sha1_file.c

   1 /*
   2  * GIT - The information manager from hell
   3  *
   4  * Copyright (C) Linus Torvalds, 2005
   5  *
   6  * This handles basic git sha1 object files - packing, unpacking,
   7  * creation etc.
   8  */
   9 #include "cache.h"
  10 #include "string-list.h"
  11 #include "lockfile.h"
  12 #include "delta.h"
  13 #include "pack.h"
  14 #include "blob.h"
  15 #include "commit.h"
  16 #include "run-command.h"
  17 #include "tag.h"
  18 #include "tree.h"
  19 #include "tree-walk.h"
  20 #include "refs.h"
  21 #include "pack-revindex.h"
  22 #include "sha1-lookup.h"
  23 #include "bulk-checkin.h"
  24 #include "streaming.h"
  25 #include "dir.h"
  26
  27 #ifndef O_NOATIME
  28 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
  29 #define O_NOATIME 01000000
  30 #else
  31 #define O_NOATIME 0
  32 #endif
  33 #endif
  34
  35 #define SZ_FMT PRIuMAX
  36 static inline uintmax_t sz_fmt(size_t s) { return s; }
  37
  38 const unsigned char null_sha1[20];
  39
  40 /*
  41  * This is meant to hold a *small* number of objects that you would
  42  * want read_sha1_file() to be able to return, but yet you do not want
  43  * to write them into the object store (e.g. a browse-only
  44  * application).
  45  */
  46 static struct cached_object {
  47         unsigned char sha1[20];
  48         enum object_type type;
  49         void *buf;
  50         unsigned long size;
  51 } *cached_objects;
  52 static int cached_object_nr, cached_object_alloc;
  53
  54 static struct cached_object empty_tree = {
  55         EMPTY_TREE_SHA1_BIN_LITERAL,
  56         OBJ_TREE,
  57         "",
  58         0
  59 };
  60
  61 /*
  62  * A pointer to the last packed_git in which an object was found.
  63  * When an object is sought, we look in this packfile first, because
  64  * objects that are looked up at similar times are often in the same
  65  * packfile as one another.
  66  */
  67 static struct packed_git *last_found_pack;
  68
  69 static struct cached_object *find_cached_object(const unsigned char *sha1)
  70 {
  71         int i;
  72         struct cached_object *co = cached_objects;
  73
  74         for (i = 0; i < cached_object_nr; i++, co++) {
  75                 if (!hashcmp(co->sha1, sha1))
  76                         return co;
  77         }
  78         if (!hashcmp(sha1, empty_tree.sha1))
  79                 return &empty_tree;
  80         return NULL;
  81 }
  82
  83 int mkdir_in_gitdir(const char *path)
  84 {
  85         if (mkdir(path, 0777)) {
  86                 int saved_errno = errno;
  87                 struct stat st;
  88                 struct strbuf sb = STRBUF_INIT;
  89
  90                 if (errno != EEXIST)
  91                         return -1;
  92                 /*
  93                  * Are we looking at a path in a symlinked worktree
  94                  * whose original repository does not yet have it?
  95                  * e.g. .git/rr-cache pointing at its original
  96                  * repository in which the user hasn't performed any
  97                  * conflict resolution yet?
