen = new.frozen; 2563 new.inuse--; 2564 if ((!new.inuse || !prior) && !was_frozen) { 2565 2566 if (kmem_cache_has_cpu_partial(s) && !prior) { 2567 2568 /* 2569 * Slab was on no list before and will be 2570 * partially empty 2571 * We can defer the list move and instead 2572 * freeze it. 2573 */ 2574 new.frozen = 1; 2575 2576 } else { /* Needs to be taken off a list */ 2577 2578 n = get_node(s, page_to_nid(page)); 2579 /* 2580 * Speculatively acquire the list_lock. 2581 * If the cmpxchg does not succeed then we may 2582 * drop the list_lock without any processing. 2583 * 2584 * Otherwise the list_lock will synchronize with 2585 * other processors updating the list of slabs. 2586 */ 2587 spin_lock_irqsave(&n->list_lock, flags); 2588 2589 } 2590 } 2591 2592 } while (!cmpxchg_double_slab(s, page, 2593 prior, counters, 2594 object, new.counters, 2595 "__slab_free")); 2596 2597 if (likely(!n)) { 2598 2599 /* 2600 * If we just froze the page then put it onto the 2601 * per cpu partial list. 2602 */ 2603 if (new.frozen && !was_frozen) { 2604 put_cpu_partial(s, page, 1); 2605 stat(s, CPU_PARTIAL_FREE); 2606 } 2607 /* 2608 * The list lock was not taken therefore no list 2609 * activity can be necessary. 2610 */ 2611 if (was_frozen) 2612 stat(s, FREE_FROZEN); 2613 return; 2614 } 2615 2616 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) 2617 goto slab_empty; 2618 2619 /* 2620 * Objects left in the slab. If it was not on the partial list before 2621 * then add it. 2622 */ 2623 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { 2624 if (kmem_cache_debug(s)) 2625 remove_full(s, n, page); 2626 add_partial(n, page, DEACTIVATE_TO_TAIL); 2627 stat(s, FREE_ADD_PARTIAL); 2628 } 2629 spin_unlock_irqrestore(&n->list_lock, flags); 2630 return; 2631 2632 slab_empty: 2633 if (prior) { 2634 /* 2635 * Slab on the partial list. 2636 */ 2637 remove_partial(n, page); 2638 stat(s, FREE_REMOVE_PARTIAL); 2639 } else { 2640 /* Slab must be on the full list */ 2641 remove_full(s, n, page); 2642 } 2643 2644 spin_unlock_irqrestore(&n->list_lock, flags); 2645 stat(s, FREE_SLAB); 2646 discard_slab(s, page); 2647 } 2648 2649 /* 2650 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 2651 * can perform fastpath freeing without additional function calls. 2652 * 2653 * The fastpath is only possible if we are freeing to the current cpu slab 2654 * of this processor. This typically the case if we have just allocated 2655 * the item before. 2656 * 2657 * If fastpath is not possible then fall back to __slab_free where we deal 2658 * with all sorts of special processing. 2659 */ 2660 static __always_inline void slab_free(struct kmem_cache *s, 2661 struct page *page, void *x, unsigned long addr) 2662 { 2663 void **object = (void *)x; 2664 struct kmem_cache_cpu *c; 2665 unsigned long tid; 2666 2667 slab_free_hook(s, x); 2668 2669 redo: 2670 /* 2671 * Determine the currently cpus per cpu slab. 2672 * The cpu may change afterward. However that does not matter since 2673 * data is retrieved via this pointer. If we are on the same cpu 2674 * during the cmpxchg then the free will succedd. 2675 */ 2676 preempt_disable(); 2677 c = __this_cpu_ptr(s->cpu_slab); 2678 2679 tid = c->tid; 2680 preempt_enable(); 2681 2682 if (likely(page == c->page)) { 2683 set_freepointer(s, object, c->freelist); 2684 2685 if (unlikely(!this_cpu_cmpxchg_double( 2686 s->cpu_slab->freelist, s->cpu_slab->tid, 2687 c->freelist, tid, 2688 object, next_tid(tid)))) { 2689 2690 note_cmpxchg_failure("slab_free", s, tid); 2691 goto redo; 2692 } 2693 stat(s, FREE_FASTPATH); 2694 } else 2695 __slab_free(s, page, x, addr); 2696 2697 } 2698 2699 void kmem_cache_free(struct kmem_cache *s, void *x) 2700 { 2701 s = cache_from_obj(s, x); 2702 if (!s) 2703 return; 2704 slab_free(s, virt_to_head_page(x), x, _RET_IP_); 2705 trace_kmem_cache_free(_RET_IP_, x); 2706 } 2707 EXPORT_SYMBOL(kmem_cache_free); 2708 2709 /* 2710 * Object placement in a slab is made very easy because we always start at 2711 * offset 0. If we tune the size of the object to the alignment then we can 2712 * get the required alignment by putting one properly sized object after 2713 * another. 2714 * 2715 * Notice that the allocation order determines the sizes of the per cpu 2716 * caches. Each processor has always one slab available for allocations. 2717 * Increasing the allocation order reduces the number of times that slabs 2718 * must be moved on and off the partial lists and is therefore a factor in 2719 * locking overhead. 2720 */ 2721 2722 /* 2723 * Mininum / Maximum order of slab pages. This influences locking overhead 2724 * and slab fragmentation. A higher order reduces the number of partial slabs 2725 * and increases the number of allocations possible without having to 2726 * take the list_lock. 2727 */ 2728 static int slub_min_order; 2729 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 2730 static int slub_min_objects; 2731 2732 /* 2733 * Merge control. If this is set then no merging of slab caches will occur. 2734 * (Could be removed. This was introduced to pacify the merge skeptics.) 2735 */ 2736 static int slub_nomerge; 2737 2738 /* 2739 * Calculate the order of allocation given an slab object size. 2740 * 2741 * The order of allocation has significant impact on performance and other 2742 * system components. Generally order 0 allocations should be preferred since 2743 * order 0 does not cause fragmentation in the page allocator. Larger objects 2744 * be problematic to put into order 0 slabs because there may be too much 2745 * unused space left. We go to a higher order if more than 1/16th of the slab 2746 * would be wasted. 2747 * 2748 * In order to reach satisfactory performance we must ensure that a minimum 2749 * number of objects is in one slab. Otherwise we may generate too much 2750 * activity on the partial lists which requires taking the list_lock. This is 2751 * less a concern for large slabs though which are rarely used. 2752 * 2753 * slub_max_order specifies the order where we begin to stop considering the 2754 * number of objects in a slab as critical. If we reach slub_max_order then 2755 * we try to keep the page order as low as possible. So we accept more waste 2756 * of space in favor of a small page order. 2757 * 2758 * Higher order allocations also allow the placement of more objects in a 2759 * slab and thereby reduce object handling overhead. If the user has 2760 * requested a higher mininum order then we start with that one instead of 2761 * the smallest order which will fit the object. 2762 */ 2763 static inline int slab_order(int size, int min_objects, 2764 int max_order, int fract_leftover, int reserved) 2765 { 2766 int order; 2767 int rem; 2768 int min_order = slub_min_order; 2769 2770 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) 2771 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 2772 2773 for (order = max(min_order, 2774 fls(min_objects * size - 1) - PAGE_SHIFT); 2775 order <= max_order; order++) { 2776 2777 unsigned long slab_size = PAGE_SIZE << order; 2778 2779 if (slab_size < min_objects * size + reserved) 2780 continue; 2781 2782 rem = (slab_size - reserved) % size; 2783 2784 if (rem <= slab_size / fract_leftover) 2785 break; 2786 2787 } 2788 2789 return order; 2790 } 2791 2792 static inline int calculate_order(int size, int reserved) 2793 { 2794 int order; 2795 int min_objects; 2796 int fraction; 2797 int max_objects; 2798 2799 /* 2800 * Attempt to find best configuration for a slab. This 2801 * works by first attempting to generate a layout with 2802 * the best configuration and backing off gradually. 2803 * 2804 * First we reduce the acceptable waste in a slab. Then 2805 * we reduce the minimum objects required in a slab. 2806 */ 2807 min_objects = slub_min_objects; 2808 if (!min_objects) 2809 min_objects = 4 * (fls(nr_cpu_ids) + 1); 2810 max_objects = order_objects(slub_max_order, size, reserved); 2811 min_objects = min(min_objects, max_objects); 2812 2813 while (min_objects > 1) { 2814 fraction = 16; 2815 while (fraction >= 4) { 2816 order = slab_order(size, min_objects, 2817 slub_max_order, fraction, reserved); 2818 if (order <= slub_max_order) 2819 return order; 2820 fraction /= 2; 2821 } 2822 min_objects--; 2823 } 2824 2825 /* 2826 * We were unable to place multiple objects in a slab. Now 2827 * lets see if we can place a single object there. 2828 */ 2829 order = slab_order(size, 1, slub_max_order, 1, reserved); 2830 if (order <= slub_max_order) 2831 return order; 2832 2833 /* 2834 * Doh this slab cannot be placed using slub_max_order. 2835 */ 2836 order = slab_order(size, 1, MAX_ORDER, 1, reserved); 2837 if (order < MAX_ORDER) 2838 return order; 2839 return -ENOSYS; 2840 } 2841 2842 static void 2843 init_kmem_cache_node(struct kmem_cache_node *n) 2844 { 2845 n->nr_partial = 0; 2846 spin_lock_init(&n->list_lock); 2847 INIT_LIST_HEAD(&n->partial); 2848 #ifdef CONFIG_SLUB_DEBUG 2849 atomic_long_set(&n->nr_slabs, 0); 2850 atomic_long_set(&n->total_objects, 0); 2851 INIT_LIST_HEAD(&n->full); 2852 #endif 2853 } 2854 2855 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 2856 { 2857 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 2858 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu)); 2859 2860 /* 2861 * Must align to double word boundary for the double cmpxchg 2862 * instructions to work; see __pcpu_double_call_return_bool(). 