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 int safe_create_leading_directories(char *path)
  39 {
  40         char *pos = path + offset_1st_component(path);
  41         struct stat st;
  42
  43         while (pos) {
  44                 pos = strchr(pos, '/');
  45                 if (!pos)
  46                         break;
  47                 while (*++pos == '/')
  48                         ;
  49                 if (!*pos)
  50                         break;
  51                 *--pos = '\0';
  52                 if (!stat(path, &st)) {
  53                         /* path exists */
  54                         if (!S_ISDIR(st.st_mode)) {
  55                                 *pos = '/';
  56                                 return -3;
  57                         }
  58                 }
  59                 else if (mkdir(path, 0777)) {
  60                         *pos = '/';
  61                         return -1;
  62                 }
  63                 else if (adjust_shared_perm(path)) {
  64                         *pos = '/';
  65                         return -2;
  66                 }
  67                 *pos++ = '/';
  68         }
  69         return 0;
  70 }
  71
  72 int safe_create_leading_directories_const(const char *path)
  73 {
  74         /* path points to cache entries, so xstrdup before messing with it */
  75         char *buf = xstrdup(path);
  76         int result = safe_create_leading_directories(buf);
  77         free(buf);
  78         return result;
  79 }
  80
  81 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
  82 {
  83         int i;
  84         for (i = 0; i < 20; i++) {
  85                 static char hex[] = "0123456789abcdef";
  86                 unsigned int val = sha1[i];
  87                 char *pos = pathbuf + i*2 + (i > 0);
  88                 *pos++ = hex[val >> 4];
  89                 *pos = hex[val & 0xf];
  90         }
  91 }
  92
  93 /*
  94  * NOTE! This returns a statically allocated buffer, so you have to be
  95  * careful about using it. Do an "xstrdup()" if you need to save the
  96  * filename.
  97  *
  98  * Also note that this returns the location for creating.  Reading
  99  * SHA1 file can happen from any alternate directory listed in the
 100  * DB_ENVIRONMENT environment variable if it is not found in
 101  * the primary object database.
 102  */
 103 char *sha1_file_name(const unsigned char *sha1)
 104 {
 105         static char *name, *base;
 106
 107         if (!base) {
 108                 const char *sha1_file_directory = get_object_directory();
 109                 int len = strlen(sha1_file_directory);
 110                 base = xmalloc(len + 60);
 111                 memcpy(base, sha1_file_directory, len);
 112                 memset(base+len, 0, 60);
 113                 base[len] = '/';
 114                 base[len+3] = '/';
 115                 name = base + len + 1;
 116         }
 117         fill_sha1_path(name, sha1);
 118         return base;
 119 }
 120
 121 static char *sha1_get_pack_name(const unsigned char *sha1,
 122                                 char **name, char **base, const char *which)
 123 {
 124         static const char hex[] = "0123456789abcdef";
 125         char *buf;
 126         int i;
 127
 128         if (!*base) {
 129                 const char *sha1_file_directory = get_object_directory();
 130                 int len = strlen(sha1_file_directory);
 131                 *base = xmalloc(len + 60);
 132                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 133                         sha1_file_directory, which);
 134                 *name = *base + len + 11;
 135         }
 136
 137         buf = *name;
 138
 139         for (i = 0; i < 20; i++) {
 140                 unsigned int val = *sha1++;
 141                 *buf++ = hex[val >> 4];
 142                 *buf++ = hex[val & 0xf];
 143         }
 144
 145         return *base;
 146 }
 147
 148 char *sha1_pack_name(const unsigned char *sha1)
 149 {
 150         static char *name, *base;
 151
 152         return sha1_get_pack_name(sha1, &name, &base, "pack");
 153 }
 154
 155 char *sha1_pack_index_name(const unsigned char *sha1)
 156 {
 157         static char *name, *base;
 158
 159         return sha1_get_pack_name(sha1, &name, &base, "idx");
 160 }
 161
 162 struct alternate_object_database *alt_odb_list;
 163 static struct alternate_object_database **alt_odb_tail;
 164
 165 static void read_info_alternates(const char * alternates, int depth);
 166
 167 /*
 168  * Prepare alternate object database registry.
 169  *
 170  * The variable alt_odb_list points at the list of struct
 171  * alternate_object_database.  The elements on this list come from
 172  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 173  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 174  * whose contents is similar to that environment variable but can be
 175  * LF separated.  Its base points at a statically allocated buffer that
 176  * contains "/the/directory/corresponding/to/.git/objects/...", while
 177  * its name points just after the slash at the end of ".git/objects/"
 178  * in the example above, and has enough space to hold 40-byte hex
 179  * SHA1, an extra slash for the first level indirection, and the
 180  * terminating NUL.
 181  */
 182 static int link_alt_odb_entry(const char * entry, int len, const char * relative_base, int depth)
 183 {
 184         const char *objdir = get_object_directory();
 185         struct alternate_object_database *ent;
 186         struct alternate_object_database *alt;
 187         /* 43 = 40-byte + 2 '/' + terminating NUL */
 188         int pfxlen = len;
 189         int entlen = pfxlen + 43;
 190         int base_len = -1;
 191
 192         if (!is_absolute_path(entry) && relative_base) {
 193                 /* Relative alt-odb */
 194                 if (base_len < 0)
 195                         base_len = strlen(relative_base) + 1;
 196                 entlen += base_len;
 197                 pfxlen += base_len;
 198         }
 199         ent = xmalloc(sizeof(*ent) + entlen);
 200
 201         if (!is_absolute_path(entry) && relative_base) {
 202                 memcpy(ent->base, relative_base, base_len - 1);
 203                 ent->base[base_len - 1] = '/';
 204                 memcpy(ent->base + base_len, entry, len);
 205         }
 206         else
 207                 memcpy(ent->base, entry, pfxlen);
 208
 209         ent->name = ent->base + pfxlen + 1;
 210         ent->base[pfxlen + 3] = '/';
 211         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 212
 213         /* Detect cases where alternate disappeared */
 214         if (!is_directory(ent->base)) {
 215                 error("object directory %s does not exist; "
 216                       "check .git/objects/info/alternates.",
 217                       ent->base);
 218                 free(ent);
 219                 return -1;
 220         }
 221
 222         /* Prevent the common mistake of listing the same
 223          * thing twice, or object directory itself.
 224          */
 225         for (alt = alt_odb_list; alt; alt = alt->next) {
 226                 if (!memcmp(ent->base, alt->base, pfxlen)) {
 227                         free(ent);
 228                         return -1;
 229                 }
 230         }
 231         if (!memcmp(ent->base, objdir, pfxlen)) {
 232                 free(ent);
 233                 return -1;
 234         }
 235
 236         /* add the alternate entry */
 237         *alt_odb_tail = ent;
 238         alt_odb_tail = &(ent->next);
 239         ent->next = NULL;
 240
 241         /* recursively add alternates */
 242         read_info_alternates(ent->base, depth + 1);
 243
 244         ent->base[pfxlen] = '/';
 245
 246         return 0;
 247 }
 248
 249 static void link_alt_odb_entries(const char *alt, const char *ep, int sep,
 250                                  const char *relative_base, int depth)
 251 {
 252         const char *cp, *last;
 253
 254         if (depth > 5) {
 255                 error("%s: ignoring alternate object stores, nesting too deep.",
 256                                 relative_base);
 257                 return;
 258         }
 259
 260         last = alt;
 261         while (last < ep) {
 262                 cp = last;
 263                 if (cp < ep && *cp == '#') {
 264                         while (cp < ep && *cp != sep)
 265                                 cp++;
 266                         last = cp + 1;
 267                         continue;
 268                 }
 269                 while (cp < ep && *cp != sep)
 270                         cp++;
 271                 if (last != cp) {
 272                         if (!is_absolute_path(last) && depth) {
 273                                 error("%s: ignoring relative alternate object store %s",
 274                                                 relative_base, last);
 275                         } else {
 276                                 link_alt_odb_entry(last, cp - last,
 277                                                 relative_base, depth);
 278                         }
 279                 }
 280                 while (cp < ep && *cp == sep)
 281                         cp++;
 282                 last = cp;
 283         }
 284 }
 285
 286 static void read_info_alternates(const char * relative_base, int depth)
 287 {
 288         char *map;
 289         size_t mapsz;
 290         struct stat st;
 291         const char alt_file_name[] = "info/alternates";
 292         /* Given that relative_base is no longer than PATH_MAX,
 293            ensure that "path" has enough space to append "/", the
 294            file name, "info/alternates", and a trailing NUL.  */
 295         char path[PATH_MAX + 1 + sizeof alt_file_name];
 296         int fd;
 297
 298         sprintf(path, "%s/%s", relative_base, alt_file_name);
 299         fd = open(path, O_RDONLY);
 300         if (fd < 0)
 301                 return;
 302         if (fstat(fd, &st) || (st.st_size == 0)) {
 303                 close(fd);
 304                 return;
 305         }
 306         mapsz = xsize_t(st.st_size);
 307         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 308         close(fd);
 309
 310         link_alt_odb_entries(map, map + mapsz, '\n', relative_base, depth);
 311
 312         munmap(map, mapsz);
 313 }
 314
 315 void add_to_alternates_file(const char *reference)
 316 {
 317         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 318         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
 319         char *alt = mkpath("%s/objects\n", reference);
 320         write_or_die(fd, alt, strlen(alt));
 321         if (commit_lock_file(lock))
 322                 die("could not close alternates file");
 323         if (alt_odb_tail)
 324                 link_alt_odb_entries(alt, alt + strlen(alt), '\n', NULL, 0);
 325 }
 326
 327 void foreach_alt_odb(alt_odb_fn fn, void *cb)
 328 {
 329         struct alternate_object_database *ent;
 330
 331         prepare_alt_odb();
 332         for (ent = alt_odb_list; ent; ent = ent->next)
 333                 if (fn(ent, cb))
 334                         return;
 335 }
 336
 337 void prepare_alt_odb(void)
 338 {
 339         const char *alt;
 340
 341         if (alt_odb_tail)
 342                 return;
 343
 344         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 345         if (!alt) alt = "";
 346
