sha1_file.c

   1 /*
   2  * GIT - The information manager from hell
   3  *
   4  * Copyright (C) Linus Torvalds, 2005
   5  *
   6  * This handles basic git sha1 object files - packing, unpacking,
   7  * creation etc.
   8  */
   9 #include "cache.h"
  10 #include "string-list.h"
  11 #include "lockfile.h"
  12 #include "delta.h"
  13 #include "pack.h"
  14 #include "blob.h"
  15 #include "commit.h"
  16 #include "run-command.h"
  17 #include "tag.h"
  18 #include "tree.h"
  19 #include "tree-walk.h"
  20 #include "refs.h"
  21 #include "pack-revindex.h"
  22 #include "sha1-lookup.h"
  23 #include "bulk-checkin.h"
  24 #include "streaming.h"
  25 #include "dir.h"
  26
  27 #ifndef O_NOATIME
  28 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
  29 #define O_NOATIME 01000000
  30 #else
  31 #define O_NOATIME 0
  32 #endif
  33 #endif
  34
  35 #define SZ_FMT PRIuMAX
  36 static inline uintmax_t sz_fmt(size_t s) { return s; }
  37
  38 const unsigned char null_sha1[20];
  39
  40 /*
  41  * This is meant to hold a *small* number of objects that you would
  42  * want read_sha1_file() to be able to return, but yet you do not want
  43  * to write them into the object store (e.g. a browse-only
  44  * application).
  45  */
  46 static struct cached_object {
  47         unsigned char sha1[20];
  48         enum object_type type;
  49         void *buf;
  50         unsigned long size;
  51 } *cached_objects;
  52 static int cached_object_nr, cached_object_alloc;
  53
  54 static struct cached_object empty_tree = {
  55         EMPTY_TREE_SHA1_BIN_LITERAL,
  56         OBJ_TREE,
  57         "",
  58         0
  59 };
  60
  61 /*
  62  * A pointer to the last packed_git in which an object was found.
  63  * When an object is sought, we look in this packfile first, because
  64  * objects that are looked up at similar times are often in the same
  65  * packfile as one another.
  66  */
  67 static struct packed_git *last_found_pack;
  68
  69 static struct cached_object *find_cached_object(const unsigned char *sha1)
  70 {
  71         int i;
  72         struct cached_object *co = cached_objects;
  73
  74         for (i = 0; i < cached_object_nr; i++, co++) {
  75                 if (!hashcmp(co->sha1, sha1))
  76                         return co;
  77         }
  78         if (!hashcmp(sha1, empty_tree.sha1))
  79                 return &empty_tree;
  80         return NULL;
  81 }
  82
  83 int mkdir_in_gitdir(const char *path)
  84 {
  85         if (mkdir(path, 0777)) {
  86                 int saved_errno = errno;
  87                 struct stat st;
  88                 struct strbuf sb = STRBUF_INIT;
  89
  90                 if (errno != EEXIST)
  91                         return -1;
  92                 /*
  93                  * Are we looking at a path in a symlinked worktree
  94                  * whose original repository does not yet have it?
  95                  * e.g. .git/rr-cache pointing at its original
  96                  * repository in which the user hasn't performed any
  97                  * conflict resolution yet?
  98                  */
  99                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
 100                     strbuf_readlink(&sb, path, st.st_size) ||
 101                     !is_absolute_path(sb.buf) ||
 102                     mkdir(sb.buf, 0777)) {
 103                         strbuf_release(&sb);
 104                         errno = saved_errno;
 105                         return -1;
 106                 }
 107                 strbuf_release(&sb);
 108         }
 109         return adjust_shared_perm(path);
 110 }
 111
 112 enum scld_error safe_create_leading_directories(char *path)
 113 {
 114         char *next_component = path + offset_1st_component(path);
 115         enum scld_error ret = SCLD_OK;
 116
 117         while (ret == SCLD_OK && next_component) {
 118                 struct stat st;
 119                 char *slash = next_component, slash_character;
 120
 121                 while (*slash && !is_dir_sep(*slash))
 122                         slash++;
 123
 124                 if (!*slash)
 125                         break;
 126
 127                 next_component = slash + 1;
 128                 while (is_dir_sep(*next_component))
 129                         next_component++;
 130                 if (!*next_component)
 131                         break;
 132
 133                 slash_character = *slash;
 134                 *slash = '\0';
 135                 if (!stat(path, &st)) {
 136                         /* path exists */
 137                         if (!S_ISDIR(st.st_mode))
 138                                 ret = SCLD_EXISTS;
 139                 } else if (mkdir(path, 0777)) {
 140                         if (errno == EEXIST &&
 141                             !stat(path, &st) && S_ISDIR(st.st_mode))
 142                                 ; /* somebody created it since we checked */
 143                         else if (errno == ENOENT)
 144                                 /*
 145                                  * Either mkdir() failed because
 146                                  * somebody just pruned the containing
 147                                  * directory, or stat() failed because
 148                                  * the file that was in our way was
 149                                  * just removed.  Either way, inform
 150                                  * the caller that it might be worth
 151                                  * trying again:
 152                                  */
 153                                 ret = SCLD_VANISHED;
 154                         else
 155                                 ret = SCLD_FAILED;
 156                 } else if (adjust_shared_perm(path)) {
 157                         ret = SCLD_PERMS;
 158                 }
 159                 *slash = slash_character;
 160         }
 161         return ret;
 162 }
 163
 164 enum scld_error safe_create_leading_directories_const(const char *path)
 165 {
 166         /* path points to cache entries, so xstrdup before messing with it */
 167         char *buf = xstrdup(path);
 168         enum scld_error result = safe_create_leading_directories(buf);
 169         free(buf);
 170         return result;
 171 }
 172
 173 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
 174 {
 175         int i;
 176         for (i = 0; i < 20; i++) {
 177                 static char hex[] = "0123456789abcdef";
 178                 unsigned int val = sha1[i];
 179                 char *pos = pathbuf + i*2 + (i > 0);
 180                 *pos++ = hex[val >> 4];
 181                 *pos = hex[val & 0xf];
 182         }
 183 }
 184
 185 const char *sha1_file_name(const unsigned char *sha1)
 186 {
 187         static char buf[PATH_MAX];
 188         const char *objdir;
 189         int len;
 190
 191         objdir = get_object_directory();
 192         len = strlen(objdir);
 193
 194         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
 195         if (len + 43 > PATH_MAX)
 196                 die("insanely long object directory %s", objdir);
 197         memcpy(buf, objdir, len);
 198         buf[len] = '/';
 199         buf[len+3] = '/';
 200         buf[len+42] = '\0';
 201         fill_sha1_path(buf + len + 1, sha1);
 202         return buf;
 203 }
 204
 205 /*
 206  * Return the name of the pack or index file with the specified sha1
 207  * in its filename.  *base and *name are scratch space that must be
 208  * provided by the caller.  which should be "pack" or "idx".
 209  */
 210 static char *sha1_get_pack_name(const unsigned char *sha1,
 211                                 char **name, char **base, const char *which)
 212 {
 213         static const char hex[] = "0123456789abcdef";
 214         char *buf;
 215         int i;
 216
 217         if (!*base) {
 218                 const char *sha1_file_directory = get_object_directory();
 219                 int len = strlen(sha1_file_directory);
 220                 *base = xmalloc(len + 60);
 221                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 222                         sha1_file_directory, which);
 223                 *name = *base + len + 11;
 224         }
 225
 226         buf = *name;
 227
 228         for (i = 0; i < 20; i++) {
 229                 unsigned int val = *sha1++;
 230                 *buf++ = hex[val >> 4];
 231                 *buf++ = hex[val & 0xf];
 232         }
 233
 234         return *base;
 235 }
 236
 237 char *sha1_pack_name(const unsigned char *sha1)
 238 {
 239         static char *name, *base;
 240
 241         return sha1_get_pack_name(sha1, &name, &base, "pack");
 242 }
 243
 244 char *sha1_pack_index_name(const unsigned char *sha1)
 245 {
 246         static char *name, *base;
 247
 248         return sha1_get_pack_name(sha1, &name, &base, "idx");
 249 }
 250
 251 struct alternate_object_database *alt_odb_list;
 252 static struct alternate_object_database **alt_odb_tail;
 253
 254 /*
 255  * Prepare alternate object database registry.
 256  *
 257  * The variable alt_odb_list points at the list of struct
 258  * alternate_object_database.  The elements on this list come from
 259  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 260  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 261  * whose contents is similar to that environment variable but can be
 262  * LF separated.  Its base points at a statically allocated buffer that
 263  * contains "/the/directory/corresponding/to/.git/objects/...", while
 264  * its name points just after the slash at the end of ".git/objects/"
 265  * in the example above, and has enough space to hold 40-byte hex
 266  * SHA1, an extra slash for the first level indirection, and the
 267  * terminating NUL.
 268  */
 269 static int link_alt_odb_entry(const char *entry, const char *relative_base,
 270         int depth, const char *normalized_objdir)
 271 {
 272         struct alternate_object_database *ent;
 273         struct alternate_object_database *alt;
 274         int pfxlen, entlen;
 275         struct strbuf pathbuf = STRBUF_INIT;
 276
 277         if (!is_absolute_path(entry) && relative_base) {
 278                 strbuf_addstr(&pathbuf, real_path(relative_base));
 279                 strbuf_addch(&pathbuf, '/');
 280         }
 281         strbuf_addstr(&pathbuf, entry);
 282
 283         normalize_path_copy(pathbuf.buf, pathbuf.buf);
 284
 285         pfxlen = strlen(pathbuf.buf);
 286
 287         /*
 288          * The trailing slash after the directory name is given by
 289          * this function at the end. Remove duplicates.
 290          */
 291         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
 292                 pfxlen -= 1;
 293
 294         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
 295         ent = xmalloc(sizeof(*ent) + entlen);
 296         memcpy(ent->base, pathbuf.buf, pfxlen);
 297         strbuf_release(&pathbuf);
 298
 299         ent->name = ent->base + pfxlen + 1;
 300         ent->base[pfxlen + 3] = '/';
 301         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 302
 303         /* Detect cases where alternate disappeared */
 304         if (!is_directory(ent->base)) {
 305                 error("object directory %s does not exist; "
 306                       "check .git/objects/info/alternates.",
 307                       ent->base);
 308                 free(ent);
 309                 return -1;
 310         }
 311
 312         /* Prevent the common mistake of listing the same
 313          * thing twice, or object directory itself.
