src/lj_alloc.c

   1 /*
   2 ** Bundled memory allocator.
   3 **
   4 ** Beware: this is a HEAVILY CUSTOMIZED version of dlmalloc.
   5 ** The original bears the following remark:
   6 **
   7 **   This is a version (aka dlmalloc) of malloc/free/realloc written by
   8 **   Doug Lea and released to the public domain, as explained at
   9 **   http://creativecommons.org/licenses/publicdomain.
  10 **
  11 **   * Version pre-2.8.4 Wed Mar 29 19:46:29 2006    (dl at gee)
  12 **
  13 ** No additional copyright is claimed over the customizations.
  14 ** Please do NOT bother the original author about this version here!
  15 **
  16 ** If you want to use dlmalloc in another project, you should get
  17 ** the original from: ftp://gee.cs.oswego.edu/pub/misc/
  18 ** For thread-safe derivatives, take a look at:
  19 ** - ptmalloc: http://www.malloc.de/
  20 ** - nedmalloc: http://www.nedprod.com/programs/portable/nedmalloc/
  21 */
  22
  23 #define lj_alloc_c
  24 #define LUA_CORE
  25
  26 /* To get the mremap prototype. Must be defined before any system includes. */
  27 #if defined(__linux__) && !defined(_GNU_SOURCE)
  28 #define _GNU_SOURCE
  29 #endif
  30
  31 #include "lj_def.h"
  32 #include "lj_arch.h"
  33 #include "lj_alloc.h"
  34
  35 #ifndef LUAJIT_USE_SYSMALLOC
  36
  37 #define MAX_SIZE_T              (~(size_t)0)
  38 #define MALLOC_ALIGNMENT        ((size_t)8U)
  39
  40 #define DEFAULT_GRANULARITY     ((size_t)128U * (size_t)1024U)
  41 #define DEFAULT_TRIM_THRESHOLD  ((size_t)2U * (size_t)1024U * (size_t)1024U)
  42 #define DEFAULT_MMAP_THRESHOLD  ((size_t)128U * (size_t)1024U)
  43 #define MAX_RELEASE_CHECK_RATE  255
  44
  45 /* ------------------- size_t and alignment properties -------------------- */
  46
  47 /* The byte and bit size of a size_t */
  48 #define SIZE_T_SIZE             (sizeof(size_t))
  49 #define SIZE_T_BITSIZE          (sizeof(size_t) << 3)
  50
  51 /* Some constants coerced to size_t */
  52 /* Annoying but necessary to avoid errors on some platforms */
  53 #define SIZE_T_ZERO             ((size_t)0)
  54 #define SIZE_T_ONE              ((size_t)1)
  55 #define SIZE_T_TWO              ((size_t)2)
  56 #define TWO_SIZE_T_SIZES        (SIZE_T_SIZE<<1)
  57 #define FOUR_SIZE_T_SIZES       (SIZE_T_SIZE<<2)
  58 #define SIX_SIZE_T_SIZES        (FOUR_SIZE_T_SIZES+TWO_SIZE_T_SIZES)
  59
  60 /* The bit mask value corresponding to MALLOC_ALIGNMENT */
  61 #define CHUNK_ALIGN_MASK        (MALLOC_ALIGNMENT - SIZE_T_ONE)
  62
  63 /* the number of bytes to offset an address to align it */
  64 #define align_offset(A)\
  65  ((((size_t)(A) & CHUNK_ALIGN_MASK) == 0)? 0 :\
  66   ((MALLOC_ALIGNMENT - ((size_t)(A) & CHUNK_ALIGN_MASK)) & CHUNK_ALIGN_MASK))
  67
  68 /* -------------------------- MMAP support ------------------------------- */
  69
  70 #define MFAIL                   ((void *)(MAX_SIZE_T))
  71 #define CMFAIL                  ((char *)(MFAIL)) /* defined for convenience */
  72
  73 #define IS_DIRECT_BIT           (SIZE_T_ONE)
  74
  75 #if LJ_TARGET_WINDOWS
  76
  77 #define WIN32_LEAN_AND_MEAN
  78 #include <windows.h>
  79
  80 #if LJ_64
  81
  82 /* Undocumented, but hey, that's what we all love so much about Windows. */
  83 typedef long (*PNTAVM)(HANDLE handle, void **addr, ULONG zbits,
  84                        size_t *size, ULONG alloctype, ULONG prot);
  85 static PNTAVM ntavm;
  86
  87 /* Number of top bits of the lower 32 bits of an address that must be zero.
  88 ** Apparently 0 gives us full 64 bit addresses and 1 gives us the lower 2GB.
  89 */
  90 #define NTAVM_ZEROBITS          1
  91
  92 static void INIT_MMAP(void)
  93 {
  94   ntavm = (PNTAVM)GetProcAddress(GetModuleHandleA("ntdll.dll"),
  95                                  "NtAllocateVirtualMemory");
  96 }
  97
  98 /* Win64 32 bit MMAP via NtAllocateVirtualMemory. */
  99 static LJ_AINLINE void *CALL_MMAP(size_t size)
 100 {
 101   DWORD olderr = GetLastError();
 102   void *ptr = NULL;
 103   long st = ntavm(INVALID_HANDLE_VALUE, &ptr, NTAVM_ZEROBITS, &size,
 104                   MEM_RESERVE|MEM_COMMIT, PAGE_READWRITE);
 105   SetLastError(olderr);
 106   return st == 0 ? ptr : MFAIL;
 107 }
 108
 109 /* For direct MMAP, use MEM_TOP_DOWN to minimize interference */
 110 static LJ_AINLINE void *DIRECT_MMAP(size_t size)
 111 {
 112   DWORD olderr = GetLastError();
 113   void *ptr = NULL;
 114   long st = ntavm(INVALID_HANDLE_VALUE, &ptr, NTAVM_ZEROBITS, &size,
 115                   MEM_RESERVE|MEM_COMMIT|MEM_TOP_DOWN, PAGE_READWRITE);
 116   SetLastError(olderr);
 117   return st == 0 ? ptr : MFAIL;
 118 }
 119
 120 #else
 121
 122 #define INIT_MMAP()             ((void)0)
 123
 124 /* Win32 MMAP via VirtualAlloc */
 125 static LJ_AINLINE void *CALL_MMAP(size_t size)
 126 {
 127   DWORD olderr = GetLastError();
 128   void *ptr = VirtualAlloc(0, size, MEM_RESERVE|MEM_COMMIT, PAGE_READWRITE);
 129   SetLastError(olderr);
 130   return ptr ? ptr : MFAIL;
 131 }
 132
 133 /* For direct MMAP, use MEM_TOP_DOWN to minimize interference */
 134 static LJ_AINLINE void *DIRECT_MMAP(size_t size)
 135 {
 136   DWORD olderr = GetLastError();
 137   void *ptr = VirtualAlloc(0, size, MEM_RESERVE|MEM_COMMIT|MEM_TOP_DOWN,
 138                            PAGE_READWRITE);
 139   SetLastError(olderr);
 140   return ptr ? ptr : MFAIL;
 141 }
 142
 143 #endif
 144
 145 /* This function supports releasing coalesed segments */
 146 static LJ_AINLINE int CALL_MUNMAP(void *ptr, size_t size)
 147 {
 148   DWORD olderr = GetLastError();
 149   MEMORY_BASIC_INFORMATION minfo;
 150   char *cptr = (char *)ptr;
 151   while (size) {
 152     if (VirtualQuery(cptr, &minfo, sizeof(minfo)) == 0)
 153       return -1;
 154     if (minfo.BaseAddress != cptr || minfo.AllocationBase != cptr ||
 155         minfo.State != MEM_COMMIT || minfo.RegionSize > size)
 156       return -1;
 157     if (VirtualFree(cptr, 0, MEM_RELEASE) == 0)
 158       return -1;
 159     cptr += minfo.RegionSize;
 160     size -= minfo.RegionSize;
 161   }
 162   SetLastError(olderr);
 163   return 0;
 164 }
 165
 166 #else
 167
 168 #include <errno.h>
 169 #include <sys/mman.h>
 170
 171 #define MMAP_PROT               (PROT_READ|PROT_WRITE)
 172 #if !defined(MAP_ANONYMOUS) && defined(MAP_ANON)
 173 #define MAP_ANONYMOUS           MAP_ANON
 174 #endif
 175 #define MMAP_FLAGS              (MAP_PRIVATE|MAP_ANONYMOUS)
 176
 177 #if LJ_64
 178 /* 64 bit mode needs special support for allocating memory in the lower 2GB. */
 179
 180 #if defined(MAP_32BIT)
 181
 182 /* Actually this only gives us max. 1GB in current Linux kernels. */
 183 static LJ_AINLINE void *CALL_MMAP(size_t size)
 184 {
 185   int olderr = errno;
 186   void *ptr = mmap(NULL, size, MMAP_PROT, MAP_32BIT|MMAP_FLAGS, -1, 0);
 187   errno = olderr;
 188   return ptr;
 189 }
 190
 191 #elif LJ_TARGET_OSX || LJ_TARGET_PS4 || defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__sun__)
 192
 193 /* OSX and FreeBSD mmap() use a naive first-fit linear search.
