1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * SLUB: A slab allocator that limits cache line use instead of queuing 4 * objects in per cpu and per node lists. 5 * 6 * The allocator synchronizes using per slab locks or atomic operatios 7 * and only uses a centralized lock to manage a pool of partial slabs. 8 * 9 * (C) 2007 SGI, Christoph Lameter 10 * (C) 2011 Linux Foundation, Christoph Lameter 11 */ 12 13 #include <linux/mm.h> 14 #include <linux/swap.h> /* struct reclaim_state */ 15 #include <linux/module.h> 16 #include <linux/bit_spinlock.h> 17 #include <linux/interrupt.h> 18 #include <linux/bitops.h> 19 #include <linux/slab.h> 20 #include "slab.h" 21 #include <linux/proc_fs.h> 22 #include <linux/seq_file.h> 23 #include <linux/kasan.h> 24 #include <linux/cpu.h> 25 #include <linux/cpuset.h> 26 #include <linux/mempolicy.h> 27 #include <linux/ctype.h> 28 #include <linux/debugobjects.h> 29 #include <linux/kallsyms.h> 30 #include <linux/memory.h> 31 #include <linux/math64.h> 32 #include <linux/fault-inject.h> 33 #include <linux/stacktrace.h> 34 #include <linux/prefetch.h> 35 #include <linux/memcontrol.h> 36 #include <linux/random.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: 55 * A. page->freelist -> List of object free in a page 56 * B. page->inuse -> Number of objects in use 57 * C. page->objects -> Number of objects in page 58 * D. 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 /* Poison object */ 195 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U) 196 /* Use cmpxchg_double */ 197 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U) 198 199 /* 200 * Tracking user of a slab. 201 */ 202 #define TRACK_ADDRS_COUNT 16 203 struct track { 204 unsigned long addr; /* Called from address */ 205 #ifdef CONFIG_STACKTRACE 206 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */ 207 #endif 208 int cpu; /* Was running on cpu */ 209 int pid; /* Pid context */ 210 unsigned long when; /* When did the operation occur */ 211 }; 212 213 enum track_item { TRACK_ALLOC, TRACK_FREE }; 214 215 #ifdef CONFIG_SYSFS 216 static int sysfs_slab_add(struct kmem_cache *); 217 static int sysfs_slab_alias(struct kmem_cache *, const char *); 218 static void memcg_propagate_slab_attrs(struct kmem_cache *s); 219 static void sysfs_slab_remove(struct kmem_cache *s); 220 #else 221 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } 222 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) 223 { return 0; } 224 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { } 225 static inline void sysfs_slab_remove(struct kmem_cache *s) { } 226 #endif 227 228 static inline void stat(const struct kmem_cache *s, enum stat_item si) 229 { 230 #ifdef CONFIG_SLUB_STATS 231 /* 232 * The rmw is racy on a preemptible kernel but this is acceptable, so 233 * avoid this_cpu_add()'s irq-disable overhead. 234 */ 235 raw_cpu_inc(s->cpu_slab->stat[si]); 236 #endif 237 } 238 239 /******************************************************************** 240 * Core slab cache functions 241 *******************************************************************/ 242 243 /* 244 * Returns freelist pointer (ptr). With hardening, this is obfuscated 245 * with an XOR of the address where the pointer is held and a per-cache 246 * random number. 247 */ 248 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr, 249 unsigned long ptr_addr) 250 { 251 #ifdef CONFIG_SLAB_FREELIST_HARDENED 252 return (void *)((unsigned long)ptr ^ s->random ^ ptr_addr); 253 #else 254 return ptr; 255 #endif 256 } 257 258 /* Returns the freelist pointer recorded at location ptr_addr. */ 259 static inline void *freelist_dereference(const struct kmem_cache *s, 260 void *ptr_addr) 261 { 262 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr), 263 (unsigned long)ptr_addr); 264 } 265 266 static inline void *get_freepointer(struct kmem_cache *s, void *object) 267 { 268 return freelist_dereference(s, object + s->offset); 269 } 270 271 static void prefetch_freepointer(const struct kmem_cache *s, void *object) 272 { 273 prefetch(object + s->offset); 274 } 275 276 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) 277 { 278 unsigned long freepointer_addr; 279 void *p; 280 281 if (!debug_pagealloc_enabled()) 282 return get_freepointer(s, object); 283 284 freepointer_addr = (unsigned long)object + s->offset; 285 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p)); 286 return freelist_ptr(s, p, freepointer_addr); 287 } 288 289 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) 290 { 291 unsigned long freeptr_addr = (unsigned long)object + s->offset; 292 293 #ifdef CONFIG_SLAB_FREELIST_HARDENED 294 BUG_ON(object == fp); /* naive detection of double free or corruption */ 295 #endif 296 297 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr); 298 } 299 300 /* Loop over all objects in a slab */ 301 #define for_each_object(__p, __s, __addr, __objects) \ 302 for (__p = fixup_red_left(__s, __addr); \ 303 __p < (__addr) + (__objects) * (__s)->size; \ 304 __p += (__s)->size) 305 306 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \ 307 for (__p = fixup_red_left(__s, __addr), __idx = 1; \ 308 __idx <= __objects; \ 309 __p += (__s)->size, __idx++) 310 311 /* Determine object index from a given position */ 312 static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr) 313 { 314 return (p - addr) / s->size; 315 } 316 317 static inline unsigned int order_objects(unsigned int order, unsigned int size) 318 { 319 return ((unsigned int)PAGE_SIZE << order) / size; 320 } 321 322 static inline struct kmem_cache_order_objects oo_make(unsigned int order, 323 unsigned int size) 324 { 325 struct kmem_cache_order_objects x = { 326 (order << OO_SHIFT) + order_objects(order, size) 327 }; 328 329 return x; 330 } 331 332 static inline unsigned int oo_order(struct kmem_cache_order_objects x) 333 { 334 return x.x >> OO_SHIFT; 335 } 336 337 static inline unsigned int oo_objects(struct kmem_cache_order_objects x) 338 { 339 return x.x & OO_MASK; 340 } 341 342 /* 343 * Per slab locking using the pagelock 344 */ 345 static __always_inline void slab_lock(struct page *page) 346 { 347 VM_BUG_ON_PAGE(PageTail(page), page); 348 bit_spin_lock(PG_locked, &page->flags); 349 } 350 351 static __always_inline void slab_unlock(struct page *page) 352 { 353 VM_BUG_ON_PAGE(PageTail(page), page); 354 __bit_spin_unlock(PG_locked, &page->flags); 355 } 356 357 /* Interrupts must be disabled (for the fallback code to work right) */ 358 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 359 void *freelist_old, unsigned long counters_old, 360 void *freelist_new, unsigned long counters_new, 361 const char *n) 362 { 363 VM_BUG_ON(!irqs_disabled()); 364 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 365 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 366 if (s->flags & __CMPXCHG_DOUBLE) { 367 if (cmpxchg_double(&page->freelist, &page->counters, 368 freelist_old, counters_old, 369 freelist_new, counters_new)) 370 return true; 371 } else 372 #endif 373 { 374 slab_lock(page); 375 if (page->freelist == freelist_old && 376 page->counters == counters_old) { 377 page->freelist = freelist_new; 378 page->counters = counters_new; 379 slab_unlock(page); 380 return true; 381 } 382 slab_unlock(page); 383 } 384 385 cpu_relax(); 386 stat(s, CMPXCHG_DOUBLE_FAIL); 387 388 #ifdef SLUB_DEBUG_CMPXCHG 389 pr_info("%s %s: cmpxchg double redo ", n, s->name); 390 #endif 391 392 return false; 393 } 394 395 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page, 396 void *freelist_old, unsigned long counters_old, 397 void *freelist_new, unsigned long counters_new, 398 const char *n) 399 { 400 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 401 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 402 if (s->flags & __CMPXCHG_DOUBLE) { 403 if (cmpxchg_double(&page->freelist, &page->counters, 404 freelist_old, counters_old, 405 freelist_new, counters_new)) 406 return true; 407 } else 408 #endif 409 { 410 unsigned long flags; 411 412 local_irq_save(flags); 413 slab_lock(page); 414 if (page->freelist == freelist_old && 415 page->counters == counters_old) { 416 page->freelist = freelist_new; 417 page->counters = counters_new; 418 slab_unlock(page); 419 local_irq_restore(flags); 420 return true; 421 } 422 slab_unlock(page); 423 local_irq_restore(flags); 424 } 425 426 cpu_relax(); 427 stat(s, CMPXCHG_DOUBLE_FAIL); 428 429 #ifdef SLUB_DEBUG_CMPXCHG 430 pr_info("%s %s: cmpxchg double redo ", n, s->name); 431 #endif 432 433 return false; 434 } 435 436 #ifdef CONFIG_SLUB_DEBUG 437 /* 438 * Determine a map of object in use on a page. 439 * 440 * Node listlock must be held to guarantee that the page does 441 * not vanish from under us. 442 */ 443 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map) 444 { 445 void *p; 446 void *addr = page_address(page); 447 448 for (p = page->freelist; p; p = get_freepointer(s, p)) 449 set_bit(slab_index(p, s, addr), map); 450 } 451 452 static inline unsigned int size_from_object(struct kmem_cache *s) 453 { 454 if (s->flags & SLAB_RED_ZONE) 455 return s->size - s->red_left_pad; 456 457 return s->size; 458 } 459 460 static inline void *restore_red_left(struct kmem_cache *s, void *p) 461 { 462 if (s->flags & SLAB_RED_ZONE) 463 p -= s->red_left_pad; 464 465 return p; 466 } 467 468 /* 469 * Debug settings: 470 */ 471 #if defined(CONFIG_SLUB_DEBUG_ON) 472 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; 473 #else 474 static slab_flags_t slub_debug; 475 #endif 476 477 static char *slub_debug_slabs; 478 static int disable_higher_order_debug; 479 480 /* 481 * slub is about to manipulate internal object metadata. This memory lies 482 * outside the range of the allocated object, so accessing it would normally 483 * be reported by kasan as a bounds error. metadata_access_enable() is used 484 * to tell kasan that these accesses are OK. 485 */ 486 static inline void metadata_access_enable(void) 487 { 488 kasan_disable_current(); 489 } 490 491 static inline void metadata_access_disable(void) 492 { 493 kasan_enable_current(); 494 } 495 496 /* 497 * Object debugging 498 */ 499 500 /* Verify that a pointer has an address that is valid within a slab page */ 501 static inline int check_valid_pointer(struct kmem_cache *s, 502 struct page *page, void *object) 503 { 504 void *base; 505 506 if (!object) 507 return 1; 508 509 base = page_address(page); 510 object = restore_red_left(s, object); 511 if (object < base || object >= base + page->objects * s->size || 512 (object - base) % s->size) { 513 return 0; 514 } 515 516 return 1; 517 } 518 519 static void print_section(char *level, char *text, u8 *addr, 520 unsigned int length) 521 { 522 metadata_access_enable(); 523 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr, 524 length, 1); 525 metadata_access_disable(); 526 } 527 528 static struct track *get_track(struct kmem_cache *s, void *object, 529 enum track_item alloc) 530 { 531 struct track *p; 532 533 if (s->offset) 534 p = object + s->offset + sizeof(void *); 535 else 536 p = object + s->inuse; 537 538 return p + alloc; 539 } 540 541 static void set_track(struct kmem_cache *s, void *object, 542 enum track_item alloc, unsigned long addr) 543 { 544 struct track *p = get_track(s, object, alloc); 545 546 if (addr) { 547 #ifdef CONFIG_STACKTRACE 548 struct stack_trace trace; 549 int i; 550 551 trace.nr_entries = 0; 552 trace.max_entries = TRACK_ADDRS_COUNT; 553 trace.entries = p->addrs; 554 trace.skip = 3; 555 metadata_access_enable(); 556 save_stack_trace(&trace); 557 metadata_access_disable(); 558 559 /* See rant in lockdep.c */ 560 if (trace.nr_entries != 0 && 561 trace.entries[trace.nr_entries - 1] == ULONG_MAX) 562 trace.nr_entries--; 563 564 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++) 565 p->addrs[i] = 0; 566 #endif 567 p->addr = addr; 568 p->cpu = smp_processor_id(); 569 p->pid = current->pid; 570 p->when = jiffies; 571 } else 572 memset(p, 0, sizeof(struct track)); 573 } 574 575 static void init_tracking(struct kmem_cache *s, void *object) 576 { 577 if (!(s->flags & SLAB_STORE_USER)) 578 return; 579 580 set_track(s, object, TRACK_FREE, 0UL); 581 set_track(s, object, TRACK_ALLOC, 0UL); 582 } 583 584 static void print_track(const char *s, struct track *t, unsigned long pr_time) 585 { 586 if (!t->addr) 587 return; 588 589 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n", 590 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid); 591 #ifdef CONFIG_STACKTRACE 592 { 593 int i; 594 for (i = 0; i < TRACK_ADDRS_COUNT; i++) 595 if (t->addrs[i]) 596 pr_err("\t%pS\n", (void *)t->addrs[i]); 597 else 598 break; 599 } 600 #endif 601 } 602 603 static void print_tracking(struct kmem_cache *s, void *object) 604 { 605 unsigned long pr_time = jiffies; 606 if (!(s->flags & SLAB_STORE_USER)) 607 return; 608 609 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time); 610 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time); 611 } 612 613 static void print_page_info(struct page *page) 614 { 615 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n", 616 page, page->objects, page->inuse, page->freelist, page->flags); 617 618 } 619 620 static void slab_bug(struct kmem_cache *s, char *fmt, ...) 621 { 622 struct va_format vaf; 623 va_list args; 624 625 va_start(args, fmt); 626 vaf.fmt = fmt; 627 vaf.va = &args; 628 pr_err("=============================================================================\n"); 629 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf); 630 pr_err("-----------------------------------------------------------------------------\n\n"); 631 632 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 633 va_end(args); 634 } 635 636 static void slab_fix(struct kmem_cache *s, char *fmt, ...) 637 { 638 struct va_format vaf; 639 va_list args; 640 641 va_start(args, fmt); 642 vaf.fmt = fmt; 643 vaf.va = &args; 644 pr_err("FIX %s: %pV\n", s->name, &vaf); 645 va_end(args); 646 } 647 648 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p) 649 { 650 unsigned int off; /* Offset of last byte */ 651 u8 *addr = page_address(page); 652 653 print_tracking(s, p); 654 655 print_page_info(page); 656 657 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n", 658 p, p - addr, get_freepointer(s, p)); 659 660 if (s->flags & SLAB_RED_ZONE) 661 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad, 662 s->red_left_pad); 663 else if (p > addr + 16) 664 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16); 665 666 print_section(KERN_ERR, "Object ", p, 667 min_t(unsigned int, s->object_size, PAGE_SIZE)); 668 if (s->flags & SLAB_RED_ZONE) 669 print_section(KERN_ERR, "Redzone ", p + s->object_size, 670 s->inuse - s->object_size); 671 672 if (s->offset) 673 off = s->offset + sizeof(void *); 674 else 675 off = s->inuse; 676 677 if (s->flags & SLAB_STORE_USER) 678 off += 2 * sizeof(struct track); 679 680 off += kasan_metadata_size(s); 681 682 if (off != size_from_object(s)) 683 /* Beginning of the filler is the free pointer */ 684 print_section(KERN_ERR, "Padding ", p + off, 685 size_from_object(s) - off); 686 687 dump_stack(); 688 } 689 690 void object_err(struct kmem_cache *s, struct page *page, 691 u8 *object, char *reason) 692 { 693 slab_bug(s, "%s", reason); 694 print_trailer(s, page, object); 695 } 696 697 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page, 698 const char *fmt, ...) 699 { 700 va_list args; 701 char buf[100]; 702 703 va_start(args, fmt); 704 vsnprintf(buf, sizeof(buf), fmt, args); 705 va_end(args); 706 slab_bug(s, "%s", buf); 707 print_page_info(page); 708 dump_stack(); 709 } 710 711 static void init_object(struct kmem_cache *s, void *object, u8 val) 712 { 713 u8 *p = object; 714 715 if (s->flags & SLAB_RED_ZONE) 716 memset(p - s->red_left_pad, val, s->red_left_pad); 717 718 if (s->flags & __OBJECT_POISON) { 719 memset(p, POISON_FREE, s->object_size - 1); 720 p[s->object_size - 1] = POISON_END; 721 } 722 723 if (s->flags & SLAB_RED_ZONE) 724 memset(p + s->object_size, val, s->inuse - s->object_size); 725 } 726 727 static void restore_bytes(struct kmem_cache *s, char *message, u8 data, 728 void *from, void *to) 729 { 730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data); 731 memset(from, data, to - from); 732 } 733 734 static int check_bytes_and_report(struct kmem_cache *s, struct page *page, 735 u8 *object, char *what, 736 u8 *start, unsigned int value, unsigned int bytes) 737 { 738 u8 *fault; 739 u8 *end; 740 741 metadata_access_enable(); 742 fault = memchr_inv(start, value, bytes); 743 metadata_access_disable(); 744 if (!fault) 745 return 1; 746 747 end = start + bytes; 748 while (end > fault && end[-1] == value) 749 end--; 750 751 slab_bug(s, "%s overwritten", what); 752 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n", 753 fault, end - 1, fault[0], value); 754 print_trailer(s, page, object); 755 756 restore_bytes(s, what, value, fault, end); 757 return 0; 758 } 759 760 /* 761 * Object layout: 762 * 763 * object address 764 * Bytes of the object to be managed. 765 * If the freepointer may overlay the object then the free 766 * pointer is the first word of the object. 767 * 768 * Poisoning uses 0x6b (POISON_FREE) and the last byte is 769 * 0xa5 (POISON_END) 770 * 771 * object + s->object_size 772 * Padding to reach word boundary. This is also used for Redzoning. 773 * Padding is extended by another word if Redzoning is enabled and 774 * object_size == inuse. 775 * 776 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with 777 * 0xcc (RED_ACTIVE) for objects in use. 778 * 779 * object + s->inuse 780 * Meta data starts here. 781 * 782 * A. Free pointer (if we cannot overwrite object on free) 783 * B. Tracking data for SLAB_STORE_USER 784 * C. Padding to reach required alignment boundary or at mininum 785 * one word if debugging is on to be able to detect writes 786 * before the word boundary. 787 * 788 * Padding is done using 0x5a (POISON_INUSE) 789 * 790 * object + s->size 791 * Nothing is used beyond s->size. 792 * 793 * If slabcaches are merged then the object_size and inuse boundaries are mostly 794 * ignored. And therefore no slab options that rely on these boundaries 795 * may be used with merged slabcaches. 796 */ 797 798 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p) 799 { 800 unsigned long off = s->inuse; /* The end of info */ 801 802 if (s->offset) 803 /* Freepointer is placed after the object. */ 804 off += sizeof(void *); 805 806 if (s->flags & SLAB_STORE_USER) 807 /* We also have user information there */ 808 off += 2 * sizeof(struct track); 809 810 off += kasan_metadata_size(s); 811 812 if (size_from_object(s) == off) 813 return 1; 814 815 return check_bytes_and_report(s, page, p, "Object padding", 816 p + off, POISON_INUSE, size_from_object(s) - off); 817 } 818 819 /* Check the pad bytes at the end of a slab page */ 820 static int slab_pad_check(struct kmem_cache *s, struct page *page) 821 { 822 u8 *start; 823 u8 *fault; 824 u8 *end; 825 u8 *pad; 826 int length; 827 int remainder; 828 829 if (!