1 /* 2 * (MPSAFE) 3 * 4 * KERN_SLABALLOC.C - Kernel SLAB memory allocator 5 * 6 * Copyright (c) 2003,2004,2010 The DragonFly Project. All rights reserved. 7 * 8 * This code is derived from software contributed to The DragonFly Project 9 * by Matthew Dillon <dillon@backplane.com> 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in 19 * the documentation and/or other materials provided with the 20 * distribution. 21 * 3. Neither the name of The DragonFly Project nor the names of its 22 * contributors may be used to endorse or promote products derived 23 * from this software without specific, prior written permission. 24 * 25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 28 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 29 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 30 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 31 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 32 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 33 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 34 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 35 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 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 * Alignment properties: 82 * - All power-of-2 sized allocations are power-of-2 aligned. 83 * - Allocations with M_POWEROF2 are power-of-2 aligned on the nearest 84 * power-of-2 round up of 'size'. 85 * - Non-power-of-2 sized allocations are zone chunk size aligned (see the 86 * above table 'Chunking' column). 87 * 88 * API REQUIREMENTS AND SIDE EFFECTS 89 * 90 * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we 91 * have remained compatible with the following API requirements: 92 * 93 * + malloc(0) is allowed and returns non-NULL (ahc driver) 94 * + ability to allocate arbitrarily large chunks of memory 95 */ 96 97 #include "opt_vm.h" 98 99 #include <sys/param.h> 100 #include <sys/systm.h> 101 #include <sys/kernel.h> 102 #include <sys/slaballoc.h> 103 #include <sys/mbuf.h> 104 #include <sys/vmmeter.h> 105 #include <sys/lock.h> 106 #include <sys/thread.h> 107 #include <sys/globaldata.h> 108 #include <sys/sysctl.h> 109 #include <sys/ktr.h> 110 111 #include <vm/vm.h> 112 #include <vm/vm_param.h> 113 #include <vm/vm_kern.h> 114 #include <vm/vm_extern.h> 115 #include <vm/vm_object.h> 116 #include <vm/pmap.h> 117 #include <vm/vm_map.h> 118 #include <vm/vm_page.h> 119 #include <vm/vm_pageout.h> 120 121 #include <machine/cpu.h> 122 123 #include <sys/thread2.h> 124 125 #define btokup(z) (&pmap_kvtom((vm_offset_t)(z))->ku_pagecnt) 126 127 #define MEMORY_STRING "ptr=%p type=%p size=%lu flags=%04x" 128 #define MEMORY_ARGS void *ptr, void *type, unsigned long size, int flags 129 130 #if !defined(KTR_MEMORY) 131 #define KTR_MEMORY KTR_ALL 132 #endif 133 KTR_INFO_MASTER(memory); 134 KTR_INFO(KTR_MEMORY, memory, malloc_beg, 0, "malloc begin"); 135 KTR_INFO(KTR_MEMORY, memory, malloc_end, 1, MEMORY_STRING, MEMORY_ARGS); 136 KTR_INFO(KTR_MEMORY, memory, free_zero, 2, MEMORY_STRING, MEMORY_ARGS); 137 KTR_INFO(KTR_MEMORY, memory, free_ovsz, 3, MEMORY_STRING, MEMORY_ARGS); 138 KTR_INFO(KTR_MEMORY, memory, free_ovsz_delayed, 4, MEMORY_STRING, MEMORY_ARGS); 139 KTR_INFO(KTR_MEMORY, memory, free_chunk, 5, MEMORY_STRING, MEMORY_ARGS); 140 #ifdef SMP 141 KTR_INFO(KTR_MEMORY, memory, free_request, 6, MEMORY_STRING, MEMORY_ARGS); 142 KTR_INFO(KTR_MEMORY, memory, free_rem_beg, 7, MEMORY_STRING, MEMORY_ARGS); 143 KTR_INFO(KTR_MEMORY, memory, free_rem_end, 8, MEMORY_STRING, MEMORY_ARGS); 144 #endif 145 KTR_INFO(KTR_MEMORY, memory, free_beg, 9, "free begin"); 146 KTR_INFO(KTR_MEMORY, memory, free_end, 10, "free end"); 147 148 #define logmemory(name, ptr, type, size, flags) \ 149 KTR_LOG(memory_ ## name, ptr, type, size, flags) 150 #define logmemory_quick(name) \ 151 KTR_LOG(memory_ ## name) 152 153 /* 154 * Fixed globals (not per-cpu) 155 */ 156 static int ZoneSize; 157 static int ZoneLimit; 158 static int ZonePageCount; 159 static uintptr_t ZoneMask; 160 static int ZoneBigAlloc; /* in KB */ 161 static int ZoneGenAlloc; /* in KB */ 162 struct malloc_type *kmemstatistics; /* exported to vmstat */ 163 static int32_t weirdary[16]; 164 165 static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags); 166 static void kmem_slab_free(void *ptr, vm_size_t bytes); 167 168 #if defined(INVARIANTS) 169 static void chunk_mark_allocated(SLZone *z, void *chunk); 170 static void chunk_mark_free(SLZone *z, void *chunk); 171 #else 172 #define chunk_mark_allocated(z, chunk) 173 #define chunk_mark_free(z, chunk) 174 #endif 175 176 /* 177 * Misc constants. Note that allocations that are exact multiples of 178 * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module. 179 */ 180 #define ZONE_RELS_THRESH 32 /* threshold number of zones */ 181 182 /* 183 * The WEIRD_ADDR is used as known text to copy into free objects to 184 * try to create deterministic failure cases if the data is accessed after 185 * free. 186 */ 187 #define WEIRD_ADDR 0xdeadc0de 188 #define MAX_COPY sizeof(weirdary) 189 #define ZERO_LENGTH_PTR ((void *)-8) 190 191 /* 192 * Misc global malloc buckets 193 */ 194 195 MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches"); 196 MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory"); 197 MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers"); 198 199 MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options"); 200 MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery"); 201 202 /* 203 * Initialize the slab memory allocator. We have to choose a zone size based 204 * on available physical memory. We choose a zone side which is approximately 205 * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of 206 * 128K. The zone size is limited to the bounds set in slaballoc.h 207 * (typically 32K min, 128K max). 