1 /* $NetBSD: jemalloc.c,v 1.20 2009/02/12 03:11:01 lukem Exp $ */ 2 3 /*- 4 * Copyright (C) 2006,2007 Jason Evans <jasone@FreeBSD.org>. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice(s), this list of conditions and the following disclaimer as 12 * the first lines of this file unmodified other than the possible 13 * addition of one or more copyright notices. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice(s), this list of conditions and the following disclaimer in 16 * the documentation and/or other materials provided with the 17 * distribution. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 20 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE 23 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR 26 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 27 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE 28 * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, 29 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 30 * 31 ******************************************************************************* 32 * 33 * This allocator implementation is designed to provide scalable performance 34 * for multi-threaded programs on multi-processor systems. The following 35 * features are included for this purpose: 36 * 37 * + Multiple arenas are used if there are multiple CPUs, which reduces lock 38 * contention and cache sloshing. 39 * 40 * + Cache line sharing between arenas is avoided for internal data 41 * structures. 42 * 43 * + Memory is managed in chunks and runs (chunks can be split into runs), 44 * rather than as individual pages. This provides a constant-time 45 * mechanism for associating allocations with particular arenas. 46 * 47 * Allocation requests are rounded up to the nearest size class, and no record 48 * of the original request size is maintained. Allocations are broken into 49 * categories according to size class. Assuming runtime defaults, 4 kB pages 50 * and a 16 byte quantum, the size classes in each category are as follows: 51 * 52 * |=====================================| 53 * | Category | Subcategory | Size | 54 * |=====================================| 55 * | Small | Tiny | 2 | 56 * | | | 4 | 57 * | | | 8 | 58 * | |----------------+---------| 59 * | | Quantum-spaced | 16 | 60 * | | | 32 | 61 * | | | 48 | 62 * | | | ... | 63 * | | | 480 | 64 * | | | 496 | 65 * | | | 512 | 66 * | |----------------+---------| 67 * | | Sub-page | 1 kB | 68 * | | | 2 kB | 69 * |=====================================| 70 * | Large | 4 kB | 71 * | | 8 kB | 72 * | | 12 kB | 73 * | | ... | 74 * | | 1012 kB | 75 * | | 1016 kB | 76 * | | 1020 kB | 77 * |=====================================| 78 * | Huge | 1 MB | 79 * | | 2 MB | 80 * | | 3 MB | 81 * | | ... | 82 * |=====================================| 83 * 84 * A different mechanism is used for each category: 85 * 86 * Small : Each size class is segregated into its own set of runs. Each run 87 * maintains a bitmap of which regions are free/allocated. 88 * 89 * Large : Each allocation is backed by a dedicated run. Metadata are stored 90 * in the associated arena chunk header maps. 91 * 92 * Huge : Each allocation is backed by a dedicated contiguous set of chunks. 93 * Metadata are stored in a separate red-black tree. 94 * 95 ******************************************************************************* 96 */ 97 98 /* LINTLIBRARY */ 99 100 #ifdef __NetBSD__ 101 # define xutrace(a, b) utrace("malloc", (a), (b)) 102 # define __DECONST(x, y) ((x)__UNCONST(y)) 103 # define NO_TLS 104 #else 105 # define xutrace(a, b) utrace((a), (b)) 106 #endif /* __NetBSD__ */ 107 108 /* 109 * MALLOC_PRODUCTION disables assertions and statistics gathering. It also 110 * defaults the A and J runtime options to off. These settings are appropriate 111 * for production systems. 112 */ 113 #define MALLOC_PRODUCTION 114 115 #ifndef MALLOC_PRODUCTION 116 # define MALLOC_DEBUG 117 #endif 118 119 #include <sys/cdefs.h> 120 /* __FBSDID("$FreeBSD: src/lib/libc/stdlib/malloc.c,v 1.147 2007/06/15 22:00:16 jasone Exp $"); */ 121 __RCSID("$NetBSD: jemalloc.c,v 1.20 2009/02/12 03:11:01 lukem Exp $"); 122 123 #ifdef __FreeBSD__ 124 #include "libc_private.h" 125 #ifdef MALLOC_DEBUG 126 # define _LOCK_DEBUG 127 #endif 128 #include "spinlock.h" 129 #endif 130 #include "namespace.h" 131 #include <sys/mman.h> 132 #include <sys/param.h> 133 #ifdef __FreeBSD__ 134 #include <sys/stddef.h> 135 #endif 136 #include <sys/time.h> 137 #include <sys/types.h> 138 #include <sys/sysctl.h> 139 #include <sys/tree.h> 140 #include <sys/uio.h> 141 #include <sys/ktrace.h> /* Must come after several other sys/ includes. */ 142 143 #ifdef __FreeBSD__ 144 #include <machine/atomic.h> 145 #include <machine/cpufunc.h> 146 #endif 147 #include <machine/vmparam.h> 148 149 #include <errno.h> 150 #include <limits.h> 151 #include <pthread.h> 152 #include <sched.h> 153 #include <stdarg.h> 154 #include <stdbool.h> 155 #include <stdio.h> 156 #include <stdint.h> 157 #include <stdlib.h> 158 #include <string.h> 159 #include <strings.h> 160 #include <unistd.h> 161 162 #ifdef __NetBSD__ 163 # include <reentrant.h> 164 # include "extern.h" 165 166 #define STRERROR_R(a, b, c) __strerror_r(a, b, c); 167 /* 168 * A non localized version of strerror, that avoids bringing in 169 * stdio and the locale code. All the malloc messages are in English 170 * so why bother? 171 */ 172 static int 173 __strerror_r(int e, char *s, size_t l) 174 { 175 int rval; 176 size_t slen; 177 178 if (e >= 0 && e < sys_nerr) { 179 slen = strlcpy(s, sys_errlist[e], l); 180 rval = 0; 181 } else { 182 slen = snprintf_ss(s, l, "Unknown error %u", e); 183 rval = EINVAL; 184 } 185 return slen >= l ? ERANGE : rval; 186 } 187 #endif 188 189 #ifdef __FreeBSD__ 190 #define STRERROR_R(a, b, c) strerror_r(a, b, c); 191 #include "un-namespace.h" 192 #endif 193 194 /* MALLOC_STATS enables statistics calculation. */ 195 #ifndef MALLOC_PRODUCTION 196 # define MALLOC_STATS 197 #endif 198 199 #ifdef MALLOC_DEBUG 200 # ifdef NDEBUG 201 # undef NDEBUG 202 # endif 203 #else 204 # ifndef NDEBUG 205 # define NDEBUG 206 # endif 207 #endif 208 #include <assert.h> 209 210 #ifdef MALLOC_DEBUG 211 /* Disable inlining to make debugging easier. */ 212 # define inline 213 #endif 214 215 /* Size of stack-allocated buffer passed to strerror_r(). */ 216 #define STRERROR_BUF 64 217 218 /* Minimum alignment of allocations is 2^QUANTUM_2POW_MIN bytes. */ 219 #ifdef __i386__ 220 # define QUANTUM_2POW_MIN 4 221 # define SIZEOF_PTR_2POW 2 222 # define USE_BRK 223 #endif 224 #ifdef __ia64__ 225 # define QUANTUM_2POW_MIN 4 226 # define SIZEOF_PTR_2POW 3 227 #endif 228 #ifdef __alpha__ 229 # define QUANTUM_2POW_MIN 4 230 # define SIZEOF_PTR_2POW 3 231 # define NO_TLS 232 #endif 233 #ifdef __sparc64__ 234 # define QUANTUM_2POW_MIN 4 235 # define SIZEOF_PTR_2POW 3 236 # define NO_TLS 237 #endif 238 #ifdef __amd64__ 239 # define QUANTUM_2POW_MIN 4 240 # define SIZEOF_PTR_2POW 3 241 #endif 242 #ifdef __arm__ 243 # define QUANTUM_2POW_MIN 3 244 # define SIZEOF_PTR_2POW 2 245 # define USE_BRK 246 # define NO_TLS 247 #endif 248 #ifdef __powerpc__ 249 # define QUANTUM_2POW_MIN 4 250 # define SIZEOF_PTR_2POW 2 251 # define USE_BRK 252 #endif 253 #if defined(__sparc__) && !defined(__sparc64__) 254 # define QUANTUM_2POW_MIN 4 255 # define SIZEOF_PTR_2POW 2 256 # define USE_BRK 257 #endif 258 #ifdef __vax__ 259 # define QUANTUM_2POW_MIN 4 260 # define SIZEOF_PTR_2POW 2 261 # define USE_BRK 262 #endif 263 #ifdef __sh__ 264 # define QUANTUM_2POW_MIN 4 265 # define SIZEOF_PTR_2POW 2 266 # define USE_BRK 267 #endif 268 #ifdef __m68k__ 269 # define QUANTUM_2POW_MIN 4 270 # define SIZEOF_PTR_2POW 2 271 # define USE_BRK 272 #endif 273 #ifdef __mips__ 274 # define QUANTUM_2POW_MIN 4 275 # define SIZEOF_PTR_2POW 2 276 # define USE_BRK 277 #endif 278 #ifdef __hppa__ 279 # define QUANTUM_2POW_MIN 4 280 # define SIZEOF_PTR_2POW 2 281 # define USE_BRK 282 #endif 283 284 #define SIZEOF_PTR (1 << SIZEOF_PTR_2POW) 285 286 /* sizeof(int) == (1 << SIZEOF_INT_2POW). */ 287 #ifndef SIZEOF_INT_2POW 288 # define SIZEOF_INT_2POW 2 289 #endif 290 291 /* We can't use TLS in non-PIC programs, since TLS relies on loader magic. */ 292 #if (!defined(PIC) && !defined(NO_TLS)) 293 # define NO_TLS 294 #endif 295 296 /* 297 * Size and alignment of memory chunks that are allocated by the OS's virtual 298 * memory system. 299 */ 300 #define CHUNK_2POW_DEFAULT 20 301 302 /* 303 * Maximum size of L1 cache line. This is used to avoid cache line aliasing, 304 * so over-estimates are okay (up to a point), but under-estimates will 305 * negatively affect performance. 306 */ 307 #define CACHELINE_2POW 6 308 #define CACHELINE ((size_t)(1 << CACHELINE_2POW)) 309 310 /* Smallest size class to support. */ 311 #define TINY_MIN_2POW 1 312 313 /* 314 * Maximum size class that is a multiple of the quantum, but not (necessarily) 315 * a power of 2. Above this size, allocations are rounded up to the nearest 316 * power of 2. 317 */ 318 #define SMALL_MAX_2POW_DEFAULT 9 319 #define SMALL_MAX_DEFAULT (1 << SMALL_MAX_2POW_DEFAULT) 320 321 /* 322 * Maximum desired run header overhead. Runs are sized as small as possible 323 * such that this setting is still honored, without violating other constraints. 324 * The goal is to make runs as small as possible without exceeding a per run 325 * external fragmentation threshold. 326 * 327 * Note that it is possible to set this low enough that it cannot be honored 328 * for some/all object sizes, since there is one bit of header overhead per 329 * object (plus a constant). In such cases, this constraint is relaxed. 330 * 331 * RUN_MAX_OVRHD_RELAX specifies the maximum number of bits per region of 332 * overhead for which RUN_MAX_OVRHD is relaxed. 333 */ 334 #define RUN_MAX_OVRHD 0.015 335 #define RUN_MAX_OVRHD_RELAX 1.5 336 337 /* Put a cap on small object run size. This overrides RUN_MAX_OVRHD. */ 338 #define RUN_MAX_SMALL_2POW 15 339 #define RUN_MAX_SMALL (1 << RUN_MAX_SMALL_2POW) 340 341 /******************************************************************************/ 342 343 #ifdef __FreeBSD__ 344 /* 345 * Mutexes based on spinlocks. We can't use normal pthread mutexes, because 346 * they require malloc()ed memory. 347 */ 348 typedef struct { 349 spinlock_t lock; 350 } malloc_mutex_t; 351 352 /* Set to true once the allocator has been initialized. */ 353 static bool malloc_initialized = false; 354 355 /* Used to avoid initialization races. */ 356 static malloc_mutex_t init_lock = {_SPINLOCK_INITIALIZER}; 357 #else 358 #define malloc_mutex_t mutex_t 359 360 /* Set to true once the allocator has been initialized. */ 361 static bool malloc_initialized = false; 362 363 /* Used to avoid initialization races. */ 364 static mutex_t init_lock = MUTEX_INITIALIZER; 365 #endif 366 367 /******************************************************************************/ 368 /* 369 * Statistics data structures. 370 */ 371 372 #ifdef MALLOC_STATS 373 374 typedef struct malloc_bin_stats_s malloc_bin_stats_t; 375 struct malloc_bin_stats_s { 376 /* 377 * Number of allocation requests that corresponded to the size of this 378 * bin. 379 */ 380 uint64_t nrequests; 381 382 /* Total number of runs created for this bin's size class. */ 383 uint64_t nruns; 384 385 /* 386 * Total number of runs reused by extracting them from the runs tree for 387 * this bin's size class. 388 */ 389 uint64_t reruns; 390 391 /* High-water mark for this bin. */ 392 unsigned long highruns; 393 394 /* Current number of runs in this bin. */ 395 unsigned long curruns; 396 }; 397 398 typedef struct arena_stats_s arena_stats_t; 399 struct arena_stats_s { 400 /* Number of bytes currently mapped. */ 401 size_t mapped; 402 403 /* Per-size-category statistics. */ 404 size_t allocated_small; 405 uint64_t nmalloc_small; 406 uint64_t ndalloc_small; 407 408 size_t allocated_large; 409 uint64_t nmalloc_large; 410 uint64_t ndalloc_large; 411 }; 412 413 typedef struct chunk_stats_s chunk_stats_t; 414 struct chunk_stats_s { 415 /* Number of chunks that were allocated. */ 416 uint64_t nchunks; 417 418 /* High-water mark for number of chunks allocated. */ 419 unsigned long highchunks; 420 421 /* 422 * Current number of chunks allocated. This value isn't maintained for 423 * any other purpose, so keep track of it in order to be able to set 424 * highchunks. 425 */ 426 unsigned long curchunks; 427 }; 428 429 #endif /* #ifdef MALLOC_STATS */ 430 431 /******************************************************************************/ 432 /* 433 * Chunk data structures. 434 */ 435 436 /* Tree of chunks. */ 437 typedef struct chunk_node_s chunk_node_t; 438 struct chunk_node_s { 439 /* Linkage for the chunk tree. */ 440 RB_ENTRY(chunk_node_s) link; 441 442 /* 443 * Pointer to the chunk that this tree node is responsible for. In some 444 * (but certainly not all) cases, this data structure is placed at the 445 * beginning of the corresponding chunk, so this field may point to this 446 * node. 447 */ 448 void *chunk; 449 450 /* Total chunk size. */ 451 size_t size; 452 }; 453 typedef struct chunk_tree_s chunk_tree_t; 454 RB_HEAD(chunk_tree_s, chunk_node_s); 455 456 /******************************************************************************/ 457 /* 458 * Arena data structures. 