  98                  */
  99                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
 100                     strbuf_readlink(&sb, path, st.st_size) ||
 101                     !is_absolute_path(sb.buf) ||
 102                     mkdir(sb.buf, 0777)) {
 103                         strbuf_release(&sb);
 104                         errno = saved_errno;
 105                         return -1;
 106                 }
 107                 strbuf_release(&sb);
 108         }
 109         return adjust_shared_perm(path);
 110 }
 111
 112 enum scld_error safe_create_leading_directories(char *path)
 113 {
 114         char *next_component = path + offset_1st_component(path);
 115         enum scld_error ret = SCLD_OK;
 116
 117         while (ret == SCLD_OK && next_component) {
 118                 struct stat st;
 119                 char *slash = next_component, slash_character;
 120
 121                 while (*slash && !is_dir_sep(*slash))
 122                         slash++;
 123
 124                 if (!*slash)
 125                         break;
 126
 127                 next_component = slash + 1;
 128                 while (is_dir_sep(*next_component))
 129                         next_component++;
 130                 if (!*next_component)
 131                         break;
 132
 133                 slash_character = *slash;
 134                 *slash = '\0';
 135                 if (!stat(path, &st)) {
 136                         /* path exists */
 137                         if (!S_ISDIR(st.st_mode))
 138                                 ret = SCLD_EXISTS;
 139                 } else if (mkdir(path, 0777)) {
 140                         if (errno == EEXIST &&
 141                             !stat(path, &st) && S_ISDIR(st.st_mode))
 142                                 ; /* somebody created it since we checked */
 143                         else if (errno == ENOENT)
 144                                 /*
 145                                  * Either mkdir() failed because
 146                                  * somebody just pruned the containing
 147                                  * directory, or stat() failed because
 148                                  * the file that was in our way was
 149                                  * just removed.  Either way, inform
 150                                  * the caller that it might be worth
 151                                  * trying again:
 152                                  */
 153                                 ret = SCLD_VANISHED;
 154                         else
 155                                 ret = SCLD_FAILED;
 156                 } else if (adjust_shared_perm(path)) {
 157                         ret = SCLD_PERMS;
 158                 }
 159                 *slash = slash_character;
 160         }
 161         return ret;
 162 }
 163
 164 enum scld_error safe_create_leading_directories_const(const char *path)
 165 {
 166         /* path points to cache entries, so xstrdup before messing with it */
 167         char *buf = xstrdup(path);
 168         enum scld_error result = safe_create_leading_directories(buf);
 169         free(buf);
 170         return result;
 171 }
 172
 173 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
 174 {
 175         int i;
 176         for (i = 0; i < 20; i++) {
 177                 static char hex[] = "0123456789abcdef";
 178                 unsigned int val = sha1[i];
 179                 char *pos = pathbuf + i*2 + (i > 0);
 180                 *pos++ = hex[val >> 4];
 181                 *pos = hex[val & 0xf];
 182         }
 183 }
 184
 185 const char *sha1_file_name(const unsigned char *sha1)
 186 {
 187         static char buf[PATH_MAX];
 188         const char *objdir;
 189         int len;
 190
 191         objdir = get_object_directory();
 192         len = strlen(objdir);
 193
 194         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
 195         if (len + 43 > PATH_MAX)
 196                 die("insanely long object directory %s", objdir);
 197         memcpy(buf, objdir, len);
 198         buf[len] = '/';
 199         buf[len+3] = '/';
 200         buf[len+42] = '\0';
 201         fill_sha1_path(buf + len + 1, sha1);
 202         return buf;
 203 }
 204
 205 /*
 206  * Return the name of the pack or index file with the specified sha1
 207  * in its filename.  *base and *name are scratch space that must be
 208  * provided by the caller.  which should be "pack" or "idx".
 209  */
 210 static char *sha1_get_pack_name(const unsigned char *sha1,
 211                                 char **name, char **base, const char *which)
 212 {
 213         static const char hex[] = "0123456789abcdef";
 214         char *buf;
 215         int i;
 216
 217         if (!*base) {
 218                 const char *sha1_file_directory = get_object_directory();
 219                 int len = strlen(sha1_file_directory);
 220                 *base = xmalloc(len + 60);
 221                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 222                         sha1_file_directory, which);
 223                 *name = *base + len + 11;
 224         }
 225
 226         buf = *name;
 227
 228         for (i = 0; i < 20; i++) {
 229                 unsigned int val = *sha1++;
 230                 *buf++ = hex[val >> 4];
 231                 *buf++ = hex[val & 0xf];
 232         }
 233
 234         return *base;
 235 }
 236
 237 char *sha1_pack_name(const unsigned char *sha1)
 238 {
 239         static char *name, *base;
 240
 241         return sha1_get_pack_name(sha1, &name, &base, "pack");
 242 }
 243
 244 char *sha1_pack_index_name(const unsigned char *sha1)
 245 {
 246         static char *name, *base;
 247
 248         return sha1_get_pack_name(sha1, &name, &base, "idx");
 249 }
 250
 251 struct alternate_object_database *alt_odb_list;
 252 static struct alternate_object_database **alt_odb_tail;
 253
 254 /*
 255  * Prepare alternate object database registry.