 194 ** That's perfect for us. Except that -pagezero_size must be set for OSX,
 195 ** otherwise the lower 4GB are blocked. And the 32GB RLIMIT_DATA needs
 196 ** to be reduced to 250MB on FreeBSD.
 197 */
 198 #if LJ_TARGET_OSX
 199 #define MMAP_REGION_START       ((uintptr_t)0x10000)
 200 #elif LJ_TARGET_PS4
 201 #define MMAP_REGION_START       ((uintptr_t)0x4000)
 202 #else
 203 #define MMAP_REGION_START       ((uintptr_t)0x10000000)
 204 #endif
 205 #define MMAP_REGION_END         ((uintptr_t)0x80000000)
 206
 207 #if (defined(__FreeBSD__) || defined(__FreeBSD_kernel__)) && !LJ_TARGET_PS4
 208 #include <sys/resource.h>
 209 #endif
 210
 211 static LJ_AINLINE void *CALL_MMAP(size_t size)
 212 {
 213   int olderr = errno;
 214   /* Hint for next allocation. Doesn't need to be thread-safe. */
 215   static uintptr_t alloc_hint = MMAP_REGION_START;
 216   int retry = 0;
 217 #if (defined(__FreeBSD__) || defined(__FreeBSD_kernel__)) && !LJ_TARGET_PS4
 218   static int rlimit_modified = 0;
 219   if (LJ_UNLIKELY(rlimit_modified == 0)) {
 220     struct rlimit rlim;
 221     rlim.rlim_cur = rlim.rlim_max = MMAP_REGION_START;
 222     setrlimit(RLIMIT_DATA, &rlim);  /* Ignore result. May fail below. */
 223     rlimit_modified = 1;
 224   }
 225 #endif
 226   for (;;) {
 227     void *p = mmap((void *)alloc_hint, size, MMAP_PROT, MMAP_FLAGS, -1, 0);
 228     if ((uintptr_t)p >= MMAP_REGION_START &&
 229         (uintptr_t)p + size < MMAP_REGION_END) {
 230       alloc_hint = (uintptr_t)p + size;
 231       errno = olderr;
 232       return p;
 233     }
 234     if (p != CMFAIL) munmap(p, size);
 235 #ifdef __sun__
 236     alloc_hint += 0x1000000;  /* Need near-exhaustive linear scan. */
 237     if (alloc_hint + size < MMAP_REGION_END) continue;
 238 #endif
 239     if (retry) break;
 240     retry = 1;
 241     alloc_hint = MMAP_REGION_START;
 242   }
 243   errno = olderr;
 244   return CMFAIL;
 245 }
 246
 247 #else
 248
 249 #error "NYI: need an equivalent of MAP_32BIT for this 64 bit OS"
 250
 251 #endif
 252
 253 #else
 254
 255 /* 32 bit mode is easy. */
 256 static LJ_AINLINE void *CALL_MMAP(size_t size)
 257 {
 258   int olderr = errno;
 259   void *ptr = mmap(NULL, size, MMAP_PROT, MMAP_FLAGS, -1, 0);
 260   errno = olderr;
 261   return ptr;
 262 }
 263
 264 #endif
 265
 266 #define INIT_MMAP()             ((void)0)
 267 #define DIRECT_MMAP(s)          CALL_MMAP(s)
 268
 269 static LJ_AINLINE int CALL_MUNMAP(void *ptr, size_t size)
 270 {
 271   int olderr = errno;
 272   int ret = munmap(ptr, size);
 273   errno = olderr;
 274   return ret;
 275 }
 276
 277 #if LJ_TARGET_LINUX
 278 /* Need to define _GNU_SOURCE to get the mremap prototype. */
 279 static LJ_AINLINE void *CALL_MREMAP_(void *ptr, size_t osz, size_t nsz,
 280                                      int flags)
 281 {
 282   int olderr = errno;
 283   ptr = mremap(ptr, osz, nsz, flags);
 284   errno = olderr;
 285   return ptr;
 286 }
 287
 288 #define CALL_MREMAP(addr, osz, nsz, mv) CALL_MREMAP_((addr), (osz), (nsz), (mv))
 289 #define CALL_MREMAP_NOMOVE      0
 290 #define CALL_MREMAP_MAYMOVE     1
 291 #if LJ_64
 292 #define CALL_MREMAP_MV          CALL_MREMAP_NOMOVE
 293 #else
 294 #define CALL_MREMAP_MV          CALL_MREMAP_MAYMOVE
 295 #endif
 296 #endif
 297
 298 #endif
 299
 300 #ifndef CALL_MREMAP
 301 #define CALL_MREMAP(addr, osz, nsz, mv) ((void)osz, MFAIL)
 302 #endif
 303
 304 /* -----------------------  Chunk representations ------------------------ */
 305
 306 struct malloc_chunk {
 307   size_t               prev_foot;  /* Size of previous chunk (if free).  */
 308   size_t               head;       /* Size and inuse bits. */
 309   struct malloc_chunk *fd;         /* double links -- used only if free. */
 310   struct malloc_chunk *bk;
 311 };
 312
 313 typedef struct malloc_chunk  mchunk;
 314 typedef struct malloc_chunk *mchunkptr;
 315 typedef struct malloc_chunk *sbinptr;  /* The type of bins of chunks */
 316 typedef size_t bindex_t;               /* Described below */
 317 typedef unsigned int binmap_t;         /* Described below */
 318 typedef unsigned int flag_t;           /* The type of various bit flag sets */
 319
 320 /* ------------------- Chunks sizes and alignments ----------------------- */
 321
 322 #define MCHUNK_SIZE             (sizeof(mchunk))
 323
 324 #define CHUNK_OVERHEAD          (SIZE_T_SIZE)
 325
 326 /* Direct chunks need a second word of overhead ... */
 327 #define DIRECT_CHUNK_OVERHEAD   (TWO_SIZE_T_SIZES)
 328 /* ... and additional padding for fake next-chunk at foot */
 329 #define DIRECT_FOOT_PAD         (FOUR_SIZE_T_SIZES)
 330
 331 /* The smallest size we can malloc is an aligned minimal chunk */
 332 #define MIN_CHUNK_SIZE\
 333   ((MCHUNK_SIZE + CHUNK_ALIGN_MASK) & ~CHUNK_ALIGN_MASK)
 334
 335 /* conversion from malloc headers to user pointers, and back */
 336 #define chunk2mem(p)            ((void *)((char *)(p) + TWO_SIZE_T_SIZES))
 337 #define mem2chunk(mem)          ((mchunkptr)((char *)(mem) - TWO_SIZE_T_SIZES))
 338 /* chunk associated with aligned address A */
 339 #define align_as_chunk(A)       (mchunkptr)((A) + align_offset(chunk2mem(A)))
 340
 341 /* Bounds on request (not chunk) sizes. */
 342 #define MAX_REQUEST             ((~MIN_CHUNK_SIZE+1) << 2)
 343 #define MIN_REQUEST             (MIN_CHUNK_SIZE - CHUNK_OVERHEAD - SIZE_T_ONE)
 344
 345 /* pad request bytes into a usable size */
 346 #define pad_request(req) \
 347    (((req) + CHUNK_OVERHEAD + CHUNK_ALIGN_MASK) & ~CHUNK_ALIGN_MASK)
 348
 349 /* pad request, checking for minimum (but not maximum) */
 350 #define request2size(req) \
 351   (((req) < MIN_REQUEST)? MIN_CHUNK_SIZE : pad_request(req))
 352
 353 /* ------------------ Operations on head and foot fields ----------------- */
 354
