| 1 | /* |
| 2 | * KERN_SLABALLOC.C - Kernel SLAB memory allocator |
| 3 | * |
| 4 | * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com> |
| 5 | * All rights reserved. |
| 6 | * |
| 7 | * Redistribution and use in source and binary forms, with or without |
| 8 | * modification, are permitted provided that the following conditions |
| 9 | * are met: |
| 10 | * 1. Redistributions of source code must retain the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer. |
| 12 | * 2. Redistributions in binary form must reproduce the above copyright |
| 13 | * notice, this list of conditions and the following disclaimer in the |
| 14 | * documentation and/or other materials provided with the distribution. |
| 15 | * |
| 16 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND |
| 17 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 18 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 19 | * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE |
| 20 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 21 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 22 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 23 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 24 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 25 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 26 | * SUCH DAMAGE. |
| 27 | * |
| 28 | * $DragonFly: src/sys/kern/kern_slaballoc.c,v 1.13 2003/10/20 16:09:00 dillon Exp $ |
| 29 | * |
| 30 | * This module implements a slab allocator drop-in replacement for the |
| 31 | * kernel malloc(). |
| 32 | * |
| 33 | * A slab allocator reserves a ZONE for each chunk size, then lays the |
| 34 | * chunks out in an array within the zone. Allocation and deallocation |
| 35 | * is nearly instantanious, and fragmentation/overhead losses are limited |
| 36 | * to a fixed worst-case amount. |
| 37 | * |
| 38 | * The downside of this slab implementation is in the chunk size |
| 39 | * multiplied by the number of zones. ~80 zones * 128K = 10MB of VM per cpu. |
| 40 | * In a kernel implementation all this memory will be physical so |
| 41 | * the zone size is adjusted downward on machines with less physical |
| 42 | * memory. The upside is that overhead is bounded... this is the *worst* |
| 43 | * case overhead. |
| 44 | * |
| 45 | * Slab management is done on a per-cpu basis and no locking or mutexes |
| 46 | * are required, only a critical section. When one cpu frees memory |
| 47 | * belonging to another cpu's slab manager an asynchronous IPI message |
| 48 | * will be queued to execute the operation. In addition, both the |
| 49 | * high level slab allocator and the low level zone allocator optimize |
| 50 | * M_ZERO requests, and the slab allocator does not have to pre initialize |
| 51 | * the linked list of chunks. |
| 52 | * |
| 53 | * XXX Balancing is needed between cpus. Balance will be handled through |
| 54 | * asynchronous IPIs primarily by reassigning the z_Cpu ownership of chunks. |
| 55 | * |
| 56 | * XXX If we have to allocate a new zone and M_USE_RESERVE is set, use of |
| 57 | * the new zone should be restricted to M_USE_RESERVE requests only. |
| 58 | * |
| 59 | * Alloc Size Chunking Number of zones |
| 60 | * 0-127 8 16 |
| 61 | * 128-255 16 8 |
| 62 | * 256-511 32 8 |
| 63 | * 512-1023 64 8 |
| 64 | * 1024-2047 128 8 |
| 65 | * 2048-4095 256 8 |
| 66 | * 4096-8191 512 8 |
| 67 | * 8192-16383 1024 8 |
| 68 | * 16384-32767 2048 8 |
| 69 | * (if PAGE_SIZE is 4K the maximum zone allocation is 16383) |
| 70 | * |
| 71 | * Allocations >= ZoneLimit go directly to kmem. |
| 72 | * |
| 73 | * API REQUIREMENTS AND SIDE EFFECTS |
| 74 | * |
| 75 | * To operate as a drop-in replacement to the FreeBSD-4.x malloc() we |
| 76 | * have remained compatible with the following API requirements: |
| 77 | * |
| 78 | * + small power-of-2 sized allocations are power-of-2 aligned (kern_tty) |
| 79 | * + all power-of-2 sized allocations are power-of-2 aligned (twe) |
| 80 | * + malloc(0) is allowed and returns non-NULL (ahc driver) |
| 81 | * + ability to allocate arbitrarily large chunks of memory |
| 82 | */ |
| 83 | |
| 84 | #include "opt_vm.h" |
| 85 | |
| 86 | #include <sys/param.h> |
| 87 | #include <sys/systm.h> |
| 88 | #include <sys/kernel.h> |