 347         alt_odb_tail = &alt_odb_list;
 348         link_alt_odb_entries(alt, alt + strlen(alt), PATH_SEP, NULL, 0);
 349
 350         read_info_alternates(get_object_directory(), 0);
 351 }
 352
 353 static int has_loose_object_local(const unsigned char *sha1)
 354 {
 355         char *name = sha1_file_name(sha1);
 356         return !access(name, F_OK);
 357 }
 358
 359 int has_loose_object_nonlocal(const unsigned char *sha1)
 360 {
 361         struct alternate_object_database *alt;
 362         prepare_alt_odb();
 363         for (alt = alt_odb_list; alt; alt = alt->next) {
 364                 fill_sha1_path(alt->name, sha1);
 365                 if (!access(alt->base, F_OK))
 366                         return 1;
 367         }
 368         return 0;
 369 }
 370
 371 static int has_loose_object(const unsigned char *sha1)
 372 {
 373         return has_loose_object_local(sha1) ||
 374                has_loose_object_nonlocal(sha1);
 375 }
 376
 377 static unsigned int pack_used_ctr;
 378 static unsigned int pack_mmap_calls;
 379 static unsigned int peak_pack_open_windows;
 380 static unsigned int pack_open_windows;
 381 static size_t peak_pack_mapped;
 382 static size_t pack_mapped;
 383 struct packed_git *packed_git;
 384
 385 void pack_report(void)
 386 {
 387         fprintf(stderr,
 388                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 389                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 390                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 391                 sz_fmt(getpagesize()),
 392                 sz_fmt(packed_git_window_size),
 393                 sz_fmt(packed_git_limit));
 394         fprintf(stderr,
 395                 "pack_report: pack_used_ctr            = %10u\n"
 396                 "pack_report: pack_mmap_calls          = %10u\n"
 397                 "pack_report: pack_open_windows        = %10u / %10u\n"
 398                 "pack_report: pack_mapped              = "
 399                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 400                 pack_used_ctr,
 401                 pack_mmap_calls,
 402                 pack_open_windows, peak_pack_open_windows,
 403                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 404 }
 405
 406 static int check_packed_git_idx(const char *path,  struct packed_git *p)
 407 {
 408         void *idx_map;
 409         struct pack_idx_header *hdr;
 410         size_t idx_size;
 411         uint32_t version, nr, i, *index;
 412         int fd = open(path, O_RDONLY);
 413         struct stat st;
 414
 415         if (fd < 0)
 416                 return -1;
 417         if (fstat(fd, &st)) {
 418                 close(fd);
 419                 return -1;
 420         }
 421         idx_size = xsize_t(st.st_size);
 422         if (idx_size < 4 * 256 + 20 + 20) {
 423                 close(fd);
 424                 return error("index file %s is too small", path);
 425         }
 426         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 427         close(fd);
 428
 429         hdr = idx_map;
 430         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 431                 version = ntohl(hdr->idx_version);
 432                 if (version < 2 || version > 2) {
 433                         munmap(idx_map, idx_size);
 434                         return error("index file %s is version %"PRIu32
 435                                      " and is not supported by this binary"
 436                                      " (try upgrading GIT to a newer version)",
 437                                      path, version);
 438                 }
 439         } else
 440                 version = 1;
 441
 442         nr = 0;
 443         index = idx_map;
 444         if (version > 1)
 445                 index += 2;  /* skip index header */
 446         for (i = 0; i < 256; i++) {
 447                 uint32_t n = ntohl(index[i]);
 448                 if (n < nr) {
 449                         munmap(idx_map, idx_size);
 450                         return error("non-monotonic index %s", path);
 451                 }
 452                 nr = n;
 453         }
 454
 455         if (version == 1) {
 456                 /*
 457                  * Total size:
 458                  *  - 256 index entries 4 bytes each
 459                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 460                  *  - 20-byte SHA1 of the packfile
 461                  *  - 20-byte SHA1 file checksum
 462                  */
 463                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 464                         munmap(idx_map, idx_size);
 465                         return error("wrong index v1 file size in %s", path);
 466                 }
 467         } else if (version == 2) {
 468                 /*
 469                  * Minimum size:
 470                  *  - 8 bytes of header
 471                  *  - 256 index entries 4 bytes each
 472                  *  - 20-byte sha1 entry * nr
 473                  *  - 4-byte crc entry * nr
 474                  *  - 4-byte offset entry * nr
 475                  *  - 20-byte SHA1 of the packfile
 476                  *  - 20-byte SHA1 file checksum
 477                  * And after the 4-byte offset table might be a
 478                  * variable sized table containing 8-byte entries
 479                  * for offsets larger than 2^31.
 480                  */
 481                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 482                 unsigned long max_size = min_size;
 483                 if (nr)
 484                         max_size += (nr - 1)*8;
 485                 if (idx_size < min_size || idx_size > max_size) {
 486                         munmap(idx_map, idx_size);
 487                         return error("wrong index v2 file size in %s", path);
 488                 }
 489                 if (idx_size != min_size &&
 490                     /*
 491                      * make sure we can deal with large pack offsets.
 492                      * 31-bit signed offset won't be enough, neither
 493                      * 32-bit unsigned one will be.
 494                      */
 495                     (sizeof(off_t) <= 4)) {
 496                         munmap(idx_map, idx_size);
 497                         return error("pack too large for current definition of off_t in %s", path);
 498                 }
 499         }
 500
 501         p->index_version = version;
 502         p->index_data = idx_map;
 503         p->index_size = idx_size;
 504         p->num_objects = nr;
 505         return 0;
 506 }
 507
 508 int open_pack_index(struct packed_git *p)
 509 {
 510         char *idx_name;
 511         int ret;
 512
 513         if (p->index_data)
 514                 return 0;
 515
 516         idx_name = xstrdup(p->pack_name);
 517         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 518         ret = check_packed_git_idx(idx_name, p);
 519         free(idx_name);
 520         return ret;
 521 }
 522
 523 static void scan_windows(struct packed_git *p,
 524         struct packed_git **lru_p,
 525         struct pack_window **lru_w,
 526         struct pack_window **lru_l)
 527 {
 528         struct pack_window *w, *w_l;
 529
 530         for (w_l = NULL, w = p->windows; w; w = w->next) {
 531                 if (!w->inuse_cnt) {
 532                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 533                                 *lru_p = p;
 534                                 *lru_w = w;
 535                                 *lru_l = w_l;
 536                         }
 537                 }
 538                 w_l = w;
 539         }
 540 }
 541
 542 static int unuse_one_window(struct packed_git *current, int keep_fd)
 543 {
 544         struct packed_git *p, *lru_p = NULL;
 545         struct pack_window *lru_w = NULL, *lru_l = NULL;
 546
 547         if (current)
 548                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 549         for (p = packed_git; p; p = p->next)
 550                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 551         if (lru_p) {
 552                 munmap(lru_w->base, lru_w->len);
 553                 pack_mapped -= lru_w->len;
 554                 if (lru_l)
 555                         lru_l->next = lru_w->next;
 556                 else {
 557                         lru_p->windows = lru_w->next;
 558                         if (!lru_p->windows && lru_p->pack_fd != keep_fd) {
 559                                 close(lru_p->pack_fd);
 560                                 lru_p->pack_fd = -1;
 561                         }
 562                 }
 563                 free(lru_w);
 564                 pack_open_windows--;
 565                 return 1;
 566         }
 567         return 0;
 568 }
 569
 570 void release_pack_memory(size_t need, int fd)
 571 {
 572         size_t cur = pack_mapped;
 573         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
 574                 ; /* nothing */
 575 }
 576
 577 void close_pack_windows(struct packed_git *p)
 578 {
 579         while (p->windows) {
 580                 struct pack_window *w = p->windows;
 581
 582                 if (w->inuse_cnt)
 583                         die("pack '%s' still has open windows to it",
 584                             p->pack_name);
 585                 munmap(w->base, w->len);
 586                 pack_mapped -= w->len;
 587                 pack_open_windows--;
 588                 p->windows = w->next;
 589                 free(w);
 590         }
 591 }
 592
 593 void unuse_pack(struct pack_window **w_cursor)
 594 {
 595         struct pack_window *w = *w_cursor;
 596         if (w) {
 597                 w->inuse_cnt--;
 598                 *w_cursor = NULL;
 599         }
 600 }
 601
 602 void close_pack_index(struct packed_git *p)
 603 {
 604         if (p->index_data) {
 605                 munmap((void *)p->index_data, p->index_size);
 606                 p->index_data = NULL;
 607         }
 608 }
 609
 610 /*
 611  * This is used by git-repack in case a newly created pack happens to
 612  * contain the same set of objects as an existing one.  In that case
 613  * the resulting file might be different even if its name would be the
 614  * same.  It is best to close any reference to the old pack before it is
 615  * replaced on disk.  Of course no index pointers nor windows for given pack
 616  * must subsist at this point.  If ever objects from this pack are requested
 617  * again, the new version of the pack will be reinitialized through
 618  * reprepare_packed_git().
 619  */
 620 void free_pack_by_name(const char *pack_name)
 621 {
 622         struct packed_git *p, **pp = &packed_git;
 623
 624         while (*pp) {
 625                 p = *pp;
 626                 if (strcmp(pack_name, p->pack_name) == 0) {
 627                         clear_delta_base_cache();
 628                         close_pack_windows(p);
 629                         if (p->pack_fd != -1)
 630                                 close(p->pack_fd);
 631                         close_pack_index(p);
 632                         free(p->bad_object_sha1);
 633                         *pp = p->next;
 634                         free(p);
 635                         return;
 636                 }
 637                 pp = &p->next;
 638         }
 639 }
 640
 641 /*
 642  * Do not call this directly as this leaks p->pack_fd on error return;
 643  * call open_packed_git() instead.
