1 /* 2 * KERN_SLABALLOC.C - Kernel SLAB memory allocator (MP SAFE) 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 1.35 2005/06/20 23:21:34 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 139 140 #define logmemory(name, ptr, type, size, flags) \ 141 KTR_LOG(memory_ ## name, ptr, type, size, flags) 142 143 /* 144 * Fixed globals (not per-cpu) 145 */ 146 static int ZoneSize; 147 static int ZoneLimit; 148 static int ZonePageCount; 149 static int ZoneMask; 150 static struct malloc_type *kmemstatistics; 151 static struct kmemusage *kmemusage; 152 static int32_t weirdary[16]; 153 154 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags); 155 static void kmem_slab_free(void *ptr, vm_size_t bytes); 156 #if defined(INVARIANTS) 157 static void chunk_mark_allocated(SLZone *z, void *chunk); 158 static void chunk_mark_free(SLZone *z, void *chunk); 159 #endif 160 161 /* 162 * Misc constants. Note that allocations that are exact multiples of 163 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module. 164 * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists. 165 */ 166 #define MIN_CHUNK_SIZE 8 /* in bytes */ 167 #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1) 168 #define ZONE_RELS_THRESH 2 /* threshold number of zones */ 169 #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK) 170 171 /* 172 * The WEIRD_ADDR is used as known text to copy into free objects to 173 * try to create deterministic failure cases if the data is accessed after 174 * free. 175 */ 176 #define WEIRD_ADDR 0xdeadc0de 177 #define MAX_COPY sizeof(weirdary) 178 #define ZERO_LENGTH_PTR ((void *)-8) 179 180 /* 181 * Misc global malloc buckets 182 */ 183 184 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches"); 185 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory"); 186 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers"); 187 188 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options"); 189 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery"); 190 191 /* 192 * Initialize the slab memory allocator. We have to choose a zone size based 193 * on available physical memory. We choose a zone side which is approximately 194 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of 195 * 128K. The zone size is limited to the bounds set in slaballoc.h 196 * (typically 32K min, 128K max). 197 */ 198 static void kmeminit(void *dummy); 199 200 SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL) 201 202 #ifdef INVARIANTS 203 /* 204 * If enabled any memory allocated without M_ZERO is initialized to -1. 205 */ 206 static int use_malloc_pattern; 207 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW, 208 &use_malloc_pattern, 0, ""); 209 #endif 210 211 static void 212 kmeminit(void *dummy) 213 { 214 vm_poff_t limsize; 215 int usesize; 216 int i; 217 vm_pindex_t npg; 218 219 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; 220 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) 221 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; 222 223 usesize = (int)(limsize / 1024); /* convert to KB */ 224 225 ZoneSize = ZALLOC_MIN_ZONE_SIZE; 226 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize) 227 ZoneSize <<= 1; 228 ZoneLimit = ZoneSize / 4; 229 if (ZoneLimit > ZALLOC_ZONE_LIMIT) 230 ZoneLimit = ZALLOC_ZONE_LIMIT; 231 ZoneMask = ZoneSize - 1; 232 ZonePageCount = ZoneSize / PAGE_SIZE; 233 234 npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE; 235 kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_WAITOK|M_ZERO); 236 237 for (i = 0; i < arysize(weirdary); ++i) 238 weirdary[i] = WEIRD_ADDR; 239 240 if (bootverbose) 241 printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024); 242 } 243 244 /* 245 * Initialize a malloc type tracking structure. 