 256  *
 257  * The variable alt_odb_list points at the list of struct
 258  * alternate_object_database.  The elements on this list come from
 259  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 260  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 261  * whose contents is similar to that environment variable but can be
 262  * LF separated.  Its base points at a statically allocated buffer that
 263  * contains "/the/directory/corresponding/to/.git/objects/...", while
 264  * its name points just after the slash at the end of ".git/objects/"
 265  * in the example above, and has enough space to hold 40-byte hex
 266  * SHA1, an extra slash for the first level indirection, and the
 267  * terminating NUL.
 268  */
 269 static int link_alt_odb_entry(const char *entry, const char *relative_base,
 270         int depth, const char *normalized_objdir)
 271 {
 272         struct alternate_object_database *ent;
 273         struct alternate_object_database *alt;
 274         int pfxlen, entlen;
 275         struct strbuf pathbuf = STRBUF_INIT;
 276
 277         if (!is_absolute_path(entry) && relative_base) {
 278                 strbuf_addstr(&pathbuf, real_path(relative_base));
 279                 strbuf_addch(&pathbuf, '/');
 280         }
 281         strbuf_addstr(&pathbuf, entry);
 282
 283         normalize_path_copy(pathbuf.buf, pathbuf.buf);
 284
 285         pfxlen = strlen(pathbuf.buf);
 286
 287         /*
 288          * The trailing slash after the directory name is given by
 289          * this function at the end. Remove duplicates.
 290          */
 291         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
 292                 pfxlen -= 1;
 293
 294         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
 295         ent = xmalloc(sizeof(*ent) + entlen);
 296         memcpy(ent->base, pathbuf.buf, pfxlen);
 297         strbuf_release(&pathbuf);
 298
 299         ent->name = ent->base + pfxlen + 1;
 300         ent->base[pfxlen + 3] = '/';
 301         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 302
 303         /* Detect cases where alternate disappeared */
 304         if (!is_directory(ent->base)) {
 305                 error("object directory %s does not exist; "
 306                       "check .git/objects/info/alternates.",
 307                       ent->base);
 308                 free(ent);
 309                 return -1;
 310         }
 311
 312         /* Prevent the common mistake of listing the same
 313          * thing twice, or object directory itself.