208 */ 209 static void kmeminit(void *dummy); 210 211 char *ZeroPage; 212 213 SYSINIT(kmem, SI_BOOT1_ALLOCATOR, SI_ORDER_FIRST, kmeminit, NULL) 214 215 #ifdef INVARIANTS 216 /* 217 * If enabled any memory allocated without M_ZERO is initialized to -1. 218 */ 219 static int use_malloc_pattern; 220 SYSCTL_INT(_debug, OID_AUTO, use_malloc_pattern, CTLFLAG_RW, 221 &use_malloc_pattern, 0, 222 "Initialize memory to -1 if M_ZERO not specified"); 223 #endif 224 225 static int ZoneRelsThresh = ZONE_RELS_THRESH; 226 SYSCTL_INT(_kern, OID_AUTO, zone_big_alloc, CTLFLAG_RD, &ZoneBigAlloc, 0, ""); 227 SYSCTL_INT(_kern, OID_AUTO, zone_gen_alloc, CTLFLAG_RD, &ZoneGenAlloc, 0, ""); 228 SYSCTL_INT(_kern, OID_AUTO, zone_cache, CTLFLAG_RW, &ZoneRelsThresh, 0, ""); 229 230 /* 231 * Returns the kernel memory size limit for the purposes of initializing 232 * various subsystem caches. The smaller of available memory and the KVM 233 * memory space is returned. 234 * 235 * The size in megabytes is returned. 236 */ 237 size_t 238 kmem_lim_size(void) 239 { 240 size_t limsize; 241 242 limsize = (size_t)vmstats.v_page_count * PAGE_SIZE; 243 if (limsize > KvaSize) 244 limsize = KvaSize; 245 return (limsize / (1024 * 1024)); 246 } 247 248 static void 249 kmeminit(void *dummy) 250 { 251 size_t limsize; 252 int usesize; 253 int i; 254 255 limsize = kmem_lim_size(); 256 usesize = (int)(limsize * 1024); /* convert to KB */ 257 258 /* 259 * If the machine has a large KVM space and more than 8G of ram, 260 * double the zone release threshold to reduce SMP invalidations. 261 * If more than 16G of ram, do it again. 262 * 263 * The BIOS eats a little ram so add some slop. We want 8G worth of 264 * memory sticks to trigger the first adjustment. 265 */ 266 if (ZoneRelsThresh == ZONE_RELS_THRESH) { 267 if (limsize >= 7 * 1024) 268 ZoneRelsThresh *= 2; 269 if (limsize >= 15 * 1024) 270 ZoneRelsThresh *= 2; 271 } 272 273 /* 274 * Calculate the zone size. This typically calculates to 275 * ZALLOC_MAX_ZONE_SIZE 276 */ 277 ZoneSize = ZALLOC_MIN_ZONE_SIZE; 278 while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize) 279 ZoneSize <<= 1; 280 ZoneLimit = ZoneSize / 4; 281 if (ZoneLimit > ZALLOC_ZONE_LIMIT) 282 ZoneLimit = ZALLOC_ZONE_LIMIT; 283 ZoneMask = ~(uintptr_t)(ZoneSize - 1); 284 ZonePageCount = ZoneSize / PAGE_SIZE; 285 286 for (i = 0; i < NELEM(weirdary); ++i) 287 weirdary[i] = WEIRD_ADDR; 288 289 ZeroPage = kmem_slab_alloc(PAGE_SIZE, PAGE_SIZE, M_WAITOK|M_ZERO); 290 291 if (bootverbose) 292 kprintf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024); 293 } 294 295 /* 296 * Initialize a malloc type tracking structure. 297 */ 298 void 299 malloc_init(void *data) 300 { 301 struct malloc_type *type = data; 302 size_t limsize; 303 304 if (type->ks_magic != M_MAGIC) 305 panic("malloc type lacks magic"); 306 307 if (type->ks_limit != 0) 308 return; 309 310 if (vmstats.v_page_count == 0) 311 panic("malloc_init not allowed before vm init"); 312 313 limsize = kmem_lim_size() * (1024 * 1024); 314 type->ks_limit = limsize / 10; 315 316 type->ks_next = kmemstatistics; 317 kmemstatistics = type; 318 } 319 320 void 321 malloc_uninit(void *data) 322 { 323 struct malloc_type *type = data; 324 struct malloc_type *t; 325 #ifdef INVARIANTS 326 int i; 327 long ttl; 328 #endif 329 330 if (type->ks_magic != M_MAGIC) 331 panic("malloc type lacks magic"); 332 333 if (vmstats.v_page_count == 0) 334 panic("malloc_uninit not allowed before vm init"); 335 336 if (type->ks_limit == 0) 337 panic("malloc_uninit on uninitialized type"); 338 339 #ifdef SMP 340 /* Make sure that all pending kfree()s are finished. */ 341 lwkt_synchronize_ipiqs("muninit"); 342 #endif 343 344 #ifdef INVARIANTS 345 /* 346 * memuse is only correct in aggregation. Due to memory being allocated 347 * on one cpu and freed on another individual array entries may be 348 * negative or positive (canceling each other out). 349 */ 350 for (i = ttl = 0; i < ncpus; ++i) 351 ttl += type->ks_memuse[i]; 352 if (ttl) { 353 kprintf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n", 354 ttl, type->ks_shortdesc, i); 355 } 356 #endif 357 if (type == kmemstatistics) { 358 kmemstatistics = type->ks_next; 359 } else { 360 for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) { 361 if (t->ks_next == type) { 362 t->ks_next = type->ks_next; 363 break; 364 } 365 } 366 } 367 type->ks_next = NULL; 368 type->ks_limit = 0; 369 } 370 371 /* 372 * Increase the kmalloc pool limit for the specified pool. No changes 373 * are the made if the pool would shrink. 374 */ 375 void 376 kmalloc_raise_limit(struct malloc_type *type, size_t bytes) 377 { 378 if (type->ks_limit == 0) 379 malloc_init(type); 380 if (bytes == 0) 381 bytes = KvaSize; 382 if (type->ks_limit < bytes) 383 type->ks_limit = bytes; 384 } 385 386 /* 387 * Dynamically create a malloc pool. This function is a NOP if *typep is 388 * already non-NULL. 389 */ 390 void 391 kmalloc_create(struct malloc_type **typep, const char *descr) 392 { 393 struct malloc_type *type; 394 395 if (*typep == NULL) { 396 type = kmalloc(sizeof(*type), M_TEMP, M_WAITOK | M_ZERO); 397 type->ks_magic = M_MAGIC; 398 type->ks_shortdesc = descr; 399 malloc_init(type); 400 *typep = type; 401 } 402 } 403 404 /* 405 * Destroy a dynamically created malloc pool. This function is a NOP if 406 * the pool has already been destroyed. 