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 #include "mru.h"
  27 #include "list.h"
  28 #include "mergesort.h"
  29 #include "quote.h"
  30
  31 #define SZ_FMT PRIuMAX
  32 static inline uintmax_t sz_fmt(size_t s) { return s; }
  33
  34 const unsigned char null_sha1[20];
  35 const struct object_id null_oid;
  36 const struct object_id empty_tree_oid = {
  37         EMPTY_TREE_SHA1_BIN_LITERAL
  38 };
  39 const struct object_id empty_blob_oid = {
  40         EMPTY_BLOB_SHA1_BIN_LITERAL
  41 };
  42
  43 /*
  44  * This is meant to hold a *small* number of objects that you would
  45  * want read_sha1_file() to be able to return, but yet you do not want
  46  * to write them into the object store (e.g. a browse-only
  47  * application).
  48  */
  49 static struct cached_object {
  50         unsigned char sha1[20];
  51         enum object_type type;
  52         void *buf;
  53         unsigned long size;
  54 } *cached_objects;
  55 static int cached_object_nr, cached_object_alloc;
  56
  57 static struct cached_object empty_tree = {
  58         EMPTY_TREE_SHA1_BIN_LITERAL,
  59         OBJ_TREE,
  60         "",
  61         0
  62 };
  63
  64 static struct cached_object *find_cached_object(const unsigned char *sha1)
  65 {
  66         int i;
  67         struct cached_object *co = cached_objects;
  68
  69         for (i = 0; i < cached_object_nr; i++, co++) {
  70                 if (!hashcmp(co->sha1, sha1))
  71                         return co;
  72         }
  73         if (!hashcmp(sha1, empty_tree.sha1))
  74                 return &empty_tree;
  75         return NULL;
  76 }
  77
  78 int mkdir_in_gitdir(const char *path)
  79 {
  80         if (mkdir(path, 0777)) {
  81                 int saved_errno = errno;
  82                 struct stat st;
  83                 struct strbuf sb = STRBUF_INIT;
  84
  85                 if (errno != EEXIST)
  86                         return -1;
  87                 /*
  88                  * Are we looking at a path in a symlinked worktree
  89                  * whose original repository does not yet have it?
  90                  * e.g. .git/rr-cache pointing at its original
  91                  * repository in which the user hasn't performed any
  92                  * conflict resolution yet?
  93                  */
  94                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
  95                     strbuf_readlink(&sb, path, st.st_size) ||
  96                     !is_absolute_path(sb.buf) ||
  97                     mkdir(sb.buf, 0777)) {
  98                         strbuf_release(&sb);
  99                         errno = saved_errno;
 100                         return -1;
 101                 }
 102                 strbuf_release(&sb);
 103         }
 104         return adjust_shared_perm(path);
 105 }
 106
 107 enum scld_error safe_create_leading_directories(char *path)
 108 {
 109         char *next_component = path + offset_1st_component(path);
 110         enum scld_error ret = SCLD_OK;
 111
 112         while (ret == SCLD_OK && next_component) {
 113                 struct stat st;
 114                 char *slash = next_component, slash_character;
 115
 116                 while (*slash && !is_dir_sep(*slash))
 117                         slash++;
 118
 119                 if (!*slash)
 120                         break;
 121
 122                 next_component = slash + 1;
 123                 while (is_dir_sep(*next_component))
 124                         next_component++;
 125                 if (!*next_component)
 126                         break;
 127
 128                 slash_character = *slash;
 129                 *slash = '\0';
 130                 if (!stat(path, &st)) {
 131                         /* path exists */
 132                         if (!S_ISDIR(st.st_mode))
 133                                 ret = SCLD_EXISTS;
 134                 } else if (mkdir(path, 0777)) {
 135                         if (errno == EEXIST &&
 136                             !stat(path, &st) && S_ISDIR(st.st_mode))
 137                                 ; /* somebody created it since we checked */
 138                         else if (errno == ENOENT)
 139                                 /*
 140                                  * Either mkdir() failed because
 141                                  * somebody just pruned the containing
 142                                  * directory, or stat() failed because
 143                                  * the file that was in our way was
 144                                  * just removed.  Either way, inform
 145                                  * the caller that it might be worth
 146                                  * trying again:
 147                                  */
 148                                 ret = SCLD_VANISHED;
 149                         else
 150                                 ret = SCLD_FAILED;
 151                 } else if (adjust_shared_perm(path)) {
 152                         ret = SCLD_PERMS;
 153                 }
 154                 *slash = slash_character;
 155         }
 156         return ret;
 157 }
 158
 159 enum scld_error safe_create_leading_directories_const(const char *path)
 160 {
 161         /* path points to cache entries, so xstrdup before messing with it */
 162         char *buf = xstrdup(path);
 163         enum scld_error result = safe_create_leading_directories(buf);
 164         free(buf);
 165         return result;
 166 }
 167
 168 static void fill_sha1_path(struct strbuf *buf, const unsigned char *sha1)
 169 {
 170         int i;
 171         for (i = 0; i < 20; i++) {
 172                 static char hex[] = "0123456789abcdef";
 173                 unsigned int val = sha1[i];
 174                 strbuf_addch(buf, hex[val >> 4]);
 175                 strbuf_addch(buf, hex[val & 0xf]);
 176                 if (!i)
 177                         strbuf_addch(buf, '/');
 178         }
 179 }
 180
 181 const char *sha1_file_name(const unsigned char *sha1)
 182 {
 183         static struct strbuf buf = STRBUF_INIT;
 184
 185         strbuf_reset(&buf);
 186         strbuf_addf(&buf, "%s/", get_object_directory());
 187
 188         fill_sha1_path(&buf, sha1);
 189         return buf.buf;
 190 }
 191
 192 struct strbuf *alt_scratch_buf(struct alternate_object_database *alt)
 193 {
 194         strbuf_setlen(&alt->scratch, alt->base_len);
 195         return &alt->scratch;
 196 }
 197
 198 static const char *alt_sha1_path(struct alternate_object_database *alt,
 199                                  const unsigned char *sha1)
 200 {
 201         struct strbuf *buf = alt_scratch_buf(alt);
 202         fill_sha1_path(buf, sha1);
 203         return buf->buf;
 204 }
 205
 206 /*
 207  * Return the name of the pack or index file with the specified sha1
 208  * in its filename.  *base and *name are scratch space that must be
 209  * provided by the caller.  which should be "pack" or "idx".
 210  */
 211 static char *sha1_get_pack_name(const unsigned char *sha1,
 212                                 struct strbuf *buf,
 213                                 const char *which)
 214 {
 215         strbuf_reset(buf);
 216         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
 217                     sha1_to_hex(sha1), which);
 218         return buf->buf;
 219 }
 220
 221 char *sha1_pack_name(const unsigned char *sha1)
 222 {
 223         static struct strbuf buf = STRBUF_INIT;
 224         return sha1_get_pack_name(sha1, &buf, "pack");
 225 }
 226
 227 char *sha1_pack_index_name(const unsigned char *sha1)
 228 {
 229         static struct strbuf buf = STRBUF_INIT;
 230         return sha1_get_pack_name(sha1, &buf, "idx");
 231 }
 232
 233 struct alternate_object_database *alt_odb_list;
 234 static struct alternate_object_database **alt_odb_tail;
 235
 236 /*
 237  * Return non-zero iff the path is usable as an alternate object database.
 238  */
 239 static int alt_odb_usable(struct strbuf *path, const char *normalized_objdir)
 240 {
 241         struct alternate_object_database *alt;
 242
 243         /* Detect cases where alternate disappeared */
 244         if (!is_directory(path->buf)) {
 245                 error("object directory %s does not exist; "
 246                       "check .git/objects/info/alternates.",
 247                       path->buf);
 248                 return 0;
 249         }
 250
 251         /*
 252          * Prevent the common mistake of listing the same
 253          * thing twice, or object directory itself.
 254          */
 255         for (alt = alt_odb_list; alt; alt = alt->next) {
 256                 if (!fspathcmp(path->buf, alt->path))
 257                         return 0;
 258         }
 259         if (!fspathcmp(path->buf, normalized_objdir))
 260                 return 0;
 261
 262         return 1;
 263 }
 264
 265 /*
 266  * Prepare alternate object database registry.
 267  *
 268  * The variable alt_odb_list points at the list of struct
 269  * alternate_object_database.  The elements on this list come from
 270  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 271  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 272  * whose contents is similar to that environment variable but can be
 273  * LF separated.  Its base points at a statically allocated buffer that
 274  * contains "/the/directory/corresponding/to/.git/objects/...", while
 275  * its name points just after the slash at the end of ".git/objects/"
 276  * in the example above, and has enough space to hold 40-byte hex
 277  * SHA1, an extra slash for the first level indirection, and the
 278  * terminating NUL.
 279  */
 280 static int link_alt_odb_entry(const char *entry, const char *relative_base,
 281         int depth, const char *normalized_objdir)
 282 {
 283         struct alternate_object_database *ent;
 284         struct strbuf pathbuf = STRBUF_INIT;
 285
 286         if (!is_absolute_path(entry) && relative_base) {
 287                 strbuf_realpath(&pathbuf, relative_base, 1);
 288                 strbuf_addch(&pathbuf, '/');
 289         }
 290         strbuf_addstr(&pathbuf, entry);
 291
 292         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
 293                 error("unable to normalize alternate object path: %s",
 294                       pathbuf.buf);
 295                 strbuf_release(&pathbuf);
 296                 return -1;
 297         }
 298
 299         /*
 300          * The trailing slash after the directory name is given by
 301          * this function at the end. Remove duplicates.
 302          */
 303         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
 304                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
 305
 306         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
 307                 strbuf_release(&pathbuf);
 308                 return -1;
 309         }
 310
 311         ent = alloc_alt_odb(pathbuf.buf);
 312
 313         /* add the alternate entry */
 314         *alt_odb_tail = ent;
 315         alt_odb_tail = &(ent->next);
 316         ent->next = NULL;
 317
 318         /* recursively add alternates */
 319         read_info_alternates(pathbuf.buf, depth + 1);
 320
 321         strbuf_release(&pathbuf);
 322         return 0;
 323 }
 324
 325 static const char *parse_alt_odb_entry(const char *string,
 326                                        int sep,
 327                                        struct strbuf *out)
 328 {
 329         const char *end;
 330
 331         strbuf_reset(out);
 332
 333         if (*string == '#') {
 334                 /* comment; consume up to next separator */
 335                 end = strchrnul(string, sep);
 336         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
 337                 /*
 338                  * quoted path; unquote_c_style has copied the
 339                  * data for us and set "end". Broken quoting (e.g.,
 340                  * an entry that doesn't end with a quote) falls
 341                  * back to the unquoted case below.
 342                  */
 343         } else {
 344                 /* normal, unquoted path */
 345                 end = strchrnul(string, sep);
 346                 strbuf_add(out, string, end - string);
 347         }
 348
 349         if (*end)
 350                 end++;
 351         return end;
 352 }
 353
 354 static void link_alt_odb_entries(const char *alt, int len, int sep,
 355                                  const char *relative_base, int depth)
 356 {
 357         struct strbuf objdirbuf = STRBUF_INIT;
 358         struct strbuf entry = STRBUF_INIT;
 359
 360         if (depth > 5) {
 361                 error("%s: ignoring alternate object stores, nesting too deep.",
 362                                 relative_base);
 363                 return;
 364         }
 365
 366         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
 367         if (strbuf_normalize_path(&objdirbuf) < 0)
 368                 die("unable to normalize object directory: %s",
 369                     objdirbuf.buf);
 370
 371         while (*alt) {
 372                 alt = parse_alt_odb_entry(alt, sep, &entry);
 373                 if (!entry.len)
 374                         continue;
 375                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
 376         }
 377         strbuf_release(&entry);
 378         strbuf_release(&objdirbuf);
 379 }
 380
 381 void read_info_alternates(const char * relative_base, int depth)
 382 {
 383         char *map;
 384         size_t mapsz;
 385         struct stat st;
 386         char *path;
 387         int fd;
 388
 389         path = xstrfmt("%s/info/alternates", relative_base);
 390         fd = git_open(path);
 391         free(path);
 392         if (fd < 0)
 393                 return;
 394         if (fstat(fd, &st) || (st.st_size == 0)) {
 395                 close(fd);
 396                 return;
 397         }
 398         mapsz = xsize_t(st.st_size);
 399         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 400         close(fd);
 401
 402         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
 403
 404         munmap(map, mapsz);
 405 }
 406
 407 struct alternate_object_database *alloc_alt_odb(const char *dir)
 408 {
 409         struct alternate_object_database *ent;
 410
 411         FLEX_ALLOC_STR(ent, path, dir);
 412         strbuf_init(&ent->scratch, 0);
 413         strbuf_addf(&ent->scratch, "%s/", dir);
 414         ent->base_len = ent->scratch.len;
 415
 416         return ent;
 417 }
 418
 419 void add_to_alternates_file(const char *reference)
 420 {
 421         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 422         char *alts = git_pathdup("objects/info/alternates");
 423         FILE *in, *out;
 424
 425         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
 426         out = fdopen_lock_file(lock, "w");
 427         if (!out)
 428                 die_errno("unable to fdopen alternates lockfile");
 429
 430         in = fopen(alts, "r");
 431         if (in) {
 432                 struct strbuf line = STRBUF_INIT;
 433                 int found = 0;
 434
 435                 while (strbuf_getline(&line, in) != EOF) {
 436                         if (!strcmp(reference, line.buf)) {
 437                                 found = 1;
 438                                 break;
 439                         }
 440                         fprintf_or_die(out, "%s\n", line.buf);
 441                 }
 442
 443                 strbuf_release(&line);
 444                 fclose(in);
 445
 446                 if (found) {
 447                         rollback_lock_file(lock);
 448                         lock = NULL;
 449                 }
 450         }
 451         else if (errno != ENOENT)
 452                 die_errno("unable to read alternates file");
 453
 454         if (lock) {
 455                 fprintf_or_die(out, "%s\n", reference);
 456                 if (commit_lock_file(lock))
 457                         die_errno("unable to move new alternates file into place");
 458                 if (alt_odb_tail)
 459                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
 460         }
 461         free(alts);
 462 }
 463
 464 void add_to_alternates_memory(const char *reference)
 465 {
 466         /*
 467          * Make sure alternates are initialized, or else our entry may be
 468          * overwritten when they are.
 469          */
 470         prepare_alt_odb();
 471
 472         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
 473 }
 474
 475 /*
 476  * Compute the exact path an alternate is at and returns it. In case of
 477  * error NULL is returned and the human readable error is added to `err`
 478  * `path` may be relative and should point to $GITDIR.
 479  * `err` must not be null.