246 */ 247 void 248 malloc_init(void *data) 249 { 250 struct malloc_type *type = data; 251 vm_poff_t limsize; 252 253 if (type->ks_magic != M_MAGIC) 254 panic("malloc type lacks magic"); 255 256 if (type->ks_limit != 0) 257 return; 258 259 if (vmstats.v_page_count == 0) 260 panic("malloc_init not allowed before vm init"); 261 262 limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; 263 if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) 264 limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; 265 type->ks_limit = limsize / 10; 266 267 type->ks_next = kmemstatistics; 268 kmemstatistics = type; 269 } 270 271 void 272 malloc_uninit(void *data) 273 { 274 struct malloc_type *type = data; 275 struct malloc_type *t; 276 #ifdef INVARIANTS 277 int i; 278 long ttl; 279 #endif 280 281 if (type->ks_magic != M_MAGIC) 282 panic("malloc type lacks magic"); 283 284 if (vmstats.v_page_count == 0) 285 panic("malloc_uninit not allowed before vm init"); 286 287 if (type->ks_limit == 0) 288 panic("malloc_uninit on uninitialized type"); 289 290 #ifdef INVARIANTS 291 /* 292 * memuse is only correct in aggregation. Due to memory being allocated 293 * on one cpu and freed on another individual array entries may be 294 * negative or positive (canceling each other out). 295 */ 296 for (i = ttl = 0; i < ncpus; ++i) 297 ttl += type->ks_memuse[i]; 298 if (ttl) { 299 printf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n", 300 ttl, type->ks_shortdesc, i); 301 } 302 #endif 303 if (type == kmemstatistics) { 304 kmemstatistics = type->ks_next; 305 } else { 306 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) { 307 if (t->ks_next == type) { 308 t->ks_next = type->ks_next; 309 break; 310 } 311 } 312 } 313 type->ks_next = NULL; 314 type->ks_limit = 0; 315 } 316 317 /* 318 * Calculate the zone index for the allocation request size and set the 319 * allocation request size to that particular zone's chunk size. 320 */ 321 static __inline int 322 zoneindex(unsigned long *bytes) 323 { 324 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */ 325 if (n < 128) { 326 *bytes = n = (n + 7) & ~7; 327 return(n / 8 - 1); /* 8 byte chunks, 16 zones */ 328 } 329 if (n < 256) { 330 *bytes = n = (n + 15) & ~15; 331 return(n / 16 + 7); 332 } 333 if (n < 8192) { 334 if (n < 512) { 335 *bytes = n = (n + 31) & ~31; 336 return(n / 32 + 15); 337 } 338 if (n < 1024) { 339 *bytes = n = (n + 63) & ~63; 340 return(n / 64 + 23); 341 } 342 if (n < 2048) { 343 *bytes = n = (n + 127) & ~127; 344 return(n / 128 + 31); 345 } 346 if (n < 4096) { 347 *bytes = n = (n + 255) & ~255; 348 return(n / 256 + 39); 349 } 350 *bytes = n = (n + 511) & ~511; 351 return(n / 512 + 47); 352 } 353 #if ZALLOC_ZONE_LIMIT > 8192 354 if (n < 16384) { 355 *bytes = n = (n + 1023) & ~1023; 356 return(n / 1024 + 55); 357 } 358 #endif 359 #if ZALLOC_ZONE_LIMIT > 16384 360 if (n < 32768) { 361 *bytes = n = (n + 2047) & ~2047; 362 return(n / 2048 + 63); 363 } 364 #endif 365 panic("Unexpected byte count %d", n); 366 return(0); 367 } 368 369 /* 370 * malloc() (SLAB ALLOCATOR) (MP SAFE) 371 * 372 * Allocate memory via the slab allocator. If the request is too large, 373 * or if it page-aligned beyond a certain size, we fall back to the 374 * KMEM subsystem. A SLAB tracking descriptor must be specified, use 375 * &SlabMisc if you don't care. 376 * 377 * M_RNOWAIT - don't block. 378 * M_NULLOK - return NULL instead of blocking. 379 * M_ZERO - zero the returned memory. 380 * M_USE_RESERVE - allow greater drawdown of the free list 381 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted 382 */ 383 void * 384 malloc(unsigned long size, struct malloc_type *type, int flags) 385 { 386 SLZone *z; 387 SLChunk *chunk; 388 SLGlobalData *slgd; 389 struct globaldata *gd; 390 int zi; 391 #ifdef INVARIANTS 392 int i; 393 #endif 394 395 gd = mycpu; 396 slgd = &gd->gd_slab; 397 398 /* 399 * XXX silly to have this in the critical path. 400 */ 401 if (type->ks_limit == 0) { 402 crit_enter(); 403 if (type->ks_limit == 0) 404 malloc_init(type); 405 crit_exit(); 406 } 407 ++type->ks_calls; 408 409 /* 410 * Handle the case where the limit is reached. Panic if we can't return 411 * NULL. The original malloc code looped, but this tended to 412 * simply deadlock the computer. 