407 */ 408 void 409 kmalloc_destroy(struct malloc_type **typep) 410 { 411 if (*typep != NULL) { 412 malloc_uninit(*typep); 413 kfree(*typep, M_TEMP); 414 *typep = NULL; 415 } 416 } 417 418 /* 419 * Calculate the zone index for the allocation request size and set the 420 * allocation request size to that particular zone's chunk size. 421 */ 422 static __inline int 423 zoneindex(unsigned long *bytes, unsigned long *align) 424 { 425 unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */ 426 if (n < 128) { 427 *bytes = n = (n + 7) & ~7; 428 *align = 8; 429 return(n / 8 - 1); /* 8 byte chunks, 16 zones */ 430 } 431 if (n < 256) { 432 *bytes = n = (n + 15) & ~15; 433 *align = 16; 434 return(n / 16 + 7); 435 } 436 if (n < 8192) { 437 if (n < 512) { 438 *bytes = n = (n + 31) & ~31; 439 *align = 32; 440 return(n / 32 + 15); 441 } 442 if (n < 1024) { 443 *bytes = n = (n + 63) & ~63; 444 *align = 64; 445 return(n / 64 + 23); 446 } 447 if (n < 2048) { 448 *bytes = n = (n + 127) & ~127; 449 *align = 128; 450 return(n / 128 + 31); 451 } 452 if (n < 4096) { 453 *bytes = n = (n + 255) & ~255; 454 *align = 256; 455 return(n / 256 + 39); 456 } 457 *bytes = n = (n + 511) & ~511; 458 *align = 512; 459 return(n / 512 + 47); 460 } 461 #if ZALLOC_ZONE_LIMIT > 8192 462 if (n < 16384) { 463 *bytes = n = (n + 1023) & ~1023; 464 *align = 1024; 465 return(n / 1024 + 55); 466 } 467 #endif 468 #if ZALLOC_ZONE_LIMIT > 16384 469 if (n < 32768) { 470 *bytes = n = (n + 2047) & ~2047; 471 *align = 2048; 472 return(n / 2048 + 63); 473 } 474 #endif 475 panic("Unexpected byte count %d", n); 476 return(0); 477 } 478 479 #ifdef SLAB_DEBUG 480 /* 481 * Used to debug memory corruption issues. Record up to (typically 32) 482 * allocation sources for this zone (for a particular chunk size). 483 */ 484 485 static void 486 slab_record_source(SLZone *z, const char *file, int line) 487 { 488 int i; 489 int b = line & (SLAB_DEBUG_ENTRIES - 1); 490 491 i = b; 492 do { 493 if (z->z_Sources[i].file == file && z->z_Sources[i].line == line) 494 return; 495 if (z->z_Sources[i].file == NULL) 496 break; 497 i = (i + 1) & (SLAB_DEBUG_ENTRIES - 1); 498 } while (i != b); 499 z->z_Sources[i].file = file; 500 z->z_Sources[i].line = line; 501 } 502 503 #endif 504 505 static __inline unsigned long 506 powerof2_size(unsigned long size) 507 { 508 int i; 509 510 if (size == 0 || powerof2(size)) 511 return size; 512 513 i = flsl(size); 514 return (1UL << i); 515 } 516 517 /* 518 * kmalloc() (SLAB ALLOCATOR) 519 * 520 * Allocate memory via the slab allocator. If the request is too large, 521 * or if it page-aligned beyond a certain size, we fall back to the 522 * KMEM subsystem. A SLAB tracking descriptor must be specified, use 523 * &SlabMisc if you don't care. 524 * 525 * M_RNOWAIT - don't block. 526 * M_NULLOK - return NULL instead of blocking. 527 * M_ZERO - zero the returned memory. 528 * M_USE_RESERVE - allow greater drawdown of the free list 529 * M_USE_INTERRUPT_RESERVE - allow the freelist to be exhausted 530 * M_POWEROF2 - roundup size to the nearest power of 2 531 * 532 * MPSAFE 533 */ 534 535 #ifdef SLAB_DEBUG 536 void * 537 kmalloc_debug(unsigned long size, struct malloc_type *type, int flags, 538 const char *file, int line) 539 #else 540 void * 541 kmalloc(unsigned long size, struct malloc_type *type, int flags) 542 #endif 543 { 544 SLZone *z; 545 SLChunk *chunk; 546 #ifdef SMP 547 SLChunk *bchunk; 548 #endif 549 SLGlobalData *slgd; 550 struct globaldata *gd; 551 unsigned long align; 552 int zi; 553 #ifdef INVARIANTS 554 int i; 555 #endif 556 557 logmemory_quick(malloc_beg); 558 gd = mycpu; 559 slgd = &gd->gd_slab; 560 561 /* 562 * XXX silly to have this in the critical path. 563 */ 564 if (type->ks_limit == 0) { 565 crit_enter(); 566 if (type->ks_limit == 0) 567 malloc_init(type); 568 crit_exit(); 569 } 570 ++type->ks_calls; 571 572 if (flags & M_POWEROF2) 573 size = powerof2_size(size); 574 575 /* 576 * Handle the case where the limit is reached. Panic if we can't return 577 * NULL. The original malloc code looped, but this tended to 578 * simply deadlock the computer. 579 * 580 * ks_loosememuse is an up-only limit that is NOT MP-synchronized, used 581 * to determine if a more complete limit check should be done. The 582 * actual memory use is tracked via ks_memuse[cpu]. 583 */ 584 while (type->ks_loosememuse >= type->ks_limit) { 585 int i; 586 long ttl; 587 588 for (i = ttl = 0; i < ncpus; ++i) 589 ttl += type->ks_memuse[i]; 590 type->ks_loosememuse = ttl; /* not MP synchronized */ 591 if ((ssize_t)ttl < 0) /* deal with occassional race */ 592 ttl = 0; 593 if (ttl >= type->ks_limit) { 594 if (flags & M_NULLOK) { 595 logmemory(malloc_end, NULL, type, size, flags); 596 return(NULL); 597 } 598 panic("%s: malloc limit exceeded", type->ks_shortdesc); 599 } 600 } 601 602 /* 603 * Handle the degenerate size == 0 case. Yes, this does happen. 604 * Return a special pointer. This is to maintain compatibility with 605 * the original malloc implementation. Certain devices, such as the 606 * adaptec driver, not only allocate 0 bytes, they check for NULL and 607 * also realloc() later on. Joy. 608 */ 609 if (size == 0) { 610 logmemory(malloc_end, ZERO_LENGTH_PTR, type, size, flags); 611 return(ZERO_LENGTH_PTR); 612 } 613 614 /* 615 * Handle hysteresis from prior frees here in malloc(). We cannot 616 * safely manipulate the kernel_map in free() due to free() possibly 617 * being called via an IPI message or from sensitive interrupt code. 618 * 619 * NOTE: ku_pagecnt must be cleared before we free the slab or we 620 * might race another cpu allocating the kva and setting 621 * ku_pagecnt. 622 */ 623 while (slgd->NFreeZones > ZoneRelsThresh && (flags & M_RNOWAIT) == 0) { 624 crit_enter(); 625 if (slgd->NFreeZones > ZoneRelsThresh) { /* crit sect race */ 626 int *kup; 627 628 z = slgd->FreeZones; 629 slgd->FreeZones = z->z_Next; 630 --slgd->NFreeZones; 631 kup = btokup(z); 632 *kup = 0; 633 kmem_slab_free(z, ZoneSize); /* may block */ 634 atomic_add_int(&ZoneGenAlloc, -ZoneSize / 1024); 635 } 636 crit_exit(); 637 } 638 639 /* 640 * XXX handle oversized frees that were queued from kfree(). 