1 /* 2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator 3 * 4 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to The DragonFly Project 7 * by Matthew Dillon <dillon@backplane.com> 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 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 17 * the documentation and/or other materials provided with the 18 * distribution. 19 * 3. Neither the name of The DragonFly Project nor the names of its 20 * contributors may be used to endorse or promote products derived 21 * from this software without specific, prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.38 2006/05/24 03:23:31 dillon Exp $ 37 * 38 * This module implements a slab allocator drop-in replacement for the 39 * kernel malloc(). 40 * 41 * A slab allocator reserves a ZONE for each chunk size, then lays the 42 * chunks out in an array within the zone. Allocation and deallocation 43 * is nearly instantanious, and fragmentation/overhead losses are limited 44 * to a fixed worst-case amount. 45 * 46 * The downside of this slab implementation is in the chunk size 47 * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu. 48 * In a kernel implementation all this memory will be physical so 49 * the zone size is adjusted downward on machines with less physical 50 * memory. The upside is that overhead is bounded... this is the *worst* 51 * case overhead. 52 * 53 * Slab management is done on a per-cpu basis and no locking or mutexes 54 * are required, only a critical section. When one cpu frees memory 55 * belonging to another cpu's slab manager an asynchronous IPI message 56 * will be queued to execute the operation. In addition, both the 57 * high level slab allocator and the low level zone allocator optimize 58 * M_ZERO requests, and the slab allocator does not have to pre initialize 59 * the linked list of chunks. 60 * 61 * XXX Balancing is needed between cpus. Balance will be handled through 62 * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks. 63 * 64 * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of 65 * the new zone should be restricted to M_USE_RESERVE requests only. 66 * 67 * Alloc Size Chunking Number of zones 68 * 0-127 8 16 69 * 128-255 16 8 70 * 256-511 32 8 71 * 512-1023 64 8 72 * 1024-2047 128 8 73 * 2048-4095 256 8 74 * 4096-8191 512 8 75 * 8192-16383 1024 8 76 * 16384-32767 2048 8 77 * (if PAGE_SIZE is 4K the maximum zone allocation is 16383) 78 * 79 * Allocations >= ZoneLimit go directly to kmem. 80 * 81 * API REQUIREMENTS AND SIDE EFFECTS 82 * 83 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we 84 * have remained compatible with the following API requirements: 85 * 86 * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty) 87 * + all power-of-2 sized allocations are power-of-2 aligned (twe) 88 * + malloc(0) is allowed and returns non-NULL (ahc driver) 89 * + ability to allocate arbitrarily large chunks of memory 90 */ 91 92 #include "opt_vm.h" 93 94 #include <sys/param.h> 95 #include <sys/systm.h> 96 #include <sys/kernel.h> 97 #include <sys/slaballoc.h> 98 #include <sys/mbuf.h> 99 #include <sys/vmmeter.h> 100 #include <sys/lock.h> 101 #include <sys/thread.h> 102 #include <sys/globaldata.h> 103 #include <sys/sysctl.h> 104 #include <sys/ktr.h> 105 106 #include <vm/vm.h> 107 #include <vm/vm_param.h> 108 #include <vm/vm_kern.h> 109 #include <vm/vm_extern.h> 110 #include <vm/vm_object.h> 111 #include <vm/pmap.h> 112 #include <vm/vm_map.h> 113 #include <vm/vm_page.h> 114 #include <vm/vm_pageout.h> 115 116 #include <machine/cpu.h> 117 118 #include <sys/thread2.h> 119 120 #define arysize(ary) (sizeof(ary)/sizeof((ary)[0])) 121 122 #define MEMORY_STRING "ptr=%p type=%p size=%d flags=%04x" 123 #define MEMORY_ARG_SIZE (sizeof(void *) * 2 + sizeof(unsigned long) + \ 124 sizeof(int)) 125 126 #if !defined(KTR_MEMORY) 127 #define KTR_MEMORY KTR_ALL 128 #endif 129 KTR_INFO_MASTER(memory); 130 KTR_INFO(KTR_MEMORY, memory, malloc, 0, MEMORY_STRING, MEMORY_ARG_SIZE); 131 KTR_INFO(KTR_MEMORY, memory, free_zero, 1, MEMORY_STRING, MEMORY_ARG_SIZE); 132 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 2, MEMORY_STRING, MEMORY_ARG_SIZE); 133 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 3, MEMORY_STRING, MEMORY_ARG_SIZE); 134 KTR_INFO(KTR_MEMORY, memory, free_chunk, 4, MEMORY_STRING, MEMORY_ARG_SIZE); 135 #ifdef SMP 136 KTR_INFO(KTR_MEMORY, memory, free_request, 5, MEMORY_STRING, MEMORY_ARG_SIZE); 