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 1.37 2005/12/07 04:49:54 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) (MPSAFE) 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 void * 389 malloc(unsigned long size, struct malloc_type *type, int flags) 390 { 391 SLZone *z; 392 SLChunk *chunk; 393 SLGlobalData *slgd; 394 struct globaldata *gd; 395 int zi; 396 #ifdef INVARIANTS 397 int i; 398 #endif 399 400 logmemory_quick(malloc_beg); 401 gd = mycpu; 402 slgd = &gd->gd_slab; 403 404 /* 405 * XXX silly to have this in the critical path. 406 */ 407 if (type->ks_limit == 0) { 408 crit_enter(); 409 if (type->ks_limit == 0) 410 malloc_init(type); 411 crit_exit(); 412 } 413 ++type->ks_calls; 414 415 /* 416 * Handle the case where the limit is reached. Panic if we can't return 417 * NULL. The original malloc code looped, but this tended to 418 * simply deadlock the computer. 419 * 420 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used 421 * to determine if a more complete limit check should be done. The 422 * actual memory use is tracked via ks_memuse[cpu]. 423 */ 424 while (type->ks_loosememuse >= type->ks_limit) { 425 int i; 426 long ttl; 427 428 for (i = ttl = 0; i < ncpus; ++i) 429 ttl += type->ks_memuse[i]; 430 type->ks_loosememuse = ttl; /* not MP synchronized */ 431 if (ttl >= type->ks_limit) { 432 if (flags & M_NULLOK) { 433 logmemory(malloc, NULL, type, size, flags); 434 return(NULL); 435 } 436 panic("%s: malloc limit exceeded", type->ks_shortdesc); 437 } 438 } 439 440 /* 441 * Handle the degenerate size == 0 case. Yes, this does happen. 442 * Return a special pointer. This is to maintain compatibility with 443 * the original malloc implementation. Certain devices, such as the 444 * adaptec driver, not only allocate 0 bytes, they check for NULL and 445 * also realloc() later on. Joy. 446 */ 447 if (size == 0) { 448 logmemory(malloc, ZERO_LENGTH_PTR, type, size, flags); 449 return(ZERO_LENGTH_PTR); 450 } 451 452 /* 453 * Handle hysteresis from prior frees here in malloc(). We cannot 454 * safely manipulate the kernel_map in free() due to free() possibly 455 * being called via an IPI message or from sensitive interrupt code. 456 */ 457 while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_RNOWAIT) == 0) { 458 crit_enter(); 459 if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */ 460 z = slgd->FreeZones; 461 slgd->FreeZones = z->z_Next; 462 --slgd->NFreeZones; 463 kmem_slab_free(z, ZoneSize); /* may block */ 464 } 465 crit_exit(); 466 } 467 /* 468 * XXX handle oversized frees that were queued from free(). 469 */ 470 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) { 471 crit_enter(); 472 if ((z = slgd->FreeOvZones) != NULL) { 473 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC); 474 slgd->FreeOvZones = z->z_Next; 475 kmem_slab_free(z, z->z_ChunkSize); /* may block */ 476 } 477 crit_exit(); 478 } 479 480 /* 481 * Handle large allocations directly. There should not be very many of 482 * these so performance is not a big issue. 483 * 484 * Guarentee page alignment for allocations in multiples of PAGE_SIZE 485 */ 486 if (size >= ZoneLimit || (size & PAGE_MASK) == 0) { 487 struct kmemusage *kup; 488 489 size = round_page(size); 490 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags); 491 if (chunk == NULL) { 492 logmemory(malloc, NULL, type, size, flags); 493 return(NULL); 494 } 495 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */ 496 flags |= M_PASSIVE_ZERO; 497 kup = btokup(chunk); 498 kup->ku_pagecnt = size / PAGE_SIZE; 499 kup->ku_cpu = gd->gd_cpuid; 500 crit_enter(); 501 goto done; 502 } 503 504 /* 505 * Attempt to allocate out of an existing zone. First try the free list, 506 * then allocate out of unallocated space. If we find a good zone move 507 * it to the head of the list so later allocations find it quickly 508 * (we might have thousands of zones in the list). 509 * 510 * Note: zoneindex() will panic of size is too large. 