459 */ 460 461 typedef struct arena_s arena_t; 462 typedef struct arena_bin_s arena_bin_t; 463 464 typedef struct arena_chunk_map_s arena_chunk_map_t; 465 struct arena_chunk_map_s { 466 /* Number of pages in run. */ 467 uint32_t npages; 468 /* 469 * Position within run. For a free run, this is POS_FREE for the first 470 * and last pages. The POS_FREE special value makes it possible to 471 * quickly coalesce free runs. 472 * 473 * This is the limiting factor for chunksize; there can be at most 2^31 474 * pages in a run. 475 */ 476 #define POS_FREE ((uint32_t)0xffffffffU) 477 uint32_t pos; 478 }; 479 480 /* Arena chunk header. */ 481 typedef struct arena_chunk_s arena_chunk_t; 482 struct arena_chunk_s { 483 /* Arena that owns the chunk. */ 484 arena_t *arena; 485 486 /* Linkage for the arena's chunk tree. */ 487 RB_ENTRY(arena_chunk_s) link; 488 489 /* 490 * Number of pages in use. This is maintained in order to make 491 * detection of empty chunks fast. 492 */ 493 uint32_t pages_used; 494 495 /* 496 * Every time a free run larger than this value is created/coalesced, 497 * this value is increased. The only way that the value decreases is if 498 * arena_run_alloc() fails to find a free run as large as advertised by 499 * this value. 500 */ 501 uint32_t max_frun_npages; 502 503 /* 504 * Every time a free run that starts at an earlier page than this value 505 * is created/coalesced, this value is decreased. It is reset in a 506 * similar fashion to max_frun_npages. 507 */ 508 uint32_t min_frun_ind; 509 510 /* 511 * Map of pages within chunk that keeps track of free/large/small. For 512 * free runs, only the map entries for the first and last pages are 513 * kept up to date, so that free runs can be quickly coalesced. 514 */ 515 arena_chunk_map_t map[1]; /* Dynamically sized. */ 516 }; 517 typedef struct arena_chunk_tree_s arena_chunk_tree_t; 518 RB_HEAD(arena_chunk_tree_s, arena_chunk_s); 519 520 typedef struct arena_run_s arena_run_t; 521 struct arena_run_s { 522 /* Linkage for run trees. */ 523 RB_ENTRY(arena_run_s) link; 524 525 #ifdef MALLOC_DEBUG 526 uint32_t magic; 527 # define ARENA_RUN_MAGIC 0x384adf93 528 #endif 529 530 /* Bin this run is associated with. */ 531 arena_bin_t *bin; 532 533 /* Index of first element that might have a free region. */ 534 unsigned regs_minelm; 535 536 /* Number of free regions in run. */ 537 unsigned nfree; 538 539 /* Bitmask of in-use regions (0: in use, 1: free). */ 540 unsigned regs_mask[1]; /* Dynamically sized. */ 541 }; 542 typedef struct arena_run_tree_s arena_run_tree_t; 543 RB_HEAD(arena_run_tree_s, arena_run_s); 544 545 struct arena_bin_s { 546 /* 547 * Current run being used to service allocations of this bin's size 548 * class. 549 */ 550 arena_run_t *runcur; 551 552 /* 553 * Tree of non-full runs. This tree is used when looking for an 554 * existing run when runcur is no longer usable. We choose the 555 * non-full run that is lowest in memory; this policy tends to keep 556 * objects packed well, and it can also help reduce the number of 557 * almost-empty chunks. 558 */ 559 arena_run_tree_t runs; 560 561 /* Size of regions in a run for this bin's size class. */ 562 size_t reg_size; 563 564 /* Total size of a run for this bin's size class. */ 565 size_t run_size; 566 567 /* Total number of regions in a run for this bin's size class. */ 568 uint32_t nregs; 569 570 /* Number of elements in a run's regs_mask for this bin's size class. */ 571 uint32_t regs_mask_nelms; 572 573 /* Offset of first region in a run for this bin's size class. */ 574 uint32_t reg0_offset; 575 576 #ifdef MALLOC_STATS 577 /* Bin statistics. */ 578 malloc_bin_stats_t stats; 579 #endif 580 }; 581 582 struct arena_s { 583 #ifdef MALLOC_DEBUG 584 uint32_t magic; 585 # define ARENA_MAGIC 0x947d3d24 586 #endif 587 588 /* All operations on this arena require that mtx be locked. */ 589 malloc_mutex_t mtx; 590 591 #ifdef MALLOC_STATS 592 arena_stats_t stats; 593 #endif 594 595 /* 596 * Tree of chunks this arena manages. 597 */ 598 arena_chunk_tree_t chunks; 599 600 /* 601 * In order to avoid rapid chunk allocation/deallocation when an arena 602 * oscillates right on the cusp of needing a new chunk, cache the most 603 * recently freed chunk. This caching is disabled by opt_hint. 604 * 605 * There is one spare chunk per arena, rather than one spare total, in 606 * order to avoid interactions between multiple threads that could make 607 * a single spare inadequate. 608 */ 609 arena_chunk_t *spare; 610 611 /* 612 * bins is used to store rings of free regions of the following sizes, 613 * assuming a 16-byte quantum, 4kB pagesize, and default MALLOC_OPTIONS. 614 * 615 * bins[i] | size | 616 * --------+------+ 617 * 0 | 2 | 618 * 1 | 4 | 619 * 2 | 8 | 620 * --------+------+ 621 * 3 | 16 | 622 * 4 | 32 | 623 * 5 | 48 | 624 * 6 | 64 | 625 * : : 626 * : : 627 * 33 | 496 | 628 * 34 | 512 | 629 * --------+------+ 630 * 35 | 1024 | 631 * 36 | 2048 | 632 * --------+------+ 633 */ 634 arena_bin_t bins[1]; /* Dynamically sized. */ 635 }; 636 637 /******************************************************************************/ 638 /* 639 * Data. 640 */ 641 642 /* Number of CPUs. */ 643 static unsigned ncpus; 644 645 /* VM page size. */ 646 static size_t pagesize; 647 static size_t pagesize_mask; 648 static int pagesize_2pow; 649 650 /* Various bin-related settings. */ 651 static size_t bin_maxclass; /* Max size class for bins. */ 652 static unsigned ntbins; /* Number of (2^n)-spaced tiny bins. */ 653 static unsigned nqbins; /* Number of quantum-spaced bins. */ 654 static unsigned nsbins; /* Number of (2^n)-spaced sub-page bins. */ 655 static size_t small_min; 656 static size_t small_max; 657 658 /* Various quantum-related settings. */ 659 static size_t quantum; 660 static size_t quantum_mask; /* (quantum - 1). */ 661 662 /* Various chunk-related settings. */ 663 static size_t chunksize; 664 static size_t chunksize_mask; /* (chunksize - 1). */ 665 static int chunksize_2pow; 666 static unsigned chunk_npages; 667 static unsigned arena_chunk_header_npages; 668 static size_t arena_maxclass; /* Max size class for arenas. */ 669 670 /********/ 671 /* 672 * Chunks. 673 */ 674 675 /* Protects chunk-related data structures. */ 676 static malloc_mutex_t chunks_mtx; 677 678 /* Tree of chunks that are stand-alone huge allocations. */ 679 static chunk_tree_t huge; 680 681 #ifdef USE_BRK 682 /* 683 * Try to use brk for chunk-size allocations, due to address space constraints. 684 */ 685 /* 686 * Protects sbrk() calls. This must be separate from chunks_mtx, since 687 * base_pages_alloc() also uses sbrk(), but cannot lock chunks_mtx (doing so 688 * could cause recursive lock acquisition). 689 */ 690 static malloc_mutex_t brk_mtx; 691 /* Result of first sbrk(0) call. */ 692 static void *brk_base; 693 /* Current end of brk, or ((void *)-1) if brk is exhausted. */ 694 static void *brk_prev; 695 /* Current upper limit on brk addresses. */ 696 static void *brk_max; 697 #endif 698 699 #ifdef MALLOC_STATS 700 /* Huge allocation statistics. */ 701 static uint64_t huge_nmalloc; 702 static uint64_t huge_ndalloc; 703 static uint64_t huge_nralloc; 704 static size_t huge_allocated; 705 #endif 706 707 /* 708 * Tree of chunks that were previously allocated. This is used when allocating 709 * chunks, in an attempt to re-use address space. 710 */ 711 static chunk_tree_t old_chunks; 712 713 /****************************/ 714 /* 715 * base (internal allocation). 716 */ 717 718 /* 719 * Current pages that are being used for internal memory allocations. These 720 * pages are carved up in cacheline-size quanta, so that there is no chance of 721 * false cache line sharing. 722 */ 723 static void *base_pages; 724 static void *base_next_addr; 725 static void *base_past_addr; /* Addr immediately past base_pages. */ 726 static chunk_node_t *base_chunk_nodes; /* LIFO cache of chunk nodes. */ 727 static malloc_mutex_t base_mtx; 728 #ifdef MALLOC_STATS 729 static size_t base_mapped; 730 #endif 731 732 /********/ 733 /* 734 * Arenas. 735 */ 736 737 /* 738 * Arenas that are used to service external requests. Not all elements of the 739 * arenas array are necessarily used; arenas are created lazily as needed. 740 */ 741 static arena_t **arenas; 742 static unsigned narenas; 743 static unsigned next_arena; 744 static malloc_mutex_t arenas_mtx; /* Protects arenas initialization. */ 745 746 #ifndef NO_TLS 747 /* 748 * Map of pthread_self() --> arenas[???], used for selecting an arena to use 749 * for allocations. 750 */ 751 static __thread arena_t *arenas_map; 752 #define get_arenas_map() (arenas_map) 753 #define set_arenas_map(x) (arenas_map = x) 754 #else 755 static thread_key_t arenas_map_key; 756 #define get_arenas_map() thr_getspecific(arenas_map_key) 757 #define set_arenas_map(x) thr_setspecific(arenas_map_key, x) 758 #endif 759 760 #ifdef MALLOC_STATS 761 /* Chunk statistics. */ 762 static chunk_stats_t stats_chunks; 763 #endif 764 765 /*******************************/ 766 /* 767 * Runtime configuration options. 768 */ 769 const char *_malloc_options; 770 771 #ifndef MALLOC_PRODUCTION 772 static bool opt_abort = true; 773 static bool opt_junk = true; 774 #else 775 static bool opt_abort = false; 776 static bool opt_junk = false; 777 #endif 778 static bool opt_hint = false; 779 static bool opt_print_stats = false; 780 static int opt_quantum_2pow = QUANTUM_2POW_MIN; 781 static int opt_small_max_2pow = SMALL_MAX_2POW_DEFAULT; 782 static int opt_chunk_2pow = CHUNK_2POW_DEFAULT; 783 static bool opt_utrace = false; 784 static bool opt_sysv = false; 785 static bool opt_xmalloc = false; 786 static bool opt_zero = false; 787 static int32_t opt_narenas_lshift = 0; 788 789 typedef struct { 790 void *p; 791 size_t s; 792 void *r; 793 } malloc_utrace_t; 794 795 #define UTRACE(a, b, c) \ 796 if (opt_utrace) { \ 797 malloc_utrace_t ut; \ 798 ut.p = a; \ 799 ut.s = b; \ 800 ut.r = c; \ 801 xutrace(&ut, sizeof(ut)); \ 802 } 803 804 /******************************************************************************/ 805 /* 806 * Begin function prototypes for non-inline static functions. 807 */ 808 809 static void wrtmessage(const char *p1, const char *p2, const char *p3, 810 const char *p4); 811 #ifdef MALLOC_STATS 812 static void malloc_printf(const char *format, ...); 813 #endif 814 static char *umax2s(uintmax_t x, char *s); 815 static bool base_pages_alloc(size_t minsize); 816 static void *base_alloc(size_t size); 817 static chunk_node_t *base_chunk_node_alloc(void); 818 static void base_chunk_node_dealloc(chunk_node_t *node); 819 #ifdef MALLOC_STATS 820 static void stats_print(arena_t *arena); 821 #endif 822 static void *pages_map(void *addr, size_t size); 823 static void *pages_map_align(void *addr, size_t size, int align); 824 static void pages_unmap(void *addr, size_t size); 825 static void *chunk_alloc(size_t size); 826 static void chunk_dealloc(void *chunk, size_t size); 827 static void arena_run_split(arena_t *arena, arena_run_t *run, size_t size); 828 static arena_chunk_t *arena_chunk_alloc(arena_t *arena); 829 static void arena_chunk_dealloc(arena_t *arena, arena_chunk_t *chunk); 830 static arena_run_t *arena_run_alloc(arena_t *arena, size_t size); 831 static void arena_run_dalloc(arena_t *arena, arena_run_t *run, size_t size); 832 static arena_run_t *arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin); 833 static void *arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin); 834 static size_t arena_bin_run_size_calc(arena_bin_t *bin, size_t min_run_size); 835 static void *arena_malloc(arena_t *arena, size_t size); 836 static void *arena_palloc(arena_t *arena, size_t alignment, size_t size, 837 size_t alloc_size); 838 static size_t arena_salloc(const void *ptr); 839 static void *arena_ralloc(void *ptr, size_t size, size_t oldsize); 840 static void arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr); 841 static bool arena_new(arena_t *arena); 842 static arena_t *arenas_extend(unsigned ind); 843 static void *huge_malloc(size_t size); 844 static void *huge_palloc(size_t alignment, size_t size); 845 static void *huge_ralloc(void *ptr, size_t size, size_t oldsize); 846 static void huge_dalloc(void *ptr); 847 static void *imalloc(size_t size); 848 static void *ipalloc(size_t alignment, size_t size); 849 static void *icalloc(size_t size); 850 static size_t isalloc(const void *ptr); 851 static void *iralloc(void *ptr, size_t size); 852 static void idalloc(void *ptr); 853 static void malloc_print_stats(void); 854 static bool malloc_init_hard(void); 855 856 /* 857 * End function prototypes. 858 */ 859 /******************************************************************************/ 860 /* 861 * Begin mutex. 862 */ 863 864 #ifdef __NetBSD__ 865 #define malloc_mutex_init(m) mutex_init(m, NULL) 866 #define malloc_mutex_lock(m) mutex_lock(m) 867 #define malloc_mutex_unlock(m) mutex_unlock(m) 868 #else /* __NetBSD__ */ 869 static inline void 870 malloc_mutex_init(malloc_mutex_t *a_mutex) 871 { 872 static const spinlock_t lock = _SPINLOCK_INITIALIZER; 873 874 a_mutex->lock = lock; 875 } 876 877 static inline void 878 malloc_mutex_lock(malloc_mutex_t *a_mutex) 879 { 880 881 if (__isthreaded) 882 _SPINLOCK(&a_mutex->lock); 883 } 884 885 static inline void 886 malloc_mutex_unlock(malloc_mutex_t *a_mutex) 887 { 888 889 if (__isthreaded) 890 _SPINUNLOCK(&a_mutex->lock); 891 } 892 #endif /* __NetBSD__ */ 893 894 /* 895 * End mutex. 896 */ 897 /******************************************************************************/ 898 /* 899 * Begin Utility functions/macros. 