137 KTR_INFO(KTR_MEMORY, memory, free_remote, 6, MEMORY_STRING, MEMORY_ARG_SIZE); 138 #endif 139 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin", 0); 140 KTR_INFO(KTR_MEMORY, memory, free_beg, 0, "free begin", 0); 141 KTR_INFO(KTR_MEMORY, memory, free_end, 0, "free end", 0); 142 143 #define logmemory(name, ptr, type, size, flags) \ 144 KTR_LOG(memory_ ## name, ptr, type, size, flags) 145 #define logmemory_quick(name) \ 146 KTR_LOG(memory_ ## name) 147 148 /* 149 * Fixed globals (not per-cpu) 150 */ 151 static int ZoneSize; 152 static int ZoneLimit; 153 static int ZonePageCount; 154 static int ZoneMask; 155 static struct malloc_type *kmemstatistics; 156 static struct kmemusage *kmemusage; 157 static int32_t weirdary[16]; 158 159 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags); 160 static void kmem_slab_free(void *ptr, vm_size_t bytes); 161 #if defined(INVARIANTS) 162 static void chunk_mark_allocated(SLZone *z, void *chunk); 163 static void chunk_mark_free(SLZone *z, void *chunk); 164 #endif 165 166 /* 167 * Misc constants. Note that allocations that are exact multiples of 168 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module. 169 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists. 170 */ 171 #define MIN_CHUNK_SIZE 8 /* in bytes */ 172 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1) 173 #define ZONE_RELS_THRESH 2 /* threshold number of zones */ 174 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK) 175 176 /* 177 * The WEIRD_ADDR is used as known text to copy into free objects to 178 * try to create deterministic failure cases if the data is accessed after 179 * free. 180 */ 181 #define WEIRD_ADDR 0xdeadc0de 182 #define MAX_COPY sizeof(weirdary) 183 #define ZERO_LENGTH_PTR ((void *)-8) 184 185 /* 186 * Misc global malloc buckets 187 */ 188 189 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches"); 190 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory"); 191 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers"); 192 193 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options"); 194 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery"); 195 196 /* 197 * Initialize the slab memory allocator. We have to choose a zone size based 198 * on available physical memory. We choose a zone side which is approximately 199 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of 200 * 128K. The zone size is limited to the bounds set in slaballoc.h 201 * (typically 32K min, 128K max). 202 */ 203 static void kmeminit(void *dummy); 204 205 SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL) 206 207 #ifdef INVARIANTS 208 /* 209 * If enabled any memory allocated without M_ZERO is initialized to -1. 210 */ 211 static int use_malloc_pattern; 212 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW, 213 &use_malloc_pattern, 0, ""); 214 #endif 215 216 static void 217 kmeminit(void *dummy) 218 { 219 vm_poff_t limsize; 220 int usesize; 221 int i; 222 vm_pindex_t npg; 223 224 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; 225 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) 226 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; 227 228 usesize = (int)(limsize / 1024); /* convert to KB */ 229 230 ZoneSize = ZALLOC_MIN_ZONE_SIZE; 231 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize) 232 ZoneSize <<= 1; 233 ZoneLimit = ZoneSize / 4; 234 if (ZoneLimit > ZALLOC_ZONE_LIMIT) 235 ZoneLimit = ZALLOC_ZONE_LIMIT; 236 ZoneMask = ZoneSize - 1; 237 ZonePageCount = ZoneSize / PAGE_SIZE; 238 239 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE; 240 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_WAITOK|M_ZERO); 241 242 for (i = 0; i < arysize(weirdary); ++i) 243 weirdary[i] = WEIRD_ADDR; 244 245 if (bootverbose) 246 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024); 247 } 248 249 /* 250 * Initialize a malloc type tracking structure. 