 644  */
 645 static int open_packed_git_1(struct packed_git *p)
 646 {
 647         struct stat st;
 648         struct pack_header hdr;
 649         unsigned char sha1[20];
 650         unsigned char *idx_sha1;
 651         long fd_flag;
 652
 653         if (!p->index_data && open_pack_index(p))
 654                 return error("packfile %s index unavailable", p->pack_name);
 655
 656         p->pack_fd = open(p->pack_name, O_RDONLY);
 657         while (p->pack_fd < 0 && errno == EMFILE && unuse_one_window(p, -1))
 658                 p->pack_fd = open(p->pack_name, O_RDONLY);
 659         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 660                 return -1;
 661
 662         /* If we created the struct before we had the pack we lack size. */
 663         if (!p->pack_size) {
 664                 if (!S_ISREG(st.st_mode))
 665                         return error("packfile %s not a regular file", p->pack_name);
 666                 p->pack_size = st.st_size;
 667         } else if (p->pack_size != st.st_size)
 668                 return error("packfile %s size changed", p->pack_name);
 669
 670         /* We leave these file descriptors open with sliding mmap;
 671          * there is no point keeping them open across exec(), though.
 672          */
 673         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 674         if (fd_flag < 0)
 675                 return error("cannot determine file descriptor flags");
 676         fd_flag |= FD_CLOEXEC;
 677         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 678                 return error("cannot set FD_CLOEXEC");
 679
 680         /* Verify we recognize this pack file format. */
 681         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 682                 return error("file %s is far too short to be a packfile", p->pack_name);
 683         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 684                 return error("file %s is not a GIT packfile", p->pack_name);
 685         if (!pack_version_ok(hdr.hdr_version))
 686                 return error("packfile %s is version %"PRIu32" and not"
 687                         " supported (try upgrading GIT to a newer version)",
 688                         p->pack_name, ntohl(hdr.hdr_version));
 689
 690         /* Verify the pack matches its index. */
 691         if (p->num_objects != ntohl(hdr.hdr_entries))
 692                 return error("packfile %s claims to have %"PRIu32" objects"
 693                              " while index indicates %"PRIu32" objects",
 694                              p->pack_name, ntohl(hdr.hdr_entries),
 695                              p->num_objects);
 696         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
 697                 return error("end of packfile %s is unavailable", p->pack_name);
 698         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
 699                 return error("packfile %s signature is unavailable", p->pack_name);
 700         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
 701         if (hashcmp(sha1, idx_sha1))
 702                 return error("packfile %s does not match index", p->pack_name);
 703         return 0;
 704 }
 705
 706 static int open_packed_git(struct packed_git *p)
 707 {
 708         if (!open_packed_git_1(p))
 709                 return 0;
 710         if (p->pack_fd != -1) {
 711                 close(p->pack_fd);
 712                 p->pack_fd = -1;
 713         }
 714         return -1;
 715 }
 716
 717 static int in_window(struct pack_window *win, off_t offset)
 718 {
 719         /* We must promise at least 20 bytes (one hash) after the
 720          * offset is available from this window, otherwise the offset
 721          * is not actually in this window and a different window (which
 722          * has that one hash excess) must be used.  This is to support
 723          * the object header and delta base parsing routines below.
 724          */
 725         off_t win_off = win->offset;
 726         return win_off <= offset
 727                 && (offset + 20) <= (win_off + win->len);
 728 }
 729
 730 unsigned char *use_pack(struct packed_git *p,
 731                 struct pack_window **w_cursor,
 732                 off_t offset,
 733                 unsigned int *left)
 734 {
 735         struct pack_window *win = *w_cursor;
 736
 737         if (p->pack_fd == -1 && open_packed_git(p))
 738                 die("packfile %s cannot be accessed", p->pack_name);
 739
 740         /* Since packfiles end in a hash of their content and it's
 741          * pointless to ask for an offset into the middle of that
 742          * hash, and the in_window function above wouldn't match
 743          * don't allow an offset too close to the end of the file.
 744          */
 745         if (offset > (p->pack_size - 20))
 746                 die("offset beyond end of packfile (truncated pack?)");
 747
 748         if (!win || !in_window(win, offset)) {
 749                 if (win)
 750                         win->inuse_cnt--;
 751                 for (win = p->windows; win; win = win->next) {
 752                         if (in_window(win, offset))
 753                                 break;
 754                 }
 755                 if (!win) {
 756                         size_t window_align = packed_git_window_size / 2;
 757                         off_t len;
 758                         win = xcalloc(1, sizeof(*win));
 759                         win->offset = (offset / window_align) * window_align;
 760                         len = p->pack_size - win->offset;
 761                         if (len > packed_git_window_size)
 762                                 len = packed_git_window_size;
 763                         win->len = (size_t)len;
 764                         pack_mapped += win->len;
 765                         while (packed_git_limit < pack_mapped
 766                                 && unuse_one_window(p, p->pack_fd))
 767                                 ; /* nothing */
 768                         win->base = xmmap(NULL, win->len,
 769                                 PROT_READ, MAP_PRIVATE,
 770                                 p->pack_fd, win->offset);
 771                         if (win->base == MAP_FAILED)
 772                                 die("packfile %s cannot be mapped: %s",
 773                                         p->pack_name,
 774                                         strerror(errno));
 775                         pack_mmap_calls++;
 776                         pack_open_windows++;
 777                         if (pack_mapped > peak_pack_mapped)
 778                                 peak_pack_mapped = pack_mapped;
 779                         if (pack_open_windows > peak_pack_open_windows)
 780                                 peak_pack_open_windows = pack_open_windows;
 781                         win->next = p->windows;
 782                         p->windows = win;
 783                 }
 784         }
 785         if (win != *w_cursor) {
 786                 win->last_used = pack_used_ctr++;
 787                 win->inuse_cnt++;
 788                 *w_cursor = win;
 789         }
 790         offset -= win->offset;
 791         if (left)
 792                 *left = win->len - xsize_t(offset);
 793         return win->base + offset;
 794 }
 795
 796 static struct packed_git *alloc_packed_git(int extra)
 797 {
 798         struct packed_git *p = xmalloc(sizeof(*p) + extra);
 799         memset(p, 0, sizeof(*p));
 800         p->pack_fd = -1;
 801         return p;
 802 }
 803
 804 struct packed_git *add_packed_git(const char *path, int path_len, int local)
 805 {
 806         struct stat st;
 807         struct packed_git *p = alloc_packed_git(path_len + 2);
 808
 809         /*
 810          * Make sure a corresponding .pack file exists and that
 811          * the index looks sane.
 812          */
 813         path_len -= strlen(".idx");
 814         if (path_len < 1) {
 815                 free(p);
 816                 return NULL;
 817         }
 818         memcpy(p->pack_name, path, path_len);
 819
 820         strcpy(p->pack_name + path_len, ".keep");
 821         if (!access(p->pack_name, F_OK))
 822                 p->pack_keep = 1;
 823
 824         strcpy(p->pack_name + path_len, ".pack");
 825         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 826                 free(p);
 827                 return NULL;
 828         }
 829
 830         /* ok, it looks sane as far as we can check without
 831          * actually mapping the pack file.
 832          */
 833         p->pack_size = st.st_size;
 834         p->pack_local = local;
 835         p->mtime = st.st_mtime;
 836         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
 837                 hashclr(p->sha1);
 838         return p;
 839 }
 840
 841 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
 842 {
 843         const char *path = sha1_pack_name(sha1);
 844         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
 845
 846         strcpy(p->pack_name, path);
 847         hashcpy(p->sha1, sha1);
 848         if (check_packed_git_idx(idx_path, p)) {
 849                 free(p);
 850                 return NULL;
 851         }
 852
 853         return p;
 854 }
 855
 856 void install_packed_git(struct packed_git *pack)
 857 {
 858         pack->next = packed_git;
 859         packed_git = pack;
 860 }
 861
 862 static void prepare_packed_git_one(char *objdir, int local)
 863 {
 864         /* Ensure that this buffer is large enough so that we can
 865            append "/pack/" without clobbering the stack even if
 866            strlen(objdir) were PATH_MAX.  */
 867         char path[PATH_MAX + 1 + 4 + 1 + 1];
 868         int len;
 869         DIR *dir;
 870         struct dirent *de;
 871
 872         sprintf(path, "%s/pack", objdir);
 873         len = strlen(path);
 874         dir = opendir(path);
 875         while (!dir && errno == EMFILE && unuse_one_window(packed_git, -1))
 876                 dir = opendir(path);
 877         if (!dir) {
 878                 if (errno != ENOENT)
 879                         error("unable to open object pack directory: %s: %s",
 880                               path, strerror(errno));
 881                 return;
 882         }
 883         path[len++] = '/';
 884         while ((de = readdir(dir)) != NULL) {
 885                 int namelen = strlen(de->d_name);
 886                 struct packed_git *p;
 887
 888                 if (!has_extension(de->d_name, ".idx"))
 889                         continue;
 890
 891                 if (len + namelen + 1 > sizeof(path))
 892                         continue;
 893
 894                 /* Don't reopen a pack we already have. */
 895                 strcpy(path + len, de->d_name);
 896                 for (p = packed_git; p; p = p->next) {
 897                         if (!memcmp(path, p->pack_name, len + namelen - 4))
 898                                 break;
 899                 }
 900                 if (p)
 901                         continue;
 902                 /* See if it really is a valid .idx file with corresponding
 903                  * .pack file that we can map.
 904                  */
 905                 p = add_packed_git(path, len + namelen, local);
 906                 if (!p)
 907                         continue;
 908                 install_packed_git(p);
 909         }
 910         closedir(dir);
 911 }
 912
 913 static int sort_pack(const void *a_, const void *b_)
 914 {
 915         struct packed_git *a = *((struct packed_git **)a_);
 916         struct packed_git *b = *((struct packed_git **)b_);
 917         int st;
 918
 919         /*
 920          * Local packs tend to contain objects specific to our
 921          * variant of the project than remote ones.  In addition,
 922          * remote ones could be on a network mounted filesystem.
