sha1_file.c

   1 /*
   2  * GIT - The information manager from hell
   3  *
   4  * Copyright (C) Linus Torvalds, 2005
   5  *
   6  * This handles basic git sha1 object files - packing, unpacking,
   7  * creation etc.
   8  */
   9 #include "cache.h"
  10 #include "delta.h"
  11 #include "pack.h"
  12 #include "blob.h"
  13 #include "commit.h"
  14 #include "tag.h"
  15 #include "tree.h"
  16 #include "refs.h"
  17 #include "pack-revindex.h"
  18 #include "sha1-lookup.h"
  19
  20 #ifndef O_NOATIME
  21 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
  22 #define O_NOATIME 01000000
  23 #else
  24 #define O_NOATIME 0
  25 #endif
  26 #endif
  27
  28 #ifdef NO_C99_FORMAT
  29 #define SZ_FMT "lu"
  30 static unsigned long sz_fmt(size_t s) { return (unsigned long)s; }
  31 #else
  32 #define SZ_FMT "zu"
  33 static size_t sz_fmt(size_t s) { return s; }
  34 #endif
  35
  36 const unsigned char null_sha1[20];
  37
  38 static inline int offset_1st_component(const char *path)
  39 {
  40         if (has_dos_drive_prefix(path))
  41                 return 2 + (path[2] == '/');
  42         return *path == '/';
  43 }
  44
  45 int mkdir_in_gitdir(const char *path)
  46 {
  47         if (mkdir(path, 0777)) {
  48                 int saved_errno = errno;
  49                 struct stat st;
  50                 struct strbuf sb = STRBUF_INIT;
  51
  52                 if (errno != EEXIST)
  53                         return -1;
  54                 /*
  55                  * Are we looking at a path in a symlinked worktree
  56                  * whose original repository does not yet have it?
  57                  * e.g. .git/rr-cache pointing at its original
  58                  * repository in which the user hasn't performed any
  59                  * conflict resolution yet?
  60                  */
  61                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
  62                     strbuf_readlink(&sb, path, st.st_size) ||
  63                     !is_absolute_path(sb.buf) ||
  64                     mkdir(sb.buf, 0777)) {
  65                         strbuf_release(&sb);
  66                         errno = saved_errno;
  67                         return -1;
  68                 }
  69                 strbuf_release(&sb);
  70         }
  71         return adjust_shared_perm(path);
  72 }
  73
  74 int safe_create_leading_directories(char *path)
  75 {
  76         char *pos = path + offset_1st_component(path);
  77         struct stat st;
  78
  79         while (pos) {
  80                 pos = strchr(pos, '/');
  81                 if (!pos)
  82                         break;
  83                 while (*++pos == '/')
  84                         ;
  85                 if (!*pos)
  86                         break;
  87                 *--pos = '\0';
  88                 if (!stat(path, &st)) {
  89                         /* path exists */
  90                         if (!S_ISDIR(st.st_mode)) {
  91                                 *pos = '/';
  92                                 return -3;
  93                         }
  94                 }
  95                 else if (mkdir(path, 0777)) {
  96                         *pos = '/';
  97                         return -1;
  98                 }
  99                 else if (adjust_shared_perm(path)) {
 100                         *pos = '/';
 101                         return -2;
 102                 }
 103                 *pos++ = '/';
 104         }
 105         return 0;
 106 }
 107
 108 int safe_create_leading_directories_const(const char *path)
 109 {
 110         /* path points to cache entries, so xstrdup before messing with it */
 111         char *buf = xstrdup(path);
 112         int result = safe_create_leading_directories(buf);
 113         free(buf);
 114         return result;
 115 }
 116
 117 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
 118 {
 119         int i;
 120         for (i = 0; i < 20; i++) {
 121                 static char hex[] = "0123456789abcdef";
 122                 unsigned int val = sha1[i];
 123                 char *pos = pathbuf + i*2 + (i > 0);
 124                 *pos++ = hex[val >> 4];
 125                 *pos = hex[val & 0xf];
 126         }
 127 }
 128
 129 /*
 130  * NOTE! This returns a statically allocated buffer, so you have to be
 131  * careful about using it. Do an "xstrdup()" if you need to save the
 132  * filename.
 133  *
 134  * Also note that this returns the location for creating.  Reading
 135  * SHA1 file can happen from any alternate directory listed in the
 136  * DB_ENVIRONMENT environment variable if it is not found in
 137  * the primary object database.
 138  */
 139 char *sha1_file_name(const unsigned char *sha1)
 140 {
 141         static char *name, *base;
 142
 143         if (!base) {
 144                 const char *sha1_file_directory = get_object_directory();
 145                 int len = strlen(sha1_file_directory);
 146                 base = xmalloc(len + 60);
 147                 memcpy(base, sha1_file_directory, len);
 148                 memset(base+len, 0, 60);
 149                 base[len] = '/';
 150                 base[len+3] = '/';
 151                 name = base + len + 1;
 152         }
 153         fill_sha1_path(name, sha1);
 154         return base;
 155 }
 156
 157 static char *sha1_get_pack_name(const unsigned char *sha1,
 158                                 char **name, char **base, const char *which)
 159 {
 160         static const char hex[] = "0123456789abcdef";
 161         char *buf;
 162         int i;
 163
 164         if (!*base) {
 165                 const char *sha1_file_directory = get_object_directory();
 166                 int len = strlen(sha1_file_directory);
 167                 *base = xmalloc(len + 60);
 168                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 169                         sha1_file_directory, which);
 170                 *name = *base + len + 11;
 171         }
 172
 173         buf = *name;
 174
 175         for (i = 0; i < 20; i++) {
 176                 unsigned int val = *sha1++;
 177                 *buf++ = hex[val >> 4];
 178                 *buf++ = hex[val & 0xf];
 179         }
 180
 181         return *base;
 182 }
 183
 184 char *sha1_pack_name(const unsigned char *sha1)
 185 {
 186         static char *name, *base;
 187
 188         return sha1_get_pack_name(sha1, &name, &base, "pack");
 189 }
 190
 191 char *sha1_pack_index_name(const unsigned char *sha1)
 192 {
 193         static char *name, *base;
 194
 195         return sha1_get_pack_name(sha1, &name, &base, "idx");
 196 }
 197
 198 struct alternate_object_database *alt_odb_list;
 199 static struct alternate_object_database **alt_odb_tail;
 200
 201 static void read_info_alternates(const char * alternates, int depth);
 202
 203 /*
 204  * Prepare alternate object database registry.
 205  *
 206  * The variable alt_odb_list points at the list of struct
 207  * alternate_object_database.  The elements on this list come from
 208  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 209  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 210  * whose contents is similar to that environment variable but can be
 211  * LF separated.  Its base points at a statically allocated buffer that
 212  * contains "/the/directory/corresponding/to/.git/objects/...", while
 213  * its name points just after the slash at the end of ".git/objects/"
 214  * in the example above, and has enough space to hold 40-byte hex
 215  * SHA1, an extra slash for the first level indirection, and the
 216  * terminating NUL.
 217  */
 218 static int link_alt_odb_entry(const char * entry, int len, const char * relative_base, int depth)
 219 {
 220         const char *objdir = get_object_directory();
 221         struct alternate_object_database *ent;
 222         struct alternate_object_database *alt;
 223         /* 43 = 40-byte + 2 '/' + terminating NUL */
 224         int pfxlen = len;
 225         int entlen = pfxlen + 43;
 226         int base_len = -1;
 227
 228         if (!is_absolute_path(entry) && relative_base) {
 229                 /* Relative alt-odb */
 230                 if (base_len < 0)
 231                         base_len = strlen(relative_base) + 1;
 232                 entlen += base_len;
 233                 pfxlen += base_len;
 234         }
 235         ent = xmalloc(sizeof(*ent) + entlen);
 236
 237         if (!is_absolute_path(entry) && relative_base) {
 238                 memcpy(ent->base, relative_base, base_len - 1);
 239                 ent->base[base_len - 1] = '/';
 240                 memcpy(ent->base + base_len, entry, len);
 241         }
 242         else
 243                 memcpy(ent->base, entry, pfxlen);
 244
 245         ent->name = ent->base + pfxlen + 1;
 246         ent->base[pfxlen + 3] = '/';
 247         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 248
 249         /* Detect cases where alternate disappeared */
 250         if (!is_directory(ent->base)) {
 251                 error("object directory %s does not exist; "
 252                       "check .git/objects/info/alternates.",
 253                       ent->base);
 254                 free(ent);
 255                 return -1;
 256         }
 257
 258         /* Prevent the common mistake of listing the same
 259          * thing twice, or object directory itself.