 314          */
 315         for (alt = alt_odb_list; alt; alt = alt->next) {
 316                 if (pfxlen == alt->name - alt->base - 1 &&
 317                     !memcmp(ent->base, alt->base, pfxlen)) {
 318                         free(ent);
 319                         return -1;
 320                 }
 321         }
 322         if (!strcmp_icase(ent->base, normalized_objdir)) {
 323                 free(ent);
 324                 return -1;
 325         }
 326
 327         /* add the alternate entry */
 328         *alt_odb_tail = ent;
 329         alt_odb_tail = &(ent->next);
 330         ent->next = NULL;
 331
 332         /* recursively add alternates */
 333         read_info_alternates(ent->base, depth + 1);
 334
 335         ent->base[pfxlen] = '/';
 336
 337         return 0;
 338 }
 339
 340 static void link_alt_odb_entries(const char *alt, int len, int sep,
 341                                  const char *relative_base, int depth)
 342 {
 343         struct string_list entries = STRING_LIST_INIT_NODUP;
 344         char *alt_copy;
 345         int i;
 346         struct strbuf objdirbuf = STRBUF_INIT;
 347
 348         if (depth > 5) {
 349                 error("%s: ignoring alternate object stores, nesting too deep.",
 350                                 relative_base);
 351                 return;
 352         }
 353
 354         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
 355         normalize_path_copy(objdirbuf.buf, objdirbuf.buf);
 356
 357         alt_copy = xmemdupz(alt, len);
 358         string_list_split_in_place(&entries, alt_copy, sep, -1);
 359         for (i = 0; i < entries.nr; i++) {
 360                 const char *entry = entries.items[i].string;
 361                 if (entry[0] == '\0' || entry[0] == '#')
 362                         continue;
 363                 if (!is_absolute_path(entry) && depth) {
 364                         error("%s: ignoring relative alternate object store %s",
 365                                         relative_base, entry);
 366                 } else {
 367                         link_alt_odb_entry(entry, relative_base, depth, objdirbuf.buf);
 368                 }
 369         }
 370         string_list_clear(&entries, 0);
 371         free(alt_copy);
 372         strbuf_release(&objdirbuf);
 373 }
 374
 375 void read_info_alternates(const char * relative_base, int depth)
 376 {
 377         char *map;
 378         size_t mapsz;
 379         struct stat st;
 380         const char alt_file_name[] = "info/alternates";
 381         /* Given that relative_base is no longer than PATH_MAX,
 382            ensure that "path" has enough space to append "/", the
 383            file name, "info/alternates", and a trailing NUL.  */
 384         char path[PATH_MAX + 1 + sizeof alt_file_name];
 385         int fd;
 386
 387         sprintf(path, "%s/%s", relative_base, alt_file_name);
 388         fd = git_open_noatime(path);
 389         if (fd < 0)
 390                 return;
 391         if (fstat(fd, &st) || (st.st_size == 0)) {
 392                 close(fd);
 393                 return;
 394         }
 395         mapsz = xsize_t(st.st_size);
 396         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 397         close(fd);
 398
 399         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
 400
 401         munmap(map, mapsz);
 402 }
 403
 404 void add_to_alternates_file(const char *reference)
 405 {
 406         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 407         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
 408         char *alt = mkpath("%s\n", reference);
 409         write_or_die(fd, alt, strlen(alt));
 410         if (commit_lock_file(lock))
 411                 die("could not close alternates file");
 412         if (alt_odb_tail)
 413                 link_alt_odb_entries(alt, strlen(alt), '\n', NULL, 0);
 414 }
 415
 416 int foreach_alt_odb(alt_odb_fn fn, void *cb)
 417 {
 418         struct alternate_object_database *ent;
 419         int r = 0;
 420
 421         prepare_alt_odb();
 422         for (ent = alt_odb_list; ent; ent = ent->next) {
 423                 r = fn(ent, cb);
 424                 if (r)
 425                         break;
 426         }
 427         return r;
 428 }
 429
 430 void prepare_alt_odb(void)
 431 {
 432         const char *alt;
 433
 434         if (alt_odb_tail)
 435                 return;
 436
 437         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 438         if (!alt) alt = "";
 439
 440         alt_odb_tail = &alt_odb_list;
 441         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
 442
 443         read_info_alternates(get_object_directory(), 0);
 444 }
 445
 446 static int freshen_file(const char *fn)
 447 {
 448         struct utimbuf t;
 449         t.actime = t.modtime = time(NULL);
 450         return !utime(fn, &t);
 451 }
 452
 453 static int check_and_freshen_file(const char *fn, int freshen)
 454 {
 455         if (access(fn, F_OK))
 456                 return 0;
 457         if (freshen && freshen_file(fn))
 458                 return 0;
 459         return 1;
 460 }
 461
 462 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
 463 {
 464         return check_and_freshen_file(sha1_file_name(sha1), freshen);
 465 }
 466
 467 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
 468 {
 469         struct alternate_object_database *alt;
 470         prepare_alt_odb();
 471         for (alt = alt_odb_list; alt; alt = alt->next) {
 472                 fill_sha1_path(alt->name, sha1);
 473                 if (check_and_freshen_file(alt->base, freshen))
 474                         return 1;
 475         }
 476         return 0;
 477 }
 478
 479 static int check_and_freshen(const unsigned char *sha1, int freshen)
 480 {
 481         return check_and_freshen_local(sha1, freshen) ||
 482                check_and_freshen_nonlocal(sha1, freshen);
 483 }
 484
 485 int has_loose_object_nonlocal(const unsigned char *sha1)
 486 {
 487         return check_and_freshen_nonlocal(sha1, 0);
 488 }
 489
 490 static int has_loose_object(const unsigned char *sha1)
 491 {
 492         return check_and_freshen(sha1, 0);
 493 }
 494
 495 static unsigned int pack_used_ctr;
 496 static unsigned int pack_mmap_calls;
 497 static unsigned int peak_pack_open_windows;
 498 static unsigned int pack_open_windows;
 499 static unsigned int pack_open_fds;
 500 static unsigned int pack_max_fds;
 501 static size_t peak_pack_mapped;
 502 static size_t pack_mapped;
 503 struct packed_git *packed_git;
 504
 505 void pack_report(void)
 506 {
 507         fprintf(stderr,
 508                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 509                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 510                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 511                 sz_fmt(getpagesize()),
 512                 sz_fmt(packed_git_window_size),
 513                 sz_fmt(packed_git_limit));
 514         fprintf(stderr,
 515                 "pack_report: pack_used_ctr            = %10u\n"
 516                 "pack_report: pack_mmap_calls          = %10u\n"
 517                 "pack_report: pack_open_windows        = %10u / %10u\n"
 518                 "pack_report: pack_mapped              = "
 519                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 520                 pack_used_ctr,
 521                 pack_mmap_calls,
 522                 pack_open_windows, peak_pack_open_windows,
 523                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 524 }
 525
 526 /*
 527  * Open and mmap the index file at path, perform a couple of
 528  * consistency checks, then record its information to p.  Return 0 on
 529  * success.
 530  */
 531 static int check_packed_git_idx(const char *path, struct packed_git *p)
 532 {
 533         void *idx_map;
 534         struct pack_idx_header *hdr;
 535         size_t idx_size;
 536         uint32_t version, nr, i, *index;
 537         int fd = git_open_noatime(path);
 538         struct stat st;
 539
 540         if (fd < 0)
 541                 return -1;
 542         if (fstat(fd, &st)) {
 543                 close(fd);
 544                 return -1;
 545         }
 546         idx_size = xsize_t(st.st_size);
 547         if (idx_size < 4 * 256 + 20 + 20) {
 548                 close(fd);
 549                 return error("index file %s is too small", path);
 550         }
 551         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 552         close(fd);
 553
 554         hdr = idx_map;
 555         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 556                 version = ntohl(hdr->idx_version);
 557                 if (version < 2 || version > 2) {
 558                         munmap(idx_map, idx_size);
 559                         return error("index file %s is version %"PRIu32
 560                                      " and is not supported by this binary"
 561                                      " (try upgrading GIT to a newer version)",
 562                                      path, version);
 563                 }
 564         } else
 565                 version = 1;
 566
 567         nr = 0;
 568         index = idx_map;
 569         if (version > 1)
 570                 index += 2;  /* skip index header */
 571         for (i = 0; i < 256; i++) {
 572                 uint32_t n = ntohl(index[i]);
 573                 if (n < nr) {
 574                         munmap(idx_map, idx_size);
 575                         return error("non-monotonic index %s", path);
 576                 }
 577                 nr = n;
 578         }
 579
 580         if (version == 1) {
 581                 /*
 582                  * Total size:
 583                  *  - 256 index entries 4 bytes each
 584                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 585                  *  - 20-byte SHA1 of the packfile
 586                  *  - 20-byte SHA1 file checksum
 587                  */
 588                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 589                         munmap(idx_map, idx_size);
 590                         return error("wrong index v1 file size in %s", path);
 591                 }
 592         } else if (version == 2) {
 593                 /*
 594                  * Minimum size:
 595                  *  - 8 bytes of header
 596                  *  - 256 index entries 4 bytes each
 597                  *  - 20-byte sha1 entry * nr
 598                  *  - 4-byte crc entry * nr
 599                  *  - 4-byte offset entry * nr
 600                  *  - 20-byte SHA1 of the packfile
 601                  *  - 20-byte SHA1 file checksum
 602                  * And after the 4-byte offset table might be a
 603                  * variable sized table containing 8-byte entries
 604                  * for offsets larger than 2^31.
 605                  */
 606                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 607                 unsigned long max_size = min_size;
 608                 if (nr)
 609                         max_size += (nr - 1)*8;
 610                 if (idx_size < min_size || idx_size > max_size) {
 611                         munmap(idx_map, idx_size);
 612                         return error("wrong index v2 file size in %s", path);
 613                 }
 614                 if (idx_size != min_size &&
 615                     /*
 616                      * make sure we can deal with large pack offsets.
 617                      * 31-bit signed offset won't be enough, neither
 618                      * 32-bit unsigned one will be.
 619                      */
 620                     (sizeof(off_t) <= 4)) {
 621                         munmap(idx_map, idx_size);
 622                         return error("pack too large for current definition of off_t in %s", path);
 623                 }
 624         }
 625
 626         p->index_version = version;
 627         p->index_data = idx_map;
 628         p->index_size = idx_size;
 629         p->num_objects = nr;
 630         return 0;
 631 }
 632
 633 int open_pack_index(struct packed_git *p)
 634 {
 635         char *idx_name;
 636         int ret;
 637
 638         if (p->index_data)
 639                 return 0;
 640
 641         idx_name = xstrdup(p->pack_name);
 642         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 643         ret = check_packed_git_idx(idx_name, p);
 644         free(idx_name);
 645         return ret;
 646 }
 647
 648 static void scan_windows(struct packed_git *p,
 649         struct packed_git **lru_p,
 650         struct pack_window **lru_w,
 651         struct pack_window **lru_l)
 652 {
 653         struct pack_window *w, *w_l;
 654
 655         for (w_l = NULL, w = p->windows; w; w = w->next) {
 656                 if (!w->inuse_cnt) {
 657                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 658                                 *lru_p = p;
 659                                 *lru_w = w;
 660                                 *lru_l = w_l;
 661                         }
 662                 }
 663                 w_l = w;
 664         }
 665 }
 666
 667 static int unuse_one_window(struct packed_git *current)
 668 {
 669         struct packed_git *p, *lru_p = NULL;
 670         struct pack_window *lru_w = NULL, *lru_l = NULL;
 671
 672         if (current)
 673                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 674         for (p = packed_git; p; p = p->next)
 675                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 676         if (lru_p) {
 677                 munmap(lru_w->base, lru_w->len);
 678                 pack_mapped -= lru_w->len;
 679                 if (lru_l)
 680                         lru_l->next = lru_w->next;
 681                 else
 682                         lru_p->windows = lru_w->next;
 683                 free(lru_w);
 684                 pack_open_windows--;
 685                 return 1;
 686         }
 687         return 0;
 688 }
 689
 690 void release_pack_memory(size_t need)
 691 {
 692         size_t cur = pack_mapped;
 693         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
 694                 ; /* nothing */
 695 }
 696
 697 static void mmap_limit_check(size_t length)
 698 {
 699         static size_t limit = 0;
 700         if (!limit) {
 701                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
 702                 if (!limit)
 703                         limit = SIZE_MAX;
 704         }
 705         if (length > limit)
 706                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
 707                     (uintmax_t)length, (uintmax_t)limit);
 708 }
 709
 710 void *xmmap(void *start, size_t length,
 711         int prot, int flags, int fd, off_t offset)
 712 {
 713         void *ret;
 714
 715         mmap_limit_check(length);
 716         ret = mmap(start, length, prot, flags, fd, offset);
 717         if (ret == MAP_FAILED) {
 718                 if (!length)
 719                         return NULL;
 720                 release_pack_memory(length);
 721                 ret = mmap(start, length, prot, flags, fd, offset);
 722                 if (ret == MAP_FAILED)
 723                         die_errno("Out of memory? mmap failed");
 724         }
 725         return ret;
 726 }
 727
 728 void close_pack_windows(struct packed_git *p)
 729 {
 730         while (p->windows) {
 731                 struct pack_window *w = p->windows;
 732
 733                 if (w->inuse_cnt)
 734                         die("pack '%s' still has open windows to it",
 735                             p->pack_name);
 736                 munmap(w->base, w->len);
 737                 pack_mapped -= w->len;
 738                 pack_open_windows--;
 739                 p->windows = w->next;
 740                 free(w);
 741         }
 742 }
 743
 744 /*
 745  * The LRU pack is the one with the oldest MRU window, preferring packs
 746  * with no used windows, or the oldest mtime if it has no windows allocated.
 747  */
 748 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 749 {
 750         struct pack_window *w, *this_mru_w;
 751         int has_windows_inuse = 0;
 752
 753         /*
 754          * Reject this pack if it has windows and the previously selected
 755          * one does not.  If this pack does not have windows, reject
 756          * it if the pack file is newer than the previously selected one.
 757          */
 758         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 759                 return;
 760
 761         for (w = this_mru_w = p->windows; w; w = w->next) {
 762                 /*
 763                  * Reject this pack if any of its windows are in use,
 764                  * but the previously selected pack did not have any
 765                  * inuse windows.  Otherwise, record that this pack
 766                  * has windows in use.
 767                  */
 768                 if (w->inuse_cnt) {
 769                         if (*accept_windows_inuse)
 770                                 has_windows_inuse = 1;
 771                         else
 772                                 return;
 773                 }
 774
 775                 if (w->last_used > this_mru_w->last_used)
 776                         this_mru_w = w;
 777
 778                 /*
 779                  * Reject this pack if it has windows that have been
 780                  * used more recently than the previously selected pack.
 781                  * If the previously selected pack had windows inuse and
 782                  * we have not encountered a window in this pack that is
 783                  * inuse, skip this check since we prefer a pack with no
 784                  * inuse windows to one that has inuse windows.
 785                  */
 786                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 787                     this_mru_w->last_used > (*mru_w)->last_used)
 788                         return;
 789         }
 790
 791         /*
 792          * Select this pack.
 793          */
 794         *mru_w = this_mru_w;
 795         *lru_p = p;
 796         *accept_windows_inuse = has_windows_inuse;
 797 }
 798
 799 static int close_one_pack(void)
 800 {
 801         struct packed_git *p, *lru_p = NULL;
 802         struct pack_window *mru_w = NULL;
 803         int accept_windows_inuse = 1;
 804
 805         for (p = packed_git; p; p = p->next) {
 806                 if (p->pack_fd == -1)
 807                         continue;
 808                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 809         }
 810
 811         if (lru_p) {
 812                 close(lru_p->pack_fd);
 813                 pack_open_fds--;
 814                 lru_p->pack_fd = -1;
 815                 return 1;
 816         }
 817
 818         return 0;
 819 }
 820
 821 void unuse_pack(struct pack_window **w_cursor)
 822 {
 823         struct pack_window *w = *w_cursor;
 824         if (w) {
 825                 w->inuse_cnt--;
 826                 *w_cursor = NULL;
 827         }
 828 }
 829
 830 void close_pack_index(struct packed_git *p)
 831 {
 832         if (p->index_data) {
 833                 munmap((void *)p->index_data, p->index_size);
 834                 p->index_data = NULL;
 835         }
 836 }
 837
 838 /*
 839  * This is used by git-repack in case a newly created pack happens to
 840  * contain the same set of objects as an existing one.  In that case
 841  * the resulting file might be different even if its name would be the
 842  * same.  It is best to close any reference to the old pack before it is
 843  * replaced on disk.  Of course no index pointers or windows for given pack
 844  * must subsist at this point.  If ever objects from this pack are requested
 845  * again, the new version of the pack will be reinitialized through
 846  * reprepare_packed_git().