 355 #define PINUSE_BIT              (SIZE_T_ONE)
 356 #define CINUSE_BIT              (SIZE_T_TWO)
 357 #define INUSE_BITS              (PINUSE_BIT|CINUSE_BIT)
 358
 359 /* Head value for fenceposts */
 360 #define FENCEPOST_HEAD          (INUSE_BITS|SIZE_T_SIZE)
 361
 362 /* extraction of fields from head words */
 363 #define cinuse(p)               ((p)->head & CINUSE_BIT)
 364 #define pinuse(p)               ((p)->head & PINUSE_BIT)
 365 #define chunksize(p)            ((p)->head & ~(INUSE_BITS))
 366
 367 #define clear_pinuse(p)         ((p)->head &= ~PINUSE_BIT)
 368 #define clear_cinuse(p)         ((p)->head &= ~CINUSE_BIT)
 369
 370 /* Treat space at ptr +/- offset as a chunk */
 371 #define chunk_plus_offset(p, s)         ((mchunkptr)(((char *)(p)) + (s)))
 372 #define chunk_minus_offset(p, s)        ((mchunkptr)(((char *)(p)) - (s)))
 373
 374 /* Ptr to next or previous physical malloc_chunk. */
 375 #define next_chunk(p)   ((mchunkptr)(((char *)(p)) + ((p)->head & ~INUSE_BITS)))
 376 #define prev_chunk(p)   ((mchunkptr)(((char *)(p)) - ((p)->prev_foot) ))
 377
 378 /* extract next chunk's pinuse bit */
 379 #define next_pinuse(p)  ((next_chunk(p)->head) & PINUSE_BIT)
 380
 381 /* Get/set size at footer */
 382 #define get_foot(p, s)  (((mchunkptr)((char *)(p) + (s)))->prev_foot)
 383 #define set_foot(p, s)  (((mchunkptr)((char *)(p) + (s)))->prev_foot = (s))
 384
 385 /* Set size, pinuse bit, and foot */
 386 #define set_size_and_pinuse_of_free_chunk(p, s)\
 387   ((p)->head = (s|PINUSE_BIT), set_foot(p, s))
 388
 389 /* Set size, pinuse bit, foot, and clear next pinuse */
 390 #define set_free_with_pinuse(p, s, n)\
 391   (clear_pinuse(n), set_size_and_pinuse_of_free_chunk(p, s))
 392
 393 #define is_direct(p)\
 394   (!((p)->head & PINUSE_BIT) && ((p)->prev_foot & IS_DIRECT_BIT))
 395
 396 /* Get the internal overhead associated with chunk p */
 397 #define overhead_for(p)\
 398  (is_direct(p)? DIRECT_CHUNK_OVERHEAD : CHUNK_OVERHEAD)
 399
 400 /* ---------------------- Overlaid data structures ----------------------- */
 401
 402 struct malloc_tree_chunk {
 403   /* The first four fields must be compatible with malloc_chunk */
 404   size_t                    prev_foot;
 405   size_t                    head;
 406   struct malloc_tree_chunk *fd;
 407   struct malloc_tree_chunk *bk;
 408
 409   struct malloc_tree_chunk *child[2];
 410   struct malloc_tree_chunk *parent;
 411   bindex_t                  index;
 412 };
 413
 414 typedef struct malloc_tree_chunk  tchunk;
 415 typedef struct malloc_tree_chunk *tchunkptr;
 416 typedef struct malloc_tree_chunk *tbinptr; /* The type of bins of trees */
 417
 418 /* A little helper macro for trees */
 419 #define leftmost_child(t) ((t)->child[0] != 0? (t)->child[0] : (t)->child[1])
 420
 421 /* ----------------------------- Segments -------------------------------- */
 422
 423 struct malloc_segment {
 424   char        *base;             /* base address */
 425   size_t       size;             /* allocated size */
 426   struct malloc_segment *next;   /* ptr to next segment */
 427 };
 428
 429 typedef struct malloc_segment  msegment;
 430 typedef struct malloc_segment *msegmentptr;
 431
 432 /* ---------------------------- malloc_state ----------------------------- */
 433
 434 /* Bin types, widths and sizes */
 435 #define NSMALLBINS              (32U)
 436 #define NTREEBINS               (32U)
 437 #define SMALLBIN_SHIFT          (3U)
 438 #define SMALLBIN_WIDTH          (SIZE_T_ONE << SMALLBIN_SHIFT)
 439 #define TREEBIN_SHIFT           (8U)
 440 #define MIN_LARGE_SIZE          (SIZE_T_ONE << TREEBIN_SHIFT)
 441 #define MAX_SMALL_SIZE          (MIN_LARGE_SIZE - SIZE_T_ONE)
 442 #define MAX_SMALL_REQUEST  (MAX_SMALL_SIZE - CHUNK_ALIGN_MASK - CHUNK_OVERHEAD)
 443
 444 struct malloc_state {
 445   binmap_t   smallmap;
 446   binmap_t   treemap;
 447   size_t     dvsize;
 448   size_t     topsize;
 449   mchunkptr  dv;
 450   mchunkptr  top;
 451   size_t     trim_check;
 452   size_t     release_checks;
 453   mchunkptr  smallbins[(NSMALLBINS+1)*2];
 454   tbinptr    treebins[NTREEBINS];
 455   msegment   seg;
 456 };
 457
 458 typedef struct malloc_state *mstate;
 459
 460 #define is_initialized(M)       ((M)->top != 0)
 461
 462 /* -------------------------- system alloc setup ------------------------- */
 463
 464 /* page-align a size */
 465 #define page_align(S)\
 466  (((S) + (LJ_PAGESIZE - SIZE_T_ONE)) & ~(LJ_PAGESIZE - SIZE_T_ONE))
 467
 468 /* granularity-align a size */
 469 #define granularity_align(S)\
 470   (((S) + (DEFAULT_GRANULARITY - SIZE_T_ONE))\
 471    & ~(DEFAULT_GRANULARITY - SIZE_T_ONE))
 472
 473 #if LJ_TARGET_WINDOWS
 474 #define mmap_align(S)   granularity_align(S)
 475 #else
 476 #define mmap_align(S)   page_align(S)
 477 #endif
 478
 479 /*  True if segment S holds address A */
 480 #define segment_holds(S, A)\
 481   ((char *)(A) >= S->base && (char *)(A) < S->base + S->size)
 482
 483 /* Return segment holding given address */
 484 static msegmentptr segment_holding(mstate m, char *addr)
 485 {
 486   msegmentptr sp = &m->seg;
 487   for (;;) {
 488     if (addr >= sp->base && addr < sp->base + sp->size)
 489       return sp;
 490     if ((sp = sp->next) == 0)
 491       return 0;
 492   }
 493 }
 494
 495 /* Return true if segment contains a segment link */
 496 static int has_segment_link(mstate m, msegmentptr ss)
 497 {
 498   msegmentptr sp = &m->seg;
 499   for (;;) {
 500     if ((char *)sp >= ss->base && (char *)sp < ss->base + ss->size)
 501       return 1;
 502     if ((sp = sp->next) == 0)
 503       return 0;
 504   }
 505 }
 506
 507 /*
 508   TOP_FOOT_SIZE is padding at the end of a segment, including space
 509   that may be needed to place segment records and fenceposts when new
 510   noncontiguous segments are added.
