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/kasan.h> 24 #include <linux/kmemcheck.h> 25 #include <linux/cpu.h> 26 #include <linux/cpuset.h> 27 #include <linux/mempolicy.h> 28 #include <linux/ctype.h> 29 #include <linux/debugobjects.h> 30 #include <linux/kallsyms.h> 31 #include <linux/memory.h> 32 #include <linux/math64.h> 33 #include <linux/fault-inject.h> 34 #include <linux/stacktrace.h> 35 #include <linux/prefetch.h> 36 #include <linux/memcontrol.h> 37 38 #include <trace/events/kmem.h> 39 40 #include "internal.h" 41 42 /* 43 * Lock order: 44 * 1. slab_mutex (Global Mutex) 45 * 2. node->list_lock 46 * 3. slab_lock(page) (Only on some arches and for debugging) 47 * 48 * slab_mutex 49 * 50 * The role of the slab_mutex is to protect the list of all the slabs 51 * and to synchronize major metadata changes to slab cache structures. 52 * 53 * The slab_lock is only used for debugging and on arches that do not 54 * have the ability to do a cmpxchg_double. It only protects the second 55 * double word in the page struct. Meaning 56 * A. page->freelist -> List of object free in a page 57 * B. page->counters -> Counters of objects 58 * C. page->frozen -> frozen state 59 * 60 * If a slab is frozen then it is exempt from list management. It is not 61 * on any list. The processor that froze the slab is the one who can 62 * perform list operations on the page. Other processors may put objects 63 * onto the freelist but the processor that froze the slab is the only 64 * one that can retrieve the objects from the page's freelist. 65 * 66 * The list_lock protects the partial and full list on each node and 67 * the partial slab counter. If taken then no new slabs may be added or 68 * removed from the lists nor make the number of partial slabs be modified. 69 * (Note that the total number of slabs is an atomic value that may be 70 * modified without taking the list lock). 71 * 72 * The list_lock is a centralized lock and thus we avoid taking it as 73 * much as possible. As long as SLUB does not have to handle partial 74 * slabs, operations can continue without any centralized lock. F.e. 75 * allocating a long series of objects that fill up slabs does not require 76 * the list lock. 77 * Interrupts are disabled during allocation and deallocation in order to 78 * make the slab allocator safe to use in the context of an irq. In addition 79 * interrupts are disabled to ensure that the processor does not change 80 * while handling per_cpu slabs, due to kernel preemption. 81 * 82 * SLUB assigns one slab for allocation to each processor. 83 * Allocations only occur from these slabs called cpu slabs. 84 * 85 * Slabs with free elements are kept on a partial list and during regular 86 * operations no list for full slabs is used. If an object in a full slab is 87 * freed then the slab will show up again on the partial lists. 88 * We track full slabs for debugging purposes though because otherwise we 89 * cannot scan all objects. 90 * 91 * Slabs are freed when they become empty. Teardown and setup is 92 * minimal so we rely on the page allocators per cpu caches for 93 * fast frees and allocs. 94 * 95 * Overloading of page flags that are otherwise used for LRU management. 96 * 97 * PageActive The slab is frozen and exempt from list processing. 98 * This means that the slab is dedicated to a purpose 99 * such as satisfying allocations for a specific 100 * processor. Objects may be freed in the slab while 101 * it is frozen but slab_free will then skip the usual 102 * list operations. It is up to the processor holding 103 * the slab to integrate the slab into the slab lists 104 * when the slab is no longer needed. 105 * 106 * One use of this flag is to mark slabs that are 107 * used for allocations. Then such a slab becomes a cpu 108 * slab. The cpu slab may be equipped with an additional 109 * freelist that allows lockless access to 110 * free objects in addition to the regular freelist 111 * that requires the slab lock. 112 * 113 * PageError Slab requires special handling due to debug 114 * options set. This moves slab handling out of 115 * the fast path and disables lockless freelists. 116 */ 117 118 static inline int kmem_cache_debug(struct kmem_cache *s) 119 { 120 #ifdef CONFIG_SLUB_DEBUG 121 return unlikely(s->flags & SLAB_DEBUG_FLAGS); 122 #else 123 return 0; 124 #endif 125 } 126 127 void *fixup_red_left(struct kmem_cache *s, void *p) 128 { 129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) 130 p += s->red_left_pad; 131 132 return p; 133 } 134 135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) 136 { 137 #ifdef CONFIG_SLUB_CPU_PARTIAL 138 return !kmem_cache_debug(s); 139 #else 140 return false; 141 #endif 142 } 143 144 /* 145 * Issues still to be resolved: 146 * 147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do. 148 * 149 * - Variable sizing of the per node arrays 150 */ 151 152 /* Enable to test recovery from slab corruption on boot */ 153 #undef SLUB_RESILIENCY_TEST 154 155 /* Enable to log cmpxchg failures */ 156 #undef SLUB_DEBUG_CMPXCHG 157 158 /* 159 * Mininum number of partial slabs. These will be left on the partial 160 * lists even if they are empty. kmem_cache_shrink may reclaim them. 161 */ 162 #define MIN_PARTIAL 5 163 164 /* 165 * Maximum number of desirable partial slabs. 166 * The existence of more partial slabs makes kmem_cache_shrink 167 * sort the partial list by the number of objects in use. 168 */ 169 #define MAX_PARTIAL 10 170 171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ 172 SLAB_POISON | SLAB_STORE_USER) 173 174 /* 175 * These debug flags cannot use CMPXCHG because there might be consistency 176 * issues when checking or reading debug information 177 */ 178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ 179 SLAB_TRACE) 180 181 182 /* 183 * Debugging flags that require metadata to be stored in the slab. These get 184 * disabled when slub_debug=O is used and a cache's min order increases with 185 * metadata. 186 */ 187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) 188 189 #define OO_SHIFT 16 190 #define OO_MASK ((1 << OO_SHIFT) - 1) 191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */ 192 193 /* Internal SLUB flags */ 194 #define __OBJECT_POISON 0x80000000UL /* Poison object */ 195 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */ 196 197 /* 198 * Tracking user of a slab. 199 */ 200 #define TRACK_ADDRS_COUNT 16 201 struct track { 202 unsigned long addr; /* Called from address */ 203 #ifdef CONFIG_STACKTRACE 204 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 205 #endif 206 int cpu; /* Was running on cpu */ 207 int pid; /* Pid context */ 208 unsigned long when; /* When did the operation occur */ 209 }; 210 211 enum track_item { TRACK_ALLOC, TRACK_FREE }; 212 213 #ifdef CONFIG_SYSFS 214 static int sysfs_slab_add(struct kmem_cache *); 215 static int sysfs_slab_alias(struct kmem_cache *, const char *); 216 static void memcg_propagate_slab_attrs(struct kmem_cache *s); 217 #else 218 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 219 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 220 { return 0; } 221 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { } 222 #endif 223 224 static inline void stat(const struct kmem_cache *s, enum stat_item si) 225 { 226 #ifdef CONFIG_SLUB_STATS 227 /* 228 * The rmw is racy on a preemptible kernel but this is acceptable, so 229 * avoid this_cpu_add()'s irq-disable overhead. 230 */ 231 raw_cpu_inc(s->cpu_slab->stat[si]); 232 #endif 233 } 234 235 /******************************************************************** 236 * Core slab cache functions 237 *******************************************************************/ 238 239 static inline void *get_freepointer(struct kmem_cache *s, void *object) 240 { 241 return *(void **)(object + s->offset); 242 } 243 244 static void prefetch_freepointer(const struct kmem_cache *s, void *object) 245 { 246 prefetch(object + s->offset); 247 } 248 249 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 250 { 251 void *p; 252 253 if (!debug_pagealloc_enabled()) 254 return get_freepointer(s, object); 255 256 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p)); 257 return p; 258 } 259 260 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 261 { 262 *(void **)(object + s->offset) = fp; 263 } 264 265 /* Loop over all objects in a slab */ 266 #define for_each_object(__p, __s, __addr, __objects) \ 267 for (__p = fixup_red_left(__s, __addr); \ 268 __p < (__addr) + (__objects) * (__s)->size; \ 269 __p += (__s)->size) 270 271 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \ 272 for (__p = fixup_red_left(__s, __addr), __idx = 1; \ 273 __idx <= __objects; \ 274 __p += (__s)->size, __idx++) 275 276 /* Determine object index from a given position */ 277 static inline int slab_index(void *p, struct kmem_cache *s, void *addr) 278 { 279 return (p - addr) / s->size; 280 } 281 282 static inline int order_objects(int order, unsigned long size, int reserved) 283 { 284 return ((PAGE_SIZE << order) - reserved) / size; 285 } 286 287 static inline struct kmem_cache_order_objects oo_make(int order, 288 unsigned long size, int reserved) 289 { 290 struct kmem_cache_order_objects x = { 291 (order << OO_SHIFT) + order_objects(order, size, reserved) 292 }; 293 294 return x; 295 } 296 297 static inline int oo_order(struct kmem_cache_order_objects x) 298 { 299 return x.x >> OO_SHIFT; 300 } 301 302 static inline int oo_objects(struct kmem_cache_order_objects x) 303 { 304 return x.x & OO_MASK; 305 } 306 307 /* 308 * Per slab locking using the pagelock 309 */ 310 static __always_inline void slab_lock(struct page *page) 311 { 312 VM_BUG_ON_PAGE(PageTail(page), page); 313 bit_spin_lock(PG_locked, &page->flags); 314 } 315 316 static __always_inline void slab_unlock(struct page *page) 317 { 318 VM_BUG_ON_PAGE(PageTail(page), page); 319 __bit_spin_unlock(PG_locked, &page->flags); 320 } 321 322 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new) 323 { 324 struct page tmp; 325 tmp.counters = counters_new; 326 /* 327 * page->counters can cover frozen/inuse/objects as well 328 * as page->_refcount. If we assign to ->counters directly 329 * we run the risk of losing updates to page->_refcount, so 330 * be careful and only assign to the fields we need. 331 */ 332 page->frozen = tmp.frozen; 333 page->inuse = tmp.inuse; 334 page->objects = tmp.objects; 335 } 336 337 /* Interrupts must be disabled (for the fallback code to work right) */ 338 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 339 void *freelist_old, unsigned long counters_old, 340 void *freelist_new, unsigned long counters_new, 341 const char *n) 342 { 343 VM_BUG_ON(!irqs_disabled()); 344 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 345 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 346 if (s->flags & __CMPXCHG_DOUBLE) { 347 if (cmpxchg_double(&page->freelist, &page->counters, 348 freelist_old, counters_old, 349 freelist_new, counters_new)) 350 return true; 351 } else 352 #endif 353 { 354 slab_lock(page); 355 if (page->freelist == freelist_old && 356 page->counters == counters_old) { 357 page->freelist = freelist_new; 358 set_page_slub_counters(page, counters_new); 359 slab_unlock(page); 360 return true; 361 } 362 slab_unlock(page); 363 } 364 365 cpu_relax(); 366 stat(s, CMPXCHG_DOUBLE_FAIL); 367 368 #ifdef SLUB_DEBUG_CMPXCHG 369 pr_info("%s %s: cmpxchg double redo ", n, s->name); 370 #endif 371 372 return false; 373 } 374 375 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 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 381 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 382 if (s->flags & __CMPXCHG_DOUBLE) { 383 if (cmpxchg_double(&page->freelist, &page->counters, 384 freelist_old, counters_old, 385 freelist_new, counters_new)) 386 return true; 387 } else 388 #endif 389 { 390 unsigned long flags; 391 392 local_irq_save(flags); 393 slab_lock(page); 394 if (page->freelist == freelist_old && 395 page->counters == counters_old) { 396 page->freelist = freelist_new; 397 set_page_slub_counters(page, counters_new); 398 slab_unlock(page); 399 local_irq_restore(flags); 400 return true; 401 } 402 slab_unlock(page); 403 local_irq_restore(flags); 404 } 405 406 cpu_relax(); 407 stat(s, CMPXCHG_DOUBLE_FAIL); 408 409 #ifdef SLUB_DEBUG_CMPXCHG 410 pr_info("%s %s: cmpxchg double redo ", n, s->name); 411 #endif 412 413 return false; 414 } 415 416 #ifdef CONFIG_SLUB_DEBUG 417 /* 418 * Determine a map of object in use on a page. 419 * 420 * Node listlock must be held to guarantee that the page does 421 * not vanish from under us. 422 */ 423 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) 424 { 425 void *p; 426 void *addr = page_address(page); 427 428 for (p = page->freelist; p; p = get_freepointer(s, p)) 429 set_bit(slab_index(p, s, addr), map); 430 } 431 432 static inline int size_from_object(struct kmem_cache *s) 433 { 434 if (s->flags & SLAB_RED_ZONE) 435 return s->size - s->red_left_pad; 436 437 return s->size; 438 } 439 440 static inline void *restore_red_left(struct kmem_cache *s, void *p) 441 { 442 if (s->flags & SLAB_RED_ZONE) 443 p -= s->red_left_pad; 444 445 return p; 446 } 447 448 /* 449 * Debug settings: 450 */ 451 #if defined(CONFIG_SLUB_DEBUG_ON) 452 static int slub_debug = DEBUG_DEFAULT_FLAGS; 453 #else 454 static int slub_debug; 455 #endif 456 457 static char *slub_debug_slabs; 458 static int disable_higher_order_debug; 459 460 /* 461 * slub is about to manipulate internal object metadata. This memory lies 462 * outside the range of the allocated object, so accessing it would normally 463 * be reported by kasan as a bounds error. metadata_access_enable() is used 464 * to tell kasan that these accesses are OK. 465 */ 466 static inline void metadata_access_enable(void) 467 { 468 kasan_disable_current(); 469 } 470 471 static inline void metadata_access_disable(void) 472 { 473 kasan_enable_current(); 474 } 475 476 /* 477 * Object debugging 478 */ 479 480 /* Verify that a pointer has an address that is valid within a slab page */ 481 static inline int check_valid_pointer(struct kmem_cache *s, 482 struct page *page, void *object) 483 { 484 void *base; 485 486 if (!object) 487 return 1; 488 489 base = page_address(page); 490 object = restore_red_left(s, object); 491 if (object < base || object >= base + page->objects * s->size || 492 (object - base) % s->size) { 493 return 0; 494 } 495 496 return 1; 497 } 498 499 static void print_section(char *level, char *text, u8 *addr, 500 unsigned int length) 501 { 502 metadata_access_enable(); 503 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr, 504 length, 1); 505 metadata_access_disable(); 506 } 507 508 static struct track *get_track(struct kmem_cache *s, void *object, 509 enum track_item alloc) 510 { 511 struct track *p; 512 513 if (s->offset) 514 p = object + s->offset + sizeof(void *); 515 else 516 p = object + s->inuse; 517 518 return p + alloc; 519 } 520 521 static void set_track(struct kmem_cache *s, void *object, 522 enum track_item alloc, unsigned long addr) 523 { 524 struct track *p = get_track(s, object, alloc); 525 526 if (addr) { 527 #ifdef CONFIG_STACKTRACE 528 struct stack_trace trace; 529 int i; 530 531 trace.nr_entries = 0; 532 trace.max_entries = TRACK_ADDRS_COUNT; 533 trace.entries = p->addrs; 534 trace.skip = 3; 535 metadata_access_enable(); 536 save_stack_trace(&trace); 537 metadata_access_disable(); 538 539 /* See rant in lockdep.c */ 540 if (trace.nr_entries != 0 && 541 trace.entries[trace.nr_entries - 1] == ULONG_MAX) 542 trace.nr_entries--; 543 544 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) 545 p->addrs[i] = 0; 546 #endif 547 p->addr = addr; 548 p->cpu = smp_processor_id(); 549 p->pid = current->pid; 550 p->when = jiffies; 551 } else 552 memset(p, 0, sizeof(struct track)); 553 } 554 555 static void init_tracking(struct kmem_cache *s, void *object) 556 { 557 if (!(s->flags & SLAB_STORE_USER)) 558 return; 559 560 set_track(s, object, TRACK_FREE, 0UL); 561 set_track(s, object, TRACK_ALLOC, 0UL); 562 } 563 564 static void print_track(const char *s, struct track *t) 565 { 566 if (!t->addr) 567 return; 568 569 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 570 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid); 571 #ifdef CONFIG_STACKTRACE 572 { 573 int i; 574 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 575 if (t->addrs[i]) 576 pr_err("\t%pS\n", (void *)t->addrs[i]); 577 else 578 break; 579 } 580 #endif 581 } 582 583 static void print_tracking(struct kmem_cache *s, void *object) 584 { 585 if (!(s->flags & SLAB_STORE_USER)) 586 return; 587 588 print_track("Allocated", get_track(s, object, TRACK_ALLOC)); 589 print_track("Freed", get_track(s, object, TRACK_FREE)); 590 } 591 592 static void print_page_info(struct page *page) 593 { 594 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 595 page, page->objects, page->inuse, page->freelist, page->flags); 596 597 } 598 599 static void slab_bug(struct kmem_cache *s, char *fmt, ...) 600 { 601 struct va_format vaf; 602 va_list args; 603 604 va_start(args, fmt); 605 vaf.fmt = fmt; 606 vaf.va = &args; 607 pr_err("=============================================================================\n"); 608 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf); 609 pr_err("-----------------------------------------------------------------------------\n\n"); 610 611 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 612 va_end(args); 613 } 614 615 static void slab_fix(struct kmem_cache *s, char *fmt, ...) 616 { 617 struct va_format vaf; 618 va_list args; 619 620 va_start(args, fmt); 621 vaf.fmt = fmt; 622 vaf.va = &args; 623 pr_err("FIX %s: %pV\n", s->name, &vaf); 624 va_end(args); 625 } 626 627 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 628 { 629 unsigned int off; /* Offset of last byte */ 630 u8 *addr = page_address(page); 631 632 print_tracking(s, p); 633 634 print_page_info(page); 635 636 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 637 p, p - addr, get_freepointer(s, p)); 638 639 if (s->flags & SLAB_RED_ZONE) 640 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad, 641 s->red_left_pad); 642 else if (p > addr + 16) 643 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16); 644 645 print_section(KERN_ERR, "Object ", p, 646 min_t(unsigned long, s->object_size, PAGE_SIZE)); 647 if (s->flags & SLAB_RED_ZONE) 648 print_section(KERN_ERR, "Redzone ", p + s->object_size, 649 s->inuse - s->object_size); 650 651 if (s->offset) 652 off = s->offset + sizeof(void *); 653 else 654 off = s->inuse; 655 656 if (s->flags & SLAB_STORE_USER) 657 off += 2 * sizeof(struct track); 658 659 off += kasan_metadata_size(s); 660 661 if (off != size_from_object(s)) 662 /* Beginning of the filler is the free pointer */ 663 print_section(KERN_ERR, "Padding ", p + off, 664 size_from_object(s) - off); 665 666 dump_stack(); 667 } 668 669 void object_err(struct kmem_cache *s, struct page *page, 670 u8 *object, char *reason) 671 { 672 slab_bug(s, "%s", reason); 673 print_trailer(s, page, object); 674 } 675 676 static void slab_err(struct kmem_cache *s, struct page *page, 677 const char *fmt, ...) 