subr_kmem.c revision 1.62.4.1
1 /* $NetBSD: subr_kmem.c,v 1.62.4.1 2017/04/21 16:54:02 bouyer Exp $ */ 2 3 /*- 4 * Copyright (c) 2009-2015 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Andrew Doran and Maxime Villard. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 29 * POSSIBILITY OF SUCH DAMAGE. 30 */ 31 32 /*- 33 * Copyright (c)2006 YAMAMOTO Takashi, 34 * All rights reserved. 35 * 36 * Redistribution and use in source and binary forms, with or without 37 * modification, are permitted provided that the following conditions 38 * are met: 39 * 1. Redistributions of source code must retain the above copyright 40 * notice, this list of conditions and the following disclaimer. 41 * 2. Redistributions in binary form must reproduce the above copyright 42 * notice, this list of conditions and the following disclaimer in the 43 * documentation and/or other materials provided with the distribution. 44 * 45 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 46 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 47 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 48 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 49 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 50 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 51 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 52 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 53 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 54 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 55 * SUCH DAMAGE. 56 */ 57 58 /* 59 * Allocator of kernel wired memory. This allocator has some debug features 60 * enabled with "option DIAGNOSTIC" and "option DEBUG". 61 */ 62 63 /* 64 * KMEM_SIZE: detect alloc/free size mismatch bugs. 65 * Prefix each allocations with a fixed-sized, aligned header and record 66 * the exact user-requested allocation size in it. When freeing, compare 67 * it with kmem_free's "size" argument. 68 * 69 * KMEM_REDZONE: detect overrun bugs. 70 * Add a 2-byte pattern (allocate one more memory chunk if needed) at the 71 * end of each allocated buffer. Check this pattern on kmem_free. 72 * 73 * These options are enabled on DIAGNOSTIC. 74 * 75 * |CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK| 76 * +-----+-----+-----+-----+-----+-----+-----+-----+-----+---+-+--+--+ 77 * |/////| | | | | | | | | |*|**|UU| 78 * |/HSZ/| | | | | | | | | |*|**|UU| 79 * |/////| | | | | | | | | |*|**|UU| 80 * +-----+-----+-----+-----+-----+-----+-----+-----+-----+---+-+--+--+ 81 * |Size | Buffer usable by the caller (requested size) |RedZ|Unused\ 82 */ 83 84 /* 85 * KMEM_POISON: detect modify-after-free bugs. 86 * Fill freed (in the sense of kmem_free) memory with a garbage pattern. 87 * Check the pattern on allocation. 88 * 89 * KMEM_GUARD 90 * A kernel with "option DEBUG" has "kmem_guard" debugging feature compiled 91 * in. See the comment below for what kind of bugs it tries to detect. Even 92 * if compiled in, it's disabled by default because it's very expensive. 93 * You can enable it on boot by: 94 * boot -d 95 * db> w kmem_guard_depth 0t30000 96 * db> c 97 * 98 * The default value of kmem_guard_depth is 0, which means disabled. 