 847  */
 848 void free_pack_by_name(const char *pack_name)
 849 {
 850         struct packed_git *p, **pp = &packed_git;
 851
 852         while (*pp) {
 853                 p = *pp;
 854                 if (strcmp(pack_name, p->pack_name) == 0) {
 855                         clear_delta_base_cache();
 856                         close_pack_windows(p);
 857                         if (p->pack_fd != -1) {
 858                                 close(p->pack_fd);
 859                                 pack_open_fds--;
 860                         }
 861                         close_pack_index(p);
 862                         free(p->bad_object_sha1);
 863                         *pp = p->next;
 864                         if (last_found_pack == p)
 865                                 last_found_pack = NULL;
 866                         free(p);
 867                         return;
 868                 }
 869                 pp = &p->next;
 870         }
 871 }
 872
 873 static unsigned int get_max_fd_limit(void)
 874 {
 875 #ifdef RLIMIT_NOFILE
 876         {
 877                 struct rlimit lim;
 878
 879                 if (!getrlimit(RLIMIT_NOFILE, &lim))
 880                         return lim.rlim_cur;
 881         }
 882 #endif
 883
 884 #ifdef _SC_OPEN_MAX
 885         {
 886                 long open_max = sysconf(_SC_OPEN_MAX);
 887                 if (0 < open_max)
 888                         return open_max;
 889                 /*
 890                  * Otherwise, we got -1 for one of the two
 891                  * reasons:
 892                  *
 893                  * (1) sysconf() did not understand _SC_OPEN_MAX
 894                  *     and signaled an error with -1; or
 895                  * (2) sysconf() said there is no limit.
 896                  *
 897                  * We _could_ clear errno before calling sysconf() to
 898                  * tell these two cases apart and return a huge number
 899                  * in the latter case to let the caller cap it to a
 900                  * value that is not so selfish, but letting the
 901                  * fallback OPEN_MAX codepath take care of these cases
 902                  * is a lot simpler.
 903                  */
 904         }
 905 #endif
 906
 907 #ifdef OPEN_MAX
 908         return OPEN_MAX;
 909 #else
 910         return 1; /* see the caller ;-) */
 911 #endif
 912 }
 913
 914 /*
 915  * Do not call this directly as this leaks p->pack_fd on error return;
 916  * call open_packed_git() instead.
 917  */
 918 static int open_packed_git_1(struct packed_git *p)
 919 {
 920         struct stat st;
 921         struct pack_header hdr;
 922         unsigned char sha1[20];
 923         unsigned char *idx_sha1;
 924         long fd_flag;
 925
 926         if (!p->index_data && open_pack_index(p))
 927                 return error("packfile %s index unavailable", p->pack_name);
 928
 929         if (!pack_max_fds) {
 930                 unsigned int max_fds = get_max_fd_limit();
 931
 932                 /* Save 3 for stdin/stdout/stderr, 22 for work */
 933                 if (25 < max_fds)
 934                         pack_max_fds = max_fds - 25;
 935                 else
 936                         pack_max_fds = 1;
 937         }
 938
 939         while (pack_max_fds <= pack_open_fds && close_one_pack())
 940                 ; /* nothing */
 941
 942         p->pack_fd = git_open_noatime(p->pack_name);
 943         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 944                 return -1;
 945         pack_open_fds++;
 946
 947         /* If we created the struct before we had the pack we lack size. */
 948         if (!p->pack_size) {
 949                 if (!S_ISREG(st.st_mode))
 950                         return error("packfile %s not a regular file", p->pack_name);
 951                 p->pack_size = st.st_size;
 952         } else if (p->pack_size != st.st_size)
 953                 return error("packfile %s size changed", p->pack_name);
 954
 955         /* We leave these file descriptors open with sliding mmap;
 956          * there is no point keeping them open across exec(), though.
 957          */
 958         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 959         if (fd_flag < 0)
 960                 return error("cannot determine file descriptor flags");
 961         fd_flag |= FD_CLOEXEC;
 962         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 963                 return error("cannot set FD_CLOEXEC");
 964
 965         /* Verify we recognize this pack file format. */
 966         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 967                 return error("file %s is far too short to be a packfile", p->pack_name);
 968         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 969                 return error("file %s is not a GIT packfile", p->pack_name);
 970         if (!pack_version_ok(hdr.hdr_version))
 971                 return error("packfile %s is version %"PRIu32" and not"
 972                         " supported (try upgrading GIT to a newer version)",
 973                         p->pack_name, ntohl(hdr.hdr_version));
 974
 975         /* Verify the pack matches its index. */
 976         if (p->num_objects != ntohl(hdr.hdr_entries))
 977                 return error("packfile %s claims to have %"PRIu32" objects"
 978                              " while index indicates %"PRIu32" objects",
 979                              p->pack_name, ntohl(hdr.hdr_entries),
 980                              p->num_objects);
 981         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
 982                 return error("end of packfile %s is unavailable", p->pack_name);
 983         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
 984                 return error("packfile %s signature is unavailable", p->pack_name);
 985         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
 986         if (hashcmp(sha1, idx_sha1))
 987                 return error("packfile %s does not match index", p->pack_name);
 988         return 0;
 989 }
 990
 991 static int open_packed_git(struct packed_git *p)
 992 {
 993         if (!open_packed_git_1(p))
 994                 return 0;
 995         if (p->pack_fd != -1) {
 996                 close(p->pack_fd);
 997                 pack_open_fds--;
 998                 p->pack_fd = -1;
 999         }
1000         return -1;
1001 }
1002
1003 static int in_window(struct pack_window *win, off_t offset)
1004 {
1005         /* We must promise at least 20 bytes (one hash) after the
1006          * offset is available from this window, otherwise the offset
1007          * is not actually in this window and a different window (which
1008          * has that one hash excess) must be used.  This is to support
1009          * the object header and delta base parsing routines below.
1010          */
1011         off_t win_off = win->offset;
1012         return win_off <= offset
1013                 && (offset + 20) <= (win_off + win->len);
1014 }
1015
1016 unsigned char *use_pack(struct packed_git *p,
1017                 struct pack_window **w_cursor,
1018                 off_t offset,
1019                 unsigned long *left)
1020 {
1021         struct pack_window *win = *w_cursor;
1022
1023         /* Since packfiles end in a hash of their content and it's
1024          * pointless to ask for an offset into the middle of that
1025          * hash, and the in_window function above wouldn't match
1026          * don't allow an offset too close to the end of the file.
1027          */
1028         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1029                 die("packfile %s cannot be accessed", p->pack_name);
1030         if (offset > (p->pack_size - 20))
1031                 die("offset beyond end of packfile (truncated pack?)");
1032
1033         if (!win || !in_window(win, offset)) {
1034                 if (win)
1035                         win->inuse_cnt--;
1036                 for (win = p->windows; win; win = win->next) {
1037                         if (in_window(win, offset))
1038                                 break;
1039                 }
1040                 if (!win) {
1041                         size_t window_align = packed_git_window_size / 2;
1042                         off_t len;
1043
1044                         if (p->pack_fd == -1 && open_packed_git(p))
1045                                 die("packfile %s cannot be accessed", p->pack_name);
1046
1047                         win = xcalloc(1, sizeof(*win));
1048                         win->offset = (offset / window_align) * window_align;
1049                         len = p->pack_size - win->offset;
1050                         if (len > packed_git_window_size)
1051                                 len = packed_git_window_size;
1052                         win->len = (size_t)len;
1053                         pack_mapped += win->len;
1054                         while (packed_git_limit < pack_mapped
1055                                 && unuse_one_window(p))
1056                                 ; /* nothing */
1057                         win->base = xmmap(NULL, win->len,
1058                                 PROT_READ, MAP_PRIVATE,
1059                                 p->pack_fd, win->offset);
1060                         if (win->base == MAP_FAILED)
1061                                 die("packfile %s cannot be mapped: %s",
1062                                         p->pack_name,
1063                                         strerror(errno));
1064                         if (!win->offset && win->len == p->pack_size
1065                                 && !p->do_not_close) {
1066                                 close(p->pack_fd);
1067                                 pack_open_fds--;
1068                                 p->pack_fd = -1;
1069                         }
1070                         pack_mmap_calls++;
1071                         pack_open_windows++;
1072                         if (pack_mapped > peak_pack_mapped)
1073                                 peak_pack_mapped = pack_mapped;
1074                         if (pack_open_windows > peak_pack_open_windows)
1075                                 peak_pack_open_windows = pack_open_windows;
1076                         win->next = p->windows;
1077                         p->windows = win;
1078                 }
1079         }
1080         if (win != *w_cursor) {
1081                 win->last_used = pack_used_ctr++;
1082                 win->inuse_cnt++;
1083                 *w_cursor = win;
1084         }
1085         offset -= win->offset;
1086         if (left)
1087                 *left = win->len - xsize_t(offset);
1088         return win->base + offset;
1089 }
1090
1091 static struct packed_git *alloc_packed_git(int extra)
1092 {
1093         struct packed_git *p = xmalloc(sizeof(*p) + extra);
1094         memset(p, 0, sizeof(*p));
1095         p->pack_fd = -1;
1096         return p;
1097 }
1098
1099 static void try_to_free_pack_memory(size_t size)
1100 {
1101         release_pack_memory(size);
1102 }
1103
1104 struct packed_git *add_packed_git(const char *path, int path_len, int local)
1105 {
1106         static int have_set_try_to_free_routine;
1107         struct stat st;
1108         struct packed_git *p = alloc_packed_git(path_len + 2);
1109
1110         if (!have_set_try_to_free_routine) {
1111                 have_set_try_to_free_routine = 1;
1112                 set_try_to_free_routine(try_to_free_pack_memory);
1113         }
1114
1115         /*
1116          * Make sure a corresponding .pack file exists and that
1117          * the index looks sane.
1118          */
1119         path_len -= strlen(".idx");
1120         if (path_len < 1) {
1121                 free(p);
1122                 return NULL;
1123         }
1124         memcpy(p->pack_name, path, path_len);
1125
1126         strcpy(p->pack_name + path_len, ".keep");
1127         if (!access(p->pack_name, F_OK))
1128                 p->pack_keep = 1;
1129
1130         strcpy(p->pack_name + path_len, ".pack");
1131         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1132                 free(p);
1133                 return NULL;
1134         }
1135
1136         /* ok, it looks sane as far as we can check without
1137          * actually mapping the pack file.
1138          */
1139         p->pack_size = st.st_size;
1140         p->pack_local = local;
1141         p->mtime = st.st_mtime;
1142         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1143                 hashclr(p->sha1);
1144         return p;
1145 }
1146
1147 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1148 {
1149         const char *path = sha1_pack_name(sha1);
1150         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
1151
1152         strcpy(p->pack_name, path);
1153         hashcpy(p->sha1, sha1);
1154         if (check_packed_git_idx(idx_path, p)) {
1155                 free(p);
1156                 return NULL;
1157         }
1158
1159         return p;
1160 }
1161
1162 void install_packed_git(struct packed_git *pack)
1163 {
1164         if (pack->pack_fd != -1)
1165                 pack_open_fds++;
1166
1167         pack->next = packed_git;
1168         packed_git = pack;
1169 }
1170
1171 void (*report_garbage)(const char *desc, const char *path);
1172
1173 static void report_helper(const struct string_list *list,
1174                           int seen_bits, int first, int last)
1175 {
1176         const char *msg;
1177         switch (seen_bits) {
1178         case 0:
1179                 msg = "no corresponding .idx or .pack";
1180                 break;
1181         case 1:
1182                 msg = "no corresponding .idx";
1183                 break;
1184         case 2:
1185                 msg = "no corresponding .pack";
1186                 break;
1187         default:
1188                 return;
1189         }
1190         for (; first < last; first++)
1191                 report_garbage(msg, list->items[first].string);
1192 }
1193
1194 static void report_pack_garbage(struct string_list *list)
1195 {
1196         int i, baselen = -1, first = 0, seen_bits = 0;
1197
1198         if (!report_garbage)
1199                 return;
1200
1201         sort_string_list(list);
1202
1203         for (i = 0; i < list->nr; i++) {
1204                 const char *path = list->items[i].string;
1205                 if (baselen != -1 &&
1206                     strncmp(path, list->items[first].string, baselen)) {
1207                         report_helper(list, seen_bits, first, i);
1208                         baselen = -1;
1209                         seen_bits = 0;
1210                 }
1211                 if (baselen == -1) {
1212                         const char *dot = strrchr(path, '.');
1213                         if (!dot) {
1214                                 report_garbage("garbage found", path);
1215                                 continue;
1216                         }
1217                         baselen = dot - path + 1;
1218                         first = i;
1219                 }
1220                 if (!strcmp(path + baselen, "pack"))
1221                         seen_bits |= 1;
1222                 else if (!strcmp(path + baselen, "idx"))
1223                         seen_bits |= 2;
1224         }
1225         report_helper(list, seen_bits, first, list->nr);
1226 }
1227
1228 static void prepare_packed_git_one(char *objdir, int local)
1229 {
1230         struct strbuf path = STRBUF_INIT;
1231         size_t dirnamelen;
1232         DIR *dir;
1233         struct dirent *de;
1234         struct string_list garbage = STRING_LIST_INIT_DUP;
1235
1236         strbuf_addstr(&path, objdir);
1237         strbuf_addstr(&path, "/pack");
1238         dir = opendir(path.buf);
1239         if (!dir) {
1240                 if (errno != ENOENT)
1241                         error("unable to open object pack directory: %s: %s",
1242                               path.buf, strerror(errno));
1243                 strbuf_release(&path);
1244                 return;
1245         }
1246         strbuf_addch(&path, '/');
1247         dirnamelen = path.len;
1248         while ((de = readdir(dir)) != NULL) {
1249                 struct packed_git *p;
1250                 size_t base_len;
1251
1252                 if (is_dot_or_dotdot(de->d_name))
1253                         continue;
1254
1255                 strbuf_setlen(&path, dirnamelen);
1256                 strbuf_addstr(&path, de->d_name);
1257
1258                 base_len = path.len;
1259                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1260                         /* Don't reopen a pack we already have. */
1261                         for (p = packed_git; p; p = p->next) {
1262                                 size_t len;
1263                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1264                                     len == base_len &&
1265                                     !memcmp(p->pack_name, path.buf, len))
1266                                         break;
1267                         }
1268                         if (p == NULL &&
1269                             /*
1270                              * See if it really is a valid .idx file with
1271                              * corresponding .pack file that we can map.