678 { 679 va_list args; 680 char buf[100]; 681 682 va_start(args, fmt); 683 vsnprintf(buf, sizeof(buf), fmt, args); 684 va_end(args); 685 slab_bug(s, "%s", buf); 686 print_page_info(page); 687 dump_stack(); 688 } 689 690 static void init_object(struct kmem_cache *s, void *object, u8 val) 691 { 692 u8 *p = object; 693 694 if (s->flags & SLAB_RED_ZONE) 695 memset(p - s->red_left_pad, val, s->red_left_pad); 696 697 if (s->flags & __OBJECT_POISON) { 698 memset(p, POISON_FREE, s->object_size - 1); 699 p[s->object_size - 1] = POISON_END; 700 } 701 702 if (s->flags & SLAB_RED_ZONE) 703 memset(p + s->object_size, val, s->inuse - s->object_size); 704 } 705 706 static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 707 void *from, void *to) 708 { 709 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 710 memset(from, data, to - from); 711 } 712 713 static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 714 u8 *object, char *what, 715 u8 *start, unsigned int value, unsigned int bytes) 716 { 717 u8 *fault; 718 u8 *end; 719 720 metadata_access_enable(); 721 fault = memchr_inv(start, value, bytes); 722 metadata_access_disable(); 723 if (!fault) 724 return 1; 725 726 end = start + bytes; 727 while (end > fault && end[-1] == value) 728 end--; 729 730 slab_bug(s, "%s overwritten", what); 731 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 732 fault, end - 1, fault[0], value); 733 print_trailer(s, page, object); 734 735 restore_bytes(s, what, value, fault, end); 736 return 0; 737 } 738 739 /* 740 * Object layout: 741 * 742 * object address 743 * Bytes of the object to be managed. 744 * If the freepointer may overlay the object then the free 745 * pointer is the first word of the object. 746 * 747 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 748 * 0xa5 (POISON_END) 749 * 750 * object + s->object_size 751 * Padding to reach word boundary. This is also used for Redzoning. 752 * Padding is extended by another word if Redzoning is enabled and 753 * object_size == inuse. 754 * 755 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 756 * 0xcc (RED_ACTIVE) for objects in use. 757 * 758 * object + s->inuse 759 * Meta data starts here. 760 * 761 * A. Free pointer (if we cannot overwrite object on free) 762 * B. Tracking data for SLAB_STORE_USER 763 * C. Padding to reach required alignment boundary or at mininum 764 * one word if debugging is on to be able to detect writes 765 * before the word boundary. 766 * 767 * Padding is done using 0x5a (POISON_INUSE) 768 * 769 * object + s->size 770 * Nothing is used beyond s->size. 771 * 772 * If slabcaches are merged then the object_size and inuse boundaries are mostly 773 * ignored. And therefore no slab options that rely on these boundaries 774 * may be used with merged slabcaches. 775 */ 776 777 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 778 { 779 unsigned long off = s->inuse; /* The end of info */ 780 781 if (s->offset) 782 /* Freepointer is placed after the object. */ 783 off += sizeof(void *); 784 785 if (s->flags & SLAB_STORE_USER) 786 /* We also have user information there */ 787 off += 2 * sizeof(struct track); 788 789 off += kasan_metadata_size(s); 790 791 if (size_from_object(s) == off) 792 return 1; 793 794 return check_bytes_and_report(s, page, p, "Object padding", 795 p + off, POISON_INUSE, size_from_object(s) - off); 796 } 797 798 /* Check the pad bytes at the end of a slab page */ 799 static int slab_pad_check(struct kmem_cache *s, struct page *page) 800 { 801 u8 *start; 802 u8 *fault; 803 u8 *end; 804 int length; 805 int remainder; 806 807 if (!(s->flags & SLAB_POISON)) 808 return 1; 809 810 start = page_address(page); 811 length = (PAGE_SIZE << compound_order(page)) - s->reserved; 812 end = start + length; 813 remainder = length % s->size; 814 if (!remainder) 815 return 1; 816 817 metadata_access_enable(); 818 fault = memchr_inv(end - remainder, POISON_INUSE, remainder); 819 metadata_access_disable(); 820 if (!fault) 821 return 1; 822 while (end > fault && end[-1] == POISON_INUSE) 823 end--; 824 825 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 826 print_section(KERN_ERR, "Padding ", end - remainder, remainder); 827 828 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end); 829 return 0; 830 } 831 832 static int check_object(struct kmem_cache *s, struct page *page, 833 void *object, u8 val) 834 { 835 u8 *p = object; 836 u8 *endobject = object + s->object_size; 837 838 if (s->flags & SLAB_RED_ZONE) { 839 if (!check_bytes_and_report(s, page, object, "Redzone", 840 object - s->red_left_pad, val, s->red_left_pad)) 841 return 0; 842 843 if (!check_bytes_and_report(s, page, object, "Redzone", 844 endobject, val, s->inuse - s->object_size)) 845 return 0; 846 } else { 847 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { 848 check_bytes_and_report(s, page, p, "Alignment padding", 849 endobject, POISON_INUSE, 850 s->inuse - s->object_size); 851 } 852 } 853 854 if (s->flags & SLAB_POISON) { 855 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 856 (!check_bytes_and_report(s, page, p, "Poison", p, 857 POISON_FREE, s->object_size - 1) || 858 !check_bytes_and_report(s, page, p, "Poison", 859 p + s->object_size - 1, POISON_END, 1))) 860 return 0; 861 /* 862 * check_pad_bytes cleans up on its own. 863 */ 864 check_pad_bytes(s, page, p); 865 } 866 867 if (!s->offset && val == SLUB_RED_ACTIVE) 868 /* 869 * Object and freepointer overlap. Cannot check 870 * freepointer while object is allocated. 871 */ 872 return 1; 873 874 /* Check free pointer validity */ 875 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 876 object_err(s, page, p, "Freepointer corrupt"); 877 /* 878 * No choice but to zap it and thus lose the remainder 879 * of the free objects in this slab. May cause 880 * another error because the object count is now wrong. 881 */ 882 set_freepointer(s, p, NULL); 883 return 0; 884 } 885 return 1; 886 } 887 888 static int check_slab(struct kmem_cache *s, struct page *page) 889 { 890 int maxobj; 891 892 VM_BUG_ON(!irqs_disabled()); 893 894 if (!PageSlab(page)) { 895 slab_err(s, page, "Not a valid slab page"); 896 return 0; 897 } 898 899 maxobj = order_objects(compound_order(page), s->size, s->reserved); 900 if (page->objects > maxobj) { 901 slab_err(s, page, "objects %u > max %u", 902 page->objects, maxobj); 903 return 0; 904 } 905 if (page->inuse > page->objects) { 906 slab_err(s, page, "inuse %u > max %u", 907 page->inuse, page->objects); 908 return 0; 909 } 910 /* Slab_pad_check fixes things up after itself */ 911 slab_pad_check(s, page); 912 return 1; 913 } 914 915 /* 916 * Determine if a certain object on a page is on the freelist. Must hold the 917 * slab lock to guarantee that the chains are in a consistent state. 918 */ 919 static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 920 { 921 int nr = 0; 922 void *fp; 923 void *object = NULL; 924 int max_objects; 925 926 fp = page->freelist; 927 while (fp && nr <= page->objects) { 928 if (fp == search) 929 return 1; 930 if (!check_valid_pointer(s, page, fp)) { 931 if (object) { 932 object_err(s, page, object, 933 "Freechain corrupt"); 934 set_freepointer(s, object, NULL); 935 } else { 936 slab_err(s, page, "Freepointer corrupt"); 937 page->freelist = NULL; 938 page->inuse = page->objects; 939 slab_fix(s, "Freelist cleared"); 940 return 0; 941 } 942 break; 943 } 944 object = fp; 945 fp = get_freepointer(s, object); 946 nr++; 947 } 948 949 max_objects = order_objects(compound_order(page), s->size, s->reserved); 950 if (max_objects > MAX_OBJS_PER_PAGE) 951 max_objects = MAX_OBJS_PER_PAGE; 952 953 if (page->objects != max_objects) { 954 slab_err(s, page, "Wrong number of objects. Found %d but should be %d", 955 page->objects, max_objects); 956 page->objects = max_objects; 957 slab_fix(s, "Number of objects adjusted."); 958 } 959 if (page->inuse != page->objects - nr) { 960 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d", 961 page->inuse, page->objects - nr); 962 page->inuse = page->objects - nr; 963 slab_fix(s, "Object count adjusted."); 964 } 965 return search == NULL; 966 } 967 968 static void trace(struct kmem_cache *s, struct page *page, void *object, 969 int alloc) 970 { 971 if (s->flags & SLAB_TRACE) { 972 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 973 s->name, 974 alloc ? "alloc" : "free", 975 object, page->inuse, 976 page->freelist); 977 978 if (!alloc) 979 print_section(KERN_INFO, "Object ", (void *)object, 980 s->object_size); 981 982 dump_stack(); 983 } 984 } 985 986 /* 987 * Tracking of fully allocated slabs for debugging purposes. 988 */ 989 static void add_full(struct kmem_cache *s, 990 struct kmem_cache_node *n, struct page *page) 991 { 992 if (!(s->flags & SLAB_STORE_USER)) 993 return; 994 995 lockdep_assert_held(&n->list_lock); 996 list_add(&page->lru, &n->full); 997 } 998 999 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page) 1000 { 1001 if (!(s->flags & SLAB_STORE_USER)) 1002 return; 1003 1004 lockdep_assert_held(&n->list_lock); 1005 list_del(&page->lru); 1006 } 1007 1008 /* Tracking of the number of slabs for debugging purposes */ 1009 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1010 { 1011 struct kmem_cache_node *n = get_node(s, node); 1012 1013 return atomic_long_read(&n->nr_slabs); 1014 } 1015 1016 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1017 { 1018 return atomic_long_read(&n->nr_slabs); 1019 } 1020 1021 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 1022 { 1023 struct kmem_cache_node *n = get_node(s, node); 1024 1025 /* 1026 * May be called early in order to allocate a slab for the 1027 * kmem_cache_node structure. Solve the chicken-egg 1028 * dilemma by deferring the increment of the count during 1029 * bootstrap (see early_kmem_cache_node_alloc). 1030 */ 1031 if (likely(n)) { 1032 atomic_long_inc(&n->nr_slabs); 1033 atomic_long_add(objects, &n->total_objects); 1034 } 1035 } 1036 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 1037 { 1038 struct kmem_cache_node *n = get_node(s, node); 1039 1040 atomic_long_dec(&n->nr_slabs); 1041 atomic_long_sub(objects, &n->total_objects); 1042 } 1043 1044 /* Object debug checks for alloc/free paths */ 1045 static void setup_object_debug(struct kmem_cache *s, struct page *page, 1046 void *object) 1047 { 1048 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 1049 return; 1050 1051 init_object(s, object, SLUB_RED_INACTIVE); 1052 init_tracking(s, object); 1053 } 1054 1055 static inline int alloc_consistency_checks(struct kmem_cache *s, 1056 struct page *page, 1057 void *object, unsigned long addr) 1058 { 1059 if (!check_slab(s, page)) 1060 return 0; 1061 1062 if (!check_valid_pointer(s, page, object)) { 1063 object_err(s, page, object, "Freelist Pointer check fails"); 1064 return 0; 1065 } 1066 1067 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1068 return 0; 1069 1070 return 1; 1071 } 1072 1073 static noinline int alloc_debug_processing(struct kmem_cache *s, 1074 struct page *page, 1075 void *object, unsigned long addr) 1076 { 1077 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1078 if (!alloc_consistency_checks(s, page, object, addr)) 1079 goto bad; 1080 } 1081 1082 /* Success perform special debug activities for allocs */ 1083 if (s->flags & SLAB_STORE_USER) 1084 set_track(s, object, TRACK_ALLOC, addr); 1085 trace(s, page, object, 1); 1086 init_object(s, object, SLUB_RED_ACTIVE); 1087 return 1; 1088 1089 bad: 1090 if (PageSlab(page)) { 1091 /* 1092 * If this is a slab page then lets do the best we can 1093 * to avoid issues in the future. Marking all objects 1094 * as used avoids touching the remaining objects. 1095 */ 1096 slab_fix(s, "Marking all objects used"); 1097 page->inuse = page->objects; 1098 page->freelist = NULL; 1099 } 1100 return 0; 1101 } 1102 1103 static inline int free_consistency_checks(struct kmem_cache *s, 1104 struct page *page, void *object, unsigned long addr) 1105 { 1106 if (!check_valid_pointer(s, page, object)) { 1107 slab_err(s, page, "Invalid object pointer 0x%p", object); 1108 return 0; 1109 } 1110 1111 if (on_freelist(s, page, object)) { 1112 object_err(s, page, object, "Object already free"); 1113 return 0; 1114 } 1115 1116 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1117 return 0; 1118 1119 if (unlikely(s != page->slab_cache)) { 1120 if (!PageSlab(page)) { 1121 slab_err(s, page, "Attempt to free object(0x%p) outside of slab", 1122 object); 1123 } else if (!page->slab_cache) { 1124 pr_err("SLUB <none>: no slab for object 0x%p.\n", 1125 object); 1126 dump_stack(); 1127 } else 1128 object_err(s, page, object, 1129 "page slab pointer corrupt."); 1130 return 0; 1131 } 1132 return 1; 1133 } 1134 1135 /* Supports checking bulk free of a constructed freelist */ 1136 static noinline int free_debug_processing( 1137 struct kmem_cache *s, struct page *page, 1138 void *head, void *tail, int bulk_cnt, 1139 unsigned long addr) 1140 { 1141 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1142 void *object = head; 1143 int cnt = 0; 1144 unsigned long uninitialized_var(flags); 1145 int ret = 0; 1146 1147 spin_lock_irqsave(&n->list_lock, flags); 1148 slab_lock(page); 1149 1150 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1151 if (!check_slab(s, page)) 1152 goto out; 1153 } 1154 1155 next_object: 1156 cnt++; 1157 1158 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1159 if (!free_consistency_checks(s, page, object, addr)) 1160 goto out; 1161 } 1162 1163 if (s->flags & SLAB_STORE_USER) 1164 set_track(s, object, TRACK_FREE, addr); 1165 trace(s, page, object, 0); 1166 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ 1167 init_object(s, object, SLUB_RED_INACTIVE); 1168 1169 /* Reached end of constructed freelist yet? */ 1170 if (object != tail) { 1171 object = get_freepointer(s, object); 1172 goto next_object; 1173 } 1174 ret = 1; 1175 1176 out: 1177 if (cnt != bulk_cnt) 1178 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n", 1179 bulk_cnt, cnt); 1180 1181 slab_unlock(page); 1182 spin_unlock_irqrestore(&n->list_lock, flags); 1183 if (!ret) 1184 slab_fix(s, "Object at 0x%p not freed", object); 1185 return ret; 1186 } 1187 1188 static int __init setup_slub_debug(char *str) 1189 { 1190 slub_debug = DEBUG_DEFAULT_FLAGS; 1191 if (*str++ != '=' || !*str) 1192 /* 1193 * No options specified. Switch on full debugging. 1194 */ 1195 goto out; 1196 1197 if (*str == ',') 1198 /* 1199 * No options but restriction on slabs. This means full 1200 * debugging for slabs matching a pattern. 1201 */ 1202 goto check_slabs; 1203 1204 slub_debug = 0; 1205 if (*str == '-') 1206 /* 1207 * Switch off all debugging measures. 1208 */ 1209 goto out; 1210 1211 /* 1212 * Determine which debug features should be switched on 1213 */ 1214 for (; *str && *str != ','; str++) { 1215 switch (tolower(*str)) { 1216 case 'f': 1217 slub_debug |= SLAB_CONSISTENCY_CHECKS; 1218 break; 1219 case 'z': 1220 slub_debug |= SLAB_RED_ZONE; 1221 break; 1222 case 'p': 1223 slub_debug |= SLAB_POISON; 1224 break; 1225 case 'u': 1226 slub_debug |= SLAB_STORE_USER; 1227 break; 1228 case 't': 1229 slub_debug |= SLAB_TRACE; 1230 break; 1231 case 'a': 1232 slub_debug |= SLAB_FAILSLAB; 1233 break; 1234 case 'o': 1235 /* 1236 * Avoid enabling debugging on caches if its minimum 1237 * order would increase as a result. 1238 */ 1239 disable_higher_order_debug = 1; 1240 break; 1241 default: 1242 pr_err("slub_debug option '%c' unknown. skipped\n", 1243 *str); 1244 } 1245 } 1246 1247 check_slabs: 1248 if (*str == ',') 1249 slub_debug_slabs = str + 1; 1250 out: 1251 return 1; 1252 } 1253 1254 __setup("slub_debug", setup_slub_debug); 1255 1256 unsigned long kmem_cache_flags(unsigned long object_size, 1257 unsigned long flags, const char *name, 1258 void (*ctor)(void *)) 1259 { 1260 /* 1261 * Enable debugging if selected on the kernel commandline. 1262 */ 1263 if (slub_debug && (!slub_debug_slabs || (name && 1264 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))) 1265 flags |= slub_debug; 1266 1267 return flags; 1268 } 1269 #else /* !CONFIG_SLUB_DEBUG */ 1270 static inline void setup_object_debug(struct kmem_cache *s, 1271 struct page *page, void *object) {} 1272 1273 static inline int alloc_debug_processing(struct kmem_cache *s, 1274 struct page *page, void *object, unsigned long addr) { return 0; } 1275 1276 static inline int free_debug_processing( 1277 struct kmem_cache *s, struct page *page, 1278 void *head, void *tail, int bulk_cnt, 1279 unsigned long addr) { return 0; } 1280 1281 static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1282 { return 1; } 1283 static inline int check_object(struct kmem_cache *s, struct page *page, 1284 void *object, u8 val) { return 1; } 1285 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, 1286 struct page *page) {} 1287 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, 1288 struct page *page) {} 1289 unsigned long kmem_cache_flags(unsigned long object_size, 1290 unsigned long flags, const char *name, 1291 void (*ctor)(void *)) 1292 { 1293 return flags; 1294 } 1295 #define slub_debug 0 1296 1297 #define disable_higher_order_debug 0 1298 1299 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1300 { return 0; } 1301 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1302 { return 0; } 1303 static inline void inc_slabs_node(struct kmem_cache *s, int node, 1304 int objects) {} 1305 static inline void dec_slabs_node(struct kmem_cache *s, int node, 1306 int objects) {} 1307 1308 #endif /* CONFIG_SLUB_DEBUG */ 1309 1310 /* 1311 * Hooks for other subsystems that check memory allocations. In a typical 1312 * production configuration these hooks all should produce no code at all. 1313 */ 1314 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) 1315 { 1316 kmemleak_alloc(ptr, size, 1, flags); 1317 kasan_kmalloc_large(ptr, size, flags); 1318 } 1319 1320 static inline void kfree_hook(const void *x) 1321 { 1322 kmemleak_free(x); 1323 kasan_kfree_large(x); 1324 } 1325 1326 static inline void *slab_free_hook(struct kmem_cache *s, void *x) 1327 { 1328 void *freeptr; 1329 1330 kmemleak_free_recursive(x, s->flags); 1331 1332 /* 1333 * Trouble is that we may no longer disable interrupts in the fast path 1334 * So in order to make the debug calls that expect irqs to be 1335 * disabled we need to disable interrupts temporarily. 1336 */ 1337 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP) 1338 { 1339 unsigned long flags; 1340 1341 local_irq_save(flags); 1342 kmemcheck_slab_free(s, x, s->object_size); 1343 debug_check_no_locks_freed(x, s->object_size); 1344 local_irq_restore(flags); 1345 } 1346 #endif 1347 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1348 debug_check_no_obj_freed(x, s->object_size); 1349 1350 freeptr = get_freepointer(s, x); 1351 /* 1352 * kasan_slab_free() may put x into memory quarantine, delaying its 1353 * reuse. In this case the object's freelist pointer is changed. 1354 */ 1355 kasan_slab_free(s, x); 1356 return freeptr; 1357 } 1358 1359 static inline void slab_free_freelist_hook(struct kmem_cache *s, 1360 void *head, void *tail) 1361 { 1362 /* 1363 * Compiler cannot detect this function can be removed if slab_free_hook() 1364 * evaluates to nothing. Thus, catch all relevant config debug options here. 1365 */ 1366 #if defined(CONFIG_KMEMCHECK) || \ 1367 defined(CONFIG_LOCKDEP) || \ 1368 defined(CONFIG_DEBUG_KMEMLEAK) || \ 1369 defined(CONFIG_DEBUG_OBJECTS_FREE) || \ 1370 defined(CONFIG_KASAN) 1371 1372 void *object = head; 1373 void *tail_obj = tail ? : head; 1374 void *freeptr; 1375 1376 do { 1377 freeptr = slab_free_hook(s, object); 1378 } while ((object != tail_obj) && (object = freeptr)); 1379 #endif 1380 } 1381 1382 static void setup_object(struct kmem_cache *s, struct page *page, 1383 void *object) 1384 { 1385 setup_object_debug(s, page, object); 1386 kasan_init_slab_obj(s, object); 1387 if (unlikely(s->ctor)) { 1388 kasan_unpoison_object_data(s, object); 1389 s->ctor(object); 1390 kasan_poison_object_data(s, object); 1391 } 1392 } 1393 1394 /* 1395 * Slab allocation and freeing 1396 */ 1397 static inline struct page *alloc_slab_page(struct kmem_cache *s, 1398 gfp_t flags, int node, struct kmem_cache_order_objects oo) 1399 { 1400 struct page *page; 1401 int order = oo_order(oo); 1402 1403 flags |= __GFP_NOTRACK; 1404 1405 if (node == NUMA_NO_NODE) 1406 page = alloc_pages(flags, order); 1407 else 1408 page = __alloc_pages_node(node, flags, order); 1409 1410 if (page && memcg_charge_slab(page, flags, order, s)) { 1411 __free_pages(page, order); 1412 page = NULL; 1413 } 1414 1415 return page; 1416 } 1417 1418 #ifdef CONFIG_SLAB_FREELIST_RANDOM 1419 /* Pre-initialize the random sequence cache */ 1420 static int init_cache_random_seq(struct kmem_cache *s) 1421 { 1422 int err; 1423 unsigned long i, count = oo_objects(s->oo); 1424 1425 err = cache_random_seq_create(s, count, GFP_KERNEL); 1426 if (err) { 1427 pr_err("SLUB: Unable to initialize free list for %s\n", 1428 s->name); 1429 return err; 1430 } 1431 1432 /* Transform to an offset on the set of pages */ 1433 if (s->random_seq) { 1434 for (i = 0; i < count; i++) 1435 s->random_seq[i] *= s->size; 1436 } 1437 return 0; 1438 } 1439 1440 /* Initialize each random sequence freelist per cache */ 1441 static void __init init_freelist_randomization(void) 1442 { 1443 struct kmem_cache *s; 1444 1445 mutex_lock(&slab_mutex); 1446 1447 list_for_each_entry(s, &slab_caches, list) 1448 init_cache_random_seq(s); 1449 1450 mutex_unlock(&slab_mutex); 1451 } 1452 1453 /* Get the next entry on the pre-computed freelist randomized */ 1454 static void *next_freelist_entry(struct kmem_cache *s, struct page *page, 1455 unsigned long *pos, void *start, 1456 unsigned long page_limit, 1457 unsigned long freelist_count) 1458 { 1459 unsigned int idx; 1460 1461 /* 1462 * If the target page allocation failed, the number of objects on the 1463 * page might be smaller than the usual size defined by the cache. 