900 */ 901 902 /* Return the chunk address for allocation address a. */ 903 #define CHUNK_ADDR2BASE(a) \ 904 ((void *)((uintptr_t)(a) & ~chunksize_mask)) 905 906 /* Return the chunk offset of address a. */ 907 #define CHUNK_ADDR2OFFSET(a) \ 908 ((size_t)((uintptr_t)(a) & chunksize_mask)) 909 910 /* Return the smallest chunk multiple that is >= s. */ 911 #define CHUNK_CEILING(s) \ 912 (((s) + chunksize_mask) & ~chunksize_mask) 913 914 /* Return the smallest cacheline multiple that is >= s. */ 915 #define CACHELINE_CEILING(s) \ 916 (((s) + (CACHELINE - 1)) & ~(CACHELINE - 1)) 917 918 /* Return the smallest quantum multiple that is >= a. */ 919 #define QUANTUM_CEILING(a) \ 920 (((a) + quantum_mask) & ~quantum_mask) 921 922 /* Return the smallest pagesize multiple that is >= s. */ 923 #define PAGE_CEILING(s) \ 924 (((s) + pagesize_mask) & ~pagesize_mask) 925 926 /* Compute the smallest power of 2 that is >= x. */ 927 static inline size_t 928 pow2_ceil(size_t x) 929 { 930 931 x--; 932 x |= x >> 1; 933 x |= x >> 2; 934 x |= x >> 4; 935 x |= x >> 8; 936 x |= x >> 16; 937 #if (SIZEOF_PTR == 8) 938 x |= x >> 32; 939 #endif 940 x++; 941 return (x); 942 } 943 944 static void 945 wrtmessage(const char *p1, const char *p2, const char *p3, const char *p4) 946 { 947 948 write(STDERR_FILENO, p1, strlen(p1)); 949 write(STDERR_FILENO, p2, strlen(p2)); 950 write(STDERR_FILENO, p3, strlen(p3)); 951 write(STDERR_FILENO, p4, strlen(p4)); 952 } 953 954 void (*_malloc_message)(const char *p1, const char *p2, const char *p3, 955 const char *p4) = wrtmessage; 956 957 #ifdef MALLOC_STATS 958 /* 959 * Print to stderr in such a way as to (hopefully) avoid memory allocation. 960 */ 961 static void 962 malloc_printf(const char *format, ...) 963 { 964 char buf[4096]; 965 va_list ap; 966 967 va_start(ap, format); 968 vsnprintf(buf, sizeof(buf), format, ap); 969 va_end(ap); 970 _malloc_message(buf, "", "", ""); 971 } 972 #endif 973 974 /* 975 * We don't want to depend on vsnprintf() for production builds, since that can 976 * cause unnecessary bloat for static binaries. umax2s() provides minimal 977 * integer printing functionality, so that malloc_printf() use can be limited to 978 * MALLOC_STATS code. 979 */ 980 #define UMAX2S_BUFSIZE 21 981 static char * 982 umax2s(uintmax_t x, char *s) 983 { 984 unsigned i; 985 986 /* Make sure UMAX2S_BUFSIZE is large enough. */ 987 /* LINTED */ 988 assert(sizeof(uintmax_t) <= 8); 989 990 i = UMAX2S_BUFSIZE - 1; 991 s[i] = '\0'; 992 do { 993 i--; 994 s[i] = "0123456789"[(int)x % 10]; 995 x /= (uintmax_t)10LL; 996 } while (x > 0); 997 998 return (&s[i]); 999 } 1000 1001 /******************************************************************************/ 1002 1003 static bool 1004 base_pages_alloc(size_t minsize) 1005 { 1006 size_t csize = 0; 1007 1008 #ifdef USE_BRK 1009 /* 1010 * Do special brk allocation here, since base allocations don't need to 1011 * be chunk-aligned. 1012 */ 1013 if (brk_prev != (void *)-1) { 1014 void *brk_cur; 1015 intptr_t incr; 1016 1017 if (minsize != 0) 1018 csize = CHUNK_CEILING(minsize); 1019 1020 malloc_mutex_lock(&brk_mtx); 1021 do { 1022 /* Get the current end of brk. */ 1023 brk_cur = sbrk(0); 1024 1025 /* 1026 * Calculate how much padding is necessary to 1027 * chunk-align the end of brk. Don't worry about 1028 * brk_cur not being chunk-aligned though. 1029 */ 1030 incr = (intptr_t)chunksize 1031 - (intptr_t)CHUNK_ADDR2OFFSET(brk_cur); 1032 assert(incr >= 0); 1033 if ((size_t)incr < minsize) 1034 incr += csize; 1035 1036 brk_prev = sbrk(incr); 1037 if (brk_prev == brk_cur) { 1038 /* Success. */ 1039 malloc_mutex_unlock(&brk_mtx); 1040 base_pages = brk_cur; 1041 base_next_addr = base_pages; 1042 base_past_addr = (void *)((uintptr_t)base_pages 1043 + incr); 1044 #ifdef MALLOC_STATS 1045 base_mapped += incr; 1046 #endif 1047 return (false); 1048 } 1049 } while (brk_prev != (void *)-1); 1050 malloc_mutex_unlock(&brk_mtx); 1051 } 1052 if (minsize == 0) { 1053 /* 1054 * Failure during initialization doesn't matter, so avoid 1055 * falling through to the mmap-based page mapping code. 1056 */ 1057 return (true); 1058 } 1059 #endif 1060 assert(minsize != 0); 1061 csize = PAGE_CEILING(minsize); 1062 base_pages = pages_map(NULL, csize); 1063 if (base_pages == NULL) 1064 return (true); 1065 base_next_addr = base_pages; 1066 base_past_addr = (void *)((uintptr_t)base_pages + csize); 1067 #ifdef MALLOC_STATS 1068 base_mapped += csize; 1069 #endif 1070 return (false); 1071 } 1072 1073 static void * 1074 base_alloc(size_t size) 1075 { 1076 void *ret; 1077 size_t csize; 1078 1079 /* Round size up to nearest multiple of the cacheline size. */ 1080 csize = CACHELINE_CEILING(size); 1081 1082 malloc_mutex_lock(&base_mtx); 1083 1084 /* Make sure there's enough space for the allocation. */ 1085 if ((uintptr_t)base_next_addr + csize > (uintptr_t)base_past_addr) { 1086 if (base_pages_alloc(csize)) { 1087 ret = NULL; 1088 goto RETURN; 1089 } 1090 } 1091 1092 /* Allocate. */ 1093 ret = base_next_addr; 1094 base_next_addr = (void *)((uintptr_t)base_next_addr + csize); 1095 1096 RETURN: 1097 malloc_mutex_unlock(&base_mtx); 1098 return (ret); 1099 } 1100 1101 static chunk_node_t * 1102 base_chunk_node_alloc(void) 1103 { 1104 chunk_node_t *ret; 1105 1106 malloc_mutex_lock(&base_mtx); 1107 if (base_chunk_nodes != NULL) { 1108 ret = base_chunk_nodes; 1109 /* LINTED */ 1110 base_chunk_nodes = *(chunk_node_t **)ret; 1111 malloc_mutex_unlock(&base_mtx); 1112 } else { 1113 malloc_mutex_unlock(&base_mtx); 1114 ret = (chunk_node_t *)base_alloc(sizeof(chunk_node_t)); 1115 } 1116 1117 return (ret); 1118 } 1119 1120 static void 1121 base_chunk_node_dealloc(chunk_node_t *node) 1122 { 1123 1124 malloc_mutex_lock(&base_mtx); 1125 /* LINTED */ 1126 *(chunk_node_t **)node = base_chunk_nodes; 1127 base_chunk_nodes = node; 1128 malloc_mutex_unlock(&base_mtx); 1129 } 1130 1131 /******************************************************************************/ 1132 1133 #ifdef MALLOC_STATS 1134 static void 1135 stats_print(arena_t *arena) 1136 { 1137 unsigned i; 1138 int gap_start; 1139 1140 malloc_printf( 1141 " allocated/mapped nmalloc ndalloc\n"); 1142 1143 malloc_printf("small: %12zu %-12s %12llu %12llu\n", 1144 arena->stats.allocated_small, "", arena->stats.nmalloc_small, 1145 arena->stats.ndalloc_small); 1146 malloc_printf("large: %12zu %-12s %12llu %12llu\n", 1147 arena->stats.allocated_large, "", arena->stats.nmalloc_large, 1148 arena->stats.ndalloc_large); 1149 malloc_printf("total: %12zu/%-12zu %12llu %12llu\n", 1150 arena->stats.allocated_small + arena->stats.allocated_large, 1151 arena->stats.mapped, 1152 arena->stats.nmalloc_small + arena->stats.nmalloc_large, 1153 arena->stats.ndalloc_small + arena->stats.ndalloc_large); 1154 1155 malloc_printf("bins: bin size regs pgs requests newruns" 1156 " reruns maxruns curruns\n"); 1157 for (i = 0, gap_start = -1; i < ntbins + nqbins + nsbins; i++) { 1158 if (arena->bins[i].stats.nrequests == 0) { 1159 if (gap_start == -1) 1160 gap_start = i; 1161 } else { 1162 if (gap_start != -1) { 1163 if (i > gap_start + 1) { 1164 /* Gap of more than one size class. */ 1165 malloc_printf("[%u..%u]\n", 1166 gap_start, i - 1); 1167 } else { 1168 /* Gap of one size class. */ 1169 malloc_printf("[%u]\n", gap_start); 1170 } 1171 gap_start = -1; 1172 } 1173 malloc_printf( 1174 "%13u %1s %4u %4u %3u %9llu %9llu" 1175 " %9llu %7lu %7lu\n", 1176 i, 1177 i < ntbins ? "T" : i < ntbins + nqbins ? "Q" : "S", 1178 arena->bins[i].reg_size, 1179 arena->bins[i].nregs, 1180 arena->bins[i].run_size >> pagesize_2pow, 1181 arena->bins[i].stats.nrequests, 1182 arena->bins[i].stats.nruns, 1183 arena->bins[i].stats.reruns, 1184 arena->bins[i].stats.highruns, 1185 arena->bins[i].stats.curruns); 1186 } 1187 } 1188 if (gap_start != -1) { 1189 if (i > gap_start + 1) { 1190 /* Gap of more than one size class. */ 1191 malloc_printf("[%u..%u]\n", gap_start, i - 1); 1192 } else { 1193 /* Gap of one size class. */ 1194 malloc_printf("[%u]\n", gap_start); 1195 } 1196 } 1197 } 1198 #endif 1199 1200 /* 1201 * End Utility functions/macros. 1202 */ 1203 /******************************************************************************/ 1204 /* 1205 * Begin chunk management functions. 1206 */ 1207 1208 #ifndef lint 1209 static inline int 1210 chunk_comp(chunk_node_t *a, chunk_node_t *b) 1211 { 1212 1213 assert(a != NULL); 1214 assert(b != NULL); 1215 1216 if ((uintptr_t)a->chunk < (uintptr_t)b->chunk) 1217 return (-1); 1218 else if (a->chunk == b->chunk) 1219 return (0); 1220 else 1221 return (1); 1222 } 1223 1224 /* Generate red-black tree code for chunks. */ 1225 RB_GENERATE_STATIC(chunk_tree_s, chunk_node_s, link, chunk_comp); 1226 #endif 1227 1228 static void * 1229 pages_map_align(void *addr, size_t size, int align) 1230 { 1231 void *ret; 1232 1233 /* 1234 * We don't use MAP_FIXED here, because it can cause the *replacement* 1235 * of existing mappings, and we only want to create new mappings. 1236 */ 1237 ret = mmap(addr, size, PROT_READ | PROT_WRITE, 1238 MAP_PRIVATE | MAP_ANON | MAP_ALIGNED(align), -1, 0); 1239 assert(ret != NULL); 1240 1241 if (ret == MAP_FAILED) 1242 ret = NULL; 1243 else if (addr != NULL && ret != addr) { 1244 /* 1245 * We succeeded in mapping memory, but not in the right place. 1246 */ 1247 if (munmap(ret, size) == -1) { 1248 char buf[STRERROR_BUF]; 1249 1250 STRERROR_R(errno, buf, sizeof(buf)); 1251 _malloc_message(getprogname(), 1252 ": (malloc) Error in munmap(): ", buf, "\n"); 1253 if (opt_abort) 1254 abort(); 1255 } 1256 ret = NULL; 1257 } 1258 1259 assert(ret == NULL || (addr == NULL && ret != addr) 1260 || (addr != NULL && ret == addr)); 1261 return (ret); 1262 } 1263 1264 static void * 1265 pages_map(void *addr, size_t size) 1266 { 1267 1268 return pages_map_align(addr, size, 0); 1269 } 1270 1271 static void 1272 pages_unmap(void *addr, size_t size) 1273 { 1274 1275 if (munmap(addr, size) == -1) { 1276 char buf[STRERROR_BUF]; 1277 1278 STRERROR_R(errno, buf, sizeof(buf)); 1279 _malloc_message(getprogname(), 1280 ": (malloc) Error in munmap(): ", buf, "\n"); 1281 if (opt_abort) 1282 abort(); 1283 } 1284 } 1285 1286 static void * 1287 chunk_alloc(size_t size) 1288 { 1289 void *ret, *chunk; 1290 chunk_node_t *tchunk, *delchunk; 1291 1292 assert(size != 0); 1293 assert((size & chunksize_mask) == 0); 1294 1295 malloc_mutex_lock(&chunks_mtx); 1296 1297 if (size == chunksize) { 1298 /* 1299 * Check for address ranges that were previously chunks and try 1300 * to use them. 1301 */ 1302 1303 /* LINTED */ 1304 tchunk = RB_MIN(chunk_tree_s, &old_chunks); 1305 while (tchunk != NULL) { 1306 /* Found an address range. Try to recycle it. */ 1307 1308 chunk = tchunk->chunk; 1309 delchunk = tchunk; 1310 /* LINTED */ 1311 tchunk = RB_NEXT(chunk_tree_s, &old_chunks, delchunk); 1312 1313 /* Remove delchunk from the tree. */ 1314 /* LINTED */ 1315 RB_REMOVE(chunk_tree_s, &old_chunks, delchunk); 1316 base_chunk_node_dealloc(delchunk); 1317 1318 #ifdef USE_BRK 1319 if ((uintptr_t)chunk >= (uintptr_t)brk_base 1320 && (uintptr_t)chunk < (uintptr_t)brk_max) { 1321 /* Re-use a previously freed brk chunk. */ 1322 ret = chunk; 1323 goto RETURN; 1324 } 1325 #endif 1326 if ((ret = pages_map(chunk, size)) != NULL) { 1327 /* Success. */ 1328 goto RETURN; 1329 } 1330 } 1331 } 1332 1333 /* 1334 * Try to over-allocate, but allow the OS to place the allocation 1335 * anywhere. Beware of size_t wrap-around. 1336 */ 1337 if (size + chunksize > size) { 1338 if ((ret = pages_map_align(NULL, size, chunksize_2pow)) 1339 != NULL) { 1340 goto RETURN; 1341 } 1342 } 1343 1344 #ifdef USE_BRK 1345 /* 1346 * Try to create allocations in brk, in order to make full use of 1347 * limited address space. 1348 */ 1349 if (brk_prev != (void *)-1) { 1350 void *brk_cur; 1351 intptr_t incr; 1352 1353 /* 1354 * The loop is necessary to recover from races with other 1355 * threads that are using brk for something other than malloc. 1356 */ 1357 malloc_mutex_lock(&brk_mtx); 1358 do { 1359 /* Get the current end of brk. */ 1360 brk_cur = sbrk(0); 1361 1362 /* 1363 * Calculate how much padding is necessary to 1364 * chunk-align the end of brk. 1365 */ 1366 incr = (intptr_t)size 1367 - (intptr_t)CHUNK_ADDR2OFFSET(brk_cur); 1368 if (incr == (intptr_t)size) { 1369 ret = brk_cur; 1370 } else { 1371 ret = (void *)((intptr_t)brk_cur + incr); 1372 incr += size; 1373 } 1374 1375 brk_prev = sbrk(incr); 1376 if (brk_prev == brk_cur) { 1377 /* Success. */ 1378 malloc_mutex_unlock(&brk_mtx); 1379 brk_max = (void *)((intptr_t)ret + size); 1380 goto RETURN; 1381 } 1382 } while (brk_prev != (void *)-1); 1383 malloc_mutex_unlock(&brk_mtx); 1384 } 1385 #endif 1386 1387 /* All strategies for allocation failed. */ 1388 ret = NULL; 1389 RETURN: 1390 if (ret != NULL) { 1391 chunk_node_t key; 1392 /* 1393 * Clean out any entries in old_chunks that overlap with the 1394 * memory we just allocated. 1395 */ 1396 key.chunk = ret; 1397 /* LINTED */ 1398 tchunk = RB_NFIND(chunk_tree_s, &old_chunks, &key); 1399 while (tchunk != NULL 1400 && (uintptr_t)tchunk->chunk >= (uintptr_t)ret 1401 && (uintptr_t)tchunk->chunk < (uintptr_t)ret + size) { 1402 delchunk = tchunk; 1403 /* LINTED */ 1404 tchunk = RB_NEXT(chunk_tree_s, &old_chunks, delchunk); 1405 /* LINTED */ 1406 RB_REMOVE(chunk_tree_s, &old_chunks, delchunk); 1407 base_chunk_node_dealloc(delchunk); 1408 } 1409 1410 } 1411 #ifdef MALLOC_STATS 1412 if (ret != NULL) { 1413 stats_chunks.nchunks += (size / chunksize); 1414 stats_chunks.curchunks += (size / chunksize); 1415 } 1416 if (stats_chunks.curchunks > stats_chunks.highchunks) 1417 stats_chunks.highchunks = stats_chunks.curchunks; 1418 #endif 1419 malloc_mutex_unlock(&chunks_mtx); 1420 1421 assert(CHUNK_ADDR2BASE(ret) == ret); 1422 return (ret); 1423 } 1424 1425 static void 1426 chunk_dealloc(void *chunk, size_t size) 1427 { 1428 chunk_node_t *node; 1429 1430 assert(chunk != NULL); 1431 assert(CHUNK_ADDR2BASE(chunk) == chunk); 1432 assert(size != 0); 1433 assert((size & chunksize_mask) == 0); 1434 1435 malloc_mutex_lock(&chunks_mtx); 1436 1437 #ifdef USE_BRK 1438 if ((uintptr_t)chunk >= (uintptr_t)brk_base 1439 && (uintptr_t)chunk < (uintptr_t)brk_max) { 1440 void *brk_cur; 1441 1442 malloc_mutex_lock(&brk_mtx); 1443 /* Get the current end of brk. */ 1444 brk_cur = sbrk(0); 1445 1446 /* 1447 * Try to shrink the data segment if this chunk is at the end 1448 * of the data segment. The sbrk() call here is subject to a 1449 * race condition with threads that use brk(2) or sbrk(2) 1450 * directly, but the alternative would be to leak memory for 1451 * the sake of poorly designed multi-threaded programs. 