1 1.64 mrg /* $NetBSD: jemalloc.c,v 1.64 2023/12/13 23:53:50 mrg Exp $ */ 2 1.2 ad 3 1.1 ad /*- 4 1.1 ad * Copyright (C) 2006,2007 Jason Evans <jasone (at) FreeBSD.org>. 5 1.58 ad * Copyright (C) 2023 Andrew Doran <ad (at) NetBSD.org>. 6 1.1 ad * All rights reserved. 7 1.1 ad * 8 1.1 ad * Redistribution and use in source and binary forms, with or without 9 1.1 ad * modification, are permitted provided that the following conditions 10 1.1 ad * are met: 11 1.1 ad * 1. Redistributions of source code must retain the above copyright 12 1.1 ad * notice(s), this list of conditions and the following disclaimer as 13 1.1 ad * the first lines of this file unmodified other than the possible 14 1.1 ad * addition of one or more copyright notices. 15 1.1 ad * 2. Redistributions in binary form must reproduce the above copyright 16 1.1 ad * notice(s), this list of conditions and the following disclaimer in 17 1.1 ad * the documentation and/or other materials provided with the 18 1.1 ad * distribution. 19 1.1 ad * 20 1.1 ad * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 21 1.1 ad * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 1.1 ad * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 23 1.1 ad * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE 24 1.1 ad * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 25 1.1 ad * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 26 1.1 ad * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR 27 1.1 ad * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 28 1.1 ad * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE 29 1.1 ad * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, 30 1.1 ad * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 31 1.1 ad * 32 1.1 ad ******************************************************************************* 33 1.1 ad * 34 1.1 ad * This allocator implementation is designed to provide scalable performance 35 1.1 ad * for multi-threaded programs on multi-processor systems. The following 36 1.1 ad * features are included for this purpose: 37 1.1 ad * 38 1.1 ad * + Multiple arenas are used if there are multiple CPUs, which reduces lock 39 1.1 ad * contention and cache sloshing. 40 1.1 ad * 41 1.1 ad * + Cache line sharing between arenas is avoided for internal data 42 1.1 ad * structures. 43 1.1 ad * 44 1.1 ad * + Memory is managed in chunks and runs (chunks can be split into runs), 45 1.1 ad * rather than as individual pages. This provides a constant-time 46 1.1 ad * mechanism for associating allocations with particular arenas. 47 1.1 ad * 48 1.1 ad * Allocation requests are rounded up to the nearest size class, and no record 49 1.1 ad * of the original request size is maintained. Allocations are broken into 50 1.1 ad * categories according to size class. Assuming runtime defaults, 4 kB pages 51 1.1 ad * and a 16 byte quantum, the size classes in each category are as follows: 52 1.1 ad * 53 1.1 ad * |=====================================| 54 1.1 ad * | Category | Subcategory | Size | 55 1.1 ad * |=====================================| 56 1.1 ad * | Small | Tiny | 2 | 57 1.1 ad * | | | 4 | 58 1.1 ad * | | | 8 | 59 1.1 ad * | |----------------+---------| 60 1.1 ad * | | Quantum-spaced | 16 | 61 1.1 ad * | | | 32 | 62 1.1 ad * | | | 48 | 63 1.1 ad * | | | ... | 64 1.1 ad * | | | 480 | 65 1.1 ad * | | | 496 | 66 1.1 ad * | | | 512 | 67 1.1 ad * | |----------------+---------| 68 1.1 ad * | | Sub-page | 1 kB | 69 1.1 ad * | | | 2 kB | 70 1.1 ad * |=====================================| 71 1.1 ad * | Large | 4 kB | 72 1.1 ad * | | 8 kB | 73 1.1 ad * | | 12 kB | 74 1.1 ad * | | ... | 75 1.1 ad * | | 1012 kB | 76 1.1 ad * | | 1016 kB | 77 1.1 ad * | | 1020 kB | 78 1.1 ad * |=====================================| 79 1.1 ad * | Huge | 1 MB | 80 1.1 ad * | | 2 MB | 81 1.1 ad * | | 3 MB | 82 1.1 ad * | | ... | 83 1.1 ad * |=====================================| 84 1.1 ad * 85 1.1 ad * A different mechanism is used for each category: 86 1.1 ad * 87 1.1 ad * Small : Each size class is segregated into its own set of runs. Each run 88 1.1 ad * maintains a bitmap of which regions are free/allocated. 89 1.1 ad * 90 1.1 ad * Large : Each allocation is backed by a dedicated run. Metadata are stored 91 1.1 ad * in the associated arena chunk header maps. 92 1.1 ad * 93 1.1 ad * Huge : Each allocation is backed by a dedicated contiguous set of chunks. 94 1.1 ad * Metadata are stored in a separate red-black tree. 95 1.1 ad * 96 1.1 ad ******************************************************************************* 97 1.1 ad */ 98 1.1 ad 99 1.2 ad /* LINTLIBRARY */ 100 1.2 ad 101 1.1 ad /* 102 1.1 ad * MALLOC_PRODUCTION disables assertions and statistics gathering. It also 103 1.1 ad * defaults the A and J runtime options to off. These settings are appropriate 104 1.1 ad * for production systems. 105 1.1 ad */ 106 1.2 ad #define MALLOC_PRODUCTION 107 1.1 ad 108 1.1 ad #ifndef MALLOC_PRODUCTION 109 1.1 ad # define MALLOC_DEBUG 110 1.1 ad #endif 111 1.1 ad 112 1.1 ad #include <sys/cdefs.h> 113 1.55 skrll /* __FBSDID("$FreeBSD: src/lib/libc/stdlib/malloc.c,v 1.147 2007/06/15 22:00:16 jasone Exp $"); */ 114 1.64 mrg __RCSID("$NetBSD: jemalloc.c,v 1.64 2023/12/13 23:53:50 mrg Exp $"); 115 1.1 ad 116 1.1 ad #include "namespace.h" 117 1.1 ad #include <sys/mman.h> 118 1.1 ad #include <sys/param.h> 119 1.1 ad #include <sys/time.h> 120 1.1 ad #include <sys/types.h> 121 1.1 ad #include <sys/sysctl.h> 122 1.58 ad #include <sys/rbtree.h> 123 1.1 ad #include <sys/uio.h> 124 1.1 ad #include <sys/ktrace.h> /* Must come after several other sys/ includes. */ 125 1.1 ad 126 1.1 ad #include <errno.h> 127 1.1 ad #include <limits.h> 128 1.1 ad #include <pthread.h> 129 1.1 ad #include <sched.h> 130 1.1 ad #include <stdarg.h> 131 1.1 ad #include <stdbool.h> 132 1.1 ad #include <stdio.h> 133 1.1 ad #include <stdint.h> 134 1.1 ad #include <stdlib.h> 135 1.1 ad #include <string.h> 136 1.1 ad #include <strings.h> 137 1.1 ad #include <unistd.h> 138 1.64 mrg #include <malloc.h> 139 1.1 ad 140 1.57 ad #include <reentrant.h> 141 1.57 ad #include "extern.h" 142 1.16 christos 143 1.41 christos #define STRERROR_R(a, b, c) strerror_r_ss(a, b, c); 144 1.1 ad 145 1.1 ad /* MALLOC_STATS enables statistics calculation. */ 146 1.1 ad #ifndef MALLOC_PRODUCTION 147 1.1 ad # define MALLOC_STATS 148 1.1 ad #endif 149 1.1 ad 150 1.1 ad #ifdef MALLOC_DEBUG 151 1.1 ad # ifdef NDEBUG 152 1.1 ad # undef NDEBUG 153 1.1 ad # endif 154 1.1 ad #else 155 1.1 ad # ifndef NDEBUG 156 1.1 ad # define NDEBUG 157 1.1 ad # endif 158 1.1 ad #endif 159 1.1 ad #include <assert.h> 160 1.1 ad 161 1.1 ad #ifdef MALLOC_DEBUG 162 1.1 ad /* Disable inlining to make debugging easier. */ 163 1.1 ad # define inline 164 1.1 ad #endif 165 1.1 ad 166 1.58 ad /* 167 1.58 ad * Compare two pointers of 64/32 bit width and produce a ternary 32-bit 168 1.58 ad * indicator without using conditionals that maintains the expected 169 1.58 ad * ordering: [negative, equal, positive]. 170 1.58 ad * 171 1.58 ad * XXX it depends on twos complement arithemetic. 172 1.58 ad * XXX maybe should be a built-in for rbtree? 173 1.58 ad */ 174 1.58 ad static inline int 175 1.58 ad ptrcmp(const void *pa, const void *pb) 176 1.58 ad { 177 1.58 ad #ifdef _LP64 178 1.60 ad const uintptr_t a = (uintptr_t)pa, b = (uintptr_t)pb; 179 1.60 ad const uintptr_t diff = a - b; 180 1.58 ad assert(((a | b) & 1) == 0); 181 1.60 ad return (int)(diff >> 32) | ((unsigned)diff >> 1); 182 1.58 ad #else 183 1.59 mrg return (intptr_t)pa - (intptr_t)pb; 184 1.58 ad #endif 185 1.58 ad } 186 1.58 ad 187 1.1 ad /* Size of stack-allocated buffer passed to strerror_r(). */ 188 1.1 ad #define STRERROR_BUF 64 189 1.1 ad 190 1.1 ad /* Minimum alignment of allocations is 2^QUANTUM_2POW_MIN bytes. */ 191 1.31 martin 192 1.31 martin /* 193 1.31 martin * If you touch the TINY_MIN_2POW definition for any architecture, please 194 1.31 martin * make sure to adjust the corresponding definition for JEMALLOC_TINY_MIN_2POW 195 1.31 martin * in the gcc 4.8 tree in dist/gcc/tree-ssa-ccp.c and verify that a native 196 1.31 martin * gcc is still buildable! 197 1.31 martin */ 198 1.31 martin 199 1.1 ad #ifdef __i386__ 200 1.1 ad # define QUANTUM_2POW_MIN 4 201 1.1 ad # define SIZEOF_PTR_2POW 2 202 1.1 ad # define USE_BRK 203 1.1 ad #endif 204 1.1 ad #ifdef __ia64__ 205 1.1 ad # define QUANTUM_2POW_MIN 4 206 1.1 ad # define SIZEOF_PTR_2POW 3 207 1.1 ad #endif 208 1.34 matt #ifdef __aarch64__ 209 1.34 matt # define QUANTUM_2POW_MIN 4 210 1.34 matt # define SIZEOF_PTR_2POW 3 211 1.47 joerg # define TINY_MIN_2POW 3 212 1.34 matt #endif 213 1.1 ad #ifdef __alpha__ 214 1.1 ad # define QUANTUM_2POW_MIN 4 215 1.1 ad # define SIZEOF_PTR_2POW 3 216 1.30 skrll # define TINY_MIN_2POW 3 217 1.1 ad #endif 218 1.1 ad #ifdef __sparc64__ 219 1.1 ad # define QUANTUM_2POW_MIN 4 220 1.1 ad # define SIZEOF_PTR_2POW 3 221 1.31 martin # define TINY_MIN_2POW 3 222 1.1 ad #endif 223 1.1 ad #ifdef __amd64__ 224 1.1 ad # define QUANTUM_2POW_MIN 4 225 1.1 ad # define SIZEOF_PTR_2POW 3 226 1.31 martin # define TINY_MIN_2POW 3 227 1.1 ad #endif 228 1.1 ad #ifdef __arm__ 229 1.1 ad # define QUANTUM_2POW_MIN 3 230 1.1 ad # define SIZEOF_PTR_2POW 2 231 1.1 ad # define USE_BRK 232 1.30 skrll # ifdef __ARM_EABI__ 233 1.30 skrll # define TINY_MIN_2POW 3 234 1.30 skrll # endif 235 1.1 ad #endif 236 1.1 ad #ifdef __powerpc__ 237 1.1 ad # define QUANTUM_2POW_MIN 4 238 1.1 ad # define SIZEOF_PTR_2POW 2 239 1.1 ad # define USE_BRK 240 1.32 martin # define TINY_MIN_2POW 3 241 1.1 ad #endif 242 1.3 he #if defined(__sparc__) && !defined(__sparc64__) 243 1.2 ad # define QUANTUM_2POW_MIN 4 244 1.2 ad # define SIZEOF_PTR_2POW 2 245 1.2 ad # define USE_BRK 246 1.2 ad #endif 247 1.35 matt #ifdef __or1k__ 248 1.35 matt # define QUANTUM_2POW_MIN 4 249 1.35 matt # define SIZEOF_PTR_2POW 2 250 1.35 matt # define USE_BRK 251 1.35 matt #endif 252 1.2 ad #ifdef __vax__ 253 1.2 ad # define QUANTUM_2POW_MIN 4 254 1.2 ad # define SIZEOF_PTR_2POW 2 255 1.2 ad # define USE_BRK 256 1.2 ad #endif 257 1.2 ad #ifdef __sh__ 258 1.2 ad # define QUANTUM_2POW_MIN 4 259 1.2 ad # define SIZEOF_PTR_2POW 2 260 1.2 ad # define USE_BRK 261 1.2 ad #endif 262 1.2 ad #ifdef __m68k__ 263 1.2 ad # define QUANTUM_2POW_MIN 4 264 1.2 ad # define SIZEOF_PTR_2POW 2 265 1.2 ad # define USE_BRK 266 1.2 ad #endif 267 1.56 skrll #if defined(__mips__) 268 1.44 mrg # ifdef _LP64 269 1.44 mrg # define SIZEOF_PTR_2POW 3 270 1.44 mrg # define TINY_MIN_2POW 3 271 1.44 mrg # else 272 1.44 mrg # define SIZEOF_PTR_2POW 2 273 1.44 mrg # endif 274 1.44 mrg # define QUANTUM_2POW_MIN 4 275 1.44 mrg # define USE_BRK 276 1.2 ad #endif 277 1.56 skrll #if defined(__riscv__) 278 1.56 skrll # ifdef _LP64 279 1.56 skrll # define SIZEOF_PTR_2POW 3 280 1.56 skrll # define TINY_MIN_2POW 3 281 1.56 skrll # else 282 1.56 skrll # define SIZEOF_PTR_2POW 2 283 1.56 skrll # endif 284 1.56 skrll # define QUANTUM_2POW_MIN 4 285 1.56 skrll # define USE_BRK 286 1.56 skrll #endif 287 1.55 skrll #ifdef __hppa__ 288 1.55 skrll # define QUANTUM_2POW_MIN 4 289 1.44 mrg # define TINY_MIN_2POW 4 290 1.55 skrll # define SIZEOF_PTR_2POW 2 291 1.55 skrll # define USE_BRK 292 1.55 skrll #endif 293 1.1 ad 294 1.1 ad #define SIZEOF_PTR (1 << SIZEOF_PTR_2POW) 295 1.1 ad 296 1.1 ad /* sizeof(int) == (1 << SIZEOF_INT_2POW). */ 297 1.1 ad #ifndef SIZEOF_INT_2POW 298 1.1 ad # define SIZEOF_INT_2POW 2 299 1.1 ad #endif 300 1.1 ad 301 1.1 ad /* 302 1.1 ad * Size and alignment of memory chunks that are allocated by the OS's virtual 303 1.1 ad * memory system. 304 1.1 ad */ 305 1.1 ad #define CHUNK_2POW_DEFAULT 20 306 1.1 ad 307 1.1 ad /* 308 1.1 ad * Maximum size of L1 cache line. This is used to avoid cache line aliasing, 309 1.1 ad * so over-estimates are okay (up to a point), but under-estimates will 310 1.1 ad * negatively affect performance. 311 1.1 ad */ 312 1.1 ad #define CACHELINE_2POW 6 313 1.1 ad #define CACHELINE ((size_t)(1 << CACHELINE_2POW)) 314 1.1 ad 315 1.1 ad /* Smallest size class to support. */ 316 1.30 skrll #ifndef TINY_MIN_2POW 317 1.30 skrll #define TINY_MIN_2POW 2 318 1.30 skrll #endif 319 1.1 ad 320 1.1 ad /* 321 1.1 ad * Maximum size class that is a multiple of the quantum, but not (necessarily) 322 1.1 ad * a power of 2. Above this size, allocations are rounded up to the nearest 323 1.1 ad * power of 2. 324 1.1 ad */ 325 1.1 ad #define SMALL_MAX_2POW_DEFAULT 9 326 1.1 ad #define SMALL_MAX_DEFAULT (1 << SMALL_MAX_2POW_DEFAULT) 327 1.1 ad 328 1.1 ad /* 329 1.22 njoly * RUN_MAX_OVRHD indicates maximum desired run header overhead. Runs are sized 330 1.22 njoly * as small as possible such that this setting is still honored, without 331 1.22 njoly * violating other constraints. The goal is to make runs as small as possible 332 1.22 njoly * without exceeding a per run external fragmentation threshold. 333 1.1 ad * 334 1.22 njoly * We use binary fixed point math for overhead computations, where the binary 335 1.22 njoly * point is implicitly RUN_BFP bits to the left. 336 1.1 ad * 337 1.22 njoly * Note that it is possible to set RUN_MAX_OVRHD low enough that it cannot be 338 1.22 njoly * honored for some/all object sizes, since there is one bit of header overhead 339 1.22 njoly * per object (plus a constant). This constraint is relaxed (ignored) for runs 340 1.22 njoly * that are so small that the per-region overhead is greater than: 341 1.22 njoly * 342 1.22 njoly * (RUN_MAX_OVRHD / (reg_size << (3+RUN_BFP)) 343 1.1 ad */ 344 1.22 njoly #define RUN_BFP 12 345 1.22 njoly /* \/ Implicit binary fixed point. */ 346 1.22 njoly #define RUN_MAX_OVRHD 0x0000003dU 347 1.22 njoly #define RUN_MAX_OVRHD_RELAX 0x00001800U 348 1.1 ad 349 1.1 ad /* Put a cap on small object run size. This overrides RUN_MAX_OVRHD. */ 350 1.1 ad #define RUN_MAX_SMALL_2POW 15 351 1.1 ad #define RUN_MAX_SMALL (1 << RUN_MAX_SMALL_2POW) 352 1.1 ad 353 1.1 ad /******************************************************************************/ 354 1.1 ad 355 1.2 ad #define malloc_mutex_t mutex_t 356 1.2 ad 357 1.2 ad /* Set to true once the allocator has been initialized. */ 358 1.2 ad static bool malloc_initialized = false; 359 1.2 ad 360 1.37 christos #ifdef _REENTRANT 361 1.2 ad /* Used to avoid initialization races. */ 362 1.2 ad static mutex_t init_lock = MUTEX_INITIALIZER; 363 1.2 ad #endif 364 1.1 ad 365 1.1 ad /******************************************************************************/ 366 1.1 ad /* 367 1.1 ad * Statistics data structures. 368 1.1 ad */ 369 1.1 ad 370 1.1 ad #ifdef MALLOC_STATS 371 1.1 ad 372 1.1 ad typedef struct malloc_bin_stats_s malloc_bin_stats_t; 373 1.1 ad struct malloc_bin_stats_s { 374 1.1 ad /* 375 1.1 ad * Number of allocation requests that corresponded to the size of this 376 1.1 ad * bin. 377 1.1 ad */ 378 1.1 ad uint64_t nrequests; 379 1.1 ad 380 1.1 ad /* Total number of runs created for this bin's size class. */ 381 1.1 ad uint64_t nruns; 382 1.1 ad 383 1.1 ad /* 384 1.1 ad * Total number of runs reused by extracting them from the runs tree for 385 1.1 ad * this bin's size class. 386 1.1 ad */ 387 1.1 ad uint64_t reruns; 388 1.1 ad 389 1.1 ad /* High-water mark for this bin. */ 390 1.1 ad unsigned long highruns; 391 1.1 ad 392 1.1 ad /* Current number of runs in this bin. */ 393 1.1 ad unsigned long curruns; 394 1.1 ad }; 395 1.1 ad 396 1.1 ad typedef struct arena_stats_s arena_stats_t; 397 1.1 ad struct arena_stats_s { 398 1.1 ad /* Number of bytes currently mapped. */ 399 1.1 ad size_t mapped; 400 1.1 ad 401 1.1 ad /* Per-size-category statistics. */ 402 1.1 ad size_t allocated_small; 403 1.1 ad uint64_t nmalloc_small; 404 1.1 ad uint64_t ndalloc_small; 405 1.1 ad 406 1.1 ad size_t allocated_large; 407 1.1 ad uint64_t nmalloc_large; 408 1.1 ad uint64_t ndalloc_large; 409 1.1 ad }; 410 1.1 ad 411 1.1 ad typedef struct chunk_stats_s chunk_stats_t; 412 1.1 ad struct chunk_stats_s { 413 1.1 ad /* Number of chunks that were allocated. */ 414 1.1 ad uint64_t nchunks; 415 1.1 ad 416 1.1 ad /* High-water mark for number of chunks allocated. */ 417 1.1 ad unsigned long highchunks; 418 1.1 ad 419 1.1 ad /* 420 1.1 ad * Current number of chunks allocated. This value isn't maintained for 421 1.1 ad * any other purpose, so keep track of it in order to be able to set 422 1.1 ad * highchunks. 423 1.1 ad */ 424 1.1 ad unsigned long curchunks; 425 1.1 ad }; 426 1.1 ad 427 1.1 ad #endif /* #ifdef MALLOC_STATS */ 428 1.1 ad 429 1.1 ad /******************************************************************************/ 430 1.1 ad /* 431 1.1 ad * Chunk data structures. 432 1.1 ad */ 433 1.1 ad 434 1.1 ad /* Tree of chunks. */ 435 1.1 ad typedef struct chunk_node_s chunk_node_t; 436 1.1 ad struct chunk_node_s { 437 1.1 ad /* Linkage for the chunk tree. */ 438 1.58 ad rb_node_t link; 439 1.1 ad 440 1.1 ad /* 441 1.1 ad * Pointer to the chunk that this tree node is responsible for. In some 442 1.1 ad * (but certainly not all) cases, this data structure is placed at the 443 1.1 ad * beginning of the corresponding chunk, so this field may point to this 444 1.1 ad * node. 445 1.1 ad */ 446 1.1 ad void *chunk; 447 1.1 ad 448 1.1 ad /* Total chunk size. */ 449 1.1 ad size_t size; 450 1.1 ad }; 451 1.1 ad typedef struct chunk_tree_s chunk_tree_t; 452 1.58 ad 453 1.58 ad static int chunk_comp(void *, const void *, const void *); 454 1.58 ad 455 1.58 ad static const rb_tree_ops_t chunk_tree_ops = { 456 1.58 ad .rbto_compare_nodes = chunk_comp, 457 1.58 ad .rbto_compare_key = chunk_comp, 458 1.58 ad .rbto_node_offset = offsetof(struct chunk_node_s, link), 459 1.58 ad }; 460 1.1 ad 461 1.1 ad /******************************************************************************/ 462 1.1 ad /* 463 1.1 ad * Arena data structures. 