| 89 | #include <sys/slaballoc.h> |
| 90 | #include <sys/mbuf.h> |
| 91 | #include <sys/vmmeter.h> |
| 92 | #include <sys/lock.h> |
| 93 | #include <sys/thread.h> |
| 94 | #include <sys/globaldata.h> |
| 95 | |
| 96 | #include <vm/vm.h> |
| 97 | #include <vm/vm_param.h> |
| 98 | #include <vm/vm_kern.h> |
| 99 | #include <vm/vm_extern.h> |
| 100 | #include <vm/vm_object.h> |
| 101 | #include <vm/pmap.h> |
| 102 | #include <vm/vm_map.h> |
| 103 | #include <vm/vm_page.h> |
| 104 | #include <vm/vm_pageout.h> |
| 105 | |
| 106 | #include <machine/cpu.h> |
| 107 | |
| 108 | #include <sys/thread2.h> |
| 109 | |
| 110 | #define arysize(ary) (sizeof(ary)/sizeof((ary)[0])) |
| 111 | |
| 112 | /* |
| 113 | * Fixed globals (not per-cpu) |
| 114 | */ |
| 115 | static int ZoneSize; |
| 116 | static int ZoneLimit; |
| 117 | static int ZonePageCount; |
| 118 | static int ZonePageLimit; |
| 119 | static int ZoneMask; |
| 120 | static struct malloc_type *kmemstatistics; |
| 121 | static struct kmemusage *kmemusage; |
| 122 | static int32_t weirdary[16]; |
| 123 | |
| 124 | static void *kmem_slab_alloc(vm_size_t bytes, vm_offset_t align, int flags); |
| 125 | static void kmem_slab_free(void *ptr, vm_size_t bytes); |
| 126 | |
| 127 | /* |
| 128 | * Misc constants. Note that allocations that are exact multiples of |
| 129 | * PAGE_SIZE, or exceed the zone limit, fall through to the kmem module. |
| 130 | * IN_SAME_PAGE_MASK is used to sanity-check the per-page free lists. |
| 131 | */ |
| 132 | #define MIN_CHUNK_SIZE 8 /* in bytes */ |
| 133 | #define MIN_CHUNK_MASK (MIN_CHUNK_SIZE - 1) |
| 134 | #define ZONE_RELS_THRESH 2 /* threshold number of zones */ |
| 135 | #define IN_SAME_PAGE_MASK (~(intptr_t)PAGE_MASK | MIN_CHUNK_MASK) |
| 136 | |
| 137 | /* |
| 138 | * The WEIRD_ADDR is used as known text to copy into free objects to |
| 139 | * try to create deterministic failure cases if the data is accessed after |
| 140 | * free. |
| 141 | */ |
| 142 | #define WEIRD_ADDR 0xdeadc0de |
| 143 | #define MAX_COPY sizeof(weirdary) |
| 144 | #define ZERO_LENGTH_PTR ((void *)-8) |
| 145 | |
| 146 | /* |
| 147 | * Misc global malloc buckets |
| 148 | */ |
| 149 | |
| 150 | MALLOC_DEFINE(M_CACHE, "cache", "Various Dynamically allocated caches"); |
| 151 | MALLOC_DEFINE(M_DEVBUF, "devbuf", "device driver memory"); |
| 152 | MALLOC_DEFINE(M_TEMP, "temp", "misc temporary data buffers"); |
| 153 | |
| 154 | MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options"); |
| 155 | MALLOC_DEFINE(M_IP6NDP, "ip6ndp", "IPv6 Neighbor Discovery"); |
| 156 | |
| 157 | /* |
| 158 | * Initialize the slab memory allocator. We have to choose a zone size based |
| 159 | * on available physical memory. We choose a zone side which is approximately |
| 160 | * 1/1024th of our memory, so if we have 128MB of ram we have a zone size of |
| 161 | * 128K. The zone size is limited to the bounds set in slaballoc.h |
| 162 | * (typically 32K min, 128K max). |
| 163 | */ |
| 164 | static void kmeminit(void *dummy); |
| 165 | |
| 166 | SYSINIT(kmem, SI_SUB_KMEM, SI_ORDER_FIRST, kmeminit, NULL) |
| 167 | |
| 168 | static void |
| 169 | kmeminit(void *dummy) |
| 170 | { |
| 171 | vm_poff_t limsize; |
| 172 | int usesize; |
| 173 | int i; |
| 174 | vm_pindex_t npg; |
| 175 | |
| 176 | limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; |
| 177 | if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) |
| 178 | limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; |
| 179 | |
| 180 | usesize = (int)(limsize / 1024); /* convert to KB */ |
| 181 | |
| 182 | ZoneSize = ZALLOC_MIN_ZONE_SIZE; |
| 183 | while (ZoneSize < ZALLOC_MAX_ZONE_SIZE && (ZoneSize << 1) < usesize) |
| 184 | ZoneSize <<= 1; |
| 185 | ZoneLimit = ZoneSize / 4; |
| 186 | if (ZoneLimit > ZALLOC_ZONE_LIMIT) |
| 187 | ZoneLimit = ZALLOC_ZONE_LIMIT; |
| 188 | ZoneMask = ZoneSize - 1; |
| 189 | ZonePageLimit = PAGE_SIZE * 4; |
| 190 | ZonePageCount = ZoneSize / PAGE_SIZE; |
| 191 | |
| 192 | npg = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE; |
| 193 | kmemusage = kmem_slab_alloc(npg * sizeof(struct kmemusage), PAGE_SIZE, M_ZERO); |
| 194 | |
| 195 | for (i = 0; i < arysize(weirdary); ++i) |
| 196 | weirdary[i] = WEIRD_ADDR; |
| 197 | |