 260          */
 261         for (alt = alt_odb_list; alt; alt = alt->next) {
 262                 if (!memcmp(ent->base, alt->base, pfxlen)) {
 263                         free(ent);
 264                         return -1;
 265                 }
 266         }
 267         if (!memcmp(ent->base, objdir, pfxlen)) {
 268                 free(ent);
 269                 return -1;
 270         }
 271
 272         /* add the alternate entry */
 273         *alt_odb_tail = ent;
 274         alt_odb_tail = &(ent->next);
 275         ent->next = NULL;
 276
 277         /* recursively add alternates */
 278         read_info_alternates(ent->base, depth + 1);
 279
 280         ent->base[pfxlen] = '/';
 281
 282         return 0;
 283 }
 284
 285 static void link_alt_odb_entries(const char *alt, const char *ep, int sep,
 286                                  const char *relative_base, int depth)
 287 {
 288         const char *cp, *last;
 289
 290         if (depth > 5) {
 291                 error("%s: ignoring alternate object stores, nesting too deep.",
 292                                 relative_base);
 293                 return;
 294         }
 295
 296         last = alt;
 297         while (last < ep) {
 298                 cp = last;
 299                 if (cp < ep && *cp == '#') {
 300                         while (cp < ep && *cp != sep)
 301                                 cp++;
 302                         last = cp + 1;
 303                         continue;
 304                 }
 305                 while (cp < ep && *cp != sep)
 306                         cp++;
 307                 if (last != cp) {
 308                         if (!is_absolute_path(last) && depth) {
 309                                 error("%s: ignoring relative alternate object store %s",
 310                                                 relative_base, last);
 311                         } else {
 312                                 link_alt_odb_entry(last, cp - last,
 313                                                 relative_base, depth);
 314                         }
 315                 }
 316                 while (cp < ep && *cp == sep)
 317                         cp++;
 318                 last = cp;
 319         }
 320 }
 321
 322 static void read_info_alternates(const char * relative_base, int depth)
 323 {
 324         char *map;
 325         size_t mapsz;
 326         struct stat st;
 327         const char alt_file_name[] = "info/alternates";
 328         /* Given that relative_base is no longer than PATH_MAX,
 329            ensure that "path" has enough space to append "/", the
 330            file name, "info/alternates", and a trailing NUL.  */
 331         char path[PATH_MAX + 1 + sizeof alt_file_name];
 332         int fd;
 333
 334         sprintf(path, "%s/%s", relative_base, alt_file_name);
 335         fd = open(path, O_RDONLY);
 336         if (fd < 0)
 337                 return;
 338         if (fstat(fd, &st) || (st.st_size == 0)) {
 339                 close(fd);
 340                 return;
 341         }
 342         mapsz = xsize_t(st.st_size);
 343         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 344         close(fd);
 345
 346         link_alt_odb_entries(map, map + mapsz, '\n', relative_base, depth);
 347
 348         munmap(map, mapsz);
 349 }
 350
 351 void add_to_alternates_file(const char *reference)
 352 {
 353         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 354         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
 355         char *alt = mkpath("%s/objects\n", reference);
 356         write_or_die(fd, alt, strlen(alt));
 357         if (commit_lock_file(lock))
 358                 die("could not close alternates file");
 359         if (alt_odb_tail)
 360                 link_alt_odb_entries(alt, alt + strlen(alt), '\n', NULL, 0);
 361 }
 362
 363 void foreach_alt_odb(alt_odb_fn fn, void *cb)
 364 {
 365         struct alternate_object_database *ent;
 366
 367         prepare_alt_odb();
 368         for (ent = alt_odb_list; ent; ent = ent->next)
 369                 if (fn(ent, cb))
 370                         return;
 371 }
 372
 373 void prepare_alt_odb(void)
 374 {
 375         const char *alt;
 376
 377         if (alt_odb_tail)
 378                 return;
 379
 380         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 381         if (!alt) alt = "";
 382
 383         alt_odb_tail = &alt_odb_list;
 384         link_alt_odb_entries(alt, alt + strlen(alt), PATH_SEP, NULL, 0);
 385
 386         read_info_alternates(get_object_directory(), 0);
 387 }
 388
 389 static int has_loose_object_local(const unsigned char *sha1)
 390 {
 391         char *name = sha1_file_name(sha1);
 392         return !access(name, F_OK);
 393 }
 394
 395 int has_loose_object_nonlocal(const unsigned char *sha1)
 396 {
 397         struct alternate_object_database *alt;
 398         prepare_alt_odb();
 399         for (alt = alt_odb_list; alt; alt = alt->next) {
 400                 fill_sha1_path(alt->name, sha1);
 401                 if (!access(alt->base, F_OK))
 402                         return 1;
 403         }
 404         return 0;
 405 }
 406
 407 static int has_loose_object(const unsigned char *sha1)
 408 {
 409         return has_loose_object_local(sha1) ||
 410                has_loose_object_nonlocal(sha1);
 411 }
 412
 413 static unsigned int pack_used_ctr;
 414 static unsigned int pack_mmap_calls;
 415 static unsigned int peak_pack_open_windows;
 416 static unsigned int pack_open_windows;
 417 static size_t peak_pack_mapped;
 418 static size_t pack_mapped;
 419 struct packed_git *packed_git;
 420
 421 void pack_report(void)
 422 {
 423         fprintf(stderr,
 424                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 425                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 426                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 427                 sz_fmt(getpagesize()),
 428                 sz_fmt(packed_git_window_size),
 429                 sz_fmt(packed_git_limit));
 430         fprintf(stderr,
 431                 "pack_report: pack_used_ctr            = %10u\n"
 432                 "pack_report: pack_mmap_calls          = %10u\n"
 433                 "pack_report: pack_open_windows        = %10u / %10u\n"
 434                 "pack_report: pack_mapped              = "
 435                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 436                 pack_used_ctr,
 437                 pack_mmap_calls,
 438                 pack_open_windows, peak_pack_open_windows,
 439                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 440 }
 441
 442 static int check_packed_git_idx(const char *path,  struct packed_git *p)
 443 {
 444         void *idx_map;
 445         struct pack_idx_header *hdr;
 446         size_t idx_size;
 447         uint32_t version, nr, i, *index;
 448         int fd = open(path, O_RDONLY);
 449         struct stat st;
 450
 451         if (fd < 0)
 452                 return -1;
 453         if (fstat(fd, &st)) {
 454                 close(fd);
 455                 return -1;
 456         }
 457         idx_size = xsize_t(st.st_size);
 458         if (idx_size < 4 * 256 + 20 + 20) {
 459                 close(fd);
 460                 return error("index file %s is too small", path);
 461         }
 462         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 463         close(fd);
 464
 465         hdr = idx_map;
 466         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 467                 version = ntohl(hdr->idx_version);
 468                 if (version < 2 || version > 2) {
 469                         munmap(idx_map, idx_size);
 470                         return error("index file %s is version %"PRIu32
 471                                      " and is not supported by this binary"
 472                                      " (try upgrading GIT to a newer version)",
 473                                      path, version);
 474                 }
 475         } else
 476                 version = 1;
 477
 478         nr = 0;
 479         index = idx_map;
 480         if (version > 1)
 481                 index += 2;  /* skip index header */
 482         for (i = 0; i < 256; i++) {
 483                 uint32_t n = ntohl(index[i]);
 484                 if (n < nr) {
 485                         munmap(idx_map, idx_size);
 486                         return error("non-monotonic index %s", path);
 487                 }
 488                 nr = n;
 489         }
 490
 491         if (version == 1) {
 492                 /*
 493                  * Total size:
 494                  *  - 256 index entries 4 bytes each
 495                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 496                  *  - 20-byte SHA1 of the packfile
 497                  *  - 20-byte SHA1 file checksum
 498                  */
 499                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 500                         munmap(idx_map, idx_size);
 501                         return error("wrong index v1 file size in %s", path);
 502                 }
 503         } else if (version == 2) {
 504                 /*
 505                  * Minimum size:
 506                  *  - 8 bytes of header
 507                  *  - 256 index entries 4 bytes each
 508                  *  - 20-byte sha1 entry * nr
 509                  *  - 4-byte crc entry * nr
 510                  *  - 4-byte offset entry * nr
 511                  *  - 20-byte SHA1 of the packfile
 512                  *  - 20-byte SHA1 file checksum
 513                  * And after the 4-byte offset table might be a
 514                  * variable sized table containing 8-byte entries
 515                  * for offsets larger than 2^31.
 516                  */
 517                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 518                 unsigned long max_size = min_size;
 519                 if (nr)
 520                         max_size += (nr - 1)*8;
 521                 if (idx_size < min_size || idx_size > max_size) {
 522                         munmap(idx_map, idx_size);
 523                         return error("wrong index v2 file size in %s", path);
 524                 }
 525                 if (idx_size != min_size &&
 526                     /*
 527                      * make sure we can deal with large pack offsets.
 528                      * 31-bit signed offset won't be enough, neither
 529                      * 32-bit unsigned one will be.
 530                      */
 531                     (sizeof(off_t) <= 4)) {
 532                         munmap(idx_map, idx_size);
 533                         return error("pack too large for current definition of off_t in %s", path);
 534                 }
 535         }
 536
 537         p->index_version = version;
 538         p->index_data = idx_map;
 539         p->index_size = idx_size;
 540         p->num_objects = nr;
 541         return 0;
 542 }
 543
 544 int open_pack_index(struct packed_git *p)
 545 {
 546         char *idx_name;
 547         int ret;
 548
 549         if (p->index_data)
 550                 return 0;
 551
 552         idx_name = xstrdup(p->pack_name);
 553         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 554         ret = check_packed_git_idx(idx_name, p);
 555         free(idx_name);
 556         return ret;
 557 }
 558
 559 static void scan_windows(struct packed_git *p,
 560         struct packed_git **lru_p,
 561         struct pack_window **lru_w,
 562         struct pack_window **lru_l)
 563 {
 564         struct pack_window *w, *w_l;
 565
 566         for (w_l = NULL, w = p->windows; w; w = w->next) {
 567                 if (!w->inuse_cnt) {
 568                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 569                                 *lru_p = p;
 570                                 *lru_w = w;
 571                                 *lru_l = w_l;
 572                         }
 573                 }
 574                 w_l = w;
 575         }
 576 }
 577
 578 static int unuse_one_window(struct packed_git *current, int keep_fd)
 579 {
 580         struct packed_git *p, *lru_p = NULL;
 581         struct pack_window *lru_w = NULL, *lru_l = NULL;
 582
 583         if (current)
 584                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 585         for (p = packed_git; p; p = p->next)
 586                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 587         if (lru_p) {
 588                 munmap(lru_w->base, lru_w->len);
 589                 pack_mapped -= lru_w->len;
 590                 if (lru_l)
 591                         lru_l->next = lru_w->next;
 592                 else {
 593                         lru_p->windows = lru_w->next;
 594                         if (!lru_p->windows && lru_p->pack_fd != keep_fd) {
 595                                 close(lru_p->pack_fd);
 596                                 lru_p->pack_fd = -1;
 597                         }
 598                 }
 599                 free(lru_w);
 600                 pack_open_windows--;
 601                 return 1;
 602         }
 603         return 0;
 604 }
 605
 606 void release_pack_memory(size_t need, int fd)
 607 {
 608         size_t cur = pack_mapped;
 609         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
 610                 ; /* nothing */
 611 }
 612
 613 void close_pack_windows(struct packed_git *p)
 614 {
 615         while (p->windows) {
 616                 struct pack_window *w = p->windows;
 617
 618                 if (w->inuse_cnt)
 619                         die("pack '%s' still has open windows to it",
 620                             p->pack_name);
 621                 munmap(w->base, w->len);
 622                 pack_mapped -= w->len;
 623                 pack_open_windows--;
 624                 p->windows = w->next;
 625                 free(w);
 626         }
 627 }
 628
 629 void unuse_pack(struct pack_window **w_cursor)
 630 {
 631         struct pack_window *w = *w_cursor;
 632         if (w) {
 633                 w->inuse_cnt--;
 634                 *w_cursor = NULL;
 635         }
 636 }
 637
 638 /*
 639  * This is used by git-repack in case a newly created pack happens to
 640  * contain the same set of objects as an existing one.  In that case
 641  * the resulting file might be different even if its name would be the
 642  * same.  It is best to close any reference to the old pack before it is
 643  * replaced on disk.  Of course no index pointers nor windows for given pack
 644  * must subsist at this point.  If ever objects from this pack are requested
 645  * again, the new version of the pack will be reinitialized through
 646  * reprepare_packed_git().