(s->flags & SLAB_POISON)) 830 return 1; 831 832 start = page_address(page); 833 length = PAGE_SIZE << compound_order(page); 834 end = start + length; 835 remainder = length % s->size; 836 if (!remainder) 837 return 1; 838 839 pad = end - remainder; 840 metadata_access_enable(); 841 fault = memchr_inv(pad, POISON_INUSE, remainder); 842 metadata_access_disable(); 843 if (!fault) 844 return 1; 845 while (end > fault && end[-1] == POISON_INUSE) 846 end--; 847 848 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1); 849 print_section(KERN_ERR, "Padding ", pad, remainder); 850 851 restore_bytes(s, "slab padding", POISON_INUSE, fault, end); 852 return 0; 853 } 854 855 static int check_object(struct kmem_cache *s, struct page *page, 856 void *object, u8 val) 857 { 858 u8 *p = object; 859 u8 *endobject = object + s->object_size; 860 861 if (s->flags & SLAB_RED_ZONE) { 862 if (!check_bytes_and_report(s, page, object, "Redzone", 863 object - s->red_left_pad, val, s->red_left_pad)) 864 return 0; 865 866 if (!check_bytes_and_report(s, page, object, "Redzone", 867 endobject, val, s->inuse - s->object_size)) 868 return 0; 869 } else { 870 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { 871 check_bytes_and_report(s, page, p, "Alignment padding", 872 endobject, POISON_INUSE, 873 s->inuse - s->object_size); 874 } 875 } 876 877 if (s->flags & SLAB_POISON) { 878 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && 879 (!check_bytes_and_report(s, page, p, "Poison", p, 880 POISON_FREE, s->object_size - 1) || 881 !check_bytes_and_report(s, page, p, "Poison", 882 p + s->object_size - 1, POISON_END, 1))) 883 return 0; 884 /* 885 * check_pad_bytes cleans up on its own. 886 */ 887 check_pad_bytes(s, page, p); 888 } 889 890 if (!s->offset && val == SLUB_RED_ACTIVE) 891 /* 892 * Object and freepointer overlap. Cannot check 893 * freepointer while object is allocated. 894 */ 895 return 1; 896 897 /* Check free pointer validity */ 898 if (!check_valid_pointer(s, page, get_freepointer(s, p))) { 899 object_err(s, page, p, "Freepointer corrupt"); 900 /* 901 * No choice but to zap it and thus lose the remainder 902 * of the free objects in this slab. May cause 903 * another error because the object count is now wrong. 904 */ 905 set_freepointer(s, p, NULL); 906 return 0; 907 } 908 return 1; 909 } 910 911 static int check_slab(struct kmem_cache *s, struct page *page) 912 { 913 int maxobj; 914 915 VM_BUG_ON(!irqs_disabled()); 916 917 if (!PageSlab(page)) { 918 slab_err(s, page, "Not a valid slab page"); 919 return 0; 920 } 921 922 maxobj = order_objects(compound_order(page), s->size); 923 if (page->objects > maxobj) { 924 slab_err(s, page, "objects %u > max %u", 925 page->objects, maxobj); 926 return 0; 927 } 928 if (page->inuse > page->objects) { 929 slab_err(s, page, "inuse %u > max %u", 930 page->inuse, page->objects); 931 return 0; 932 } 933 /* Slab_pad_check fixes things up after itself */ 934 slab_pad_check(s, page); 935 return 1; 936 } 937 938 /* 939 * Determine if a certain object on a page is on the freelist. Must hold the 940 * slab lock to guarantee that the chains are in a consistent state. 941 */ 942 static int on_freelist(struct kmem_cache *s, struct page *page, void *search) 943 { 944 int nr = 0; 945 void *fp; 946 void *object = NULL; 947 int max_objects; 948 949 fp = page->freelist; 950 while (fp && nr <= page->objects) { 951 if (fp == search) 952 return 1; 953 if (!check_valid_pointer(s, page, fp)) { 954 if (object) { 955 object_err(s, page, object, 956 "Freechain corrupt"); 957 set_freepointer(s, object, NULL); 958 } else { 959 slab_err(s, page, "Freepointer corrupt"); 960 page->freelist = NULL; 961 page->inuse = page->objects; 962 slab_fix(s, "Freelist cleared"); 963 return 0; 964 } 965 break; 966 } 967 object = fp; 968 fp = get_freepointer(s, object); 969 nr++; 970 } 971 972 max_objects = order_objects(compound_order(page), s->size); 973 if (max_objects > MAX_OBJS_PER_PAGE) 974 max_objects = MAX_OBJS_PER_PAGE; 975 976 if (page->objects != max_objects) { 977 slab_err(s, page, "Wrong number of objects. Found %d but should be %d", 978 page->objects, max_objects); 979 page->objects = max_objects; 980 slab_fix(s, "Number of objects adjusted."); 981 } 982 if (page->inuse != page->objects - nr) { 983 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d", 984 page->inuse, page->objects - nr); 985 page->inuse = page->objects - nr; 986 slab_fix(s, "Object count adjusted."); 987 } 988 return search == NULL; 989 } 990 991 static void trace(struct kmem_cache *s, struct page *page, void *object, 992 int alloc) 993 { 994 if (s->flags & SLAB_TRACE) { 995 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n", 996 s->name, 997 alloc ? "alloc" : "free", 998 object, page->inuse, 999 page->freelist); 1000 1001 if (!alloc) 1002 print_section(KERN_INFO, "Object ", (void *)object, 1003 s->object_size); 1004 1005 dump_stack(); 1006 } 1007 } 1008 1009 /* 1010 * Tracking of fully allocated slabs for debugging purposes. 1011 */ 1012 static void add_full(struct kmem_cache *s, 1013 struct kmem_cache_node *n, struct page *page) 1014 { 1015 if (!(s->flags & SLAB_STORE_USER)) 1016 return; 1017 1018 lockdep_assert_held(&n->list_lock); 1019 list_add(&page->lru, &n->full); 1020 } 1021 1022 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page) 1023 { 1024 if (!(s->flags & SLAB_STORE_USER)) 1025 return; 1026 1027 lockdep_assert_held(&n->list_lock); 1028 list_del(&page->lru); 1029 } 1030 1031 /* Tracking of the number of slabs for debugging purposes */ 1032 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1033 { 1034 struct kmem_cache_node *n = get_node(s, node); 1035 1036 return atomic_long_read(&n->nr_slabs); 1037 } 1038 1039 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1040 { 1041 return atomic_long_read(&n->nr_slabs); 1042 } 1043 1044 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) 1045 { 1046 struct kmem_cache_node *n = get_node(s, node); 1047 1048 /* 1049 * May be called early in order to allocate a slab for the 1050 * kmem_cache_node structure. Solve the chicken-egg 1051 * dilemma by deferring the increment of the count during 1052 * bootstrap (see early_kmem_cache_node_alloc). 1053 */ 1054 if (likely(n)) { 1055 atomic_long_inc(&n->nr_slabs); 1056 atomic_long_add(objects, &n->total_objects); 1057 } 1058 } 1059 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) 1060 { 1061 struct kmem_cache_node *n = get_node(s, node); 1062 1063 atomic_long_dec(&n->nr_slabs); 1064 atomic_long_sub(objects, &n->total_objects); 1065 } 1066 1067 /* Object debug checks for alloc/free paths */ 1068 static void setup_object_debug(struct kmem_cache *s, struct page *page, 1069 void *object) 1070 { 1071 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))) 1072 return; 1073 1074 init_object(s, object, SLUB_RED_INACTIVE); 1075 init_tracking(s, object); 1076 } 1077 1078 static inline int alloc_consistency_checks(struct kmem_cache *s, 1079 struct page *page, 1080 void *object, unsigned long addr) 1081 { 1082 if (!check_slab(s, page)) 1083 return 0; 1084 1085 if (!check_valid_pointer(s, page, object)) { 1086 object_err(s, page, object, "Freelist Pointer check fails"); 1087 return 0; 1088 } 1089 1090 if (!check_object(s, page, object, SLUB_RED_INACTIVE)) 1091 return 0; 1092 1093 return 1; 1094 } 1095 1096 static noinline int alloc_debug_processing(struct kmem_cache *s, 1097 struct page *page, 1098 void *object, unsigned long addr) 1099 { 1100 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1101 if (!alloc_consistency_checks(s, page, object, addr)) 1102 goto bad; 1103 } 1104 1105 /* Success perform special debug activities for allocs */ 1106 if (s->flags & SLAB_STORE_USER) 1107 set_track(s, object, TRACK_ALLOC, addr); 1108 trace(s, page, object, 1); 1109 init_object(s, object, SLUB_RED_ACTIVE); 1110 return 1; 1111 1112 bad: 1113 if (PageSlab(page)) { 1114 /* 1115 * If this is a slab page then lets do the best we can 1116 * to avoid issues in the future. Marking all objects 1117 * as used avoids touching the remaining objects. 1118 */ 1119 slab_fix(s, "Marking all objects used"); 1120 page->inuse = page->objects; 1121 page->freelist = NULL; 1122 } 1123 return 0; 1124 } 1125 1126 static inline int free_consistency_checks(struct kmem_cache *s, 1127 struct page *page, void *object, unsigned long addr) 1128 { 1129 if (!check_valid_pointer(s, page, object)) { 1130 slab_err(s, page, "Invalid object pointer 0x%p", object); 1131 return 0; 1132 } 1133 1134 if (on_freelist(s, page, object)) { 1135 object_err(s, page, object, "Object already free"); 1136 return 0; 1137 } 1138 1139 if (!check_object(s, page, object, SLUB_RED_ACTIVE)) 1140 return 0; 1141 1142 if (unlikely(s != page->slab_cache)) { 1143 if (!PageSlab(page)) { 1144 slab_err(s, page, "Attempt to free object(0x%p) outside of slab", 1145 object); 1146 } else if (!page->slab_cache) { 1147 pr_err("SLUB <none>: no slab for object 0x%p.\n", 1148 object); 1149 dump_stack(); 1150 } else 1151 object_err(s, page, object, 1152 "page slab pointer corrupt."); 1153 return 0; 1154 } 1155 return 1; 1156 } 1157 1158 /* Supports checking bulk free of a constructed freelist */ 1159 static noinline int free_debug_processing( 1160 struct kmem_cache *s, struct page *page, 1161 void *head, void *tail, int bulk_cnt, 1162 unsigned long addr) 1163 { 1164 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 1165 void *object = head; 1166 int cnt = 0; 1167 unsigned long uninitialized_var(flags); 1168 int ret = 0; 1169 1170 spin_lock_irqsave(&n->list_lock, flags); 1171 slab_lock(page); 1172 1173 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1174 if (!check_slab(s, page)) 1175 goto out; 1176 } 1177 1178 next_object: 1179 cnt++; 1180 1181 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1182 if (!free_consistency_checks(s, page, object, addr)) 1183 goto out; 1184 } 1185 1186 if (s->flags & SLAB_STORE_USER) 1187 set_track(s, object, TRACK_FREE, addr); 1188 trace(s, page, object, 0); 1189 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ 1190 init_object(s, object, SLUB_RED_INACTIVE); 1191 1192 /* Reached end of constructed freelist yet? */ 1193 if (object != tail) { 1194 object = get_freepointer(s, object); 1195 goto next_object; 1196 } 1197 ret = 1; 1198 1199 out: 1200 if (cnt != bulk_cnt) 1201 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n", 1202 bulk_cnt, cnt); 1203 1204 slab_unlock(page); 1205 spin_unlock_irqrestore(&n->list_lock, flags); 1206 if (!ret) 1207 slab_fix(s, "Object at 0x%p not freed", object); 1208 return ret; 1209 } 1210 1211 static int __init setup_slub_debug(char *str) 1212 { 1213 slub_debug = DEBUG_DEFAULT_FLAGS; 1214 if (*str++ != '=' || !*str) 1215 /* 1216 * No options specified. Switch on full debugging. 1217 */ 1218 goto out; 1219 1220 if (*str == ',') 1221 /* 1222 * No options but restriction on slabs. This means full 1223 * debugging for slabs matching a pattern. 1224 */ 1225 goto check_slabs; 1226 1227 slub_debug = 0; 1228 if (*str == '-') 1229 /* 1230 * Switch off all debugging measures. 1231 */ 1232 goto out; 1233 1234 /* 1235 * Determine which debug features should be switched on 1236 */ 1237 for (; *str && *str != ','; str++) { 1238 switch (tolower(*str)) { 1239 case 'f': 1240 slub_debug |= SLAB_CONSISTENCY_CHECKS; 1241 break; 1242 case 'z': 1243 slub_debug |= SLAB_RED_ZONE; 1244 break; 1245 case 'p': 1246 slub_debug |= SLAB_POISON; 1247 break; 1248 case 'u': 1249 slub_debug |= SLAB_STORE_USER; 1250 break; 1251 case 't': 1252 slub_debug |= SLAB_TRACE; 1253 break; 1254 case 'a': 1255 slub_debug |= SLAB_FAILSLAB; 1256 break; 1257 case 'o': 1258 /* 1259 * Avoid enabling debugging on caches if its minimum 1260 * order would increase as a result. 1261 */ 1262 disable_higher_order_debug = 1; 1263 break; 1264 default: 1265 pr_err("slub_debug option '%c' unknown. skipped\n", 1266 *str); 1267 } 1268 } 1269 1270 check_slabs: 1271 if (*str == ',') 1272 slub_debug_slabs = str + 1; 1273 out: 1274 return 1; 1275 } 1276 1277 __setup("slub_debug", setup_slub_debug); 1278 1279 /* 1280 * kmem_cache_flags - apply debugging options to the cache 1281 * @object_size: the size of an object without meta data 1282 * @flags: flags to set 1283 * @name: name of the cache 1284 * @ctor: constructor function 1285 * 1286 * Debug option(s) are applied to @flags. In addition to the debug 1287 * option(s), if a slab name (or multiple) is specified i.e. 1288 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ... 1289 * then only the select slabs will receive the debug option(s). 1290 */ 1291 slab_flags_t kmem_cache_flags(unsigned int object_size, 1292 slab_flags_t flags, const char *name, 1293 void (*ctor)(void *)) 1294 { 1295 char *iter; 1296 size_t len; 1297 1298 /* If slub_debug = 0, it folds into the if conditional. */ 1299 if (!slub_debug_slabs) 1300 return flags | slub_debug; 1301 1302 len = strlen(name); 1303 iter = slub_debug_slabs; 1304 while (*iter) { 1305 char *end, *glob; 1306 size_t cmplen; 1307 1308 end = strchr(iter, ','); 1309 if (!end) 1310 end = iter + strlen(iter); 1311 1312 glob = strnchr(iter, end - iter, '*'); 1313 if (glob) 1314 cmplen = glob - iter; 1315 else 1316 cmplen = max_t(size_t, len, (end - iter)); 1317 1318 if (!strncmp(name, iter, cmplen)) { 1319 flags |= slub_debug; 1320 break; 1321 } 1322 1323 if (!*end) 1324 break; 1325 iter = end + 1; 1326 } 1327 1328 return flags; 1329 } 1330 #else /* !CONFIG_SLUB_DEBUG */ 1331 static inline void setup_object_debug(struct kmem_cache *s, 1332 struct page *page, void *object) {} 1333 1334 static inline int alloc_debug_processing(struct kmem_cache *s, 1335 struct page *page, void *object, unsigned long addr) { return 0; } 1336 1337 static inline int free_debug_processing( 1338 struct kmem_cache *s, struct page *page, 1339 void *head, void *tail, int bulk_cnt, 1340 unsigned long addr) { return 0; } 1341 1342 static inline int slab_pad_check(struct kmem_cache *s, struct page *page) 1343 { return 1; } 1344 static inline int check_object(struct kmem_cache *s, struct page *page, 1345 void *object, u8 val) { return 1; } 1346 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, 1347 struct page *page) {} 1348 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, 1349 struct page *page) {} 1350 slab_flags_t kmem_cache_flags(unsigned int object_size, 1351 slab_flags_t flags, const char *name, 1352 void (*ctor)(void *)) 1353 { 1354 return flags; 1355 } 1356 #define slub_debug 0 1357 1358 #define disable_higher_order_debug 0 1359 1360 static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1361 { return 0; } 1362 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1363 { return 0; } 1364 static inline void inc_slabs_node(struct kmem_cache *s, int node, 1365 int objects) {} 1366 static inline void dec_slabs_node(struct kmem_cache *s, int node, 1367 int objects) {} 1368 1369 #endif /* CONFIG_SLUB_DEBUG */ 1370 1371 /* 1372 * Hooks for other subsystems that check memory allocations. In a typical 1373 * production configuration these hooks all should produce no code at all. 1374 */ 1375 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags) 1376 { 1377 kmemleak_alloc(ptr, size, 1, flags); 1378 return kasan_kmalloc_large(ptr, size, flags); 1379 } 1380 1381 static __always_inline void kfree_hook(void *x) 1382 { 1383 kmemleak_free(x); 1384 kasan_kfree_large(x, _RET_IP_); 1385 } 1386 1387 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x) 1388 { 1389 kmemleak_free_recursive(x, s->flags); 1390 1391 /* 1392 * Trouble is that we may no longer disable interrupts in the fast path 1393 * So in order to make the debug calls that expect irqs to be 1394 * disabled we need to disable interrupts temporarily. 1395 */ 1396 #ifdef CONFIG_LOCKDEP 1397 { 1398 unsigned long flags; 1399 1400 local_irq_save(flags); 1401 debug_check_no_locks_freed(x, s->object_size); 1402 local_irq_restore(flags); 1403 } 1404 #endif 1405 if (!(s->flags & SLAB_DEBUG_OBJECTS)) 1406 debug_check_no_obj_freed(x, s->object_size); 1407 1408 /* KASAN might put x into memory quarantine, delaying its reuse */ 1409 return kasan_slab_free(s, x, _RET_IP_); 1410 } 1411 1412 static inline bool slab_free_freelist_hook(struct kmem_cache *s, 1413 void **head, void **tail) 1414 { 1415 /* 1416 * Compiler cannot detect this function can be removed if slab_free_hook() 1417 * evaluates to nothing. Thus, catch all relevant config debug options here. 1418 */ 1419 #if defined(CONFIG_LOCKDEP) || \ 1420 defined(CONFIG_DEBUG_KMEMLEAK) || \ 1421 defined(CONFIG_DEBUG_OBJECTS_FREE) || \ 1422 defined(CONFIG_KASAN) 1423 1424 void *object; 1425 void *next = *head; 1426 void *old_tail = *tail ? *tail : *head; 1427 1428 /* Head and tail of the reconstructed freelist */ 1429 *head = NULL; 1430 *tail = NULL; 1431 1432 do { 1433 object = next; 1434 next = get_freepointer(s, object); 1435 /* If object's reuse doesn't have to be delayed */ 1436 if (!slab_free_hook(s, object)) { 1437 /* Move object to the new freelist */ 1438 set_freepointer(s, object, *head); 1439 *head = object; 1440 if (!*tail) 1441 *tail = object; 1442 } 1443 } while (object != old_tail); 1444 1445 if (*head == *tail) 1446 *tail = NULL; 1447 1448 return *head != NULL; 1449 #else 1450 return true; 1451 #endif 1452 } 1453 1454 static void *setup_object(struct kmem_cache *s, struct page *page, 1455 void *object) 1456 { 1457 setup_object_debug(s, page, object); 1458 object = kasan_init_slab_obj(s, object); 1459 if (unlikely(s->ctor)) { 1460 kasan_unpoison_object_data(s, object); 1461 s->ctor(object); 1462 kasan_poison_object_data(s, object); 1463 } 1464 return object; 1465 } 1466 1467 /* 1468 * Slab allocation and freeing 1469 */ 1470 static inline struct page *alloc_slab_page(struct kmem_cache *s, 1471 gfp_t flags, int node, struct kmem_cache_order_objects oo) 1472 { 1473 struct page *page; 1474 unsigned int order = oo_order(oo); 1475 1476 if (node == NUMA_NO_NODE) 1477 page = alloc_pages(flags, order); 1478 else 1479 page = __alloc_pages_node(node, flags, order); 1480 1481 if (page && memcg_charge_slab(page, flags, order, s)) { 1482 __free_pages(page, order); 1483 page = NULL; 1484 } 1485 1486 return page; 1487 } 1488 1489 #ifdef CONFIG_SLAB_FREELIST_RANDOM 1490 /* Pre-initialize the random sequence cache */ 1491 static int init_cache_random_seq(struct kmem_cache *s) 1492 { 1493 unsigned int count = oo_objects(s->oo); 1494 int err; 1495 1496 /* Bailout if already initialised */ 1497 if (s->random_seq) 1498 return 0; 1499 1500 err = cache_random_seq_create(s, count, GFP_KERNEL); 1501 if (err) { 1502 pr_err("SLUB: Unable to initialize free list for %s\n", 1503 s->name); 1504 return err; 1505 } 1506 1507 /* Transform to an offset on the set of pages */ 1508 if (s->random_seq) { 1509 unsigned int i; 1510 1511 for (i = 0; i < count; i++) 1512 s->random_seq[i] *= s->size; 1513 } 1514 return 0; 1515 } 1516 1517 /* Initialize each random sequence freelist per cache */ 1518 static void __init init_freelist_randomization(void) 1519 { 1520 struct kmem_cache *s; 1521 1522 mutex_lock(&slab_mutex); 1523 1524 list_for_each_entry(s, &slab_caches, list) 1525 init_cache_random_seq(s); 1526 1527 mutex_unlock(&slab_mutex); 1528 } 1529 1530 /* Get the next entry on the pre-computed freelist randomized */ 1531 static void *next_freelist_entry(struct kmem_cache *s, struct page *page, 1532 unsigned long *pos, void *start, 1533 unsigned long page_limit, 1534 unsigned long freelist_count) 1535 { 1536 unsigned int idx; 1537 1538 /* 1539 * If the target page allocation failed, the number of objects on the 1540 * page might be smaller than the usual size defined by the cache. 1541 */ 1542 do { 1543 idx = s->random_seq[*pos]; 1544 *pos += 1; 1545 if (*pos >= freelist_count) 1546 *pos = 0; 1547 } while (unlikely(idx >= page_limit)); 1548 1549 return (char *)start + idx; 1550 } 1551 1552 /* Shuffle the single linked freelist based on a random pre-computed sequence */ 1553 static bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1554 { 1555 void *start; 1556 void *cur; 1557 void *next; 1558 unsigned long idx, pos, page_limit, freelist_count; 1559 1560 if (page->objects < 2 || !s->random_seq) 1561 return false; 1562 1563 freelist_count = oo_objects(s->oo); 1564 pos = get_random_int() % freelist_count; 1565 1566 page_limit = page->objects * s->size; 1567 start = fixup_red_left(s, page_address(page)); 1568 1569 /* First entry is used as the base of the freelist */ 1570 cur = next_freelist_entry(s, page, &pos, start, page_limit, 1571 freelist_count); 1572 cur = setup_object(s, page, cur); 1573 page->freelist = cur; 1574 1575 for (idx = 1; idx < page->objects; idx++) { 1576 next = next_freelist_entry(s, page, &pos, start, page_limit, 1577 freelist_count); 1578 next = setup_object(s, page, next); 1579 set_freepointer(s, cur, next); 1580 cur = next; 1581 } 1582 set_freepointer(s, cur, NULL); 1583 1584 return true; 1585 } 1586 #else 1587 static inline int init_cache_random_seq(struct kmem_cache *s) 1588 { 1589 return 0; 1590 } 1591 static inline void init_freelist_randomization(void) { } 1592 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page) 1593 { 1594 return false; 1595 } 1596 #endif /* CONFIG_SLAB_FREELIST_RANDOM */ 1597 1598 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) 1599 { 1600 struct page *page; 1601 struct kmem_cache_order_objects oo = s->oo; 1602 gfp_t alloc_gfp; 1603 void *start, *p, *next; 1604 int idx, order; 1605 bool shuffle; 1606 1607 flags &= gfp_allowed_mask; 1608 1609 if (gfpflags_allow_blocking(flags)) 1610 local_irq_enable(); 1611 1612 flags |= s->allocflags; 1613 1614 /* 1615 * Let the initial higher-order allocation fail under memory pressure 1616 * so we fall-back to the minimum order allocation. 