511 */ 512 zi = zoneindex(&size); 513 KKASSERT(zi < NZONES); 514 crit_enter(); 515 if ((z = slgd->ZoneAry[zi]) != NULL) { 516 KKASSERT(z->z_NFree > 0); 517 518 /* 519 * Remove us from the ZoneAry[] when we become empty 520 */ 521 if (--z->z_NFree == 0) { 522 slgd->ZoneAry[zi] = z->z_Next; 523 z->z_Next = NULL; 524 } 525 526 /* 527 * Locate a chunk in a free page. This attempts to localize 528 * reallocations into earlier pages without us having to sort 529 * the chunk list. A chunk may still overlap a page boundary. 530 */ 531 while (z->z_FirstFreePg < ZonePageCount) { 532 if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) { 533 #ifdef DIAGNOSTIC 534 /* 535 * Diagnostic: c_Next is not total garbage. 536 */ 537 KKASSERT(chunk->c_Next == NULL || 538 ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) == 539 ((intptr_t)chunk & IN_SAME_PAGE_MASK)); 540 #endif 541 #ifdef INVARIANTS 542 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) 543 panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount); 544 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) 545 panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount); 546 chunk_mark_allocated(z, chunk); 547 #endif 548 z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next; 549 goto done; 550 } 551 ++z->z_FirstFreePg; 552 } 553 554 /* 555 * No chunks are available but NFree said we had some memory, so 556 * it must be available in the never-before-used-memory area 557 * governed by UIndex. The consequences are very serious if our zone 558 * got corrupted so we use an explicit panic rather then a KASSERT. 559 */ 560 if (z->z_UIndex + 1 != z->z_NMax) 561 z->z_UIndex = z->z_UIndex + 1; 562 else 563 z->z_UIndex = 0; 564 if (z->z_UIndex == z->z_UEndIndex) 565 panic("slaballoc: corrupted zone"); 566 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 567 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 568 flags &= ~M_ZERO; 569 flags |= M_PASSIVE_ZERO; 570 } 571 #if defined(INVARIANTS) 572 chunk_mark_allocated(z, chunk); 573 #endif 574 goto done; 575 } 576 577 /* 578 * If all zones are exhausted we need to allocate a new zone for this 579 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see 580 * UAlloc use above in regards to M_ZERO. Note that when we are reusing 581 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and 582 * we do not pre-zero it because we do not want to mess up the L1 cache. 583 * 584 * At least one subsystem, the tty code (see CROUND) expects power-of-2 585 * allocations to be power-of-2 aligned. We maintain compatibility by 586 * adjusting the base offset below. 587 */ 588 { 589 int off; 590 591 if ((z = slgd->FreeZones) != NULL) { 592 slgd->FreeZones = z->z_Next; 593 --slgd->NFreeZones; 594 bzero(z, sizeof(SLZone)); 595 z->z_Flags |= SLZF_UNOTZEROD; 596 } else { 597 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO); 598 if (z == NULL) 599 goto fail; 600 } 601 602 /* 603 * How big is the base structure? 604 */ 605 #if defined(INVARIANTS) 606 /* 607 * Make room for z_Bitmap. An exact calculation is somewhat more 608 * complicated so don't make an exact calculation. 609 */ 610 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]); 611 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8); 612 #else 613 off = sizeof(SLZone); 614 #endif 615 616 /* 617 * Guarentee power-of-2 alignment for power-of-2-sized chunks. 618 * Otherwise just 8-byte align the data. 619 */ 620 if ((size | (size - 1)) + 1 == (size << 1)) 621 off = (off + size - 1) & ~(size - 1); 622 else 623 off = (off + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK; 624 z->z_Magic = ZALLOC_SLAB_MAGIC; 625 z->z_ZoneIndex = zi; 626 z->z_NMax = (ZoneSize - off) / size; 627 z->z_NFree = z->z_NMax - 1; 628 z->z_BasePtr = (char *)z + off; 629 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax; 630 z->z_ChunkSize = size; 631 z->z_FirstFreePg = ZonePageCount; 632 z->z_CpuGd = gd; 633 z->z_Cpu = gd->gd_cpuid; 634 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 635 z->z_Next = slgd->ZoneAry[zi]; 636 slgd->ZoneAry[zi] = z; 637 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 638 flags &= ~M_ZERO; /* already zero'd */ 639 flags |= M_PASSIVE_ZERO; 640 } 641 #if defined(INVARIANTS) 642 chunk_mark_allocated(z, chunk); 643 #endif 644 645 /* 646 * Slide the base index for initial allocations out of the next 647 * zone we create so we do not over-weight the lower part of the 648 * cpu memory caches. 