 923          * Favor local ones for these reasons.
 924          */
 925         st = a->pack_local - b->pack_local;
 926         if (st)
 927                 return -st;
 928
 929         /*
 930          * Younger packs tend to contain more recent objects,
 931          * and more recent objects tend to get accessed more
 932          * often.
 933          */
 934         if (a->mtime < b->mtime)
 935                 return 1;
 936         else if (a->mtime == b->mtime)
 937                 return 0;
 938         return -1;
 939 }
 940
 941 static void rearrange_packed_git(void)
 942 {
 943         struct packed_git **ary, *p;
 944         int i, n;
 945
 946         for (n = 0, p = packed_git; p; p = p->next)
 947                 n++;
 948         if (n < 2)
 949                 return;
 950
 951         /* prepare an array of packed_git for easier sorting */
 952         ary = xcalloc(n, sizeof(struct packed_git *));
 953         for (n = 0, p = packed_git; p; p = p->next)
 954                 ary[n++] = p;
 955
 956         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
 957
 958         /* link them back again */
 959         for (i = 0; i < n - 1; i++)
 960                 ary[i]->next = ary[i + 1];
 961         ary[n - 1]->next = NULL;
 962         packed_git = ary[0];
 963
 964         free(ary);
 965 }
 966
 967 static int prepare_packed_git_run_once = 0;
 968 void prepare_packed_git(void)
 969 {
 970         struct alternate_object_database *alt;
 971
 972         if (prepare_packed_git_run_once)
 973                 return;
 974         prepare_packed_git_one(get_object_directory(), 1);
 975         prepare_alt_odb();
 976         for (alt = alt_odb_list; alt; alt = alt->next) {
 977                 alt->name[-1] = 0;
 978                 prepare_packed_git_one(alt->base, 0);
 979                 alt->name[-1] = '/';
 980         }
 981         rearrange_packed_git();
 982         prepare_packed_git_run_once = 1;
 983 }
 984
 985 void reprepare_packed_git(void)
 986 {
 987         discard_revindex();
 988         prepare_packed_git_run_once = 0;
 989         prepare_packed_git();
 990 }
 991
 992 static void mark_bad_packed_object(struct packed_git *p,
 993                                    const unsigned char *sha1)
 994 {
 995         unsigned i;
 996         for (i = 0; i < p->num_bad_objects; i++)
 997                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
 998                         return;
 999         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1000         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1001         p->num_bad_objects++;
1002 }
1003
1004 static int has_packed_and_bad(const unsigned char *sha1)
1005 {
1006         struct packed_git *p;
1007         unsigned i;
1008
1009         for (p = packed_git; p; p = p->next)
1010                 for (i = 0; i < p->num_bad_objects; i++)
1011                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1012                                 return 1;
1013         return 0;
1014 }
1015
1016 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1017 {
1018         unsigned char real_sha1[20];
1019         hash_sha1_file(map, size, type, real_sha1);
1020         return hashcmp(sha1, real_sha1) ? -1 : 0;
1021 }
1022
1023 static int git_open_noatime(const char *name)
1024 {
1025         static int sha1_file_open_flag = O_NOATIME;
1026         int fd = open(name, O_RDONLY | sha1_file_open_flag);
1027
1028         /* Might the failure be due to O_NOATIME? */
1029         if (fd < 0 && errno != ENOENT && sha1_file_open_flag) {
1030                 fd = open(name, O_RDONLY);
1031                 if (fd >= 0)
1032                         sha1_file_open_flag = 0;
1033         }
1034         return fd;
1035 }
1036
1037 static int open_sha1_file(const unsigned char *sha1)
1038 {
1039         int fd;
1040         char *name = sha1_file_name(sha1);
1041         struct alternate_object_database *alt;
1042
1043         fd = git_open_noatime(name);
1044         if (fd >= 0)
1045                 return fd;
1046
1047         prepare_alt_odb();
1048         errno = ENOENT;
1049         for (alt = alt_odb_list; alt; alt = alt->next) {
1050                 name = alt->name;
1051                 fill_sha1_path(name, sha1);
1052                 fd = git_open_noatime(alt->base);
1053                 if (fd >= 0)
1054                         return fd;
1055         }
1056         return -1;
1057 }
1058
1059 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1060 {
1061         void *map;
1062         int fd;
1063
1064         fd = open_sha1_file(sha1);
1065         map = NULL;
1066         if (fd >= 0) {
1067                 struct stat st;
1068
1069                 if (!fstat(fd, &st)) {
1070                         *size = xsize_t(st.st_size);
1071                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1072                 }
1073                 close(fd);
1074         }
1075         return map;
1076 }
1077
1078 static int legacy_loose_object(unsigned char *map)
1079 {
1080         unsigned int word;
1081
1082         /*
1083          * Is it a zlib-compressed buffer? If so, the first byte
1084          * must be 0x78 (15-bit window size, deflated), and the
1085          * first 16-bit word is evenly divisible by 31
1086          */
1087         word = (map[0] << 8) + map[1];
1088         if (map[0] == 0x78 && !(word % 31))
1089                 return 1;
1090         else
1091                 return 0;
1092 }
1093
1094 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1095                 unsigned long len, enum object_type *type, unsigned long *sizep)
1096 {
1097         unsigned shift;
1098         unsigned long size, c;
1099         unsigned long used = 0;
1100
1101         c = buf[used++];
1102         *type = (c >> 4) & 7;
1103         size = c & 15;
1104         shift = 4;
1105         while (c & 0x80) {
1106                 if (len <= used || bitsizeof(long) <= shift) {
1107                         error("bad object header");
1108                         return 0;
1109                 }
1110                 c = buf[used++];
1111                 size += (c & 0x7f) << shift;
1112                 shift += 7;
1113         }
1114         *sizep = size;
1115         return used;
1116 }
1117
1118 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1119 {
1120         unsigned long size, used;
1121         static const char valid_loose_object_type[8] = {
1122                 0, /* OBJ_EXT */
1123                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1124                 0, /* "delta" and others are invalid in a loose object */
1125         };
1126         enum object_type type;
1127
1128         /* Get the data stream */
1129         memset(stream, 0, sizeof(*stream));
1130         stream->next_in = map;
1131         stream->avail_in = mapsize;
1132         stream->next_out = buffer;
1133         stream->avail_out = bufsiz;
1134
1135         if (legacy_loose_object(map)) {
1136                 git_inflate_init(stream);
1137                 return git_inflate(stream, 0);
1138         }
1139
1140
1141         /*
1142          * There used to be a second loose object header format which
1143          * was meant to mimic the in-pack format, allowing for direct
1144          * copy of the object data.  This format turned up not to be
1145          * really worth it and we don't write it any longer.  But we
1146          * can still read it.
1147          */
1148         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1149         if (!used || !valid_loose_object_type[type])
1150                 return -1;
1151         map += used;
1152         mapsize -= used;
1153
1154         /* Set up the stream for the rest.. */
1155         stream->next_in = map;
1156         stream->avail_in = mapsize;
1157         git_inflate_init(stream);
1158
1159         /* And generate the fake traditional header */
1160         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1161                                          typename(type), size);
1162         return 0;
1163 }
1164
1165 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1166 {
1167         int bytes = strlen(buffer) + 1;
1168         unsigned char *buf = xmallocz(size);
1169         unsigned long n;
1170         int status = Z_OK;
1171
1172         n = stream->total_out - bytes;
1173         if (n > size)
1174                 n = size;
1175         memcpy(buf, (char *) buffer + bytes, n);
1176         bytes = n;
1177         if (bytes <= size) {
1178                 /*
1179                  * The above condition must be (bytes <= size), not
1180                  * (bytes < size).  In other words, even though we
1181                  * expect no more output and set avail_out to zer0,
1182                  * the input zlib stream may have bytes that express
1183                  * "this concludes the stream", and we *do* want to
1184                  * eat that input.
1185                  *
1186                  * Otherwise we would not be able to test that we
1187                  * consumed all the input to reach the expected size;
1188                  * we also want to check that zlib tells us that all
1189                  * went well with status == Z_STREAM_END at the end.
1190                  */
1191                 stream->next_out = buf + bytes;
1192                 stream->avail_out = size - bytes;
1193                 while (status == Z_OK)
1194                         status = git_inflate(stream, Z_FINISH);
1195         }
1196         if (status == Z_STREAM_END && !stream->avail_in) {
1197                 git_inflate_end(stream);
1198                 return buf;
1199         }
1200
1201         if (status < 0)
1202                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1203         else if (stream->avail_in)
1204                 error("garbage at end of loose object '%s'",
1205                       sha1_to_hex(sha1));
1206         free(buf);
1207         return NULL;
1208 }
1209
1210 /*
1211  * We used to just use "sscanf()", but that's actually way
1212  * too permissive for what we want to check. So do an anal
1213  * object header parse by hand.
1214  */
1215 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1216 {
1217         char type[10];
1218         int i;
1219         unsigned long size;
1220
1221         /*
1222          * The type can be at most ten bytes (including the
1223          * terminating '\0' that we add), and is followed by
1224          * a space.
1225          */
1226         i = 0;
1227         for (;;) {
1228                 char c = *hdr++;
1229                 if (c == ' ')
1230                         break;
1231                 type[i++] = c;
1232                 if (i >= sizeof(type))
1233                         return -1;
1234         }
1235         type[i] = 0;
1236
1237         /*
1238          * The length must follow immediately, and be in canonical
1239          * decimal format (ie "010" is not valid).