1617 */ 1618 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; 1619 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min)) 1620 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL); 1621 1622 page = alloc_slab_page(s, alloc_gfp, node, oo); 1623 if (unlikely(!page)) { 1624 oo = s->min; 1625 alloc_gfp = flags; 1626 /* 1627 * Allocation may have failed due to fragmentation. 1628 * Try a lower order alloc if possible 1629 */ 1630 page = alloc_slab_page(s, alloc_gfp, node, oo); 1631 if (unlikely(!page)) 1632 goto out; 1633 stat(s, ORDER_FALLBACK); 1634 } 1635 1636 page->objects = oo_objects(oo); 1637 1638 order = compound_order(page); 1639 page->slab_cache = s; 1640 __SetPageSlab(page); 1641 if (page_is_pfmemalloc(page)) 1642 SetPageSlabPfmemalloc(page); 1643 1644 start = page_address(page); 1645 1646 if (unlikely(s->flags & SLAB_POISON)) 1647 memset(start, POISON_INUSE, PAGE_SIZE << order); 1648 1649 kasan_poison_slab(page); 1650 1651 shuffle = shuffle_freelist(s, page); 1652 1653 if (!shuffle) { 1654 for_each_object_idx(p, idx, s, start, page->objects) { 1655 if (likely(idx < page->objects)) { 1656 next = p + s->size; 1657 next = setup_object(s, page, next); 1658 set_freepointer(s, p, next); 1659 } else 1660 set_freepointer(s, p, NULL); 1661 } 1662 start = fixup_red_left(s, start); 1663 start = setup_object(s, page, start); 1664 page->freelist = start; 1665 } 1666 1667 page->inuse = page->objects; 1668 page->frozen = 1; 1669 1670 out: 1671 if (gfpflags_allow_blocking(flags)) 1672 local_irq_disable(); 1673 if (!page) 1674 return NULL; 1675 1676 mod_lruvec_page_state(page, 1677 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1678 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1679 1 << oo_order(oo)); 1680 1681 inc_slabs_node(s, page_to_nid(page), page->objects); 1682 1683 return page; 1684 } 1685 1686 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node) 1687 { 1688 if (unlikely(flags & GFP_SLAB_BUG_MASK)) { 1689 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK; 1690 flags &= ~GFP_SLAB_BUG_MASK; 1691 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n", 1692 invalid_mask, &invalid_mask, flags, &flags); 1693 dump_stack(); 1694 } 1695 1696 return allocate_slab(s, 1697 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); 1698 } 1699 1700 static void __free_slab(struct kmem_cache *s, struct page *page) 1701 { 1702 int order = compound_order(page); 1703 int pages = 1 << order; 1704 1705 if (s->flags & SLAB_CONSISTENCY_CHECKS) { 1706 void *p; 1707 1708 slab_pad_check(s, page); 1709 for_each_object(p, s, page_address(page), 1710 page->objects) 1711 check_object(s, page, p, SLUB_RED_INACTIVE); 1712 } 1713 1714 mod_lruvec_page_state(page, 1715 (s->flags & SLAB_RECLAIM_ACCOUNT) ? 1716 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE, 1717 -pages); 1718 1719 __ClearPageSlabPfmemalloc(page); 1720 __ClearPageSlab(page); 1721 1722 page->mapping = NULL; 1723 if (current->reclaim_state) 1724 current->reclaim_state->reclaimed_slab += pages; 1725 memcg_uncharge_slab(page, order, s); 1726 __free_pages(page, order); 1727 } 1728 1729 static void rcu_free_slab(struct rcu_head *h) 1730 { 1731 struct page *page = container_of(h, struct page, rcu_head); 1732 1733 __free_slab(page->slab_cache, page); 1734 } 1735 1736 static void free_slab(struct kmem_cache *s, struct page *page) 1737 { 1738 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) { 1739 call_rcu(&page->rcu_head, rcu_free_slab); 1740 } else 1741 __free_slab(s, page); 1742 } 1743 1744 static void discard_slab(struct kmem_cache *s, struct page *page) 1745 { 1746 dec_slabs_node(s, page_to_nid(page), page->objects); 1747 free_slab(s, page); 1748 } 1749 1750 /* 1751 * Management of partially allocated slabs. 1752 */ 1753 static inline void 1754 __add_partial(struct kmem_cache_node *n, struct page *page, int tail) 1755 { 1756 n->nr_partial++; 1757 if (tail == DEACTIVATE_TO_TAIL) 1758 list_add_tail(&page->lru, &n->partial); 1759 else 1760 list_add(&page->lru, &n->partial); 1761 } 1762 1763 static inline void add_partial(struct kmem_cache_node *n, 1764 struct page *page, int tail) 1765 { 1766 lockdep_assert_held(&n->list_lock); 1767 __add_partial(n, page, tail); 1768 } 1769 1770 static inline void remove_partial(struct kmem_cache_node *n, 1771 struct page *page) 1772 { 1773 lockdep_assert_held(&n->list_lock); 1774 list_del(&page->lru); 1775 n->nr_partial--; 1776 } 1777 1778 /* 1779 * Remove slab from the partial list, freeze it and 1780 * return the pointer to the freelist. 1781 * 1782 * Returns a list of objects or NULL if it fails. 1783 */ 1784 static inline void *acquire_slab(struct kmem_cache *s, 1785 struct kmem_cache_node *n, struct page *page, 1786 int mode, int *objects) 1787 { 1788 void *freelist; 1789 unsigned long counters; 1790 struct page new; 1791 1792 lockdep_assert_held(&n->list_lock); 1793 1794 /* 1795 * Zap the freelist and set the frozen bit. 1796 * The old freelist is the list of objects for the 1797 * per cpu allocation list. 1798 */ 1799 freelist = page->freelist; 1800 counters = page->counters; 1801 new.counters = counters; 1802 *objects = new.objects - new.inuse; 1803 if (mode) { 1804 new.inuse = page->objects; 1805 new.freelist = NULL; 1806 } else { 1807 new.freelist = freelist; 1808 } 1809 1810 VM_BUG_ON(new.frozen); 1811 new.frozen = 1; 1812 1813 if (!__cmpxchg_double_slab(s, page, 1814 freelist, counters, 1815 new.freelist, new.counters, 1816 "acquire_slab")) 1817 return NULL; 1818 1819 remove_partial(n, page); 1820 WARN_ON(!freelist); 1821 return freelist; 1822 } 1823 1824 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain); 1825 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags); 1826 1827 /* 1828 * Try to allocate a partial slab from a specific node. 1829 */ 1830 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, 1831 struct kmem_cache_cpu *c, gfp_t flags) 1832 { 1833 struct page *page, *page2; 1834 void *object = NULL; 1835 unsigned int available = 0; 1836 int objects; 1837 1838 /* 1839 * Racy check. If we mistakenly see no partial slabs then we 1840 * just allocate an empty slab. If we mistakenly try to get a 1841 * partial slab and there is none available then get_partials() 1842 * will return NULL. 1843 */ 1844 if (!n || !n->nr_partial) 1845 return NULL; 1846 1847 spin_lock(&n->list_lock); 1848 list_for_each_entry_safe(page, page2, &n->partial, lru) { 1849 void *t; 1850 1851 if (!pfmemalloc_match(page, flags)) 1852 continue; 1853 1854 t = acquire_slab(s, n, page, object == NULL, &objects); 1855 if (!t) 1856 break; 1857 1858 available += objects; 1859 if (!object) { 1860 c->page = page; 1861 stat(s, ALLOC_FROM_PARTIAL); 1862 object = t; 1863 } else { 1864 put_cpu_partial(s, page, 0); 1865 stat(s, CPU_PARTIAL_NODE); 1866 } 1867 if (!kmem_cache_has_cpu_partial(s) 1868 || available > slub_cpu_partial(s) / 2) 1869 break; 1870 1871 } 1872 spin_unlock(&n->list_lock); 1873 return object; 1874 } 1875 1876 /* 1877 * Get a page from somewhere. Search in increasing NUMA distances. 1878 */ 1879 static void *get_any_partial(struct kmem_cache *s, gfp_t flags, 1880 struct kmem_cache_cpu *c) 1881 { 1882 #ifdef CONFIG_NUMA 1883 struct zonelist *zonelist; 1884 struct zoneref *z; 1885 struct zone *zone; 1886 enum zone_type high_zoneidx = gfp_zone(flags); 1887 void *object; 1888 unsigned int cpuset_mems_cookie; 1889 1890 /* 1891 * The defrag ratio allows a configuration of the tradeoffs between 1892 * inter node defragmentation and node local allocations. A lower 1893 * defrag_ratio increases the tendency to do local allocations 1894 * instead of attempting to obtain partial slabs from other nodes. 1895 * 1896 * If the defrag_ratio is set to 0 then kmalloc() always 1897 * returns node local objects. If the ratio is higher then kmalloc() 1898 * may return off node objects because partial slabs are obtained 1899 * from other nodes and filled up. 1900 * 1901 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 1902 * (which makes defrag_ratio = 1000) then every (well almost) 1903 * allocation will first attempt to defrag slab caches on other nodes. 1904 * This means scanning over all nodes to look for partial slabs which 1905 * may be expensive if we do it every time we are trying to find a slab 1906 * with available objects. 1907 */ 1908 if (!s->remote_node_defrag_ratio || 1909 get_cycles() % 1024 > s->remote_node_defrag_ratio) 1910 return NULL; 1911 1912 do { 1913 cpuset_mems_cookie = read_mems_allowed_begin(); 1914 zonelist = node_zonelist(mempolicy_slab_node(), flags); 1915 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { 1916 struct kmem_cache_node *n; 1917 1918 n = get_node(s, zone_to_nid(zone)); 1919 1920 if (n && cpuset_zone_allowed(zone, flags) && 1921 n->nr_partial > s->min_partial) { 1922 object = get_partial_node(s, n, c, flags); 1923 if (object) { 1924 /* 1925 * Don't check read_mems_allowed_retry() 1926 * here - if mems_allowed was updated in 1927 * parallel, that was a harmless race 1928 * between allocation and the cpuset 1929 * update 1930 */ 1931 return object; 1932 } 1933 } 1934 } 1935 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 1936 #endif 1937 return NULL; 1938 } 1939 1940 /* 1941 * Get a partial page, lock it and return it. 1942 */ 1943 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node, 1944 struct kmem_cache_cpu *c) 1945 { 1946 void *object; 1947 int searchnode = node; 1948 1949 if (node == NUMA_NO_NODE) 1950 searchnode = numa_mem_id(); 1951 else if (!node_present_pages(node)) 1952 searchnode = node_to_mem_node(node); 1953 1954 object = get_partial_node(s, get_node(s, searchnode), c, flags); 1955 if (object || node != NUMA_NO_NODE) 1956 return object; 1957 1958 return get_any_partial(s, flags, c); 1959 } 1960 1961 #ifdef CONFIG_PREEMPT 1962 /* 1963 * Calculate the next globally unique transaction for disambiguiation 1964 * during cmpxchg. The transactions start with the cpu number and are then 1965 * incremented by CONFIG_NR_CPUS. 1966 */ 1967 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) 1968 #else 1969 /* 1970 * No preemption supported therefore also no need to check for 1971 * different cpus. 1972 */ 1973 #define TID_STEP 1 1974 #endif 1975 1976 static inline unsigned long next_tid(unsigned long tid) 1977 { 1978 return tid + TID_STEP; 1979 } 1980 1981 static inline unsigned int tid_to_cpu(unsigned long tid) 1982 { 1983 return tid % TID_STEP; 1984 } 1985 1986 static inline unsigned long tid_to_event(unsigned long tid) 1987 { 1988 return tid / TID_STEP; 1989 } 1990 1991 static inline unsigned int init_tid(int cpu) 1992 { 1993 return cpu; 1994 } 1995 1996 static inline void note_cmpxchg_failure(const char *n, 1997 const struct kmem_cache *s, unsigned long tid) 1998 { 1999 #ifdef SLUB_DEBUG_CMPXCHG 2000 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); 2001 2002 pr_info("%s %s: cmpxchg redo ", n, s->name); 2003 2004 #ifdef CONFIG_PREEMPT 2005 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) 2006 pr_warn("due to cpu change %d -> %d\n", 2007 tid_to_cpu(tid), tid_to_cpu(actual_tid)); 2008 else 2009 #endif 2010 if (tid_to_event(tid) != tid_to_event(actual_tid)) 2011 pr_warn("due to cpu running other code. Event %ld->%ld\n", 2012 tid_to_event(tid), tid_to_event(actual_tid)); 2013 else 2014 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n", 2015 actual_tid, tid, next_tid(tid)); 2016 #endif 2017 stat(s, CMPXCHG_DOUBLE_CPU_FAIL); 2018 } 2019 2020 static void init_kmem_cache_cpus(struct kmem_cache *s) 2021 { 2022 int cpu; 2023 2024 for_each_possible_cpu(cpu) 2025 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu); 2026 } 2027 2028 /* 2029 * Remove the cpu slab 2030 */ 2031 static void deactivate_slab(struct kmem_cache *s, struct page *page, 2032 void *freelist, struct kmem_cache_cpu *c) 2033 { 2034 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE }; 2035 struct kmem_cache_node *n = get_node(s, page_to_nid(page)); 2036 int lock = 0; 2037 enum slab_modes l = M_NONE, m = M_NONE; 2038 void *nextfree; 2039 int tail = DEACTIVATE_TO_HEAD; 2040 struct page new; 2041 struct page old; 2042 2043 if (page->freelist) { 2044 stat(s, DEACTIVATE_REMOTE_FREES); 2045 tail = DEACTIVATE_TO_TAIL; 2046 } 2047 2048 /* 2049 * Stage one: Free all available per cpu objects back 2050 * to the page freelist while it is still frozen. Leave the 2051 * last one. 2052 * 2053 * There is no need to take the list->lock because the page 2054 * is still frozen. 2055 */ 2056 while (freelist && (nextfree = get_freepointer(s, freelist))) { 2057 void *prior; 2058 unsigned long counters; 2059 2060 do { 2061 prior = page->freelist; 2062 counters = page->counters; 2063 set_freepointer(s, freelist, prior); 2064 new.counters = counters; 2065 new.inuse--; 2066 VM_BUG_ON(!new.frozen); 2067 2068 } while (!__cmpxchg_double_slab(s, page, 2069 prior, counters, 2070 freelist, new.counters, 2071 "drain percpu freelist")); 2072 2073 freelist = nextfree; 2074 } 2075 2076 /* 2077 * Stage two: Ensure that the page is unfrozen while the 2078 * list presence reflects the actual number of objects 2079 * during unfreeze. 2080 * 2081 * We setup the list membership and then perform a cmpxchg 2082 * with the count. If there is a mismatch then the page 2083 * is not unfrozen but the page is on the wrong list. 2084 * 2085 * Then we restart the process which may have to remove 2086 * the page from the list that we just put it on again 2087 * because the number of objects in the slab may have 2088 * changed. 2089 */ 2090 redo: 2091 2092 old.freelist = page->freelist; 2093 old.counters = page->counters; 2094 VM_BUG_ON(!old.frozen); 2095 2096 /* Determine target state of the slab */ 2097 new.counters = old.counters; 2098 if (freelist) { 2099 new.inuse--; 2100 set_freepointer(s, freelist, old.freelist); 2101 new.freelist = freelist; 2102 } else 2103 new.freelist = old.freelist; 2104 2105 new.frozen = 0; 2106 2107 if (!new.inuse && n->nr_partial >= s->min_partial) 2108 m = M_FREE; 2109 else if (new.freelist) { 2110 m = M_PARTIAL; 2111 if (!lock) { 2112 lock = 1; 2113 /* 2114 * Taking the spinlock removes the possiblity 2115 * that acquire_slab() will see a slab page that 2116 * is frozen 2117 */ 2118 spin_lock(&n->list_lock); 2119 } 2120 } else { 2121 m = M_FULL; 2122 if (kmem_cache_debug(s) && !lock) { 2123 lock = 1; 2124 /* 2125 * This also ensures that the scanning of full 2126 * slabs from diagnostic functions will not see 2127 * any frozen slabs. 2128 */ 2129 spin_lock(&n->list_lock); 2130 } 2131 } 2132 2133 if (l != m) { 2134 2135 if (l == M_PARTIAL) 2136 2137 remove_partial(n, page); 2138 2139 else if (l == M_FULL) 2140 2141 remove_full(s, n, page); 2142 2143 if (m == M_PARTIAL) { 2144 2145 add_partial(n, page, tail); 2146 stat(s, tail); 2147 2148 } else if (m == M_FULL) { 2149 2150 stat(s, DEACTIVATE_FULL); 2151 add_full(s, n, page); 2152 2153 } 2154 } 2155 2156 l = m; 2157 if (!__cmpxchg_double_slab(s, page, 2158 old.freelist, old.counters, 2159 new.freelist, new.counters, 2160 "unfreezing slab")) 2161 goto redo; 2162 2163 if (lock) 2164 spin_unlock(&n->list_lock); 2165 2166 if (m == M_FREE) { 2167 stat(s, DEACTIVATE_EMPTY); 2168 discard_slab(s, page); 2169 stat(s, FREE_SLAB); 2170 } 2171 2172 c->page = NULL; 2173 c->freelist = NULL; 2174 } 2175 2176 /* 2177 * Unfreeze all the cpu partial slabs. 2178 * 2179 * This function must be called with interrupts disabled 2180 * for the cpu using c (or some other guarantee must be there 2181 * to guarantee no concurrent accesses). 2182 */ 2183 static void unfreeze_partials(struct kmem_cache *s, 2184 struct kmem_cache_cpu *c) 2185 { 2186 #ifdef CONFIG_SLUB_CPU_PARTIAL 2187 struct kmem_cache_node *n = NULL, *n2 = NULL; 2188 struct page *page, *discard_page = NULL; 2189 2190 while ((page = c->partial)) { 2191 struct page new; 2192 struct page old; 2193 2194 c->partial = page->next; 2195 2196 n2 = get_node(s, page_to_nid(page)); 2197 if (n != n2) { 2198 if (n) 2199 spin_unlock(&n->list_lock); 2200 2201 n = n2; 2202 spin_lock(&n->list_lock); 2203 } 2204 2205 do { 2206 2207 old.freelist = page->freelist; 2208 old.counters = page->counters; 2209 VM_BUG_ON(!old.frozen); 2210 2211 new.counters = old.counters; 2212 new.freelist = old.freelist; 2213 2214 new.frozen = 0; 2215 2216 } while (!__cmpxchg_double_slab(s, page, 2217 old.freelist, old.counters, 2218 new.freelist, new.counters, 2219 "unfreezing slab")); 2220 2221 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { 2222 page->next = discard_page; 2223 discard_page = page; 2224 } else { 2225 add_partial(n, page, DEACTIVATE_TO_TAIL); 2226 stat(s, FREE_ADD_PARTIAL); 2227 } 2228 } 2229 2230 if (n) 2231 spin_unlock(&n->list_lock); 2232 2233 while (discard_page) { 2234 page = discard_page; 2235 discard_page = discard_page->next; 2236 2237 stat(s, DEACTIVATE_EMPTY); 2238 discard_slab(s, page); 2239 stat(s, FREE_SLAB); 2240 } 2241 #endif 2242 } 2243 2244 /* 2245 * Put a page that was just frozen (in __slab_free) into a partial page 2246 * slot if available. 2247 * 2248 * If we did not find a slot then simply move all the partials to the 2249 * per node partial list. 2250 */ 2251 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain) 2252 { 2253 #ifdef CONFIG_SLUB_CPU_PARTIAL 2254 struct page *oldpage; 2255 int pages; 2256 int pobjects; 2257 2258 preempt_disable(); 2259 do { 2260 pages = 0; 2261 pobjects = 0; 2262 oldpage = this_cpu_read(s->cpu_slab->partial); 2263 2264 if (oldpage) { 2265 pobjects = oldpage->pobjects; 2266 pages = oldpage->pages; 2267 if (drain && pobjects > s->cpu_partial) { 2268 unsigned long flags; 2269 /* 2270 * partial array is full. Move the existing 2271 * set to the per node partial list. 2272 */ 2273 local_irq_save(flags); 2274 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2275 local_irq_restore(flags); 2276 oldpage = NULL; 2277 pobjects = 0; 2278 pages = 0; 2279 stat(s, CPU_PARTIAL_DRAIN); 2280 } 2281 } 2282 2283 pages++; 2284 pobjects += page->objects - page->inuse; 2285 2286 page->pages = pages; 2287 page->pobjects = pobjects; 2288 page->next = oldpage; 2289 2290 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) 2291 != oldpage); 2292 if (unlikely(!s->cpu_partial)) { 2293 unsigned long flags; 2294 2295 local_irq_save(flags); 2296 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab)); 2297 local_irq_restore(flags); 2298 } 2299 preempt_enable(); 2300 #endif 2301 } 2302 2303 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) 2304 { 2305 stat(s, CPUSLAB_FLUSH); 2306 deactivate_slab(s, c->page, c->freelist, c); 2307 2308 c->tid = next_tid(c->tid); 2309 } 2310 2311 /* 2312 * Flush cpu slab. 2313 * 2314 * Called from IPI handler with interrupts disabled. 