 480  */
 481 char *compute_alternate_path(const char *path, struct strbuf *err)
 482 {
 483         char *ref_git = NULL;
 484         const char *repo, *ref_git_s;
 485         int seen_error = 0;
 486
 487         ref_git_s = real_path_if_valid(path);
 488         if (!ref_git_s) {
 489                 seen_error = 1;
 490                 strbuf_addf(err, _("path '%s' does not exist"), path);
 491                 goto out;
 492         } else
 493                 /*
 494                  * Beware: read_gitfile(), real_path() and mkpath()
 495                  * return static buffer
 496                  */
 497                 ref_git = xstrdup(ref_git_s);
 498
 499         repo = read_gitfile(ref_git);
 500         if (!repo)
 501                 repo = read_gitfile(mkpath("%s/.git", ref_git));
 502         if (repo) {
 503                 free(ref_git);
 504                 ref_git = xstrdup(repo);
 505         }
 506
 507         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
 508                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
 509                 free(ref_git);
 510                 ref_git = ref_git_git;
 511         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
 512                 struct strbuf sb = STRBUF_INIT;
 513                 seen_error = 1;
 514                 if (get_common_dir(&sb, ref_git)) {
 515                         strbuf_addf(err,
 516                                     _("reference repository '%s' as a linked "
 517                                       "checkout is not supported yet."),
 518                                     path);
 519                         goto out;
 520                 }
 521
 522                 strbuf_addf(err, _("reference repository '%s' is not a "
 523                                         "local repository."), path);
 524                 goto out;
 525         }
 526
 527         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
 528                 strbuf_addf(err, _("reference repository '%s' is shallow"),
 529                             path);
 530                 seen_error = 1;
 531                 goto out;
 532         }
 533
 534         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
 535                 strbuf_addf(err,
 536                             _("reference repository '%s' is grafted"),
 537                             path);
 538                 seen_error = 1;
 539                 goto out;
 540         }
 541
 542 out:
 543         if (seen_error) {
 544                 free(ref_git);
 545                 ref_git = NULL;
 546         }
 547
 548         return ref_git;
 549 }
 550
 551 int foreach_alt_odb(alt_odb_fn fn, void *cb)
 552 {
 553         struct alternate_object_database *ent;
 554         int r = 0;
 555
 556         prepare_alt_odb();
 557         for (ent = alt_odb_list; ent; ent = ent->next) {
 558                 r = fn(ent, cb);
 559                 if (r)
 560                         break;
 561         }
 562         return r;
 563 }
 564
 565 void prepare_alt_odb(void)
 566 {
 567         const char *alt;
 568
 569         if (alt_odb_tail)
 570                 return;
 571
 572         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 573         if (!alt) alt = "";
 574
 575         alt_odb_tail = &alt_odb_list;
 576         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
 577
 578         read_info_alternates(get_object_directory(), 0);
 579 }
 580
 581 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
 582 static int freshen_file(const char *fn)
 583 {
 584         struct utimbuf t;
 585         t.actime = t.modtime = time(NULL);
 586         return !utime(fn, &t);
 587 }
 588
 589 /*
 590  * All of the check_and_freshen functions return 1 if the file exists and was
 591  * freshened (if freshening was requested), 0 otherwise. If they return
 592  * 0, you should not assume that it is safe to skip a write of the object (it
 593  * either does not exist on disk, or has a stale mtime and may be subject to
 594  * pruning).
 595  */
 596 static int check_and_freshen_file(const char *fn, int freshen)
 597 {
 598         if (access(fn, F_OK))
 599                 return 0;
 600         if (freshen && !freshen_file(fn))
 601                 return 0;
 602         return 1;
 603 }
 604
 605 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
 606 {
 607         return check_and_freshen_file(sha1_file_name(sha1), freshen);
 608 }
 609
 610 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
 611 {
 612         struct alternate_object_database *alt;
 613         prepare_alt_odb();
 614         for (alt = alt_odb_list; alt; alt = alt->next) {
 615                 const char *path = alt_sha1_path(alt, sha1);
 616                 if (check_and_freshen_file(path, freshen))
 617                         return 1;
 618         }
 619         return 0;
 620 }
 621
 622 static int check_and_freshen(const unsigned char *sha1, int freshen)
 623 {
 624         return check_and_freshen_local(sha1, freshen) ||
 625                check_and_freshen_nonlocal(sha1, freshen);
 626 }
 627
 628 int has_loose_object_nonlocal(const unsigned char *sha1)
 629 {
 630         return check_and_freshen_nonlocal(sha1, 0);
 631 }
 632
 633 static int has_loose_object(const unsigned char *sha1)
 634 {
 635         return check_and_freshen(sha1, 0);
 636 }
 637
 638 static unsigned int pack_used_ctr;
 639 static unsigned int pack_mmap_calls;
 640 static unsigned int peak_pack_open_windows;
 641 static unsigned int pack_open_windows;
 642 static unsigned int pack_open_fds;
 643 static unsigned int pack_max_fds;
 644 static size_t peak_pack_mapped;
 645 static size_t pack_mapped;
 646 struct packed_git *packed_git;
 647
 648 static struct mru packed_git_mru_storage;
 649 struct mru *packed_git_mru = &packed_git_mru_storage;
 650
 651 void pack_report(void)
 652 {
 653         fprintf(stderr,
 654                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 655                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 656                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 657                 sz_fmt(getpagesize()),
 658                 sz_fmt(packed_git_window_size),
 659                 sz_fmt(packed_git_limit));
 660         fprintf(stderr,
 661                 "pack_report: pack_used_ctr            = %10u\n"
 662                 "pack_report: pack_mmap_calls          = %10u\n"
 663                 "pack_report: pack_open_windows        = %10u / %10u\n"
 664                 "pack_report: pack_mapped              = "
 665                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 666                 pack_used_ctr,
 667                 pack_mmap_calls,
 668                 pack_open_windows, peak_pack_open_windows,
 669                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 670 }
 671
 672 /*
 673  * Open and mmap the index file at path, perform a couple of
 674  * consistency checks, then record its information to p.  Return 0 on
 675  * success.
 676  */
 677 static int check_packed_git_idx(const char *path, struct packed_git *p)
 678 {
 679         void *idx_map;
 680         struct pack_idx_header *hdr;
 681         size_t idx_size;
 682         uint32_t version, nr, i, *index;
 683         int fd = git_open(path);
 684         struct stat st;
 685
 686         if (fd < 0)
 687                 return -1;
 688         if (fstat(fd, &st)) {
 689                 close(fd);
 690                 return -1;
 691         }
 692         idx_size = xsize_t(st.st_size);
 693         if (idx_size < 4 * 256 + 20 + 20) {
 694                 close(fd);
 695                 return error("index file %s is too small", path);
 696         }
 697         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 698         close(fd);
 699
 700         hdr = idx_map;
 701         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 702                 version = ntohl(hdr->idx_version);
 703                 if (version < 2 || version > 2) {
 704                         munmap(idx_map, idx_size);
 705                         return error("index file %s is version %"PRIu32
 706                                      " and is not supported by this binary"
 707                                      " (try upgrading GIT to a newer version)",
 708                                      path, version);
 709                 }
 710         } else
 711                 version = 1;
 712
 713         nr = 0;
 714         index = idx_map;
 715         if (version > 1)
 716                 index += 2;  /* skip index header */
 717         for (i = 0; i < 256; i++) {
 718                 uint32_t n = ntohl(index[i]);
 719                 if (n < nr) {
 720                         munmap(idx_map, idx_size);
 721                         return error("non-monotonic index %s", path);
 722                 }
 723                 nr = n;
 724         }
 725
 726         if (version == 1) {
 727                 /*
 728                  * Total size:
 729                  *  - 256 index entries 4 bytes each
 730                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 731                  *  - 20-byte SHA1 of the packfile
 732                  *  - 20-byte SHA1 file checksum
 733                  */
 734                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 735                         munmap(idx_map, idx_size);
 736                         return error("wrong index v1 file size in %s", path);
 737                 }
 738         } else if (version == 2) {
 739                 /*
 740                  * Minimum size:
 741                  *  - 8 bytes of header
 742                  *  - 256 index entries 4 bytes each
 743                  *  - 20-byte sha1 entry * nr
 744                  *  - 4-byte crc entry * nr
 745                  *  - 4-byte offset entry * nr
 746                  *  - 20-byte SHA1 of the packfile
 747                  *  - 20-byte SHA1 file checksum
 748                  * And after the 4-byte offset table might be a
 749                  * variable sized table containing 8-byte entries
 750                  * for offsets larger than 2^31.
 751                  */
 752                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 753                 unsigned long max_size = min_size;
 754                 if (nr)
 755                         max_size += (nr - 1)*8;
 756                 if (idx_size < min_size || idx_size > max_size) {
 757                         munmap(idx_map, idx_size);
 758                         return error("wrong index v2 file size in %s", path);
 759                 }
 760                 if (idx_size != min_size &&
 761                     /*
 762                      * make sure we can deal with large pack offsets.
 763                      * 31-bit signed offset won't be enough, neither
 764                      * 32-bit unsigned one will be.
 765                      */
 766                     (sizeof(off_t) <= 4)) {
 767                         munmap(idx_map, idx_size);
 768                         return error("pack too large for current definition of off_t in %s", path);
 769                 }
 770         }
 771
 772         p->index_version = version;
 773         p->index_data = idx_map;
 774         p->index_size = idx_size;
 775         p->num_objects = nr;
 776         return 0;
 777 }
 778
 779 int open_pack_index(struct packed_git *p)
 780 {
 781         char *idx_name;
 782         size_t len;
 783         int ret;
 784
 785         if (p->index_data)
 786                 return 0;
 787
 788         if (!strip_suffix(p->pack_name, ".pack", &len))
 789                 die("BUG: pack_name does not end in .pack");
 790         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
 791         ret = check_packed_git_idx(idx_name, p);
 792         free(idx_name);
 793         return ret;
 794 }
 795
 796 static void scan_windows(struct packed_git *p,
 797         struct packed_git **lru_p,
 798         struct pack_window **lru_w,
 799         struct pack_window **lru_l)
 800 {
 801         struct pack_window *w, *w_l;
 802
 803         for (w_l = NULL, w = p->windows; w; w = w->next) {
 804                 if (!w->inuse_cnt) {
 805                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 806                                 *lru_p = p;
 807                                 *lru_w = w;
 808                                 *lru_l = w_l;
 809                         }
 810                 }
 811                 w_l = w;
 812         }
 813 }
 814
 815 static int unuse_one_window(struct packed_git *current)
 816 {
 817         struct packed_git *p, *lru_p = NULL;
 818         struct pack_window *lru_w = NULL, *lru_l = NULL;
 819
 820         if (current)
 821                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 822         for (p = packed_git; p; p = p->next)
 823                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 824         if (lru_p) {
 825                 munmap(lru_w->base, lru_w->len);
 826                 pack_mapped -= lru_w->len;
 827                 if (lru_l)
 828                         lru_l->next = lru_w->next;
 829                 else
 830                         lru_p->windows = lru_w->next;
 831                 free(lru_w);
 832                 pack_open_windows--;
 833                 return 1;
 834         }
 835         return 0;
 836 }
 837
 838 void release_pack_memory(size_t need)
 839 {
 840         size_t cur = pack_mapped;
 841         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
 842                 ; /* nothing */
 843 }
 844
 845 static void mmap_limit_check(size_t length)
 846 {
 847         static size_t limit = 0;
 848         if (!limit) {
 849                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
 850                 if (!limit)
 851                         limit = SIZE_MAX;
 852         }
 853         if (length > limit)
 854                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
 855                     (uintmax_t)length, (uintmax_t)limit);
 856 }
 857
 858 void *xmmap_gently(void *start, size_t length,
 859                   int prot, int flags, int fd, off_t offset)
 860 {
 861         void *ret;
 862
 863         mmap_limit_check(length);
 864         ret = mmap(start, length, prot, flags, fd, offset);
 865         if (ret == MAP_FAILED) {
 866                 if (!length)
 867                         return NULL;
 868                 release_pack_memory(length);
 869                 ret = mmap(start, length, prot, flags, fd, offset);
 870         }
 871         return ret;
 872 }
 873
 874 void *xmmap(void *start, size_t length,
 875         int prot, int flags, int fd, off_t offset)
 876 {
 877         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
 878         if (ret == MAP_FAILED)
 879                 die_errno("mmap failed");
 880         return ret;
 881 }
 882
 883 void close_pack_windows(struct packed_git *p)
 884 {
 885         while (p->windows) {
 886                 struct pack_window *w = p->windows;
 887
 888                 if (w->inuse_cnt)
 889                         die("pack '%s' still has open windows to it",
 890                             p->pack_name);
 891                 munmap(w->base, w->len);
 892                 pack_mapped -= w->len;
 893                 pack_open_windows--;
 894                 p->windows = w->next;
 895                 free(w);
 896         }
 897 }
 898
 899 static int close_pack_fd(struct packed_git *p)
 900 {
 901         if (p->pack_fd < 0)
 902                 return 0;
 903
 904         close(p->pack_fd);
 905         pack_open_fds--;
 906         p->pack_fd = -1;
 907
 908         return 1;
 909 }
 910
 911 static void close_pack(struct packed_git *p)
 912 {
 913         close_pack_windows(p);
 914         close_pack_fd(p);
 915         close_pack_index(p);
 916 }
 917
 918 void close_all_packs(void)
 919 {
 920         struct packed_git *p;
 921
 922         for (p = packed_git; p; p = p->next)
 923                 if (p->do_not_close)
 924                         die("BUG: want to close pack marked 'do-not-close'");
 925                 else
 926                         close_pack(p);
 927 }
 928
 929
 930 /*
 931  * The LRU pack is the one with the oldest MRU window, preferring packs
 932  * with no used windows, or the oldest mtime if it has no windows allocated.
 933  */
 934 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 935 {
 936         struct pack_window *w, *this_mru_w;
 937         int has_windows_inuse = 0;
 938
 939         /*
 940          * Reject this pack if it has windows and the previously selected
 941          * one does not.  If this pack does not have windows, reject
 942          * it if the pack file is newer than the previously selected one.
 943          */
 944         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 945                 return;
 946
 947         for (w = this_mru_w = p->windows; w; w = w->next) {
 948                 /*
 949                  * Reject this pack if any of its windows are in use,
 950                  * but the previously selected pack did not have any
 951                  * inuse windows.  Otherwise, record that this pack
 952                  * has windows in use.
 953                  */
 954                 if (w->inuse_cnt) {
 955                         if (*accept_windows_inuse)
 956                                 has_windows_inuse = 1;
 957                         else
 958                                 return;
 959                 }
 960
 961                 if (w->last_used > this_mru_w->last_used)
 962                         this_mru_w = w;
 963
 964                 /*
 965                  * Reject this pack if it has windows that have been
 966                  * used more recently than the previously selected pack.
 967                  * If the previously selected pack had windows inuse and
 968                  * we have not encountered a window in this pack that is
 969                  * inuse, skip this check since we prefer a pack with no
 970                  * inuse windows to one that has inuse windows.
 971                  */
 972                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 973                     this_mru_w->last_used > (*mru_w)->last_used)
 974                         return;
 975         }
 976
 977         /*
 978          * Select this pack.