2863 */ 2864 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2865 2 * sizeof(void *)); 2866 2867 if (!s->cpu_slab) 2868 return 0; 2869 2870 init_kmem_cache_cpus(s); 2871 2872 return 1; 2873 } 2874 2875 static struct kmem_cache *kmem_cache_node; 2876 2877 /* 2878 * No kmalloc_node yet so do it by hand. We know that this is the first 2879 * slab on the node for this slabcache. There are no concurrent accesses 2880 * possible. 2881 * 2882 * Note that this function only works on the kmem_cache_node 2883 * when allocating for the kmem_cache_node. This is used for bootstrapping 2884 * memory on a fresh node that has no slab structures yet. 2885 */ 2886 static void early_kmem_cache_node_alloc(int node) 2887 { 2888 struct page *page; 2889 struct kmem_cache_node *n; 2890 2891 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 2892 2893 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 2894 2895 BUG_ON(!page); 2896 if (page_to_nid(page) != node) { 2897 printk(KERN_ERR "SLUB: Unable to allocate memory from " 2898 "node %d\n", node); 2899 printk(KERN_ERR "SLUB: Allocating a useless per node structure " 2900 "in order to be able to continue\n"); 2901 } 2902 2903 n = page->freelist; 2904 BUG_ON(!n); 2905 page->freelist = get_freepointer(kmem_cache_node, n); 2906 page->inuse = 1; 2907 page->frozen = 0; 2908 kmem_cache_node->node[node] = n; 2909 #ifdef CONFIG_SLUB_DEBUG 2910 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 2911 init_tracking(kmem_cache_node, n); 2912 #endif 2913 init_kmem_cache_node(n); 2914 inc_slabs_node(kmem_cache_node, node, page->objects); 2915 2916 /* 2917 * No locks need to be taken here as it has just been 2918 * initialized and there is no concurrent access. 2919 */ 2920 __add_partial(n, page, DEACTIVATE_TO_HEAD); 2921 } 2922 2923 static void free_kmem_cache_nodes(struct kmem_cache *s) 2924 { 2925 int node; 2926 2927 for_each_node_state(node, N_NORMAL_MEMORY) { 2928 struct kmem_cache_node *n = s->node[node]; 2929 2930 if (n) 2931 kmem_cache_free(kmem_cache_node, n); 2932 2933 s->node[node] = NULL; 2934 } 2935 } 2936 2937 static int init_kmem_cache_nodes(struct kmem_cache *s) 2938 { 2939 int node; 2940 2941 for_each_node_state(node, N_NORMAL_MEMORY) { 2942 struct kmem_cache_node *n; 2943 2944 if (slab_state == DOWN) { 2945 early_kmem_cache_node_alloc(node); 2946 continue; 2947 } 2948 n = kmem_cache_alloc_node(kmem_cache_node, 2949 GFP_KERNEL, node); 2950 2951 if (!n) { 2952 free_kmem_cache_nodes(s); 2953 return 0; 2954 } 2955 2956 s->node[node] = n; 2957 init_kmem_cache_node(n); 2958 } 2959 return 1; 2960 } 2961 2962 static void set_min_partial(struct kmem_cache *s, unsigned long min) 2963 { 2964 if (min < MIN_PARTIAL) 2965 min = MIN_PARTIAL; 2966 else if (min > MAX_PARTIAL) 2967 min = MAX_PARTIAL; 2968 s->min_partial = min; 2969 } 2970 2971 /* 2972 * calculate_sizes() determines the order and the distribution of data within 2973 * a slab object. 2974 */ 2975 static int calculate_sizes(struct kmem_cache *s, int forced_order) 2976 { 2977 unsigned long flags = s->flags; 2978 unsigned long size = s->object_size; 2979 int order; 2980 2981 /* 2982 * Round up object size to the next word boundary. We can only 2983 * place the free pointer at word boundaries and this determines 2984 * the possible location of the free pointer. 2985 */ 2986 size = ALIGN(size, sizeof(void *)); 2987 2988 #ifdef CONFIG_SLUB_DEBUG 2989 /* 2990 * Determine if we can poison the object itself. If the user of 2991 * the slab may touch the object after free or before allocation 2992 * then we should never poison the object itself. 2993 */ 2994 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 2995 !s->ctor) 2996 s->flags |= __OBJECT_POISON; 2997 else 2998 s->flags &= ~__OBJECT_POISON; 2999 3000 3001 /* 3002 * If we are Redzoning then check if there is some space between the 3003 * end of the object and the free pointer. If not then add an 3004 * additional word to have some bytes to store Redzone information. 3005 */ 3006 if ((flags & SLAB_RED_ZONE) && size == s->object_size) 3007 size += sizeof(void *); 3008 #endif 3009 3010 /* 3011 * With that we have determined the number of bytes in actual use 3012 * by the object. This is the potential offset to the free pointer. 3013 */ 3014 s->inuse = size; 3015 3016 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 3017 s->ctor)) { 3018 /* 3019 * Relocate free pointer after the object if it is not 3020 * permitted to overwrite the first word of the object on 3021 * kmem_cache_free. 3022 * 3023 * This is the case if we do RCU, have a constructor or 3024 * destructor or are poisoning the objects. 3025 */ 3026 s->offset = size; 3027 size += sizeof(void *); 3028 } 3029 3030 #ifdef CONFIG_SLUB_DEBUG 3031 if (flags & SLAB_STORE_USER) 3032 /* 3033 * Need to store information about allocs and frees after 3034 * the object. 3035 */ 3036 size += 2 * sizeof(struct track); 3037 3038 if (flags & SLAB_RED_ZONE) 3039 /* 3040 * Add some empty padding so that we can catch 3041 * overwrites from earlier objects rather than let 3042 * tracking information or the free pointer be 3043 * corrupted if a user writes before the start 3044 * of the object. 3045 */ 3046 size += sizeof(void *); 3047 #endif 3048 3049 /* 3050 * SLUB stores one object immediately after another beginning from 3051 * offset 0. In order to align the objects we have to simply size 3052 * each object to conform to the alignment. 3053 */ 3054 size = ALIGN(size, s->align); 3055 s->size = size; 3056 if (forced_order >= 0) 3057 order = forced_order; 3058 else 3059 order = calculate_order(size, s->reserved); 3060 3061 if (order < 0) 3062 return 0; 3063 3064 s->allocflags = 0; 3065 if (order) 3066 s->allocflags |= __GFP_COMP; 3067 3068 if (s->flags & SLAB_CACHE_DMA) 3069 s->allocflags |= GFP_DMA; 3070 3071 if (s->flags & SLAB_RECLAIM_ACCOUNT) 3072 s->allocflags |= __GFP_RECLAIMABLE; 3073 3074 /* 3075 * Determine the number of objects per slab 3076 */ 3077 s->oo = oo_make(order, size, s->reserved); 3078 s->min = oo_make(get_order(size), size, s->reserved); 3079 if (oo_objects(s->oo) > oo_objects(s->max)) 3080 s->max = s->oo; 3081 3082 return !!oo_objects(s->oo); 3083 } 3084 3085 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags) 3086 { 3087 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor); 3088 s->reserved = 0; 3089 3090 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) 3091 s->reserved = sizeof(struct rcu_head); 3092 3093 if (!calculate_sizes(s, -1)) 3094 goto error; 3095 if (disable_higher_order_debug) { 3096 /* 3097 * Disable debugging flags that store metadata if the min slab 3098 * order increased. 3099 */ 3100 if (get_order(s->size) > get_order(s->object_size)) { 3101 s->flags &= ~DEBUG_METADATA_FLAGS; 3102 s->offset = 0; 3103 if (!calculate_sizes(s, -1)) 3104 goto error; 3105 } 3106 } 3107 3108 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 3109 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 3110 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0) 3111 /* Enable fast mode */ 3112 s->flags |= __CMPXCHG_DOUBLE; 3113 #endif 3114 3115 /* 3116 * The larger the object size is, the more pages we want on the partial 3117 * list to avoid pounding the page allocator excessively. 3118 */ 3119 set_min_partial(s, ilog2(s->size) / 2); 3120 3121 /* 3122 * cpu_partial determined the maximum number of objects kept in the 3123 * per cpu partial lists of a processor. 3124 * 3125 * Per cpu partial lists mainly contain slabs that just have one 3126 * object freed. If they are used for allocation then they can be 3127 * filled up again with minimal effort. The slab will never hit the 3128 * per node partial lists and therefore no locking will be required. 3129 * 3130 * This setting also determines 3131 * 3132 * A) The number of objects from per cpu partial slabs dumped to the 3133 * per node list when we reach the limit. 3134 * B) The number of objects in cpu partial slabs to extract from the 3135 * per node list when we run out of per cpu objects. We only fetch 3136 * 50% to keep some capacity around for frees. 3137 */ 3138 if (!kmem_cache_has_cpu_partial(s)) 3139 s->cpu_partial = 0; 3140 else if (s->size >= PAGE_SIZE) 3141 s->cpu_partial = 2; 3142 else if (s->size >= 1024) 3143 s->cpu_partial = 6; 3144 else if (s->size >= 256) 3145 s->cpu_partial = 13; 3146 else 3147 s->cpu_partial = 30; 3148 3149 #ifdef CONFIG_NUMA 3150 s->remote_node_defrag_ratio = 1000; 3151 #endif 3152 if (!init_kmem_cache_nodes(s)) 3153 goto error; 3154 3155 if (alloc_kmem_cache_cpus(s)) 3156 return 0; 3157 3158 free_kmem_cache_nodes(s); 3159 error: 3160 if (flags & SLAB_PANIC) 3161 panic("Cannot create slab %s size=%lu realsize=%u " 3162 "order=%u offset=%u flags=%lx\n", 3163 s->name, (unsigned long)s->size, s->size, 3164 oo_order(s->oo), s->offset, flags); 3165 return -EINVAL; 3166 } 3167 3168 static void list_slab_objects(struct kmem_cache *s, struct page *page, 3169 const char *text) 3170 { 3171 #ifdef CONFIG_SLUB_DEBUG 3172 void *addr = page_address(page); 3173 void *p; 3174 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * 3175 sizeof(long), GFP_ATOMIC); 3176 if (!map) 3177 return; 3178 slab_err(s, page, text, s->name); 3179 slab_lock(page); 3180 3181 get_map(s, page, map); 3182 for_each_object(p, s, addr, page->objects) { 3183 3184 if (!test_bit(slab_index(p, s, addr), map)) { 3185 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n", 3186 p, p - addr); 3187 print_tracking(s, p); 3188 } 3189 } 3190 slab_unlock(page); 3191 kfree(map); 3192 #endif 3193 } 3194 3195 /* 3196 * Attempt to free all partial slabs on a node. 3197 * This is called from kmem_cache_close(). We must be the last thread 3198 * using the cache and therefore we do not need to lock anymore. 