99 * It can be changed by KMEM_GUARD_DEPTH kernel config option. 100 */ 101 102 #include <sys/cdefs.h> 103 __KERNEL_RCSID(0, "$NetBSD: subr_kmem.c,v 1.62.4.1 2017/04/21 16:54:02 bouyer Exp $"); 104 105 #ifdef _KERNEL_OPT 106 #include "opt_kmem.h" 107 #endif 108 109 #include <sys/param.h> 110 #include <sys/callback.h> 111 #include <sys/kmem.h> 112 #include <sys/pool.h> 113 #include <sys/debug.h> 114 #include <sys/lockdebug.h> 115 #include <sys/cpu.h> 116 117 #include <uvm/uvm_extern.h> 118 #include <uvm/uvm_map.h> 119 120 #include <lib/libkern/libkern.h> 121 122 struct kmem_cache_info { 123 size_t kc_size; 124 const char * kc_name; 125 }; 126 127 static const struct kmem_cache_info kmem_cache_sizes[] = { 128 { 8, "kmem-8" }, 129 { 16, "kmem-16" }, 130 { 24, "kmem-24" }, 131 { 32, "kmem-32" }, 132 { 40, "kmem-40" }, 133 { 48, "kmem-48" }, 134 { 56, "kmem-56" }, 135 { 64, "kmem-64" }, 136 { 80, "kmem-80" }, 137 { 96, "kmem-96" }, 138 { 112, "kmem-112" }, 139 { 128, "kmem-128" }, 140 { 160, "kmem-160" }, 141 { 192, "kmem-192" }, 142 { 224, "kmem-224" }, 143 { 256, "kmem-256" }, 144 { 320, "kmem-320" }, 145 { 384, "kmem-384" }, 146 { 448, "kmem-448" }, 147 { 512, "kmem-512" }, 148 { 768, "kmem-768" }, 149 { 1024, "kmem-1024" }, 150 { 0, NULL } 151 }; 152 153 static const struct kmem_cache_info kmem_cache_big_sizes[] = { 154 { 2048, "kmem-2048" }, 155 { 4096, "kmem-4096" }, 156 { 8192, "kmem-8192" }, 157 { 16384, "kmem-16384" }, 158 { 0, NULL } 159 }; 160 161 /* 162 * KMEM_ALIGN is the smallest guaranteed alignment and also the 163 * smallest allocateable quantum. 164 * Every cache size >= CACHE_LINE_SIZE gets CACHE_LINE_SIZE alignment. 165 */ 166 #define KMEM_ALIGN 8 167 #define KMEM_SHIFT 3 168 #define KMEM_MAXSIZE 1024 169 #define KMEM_CACHE_COUNT (KMEM_MAXSIZE >> KMEM_SHIFT) 170 171 static pool_cache_t kmem_cache[KMEM_CACHE_COUNT] __cacheline_aligned; 172 static size_t kmem_cache_maxidx __read_mostly; 173 174 #define KMEM_BIG_ALIGN 2048 175 #define KMEM_BIG_SHIFT 11 176 #define KMEM_BIG_MAXSIZE 16384 177 #define KMEM_CACHE_BIG_COUNT (KMEM_BIG_MAXSIZE >> KMEM_BIG_SHIFT) 178 179 static pool_cache_t kmem_cache_big[KMEM_CACHE_BIG_COUNT] __cacheline_aligned; 180 static size_t kmem_cache_big_maxidx __read_mostly; 181 182 #if defined(DIAGNOSTIC) && defined(_HARDKERNEL) 183 #define KMEM_SIZE 184 #define KMEM_REDZONE 185 #endif /* defined(DIAGNOSTIC) */ 186 187 #if defined(DEBUG) && defined(_HARDKERNEL) 188 #define KMEM_SIZE 189 #define KMEM_POISON 190 #define KMEM_GUARD 191 static void *kmem_freecheck; 192 #endif /* defined(DEBUG) */ 193 194 #if defined(KMEM_POISON) 195 static int kmem_poison_ctor(void *, void *, int); 196 static void kmem_poison_fill(void *, size_t); 197 static void kmem_poison_check(void *, size_t); 198 #else /* defined(KMEM_POISON) */ 199 #define kmem_poison_fill(p, sz) /* nothing */ 200 #define kmem_poison_check(p, sz) /* nothing */ 201 #endif /* defined(KMEM_POISON) */ 202 203 #if defined(KMEM_REDZONE) 204 #define REDZONE_SIZE 2 205 static void kmem_redzone_fill(void *, size_t); 206 static void kmem_redzone_check(void *, size_t); 207 #else /* defined(KMEM_REDZONE) */ 