 979          */
 980         *mru_w = this_mru_w;
 981         *lru_p = p;
 982         *accept_windows_inuse = has_windows_inuse;
 983 }
 984
 985 static int close_one_pack(void)
 986 {
 987         struct packed_git *p, *lru_p = NULL;
 988         struct pack_window *mru_w = NULL;
 989         int accept_windows_inuse = 1;
 990
 991         for (p = packed_git; p; p = p->next) {
 992                 if (p->pack_fd == -1)
 993                         continue;
 994                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 995         }
 996
 997         if (lru_p)
 998                 return close_pack_fd(lru_p);
 999
1000         return 0;
1001 }
1002
1003 void unuse_pack(struct pack_window **w_cursor)
1004 {
1005         struct pack_window *w = *w_cursor;
1006         if (w) {
1007                 w->inuse_cnt--;
1008                 *w_cursor = NULL;
1009         }
1010 }
1011
1012 void close_pack_index(struct packed_git *p)
1013 {
1014         if (p->index_data) {
1015                 munmap((void *)p->index_data, p->index_size);
1016                 p->index_data = NULL;
1017         }
1018 }
1019
1020 static unsigned int get_max_fd_limit(void)
1021 {
1022 #ifdef RLIMIT_NOFILE
1023         {
1024                 struct rlimit lim;
1025
1026                 if (!getrlimit(RLIMIT_NOFILE, &lim))
1027                         return lim.rlim_cur;
1028         }
1029 #endif
1030
1031 #ifdef _SC_OPEN_MAX
1032         {
1033                 long open_max = sysconf(_SC_OPEN_MAX);
1034                 if (0 < open_max)
1035                         return open_max;
1036                 /*
1037                  * Otherwise, we got -1 for one of the two
1038                  * reasons:
1039                  *
1040                  * (1) sysconf() did not understand _SC_OPEN_MAX
1041                  *     and signaled an error with -1; or
1042                  * (2) sysconf() said there is no limit.
1043                  *
1044                  * We _could_ clear errno before calling sysconf() to
1045                  * tell these two cases apart and return a huge number
1046                  * in the latter case to let the caller cap it to a
1047                  * value that is not so selfish, but letting the
1048                  * fallback OPEN_MAX codepath take care of these cases
1049                  * is a lot simpler.
1050                  */
1051         }
1052 #endif
1053
1054 #ifdef OPEN_MAX
1055         return OPEN_MAX;
1056 #else
1057         return 1; /* see the caller ;-) */
1058 #endif
1059 }
1060
1061 /*
1062  * Do not call this directly as this leaks p->pack_fd on error return;
1063  * call open_packed_git() instead.
1064  */
1065 static int open_packed_git_1(struct packed_git *p)
1066 {
1067         struct stat st;
1068         struct pack_header hdr;
1069         unsigned char sha1[20];
1070         unsigned char *idx_sha1;
1071         long fd_flag;
1072
1073         if (!p->index_data && open_pack_index(p))
1074                 return error("packfile %s index unavailable", p->pack_name);
1075
1076         if (!pack_max_fds) {
1077                 unsigned int max_fds = get_max_fd_limit();
1078
1079                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1080                 if (25 < max_fds)
1081                         pack_max_fds = max_fds - 25;
1082                 else
1083                         pack_max_fds = 1;
1084         }
1085
1086         while (pack_max_fds <= pack_open_fds && close_one_pack())
1087                 ; /* nothing */
1088
1089         p->pack_fd = git_open(p->pack_name);
1090         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1091                 return -1;
1092         pack_open_fds++;
1093
1094         /* If we created the struct before we had the pack we lack size. */
1095         if (!p->pack_size) {
1096                 if (!S_ISREG(st.st_mode))
1097                         return error("packfile %s not a regular file", p->pack_name);
1098                 p->pack_size = st.st_size;
1099         } else if (p->pack_size != st.st_size)
1100                 return error("packfile %s size changed", p->pack_name);
1101
1102         /* We leave these file descriptors open with sliding mmap;
1103          * there is no point keeping them open across exec(), though.
1104          */
1105         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1106         if (fd_flag < 0)
1107                 return error("cannot determine file descriptor flags");
1108         fd_flag |= FD_CLOEXEC;
1109         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1110                 return error("cannot set FD_CLOEXEC");
1111
1112         /* Verify we recognize this pack file format. */
1113         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1114                 return error("file %s is far too short to be a packfile", p->pack_name);
1115         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1116                 return error("file %s is not a GIT packfile", p->pack_name);
1117         if (!pack_version_ok(hdr.hdr_version))
1118                 return error("packfile %s is version %"PRIu32" and not"
1119                         " supported (try upgrading GIT to a newer version)",
1120                         p->pack_name, ntohl(hdr.hdr_version));
1121
1122         /* Verify the pack matches its index. */
1123         if (p->num_objects != ntohl(hdr.hdr_entries))
1124                 return error("packfile %s claims to have %"PRIu32" objects"
1125                              " while index indicates %"PRIu32" objects",
1126                              p->pack_name, ntohl(hdr.hdr_entries),
1127                              p->num_objects);
1128         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1129                 return error("end of packfile %s is unavailable", p->pack_name);
1130         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1131                 return error("packfile %s signature is unavailable", p->pack_name);
1132         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1133         if (hashcmp(sha1, idx_sha1))
1134                 return error("packfile %s does not match index", p->pack_name);
1135         return 0;
1136 }
1137
1138 static int open_packed_git(struct packed_git *p)
1139 {
1140         if (!open_packed_git_1(p))
1141                 return 0;
1142         close_pack_fd(p);
1143         return -1;
1144 }
1145
1146 static int in_window(struct pack_window *win, off_t offset)
1147 {
1148         /* We must promise at least 20 bytes (one hash) after the
1149          * offset is available from this window, otherwise the offset
1150          * is not actually in this window and a different window (which
1151          * has that one hash excess) must be used.  This is to support
1152          * the object header and delta base parsing routines below.
1153          */
1154         off_t win_off = win->offset;
1155         return win_off <= offset
1156                 && (offset + 20) <= (win_off + win->len);
1157 }
1158
1159 unsigned char *use_pack(struct packed_git *p,
1160                 struct pack_window **w_cursor,
1161                 off_t offset,
1162                 unsigned long *left)
1163 {
1164         struct pack_window *win = *w_cursor;
1165
1166         /* Since packfiles end in a hash of their content and it's
1167          * pointless to ask for an offset into the middle of that
1168          * hash, and the in_window function above wouldn't match
1169          * don't allow an offset too close to the end of the file.
1170          */
1171         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1172                 die("packfile %s cannot be accessed", p->pack_name);
1173         if (offset > (p->pack_size - 20))
1174                 die("offset beyond end of packfile (truncated pack?)");
1175         if (offset < 0)
1176                 die(_("offset before end of packfile (broken .idx?)"));
1177
1178         if (!win || !in_window(win, offset)) {
1179                 if (win)
1180                         win->inuse_cnt--;
1181                 for (win = p->windows; win; win = win->next) {
1182                         if (in_window(win, offset))
1183                                 break;
1184                 }
1185                 if (!win) {
1186                         size_t window_align = packed_git_window_size / 2;
1187                         off_t len;
1188
1189                         if (p->pack_fd == -1 && open_packed_git(p))
1190                                 die("packfile %s cannot be accessed", p->pack_name);
1191
1192                         win = xcalloc(1, sizeof(*win));
1193                         win->offset = (offset / window_align) * window_align;
1194                         len = p->pack_size - win->offset;
1195                         if (len > packed_git_window_size)
1196                                 len = packed_git_window_size;
1197                         win->len = (size_t)len;
1198                         pack_mapped += win->len;
1199                         while (packed_git_limit < pack_mapped
1200                                 && unuse_one_window(p))
1201                                 ; /* nothing */
1202                         win->base = xmmap(NULL, win->len,
1203                                 PROT_READ, MAP_PRIVATE,
1204                                 p->pack_fd, win->offset);
1205                         if (win->base == MAP_FAILED)
1206                                 die_errno("packfile %s cannot be mapped",
1207                                           p->pack_name);
1208                         if (!win->offset && win->len == p->pack_size
1209                                 && !p->do_not_close)
1210                                 close_pack_fd(p);
1211                         pack_mmap_calls++;
1212                         pack_open_windows++;
1213                         if (pack_mapped > peak_pack_mapped)
1214                                 peak_pack_mapped = pack_mapped;
1215                         if (pack_open_windows > peak_pack_open_windows)
1216                                 peak_pack_open_windows = pack_open_windows;
1217                         win->next = p->windows;
1218                         p->windows = win;
1219                 }
1220         }
1221         if (win != *w_cursor) {
1222                 win->last_used = pack_used_ctr++;
1223                 win->inuse_cnt++;
1224                 *w_cursor = win;
1225         }
1226         offset -= win->offset;
1227         if (left)
1228                 *left = win->len - xsize_t(offset);
1229         return win->base + offset;
1230 }
1231
1232 static struct packed_git *alloc_packed_git(int extra)
1233 {
1234         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1235         memset(p, 0, sizeof(*p));
1236         p->pack_fd = -1;
1237         return p;
1238 }
1239
1240 static void try_to_free_pack_memory(size_t size)
1241 {
1242         release_pack_memory(size);
1243 }
1244
1245 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1246 {
1247         static int have_set_try_to_free_routine;
1248         struct stat st;
1249         size_t alloc;
1250         struct packed_git *p;
1251
1252         if (!have_set_try_to_free_routine) {
1253                 have_set_try_to_free_routine = 1;
1254                 set_try_to_free_routine(try_to_free_pack_memory);
1255         }
1256
1257         /*
1258          * Make sure a corresponding .pack file exists and that
1259          * the index looks sane.
1260          */
1261         if (!strip_suffix_mem(path, &path_len, ".idx"))
1262                 return NULL;
1263
1264         /*
1265          * ".pack" is long enough to hold any suffix we're adding (and
1266          * the use xsnprintf double-checks that)
1267          */
1268         alloc = st_add3(path_len, strlen(".pack"), 1);
1269         p = alloc_packed_git(alloc);
1270         memcpy(p->pack_name, path, path_len);
1271
1272         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1273         if (!access(p->pack_name, F_OK))
1274                 p->pack_keep = 1;
1275
1276         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1277         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1278                 free(p);
1279                 return NULL;
1280         }
1281
1282         /* ok, it looks sane as far as we can check without
1283          * actually mapping the pack file.
1284          */
1285         p->pack_size = st.st_size;
1286         p->pack_local = local;
1287         p->mtime = st.st_mtime;
1288         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1289                 hashclr(p->sha1);
1290         return p;
1291 }
1292
1293 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1294 {
1295         const char *path = sha1_pack_name(sha1);
1296         size_t alloc = st_add(strlen(path), 1);
1297         struct packed_git *p = alloc_packed_git(alloc);
1298
1299         memcpy(p->pack_name, path, alloc); /* includes NUL */
1300         hashcpy(p->sha1, sha1);
1301         if (check_packed_git_idx(idx_path, p)) {
1302                 free(p);
1303                 return NULL;
1304         }
1305
1306         return p;
1307 }
1308
1309 void install_packed_git(struct packed_git *pack)
1310 {
1311         if (pack->pack_fd != -1)
1312                 pack_open_fds++;
1313
1314         pack->next = packed_git;
1315         packed_git = pack;
1316 }
1317
1318 void (*report_garbage)(unsigned seen_bits, const char *path);
1319
1320 static void report_helper(const struct string_list *list,
1321                           int seen_bits, int first, int last)
1322 {
1323         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1324                 return;
1325
1326         for (; first < last; first++)
1327                 report_garbage(seen_bits, list->items[first].string);
1328 }
1329
1330 static void report_pack_garbage(struct string_list *list)
1331 {
1332         int i, baselen = -1, first = 0, seen_bits = 0;
1333
1334         if (!report_garbage)
1335                 return;
1336
1337         string_list_sort(list);
1338
1339         for (i = 0; i < list->nr; i++) {
1340                 const char *path = list->items[i].string;
1341                 if (baselen != -1 &&
1342                     strncmp(path, list->items[first].string, baselen)) {
1343                         report_helper(list, seen_bits, first, i);
1344                         baselen = -1;
1345                         seen_bits = 0;
1346                 }
1347                 if (baselen == -1) {
1348                         const char *dot = strrchr(path, '.');
1349                         if (!dot) {
1350                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1351                                 continue;
1352                         }
1353                         baselen = dot - path + 1;
1354                         first = i;
1355                 }
1356                 if (!strcmp(path + baselen, "pack"))
1357                         seen_bits |= 1;
1358                 else if (!strcmp(path + baselen, "idx"))
1359                         seen_bits |= 2;
1360         }
1361         report_helper(list, seen_bits, first, list->nr);
1362 }
1363
1364 static void prepare_packed_git_one(char *objdir, int local)
1365 {
1366         struct strbuf path = STRBUF_INIT;
1367         size_t dirnamelen;
1368         DIR *dir;
1369         struct dirent *de;
1370         struct string_list garbage = STRING_LIST_INIT_DUP;
1371
1372         strbuf_addstr(&path, objdir);
1373         strbuf_addstr(&path, "/pack");
1374         dir = opendir(path.buf);
1375         if (!dir) {
1376                 if (errno != ENOENT)
1377                         error_errno("unable to open object pack directory: %s",
1378                                     path.buf);
1379                 strbuf_release(&path);
1380                 return;
1381         }
1382         strbuf_addch(&path, '/');
1383         dirnamelen = path.len;
1384         while ((de = readdir(dir)) != NULL) {
1385                 struct packed_git *p;
1386                 size_t base_len;
1387
1388                 if (is_dot_or_dotdot(de->d_name))
1389                         continue;
1390
1391                 strbuf_setlen(&path, dirnamelen);
1392                 strbuf_addstr(&path, de->d_name);
1393
1394                 base_len = path.len;
1395                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1396                         /* Don't reopen a pack we already have. */
1397                         for (p = packed_git; p; p = p->next) {
1398                                 size_t len;
1399                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1400                                     len == base_len &&
1401                                     !memcmp(p->pack_name, path.buf, len))
1402                                         break;
1403                         }
1404                         if (p == NULL &&
1405                             /*
1406                              * See if it really is a valid .idx file with
1407                              * corresponding .pack file that we can map.
1408                              */
1409                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1410                                 install_packed_git(p);
1411                 }
1412
1413                 if (!report_garbage)
1414                         continue;
1415
1416                 if (ends_with(de->d_name, ".idx") ||
1417                     ends_with(de->d_name, ".pack") ||
1418                     ends_with(de->d_name, ".bitmap") ||
1419                     ends_with(de->d_name, ".keep"))
1420                         string_list_append(&garbage, path.buf);
1421                 else
1422                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1423         }
1424         closedir(dir);
1425         report_pack_garbage(&garbage);
1426         string_list_clear(&garbage, 0);
1427         strbuf_release(&path);
1428 }
1429
1430 static int approximate_object_count_valid;
1431
1432 /*
1433  * Give a fast, rough count of the number of objects in the repository. This
1434  * ignores loose objects completely. If you have a lot of them, then either
1435  * you should repack because your performance will be awful, or they are
1436  * all unreachable objects about to be pruned, in which case they're not really
1437  * interesting as a measure of repo size in the first place.