3199 */ 3200 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 3201 { 3202 struct page *page, *h; 3203 3204 list_for_each_entry_safe(page, h, &n->partial, lru) { 3205 if (!page->inuse) { 3206 __remove_partial(n, page); 3207 discard_slab(s, page); 3208 } else { 3209 list_slab_objects(s, page, 3210 "Objects remaining in %s on kmem_cache_close()"); 3211 } 3212 } 3213 } 3214 3215 /* 3216 * Release all resources used by a slab cache. 3217 */ 3218 static inline int kmem_cache_close(struct kmem_cache *s) 3219 { 3220 int node; 3221 3222 flush_all(s); 3223 /* Attempt to free all objects */ 3224 for_each_node_state(node, N_NORMAL_MEMORY) { 3225 struct kmem_cache_node *n = get_node(s, node); 3226 3227 free_partial(s, n); 3228 if (n->nr_partial || slabs_node(s, node)) 3229 return 1; 3230 } 3231 free_percpu(s->cpu_slab); 3232 free_kmem_cache_nodes(s); 3233 return 0; 3234 } 3235 3236 int __kmem_cache_shutdown(struct kmem_cache *s) 3237 { 3238 return kmem_cache_close(s); 3239 } 3240 3241 /******************************************************************** 3242 * Kmalloc subsystem 3243 *******************************************************************/ 3244 3245 static int __init setup_slub_min_order(char *str) 3246 { 3247 get_option(&str, &slub_min_order); 3248 3249 return 1; 3250 } 3251 3252 __setup("slub_min_order=", setup_slub_min_order); 3253 3254 static int __init setup_slub_max_order(char *str) 3255 { 3256 get_option(&str, &slub_max_order); 3257 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 3258 3259 return 1; 3260 } 3261 3262 __setup("slub_max_order=", setup_slub_max_order); 3263 3264 static int __init setup_slub_min_objects(char *str) 3265 { 3266 get_option(&str, &slub_min_objects); 3267 3268 return 1; 3269 } 3270 3271 __setup("slub_min_objects=", setup_slub_min_objects); 3272 3273 static int __init setup_slub_nomerge(char *str) 3274 { 3275 slub_nomerge = 1; 3276 return 1; 3277 } 3278 3279 __setup("slub_nomerge", setup_slub_nomerge); 3280 3281 void *__kmalloc(size_t size, gfp_t flags) 3282 { 3283 struct kmem_cache *s; 3284 void *ret; 3285 3286 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 3287 return kmalloc_large(size, flags); 3288 3289 s = kmalloc_slab(size, flags); 3290 3291 if (unlikely(ZERO_OR_NULL_PTR(s))) 3292 return s; 3293 3294 ret = slab_alloc(s, flags, _RET_IP_); 3295 3296 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 3297 3298 return ret; 3299 } 3300 EXPORT_SYMBOL(__kmalloc); 3301 3302 #ifdef CONFIG_NUMA 3303 static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 3304 { 3305 struct page *page; 3306 void *ptr = NULL; 3307 3308 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG; 3309 page = alloc_pages_node(node, flags, get_order(size)); 3310 if (page) 3311 ptr = page_address(page); 3312 3313 kmalloc_large_node_hook(ptr, size, flags); 3314 return ptr; 3315 } 3316 3317 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3318 { 3319 struct kmem_cache *s; 3320 void *ret; 3321 3322 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 3323 ret = kmalloc_large_node(size, flags, node); 3324 3325 trace_kmalloc_node(_RET_IP_, ret, 3326 size, PAGE_SIZE << get_order(size), 3327 flags, node); 3328 3329 return ret; 3330 } 3331 3332 s = kmalloc_slab(size, flags); 3333 3334 if (unlikely(ZERO_OR_NULL_PTR(s))) 3335 return s; 3336 3337 ret = slab_alloc_node(s, flags, node, _RET_IP_); 3338 3339 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 3340 3341 return ret; 3342 } 3343 EXPORT_SYMBOL(__kmalloc_node); 3344 #endif 3345 3346 size_t ksize(const void *object) 3347 { 3348 struct page *page; 3349 3350 if (unlikely(object == ZERO_SIZE_PTR)) 3351 return 0; 3352 3353 page = virt_to_head_page(object); 3354 3355 if (unlikely(!PageSlab(page))) { 3356 WARN_ON(!PageCompound(page)); 3357 return PAGE_SIZE << compound_order(page); 3358 } 3359 3360 return slab_ksize(page->slab_cache); 3361 } 3362 EXPORT_SYMBOL(ksize); 3363 3364 void kfree(const void *x) 3365 { 3366 struct page *page; 3367 void *object = (void *)x; 3368 3369 trace_kfree(_RET_IP_, x); 3370 3371 if (unlikely(ZERO_OR_NULL_PTR(x))) 3372 return; 3373 3374 page = virt_to_head_page(x); 3375 if (unlikely(!PageSlab(page))) { 3376 BUG_ON(!PageCompound(page)); 3377 kfree_hook(x); 3378 __free_memcg_kmem_pages(page, compound_order(page)); 3379 return; 3380 } 3381 slab_free(page->slab_cache, page, object, _RET_IP_); 3382 } 3383 EXPORT_SYMBOL(kfree); 3384 3385 /* 3386 * kmem_cache_shrink removes empty slabs from the partial lists and sorts 3387 * the remaining slabs by the number of items in use. The slabs with the 3388 * most items in use come first. New allocations will then fill those up 3389 * and thus they can be removed from the partial lists. 3390 * 3391 * The slabs with the least items are placed last. This results in them 3392 * being allocated from last increasing the chance that the last objects 3393 * are freed in them. 3394 */ 3395 int kmem_cache_shrink(struct kmem_cache *s) 3396 { 3397 int node; 3398 int i; 3399 struct kmem_cache_node *n; 3400 struct page *page; 3401 struct page *t; 3402 int objects = oo_objects(s->max); 3403 struct list_head *slabs_by_inuse = 3404 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL); 3405 unsigned long flags; 3406 3407 if (!slabs_by_inuse) 3408 return -ENOMEM; 3409 3410 flush_all(s); 3411 for_each_node_state(node, N_NORMAL_MEMORY) { 3412 n = get_node(s, node); 3413 3414 if (!n->nr_partial) 3415 continue; 3416 3417 for (i = 0; i < objects; i++) 3418 INIT_LIST_HEAD(slabs_by_inuse + i); 3419 3420 spin_lock_irqsave(&n->list_lock, flags); 3421 3422 /* 3423 * Build lists indexed by the items in use in each slab. 3424 * 3425 * Note that concurrent frees may occur while we hold the 3426 * list_lock. page->inuse here is the upper limit. 3427 */ 3428 list_for_each_entry_safe(page, t, &n->partial, lru) { 3429 list_move(&page->lru, slabs_by_inuse + page->inuse); 3430 if (!page->inuse) 3431 n->nr_partial--; 3432 } 3433 3434 /* 3435 * Rebuild the partial list with the slabs filled up most 3436 * first and the least used slabs at the end. 3437 */ 3438 for (i = objects - 1; i > 0; i--) 3439 list_splice(slabs_by_inuse + i, n->partial.prev); 3440 3441 spin_unlock_irqrestore(&n->list_lock, flags); 3442 3443 /* Release empty slabs */ 3444 list_for_each_entry_safe(page, t, slabs_by_inuse, lru) 3445 discard_slab(s, page); 3446 } 3447 3448 kfree(slabs_by_inuse); 3449 return 0; 3450 } 3451 EXPORT_SYMBOL(kmem_cache_shrink); 3452 3453 static int slab_mem_going_offline_callback(void *arg) 3454 { 3455 struct kmem_cache *s; 3456 3457 mutex_lock(&slab_mutex); 3458 list_for_each_entry(s, &slab_caches, list) 3459 kmem_cache_shrink(s); 3460 mutex_unlock(&slab_mutex); 3461 3462 return 0; 3463 } 3464 3465 static void slab_mem_offline_callback(void *arg) 3466 { 3467 struct kmem_cache_node *n; 3468 struct kmem_cache *s; 3469 struct memory_notify *marg = arg; 3470 int offline_node; 3471 3472 offline_node = marg->status_change_nid_normal; 3473 3474 /* 3475 * If the node still has available memory. we need kmem_cache_node 3476 * for it yet. 3477 */ 3478 if (offline_node < 0) 3479 return; 3480 3481 mutex_lock(&slab_mutex); 3482 list_for_each_entry(s, &slab_caches, list) { 3483 n = get_node(s, offline_node); 3484 if (n) { 3485 /* 3486 * if n->nr_slabs > 0, slabs still exist on the node 3487 * that is going down. We were unable to free them, 3488 * and offline_pages() function shouldn't call this 3489 * callback. So, we must fail. 3490 */ 3491 BUG_ON(slabs_node(s, offline_node)); 3492 3493 s->node[offline_node] = NULL; 3494 kmem_cache_free(kmem_cache_node, n); 3495 } 3496 } 3497 mutex_unlock(&slab_mutex); 3498 } 3499 3500 static int slab_mem_going_online_callback(void *arg) 3501 { 3502 struct kmem_cache_node *n; 3503 struct kmem_cache *s; 3504 struct memory_notify *marg = arg; 3505 int nid = marg->status_change_nid_normal; 3506 int ret = 0; 3507 3508 /* 3509 * If the node's memory is already available, then kmem_cache_node is 3510 * already created. Nothing to do. 3511 */ 3512 if (nid < 0) 3513 return 0; 3514 3515 /* 3516 * We are bringing a node online. No memory is available yet. We must 3517 * allocate a kmem_cache_node structure in order to bring the node 3518 * online. 3519 */ 3520 mutex_lock(&slab_mutex); 3521 list_for_each_entry(s, &slab_caches, list) { 3522 /* 3523 * XXX: kmem_cache_alloc_node will fallback to other nodes 3524 * since memory is not yet available from the node that 3525 * is brought up. 3526 */ 3527 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 3528 if (!n) { 3529 ret = -ENOMEM; 3530 goto out; 3531 } 3532 init_kmem_cache_node(n); 3533 s->node[nid] = n; 3534 } 3535 out: 3536 mutex_unlock(&slab_mutex); 3537 return ret; 3538 } 3539 3540 static int slab_memory_callback(struct notifier_block *self, 3541 unsigned long action, void *arg) 3542 { 3543 int ret = 0; 3544 3545 switch (action) { 3546 case MEM_GOING_ONLINE: 3547 ret = slab_mem_going_online_callback(arg); 3548 break; 3549 case MEM_GOING_OFFLINE: 3550 ret = slab_mem_going_offline_callback(arg); 3551 break; 3552 case MEM_OFFLINE: 3553 case MEM_CANCEL_ONLINE: 3554 slab_mem_offline_callback(arg); 3555 break; 3556 case MEM_ONLINE: 3557 case MEM_CANCEL_OFFLINE: 3558 break; 3559 } 3560 if (ret) 3561 ret = notifier_from_errno(ret); 3562 else 3563 ret = NOTIFY_OK; 3564 return ret; 3565 } 3566 3567 static struct notifier_block slab_memory_callback_nb = { 3568 .notifier_call = slab_memory_callback, 3569 .priority = SLAB_CALLBACK_PRI, 3570 }; 3571 3572 /******************************************************************** 3573 * Basic setup of slabs 3574 *******************************************************************/ 3575 3576 /* 3577 * Used for early kmem_cache structures that were allocated using 3578 * the page allocator. Allocate them properly then fix up the pointers 3579 * that may be pointing to the wrong kmem_cache structure. 