251 */ 252 void 253 malloc_init(void *data) 254 { 255 struct malloc_type *type = data; 256 vm_poff_t limsize; 257 258 if (type->ks_magic != M_MAGIC) 259 panic("malloc type lacks magic"); 260 261 if (type->ks_limit != 0) 262 return; 263 264 if (vmstats.v_page_count == 0) 265 panic("malloc_init not allowed before vm init"); 266 267 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; 268 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) 269 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; 270 type->ks_limit = limsize / 10; 271 272 type->ks_next = kmemstatistics; 273 kmemstatistics = type; 274 } 275 276 void 277 malloc_uninit(void *data) 278 { 279 struct malloc_type *type = data; 280 struct malloc_type *t; 281 #ifdef INVARIANTS 282 int i; 283 long ttl; 284 #endif 285 286 if (type->ks_magic != M_MAGIC) 287 panic("malloc type lacks magic"); 288 289 if (vmstats.v_page_count == 0) 290 panic("malloc_uninit not allowed before vm init"); 291 292 if (type->ks_limit == 0) 293 panic("malloc_uninit on uninitialized type"); 294 295 #ifdef INVARIANTS 296 /* 297 * memuse is only correct in aggregation. Due to memory being allocated 298 * on one cpu and freed on another individual array entries may be 299 * negative or positive (canceling each other out). 300 */ 301 for (i = ttl = 0; i < ncpus; ++i) 302 ttl += type->ks_memuse[i]; 303 if (ttl) { 304 printf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n", 305 ttl, type->ks_shortdesc, i); 306 } 307 #endif 308 if (type == kmemstatistics) { 309 kmemstatistics = type->ks_next; 310 } else { 311 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) { 312 if (t->ks_next == type) { 313 t->ks_next = type->ks_next; 314 break; 315 } 316 } 317 } 318 type->ks_next = NULL; 319 type->ks_limit = 0; 320 } 321 322 /* 323 * Calculate the zone index for the allocation request size and set the 324 * allocation request size to that particular zone's chunk size. 325 */ 326 static __inline int 327 zoneindex(unsigned long *bytes) 328 { 329 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */ 330 if (n < 128) { 331 *bytes = n = (n + 7) & ~7; 332 return(n / 8 - 1); /* 8 byte chunks, 16 zones */ 333 } 334 if (n < 256) { 335 *bytes = n = (n + 15) & ~15; 336 return(n / 16 + 7); 337 } 338 if (n < 8192) { 339 if (n < 512) { 340 *bytes = n = (n + 31) & ~31; 341 return(n / 32 + 15); 342 } 343 if (n < 1024) { 344 *bytes = n = (n + 63) & ~63; 345 return(n / 64 + 23); 346 } 347 if (n < 2048) { 348 *bytes = n = (n + 127) & ~127; 349 return(n / 128 + 31); 350 } 351 if (n < 4096) { 352 *bytes = n = (n + 255) & ~255; 353 return(n / 256 + 39); 354 } 355 *bytes = n = (n + 511) & ~511; 356 return(n / 512 + 47); 357 } 358 #if ZALLOC_ZONE_LIMIT > 8192 359 if (n < 16384) { 360 *bytes = n = (n + 1023) & ~1023; 361 return(n / 1024 + 55); 362 } 363 #endif 364 #if ZALLOC_ZONE_LIMIT > 16384 365 if (n < 32768) { 366 *bytes = n = (n + 2047) & ~2047; 367 return(n / 2048 + 63); 368 } 369 #endif 370 panic("Unexpected byte count %d", n); 371 return(0); 372 } 373 374 /* 375 * malloc() (SLAB ALLOCATOR) 376 * 377 * Allocate memory via the slab allocator. If the request is too large, 378 * or if it page-aligned beyond a certain size, we fall back to the 379 * KMEM subsystem. A SLAB tracking descriptor must be specified, use 380 * &SlabMisc if you don't care. 381 * 382 * M_RNOWAIT - don't block. 383 * M_NULLOK - return NULL instead of blocking. 384 * M_ZERO - zero the returned memory. 385 * M_USE_RESERVE - allow greater drawdown of the free list 386 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted 387 * 388 * MPSAFE 389 */ 390 void * 391 malloc(unsigned long size, struct malloc_type *type, int flags) 392 { 393 SLZone *z; 394 SLChunk *chunk; 395 SLGlobalData *slgd; 396 struct globaldata *gd; 397 int zi; 398 #ifdef INVARIANTS 399 int i; 400 #endif 401 402 logmemory_quick(malloc_beg); 403 gd = mycpu; 404 slgd = &gd->gd_slab; 405 406 /* 407 * XXX silly to have this in the critical path. 408 */ 409 if (type->ks_limit == 0) { 410 crit_enter(); 411 if (type->ks_limit == 0) 412 malloc_init(type); 413 crit_exit(); 414 } 415 ++type->ks_calls; 416 417 /* 418 * Handle the case where the limit is reached. Panic if we can't return 419 * NULL. The original malloc code looped, but this tended to 420 * simply deadlock the computer. 421 * 422 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used 423 * to determine if a more complete limit check should be done. The 424 * actual memory use is tracked via ks_memuse[cpu]. 425 */ 426 while (type->ks_loosememuse >= type->ks_limit) { 427 int i; 428 long ttl; 429 430 for (i = ttl = 0; i < ncpus; ++i) 431 ttl += type->ks_memuse[i]; 432 type->ks_loosememuse = ttl; /* not MP synchronized */ 433 if (ttl >= type->ks_limit) { 434 if (flags & M_NULLOK) { 435 logmemory(malloc, NULL, type, size, flags); 436 return(NULL); 437 } 438 panic("%s: malloc limit exceeded", type->ks_shortdesc); 439 } 440 } 441 442 /* 443 * Handle the degenerate size == 0 case. Yes, this does happen. 