1438  */
1439 unsigned long approximate_object_count(void)
1440 {
1441         static unsigned long count;
1442         if (!approximate_object_count_valid) {
1443                 struct packed_git *p;
1444
1445                 prepare_packed_git();
1446                 count = 0;
1447                 for (p = packed_git; p; p = p->next) {
1448                         if (open_pack_index(p))
1449                                 continue;
1450                         count += p->num_objects;
1451                 }
1452         }
1453         return count;
1454 }
1455
1456 static void *get_next_packed_git(const void *p)
1457 {
1458         return ((const struct packed_git *)p)->next;
1459 }
1460
1461 static void set_next_packed_git(void *p, void *next)
1462 {
1463         ((struct packed_git *)p)->next = next;
1464 }
1465
1466 static int sort_pack(const void *a_, const void *b_)
1467 {
1468         const struct packed_git *a = a_;
1469         const struct packed_git *b = b_;
1470         int st;
1471
1472         /*
1473          * Local packs tend to contain objects specific to our
1474          * variant of the project than remote ones.  In addition,
1475          * remote ones could be on a network mounted filesystem.
1476          * Favor local ones for these reasons.
1477          */
1478         st = a->pack_local - b->pack_local;
1479         if (st)
1480                 return -st;
1481
1482         /*
1483          * Younger packs tend to contain more recent objects,
1484          * and more recent objects tend to get accessed more
1485          * often.
1486          */
1487         if (a->mtime < b->mtime)
1488                 return 1;
1489         else if (a->mtime == b->mtime)
1490                 return 0;
1491         return -1;
1492 }
1493
1494 static void rearrange_packed_git(void)
1495 {
1496         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1497                                      set_next_packed_git, sort_pack);
1498 }
1499
1500 static void prepare_packed_git_mru(void)
1501 {
1502         struct packed_git *p;
1503
1504         mru_clear(packed_git_mru);
1505         for (p = packed_git; p; p = p->next)
1506                 mru_append(packed_git_mru, p);
1507 }
1508
1509 static int prepare_packed_git_run_once = 0;
1510 void prepare_packed_git(void)
1511 {
1512         struct alternate_object_database *alt;
1513
1514         if (prepare_packed_git_run_once)
1515                 return;
1516         prepare_packed_git_one(get_object_directory(), 1);
1517         prepare_alt_odb();
1518         for (alt = alt_odb_list; alt; alt = alt->next)
1519                 prepare_packed_git_one(alt->path, 0);
1520         rearrange_packed_git();
1521         prepare_packed_git_mru();
1522         prepare_packed_git_run_once = 1;
1523 }
1524
1525 void reprepare_packed_git(void)
1526 {
1527         approximate_object_count_valid = 0;
1528         prepare_packed_git_run_once = 0;
1529         prepare_packed_git();
1530 }
1531
1532 static void mark_bad_packed_object(struct packed_git *p,
1533                                    const unsigned char *sha1)
1534 {
1535         unsigned i;
1536         for (i = 0; i < p->num_bad_objects; i++)
1537                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1538                         return;
1539         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1540                                       st_mult(GIT_SHA1_RAWSZ,
1541                                               st_add(p->num_bad_objects, 1)));
1542         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1543         p->num_bad_objects++;
1544 }
1545
1546 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1547 {
1548         struct packed_git *p;
1549         unsigned i;
1550
1551         for (p = packed_git; p; p = p->next)
1552                 for (i = 0; i < p->num_bad_objects; i++)
1553                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1554                                 return p;
1555         return NULL;
1556 }
1557
1558 /*
1559  * With an in-core object data in "map", rehash it to make sure the
1560  * object name actually matches "sha1" to detect object corruption.
1561  * With "map" == NULL, try reading the object named with "sha1" using
1562  * the streaming interface and rehash it to do the same.
1563  */
1564 int check_sha1_signature(const unsigned char *sha1, void *map,
1565                          unsigned long size, const char *type)
1566 {
1567         unsigned char real_sha1[20];
1568         enum object_type obj_type;
1569         struct git_istream *st;
1570         git_SHA_CTX c;
1571         char hdr[32];
1572         int hdrlen;
1573
1574         if (map) {
1575                 hash_sha1_file(map, size, type, real_sha1);
1576                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1577         }
1578
1579         st = open_istream(sha1, &obj_type, &size, NULL);
1580         if (!st)
1581                 return -1;
1582
1583         /* Generate the header */
1584         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1585
1586         /* Sha1.. */
1587         git_SHA1_Init(&c);
1588         git_SHA1_Update(&c, hdr, hdrlen);
1589         for (;;) {
1590                 char buf[1024 * 16];
1591                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1592
1593                 if (readlen < 0) {
1594                         close_istream(st);
1595                         return -1;
1596                 }
1597                 if (!readlen)
1598                         break;
1599                 git_SHA1_Update(&c, buf, readlen);
1600         }
1601         git_SHA1_Final(real_sha1, &c);
1602         close_istream(st);
1603         return hashcmp(sha1, real_sha1) ? -1 : 0;
1604 }
1605
1606 int git_open_cloexec(const char *name, int flags)
1607 {
1608         int fd;
1609         static int o_cloexec = O_CLOEXEC;
1610
1611         fd = open(name, flags | o_cloexec);
1612         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
1613                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
1614                 o_cloexec &= ~O_CLOEXEC;
1615                 fd = open(name, flags | o_cloexec);
1616         }
1617
1618 #if defined(F_GETFL) && defined(F_SETFL) && defined(FD_CLOEXEC)
1619         {
1620                 static int fd_cloexec = FD_CLOEXEC;
1621
1622                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
1623                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
1624                         int flags = fcntl(fd, F_GETFL);
1625                         if (fcntl(fd, F_SETFL, flags | fd_cloexec))
1626                                 fd_cloexec = 0;
1627                 }
1628         }
1629 #endif
1630         return fd;
1631 }
1632
1633 /*
1634  * Find "sha1" as a loose object in the local repository or in an alternate.
1635  * Returns 0 on success, negative on failure.
1636  *
1637  * The "path" out-parameter will give the path of the object we found (if any).
1638  * Note that it may point to static storage and is only valid until another
1639  * call to sha1_file_name(), etc.
1640  */
1641 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
1642                           const char **path)
1643 {
1644         struct alternate_object_database *alt;
1645
1646         *path = sha1_file_name(sha1);
1647         if (!lstat(*path, st))
1648                 return 0;
1649
1650         prepare_alt_odb();
1651         errno = ENOENT;
1652         for (alt = alt_odb_list; alt; alt = alt->next) {
1653                 *path = alt_sha1_path(alt, sha1);
1654                 if (!lstat(*path, st))
1655                         return 0;
1656         }
1657
1658         return -1;
1659 }
1660
1661 /*
1662  * Like stat_sha1_file(), but actually open the object and return the
1663  * descriptor. See the caveats on the "path" parameter above.
1664  */
1665 static int open_sha1_file(const unsigned char *sha1, const char **path)
1666 {
1667         int fd;
1668         struct alternate_object_database *alt;
1669         int most_interesting_errno;
1670
1671         *path = sha1_file_name(sha1);
1672         fd = git_open(*path);
1673         if (fd >= 0)
1674                 return fd;
1675         most_interesting_errno = errno;
1676
1677         prepare_alt_odb();
1678         for (alt = alt_odb_list; alt; alt = alt->next) {
1679                 *path = alt_sha1_path(alt, sha1);
1680                 fd = git_open(*path);
1681                 if (fd >= 0)
1682                         return fd;
1683                 if (most_interesting_errno == ENOENT)
1684                         most_interesting_errno = errno;
1685         }
1686         errno = most_interesting_errno;
1687         return -1;
1688 }
1689
1690 /*
1691  * Map the loose object at "path" if it is not NULL, or the path found by
1692  * searching for a loose object named "sha1".
1693  */
1694 static void *map_sha1_file_1(const char *path,
1695                              const unsigned char *sha1,
1696                              unsigned long *size)
1697 {
1698         void *map;
1699         int fd;
1700
1701         if (path)
1702                 fd = git_open(path);
1703         else
1704                 fd = open_sha1_file(sha1, &path);
1705         map = NULL;
1706         if (fd >= 0) {
1707                 struct stat st;
1708
1709                 if (!fstat(fd, &st)) {
1710                         *size = xsize_t(st.st_size);
1711                         if (!*size) {
1712                                 /* mmap() is forbidden on empty files */
1713                                 error("object file %s is empty", path);
1714                                 return NULL;
1715                         }
1716                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1717                 }
1718                 close(fd);
1719         }
1720         return map;
1721 }
1722
1723 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1724 {
1725         return map_sha1_file_1(NULL, sha1, size);
1726 }
1727
1728 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1729                 unsigned long len, enum object_type *type, unsigned long *sizep)
1730 {
1731         unsigned shift;
1732         unsigned long size, c;
1733         unsigned long used = 0;
1734
1735         c = buf[used++];
1736         *type = (c >> 4) & 7;
1737         size = c & 15;
1738         shift = 4;
1739         while (c & 0x80) {
1740                 if (len <= used || bitsizeof(long) <= shift) {
1741                         error("bad object header");
1742                         size = used = 0;
1743                         break;
1744                 }
1745                 c = buf[used++];
1746                 size += (c & 0x7f) << shift;
1747                 shift += 7;
1748         }
1749         *sizep = size;
1750         return used;
1751 }
1752
1753 static int unpack_sha1_short_header(git_zstream *stream,
1754                                     unsigned char *map, unsigned long mapsize,
1755                                     void *buffer, unsigned long bufsiz)
1756 {
1757         /* Get the data stream */
1758         memset(stream, 0, sizeof(*stream));
1759         stream->next_in = map;
1760         stream->avail_in = mapsize;
1761         stream->next_out = buffer;
1762         stream->avail_out = bufsiz;
1763
1764         git_inflate_init(stream);
1765         return git_inflate(stream, 0);
1766 }
1767
1768 int unpack_sha1_header(git_zstream *stream,
1769                        unsigned char *map, unsigned long mapsize,
1770                        void *buffer, unsigned long bufsiz)
1771 {
1772         int status = unpack_sha1_short_header(stream, map, mapsize,
1773                                               buffer, bufsiz);
1774
1775         if (status < Z_OK)
1776                 return status;
1777
1778         /* Make sure we have the terminating NUL */
1779         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1780                 return -1;
1781         return 0;
1782 }
1783
1784 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1785                                         unsigned long mapsize, void *buffer,
1786                                         unsigned long bufsiz, struct strbuf *header)
1787 {
1788         int status;
1789
1790         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
1791         if (status < Z_OK)
1792                 return -1;
1793
1794         /*
1795          * Check if entire header is unpacked in the first iteration.
1796          */
1797         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1798                 return 0;
1799
1800         /*
1801          * buffer[0..bufsiz] was not large enough.  Copy the partial
1802          * result out to header, and then append the result of further
1803          * reading the stream.
1804          */
1805         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1806         stream->next_out = buffer;
1807         stream->avail_out = bufsiz;
1808
1809         do {
1810                 status = git_inflate(stream, 0);
1811                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1812                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1813                         return 0;
1814                 stream->next_out = buffer;
1815                 stream->avail_out = bufsiz;
1816         } while (status != Z_STREAM_END);
1817         return -1;
1818 }
1819
1820 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1821 {
1822         int bytes = strlen(buffer) + 1;
1823         unsigned char *buf = xmallocz(size);
1824         unsigned long n;
1825         int status = Z_OK;
1826
1827         n = stream->total_out - bytes;
1828         if (n > size)
1829                 n = size;
1830         memcpy(buf, (char *) buffer + bytes, n);
1831         bytes = n;
1832         if (bytes <= size) {
1833                 /*
1834                  * The above condition must be (bytes <= size), not
1835                  * (bytes < size).  In other words, even though we
1836                  * expect no more output and set avail_out to zero,
1837                  * the input zlib stream may have bytes that express
1838                  * "this concludes the stream", and we *do* want to
1839                  * eat that input.
1840                  *
1841                  * Otherwise we would not be able to test that we
1842                  * consumed all the input to reach the expected size;
1843                  * we also want to check that zlib tells us that all
1844                  * went well with status == Z_STREAM_END at the end.
1845                  */
1846                 stream->next_out = buf + bytes;
1847                 stream->avail_out = size - bytes;
1848                 while (status == Z_OK)
1849                         status = git_inflate(stream, Z_FINISH);
1850         }
1851         if (status == Z_STREAM_END && !stream->avail_in) {
1852                 git_inflate_end(stream);
1853                 return buf;
1854         }
1855
1856         if (status < 0)
1857                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1858         else if (stream->avail_in)
1859                 error("garbage at end of loose object '%s'",
1860                       sha1_to_hex(sha1));
1861         free(buf);
1862         return NULL;
1863 }
1864
1865 /*
1866  * We used to just use "sscanf()", but that's actually way
1867  * too permissive for what we want to check. So do an anal
1868  * object header parse by hand.
1869  */
1870 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1871                                unsigned int flags)
1872 {
1873         const char *type_buf = hdr;
1874         unsigned long size;
1875         int type, type_len = 0;
1876
1877         /*
1878          * The type can be of any size but is followed by
1879          * a space.
1880          */
1881         for (;;) {
1882                 char c = *hdr++;
1883                 if (!c)
1884                         return -1;
1885                 if (c == ' ')
1886                         break;
1887                 type_len++;
1888         }
1889
1890         type = type_from_string_gently(type_buf, type_len, 1);
1891         if (oi->typename)
1892                 strbuf_add(oi->typename, type_buf, type_len);
1893         /*
1894          * Set type to 0 if its an unknown object and
1895          * we're obtaining the type using '--allow-unknown-type'
1896          * option.
1897          */
1898         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1899                 type = 0;
1900         else if (type < 0)
1901                 die("invalid object type");
1902         if (oi->typep)
1903                 *oi->typep = type;
1904
1905         /*
1906          * The length must follow immediately, and be in canonical
1907          * decimal format (ie "010" is not valid).
1908          */
1909         size = *hdr++ - '0';
1910         if (size > 9)
1911                 return -1;
1912         if (size) {
1913                 for (;;) {
1914                         unsigned long c = *hdr - '0';
1915                         if (c > 9)
1916                                 break;
1917                         hdr++;
1918                         size = size * 10 + c;
1919                 }
1920         }
1921
1922         if (oi->sizep)
1923                 *oi->sizep = size;
1924
1925         /*
1926          * The length must be followed by a zero byte
1927          */
1928         return *hdr ? -1 : type;
1929 }
1930
1931 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1932 {
1933         struct object_info oi = OBJECT_INFO_INIT;
1934
1935         oi.sizep = sizep;
1936         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
1937 }
1938
1939 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1940 {
1941         int ret;
1942         git_zstream stream;
1943         char hdr[8192];
1944
1945         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1946         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1947                 return NULL;
1948
1949         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1950 }
1951
1952 unsigned long get_size_from_delta(struct packed_git *p,
1953                                   struct pack_window **w_curs,
1954                                   off_t curpos)
1955 {
1956         const unsigned char *data;
1957         unsigned char delta_head[20], *in;
1958         git_zstream stream;
1959         int st;
1960
1961         memset(&stream, 0, sizeof(stream));
1962         stream.next_out = delta_head;
1963         stream.avail_out = sizeof(delta_head);
1964
1965         git_inflate_init(&stream);
1966         do {
1967                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1968                 stream.next_in = in;
1969                 st = git_inflate(&stream, Z_FINISH);
1970                 curpos += stream.next_in - in;
1971         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1972                  stream.total_out < sizeof(delta_head));
1973         git_inflate_end(&stream);
1974         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1975                 error("delta data unpack-initial failed");
1976                 return 0;
1977         }
1978
1979         /* Examine the initial part of the delta to figure out
1980          * the result size.