1272                              */
1273                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1274                                 install_packed_git(p);
1275                 }
1276
1277                 if (!report_garbage)
1278                         continue;
1279
1280                 if (ends_with(de->d_name, ".idx") ||
1281                     ends_with(de->d_name, ".pack") ||
1282                     ends_with(de->d_name, ".bitmap") ||
1283                     ends_with(de->d_name, ".keep"))
1284                         string_list_append(&garbage, path.buf);
1285                 else
1286                         report_garbage("garbage found", path.buf);
1287         }
1288         closedir(dir);
1289         report_pack_garbage(&garbage);
1290         string_list_clear(&garbage, 0);
1291         strbuf_release(&path);
1292 }
1293
1294 static int sort_pack(const void *a_, const void *b_)
1295 {
1296         struct packed_git *a = *((struct packed_git **)a_);
1297         struct packed_git *b = *((struct packed_git **)b_);
1298         int st;
1299
1300         /*
1301          * Local packs tend to contain objects specific to our
1302          * variant of the project than remote ones.  In addition,
1303          * remote ones could be on a network mounted filesystem.
1304          * Favor local ones for these reasons.
1305          */
1306         st = a->pack_local - b->pack_local;
1307         if (st)
1308                 return -st;
1309
1310         /*
1311          * Younger packs tend to contain more recent objects,
1312          * and more recent objects tend to get accessed more
1313          * often.
1314          */
1315         if (a->mtime < b->mtime)
1316                 return 1;
1317         else if (a->mtime == b->mtime)
1318                 return 0;
1319         return -1;
1320 }
1321
1322 static void rearrange_packed_git(void)
1323 {
1324         struct packed_git **ary, *p;
1325         int i, n;
1326
1327         for (n = 0, p = packed_git; p; p = p->next)
1328                 n++;
1329         if (n < 2)
1330                 return;
1331
1332         /* prepare an array of packed_git for easier sorting */
1333         ary = xcalloc(n, sizeof(struct packed_git *));
1334         for (n = 0, p = packed_git; p; p = p->next)
1335                 ary[n++] = p;
1336
1337         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1338
1339         /* link them back again */
1340         for (i = 0; i < n - 1; i++)
1341                 ary[i]->next = ary[i + 1];
1342         ary[n - 1]->next = NULL;
1343         packed_git = ary[0];
1344
1345         free(ary);
1346 }
1347
1348 static int prepare_packed_git_run_once = 0;
1349 void prepare_packed_git(void)
1350 {
1351         struct alternate_object_database *alt;
1352
1353         if (prepare_packed_git_run_once)
1354                 return;
1355         prepare_packed_git_one(get_object_directory(), 1);
1356         prepare_alt_odb();
1357         for (alt = alt_odb_list; alt; alt = alt->next) {
1358                 alt->name[-1] = 0;
1359                 prepare_packed_git_one(alt->base, 0);
1360                 alt->name[-1] = '/';
1361         }
1362         rearrange_packed_git();
1363         prepare_packed_git_run_once = 1;
1364 }
1365
1366 void reprepare_packed_git(void)
1367 {
1368         prepare_packed_git_run_once = 0;
1369         prepare_packed_git();
1370 }
1371
1372 static void mark_bad_packed_object(struct packed_git *p,
1373                                    const unsigned char *sha1)
1374 {
1375         unsigned i;
1376         for (i = 0; i < p->num_bad_objects; i++)
1377                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1378                         return;
1379         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1380         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1381         p->num_bad_objects++;
1382 }
1383
1384 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1385 {
1386         struct packed_git *p;
1387         unsigned i;
1388
1389         for (p = packed_git; p; p = p->next)
1390                 for (i = 0; i < p->num_bad_objects; i++)
1391                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1392                                 return p;
1393         return NULL;
1394 }
1395
1396 /*
1397  * With an in-core object data in "map", rehash it to make sure the
1398  * object name actually matches "sha1" to detect object corruption.
1399  * With "map" == NULL, try reading the object named with "sha1" using
1400  * the streaming interface and rehash it to do the same.
1401  */
1402 int check_sha1_signature(const unsigned char *sha1, void *map,
1403                          unsigned long size, const char *type)
1404 {
1405         unsigned char real_sha1[20];
1406         enum object_type obj_type;
1407         struct git_istream *st;
1408         git_SHA_CTX c;
1409         char hdr[32];
1410         int hdrlen;
1411
1412         if (map) {
1413                 hash_sha1_file(map, size, type, real_sha1);
1414                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1415         }
1416
1417         st = open_istream(sha1, &obj_type, &size, NULL);
1418         if (!st)
1419                 return -1;
1420
1421         /* Generate the header */
1422         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1423
1424         /* Sha1.. */
1425         git_SHA1_Init(&c);
1426         git_SHA1_Update(&c, hdr, hdrlen);
1427         for (;;) {
1428                 char buf[1024 * 16];
1429                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1430
1431                 if (readlen < 0) {
1432                         close_istream(st);
1433                         return -1;
1434                 }
1435                 if (!readlen)
1436                         break;
1437                 git_SHA1_Update(&c, buf, readlen);
1438         }
1439         git_SHA1_Final(real_sha1, &c);
1440         close_istream(st);
1441         return hashcmp(sha1, real_sha1) ? -1 : 0;
1442 }
1443
1444 int git_open_noatime(const char *name)
1445 {
1446         static int sha1_file_open_flag = O_NOATIME;
1447
1448         for (;;) {
1449                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1450                 if (fd >= 0)
1451                         return fd;
1452
1453                 /* Might the failure be due to O_NOATIME? */
1454                 if (errno != ENOENT && sha1_file_open_flag) {
1455                         sha1_file_open_flag = 0;
1456                         continue;
1457                 }
1458
1459                 return -1;
1460         }
1461 }
1462
1463 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1464 {
1465         struct alternate_object_database *alt;
1466
1467         if (!lstat(sha1_file_name(sha1), st))
1468                 return 0;
1469
1470         prepare_alt_odb();
1471         errno = ENOENT;
1472         for (alt = alt_odb_list; alt; alt = alt->next) {
1473                 fill_sha1_path(alt->name, sha1);
1474                 if (!lstat(alt->base, st))
1475                         return 0;
1476         }
1477
1478         return -1;
1479 }
1480
1481 static int open_sha1_file(const unsigned char *sha1)
1482 {
1483         int fd;
1484         struct alternate_object_database *alt;
1485         int most_interesting_errno;
1486
1487         fd = git_open_noatime(sha1_file_name(sha1));
1488         if (fd >= 0)
1489                 return fd;
1490         most_interesting_errno = errno;
1491
1492         prepare_alt_odb();
1493         for (alt = alt_odb_list; alt; alt = alt->next) {
1494                 fill_sha1_path(alt->name, sha1);
1495                 fd = git_open_noatime(alt->base);
1496                 if (fd >= 0)
1497                         return fd;
1498                 if (most_interesting_errno == ENOENT)
1499                         most_interesting_errno = errno;
1500         }
1501         errno = most_interesting_errno;
1502         return -1;
1503 }
1504
1505 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1506 {
1507         void *map;
1508         int fd;
1509
1510         fd = open_sha1_file(sha1);
1511         map = NULL;
1512         if (fd >= 0) {
1513                 struct stat st;
1514
1515                 if (!fstat(fd, &st)) {
1516                         *size = xsize_t(st.st_size);
1517                         if (!*size) {
1518                                 /* mmap() is forbidden on empty files */
1519                                 error("object file %s is empty", sha1_file_name(sha1));
1520                                 return NULL;
1521                         }
1522                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1523                 }
1524                 close(fd);
1525         }
1526         return map;
1527 }
1528
1529 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1530                 unsigned long len, enum object_type *type, unsigned long *sizep)
1531 {
1532         unsigned shift;
1533         unsigned long size, c;
1534         unsigned long used = 0;
1535
1536         c = buf[used++];
1537         *type = (c >> 4) & 7;
1538         size = c & 15;
1539         shift = 4;
1540         while (c & 0x80) {
1541                 if (len <= used || bitsizeof(long) <= shift) {
1542                         error("bad object header");
1543                         size = used = 0;
1544                         break;
1545                 }
1546                 c = buf[used++];
1547                 size += (c & 0x7f) << shift;
1548                 shift += 7;
1549         }
1550         *sizep = size;
1551         return used;
1552 }
1553
1554 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1555 {
1556         /* Get the data stream */
1557         memset(stream, 0, sizeof(*stream));
1558         stream->next_in = map;
1559         stream->avail_in = mapsize;
1560         stream->next_out = buffer;
1561         stream->avail_out = bufsiz;
1562
1563         git_inflate_init(stream);
1564         return git_inflate(stream, 0);
1565 }
1566
1567 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1568 {
1569         int bytes = strlen(buffer) + 1;
1570         unsigned char *buf = xmallocz(size);
1571         unsigned long n;
1572         int status = Z_OK;
1573
1574         n = stream->total_out - bytes;
1575         if (n > size)
1576                 n = size;
1577         memcpy(buf, (char *) buffer + bytes, n);
1578         bytes = n;
1579         if (bytes <= size) {
1580                 /*
1581                  * The above condition must be (bytes <= size), not
1582                  * (bytes < size).  In other words, even though we
1583                  * expect no more output and set avail_out to zero,
1584                  * the input zlib stream may have bytes that express
1585                  * "this concludes the stream", and we *do* want to
1586                  * eat that input.
1587                  *
1588                  * Otherwise we would not be able to test that we
1589                  * consumed all the input to reach the expected size;
1590                  * we also want to check that zlib tells us that all
1591                  * went well with status == Z_STREAM_END at the end.
1592                  */
1593                 stream->next_out = buf + bytes;
1594                 stream->avail_out = size - bytes;
1595                 while (status == Z_OK)
1596                         status = git_inflate(stream, Z_FINISH);
1597         }
1598         if (status == Z_STREAM_END && !stream->avail_in) {
1599                 git_inflate_end(stream);
1600                 return buf;
1601         }
1602
1603         if (status < 0)
1604                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1605         else if (stream->avail_in)
1606                 error("garbage at end of loose object '%s'",
1607                       sha1_to_hex(sha1));
1608         free(buf);
1609         return NULL;
1610 }
1611
1612 /*
1613  * We used to just use "sscanf()", but that's actually way
1614  * too permissive for what we want to check. So do an anal
1615  * object header parse by hand.
1616  */
1617 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1618 {
1619         char type[10];
1620         int i;
1621         unsigned long size;
1622
1623         /*
1624          * The type can be at most ten bytes (including the
1625          * terminating '\0' that we add), and is followed by
1626          * a space.
1627          */
1628         i = 0;
1629         for (;;) {
1630                 char c = *hdr++;
1631                 if (c == ' ')
1632                         break;
1633                 type[i++] = c;
1634                 if (i >= sizeof(type))
1635                         return -1;
1636         }
1637         type[i] = 0;
1638
1639         /*
1640          * The length must follow immediately, and be in canonical
1641          * decimal format (ie "010" is not valid).
1642          */
1643         size = *hdr++ - '0';
1644         if (size > 9)
1645                 return -1;
1646         if (size) {
1647                 for (;;) {
1648                         unsigned long c = *hdr - '0';
1649                         if (c > 9)
1650                                 break;
1651                         hdr++;
1652                         size = size * 10 + c;
1653                 }
1654         }
1655         *sizep = size;
1656
1657         /*
1658          * The length must be followed by a zero byte
1659          */
1660         return *hdr ? -1 : type_from_string(type);
1661 }
1662
1663 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1664 {
1665         int ret;
1666         git_zstream stream;
1667         char hdr[8192];
1668
1669         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1670         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1671                 return NULL;
1672
1673         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1674 }
1675
1676 unsigned long get_size_from_delta(struct packed_git *p,
1677                                   struct pack_window **w_curs,
1678                                   off_t curpos)
1679 {
1680         const unsigned char *data;
1681         unsigned char delta_head[20], *in;
1682         git_zstream stream;
1683         int st;
1684
1685         memset(&stream, 0, sizeof(stream));
1686         stream.next_out = delta_head;
1687         stream.avail_out = sizeof(delta_head);
1688
1689         git_inflate_init(&stream);
1690         do {
1691                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1692                 stream.next_in = in;
1693                 st = git_inflate(&stream, Z_FINISH);
1694                 curpos += stream.next_in - in;
1695         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1696                  stream.total_out < sizeof(delta_head));
1697         git_inflate_end(&stream);
1698         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1699                 error("delta data unpack-initial failed");
1700                 return 0;
1701         }
1702
1703         /* Examine the initial part of the delta to figure out
1704          * the result size.