1452 */ 1453 if (brk_cur == brk_max 1454 && (void *)((uintptr_t)chunk + size) == brk_max 1455 && sbrk(-(intptr_t)size) == brk_max) { 1456 malloc_mutex_unlock(&brk_mtx); 1457 if (brk_prev == brk_max) { 1458 /* Success. */ 1459 brk_prev = (void *)((intptr_t)brk_max 1460 - (intptr_t)size); 1461 brk_max = brk_prev; 1462 } 1463 } else { 1464 size_t offset; 1465 1466 malloc_mutex_unlock(&brk_mtx); 1467 madvise(chunk, size, MADV_FREE); 1468 1469 /* 1470 * Iteratively create records of each chunk-sized 1471 * memory region that 'chunk' is comprised of, so that 1472 * the address range can be recycled if memory usage 1473 * increases later on. 1474 */ 1475 for (offset = 0; offset < size; offset += chunksize) { 1476 node = base_chunk_node_alloc(); 1477 if (node == NULL) 1478 break; 1479 1480 node->chunk = (void *)((uintptr_t)chunk 1481 + (uintptr_t)offset); 1482 node->size = chunksize; 1483 /* LINTED */ 1484 RB_INSERT(chunk_tree_s, &old_chunks, node); 1485 } 1486 } 1487 } else { 1488 #endif 1489 pages_unmap(chunk, size); 1490 1491 /* 1492 * Make a record of the chunk's address, so that the address 1493 * range can be recycled if memory usage increases later on. 1494 * Don't bother to create entries if (size > chunksize), since 1495 * doing so could cause scalability issues for truly gargantuan 1496 * objects (many gigabytes or larger). 1497 */ 1498 if (size == chunksize) { 1499 node = base_chunk_node_alloc(); 1500 if (node != NULL) { 1501 node->chunk = (void *)(uintptr_t)chunk; 1502 node->size = chunksize; 1503 /* LINTED */ 1504 RB_INSERT(chunk_tree_s, &old_chunks, node); 1505 } 1506 } 1507 #ifdef USE_BRK 1508 } 1509 #endif 1510 1511 #ifdef MALLOC_STATS 1512 stats_chunks.curchunks -= (size / chunksize); 1513 #endif 1514 malloc_mutex_unlock(&chunks_mtx); 1515 } 1516 1517 /* 1518 * End chunk management functions. 1519 */ 1520 /******************************************************************************/ 1521 /* 1522 * Begin arena. 1523 */ 1524 1525 /* 1526 * Choose an arena based on a per-thread and (optimistically) per-CPU value. 1527 * 1528 * We maintain at least one block of arenas. Usually there are more. 1529 * The blocks are $ncpu arenas in size. Whole blocks are 'hashed' 1530 * amongst threads. To accomplish this, next_arena advances only in 1531 * ncpu steps. 1532 */ 1533 static __noinline arena_t * 1534 choose_arena_hard(void) 1535 { 1536 unsigned i, curcpu; 1537 arena_t **map; 1538 1539 /* Initialize the current block of arenas and advance to next. */ 1540 malloc_mutex_lock(&arenas_mtx); 1541 assert(next_arena % ncpus == 0); 1542 assert(narenas % ncpus == 0); 1543 map = &arenas[next_arena]; 1544 set_arenas_map(map); 1545 for (i = 0; i < ncpus; i++) { 1546 if (arenas[next_arena] == NULL) 1547 arenas_extend(next_arena); 1548 next_arena = (next_arena + 1) % narenas; 1549 } 1550 malloc_mutex_unlock(&arenas_mtx); 1551 1552 /* 1553 * If we were unable to allocate an arena above, then default to 1554 * the first arena, which is always present. 1555 */ 1556 curcpu = thr_curcpu(); 1557 if (map[curcpu] != NULL) 1558 return map[curcpu]; 1559 return arenas[0]; 1560 } 1561 1562 static inline arena_t * 1563 choose_arena(void) 1564 { 1565 unsigned curcpu; 1566 arena_t **map; 1567 1568 map = get_arenas_map(); 1569 curcpu = thr_curcpu(); 1570 if (__predict_true(map != NULL && map[curcpu] != NULL)) 1571 return map[curcpu]; 1572 1573 return choose_arena_hard(); 1574 } 1575 1576 #ifndef lint 1577 static inline int 1578 arena_chunk_comp(arena_chunk_t *a, arena_chunk_t *b) 1579 { 1580 1581 assert(a != NULL); 1582 assert(b != NULL); 1583 1584 if ((uintptr_t)a < (uintptr_t)b) 1585 return (-1); 1586 else if (a == b) 1587 return (0); 1588 else 1589 return (1); 1590 } 1591 1592 /* Generate red-black tree code for arena chunks. */ 1593 RB_GENERATE_STATIC(arena_chunk_tree_s, arena_chunk_s, link, arena_chunk_comp); 1594 #endif 1595 1596 #ifndef lint 1597 static inline int 1598 arena_run_comp(arena_run_t *a, arena_run_t *b) 1599 { 1600 1601 assert(a != NULL); 1602 assert(b != NULL); 1603 1604 if ((uintptr_t)a < (uintptr_t)b) 1605 return (-1); 1606 else if (a == b) 1607 return (0); 1608 else 1609 return (1); 1610 } 1611 1612 /* Generate red-black tree code for arena runs. */ 1613 RB_GENERATE_STATIC(arena_run_tree_s, arena_run_s, link, arena_run_comp); 1614 #endif 1615 1616 static inline void * 1617 arena_run_reg_alloc(arena_run_t *run, arena_bin_t *bin) 1618 { 1619 void *ret; 1620 unsigned i, mask, bit, regind; 1621 1622 assert(run->magic == ARENA_RUN_MAGIC); 1623 assert(run->regs_minelm < bin->regs_mask_nelms); 1624 1625 /* 1626 * Move the first check outside the loop, so that run->regs_minelm can 1627 * be updated unconditionally, without the possibility of updating it 1628 * multiple times. 1629 */ 1630 i = run->regs_minelm; 1631 mask = run->regs_mask[i]; 1632 if (mask != 0) { 1633 /* Usable allocation found. */ 1634 bit = ffs((int)mask) - 1; 1635 1636 regind = ((i << (SIZEOF_INT_2POW + 3)) + bit); 1637 ret = (void *)(((uintptr_t)run) + bin->reg0_offset 1638 + (bin->reg_size * regind)); 1639 1640 /* Clear bit. */ 1641 mask ^= (1 << bit); 1642 run->regs_mask[i] = mask; 1643 1644 return (ret); 1645 } 1646 1647 for (i++; i < bin->regs_mask_nelms; i++) { 1648 mask = run->regs_mask[i]; 1649 if (mask != 0) { 1650 /* Usable allocation found. */ 1651 bit = ffs((int)mask) - 1; 1652 1653 regind = ((i << (SIZEOF_INT_2POW + 3)) + bit); 1654 ret = (void *)(((uintptr_t)run) + bin->reg0_offset 1655 + (bin->reg_size * regind)); 1656 1657 /* Clear bit. */ 1658 mask ^= (1 << bit); 1659 run->regs_mask[i] = mask; 1660 1661 /* 1662 * Make a note that nothing before this element 1663 * contains a free region. 1664 */ 1665 run->regs_minelm = i; /* Low payoff: + (mask == 0); */ 1666 1667 return (ret); 1668 } 1669 } 1670 /* Not reached. */ 1671 /* LINTED */ 1672 assert(0); 1673 return (NULL); 1674 } 1675 1676 static inline void 1677 arena_run_reg_dalloc(arena_run_t *run, arena_bin_t *bin, void *ptr, size_t size) 1678 { 1679 /* 1680 * To divide by a number D that is not a power of two we multiply 1681 * by (2^21 / D) and then right shift by 21 positions. 1682 * 1683 * X / D 1684 * 1685 * becomes 1686 * 1687 * (X * size_invs[(D >> QUANTUM_2POW_MIN) - 3]) >> SIZE_INV_SHIFT 1688 */ 1689 #define SIZE_INV_SHIFT 21 1690 #define SIZE_INV(s) (((1 << SIZE_INV_SHIFT) / (s << QUANTUM_2POW_MIN)) + 1) 1691 static const unsigned size_invs[] = { 1692 SIZE_INV(3), 1693 SIZE_INV(4), SIZE_INV(5), SIZE_INV(6), SIZE_INV(7), 1694 SIZE_INV(8), SIZE_INV(9), SIZE_INV(10), SIZE_INV(11), 1695 SIZE_INV(12),SIZE_INV(13), SIZE_INV(14), SIZE_INV(15), 1696 SIZE_INV(16),SIZE_INV(17), SIZE_INV(18), SIZE_INV(19), 1697 SIZE_INV(20),SIZE_INV(21), SIZE_INV(22), SIZE_INV(23), 1698 SIZE_INV(24),SIZE_INV(25), SIZE_INV(26), SIZE_INV(27), 1699 SIZE_INV(28),SIZE_INV(29), SIZE_INV(30), SIZE_INV(31) 1700 #if (QUANTUM_2POW_MIN < 4) 1701 , 1702 SIZE_INV(32), SIZE_INV(33), SIZE_INV(34), SIZE_INV(35), 1703 SIZE_INV(36), SIZE_INV(37), SIZE_INV(38), SIZE_INV(39), 1704 SIZE_INV(40), SIZE_INV(41), SIZE_INV(42), SIZE_INV(43), 1705 SIZE_INV(44), SIZE_INV(45), SIZE_INV(46), SIZE_INV(47), 1706 SIZE_INV(48), SIZE_INV(49), SIZE_INV(50), SIZE_INV(51), 1707 SIZE_INV(52), SIZE_INV(53), SIZE_INV(54), SIZE_INV(55), 1708 SIZE_INV(56), SIZE_INV(57), SIZE_INV(58), SIZE_INV(59), 1709 SIZE_INV(60), SIZE_INV(61), SIZE_INV(62), SIZE_INV(63) 1710 #endif 1711 }; 1712 unsigned diff, regind, elm, bit; 1713 1714 /* LINTED */ 1715 assert(run->magic == ARENA_RUN_MAGIC); 1716 assert(((sizeof(size_invs)) / sizeof(unsigned)) + 3 1717 >= (SMALL_MAX_DEFAULT >> QUANTUM_2POW_MIN)); 1718 1719 /* 1720 * Avoid doing division with a variable divisor if possible. Using 1721 * actual division here can reduce allocator throughput by over 20%! 1722 */ 1723 diff = (unsigned)((uintptr_t)ptr - (uintptr_t)run - bin->reg0_offset); 1724 if ((size & (size - 1)) == 0) { 1725 /* 1726 * log2_table allows fast division of a power of two in the 1727 * [1..128] range. 1728 * 1729 * (x / divisor) becomes (x >> log2_table[divisor - 1]). 1730 */ 1731 static const unsigned char log2_table[] = { 1732 0, 1, 0, 2, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 4, 1733 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 1734 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1735 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6, 1736 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1737 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1738 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1739 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7 1740 }; 1741 1742 if (size <= 128) 1743 regind = (diff >> log2_table[size - 1]); 1744 else if (size <= 32768) 1745 regind = diff >> (8 + log2_table[(size >> 8) - 1]); 1746 else { 1747 /* 1748 * The page size is too large for us to use the lookup 1749 * table. Use real division. 1750 */ 1751 regind = (unsigned)(diff / size); 1752 } 1753 } else if (size <= ((sizeof(size_invs) / sizeof(unsigned)) 1754 << QUANTUM_2POW_MIN) + 2) { 1755 regind = size_invs[(size >> QUANTUM_2POW_MIN) - 3] * diff; 1756 regind >>= SIZE_INV_SHIFT; 1757 } else { 1758 /* 1759 * size_invs isn't large enough to handle this size class, so 1760 * calculate regind using actual division. This only happens 1761 * if the user increases small_max via the 'S' runtime 1762 * configuration option. 1763 */ 1764 regind = (unsigned)(diff / size); 1765 }; 1766 assert(diff == regind * size); 1767 assert(regind < bin->nregs); 1768 1769 elm = regind >> (SIZEOF_INT_2POW + 3); 1770 if (elm < run->regs_minelm) 1771 run->regs_minelm = elm; 1772 bit = regind - (elm << (SIZEOF_INT_2POW + 3)); 1773 assert((run->regs_mask[elm] & (1 << bit)) == 0); 1774 run->regs_mask[elm] |= (1 << bit); 1775 #undef SIZE_INV 1776 #undef SIZE_INV_SHIFT 1777 } 1778 1779 static void 1780 arena_run_split(arena_t *arena, arena_run_t *run, size_t size) 1781 { 1782 arena_chunk_t *chunk; 1783 unsigned run_ind, map_offset, total_pages, need_pages, rem_pages; 1784 unsigned i; 1785 1786 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); 1787 run_ind = (unsigned)(((uintptr_t)run - (uintptr_t)chunk) 1788 >> pagesize_2pow); 1789 total_pages = chunk->map[run_ind].npages; 1790 need_pages = (unsigned)(size >> pagesize_2pow); 1791 assert(need_pages <= total_pages); 1792 rem_pages = total_pages - need_pages; 1793 1794 /* Split enough pages from the front of run to fit allocation size. */ 1795 map_offset = run_ind; 1796 for (i = 0; i < need_pages; i++) { 1797 chunk->map[map_offset + i].npages = need_pages; 1798 chunk->map[map_offset + i].pos = i; 1799 } 1800 1801 /* Keep track of trailing unused pages for later use. */ 1802 if (rem_pages > 0) { 1803 /* Update map for trailing pages. */ 1804 map_offset += need_pages; 1805 chunk->map[map_offset].npages = rem_pages; 1806 chunk->map[map_offset].pos = POS_FREE; 1807 chunk->map[map_offset + rem_pages - 1].npages = rem_pages; 1808 chunk->map[map_offset + rem_pages - 1].pos = POS_FREE; 1809 } 1810 1811 chunk->pages_used += need_pages; 1812 } 1813 1814 static arena_chunk_t * 1815 arena_chunk_alloc(arena_t *arena) 1816 { 1817 arena_chunk_t *chunk; 1818 1819 if (arena->spare != NULL) { 1820 chunk = arena->spare; 1821 arena->spare = NULL; 1822 1823 /* LINTED */ 1824 RB_INSERT(arena_chunk_tree_s, &arena->chunks, chunk); 1825 } else { 1826 chunk = (arena_chunk_t *)chunk_alloc(chunksize); 1827 if (chunk == NULL) 1828 return (NULL); 1829 #ifdef MALLOC_STATS 1830 arena->stats.mapped += chunksize; 1831 #endif 1832 1833 chunk->arena = arena; 1834 1835 /* LINTED */ 1836 RB_INSERT(arena_chunk_tree_s, &arena->chunks, chunk); 1837 1838 /* 1839 * Claim that no pages are in use, since the header is merely 1840 * overhead. 