1981          */
1982         data = delta_head;
1983
1984         /* ignore base size */
1985         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1986
1987         /* Read the result size */
1988         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1989 }
1990
1991 static off_t get_delta_base(struct packed_git *p,
1992                                     struct pack_window **w_curs,
1993                                     off_t *curpos,
1994                                     enum object_type type,
1995                                     off_t delta_obj_offset)
1996 {
1997         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1998         off_t base_offset;
1999
2000         /* use_pack() assured us we have [base_info, base_info + 20)
2001          * as a range that we can look at without walking off the
2002          * end of the mapped window.  Its actually the hash size
2003          * that is assured.  An OFS_DELTA longer than the hash size
2004          * is stupid, as then a REF_DELTA would be smaller to store.
2005          */
2006         if (type == OBJ_OFS_DELTA) {
2007                 unsigned used = 0;
2008                 unsigned char c = base_info[used++];
2009                 base_offset = c & 127;
2010                 while (c & 128) {
2011                         base_offset += 1;
2012                         if (!base_offset || MSB(base_offset, 7))
2013                                 return 0;  /* overflow */
2014                         c = base_info[used++];
2015                         base_offset = (base_offset << 7) + (c & 127);
2016                 }
2017                 base_offset = delta_obj_offset - base_offset;
2018                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
2019                         return 0;  /* out of bound */
2020                 *curpos += used;
2021         } else if (type == OBJ_REF_DELTA) {
2022                 /* The base entry _must_ be in the same pack */
2023                 base_offset = find_pack_entry_one(base_info, p);
2024                 *curpos += 20;
2025         } else
2026                 die("I am totally screwed");
2027         return base_offset;
2028 }
2029
2030 /*
2031  * Like get_delta_base above, but we return the sha1 instead of the pack
2032  * offset. This means it is cheaper for REF deltas (we do not have to do
2033  * the final object lookup), but more expensive for OFS deltas (we
2034  * have to load the revidx to convert the offset back into a sha1).
2035  */
2036 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
2037                                                 struct pack_window **w_curs,
2038                                                 off_t curpos,
2039                                                 enum object_type type,
2040                                                 off_t delta_obj_offset)
2041 {
2042         if (type == OBJ_REF_DELTA) {
2043                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
2044                 return base;
2045         } else if (type == OBJ_OFS_DELTA) {
2046                 struct revindex_entry *revidx;
2047                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
2048                                                    type, delta_obj_offset);
2049
2050                 if (!base_offset)
2051                         return NULL;
2052
2053                 revidx = find_pack_revindex(p, base_offset);
2054                 if (!revidx)
2055                         return NULL;
2056
2057                 return nth_packed_object_sha1(p, revidx->nr);
2058         } else
2059                 return NULL;
2060 }
2061
2062 int unpack_object_header(struct packed_git *p,
2063                          struct pack_window **w_curs,
2064                          off_t *curpos,
2065                          unsigned long *sizep)
2066 {
2067         unsigned char *base;
2068         unsigned long left;
2069         unsigned long used;
2070         enum object_type type;
2071
2072         /* use_pack() assures us we have [base, base + 20) available
2073          * as a range that we can look at.  (Its actually the hash
2074          * size that is assured.)  With our object header encoding
2075          * the maximum deflated object size is 2^137, which is just
2076          * insane, so we know won't exceed what we have been given.
2077          */
2078         base = use_pack(p, w_curs, *curpos, &left);
2079         used = unpack_object_header_buffer(base, left, &type, sizep);
2080         if (!used) {
2081                 type = OBJ_BAD;
2082         } else
2083                 *curpos += used;
2084
2085         return type;
2086 }
2087
2088 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
2089 {
2090         int type;
2091         struct revindex_entry *revidx;
2092         const unsigned char *sha1;
2093         revidx = find_pack_revindex(p, obj_offset);
2094         if (!revidx)
2095                 return OBJ_BAD;
2096         sha1 = nth_packed_object_sha1(p, revidx->nr);
2097         mark_bad_packed_object(p, sha1);
2098         type = sha1_object_info(sha1, NULL);
2099         if (type <= OBJ_NONE)
2100                 return OBJ_BAD;
2101         return type;
2102 }
2103
2104 #define POI_STACK_PREALLOC 64
2105
2106 static enum object_type packed_to_object_type(struct packed_git *p,
2107                                               off_t obj_offset,
2108                                               enum object_type type,
2109                                               struct pack_window **w_curs,
2110                                               off_t curpos)
2111 {
2112         off_t small_poi_stack[POI_STACK_PREALLOC];
2113         off_t *poi_stack = small_poi_stack;
2114         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
2115
2116         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2117                 off_t base_offset;
2118                 unsigned long size;
2119                 /* Push the object we're going to leave behind */
2120                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
2121                         poi_stack_alloc = alloc_nr(poi_stack_nr);
2122                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
2123                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2124                 } else {
2125                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2126                 }
2127                 poi_stack[poi_stack_nr++] = obj_offset;
2128                 /* If parsing the base offset fails, just unwind */
2129                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2130                 if (!base_offset)
2131                         goto unwind;
2132                 curpos = obj_offset = base_offset;
2133                 type = unpack_object_header(p, w_curs, &curpos, &size);
2134                 if (type <= OBJ_NONE) {
2135                         /* If getting the base itself fails, we first
2136                          * retry the base, otherwise unwind */
2137                         type = retry_bad_packed_offset(p, base_offset);
2138                         if (type > OBJ_NONE)
2139                                 goto out;
2140                         goto unwind;
2141                 }
2142         }
2143
2144         switch (type) {
2145         case OBJ_BAD:
2146         case OBJ_COMMIT:
2147         case OBJ_TREE:
2148         case OBJ_BLOB:
2149         case OBJ_TAG:
2150                 break;
2151         default:
2152                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2153                       type, (uintmax_t)obj_offset, p->pack_name);
2154                 type = OBJ_BAD;
2155         }
2156
2157 out:
2158         if (poi_stack != small_poi_stack)
2159                 free(poi_stack);
2160         return type;
2161
2162 unwind:
2163         while (poi_stack_nr) {
2164                 obj_offset = poi_stack[--poi_stack_nr];
2165                 type = retry_bad_packed_offset(p, obj_offset);
2166                 if (type > OBJ_NONE)
2167                         goto out;
2168         }
2169         type = OBJ_BAD;
2170         goto out;
2171 }
2172
2173 int packed_object_info(struct packed_git *p, off_t obj_offset,
2174                        struct object_info *oi)
2175 {
2176         struct pack_window *w_curs = NULL;
2177         unsigned long size;
2178         off_t curpos = obj_offset;
2179         enum object_type type;
2180
2181         /*
2182          * We always get the representation type, but only convert it to
2183          * a "real" type later if the caller is interested.
2184          */
2185         type = unpack_object_header(p, &w_curs, &curpos, &size);
2186
2187         if (oi->sizep) {
2188                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2189                         off_t tmp_pos = curpos;
2190                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2191                                                            type, obj_offset);
2192                         if (!base_offset) {
2193                                 type = OBJ_BAD;
2194                                 goto out;
2195                         }
2196                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2197                         if (*oi->sizep == 0) {
2198                                 type = OBJ_BAD;
2199                                 goto out;
2200                         }
2201                 } else {
2202                         *oi->sizep = size;
2203                 }
2204         }
2205
2206         if (oi->disk_sizep) {
2207                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2208                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2209         }
2210
2211         if (oi->typep) {
2212                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2213                 if (*oi->typep < 0) {
2214                         type = OBJ_BAD;
2215                         goto out;
2216                 }
2217         }
2218
2219         if (oi->delta_base_sha1) {
2220                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2221                         const unsigned char *base;
2222
2223                         base = get_delta_base_sha1(p, &w_curs, curpos,
2224                                                    type, obj_offset);
2225                         if (!base) {
2226                                 type = OBJ_BAD;
2227                                 goto out;
2228                         }
2229
2230                         hashcpy(oi->delta_base_sha1, base);
2231                 } else
2232                         hashclr(oi->delta_base_sha1);
2233         }
2234
2235 out:
2236         unuse_pack(&w_curs);
2237         return type;
2238 }
2239
2240 static void *unpack_compressed_entry(struct packed_git *p,
2241                                     struct pack_window **w_curs,
2242                                     off_t curpos,
2243                                     unsigned long size)
2244 {
2245         int st;
2246         git_zstream stream;
2247         unsigned char *buffer, *in;
2248
2249         buffer = xmallocz_gently(size);
2250         if (!buffer)
2251                 return NULL;
2252         memset(&stream, 0, sizeof(stream));
2253         stream.next_out = buffer;
2254         stream.avail_out = size + 1;
2255
2256         git_inflate_init(&stream);
2257         do {
2258                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2259                 stream.next_in = in;
2260                 st = git_inflate(&stream, Z_FINISH);
2261                 if (!stream.avail_out)
2262                         break; /* the payload is larger than it should be */
2263                 curpos += stream.next_in - in;
2264         } while (st == Z_OK || st == Z_BUF_ERROR);
2265         git_inflate_end(&stream);
2266         if ((st != Z_STREAM_END) || stream.total_out != size) {
2267                 free(buffer);
2268                 return NULL;
2269         }
2270
2271         return buffer;
2272 }
2273
2274 static struct hashmap delta_base_cache;
2275 static size_t delta_base_cached;
2276
2277 static LIST_HEAD(delta_base_cache_lru);
2278
2279 struct delta_base_cache_key {
2280         struct packed_git *p;
2281         off_t base_offset;
2282 };
2283
2284 struct delta_base_cache_entry {
2285         struct hashmap hash;
2286         struct delta_base_cache_key key;
2287         struct list_head lru;
2288         void *data;
2289         unsigned long size;
2290         enum object_type type;
2291 };
2292
2293 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2294 {
2295         unsigned int hash;
2296
2297         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2298         hash += (hash >> 8) + (hash >> 16);
2299         return hash;
2300 }
2301
2302 static struct delta_base_cache_entry *
2303 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2304 {
2305         struct hashmap_entry entry;
2306         struct delta_base_cache_key key;
2307
2308         if (!delta_base_cache.cmpfn)
2309                 return NULL;
2310
2311         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2312         key.p = p;
2313         key.base_offset = base_offset;
2314         return hashmap_get(&delta_base_cache, &entry, &key);
2315 }
2316
2317 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2318                                    const struct delta_base_cache_key *b)
2319 {
2320         return a->p == b->p && a->base_offset == b->base_offset;
2321 }
2322
2323 static int delta_base_cache_hash_cmp(const void *va, const void *vb,
2324                                      const void *vkey)
2325 {
2326         const struct delta_base_cache_entry *a = va, *b = vb;
2327         const struct delta_base_cache_key *key = vkey;
2328         if (key)
2329                 return !delta_base_cache_key_eq(&a->key, key);
2330         else
2331                 return !delta_base_cache_key_eq(&a->key, &b->key);
2332 }
2333
2334 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2335 {
2336         return !!get_delta_base_cache_entry(p, base_offset);
2337 }
2338
2339 /*
2340  * Remove the entry from the cache, but do _not_ free the associated
2341  * entry data. The caller takes ownership of the "data" buffer, and
2342  * should copy out any fields it wants before detaching.
2343  */
2344 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2345 {
2346         hashmap_remove(&delta_base_cache, ent, &ent->key);
2347         list_del(&ent->lru);
2348         delta_base_cached -= ent->size;
2349         free(ent);
2350 }
2351
2352 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2353         unsigned long *base_size, enum object_type *type)
2354 {
2355         struct delta_base_cache_entry *ent;
2356
2357         ent = get_delta_base_cache_entry(p, base_offset);
2358         if (!ent)
2359                 return unpack_entry(p, base_offset, type, base_size);
2360
2361         *type = ent->type;
2362         *base_size = ent->size;
2363         return xmemdupz(ent->data, ent->size);
2364 }
2365
2366 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2367 {
2368         free(ent->data);
2369         detach_delta_base_cache_entry(ent);
2370 }
2371
2372 void clear_delta_base_cache(void)
2373 {
2374         struct list_head *lru, *tmp;
2375         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2376                 struct delta_base_cache_entry *entry =
2377                         list_entry(lru, struct delta_base_cache_entry, lru);
2378                 release_delta_base_cache(entry);
2379         }
2380 }
2381
2382 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2383         void *base, unsigned long base_size, enum object_type type)
2384 {
2385         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2386         struct list_head *lru, *tmp;
2387
2388         delta_base_cached += base_size;
2389
2390         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2391                 struct delta_base_cache_entry *f =
2392                         list_entry(lru, struct delta_base_cache_entry, lru);
2393                 if (delta_base_cached <= delta_base_cache_limit)
2394                         break;
2395                 release_delta_base_cache(f);
2396         }
2397
2398         ent->key.p = p;
2399         ent->key.base_offset = base_offset;
2400         ent->type = type;
2401         ent->data = base;
2402         ent->size = base_size;
2403         list_add_tail(&ent->lru, &delta_base_cache_lru);
2404
2405         if (!delta_base_cache.cmpfn)
2406                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, 0);
2407         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2408         hashmap_add(&delta_base_cache, ent);
2409 }
2410
2411 static void *read_object(const unsigned char *sha1, enum object_type *type,
2412                          unsigned long *size);
2413
2414 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2415 {
2416         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2417         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2418                          p->pack_name, (uintmax_t)obj_offset);
2419 }
2420
2421 int do_check_packed_object_crc;
2422
2423 #define UNPACK_ENTRY_STACK_PREALLOC 64
2424 struct unpack_entry_stack_ent {
2425         off_t obj_offset;
2426         off_t curpos;
2427         unsigned long size;
2428 };
2429
2430 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2431                    enum object_type *final_type, unsigned long *final_size)
2432 {
2433         struct pack_window *w_curs = NULL;
2434         off_t curpos = obj_offset;
2435         void *data = NULL;
2436         unsigned long size;
2437         enum object_type type;
2438         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2439         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2440         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2441         int base_from_cache = 0;
2442
2443         write_pack_access_log(p, obj_offset);
2444
2445         /* PHASE 1: drill down to the innermost base object */
2446         for (;;) {
2447                 off_t base_offset;
2448                 int i;
2449                 struct delta_base_cache_entry *ent;
2450
2451                 ent = get_delta_base_cache_entry(p, curpos);
2452                 if (ent) {
2453                         type = ent->type;
2454                         data = ent->data;
2455                         size = ent->size;
2456                         detach_delta_base_cache_entry(ent);
2457                         base_from_cache = 1;
2458                         break;
2459                 }
2460
2461                 if (do_check_packed_object_crc && p->index_version > 1) {
2462                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2463                         off_t len = revidx[1].offset - obj_offset;
2464                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2465                                 const unsigned char *sha1 =
2466                                         nth_packed_object_sha1(p, revidx->nr);
2467                                 error("bad packed object CRC for %s",
2468                                       sha1_to_hex(sha1));
2469                                 mark_bad_packed_object(p, sha1);
2470                                 unuse_pack(&w_curs);
2471                                 return NULL;
2472                         }
2473                 }
2474
2475                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2476                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2477                         break;
2478
2479                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2480                 if (!base_offset) {
2481                         error("failed to validate delta base reference "
2482                               "at offset %"PRIuMAX" from %s",
2483                               (uintmax_t)curpos, p->pack_name);
2484                         /* bail to phase 2, in hopes of recovery */
2485                         data = NULL;
2486                         break;
2487                 }
2488
2489                 /* push object, proceed to base */
2490                 if (delta_stack_nr >= delta_stack_alloc
2491                     && delta_stack == small_delta_stack) {
2492                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2493                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2494                         memcpy(delta_stack, small_delta_stack,
2495                                sizeof(*delta_stack)*delta_stack_nr);
2496                 } else {
2497                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2498                 }
2499                 i = delta_stack_nr++;
2500                 delta_stack[i].obj_offset = obj_offset;
2501                 delta_stack[i].curpos = curpos;
2502                 delta_stack[i].size = size;
2503
2504                 curpos = obj_offset = base_offset;
2505         }
2506
2507         /* PHASE 2: handle the base */
2508         switch (type) {
2509         case OBJ_OFS_DELTA:
2510         case OBJ_REF_DELTA:
2511                 if (data)
2512                         die("BUG: unpack_entry: left loop at a valid delta");
2513                 break;
2514         case OBJ_COMMIT:
2515         case OBJ_TREE:
2516         case OBJ_BLOB:
2517         case OBJ_TAG:
2518                 if (!base_from_cache)
2519                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2520                 break;
2521         default:
2522                 data = NULL;
2523                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2524                       type, (uintmax_t)obj_offset, p->pack_name);
2525         }
2526
2527         /* PHASE 3: apply deltas in order */
2528
2529         /* invariants:
2530          *   'data' holds the base data, or NULL if there was corruption
2531          */
2532         while (delta_stack_nr) {
2533                 void *delta_data;
2534                 void *base = data;
2535                 void *external_base = NULL;
2536                 unsigned long delta_size, base_size = size;
2537                 int i;
2538
2539                 data = NULL;
2540
2541                 if (base)
2542                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2543
2544                 if (!base) {
2545                         /*
2546                          * We're probably in deep shit, but let's try to fetch
2547                          * the required base anyway from another pack or loose.