1705          */
1706         data = delta_head;
1707
1708         /* ignore base size */
1709         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1710
1711         /* Read the result size */
1712         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1713 }
1714
1715 static off_t get_delta_base(struct packed_git *p,
1716                                     struct pack_window **w_curs,
1717                                     off_t *curpos,
1718                                     enum object_type type,
1719                                     off_t delta_obj_offset)
1720 {
1721         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1722         off_t base_offset;
1723
1724         /* use_pack() assured us we have [base_info, base_info + 20)
1725          * as a range that we can look at without walking off the
1726          * end of the mapped window.  Its actually the hash size
1727          * that is assured.  An OFS_DELTA longer than the hash size
1728          * is stupid, as then a REF_DELTA would be smaller to store.
1729          */
1730         if (type == OBJ_OFS_DELTA) {
1731                 unsigned used = 0;
1732                 unsigned char c = base_info[used++];
1733                 base_offset = c & 127;
1734                 while (c & 128) {
1735                         base_offset += 1;
1736                         if (!base_offset || MSB(base_offset, 7))
1737                                 return 0;  /* overflow */
1738                         c = base_info[used++];
1739                         base_offset = (base_offset << 7) + (c & 127);
1740                 }
1741                 base_offset = delta_obj_offset - base_offset;
1742                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1743                         return 0;  /* out of bound */
1744                 *curpos += used;
1745         } else if (type == OBJ_REF_DELTA) {
1746                 /* The base entry _must_ be in the same pack */
1747                 base_offset = find_pack_entry_one(base_info, p);
1748                 *curpos += 20;
1749         } else
1750                 die("I am totally screwed");
1751         return base_offset;
1752 }
1753
1754 /*
1755  * Like get_delta_base above, but we return the sha1 instead of the pack
1756  * offset. This means it is cheaper for REF deltas (we do not have to do
1757  * the final object lookup), but more expensive for OFS deltas (we
1758  * have to load the revidx to convert the offset back into a sha1).
1759  */
1760 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1761                                                 struct pack_window **w_curs,
1762                                                 off_t curpos,
1763                                                 enum object_type type,
1764                                                 off_t delta_obj_offset)
1765 {
1766         if (type == OBJ_REF_DELTA) {
1767                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1768                 return base;
1769         } else if (type == OBJ_OFS_DELTA) {
1770                 struct revindex_entry *revidx;
1771                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1772                                                    type, delta_obj_offset);
1773
1774                 if (!base_offset)
1775                         return NULL;
1776
1777                 revidx = find_pack_revindex(p, base_offset);
1778                 if (!revidx)
1779                         return NULL;
1780
1781                 return nth_packed_object_sha1(p, revidx->nr);
1782         } else
1783                 return NULL;
1784 }
1785
1786 int unpack_object_header(struct packed_git *p,
1787                          struct pack_window **w_curs,
1788                          off_t *curpos,
1789                          unsigned long *sizep)
1790 {
1791         unsigned char *base;
1792         unsigned long left;
1793         unsigned long used;
1794         enum object_type type;
1795
1796         /* use_pack() assures us we have [base, base + 20) available
1797          * as a range that we can look at.  (Its actually the hash
1798          * size that is assured.)  With our object header encoding
1799          * the maximum deflated object size is 2^137, which is just
1800          * insane, so we know won't exceed what we have been given.
1801          */
1802         base = use_pack(p, w_curs, *curpos, &left);
1803         used = unpack_object_header_buffer(base, left, &type, sizep);
1804         if (!used) {
1805                 type = OBJ_BAD;
1806         } else
1807                 *curpos += used;
1808
1809         return type;
1810 }
1811
1812 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1813 {
1814         int type;
1815         struct revindex_entry *revidx;
1816         const unsigned char *sha1;
1817         revidx = find_pack_revindex(p, obj_offset);
1818         if (!revidx)
1819                 return OBJ_BAD;
1820         sha1 = nth_packed_object_sha1(p, revidx->nr);
1821         mark_bad_packed_object(p, sha1);
1822         type = sha1_object_info(sha1, NULL);
1823         if (type <= OBJ_NONE)
1824                 return OBJ_BAD;
1825         return type;
1826 }
1827
1828 #define POI_STACK_PREALLOC 64
1829
1830 static enum object_type packed_to_object_type(struct packed_git *p,
1831                                               off_t obj_offset,
1832                                               enum object_type type,
1833                                               struct pack_window **w_curs,
1834                                               off_t curpos)
1835 {
1836         off_t small_poi_stack[POI_STACK_PREALLOC];
1837         off_t *poi_stack = small_poi_stack;
1838         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1839
1840         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1841                 off_t base_offset;
1842                 unsigned long size;
1843                 /* Push the object we're going to leave behind */
1844                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1845                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1846                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1847                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1848                 } else {
1849                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1850                 }
1851                 poi_stack[poi_stack_nr++] = obj_offset;
1852                 /* If parsing the base offset fails, just unwind */
1853                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1854                 if (!base_offset)
1855                         goto unwind;
1856                 curpos = obj_offset = base_offset;
1857                 type = unpack_object_header(p, w_curs, &curpos, &size);
1858                 if (type <= OBJ_NONE) {
1859                         /* If getting the base itself fails, we first
1860                          * retry the base, otherwise unwind */
1861                         type = retry_bad_packed_offset(p, base_offset);
1862                         if (type > OBJ_NONE)
1863                                 goto out;
1864                         goto unwind;
1865                 }
1866         }
1867
1868         switch (type) {
1869         case OBJ_BAD:
1870         case OBJ_COMMIT:
1871         case OBJ_TREE:
1872         case OBJ_BLOB:
1873         case OBJ_TAG:
1874                 break;
1875         default:
1876                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1877                       type, (uintmax_t)obj_offset, p->pack_name);
1878                 type = OBJ_BAD;
1879         }
1880
1881 out:
1882         if (poi_stack != small_poi_stack)
1883                 free(poi_stack);
1884         return type;
1885
1886 unwind:
1887         while (poi_stack_nr) {
1888                 obj_offset = poi_stack[--poi_stack_nr];
1889                 type = retry_bad_packed_offset(p, obj_offset);
1890                 if (type > OBJ_NONE)
1891                         goto out;
1892         }
1893         type = OBJ_BAD;
1894         goto out;
1895 }
1896
1897 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1898                               struct object_info *oi)
1899 {
1900         struct pack_window *w_curs = NULL;
1901         unsigned long size;
1902         off_t curpos = obj_offset;
1903         enum object_type type;
1904
1905         /*
1906          * We always get the representation type, but only convert it to
1907          * a "real" type later if the caller is interested.
1908          */
1909         type = unpack_object_header(p, &w_curs, &curpos, &size);
1910
1911         if (oi->sizep) {
1912                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1913                         off_t tmp_pos = curpos;
1914                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1915                                                            type, obj_offset);
1916                         if (!base_offset) {
1917                                 type = OBJ_BAD;
1918                                 goto out;
1919                         }
1920                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1921                         if (*oi->sizep == 0) {
1922                                 type = OBJ_BAD;
1923                                 goto out;
1924                         }
1925                 } else {
1926                         *oi->sizep = size;
1927                 }
1928         }
1929
1930         if (oi->disk_sizep) {
1931                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1932                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1933         }
1934
1935         if (oi->typep) {
1936                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
1937                 if (*oi->typep < 0) {
1938                         type = OBJ_BAD;
1939                         goto out;
1940                 }
1941         }
1942
1943         if (oi->delta_base_sha1) {
1944                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1945                         const unsigned char *base;
1946
1947                         base = get_delta_base_sha1(p, &w_curs, curpos,
1948                                                    type, obj_offset);
1949                         if (!base) {
1950                                 type = OBJ_BAD;
1951                                 goto out;
1952                         }
1953
1954                         hashcpy(oi->delta_base_sha1, base);
1955                 } else
1956                         hashclr(oi->delta_base_sha1);
1957         }
1958
1959 out:
1960         unuse_pack(&w_curs);
1961         return type;
1962 }
1963
1964 static void *unpack_compressed_entry(struct packed_git *p,
1965                                     struct pack_window **w_curs,
1966                                     off_t curpos,
1967                                     unsigned long size)
1968 {
1969         int st;
1970         git_zstream stream;
1971         unsigned char *buffer, *in;
1972
1973         buffer = xmallocz_gently(size);
1974         if (!buffer)
1975                 return NULL;
1976         memset(&stream, 0, sizeof(stream));
1977         stream.next_out = buffer;
1978         stream.avail_out = size + 1;
1979
1980         git_inflate_init(&stream);
1981         do {
1982                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1983                 stream.next_in = in;
1984                 st = git_inflate(&stream, Z_FINISH);
1985                 if (!stream.avail_out)
1986                         break; /* the payload is larger than it should be */
1987                 curpos += stream.next_in - in;
1988         } while (st == Z_OK || st == Z_BUF_ERROR);
1989         git_inflate_end(&stream);
1990         if ((st != Z_STREAM_END) || stream.total_out != size) {
1991                 free(buffer);
1992                 return NULL;
1993         }
1994
1995         return buffer;
1996 }
1997
1998 #define MAX_DELTA_CACHE (256)
1999
2000 static size_t delta_base_cached;
2001
2002 static struct delta_base_cache_lru_list {
2003         struct delta_base_cache_lru_list *prev;
2004         struct delta_base_cache_lru_list *next;
2005 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
2006
2007 static struct delta_base_cache_entry {
2008         struct delta_base_cache_lru_list lru;
2009         void *data;
2010         struct packed_git *p;
2011         off_t base_offset;
2012         unsigned long size;
2013         enum object_type type;
2014 } delta_base_cache[MAX_DELTA_CACHE];
2015
2016 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
2017 {
2018         unsigned long hash;
2019
2020         hash = (unsigned long)p + (unsigned long)base_offset;
2021         hash += (hash >> 8) + (hash >> 16);
2022         return hash % MAX_DELTA_CACHE;
2023 }
2024
2025 static struct delta_base_cache_entry *
2026 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2027 {
2028         unsigned long hash = pack_entry_hash(p, base_offset);
2029         return delta_base_cache + hash;
2030 }
2031
2032 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
2033                                      struct packed_git *p, off_t base_offset)
2034 {
2035         return (ent->data && ent->p == p && ent->base_offset == base_offset);
2036 }
2037
2038 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2039 {
2040         struct delta_base_cache_entry *ent;
2041         ent = get_delta_base_cache_entry(p, base_offset);
2042         return eq_delta_base_cache_entry(ent, p, base_offset);
2043 }
2044
2045 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2046 {
2047         ent->data = NULL;
2048         ent->lru.next->prev = ent->lru.prev;
2049         ent->lru.prev->next = ent->lru.next;
2050         delta_base_cached -= ent->size;
2051 }
2052
2053 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2054         unsigned long *base_size, enum object_type *type, int keep_cache)
2055 {
2056         struct delta_base_cache_entry *ent;
2057         void *ret;
2058
2059         ent = get_delta_base_cache_entry(p, base_offset);
2060
2061         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2062                 return unpack_entry(p, base_offset, type, base_size);
2063
2064         ret = ent->data;
2065
2066         if (!keep_cache)
2067                 clear_delta_base_cache_entry(ent);
2068         else
2069                 ret = xmemdupz(ent->data, ent->size);
2070         *type = ent->type;
2071         *base_size = ent->size;
2072         return ret;
2073 }
2074
2075 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2076 {
2077         if (ent->data) {
2078                 free(ent->data);
2079                 ent->data = NULL;
2080                 ent->lru.next->prev = ent->lru.prev;
2081                 ent->lru.prev->next = ent->lru.next;
2082                 delta_base_cached -= ent->size;
2083         }
2084 }
2085
2086 void clear_delta_base_cache(void)
2087 {
2088         unsigned long p;
2089         for (p = 0; p < MAX_DELTA_CACHE; p++)
2090                 release_delta_base_cache(&delta_base_cache[p]);
2091 }
2092
2093 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2094         void *base, unsigned long base_size, enum object_type type)
2095 {
2096         unsigned long hash = pack_entry_hash(p, base_offset);
2097         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2098         struct delta_base_cache_lru_list *lru;
2099
2100         release_delta_base_cache(ent);
2101         delta_base_cached += base_size;
2102
2103         for (lru = delta_base_cache_lru.next;
2104              delta_base_cached > delta_base_cache_limit
2105              && lru != &delta_base_cache_lru;
2106              lru = lru->next) {
2107                 struct delta_base_cache_entry *f = (void *)lru;
2108                 if (f->type == OBJ_BLOB)
2109                         release_delta_base_cache(f);
2110         }
2111         for (lru = delta_base_cache_lru.next;
2112              delta_base_cached > delta_base_cache_limit
2113              && lru != &delta_base_cache_lru;
2114              lru = lru->next) {
2115                 struct delta_base_cache_entry *f = (void *)lru;
2116                 release_delta_base_cache(f);
2117         }
2118
2119         ent->p = p;
2120         ent->base_offset = base_offset;
2121         ent->type = type;
2122         ent->data = base;
2123         ent->size = base_size;