413 * 414 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used 415 * to determine if a more complete limit check should be done. The 416 * actual memory use is tracked via ks_memuse[cpu]. 417 */ 418 while (type->ks_loosememuse >= type->ks_limit) { 419 int i; 420 long ttl; 421 422 for (i = ttl = 0; i < ncpus; ++i) 423 ttl += type->ks_memuse[i]; 424 type->ks_loosememuse = ttl; /* not MP synchronized */ 425 if (ttl >= type->ks_limit) { 426 if (flags & M_NULLOK) { 427 logmemory(malloc, NULL, type, size, flags); 428 return(NULL); 429 } 430 panic("%s: malloc limit exceeded", type->ks_shortdesc); 431 } 432 } 433 434 /* 435 * Handle the degenerate size == 0 case. Yes, this does happen. 436 * Return a special pointer. This is to maintain compatibility with 437 * the original malloc implementation. Certain devices, such as the 438 * adaptec driver, not only allocate 0 bytes, they check for NULL and 439 * also realloc() later on. Joy. 440 */ 441 if (size == 0) { 442 logmemory(malloc, ZERO_LENGTH_PTR, type, size, flags); 443 return(ZERO_LENGTH_PTR); 444 } 445 446 /* 447 * Handle hysteresis from prior frees here in malloc(). We cannot 448 * safely manipulate the kernel_map in free() due to free() possibly 449 * being called via an IPI message or from sensitive interrupt code. 450 */ 451 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) { 452 crit_enter(); 453 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */ 454 z = slgd->FreeZones; 455 slgd->FreeZones = z->z_Next; 456 --slgd->NFreeZones; 457 kmem_slab_free(z, ZoneSize); /* may block */ 458 } 459 crit_exit(); 460 } 461 /* 462 * XXX handle oversized frees that were queued from free(). 463 */ 464 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) { 465 crit_enter(); 466 if ((z = slgd->FreeOvZones) != NULL) { 467 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC); 468 slgd->FreeOvZones = z->z_Next; 469 kmem_slab_free(z, z->z_ChunkSize); /* may block */ 470 } 471 crit_exit(); 472 } 473 474 /* 475 * Handle large allocations directly. There should not be very many of 476 * these so performance is not a big issue. 477 * 478 * Guarentee page alignment for allocations in multiples of PAGE_SIZE 479 */ 480 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) { 481 struct kmemusage *kup; 482 483 size = round_page(size); 484 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags); 485 if (chunk == NULL) { 486 logmemory(malloc, NULL, type, size, flags); 487 return(NULL); 488 } 489 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */ 490 flags |= M_PASSIVE_ZERO; 491 kup = btokup(chunk); 492 kup->ku_pagecnt = size / PAGE_SIZE; 493 kup->ku_cpu = gd->gd_cpuid; 494 crit_enter(); 495 goto done; 496 } 497 498 /* 499 * Attempt to allocate out of an existing zone. First try the free list, 500 * then allocate out of unallocated space. If we find a good zone move 501 * it to the head of the list so later allocations find it quickly 502 * (we might have thousands of zones in the list). 503 * 504 * Note: zoneindex() will panic of size is too large. 505 */ 506 zi = zoneindex(&size); 507 KKASSERT(zi < NZONES); 508 crit_enter(); 509 if ((z = slgd->ZoneAry[zi]) != NULL) { 510 KKASSERT(z->z_NFree > 0); 511 512 /* 513 * Remove us from the ZoneAry[] when we become empty 514 */ 515 if (--z->z_NFree == 0) { 516 slgd->ZoneAry[zi] = z->z_Next; 517 z->z_Next = NULL; 518 } 519 520 /* 521 * Locate a chunk in a free page. This attempts to localize 522 * reallocations into earlier pages without us having to sort 523 * the chunk list. A chunk may still overlap a page boundary. 