3580 */ 3581 3582 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) 3583 { 3584 int node; 3585 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); 3586 3587 memcpy(s, static_cache, kmem_cache->object_size); 3588 3589 /* 3590 * This runs very early, and only the boot processor is supposed to be 3591 * up. Even if it weren't true, IRQs are not up so we couldn't fire 3592 * IPIs around. 3593 */ 3594 __flush_cpu_slab(s, smp_processor_id()); 3595 for_each_node_state(node, N_NORMAL_MEMORY) { 3596 struct kmem_cache_node *n = get_node(s, node); 3597 struct page *p; 3598 3599 if (n) { 3600 list_for_each_entry(p, &n->partial, lru) 3601 p->slab_cache = s; 3602 3603 #ifdef CONFIG_SLUB_DEBUG 3604 list_for_each_entry(p, &n->full, lru) 3605 p->slab_cache = s; 3606 #endif 3607 } 3608 } 3609 list_add(&s->list, &slab_caches); 3610 return s; 3611 } 3612 3613 void __init kmem_cache_init(void) 3614 { 3615 static __initdata struct kmem_cache boot_kmem_cache, 3616 boot_kmem_cache_node; 3617 3618 if (debug_guardpage_minorder()) 3619 slub_max_order = 0; 3620 3621 kmem_cache_node = &boot_kmem_cache_node; 3622 kmem_cache = &boot_kmem_cache; 3623 3624 create_boot_cache(kmem_cache_node, "kmem_cache_node", 3625 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN); 3626 3627 register_hotmemory_notifier(&slab_memory_callback_nb); 3628 3629 /* Able to allocate the per node structures */ 3630 slab_state = PARTIAL; 3631 3632 create_boot_cache(kmem_cache, "kmem_cache", 3633 offsetof(struct kmem_cache, node) + 3634 nr_node_ids * sizeof(struct kmem_cache_node *), 3635 SLAB_HWCACHE_ALIGN); 3636 3637 kmem_cache = bootstrap(&boot_kmem_cache); 3638 3639 /* 3640 * Allocate kmem_cache_node properly from the kmem_cache slab. 3641 * kmem_cache_node is separately allocated so no need to 3642 * update any list pointers. 3643 */ 3644 kmem_cache_node = bootstrap(&boot_kmem_cache_node); 3645 3646 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 3647 create_kmalloc_caches(0); 3648 3649 #ifdef CONFIG_SMP 3650 register_cpu_notifier(&slab_notifier); 3651 #endif 3652 3653 printk(KERN_INFO 3654 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d," 3655 " CPUs=%d, Nodes=%d\n", 3656 cache_line_size(), 3657 slub_min_order, slub_max_order, slub_min_objects, 3658 nr_cpu_ids, nr_node_ids); 3659 } 3660 3661 void __init kmem_cache_init_late(void) 3662 { 3663 } 3664 3665 /* 3666 * Find a mergeable slab cache 3667 */ 3668 static int slab_unmergeable(struct kmem_cache *s) 3669 { 3670 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE)) 3671 return 1; 3672 3673 if (!is_root_cache(s)) 3674 return 1; 3675 3676 if (s->ctor) 3677 return 1; 3678 3679 /* 3680 * We may have set a slab to be unmergeable during bootstrap. 3681 */ 3682 if (s->refcount < 0) 3683 return 1; 3684 3685 return 0; 3686 } 3687 3688 static struct kmem_cache *find_mergeable(size_t size, size_t align, 3689 unsigned long flags, const char *name, void (*ctor)(void *)) 3690 { 3691 struct kmem_cache *s; 3692 3693 if (slub_nomerge || (flags & SLUB_NEVER_MERGE)) 3694 return NULL; 3695 3696 if (ctor) 3697 return NULL; 3698 3699 size = ALIGN(size, sizeof(void *)); 3700 align = calculate_alignment(flags, align, size); 3701 size = ALIGN(size, align); 3702 flags = kmem_cache_flags(size, flags, name, NULL); 3703 3704 list_for_each_entry(s, &slab_caches, list) { 3705 if (slab_unmergeable(s)) 3706 continue; 3707 3708 if (size > s->size) 3709 continue; 3710 3711 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME)) 3712 continue; 3713 /* 3714 * Check if alignment is compatible. 3715 * Courtesy of Adrian Drzewiecki 3716 */ 3717 if ((s->size & ~(align - 1)) != s->size) 3718 continue; 3719 3720 if (s->size - size >= sizeof(void *)) 3721 continue; 3722 3723 return s; 3724 } 3725 return NULL; 3726 } 3727 3728 struct kmem_cache * 3729 __kmem_cache_alias(const char *name, size_t size, size_t align, 3730 unsigned long flags, void (*ctor)(void *)) 3731 { 3732 struct kmem_cache *s; 3733 3734 s = find_mergeable(size, align, flags, name, ctor); 3735 if (s) { 3736 int i; 3737 struct kmem_cache *c; 3738 3739 s->refcount++; 3740 3741 /* 3742 * Adjust the object sizes so that we clear 3743 * the complete object on kzalloc. 3744 */ 3745 s->object_size = max(s->object_size, (int)size); 3746 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 3747 3748 for_each_memcg_cache_index(i) { 3749 c = cache_from_memcg_idx(s, i); 3750 if (!c) 3751 continue; 3752 c->object_size = s->object_size; 3753 c->inuse = max_t(int, c->inuse, 3754 ALIGN(size, sizeof(void *))); 3755 } 3756 3757 if (sysfs_slab_alias(s, name)) { 3758 s->refcount--; 3759 s = NULL; 3760 } 3761 } 3762 3763 return s; 3764 } 3765 3766 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags) 3767 { 3768 int err; 3769 3770 err = kmem_cache_open(s, flags); 3771 if (err) 3772 return err; 3773 3774 /* Mutex is not taken during early boot */ 3775 if (slab_state <= UP) 3776 return 0; 3777 3778 memcg_propagate_slab_attrs(s); 3779 err = sysfs_slab_add(s); 3780 if (err) 3781 kmem_cache_close(s); 3782 3783 return err; 3784 } 3785 3786 #ifdef CONFIG_SMP 3787 /* 3788 * Use the cpu notifier to insure that the cpu slabs are flushed when 3789 * necessary. 3790 */ 3791 static int slab_cpuup_callback(struct notifier_block *nfb, 3792 unsigned long action, void *hcpu) 3793 { 3794 long cpu = (long)hcpu; 3795 struct kmem_cache *s; 3796 unsigned long flags; 3797 3798 switch (action) { 3799 case CPU_UP_CANCELED: 3800 case CPU_UP_CANCELED_FROZEN: 3801 case CPU_DEAD: 3802 case CPU_DEAD_FROZEN: 3803 mutex_lock(&slab_mutex); 3804 list_for_each_entry(s, &slab_caches, list) { 3805 local_irq_save(flags); 3806 __flush_cpu_slab(s, cpu); 3807 local_irq_restore(flags); 3808 } 3809 mutex_unlock(&slab_mutex); 3810 break; 3811 default: 3812 break; 3813 } 3814 return NOTIFY_OK; 3815 } 3816 3817 static struct notifier_block slab_notifier = { 3818 .notifier_call = slab_cpuup_callback 3819 }; 3820 3821 #endif 3822 3823 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 3824 { 3825 struct kmem_cache *s; 3826 void *ret; 3827 3828 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 3829 return kmalloc_large(size, gfpflags); 3830 3831 s = kmalloc_slab(size, gfpflags); 3832 3833 if (unlikely(ZERO_OR_NULL_PTR(s))) 3834 return s; 3835 3836 ret = slab_alloc(s, gfpflags, caller); 3837 3838 /* Honor the call site pointer we received. */ 3839 trace_kmalloc(caller, ret, size, s->size, gfpflags); 3840 3841 return ret; 3842 } 3843 3844 #ifdef CONFIG_NUMA 3845 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 3846 int node, unsigned long caller) 3847 { 3848 struct kmem_cache *s; 3849 void *ret; 3850 3851 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 3852 ret = kmalloc_large_node(size, gfpflags, node); 3853 3854 trace_kmalloc_node(caller, ret, 3855 size, PAGE_SIZE << get_order(size), 3856 gfpflags, node); 3857 3858 return ret; 3859 } 3860 3861 s = kmalloc_slab(size, gfpflags); 3862 3863 if (unlikely(ZERO_OR_NULL_PTR(s))) 3864 return s; 3865 3866 ret = slab_alloc_node(s, gfpflags, node, caller); 3867 3868 /* Honor the call site pointer we received. */ 3869 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 3870 3871 return ret; 3872 } 3873 #endif 3874 3875 #ifdef CONFIG_SYSFS 3876 static int count_inuse(struct page *page) 3877 { 3878 return page->inuse; 3879 } 3880 3881 static int count_total(struct page *page) 3882 { 3883 return page->objects; 3884 } 3885 #endif 3886 3887 #ifdef CONFIG_SLUB_DEBUG 3888 static int validate_slab(struct kmem_cache *s, struct page *page, 3889 unsigned long *map) 3890 { 3891 void *p; 3892 void *addr = page_address(page); 3893 3894 if (!check_slab(s, page) || 3895 !on_freelist(s, page, NULL)) 3896 return 0; 3897 3898 /* Now we know that a valid freelist exists */ 3899 bitmap_zero(map, page->objects); 3900 3901 get_map(s, page, map); 3902 for_each_object(p, s, addr, page->objects) { 3903 if (test_bit(slab_index(p, s, addr), map)) 3904 if (!check_object(s, page, p, SLUB_RED_INACTIVE)) 3905 return 0; 3906 } 3907 3908 for_each_object(p, s, addr, page->objects) 3909 if (!test_bit(slab_index(p, s, addr), map)) 3910 if (!check_object(s, page, p, SLUB_RED_ACTIVE)) 3911 return 0; 3912 return 1; 3913 } 3914 3915 static void validate_slab_slab(struct kmem_cache *s, struct page *page, 3916 unsigned long *map) 3917 { 3918 slab_lock(page); 3919 validate_slab(s, page, map); 3920 slab_unlock(page); 3921 } 3922 3923 static int validate_slab_node(struct kmem_cache *s, 3924 struct kmem_cache_node *n, unsigned long *map) 3925 { 3926 unsigned long count = 0; 3927 struct page *page; 3928 unsigned long flags; 3929 3930 spin_lock_irqsave(&n->list_lock, flags); 3931 3932 list_for_each_entry(page, &n->partial, lru) { 3933 validate_slab_slab(s, page, map); 3934 count++; 3935 } 3936 if (count != n->nr_partial) 3937 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but " 3938 "counter=%ld\n", s->name, count, n->nr_partial); 3939 3940 if (!