444 * Return a special pointer. This is to maintain compatibility with 445 * the original malloc implementation. Certain devices, such as the 446 * adaptec driver, not only allocate 0 bytes, they check for NULL and 447 * also realloc() later on. Joy. 448 */ 449 if (size == 0) { 450 logmemory(malloc, ZERO_LENGTH_PTR, type, size, flags); 451 return(ZERO_LENGTH_PTR); 452 } 453 454 /* 455 * Handle hysteresis from prior frees here in malloc(). We cannot 456 * safely manipulate the kernel_map in free() due to free() possibly 457 * being called via an IPI message or from sensitive interrupt code. 458 */ 459 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) { 460 crit_enter(); 461 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */ 462 z = slgd->FreeZones; 463 slgd->FreeZones = z->z_Next; 464 --slgd->NFreeZones; 465 kmem_slab_free(z, ZoneSize); /* may block */ 466 } 467 crit_exit(); 468 } 469 /* 470 * XXX handle oversized frees that were queued from free(). 471 */ 472 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) { 473 crit_enter(); 474 if ((z = slgd->FreeOvZones) != NULL) { 475 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC); 476 slgd->FreeOvZones = z->z_Next; 477 kmem_slab_free(z, z->z_ChunkSize); /* may block */ 478 } 479 crit_exit(); 480 } 481 482 /* 483 * Handle large allocations directly. There should not be very many of 484 * these so performance is not a big issue. 485 * 486 * Guarentee page alignment for allocations in multiples of PAGE_SIZE 487 */ 488 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) { 489 struct kmemusage *kup; 490 491 size = round_page(size); 492 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags); 493 if (chunk == NULL) { 494 logmemory(malloc, NULL, type, size, flags); 495 return(NULL); 496 } 497 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */ 498 flags |= M_PASSIVE_ZERO; 499 kup = btokup(chunk); 500 kup->ku_pagecnt = size / PAGE_SIZE; 501 kup->ku_cpu = gd->gd_cpuid; 502 crit_enter(); 503 goto done; 504 } 505 506 /* 507 * Attempt to allocate out of an existing zone. First try the free list, 508 * then allocate out of unallocated space. If we find a good zone move 509 * it to the head of the list so later allocations find it quickly 510 * (we might have thousands of zones in the list). 511 * 512 * Note: zoneindex() will panic of size is too large. 513 */ 514 zi = zoneindex(&size); 515 KKASSERT(zi < NZONES); 516 crit_enter(); 517 if ((z = slgd->ZoneAry[zi]) != NULL) { 518 KKASSERT(z->z_NFree > 0); 519 520 /* 521 * Remove us from the ZoneAry[] when we become empty 522 */ 523 if (--z->z_NFree == 0) { 524 slgd->ZoneAry[zi] = z->z_Next; 525 z->z_Next = NULL; 526 } 527 528 /* 529 * Locate a chunk in a free page. This attempts to localize 530 * reallocations into earlier pages without us having to sort 531 * the chunk list. A chunk may still overlap a page boundary. 532 */ 533 while (z->z_FirstFreePg < ZonePageCount) { 534 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) { 535 #ifdef DIAGNOSTIC 536 /* 537 * Diagnostic: c_Next is not total garbage. 538 */ 539 KKASSERT(chunk->c_Next == NULL || 540 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) == 541 ((intptr_t)chunk & IN_SAME_PAGE_MASK)); 542 #endif 543 #ifdef INVARIANTS 544 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) 545 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount); 546 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) 547 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount); 548 chunk_mark_allocated(z, chunk); 549 #endif 550 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next; 551 goto done; 552 } 553 ++z->z_FirstFreePg; 554 } 555 556 /* 557 * No chunks are available but NFree said we had some memory, so 558 * it must be available in the never-before-used-memory area 559 * governed by UIndex. The consequences are very serious if our zone 560 * got corrupted so we use an explicit panic rather then a KASSERT. 561 */ 562 if (z->z_UIndex + 1 != z->z_NMax) 563 z->z_UIndex = z->z_UIndex + 1; 564 else 565 z->z_UIndex = 0; 566 if (z->z_UIndex == z->z_UEndIndex) 567 panic("slaballoc: corrupted zone"); 568 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 569 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 570 flags &= ~M_ZERO; 571 flags |= M_PASSIVE_ZERO; 572 } 573 #if defined(INVARIANTS) 574 chunk_mark_allocated(z, chunk); 575 #endif 576 goto done; 577 } 578 579 /* 580 * If all zones are exhausted we need to allocate a new zone for this 581 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see 582 * UAlloc use above in regards to M_ZERO. Note that when we are reusing 583 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and 584 * we do not pre-zero it because we do not want to mess up the L1 cache. 