524 */ 525 while (z->z_FirstFreePg < ZonePageCount) { 526 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) { 527 #ifdef DIAGNOSTIC 528 /* 529 * Diagnostic: c_Next is not total garbage. 530 */ 531 KKASSERT(chunk->c_Next == NULL || 532 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) == 533 ((intptr_t)chunk & IN_SAME_PAGE_MASK)); 534 #endif 535 #ifdef INVARIANTS 536 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) 537 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount); 538 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) 539 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount); 540 chunk_mark_allocated(z, chunk); 541 #endif 542 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next; 543 goto done; 544 } 545 ++z->z_FirstFreePg; 546 } 547 548 /* 549 * No chunks are available but NFree said we had some memory, so 550 * it must be available in the never-before-used-memory area 551 * governed by UIndex. The consequences are very serious if our zone 552 * got corrupted so we use an explicit panic rather then a KASSERT. 553 */ 554 if (z->z_UIndex + 1 != z->z_NMax) 555 z->z_UIndex = z->z_UIndex + 1; 556 else 557 z->z_UIndex = 0; 558 if (z->z_UIndex == z->z_UEndIndex) 559 panic("slaballoc: corrupted zone"); 560 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 561 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 562 flags &= ~M_ZERO; 563 flags |= M_PASSIVE_ZERO; 564 } 565 #if defined(INVARIANTS) 566 chunk_mark_allocated(z, chunk); 567 #endif 568 goto done; 569 } 570 571 /* 572 * If all zones are exhausted we need to allocate a new zone for this 573 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see 574 * UAlloc use above in regards to M_ZERO. Note that when we are reusing 575 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and 576 * we do not pre-zero it because we do not want to mess up the L1 cache. 577 * 578 * At least one subsystem, the tty code (see CROUND) expects power-of-2 579 * allocations to be power-of-2 aligned. We maintain compatibility by 580 * adjusting the base offset below. 581 */ 582 { 583 int off; 584 585 if ((z = slgd->FreeZones) != NULL) { 586 slgd->FreeZones = z->z_Next; 587 --slgd->NFreeZones; 588 bzero(z, sizeof(SLZone)); 589 z->z_Flags |= SLZF_UNOTZEROD; 590 } else { 591 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO); 592 if (z == NULL) 593 goto fail; 594 } 595 596 /* 597 * How big is the base structure? 598 */ 599 #if defined(INVARIANTS) 600 /* 601 * Make room for z_Bitmap. An exact calculation is somewhat more 602 * complicated so don't make an exact calculation. 603 */ 604 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]); 605 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8); 606 #else 607 off = sizeof(SLZone); 608 #endif 609 610 /* 611 * Guarentee power-of-2 alignment for power-of-2-sized chunks. 612 * Otherwise just 8-byte align the data. 613 */ 614 if ((size | (size - 1)) + 1 == (size << 1)) 615 off = (off + size - 1) & ~(size - 1); 616 else 617 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK; 618 z->z_Magic = ZALLOC_SLAB_MAGIC; 619 z->z_ZoneIndex = zi; 620 z->z_NMax = (ZoneSize - off) / size; 621 z->z_NFree = z->z_NMax - 1; 622 z->z_BasePtr = (char *)z + off; 623 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax; 624 z->z_ChunkSize = size; 625 z->z_FirstFreePg = ZonePageCount; 626 z->z_CpuGd = gd; 627 z->z_Cpu = gd->gd_cpuid; 628 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 629 z->z_Next = slgd->ZoneAry[zi]; 630 slgd->ZoneAry[zi] = z; 631 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 632 flags &= ~M_ZERO; /* already zero'd */ 633 flags |= M_PASSIVE_ZERO; 634 } 635 #if defined(INVARIANTS) 636 chunk_mark_allocated(z, chunk); 637 #endif 638 639 /* 640 * Slide the base index for initial allocations out of the next 641 * zone we create so we do not over-weight the lower part of the 642 * cpu memory caches. 