649 */ 650 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE) 651 & (ZALLOC_MAX_ZONE_SIZE - 1); 652 } 653 done: 654 ++type->ks_inuse[gd->gd_cpuid]; 655 type->ks_memuse[gd->gd_cpuid] += size; 656 type->ks_loosememuse += size; /* not MP synchronized */ 657 crit_exit(); 658 if (flags & M_ZERO) 659 bzero(chunk, size); 660 #ifdef INVARIANTS 661 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) { 662 if (use_malloc_pattern) { 663 for (i = 0; i < size; i += sizeof(int)) { 664 *(int *)((char *)chunk + i) = -1; 665 } 666 } 667 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */ 668 } 669 #endif 670 logmemory(malloc, chunk, type, size, flags); 671 return(chunk); 672 fail: 673 crit_exit(); 674 logmemory(malloc, NULL, type, size, flags); 675 return(NULL); 676 } 677 678 /* 679 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE) 680 * 681 * Generally speaking this routine is not called very often and we do 682 * not attempt to optimize it beyond reusing the same pointer if the 683 * new size fits within the chunking of the old pointer's zone. 684 */ 685 void * 686 realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags) 687 { 688 SLZone *z; 689 void *nptr; 690 unsigned long osize; 691 692 KKASSERT((flags & M_ZERO) == 0); /* not supported */ 693 694 if (ptr == NULL || ptr == ZERO_LENGTH_PTR) 695 return(malloc(size, type, flags)); 696 if (size == 0) { 697 free(ptr, type); 698 return(NULL); 699 } 700 701 /* 702 * Handle oversized allocations. XXX we really should require that a 703 * size be passed to free() instead of this nonsense. 704 */ 705 { 706 struct kmemusage *kup; 707 708 kup = btokup(ptr); 709 if (kup->ku_pagecnt) { 710 osize = kup->ku_pagecnt << PAGE_SHIFT; 711 if (osize == round_page(size)) 712 return(ptr); 713 if ((nptr = malloc(size, type, flags)) == NULL) 714 return(NULL); 715 bcopy(ptr, nptr, min(size, osize)); 716 free(ptr, type); 717 return(nptr); 718 } 719 } 720 721 /* 722 * Get the original allocation's zone. If the new request winds up 723 * using the same chunk size we do not have to do anything. 724 */ 725 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 726 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 727 728 zoneindex(&size); 729 if (z->z_ChunkSize == size) 730 return(ptr); 731 732 /* 733 * Allocate memory for the new request size. Note that zoneindex has 734 * already adjusted the request size to the appropriate chunk size, which 735 * should optimize our bcopy(). Then copy and return the new pointer. 736 */ 737 if ((nptr = malloc(size, type, flags)) == NULL) 738 return(NULL); 739 bcopy(ptr, nptr, min(size, z->z_ChunkSize)); 740 free(ptr, type); 741 return(nptr); 742 } 743 744 /* 745 * Allocate a copy of the specified string. 746 * 747 * (MP SAFE) (MAY BLOCK) 748 */ 749 char * 750 strdup(const char *str, struct malloc_type *type) 751 { 752 int zlen; /* length inclusive of terminating NUL */ 753 char *nstr; 754 755 if (str == NULL) 756 return(NULL); 757 zlen = strlen(str) + 1; 758 nstr = malloc(zlen, type, M_WAITOK); 759 bcopy(str, nstr, zlen); 760 return(nstr); 761 } 762 763 #ifdef SMP 764 /* 765 * free() (SLAB ALLOCATOR) 766 * 767 * Free the specified chunk of memory. 768 */ 769 static 770 void 771 free_remote(void *ptr) 772 { 773 logmemory(free_remote, ptr, *(struct malloc_type **)ptr, -1, 0); 774 free(ptr, *(struct malloc_type **)ptr); 775 } 776 777 #endif 778 779 /* 780 * free (SLAB ALLOCATOR) (MP SAFE) 781 * 782 * Free a memory block previously allocated by malloc. Note that we do not 783 * attempt to uplodate ks_loosememuse as MP races could prevent us from 784 * checking memory limits in malloc. 