1240          */
1241         size = *hdr++ - '0';
1242         if (size > 9)
1243                 return -1;
1244         if (size) {
1245                 for (;;) {
1246                         unsigned long c = *hdr - '0';
1247                         if (c > 9)
1248                                 break;
1249                         hdr++;
1250                         size = size * 10 + c;
1251                 }
1252         }
1253         *sizep = size;
1254
1255         /*
1256          * The length must be followed by a zero byte
1257          */
1258         return *hdr ? -1 : type_from_string(type);
1259 }
1260
1261 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1262 {
1263         int ret;
1264         z_stream stream;
1265         char hdr[8192];
1266
1267         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1268         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1269                 return NULL;
1270
1271         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1272 }
1273
1274 unsigned long get_size_from_delta(struct packed_git *p,
1275                                   struct pack_window **w_curs,
1276                                   off_t curpos)
1277 {
1278         const unsigned char *data;
1279         unsigned char delta_head[20], *in;
1280         z_stream stream;
1281         int st;
1282
1283         memset(&stream, 0, sizeof(stream));
1284         stream.next_out = delta_head;
1285         stream.avail_out = sizeof(delta_head);
1286
1287         git_inflate_init(&stream);
1288         do {
1289                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1290                 stream.next_in = in;
1291                 st = git_inflate(&stream, Z_FINISH);
1292                 curpos += stream.next_in - in;
1293         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1294                  stream.total_out < sizeof(delta_head));
1295         git_inflate_end(&stream);
1296         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1297                 error("delta data unpack-initial failed");
1298                 return 0;
1299         }
1300
1301         /* Examine the initial part of the delta to figure out
1302          * the result size.
1303          */
1304         data = delta_head;
1305
1306         /* ignore base size */
1307         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1308
1309         /* Read the result size */
1310         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1311 }
1312
1313 static off_t get_delta_base(struct packed_git *p,
1314                                     struct pack_window **w_curs,
1315                                     off_t *curpos,
1316                                     enum object_type type,
1317                                     off_t delta_obj_offset)
1318 {
1319         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1320         off_t base_offset;
1321
1322         /* use_pack() assured us we have [base_info, base_info + 20)
1323          * as a range that we can look at without walking off the
1324          * end of the mapped window.  Its actually the hash size
1325          * that is assured.  An OFS_DELTA longer than the hash size
1326          * is stupid, as then a REF_DELTA would be smaller to store.
1327          */
1328         if (type == OBJ_OFS_DELTA) {
1329                 unsigned used = 0;
1330                 unsigned char c = base_info[used++];
1331                 base_offset = c & 127;
1332                 while (c & 128) {
1333                         base_offset += 1;
1334                         if (!base_offset || MSB(base_offset, 7))
1335                                 return 0;  /* overflow */
1336                         c = base_info[used++];
1337                         base_offset = (base_offset << 7) + (c & 127);
1338                 }
1339                 base_offset = delta_obj_offset - base_offset;
1340                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1341                         return 0;  /* out of bound */
1342                 *curpos += used;
1343         } else if (type == OBJ_REF_DELTA) {
1344                 /* The base entry _must_ be in the same pack */
1345                 base_offset = find_pack_entry_one(base_info, p);
1346                 *curpos += 20;
1347         } else
1348                 die("I am totally screwed");
1349         return base_offset;
1350 }
1351
1352 /* forward declaration for a mutually recursive function */
1353 static int packed_object_info(struct packed_git *p, off_t offset,
1354                               unsigned long *sizep);
1355
1356 static int packed_delta_info(struct packed_git *p,
1357                              struct pack_window **w_curs,
1358                              off_t curpos,
1359                              enum object_type type,
1360                              off_t obj_offset,
1361                              unsigned long *sizep)
1362 {
1363         off_t base_offset;
1364
1365         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1366         if (!base_offset)
1367                 return OBJ_BAD;
1368         type = packed_object_info(p, base_offset, NULL);
1369         if (type <= OBJ_NONE) {
1370                 struct revindex_entry *revidx;
1371                 const unsigned char *base_sha1;
1372                 revidx = find_pack_revindex(p, base_offset);
1373                 if (!revidx)
1374                         return OBJ_BAD;
1375                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1376                 mark_bad_packed_object(p, base_sha1);
1377                 type = sha1_object_info(base_sha1, NULL);
1378                 if (type <= OBJ_NONE)
1379                         return OBJ_BAD;
1380         }
1381
1382         /* We choose to only get the type of the base object and
1383          * ignore potentially corrupt pack file that expects the delta
1384          * based on a base with a wrong size.  This saves tons of
1385          * inflate() calls.
1386          */
1387         if (sizep) {
1388                 *sizep = get_size_from_delta(p, w_curs, curpos);
1389                 if (*sizep == 0)
1390                         type = OBJ_BAD;
1391         }
1392
1393         return type;
1394 }
1395
1396 static int unpack_object_header(struct packed_git *p,
1397                                 struct pack_window **w_curs,
1398                                 off_t *curpos,
1399                                 unsigned long *sizep)
1400 {
1401         unsigned char *base;
1402         unsigned int left;
1403         unsigned long used;
1404         enum object_type type;
1405
1406         /* use_pack() assures us we have [base, base + 20) available
1407          * as a range that we can look at at.  (Its actually the hash
1408          * size that is assured.)  With our object header encoding
1409          * the maximum deflated object size is 2^137, which is just
1410          * insane, so we know won't exceed what we have been given.
1411          */
1412         base = use_pack(p, w_curs, *curpos, &left);
1413         used = unpack_object_header_buffer(base, left, &type, sizep);
1414         if (!used) {
1415                 type = OBJ_BAD;
1416         } else
1417                 *curpos += used;
1418
1419         return type;
1420 }
1421
1422 const char *packed_object_info_detail(struct packed_git *p,
1423                                       off_t obj_offset,
1424                                       unsigned long *size,
1425                                       unsigned long *store_size,
1426                                       unsigned int *delta_chain_length,
1427                                       unsigned char *base_sha1)
1428 {
1429         struct pack_window *w_curs = NULL;
1430         off_t curpos;
1431         unsigned long dummy;
1432         unsigned char *next_sha1;
1433         enum object_type type;
1434         struct revindex_entry *revidx;
1435
1436         *delta_chain_length = 0;
1437         curpos = obj_offset;
1438         type = unpack_object_header(p, &w_curs, &curpos, size);
1439
1440         revidx = find_pack_revindex(p, obj_offset);
1441         *store_size = revidx[1].offset - obj_offset;
1442
1443         for (;;) {
1444                 switch (type) {
1445                 default:
1446                         die("pack %s contains unknown object type %d",
1447                             p->pack_name, type);
1448                 case OBJ_COMMIT:
1449                 case OBJ_TREE:
1450                 case OBJ_BLOB:
1451                 case OBJ_TAG:
1452                         unuse_pack(&w_curs);
1453                         return typename(type);
1454                 case OBJ_OFS_DELTA:
1455                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1456                         if (!obj_offset)
1457                                 die("pack %s contains bad delta base reference of type %s",
1458                                     p->pack_name, typename(type));
1459                         if (*delta_chain_length == 0) {
1460                                 revidx = find_pack_revindex(p, obj_offset);
1461                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1462                         }
1463                         break;
1464                 case OBJ_REF_DELTA:
1465                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1466                         if (*delta_chain_length == 0)
1467                                 hashcpy(base_sha1, next_sha1);
1468                         obj_offset = find_pack_entry_one(next_sha1, p);
1469                         break;
1470                 }
1471                 (*delta_chain_length)++;
1472                 curpos = obj_offset;
1473                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1474         }
1475 }
1476
1477 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1478                               unsigned long *sizep)
1479 {
1480         struct pack_window *w_curs = NULL;
1481         unsigned long size;
1482         off_t curpos = obj_offset;
1483         enum object_type type;
1484
1485         type = unpack_object_header(p, &w_curs, &curpos, &size);
1486
1487         switch (type) {
1488         case OBJ_OFS_DELTA:
1489         case OBJ_REF_DELTA:
1490                 type = packed_delta_info(p, &w_curs, curpos,
1491                                          type, obj_offset, sizep);
1492                 break;
1493         case OBJ_COMMIT:
1494         case OBJ_TREE:
1495         case OBJ_BLOB:
1496         case OBJ_TAG:
1497                 if (sizep)
1498                         *sizep = size;
1499                 break;
1500         default:
1501                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1502                       type, (uintmax_t)obj_offset, p->pack_name);
1503                 type = OBJ_BAD;
1504         }
1505         unuse_pack(&w_curs);
1506         return type;
1507 }
1508
1509 static void *unpack_compressed_entry(struct packed_git *p,
1510                                     struct pack_window **w_curs,
1511                                     off_t curpos,
1512                                     unsigned long size)
1513 {
1514         int st;
1515         z_stream stream;
1516         unsigned char *buffer, *in;
1517
1518         buffer = xmallocz(size);
1519         memset(&stream, 0, sizeof(stream));
1520         stream.next_out = buffer;
1521         stream.avail_out = size + 1;
1522
1523         git_inflate_init(&stream);
1524         do {
1525                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1526                 stream.next_in = in;
1527                 st = git_inflate(&stream, Z_FINISH);
1528                 if (!stream.avail_out)
1529                         break; /* the payload is larger than it should be */
1530                 curpos += stream.next_in - in;
1531         } while (st == Z_OK || st == Z_BUF_ERROR);
1532         git_inflate_end(&stream);
1533         if ((st != Z_STREAM_END) || stream.total_out != size) {
1534                 free(buffer);
1535                 return NULL;
1536         }
1537
1538         return buffer;
1539 }
1540
1541 #define MAX_DELTA_CACHE (256)
1542
1543 static size_t delta_base_cached;
1544
1545 static struct delta_base_cache_lru_list {
1546         struct delta_base_cache_lru_list *prev;
1547         struct delta_base_cache_lru_list *next;
1548 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1549
1550 static struct delta_base_cache_entry {
1551         struct delta_base_cache_lru_list lru;
1552         void *data;
1553         struct packed_git *p;
1554         off_t base_offset;
1555         unsigned long size;
1556         enum object_type type;
1557 } delta_base_cache[MAX_DELTA_CACHE];
1558
1559 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1560 {
1561         unsigned long hash;
1562
1563         hash = (unsigned long)p + (unsigned long)base_offset;
1564         hash += (hash >> 8) + (hash >> 16);