2124         ent->lru.next = &delta_base_cache_lru;
2125         ent->lru.prev = delta_base_cache_lru.prev;
2126         delta_base_cache_lru.prev->next = &ent->lru;
2127         delta_base_cache_lru.prev = &ent->lru;
2128 }
2129
2130 static void *read_object(const unsigned char *sha1, enum object_type *type,
2131                          unsigned long *size);
2132
2133 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2134 {
2135         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2136         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2137                          p->pack_name, (uintmax_t)obj_offset);
2138 }
2139
2140 int do_check_packed_object_crc;
2141
2142 #define UNPACK_ENTRY_STACK_PREALLOC 64
2143 struct unpack_entry_stack_ent {
2144         off_t obj_offset;
2145         off_t curpos;
2146         unsigned long size;
2147 };
2148
2149 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2150                    enum object_type *final_type, unsigned long *final_size)
2151 {
2152         struct pack_window *w_curs = NULL;
2153         off_t curpos = obj_offset;
2154         void *data = NULL;
2155         unsigned long size;
2156         enum object_type type;
2157         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2158         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2159         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2160         int base_from_cache = 0;
2161
2162         write_pack_access_log(p, obj_offset);
2163
2164         /* PHASE 1: drill down to the innermost base object */
2165         for (;;) {
2166                 off_t base_offset;
2167                 int i;
2168                 struct delta_base_cache_entry *ent;
2169
2170                 ent = get_delta_base_cache_entry(p, curpos);
2171                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2172                         type = ent->type;
2173                         data = ent->data;
2174                         size = ent->size;
2175                         clear_delta_base_cache_entry(ent);
2176                         base_from_cache = 1;
2177                         break;
2178                 }
2179
2180                 if (do_check_packed_object_crc && p->index_version > 1) {
2181                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2182                         unsigned long len = revidx[1].offset - obj_offset;
2183                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2184                                 const unsigned char *sha1 =
2185                                         nth_packed_object_sha1(p, revidx->nr);
2186                                 error("bad packed object CRC for %s",
2187                                       sha1_to_hex(sha1));
2188                                 mark_bad_packed_object(p, sha1);
2189                                 unuse_pack(&w_curs);
2190                                 return NULL;
2191                         }
2192                 }
2193
2194                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2195                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2196                         break;
2197
2198                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2199                 if (!base_offset) {
2200                         error("failed to validate delta base reference "
2201                               "at offset %"PRIuMAX" from %s",
2202                               (uintmax_t)curpos, p->pack_name);
2203                         /* bail to phase 2, in hopes of recovery */
2204                         data = NULL;
2205                         break;
2206                 }
2207
2208                 /* push object, proceed to base */
2209                 if (delta_stack_nr >= delta_stack_alloc
2210                     && delta_stack == small_delta_stack) {
2211                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2212                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2213                         memcpy(delta_stack, small_delta_stack,
2214                                sizeof(*delta_stack)*delta_stack_nr);
2215                 } else {
2216                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2217                 }
2218                 i = delta_stack_nr++;
2219                 delta_stack[i].obj_offset = obj_offset;
2220                 delta_stack[i].curpos = curpos;
2221                 delta_stack[i].size = size;
2222
2223                 curpos = obj_offset = base_offset;
2224         }
2225
2226         /* PHASE 2: handle the base */
2227         switch (type) {
2228         case OBJ_OFS_DELTA:
2229         case OBJ_REF_DELTA:
2230                 if (data)
2231                         die("BUG in unpack_entry: left loop at a valid delta");
2232                 break;
2233         case OBJ_COMMIT:
2234         case OBJ_TREE:
2235         case OBJ_BLOB:
2236         case OBJ_TAG:
2237                 if (!base_from_cache)
2238                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2239                 break;
2240         default:
2241                 data = NULL;
2242                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2243                       type, (uintmax_t)obj_offset, p->pack_name);
2244         }
2245
2246         /* PHASE 3: apply deltas in order */
2247
2248         /* invariants:
2249          *   'data' holds the base data, or NULL if there was corruption
2250          */
2251         while (delta_stack_nr) {
2252                 void *delta_data;
2253                 void *base = data;
2254                 unsigned long delta_size, base_size = size;
2255                 int i;
2256
2257                 data = NULL;
2258
2259                 if (base)
2260                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2261
2262                 if (!base) {
2263                         /*
2264                          * We're probably in deep shit, but let's try to fetch
2265                          * the required base anyway from another pack or loose.
2266                          * This is costly but should happen only in the presence
2267                          * of a corrupted pack, and is better than failing outright.
2268                          */
2269                         struct revindex_entry *revidx;
2270                         const unsigned char *base_sha1;
2271                         revidx = find_pack_revindex(p, obj_offset);
2272                         if (revidx) {
2273                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2274                                 error("failed to read delta base object %s"
2275                                       " at offset %"PRIuMAX" from %s",
2276                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2277                                       p->pack_name);
2278                                 mark_bad_packed_object(p, base_sha1);
2279                                 base = read_object(base_sha1, &type, &base_size);
2280                         }
2281                 }
2282
2283                 i = --delta_stack_nr;
2284                 obj_offset = delta_stack[i].obj_offset;
2285                 curpos = delta_stack[i].curpos;
2286                 delta_size = delta_stack[i].size;
2287
2288                 if (!base)
2289                         continue;
2290
2291                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2292
2293                 if (!delta_data) {
2294                         error("failed to unpack compressed delta "
2295                               "at offset %"PRIuMAX" from %s",
2296                               (uintmax_t)curpos, p->pack_name);
2297                         data = NULL;
2298                         continue;
2299                 }
2300
2301                 data = patch_delta(base, base_size,
2302                                    delta_data, delta_size,
2303                                    &size);
2304
2305                 /*
2306                  * We could not apply the delta; warn the user, but keep going.
2307                  * Our failure will be noticed either in the next iteration of
2308                  * the loop, or if this is the final delta, in the caller when
2309                  * we return NULL. Those code paths will take care of making
2310                  * a more explicit warning and retrying with another copy of
2311                  * the object.
2312                  */
2313                 if (!data)
2314                         error("failed to apply delta");
2315
2316                 free(delta_data);
2317         }
2318
2319         *final_type = type;
2320         *final_size = size;
2321
2322         unuse_pack(&w_curs);
2323
2324         if (delta_stack != small_delta_stack)
2325                 free(delta_stack);
2326
2327         return data;
2328 }
2329
2330 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2331                                             uint32_t n)
2332 {
2333         const unsigned char *index = p->index_data;
2334         if (!index) {
2335                 if (open_pack_index(p))
2336                         return NULL;
2337                 index = p->index_data;
2338         }
2339         if (n >= p->num_objects)
2340                 return NULL;
2341         index += 4 * 256;
2342         if (p->index_version == 1) {
2343                 return index + 24 * n + 4;
2344         } else {
2345                 index += 8;
2346                 return index + 20 * n;
2347         }
2348 }
2349
2350 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2351 {
2352         const unsigned char *index = p->index_data;
2353         index += 4 * 256;
2354         if (p->index_version == 1) {
2355                 return ntohl(*((uint32_t *)(index + 24 * n)));
2356         } else {
2357                 uint32_t off;
2358                 index += 8 + p->num_objects * (20 + 4);
2359                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2360                 if (!(off & 0x80000000))
2361                         return off;
2362                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2363                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2364                                    ntohl(*((uint32_t *)(index + 4)));
2365         }
2366 }
2367
2368 off_t find_pack_entry_one(const unsigned char *sha1,
2369                                   struct packed_git *p)
2370 {
2371         const uint32_t *level1_ofs = p->index_data;
2372         const unsigned char *index = p->index_data;
2373         unsigned hi, lo, stride;
2374         static int use_lookup = -1;
2375         static int debug_lookup = -1;
2376
2377         if (debug_lookup < 0)
2378                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2379
2380         if (!index) {
2381                 if (open_pack_index(p))
2382                         return 0;
2383                 level1_ofs = p->index_data;
2384                 index = p->index_data;
2385         }
2386         if (p->index_version > 1) {
2387                 level1_ofs += 2;
2388                 index += 8;
2389         }
2390         index += 4 * 256;
2391         hi = ntohl(level1_ofs[*sha1]);
2392         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2393         if (p->index_version > 1) {
2394                 stride = 20;
2395         } else {
2396                 stride = 24;
2397                 index += 4;
2398         }
2399
2400         if (debug_lookup)
2401                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2402                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2403
2404         if (use_lookup < 0)
2405                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2406         if (use_lookup) {
2407                 int pos = sha1_entry_pos(index, stride, 0,
2408                                          lo, hi, p->num_objects, sha1);
2409                 if (pos < 0)
2410                         return 0;
2411                 return nth_packed_object_offset(p, pos);
2412         }
2413
2414         do {
2415                 unsigned mi = (lo + hi) / 2;
2416                 int cmp = hashcmp(index + mi * stride, sha1);
2417
2418                 if (debug_lookup)
2419                         printf("lo %u hi %u rg %u mi %u\n",
2420                                lo, hi, hi - lo, mi);
2421                 if (!cmp)
2422                         return nth_packed_object_offset(p, mi);
2423                 if (cmp > 0)
2424                         hi = mi;
2425                 else
2426                         lo = mi+1;
2427         } while (lo < hi);
2428         return 0;
2429 }
2430
2431 int is_pack_valid(struct packed_git *p)
2432 {
2433         /* An already open pack is known to be valid. */
2434         if (p->pack_fd != -1)
2435                 return 1;
2436
2437         /* If the pack has one window completely covering the
2438          * file size, the pack is known to be valid even if
2439          * the descriptor is not currently open.
2440          */
2441         if (p->windows) {
2442                 struct pack_window *w = p->windows;
2443
2444                 if (!w->offset && w->len == p->pack_size)
2445                         return 1;
2446         }
2447
2448         /* Force the pack to open to prove its valid. */
2449         return !open_packed_git(p);
2450 }
2451
2452 static int fill_pack_entry(const unsigned char *sha1,
2453                            struct pack_entry *e,
2454                            struct packed_git *p)
2455 {
2456         off_t offset;
2457
2458         if (p->num_bad_objects) {
2459                 unsigned i;
2460                 for (i = 0; i < p->num_bad_objects; i++)
2461                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2462                                 return 0;
2463         }
2464
2465         offset = find_pack_entry_one(sha1, p);
2466         if (!offset)
2467                 return 0;
2468
2469         /*
2470          * We are about to tell the caller where they can locate the
2471          * requested object.  We better make sure the packfile is
2472          * still here and can be accessed before supplying that
2473          * answer, as it may have been deleted since the index was
2474          * loaded!
2475          */
2476         if (!is_pack_valid(p)) {
2477                 warning("packfile %s cannot be accessed", p->pack_name);
2478                 return 0;
2479         }
2480         e->offset = offset;
2481         e->p = p;
2482         hashcpy(e->sha1, sha1);
2483         return 1;
2484 }
2485
2486 /*
2487  * Iff a pack file contains the object named by sha1, return true and
2488  * store its location to e.
2489  */
2490 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2491 {
2492         struct packed_git *p;
2493
2494         prepare_packed_git();
2495         if (!packed_git)
2496                 return 0;
2497
2498         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2499                 return 1;
2500
2501         for (p = packed_git; p; p = p->next) {
2502                 if (p == last_found_pack)
2503                         continue; /* we already checked this one */
2504
2505                 if (fill_pack_entry(sha1, e, p)) {
2506                         last_found_pack = p;
2507                         return 1;
2508                 }
2509         }
2510         return 0;
2511 }
2512
2513 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2514                                   struct packed_git *packs)
2515 {
2516         struct packed_git *p;
2517
2518         for (p = packs; p; p = p->next) {
2519                 if (find_pack_entry_one(sha1, p))
2520                         return p;
2521         }
2522         return NULL;
2523
2524 }
2525
2526 static int sha1_loose_object_info(const unsigned char *sha1,
2527                                   struct object_info *oi)
2528 {
2529         int status;
2530         unsigned long mapsize, size;
2531         void *map;
2532         git_zstream stream;
2533         char hdr[32];
2534
2535         if (oi->delta_base_sha1)
2536                 hashclr(oi->delta_base_sha1);
2537
2538         /*
2539          * If we don't care about type or size, then we don't
2540          * need to look inside the object at all. Note that we
2541          * do not optimize out the stat call, even if the
2542          * caller doesn't care about the disk-size, since our
2543          * return value implicitly indicates whether the
2544          * object even exists.