2315 */ 2316 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) 2317 { 2318 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2319 2320 if (likely(c)) { 2321 if (c->page) 2322 flush_slab(s, c); 2323 2324 unfreeze_partials(s, c); 2325 } 2326 } 2327 2328 static void flush_cpu_slab(void *d) 2329 { 2330 struct kmem_cache *s = d; 2331 2332 __flush_cpu_slab(s, smp_processor_id()); 2333 } 2334 2335 static bool has_cpu_slab(int cpu, void *info) 2336 { 2337 struct kmem_cache *s = info; 2338 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); 2339 2340 return c->page || slub_percpu_partial(c); 2341 } 2342 2343 static void flush_all(struct kmem_cache *s) 2344 { 2345 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC); 2346 } 2347 2348 /* 2349 * Use the cpu notifier to insure that the cpu slabs are flushed when 2350 * necessary. 2351 */ 2352 static int slub_cpu_dead(unsigned int cpu) 2353 { 2354 struct kmem_cache *s; 2355 unsigned long flags; 2356 2357 mutex_lock(&slab_mutex); 2358 list_for_each_entry(s, &slab_caches, list) { 2359 local_irq_save(flags); 2360 __flush_cpu_slab(s, cpu); 2361 local_irq_restore(flags); 2362 } 2363 mutex_unlock(&slab_mutex); 2364 return 0; 2365 } 2366 2367 /* 2368 * Check if the objects in a per cpu structure fit numa 2369 * locality expectations. 2370 */ 2371 static inline int node_match(struct page *page, int node) 2372 { 2373 #ifdef CONFIG_NUMA 2374 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node)) 2375 return 0; 2376 #endif 2377 return 1; 2378 } 2379 2380 #ifdef CONFIG_SLUB_DEBUG 2381 static int count_free(struct page *page) 2382 { 2383 return page->objects - page->inuse; 2384 } 2385 2386 static inline unsigned long node_nr_objs(struct kmem_cache_node *n) 2387 { 2388 return atomic_long_read(&n->total_objects); 2389 } 2390 #endif /* CONFIG_SLUB_DEBUG */ 2391 2392 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS) 2393 static unsigned long count_partial(struct kmem_cache_node *n, 2394 int (*get_count)(struct page *)) 2395 { 2396 unsigned long flags; 2397 unsigned long x = 0; 2398 struct page *page; 2399 2400 spin_lock_irqsave(&n->list_lock, flags); 2401 list_for_each_entry(page, &n->partial, lru) 2402 x += get_count(page); 2403 spin_unlock_irqrestore(&n->list_lock, flags); 2404 return x; 2405 } 2406 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */ 2407 2408 static noinline void 2409 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) 2410 { 2411 #ifdef CONFIG_SLUB_DEBUG 2412 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, 2413 DEFAULT_RATELIMIT_BURST); 2414 int node; 2415 struct kmem_cache_node *n; 2416 2417 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) 2418 return; 2419 2420 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", 2421 nid, gfpflags, &gfpflags); 2422 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n", 2423 s->name, s->object_size, s->size, oo_order(s->oo), 2424 oo_order(s->min)); 2425 2426 if (oo_order(s->min) > get_order(s->object_size)) 2427 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n", 2428 s->name); 2429 2430 for_each_kmem_cache_node(s, node, n) { 2431 unsigned long nr_slabs; 2432 unsigned long nr_objs; 2433 unsigned long nr_free; 2434 2435 nr_free = count_partial(n, count_free); 2436 nr_slabs = node_nr_slabs(n); 2437 nr_objs = node_nr_objs(n); 2438 2439 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n", 2440 node, nr_slabs, nr_objs, nr_free); 2441 } 2442 #endif 2443 } 2444 2445 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags, 2446 int node, struct kmem_cache_cpu **pc) 2447 { 2448 void *freelist; 2449 struct kmem_cache_cpu *c = *pc; 2450 struct page *page; 2451 2452 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO)); 2453 2454 freelist = get_partial(s, flags, node, c); 2455 2456 if (freelist) 2457 return freelist; 2458 2459 page = new_slab(s, flags, node); 2460 if (page) { 2461 c = raw_cpu_ptr(s->cpu_slab); 2462 if (c->page) 2463 flush_slab(s, c); 2464 2465 /* 2466 * No other reference to the page yet so we can 2467 * muck around with it freely without cmpxchg 2468 */ 2469 freelist = page->freelist; 2470 page->freelist = NULL; 2471 2472 stat(s, ALLOC_SLAB); 2473 c->page = page; 2474 *pc = c; 2475 } else 2476 freelist = NULL; 2477 2478 return freelist; 2479 } 2480 2481 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags) 2482 { 2483 if (unlikely(PageSlabPfmemalloc(page))) 2484 return gfp_pfmemalloc_allowed(gfpflags); 2485 2486 return true; 2487 } 2488 2489 /* 2490 * Check the page->freelist of a page and either transfer the freelist to the 2491 * per cpu freelist or deactivate the page. 2492 * 2493 * The page is still frozen if the return value is not NULL. 2494 * 2495 * If this function returns NULL then the page has been unfrozen. 2496 * 2497 * This function must be called with interrupt disabled. 2498 */ 2499 static inline void *get_freelist(struct kmem_cache *s, struct page *page) 2500 { 2501 struct page new; 2502 unsigned long counters; 2503 void *freelist; 2504 2505 do { 2506 freelist = page->freelist; 2507 counters = page->counters; 2508 2509 new.counters = counters; 2510 VM_BUG_ON(!new.frozen); 2511 2512 new.inuse = page->objects; 2513 new.frozen = freelist != NULL; 2514 2515 } while (!__cmpxchg_double_slab(s, page, 2516 freelist, counters, 2517 NULL, new.counters, 2518 "get_freelist")); 2519 2520 return freelist; 2521 } 2522 2523 /* 2524 * Slow path. The lockless freelist is empty or we need to perform 2525 * debugging duties. 2526 * 2527 * Processing is still very fast if new objects have been freed to the 2528 * regular freelist. In that case we simply take over the regular freelist 2529 * as the lockless freelist and zap the regular freelist. 2530 * 2531 * If that is not working then we fall back to the partial lists. We take the 2532 * first element of the freelist as the object to allocate now and move the 2533 * rest of the freelist to the lockless freelist. 2534 * 2535 * And if we were unable to get a new slab from the partial slab lists then 2536 * we need to allocate a new slab. This is the slowest path since it involves 2537 * a call to the page allocator and the setup of a new slab. 2538 * 2539 * Version of __slab_alloc to use when we know that interrupts are 2540 * already disabled (which is the case for bulk allocation). 2541 */ 2542 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2543 unsigned long addr, struct kmem_cache_cpu *c) 2544 { 2545 void *freelist; 2546 struct page *page; 2547 2548 page = c->page; 2549 if (!page) 2550 goto new_slab; 2551 redo: 2552 2553 if (unlikely(!node_match(page, node))) { 2554 int searchnode = node; 2555 2556 if (node != NUMA_NO_NODE && !node_present_pages(node)) 2557 searchnode = node_to_mem_node(node); 2558 2559 if (unlikely(!node_match(page, searchnode))) { 2560 stat(s, ALLOC_NODE_MISMATCH); 2561 deactivate_slab(s, page, c->freelist, c); 2562 goto new_slab; 2563 } 2564 } 2565 2566 /* 2567 * By rights, we should be searching for a slab page that was 2568 * PFMEMALLOC but right now, we are losing the pfmemalloc 2569 * information when the page leaves the per-cpu allocator 2570 */ 2571 if (unlikely(!pfmemalloc_match(page, gfpflags))) { 2572 deactivate_slab(s, page, c->freelist, c); 2573 goto new_slab; 2574 } 2575 2576 /* must check again c->freelist in case of cpu migration or IRQ */ 2577 freelist = c->freelist; 2578 if (freelist) 2579 goto load_freelist; 2580 2581 freelist = get_freelist(s, page); 2582 2583 if (!freelist) { 2584 c->page = NULL; 2585 stat(s, DEACTIVATE_BYPASS); 2586 goto new_slab; 2587 } 2588 2589 stat(s, ALLOC_REFILL); 2590 2591 load_freelist: 2592 /* 2593 * freelist is pointing to the list of objects to be used. 2594 * page is pointing to the page from which the objects are obtained. 2595 * That page must be frozen for per cpu allocations to work. 2596 */ 2597 VM_BUG_ON(!c->page->frozen); 2598 c->freelist = get_freepointer(s, freelist); 2599 c->tid = next_tid(c->tid); 2600 return freelist; 2601 2602 new_slab: 2603 2604 if (slub_percpu_partial(c)) { 2605 page = c->page = slub_percpu_partial(c); 2606 slub_set_percpu_partial(c, page); 2607 stat(s, CPU_PARTIAL_ALLOC); 2608 goto redo; 2609 } 2610 2611 freelist = new_slab_objects(s, gfpflags, node, &c); 2612 2613 if (unlikely(!freelist)) { 2614 slab_out_of_memory(s, gfpflags, node); 2615 return NULL; 2616 } 2617 2618 page = c->page; 2619 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags))) 2620 goto load_freelist; 2621 2622 /* Only entered in the debug case */ 2623 if (kmem_cache_debug(s) && 2624 !alloc_debug_processing(s, page, freelist, addr)) 2625 goto new_slab; /* Slab failed checks. Next slab needed */ 2626 2627 deactivate_slab(s, page, get_freepointer(s, freelist), c); 2628 return freelist; 2629 } 2630 2631 /* 2632 * Another one that disabled interrupt and compensates for possible 2633 * cpu changes by refetching the per cpu area pointer. 2634 */ 2635 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, 2636 unsigned long addr, struct kmem_cache_cpu *c) 2637 { 2638 void *p; 2639 unsigned long flags; 2640 2641 local_irq_save(flags); 2642 #ifdef CONFIG_PREEMPT 2643 /* 2644 * We may have been preempted and rescheduled on a different 2645 * cpu before disabling interrupts. Need to reload cpu area 2646 * pointer. 2647 */ 2648 c = this_cpu_ptr(s->cpu_slab); 2649 #endif 2650 2651 p = ___slab_alloc(s, gfpflags, node, addr, c); 2652 local_irq_restore(flags); 2653 return p; 2654 } 2655 2656 /* 2657 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) 2658 * have the fastpath folded into their functions. So no function call 2659 * overhead for requests that can be satisfied on the fastpath. 2660 * 2661 * The fastpath works by first checking if the lockless freelist can be used. 2662 * If not then __slab_alloc is called for slow processing. 2663 * 2664 * Otherwise we can simply pick the next object from the lockless free list. 2665 */ 2666 static __always_inline void *slab_alloc_node(struct kmem_cache *s, 2667 gfp_t gfpflags, int node, unsigned long addr) 2668 { 2669 void *object; 2670 struct kmem_cache_cpu *c; 2671 struct page *page; 2672 unsigned long tid; 2673 2674 s = slab_pre_alloc_hook(s, gfpflags); 2675 if (!s) 2676 return NULL; 2677 redo: 2678 /* 2679 * Must read kmem_cache cpu data via this cpu ptr. Preemption is 2680 * enabled. We may switch back and forth between cpus while 2681 * reading from one cpu area. That does not matter as long 2682 * as we end up on the original cpu again when doing the cmpxchg. 2683 * 2684 * We should guarantee that tid and kmem_cache are retrieved on 2685 * the same cpu. It could be different if CONFIG_PREEMPT so we need 2686 * to check if it is matched or not. 2687 */ 2688 do { 2689 tid = this_cpu_read(s->cpu_slab->tid); 2690 c = raw_cpu_ptr(s->cpu_slab); 2691 } while (IS_ENABLED(CONFIG_PREEMPT) && 2692 unlikely(tid != READ_ONCE(c->tid))); 2693 2694 /* 2695 * Irqless object alloc/free algorithm used here depends on sequence 2696 * of fetching cpu_slab's data. tid should be fetched before anything 2697 * on c to guarantee that object and page associated with previous tid 2698 * won't be used with current tid. If we fetch tid first, object and 2699 * page could be one associated with next tid and our alloc/free 2700 * request will be failed. In this case, we will retry. So, no problem. 2701 */ 2702 barrier(); 2703 2704 /* 2705 * The transaction ids are globally unique per cpu and per operation on 2706 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double 2707 * occurs on the right processor and that there was no operation on the 2708 * linked list in between. 2709 */ 2710 2711 object = c->freelist; 2712 page = c->page; 2713 if (unlikely(!object || !node_match(page, node))) { 2714 object = __slab_alloc(s, gfpflags, node, addr, c); 2715 stat(s, ALLOC_SLOWPATH); 2716 } else { 2717 void *next_object = get_freepointer_safe(s, object); 2718 2719 /* 2720 * The cmpxchg will only match if there was no additional 2721 * operation and if we are on the right processor. 2722 * 2723 * The cmpxchg does the following atomically (without lock 2724 * semantics!) 2725 * 1. Relocate first pointer to the current per cpu area. 2726 * 2. Verify that tid and freelist have not been changed 2727 * 3. If they were not changed replace tid and freelist 2728 * 2729 * Since this is without lock semantics the protection is only 2730 * against code executing on this cpu *not* from access by 2731 * other cpus. 2732 */ 2733 if (unlikely(!this_cpu_cmpxchg_double( 2734 s->cpu_slab->freelist, s->cpu_slab->tid, 2735 object, tid, 2736 next_object, next_tid(tid)))) { 2737 2738 note_cmpxchg_failure("slab_alloc", s, tid); 2739 goto redo; 2740 } 2741 prefetch_freepointer(s, next_object); 2742 stat(s, ALLOC_FASTPATH); 2743 } 2744 2745 if (unlikely(gfpflags & __GFP_ZERO) && object) 2746 memset(object, 0, s->object_size); 2747 2748 slab_post_alloc_hook(s, gfpflags, 1, &object); 2749 2750 return object; 2751 } 2752 2753 static __always_inline void *slab_alloc(struct kmem_cache *s, 2754 gfp_t gfpflags, unsigned long addr) 2755 { 2756 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr); 2757 } 2758 2759 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) 2760 { 2761 void *ret = slab_alloc(s, gfpflags, _RET_IP_); 2762 2763 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, 2764 s->size, gfpflags); 2765 2766 return ret; 2767 } 2768 EXPORT_SYMBOL(kmem_cache_alloc); 2769 2770 #ifdef CONFIG_TRACING 2771 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) 2772 { 2773 void *ret = slab_alloc(s, gfpflags, _RET_IP_); 2774 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags); 2775 ret = kasan_kmalloc(s, ret, size, gfpflags); 2776 return ret; 2777 } 2778 EXPORT_SYMBOL(kmem_cache_alloc_trace); 2779 #endif 2780 2781 #ifdef CONFIG_NUMA 2782 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) 2783 { 2784 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); 2785 2786 trace_kmem_cache_alloc_node(_RET_IP_, ret, 2787 s->object_size, s->size, gfpflags, node); 2788 2789 return ret; 2790 } 2791 EXPORT_SYMBOL(kmem_cache_alloc_node); 2792 2793 #ifdef CONFIG_TRACING 2794 void *kmem_cache_alloc_node_trace(struct kmem_cache *s, 2795 gfp_t gfpflags, 2796 int node, size_t size) 2797 { 2798 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_); 2799 2800 trace_kmalloc_node(_RET_IP_, ret, 2801 size, s->size, gfpflags, node); 2802 2803 ret = kasan_kmalloc(s, ret, size, gfpflags); 2804 return ret; 2805 } 2806 EXPORT_SYMBOL(kmem_cache_alloc_node_trace); 2807 #endif 2808 #endif 2809 2810 /* 2811 * Slow path handling. This may still be called frequently since objects 2812 * have a longer lifetime than the cpu slabs in most processing loads. 2813 * 2814 * So we still attempt to reduce cache line usage. Just take the slab 2815 * lock and free the item. If there is no additional partial page 2816 * handling required then we can return immediately. 2817 */ 2818 static void __slab_free(struct kmem_cache *s, struct page *page, 2819 void *head, void *tail, int cnt, 2820 unsigned long addr) 2821 2822 { 2823 void *prior; 2824 int was_frozen; 2825 struct page new; 2826 unsigned long counters; 2827 struct kmem_cache_node *n = NULL; 2828 unsigned long uninitialized_var(flags); 2829 2830 stat(s, FREE_SLOWPATH); 2831 2832 if (kmem_cache_debug(s) && 2833 !free_debug_processing(s, page, head, tail, cnt, addr)) 2834 return; 2835 2836 do { 2837 if (unlikely(n)) { 2838 spin_unlock_irqrestore(&n->list_lock, flags); 2839 n = NULL; 2840 } 2841 prior = page->freelist; 2842 counters = page->counters; 2843 set_freepointer(s, tail, prior); 2844 new.counters = counters; 2845 was_frozen = new.frozen; 2846 new.inuse -= cnt; 2847 if ((!new.inuse || !prior) && !was_frozen) { 2848 2849 if (kmem_cache_has_cpu_partial(s) && !prior) { 2850 2851 /* 2852 * Slab was on no list before and will be 2853 * partially empty 2854 * We can defer the list move and instead 2855 * freeze it. 2856 */ 2857 new.frozen = 1; 2858 2859 } else { /* Needs to be taken off a list */ 2860 2861 n = get_node(s, page_to_nid(page)); 2862 /* 2863 * Speculatively acquire the list_lock. 2864 * If the cmpxchg does not succeed then we may 2865 * drop the list_lock without any processing. 2866 * 2867 * Otherwise the list_lock will synchronize with 2868 * other processors updating the list of slabs. 2869 */ 2870 spin_lock_irqsave(&n->list_lock, flags); 2871 2872 } 2873 } 2874 2875 } while (!cmpxchg_double_slab(s, page, 2876 prior, counters, 2877 head, new.counters, 2878 "__slab_free")); 2879 2880 if (likely(!n)) { 2881 2882 /* 2883 * If we just froze the page then put it onto the 2884 * per cpu partial list. 2885 */ 2886 if (new.frozen && !was_frozen) { 2887 put_cpu_partial(s, page, 1); 2888 stat(s, CPU_PARTIAL_FREE); 2889 } 2890 /* 2891 * The list lock was not taken therefore no list 2892 * activity can be necessary. 2893 */ 2894 if (was_frozen) 2895 stat(s, FREE_FROZEN); 2896 return; 2897 } 2898 2899 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) 2900 goto slab_empty; 2901 2902 /* 2903 * Objects left in the slab. If it was not on the partial list before 2904 * then add it. 2905 */ 2906 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { 2907 if (kmem_cache_debug(s)) 2908 remove_full(s, n, page); 2909 add_partial(n, page, DEACTIVATE_TO_TAIL); 2910 stat(s, FREE_ADD_PARTIAL); 2911 } 2912 spin_unlock_irqrestore(&n->list_lock, flags); 2913 return; 2914 2915 slab_empty: 2916 if (prior) { 2917 /* 2918 * Slab on the partial list. 2919 */ 2920 remove_partial(n, page); 2921 stat(s, FREE_REMOVE_PARTIAL); 2922 } else { 2923 /* Slab must be on the full list */ 2924 remove_full(s, n, page); 2925 } 2926 2927 spin_unlock_irqrestore(&n->list_lock, flags); 2928 stat(s, FREE_SLAB); 2929 discard_slab(s, page); 2930 } 2931 2932 /* 2933 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that 2934 * can perform fastpath freeing without additional function calls. 2935 * 2936 * The fastpath is only possible if we are freeing to the current cpu slab 2937 * of this processor. This typically the case if we have just allocated 2938 * the item before. 2939 * 2940 * If fastpath is not possible then fall back to __slab_free where we deal 2941 * with all sorts of special processing. 2942 * 2943 * Bulk free of a freelist with several objects (all pointing to the 2944 * same page) possible by specifying head and tail ptr, plus objects 2945 * count (cnt). Bulk free indicated by tail pointer being set. 2946 */ 2947 static __always_inline void do_slab_free(struct kmem_cache *s, 2948 struct page *page, void *head, void *tail, 2949 int cnt, unsigned long addr) 2950 { 2951 void *tail_obj = tail ? : head; 2952 struct kmem_cache_cpu *c; 2953 unsigned long tid; 2954 redo: 2955 /* 2956 * Determine the currently cpus per cpu slab. 2957 * The cpu may change afterward. However that does not matter since 2958 * data is retrieved via this pointer. If we are on the same cpu 2959 * during the cmpxchg then the free will succeed. 2960 */ 2961 do { 2962 tid = this_cpu_read(s->cpu_slab->tid); 2963 c = raw_cpu_ptr(s->cpu_slab); 2964 } while (IS_ENABLED(CONFIG_PREEMPT) && 2965 unlikely(tid != READ_ONCE(c->tid))); 2966 2967 /* Same with comment on barrier() in slab_alloc_node() */ 2968 barrier(); 2969 2970 if (likely(page == c->page)) { 2971 set_freepointer(s, tail_obj, c->freelist); 2972 2973 if (unlikely(!this_cpu_cmpxchg_double( 2974 s->cpu_slab->freelist, s->cpu_slab->tid, 2975 c->freelist, tid, 2976 head, next_tid(tid)))) { 2977 2978 note_cmpxchg_failure("slab_free", s, tid); 2979 goto redo; 2980 } 2981 stat(s, FREE_FASTPATH); 2982 } else 2983 __slab_free(s, page, head, tail_obj, cnt, addr); 2984 2985 } 2986 2987 static __always_inline void slab_free(struct kmem_cache *s, struct page *page, 2988 void *head, void *tail, int cnt, 2989 unsigned long addr) 2990 { 2991 /* 2992 * With KASAN enabled slab_free_freelist_hook modifies the freelist 2993 * to remove objects, whose reuse must be delayed. 