2548                          * This is costly but should happen only in the presence
2549                          * of a corrupted pack, and is better than failing outright.
2550                          */
2551                         struct revindex_entry *revidx;
2552                         const unsigned char *base_sha1;
2553                         revidx = find_pack_revindex(p, obj_offset);
2554                         if (revidx) {
2555                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2556                                 error("failed to read delta base object %s"
2557                                       " at offset %"PRIuMAX" from %s",
2558                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2559                                       p->pack_name);
2560                                 mark_bad_packed_object(p, base_sha1);
2561                                 base = read_object(base_sha1, &type, &base_size);
2562                                 external_base = base;
2563                         }
2564                 }
2565
2566                 i = --delta_stack_nr;
2567                 obj_offset = delta_stack[i].obj_offset;
2568                 curpos = delta_stack[i].curpos;
2569                 delta_size = delta_stack[i].size;
2570
2571                 if (!base)
2572                         continue;
2573
2574                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2575
2576                 if (!delta_data) {
2577                         error("failed to unpack compressed delta "
2578                               "at offset %"PRIuMAX" from %s",
2579                               (uintmax_t)curpos, p->pack_name);
2580                         data = NULL;
2581                         free(external_base);
2582                         continue;
2583                 }
2584
2585                 data = patch_delta(base, base_size,
2586                                    delta_data, delta_size,
2587                                    &size);
2588
2589                 /*
2590                  * We could not apply the delta; warn the user, but keep going.
2591                  * Our failure will be noticed either in the next iteration of
2592                  * the loop, or if this is the final delta, in the caller when
2593                  * we return NULL. Those code paths will take care of making
2594                  * a more explicit warning and retrying with another copy of
2595                  * the object.
2596                  */
2597                 if (!data)
2598                         error("failed to apply delta");
2599
2600                 free(delta_data);
2601                 free(external_base);
2602         }
2603
2604         *final_type = type;
2605         *final_size = size;
2606
2607         unuse_pack(&w_curs);
2608
2609         if (delta_stack != small_delta_stack)
2610                 free(delta_stack);
2611
2612         return data;
2613 }
2614
2615 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2616                                             uint32_t n)
2617 {
2618         const unsigned char *index = p->index_data;
2619         if (!index) {
2620                 if (open_pack_index(p))
2621                         return NULL;
2622                 index = p->index_data;
2623         }
2624         if (n >= p->num_objects)
2625                 return NULL;
2626         index += 4 * 256;
2627         if (p->index_version == 1) {
2628                 return index + 24 * n + 4;
2629         } else {
2630                 index += 8;
2631                 return index + 20 * n;
2632         }
2633 }
2634
2635 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2636 {
2637         const unsigned char *ptr = vptr;
2638         const unsigned char *start = p->index_data;
2639         const unsigned char *end = start + p->index_size;
2640         if (ptr < start)
2641                 die(_("offset before start of pack index for %s (corrupt index?)"),
2642                     p->pack_name);
2643         /* No need to check for underflow; .idx files must be at least 8 bytes */
2644         if (ptr >= end - 8)
2645                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2646                     p->pack_name);
2647 }
2648
2649 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2650 {
2651         const unsigned char *index = p->index_data;
2652         index += 4 * 256;
2653         if (p->index_version == 1) {
2654                 return ntohl(*((uint32_t *)(index + 24 * n)));
2655         } else {
2656                 uint32_t off;
2657                 index += 8 + p->num_objects * (20 + 4);
2658                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2659                 if (!(off & 0x80000000))
2660                         return off;
2661                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2662                 check_pack_index_ptr(p, index);
2663                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2664                                    ntohl(*((uint32_t *)(index + 4)));
2665         }
2666 }
2667
2668 off_t find_pack_entry_one(const unsigned char *sha1,
2669                                   struct packed_git *p)
2670 {
2671         const uint32_t *level1_ofs = p->index_data;
2672         const unsigned char *index = p->index_data;
2673         unsigned hi, lo, stride;
2674         static int use_lookup = -1;
2675         static int debug_lookup = -1;
2676
2677         if (debug_lookup < 0)
2678                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2679
2680         if (!index) {
2681                 if (open_pack_index(p))
2682                         return 0;
2683                 level1_ofs = p->index_data;
2684                 index = p->index_data;
2685         }
2686         if (p->index_version > 1) {
2687                 level1_ofs += 2;
2688                 index += 8;
2689         }
2690         index += 4 * 256;
2691         hi = ntohl(level1_ofs[*sha1]);
2692         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2693         if (p->index_version > 1) {
2694                 stride = 20;
2695         } else {
2696                 stride = 24;
2697                 index += 4;
2698         }
2699
2700         if (debug_lookup)
2701                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2702                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2703
2704         if (use_lookup < 0)
2705                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2706         if (use_lookup) {
2707                 int pos = sha1_entry_pos(index, stride, 0,
2708                                          lo, hi, p->num_objects, sha1);
2709                 if (pos < 0)
2710                         return 0;
2711                 return nth_packed_object_offset(p, pos);
2712         }
2713
2714         do {
2715                 unsigned mi = (lo + hi) / 2;
2716                 int cmp = hashcmp(index + mi * stride, sha1);
2717
2718                 if (debug_lookup)
2719                         printf("lo %u hi %u rg %u mi %u\n",
2720                                lo, hi, hi - lo, mi);
2721                 if (!cmp)
2722                         return nth_packed_object_offset(p, mi);
2723                 if (cmp > 0)
2724                         hi = mi;
2725                 else
2726                         lo = mi+1;
2727         } while (lo < hi);
2728         return 0;
2729 }
2730
2731 int is_pack_valid(struct packed_git *p)
2732 {
2733         /* An already open pack is known to be valid. */
2734         if (p->pack_fd != -1)
2735                 return 1;
2736
2737         /* If the pack has one window completely covering the
2738          * file size, the pack is known to be valid even if
2739          * the descriptor is not currently open.
2740          */
2741         if (p->windows) {
2742                 struct pack_window *w = p->windows;
2743
2744                 if (!w->offset && w->len == p->pack_size)
2745                         return 1;
2746         }
2747
2748         /* Force the pack to open to prove its valid. */
2749         return !open_packed_git(p);
2750 }
2751
2752 static int fill_pack_entry(const unsigned char *sha1,
2753                            struct pack_entry *e,
2754                            struct packed_git *p)
2755 {
2756         off_t offset;
2757
2758         if (p->num_bad_objects) {
2759                 unsigned i;
2760                 for (i = 0; i < p->num_bad_objects; i++)
2761                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2762                                 return 0;
2763         }
2764
2765         offset = find_pack_entry_one(sha1, p);
2766         if (!offset)
2767                 return 0;
2768
2769         /*
2770          * We are about to tell the caller where they can locate the
2771          * requested object.  We better make sure the packfile is
2772          * still here and can be accessed before supplying that
2773          * answer, as it may have been deleted since the index was
2774          * loaded!
2775          */
2776         if (!is_pack_valid(p))
2777                 return 0;
2778         e->offset = offset;
2779         e->p = p;
2780         hashcpy(e->sha1, sha1);
2781         return 1;
2782 }
2783
2784 /*
2785  * Iff a pack file contains the object named by sha1, return true and
2786  * store its location to e.
2787  */
2788 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2789 {
2790         struct mru_entry *p;
2791
2792         prepare_packed_git();
2793         if (!packed_git)
2794                 return 0;
2795
2796         for (p = packed_git_mru->head; p; p = p->next) {
2797                 if (fill_pack_entry(sha1, e, p->item)) {
2798                         mru_mark(packed_git_mru, p);
2799                         return 1;
2800                 }
2801         }
2802         return 0;
2803 }
2804
2805 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2806                                   struct packed_git *packs)
2807 {
2808         struct packed_git *p;
2809
2810         for (p = packs; p; p = p->next) {
2811                 if (find_pack_entry_one(sha1, p))
2812                         return p;
2813         }
2814         return NULL;
2815
2816 }
2817
2818 static int sha1_loose_object_info(const unsigned char *sha1,
2819                                   struct object_info *oi,
2820                                   int flags)
2821 {
2822         int status = 0;
2823         unsigned long mapsize;
2824         void *map;
2825         git_zstream stream;
2826         char hdr[32];
2827         struct strbuf hdrbuf = STRBUF_INIT;
2828
2829         if (oi->delta_base_sha1)
2830                 hashclr(oi->delta_base_sha1);
2831
2832         /*
2833          * If we don't care about type or size, then we don't
2834          * need to look inside the object at all. Note that we
2835          * do not optimize out the stat call, even if the
2836          * caller doesn't care about the disk-size, since our
2837          * return value implicitly indicates whether the
2838          * object even exists.
2839          */
2840         if (!oi->typep && !oi->typename && !oi->sizep) {
2841                 const char *path;
2842                 struct stat st;
2843                 if (stat_sha1_file(sha1, &st, &path) < 0)
2844                         return -1;
2845                 if (oi->disk_sizep)
2846                         *oi->disk_sizep = st.st_size;
2847                 return 0;
2848         }
2849
2850         map = map_sha1_file(sha1, &mapsize);
2851         if (!map)
2852                 return -1;
2853         if (oi->disk_sizep)
2854                 *oi->disk_sizep = mapsize;
2855         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2856                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2857                         status = error("unable to unpack %s header with --allow-unknown-type",
2858                                        sha1_to_hex(sha1));
2859         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2860                 status = error("unable to unpack %s header",
2861                                sha1_to_hex(sha1));
2862         if (status < 0)
2863                 ; /* Do nothing */
2864         else if (hdrbuf.len) {
2865                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2866                         status = error("unable to parse %s header with --allow-unknown-type",
2867                                        sha1_to_hex(sha1));
2868         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2869                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2870         git_inflate_end(&stream);
2871         munmap(map, mapsize);
2872         if (status && oi->typep)
2873                 *oi->typep = status;
2874         strbuf_release(&hdrbuf);
2875         return 0;
2876 }
2877
2878 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2879 {
2880         struct cached_object *co;
2881         struct pack_entry e;
2882         int rtype;
2883         enum object_type real_type;
2884         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2885
2886         co = find_cached_object(real);
2887         if (co) {
2888                 if (oi->typep)
2889                         *(oi->typep) = co->type;
2890                 if (oi->sizep)
2891                         *(oi->sizep) = co->size;
2892                 if (oi->disk_sizep)
2893                         *(oi->disk_sizep) = 0;
2894                 if (oi->delta_base_sha1)
2895                         hashclr(oi->delta_base_sha1);
2896                 if (oi->typename)
2897                         strbuf_addstr(oi->typename, typename(co->type));
2898                 oi->whence = OI_CACHED;
2899                 return 0;
2900         }
2901
2902         if (!find_pack_entry(real, &e)) {
2903                 /* Most likely it's a loose object. */
2904                 if (!sha1_loose_object_info(real, oi, flags)) {
2905                         oi->whence = OI_LOOSE;
2906                         return 0;
2907                 }
2908
2909                 /* Not a loose object; someone else may have just packed it. */
2910                 reprepare_packed_git();
2911                 if (!find_pack_entry(real, &e))
2912                         return -1;
2913         }
2914
2915         /*
2916          * packed_object_info() does not follow the delta chain to
2917          * find out the real type, unless it is given oi->typep.
2918          */
2919         if (oi->typename && !oi->typep)
2920                 oi->typep = &real_type;
2921
2922         rtype = packed_object_info(e.p, e.offset, oi);
2923         if (rtype < 0) {
2924                 mark_bad_packed_object(e.p, real);
2925                 if (oi->typep == &real_type)
2926                         oi->typep = NULL;
2927                 return sha1_object_info_extended(real, oi, 0);
2928         } else if (in_delta_base_cache(e.p, e.offset)) {
2929                 oi->whence = OI_DBCACHED;
2930         } else {
2931                 oi->whence = OI_PACKED;
2932                 oi->u.packed.offset = e.offset;
2933                 oi->u.packed.pack = e.p;
2934                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2935                                          rtype == OBJ_OFS_DELTA);
2936         }
2937         if (oi->typename)
2938                 strbuf_addstr(oi->typename, typename(*oi->typep));
2939         if (oi->typep == &real_type)
2940                 oi->typep = NULL;
2941
2942         return 0;
2943 }
2944
2945 /* returns enum object_type or negative */
2946 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2947 {
2948         enum object_type type;
2949         struct object_info oi = OBJECT_INFO_INIT;
2950
2951         oi.typep = &type;
2952         oi.sizep = sizep;
2953         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2954                 return -1;
2955         return type;
2956 }
2957
2958 static void *read_packed_sha1(const unsigned char *sha1,
2959                               enum object_type *type, unsigned long *size)
2960 {
2961         struct pack_entry e;
2962         void *data;
2963
2964         if (!find_pack_entry(sha1, &e))
2965                 return NULL;
2966         data = cache_or_unpack_entry(e.p, e.offset, size, type);
2967         if (!data) {
2968                 /*
2969                  * We're probably in deep shit, but let's try to fetch
2970                  * the required object anyway from another pack or loose.