| 198 | if (bootverbose) |
| 199 | printf("Slab ZoneSize set to %dKB\n", ZoneSize / 1024); |
| 200 | } |
| 201 | |
| 202 | /* |
| 203 | * Initialize a malloc type tracking structure. |
| 204 | */ |
| 205 | void |
| 206 | malloc_init(void *data) |
| 207 | { |
| 208 | struct malloc_type *type = data; |
| 209 | vm_poff_t limsize; |
| 210 | |
| 211 | if (type->ks_magic != M_MAGIC) |
| 212 | panic("malloc type lacks magic"); |
| 213 | |
| 214 | if (type->ks_limit != 0) |
| 215 | return; |
| 216 | |
| 217 | if (vmstats.v_page_count == 0) |
| 218 | panic("malloc_init not allowed before vm init"); |
| 219 | |
| 220 | limsize = (vm_poff_t)vmstats.v_page_count * PAGE_SIZE; |
| 221 | if (limsize > VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) |
| 222 | limsize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; |
| 223 | type->ks_limit = limsize / 10; |
| 224 | |
| 225 | type->ks_next = kmemstatistics; |
| 226 | kmemstatistics = type; |
| 227 | } |
| 228 | |
| 229 | void |
| 230 | malloc_uninit(void *data) |
| 231 | { |
| 232 | struct malloc_type *type = data; |
| 233 | struct malloc_type *t; |
| 234 | #ifdef INVARIANTS |
| 235 | int i; |
| 236 | long ttl; |
| 237 | #endif |
| 238 | |
| 239 | if (type->ks_magic != M_MAGIC) |
| 240 | panic("malloc type lacks magic"); |
| 241 | |
| 242 | if (vmstats.v_page_count == 0) |
| 243 | panic("malloc_uninit not allowed before vm init"); |
| 244 | |
| 245 | if (type->ks_limit == 0) |
| 246 | panic("malloc_uninit on uninitialized type"); |
| 247 | |
| 248 | #ifdef INVARIANTS |
| 249 | /* |
| 250 | * memuse is only correct in aggregation. Due to memory being allocated |
| 251 | * on one cpu and freed on another individual array entries may be |
| 252 | * negative or positive (canceling each other out). |
| 253 | */ |
| 254 | for (i = ttl = 0; i < ncpus; ++i) |
| 255 | ttl += type->ks_memuse[i]; |
| 256 | if (ttl) { |
| 257 | printf("malloc_uninit: %ld bytes of '%s' still allocated on cpu %d\n", |
| 258 | ttl, type->ks_shortdesc, i); |
| 259 | } |
| 260 | #endif |
| 261 | if (type == kmemstatistics) { |
| 262 | kmemstatistics = type->ks_next; |
| 263 | } else { |
| 264 | for (t = kmemstatistics; t->ks_next != NULL; t = t->ks_next) { |
| 265 | if (t->ks_next == type) { |
| 266 | t->ks_next = type->ks_next; |
| 267 | break; |
| 268 | } |
| 269 | } |
| 270 | } |
| 271 | type->ks_next = NULL; |
| 272 | type->ks_limit = 0; |
| 273 | } |
| 274 | |
| 275 | /* |
| 276 | * Calculate the zone index for the allocation request size and set the |
| 277 | * allocation request size to that particular zone's chunk size. |
| 278 | */ |
| 279 | static __inline int |
| 280 | zoneindex(unsigned long *bytes) |
| 281 | { |
| 282 | unsigned int n = (unsigned int)*bytes; /* unsigned for shift opt */ |
| 283 | if (n < 128) { |
| 284 | *bytes = n = (n + 7) & ~7; |
| 285 | return(n / 8 - 1); /* 8 byte chunks, 16 zones */ |
| 286 | } |
| 287 | if (n < 256) { |
| 288 | *bytes = n = (n + 15) & ~15; |
| 289 | return(n / 16 + 7); |
| 290 | } |
| 291 | if (n < 8192) { |
| 292 | if (n < 512) { |
| 293 | *bytes = n = (n + 31) & ~31; |
| 294 | return(n / 32 + 15); |
| 295 | } |
| 296 | if (n < 1024) { |
| 297 | *bytes = n = (n + 63) & ~63; |
| 298 | return(n / 64 + 23); |
| 299 | } |
| 300 | if (n < 2048) { |
| 301 | *bytes = n = (n + 127) & ~127; |
| 302 | return(n / 128 + 31); |
| 303 | } |
| 304 | if (n < 4096) { |
| 305 | *bytes = n = (n + 255) & ~255; |
| 306 | return(n / 256 + 39); |
| 307 | } |
| 308 | *bytes = n = (n + 511) & ~511; |
| 309 | return(n / 512 + 47); |
| 310 | } |
| 311 | #if ZALLOC_ZONE_LIMIT > 8192 |
| 312 | if (n < 16384) { |
| 313 | *bytes = n = (n + 1023) & ~1023; |
| 314 | return(n / 1024 + 55); |
| 315 | } |
| 316 | #endif |
| 317 | #if ZALLOC_ZONE_LIMIT > 16384 |
| 318 | if (n < 32768) { |
| 319 | *bytes = n = (n + 2047) & ~2047; |
| 320 | return(n / 2048 + 63); |
| 321 | } |
| 322 | #endif |
| 323 | panic("Unexpected byte count %d", n); |
| 324 | return(0); |
| 325 | } |
| 326 | |
| 327 | /* |
| 328 | * malloc() (SLAB ALLOCATOR) |
| 329 | * |
| 330 | * Allocate memory via the slab allocator. If the request is too large, |
| 331 | * or if it page-aligned beyond a certain size, we fall back to the |
