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 "string-list.h"
  11 #include "lockfile.h"
  12 #include "delta.h"
  13 #include "pack.h"
  14 #include "blob.h"
  15 #include "commit.h"
  16 #include "run-command.h"
  17 #include "tag.h"
  18 #include "tree.h"
  19 #include "tree-walk.h"
  20 #include "refs.h"
  21 #include "pack-revindex.h"
  22 #include "sha1-lookup.h"
  23 #include "bulk-checkin.h"
  24 #include "streaming.h"
  25 #include "dir.h"
  26
  27 #ifndef O_NOATIME
  28 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
  29 #define O_NOATIME 01000000
  30 #else
  31 #define O_NOATIME 0
  32 #endif
  33 #endif
  34
  35 #define SZ_FMT PRIuMAX
  36 static inline uintmax_t sz_fmt(size_t s) { return s; }
  37
  38 const unsigned char null_sha1[20];
  39
  40 /*
  41  * This is meant to hold a *small* number of objects that you would
  42  * want read_sha1_file() to be able to return, but yet you do not want
  43  * to write them into the object store (e.g. a browse-only
  44  * application).
  45  */
  46 static struct cached_object {
  47         unsigned char sha1[20];
  48         enum object_type type;
  49         void *buf;
  50         unsigned long size;
  51 } *cached_objects;
  52 static int cached_object_nr, cached_object_alloc;
  53
  54 static struct cached_object empty_tree = {
  55         EMPTY_TREE_SHA1_BIN_LITERAL,
  56         OBJ_TREE,
  57         "",
  58         0
  59 };
  60
  61 /*
  62  * A pointer to the last packed_git in which an object was found.
  63  * When an object is sought, we look in this packfile first, because
  64  * objects that are looked up at similar times are often in the same
  65  * packfile as one another.
  66  */
  67 static struct packed_git *last_found_pack;
  68
  69 static struct cached_object *find_cached_object(const unsigned char *sha1)
  70 {
  71         int i;
  72         struct cached_object *co = cached_objects;
  73
  74         for (i = 0; i < cached_object_nr; i++, co++) {
  75                 if (!hashcmp(co->sha1, sha1))
  76                         return co;
  77         }
  78         if (!hashcmp(sha1, empty_tree.sha1))
  79                 return &empty_tree;
  80         return NULL;
  81 }
  82
  83 int mkdir_in_gitdir(const char *path)
  84 {
  85         if (mkdir(path, 0777)) {
  86                 int saved_errno = errno;
  87                 struct stat st;
  88                 struct strbuf sb = STRBUF_INIT;
  89
  90                 if (errno != EEXIST)
  91                         return -1;
  92                 /*
  93                  * Are we looking at a path in a symlinked worktree
  94                  * whose original repository does not yet have it?
  95                  * e.g. .git/rr-cache pointing at its original
  96                  * repository in which the user hasn't performed any
  97                  * conflict resolution yet?
  98                  */
  99                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
 100                     strbuf_readlink(&sb, path, st.st_size) ||
 101                     !is_absolute_path(sb.buf) ||
 102                     mkdir(sb.buf, 0777)) {
 103                         strbuf_release(&sb);
 104                         errno = saved_errno;
 105                         return -1;
 106                 }
 107                 strbuf_release(&sb);
 108         }
 109         return adjust_shared_perm(path);
 110 }
 111
 112 enum scld_error safe_create_leading_directories(char *path)
 113 {
 114         char *next_component = path + offset_1st_component(path);
 115         enum scld_error ret = SCLD_OK;
 116
 117         while (ret == SCLD_OK && next_component) {
 118                 struct stat st;
 119                 char *slash = next_component, slash_character;
 120
 121                 while (*slash && !is_dir_sep(*slash))
 122                         slash++;
 123
 124                 if (!*slash)
 125                         break;
 126
 127                 next_component = slash + 1;
 128                 while (is_dir_sep(*next_component))
 129                         next_component++;
 130                 if (!*next_component)
 131                         break;
 132
 133                 slash_character = *slash;
 134                 *slash = '\0';
 135                 if (!stat(path, &st)) {
 136                         /* path exists */
 137                         if (!S_ISDIR(st.st_mode))
 138                                 ret = SCLD_EXISTS;
 139                 } else if (mkdir(path, 0777)) {
 140                         if (errno == EEXIST &&
 141                             !stat(path, &st) && S_ISDIR(st.st_mode))
 142                                 ; /* somebody created it since we checked */
 143                         else if (errno == ENOENT)
 144                                 /*
 145                                  * Either mkdir() failed because
 146                                  * somebody just pruned the containing
 147                                  * directory, or stat() failed because
 148                                  * the file that was in our way was
 149                                  * just removed.  Either way, inform
 150                                  * the caller that it might be worth
 151                                  * trying again:
 152                                  */
 153                                 ret = SCLD_VANISHED;
 154                         else
 155                                 ret = SCLD_FAILED;
 156                 } else if (adjust_shared_perm(path)) {
 157                         ret = SCLD_PERMS;
 158                 }
 159                 *slash = slash_character;
 160         }
 161         return ret;
 162 }
 163
 164 enum scld_error safe_create_leading_directories_const(const char *path)
 165 {
 166         /* path points to cache entries, so xstrdup before messing with it */
 167         char *buf = xstrdup(path);
 168         enum scld_error result = safe_create_leading_directories(buf);
 169         free(buf);
 170         return result;
 171 }
 172
 173 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
 174 {
 175         int i;
 176         for (i = 0; i < 20; i++) {
 177                 static char hex[] = "0123456789abcdef";
 178                 unsigned int val = sha1[i];
 179                 char *pos = pathbuf + i*2 + (i > 0);
 180                 *pos++ = hex[val >> 4];
 181                 *pos = hex[val & 0xf];
 182         }
 183 }
 184
 185 const char *sha1_file_name(const unsigned char *sha1)
 186 {
 187         static char buf[PATH_MAX];
 188         const char *objdir;
 189         int len;
 190
 191         objdir = get_object_directory();
 192         len = strlen(objdir);
 193
 194         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
 195         if (len + 43 > PATH_MAX)
 196                 die("insanely long object directory %s", objdir);
 197         memcpy(buf, objdir, len);
 198         buf[len] = '/';
 199         buf[len+3] = '/';
 200         buf[len+42] = '\0';
 201         fill_sha1_path(buf + len + 1, sha1);
 202         return buf;
 203 }
 204
 205 /*
 206  * Return the name of the pack or index file with the specified sha1
 207  * in its filename.  *base and *name are scratch space that must be
 208  * provided by the caller.  which should be "pack" or "idx".
 209  */
 210 static char *sha1_get_pack_name(const unsigned char *sha1,
 211                                 char **name, char **base, const char *which)
 212 {
 213         static const char hex[] = "0123456789abcdef";
 214         char *buf;
 215         int i;
 216
 217         if (!*base) {
 218                 const char *sha1_file_directory = get_object_directory();
 219                 int len = strlen(sha1_file_directory);
 220                 *base = xmalloc(len + 60);
 221                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 222                         sha1_file_directory, which);
 223                 *name = *base + len + 11;
 224         }
 225
 226         buf = *name;
 227
 228         for (i = 0; i < 20; i++) {
 229                 unsigned int val = *sha1++;
 230                 *buf++ = hex[val >> 4];
 231                 *buf++ = hex[val & 0xf];
 232         }
 233
 234         return *base;
 235 }
 236
 237 char *sha1_pack_name(const unsigned char *sha1)
 238 {
 239         static char *name, *base;
 240
 241         return sha1_get_pack_name(sha1, &name, &base, "pack");
 242 }
 243
 244 char *sha1_pack_index_name(const unsigned char *sha1)
 245 {
 246         static char *name, *base;
 247
 248         return sha1_get_pack_name(sha1, &name, &base, "idx");
 249 }
 250
 251 struct alternate_object_database *alt_odb_list;
 252 static struct alternate_object_database **alt_odb_tail;
 253
 254 /*
 255  * Prepare alternate object database registry.
 256  *
 257  * The variable alt_odb_list points at the list of struct
 258  * alternate_object_database.  The elements on this list come from
 259  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 260  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 261  * whose contents is similar to that environment variable but can be
 262  * LF separated.  Its base points at a statically allocated buffer that
 263  * contains "/the/directory/corresponding/to/.git/objects/...", while
 264  * its name points just after the slash at the end of ".git/objects/"
 265  * in the example above, and has enough space to hold 40-byte hex
 266  * SHA1, an extra slash for the first level indirection, and the
 267  * terminating NUL.
 268  */
 269 static int link_alt_odb_entry(const char *entry, const char *relative_base,
 270         int depth, const char *normalized_objdir)
 271 {
 272         struct alternate_object_database *ent;
 273         struct alternate_object_database *alt;
 274         int pfxlen, entlen;
 275         struct strbuf pathbuf = STRBUF_INIT;
 276
 277         if (!is_absolute_path(entry) && relative_base) {
 278                 strbuf_addstr(&pathbuf, real_path(relative_base));
 279                 strbuf_addch(&pathbuf, '/');
 280         }
 281         strbuf_addstr(&pathbuf, entry);
 282
 283         normalize_path_copy(pathbuf.buf, pathbuf.buf);
 284
 285         pfxlen = strlen(pathbuf.buf);
 286
 287         /*
 288          * The trailing slash after the directory name is given by
 289          * this function at the end. Remove duplicates.
 290          */
 291         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
 292                 pfxlen -= 1;
 293
 294         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
 295         ent = xmalloc(sizeof(*ent) + entlen);
 296         memcpy(ent->base, pathbuf.buf, pfxlen);
 297         strbuf_release(&pathbuf);
 298
 299         ent->name = ent->base + pfxlen + 1;
 300         ent->base[pfxlen + 3] = '/';
 301         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 302
 303         /* Detect cases where alternate disappeared */
 304         if (!is_directory(ent->base)) {
 305                 error("object directory %s does not exist; "
 306                       "check .git/objects/info/alternates.",
 307                       ent->base);
 308                 free(ent);
 309                 return -1;
 310         }
 311
 312         /* Prevent the common mistake of listing the same
 313          * thing twice, or object directory itself.
 314          */
 315         for (alt = alt_odb_list; alt; alt = alt->next) {
 316                 if (pfxlen == alt->name - alt->base - 1 &&
 317                     !memcmp(ent->base, alt->base, pfxlen)) {
 318                         free(ent);
 319                         return -1;
 320                 }
 321         }
 322         if (!strcmp_icase(ent->base, normalized_objdir)) {
 323                 free(ent);
 324                 return -1;
 325         }
 326
 327         /* add the alternate entry */
 328         *alt_odb_tail = ent;
 329         alt_odb_tail = &(ent->next);
 330         ent->next = NULL;
 331
 332         /* recursively add alternates */
 333         read_info_alternates(ent->base, depth + 1);
 334
 335         ent->base[pfxlen] = '/';
 336
 337         return 0;
 338 }
 339
 340 static void link_alt_odb_entries(const char *alt, int len, int sep,
 341                                  const char *relative_base, int depth)
 342 {
 343         struct string_list entries = STRING_LIST_INIT_NODUP;
 344         char *alt_copy;
 345         int i;
 346         struct strbuf objdirbuf = STRBUF_INIT;
 347
 348         if (depth > 5) {
 349                 error("%s: ignoring alternate object stores, nesting too deep.",
 350                                 relative_base);
 351                 return;
 352         }
 353
 354         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
 355         normalize_path_copy(objdirbuf.buf, objdirbuf.buf);
 356
 357         alt_copy = xmemdupz(alt, len);
 358         string_list_split_in_place(&entries, alt_copy, sep, -1);
 359         for (i = 0; i < entries.nr; i++) {
 360                 const char *entry = entries.items[i].string;
 361                 if (entry[0] == '\0' || entry[0] == '#')
 362                         continue;
 363                 if (!is_absolute_path(entry) && depth) {
 364                         error("%s: ignoring relative alternate object store %s",
 365                                         relative_base, entry);
 366                 } else {
 367                         link_alt_odb_entry(entry, relative_base, depth, objdirbuf.buf);
 368                 }
 369         }
 370         string_list_clear(&entries, 0);
 371         free(alt_copy);
 372         strbuf_release(&objdirbuf);
 373 }
 374
 375 void read_info_alternates(const char * relative_base, int depth)
 376 {
 377         char *map;
 378         size_t mapsz;
 379         struct stat st;
 380         char *path;
 381         int fd;
 382
 383         path = xstrfmt("%s/info/alternates", relative_base);
 384         fd = git_open_noatime(path);
 385         free(path);
 386         if (fd < 0)
 387                 return;
 388         if (fstat(fd, &st) || (st.st_size == 0)) {
 389                 close(fd);
 390                 return;
 391         }
 392         mapsz = xsize_t(st.st_size);
 393         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 394         close(fd);
 395
 396         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
 397
 398         munmap(map, mapsz);
 399 }
 400
 401 void add_to_alternates_file(const char *reference)
 402 {
 403         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 404         char *alts = git_pathdup("objects/info/alternates");
 405         FILE *in, *out;
 406
 407         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
 408         out = fdopen_lock_file(lock, "w");
 409         if (!out)
 410                 die_errno("unable to fdopen alternates lockfile");
 411
 412         in = fopen(alts, "r");
 413         if (in) {
 414                 struct strbuf line = STRBUF_INIT;
 415                 int found = 0;
 416
 417                 while (strbuf_getline(&line, in, '\n') != EOF) {
 418                         if (!strcmp(reference, line.buf)) {
 419                                 found = 1;
 420                                 break;
 421                         }
 422                         fprintf_or_die(out, "%s\n", line.buf);
 423                 }
 424
 425                 strbuf_release(&line);
 426                 fclose(in);
 427
 428                 if (found) {
 429                         rollback_lock_file(lock);
 430                         lock = NULL;
 431                 }
 432         }
 433         else if (errno != ENOENT)
 434                 die_errno("unable to read alternates file");
 435
 436         if (lock) {
 437                 fprintf_or_die(out, "%s\n", reference);
 438                 if (commit_lock_file(lock))
 439                         die_errno("unable to move new alternates file into place");
 440                 if (alt_odb_tail)
 441                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
 442         }
 443         free(alts);
 444 }
 445
 446 int foreach_alt_odb(alt_odb_fn fn, void *cb)
 447 {
 448         struct alternate_object_database *ent;
 449         int r = 0;
 450
 451         prepare_alt_odb();
 452         for (ent = alt_odb_list; ent; ent = ent->next) {
 453                 r = fn(ent, cb);
 454                 if (r)
 455                         break;
 456         }
 457         return r;
 458 }
 459
 460 void prepare_alt_odb(void)
 461 {
 462         const char *alt;
 463
 464         if (alt_odb_tail)
 465                 return;
 466
 467         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 468         if (!alt) alt = "";
 469
 470         alt_odb_tail = &alt_odb_list;
 471         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
 472
 473         read_info_alternates(get_object_directory(), 0);
 474 }
 475
 476 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
 477 static int freshen_file(const char *fn)
 478 {
 479         struct utimbuf t;
 480         t.actime = t.modtime = time(NULL);
 481         return !utime(fn, &t);
 482 }
 483
 484 /*
 485  * All of the check_and_freshen functions return 1 if the file exists and was
 486  * freshened (if freshening was requested), 0 otherwise. If they return
 487  * 0, you should not assume that it is safe to skip a write of the object (it
 488  * either does not exist on disk, or has a stale mtime and may be subject to
 489  * pruning).