1841 */ 1842 chunk->pages_used = 0; 1843 1844 chunk->max_frun_npages = chunk_npages - 1845 arena_chunk_header_npages; 1846 chunk->min_frun_ind = arena_chunk_header_npages; 1847 1848 /* 1849 * Initialize enough of the map to support one maximal free run. 1850 */ 1851 chunk->map[arena_chunk_header_npages].npages = chunk_npages - 1852 arena_chunk_header_npages; 1853 chunk->map[arena_chunk_header_npages].pos = POS_FREE; 1854 chunk->map[chunk_npages - 1].npages = chunk_npages - 1855 arena_chunk_header_npages; 1856 chunk->map[chunk_npages - 1].pos = POS_FREE; 1857 } 1858 1859 return (chunk); 1860 } 1861 1862 static void 1863 arena_chunk_dealloc(arena_t *arena, arena_chunk_t *chunk) 1864 { 1865 1866 /* 1867 * Remove chunk from the chunk tree, regardless of whether this chunk 1868 * will be cached, so that the arena does not use it. 1869 */ 1870 /* LINTED */ 1871 RB_REMOVE(arena_chunk_tree_s, &chunk->arena->chunks, chunk); 1872 1873 if (opt_hint == false) { 1874 if (arena->spare != NULL) { 1875 chunk_dealloc((void *)arena->spare, chunksize); 1876 #ifdef MALLOC_STATS 1877 arena->stats.mapped -= chunksize; 1878 #endif 1879 } 1880 arena->spare = chunk; 1881 } else { 1882 assert(arena->spare == NULL); 1883 chunk_dealloc((void *)chunk, chunksize); 1884 #ifdef MALLOC_STATS 1885 arena->stats.mapped -= chunksize; 1886 #endif 1887 } 1888 } 1889 1890 static arena_run_t * 1891 arena_run_alloc(arena_t *arena, size_t size) 1892 { 1893 arena_chunk_t *chunk; 1894 arena_run_t *run; 1895 unsigned need_npages, limit_pages, compl_need_npages; 1896 1897 assert(size <= (chunksize - (arena_chunk_header_npages << 1898 pagesize_2pow))); 1899 assert((size & pagesize_mask) == 0); 1900 1901 /* 1902 * Search through arena's chunks in address order for a free run that is 1903 * large enough. Look for the first fit. 1904 */ 1905 need_npages = (unsigned)(size >> pagesize_2pow); 1906 limit_pages = chunk_npages - arena_chunk_header_npages; 1907 compl_need_npages = limit_pages - need_npages; 1908 /* LINTED */ 1909 RB_FOREACH(chunk, arena_chunk_tree_s, &arena->chunks) { 1910 /* 1911 * Avoid searching this chunk if there are not enough 1912 * contiguous free pages for there to possibly be a large 1913 * enough free run. 1914 */ 1915 if (chunk->pages_used <= compl_need_npages && 1916 need_npages <= chunk->max_frun_npages) { 1917 arena_chunk_map_t *mapelm; 1918 unsigned i; 1919 unsigned max_frun_npages = 0; 1920 unsigned min_frun_ind = chunk_npages; 1921 1922 assert(chunk->min_frun_ind >= 1923 arena_chunk_header_npages); 1924 for (i = chunk->min_frun_ind; i < chunk_npages;) { 1925 mapelm = &chunk->map[i]; 1926 if (mapelm->pos == POS_FREE) { 1927 if (mapelm->npages >= need_npages) { 1928 run = (arena_run_t *) 1929 ((uintptr_t)chunk + (i << 1930 pagesize_2pow)); 1931 /* Update page map. */ 1932 arena_run_split(arena, run, 1933 size); 1934 return (run); 1935 } 1936 if (mapelm->npages > 1937 max_frun_npages) { 1938 max_frun_npages = 1939 mapelm->npages; 1940 } 1941 if (i < min_frun_ind) { 1942 min_frun_ind = i; 1943 if (i < chunk->min_frun_ind) 1944 chunk->min_frun_ind = i; 1945 } 1946 } 1947 i += mapelm->npages; 1948 } 1949 /* 1950 * Search failure. Reset cached chunk->max_frun_npages. 1951 * chunk->min_frun_ind was already reset above (if 1952 * necessary). 1953 */ 1954 chunk->max_frun_npages = max_frun_npages; 1955 } 1956 } 1957 1958 /* 1959 * No usable runs. Create a new chunk from which to allocate the run. 1960 */ 1961 chunk = arena_chunk_alloc(arena); 1962 if (chunk == NULL) 1963 return (NULL); 1964 run = (arena_run_t *)((uintptr_t)chunk + (arena_chunk_header_npages << 1965 pagesize_2pow)); 1966 /* Update page map. */ 1967 arena_run_split(arena, run, size); 1968 return (run); 1969 } 1970 1971 static void 1972 arena_run_dalloc(arena_t *arena, arena_run_t *run, size_t size) 1973 { 1974 arena_chunk_t *chunk; 1975 unsigned run_ind, run_pages; 1976 1977 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(run); 1978 1979 run_ind = (unsigned)(((uintptr_t)run - (uintptr_t)chunk) 1980 >> pagesize_2pow); 1981 assert(run_ind >= arena_chunk_header_npages); 1982 assert(run_ind < (chunksize >> pagesize_2pow)); 1983 run_pages = (unsigned)(size >> pagesize_2pow); 1984 assert(run_pages == chunk->map[run_ind].npages); 1985 1986 /* Subtract pages from count of pages used in chunk. */ 1987 chunk->pages_used -= run_pages; 1988 1989 /* Mark run as deallocated. */ 1990 assert(chunk->map[run_ind].npages == run_pages); 1991 chunk->map[run_ind].pos = POS_FREE; 1992 assert(chunk->map[run_ind + run_pages - 1].npages == run_pages); 1993 chunk->map[run_ind + run_pages - 1].pos = POS_FREE; 1994 1995 /* 1996 * Tell the kernel that we don't need the data in this run, but only if 1997 * requested via runtime configuration. 1998 */ 1999 if (opt_hint) 2000 madvise(run, size, MADV_FREE); 2001 2002 /* Try to coalesce with neighboring runs. */ 2003 if (run_ind > arena_chunk_header_npages && 2004 chunk->map[run_ind - 1].pos == POS_FREE) { 2005 unsigned prev_npages; 2006 2007 /* Coalesce with previous run. */ 2008 prev_npages = chunk->map[run_ind - 1].npages; 2009 run_ind -= prev_npages; 2010 assert(chunk->map[run_ind].npages == prev_npages); 2011 assert(chunk->map[run_ind].pos == POS_FREE); 2012 run_pages += prev_npages; 2013 2014 chunk->map[run_ind].npages = run_pages; 2015 assert(chunk->map[run_ind].pos == POS_FREE); 2016 chunk->map[run_ind + run_pages - 1].npages = run_pages; 2017 assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE); 2018 } 2019 2020 if (run_ind + run_pages < chunk_npages && 2021 chunk->map[run_ind + run_pages].pos == POS_FREE) { 2022 unsigned next_npages; 2023 2024 /* Coalesce with next run. */ 2025 next_npages = chunk->map[run_ind + run_pages].npages; 2026 run_pages += next_npages; 2027 assert(chunk->map[run_ind + run_pages - 1].npages == 2028 next_npages); 2029 assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE); 2030 2031 chunk->map[run_ind].npages = run_pages; 2032 chunk->map[run_ind].pos = POS_FREE; 2033 chunk->map[run_ind + run_pages - 1].npages = run_pages; 2034 assert(chunk->map[run_ind + run_pages - 1].pos == POS_FREE); 2035 } 2036 2037 if (chunk->map[run_ind].npages > chunk->max_frun_npages) 2038 chunk->max_frun_npages = chunk->map[run_ind].npages; 2039 if (run_ind < chunk->min_frun_ind) 2040 chunk->min_frun_ind = run_ind; 2041 2042 /* Deallocate chunk if it is now completely unused. */ 2043 if (chunk->pages_used == 0) 2044 arena_chunk_dealloc(arena, chunk); 2045 } 2046 2047 static arena_run_t * 2048 arena_bin_nonfull_run_get(arena_t *arena, arena_bin_t *bin) 2049 { 2050 arena_run_t *run; 2051 unsigned i, remainder; 2052 2053 /* Look for a usable run. */ 2054 /* LINTED */ 2055 if ((run = RB_MIN(arena_run_tree_s, &bin->runs)) != NULL) { 2056 /* run is guaranteed to have available space. */ 2057 /* LINTED */ 2058 RB_REMOVE(arena_run_tree_s, &bin->runs, run); 2059 #ifdef MALLOC_STATS 2060 bin->stats.reruns++; 2061 #endif 2062 return (run); 2063 } 2064 /* No existing runs have any space available. */ 2065 2066 /* Allocate a new run. */ 2067 run = arena_run_alloc(arena, bin->run_size); 2068 if (run == NULL) 2069 return (NULL); 2070 2071 /* Initialize run internals. */ 2072 run->bin = bin; 2073 2074 for (i = 0; i < bin->regs_mask_nelms; i++) 2075 run->regs_mask[i] = UINT_MAX; 2076 remainder = bin->nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1); 2077 if (remainder != 0) { 2078 /* The last element has spare bits that need to be unset. */ 2079 run->regs_mask[i] = (UINT_MAX >> ((1 << (SIZEOF_INT_2POW + 3)) 2080 - remainder)); 2081 } 2082 2083 run->regs_minelm = 0; 2084 2085 run->nfree = bin->nregs; 2086 #ifdef MALLOC_DEBUG 2087 run->magic = ARENA_RUN_MAGIC; 2088 #endif 2089 2090 #ifdef MALLOC_STATS 2091 bin->stats.nruns++; 2092 bin->stats.curruns++; 2093 if (bin->stats.curruns > bin->stats.highruns) 2094 bin->stats.highruns = bin->stats.curruns; 2095 #endif 2096 return (run); 2097 } 2098 2099 /* bin->runcur must have space available before this function is called. */ 2100 static inline void * 2101 arena_bin_malloc_easy(arena_t *arena, arena_bin_t *bin, arena_run_t *run) 2102 { 2103 void *ret; 2104 2105 assert(run->magic == ARENA_RUN_MAGIC); 2106 assert(run->nfree > 0); 2107 2108 ret = arena_run_reg_alloc(run, bin); 2109 assert(ret != NULL); 2110 run->nfree--; 2111 2112 return (ret); 2113 } 2114 2115 /* Re-fill bin->runcur, then call arena_bin_malloc_easy(). */ 2116 static void * 2117 arena_bin_malloc_hard(arena_t *arena, arena_bin_t *bin) 2118 { 2119 2120 bin->runcur = arena_bin_nonfull_run_get(arena, bin); 2121 if (bin->runcur == NULL) 2122 return (NULL); 2123 assert(bin->runcur->magic == ARENA_RUN_MAGIC); 2124 assert(bin->runcur->nfree > 0); 2125 2126 return (arena_bin_malloc_easy(arena, bin, bin->runcur)); 2127 } 2128 2129 /* 2130 * Calculate bin->run_size such that it meets the following constraints: 2131 * 2132 * *) bin->run_size >= min_run_size 2133 * *) bin->run_size <= arena_maxclass 2134 * *) bin->run_size <= RUN_MAX_SMALL 2135 * *) run header overhead <= RUN_MAX_OVRHD (or header overhead relaxed). 2136 * 2137 * bin->nregs, bin->regs_mask_nelms, and bin->reg0_offset are 2138 * also calculated here, since these settings are all interdependent. 2139 */ 2140 static size_t 2141 arena_bin_run_size_calc(arena_bin_t *bin, size_t min_run_size) 2142 { 2143 size_t try_run_size, good_run_size; 2144 unsigned good_nregs, good_mask_nelms, good_reg0_offset; 2145 unsigned try_nregs, try_mask_nelms, try_reg0_offset; 2146 float max_ovrhd = RUN_MAX_OVRHD; 2147 2148 assert(min_run_size >= pagesize); 2149 assert(min_run_size <= arena_maxclass); 2150 assert(min_run_size <= RUN_MAX_SMALL); 2151 2152 /* 2153 * Calculate known-valid settings before entering the run_size 2154 * expansion loop, so that the first part of the loop always copies 2155 * valid settings. 2156 * 2157 * The do..while loop iteratively reduces the number of regions until 2158 * the run header and the regions no longer overlap. A closed formula 2159 * would be quite messy, since there is an interdependency between the 2160 * header's mask length and the number of regions. 2161 */ 2162 try_run_size = min_run_size; 2163 try_nregs = (unsigned)(((try_run_size - sizeof(arena_run_t)) / 2164 bin->reg_size) + 1); /* Counter-act the first line of the loop. */ 2165 do { 2166 try_nregs--; 2167 try_mask_nelms = (try_nregs >> (SIZEOF_INT_2POW + 3)) + 2168 ((try_nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1)) ? 1 : 0); 2169 try_reg0_offset = (unsigned)(try_run_size - 2170 (try_nregs * bin->reg_size)); 2171 } while (sizeof(arena_run_t) + (sizeof(unsigned) * (try_mask_nelms - 1)) 2172 > try_reg0_offset); 2173 2174 /* run_size expansion loop. */ 2175 do { 2176 /* 2177 * Copy valid settings before trying more aggressive settings. 2178 */ 2179 good_run_size = try_run_size; 2180 good_nregs = try_nregs; 2181 good_mask_nelms = try_mask_nelms; 2182 good_reg0_offset = try_reg0_offset; 2183 2184 /* Try more aggressive settings. */ 2185 try_run_size += pagesize; 2186 try_nregs = (unsigned)(((try_run_size - sizeof(arena_run_t)) / 2187 bin->reg_size) + 1); /* Counter-act try_nregs-- in loop. */ 2188 do { 2189 try_nregs--; 2190 try_mask_nelms = (try_nregs >> (SIZEOF_INT_2POW + 3)) + 2191 ((try_nregs & ((1 << (SIZEOF_INT_2POW + 3)) - 1)) ? 2192 1 : 0); 2193 try_reg0_offset = (unsigned)(try_run_size - (try_nregs * 2194 bin->reg_size)); 2195 } while (sizeof(arena_run_t) + (sizeof(unsigned) * 2196 (try_mask_nelms - 1)) > try_reg0_offset); 2197 } while (try_run_size <= arena_maxclass && try_run_size <= RUN_MAX_SMALL 2198 && max_ovrhd > RUN_MAX_OVRHD_RELAX / ((float)(bin->reg_size << 3)) 2199 && ((float)(try_reg0_offset)) / ((float)(try_run_size)) > 2200 max_ovrhd); 2201 2202 assert(sizeof(arena_run_t) + (sizeof(unsigned) * (good_mask_nelms - 1)) 2203 <= good_reg0_offset); 2204 assert((good_mask_nelms << (SIZEOF_INT_2POW + 3)) >= good_nregs); 2205 2206 /* Copy final settings. */ 2207 bin->run_size = good_run_size; 2208 bin->nregs = good_nregs; 2209 bin->regs_mask_nelms = good_mask_nelms; 2210 bin->reg0_offset = good_reg0_offset; 2211 2212 return (good_run_size); 2213 } 2214 2215 static void * 2216 arena_malloc(arena_t *arena, size_t size) 2217 { 2218 void *ret; 2219 2220 assert(arena != NULL); 2221 assert(arena->magic == ARENA_MAGIC); 2222 assert(size != 0); 2223 assert(QUANTUM_CEILING(size) <= arena_maxclass); 2224 2225 if (size <= bin_maxclass) { 2226 arena_bin_t *bin; 2227 arena_run_t *run; 2228 2229 /* Small allocation. */ 2230 2231 if (size < small_min) { 2232 /* Tiny. */ 2233 size = pow2_ceil(size); 2234 bin = &arena->bins[ffs((int)(size >> (TINY_MIN_2POW + 2235 1)))]; 2236 #if (!defined(NDEBUG) || defined(MALLOC_STATS)) 2237 /* 2238 * Bin calculation is always correct, but we may need 2239 * to fix size for the purposes of assertions and/or 2240 * stats accuracy. 2241 */ 2242 if (size < (1 << TINY_MIN_2POW)) 2243 size = (1 << TINY_MIN_2POW); 2244 #endif 2245 } else if (size <= small_max) { 2246 /* Quantum-spaced. */ 2247 size = QUANTUM_CEILING(size); 2248 bin = &arena->bins[ntbins + (size >> opt_quantum_2pow) 2249 - 1]; 2250 } else { 2251 /* Sub-page. */ 2252 size = pow2_ceil(size); 2253 bin = &arena->bins[ntbins + nqbins 2254 + (ffs((int)(size >> opt_small_max_2pow)) - 2)]; 2255 } 2256 assert(size == bin->reg_size); 2257 2258 malloc_mutex_lock(&arena->mtx); 2259 if ((run = bin->runcur) != NULL && run->nfree > 0) 2260 ret = arena_bin_malloc_easy(arena, bin, run); 2261 else 2262 ret = arena_bin_malloc_hard(arena, bin); 2263 2264 if (ret == NULL) { 2265 malloc_mutex_unlock(&arena->mtx); 2266 return (NULL); 2267 } 2268 2269 #ifdef MALLOC_STATS 2270 bin->stats.nrequests++; 2271 arena->stats.nmalloc_small++; 2272 arena->stats.allocated_small += size; 2273 #endif 2274 } else { 2275 /* Large allocation. */ 2276 size = PAGE_CEILING(size); 2277 malloc_mutex_lock(&arena->mtx); 2278 ret = (void *)arena_run_alloc(arena, size); 2279 if (ret == NULL) { 2280 malloc_mutex_unlock(&arena->mtx); 2281 return (NULL); 2282 } 2283 #ifdef MALLOC_STATS 2284 arena->stats.nmalloc_large++; 2285 arena->stats.allocated_large += size; 2286 #endif 2287 } 2288 2289 malloc_mutex_unlock(&arena->mtx); 2290 2291 if (opt_junk) 2292 memset(ret, 0xa5, size); 2293 else if (opt_zero) 2294 memset(ret, 0, size); 2295 return (ret); 2296 } 2297 2298 static inline void 2299 arena_palloc_trim(arena_t *arena, arena_chunk_t *chunk, unsigned pageind, 2300 unsigned npages) 2301 { 2302 unsigned i; 2303 2304 assert(npages > 0); 2305 2306 /* 2307 * Modifiy the map such that arena_run_dalloc() sees the run as 2308 * separately allocated. 