(s->flags & SLAB_STORE_USER)) 3941 goto out; 3942 3943 list_for_each_entry(page, &n->full, lru) { 3944 validate_slab_slab(s, page, map); 3945 count++; 3946 } 3947 if (count != atomic_long_read(&n->nr_slabs)) 3948 printk(KERN_ERR "SLUB: %s %ld slabs counted but " 3949 "counter=%ld\n", s->name, count, 3950 atomic_long_read(&n->nr_slabs)); 3951 3952 out: 3953 spin_unlock_irqrestore(&n->list_lock, flags); 3954 return count; 3955 } 3956 3957 static long validate_slab_cache(struct kmem_cache *s) 3958 { 3959 int node; 3960 unsigned long count = 0; 3961 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 3962 sizeof(unsigned long), GFP_KERNEL); 3963 3964 if (!map) 3965 return -ENOMEM; 3966 3967 flush_all(s); 3968 for_each_node_state(node, N_NORMAL_MEMORY) { 3969 struct kmem_cache_node *n = get_node(s, node); 3970 3971 count += validate_slab_node(s, n, map); 3972 } 3973 kfree(map); 3974 return count; 3975 } 3976 /* 3977 * Generate lists of code addresses where slabcache objects are allocated 3978 * and freed. 3979 */ 3980 3981 struct location { 3982 unsigned long count; 3983 unsigned long addr; 3984 long long sum_time; 3985 long min_time; 3986 long max_time; 3987 long min_pid; 3988 long max_pid; 3989 DECLARE_BITMAP(cpus, NR_CPUS); 3990 nodemask_t nodes; 3991 }; 3992 3993 struct loc_track { 3994 unsigned long max; 3995 unsigned long count; 3996 struct location *loc; 3997 }; 3998 3999 static void free_loc_track(struct loc_track *t) 4000 { 4001 if (t->max) 4002 free_pages((unsigned long)t->loc, 4003 get_order(sizeof(struct location) * t->max)); 4004 } 4005 4006 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 4007 { 4008 struct location *l; 4009 int order; 4010 4011 order = get_order(sizeof(struct location) * max); 4012 4013 l = (void *)__get_free_pages(flags, order); 4014 if (!l) 4015 return 0; 4016 4017 if (t->count) { 4018 memcpy(l, t->loc, sizeof(struct location) * t->count); 4019 free_loc_track(t); 4020 } 4021 t->max = max; 4022 t->loc = l; 4023 return 1; 4024 } 4025 4026 static int add_location(struct loc_track *t, struct kmem_cache *s, 4027 const struct track *track) 4028 { 4029 long start, end, pos; 4030 struct location *l; 4031 unsigned long caddr; 4032 unsigned long age = jiffies - track->when; 4033 4034 start = -1; 4035 end = t->count; 4036 4037 for ( ; ; ) { 4038 pos = start + (end - start + 1) / 2; 4039 4040 /* 4041 * There is nothing at "end". If we end up there 4042 * we need to add something to before end. 4043 */ 4044 if (pos == end) 4045 break; 4046 4047 caddr = t->loc[pos].addr; 4048 if (track->addr == caddr) { 4049 4050 l = &t->loc[pos]; 4051 l->count++; 4052 if (track->when) { 4053 l->sum_time += age; 4054 if (age < l->min_time) 4055 l->min_time = age; 4056 if (age > l->max_time) 4057 l->max_time = age; 4058 4059 if (track->pid < l->min_pid) 4060 l->min_pid = track->pid; 4061 if (track->pid > l->max_pid) 4062 l->max_pid = track->pid; 4063 4064 cpumask_set_cpu(track->cpu, 4065 to_cpumask(l->cpus)); 4066 } 4067 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4068 return 1; 4069 } 4070 4071 if (track->addr < caddr) 4072 end = pos; 4073 else 4074 start = pos; 4075 } 4076 4077 /* 4078 * Not found. Insert new tracking element. 4079 */ 4080 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 4081 return 0; 4082 4083 l = t->loc + pos; 4084 if (pos < t->count) 4085 memmove(l + 1, l, 4086 (t->count - pos) * sizeof(struct location)); 4087 t->count++; 4088 l->count = 1; 4089 l->addr = track->addr; 4090 l->sum_time = age; 4091 l->min_time = age; 4092 l->max_time = age; 4093 l->min_pid = track->pid; 4094 l->max_pid = track->pid; 4095 cpumask_clear(to_cpumask(l->cpus)); 4096 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 4097 nodes_clear(l->nodes); 4098 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4099 return 1; 4100 } 4101 4102 static void process_slab(struct loc_track *t, struct kmem_cache *s, 4103 struct page *page, enum track_item alloc, 4104 unsigned long *map) 4105 { 4106 void *addr = page_address(page); 4107 void *p; 4108 4109 bitmap_zero(map, page->objects); 4110 get_map(s, page, map); 4111 4112 for_each_object(p, s, addr, page->objects) 4113 if (!test_bit(slab_index(p, s, addr), map)) 4114 add_location(t, s, get_track(s, p, alloc)); 4115 } 4116 4117 static int list_locations(struct kmem_cache *s, char *buf, 4118 enum track_item alloc) 4119 { 4120 int len = 0; 4121 unsigned long i; 4122 struct loc_track t = { 0, 0, NULL }; 4123 int node; 4124 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 4125 sizeof(unsigned long), GFP_KERNEL); 4126 4127 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 4128 GFP_TEMPORARY)) { 4129 kfree(map); 4130 return sprintf(buf, "Out of memory\n"); 4131 } 4132 /* Push back cpu slabs */ 4133 flush_all(s); 4134 4135 for_each_node_state(node, N_NORMAL_MEMORY) { 4136 struct kmem_cache_node *n = get_node(s, node); 4137 unsigned long flags; 4138 struct page *page; 4139 4140 if (!atomic_long_read(&n->nr_slabs)) 4141 continue; 4142 4143 spin_lock_irqsave(&n->list_lock, flags); 4144 list_for_each_entry(page, &n->partial, lru) 4145 process_slab(&t, s, page, alloc, map); 4146 list_for_each_entry(page, &n->full, lru) 4147 process_slab(&t, s, page, alloc, map); 4148 spin_unlock_irqrestore(&n->list_lock, flags); 4149 } 4150 4151 for (i = 0; i < t.count; i++) { 4152 struct location *l = &t.loc[i]; 4153 4154 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 4155 break; 4156 len += sprintf(buf + len, "%7ld ", l->count); 4157 4158 if (l->addr) 4159 len += sprintf(buf + len, "%pS", (void *)l->addr); 4160 else 4161 len += sprintf(buf + len, "<not-available>"); 4162 4163 if (l->sum_time != l->min_time) { 4164 len += sprintf(buf + len, " age=%ld/%ld/%ld", 4165 l->min_time, 4166 (long)div_u64(l->sum_time, l->count), 4167 l->max_time); 4168 } else 4169 len += sprintf(buf + len, " age=%ld", 4170 l->min_time); 4171 4172 if (l->min_pid != l->max_pid) 4173 len += sprintf(buf + len, " pid=%ld-%ld", 4174 l->min_pid, l->max_pid); 4175 else 4176 len += sprintf(buf + len, " pid=%ld", 4177 l->min_pid); 4178 4179 if (num_online_cpus() > 1 && 4180 !cpumask_empty(to_cpumask(l->cpus)) && 4181 len < PAGE_SIZE - 60) { 4182 len += sprintf(buf + len, " cpus="); 4183 len += cpulist_scnprintf(buf + len, 4184 PAGE_SIZE - len - 50, 4185 to_cpumask(l->cpus)); 4186 } 4187 4188 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 4189 len < PAGE_SIZE - 60) { 4190 len += sprintf(buf + len, " nodes="); 4191 len += nodelist_scnprintf(buf + len, 4192 PAGE_SIZE - len - 50, 4193 l->nodes); 4194 } 4195 4196 len += sprintf(buf + len, "\n"); 4197 } 4198 4199 free_loc_track(&t); 4200 kfree(map); 4201 if (!t.count) 4202 len += sprintf(buf, "No data\n"); 4203 return len; 4204 } 4205 #endif 4206 4207 #ifdef SLUB_RESILIENCY_TEST 4208 static void resiliency_test(void) 4209 { 4210 u8 *p; 4211 4212 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10); 4213 4214 printk(KERN_ERR "SLUB resiliency testing\n"); 4215 printk(KERN_ERR "-----------------------\n"); 4216 printk(KERN_ERR "A. Corruption after allocation\n"); 4217 4218 p = kzalloc(16, GFP_KERNEL); 4219 p[16] = 0x12; 4220 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer" 4221 " 0x12->0x%p\n\n", p + 16); 4222 4223 validate_slab_cache(kmalloc_caches[4]); 4224 4225 /* Hmmm... The next two are dangerous */ 4226 p = kzalloc(32, GFP_KERNEL); 4227 p[32 + sizeof(void *)] = 0x34; 4228 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab" 4229 " 0x34 -> -0x%p\n", p); 4230 printk(KERN_ERR 4231 "If allocated object is overwritten then not detectable\n\n"); 4232 4233 validate_slab_cache(kmalloc_caches[5]); 4234 p = kzalloc(64, GFP_KERNEL); 4235 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 4236 *p = 0x56; 4237 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 4238 p); 4239 printk(KERN_ERR 4240 "If allocated object is overwritten then not detectable\n\n"); 4241 validate_slab_cache(kmalloc_caches[6]); 4242 4243 printk(KERN_ERR "\nB. Corruption after free\n"); 4244 p = kzalloc(128, GFP_KERNEL); 4245 kfree(p); 4246 *p = 0x78; 4247 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 4248 validate_slab_cache(kmalloc_caches[7]); 4249 4250 p = kzalloc(256, GFP_KERNEL); 4251 kfree(p); 4252 p[50] = 0x9a; 4253 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", 4254 p); 4255 validate_slab_cache(kmalloc_caches[8]); 4256 4257 p = kzalloc(512, GFP_KERNEL); 4258 kfree(p); 4259 p[512] = 0xab; 4260 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 4261 validate_slab_cache(kmalloc_caches[9]); 4262 } 4263 #else 4264 #ifdef CONFIG_SYSFS 4265 static void resiliency_test(void) {}; 4266 #endif 4267 #endif 4268 4269 #ifdef CONFIG_SYSFS 4270 enum slab_stat_type { 4271 SL_ALL, /* All slabs */ 4272 SL_PARTIAL, /* Only partially allocated slabs */ 4273 SL_CPU, /* Only slabs used for cpu caches */ 4274 SL_OBJECTS, /* Determine allocated objects not slabs */ 4275 SL_TOTAL /* Determine object capacity not slabs */ 4276 }; 4277 4278 #define SO_ALL (1 << SL_ALL) 4279 #define SO_PARTIAL (1 << SL_PARTIAL) 4280 #define SO_CPU (1 << SL_CPU) 4281 #define SO_OBJECTS (1 << SL_OBJECTS) 4282 #define SO_TOTAL (1 << SL_TOTAL) 4283 4284 static ssize_t