 490  */
 491 static int check_and_freshen_file(const char *fn, int freshen)
 492 {
 493         if (access(fn, F_OK))
 494                 return 0;
 495         if (freshen && !freshen_file(fn))
 496                 return 0;
 497         return 1;
 498 }
 499
 500 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
 501 {
 502         return check_and_freshen_file(sha1_file_name(sha1), freshen);
 503 }
 504
 505 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
 506 {
 507         struct alternate_object_database *alt;
 508         prepare_alt_odb();
 509         for (alt = alt_odb_list; alt; alt = alt->next) {
 510                 fill_sha1_path(alt->name, sha1);
 511                 if (check_and_freshen_file(alt->base, freshen))
 512                         return 1;
 513         }
 514         return 0;
 515 }
 516
 517 static int check_and_freshen(const unsigned char *sha1, int freshen)
 518 {
 519         return check_and_freshen_local(sha1, freshen) ||
 520                check_and_freshen_nonlocal(sha1, freshen);
 521 }
 522
 523 int has_loose_object_nonlocal(const unsigned char *sha1)
 524 {
 525         return check_and_freshen_nonlocal(sha1, 0);
 526 }
 527
 528 static int has_loose_object(const unsigned char *sha1)
 529 {
 530         return check_and_freshen(sha1, 0);
 531 }
 532
 533 static unsigned int pack_used_ctr;
 534 static unsigned int pack_mmap_calls;
 535 static unsigned int peak_pack_open_windows;
 536 static unsigned int pack_open_windows;
 537 static unsigned int pack_open_fds;
 538 static unsigned int pack_max_fds;
 539 static size_t peak_pack_mapped;
 540 static size_t pack_mapped;
 541 struct packed_git *packed_git;
 542
 543 void pack_report(void)
 544 {
 545         fprintf(stderr,
 546                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 547                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 548                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 549                 sz_fmt(getpagesize()),
 550                 sz_fmt(packed_git_window_size),
 551                 sz_fmt(packed_git_limit));
 552         fprintf(stderr,
 553                 "pack_report: pack_used_ctr            = %10u\n"
 554                 "pack_report: pack_mmap_calls          = %10u\n"
 555                 "pack_report: pack_open_windows        = %10u / %10u\n"
 556                 "pack_report: pack_mapped              = "
 557                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 558                 pack_used_ctr,
 559                 pack_mmap_calls,
 560                 pack_open_windows, peak_pack_open_windows,
 561                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 562 }
 563
 564 /*
 565  * Open and mmap the index file at path, perform a couple of
 566  * consistency checks, then record its information to p.  Return 0 on
 567  * success.
 568  */
 569 static int check_packed_git_idx(const char *path, struct packed_git *p)
 570 {
 571         void *idx_map;
 572         struct pack_idx_header *hdr;
 573         size_t idx_size;
 574         uint32_t version, nr, i, *index;
 575         int fd = git_open_noatime(path);
 576         struct stat st;
 577
 578         if (fd < 0)
 579                 return -1;
 580         if (fstat(fd, &st)) {
 581                 close(fd);
 582                 return -1;
 583         }
 584         idx_size = xsize_t(st.st_size);
 585         if (idx_size < 4 * 256 + 20 + 20) {
 586                 close(fd);
 587                 return error("index file %s is too small", path);
 588         }
 589         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 590         close(fd);
 591
 592         hdr = idx_map;
 593         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 594                 version = ntohl(hdr->idx_version);
 595                 if (version < 2 || version > 2) {
 596                         munmap(idx_map, idx_size);
 597                         return error("index file %s is version %"PRIu32
 598                                      " and is not supported by this binary"
 599                                      " (try upgrading GIT to a newer version)",
 600                                      path, version);
 601                 }
 602         } else
 603                 version = 1;
 604
 605         nr = 0;
 606         index = idx_map;
 607         if (version > 1)
 608                 index += 2;  /* skip index header */
 609         for (i = 0; i < 256; i++) {
 610                 uint32_t n = ntohl(index[i]);
 611                 if (n < nr) {
 612                         munmap(idx_map, idx_size);
 613                         return error("non-monotonic index %s", path);
 614                 }
 615                 nr = n;
 616         }
 617
 618         if (version == 1) {
 619                 /*
 620                  * Total size:
 621                  *  - 256 index entries 4 bytes each
 622                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 623                  *  - 20-byte SHA1 of the packfile
 624                  *  - 20-byte SHA1 file checksum
 625                  */
 626                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 627                         munmap(idx_map, idx_size);
 628                         return error("wrong index v1 file size in %s", path);
 629                 }
 630         } else if (version == 2) {
 631                 /*
 632                  * Minimum size:
 633                  *  - 8 bytes of header
 634                  *  - 256 index entries 4 bytes each
 635                  *  - 20-byte sha1 entry * nr
 636                  *  - 4-byte crc entry * nr
 637                  *  - 4-byte offset entry * nr
 638                  *  - 20-byte SHA1 of the packfile
 639                  *  - 20-byte SHA1 file checksum
 640                  * And after the 4-byte offset table might be a
 641                  * variable sized table containing 8-byte entries
 642                  * for offsets larger than 2^31.
 643                  */
 644                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 645                 unsigned long max_size = min_size;
 646                 if (nr)
 647                         max_size += (nr - 1)*8;
 648                 if (idx_size < min_size || idx_size > max_size) {
 649                         munmap(idx_map, idx_size);
 650                         return error("wrong index v2 file size in %s", path);
 651                 }
 652                 if (idx_size != min_size &&
 653                     /*
 654                      * make sure we can deal with large pack offsets.
 655                      * 31-bit signed offset won't be enough, neither
 656                      * 32-bit unsigned one will be.
 657                      */
 658                     (sizeof(off_t) <= 4)) {
 659                         munmap(idx_map, idx_size);
 660                         return error("pack too large for current definition of off_t in %s", path);
 661                 }
 662         }
 663
 664         p->index_version = version;
 665         p->index_data = idx_map;
 666         p->index_size = idx_size;
 667         p->num_objects = nr;
 668         return 0;
 669 }
 670
 671 int open_pack_index(struct packed_git *p)
 672 {
 673         char *idx_name;
 674         int ret;
 675
 676         if (p->index_data)
 677                 return 0;
 678
 679         idx_name = xstrdup(p->pack_name);
 680         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 681         ret = check_packed_git_idx(idx_name, p);
 682         free(idx_name);
 683         return ret;
 684 }
 685
 686 static void scan_windows(struct packed_git *p,
 687         struct packed_git **lru_p,
 688         struct pack_window **lru_w,
 689         struct pack_window **lru_l)
 690 {
 691         struct pack_window *w, *w_l;
 692
 693         for (w_l = NULL, w = p->windows; w; w = w->next) {
 694                 if (!w->inuse_cnt) {
 695                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 696                                 *lru_p = p;
 697                                 *lru_w = w;
 698                                 *lru_l = w_l;
 699                         }
 700                 }
 701                 w_l = w;
 702         }
 703 }
 704
 705 static int unuse_one_window(struct packed_git *current)
 706 {
 707         struct packed_git *p, *lru_p = NULL;
 708         struct pack_window *lru_w = NULL, *lru_l = NULL;
 709
 710         if (current)
 711                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 712         for (p = packed_git; p; p = p->next)
 713                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 714         if (lru_p) {
 715                 munmap(lru_w->base, lru_w->len);
 716                 pack_mapped -= lru_w->len;
 717                 if (lru_l)
 718                         lru_l->next = lru_w->next;
 719                 else
 720                         lru_p->windows = lru_w->next;
 721                 free(lru_w);
 722                 pack_open_windows--;
 723                 return 1;
 724         }
 725         return 0;
 726 }
 727
 728 void release_pack_memory(size_t need)
 729 {
 730         size_t cur = pack_mapped;
 731         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
 732                 ; /* nothing */
 733 }
 734
 735 static void mmap_limit_check(size_t length)
 736 {
 737         static size_t limit = 0;
 738         if (!limit) {
 739                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
 740                 if (!limit)
 741                         limit = SIZE_MAX;
 742         }
 743         if (length > limit)
 744                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
 745                     (uintmax_t)length, (uintmax_t)limit);
 746 }
 747
 748 void *xmmap_gently(void *start, size_t length,
 749                   int prot, int flags, int fd, off_t offset)
 750 {
 751         void *ret;
 752
 753         mmap_limit_check(length);
 754         ret = mmap(start, length, prot, flags, fd, offset);
 755         if (ret == MAP_FAILED) {
 756                 if (!length)
 757                         return NULL;
 758                 release_pack_memory(length);
 759                 ret = mmap(start, length, prot, flags, fd, offset);
 760         }
 761         return ret;
 762 }
 763
 764 void *xmmap(void *start, size_t length,
 765         int prot, int flags, int fd, off_t offset)
 766 {
 767         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
 768         if (ret == MAP_FAILED)
 769                 die_errno("mmap failed");
 770         return ret;
 771 }
 772
 773 void close_pack_windows(struct packed_git *p)
 774 {
 775         while (p->windows) {
 776                 struct pack_window *w = p->windows;
 777
 778                 if (w->inuse_cnt)
 779                         die("pack '%s' still has open windows to it",
 780                             p->pack_name);
 781                 munmap(w->base, w->len);
 782                 pack_mapped -= w->len;
 783                 pack_open_windows--;
 784                 p->windows = w->next;
 785                 free(w);
 786         }
 787 }
 788
 789 static int close_pack_fd(struct packed_git *p)
 790 {
 791         if (p->pack_fd < 0)
 792                 return 0;
 793
 794         close(p->pack_fd);
 795         pack_open_fds--;
 796         p->pack_fd = -1;
 797
 798         return 1;
 799 }
 800
 801 static void close_pack(struct packed_git *p)
 802 {
 803         close_pack_windows(p);
 804         close_pack_fd(p);
 805         close_pack_index(p);
 806 }
 807
 808 void close_all_packs(void)
 809 {
 810         struct packed_git *p;
 811
 812         for (p = packed_git; p; p = p->next)
 813                 if (p->do_not_close)
 814                         die("BUG! Want to close pack marked 'do-not-close'");
 815                 else
 816                         close_pack(p);
 817 }
 818
 819
 820 /*
 821  * The LRU pack is the one with the oldest MRU window, preferring packs
 822  * with no used windows, or the oldest mtime if it has no windows allocated.
 823  */
 824 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 825 {
 826         struct pack_window *w, *this_mru_w;
 827         int has_windows_inuse = 0;
 828
 829         /*
 830          * Reject this pack if it has windows and the previously selected
 831          * one does not.  If this pack does not have windows, reject
 832          * it if the pack file is newer than the previously selected one.
 833          */
 834         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 835                 return;
 836
 837         for (w = this_mru_w = p->windows; w; w = w->next) {
 838                 /*
 839                  * Reject this pack if any of its windows are in use,
 840                  * but the previously selected pack did not have any
 841                  * inuse windows.  Otherwise, record that this pack
 842                  * has windows in use.
 843                  */
 844                 if (w->inuse_cnt) {
 845                         if (*accept_windows_inuse)
 846                                 has_windows_inuse = 1;
 847                         else
 848                                 return;
 849                 }
 850
 851                 if (w->last_used > this_mru_w->last_used)
 852                         this_mru_w = w;
 853
 854                 /*
 855                  * Reject this pack if it has windows that have been
 856                  * used more recently than the previously selected pack.
 857                  * If the previously selected pack had windows inuse and
 858                  * we have not encountered a window in this pack that is
 859                  * inuse, skip this check since we prefer a pack with no
 860                  * inuse windows to one that has inuse windows.
 861                  */
 862                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 863                     this_mru_w->last_used > (*mru_w)->last_used)
 864                         return;
 865         }
 866
 867         /*
 868          * Select this pack.