1565         return hash % MAX_DELTA_CACHE;
1566 }
1567
1568 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1569         unsigned long *base_size, enum object_type *type, int keep_cache)
1570 {
1571         void *ret;
1572         unsigned long hash = pack_entry_hash(p, base_offset);
1573         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1574
1575         ret = ent->data;
1576         if (!ret || ent->p != p || ent->base_offset != base_offset)
1577                 return unpack_entry(p, base_offset, type, base_size);
1578
1579         if (!keep_cache) {
1580                 ent->data = NULL;
1581                 ent->lru.next->prev = ent->lru.prev;
1582                 ent->lru.prev->next = ent->lru.next;
1583                 delta_base_cached -= ent->size;
1584         } else {
1585                 ret = xmemdupz(ent->data, ent->size);
1586         }
1587         *type = ent->type;
1588         *base_size = ent->size;
1589         return ret;
1590 }
1591
1592 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1593 {
1594         if (ent->data) {
1595                 free(ent->data);
1596                 ent->data = NULL;
1597                 ent->lru.next->prev = ent->lru.prev;
1598                 ent->lru.prev->next = ent->lru.next;
1599                 delta_base_cached -= ent->size;
1600         }
1601 }
1602
1603 void clear_delta_base_cache(void)
1604 {
1605         unsigned long p;
1606         for (p = 0; p < MAX_DELTA_CACHE; p++)
1607                 release_delta_base_cache(&delta_base_cache[p]);
1608 }
1609
1610 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1611         void *base, unsigned long base_size, enum object_type type)
1612 {
1613         unsigned long hash = pack_entry_hash(p, base_offset);
1614         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1615         struct delta_base_cache_lru_list *lru;
1616
1617         release_delta_base_cache(ent);
1618         delta_base_cached += base_size;
1619
1620         for (lru = delta_base_cache_lru.next;
1621              delta_base_cached > delta_base_cache_limit
1622              && lru != &delta_base_cache_lru;
1623              lru = lru->next) {
1624                 struct delta_base_cache_entry *f = (void *)lru;
1625                 if (f->type == OBJ_BLOB)
1626                         release_delta_base_cache(f);
1627         }
1628         for (lru = delta_base_cache_lru.next;
1629              delta_base_cached > delta_base_cache_limit
1630              && lru != &delta_base_cache_lru;
1631              lru = lru->next) {
1632                 struct delta_base_cache_entry *f = (void *)lru;
1633                 release_delta_base_cache(f);
1634         }
1635
1636         ent->p = p;
1637         ent->base_offset = base_offset;
1638         ent->type = type;
1639         ent->data = base;
1640         ent->size = base_size;
1641         ent->lru.next = &delta_base_cache_lru;
1642         ent->lru.prev = delta_base_cache_lru.prev;
1643         delta_base_cache_lru.prev->next = &ent->lru;
1644         delta_base_cache_lru.prev = &ent->lru;
1645 }
1646
1647 static void *read_object(const unsigned char *sha1, enum object_type *type,
1648                          unsigned long *size);
1649
1650 static void *unpack_delta_entry(struct packed_git *p,
1651                                 struct pack_window **w_curs,
1652                                 off_t curpos,
1653                                 unsigned long delta_size,
1654                                 off_t obj_offset,
1655                                 enum object_type *type,
1656                                 unsigned long *sizep)
1657 {
1658         void *delta_data, *result, *base;
1659         unsigned long base_size;
1660         off_t base_offset;
1661
1662         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1663         if (!base_offset) {
1664                 error("failed to validate delta base reference "
1665                       "at offset %"PRIuMAX" from %s",
1666                       (uintmax_t)curpos, p->pack_name);
1667                 return NULL;
1668         }
1669         unuse_pack(w_curs);
1670         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1671         if (!base) {
1672                 /*
1673                  * We're probably in deep shit, but let's try to fetch
1674                  * the required base anyway from another pack or loose.
1675                  * This is costly but should happen only in the presence
1676                  * of a corrupted pack, and is better than failing outright.
1677                  */
1678                 struct revindex_entry *revidx;
1679                 const unsigned char *base_sha1;
1680                 revidx = find_pack_revindex(p, base_offset);
1681                 if (!revidx)
1682                         return NULL;
1683                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1684                 error("failed to read delta base object %s"
1685                       " at offset %"PRIuMAX" from %s",
1686                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1687                       p->pack_name);
1688                 mark_bad_packed_object(p, base_sha1);
1689                 base = read_object(base_sha1, type, &base_size);
1690                 if (!base)
1691                         return NULL;
1692         }
1693
1694         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1695         if (!delta_data) {
1696                 error("failed to unpack compressed delta "
1697                       "at offset %"PRIuMAX" from %s",
1698                       (uintmax_t)curpos, p->pack_name);
1699                 free(base);
1700                 return NULL;
1701         }
1702         result = patch_delta(base, base_size,
1703                              delta_data, delta_size,
1704                              sizep);
1705         if (!result)
1706                 die("failed to apply delta");
1707         free(delta_data);
1708         add_delta_base_cache(p, base_offset, base, base_size, *type);
1709         return result;
1710 }
1711
1712 int do_check_packed_object_crc;
1713
1714 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1715                    enum object_type *type, unsigned long *sizep)
1716 {
1717         struct pack_window *w_curs = NULL;
1718         off_t curpos = obj_offset;
1719         void *data;
1720
1721         if (do_check_packed_object_crc && p->index_version > 1) {
1722                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1723                 unsigned long len = revidx[1].offset - obj_offset;
1724                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1725                         const unsigned char *sha1 =
1726                                 nth_packed_object_sha1(p, revidx->nr);
1727                         error("bad packed object CRC for %s",
1728                               sha1_to_hex(sha1));
1729                         mark_bad_packed_object(p, sha1);
1730                         unuse_pack(&w_curs);
1731                         return NULL;
1732                 }
1733         }
1734
1735         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1736         switch (*type) {
1737         case OBJ_OFS_DELTA:
1738         case OBJ_REF_DELTA:
1739                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1740                                           obj_offset, type, sizep);
1741                 break;
1742         case OBJ_COMMIT:
1743         case OBJ_TREE:
1744         case OBJ_BLOB:
1745         case OBJ_TAG:
1746                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1747                 break;
1748         default:
1749                 data = NULL;
1750                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1751                       *type, (uintmax_t)obj_offset, p->pack_name);
1752         }
1753         unuse_pack(&w_curs);
1754         return data;
1755 }
1756
1757 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1758                                             uint32_t n)
1759 {
1760         const unsigned char *index = p->index_data;
1761         if (!index) {
1762                 if (open_pack_index(p))
1763                         return NULL;
1764                 index = p->index_data;
1765         }
1766         if (n >= p->num_objects)
1767                 return NULL;
1768         index += 4 * 256;
1769         if (p->index_version == 1) {
1770                 return index + 24 * n + 4;
1771         } else {
1772                 index += 8;
1773                 return index + 20 * n;
1774         }
1775 }
1776
1777 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1778 {
1779         const unsigned char *index = p->index_data;
1780         index += 4 * 256;
1781         if (p->index_version == 1) {
1782                 return ntohl(*((uint32_t *)(index + 24 * n)));
1783         } else {
1784                 uint32_t off;
1785                 index += 8 + p->num_objects * (20 + 4);
1786                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1787                 if (!(off & 0x80000000))
1788                         return off;
1789                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1790                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1791                                    ntohl(*((uint32_t *)(index + 4)));
1792         }
1793 }
1794
1795 off_t find_pack_entry_one(const unsigned char *sha1,
1796                                   struct packed_git *p)
1797 {
1798         const uint32_t *level1_ofs = p->index_data;
1799         const unsigned char *index = p->index_data;
1800         unsigned hi, lo, stride;
1801         static int use_lookup = -1;
1802         static int debug_lookup = -1;
1803
1804         if (debug_lookup < 0)
1805                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1806
1807         if (!index) {
1808                 if (open_pack_index(p))
1809                         return 0;
1810                 level1_ofs = p->index_data;
1811                 index = p->index_data;
1812         }
1813         if (p->index_version > 1) {
1814                 level1_ofs += 2;
1815                 index += 8;
1816         }
1817         index += 4 * 256;
1818         hi = ntohl(level1_ofs[*sha1]);
1819         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1820         if (p->index_version > 1) {
1821                 stride = 20;
1822         } else {
1823                 stride = 24;
1824                 index += 4;
1825         }
1826
1827         if (debug_lookup)
1828                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1829                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1830
1831         if (use_lookup < 0)
1832                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1833         if (use_lookup) {
1834                 int pos = sha1_entry_pos(index, stride, 0,
1835                                          lo, hi, p->num_objects, sha1);
1836                 if (pos < 0)
1837                         return 0;
1838                 return nth_packed_object_offset(p, pos);
1839         }
1840
1841         do {
1842                 unsigned mi = (lo + hi) / 2;
1843                 int cmp = hashcmp(index + mi * stride, sha1);
1844
1845                 if (debug_lookup)
1846                         printf("lo %u hi %u rg %u mi %u\n",
1847                                lo, hi, hi - lo, mi);
1848                 if (!cmp)
1849                         return nth_packed_object_offset(p, mi);
1850                 if (cmp > 0)
1851                         hi = mi;
1852                 else
1853                         lo = mi+1;
1854         } while (lo < hi);
1855         return 0;
1856 }
1857
1858 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1859 {
1860         static struct packed_git *last_found = (void *)1;
1861         struct packed_git *p;
1862         off_t offset;
1863
1864         prepare_packed_git();
1865         if (!packed_git)
1866                 return 0;
1867         p = (last_found == (void *)1) ? packed_git : last_found;
1868
1869         do {
1870                 if (p->num_bad_objects) {
1871                         unsigned i;
1872                         for (i = 0; i < p->num_bad_objects; i++)
1873                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1874                                         goto next;
1875                 }
1876
1877                 offset = find_pack_entry_one(sha1, p);
1878                 if (offset) {
1879                         /*
1880                          * We are about to tell the caller where they can
1881                          * locate the requested object.  We better make
1882                          * sure the packfile is still here and can be
1883                          * accessed before supplying that answer, as
1884                          * it may have been deleted since the index
1885                          * was loaded!