208 #define REDZONE_SIZE 0 209 #define kmem_redzone_fill(p, sz) /* nothing */ 210 #define kmem_redzone_check(p, sz) /* nothing */ 211 #endif /* defined(KMEM_REDZONE) */ 212 213 #if defined(KMEM_SIZE) 214 struct kmem_header { 215 size_t size; 216 } __aligned(KMEM_ALIGN); 217 #define SIZE_SIZE sizeof(struct kmem_header) 218 static void kmem_size_set(void *, size_t); 219 static void kmem_size_check(void *, size_t); 220 #else 221 #define SIZE_SIZE 0 222 #define kmem_size_set(p, sz) /* nothing */ 223 #define kmem_size_check(p, sz) /* nothing */ 224 #endif 225 226 #if defined(KMEM_GUARD) 227 #ifndef KMEM_GUARD_DEPTH 228 #define KMEM_GUARD_DEPTH 0 229 #endif 230 struct kmem_guard { 231 u_int kg_depth; 232 intptr_t * kg_fifo; 233 u_int kg_rotor; 234 vmem_t * kg_vmem; 235 }; 236 237 static bool kmem_guard_init(struct kmem_guard *, u_int, vmem_t *); 238 static void *kmem_guard_alloc(struct kmem_guard *, size_t, bool); 239 static void kmem_guard_free(struct kmem_guard *, size_t, void *); 240 241 int kmem_guard_depth = KMEM_GUARD_DEPTH; 242 static bool kmem_guard_enabled; 243 static struct kmem_guard kmem_guard; 244 #endif /* defined(KMEM_GUARD) */ 245 246 CTASSERT(KM_SLEEP == PR_WAITOK); 247 CTASSERT(KM_NOSLEEP == PR_NOWAIT); 248 249 /* 250 * kmem_intr_alloc: allocate wired memory. 251 */ 252 253 void * 254 kmem_intr_alloc(size_t requested_size, km_flag_t kmflags) 255 { 256 size_t allocsz, index; 257 size_t size; 258 pool_cache_t pc; 259 uint8_t *p; 260 261 KASSERT(requested_size > 0); 262 263 #ifdef KMEM_GUARD 264 if (kmem_guard_enabled) { 265 return kmem_guard_alloc(&kmem_guard, requested_size, 266 (kmflags & KM_SLEEP) != 0); 267 } 268 #endif 269 size = kmem_roundup_size(requested_size); 270 allocsz = size + SIZE_SIZE; 271 272 #ifdef KMEM_REDZONE 273 if (size - requested_size < REDZONE_SIZE) { 274 /* If there isn't enough space in the padding, allocate 275 * one more memory chunk for the red zone. */ 276 allocsz += kmem_roundup_size(REDZONE_SIZE); 277 } 278 #endif 279 280 if ((index = ((allocsz -1) >> KMEM_SHIFT)) 281 < kmem_cache_maxidx) { 282 pc = kmem_cache[index]; 283 } else if ((index = ((allocsz - 1) >> KMEM_BIG_SHIFT)) 284 < kmem_cache_big_maxidx) { 285 pc = kmem_cache_big[index]; 286 } else { 287 int ret = uvm_km_kmem_alloc(kmem_va_arena, 288 (vsize_t)round_page(size), 289 ((kmflags & KM_SLEEP) ? VM_SLEEP : VM_NOSLEEP) 290 | VM_INSTANTFIT, (vmem_addr_t *)&p); 291 if (ret) { 292 return NULL; 293 } 294 FREECHECK_OUT(&kmem_freecheck, p); 295 return p; 296 } 297 298 p = pool_cache_get(pc, kmflags); 299 300 if (__predict_true(p != NULL)) { 301 kmem_poison_check(p, allocsz); 302 FREECHECK_OUT(&kmem_freecheck, p); 303 kmem_size_set(p, requested_size); 304 kmem_redzone_fill(p, requested_size + SIZE_SIZE); 305 306 return p + SIZE_SIZE; 307 } 308 return p; 309 } 310 311 /* 312 * kmem_intr_zalloc: allocate zeroed wired memory. 313 */ 314 315 void * 316 kmem_intr_zalloc(size_t size, km_flag_t kmflags) 317 { 318 void *p; 319 320 p = kmem_intr_alloc(size, kmflags); 321 if (p != NULL) { 322 memset(p, 0, size); 323 } 324 return p; 325 } 326 327 /* 328 * kmem_intr_free: free wired memory allocated by kmem_alloc. 