2994 */ 2995 if (slab_free_freelist_hook(s, &head, &tail)) 2996 do_slab_free(s, page, head, tail, cnt, addr); 2997 } 2998 2999 #ifdef CONFIG_KASAN_GENERIC 3000 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) 3001 { 3002 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr); 3003 } 3004 #endif 3005 3006 void kmem_cache_free(struct kmem_cache *s, void *x) 3007 { 3008 s = cache_from_obj(s, x); 3009 if (!s) 3010 return; 3011 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_); 3012 trace_kmem_cache_free(_RET_IP_, x); 3013 } 3014 EXPORT_SYMBOL(kmem_cache_free); 3015 3016 struct detached_freelist { 3017 struct page *page; 3018 void *tail; 3019 void *freelist; 3020 int cnt; 3021 struct kmem_cache *s; 3022 }; 3023 3024 /* 3025 * This function progressively scans the array with free objects (with 3026 * a limited look ahead) and extract objects belonging to the same 3027 * page. It builds a detached freelist directly within the given 3028 * page/objects. This can happen without any need for 3029 * synchronization, because the objects are owned by running process. 3030 * The freelist is build up as a single linked list in the objects. 3031 * The idea is, that this detached freelist can then be bulk 3032 * transferred to the real freelist(s), but only requiring a single 3033 * synchronization primitive. Look ahead in the array is limited due 3034 * to performance reasons. 3035 */ 3036 static inline 3037 int build_detached_freelist(struct kmem_cache *s, size_t size, 3038 void **p, struct detached_freelist *df) 3039 { 3040 size_t first_skipped_index = 0; 3041 int lookahead = 3; 3042 void *object; 3043 struct page *page; 3044 3045 /* Always re-init detached_freelist */ 3046 df->page = NULL; 3047 3048 do { 3049 object = p[--size]; 3050 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */ 3051 } while (!object && size); 3052 3053 if (!object) 3054 return 0; 3055 3056 page = virt_to_head_page(object); 3057 if (!s) { 3058 /* Handle kalloc'ed objects */ 3059 if (unlikely(!PageSlab(page))) { 3060 BUG_ON(!PageCompound(page)); 3061 kfree_hook(object); 3062 __free_pages(page, compound_order(page)); 3063 p[size] = NULL; /* mark object processed */ 3064 return size; 3065 } 3066 /* Derive kmem_cache from object */ 3067 df->s = page->slab_cache; 3068 } else { 3069 df->s = cache_from_obj(s, object); /* Support for memcg */ 3070 } 3071 3072 /* Start new detached freelist */ 3073 df->page = page; 3074 set_freepointer(df->s, object, NULL); 3075 df->tail = object; 3076 df->freelist = object; 3077 p[size] = NULL; /* mark object processed */ 3078 df->cnt = 1; 3079 3080 while (size) { 3081 object = p[--size]; 3082 if (!object) 3083 continue; /* Skip processed objects */ 3084 3085 /* df->page is always set at this point */ 3086 if (df->page == virt_to_head_page(object)) { 3087 /* Opportunity build freelist */ 3088 set_freepointer(df->s, object, df->freelist); 3089 df->freelist = object; 3090 df->cnt++; 3091 p[size] = NULL; /* mark object processed */ 3092 3093 continue; 3094 } 3095 3096 /* Limit look ahead search */ 3097 if (!--lookahead) 3098 break; 3099 3100 if (!first_skipped_index) 3101 first_skipped_index = size + 1; 3102 } 3103 3104 return first_skipped_index; 3105 } 3106 3107 /* Note that interrupts must be enabled when calling this function. */ 3108 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) 3109 { 3110 if (WARN_ON(!size)) 3111 return; 3112 3113 do { 3114 struct detached_freelist df; 3115 3116 size = build_detached_freelist(s, size, p, &df); 3117 if (!df.page) 3118 continue; 3119 3120 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_); 3121 } while (likely(size)); 3122 } 3123 EXPORT_SYMBOL(kmem_cache_free_bulk); 3124 3125 /* Note that interrupts must be enabled when calling this function. */ 3126 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, 3127 void **p) 3128 { 3129 struct kmem_cache_cpu *c; 3130 int i; 3131 3132 /* memcg and kmem_cache debug support */ 3133 s = slab_pre_alloc_hook(s, flags); 3134 if (unlikely(!s)) 3135 return false; 3136 /* 3137 * Drain objects in the per cpu slab, while disabling local 3138 * IRQs, which protects against PREEMPT and interrupts 3139 * handlers invoking normal fastpath. 3140 */ 3141 local_irq_disable(); 3142 c = this_cpu_ptr(s->cpu_slab); 3143 3144 for (i = 0; i < size; i++) { 3145 void *object = c->freelist; 3146 3147 if (unlikely(!object)) { 3148 /* 3149 * Invoking slow path likely have side-effect 3150 * of re-populating per CPU c->freelist 3151 */ 3152 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, 3153 _RET_IP_, c); 3154 if (unlikely(!p[i])) 3155 goto error; 3156 3157 c = this_cpu_ptr(s->cpu_slab); 3158 continue; /* goto for-loop */ 3159 } 3160 c->freelist = get_freepointer(s, object); 3161 p[i] = object; 3162 } 3163 c->tid = next_tid(c->tid); 3164 local_irq_enable(); 3165 3166 /* Clear memory outside IRQ disabled fastpath loop */ 3167 if (unlikely(flags & __GFP_ZERO)) { 3168 int j; 3169 3170 for (j = 0; j < i; j++) 3171 memset(p[j], 0, s->object_size); 3172 } 3173 3174 /* memcg and kmem_cache debug support */ 3175 slab_post_alloc_hook(s, flags, size, p); 3176 return i; 3177 error: 3178 local_irq_enable(); 3179 slab_post_alloc_hook(s, flags, i, p); 3180 __kmem_cache_free_bulk(s, i, p); 3181 return 0; 3182 } 3183 EXPORT_SYMBOL(kmem_cache_alloc_bulk); 3184 3185 3186 /* 3187 * Object placement in a slab is made very easy because we always start at 3188 * offset 0. If we tune the size of the object to the alignment then we can 3189 * get the required alignment by putting one properly sized object after 3190 * another. 3191 * 3192 * Notice that the allocation order determines the sizes of the per cpu 3193 * caches. Each processor has always one slab available for allocations. 3194 * Increasing the allocation order reduces the number of times that slabs 3195 * must be moved on and off the partial lists and is therefore a factor in 3196 * locking overhead. 3197 */ 3198 3199 /* 3200 * Mininum / Maximum order of slab pages. This influences locking overhead 3201 * and slab fragmentation. A higher order reduces the number of partial slabs 3202 * and increases the number of allocations possible without having to 3203 * take the list_lock. 3204 */ 3205 static unsigned int slub_min_order; 3206 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER; 3207 static unsigned int slub_min_objects; 3208 3209 /* 3210 * Calculate the order of allocation given an slab object size. 3211 * 3212 * The order of allocation has significant impact on performance and other 3213 * system components. Generally order 0 allocations should be preferred since 3214 * order 0 does not cause fragmentation in the page allocator. Larger objects 3215 * be problematic to put into order 0 slabs because there may be too much 3216 * unused space left. We go to a higher order if more than 1/16th of the slab 3217 * would be wasted. 3218 * 3219 * In order to reach satisfactory performance we must ensure that a minimum 3220 * number of objects is in one slab. Otherwise we may generate too much 3221 * activity on the partial lists which requires taking the list_lock. This is 3222 * less a concern for large slabs though which are rarely used. 3223 * 3224 * slub_max_order specifies the order where we begin to stop considering the 3225 * number of objects in a slab as critical. If we reach slub_max_order then 3226 * we try to keep the page order as low as possible. So we accept more waste 3227 * of space in favor of a small page order. 3228 * 3229 * Higher order allocations also allow the placement of more objects in a 3230 * slab and thereby reduce object handling overhead. If the user has 3231 * requested a higher mininum order then we start with that one instead of 3232 * the smallest order which will fit the object. 3233 */ 3234 static inline unsigned int slab_order(unsigned int size, 3235 unsigned int min_objects, unsigned int max_order, 3236 unsigned int fract_leftover) 3237 { 3238 unsigned int min_order = slub_min_order; 3239 unsigned int order; 3240 3241 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE) 3242 return get_order(size * MAX_OBJS_PER_PAGE) - 1; 3243 3244 for (order = max(min_order, (unsigned int)get_order(min_objects * size)); 3245 order <= max_order; order++) { 3246 3247 unsigned int slab_size = (unsigned int)PAGE_SIZE << order; 3248 unsigned int rem; 3249 3250 rem = slab_size % size; 3251 3252 if (rem <= slab_size / fract_leftover) 3253 break; 3254 } 3255 3256 return order; 3257 } 3258 3259 static inline int calculate_order(unsigned int size) 3260 { 3261 unsigned int order; 3262 unsigned int min_objects; 3263 unsigned int max_objects; 3264 3265 /* 3266 * Attempt to find best configuration for a slab. This 3267 * works by first attempting to generate a layout with 3268 * the best configuration and backing off gradually. 3269 * 3270 * First we increase the acceptable waste in a slab. Then 3271 * we reduce the minimum objects required in a slab. 3272 */ 3273 min_objects = slub_min_objects; 3274 if (!min_objects) 3275 min_objects = 4 * (fls(nr_cpu_ids) + 1); 3276 max_objects = order_objects(slub_max_order, size); 3277 min_objects = min(min_objects, max_objects); 3278 3279 while (min_objects > 1) { 3280 unsigned int fraction; 3281 3282 fraction = 16; 3283 while (fraction >= 4) { 3284 order = slab_order(size, min_objects, 3285 slub_max_order, fraction); 3286 if (order <= slub_max_order) 3287 return order; 3288 fraction /= 2; 3289 } 3290 min_objects--; 3291 } 3292 3293 /* 3294 * We were unable to place multiple objects in a slab. Now 3295 * lets see if we can place a single object there. 3296 */ 3297 order = slab_order(size, 1, slub_max_order, 1); 3298 if (order <= slub_max_order) 3299 return order; 3300 3301 /* 3302 * Doh this slab cannot be placed using slub_max_order. 3303 */ 3304 order = slab_order(size, 1, MAX_ORDER, 1); 3305 if (order < MAX_ORDER) 3306 return order; 3307 return -ENOSYS; 3308 } 3309 3310 static void 3311 init_kmem_cache_node(struct kmem_cache_node *n) 3312 { 3313 n->nr_partial = 0; 3314 spin_lock_init(&n->list_lock); 3315 INIT_LIST_HEAD(&n->partial); 3316 #ifdef CONFIG_SLUB_DEBUG 3317 atomic_long_set(&n->nr_slabs, 0); 3318 atomic_long_set(&n->total_objects, 0); 3319 INIT_LIST_HEAD(&n->full); 3320 #endif 3321 } 3322 3323 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) 3324 { 3325 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < 3326 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu)); 3327 3328 /* 3329 * Must align to double word boundary for the double cmpxchg 3330 * instructions to work; see __pcpu_double_call_return_bool(). 3331 */ 3332 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 3333 2 * sizeof(void *)); 3334 3335 if (!s->cpu_slab) 3336 return 0; 3337 3338 init_kmem_cache_cpus(s); 3339 3340 return 1; 3341 } 3342 3343 static struct kmem_cache *kmem_cache_node; 3344 3345 /* 3346 * No kmalloc_node yet so do it by hand. We know that this is the first 3347 * slab on the node for this slabcache. There are no concurrent accesses 3348 * possible. 3349 * 3350 * Note that this function only works on the kmem_cache_node 3351 * when allocating for the kmem_cache_node. This is used for bootstrapping 3352 * memory on a fresh node that has no slab structures yet. 3353 */ 3354 static void early_kmem_cache_node_alloc(int node) 3355 { 3356 struct page *page; 3357 struct kmem_cache_node *n; 3358 3359 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); 3360 3361 page = new_slab(kmem_cache_node, GFP_NOWAIT, node); 3362 3363 BUG_ON(!page); 3364 if (page_to_nid(page) != node) { 3365 pr_err("SLUB: Unable to allocate memory from node %d\n", node); 3366 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n"); 3367 } 3368 3369 n = page->freelist; 3370 BUG_ON(!n); 3371 #ifdef CONFIG_SLUB_DEBUG 3372 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); 3373 init_tracking(kmem_cache_node, n); 3374 #endif 3375 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node), 3376 GFP_KERNEL); 3377 page->freelist = get_freepointer(kmem_cache_node, n); 3378 page->inuse = 1; 3379 page->frozen = 0; 3380 kmem_cache_node->node[node] = n; 3381 init_kmem_cache_node(n); 3382 inc_slabs_node(kmem_cache_node, node, page->objects); 3383 3384 /* 3385 * No locks need to be taken here as it has just been 3386 * initialized and there is no concurrent access. 3387 */ 3388 __add_partial(n, page, DEACTIVATE_TO_HEAD); 3389 } 3390 3391 static void free_kmem_cache_nodes(struct kmem_cache *s) 3392 { 3393 int node; 3394 struct kmem_cache_node *n; 3395 3396 for_each_kmem_cache_node(s, node, n) { 3397 s->node[node] = NULL; 3398 kmem_cache_free(kmem_cache_node, n); 3399 } 3400 } 3401 3402 void __kmem_cache_release(struct kmem_cache *s) 3403 { 3404 cache_random_seq_destroy(s); 3405 free_percpu(s->cpu_slab); 3406 free_kmem_cache_nodes(s); 3407 } 3408 3409 static int init_kmem_cache_nodes(struct kmem_cache *s) 3410 { 3411 int node; 3412 3413 for_each_node_state(node, N_NORMAL_MEMORY) { 3414 struct kmem_cache_node *n; 3415 3416 if (slab_state == DOWN) { 3417 early_kmem_cache_node_alloc(node); 3418 continue; 3419 } 3420 n = kmem_cache_alloc_node(kmem_cache_node, 3421 GFP_KERNEL, node); 3422 3423 if (!n) { 3424 free_kmem_cache_nodes(s); 3425 return 0; 3426 } 3427 3428 init_kmem_cache_node(n); 3429 s->node[node] = n; 3430 } 3431 return 1; 3432 } 3433 3434 static void set_min_partial(struct kmem_cache *s, unsigned long min) 3435 { 3436 if (min < MIN_PARTIAL) 3437 min = MIN_PARTIAL; 3438 else if (min > MAX_PARTIAL) 3439 min = MAX_PARTIAL; 3440 s->min_partial = min; 3441 } 3442 3443 static void set_cpu_partial(struct kmem_cache *s) 3444 { 3445 #ifdef CONFIG_SLUB_CPU_PARTIAL 3446 /* 3447 * cpu_partial determined the maximum number of objects kept in the 3448 * per cpu partial lists of a processor. 3449 * 3450 * Per cpu partial lists mainly contain slabs that just have one 3451 * object freed. If they are used for allocation then they can be 3452 * filled up again with minimal effort. The slab will never hit the 3453 * per node partial lists and therefore no locking will be required. 3454 * 3455 * This setting also determines 3456 * 3457 * A) The number of objects from per cpu partial slabs dumped to the 3458 * per node list when we reach the limit. 3459 * B) The number of objects in cpu partial slabs to extract from the 3460 * per node list when we run out of per cpu objects. We only fetch 3461 * 50% to keep some capacity around for frees. 3462 */ 3463 if (!kmem_cache_has_cpu_partial(s)) 3464 s->cpu_partial = 0; 3465 else if (s->size >= PAGE_SIZE) 3466 s->cpu_partial = 2; 3467 else if (s->size >= 1024) 3468 s->cpu_partial = 6; 3469 else if (s->size >= 256) 3470 s->cpu_partial = 13; 3471 else 3472 s->cpu_partial = 30; 3473 #endif 3474 } 3475 3476 /* 3477 * calculate_sizes() determines the order and the distribution of data within 3478 * a slab object. 3479 */ 3480 static int calculate_sizes(struct kmem_cache *s, int forced_order) 3481 { 3482 slab_flags_t flags = s->flags; 3483 unsigned int size = s->object_size; 3484 unsigned int order; 3485 3486 /* 3487 * Round up object size to the next word boundary. We can only 3488 * place the free pointer at word boundaries and this determines 3489 * the possible location of the free pointer. 3490 */ 3491 size = ALIGN(size, sizeof(void *)); 3492 3493 #ifdef CONFIG_SLUB_DEBUG 3494 /* 3495 * Determine if we can poison the object itself. If the user of 3496 * the slab may touch the object after free or before allocation 3497 * then we should never poison the object itself. 3498 */ 3499 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && 3500 !s->ctor) 3501 s->flags |= __OBJECT_POISON; 3502 else 3503 s->flags &= ~__OBJECT_POISON; 3504 3505 3506 /* 3507 * If we are Redzoning then check if there is some space between the 3508 * end of the object and the free pointer. If not then add an 3509 * additional word to have some bytes to store Redzone information. 3510 */ 3511 if ((flags & SLAB_RED_ZONE) && size == s->object_size) 3512 size += sizeof(void *); 3513 #endif 3514 3515 /* 3516 * With that we have determined the number of bytes in actual use 3517 * by the object. This is the potential offset to the free pointer. 3518 */ 3519 s->inuse = size; 3520 3521 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || 3522 s->ctor)) { 3523 /* 3524 * Relocate free pointer after the object if it is not 3525 * permitted to overwrite the first word of the object on 3526 * kmem_cache_free. 3527 * 3528 * This is the case if we do RCU, have a constructor or 3529 * destructor or are poisoning the objects. 3530 */ 3531 s->offset = size; 3532 size += sizeof(void *); 3533 } 3534 3535 #ifdef CONFIG_SLUB_DEBUG 3536 if (flags & SLAB_STORE_USER) 3537 /* 3538 * Need to store information about allocs and frees after 3539 * the object. 3540 */ 3541 size += 2 * sizeof(struct track); 3542 #endif 3543 3544 kasan_cache_create(s, &size, &s->flags); 3545 #ifdef CONFIG_SLUB_DEBUG 3546 if (flags & SLAB_RED_ZONE) { 3547 /* 3548 * Add some empty padding so that we can catch 3549 * overwrites from earlier objects rather than let 3550 * tracking information or the free pointer be 3551 * corrupted if a user writes before the start 3552 * of the object. 3553 */ 3554 size += sizeof(void *); 3555 3556 s->red_left_pad = sizeof(void *); 3557 s->red_left_pad = ALIGN(s->red_left_pad, s->align); 3558 size += s->red_left_pad; 3559 } 3560 #endif 3561 3562 /* 3563 * SLUB stores one object immediately after another beginning from 3564 * offset 0. In order to align the objects we have to simply size 3565 * each object to conform to the alignment. 3566 */ 3567 size = ALIGN(size, s->align); 3568 s->size = size; 3569 if (forced_order >= 0) 3570 order = forced_order; 3571 else 3572 order = calculate_order(size); 3573 3574 if ((int)order < 0) 3575 return 0; 3576 3577 s->allocflags = 0; 3578 if (order) 3579 s->allocflags |= __GFP_COMP; 3580 3581 if (s->flags & SLAB_CACHE_DMA) 3582 s->allocflags |= GFP_DMA; 3583 3584 if (s->flags & SLAB_RECLAIM_ACCOUNT) 3585 s->allocflags |= __GFP_RECLAIMABLE; 3586 3587 /* 3588 * Determine the number of objects per slab 3589 */ 3590 s->oo = oo_make(order, size); 3591 s->min = oo_make(get_order(size), size); 3592 if (oo_objects(s->oo) > oo_objects(s->max)) 3593 s->max = s->oo; 3594 3595 return !!oo_objects(s->oo); 3596 } 3597 3598 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) 3599 { 3600 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor); 3601 #ifdef CONFIG_SLAB_FREELIST_HARDENED 3602 s->random = get_random_long(); 3603 #endif 3604 3605 if (!calculate_sizes(s, -1)) 3606 goto error; 3607 if (disable_higher_order_debug) { 3608 /* 3609 * Disable debugging flags that store metadata if the min slab 3610 * order increased. 3611 */ 3612 if (get_order(s->size) > get_order(s->object_size)) { 3613 s->flags &= ~DEBUG_METADATA_FLAGS; 3614 s->offset = 0; 3615 if (!calculate_sizes(s, -1)) 3616 goto error; 3617 } 3618 } 3619 3620 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \ 3621 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE) 3622 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0) 3623 /* Enable fast mode */ 3624 s->flags |= __CMPXCHG_DOUBLE; 3625 #endif 3626 3627 /* 3628 * The larger the object size is, the more pages we want on the partial 3629 * list to avoid pounding the page allocator excessively. 