464 1.1 ad */ 465 1.1 ad 466 1.1 ad typedef struct arena_s arena_t; 467 1.1 ad typedef struct arena_bin_s arena_bin_t; 468 1.1 ad 469 1.1 ad typedef struct arena_chunk_map_s arena_chunk_map_t; 470 1.1 ad struct arena_chunk_map_s { 471 1.1 ad /* Number of pages in run. */ 472 1.1 ad uint32_t npages; 473 1.1 ad /* 474 1.1 ad * Position within run. For a free run, this is POS_FREE for the first 475 1.1 ad * and last pages. The POS_FREE special value makes it possible to 476 1.1 ad * quickly coalesce free runs. 477 1.1 ad * 478 1.1 ad * This is the limiting factor for chunksize; there can be at most 2^31 479 1.1 ad * pages in a run. 480 1.1 ad */ 481 1.1 ad #define POS_FREE ((uint32_t)0xffffffffU) 482 1.1 ad uint32_t pos; 483 1.1 ad }; 484 1.1 ad 485 1.1 ad /* Arena chunk header. */ 486 1.1 ad typedef struct arena_chunk_s arena_chunk_t; 487 1.1 ad struct arena_chunk_s { 488 1.58 ad /* Linkage for the arena's chunk tree. */ 489 1.58 ad rb_node_t link; 490 1.58 ad 491 1.1 ad /* Arena that owns the chunk. */ 492 1.1 ad arena_t *arena; 493 1.1 ad 494 1.1 ad /* 495 1.1 ad * Number of pages in use. This is maintained in order to make 496 1.1 ad * detection of empty chunks fast. 497 1.1 ad */ 498 1.1 ad uint32_t pages_used; 499 1.1 ad 500 1.1 ad /* 501 1.1 ad * Every time a free run larger than this value is created/coalesced, 502 1.1 ad * this value is increased. The only way that the value decreases is if 503 1.1 ad * arena_run_alloc() fails to find a free run as large as advertised by 504 1.1 ad * this value. 505 1.1 ad */ 506 1.1 ad uint32_t max_frun_npages; 507 1.1 ad 508 1.1 ad /* 509 1.1 ad * Every time a free run that starts at an earlier page than this value 510 1.1 ad * is created/coalesced, this value is decreased. It is reset in a 511 1.1 ad * similar fashion to max_frun_npages. 512 1.1 ad */ 513 1.1 ad uint32_t min_frun_ind; 514 1.1 ad 515 1.1 ad /* 516 1.1 ad * Map of pages within chunk that keeps track of free/large/small. For 517 1.1 ad * free runs, only the map entries for the first and last pages are 518 1.1 ad * kept up to date, so that free runs can be quickly coalesced. 519 1.1 ad */ 520 1.1 ad arena_chunk_map_t map[1]; /* Dynamically sized. */ 521 1.1 ad }; 522 1.1 ad typedef struct arena_chunk_tree_s arena_chunk_tree_t; 523 1.58 ad 524 1.58 ad static int arena_chunk_comp(void *, const void *, const void *); 525 1.58 ad 526 1.58 ad static const rb_tree_ops_t arena_chunk_tree_ops = { 527 1.58 ad .rbto_compare_nodes = arena_chunk_comp, 528 1.58 ad .rbto_compare_key = arena_chunk_comp, 529 1.58 ad .rbto_node_offset = offsetof(struct arena_chunk_s, link), 530 1.58 ad }; 531 1.1 ad 532 1.1 ad typedef struct arena_run_s arena_run_t; 533 1.1 ad struct arena_run_s { 534 1.1 ad /* Linkage for run trees. */ 535 1.58 ad rb_node_t link; 536 1.1 ad 537 1.1 ad #ifdef MALLOC_DEBUG 538 1.1 ad uint32_t magic; 539 1.1 ad # define ARENA_RUN_MAGIC 0x384adf93 540 1.1 ad #endif 541 1.1 ad 542 1.1 ad /* Bin this run is associated with. */ 543 1.1 ad arena_bin_t *bin; 544 1.1 ad 545 1.1 ad /* Index of first element that might have a free region. */ 546 1.1 ad unsigned regs_minelm; 547 1.1 ad 548 1.1 ad /* Number of free regions in run. */ 549 1.1 ad unsigned nfree; 550 1.1 ad 551 1.1 ad /* Bitmask of in-use regions (0: in use, 1: free). */ 552 1.1 ad unsigned regs_mask[1]; /* Dynamically sized. */ 553 1.1 ad }; 554 1.1 ad typedef struct arena_run_tree_s arena_run_tree_t; 555 1.58 ad 556 1.58 ad static int arena_run_comp(void *, const void *, const void *); 557 1.58 ad 558 1.58 ad static const rb_tree_ops_t arena_run_tree_ops = { 559 1.58 ad .rbto_compare_nodes = arena_run_comp, 560 1.58 ad .rbto_compare_key = arena_run_comp, 561 1.58 ad .rbto_node_offset = offsetof(struct arena_run_s, link), 562 1.58 ad }; 563 1.1 ad 564 1.1 ad struct arena_bin_s { 565 1.1 ad /* 566 1.1 ad * Current run being used to service allocations of this bin's size 567 1.1 ad * class. 568 1.1 ad */ 569 1.1 ad arena_run_t *runcur; 570 1.1 ad 571 1.1 ad /* 572 1.1 ad * Tree of non-full runs. This tree is used when looking for an 573 1.1 ad * existing run when runcur is no longer usable. We choose the 574 1.1 ad * non-full run that is lowest in memory; this policy tends to keep 575 1.1 ad * objects packed well, and it can also help reduce the number of 576 1.1 ad * almost-empty chunks. 577 1.1 ad */ 578 1.58 ad rb_tree_t runs; 579 1.1 ad 580 1.1 ad /* Size of regions in a run for this bin's size class. */ 581 1.1 ad size_t reg_size; 582 1.1 ad 583 1.1 ad /* Total size of a run for this bin's size class. */ 584 1.1 ad size_t run_size; 585 1.1 ad 586 1.1 ad /* Total number of regions in a run for this bin's size class. */ 587 1.1 ad uint32_t nregs; 588 1.1 ad 589 1.1 ad /* Number of elements in a run's regs_mask for this bin's size class. */ 590 1.1 ad uint32_t regs_mask_nelms; 591 1.1 ad 592 1.1 ad /* Offset of first region in a run for this bin's size class. */ 593 1.1 ad uint32_t reg0_offset; 594 1.1 ad 595 1.1 ad #ifdef MALLOC_STATS 596 1.1 ad /* Bin statistics. */ 597 1.1 ad malloc_bin_stats_t stats; 598 1.1 ad #endif 599 1.1 ad }; 600 1.1 ad 601 1.1 ad struct arena_s { 602 1.1 ad #ifdef MALLOC_DEBUG 603 1.1 ad uint32_t magic; 604 1.1 ad # define ARENA_MAGIC 0x947d3d24 605 1.1 ad #endif 606 1.1 ad 607 1.1 ad /* All operations on this arena require that mtx be locked. */ 608 1.1 ad malloc_mutex_t mtx; 609 1.1 ad 610 1.1 ad #ifdef MALLOC_STATS 611 1.1 ad arena_stats_t stats; 612 1.1 ad #endif 613 1.1 ad 614 1.1 ad /* 615 1.1 ad * Tree of chunks this arena manages. 616 1.1 ad */ 617 1.58 ad rb_tree_t chunks; 618 1.1 ad 619 1.1 ad /* 620 1.1 ad * In order to avoid rapid chunk allocation/deallocation when an arena 621 1.1 ad * oscillates right on the cusp of needing a new chunk, cache the most 622 1.1 ad * recently freed chunk. This caching is disabled by opt_hint. 623 1.1 ad * 624 1.1 ad * There is one spare chunk per arena, rather than one spare total, in 625 1.1 ad * order to avoid interactions between multiple threads that could make 626 1.1 ad * a single spare inadequate. 627 1.1 ad */ 628 1.1 ad arena_chunk_t *spare; 629 1.1 ad 630 1.1 ad /* 631 1.1 ad * bins is used to store rings of free regions of the following sizes, 632 1.1 ad * assuming a 16-byte quantum, 4kB pagesize, and default MALLOC_OPTIONS. 633 1.1 ad * 634 1.1 ad * bins[i] | size | 635 1.1 ad * --------+------+ 636 1.1 ad * 0 | 2 | 637 1.1 ad * 1 | 4 | 638 1.1 ad * 2 | 8 | 639 1.1 ad * --------+------+ 640 1.1 ad * 3 | 16 | 641 1.1 ad * 4 | 32 | 642 1.1 ad * 5 | 48 | 643 1.1 ad * 6 | 64 | 644 1.1 ad * : : 645 1.1 ad * : : 646 1.1 ad * 33 | 496 | 647 1.1 ad * 34 | 512 | 648 1.1 ad * --------+------+ 649 1.1 ad * 35 | 1024 | 650 1.1 ad * 36 | 2048 | 651 1.1 ad * --------+------+ 652 1.1 ad */ 653 1.1 ad arena_bin_t bins[1]; /* Dynamically sized. */ 654 1.1 ad }; 655 1.1 ad 656 1.1 ad /******************************************************************************/ 657 1.1 ad /* 658 1.1 ad * Data. 659 1.1 ad */ 660 1.1 ad 661 1.1 ad /* Number of CPUs. */ 662 1.1 ad static unsigned ncpus; 663 1.1 ad 664 1.1 ad /* VM page size. */ 665 1.1 ad static size_t pagesize; 666 1.1 ad static size_t pagesize_mask; 667 1.9 christos static int pagesize_2pow; 668 1.1 ad 669 1.1 ad /* Various bin-related settings. */ 670 1.1 ad static size_t bin_maxclass; /* Max size class for bins. */ 671 1.1 ad static unsigned ntbins; /* Number of (2^n)-spaced tiny bins. */ 672 1.1 ad static unsigned nqbins; /* Number of quantum-spaced bins. */ 673 1.1 ad static unsigned nsbins; /* Number of (2^n)-spaced sub-page bins. */ 674 1.1 ad static size_t small_min; 675 1.1 ad static size_t small_max; 676 1.1 ad 677 1.1 ad /* Various quantum-related settings. */ 678 1.1 ad static size_t quantum; 679 1.1 ad static size_t quantum_mask; /* (quantum - 1). */ 680 1.1 ad 681 1.1 ad /* Various chunk-related settings. */ 682 1.1 ad static size_t chunksize; 683 1.1 ad static size_t chunksize_mask; /* (chunksize - 1). */ 684 1.5 yamt static int chunksize_2pow; 685 1.1 ad static unsigned chunk_npages; 686 1.1 ad static unsigned arena_chunk_header_npages; 687 1.1 ad static size_t arena_maxclass; /* Max size class for arenas. */ 688 1.1 ad 689 1.1 ad /********/ 690 1.1 ad /* 691 1.1 ad * Chunks. 692 1.1 ad */ 693 1.1 ad 694 1.37 christos #ifdef _REENTRANT 695 1.1 ad /* Protects chunk-related data structures. */ 696 1.1 ad static malloc_mutex_t chunks_mtx; 697 1.37 christos #endif 698 1.1 ad 699 1.1 ad /* Tree of chunks that are stand-alone huge allocations. */ 700 1.58 ad static rb_tree_t huge; 701 1.1 ad 702 1.1 ad #ifdef USE_BRK 703 1.1 ad /* 704 1.1 ad * Try to use brk for chunk-size allocations, due to address space constraints. 705 1.1 ad */ 706 1.1 ad /* 707 1.1 ad * Protects sbrk() calls. This must be separate from chunks_mtx, since 708 1.1 ad * base_pages_alloc() also uses sbrk(), but cannot lock chunks_mtx (doing so 709 1.1 ad * could cause recursive lock acquisition). 710 1.1 ad */ 711 1.46 christos #ifdef _REENTRANT 712 1.1 ad static malloc_mutex_t brk_mtx; 713 1.46 christos #endif 714 1.1 ad /* Result of first sbrk(0) call. */ 715 1.1 ad static void *brk_base; 716 1.1 ad /* Current end of brk, or ((void *)-1) if brk is exhausted. */ 717 1.1 ad static void *brk_prev; 718 1.1 ad /* Current upper limit on brk addresses. */ 719 1.1 ad static void *brk_max; 720 1.1 ad #endif 721 1.1 ad 722 1.1 ad #ifdef MALLOC_STATS 723 1.1 ad /* Huge allocation statistics. */ 724 1.1 ad static uint64_t huge_nmalloc; 725 1.1 ad static uint64_t huge_ndalloc; 726 1.8 yamt static uint64_t huge_nralloc; 727 1.1 ad static size_t huge_allocated; 728 1.1 ad #endif 729 1.1 ad 730 1.1 ad /* 731 1.1 ad * Tree of chunks that were previously allocated. This is used when allocating 732 1.1 ad * chunks, in an attempt to re-use address space. 733 1.1 ad */ 734 1.58 ad static rb_tree_t old_chunks; 735 1.1 ad 736 1.1 ad /****************************/ 737 1.1 ad /* 738 1.1 ad * base (internal allocation). 739 1.1 ad */ 740 1.1 ad 741 1.1 ad /* 742 1.1 ad * Current pages that are being used for internal memory allocations. These 743 1.1 ad * pages are carved up in cacheline-size quanta, so that there is no chance of 744 1.1 ad * false cache line sharing. 745 1.1 ad */ 746 1.1 ad static void *base_pages; 747 1.1 ad static void *base_next_addr; 748 1.1 ad static void *base_past_addr; /* Addr immediately past base_pages. */ 749 1.1 ad static chunk_node_t *base_chunk_nodes; /* LIFO cache of chunk nodes. */ 750 1.37 christos #ifdef _REENTRANT 751 1.1 ad static malloc_mutex_t base_mtx; 752 1.37 christos #endif 753 1.1 ad #ifdef MALLOC_STATS 754 1.1 ad static size_t base_mapped; 755 1.1 ad #endif 756 1.1 ad 757 1.1 ad /********/ 758 1.1 ad /* 759 1.1 ad * Arenas. 760 1.1 ad */ 761 1.1 ad 762 1.1 ad /* 763 1.1 ad * Arenas that are used to service external requests. Not all elements of the 764 1.1 ad * arenas array are necessarily used; arenas are created lazily as needed. 765 1.1 ad */ 766 1.1 ad static arena_t **arenas; 767 1.37 christos #ifdef _REENTRANT 768 1.1 ad static malloc_mutex_t arenas_mtx; /* Protects arenas initialization. */ 769 1.37 christos #endif 770 1.1 ad 771 1.1 ad #ifdef MALLOC_STATS 772 1.1 ad /* Chunk statistics. */ 773 1.1 ad static chunk_stats_t stats_chunks; 774 1.1 ad #endif 775 1.1 ad 776 1.1 ad /*******************************/ 777 1.1 ad /* 778 1.1 ad * Runtime configuration options. 779 1.1 ad */ 780 1.1 ad const char *_malloc_options; 781 1.1 ad 782 1.1 ad #ifndef MALLOC_PRODUCTION 783 1.1 ad static bool opt_abort = true; 784 1.1 ad static bool opt_junk = true; 785 1.1 ad #else 786 1.1 ad static bool opt_abort = false; 787 1.1 ad static bool opt_junk = false; 788 1.1 ad #endif 789 1.1 ad static bool opt_hint = false; 790 1.1 ad static bool opt_print_stats = false; 791 1.9 christos static int opt_quantum_2pow = QUANTUM_2POW_MIN; 792 1.9 christos static int opt_small_max_2pow = SMALL_MAX_2POW_DEFAULT; 793 1.9 christos static int opt_chunk_2pow = CHUNK_2POW_DEFAULT; 794 1.1 ad static bool opt_utrace = false; 795 1.1 ad static bool opt_sysv = false; 796 1.1 ad static bool opt_xmalloc = false; 797 1.1 ad static bool opt_zero = false; 798 1.1 ad 799 1.1 ad typedef struct { 800 1.1 ad void *p; 801 1.1 ad size_t s; 802 1.1 ad void *r; 803 1.1 ad } malloc_utrace_t; 804 1.1 ad 805 1.57 ad /* Sprinkle branch hints for the compiler and CPU. */ 806 1.57 ad #define OPT(a) __predict_false(opt_##a) 807 1.57 ad #define NOT_OPT(a) __predict_true(!opt_##a) 808 1.57 ad 809 1.57 ad /* Trace malloc/free for ktrace/kdump. */ 810 1.1 ad #define UTRACE(a, b, c) \ 811 1.57 ad if (OPT(utrace)) { \ 812 1.2 ad malloc_utrace_t ut; \ 813 1.2 ad ut.p = a; \ 814 1.2 ad ut.s = b; \ 815 1.2 ad ut.r = c; \ 816 1.57 ad utrace("malloc", &ut, sizeof(ut)); \ 817 1.1 ad } 818 1.1 ad 819 1.1 ad /******************************************************************************/ 820 1.1 ad /* 821 1.1 ad * Begin function prototypes for non-inline static functions. 822 1.1 ad */ 823 1.1 ad 824 1.1 ad static void wrtmessage(const char *p1, const char *p2, const char *p3, 825 1.1 ad const char *p4); 826 1.1 ad #ifdef MALLOC_STATS 827 1.1 ad static void malloc_printf(const char *format, ...); 828 1.1 ad #endif 829 1.28 christos static char *size_t2s(size_t x, char *s); 830 1.1 ad static bool base_pages_alloc(size_t minsize); 831 1.1 ad static void *base_alloc(size_t size); 832 1.1 ad static chunk_node_t *base_chunk_node_alloc(void); 833 1.1 ad static void base_chunk_node_dealloc(chunk_node_t *node); 834 1.1 ad #ifdef MALLOC_STATS 835 1.1 ad static void stats_print(arena_t *arena); 836 1.1 ad #endif 837 1.1 ad static void *pages_map(void *addr, size_t size); 838 1.5 yamt static void *pages_map_align(void *addr, size_t size, int align); 839 1.1 ad static void pages_unmap(void *addr, size_t size); 840 1.1 ad static void *chunk_alloc(size_t size); 841 1.1 ad static void chunk_dealloc(void *chunk, size_t size); 842 1.1 ad static void arena_run_split(arena_t *arena, arena_run_t *run, size_t size); 843 1.1 ad static arena_chunk_t *arena_chunk_alloc(arena_t *arena); 844 1.1 ad static void arena_chunk_dealloc(arena_t *arena, arena_chunk_t *chunk); 845 1.1 ad static arena_run_t *arena_run_alloc(arena_t *arena, size_t size); 846 1.1 ad static void arena_run_dalloc(arena_t *arena, arena_run_t *run, size_t size); 847 1.1 ad static arena_run_t *arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin); 848 1.1 ad static void *arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin); 849 1.1 ad static size_t arena_bin_run_size_calc(arena_bin_t *bin, size_t min_run_size); 850 1.1 ad static void *arena_malloc(arena_t *arena, size_t size); 851 1.1 ad static void *arena_palloc(arena_t *arena, size_t alignment, size_t size, 852 1.1 ad size_t alloc_size); 853 1.1 ad static size_t arena_salloc(const void *ptr); 854 1.1 ad static void *arena_ralloc(void *ptr, size_t size, size_t oldsize); 855 1.1 ad static void arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr); 856 1.57 ad static void arena_new(arena_t *arena); 857 1.57 ad static arena_t *arenas_extend(void); 858 1.1 ad static void *huge_malloc(size_t size); 859 1.1 ad static void *huge_palloc(size_t alignment, size_t size); 860 1.1 ad static void *huge_ralloc(void *ptr, size_t size, size_t oldsize); 861 1.1 ad static void huge_dalloc(void *ptr); 862 1.1 ad static void *imalloc(size_t size); 863 1.1 ad static void *ipalloc(size_t alignment, size_t size); 864 1.1 ad static void *icalloc(size_t size); 865 1.1 ad static size_t isalloc(const void *ptr); 866 1.1 ad static void *iralloc(void *ptr, size_t size); 867 1.1 ad static void idalloc(void *ptr); 868 1.1 ad static void malloc_print_stats(void); 869 1.1 ad static bool malloc_init_hard(void); 870 1.1 ad 871 1.1 ad /* 872 1.1 ad * End function prototypes. 873 1.1 ad */ 874 1.1 ad /******************************************************************************/ 875 1.1 ad /* 876 1.1 ad * Begin mutex. 877 1.1 ad */ 878 1.1 ad 879 1.2 ad #define malloc_mutex_init(m) mutex_init(m, NULL) 880 1.2 ad #define malloc_mutex_lock(m) mutex_lock(m) 881 1.2 ad #define malloc_mutex_unlock(m) mutex_unlock(m) 882 1.1 ad 883 1.1 ad /* 884 1.1 ad * End mutex. 885 1.1 ad */ 886 1.1 ad /******************************************************************************/ 887 1.1 ad /* 888 1.1 ad * Begin Utility functions/macros. 