 869          */
 870         *mru_w = this_mru_w;
 871         *lru_p = p;
 872         *accept_windows_inuse = has_windows_inuse;
 873 }
 874
 875 static int close_one_pack(void)
 876 {
 877         struct packed_git *p, *lru_p = NULL;
 878         struct pack_window *mru_w = NULL;
 879         int accept_windows_inuse = 1;
 880
 881         for (p = packed_git; p; p = p->next) {
 882                 if (p->pack_fd == -1)
 883                         continue;
 884                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 885         }
 886
 887         if (lru_p)
 888                 return close_pack_fd(lru_p);
 889
 890         return 0;
 891 }
 892
 893 void unuse_pack(struct pack_window **w_cursor)
 894 {
 895         struct pack_window *w = *w_cursor;
 896         if (w) {
 897                 w->inuse_cnt--;
 898                 *w_cursor = NULL;
 899         }
 900 }
 901
 902 void close_pack_index(struct packed_git *p)
 903 {
 904         if (p->index_data) {
 905                 munmap((void *)p->index_data, p->index_size);
 906                 p->index_data = NULL;
 907         }
 908 }
 909
 910 /*
 911  * This is used by git-repack in case a newly created pack happens to
 912  * contain the same set of objects as an existing one.  In that case
 913  * the resulting file might be different even if its name would be the
 914  * same.  It is best to close any reference to the old pack before it is
 915  * replaced on disk.  Of course no index pointers or windows for given pack
 916  * must subsist at this point.  If ever objects from this pack are requested
 917  * again, the new version of the pack will be reinitialized through
 918  * reprepare_packed_git().
 919  */
 920 void free_pack_by_name(const char *pack_name)
 921 {
 922         struct packed_git *p, **pp = &packed_git;
 923
 924         while (*pp) {
 925                 p = *pp;
 926                 if (strcmp(pack_name, p->pack_name) == 0) {
 927                         clear_delta_base_cache();
 928                         close_pack(p);
 929                         free(p->bad_object_sha1);
 930                         *pp = p->next;
 931                         if (last_found_pack == p)
 932                                 last_found_pack = NULL;
 933                         free(p);
 934                         return;
 935                 }
 936                 pp = &p->next;
 937         }
 938 }
 939
 940 static unsigned int get_max_fd_limit(void)
 941 {
 942 #ifdef RLIMIT_NOFILE
 943         {
 944                 struct rlimit lim;
 945
 946                 if (!getrlimit(RLIMIT_NOFILE, &lim))
 947                         return lim.rlim_cur;
 948         }
 949 #endif
 950
 951 #ifdef _SC_OPEN_MAX
 952         {
 953                 long open_max = sysconf(_SC_OPEN_MAX);
 954                 if (0 < open_max)
 955                         return open_max;
 956                 /*
 957                  * Otherwise, we got -1 for one of the two
 958                  * reasons:
 959                  *
 960                  * (1) sysconf() did not understand _SC_OPEN_MAX
 961                  *     and signaled an error with -1; or
 962                  * (2) sysconf() said there is no limit.
 963                  *
 964                  * We _could_ clear errno before calling sysconf() to
 965                  * tell these two cases apart and return a huge number
 966                  * in the latter case to let the caller cap it to a
 967                  * value that is not so selfish, but letting the
 968                  * fallback OPEN_MAX codepath take care of these cases
 969                  * is a lot simpler.
 970                  */
 971         }
 972 #endif
 973
 974 #ifdef OPEN_MAX
 975         return OPEN_MAX;
 976 #else
 977         return 1; /* see the caller ;-) */
 978 #endif
 979 }
 980
 981 /*
 982  * Do not call this directly as this leaks p->pack_fd on error return;
 983  * call open_packed_git() instead.
 984  */
 985 static int open_packed_git_1(struct packed_git *p)
 986 {
 987         struct stat st;
 988         struct pack_header hdr;
 989         unsigned char sha1[20];
 990         unsigned char *idx_sha1;
 991         long fd_flag;
 992
 993         if (!p->index_data && open_pack_index(p))
 994                 return error("packfile %s index unavailable", p->pack_name);
 995
 996         if (!pack_max_fds) {
 997                 unsigned int max_fds = get_max_fd_limit();
 998
 999                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1000                 if (25 < max_fds)
1001                         pack_max_fds = max_fds - 25;
1002                 else
1003                         pack_max_fds = 1;
1004         }
1005
1006         while (pack_max_fds <= pack_open_fds && close_one_pack())
1007                 ; /* nothing */
1008
1009         p->pack_fd = git_open_noatime(p->pack_name);
1010         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1011                 return -1;
1012         pack_open_fds++;
1013
1014         /* If we created the struct before we had the pack we lack size. */
1015         if (!p->pack_size) {
1016                 if (!S_ISREG(st.st_mode))
1017                         return error("packfile %s not a regular file", p->pack_name);
1018                 p->pack_size = st.st_size;
1019         } else if (p->pack_size != st.st_size)
1020                 return error("packfile %s size changed", p->pack_name);
1021
1022         /* We leave these file descriptors open with sliding mmap;
1023          * there is no point keeping them open across exec(), though.
1024          */
1025         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1026         if (fd_flag < 0)
1027                 return error("cannot determine file descriptor flags");
1028         fd_flag |= FD_CLOEXEC;
1029         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1030                 return error("cannot set FD_CLOEXEC");
1031
1032         /* Verify we recognize this pack file format. */
1033         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1034                 return error("file %s is far too short to be a packfile", p->pack_name);
1035         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1036                 return error("file %s is not a GIT packfile", p->pack_name);
1037         if (!pack_version_ok(hdr.hdr_version))
1038                 return error("packfile %s is version %"PRIu32" and not"
1039                         " supported (try upgrading GIT to a newer version)",
1040                         p->pack_name, ntohl(hdr.hdr_version));
1041
1042         /* Verify the pack matches its index. */
1043         if (p->num_objects != ntohl(hdr.hdr_entries))
1044                 return error("packfile %s claims to have %"PRIu32" objects"
1045                              " while index indicates %"PRIu32" objects",
1046                              p->pack_name, ntohl(hdr.hdr_entries),
1047                              p->num_objects);
1048         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1049                 return error("end of packfile %s is unavailable", p->pack_name);
1050         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1051                 return error("packfile %s signature is unavailable", p->pack_name);
1052         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1053         if (hashcmp(sha1, idx_sha1))
1054                 return error("packfile %s does not match index", p->pack_name);
1055         return 0;
1056 }
1057
1058 static int open_packed_git(struct packed_git *p)
1059 {
1060         if (!open_packed_git_1(p))
1061                 return 0;
1062         close_pack_fd(p);
1063         return -1;
1064 }
1065
1066 static int in_window(struct pack_window *win, off_t offset)
1067 {
1068         /* We must promise at least 20 bytes (one hash) after the
1069          * offset is available from this window, otherwise the offset
1070          * is not actually in this window and a different window (which
1071          * has that one hash excess) must be used.  This is to support
1072          * the object header and delta base parsing routines below.
1073          */
1074         off_t win_off = win->offset;
1075         return win_off <= offset
1076                 && (offset + 20) <= (win_off + win->len);
1077 }
1078
1079 unsigned char *use_pack(struct packed_git *p,
1080                 struct pack_window **w_cursor,
1081                 off_t offset,
1082                 unsigned long *left)
1083 {
1084         struct pack_window *win = *w_cursor;
1085
1086         /* Since packfiles end in a hash of their content and it's
1087          * pointless to ask for an offset into the middle of that
1088          * hash, and the in_window function above wouldn't match
1089          * don't allow an offset too close to the end of the file.
1090          */
1091         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1092                 die("packfile %s cannot be accessed", p->pack_name);
1093         if (offset > (p->pack_size - 20))
1094                 die("offset beyond end of packfile (truncated pack?)");
1095
1096         if (!win || !in_window(win, offset)) {
1097                 if (win)
1098                         win->inuse_cnt--;
1099                 for (win = p->windows; win; win = win->next) {
1100                         if (in_window(win, offset))
1101                                 break;
1102                 }
1103                 if (!win) {
1104                         size_t window_align = packed_git_window_size / 2;
1105                         off_t len;
1106
1107                         if (p->pack_fd == -1 && open_packed_git(p))
1108                                 die("packfile %s cannot be accessed", p->pack_name);
1109
1110                         win = xcalloc(1, sizeof(*win));
1111                         win->offset = (offset / window_align) * window_align;
1112                         len = p->pack_size - win->offset;
1113                         if (len > packed_git_window_size)
1114                                 len = packed_git_window_size;
1115                         win->len = (size_t)len;
1116                         pack_mapped += win->len;
1117                         while (packed_git_limit < pack_mapped
1118                                 && unuse_one_window(p))
1119                                 ; /* nothing */
1120                         win->base = xmmap(NULL, win->len,
1121                                 PROT_READ, MAP_PRIVATE,
1122                                 p->pack_fd, win->offset);
1123                         if (win->base == MAP_FAILED)
1124                                 die("packfile %s cannot be mapped: %s",
1125                                         p->pack_name,
1126                                         strerror(errno));
1127                         if (!win->offset && win->len == p->pack_size
1128                                 && !p->do_not_close)
1129                                 close_pack_fd(p);
1130                         pack_mmap_calls++;
1131                         pack_open_windows++;
1132                         if (pack_mapped > peak_pack_mapped)
1133                                 peak_pack_mapped = pack_mapped;
1134                         if (pack_open_windows > peak_pack_open_windows)
1135                                 peak_pack_open_windows = pack_open_windows;
1136                         win->next = p->windows;
1137                         p->windows = win;
1138                 }
1139         }
1140         if (win != *w_cursor) {
1141                 win->last_used = pack_used_ctr++;
1142                 win->inuse_cnt++;
1143                 *w_cursor = win;
1144         }
1145         offset -= win->offset;
1146         if (left)
1147                 *left = win->len - xsize_t(offset);
1148         return win->base + offset;
1149 }
1150
1151 static struct packed_git *alloc_packed_git(int extra)
1152 {
1153         struct packed_git *p = xmalloc(sizeof(*p) + extra);
1154         memset(p, 0, sizeof(*p));
1155         p->pack_fd = -1;
1156         return p;
1157 }
1158
1159 static void try_to_free_pack_memory(size_t size)
1160 {
1161         release_pack_memory(size);
1162 }
1163
1164 struct packed_git *add_packed_git(const char *path, int path_len, int local)
1165 {
1166         static int have_set_try_to_free_routine;
1167         struct stat st;
1168         struct packed_git *p = alloc_packed_git(path_len + 2);
1169
1170         if (!have_set_try_to_free_routine) {
1171                 have_set_try_to_free_routine = 1;
1172                 set_try_to_free_routine(try_to_free_pack_memory);
1173         }
1174
1175         /*
1176          * Make sure a corresponding .pack file exists and that
1177          * the index looks sane.
1178          */
1179         path_len -= strlen(".idx");
1180         if (path_len < 1) {
1181                 free(p);
1182                 return NULL;
1183         }
1184         memcpy(p->pack_name, path, path_len);
1185
1186         strcpy(p->pack_name + path_len, ".keep");
1187         if (!access(p->pack_name, F_OK))
1188                 p->pack_keep = 1;
1189
1190         strcpy(p->pack_name + path_len, ".pack");
1191         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1192                 free(p);
1193                 return NULL;
1194         }
1195
1196         /* ok, it looks sane as far as we can check without
1197          * actually mapping the pack file.
1198          */
1199         p->pack_size = st.st_size;
1200         p->pack_local = local;
1201         p->mtime = st.st_mtime;
1202         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1203                 hashclr(p->sha1);
1204         return p;
1205 }
1206
1207 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1208 {
1209         const char *path = sha1_pack_name(sha1);
1210         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
1211
1212         strcpy(p->pack_name, path);
1213         hashcpy(p->sha1, sha1);
1214         if (check_packed_git_idx(idx_path, p)) {
1215                 free(p);
1216                 return NULL;
1217         }
1218
1219         return p;
1220 }
1221
1222 void install_packed_git(struct packed_git *pack)
1223 {
1224         if (pack->pack_fd != -1)
1225                 pack_open_fds++;
1226
1227         pack->next = packed_git;
1228         packed_git = pack;
1229 }
1230
1231 void (*report_garbage)(unsigned seen_bits, const char *path);
1232
1233 static void report_helper(const struct string_list *list,
1234                           int seen_bits, int first, int last)
1235 {
1236         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1237                 return;
1238
1239         for (; first < last; first++)
1240                 report_garbage(seen_bits, list->items[first].string);
1241 }
1242
1243 static void report_pack_garbage(struct string_list *list)
1244 {
1245         int i, baselen = -1, first = 0, seen_bits = 0;
1246
1247         if (!report_garbage)
1248                 return;
1249
1250         string_list_sort(list);
1251
1252         for (i = 0; i < list->nr; i++) {
1253                 const char *path = list->items[i].string;
1254                 if (baselen != -1 &&
1255                     strncmp(path, list->items[first].string, baselen)) {
1256                         report_helper(list, seen_bits, first, i);
1257                         baselen = -1;
1258                         seen_bits = 0;
1259                 }
1260                 if (baselen == -1) {
1261                         const char *dot = strrchr(path, '.');
1262                         if (!dot) {
1263                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1264                                 continue;
1265                         }
1266                         baselen = dot - path + 1;
1267                         first = i;
1268                 }
1269                 if (!strcmp(path + baselen, "pack"))
1270                         seen_bits |= 1;
1271                 else if (!strcmp(path + baselen, "idx"))
1272                         seen_bits |= 2;
1273         }
1274         report_helper(list, seen_bits, first, list->nr);
1275 }
1276
1277 static void prepare_packed_git_one(char *objdir, int local)
1278 {
1279         struct strbuf path = STRBUF_INIT;
1280         size_t dirnamelen;
1281         DIR *dir;
1282         struct dirent *de;
1283         struct string_list garbage = STRING_LIST_INIT_DUP;
1284
1285         strbuf_addstr(&path, objdir);
1286         strbuf_addstr(&path, "/pack");
1287         dir = opendir(path.buf);
1288         if (!dir) {
1289                 if (errno != ENOENT)
1290                         error("unable to open object pack directory: %s: %s",
1291                               path.buf, strerror(errno));
1292                 strbuf_release(&path);
1293                 return;
1294         }
1295         strbuf_addch(&path, '/');
1296         dirnamelen = path.len;
1297         while ((de = readdir(dir)) != NULL) {
1298                 struct packed_git *p;
1299                 size_t base_len;
1300
1301                 if (is_dot_or_dotdot(de->d_name))
1302                         continue;
1303
1304                 strbuf_setlen(&path, dirnamelen);
1305                 strbuf_addstr(&path, de->d_name);
1306
1307                 base_len = path.len;
1308                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1309                         /* Don't reopen a pack we already have. */
1310                         for (p = packed_git; p; p = p->next) {
1311                                 size_t len;
1312                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1313                                     len == base_len &&
1314                                     !memcmp(p->pack_name, path.buf, len))
1315                                         break;
1316                         }
1317                         if (p == NULL &&
1318                             /*
1319                              * See if it really is a valid .idx file with
1320                              * corresponding .pack file that we can map.