585 * 586 * At least one subsystem, the tty code (see CROUND) expects power-of-2 587 * allocations to be power-of-2 aligned. We maintain compatibility by 588 * adjusting the base offset below. 589 */ 590 { 591 int off; 592 593 if ((z = slgd->FreeZones) != NULL) { 594 slgd->FreeZones = z->z_Next; 595 --slgd->NFreeZones; 596 bzero(z, sizeof(SLZone)); 597 z->z_Flags |= SLZF_UNOTZEROD; 598 } else { 599 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO); 600 if (z == NULL) 601 goto fail; 602 } 603 604 /* 605 * How big is the base structure? 606 */ 607 #if defined(INVARIANTS) 608 /* 609 * Make room for z_Bitmap. An exact calculation is somewhat more 610 * complicated so don't make an exact calculation. 611 */ 612 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]); 613 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8); 614 #else 615 off = sizeof(SLZone); 616 #endif 617 618 /* 619 * Guarentee power-of-2 alignment for power-of-2-sized chunks. 620 * Otherwise just 8-byte align the data. 621 */ 622 if ((size | (size - 1)) + 1 == (size << 1)) 623 off = (off + size - 1) & ~(size - 1); 624 else 625 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK; 626 z->z_Magic = ZALLOC_SLAB_MAGIC; 627 z->z_ZoneIndex = zi; 628 z->z_NMax = (ZoneSize - off) / size; 629 z->z_NFree = z->z_NMax - 1; 630 z->z_BasePtr = (char *)z + off; 631 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax; 632 z->z_ChunkSize = size; 633 z->z_FirstFreePg = ZonePageCount; 634 z->z_CpuGd = gd; 635 z->z_Cpu = gd->gd_cpuid; 636 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 637 z->z_Next = slgd->ZoneAry[zi]; 638 slgd->ZoneAry[zi] = z; 639 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 640 flags &= ~M_ZERO; /* already zero'd */ 641 flags |= M_PASSIVE_ZERO; 642 } 643 #if defined(INVARIANTS) 644 chunk_mark_allocated(z, chunk); 645 #endif 646 647 /* 648 * Slide the base index for initial allocations out of the next 649 * zone we create so we do not over-weight the lower part of the 650 * cpu memory caches. 651 */ 652 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE) 653 & (ZALLOC_MAX_ZONE_SIZE - 1); 654 } 655 done: 656 ++type->ks_inuse[gd->gd_cpuid]; 657 type->ks_memuse[gd->gd_cpuid] += size; 658 type->ks_loosememuse += size; /* not MP synchronized */ 659 crit_exit(); 660 if (flags & M_ZERO) 661 bzero(chunk, size); 662 #ifdef INVARIANTS 663 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) { 664 if (use_malloc_pattern) { 665 for (i = 0; i < size; i += sizeof(int)) { 666 *(int *)((char *)chunk + i) = -1; 667 } 668 } 669 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */ 670 } 671 #endif 672 logmemory(malloc, chunk, type, size, flags); 673 return(chunk); 674 fail: 675 crit_exit(); 676 logmemory(malloc, NULL, type, size, flags); 677 return(NULL); 678 } 679 680 /* 681 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE) 682 * 683 * Generally speaking this routine is not called very often and we do 684 * not attempt to optimize it beyond reusing the same pointer if the 685 * new size fits within the chunking of the old pointer's zone. 686 */ 687 void * 688 realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags) 689 { 690 SLZone *z; 691 void *nptr; 692 unsigned long osize; 693 694 KKASSERT((flags & M_ZERO) == 0); /* not supported */ 695 696 if (ptr == NULL || ptr == ZERO_LENGTH_PTR) 697 return(malloc(size, type, flags)); 698 if (size == 0) { 699 free(ptr, type); 700 return(NULL); 701 } 702 703 /* 704 * Handle oversized allocations. XXX we really should require that a 705 * size be passed to free() instead of this nonsense. 706 */ 707 { 708 struct kmemusage *kup; 709 710 kup = btokup(ptr); 711 if (kup->ku_pagecnt) { 712 osize = kup->ku_pagecnt << PAGE_SHIFT; 713 if (osize == round_page(size)) 714 return(ptr); 715 if ((nptr = malloc(size, type, flags)) == NULL) 716 return(NULL); 717 bcopy(ptr, nptr, min(size, osize)); 718 free(ptr, type); 719 return(nptr); 720 } 721 } 722 723 /* 724 * Get the original allocation's zone. If the new request winds up 725 * using the same chunk size we do not have to do anything. 