889 1.1 ad */ 890 1.1 ad 891 1.1 ad /* Return the chunk address for allocation address a. */ 892 1.1 ad #define CHUNK_ADDR2BASE(a) \ 893 1.1 ad ((void *)((uintptr_t)(a) & ~chunksize_mask)) 894 1.1 ad 895 1.1 ad /* Return the chunk offset of address a. */ 896 1.1 ad #define CHUNK_ADDR2OFFSET(a) \ 897 1.1 ad ((size_t)((uintptr_t)(a) & chunksize_mask)) 898 1.1 ad 899 1.1 ad /* Return the smallest chunk multiple that is >= s. */ 900 1.1 ad #define CHUNK_CEILING(s) \ 901 1.1 ad (((s) + chunksize_mask) & ~chunksize_mask) 902 1.1 ad 903 1.1 ad /* Return the smallest cacheline multiple that is >= s. */ 904 1.1 ad #define CACHELINE_CEILING(s) \ 905 1.1 ad (((s) + (CACHELINE - 1)) & ~(CACHELINE - 1)) 906 1.1 ad 907 1.1 ad /* Return the smallest quantum multiple that is >= a. */ 908 1.1 ad #define QUANTUM_CEILING(a) \ 909 1.1 ad (((a) + quantum_mask) & ~quantum_mask) 910 1.1 ad 911 1.1 ad /* Return the smallest pagesize multiple that is >= s. */ 912 1.1 ad #define PAGE_CEILING(s) \ 913 1.1 ad (((s) + pagesize_mask) & ~pagesize_mask) 914 1.1 ad 915 1.1 ad /* Compute the smallest power of 2 that is >= x. */ 916 1.1 ad static inline size_t 917 1.1 ad pow2_ceil(size_t x) 918 1.1 ad { 919 1.1 ad 920 1.1 ad x--; 921 1.1 ad x |= x >> 1; 922 1.1 ad x |= x >> 2; 923 1.1 ad x |= x >> 4; 924 1.1 ad x |= x >> 8; 925 1.1 ad x |= x >> 16; 926 1.1 ad #if (SIZEOF_PTR == 8) 927 1.1 ad x |= x >> 32; 928 1.1 ad #endif 929 1.1 ad x++; 930 1.1 ad return (x); 931 1.1 ad } 932 1.1 ad 933 1.1 ad static void 934 1.1 ad wrtmessage(const char *p1, const char *p2, const char *p3, const char *p4) 935 1.1 ad { 936 1.1 ad 937 1.16 christos write(STDERR_FILENO, p1, strlen(p1)); 938 1.16 christos write(STDERR_FILENO, p2, strlen(p2)); 939 1.16 christos write(STDERR_FILENO, p3, strlen(p3)); 940 1.16 christos write(STDERR_FILENO, p4, strlen(p4)); 941 1.1 ad } 942 1.1 ad 943 1.1 ad void (*_malloc_message)(const char *p1, const char *p2, const char *p3, 944 1.1 ad const char *p4) = wrtmessage; 945 1.1 ad 946 1.1 ad #ifdef MALLOC_STATS 947 1.1 ad /* 948 1.1 ad * Print to stderr in such a way as to (hopefully) avoid memory allocation. 949 1.1 ad */ 950 1.1 ad static void 951 1.1 ad malloc_printf(const char *format, ...) 952 1.1 ad { 953 1.1 ad char buf[4096]; 954 1.1 ad va_list ap; 955 1.1 ad 956 1.1 ad va_start(ap, format); 957 1.1 ad vsnprintf(buf, sizeof(buf), format, ap); 958 1.1 ad va_end(ap); 959 1.1 ad _malloc_message(buf, "", "", ""); 960 1.1 ad } 961 1.1 ad #endif 962 1.1 ad 963 1.1 ad /* 964 1.1 ad * We don't want to depend on vsnprintf() for production builds, since that can 965 1.28 christos * cause unnecessary bloat for static binaries. size_t2s() provides minimal 966 1.1 ad * integer printing functionality, so that malloc_printf() use can be limited to 967 1.1 ad * MALLOC_STATS code. 968 1.1 ad */ 969 1.1 ad #define UMAX2S_BUFSIZE 21 970 1.1 ad static char * 971 1.28 christos size_t2s(size_t x, char *s) 972 1.1 ad { 973 1.1 ad unsigned i; 974 1.1 ad 975 1.1 ad /* Make sure UMAX2S_BUFSIZE is large enough. */ 976 1.7 yamt /* LINTED */ 977 1.25 christos assert(sizeof(size_t) <= 8); 978 1.1 ad 979 1.1 ad i = UMAX2S_BUFSIZE - 1; 980 1.1 ad s[i] = '\0'; 981 1.1 ad do { 982 1.1 ad i--; 983 1.2 ad s[i] = "0123456789"[(int)x % 10]; 984 1.2 ad x /= (uintmax_t)10LL; 985 1.1 ad } while (x > 0); 986 1.1 ad 987 1.1 ad return (&s[i]); 988 1.1 ad } 989 1.1 ad 990 1.1 ad /******************************************************************************/ 991 1.1 ad 992 1.1 ad static bool 993 1.1 ad base_pages_alloc(size_t minsize) 994 1.1 ad { 995 1.2 ad size_t csize = 0; 996 1.1 ad 997 1.1 ad #ifdef USE_BRK 998 1.1 ad /* 999 1.1 ad * Do special brk allocation here, since base allocations don't need to 1000 1.1 ad * be chunk-aligned. 1001 1.1 ad */ 1002 1.1 ad if (brk_prev != (void *)-1) { 1003 1.1 ad void *brk_cur; 1004 1.1 ad intptr_t incr; 1005 1.1 ad 1006 1.1 ad if (minsize != 0) 1007 1.1 ad csize = CHUNK_CEILING(minsize); 1008 1.1 ad 1009 1.1 ad malloc_mutex_lock(&brk_mtx); 1010 1.1 ad do { 1011 1.1 ad /* Get the current end of brk. */ 1012 1.1 ad brk_cur = sbrk(0); 1013 1.1 ad 1014 1.1 ad /* 1015 1.1 ad * Calculate how much padding is necessary to 1016 1.1 ad * chunk-align the end of brk. Don't worry about 1017 1.1 ad * brk_cur not being chunk-aligned though. 1018 1.1 ad */ 1019 1.1 ad incr = (intptr_t)chunksize 1020 1.1 ad - (intptr_t)CHUNK_ADDR2OFFSET(brk_cur); 1021 1.20 lukem assert(incr >= 0); 1022 1.20 lukem if ((size_t)incr < minsize) 1023 1.1 ad incr += csize; 1024 1.1 ad 1025 1.1 ad brk_prev = sbrk(incr); 1026 1.1 ad if (brk_prev == brk_cur) { 1027 1.1 ad /* Success. */ 1028 1.1 ad malloc_mutex_unlock(&brk_mtx); 1029 1.1 ad base_pages = brk_cur; 1030 1.1 ad base_next_addr = base_pages; 1031 1.1 ad base_past_addr = (void *)((uintptr_t)base_pages 1032 1.1 ad + incr); 1033 1.1 ad #ifdef MALLOC_STATS 1034 1.1 ad base_mapped += incr; 1035 1.1 ad #endif 1036 1.1 ad return (false); 1037 1.1 ad } 1038 1.1 ad } while (brk_prev != (void *)-1); 1039 1.1 ad malloc_mutex_unlock(&brk_mtx); 1040 1.1 ad } 1041 1.1 ad if (minsize == 0) { 1042 1.1 ad /* 1043 1.1 ad * Failure during initialization doesn't matter, so avoid 1044 1.1 ad * falling through to the mmap-based page mapping code. 1045 1.1 ad */ 1046 1.1 ad return (true); 1047 1.1 ad } 1048 1.1 ad #endif 1049 1.1 ad assert(minsize != 0); 1050 1.1 ad csize = PAGE_CEILING(minsize); 1051 1.1 ad base_pages = pages_map(NULL, csize); 1052 1.1 ad if (base_pages == NULL) 1053 1.1 ad return (true); 1054 1.1 ad base_next_addr = base_pages; 1055 1.1 ad base_past_addr = (void *)((uintptr_t)base_pages + csize); 1056 1.1 ad #ifdef MALLOC_STATS 1057 1.1 ad base_mapped += csize; 1058 1.1 ad #endif 1059 1.1 ad return (false); 1060 1.1 ad } 1061 1.1 ad 1062 1.1 ad static void * 1063 1.1 ad base_alloc(size_t size) 1064 1.1 ad { 1065 1.1 ad void *ret; 1066 1.1 ad size_t csize; 1067 1.1 ad 1068 1.1 ad /* Round size up to nearest multiple of the cacheline size. */ 1069 1.1 ad csize = CACHELINE_CEILING(size); 1070 1.1 ad 1071 1.1 ad malloc_mutex_lock(&base_mtx); 1072 1.1 ad 1073 1.1 ad /* Make sure there's enough space for the allocation. */ 1074 1.1 ad if ((uintptr_t)base_next_addr + csize > (uintptr_t)base_past_addr) { 1075 1.1 ad if (base_pages_alloc(csize)) { 1076 1.1 ad ret = NULL; 1077 1.1 ad goto RETURN; 1078 1.1 ad } 1079 1.1 ad } 1080 1.1 ad 1081 1.1 ad /* Allocate. */ 1082 1.1 ad ret = base_next_addr; 1083 1.1 ad base_next_addr = (void *)((uintptr_t)base_next_addr + csize); 1084 1.1 ad 1085 1.1 ad RETURN: 1086 1.1 ad malloc_mutex_unlock(&base_mtx); 1087 1.1 ad return (ret); 1088 1.1 ad } 1089 1.1 ad 1090 1.1 ad static chunk_node_t * 1091 1.1 ad base_chunk_node_alloc(void) 1092 1.1 ad { 1093 1.1 ad chunk_node_t *ret; 1094 1.1 ad 1095 1.1 ad malloc_mutex_lock(&base_mtx); 1096 1.1 ad if (base_chunk_nodes != NULL) { 1097 1.1 ad ret = base_chunk_nodes; 1098 1.2 ad /* LINTED */ 1099 1.1 ad base_chunk_nodes = *(chunk_node_t **)ret; 1100 1.1 ad malloc_mutex_unlock(&base_mtx); 1101 1.1 ad } else { 1102 1.1 ad malloc_mutex_unlock(&base_mtx); 1103 1.1 ad ret = (chunk_node_t *)base_alloc(sizeof(chunk_node_t)); 1104 1.1 ad } 1105 1.1 ad 1106 1.1 ad return (ret); 1107 1.1 ad } 1108 1.1 ad 1109 1.1 ad static void 1110 1.1 ad base_chunk_node_dealloc(chunk_node_t *node) 1111 1.1 ad { 1112 1.1 ad 1113 1.1 ad malloc_mutex_lock(&base_mtx); 1114 1.2 ad /* LINTED */ 1115 1.1 ad *(chunk_node_t **)node = base_chunk_nodes; 1116 1.1 ad base_chunk_nodes = node; 1117 1.1 ad malloc_mutex_unlock(&base_mtx); 1118 1.1 ad } 1119 1.1 ad 1120 1.1 ad /******************************************************************************/ 1121 1.1 ad 1122 1.1 ad #ifdef MALLOC_STATS 1123 1.1 ad static void 1124 1.1 ad stats_print(arena_t *arena) 1125 1.1 ad { 1126 1.57 ad const unsigned minusone = (unsigned)-1; 1127 1.57 ad unsigned i, gap_start; 1128 1.1 ad 1129 1.1 ad malloc_printf( 1130 1.1 ad " allocated/mapped nmalloc ndalloc\n"); 1131 1.2 ad 1132 1.2 ad malloc_printf("small: %12zu %-12s %12llu %12llu\n", 1133 1.1 ad arena->stats.allocated_small, "", arena->stats.nmalloc_small, 1134 1.1 ad arena->stats.ndalloc_small); 1135 1.2 ad malloc_printf("large: %12zu %-12s %12llu %12llu\n", 1136 1.1 ad arena->stats.allocated_large, "", arena->stats.nmalloc_large, 1137 1.1 ad arena->stats.ndalloc_large); 1138 1.2 ad malloc_printf("total: %12zu/%-12zu %12llu %12llu\n", 1139 1.1 ad arena->stats.allocated_small + arena->stats.allocated_large, 1140 1.1 ad arena->stats.mapped, 1141 1.1 ad arena->stats.nmalloc_small + arena->stats.nmalloc_large, 1142 1.1 ad arena->stats.ndalloc_small + arena->stats.ndalloc_large); 1143 1.1 ad 1144 1.1 ad malloc_printf("bins: bin size regs pgs requests newruns" 1145 1.1 ad " reruns maxruns curruns\n"); 1146 1.57 ad for (i = 0, gap_start = minusone; i < ntbins + nqbins + nsbins; i++) { 1147 1.1 ad if (arena->bins[i].stats.nrequests == 0) { 1148 1.57 ad if (gap_start == minusone) 1149 1.1 ad gap_start = i; 1150 1.1 ad } else { 1151 1.57 ad if (gap_start != minusone) { 1152 1.1 ad if (i > gap_start + 1) { 1153 1.1 ad /* Gap of more than one size class. */ 1154 1.1 ad malloc_printf("[%u..%u]\n", 1155 1.1 ad gap_start, i - 1); 1156 1.1 ad } else { 1157 1.1 ad /* Gap of one size class. */ 1158 1.1 ad malloc_printf("[%u]\n", gap_start); 1159 1.1 ad } 1160 1.57 ad gap_start = minusone; 1161 1.1 ad } 1162 1.1 ad malloc_printf( 1163 1.1 ad "%13u %1s %4u %4u %3u %9llu %9llu" 1164 1.1 ad " %9llu %7lu %7lu\n", 1165 1.1 ad i, 1166 1.1 ad i < ntbins ? "T" : i < ntbins + nqbins ? "Q" : "S", 1167 1.1 ad arena->bins[i].reg_size, 1168 1.1 ad arena->bins[i].nregs, 1169 1.1 ad arena->bins[i].run_size >> pagesize_2pow, 1170 1.1 ad arena->bins[i].stats.nrequests, 1171 1.1 ad arena->bins[i].stats.nruns, 1172 1.1 ad arena->bins[i].stats.reruns, 1173 1.1 ad arena->bins[i].stats.highruns, 1174 1.1 ad arena->bins[i].stats.curruns); 1175 1.1 ad } 1176 1.1 ad } 1177 1.57 ad if (gap_start != minusone) { 1178 1.1 ad if (i > gap_start + 1) { 1179 1.1 ad /* Gap of more than one size class. */ 1180 1.1 ad malloc_printf("[%u..%u]\n", gap_start, i - 1); 1181 1.1 ad } else { 1182 1.1 ad /* Gap of one size class. */ 1183 1.1 ad malloc_printf("[%u]\n", gap_start); 1184 1.1 ad } 1185 1.1 ad } 1186 1.1 ad } 1187 1.1 ad #endif 1188 1.1 ad 1189 1.1 ad /* 1190 1.1 ad * End Utility functions/macros. 1191 1.1 ad */ 1192 1.1 ad /******************************************************************************/ 1193 1.1 ad /* 1194 1.1 ad * Begin chunk management functions. 1195 1.1 ad */ 1196 1.1 ad 1197 1.58 ad static int 1198 1.58 ad chunk_comp(void *context, const void *va, const void *vb) 1199 1.1 ad { 1200 1.58 ad const chunk_node_t *a = va, *b = vb; 1201 1.1 ad 1202 1.1 ad assert(a != NULL); 1203 1.1 ad assert(b != NULL); 1204 1.1 ad 1205 1.58 ad return ptrcmp(a->chunk, b->chunk); 1206 1.1 ad } 1207 1.1 ad 1208 1.1 ad static void * 1209 1.5 yamt pages_map_align(void *addr, size_t size, int align) 1210 1.1 ad { 1211 1.1 ad void *ret; 1212 1.1 ad 1213 1.1 ad /* 1214 1.1 ad * We don't use MAP_FIXED here, because it can cause the *replacement* 1215 1.1 ad * of existing mappings, and we only want to create new mappings. 1216 1.1 ad */ 1217 1.5 yamt ret = mmap(addr, size, PROT_READ | PROT_WRITE, 1218 1.5 yamt MAP_PRIVATE | MAP_ANON | MAP_ALIGNED(align), -1, 0); 1219 1.1 ad assert(ret != NULL); 1220 1.1 ad 1221 1.1 ad if (ret == MAP_FAILED) 1222 1.1 ad ret = NULL; 1223 1.1 ad else if (addr != NULL && ret != addr) { 1224 1.1 ad /* 1225 1.1 ad * We succeeded in mapping memory, but not in the right place. 1226 1.1 ad */ 1227 1.1 ad if (munmap(ret, size) == -1) { 1228 1.1 ad char buf[STRERROR_BUF]; 1229 1.1 ad 1230 1.16 christos STRERROR_R(errno, buf, sizeof(buf)); 1231 1.16 christos _malloc_message(getprogname(), 1232 1.1 ad ": (malloc) Error in munmap(): ", buf, "\n"); 1233 1.57 ad if (OPT(abort)) 1234 1.1 ad abort(); 1235 1.1 ad } 1236 1.1 ad ret = NULL; 1237 1.1 ad } 1238 1.1 ad 1239 1.1 ad assert(ret == NULL || (addr == NULL && ret != addr) 1240 1.1 ad || (addr != NULL && ret == addr)); 1241 1.1 ad return (ret); 1242 1.1 ad } 1243 1.1 ad 1244 1.5 yamt static void * 1245 1.5 yamt pages_map(void *addr, size_t size) 1246 1.5 yamt { 1247 1.5 yamt 1248 1.5 yamt return pages_map_align(addr, size, 0); 1249 1.5 yamt } 1250 1.5 yamt 1251 1.1 ad static void 1252 1.1 ad pages_unmap(void *addr, size_t size) 1253 1.1 ad { 1254 1.1 ad 1255 1.1 ad if (munmap(addr, size) == -1) { 1256 1.1 ad char buf[STRERROR_BUF]; 1257 1.1 ad 1258 1.16 christos STRERROR_R(errno, buf, sizeof(buf)); 1259 1.16 christos _malloc_message(getprogname(), 1260 1.1 ad ": (malloc) Error in munmap(): ", buf, "\n"); 1261 1.57 ad if (OPT(abort)) 1262 1.1 ad abort(); 1263 1.1 ad } 1264 1.1 ad } 1265 1.1 ad 1266 1.1 ad static void * 1267 1.1 ad chunk_alloc(size_t size) 1268 1.1 ad { 1269 1.1 ad void *ret, *chunk; 1270 1.1 ad chunk_node_t *tchunk, *delchunk; 1271 1.1 ad 1272 1.1 ad assert(size != 0); 1273 1.1 ad assert((size & chunksize_mask) == 0); 1274 1.1 ad 1275 1.1 ad malloc_mutex_lock(&chunks_mtx); 1276 1.1 ad 1277 1.1 ad if (size == chunksize) { 1278 1.1 ad /* 1279 1.1 ad * Check for address ranges that were previously chunks and try 1280 1.1 ad * to use them. 1281 1.1 ad */ 1282 1.1 ad 1283 1.58 ad tchunk = RB_TREE_MIN(&old_chunks); 1284 1.1 ad while (tchunk != NULL) { 1285 1.1 ad /* Found an address range. Try to recycle it. */ 1286 1.1 ad 1287 1.1 ad chunk = tchunk->chunk; 1288 1.1 ad delchunk = tchunk; 1289 1.58 ad tchunk = RB_TREE_NEXT(&old_chunks, delchunk); 1290 1.1 ad 1291 1.1 ad /* Remove delchunk from the tree. */ 1292 1.58 ad rb_tree_remove_node(&old_chunks, delchunk); 1293 1.1 ad base_chunk_node_dealloc(delchunk); 1294 1.1 ad 1295 1.1 ad #ifdef USE_BRK 1296 1.1 ad if ((uintptr_t)chunk >= (uintptr_t)brk_base 1297 1.1 ad && (uintptr_t)chunk < (uintptr_t)brk_max) { 1298 1.1 ad /* Re-use a previously freed brk chunk. */ 1299 1.1 ad ret = chunk; 1300 1.1 ad goto RETURN; 1301 1.1 ad } 1302 1.1 ad #endif 1303 1.1 ad if ((ret = pages_map(chunk, size)) != NULL) { 1304 1.1 ad /* Success. */ 1305 1.1 ad goto RETURN; 1306 1.1 ad } 1307 1.1 ad } 1308 1.1 ad } 1309 1.1 ad 1310 1.1 ad /* 1311 1.1 ad * Try to over-allocate, but allow the OS to place the allocation 1312 1.1 ad * anywhere. Beware of size_t wrap-around. 1313 1.1 ad */ 1314 1.1 ad if (size + chunksize > size) { 1315 1.5 yamt if ((ret = pages_map_align(NULL, size, chunksize_2pow)) 1316 1.5 yamt != NULL) { 1317 1.1 ad goto RETURN; 1318 1.1 ad } 1319 1.1 ad } 1320 1.1 ad 1321 1.1 ad #ifdef USE_BRK 1322 1.1 ad /* 1323 1.1 ad * Try to create allocations in brk, in order to make full use of 1324 1.1 ad * limited address space. 1325 1.1 ad */ 1326 1.1 ad if (brk_prev != (void *)-1) { 1327 1.1 ad void *brk_cur; 1328 1.1 ad intptr_t incr; 1329 1.1 ad 1330 1.1 ad /* 1331 1.1 ad * The loop is necessary to recover from races with other 1332 1.1 ad * threads that are using brk for something other than malloc. 1333 1.1 ad */ 1334 1.1 ad malloc_mutex_lock(&brk_mtx); 1335 1.1 ad do { 1336 1.1 ad /* Get the current end of brk. */ 1337 1.1 ad brk_cur = sbrk(0); 1338 1.1 ad 1339 1.1 ad /* 1340 1.1 ad * Calculate how much padding is necessary to 1341 1.1 ad * chunk-align the end of brk. 1342 1.1 ad */ 1343 1.1 ad incr = (intptr_t)size 1344 1.1 ad - (intptr_t)CHUNK_ADDR2OFFSET(brk_cur); 1345 1.20 lukem if (incr == (intptr_t)size) { 1346 1.1 ad ret = brk_cur; 1347 1.1 ad } else { 1348 1.1 ad ret = (void *)((intptr_t)brk_cur + incr); 1349 1.1 ad incr += size; 1350 1.1 ad } 1351 1.1 ad 1352 1.1 ad brk_prev = sbrk(incr); 1353 1.1 ad if (brk_prev == brk_cur) { 1354 1.1 ad /* Success. */ 1355 1.1 ad malloc_mutex_unlock(&brk_mtx); 1356 1.1 ad brk_max = (void *)((intptr_t)ret + size); 1357 1.1 ad goto RETURN; 1358 1.1 ad } 1359 1.1 ad } while (brk_prev != (void *)-1); 1360 1.1 ad malloc_mutex_unlock(&brk_mtx); 1361 1.1 ad } 1362 1.1 ad #endif 1363 1.1 ad 1364 1.1 ad /* All strategies for allocation failed. */ 1365 1.1 ad ret = NULL; 1366 1.1 ad RETURN: 1367 1.1 ad if (ret != NULL) { 1368 1.1 ad chunk_node_t key; 1369 1.1 ad /* 1370 1.1 ad * Clean out any entries in old_chunks that overlap with the 1371 1.1 ad * memory we just allocated. 1372 1.1 ad */ 1373 1.1 ad key.chunk = ret; 1374 1.58 ad tchunk = rb_tree_find_node_geq(&old_chunks, &key); 1375 1.1 ad while (tchunk != NULL 1376 1.1 ad && (uintptr_t)tchunk->chunk >= (uintptr_t)ret 1377 1.1 ad && (uintptr_t)tchunk->chunk < (uintptr_t)ret + size) { 1378 1.1 ad delchunk = tchunk; 1379 1.58 ad tchunk = RB_TREE_NEXT(&old_chunks, delchunk); 1380 1.58 ad rb_tree_remove_node(&old_chunks, delchunk); 1381 1.1 ad base_chunk_node_dealloc(delchunk); 1382 1.1 ad } 1383 1.1 ad 1384 1.1 ad } 1385 1.1 ad #ifdef MALLOC_STATS 1386 1.1 ad if (ret != NULL) { 1387 1.1 ad stats_chunks.nchunks += (size / chunksize); 1388 1.1 ad stats_chunks.curchunks += (size / chunksize); 1389 1.1 ad } 1390 1.1 ad if (stats_chunks.curchunks > stats_chunks.highchunks) 1391 1.1 ad stats_chunks.highchunks = stats_chunks.curchunks; 1392 1.1 ad #endif 1393 1.1 ad malloc_mutex_unlock(&chunks_mtx); 1394 1.1 ad 1395 1.1 ad assert(CHUNK_ADDR2BASE(ret) == ret); 1396 1.1 ad return (ret); 1397 1.1 ad } 1398 1.1 ad 1399 1.1 ad static void 1400 1.1 ad chunk_dealloc(void *chunk, size_t size) 1401 1.1 ad { 1402 1.1 ad chunk_node_t *node; 1403 1.1 ad 1404 1.1 ad assert(chunk != NULL); 1405 1.1 ad assert(CHUNK_ADDR2BASE(chunk) == chunk); 1406 1.1 ad assert(size != 0); 1407 1.1 ad assert((size & chunksize_mask) == 0); 1408 1.1 ad 1409 1.1 ad malloc_mutex_lock(&chunks_mtx); 1410 1.1 ad 1411 1.1 ad #ifdef USE_BRK 1412 1.1 ad if ((uintptr_t)chunk >= (uintptr_t)brk_base 1413 1.1 ad && (uintptr_t)chunk < (uintptr_t)brk_max) { 1414 1.1 ad void *brk_cur; 1415 1.1 ad 1416 1.1 ad malloc_mutex_lock(&brk_mtx); 1417 1.1 ad /* Get the current end of brk. */ 1418 1.1 ad brk_cur = sbrk(0); 1419 1.1 ad 1420 1.1 ad /* 1421 1.1 ad * Try to shrink the data segment if this chunk is at the end 1422 1.1 ad * of the data segment. The sbrk() call here is subject to a 1423 1.1 ad * race condition with threads that use brk(2) or sbrk(2) 1424 1.1 ad * directly, but the alternative would be to leak memory for 1425 1.1 ad * the sake of poorly designed multi-threaded programs. 