1464 */ 1465 do { 1466 idx = s->random_seq[*pos]; 1467 *pos += 1; 1468 if (*pos >= freelist_count) 1469 *pos = 0; 1470 } while (unlikely(idx >= page_limit)); 1471 1472 return (char *)start + idx; 1473 } 1474 1475 /* Shuffle the single linked freelist based on a random pre-computed sequence */ 1476 static bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1477 { 1478 void *start; 1479 void *cur; 1480 void *next; 1481 unsigned long idx, pos, page_limit, freelist_count; 1482 1483 if (page->objects < 2 || !s->random_seq) 1484 return false; 1485 1486 freelist_count = oo_objects(s->oo); 1487 pos = get_random_int() % freelist_count; 1488 1489 page_limit = page->objects * s->size; 1490 start = fixup_red_left(s, page_address(page)); 1491 1492 /* First entry is used as the base of the freelist */ 1493 cur = next_freelist_entry(s, page, &pos, start, page_limit, 1494 freelist_count); 1495 page->freelist = cur; 1496 1497 for (idx = 1; idx < page->objects; idx++) { 1498 setup_object(s, page, cur); 1499 next = next_freelist_entry(s, page, &pos, start, page_limit, 1500 freelist_count); 1501 set_freepointer(s, cur, next); 1502 cur = next; 1503 } 1504 setup_object(s, page, cur); 1505 set_freepointer(s, cur, NULL); 1506 1507 return true; 1508 } 1509 #else 1510 static inline int init_cache_random_seq(struct kmem_cache *s) 1511 { 1512 return 0; 1513 } 1514 static inline void init_freelist_randomization(void) { } 1515 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1516 { 1517 return false; 1518 } 1519 #endif /* CONFIG_SLAB_FREELIST_RANDOM */ 1520 1521 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1522 { 1523 struct page *page; 1524 struct kmem_cache_order_objects oo = s->oo; 1525 gfp_t alloc_gfp; 1526 void *start, *p; 1527 int idx, order; 1528 bool shuffle; 1529 1530 flags &= gfp_allowed_mask; 1531 1532 if (gfpflags_allow_blocking(flags)) 1533 local_irq_enable(); 1534 1535 flags |= s->allocflags; 1536 1537 /* 1538 * Let the initial higher-order allocation fail under memory pressure 1539 * so we fall-back to the minimum order allocation. 1540 */ 1541 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1542 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min)) 1543 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL); 1544 1545 page = alloc_slab_page(s, alloc_gfp, node, oo); 1546 if (unlikely(!page)) { 1547 oo = s->min; 1548 alloc_gfp = flags; 1549 /* 1550 * Allocation may have failed due to fragmentation. 1551 * Try a lower order alloc if possible 1552 */ 1553 page = alloc_slab_page(s, alloc_gfp, node, oo); 1554 if (unlikely(!page)) 1555 goto out; 1556 stat(s, ORDER_FALLBACK); 1557 } 1558 1559 if (kmemcheck_enabled && 1560 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) { 1561 int pages = 1 << oo_order(oo); 1562 1563 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node); 1564 1565 /* 1566 * Objects from caches that have a constructor don't get 1567 * cleared when they're allocated, so we need to do it here. 1568 */ 1569 if (s->ctor) 1570 kmemcheck_mark_uninitialized_pages(page, pages); 1571 else 1572 kmemcheck_mark_unallocated_pages(page, pages); 1573 } 1574 1575 page->objects = oo_objects(oo); 1576 1577 order = compound_order(page); 1578 page->slab_cache = s; 1579 __SetPageSlab(page); 1580 if (page_is_pfmemalloc(page)) 1581 SetPageSlabPfmemalloc(page); 1582 1583 start = page_address(page); 1584 1585 if (unlikely(s->flags & SLAB_POISON)) 1586 memset(start, POISON_INUSE, PAGE_SIZE << order); 1587 1588 kasan_poison_slab(page); 1589 1590 shuffle = shuffle_freelist(s, page); 1591 1592 if (!shuffle) { 1593 for_each_object_idx(p, idx, s, start, page->objects) { 1594 setup_object(s, page, p); 1595 if (likely(idx < page->objects)) 1596 set_freepointer(s, p, p + s->size); 1597 else 1598 set_freepointer(s, p, NULL); 1599 } 1600 page->freelist = fixup_red_left(s, start); 1601 } 1602 1603 page->inuse = page->objects; 1604 page->frozen = 1; 1605 1606 out: 1607 if (gfpflags_allow_blocking(flags)) 1608 local_irq_disable(); 1609 if (!page) 1610 return NULL; 1611 1612 mod_zone_page_state(page_zone(page), 1613 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1614 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1615 1 << oo_order(oo)); 1616 1617 inc_slabs_node(s, page_to_nid(page), page->objects); 1618 1619 return page; 1620 } 1621 1622 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1623 { 1624 if (unlikely(flags & GFP_SLAB_BUG_MASK)) { 1625 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK; 1626 flags &= ~GFP_SLAB_BUG_MASK; 1627 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n", 1628 invalid_mask, &invalid_mask, flags, &flags); 1629 } 1630 1631 return allocate_slab(s, 1632 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1633 } 1634 1635 static void __free_slab(struct kmem_cache *s, struct page *page) 1636 { 1637 int order = compound_order(page); 1638 int pages = 1 << order; 1639 1640 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1641 void *p; 1642 1643 slab_pad_check(s, page); 1644 for_each_object(p, s, page_address(page), 1645 page->objects) 1646 check_object(s, page, p, SLUB_RED_INACTIVE); 1647 } 1648 1649 kmemcheck_free_shadow(page, compound_order(page)); 1650 1651 mod_zone_page_state(page_zone(page), 1652 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1653 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1654 -pages); 1655 1656 __ClearPageSlabPfmemalloc(page); 1657 __ClearPageSlab(page); 1658 1659 page_mapcount_reset(page); 1660 if (current->reclaim_state) 1661 current->reclaim_state->reclaimed_slab += pages; 1662 memcg_uncharge_slab(page, order, s); 1663 __free_pages(page, order); 1664 } 1665 1666 #define need_reserve_slab_rcu \ 1667 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head)) 1668 1669 static void rcu_free_slab(struct rcu_head *h) 1670 { 1671 struct page *page; 1672 1673 if (need_reserve_slab_rcu) 1674 page = virt_to_head_page(h); 1675 else 1676 page = container_of((struct list_head *)h, struct page, lru); 1677 1678 __free_slab(page->slab_cache, page); 1679 } 1680 1681 static void free_slab(struct kmem_cache *s, struct page *page) 1682 { 1683 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) { 1684 struct rcu_head *head; 1685 1686 if (need_reserve_slab_rcu) { 1687 int order = compound_order(page); 1688 int offset = (PAGE_SIZE << order) - s->reserved; 1689 1690 VM_BUG_ON(s->reserved != sizeof(*head)); 1691 head = page_address(page) + offset; 1692 } else { 1693 head = &page->rcu_head; 1694 } 1695 1696 call_rcu(head, rcu_free_slab); 1697 } else 1698 __free_slab(s, page); 1699 } 1700 1701 static void discard_slab(struct kmem_cache *s, struct page *page) 1702 { 1703 dec_slabs_node(s, page_to_nid(page), page->objects); 1704 free_slab(s, page); 1705 } 1706 1707 /* 1708 * Management of partially allocated slabs. 1709 */ 1710 static inline void 1711 __add_partial(struct kmem_cache_node *n, struct page *page, int tail) 1712 { 1713 n->nr_partial++; 1714 if (tail == DEACTIVATE_TO_TAIL) 1715 list_add_tail(&page->lru, &n->partial); 1716 else 1717 list_add(&page->lru, &n->partial); 1718 } 1719 1720 static inline void add_partial(struct kmem_cache_node *n, 1721 struct page *page, int tail) 1722 { 1723 lockdep_assert_held(&n->list_lock); 1724 __add_partial(n, page, tail); 1725 } 1726 1727 static inline void remove_partial(struct kmem_cache_node *n, 1728 struct page *page) 1729 { 1730 lockdep_assert_held(&n->list_lock); 1731 list_del(&page->lru); 1732 n->nr_partial--; 1733 } 1734 1735 /* 1736 * Remove slab from the partial list, freeze it and 1737 * return the pointer to the freelist. 1738 * 1739 * Returns a list of objects or NULL if it fails. 1740 */ 1741 static inline void *acquire_slab(struct kmem_cache *s, 1742 struct kmem_cache_node *n, struct page *page, 1743 int mode, int *objects) 1744 { 1745 void *freelist; 1746 unsigned long counters; 1747 struct page new; 1748 1749 lockdep_assert_held(&n->list_lock); 1750 1751 /* 1752 * Zap the freelist and set the frozen bit. 1753 * The old freelist is the list of objects for the 1754 * per cpu allocation list. 1755 */ 1756 freelist = page->freelist; 1757 counters = page->counters; 1758 new.counters = counters; 1759 *objects = new.objects - new.inuse; 1760 if (mode) { 1761 new.inuse = page->objects; 1762 new.freelist = NULL; 1763 } else { 1764 new.freelist = freelist; 1765 } 1766 1767 VM_BUG_ON(new.frozen); 1768 new.frozen = 1; 1769 1770 if (!__cmpxchg_double_slab(s, page, 1771 freelist, counters, 1772 new.freelist, new.counters, 1773 "acquire_slab")) 1774 return NULL; 1775 1776 remove_partial(n, page); 1777 WARN_ON(!freelist); 1778 return freelist; 1779 } 1780 1781 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); 1782 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags); 1783 1784 /* 1785 * Try to allocate a partial slab from a specific node. 1786 */ 1787 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, 1788 struct kmem_cache_cpu *c, gfp_t flags) 1789 { 1790 struct page *page, *page2; 1791 void *object = NULL; 1792 int available = 0; 1793 int objects; 1794 1795 /* 1796 * Racy check. If we mistakenly see no partial slabs then we 1797 * just allocate an empty slab. If we mistakenly try to get a 1798 * partial slab and there is none available then get_partials() 1799 * will return NULL. 1800 */ 1801 if (!n || !n->nr_partial) 1802 return NULL; 1803 1804 spin_lock(&n->list_lock); 1805 list_for_each_entry_safe(page, page2, &n->partial, lru) { 1806 void *t; 1807 1808 if (!pfmemalloc_match(page, flags)) 1809 continue; 1810 1811 t = acquire_slab(s, n, page, object == NULL, &objects); 1812 if (!t) 1813 break; 1814 1815 available += objects; 1816 if (!object) { 1817 c->page = page; 1818 stat(s, ALLOC_FROM_PARTIAL); 1819 object = t; 1820 } else { 1821 put_cpu_partial(s, page, 0); 1822 stat(s, CPU_PARTIAL_NODE); 1823 } 1824 if (!kmem_cache_has_cpu_partial(s) 1825 || available > s->cpu_partial / 2) 1826 break; 1827 1828 } 1829 spin_unlock(&n->list_lock); 1830 return object; 1831 } 1832 1833 /* 1834 * Get a page from somewhere. Search in increasing NUMA distances. 1835 */ 1836 static void *get_any_partial(struct kmem_cache *s, gfp_t flags, 1837 struct kmem_cache_cpu *c) 1838 { 1839 #ifdef CONFIG_NUMA 1840 struct zonelist *zonelist; 1841 struct zoneref *z; 1842 struct zone *zone; 1843 enum zone_type high_zoneidx = gfp_zone(flags); 1844 void *object; 1845 unsigned int cpuset_mems_cookie; 1846 1847 /* 1848 * The defrag ratio allows a configuration of the tradeoffs between 1849 * inter node defragmentation and node local allocations. A lower 1850 * defrag_ratio increases the tendency to do local allocations 1851 * instead of attempting to obtain partial slabs from other nodes. 1852 * 1853 * If the defrag_ratio is set to 0 then kmalloc() always 1854 * returns node local objects. If the ratio is higher then kmalloc() 1855 * may return off node objects because partial slabs are obtained 1856 * from other nodes and filled up. 1857 * 1858 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 1859 * (which makes defrag_ratio = 1000) then every (well almost) 1860 * allocation will first attempt to defrag slab caches on other nodes. 1861 * This means scanning over all nodes to look for partial slabs which 1862 * may be expensive if we do it every time we are trying to find a slab 1863 * with available objects. 1864 */ 1865 if (!s->remote_node_defrag_ratio || 1866 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1867 return NULL; 1868 1869 do { 1870 cpuset_mems_cookie = read_mems_allowed_begin(); 1871 zonelist = node_zonelist(mempolicy_slab_node(), flags); 1872 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1873 struct kmem_cache_node *n; 1874 1875 n = get_node(s, zone_to_nid(zone)); 1876 1877 if (n && cpuset_zone_allowed(zone, flags) && 1878 n->nr_partial > s->min_partial) { 1879 object = get_partial_node(s, n, c, flags); 1880 if (object) { 1881 /* 1882 * Don't check read_mems_allowed_retry() 1883 * here - if mems_allowed was updated in 1884 * parallel, that was a harmless race 1885 * between allocation and the cpuset 1886 * update 1887 */ 1888 return object; 1889 } 1890 } 1891 } 1892 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1893 #endif 1894 return NULL; 1895 } 1896 1897 /* 1898 * Get a partial page, lock it and return it. 1899 */ 1900 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, 1901 struct kmem_cache_cpu *c) 1902 { 1903 void *object; 1904 int searchnode = node; 1905 1906 if (node == NUMA_NO_NODE) 1907 searchnode = numa_mem_id(); 1908 else if (!node_present_pages(node)) 1909 searchnode = node_to_mem_node(node); 1910 1911 object = get_partial_node(s, get_node(s, searchnode), c, flags); 1912 if (object || node != NUMA_NO_NODE) 1913 return object; 1914 1915 return get_any_partial(s, flags, c); 1916 } 1917 1918 #ifdef CONFIG_PREEMPT 1919 /* 1920 * Calculate the next globally unique transaction for disambiguiation 1921 * during cmpxchg. The transactions start with the cpu number and are then 1922 * incremented by CONFIG_NR_CPUS. 1923 */ 1924 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 1925 #else 1926 /* 1927 * No preemption supported therefore also no need to check for 1928 * different cpus. 1929 */ 1930 #define TID_STEP 1 1931 #endif 1932 1933 static inline unsigned long next_tid(unsigned long tid) 1934 { 1935 return tid + TID_STEP; 1936 } 1937 1938 static inline unsigned int tid_to_cpu(unsigned long tid) 1939 { 1940 return tid % TID_STEP; 1941 } 1942 1943 static inline unsigned long tid_to_event(unsigned long tid) 1944 { 1945 return tid / TID_STEP; 1946 } 1947 1948 static inline unsigned int init_tid(int cpu) 1949 { 1950 return cpu; 1951 } 1952 1953 static inline void note_cmpxchg_failure(const char *n, 1954 const struct kmem_cache *s, unsigned long tid) 1955 { 1956 #ifdef SLUB_DEBUG_CMPXCHG 1957 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 1958 1959 pr_info("%s %s: cmpxchg redo ", n, s->name); 1960 1961 #ifdef CONFIG_PREEMPT 1962 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 1963 pr_warn("due to cpu change %d -> %d\n", 1964 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 1965 else 1966 #endif 1967 if (tid_to_event(tid) != tid_to_event(actual_tid)) 1968 pr_warn("due to cpu running other code. Event %ld->%ld\n", 1969 tid_to_event(tid), tid_to_event(actual_tid)); 1970 else 1971 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n", 1972 actual_tid, tid, next_tid(tid)); 1973 #endif 1974 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 1975 } 1976 1977 static void init_kmem_cache_cpus(struct kmem_cache *s) 1978 { 1979 int cpu; 1980 1981 for_each_possible_cpu(cpu) 1982 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 1983 } 1984 1985 /* 1986 * Remove the cpu slab 1987 */ 1988 static void deactivate_slab(struct kmem_cache *s, struct page *page, 1989 void *freelist) 1990 { 1991 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; 1992 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1993 int lock = 0; 1994 enum slab_modes l = M_NONE, m = M_NONE; 1995 void *nextfree; 1996 int tail = DEACTIVATE_TO_HEAD; 1997 struct page new; 1998 struct page old; 1999 2000 if (page->freelist) { 2001 stat(s, DEACTIVATE_REMOTE_FREES); 2002 tail = DEACTIVATE_TO_TAIL; 2003 } 2004 2005 /* 2006 * Stage one: Free all available per cpu objects back 2007 * to the page freelist while it is still frozen. Leave the 2008 * last one. 2009 * 2010 * There is no need to take the list->lock because the page 2011 * is still frozen. 2012 */ 2013 while (freelist && (nextfree = get_freepointer(s, freelist))) { 2014 void *prior; 2015 unsigned long counters; 2016 2017 do { 2018 prior = page->freelist; 2019 counters = page->counters; 2020 set_freepointer(s, freelist, prior); 2021 new.counters = counters; 2022 new.inuse--; 2023 VM_BUG_ON(!new.frozen); 2024 2025 } while (!__cmpxchg_double_slab(s, page, 2026 prior, counters, 2027 freelist, new.counters, 2028 "drain percpu freelist")); 2029 2030 freelist = nextfree; 2031 } 2032 2033 /* 2034 * Stage two: Ensure that the page is unfrozen while the 2035 * list presence reflects the actual number of objects 2036 * during unfreeze. 2037 * 2038 * We setup the list membership and then perform a cmpxchg 2039 * with the count. If there is a mismatch then the page 2040 * is not unfrozen but the page is on the wrong list. 2041 * 2042 * Then we restart the process which may have to remove 2043 * the page from the list that we just put it on again 2044 * because the number of objects in the slab may have 2045 * changed. 2046 */ 2047 redo: 2048 2049 old.freelist = page->freelist; 2050 old.counters = page->counters; 2051 VM_BUG_ON(!old.frozen); 2052 2053 /* Determine target state of the slab */ 2054 new.counters = old.counters; 2055 if (freelist) { 2056 new.inuse--; 2057 set_freepointer(s, freelist, old.freelist); 2058 new.freelist = freelist; 2059 } else 2060 new.freelist = old.freelist; 2061 2062 new.frozen = 0; 2063 2064 if (!new.inuse && n->nr_partial >= s->min_partial) 2065 m = M_FREE; 2066 else if (new.freelist) { 2067 m = M_PARTIAL; 2068 if (!lock) { 2069 lock = 1; 2070 /* 2071 * Taking the spinlock removes the possiblity 2072 * that acquire_slab() will see a slab page that 2073 * is frozen 2074 */ 2075 spin_lock(&n->list_lock); 2076 } 2077 } else { 2078 m = M_FULL; 2079 if (kmem_cache_debug(s) && !lock) { 2080 lock = 1; 2081 /* 2082 * This also ensures that the scanning of full 2083 * slabs from diagnostic functions will not see 2084 * any frozen slabs. 2085 */ 2086 spin_lock(&n->list_lock); 2087 } 2088 } 2089 2090 if (l != m) { 2091 2092 if (l == M_PARTIAL) 2093 2094 remove_partial(n, page); 2095 2096 else if (l == M_FULL) 2097 2098 remove_full(s, n, page); 2099 2100 if (m == M_PARTIAL) { 2101 2102 add_partial(n, page, tail); 2103 stat(s, tail); 2104 2105 } else if (m == M_FULL) { 2106 2107 stat(s, DEACTIVATE_FULL); 2108 add_full(s, n, page); 2109 2110 } 2111 } 2112 2113 l = m; 2114 if (!__cmpxchg_double_slab(s, page, 2115 old.freelist, old.counters, 2116 new.freelist, new.counters, 2117 "unfreezing slab")) 2118 goto redo; 2119 2120 if (lock) 2121 spin_unlock(&n->list_lock); 2122 2123 if (m == M_FREE) { 2124 stat(s, DEACTIVATE_EMPTY); 2125 discard_slab(s, page); 2126 stat(s, FREE_SLAB); 2127 } 2128 } 2129 2130 /* 2131 * Unfreeze all the cpu partial slabs. 2132 * 2133 * This function must be called with interrupts disabled 2134 * for the cpu using c (or some other guarantee must be there 2135 * to guarantee no concurrent accesses). 