1321                              */
1322                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1323                                 install_packed_git(p);
1324                 }
1325
1326                 if (!report_garbage)
1327                         continue;
1328
1329                 if (ends_with(de->d_name, ".idx") ||
1330                     ends_with(de->d_name, ".pack") ||
1331                     ends_with(de->d_name, ".bitmap") ||
1332                     ends_with(de->d_name, ".keep"))
1333                         string_list_append(&garbage, path.buf);
1334                 else
1335                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1336         }
1337         closedir(dir);
1338         report_pack_garbage(&garbage);
1339         string_list_clear(&garbage, 0);
1340         strbuf_release(&path);
1341 }
1342
1343 static int sort_pack(const void *a_, const void *b_)
1344 {
1345         struct packed_git *a = *((struct packed_git **)a_);
1346         struct packed_git *b = *((struct packed_git **)b_);
1347         int st;
1348
1349         /*
1350          * Local packs tend to contain objects specific to our
1351          * variant of the project than remote ones.  In addition,
1352          * remote ones could be on a network mounted filesystem.
1353          * Favor local ones for these reasons.
1354          */
1355         st = a->pack_local - b->pack_local;
1356         if (st)
1357                 return -st;
1358
1359         /*
1360          * Younger packs tend to contain more recent objects,
1361          * and more recent objects tend to get accessed more
1362          * often.
1363          */
1364         if (a->mtime < b->mtime)
1365                 return 1;
1366         else if (a->mtime == b->mtime)
1367                 return 0;
1368         return -1;
1369 }
1370
1371 static void rearrange_packed_git(void)
1372 {
1373         struct packed_git **ary, *p;
1374         int i, n;
1375
1376         for (n = 0, p = packed_git; p; p = p->next)
1377                 n++;
1378         if (n < 2)
1379                 return;
1380
1381         /* prepare an array of packed_git for easier sorting */
1382         ary = xcalloc(n, sizeof(struct packed_git *));
1383         for (n = 0, p = packed_git; p; p = p->next)
1384                 ary[n++] = p;
1385
1386         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1387
1388         /* link them back again */
1389         for (i = 0; i < n - 1; i++)
1390                 ary[i]->next = ary[i + 1];
1391         ary[n - 1]->next = NULL;
1392         packed_git = ary[0];
1393
1394         free(ary);
1395 }
1396
1397 static int prepare_packed_git_run_once = 0;
1398 void prepare_packed_git(void)
1399 {
1400         struct alternate_object_database *alt;
1401
1402         if (prepare_packed_git_run_once)
1403                 return;
1404         prepare_packed_git_one(get_object_directory(), 1);
1405         prepare_alt_odb();
1406         for (alt = alt_odb_list; alt; alt = alt->next) {
1407                 alt->name[-1] = 0;
1408                 prepare_packed_git_one(alt->base, 0);
1409                 alt->name[-1] = '/';
1410         }
1411         rearrange_packed_git();
1412         prepare_packed_git_run_once = 1;
1413 }
1414
1415 void reprepare_packed_git(void)
1416 {
1417         prepare_packed_git_run_once = 0;
1418         prepare_packed_git();
1419 }
1420
1421 static void mark_bad_packed_object(struct packed_git *p,
1422                                    const unsigned char *sha1)
1423 {
1424         unsigned i;
1425         for (i = 0; i < p->num_bad_objects; i++)
1426                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1427                         return;
1428         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1429         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1430         p->num_bad_objects++;
1431 }
1432
1433 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1434 {
1435         struct packed_git *p;
1436         unsigned i;
1437
1438         for (p = packed_git; p; p = p->next)
1439                 for (i = 0; i < p->num_bad_objects; i++)
1440                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1441                                 return p;
1442         return NULL;
1443 }
1444
1445 /*
1446  * With an in-core object data in "map", rehash it to make sure the
1447  * object name actually matches "sha1" to detect object corruption.
1448  * With "map" == NULL, try reading the object named with "sha1" using
1449  * the streaming interface and rehash it to do the same.
1450  */
1451 int check_sha1_signature(const unsigned char *sha1, void *map,
1452                          unsigned long size, const char *type)
1453 {
1454         unsigned char real_sha1[20];
1455         enum object_type obj_type;
1456         struct git_istream *st;
1457         git_SHA_CTX c;
1458         char hdr[32];
1459         int hdrlen;
1460
1461         if (map) {
1462                 hash_sha1_file(map, size, type, real_sha1);
1463                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1464         }
1465
1466         st = open_istream(sha1, &obj_type, &size, NULL);
1467         if (!st)
1468                 return -1;
1469
1470         /* Generate the header */
1471         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1472
1473         /* Sha1.. */
1474         git_SHA1_Init(&c);
1475         git_SHA1_Update(&c, hdr, hdrlen);
1476         for (;;) {
1477                 char buf[1024 * 16];
1478                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1479
1480                 if (readlen < 0) {
1481                         close_istream(st);
1482                         return -1;
1483                 }
1484                 if (!readlen)
1485                         break;
1486                 git_SHA1_Update(&c, buf, readlen);
1487         }
1488         git_SHA1_Final(real_sha1, &c);
1489         close_istream(st);
1490         return hashcmp(sha1, real_sha1) ? -1 : 0;
1491 }
1492
1493 int git_open_noatime(const char *name)
1494 {
1495         static int sha1_file_open_flag = O_NOATIME;
1496
1497         for (;;) {
1498                 int fd;
1499
1500                 errno = 0;
1501                 fd = open(name, O_RDONLY | sha1_file_open_flag);
1502                 if (fd >= 0)
1503                         return fd;
1504
1505                 /* Might the failure be due to O_NOATIME? */
1506                 if (errno != ENOENT && sha1_file_open_flag) {
1507                         sha1_file_open_flag = 0;
1508                         continue;
1509                 }
1510
1511                 return -1;
1512         }
1513 }
1514
1515 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1516 {
1517         struct alternate_object_database *alt;
1518
1519         if (!lstat(sha1_file_name(sha1), st))
1520                 return 0;
1521
1522         prepare_alt_odb();
1523         errno = ENOENT;
1524         for (alt = alt_odb_list; alt; alt = alt->next) {
1525                 fill_sha1_path(alt->name, sha1);
1526                 if (!lstat(alt->base, st))
1527                         return 0;
1528         }
1529
1530         return -1;
1531 }
1532
1533 static int open_sha1_file(const unsigned char *sha1)
1534 {
1535         int fd;
1536         struct alternate_object_database *alt;
1537         int most_interesting_errno;
1538
1539         fd = git_open_noatime(sha1_file_name(sha1));
1540         if (fd >= 0)
1541                 return fd;
1542         most_interesting_errno = errno;
1543
1544         prepare_alt_odb();
1545         for (alt = alt_odb_list; alt; alt = alt->next) {
1546                 fill_sha1_path(alt->name, sha1);
1547                 fd = git_open_noatime(alt->base);
1548                 if (fd >= 0)
1549                         return fd;
1550                 if (most_interesting_errno == ENOENT)
1551                         most_interesting_errno = errno;
1552         }
1553         errno = most_interesting_errno;
1554         return -1;
1555 }
1556
1557 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1558 {
1559         void *map;
1560         int fd;
1561
1562         fd = open_sha1_file(sha1);
1563         map = NULL;
1564         if (fd >= 0) {
1565                 struct stat st;
1566
1567                 if (!fstat(fd, &st)) {
1568                         *size = xsize_t(st.st_size);
1569                         if (!*size) {
1570                                 /* mmap() is forbidden on empty files */
1571                                 error("object file %s is empty", sha1_file_name(sha1));
1572                                 return NULL;
1573                         }
1574                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1575                 }
1576                 close(fd);
1577         }
1578         return map;
1579 }
1580
1581 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1582                 unsigned long len, enum object_type *type, unsigned long *sizep)
1583 {
1584         unsigned shift;
1585         unsigned long size, c;
1586         unsigned long used = 0;
1587
1588         c = buf[used++];
1589         *type = (c >> 4) & 7;
1590         size = c & 15;
1591         shift = 4;
1592         while (c & 0x80) {
1593                 if (len <= used || bitsizeof(long) <= shift) {
1594                         error("bad object header");
1595                         size = used = 0;
1596                         break;
1597                 }
1598                 c = buf[used++];
1599                 size += (c & 0x7f) << shift;
1600                 shift += 7;
1601         }
1602         *sizep = size;
1603         return used;
1604 }
1605
1606 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1607 {
1608         /* Get the data stream */
1609         memset(stream, 0, sizeof(*stream));
1610         stream->next_in = map;
1611         stream->avail_in = mapsize;
1612         stream->next_out = buffer;
1613         stream->avail_out = bufsiz;
1614
1615         git_inflate_init(stream);
1616         return git_inflate(stream, 0);
1617 }
1618
1619 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1620                                         unsigned long mapsize, void *buffer,
1621                                         unsigned long bufsiz, struct strbuf *header)
1622 {
1623         int status;
1624
1625         status = unpack_sha1_header(stream, map, mapsize, buffer, bufsiz);
1626
1627         /*
1628          * Check if entire header is unpacked in the first iteration.
1629          */
1630         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1631                 return 0;
1632
1633         /*
1634          * buffer[0..bufsiz] was not large enough.  Copy the partial
1635          * result out to header, and then append the result of further
1636          * reading the stream.
1637          */
1638         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1639         stream->next_out = buffer;
1640         stream->avail_out = bufsiz;
1641
1642         do {
1643                 status = git_inflate(stream, 0);
1644                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1645                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1646                         return 0;
1647                 stream->next_out = buffer;
1648                 stream->avail_out = bufsiz;
1649         } while (status != Z_STREAM_END);
1650         return -1;
1651 }
1652
1653 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1654 {
1655         int bytes = strlen(buffer) + 1;
1656         unsigned char *buf = xmallocz(size);
1657         unsigned long n;
1658         int status = Z_OK;
1659
1660         n = stream->total_out - bytes;
1661         if (n > size)
1662                 n = size;
1663         memcpy(buf, (char *) buffer + bytes, n);
1664         bytes = n;
1665         if (bytes <= size) {
1666                 /*
1667                  * The above condition must be (bytes <= size), not
1668                  * (bytes < size).  In other words, even though we
1669                  * expect no more output and set avail_out to zero,
1670                  * the input zlib stream may have bytes that express
1671                  * "this concludes the stream", and we *do* want to
1672                  * eat that input.
1673                  *
1674                  * Otherwise we would not be able to test that we
1675                  * consumed all the input to reach the expected size;
1676                  * we also want to check that zlib tells us that all
1677                  * went well with status == Z_STREAM_END at the end.
1678                  */
1679                 stream->next_out = buf + bytes;
1680                 stream->avail_out = size - bytes;
1681                 while (status == Z_OK)
1682                         status = git_inflate(stream, Z_FINISH);
1683         }
1684         if (status == Z_STREAM_END && !stream->avail_in) {
1685                 git_inflate_end(stream);
1686                 return buf;
1687         }
1688
1689         if (status < 0)
1690                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1691         else if (stream->avail_in)
1692                 error("garbage at end of loose object '%s'",
1693                       sha1_to_hex(sha1));
1694         free(buf);
1695         return NULL;
1696 }
1697
1698 /*
1699  * We used to just use "sscanf()", but that's actually way
1700  * too permissive for what we want to check. So do an anal
1701  * object header parse by hand.
1702  */
1703 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1704                                unsigned int flags)
1705 {
1706         const char *type_buf = hdr;
1707         unsigned long size;
1708         int type, type_len = 0;
1709
1710         /*
1711          * The type can be of any size but is followed by
1712          * a space.
1713          */
1714         for (;;) {
1715                 char c = *hdr++;
1716                 if (c == ' ')
1717                         break;
1718                 type_len++;
1719         }
1720
1721         type = type_from_string_gently(type_buf, type_len, 1);
1722         if (oi->typename)
1723                 strbuf_add(oi->typename, type_buf, type_len);
1724         /*
1725          * Set type to 0 if its an unknown object and
1726          * we're obtaining the type using '--allow-unkown-type'
1727          * option.
1728          */
1729         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1730                 type = 0;
1731         else if (type < 0)
1732                 die("invalid object type");
1733         if (oi->typep)
1734                 *oi->typep = type;
1735
1736         /*
1737          * The length must follow immediately, and be in canonical
1738          * decimal format (ie "010" is not valid).
1739          */
1740         size = *hdr++ - '0';
1741         if (size > 9)
1742                 return -1;
1743         if (size) {
1744                 for (;;) {
1745                         unsigned long c = *hdr - '0';
1746                         if (c > 9)
1747                                 break;
1748                         hdr++;
1749                         size = size * 10 + c;
1750                 }
1751         }
1752
1753         if (oi->sizep)
1754                 *oi->sizep = size;
1755
1756         /*
1757          * The length must be followed by a zero byte
1758          */
1759         return *hdr ? -1 : type;
1760 }
1761
1762 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1763 {
1764         struct object_info oi;
1765
1766         oi.sizep = sizep;
1767         oi.typename = NULL;
1768         oi.typep = NULL;
1769         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
1770 }
1771
1772 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1773 {
1774         int ret;
1775         git_zstream stream;
1776         char hdr[8192];
1777
1778         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1779         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1780                 return NULL;
1781
1782         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1783 }
1784
1785 unsigned long get_size_from_delta(struct packed_git *p,
1786                                   struct pack_window **w_curs,
1787                                   off_t curpos)
1788 {
1789         const unsigned char *data;
1790         unsigned char delta_head[20], *in;
1791         git_zstream stream;
1792         int st;
1793
1794         memset(&stream, 0, sizeof(stream));
1795         stream.next_out = delta_head;
1796         stream.avail_out = sizeof(delta_head);
1797
1798         git_inflate_init(&stream);
1799         do {
1800                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1801                 stream.next_in = in;
1802                 st = git_inflate(&stream, Z_FINISH);
1803                 curpos += stream.next_in - in;
1804         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1805                  stream.total_out < sizeof(delta_head));
1806         git_inflate_end(&stream);
1807         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1808                 error("delta data unpack-initial failed");
1809                 return 0;
1810         }
1811
1812         /* Examine the initial part of the delta to figure out
1813          * the result size.