1426 1.1 ad */ 1427 1.1 ad if (brk_cur == brk_max 1428 1.1 ad && (void *)((uintptr_t)chunk + size) == brk_max 1429 1.1 ad && sbrk(-(intptr_t)size) == brk_max) { 1430 1.1 ad malloc_mutex_unlock(&brk_mtx); 1431 1.1 ad if (brk_prev == brk_max) { 1432 1.1 ad /* Success. */ 1433 1.1 ad brk_prev = (void *)((intptr_t)brk_max 1434 1.1 ad - (intptr_t)size); 1435 1.1 ad brk_max = brk_prev; 1436 1.1 ad } 1437 1.1 ad } else { 1438 1.1 ad size_t offset; 1439 1.1 ad 1440 1.1 ad malloc_mutex_unlock(&brk_mtx); 1441 1.1 ad madvise(chunk, size, MADV_FREE); 1442 1.1 ad 1443 1.1 ad /* 1444 1.1 ad * Iteratively create records of each chunk-sized 1445 1.1 ad * memory region that 'chunk' is comprised of, so that 1446 1.1 ad * the address range can be recycled if memory usage 1447 1.1 ad * increases later on. 1448 1.1 ad */ 1449 1.1 ad for (offset = 0; offset < size; offset += chunksize) { 1450 1.1 ad node = base_chunk_node_alloc(); 1451 1.1 ad if (node == NULL) 1452 1.1 ad break; 1453 1.1 ad 1454 1.1 ad node->chunk = (void *)((uintptr_t)chunk 1455 1.1 ad + (uintptr_t)offset); 1456 1.1 ad node->size = chunksize; 1457 1.58 ad rb_tree_insert_node(&old_chunks, node); 1458 1.1 ad } 1459 1.1 ad } 1460 1.1 ad } else { 1461 1.1 ad #endif 1462 1.1 ad pages_unmap(chunk, size); 1463 1.1 ad 1464 1.1 ad /* 1465 1.1 ad * Make a record of the chunk's address, so that the address 1466 1.1 ad * range can be recycled if memory usage increases later on. 1467 1.1 ad * Don't bother to create entries if (size > chunksize), since 1468 1.1 ad * doing so could cause scalability issues for truly gargantuan 1469 1.1 ad * objects (many gigabytes or larger). 1470 1.1 ad */ 1471 1.1 ad if (size == chunksize) { 1472 1.1 ad node = base_chunk_node_alloc(); 1473 1.1 ad if (node != NULL) { 1474 1.1 ad node->chunk = (void *)(uintptr_t)chunk; 1475 1.1 ad node->size = chunksize; 1476 1.58 ad rb_tree_insert_node(&old_chunks, node); 1477 1.1 ad } 1478 1.1 ad } 1479 1.1 ad #ifdef USE_BRK 1480 1.1 ad } 1481 1.1 ad #endif 1482 1.1 ad 1483 1.1 ad #ifdef MALLOC_STATS 1484 1.1 ad stats_chunks.curchunks -= (size / chunksize); 1485 1.1 ad #endif 1486 1.1 ad malloc_mutex_unlock(&chunks_mtx); 1487 1.1 ad } 1488 1.1 ad 1489 1.1 ad /* 1490 1.1 ad * End chunk management functions. 1491 1.1 ad */ 1492 1.1 ad /******************************************************************************/ 1493 1.1 ad /* 1494 1.1 ad * Begin arena. 1495 1.1 ad */ 1496 1.1 ad 1497 1.1 ad /* 1498 1.57 ad * Choose a per-CPU arena. 1499 1.1 ad */ 1500 1.19 ad static __noinline arena_t * 1501 1.19 ad choose_arena_hard(void) 1502 1.1 ad { 1503 1.1 ad 1504 1.57 ad assert(arenas[0] != NULL); 1505 1.57 ad 1506 1.1 ad malloc_mutex_lock(&arenas_mtx); 1507 1.57 ad for (unsigned i = 1; i < ncpus; i++) 1508 1.57 ad if (arenas[i] == NULL) 1509 1.57 ad arenas[i] = arenas_extend(); 1510 1.1 ad malloc_mutex_unlock(&arenas_mtx); 1511 1.1 ad 1512 1.57 ad return arenas[thr_curcpu()]; 1513 1.1 ad } 1514 1.1 ad 1515 1.19 ad static inline arena_t * 1516 1.19 ad choose_arena(void) 1517 1.19 ad { 1518 1.57 ad arena_t *arena; 1519 1.19 ad 1520 1.57 ad /* NB: when libpthread is absent, thr_curcpu() always returns zero. */ 1521 1.57 ad arena = arenas[thr_curcpu()]; 1522 1.57 ad if (__predict_true(arena != NULL)) 1523 1.57 ad return arena; 1524 1.19 ad 1525 1.57 ad return choose_arena_hard(); 1526 1.19 ad } 1527 1.19 ad 1528 1.58 ad static int 1529 1.58 ad arena_chunk_comp(void *context, const void *va, const void *vb) 1530 1.1 ad { 1531 1.58 ad const arena_chunk_t *a = va, *b = vb; 1532 1.58 ad int diff; 1533 1.1 ad 1534 1.1 ad assert(a != NULL); 1535 1.1 ad assert(b != NULL); 1536 1.1 ad 1537 1.58 ad if ((diff = a->max_frun_npages - b->max_frun_npages) != 0) 1538 1.58 ad return diff; 1539 1.58 ad return ptrcmp(a, b); 1540 1.1 ad } 1541 1.1 ad 1542 1.58 ad static int 1543 1.58 ad arena_run_comp(void *context, const void *a, const void *b) 1544 1.1 ad { 1545 1.1 ad 1546 1.1 ad assert(a != NULL); 1547 1.1 ad assert(b != NULL); 1548 1.1 ad 1549 1.58 ad return ptrcmp(a, b); 1550 1.1 ad } 1551 1.1 ad 1552 1.1 ad static inline void * 1553 1.1 ad arena_run_reg_alloc(arena_run_t *run, arena_bin_t *bin) 1554 1.1 ad { 1555 1.1 ad void *ret; 1556 1.1 ad unsigned i, mask, bit, regind; 1557 1.1 ad 1558 1.1 ad assert(run->magic == ARENA_RUN_MAGIC); 1559 1.1 ad assert(run->regs_minelm < bin->regs_mask_nelms); 1560 1.1 ad 1561 1.1 ad /* 1562 1.1 ad * Move the first check outside the loop, so that run->regs_minelm can 1563 1.1 ad * be updated unconditionally, without the possibility of updating it 1564 1.1 ad * multiple times. 1565 1.1 ad */ 1566 1.1 ad i = run->regs_minelm; 1567 1.1 ad mask = run->regs_mask[i]; 1568 1.1 ad if (mask != 0) { 1569 1.1 ad /* Usable allocation found. */ 1570 1.1 ad bit = ffs((int)mask) - 1; 1571 1.1 ad 1572 1.1 ad regind = ((i << (SIZEOF_INT_2POW + 3)) + bit); 1573 1.1 ad ret = (void *)(((uintptr_t)run) + bin->reg0_offset 1574 1.1 ad + (bin->reg_size * regind)); 1575 1.1 ad 1576 1.1 ad /* Clear bit. */ 1577 1.45 kamil mask ^= (1U << bit); 1578 1.1 ad run->regs_mask[i] = mask; 1579 1.1 ad 1580 1.1 ad return (ret); 1581 1.1 ad } 1582 1.1 ad 1583 1.1 ad for (i++; i < bin->regs_mask_nelms; i++) { 1584 1.1 ad mask = run->regs_mask[i]; 1585 1.1 ad if (mask != 0) { 1586 1.1 ad /* Usable allocation found. */ 1587 1.1 ad bit = ffs((int)mask) - 1; 1588 1.1 ad 1589 1.1 ad regind = ((i << (SIZEOF_INT_2POW + 3)) + bit); 1590 1.1 ad ret = (void *)(((uintptr_t)run) + bin->reg0_offset 1591 1.1 ad + (bin->reg_size * regind)); 1592 1.1 ad 1593 1.1 ad /* Clear bit. */ 1594 1.45 kamil mask ^= (1U << bit); 1595 1.1 ad run->regs_mask[i] = mask; 1596 1.1 ad 1597 1.1 ad /* 1598 1.1 ad * Make a note that nothing before this element 1599 1.1 ad * contains a free region. 1600 1.1 ad */ 1601 1.1 ad run->regs_minelm = i; /* Low payoff: + (mask == 0); */ 1602 1.1 ad 1603 1.1 ad return (ret); 1604 1.1 ad } 1605 1.1 ad } 1606 1.1 ad /* Not reached. */ 1607 1.7 yamt /* LINTED */ 1608 1.1 ad assert(0); 1609 1.1 ad return (NULL); 1610 1.1 ad } 1611 1.1 ad 1612 1.1 ad static inline void 1613 1.1 ad arena_run_reg_dalloc(arena_run_t *run, arena_bin_t *bin, void *ptr, size_t size) 1614 1.1 ad { 1615 1.1 ad /* 1616 1.1 ad * To divide by a number D that is not a power of two we multiply 1617 1.1 ad * by (2^21 / D) and then right shift by 21 positions. 1618 1.1 ad * 1619 1.1 ad * X / D 1620 1.1 ad * 1621 1.1 ad * becomes 1622 1.1 ad * 1623 1.1 ad * (X * size_invs[(D >> QUANTUM_2POW_MIN) - 3]) >> SIZE_INV_SHIFT 1624 1.1 ad */ 1625 1.1 ad #define SIZE_INV_SHIFT 21 1626 1.1 ad #define SIZE_INV(s) (((1 << SIZE_INV_SHIFT) / (s << QUANTUM_2POW_MIN)) + 1) 1627 1.1 ad static const unsigned size_invs[] = { 1628 1.1 ad SIZE_INV(3), 1629 1.1 ad SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7), 1630 1.1 ad SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11), 1631 1.1 ad SIZE_INV(12),SIZE_INV(13), SIZE_INV(14), SIZE_INV(15), 1632 1.1 ad SIZE_INV(16),SIZE_INV(17), SIZE_INV(18), SIZE_INV(19), 1633 1.1 ad SIZE_INV(20),SIZE_INV(21), SIZE_INV(22), SIZE_INV(23), 1634 1.1 ad SIZE_INV(24),SIZE_INV(25), SIZE_INV(26), SIZE_INV(27), 1635 1.1 ad SIZE_INV(28),SIZE_INV(29), SIZE_INV(30), SIZE_INV(31) 1636 1.1 ad #if (QUANTUM_2POW_MIN < 4) 1637 1.1 ad , 1638 1.1 ad SIZE_INV(32), SIZE_INV(33), SIZE_INV(34), SIZE_INV(35), 1639 1.1 ad SIZE_INV(36), SIZE_INV(37), SIZE_INV(38), SIZE_INV(39), 1640 1.1 ad SIZE_INV(40), SIZE_INV(41), SIZE_INV(42), SIZE_INV(43), 1641 1.1 ad SIZE_INV(44), SIZE_INV(45), SIZE_INV(46), SIZE_INV(47), 1642 1.1 ad SIZE_INV(48), SIZE_INV(49), SIZE_INV(50), SIZE_INV(51), 1643 1.1 ad SIZE_INV(52), SIZE_INV(53), SIZE_INV(54), SIZE_INV(55), 1644 1.1 ad SIZE_INV(56), SIZE_INV(57), SIZE_INV(58), SIZE_INV(59), 1645 1.1 ad SIZE_INV(60), SIZE_INV(61), SIZE_INV(62), SIZE_INV(63) 1646 1.1 ad #endif 1647 1.1 ad }; 1648 1.1 ad unsigned diff, regind, elm, bit; 1649 1.1 ad 1650 1.7 yamt /* LINTED */ 1651 1.1 ad assert(run->magic == ARENA_RUN_MAGIC); 1652 1.1 ad assert(((sizeof(size_invs)) / sizeof(unsigned)) + 3 1653 1.1 ad >= (SMALL_MAX_DEFAULT >> QUANTUM_2POW_MIN)); 1654 1.1 ad 1655 1.1 ad /* 1656 1.1 ad * Avoid doing division with a variable divisor if possible. Using 1657 1.1 ad * actual division here can reduce allocator throughput by over 20%! 1658 1.1 ad */ 1659 1.1 ad diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - bin->reg0_offset); 1660 1.1 ad if ((size & (size - 1)) == 0) { 1661 1.1 ad /* 1662 1.1 ad * log2_table allows fast division of a power of two in the 1663 1.1 ad * [1..128] range. 1664 1.1 ad * 1665 1.1 ad * (x / divisor) becomes (x >> log2_table[divisor - 1]). 1666 1.1 ad */ 1667 1.1 ad static const unsigned char log2_table[] = { 1668 1.1 ad 0, 1, 0, 2, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 4, 1669 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 1670 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1671 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6, 1672 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1673 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1674 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1675 1.1 ad 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7 1676 1.1 ad }; 1677 1.1 ad 1678 1.1 ad if (size <= 128) 1679 1.1 ad regind = (diff >> log2_table[size - 1]); 1680 1.1 ad else if (size <= 32768) 1681 1.1 ad regind = diff >> (8 + log2_table[(size >> 8) - 1]); 1682 1.1 ad else { 1683 1.1 ad /* 1684 1.1 ad * The page size is too large for us to use the lookup 1685 1.1 ad * table. Use real division. 1686 1.1 ad */ 1687 1.9 christos regind = (unsigned)(diff / size); 1688 1.1 ad } 1689 1.61 ad } else if (size <= (((sizeof(size_invs) / sizeof(unsigned)) + 2) 1690 1.61 ad << QUANTUM_2POW_MIN)) { 1691 1.1 ad regind = size_invs[(size >> QUANTUM_2POW_MIN) - 3] * diff; 1692 1.1 ad regind >>= SIZE_INV_SHIFT; 1693 1.1 ad } else { 1694 1.1 ad /* 1695 1.1 ad * size_invs isn't large enough to handle this size class, so 1696 1.1 ad * calculate regind using actual division. This only happens 1697 1.1 ad * if the user increases small_max via the 'S' runtime 1698 1.1 ad * configuration option. 1699 1.1 ad */ 1700 1.9 christos regind = (unsigned)(diff / size); 1701 1.1 ad }; 1702 1.1 ad assert(diff == regind * size); 1703 1.1 ad assert(regind < bin->nregs); 1704 1.1 ad 1705 1.1 ad elm = regind >> (SIZEOF_INT_2POW + 3); 1706 1.1 ad if (elm < run->regs_minelm) 1707 1.1 ad run->regs_minelm = elm; 1708 1.1 ad bit = regind - (elm << (SIZEOF_INT_2POW + 3)); 1709 1.45 kamil assert((run->regs_mask[elm] & (1U << bit)) == 0); 1710 1.45 kamil run->regs_mask[elm] |= (1U << bit); 1711 1.1 ad #undef SIZE_INV 1712 1.1 ad #undef SIZE_INV_SHIFT 1713 1.1 ad } 1714 1.1 ad 1715 1.1 ad static void 1716 1.1 ad arena_run_split(arena_t *arena, arena_run_t *run, size_t size) 1717 1.1 ad { 1718 1.1 ad arena_chunk_t *chunk; 1719 1.1 ad unsigned run_ind, map_offset, total_pages, need_pages, rem_pages; 1720 1.1 ad unsigned i; 1721 1.1 ad 1722 1.1 ad chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); 1723 1.1 ad run_ind = (unsigned)(((uintptr_t)run - (uintptr_t)chunk) 1724 1.1 ad >> pagesize_2pow); 1725 1.1 ad total_pages = chunk->map[run_ind].npages; 1726 1.9 christos need_pages = (unsigned)(size >> pagesize_2pow); 1727 1.1 ad assert(need_pages <= total_pages); 1728 1.1 ad rem_pages = total_pages - need_pages; 1729 1.1 ad 1730 1.1 ad /* Split enough pages from the front of run to fit allocation size. */ 1731 1.1 ad map_offset = run_ind; 1732 1.1 ad for (i = 0; i < need_pages; i++) { 1733 1.1 ad chunk->map[map_offset + i].npages = need_pages; 1734 1.1 ad chunk->map[map_offset + i].pos = i; 1735 1.1 ad } 1736 1.1 ad 1737 1.1 ad /* Keep track of trailing unused pages for later use. */ 1738 1.1 ad if (rem_pages > 0) { 1739 1.1 ad /* Update map for trailing pages. */ 1740 1.1 ad map_offset += need_pages; 1741 1.1 ad chunk->map[map_offset].npages = rem_pages; 1742 1.1 ad chunk->map[map_offset].pos = POS_FREE; 1743 1.1 ad chunk->map[map_offset + rem_pages - 1].npages = rem_pages; 1744 1.1 ad chunk->map[map_offset + rem_pages - 1].pos = POS_FREE; 1745 1.1 ad } 1746 1.1 ad 1747 1.1 ad chunk->pages_used += need_pages; 1748 1.1 ad } 1749 1.1 ad 1750 1.1 ad static arena_chunk_t * 1751 1.1 ad arena_chunk_alloc(arena_t *arena) 1752 1.1 ad { 1753 1.1 ad arena_chunk_t *chunk; 1754 1.1 ad 1755 1.1 ad if (arena->spare != NULL) { 1756 1.1 ad chunk = arena->spare; 1757 1.1 ad arena->spare = NULL; 1758 1.1 ad 1759 1.58 ad rb_tree_insert_node(&arena->chunks, chunk); 1760 1.1 ad } else { 1761 1.1 ad chunk = (arena_chunk_t *)chunk_alloc(chunksize); 1762 1.1 ad if (chunk == NULL) 1763 1.1 ad return (NULL); 1764 1.1 ad #ifdef MALLOC_STATS 1765 1.1 ad arena->stats.mapped += chunksize; 1766 1.1 ad #endif 1767 1.1 ad 1768 1.1 ad chunk->arena = arena; 1769 1.1 ad 1770 1.1 ad /* 1771 1.1 ad * Claim that no pages are in use, since the header is merely 1772 1.1 ad * overhead. 1773 1.1 ad */ 1774 1.1 ad chunk->pages_used = 0; 1775 1.1 ad 1776 1.1 ad chunk->max_frun_npages = chunk_npages - 1777 1.1 ad arena_chunk_header_npages; 1778 1.1 ad chunk->min_frun_ind = arena_chunk_header_npages; 1779 1.1 ad 1780 1.1 ad /* 1781 1.1 ad * Initialize enough of the map to support one maximal free run. 1782 1.1 ad */ 1783 1.1 ad chunk->map[arena_chunk_header_npages].npages = chunk_npages - 1784 1.1 ad arena_chunk_header_npages; 1785 1.1 ad chunk->map[arena_chunk_header_npages].pos = POS_FREE; 1786 1.1 ad chunk->map[chunk_npages - 1].npages = chunk_npages - 1787 1.1 ad arena_chunk_header_npages; 1788 1.1 ad chunk->map[chunk_npages - 1].pos = POS_FREE; 1789 1.40 joerg 1790 1.58 ad rb_tree_insert_node(&arena->chunks, chunk); 1791 1.1 ad } 1792 1.1 ad 1793 1.1 ad return (chunk); 1794 1.1 ad } 1795 1.1 ad 1796 1.1 ad static void 1797 1.1 ad arena_chunk_dealloc(arena_t *arena, arena_chunk_t *chunk) 1798 1.1 ad { 1799 1.1 ad 1800 1.1 ad /* 1801 1.1 ad * Remove chunk from the chunk tree, regardless of whether this chunk 1802 1.1 ad * will be cached, so that the arena does not use it. 1803 1.1 ad */ 1804 1.58 ad rb_tree_remove_node(&chunk->arena->chunks, chunk); 1805 1.1 ad 1806 1.57 ad if (NOT_OPT(hint)) { 1807 1.1 ad if (arena->spare != NULL) { 1808 1.1 ad chunk_dealloc((void *)arena->spare, chunksize); 1809 1.1 ad #ifdef MALLOC_STATS 1810 1.1 ad arena->stats.mapped -= chunksize; 1811 1.1 ad #endif 1812 1.1 ad } 1813 1.1 ad arena->spare = chunk; 1814 1.1 ad } else { 1815 1.1 ad assert(arena->spare == NULL); 1816 1.1 ad chunk_dealloc((void *)chunk, chunksize); 1817 1.1 ad #ifdef MALLOC_STATS 1818 1.1 ad arena->stats.mapped -= chunksize; 1819 1.1 ad #endif 1820 1.1 ad } 1821 1.1 ad } 1822 1.1 ad 1823 1.1 ad static arena_run_t * 1824 1.1 ad arena_run_alloc(arena_t *arena, size_t size) 1825 1.1 ad { 1826 1.58 ad arena_chunk_t *chunk; 1827 1.1 ad arena_run_t *run; 1828 1.40 joerg unsigned need_npages; 1829 1.1 ad 1830 1.1 ad assert(size <= (chunksize - (arena_chunk_header_npages << 1831 1.1 ad pagesize_2pow))); 1832 1.1 ad assert((size & pagesize_mask) == 0); 1833 1.1 ad 1834 1.1 ad /* 1835 1.40 joerg * Search through the arena chunk tree for a large enough free run. 1836 1.40 joerg * Tree order ensures that any exact fit is picked immediately or 1837 1.40 joerg * otherwise the lowest address of the next size. 1838 1.1 ad */ 1839 1.9 christos need_npages = (unsigned)(size >> pagesize_2pow); 1840 1.2 ad /* LINTED */ 1841 1.40 joerg for (;;) { 1842 1.58 ad rb_node_t *node = arena->chunks.rbt_root; 1843 1.40 joerg chunk = NULL; 1844 1.58 ad while (!RB_SENTINEL_P(node)) { 1845 1.58 ad assert(offsetof(struct arena_chunk_s, link) == 0); 1846 1.58 ad arena_chunk_t *chunk_tmp = (arena_chunk_t *)node; 1847 1.40 joerg if (chunk_tmp->max_frun_npages == need_npages) { 1848 1.40 joerg chunk = chunk_tmp; 1849 1.40 joerg break; 1850 1.40 joerg } 1851 1.40 joerg if (chunk_tmp->max_frun_npages < need_npages) { 1852 1.58 ad node = node->rb_nodes[1]; 1853 1.40 joerg continue; 1854 1.40 joerg } 1855 1.40 joerg chunk = chunk_tmp; 1856 1.58 ad node = node->rb_nodes[0]; 1857 1.40 joerg } 1858 1.40 joerg if (chunk == NULL) 1859 1.40 joerg break; 1860 1.1 ad /* 1861 1.40 joerg * At this point, the chunk must have a cached run size large 1862 1.40 joerg * enough to fit the allocation. 