785 */ 786 void 787 free(void *ptr, struct malloc_type *type) 788 { 789 SLZone *z; 790 SLChunk *chunk; 791 SLGlobalData *slgd; 792 struct globaldata *gd; 793 int pgno; 794 795 logmemory_quick(free_beg); 796 gd = mycpu; 797 slgd = &gd->gd_slab; 798 799 if (ptr == NULL) 800 panic("trying to free NULL pointer"); 801 802 /* 803 * Handle special 0-byte allocations 804 */ 805 if (ptr == ZERO_LENGTH_PTR) { 806 logmemory(free_zero, ptr, type, -1, 0); 807 logmemory_quick(free_end); 808 return; 809 } 810 811 /* 812 * Handle oversized allocations. XXX we really should require that a 813 * size be passed to free() instead of this nonsense. 814 * 815 * This code is never called via an ipi. 816 */ 817 { 818 struct kmemusage *kup; 819 unsigned long size; 820 821 kup = btokup(ptr); 822 if (kup->ku_pagecnt) { 823 size = kup->ku_pagecnt << PAGE_SHIFT; 824 kup->ku_pagecnt = 0; 825 #ifdef INVARIANTS 826 KKASSERT(sizeof(weirdary) <= size); 827 bcopy(weirdary, ptr, sizeof(weirdary)); 828 #endif 829 /* 830 * note: we always adjust our cpu's slot, not the originating 831 * cpu (kup->ku_cpuid). The statistics are in aggregate. 832 * 833 * note: XXX we have still inherited the interrupts-can't-block 834 * assumption. An interrupt thread does not bump 835 * gd_intr_nesting_level so check TDF_INTTHREAD. This is 836 * primarily until we can fix softupdate's assumptions about free(). 837 */ 838 crit_enter(); 839 --type->ks_inuse[gd->gd_cpuid]; 840 type->ks_memuse[gd->gd_cpuid] -= size; 841 if (mycpu->gd_intr_nesting_level || (gd->gd_curthread->td_flags & TDF_INTTHREAD)) { 842 logmemory(free_ovsz_delayed, ptr, type, size, 0); 843 z = (SLZone *)ptr; 844 z->z_Magic = ZALLOC_OVSZ_MAGIC; 845 z->z_Next = slgd->FreeOvZones; 846 z->z_ChunkSize = size; 847 slgd->FreeOvZones = z; 848 crit_exit(); 849 } else { 850 crit_exit(); 851 logmemory(free_ovsz, ptr, type, size, 0); 852 kmem_slab_free(ptr, size); /* may block */ 853 } 854 logmemory_quick(free_end); 855 return; 856 } 857 } 858 859 /* 860 * Zone case. Figure out the zone based on the fact that it is 861 * ZoneSize aligned. 862 */ 863 z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); 864 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 865 866 /* 867 * If we do not own the zone then forward the request to the 868 * cpu that does. Since the timing is non-critical, a passive 869 * message is sent. 870 */ 871 if (z->z_CpuGd != gd) { 872 *(struct malloc_type **)ptr = type; 873 #ifdef SMP 874 logmemory(free_request, ptr, type, z->z_ChunkSize, 0); 875 lwkt_send_ipiq_passive(z->z_CpuGd, free_remote, ptr); 876 #else 877 panic("Corrupt SLZone"); 878 #endif 879 logmemory_quick(free_end); 880 return; 881 } 882 883 logmemory(free_chunk, ptr, type, z->z_ChunkSize, 0); 884 885 if (type->ks_magic != M_MAGIC) 886 panic("free: malloc type lacks magic"); 887 888 crit_enter(); 889 pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT; 890 chunk = ptr; 891 892 #ifdef INVARIANTS 893 /* 894 * Attempt to detect a double-free. To reduce overhead we only check 895 * if there appears to be link pointer at the base of the data. 896 */ 897 if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) { 898 SLChunk *scan; 899 for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) { 900 if (scan == chunk) 901 panic("Double free at %p", chunk); 902 } 903 } 904 chunk_mark_free(z, chunk); 905 #endif 906 907 /* 908 * Put weird data into the memory to detect modifications after freeing, 909 * illegal pointer use after freeing (we should fault on the odd address), 910 * and so forth. XXX needs more work, see the old malloc code. 911 */ 912 #ifdef INVARIANTS 913 if (z->z_ChunkSize < sizeof(weirdary)) 914 bcopy(weirdary, chunk, z->z_ChunkSize); 915 else 916 bcopy(weirdary, chunk, sizeof(weirdary)); 917 #endif 918 919 /* 920 * Add this free non-zero'd chunk to a linked list for reuse, adjust 921 * z_FirstFreePg. 