 314          */
 315         for (alt = alt_odb_list; alt; alt = alt->next) {
 316                 if (pfxlen == alt->name - alt->base - 1 &&
 317                     !memcmp(ent->base, alt->base, pfxlen)) {
 318                         free(ent);
 319                         return -1;
 320                 }
 321         }
 322         if (!strcmp_icase(ent->base, normalized_objdir)) {
 323                 free(ent);
 324                 return -1;
 325         }
 326
 327         /* add the alternate entry */
 328         *alt_odb_tail = ent;
 329         alt_odb_tail = &(ent->next);
 330         ent->next = NULL;
 331
 332         /* recursively add alternates */
 333         read_info_alternates(ent->base, depth + 1);
 334
 335         ent->base[pfxlen] = '/';
 336
 337         return 0;
 338 }
 339
 340 static void link_alt_odb_entries(const char *alt, int len, int sep,
 341                                  const char *relative_base, int depth)
 342 {
 343         struct string_list entries = STRING_LIST_INIT_NODUP;
 344         char *alt_copy;
 345         int i;
 346         struct strbuf objdirbuf = STRBUF_INIT;
 347
 348         if (depth > 5) {
 349                 error("%s: ignoring alternate object stores, nesting too deep.",
 350                                 relative_base);
 351                 return;
 352         }
 353
 354         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
 355         normalize_path_copy(objdirbuf.buf, objdirbuf.buf);
 356
 357         alt_copy = xmemdupz(alt, len);
 358         string_list_split_in_place(&entries, alt_copy, sep, -1);
 359         for (i = 0; i < entries.nr; i++) {
 360                 const char *entry = entries.items[i].string;
 361                 if (entry[0] == '\0' || entry[0] == '#')
 362                         continue;
 363                 if (!is_absolute_path(entry) && depth) {
 364                         error("%s: ignoring relative alternate object store %s",
 365                                         relative_base, entry);
 366                 } else {
 367                         link_alt_odb_entry(entry, relative_base, depth, objdirbuf.buf);
 368                 }
 369         }
 370         string_list_clear(&entries, 0);
 371         free(alt_copy);
 372         strbuf_release(&objdirbuf);
 373 }
 374
 375 void read_info_alternates(const char * relative_base, int depth)
 376 {
 377         char *map;
 378         size_t mapsz;
 379         struct stat st;
 380         char *path;
 381         int fd;
 382
 383         path = xstrfmt("%s/info/alternates", relative_base);
 384         fd = git_open_noatime(path);
 385         free(path);
 386         if (fd < 0)
 387                 return;
 388         if (fstat(fd, &st) || (st.st_size == 0)) {
 389                 close(fd);
 390                 return;
 391         }
 392         mapsz = xsize_t(st.st_size);
 393         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 394         close(fd);
 395
 396         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
 397
 398         munmap(map, mapsz);
 399 }
 400
 401 void add_to_alternates_file(const char *reference)
 402 {
 403         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 404         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
 405         char *alt = mkpath("%s\n", reference);
 406         write_or_die(fd, alt, strlen(alt));
 407         if (commit_lock_file(lock))
 408                 die("could not close alternates file");
 409         if (alt_odb_tail)
 410                 link_alt_odb_entries(alt, strlen(alt), '\n', NULL, 0);
 411 }
 412
 413 int foreach_alt_odb(alt_odb_fn fn, void *cb)
 414 {
 415         struct alternate_object_database *ent;
 416         int r = 0;
 417
 418         prepare_alt_odb();
 419         for (ent = alt_odb_list; ent; ent = ent->next) {
 420                 r = fn(ent, cb);
 421                 if (r)
 422                         break;
 423         }
 424         return r;
 425 }
 426
 427 void prepare_alt_odb(void)
 428 {
 429         const char *alt;
 430
 431         if (alt_odb_tail)
 432                 return;
 433
 434         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 435         if (!alt) alt = "";
 436
 437         alt_odb_tail = &alt_odb_list;
 438         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
 439
 440         read_info_alternates(get_object_directory(), 0);
 441 }
 442
 443 static int freshen_file(const char *fn)
 444 {
 445         struct utimbuf t;
 446         t.actime = t.modtime = time(NULL);
 447         return !utime(fn, &t);
 448 }
 449
 450 static int check_and_freshen_file(const char *fn, int freshen)
 451 {
 452         if (access(fn, F_OK))
 453                 return 0;
 454         if (freshen && freshen_file(fn))
 455                 return 0;
 456         return 1;
 457 }
 458
 459 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
 460 {
 461         return check_and_freshen_file(sha1_file_name(sha1), freshen);
 462 }
 463
 464 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
 465 {
 466         struct alternate_object_database *alt;
 467         prepare_alt_odb();
 468         for (alt = alt_odb_list; alt; alt = alt->next) {
 469                 fill_sha1_path(alt->name, sha1);
 470                 if (check_and_freshen_file(alt->base, freshen))