 647  */
 648 void free_pack_by_name(const char *pack_name)
 649 {
 650         struct packed_git *p, **pp = &packed_git;
 651
 652         while (*pp) {
 653                 p = *pp;
 654                 if (strcmp(pack_name, p->pack_name) == 0) {
 655                         clear_delta_base_cache();
 656                         close_pack_windows(p);
 657                         if (p->pack_fd != -1)
 658                                 close(p->pack_fd);
 659                         if (p->index_data)
 660                                 munmap((void *)p->index_data, p->index_size);
 661                         free(p->bad_object_sha1);
 662                         *pp = p->next;
 663                         free(p);
 664                         return;
 665                 }
 666                 pp = &p->next;
 667         }
 668 }
 669
 670 /*
 671  * Do not call this directly as this leaks p->pack_fd on error return;
 672  * call open_packed_git() instead.
 673  */
 674 static int open_packed_git_1(struct packed_git *p)
 675 {
 676         struct stat st;
 677         struct pack_header hdr;
 678         unsigned char sha1[20];
 679         unsigned char *idx_sha1;
 680         long fd_flag;
 681
 682         if (!p->index_data && open_pack_index(p))
 683                 return error("packfile %s index unavailable", p->pack_name);
 684
 685         p->pack_fd = open(p->pack_name, O_RDONLY);
 686         while (p->pack_fd < 0 && errno == EMFILE && unuse_one_window(p, -1))
 687                 p->pack_fd = open(p->pack_name, O_RDONLY);
 688         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 689                 return -1;
 690
 691         /* If we created the struct before we had the pack we lack size. */
 692         if (!p->pack_size) {
 693                 if (!S_ISREG(st.st_mode))
 694                         return error("packfile %s not a regular file", p->pack_name);
 695                 p->pack_size = st.st_size;
 696         } else if (p->pack_size != st.st_size)
 697                 return error("packfile %s size changed", p->pack_name);
 698
 699         /* We leave these file descriptors open with sliding mmap;
 700          * there is no point keeping them open across exec(), though.
 701          */
 702         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 703         if (fd_flag < 0)
 704                 return error("cannot determine file descriptor flags");
 705         fd_flag |= FD_CLOEXEC;
 706         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 707                 return error("cannot set FD_CLOEXEC");
 708
 709         /* Verify we recognize this pack file format. */
 710         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 711                 return error("file %s is far too short to be a packfile", p->pack_name);
 712         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 713                 return error("file %s is not a GIT packfile", p->pack_name);
 714         if (!pack_version_ok(hdr.hdr_version))
 715                 return error("packfile %s is version %"PRIu32" and not"
 716                         " supported (try upgrading GIT to a newer version)",
 717                         p->pack_name, ntohl(hdr.hdr_version));
 718
 719         /* Verify the pack matches its index. */
 720         if (p->num_objects != ntohl(hdr.hdr_entries))
 721                 return error("packfile %s claims to have %"PRIu32" objects"
 722                              " while index indicates %"PRIu32" objects",
 723                              p->pack_name, ntohl(hdr.hdr_entries),
 724                              p->num_objects);
 725         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
 726                 return error("end of packfile %s is unavailable", p->pack_name);
 727         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
 728                 return error("packfile %s signature is unavailable", p->pack_name);
 729         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
 730         if (hashcmp(sha1, idx_sha1))
 731                 return error("packfile %s does not match index", p->pack_name);
 732         return 0;
 733 }
 734
 735 static int open_packed_git(struct packed_git *p)
 736 {
 737         if (!open_packed_git_1(p))
 738                 return 0;
 739         if (p->pack_fd != -1) {
 740                 close(p->pack_fd);
 741                 p->pack_fd = -1;
 742         }
 743         return -1;
 744 }
 745
 746 static int in_window(struct pack_window *win, off_t offset)
 747 {
 748         /* We must promise at least 20 bytes (one hash) after the
 749          * offset is available from this window, otherwise the offset
 750          * is not actually in this window and a different window (which
 751          * has that one hash excess) must be used.  This is to support
 752          * the object header and delta base parsing routines below.
 753          */
 754         off_t win_off = win->offset;
 755         return win_off <= offset
 756                 && (offset + 20) <= (win_off + win->len);
 757 }
 758
 759 unsigned char *use_pack(struct packed_git *p,
 760                 struct pack_window **w_cursor,
 761                 off_t offset,
 762                 unsigned int *left)
 763 {
 764         struct pack_window *win = *w_cursor;
 765
 766         if (p->pack_fd == -1 && open_packed_git(p))
 767                 die("packfile %s cannot be accessed", p->pack_name);
 768
 769         /* Since packfiles end in a hash of their content and it's
 770          * pointless to ask for an offset into the middle of that
 771          * hash, and the in_window function above wouldn't match
 772          * don't allow an offset too close to the end of the file.
 773          */
 774         if (offset > (p->pack_size - 20))
 775                 die("offset beyond end of packfile (truncated pack?)");
 776
 777         if (!win || !in_window(win, offset)) {
 778                 if (win)
 779                         win->inuse_cnt--;
 780                 for (win = p->windows; win; win = win->next) {
 781                         if (in_window(win, offset))
 782                                 break;
 783                 }
 784                 if (!win) {
 785                         size_t window_align = packed_git_window_size / 2;
 786                         off_t len;
 787                         win = xcalloc(1, sizeof(*win));
 788                         win->offset = (offset / window_align) * window_align;
 789                         len = p->pack_size - win->offset;
 790                         if (len > packed_git_window_size)
 791                                 len = packed_git_window_size;
 792                         win->len = (size_t)len;
 793                         pack_mapped += win->len;
 794                         while (packed_git_limit < pack_mapped
 795                                 && unuse_one_window(p, p->pack_fd))
 796                                 ; /* nothing */
 797                         win->base = xmmap(NULL, win->len,
 798                                 PROT_READ, MAP_PRIVATE,
 799                                 p->pack_fd, win->offset);
 800                         if (win->base == MAP_FAILED)
 801                                 die("packfile %s cannot be mapped: %s",
 802                                         p->pack_name,
 803                                         strerror(errno));
 804                         pack_mmap_calls++;
 805                         pack_open_windows++;
 806                         if (pack_mapped > peak_pack_mapped)
 807                                 peak_pack_mapped = pack_mapped;
 808                         if (pack_open_windows > peak_pack_open_windows)
 809                                 peak_pack_open_windows = pack_open_windows;
 810                         win->next = p->windows;
 811                         p->windows = win;
 812                 }
 813         }
 814         if (win != *w_cursor) {
 815                 win->last_used = pack_used_ctr++;
 816                 win->inuse_cnt++;
 817                 *w_cursor = win;
 818         }
 819         offset -= win->offset;
 820         if (left)
 821                 *left = win->len - xsize_t(offset);
 822         return win->base + offset;
 823 }
 824
 825 static struct packed_git *alloc_packed_git(int extra)
 826 {
 827         struct packed_git *p = xmalloc(sizeof(*p) + extra);
 828         memset(p, 0, sizeof(*p));
 829         p->pack_fd = -1;
 830         return p;
 831 }
 832
 833 struct packed_git *add_packed_git(const char *path, int path_len, int local)
 834 {
 835         struct stat st;
 836         struct packed_git *p = alloc_packed_git(path_len + 2);
 837
 838         /*
 839          * Make sure a corresponding .pack file exists and that
 840          * the index looks sane.
 841          */
 842         path_len -= strlen(".idx");
 843         if (path_len < 1) {
 844                 free(p);
 845                 return NULL;
 846         }
 847         memcpy(p->pack_name, path, path_len);
 848
 849         strcpy(p->pack_name + path_len, ".keep");
 850         if (!access(p->pack_name, F_OK))
 851                 p->pack_keep = 1;
 852
 853         strcpy(p->pack_name + path_len, ".pack");
 854         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 855                 free(p);
 856                 return NULL;
 857         }
 858
 859         /* ok, it looks sane as far as we can check without
 860          * actually mapping the pack file.
 861          */
 862         p->pack_size = st.st_size;
 863         p->pack_local = local;
 864         p->mtime = st.st_mtime;
 865         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
 866                 hashclr(p->sha1);
 867         return p;
 868 }
 869
 870 struct packed_git *parse_pack_index(unsigned char *sha1)
 871 {
 872         const char *idx_path = sha1_pack_index_name(sha1);
 873         const char *path = sha1_pack_name(sha1);
 874         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
 875
 876         strcpy(p->pack_name, path);
 877         hashcpy(p->sha1, sha1);
 878         if (check_packed_git_idx(idx_path, p)) {
 879                 free(p);
 880                 return NULL;
 881         }
 882
 883         return p;
 884 }
 885
 886 void install_packed_git(struct packed_git *pack)
 887 {
 888         pack->next = packed_git;
 889         packed_git = pack;
 890 }
 891
 892 static void prepare_packed_git_one(char *objdir, int local)
 893 {
 894         /* Ensure that this buffer is large enough so that we can
 895            append "/pack/" without clobbering the stack even if
 896            strlen(objdir) were PATH_MAX.  */
 897         char path[PATH_MAX + 1 + 4 + 1 + 1];
 898         int len;
 899         DIR *dir;
 900         struct dirent *de;
 901
 902         sprintf(path, "%s/pack", objdir);
 903         len = strlen(path);
 904         dir = opendir(path);
 905         while (!dir && errno == EMFILE && unuse_one_window(packed_git, -1))
 906                 dir = opendir(path);
 907         if (!dir) {
 908                 if (errno != ENOENT)
 909                         error("unable to open object pack directory: %s: %s",
 910                               path, strerror(errno));
 911                 return;
 912         }
 913         path[len++] = '/';
 914         while ((de = readdir(dir)) != NULL) {
 915                 int namelen = strlen(de->d_name);
 916                 struct packed_git *p;
 917
 918                 if (!has_extension(de->d_name, ".idx"))
 919                         continue;
 920
 921                 if (len + namelen + 1 > sizeof(path))
 922                         continue;
 923
 924                 /* Don't reopen a pack we already have. */
 925                 strcpy(path + len, de->d_name);
 926                 for (p = packed_git; p; p = p->next) {
 927                         if (!memcmp(path, p->pack_name, len + namelen - 4))
 928                                 break;
 929                 }
 930                 if (p)
 931                         continue;
 932                 /* See if it really is a valid .idx file with corresponding
 933                  * .pack file that we can map.