641 */ 642 while (slgd->FreeOvZones && (flags & M_RNOWAIT) == 0) { 643 crit_enter(); 644 if ((z = slgd->FreeOvZones) != NULL) { 645 vm_size_t tsize; 646 647 KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC); 648 slgd->FreeOvZones = z->z_Next; 649 tsize = z->z_ChunkSize; 650 kmem_slab_free(z, tsize); /* may block */ 651 atomic_add_int(&ZoneBigAlloc, -(int)tsize / 1024); 652 } 653 crit_exit(); 654 } 655 656 /* 657 * Handle large allocations directly. There should not be very many of 658 * these so performance is not a big issue. 659 * 660 * The backend allocator is pretty nasty on a SMP system. Use the 661 * slab allocator for one and two page-sized chunks even though we lose 662 * some efficiency. XXX maybe fix mmio and the elf loader instead. 663 */ 664 if (size >= ZoneLimit || ((size & PAGE_MASK) == 0 && size > PAGE_SIZE*2)) { 665 int *kup; 666 667 size = round_page(size); 668 chunk = kmem_slab_alloc(size, PAGE_SIZE, flags); 669 if (chunk == NULL) { 670 logmemory(malloc_end, NULL, type, size, flags); 671 return(NULL); 672 } 673 atomic_add_int(&ZoneBigAlloc, (int)size / 1024); 674 flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */ 675 flags |= M_PASSIVE_ZERO; 676 kup = btokup(chunk); 677 *kup = size / PAGE_SIZE; 678 crit_enter(); 679 goto done; 680 } 681 682 /* 683 * Attempt to allocate out of an existing zone. First try the free list, 684 * then allocate out of unallocated space. If we find a good zone move 685 * it to the head of the list so later allocations find it quickly 686 * (we might have thousands of zones in the list). 687 * 688 * Note: zoneindex() will panic of size is too large. 689 */ 690 zi = zoneindex(&size, &align); 691 KKASSERT(zi < NZONES); 692 crit_enter(); 693 694 if ((z = slgd->ZoneAry[zi]) != NULL) { 695 /* 696 * Locate a chunk - we have to have at least one. If this is the 697 * last chunk go ahead and do the work to retrieve chunks freed 698 * from remote cpus, and if the zone is still empty move it off 699 * the ZoneAry. 700 */ 701 if (--z->z_NFree <= 0) { 702 KKASSERT(z->z_NFree == 0); 703 704 #ifdef SMP 705 /* 706 * WARNING! This code competes with other cpus. It is ok 707 * for us to not drain RChunks here but we might as well, and 708 * it is ok if more accumulate after we're done. 709 * 710 * Set RSignal before pulling rchunks off, indicating that we 711 * will be moving ourselves off of the ZoneAry. Remote ends will 712 * read RSignal before putting rchunks on thus interlocking 713 * their IPI signaling. 714 */ 715 if (z->z_RChunks == NULL) 716 atomic_swap_int(&z->z_RSignal, 1); 717 718 while ((bchunk = z->z_RChunks) != NULL) { 719 cpu_ccfence(); 720 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) { 721 *z->z_LChunksp = bchunk; 722 while (bchunk) { 723 chunk_mark_free(z, bchunk); 724 z->z_LChunksp = &bchunk->c_Next; 725 bchunk = bchunk->c_Next; 726 ++z->z_NFree; 727 } 728 break; 729 } 730 } 731 #endif 732 /* 733 * Remove from the zone list if no free chunks remain. 734 * Clear RSignal 735 */ 736 if (z->z_NFree == 0) { 737 slgd->ZoneAry[zi] = z->z_Next; 738 z->z_Next = NULL; 739 } else { 740 z->z_RSignal = 0; 741 } 742 } 743 744 /* 745 * Fast path, we have chunks available in z_LChunks. 746 */ 747 chunk = z->z_LChunks; 748 if (chunk) { 749 chunk_mark_allocated(z, chunk); 750 z->z_LChunks = chunk->c_Next; 751 if (z->z_LChunks == NULL) 752 z->z_LChunksp = &z->z_LChunks; 753 #ifdef SLAB_DEBUG 754 slab_record_source(z, file, line); 755 #endif 756 goto done; 757 } 758 759 /* 760 * No chunks are available in LChunks, the free chunk MUST be 761 * in the never-before-used memory area, controlled by UIndex. 762 * 763 * The consequences are very serious if our zone got corrupted so 764 * we use an explicit panic rather than a KASSERT. 765 */ 766 if (z->z_UIndex + 1 != z->z_NMax) 767 ++z->z_UIndex; 768 else 769 z->z_UIndex = 0; 770 771 if (z->z_UIndex == z->z_UEndIndex) 772 panic("slaballoc: corrupted zone"); 773 774 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 775 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 776 flags &= ~M_ZERO; 777 flags |= M_PASSIVE_ZERO; 778 } 779 chunk_mark_allocated(z, chunk); 780 #ifdef SLAB_DEBUG 781 slab_record_source(z, file, line); 782 #endif 783 goto done; 784 } 785 786 /* 787 * If all zones are exhausted we need to allocate a new zone for this 788 * index. Use M_ZERO to take advantage of pre-zerod pages. Also see 789 * UAlloc use above in regards to M_ZERO. Note that when we are reusing 790 * a zone from the FreeZones list UAlloc'd data will not be zero'd, and 791 * we do not pre-zero it because we do not want to mess up the L1 cache. 792 * 793 * At least one subsystem, the tty code (see CROUND) expects power-of-2 794 * allocations to be power-of-2 aligned. We maintain compatibility by 795 * adjusting the base offset below. 796 */ 797 { 798 int off; 799 int *kup; 800 801 if ((z = slgd->FreeZones) != NULL) { 802 slgd->FreeZones = z->z_Next; 803 --slgd->NFreeZones; 804 bzero(z, sizeof(SLZone)); 805 z->z_Flags |= SLZF_UNOTZEROD; 806 } else { 807 z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO); 808 if (z == NULL) 809 goto fail; 810 atomic_add_int(&ZoneGenAlloc, ZoneSize / 1024); 811 } 812 813 /* 814 * How big is the base structure? 815 */ 816 #if defined(INVARIANTS) 817 /* 818 * Make room for z_Bitmap. An exact calculation is somewhat more 819 * complicated so don't make an exact calculation. 820 */ 821 off = offsetof(SLZone, z_Bitmap[(ZoneSize / size + 31) / 32]); 822 bzero(z->z_Bitmap, (ZoneSize / size + 31) / 8); 823 #else 824 off = sizeof(SLZone); 825 #endif 826 827 /* 828 * Guarentee power-of-2 alignment for power-of-2-sized chunks. 