2309 */ 2310 for (i = 0; i < npages; i++) { 2311 chunk->map[pageind + i].npages = npages; 2312 chunk->map[pageind + i].pos = i; 2313 } 2314 arena_run_dalloc(arena, (arena_run_t *)((uintptr_t)chunk + (pageind << 2315 pagesize_2pow)), npages << pagesize_2pow); 2316 } 2317 2318 /* Only handles large allocations that require more than page alignment. */ 2319 static void * 2320 arena_palloc(arena_t *arena, size_t alignment, size_t size, size_t alloc_size) 2321 { 2322 void *ret; 2323 size_t offset; 2324 arena_chunk_t *chunk; 2325 unsigned pageind, i, npages; 2326 2327 assert((size & pagesize_mask) == 0); 2328 assert((alignment & pagesize_mask) == 0); 2329 2330 npages = (unsigned)(size >> pagesize_2pow); 2331 2332 malloc_mutex_lock(&arena->mtx); 2333 ret = (void *)arena_run_alloc(arena, alloc_size); 2334 if (ret == NULL) { 2335 malloc_mutex_unlock(&arena->mtx); 2336 return (NULL); 2337 } 2338 2339 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ret); 2340 2341 offset = (uintptr_t)ret & (alignment - 1); 2342 assert((offset & pagesize_mask) == 0); 2343 assert(offset < alloc_size); 2344 if (offset == 0) { 2345 pageind = (unsigned)(((uintptr_t)ret - (uintptr_t)chunk) >> 2346 pagesize_2pow); 2347 2348 /* Update the map for the run to be kept. */ 2349 for (i = 0; i < npages; i++) { 2350 chunk->map[pageind + i].npages = npages; 2351 assert(chunk->map[pageind + i].pos == i); 2352 } 2353 2354 /* Trim trailing space. */ 2355 arena_palloc_trim(arena, chunk, pageind + npages, 2356 (unsigned)((alloc_size - size) >> pagesize_2pow)); 2357 } else { 2358 size_t leadsize, trailsize; 2359 2360 leadsize = alignment - offset; 2361 ret = (void *)((uintptr_t)ret + leadsize); 2362 pageind = (unsigned)(((uintptr_t)ret - (uintptr_t)chunk) >> 2363 pagesize_2pow); 2364 2365 /* Update the map for the run to be kept. */ 2366 for (i = 0; i < npages; i++) { 2367 chunk->map[pageind + i].npages = npages; 2368 chunk->map[pageind + i].pos = i; 2369 } 2370 2371 /* Trim leading space. */ 2372 arena_palloc_trim(arena, chunk, 2373 (unsigned)(pageind - (leadsize >> pagesize_2pow)), 2374 (unsigned)(leadsize >> pagesize_2pow)); 2375 2376 trailsize = alloc_size - leadsize - size; 2377 if (trailsize != 0) { 2378 /* Trim trailing space. */ 2379 assert(trailsize < alloc_size); 2380 arena_palloc_trim(arena, chunk, pageind + npages, 2381 (unsigned)(trailsize >> pagesize_2pow)); 2382 } 2383 } 2384 2385 #ifdef MALLOC_STATS 2386 arena->stats.nmalloc_large++; 2387 arena->stats.allocated_large += size; 2388 #endif 2389 malloc_mutex_unlock(&arena->mtx); 2390 2391 if (opt_junk) 2392 memset(ret, 0xa5, size); 2393 else if (opt_zero) 2394 memset(ret, 0, size); 2395 return (ret); 2396 } 2397 2398 /* Return the size of the allocation pointed to by ptr. */ 2399 static size_t 2400 arena_salloc(const void *ptr) 2401 { 2402 size_t ret; 2403 arena_chunk_t *chunk; 2404 arena_chunk_map_t *mapelm; 2405 unsigned pageind; 2406 2407 assert(ptr != NULL); 2408 assert(CHUNK_ADDR2BASE(ptr) != ptr); 2409 2410 /* 2411 * No arena data structures that we query here can change in a way that 2412 * affects this function, so we don't need to lock. 2413 */ 2414 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 2415 pageind = (unsigned)(((uintptr_t)ptr - (uintptr_t)chunk) >> 2416 pagesize_2pow); 2417 mapelm = &chunk->map[pageind]; 2418 if (mapelm->pos != 0 || ptr != (char *)((uintptr_t)chunk) + (pageind << 2419 pagesize_2pow)) { 2420 arena_run_t *run; 2421 2422 pageind -= mapelm->pos; 2423 2424 run = (arena_run_t *)((uintptr_t)chunk + (pageind << 2425 pagesize_2pow)); 2426 assert(run->magic == ARENA_RUN_MAGIC); 2427 ret = run->bin->reg_size; 2428 } else 2429 ret = mapelm->npages << pagesize_2pow; 2430 2431 return (ret); 2432 } 2433 2434 static void * 2435 arena_ralloc(void *ptr, size_t size, size_t oldsize) 2436 { 2437 void *ret; 2438 2439 /* Avoid moving the allocation if the size class would not change. */ 2440 if (size < small_min) { 2441 if (oldsize < small_min && 2442 ffs((int)(pow2_ceil(size) >> (TINY_MIN_2POW + 1))) 2443 == ffs((int)(pow2_ceil(oldsize) >> (TINY_MIN_2POW + 1)))) 2444 goto IN_PLACE; 2445 } else if (size <= small_max) { 2446 if (oldsize >= small_min && oldsize <= small_max && 2447 (QUANTUM_CEILING(size) >> opt_quantum_2pow) 2448 == (QUANTUM_CEILING(oldsize) >> opt_quantum_2pow)) 2449 goto IN_PLACE; 2450 } else { 2451 /* 2452 * We make no attempt to resize runs here, though it would be 2453 * possible to do so. 2454 */ 2455 if (oldsize > small_max && PAGE_CEILING(size) == oldsize) 2456 goto IN_PLACE; 2457 } 2458 2459 /* 2460 * If we get here, then size and oldsize are different enough that we 2461 * need to use a different size class. In that case, fall back to 2462 * allocating new space and copying. 2463 */ 2464 ret = arena_malloc(choose_arena(), size); 2465 if (ret == NULL) 2466 return (NULL); 2467 2468 /* Junk/zero-filling were already done by arena_malloc(). */ 2469 if (size < oldsize) 2470 memcpy(ret, ptr, size); 2471 else 2472 memcpy(ret, ptr, oldsize); 2473 idalloc(ptr); 2474 return (ret); 2475 IN_PLACE: 2476 if (opt_junk && size < oldsize) 2477 memset((void *)((uintptr_t)ptr + size), 0x5a, oldsize - size); 2478 else if (opt_zero && size > oldsize) 2479 memset((void *)((uintptr_t)ptr + oldsize), 0, size - oldsize); 2480 return (ptr); 2481 } 2482 2483 static void 2484 arena_dalloc(arena_t *arena, arena_chunk_t *chunk, void *ptr) 2485 { 2486 unsigned pageind; 2487 arena_chunk_map_t *mapelm; 2488 size_t size; 2489 2490 assert(arena != NULL); 2491 assert(arena->magic == ARENA_MAGIC); 2492 assert(chunk->arena == arena); 2493 assert(ptr != NULL); 2494 assert(CHUNK_ADDR2BASE(ptr) != ptr); 2495 2496 pageind = (unsigned)(((uintptr_t)ptr - (uintptr_t)chunk) >> 2497 pagesize_2pow); 2498 mapelm = &chunk->map[pageind]; 2499 if (mapelm->pos != 0 || ptr != (char *)((uintptr_t)chunk) + (pageind << 2500 pagesize_2pow)) { 2501 arena_run_t *run; 2502 arena_bin_t *bin; 2503 2504 /* Small allocation. */ 2505 2506 pageind -= mapelm->pos; 2507 2508 run = (arena_run_t *)((uintptr_t)chunk + (pageind << 2509 pagesize_2pow)); 2510 assert(run->magic == ARENA_RUN_MAGIC); 2511 bin = run->bin; 2512 size = bin->reg_size; 2513 2514 if (opt_junk) 2515 memset(ptr, 0x5a, size); 2516 2517 malloc_mutex_lock(&arena->mtx); 2518 arena_run_reg_dalloc(run, bin, ptr, size); 2519 run->nfree++; 2520 2521 if (run->nfree == bin->nregs) { 2522 /* Deallocate run. */ 2523 if (run == bin->runcur) 2524 bin->runcur = NULL; 2525 else if (bin->nregs != 1) { 2526 /* 2527 * This block's conditional is necessary because 2528 * if the run only contains one region, then it 2529 * never gets inserted into the non-full runs 2530 * tree. 2531 */ 2532 /* LINTED */ 2533 RB_REMOVE(arena_run_tree_s, &bin->runs, run); 2534 } 2535 #ifdef MALLOC_DEBUG 2536 run->magic = 0; 2537 #endif 2538 arena_run_dalloc(arena, run, bin->run_size); 2539 #ifdef MALLOC_STATS 2540 bin->stats.curruns--; 2541 #endif 2542 } else if (run->nfree == 1 && run != bin->runcur) { 2543 /* 2544 * Make sure that bin->runcur always refers to the 2545 * lowest non-full run, if one exists. 2546 */ 2547 if (bin->runcur == NULL) 2548 bin->runcur = run; 2549 else if ((uintptr_t)run < (uintptr_t)bin->runcur) { 2550 /* Switch runcur. */ 2551 if (bin->runcur->nfree > 0) { 2552 /* Insert runcur. */ 2553 /* LINTED */ 2554 RB_INSERT(arena_run_tree_s, &bin->runs, 2555 bin->runcur); 2556 } 2557 bin->runcur = run; 2558 } else { 2559 /* LINTED */ 2560 RB_INSERT(arena_run_tree_s, &bin->runs, run); 2561 } 2562 } 2563 #ifdef MALLOC_STATS 2564 arena->stats.allocated_small -= size; 2565 arena->stats.ndalloc_small++; 2566 #endif 2567 } else { 2568 /* Large allocation. */ 2569 2570 size = mapelm->npages << pagesize_2pow; 2571 assert((((uintptr_t)ptr) & pagesize_mask) == 0); 2572 2573 if (opt_junk) 2574 memset(ptr, 0x5a, size); 2575 2576 malloc_mutex_lock(&arena->mtx); 2577 arena_run_dalloc(arena, (arena_run_t *)ptr, size); 2578 #ifdef MALLOC_STATS 2579 arena->stats.allocated_large -= size; 2580 arena->stats.ndalloc_large++; 2581 #endif 2582 } 2583 2584 malloc_mutex_unlock(&arena->mtx); 2585 } 2586 2587 static bool 2588 arena_new(arena_t *arena) 2589 { 2590 unsigned i; 2591 arena_bin_t *bin; 2592 size_t prev_run_size; 2593 2594 malloc_mutex_init(&arena->mtx); 2595 2596 #ifdef MALLOC_STATS 2597 memset(&arena->stats, 0, sizeof(arena_stats_t)); 2598 #endif 2599 2600 /* Initialize chunks. */ 2601 RB_INIT(&arena->chunks); 2602 arena->spare = NULL; 2603 2604 /* Initialize bins. */ 2605 prev_run_size = pagesize; 2606 2607 /* (2^n)-spaced tiny bins. */ 2608 for (i = 0; i < ntbins; i++) { 2609 bin = &arena->bins[i]; 2610 bin->runcur = NULL; 2611 RB_INIT(&bin->runs); 2612 2613 bin->reg_size = (1 << (TINY_MIN_2POW + i)); 2614 prev_run_size = arena_bin_run_size_calc(bin, prev_run_size); 2615 2616 #ifdef MALLOC_STATS 2617 memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); 2618 #endif 2619 } 2620 2621 /* Quantum-spaced bins. */ 2622 for (; i < ntbins + nqbins; i++) { 2623 bin = &arena->bins[i]; 2624 bin->runcur = NULL; 2625 RB_INIT(&bin->runs); 2626 2627 bin->reg_size = quantum * (i - ntbins + 1); 2628 /* 2629 pow2_size = pow2_ceil(quantum * (i - ntbins + 1)); 2630 */ 2631 prev_run_size = arena_bin_run_size_calc(bin, prev_run_size); 2632 2633 #ifdef MALLOC_STATS 2634 memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); 2635 #endif 2636 } 2637 2638 /* (2^n)-spaced sub-page bins. */ 2639 for (; i < ntbins + nqbins + nsbins; i++) { 2640 bin = &arena->bins[i]; 2641 bin->runcur = NULL; 2642 RB_INIT(&bin->runs); 2643 2644 bin->reg_size = (small_max << (i - (ntbins + nqbins) + 1)); 2645 2646 prev_run_size = arena_bin_run_size_calc(bin, prev_run_size); 2647 2648 #ifdef MALLOC_STATS 2649 memset(&bin->stats, 0, sizeof(malloc_bin_stats_t)); 2650 #endif 2651 } 2652 2653 #ifdef MALLOC_DEBUG 2654 arena->magic = ARENA_MAGIC; 2655 #endif 2656 2657 return (false); 2658 } 2659 2660 /* Create a new arena and insert it into the arenas array at index ind. */ 2661 static arena_t * 2662 arenas_extend(unsigned ind) 2663 { 2664 arena_t *ret; 2665 2666 /* Allocate enough space for trailing bins. */ 2667 ret = (arena_t *)base_alloc(sizeof(arena_t) 2668 + (sizeof(arena_bin_t) * (ntbins + nqbins + nsbins - 1))); 2669 if (ret != NULL && arena_new(ret) == false) { 2670 arenas[ind] = ret; 2671 return (ret); 2672 } 2673 /* Only reached if there is an OOM error. */ 2674 2675 /* 2676 * OOM here is quite inconvenient to propagate, since dealing with it 2677 * would require a check for failure in the fast path. Instead, punt 2678 * by using arenas[0]. In practice, this is an extremely unlikely 2679 * failure. 2680 */ 2681 _malloc_message(getprogname(), 2682 ": (malloc) Error initializing arena\n", "", ""); 2683 if (opt_abort) 2684 abort(); 2685 2686 return (arenas[0]); 2687 } 2688 2689 /* 2690 * End arena. 2691 */ 2692 /******************************************************************************/ 2693 /* 2694 * Begin general internal functions. 2695 */ 2696 2697 static void * 2698 huge_malloc(size_t size) 2699 { 2700 void *ret; 2701 size_t csize; 2702 chunk_node_t *node; 2703 2704 /* Allocate one or more contiguous chunks for this request. */ 2705 2706 csize = CHUNK_CEILING(size); 2707 if (csize == 0) { 2708 /* size is large enough to cause size_t wrap-around. */ 2709 return (NULL); 2710 } 2711 2712 /* Allocate a chunk node with which to track the chunk. */ 2713 node = base_chunk_node_alloc(); 2714 if (node == NULL) 2715 return (NULL); 2716 2717 ret = chunk_alloc(csize); 2718 if (ret == NULL) { 2719 base_chunk_node_dealloc(node); 2720 return (NULL); 2721 } 2722 2723 /* Insert node into huge. */ 2724 node->chunk = ret; 2725 node->size = csize; 2726 2727 malloc_mutex_lock(&chunks_mtx); 2728 RB_INSERT(chunk_tree_s, &huge, node); 2729 #ifdef MALLOC_STATS 2730 huge_nmalloc++; 2731 huge_allocated += csize; 2732 #endif 2733 malloc_mutex_unlock(&chunks_mtx); 2734 2735 if (opt_junk) 2736 memset(ret, 0xa5, csize); 2737 else if (opt_zero) 2738 memset(ret, 0, csize); 2739 2740 return (ret); 2741 } 2742 2743 /* Only handles large allocations that require more than chunk alignment. */ 2744 static void * 2745 huge_palloc(size_t alignment, size_t size) 2746 { 2747 void *ret; 2748 size_t alloc_size, chunk_size, offset; 2749 chunk_node_t *node; 2750 2751 /* 2752 * This allocation requires alignment that is even larger than chunk 2753 * alignment. This means that huge_malloc() isn't good enough. 2754 * 2755 * Allocate almost twice as many chunks as are demanded by the size or 2756 * alignment, in order to assure the alignment can be achieved, then 2757 * unmap leading and trailing chunks. 