2971                  * This should happen only in the presence of a corrupted
2972                  * pack, and is better than failing outright.
2973                  */
2974                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2975                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2976                 mark_bad_packed_object(e.p, sha1);
2977                 data = read_object(sha1, type, size);
2978         }
2979         return data;
2980 }
2981
2982 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2983                       unsigned char *sha1)
2984 {
2985         struct cached_object *co;
2986
2987         hash_sha1_file(buf, len, typename(type), sha1);
2988         if (has_sha1_file(sha1) || find_cached_object(sha1))
2989                 return 0;
2990         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2991         co = &cached_objects[cached_object_nr++];
2992         co->size = len;
2993         co->type = type;
2994         co->buf = xmalloc(len);
2995         memcpy(co->buf, buf, len);
2996         hashcpy(co->sha1, sha1);
2997         return 0;
2998 }
2999
3000 static void *read_object(const unsigned char *sha1, enum object_type *type,
3001                          unsigned long *size)
3002 {
3003         unsigned long mapsize;
3004         void *map, *buf;
3005         struct cached_object *co;
3006
3007         co = find_cached_object(sha1);
3008         if (co) {
3009                 *type = co->type;
3010                 *size = co->size;
3011                 return xmemdupz(co->buf, co->size);
3012         }
3013
3014         buf = read_packed_sha1(sha1, type, size);
3015         if (buf)
3016                 return buf;
3017         map = map_sha1_file(sha1, &mapsize);
3018         if (map) {
3019                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
3020                 munmap(map, mapsize);
3021                 return buf;
3022         }
3023         reprepare_packed_git();
3024         return read_packed_sha1(sha1, type, size);
3025 }
3026
3027 /*
3028  * This function dies on corrupt objects; the callers who want to
3029  * deal with them should arrange to call read_object() and give error
3030  * messages themselves.
3031  */
3032 void *read_sha1_file_extended(const unsigned char *sha1,
3033                               enum object_type *type,
3034                               unsigned long *size,
3035                               unsigned flag)
3036 {
3037         void *data;
3038         const struct packed_git *p;
3039         const char *path;
3040         struct stat st;
3041         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
3042
3043         errno = 0;
3044         data = read_object(repl, type, size);
3045         if (data)
3046                 return data;
3047
3048         if (errno && errno != ENOENT)
3049                 die_errno("failed to read object %s", sha1_to_hex(sha1));
3050
3051         /* die if we replaced an object with one that does not exist */
3052         if (repl != sha1)
3053                 die("replacement %s not found for %s",
3054                     sha1_to_hex(repl), sha1_to_hex(sha1));
3055
3056         if (!stat_sha1_file(repl, &st, &path))
3057                 die("loose object %s (stored in %s) is corrupt",
3058                     sha1_to_hex(repl), path);
3059
3060         if ((p = has_packed_and_bad(repl)) != NULL)
3061                 die("packed object %s (stored in %s) is corrupt",
3062                     sha1_to_hex(repl), p->pack_name);
3063
3064         return NULL;
3065 }
3066
3067 void *read_object_with_reference(const unsigned char *sha1,
3068                                  const char *required_type_name,
3069                                  unsigned long *size,
3070                                  unsigned char *actual_sha1_return)
3071 {
3072         enum object_type type, required_type;
3073         void *buffer;
3074         unsigned long isize;
3075         unsigned char actual_sha1[20];
3076
3077         required_type = type_from_string(required_type_name);
3078         hashcpy(actual_sha1, sha1);
3079         while (1) {
3080                 int ref_length = -1;
3081                 const char *ref_type = NULL;
3082
3083                 buffer = read_sha1_file(actual_sha1, &type, &isize);
3084                 if (!buffer)
3085                         return NULL;
3086                 if (type == required_type) {
3087                         *size = isize;
3088                         if (actual_sha1_return)
3089                                 hashcpy(actual_sha1_return, actual_sha1);
3090                         return buffer;
3091                 }
3092                 /* Handle references */
3093                 else if (type == OBJ_COMMIT)
3094                         ref_type = "tree ";
3095                 else if (type == OBJ_TAG)
3096                         ref_type = "object ";
3097                 else {
3098                         free(buffer);
3099                         return NULL;
3100                 }
3101                 ref_length = strlen(ref_type);
3102
3103                 if (ref_length + 40 > isize ||
3104                     memcmp(buffer, ref_type, ref_length) ||
3105                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
3106                         free(buffer);
3107                         return NULL;
3108                 }
3109                 free(buffer);
3110                 /* Now we have the ID of the referred-to object in
3111                  * actual_sha1.  Check again. */
3112         }
3113 }
3114
3115 static void write_sha1_file_prepare(const void *buf, unsigned long len,
3116                                     const char *type, unsigned char *sha1,
3117                                     char *hdr, int *hdrlen)
3118 {
3119         git_SHA_CTX c;
3120
3121         /* Generate the header */
3122         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
3123
3124         /* Sha1.. */
3125         git_SHA1_Init(&c);
3126         git_SHA1_Update(&c, hdr, *hdrlen);
3127         git_SHA1_Update(&c, buf, len);
3128         git_SHA1_Final(sha1, &c);
3129 }
3130
3131 /*
3132  * Move the just written object into its final resting place.
3133  */
3134 int finalize_object_file(const char *tmpfile, const char *filename)
3135 {
3136         int ret = 0;
3137
3138         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3139                 goto try_rename;
3140         else if (link(tmpfile, filename))
3141                 ret = errno;
3142
3143         /*
3144          * Coda hack - coda doesn't like cross-directory links,
3145          * so we fall back to a rename, which will mean that it
3146          * won't be able to check collisions, but that's not a
3147          * big deal.
3148          *
3149          * The same holds for FAT formatted media.
3150          *
3151          * When this succeeds, we just return.  We have nothing
3152          * left to unlink.
3153          */
3154         if (ret && ret != EEXIST) {
3155         try_rename:
3156                 if (!rename(tmpfile, filename))
3157                         goto out;
3158                 ret = errno;
3159         }
3160         unlink_or_warn(tmpfile);
3161         if (ret) {
3162                 if (ret != EEXIST) {
3163                         return error_errno("unable to write sha1 filename %s", filename);
3164                 }
3165                 /* FIXME!!! Collision check here ? */
3166         }
3167
3168 out:
3169         if (adjust_shared_perm(filename))
3170                 return error("unable to set permission to '%s'", filename);
3171         return 0;
3172 }
3173
3174 static int write_buffer(int fd, const void *buf, size_t len)
3175 {
3176         if (write_in_full(fd, buf, len) < 0)
3177                 return error_errno("file write error");
3178         return 0;
3179 }
3180
3181 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3182                    unsigned char *sha1)
3183 {
3184         char hdr[32];
3185         int hdrlen = sizeof(hdr);
3186         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3187         return 0;
3188 }
3189
3190 /* Finalize a file on disk, and close it. */
3191 static void close_sha1_file(int fd)
3192 {
3193         if (fsync_object_files)
3194                 fsync_or_die(fd, "sha1 file");
3195         if (close(fd) != 0)
3196                 die_errno("error when closing sha1 file");
3197 }
3198
3199 /* Size of directory component, including the ending '/' */
3200 static inline int directory_size(const char *filename)
3201 {
3202         const char *s = strrchr(filename, '/');
3203         if (!s)
3204                 return 0;
3205         return s - filename + 1;
3206 }
3207
3208 /*
3209  * This creates a temporary file in the same directory as the final
3210  * 'filename'
3211  *
3212  * We want to avoid cross-directory filename renames, because those
3213  * can have problems on various filesystems (FAT, NFS, Coda).
3214  */
3215 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3216 {
3217         int fd, dirlen = directory_size(filename);
3218
3219         strbuf_reset(tmp);
3220         strbuf_add(tmp, filename, dirlen);
3221         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3222         fd = git_mkstemp_mode(tmp->buf, 0444);
3223         if (fd < 0 && dirlen && errno == ENOENT) {
3224                 /*
3225                  * Make sure the directory exists; note that the contents
3226                  * of the buffer are undefined after mkstemp returns an
3227                  * error, so we have to rewrite the whole buffer from
3228                  * scratch.
3229                  */
3230                 strbuf_reset(tmp);
3231                 strbuf_add(tmp, filename, dirlen - 1);
3232                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3233                         return -1;
3234                 if (adjust_shared_perm(tmp->buf))
3235                         return -1;
3236
3237                 /* Try again */
3238                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3239                 fd = git_mkstemp_mode(tmp->buf, 0444);
3240         }
3241         return fd;
3242 }
3243
3244 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3245                               const void *buf, unsigned long len, time_t mtime)
3246 {
3247         int fd, ret;
3248         unsigned char compressed[4096];
3249         git_zstream stream;
3250         git_SHA_CTX c;
3251         unsigned char parano_sha1[20];
3252         static struct strbuf tmp_file = STRBUF_INIT;
3253         const char *filename = sha1_file_name(sha1);
3254
3255         fd = create_tmpfile(&tmp_file, filename);
3256         if (fd < 0) {
3257                 if (errno == EACCES)
3258                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3259                 else
3260                         return error_errno("unable to create temporary file");
3261         }
3262
3263         /* Set it up */
3264         git_deflate_init(&stream, zlib_compression_level);
3265         stream.next_out = compressed;
3266         stream.avail_out = sizeof(compressed);
3267         git_SHA1_Init(&c);
3268
3269         /* First header.. */
3270         stream.next_in = (unsigned char *)hdr;
3271         stream.avail_in = hdrlen;
3272         while (git_deflate(&stream, 0) == Z_OK)
3273                 ; /* nothing */
3274         git_SHA1_Update(&c, hdr, hdrlen);
3275
3276         /* Then the data itself.. */
3277         stream.next_in = (void *)buf;
3278         stream.avail_in = len;
3279         do {
3280                 unsigned char *in0 = stream.next_in;
3281                 ret = git_deflate(&stream, Z_FINISH);
3282                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3283                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3284                         die("unable to write sha1 file");
3285                 stream.next_out = compressed;
3286                 stream.avail_out = sizeof(compressed);
3287         } while (ret == Z_OK);
3288
3289         if (ret != Z_STREAM_END)
3290                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3291         ret = git_deflate_end_gently(&stream);
3292         if (ret != Z_OK)
3293                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3294         git_SHA1_Final(parano_sha1, &c);
3295         if (hashcmp(sha1, parano_sha1) != 0)
3296                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3297
3298         close_sha1_file(fd);
3299
3300         if (mtime) {
3301                 struct utimbuf utb;
3302                 utb.actime = mtime;
3303                 utb.modtime = mtime;
3304                 if (utime(tmp_file.buf, &utb) < 0)
3305                         warning_errno("failed utime() on %s", tmp_file.buf);
3306         }
3307
3308         return finalize_object_file(tmp_file.buf, filename);
3309 }
3310
3311 static int freshen_loose_object(const unsigned char *sha1)
3312 {
3313         return check_and_freshen(sha1, 1);
3314 }
3315
3316 static int freshen_packed_object(const unsigned char *sha1)
3317 {
3318         struct pack_entry e;
3319         if (!find_pack_entry(sha1, &e))
3320                 return 0;
3321         if (e.p->freshened)
3322                 return 1;
3323         if (!freshen_file(e.p->pack_name))
3324                 return 0;
3325         e.p->freshened = 1;
3326         return 1;
3327 }
3328
3329 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3330 {
3331         char hdr[32];
3332         int hdrlen = sizeof(hdr);
3333
3334         /* Normally if we have it in the pack then we do not bother writing
3335          * it out into .git/objects/??/?{38} file.