922 */ 923 #ifdef INVARIANTS 924 if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) 925 panic("BADFREE %p", chunk); 926 #endif 927 chunk->c_Next = z->z_PageAry[pgno]; 928 z->z_PageAry[pgno] = chunk; 929 #ifdef INVARIANTS 930 if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) 931 panic("BADFREE2"); 932 #endif 933 if (z->z_FirstFreePg > pgno) 934 z->z_FirstFreePg = pgno; 935 936 /* 937 * Bump the number of free chunks. If it becomes non-zero the zone 938 * must be added back onto the appropriate list. 939 */ 940 if (z->z_NFree++ == 0) { 941 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; 942 slgd->ZoneAry[z->z_ZoneIndex] = z; 943 } 944 945 --type->ks_inuse[z->z_Cpu]; 946 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize; 947 948 /* 949 * If the zone becomes totally free, and there are other zones we 950 * can allocate from, move this zone to the FreeZones list. Since 951 * this code can be called from an IPI callback, do *NOT* try to mess 952 * with kernel_map here. Hysteresis will be performed at malloc() time. 953 */ 954 if (z->z_NFree == z->z_NMax && 955 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) 956 ) { 957 SLZone **pz; 958 959 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next) 960 ; 961 *pz = z->z_Next; 962 z->z_Magic = -1; 963 z->z_Next = slgd->FreeZones; 964 slgd->FreeZones = z; 965 ++slgd->NFreeZones; 966 } 967 logmemory_quick(free_end); 968 crit_exit(); 969 } 970 971 #if defined(INVARIANTS) 972 /* 973 * Helper routines for sanity checks 974 */ 975 static 976 void 977 chunk_mark_allocated(SLZone *z, void *chunk) 978 { 979 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 980 __uint32_t *bitptr; 981 982 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal", chunk, bitdex)); 983 bitptr = &z->z_Bitmap[bitdex >> 5]; 984 bitdex &= 31; 985 KASSERT((*bitptr & (1 << bitdex)) == 0, ("memory chunk %p is already allocated!", chunk)); 986 *bitptr |= 1 << bitdex; 987 } 988 989 static 990 void 991 chunk_mark_free(SLZone *z, void *chunk) 992 { 993 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 994 __uint32_t *bitptr; 995 996 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, ("memory chunk %p bit index %d is illegal!", chunk, bitdex)); 997 bitptr = &z->z_Bitmap[bitdex >> 5]; 998 bitdex &= 31; 999 KASSERT((*bitptr & (1 << bitdex)) != 0, ("memory chunk %p is already free!", chunk)); 1000 *bitptr &= ~(1 << bitdex); 1001 } 1002 1003 #endif 1004 1005 /* 1006 * kmem_slab_alloc() (MP SAFE) (GETS BGL) 1007 * 1008 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the 1009 * specified alignment. M_* flags are expected in the flags field. 1010 * 1011 * Alignment must be a multiple of PAGE_SIZE. 1012 * 1013 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(), 1014 * but when we move zalloc() over to use this function as its backend 1015 * we will have to switch to kreserve/krelease and call reserve(0) 1016 * after the new space is made available. 1017 * 1018 * Interrupt code which has preempted other code is not allowed to 1019 * use PQ_CACHE pages. However, if an interrupt thread is run 1020 * non-preemptively or blocks and then runs non-preemptively, then 1021 * it is free to use PQ_CACHE pages. 1022 * 1023 * This routine will currently obtain the BGL. 1024 */ 1025 static void * 1026 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags) 1027 { 1028 vm_size_t i; 1029 vm_offset_t addr; 1030 vm_offset_t offset; 1031 int count, vmflags, base_vmflags; 1032 thread_t td; 1033 vm_map_t map = kernel_map; 1034 1035 size = round_page(size); 1036 addr = vm_map_min(map); 1037 1038 /* 1039 * Reserve properly aligned space from kernel_map. RNOWAIT allocations 1040 * cannot block. 