 934                  */
 935                 p = add_packed_git(path, len + namelen, local);
 936                 if (!p)
 937                         continue;
 938                 install_packed_git(p);
 939         }
 940         closedir(dir);
 941 }
 942
 943 static int sort_pack(const void *a_, const void *b_)
 944 {
 945         struct packed_git *a = *((struct packed_git **)a_);
 946         struct packed_git *b = *((struct packed_git **)b_);
 947         int st;
 948
 949         /*
 950          * Local packs tend to contain objects specific to our
 951          * variant of the project than remote ones.  In addition,
 952          * remote ones could be on a network mounted filesystem.
 953          * Favor local ones for these reasons.
 954          */
 955         st = a->pack_local - b->pack_local;
 956         if (st)
 957                 return -st;
 958
 959         /*
 960          * Younger packs tend to contain more recent objects,
 961          * and more recent objects tend to get accessed more
 962          * often.
 963          */
 964         if (a->mtime < b->mtime)
 965                 return 1;
 966         else if (a->mtime == b->mtime)
 967                 return 0;
 968         return -1;
 969 }
 970
 971 static void rearrange_packed_git(void)
 972 {
 973         struct packed_git **ary, *p;
 974         int i, n;
 975
 976         for (n = 0, p = packed_git; p; p = p->next)
 977                 n++;
 978         if (n < 2)
 979                 return;
 980
 981         /* prepare an array of packed_git for easier sorting */
 982         ary = xcalloc(n, sizeof(struct packed_git *));
 983         for (n = 0, p = packed_git; p; p = p->next)
 984                 ary[n++] = p;
 985
 986         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
 987
 988         /* link them back again */
 989         for (i = 0; i < n - 1; i++)
 990                 ary[i]->next = ary[i + 1];
 991         ary[n - 1]->next = NULL;
 992         packed_git = ary[0];
 993
 994         free(ary);
 995 }
 996
 997 static int prepare_packed_git_run_once = 0;
 998 void prepare_packed_git(void)
 999 {
1000         struct alternate_object_database *alt;
1001
1002         if (prepare_packed_git_run_once)
1003                 return;
1004         prepare_packed_git_one(get_object_directory(), 1);
1005         prepare_alt_odb();
1006         for (alt = alt_odb_list; alt; alt = alt->next) {
1007                 alt->name[-1] = 0;
1008                 prepare_packed_git_one(alt->base, 0);
1009                 alt->name[-1] = '/';
1010         }
1011         rearrange_packed_git();
1012         prepare_packed_git_run_once = 1;
1013 }
1014
1015 void reprepare_packed_git(void)
1016 {
1017         discard_revindex();
1018         prepare_packed_git_run_once = 0;
1019         prepare_packed_git();
1020 }
1021
1022 static void mark_bad_packed_object(struct packed_git *p,
1023                                    const unsigned char *sha1)
1024 {
1025         unsigned i;
1026         for (i = 0; i < p->num_bad_objects; i++)
1027                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1028                         return;
1029         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1030         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1031         p->num_bad_objects++;
1032 }
1033
1034 static int has_packed_and_bad(const unsigned char *sha1)
1035 {
1036         struct packed_git *p;
1037         unsigned i;
1038
1039         for (p = packed_git; p; p = p->next)
1040                 for (i = 0; i < p->num_bad_objects; i++)
1041                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1042                                 return 1;
1043         return 0;
1044 }
1045
1046 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1047 {
1048         unsigned char real_sha1[20];
1049         hash_sha1_file(map, size, type, real_sha1);
1050         return hashcmp(sha1, real_sha1) ? -1 : 0;
1051 }
1052
1053 static int git_open_noatime(const char *name)
1054 {
1055         static int sha1_file_open_flag = O_NOATIME;
1056         int fd = open(name, O_RDONLY | sha1_file_open_flag);
1057
1058         /* Might the failure be due to O_NOATIME? */
1059         if (fd < 0 && errno != ENOENT && sha1_file_open_flag) {
1060                 fd = open(name, O_RDONLY);
1061                 if (fd >= 0)
1062                         sha1_file_open_flag = 0;
1063         }
1064         return fd;
1065 }
1066
1067 static int open_sha1_file(const unsigned char *sha1)
1068 {
1069         int fd;
1070         char *name = sha1_file_name(sha1);
1071         struct alternate_object_database *alt;
1072
1073         fd = git_open_noatime(name);
1074         if (fd >= 0)
1075                 return fd;
1076
1077         prepare_alt_odb();
1078         errno = ENOENT;
1079         for (alt = alt_odb_list; alt; alt = alt->next) {
1080                 name = alt->name;
1081                 fill_sha1_path(name, sha1);
1082                 fd = git_open_noatime(alt->base);
1083                 if (fd >= 0)
1084                         return fd;
1085         }
1086         return -1;
1087 }
1088
1089 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1090 {
1091         void *map;
1092         int fd;
1093
1094         fd = open_sha1_file(sha1);
1095         map = NULL;
1096         if (fd >= 0) {
1097                 struct stat st;
1098
1099                 if (!fstat(fd, &st)) {
1100                         *size = xsize_t(st.st_size);
1101                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1102                 }
1103                 close(fd);
1104         }
1105         return map;
1106 }
1107
1108 static int legacy_loose_object(unsigned char *map)
1109 {
1110         unsigned int word;
1111
1112         /*
1113          * Is it a zlib-compressed buffer? If so, the first byte
1114          * must be 0x78 (15-bit window size, deflated), and the
1115          * first 16-bit word is evenly divisible by 31
1116          */
1117         word = (map[0] << 8) + map[1];
1118         if (map[0] == 0x78 && !(word % 31))
1119                 return 1;
1120         else
1121                 return 0;
1122 }
1123
1124 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1125                 unsigned long len, enum object_type *type, unsigned long *sizep)
1126 {
1127         unsigned shift;
1128         unsigned long size, c;
1129         unsigned long used = 0;
1130
1131         c = buf[used++];
1132         *type = (c >> 4) & 7;
1133         size = c & 15;
1134         shift = 4;
1135         while (c & 0x80) {
1136                 if (len <= used || bitsizeof(long) <= shift) {
1137                         error("bad object header");
1138                         return 0;
1139                 }
1140                 c = buf[used++];
1141                 size += (c & 0x7f) << shift;
1142                 shift += 7;
1143         }
1144         *sizep = size;
1145         return used;
1146 }
1147
1148 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1149 {
1150         unsigned long size, used;
1151         static const char valid_loose_object_type[8] = {
1152                 0, /* OBJ_EXT */
1153                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1154                 0, /* "delta" and others are invalid in a loose object */
1155         };
1156         enum object_type type;
1157
1158         /* Get the data stream */
1159         memset(stream, 0, sizeof(*stream));
1160         stream->next_in = map;
1161         stream->avail_in = mapsize;
1162         stream->next_out = buffer;
1163         stream->avail_out = bufsiz;
1164
1165         if (legacy_loose_object(map)) {
1166                 git_inflate_init(stream);
1167                 return git_inflate(stream, 0);
1168         }
1169
1170
1171         /*
1172          * There used to be a second loose object header format which
1173          * was meant to mimic the in-pack format, allowing for direct
1174          * copy of the object data.  This format turned up not to be
1175          * really worth it and we don't write it any longer.  But we
1176          * can still read it.
1177          */
1178         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1179         if (!used || !valid_loose_object_type[type])
1180                 return -1;
1181         map += used;
1182         mapsize -= used;
1183
1184         /* Set up the stream for the rest.. */
1185         stream->next_in = map;
1186         stream->avail_in = mapsize;
1187         git_inflate_init(stream);
1188
1189         /* And generate the fake traditional header */
1190         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1191                                          typename(type), size);
1192         return 0;
1193 }
1194
1195 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1196 {
1197         int bytes = strlen(buffer) + 1;
1198         unsigned char *buf = xmallocz(size);
1199         unsigned long n;
1200         int status = Z_OK;
1201
1202         n = stream->total_out - bytes;
1203         if (n > size)
1204                 n = size;
1205         memcpy(buf, (char *) buffer + bytes, n);
1206         bytes = n;
1207         if (bytes <= size) {
1208                 /*
1209                  * The above condition must be (bytes <= size), not
1210                  * (bytes < size).  In other words, even though we
1211                  * expect no more output and set avail_out to zer0,
1212                  * the input zlib stream may have bytes that express
1213                  * "this concludes the stream", and we *do* want to
1214                  * eat that input.
1215                  *
1216                  * Otherwise we would not be able to test that we
1217                  * consumed all the input to reach the expected size;
1218                  * we also want to check that zlib tells us that all
1219                  * went well with status == Z_STREAM_END at the end.
1220                  */
1221                 stream->next_out = buf + bytes;
1222                 stream->avail_out = size - bytes;
1223                 while (status == Z_OK)
1224                         status = git_inflate(stream, Z_FINISH);
1225         }
1226         if (status == Z_STREAM_END && !stream->avail_in) {
1227                 git_inflate_end(stream);
1228                 return buf;
1229         }
1230
1231         if (status < 0)
1232                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1233         else if (stream->avail_in)
1234                 error("garbage at end of loose object '%s'",
1235                       sha1_to_hex(sha1));
1236         free(buf);
1237         return NULL;
1238 }
1239
1240 /*
1241  * We used to just use "sscanf()", but that's actually way
1242  * too permissive for what we want to check. So do an anal
1243  * object header parse by hand.