643 */ 644 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE) 645 & (ZALLOC_MAX_ZONE_SIZE - 1); 646 } 647 done: 648 ++type->ks_inuse[gd->gd_cpuid]; 649 type->ks_memuse[gd->gd_cpuid] += size; 650 type->ks_loosememuse += size; /* not MP synchronized */ 651 crit_exit(); 652 if (flags & M_ZERO) 653 bzero(chunk, size); 654 #ifdef INVARIANTS 655 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) { 656 if (use_malloc_pattern) { 657 for (i = 0; i < size; i += sizeof(int)) { 658 *(int *)((char *)chunk + i) = -1; 659 } 660 } 661 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */ 662 } 663 #endif 664 logmemory(malloc, chunk, type, size, flags); 665 return(chunk); 666 fail: 667 crit_exit(); 668 logmemory(malloc, NULL, type, size, flags); 669 return(NULL); 670 } 671 672 /* 673 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE) 674 * 675 * Generally speaking this routine is not called very often and we do 676 * not attempt to optimize it beyond reusing the same pointer if the 677 * new size fits within the chunking of the old pointer's zone. 678 */ 679 void * 680 realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags) 681 { 682 SLZone *z; 683 void *nptr; 684 unsigned long osize; 685 686 KKASSERT((flags & M_ZERO) == 0); /* not supported */ 687 688 if (ptr == NULL || ptr == ZERO_LENGTH_PTR) 689 return(malloc(size, type, flags)); 690 if (size == 0) { 691 free(ptr, type); 692 return(NULL); 693 } 694 695 /* 696 * Handle oversized allocations. XXX we really should require that a 697 * size be passed to free() instead of this nonsense. 698 */ 699 { 700 struct kmemusage *kup; 701 702 kup = btokup(ptr); 703 if (kup->ku_pagecnt) { 704 osize = kup->ku_pagecnt << PAGE_SHIFT; 705 if (osize == round_page(size)) 706 return(ptr); 707 if ((nptr = malloc(size, type, flags)) == NULL) 708 return(NULL); 709 bcopy(ptr, nptr, min(size, osize)); 710 free(ptr, type); 711 return(nptr); 712 } 713 } 714 715 /* 716 * Get the original allocation's zone. If the new request winds up 717 * using the same chunk size we do not have to do anything. 718 */ 719 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 720 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 721 722 zoneindex(&size); 723 if (z->z_ChunkSize == size) 724 return(ptr); 725 726 /* 727 * Allocate memory for the new request size. Note that zoneindex has 728 * already adjusted the request size to the appropriate chunk size, which 729 * should optimize our bcopy(). Then copy and return the new pointer. 730 */ 731 if ((nptr = malloc(size, type, flags)) == NULL) 732 return(NULL); 733 bcopy(ptr, nptr, min(size, z->z_ChunkSize)); 734 free(ptr, type); 735 return(nptr); 736 } 737 738 /* 739 * Allocate a copy of the specified string. 740 * 741 * (MP SAFE) (MAY BLOCK) 742 */ 743 char * 744 strdup(const char *str, struct malloc_type *type) 745 { 746 int zlen; /* length inclusive of terminating NUL */ 747 char *nstr; 748 749 if (str == NULL) 750 return(NULL); 751 zlen = strlen(str) + 1; 752 nstr = malloc(zlen, type, M_WAITOK); 753 bcopy(str, nstr, zlen); 754 return(nstr); 755 } 756 757 #ifdef SMP 758 /* 759 * free() (SLAB ALLOCATOR) 760 * 761 * Free the specified chunk of memory. 762 */ 763 static 764 void 765 free_remote(void *ptr) 766 { 767 logmemory(free_remote, ptr, *(struct malloc_type **)ptr, -1, 0); 768 free(ptr, *(struct malloc_type **)ptr); 769 } 770 771 #endif 772 773 /* 774 * free (SLAB ALLOCATOR) (MP SAFE) 775 * 776 * Free a memory block previously allocated by malloc. Note that we do not 777 * attempt to uplodate ks_loosememuse as MP races could prevent us from 778 * checking memory limits in malloc. 