3630 */ 3631 set_min_partial(s, ilog2(s->size) / 2); 3632 3633 set_cpu_partial(s); 3634 3635 #ifdef CONFIG_NUMA 3636 s->remote_node_defrag_ratio = 1000; 3637 #endif 3638 3639 /* Initialize the pre-computed randomized freelist if slab is up */ 3640 if (slab_state >= UP) { 3641 if (init_cache_random_seq(s)) 3642 goto error; 3643 } 3644 3645 if (!init_kmem_cache_nodes(s)) 3646 goto error; 3647 3648 if (alloc_kmem_cache_cpus(s)) 3649 return 0; 3650 3651 free_kmem_cache_nodes(s); 3652 error: 3653 if (flags & SLAB_PANIC) 3654 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n", 3655 s->name, s->size, s->size, 3656 oo_order(s->oo), s->offset, (unsigned long)flags); 3657 return -EINVAL; 3658 } 3659 3660 static void list_slab_objects(struct kmem_cache *s, struct page *page, 3661 const char *text) 3662 { 3663 #ifdef CONFIG_SLUB_DEBUG 3664 void *addr = page_address(page); 3665 void *p; 3666 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC); 3667 if (!map) 3668 return; 3669 slab_err(s, page, text, s->name); 3670 slab_lock(page); 3671 3672 get_map(s, page, map); 3673 for_each_object(p, s, addr, page->objects) { 3674 3675 if (!test_bit(slab_index(p, s, addr), map)) { 3676 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr); 3677 print_tracking(s, p); 3678 } 3679 } 3680 slab_unlock(page); 3681 bitmap_free(map); 3682 #endif 3683 } 3684 3685 /* 3686 * Attempt to free all partial slabs on a node. 3687 * This is called from __kmem_cache_shutdown(). We must take list_lock 3688 * because sysfs file might still access partial list after the shutdowning. 3689 */ 3690 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) 3691 { 3692 LIST_HEAD(discard); 3693 struct page *page, *h; 3694 3695 BUG_ON(irqs_disabled()); 3696 spin_lock_irq(&n->list_lock); 3697 list_for_each_entry_safe(page, h, &n->partial, lru) { 3698 if (!page->inuse) { 3699 remove_partial(n, page); 3700 list_add(&page->lru, &discard); 3701 } else { 3702 list_slab_objects(s, page, 3703 "Objects remaining in %s on __kmem_cache_shutdown()"); 3704 } 3705 } 3706 spin_unlock_irq(&n->list_lock); 3707 3708 list_for_each_entry_safe(page, h, &discard, lru) 3709 discard_slab(s, page); 3710 } 3711 3712 bool __kmem_cache_empty(struct kmem_cache *s) 3713 { 3714 int node; 3715 struct kmem_cache_node *n; 3716 3717 for_each_kmem_cache_node(s, node, n) 3718 if (n->nr_partial || slabs_node(s, node)) 3719 return false; 3720 return true; 3721 } 3722 3723 /* 3724 * Release all resources used by a slab cache. 3725 */ 3726 int __kmem_cache_shutdown(struct kmem_cache *s) 3727 { 3728 int node; 3729 struct kmem_cache_node *n; 3730 3731 flush_all(s); 3732 /* Attempt to free all objects */ 3733 for_each_kmem_cache_node(s, node, n) { 3734 free_partial(s, n); 3735 if (n->nr_partial || slabs_node(s, node)) 3736 return 1; 3737 } 3738 sysfs_slab_remove(s); 3739 return 0; 3740 } 3741 3742 /******************************************************************** 3743 * Kmalloc subsystem 3744 *******************************************************************/ 3745 3746 static int __init setup_slub_min_order(char *str) 3747 { 3748 get_option(&str, (int *)&slub_min_order); 3749 3750 return 1; 3751 } 3752 3753 __setup("slub_min_order=", setup_slub_min_order); 3754 3755 static int __init setup_slub_max_order(char *str) 3756 { 3757 get_option(&str, (int *)&slub_max_order); 3758 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1); 3759 3760 return 1; 3761 } 3762 3763 __setup("slub_max_order=", setup_slub_max_order); 3764 3765 static int __init setup_slub_min_objects(char *str) 3766 { 3767 get_option(&str, (int *)&slub_min_objects); 3768 3769 return 1; 3770 } 3771 3772 __setup("slub_min_objects=", setup_slub_min_objects); 3773 3774 void *__kmalloc(size_t size, gfp_t flags) 3775 { 3776 struct kmem_cache *s; 3777 void *ret; 3778 3779 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 3780 return kmalloc_large(size, flags); 3781 3782 s = kmalloc_slab(size, flags); 3783 3784 if (unlikely(ZERO_OR_NULL_PTR(s))) 3785 return s; 3786 3787 ret = slab_alloc(s, flags, _RET_IP_); 3788 3789 trace_kmalloc(_RET_IP_, ret, size, s->size, flags); 3790 3791 ret = kasan_kmalloc(s, ret, size, flags); 3792 3793 return ret; 3794 } 3795 EXPORT_SYMBOL(__kmalloc); 3796 3797 #ifdef CONFIG_NUMA 3798 static void *kmalloc_large_node(size_t size, gfp_t flags, int node) 3799 { 3800 struct page *page; 3801 void *ptr = NULL; 3802 3803 flags |= __GFP_COMP; 3804 page = alloc_pages_node(node, flags, get_order(size)); 3805 if (page) 3806 ptr = page_address(page); 3807 3808 return kmalloc_large_node_hook(ptr, size, flags); 3809 } 3810 3811 void *__kmalloc_node(size_t size, gfp_t flags, int node) 3812 { 3813 struct kmem_cache *s; 3814 void *ret; 3815 3816 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 3817 ret = kmalloc_large_node(size, flags, node); 3818 3819 trace_kmalloc_node(_RET_IP_, ret, 3820 size, PAGE_SIZE << get_order(size), 3821 flags, node); 3822 3823 return ret; 3824 } 3825 3826 s = kmalloc_slab(size, flags); 3827 3828 if (unlikely(ZERO_OR_NULL_PTR(s))) 3829 return s; 3830 3831 ret = slab_alloc_node(s, flags, node, _RET_IP_); 3832 3833 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node); 3834 3835 ret = kasan_kmalloc(s, ret, size, flags); 3836 3837 return ret; 3838 } 3839 EXPORT_SYMBOL(__kmalloc_node); 3840 #endif 3841 3842 #ifdef CONFIG_HARDENED_USERCOPY 3843 /* 3844 * Rejects incorrectly sized objects and objects that are to be copied 3845 * to/from userspace but do not fall entirely within the containing slab 3846 * cache's usercopy region. 3847 * 3848 * Returns NULL if check passes, otherwise const char * to name of cache 3849 * to indicate an error. 3850 */ 3851 void __check_heap_object(const void *ptr, unsigned long n, struct page *page, 3852 bool to_user) 3853 { 3854 struct kmem_cache *s; 3855 unsigned int offset; 3856 size_t object_size; 3857 3858 /* Find object and usable object size. */ 3859 s = page->slab_cache; 3860 3861 /* Reject impossible pointers. */ 3862 if (ptr < page_address(page)) 3863 usercopy_abort("SLUB object not in SLUB page?!", NULL, 3864 to_user, 0, n); 3865 3866 /* Find offset within object. */ 3867 offset = (ptr - page_address(page)) % s->size; 3868 3869 /* Adjust for redzone and reject if within the redzone. */ 3870 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) { 3871 if (offset < s->red_left_pad) 3872 usercopy_abort("SLUB object in left red zone", 3873 s->name, to_user, offset, n); 3874 offset -= s->red_left_pad; 3875 } 3876 3877 /* Allow address range falling entirely within usercopy region. */ 3878 if (offset >= s->useroffset && 3879 offset - s->useroffset <= s->usersize && 3880 n <= s->useroffset - offset + s->usersize) 3881 return; 3882 3883 /* 3884 * If the copy is still within the allocated object, produce 3885 * a warning instead of rejecting the copy. This is intended 3886 * to be a temporary method to find any missing usercopy 3887 * whitelists. 3888 */ 3889 object_size = slab_ksize(s); 3890 if (usercopy_fallback && 3891 offset <= object_size && n <= object_size - offset) { 3892 usercopy_warn("SLUB object", s->name, to_user, offset, n); 3893 return; 3894 } 3895 3896 usercopy_abort("SLUB object", s->name, to_user, offset, n); 3897 } 3898 #endif /* CONFIG_HARDENED_USERCOPY */ 3899 3900 static size_t __ksize(const void *object) 3901 { 3902 struct page *page; 3903 3904 if (unlikely(object == ZERO_SIZE_PTR)) 3905 return 0; 3906 3907 page = virt_to_head_page(object); 3908 3909 if (unlikely(!PageSlab(page))) { 3910 WARN_ON(!PageCompound(page)); 3911 return PAGE_SIZE << compound_order(page); 3912 } 3913 3914 return slab_ksize(page->slab_cache); 3915 } 3916 3917 size_t ksize(const void *object) 3918 { 3919 size_t size = __ksize(object); 3920 /* We assume that ksize callers could use whole allocated area, 3921 * so we need to unpoison this area. 3922 */ 3923 kasan_unpoison_shadow(object, size); 3924 return size; 3925 } 3926 EXPORT_SYMBOL(ksize); 3927 3928 void kfree(const void *x) 3929 { 3930 struct page *page; 3931 void *object = (void *)x; 3932 3933 trace_kfree(_RET_IP_, x); 3934 3935 if (unlikely(ZERO_OR_NULL_PTR(x))) 3936 return; 3937 3938 page = virt_to_head_page(x); 3939 if (unlikely(!PageSlab(page))) { 3940 BUG_ON(!PageCompound(page)); 3941 kfree_hook(object); 3942 __free_pages(page, compound_order(page)); 3943 return; 3944 } 3945 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_); 3946 } 3947 EXPORT_SYMBOL(kfree); 3948 3949 #define SHRINK_PROMOTE_MAX 32 3950 3951 /* 3952 * kmem_cache_shrink discards empty slabs and promotes the slabs filled 3953 * up most to the head of the partial lists. New allocations will then 3954 * fill those up and thus they can be removed from the partial lists. 3955 * 3956 * The slabs with the least items are placed last. This results in them 3957 * being allocated from last increasing the chance that the last objects 3958 * are freed in them. 3959 */ 3960 int __kmem_cache_shrink(struct kmem_cache *s) 3961 { 3962 int node; 3963 int i; 3964 struct kmem_cache_node *n; 3965 struct page *page; 3966 struct page *t; 3967 struct list_head discard; 3968 struct list_head promote[SHRINK_PROMOTE_MAX]; 3969 unsigned long flags; 3970 int ret = 0; 3971 3972 flush_all(s); 3973 for_each_kmem_cache_node(s, node, n) { 3974 INIT_LIST_HEAD(&discard); 3975 for (i = 0; i < SHRINK_PROMOTE_MAX; i++) 3976 INIT_LIST_HEAD(promote + i); 3977 3978 spin_lock_irqsave(&n->list_lock, flags); 3979 3980 /* 3981 * Build lists of slabs to discard or promote. 3982 * 3983 * Note that concurrent frees may occur while we hold the 3984 * list_lock. page->inuse here is the upper limit. 3985 */ 3986 list_for_each_entry_safe(page, t, &n->partial, lru) { 3987 int free = page->objects - page->inuse; 3988 3989 /* Do not reread page->inuse */ 3990 barrier(); 3991 3992 /* We do not keep full slabs on the list */ 3993 BUG_ON(free <= 0); 3994 3995 if (free == page->objects) { 3996 list_move(&page->lru, &discard); 3997 n->nr_partial--; 3998 } else if (free <= SHRINK_PROMOTE_MAX) 3999 list_move(&page->lru, promote + free - 1); 4000 } 4001 4002 /* 4003 * Promote the slabs filled up most to the head of the 4004 * partial list. 4005 */ 4006 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) 4007 list_splice(promote + i, &n->partial); 4008 4009 spin_unlock_irqrestore(&n->list_lock, flags); 4010 4011 /* Release empty slabs */ 4012 list_for_each_entry_safe(page, t, &discard, lru) 4013 discard_slab(s, page); 4014 4015 if (slabs_node(s, node)) 4016 ret = 1; 4017 } 4018 4019 return ret; 4020 } 4021 4022 #ifdef CONFIG_MEMCG 4023 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s) 4024 { 4025 /* 4026 * Called with all the locks held after a sched RCU grace period. 4027 * Even if @s becomes empty after shrinking, we can't know that @s 4028 * doesn't have allocations already in-flight and thus can't 4029 * destroy @s until the associated memcg is released. 4030 * 4031 * However, let's remove the sysfs files for empty caches here. 4032 * Each cache has a lot of interface files which aren't 4033 * particularly useful for empty draining caches; otherwise, we can 4034 * easily end up with millions of unnecessary sysfs files on 4035 * systems which have a lot of memory and transient cgroups. 4036 */ 4037 if (!__kmem_cache_shrink(s)) 4038 sysfs_slab_remove(s); 4039 } 4040 4041 void __kmemcg_cache_deactivate(struct kmem_cache *s) 4042 { 4043 /* 4044 * Disable empty slabs caching. Used to avoid pinning offline 4045 * memory cgroups by kmem pages that can be freed. 4046 */ 4047 slub_set_cpu_partial(s, 0); 4048 s->min_partial = 0; 4049 4050 /* 4051 * s->cpu_partial is checked locklessly (see put_cpu_partial), so 4052 * we have to make sure the change is visible before shrinking. 4053 */ 4054 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu); 4055 } 4056 #endif 4057 4058 static int slab_mem_going_offline_callback(void *arg) 4059 { 4060 struct kmem_cache *s; 4061 4062 mutex_lock(&slab_mutex); 4063 list_for_each_entry(s, &slab_caches, list) 4064 __kmem_cache_shrink(s); 4065 mutex_unlock(&slab_mutex); 4066 4067 return 0; 4068 } 4069 4070 static void slab_mem_offline_callback(void *arg) 4071 { 4072 struct kmem_cache_node *n; 4073 struct kmem_cache *s; 4074 struct memory_notify *marg = arg; 4075 int offline_node; 4076 4077 offline_node = marg->status_change_nid_normal; 4078 4079 /* 4080 * If the node still has available memory. we need kmem_cache_node 4081 * for it yet. 4082 */ 4083 if (offline_node < 0) 4084 return; 4085 4086 mutex_lock(&slab_mutex); 4087 list_for_each_entry(s, &slab_caches, list) { 4088 n = get_node(s, offline_node); 4089 if (n) { 4090 /* 4091 * if n->nr_slabs > 0, slabs still exist on the node 4092 * that is going down. We were unable to free them, 4093 * and offline_pages() function shouldn't call this 4094 * callback. So, we must fail. 4095 */ 4096 BUG_ON(slabs_node(s, offline_node)); 4097 4098 s->node[offline_node] = NULL; 4099 kmem_cache_free(kmem_cache_node, n); 4100 } 4101 } 4102 mutex_unlock(&slab_mutex); 4103 } 4104 4105 static int slab_mem_going_online_callback(void *arg) 4106 { 4107 struct kmem_cache_node *n; 4108 struct kmem_cache *s; 4109 struct memory_notify *marg = arg; 4110 int nid = marg->status_change_nid_normal; 4111 int ret = 0; 4112 4113 /* 4114 * If the node's memory is already available, then kmem_cache_node is 4115 * already created. Nothing to do. 4116 */ 4117 if (nid < 0) 4118 return 0; 4119 4120 /* 4121 * We are bringing a node online. No memory is available yet. We must 4122 * allocate a kmem_cache_node structure in order to bring the node 4123 * online. 4124 */ 4125 mutex_lock(&slab_mutex); 4126 list_for_each_entry(s, &slab_caches, list) { 4127 /* 4128 * XXX: kmem_cache_alloc_node will fallback to other nodes 4129 * since memory is not yet available from the node that 4130 * is brought up. 4131 */ 4132 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); 4133 if (!n) { 4134 ret = -ENOMEM; 4135 goto out; 4136 } 4137 init_kmem_cache_node(n); 4138 s->node[nid] = n; 4139 } 4140 out: 4141 mutex_unlock(&slab_mutex); 4142 return ret; 4143 } 4144 4145 static int slab_memory_callback(struct notifier_block *self, 4146 unsigned long action, void *arg) 4147 { 4148 int ret = 0; 4149 4150 switch (action) { 4151 case MEM_GOING_ONLINE: 4152 ret = slab_mem_going_online_callback(arg); 4153 break; 4154 case MEM_GOING_OFFLINE: 4155 ret = slab_mem_going_offline_callback(arg); 4156 break; 4157 case MEM_OFFLINE: 4158 case MEM_CANCEL_ONLINE: 4159 slab_mem_offline_callback(arg); 4160 break; 4161 case MEM_ONLINE: 4162 case MEM_CANCEL_OFFLINE: 4163 break; 4164 } 4165 if (ret) 4166 ret = notifier_from_errno(ret); 4167 else 4168 ret = NOTIFY_OK; 4169 return ret; 4170 } 4171 4172 static struct notifier_block slab_memory_callback_nb = { 4173 .notifier_call = slab_memory_callback, 4174 .priority = SLAB_CALLBACK_PRI, 4175 }; 4176 4177 /******************************************************************** 4178 * Basic setup of slabs 4179 *******************************************************************/ 4180 4181 /* 4182 * Used for early kmem_cache structures that were allocated using 4183 * the page allocator. Allocate them properly then fix up the pointers 4184 * that may be pointing to the wrong kmem_cache structure. 4185 */ 4186 4187 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) 4188 { 4189 int node; 4190 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); 4191 struct kmem_cache_node *n; 4192 4193 memcpy(s, static_cache, kmem_cache->object_size); 4194 4195 /* 4196 * This runs very early, and only the boot processor is supposed to be 4197 * up. Even if it weren't true, IRQs are not up so we couldn't fire 4198 * IPIs around. 4199 */ 4200 __flush_cpu_slab(s, smp_processor_id()); 4201 for_each_kmem_cache_node(s, node, n) { 4202 struct page *p; 4203 4204 list_for_each_entry(p, &n->partial, lru) 4205 p->slab_cache = s; 4206 4207 #ifdef CONFIG_SLUB_DEBUG 4208 list_for_each_entry(p, &n->full, lru) 4209 p->slab_cache = s; 4210 #endif 4211 } 4212 slab_init_memcg_params(s); 4213 list_add(&s->list, &slab_caches); 4214 memcg_link_cache(s); 4215 return s; 4216 } 4217 4218 void __init kmem_cache_init(void) 4219 { 4220 static __initdata struct kmem_cache boot_kmem_cache, 4221 boot_kmem_cache_node; 4222 4223 if (debug_guardpage_minorder()) 4224 slub_max_order = 0; 4225 4226 kmem_cache_node = &boot_kmem_cache_node; 4227 kmem_cache = &boot_kmem_cache; 4228 4229 create_boot_cache(kmem_cache_node, "kmem_cache_node", 4230 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0); 4231 4232 register_hotmemory_notifier(&slab_memory_callback_nb); 4233 4234 /* Able to allocate the per node structures */ 4235 slab_state = PARTIAL; 4236 4237 create_boot_cache(kmem_cache, "kmem_cache", 4238 offsetof(struct kmem_cache, node) + 4239 nr_node_ids * sizeof(struct kmem_cache_node *), 4240 SLAB_HWCACHE_ALIGN, 0, 0); 4241 4242 kmem_cache = bootstrap(&boot_kmem_cache); 4243 kmem_cache_node = bootstrap(&boot_kmem_cache_node); 4244 4245 /* Now we can use the kmem_cache to allocate kmalloc slabs */ 4246 setup_kmalloc_cache_index_table(); 4247 create_kmalloc_caches(0); 4248 4249 /* Setup random freelists for each cache */ 4250 init_freelist_randomization(); 4251 4252 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL, 4253 slub_cpu_dead); 4254 4255 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%d\n", 4256 cache_line_size(), 4257 slub_min_order, slub_max_order, slub_min_objects, 4258 nr_cpu_ids, nr_node_ids); 4259 } 4260 4261 void __init kmem_cache_init_late(void) 4262 { 4263 } 4264 4265 struct kmem_cache * 4266 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, 4267 slab_flags_t flags, void (*ctor)(void *)) 4268 { 4269 struct kmem_cache *s, *c; 4270 4271 s = find_mergeable(size, align, flags, name, ctor); 4272 if (s) { 4273 s->refcount++; 4274 4275 /* 4276 * Adjust the object sizes so that we clear 4277 * the complete object on kzalloc. 4278 */ 4279 s->object_size = max(s->object_size, size); 4280 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *))); 4281 4282 for_each_memcg_cache(c, s) { 4283 c->object_size = s->object_size; 4284 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *))); 4285 } 4286 4287 if (sysfs_slab_alias(s, name)) { 4288 s->refcount--; 4289 s = NULL; 4290 } 4291 } 4292 4293 return s; 4294 } 4295 4296 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) 4297 { 4298 int err; 4299 4300 err = kmem_cache_open(s, flags); 4301 if (err) 4302 return err; 4303 4304 /* Mutex is not taken during early boot */ 4305 if (slab_state <= UP) 4306 return 0; 4307 4308 memcg_propagate_slab_attrs(s); 4309 err = sysfs_slab_add(s); 4310 if (err) 4311 __kmem_cache_release(s); 4312 4313 return err; 4314 } 4315 4316 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller) 4317 { 4318 struct kmem_cache *s; 4319 void *ret; 4320 4321 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) 4322 return kmalloc_large(size, gfpflags); 4323 4324 s = kmalloc_slab(size, gfpflags); 4325 4326 if (unlikely(ZERO_OR_NULL_PTR(s))) 4327 return s; 4328 4329 ret = slab_alloc(s, gfpflags, caller); 4330 4331 /* Honor the call site pointer we received. */ 4332 trace_kmalloc(caller, ret, size, s->size, gfpflags); 4333 4334 return ret; 4335 } 4336 4337 #ifdef CONFIG_NUMA 4338 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags, 4339 int node, unsigned long caller) 4340 { 4341 struct kmem_cache *s; 4342 void *ret; 4343 4344 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { 4345 ret = kmalloc_large_node(size, gfpflags, node); 4346 4347 trace_kmalloc_node(caller, ret, 4348 size, PAGE_SIZE << get_order(size), 4349 gfpflags, node); 4350 4351 return ret; 4352 } 4353 4354 s = kmalloc_slab(size, gfpflags); 4355 4356 if (unlikely(ZERO_OR_NULL_PTR(s))) 4357 return s; 4358 4359 ret = slab_alloc_node(s, gfpflags, node, caller); 4360 4361 /* Honor the call site pointer we received. */ 4362 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node); 4363 4364 return ret; 4365 } 4366 #endif 4367 4368 #ifdef CONFIG_SYSFS 4369 static int count_inuse(struct page *page) 4370 { 4371 return page->inuse; 4372 } 4373 4374 static int count_total(struct page *page) 4375 { 4376 return page->objects; 4377 } 4378 #endif 4379 4380 #ifdef CONFIG_SLUB_DEBUG 4381 static int validate_slab(struct kmem_cache *s, struct page *page, 4382 unsigned long *map) 4383 { 4384 void *p; 4385 void *addr = page_address(page); 4386 4387 if (!check_slab(s, page) || 4388 !on_freelist(s, page, NULL)) 4389 return 0; 4390 4391 /* Now we know that a valid freelist exists */ 4392 bitmap_zero(map, page->objects); 4393 4394 get_map(s, page, map); 4395 for_each_object(p, s, addr, page->objects) { 4396 if (test_bit(slab_index(p, s, addr), map)) 4397 if (!check_object(s, page, p, SLUB_RED_INACTIVE)) 4398 return 0; 4399 } 4400 4401 for_each_object(p, s, addr, page->objects) 4402 if (!test_bit(slab_index(p, s, addr), map)) 4403 if (!check_object(s, page, p, SLUB_RED_ACTIVE)) 4404 return 0; 4405 return 1; 4406 } 4407 4408 static void validate_slab_slab(struct kmem_cache *s, struct page *page, 4409 unsigned long *map) 4410 { 4411 slab_lock(page); 4412 validate_slab(s, page, map); 4413 slab_unlock(page); 4414 } 4415 4416 static int validate_slab_node(struct kmem_cache *s, 4417 struct kmem_cache_node *n, unsigned long *map) 4418 { 4419 unsigned long count = 0; 4420 struct page *page; 4421 unsigned long flags; 4422 4423 spin_lock_irqsave(&n->list_lock, flags); 4424 4425 list_for_each_entry(page, &n->partial, lru) { 4426 validate_slab_slab(s, page, map); 4427 count++; 4428 } 4429 if (count != n->nr_partial) 4430 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n", 4431 s->name, count, n->nr_partial); 4432 4433 if (!