1814          */
1815         data = delta_head;
1816
1817         /* ignore base size */
1818         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1819
1820         /* Read the result size */
1821         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1822 }
1823
1824 static off_t get_delta_base(struct packed_git *p,
1825                                     struct pack_window **w_curs,
1826                                     off_t *curpos,
1827                                     enum object_type type,
1828                                     off_t delta_obj_offset)
1829 {
1830         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1831         off_t base_offset;
1832
1833         /* use_pack() assured us we have [base_info, base_info + 20)
1834          * as a range that we can look at without walking off the
1835          * end of the mapped window.  Its actually the hash size
1836          * that is assured.  An OFS_DELTA longer than the hash size
1837          * is stupid, as then a REF_DELTA would be smaller to store.
1838          */
1839         if (type == OBJ_OFS_DELTA) {
1840                 unsigned used = 0;
1841                 unsigned char c = base_info[used++];
1842                 base_offset = c & 127;
1843                 while (c & 128) {
1844                         base_offset += 1;
1845                         if (!base_offset || MSB(base_offset, 7))
1846                                 return 0;  /* overflow */
1847                         c = base_info[used++];
1848                         base_offset = (base_offset << 7) + (c & 127);
1849                 }
1850                 base_offset = delta_obj_offset - base_offset;
1851                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1852                         return 0;  /* out of bound */
1853                 *curpos += used;
1854         } else if (type == OBJ_REF_DELTA) {
1855                 /* The base entry _must_ be in the same pack */
1856                 base_offset = find_pack_entry_one(base_info, p);
1857                 *curpos += 20;
1858         } else
1859                 die("I am totally screwed");
1860         return base_offset;
1861 }
1862
1863 /*
1864  * Like get_delta_base above, but we return the sha1 instead of the pack
1865  * offset. This means it is cheaper for REF deltas (we do not have to do
1866  * the final object lookup), but more expensive for OFS deltas (we
1867  * have to load the revidx to convert the offset back into a sha1).
1868  */
1869 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1870                                                 struct pack_window **w_curs,
1871                                                 off_t curpos,
1872                                                 enum object_type type,
1873                                                 off_t delta_obj_offset)
1874 {
1875         if (type == OBJ_REF_DELTA) {
1876                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1877                 return base;
1878         } else if (type == OBJ_OFS_DELTA) {
1879                 struct revindex_entry *revidx;
1880                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1881                                                    type, delta_obj_offset);
1882
1883                 if (!base_offset)
1884                         return NULL;
1885
1886                 revidx = find_pack_revindex(p, base_offset);
1887                 if (!revidx)
1888                         return NULL;
1889
1890                 return nth_packed_object_sha1(p, revidx->nr);
1891         } else
1892                 return NULL;
1893 }
1894
1895 int unpack_object_header(struct packed_git *p,
1896                          struct pack_window **w_curs,
1897                          off_t *curpos,
1898                          unsigned long *sizep)
1899 {
1900         unsigned char *base;
1901         unsigned long left;
1902         unsigned long used;
1903         enum object_type type;
1904
1905         /* use_pack() assures us we have [base, base + 20) available
1906          * as a range that we can look at.  (Its actually the hash
1907          * size that is assured.)  With our object header encoding
1908          * the maximum deflated object size is 2^137, which is just
1909          * insane, so we know won't exceed what we have been given.
1910          */
1911         base = use_pack(p, w_curs, *curpos, &left);
1912         used = unpack_object_header_buffer(base, left, &type, sizep);
1913         if (!used) {
1914                 type = OBJ_BAD;
1915         } else
1916                 *curpos += used;
1917
1918         return type;
1919 }
1920
1921 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1922 {
1923         int type;
1924         struct revindex_entry *revidx;
1925         const unsigned char *sha1;
1926         revidx = find_pack_revindex(p, obj_offset);
1927         if (!revidx)
1928                 return OBJ_BAD;
1929         sha1 = nth_packed_object_sha1(p, revidx->nr);
1930         mark_bad_packed_object(p, sha1);
1931         type = sha1_object_info(sha1, NULL);
1932         if (type <= OBJ_NONE)
1933                 return OBJ_BAD;
1934         return type;
1935 }
1936
1937 #define POI_STACK_PREALLOC 64
1938
1939 static enum object_type packed_to_object_type(struct packed_git *p,
1940                                               off_t obj_offset,
1941                                               enum object_type type,
1942                                               struct pack_window **w_curs,
1943                                               off_t curpos)
1944 {
1945         off_t small_poi_stack[POI_STACK_PREALLOC];
1946         off_t *poi_stack = small_poi_stack;
1947         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1948
1949         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1950                 off_t base_offset;
1951                 unsigned long size;
1952                 /* Push the object we're going to leave behind */
1953                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1954                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1955                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1956                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1957                 } else {
1958                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1959                 }
1960                 poi_stack[poi_stack_nr++] = obj_offset;
1961                 /* If parsing the base offset fails, just unwind */
1962                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1963                 if (!base_offset)
1964                         goto unwind;
1965                 curpos = obj_offset = base_offset;
1966                 type = unpack_object_header(p, w_curs, &curpos, &size);
1967                 if (type <= OBJ_NONE) {
1968                         /* If getting the base itself fails, we first
1969                          * retry the base, otherwise unwind */
1970                         type = retry_bad_packed_offset(p, base_offset);
1971                         if (type > OBJ_NONE)
1972                                 goto out;
1973                         goto unwind;
1974                 }
1975         }
1976
1977         switch (type) {
1978         case OBJ_BAD:
1979         case OBJ_COMMIT:
1980         case OBJ_TREE:
1981         case OBJ_BLOB:
1982         case OBJ_TAG:
1983                 break;
1984         default:
1985                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1986                       type, (uintmax_t)obj_offset, p->pack_name);
1987                 type = OBJ_BAD;
1988         }
1989
1990 out:
1991         if (poi_stack != small_poi_stack)
1992                 free(poi_stack);
1993         return type;
1994
1995 unwind:
1996         while (poi_stack_nr) {
1997                 obj_offset = poi_stack[--poi_stack_nr];
1998                 type = retry_bad_packed_offset(p, obj_offset);
1999                 if (type > OBJ_NONE)
2000                         goto out;
2001         }
2002         type = OBJ_BAD;
2003         goto out;
2004 }
2005
2006 static int packed_object_info(struct packed_git *p, off_t obj_offset,
2007                               struct object_info *oi)
2008 {
2009         struct pack_window *w_curs = NULL;
2010         unsigned long size;
2011         off_t curpos = obj_offset;
2012         enum object_type type;
2013
2014         /*
2015          * We always get the representation type, but only convert it to
2016          * a "real" type later if the caller is interested.
2017          */
2018         type = unpack_object_header(p, &w_curs, &curpos, &size);
2019
2020         if (oi->sizep) {
2021                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2022                         off_t tmp_pos = curpos;
2023                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2024                                                            type, obj_offset);
2025                         if (!base_offset) {
2026                                 type = OBJ_BAD;
2027                                 goto out;
2028                         }
2029                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2030                         if (*oi->sizep == 0) {
2031                                 type = OBJ_BAD;
2032                                 goto out;
2033                         }
2034                 } else {
2035                         *oi->sizep = size;
2036                 }
2037         }
2038
2039         if (oi->disk_sizep) {
2040                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2041                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2042         }
2043
2044         if (oi->typep) {
2045                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2046                 if (*oi->typep < 0) {
2047                         type = OBJ_BAD;
2048                         goto out;
2049                 }
2050         }
2051
2052         if (oi->delta_base_sha1) {
2053                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2054                         const unsigned char *base;
2055
2056                         base = get_delta_base_sha1(p, &w_curs, curpos,
2057                                                    type, obj_offset);
2058                         if (!base) {
2059                                 type = OBJ_BAD;
2060                                 goto out;
2061                         }
2062
2063                         hashcpy(oi->delta_base_sha1, base);
2064                 } else
2065                         hashclr(oi->delta_base_sha1);
2066         }
2067
2068 out:
2069         unuse_pack(&w_curs);
2070         return type;
2071 }
2072
2073 static void *unpack_compressed_entry(struct packed_git *p,
2074                                     struct pack_window **w_curs,
2075                                     off_t curpos,
2076                                     unsigned long size)
2077 {
2078         int st;
2079         git_zstream stream;
2080         unsigned char *buffer, *in;
2081
2082         buffer = xmallocz_gently(size);
2083         if (!buffer)
2084                 return NULL;
2085         memset(&stream, 0, sizeof(stream));
2086         stream.next_out = buffer;
2087         stream.avail_out = size + 1;
2088
2089         git_inflate_init(&stream);
2090         do {
2091                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2092                 stream.next_in = in;
2093                 st = git_inflate(&stream, Z_FINISH);
2094                 if (!stream.avail_out)
2095                         break; /* the payload is larger than it should be */
2096                 curpos += stream.next_in - in;
2097         } while (st == Z_OK || st == Z_BUF_ERROR);
2098         git_inflate_end(&stream);
2099         if ((st != Z_STREAM_END) || stream.total_out != size) {
2100                 free(buffer);
2101                 return NULL;
2102         }
2103
2104         return buffer;
2105 }
2106
2107 #define MAX_DELTA_CACHE (256)
2108
2109 static size_t delta_base_cached;
2110
2111 static struct delta_base_cache_lru_list {
2112         struct delta_base_cache_lru_list *prev;
2113         struct delta_base_cache_lru_list *next;
2114 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
2115
2116 static struct delta_base_cache_entry {
2117         struct delta_base_cache_lru_list lru;
2118         void *data;
2119         struct packed_git *p;
2120         off_t base_offset;
2121         unsigned long size;
2122         enum object_type type;
2123 } delta_base_cache[MAX_DELTA_CACHE];
2124
2125 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
2126 {
2127         unsigned long hash;
2128
2129         hash = (unsigned long)(intptr_t)p + (unsigned long)base_offset;
2130         hash += (hash >> 8) + (hash >> 16);
2131         return hash % MAX_DELTA_CACHE;
2132 }
2133
2134 static struct delta_base_cache_entry *
2135 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2136 {
2137         unsigned long hash = pack_entry_hash(p, base_offset);
2138         return delta_base_cache + hash;
2139 }
2140
2141 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
2142                                      struct packed_git *p, off_t base_offset)
2143 {
2144         return (ent->data && ent->p == p && ent->base_offset == base_offset);
2145 }
2146
2147 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2148 {
2149         struct delta_base_cache_entry *ent;
2150         ent = get_delta_base_cache_entry(p, base_offset);
2151         return eq_delta_base_cache_entry(ent, p, base_offset);
2152 }
2153
2154 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2155 {
2156         ent->data = NULL;
2157         ent->lru.next->prev = ent->lru.prev;
2158         ent->lru.prev->next = ent->lru.next;
2159         delta_base_cached -= ent->size;
2160 }
2161
2162 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2163         unsigned long *base_size, enum object_type *type, int keep_cache)
2164 {
2165         struct delta_base_cache_entry *ent;
2166         void *ret;
2167
2168         ent = get_delta_base_cache_entry(p, base_offset);
2169
2170         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2171                 return unpack_entry(p, base_offset, type, base_size);
2172
2173         ret = ent->data;
2174
2175         if (!keep_cache)
2176                 clear_delta_base_cache_entry(ent);
2177         else
2178                 ret = xmemdupz(ent->data, ent->size);
2179         *type = ent->type;
2180         *base_size = ent->size;
2181         return ret;
2182 }
2183
2184 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2185 {
2186         if (ent->data) {
2187                 free(ent->data);
2188                 ent->data = NULL;
2189                 ent->lru.next->prev = ent->lru.prev;
2190                 ent->lru.prev->next = ent->lru.next;
2191                 delta_base_cached -= ent->size;
2192         }
2193 }
2194
2195 void clear_delta_base_cache(void)
2196 {
2197         unsigned long p;
2198         for (p = 0; p < MAX_DELTA_CACHE; p++)
2199                 release_delta_base_cache(&delta_base_cache[p]);
2200 }
2201
2202 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2203         void *base, unsigned long base_size, enum object_type type)
2204 {
2205         unsigned long hash = pack_entry_hash(p, base_offset);
2206         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2207         struct delta_base_cache_lru_list *lru;
2208
2209         release_delta_base_cache(ent);
2210         delta_base_cached += base_size;
2211
2212         for (lru = delta_base_cache_lru.next;
2213              delta_base_cached > delta_base_cache_limit
2214              && lru != &delta_base_cache_lru;
2215              lru = lru->next) {
2216                 struct delta_base_cache_entry *f = (void *)lru;
2217                 if (f->type == OBJ_BLOB)
2218                         release_delta_base_cache(f);
2219         }
2220         for (lru = delta_base_cache_lru.next;
2221              delta_base_cached > delta_base_cache_limit
2222              && lru != &delta_base_cache_lru;
2223              lru = lru->next) {
2224                 struct delta_base_cache_entry *f = (void *)lru;
2225                 release_delta_base_cache(f);
2226         }
2227
2228         ent->p = p;
2229         ent->base_offset = base_offset;
2230         ent->type = type;
2231         ent->data = base;
2232         ent->size = base_size;
2233         ent->lru.next = &delta_base_cache_lru;
2234         ent->lru.prev = delta_base_cache_lru.prev;
2235         delta_base_cache_lru.prev->next = &ent->lru;
2236         delta_base_cache_lru.prev = &ent->lru;
2237 }
2238
2239 static void *read_object(const unsigned char *sha1, enum object_type *type,
2240                          unsigned long *size);
2241
2242 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2243 {
2244         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2245         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2246                          p->pack_name, (uintmax_t)obj_offset);
2247 }
2248
2249 int do_check_packed_object_crc;
2250
2251 #define UNPACK_ENTRY_STACK_PREALLOC 64
2252 struct unpack_entry_stack_ent {
2253         off_t obj_offset;
2254         off_t curpos;
2255         unsigned long size;
2256 };
2257
2258 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2259                    enum object_type *final_type, unsigned long *final_size)
2260 {
2261         struct pack_window *w_curs = NULL;
2262         off_t curpos = obj_offset;
2263         void *data = NULL;
2264         unsigned long size;
2265         enum object_type type;
2266         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2267         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2268         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2269         int base_from_cache = 0;
2270
2271         write_pack_access_log(p, obj_offset);
2272
2273         /* PHASE 1: drill down to the innermost base object */
2274         for (;;) {
2275                 off_t base_offset;
2276                 int i;
2277                 struct delta_base_cache_entry *ent;