2136 */ 2137 static void unfreeze_partials(struct kmem_cache *s, 2138 struct kmem_cache_cpu *c) 2139 { 2140 #ifdef CONFIG_SLUB_CPU_PARTIAL 2141 struct kmem_cache_node *n = NULL, *n2 = NULL; 2142 struct page *page, *discard_page = NULL; 2143 2144 while ((page = c->partial)) { 2145 struct page new; 2146 struct page old; 2147 2148 c->partial = page->next; 2149 2150 n2 = get_node(s, page_to_nid(page)); 2151 if (n != n2) { 2152 if (n) 2153 spin_unlock(&n->list_lock); 2154 2155 n = n2; 2156 spin_lock(&n->list_lock); 2157 } 2158 2159 do { 2160 2161 old.freelist = page->freelist; 2162 old.counters = page->counters; 2163 VM_BUG_ON(!old.frozen); 2164 2165 new.counters = old.counters; 2166 new.freelist = old.freelist; 2167 2168 new.frozen = 0; 2169 2170 } while (!__cmpxchg_double_slab(s, page, 2171 old.freelist, old.counters, 2172 new.freelist, new.counters, 2173 "unfreezing slab")); 2174 2175 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { 2176 page->next = discard_page; 2177 discard_page = page; 2178 } else { 2179 add_partial(n, page, DEACTIVATE_TO_TAIL); 2180 stat(s, FREE_ADD_PARTIAL); 2181 } 2182 } 2183 2184 if (n) 2185 spin_unlock(&n->list_lock); 2186 2187 while (discard_page) { 2188 page = discard_page; 2189 discard_page = discard_page->next; 2190 2191 stat(s, DEACTIVATE_EMPTY); 2192 discard_slab(s, page); 2193 stat(s, FREE_SLAB); 2194 } 2195 #endif 2196 } 2197 2198 /* 2199 * Put a page that was just frozen (in __slab_free) into a partial page 2200 * slot if available. This is done without interrupts disabled and without 2201 * preemption disabled. The cmpxchg is racy and may put the partial page 2202 * onto a random cpus partial slot. 2203 * 2204 * If we did not find a slot then simply move all the partials to the 2205 * per node partial list. 2206 */ 2207 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) 2208 { 2209 #ifdef CONFIG_SLUB_CPU_PARTIAL 2210 struct page *oldpage; 2211 int pages; 2212 int pobjects; 2213 2214 preempt_disable(); 2215 do { 2216 pages = 0; 2217 pobjects = 0; 2218 oldpage = this_cpu_read(s->cpu_slab->partial); 2219 2220 if (oldpage) { 2221 pobjects = oldpage->pobjects; 2222 pages = oldpage->pages; 2223 if (drain && pobjects > s->cpu_partial) { 2224 unsigned long flags; 2225 /* 2226 * partial array is full. Move the existing 2227 * set to the per node partial list. 2228 */ 2229 local_irq_save(flags); 2230 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2231 local_irq_restore(flags); 2232 oldpage = NULL; 2233 pobjects = 0; 2234 pages = 0; 2235 stat(s, CPU_PARTIAL_DRAIN); 2236 } 2237 } 2238 2239 pages++; 2240 pobjects += page->objects - page->inuse; 2241 2242 page->pages = pages; 2243 page->pobjects = pobjects; 2244 page->next = oldpage; 2245 2246 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) 2247 != oldpage); 2248 if (unlikely(!s->cpu_partial)) { 2249 unsigned long flags; 2250 2251 local_irq_save(flags); 2252 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2253 local_irq_restore(flags); 2254 } 2255 preempt_enable(); 2256 #endif 2257 } 2258 2259 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 2260 { 2261 stat(s, CPUSLAB_FLUSH); 2262 deactivate_slab(s, c->page, c->freelist); 2263 2264 c->tid = next_tid(c->tid); 2265 c->page = NULL; 2266 c->freelist = NULL; 2267 } 2268 2269 /* 2270 * Flush cpu slab. 2271 * 2272 * Called from IPI handler with interrupts disabled. 2273 */ 2274 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 2275 { 2276 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2277 2278 if (likely(c)) { 2279 if (c->page) 2280 flush_slab(s, c); 2281 2282 unfreeze_partials(s, c); 2283 } 2284 } 2285 2286 static void flush_cpu_slab(void *d) 2287 { 2288 struct kmem_cache *s = d; 2289 2290 __flush_cpu_slab(s, smp_processor_id()); 2291 } 2292 2293 static bool has_cpu_slab(int cpu, void *info) 2294 { 2295 struct kmem_cache *s = info; 2296 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2297 2298 return c->page || c->partial; 2299 } 2300 2301 static void flush_all(struct kmem_cache *s) 2302 { 2303 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC); 2304 } 2305 2306 /* 2307 * Use the cpu notifier to insure that the cpu slabs are flushed when 2308 * necessary. 2309 */ 2310 static int slub_cpu_dead(unsigned int cpu) 2311 { 2312 struct kmem_cache *s; 2313 unsigned long flags; 2314 2315 mutex_lock(&slab_mutex); 2316 list_for_each_entry(s, &slab_caches, list) { 2317 local_irq_save(flags); 2318 __flush_cpu_slab(s, cpu); 2319 local_irq_restore(flags); 2320 } 2321 mutex_unlock(&slab_mutex); 2322 return 0; 2323 } 2324 2325 /* 2326 * Check if the objects in a per cpu structure fit numa 2327 * locality expectations. 2328 */ 2329 static inline int node_match(struct page *page, int node) 2330 { 2331 #ifdef CONFIG_NUMA 2332 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node)) 2333 return 0; 2334 #endif 2335 return 1; 2336 } 2337 2338 #ifdef CONFIG_SLUB_DEBUG 2339 static int count_free(struct page *page) 2340 { 2341 return page->objects - page->inuse; 2342 } 2343 2344 static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 2345 { 2346 return atomic_long_read(&n->total_objects); 2347 } 2348 #endif /* CONFIG_SLUB_DEBUG */ 2349 2350 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS) 2351 static unsigned long count_partial(struct kmem_cache_node *n, 2352 int (*get_count)(struct page *)) 2353 { 2354 unsigned long flags; 2355 unsigned long x = 0; 2356 struct page *page; 2357 2358 spin_lock_irqsave(&n->list_lock, flags); 2359 list_for_each_entry(page, &n->partial, lru) 2360 x += get_count(page); 2361 spin_unlock_irqrestore(&n->list_lock, flags); 2362 return x; 2363 } 2364 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */ 2365 2366 static noinline void 2367 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 2368 { 2369 #ifdef CONFIG_SLUB_DEBUG 2370 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, 2371 DEFAULT_RATELIMIT_BURST); 2372 int node; 2373 struct kmem_cache_node *n; 2374 2375 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) 2376 return; 2377 2378 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", 2379 nid, gfpflags, &gfpflags); 2380 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n", 2381 s->name, s->object_size, s->size, oo_order(s->oo), 2382 oo_order(s->min)); 2383 2384 if (oo_order(s->min) > get_order(s->object_size)) 2385 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n", 2386 s->name); 2387 2388 for_each_kmem_cache_node(s, node, n) { 2389 unsigned long nr_slabs; 2390 unsigned long nr_objs; 2391 unsigned long nr_free; 2392 2393 nr_free = count_partial(n, count_free); 2394 nr_slabs = node_nr_slabs(n); 2395 nr_objs = node_nr_objs(n); 2396 2397 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n", 2398 node, nr_slabs, nr_objs, nr_free); 2399 } 2400 #endif 2401 } 2402 2403 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, 2404 int node, struct kmem_cache_cpu **pc) 2405 { 2406 void *freelist; 2407 struct kmem_cache_cpu *c = *pc; 2408 struct page *page; 2409 2410 freelist = get_partial(s, flags, node, c); 2411 2412 if (freelist) 2413 return freelist; 2414 2415 page = new_slab(s, flags, node); 2416 if (page) { 2417 c = raw_cpu_ptr(s->cpu_slab); 2418 if (c->page) 2419 flush_slab(s, c); 2420 2421 /* 2422 * No other reference to the page yet so we can 2423 * muck around with it freely without cmpxchg 2424 */ 2425 freelist = page->freelist; 2426 page->freelist = NULL; 2427 2428 stat(s, ALLOC_SLAB); 2429 c->page = page; 2430 *pc = c; 2431 } else 2432 freelist = NULL; 2433 2434 return freelist; 2435 } 2436 2437 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags) 2438 { 2439 if (unlikely(PageSlabPfmemalloc(page))) 2440 return gfp_pfmemalloc_allowed(gfpflags); 2441 2442 return true; 2443 } 2444 2445 /* 2446 * Check the page->freelist of a page and either transfer the freelist to the 2447 * per cpu freelist or deactivate the page. 2448 * 2449 * The page is still frozen if the return value is not NULL. 2450 * 2451 * If this function returns NULL then the page has been unfrozen. 2452 * 2453 * This function must be called with interrupt disabled. 2454 */ 2455 static inline void *get_freelist(struct kmem_cache *s, struct page *page) 2456 { 2457 struct page new; 2458 unsigned long counters; 2459 void *freelist; 2460 2461 do { 2462 freelist = page->freelist; 2463 counters = page->counters; 2464 2465 new.counters = counters; 2466 VM_BUG_ON(!new.frozen); 2467 2468 new.inuse = page->objects; 2469 new.frozen = freelist != NULL; 2470 2471 } while (!__cmpxchg_double_slab(s, page, 2472 freelist, counters, 2473 NULL, new.counters, 2474 "get_freelist")); 2475 2476 return freelist; 2477 } 2478 2479 /* 2480 * Slow path. The lockless freelist is empty or we need to perform 2481 * debugging duties. 2482 * 2483 * Processing is still very fast if new objects have been freed to the 2484 * regular freelist. In that case we simply take over the regular freelist 2485 * as the lockless freelist and zap the regular freelist. 2486 * 2487 * If that is not working then we fall back to the partial lists. We take the 2488 * first element of the freelist as the object to allocate now and move the 2489 * rest of the freelist to the lockless freelist. 2490 * 2491 * And if we were unable to get a new slab from the partial slab lists then 2492 * we need to allocate a new slab. This is the slowest path since it involves 2493 * a call to the page allocator and the setup of a new slab. 2494 * 2495 * Version of __slab_alloc to use when we know that interrupts are 2496 * already disabled (which is the case for bulk allocation). 2497 */ 2498 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2499 unsigned long addr, struct kmem_cache_cpu *c) 2500 { 2501 void *freelist; 2502 struct page *page; 2503 2504 page = c->page; 2505 if (!page) 2506 goto new_slab; 2507 redo: 2508 2509 if (unlikely(!node_match(page, node))) { 2510 int searchnode = node; 2511 2512 if (node != NUMA_NO_NODE && !node_present_pages(node)) 2513 searchnode = node_to_mem_node(node); 2514 2515 if (unlikely(!node_match(page, searchnode))) { 2516 stat(s, ALLOC_NODE_MISMATCH); 2517 deactivate_slab(s, page, c->freelist); 2518 c->page = NULL; 2519 c->freelist = NULL; 2520 goto new_slab; 2521 } 2522 } 2523 2524 /* 2525 * By rights, we should be searching for a slab page that was 2526 * PFMEMALLOC but right now, we are losing the pfmemalloc 2527 * information when the page leaves the per-cpu allocator 2528 */ 2529 if (unlikely(!pfmemalloc_match(page, gfpflags))) { 2530 deactivate_slab(s, page, c->freelist); 2531 c->page = NULL; 2532 c->freelist = NULL; 2533 goto new_slab; 2534 } 2535 2536 /* must check again c->freelist in case of cpu migration or IRQ */ 2537 freelist = c->freelist; 2538 if (freelist) 2539 goto load_freelist; 2540 2541 freelist = get_freelist(s, page); 2542 2543 if (!freelist) { 2544 c->page = NULL; 2545 stat(s, DEACTIVATE_BYPASS); 2546 goto new_slab; 2547 } 2548 2549 stat(s, ALLOC_REFILL); 2550 2551 load_freelist: 2552 /* 2553 * freelist is pointing to the list of objects to be used. 2554 * page is pointing to the page from which the objects are obtained. 2555 * That page must be frozen for per cpu allocations to work. 2556 */ 2557 VM_BUG_ON(!c->page->frozen); 2558 c->freelist = get_freepointer(s, freelist); 2559 c->tid = next_tid(c->tid); 2560 return freelist; 2561 2562 new_slab: 2563 2564 if (c->partial) { 2565 page = c->page = c->partial; 2566 c->partial = page->next; 2567 stat(s, CPU_PARTIAL_ALLOC); 2568 c->freelist = NULL; 2569 goto redo; 2570 } 2571 2572 freelist = new_slab_objects(s, gfpflags, node, &c); 2573 2574 if (unlikely(!freelist)) { 2575 slab_out_of_memory(s, gfpflags, node); 2576 return NULL; 2577 } 2578 2579 page = c->page; 2580 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags))) 2581 goto load_freelist; 2582 2583 /* Only entered in the debug case */ 2584 if (kmem_cache_debug(s) && 2585 !alloc_debug_processing(s, page, freelist, addr)) 2586 goto new_slab; /* Slab failed checks. Next slab needed */ 2587 2588 deactivate_slab(s, page, get_freepointer(s, freelist)); 2589 c->page = NULL; 2590 c->freelist = NULL; 2591 return freelist; 2592 } 2593 2594 /* 2595 * Another one that disabled interrupt and compensates for possible 2596 * cpu changes by refetching the per cpu area pointer. 2597 */ 2598 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2599 unsigned long addr, struct kmem_cache_cpu *c) 2600 { 2601 void *p; 2602 unsigned long flags; 2603 2604 local_irq_save(flags); 2605 #ifdef CONFIG_PREEMPT 2606 /* 2607 * We may have been preempted and rescheduled on a different 2608 * cpu before disabling interrupts. Need to reload cpu area 2609 * pointer. 2610 */ 2611 c = this_cpu_ptr(s->cpu_slab); 2612 #endif 2613 2614 p = ___slab_alloc(s, gfpflags, node, addr, c); 2615 local_irq_restore(flags); 2616 return p; 2617 } 2618 2619 /* 2620 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 2621 * have the fastpath folded into their functions. So no function call 2622 * overhead for requests that can be satisfied on the fastpath. 2623 * 2624 * The fastpath works by first checking if the lockless freelist can be used. 2625 * If not then __slab_alloc is called for slow processing. 2626 * 2627 * Otherwise we can simply pick the next object from the lockless free list. 2628 */ 2629 static __always_inline void *slab_alloc_node(struct kmem_cache *s, 2630 gfp_t gfpflags, int node, unsigned long addr) 2631 { 2632 void *object; 2633 struct kmem_cache_cpu *c; 2634 struct page *page; 2635 unsigned long tid; 2636 2637 s = slab_pre_alloc_hook(s, gfpflags); 2638 if (!s) 2639 return NULL; 2640 redo: 2641 /* 2642 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 2643 * enabled. We may switch back and forth between cpus while 2644 * reading from one cpu area. That does not matter as long 2645 * as we end up on the original cpu again when doing the cmpxchg. 2646 * 2647 * We should guarantee that tid and kmem_cache are retrieved on 2648 * the same cpu. It could be different if CONFIG_PREEMPT so we need 2649 * to check if it is matched or not. 2650 */ 2651 do { 2652 tid = this_cpu_read(s->cpu_slab->tid); 2653 c = raw_cpu_ptr(s->cpu_slab); 2654 } while (IS_ENABLED(CONFIG_PREEMPT) && 2655 unlikely(tid != READ_ONCE(c->tid))); 2656 2657 /* 2658 * Irqless object alloc/free algorithm used here depends on sequence 2659 * of fetching cpu_slab's data. tid should be fetched before anything 2660 * on c to guarantee that object and page associated with previous tid 2661 * won't be used with current tid. If we fetch tid first, object and 2662 * page could be one associated with next tid and our alloc/free 2663 * request will be failed. In this case, we will retry. So, no problem. 2664 */ 2665 barrier(); 2666 2667 /* 2668 * The transaction ids are globally unique per cpu and per operation on 2669 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 2670 * occurs on the right processor and that there was no operation on the 2671 * linked list in between. 2672 */ 2673 2674 object = c->freelist; 2675 page = c->page; 2676 if (unlikely(!object || !node_match(page, node))) { 2677 object = __slab_alloc(s, gfpflags, node, addr, c); 2678 stat(s, ALLOC_SLOWPATH); 2679 } else { 2680 void *next_object = get_freepointer_safe(s, object); 2681 2682 /* 2683 * The cmpxchg will only match if there was no additional 2684 * operation and if we are on the right processor. 2685 * 2686 * The cmpxchg does the following atomically (without lock 2687 * semantics!) 2688 * 1. Relocate first pointer to the current per cpu area. 2689 * 2. Verify that tid and freelist have not been changed 2690 * 3. If they were not changed replace tid and freelist 2691 * 2692 * Since this is without lock semantics the protection is only 2693 * against code executing on this cpu *not* from access by 2694 * other cpus. 2695 */ 2696 if (unlikely(!this_cpu_cmpxchg_double( 2697 s->cpu_slab->freelist, s->cpu_slab->tid, 2698 object, tid, 2699 next_object, next_tid(tid)))) { 2700 2701 note_cmpxchg_failure("slab_alloc", s, tid); 2702 goto redo; 2703 } 2704 prefetch_freepointer(s, next_object); 2705 stat(s, ALLOC_FASTPATH); 2706 } 2707 2708 if (unlikely(gfpflags & __GFP_ZERO) && object) 2709 memset(object, 0, s->object_size); 2710 2711 slab_post_alloc_hook(s, gfpflags, 1, &object); 2712 2713 return object; 2714 } 2715 2716 static __always_inline void *slab_alloc(struct kmem_cache *s, 2717 gfp_t gfpflags, unsigned long addr) 2718 { 2719 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr); 2720 } 2721 2722 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 2723 { 2724 void *ret = slab_alloc(s, gfpflags, _RET_IP_); 2725 2726 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, 2727 s->size, gfpflags); 2728 2729 return ret; 2730 } 2731 EXPORT_SYMBOL(kmem_cache_alloc); 2732 2733 #ifdef CONFIG_TRACING 2734 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2735 { 2736 void *ret = slab_alloc(s, gfpflags, _RET_IP_); 2737 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2738 kasan_kmalloc(s, ret, size, gfpflags); 2739 return ret; 2740 } 2741 EXPORT_SYMBOL(kmem_cache_alloc_trace); 2742 #endif 2743 2744 #ifdef CONFIG_NUMA 2745 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2746 { 2747 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); 2748 2749 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2750 s->object_size, s->size, gfpflags, node); 2751 2752 return ret; 2753 } 2754 EXPORT_SYMBOL(kmem_cache_alloc_node); 2755 2756 #ifdef CONFIG_TRACING 2757 void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2758 gfp_t gfpflags, 2759 int node, size_t size) 2760 { 2761 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); 2762 2763 trace_kmalloc_node(_RET_IP_, ret, 2764 size, s->size, gfpflags, node); 2765 2766 kasan_kmalloc(s, ret, size, gfpflags); 2767 return ret; 2768 } 2769 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2770 #endif 2771 #endif 2772 2773 /* 2774 * Slow path handling. This may still be called frequently since objects 2775 * have a longer lifetime than the cpu slabs in most processing loads. 2776 * 2777 * So we still attempt to reduce cache line usage. Just take the slab 2778 * lock and free the item. If there is no additional partial page 2779 * handling required then we can return immediately. 2780 */ 2781 static void __slab_free(struct kmem_cache *s, struct page *page, 2782 void *head, void *tail, int cnt, 2783 unsigned long addr) 2784 2785 { 2786 void *prior; 2787 int was_frozen; 2788 struct page new; 2789 unsigned long counters; 2790 struct kmem_cache_node *n = NULL; 2791 unsigned long uninitialized_var(flags); 2792 2793 stat(s, FREE_SLOWPATH); 2794 2795 if (kmem_cache_debug(s) && 2796 !free_debug_processing(s, page, head, tail, cnt, addr)) 2797 return; 2798 2799 do { 2800 if (unlikely(n)) { 2801 spin_unlock_irqrestore(&n->list_lock, flags); 2802 n = NULL; 2803 } 2804 prior = page->freelist; 2805 counters = page->counters; 2806 set_freepointer(s, tail, prior); 2807 new.counters = counters; 2808 was_frozen = new.frozen; 2809 new.inuse -= cnt; 2810 if ((!new.inuse || !prior) && !was_frozen) { 2811 2812 if (kmem_cache_has_cpu_partial(s) && !prior) { 2813 2814 /* 2815 * Slab was on no list before and will be 2816 * partially empty 2817 * We can defer the list move and instead 2818 * freeze it. 2819 */ 2820 new.frozen = 1; 2821 2822 } else { /* Needs to be taken off a list */ 2823 2824 n = get_node(s, page_to_nid(page)); 2825 /* 2826 * Speculatively acquire the list_lock. 2827 * If the cmpxchg does not succeed then we may 2828 * drop the list_lock without any processing. 2829 * 2830 * Otherwise the list_lock will synchronize with 2831 * other processors updating the list of slabs. 