1041 */ 1042 if (flags & M_RNOWAIT) { 1043 if (try_mplock() == 0) 1044 return(NULL); 1045 } else { 1046 get_mplock(); 1047 } 1048 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1049 crit_enter(); 1050 vm_map_lock(map); 1051 if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) { 1052 vm_map_unlock(map); 1053 if ((flags & M_NULLOK) == 0) 1054 panic("kmem_slab_alloc(): kernel_map ran out of space!"); 1055 crit_exit(); 1056 vm_map_entry_release(count); 1057 rel_mplock(); 1058 return(NULL); 1059 } 1060 offset = addr - VM_MIN_KERNEL_ADDRESS; 1061 vm_object_reference(kernel_object); 1062 vm_map_insert(map, &count, 1063 kernel_object, offset, addr, addr + size, 1064 VM_PROT_ALL, VM_PROT_ALL, 0); 1065 1066 td = curthread; 1067 1068 base_vmflags = 0; 1069 if (flags & M_ZERO) 1070 base_vmflags |= VM_ALLOC_ZERO; 1071 if (flags & M_USE_RESERVE) 1072 base_vmflags |= VM_ALLOC_SYSTEM; 1073 if (flags & M_USE_INTERRUPT_RESERVE) 1074 base_vmflags |= VM_ALLOC_INTERRUPT; 1075 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) 1076 panic("kmem_slab_alloc: bad flags %08x (%p)", flags, ((int **)&size)[-1]); 1077 1078 1079 /* 1080 * Allocate the pages. Do not mess with the PG_ZERO flag yet. 1081 */ 1082 for (i = 0; i < size; i += PAGE_SIZE) { 1083 vm_page_t m; 1084 vm_pindex_t idx = OFF_TO_IDX(offset + i); 1085 1086 /* 1087 * VM_ALLOC_NORMAL can only be set if we are not preempting. 1088 * 1089 * VM_ALLOC_SYSTEM is automatically set if we are preempting and 1090 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is 1091 * implied in this case), though I'm sure if we really need to do 1092 * that. 1093 */ 1094 vmflags = base_vmflags; 1095 if (flags & M_WAITOK) { 1096 if (td->td_preempted) 1097 vmflags |= VM_ALLOC_SYSTEM; 1098 else 1099 vmflags |= VM_ALLOC_NORMAL; 1100 } 1101 1102 m = vm_page_alloc(kernel_object, idx, vmflags); 1103 1104 /* 1105 * If the allocation failed we either return NULL or we retry. 1106 * 1107 * If M_WAITOK is specified we wait for more memory and retry. 1108 * If M_WAITOK is specified from a preemption we yield instead of 1109 * wait. Livelock will not occur because the interrupt thread 1110 * will not be preempting anyone the second time around after the 1111 * yield. 1112 */ 1113 if (m == NULL) { 1114 if (flags & M_WAITOK) { 1115 if (td->td_preempted) { 1116 vm_map_unlock(map); 1117 lwkt_yield(); 1118 vm_map_lock(map); 1119 } else { 1120 vm_map_unlock(map); 1121 vm_wait(); 1122 vm_map_lock(map); 1123 } 1124 i -= PAGE_SIZE; /* retry */ 1125 continue; 1126 } 1127 1128 /* 1129 * We were unable to recover, cleanup and return NULL 1130 */ 1131 while (i != 0) { 1132 i -= PAGE_SIZE; 1133 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); 1134 vm_page_free(m); 1135 } 1136 vm_map_delete(map, addr, addr + size, &count); 1137 vm_map_unlock(map); 1138 crit_exit(); 1139 vm_map_entry_release(count); 1140 rel_mplock(); 1141 return(NULL); 1142 } 1143 } 1144 1145 /* 1146 * Success! 1147 * 1148 * Mark the map entry as non-pageable using a routine that allows us to 1149 * populate the underlying pages. 1150 */ 1151 vm_map_set_wired_quick(map, addr, size, &count); 1152 crit_exit(); 1153 1154 /* 1155 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO. 1156 */ 1157 for (i = 0; i < size; i += PAGE_SIZE) { 1158 vm_page_t m; 1159 1160 m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); 1161 m->valid = VM_PAGE_BITS_ALL; 1162 vm_page_wire(m); 1163 vm_page_wakeup(m); 1164 pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1); 1165 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO)) 1166 bzero((char *)addr + i, PAGE_SIZE); 1167 vm_page_flag_clear(m, PG_ZERO); 1168 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED); 1169 } 1170 vm_map_unlock(map); 1171 vm_map_entry_release(count); 1172 rel_mplock(); 1173 return((void *)addr); 1174 } 1175 1176 /* 1177 * kmem_slab_free() (MP SAFE) (GETS BGL) 1178 */ 1179 static void 1180 kmem_slab_free(void *ptr, vm_size_t size) 1181 { 1182 get_mplock(); 1183 crit_enter(); 1184 vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size); 1185 crit_exit(); 1186 rel_mplock(); 1187 } 1188 1189