1886                          */
1887                         if (p->pack_fd == -1 && open_packed_git(p)) {
1888                                 error("packfile %s cannot be accessed", p->pack_name);
1889                                 goto next;
1890                         }
1891                         e->offset = offset;
1892                         e->p = p;
1893                         hashcpy(e->sha1, sha1);
1894                         last_found = p;
1895                         return 1;
1896                 }
1897
1898                 next:
1899                 if (p == last_found)
1900                         p = packed_git;
1901                 else
1902                         p = p->next;
1903                 if (p == last_found)
1904                         p = p->next;
1905         } while (p);
1906         return 0;
1907 }
1908
1909 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1910                                   struct packed_git *packs)
1911 {
1912         struct packed_git *p;
1913
1914         for (p = packs; p; p = p->next) {
1915                 if (find_pack_entry_one(sha1, p))
1916                         return p;
1917         }
1918         return NULL;
1919
1920 }
1921
1922 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1923 {
1924         int status;
1925         unsigned long mapsize, size;
1926         void *map;
1927         z_stream stream;
1928         char hdr[32];
1929
1930         map = map_sha1_file(sha1, &mapsize);
1931         if (!map)
1932                 return error("unable to find %s", sha1_to_hex(sha1));
1933         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1934                 status = error("unable to unpack %s header",
1935                                sha1_to_hex(sha1));
1936         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1937                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1938         else if (sizep)
1939                 *sizep = size;
1940         git_inflate_end(&stream);
1941         munmap(map, mapsize);
1942         return status;
1943 }
1944
1945 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1946 {
1947         struct pack_entry e;
1948         int status;
1949
1950         if (!find_pack_entry(sha1, &e)) {
1951                 /* Most likely it's a loose object. */
1952                 status = sha1_loose_object_info(sha1, sizep);
1953                 if (status >= 0)
1954                         return status;
1955
1956                 /* Not a loose object; someone else may have just packed it. */
1957                 reprepare_packed_git();
1958                 if (!find_pack_entry(sha1, &e))
1959                         return status;
1960         }
1961
1962         status = packed_object_info(e.p, e.offset, sizep);
1963         if (status < 0) {
1964                 mark_bad_packed_object(e.p, sha1);
1965                 status = sha1_object_info(sha1, sizep);
1966         }
1967
1968         return status;
1969 }
1970
1971 static void *read_packed_sha1(const unsigned char *sha1,
1972                               enum object_type *type, unsigned long *size)
1973 {
1974         struct pack_entry e;
1975         void *data;
1976
1977         if (!find_pack_entry(sha1, &e))
1978                 return NULL;
1979         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
1980         if (!data) {
1981                 /*
1982                  * We're probably in deep shit, but let's try to fetch
1983                  * the required object anyway from another pack or loose.
1984                  * This should happen only in the presence of a corrupted
1985                  * pack, and is better than failing outright.
1986                  */
1987                 error("failed to read object %s at offset %"PRIuMAX" from %s",
1988                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
1989                 mark_bad_packed_object(e.p, sha1);
1990                 data = read_object(sha1, type, size);
1991         }
1992         return data;
1993 }
1994
1995 /*
1996  * This is meant to hold a *small* number of objects that you would
1997  * want read_sha1_file() to be able to return, but yet you do not want
1998  * to write them into the object store (e.g. a browse-only
1999  * application).
2000  */
2001 static struct cached_object {
2002         unsigned char sha1[20];
2003         enum object_type type;
2004         void *buf;
2005         unsigned long size;
2006 } *cached_objects;
2007 static int cached_object_nr, cached_object_alloc;
2008
2009 static struct cached_object empty_tree = {
2010         EMPTY_TREE_SHA1_BIN,
2011         OBJ_TREE,
2012         "",
2013         0
2014 };
2015
2016 static struct cached_object *find_cached_object(const unsigned char *sha1)
2017 {
2018         int i;
2019         struct cached_object *co = cached_objects;
2020
2021         for (i = 0; i < cached_object_nr; i++, co++) {
2022                 if (!hashcmp(co->sha1, sha1))
2023                         return co;
2024         }
2025         if (!hashcmp(sha1, empty_tree.sha1))
2026                 return &empty_tree;
2027         return NULL;
2028 }
2029
2030 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2031                       unsigned char *sha1)
2032 {
2033         struct cached_object *co;
2034
2035         hash_sha1_file(buf, len, typename(type), sha1);
2036         if (has_sha1_file(sha1) || find_cached_object(sha1))
2037                 return 0;
2038         if (cached_object_alloc <= cached_object_nr) {
2039                 cached_object_alloc = alloc_nr(cached_object_alloc);
2040                 cached_objects = xrealloc(cached_objects,
2041                                           sizeof(*cached_objects) *
2042                                           cached_object_alloc);
2043         }
2044         co = &cached_objects[cached_object_nr++];
2045         co->size = len;
2046         co->type = type;
2047         co->buf = xmalloc(len);
2048         memcpy(co->buf, buf, len);
2049         hashcpy(co->sha1, sha1);
2050         return 0;
2051 }
2052
2053 static void *read_object(const unsigned char *sha1, enum object_type *type,
2054                          unsigned long *size)
2055 {
2056         unsigned long mapsize;
2057         void *map, *buf;
2058         struct cached_object *co;
2059
2060         co = find_cached_object(sha1);
2061         if (co) {
2062                 *type = co->type;
2063                 *size = co->size;
2064                 return xmemdupz(co->buf, co->size);
2065         }
2066
2067         buf = read_packed_sha1(sha1, type, size);
2068         if (buf)
2069                 return buf;
2070         map = map_sha1_file(sha1, &mapsize);
2071         if (map) {
2072                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2073                 munmap(map, mapsize);
2074                 return buf;
2075         }
2076         reprepare_packed_git();
2077         return read_packed_sha1(sha1, type, size);
2078 }
2079
2080 void *read_sha1_file_repl(const unsigned char *sha1,
2081                           enum object_type *type,
2082                           unsigned long *size,
2083                           const unsigned char **replacement)
2084 {
2085         const unsigned char *repl = lookup_replace_object(sha1);
2086         void *data = read_object(repl, type, size);
2087
2088         /* die if we replaced an object with one that does not exist */
2089         if (!data && repl != sha1)
2090                 die("replacement %s not found for %s",
2091                     sha1_to_hex(repl), sha1_to_hex(sha1));
2092
2093         /* legacy behavior is to die on corrupted objects */
2094         if (!data && (has_loose_object(repl) || has_packed_and_bad(repl)))
2095                 die("object %s is corrupted", sha1_to_hex(repl));
2096
2097         if (replacement)
2098                 *replacement = repl;
2099
2100         return data;
2101 }
2102
2103 void *read_object_with_reference(const unsigned char *sha1,
2104                                  const char *required_type_name,
2105                                  unsigned long *size,
2106                                  unsigned char *actual_sha1_return)
2107 {
2108         enum object_type type, required_type;
2109         void *buffer;
2110         unsigned long isize;
2111         unsigned char actual_sha1[20];
2112
2113         required_type = type_from_string(required_type_name);
2114         hashcpy(actual_sha1, sha1);
2115         while (1) {
2116                 int ref_length = -1;
2117                 const char *ref_type = NULL;
2118
2119                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2120                 if (!buffer)
2121                         return NULL;
2122                 if (type == required_type) {
2123                         *size = isize;
2124                         if (actual_sha1_return)
2125                                 hashcpy(actual_sha1_return, actual_sha1);
2126                         return buffer;
2127                 }
2128                 /* Handle references */
2129                 else if (type == OBJ_COMMIT)
2130                         ref_type = "tree ";
2131                 else if (type == OBJ_TAG)
2132                         ref_type = "object ";
2133                 else {
2134                         free(buffer);
2135                         return NULL;
2136                 }
2137                 ref_length = strlen(ref_type);
2138
2139                 if (ref_length + 40 > isize ||
2140                     memcmp(buffer, ref_type, ref_length) ||
2141                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2142                         free(buffer);
2143                         return NULL;
2144                 }
2145                 free(buffer);
2146                 /* Now we have the ID of the referred-to object in
2147                  * actual_sha1.  Check again. */
2148         }
2149 }
2150
2151 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2152                                     const char *type, unsigned char *sha1,
2153                                     char *hdr, int *hdrlen)
2154 {
2155         git_SHA_CTX c;
2156
2157         /* Generate the header */
2158         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2159
2160         /* Sha1.. */
2161         git_SHA1_Init(&c);
2162         git_SHA1_Update(&c, hdr, *hdrlen);
2163         git_SHA1_Update(&c, buf, len);
2164         git_SHA1_Final(sha1, &c);
2165 }
2166
2167 /*
2168  * Move the just written object into its final resting place.
2169  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2170  * "moving" is only a part of what it does, when no patch between
2171  * master to pu changes the call sites of this function.
2172  */
2173 int move_temp_to_file(const char *tmpfile, const char *filename)
2174 {
2175         int ret = 0;
2176
2177         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2178                 goto try_rename;
2179         else if (link(tmpfile, filename))
2180                 ret = errno;
2181
2182         /*
2183          * Coda hack - coda doesn't like cross-directory links,
2184          * so we fall back to a rename, which will mean that it
2185          * won't be able to check collisions, but that's not a
2186          * big deal.
2187          *
2188          * The same holds for FAT formatted media.
2189          *
2190          * When this succeeds, we just return.  We have nothing
2191          * left to unlink.