2758 */ 2759 assert(alignment >= chunksize); 2760 2761 chunk_size = CHUNK_CEILING(size); 2762 2763 if (size >= alignment) 2764 alloc_size = chunk_size + alignment - chunksize; 2765 else 2766 alloc_size = (alignment << 1) - chunksize; 2767 2768 /* Allocate a chunk node with which to track the chunk. */ 2769 node = base_chunk_node_alloc(); 2770 if (node == NULL) 2771 return (NULL); 2772 2773 ret = chunk_alloc(alloc_size); 2774 if (ret == NULL) { 2775 base_chunk_node_dealloc(node); 2776 return (NULL); 2777 } 2778 2779 offset = (uintptr_t)ret & (alignment - 1); 2780 assert((offset & chunksize_mask) == 0); 2781 assert(offset < alloc_size); 2782 if (offset == 0) { 2783 /* Trim trailing space. */ 2784 chunk_dealloc((void *)((uintptr_t)ret + chunk_size), alloc_size 2785 - chunk_size); 2786 } else { 2787 size_t trailsize; 2788 2789 /* Trim leading space. */ 2790 chunk_dealloc(ret, alignment - offset); 2791 2792 ret = (void *)((uintptr_t)ret + (alignment - offset)); 2793 2794 trailsize = alloc_size - (alignment - offset) - chunk_size; 2795 if (trailsize != 0) { 2796 /* Trim trailing space. */ 2797 assert(trailsize < alloc_size); 2798 chunk_dealloc((void *)((uintptr_t)ret + chunk_size), 2799 trailsize); 2800 } 2801 } 2802 2803 /* Insert node into huge. */ 2804 node->chunk = ret; 2805 node->size = chunk_size; 2806 2807 malloc_mutex_lock(&chunks_mtx); 2808 RB_INSERT(chunk_tree_s, &huge, node); 2809 #ifdef MALLOC_STATS 2810 huge_nmalloc++; 2811 huge_allocated += chunk_size; 2812 #endif 2813 malloc_mutex_unlock(&chunks_mtx); 2814 2815 if (opt_junk) 2816 memset(ret, 0xa5, chunk_size); 2817 else if (opt_zero) 2818 memset(ret, 0, chunk_size); 2819 2820 return (ret); 2821 } 2822 2823 static void * 2824 huge_ralloc(void *ptr, size_t size, size_t oldsize) 2825 { 2826 void *ret; 2827 2828 /* Avoid moving the allocation if the size class would not change. */ 2829 if (oldsize > arena_maxclass && 2830 CHUNK_CEILING(size) == CHUNK_CEILING(oldsize)) { 2831 if (opt_junk && size < oldsize) { 2832 memset((void *)((uintptr_t)ptr + size), 0x5a, oldsize 2833 - size); 2834 } else if (opt_zero && size > oldsize) { 2835 memset((void *)((uintptr_t)ptr + oldsize), 0, size 2836 - oldsize); 2837 } 2838 return (ptr); 2839 } 2840 2841 if (CHUNK_ADDR2BASE(ptr) == ptr 2842 #ifdef USE_BRK 2843 && ((uintptr_t)ptr < (uintptr_t)brk_base 2844 || (uintptr_t)ptr >= (uintptr_t)brk_max) 2845 #endif 2846 ) { 2847 chunk_node_t *node, key; 2848 void *newptr; 2849 size_t oldcsize; 2850 size_t newcsize; 2851 2852 newcsize = CHUNK_CEILING(size); 2853 oldcsize = CHUNK_CEILING(oldsize); 2854 assert(oldcsize != newcsize); 2855 if (newcsize == 0) { 2856 /* size_t wrap-around */ 2857 return (NULL); 2858 } 2859 newptr = mremap(ptr, oldcsize, NULL, newcsize, 2860 MAP_ALIGNED(chunksize_2pow)); 2861 if (newptr != MAP_FAILED) { 2862 assert(CHUNK_ADDR2BASE(newptr) == newptr); 2863 2864 /* update tree */ 2865 malloc_mutex_lock(&chunks_mtx); 2866 key.chunk = __DECONST(void *, ptr); 2867 /* LINTED */ 2868 node = RB_FIND(chunk_tree_s, &huge, &key); 2869 assert(node != NULL); 2870 assert(node->chunk == ptr); 2871 assert(node->size == oldcsize); 2872 node->size = newcsize; 2873 if (ptr != newptr) { 2874 RB_REMOVE(chunk_tree_s, &huge, node); 2875 node->chunk = newptr; 2876 RB_INSERT(chunk_tree_s, &huge, node); 2877 } 2878 #ifdef MALLOC_STATS 2879 huge_nralloc++; 2880 huge_allocated += newcsize - oldcsize; 2881 if (newcsize > oldcsize) { 2882 stats_chunks.curchunks += 2883 (newcsize - oldcsize) / chunksize; 2884 if (stats_chunks.curchunks > 2885 stats_chunks.highchunks) 2886 stats_chunks.highchunks = 2887 stats_chunks.curchunks; 2888 } else { 2889 stats_chunks.curchunks -= 2890 (oldcsize - newcsize) / chunksize; 2891 } 2892 #endif 2893 malloc_mutex_unlock(&chunks_mtx); 2894 2895 if (opt_junk && size < oldsize) { 2896 memset((void *)((uintptr_t)newptr + size), 0x5a, 2897 newcsize - size); 2898 } else if (opt_zero && size > oldsize) { 2899 memset((void *)((uintptr_t)newptr + oldsize), 0, 2900 size - oldsize); 2901 } 2902 return (newptr); 2903 } 2904 } 2905 2906 /* 2907 * If we get here, then size and oldsize are different enough that we 2908 * need to use a different size class. In that case, fall back to 2909 * allocating new space and copying. 2910 */ 2911 ret = huge_malloc(size); 2912 if (ret == NULL) 2913 return (NULL); 2914 2915 if (CHUNK_ADDR2BASE(ptr) == ptr) { 2916 /* The old allocation is a chunk. */ 2917 if (size < oldsize) 2918 memcpy(ret, ptr, size); 2919 else 2920 memcpy(ret, ptr, oldsize); 2921 } else { 2922 /* The old allocation is a region. */ 2923 assert(oldsize < size); 2924 memcpy(ret, ptr, oldsize); 2925 } 2926 idalloc(ptr); 2927 return (ret); 2928 } 2929 2930 static void 2931 huge_dalloc(void *ptr) 2932 { 2933 chunk_node_t key; 2934 chunk_node_t *node; 2935 2936 malloc_mutex_lock(&chunks_mtx); 2937 2938 /* Extract from tree of huge allocations. */ 2939 key.chunk = ptr; 2940 /* LINTED */ 2941 node = RB_FIND(chunk_tree_s, &huge, &key); 2942 assert(node != NULL); 2943 assert(node->chunk == ptr); 2944 /* LINTED */ 2945 RB_REMOVE(chunk_tree_s, &huge, node); 2946 2947 #ifdef MALLOC_STATS 2948 huge_ndalloc++; 2949 huge_allocated -= node->size; 2950 #endif 2951 2952 malloc_mutex_unlock(&chunks_mtx); 2953 2954 /* Unmap chunk. */ 2955 #ifdef USE_BRK 2956 if (opt_junk) 2957 memset(node->chunk, 0x5a, node->size); 2958 #endif 2959 chunk_dealloc(node->chunk, node->size); 2960 2961 base_chunk_node_dealloc(node); 2962 } 2963 2964 static void * 2965 imalloc(size_t size) 2966 { 2967 void *ret; 2968 2969 assert(size != 0); 2970 2971 if (size <= arena_maxclass) 2972 ret = arena_malloc(choose_arena(), size); 2973 else 2974 ret = huge_malloc(size); 2975 2976 return (ret); 2977 } 2978 2979 static void * 2980 ipalloc(size_t alignment, size_t size) 2981 { 2982 void *ret; 2983 size_t ceil_size; 2984 2985 /* 2986 * Round size up to the nearest multiple of alignment. 2987 * 2988 * This done, we can take advantage of the fact that for each small 2989 * size class, every object is aligned at the smallest power of two 2990 * that is non-zero in the base two representation of the size. For 2991 * example: 2992 * 2993 * Size | Base 2 | Minimum alignment 2994 * -----+----------+------------------ 2995 * 96 | 1100000 | 32 2996 * 144 | 10100000 | 32 2997 * 192 | 11000000 | 64 2998 * 2999 * Depending on runtime settings, it is possible that arena_malloc() 3000 * will further round up to a power of two, but that never causes 3001 * correctness issues. 3002 */ 3003 ceil_size = (size + (alignment - 1)) & (-alignment); 3004 /* 3005 * (ceil_size < size) protects against the combination of maximal 3006 * alignment and size greater than maximal alignment. 3007 */ 3008 if (ceil_size < size) { 3009 /* size_t overflow. */ 3010 return (NULL); 3011 } 3012 3013 if (ceil_size <= pagesize || (alignment <= pagesize 3014 && ceil_size <= arena_maxclass)) 3015 ret = arena_malloc(choose_arena(), ceil_size); 3016 else { 3017 size_t run_size; 3018 3019 /* 3020 * We can't achieve sub-page alignment, so round up alignment 3021 * permanently; it makes later calculations simpler. 3022 */ 3023 alignment = PAGE_CEILING(alignment); 3024 ceil_size = PAGE_CEILING(size); 3025 /* 3026 * (ceil_size < size) protects against very large sizes within 3027 * pagesize of SIZE_T_MAX. 3028 * 3029 * (ceil_size + alignment < ceil_size) protects against the 3030 * combination of maximal alignment and ceil_size large enough 3031 * to cause overflow. This is similar to the first overflow 3032 * check above, but it needs to be repeated due to the new 3033 * ceil_size value, which may now be *equal* to maximal 3034 * alignment, whereas before we only detected overflow if the 3035 * original size was *greater* than maximal alignment. 3036 */ 3037 if (ceil_size < size || ceil_size + alignment < ceil_size) { 3038 /* size_t overflow. */ 3039 return (NULL); 3040 } 3041 3042 /* 3043 * Calculate the size of the over-size run that arena_palloc() 3044 * would need to allocate in order to guarantee the alignment. 3045 */ 3046 if (ceil_size >= alignment) 3047 run_size = ceil_size + alignment - pagesize; 3048 else { 3049 /* 3050 * It is possible that (alignment << 1) will cause 3051 * overflow, but it doesn't matter because we also 3052 * subtract pagesize, which in the case of overflow 3053 * leaves us with a very large run_size. That causes 3054 * the first conditional below to fail, which means 3055 * that the bogus run_size value never gets used for 3056 * anything important. 3057 */ 3058 run_size = (alignment << 1) - pagesize; 3059 } 3060 3061 if (run_size <= arena_maxclass) { 3062 ret = arena_palloc(choose_arena(), alignment, ceil_size, 3063 run_size); 3064 } else if (alignment <= chunksize) 3065 ret = huge_malloc(ceil_size); 3066 else 3067 ret = huge_palloc(alignment, ceil_size); 3068 } 3069 3070 assert(((uintptr_t)ret & (alignment - 1)) == 0); 3071 return (ret); 3072 } 3073 3074 static void * 3075 icalloc(size_t size) 3076 { 3077 void *ret; 3078 3079 if (size <= arena_maxclass) { 3080 ret = arena_malloc(choose_arena(), size); 3081 if (ret == NULL) 3082 return (NULL); 3083 memset(ret, 0, size); 3084 } else { 3085 /* 3086 * The virtual memory system provides zero-filled pages, so 3087 * there is no need to do so manually, unless opt_junk is 3088 * enabled, in which case huge_malloc() fills huge allocations 3089 * with junk. 3090 */ 3091 ret = huge_malloc(size); 3092 if (ret == NULL) 3093 return (NULL); 3094 3095 if (opt_junk) 3096 memset(ret, 0, size); 3097 #ifdef USE_BRK 3098 else if ((uintptr_t)ret >= (uintptr_t)brk_base 3099 && (uintptr_t)ret < (uintptr_t)brk_max) { 3100 /* 3101 * This may be a re-used brk chunk. Therefore, zero 3102 * the memory. 3103 */ 3104 memset(ret, 0, size); 3105 } 3106 #endif 3107 } 3108 3109 return (ret); 3110 } 3111 3112 static size_t 3113 isalloc(const void *ptr) 3114 { 3115 size_t ret; 3116 arena_chunk_t *chunk; 3117 3118 assert(ptr != NULL); 3119 3120 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 3121 if (chunk != ptr) { 3122 /* Region. */ 3123 assert(chunk->arena->magic == ARENA_MAGIC); 3124 3125 ret = arena_salloc(ptr); 3126 } else { 3127 chunk_node_t *node, key; 3128 3129 /* Chunk (huge allocation). */ 3130 3131 malloc_mutex_lock(&chunks_mtx); 3132 3133 /* Extract from tree of huge allocations. */ 3134 key.chunk = __DECONST(void *, ptr); 3135 /* LINTED */ 3136 node = RB_FIND(chunk_tree_s, &huge, &key); 3137 assert(node != NULL); 3138 3139 ret = node->size; 3140 3141 malloc_mutex_unlock(&chunks_mtx); 3142 } 3143 3144 return (ret); 3145 } 3146 3147 static void * 3148 iralloc(void *ptr, size_t size) 3149 { 3150 void *ret; 3151 size_t oldsize; 3152 3153 assert(ptr != NULL); 3154 assert(size != 0); 3155 3156 oldsize = isalloc(ptr); 3157 3158 if (size <= arena_maxclass) 3159 ret = arena_ralloc(ptr, size, oldsize); 3160 else 3161 ret = huge_ralloc(ptr, size, oldsize); 3162 3163 return (ret); 3164 } 3165 3166 static void 3167 idalloc(void *ptr) 3168 { 3169 arena_chunk_t *chunk; 3170 3171 assert(ptr != NULL); 3172 3173 chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr); 3174 if (chunk != ptr) { 3175 /* Region. */ 3176 arena_dalloc(chunk->arena, chunk, ptr); 3177 } else 3178 huge_dalloc(ptr); 3179 } 3180 3181 static void 3182 malloc_print_stats(void) 3183 { 3184 3185 if (opt_print_stats) { 3186 char s[UMAX2S_BUFSIZE]; 3187 _malloc_message("___ Begin malloc statistics ___\n", "", "", 3188 ""); 3189 _malloc_message("Assertions ", 3190 #ifdef NDEBUG 3191 "disabled", 3192 #else 3193 "enabled", 3194 #endif 3195 "\n", ""); 3196 _malloc_message("Boolean MALLOC_OPTIONS: ", 3197 opt_abort ? "A" : "a", 3198 opt_junk ? "J" : "j", 3199 opt_hint ? "H" : "h"); 3200 _malloc_message(opt_utrace ? "PU" : "Pu", 3201 opt_sysv ? "V" : "v", 3202 opt_xmalloc ? "X" : "x", 3203 opt_zero ? "Z\n" : "z\n"); 3204 3205 _malloc_message("CPUs: ", umax2s(ncpus, s), "\n", ""); 3206 _malloc_message("Max arenas: ", umax2s(narenas, s), "\n", ""); 3207 _malloc_message("Pointer size: ", umax2s(sizeof(void *), s), 3208 "\n", ""); 3209 _malloc_message("Quantum size: ", umax2s(quantum, s), "\n", ""); 3210 _malloc_message("Max small size: ", umax2s(small_max, s), "\n", 3211 ""); 3212 3213 _malloc_message("Chunk size: ", umax2s(chunksize, s), "", ""); 3214 _malloc_message(" (2^", umax2s(opt_chunk_2pow, s), ")\n", ""); 3215 3216 #ifdef MALLOC_STATS 3217 { 3218 size_t allocated, mapped; 3219 unsigned i; 3220 arena_t *arena; 3221 3222 /* Calculate and print allocated/mapped stats. */ 3223 3224 /* arenas. */ 3225 for (i = 0, allocated = 0; i < narenas; i++) { 3226 if (arenas[i] != NULL) { 3227 malloc_mutex_lock(&arenas[i]->mtx); 3228 allocated += 3229 arenas[i]->stats.allocated_small; 3230 allocated += 3231 arenas[i]->stats.allocated_large; 3232 malloc_mutex_unlock(&arenas[i]->mtx); 3233 } 3234 } 3235 3236 /* huge/base. */ 3237 malloc_mutex_lock(&chunks_mtx); 3238 allocated += huge_allocated; 3239 mapped = stats_chunks.curchunks * chunksize; 3240 malloc_mutex_unlock(&chunks_mtx); 3241 3242 malloc_mutex_lock(&base_mtx); 3243 mapped += base_mapped; 3244 malloc_mutex_unlock(&base_mtx); 3245 3246 malloc_printf("Allocated: %zu, mapped: %zu\n", 3247 allocated, mapped); 3248 3249 /* Print chunk stats. */ 3250 { 3251 chunk_stats_t chunks_stats; 3252 3253 malloc_mutex_lock(&chunks_mtx); 3254 chunks_stats = stats_chunks; 3255 malloc_mutex_unlock(&chunks_mtx); 3256 3257 malloc_printf("chunks: nchunks " 3258 "highchunks curchunks\n"); 3259 malloc_printf(" %13llu%13lu%13lu\n", 3260 chunks_stats.nchunks, 3261 chunks_stats.highchunks, 3262 chunks_stats.curchunks); 3263 } 3264 3265 /* Print chunk stats. */ 3266 malloc_printf( 3267 "huge: nmalloc ndalloc " 3268 "nralloc allocated\n"); 3269 malloc_printf(" %12llu %12llu %12llu %12zu\n", 3270 huge_nmalloc, huge_ndalloc, huge_nralloc, 3271 huge_allocated); 3272 3273 /* Print stats for each arena. */ 3274 for (i = 0; i < narenas; i++) { 3275 arena = arenas[i]; 3276 if (arena != NULL) { 3277 malloc_printf( 3278 "\narenas[%u] @ %p\n", i, arena); 3279 malloc_mutex_lock(&arena->mtx); 3280 stats_print(arena); 3281 malloc_mutex_unlock(&arena->mtx); 3282 } 3283 } 3284 } 3285 #endif /* #ifdef MALLOC_STATS */ 3286 _malloc_message("--- End malloc statistics ---\n", "", "", ""); 3287 } 3288 } 3289 3290 /* 3291 * FreeBSD's pthreads implementation calls malloc(3), so the malloc 3292 * implementation has to take pains to avoid infinite recursion during 3293 * initialization. 3294 */ 3295 static inline bool 3296 malloc_init(void) 3297 { 3298 3299 if (malloc_initialized == false) 3300 return (malloc_init_hard()); 3301 3302 return (false); 3303 } 3304 3305 static bool 3306 malloc_init_hard(void) 3307 { 3308 unsigned i, j; 3309 ssize_t linklen; 3310 char buf[PATH_MAX + 1]; 3311 const char *opts = ""; 3312 3313 malloc_mutex_lock(&init_lock); 3314 if (malloc_initialized) { 3315 /* 3316 * Another thread initialized the allocator before this one 3317 * acquired init_lock. 3318 */ 3319 malloc_mutex_unlock(&init_lock); 3320 return (false); 3321 } 3322 3323 /* Get number of CPUs. */ 3324 { 3325 int mib[2]; 3326 size_t len; 3327 3328 mib[0] = CTL_HW; 3329 mib[1] = HW_NCPU; 3330 len = sizeof(ncpus); 3331 if (sysctl(mib, 2, &ncpus, &len, (void *) 0, 0) == -1) { 3332 /* Error. */ 3333 ncpus = 1; 3334 } 3335 } 3336 3337 /* Get page size. */ 3338 { 3339 long result; 3340 3341 result = sysconf(_SC_PAGESIZE); 3342 assert(result != -1); 3343 pagesize = (unsigned) result; 3344 3345 /* 3346 * We assume that pagesize is a power of 2 when calculating 3347 * pagesize_mask and pagesize_2pow. 3348 */ 3349 assert(((result - 1) & result) == 0); 3350 pagesize_mask = result - 1; 3351 pagesize_2pow = ffs((int)result) - 1; 3352 } 3353 3354 for (i = 0; i < 3; i++) { 3355 /* Get runtime configuration. */ 3356 switch (i) { 3357 case 0: 3358 if ((linklen = readlink("/etc/malloc.conf", buf, 3359 sizeof(buf) - 1)) != -1) { 3360 /* 3361 * Use the contents of the "/etc/malloc.conf" 3362 * symbolic link's name. 3363 */ 3364 buf[linklen] = '\0'; 3365 opts = buf; 3366 } else { 3367 /* No configuration specified. */ 3368 buf[0] = '\0'; 3369 opts = buf; 3370 } 3371 break; 3372 case 1: 3373 if ((opts = getenv("MALLOC_OPTIONS")) != NULL && 3374 issetugid() == 0) { 3375 /* 3376 * Do nothing; opts is already initialized to 3377 * the value of the MALLOC_OPTIONS environment 3378 * variable. 