726 */ 727 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 728 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 729 730 zoneindex(&size); 731 if (z->z_ChunkSize == size) 732 return(ptr); 733 734 /* 735 * Allocate memory for the new request size. Note that zoneindex has 736 * already adjusted the request size to the appropriate chunk size, which 737 * should optimize our bcopy(). Then copy and return the new pointer. 738 */ 739 if ((nptr = malloc(size, type, flags)) == NULL) 740 return(NULL); 741 bcopy(ptr, nptr, min(size, z->z_ChunkSize)); 742 free(ptr, type); 743 return(nptr); 744 } 745 746 /* 747 * Allocate a copy of the specified string. 748 * 749 * (MP SAFE) (MAY BLOCK) 750 */ 751 char * 752 strdup(const char *str, struct malloc_type *type) 753 { 754 int zlen; /* length inclusive of terminating NUL */ 755 char *nstr; 756 757 if (str == NULL) 758 return(NULL); 759 zlen = strlen(str) + 1; 760 nstr = malloc(zlen, type, M_WAITOK); 761 bcopy(str, nstr, zlen); 762 return(nstr); 763 } 764 765 #ifdef SMP 766 /* 767 * free() (SLAB ALLOCATOR) 768 * 769 * Free the specified chunk of memory. 770 */ 771 static 772 void 773 free_remote(void *ptr) 774 { 775 logmemory(free_remote, ptr, *(struct malloc_type **)ptr, -1, 0); 776 free(ptr, *(struct malloc_type **)ptr); 777 } 778 779 #endif 780 781 /* 782 * free (SLAB ALLOCATOR) 783 * 784 * Free a memory block previously allocated by malloc. Note that we do not 785 * attempt to uplodate ks_loosememuse as MP races could prevent us from 786 * checking memory limits in malloc. 787 * 788 * MPSAFE 789 */ 790 void 791 free(void *ptr, struct malloc_type *type) 792 { 793 SLZone *z; 794 SLChunk *chunk; 795 SLGlobalData *slgd; 796 struct globaldata *gd; 797 int pgno; 798 799 logmemory_quick(free_beg); 800 gd = mycpu; 801 slgd = &gd->gd_slab; 802 803 if (ptr == NULL) 804 panic("trying to free NULL pointer"); 805 806 /* 807 * Handle special 0-byte allocations 808 */ 809 if (ptr == ZERO_LENGTH_PTR) { 810 logmemory(free_zero, ptr, type, -1, 0); 811 logmemory_quick(free_end); 812 return; 813 } 814 815 /* 816 * Handle oversized allocations. XXX we really should require that a 817 * size be passed to free() instead of this nonsense. 818 * 819 * This code is never called via an ipi. 820 */ 821 { 822 struct kmemusage *kup; 823 unsigned long size; 824 825 kup = btokup(ptr); 826 if (kup->ku_pagecnt) { 827 size = kup->ku_pagecnt << PAGE_SHIFT; 828 kup->ku_pagecnt = 0; 829 #ifdef INVARIANTS 830 KKASSERT(sizeof(weirdary) <= size); 831 bcopy(weirdary, ptr, sizeof(weirdary)); 832 #endif 833 /* 834 * note: we always adjust our cpu's slot, not the originating 835 * cpu (kup->ku_cpuid). The statistics are in aggregate. 836 * 837 * note: XXX we have still inherited the interrupts-can't-block 838 * assumption. An interrupt thread does not bump 839 * gd_intr_nesting_level so check TDF_INTTHREAD. This is 840 * primarily until we can fix softupdate's assumptions about free(). 841 */ 842 crit_enter(); 843 --type->ks_inuse[gd->gd_cpuid]; 844 type->ks_memuse[gd->gd_cpuid] -= size; 845 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) { 846 logmemory(free_ovsz_delayed, ptr, type, size, 0); 847 z = (SLZone *)ptr; 848 z->z_Magic = ZALLOC_OVSZ_MAGIC; 849 z->z_Next = slgd->FreeOvZones; 850 z->z_ChunkSize = size; 851 slgd->FreeOvZones = z; 852 crit_exit(); 853 } else { 854 crit_exit(); 855 logmemory(free_ovsz, ptr, type, size, 0); 856 kmem_slab_free(ptr, size); /* may block */ 857 } 858 logmemory_quick(free_end); 859 return; 860 } 861 } 862 863 /* 864 * Zone case. Figure out the zone based on the fact that it is 865 * ZoneSize aligned. 866 */ 867 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 868 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 869 870 /* 871 * If we do not own the zone then forward the request to the 872 * cpu that does. Since the timing is non-critical, a passive 873 * message is sent. 874 */ 875 if (z->z_CpuGd != gd) { 876 *(struct malloc_type **)ptr = type; 877 #ifdef SMP 878 logmemory(free_request, ptr, type, z->z_ChunkSize, 0); 879 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr); 880 #else 881 panic("Corrupt SLZone"); 882 #endif 883 logmemory_quick(free_end); 884 return; 885 } 886 887 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0); 888 889 if (type->ks_magic != M_MAGIC) 890 panic("free: malloc type lacks magic"); 891 