1244  */
1245 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1246 {
1247         char type[10];
1248         int i;
1249         unsigned long size;
1250
1251         /*
1252          * The type can be at most ten bytes (including the
1253          * terminating '\0' that we add), and is followed by
1254          * a space.
1255          */
1256         i = 0;
1257         for (;;) {
1258                 char c = *hdr++;
1259                 if (c == ' ')
1260                         break;
1261                 type[i++] = c;
1262                 if (i >= sizeof(type))
1263                         return -1;
1264         }
1265         type[i] = 0;
1266
1267         /*
1268          * The length must follow immediately, and be in canonical
1269          * decimal format (ie "010" is not valid).
1270          */
1271         size = *hdr++ - '0';
1272         if (size > 9)
1273                 return -1;
1274         if (size) {
1275                 for (;;) {
1276                         unsigned long c = *hdr - '0';
1277                         if (c > 9)
1278                                 break;
1279                         hdr++;
1280                         size = size * 10 + c;
1281                 }
1282         }
1283         *sizep = size;
1284
1285         /*
1286          * The length must be followed by a zero byte
1287          */
1288         return *hdr ? -1 : type_from_string(type);
1289 }
1290
1291 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1292 {
1293         int ret;
1294         z_stream stream;
1295         char hdr[8192];
1296
1297         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1298         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1299                 return NULL;
1300
1301         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1302 }
1303
1304 unsigned long get_size_from_delta(struct packed_git *p,
1305                                   struct pack_window **w_curs,
1306                                   off_t curpos)
1307 {
1308         const unsigned char *data;
1309         unsigned char delta_head[20], *in;
1310         z_stream stream;
1311         int st;
1312
1313         memset(&stream, 0, sizeof(stream));
1314         stream.next_out = delta_head;
1315         stream.avail_out = sizeof(delta_head);
1316
1317         git_inflate_init(&stream);
1318         do {
1319                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1320                 stream.next_in = in;
1321                 st = git_inflate(&stream, Z_FINISH);
1322                 curpos += stream.next_in - in;
1323         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1324                  stream.total_out < sizeof(delta_head));
1325         git_inflate_end(&stream);
1326         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1327                 error("delta data unpack-initial failed");
1328                 return 0;
1329         }
1330
1331         /* Examine the initial part of the delta to figure out
1332          * the result size.
1333          */
1334         data = delta_head;
1335
1336         /* ignore base size */
1337         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1338
1339         /* Read the result size */
1340         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1341 }
1342
1343 static off_t get_delta_base(struct packed_git *p,
1344                                     struct pack_window **w_curs,
1345                                     off_t *curpos,
1346                                     enum object_type type,
1347                                     off_t delta_obj_offset)
1348 {
1349         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1350         off_t base_offset;
1351
1352         /* use_pack() assured us we have [base_info, base_info + 20)
1353          * as a range that we can look at without walking off the
1354          * end of the mapped window.  Its actually the hash size
1355          * that is assured.  An OFS_DELTA longer than the hash size
1356          * is stupid, as then a REF_DELTA would be smaller to store.
1357          */
1358         if (type == OBJ_OFS_DELTA) {
1359                 unsigned used = 0;
1360                 unsigned char c = base_info[used++];
1361                 base_offset = c & 127;
1362                 while (c & 128) {
1363                         base_offset += 1;
1364                         if (!base_offset || MSB(base_offset, 7))
1365                                 return 0;  /* overflow */
1366                         c = base_info[used++];
1367                         base_offset = (base_offset << 7) + (c & 127);
1368                 }
1369                 base_offset = delta_obj_offset - base_offset;
1370                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1371                         return 0;  /* out of bound */
1372                 *curpos += used;
1373         } else if (type == OBJ_REF_DELTA) {
1374                 /* The base entry _must_ be in the same pack */
1375                 base_offset = find_pack_entry_one(base_info, p);
1376                 *curpos += 20;
1377         } else
1378                 die("I am totally screwed");
1379         return base_offset;
1380 }
1381
1382 /* forward declaration for a mutually recursive function */
1383 static int packed_object_info(struct packed_git *p, off_t offset,
1384                               unsigned long *sizep);
1385
1386 static int packed_delta_info(struct packed_git *p,
1387                              struct pack_window **w_curs,
1388                              off_t curpos,
1389                              enum object_type type,
1390                              off_t obj_offset,
1391                              unsigned long *sizep)
1392 {
1393         off_t base_offset;
1394
1395         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1396         if (!base_offset)
1397                 return OBJ_BAD;
1398         type = packed_object_info(p, base_offset, NULL);
1399         if (type <= OBJ_NONE) {
1400                 struct revindex_entry *revidx;
1401                 const unsigned char *base_sha1;
1402                 revidx = find_pack_revindex(p, base_offset);
1403                 if (!revidx)
1404                         return OBJ_BAD;
1405                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1406                 mark_bad_packed_object(p, base_sha1);
1407                 type = sha1_object_info(base_sha1, NULL);
1408                 if (type <= OBJ_NONE)
1409                         return OBJ_BAD;
1410         }
1411
1412         /* We choose to only get the type of the base object and
1413          * ignore potentially corrupt pack file that expects the delta
1414          * based on a base with a wrong size.  This saves tons of
1415          * inflate() calls.
1416          */
1417         if (sizep) {
1418                 *sizep = get_size_from_delta(p, w_curs, curpos);
1419                 if (*sizep == 0)
1420                         type = OBJ_BAD;
1421         }
1422
1423         return type;
1424 }
1425
1426 static int unpack_object_header(struct packed_git *p,
1427                                 struct pack_window **w_curs,
1428                                 off_t *curpos,
1429                                 unsigned long *sizep)
1430 {
1431         unsigned char *base;
1432         unsigned int left;
1433         unsigned long used;
1434         enum object_type type;
1435
1436         /* use_pack() assures us we have [base, base + 20) available
1437          * as a range that we can look at at.  (Its actually the hash
1438          * size that is assured.)  With our object header encoding
1439          * the maximum deflated object size is 2^137, which is just
1440          * insane, so we know won't exceed what we have been given.
1441          */
1442         base = use_pack(p, w_curs, *curpos, &left);
1443         used = unpack_object_header_buffer(base, left, &type, sizep);
1444         if (!used) {
1445                 type = OBJ_BAD;
1446         } else
1447                 *curpos += used;
1448
1449         return type;
1450 }
1451
1452 const char *packed_object_info_detail(struct packed_git *p,
1453                                       off_t obj_offset,
1454                                       unsigned long *size,
1455                                       unsigned long *store_size,
1456                                       unsigned int *delta_chain_length,
1457                                       unsigned char *base_sha1)
1458 {
1459         struct pack_window *w_curs = NULL;
1460         off_t curpos;
1461         unsigned long dummy;
1462         unsigned char *next_sha1;
1463         enum object_type type;
1464         struct revindex_entry *revidx;
1465
1466         *delta_chain_length = 0;
1467         curpos = obj_offset;
1468         type = unpack_object_header(p, &w_curs, &curpos, size);
1469
1470         revidx = find_pack_revindex(p, obj_offset);
1471         *store_size = revidx[1].offset - obj_offset;
1472
1473         for (;;) {
1474                 switch (type) {
1475                 default:
1476                         die("pack %s contains unknown object type %d",
1477                             p->pack_name, type);
1478                 case OBJ_COMMIT:
1479                 case OBJ_TREE:
1480                 case OBJ_BLOB:
1481                 case OBJ_TAG:
1482                         unuse_pack(&w_curs);
1483                         return typename(type);
1484                 case OBJ_OFS_DELTA:
1485                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1486                         if (!obj_offset)
1487                                 die("pack %s contains bad delta base reference of type %s",
1488                                     p->pack_name, typename(type));
1489                         if (*delta_chain_length == 0) {
1490                                 revidx = find_pack_revindex(p, obj_offset);
1491                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1492                         }
1493                         break;
1494                 case OBJ_REF_DELTA:
1495                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1496                         if (*delta_chain_length == 0)
1497                                 hashcpy(base_sha1, next_sha1);
1498                         obj_offset = find_pack_entry_one(next_sha1, p);
1499                         break;
1500                 }
1501                 (*delta_chain_length)++;
1502                 curpos = obj_offset;
1503                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1504         }
1505 }
1506
1507 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1508                               unsigned long *sizep)
1509 {
1510         struct pack_window *w_curs = NULL;
1511         unsigned long size;
1512         off_t curpos = obj_offset;
1513         enum object_type type;
1514
1515         type = unpack_object_header(p, &w_curs, &curpos, &size);
1516
1517         switch (type) {
1518         case OBJ_OFS_DELTA:
1519         case OBJ_REF_DELTA:
1520                 type = packed_delta_info(p, &w_curs, curpos,
1521                                          type, obj_offset, sizep);
1522                 break;
1523         case OBJ_COMMIT:
1524         case OBJ_TREE:
1525         case OBJ_BLOB:
1526         case OBJ_TAG:
1527                 if (sizep)
1528                         *sizep = size;
1529                 break;
1530         default:
1531                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1532                       type, (uintmax_t)obj_offset, p->pack_name);
1533                 type = OBJ_BAD;
1534         }
1535         unuse_pack(&w_curs);
1536         return type;
1537 }
1538
1539 static void *unpack_compressed_entry(struct packed_git *p,
1540                                     struct pack_window **w_curs,
1541                                     off_t curpos,
1542                                     unsigned long size)
1543 {
1544         int st;
1545         z_stream stream;
1546         unsigned char *buffer, *in;
1547
1548         buffer = xmallocz(size);
1549         memset(&stream, 0, sizeof(stream));
1550         stream.next_out = buffer;
1551         stream.avail_out = size + 1;
1552
1553         git_inflate_init(&stream);
1554         do {
1555                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1556                 stream.next_in = in;
1557                 st = git_inflate(&stream, Z_FINISH);
1558                 if (!stream.avail_out)
1559                         break; /* the payload is larger than it should be */
1560                 curpos += stream.next_in - in;
1561         } while (st == Z_OK || st == Z_BUF_ERROR);
1562         git_inflate_end(&stream);
1563         if ((st != Z_STREAM_END) || stream.total_out != size) {
1564                 free(buffer);
1565                 return NULL;
1566         }
1567
1568         return buffer;
1569 }
1570
1571 #define MAX_DELTA_CACHE (256)
1572
1573 static size_t delta_base_cached;
1574
1575 static struct delta_base_cache_lru_list {
1576         struct delta_base_cache_lru_list *prev;
1577         struct delta_base_cache_lru_list *next;
1578 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1579
1580 static struct delta_base_cache_entry {
1581         struct delta_base_cache_lru_list lru;
1582         void *data;