329 */ 330 331 void 332 kmem_intr_free(void *p, size_t requested_size) 333 { 334 size_t allocsz, index; 335 size_t size; 336 pool_cache_t pc; 337 338 KASSERT(p != NULL); 339 KASSERT(requested_size > 0); 340 341 #ifdef KMEM_GUARD 342 if (kmem_guard_enabled) { 343 kmem_guard_free(&kmem_guard, requested_size, p); 344 return; 345 } 346 #endif 347 348 size = kmem_roundup_size(requested_size); 349 allocsz = size + SIZE_SIZE; 350 351 #ifdef KMEM_REDZONE 352 if (size - requested_size < REDZONE_SIZE) { 353 allocsz += kmem_roundup_size(REDZONE_SIZE); 354 } 355 #endif 356 357 if ((index = ((allocsz -1) >> KMEM_SHIFT)) 358 < kmem_cache_maxidx) { 359 pc = kmem_cache[index]; 360 } else if ((index = ((allocsz - 1) >> KMEM_BIG_SHIFT)) 361 < kmem_cache_big_maxidx) { 362 pc = kmem_cache_big[index]; 363 } else { 364 FREECHECK_IN(&kmem_freecheck, p); 365 uvm_km_kmem_free(kmem_va_arena, (vaddr_t)p, 366 round_page(size)); 367 return; 368 } 369 370 p = (uint8_t *)p - SIZE_SIZE; 371 kmem_size_check(p, requested_size); 372 kmem_redzone_check(p, requested_size + SIZE_SIZE); 373 FREECHECK_IN(&kmem_freecheck, p); 374 LOCKDEBUG_MEM_CHECK(p, size); 375 kmem_poison_fill(p, allocsz); 376 377 pool_cache_put(pc, p); 378 } 379 380 /* ---- kmem API */ 381 382 /* 383 * kmem_alloc: allocate wired memory. 384 * => must not be called from interrupt context. 385 */ 386 387 void * 388 kmem_alloc(size_t size, km_flag_t kmflags) 389 { 390 void *v; 391 392 KASSERTMSG((!cpu_intr_p() && !cpu_softintr_p()), 393 "kmem(9) should not be used from the interrupt context"); 394 v = kmem_intr_alloc(size, kmflags); 395 KASSERT(v || (kmflags & KM_NOSLEEP) != 0); 396 return v; 397 } 398 399 /* 400 * kmem_zalloc: allocate zeroed wired memory. 401 * => must not be called from interrupt context. 402 */ 403 404 void * 405 kmem_zalloc(size_t size, km_flag_t kmflags) 406 { 407 void *v; 408 409 KASSERTMSG((!cpu_intr_p() && !cpu_softintr_p()), 410 "kmem(9) should not be used from the interrupt context"); 411 v = kmem_intr_zalloc(size, kmflags); 412 KASSERT(v || (kmflags & KM_NOSLEEP) != 0); 413 return v; 414 } 415 416 /* 417 * kmem_free: free wired memory allocated by kmem_alloc. 418 * => must not be called from interrupt context. 419 */ 420 421 void 422 kmem_free(void *p, size_t size) 423 { 424 KASSERT(!cpu_intr_p()); 425 KASSERT(!cpu_softintr_p()); 426 kmem_intr_free(p, size); 427 } 428 429 static size_t 430 kmem_create_caches(const struct kmem_cache_info *array, 431 pool_cache_t alloc_table[], size_t maxsize, int shift, int ipl) 432 { 433 size_t maxidx = 0; 434 size_t table_unit = (1 << shift); 435 size_t size = table_unit; 436 int i; 437 438 for (i = 0; array[i].kc_size != 0 ; i++) { 439 const char *name = array[i].kc_name; 440 size_t cache_size = array[i].kc_size; 441 struct pool_allocator *pa; 442 int flags = PR_NOALIGN; 443 pool_cache_t pc; 444 size_t align; 445 446 if ((cache_size & (CACHE_LINE_SIZE - 1)) == 0) 447 align = CACHE_LINE_SIZE; 448 else if ((cache_size & (PAGE_SIZE - 1)) == 0) 449 align = PAGE_SIZE; 450 else 451 align = KMEM_ALIGN; 452 453 if (cache_size < CACHE_LINE_SIZE) 454 flags |= PR_NOTOUCH; 455 456 /* check