892 crit_enter(); 893 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT; 894 chunk = ptr; 895 896 #ifdef INVARIANTS 897 /* 898 * Attempt to detect a double-free. To reduce overhead we only check 899 * if there appears to be link pointer at the base of the data. 900 */ 901 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) { 902 SLChunk *scan; 903 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) { 904 if (scan == chunk) 905 panic("Double free at %p", chunk); 906 } 907 } 908 chunk_mark_free(z, chunk); 909 #endif 910 911 /* 912 * Put weird data into the memory to detect modifications after freeing, 913 * illegal pointer use after freeing (we should fault on the odd address), 914 * and so forth. XXX needs more work, see the old malloc code. 915 */ 916 #ifdef INVARIANTS 917 if (z->z_ChunkSize < sizeof(weirdary)) 918 bcopy(weirdary, chunk, z->z_ChunkSize); 919 else 920 bcopy(weirdary, chunk, sizeof(weirdary)); 921 #endif 922 923 /* 924 * Add this free non-zero'd chunk to a linked list for reuse, adjust 925 * z_FirstFreePg. 926 */ 927 #ifdef INVARIANTS 928 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) 929 panic("BADFREE %p", chunk); 930 #endif 931 chunk->c_Next = z->z_PageAry[pgno]; 932 z->z_PageAry[pgno] = chunk; 933 #ifdef INVARIANTS 934 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) 935 panic("BADFREE2"); 936 #endif 937 if (z->z_FirstFreePg > pgno) 938 z->z_FirstFreePg = pgno; 939 940 /* 941 * Bump the number of free chunks. If it becomes non-zero the zone 942 * must be added back onto the appropriate list. 943 */ 944 if (z->z_NFree++ == 0) { 945 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; 946 slgd->ZoneAry[z->z_ZoneIndex] = z; 947 } 948 949 --type->ks_inuse[z->z_Cpu]; 950 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize; 951 952 /* 953 * If the zone becomes totally free, and there are other zones we 954 * can allocate from, move this zone to the FreeZones list. Since 955 * this code can be called from an IPI callback, do *NOT* try to mess 956 * with kernel_map here. Hysteresis will be performed at malloc() time. 957 */ 958 if (z->z_NFree == z->z_NMax && 959 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) 960 ) { 961 SLZone **pz; 962 963 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next) 964 ; 965 *pz = z->z_Next; 966 z->z_Magic = -1; 967 z->z_Next = slgd->FreeZones; 968 slgd->FreeZones = z; 969 ++slgd->NFreeZones; 970 } 971 logmemory_quick(free_end); 972 crit_exit(); 973 } 974 975 #if defined(INVARIANTS) 976 /* 977 * Helper routines for sanity checks 978 */ 979 static 980 void 981 chunk_mark_allocated(SLZone *z, void *chunk) 982 { 983 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 984 __uint32_t *bitptr; 985 986 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal", chunk, bitdex)); 987 bitptr = &z->z_Bitmap[bitdex >> 5]; 988 bitdex &= 31; 989 KASSERT((*bitptr & (1 << bitdex)) == 0, ("memory chunk %p is already allocated!", chunk)); 990 *bitptr |= 1 << bitdex; 991 } 992 993 static 994 void 995 chunk_mark_free(SLZone *z, void *chunk) 996 { 997 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 998 __uint32_t *bitptr; 999 1000 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal!", chunk, bitdex)); 1001 bitptr = &z->z_Bitmap[bitdex >> 5]; 1002 bitdex &= 31; 1003 KASSERT((*bitptr & (1 << bitdex)) != 0, ("memory chunk %p is already free!", chunk)); 1004 *bitptr &= ~(1 << bitdex); 1005 } 1006 1007 #endif 1008 1009 /* 1010 * kmem_slab_alloc() 1011 * 1012 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the 1013 * specified alignment. M_* flags are expected in the flags field. 1014 * 1015 * Alignment must be a multiple of PAGE_SIZE. 1016 * 1017 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(), 1018 * but when we move zalloc() over to use this function as its backend 1019 * we will have to switch to kreserve/krelease and call reserve(0) 1020 * after the new space is made available. 1021 * 1022 * Interrupt code which has preempted other code is not allowed to 1023 * use PQ_CACHE pages. However, if an interrupt thread is run 1024 * non-preemptively or blocks and then runs non-preemptively, then 1025 * it is free to use PQ_CACHE pages. 