 511 */
 512 #define TOP_FOOT_SIZE\
 513   (align_offset(chunk2mem(0))+pad_request(sizeof(struct malloc_segment))+MIN_CHUNK_SIZE)
 514
 515 /* ---------------------------- Indexing Bins ---------------------------- */
 516
 517 #define is_small(s)             (((s) >> SMALLBIN_SHIFT) < NSMALLBINS)
 518 #define small_index(s)          ((s)  >> SMALLBIN_SHIFT)
 519 #define small_index2size(i)     ((i)  << SMALLBIN_SHIFT)
 520 #define MIN_SMALL_INDEX         (small_index(MIN_CHUNK_SIZE))
 521
 522 /* addressing by index. See above about smallbin repositioning */
 523 #define smallbin_at(M, i)       ((sbinptr)((char *)&((M)->smallbins[(i)<<1])))
 524 #define treebin_at(M,i)         (&((M)->treebins[i]))
 525
 526 /* assign tree index for size S to variable I */
 527 #define compute_tree_index(S, I)\
 528 {\
 529   unsigned int X = (unsigned int)(S >> TREEBIN_SHIFT);\
 530   if (X == 0) {\
 531     I = 0;\
 532   } else if (X > 0xFFFF) {\
 533     I = NTREEBINS-1;\
 534   } else {\
 535     unsigned int K = lj_fls(X);\
 536     I =  (bindex_t)((K << 1) + ((S >> (K + (TREEBIN_SHIFT-1)) & 1)));\
 537   }\
 538 }
 539
 540 /* Bit representing maximum resolved size in a treebin at i */
 541 #define bit_for_tree_index(i) \
 542    (i == NTREEBINS-1)? (SIZE_T_BITSIZE-1) : (((i) >> 1) + TREEBIN_SHIFT - 2)
 543
 544 /* Shift placing maximum resolved bit in a treebin at i as sign bit */
 545 #define leftshift_for_tree_index(i) \
 546    ((i == NTREEBINS-1)? 0 : \
 547     ((SIZE_T_BITSIZE-SIZE_T_ONE) - (((i) >> 1) + TREEBIN_SHIFT - 2)))
 548
 549 /* The size of the smallest chunk held in bin with index i */
 550 #define minsize_for_tree_index(i) \
 551    ((SIZE_T_ONE << (((i) >> 1) + TREEBIN_SHIFT)) |  \
 552    (((size_t)((i) & SIZE_T_ONE)) << (((i) >> 1) + TREEBIN_SHIFT - 1)))
 553
 554 /* ------------------------ Operations on bin maps ----------------------- */
 555
 556 /* bit corresponding to given index */
 557 #define idx2bit(i)              ((binmap_t)(1) << (i))
 558
 559 /* Mark/Clear bits with given index */
 560 #define mark_smallmap(M,i)      ((M)->smallmap |=  idx2bit(i))
 561 #define clear_smallmap(M,i)     ((M)->smallmap &= ~idx2bit(i))
 562 #define smallmap_is_marked(M,i) ((M)->smallmap &   idx2bit(i))
 563
 564 #define mark_treemap(M,i)       ((M)->treemap  |=  idx2bit(i))
 565 #define clear_treemap(M,i)      ((M)->treemap  &= ~idx2bit(i))
 566 #define treemap_is_marked(M,i)  ((M)->treemap  &   idx2bit(i))
 567
 568 /* mask with all bits to left of least bit of x on */
 569 #define left_bits(x)            ((x<<1) | (~(x<<1)+1))
 570
 571 /* Set cinuse bit and pinuse bit of next chunk */
 572 #define set_inuse(M,p,s)\
 573   ((p)->head = (((p)->head & PINUSE_BIT)|s|CINUSE_BIT),\
 574   ((mchunkptr)(((char *)(p)) + (s)))->head |= PINUSE_BIT)
 575
 576 /* Set cinuse and pinuse of this chunk and pinuse of next chunk */
 577 #define set_inuse_and_pinuse(M,p,s)\
 578   ((p)->head = (s|PINUSE_BIT|CINUSE_BIT),\
 579   ((mchunkptr)(((char *)(p)) + (s)))->head |= PINUSE_BIT)
 580
 581 /* Set size, cinuse and pinuse bit of this chunk */
 582 #define set_size_and_pinuse_of_inuse_chunk(M, p, s)\
 583   ((p)->head = (s|PINUSE_BIT|CINUSE_BIT))
 584
 585 /* ----------------------- Operations on smallbins ----------------------- */
 586
 587 /* Link a free chunk into a smallbin  */
 588 #define insert_small_chunk(M, P, S) {\
 589   bindex_t I = small_index(S);\
 590   mchunkptr B = smallbin_at(M, I);\
 591   mchunkptr F = B;\
 592   if (!smallmap_is_marked(M, I))\
 593     mark_smallmap(M, I);\
 594   else\
 595     F = B->fd;\
 596   B->fd = P;\
 597   F->bk = P;\
 598   P->fd = F;\
 599   P->bk = B;\
 600 }
 601
 602 /* Unlink a chunk from a smallbin  */
 603 #define unlink_small_chunk(M, P, S) {\
 604   mchunkptr F = P->fd;\
 605   mchunkptr B = P->bk;\
 606   bindex_t I = small_index(S);\
 607   if (F == B) {\
 608     clear_smallmap(M, I);\
 609   } else {\
 610     F->bk = B;\
 611     B->fd = F;\
 612   }\
 613 }
 614
 615 /* Unlink the first chunk from a smallbin */
 616 #define unlink_first_small_chunk(M, B, P, I) {\
 617   mchunkptr F = P->fd;\
 618   if (B == F) {\
 619     clear_smallmap(M, I);\
 620   } else {\
 621     B->fd = F;\
 622     F->bk = B;\
 623   }\
 624 }
 625
 626 /* Replace dv node, binning the old one */
 627 /* Used only when dvsize known to be small */
 628 #define replace_dv(M, P, S) {\
 629   size_t DVS = M->dvsize;\
 630   if (DVS != 0) {\
 631     mchunkptr DV = M->dv;\
 632     insert_small_chunk(M, DV, DVS);\
 633   }\
 634   M->dvsize = S;\
 635   M->dv = P;\
 636 }
 637
 638 /* ------------------------- Operations on trees ------------------------- */
 639
 640 /* Insert chunk into tree */
 641 #define insert_large_chunk(M, X, S) {\
 642   tbinptr *H;\
 643   bindex_t I;\
 644   compute_tree_index(S, I);\
 645   H = treebin_at(M, I);\
 646   X->index = I;\
 647   X->child[0] = X->child[1] = 0;\
 648   if (!treemap_is_marked(M, I)) {\
 649     mark_treemap(M, I);\
 650     *H = X;\
 651     X->parent = (tchunkptr)H;\
 652     X->fd = X->bk = X;\
 653   } else {\
 654     tchunkptr T = *H;\
 655     size_t K = S << leftshift_for_tree_index(I);\
 656     for (;;) {\
 657       if (chunksize(T) != S) {\
 658         tchunkptr *C = &(T->child[(K >> (SIZE_T_BITSIZE-SIZE_T_ONE)) & 1]);\
 659         K <<= 1;\
 660         if (*C != 0) {\
 661           T = *C;\
 662         } else {\
 663           *C = X;\
 664           X->parent = T;\
 665           X->fd = X->bk = X;\
 666           break;\
 667         }\
 668       } else {\
 669         tchunkptr F = T->fd;\