1863 1.1 ad */ 1864 1.40 joerg assert(need_npages <= chunk->max_frun_npages); 1865 1.40 joerg { 1866 1.1 ad arena_chunk_map_t *mapelm; 1867 1.1 ad unsigned i; 1868 1.1 ad unsigned max_frun_npages = 0; 1869 1.1 ad unsigned min_frun_ind = chunk_npages; 1870 1.1 ad 1871 1.1 ad assert(chunk->min_frun_ind >= 1872 1.1 ad arena_chunk_header_npages); 1873 1.1 ad for (i = chunk->min_frun_ind; i < chunk_npages;) { 1874 1.1 ad mapelm = &chunk->map[i]; 1875 1.1 ad if (mapelm->pos == POS_FREE) { 1876 1.1 ad if (mapelm->npages >= need_npages) { 1877 1.1 ad run = (arena_run_t *) 1878 1.1 ad ((uintptr_t)chunk + (i << 1879 1.1 ad pagesize_2pow)); 1880 1.1 ad /* Update page map. */ 1881 1.1 ad arena_run_split(arena, run, 1882 1.1 ad size); 1883 1.1 ad return (run); 1884 1.1 ad } 1885 1.1 ad if (mapelm->npages > 1886 1.1 ad max_frun_npages) { 1887 1.1 ad max_frun_npages = 1888 1.1 ad mapelm->npages; 1889 1.1 ad } 1890 1.1 ad if (i < min_frun_ind) { 1891 1.1 ad min_frun_ind = i; 1892 1.1 ad if (i < chunk->min_frun_ind) 1893 1.1 ad chunk->min_frun_ind = i; 1894 1.1 ad } 1895 1.1 ad } 1896 1.1 ad i += mapelm->npages; 1897 1.1 ad } 1898 1.1 ad /* 1899 1.1 ad * Search failure. Reset cached chunk->max_frun_npages. 1900 1.1 ad * chunk->min_frun_ind was already reset above (if 1901 1.1 ad * necessary). 1902 1.1 ad */ 1903 1.58 ad rb_tree_remove_node(&arena->chunks, chunk); 1904 1.1 ad chunk->max_frun_npages = max_frun_npages; 1905 1.58 ad rb_tree_insert_node(&arena->chunks, chunk); 1906 1.1 ad } 1907 1.1 ad } 1908 1.1 ad 1909 1.1 ad /* 1910 1.1 ad * No usable runs. Create a new chunk from which to allocate the run. 1911 1.1 ad */ 1912 1.1 ad chunk = arena_chunk_alloc(arena); 1913 1.1 ad if (chunk == NULL) 1914 1.1 ad return (NULL); 1915 1.1 ad run = (arena_run_t *)((uintptr_t)chunk + (arena_chunk_header_npages << 1916 1.1 ad pagesize_2pow)); 1917 1.1 ad /* Update page map. */ 1918 1.1 ad arena_run_split(arena, run, size); 1919 1.1 ad return (run); 1920 1.1 ad } 1921 1.1 ad 1922 1.1 ad static void 1923 1.1 ad arena_run_dalloc(arena_t *arena, arena_run_t *run, size_t size) 1924 1.1 ad { 1925 1.1 ad arena_chunk_t *chunk; 1926 1.1 ad unsigned run_ind, run_pages; 1927 1.1 ad 1928 1.1 ad chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); 1929 1.1 ad 1930 1.1 ad run_ind = (unsigned)(((uintptr_t)run - (uintptr_t)chunk) 1931 1.1 ad >> pagesize_2pow); 1932 1.1 ad assert(run_ind >= arena_chunk_header_npages); 1933 1.1 ad assert(run_ind < (chunksize >> pagesize_2pow)); 1934 1.9 christos run_pages = (unsigned)(size >> pagesize_2pow); 1935 1.1 ad assert(run_pages == chunk->map[run_ind].npages); 1936 1.1 ad 1937 1.1 ad /* Subtract pages from count of pages used in chunk. */ 1938 1.1 ad chunk->pages_used -= run_pages; 1939 1.1 ad 1940 1.1 ad /* Mark run as deallocated. */ 1941 1.1 ad assert(chunk->map[run_ind].npages == run_pages); 1942 1.1 ad chunk->map[run_ind].pos = POS_FREE; 1943 1.1 ad assert(chunk->map[run_ind + run_pages - 1].npages == run_pages); 1944 1.1 ad chunk->map[run_ind + run_pages - 1].pos = POS_FREE; 1945 1.1 ad 1946 1.1 ad /* 1947 1.1 ad * Tell the kernel that we don't need the data in this run, but only if 1948 1.1 ad * requested via runtime configuration. 1949 1.1 ad */ 1950 1.57 ad if (OPT(hint)) 1951 1.1 ad madvise(run, size, MADV_FREE); 1952 1.1 ad 1953 1.1 ad /* Try to coalesce with neighboring runs. */ 1954 1.1 ad if (run_ind > arena_chunk_header_npages && 1955 1.1 ad chunk->map[run_ind - 1].pos == POS_FREE) { 1956 1.1 ad unsigned prev_npages; 1957 1.1 ad 1958 1.1 ad /* Coalesce with previous run. */ 1959 1.1 ad prev_npages = chunk->map[run_ind - 1].npages; 1960 1.1 ad run_ind -= prev_npages; 1961 1.1 ad assert(chunk->map[run_ind].npages == prev_npages); 1962 1.1 ad assert(chunk->map[run_ind].pos == POS_FREE); 1963 1.1 ad run_pages += prev_npages; 1964 1.1 ad 1965 1.1 ad chunk->map[run_ind].npages = run_pages; 1966 1.1 ad assert(chunk->map[run_ind].pos == POS_FREE); 1967 1.1 ad chunk->map[run_ind + run_pages - 1].npages = run_pages; 1968 1.1 ad assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE); 1969 1.1 ad } 1970 1.1 ad 1971 1.1 ad if (run_ind + run_pages < chunk_npages && 1972 1.1 ad chunk->map[run_ind + run_pages].pos == POS_FREE) { 1973 1.1 ad unsigned next_npages; 1974 1.1 ad 1975 1.1 ad /* Coalesce with next run. */ 1976 1.1 ad next_npages = chunk->map[run_ind + run_pages].npages; 1977 1.1 ad run_pages += next_npages; 1978 1.1 ad assert(chunk->map[run_ind + run_pages - 1].npages == 1979 1.1 ad next_npages); 1980 1.1 ad assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE); 1981 1.1 ad 1982 1.1 ad chunk->map[run_ind].npages = run_pages; 1983 1.1 ad chunk->map[run_ind].pos = POS_FREE; 1984 1.1 ad chunk->map[run_ind + run_pages - 1].npages = run_pages; 1985 1.1 ad assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE); 1986 1.1 ad } 1987 1.1 ad 1988 1.40 joerg if (chunk->map[run_ind].npages > chunk->max_frun_npages) { 1989 1.58 ad rb_tree_remove_node(&arena->chunks, chunk); 1990 1.1 ad chunk->max_frun_npages = chunk->map[run_ind].npages; 1991 1.58 ad rb_tree_insert_node(&arena->chunks, chunk); 1992 1.40 joerg } 1993 1.1 ad if (run_ind < chunk->min_frun_ind) 1994 1.1 ad chunk->min_frun_ind = run_ind; 1995 1.1 ad 1996 1.1 ad /* Deallocate chunk if it is now completely unused. */ 1997 1.1 ad if (chunk->pages_used == 0) 1998 1.1 ad arena_chunk_dealloc(arena, chunk); 1999 1.1 ad } 2000 1.1 ad 2001 1.1 ad static arena_run_t * 2002 1.1 ad arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin) 2003 1.1 ad { 2004 1.1 ad arena_run_t *run; 2005 1.1 ad unsigned i, remainder; 2006 1.1 ad 2007 1.1 ad /* Look for a usable run. */ 2008 1.58 ad if ((run = RB_TREE_MIN(&bin->runs)) != NULL) { 2009 1.1 ad /* run is guaranteed to have available space. */ 2010 1.58 ad rb_tree_remove_node(&bin->runs, run); 2011 1.1 ad #ifdef MALLOC_STATS 2012 1.1 ad bin->stats.reruns++; 2013 1.1 ad #endif 2014 1.1 ad return (run); 2015 1.1 ad } 2016 1.1 ad /* No existing runs have any space available. */ 2017 1.1 ad 2018 1.1 ad /* Allocate a new run. */ 2019 1.1 ad run = arena_run_alloc(arena, bin->run_size); 2020 1.1 ad if (run == NULL) 2021 1.1 ad return (NULL); 2022 1.1 ad 2023 1.1 ad /* Initialize run internals. */ 2024 1.1 ad run->bin = bin; 2025 1.1 ad 2026 1.1 ad for (i = 0; i < bin->regs_mask_nelms; i++) 2027 1.1 ad run->regs_mask[i] = UINT_MAX; 2028 1.1 ad remainder = bin->nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1); 2029 1.1 ad if (remainder != 0) { 2030 1.1 ad /* The last element has spare bits that need to be unset. */ 2031 1.1 ad run->regs_mask[i] = (UINT_MAX >> ((1 << (SIZEOF_INT_2POW + 3)) 2032 1.1 ad - remainder)); 2033 1.1 ad } 2034 1.1 ad 2035 1.1 ad run->regs_minelm = 0; 2036 1.1 ad 2037 1.1 ad run->nfree = bin->nregs; 2038 1.1 ad #ifdef MALLOC_DEBUG 2039 1.1 ad run->magic = ARENA_RUN_MAGIC; 2040 1.1 ad #endif 2041 1.1 ad 2042 1.1 ad #ifdef MALLOC_STATS 2043 1.1 ad bin->stats.nruns++; 2044 1.1 ad bin->stats.curruns++; 2045 1.1 ad if (bin->stats.curruns > bin->stats.highruns) 2046 1.1 ad bin->stats.highruns = bin->stats.curruns; 2047 1.1 ad #endif 2048 1.1 ad return (run); 2049 1.1 ad } 2050 1.1 ad 2051 1.1 ad /* bin->runcur must have space available before this function is called. */ 2052 1.1 ad static inline void * 2053 1.1 ad arena_bin_malloc_easy(arena_t *arena, arena_bin_t *bin, arena_run_t *run) 2054 1.1 ad { 2055 1.1 ad void *ret; 2056 1.1 ad 2057 1.1 ad assert(run->magic == ARENA_RUN_MAGIC); 2058 1.1 ad assert(run->nfree > 0); 2059 1.1 ad 2060 1.1 ad ret = arena_run_reg_alloc(run, bin); 2061 1.1 ad assert(ret != NULL); 2062 1.1 ad run->nfree--; 2063 1.1 ad 2064 1.1 ad return (ret); 2065 1.1 ad } 2066 1.1 ad 2067 1.1 ad /* Re-fill bin->runcur, then call arena_bin_malloc_easy(). */ 2068 1.1 ad static void * 2069 1.1 ad arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin) 2070 1.1 ad { 2071 1.1 ad 2072 1.1 ad bin->runcur = arena_bin_nonfull_run_get(arena, bin); 2073 1.1 ad if (bin->runcur == NULL) 2074 1.1 ad return (NULL); 2075 1.1 ad assert(bin->runcur->magic == ARENA_RUN_MAGIC); 2076 1.1 ad assert(bin->runcur->nfree > 0); 2077 1.1 ad 2078 1.1 ad return (arena_bin_malloc_easy(arena, bin, bin->runcur)); 2079 1.1 ad } 2080 1.1 ad 2081 1.1 ad /* 2082 1.1 ad * Calculate bin->run_size such that it meets the following constraints: 2083 1.1 ad * 2084 1.1 ad * *) bin->run_size >= min_run_size 2085 1.1 ad * *) bin->run_size <= arena_maxclass 2086 1.1 ad * *) bin->run_size <= RUN_MAX_SMALL 2087 1.1 ad * *) run header overhead <= RUN_MAX_OVRHD (or header overhead relaxed). 2088 1.1 ad * 2089 1.1 ad * bin->nregs, bin->regs_mask_nelms, and bin->reg0_offset are 2090 1.1 ad * also calculated here, since these settings are all interdependent. 2091 1.1 ad */ 2092 1.1 ad static size_t 2093 1.1 ad arena_bin_run_size_calc(arena_bin_t *bin, size_t min_run_size) 2094 1.1 ad { 2095 1.1 ad size_t try_run_size, good_run_size; 2096 1.1 ad unsigned good_nregs, good_mask_nelms, good_reg0_offset; 2097 1.1 ad unsigned try_nregs, try_mask_nelms, try_reg0_offset; 2098 1.1 ad 2099 1.1 ad assert(min_run_size >= pagesize); 2100 1.1 ad assert(min_run_size <= arena_maxclass); 2101 1.1 ad assert(min_run_size <= RUN_MAX_SMALL); 2102 1.1 ad 2103 1.1 ad /* 2104 1.1 ad * Calculate known-valid settings before entering the run_size 2105 1.1 ad * expansion loop, so that the first part of the loop always copies 2106 1.1 ad * valid settings. 2107 1.1 ad * 2108 1.1 ad * The do..while loop iteratively reduces the number of regions until 2109 1.1 ad * the run header and the regions no longer overlap. A closed formula 2110 1.1 ad * would be quite messy, since there is an interdependency between the 2111 1.1 ad * header's mask length and the number of regions. 2112 1.1 ad */ 2113 1.1 ad try_run_size = min_run_size; 2114 1.9 christos try_nregs = (unsigned)(((try_run_size - sizeof(arena_run_t)) / 2115 1.22 njoly bin->reg_size) + 1); /* Counter-act try_nregs-- in loop. */ 2116 1.1 ad do { 2117 1.1 ad try_nregs--; 2118 1.1 ad try_mask_nelms = (try_nregs >> (SIZEOF_INT_2POW + 3)) + 2119 1.1 ad ((try_nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1)) ? 1 : 0); 2120 1.9 christos try_reg0_offset = (unsigned)(try_run_size - 2121 1.9 christos (try_nregs * bin->reg_size)); 2122 1.1 ad } while (sizeof(arena_run_t) + (sizeof(unsigned) * (try_mask_nelms - 1)) 2123 1.1 ad > try_reg0_offset); 2124 1.1 ad 2125 1.1 ad /* run_size expansion loop. */ 2126 1.1 ad do { 2127 1.1 ad /* 2128 1.1 ad * Copy valid settings before trying more aggressive settings. 2129 1.1 ad */ 2130 1.1 ad good_run_size = try_run_size; 2131 1.1 ad good_nregs = try_nregs; 2132 1.1 ad good_mask_nelms = try_mask_nelms; 2133 1.1 ad good_reg0_offset = try_reg0_offset; 2134 1.1 ad 2135 1.1 ad /* Try more aggressive settings. */ 2136 1.1 ad try_run_size += pagesize; 2137 1.9 christos try_nregs = (unsigned)(((try_run_size - sizeof(arena_run_t)) / 2138 1.9 christos bin->reg_size) + 1); /* Counter-act try_nregs-- in loop. */ 2139 1.1 ad do { 2140 1.1 ad try_nregs--; 2141 1.1 ad try_mask_nelms = (try_nregs >> (SIZEOF_INT_2POW + 3)) + 2142 1.1 ad ((try_nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1)) ? 2143 1.1 ad 1 : 0); 2144 1.9 christos try_reg0_offset = (unsigned)(try_run_size - (try_nregs * 2145 1.9 christos bin->reg_size)); 2146 1.1 ad } while (sizeof(arena_run_t) + (sizeof(unsigned) * 2147 1.1 ad (try_mask_nelms - 1)) > try_reg0_offset); 2148 1.1 ad } while (try_run_size <= arena_maxclass && try_run_size <= RUN_MAX_SMALL 2149 1.22 njoly && RUN_MAX_OVRHD * (bin->reg_size << 3) > RUN_MAX_OVRHD_RELAX 2150 1.22 njoly && (try_reg0_offset << RUN_BFP) > RUN_MAX_OVRHD * try_run_size); 2151 1.1 ad 2152 1.1 ad assert(sizeof(arena_run_t) + (sizeof(unsigned) * (good_mask_nelms - 1)) 2153 1.1 ad <= good_reg0_offset); 2154 1.1 ad assert((good_mask_nelms << (SIZEOF_INT_2POW + 3)) >= good_nregs); 2155 1.1 ad 2156 1.1 ad /* Copy final settings. */ 2157 1.1 ad bin->run_size = good_run_size; 2158 1.1 ad bin->nregs = good_nregs; 2159 1.1 ad bin->regs_mask_nelms = good_mask_nelms; 2160 1.1 ad bin->reg0_offset = good_reg0_offset; 2161 1.1 ad 2162 1.1 ad return (good_run_size); 2163 1.1 ad } 2164 1.1 ad 2165 1.1 ad static void * 2166 1.1 ad arena_malloc(arena_t *arena, size_t size) 2167 1.1 ad { 2168 1.1 ad void *ret; 2169 1.1 ad 2170 1.1 ad assert(arena != NULL); 2171 1.1 ad assert(arena->magic == ARENA_MAGIC); 2172 1.1 ad assert(size != 0); 2173 1.1 ad assert(QUANTUM_CEILING(size) <= arena_maxclass); 2174 1.1 ad 2175 1.1 ad if (size <= bin_maxclass) { 2176 1.1 ad arena_bin_t *bin; 2177 1.1 ad arena_run_t *run; 2178 1.1 ad 2179 1.1 ad /* Small allocation. */ 2180 1.1 ad 2181 1.1 ad if (size < small_min) { 2182 1.1 ad /* Tiny. */ 2183 1.1 ad size = pow2_ceil(size); 2184 1.1 ad bin = &arena->bins[ffs((int)(size >> (TINY_MIN_2POW + 2185 1.1 ad 1)))]; 2186 1.1 ad #if (!defined(NDEBUG) || defined(MALLOC_STATS)) 2187 1.1 ad /* 2188 1.1 ad * Bin calculation is always correct, but we may need 2189 1.1 ad * to fix size for the purposes of assertions and/or 2190 1.1 ad * stats accuracy. 2191 1.1 ad */ 2192 1.1 ad if (size < (1 << TINY_MIN_2POW)) 2193 1.1 ad size = (1 << TINY_MIN_2POW); 2194 1.1 ad #endif 2195 1.1 ad } else if (size <= small_max) { 2196 1.1 ad /* Quantum-spaced. */ 2197 1.1 ad size = QUANTUM_CEILING(size); 2198 1.1 ad bin = &arena->bins[ntbins + (size >> opt_quantum_2pow) 2199 1.1 ad - 1]; 2200 1.1 ad } else { 2201 1.1 ad /* Sub-page. */ 2202 1.1 ad size = pow2_ceil(size); 2203 1.1 ad bin = &arena->bins[ntbins + nqbins 2204 1.1 ad + (ffs((int)(size >> opt_small_max_2pow)) - 2)]; 2205 1.1 ad } 2206 1.1 ad assert(size == bin->reg_size); 2207 1.1 ad 2208 1.1 ad malloc_mutex_lock(&arena->mtx); 2209 1.1 ad if ((run = bin->runcur) != NULL && run->nfree > 0) 2210 1.1 ad ret = arena_bin_malloc_easy(arena, bin, run); 2211 1.1 ad else 2212 1.1 ad ret = arena_bin_malloc_hard(arena, bin); 2213 1.1 ad 2214 1.1 ad if (ret == NULL) { 2215 1.1 ad malloc_mutex_unlock(&arena->mtx); 2216 1.1 ad return (NULL); 2217 1.1 ad } 2218 1.1 ad 2219 1.1 ad #ifdef MALLOC_STATS 2220 1.1 ad bin->stats.nrequests++; 2221 1.1 ad arena->stats.nmalloc_small++; 2222 1.1 ad arena->stats.allocated_small += size; 2223 1.1 ad #endif 2224 1.1 ad } else { 2225 1.1 ad /* Large allocation. */ 2226 1.1 ad size = PAGE_CEILING(size); 2227 1.1 ad malloc_mutex_lock(&arena->mtx); 2228 1.1 ad ret = (void *)arena_run_alloc(arena, size); 2229 1.1 ad if (ret == NULL) { 2230 1.1 ad malloc_mutex_unlock(&arena->mtx); 2231 1.1 ad return (NULL); 2232 1.1 ad } 2233 1.1 ad #ifdef MALLOC_STATS 2234 1.1 ad arena->stats.nmalloc_large++; 2235 1.1 ad arena->stats.allocated_large += size; 2236 1.1 ad #endif 2237 1.1 ad } 2238 1.1 ad 2239 1.1 ad malloc_mutex_unlock(&arena->mtx); 2240 1.1 ad 2241 1.57 ad if (OPT(junk)) 2242 1.1 ad memset(ret, 0xa5, size); 2243 1.57 ad else if (OPT(zero)) 2244 1.1 ad memset(ret, 0, size); 2245 1.1 ad return (ret); 2246 1.1 ad } 2247 1.1 ad 2248 1.1 ad static inline void 2249 1.1 ad arena_palloc_trim(arena_t *arena, arena_chunk_t *chunk, unsigned pageind, 2250 1.1 ad unsigned npages) 2251 1.1 ad { 2252 1.1 ad unsigned i; 2253 1.1 ad 2254 1.1 ad assert(npages > 0); 2255 1.1 ad 2256 1.1 ad /* 2257 1.1 ad * Modifiy the map such that arena_run_dalloc() sees the run as 2258 1.1 ad * separately allocated. 