1583         struct packed_git *p;
1584         off_t base_offset;
1585         unsigned long size;
1586         enum object_type type;
1587 } delta_base_cache[MAX_DELTA_CACHE];
1588
1589 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1590 {
1591         unsigned long hash;
1592
1593         hash = (unsigned long)p + (unsigned long)base_offset;
1594         hash += (hash >> 8) + (hash >> 16);
1595         return hash % MAX_DELTA_CACHE;
1596 }
1597
1598 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1599         unsigned long *base_size, enum object_type *type, int keep_cache)
1600 {
1601         void *ret;
1602         unsigned long hash = pack_entry_hash(p, base_offset);
1603         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1604
1605         ret = ent->data;
1606         if (!ret || ent->p != p || ent->base_offset != base_offset)
1607                 return unpack_entry(p, base_offset, type, base_size);
1608
1609         if (!keep_cache) {
1610                 ent->data = NULL;
1611                 ent->lru.next->prev = ent->lru.prev;
1612                 ent->lru.prev->next = ent->lru.next;
1613                 delta_base_cached -= ent->size;
1614         } else {
1615                 ret = xmemdupz(ent->data, ent->size);
1616         }
1617         *type = ent->type;
1618         *base_size = ent->size;
1619         return ret;
1620 }
1621
1622 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1623 {
1624         if (ent->data) {
1625                 free(ent->data);
1626                 ent->data = NULL;
1627                 ent->lru.next->prev = ent->lru.prev;
1628                 ent->lru.prev->next = ent->lru.next;
1629                 delta_base_cached -= ent->size;
1630         }
1631 }
1632
1633 void clear_delta_base_cache(void)
1634 {
1635         unsigned long p;
1636         for (p = 0; p < MAX_DELTA_CACHE; p++)
1637                 release_delta_base_cache(&delta_base_cache[p]);
1638 }
1639
1640 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1641         void *base, unsigned long base_size, enum object_type type)
1642 {
1643         unsigned long hash = pack_entry_hash(p, base_offset);
1644         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1645         struct delta_base_cache_lru_list *lru;
1646
1647         release_delta_base_cache(ent);
1648         delta_base_cached += base_size;
1649
1650         for (lru = delta_base_cache_lru.next;
1651              delta_base_cached > delta_base_cache_limit
1652              && lru != &delta_base_cache_lru;
1653              lru = lru->next) {
1654                 struct delta_base_cache_entry *f = (void *)lru;
1655                 if (f->type == OBJ_BLOB)
1656                         release_delta_base_cache(f);
1657         }
1658         for (lru = delta_base_cache_lru.next;
1659              delta_base_cached > delta_base_cache_limit
1660              && lru != &delta_base_cache_lru;
1661              lru = lru->next) {
1662                 struct delta_base_cache_entry *f = (void *)lru;
1663                 release_delta_base_cache(f);
1664         }
1665
1666         ent->p = p;
1667         ent->base_offset = base_offset;
1668         ent->type = type;
1669         ent->data = base;
1670         ent->size = base_size;
1671         ent->lru.next = &delta_base_cache_lru;
1672         ent->lru.prev = delta_base_cache_lru.prev;
1673         delta_base_cache_lru.prev->next = &ent->lru;
1674         delta_base_cache_lru.prev = &ent->lru;
1675 }
1676
1677 static void *read_object(const unsigned char *sha1, enum object_type *type,
1678                          unsigned long *size);
1679
1680 static void *unpack_delta_entry(struct packed_git *p,
1681                                 struct pack_window **w_curs,
1682                                 off_t curpos,
1683                                 unsigned long delta_size,
1684                                 off_t obj_offset,
1685                                 enum object_type *type,
1686                                 unsigned long *sizep)
1687 {
1688         void *delta_data, *result, *base;
1689         unsigned long base_size;
1690         off_t base_offset;
1691
1692         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1693         if (!base_offset) {
1694                 error("failed to validate delta base reference "
1695                       "at offset %"PRIuMAX" from %s",
1696                       (uintmax_t)curpos, p->pack_name);
1697                 return NULL;
1698         }
1699         unuse_pack(w_curs);
1700         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1701         if (!base) {
1702                 /*
1703                  * We're probably in deep shit, but let's try to fetch
1704                  * the required base anyway from another pack or loose.
1705                  * This is costly but should happen only in the presence
1706                  * of a corrupted pack, and is better than failing outright.
1707                  */
1708                 struct revindex_entry *revidx;
1709                 const unsigned char *base_sha1;
1710                 revidx = find_pack_revindex(p, base_offset);
1711                 if (!revidx)
1712                         return NULL;
1713                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1714                 error("failed to read delta base object %s"
1715                       " at offset %"PRIuMAX" from %s",
1716                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1717                       p->pack_name);
1718                 mark_bad_packed_object(p, base_sha1);
1719                 base = read_object(base_sha1, type, &base_size);
1720                 if (!base)
1721                         return NULL;
1722         }
1723
1724         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1725         if (!delta_data) {
1726                 error("failed to unpack compressed delta "
1727                       "at offset %"PRIuMAX" from %s",
1728                       (uintmax_t)curpos, p->pack_name);
1729                 free(base);
1730                 return NULL;
1731         }
1732         result = patch_delta(base, base_size,
1733                              delta_data, delta_size,
1734                              sizep);
1735         if (!result)
1736                 die("failed to apply delta");
1737         free(delta_data);
1738         add_delta_base_cache(p, base_offset, base, base_size, *type);
1739         return result;
1740 }
1741
1742 int do_check_packed_object_crc;
1743
1744 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1745                    enum object_type *type, unsigned long *sizep)
1746 {
1747         struct pack_window *w_curs = NULL;
1748         off_t curpos = obj_offset;
1749         void *data;
1750
1751         if (do_check_packed_object_crc && p->index_version > 1) {
1752                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1753                 unsigned long len = revidx[1].offset - obj_offset;
1754                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1755                         const unsigned char *sha1 =
1756                                 nth_packed_object_sha1(p, revidx->nr);
1757                         error("bad packed object CRC for %s",
1758                               sha1_to_hex(sha1));
1759                         mark_bad_packed_object(p, sha1);
1760                         unuse_pack(&w_curs);
1761                         return NULL;
1762                 }
1763         }
1764
1765         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1766         switch (*type) {
1767         case OBJ_OFS_DELTA:
1768         case OBJ_REF_DELTA:
1769                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1770                                           obj_offset, type, sizep);
1771                 break;
1772         case OBJ_COMMIT:
1773         case OBJ_TREE:
1774         case OBJ_BLOB:
1775         case OBJ_TAG:
1776                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1777                 break;
1778         default:
1779                 data = NULL;
1780                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1781                       *type, (uintmax_t)obj_offset, p->pack_name);
1782         }
1783         unuse_pack(&w_curs);
1784         return data;
1785 }
1786
1787 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1788                                             uint32_t n)
1789 {
1790         const unsigned char *index = p->index_data;
1791         if (!index) {
1792                 if (open_pack_index(p))
1793                         return NULL;
1794                 index = p->index_data;
1795         }
1796         if (n >= p->num_objects)
1797                 return NULL;
1798         index += 4 * 256;
1799         if (p->index_version == 1) {
1800                 return index + 24 * n + 4;
1801         } else {
1802                 index += 8;
1803                 return index + 20 * n;
1804         }
1805 }
1806
1807 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1808 {
1809         const unsigned char *index = p->index_data;
1810         index += 4 * 256;
1811         if (p->index_version == 1) {
1812                 return ntohl(*((uint32_t *)(index + 24 * n)));
1813         } else {
1814                 uint32_t off;
1815                 index += 8 + p->num_objects * (20 + 4);
1816                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1817                 if (!(off & 0x80000000))
1818                         return off;
1819                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1820                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1821                                    ntohl(*((uint32_t *)(index + 4)));
1822         }
1823 }
1824
1825 off_t find_pack_entry_one(const unsigned char *sha1,
1826                                   struct packed_git *p)
1827 {
1828         const uint32_t *level1_ofs = p->index_data;
1829         const unsigned char *index = p->index_data;
1830         unsigned hi, lo, stride;
1831         static int use_lookup = -1;
1832         static int debug_lookup = -1;
1833
1834         if (debug_lookup < 0)
1835                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1836
1837         if (!index) {
1838                 if (open_pack_index(p))
1839                         return 0;
1840                 level1_ofs = p->index_data;
1841                 index = p->index_data;
1842         }
1843         if (p->index_version > 1) {
1844                 level1_ofs += 2;
1845                 index += 8;
1846         }
1847         index += 4 * 256;
1848         hi = ntohl(level1_ofs[*sha1]);
1849         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1850         if (p->index_version > 1) {
1851                 stride = 20;
1852         } else {
1853                 stride = 24;
1854                 index += 4;
1855         }
1856
1857         if (debug_lookup)
1858                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1859                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1860
1861         if (use_lookup < 0)
1862                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1863         if (use_lookup) {
1864                 int pos = sha1_entry_pos(index, stride, 0,
1865                                          lo, hi, p->num_objects, sha1);
1866                 if (pos < 0)
1867                         return 0;
1868                 return nth_packed_object_offset(p, pos);
1869         }
1870
1871         do {
1872                 unsigned mi = (lo + hi) / 2;
1873                 int cmp = hashcmp(index + mi * stride, sha1);
1874
1875                 if (debug_lookup)
1876                         printf("lo %u hi %u rg %u mi %u\n",
1877                                lo, hi, hi - lo, mi);
1878                 if (!cmp)
1879                         return nth_packed_object_offset(p, mi);
1880                 if (cmp > 0)
1881                         hi = mi;
1882                 else
1883                         lo = mi+1;
1884         } while (lo < hi);
1885         return 0;
1886 }
1887
1888 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1889 {
1890         static struct packed_git *last_found = (void *)1;
1891         struct packed_git *p;
1892         off_t offset;
1893
1894         prepare_packed_git();
1895         if (!packed_git)
1896                 return 0;
1897         p = (last_found == (void *)1) ? packed_git : last_found;
1898
1899         do {
1900                 if (p->num_bad_objects) {
1901                         unsigned i;
1902                         for (i = 0; i < p->num_bad_objects; i++)
1903                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1904                                         goto next;
1905                 }
1906
1907                 offset = find_pack_entry_one(sha1, p);
1908                 if (offset) {
1909                         /*
1910                          * We are about to tell the caller where they can
1911                          * locate the requested object.  We better make
1912                          * sure the packfile is still here and can be
1913                          * accessed before supplying that answer, as
1914                          * it may have been deleted since the index
1915                          * was loaded!