779 */ 780 void 781 free(void *ptr, struct malloc_type *type) 782 { 783 SLZone *z; 784 SLChunk *chunk; 785 SLGlobalData *slgd; 786 struct globaldata *gd; 787 int pgno; 788 789 gd = mycpu; 790 slgd = &gd->gd_slab; 791 792 if (ptr == NULL) 793 panic("trying to free NULL pointer"); 794 795 /* 796 * Handle special 0-byte allocations 797 */ 798 if (ptr == ZERO_LENGTH_PTR) { 799 logmemory(free_zero, ptr, type, -1, 0); 800 return; 801 } 802 803 /* 804 * Handle oversized allocations. XXX we really should require that a 805 * size be passed to free() instead of this nonsense. 806 * 807 * This code is never called via an ipi. 808 */ 809 { 810 struct kmemusage *kup; 811 unsigned long size; 812 813 kup = btokup(ptr); 814 if (kup->ku_pagecnt) { 815 size = kup->ku_pagecnt << PAGE_SHIFT; 816 kup->ku_pagecnt = 0; 817 #ifdef INVARIANTS 818 KKASSERT(sizeof(weirdary) <= size); 819 bcopy(weirdary, ptr, sizeof(weirdary)); 820 #endif 821 /* 822 * note: we always adjust our cpu's slot, not the originating 823 * cpu (kup->ku_cpuid). The statistics are in aggregate. 824 * 825 * note: XXX we have still inherited the interrupts-can't-block 826 * assumption. An interrupt thread does not bump 827 * gd_intr_nesting_level so check TDF_INTTHREAD. This is 828 * primarily until we can fix softupdate's assumptions about free(). 829 */ 830 crit_enter(); 831 --type->ks_inuse[gd->gd_cpuid]; 832 type->ks_memuse[gd->gd_cpuid] -= size; 833 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) { 834 logmemory(free_ovsz_delayed, ptr, type, size, 0); 835 z = (SLZone *)ptr; 836 z->z_Magic = ZALLOC_OVSZ_MAGIC; 837 z->z_Next = slgd->FreeOvZones; 838 z->z_ChunkSize = size; 839 slgd->FreeOvZones = z; 840 crit_exit(); 841 } else { 842 crit_exit(); 843 logmemory(free_ovsz, ptr, type, size, 0); 844 kmem_slab_free(ptr, size); /* may block */ 845 } 846 return; 847 } 848 } 849 850 /* 851 * Zone case. Figure out the zone based on the fact that it is 852 * ZoneSize aligned. 853 */ 854 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 855 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 856 857 /* 858 * If we do not own the zone then forward the request to the 859 * cpu that does. Since the timing is non-critical, a passive 860 * message is sent. 861 */ 862 if (z->z_CpuGd != gd) { 863 *(struct malloc_type **)ptr = type; 864 #ifdef SMP 865 logmemory(free_request, ptr, type, z->z_ChunkSize, 0); 866 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr); 867 #else 868 panic("Corrupt SLZone"); 869 #endif 870 return; 871 } 872 873 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0); 874 875 if (type->ks_magic != M_MAGIC) 876 panic("free: malloc type lacks magic"); 877 878 crit_enter(); 879 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT; 880 chunk = ptr; 881 882 #ifdef INVARIANTS 883 /* 884 * Attempt to detect a double-free. To reduce overhead we only check 885 * if there appears to be link pointer at the base of the data. 886 */ 887 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) { 888 SLChunk *scan; 889 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) { 890 if (scan == chunk) 891 panic("Double free at %p", chunk); 892 } 893 } 894 chunk_mark_free(z, chunk); 895 #endif 896 897 /* 898 * Put weird data into the memory to detect modifications after freeing, 899 * illegal pointer use after freeing (we should fault on the odd address), 900 * and so forth. XXX needs more work, see the old malloc code. 901 */ 902 #ifdef INVARIANTS 903 if (z->z_ChunkSize < sizeof(weirdary)) 904 bcopy(weirdary, chunk, z->z_ChunkSize); 905 else 906 bcopy(weirdary, chunk, sizeof(weirdary)); 907 #endif 908 909 /* 910 * Add this free non-zero'd chunk to a linked list for reuse, adjust 911 * z_FirstFreePg. 