2259 1.1 ad */ 2260 1.1 ad for (i = 0; i < npages; i++) { 2261 1.1 ad chunk->map[pageind + i].npages = npages; 2262 1.1 ad chunk->map[pageind + i].pos = i; 2263 1.1 ad } 2264 1.1 ad arena_run_dalloc(arena, (arena_run_t *)((uintptr_t)chunk + (pageind << 2265 1.1 ad pagesize_2pow)), npages << pagesize_2pow); 2266 1.1 ad } 2267 1.1 ad 2268 1.1 ad /* Only handles large allocations that require more than page alignment. */ 2269 1.1 ad static void * 2270 1.1 ad arena_palloc(arena_t *arena, size_t alignment, size_t size, size_t alloc_size) 2271 1.1 ad { 2272 1.1 ad void *ret; 2273 1.1 ad size_t offset; 2274 1.1 ad arena_chunk_t *chunk; 2275 1.1 ad unsigned pageind, i, npages; 2276 1.1 ad 2277 1.1 ad assert((size & pagesize_mask) == 0); 2278 1.1 ad assert((alignment & pagesize_mask) == 0); 2279 1.1 ad 2280 1.9 christos npages = (unsigned)(size >> pagesize_2pow); 2281 1.1 ad 2282 1.1 ad malloc_mutex_lock(&arena->mtx); 2283 1.1 ad ret = (void *)arena_run_alloc(arena, alloc_size); 2284 1.1 ad if (ret == NULL) { 2285 1.1 ad malloc_mutex_unlock(&arena->mtx); 2286 1.1 ad return (NULL); 2287 1.1 ad } 2288 1.1 ad 2289 1.1 ad chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ret); 2290 1.1 ad 2291 1.1 ad offset = (uintptr_t)ret & (alignment - 1); 2292 1.1 ad assert((offset & pagesize_mask) == 0); 2293 1.1 ad assert(offset < alloc_size); 2294 1.1 ad if (offset == 0) { 2295 1.9 christos pageind = (unsigned)(((uintptr_t)ret - (uintptr_t)chunk) >> 2296 1.1 ad pagesize_2pow); 2297 1.1 ad 2298 1.1 ad /* Update the map for the run to be kept. */ 2299 1.1 ad for (i = 0; i < npages; i++) { 2300 1.1 ad chunk->map[pageind + i].npages = npages; 2301 1.1 ad assert(chunk->map[pageind + i].pos == i); 2302 1.1 ad } 2303 1.1 ad 2304 1.1 ad /* Trim trailing space. */ 2305 1.1 ad arena_palloc_trim(arena, chunk, pageind + npages, 2306 1.9 christos (unsigned)((alloc_size - size) >> pagesize_2pow)); 2307 1.1 ad } else { 2308 1.1 ad size_t leadsize, trailsize; 2309 1.1 ad 2310 1.1 ad leadsize = alignment - offset; 2311 1.1 ad ret = (void *)((uintptr_t)ret + leadsize); 2312 1.9 christos pageind = (unsigned)(((uintptr_t)ret - (uintptr_t)chunk) >> 2313 1.1 ad pagesize_2pow); 2314 1.1 ad 2315 1.1 ad /* Update the map for the run to be kept. */ 2316 1.1 ad for (i = 0; i < npages; i++) { 2317 1.1 ad chunk->map[pageind + i].npages = npages; 2318 1.1 ad chunk->map[pageind + i].pos = i; 2319 1.1 ad } 2320 1.1 ad 2321 1.1 ad /* Trim leading space. */ 2322 1.9 christos arena_palloc_trim(arena, chunk, 2323 1.9 christos (unsigned)(pageind - (leadsize >> pagesize_2pow)), 2324 1.9 christos (unsigned)(leadsize >> pagesize_2pow)); 2325 1.1 ad 2326 1.1 ad trailsize = alloc_size - leadsize - size; 2327 1.1 ad if (trailsize != 0) { 2328 1.1 ad /* Trim trailing space. */ 2329 1.1 ad assert(trailsize < alloc_size); 2330 1.1 ad arena_palloc_trim(arena, chunk, pageind + npages, 2331 1.9 christos (unsigned)(trailsize >> pagesize_2pow)); 2332 1.1 ad } 2333 1.1 ad } 2334 1.1 ad 2335 1.1 ad #ifdef MALLOC_STATS 2336 1.1 ad arena->stats.nmalloc_large++; 2337 1.1 ad arena->stats.allocated_large += size; 2338 1.1 ad #endif 2339 1.1 ad malloc_mutex_unlock(&arena->mtx); 2340 1.1 ad 2341 1.57 ad if (OPT(junk)) 2342 1.1 ad memset(ret, 0xa5, size); 2343 1.57 ad else if (OPT(zero)) 2344 1.1 ad memset(ret, 0, size); 2345 1.1 ad return (ret); 2346 1.1 ad } 2347 1.1 ad 2348 1.1 ad /* Return the size of the allocation pointed to by ptr. */ 2349 1.1 ad static size_t 2350 1.1 ad arena_salloc(const void *ptr) 2351 1.1 ad { 2352 1.1 ad size_t ret; 2353 1.1 ad arena_chunk_t *chunk; 2354 1.1 ad arena_chunk_map_t *mapelm; 2355 1.1 ad unsigned pageind; 2356 1.1 ad 2357 1.1 ad assert(ptr != NULL); 2358 1.1 ad assert(CHUNK_ADDR2BASE(ptr) != ptr); 2359 1.1 ad 2360 1.1 ad /* 2361 1.1 ad * No arena data structures that we query here can change in a way that 2362 1.1 ad * affects this function, so we don't need to lock. 2363 1.1 ad */ 2364 1.1 ad chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 2365 1.9 christos pageind = (unsigned)(((uintptr_t)ptr - (uintptr_t)chunk) >> 2366 1.9 christos pagesize_2pow); 2367 1.1 ad mapelm = &chunk->map[pageind]; 2368 1.10 simonb if (mapelm->pos != 0 || ptr != (char *)((uintptr_t)chunk) + (pageind << 2369 1.10 simonb pagesize_2pow)) { 2370 1.10 simonb arena_run_t *run; 2371 1.1 ad 2372 1.1 ad pageind -= mapelm->pos; 2373 1.1 ad 2374 1.10 simonb run = (arena_run_t *)((uintptr_t)chunk + (pageind << 2375 1.10 simonb pagesize_2pow)); 2376 1.1 ad assert(run->magic == ARENA_RUN_MAGIC); 2377 1.1 ad ret = run->bin->reg_size; 2378 1.1 ad } else 2379 1.1 ad ret = mapelm->npages << pagesize_2pow; 2380 1.1 ad 2381 1.1 ad return (ret); 2382 1.1 ad } 2383 1.1 ad 2384 1.1 ad static void * 2385 1.1 ad arena_ralloc(void *ptr, size_t size, size_t oldsize) 2386 1.1 ad { 2387 1.1 ad void *ret; 2388 1.1 ad 2389 1.1 ad /* Avoid moving the allocation if the size class would not change. */ 2390 1.1 ad if (size < small_min) { 2391 1.1 ad if (oldsize < small_min && 2392 1.1 ad ffs((int)(pow2_ceil(size) >> (TINY_MIN_2POW + 1))) 2393 1.1 ad == ffs((int)(pow2_ceil(oldsize) >> (TINY_MIN_2POW + 1)))) 2394 1.1 ad goto IN_PLACE; 2395 1.1 ad } else if (size <= small_max) { 2396 1.1 ad if (oldsize >= small_min && oldsize <= small_max && 2397 1.1 ad (QUANTUM_CEILING(size) >> opt_quantum_2pow) 2398 1.1 ad == (QUANTUM_CEILING(oldsize) >> opt_quantum_2pow)) 2399 1.1 ad goto IN_PLACE; 2400 1.1 ad } else { 2401 1.1 ad /* 2402 1.1 ad * We make no attempt to resize runs here, though it would be 2403 1.1 ad * possible to do so. 2404 1.1 ad */ 2405 1.1 ad if (oldsize > small_max && PAGE_CEILING(size) == oldsize) 2406 1.1 ad goto IN_PLACE; 2407 1.1 ad } 2408 1.1 ad 2409 1.1 ad /* 2410 1.1 ad * If we get here, then size and oldsize are different enough that we 2411 1.1 ad * need to use a different size class. In that case, fall back to 2412 1.1 ad * allocating new space and copying. 2413 1.1 ad */ 2414 1.1 ad ret = arena_malloc(choose_arena(), size); 2415 1.1 ad if (ret == NULL) 2416 1.1 ad return (NULL); 2417 1.1 ad 2418 1.1 ad /* Junk/zero-filling were already done by arena_malloc(). */ 2419 1.1 ad if (size < oldsize) 2420 1.1 ad memcpy(ret, ptr, size); 2421 1.1 ad else 2422 1.1 ad memcpy(ret, ptr, oldsize); 2423 1.1 ad idalloc(ptr); 2424 1.1 ad return (ret); 2425 1.1 ad IN_PLACE: 2426 1.57 ad if (OPT(junk) && size < oldsize) 2427 1.1 ad memset((void *)((uintptr_t)ptr + size), 0x5a, oldsize - size); 2428 1.57 ad else if (OPT(zero) && size > oldsize) 2429 1.1 ad memset((void *)((uintptr_t)ptr + oldsize), 0, size - oldsize); 2430 1.1 ad return (ptr); 2431 1.1 ad } 2432 1.1 ad 2433 1.1 ad static void 2434 1.1 ad arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr) 2435 1.1 ad { 2436 1.1 ad unsigned pageind; 2437 1.1 ad arena_chunk_map_t *mapelm; 2438 1.1 ad size_t size; 2439 1.1 ad 2440 1.1 ad assert(arena != NULL); 2441 1.1 ad assert(arena->magic == ARENA_MAGIC); 2442 1.1 ad assert(chunk->arena == arena); 2443 1.1 ad assert(ptr != NULL); 2444 1.1 ad assert(CHUNK_ADDR2BASE(ptr) != ptr); 2445 1.1 ad 2446 1.9 christos pageind = (unsigned)(((uintptr_t)ptr - (uintptr_t)chunk) >> 2447 1.9 christos pagesize_2pow); 2448 1.1 ad mapelm = &chunk->map[pageind]; 2449 1.10 simonb if (mapelm->pos != 0 || ptr != (char *)((uintptr_t)chunk) + (pageind << 2450 1.10 simonb pagesize_2pow)) { 2451 1.10 simonb arena_run_t *run; 2452 1.1 ad arena_bin_t *bin; 2453 1.1 ad 2454 1.1 ad /* Small allocation. */ 2455 1.1 ad 2456 1.1 ad pageind -= mapelm->pos; 2457 1.1 ad 2458 1.10 simonb run = (arena_run_t *)((uintptr_t)chunk + (pageind << 2459 1.10 simonb pagesize_2pow)); 2460 1.1 ad assert(run->magic == ARENA_RUN_MAGIC); 2461 1.1 ad bin = run->bin; 2462 1.1 ad size = bin->reg_size; 2463 1.1 ad 2464 1.57 ad if (OPT(junk)) 2465 1.1 ad memset(ptr, 0x5a, size); 2466 1.1 ad 2467 1.1 ad malloc_mutex_lock(&arena->mtx); 2468 1.1 ad arena_run_reg_dalloc(run, bin, ptr, size); 2469 1.1 ad run->nfree++; 2470 1.1 ad 2471 1.1 ad if (run->nfree == bin->nregs) { 2472 1.1 ad /* Deallocate run. */ 2473 1.1 ad if (run == bin->runcur) 2474 1.1 ad bin->runcur = NULL; 2475 1.1 ad else if (bin->nregs != 1) { 2476 1.1 ad /* 2477 1.1 ad * This block's conditional is necessary because 2478 1.1 ad * if the run only contains one region, then it 2479 1.1 ad * never gets inserted into the non-full runs 2480 1.1 ad * tree. 2481 1.1 ad */ 2482 1.58 ad rb_tree_remove_node(&bin->runs, run); 2483 1.1 ad } 2484 1.1 ad #ifdef MALLOC_DEBUG 2485 1.1 ad run->magic = 0; 2486 1.1 ad #endif 2487 1.1 ad arena_run_dalloc(arena, run, bin->run_size); 2488 1.1 ad #ifdef MALLOC_STATS 2489 1.1 ad bin->stats.curruns--; 2490 1.1 ad #endif 2491 1.1 ad } else if (run->nfree == 1 && run != bin->runcur) { 2492 1.1 ad /* 2493 1.1 ad * Make sure that bin->runcur always refers to the 2494 1.1 ad * lowest non-full run, if one exists. 2495 1.1 ad */ 2496 1.1 ad if (bin->runcur == NULL) 2497 1.1 ad bin->runcur = run; 2498 1.1 ad else if ((uintptr_t)run < (uintptr_t)bin->runcur) { 2499 1.1 ad /* Switch runcur. */ 2500 1.1 ad if (bin->runcur->nfree > 0) { 2501 1.1 ad /* Insert runcur. */ 2502 1.58 ad rb_tree_insert_node(&bin->runs, bin->runcur); 2503 1.1 ad } 2504 1.1 ad bin->runcur = run; 2505 1.2 ad } else { 2506 1.58 ad rb_tree_insert_node(&bin->runs, run); 2507 1.2 ad } 2508 1.1 ad } 2509 1.1 ad #ifdef MALLOC_STATS 2510 1.1 ad arena->stats.allocated_small -= size; 2511 1.1 ad arena->stats.ndalloc_small++; 2512 1.1 ad #endif 2513 1.1 ad } else { 2514 1.1 ad /* Large allocation. */ 2515 1.1 ad 2516 1.1 ad size = mapelm->npages << pagesize_2pow; 2517 1.1 ad assert((((uintptr_t)ptr) & pagesize_mask) == 0); 2518 1.1 ad 2519 1.57 ad if (OPT(junk)) 2520 1.1 ad memset(ptr, 0x5a, size); 2521 1.1 ad 2522 1.1 ad malloc_mutex_lock(&arena->mtx); 2523 1.1 ad arena_run_dalloc(arena, (arena_run_t *)ptr, size); 2524 1.1 ad #ifdef MALLOC_STATS 2525 1.1 ad arena->stats.allocated_large -= size; 2526 1.1 ad arena->stats.ndalloc_large++; 2527 1.1 ad #endif 2528 1.1 ad } 2529 1.1 ad 2530 1.1 ad malloc_mutex_unlock(&arena->mtx); 2531 1.1 ad } 2532 1.1 ad 2533 1.57 ad static void 2534 1.1 ad arena_new(arena_t *arena) 2535 1.1 ad { 2536 1.1 ad unsigned i; 2537 1.1 ad arena_bin_t *bin; 2538 1.2 ad size_t prev_run_size; 2539 1.1 ad 2540 1.1 ad malloc_mutex_init(&arena->mtx); 2541 1.1 ad 2542 1.1 ad #ifdef MALLOC_STATS 2543 1.1 ad memset(&arena->stats, 0, sizeof(arena_stats_t)); 2544 1.1 ad #endif 2545 1.1 ad 2546 1.1 ad /* Initialize chunks. */ 2547 1.58 ad rb_tree_init(&arena->chunks, &arena_chunk_tree_ops); 2548 1.1 ad arena->spare = NULL; 2549 1.1 ad 2550 1.1 ad /* Initialize bins. */ 2551 1.1 ad prev_run_size = pagesize; 2552 1.1 ad 2553 1.1 ad /* (2^n)-spaced tiny bins. */ 2554 1.1 ad for (i = 0; i < ntbins; i++) { 2555 1.1 ad bin = &arena->bins[i]; 2556 1.1 ad bin->runcur = NULL; 2557 1.58 ad rb_tree_init(&bin->runs, &arena_run_tree_ops); 2558 1.1 ad 2559 1.1 ad bin->reg_size = (1 << (TINY_MIN_2POW + i)); 2560 1.1 ad prev_run_size = arena_bin_run_size_calc(bin, prev_run_size); 2561 1.1 ad 2562 1.1 ad #ifdef MALLOC_STATS 2563 1.1 ad memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); 2564 1.1 ad #endif 2565 1.1 ad } 2566 1.1 ad 2567 1.1 ad /* Quantum-spaced bins. */ 2568 1.1 ad for (; i < ntbins + nqbins; i++) { 2569 1.1 ad bin = &arena->bins[i]; 2570 1.1 ad bin->runcur = NULL; 2571 1.58 ad rb_tree_init(&bin->runs, &arena_run_tree_ops); 2572 1.1 ad 2573 1.1 ad bin->reg_size = quantum * (i - ntbins + 1); 2574 1.2 ad /* 2575 1.1 ad pow2_size = pow2_ceil(quantum * (i - ntbins + 1)); 2576 1.2 ad */ 2577 1.1 ad prev_run_size = arena_bin_run_size_calc(bin, prev_run_size); 2578 1.1 ad 2579 1.1 ad #ifdef MALLOC_STATS 2580 1.1 ad memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); 2581 1.1 ad #endif 2582 1.1 ad } 2583 1.1 ad 2584 1.1 ad /* (2^n)-spaced sub-page bins. */ 2585 1.1 ad for (; i < ntbins + nqbins + nsbins; i++) { 2586 1.1 ad bin = &arena->bins[i]; 2587 1.1 ad bin->runcur = NULL; 2588 1.58 ad rb_tree_init(&bin->runs, &arena_run_tree_ops); 2589 1.1 ad 2590 1.1 ad bin->reg_size = (small_max << (i - (ntbins + nqbins) + 1)); 2591 1.1 ad 2592 1.1 ad prev_run_size = arena_bin_run_size_calc(bin, prev_run_size); 2593 1.1 ad 2594 1.1 ad #ifdef MALLOC_STATS 2595 1.1 ad memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); 2596 1.1 ad #endif 2597 1.1 ad } 2598 1.1 ad 2599 1.1 ad #ifdef MALLOC_DEBUG 2600 1.1 ad arena->magic = ARENA_MAGIC; 2601 1.1 ad #endif 2602 1.1 ad } 2603 1.1 ad 2604 1.1 ad /* Create a new arena and insert it into the arenas array at index ind. */ 2605 1.1 ad static arena_t * 2606 1.57 ad arenas_extend(void) 2607 1.1 ad { 2608 1.1 ad arena_t *ret; 2609 1.1 ad 2610 1.1 ad /* Allocate enough space for trailing bins. */ 2611 1.1 ad ret = (arena_t *)base_alloc(sizeof(arena_t) 2612 1.1 ad + (sizeof(arena_bin_t) * (ntbins + nqbins + nsbins - 1))); 2613 1.57 ad if (ret != NULL) { 2614 1.57 ad arena_new(ret); 2615 1.1 ad return (ret); 2616 1.1 ad } 2617 1.1 ad /* Only reached if there is an OOM error. */ 2618 1.1 ad 2619 1.1 ad /* 2620 1.1 ad * OOM here is quite inconvenient to propagate, since dealing with it 2621 1.1 ad * would require a check for failure in the fast path. Instead, punt 2622 1.1 ad * by using arenas[0]. In practice, this is an extremely unlikely 2623 1.1 ad * failure. 2624 1.1 ad */ 2625 1.16 christos _malloc_message(getprogname(), 2626 1.1 ad ": (malloc) Error initializing arena\n", "", ""); 2627 1.57 ad if (OPT(abort)) 2628 1.1 ad abort(); 2629 1.1 ad 2630 1.1 ad return (arenas[0]); 2631 1.1 ad } 2632 1.1 ad 2633 1.1 ad /* 2634 1.1 ad * End arena. 2635 1.1 ad */ 2636 1.1 ad /******************************************************************************/ 2637 1.1 ad /* 2638 1.1 ad * Begin general internal functions. 2639 1.1 ad */ 2640 1.1 ad 2641 1.1 ad static void * 2642 1.1 ad huge_malloc(size_t size) 2643 1.1 ad { 2644 1.1 ad void *ret; 2645 1.1 ad size_t csize; 2646 1.1 ad chunk_node_t *node; 2647 1.1 ad 2648 1.1 ad /* Allocate one or more contiguous chunks for this request. */ 2649 1.1 ad 2650 1.1 ad csize = CHUNK_CEILING(size); 2651 1.1 ad if (csize == 0) { 2652 1.1 ad /* size is large enough to cause size_t wrap-around. */ 2653 1.1 ad return (NULL); 2654 1.1 ad } 2655 1.1 ad 2656 1.1 ad /* Allocate a chunk node with which to track the chunk. */ 2657 1.1 ad node = base_chunk_node_alloc(); 2658 1.1 ad if (node == NULL) 2659 1.1 ad return (NULL); 2660 1.1 ad 2661 1.1 ad ret = chunk_alloc(csize); 2662 1.1 ad if (ret == NULL) { 2663 1.1 ad base_chunk_node_dealloc(node); 2664 1.1 ad return (NULL); 2665 1.1 ad } 2666 1.1 ad 2667 1.1 ad /* Insert node into huge. */ 2668 1.1 ad node->chunk = ret; 2669 1.1 ad node->size = csize; 2670 1.1 ad 2671 1.1 ad malloc_mutex_lock(&chunks_mtx); 2672 1.58 ad rb_tree_insert_node(&huge, node); 2673 1.1 ad #ifdef MALLOC_STATS 2674 1.1 ad huge_nmalloc++; 2675 1.1 ad huge_allocated += csize; 2676 1.1 ad #endif 2677 1.1 ad malloc_mutex_unlock(&chunks_mtx); 2678 1.1 ad 2679 1.57 ad if (OPT(junk)) 2680 1.1 ad memset(ret, 0xa5, csize); 2681 1.57 ad else if (OPT(zero)) 2682 1.1 ad memset(ret, 0, csize); 2683 1.1 ad 2684 1.1 ad return (ret); 2685 1.1 ad } 2686 1.1 ad 2687 1.1 ad /* Only handles large allocations that require more than chunk alignment. */ 2688 1.1 ad static void * 2689 1.1 ad huge_palloc(size_t alignment, size_t size) 2690 1.1 ad { 2691 1.1 ad void *ret; 2692 1.1 ad size_t alloc_size, chunk_size, offset; 2693 1.1 ad chunk_node_t *node; 2694 1.1 ad 2695 1.1 ad /* 2696 1.1 ad * This allocation requires alignment that is even larger than chunk 2697 1.1 ad * alignment. This means that huge_malloc() isn't good enough. 2698 1.1 ad * 2699 1.1 ad * Allocate almost twice as many chunks as are demanded by the size or 2700 1.1 ad * alignment, in order to assure the alignment can be achieved, then 2701 1.1 ad * unmap leading and trailing chunks. 2702 1.1 ad */ 2703 1.1 ad assert(alignment >= chunksize); 2704 1.1 ad 2705 1.1 ad chunk_size = CHUNK_CEILING(size); 2706 1.1 ad 2707 1.1 ad if (size >= alignment) 2708 1.1 ad alloc_size = chunk_size + alignment - chunksize; 2709 1.1 ad else 2710 1.1 ad alloc_size = (alignment << 1) - chunksize; 2711 1.1 ad 2712 1.1 ad /* Allocate a chunk node with which to track the chunk. */ 2713 1.1 ad node = base_chunk_node_alloc(); 2714 1.1 ad if (node == NULL) 2715 1.1 ad return (NULL); 2716 1.1 ad 2717 1.1 ad ret = chunk_alloc(alloc_size); 2718 1.1 ad if (ret == NULL) { 2719 1.1 ad base_chunk_node_dealloc(node); 2720 1.1 ad return (NULL); 2721 1.1 ad } 2722 1.1 ad 2723 1.1 ad offset = (uintptr_t)ret & (alignment - 1); 2724 1.1 ad assert((offset & chunksize_mask) == 0); 2725 1.1 ad assert(offset < alloc_size); 2726 1.1 ad if (offset == 0) { 2727 1.1 ad /* Trim trailing space. */ 2728 1.1 ad chunk_dealloc((void *)((uintptr_t)ret + chunk_size), alloc_size 2729 1.1 ad - chunk_size); 2730 1.1 ad } else { 2731 1.1 ad size_t trailsize; 2732 1.1 ad 2733 1.1 ad /* Trim leading space. */ 2734 1.1 ad chunk_dealloc(ret, alignment - offset); 2735 1.1 ad 2736 1.1 ad ret = (void *)((uintptr_t)ret + (alignment - offset)); 2737 1.1 ad 2738 1.1 ad trailsize = alloc_size - (alignment - offset) - chunk_size; 2739 1.1 ad if (trailsize != 0) { 2740 1.1 ad /* Trim trailing space. */ 2741 1.1 ad assert(trailsize < alloc_size); 2742 1.1 ad chunk_dealloc((void *)((uintptr_t)ret + chunk_size), 2743 1.1 ad trailsize); 2744 1.1 ad } 2745 1.1 ad } 2746 1.1 ad 2747 1.1 ad /* Insert node into huge. */ 2748 1.1 ad node->chunk = ret; 2749 1.1 ad node->size = chunk_size; 2750 1.1 ad 2751 1.1 ad malloc_mutex_lock(&chunks_mtx); 2752 1.58 ad rb_tree_insert_node(&huge, node); 2753 1.1 ad #ifdef MALLOC_STATS 2754 1.1 ad huge_nmalloc++; 2755 1.1 ad huge_allocated += chunk_size; 2756 1.1 ad #endif 2757 1.1 ad malloc_mutex_unlock(&chunks_mtx); 2758 1.1 ad 2759 1.57 ad if (OPT(junk)) 2760 1.1 ad memset(ret, 0xa5, chunk_size); 2761 1.57 ad else if (OPT(zero)) 2762 1.1 ad memset(ret, 0, chunk_size); 2763 1.1 ad 2764 1.1 ad return (ret); 2765 1.1 ad } 2766 1.1 ad 2767 1.1 ad static void * 2768 1.1 ad huge_ralloc(void *ptr, size_t size, size_t oldsize) 2769 1.1 ad { 2770 1.1 ad void *ret; 2771 1.1 ad 2772 1.1 ad /* Avoid moving the allocation if the size class would not change. */ 2773 1.1 ad if (oldsize > arena_maxclass && 2774 1.1 ad CHUNK_CEILING(size) == CHUNK_CEILING(oldsize)) { 2775 1.57 ad if (OPT(junk) && size < oldsize) { 2776 1.1 ad memset((void *)((uintptr_t)ptr + size), 0x5a, oldsize 2777 1.1 ad - size); 2778 1.57 ad } else if (OPT(zero) && size > oldsize) { 2779 1.1 ad memset((void *)((uintptr_t)ptr + oldsize), 0, size 2780 1.1 ad - oldsize); 2781 1.1 ad } 2782 1.1 ad return (ptr); 2783 1.1 ad } 2784 1.1 ad 2785 1.8 yamt if (CHUNK_ADDR2BASE(ptr) == ptr 2786 1.8 yamt #ifdef USE_BRK 2787 1.8 yamt && ((uintptr_t)ptr < (uintptr_t)brk_base 2788 1.8 yamt || (uintptr_t)ptr >= (uintptr_t)brk_max) 2789 1.8 yamt #endif 2790 1.8 yamt ) { 2791 1.8 yamt chunk_node_t *node, key; 2792 1.8 yamt void *newptr; 2793 1.8 yamt size_t oldcsize; 2794 1.8 yamt size_t newcsize; 2795 1.8 yamt 2796 1.8 yamt newcsize = CHUNK_CEILING(size); 2797 1.8 yamt oldcsize = CHUNK_CEILING(oldsize); 2798 1.8 yamt assert(oldcsize != newcsize); 2799 1.8 yamt if (newcsize == 0) { 2800 1.8 yamt /* size_t wrap-around */ 2801 1.8 yamt return (NULL); 2802 1.8 yamt } 2803 1.21 enami 2804 1.21 enami /* 2805 1.21 enami * Remove the old region from the tree now. If mremap() 2806 1.21 enami * returns the region to the system, other thread may 2807 1.21 enami * map it for same huge allocation and insert it to the 2808 1.21 enami * tree before we acquire the mutex lock again. 