show_slab_objects(struct kmem_cache *s, 4285 char *buf, unsigned long flags) 4286 { 4287 unsigned long total = 0; 4288 int node; 4289 int x; 4290 unsigned long *nodes; 4291 4292 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 4293 if (!nodes) 4294 return -ENOMEM; 4295 4296 if (flags & SO_CPU) { 4297 int cpu; 4298 4299 for_each_possible_cpu(cpu) { 4300 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, 4301 cpu); 4302 int node; 4303 struct page *page; 4304 4305 page = ACCESS_ONCE(c->page); 4306 if (!page) 4307 continue; 4308 4309 node = page_to_nid(page); 4310 if (flags & SO_TOTAL) 4311 x = page->objects; 4312 else if (flags & SO_OBJECTS) 4313 x = page->inuse; 4314 else 4315 x = 1; 4316 4317 total += x; 4318 nodes[node] += x; 4319 4320 page = ACCESS_ONCE(c->partial); 4321 if (page) { 4322 node = page_to_nid(page); 4323 if (flags & SO_TOTAL) 4324 WARN_ON_ONCE(1); 4325 else if (flags & SO_OBJECTS) 4326 WARN_ON_ONCE(1); 4327 else 4328 x = page->pages; 4329 total += x; 4330 nodes[node] += x; 4331 } 4332 } 4333 } 4334 4335 lock_memory_hotplug(); 4336 #ifdef CONFIG_SLUB_DEBUG 4337 if (flags & SO_ALL) { 4338 for_each_node_state(node, N_NORMAL_MEMORY) { 4339 struct kmem_cache_node *n = get_node(s, node); 4340 4341 if (flags & SO_TOTAL) 4342 x = atomic_long_read(&n->total_objects); 4343 else if (flags & SO_OBJECTS) 4344 x = atomic_long_read(&n->total_objects) - 4345 count_partial(n, count_free); 4346 else 4347 x = atomic_long_read(&n->nr_slabs); 4348 total += x; 4349 nodes[node] += x; 4350 } 4351 4352 } else 4353 #endif 4354 if (flags & SO_PARTIAL) { 4355 for_each_node_state(node, N_NORMAL_MEMORY) { 4356 struct kmem_cache_node *n = get_node(s, node); 4357 4358 if (flags & SO_TOTAL) 4359 x = count_partial(n, count_total); 4360 else if (flags & SO_OBJECTS) 4361 x = count_partial(n, count_inuse); 4362 else 4363 x = n->nr_partial; 4364 total += x; 4365 nodes[node] += x; 4366 } 4367 } 4368 x = sprintf(buf, "%lu", total); 4369 #ifdef CONFIG_NUMA 4370 for_each_node_state(node, N_NORMAL_MEMORY) 4371 if (nodes[node]) 4372 x += sprintf(buf + x, " N%d=%lu", 4373 node, nodes[node]); 4374 #endif 4375 unlock_memory_hotplug(); 4376 kfree(nodes); 4377 return x + sprintf(buf + x, "\n"); 4378 } 4379 4380 #ifdef CONFIG_SLUB_DEBUG 4381 static int any_slab_objects(struct kmem_cache *s) 4382 { 4383 int node; 4384 4385 for_each_online_node(node) { 4386 struct kmem_cache_node *n = get_node(s, node); 4387 4388 if (!n) 4389 continue; 4390 4391 if (atomic_long_read(&n->total_objects)) 4392 return 1; 4393 } 4394 return 0; 4395 } 4396 #endif 4397 4398 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 4399 #define to_slab(n) container_of(n, struct kmem_cache, kobj) 4400 4401 struct slab_attribute { 4402 struct attribute attr; 4403 ssize_t (*show)(struct kmem_cache *s, char *buf); 4404 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 4405 }; 4406 4407 #define SLAB_ATTR_RO(_name) \ 4408 static struct slab_attribute _name##_attr = \ 4409 __ATTR(_name, 0400, _name##_show, NULL) 4410 4411 #define SLAB_ATTR(_name) \ 4412 static struct slab_attribute _name##_attr = \ 4413 __ATTR(_name, 0600, _name##_show, _name##_store) 4414 4415 static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 4416 { 4417 return sprintf(buf, "%d\n", s->size); 4418 } 4419 SLAB_ATTR_RO(slab_size); 4420 4421 static ssize_t align_show(struct kmem_cache *s, char *buf) 4422 { 4423 return sprintf(buf, "%d\n", s->align); 4424 } 4425 SLAB_ATTR_RO(align); 4426 4427 static ssize_t object_size_show(struct kmem_cache *s, char *buf) 4428 { 4429 return sprintf(buf, "%d\n", s->object_size); 4430 } 4431 SLAB_ATTR_RO(object_size); 4432 4433 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 4434 { 4435 return sprintf(buf, "%d\n", oo_objects(s->oo)); 4436 } 4437 SLAB_ATTR_RO(objs_per_slab); 4438 4439 static ssize_t order_store(struct kmem_cache *s, 4440 const char *buf, size_t length) 4441 { 4442 unsigned long order; 4443 int err; 4444 4445 err = kstrtoul(buf, 10, &order); 4446 if (err) 4447 return err; 4448 4449 if (order > slub_max_order || order < slub_min_order) 4450 return -EINVAL; 4451 4452 calculate_sizes(s, order); 4453 return length; 4454 } 4455 4456 static ssize_t order_show(struct kmem_cache *s, char *buf) 4457 { 4458 return sprintf(buf, "%d\n", oo_order(s->oo)); 4459 } 4460 SLAB_ATTR(order); 4461 4462 static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 4463 { 4464 return sprintf(buf, "%lu\n", s->min_partial); 4465 } 4466 4467 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 4468 size_t length) 4469 { 4470 unsigned long min; 4471 int err; 4472 4473 err = kstrtoul(buf, 10, &min); 4474 if (err) 4475 return err; 4476 4477 set_min_partial(s, min); 4478 return length; 4479 } 4480 SLAB_ATTR(min_partial); 4481 4482 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) 4483 { 4484 return sprintf(buf, "%u\n", s->cpu_partial); 4485 } 4486 4487 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, 4488 size_t length) 4489 { 4490 unsigned long objects; 4491 int err; 4492 4493 err = kstrtoul(buf, 10, &objects); 4494 if (err) 4495 return err; 4496 if (objects && !kmem_cache_has_cpu_partial(s)) 4497 return -EINVAL; 4498 4499 s->cpu_partial = objects; 4500 flush_all(s); 4501 return length; 4502 } 4503 SLAB_ATTR(cpu_partial); 4504 4505 static ssize_t ctor_show(struct kmem_cache *s, char *buf) 4506 { 4507 if (!s->ctor) 4508 return 0; 4509 return sprintf(buf, "%pS\n", s->ctor); 4510 } 4511 SLAB_ATTR_RO(ctor); 4512 4513 static ssize_t aliases_show(struct kmem_cache *s, char *buf) 4514 { 4515 return sprintf(buf, "%d\n", s->refcount - 1); 4516 } 4517 SLAB_ATTR_RO(aliases); 4518 4519 static ssize_t partial_show(struct kmem_cache *s, char *buf) 4520 { 4521 return show_slab_objects(s, buf, SO_PARTIAL); 4522 } 4523 SLAB_ATTR_RO(partial); 4524 4525 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4526 { 4527 return show_slab_objects(s, buf, SO_CPU); 4528 } 4529 SLAB_ATTR_RO(cpu_slabs); 4530 4531 static ssize_t objects_show(struct kmem_cache *s, char *buf) 4532 { 4533 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4534 } 4535 SLAB_ATTR_RO(objects); 4536 4537 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4538 { 4539 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4540 } 4541 SLAB_ATTR_RO(objects_partial); 4542 4543 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) 4544 { 4545 int objects = 0; 4546 int pages = 0; 4547 int cpu; 4548 int len; 4549 4550 for_each_online_cpu(cpu) { 4551 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial; 4552 4553 if (page) { 4554 pages += page->pages; 4555 objects += page->pobjects; 4556 } 4557 } 4558 4559 len = sprintf(buf, "%d(%d)", objects, pages); 4560 4561 #ifdef CONFIG_SMP 4562 for_each_online_cpu(cpu) { 4563 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial; 4564 4565 if (page && len < PAGE_SIZE - 20) 4566 len += sprintf(buf + len, " C%d=%d(%d)", cpu, 4567 page->pobjects, page->pages); 4568 } 4569 #endif 4570 return len + sprintf(buf + len, "\n"); 4571 } 4572 SLAB_ATTR_RO(slabs_cpu_partial); 4573 4574 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4575 { 4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4577 } 4578 4579 static ssize_t reclaim_account_store(struct kmem_cache *s, 4580 const char *buf, size_t length) 4581 { 4582 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4583 if (buf[0] == '1') 4584 s->flags |= SLAB_RECLAIM_ACCOUNT; 4585 return length; 4586 } 4587 SLAB_ATTR(reclaim_account); 4588 4589 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4590 { 4591 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4592 } 4593 SLAB_ATTR_RO(hwcache_align); 4594 4595 #ifdef CONFIG_ZONE_DMA 4596 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4597 { 4598 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4599 } 4600 SLAB_ATTR_RO(cache_dma); 4601 #endif 4602 4603 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4604 { 4605 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4606 } 4607 SLAB_ATTR_RO(destroy_by_rcu); 4608 4609 static ssize_t reserved_show(struct kmem_cache *s, char *buf) 4610 { 4611 return sprintf(buf, "%d\n", s->reserved); 4612 } 4613 SLAB_ATTR_RO(reserved); 4614 4615 #ifdef CONFIG_SLUB_DEBUG 4616 static ssize_t slabs_show(struct kmem_cache *s, char *buf) 4617 { 4618 return show_slab_objects(s, buf, SO_ALL); 4619 } 4620 SLAB_ATTR_RO(slabs); 4621 4622 static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 4623 { 4624 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 4625 } 4626 SLAB_ATTR_RO(total_objects); 4627 4628 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 4629 { 4630 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE)); 4631 } 4632 4633 static ssize_t sanity_checks_store(struct kmem_cache *s, 4634 const char *buf, size_t length) 4635 { 4636 s->flags &= ~SLAB_DEBUG_FREE; 4637 if (buf[0] == '1') { 4638 s->flags &= ~__CMPXCHG_DOUBLE; 4639 s->flags |= SLAB_DEBUG_FREE; 4640 } 4641 return length; 4642 } 4643 SLAB_ATTR(sanity_checks); 4644 4645 static ssize_t trace_show(struct kmem_cache *s, char *buf) 4646 { 4647 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 4648 } 4649 4650 static ssize_t trace_store(struct kmem_cache *s, const char *buf, 4651 size_t length) 4652 { 4653 s->flags &= ~SLAB_TRACE; 4654 if (buf[0] == '1') { 4655 s->flags &= ~__CMPXCHG_DOUBLE; 4656 s->flags |= SLAB_TRACE; 4657 } 4658 return length; 4659 } 4660 SLAB_ATTR(trace); 4661 4662 