1026 * 1027 * This routine will currently obtain the BGL. 1028 * 1029 * MPALMOSTSAFE - acquires mplock 1030 */ 1031 static void * 1032 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags) 1033 { 1034 vm_size_t i; 1035 vm_offset_t addr; 1036 vm_offset_t offset; 1037 int count, vmflags, base_vmflags; 1038 thread_t td; 1039 vm_map_t map = kernel_map; 1040 1041 size = round_page(size); 1042 addr = vm_map_min(map); 1043 1044 /* 1045 * Reserve properly aligned space from kernel_map. RNOWAIT allocations 1046 * cannot block. 1047 */ 1048 if (flags & M_RNOWAIT) { 1049 if (try_mplock() == 0) 1050 return(NULL); 1051 } else { 1052 get_mplock(); 1053 } 1054 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1055 crit_enter(); 1056 vm_map_lock(map); 1057 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) { 1058 vm_map_unlock(map); 1059 if ((flags & M_NULLOK) == 0) 1060 panic("kmem_slab_alloc(): kernel_map ran out of space!"); 1061 crit_exit(); 1062 vm_map_entry_release(count); 1063 rel_mplock(); 1064 return(NULL); 1065 } 1066 offset = addr - VM_MIN_KERNEL_ADDRESS; 1067 vm_object_reference(kernel_object); 1068 vm_map_insert(map, &count, 1069 kernel_object, offset, addr, addr + size, 1070 VM_PROT_ALL, VM_PROT_ALL, 0); 1071 1072 td = curthread; 1073 1074 base_vmflags = 0; 1075 if (flags & M_ZERO) 1076 base_vmflags |= VM_ALLOC_ZERO; 1077 if (flags & M_USE_RESERVE) 1078 base_vmflags |= VM_ALLOC_SYSTEM; 1079 if (flags & M_USE_INTERRUPT_RESERVE) 1080 base_vmflags |= VM_ALLOC_INTERRUPT; 1081 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) 1082 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]); 1083 1084 1085 /* 1086 * Allocate the pages. Do not mess with the PG_ZERO flag yet. 1087 */ 1088 for (i = 0; i < size; i += PAGE_SIZE) { 1089 vm_page_t m; 1090 vm_pindex_t idx = OFF_TO_IDX(offset + i); 1091 1092 /* 1093 * VM_ALLOC_NORMAL can only be set if we are not preempting. 1094 * 1095 * VM_ALLOC_SYSTEM is automatically set if we are preempting and 1096 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is 1097 * implied in this case), though I'm sure if we really need to do 1098 * that. 1099 */ 1100 vmflags = base_vmflags; 1101 if (flags & M_WAITOK) { 1102 if (td->td_preempted) 1103 vmflags |= VM_ALLOC_SYSTEM; 1104 else 1105 vmflags |= VM_ALLOC_NORMAL; 1106 } 1107 1108 m = vm_page_alloc(kernel_object, idx, vmflags); 1109 1110 /* 1111 * If the allocation failed we either return NULL or we retry. 1112 * 1113 * If M_WAITOK is specified we wait for more memory and retry. 1114 * If M_WAITOK is specified from a preemption we yield instead of 1115 * wait. Livelock will not occur because the interrupt thread 1116 * will not be preempting anyone the second time around after the 1117 * yield. 1118 */ 1119 if (m == NULL) { 1120 if (flags & M_WAITOK) { 1121 if (td->td_preempted) { 1122 vm_map_unlock(map); 1123 lwkt_yield(); 1124 vm_map_lock(map); 1125 } else { 1126 vm_map_unlock(map); 1127 vm_wait(); 1128 vm_map_lock(map); 1129 } 1130 i -= PAGE_SIZE; /* retry */ 1131 continue; 1132 } 1133 1134 /* 1135 * We were unable to recover, cleanup and return NULL 1136 */ 1137 while (i != 0) { 1138 i -= PAGE_SIZE; 1139 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); 1140 vm_page_free(m); 1141 } 1142 vm_map_delete(map, addr, addr + size, &count); 1143 vm_map_unlock(map); 1144 crit_exit(); 1145 vm_map_entry_release(count); 1146 rel_mplock(); 1147 return(NULL); 1148 } 1149 } 1150 1151 /* 1152 * Success! 1153 * 1154 * Mark the map entry as non-pageable using a routine that allows us to 1155 * populate the underlying pages. 1156 */ 1157 vm_map_set_wired_quick(map, addr, size, &count); 1158 crit_exit(); 1159 1160 /* 1161 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO. 1162 */ 1163 for (i = 0; i < size; i += PAGE_SIZE) { 1164 vm_page_t m; 1165 1166 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); 1167 m->valid = VM_PAGE_BITS_ALL; 1168 vm_page_wire(m); 1169 vm_page_wakeup(m); 1170 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1); 1171 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO)) 1172 bzero((char *)addr + i, PAGE_SIZE); 1173 vm_page_flag_clear(m, PG_ZERO); 1174 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED); 1175 } 1176 vm_map_unlock(map); 1177 vm_map_entry_release(count); 1178 rel_mplock(); 1179 return((void *)addr); 1180 } 1181 1182 /* 1183 * kmem_slab_free() 1184 * 1185 * MPALMOSTSAFE - acquires mplock 1186 */ 1187 static void 1188 kmem_slab_free(void *ptr, vm_size_t size) 1189 { 1190 get_mplock(); 1191 crit_enter(); 1192 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size); 1193 crit_exit(); 1194 rel_mplock(); 1195 } 1196 1197