2545          */
2546         if (!oi->typep && !oi->sizep) {
2547                 struct stat st;
2548                 if (stat_sha1_file(sha1, &st) < 0)
2549                         return -1;
2550                 if (oi->disk_sizep)
2551                         *oi->disk_sizep = st.st_size;
2552                 return 0;
2553         }
2554
2555         map = map_sha1_file(sha1, &mapsize);
2556         if (!map)
2557                 return -1;
2558         if (oi->disk_sizep)
2559                 *oi->disk_sizep = mapsize;
2560         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2561                 status = error("unable to unpack %s header",
2562                                sha1_to_hex(sha1));
2563         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2564                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2565         else if (oi->sizep)
2566                 *oi->sizep = size;
2567         git_inflate_end(&stream);
2568         munmap(map, mapsize);
2569         if (oi->typep)
2570                 *oi->typep = status;
2571         return 0;
2572 }
2573
2574 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2575 {
2576         struct cached_object *co;
2577         struct pack_entry e;
2578         int rtype;
2579         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2580
2581         co = find_cached_object(real);
2582         if (co) {
2583                 if (oi->typep)
2584                         *(oi->typep) = co->type;
2585                 if (oi->sizep)
2586                         *(oi->sizep) = co->size;
2587                 if (oi->disk_sizep)
2588                         *(oi->disk_sizep) = 0;
2589                 if (oi->delta_base_sha1)
2590                         hashclr(oi->delta_base_sha1);
2591                 oi->whence = OI_CACHED;
2592                 return 0;
2593         }
2594
2595         if (!find_pack_entry(real, &e)) {
2596                 /* Most likely it's a loose object. */
2597                 if (!sha1_loose_object_info(real, oi)) {
2598                         oi->whence = OI_LOOSE;
2599                         return 0;
2600                 }
2601
2602                 /* Not a loose object; someone else may have just packed it. */
2603                 reprepare_packed_git();
2604                 if (!find_pack_entry(real, &e))
2605                         return -1;
2606         }
2607
2608         rtype = packed_object_info(e.p, e.offset, oi);
2609         if (rtype < 0) {
2610                 mark_bad_packed_object(e.p, real);
2611                 return sha1_object_info_extended(real, oi, 0);
2612         } else if (in_delta_base_cache(e.p, e.offset)) {
2613                 oi->whence = OI_DBCACHED;
2614         } else {
2615                 oi->whence = OI_PACKED;
2616                 oi->u.packed.offset = e.offset;
2617                 oi->u.packed.pack = e.p;
2618                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2619                                          rtype == OBJ_OFS_DELTA);
2620         }
2621
2622         return 0;
2623 }
2624
2625 /* returns enum object_type or negative */
2626 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2627 {
2628         enum object_type type;
2629         struct object_info oi = {NULL};
2630
2631         oi.typep = &type;
2632         oi.sizep = sizep;
2633         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2634                 return -1;
2635         return type;
2636 }
2637
2638 static void *read_packed_sha1(const unsigned char *sha1,
2639                               enum object_type *type, unsigned long *size)
2640 {
2641         struct pack_entry e;
2642         void *data;
2643
2644         if (!find_pack_entry(sha1, &e))
2645                 return NULL;
2646         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2647         if (!data) {
2648                 /*
2649                  * We're probably in deep shit, but let's try to fetch
2650                  * the required object anyway from another pack or loose.
2651                  * This should happen only in the presence of a corrupted
2652                  * pack, and is better than failing outright.
2653                  */
2654                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2655                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2656                 mark_bad_packed_object(e.p, sha1);
2657                 data = read_object(sha1, type, size);
2658         }
2659         return data;
2660 }
2661
2662 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2663                       unsigned char *sha1)
2664 {
2665         struct cached_object *co;
2666
2667         hash_sha1_file(buf, len, typename(type), sha1);
2668         if (has_sha1_file(sha1) || find_cached_object(sha1))
2669                 return 0;
2670         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2671         co = &cached_objects[cached_object_nr++];
2672         co->size = len;
2673         co->type = type;
2674         co->buf = xmalloc(len);
2675         memcpy(co->buf, buf, len);
2676         hashcpy(co->sha1, sha1);
2677         return 0;
2678 }
2679
2680 static void *read_object(const unsigned char *sha1, enum object_type *type,
2681                          unsigned long *size)
2682 {
2683         unsigned long mapsize;
2684         void *map, *buf;
2685         struct cached_object *co;
2686
2687         co = find_cached_object(sha1);
2688         if (co) {
2689                 *type = co->type;
2690                 *size = co->size;
2691                 return xmemdupz(co->buf, co->size);
2692         }
2693
2694         buf = read_packed_sha1(sha1, type, size);
2695         if (buf)
2696                 return buf;
2697         map = map_sha1_file(sha1, &mapsize);
2698         if (map) {
2699                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2700                 munmap(map, mapsize);
2701                 return buf;
2702         }
2703         reprepare_packed_git();
2704         return read_packed_sha1(sha1, type, size);
2705 }
2706
2707 /*
2708  * This function dies on corrupt objects; the callers who want to
2709  * deal with them should arrange to call read_object() and give error
2710  * messages themselves.
2711  */
2712 void *read_sha1_file_extended(const unsigned char *sha1,
2713                               enum object_type *type,
2714                               unsigned long *size,
2715                               unsigned flag)
2716 {
2717         void *data;
2718         const struct packed_git *p;
2719         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2720
2721         errno = 0;
2722         data = read_object(repl, type, size);
2723         if (data)
2724                 return data;
2725
2726         if (errno && errno != ENOENT)
2727                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2728
2729         /* die if we replaced an object with one that does not exist */
2730         if (repl != sha1)
2731                 die("replacement %s not found for %s",
2732                     sha1_to_hex(repl), sha1_to_hex(sha1));
2733
2734         if (has_loose_object(repl)) {
2735                 const char *path = sha1_file_name(sha1);
2736
2737                 die("loose object %s (stored in %s) is corrupt",
2738                     sha1_to_hex(repl), path);
2739         }
2740
2741         if ((p = has_packed_and_bad(repl)) != NULL)
2742                 die("packed object %s (stored in %s) is corrupt",
2743                     sha1_to_hex(repl), p->pack_name);
2744
2745         return NULL;
2746 }
2747
2748 void *read_object_with_reference(const unsigned char *sha1,
2749                                  const char *required_type_name,
2750                                  unsigned long *size,
2751                                  unsigned char *actual_sha1_return)
2752 {
2753         enum object_type type, required_type;
2754         void *buffer;
2755         unsigned long isize;
2756         unsigned char actual_sha1[20];
2757
2758         required_type = type_from_string(required_type_name);
2759         hashcpy(actual_sha1, sha1);
2760         while (1) {
2761                 int ref_length = -1;
2762                 const char *ref_type = NULL;
2763
2764                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2765                 if (!buffer)
2766                         return NULL;
2767                 if (type == required_type) {
2768                         *size = isize;
2769                         if (actual_sha1_return)
2770                                 hashcpy(actual_sha1_return, actual_sha1);
2771                         return buffer;
2772                 }
2773                 /* Handle references */
2774                 else if (type == OBJ_COMMIT)
2775                         ref_type = "tree ";
2776                 else if (type == OBJ_TAG)
2777                         ref_type = "object ";
2778                 else {
2779                         free(buffer);
2780                         return NULL;
2781                 }
2782                 ref_length = strlen(ref_type);
2783
2784                 if (ref_length + 40 > isize ||
2785                     memcmp(buffer, ref_type, ref_length) ||
2786                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2787                         free(buffer);
2788                         return NULL;
2789                 }
2790                 free(buffer);
2791                 /* Now we have the ID of the referred-to object in
2792                  * actual_sha1.  Check again. */
2793         }
2794 }
2795
2796 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2797                                     const char *type, unsigned char *sha1,
2798                                     char *hdr, int *hdrlen)
2799 {
2800         git_SHA_CTX c;
2801
2802         /* Generate the header */
2803         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2804
2805         /* Sha1.. */
2806         git_SHA1_Init(&c);
2807         git_SHA1_Update(&c, hdr, *hdrlen);
2808         git_SHA1_Update(&c, buf, len);
2809         git_SHA1_Final(sha1, &c);
2810 }
2811
2812 /*
2813  * Move the just written object into its final resting place.
2814  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2815  * "moving" is only a part of what it does, when no patch between
2816  * master to pu changes the call sites of this function.
2817  */
2818 int move_temp_to_file(const char *tmpfile, const char *filename)
2819 {
2820         int ret = 0;
2821
2822         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2823                 goto try_rename;
2824         else if (link(tmpfile, filename))
2825                 ret = errno;
2826
2827         /*
2828          * Coda hack - coda doesn't like cross-directory links,
2829          * so we fall back to a rename, which will mean that it
2830          * won't be able to check collisions, but that's not a
2831          * big deal.
2832          *
2833          * The same holds for FAT formatted media.
2834          *
2835          * When this succeeds, we just return.  We have nothing
2836          * left to unlink.
2837          */
2838         if (ret && ret != EEXIST) {
2839         try_rename:
2840                 if (!rename(tmpfile, filename))
2841                         goto out;
2842                 ret = errno;
2843         }
2844         unlink_or_warn(tmpfile);
2845         if (ret) {
2846                 if (ret != EEXIST) {
2847                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2848                 }
2849                 /* FIXME!!! Collision check here ? */
2850         }
2851
2852 out:
2853         if (adjust_shared_perm(filename))
2854                 return error("unable to set permission to '%s'", filename);
2855         return 0;
2856 }
2857
2858 static int write_buffer(int fd, const void *buf, size_t len)
2859 {
2860         if (write_in_full(fd, buf, len) < 0)
2861                 return error("file write error (%s)", strerror(errno));
2862         return 0;
2863 }
2864
2865 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2866                    unsigned char *sha1)
2867 {
2868         char hdr[32];
2869         int hdrlen;
2870         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2871         return 0;
2872 }
2873
2874 /* Finalize a file on disk, and close it. */
2875 static void close_sha1_file(int fd)
2876 {
2877         if (fsync_object_files)
2878                 fsync_or_die(fd, "sha1 file");
2879         if (close(fd) != 0)
2880                 die_errno("error when closing sha1 file");
2881 }
2882
2883 /* Size of directory component, including the ending '/' */
2884 static inline int directory_size(const char *filename)
2885 {
2886         const char *s = strrchr(filename, '/');
2887         if (!s)
2888                 return 0;
2889         return s - filename + 1;
2890 }
2891
2892 /*
2893  * This creates a temporary file in the same directory as the final
2894  * 'filename'
2895  *
2896  * We want to avoid cross-directory filename renames, because those
2897  * can have problems on various filesystems (FAT, NFS, Coda).
2898  */
2899 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2900 {
2901         int fd, dirlen = directory_size(filename);
2902
2903         if (dirlen + 20 > bufsiz) {
2904                 errno = ENAMETOOLONG;
2905                 return -1;
2906         }
2907         memcpy(buffer, filename, dirlen);
2908         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2909         fd = git_mkstemp_mode(buffer, 0444);
2910         if (fd < 0 && dirlen && errno == ENOENT) {
2911                 /* Make sure the directory exists */
2912                 memcpy(buffer, filename, dirlen);
2913                 buffer[dirlen-1] = 0;
2914                 if (mkdir(buffer, 0777) && errno != EEXIST)
2915                         return -1;
2916                 if (adjust_shared_perm(buffer))
2917                         return -1;
2918
2919                 /* Try again */
2920                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2921                 fd = git_mkstemp_mode(buffer, 0444);
2922         }
2923         return fd;
2924 }
2925
2926 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2927                               const void *buf, unsigned long len, time_t mtime)
2928 {
2929         int fd, ret;
2930         unsigned char compressed[4096];
2931         git_zstream stream;
2932         git_SHA_CTX c;
2933         unsigned char parano_sha1[20];
2934         static char tmp_file[PATH_MAX];
2935         const char *filename = sha1_file_name(sha1);
2936
2937         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
2938         if (fd < 0) {
2939                 if (errno == EACCES)
2940                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
2941                 else
2942                         return error("unable to create temporary file: %s", strerror(errno));
2943         }
2944
2945         /* Set it up */
2946         memset(&stream, 0, sizeof(stream));
2947         git_deflate_init(&stream, zlib_compression_level);
2948         stream.next_out = compressed;
2949         stream.avail_out = sizeof(compressed);
2950         git_SHA1_Init(&c);
2951
2952         /* First header.. */
2953         stream.next_in = (unsigned char *)hdr;
2954         stream.avail_in = hdrlen;
2955         while (git_deflate(&stream, 0) == Z_OK)
2956                 ; /* nothing */
2957         git_SHA1_Update(&c, hdr, hdrlen);
2958
2959         /* Then the data itself.. */
2960         stream.next_in = (void *)buf;
2961         stream.avail_in = len;
2962         do {
2963                 unsigned char *in0 = stream.next_in;
2964                 ret = git_deflate(&stream, Z_FINISH);
2965                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2966                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2967                         die("unable to write sha1 file");
2968                 stream.next_out = compressed;
2969                 stream.avail_out = sizeof(compressed);
2970         } while (ret == Z_OK);
2971
2972         if (ret != Z_STREAM_END)
2973                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2974         ret = git_deflate_end_gently(&stream);
2975         if (ret != Z_OK)
2976                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2977         git_SHA1_Final(parano_sha1, &c);
2978         if (hashcmp(sha1, parano_sha1) != 0)
2979                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2980
2981         close_sha1_file(fd);
2982
2983         if (mtime) {
2984                 struct utimbuf utb;
2985                 utb.actime = mtime;
2986                 utb.modtime = mtime;
2987                 if (utime(tmp_file, &utb) < 0)
2988                         warning("failed utime() on %s: %s",
2989                                 tmp_file, strerror(errno));
2990         }
2991
2992         return move_temp_to_file(tmp_file, filename);
2993 }
2994
2995 static int freshen_loose_object(const unsigned char *sha1)
2996 {
2997         return check_and_freshen(sha1, 1);
2998 }
2999
3000 static int freshen_packed_object(const unsigned char *sha1)
3001 {
3002         struct pack_entry e;
3003         return find_pack_entry(sha1, &e) && freshen_file(e.p->pack_name);
3004 }
3005
3006 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
3007 {
3008         unsigned char sha1[20];
3009         char hdr[32];
3010         int hdrlen;
3011
3012         /* Normally if we have it in the pack then we do not bother writing
3013          * it out into .git/objects/??/?{38} file.