static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 4663 { 4664 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 4665 } 4666 4667 static ssize_t red_zone_store(struct kmem_cache *s, 4668 const char *buf, size_t length) 4669 { 4670 if (any_slab_objects(s)) 4671 return -EBUSY; 4672 4673 s->flags &= ~SLAB_RED_ZONE; 4674 if (buf[0] == '1') { 4675 s->flags &= ~__CMPXCHG_DOUBLE; 4676 s->flags |= SLAB_RED_ZONE; 4677 } 4678 calculate_sizes(s, -1); 4679 return length; 4680 } 4681 SLAB_ATTR(red_zone); 4682 4683 static ssize_t poison_show(struct kmem_cache *s, char *buf) 4684 { 4685 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 4686 } 4687 4688 static ssize_t poison_store(struct kmem_cache *s, 4689 const char *buf, size_t length) 4690 { 4691 if (any_slab_objects(s)) 4692 return -EBUSY; 4693 4694 s->flags &= ~SLAB_POISON; 4695 if (buf[0] == '1') { 4696 s->flags &= ~__CMPXCHG_DOUBLE; 4697 s->flags |= SLAB_POISON; 4698 } 4699 calculate_sizes(s, -1); 4700 return length; 4701 } 4702 SLAB_ATTR(poison); 4703 4704 static ssize_t store_user_show(struct kmem_cache *s, char *buf) 4705 { 4706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 4707 } 4708 4709 static ssize_t store_user_store(struct kmem_cache *s, 4710 const char *buf, size_t length) 4711 { 4712 if (any_slab_objects(s)) 4713 return -EBUSY; 4714 4715 s->flags &= ~SLAB_STORE_USER; 4716 if (buf[0] == '1') { 4717 s->flags &= ~__CMPXCHG_DOUBLE; 4718 s->flags |= SLAB_STORE_USER; 4719 } 4720 calculate_sizes(s, -1); 4721 return length; 4722 } 4723 SLAB_ATTR(store_user); 4724 4725 static ssize_t validate_show(struct kmem_cache *s, char *buf) 4726 { 4727 return 0; 4728 } 4729 4730 static ssize_t validate_store(struct kmem_cache *s, 4731 const char *buf, size_t length) 4732 { 4733 int ret = -EINVAL; 4734 4735 if (buf[0] == '1') { 4736 ret = validate_slab_cache(s); 4737 if (ret >= 0) 4738 ret = length; 4739 } 4740 return ret; 4741 } 4742 SLAB_ATTR(validate); 4743 4744 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 4745 { 4746 if (!(s->flags & SLAB_STORE_USER)) 4747 return -ENOSYS; 4748 return list_locations(s, buf, TRACK_ALLOC); 4749 } 4750 SLAB_ATTR_RO(alloc_calls); 4751 4752 static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 4753 { 4754 if (!(s->flags & SLAB_STORE_USER)) 4755 return -ENOSYS; 4756 return list_locations(s, buf, TRACK_FREE); 4757 } 4758 SLAB_ATTR_RO(free_calls); 4759 #endif /* CONFIG_SLUB_DEBUG */ 4760 4761 #ifdef CONFIG_FAILSLAB 4762 static ssize_t failslab_show(struct kmem_cache *s, char *buf) 4763 { 4764 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 4765 } 4766 4767 static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 4768 size_t length) 4769 { 4770 s->flags &= ~SLAB_FAILSLAB; 4771 if (buf[0] == '1') 4772 s->flags |= SLAB_FAILSLAB; 4773 return length; 4774 } 4775 SLAB_ATTR(failslab); 4776 #endif 4777 4778 static ssize_t shrink_show(struct kmem_cache *s, char *buf) 4779 { 4780 return 0; 4781 } 4782 4783 static ssize_t shrink_store(struct kmem_cache *s, 4784 const char *buf, size_t length) 4785 { 4786 if (buf[0] == '1') { 4787 int rc = kmem_cache_shrink(s); 4788 4789 if (rc) 4790 return rc; 4791 } else 4792 return -EINVAL; 4793 return length; 4794 } 4795 SLAB_ATTR(shrink); 4796 4797 #ifdef CONFIG_NUMA 4798 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 4799 { 4800 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 4801 } 4802 4803 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 4804 const char *buf, size_t length) 4805 { 4806 unsigned long ratio; 4807 int err; 4808 4809 err = kstrtoul(buf, 10, &ratio); 4810 if (err) 4811 return err; 4812 4813 if (ratio <= 100) 4814 s->remote_node_defrag_ratio = ratio * 10; 4815 4816 return length; 4817 } 4818 SLAB_ATTR(remote_node_defrag_ratio); 4819 #endif 4820 4821 #ifdef CONFIG_SLUB_STATS 4822 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 4823 { 4824 unsigned long sum = 0; 4825 int cpu; 4826 int len; 4827 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 4828 4829 if (!data) 4830 return -ENOMEM; 4831 4832 for_each_online_cpu(cpu) { 4833 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 4834 4835 data[cpu] = x; 4836 sum += x; 4837 } 4838 4839 len = sprintf(buf, "%lu", sum); 4840 4841 #ifdef CONFIG_SMP 4842 for_each_online_cpu(cpu) { 4843 if (data[cpu] && len < PAGE_SIZE - 20) 4844 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 4845 } 4846 #endif 4847 kfree(data); 4848 return len + sprintf(buf + len, "\n"); 4849 } 4850 4851 static void clear_stat(struct kmem_cache *s, enum stat_item si) 4852 { 4853 int cpu; 4854 4855 for_each_online_cpu(cpu) 4856 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 4857 } 4858 4859 #define STAT_ATTR(si, text) \ 4860 static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 4861 { \ 4862 return show_stat(s, buf, si); \ 4863 } \ 4864 static ssize_t text##_store(struct kmem_cache *s, \ 4865 const char *buf, size_t length) \ 4866 { \ 4867 if (buf[0] != '0') \ 4868 return -EINVAL; \ 4869 clear_stat(s, si); \ 4870 return length; \ 4871 } \ 4872 SLAB_ATTR(text); \ 4873 4874 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 4875 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 4876 STAT_ATTR(FREE_FASTPATH, free_fastpath); 4877 STAT_ATTR(FREE_SLOWPATH, free_slowpath); 4878 STAT_ATTR(FREE_FROZEN, free_frozen); 4879 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 4880 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 4881 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 4882 STAT_ATTR(ALLOC_SLAB, alloc_slab); 4883 STAT_ATTR(ALLOC_REFILL, alloc_refill); 4884 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); 4885 STAT_ATTR(FREE_SLAB, free_slab); 4886 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 4887 STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 4888 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 4889 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 4890 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 4891 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 4892 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); 4893 STAT_ATTR(ORDER_FALLBACK, order_fallback); 4894 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); 4895 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); 4896 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); 4897 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); 4898 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); 4899 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); 4900 #endif 4901 4902 static struct attribute *slab_attrs[] = { 4903 &slab_size_attr.attr, 4904 &object_size_attr.attr, 4905 &objs_per_slab_attr.attr, 4906 &order_attr.attr, 4907 &min_partial_attr.attr, 4908 &cpu_partial_attr.attr, 4909 &objects_attr.attr, 4910 &objects_partial_attr.attr, 4911 &partial_attr.attr, 4912 &cpu_slabs_attr.attr, 4913 &ctor_attr.attr, 4914 &aliases_attr.attr, 4915 &align_attr.attr, 4916 &hwcache_align_attr.attr, 4917 &reclaim_account_attr.attr, 4918 &destroy_by_rcu_attr.attr, 4919 &shrink_attr.attr, 4920 &reserved_attr.attr, 4921 &slabs_cpu_partial_attr.attr, 4922 #ifdef CONFIG_SLUB_DEBUG 4923 &total_objects_attr.attr, 4924 &slabs_attr.attr, 4925 &sanity_checks_attr.attr, 4926 &trace_attr.attr, 4927 &red_zone_attr.attr, 4928 &poison_attr.attr, 4929 &store_user_attr.attr, 4930 &validate_attr.attr, 4931 &alloc_calls_attr.attr, 4932 &free_calls_attr.attr, 4933 #endif 4934 #ifdef CONFIG_ZONE_DMA 4935 &cache_dma_attr.attr, 4936 #endif 4937 #ifdef CONFIG_NUMA 4938 &remote_node_defrag_ratio_attr.attr, 4939 #endif 4940 #ifdef CONFIG_SLUB_STATS 4941 &alloc_fastpath_attr.attr, 4942 &alloc_slowpath_attr.attr, 4943 &free_fastpath_attr.attr, 4944 &free_slowpath_attr.attr, 4945 &free_frozen_attr.attr, 4946 &free_add_partial_attr.attr, 4947 &free_remove_partial_attr.attr, 4948 &alloc_from_partial_attr.attr, 4949 &alloc_slab_attr.attr, 4950 &alloc_refill_attr.attr, 4951 &alloc_node_mismatch_attr.attr, 4952 &free_slab_attr.attr, 4953 &cpuslab_flush_attr.attr, 4954 &deactivate_full_attr.attr, 4955 &deactivate_empty_attr.attr, 4956 &deactivate_to_head_attr.attr, 4957 &deactivate_to_tail_attr.attr, 4958 &deactivate_remote_frees_attr.attr, 4959 &deactivate_bypass_attr.attr, 4960 &order_fallback_attr.attr, 4961 &cmpxchg_double_fail_attr.attr, 4962 &cmpxchg_double_cpu_fail_attr.attr, 4963 &cpu_partial_alloc_attr.attr, 4964 &cpu_partial_free_attr.attr, 4965 &cpu_partial_node_attr.attr, 4966 &cpu_partial_drain_attr.attr, 4967 #endif 4968 #ifdef CONFIG_FAILSLAB 4969 &failslab_attr.attr, 4970 #endif 4971 4972 NULL 4973 }; 4974 4975 static struct attribute_group slab_attr_group = { 4976 .attrs = slab_attrs, 4977 }; 4978 4979 static ssize_t slab_attr_show(struct kobject *kobj, 4980 struct attribute *attr, 4981 char *buf) 4982 { 4983 struct slab_attribute *attribute; 4984 struct kmem_cache *s; 4985 int err; 4986 4987 attribute = to_slab_attr(attr); 4988 s = to_slab(kobj); 4989 4990 if (!attribute->show) 4991 return -EIO; 4992 4993 err = attribute->show(s, buf); 4994 4995 return err; 4996 } 4997 4998 static ssize_t slab_attr_store(struct kobject *kobj, 4999 struct attribute *attr, 5000 const char *buf, size_t len) 5001 { 5002 struct slab_attribute *attribute; 5003 struct kmem_cache *s; 5004 int err; 5005 5006 attribute = to_slab_attr(attr); 5007 s = to_slab(kobj); 5008 5009 if (!attribute->store) 5010 return -EIO; 5011 5012 err = attribute->store(s, buf, len); 5013 #ifdef CONFIG_MEMCG_KMEM 5014 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) { 5015 int i; 5016 5017 mutex_lock(&slab_mutex); 5018 if (s->max_attr_size < len) 5019 s->max_attr_size = len; 5020 5021 /* 5022 * This is a best effort propagation, so this function's return 5023 * value will be determined by the parent cache only. This is 5024 * basically because not all attributes will have a well 5025 * defined semantics for rollbacks - most of the actions will 5026 * have permanent effects. 