 471                         return 1;
 472         }
 473         return 0;
 474 }
 475
 476 static int check_and_freshen(const unsigned char *sha1, int freshen)
 477 {
 478         return check_and_freshen_local(sha1, freshen) ||
 479                check_and_freshen_nonlocal(sha1, freshen);
 480 }
 481
 482 int has_loose_object_nonlocal(const unsigned char *sha1)
 483 {
 484         return check_and_freshen_nonlocal(sha1, 0);
 485 }
 486
 487 static int has_loose_object(const unsigned char *sha1)
 488 {
 489         return check_and_freshen(sha1, 0);
 490 }
 491
 492 static unsigned int pack_used_ctr;
 493 static unsigned int pack_mmap_calls;
 494 static unsigned int peak_pack_open_windows;
 495 static unsigned int pack_open_windows;
 496 static unsigned int pack_open_fds;
 497 static unsigned int pack_max_fds;
 498 static size_t peak_pack_mapped;
 499 static size_t pack_mapped;
 500 struct packed_git *packed_git;
 501
 502 void pack_report(void)
 503 {
 504         fprintf(stderr,
 505                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 506                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 507                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 508                 sz_fmt(getpagesize()),
 509                 sz_fmt(packed_git_window_size),
 510                 sz_fmt(packed_git_limit));
 511         fprintf(stderr,
 512                 "pack_report: pack_used_ctr            = %10u\n"
 513                 "pack_report: pack_mmap_calls          = %10u\n"
 514                 "pack_report: pack_open_windows        = %10u / %10u\n"
 515                 "pack_report: pack_mapped              = "
 516                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 517                 pack_used_ctr,
 518                 pack_mmap_calls,
 519                 pack_open_windows, peak_pack_open_windows,
 520                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 521 }
 522
 523 /*
 524  * Open and mmap the index file at path, perform a couple of
 525  * consistency checks, then record its information to p.  Return 0 on
 526  * success.
 527  */
 528 static int check_packed_git_idx(const char *path, struct packed_git *p)
 529 {
 530         void *idx_map;
 531         struct pack_idx_header *hdr;
 532         size_t idx_size;
 533         uint32_t version, nr, i, *index;
 534         int fd = git_open_noatime(path);
 535         struct stat st;
 536
 537         if (fd < 0)
 538                 return -1;
 539         if (fstat(fd, &st)) {
 540                 close(fd);
 541                 return -1;
 542         }
 543         idx_size = xsize_t(st.st_size);
 544         if (idx_size < 4 * 256 + 20 + 20) {
 545                 close(fd);
 546                 return error("index file %s is too small", path);
 547         }
 548         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 549         close(fd);
 550
 551         hdr = idx_map;
 552         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 553                 version = ntohl(hdr->idx_version);
 554                 if (version < 2 || version > 2) {
 555                         munmap(idx_map, idx_size);
 556                         return error("index file %s is version %"PRIu32
 557                                      " and is not supported by this binary"
 558                                      " (try upgrading GIT to a newer version)",
 559                                      path, version);
 560                 }
 561         } else
 562                 version = 1;
 563
 564         nr = 0;
 565         index = idx_map;
 566         if (version > 1)
 567                 index += 2;  /* skip index header */
 568         for (i = 0; i < 256; i++) {
 569                 uint32_t n = ntohl(index[i]);
 570                 if (n < nr) {
 571                         munmap(idx_map, idx_size);
 572                         return error("non-monotonic index %s", path);
 573                 }
 574                 nr = n;
 575         }
 576
 577         if (version == 1) {
 578                 /*
 579                  * Total size:
 580                  *  - 256 index entries 4 bytes each
 581                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 582                  *  - 20-byte SHA1 of the packfile
 583                  *  - 20-byte SHA1 file checksum
 584                  */
 585                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 586                         munmap(idx_map, idx_size);
 587                         return error("wrong index v1 file size in %s", path);
 588                 }
 589         } else if (version == 2) {
 590                 /*
 591                  * Minimum size:
 592                  *  - 8 bytes of header
 593                  *  - 256 index entries 4 bytes each
 594                  *  - 20-byte sha1 entry * nr
 595                  *  - 4-byte crc entry * nr
 596                  *  - 4-byte offset entry * nr
 597                  *  - 20-byte SHA1 of the packfile
 598                  *  - 20-byte SHA1 file checksum
 599                  * And after the 4-byte offset table might be a
 600                  * variable sized table containing 8-byte entries
 601                  * for offsets larger than 2^31.