2809 1.21 enami */ 2810 1.21 enami malloc_mutex_lock(&chunks_mtx); 2811 1.57 ad key.chunk = __UNCONST(ptr); 2812 1.58 ad node = rb_tree_find_node(&huge, &key); 2813 1.21 enami assert(node != NULL); 2814 1.21 enami assert(node->chunk == ptr); 2815 1.21 enami assert(node->size == oldcsize); 2816 1.58 ad rb_tree_remove_node(&huge, node); 2817 1.21 enami malloc_mutex_unlock(&chunks_mtx); 2818 1.21 enami 2819 1.8 yamt newptr = mremap(ptr, oldcsize, NULL, newcsize, 2820 1.8 yamt MAP_ALIGNED(chunksize_2pow)); 2821 1.21 enami if (newptr == MAP_FAILED) { 2822 1.21 enami /* We still own the old region. */ 2823 1.21 enami malloc_mutex_lock(&chunks_mtx); 2824 1.58 ad rb_tree_insert_node(&huge, node); 2825 1.21 enami malloc_mutex_unlock(&chunks_mtx); 2826 1.21 enami } else { 2827 1.8 yamt assert(CHUNK_ADDR2BASE(newptr) == newptr); 2828 1.8 yamt 2829 1.21 enami /* Insert new or resized old region. */ 2830 1.8 yamt malloc_mutex_lock(&chunks_mtx); 2831 1.8 yamt node->size = newcsize; 2832 1.21 enami node->chunk = newptr; 2833 1.58 ad rb_tree_insert_node(&huge, node); 2834 1.8 yamt #ifdef MALLOC_STATS 2835 1.8 yamt huge_nralloc++; 2836 1.8 yamt huge_allocated += newcsize - oldcsize; 2837 1.8 yamt if (newcsize > oldcsize) { 2838 1.8 yamt stats_chunks.curchunks += 2839 1.8 yamt (newcsize - oldcsize) / chunksize; 2840 1.8 yamt if (stats_chunks.curchunks > 2841 1.8 yamt stats_chunks.highchunks) 2842 1.8 yamt stats_chunks.highchunks = 2843 1.8 yamt stats_chunks.curchunks; 2844 1.8 yamt } else { 2845 1.8 yamt stats_chunks.curchunks -= 2846 1.8 yamt (oldcsize - newcsize) / chunksize; 2847 1.8 yamt } 2848 1.8 yamt #endif 2849 1.8 yamt malloc_mutex_unlock(&chunks_mtx); 2850 1.8 yamt 2851 1.57 ad if (OPT(junk) && size < oldsize) { 2852 1.8 yamt memset((void *)((uintptr_t)newptr + size), 0x5a, 2853 1.8 yamt newcsize - size); 2854 1.57 ad } else if (OPT(zero) && size > oldsize) { 2855 1.8 yamt memset((void *)((uintptr_t)newptr + oldsize), 0, 2856 1.8 yamt size - oldsize); 2857 1.8 yamt } 2858 1.8 yamt return (newptr); 2859 1.8 yamt } 2860 1.8 yamt } 2861 1.8 yamt 2862 1.1 ad /* 2863 1.1 ad * If we get here, then size and oldsize are different enough that we 2864 1.1 ad * need to use a different size class. In that case, fall back to 2865 1.1 ad * allocating new space and copying. 2866 1.1 ad */ 2867 1.1 ad ret = huge_malloc(size); 2868 1.1 ad if (ret == NULL) 2869 1.1 ad return (NULL); 2870 1.1 ad 2871 1.1 ad if (CHUNK_ADDR2BASE(ptr) == ptr) { 2872 1.1 ad /* The old allocation is a chunk. */ 2873 1.1 ad if (size < oldsize) 2874 1.1 ad memcpy(ret, ptr, size); 2875 1.1 ad else 2876 1.1 ad memcpy(ret, ptr, oldsize); 2877 1.1 ad } else { 2878 1.1 ad /* The old allocation is a region. */ 2879 1.1 ad assert(oldsize < size); 2880 1.1 ad memcpy(ret, ptr, oldsize); 2881 1.1 ad } 2882 1.1 ad idalloc(ptr); 2883 1.1 ad return (ret); 2884 1.1 ad } 2885 1.1 ad 2886 1.1 ad static void 2887 1.1 ad huge_dalloc(void *ptr) 2888 1.1 ad { 2889 1.1 ad chunk_node_t key; 2890 1.1 ad chunk_node_t *node; 2891 1.1 ad 2892 1.1 ad malloc_mutex_lock(&chunks_mtx); 2893 1.1 ad 2894 1.1 ad /* Extract from tree of huge allocations. */ 2895 1.1 ad key.chunk = ptr; 2896 1.58 ad node = rb_tree_find_node(&huge, &key); 2897 1.1 ad assert(node != NULL); 2898 1.1 ad assert(node->chunk == ptr); 2899 1.58 ad rb_tree_remove_node(&huge, node); 2900 1.1 ad 2901 1.1 ad #ifdef MALLOC_STATS 2902 1.1 ad huge_ndalloc++; 2903 1.1 ad huge_allocated -= node->size; 2904 1.1 ad #endif 2905 1.1 ad 2906 1.1 ad malloc_mutex_unlock(&chunks_mtx); 2907 1.1 ad 2908 1.1 ad /* Unmap chunk. */ 2909 1.1 ad #ifdef USE_BRK 2910 1.57 ad if (OPT(junk)) 2911 1.1 ad memset(node->chunk, 0x5a, node->size); 2912 1.1 ad #endif 2913 1.1 ad chunk_dealloc(node->chunk, node->size); 2914 1.1 ad 2915 1.1 ad base_chunk_node_dealloc(node); 2916 1.1 ad } 2917 1.1 ad 2918 1.1 ad static void * 2919 1.1 ad imalloc(size_t size) 2920 1.1 ad { 2921 1.1 ad void *ret; 2922 1.1 ad 2923 1.1 ad assert(size != 0); 2924 1.1 ad 2925 1.1 ad if (size <= arena_maxclass) 2926 1.1 ad ret = arena_malloc(choose_arena(), size); 2927 1.1 ad else 2928 1.1 ad ret = huge_malloc(size); 2929 1.1 ad 2930 1.1 ad return (ret); 2931 1.1 ad } 2932 1.1 ad 2933 1.1 ad static void * 2934 1.1 ad ipalloc(size_t alignment, size_t size) 2935 1.1 ad { 2936 1.1 ad void *ret; 2937 1.1 ad size_t ceil_size; 2938 1.1 ad 2939 1.1 ad /* 2940 1.1 ad * Round size up to the nearest multiple of alignment. 2941 1.1 ad * 2942 1.1 ad * This done, we can take advantage of the fact that for each small 2943 1.1 ad * size class, every object is aligned at the smallest power of two 2944 1.1 ad * that is non-zero in the base two representation of the size. For 2945 1.1 ad * example: 2946 1.1 ad * 2947 1.1 ad * Size | Base 2 | Minimum alignment 2948 1.1 ad * -----+----------+------------------ 2949 1.1 ad * 96 | 1100000 | 32 2950 1.1 ad * 144 | 10100000 | 32 2951 1.1 ad * 192 | 11000000 | 64 2952 1.1 ad * 2953 1.1 ad * Depending on runtime settings, it is possible that arena_malloc() 2954 1.1 ad * will further round up to a power of two, but that never causes 2955 1.1 ad * correctness issues. 2956 1.1 ad */ 2957 1.1 ad ceil_size = (size + (alignment - 1)) & (-alignment); 2958 1.1 ad /* 2959 1.1 ad * (ceil_size < size) protects against the combination of maximal 2960 1.1 ad * alignment and size greater than maximal alignment. 2961 1.1 ad */ 2962 1.1 ad if (ceil_size < size) { 2963 1.1 ad /* size_t overflow. */ 2964 1.1 ad return (NULL); 2965 1.1 ad } 2966 1.1 ad 2967 1.1 ad if (ceil_size <= pagesize || (alignment <= pagesize 2968 1.1 ad && ceil_size <= arena_maxclass)) 2969 1.1 ad ret = arena_malloc(choose_arena(), ceil_size); 2970 1.1 ad else { 2971 1.1 ad size_t run_size; 2972 1.1 ad 2973 1.1 ad /* 2974 1.1 ad * We can't achieve sub-page alignment, so round up alignment 2975 1.1 ad * permanently; it makes later calculations simpler. 2976 1.1 ad */ 2977 1.1 ad alignment = PAGE_CEILING(alignment); 2978 1.1 ad ceil_size = PAGE_CEILING(size); 2979 1.1 ad /* 2980 1.1 ad * (ceil_size < size) protects against very large sizes within 2981 1.1 ad * pagesize of SIZE_T_MAX. 2982 1.1 ad * 2983 1.1 ad * (ceil_size + alignment < ceil_size) protects against the 2984 1.1 ad * combination of maximal alignment and ceil_size large enough 2985 1.1 ad * to cause overflow. This is similar to the first overflow 2986 1.1 ad * check above, but it needs to be repeated due to the new 2987 1.1 ad * ceil_size value, which may now be *equal* to maximal 2988 1.1 ad * alignment, whereas before we only detected overflow if the 2989 1.1 ad * original size was *greater* than maximal alignment. 2990 1.1 ad */ 2991 1.1 ad if (ceil_size < size || ceil_size + alignment < ceil_size) { 2992 1.1 ad /* size_t overflow. */ 2993 1.1 ad return (NULL); 2994 1.1 ad } 2995 1.1 ad 2996 1.1 ad /* 2997 1.1 ad * Calculate the size of the over-size run that arena_palloc() 2998 1.1 ad * would need to allocate in order to guarantee the alignment. 2999 1.1 ad */ 3000 1.1 ad if (ceil_size >= alignment) 3001 1.1 ad run_size = ceil_size + alignment - pagesize; 3002 1.1 ad else { 3003 1.1 ad /* 3004 1.1 ad * It is possible that (alignment << 1) will cause 3005 1.1 ad * overflow, but it doesn't matter because we also 3006 1.1 ad * subtract pagesize, which in the case of overflow 3007 1.1 ad * leaves us with a very large run_size. That causes 3008 1.1 ad * the first conditional below to fail, which means 3009 1.1 ad * that the bogus run_size value never gets used for 3010 1.1 ad * anything important. 3011 1.1 ad */ 3012 1.1 ad run_size = (alignment << 1) - pagesize; 3013 1.1 ad } 3014 1.1 ad 3015 1.1 ad if (run_size <= arena_maxclass) { 3016 1.1 ad ret = arena_palloc(choose_arena(), alignment, ceil_size, 3017 1.1 ad run_size); 3018 1.1 ad } else if (alignment <= chunksize) 3019 1.1 ad ret = huge_malloc(ceil_size); 3020 1.1 ad else 3021 1.1 ad ret = huge_palloc(alignment, ceil_size); 3022 1.1 ad } 3023 1.1 ad 3024 1.1 ad assert(((uintptr_t)ret & (alignment - 1)) == 0); 3025 1.1 ad return (ret); 3026 1.1 ad } 3027 1.1 ad 3028 1.1 ad static void * 3029 1.1 ad icalloc(size_t size) 3030 1.1 ad { 3031 1.1 ad void *ret; 3032 1.1 ad 3033 1.1 ad if (size <= arena_maxclass) { 3034 1.1 ad ret = arena_malloc(choose_arena(), size); 3035 1.1 ad if (ret == NULL) 3036 1.1 ad return (NULL); 3037 1.1 ad memset(ret, 0, size); 3038 1.1 ad } else { 3039 1.1 ad /* 3040 1.1 ad * The virtual memory system provides zero-filled pages, so 3041 1.1 ad * there is no need to do so manually, unless opt_junk is 3042 1.1 ad * enabled, in which case huge_malloc() fills huge allocations 3043 1.1 ad * with junk. 3044 1.1 ad */ 3045 1.1 ad ret = huge_malloc(size); 3046 1.1 ad if (ret == NULL) 3047 1.1 ad return (NULL); 3048 1.1 ad 3049 1.57 ad if (OPT(junk)) 3050 1.1 ad memset(ret, 0, size); 3051 1.1 ad #ifdef USE_BRK 3052 1.1 ad else if ((uintptr_t)ret >= (uintptr_t)brk_base 3053 1.1 ad && (uintptr_t)ret < (uintptr_t)brk_max) { 3054 1.1 ad /* 3055 1.1 ad * This may be a re-used brk chunk. Therefore, zero 3056 1.1 ad * the memory. 3057 1.1 ad */ 3058 1.1 ad memset(ret, 0, size); 3059 1.1 ad } 3060 1.1 ad #endif 3061 1.1 ad } 3062 1.1 ad 3063 1.1 ad return (ret); 3064 1.1 ad } 3065 1.1 ad 3066 1.1 ad static size_t 3067 1.1 ad isalloc(const void *ptr) 3068 1.1 ad { 3069 1.1 ad size_t ret; 3070 1.1 ad arena_chunk_t *chunk; 3071 1.1 ad 3072 1.1 ad assert(ptr != NULL); 3073 1.1 ad 3074 1.1 ad chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 3075 1.1 ad if (chunk != ptr) { 3076 1.1 ad /* Region. */ 3077 1.1 ad assert(chunk->arena->magic == ARENA_MAGIC); 3078 1.1 ad 3079 1.1 ad ret = arena_salloc(ptr); 3080 1.1 ad } else { 3081 1.1 ad chunk_node_t *node, key; 3082 1.1 ad 3083 1.1 ad /* Chunk (huge allocation). */ 3084 1.1 ad 3085 1.1 ad malloc_mutex_lock(&chunks_mtx); 3086 1.1 ad 3087 1.1 ad /* Extract from tree of huge allocations. */ 3088 1.57 ad key.chunk = __UNCONST(ptr); 3089 1.58 ad node = rb_tree_find_node(&huge, &key); 3090 1.1 ad assert(node != NULL); 3091 1.1 ad 3092 1.1 ad ret = node->size; 3093 1.1 ad 3094 1.1 ad malloc_mutex_unlock(&chunks_mtx); 3095 1.1 ad } 3096 1.1 ad 3097 1.1 ad return (ret); 3098 1.1 ad } 3099 1.1 ad 3100 1.1 ad static void * 3101 1.1 ad iralloc(void *ptr, size_t size) 3102 1.1 ad { 3103 1.1 ad void *ret; 3104 1.1 ad size_t oldsize; 3105 1.1 ad 3106 1.1 ad assert(ptr != NULL); 3107 1.1 ad assert(size != 0); 3108 1.1 ad 3109 1.1 ad oldsize = isalloc(ptr); 3110 1.1 ad 3111 1.1 ad if (size <= arena_maxclass) 3112 1.1 ad ret = arena_ralloc(ptr, size, oldsize); 3113 1.1 ad else 3114 1.1 ad ret = huge_ralloc(ptr, size, oldsize); 3115 1.1 ad 3116 1.1 ad return (ret); 3117 1.1 ad } 3118 1.1 ad 3119 1.1 ad static void 3120 1.1 ad idalloc(void *ptr) 3121 1.1 ad { 3122 1.1 ad arena_chunk_t *chunk; 3123 1.1 ad 3124 1.1 ad assert(ptr != NULL); 3125 1.1 ad 3126 1.1 ad chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 3127 1.1 ad if (chunk != ptr) { 3128 1.1 ad /* Region. */ 3129 1.1 ad arena_dalloc(chunk->arena, chunk, ptr); 3130 1.1 ad } else 3131 1.1 ad huge_dalloc(ptr); 3132 1.1 ad } 3133 1.1 ad 3134 1.1 ad static void 3135 1.1 ad malloc_print_stats(void) 3136 1.1 ad { 3137 1.1 ad 3138 1.57 ad if (OPT(print_stats)) { 3139 1.1 ad char s[UMAX2S_BUFSIZE]; 3140 1.1 ad _malloc_message("___ Begin malloc statistics ___\n", "", "", 3141 1.1 ad ""); 3142 1.1 ad _malloc_message("Assertions ", 3143 1.1 ad #ifdef NDEBUG 3144 1.1 ad "disabled", 3145 1.1 ad #else 3146 1.1 ad "enabled", 3147 1.1 ad #endif 3148 1.1 ad "\n", ""); 3149 1.1 ad _malloc_message("Boolean MALLOC_OPTIONS: ", 3150 1.1 ad opt_abort ? "A" : "a", 3151 1.1 ad opt_junk ? "J" : "j", 3152 1.1 ad opt_hint ? "H" : "h"); 3153 1.57 ad _malloc_message(OPT(utrace) ? "PU" : "Pu", 3154 1.1 ad opt_sysv ? "V" : "v", 3155 1.1 ad opt_xmalloc ? "X" : "x", 3156 1.1 ad opt_zero ? "Z\n" : "z\n"); 3157 1.1 ad 3158 1.28 christos _malloc_message("CPUs: ", size_t2s(ncpus, s), "\n", ""); 3159 1.28 christos _malloc_message("Pointer size: ", size_t2s(sizeof(void *), s), 3160 1.1 ad "\n", ""); 3161 1.28 christos _malloc_message("Quantum size: ", size_t2s(quantum, s), "\n", ""); 3162 1.28 christos _malloc_message("Max small size: ", size_t2s(small_max, s), "\n", 3163 1.1 ad ""); 3164 1.1 ad 3165 1.28 christos _malloc_message("Chunk size: ", size_t2s(chunksize, s), "", ""); 3166 1.28 christos _malloc_message(" (2^", size_t2s((size_t)opt_chunk_2pow, s), 3167 1.27 matt ")\n", ""); 3168 1.1 ad 3169 1.1 ad #ifdef MALLOC_STATS 3170 1.1 ad { 3171 1.1 ad size_t allocated, mapped; 3172 1.1 ad unsigned i; 3173 1.1 ad arena_t *arena; 3174 1.1 ad 3175 1.1 ad /* Calculate and print allocated/mapped stats. */ 3176 1.1 ad 3177 1.1 ad /* arenas. */ 3178 1.57 ad for (i = 0, allocated = 0; i < ncpus; i++) { 3179 1.1 ad if (arenas[i] != NULL) { 3180 1.1 ad malloc_mutex_lock(&arenas[i]->mtx); 3181 1.1 ad allocated += 3182 1.1 ad arenas[i]->stats.allocated_small; 3183 1.1 ad allocated += 3184 1.1 ad arenas[i]->stats.allocated_large; 3185 1.1 ad malloc_mutex_unlock(&arenas[i]->mtx); 3186 1.1 ad } 3187 1.1 ad } 3188 1.1 ad 3189 1.1 ad /* huge/base. */ 3190 1.1 ad malloc_mutex_lock(&chunks_mtx); 3191 1.1 ad allocated += huge_allocated; 3192 1.1 ad mapped = stats_chunks.curchunks * chunksize; 3193 1.1 ad malloc_mutex_unlock(&chunks_mtx); 3194 1.1 ad 3195 1.1 ad malloc_mutex_lock(&base_mtx); 3196 1.1 ad mapped += base_mapped; 3197 1.1 ad malloc_mutex_unlock(&base_mtx); 3198 1.1 ad 3199 1.1 ad malloc_printf("Allocated: %zu, mapped: %zu\n", 3200 1.1 ad allocated, mapped); 3201 1.1 ad 3202 1.1 ad /* Print chunk stats. */ 3203 1.1 ad { 3204 1.1 ad chunk_stats_t chunks_stats; 3205 1.1 ad 3206 1.1 ad malloc_mutex_lock(&chunks_mtx); 3207 1.1 ad chunks_stats = stats_chunks; 3208 1.1 ad malloc_mutex_unlock(&chunks_mtx); 3209 1.1 ad 3210 1.1 ad malloc_printf("chunks: nchunks " 3211 1.1 ad "highchunks curchunks\n"); 3212 1.1 ad malloc_printf(" %13llu%13lu%13lu\n", 3213 1.1 ad chunks_stats.nchunks, 3214 1.1 ad chunks_stats.highchunks, 3215 1.1 ad chunks_stats.curchunks); 3216 1.1 ad } 3217 1.1 ad 3218 1.1 ad /* Print chunk stats. */ 3219 1.1 ad malloc_printf( 3220 1.8 yamt "huge: nmalloc ndalloc " 3221 1.8 yamt "nralloc allocated\n"); 3222 1.8 yamt malloc_printf(" %12llu %12llu %12llu %12zu\n", 3223 1.8 yamt huge_nmalloc, huge_ndalloc, huge_nralloc, 3224 1.8 yamt huge_allocated); 3225 1.1 ad 3226 1.1 ad /* Print stats for each arena. */ 3227 1.57 ad for (i = 0; i < ncpus; i++) { 3228 1.1 ad arena = arenas[i]; 3229 1.1 ad if (arena != NULL) { 3230 1.1 ad malloc_printf( 3231 1.2 ad "\narenas[%u] @ %p\n", i, arena); 3232 1.1 ad malloc_mutex_lock(&arena->mtx); 3233 1.1 ad stats_print(arena); 3234 1.1 ad malloc_mutex_unlock(&arena->mtx); 3235 1.1 ad } 3236 1.1 ad } 3237 1.1 ad } 3238 1.1 ad #endif /* #ifdef MALLOC_STATS */ 3239 1.1 ad _malloc_message("--- End malloc statistics ---\n", "", "", ""); 3240 1.1 ad } 3241 1.1 ad } 3242 1.1 ad 3243 1.1 ad /* 3244 1.57 ad * libpthread might call malloc(3), so the malloc implementation has to take 3245 1.57 ad * pains to avoid infinite recursion during initialization. 3246 1.1 ad */ 3247 1.1 ad static inline bool 3248 1.1 ad malloc_init(void) 3249 1.1 ad { 3250 1.1 ad 3251 1.57 ad if (__predict_false(malloc_initialized == false)) 3252 1.1 ad return (malloc_init_hard()); 3253 1.1 ad 3254 1.1 ad return (false); 3255 1.1 ad } 3256 1.1 ad 3257 1.1 ad static bool 3258 1.1 ad malloc_init_hard(void) 3259 1.1 ad { 3260 1.1 ad unsigned i, j; 3261 1.9 christos ssize_t linklen; 3262 1.1 ad char buf[PATH_MAX + 1]; 3263 1.2 ad const char *opts = ""; 3264 1.23 christos int serrno; 3265 1.1 ad 3266 1.1 ad malloc_mutex_lock(&init_lock); 3267 1.1 ad if (malloc_initialized) { 3268 1.1 ad /* 3269 1.1 ad * Another thread initialized the allocator before this one 3270 1.1 ad * acquired init_lock. 3271 1.1 ad */ 3272 1.1 ad malloc_mutex_unlock(&init_lock); 3273 1.1 ad return (false); 3274 1.1 ad } 3275 1.1 ad 3276 1.24 christos serrno = errno; 3277 1.1 ad /* Get number of CPUs. */ 3278 1.1 ad { 3279 1.1 ad int mib[2]; 3280 1.1 ad size_t len; 3281 1.1 ad 3282 1.1 ad mib[0] = CTL_HW; 3283 1.1 ad mib[1] = HW_NCPU; 3284 1.1 ad len = sizeof(ncpus); 3285 1.1 ad if (sysctl(mib, 2, &ncpus, &len, (void *) 0, 0) == -1) { 3286 1.1 ad /* Error. */ 3287 1.1 ad ncpus = 1; 3288 1.1 ad } 3289 1.1 ad } 3290 1.1 ad 3291 1.1 ad /* Get page size. */ 3292 1.1 ad { 3293 1.1 ad long result; 3294 1.1 ad 3295 1.57 ad result = getpagesize(); 3296 1.1 ad assert(result != -1); 3297 1.1 ad pagesize = (unsigned) result; 3298 1.1 ad 3299 1.1 ad /* 3300 1.1 ad * We assume that pagesize is a power of 2 when calculating 3301 1.1 ad * pagesize_mask and pagesize_2pow. 3302 1.1 ad */ 3303 1.1 ad assert(((result - 1) & result) == 0); 3304 1.1 ad pagesize_mask = result - 1; 3305 1.1 ad pagesize_2pow = ffs((int)result) - 1; 3306 1.1 ad } 3307 1.1 ad 3308 1.1 ad for (i = 0; i < 3; i++) { 3309 1.1 ad /* Get runtime configuration. */ 3310 1.1 ad switch (i) { 3311 1.1 ad case 0: 3312 1.1 ad if ((linklen = readlink("/etc/malloc.conf", buf, 3313 1.57 ad sizeof(buf) - 1)) != -1) { 3314 1.1 ad /* 3315 1.1 ad * Use the contents of the "/etc/malloc.conf" 3316 1.1 ad * symbolic link's name. 3317 1.1 ad */ 3318 1.1 ad buf[linklen] = '\0'; 3319 1.1 ad opts = buf; 3320 1.1 ad } else { 3321 1.1 ad /* No configuration specified. */ 3322 1.1 ad buf[0] = '\0'; 3323 1.1 ad opts = buf; 3324 1.1 ad } 3325 1.1 ad break; 3326 1.1 ad case 1: 3327 1.18 ad if ((opts = getenv("MALLOC_OPTIONS")) != NULL && 3328 1.18 ad issetugid() == 0) { 3329 1.1 ad /* 3330 1.1 ad * Do nothing; opts is already initialized to 3331 1.1 ad * the value of the MALLOC_OPTIONS environment 3332 1.1 ad * variable. 