912 */ 913 #ifdef INVARIANTS 914 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) 915 panic("BADFREE %p", chunk); 916 #endif 917 chunk->c_Next = z->z_PageAry[pgno]; 918 z->z_PageAry[pgno] = chunk; 919 #ifdef INVARIANTS 920 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) 921 panic("BADFREE2"); 922 #endif 923 if (z->z_FirstFreePg > pgno) 924 z->z_FirstFreePg = pgno; 925 926 /* 927 * Bump the number of free chunks. If it becomes non-zero the zone 928 * must be added back onto the appropriate list. 929 */ 930 if (z->z_NFree++ == 0) { 931 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; 932 slgd->ZoneAry[z->z_ZoneIndex] = z; 933 } 934 935 --type->ks_inuse[z->z_Cpu]; 936 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize; 937 938 /* 939 * If the zone becomes totally free, and there are other zones we 940 * can allocate from, move this zone to the FreeZones list. Since 941 * this code can be called from an IPI callback, do *NOT* try to mess 942 * with kernel_map here. Hysteresis will be performed at malloc() time. 943 */ 944 if (z->z_NFree == z->z_NMax && 945 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) 946 ) { 947 SLZone **pz; 948 949 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next) 950 ; 951 *pz = z->z_Next; 952 z->z_Magic = -1; 953 z->z_Next = slgd->FreeZones; 954 slgd->FreeZones = z; 955 ++slgd->NFreeZones; 956 } 957 crit_exit(); 958 } 959 960 #if defined(INVARIANTS) 961 /* 962 * Helper routines for sanity checks 963 */ 964 static 965 void 966 chunk_mark_allocated(SLZone *z, void *chunk) 967 { 968 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 969 __uint32_t *bitptr; 970 971 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal", chunk, bitdex)); 972 bitptr = &z->z_Bitmap[bitdex >> 5]; 973 bitdex &= 31; 974 KASSERT((*bitptr & (1 << bitdex)) == 0, ("memory chunk %p is already allocated!", chunk)); 975 *bitptr |= 1 << bitdex; 976 } 977 978 static 979 void 980 chunk_mark_free(SLZone *z, void *chunk) 981 { 982 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 983 __uint32_t *bitptr; 984 985 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal!", chunk, bitdex)); 986 bitptr = &z->z_Bitmap[bitdex >> 5]; 987 bitdex &= 31; 988 KASSERT((*bitptr & (1 << bitdex)) != 0, ("memory chunk %p is already free!", chunk)); 989 *bitptr &= ~(1 << bitdex); 990 } 991 992 #endif 993 994 /* 995 * kmem_slab_alloc() (MP SAFE) (GETS BGL) 996 * 997 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the 998 * specified alignment. M_* flags are expected in the flags field. 999 * 1000 * Alignment must be a multiple of PAGE_SIZE. 1001 * 1002 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(), 1003 * but when we move zalloc() over to use this function as its backend 1004 * we will have to switch to kreserve/krelease and call reserve(0) 1005 * after the new space is made available. 1006 * 1007 * Interrupt code which has preempted other code is not allowed to 1008 * use PQ_CACHE pages. However, if an interrupt thread is run 1009 * non-preemptively or blocks and then runs non-preemptively, then 1010 * it is free to use PQ_CACHE pages. 1011 * 1012 * This routine will currently obtain the BGL. 1013 */ 1014 static void * 1015 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags) 1016 { 1017 vm_size_t i; 1018 vm_offset_t addr; 1019 vm_offset_t offset; 1020 int count, vmflags, base_vmflags; 1021 thread_t td; 1022 vm_map_t map = kernel_map; 1023 1024 size = round_page(size); 1025 addr = vm_map_min(map); 1026 1027 /* 1028 * Reserve properly aligned space from kernel_map. RNOWAIT allocations 1029 * cannot block. 