| 332 | * KMEM subsystem. A SLAB tracking descriptor must be specified, use |
| 333 | * &SlabMisc if you don't care. |
| 334 | * |
| 335 | * M_NOWAIT - return NULL instead of blocking. |
| 336 | * M_ZERO - zero the returned memory. |
| 337 | * M_USE_RESERVE - allocate out of the system reserve if necessary |
| 338 | */ |
| 339 | void * |
| 340 | malloc(unsigned long size, struct malloc_type *type, int flags) |
| 341 | { |
| 342 | SLZone *z; |
| 343 | SLChunk *chunk; |
| 344 | SLGlobalData *slgd; |
| 345 | struct globaldata *gd; |
| 346 | int zi; |
| 347 | |
| 348 | gd = mycpu; |
| 349 | slgd = &gd->gd_slab; |
| 350 | |
| 351 | /* |
| 352 | * XXX silly to have this in the critical path. |
| 353 | */ |
| 354 | if (type->ks_limit == 0) { |
| 355 | crit_enter(); |
| 356 | if (type->ks_limit == 0) |
| 357 | malloc_init(type); |
| 358 | crit_exit(); |
| 359 | } |
| 360 | ++type->ks_calls; |
| 361 | |
| 362 | /* |
| 363 | * Handle the case where the limit is reached. Panic if can't return |
| 364 | * NULL. XXX the original malloc code looped, but this tended to |
| 365 | * simply deadlock the computer. |
| 366 | */ |
| 367 | while (type->ks_loosememuse >= type->ks_limit) { |
| 368 | int i; |
| 369 | long ttl; |
| 370 | |
| 371 | for (i = ttl = 0; i < ncpus; ++i) |
| 372 | ttl += type->ks_memuse[i]; |
| 373 | type->ks_loosememuse = ttl; |
| 374 | if (ttl >= type->ks_limit) { |
| 375 | if (flags & (M_NOWAIT|M_NULLOK)) |
| 376 | return(NULL); |
| 377 | panic("%s: malloc limit exceeded", type->ks_shortdesc); |
| 378 | } |
| 379 | } |
| 380 | |
| 381 | /* |
| 382 | * Handle the degenerate size == 0 case. Yes, this does happen. |
| 383 | * Return a special pointer. This is to maintain compatibility with |
| 384 | * the original malloc implementation. Certain devices, such as the |
| 385 | * adaptec driver, not only allocate 0 bytes, they check for NULL and |
| 386 | * also realloc() later on. Joy. |
| 387 | */ |
| 388 | if (size == 0) |
| 389 | return(ZERO_LENGTH_PTR); |
| 390 | |
| 391 | /* |
| 392 | * Handle hysteresis from prior frees here in malloc(). We cannot |
| 393 | * safely manipulate the kernel_map in free() due to free() possibly |
| 394 | * being called via an IPI message or from sensitive interrupt code. |
| 395 | */ |
| 396 | while (slgd->NFreeZones > ZONE_RELS_THRESH && (flags & M_NOWAIT) == 0) { |
| 397 | crit_enter(); |
| 398 | if (slgd->NFreeZones > ZONE_RELS_THRESH) { /* crit sect race */ |
| 399 | z = slgd->FreeZones; |
| 400 | slgd->FreeZones = z->z_Next; |
| 401 | --slgd->NFreeZones; |
| 402 | kmem_slab_free(z, ZoneSize); /* may block */ |
| 403 | } |
| 404 | crit_exit(); |
| 405 | } |
| 406 | /* |
| 407 | * XXX handle oversized frees that were queued from free(). |
| 408 | */ |
| 409 | while (slgd->FreeOvZones && (flags & M_NOWAIT) == 0) { |
| 410 | crit_enter(); |
| 411 | if ((z = slgd->FreeOvZones) != NULL) { |
| 412 | KKASSERT(z->z_Magic == ZALLOC_OVSZ_MAGIC); |
| 413 | slgd->FreeOvZones = z->z_Next; |
| 414 | kmem_slab_free(z, z->z_ChunkSize); /* may block */ |
| 415 | } |
| 416 | crit_exit(); |
| 417 | } |
| 418 | |
| 419 | /* |
| 420 | * Handle large allocations directly. There should not be very many of |
| 421 | * these so performance is not a big issue. |
| 422 | * |
| 423 | * Guarentee page alignment for allocations in multiples of PAGE_SIZE |
| 424 | */ |
| 425 | if (size >= ZoneLimit || (size & PAGE_MASK) == 0) { |
| 426 | struct kmemusage *kup; |
| 427 | |
| 428 | size = round_page(size); |
| 429 | chunk = kmem_slab_alloc(size, PAGE_SIZE, flags); |
| 430 | if (chunk == NULL) |
| 431 | return(NULL); |
| 432 | flags &= ~M_ZERO; /* result already zero'd if M_ZERO was set */ |
| 433 | flags |= M_PASSIVE_ZERO; |
| 434 | kup = btokup(chunk); |
| 435 | kup->ku_pagecnt = size / PAGE_SIZE; |
| 436 | kup->ku_cpu = gd->gd_cpuid; |
| 437 | crit_enter(); |
| 438 | goto done; |
| 439 | } |
| 440 | |
| 441 | /* |
| 442 | * Attempt to allocate out of an existing zone. First try the free list, |
| 443 | * then allocate out of unallocated space. If we find a good zone move |
| 444 | * it to the head of the list so later allocations find it quickly |
| 445 | * (we might have thousands of zones in the list). |
| 446 | * |
| 447 | * Note: zoneindex() will panic of size is too large. |