2192          */
2193         if (ret && ret != EEXIST) {
2194         try_rename:
2195                 if (!rename(tmpfile, filename))
2196                         goto out;
2197                 ret = errno;
2198         }
2199         unlink_or_warn(tmpfile);
2200         if (ret) {
2201                 if (ret != EEXIST) {
2202                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2203                 }
2204                 /* FIXME!!! Collision check here ? */
2205         }
2206
2207 out:
2208         if (adjust_shared_perm(filename))
2209                 return error("unable to set permission to '%s'", filename);
2210         return 0;
2211 }
2212
2213 static int write_buffer(int fd, const void *buf, size_t len)
2214 {
2215         if (write_in_full(fd, buf, len) < 0)
2216                 return error("file write error (%s)", strerror(errno));
2217         return 0;
2218 }
2219
2220 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2221                    unsigned char *sha1)
2222 {
2223         char hdr[32];
2224         int hdrlen;
2225         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2226         return 0;
2227 }
2228
2229 /* Finalize a file on disk, and close it. */
2230 static void close_sha1_file(int fd)
2231 {
2232         if (fsync_object_files)
2233                 fsync_or_die(fd, "sha1 file");
2234         if (close(fd) != 0)
2235                 die_errno("error when closing sha1 file");
2236 }
2237
2238 /* Size of directory component, including the ending '/' */
2239 static inline int directory_size(const char *filename)
2240 {
2241         const char *s = strrchr(filename, '/');
2242         if (!s)
2243                 return 0;
2244         return s - filename + 1;
2245 }
2246
2247 /*
2248  * This creates a temporary file in the same directory as the final
2249  * 'filename'
2250  *
2251  * We want to avoid cross-directory filename renames, because those
2252  * can have problems on various filesystems (FAT, NFS, Coda).
2253  */
2254 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2255 {
2256         int fd, dirlen = directory_size(filename);
2257
2258         if (dirlen + 20 > bufsiz) {
2259                 errno = ENAMETOOLONG;
2260                 return -1;
2261         }
2262         memcpy(buffer, filename, dirlen);
2263         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2264         fd = git_mkstemp_mode(buffer, 0444);
2265         if (fd < 0 && dirlen && errno == ENOENT) {
2266                 /* Make sure the directory exists */
2267                 memcpy(buffer, filename, dirlen);
2268                 buffer[dirlen-1] = 0;
2269                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2270                         return -1;
2271
2272                 /* Try again */
2273                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2274                 fd = git_mkstemp_mode(buffer, 0444);
2275         }
2276         return fd;
2277 }
2278
2279 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2280                               const void *buf, unsigned long len, time_t mtime)
2281 {
2282         int fd, ret;
2283         unsigned char compressed[4096];
2284         z_stream stream;
2285         git_SHA_CTX c;
2286         unsigned char parano_sha1[20];
2287         char *filename;
2288         static char tmpfile[PATH_MAX];
2289
2290         filename = sha1_file_name(sha1);
2291         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2292         while (fd < 0 && errno == EMFILE && unuse_one_window(packed_git, -1))
2293                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2294         if (fd < 0) {
2295                 if (errno == EACCES)
2296                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2297                 else
2298                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2299         }
2300
2301         /* Set it up */
2302         memset(&stream, 0, sizeof(stream));
2303         deflateInit(&stream, zlib_compression_level);
2304         stream.next_out = compressed;
2305         stream.avail_out = sizeof(compressed);
2306         git_SHA1_Init(&c);
2307
2308         /* First header.. */
2309         stream.next_in = (unsigned char *)hdr;
2310         stream.avail_in = hdrlen;
2311         while (deflate(&stream, 0) == Z_OK)
2312                 /* nothing */;
2313         git_SHA1_Update(&c, hdr, hdrlen);
2314
2315         /* Then the data itself.. */
2316         stream.next_in = (void *)buf;
2317         stream.avail_in = len;
2318         do {
2319                 unsigned char *in0 = stream.next_in;
2320                 ret = deflate(&stream, Z_FINISH);
2321                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2322                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2323                         die("unable to write sha1 file");
2324                 stream.next_out = compressed;
2325                 stream.avail_out = sizeof(compressed);
2326         } while (ret == Z_OK);
2327
2328         if (ret != Z_STREAM_END)
2329                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2330         ret = deflateEnd(&stream);
2331         if (ret != Z_OK)
2332                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2333         git_SHA1_Final(parano_sha1, &c);
2334         if (hashcmp(sha1, parano_sha1) != 0)
2335                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2336
2337         close_sha1_file(fd);
2338
2339         if (mtime) {
2340                 struct utimbuf utb;
2341                 utb.actime = mtime;
2342                 utb.modtime = mtime;
2343                 if (utime(tmpfile, &utb) < 0)
2344                         warning("failed utime() on %s: %s",
2345                                 tmpfile, strerror(errno));
2346         }
2347
2348         return move_temp_to_file(tmpfile, filename);
2349 }
2350
2351 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2352 {
2353         unsigned char sha1[20];
2354         char hdr[32];
2355         int hdrlen;
2356
2357         /* Normally if we have it in the pack then we do not bother writing
2358          * it out into .git/objects/??/?{38} file.
2359          */
2360         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2361         if (returnsha1)
2362                 hashcpy(returnsha1, sha1);
2363         if (has_sha1_file(sha1))
2364                 return 0;
2365         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2366 }
2367
2368 int force_object_loose(const unsigned char *sha1, time_t mtime)
2369 {
2370         void *buf;
2371         unsigned long len;
2372         enum object_type type;
2373         char hdr[32];
2374         int hdrlen;
2375         int ret;
2376
2377         if (has_loose_object(sha1))
2378                 return 0;
2379         buf = read_packed_sha1(sha1, &type, &len);
2380         if (!buf)
2381                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2382         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2383         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2384         free(buf);
2385
2386         return ret;
2387 }
2388
2389 int has_pack_index(const unsigned char *sha1)
2390 {
2391         struct stat st;
2392         if (stat(sha1_pack_index_name(sha1), &st))
2393                 return 0;
2394         return 1;
2395 }
2396
2397 int has_sha1_pack(const unsigned char *sha1)
2398 {
2399         struct pack_entry e;
2400         return find_pack_entry(sha1, &e);
2401 }
2402
2403 int has_sha1_file(const unsigned char *sha1)
2404 {
2405         struct pack_entry e;
2406
2407         if (find_pack_entry(sha1, &e))
2408                 return 1;
2409         return has_loose_object(sha1);
2410 }
2411
2412 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2413                      int write_object, enum object_type type, const char *path)
2414 {
2415         int ret, re_allocated = 0;
2416
2417         if (!type)
2418                 type = OBJ_BLOB;
2419
2420         /*
2421          * Convert blobs to git internal format
2422          */
2423         if ((type == OBJ_BLOB) && path) {
2424                 struct strbuf nbuf = STRBUF_INIT;
2425                 if (convert_to_git(path, buf, size, &nbuf,
2426                                    write_object ? safe_crlf : 0)) {
2427                         buf = strbuf_detach(&nbuf, &size);
2428                         re_allocated = 1;
2429                 }
2430         }
2431
2432         if (write_object)
2433                 ret = write_sha1_file(buf, size, typename(type), sha1);
2434         else
2435                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2436         if (re_allocated)
2437                 free(buf);
2438         return ret;
2439 }
2440
2441 #define SMALL_FILE_SIZE (32*1024)
2442
2443 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2444              enum object_type type, const char *path)
2445 {
2446         int ret;
2447         size_t size = xsize_t(st->st_size);
2448
2449         if (!S_ISREG(st->st_mode)) {
2450                 struct strbuf sbuf = STRBUF_INIT;
2451                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2452                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2453                                         type, path);
2454                 else
2455                         ret = -1;
2456                 strbuf_release(&sbuf);
2457         } else if (!size) {
2458                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2459         } else if (size <= SMALL_FILE_SIZE) {
2460                 char *buf = xmalloc(size);
2461                 if (size == read_in_full(fd, buf, size))
2462                         ret = index_mem(sha1, buf, size, write_object, type,
2463                                         path);
2464                 else
2465                         ret = error("short read %s", strerror(errno));
2466                 free(buf);
2467         } else {
2468                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2469                 ret = index_mem(sha1, buf, size, write_object, type, path);
2470                 munmap(buf, size);
2471         }
2472         close(fd);
2473         return ret;
2474 }
2475
2476 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2477 {
2478         int fd;
2479         struct strbuf sb = STRBUF_INIT;
2480
2481         switch (st->st_mode & S_IFMT) {
2482         case S_IFREG:
2483                 fd = open(path, O_RDONLY);
2484                 if (fd < 0)
2485                         return error("open(\"%s\"): %s", path,
2486                                      strerror(errno));
2487                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2488                         return error("%s: failed to insert into database",
2489                                      path);
2490                 break;
2491         case S_IFLNK:
2492                 if (strbuf_readlink(&sb, path, st->st_size)) {
2493                         char *errstr = strerror(errno);
2494                         return error("readlink(\"%s\"): %s", path,
2495                                      errstr);
2496                 }
2497                 if (!write_object)
2498                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2499                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2500                         return error("%s: failed to insert into database",
2501                                      path);
2502                 strbuf_release(&sb);
2503                 break;
2504         case S_IFDIR:
2505                 return resolve_gitlink_ref(path, "HEAD", sha1);
2506         default:
2507                 return error("%s: unsupported file type", path);
2508         }
2509         return 0;
2510 }
2511
2512 int read_pack_header(int fd, struct pack_header *header)
2513 {
2514         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2515                 /* "eof before pack header was fully read" */
2516                 return PH_ERROR_EOF;
2517
2518         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2519                 /* "protocol error (pack signature mismatch detected)" */
2520                 return PH_ERROR_PACK_SIGNATURE;
2521         if (!pack_version_ok(header->hdr_version))
2522                 /* "protocol error (pack version unsupported)" */
2523                 return PH_ERROR_PROTOCOL;
2524         return 0;
2525 }
2526
2527 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
2528 {
2529         enum object_type type = sha1_object_info(sha1, NULL);
2530         if (type < 0)
2531                 die("%s is not a valid object", sha1_to_hex(sha1));
2532         if (type != expect)
2533                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
2534                     typename(expect));
2535 }