(s->flags & SLAB_STORE_USER)) 4434 goto out; 4435 4436 list_for_each_entry(page, &n->full, lru) { 4437 validate_slab_slab(s, page, map); 4438 count++; 4439 } 4440 if (count != atomic_long_read(&n->nr_slabs)) 4441 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n", 4442 s->name, count, atomic_long_read(&n->nr_slabs)); 4443 4444 out: 4445 spin_unlock_irqrestore(&n->list_lock, flags); 4446 return count; 4447 } 4448 4449 static long validate_slab_cache(struct kmem_cache *s) 4450 { 4451 int node; 4452 unsigned long count = 0; 4453 struct kmem_cache_node *n; 4454 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL); 4455 4456 if (!map) 4457 return -ENOMEM; 4458 4459 flush_all(s); 4460 for_each_kmem_cache_node(s, node, n) 4461 count += validate_slab_node(s, n, map); 4462 bitmap_free(map); 4463 return count; 4464 } 4465 /* 4466 * Generate lists of code addresses where slabcache objects are allocated 4467 * and freed. 4468 */ 4469 4470 struct location { 4471 unsigned long count; 4472 unsigned long addr; 4473 long long sum_time; 4474 long min_time; 4475 long max_time; 4476 long min_pid; 4477 long max_pid; 4478 DECLARE_BITMAP(cpus, NR_CPUS); 4479 nodemask_t nodes; 4480 }; 4481 4482 struct loc_track { 4483 unsigned long max; 4484 unsigned long count; 4485 struct location *loc; 4486 }; 4487 4488 static void free_loc_track(struct loc_track *t) 4489 { 4490 if (t->max) 4491 free_pages((unsigned long)t->loc, 4492 get_order(sizeof(struct location) * t->max)); 4493 } 4494 4495 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) 4496 { 4497 struct location *l; 4498 int order; 4499 4500 order = get_order(sizeof(struct location) * max); 4501 4502 l = (void *)__get_free_pages(flags, order); 4503 if (!l) 4504 return 0; 4505 4506 if (t->count) { 4507 memcpy(l, t->loc, sizeof(struct location) * t->count); 4508 free_loc_track(t); 4509 } 4510 t->max = max; 4511 t->loc = l; 4512 return 1; 4513 } 4514 4515 static int add_location(struct loc_track *t, struct kmem_cache *s, 4516 const struct track *track) 4517 { 4518 long start, end, pos; 4519 struct location *l; 4520 unsigned long caddr; 4521 unsigned long age = jiffies - track->when; 4522 4523 start = -1; 4524 end = t->count; 4525 4526 for ( ; ; ) { 4527 pos = start + (end - start + 1) / 2; 4528 4529 /* 4530 * There is nothing at "end". If we end up there 4531 * we need to add something to before end. 4532 */ 4533 if (pos == end) 4534 break; 4535 4536 caddr = t->loc[pos].addr; 4537 if (track->addr == caddr) { 4538 4539 l = &t->loc[pos]; 4540 l->count++; 4541 if (track->when) { 4542 l->sum_time += age; 4543 if (age < l->min_time) 4544 l->min_time = age; 4545 if (age > l->max_time) 4546 l->max_time = age; 4547 4548 if (track->pid < l->min_pid) 4549 l->min_pid = track->pid; 4550 if (track->pid > l->max_pid) 4551 l->max_pid = track->pid; 4552 4553 cpumask_set_cpu(track->cpu, 4554 to_cpumask(l->cpus)); 4555 } 4556 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4557 return 1; 4558 } 4559 4560 if (track->addr < caddr) 4561 end = pos; 4562 else 4563 start = pos; 4564 } 4565 4566 /* 4567 * Not found. Insert new tracking element. 4568 */ 4569 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) 4570 return 0; 4571 4572 l = t->loc + pos; 4573 if (pos < t->count) 4574 memmove(l + 1, l, 4575 (t->count - pos) * sizeof(struct location)); 4576 t->count++; 4577 l->count = 1; 4578 l->addr = track->addr; 4579 l->sum_time = age; 4580 l->min_time = age; 4581 l->max_time = age; 4582 l->min_pid = track->pid; 4583 l->max_pid = track->pid; 4584 cpumask_clear(to_cpumask(l->cpus)); 4585 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); 4586 nodes_clear(l->nodes); 4587 node_set(page_to_nid(virt_to_page(track)), l->nodes); 4588 return 1; 4589 } 4590 4591 static void process_slab(struct loc_track *t, struct kmem_cache *s, 4592 struct page *page, enum track_item alloc, 4593 unsigned long *map) 4594 { 4595 void *addr = page_address(page); 4596 void *p; 4597 4598 bitmap_zero(map, page->objects); 4599 get_map(s, page, map); 4600 4601 for_each_object(p, s, addr, page->objects) 4602 if (!test_bit(slab_index(p, s, addr), map)) 4603 add_location(t, s, get_track(s, p, alloc)); 4604 } 4605 4606 static int list_locations(struct kmem_cache *s, char *buf, 4607 enum track_item alloc) 4608 { 4609 int len = 0; 4610 unsigned long i; 4611 struct loc_track t = { 0, 0, NULL }; 4612 int node; 4613 struct kmem_cache_node *n; 4614 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL); 4615 4616 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location), 4617 GFP_KERNEL)) { 4618 bitmap_free(map); 4619 return sprintf(buf, "Out of memory\n"); 4620 } 4621 /* Push back cpu slabs */ 4622 flush_all(s); 4623 4624 for_each_kmem_cache_node(s, node, n) { 4625 unsigned long flags; 4626 struct page *page; 4627 4628 if (!atomic_long_read(&n->nr_slabs)) 4629 continue; 4630 4631 spin_lock_irqsave(&n->list_lock, flags); 4632 list_for_each_entry(page, &n->partial, lru) 4633 process_slab(&t, s, page, alloc, map); 4634 list_for_each_entry(page, &n->full, lru) 4635 process_slab(&t, s, page, alloc, map); 4636 spin_unlock_irqrestore(&n->list_lock, flags); 4637 } 4638 4639 for (i = 0; i < t.count; i++) { 4640 struct location *l = &t.loc[i]; 4641 4642 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100) 4643 break; 4644 len += sprintf(buf + len, "%7ld ", l->count); 4645 4646 if (l->addr) 4647 len += sprintf(buf + len, "%pS", (void *)l->addr); 4648 else 4649 len += sprintf(buf + len, "<not-available>"); 4650 4651 if (l->sum_time != l->min_time) { 4652 len += sprintf(buf + len, " age=%ld/%ld/%ld", 4653 l->min_time, 4654 (long)div_u64(l->sum_time, l->count), 4655 l->max_time); 4656 } else 4657 len += sprintf(buf + len, " age=%ld", 4658 l->min_time); 4659 4660 if (l->min_pid != l->max_pid) 4661 len += sprintf(buf + len, " pid=%ld-%ld", 4662 l->min_pid, l->max_pid); 4663 else 4664 len += sprintf(buf + len, " pid=%ld", 4665 l->min_pid); 4666 4667 if (num_online_cpus() > 1 && 4668 !cpumask_empty(to_cpumask(l->cpus)) && 4669 len < PAGE_SIZE - 60) 4670 len += scnprintf(buf + len, PAGE_SIZE - len - 50, 4671 " cpus=%*pbl", 4672 cpumask_pr_args(to_cpumask(l->cpus))); 4673 4674 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) && 4675 len < PAGE_SIZE - 60) 4676 len += scnprintf(buf + len, PAGE_SIZE - len - 50, 4677 " nodes=%*pbl", 4678 nodemask_pr_args(&l->nodes)); 4679 4680 len += sprintf(buf + len, "\n"); 4681 } 4682 4683 free_loc_track(&t); 4684 bitmap_free(map); 4685 if (!t.count) 4686 len += sprintf(buf, "No data\n"); 4687 return len; 4688 } 4689 #endif 4690 4691 #ifdef SLUB_RESILIENCY_TEST 4692 static void __init resiliency_test(void) 4693 { 4694 u8 *p; 4695 int type = KMALLOC_NORMAL; 4696 4697 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10); 4698 4699 pr_err("SLUB resiliency testing\n"); 4700 pr_err("-----------------------\n"); 4701 pr_err("A. Corruption after allocation\n"); 4702 4703 p = kzalloc(16, GFP_KERNEL); 4704 p[16] = 0x12; 4705 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n", 4706 p + 16); 4707 4708 validate_slab_cache(kmalloc_caches[type][4]); 4709 4710 /* Hmmm... The next two are dangerous */ 4711 p = kzalloc(32, GFP_KERNEL); 4712 p[32 + sizeof(void *)] = 0x34; 4713 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n", 4714 p); 4715 pr_err("If allocated object is overwritten then not detectable\n\n"); 4716 4717 validate_slab_cache(kmalloc_caches[type][5]); 4718 p = kzalloc(64, GFP_KERNEL); 4719 p += 64 + (get_cycles() & 0xff) * sizeof(void *); 4720 *p = 0x56; 4721 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n", 4722 p); 4723 pr_err("If allocated object is overwritten then not detectable\n\n"); 4724 validate_slab_cache(kmalloc_caches[type][6]); 4725 4726 pr_err("\nB. Corruption after free\n"); 4727 p = kzalloc(128, GFP_KERNEL); 4728 kfree(p); 4729 *p = 0x78; 4730 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p); 4731 validate_slab_cache(kmalloc_caches[type][7]); 4732 4733 p = kzalloc(256, GFP_KERNEL); 4734 kfree(p); 4735 p[50] = 0x9a; 4736 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p); 4737 validate_slab_cache(kmalloc_caches[type][8]); 4738 4739 p = kzalloc(512, GFP_KERNEL); 4740 kfree(p); 4741 p[512] = 0xab; 4742 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p); 4743 validate_slab_cache(kmalloc_caches[type][9]); 4744 } 4745 #else 4746 #ifdef CONFIG_SYSFS 4747 static void resiliency_test(void) {}; 4748 #endif 4749 #endif 4750 4751 #ifdef CONFIG_SYSFS 4752 enum slab_stat_type { 4753 SL_ALL, /* All slabs */ 4754 SL_PARTIAL, /* Only partially allocated slabs */ 4755 SL_CPU, /* Only slabs used for cpu caches */ 4756 SL_OBJECTS, /* Determine allocated objects not slabs */ 4757 SL_TOTAL /* Determine object capacity not slabs */ 4758 }; 4759 4760 #define SO_ALL (1 << SL_ALL) 4761 #define SO_PARTIAL (1 << SL_PARTIAL) 4762 #define SO_CPU (1 << SL_CPU) 4763 #define SO_OBJECTS (1 << SL_OBJECTS) 4764 #define SO_TOTAL (1 << SL_TOTAL) 4765 4766 #ifdef CONFIG_MEMCG 4767 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON); 4768 4769 static int __init setup_slub_memcg_sysfs(char *str) 4770 { 4771 int v; 4772 4773 if (get_option(&str, &v) > 0) 4774 memcg_sysfs_enabled = v; 4775 4776 return 1; 4777 } 4778 4779 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs); 4780 #endif 4781 4782 static ssize_t show_slab_objects(struct kmem_cache *s, 4783 char *buf, unsigned long flags) 4784 { 4785 unsigned long total = 0; 4786 int node; 4787 int x; 4788 unsigned long *nodes; 4789 4790 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL); 4791 if (!nodes) 4792 return -ENOMEM; 4793 4794 if (flags & SO_CPU) { 4795 int cpu; 4796 4797 for_each_possible_cpu(cpu) { 4798 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, 4799 cpu); 4800 int node; 4801 struct page *page; 4802 4803 page = READ_ONCE(c->page); 4804 if (!page) 4805 continue; 4806 4807 node = page_to_nid(page); 4808 if (flags & SO_TOTAL) 4809 x = page->objects; 4810 else if (flags & SO_OBJECTS) 4811 x = page->inuse; 4812 else 4813 x = 1; 4814 4815 total += x; 4816 nodes[node] += x; 4817 4818 page = slub_percpu_partial_read_once(c); 4819 if (page) { 4820 node = page_to_nid(page); 4821 if (flags & SO_TOTAL) 4822 WARN_ON_ONCE(1); 4823 else if (flags & SO_OBJECTS) 4824 WARN_ON_ONCE(1); 4825 else 4826 x = page->pages; 4827 total += x; 4828 nodes[node] += x; 4829 } 4830 } 4831 } 4832 4833 get_online_mems(); 4834 #ifdef CONFIG_SLUB_DEBUG 4835 if (flags & SO_ALL) { 4836 struct kmem_cache_node *n; 4837 4838 for_each_kmem_cache_node(s, node, n) { 4839 4840 if (flags & SO_TOTAL) 4841 x = atomic_long_read(&n->total_objects); 4842 else if (flags & SO_OBJECTS) 4843 x = atomic_long_read(&n->total_objects) - 4844 count_partial(n, count_free); 4845 else 4846 x = atomic_long_read(&n->nr_slabs); 4847 total += x; 4848 nodes[node] += x; 4849 } 4850 4851 } else 4852 #endif 4853 if (flags & SO_PARTIAL) { 4854 struct kmem_cache_node *n; 4855 4856 for_each_kmem_cache_node(s, node, n) { 4857 if (flags & SO_TOTAL) 4858 x = count_partial(n, count_total); 4859 else if (flags & SO_OBJECTS) 4860 x = count_partial(n, count_inuse); 4861 else 4862 x = n->nr_partial; 4863 total += x; 4864 nodes[node] += x; 4865 } 4866 } 4867 x = sprintf(buf, "%lu", total); 4868 #ifdef CONFIG_NUMA 4869 for (node = 0; node < nr_node_ids; node++) 4870 if (nodes[node]) 4871 x += sprintf(buf + x, " N%d=%lu", 4872 node, nodes[node]); 4873 #endif 4874 put_online_mems(); 4875 kfree(nodes); 4876 return x + sprintf(buf + x, "\n"); 4877 } 4878 4879 #ifdef CONFIG_SLUB_DEBUG 4880 static int any_slab_objects(struct kmem_cache *s) 4881 { 4882 int node; 4883 struct kmem_cache_node *n; 4884 4885 for_each_kmem_cache_node(s, node, n) 4886 if (atomic_long_read(&n->total_objects)) 4887 return 1; 4888 4889 return 0; 4890 } 4891 #endif 4892 4893 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) 4894 #define to_slab(n) container_of(n, struct kmem_cache, kobj) 4895 4896 struct slab_attribute { 4897 struct attribute attr; 4898 ssize_t (*show)(struct kmem_cache *s, char *buf); 4899 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); 4900 }; 4901 4902 #define SLAB_ATTR_RO(_name) \ 4903 static struct slab_attribute _name##_attr = \ 4904 __ATTR(_name, 0400, _name##_show, NULL) 4905 4906 #define SLAB_ATTR(_name) \ 4907 static struct slab_attribute _name##_attr = \ 4908 __ATTR(_name, 0600, _name##_show, _name##_store) 4909 4910 static ssize_t slab_size_show(struct kmem_cache *s, char *buf) 4911 { 4912 return sprintf(buf, "%u\n", s->size); 4913 } 4914 SLAB_ATTR_RO(slab_size); 4915 4916 static ssize_t align_show(struct kmem_cache *s, char *buf) 4917 { 4918 return sprintf(buf, "%u\n", s->align); 4919 } 4920 SLAB_ATTR_RO(align); 4921 4922 static ssize_t object_size_show(struct kmem_cache *s, char *buf) 4923 { 4924 return sprintf(buf, "%u\n", s->object_size); 4925 } 4926 SLAB_ATTR_RO(object_size); 4927 4928 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) 4929 { 4930 return sprintf(buf, "%u\n", oo_objects(s->oo)); 4931 } 4932 SLAB_ATTR_RO(objs_per_slab); 4933 4934 static ssize_t order_store(struct kmem_cache *s, 4935 const char *buf, size_t length) 4936 { 4937 unsigned int order; 4938 int err; 4939 4940 err = kstrtouint(buf, 10, &order); 4941 if (err) 4942 return err; 4943 4944 if (order > slub_max_order || order < slub_min_order) 4945 return -EINVAL; 4946 4947 calculate_sizes(s, order); 4948 return length; 4949 } 4950 4951 static ssize_t order_show(struct kmem_cache *s, char *buf) 4952 { 4953 return sprintf(buf, "%u\n", oo_order(s->oo)); 4954 } 4955 SLAB_ATTR(order); 4956 4957 static ssize_t min_partial_show(struct kmem_cache *s, char *buf) 4958 { 4959 return sprintf(buf, "%lu\n", s->min_partial); 4960 } 4961 4962 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, 4963 size_t length) 4964 { 4965 unsigned long min; 4966 int err; 4967 4968 err = kstrtoul(buf, 10, &min); 4969 if (err) 4970 return err; 4971 4972 set_min_partial(s, min); 4973 return length; 4974 } 4975 SLAB_ATTR(min_partial); 4976 4977 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) 4978 { 4979 return sprintf(buf, "%u\n", slub_cpu_partial(s)); 4980 } 4981 4982 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, 4983 size_t length) 4984 { 4985 unsigned int objects; 4986 int err; 4987 4988 err = kstrtouint(buf, 10, &objects); 4989 if (err) 4990 return err; 4991 if (objects && !kmem_cache_has_cpu_partial(s)) 4992 return -EINVAL; 4993 4994 slub_set_cpu_partial(s, objects); 4995 flush_all(s); 4996 return length; 4997 } 4998 SLAB_ATTR(cpu_partial); 4999 5000 static ssize_t ctor_show(struct kmem_cache *s, char *buf) 5001 { 5002 if (!s->ctor) 5003 return 0; 5004 return sprintf(buf, "%pS\n", s->ctor); 5005 } 5006 SLAB_ATTR_RO(ctor); 5007 5008 static ssize_t aliases_show(struct kmem_cache *s, char *buf) 5009 { 5010 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1); 5011 } 5012 SLAB_ATTR_RO(aliases); 5013 5014 static ssize_t partial_show(struct kmem_cache *s, char *buf) 5015 { 5016 return show_slab_objects(s, buf, SO_PARTIAL); 5017 } 5018 SLAB_ATTR_RO(partial); 5019 5020 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) 5021 { 5022 return show_slab_objects(s, buf, SO_CPU); 5023 } 5024 SLAB_ATTR_RO(cpu_slabs); 5025 5026 static ssize_t objects_show(struct kmem_cache *s, char *buf) 5027 { 5028 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); 5029 } 5030 SLAB_ATTR_RO(objects); 5031 5032 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) 5033 { 5034 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); 5035 } 5036 SLAB_ATTR_RO(objects_partial); 5037 5038 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) 5039 { 5040 int objects = 0; 5041 int pages = 0; 5042 int cpu; 5043 int len; 5044 5045 for_each_online_cpu(cpu) { 5046 struct page *page; 5047 5048 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); 5049 5050 if (page) { 5051 pages += page->pages; 5052 objects += page->pobjects; 5053 } 5054 } 5055 5056 len = sprintf(buf, "%d(%d)", objects, pages); 5057 5058 #ifdef CONFIG_SMP 5059 for_each_online_cpu(cpu) { 5060 struct page *page; 5061 5062 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); 5063 5064 if (page && len < PAGE_SIZE - 20) 5065 len += sprintf(buf + len, " C%d=%d(%d)", cpu, 5066 page->pobjects, page->pages); 5067 } 5068 #endif 5069 return len + sprintf(buf + len, "\n"); 5070 } 5071 SLAB_ATTR_RO(slabs_cpu_partial); 5072 5073 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) 5074 { 5075 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); 5076 } 5077 5078 static ssize_t reclaim_account_store(struct kmem_cache *s, 5079 const char *buf, size_t length) 5080 { 5081 s->flags &= ~SLAB_RECLAIM_ACCOUNT; 5082 if (buf[0] == '1') 5083 s->flags |= SLAB_RECLAIM_ACCOUNT; 5084 return length; 5085 } 5086 SLAB_ATTR(reclaim_account); 5087 5088 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) 5089 { 5090 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); 5091 } 5092 SLAB_ATTR_RO(hwcache_align); 5093 5094 #ifdef CONFIG_ZONE_DMA 5095 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) 5096 { 5097 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); 5098 } 5099 SLAB_ATTR_RO(cache_dma); 5100 #endif 5101 5102 static ssize_t usersize_show(struct kmem_cache *s, char *buf) 5103 { 5104 return sprintf(buf, "%u\n", s->usersize); 5105 } 5106 SLAB_ATTR_RO(usersize); 5107 5108 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) 5109 { 5110 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU)); 5111 } 5112 SLAB_ATTR_RO(destroy_by_rcu); 5113 5114 #ifdef CONFIG_SLUB_DEBUG 5115 static ssize_t slabs_show(struct kmem_cache *s, char *buf) 5116 { 5117 return show_slab_objects(s, buf, SO_ALL); 5118 } 5119 SLAB_ATTR_RO(slabs); 5120 5121 static ssize_t total_objects_show(struct kmem_cache *s, char *buf) 5122 { 5123 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); 5124 } 5125 SLAB_ATTR_RO(total_objects); 5126 5127 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) 5128 { 5129 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS)); 5130 } 5131 5132 static ssize_t sanity_checks_store(struct kmem_cache *s, 5133 const char *buf, size_t length) 5134 { 5135 s->flags &= ~SLAB_CONSISTENCY_CHECKS; 5136 if (buf[0] == '1') { 5137 s->flags &= ~__CMPXCHG_DOUBLE; 5138 s->flags |= SLAB_CONSISTENCY_CHECKS; 5139 } 5140 return length; 5141 } 5142 SLAB_ATTR(sanity_checks); 5143 5144 static ssize_t trace_show(struct kmem_cache *s, char *buf) 5145 { 5146 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE)); 5147 } 5148 5149 static ssize_t trace_store(struct kmem_cache *s, const char *buf, 5150 size_t length) 5151 { 5152 /* 5153 * Tracing a merged cache is going to give confusing results 5154 * as well as cause other issues like converting a mergeable 5155 * cache into an umergeable one. 5156 */ 5157 if (s->refcount > 1) 5158 return -EINVAL; 5159 5160 s->flags &= ~SLAB_TRACE; 5161 if (buf[0] == '1') { 5162 s->flags &= ~__CMPXCHG_DOUBLE; 5163 s->flags |= SLAB_TRACE; 5164 } 5165 return length; 5166 } 5167 SLAB_ATTR(trace); 5168 5169 static ssize_t red_zone_show(struct kmem_cache *s, char *buf) 5170 { 5171 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); 5172 } 5173 5174 static ssize_t red_zone_store(struct kmem_cache *s, 5175 const char *buf, size_t length) 5176 { 5177 if (any_slab_objects(s)) 5178 return -EBUSY; 5179 5180 s->flags &= ~SLAB_RED_ZONE; 5181 if (buf[0] == '1') { 5182 s->flags |= SLAB_RED_ZONE; 5183 } 5184 calculate_sizes(s, -1); 5185 return length; 5186 } 5187 SLAB_ATTR(red_zone); 5188 5189 static ssize_t poison_show(struct kmem_cache *s, char *buf) 5190 { 5191 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON)); 5192 } 5193 5194 static ssize_t poison_store(struct kmem_cache *s, 5195 const char *buf, size_t length) 5196 { 5197 if (any_slab_objects(s)) 5198 return -EBUSY; 5199 5200 s->flags &= ~SLAB_POISON; 5201 if (buf[0] == '1') { 5202 s->flags |= SLAB_POISON; 5203 } 5204 calculate_sizes(s, -1); 5205 return length; 5206 } 5207 SLAB_ATTR(poison); 5208 5209 static ssize_t store_user_show(struct kmem_cache *s, char *buf) 5210 { 5211 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); 5212 } 5213 5214 static ssize_t store_user_store(struct kmem_cache *s, 5215 const char *buf, size_t length) 5216 { 5217 if (any_slab_objects(s)) 5218 return -EBUSY; 5219 5220 s->flags &= ~SLAB_STORE_USER; 5221 if (buf[0] == '1') { 5222 s->flags &= ~__CMPXCHG_DOUBLE; 5223 s->flags |= SLAB_STORE_USER; 5224 } 5225 calculate_sizes(s, -1); 5226 return length; 5227 } 5228 SLAB_ATTR(store_user); 5229 5230 static ssize_t validate_show(struct kmem_cache *s, char *buf) 5231 { 5232 return 0; 5233 } 5234 5235 static ssize_t validate_store(struct kmem_cache *s, 5236 const char *buf, size_t length) 5237 { 5238 int ret = -EINVAL; 5239 5240 if (buf[0] == '1') { 5241 ret = validate_slab_cache(s); 5242 if (ret >= 0) 5243 ret = length; 5244 } 5245 return ret; 5246 } 5247 SLAB_ATTR(validate); 5248 5249 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf) 5250 { 5251 if (!