3014          */
3015         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3016         if (returnsha1)
3017                 hashcpy(returnsha1, sha1);
3018         if (freshen_loose_object(sha1) || freshen_packed_object(sha1))
3019                 return 0;
3020         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3021 }
3022
3023 int force_object_loose(const unsigned char *sha1, time_t mtime)
3024 {
3025         void *buf;
3026         unsigned long len;
3027         enum object_type type;
3028         char hdr[32];
3029         int hdrlen;
3030         int ret;
3031
3032         if (has_loose_object(sha1))
3033                 return 0;
3034         buf = read_packed_sha1(sha1, &type, &len);
3035         if (!buf)
3036                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3037         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
3038         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3039         free(buf);
3040
3041         return ret;
3042 }
3043
3044 int has_pack_index(const unsigned char *sha1)
3045 {
3046         struct stat st;
3047         if (stat(sha1_pack_index_name(sha1), &st))
3048                 return 0;
3049         return 1;
3050 }
3051
3052 int has_sha1_pack(const unsigned char *sha1)
3053 {
3054         struct pack_entry e;
3055         return find_pack_entry(sha1, &e);
3056 }
3057
3058 int has_sha1_file(const unsigned char *sha1)
3059 {
3060         struct pack_entry e;
3061
3062         if (find_pack_entry(sha1, &e))
3063                 return 1;
3064         if (has_loose_object(sha1))
3065                 return 1;
3066         reprepare_packed_git();
3067         return find_pack_entry(sha1, &e);
3068 }
3069
3070 static void check_tree(const void *buf, size_t size)
3071 {
3072         struct tree_desc desc;
3073         struct name_entry entry;
3074
3075         init_tree_desc(&desc, buf, size);
3076         while (tree_entry(&desc, &entry))
3077                 /* do nothing
3078                  * tree_entry() will die() on malformed entries */
3079                 ;
3080 }
3081
3082 static void check_commit(const void *buf, size_t size)
3083 {
3084         struct commit c;
3085         memset(&c, 0, sizeof(c));
3086         if (parse_commit_buffer(&c, buf, size))
3087                 die("corrupt commit");
3088 }
3089
3090 static void check_tag(const void *buf, size_t size)
3091 {
3092         struct tag t;
3093         memset(&t, 0, sizeof(t));
3094         if (parse_tag_buffer(&t, buf, size))
3095                 die("corrupt tag");
3096 }
3097
3098 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3099                      enum object_type type,
3100                      const char *path, unsigned flags)
3101 {
3102         int ret, re_allocated = 0;
3103         int write_object = flags & HASH_WRITE_OBJECT;
3104
3105         if (!type)
3106                 type = OBJ_BLOB;
3107
3108         /*
3109          * Convert blobs to git internal format
3110          */
3111         if ((type == OBJ_BLOB) && path) {
3112                 struct strbuf nbuf = STRBUF_INIT;
3113                 if (convert_to_git(path, buf, size, &nbuf,
3114                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3115                         buf = strbuf_detach(&nbuf, &size);
3116                         re_allocated = 1;
3117                 }
3118         }
3119         if (flags & HASH_FORMAT_CHECK) {
3120                 if (type == OBJ_TREE)
3121                         check_tree(buf, size);
3122                 if (type == OBJ_COMMIT)
3123                         check_commit(buf, size);
3124                 if (type == OBJ_TAG)
3125                         check_tag(buf, size);
3126         }
3127
3128         if (write_object)
3129                 ret = write_sha1_file(buf, size, typename(type), sha1);
3130         else
3131                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3132         if (re_allocated)
3133                 free(buf);
3134         return ret;
3135 }
3136
3137 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3138                                      const char *path, unsigned flags)
3139 {
3140         int ret;
3141         const int write_object = flags & HASH_WRITE_OBJECT;
3142         struct strbuf sbuf = STRBUF_INIT;
3143
3144         assert(path);
3145         assert(would_convert_to_git_filter_fd(path));
3146
3147         convert_to_git_filter_fd(path, fd, &sbuf,
3148                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3149
3150         if (write_object)
3151                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3152                                       sha1);
3153         else
3154                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3155                                      sha1);
3156         strbuf_release(&sbuf);
3157         return ret;
3158 }
3159
3160 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3161                       const char *path, unsigned flags)
3162 {
3163         struct strbuf sbuf = STRBUF_INIT;
3164         int ret;
3165
3166         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3167                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3168         else
3169                 ret = -1;
3170         strbuf_release(&sbuf);
3171         return ret;
3172 }
3173
3174 #define SMALL_FILE_SIZE (32*1024)
3175
3176 static int index_core(unsigned char *sha1, int fd, size_t size,
3177                       enum object_type type, const char *path,
3178                       unsigned flags)
3179 {
3180         int ret;
3181
3182         if (!size) {
3183                 ret = index_mem(sha1, NULL, size, type, path, flags);
3184         } else if (size <= SMALL_FILE_SIZE) {
3185                 char *buf = xmalloc(size);
3186                 if (size == read_in_full(fd, buf, size))
3187                         ret = index_mem(sha1, buf, size, type, path, flags);
3188                 else
3189                         ret = error("short read %s", strerror(errno));
3190                 free(buf);
3191         } else {
3192                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3193                 ret = index_mem(sha1, buf, size, type, path, flags);
3194                 munmap(buf, size);
3195         }
3196         return ret;
3197 }
3198
3199 /*
3200  * This creates one packfile per large blob unless bulk-checkin
3201  * machinery is "plugged".
3202  *
3203  * This also bypasses the usual "convert-to-git" dance, and that is on
3204  * purpose. We could write a streaming version of the converting
3205  * functions and insert that before feeding the data to fast-import
3206  * (or equivalent in-core API described above). However, that is
3207  * somewhat complicated, as we do not know the size of the filter
3208  * result, which we need to know beforehand when writing a git object.
3209  * Since the primary motivation for trying to stream from the working
3210  * tree file and to avoid mmaping it in core is to deal with large
3211  * binary blobs, they generally do not want to get any conversion, and
3212  * callers should avoid this code path when filters are requested.
3213  */
3214 static int index_stream(unsigned char *sha1, int fd, size_t size,
3215                         enum object_type type, const char *path,
3216                         unsigned flags)
3217 {
3218         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3219 }
3220
3221 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3222              enum object_type type, const char *path, unsigned flags)
3223 {
3224         int ret;
3225
3226         /*
3227          * Call xsize_t() only when needed to avoid potentially unnecessary
3228          * die() for large files.
3229          */
3230         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3231                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3232         else if (!S_ISREG(st->st_mode))
3233                 ret = index_pipe(sha1, fd, type, path, flags);
3234         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3235                  (path && would_convert_to_git(path)))
3236                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3237                                  flags);
3238         else
3239                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3240                                    flags);
3241         close(fd);
3242         return ret;
3243 }
3244
3245 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3246 {
3247         int fd;
3248         struct strbuf sb = STRBUF_INIT;
3249
3250         switch (st->st_mode & S_IFMT) {
3251         case S_IFREG:
3252                 fd = open(path, O_RDONLY);
3253                 if (fd < 0)
3254                         return error("open(\"%s\"): %s", path,
3255                                      strerror(errno));
3256                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3257                         return error("%s: failed to insert into database",
3258                                      path);
3259                 break;
3260         case S_IFLNK:
3261                 if (strbuf_readlink(&sb, path, st->st_size)) {
3262                         char *errstr = strerror(errno);
3263                         return error("readlink(\"%s\"): %s", path,
3264                                      errstr);
3265                 }
3266                 if (!(flags & HASH_WRITE_OBJECT))
3267                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3268                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3269                         return error("%s: failed to insert into database",
3270                                      path);
3271                 strbuf_release(&sb);
3272                 break;
3273         case S_IFDIR:
3274                 return resolve_gitlink_ref(path, "HEAD", sha1);
3275         default:
3276                 return error("%s: unsupported file type", path);
3277         }
3278         return 0;
3279 }
3280
3281 int read_pack_header(int fd, struct pack_header *header)
3282 {
3283         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3284                 /* "eof before pack header was fully read" */
3285                 return PH_ERROR_EOF;
3286
3287         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3288                 /* "protocol error (pack signature mismatch detected)" */
3289                 return PH_ERROR_PACK_SIGNATURE;
3290         if (!pack_version_ok(header->hdr_version))
3291                 /* "protocol error (pack version unsupported)" */
3292                 return PH_ERROR_PROTOCOL;
3293         return 0;
3294 }
3295
3296 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3297 {
3298         enum object_type type = sha1_object_info(sha1, NULL);
3299         if (type < 0)
3300                 die("%s is not a valid object", sha1_to_hex(sha1));
3301         if (type != expect)
3302                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3303                     typename(expect));
3304 }
3305
3306 static int for_each_file_in_obj_subdir(int subdir_nr,
3307                                        struct strbuf *path,
3308                                        each_loose_object_fn obj_cb,
3309                                        each_loose_cruft_fn cruft_cb,
3310                                        each_loose_subdir_fn subdir_cb,
3311                                        void *data)
3312 {
3313         size_t baselen = path->len;
3314         DIR *dir = opendir(path->buf);
3315         struct dirent *de;
3316         int r = 0;
3317
3318         if (!dir) {
3319                 if (errno == ENOENT)
3320                         return 0;
3321                 return error("unable to open %s: %s", path->buf, strerror(errno));
3322         }
3323
3324         while ((de = readdir(dir))) {
3325                 if (is_dot_or_dotdot(de->d_name))
3326                         continue;
3327
3328                 strbuf_setlen(path, baselen);
3329                 strbuf_addf(path, "/%s", de->d_name);
3330
3331                 if (strlen(de->d_name) == 38)  {
3332                         char hex[41];
3333                         unsigned char sha1[20];
3334
3335                         snprintf(hex, sizeof(hex), "%02x%s",
3336                                  subdir_nr, de->d_name);
3337                         if (!get_sha1_hex(hex, sha1)) {
3338                                 if (obj_cb) {
3339                                         r = obj_cb(sha1, path->buf, data);
3340                                         if (r)
3341                                                 break;
3342                                 }
3343                                 continue;
3344                         }
3345                 }
3346
3347                 if (cruft_cb) {
3348                         r = cruft_cb(de->d_name, path->buf, data);
3349                         if (r)
3350                                 break;
3351                 }
3352         }
3353         strbuf_setlen(path, baselen);
3354
3355         if (!r && subdir_cb)
3356                 r = subdir_cb(subdir_nr, path->buf, data);
3357
3358         closedir(dir);
3359         return r;
3360 }
3361
3362 int for_each_loose_file_in_objdir(const char *path,
3363                             each_loose_object_fn obj_cb,
3364                             each_loose_cruft_fn cruft_cb,
3365                             each_loose_subdir_fn subdir_cb,
3366                             void *data)
3367 {
3368         struct strbuf buf = STRBUF_INIT;
3369         size_t baselen;
3370         int r = 0;
3371         int i;
3372
3373         strbuf_addstr(&buf, path);
3374         strbuf_addch(&buf, '/');
3375         baselen = buf.len;
3376
3377         for (i = 0; i < 256; i++) {
3378                 strbuf_addf(&buf, "%02x", i);
3379                 r = for_each_file_in_obj_subdir(i, &buf, obj_cb, cruft_cb,
3380                                                 subdir_cb, data);
3381                 strbuf_setlen(&buf, baselen);
3382                 if (r)
3383                         break;
3384         }
3385
3386         strbuf_release(&buf);
3387         return r;
3388 }
3389
3390 struct loose_alt_odb_data {
3391         each_loose_object_fn *cb;
3392         void *data;
3393 };
3394
3395 static int loose_from_alt_odb(struct alternate_object_database *alt,
3396                               void *vdata)
3397 {
3398         struct loose_alt_odb_data *data = vdata;
3399         return for_each_loose_file_in_objdir(alt->base,
3400                                              data->cb, NULL, NULL,
3401                                              data->data);
3402 }
3403
3404 int for_each_loose_object(each_loose_object_fn cb, void *data)
3405 {
3406         struct loose_alt_odb_data alt;
3407         int r;
3408
3409         r = for_each_loose_file_in_objdir(get_object_directory(),
3410                                           cb, NULL, NULL, data);
3411         if (r)
3412                 return r;
3413
3414         alt.cb = cb;
3415         alt.data = data;
3416         return foreach_alt_odb(loose_from_alt_odb, &alt);
3417 }
3418
3419 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3420 {
3421         uint32_t i;
3422         int r = 0;
3423
3424         for (i = 0; i < p->num_objects; i++) {
3425                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3426
3427                 if (!sha1)
3428                         return error("unable to get sha1 of object %u in %s",
3429                                      i, p->pack_name);
3430
3431                 r = cb(sha1, p, i, data);
3432                 if (r)
3433                         break;
3434         }
3435         return r;
3436 }
3437
3438 int for_each_packed_object(each_packed_object_fn cb, void *data)
3439 {
3440         struct packed_git *p;
3441         int r = 0;
3442
3443         prepare_packed_git();
3444         for (p = packed_git; p; p = p->next) {
3445                 r = for_each_object_in_pack(p, cb, data);
3446                 if (r)
3447                         break;
3448         }
3449         return r;
3450 }