| 448 | */ |
| 449 | zi = zoneindex(&size); |
| 450 | KKASSERT(zi < NZONES); |
| 451 | crit_enter(); |
| 452 | if ((z = slgd->ZoneAry[zi]) != NULL) { |
| 453 | KKASSERT(z->z_NFree > 0); |
| 454 | |
| 455 | /* |
| 456 | * Remove us from the ZoneAry[] when we become empty |
| 457 | */ |
| 458 | if (--z->z_NFree == 0) { |
| 459 | slgd->ZoneAry[zi] = z->z_Next; |
| 460 | z->z_Next = NULL; |
| 461 | } |
| 462 | |
| 463 | /* |
| 464 | * Locate a chunk in a free page. This attempts to localize |
| 465 | * reallocations into earlier pages without us having to sort |
| 466 | * the chunk list. A chunk may still overlap a page boundary. |
| 467 | */ |
| 468 | while (z->z_FirstFreePg < ZonePageCount) { |
| 469 | if ((chunk = z->z_PageAry[z->z_FirstFreePg]) != NULL) { |
| 470 | #ifdef DIAGNOSTIC |
| 471 | /* |
| 472 | * Diagnostic: c_Next is not total garbage. |
| 473 | */ |
| 474 | KKASSERT(chunk->c_Next == NULL || |
| 475 | ((intptr_t)chunk->c_Next & IN_SAME_PAGE_MASK) == |
| 476 | ((intptr_t)chunk & IN_SAME_PAGE_MASK)); |
| 477 | #endif |
| 478 | #ifdef INVARIANTS |
| 479 | if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) |
| 480 | panic("chunk %p FFPG %d/%d", chunk, z->z_FirstFreePg, ZonePageCount); |
| 481 | if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) |
| 482 | panic("chunkNEXT %p %p FFPG %d/%d", chunk, chunk->c_Next, z->z_FirstFreePg, ZonePageCount); |
| 483 | #endif |
| 484 | z->z_PageAry[z->z_FirstFreePg] = chunk->c_Next; |
| 485 | goto done; |
| 486 | } |
| 487 | ++z->z_FirstFreePg; |
| 488 | } |
| 489 | |
| 490 | /* |
| 491 | * No chunks are available but NFree said we had some memory, so |
| 492 | * it must be available in the never-before-used-memory area |
| 493 | * governed by UIndex. The consequences are very serious if our zone |
| 494 | * got corrupted so we use an explicit panic rather then a KASSERT. |
| 495 | */ |
| 496 | if (z->z_UIndex + 1 != z->z_NMax) |
| 497 | z->z_UIndex = z->z_UIndex + 1; |
| 498 | else |
| 499 | z->z_UIndex = 0; |
| 500 | if (z->z_UIndex == z->z_UEndIndex) |
| 501 | panic("slaballoc: corrupted zone"); |
| 502 | chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); |
| 503 | if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { |
| 504 | flags &= ~M_ZERO; |
| 505 | flags |= M_PASSIVE_ZERO; |
| 506 | } |
| 507 | goto done; |
| 508 | } |
| 509 | |
| 510 | /* |
| 511 | * If all zones are exhausted we need to allocate a new zone for this |
| 512 | * index. Use M_ZERO to take advantage of pre-zerod pages. Also see |
| 513 | * UAlloc use above in regards to M_ZERO. Note that when we are reusing |
| 514 | * a zone from the FreeZones list UAlloc'd data will not be zero'd, and |
| 515 | * we do not pre-zero it because we do not want to mess up the L1 cache. |
| 516 | * |
| 517 | * At least one subsystem, the tty code (see CROUND) expects power-of-2 |
| 518 | * allocations to be power-of-2 aligned. We maintain compatibility by |
| 519 | * adjusting the base offset below. |
| 520 | */ |
| 521 | { |
| 522 | int off; |
| 523 | |
| 524 | if ((z = slgd->FreeZones) != NULL) { |
| 525 | slgd->FreeZones = z->z_Next; |
| 526 | --slgd->NFreeZones; |
| 527 | bzero(z, sizeof(SLZone)); |
| 528 | z->z_Flags |= SLZF_UNOTZEROD; |
| 529 | } else { |
| 530 | z = kmem_slab_alloc(ZoneSize, ZoneSize, flags|M_ZERO); |
| 531 | if (z == NULL) |
| 532 | goto fail; |
| 533 | } |
| 534 | |
| 535 | /* |
| 536 | * Guarentee power-of-2 alignment for power-of-2-sized chunks. |
| 537 | * Otherwise just 8-byte align the data. |
| 538 | */ |
| 539 | if ((size | (size - 1)) + 1 == (size << 1)) |
| 540 | off = (sizeof(SLZone) + size - 1) & ~(size - 1); |
| 541 | else |
| 542 | off = (sizeof(SLZone) + MIN_CHUNK_MASK) & ~MIN_CHUNK_MASK; |
| 543 | z->z_Magic = ZALLOC_SLAB_MAGIC; |
| 544 | z->z_ZoneIndex = zi; |
| 545 | z->z_NMax = (ZoneSize - off) / size; |
| 546 | z->z_NFree = z->z_NMax - 1; |
| 547 | z->z_BasePtr = (char *)z + off; |
| 548 | z->z_UIndex = z->z_UEndIndex = slgd->JunkIndex % z->z_NMax; |
| 549 | z->z_ChunkSize = size; |
| 550 | z->z_FirstFreePg = ZonePageCount; |
| 551 | z->z_Cpu = gd->gd_cpuid; |
| 552 | chunk = (SLChunk *)(z->z_BasePtr + z->z_UIndex * size); |
| 553 | z->z_Next = slgd->ZoneAry[zi]; |
| 554 | slgd->ZoneAry[zi] = z; |