2832 */ 2833 spin_lock_irqsave(&n->list_lock, flags); 2834 2835 } 2836 } 2837 2838 } while (!cmpxchg_double_slab(s, page, 2839 prior, counters, 2840 head, new.counters, 2841 "__slab_free")); 2842 2843 if (likely(!n)) { 2844 2845 /* 2846 * If we just froze the page then put it onto the 2847 * per cpu partial list. 2848 */ 2849 if (new.frozen && !was_frozen) { 2850 put_cpu_partial(s, page, 1); 2851 stat(s, CPU_PARTIAL_FREE); 2852 } 2853 /* 2854 * The list lock was not taken therefore no list 2855 * activity can be necessary. 2856 */ 2857 if (was_frozen) 2858 stat(s, FREE_FROZEN); 2859 return; 2860 } 2861 2862 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) 2863 goto slab_empty; 2864 2865 /* 2866 * Objects left in the slab. If it was not on the partial list before 2867 * then add it. 2868 */ 2869 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { 2870 if (kmem_cache_debug(s)) 2871 remove_full(s, n, page); 2872 add_partial(n, page, DEACTIVATE_TO_TAIL); 2873 stat(s, FREE_ADD_PARTIAL); 2874 } 2875 spin_unlock_irqrestore(&n->list_lock, flags); 2876 return; 2877 2878 slab_empty: 2879 if (prior) { 2880 /* 2881 * Slab on the partial list. 2882 */ 2883 remove_partial(n, page); 2884 stat(s, FREE_REMOVE_PARTIAL); 2885 } else { 2886 /* Slab must be on the full list */ 2887 remove_full(s, n, page); 2888 } 2889 2890 spin_unlock_irqrestore(&n->list_lock, flags); 2891 stat(s, FREE_SLAB); 2892 discard_slab(s, page); 2893 } 2894 2895 /* 2896 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 2897 * can perform fastpath freeing without additional function calls. 2898 * 2899 * The fastpath is only possible if we are freeing to the current cpu slab 2900 * of this processor. This typically the case if we have just allocated 2901 * the item before. 2902 * 2903 * If fastpath is not possible then fall back to __slab_free where we deal 2904 * with all sorts of special processing. 2905 * 2906 * Bulk free of a freelist with several objects (all pointing to the 2907 * same page) possible by specifying head and tail ptr, plus objects 2908 * count (cnt). Bulk free indicated by tail pointer being set. 2909 */ 2910 static __always_inline void do_slab_free(struct kmem_cache *s, 2911 struct page *page, void *head, void *tail, 2912 int cnt, unsigned long addr) 2913 { 2914 void *tail_obj = tail ? : head; 2915 struct kmem_cache_cpu *c; 2916 unsigned long tid; 2917 redo: 2918 /* 2919 * Determine the currently cpus per cpu slab. 2920 * The cpu may change afterward. However that does not matter since 2921 * data is retrieved via this pointer. If we are on the same cpu 2922 * during the cmpxchg then the free will succeed. 2923 */ 2924 do { 2925 tid = this_cpu_read(s->cpu_slab->tid); 2926 c = raw_cpu_ptr(s->cpu_slab); 2927 } while (IS_ENABLED(CONFIG_PREEMPT) && 2928 unlikely(tid != READ_ONCE(c->tid))); 2929 2930 /* Same with comment on barrier() in slab_alloc_node() */ 2931 barrier(); 2932 2933 if (likely(page == c->page)) { 2934 set_freepointer(s, tail_obj, c->freelist); 2935 2936 if (unlikely(!this_cpu_cmpxchg_double( 2937 s->cpu_slab->freelist, s->cpu_slab->tid, 2938 c->freelist, tid, 2939 head, next_tid(tid)))) { 2940 2941 note_cmpxchg_failure("slab_free", s, tid); 2942 goto redo; 2943 } 2944 stat(s, FREE_FASTPATH); 2945 } else 2946 __slab_free(s, page, head, tail_obj, cnt, addr); 2947 2948 } 2949 2950 static __always_inline void slab_free(struct kmem_cache *s, struct page *page, 2951 void *head, void *tail, int cnt, 2952 unsigned long addr) 2953 { 2954 slab_free_freelist_hook(s, head, tail); 2955 /* 2956 * slab_free_freelist_hook() could have put the items into quarantine. 2957 * If so, no need to free them. 2958 */ 2959 if (s->flags & SLAB_KASAN && !(s->flags & SLAB_DESTROY_BY_RCU)) 2960 return; 2961 do_slab_free(s, page, head, tail, cnt, addr); 2962 } 2963 2964 #ifdef CONFIG_KASAN 2965 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) 2966 { 2967 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr); 2968 } 2969 #endif 2970 2971 void kmem_cache_free(struct kmem_cache *s, void *x) 2972 { 2973 s = cache_from_obj(s, x); 2974 if (!s) 2975 return; 2976 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_); 2977 trace_kmem_cache_free(_RET_IP_, x); 2978 } 2979 EXPORT_SYMBOL(kmem_cache_free); 2980 2981 struct detached_freelist { 2982 struct page *page; 2983 void *tail; 2984 void *freelist; 2985 int cnt; 2986 struct kmem_cache *s; 2987 }; 2988 2989 /* 2990 * This function progressively scans the array with free objects (with 2991 * a limited look ahead) and extract objects belonging to the same 2992 * page. It builds a detached freelist directly within the given 2993 * page/objects. This can happen without any need for 2994 * synchronization, because the objects are owned by running process. 2995 * The freelist is build up as a single linked list in the objects. 2996 * The idea is, that this detached freelist can then be bulk 2997 * transferred to the real freelist(s), but only requiring a single 2998 * synchronization primitive. Look ahead in the array is limited due 2999 * to performance reasons. 3000 */ 3001 static inline 3002 int build_detached_freelist(struct kmem_cache *s, size_t size, 3003 void **p, struct detached_freelist *df) 3004 { 3005 size_t first_skipped_index = 0; 3006 int lookahead = 3; 3007 void *object; 3008 struct page *page; 3009 3010 /* Always re-init detached_freelist */ 3011 df->page = NULL; 3012 3013 do { 3014 object = p[--size]; 3015 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */ 3016 } while (!object && size); 3017 3018 if (!object) 3019 return 0; 3020 3021 page = virt_to_head_page(object); 3022 if (!s) { 3023 /* Handle kalloc'ed objects */ 3024 if (unlikely(!PageSlab(page))) { 3025 BUG_ON(!PageCompound(page)); 3026 kfree_hook(object); 3027 __free_pages(page, compound_order(page)); 3028 p[size] = NULL; /* mark object processed */ 3029 return size; 3030 } 3031 /* Derive kmem_cache from object */ 3032 df->s = page->slab_cache; 3033 } else { 3034 df->s = cache_from_obj(s, object); /* Support for memcg */ 3035 } 3036 3037 /* Start new detached freelist */ 3038 df->page = page; 3039 set_freepointer(df->s, object, NULL); 3040 df->tail = object; 3041 df->freelist = object; 3042 p[size] = NULL; /* mark object processed */ 3043 df->cnt = 1; 3044 3045 while (size) { 3046 object = p[--size]; 3047 if (!object) 3048 continue; /* Skip processed objects */ 3049 3050 /* df->page is always set at this point */ 3051 if (df->page == virt_to_head_page(object)) { 3052 /* Opportunity build freelist */ 3053 set_freepointer(df->s, object, df->freelist); 3054 df->freelist = object; 3055 df->cnt++; 3056 p[size] = NULL; /* mark object processed */ 3057 3058 continue; 3059 } 3060 3061 /* Limit look ahead search */ 3062 if (!--lookahead) 3063 break; 3064 3065 if (!first_skipped_index) 3066 first_skipped_index = size + 1; 3067 } 3068 3069 return first_skipped_index; 3070 } 3071 3072 /* Note that interrupts must be enabled when calling this function. */ 3073 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) 3074 { 3075 if (WARN_ON(!size)) 3076 return; 3077 3078 do { 3079 struct detached_freelist df; 3080 3081 size = build_detached_freelist(s, size, p, &df); 3082 if (!df.page) 3083 continue; 3084 3085 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_); 3086 } while (likely(size)); 3087 } 3088 EXPORT_SYMBOL(kmem_cache_free_bulk); 3089 3090 /* Note that interrupts must be enabled when calling this function. */ 3091 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, 3092 void **p) 3093 { 3094 struct kmem_cache_cpu *c; 3095 int i; 3096 3097 /* memcg and kmem_cache debug support */ 3098 s = slab_pre_alloc_hook(s, flags); 3099 if (unlikely(!s)) 3100 return false; 3101 /* 3102 * Drain objects in the per cpu slab, while disabling local 3103 * IRQs, which protects against PREEMPT and interrupts 3104 * handlers invoking normal fastpath. 3105 */ 3106 local_irq_disable(); 3107 c = this_cpu_ptr(s->cpu_slab); 3108 3109 for (i = 0; i < size; i++) { 3110 void *object = c->freelist; 3111 3112 if (unlikely(!object)) { 3113 /* 3114 * Invoking slow path likely have side-effect 3115 * of re-populating per CPU c->freelist 3116 */ 3117 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, 3118 _RET_IP_, c); 3119 if (unlikely(!p[i])) 3120 goto error; 3121 3122 c = this_cpu_ptr(s->cpu_slab); 3123 continue; /* goto for-loop */ 3124 } 3125 c->freelist = get_freepointer(s, object); 3126 p[i] = object; 3127 } 3128 c->tid = next_tid(c->tid); 3129 local_irq_enable(); 3130 3131 /* Clear memory outside IRQ disabled fastpath loop */ 3132 if (unlikely(flags & __GFP_ZERO)) { 3133 int j; 3134 3135 for (j = 0; j < i; j++) 3136 memset(p[j], 0, s->object_size); 3137 } 3138 3139 /* memcg and kmem_cache debug support */ 3140 slab_post_alloc_hook(s, flags, size, p); 3141 return i; 3142 error: 3143 local_irq_enable(); 3144 slab_post_alloc_hook(s, flags, i, p); 3145 __kmem_cache_free_bulk(s, i, p); 3146 return 0; 3147 } 3148 EXPORT_SYMBOL(kmem_cache_alloc_bulk); 3149 3150 3151 /* 3152 * Object placement in a slab is made very easy because we always start at 3153 * offset 0. If we tune the size of the object to the alignment then we can 3154 * get the required alignment by putting one properly sized object after 3155 * another. 3156 * 3157 * Notice that the allocation order determines the sizes of the per cpu 3158 * caches. Each processor has always one slab available for allocations. 3159 * Increasing the allocation order reduces the number of times that slabs 3160 * must be moved on and off the partial lists and is therefore a factor in 3161 * locking overhead. 3162 */ 3163 3164 /* 3165 * Mininum / Maximum order of slab pages. This influences locking overhead 3166 * and slab fragmentation. A higher order reduces the number of partial slabs 3167 * and increases the number of allocations possible without having to 3168 * take the list_lock. 3169 */ 3170 static int slub_min_order; 3171 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 3172 static int slub_min_objects; 3173 3174 /* 3175 * Calculate the order of allocation given an slab object size. 3176 * 3177 * The order of allocation has significant impact on performance and other 3178 * system components. Generally order 0 allocations should be preferred since 3179 * order 0 does not cause fragmentation in the page allocator. Larger objects 3180 * be problematic to put into order 0 slabs because there may be too much 3181 * unused space left. We go to a higher order if more than 1/16th of the slab 3182 * would be wasted. 3183 * 3184 * In order to reach satisfactory performance we must ensure that a minimum 3185 * number of objects is in one slab. Otherwise we may generate too much 3186 * activity on the partial lists which requires taking the list_lock. This is 3187 * less a concern for large slabs though which are rarely used. 3188 * 3189 * slub_max_order specifies the order where we begin to stop considering the 3190 * number of objects in a slab as critical. If we reach slub_max_order then 3191 * we try to keep the page order as low as possible. So we accept more waste 3192 * of space in favor of a small page order. 3193 * 3194 * Higher order allocations also allow the placement of more objects in a 3195 * slab and thereby reduce object handling overhead. If the user has 3196 * requested a higher mininum order then we start with that one instead of 3197 * the smallest order which will fit the object. 3198 */ 3199 static inline int slab_order(int size, int min_objects, 3200 int max_order, int fract_leftover, int reserved) 3201 { 3202 int order; 3203 int rem; 3204 int min_order = slub_min_order; 3205 3206 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE) 3207 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 3208 3209 for (order = max(min_order, get_order(min_objects * size + reserved)); 3210 order <= max_order; order++) { 3211 3212 unsigned long slab_size = PAGE_SIZE << order; 3213 3214 rem = (slab_size - reserved) % size; 3215 3216 if (rem <= slab_size / fract_leftover) 3217 break; 3218 } 3219 3220 return order; 3221 } 3222 3223 static inline int calculate_order(int size, int reserved) 3224 { 3225 int order; 3226 int min_objects; 3227 int fraction; 3228 int max_objects; 3229 3230 /* 3231 * Attempt to find best configuration for a slab. This 3232 * works by first attempting to generate a layout with 3233 * the best configuration and backing off gradually. 3234 * 3235 * First we increase the acceptable waste in a slab. Then 3236 * we reduce the minimum objects required in a slab. 3237 */ 3238 min_objects = slub_min_objects; 3239 if (!min_objects) 3240 min_objects = 4 * (fls(nr_cpu_ids) + 1); 3241 max_objects = order_objects(slub_max_order, size, reserved); 3242 min_objects = min(min_objects, max_objects); 3243 3244 while (min_objects > 1) { 3245 fraction = 16; 3246 while (fraction >= 4) { 3247 order = slab_order(size, min_objects, 3248 slub_max_order, fraction, reserved); 3249 if (order <= slub_max_order) 3250 return order; 3251 fraction /= 2; 3252 } 3253 min_objects--; 3254 } 3255 3256 /* 3257 * We were unable to place multiple objects in a slab. Now 3258 * lets see if we can place a single object there. 3259 */ 3260 order = slab_order(size, 1, slub_max_order, 1, reserved); 3261 if (order <= slub_max_order) 3262 return order; 3263 3264 /* 3265 * Doh this slab cannot be placed using slub_max_order. 3266 */ 3267 order = slab_order(size, 1, MAX_ORDER, 1, reserved); 3268 if (order < MAX_ORDER) 3269 return order; 3270 return -ENOSYS; 3271 } 3272 3273 static void 3274 init_kmem_cache_node(struct kmem_cache_node *n) 3275 { 3276 n->nr_partial = 0; 3277 spin_lock_init(&n->list_lock); 3278 INIT_LIST_HEAD(&n->partial); 3279 #ifdef CONFIG_SLUB_DEBUG 3280 atomic_long_set(&n->nr_slabs, 0); 3281 atomic_long_set(&n->total_objects, 0); 3282 INIT_LIST_HEAD(&n->full); 3283 #endif 3284 } 3285 3286 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 3287 { 3288 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 3289 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu)); 3290 3291 /* 3292 * Must align to double word boundary for the double cmpxchg 3293 * instructions to work; see __pcpu_double_call_return_bool(). 3294 */ 3295 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 3296 2 * sizeof(void *)); 3297 3298 if (!s->cpu_slab) 3299 return 0; 3300 3301 init_kmem_cache_cpus(s); 3302 3303 return 1; 3304 } 3305 3306 static struct kmem_cache *kmem_cache_node; 3307 3308 /* 3309 * No kmalloc_node yet so do it by hand. We know that this is the first 3310 * slab on the node for this slabcache. There are no concurrent accesses 3311 * possible. 3312 * 3313 * Note that this function only works on the kmem_cache_node 3314 * when allocating for the kmem_cache_node. This is used for bootstrapping 3315 * memory on a fresh node that has no slab structures yet. 3316 */ 3317 static void early_kmem_cache_node_alloc(int node) 3318 { 3319 struct page *page; 3320 struct kmem_cache_node *n; 3321 3322 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 3323 3324 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 3325 3326 BUG_ON(!page); 3327 if (page_to_nid(page) != node) { 3328 pr_err("SLUB: Unable to allocate memory from node %d\n", node); 3329 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n"); 3330 } 3331 3332 n = page->freelist; 3333 BUG_ON(!n); 3334 page->freelist = get_freepointer(kmem_cache_node, n); 3335 page->inuse = 1; 3336 page->frozen = 0; 3337 kmem_cache_node->node[node] = n; 3338 #ifdef CONFIG_SLUB_DEBUG 3339 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 3340 init_tracking(kmem_cache_node, n); 3341 #endif 3342 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node), 3343 GFP_KERNEL); 3344 init_kmem_cache_node(n); 3345 inc_slabs_node(kmem_cache_node, node, page->objects); 3346 3347 /* 3348 * No locks need to be taken here as it has just been 3349 * initialized and there is no concurrent access. 3350 */ 3351 __add_partial(n, page, DEACTIVATE_TO_HEAD); 3352 } 3353 3354 static void free_kmem_cache_nodes(struct kmem_cache *s) 3355 { 3356 int node; 3357 struct kmem_cache_node *n; 3358 3359 for_each_kmem_cache_node(s, node, n) { 3360 kmem_cache_free(kmem_cache_node, n); 3361 s->node[node] = NULL; 3362 } 3363 } 3364 3365 void __kmem_cache_release(struct kmem_cache *s) 3366 { 3367 cache_random_seq_destroy(s); 3368 free_percpu(s->cpu_slab); 3369 free_kmem_cache_nodes(s); 3370 } 3371 3372 static int init_kmem_cache_nodes(struct kmem_cache *s) 3373 { 3374 int node; 3375 3376 for_each_node_state(node, N_NORMAL_MEMORY) { 3377 struct kmem_cache_node *n; 3378 3379 if (slab_state == DOWN) { 3380 early_kmem_cache_node_alloc(node); 3381 continue; 3382 } 3383 n = kmem_cache_alloc_node(kmem_cache_node, 3384 GFP_KERNEL, node); 3385 3386 if (!n) { 3387 free_kmem_cache_nodes(s); 3388 return 0; 3389 } 3390 3391 s->node[node] = n; 3392 init_kmem_cache_node(n); 3393 } 3394 return 1; 3395 } 3396 3397 static void set_min_partial(struct kmem_cache *s, unsigned long min) 3398 { 3399 if (min < MIN_PARTIAL) 3400 min = MIN_PARTIAL; 3401 else if (min > MAX_PARTIAL) 3402 min = MAX_PARTIAL; 3403 s->min_partial = min; 3404 } 3405 3406 /* 3407 * calculate_sizes() determines the order and the distribution of data within 3408 * a slab object. 3409 */ 3410 static int calculate_sizes(struct kmem_cache *s, int forced_order) 3411 { 3412 unsigned long flags = s->flags; 3413 size_t size = s->object_size; 3414 int order; 3415 3416 /* 3417 * Round up object size to the next word boundary. We can only 3418 * place the free pointer at word boundaries and this determines 3419 * the possible location of the free pointer. 3420 */ 3421 size = ALIGN(size, sizeof(void *)); 3422 3423 #ifdef CONFIG_SLUB_DEBUG 3424 /* 3425 * Determine if we can poison the object itself. If the user of 3426 * the slab may touch the object after free or before allocation 3427 * then we should never poison the object itself. 3428 */ 3429 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) && 3430 !s->ctor) 3431 s->flags |= __OBJECT_POISON; 3432 else 3433 s->flags &= ~__OBJECT_POISON; 3434 3435 3436 /* 3437 * If we are Redzoning then check if there is some space between the 3438 * end of the object and the free pointer. If not then add an 3439 * additional word to have some bytes to store Redzone information. 3440 */ 3441 if ((flags & SLAB_RED_ZONE) && size == s->object_size) 3442 size += sizeof(void *); 3443 #endif 3444 3445 /* 3446 * With that we have determined the number of bytes in actual use 3447 * by the object. This is the potential offset to the free pointer. 3448 */ 3449 s->inuse = size; 3450 3451 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) || 3452 s->ctor)) { 3453 /* 3454 * Relocate free pointer after the object if it is not 3455 * permitted to overwrite the first word of the object on 3456 * kmem_cache_free. 3457 * 3458 * This is the case if we do RCU, have a constructor or 3459 * destructor or are poisoning the objects. 3460 */ 3461 s->offset = size; 3462 size += sizeof(void *); 3463 } 3464 3465 #ifdef CONFIG_SLUB_DEBUG 3466 if (flags & SLAB_STORE_USER) 3467 /* 3468 * Need to store information about allocs and frees after 3469 * the object. 3470 */ 3471 size += 2 * sizeof(struct track); 3472 #endif 3473 3474 kasan_cache_create(s, &size, &s->flags); 3475 #ifdef CONFIG_SLUB_DEBUG 3476 if (flags & SLAB_RED_ZONE) { 3477 /* 3478 * Add some empty padding so that we can catch 3479 * overwrites from earlier objects rather than let 3480 * tracking information or the free pointer be 3481 * corrupted if a user writes before the start 3482 * of the object. 3483 */ 3484 size += sizeof(void *); 3485 3486 s->red_left_pad = sizeof(void *); 3487 s->red_left_pad = ALIGN(s->red_left_pad, s->align); 3488 size += s->red_left_pad; 3489 } 3490 #endif 3491 3492 /* 3493 * SLUB stores one object immediately after another beginning from 3494 * offset 0. In order to align the objects we have to simply size 3495 * each object to conform to the alignment. 