(s->flags & SLAB_STORE_USER)) 5252 return -ENOSYS; 5253 return list_locations(s, buf, TRACK_ALLOC); 5254 } 5255 SLAB_ATTR_RO(alloc_calls); 5256 5257 static ssize_t free_calls_show(struct kmem_cache *s, char *buf) 5258 { 5259 if (!(s->flags & SLAB_STORE_USER)) 5260 return -ENOSYS; 5261 return list_locations(s, buf, TRACK_FREE); 5262 } 5263 SLAB_ATTR_RO(free_calls); 5264 #endif /* CONFIG_SLUB_DEBUG */ 5265 5266 #ifdef CONFIG_FAILSLAB 5267 static ssize_t failslab_show(struct kmem_cache *s, char *buf) 5268 { 5269 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); 5270 } 5271 5272 static ssize_t failslab_store(struct kmem_cache *s, const char *buf, 5273 size_t length) 5274 { 5275 if (s->refcount > 1) 5276 return -EINVAL; 5277 5278 s->flags &= ~SLAB_FAILSLAB; 5279 if (buf[0] == '1') 5280 s->flags |= SLAB_FAILSLAB; 5281 return length; 5282 } 5283 SLAB_ATTR(failslab); 5284 #endif 5285 5286 static ssize_t shrink_show(struct kmem_cache *s, char *buf) 5287 { 5288 return 0; 5289 } 5290 5291 static ssize_t shrink_store(struct kmem_cache *s, 5292 const char *buf, size_t length) 5293 { 5294 if (buf[0] == '1') 5295 kmem_cache_shrink(s); 5296 else 5297 return -EINVAL; 5298 return length; 5299 } 5300 SLAB_ATTR(shrink); 5301 5302 #ifdef CONFIG_NUMA 5303 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) 5304 { 5305 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10); 5306 } 5307 5308 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, 5309 const char *buf, size_t length) 5310 { 5311 unsigned int ratio; 5312 int err; 5313 5314 err = kstrtouint(buf, 10, &ratio); 5315 if (err) 5316 return err; 5317 if (ratio > 100) 5318 return -ERANGE; 5319 5320 s->remote_node_defrag_ratio = ratio * 10; 5321 5322 return length; 5323 } 5324 SLAB_ATTR(remote_node_defrag_ratio); 5325 #endif 5326 5327 #ifdef CONFIG_SLUB_STATS 5328 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) 5329 { 5330 unsigned long sum = 0; 5331 int cpu; 5332 int len; 5333 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL); 5334 5335 if (!data) 5336 return -ENOMEM; 5337 5338 for_each_online_cpu(cpu) { 5339 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; 5340 5341 data[cpu] = x; 5342 sum += x; 5343 } 5344 5345 len = sprintf(buf, "%lu", sum); 5346 5347 #ifdef CONFIG_SMP 5348 for_each_online_cpu(cpu) { 5349 if (data[cpu] && len < PAGE_SIZE - 20) 5350 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]); 5351 } 5352 #endif 5353 kfree(data); 5354 return len + sprintf(buf + len, "\n"); 5355 } 5356 5357 static void clear_stat(struct kmem_cache *s, enum stat_item si) 5358 { 5359 int cpu; 5360 5361 for_each_online_cpu(cpu) 5362 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; 5363 } 5364 5365 #define STAT_ATTR(si, text) \ 5366 static ssize_t text##_show(struct kmem_cache *s, char *buf) \ 5367 { \ 5368 return show_stat(s, buf, si); \ 5369 } \ 5370 static ssize_t text##_store(struct kmem_cache *s, \ 5371 const char *buf, size_t length) \ 5372 { \ 5373 if (buf[0] != '0') \ 5374 return -EINVAL; \ 5375 clear_stat(s, si); \ 5376 return length; \ 5377 } \ 5378 SLAB_ATTR(text); \ 5379 5380 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); 5381 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); 5382 STAT_ATTR(FREE_FASTPATH, free_fastpath); 5383 STAT_ATTR(FREE_SLOWPATH, free_slowpath); 5384 STAT_ATTR(FREE_FROZEN, free_frozen); 5385 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); 5386 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); 5387 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); 5388 STAT_ATTR(ALLOC_SLAB, alloc_slab); 5389 STAT_ATTR(ALLOC_REFILL, alloc_refill); 5390 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); 5391 STAT_ATTR(FREE_SLAB, free_slab); 5392 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); 5393 STAT_ATTR(DEACTIVATE_FULL, deactivate_full); 5394 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); 5395 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); 5396 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); 5397 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); 5398 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); 5399 STAT_ATTR(ORDER_FALLBACK, order_fallback); 5400 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); 5401 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); 5402 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); 5403 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); 5404 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); 5405 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); 5406 #endif 5407 5408 static struct attribute *slab_attrs[] = { 5409 &slab_size_attr.attr, 5410 &object_size_attr.attr, 5411 &objs_per_slab_attr.attr, 5412 &order_attr.attr, 5413 &min_partial_attr.attr, 5414 &cpu_partial_attr.attr, 5415 &objects_attr.attr, 5416 &objects_partial_attr.attr, 5417 &partial_attr.attr, 5418 &cpu_slabs_attr.attr, 5419 &ctor_attr.attr, 5420 &aliases_attr.attr, 5421 &align_attr.attr, 5422 &hwcache_align_attr.attr, 5423 &reclaim_account_attr.attr, 5424 &destroy_by_rcu_attr.attr, 5425 &shrink_attr.attr, 5426 &slabs_cpu_partial_attr.attr, 5427 #ifdef CONFIG_SLUB_DEBUG 5428 &total_objects_attr.attr, 5429 &slabs_attr.attr, 5430 &sanity_checks_attr.attr, 5431 &trace_attr.attr, 5432 &red_zone_attr.attr, 5433 &poison_attr.attr, 5434 &store_user_attr.attr, 5435 &validate_attr.attr, 5436 &alloc_calls_attr.attr, 5437 &free_calls_attr.attr, 5438 #endif 5439 #ifdef CONFIG_ZONE_DMA 5440 &cache_dma_attr.attr, 5441 #endif 5442 #ifdef CONFIG_NUMA 5443 &remote_node_defrag_ratio_attr.attr, 5444 #endif 5445 #ifdef CONFIG_SLUB_STATS 5446 &alloc_fastpath_attr.attr, 5447 &alloc_slowpath_attr.attr, 5448 &free_fastpath_attr.attr, 5449 &free_slowpath_attr.attr, 5450 &free_frozen_attr.attr, 5451 &free_add_partial_attr.attr, 5452 &free_remove_partial_attr.attr, 5453 &alloc_from_partial_attr.attr, 5454 &alloc_slab_attr.attr, 5455 &alloc_refill_attr.attr, 5456 &alloc_node_mismatch_attr.attr, 5457 &free_slab_attr.attr, 5458 &cpuslab_flush_attr.attr, 5459 &deactivate_full_attr.attr, 5460 &deactivate_empty_attr.attr, 5461 &deactivate_to_head_attr.attr, 5462 &deactivate_to_tail_attr.attr, 5463 &deactivate_remote_frees_attr.attr, 5464 &deactivate_bypass_attr.attr, 5465 &order_fallback_attr.attr, 5466 &cmpxchg_double_fail_attr.attr, 5467 &cmpxchg_double_cpu_fail_attr.attr, 5468 &cpu_partial_alloc_attr.attr, 5469 &cpu_partial_free_attr.attr, 5470 &cpu_partial_node_attr.attr, 5471 &cpu_partial_drain_attr.attr, 5472 #endif 5473 #ifdef CONFIG_FAILSLAB 5474 &failslab_attr.attr, 5475 #endif 5476 &usersize_attr.attr, 5477 5478 NULL 5479 }; 5480 5481 static const struct attribute_group slab_attr_group = { 5482 .attrs = slab_attrs, 5483 }; 5484 5485 static ssize_t slab_attr_show(struct kobject *kobj, 5486 struct attribute *attr, 5487 char *buf) 5488 { 5489 struct slab_attribute *attribute; 5490 struct kmem_cache *s; 5491 int err; 5492 5493 attribute = to_slab_attr(attr); 5494 s = to_slab(kobj); 5495 5496 if (!attribute->show) 5497 return -EIO; 5498 5499 err = attribute->show(s, buf); 5500 5501 return err; 5502 } 5503 5504 static ssize_t slab_attr_store(struct kobject *kobj, 5505 struct attribute *attr, 5506 const char *buf, size_t len) 5507 { 5508 struct slab_attribute *attribute; 5509 struct kmem_cache *s; 5510 int err; 5511 5512 attribute = to_slab_attr(attr); 5513 s = to_slab(kobj); 5514 5515 if (!attribute->store) 5516 return -EIO; 5517 5518 err = attribute->store(s, buf, len); 5519 #ifdef CONFIG_MEMCG 5520 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) { 5521 struct kmem_cache *c; 5522 5523 mutex_lock(&slab_mutex); 5524 if (s->max_attr_size < len) 5525 s->max_attr_size = len; 5526 5527 /* 5528 * This is a best effort propagation, so this function's return 5529 * value will be determined by the parent cache only. This is 5530 * basically because not all attributes will have a well 5531 * defined semantics for rollbacks - most of the actions will 5532 * have permanent effects. 5533 * 5534 * Returning the error value of any of the children that fail 5535 * is not 100 % defined, in the sense that users seeing the 5536 * error code won't be able to know anything about the state of 5537 * the cache. 5538 * 5539 * Only returning the error code for the parent cache at least 5540 * has well defined semantics. The cache being written to 5541 * directly either failed or succeeded, in which case we loop 5542 * through the descendants with best-effort propagation. 5543 */ 5544 for_each_memcg_cache(c, s) 5545 attribute->store(c, buf, len); 5546 mutex_unlock(&slab_mutex); 5547 } 5548 #endif 5549 return err; 5550 } 5551 5552 static void memcg_propagate_slab_attrs(struct kmem_cache *s) 5553 { 5554 #ifdef CONFIG_MEMCG 5555 int i; 5556 char *buffer = NULL; 5557 struct kmem_cache *root_cache; 5558 5559 if (is_root_cache(s)) 5560 return; 5561 5562 root_cache = s->memcg_params.root_cache; 5563 5564 /* 5565 * This mean this cache had no attribute written. Therefore, no point 5566 * in copying default values around 5567 */ 5568 if (!root_cache->max_attr_size) 5569 return; 5570 5571 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) { 5572 char mbuf[64]; 5573 char *buf; 5574 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]); 5575 ssize_t len; 5576 5577 if (!attr || !attr->store || !attr->show) 5578 continue; 5579 5580 /* 5581 * It is really bad that we have to allocate here, so we will 5582 * do it only as a fallback. If we actually allocate, though, 5583 * we can just use the allocated buffer until the end. 5584 * 5585 * Most of the slub attributes will tend to be very small in 5586 * size, but sysfs allows buffers up to a page, so they can 5587 * theoretically happen. 5588 */ 5589 if (buffer) 5590 buf = buffer; 5591 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf)) 5592 buf = mbuf; 5593 else { 5594 buffer = (char *) get_zeroed_page(GFP_KERNEL); 5595 if (WARN_ON(!buffer)) 5596 continue; 5597 buf = buffer; 5598 } 5599 5600 len = attr->show(root_cache, buf); 5601 if (len > 0) 5602 attr->store(s, buf, len); 5603 } 5604 5605 if (buffer) 5606 free_page((unsigned long)buffer); 5607 #endif 5608 } 5609 5610 static void kmem_cache_release(struct kobject *k) 5611 { 5612 slab_kmem_cache_release(to_slab(k)); 5613 } 5614 5615 static const struct sysfs_ops slab_sysfs_ops = { 5616 .show = slab_attr_show, 5617 .store = slab_attr_store, 5618 }; 5619 5620 static struct kobj_type slab_ktype = { 5621 .sysfs_ops = &slab_sysfs_ops, 5622 .release = kmem_cache_release, 5623 }; 5624 5625 static int uevent_filter(struct kset *kset, struct kobject *kobj) 5626 { 5627 struct kobj_type *ktype = get_ktype(kobj); 5628 5629 if (ktype == &slab_ktype) 5630 return 1; 5631 return 0; 5632 } 5633 5634 static const struct kset_uevent_ops slab_uevent_ops = { 5635 .filter = uevent_filter, 5636 }; 5637 5638 static struct kset *slab_kset; 5639 5640 static inline struct kset *cache_kset(struct kmem_cache *s) 5641 { 5642 #ifdef CONFIG_MEMCG 5643 if (!is_root_cache(s)) 5644 return s->memcg_params.root_cache->memcg_kset; 5645 #endif 5646 return slab_kset; 5647 } 5648 5649 #define ID_STR_LENGTH 64 5650 5651 /* Create a unique string id for a slab cache: 5652 * 5653 * Format :[flags-]size 5654 */ 5655 static char *create_unique_id(struct kmem_cache *s) 5656 { 5657 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); 5658 char *p = name; 5659 5660 BUG_ON(!name); 5661 5662 *p++ = ':'; 5663 /* 5664 * First flags affecting slabcache operations. We will only 5665 * get here for aliasable slabs so we do not need to support 5666 * too many flags. The flags here must cover all flags that 5667 * are matched during merging to guarantee that the id is 5668 * unique. 5669 */ 5670 if (s->flags & SLAB_CACHE_DMA) 5671 *p++ = 'd'; 5672 if (s->flags & SLAB_RECLAIM_ACCOUNT) 5673 *p++ = 'a'; 5674 if (s->flags & SLAB_CONSISTENCY_CHECKS) 5675 *p++ = 'F'; 5676 if (s->flags & SLAB_ACCOUNT) 5677 *p++ = 'A'; 5678 if (p != name + 1) 5679 *p++ = '-'; 5680 p += sprintf(p, "%07u", s->size); 5681 5682 BUG_ON(p > name + ID_STR_LENGTH - 1); 5683 return name; 5684 } 5685 5686 static void sysfs_slab_remove_workfn(struct work_struct *work) 5687 { 5688 struct kmem_cache *s = 5689 container_of(work, struct kmem_cache, kobj_remove_work); 5690 5691 if (!s->kobj.state_in_sysfs) 5692 /* 5693 * For a memcg cache, this may be called during 5694 * deactivation and again on shutdown. Remove only once. 5695 * A cache is never shut down before deactivation is 5696 * complete, so no need to worry about synchronization. 5697 */ 5698 goto out; 5699 5700 #ifdef CONFIG_MEMCG 5701 kset_unregister(s->memcg_kset); 5702 #endif 5703 kobject_uevent(&s->kobj, KOBJ_REMOVE); 5704 out: 5705 kobject_put(&s->kobj); 5706 } 5707 5708 static int sysfs_slab_add(struct kmem_cache *s) 5709 { 5710 int err; 5711 const char *name; 5712 struct kset *kset = cache_kset(s); 5713 int unmergeable = slab_unmergeable(s); 5714 5715 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn); 5716 5717 if (!kset) { 5718 kobject_init(&s->kobj, &slab_ktype); 5719 return 0; 5720 } 5721 5722 if (!unmergeable && disable_higher_order_debug && 5723 (slub_debug & DEBUG_METADATA_FLAGS)) 5724 unmergeable = 1; 5725 5726 if (unmergeable) { 5727 /* 5728 * Slabcache can never be merged so we can use the name proper. 5729 * This is typically the case for debug situations. In that 5730 * case we can catch duplicate names easily. 5731 */ 5732 sysfs_remove_link(&slab_kset->kobj, s->name); 5733 name = s->name; 5734 } else { 5735 /* 5736 * Create a unique name for the slab as a target 5737 * for the symlinks. 5738 */ 5739 name = create_unique_id(s); 5740 } 5741 5742 s->kobj.kset = kset; 5743 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); 5744 if (err) 5745 goto out; 5746 5747 err = sysfs_create_group(&s->kobj, &slab_attr_group); 5748 if (err) 5749 goto out_del_kobj; 5750 5751 #ifdef CONFIG_MEMCG 5752 if (is_root_cache(s) && memcg_sysfs_enabled) { 5753 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj); 5754 if (!s->memcg_kset) { 5755 err = -ENOMEM; 5756 goto out_del_kobj; 5757 } 5758 } 5759 #endif 5760 5761 kobject_uevent(&s->kobj, KOBJ_ADD); 5762 if (!unmergeable) { 5763 /* Setup first alias */ 5764 sysfs_slab_alias(s, s->name); 5765 } 5766 out: 5767 if (!unmergeable) 5768 kfree(name); 5769 return err; 5770 out_del_kobj: 5771 kobject_del(&s->kobj); 5772 goto out; 5773 } 5774 5775 static void sysfs_slab_remove(struct kmem_cache *s) 5776 { 5777 if (slab_state < FULL) 5778 /* 5779 * Sysfs has not been setup yet so no need to remove the 5780 * cache from sysfs. 5781 */ 5782 return; 5783 5784 kobject_get(&s->kobj); 5785 schedule_work(&s->kobj_remove_work); 5786 } 5787 5788 void sysfs_slab_unlink(struct kmem_cache *s) 5789 { 5790 if (slab_state >= FULL) 5791 kobject_del(&s->kobj); 5792 } 5793 5794 void sysfs_slab_release(struct kmem_cache *s) 5795 { 5796 if (slab_state >= FULL) 5797 kobject_put(&s->kobj); 5798 } 5799 5800 /* 5801 * Need to buffer aliases during bootup until sysfs becomes 5802 * available lest we lose that information. 5803 */ 5804 struct saved_alias { 5805 struct kmem_cache *s; 5806 const char *name; 5807 struct saved_alias *next; 5808 }; 5809 5810 static struct saved_alias *alias_list; 5811 5812 static int sysfs_slab_alias(struct kmem_cache *s, const char *name) 5813 { 5814 struct saved_alias *al; 5815 5816 if (slab_state == FULL) { 5817 /* 5818 * If we have a leftover link then remove it. 5819 */ 5820 sysfs_remove_link(&slab_kset->kobj, name); 5821 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); 5822 } 5823 5824 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); 5825 if (!al) 5826 return -ENOMEM; 5827 5828 al->s = s; 5829 al->name = name; 5830 al->next = alias_list; 5831 alias_list = al; 5832 return 0; 5833 } 5834 5835 static int __init slab_sysfs_init(void) 5836 { 5837 struct kmem_cache *s; 5838 int err; 5839 5840 mutex_lock(&slab_mutex); 5841 5842 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj); 5843 if (!slab_kset) { 5844 mutex_unlock(&slab_mutex); 5845 pr_err("Cannot register slab subsystem.\n"); 5846 return -ENOSYS; 5847 } 5848 5849 slab_state = FULL; 5850 5851 list_for_each_entry(s, &slab_caches, list) { 5852 err = sysfs_slab_add(s); 5853 if (err) 5854 pr_err("SLUB: Unable to add boot slab %s to sysfs\n", 5855 s->name); 5856 } 5857 5858 while (alias_list) { 5859 struct saved_alias *al = alias_list; 5860 5861 alias_list = alias_list->next; 5862 err = sysfs_slab_alias(al->s, al->name); 5863 if (err) 5864 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n", 5865 al->name); 5866 kfree(al); 5867 } 5868 5869 mutex_unlock(&slab_mutex); 5870 resiliency_test(); 5871 return 0; 5872 } 5873 5874 __initcall(slab_sysfs_init); 5875 #endif /* CONFIG_SYSFS */ 5876 5877 /* 5878 * The /proc/slabinfo ABI 5879 */ 5880 #ifdef CONFIG_SLUB_DEBUG 5881 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) 5882 { 5883 unsigned long nr_slabs = 0; 5884 unsigned long nr_objs = 0; 5885 unsigned long nr_free = 0; 5886 int node; 5887 struct kmem_cache_node *n; 5888 5889 for_each_kmem_cache_node(s, node, n) { 5890 nr_slabs += node_nr_slabs(n); 5891 nr_objs += node_nr_objs(n); 5892 nr_free += count_partial(n, count_free); 5893 } 5894 5895 sinfo->active_objs = nr_objs - nr_free; 5896 sinfo->num_objs = nr_objs; 5897 sinfo->active_slabs = nr_slabs; 5898 sinfo->num_slabs = nr_slabs; 5899 sinfo->objects_per_slab = oo_objects(s->oo); 5900 sinfo->cache_order = oo_order(s->oo); 5901 } 5902 5903 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) 5904 { 5905 } 5906 5907 ssize_t slabinfo_write(struct file *file, const char __user *buffer, 5908 size_t count, loff_t *ppos) 5909 { 5910 return -EIO; 5911 } 5912 #endif /* CONFIG_SLUB_DEBUG */ 5913