3333 1.1 ad */ 3334 1.1 ad } else { 3335 1.1 ad /* No configuration specified. */ 3336 1.1 ad buf[0] = '\0'; 3337 1.1 ad opts = buf; 3338 1.1 ad } 3339 1.1 ad break; 3340 1.1 ad case 2: 3341 1.1 ad if (_malloc_options != NULL) { 3342 1.57 ad /* 3343 1.57 ad * Use options that were compiled into the program. 3344 1.57 ad */ 3345 1.57 ad opts = _malloc_options; 3346 1.1 ad } else { 3347 1.1 ad /* No configuration specified. */ 3348 1.1 ad buf[0] = '\0'; 3349 1.1 ad opts = buf; 3350 1.1 ad } 3351 1.1 ad break; 3352 1.1 ad default: 3353 1.1 ad /* NOTREACHED */ 3354 1.7 yamt /* LINTED */ 3355 1.1 ad assert(false); 3356 1.1 ad } 3357 1.1 ad 3358 1.1 ad for (j = 0; opts[j] != '\0'; j++) { 3359 1.1 ad switch (opts[j]) { 3360 1.1 ad case 'a': 3361 1.1 ad opt_abort = false; 3362 1.1 ad break; 3363 1.1 ad case 'A': 3364 1.1 ad opt_abort = true; 3365 1.1 ad break; 3366 1.1 ad case 'h': 3367 1.1 ad opt_hint = false; 3368 1.1 ad break; 3369 1.1 ad case 'H': 3370 1.1 ad opt_hint = true; 3371 1.1 ad break; 3372 1.1 ad case 'j': 3373 1.1 ad opt_junk = false; 3374 1.1 ad break; 3375 1.1 ad case 'J': 3376 1.1 ad opt_junk = true; 3377 1.1 ad break; 3378 1.1 ad case 'k': 3379 1.1 ad /* 3380 1.1 ad * Chunks always require at least one header 3381 1.1 ad * page, so chunks can never be smaller than 3382 1.1 ad * two pages. 3383 1.1 ad */ 3384 1.1 ad if (opt_chunk_2pow > pagesize_2pow + 1) 3385 1.1 ad opt_chunk_2pow--; 3386 1.1 ad break; 3387 1.1 ad case 'K': 3388 1.20 lukem if (opt_chunk_2pow + 1 < 3389 1.20 lukem (int)(sizeof(size_t) << 3)) 3390 1.1 ad opt_chunk_2pow++; 3391 1.1 ad break; 3392 1.1 ad case 'p': 3393 1.1 ad opt_print_stats = false; 3394 1.1 ad break; 3395 1.1 ad case 'P': 3396 1.1 ad opt_print_stats = true; 3397 1.1 ad break; 3398 1.1 ad case 'q': 3399 1.1 ad if (opt_quantum_2pow > QUANTUM_2POW_MIN) 3400 1.1 ad opt_quantum_2pow--; 3401 1.1 ad break; 3402 1.1 ad case 'Q': 3403 1.1 ad if (opt_quantum_2pow < pagesize_2pow - 1) 3404 1.1 ad opt_quantum_2pow++; 3405 1.1 ad break; 3406 1.1 ad case 's': 3407 1.1 ad if (opt_small_max_2pow > QUANTUM_2POW_MIN) 3408 1.1 ad opt_small_max_2pow--; 3409 1.1 ad break; 3410 1.1 ad case 'S': 3411 1.1 ad if (opt_small_max_2pow < pagesize_2pow - 1) 3412 1.1 ad opt_small_max_2pow++; 3413 1.1 ad break; 3414 1.1 ad case 'u': 3415 1.1 ad opt_utrace = false; 3416 1.1 ad break; 3417 1.1 ad case 'U': 3418 1.1 ad opt_utrace = true; 3419 1.1 ad break; 3420 1.1 ad case 'v': 3421 1.1 ad opt_sysv = false; 3422 1.1 ad break; 3423 1.1 ad case 'V': 3424 1.1 ad opt_sysv = true; 3425 1.1 ad break; 3426 1.1 ad case 'x': 3427 1.1 ad opt_xmalloc = false; 3428 1.1 ad break; 3429 1.1 ad case 'X': 3430 1.1 ad opt_xmalloc = true; 3431 1.1 ad break; 3432 1.1 ad case 'z': 3433 1.1 ad opt_zero = false; 3434 1.1 ad break; 3435 1.1 ad case 'Z': 3436 1.1 ad opt_zero = true; 3437 1.1 ad break; 3438 1.1 ad default: { 3439 1.1 ad char cbuf[2]; 3440 1.55 skrll 3441 1.1 ad cbuf[0] = opts[j]; 3442 1.1 ad cbuf[1] = '\0'; 3443 1.16 christos _malloc_message(getprogname(), 3444 1.1 ad ": (malloc) Unsupported character in " 3445 1.1 ad "malloc options: '", cbuf, "'\n"); 3446 1.1 ad } 3447 1.1 ad } 3448 1.1 ad } 3449 1.1 ad } 3450 1.24 christos errno = serrno; 3451 1.1 ad 3452 1.1 ad /* Take care to call atexit() only once. */ 3453 1.57 ad if (OPT(print_stats)) { 3454 1.1 ad /* Print statistics at exit. */ 3455 1.1 ad atexit(malloc_print_stats); 3456 1.1 ad } 3457 1.1 ad 3458 1.1 ad /* Set variables according to the value of opt_small_max_2pow. */ 3459 1.1 ad if (opt_small_max_2pow < opt_quantum_2pow) 3460 1.1 ad opt_small_max_2pow = opt_quantum_2pow; 3461 1.57 ad small_max = (1U << opt_small_max_2pow); 3462 1.1 ad 3463 1.1 ad /* Set bin-related variables. */ 3464 1.1 ad bin_maxclass = (pagesize >> 1); 3465 1.1 ad assert(opt_quantum_2pow >= TINY_MIN_2POW); 3466 1.9 christos ntbins = (unsigned)(opt_quantum_2pow - TINY_MIN_2POW); 3467 1.57 ad assert(ntbins <= (unsigned)opt_quantum_2pow); 3468 1.9 christos nqbins = (unsigned)(small_max >> opt_quantum_2pow); 3469 1.9 christos nsbins = (unsigned)(pagesize_2pow - opt_small_max_2pow - 1); 3470 1.1 ad 3471 1.1 ad /* Set variables according to the value of opt_quantum_2pow. */ 3472 1.1 ad quantum = (1 << opt_quantum_2pow); 3473 1.1 ad quantum_mask = quantum - 1; 3474 1.1 ad if (ntbins > 0) 3475 1.1 ad small_min = (quantum >> 1) + 1; 3476 1.1 ad else 3477 1.1 ad small_min = 1; 3478 1.1 ad assert(small_min <= quantum); 3479 1.1 ad 3480 1.1 ad /* Set variables according to the value of opt_chunk_2pow. */ 3481 1.1 ad chunksize = (1LU << opt_chunk_2pow); 3482 1.1 ad chunksize_mask = chunksize - 1; 3483 1.9 christos chunksize_2pow = (unsigned)opt_chunk_2pow; 3484 1.9 christos chunk_npages = (unsigned)(chunksize >> pagesize_2pow); 3485 1.1 ad { 3486 1.1 ad unsigned header_size; 3487 1.1 ad 3488 1.9 christos header_size = (unsigned)(sizeof(arena_chunk_t) + 3489 1.9 christos (sizeof(arena_chunk_map_t) * (chunk_npages - 1))); 3490 1.1 ad arena_chunk_header_npages = (header_size >> pagesize_2pow); 3491 1.1 ad if ((header_size & pagesize_mask) != 0) 3492 1.1 ad arena_chunk_header_npages++; 3493 1.1 ad } 3494 1.1 ad arena_maxclass = chunksize - (arena_chunk_header_npages << 3495 1.1 ad pagesize_2pow); 3496 1.1 ad 3497 1.1 ad UTRACE(0, 0, 0); 3498 1.1 ad 3499 1.1 ad #ifdef MALLOC_STATS 3500 1.1 ad memset(&stats_chunks, 0, sizeof(chunk_stats_t)); 3501 1.1 ad #endif 3502 1.1 ad 3503 1.1 ad /* Various sanity checks that regard configuration. */ 3504 1.1 ad assert(quantum >= sizeof(void *)); 3505 1.1 ad assert(quantum <= pagesize); 3506 1.1 ad assert(chunksize >= pagesize); 3507 1.1 ad assert(quantum * 4 <= chunksize); 3508 1.1 ad 3509 1.1 ad /* Initialize chunks data. */ 3510 1.1 ad malloc_mutex_init(&chunks_mtx); 3511 1.58 ad rb_tree_init(&huge, &chunk_tree_ops); 3512 1.1 ad #ifdef USE_BRK 3513 1.1 ad malloc_mutex_init(&brk_mtx); 3514 1.1 ad brk_base = sbrk(0); 3515 1.1 ad brk_prev = brk_base; 3516 1.1 ad brk_max = brk_base; 3517 1.1 ad #endif 3518 1.1 ad #ifdef MALLOC_STATS 3519 1.1 ad huge_nmalloc = 0; 3520 1.1 ad huge_ndalloc = 0; 3521 1.8 yamt huge_nralloc = 0; 3522 1.1 ad huge_allocated = 0; 3523 1.1 ad #endif 3524 1.58 ad rb_tree_init(&old_chunks, &chunk_tree_ops); 3525 1.1 ad 3526 1.1 ad /* Initialize base allocation data structures. */ 3527 1.1 ad #ifdef MALLOC_STATS 3528 1.1 ad base_mapped = 0; 3529 1.1 ad #endif 3530 1.1 ad #ifdef USE_BRK 3531 1.1 ad /* 3532 1.1 ad * Allocate a base chunk here, since it doesn't actually have to be 3533 1.1 ad * chunk-aligned. Doing this before allocating any other chunks allows 3534 1.1 ad * the use of space that would otherwise be wasted. 3535 1.1 ad */ 3536 1.1 ad base_pages_alloc(0); 3537 1.1 ad #endif 3538 1.1 ad base_chunk_nodes = NULL; 3539 1.1 ad malloc_mutex_init(&base_mtx); 3540 1.1 ad 3541 1.1 ad /* Allocate and initialize arenas. */ 3542 1.57 ad arenas = (arena_t **)base_alloc(sizeof(arena_t *) * ncpus); 3543 1.1 ad if (arenas == NULL) { 3544 1.1 ad malloc_mutex_unlock(&init_lock); 3545 1.1 ad return (true); 3546 1.1 ad } 3547 1.1 ad /* 3548 1.1 ad * Zero the array. In practice, this should always be pre-zeroed, 3549 1.1 ad * since it was just mmap()ed, but let's be sure. 3550 1.1 ad */ 3551 1.57 ad memset(arenas, 0, sizeof(arena_t *) * ncpus); 3552 1.1 ad 3553 1.1 ad /* 3554 1.1 ad * Initialize one arena here. The rest are lazily created in 3555 1.1 ad * arena_choose_hard(). 3556 1.1 ad */ 3557 1.57 ad if ((arenas[0] = arenas_extend()) == NULL) { 3558 1.1 ad malloc_mutex_unlock(&init_lock); 3559 1.1 ad return (true); 3560 1.1 ad } 3561 1.1 ad 3562 1.1 ad malloc_mutex_init(&arenas_mtx); 3563 1.1 ad 3564 1.1 ad malloc_initialized = true; 3565 1.1 ad malloc_mutex_unlock(&init_lock); 3566 1.1 ad return (false); 3567 1.1 ad } 3568 1.1 ad 3569 1.1 ad /* 3570 1.1 ad * End general internal functions. 3571 1.1 ad */ 3572 1.1 ad /******************************************************************************/ 3573 1.1 ad /* 3574 1.1 ad * Begin malloc(3)-compatible functions. 3575 1.1 ad */ 3576 1.1 ad 3577 1.1 ad void * 3578 1.1 ad malloc(size_t size) 3579 1.1 ad { 3580 1.1 ad void *ret; 3581 1.1 ad 3582 1.57 ad if (__predict_false(malloc_init())) { 3583 1.1 ad ret = NULL; 3584 1.1 ad goto RETURN; 3585 1.1 ad } 3586 1.1 ad 3587 1.57 ad if (__predict_false(size == 0)) { 3588 1.57 ad if (NOT_OPT(sysv)) 3589 1.1 ad size = 1; 3590 1.1 ad else { 3591 1.1 ad ret = NULL; 3592 1.1 ad goto RETURN; 3593 1.1 ad } 3594 1.1 ad } 3595 1.1 ad 3596 1.1 ad ret = imalloc(size); 3597 1.1 ad 3598 1.1 ad RETURN: 3599 1.57 ad if (__predict_false(ret == NULL)) { 3600 1.57 ad if (OPT(xmalloc)) { 3601 1.16 christos _malloc_message(getprogname(), 3602 1.1 ad ": (malloc) Error in malloc(): out of memory\n", "", 3603 1.1 ad ""); 3604 1.1 ad abort(); 3605 1.1 ad } 3606 1.1 ad errno = ENOMEM; 3607 1.1 ad } 3608 1.1 ad 3609 1.1 ad UTRACE(0, size, ret); 3610 1.1 ad return (ret); 3611 1.1 ad } 3612 1.1 ad 3613 1.1 ad int 3614 1.1 ad posix_memalign(void **memptr, size_t alignment, size_t size) 3615 1.1 ad { 3616 1.1 ad int ret; 3617 1.1 ad void *result; 3618 1.1 ad 3619 1.57 ad if (__predict_false(malloc_init())) 3620 1.1 ad result = NULL; 3621 1.1 ad else { 3622 1.1 ad /* Make sure that alignment is a large enough power of 2. */ 3623 1.1 ad if (((alignment - 1) & alignment) != 0 3624 1.1 ad || alignment < sizeof(void *)) { 3625 1.57 ad if (OPT(xmalloc)) { 3626 1.16 christos _malloc_message(getprogname(), 3627 1.1 ad ": (malloc) Error in posix_memalign(): " 3628 1.1 ad "invalid alignment\n", "", ""); 3629 1.1 ad abort(); 3630 1.1 ad } 3631 1.1 ad result = NULL; 3632 1.1 ad ret = EINVAL; 3633 1.1 ad goto RETURN; 3634 1.1 ad } 3635 1.1 ad 3636 1.62 ad if (size == 0) { 3637 1.62 ad if (NOT_OPT(sysv)) 3638 1.62 ad size = 1; 3639 1.62 ad else { 3640 1.62 ad result = NULL; 3641 1.62 ad ret = 0; 3642 1.62 ad goto RETURN; 3643 1.62 ad } 3644 1.62 ad } 3645 1.1 ad result = ipalloc(alignment, size); 3646 1.1 ad } 3647 1.1 ad 3648 1.57 ad if (__predict_false(result == NULL)) { 3649 1.57 ad if (OPT(xmalloc)) { 3650 1.16 christos _malloc_message(getprogname(), 3651 1.1 ad ": (malloc) Error in posix_memalign(): out of memory\n", 3652 1.1 ad "", ""); 3653 1.1 ad abort(); 3654 1.1 ad } 3655 1.1 ad ret = ENOMEM; 3656 1.1 ad goto RETURN; 3657 1.1 ad } 3658 1.1 ad 3659 1.1 ad *memptr = result; 3660 1.1 ad ret = 0; 3661 1.1 ad 3662 1.1 ad RETURN: 3663 1.1 ad UTRACE(0, size, result); 3664 1.1 ad return (ret); 3665 1.1 ad } 3666 1.1 ad 3667 1.1 ad void * 3668 1.1 ad calloc(size_t num, size_t size) 3669 1.1 ad { 3670 1.1 ad void *ret; 3671 1.1 ad size_t num_size; 3672 1.1 ad 3673 1.57 ad if (__predict_false(malloc_init())) { 3674 1.1 ad num_size = 0; 3675 1.1 ad ret = NULL; 3676 1.1 ad goto RETURN; 3677 1.1 ad } 3678 1.1 ad 3679 1.1 ad num_size = num * size; 3680 1.57 ad if (__predict_false(num_size == 0)) { 3681 1.57 ad if (NOT_OPT(sysv) && ((num == 0) || (size == 0))) 3682 1.1 ad num_size = 1; 3683 1.1 ad else { 3684 1.1 ad ret = NULL; 3685 1.1 ad goto RETURN; 3686 1.1 ad } 3687 1.1 ad /* 3688 1.1 ad * Try to avoid division here. We know that it isn't possible to 3689 1.1 ad * overflow during multiplication if neither operand uses any of the 3690 1.1 ad * most significant half of the bits in a size_t. 3691 1.1 ad */ 3692 1.2 ad } else if ((unsigned long long)((num | size) & 3693 1.2 ad ((unsigned long long)SIZE_T_MAX << (sizeof(size_t) << 2))) && 3694 1.2 ad (num_size / size != num)) { 3695 1.1 ad /* size_t overflow. */ 3696 1.1 ad ret = NULL; 3697 1.1 ad goto RETURN; 3698 1.1 ad } 3699 1.1 ad 3700 1.1 ad ret = icalloc(num_size); 3701 1.1 ad 3702 1.1 ad RETURN: 3703 1.63 ad if (__predict_false(ret == NULL)) { 3704 1.57 ad if (OPT(xmalloc)) { 3705 1.16 christos _malloc_message(getprogname(), 3706 1.1 ad ": (malloc) Error in calloc(): out of memory\n", "", 3707 1.1 ad ""); 3708 1.1 ad abort(); 3709 1.1 ad } 3710 1.1 ad errno = ENOMEM; 3711 1.1 ad } 3712 1.1 ad 3713 1.1 ad UTRACE(0, num_size, ret); 3714 1.1 ad return (ret); 3715 1.1 ad } 3716 1.1 ad 3717 1.1 ad void * 3718 1.1 ad realloc(void *ptr, size_t size) 3719 1.1 ad { 3720 1.1 ad void *ret; 3721 1.1 ad 3722 1.57 ad if (__predict_false(size == 0)) { 3723 1.57 ad if (NOT_OPT(sysv)) 3724 1.1 ad size = 1; 3725 1.1 ad else { 3726 1.1 ad if (ptr != NULL) 3727 1.1 ad idalloc(ptr); 3728 1.1 ad ret = NULL; 3729 1.1 ad goto RETURN; 3730 1.1 ad } 3731 1.1 ad } 3732 1.1 ad 3733 1.57 ad if (__predict_true(ptr != NULL)) { 3734 1.1 ad assert(malloc_initialized); 3735 1.1 ad 3736 1.1 ad ret = iralloc(ptr, size); 3737 1.1 ad 3738 1.63 ad if (__predict_false(ret == NULL)) { 3739 1.57 ad if (OPT(xmalloc)) { 3740 1.16 christos _malloc_message(getprogname(), 3741 1.1 ad ": (malloc) Error in realloc(): out of " 3742 1.1 ad "memory\n", "", ""); 3743 1.1 ad abort(); 3744 1.1 ad } 3745 1.1 ad errno = ENOMEM; 3746 1.1 ad } 3747 1.1 ad } else { 3748 1.57 ad if (__predict_false(malloc_init())) 3749 1.1 ad ret = NULL; 3750 1.1 ad else 3751 1.1 ad ret = imalloc(size); 3752 1.1 ad 3753 1.63 ad if (__predict_false(ret == NULL)) { 3754 1.57 ad if (OPT(xmalloc)) { 3755 1.16 christos _malloc_message(getprogname(), 3756 1.1 ad ": (malloc) Error in realloc(): out of " 3757 1.1 ad "memory\n", "", ""); 3758 1.1 ad abort(); 3759 1.1 ad } 3760 1.1 ad errno = ENOMEM; 3761 1.1 ad } 3762 1.1 ad } 3763 1.1 ad 3764 1.1 ad RETURN: 3765 1.1 ad UTRACE(ptr, size, ret); 3766 1.1 ad return (ret); 3767 1.1 ad } 3768 1.1 ad 3769 1.1 ad void 3770 1.1 ad free(void *ptr) 3771 1.1 ad { 3772 1.1 ad 3773 1.1 ad UTRACE(ptr, 0, 0); 3774 1.63 ad if (__predict_true(ptr != NULL)) { 3775 1.1 ad assert(malloc_initialized); 3776 1.1 ad 3777 1.1 ad idalloc(ptr); 3778 1.1 ad } 3779 1.1 ad } 3780 1.1 ad 3781 1.1 ad /* 3782 1.1 ad * End malloc(3)-compatible functions. 3783 1.1 ad */ 3784 1.1 ad /******************************************************************************/ 3785 1.1 ad /* 3786 1.1 ad * Begin non-standard functions. 3787 1.1 ad */ 3788 1.1 ad size_t 3789 1.1 ad malloc_usable_size(const void *ptr) 3790 1.1 ad { 3791 1.1 ad 3792 1.1 ad assert(ptr != NULL); 3793 1.1 ad 3794 1.1 ad return (isalloc(ptr)); 3795 1.1 ad } 3796 1.1 ad 3797 1.1 ad /* 3798 1.1 ad * End non-standard functions. 3799 1.1 ad */ 3800 1.1 ad /******************************************************************************/ 3801 1.1 ad /* 3802 1.1 ad * Begin library-private functions, used by threading libraries for protection 3803 1.1 ad * of malloc during fork(). These functions are only called if the program is 3804 1.1 ad * running in threaded mode, so there is no need to check whether the program 3805 1.1 ad * is threaded here. 3806 1.1 ad */ 3807 1.1 ad 3808 1.1 ad void 3809 1.1 ad _malloc_prefork(void) 3810 1.1 ad { 3811 1.1 ad unsigned i; 3812 1.1 ad 3813 1.1 ad /* Acquire all mutexes in a safe order. */ 3814 1.48 joerg malloc_mutex_lock(&init_lock); 3815 1.1 ad malloc_mutex_lock(&arenas_mtx); 3816 1.57 ad for (i = 0; i < ncpus; i++) { 3817 1.1 ad if (arenas[i] != NULL) 3818 1.1 ad malloc_mutex_lock(&arenas[i]->mtx); 3819 1.1 ad } 3820 1.48 joerg malloc_mutex_lock(&chunks_mtx); 3821 1.1 ad malloc_mutex_lock(&base_mtx); 3822 1.48 joerg #ifdef USE_BRK 3823 1.48 joerg malloc_mutex_lock(&brk_mtx); 3824 1.48 joerg #endif 3825 1.1 ad } 3826 1.1 ad 3827 1.1 ad void 3828 1.1 ad _malloc_postfork(void) 3829 1.1 ad { 3830 1.1 ad unsigned i; 3831 1.1 ad 3832 1.1 ad /* Release all mutexes, now that fork() has completed. */ 3833 1.48 joerg #ifdef USE_BRK 3834 1.48 joerg malloc_mutex_unlock(&brk_mtx); 3835 1.48 joerg #endif 3836 1.48 joerg malloc_mutex_unlock(&base_mtx); 3837 1.1 ad malloc_mutex_unlock(&chunks_mtx); 3838 1.1 ad 3839 1.57 ad for (i = ncpus; i-- > 0; ) { 3840 1.1 ad if (arenas[i] != NULL) 3841 1.1 ad malloc_mutex_unlock(&arenas[i]->mtx); 3842 1.1 ad } 3843 1.1 ad malloc_mutex_unlock(&arenas_mtx); 3844 1.48 joerg malloc_mutex_unlock(&init_lock); 3845 1.1 ad } 3846 1.1 ad 3847 1.53 joerg void 3848 1.53 joerg _malloc_postfork_child(void) 3849 1.53 joerg { 3850 1.53 joerg unsigned i; 3851 1.53 joerg 3852 1.53 joerg /* Release all mutexes, now that fork() has completed. */ 3853 1.53 joerg #ifdef USE_BRK 3854 1.53 joerg malloc_mutex_init(&brk_mtx); 3855 1.53 joerg #endif 3856 1.53 joerg malloc_mutex_init(&base_mtx); 3857 1.53 joerg malloc_mutex_init(&chunks_mtx); 3858 1.53 joerg 3859 1.57 ad for (i = ncpus; i-- > 0; ) { 3860 1.53 joerg if (arenas[i] != NULL) 3861 1.53 joerg malloc_mutex_init(&arenas[i]->mtx); 3862 1.53 joerg } 3863 1.53 joerg malloc_mutex_init(&arenas_mtx); 3864 1.53 joerg malloc_mutex_init(&init_lock); 3865 1.53 joerg } 3866 1.53 joerg 3867 1.1 ad /* 3868 1.1 ad * End library-private functions. 3869 1.1 ad */ 3870 1.1 ad /******************************************************************************/ 3871