| 555 | if ((z->z_Flags & SLZF_UNOTZEROD) == 0) { |
| 556 | flags &= ~M_ZERO; /* already zero'd */ |
| 557 | flags |= M_PASSIVE_ZERO; |
| 558 | } |
| 559 | |
| 560 | /* |
| 561 | * Slide the base index for initial allocations out of the next |
| 562 | * zone we create so we do not over-weight the lower part of the |
| 563 | * cpu memory caches. |
| 564 | */ |
| 565 | slgd->JunkIndex = (slgd->JunkIndex + ZALLOC_SLAB_SLIDE) |
| 566 | & (ZALLOC_MAX_ZONE_SIZE - 1); |
| 567 | } |
| 568 | done: |
| 569 | ++type->ks_inuse[gd->gd_cpuid]; |
| 570 | type->ks_memuse[gd->gd_cpuid] += size; |
| 571 | type->ks_loosememuse += size; |
| 572 | crit_exit(); |
| 573 | if (flags & M_ZERO) |
| 574 | bzero(chunk, size); |
| 575 | #ifdef INVARIANTS |
| 576 | else if ((flags & (M_ZERO|M_PASSIVE_ZERO)) == 0) |
| 577 | chunk->c_Next = (void *)-1; /* avoid accidental double-free check */ |
| 578 | #endif |
| 579 | return(chunk); |
| 580 | fail: |
| 581 | crit_exit(); |
| 582 | return(NULL); |
| 583 | } |
| 584 | |
| 585 | void * |
| 586 | realloc(void *ptr, unsigned long size, struct malloc_type *type, int flags) |
| 587 | { |
| 588 | SLZone *z; |
| 589 | void *nptr; |
| 590 | unsigned long osize; |
| 591 | |
| 592 | if (ptr == NULL || ptr == ZERO_LENGTH_PTR) |
| 593 | return(malloc(size, type, flags)); |
| 594 | if (size == 0) { |
| 595 | free(ptr, type); |
| 596 | return(NULL); |
| 597 | } |
| 598 | |
| 599 | /* |
| 600 | * Handle oversized allocations. XXX we really should require that a |
| 601 | * size be passed to free() instead of this nonsense. |
| 602 | */ |
| 603 | { |
| 604 | struct kmemusage *kup; |
| 605 | |
| 606 | kup = btokup(ptr); |
| 607 | if (kup->ku_pagecnt) { |
| 608 | osize = kup->ku_pagecnt << PAGE_SHIFT; |
| 609 | if (osize == round_page(size)) |
| 610 | return(ptr); |
| 611 | if ((nptr = malloc(size, type, flags)) == NULL) |
| 612 | return(NULL); |
| 613 | bcopy(ptr, nptr, min(size, osize)); |
| 614 | free(ptr, type); |
| 615 | return(nptr); |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | /* |
| 620 | * Get the original allocation's zone. If the new request winds up |
| 621 | * using the same chunk size we do not have to do anything. |
| 622 | */ |
| 623 | z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); |
| 624 | KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); |
| 625 | |
| 626 | zoneindex(&size); |
| 627 | if (z->z_ChunkSize == size) |
| 628 | return(ptr); |
| 629 | |
| 630 | /* |
| 631 | * Allocate memory for the new request size. Note that zoneindex has |
| 632 | * already adjusted the request size to the appropriate chunk size, which |
| 633 | * should optimize our bcopy(). Then copy and return the new pointer. |
| 634 | */ |
| 635 | if ((nptr = malloc(size, type, flags)) == NULL) |
| 636 | return(NULL); |
| 637 | bcopy(ptr, nptr, min(size, z->z_ChunkSize)); |
| 638 | free(ptr, type); |
| 639 | return(nptr); |
| 640 | } |
| 641 | |
| 642 | #ifdef SMP |
| 643 | /* |
| 644 | * free() (SLAB ALLOCATOR) |
| 645 | * |
| 646 | * Free the specified chunk of memory. |
| 647 | */ |
| 648 | static |
| 649 | void |
| 650 | free_remote(void *ptr) |
| 651 | { |
| 652 | free(ptr, *(struct malloc_type **)ptr); |
| 653 | } |
| 654 | |
| 655 | #endif |
| 656 | |
| 657 | void |
| 658 | free(void *ptr, struct malloc_type *type) |
| 659 | { |
| 660 | SLZone *z; |
| 661 | SLChunk *chunk; |
| 662 | SLGlobalData *slgd; |
| 663 | struct globaldata *gd; |
| 664 | int pgno; |
| 665 | |
| 666 | gd = mycpu; |
| 667 | slgd = &gd->gd_slab; |
| 668 | |
| 669 | /* |
| 670 | * Handle special 0-byte allocations |
| 671 | */ |
| 672 | if (ptr == ZERO_LENGTH_PTR) |
| 673 | return; |
| 674 | |
| 675 | /* |
| 676 | * Handle oversized allocations. XXX we really should require that a |
| 677 | * size be passed to free() instead of this nonsense. |
| 678 | * |
| 679 | * This code is never called via an ipi. |
| 680 | */ |
| 681 | { |
| 682 | struct kmemusage *kup; |
| 683 | unsigned long size; |
| 684 | |
| 685 | kup = btokup(ptr); |
| 686 | if (kup->ku_pagecnt) { |
| 687 | size = kup->ku_pagecnt << PAGE_SHIFT; |
| 688 | kup->ku_pagecnt = 0; |
| 689 | #ifdef INVARIANTS |
| 690 | KKASSERT(sizeof(weirdary) <= size); |
| 691 | bcopy(weirdary, ptr, sizeof(weirdary)); |
| 692 | #endif |