 670         T->fd = F->bk = X;\
 671         X->fd = F;\
 672         X->bk = T;\
 673         X->parent = 0;\
 674         break;\
 675       }\
 676     }\
 677   }\
 678 }
 679
 680 #define unlink_large_chunk(M, X) {\
 681   tchunkptr XP = X->parent;\
 682   tchunkptr R;\
 683   if (X->bk != X) {\
 684     tchunkptr F = X->fd;\
 685     R = X->bk;\
 686     F->bk = R;\
 687     R->fd = F;\
 688   } else {\
 689     tchunkptr *RP;\
 690     if (((R = *(RP = &(X->child[1]))) != 0) ||\
 691         ((R = *(RP = &(X->child[0]))) != 0)) {\
 692       tchunkptr *CP;\
 693       while ((*(CP = &(R->child[1])) != 0) ||\
 694              (*(CP = &(R->child[0])) != 0)) {\
 695         R = *(RP = CP);\
 696       }\
 697       *RP = 0;\
 698     }\
 699   }\
 700   if (XP != 0) {\
 701     tbinptr *H = treebin_at(M, X->index);\
 702     if (X == *H) {\
 703       if ((*H = R) == 0) \
 704         clear_treemap(M, X->index);\
 705     } else {\
 706       if (XP->child[0] == X) \
 707         XP->child[0] = R;\
 708       else \
 709         XP->child[1] = R;\
 710     }\
 711     if (R != 0) {\
 712       tchunkptr C0, C1;\
 713       R->parent = XP;\
 714       if ((C0 = X->child[0]) != 0) {\
 715         R->child[0] = C0;\
 716         C0->parent = R;\
 717       }\
 718       if ((C1 = X->child[1]) != 0) {\
 719         R->child[1] = C1;\
 720         C1->parent = R;\
 721       }\
 722     }\
 723   }\
 724 }
 725
 726 /* Relays to large vs small bin operations */
 727
 728 #define insert_chunk(M, P, S)\
 729   if (is_small(S)) { insert_small_chunk(M, P, S)\
 730   } else { tchunkptr TP = (tchunkptr)(P); insert_large_chunk(M, TP, S); }
 731
 732 #define unlink_chunk(M, P, S)\
 733   if (is_small(S)) { unlink_small_chunk(M, P, S)\
 734   } else { tchunkptr TP = (tchunkptr)(P); unlink_large_chunk(M, TP); }
 735
 736 /* -----------------------  Direct-mmapping chunks ----------------------- */
 737
 738 static void *direct_alloc(size_t nb)
 739 {
 740   size_t mmsize = mmap_align(nb + SIX_SIZE_T_SIZES + CHUNK_ALIGN_MASK);
 741   if (LJ_LIKELY(mmsize > nb)) {     /* Check for wrap around 0 */
 742     char *mm = (char *)(DIRECT_MMAP(mmsize));
 743     if (mm != CMFAIL) {
 744       size_t offset = align_offset(chunk2mem(mm));
 745       size_t psize = mmsize - offset - DIRECT_FOOT_PAD;
 746       mchunkptr p = (mchunkptr)(mm + offset);
 747       p->prev_foot = offset | IS_DIRECT_BIT;
 748       p->head = psize|CINUSE_BIT;
 749       chunk_plus_offset(p, psize)->head = FENCEPOST_HEAD;
 750       chunk_plus_offset(p, psize+SIZE_T_SIZE)->head = 0;
 751       return chunk2mem(p);
 752     }
 753   }
 754   return NULL;
 755 }
 756
 757 static mchunkptr direct_resize(mchunkptr oldp, size_t nb)
 758 {
 759   size_t oldsize = chunksize(oldp);
 760   if (is_small(nb)) /* Can't shrink direct regions below small size */
 761     return NULL;
 762   /* Keep old chunk if big enough but not too big */
 763   if (oldsize >= nb + SIZE_T_SIZE &&
 764       (oldsize - nb) <= (DEFAULT_GRANULARITY >> 1)) {
 765     return oldp;
 766   } else {
 767     size_t offset = oldp->prev_foot & ~IS_DIRECT_BIT;
 768     size_t oldmmsize = oldsize + offset + DIRECT_FOOT_PAD;
 769     size_t newmmsize = mmap_align(nb + SIX_SIZE_T_SIZES + CHUNK_ALIGN_MASK);
 770     char *cp = (char *)CALL_MREMAP((char *)oldp - offset,
 771                                    oldmmsize, newmmsize, CALL_MREMAP_MV);
 772     if (cp != CMFAIL) {
 773       mchunkptr newp = (mchunkptr)(cp + offset);
 774       size_t psize = newmmsize - offset - DIRECT_FOOT_PAD;
 775       newp->head = psize|CINUSE_BIT;
 776       chunk_plus_offset(newp, psize)->head = FENCEPOST_HEAD;
 777       chunk_plus_offset(newp, psize+SIZE_T_SIZE)->head = 0;
 778       return newp;
 779     }
 780   }
 781   return NULL;
 782 }
 783
 784 /* -------------------------- mspace management -------------------------- */
 785
 786 /* Initialize top chunk and its size */
 787 static void init_top(mstate m, mchunkptr p, size_t psize)
 788 {
 789   /* Ensure alignment */
 790   size_t offset = align_offset(chunk2mem(p));
 791   p = (mchunkptr)((char *)p + offset);
 792   psize -= offset;
 793
 794   m->top = p;
 795   m->topsize = psize;
 796   p->head = psize | PINUSE_BIT;
 797   /* set size of fake trailing chunk holding overhead space only once */
 798   chunk_plus_offset(p, psize)->head = TOP_FOOT_SIZE;
 799   m->trim_check = DEFAULT_TRIM_THRESHOLD; /* reset on each update */
 800 }
 801
 802 /* Initialize bins for a new mstate that is otherwise zeroed out */
 803 static void init_bins(mstate m)
 804 {
 805   /* Establish circular links for smallbins */
 806   bindex_t i;
 807   for (i = 0; i < NSMALLBINS; i++) {
 808     sbinptr bin = smallbin_at(m,i);
 809     bin->fd = bin->bk = bin;
 810   }
 811 }
 812
 813 /* Allocate chunk and prepend remainder with chunk in successor base. */
 814 static void *prepend_alloc(mstate m, char *newbase, char *oldbase, size_t nb)
 815 {
 816   mchunkptr p = align_as_chunk(newbase);
 817   mchunkptr oldfirst = align_as_chunk(oldbase);
 818   size_t psize = (size_t)((char *)oldfirst - (char *)p);
 819   mchunkptr q = chunk_plus_offset(p, nb);
 820   size_t qsize = psize - nb;
 821   set_size_and_pinuse_of_inuse_chunk(m, p, nb);
 822
 823   /* consolidate remainder with first chunk of old base */
 824   if (oldfirst == m->top) {
 825     size_t tsize = m->topsize += qsize;
 826     m->top = q;
 827     q->head = tsize | PINUSE_BIT;
 828   } else if (oldfirst == m->dv) {
 829     size_t dsize = m->dvsize += qsize;