1916                          */
1917                         if (p->pack_fd == -1 && open_packed_git(p)) {
1918                                 error("packfile %s cannot be accessed", p->pack_name);
1919                                 goto next;
1920                         }
1921                         e->offset = offset;
1922                         e->p = p;
1923                         hashcpy(e->sha1, sha1);
1924                         last_found = p;
1925                         return 1;
1926                 }
1927
1928                 next:
1929                 if (p == last_found)
1930                         p = packed_git;
1931                 else
1932                         p = p->next;
1933                 if (p == last_found)
1934                         p = p->next;
1935         } while (p);
1936         return 0;
1937 }
1938
1939 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1940                                   struct packed_git *packs)
1941 {
1942         struct packed_git *p;
1943
1944         for (p = packs; p; p = p->next) {
1945                 if (find_pack_entry_one(sha1, p))
1946                         return p;
1947         }
1948         return NULL;
1949
1950 }
1951
1952 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1953 {
1954         int status;
1955         unsigned long mapsize, size;
1956         void *map;
1957         z_stream stream;
1958         char hdr[32];
1959
1960         map = map_sha1_file(sha1, &mapsize);
1961         if (!map)
1962                 return error("unable to find %s", sha1_to_hex(sha1));
1963         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1964                 status = error("unable to unpack %s header",
1965                                sha1_to_hex(sha1));
1966         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1967                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1968         else if (sizep)
1969                 *sizep = size;
1970         git_inflate_end(&stream);
1971         munmap(map, mapsize);
1972         return status;
1973 }
1974
1975 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1976 {
1977         struct pack_entry e;
1978         int status;
1979
1980         if (!find_pack_entry(sha1, &e)) {
1981                 /* Most likely it's a loose object. */
1982                 status = sha1_loose_object_info(sha1, sizep);
1983                 if (status >= 0)
1984                         return status;
1985
1986                 /* Not a loose object; someone else may have just packed it. */
1987                 reprepare_packed_git();
1988                 if (!find_pack_entry(sha1, &e))
1989                         return status;
1990         }
1991
1992         status = packed_object_info(e.p, e.offset, sizep);
1993         if (status < 0) {
1994                 mark_bad_packed_object(e.p, sha1);
1995                 status = sha1_object_info(sha1, sizep);
1996         }
1997
1998         return status;
1999 }
2000
2001 static void *read_packed_sha1(const unsigned char *sha1,
2002                               enum object_type *type, unsigned long *size)
2003 {
2004         struct pack_entry e;
2005         void *data;
2006
2007         if (!find_pack_entry(sha1, &e))
2008                 return NULL;
2009         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2010         if (!data) {
2011                 /*
2012                  * We're probably in deep shit, but let's try to fetch
2013                  * the required object anyway from another pack or loose.
2014                  * This should happen only in the presence of a corrupted
2015                  * pack, and is better than failing outright.
2016                  */
2017                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2018                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2019                 mark_bad_packed_object(e.p, sha1);
2020                 data = read_object(sha1, type, size);
2021         }
2022         return data;
2023 }
2024
2025 /*
2026  * This is meant to hold a *small* number of objects that you would
2027  * want read_sha1_file() to be able to return, but yet you do not want
2028  * to write them into the object store (e.g. a browse-only
2029  * application).
2030  */
2031 static struct cached_object {
2032         unsigned char sha1[20];
2033         enum object_type type;
2034         void *buf;
2035         unsigned long size;
2036 } *cached_objects;
2037 static int cached_object_nr, cached_object_alloc;
2038
2039 static struct cached_object empty_tree = {
2040         EMPTY_TREE_SHA1_BIN,
2041         OBJ_TREE,
2042         "",
2043         0
2044 };
2045
2046 static struct cached_object *find_cached_object(const unsigned char *sha1)
2047 {
2048         int i;
2049         struct cached_object *co = cached_objects;
2050
2051         for (i = 0; i < cached_object_nr; i++, co++) {
2052                 if (!hashcmp(co->sha1, sha1))
2053                         return co;
2054         }
2055         if (!hashcmp(sha1, empty_tree.sha1))
2056                 return &empty_tree;
2057         return NULL;
2058 }
2059
2060 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2061                       unsigned char *sha1)
2062 {
2063         struct cached_object *co;
2064
2065         hash_sha1_file(buf, len, typename(type), sha1);
2066         if (has_sha1_file(sha1) || find_cached_object(sha1))
2067                 return 0;
2068         if (cached_object_alloc <= cached_object_nr) {
2069                 cached_object_alloc = alloc_nr(cached_object_alloc);
2070                 cached_objects = xrealloc(cached_objects,
2071                                           sizeof(*cached_objects) *
2072                                           cached_object_alloc);
2073         }
2074         co = &cached_objects[cached_object_nr++];
2075         co->size = len;
2076         co->type = type;
2077         co->buf = xmalloc(len);
2078         memcpy(co->buf, buf, len);
2079         hashcpy(co->sha1, sha1);
2080         return 0;
2081 }
2082
2083 static void *read_object(const unsigned char *sha1, enum object_type *type,
2084                          unsigned long *size)
2085 {
2086         unsigned long mapsize;
2087         void *map, *buf;
2088         struct cached_object *co;
2089
2090         co = find_cached_object(sha1);
2091         if (co) {
2092                 *type = co->type;
2093                 *size = co->size;
2094                 return xmemdupz(co->buf, co->size);
2095         }
2096
2097         buf = read_packed_sha1(sha1, type, size);
2098         if (buf)
2099                 return buf;
2100         map = map_sha1_file(sha1, &mapsize);
2101         if (map) {
2102                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2103                 munmap(map, mapsize);
2104                 return buf;
2105         }
2106         reprepare_packed_git();
2107         return read_packed_sha1(sha1, type, size);
2108 }
2109
2110 void *read_sha1_file_repl(const unsigned char *sha1,
2111                           enum object_type *type,
2112                           unsigned long *size,
2113                           const unsigned char **replacement)
2114 {
2115         const unsigned char *repl = lookup_replace_object(sha1);
2116         void *data = read_object(repl, type, size);
2117
2118         /* die if we replaced an object with one that does not exist */
2119         if (!data && repl != sha1)
2120                 die("replacement %s not found for %s",
2121                     sha1_to_hex(repl), sha1_to_hex(sha1));
2122
2123         /* legacy behavior is to die on corrupted objects */
2124         if (!data && (has_loose_object(repl) || has_packed_and_bad(repl)))
2125                 die("object %s is corrupted", sha1_to_hex(repl));
2126
2127         if (replacement)
2128                 *replacement = repl;
2129
2130         return data;
2131 }
2132
2133 void *read_object_with_reference(const unsigned char *sha1,
2134                                  const char *required_type_name,
2135                                  unsigned long *size,
2136                                  unsigned char *actual_sha1_return)
2137 {
2138         enum object_type type, required_type;
2139         void *buffer;
2140         unsigned long isize;
2141         unsigned char actual_sha1[20];
2142
2143         required_type = type_from_string(required_type_name);
2144         hashcpy(actual_sha1, sha1);
2145         while (1) {
2146                 int ref_length = -1;
2147                 const char *ref_type = NULL;
2148
2149                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2150                 if (!buffer)
2151                         return NULL;
2152                 if (type == required_type) {
2153                         *size = isize;
2154                         if (actual_sha1_return)
2155                                 hashcpy(actual_sha1_return, actual_sha1);
2156                         return buffer;
2157                 }
2158                 /* Handle references */
2159                 else if (type == OBJ_COMMIT)
2160                         ref_type = "tree ";
2161                 else if (type == OBJ_TAG)
2162                         ref_type = "object ";
2163                 else {
2164                         free(buffer);
2165                         return NULL;
2166                 }
2167                 ref_length = strlen(ref_type);
2168
2169                 if (ref_length + 40 > isize ||
2170                     memcmp(buffer, ref_type, ref_length) ||
2171                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2172                         free(buffer);
2173                         return NULL;
2174                 }
2175                 free(buffer);
2176                 /* Now we have the ID of the referred-to object in
2177                  * actual_sha1.  Check again. */
2178         }
2179 }
2180
2181 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2182                                     const char *type, unsigned char *sha1,
2183                                     char *hdr, int *hdrlen)
2184 {
2185         git_SHA_CTX c;
2186
2187         /* Generate the header */
2188         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2189
2190         /* Sha1.. */
2191         git_SHA1_Init(&c);
2192         git_SHA1_Update(&c, hdr, *hdrlen);
2193         git_SHA1_Update(&c, buf, len);
2194         git_SHA1_Final(sha1, &c);
2195 }
2196
2197 /*
2198  * Move the just written object into its final resting place.
2199  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2200  * "moving" is only a part of what it does, when no patch between
2201  * master to pu changes the call sites of this function.
2202  */
2203 int move_temp_to_file(const char *tmpfile, const char *filename)
2204 {
2205         int ret = 0;
2206
2207         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2208                 goto try_rename;
2209         else if (link(tmpfile, filename))
2210                 ret = errno;
2211
2212         /*
2213          * Coda hack - coda doesn't like cross-directory links,
2214          * so we fall back to a rename, which will mean that it
2215          * won't be able to check collisions, but that's not a
2216          * big deal.