2278
2279                 ent = get_delta_base_cache_entry(p, curpos);
2280                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2281                         type = ent->type;
2282                         data = ent->data;
2283                         size = ent->size;
2284                         clear_delta_base_cache_entry(ent);
2285                         base_from_cache = 1;
2286                         break;
2287                 }
2288
2289                 if (do_check_packed_object_crc && p->index_version > 1) {
2290                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2291                         unsigned long len = revidx[1].offset - obj_offset;
2292                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2293                                 const unsigned char *sha1 =
2294                                         nth_packed_object_sha1(p, revidx->nr);
2295                                 error("bad packed object CRC for %s",
2296                                       sha1_to_hex(sha1));
2297                                 mark_bad_packed_object(p, sha1);
2298                                 unuse_pack(&w_curs);
2299                                 return NULL;
2300                         }
2301                 }
2302
2303                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2304                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2305                         break;
2306
2307                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2308                 if (!base_offset) {
2309                         error("failed to validate delta base reference "
2310                               "at offset %"PRIuMAX" from %s",
2311                               (uintmax_t)curpos, p->pack_name);
2312                         /* bail to phase 2, in hopes of recovery */
2313                         data = NULL;
2314                         break;
2315                 }
2316
2317                 /* push object, proceed to base */
2318                 if (delta_stack_nr >= delta_stack_alloc
2319                     && delta_stack == small_delta_stack) {
2320                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2321                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2322                         memcpy(delta_stack, small_delta_stack,
2323                                sizeof(*delta_stack)*delta_stack_nr);
2324                 } else {
2325                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2326                 }
2327                 i = delta_stack_nr++;
2328                 delta_stack[i].obj_offset = obj_offset;
2329                 delta_stack[i].curpos = curpos;
2330                 delta_stack[i].size = size;
2331
2332                 curpos = obj_offset = base_offset;
2333         }
2334
2335         /* PHASE 2: handle the base */
2336         switch (type) {
2337         case OBJ_OFS_DELTA:
2338         case OBJ_REF_DELTA:
2339                 if (data)
2340                         die("BUG in unpack_entry: left loop at a valid delta");
2341                 break;
2342         case OBJ_COMMIT:
2343         case OBJ_TREE:
2344         case OBJ_BLOB:
2345         case OBJ_TAG:
2346                 if (!base_from_cache)
2347                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2348                 break;
2349         default:
2350                 data = NULL;
2351                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2352                       type, (uintmax_t)obj_offset, p->pack_name);
2353         }
2354
2355         /* PHASE 3: apply deltas in order */
2356
2357         /* invariants:
2358          *   'data' holds the base data, or NULL if there was corruption
2359          */
2360         while (delta_stack_nr) {
2361                 void *delta_data;
2362                 void *base = data;
2363                 unsigned long delta_size, base_size = size;
2364                 int i;
2365
2366                 data = NULL;
2367
2368                 if (base)
2369                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2370
2371                 if (!base) {
2372                         /*
2373                          * We're probably in deep shit, but let's try to fetch
2374                          * the required base anyway from another pack or loose.
2375                          * This is costly but should happen only in the presence
2376                          * of a corrupted pack, and is better than failing outright.
2377                          */
2378                         struct revindex_entry *revidx;
2379                         const unsigned char *base_sha1;
2380                         revidx = find_pack_revindex(p, obj_offset);
2381                         if (revidx) {
2382                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2383                                 error("failed to read delta base object %s"
2384                                       " at offset %"PRIuMAX" from %s",
2385                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2386                                       p->pack_name);
2387                                 mark_bad_packed_object(p, base_sha1);
2388                                 base = read_object(base_sha1, &type, &base_size);
2389                         }
2390                 }
2391
2392                 i = --delta_stack_nr;
2393                 obj_offset = delta_stack[i].obj_offset;
2394                 curpos = delta_stack[i].curpos;
2395                 delta_size = delta_stack[i].size;
2396
2397                 if (!base)
2398                         continue;
2399
2400                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2401
2402                 if (!delta_data) {
2403                         error("failed to unpack compressed delta "
2404                               "at offset %"PRIuMAX" from %s",
2405                               (uintmax_t)curpos, p->pack_name);
2406                         data = NULL;
2407                         continue;
2408                 }
2409
2410                 data = patch_delta(base, base_size,
2411                                    delta_data, delta_size,
2412                                    &size);
2413
2414                 /*
2415                  * We could not apply the delta; warn the user, but keep going.
2416                  * Our failure will be noticed either in the next iteration of
2417                  * the loop, or if this is the final delta, in the caller when
2418                  * we return NULL. Those code paths will take care of making
2419                  * a more explicit warning and retrying with another copy of
2420                  * the object.
2421                  */
2422                 if (!data)
2423                         error("failed to apply delta");
2424
2425                 free(delta_data);
2426         }
2427
2428         *final_type = type;
2429         *final_size = size;
2430
2431         unuse_pack(&w_curs);
2432
2433         if (delta_stack != small_delta_stack)
2434                 free(delta_stack);
2435
2436         return data;
2437 }
2438
2439 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2440                                             uint32_t n)
2441 {
2442         const unsigned char *index = p->index_data;
2443         if (!index) {
2444                 if (open_pack_index(p))
2445                         return NULL;
2446                 index = p->index_data;
2447         }
2448         if (n >= p->num_objects)
2449                 return NULL;
2450         index += 4 * 256;
2451         if (p->index_version == 1) {
2452                 return index + 24 * n + 4;
2453         } else {
2454                 index += 8;
2455                 return index + 20 * n;
2456         }
2457 }
2458
2459 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2460 {
2461         const unsigned char *index = p->index_data;
2462         index += 4 * 256;
2463         if (p->index_version == 1) {
2464                 return ntohl(*((uint32_t *)(index + 24 * n)));
2465         } else {
2466                 uint32_t off;
2467                 index += 8 + p->num_objects * (20 + 4);
2468                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2469                 if (!(off & 0x80000000))
2470                         return off;
2471                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2472                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2473                                    ntohl(*((uint32_t *)(index + 4)));
2474         }
2475 }
2476
2477 off_t find_pack_entry_one(const unsigned char *sha1,
2478                                   struct packed_git *p)
2479 {
2480         const uint32_t *level1_ofs = p->index_data;
2481         const unsigned char *index = p->index_data;
2482         unsigned hi, lo, stride;
2483         static int use_lookup = -1;
2484         static int debug_lookup = -1;
2485
2486         if (debug_lookup < 0)
2487                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2488
2489         if (!index) {
2490                 if (open_pack_index(p))
2491                         return 0;
2492                 level1_ofs = p->index_data;
2493                 index = p->index_data;
2494         }
2495         if (p->index_version > 1) {
2496                 level1_ofs += 2;
2497                 index += 8;
2498         }
2499         index += 4 * 256;
2500         hi = ntohl(level1_ofs[*sha1]);
2501         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2502         if (p->index_version > 1) {
2503                 stride = 20;
2504         } else {
2505                 stride = 24;
2506                 index += 4;
2507         }
2508
2509         if (debug_lookup)
2510                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2511                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2512
2513         if (use_lookup < 0)
2514                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2515         if (use_lookup) {
2516                 int pos = sha1_entry_pos(index, stride, 0,
2517                                          lo, hi, p->num_objects, sha1);
2518                 if (pos < 0)
2519                         return 0;
2520                 return nth_packed_object_offset(p, pos);
2521         }
2522
2523         do {
2524                 unsigned mi = (lo + hi) / 2;
2525                 int cmp = hashcmp(index + mi * stride, sha1);
2526
2527                 if (debug_lookup)
2528                         printf("lo %u hi %u rg %u mi %u\n",
2529                                lo, hi, hi - lo, mi);
2530                 if (!cmp)
2531                         return nth_packed_object_offset(p, mi);
2532                 if (cmp > 0)
2533                         hi = mi;
2534                 else
2535                         lo = mi+1;
2536         } while (lo < hi);
2537         return 0;
2538 }
2539
2540 int is_pack_valid(struct packed_git *p)
2541 {
2542         /* An already open pack is known to be valid. */
2543         if (p->pack_fd != -1)
2544                 return 1;
2545
2546         /* If the pack has one window completely covering the
2547          * file size, the pack is known to be valid even if
2548          * the descriptor is not currently open.
2549          */
2550         if (p->windows) {
2551                 struct pack_window *w = p->windows;
2552
2553                 if (!w->offset && w->len == p->pack_size)
2554                         return 1;
2555         }
2556
2557         /* Force the pack to open to prove its valid. */
2558         return !open_packed_git(p);
2559 }
2560
2561 static int fill_pack_entry(const unsigned char *sha1,
2562                            struct pack_entry *e,
2563                            struct packed_git *p)
2564 {
2565         off_t offset;
2566
2567         if (p->num_bad_objects) {
2568                 unsigned i;
2569                 for (i = 0; i < p->num_bad_objects; i++)
2570                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2571                                 return 0;
2572         }
2573
2574         offset = find_pack_entry_one(sha1, p);
2575         if (!offset)
2576                 return 0;
2577
2578         /*
2579          * We are about to tell the caller where they can locate the
2580          * requested object.  We better make sure the packfile is
2581          * still here and can be accessed before supplying that
2582          * answer, as it may have been deleted since the index was
2583          * loaded!
2584          */
2585         if (!is_pack_valid(p))
2586                 return 0;
2587         e->offset = offset;
2588         e->p = p;
2589         hashcpy(e->sha1, sha1);
2590         return 1;
2591 }
2592
2593 /*
2594  * Iff a pack file contains the object named by sha1, return true and
2595  * store its location to e.
2596  */
2597 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2598 {
2599         struct packed_git *p;
2600
2601         prepare_packed_git();
2602         if (!packed_git)
2603                 return 0;
2604
2605         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2606                 return 1;
2607
2608         for (p = packed_git; p; p = p->next) {
2609                 if (p == last_found_pack)
2610                         continue; /* we already checked this one */
2611
2612                 if (fill_pack_entry(sha1, e, p)) {
2613                         last_found_pack = p;
2614                         return 1;
2615                 }
2616         }
2617         return 0;
2618 }
2619
2620 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2621                                   struct packed_git *packs)
2622 {
2623         struct packed_git *p;
2624
2625         for (p = packs; p; p = p->next) {
2626                 if (find_pack_entry_one(sha1, p))
2627                         return p;
2628         }
2629         return NULL;
2630
2631 }
2632
2633 static int sha1_loose_object_info(const unsigned char *sha1,
2634                                   struct object_info *oi,
2635                                   int flags)
2636 {
2637         int status = 0;
2638         unsigned long mapsize;
2639         void *map;
2640         git_zstream stream;
2641         char hdr[32];
2642         struct strbuf hdrbuf = STRBUF_INIT;
2643
2644         if (oi->delta_base_sha1)
2645                 hashclr(oi->delta_base_sha1);
2646
2647         /*
2648          * If we don't care about type or size, then we don't
2649          * need to look inside the object at all. Note that we
2650          * do not optimize out the stat call, even if the
2651          * caller doesn't care about the disk-size, since our
2652          * return value implicitly indicates whether the
2653          * object even exists.
2654          */
2655         if (!oi->typep && !oi->typename && !oi->sizep) {
2656                 struct stat st;
2657                 if (stat_sha1_file(sha1, &st) < 0)
2658                         return -1;
2659                 if (oi->disk_sizep)
2660                         *oi->disk_sizep = st.st_size;
2661                 return 0;
2662         }
2663
2664         map = map_sha1_file(sha1, &mapsize);
2665         if (!map)
2666                 return -1;
2667         if (oi->disk_sizep)
2668                 *oi->disk_sizep = mapsize;
2669         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2670                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2671                         status = error("unable to unpack %s header with --allow-unknown-type",
2672                                        sha1_to_hex(sha1));
2673         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2674                 status = error("unable to unpack %s header",
2675                                sha1_to_hex(sha1));
2676         if (status < 0)
2677                 ; /* Do nothing */
2678         else if (hdrbuf.len) {
2679                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2680                         status = error("unable to parse %s header with --allow-unknown-type",
2681                                        sha1_to_hex(sha1));
2682         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2683                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2684         git_inflate_end(&stream);
2685         munmap(map, mapsize);
2686         if (status && oi->typep)
2687                 *oi->typep = status;
2688         strbuf_release(&hdrbuf);
2689         return 0;
2690 }
2691
2692 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2693 {
2694         struct cached_object *co;
2695         struct pack_entry e;
2696         int rtype;
2697         enum object_type real_type;
2698         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2699
2700         co = find_cached_object(real);
2701         if (co) {
2702                 if (oi->typep)
2703                         *(oi->typep) = co->type;
2704                 if (oi->sizep)
2705                         *(oi->sizep) = co->size;
2706                 if (oi->disk_sizep)
2707                         *(oi->disk_sizep) = 0;
2708                 if (oi->delta_base_sha1)
2709                         hashclr(oi->delta_base_sha1);
2710                 if (oi->typename)
2711                         strbuf_addstr(oi->typename, typename(co->type));
2712                 oi->whence = OI_CACHED;
2713                 return 0;
2714         }
2715
2716         if (!find_pack_entry(real, &e)) {
2717                 /* Most likely it's a loose object. */
2718                 if (!sha1_loose_object_info(real, oi, flags)) {
2719                         oi->whence = OI_LOOSE;
2720                         return 0;
2721                 }
2722
2723                 /* Not a loose object; someone else may have just packed it. */
2724                 reprepare_packed_git();
2725                 if (!find_pack_entry(real, &e))
2726                         return -1;
2727         }
2728
2729         /*
2730          * packed_object_info() does not follow the delta chain to
2731          * find out the real type, unless it is given oi->typep.