829 * Otherwise properly align the data according to the chunk size. 830 */ 831 if (powerof2(size)) 832 align = size; 833 off = (off + align - 1) & ~(align - 1); 834 835 z->z_Magic = ZALLOC_SLAB_MAGIC; 836 z->z_ZoneIndex = zi; 837 z->z_NMax = (ZoneSize - off) / size; 838 z->z_NFree = z->z_NMax - 1; 839 z->z_BasePtr = (char *)z + off; 840 z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax; 841 z->z_ChunkSize = size; 842 z->z_CpuGd = gd; 843 z->z_Cpu = gd->gd_cpuid; 844 z->z_LChunksp = &z->z_LChunks; 845 #ifdef SLAB_DEBUG 846 bcopy(z->z_Sources, z->z_AltSources, sizeof(z->z_Sources)); 847 bzero(z->z_Sources, sizeof(z->z_Sources)); 848 #endif 849 chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); 850 z->z_Next = slgd->ZoneAry[zi]; 851 slgd->ZoneAry[zi] = z; 852 if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { 853 flags &= ~M_ZERO; /* already zero'd */ 854 flags |= M_PASSIVE_ZERO; 855 } 856 kup = btokup(z); 857 *kup = -(z->z_Cpu + 1); /* -1 to -(N+1) */ 858 chunk_mark_allocated(z, chunk); 859 #ifdef SLAB_DEBUG 860 slab_record_source(z, file, line); 861 #endif 862 863 /* 864 * Slide the base index for initial allocations out of the next 865 * zone we create so we do not over-weight the lower part of the 866 * cpu memory caches. 867 */ 868 slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE) 869 & (ZALLOC_MAX_ZONE_SIZE - 1); 870 } 871 872 done: 873 ++type->ks_inuse[gd->gd_cpuid]; 874 type->ks_memuse[gd->gd_cpuid] += size; 875 type->ks_loosememuse += size; /* not MP synchronized */ 876 crit_exit(); 877 878 if (flags & M_ZERO) 879 bzero(chunk, size); 880 #ifdef INVARIANTS 881 else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) { 882 if (use_malloc_pattern) { 883 for (i = 0; i < size; i += sizeof(int)) { 884 *(int *)((char *)chunk + i) = -1; 885 } 886 } 887 chunk->c_Next = (void *)-1; /* avoid accidental double-free check */ 888 } 889 #endif 890 logmemory(malloc_end, chunk, type, size, flags); 891 return(chunk); 892 fail: 893 crit_exit(); 894 logmemory(malloc_end, NULL, type, size, flags); 895 return(NULL); 896 } 897 898 /* 899 * kernel realloc. (SLAB ALLOCATOR) (MP SAFE) 900 * 901 * Generally speaking this routine is not called very often and we do 902 * not attempt to optimize it beyond reusing the same pointer if the 903 * new size fits within the chunking of the old pointer's zone. 904 */ 905 #ifdef SLAB_DEBUG 906 void * 907 krealloc_debug(void *ptr, unsigned long size, 908 struct malloc_type *type, int flags, 909 const char *file, int line) 910 #else 911 void * 912 krealloc(void *ptr, unsigned long size, struct malloc_type *type, int flags) 913 #endif 914 { 915 unsigned long osize; 916 unsigned long align; 917 SLZone *z; 918 void *nptr; 919 int *kup; 920 921 KKASSERT((flags & M_ZERO) == 0); /* not supported */ 922 923 if (ptr == NULL || ptr == ZERO_LENGTH_PTR) 924 return(kmalloc_debug(size, type, flags, file, line)); 925 if (size == 0) { 926 kfree(ptr, type); 927 return(NULL); 928 } 929 930 /* 931 * Handle oversized allocations. XXX we really should require that a 932 * size be passed to free() instead of this nonsense. 933 */ 934 kup = btokup(ptr); 935 if (*kup > 0) { 936 osize = *kup << PAGE_SHIFT; 937 if (osize == round_page(size)) 938 return(ptr); 939 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL) 940 return(NULL); 941 bcopy(ptr, nptr, min(size, osize)); 942 kfree(ptr, type); 943 return(nptr); 944 } 945 946 /* 947 * Get the original allocation's zone. If the new request winds up 948 * using the same chunk size we do not have to do anything. 949 */ 950 z = (SLZone *)((uintptr_t)ptr & ZoneMask); 951 kup = btokup(z); 952 KKASSERT(*kup < 0); 953 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 954 955 /* 956 * Allocate memory for the new request size. Note that zoneindex has 957 * already adjusted the request size to the appropriate chunk size, which 958 * should optimize our bcopy(). Then copy and return the new pointer. 959 * 960 * Resizing a non-power-of-2 allocation to a power-of-2 size does not 961 * necessary align the result. 962 * 963 * We can only zoneindex (to align size to the chunk size) if the new 964 * size is not too large. 965 */ 966 if (size < ZoneLimit) { 967 zoneindex(&size, &align); 968 if (z->z_ChunkSize == size) 969 return(ptr); 970 } 971 if ((nptr = kmalloc_debug(size, type, flags, file, line)) == NULL) 972 return(NULL); 973 bcopy(ptr, nptr, min(size, z->z_ChunkSize)); 974 kfree(ptr, type); 975 return(nptr); 976 } 977 978 /* 979 * Return the kmalloc limit for this type, in bytes. 980 */ 981 long 982 kmalloc_limit(struct malloc_type *type) 983 { 984 if (type->ks_limit == 0) { 985 crit_enter(); 986 if (type->ks_limit == 0) 987 malloc_init(type); 988 crit_exit(); 989 } 990 return(type->ks_limit); 991 } 992 993 /* 994 * Allocate a copy of the specified string. 995 * 996 * (MP SAFE) (MAY BLOCK) 997 */ 998 #ifdef SLAB_DEBUG 999 char * 1000 kstrdup_debug(const char *str, struct malloc_type *type, 1001 const char *file, int line) 1002 #else 1003 char * 1004 kstrdup(const char *str, struct malloc_type *type) 1005 #endif 1006 { 1007 int zlen; /* length inclusive of terminating NUL */ 1008 char *nstr; 1009 1010 if (str == NULL) 1011 return(NULL); 1012 zlen = strlen(str) + 1; 1013 nstr = kmalloc_debug(zlen, type, M_WAITOK, file, line); 1014 bcopy(str, nstr, zlen); 1015 return(nstr); 1016 } 1017 1018 #ifdef SMP 1019 /* 1020 * Notify our cpu that a remote cpu has freed some chunks in a zone that 1021 * we own. RCount will be bumped so the memory should be good, but validate 1022 * that it really is. 1023 */ 1024 static 1025 void 1026 kfree_remote(void *ptr) 1027 { 1028 SLGlobalData *slgd; 1029 SLChunk *bchunk; 1030 SLZone *z; 1031 int nfree; 1032 int *kup; 1033 1034 slgd = &mycpu->gd_slab; 1035 z = ptr; 1036 kup = btokup(z); 1037 KKASSERT(*kup == -((int)mycpuid + 1)); 1038 KKASSERT(z->z_RCount > 0); 1039 atomic_subtract_int(&z->z_RCount, 1); 1040 1041 logmemory(free_rem_beg, z, NULL, 0L, 0); 1042 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 1043 KKASSERT(z->z_Cpu == mycpu->gd_cpuid); 1044 nfree = z->z_NFree; 1045 1046 /* 1047 * Indicate that we will no longer be off of the ZoneAry by 1048 * clearing RSignal. 