| 693 | /* |
| 694 | * note: we always adjust our cpu's slot, not the originating |
| 695 | * cpu (kup->ku_cpuid). The statistics are in aggregate. |
| 696 | */ |
| 697 | crit_enter(); |
| 698 | --type->ks_inuse[gd->gd_cpuid]; |
| 699 | type->ks_memuse[gd->gd_cpuid] -= size; |
| 700 | if (mycpu->gd_intr_nesting_level) { |
| 701 | z = (SLZone *)ptr; |
| 702 | z->z_Magic = ZALLOC_OVSZ_MAGIC; |
| 703 | z->z_Next = slgd->FreeOvZones; |
| 704 | z->z_ChunkSize = size; |
| 705 | slgd->FreeOvZones = z; |
| 706 | crit_exit(); |
| 707 | } else { |
| 708 | crit_exit(); |
| 709 | kmem_slab_free(ptr, size); /* may block */ |
| 710 | } |
| 711 | return; |
| 712 | } |
| 713 | } |
| 714 | |
| 715 | /* |
| 716 | * Zone case. Figure out the zone based on the fact that it is |
| 717 | * ZoneSize aligned. |
| 718 | */ |
| 719 | z = (SLZone *)((uintptr_t)ptr & ~(uintptr_t)ZoneMask); |
| 720 | KKASSERT(z->z_Magic == ZALLOC_SLAB_MAGIC); |
| 721 | |
| 722 | /* |
| 723 | * If we do not own the zone then forward the request to the |
| 724 | * cpu that does. The freeing code does not need the byte count |
| 725 | * unless DIAGNOSTIC is set. |
| 726 | */ |
| 727 | if (z->z_Cpu != gd->gd_cpuid) { |
| 728 | *(struct malloc_type **)ptr = type; |
| 729 | #ifdef SMP |
| 730 | lwkt_send_ipiq(z->z_Cpu, free_remote, ptr); |
| 731 | #else |
| 732 | panic("Corrupt SLZone"); |
| 733 | #endif |
| 734 | return; |
| 735 | } |
| 736 | |
| 737 | if (type->ks_magic != M_MAGIC) |
| 738 | panic("free: malloc type lacks magic"); |
| 739 | |
| 740 | crit_enter(); |
| 741 | pgno = ((char *)ptr - (char *)z) >> PAGE_SHIFT; |
| 742 | chunk = ptr; |
| 743 | |
| 744 | #ifdef INVARIANTS |
| 745 | /* |
| 746 | * Attempt to detect a double-free. To reduce overhead we only check |
| 747 | * if there appears to be link pointer at the base of the data. |
| 748 | */ |
| 749 | if (((intptr_t)chunk->c_Next - (intptr_t)z) >> PAGE_SHIFT == pgno) { |
| 750 | SLChunk *scan; |
| 751 | for (scan = z->z_PageAry[pgno]; scan; scan = scan->c_Next) { |
| 752 | if (scan == chunk) |
| 753 | panic("Double free at %p", chunk); |
| 754 | } |
| 755 | } |
| 756 | #endif |
| 757 | |
| 758 | /* |
| 759 | * Put weird data into the memory to detect modifications after freeing, |
| 760 | * illegal pointer use after freeing (we should fault on the odd address), |
| 761 | * and so forth. XXX needs more work, see the old malloc code. |
| 762 | */ |
| 763 | #ifdef INVARIANTS |
| 764 | if (z->z_ChunkSize < sizeof(weirdary)) |
| 765 | bcopy(weirdary, chunk, z->z_ChunkSize); |
| 766 | else |
| 767 | bcopy(weirdary, chunk, sizeof(weirdary)); |
| 768 | #endif |
| 769 | |
| 770 | /* |
| 771 | * Add this free non-zero'd chunk to a linked list for reuse, adjust |
| 772 | * z_FirstFreePg. |
| 773 | */ |
| 774 | #ifdef INVARIANTS |
| 775 | if ((uintptr_t)chunk < VM_MIN_KERNEL_ADDRESS) |
| 776 | panic("BADFREE %p\n", chunk); |
| 777 | #endif |
| 778 | chunk->c_Next = z->z_PageAry[pgno]; |
| 779 | z->z_PageAry[pgno] = chunk; |
| 780 | #ifdef INVARIANTS |
| 781 | if (chunk->c_Next && (uintptr_t)chunk->c_Next < VM_MIN_KERNEL_ADDRESS) |
| 782 | panic("BADFREE2"); |
| 783 | #endif |
| 784 | if (z->z_FirstFreePg > pgno) |
| 785 | z->z_FirstFreePg = pgno; |
| 786 | |
| 787 | /* |
| 788 | * Bump the number of free chunks. If it becomes non-zero the zone |
| 789 | * must be added back onto the appropriate list. |
| 790 | */ |
| 791 | if (z->z_NFree++ == 0) { |
| 792 | z->z_Next = slgd->ZoneAry[z->z_ZoneIndex]; |
| 793 | slgd->ZoneAry[z->z_ZoneIndex] = z; |
| 794 | } |
| 795 | |
| 796 | --type->ks_inuse[z->z_Cpu]; |
| 797 | type->ks_memuse[z->z_Cpu] -= z->z_ChunkSize; |
| 798 | |
| 799 | /* |
| 800 | * If the zone becomes totally free, and there are other zones we |
| 801 | * can allocate from, move this zone to the FreeZones list. Since |
| 802 | * this code can be called from an IPI callback, do *NOT* try to mess |
| 803 | * with kernel_map here. Hysteresis will be performed at malloc() time. |
| 804 | */ |
| 805 | if (z->z_NFree == z->z_NMax && |
| 806 | (z->z_Next || slgd->ZoneAry[z->z_ZoneIndex] != z) |
| 807 | ) { |
| 808 | SLZone **pz; |
| 809 | |
| 810 | for (pz = &slgd->ZoneAry[z->z_ZoneIndex]; z != *pz; pz = &(*pz)->z_Next) |
| 811 | ; |
| 812 | *pz = z->z_Next; |