if we reached the requested size */ 457 if (cache_size > maxsize || cache_size > PAGE_SIZE) { 458 break; 459 } 460 if ((cache_size >> shift) > maxidx) { 461 maxidx = cache_size >> shift; 462 } 463 464 if ((cache_size >> shift) > maxidx) { 465 maxidx = cache_size >> shift; 466 } 467 468 pa = &pool_allocator_kmem; 469 #if defined(KMEM_POISON) 470 pc = pool_cache_init(cache_size, align, 0, flags, 471 name, pa, ipl, kmem_poison_ctor, 472 NULL, (void *)cache_size); 473 #else /* defined(KMEM_POISON) */ 474 pc = pool_cache_init(cache_size, align, 0, flags, 475 name, pa, ipl, NULL, NULL, NULL); 476 #endif /* defined(KMEM_POISON) */ 477 478 while (size <= cache_size) { 479 alloc_table[(size - 1) >> shift] = pc; 480 size += table_unit; 481 } 482 } 483 return maxidx; 484 } 485 486 void 487 kmem_init(void) 488 { 489 #ifdef KMEM_GUARD 490 kmem_guard_enabled = kmem_guard_init(&kmem_guard, kmem_guard_depth, 491 kmem_va_arena); 492 #endif 493 kmem_cache_maxidx = kmem_create_caches(kmem_cache_sizes, 494 kmem_cache, KMEM_MAXSIZE, KMEM_SHIFT, IPL_VM); 495 kmem_cache_big_maxidx = kmem_create_caches(kmem_cache_big_sizes, 496 kmem_cache_big, PAGE_SIZE, KMEM_BIG_SHIFT, IPL_VM); 497 } 498 499 size_t 500 kmem_roundup_size(size_t size) 501 { 502 return (size + (KMEM_ALIGN - 1)) & ~(KMEM_ALIGN - 1); 503 } 504 505 /* 506 * Used to dynamically allocate string with kmem accordingly to format. 507 */ 508 char * 509 kmem_asprintf(const char *fmt, ...) 510 { 511 int size __diagused, len; 512 va_list va; 513 char *str; 514 515 va_start(va, fmt); 516 len = vsnprintf(NULL, 0, fmt, va); 517 va_end(va); 518 519 str = kmem_alloc(len + 1, KM_SLEEP); 520 521 va_start(va, fmt); 522 size = vsnprintf(str, len + 1, fmt, va); 523 va_end(va); 524 525 KASSERT(size == len); 526 527 return str; 528 } 529 530 /* ------------------ DEBUG / DIAGNOSTIC ------------------ */ 531 532 #if defined(KMEM_POISON) || defined(KMEM_REDZONE) 533 #if defined(_LP64) 534 #define PRIME 0x9e37fffffffc0000UL 535 #else /* defined(_LP64) */ 536 #define PRIME 0x9e3779b1 537 #endif /* defined(_LP64) */ 538 539 static inline uint8_t 540 kmem_pattern_generate(const void *p) 541 { 542 return (uint8_t)(((uintptr_t)p) * PRIME 543 >> ((sizeof(uintptr_t) - sizeof(uint8_t))) * CHAR_BIT); 544 } 545 #endif /* defined(KMEM_POISON) || defined(KMEM_REDZONE) */ 546 547 #if defined(KMEM_POISON) 548 static int 549 kmem_poison_ctor(void *arg, void *obj, int flag) 550 { 551 size_t sz = (size_t)arg; 552 553 kmem_poison_fill(obj, sz); 554 555 return 0; 556 } 557 558 static void 559 kmem_poison_fill(void *p, size_t sz) 560 { 561 uint8_t *cp; 562 const uint8_t *ep; 563 564 cp = p; 565 ep = cp + sz; 566 while (cp < ep) { 567 *cp = kmem_pattern_generate(cp); 568 cp++; 569 } 570 } 571 572 static void 573 kmem_poison_check(void *p, size_t sz) 574 { 575 uint8_t *cp; 576 const uint8_t *ep; 577 578 cp = p; 579 ep = cp + sz; 580 while (cp < ep) { 581 const uint8_t expected = kmem_pattern_generate(cp); 582 583 if (*cp != expected) { 584 panic("%s: %p: 0x%02x != 0x%02x\n", 585 __func__, cp, *cp, expected); 586 } 587 cp++; 588 } 589 } 590 #endif /* defined(KMEM_POISON) */ 591 592 #if