 830     m->dv = q;
 831     set_size_and_pinuse_of_free_chunk(q, dsize);
 832   } else {
 833     if (!cinuse(oldfirst)) {
 834       size_t nsize = chunksize(oldfirst);
 835       unlink_chunk(m, oldfirst, nsize);
 836       oldfirst = chunk_plus_offset(oldfirst, nsize);
 837       qsize += nsize;
 838     }
 839     set_free_with_pinuse(q, qsize, oldfirst);
 840     insert_chunk(m, q, qsize);
 841   }
 842
 843   return chunk2mem(p);
 844 }
 845
 846 /* Add a segment to hold a new noncontiguous region */
 847 static void add_segment(mstate m, char *tbase, size_t tsize)
 848 {
 849   /* Determine locations and sizes of segment, fenceposts, old top */
 850   char *old_top = (char *)m->top;
 851   msegmentptr oldsp = segment_holding(m, old_top);
 852   char *old_end = oldsp->base + oldsp->size;
 853   size_t ssize = pad_request(sizeof(struct malloc_segment));
 854   char *rawsp = old_end - (ssize + FOUR_SIZE_T_SIZES + CHUNK_ALIGN_MASK);
 855   size_t offset = align_offset(chunk2mem(rawsp));
 856   char *asp = rawsp + offset;
 857   char *csp = (asp < (old_top + MIN_CHUNK_SIZE))? old_top : asp;
 858   mchunkptr sp = (mchunkptr)csp;
 859   msegmentptr ss = (msegmentptr)(chunk2mem(sp));
 860   mchunkptr tnext = chunk_plus_offset(sp, ssize);
 861   mchunkptr p = tnext;
 862
 863   /* reset top to new space */
 864   init_top(m, (mchunkptr)tbase, tsize - TOP_FOOT_SIZE);
 865
 866   /* Set up segment record */
 867   set_size_and_pinuse_of_inuse_chunk(m, sp, ssize);
 868   *ss = m->seg; /* Push current record */
 869   m->seg.base = tbase;
 870   m->seg.size = tsize;
 871   m->seg.next = ss;
 872
 873   /* Insert trailing fenceposts */
 874   for (;;) {
 875     mchunkptr nextp = chunk_plus_offset(p, SIZE_T_SIZE);
 876     p->head = FENCEPOST_HEAD;
 877     if ((char *)(&(nextp->head)) < old_end)
 878       p = nextp;
 879     else
 880       break;
 881   }
 882
 883   /* Insert the rest of old top into a bin as an ordinary free chunk */
 884   if (csp != old_top) {
 885     mchunkptr q = (mchunkptr)old_top;
 886     size_t psize = (size_t)(csp - old_top);
 887     mchunkptr tn = chunk_plus_offset(q, psize);
 888     set_free_with_pinuse(q, psize, tn);
 889     insert_chunk(m, q, psize);
 890   }
 891 }
 892
 893 /* -------------------------- System allocation -------------------------- */
 894
 895 static void *alloc_sys(mstate m, size_t nb)
 896 {
 897   char *tbase = CMFAIL;
 898   size_t tsize = 0;
 899
 900   /* Directly map large chunks */
 901   if (LJ_UNLIKELY(nb >= DEFAULT_MMAP_THRESHOLD)) {
 902     void *mem = direct_alloc(nb);
 903     if (mem != 0)
 904       return mem;
 905   }
 906
 907   {
 908     size_t req = nb + TOP_FOOT_SIZE + SIZE_T_ONE;
 909     size_t rsize = granularity_align(req);
 910     if (LJ_LIKELY(rsize > nb)) { /* Fail if wraps around zero */
 911       char *mp = (char *)(CALL_MMAP(rsize));
 912       if (mp != CMFAIL) {
 913         tbase = mp;
 914         tsize = rsize;
 915       }
 916     }
 917   }
 918
 919   if (tbase != CMFAIL) {
 920     msegmentptr sp = &m->seg;
 921     /* Try to merge with an existing segment */
 922     while (sp != 0 && tbase != sp->base + sp->size)
 923       sp = sp->next;
 924     if (sp != 0 && segment_holds(sp, m->top)) { /* append */
 925       sp->size += tsize;
 926       init_top(m, m->top, m->topsize + tsize);
 927     } else {
 928       sp = &m->seg;
 929       while (sp != 0 && sp->base != tbase + tsize)
 930         sp = sp->next;
 931       if (sp != 0) {
 932         char *oldbase = sp->base;
 933         sp->base = tbase;
 934         sp->size += tsize;
 935         return prepend_alloc(m, tbase, oldbase, nb);
 936       } else {
 937         add_segment(m, tbase, tsize);
 938       }
 939     }
 940
 941     if (nb < m->topsize) { /* Allocate from new or extended top space */
 942       size_t rsize = m->topsize -= nb;
 943       mchunkptr p = m->top;
 944       mchunkptr r = m->top = chunk_plus_offset(p, nb);
 945       r->head = rsize | PINUSE_BIT;
 946       set_size_and_pinuse_of_inuse_chunk(m, p, nb);
 947       return chunk2mem(p);
 948     }
 949   }
 950
 951   return NULL;
 952 }
 953
 954 /* -----------------------  system deallocation -------------------------- */
 955
 956 /* Unmap and unlink any mmapped segments that don't contain used chunks */
 957 static size_t release_unused_segments(mstate m)
 958 {
 959   size_t released = 0;
 960   size_t nsegs = 0;
 961   msegmentptr pred = &m->seg;
 962   msegmentptr sp = pred->next;
 963   while (sp != 0) {
 964     char *base = sp->base;
 965     size_t size = sp->size;
 966     msegmentptr next = sp->next;
 967     nsegs++;
 968     {
 969       mchunkptr p = align_as_chunk(base);
 970       size_t psize = chunksize(p);
 971       /* Can unmap if first chunk holds entire segment and not pinned */
 972       if (!cinuse(p) && (char *)p + psize >= base + size - TOP_FOOT_SIZE) {
 973         tchunkptr tp = (tchunkptr)p;
 974         if (p == m->dv) {
 975           m->dv = 0;
 976           m->dvsize = 0;
 977         } else {
 978           unlink_large_chunk(m, tp);
 979         }
 980         if (CALL_MUNMAP(base, size) == 0) {
 981           released += size;
 982           /* unlink obsoleted record */
 983           sp = pred;
 984           sp->next = next;
 985         } else { /* back out if cannot unmap */
 986           insert_large_chunk(m, tp, psize);
 987         }
 988       }
 989     }
 990     pred = sp;
 991     sp = next;
 992   }
 993   /* Reset check counter */
 994   m->release_checks = nsegs > MAX_RELEASE_CHECK_RATE ?