| 813 | z->z_Magic = -1; |
| 814 | z->z_Next = slgd->FreeZones; |
| 815 | slgd->FreeZones = z; |
| 816 | ++slgd->NFreeZones; |
| 817 | } |
| 818 | crit_exit(); |
| 819 | } |
| 820 | |
| 821 | /* |
| 822 | * kmem_slab_alloc() |
| 823 | * |
| 824 | * Directly allocate and wire kernel memory in PAGE_SIZE chunks with the |
| 825 | * specified alignment. M_* flags are expected in the flags field. |
| 826 | * |
| 827 | * Alignment must be a multiple of PAGE_SIZE. |
| 828 | * |
| 829 | * NOTE! XXX For the moment we use vm_map_entry_reserve/release(), |
| 830 | * but when we move zalloc() over to use this function as its backend |
| 831 | * we will have to switch to kreserve/krelease and call reserve(0) |
| 832 | * after the new space is made available. |
| 833 | */ |
| 834 | static void * |
| 835 | kmem_slab_alloc(vm_size_t size, vm_offset_t align, int flags) |
| 836 | { |
| 837 | vm_size_t i; |
| 838 | vm_offset_t addr; |
| 839 | vm_offset_t offset; |
| 840 | int count; |
| 841 | vm_map_t map = kernel_map; |
| 842 | |
| 843 | size = round_page(size); |
| 844 | addr = vm_map_min(map); |
| 845 | |
| 846 | /* |
| 847 | * Reserve properly aligned space from kernel_map |
| 848 | */ |
| 849 | count = vm_map_entry_reserve(MAP_RESERVE_COUNT); |
| 850 | crit_enter(); |
| 851 | vm_map_lock(map); |
| 852 | if (vm_map_findspace(map, vm_map_min(map), size, align, &addr)) { |
| 853 | vm_map_unlock(map); |
| 854 | if ((flags & (M_NOWAIT|M_NULLOK)) == 0) |
| 855 | panic("kmem_slab_alloc(): kernel_map ran out of space!"); |
| 856 | crit_exit(); |
| 857 | vm_map_entry_release(count); |
| 858 | return(NULL); |
| 859 | } |
| 860 | offset = addr - VM_MIN_KERNEL_ADDRESS; |
| 861 | vm_object_reference(kernel_object); |
| 862 | vm_map_insert(map, &count, |
| 863 | kernel_object, offset, addr, addr + size, |
| 864 | VM_PROT_ALL, VM_PROT_ALL, 0); |
| 865 | |
| 866 | /* |
| 867 | * Allocate the pages. Do not mess with the PG_ZERO flag yet. |
| 868 | */ |
| 869 | for (i = 0; i < size; i += PAGE_SIZE) { |
| 870 | vm_page_t m; |
| 871 | vm_pindex_t idx = OFF_TO_IDX(offset + i); |
| 872 | int zero = (flags & M_ZERO) ? VM_ALLOC_ZERO : 0; |
| 873 | |
| 874 | if ((flags & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT) |
| 875 | m = vm_page_alloc(kernel_object, idx, VM_ALLOC_INTERRUPT|zero); |
| 876 | else |
| 877 | m = vm_page_alloc(kernel_object, idx, VM_ALLOC_SYSTEM|zero); |
| 878 | if (m == NULL) { |
| 879 | if ((flags & M_NOWAIT) == 0) { |
| 880 | vm_map_unlock(map); |
| 881 | vm_wait(); |
| 882 | vm_map_lock(map); |
| 883 | i -= PAGE_SIZE; /* retry */ |
| 884 | continue; |
| 885 | } |
| 886 | while (i != 0) { |
| 887 | i -= PAGE_SIZE; |
| 888 | m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); |
| 889 | vm_page_free(m); |
| 890 | } |
| 891 | vm_map_delete(map, addr, addr + size, &count); |
| 892 | vm_map_unlock(map); |
| 893 | crit_exit(); |
| 894 | vm_map_entry_release(count); |
| 895 | return(NULL); |
| 896 | } |
| 897 | } |
| 898 | |
| 899 | /* |
| 900 | * Mark the map entry as non-pageable using a routine that allows us to |
| 901 | * populate the underlying pages. |
| 902 | */ |
| 903 | vm_map_set_wired_quick(map, addr, size, &count); |
| 904 | crit_exit(); |
| 905 | |
| 906 | /* |
| 907 | * Enter the pages into the pmap and deal with PG_ZERO and M_ZERO. |
| 908 | */ |
| 909 | for (i = 0; i < size; i += PAGE_SIZE) { |
| 910 | vm_page_t m; |
| 911 | |
| 912 | m = vm_page_lookup(kernel_object, OFF_TO_IDX(offset + i)); |
| 913 | m->valid = VM_PAGE_BITS_ALL; |
| 914 | vm_page_wire(m); |
| 915 | vm_page_wakeup(m); |
| 916 | pmap_enter(kernel_pmap, addr + i, m, VM_PROT_ALL, 1); |
| 917 | if ((m->flags & PG_ZERO) == 0 && (flags & M_ZERO)) |
| 918 | bzero((char *)addr + i, PAGE_SIZE); |
| 919 | vm_page_flag_clear(m, PG_ZERO); |
| 920 | vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE | PG_REFERENCED); |
| 921 | } |
| 922 | vm_map_unlock(map); |
| 923 | vm_map_entry_release(count); |
| 924 | return((void *)addr); |
| 925 | } |
| 926 | |
| 927 | static void |
| 928 | kmem_slab_free(void *ptr, vm_size_t size) |
| 929 | { |
| 930 | crit_enter(); |
| 931 | vm_map_remove(kernel_map, (vm_offset_t)ptr, (vm_offset_t)ptr + size); |
| 932 | crit_exit(); |
| 933 | } |
| 934 | |