5027 * 5028 * Returning the error value of any of the children that fail 5029 * is not 100 % defined, in the sense that users seeing the 5030 * error code won't be able to know anything about the state of 5031 * the cache. 5032 * 5033 * Only returning the error code for the parent cache at least 5034 * has well defined semantics. The cache being written to 5035 * directly either failed or succeeded, in which case we loop 5036 * through the descendants with best-effort propagation. 5037 */ 5038 for_each_memcg_cache_index(i) { 5039 struct kmem_cache *c = cache_from_memcg_idx(s, i); 5040 if (c) 5041 attribute->store(c, buf, len); 5042 } 5043 mutex_unlock(&slab_mutex); 5044 } 5045 #endif 5046 return err; 5047 } 5048 5049 static void memcg_propagate_slab_attrs(struct kmem_cache *s) 5050 { 5051 #ifdef CONFIG_MEMCG_KMEM 5052 int i; 5053 char *buffer = NULL; 5054 struct kmem_cache *root_cache; 5055 5056 if (is_root_cache(s)) 5057 return; 5058 5059 root_cache = s->memcg_params->root_cache; 5060 5061 /* 5062 * This mean this cache had no attribute written. Therefore, no point 5063 * in copying default values around 5064 */ 5065 if (!root_cache->max_attr_size) 5066 return; 5067 5068 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) { 5069 char mbuf[64]; 5070 char *buf; 5071 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]); 5072 5073 if (!attr || !attr->store || !attr->show) 5074 continue; 5075 5076 /* 5077 * It is really bad that we have to allocate here, so we will 5078 * do it only as a fallback. If we actually allocate, though, 5079 * we can just use the allocated buffer until the end. 5080 * 5081 * Most of the slub attributes will tend to be very small in 5082 * size, but sysfs allows buffers up to a page, so they can 5083 * theoretically happen. 5084 */ 5085 if (buffer) 5086 buf = buffer; 5087 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf)) 5088 buf = mbuf; 5089 else { 5090 buffer = (char *) get_zeroed_page(GFP_KERNEL); 5091 if (WARN_ON(!buffer)) 5092 continue; 5093 buf = buffer; 5094 } 5095 5096 attr->show(root_cache, buf); 5097 attr->store(s, buf, strlen(buf)); 5098 } 5099 5100 if (buffer) 5101 free_page((unsigned long)buffer); 5102 #endif 5103 } 5104 5105 static void kmem_cache_release(struct kobject *k) 5106 { 5107 slab_kmem_cache_release(to_slab(k)); 5108 } 5109 5110 static const struct sysfs_ops slab_sysfs_ops = { 5111 .show = slab_attr_show, 5112 .store = slab_attr_store, 5113 }; 5114 5115 static struct kobj_type slab_ktype = { 5116 .sysfs_ops = &slab_sysfs_ops, 5117 .release = kmem_cache_release, 5118 }; 5119 5120 static int uevent_filter(struct kset *kset, struct kobject *kobj) 5121 { 5122 struct kobj_type *ktype = get_ktype(kobj); 5123 5124 if (ktype == &slab_ktype) 5125 return 1; 5126 return 0; 5127 } 5128 5129 static const struct kset_uevent_ops slab_uevent_ops = { 5130 .filter = uevent_filter, 5131 }; 5132 5133 static struct kset *slab_kset; 5134 5135 static inline struct kset *cache_kset(struct kmem_cache *s) 5136 { 5137 #ifdef CONFIG_MEMCG_KMEM 5138 if (!is_root_cache(s)) 5139 return s->memcg_params->root_cache->memcg_kset; 5140 #endif 5141 return slab_kset; 5142 } 5143 5144 #define ID_STR_LENGTH 64 5145 5146 /* Create a unique string id for a slab cache: 5147 * 5148 * Format :[flags-]size 5149 */ 5150 static char *create_unique_id(struct kmem_cache *s) 5151 { 5152 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 5153 char *p = name; 5154 5155 BUG_ON(!name); 5156 5157 *p++ = ':'; 5158 /* 5159 * First flags affecting slabcache operations. We will only 5160 * get here for aliasable slabs so we do not need to support 5161 * too many flags. The flags here must cover all flags that 5162 * are matched during merging to guarantee that the id is 5163 * unique. 5164 */ 5165 if (s->flags & SLAB_CACHE_DMA) 5166 *p++ = 'd'; 5167 if (s->flags & SLAB_RECLAIM_ACCOUNT) 5168 *p++ = 'a'; 5169 if (s->flags & SLAB_DEBUG_FREE) 5170 *p++ = 'F'; 5171 if (!(s->flags & SLAB_NOTRACK)) 5172 *p++ = 't'; 5173 if (p != name + 1) 5174 *p++ = '-'; 5175 p += sprintf(p, "%07d", s->size); 5176 5177 #ifdef CONFIG_MEMCG_KMEM 5178 if (!is_root_cache(s)) 5179 p += sprintf(p, "-%08d", 5180 memcg_cache_id(s->memcg_params->memcg)); 5181 #endif 5182 5183 BUG_ON(p > name + ID_STR_LENGTH - 1); 5184 return name; 5185 } 5186 5187 static int sysfs_slab_add(struct kmem_cache *s) 5188 { 5189 int err; 5190 const char *name; 5191 int unmergeable = slab_unmergeable(s); 5192 5193 if (unmergeable) { 5194 /* 5195 * Slabcache can never be merged so we can use the name proper. 5196 * This is typically the case for debug situations. In that 5197 * case we can catch duplicate names easily. 5198 */ 5199 sysfs_remove_link(&slab_kset->kobj, s->name); 5200 name = s->name; 5201 } else { 5202 /* 5203 * Create a unique name for the slab as a target 5204 * for the symlinks. 5205 */ 5206 name = create_unique_id(s); 5207 } 5208 5209 s->kobj.kset = cache_kset(s); 5210 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); 5211 if (err) 5212 goto out_put_kobj; 5213 5214 err = sysfs_create_group(&s->kobj, &slab_attr_group); 5215 if (err) 5216 goto out_del_kobj; 5217 5218 #ifdef CONFIG_MEMCG_KMEM 5219 if (is_root_cache(s)) { 5220 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj); 5221 if (!s->memcg_kset) { 5222 err = -ENOMEM; 5223 goto out_del_kobj; 5224 } 5225 } 5226 #endif 5227 5228 kobject_uevent(&s->kobj, KOBJ_ADD); 5229 if (!unmergeable) { 5230 /* Setup first alias */ 5231 sysfs_slab_alias(s, s->name); 5232 } 5233 out: 5234 if (!unmergeable) 5235 kfree(name); 5236 return err; 5237 out_del_kobj: 5238 kobject_del(&s->kobj); 5239 out_put_kobj: 5240 kobject_put(&s->kobj); 5241 goto out; 5242 } 5243 5244 void sysfs_slab_remove(struct kmem_cache *s) 5245 { 5246 if (slab_state < FULL) 5247 /* 5248 * Sysfs has not been setup yet so no need to remove the 5249 * cache from sysfs. 5250 */ 5251 return; 5252 5253 #ifdef CONFIG_MEMCG_KMEM 5254 kset_unregister(s->memcg_kset); 5255 #endif 5256 kobject_uevent(&s->kobj, KOBJ_REMOVE); 5257 kobject_del(&s->kobj); 5258 kobject_put(&s->kobj); 5259 } 5260 5261 /* 5262 * Need to buffer aliases during bootup until sysfs becomes 5263 * available lest we lose that information. 5264 */ 5265 struct saved_alias { 5266 struct kmem_cache *s; 5267 const char *name; 5268 struct saved_alias *next; 5269 }; 5270 5271 static struct saved_alias *alias_list; 5272 5273 static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 5274 { 5275 struct saved_alias *al; 5276 5277 if (slab_state == FULL) { 5278 /* 5279 * If we have a leftover link then remove it. 5280 */ 5281 sysfs_remove_link(&slab_kset->kobj, name); 5282 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 5283 } 5284 5285 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 5286 if (!al) 5287 return -ENOMEM; 5288 5289 al->s = s; 5290 al->name = name; 5291 al->next = alias_list; 5292 alias_list = al; 5293 return 0; 5294 } 5295 5296 static int __init slab_sysfs_init(void) 5297 { 5298 struct kmem_cache *s; 5299 int err; 5300 5301 mutex_lock(&slab_mutex); 5302 5303 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 5304 if (!slab_kset) { 5305 mutex_unlock(&slab_mutex); 5306 printk(KERN_ERR "Cannot register slab subsystem.\n"); 5307 return -ENOSYS; 5308 } 5309 5310 slab_state = FULL; 5311 5312 list_for_each_entry(s, &slab_caches, list) { 5313 err = sysfs_slab_add(s); 5314 if (err) 5315 printk(KERN_ERR "SLUB: Unable to add boot slab %s" 5316 " to sysfs\n", s->name); 5317 } 5318 5319 while (alias_list) { 5320 struct saved_alias *al = alias_list; 5321 5322 alias_list = alias_list->next; 5323 err = sysfs_slab_alias(al->s, al->name); 5324 if (err) 5325 printk(KERN_ERR "SLUB: Unable to add boot slab alias" 5326 " %s to sysfs\n", al->name); 5327 kfree(al); 5328 } 5329 5330 mutex_unlock(&slab_mutex); 5331 resiliency_test(); 5332 return 0; 5333 } 5334 5335 __initcall(slab_sysfs_init); 5336 #endif /* CONFIG_SYSFS */ 5337 5338 /* 5339 * The /proc/slabinfo ABI 5340 */ 5341 #ifdef CONFIG_SLABINFO 5342 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) 5343 { 5344 unsigned long nr_slabs = 0; 5345 unsigned long nr_objs = 0; 5346 unsigned long nr_free = 0; 5347 int node; 5348 5349 for_each_online_node(node) { 5350 struct kmem_cache_node *n = get_node(s, node); 5351 5352 if (!n) 5353 continue; 5354 5355 nr_slabs += node_nr_slabs(n); 5356 nr_objs += node_nr_objs(n); 5357 nr_free += count_partial(n, count_free); 5358 } 5359 5360 sinfo->active_objs = nr_objs - nr_free; 5361 sinfo->num_objs = nr_objs; 5362 sinfo->active_slabs = nr_slabs; 5363 sinfo->num_slabs = nr_slabs; 5364 sinfo->objects_per_slab = oo_objects(s->oo); 5365 sinfo->cache_order = oo_order(s->oo); 5366 } 5367 5368 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) 5369 { 5370 } 5371 5372 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 5373 size_t count, loff_t *ppos) 5374 { 5375 return -EIO; 5376 } 5377 #endif /* CONFIG_SLABINFO */ 5378