 602                  */
 603                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 604                 unsigned long max_size = min_size;
 605                 if (nr)
 606                         max_size += (nr - 1)*8;
 607                 if (idx_size < min_size || idx_size > max_size) {
 608                         munmap(idx_map, idx_size);
 609                         return error("wrong index v2 file size in %s", path);
 610                 }
 611                 if (idx_size != min_size &&
 612                     /*
 613                      * make sure we can deal with large pack offsets.
 614                      * 31-bit signed offset won't be enough, neither
 615                      * 32-bit unsigned one will be.
 616                      */
 617                     (sizeof(off_t) <= 4)) {
 618                         munmap(idx_map, idx_size);
 619                         return error("pack too large for current definition of off_t in %s", path);
 620                 }
 621         }
 622
 623         p->index_version = version;
 624         p->index_data = idx_map;
 625         p->index_size = idx_size;
 626         p->num_objects = nr;
 627         return 0;
 628 }
 629
 630 int open_pack_index(struct packed_git *p)
 631 {
 632         char *idx_name;
 633         int ret;
 634
 635         if (p->index_data)
 636                 return 0;
 637
 638         idx_name = xstrdup(p->pack_name);
 639         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 640         ret = check_packed_git_idx(idx_name, p);
 641         free(idx_name);
 642         return ret;
 643 }
 644
 645 static void scan_windows(struct packed_git *p,
 646         struct packed_git **lru_p,
 647         struct pack_window **lru_w,
 648         struct pack_window **lru_l)
 649 {
 650         struct pack_window *w, *w_l;
 651
 652         for (w_l = NULL, w = p->windows; w; w = w->next) {
 653                 if (!w->inuse_cnt) {
 654                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 655                                 *lru_p = p;
 656                                 *lru_w = w;
 657                                 *lru_l = w_l;
 658                         }
 659                 }
 660                 w_l = w;
 661         }
 662 }
 663
 664 static int unuse_one_window(struct packed_git *current)
 665 {
 666         struct packed_git *p, *lru_p = NULL;
 667         struct pack_window *lru_w = NULL, *lru_l = NULL;
 668
 669         if (current)
 670                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 671         for (p = packed_git; p; p = p->next)
 672                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 673         if (lru_p) {
 674                 munmap(lru_w->base, lru_w->len);
 675                 pack_mapped -= lru_w->len;
 676                 if (lru_l)
 677                         lru_l->next = lru_w->next;
 678                 else
 679                         lru_p->windows = lru_w->next;
 680                 free(lru_w);
 681                 pack_open_windows--;
 682                 return 1;
 683         }
 684         return 0;
 685 }
 686
 687 void release_pack_memory(size_t need)
 688 {
 689         size_t cur = pack_mapped;
 690         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
 691                 ; /* nothing */
 692 }
 693
 694 static void mmap_limit_check(size_t length)
 695 {
 696         static size_t limit = 0;
 697         if (!limit) {
 698                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
 699                 if (!limit)
 700                         limit = SIZE_MAX;
 701         }
 702         if (length > limit)
 703                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
 704                     (uintmax_t)length, (uintmax_t)limit);
 705 }
 706
 707 void *xmmap(void *start, size_t length,
 708         int prot, int flags, int fd, off_t offset)
 709 {
 710         void *ret;
 711
 712         mmap_limit_check(length);
 713         ret = mmap(start, length, prot, flags, fd, offset);
 714         if (ret == MAP_FAILED) {
 715                 if (!length)
 716                         return NULL;
 717                 release_pack_memory(length);
 718                 ret = mmap(start, length, prot, flags, fd, offset);
 719                 if (ret == MAP_FAILED)
 720                         die_errno("Out of memory? mmap failed");
 721         }
 722         return ret;
 723 }
 724
 725 void close_pack_windows(struct packed_git *p)
 726 {
 727         while (p->windows) {
 728                 struct pack_window *w = p->windows;
 729
 730                 if (w->inuse_cnt)
 731                         die("pack '%s' still has open windows to it",
 732                             p->pack_name);
 733                 munmap(w->base, w->len);
 734                 pack_mapped -= w->len;
 735                 pack_open_windows--;
 736                 p->windows = w->next;
 737                 free(w);
 738         }
 739 }
 740
 741 /*
 742  * The LRU pack is the one with the oldest MRU window, preferring packs
 743  * with no used windows, or the oldest mtime if it has no windows allocated.