1049 */ 1050 if (z->z_RChunks) 1051 z->z_RSignal = 0; 1052 1053 /* 1054 * Atomically extract the bchunks list and then process it back 1055 * into the lchunks list. We want to append our bchunks to the 1056 * lchunks list and not prepend since we likely do not have 1057 * cache mastership of the related data (not that it helps since 1058 * we are using c_Next). 1059 */ 1060 while ((bchunk = z->z_RChunks) != NULL) { 1061 cpu_ccfence(); 1062 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, NULL)) { 1063 *z->z_LChunksp = bchunk; 1064 while (bchunk) { 1065 chunk_mark_free(z, bchunk); 1066 z->z_LChunksp = &bchunk->c_Next; 1067 bchunk = bchunk->c_Next; 1068 ++z->z_NFree; 1069 } 1070 break; 1071 } 1072 } 1073 if (z->z_NFree && nfree == 0) { 1074 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; 1075 slgd->ZoneAry[z->z_ZoneIndex] = z; 1076 } 1077 1078 /* 1079 * If the zone becomes totally free, and there are other zones we 1080 * can allocate from, move this zone to the FreeZones list. Since 1081 * this code can be called from an IPI callback, do *NOT* try to mess 1082 * with kernel_map here. Hysteresis will be performed at malloc() time. 1083 * 1084 * Do not move the zone if there is an IPI inflight, otherwise MP 1085 * races can result in our free_remote code accessing a destroyed 1086 * zone. 1087 */ 1088 if (z->z_NFree == z->z_NMax && 1089 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) && 1090 z->z_RCount == 0 1091 ) { 1092 SLZone **pz; 1093 int *kup; 1094 1095 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; 1096 z != *pz; 1097 pz = &(*pz)->z_Next) { 1098 ; 1099 } 1100 *pz = z->z_Next; 1101 z->z_Magic = -1; 1102 z->z_Next = slgd->FreeZones; 1103 slgd->FreeZones = z; 1104 ++slgd->NFreeZones; 1105 kup = btokup(z); 1106 *kup = 0; 1107 } 1108 logmemory(free_rem_end, z, bchunk, 0L, 0); 1109 } 1110 1111 #endif 1112 1113 /* 1114 * free (SLAB ALLOCATOR) 1115 * 1116 * Free a memory block previously allocated by malloc. Note that we do not 1117 * attempt to update ks_loosememuse as MP races could prevent us from 1118 * checking memory limits in malloc. 1119 * 1120 * MPSAFE 1121 */ 1122 void 1123 kfree(void *ptr, struct malloc_type *type) 1124 { 1125 SLZone *z; 1126 SLChunk *chunk; 1127 SLGlobalData *slgd; 1128 struct globaldata *gd; 1129 int *kup; 1130 unsigned long size; 1131 #ifdef SMP 1132 SLChunk *bchunk; 1133 int rsignal; 1134 #endif 1135 1136 logmemory_quick(free_beg); 1137 gd = mycpu; 1138 slgd = &gd->gd_slab; 1139 1140 if (ptr == NULL) 1141 panic("trying to free NULL pointer"); 1142 1143 /* 1144 * Handle special 0-byte allocations 1145 */ 1146 if (ptr == ZERO_LENGTH_PTR) { 1147 logmemory(free_zero, ptr, type, -1UL, 0); 1148 logmemory_quick(free_end); 1149 return; 1150 } 1151 1152 /* 1153 * Panic on bad malloc type 1154 */ 1155 if (type->ks_magic != M_MAGIC) 1156 panic("free: malloc type lacks magic"); 1157 1158 /* 1159 * Handle oversized allocations. XXX we really should require that a 1160 * size be passed to free() instead of this nonsense. 1161 * 1162 * This code is never called via an ipi. 1163 */ 1164 kup = btokup(ptr); 1165 if (*kup > 0) { 1166 size = *kup << PAGE_SHIFT; 1167 *kup = 0; 1168 #ifdef INVARIANTS 1169 KKASSERT(sizeof(weirdary) <= size); 1170 bcopy(weirdary, ptr, sizeof(weirdary)); 1171 #endif 1172 /* 1173 * NOTE: For oversized allocations we do not record the 1174 * originating cpu. It gets freed on the cpu calling 1175 * kfree(). The statistics are in aggregate. 1176 * 1177 * note: XXX we have still inherited the interrupts-can't-block 1178 * assumption. An interrupt thread does not bump 1179 * gd_intr_nesting_level so check TDF_INTTHREAD. This is 1180 * primarily until we can fix softupdate's assumptions about free(). 1181 */ 1182 crit_enter(); 1183 --type->ks_inuse[gd->gd_cpuid]; 1184 type->ks_memuse[gd->gd_cpuid] -= size; 1185 if (mycpu->gd_intr_nesting_level || 1186 (gd->gd_curthread->td_flags & TDF_INTTHREAD)) 1187 { 1188 logmemory(free_ovsz_delayed, ptr, type, size, 0); 1189 z = (SLZone *)ptr; 1190 z->z_Magic = ZALLOC_OVSZ_MAGIC; 1191 z->z_Next = slgd->FreeOvZones; 1192 z->z_ChunkSize = size; 1193 slgd->FreeOvZones = z; 1194 crit_exit(); 1195 } else { 1196 crit_exit(); 1197 logmemory(free_ovsz, ptr, type, size, 0); 1198 kmem_slab_free(ptr, size); /* may block */ 1199 atomic_add_int(&ZoneBigAlloc, -(int)size / 1024); 1200 } 1201 logmemory_quick(free_end); 1202 return; 1203 } 1204 1205 /* 1206 * Zone case. Figure out the zone based on the fact that it is 1207 * ZoneSize aligned. 1208 */ 1209 z = (SLZone *)((uintptr_t)ptr & ZoneMask); 1210 kup = btokup(z); 1211 KKASSERT(*kup < 0); 1212 KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); 1213 1214 /* 1215 * If we do not own the zone then use atomic ops to free to the 1216 * remote cpu linked list and notify the target zone using a 1217 * passive message. 1218 * 1219 * The target zone cannot be deallocated while we own a chunk of it, 1220 * so the zone header's storage is stable until the very moment 1221 * we adjust z_RChunks. After that we cannot safely dereference (z). 