3379 */ 3380 } else { 3381 /* No configuration specified. */ 3382 buf[0] = '\0'; 3383 opts = buf; 3384 } 3385 break; 3386 case 2: 3387 if (_malloc_options != NULL) { 3388 /* 3389 * Use options that were compiled into the program. 3390 */ 3391 opts = _malloc_options; 3392 } else { 3393 /* No configuration specified. */ 3394 buf[0] = '\0'; 3395 opts = buf; 3396 } 3397 break; 3398 default: 3399 /* NOTREACHED */ 3400 /* LINTED */ 3401 assert(false); 3402 } 3403 3404 for (j = 0; opts[j] != '\0'; j++) { 3405 switch (opts[j]) { 3406 case 'a': 3407 opt_abort = false; 3408 break; 3409 case 'A': 3410 opt_abort = true; 3411 break; 3412 case 'h': 3413 opt_hint = false; 3414 break; 3415 case 'H': 3416 opt_hint = true; 3417 break; 3418 case 'j': 3419 opt_junk = false; 3420 break; 3421 case 'J': 3422 opt_junk = true; 3423 break; 3424 case 'k': 3425 /* 3426 * Chunks always require at least one header 3427 * page, so chunks can never be smaller than 3428 * two pages. 3429 */ 3430 if (opt_chunk_2pow > pagesize_2pow + 1) 3431 opt_chunk_2pow--; 3432 break; 3433 case 'K': 3434 if (opt_chunk_2pow + 1 < 3435 (int)(sizeof(size_t) << 3)) 3436 opt_chunk_2pow++; 3437 break; 3438 case 'n': 3439 opt_narenas_lshift--; 3440 break; 3441 case 'N': 3442 opt_narenas_lshift++; 3443 break; 3444 case 'p': 3445 opt_print_stats = false; 3446 break; 3447 case 'P': 3448 opt_print_stats = true; 3449 break; 3450 case 'q': 3451 if (opt_quantum_2pow > QUANTUM_2POW_MIN) 3452 opt_quantum_2pow--; 3453 break; 3454 case 'Q': 3455 if (opt_quantum_2pow < pagesize_2pow - 1) 3456 opt_quantum_2pow++; 3457 break; 3458 case 's': 3459 if (opt_small_max_2pow > QUANTUM_2POW_MIN) 3460 opt_small_max_2pow--; 3461 break; 3462 case 'S': 3463 if (opt_small_max_2pow < pagesize_2pow - 1) 3464 opt_small_max_2pow++; 3465 break; 3466 case 'u': 3467 opt_utrace = false; 3468 break; 3469 case 'U': 3470 opt_utrace = true; 3471 break; 3472 case 'v': 3473 opt_sysv = false; 3474 break; 3475 case 'V': 3476 opt_sysv = true; 3477 break; 3478 case 'x': 3479 opt_xmalloc = false; 3480 break; 3481 case 'X': 3482 opt_xmalloc = true; 3483 break; 3484 case 'z': 3485 opt_zero = false; 3486 break; 3487 case 'Z': 3488 opt_zero = true; 3489 break; 3490 default: { 3491 char cbuf[2]; 3492 3493 cbuf[0] = opts[j]; 3494 cbuf[1] = '\0'; 3495 _malloc_message(getprogname(), 3496 ": (malloc) Unsupported character in " 3497 "malloc options: '", cbuf, "'\n"); 3498 } 3499 } 3500 } 3501 } 3502 3503 /* Take care to call atexit() only once. */ 3504 if (opt_print_stats) { 3505 /* Print statistics at exit. */ 3506 atexit(malloc_print_stats); 3507 } 3508 3509 /* Set variables according to the value of opt_small_max_2pow. */ 3510 if (opt_small_max_2pow < opt_quantum_2pow) 3511 opt_small_max_2pow = opt_quantum_2pow; 3512 small_max = (1 << opt_small_max_2pow); 3513 3514 /* Set bin-related variables. */ 3515 bin_maxclass = (pagesize >> 1); 3516 assert(opt_quantum_2pow >= TINY_MIN_2POW); 3517 ntbins = (unsigned)(opt_quantum_2pow - TINY_MIN_2POW); 3518 assert(ntbins <= opt_quantum_2pow); 3519 nqbins = (unsigned)(small_max >> opt_quantum_2pow); 3520 nsbins = (unsigned)(pagesize_2pow - opt_small_max_2pow - 1); 3521 3522 /* Set variables according to the value of opt_quantum_2pow. */ 3523 quantum = (1 << opt_quantum_2pow); 3524 quantum_mask = quantum - 1; 3525 if (ntbins > 0) 3526 small_min = (quantum >> 1) + 1; 3527 else 3528 small_min = 1; 3529 assert(small_min <= quantum); 3530 3531 /* Set variables according to the value of opt_chunk_2pow. */ 3532 chunksize = (1LU << opt_chunk_2pow); 3533 chunksize_mask = chunksize - 1; 3534 chunksize_2pow = (unsigned)opt_chunk_2pow; 3535 chunk_npages = (unsigned)(chunksize >> pagesize_2pow); 3536 { 3537 unsigned header_size; 3538 3539 header_size = (unsigned)(sizeof(arena_chunk_t) + 3540 (sizeof(arena_chunk_map_t) * (chunk_npages - 1))); 3541 arena_chunk_header_npages = (header_size >> pagesize_2pow); 3542 if ((header_size & pagesize_mask) != 0) 3543 arena_chunk_header_npages++; 3544 } 3545 arena_maxclass = chunksize - (arena_chunk_header_npages << 3546 pagesize_2pow); 3547 3548 UTRACE(0, 0, 0); 3549 3550 #ifdef MALLOC_STATS 3551 memset(&stats_chunks, 0, sizeof(chunk_stats_t)); 3552 #endif 3553 3554 /* Various sanity checks that regard configuration. */ 3555 assert(quantum >= sizeof(void *)); 3556 assert(quantum <= pagesize); 3557 assert(chunksize >= pagesize); 3558 assert(quantum * 4 <= chunksize); 3559 3560 /* Initialize chunks data. */ 3561 malloc_mutex_init(&chunks_mtx); 3562 RB_INIT(&huge); 3563 #ifdef USE_BRK 3564 malloc_mutex_init(&brk_mtx); 3565 brk_base = sbrk(0); 3566 brk_prev = brk_base; 3567 brk_max = brk_base; 3568 #endif 3569 #ifdef MALLOC_STATS 3570 huge_nmalloc = 0; 3571 huge_ndalloc = 0; 3572 huge_nralloc = 0; 3573 huge_allocated = 0; 3574 #endif 3575 RB_INIT(&old_chunks); 3576 3577 /* Initialize base allocation data structures. */ 3578 #ifdef MALLOC_STATS 3579 base_mapped = 0; 3580 #endif 3581 #ifdef USE_BRK 3582 /* 3583 * Allocate a base chunk here, since it doesn't actually have to be 3584 * chunk-aligned. Doing this before allocating any other chunks allows 3585 * the use of space that would otherwise be wasted. 3586 */ 3587 base_pages_alloc(0); 3588 #endif 3589 base_chunk_nodes = NULL; 3590 malloc_mutex_init(&base_mtx); 3591 3592 if (ncpus > 1) { 3593 /* 3594 * For SMP systems, create four times as many arenas as there 3595 * are CPUs by default. 3596 */ 3597 opt_narenas_lshift += 2; 3598 } 3599 3600 #ifdef NO_TLS 3601 /* Initialize arena key. */ 3602 (void)thr_keycreate(&arenas_map_key, NULL); 3603 #endif 3604 3605 /* Determine how many arenas to use. */ 3606 narenas = ncpus; 3607 if (opt_narenas_lshift > 0) { 3608 if ((narenas << opt_narenas_lshift) > narenas) 3609 narenas <<= opt_narenas_lshift; 3610 /* 3611 * Make sure not to exceed the limits of what base_malloc() 3612 * can handle. 3613 */ 3614 if (narenas * sizeof(arena_t *) > chunksize) 3615 narenas = (unsigned)(chunksize / sizeof(arena_t *)); 3616 } else if (opt_narenas_lshift < 0) { 3617 if ((narenas << opt_narenas_lshift) < narenas) 3618 narenas <<= opt_narenas_lshift; 3619 /* Make sure there is at least one arena. */ 3620 if (narenas == 0) 3621 narenas = 1; 3622 } 3623 3624 next_arena = 0; 3625 3626 /* Allocate and initialize arenas. */ 3627 arenas = (arena_t **)base_alloc(sizeof(arena_t *) * narenas); 3628 if (arenas == NULL) { 3629 malloc_mutex_unlock(&init_lock); 3630 return (true); 3631 } 3632 /* 3633 * Zero the array. In practice, this should always be pre-zeroed, 3634 * since it was just mmap()ed, but let's be sure. 3635 */ 3636 memset(arenas, 0, sizeof(arena_t *) * narenas); 3637 3638 /* 3639 * Initialize one arena here. The rest are lazily created in 3640 * arena_choose_hard(). 3641 */ 3642 arenas_extend(0); 3643 if (arenas[0] == NULL) { 3644 malloc_mutex_unlock(&init_lock); 3645 return (true); 3646 } 3647 3648 malloc_mutex_init(&arenas_mtx); 3649 3650 malloc_initialized = true; 3651 malloc_mutex_unlock(&init_lock); 3652 return (false); 3653 } 3654 3655 /* 3656 * End general internal functions. 3657 */ 3658 /******************************************************************************/ 3659 /* 3660 * Begin malloc(3)-compatible functions. 3661 */ 3662 3663 void * 3664 malloc(size_t size) 3665 { 3666 void *ret; 3667 3668 if (malloc_init()) { 3669 ret = NULL; 3670 goto RETURN; 3671 } 3672 3673 if (size == 0) { 3674 if (opt_sysv == false) 3675 size = 1; 3676 else { 3677 ret = NULL; 3678 goto RETURN; 3679 } 3680 } 3681 3682 ret = imalloc(size); 3683 3684 RETURN: 3685 if (ret == NULL) { 3686 if (opt_xmalloc) { 3687 _malloc_message(getprogname(), 3688 ": (malloc) Error in malloc(): out of memory\n", "", 3689 ""); 3690 abort(); 3691 } 3692 errno = ENOMEM; 3693 } 3694 3695 UTRACE(0, size, ret); 3696 return (ret); 3697 } 3698 3699 int 3700 posix_memalign(void **memptr, size_t alignment, size_t size) 3701 { 3702 int ret; 3703 void *result; 3704 3705 if (malloc_init()) 3706 result = NULL; 3707 else { 3708 /* Make sure that alignment is a large enough power of 2. */ 3709 if (((alignment - 1) & alignment) != 0 3710 || alignment < sizeof(void *)) { 3711 if (opt_xmalloc) { 3712 _malloc_message(getprogname(), 3713 ": (malloc) Error in posix_memalign(): " 3714 "invalid alignment\n", "", ""); 3715 abort(); 3716 } 3717 result = NULL; 3718 ret = EINVAL; 3719 goto RETURN; 3720 } 3721 3722 result = ipalloc(alignment, size); 3723 } 3724 3725 if (result == NULL) { 3726 if (opt_xmalloc) { 3727 _malloc_message(getprogname(), 3728 ": (malloc) Error in posix_memalign(): out of memory\n", 3729 "", ""); 3730 abort(); 3731 } 3732 ret = ENOMEM; 3733 goto RETURN; 3734 } 3735 3736 *memptr = result; 3737 ret = 0; 3738 3739 RETURN: 3740 UTRACE(0, size, result); 3741 return (ret); 3742 } 3743 3744 void * 3745 calloc(size_t num, size_t size) 3746 { 3747 void *ret; 3748 size_t num_size; 3749 3750 if (malloc_init()) { 3751 num_size = 0; 3752 ret = NULL; 3753 goto RETURN; 3754 } 3755 3756 num_size = num * size; 3757 if (num_size == 0) { 3758 if ((opt_sysv == false) && ((num == 0) || (size == 0))) 3759 num_size = 1; 3760 else { 3761 ret = NULL; 3762 goto RETURN; 3763 } 3764 /* 3765 * Try to avoid division here. We know that it isn't possible to 3766 * overflow during multiplication if neither operand uses any of the 3767 * most significant half of the bits in a size_t. 3768 */ 3769 } else if ((unsigned long long)((num | size) & 3770 ((unsigned long long)SIZE_T_MAX << (sizeof(size_t) << 2))) && 3771 (num_size / size != num)) { 3772 /* size_t overflow. */ 3773 ret = NULL; 3774 goto RETURN; 3775 } 3776 3777 ret = icalloc(num_size); 3778 3779 RETURN: 3780 if (ret == NULL) { 3781 if (opt_xmalloc) { 3782 _malloc_message(getprogname(), 3783 ": (malloc) Error in calloc(): out of memory\n", "", 3784 ""); 3785 abort(); 3786 } 3787 errno = ENOMEM; 3788 } 3789 3790 UTRACE(0, num_size, ret); 3791 return (ret); 3792 } 3793 3794 void * 3795 realloc(void *ptr, size_t size) 3796 { 3797 void *ret; 3798 3799 if (size == 0) { 3800 if (opt_sysv == false) 3801 size = 1; 3802 else { 3803 if (ptr != NULL) 3804 idalloc(ptr); 3805 ret = NULL; 3806 goto RETURN; 3807 } 3808 } 3809 3810 if (ptr != NULL) { 3811 assert(malloc_initialized); 3812 3813 ret = iralloc(ptr, size); 3814 3815 if (ret == NULL) { 3816 if (opt_xmalloc) { 3817 _malloc_message(getprogname(), 3818 ": (malloc) Error in realloc(): out of " 3819 "memory\n", "", ""); 3820 abort(); 3821 } 3822 errno = ENOMEM; 3823 } 3824 } else { 3825 if (malloc_init()) 3826 ret = NULL; 3827 else 3828 ret = imalloc(size); 3829 3830 if (ret == NULL) { 3831 if (opt_xmalloc) { 3832 _malloc_message(getprogname(), 3833 ": (malloc) Error in realloc(): out of " 3834 "memory\n", "", ""); 3835 abort(); 3836 } 3837 errno = ENOMEM; 3838 } 3839 } 3840 3841 RETURN: 3842 UTRACE(ptr, size, ret); 3843 return (ret); 3844 } 3845 3846 void 3847 free(void *ptr) 3848 { 3849 3850 UTRACE(ptr, 0, 0); 3851 if (ptr != NULL) { 3852 assert(malloc_initialized); 3853 3854 idalloc(ptr); 3855 } 3856 } 3857 3858 /* 3859 * End malloc(3)-compatible functions. 3860 */ 3861 /******************************************************************************/ 3862 /* 3863 * Begin non-standard functions. 3864 */ 3865 #ifndef __NetBSD__ 3866 size_t 3867 malloc_usable_size(const void *ptr) 3868 { 3869 3870 assert(ptr != NULL); 3871 3872 return (isalloc(ptr)); 3873 } 3874 #endif 3875 3876 /* 3877 * End non-standard functions. 3878 */ 3879 /******************************************************************************/ 3880 /* 3881 * Begin library-private functions, used by threading libraries for protection 3882 * of malloc during fork(). These functions are only called if the program is 3883 * running in threaded mode, so there is no need to check whether the program 3884 * is threaded here. 3885 */ 3886 3887 void 3888 _malloc_prefork(void) 3889 { 3890 unsigned i; 3891 3892 /* Acquire all mutexes in a safe order. */ 3893 3894 malloc_mutex_lock(&arenas_mtx); 3895 for (i = 0; i < narenas; i++) { 3896 if (arenas[i] != NULL) 3897 malloc_mutex_lock(&arenas[i]->mtx); 3898 } 3899 malloc_mutex_unlock(&arenas_mtx); 3900 3901 malloc_mutex_lock(&base_mtx); 3902 3903 malloc_mutex_lock(&chunks_mtx); 3904 } 3905 3906 void 3907 _malloc_postfork(void) 3908 { 3909 unsigned i; 3910 3911 /* Release all mutexes, now that fork() has completed. */ 3912 3913 malloc_mutex_unlock(&chunks_mtx); 3914 3915 malloc_mutex_unlock(&base_mtx); 3916 3917 malloc_mutex_lock(&arenas_mtx); 3918 for (i = 0; i < narenas; i++) { 3919 if (arenas[i] != NULL) 3920 malloc_mutex_unlock(&arenas[i]->mtx); 3921 } 3922 malloc_mutex_unlock(&arenas_mtx); 3923 } 3924 3925 /* 3926 * End library-private functions. 3927 */ 3928 /******************************************************************************/ 3929