 744  */
 745 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 746 {
 747         struct pack_window *w, *this_mru_w;
 748         int has_windows_inuse = 0;
 749
 750         /*
 751          * Reject this pack if it has windows and the previously selected
 752          * one does not.  If this pack does not have windows, reject
 753          * it if the pack file is newer than the previously selected one.
 754          */
 755         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 756                 return;
 757
 758         for (w = this_mru_w = p->windows; w; w = w->next) {
 759                 /*
 760                  * Reject this pack if any of its windows are in use,
 761                  * but the previously selected pack did not have any
 762                  * inuse windows.  Otherwise, record that this pack
 763                  * has windows in use.
 764                  */
 765                 if (w->inuse_cnt) {
 766                         if (*accept_windows_inuse)
 767                                 has_windows_inuse = 1;
 768                         else
 769                                 return;
 770                 }
 771
 772                 if (w->last_used > this_mru_w->last_used)
 773                         this_mru_w = w;
 774
 775                 /*
 776                  * Reject this pack if it has windows that have been
 777                  * used more recently than the previously selected pack.
 778                  * If the previously selected pack had windows inuse and
 779                  * we have not encountered a window in this pack that is
 780                  * inuse, skip this check since we prefer a pack with no
 781                  * inuse windows to one that has inuse windows.
 782                  */
 783                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 784                     this_mru_w->last_used > (*mru_w)->last_used)
 785                         return;
 786         }
 787
 788         /*
 789          * Select this pack.
 790          */
 791         *mru_w = this_mru_w;
 792         *lru_p = p;
 793         *accept_windows_inuse = has_windows_inuse;
 794 }
 795
 796 static int close_one_pack(void)
 797 {
 798         struct packed_git *p, *lru_p = NULL;
 799         struct pack_window *mru_w = NULL;
 800         int accept_windows_inuse = 1;
 801
 802         for (p = packed_git; p; p = p->next) {
 803                 if (p->pack_fd == -1)
 804                         continue;
 805                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 806         }
 807
 808         if (lru_p) {
 809                 close(lru_p->pack_fd);
 810                 pack_open_fds--;
 811                 lru_p->pack_fd = -1;
 812                 return 1;
 813         }
 814
 815         return 0;
 816 }
 817
 818 void unuse_pack(struct pack_window **w_cursor)
 819 {
 820         struct pack_window *w = *w_cursor;
 821         if (w) {
 822                 w->inuse_cnt--;
 823                 *w_cursor = NULL;
 824         }
 825 }
 826
 827 void close_pack_index(struct packed_git *p)
 828 {
 829         if (p->index_data) {
 830                 munmap((void *)p->index_data, p->index_size);
 831                 p->index_data = NULL;
 832         }
 833 }
 834
 835 /*
 836  * This is used by git-repack in case a newly created pack happens to
 837  * contain the same set of objects as an existing one.  In that case
 838  * the resulting file might be different even if its name would be the
 839  * same.  It is best to close any reference to the old pack before it is
 840  * replaced on disk.  Of course no index pointers or windows for given pack
 841  * must subsist at this point.  If ever objects from this pack are requested
 842  * again, the new version of the pack will be reinitialized through
 843  * reprepare_packed_git().