1222 * 1223 * (no critical section needed) 1224 */ 1225 if (z->z_CpuGd != gd) { 1226 #ifdef SMP 1227 /* 1228 * Making these adjustments now allow us to avoid passing (type) 1229 * to the remote cpu. Note that ks_inuse/ks_memuse is being 1230 * adjusted on OUR cpu, not the zone cpu, but it should all still 1231 * sum up properly and cancel out. 1232 */ 1233 crit_enter(); 1234 --type->ks_inuse[gd->gd_cpuid]; 1235 type->ks_memuse[gd->gd_cpuid] -= z->z_ChunkSize; 1236 crit_exit(); 1237 1238 /* 1239 * WARNING! This code competes with other cpus. Once we 1240 * successfully link the chunk to RChunks the remote 1241 * cpu can rip z's storage out from under us. 1242 * 1243 * Bumping RCount prevents z's storage from getting 1244 * ripped out. 1245 */ 1246 rsignal = z->z_RSignal; 1247 cpu_lfence(); 1248 if (rsignal) 1249 atomic_add_int(&z->z_RCount, 1); 1250 1251 chunk = ptr; 1252 for (;;) { 1253 bchunk = z->z_RChunks; 1254 cpu_ccfence(); 1255 chunk->c_Next = bchunk; 1256 cpu_sfence(); 1257 1258 if (atomic_cmpset_ptr(&z->z_RChunks, bchunk, chunk)) 1259 break; 1260 } 1261 1262 /* 1263 * We have to signal the remote cpu if our actions will cause 1264 * the remote zone to be placed back on ZoneAry so it can 1265 * move the zone back on. 1266 * 1267 * We only need to deal with NULL->non-NULL RChunk transitions 1268 * and only if z_RSignal is set. We interlock by reading rsignal 1269 * before adding our chunk to RChunks. This should result in 1270 * virtually no IPI traffic. 1271 * 1272 * We can use a passive IPI to reduce overhead even further. 1273 */ 1274 if (bchunk == NULL && rsignal) { 1275 logmemory(free_request, ptr, type, (unsigned long)z->z_ChunkSize, 0); 1276 lwkt_send_ipiq_passive(z->z_CpuGd, kfree_remote, z); 1277 /* z can get ripped out from under us from this point on */ 1278 } else if (rsignal) { 1279 atomic_subtract_int(&z->z_RCount, 1); 1280 /* z can get ripped out from under us from this point on */ 1281 } 1282 #else 1283 panic("Corrupt SLZone"); 1284 #endif 1285 logmemory_quick(free_end); 1286 return; 1287 } 1288 1289 /* 1290 * kfree locally 1291 */ 1292 logmemory(free_chunk, ptr, type, (unsigned long)z->z_ChunkSize, 0); 1293 1294 crit_enter(); 1295 chunk = ptr; 1296 chunk_mark_free(z, chunk); 1297 1298 /* 1299 * Put weird data into the memory to detect modifications after freeing, 1300 * illegal pointer use after freeing (we should fault on the odd address), 1301 * and so forth. XXX needs more work, see the old malloc code. 1302 */ 1303 #ifdef INVARIANTS 1304 if (z->z_ChunkSize < sizeof(weirdary)) 1305 bcopy(weirdary, chunk, z->z_ChunkSize); 1306 else 1307 bcopy(weirdary, chunk, sizeof(weirdary)); 1308 #endif 1309 1310 /* 1311 * Add this free non-zero'd chunk to a linked list for reuse. Add 1312 * to the front of the linked list so it is more likely to be 1313 * reallocated, since it is already in our L1 cache. 1314 */ 1315 #ifdef INVARIANTS 1316 if ((vm_offset_t)chunk < KvaStart || (vm_offset_t)chunk >= KvaEnd) 1317 panic("BADFREE %p", chunk); 1318 #endif 1319 chunk->c_Next = z->z_LChunks; 1320 z->z_LChunks = chunk; 1321 if (chunk->c_Next == NULL) 1322 z->z_LChunksp = &chunk->c_Next; 1323 1324 #ifdef INVARIANTS 1325 if (chunk->c_Next && (vm_offset_t)chunk->c_Next < KvaStart) 1326 panic("BADFREE2"); 1327 #endif 1328 1329 /* 1330 * Bump the number of free chunks. If it becomes non-zero the zone 1331 * must be added back onto the appropriate list. 1332 */ 1333 if (z->z_NFree++ == 0) { 1334 z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; 1335 slgd->ZoneAry[z->z_ZoneIndex] = z; 1336 } 1337 1338 --type->ks_inuse[z->z_Cpu]; 1339 type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize; 1340 1341 /* 1342 * If the zone becomes totally free, and there are other zones we 1343 * can allocate from, move this zone to the FreeZones list. Since 1344 * this code can be called from an IPI callback, do *NOT* try to mess 1345 * with kernel_map here. Hysteresis will be performed at malloc() time. 1346 */ 1347 if (z->z_NFree == z->z_NMax && 1348 (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) && 1349 z->z_RCount == 0 1350 ) { 1351 SLZone **pz; 1352 int *kup; 1353 1354 for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next) 1355 ; 1356 *pz = z->z_Next; 1357 z->z_Magic = -1; 1358 z->z_Next = slgd->FreeZones; 1359 slgd->FreeZones = z; 1360 ++slgd->NFreeZones; 1361 kup = btokup(z); 1362 *kup = 0; 1363 } 1364 logmemory_quick(free_end); 1365 crit_exit(); 1366 } 1367 1368 #if defined(INVARIANTS) 1369 1370 /* 1371 * Helper routines for sanity checks 1372 */ 1373 static 1374 void 1375 chunk_mark_allocated(SLZone *z, void *chunk) 1376 { 1377 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 1378 __uint32_t *bitptr; 1379 1380 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0); 1381 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, 1382 ("memory chunk %p bit index %d is illegal", chunk, bitdex)); 1383 bitptr = &z->z_Bitmap[bitdex >> 5]; 1384 bitdex &= 31; 1385 KASSERT((*bitptr & (1 << bitdex)) == 0, 1386 ("memory chunk %p is already allocated!", chunk)); 1387 *bitptr |= 1 << bitdex; 1388 } 1389 1390 static 1391 void 1392 chunk_mark_free(SLZone *z, void *chunk) 1393 { 1394 int bitdex = ((char *)chunk - (char *)z->z_BasePtr) / z->z_ChunkSize; 1395 __uint32_t *bitptr; 1396 1397 KKASSERT((((intptr_t)chunk ^ (intptr_t)z) & ZoneMask) == 0); 1398 KASSERT(bitdex >= 0 && bitdex < z->z_NMax, 1399 ("memory chunk %p bit index %d is illegal!", chunk, bitdex)); 1400 bitptr = &z->z_Bitmap[bitdex >> 5]; 1401 bitdex &= 31; 1402 KASSERT((*bitptr & (1 << bitdex)) != 0, 1403 ("memory chunk %p is already free!", chunk)); 1404 *bitptr &= ~(1 << bitdex); 1405 } 1406 1407 #endif 1408 1409 /* 1410 * kmem_slab_alloc() 1411 * 1412 * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the 1413 * specified alignment. M_* flags are expected in the flags field. 1414 * 1415 * Alignment must be a multiple of PAGE_SIZE. 1416 * 1417 * NOTE! XXX For the moment we use vm_map_entry_reserve/release(), 1418 * but when we move zalloc() over to use this function as its backend 1419 * we will have to switch to kreserve/krelease and call reserve(0) 1420 * after the new space is made available. 