3496 */ 3497 size = ALIGN(size, s->align); 3498 s->size = size; 3499 if (forced_order >= 0) 3500 order = forced_order; 3501 else 3502 order = calculate_order(size, s->reserved); 3503 3504 if (order < 0) 3505 return 0; 3506 3507 s->allocflags = 0; 3508 if (order) 3509 s->allocflags |= __GFP_COMP; 3510 3511 if (s->flags & SLAB_CACHE_DMA) 3512 s->allocflags |= GFP_DMA; 3513 3514 if (s->flags & SLAB_RECLAIM_ACCOUNT) 3515 s->allocflags |= __GFP_RECLAIMABLE; 3516 3517 /* 3518 * Determine the number of objects per slab 3519 */ 3520 s->oo = oo_make(order, size, s->reserved); 3521 s->min = oo_make(get_order(size), size, s->reserved); 3522 if (oo_objects(s->oo) > oo_objects(s->max)) 3523 s->max = s->oo; 3524 3525 return !!oo_objects(s->oo); 3526 } 3527 3528 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags) 3529 { 3530 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor); 3531 s->reserved = 0; 3532 3533 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU)) 3534 s->reserved = sizeof(struct rcu_head); 3535 3536 if (!calculate_sizes(s, -1)) 3537 goto error; 3538 if (disable_higher_order_debug) { 3539 /* 3540 * Disable debugging flags that store metadata if the min slab 3541 * order increased. 3542 */ 3543 if (get_order(s->size) > get_order(s->object_size)) { 3544 s->flags &= ~DEBUG_METADATA_FLAGS; 3545 s->offset = 0; 3546 if (!calculate_sizes(s, -1)) 3547 goto error; 3548 } 3549 } 3550 3551 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 3552 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 3553 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0) 3554 /* Enable fast mode */ 3555 s->flags |= __CMPXCHG_DOUBLE; 3556 #endif 3557 3558 /* 3559 * The larger the object size is, the more pages we want on the partial 3560 * list to avoid pounding the page allocator excessively. 3561 */ 3562 set_min_partial(s, ilog2(s->size) / 2); 3563 3564 /* 3565 * cpu_partial determined the maximum number of objects kept in the 3566 * per cpu partial lists of a processor. 3567 * 3568 * Per cpu partial lists mainly contain slabs that just have one 3569 * object freed. If they are used for allocation then they can be 3570 * filled up again with minimal effort. The slab will never hit the 3571 * per node partial lists and therefore no locking will be required. 3572 * 3573 * This setting also determines 3574 * 3575 * A) The number of objects from per cpu partial slabs dumped to the 3576 * per node list when we reach the limit. 3577 * B) The number of objects in cpu partial slabs to extract from the 3578 * per node list when we run out of per cpu objects. We only fetch 3579 * 50% to keep some capacity around for frees. 3580 */ 3581 if (!kmem_cache_has_cpu_partial(s)) 3582 s->cpu_partial = 0; 3583 else if (s->size >= PAGE_SIZE) 3584 s->cpu_partial = 2; 3585 else if (s->size >= 1024) 3586 s->cpu_partial = 6; 3587 else if (s->size >= 256) 3588 s->cpu_partial = 13; 3589 else 3590 s->cpu_partial = 30; 3591 3592 #ifdef CONFIG_NUMA 3593 s->remote_node_defrag_ratio = 1000; 3594 #endif 3595 3596 /* Initialize the pre-computed randomized freelist if slab is up */ 3597 if (slab_state >= UP) { 3598 if (init_cache_random_seq(s)) 3599 goto error; 3600 } 3601 3602 if (!init_kmem_cache_nodes(s)) 3603 goto error; 3604 3605 if (alloc_kmem_cache_cpus(s)) 3606 return 0; 3607 3608 free_kmem_cache_nodes(s); 3609 error: 3610 if (flags & SLAB_PANIC) 3611 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n", 3612 s->name, (unsigned long)s->size, s->size, 3613 oo_order(s->oo), s->offset, flags); 3614 return -EINVAL; 3615 } 3616 3617 static void list_slab_objects(struct kmem_cache *s, struct page *page, 3618 const char *text) 3619 { 3620 #ifdef CONFIG_SLUB_DEBUG 3621 void *addr = page_address(page); 3622 void *p; 3623 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) * 3624 sizeof(long), GFP_ATOMIC); 3625 if (!map) 3626 return; 3627 slab_err(s, page, text, s->name); 3628 slab_lock(page); 3629 3630 get_map(s, page, map); 3631 for_each_object(p, s, addr, page->objects) { 3632 3633 if (!test_bit(slab_index(p, s, addr), map)) { 3634 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr); 3635 print_tracking(s, p); 3636 } 3637 } 3638 slab_unlock(page); 3639 kfree(map); 3640 #endif 3641 } 3642 3643 /* 3644 * Attempt to free all partial slabs on a node. 3645 * This is called from __kmem_cache_shutdown(). We must take list_lock 3646 * because sysfs file might still access partial list after the shutdowning. 3647 */ 3648 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 3649 { 3650 LIST_HEAD(discard); 3651 struct page *page, *h; 3652 3653 BUG_ON(irqs_disabled()); 3654 spin_lock_irq(&n->list_lock); 3655 list_for_each_entry_safe(page, h, &n->partial, lru) { 3656 if (!page->inuse) { 3657 remove_partial(n, page); 3658 list_add(&page->lru, &discard); 3659 } else { 3660 list_slab_objects(s, page, 3661 "Objects remaining in %s on __kmem_cache_shutdown()"); 3662 } 3663 } 3664 spin_unlock_irq(&n->list_lock); 3665 3666 list_for_each_entry_safe(page, h, &discard, lru) 3667 discard_slab(s, page); 3668 } 3669 3670 /* 3671 * Release all resources used by a slab cache. 3672 */ 3673 int __kmem_cache_shutdown(struct kmem_cache *s) 3674 { 3675 int node; 3676 struct kmem_cache_node *n; 3677 3678 flush_all(s); 3679 /* Attempt to free all objects */ 3680 for_each_kmem_cache_node(s, node, n) { 3681 free_partial(s, n); 3682 if (n->nr_partial || slabs_node(s, node)) 3683 return 1; 3684 } 3685 return 0; 3686 } 3687 3688 /******************************************************************** 3689 * Kmalloc subsystem 3690 *******************************************************************/ 3691 3692 static int __init setup_slub_min_order(char *str) 3693 { 3694 get_option(&str, &slub_min_order); 3695 3696 return 1; 3697 } 3698 3699 __setup("slub_min_order=", setup_slub_min_order); 3700 3701 static int __init setup_slub_max_order(char *str) 3702 { 3703 get_option(&str, &slub_max_order); 3704 slub_max_order = min(slub_max_order, MAX_ORDER - 1); 3705 3706 return 1; 3707 } 3708 3709 __setup("slub_max_order=", setup_slub_max_order); 3710 3711 static int __init setup_slub_min_objects(char *str) 3712 { 3713 get_option(&str, &slub_min_objects); 3714 3715 return 1; 3716 } 3717 3718 __setup("slub_min_objects=", setup_slub_min_objects); 3719 3720 void *__kmalloc(size_t size, gfp_t flags) 3721 { 3722 struct kmem_cache *s; 3723 void *ret; 3724 3725 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 3726 return kmalloc_large(size, flags); 3727 3728 s = kmalloc_slab(size, flags); 3729 3730 if (unlikely(ZERO_OR_NULL_PTR(s))) 3731 return s; 3732 3733 ret = slab_alloc(s, flags, _RET_IP_); 3734 3735 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 3736 3737 kasan_kmalloc(s, ret, size, flags); 3738 3739 return ret; 3740 } 3741 EXPORT_SYMBOL(__kmalloc); 3742 3743 #ifdef CONFIG_NUMA 3744 static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 3745 { 3746 struct page *page; 3747 void *ptr = NULL; 3748 3749 flags |= __GFP_COMP | __GFP_NOTRACK; 3750 page = alloc_pages_node(node, flags, get_order(size)); 3751 if (page) 3752 ptr = page_address(page); 3753 3754 kmalloc_large_node_hook(ptr, size, flags); 3755 return ptr; 3756 } 3757 3758 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3759 { 3760 struct kmem_cache *s; 3761 void *ret; 3762 3763 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 3764 ret = kmalloc_large_node(size, flags, node); 3765 3766 trace_kmalloc_node(_RET_IP_, ret, 3767 size, PAGE_SIZE << get_order(size), 3768 flags, node); 3769 3770 return ret; 3771 } 3772 3773 s = kmalloc_slab(size, flags); 3774 3775 if (unlikely(ZERO_OR_NULL_PTR(s))) 3776 return s; 3777 3778 ret = slab_alloc_node(s, flags, node, _RET_IP_); 3779 3780 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 3781 3782 kasan_kmalloc(s, ret, size, flags); 3783 3784 return ret; 3785 } 3786 EXPORT_SYMBOL(__kmalloc_node); 3787 #endif 3788 3789 #ifdef CONFIG_HARDENED_USERCOPY 3790 /* 3791 * Rejects objects that are incorrectly sized. 3792 * 3793 * Returns NULL if check passes, otherwise const char * to name of cache 3794 * to indicate an error. 3795 */ 3796 const char *__check_heap_object(const void *ptr, unsigned long n, 3797 struct page *page) 3798 { 3799 struct kmem_cache *s; 3800 unsigned long offset; 3801 size_t object_size; 3802 3803 /* Find object and usable object size. */ 3804 s = page->slab_cache; 3805 object_size = slab_ksize(s); 3806 3807 /* Reject impossible pointers. */ 3808 if (ptr < page_address(page)) 3809 return s->name; 3810 3811 /* Find offset within object. */ 3812 offset = (ptr - page_address(page)) % s->size; 3813 3814 /* Adjust for redzone and reject if within the redzone. */ 3815 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) { 3816 if (offset < s->red_left_pad) 3817 return s->name; 3818 offset -= s->red_left_pad; 3819 } 3820 3821 /* Allow address range falling entirely within object size. */ 3822 if (offset <= object_size && n <= object_size - offset) 3823 return NULL; 3824 3825 return s->name; 3826 } 3827 #endif /* CONFIG_HARDENED_USERCOPY */ 3828 3829 static size_t __ksize(const void *object) 3830 { 3831 struct page *page; 3832 3833 if (unlikely(object == ZERO_SIZE_PTR)) 3834 return 0; 3835 3836 page = virt_to_head_page(object); 3837 3838 if (unlikely(!PageSlab(page))) { 3839 WARN_ON(!PageCompound(page)); 3840 return PAGE_SIZE << compound_order(page); 3841 } 3842 3843 return slab_ksize(page->slab_cache); 3844 } 3845 3846 size_t ksize(const void *object) 3847 { 3848 size_t size = __ksize(object); 3849 /* We assume that ksize callers could use whole allocated area, 3850 * so we need to unpoison this area. 3851 */ 3852 kasan_unpoison_shadow(object, size); 3853 return size; 3854 } 3855 EXPORT_SYMBOL(ksize); 3856 3857 void kfree(const void *x) 3858 { 3859 struct page *page; 3860 void *object = (void *)x; 3861 3862 trace_kfree(_RET_IP_, x); 3863 3864 if (unlikely(ZERO_OR_NULL_PTR(x))) 3865 return; 3866 3867 page = virt_to_head_page(x); 3868 if (unlikely(!PageSlab(page))) { 3869 BUG_ON(!PageCompound(page)); 3870 kfree_hook(x); 3871 __free_pages(page, compound_order(page)); 3872 return; 3873 } 3874 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_); 3875 } 3876 EXPORT_SYMBOL(kfree); 3877 3878 #define SHRINK_PROMOTE_MAX 32 3879 3880 /* 3881 * kmem_cache_shrink discards empty slabs and promotes the slabs filled 3882 * up most to the head of the partial lists. New allocations will then 3883 * fill those up and thus they can be removed from the partial lists. 3884 * 3885 * The slabs with the least items are placed last. This results in them 3886 * being allocated from last increasing the chance that the last objects 3887 * are freed in them. 3888 */ 3889 int __kmem_cache_shrink(struct kmem_cache *s) 3890 { 3891 int node; 3892 int i; 3893 struct kmem_cache_node *n; 3894 struct page *page; 3895 struct page *t; 3896 struct list_head discard; 3897 struct list_head promote[SHRINK_PROMOTE_MAX]; 3898 unsigned long flags; 3899 int ret = 0; 3900 3901 flush_all(s); 3902 for_each_kmem_cache_node(s, node, n) { 3903 INIT_LIST_HEAD(&discard); 3904 for (i = 0; i < SHRINK_PROMOTE_MAX; i++) 3905 INIT_LIST_HEAD(promote + i); 3906 3907 spin_lock_irqsave(&n->list_lock, flags); 3908 3909 /* 3910 * Build lists of slabs to discard or promote. 3911 * 3912 * Note that concurrent frees may occur while we hold the 3913 * list_lock. page->inuse here is the upper limit. 3914 */ 3915 list_for_each_entry_safe(page, t, &n->partial, lru) { 3916 int free = page->objects - page->inuse; 3917 3918 /* Do not reread page->inuse */ 3919 barrier(); 3920 3921 /* We do not keep full slabs on the list */ 3922 BUG_ON(free <= 0); 3923 3924 if (free == page->objects) { 3925 list_move(&page->lru, &discard); 3926 n->nr_partial--; 3927 } else if (free <= SHRINK_PROMOTE_MAX) 3928 list_move(&page->lru, promote + free - 1); 3929 } 3930 3931 /* 3932 * Promote the slabs filled up most to the head of the 3933 * partial list. 3934 */ 3935 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) 3936 list_splice(promote + i, &n->partial); 3937 3938 spin_unlock_irqrestore(&n->list_lock, flags); 3939 3940 /* Release empty slabs */ 3941 list_for_each_entry_safe(page, t, &discard, lru) 3942 discard_slab(s, page); 3943 3944 if (slabs_node(s, node)) 3945 ret = 1; 3946 } 3947 3948 return ret; 3949 } 3950 3951 static int slab_mem_going_offline_callback(void *arg) 3952 { 3953 struct kmem_cache *s; 3954 3955 mutex_lock(&slab_mutex); 3956 list_for_each_entry(s, &slab_caches, list) 3957 __kmem_cache_shrink(s); 3958 mutex_unlock(&slab_mutex); 3959 3960 return 0; 3961 } 3962 3963 static void slab_mem_offline_callback(void *arg) 3964 { 3965 struct kmem_cache_node *n; 3966 struct kmem_cache *s; 3967 struct memory_notify *marg = arg; 3968 int offline_node; 3969 3970 offline_node = marg->status_change_nid_normal; 3971 3972 /* 3973 * If the node still has available memory. we need kmem_cache_node 3974 * for it yet. 3975 */ 3976 if (offline_node < 0) 3977 return; 3978 3979 mutex_lock(&slab_mutex); 3980 list_for_each_entry(s, &slab_caches, list) { 3981 n = get_node(s, offline_node); 3982 if (n) { 3983 /* 3984 * if n->nr_slabs > 0, slabs still exist on the node 3985 * that is going down. We were unable to free them, 3986 * and offline_pages() function shouldn't call this 3987 * callback. So, we must fail. 3988 */ 3989 BUG_ON(slabs_node(s, offline_node)); 3990 3991 s->node[offline_node] = NULL; 3992 kmem_cache_free(kmem_cache_node, n); 3993 } 3994 } 3995 mutex_unlock(&slab_mutex); 3996 } 3997 3998 static int slab_mem_going_online_callback(void *arg) 3999 { 4000 struct kmem_cache_node *n; 4001 struct kmem_cache *s; 4002 struct memory_notify *marg = arg; 4003 int nid = marg->status_change_nid_normal; 4004 int ret = 0; 4005 4006 /* 4007 * If the node's memory is already available, then kmem_cache_node is 4008 * already created. Nothing to do. 4009 */ 4010 if (nid < 0) 4011 return 0; 4012 4013 /* 4014 * We are bringing a node online. No memory is available yet. We must 4015 * allocate a kmem_cache_node structure in order to bring the node 4016 * online. 4017 */ 4018 mutex_lock(&slab_mutex); 4019 list_for_each_entry(s, &slab_caches, list) { 4020 /* 4021 * XXX: kmem_cache_alloc_node will fallback to other nodes 4022 * since memory is not yet available from the node that 4023 * is brought up. 4024 */ 4025 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 4026 if (!n) { 4027 ret = -ENOMEM; 4028 goto out; 4029 } 4030 init_kmem_cache_node(n); 4031 s->node[nid] = n; 4032 } 4033 out: 4034 mutex_unlock(&slab_mutex); 4035 return ret; 4036 } 4037 4038 static int slab_memory_callback(struct notifier_block *self, 4039 unsigned long action, void *arg) 4040 { 4041 int ret = 0; 4042 4043 switch (action) { 4044 case MEM_GOING_ONLINE: 4045 ret = slab_mem_going_online_callback(arg); 4046 break; 4047 case MEM_GOING_OFFLINE: 4048 ret = slab_mem_going_offline_callback(arg); 4049 break; 4050 case MEM_OFFLINE: 4051 case MEM_CANCEL_ONLINE: 4052 slab_mem_offline_callback(arg); 4053 break; 4054 case MEM_ONLINE: 4055 case MEM_CANCEL_OFFLINE: 4056 break; 4057 } 4058 if (ret) 4059 ret = notifier_from_errno(ret); 4060 else 4061 ret = NOTIFY_OK; 4062 return ret; 4063 } 4064 4065 static struct notifier_block slab_memory_callback_nb = { 4066 .notifier_call = slab_memory_callback, 4067 .priority = SLAB_CALLBACK_PRI, 4068 }; 4069 4070 /******************************************************************** 4071 * Basic setup of slabs 4072 *******************************************************************/ 4073 4074 /* 4075 * Used for early kmem_cache structures that were allocated using 4076 * the page allocator. Allocate them properly then fix up the pointers 4077 * that may be pointing to the wrong kmem_cache structure. 4078 */ 4079 4080 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) 4081 { 4082 int node; 4083 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); 4084 struct kmem_cache_node *n; 4085 4086 memcpy(s, static_cache, kmem_cache->object_size); 4087 4088 /* 4089 * This runs very early, and only the boot processor is supposed to be 4090 * up. Even if it weren't true, IRQs are not up so we couldn't fire 4091 * IPIs around. 4092 */ 4093 __flush_cpu_slab(s, smp_processor_id()); 4094 for_each_kmem_cache_node(s, node, n) { 4095 struct page *p; 4096 4097 list_for_each_entry(p, &n->partial, lru) 4098 p->slab_cache = s; 4099 4100 #ifdef CONFIG_SLUB_DEBUG 4101 list_for_each_entry(p, &n->full, lru) 4102 p->slab_cache = s; 4103 #endif 4104 } 4105 slab_init_memcg_params(s); 4106 list_add(&s->list, &slab_caches); 4107 return s; 4108 } 4109 4110 void __init kmem_cache_init(void) 4111 { 4112 static __initdata struct kmem_cache boot_kmem_cache, 4113 boot_kmem_cache_node; 4114 4115 if (debug_guardpage_minorder()) 4116 slub_max_order = 0; 4117 4118 kmem_cache_node = &boot_kmem_cache_node; 4119 kmem_cache = &boot_kmem_cache; 4120 4121 create_boot_cache(kmem_cache_node, "kmem_cache_node", 4122 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN); 4123 4124 register_hotmemory_notifier(&slab_memory_callback_nb); 4125 4126 /* Able to allocate the per node structures */ 4127 slab_state = PARTIAL; 4128 4129 create_boot_cache(kmem_cache, "kmem_cache", 4130 offsetof(struct kmem_cache, node) + 4131 nr_node_ids * sizeof(struct kmem_cache_node *), 4132 SLAB_HWCACHE_ALIGN); 4133 4134 kmem_cache = bootstrap(&boot_kmem_cache); 4135 4136 /* 4137 * Allocate kmem_cache_node properly from the kmem_cache slab. 4138 * kmem_cache_node is separately allocated so no need to 4139 * update any list pointers. 4140 */ 4141 kmem_cache_node = bootstrap(&boot_kmem_cache_node); 4142 4143 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 4144 setup_kmalloc_cache_index_table(); 4145 create_kmalloc_caches(0); 4146 4147 /* Setup random freelists for each cache */ 4148 init_freelist_randomization(); 4149 4150 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL, 4151 slub_cpu_dead); 4152 4153 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n", 4154 cache_line_size(), 4155 slub_min_order, slub_max_order, slub_min_objects, 4156 nr_cpu_ids, nr_node_ids); 4157 } 4158 4159 void __init kmem_cache_init_late(void) 4160 { 4161 } 4162 4163 struct kmem_cache * 4164 __kmem_cache_alias(const char *name, size_t size, size_t align, 4165 unsigned long flags, void (*ctor)(void *)) 4166 { 4167 struct kmem_cache *s, *c; 4168 4169 s = find_mergeable(size, align, flags, name, ctor); 4170 if (s) { 4171 s->refcount++; 4172 4173 /* 4174 * Adjust the object sizes so that we clear 4175 * the complete object on kzalloc. 