2732          */
2733         if (oi->typename && !oi->typep)
2734                 oi->typep = &real_type;
2735
2736         rtype = packed_object_info(e.p, e.offset, oi);
2737         if (rtype < 0) {
2738                 mark_bad_packed_object(e.p, real);
2739                 if (oi->typep == &real_type)
2740                         oi->typep = NULL;
2741                 return sha1_object_info_extended(real, oi, 0);
2742         } else if (in_delta_base_cache(e.p, e.offset)) {
2743                 oi->whence = OI_DBCACHED;
2744         } else {
2745                 oi->whence = OI_PACKED;
2746                 oi->u.packed.offset = e.offset;
2747                 oi->u.packed.pack = e.p;
2748                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2749                                          rtype == OBJ_OFS_DELTA);
2750         }
2751         if (oi->typename)
2752                 strbuf_addstr(oi->typename, typename(*oi->typep));
2753         if (oi->typep == &real_type)
2754                 oi->typep = NULL;
2755
2756         return 0;
2757 }
2758
2759 /* returns enum object_type or negative */
2760 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2761 {
2762         enum object_type type;
2763         struct object_info oi = {NULL};
2764
2765         oi.typep = &type;
2766         oi.sizep = sizep;
2767         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2768                 return -1;
2769         return type;
2770 }
2771
2772 static void *read_packed_sha1(const unsigned char *sha1,
2773                               enum object_type *type, unsigned long *size)
2774 {
2775         struct pack_entry e;
2776         void *data;
2777
2778         if (!find_pack_entry(sha1, &e))
2779                 return NULL;
2780         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2781         if (!data) {
2782                 /*
2783                  * We're probably in deep shit, but let's try to fetch
2784                  * the required object anyway from another pack or loose.
2785                  * This should happen only in the presence of a corrupted
2786                  * pack, and is better than failing outright.
2787                  */
2788                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2789                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2790                 mark_bad_packed_object(e.p, sha1);
2791                 data = read_object(sha1, type, size);
2792         }
2793         return data;
2794 }
2795
2796 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2797                       unsigned char *sha1)
2798 {
2799         struct cached_object *co;
2800
2801         hash_sha1_file(buf, len, typename(type), sha1);
2802         if (has_sha1_file(sha1) || find_cached_object(sha1))
2803                 return 0;
2804         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2805         co = &cached_objects[cached_object_nr++];
2806         co->size = len;
2807         co->type = type;
2808         co->buf = xmalloc(len);
2809         memcpy(co->buf, buf, len);
2810         hashcpy(co->sha1, sha1);
2811         return 0;
2812 }
2813
2814 static void *read_object(const unsigned char *sha1, enum object_type *type,
2815                          unsigned long *size)
2816 {
2817         unsigned long mapsize;
2818         void *map, *buf;
2819         struct cached_object *co;
2820
2821         co = find_cached_object(sha1);
2822         if (co) {
2823                 *type = co->type;
2824                 *size = co->size;
2825                 return xmemdupz(co->buf, co->size);
2826         }
2827
2828         buf = read_packed_sha1(sha1, type, size);
2829         if (buf)
2830                 return buf;
2831         map = map_sha1_file(sha1, &mapsize);
2832         if (map) {
2833                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2834                 munmap(map, mapsize);
2835                 return buf;
2836         }
2837         reprepare_packed_git();
2838         return read_packed_sha1(sha1, type, size);
2839 }
2840
2841 /*
2842  * This function dies on corrupt objects; the callers who want to
2843  * deal with them should arrange to call read_object() and give error
2844  * messages themselves.
2845  */
2846 void *read_sha1_file_extended(const unsigned char *sha1,
2847                               enum object_type *type,
2848                               unsigned long *size,
2849                               unsigned flag)
2850 {
2851         void *data;
2852         const struct packed_git *p;
2853         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2854
2855         errno = 0;
2856         data = read_object(repl, type, size);
2857         if (data)
2858                 return data;
2859
2860         if (errno && errno != ENOENT)
2861                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2862
2863         /* die if we replaced an object with one that does not exist */
2864         if (repl != sha1)
2865                 die("replacement %s not found for %s",
2866                     sha1_to_hex(repl), sha1_to_hex(sha1));
2867
2868         if (has_loose_object(repl)) {
2869                 const char *path = sha1_file_name(sha1);
2870
2871                 die("loose object %s (stored in %s) is corrupt",
2872                     sha1_to_hex(repl), path);
2873         }
2874
2875         if ((p = has_packed_and_bad(repl)) != NULL)
2876                 die("packed object %s (stored in %s) is corrupt",
2877                     sha1_to_hex(repl), p->pack_name);
2878
2879         return NULL;
2880 }
2881
2882 void *read_object_with_reference(const unsigned char *sha1,
2883                                  const char *required_type_name,
2884                                  unsigned long *size,
2885                                  unsigned char *actual_sha1_return)
2886 {
2887         enum object_type type, required_type;
2888         void *buffer;
2889         unsigned long isize;
2890         unsigned char actual_sha1[20];
2891
2892         required_type = type_from_string(required_type_name);
2893         hashcpy(actual_sha1, sha1);
2894         while (1) {
2895                 int ref_length = -1;
2896                 const char *ref_type = NULL;
2897
2898                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2899                 if (!buffer)
2900                         return NULL;
2901                 if (type == required_type) {
2902                         *size = isize;
2903                         if (actual_sha1_return)
2904                                 hashcpy(actual_sha1_return, actual_sha1);
2905                         return buffer;
2906                 }
2907                 /* Handle references */
2908                 else if (type == OBJ_COMMIT)
2909                         ref_type = "tree ";
2910                 else if (type == OBJ_TAG)
2911                         ref_type = "object ";
2912                 else {
2913                         free(buffer);
2914                         return NULL;
2915                 }
2916                 ref_length = strlen(ref_type);
2917
2918                 if (ref_length + 40 > isize ||
2919                     memcmp(buffer, ref_type, ref_length) ||
2920                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2921                         free(buffer);
2922                         return NULL;
2923                 }
2924                 free(buffer);
2925                 /* Now we have the ID of the referred-to object in
2926                  * actual_sha1.  Check again. */
2927         }
2928 }
2929
2930 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2931                                     const char *type, unsigned char *sha1,
2932                                     char *hdr, int *hdrlen)
2933 {
2934         git_SHA_CTX c;
2935
2936         /* Generate the header */
2937         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2938
2939         /* Sha1.. */
2940         git_SHA1_Init(&c);
2941         git_SHA1_Update(&c, hdr, *hdrlen);
2942         git_SHA1_Update(&c, buf, len);
2943         git_SHA1_Final(sha1, &c);
2944 }
2945
2946 /*
2947  * Move the just written object into its final resting place.
2948  */
2949 int finalize_object_file(const char *tmpfile, const char *filename)
2950 {
2951         int ret = 0;
2952
2953         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2954                 goto try_rename;
2955         else if (link(tmpfile, filename))
2956                 ret = errno;
2957
2958         /*
2959          * Coda hack - coda doesn't like cross-directory links,
2960          * so we fall back to a rename, which will mean that it
2961          * won't be able to check collisions, but that's not a
2962          * big deal.
2963          *
2964          * The same holds for FAT formatted media.
2965          *
2966          * When this succeeds, we just return.  We have nothing
2967          * left to unlink.
2968          */
2969         if (ret && ret != EEXIST) {
2970         try_rename:
2971                 if (!rename(tmpfile, filename))
2972                         goto out;
2973                 ret = errno;
2974         }
2975         unlink_or_warn(tmpfile);
2976         if (ret) {
2977                 if (ret != EEXIST) {
2978                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2979                 }
2980                 /* FIXME!!! Collision check here ? */
2981         }
2982
2983 out:
2984         if (adjust_shared_perm(filename))
2985                 return error("unable to set permission to '%s'", filename);
2986         return 0;
2987 }
2988
2989 static int write_buffer(int fd, const void *buf, size_t len)
2990 {
2991         if (write_in_full(fd, buf, len) < 0)
2992                 return error("file write error (%s)", strerror(errno));
2993         return 0;
2994 }
2995
2996 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2997                    unsigned char *sha1)
2998 {
2999         char hdr[32];
3000         int hdrlen;
3001         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3002         return 0;
3003 }
3004
3005 /* Finalize a file on disk, and close it. */
3006 static void close_sha1_file(int fd)
3007 {
3008         if (fsync_object_files)
3009                 fsync_or_die(fd, "sha1 file");
3010         if (close(fd) != 0)
3011                 die_errno("error when closing sha1 file");
3012 }
3013
3014 /* Size of directory component, including the ending '/' */
3015 static inline int directory_size(const char *filename)
3016 {
3017         const char *s = strrchr(filename, '/');
3018         if (!s)
3019                 return 0;
3020         return s - filename + 1;
3021 }
3022
3023 /*
3024  * This creates a temporary file in the same directory as the final
3025  * 'filename'
3026  *
3027  * We want to avoid cross-directory filename renames, because those
3028  * can have problems on various filesystems (FAT, NFS, Coda).
3029  */
3030 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
3031 {
3032         int fd, dirlen = directory_size(filename);
3033
3034         if (dirlen + 20 > bufsiz) {
3035                 errno = ENAMETOOLONG;
3036                 return -1;
3037         }
3038         memcpy(buffer, filename, dirlen);
3039         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
3040         fd = git_mkstemp_mode(buffer, 0444);
3041         if (fd < 0 && dirlen && errno == ENOENT) {
3042                 /* Make sure the directory exists */
3043                 memcpy(buffer, filename, dirlen);
3044                 buffer[dirlen-1] = 0;
3045                 if (mkdir(buffer, 0777) && errno != EEXIST)
3046                         return -1;
3047                 if (adjust_shared_perm(buffer))
3048                         return -1;
3049
3050                 /* Try again */
3051                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
3052                 fd = git_mkstemp_mode(buffer, 0444);
3053         }
3054         return fd;
3055 }
3056
3057 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3058                               const void *buf, unsigned long len, time_t mtime)
3059 {
3060         int fd, ret;
3061         unsigned char compressed[4096];
3062         git_zstream stream;
3063         git_SHA_CTX c;
3064         unsigned char parano_sha1[20];
3065         static char tmp_file[PATH_MAX];
3066         const char *filename = sha1_file_name(sha1);
3067
3068         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
3069         if (fd < 0) {
3070                 if (errno == EACCES)
3071                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3072                 else
3073                         return error("unable to create temporary file: %s", strerror(errno));
3074         }
3075
3076         /* Set it up */
3077         git_deflate_init(&stream, zlib_compression_level);
3078         stream.next_out = compressed;
3079         stream.avail_out = sizeof(compressed);
3080         git_SHA1_Init(&c);
3081
3082         /* First header.. */
3083         stream.next_in = (unsigned char *)hdr;
3084         stream.avail_in = hdrlen;
3085         while (git_deflate(&stream, 0) == Z_OK)
3086                 ; /* nothing */
3087         git_SHA1_Update(&c, hdr, hdrlen);
3088
3089         /* Then the data itself.. */
3090         stream.next_in = (void *)buf;
3091         stream.avail_in = len;
3092         do {
3093                 unsigned char *in0 = stream.next_in;
3094                 ret = git_deflate(&stream, Z_FINISH);
3095                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3096                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3097                         die("unable to write sha1 file");
3098                 stream.next_out = compressed;
3099                 stream.avail_out = sizeof(compressed);
3100         } while (ret == Z_OK);
3101
3102         if (ret != Z_STREAM_END)
3103                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3104         ret = git_deflate_end_gently(&stream);
3105         if (ret != Z_OK)
3106                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3107         git_SHA1_Final(parano_sha1, &c);
3108         if (hashcmp(sha1, parano_sha1) != 0)
3109                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3110
3111         close_sha1_file(fd);
3112
3113         if (mtime) {
3114                 struct utimbuf utb;
3115                 utb.actime = mtime;
3116                 utb.modtime = mtime;
3117                 if (utime(tmp_file, &utb) < 0)
3118                         warning("failed utime() on %s: %s",
3119                                 tmp_file, strerror(errno));
3120         }
3121
3122         return finalize_object_file(tmp_file, filename);
3123 }
3124
3125 static int freshen_loose_object(const unsigned char *sha1)
3126 {
3127         return check_and_freshen(sha1, 1);
3128 }
3129
3130 static int freshen_packed_object(const unsigned char *sha1)
3131 {
3132         struct pack_entry e;
3133         if (!find_pack_entry(sha1, &e))
3134                 return 0;
3135         if (e.p->freshened)
3136                 return 1;
3137         if (!freshen_file(e.p->pack_name))
3138                 return 0;
3139         e.p->freshened = 1;
3140         return 1;
3141 }
3142
3143 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3144 {
3145         char hdr[32];
3146         int hdrlen;
3147
3148         /* Normally if we have it in the pack then we do not bother writing
3149          * it out into .git/objects/??/?{38} file.