1421 * 1422 * Interrupt code which has preempted other code is not allowed to 1423 * use PQ_CACHE pages. However, if an interrupt thread is run 1424 * non-preemptively or blocks and then runs non-preemptively, then 1425 * it is free to use PQ_CACHE pages. <--- may not apply any longer XXX 1426 */ 1427 static void * 1428 kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags) 1429 { 1430 vm_size_t i; 1431 vm_offset_t addr; 1432 int count, vmflags, base_vmflags; 1433 vm_page_t mbase = NULL; 1434 vm_page_t m; 1435 thread_t td; 1436 1437 size = round_page(size); 1438 addr = vm_map_min(&kernel_map); 1439 1440 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1441 crit_enter(); 1442 vm_map_lock(&kernel_map); 1443 if (vm_map_findspace(&kernel_map, addr, size, align, 0, &addr)) { 1444 vm_map_unlock(&kernel_map); 1445 if ((flags & M_NULLOK) == 0) 1446 panic("kmem_slab_alloc(): kernel_map ran out of space!"); 1447 vm_map_entry_release(count); 1448 crit_exit(); 1449 return(NULL); 1450 } 1451 1452 /* 1453 * kernel_object maps 1:1 to kernel_map. 1454 */ 1455 vm_object_hold(&kernel_object); 1456 vm_object_reference_locked(&kernel_object); 1457 vm_map_insert(&kernel_map, &count, 1458 &kernel_object, addr, addr, addr + size, 1459 VM_MAPTYPE_NORMAL, 1460 VM_PROT_ALL, VM_PROT_ALL, 1461 0); 1462 vm_object_drop(&kernel_object); 1463 vm_map_set_wired_quick(&kernel_map, addr, size, &count); 1464 vm_map_unlock(&kernel_map); 1465 1466 td = curthread; 1467 1468 base_vmflags = 0; 1469 if (flags & M_ZERO) 1470 base_vmflags |= VM_ALLOC_ZERO; 1471 if (flags & M_USE_RESERVE) 1472 base_vmflags |= VM_ALLOC_SYSTEM; 1473 if (flags & M_USE_INTERRUPT_RESERVE) 1474 base_vmflags |= VM_ALLOC_INTERRUPT; 1475 if ((flags & (M_RNOWAIT|M_WAITOK)) == 0) { 1476 panic("kmem_slab_alloc: bad flags %08x (%p)", 1477 flags, ((int **)&size)[-1]); 1478 } 1479 1480 /* 1481 * Allocate the pages. Do not mess with the PG_ZERO flag or map 1482 * them yet. VM_ALLOC_NORMAL can only be set if we are not preempting. 1483 * 1484 * VM_ALLOC_SYSTEM is automatically set if we are preempting and 1485 * M_WAITOK was specified as an alternative (i.e. M_USE_RESERVE is 1486 * implied in this case), though I'm not sure if we really need to 1487 * do that. 1488 */ 1489 vmflags = base_vmflags; 1490 if (flags & M_WAITOK) { 1491 if (td->td_preempted) 1492 vmflags |= VM_ALLOC_SYSTEM; 1493 else 1494 vmflags |= VM_ALLOC_NORMAL; 1495 } 1496 1497 vm_object_hold(&kernel_object); 1498 for (i = 0; i < size; i += PAGE_SIZE) { 1499 m = vm_page_alloc(&kernel_object, OFF_TO_IDX(addr + i), vmflags); 1500 if (i == 0) 1501 mbase = m; 1502 1503 /* 1504 * If the allocation failed we either return NULL or we retry. 1505 * 1506 * If M_WAITOK is specified we wait for more memory and retry. 1507 * If M_WAITOK is specified from a preemption we yield instead of 1508 * wait. Livelock will not occur because the interrupt thread 1509 * will not be preempting anyone the second time around after the 1510 * yield. 1511 */ 1512 if (m == NULL) { 1513 if (flags & M_WAITOK) { 1514 if (td->td_preempted) { 1515 lwkt_switch(); 1516 } else { 1517 vm_wait(0); 1518 } 1519 i -= PAGE_SIZE; /* retry */ 1520 continue; 1521 } 1522 break; 1523 } 1524 } 1525 1526 /* 1527 * Check and deal with an allocation failure 1528 */ 1529 if (i != size) { 1530 while (i != 0) { 1531 i -= PAGE_SIZE; 1532 m = vm_page_lookup(&kernel_object, OFF_TO_IDX(addr + i)); 1533 /* page should already be busy */ 1534 vm_page_free(m); 1535 } 1536 vm_map_lock(&kernel_map); 1537 vm_map_delete(&kernel_map, addr, addr + size, &count); 1538 vm_map_unlock(&kernel_map); 1539 vm_object_drop(&kernel_object); 1540 1541 vm_map_entry_release(count); 1542 crit_exit(); 1543 return(NULL); 1544 } 1545 1546 /* 1547 * Success! 1548 * 1549 * NOTE: The VM pages are still busied. mbase points to the first one 1550 * but we have to iterate via vm_page_next() 1551 */ 1552 vm_object_drop(&kernel_object); 1553 crit_exit(); 1554 1555 /* 1556 * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO. 1557 */ 1558 m = mbase; 1559 i = 0; 1560 1561 while (i < size) { 1562 /* 1563 * page should already be busy 1564 */ 1565 m->valid = VM_PAGE_BITS_ALL; 1566 vm_page_wire(m); 1567 pmap_enter(&kernel_pmap, addr + i, m, VM_PROT_ALL | VM_PROT_NOSYNC, 1568 1, NULL); 1569 if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO)) 1570 bzero((char *)addr + i, PAGE_SIZE); 1571 vm_page_flag_clear(m, PG_ZERO); 1572 KKASSERT(m->flags & (PG_WRITEABLE | PG_MAPPED)); 1573 vm_page_flag_set(m, PG_REFERENCED); 1574 vm_page_wakeup(m); 1575 1576 i += PAGE_SIZE; 1577 vm_object_hold(&kernel_object); 1578 m = vm_page_next(m); 1579 vm_object_drop(&kernel_object); 1580 } 1581 smp_invltlb(); 1582 vm_map_entry_release(count); 1583 return((void *)addr); 1584 } 1585 1586 /* 1587 * kmem_slab_free() 1588 */ 1589 static void 1590 kmem_slab_free(void *ptr, vm_size_t size) 1591 { 1592 crit_enter(); 1593 vm_map_remove(&kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size); 1594 crit_exit(); 1595 } 1596 1597 void * 1598 kmalloc_cachealign(unsigned long size_alloc, struct malloc_type *type, 1599 int flags) 1600 { 1601 #if (__VM_CACHELINE_SIZE == 32) 1602 #define CAN_CACHEALIGN(sz) ((sz) >= 256) 1603 #elif (__VM_CACHELINE_SIZE == 64) 1604 #define CAN_CACHEALIGN(sz) ((sz) >= 512) 1605 #elif (__VM_CACHELINE_SIZE == 128) 1606 #define CAN_CACHEALIGN(sz) ((sz) >= 1024) 1607 #else 1608 #error "unsupported cacheline size" 1609 #endif 1610 1611 void *ret; 1612 1613 if (size_alloc < __VM_CACHELINE_SIZE) 1614 size_alloc = __VM_CACHELINE_SIZE; 1615 else if (!CAN_CACHEALIGN(size_alloc)) 1616 flags |= M_POWEROF2; 1617 1618 ret = kmalloc(size_alloc, type, flags); 1619 KASSERT(((uintptr_t)ret & (__VM_CACHELINE_SIZE - 1)) == 0, 1620 ("%p(%lu) not cacheline %d aligned", 1621 ret, size_alloc, __VM_CACHELINE_SIZE)); 1622 return ret; 1623 1624 #undef CAN_CACHEALIGN 1625 } 1626