 995                       nsegs : MAX_RELEASE_CHECK_RATE;
 996   return released;
 997 }
 998
 999 static int alloc_trim(mstate m, size_t pad)
1000 {
1001   size_t released = 0;
1002   if (pad < MAX_REQUEST && is_initialized(m)) {
1003     pad += TOP_FOOT_SIZE; /* ensure enough room for segment overhead */
1004
1005     if (m->topsize > pad) {
1006       /* Shrink top space in granularity-size units, keeping at least one */
1007       size_t unit = DEFAULT_GRANULARITY;
1008       size_t extra = ((m->topsize - pad + (unit - SIZE_T_ONE)) / unit -
1009                       SIZE_T_ONE) * unit;
1010       msegmentptr sp = segment_holding(m, (char *)m->top);
1011
1012       if (sp->size >= extra &&
1013           !has_segment_link(m, sp)) { /* can't shrink if pinned */
1014         size_t newsize = sp->size - extra;
1015         /* Prefer mremap, fall back to munmap */
1016         if ((CALL_MREMAP(sp->base, sp->size, newsize, CALL_MREMAP_NOMOVE) != MFAIL) ||
1017             (CALL_MUNMAP(sp->base + newsize, extra) == 0)) {
1018           released = extra;
1019         }
1020       }
1021
1022       if (released != 0) {
1023         sp->size -= released;
1024         init_top(m, m->top, m->topsize - released);
1025       }
1026     }
1027
1028     /* Unmap any unused mmapped segments */
1029     released += release_unused_segments(m);
1030
1031     /* On failure, disable autotrim to avoid repeated failed future calls */
1032     if (released == 0 && m->topsize > m->trim_check)
1033       m->trim_check = MAX_SIZE_T;
1034   }
1035
1036   return (released != 0)? 1 : 0;
1037 }
1038
1039 /* ---------------------------- malloc support --------------------------- */
1040
1041 /* allocate a large request from the best fitting chunk in a treebin */
1042 static void *tmalloc_large(mstate m, size_t nb)
1043 {
1044   tchunkptr v = 0;
1045   size_t rsize = ~nb+1; /* Unsigned negation */
1046   tchunkptr t;
1047   bindex_t idx;
1048   compute_tree_index(nb, idx);
1049
1050   if ((t = *treebin_at(m, idx)) != 0) {
1051     /* Traverse tree for this bin looking for node with size == nb */
1052     size_t sizebits = nb << leftshift_for_tree_index(idx);
1053     tchunkptr rst = 0;  /* The deepest untaken right subtree */
1054     for (;;) {
1055       tchunkptr rt;
1056       size_t trem = chunksize(t) - nb;
1057       if (trem < rsize) {
1058         v = t;
1059         if ((rsize = trem) == 0)
1060           break;
1061       }
1062       rt = t->child[1];
1063       t = t->child[(sizebits >> (SIZE_T_BITSIZE-SIZE_T_ONE)) & 1];
1064       if (rt != 0 && rt != t)
1065         rst = rt;
1066       if (t == 0) {
1067         t = rst; /* set t to least subtree holding sizes > nb */
1068         break;
1069       }
1070       sizebits <<= 1;
1071     }
1072   }
1073
1074   if (t == 0 && v == 0) { /* set t to root of next non-empty treebin */
1075     binmap_t leftbits = left_bits(idx2bit(idx)) & m->treemap;
1076     if (leftbits != 0)
1077       t = *treebin_at(m, lj_ffs(leftbits));
1078   }
1079
1080   while (t != 0) { /* find smallest of tree or subtree */
1081     size_t trem = chunksize(t) - nb;
1082     if (trem < rsize) {
1083       rsize = trem;
1084       v = t;
1085     }
1086     t = leftmost_child(t);
1087   }
1088
1089   /*  If dv is a better fit, return NULL so malloc will use it */
1090   if (v != 0 && rsize < (size_t)(m->dvsize - nb)) {
1091     mchunkptr r = chunk_plus_offset(v, nb);
1092     unlink_large_chunk(m, v);
1093     if (rsize < MIN_CHUNK_SIZE) {
1094       set_inuse_and_pinuse(m, v, (rsize + nb));
1095     } else {
1096       set_size_and_pinuse_of_inuse_chunk(m, v, nb);
1097       set_size_and_pinuse_of_free_chunk(r, rsize);
1098       insert_chunk(m, r, rsize);
1099     }
1100     return chunk2mem(v);
1101   }
1102   return NULL;
1103 }
1104
1105 /* allocate a small request from the best fitting chunk in a treebin */
1106 static void *tmalloc_small(mstate m, size_t nb)
1107 {
1108   tchunkptr t, v;
1109   mchunkptr r;
1110   size_t rsize;
1111   bindex_t i = lj_ffs(m->treemap);
1112
1113   v = t = *treebin_at(m, i);
1114   rsize = chunksize(t) - nb;
1115
1116   while ((t = leftmost_child(t)) != 0) {
1117     size_t trem = chunksize(t) - nb;
1118     if (trem < rsize) {
1119       rsize = trem;
1120       v = t;
1121     }
1122   }
1123
1124   r = chunk_plus_offset(v, nb);
1125   unlink_large_chunk(m, v);
1126   if (rsize < MIN_CHUNK_SIZE) {
1127     set_inuse_and_pinuse(m, v, (rsize + nb));
1128   } else {
1129     set_size_and_pinuse_of_inuse_chunk(m, v, nb);
1130     set_size_and_pinuse_of_free_chunk(r, rsize);
1131     replace_dv(m, r, rsize);
1132   }
1133   return chunk2mem(v);
1134 }
1135
1136 /* ----------------------------------------------------------------------- */
1137
1138 void *lj_alloc_create(void)
1139 {
1140   size_t tsize = DEFAULT_GRANULARITY;
1141   char *tbase;
1142   INIT_MMAP();
1143   tbase = (char *)(CALL_MMAP(tsize));
1144   if (tbase != CMFAIL) {
1145     size_t msize = pad_request(sizeof(struct malloc_state));
1146     mchunkptr mn;
1147     mchunkptr msp = align_as_chunk(tbase);
1148     mstate m = (mstate)(chunk2mem(msp));
1149     memset(m, 0, msize);
1150     msp->head = (msize|PINUSE_BIT|CINUSE_BIT);
1151     m->seg.base = tbase;
1152     m->seg.size = tsize;
1153     m->release_checks = MAX_RELEASE_CHECK_RATE;
1154     init_bins(m);
1155     mn = next_chunk(mem2chunk(m));
1156     init_top(m, mn, (size_t)((tbase + tsize) - (char *)mn) - TOP_FOOT_SIZE);
1157     return m;
1158   }
1159   return NULL;
1160 }
1161
1162 void lj_alloc_destroy(void *msp)
1163 {
1164   mstate ms = (mstate)msp;
1165   msegmentptr sp = &ms->seg;
1166   while (sp != 0) {
1167     char *base = sp->base;
1168     size_t size = sp->size;
1169     sp = sp->next;
1170     CALL_MUNMAP(base, size);
1171   }
1172 }
1173
1174 static LJ_NOINLINE void *lj_alloc_malloc(void *msp, size_t nsize)
1175 {
1176   mstate ms = (mstate)msp;
1177   void *mem;
1178   size_t nb;
1179   if (nsize <= MAX_SMALL_REQUEST) {
1180     bindex_t idx;
1181     binmap_t smallbits;
1182     nb = (nsize < MIN_REQUEST)? MIN_CHUNK_SIZE : pad_request(nsize);
1183     idx = small_index(nb);
1184     smallbits = ms->smallmap >> idx;
1185
1186     if ((smallbits & 0x3U) != 0) { /* Remainderless fit to a smallbin. */
1187       mchunkptr b, p;
1188       idx += ~smallbits & 1;       /* Uses next bin if idx empty */
1189       b = smallbin_at(ms, idx);
1190       p = b->fd;
1191       unlink_first_small_chunk(ms, b, p, idx);
1192       set_inuse_and_pinuse(ms, p, small_index2size(idx));
1193       mem = chunk2mem(p);
1194       return mem;
1195     } else if (nb > ms->dvsize) {
1196       if (smallbits != 0) { /* Use chunk in next nonempty smallbin */