defined(KMEM_SIZE) 593 static void 594 kmem_size_set(void *p, size_t sz) 595 { 596 struct kmem_header *hd; 597 hd = (struct kmem_header *)p; 598 hd->size = sz; 599 } 600 601 static void 602 kmem_size_check(void *p, size_t sz) 603 { 604 struct kmem_header *hd; 605 size_t hsz; 606 607 hd = (struct kmem_header *)p; 608 hsz = hd->size; 609 610 if (hsz != sz) { 611 panic("kmem_free(%p, %zu) != allocated size %zu", 612 (const uint8_t *)p + SIZE_SIZE, sz, hsz); 613 } 614 } 615 #endif /* defined(KMEM_SIZE) */ 616 617 #if defined(KMEM_REDZONE) 618 #define STATIC_BYTE 0xFE 619 CTASSERT(REDZONE_SIZE > 1); 620 static void 621 kmem_redzone_fill(void *p, size_t sz) 622 { 623 uint8_t *cp, pat; 624 const uint8_t *ep; 625 626 cp = (uint8_t *)p + sz; 627 ep = cp + REDZONE_SIZE; 628 629 /* 630 * We really don't want the first byte of the red zone to be '\0'; 631 * an off-by-one in a string may not be properly detected. 632 */ 633 pat = kmem_pattern_generate(cp); 634 *cp = (pat == '\0') ? STATIC_BYTE: pat; 635 cp++; 636 637 while (cp < ep) { 638 *cp = kmem_pattern_generate(cp); 639 cp++; 640 } 641 } 642 643 static void 644 kmem_redzone_check(void *p, size_t sz) 645 { 646 uint8_t *cp, pat, expected; 647 const uint8_t *ep; 648 649 cp = (uint8_t *)p + sz; 650 ep = cp + REDZONE_SIZE; 651 652 pat = kmem_pattern_generate(cp); 653 expected = (pat == '\0') ? STATIC_BYTE: pat; 654 if (expected != *cp) { 655 panic("%s: %p: 0x%02x != 0x%02x\n", 656 __func__, cp, *cp, expected); 657 } 658 cp++; 659 660 while (cp < ep) { 661 expected = kmem_pattern_generate(cp); 662 if (*cp != expected) { 663 panic("%s: %p: 0x%02x != 0x%02x\n", 664 __func__, cp, *cp, expected); 665 } 666 cp++; 667 } 668 } 669 #endif /* defined(KMEM_REDZONE) */ 670 671 672 #if defined(KMEM_GUARD) 673 /* 674 * The ultimate memory allocator for debugging, baby. It tries to catch: 675 * 676 * 1. Overflow, in realtime. A guard page sits immediately after the 677 * requested area; a read/write overflow therefore triggers a page 678 * fault. 679 * 2. Invalid pointer/size passed, at free. A kmem_header structure sits 680 * just before the requested area, and holds the allocated size. Any 681 * difference with what is given at free triggers a panic. 682 * 3. Underflow, at free. If an underflow occurs, the kmem header will be 683 * modified, and 2. will trigger a panic. 684 * 4. Use-after-free. When freeing, the memory is unmapped, and depending 685 * on the value of kmem_guard_depth, the kernel will more or less delay 686 * the recycling of that memory. Which means that any ulterior read/write 687 * access to the memory will trigger a page fault, given it hasn't been 688 * recycled yet. 689 */ 690 691 #include <sys/atomic.h> 692 #include <uvm/uvm.h> 693 694 static bool 695 kmem_guard_init(struct kmem_guard *kg, u_int depth, vmem_t *vm) 696 { 697 vaddr_t va; 698 699 /* If not enabled, we have nothing to do. */ 700 if (depth == 0) { 701 return false; 702 } 703 depth = roundup(depth, PAGE_SIZE / sizeof(void *)); 704 KASSERT(depth != 0); 705 706 /* 707 * Allocate fifo. 708 */ 709 va = uvm_km_alloc(kernel_map, depth * sizeof(void *), PAGE_SIZE, 710 UVM_KMF_WIRED | UVM_KMF_ZERO); 711 if (va == 0) { 712 return false; 713 } 714 715 /* 716 * Init object. 