3336          */
3337         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3338         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3339                 return 0;
3340         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3341 }
3342
3343 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3344                              unsigned char *sha1, unsigned flags)
3345 {
3346         char *header;
3347         int hdrlen, status = 0;
3348
3349         /* type string, SP, %lu of the length plus NUL must fit this */
3350         hdrlen = strlen(type) + 32;
3351         header = xmalloc(hdrlen);
3352         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3353
3354         if (!(flags & HASH_WRITE_OBJECT))
3355                 goto cleanup;
3356         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3357                 goto cleanup;
3358         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3359
3360 cleanup:
3361         free(header);
3362         return status;
3363 }
3364
3365 int force_object_loose(const unsigned char *sha1, time_t mtime)
3366 {
3367         void *buf;
3368         unsigned long len;
3369         enum object_type type;
3370         char hdr[32];
3371         int hdrlen;
3372         int ret;
3373
3374         if (has_loose_object(sha1))
3375                 return 0;
3376         buf = read_packed_sha1(sha1, &type, &len);
3377         if (!buf)
3378                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3379         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3380         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3381         free(buf);
3382
3383         return ret;
3384 }
3385
3386 int has_pack_index(const unsigned char *sha1)
3387 {
3388         struct stat st;
3389         if (stat(sha1_pack_index_name(sha1), &st))
3390                 return 0;
3391         return 1;
3392 }
3393
3394 int has_sha1_pack(const unsigned char *sha1)
3395 {
3396         struct pack_entry e;
3397         return find_pack_entry(sha1, &e);
3398 }
3399
3400 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3401 {
3402         struct pack_entry e;
3403
3404         if (find_pack_entry(sha1, &e))
3405                 return 1;
3406         if (has_loose_object(sha1))
3407                 return 1;
3408         if (flags & HAS_SHA1_QUICK)
3409                 return 0;
3410         reprepare_packed_git();
3411         return find_pack_entry(sha1, &e);
3412 }
3413
3414 int has_object_file(const struct object_id *oid)
3415 {
3416         return has_sha1_file(oid->hash);
3417 }
3418
3419 int has_object_file_with_flags(const struct object_id *oid, int flags)
3420 {
3421         return has_sha1_file_with_flags(oid->hash, flags);
3422 }
3423
3424 static void check_tree(const void *buf, size_t size)
3425 {
3426         struct tree_desc desc;
3427         struct name_entry entry;
3428
3429         init_tree_desc(&desc, buf, size);
3430         while (tree_entry(&desc, &entry))
3431                 /* do nothing
3432                  * tree_entry() will die() on malformed entries */
3433                 ;
3434 }
3435
3436 static void check_commit(const void *buf, size_t size)
3437 {
3438         struct commit c;
3439         memset(&c, 0, sizeof(c));
3440         if (parse_commit_buffer(&c, buf, size))
3441                 die("corrupt commit");
3442 }
3443
3444 static void check_tag(const void *buf, size_t size)
3445 {
3446         struct tag t;
3447         memset(&t, 0, sizeof(t));
3448         if (parse_tag_buffer(&t, buf, size))
3449                 die("corrupt tag");
3450 }
3451
3452 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3453                      enum object_type type,
3454                      const char *path, unsigned flags)
3455 {
3456         int ret, re_allocated = 0;
3457         int write_object = flags & HASH_WRITE_OBJECT;
3458
3459         if (!type)
3460                 type = OBJ_BLOB;
3461
3462         /*
3463          * Convert blobs to git internal format
3464          */
3465         if ((type == OBJ_BLOB) && path) {
3466                 struct strbuf nbuf = STRBUF_INIT;
3467                 if (convert_to_git(path, buf, size, &nbuf,
3468                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3469                         buf = strbuf_detach(&nbuf, &size);
3470                         re_allocated = 1;
3471                 }
3472         }
3473         if (flags & HASH_FORMAT_CHECK) {
3474                 if (type == OBJ_TREE)
3475                         check_tree(buf, size);
3476                 if (type == OBJ_COMMIT)
3477                         check_commit(buf, size);
3478                 if (type == OBJ_TAG)
3479                         check_tag(buf, size);
3480         }
3481
3482         if (write_object)
3483                 ret = write_sha1_file(buf, size, typename(type), sha1);
3484         else
3485                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3486         if (re_allocated)
3487                 free(buf);
3488         return ret;
3489 }
3490
3491 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3492                                      const char *path, unsigned flags)
3493 {
3494         int ret;
3495         const int write_object = flags & HASH_WRITE_OBJECT;
3496         struct strbuf sbuf = STRBUF_INIT;
3497
3498         assert(path);
3499         assert(would_convert_to_git_filter_fd(path));
3500
3501         convert_to_git_filter_fd(path, fd, &sbuf,
3502                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3503
3504         if (write_object)
3505                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3506                                       sha1);
3507         else
3508                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3509                                      sha1);
3510         strbuf_release(&sbuf);
3511         return ret;
3512 }
3513
3514 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3515                       const char *path, unsigned flags)
3516 {
3517         struct strbuf sbuf = STRBUF_INIT;
3518         int ret;
3519
3520         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3521                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3522         else
3523                 ret = -1;
3524         strbuf_release(&sbuf);
3525         return ret;
3526 }
3527
3528 #define SMALL_FILE_SIZE (32*1024)
3529
3530 static int index_core(unsigned char *sha1, int fd, size_t size,
3531                       enum object_type type, const char *path,
3532                       unsigned flags)
3533 {
3534         int ret;
3535
3536         if (!size) {
3537                 ret = index_mem(sha1, "", size, type, path, flags);
3538         } else if (size <= SMALL_FILE_SIZE) {
3539                 char *buf = xmalloc(size);
3540                 if (size == read_in_full(fd, buf, size))
3541                         ret = index_mem(sha1, buf, size, type, path, flags);
3542                 else
3543                         ret = error_errno("short read");
3544                 free(buf);
3545         } else {
3546                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3547                 ret = index_mem(sha1, buf, size, type, path, flags);
3548                 munmap(buf, size);
3549         }
3550         return ret;
3551 }
3552
3553 /*
3554  * This creates one packfile per large blob unless bulk-checkin
3555  * machinery is "plugged".
3556  *
3557  * This also bypasses the usual "convert-to-git" dance, and that is on
3558  * purpose. We could write a streaming version of the converting
3559  * functions and insert that before feeding the data to fast-import
3560  * (or equivalent in-core API described above). However, that is
3561  * somewhat complicated, as we do not know the size of the filter
3562  * result, which we need to know beforehand when writing a git object.
3563  * Since the primary motivation for trying to stream from the working
3564  * tree file and to avoid mmaping it in core is to deal with large
3565  * binary blobs, they generally do not want to get any conversion, and
3566  * callers should avoid this code path when filters are requested.
3567  */
3568 static int index_stream(unsigned char *sha1, int fd, size_t size,
3569                         enum object_type type, const char *path,
3570                         unsigned flags)
3571 {
3572         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3573 }
3574
3575 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3576              enum object_type type, const char *path, unsigned flags)
3577 {
3578         int ret;
3579
3580         /*
3581          * Call xsize_t() only when needed to avoid potentially unnecessary
3582          * die() for large files.
3583          */
3584         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3585                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3586         else if (!S_ISREG(st->st_mode))
3587                 ret = index_pipe(sha1, fd, type, path, flags);
3588         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3589                  (path && would_convert_to_git(path)))
3590                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3591                                  flags);
3592         else
3593                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3594                                    flags);
3595         close(fd);
3596         return ret;
3597 }
3598
3599 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3600 {
3601         int fd;
3602         struct strbuf sb = STRBUF_INIT;
3603
3604         switch (st->st_mode & S_IFMT) {
3605         case S_IFREG:
3606                 fd = open(path, O_RDONLY);
3607                 if (fd < 0)
3608                         return error_errno("open(\"%s\")", path);
3609                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3610                         return error("%s: failed to insert into database",
3611                                      path);
3612                 break;
3613         case S_IFLNK:
3614                 if (strbuf_readlink(&sb, path, st->st_size))
3615                         return error_errno("readlink(\"%s\")", path);
3616                 if (!(flags & HASH_WRITE_OBJECT))
3617                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3618                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3619                         return error("%s: failed to insert into database",
3620                                      path);
3621                 strbuf_release(&sb);
3622                 break;
3623         case S_IFDIR:
3624                 return resolve_gitlink_ref(path, "HEAD", sha1);
3625         default:
3626                 return error("%s: unsupported file type", path);
3627         }
3628         return 0;
3629 }
3630
3631 int read_pack_header(int fd, struct pack_header *header)
3632 {
3633         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3634                 /* "eof before pack header was fully read" */
3635                 return PH_ERROR_EOF;
3636
3637         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3638                 /* "protocol error (pack signature mismatch detected)" */
3639                 return PH_ERROR_PACK_SIGNATURE;
3640         if (!pack_version_ok(header->hdr_version))
3641                 /* "protocol error (pack version unsupported)" */
3642                 return PH_ERROR_PROTOCOL;
3643         return 0;
3644 }
3645
3646 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3647 {
3648         enum object_type type = sha1_object_info(sha1, NULL);
3649         if (type < 0)
3650                 die("%s is not a valid object", sha1_to_hex(sha1));
3651         if (type != expect)
3652                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3653                     typename(expect));
3654 }
3655
3656 static int for_each_file_in_obj_subdir(int subdir_nr,
3657                                        struct strbuf *path,
3658                                        each_loose_object_fn obj_cb,
3659                                        each_loose_cruft_fn cruft_cb,
3660                                        each_loose_subdir_fn subdir_cb,
3661                                        void *data)
3662 {
3663         size_t baselen = path->len;
3664         DIR *dir = opendir(path->buf);
3665         struct dirent *de;
3666         int r = 0;
3667
3668         if (!dir) {
3669                 if (errno == ENOENT)
3670                         return 0;
3671                 return error_errno("unable to open %s", path->buf);
3672         }
3673
3674         while ((de = readdir(dir))) {
3675                 if (is_dot_or_dotdot(de->d_name))
3676                         continue;
3677
3678                 strbuf_setlen(path, baselen);
3679                 strbuf_addf(path, "/%s", de->d_name);
3680
3681                 if (strlen(de->d_name) == 38)  {
3682                         char hex[41];
3683                         unsigned char sha1[20];
3684
3685                         snprintf(hex, sizeof(hex), "%02x%s",
3686                                  subdir_nr, de->d_name);
3687                         if (!get_sha1_hex(hex, sha1)) {
3688                                 if (obj_cb) {
3689                                         r = obj_cb(sha1, path->buf, data);
3690                                         if (r)
3691                                                 break;
3692                                 }
3693                                 continue;
3694                         }
3695                 }
3696
3697                 if (cruft_cb) {
3698                         r = cruft_cb(de->d_name, path->buf, data);
3699                         if (r)
3700                                 break;
3701                 }
3702         }
3703         closedir(dir);
3704
3705         strbuf_setlen(path, baselen);
3706         if (!r && subdir_cb)
3707                 r = subdir_cb(subdir_nr, path->buf, data);
3708
3709         return r;
3710 }
3711
3712 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3713                             each_loose_object_fn obj_cb,
3714                             each_loose_cruft_fn cruft_cb,
3715                             each_loose_subdir_fn subdir_cb,
3716                             void *data)
3717 {
3718         size_t baselen = path->len;
3719         int r = 0;
3720         int i;
3721
3722         for (i = 0; i < 256; i++) {
3723                 strbuf_addf(path, "/%02x", i);
3724                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3725                                                 subdir_cb, data);
3726                 strbuf_setlen(path, baselen);
3727                 if (r)
3728                         break;
3729         }
3730
3731         return r;
3732 }
3733
3734 int for_each_loose_file_in_objdir(const char *path,
3735                                   each_loose_object_fn obj_cb,
3736                                   each_loose_cruft_fn cruft_cb,
3737                                   each_loose_subdir_fn subdir_cb,
3738                                   void *data)
3739 {
3740         struct strbuf buf = STRBUF_INIT;
3741         int r;
3742
3743         strbuf_addstr(&buf, path);
3744         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3745                                               subdir_cb, data);
3746         strbuf_release(&buf);
3747
3748         return r;
3749 }
3750
3751 struct loose_alt_odb_data {
3752         each_loose_object_fn *cb;
3753         void *data;
3754 };
3755
3756 static int loose_from_alt_odb(struct alternate_object_database *alt,
3757                               void *vdata)
3758 {
3759         struct loose_alt_odb_data *data = vdata;
3760         struct strbuf buf = STRBUF_INIT;
3761         int r;
3762
3763         strbuf_addstr(&buf, alt->path);
3764         r = for_each_loose_file_in_objdir_buf(&buf,
3765                                               data->cb, NULL, NULL,
3766                                               data->data);
3767         strbuf_release(&buf);
3768         return r;
3769 }
3770
3771 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3772 {
3773         struct loose_alt_odb_data alt;
3774         int r;
3775
3776         r = for_each_loose_file_in_objdir(get_object_directory(),
3777                                           cb, NULL, NULL, data);
3778         if (r)
3779                 return r;
3780
3781         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3782                 return 0;
3783
3784         alt.cb = cb;
3785         alt.data = data;
3786         return foreach_alt_odb(loose_from_alt_odb, &alt);
3787 }
3788
3789 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3790 {
3791         uint32_t i;
3792         int r = 0;
3793
3794         for (i = 0; i < p->num_objects; i++) {
3795                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3796
3797                 if (!sha1)
3798                         return error("unable to get sha1 of object %u in %s",
3799                                      i, p->pack_name);
3800
3801                 r = cb(sha1, p, i, data);
3802                 if (r)
3803                         break;
3804         }
3805         return r;
3806 }
3807
3808 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3809 {
3810         struct packed_git *p;
3811         int r = 0;
3812         int pack_errors = 0;
3813
3814         prepare_packed_git();
3815         for (p = packed_git; p; p = p->next) {
3816                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3817                         continue;
3818                 if (open_pack_index(p)) {
3819                         pack_errors = 1;
3820                         continue;
3821                 }
3822                 r = for_each_object_in_pack(p, cb, data);
3823                 if (r)
3824                         break;
3825         }
3826         return r ? r : pack_errors;
3827 }
3828
3829 static int check_stream_sha1(git_zstream *stream,
3830                              const char *hdr,
3831                              unsigned long size,
3832                              const char *path,
3833                              const unsigned char *expected_sha1)
3834 {
3835         git_SHA_CTX c;
3836         unsigned char real_sha1[GIT_SHA1_RAWSZ];
3837         unsigned char buf[4096];
3838         unsigned long total_read;
3839         int status = Z_OK;
3840
3841         git_SHA1_Init(&c);
3842         git_SHA1_Update(&c, hdr, stream->total_out);
3843
3844         /*
3845          * We already read some bytes into hdr, but the ones up to the NUL
3846          * do not count against the object's content size.
3847          */
3848         total_read = stream->total_out - strlen(hdr) - 1;
3849
3850         /*
3851          * This size comparison must be "<=" to read the final zlib packets;
3852          * see the comment in unpack_sha1_rest for details.
3853          */
3854         while (total_read <= size &&
3855                (status == Z_OK || status == Z_BUF_ERROR)) {
3856                 stream->next_out = buf;
3857                 stream->avail_out = sizeof(buf);
3858                 if (size - total_read < stream->avail_out)
3859                         stream->avail_out = size - total_read;
3860                 status = git_inflate(stream, Z_FINISH);
3861                 git_SHA1_Update(&c, buf, stream->next_out - buf);
3862                 total_read += stream->next_out - buf;
3863         }
3864         git_inflate_end(stream);
3865
3866         if (status != Z_STREAM_END) {
3867                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
3868                 return -1;
3869         }
3870         if (stream->avail_in) {
3871                 error("garbage at end of loose object '%s'",
3872                       sha1_to_hex(expected_sha1));
3873                 return -1;
3874         }
3875
3876         git_SHA1_Final(real_sha1, &c);
3877         if (hashcmp(expected_sha1, real_sha1)) {
3878                 error("sha1 mismatch for %s (expected %s)", path,
3879                       sha1_to_hex(expected_sha1));
3880                 return -1;
3881         }
3882
3883         return 0;
3884 }
3885
3886 int read_loose_object(const char *path,
3887                       const unsigned char *expected_sha1,
3888                       enum object_type *type,
3889                       unsigned long *size,
3890                       void **contents)
3891 {
3892         int ret = -1;
3893         int fd = -1;
3894         void *map = NULL;
3895         unsigned long mapsize;
3896         git_zstream stream;
3897         char hdr[32];
3898
3899         *contents = NULL;
3900
3901         map = map_sha1_file_1(path, NULL, &mapsize);
3902         if (!map) {
3903                 error_errno("unable to mmap %s", path);
3904                 goto out;
3905         }
3906
3907         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
3908                 error("unable to unpack header of %s", path);
3909                 goto out;
3910         }
3911
3912         *type = parse_sha1_header(hdr, size);
3913         if (*type < 0) {
3914                 error("unable to parse header of %s", path);
3915                 git_inflate_end(&stream);
3916                 goto out;
3917         }
3918
3919         if (*type == OBJ_BLOB) {
3920                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
3921                         goto out;
3922         } else {
3923                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
3924                 if (!*contents) {
3925                         error("unable to unpack contents of %s", path);
3926                         git_inflate_end(&stream);
3927                         goto out;
3928                 }
3929                 if (check_sha1_signature(expected_sha1, *contents,
3930                                          *size, typename(*type))) {
3931                         error("sha1 mismatch for %s (expected %s)", path,
3932                               sha1_to_hex(expected_sha1));
3933                         free(*contents);
3934                         goto out;
3935                 }
3936         }
3937
3938         ret = 0; /* everything checks out */
3939
3940 out:
3941         if (map)
3942                 munmap(map, mapsize);
3943         if (fd >= 0)
3944                 close(fd);
3945         return ret;
3946 }