1197         mchunkptr b, p, r;
1198         size_t rsize;
1199         binmap_t leftbits = (smallbits << idx) & left_bits(idx2bit(idx));
1200         bindex_t i = lj_ffs(leftbits);
1201         b = smallbin_at(ms, i);
1202         p = b->fd;
1203         unlink_first_small_chunk(ms, b, p, i);
1204         rsize = small_index2size(i) - nb;
1205         /* Fit here cannot be remainderless if 4byte sizes */
1206         if (SIZE_T_SIZE != 4 && rsize < MIN_CHUNK_SIZE) {
1207           set_inuse_and_pinuse(ms, p, small_index2size(i));
1208         } else {
1209           set_size_and_pinuse_of_inuse_chunk(ms, p, nb);
1210           r = chunk_plus_offset(p, nb);
1211           set_size_and_pinuse_of_free_chunk(r, rsize);
1212           replace_dv(ms, r, rsize);
1213         }
1214         mem = chunk2mem(p);
1215         return mem;
1216       } else if (ms->treemap != 0 && (mem = tmalloc_small(ms, nb)) != 0) {
1217         return mem;
1218       }
1219     }
1220   } else if (nsize >= MAX_REQUEST) {
1221     nb = MAX_SIZE_T; /* Too big to allocate. Force failure (in sys alloc) */
1222   } else {
1223     nb = pad_request(nsize);
1224     if (ms->treemap != 0 && (mem = tmalloc_large(ms, nb)) != 0) {
1225       return mem;
1226     }
1227   }
1228
1229   if (nb <= ms->dvsize) {
1230     size_t rsize = ms->dvsize - nb;
1231     mchunkptr p = ms->dv;
1232     if (rsize >= MIN_CHUNK_SIZE) { /* split dv */
1233       mchunkptr r = ms->dv = chunk_plus_offset(p, nb);
1234       ms->dvsize = rsize;
1235       set_size_and_pinuse_of_free_chunk(r, rsize);
1236       set_size_and_pinuse_of_inuse_chunk(ms, p, nb);
1237     } else { /* exhaust dv */
1238       size_t dvs = ms->dvsize;
1239       ms->dvsize = 0;
1240       ms->dv = 0;
1241       set_inuse_and_pinuse(ms, p, dvs);
1242     }
1243     mem = chunk2mem(p);
1244     return mem;
1245   } else if (nb < ms->topsize) { /* Split top */
1246     size_t rsize = ms->topsize -= nb;
1247     mchunkptr p = ms->top;
1248     mchunkptr r = ms->top = chunk_plus_offset(p, nb);
1249     r->head = rsize | PINUSE_BIT;
1250     set_size_and_pinuse_of_inuse_chunk(ms, p, nb);
1251     mem = chunk2mem(p);
1252     return mem;
1253   }
1254   return alloc_sys(ms, nb);
1255 }
1256
1257 static LJ_NOINLINE void *lj_alloc_free(void *msp, void *ptr)
1258 {
1259   if (ptr != 0) {
1260     mchunkptr p = mem2chunk(ptr);
1261     mstate fm = (mstate)msp;
1262     size_t psize = chunksize(p);
1263     mchunkptr next = chunk_plus_offset(p, psize);
1264     if (!pinuse(p)) {
1265       size_t prevsize = p->prev_foot;
1266       if ((prevsize & IS_DIRECT_BIT) != 0) {
1267         prevsize &= ~IS_DIRECT_BIT;
1268         psize += prevsize + DIRECT_FOOT_PAD;
1269         CALL_MUNMAP((char *)p - prevsize, psize);
1270         return NULL;
1271       } else {
1272         mchunkptr prev = chunk_minus_offset(p, prevsize);
1273         psize += prevsize;
1274         p = prev;
1275         /* consolidate backward */
1276         if (p != fm->dv) {
1277           unlink_chunk(fm, p, prevsize);
1278         } else if ((next->head & INUSE_BITS) == INUSE_BITS) {
1279           fm->dvsize = psize;
1280           set_free_with_pinuse(p, psize, next);
1281           return NULL;
1282         }
1283       }
1284     }
1285     if (!cinuse(next)) {  /* consolidate forward */
1286       if (next == fm->top) {
1287         size_t tsize = fm->topsize += psize;
1288         fm->top = p;
1289         p->head = tsize | PINUSE_BIT;
1290         if (p == fm->dv) {
1291           fm->dv = 0;
1292           fm->dvsize = 0;
1293         }
1294         if (tsize > fm->trim_check)
1295           alloc_trim(fm, 0);
1296         return NULL;
1297       } else if (next == fm->dv) {
1298         size_t dsize = fm->dvsize += psize;
1299         fm->dv = p;
1300         set_size_and_pinuse_of_free_chunk(p, dsize);
1301         return NULL;
1302       } else {
1303         size_t nsize = chunksize(next);
1304         psize += nsize;
1305         unlink_chunk(fm, next, nsize);
1306         set_size_and_pinuse_of_free_chunk(p, psize);
1307         if (p == fm->dv) {
1308           fm->dvsize = psize;
1309           return NULL;
1310         }
1311       }
1312     } else {
1313       set_free_with_pinuse(p, psize, next);
1314     }
1315
1316     if (is_small(psize)) {
1317       insert_small_chunk(fm, p, psize);
1318     } else {
1319       tchunkptr tp = (tchunkptr)p;
1320       insert_large_chunk(fm, tp, psize);
1321       if (--fm->release_checks == 0)
1322         release_unused_segments(fm);
1323     }
1324   }
1325   return NULL;
1326 }
1327
1328 static LJ_NOINLINE void *lj_alloc_realloc(void *msp, void *ptr, size_t nsize)
1329 {
1330   if (nsize >= MAX_REQUEST) {
1331     return NULL;
1332   } else {
1333     mstate m = (mstate)msp;
1334     mchunkptr oldp = mem2chunk(ptr);
1335     size_t oldsize = chunksize(oldp);
1336     mchunkptr next = chunk_plus_offset(oldp, oldsize);
1337     mchunkptr newp = 0;
1338     size_t nb = request2size(nsize);
1339
1340     /* Try to either shrink or extend into top. Else malloc-copy-free */
1341     if (is_direct(oldp)) {
1342       newp = direct_resize(oldp, nb);  /* this may return NULL. */
1343     } else if (oldsize >= nb) { /* already big enough */
1344       size_t rsize = oldsize - nb;
1345       newp = oldp;
1346       if (rsize >= MIN_CHUNK_SIZE) {
1347         mchunkptr rem = chunk_plus_offset(newp, nb);
1348         set_inuse(m, newp, nb);
1349         set_inuse(m, rem, rsize);
1350         lj_alloc_free(m, chunk2mem(rem));
1351       }
1352     } else if (next == m->top && oldsize + m->topsize > nb) {
1353       /* Expand into top */
1354       size_t newsize = oldsize + m->topsize;
1355       size_t newtopsize = newsize - nb;
1356       mchunkptr newtop = chunk_plus_offset(oldp, nb);
1357       set_inuse(m, oldp, nb);
1358       newtop->head = newtopsize |PINUSE_BIT;
1359       m->top = newtop;
1360       m->topsize = newtopsize;
1361       newp = oldp;
1362     }
1363
1364     if (newp != 0) {
1365       return chunk2mem(newp);
1366     } else {
1367       void *newmem = lj_alloc_malloc(m, nsize);
1368       if (newmem != 0) {
1369         size_t oc = oldsize - overhead_for(oldp);
1370         memcpy(newmem, ptr, oc < nsize ? oc : nsize);
1371         lj_alloc_free(m, ptr);
1372       }
1373       return newmem;
1374     }
1375   }
1376 }
1377
1378 void *lj_alloc_f(void *msp, void *ptr, size_t osize, size_t nsize)
1379 {
1380   (void)osize;
1381   if (nsize == 0) {
1382     return lj_alloc_free(msp, ptr);
1383   } else if (ptr == NULL) {
1384     return lj_alloc_malloc(msp, nsize);
1385   } else {
1386     return lj_alloc_realloc(msp, ptr, nsize);
1387   }
1388 }
1389
1390 #endif