1030 */ 1031 if (flags & M_RNOWAIT) { 1032 if (try_mplock() == 0) 1033 return(NULL); 1034 } else { 1035 get_mplock(); 1036 } 1037 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1038 crit_enter(); 1039 vm_map_lock(map); 1040 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) { 1041 vm_map_unlock(map); 1042 if ((flags & M_NULLOK) == 0) 1043 panic("kmem_slab_alloc(): kernel_map ran out of space!"); 1044 crit_exit(); 1045 vm_map_entry_release(count); 1046 rel_mplock(); 1047 return(NULL); 1048 } 1049 offset = addr - VM_MIN_KERNEL_ADDRESS; 1050 vm_object_reference(kernel_object); 1051 vm_map_insert(map, &count, 1052 kernel_object, offset, addr, addr + size, 1053 VM_PROT_ALL, VM_PROT_ALL, 0); 1054 1055 td = curthread; 1056 1057 base_vmflags = 0; 1058 if (flags & M_ZERO) 1059 base_vmflags |= VM_ALLOC_ZERO; 1060 if (flags & M_USE_RESERVE) 1061 base_vmflags |= VM_ALLOC_SYSTEM; 1062 if (flags & M_USE_INTERRUPT_RESERVE) 1063 base_vmflags |= VM_ALLOC_INTERRUPT; 1064 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) 1065 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]); 1066 1067 1068 /* 1069 * Allocate the pages. Do not mess with the PG_ZERO flag yet. 1070 */ 1071 for (i = 0; i < size; i += PAGE_SIZE) { 1072 vm_page_t m; 1073 vm_pindex_t idx = OFF_TO_IDX(offset + i); 1074 1075 /* 1076 * VM_ALLOC_NORMAL can only be set if we are not preempting. 1077 * 1078 * VM_ALLOC_SYSTEM is automatically set if we are preempting and 1079 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is 1080 * implied in this case), though I'm sure if we really need to do 1081 * that. 1082 */ 1083 vmflags = base_vmflags; 1084 if (flags & M_WAITOK) { 1085 if (td->td_preempted) 1086 vmflags |= VM_ALLOC_SYSTEM; 1087 else 1088 vmflags |= VM_ALLOC_NORMAL; 1089 } 1090 1091 m = vm_page_alloc(kernel_object, idx, vmflags); 1092 1093 /* 1094 * If the allocation failed we either return NULL or we retry. 1095 * 1096 * If M_WAITOK is specified we wait for more memory and retry. 1097 * If M_WAITOK is specified from a preemption we yield instead of 1098 * wait. Livelock will not occur because the interrupt thread 1099 * will not be preempting anyone the second time around after the 1100 * yield. 1101 */ 1102 if (m == NULL) { 1103 if (flags & M_WAITOK) { 1104 if (td->td_preempted) { 1105 vm_map_unlock(map); 1106 lwkt_yield(); 1107 vm_map_lock(map); 1108 } else { 1109 vm_map_unlock(map); 1110 vm_wait(); 1111 vm_map_lock(map); 1112 } 1113 i -= PAGE_SIZE; /* retry */ 1114 continue; 1115 } 1116 1117 /* 1118 * We were unable to recover, cleanup and return NULL 1119 */ 1120 while (i != 0) { 1121 i -= PAGE_SIZE; 1122 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); 1123 vm_page_free(m); 1124 } 1125 vm_map_delete(map, addr, addr + size, &count); 1126 vm_map_unlock(map); 1127 crit_exit(); 1128 vm_map_entry_release(count); 1129 rel_mplock(); 1130 return(NULL); 1131 } 1132 } 1133 1134 /* 1135 * Success! 1136 * 1137 * Mark the map entry as non-pageable using a routine that allows us to 1138 * populate the underlying pages. 1139 */ 1140 vm_map_set_wired_quick(map, addr, size, &count); 1141 crit_exit(); 1142 1143 /* 1144 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO. 1145 */ 1146 for (i = 0; i < size; i += PAGE_SIZE) { 1147 vm_page_t m; 1148 1149 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); 1150 m->valid = VM_PAGE_BITS_ALL; 1151 vm_page_wire(m); 1152 vm_page_wakeup(m); 1153 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1); 1154 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO)) 1155 bzero((char *)addr + i, PAGE_SIZE); 1156 vm_page_flag_clear(m, PG_ZERO); 1157 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED); 1158 } 1159 vm_map_unlock(map); 1160 vm_map_entry_release(count); 1161 rel_mplock(); 1162 return((void *)addr); 1163 } 1164 1165 /* 1166 * kmem_slab_free() (MP SAFE) (GETS BGL) 1167 */ 1168 static void 1169 kmem_slab_free(void *ptr, vm_size_t size) 1170 { 1171 get_mplock(); 1172 crit_enter(); 1173 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size); 1174 crit_exit(); 1175 rel_mplock(); 1176 } 1177 1178