3150          */
3151         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3152         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3153                 return 0;
3154         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3155 }
3156
3157 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3158                              unsigned char *sha1, unsigned flags)
3159 {
3160         char *header;
3161         int hdrlen, status = 0;
3162
3163         /* type string, SP, %lu of the length plus NUL must fit this */
3164         header = xmalloc(strlen(type) + 32);
3165         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3166
3167         if (!(flags & HASH_WRITE_OBJECT))
3168                 goto cleanup;
3169         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3170                 goto cleanup;
3171         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3172
3173 cleanup:
3174         free(header);
3175         return status;
3176 }
3177
3178 int force_object_loose(const unsigned char *sha1, time_t mtime)
3179 {
3180         void *buf;
3181         unsigned long len;
3182         enum object_type type;
3183         char hdr[32];
3184         int hdrlen;
3185         int ret;
3186
3187         if (has_loose_object(sha1))
3188                 return 0;
3189         buf = read_packed_sha1(sha1, &type, &len);
3190         if (!buf)
3191                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3192         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
3193         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3194         free(buf);
3195
3196         return ret;
3197 }
3198
3199 int has_pack_index(const unsigned char *sha1)
3200 {
3201         struct stat st;
3202         if (stat(sha1_pack_index_name(sha1), &st))
3203                 return 0;
3204         return 1;
3205 }
3206
3207 int has_sha1_pack(const unsigned char *sha1)
3208 {
3209         struct pack_entry e;
3210         return find_pack_entry(sha1, &e);
3211 }
3212
3213 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3214 {
3215         struct pack_entry e;
3216
3217         if (find_pack_entry(sha1, &e))
3218                 return 1;
3219         if (has_loose_object(sha1))
3220                 return 1;
3221         if (flags & HAS_SHA1_QUICK)
3222                 return 0;
3223         reprepare_packed_git();
3224         return find_pack_entry(sha1, &e);
3225 }
3226
3227 static void check_tree(const void *buf, size_t size)
3228 {
3229         struct tree_desc desc;
3230         struct name_entry entry;
3231
3232         init_tree_desc(&desc, buf, size);
3233         while (tree_entry(&desc, &entry))
3234                 /* do nothing
3235                  * tree_entry() will die() on malformed entries */
3236                 ;
3237 }
3238
3239 static void check_commit(const void *buf, size_t size)
3240 {
3241         struct commit c;
3242         memset(&c, 0, sizeof(c));
3243         if (parse_commit_buffer(&c, buf, size))
3244                 die("corrupt commit");
3245 }
3246
3247 static void check_tag(const void *buf, size_t size)
3248 {
3249         struct tag t;
3250         memset(&t, 0, sizeof(t));
3251         if (parse_tag_buffer(&t, buf, size))
3252                 die("corrupt tag");
3253 }
3254
3255 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3256                      enum object_type type,
3257                      const char *path, unsigned flags)
3258 {
3259         int ret, re_allocated = 0;
3260         int write_object = flags & HASH_WRITE_OBJECT;
3261
3262         if (!type)
3263                 type = OBJ_BLOB;
3264
3265         /*
3266          * Convert blobs to git internal format
3267          */
3268         if ((type == OBJ_BLOB) && path) {
3269                 struct strbuf nbuf = STRBUF_INIT;
3270                 if (convert_to_git(path, buf, size, &nbuf,
3271                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3272                         buf = strbuf_detach(&nbuf, &size);
3273                         re_allocated = 1;
3274                 }
3275         }
3276         if (flags & HASH_FORMAT_CHECK) {
3277                 if (type == OBJ_TREE)
3278                         check_tree(buf, size);
3279                 if (type == OBJ_COMMIT)
3280                         check_commit(buf, size);
3281                 if (type == OBJ_TAG)
3282                         check_tag(buf, size);
3283         }
3284
3285         if (write_object)
3286                 ret = write_sha1_file(buf, size, typename(type), sha1);
3287         else
3288                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3289         if (re_allocated)
3290                 free(buf);
3291         return ret;
3292 }
3293
3294 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3295                                      const char *path, unsigned flags)
3296 {
3297         int ret;
3298         const int write_object = flags & HASH_WRITE_OBJECT;
3299         struct strbuf sbuf = STRBUF_INIT;
3300
3301         assert(path);
3302         assert(would_convert_to_git_filter_fd(path));
3303
3304         convert_to_git_filter_fd(path, fd, &sbuf,
3305                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3306
3307         if (write_object)
3308                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3309                                       sha1);
3310         else
3311                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3312                                      sha1);
3313         strbuf_release(&sbuf);
3314         return ret;
3315 }
3316
3317 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3318                       const char *path, unsigned flags)
3319 {
3320         struct strbuf sbuf = STRBUF_INIT;
3321         int ret;
3322
3323         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3324                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3325         else
3326                 ret = -1;
3327         strbuf_release(&sbuf);
3328         return ret;
3329 }
3330
3331 #define SMALL_FILE_SIZE (32*1024)
3332
3333 static int index_core(unsigned char *sha1, int fd, size_t size,
3334                       enum object_type type, const char *path,
3335                       unsigned flags)
3336 {
3337         int ret;
3338
3339         if (!size) {
3340                 ret = index_mem(sha1, "", size, type, path, flags);
3341         } else if (size <= SMALL_FILE_SIZE) {
3342                 char *buf = xmalloc(size);
3343                 if (size == read_in_full(fd, buf, size))
3344                         ret = index_mem(sha1, buf, size, type, path, flags);
3345                 else
3346                         ret = error("short read %s", strerror(errno));
3347                 free(buf);
3348         } else {
3349                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3350                 ret = index_mem(sha1, buf, size, type, path, flags);
3351                 munmap(buf, size);
3352         }
3353         return ret;
3354 }
3355
3356 /*
3357  * This creates one packfile per large blob unless bulk-checkin
3358  * machinery is "plugged".
3359  *
3360  * This also bypasses the usual "convert-to-git" dance, and that is on
3361  * purpose. We could write a streaming version of the converting
3362  * functions and insert that before feeding the data to fast-import
3363  * (or equivalent in-core API described above). However, that is
3364  * somewhat complicated, as we do not know the size of the filter
3365  * result, which we need to know beforehand when writing a git object.
3366  * Since the primary motivation for trying to stream from the working
3367  * tree file and to avoid mmaping it in core is to deal with large
3368  * binary blobs, they generally do not want to get any conversion, and
3369  * callers should avoid this code path when filters are requested.
3370  */
3371 static int index_stream(unsigned char *sha1, int fd, size_t size,
3372                         enum object_type type, const char *path,
3373                         unsigned flags)
3374 {
3375         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3376 }
3377
3378 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3379              enum object_type type, const char *path, unsigned flags)
3380 {
3381         int ret;
3382
3383         /*
3384          * Call xsize_t() only when needed to avoid potentially unnecessary
3385          * die() for large files.
3386          */
3387         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3388                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3389         else if (!S_ISREG(st->st_mode))
3390                 ret = index_pipe(sha1, fd, type, path, flags);
3391         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3392                  (path && would_convert_to_git(path)))
3393                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3394                                  flags);
3395         else
3396                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3397                                    flags);
3398         close(fd);
3399         return ret;
3400 }
3401
3402 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3403 {
3404         int fd;
3405         struct strbuf sb = STRBUF_INIT;
3406
3407         switch (st->st_mode & S_IFMT) {
3408         case S_IFREG:
3409                 fd = open(path, O_RDONLY);
3410                 if (fd < 0)
3411                         return error("open(\"%s\"): %s", path,
3412                                      strerror(errno));
3413                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3414                         return error("%s: failed to insert into database",
3415                                      path);
3416                 break;
3417         case S_IFLNK:
3418                 if (strbuf_readlink(&sb, path, st->st_size)) {
3419                         char *errstr = strerror(errno);
3420                         return error("readlink(\"%s\"): %s", path,
3421                                      errstr);
3422                 }
3423                 if (!(flags & HASH_WRITE_OBJECT))
3424                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3425                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3426                         return error("%s: failed to insert into database",
3427                                      path);
3428                 strbuf_release(&sb);
3429                 break;
3430         case S_IFDIR:
3431                 return resolve_gitlink_ref(path, "HEAD", sha1);
3432         default:
3433                 return error("%s: unsupported file type", path);
3434         }
3435         return 0;
3436 }
3437
3438 int read_pack_header(int fd, struct pack_header *header)
3439 {
3440         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3441                 /* "eof before pack header was fully read" */
3442                 return PH_ERROR_EOF;
3443
3444         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3445                 /* "protocol error (pack signature mismatch detected)" */
3446                 return PH_ERROR_PACK_SIGNATURE;
3447         if (!pack_version_ok(header->hdr_version))
3448                 /* "protocol error (pack version unsupported)" */
3449                 return PH_ERROR_PROTOCOL;
3450         return 0;
3451 }
3452
3453 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3454 {
3455         enum object_type type = sha1_object_info(sha1, NULL);
3456         if (type < 0)
3457                 die("%s is not a valid object", sha1_to_hex(sha1));
3458         if (type != expect)
3459                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3460                     typename(expect));
3461 }
3462
3463 static int for_each_file_in_obj_subdir(int subdir_nr,
3464                                        struct strbuf *path,
3465                                        each_loose_object_fn obj_cb,
3466                                        each_loose_cruft_fn cruft_cb,
3467                                        each_loose_subdir_fn subdir_cb,
3468                                        void *data)
3469 {
3470         size_t baselen = path->len;
3471         DIR *dir = opendir(path->buf);
3472         struct dirent *de;
3473         int r = 0;
3474
3475         if (!dir) {
3476                 if (errno == ENOENT)
3477                         return 0;
3478                 return error("unable to open %s: %s", path->buf, strerror(errno));
3479         }
3480
3481         while ((de = readdir(dir))) {
3482                 if (is_dot_or_dotdot(de->d_name))
3483                         continue;
3484
3485                 strbuf_setlen(path, baselen);
3486                 strbuf_addf(path, "/%s", de->d_name);
3487
3488                 if (strlen(de->d_name) == 38)  {
3489                         char hex[41];
3490                         unsigned char sha1[20];
3491
3492                         snprintf(hex, sizeof(hex), "%02x%s",
3493                                  subdir_nr, de->d_name);
3494                         if (!get_sha1_hex(hex, sha1)) {
3495                                 if (obj_cb) {
3496                                         r = obj_cb(sha1, path->buf, data);
3497                                         if (r)
3498                                                 break;
3499                                 }
3500                                 continue;
3501                         }
3502                 }
3503
3504                 if (cruft_cb) {
3505                         r = cruft_cb(de->d_name, path->buf, data);
3506                         if (r)
3507                                 break;
3508                 }
3509         }
3510         strbuf_setlen(path, baselen);
3511
3512         if (!r && subdir_cb)
3513                 r = subdir_cb(subdir_nr, path->buf, data);
3514
3515         closedir(dir);
3516         return r;
3517 }
3518
3519 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3520                             each_loose_object_fn obj_cb,
3521                             each_loose_cruft_fn cruft_cb,
3522                             each_loose_subdir_fn subdir_cb,
3523                             void *data)
3524 {
3525         size_t baselen = path->len;
3526         int r = 0;
3527         int i;
3528
3529         for (i = 0; i < 256; i++) {
3530                 strbuf_addf(path, "/%02x", i);
3531                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3532                                                 subdir_cb, data);
3533                 strbuf_setlen(path, baselen);
3534                 if (r)
3535                         break;
3536         }
3537
3538         return r;
3539 }
3540
3541 int for_each_loose_file_in_objdir(const char *path,
3542                                   each_loose_object_fn obj_cb,
3543                                   each_loose_cruft_fn cruft_cb,
3544                                   each_loose_subdir_fn subdir_cb,
3545                                   void *data)
3546 {
3547         struct strbuf buf = STRBUF_INIT;
3548         int r;
3549
3550         strbuf_addstr(&buf, path);
3551         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3552                                               subdir_cb, data);
3553         strbuf_release(&buf);
3554
3555         return r;
3556 }
3557
3558 struct loose_alt_odb_data {
3559         each_loose_object_fn *cb;
3560         void *data;
3561 };
3562
3563 static int loose_from_alt_odb(struct alternate_object_database *alt,
3564                               void *vdata)
3565 {
3566         struct loose_alt_odb_data *data = vdata;
3567         struct strbuf buf = STRBUF_INIT;
3568         int r;
3569
3570         /* copy base not including trailing '/' */
3571         strbuf_add(&buf, alt->base, alt->name - alt->base - 1);
3572         r = for_each_loose_file_in_objdir_buf(&buf,
3573                                               data->cb, NULL, NULL,
3574                                               data->data);
3575         strbuf_release(&buf);
3576         return r;
3577 }
3578
3579 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3580 {
3581         struct loose_alt_odb_data alt;
3582         int r;
3583
3584         r = for_each_loose_file_in_objdir(get_object_directory(),
3585                                           cb, NULL, NULL, data);
3586         if (r)
3587                 return r;
3588
3589         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3590                 return 0;
3591
3592         alt.cb = cb;
3593         alt.data = data;
3594         return foreach_alt_odb(loose_from_alt_odb, &alt);
3595 }
3596
3597 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3598 {
3599         uint32_t i;
3600         int r = 0;
3601
3602         for (i = 0; i < p->num_objects; i++) {
3603                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3604
3605                 if (!sha1)
3606                         return error("unable to get sha1 of object %u in %s",
3607                                      i, p->pack_name);
3608
3609                 r = cb(sha1, p, i, data);
3610                 if (r)
3611                         break;
3612         }
3613         return r;
3614 }
3615
3616 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3617 {
3618         struct packed_git *p;
3619         int r = 0;
3620         int pack_errors = 0;
3621
3622         prepare_packed_git();
3623         for (p = packed_git; p; p = p->next) {
3624                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3625                         continue;
3626                 if (open_pack_index(p)) {
3627                         pack_errors = 1;
3628                         continue;
3629                 }
3630                 r = for_each_object_in_pack(p, cb, data);
3631                 if (r)
3632                         break;
3633         }
3634         return r ? r : pack_errors;
3635 }