4176 */ 4177 s->object_size = max(s->object_size, (int)size); 4178 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *))); 4179 4180 for_each_memcg_cache(c, s) { 4181 c->object_size = s->object_size; 4182 c->inuse = max_t(int, c->inuse, 4183 ALIGN(size, sizeof(void *))); 4184 } 4185 4186 if (sysfs_slab_alias(s, name)) { 4187 s->refcount--; 4188 s = NULL; 4189 } 4190 } 4191 4192 return s; 4193 } 4194 4195 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags) 4196 { 4197 int err; 4198 4199 err = kmem_cache_open(s, flags); 4200 if (err) 4201 return err; 4202 4203 /* Mutex is not taken during early boot */ 4204 if (slab_state <= UP) 4205 return 0; 4206 4207 memcg_propagate_slab_attrs(s); 4208 err = sysfs_slab_add(s); 4209 if (err) 4210 __kmem_cache_release(s); 4211 4212 return err; 4213 } 4214 4215 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 4216 { 4217 struct kmem_cache *s; 4218 void *ret; 4219 4220 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 4221 return kmalloc_large(size, gfpflags); 4222 4223 s = kmalloc_slab(size, gfpflags); 4224 4225 if (unlikely(ZERO_OR_NULL_PTR(s))) 4226 return s; 4227 4228 ret = slab_alloc(s, gfpflags, caller); 4229 4230 /* Honor the call site pointer we received. */ 4231 trace_kmalloc(caller, ret, size, s->size, gfpflags); 4232 4233 return ret; 4234 } 4235 4236 #ifdef CONFIG_NUMA 4237 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 4238 int node, unsigned long caller) 4239 { 4240 struct kmem_cache *s; 4241 void *ret; 4242 4243 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 4244 ret = kmalloc_large_node(size, gfpflags, node); 4245 4246 trace_kmalloc_node(caller, ret, 4247 size, PAGE_SIZE << get_order(size), 4248 gfpflags, node); 4249 4250 return ret; 4251 } 4252 4253 s = kmalloc_slab(size, gfpflags); 4254 4255 if (unlikely(ZERO_OR_NULL_PTR(s))) 4256 return s; 4257 4258 ret = slab_alloc_node(s, gfpflags, node, caller); 4259 4260 /* Honor the call site pointer we received. */ 4261 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 4262 4263 return ret; 4264 } 4265 #endif 4266 4267 #ifdef CONFIG_SYSFS 4268 static int count_inuse(struct page *page) 4269 { 4270 return page->inuse; 4271 } 4272 4273 static int count_total(struct page *page) 4274 { 4275 return page->objects; 4276 } 4277 #endif 4278 4279 #ifdef CONFIG_SLUB_DEBUG 4280 static int validate_slab(struct kmem_cache *s, struct page *page, 4281 unsigned long *map) 4282 { 4283 void *p; 4284 void *addr = page_address(page); 4285 4286 if (!check_slab(s, page) || 4287 !on_freelist(s, page, NULL)) 4288 return 0; 4289 4290 /* Now we know that a valid freelist exists */ 4291 bitmap_zero(map, page->objects); 4292 4293 get_map(s, page, map); 4294 for_each_object(p, s, addr, page->objects) { 4295 if (test_bit(slab_index(p, s, addr), map)) 4296 if (!check_object(s, page, p, SLUB_RED_INACTIVE)) 4297 return 0; 4298 } 4299 4300 for_each_object(p, s, addr, page->objects) 4301 if (!test_bit(slab_index(p, s, addr), map)) 4302 if (!check_object(s, page, p, SLUB_RED_ACTIVE)) 4303 return 0; 4304 return 1; 4305 } 4306 4307 static void validate_slab_slab(struct kmem_cache *s, struct page *page, 4308 unsigned long *map) 4309 { 4310 slab_lock(page); 4311 validate_slab(s, page, map); 4312 slab_unlock(page); 4313 } 4314 4315 static int validate_slab_node(struct kmem_cache *s, 4316 struct kmem_cache_node *n, unsigned long *map) 4317 { 4318 unsigned long count = 0; 4319 struct page *page; 4320 unsigned long flags; 4321 4322 spin_lock_irqsave(&n->list_lock, flags); 4323 4324 list_for_each_entry(page, &n->partial, lru) { 4325 validate_slab_slab(s, page, map); 4326 count++; 4327 } 4328 if (count != n->nr_partial) 4329 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n", 4330 s->name, count, n->nr_partial); 4331 4332 if (!(s->flags & SLAB_STORE_USER)) 4333 goto out; 4334 4335 list_for_each_entry(page, &n->full, lru) { 4336 validate_slab_slab(s, page, map); 4337 count++; 4338 } 4339 if (count != atomic_long_read(&n->nr_slabs)) 4340 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n", 4341 s->name, count, atomic_long_read(&n->nr_slabs)); 4342 4343 out: 4344 spin_unlock_irqrestore(&n->list_lock, flags); 4345 return count; 4346 } 4347 4348 static long validate_slab_cache(struct kmem_cache *s) 4349 { 4350 int node; 4351 unsigned long count = 0; 4352 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 4353 sizeof(unsigned long), GFP_KERNEL); 4354 struct kmem_cache_node *n; 4355 4356 if (!map) 4357 return -ENOMEM; 4358 4359 flush_all(s); 4360 for_each_kmem_cache_node(s, node, n) 4361 count += validate_slab_node(s, n, map); 4362 kfree(map); 4363 return count; 4364 } 4365 /* 4366 * Generate lists of code addresses where slabcache objects are allocated 4367 * and freed. 4368 */ 4369 4370 struct location { 4371 unsigned long count; 4372 unsigned long addr; 4373 long long sum_time; 4374 long min_time; 4375 long max_time; 4376 long min_pid; 4377 long max_pid; 4378 DECLARE_BITMAP(cpus, NR_CPUS); 4379 nodemask_t nodes; 4380 }; 4381 4382 struct loc_track { 4383 unsigned long max; 4384 unsigned long count; 4385 struct location *loc; 4386 }; 4387 4388 static void free_loc_track(struct loc_track *t) 4389 { 4390 if (t->max) 4391 free_pages((unsigned long)t->loc, 4392 get_order(sizeof(struct location) * t->max)); 4393 } 4394 4395 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 4396 { 4397 struct location *l; 4398 int order; 4399 4400 order = get_order(sizeof(struct location) * max); 4401 4402 l = (void *)__get_free_pages(flags, order); 4403 if (!l) 4404 return 0; 4405 4406 if (t->count) { 4407 memcpy(l, t->loc, sizeof(struct location) * t->count); 4408 free_loc_track(t); 4409 } 4410 t->max = max; 4411 t->loc = l; 4412 return 1; 4413 } 4414 4415 static int add_location(struct loc_track *t, struct kmem_cache *s, 4416 const struct track *track) 4417 { 4418 long start, end, pos; 4419 struct location *l; 4420 unsigned long caddr; 4421 unsigned long age = jiffies - track->when; 4422 4423 start = -1; 4424 end = t->count; 4425 4426 for ( ; ; ) { 4427 pos = start + (end - start + 1) / 2; 4428 4429 /* 4430 * There is nothing at "end". If we end up there 4431 * we need to add something to before end. 4432 */ 4433 if (pos == end) 4434 break; 4435 4436 caddr = t->loc[pos].addr; 4437 if (track->addr == caddr) { 4438 4439 l = &t->loc[pos]; 4440 l->count++; 4441 if (track->when) { 4442 l->sum_time += age; 4443 if (age < l->min_time) 4444 l->min_time = age; 4445 if (age > l->max_time) 4446 l->max_time = age; 4447 4448 if (track->pid < l->min_pid) 4449 l->min_pid = track->pid; 4450 if (track->pid > l->max_pid) 4451 l->max_pid = track->pid; 4452 4453 cpumask_set_cpu(track->cpu, 4454 to_cpumask(l->cpus)); 4455 } 4456 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4457 return 1; 4458 } 4459 4460 if (track->addr < caddr) 4461 end = pos; 4462 else 4463 start = pos; 4464 } 4465 4466 /* 4467 * Not found. Insert new tracking element. 4468 */ 4469 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 4470 return 0; 4471 4472 l = t->loc + pos; 4473 if (pos < t->count) 4474 memmove(l + 1, l, 4475 (t->count - pos) * sizeof(struct location)); 4476 t->count++; 4477 l->count = 1; 4478 l->addr = track->addr; 4479 l->sum_time = age; 4480 l->min_time = age; 4481 l->max_time = age; 4482 l->min_pid = track->pid; 4483 l->max_pid = track->pid; 4484 cpumask_clear(to_cpumask(l->cpus)); 4485 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 4486 nodes_clear(l->nodes); 4487 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4488 return 1; 4489 } 4490 4491 static void process_slab(struct loc_track *t, struct kmem_cache *s, 4492 struct page *page, enum track_item alloc, 4493 unsigned long *map) 4494 { 4495 void *addr = page_address(page); 4496 void *p; 4497 4498 bitmap_zero(map, page->objects); 4499 get_map(s, page, map); 4500 4501 for_each_object(p, s, addr, page->objects) 4502 if (!test_bit(slab_index(p, s, addr), map)) 4503 add_location(t, s, get_track(s, p, alloc)); 4504 } 4505 4506 static int list_locations(struct kmem_cache *s, char *buf, 4507 enum track_item alloc) 4508 { 4509 int len = 0; 4510 unsigned long i; 4511 struct loc_track t = { 0, 0, NULL }; 4512 int node; 4513 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) * 4514 sizeof(unsigned long), GFP_KERNEL); 4515 struct kmem_cache_node *n; 4516 4517 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 4518 GFP_TEMPORARY)) { 4519 kfree(map); 4520 return sprintf(buf, "Out of memory\n"); 4521 } 4522 /* Push back cpu slabs */ 4523 flush_all(s); 4524 4525 for_each_kmem_cache_node(s, node, n) { 4526 unsigned long flags; 4527 struct page *page; 4528 4529 if (!atomic_long_read(&n->nr_slabs)) 4530 continue; 4531 4532 spin_lock_irqsave(&n->list_lock, flags); 4533 list_for_each_entry(page, &n->partial, lru) 4534 process_slab(&t, s, page, alloc, map); 4535 list_for_each_entry(page, &n->full, lru) 4536 process_slab(&t, s, page, alloc, map); 4537 spin_unlock_irqrestore(&n->list_lock, flags); 4538 } 4539 4540 for (i = 0; i < t.count; i++) { 4541 struct location *l = &t.loc[i]; 4542 4543 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 4544 break; 4545 len += sprintf(buf + len, "%7ld ", l->count); 4546 4547 if (l->addr) 4548 len += sprintf(buf + len, "%pS", (void *)l->addr); 4549 else 4550 len += sprintf(buf + len, "<not-available>"); 4551 4552 if (l->sum_time != l->min_time) { 4553 len += sprintf(buf + len, " age=%ld/%ld/%ld", 4554 l->min_time, 4555 (long)div_u64(l->sum_time, l->count), 4556 l->max_time); 4557 } else 4558 len += sprintf(buf + len, " age=%ld", 4559 l->min_time); 4560 4561 if (l->min_pid != l->max_pid) 4562 len += sprintf(buf + len, " pid=%ld-%ld", 4563 l->min_pid, l->max_pid); 4564 else 4565 len += sprintf(buf + len, " pid=%ld", 4566 l->min_pid); 4567 4568 if (num_online_cpus() > 1 && 4569 !cpumask_empty(to_cpumask(l->cpus)) && 4570 len < PAGE_SIZE - 60) 4571 len += scnprintf(buf + len, PAGE_SIZE - len - 50, 4572 " cpus=%*pbl", 4573 cpumask_pr_args(to_cpumask(l->cpus))); 4574 4575 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 4576 len < PAGE_SIZE - 60) 4577 len += scnprintf(buf + len, PAGE_SIZE - len - 50, 4578 " nodes=%*pbl", 4579 nodemask_pr_args(&l->nodes)); 4580 4581 len += sprintf(buf + len, "\n"); 4582 } 4583 4584 free_loc_track(&t); 4585 kfree(map); 4586 if (!t.count) 4587 len += sprintf(buf, "No data\n"); 4588 return len; 4589 } 4590 #endif 4591 4592 #ifdef SLUB_RESILIENCY_TEST 4593 static void __init resiliency_test(void) 4594 { 4595 u8 *p; 4596 4597 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10); 4598 4599 pr_err("SLUB resiliency testing\n"); 4600 pr_err("-----------------------\n"); 4601 pr_err("A. Corruption after allocation\n"); 4602 4603 p = kzalloc(16, GFP_KERNEL); 4604 p[16] = 0x12; 4605 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n", 4606 p + 16); 4607 4608 validate_slab_cache(kmalloc_caches[4]); 4609 4610 /* Hmmm... The next two are dangerous */ 4611 p = kzalloc(32, GFP_KERNEL); 4612 p[32 + sizeof(void *)] = 0x34; 4613 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n", 4614 p); 4615 pr_err("If allocated object is overwritten then not detectable\n\n"); 4616 4617 validate_slab_cache(kmalloc_caches[5]); 4618 p = kzalloc(64, GFP_KERNEL); 4619 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 4620 *p = 0x56; 4621 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 4622 p); 4623 pr_err("If allocated object is overwritten then not detectable\n\n"); 4624 validate_slab_cache(kmalloc_caches[6]); 4625 4626 pr_err("\nB. Corruption after free\n"); 4627 p = kzalloc(128, GFP_KERNEL); 4628 kfree(p); 4629 *p = 0x78; 4630 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 4631 validate_slab_cache(kmalloc_caches[7]); 4632 4633 p = kzalloc(256, GFP_KERNEL); 4634 kfree(p); 4635 p[50] = 0x9a; 4636 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); 4637 validate_slab_cache(kmalloc_caches[8]); 4638 4639 p = kzalloc(512, GFP_KERNEL); 4640 kfree(p); 4641 p[512] = 0xab; 4642 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 4643 validate_slab_cache(kmalloc_caches[9]); 4644 } 4645 #else 4646 #ifdef CONFIG_SYSFS 4647 static void resiliency_test(void) {}; 4648 #endif 4649 #endif 4650 4651 #ifdef CONFIG_SYSFS 4652 enum slab_stat_type { 4653 SL_ALL, /* All slabs */ 4654 SL_PARTIAL, /* Only partially allocated slabs */ 4655 SL_CPU, /* Only slabs used for cpu caches */ 4656 SL_OBJECTS, /* Determine allocated objects not slabs */ 4657 SL_TOTAL /* Determine object capacity not slabs */ 4658 }; 4659 4660 #define SO_ALL (1 << SL_ALL) 4661 #define SO_PARTIAL (1 << SL_PARTIAL) 4662 #define SO_CPU (1 << SL_CPU) 4663 #define SO_OBJECTS (1 << SL_OBJECTS) 4664 #define SO_TOTAL (1 << SL_TOTAL) 4665 4666 static ssize_t show_slab_objects(struct kmem_cache *s, 4667 char *buf, unsigned long flags) 4668 { 4669 unsigned long total = 0; 4670 int node; 4671 int x; 4672 unsigned long *nodes; 4673 4674 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL); 4675 if (!nodes) 4676 return -ENOMEM; 4677 4678 if (flags & SO_CPU) { 4679 int cpu; 4680 4681 for_each_possible_cpu(cpu) { 4682 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, 4683 cpu); 4684 int node; 4685 struct page *page; 4686 4687 page = READ_ONCE(c->page); 4688 if (!page) 4689 continue; 4690 4691 node = page_to_nid(page); 4692 if (flags & SO_TOTAL) 4693 x = page->objects; 4694 else if (flags & SO_OBJECTS) 4695 x = page->inuse; 4696 else 4697 x = 1; 4698 4699 total += x; 4700 nodes[node] += x; 4701 4702 page = READ_ONCE(c->partial); 4703 if (page) { 4704 node = page_to_nid(page); 4705 if (flags & SO_TOTAL) 4706 WARN_ON_ONCE(1); 4707 else if (flags & SO_OBJECTS) 4708 WARN_ON_ONCE(1); 4709 else 4710 x = page->pages; 4711 total += x; 4712 nodes[node] += x; 4713 } 4714 } 4715 } 4716 4717 get_online_mems(); 4718 #ifdef CONFIG_SLUB_DEBUG 4719 if (flags & SO_ALL) { 4720 struct kmem_cache_node *n; 4721 4722 for_each_kmem_cache_node(s, node, n) { 4723 4724 if (flags & SO_TOTAL) 4725 x = atomic_long_read(&n->total_objects); 4726 else if (flags & SO_OBJECTS) 4727 x = atomic_long_read(&n->total_objects) - 4728 count_partial(n, count_free); 4729 else 4730 x = atomic_long_read(&n->nr_slabs); 4731 total += x; 4732 nodes[node] += x; 4733 } 4734 4735 } else 4736 #endif 4737 if (flags & SO_PARTIAL) { 4738 struct kmem_cache_node *n; 4739 4740 for_each_kmem_cache_node(s, node, n) { 4741 if (flags & SO_TOTAL) 4742 x = count_partial(n, count_total); 4743 else if (flags & SO_OBJECTS) 4744 x = count_partial(n, count_inuse); 4745 else 4746 x = n->nr_partial; 4747 total += x; 4748 nodes[node] += x; 4749 } 4750 } 4751 x = sprintf(buf, "%lu", total); 4752 #ifdef CONFIG_NUMA 4753 for (node = 0; node < nr_node_ids; node++) 4754 if (nodes[node]) 4755 x += sprintf(buf + x, " N%d=%lu", 4756 node, nodes[node]); 4757 #endif 4758 put_online_mems(); 4759 kfree(nodes); 4760 return x + sprintf(buf + x, "\n"); 4761 } 4762 4763 #ifdef CONFIG_SLUB_DEBUG 4764 static int any_slab_objects(struct kmem_cache *s) 4765 { 4766 int node; 4767 struct kmem_cache_node *n; 4768 4769 for_each_kmem_cache_node(s, node, n) 4770 if (atomic_long_read(&n->total_objects)) 4771 return 1; 4772 4773 return 0; 4774 } 4775 #endif 4776 4777 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 4778 #define to_slab(n) container_of(n, struct kmem_cache, kobj) 4779 4780 struct slab_attribute { 4781 struct attribute attr; 4782 ssize_t (*show)(struct kmem_cache *s, char *buf); 4783 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 4784 }; 4785 4786 #define SLAB_ATTR_RO(_name) \ 4787 static struct slab_attribute _name##_attr = \ 4788 __ATTR(_name, 0400, _name##_show, NULL) 4789 4790 #define SLAB_ATTR(_name) \ 4791 static struct slab_attribute _name##_attr = \ 4792 __ATTR(_name, 0600, _name##_show, _name##_store) 4793 4794 static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 4795 { 4796 return sprintf(buf, "%d\n", s->size); 4797 } 4798 SLAB_ATTR_RO(slab_size); 4799 4800 static ssize_t align_show(struct kmem_cache *s, char *buf) 4801 { 4802 return sprintf(buf, "%d\n", s->align); 4803 } 4804 SLAB_ATTR_RO(align); 4805 4806 static ssize_t object_size_show(struct kmem_cache *s, char *buf) 4807 { 4808 return sprintf(buf, "%d\n", s->object_size); 4809 } 4810 SLAB_ATTR_RO(object_size); 4811 4812 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 4813 { 4814 return sprintf(buf, "%d\n", oo_objects(s->oo)); 4815 } 4816 SLAB_ATTR_RO(objs_per_slab); 4817 4818 static ssize_t order_store(struct kmem_cache *s, 4819 const char *buf, size_t length) 4820 { 4821 unsigned long order; 4822 int err; 4823 4824 err = kstrtoul(buf, 10, &order); 4825 if (err) 4826 return err; 4827 4828 if (order > slub_max_order || order < slub_min_order) 4829 return -EINVAL; 4830 4831 calculate_sizes(s, order); 4832 return length; 4833 } 4834 4835 static ssize_t order_show(struct kmem_cache *s, char *buf) 4836 { 4837 return sprintf(buf, "%d\n", oo_order(s->oo)); 4838 } 4839 SLAB_ATTR(order); 4840 4841 static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 4842 { 4843 return sprintf(buf, "%lu\n", s->min_partial); 4844 } 4845 4846 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 4847 size_t length) 4848 { 4849 unsigned long min; 4850 int err; 4851 4852 err = kstrtoul(buf, 10, &min); 4853 if (err) 4854 return err; 4855 4856 set_min_partial(s, min); 4857 return length; 4858 } 4859 SLAB_ATTR(min_partial); 4860 4861 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) 4862 { 4863 return sprintf(buf, "%u\n", s->cpu_partial); 4864 } 4865 4866 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, 4867 size_t length) 4868 { 4869 unsigned long objects; 4870 int err; 4871 4872 err = kstrtoul(buf, 10, &objects); 4873 if (err) 4874 return err; 4875 if (objects && !kmem_cache_has_cpu_partial(s)) 4876 return -EINVAL; 4877 4878 s->cpu_partial = objects; 4879 flush_all(s); 4880 return length; 4881 } 4882 SLAB_ATTR(cpu_partial); 4883 4884 static ssize_t ctor_show(struct kmem_cache *s, char *buf) 4885 { 4886 if (!s->ctor) 4887 return 0; 4888 return sprintf(buf, "%pS\n", s->ctor); 4889 } 4890 SLAB_ATTR_RO(ctor); 4891 4892 static ssize_t aliases_show(struct kmem_cache *s, char *buf) 4893 { 4894 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1); 4895 } 4896 SLAB_ATTR_RO(aliases); 4897 4898 static ssize_t partial_show(struct kmem_cache *s, char *buf) 4899 { 4900 return show_slab_objects(s, buf, SO_PARTIAL); 4901 } 4902 SLAB_ATTR_RO(partial); 4903 4904 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 4905 { 4906 return show_slab_objects(s, buf, SO_CPU); 4907 } 4908 SLAB_ATTR_RO(cpu_slabs); 4909 4910 static ssize_t objects_show(struct kmem_cache *s, char *buf) 4911 { 4912 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 4913 } 4914 SLAB_ATTR_RO(objects); 4915 4916 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 4917 { 4918 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 4919 } 4920 SLAB_ATTR_RO(objects_partial); 4921 4922 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) 4923 { 4924 int objects = 0; 4925 int pages = 0; 4926 int cpu; 4927 int len; 4928 4929 for_each_online_cpu(cpu) { 4930 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial; 4931 4932 if (page) { 4933 pages += page->pages; 4934 objects += page->pobjects; 4935 } 4936 } 4937 4938 len = sprintf(buf, "%d(%d)", objects, pages); 4939 4940 #ifdef CONFIG_SMP 4941 for_each_online_cpu(cpu) { 4942 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial; 4943 4944 if (page && len < PAGE_SIZE - 20) 4945 len += sprintf(buf + len, " C%d=%d(%d)", cpu, 4946 page->pobjects, page->pages); 4947 } 4948 #endif 4949 return len + sprintf(buf + len, "\n"); 4950 } 4951 SLAB_ATTR_RO(slabs_cpu_partial); 4952 4953 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 4954 { 4955 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 4956 } 4957 4958 static ssize_t reclaim_account_store(struct kmem_cache *s, 4959 const char *buf, size_t length) 4960 { 4961 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 4962 if (buf[0] == '1') 4963 s->flags |= SLAB_RECLAIM_ACCOUNT; 4964 return length; 4965 } 4966 SLAB_ATTR(reclaim_account); 4967 4968 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 4969 { 4970 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 4971 } 4972 SLAB_ATTR_RO(hwcache_align); 4973 4974 #ifdef CONFIG_ZONE_DMA 4975 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 4976 { 4977 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 4978 } 4979 SLAB_ATTR_RO(cache_dma); 4980 #endif 4981 4982 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 4983 { 4984 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU)); 4985 } 4986 SLAB_ATTR_RO(destroy_by_rcu); 4987 4988 static ssize_t reserved_show(struct kmem_cache *s, char *buf) 4989 { 4990 return sprintf(buf, "%d\n", s->reserved); 4991 } 4992 SLAB_ATTR_RO(reserved); 4993 4994 #ifdef CONFIG_SLUB_DEBUG 4995 static ssize_t slabs_show(struct kmem_cache *s, char *buf) 4996 { 4997 return show_slab_objects(s, buf, SO_ALL); 