717 */ 718 kg->kg_vmem = vm; 719 kg->kg_fifo = (void *)va; 720 kg->kg_depth = depth; 721 kg->kg_rotor = 0; 722 723 printf("kmem_guard(%p): depth %d\n", kg, depth); 724 return true; 725 } 726 727 static void * 728 kmem_guard_alloc(struct kmem_guard *kg, size_t requested_size, bool waitok) 729 { 730 struct vm_page *pg; 731 vm_flag_t flags; 732 vmem_addr_t va; 733 vaddr_t loopva; 734 vsize_t loopsize; 735 size_t size; 736 void **p; 737 738 /* 739 * Compute the size: take the kmem header into account, and add a guard 740 * page at the end. 741 */ 742 size = round_page(requested_size + SIZE_SIZE) + PAGE_SIZE; 743 744 /* Allocate pages of kernel VA, but do not map anything in yet. */ 745 flags = VM_BESTFIT | (waitok ? VM_SLEEP : VM_NOSLEEP); 746 if (vmem_alloc(kg->kg_vmem, size, flags, &va) != 0) { 747 return NULL; 748 } 749 750 loopva = va; 751 loopsize = size - PAGE_SIZE; 752 753 while (loopsize) { 754 pg = uvm_pagealloc(NULL, loopva, NULL, 0); 755 if (__predict_false(pg == NULL)) { 756 if (waitok) { 757 uvm_wait("kmem_guard"); 758 continue; 759 } else { 760 uvm_km_pgremove_intrsafe(kernel_map, va, 761 va + size); 762 vmem_free(kg->kg_vmem, va, size); 763 return NULL; 764 } 765 } 766 767 pg->flags &= ~PG_BUSY; /* new page */ 768 UVM_PAGE_OWN(pg, NULL); 769 pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg), 770 VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE); 771 772 loopva += PAGE_SIZE; 773 loopsize -= PAGE_SIZE; 774 } 775 776 pmap_update(pmap_kernel()); 777 778 /* 779 * Offset the returned pointer so that the unmapped guard page sits 780 * immediately after the returned object. 781 */ 782 p = (void **)((va + (size - PAGE_SIZE) - requested_size) & ~(uintptr_t)ALIGNBYTES); 783 kmem_size_set((uint8_t *)p - SIZE_SIZE, requested_size); 784 return (void *)p; 785 } 786 787 static void 788 kmem_guard_free(struct kmem_guard *kg, size_t requested_size, void *p) 789 { 790 vaddr_t va; 791 u_int rotor; 792 size_t size; 793 uint8_t *ptr; 794 795 ptr = (uint8_t *)p - SIZE_SIZE; 796 kmem_size_check(ptr, requested_size); 797 va = trunc_page((vaddr_t)ptr); 798 size = round_page(requested_size + SIZE_SIZE) + PAGE_SIZE; 799 800 KASSERT(pmap_extract(pmap_kernel(), va, NULL)); 801 KASSERT(!pmap_extract(pmap_kernel(), va + (size - PAGE_SIZE), NULL)); 802 803 /* 804 * Unmap and free the pages. The last one is never allocated. 805 */ 806 uvm_km_pgremove_intrsafe(kernel_map, va, va + size); 807 pmap_update(pmap_kernel()); 808 809 #if 0 810 /* 811 * XXX: Here, we need to atomically register the va and its size in the 812 * fifo. 813 */ 814 815 /* 816 * Put the VA allocation into the list and swap an old one out to free. 817 * This behaves mostly like a fifo. 818 */ 819 rotor = atomic_inc_uint_nv(&kg->kg_rotor) % kg->kg_depth; 820 va = (vaddr_t)atomic_swap_ptr(&kg->kg_fifo[rotor], (void *)va); 821 if (va != 0) { 822 vmem_free(kg->kg_vmem, va, size); 823 } 824 #else 825 (void)rotor; 826 vmem_free(kg->kg_vmem, va, size); 827 #endif 828 } 829 830 #endif /* defined(KMEM_GUARD) */ 831
Indexes created Sun Sep 21 20:09:37 GMT 2025