2217          *
2218          * The same holds for FAT formatted media.
2219          *
2220          * When this succeeds, we just return.  We have nothing
2221          * left to unlink.
2222          */
2223         if (ret && ret != EEXIST) {
2224         try_rename:
2225                 if (!rename(tmpfile, filename))
2226                         goto out;
2227                 ret = errno;
2228         }
2229         unlink_or_warn(tmpfile);
2230         if (ret) {
2231                 if (ret != EEXIST) {
2232                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2233                 }
2234                 /* FIXME!!! Collision check here ? */
2235         }
2236
2237 out:
2238         if (adjust_shared_perm(filename))
2239                 return error("unable to set permission to '%s'", filename);
2240         return 0;
2241 }
2242
2243 static int write_buffer(int fd, const void *buf, size_t len)
2244 {
2245         if (write_in_full(fd, buf, len) < 0)
2246                 return error("file write error (%s)", strerror(errno));
2247         return 0;
2248 }
2249
2250 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2251                    unsigned char *sha1)
2252 {
2253         char hdr[32];
2254         int hdrlen;
2255         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2256         return 0;
2257 }
2258
2259 /* Finalize a file on disk, and close it. */
2260 static void close_sha1_file(int fd)
2261 {
2262         if (fsync_object_files)
2263                 fsync_or_die(fd, "sha1 file");
2264         if (close(fd) != 0)
2265                 die_errno("error when closing sha1 file");
2266 }
2267
2268 /* Size of directory component, including the ending '/' */
2269 static inline int directory_size(const char *filename)
2270 {
2271         const char *s = strrchr(filename, '/');
2272         if (!s)
2273                 return 0;
2274         return s - filename + 1;
2275 }
2276
2277 /*
2278  * This creates a temporary file in the same directory as the final
2279  * 'filename'
2280  *
2281  * We want to avoid cross-directory filename renames, because those
2282  * can have problems on various filesystems (FAT, NFS, Coda).
2283  */
2284 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2285 {
2286         int fd, dirlen = directory_size(filename);
2287
2288         if (dirlen + 20 > bufsiz) {
2289                 errno = ENAMETOOLONG;
2290                 return -1;
2291         }
2292         memcpy(buffer, filename, dirlen);
2293         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2294         fd = git_mkstemp_mode(buffer, 0444);
2295         if (fd < 0 && dirlen && errno == ENOENT) {
2296                 /* Make sure the directory exists */
2297                 memcpy(buffer, filename, dirlen);
2298                 buffer[dirlen-1] = 0;
2299                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2300                         return -1;
2301
2302                 /* Try again */
2303                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2304                 fd = git_mkstemp_mode(buffer, 0444);
2305         }
2306         return fd;
2307 }
2308
2309 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2310                               void *buf, unsigned long len, time_t mtime)
2311 {
2312         int fd, ret;
2313         unsigned char compressed[4096];
2314         z_stream stream;
2315         git_SHA_CTX c;
2316         unsigned char parano_sha1[20];
2317         char *filename;
2318         static char tmpfile[PATH_MAX];
2319
2320         filename = sha1_file_name(sha1);
2321         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2322         while (fd < 0 && errno == EMFILE && unuse_one_window(packed_git, -1))
2323                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2324         if (fd < 0) {
2325                 if (errno == EACCES)
2326                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2327                 else
2328                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2329         }
2330
2331         /* Set it up */
2332         memset(&stream, 0, sizeof(stream));
2333         deflateInit(&stream, zlib_compression_level);
2334         stream.next_out = compressed;
2335         stream.avail_out = sizeof(compressed);
2336         git_SHA1_Init(&c);
2337
2338         /* First header.. */
2339         stream.next_in = (unsigned char *)hdr;
2340         stream.avail_in = hdrlen;
2341         while (deflate(&stream, 0) == Z_OK)
2342                 /* nothing */;
2343         git_SHA1_Update(&c, hdr, hdrlen);
2344
2345         /* Then the data itself.. */
2346         stream.next_in = buf;
2347         stream.avail_in = len;
2348         do {
2349                 unsigned char *in0 = stream.next_in;
2350                 ret = deflate(&stream, Z_FINISH);
2351                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2352                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2353                         die("unable to write sha1 file");
2354                 stream.next_out = compressed;
2355                 stream.avail_out = sizeof(compressed);
2356         } while (ret == Z_OK);
2357
2358         if (ret != Z_STREAM_END)
2359                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2360         ret = deflateEnd(&stream);
2361         if (ret != Z_OK)
2362                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2363         git_SHA1_Final(parano_sha1, &c);
2364         if (hashcmp(sha1, parano_sha1) != 0)
2365                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2366
2367         close_sha1_file(fd);
2368
2369         if (mtime) {
2370                 struct utimbuf utb;
2371                 utb.actime = mtime;
2372                 utb.modtime = mtime;
2373                 if (utime(tmpfile, &utb) < 0)
2374                         warning("failed utime() on %s: %s",
2375                                 tmpfile, strerror(errno));
2376         }
2377
2378         return move_temp_to_file(tmpfile, filename);
2379 }
2380
2381 int write_sha1_file(void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2382 {
2383         unsigned char sha1[20];
2384         char hdr[32];
2385         int hdrlen;
2386
2387         /* Normally if we have it in the pack then we do not bother writing
2388          * it out into .git/objects/??/?{38} file.
2389          */
2390         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2391         if (returnsha1)
2392                 hashcpy(returnsha1, sha1);
2393         if (has_sha1_file(sha1))
2394                 return 0;
2395         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2396 }
2397
2398 int force_object_loose(const unsigned char *sha1, time_t mtime)
2399 {
2400         void *buf;
2401         unsigned long len;
2402         enum object_type type;
2403         char hdr[32];
2404         int hdrlen;
2405         int ret;
2406
2407         if (has_loose_object(sha1))
2408                 return 0;
2409         buf = read_packed_sha1(sha1, &type, &len);
2410         if (!buf)
2411                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2412         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2413         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2414         free(buf);
2415
2416         return ret;
2417 }
2418
2419 int has_pack_index(const unsigned char *sha1)
2420 {
2421         struct stat st;
2422         if (stat(sha1_pack_index_name(sha1), &st))
2423                 return 0;
2424         return 1;
2425 }
2426
2427 int has_sha1_pack(const unsigned char *sha1)
2428 {
2429         struct pack_entry e;
2430         return find_pack_entry(sha1, &e);
2431 }
2432
2433 int has_sha1_file(const unsigned char *sha1)
2434 {
2435         struct pack_entry e;
2436
2437         if (find_pack_entry(sha1, &e))
2438                 return 1;
2439         return has_loose_object(sha1);
2440 }
2441
2442 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2443                      int write_object, enum object_type type, const char *path)
2444 {
2445         int ret, re_allocated = 0;
2446
2447         if (!type)
2448                 type = OBJ_BLOB;
2449
2450         /*
2451          * Convert blobs to git internal format
2452          */
2453         if ((type == OBJ_BLOB) && path) {
2454                 struct strbuf nbuf = STRBUF_INIT;
2455                 if (convert_to_git(path, buf, size, &nbuf,
2456                                    write_object ? safe_crlf : 0)) {
2457                         buf = strbuf_detach(&nbuf, &size);
2458                         re_allocated = 1;
2459                 }
2460         }
2461
2462         if (write_object)
2463                 ret = write_sha1_file(buf, size, typename(type), sha1);
2464         else
2465                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2466         if (re_allocated)
2467                 free(buf);
2468         return ret;
2469 }
2470
2471 #define SMALL_FILE_SIZE (32*1024)
2472
2473 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2474              enum object_type type, const char *path)
2475 {
2476         int ret;
2477         size_t size = xsize_t(st->st_size);
2478
2479         if (!S_ISREG(st->st_mode)) {
2480                 struct strbuf sbuf = STRBUF_INIT;
2481                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2482                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2483                                         type, path);
2484                 else
2485                         ret = -1;
2486                 strbuf_release(&sbuf);
2487         } else if (size <= SMALL_FILE_SIZE) {
2488                 char *buf = xmalloc(size);
2489                 if (size == read_in_full(fd, buf, size))
2490                         ret = index_mem(sha1, buf, size, write_object, type,
2491                                         path);
2492                 else
2493                         ret = error("short read %s", strerror(errno));
2494                 free(buf);
2495         } else if (size) {
2496                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2497                 ret = index_mem(sha1, buf, size, write_object, type, path);
2498                 munmap(buf, size);
2499         } else
2500                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2501         close(fd);
2502         return ret;
2503 }
2504
2505 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2506 {
2507         int fd;
2508         struct strbuf sb = STRBUF_INIT;
2509
2510         switch (st->st_mode & S_IFMT) {
2511         case S_IFREG:
2512                 fd = open(path, O_RDONLY);
2513                 if (fd < 0)
2514                         return error("open(\"%s\"): %s", path,
2515                                      strerror(errno));
2516                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2517                         return error("%s: failed to insert into database",
2518                                      path);
2519                 break;
2520         case S_IFLNK:
2521                 if (strbuf_readlink(&sb, path, st->st_size)) {
2522                         char *errstr = strerror(errno);
2523                         return error("readlink(\"%s\"): %s", path,
2524                                      errstr);
2525                 }
2526                 if (!write_object)
2527                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2528                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2529                         return error("%s: failed to insert into database",
2530                                      path);
2531                 strbuf_release(&sb);
2532                 break;
2533         case S_IFDIR:
2534                 return resolve_gitlink_ref(path, "HEAD", sha1);
2535         default:
2536                 return error("%s: unsupported file type", path);
2537         }
2538         return 0;
2539 }
2540
2541 int read_pack_header(int fd, struct pack_header *header)
2542 {
2543         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2544                 /* "eof before pack header was fully read" */
2545                 return PH_ERROR_EOF;
2546
2547         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2548                 /* "protocol error (pack signature mismatch detected)" */
2549                 return PH_ERROR_PACK_SIGNATURE;
2550         if (!pack_version_ok(header->hdr_version))
2551                 /* "protocol error (pack version unsupported)" */
2552                 return PH_ERROR_PROTOCOL;
2553         return 0;
2554 }