4998 } 4999 SLAB_ATTR_RO(slabs); 5000 5001 static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 5002 { 5003 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 5004 } 5005 SLAB_ATTR_RO(total_objects); 5006 5007 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 5008 { 5009 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS)); 5010 } 5011 5012 static ssize_t sanity_checks_store(struct kmem_cache *s, 5013 const char *buf, size_t length) 5014 { 5015 s->flags &= ~SLAB_CONSISTENCY_CHECKS; 5016 if (buf[0] == '1') { 5017 s->flags &= ~__CMPXCHG_DOUBLE; 5018 s->flags |= SLAB_CONSISTENCY_CHECKS; 5019 } 5020 return length; 5021 } 5022 SLAB_ATTR(sanity_checks); 5023 5024 static ssize_t trace_show(struct kmem_cache *s, char *buf) 5025 { 5026 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 5027 } 5028 5029 static ssize_t trace_store(struct kmem_cache *s, const char *buf, 5030 size_t length) 5031 { 5032 /* 5033 * Tracing a merged cache is going to give confusing results 5034 * as well as cause other issues like converting a mergeable 5035 * cache into an umergeable one. 5036 */ 5037 if (s->refcount > 1) 5038 return -EINVAL; 5039 5040 s->flags &= ~SLAB_TRACE; 5041 if (buf[0] == '1') { 5042 s->flags &= ~__CMPXCHG_DOUBLE; 5043 s->flags |= SLAB_TRACE; 5044 } 5045 return length; 5046 } 5047 SLAB_ATTR(trace); 5048 5049 static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 5050 { 5051 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 5052 } 5053 5054 static ssize_t red_zone_store(struct kmem_cache *s, 5055 const char *buf, size_t length) 5056 { 5057 if (any_slab_objects(s)) 5058 return -EBUSY; 5059 5060 s->flags &= ~SLAB_RED_ZONE; 5061 if (buf[0] == '1') { 5062 s->flags |= SLAB_RED_ZONE; 5063 } 5064 calculate_sizes(s, -1); 5065 return length; 5066 } 5067 SLAB_ATTR(red_zone); 5068 5069 static ssize_t poison_show(struct kmem_cache *s, char *buf) 5070 { 5071 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 5072 } 5073 5074 static ssize_t poison_store(struct kmem_cache *s, 5075 const char *buf, size_t length) 5076 { 5077 if (any_slab_objects(s)) 5078 return -EBUSY; 5079 5080 s->flags &= ~SLAB_POISON; 5081 if (buf[0] == '1') { 5082 s->flags |= SLAB_POISON; 5083 } 5084 calculate_sizes(s, -1); 5085 return length; 5086 } 5087 SLAB_ATTR(poison); 5088 5089 static ssize_t store_user_show(struct kmem_cache *s, char *buf) 5090 { 5091 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 5092 } 5093 5094 static ssize_t store_user_store(struct kmem_cache *s, 5095 const char *buf, size_t length) 5096 { 5097 if (any_slab_objects(s)) 5098 return -EBUSY; 5099 5100 s->flags &= ~SLAB_STORE_USER; 5101 if (buf[0] == '1') { 5102 s->flags &= ~__CMPXCHG_DOUBLE; 5103 s->flags |= SLAB_STORE_USER; 5104 } 5105 calculate_sizes(s, -1); 5106 return length; 5107 } 5108 SLAB_ATTR(store_user); 5109 5110 static ssize_t validate_show(struct kmem_cache *s, char *buf) 5111 { 5112 return 0; 5113 } 5114 5115 static ssize_t validate_store(struct kmem_cache *s, 5116 const char *buf, size_t length) 5117 { 5118 int ret = -EINVAL; 5119 5120 if (buf[0] == '1') { 5121 ret = validate_slab_cache(s); 5122 if (ret >= 0) 5123 ret = length; 5124 } 5125 return ret; 5126 } 5127 SLAB_ATTR(validate); 5128 5129 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 5130 { 5131 if (!(s->flags & SLAB_STORE_USER)) 5132 return -ENOSYS; 5133 return list_locations(s, buf, TRACK_ALLOC); 5134 } 5135 SLAB_ATTR_RO(alloc_calls); 5136 5137 static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 5138 { 5139 if (!(s->flags & SLAB_STORE_USER)) 5140 return -ENOSYS; 5141 return list_locations(s, buf, TRACK_FREE); 5142 } 5143 SLAB_ATTR_RO(free_calls); 5144 #endif /* CONFIG_SLUB_DEBUG */ 5145 5146 #ifdef CONFIG_FAILSLAB 5147 static ssize_t failslab_show(struct kmem_cache *s, char *buf) 5148 { 5149 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 5150 } 5151 5152 static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 5153 size_t length) 5154 { 5155 if (s->refcount > 1) 5156 return -EINVAL; 5157 5158 s->flags &= ~SLAB_FAILSLAB; 5159 if (buf[0] == '1') 5160 s->flags |= SLAB_FAILSLAB; 5161 return length; 5162 } 5163 SLAB_ATTR(failslab); 5164 #endif 5165 5166 static ssize_t shrink_show(struct kmem_cache *s, char *buf) 5167 { 5168 return 0; 5169 } 5170 5171 static ssize_t shrink_store(struct kmem_cache *s, 5172 const char *buf, size_t length) 5173 { 5174 if (buf[0] == '1') 5175 kmem_cache_shrink(s); 5176 else 5177 return -EINVAL; 5178 return length; 5179 } 5180 SLAB_ATTR(shrink); 5181 5182 #ifdef CONFIG_NUMA 5183 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 5184 { 5185 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10); 5186 } 5187 5188 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 5189 const char *buf, size_t length) 5190 { 5191 unsigned long ratio; 5192 int err; 5193 5194 err = kstrtoul(buf, 10, &ratio); 5195 if (err) 5196 return err; 5197 5198 if (ratio <= 100) 5199 s->remote_node_defrag_ratio = ratio * 10; 5200 5201 return length; 5202 } 5203 SLAB_ATTR(remote_node_defrag_ratio); 5204 #endif 5205 5206 #ifdef CONFIG_SLUB_STATS 5207 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 5208 { 5209 unsigned long sum = 0; 5210 int cpu; 5211 int len; 5212 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL); 5213 5214 if (!data) 5215 return -ENOMEM; 5216 5217 for_each_online_cpu(cpu) { 5218 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 5219 5220 data[cpu] = x; 5221 sum += x; 5222 } 5223 5224 len = sprintf(buf, "%lu", sum); 5225 5226 #ifdef CONFIG_SMP 5227 for_each_online_cpu(cpu) { 5228 if (data[cpu] && len < PAGE_SIZE - 20) 5229 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 5230 } 5231 #endif 5232 kfree(data); 5233 return len + sprintf(buf + len, "\n"); 5234 } 5235 5236 static void clear_stat(struct kmem_cache *s, enum stat_item si) 5237 { 5238 int cpu; 5239 5240 for_each_online_cpu(cpu) 5241 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 5242 } 5243 5244 #define STAT_ATTR(si, text) \ 5245 static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 5246 { \ 5247 return show_stat(s, buf, si); \ 5248 } \ 5249 static ssize_t text##_store(struct kmem_cache *s, \ 5250 const char *buf, size_t length) \ 5251 { \ 5252 if (buf[0] != '0') \ 5253 return -EINVAL; \ 5254 clear_stat(s, si); \ 5255 return length; \ 5256 } \ 5257 SLAB_ATTR(text); \ 5258 5259 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 5260 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 5261 STAT_ATTR(FREE_FASTPATH, free_fastpath); 5262 STAT_ATTR(FREE_SLOWPATH, free_slowpath); 5263 STAT_ATTR(FREE_FROZEN, free_frozen); 5264 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 5265 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 5266 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 5267 STAT_ATTR(ALLOC_SLAB, alloc_slab); 5268 STAT_ATTR(ALLOC_REFILL, alloc_refill); 5269 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); 5270 STAT_ATTR(FREE_SLAB, free_slab); 5271 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 5272 STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 5273 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 5274 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 5275 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 5276 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 5277 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); 5278 STAT_ATTR(ORDER_FALLBACK, order_fallback); 5279 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); 5280 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); 5281 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); 5282 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); 5283 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); 5284 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); 5285 #endif 5286 5287 static struct attribute *slab_attrs[] = { 5288 &slab_size_attr.attr, 5289 &object_size_attr.attr, 5290 &objs_per_slab_attr.attr, 5291 &order_attr.attr, 5292 &min_partial_attr.attr, 5293 &cpu_partial_attr.attr, 5294 &objects_attr.attr, 5295 &objects_partial_attr.attr, 5296 &partial_attr.attr, 5297 &cpu_slabs_attr.attr, 5298 &ctor_attr.attr, 5299 &aliases_attr.attr, 5300 &align_attr.attr, 5301 &hwcache_align_attr.attr, 5302 &reclaim_account_attr.attr, 5303 &destroy_by_rcu_attr.attr, 5304 &shrink_attr.attr, 5305 &reserved_attr.attr, 5306 &slabs_cpu_partial_attr.attr, 5307 #ifdef CONFIG_SLUB_DEBUG 5308 &total_objects_attr.attr, 5309 &slabs_attr.attr, 5310 &sanity_checks_attr.attr, 5311 &trace_attr.attr, 5312 &red_zone_attr.attr, 5313 &poison_attr.attr, 5314 &store_user_attr.attr, 5315 &validate_attr.attr, 5316 &alloc_calls_attr.attr, 5317 &free_calls_attr.attr, 5318 #endif 5319 #ifdef CONFIG_ZONE_DMA 5320 &cache_dma_attr.attr, 5321 #endif 5322 #ifdef CONFIG_NUMA 5323 &remote_node_defrag_ratio_attr.attr, 5324 #endif 5325 #ifdef CONFIG_SLUB_STATS 5326 &alloc_fastpath_attr.attr, 5327 &alloc_slowpath_attr.attr, 5328 &free_fastpath_attr.attr, 5329 &free_slowpath_attr.attr, 5330 &free_frozen_attr.attr, 5331 &free_add_partial_attr.attr, 5332 &free_remove_partial_attr.attr, 5333 &alloc_from_partial_attr.attr, 5334 &alloc_slab_attr.attr, 5335 &alloc_refill_attr.attr, 5336 &alloc_node_mismatch_attr.attr, 5337 &free_slab_attr.attr, 5338 &cpuslab_flush_attr.attr, 5339 &deactivate_full_attr.attr, 5340 &deactivate_empty_attr.attr, 5341 &deactivate_to_head_attr.attr, 5342 &deactivate_to_tail_attr.attr, 5343 &deactivate_remote_frees_attr.attr, 5344 &deactivate_bypass_attr.attr, 5345 &order_fallback_attr.attr, 5346 &cmpxchg_double_fail_attr.attr, 5347 &cmpxchg_double_cpu_fail_attr.attr, 5348 &cpu_partial_alloc_attr.attr, 5349 &cpu_partial_free_attr.attr, 5350 &cpu_partial_node_attr.attr, 5351 &cpu_partial_drain_attr.attr, 5352 #endif 5353 #ifdef CONFIG_FAILSLAB 5354 &failslab_attr.attr, 5355 #endif 5356 5357 NULL 5358 }; 5359 5360 static struct attribute_group slab_attr_group = { 5361 .attrs = slab_attrs, 5362 }; 5363 5364 static ssize_t slab_attr_show(struct kobject *kobj, 5365 struct attribute *attr, 5366 char *buf) 5367 { 5368 struct slab_attribute *attribute; 5369 struct kmem_cache *s; 5370 int err; 5371 5372 attribute = to_slab_attr(attr); 5373 s = to_slab(kobj); 5374 5375 if (!attribute->show) 5376 return -EIO; 5377 5378 err = attribute->show(s, buf); 5379 5380 return err; 5381 } 5382 5383 static ssize_t slab_attr_store(struct kobject *kobj, 5384 struct attribute *attr, 5385 const char *buf, size_t len) 5386 { 5387 struct slab_attribute *attribute; 5388 struct kmem_cache *s; 5389 int err; 5390 5391 attribute = to_slab_attr(attr); 5392 s = to_slab(kobj); 5393 5394 if (!attribute->store) 5395 return -EIO; 5396 5397 err = attribute->store(s, buf, len); 5398 #ifdef CONFIG_MEMCG 5399 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) { 5400 struct kmem_cache *c; 5401 5402 mutex_lock(&slab_mutex); 5403 if (s->max_attr_size < len) 5404 s->max_attr_size = len; 5405 5406 /* 5407 * This is a best effort propagation, so this function's return 5408 * value will be determined by the parent cache only. This is 5409 * basically because not all attributes will have a well 5410 * defined semantics for rollbacks - most of the actions will 5411 * have permanent effects. 5412 * 5413 * Returning the error value of any of the children that fail 5414 * is not 100 % defined, in the sense that users seeing the 5415 * error code won't be able to know anything about the state of 5416 * the cache. 5417 * 5418 * Only returning the error code for the parent cache at least 5419 * has well defined semantics. The cache being written to 5420 * directly either failed or succeeded, in which case we loop 5421 * through the descendants with best-effort propagation. 5422 */ 5423 for_each_memcg_cache(c, s) 5424 attribute->store(c, buf, len); 5425 mutex_unlock(&slab_mutex); 5426 } 5427 #endif 5428 return err; 5429 } 5430 5431 static void memcg_propagate_slab_attrs(struct kmem_cache *s) 5432 { 5433 #ifdef CONFIG_MEMCG 5434 int i; 5435 char *buffer = NULL; 5436 struct kmem_cache *root_cache; 5437 5438 if (is_root_cache(s)) 5439 return; 5440 5441 root_cache = s->memcg_params.root_cache; 5442 5443 /* 5444 * This mean this cache had no attribute written. Therefore, no point 5445 * in copying default values around 5446 */ 5447 if (!root_cache->max_attr_size) 5448 return; 5449 5450 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) { 5451 char mbuf[64]; 5452 char *buf; 5453 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]); 5454 5455 if (!attr || !attr->store || !attr->show) 5456 continue; 5457 5458 /* 5459 * It is really bad that we have to allocate here, so we will 5460 * do it only as a fallback. If we actually allocate, though, 5461 * we can just use the allocated buffer until the end. 5462 * 5463 * Most of the slub attributes will tend to be very small in 5464 * size, but sysfs allows buffers up to a page, so they can 5465 * theoretically happen. 5466 */ 5467 if (buffer) 5468 buf = buffer; 5469 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf)) 5470 buf = mbuf; 5471 else { 5472 buffer = (char *) get_zeroed_page(GFP_KERNEL); 5473 if (WARN_ON(!buffer)) 5474 continue; 5475 buf = buffer; 5476 } 5477 5478 attr->show(root_cache, buf); 5479 attr->store(s, buf, strlen(buf)); 5480 } 5481 5482 if (buffer) 5483 free_page((unsigned long)buffer); 5484 #endif 5485 } 5486 5487 static void kmem_cache_release(struct kobject *k) 5488 { 5489 slab_kmem_cache_release(to_slab(k)); 5490 } 5491 5492 static const struct sysfs_ops slab_sysfs_ops = { 5493 .show = slab_attr_show, 5494 .store = slab_attr_store, 5495 }; 5496 5497 static struct kobj_type slab_ktype = { 5498 .sysfs_ops = &slab_sysfs_ops, 5499 .release = kmem_cache_release, 5500 }; 5501 5502 static int uevent_filter(struct kset *kset, struct kobject *kobj) 5503 { 5504 struct kobj_type *ktype = get_ktype(kobj); 5505 5506 if (ktype == &slab_ktype) 5507 return 1; 5508 return 0; 5509 } 5510 5511 static const struct kset_uevent_ops slab_uevent_ops = { 5512 .filter = uevent_filter, 5513 }; 5514 5515 static struct kset *slab_kset; 5516 5517 static inline struct kset *cache_kset(struct kmem_cache *s) 5518 { 5519 #ifdef CONFIG_MEMCG 5520 if (!is_root_cache(s)) 5521 return s->memcg_params.root_cache->memcg_kset; 5522 #endif 5523 return slab_kset; 5524 } 5525 5526 #define ID_STR_LENGTH 64 5527 5528 /* Create a unique string id for a slab cache: 5529 * 5530 * Format :[flags-]size 5531 */ 5532 static char *create_unique_id(struct kmem_cache *s) 5533 { 5534 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 5535 char *p = name; 5536 5537 BUG_ON(!name); 5538 5539 *p++ = ':'; 5540 /* 5541 * First flags affecting slabcache operations. We will only 5542 * get here for aliasable slabs so we do not need to support 5543 * too many flags. The flags here must cover all flags that 5544 * are matched during merging to guarantee that the id is 5545 * unique. 5546 */ 5547 if (s->flags & SLAB_CACHE_DMA) 5548 *p++ = 'd'; 5549 if (s->flags & SLAB_RECLAIM_ACCOUNT) 5550 *p++ = 'a'; 5551 if (s->flags & SLAB_CONSISTENCY_CHECKS) 5552 *p++ = 'F'; 5553 if (!(s->flags & SLAB_NOTRACK)) 5554 *p++ = 't'; 5555 if (s->flags & SLAB_ACCOUNT) 5556 *p++ = 'A'; 5557 if (p != name + 1) 5558 *p++ = '-'; 5559 p += sprintf(p, "%07d", s->size); 5560 5561 BUG_ON(p > name + ID_STR_LENGTH - 1); 5562 return name; 5563 } 5564 5565 static int sysfs_slab_add(struct kmem_cache *s) 5566 { 5567 int err; 5568 const char *name; 5569 int unmergeable = slab_unmergeable(s); 5570 5571 if (unmergeable) { 5572 /* 5573 * Slabcache can never be merged so we can use the name proper. 5574 * This is typically the case for debug situations. In that 5575 * case we can catch duplicate names easily. 5576 */ 5577 sysfs_remove_link(&slab_kset->kobj, s->name); 5578 name = s->name; 5579 } else { 5580 /* 5581 * Create a unique name for the slab as a target 5582 * for the symlinks. 5583 */ 5584 name = create_unique_id(s); 5585 } 5586 5587 s->kobj.kset = cache_kset(s); 5588 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); 5589 if (err) 5590 goto out; 5591 5592 err = sysfs_create_group(&s->kobj, &slab_attr_group); 5593 if (err) 5594 goto out_del_kobj; 5595 5596 #ifdef CONFIG_MEMCG 5597 if (is_root_cache(s)) { 5598 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj); 5599 if (!s->memcg_kset) { 5600 err = -ENOMEM; 5601 goto out_del_kobj; 5602 } 5603 } 5604 #endif 5605 5606 kobject_uevent(&s->kobj, KOBJ_ADD); 5607 if (!unmergeable) { 5608 /* Setup first alias */ 5609 sysfs_slab_alias(s, s->name); 5610 } 5611 out: 5612 if (!unmergeable) 5613 kfree(name); 5614 return err; 5615 out_del_kobj: 5616 kobject_del(&s->kobj); 5617 goto out; 5618 } 5619 5620 void sysfs_slab_remove(struct kmem_cache *s) 5621 { 5622 if (slab_state < FULL) 5623 /* 5624 * Sysfs has not been setup yet so no need to remove the 5625 * cache from sysfs. 5626 */ 5627 return; 5628 5629 #ifdef CONFIG_MEMCG 5630 kset_unregister(s->memcg_kset); 5631 #endif 5632 kobject_uevent(&s->kobj, KOBJ_REMOVE); 5633 kobject_del(&s->kobj); 5634 kobject_put(&s->kobj); 5635 } 5636 5637 /* 5638 * Need to buffer aliases during bootup until sysfs becomes 5639 * available lest we lose that information. 5640 */ 5641 struct saved_alias { 5642 struct kmem_cache *s; 5643 const char *name; 5644 struct saved_alias *next; 5645 }; 5646 5647 static struct saved_alias *alias_list; 5648 5649 static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 5650 { 5651 struct saved_alias *al; 5652 5653 if (slab_state == FULL) { 5654 /* 5655 * If we have a leftover link then remove it. 5656 */ 5657 sysfs_remove_link(&slab_kset->kobj, name); 5658 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 5659 } 5660 5661 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 5662 if (!al) 5663 return -ENOMEM; 5664 5665 al->s = s; 5666 al->name = name; 5667 al->next = alias_list; 5668 alias_list = al; 5669 return 0; 5670 } 5671 5672 static int __init slab_sysfs_init(void) 5673 { 5674 struct kmem_cache *s; 5675 int err; 5676 5677 mutex_lock(&slab_mutex); 5678 5679 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 5680 if (!slab_kset) { 5681 mutex_unlock(&slab_mutex); 5682 pr_err("Cannot register slab subsystem.\n"); 5683 return -ENOSYS; 5684 } 5685 5686 slab_state = FULL; 5687 5688 list_for_each_entry(s, &slab_caches, list) { 5689 err = sysfs_slab_add(s); 5690 if (err) 5691 pr_err("SLUB: Unable to add boot slab %s to sysfs\n", 5692 s->name); 5693 } 5694 5695 while (alias_list) { 5696 struct saved_alias *al = alias_list; 5697 5698 alias_list = alias_list->next; 5699 err = sysfs_slab_alias(al->s, al->name); 5700 if (err) 5701 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n", 5702 al->name); 5703 kfree(al); 5704 } 5705 5706 mutex_unlock(&slab_mutex); 5707 resiliency_test(); 5708 return 0; 5709 } 5710 5711 __initcall(slab_sysfs_init); 5712 #endif /* CONFIG_SYSFS */ 5713 5714 /* 5715 * The /proc/slabinfo ABI 5716 */ 5717 #ifdef CONFIG_SLABINFO 5718 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) 5719 { 5720 unsigned long nr_slabs = 0; 5721 unsigned long nr_objs = 0; 5722 unsigned long nr_free = 0; 5723 int node; 5724 struct kmem_cache_node *n; 5725 5726 for_each_kmem_cache_node(s, node, n) { 5727 nr_slabs += node_nr_slabs(n); 5728 nr_objs += node_nr_objs(n); 5729 nr_free += count_partial(n, count_free); 5730 } 5731 5732 sinfo->active_objs = nr_objs - nr_free; 5733 sinfo->num_objs = nr_objs; 5734 sinfo->active_slabs = nr_slabs; 5735 sinfo->num_slabs = nr_slabs; 5736 sinfo->objects_per_slab = oo_objects(s->oo); 5737 sinfo->cache_order = oo_order(s->oo); 5738 } 5739 5740 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) 5741 { 5742 } 5743 5744 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 5745 size_t count, loff_t *ppos) 5746 { 5747 return -EIO; 5748 } 5749 #endif /* CONFIG_SLABINFO */ 5750