 844  */
 845 void free_pack_by_name(const char *pack_name)
 846 {
 847         struct packed_git *p, **pp = &packed_git;
 848
 849         while (*pp) {
 850                 p = *pp;
 851                 if (strcmp(pack_name, p->pack_name) == 0) {
 852                         clear_delta_base_cache();
 853                         close_pack_windows(p);
 854                         if (p->pack_fd != -1) {
 855                                 close(p->pack_fd);
 856                                 pack_open_fds--;
 857                         }
 858                         close_pack_index(p);
 859                         free(p->bad_object_sha1);
 860                         *pp = p->next;
 861                         if (last_found_pack == p)
 862                                 last_found_pack = NULL;
 863                         free(p);
 864                         return;
 865                 }
 866                 pp = &p->next;
 867         }
 868 }
 869
 870 static unsigned int get_max_fd_limit(void)
 871 {
 872 #ifdef RLIMIT_NOFILE
 873         {
 874                 struct rlimit lim;
 875
 876                 if (!getrlimit(RLIMIT_NOFILE, &lim))
 877                         return lim.rlim_cur;
 878         }
 879 #endif
 880
 881 #ifdef _SC_OPEN_MAX
 882         {
 883                 long open_max = sysconf(_SC_OPEN_MAX);
 884                 if (0 < open_max)
 885                         return open_max;
 886                 /*
 887                  * Otherwise, we got -1 for one of the two
 888                  * reasons:
 889                  *
 890                  * (1) sysconf() did not understand _SC_OPEN_MAX
 891                  *     and signaled an error with -1; or
 892                  * (2) sysconf() said there is no limit.
 893                  *
 894                  * We _could_ clear errno before calling sysconf() to
 895                  * tell these two cases apart and return a huge number
 896                  * in the latter case to let the caller cap it to a
 897                  * value that is not so selfish, but letting the
 898                  * fallback OPEN_MAX codepath take care of these cases
 899                  * is a lot simpler.
 900                  */
 901         }
 902 #endif
 903
 904 #ifdef OPEN_MAX
 905         return OPEN_MAX;
 906 #else
 907         return 1; /* see the caller ;-) */
 908 #endif
 909 }
 910
 911 /*
 912  * Do not call this directly as this leaks p->pack_fd on error return;
 913  * call open_packed_git() instead.
 914  */
 915 static int open_packed_git_1(struct packed_git *p)
 916 {
 917         struct stat st;
 918         struct pack_header hdr;
 919         unsigned char sha1[20];
 920         unsigned char *idx_sha1;
 921         long fd_flag;
 922
 923         if (!p->index_data && open_pack_index(p))
 924                 return error("packfile %s index unavailable", p->pack_name);
 925
 926         if (!pack_max_fds) {
 927                 unsigned int max_fds = get_max_fd_limit();
 928
 929                 /* Save 3 for stdin/stdout/stderr, 22 for work */
 930                 if (25 < max_fds)
 931                         pack_max_fds = max_fds - 25;
 932                 else
 933                         pack_max_fds = 1;
 934         }
 935
 936         while (pack_max_fds <= pack_open_fds && close_one_pack())
 937                 ; /* nothing */
 938
 939         p->pack_fd = git_open_noatime(p->pack_name);
 940         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 941                 return -1;
 942         pack_open_fds++;
 943
 944         /* If we created the struct before we had the pack we lack size. */
 945         if (!p->pack_size) {
 946                 if (!S_ISREG(st.st_mode))
 947                         return error("packfile %s not a regular file", p->pack_name);
 948                 p->pack_size = st.st_size;
 949         } else if (p->pack_size != st.st_size)
 950                 return error("packfile %s size changed", p->pack_name);
 951
 952         /* We leave these file descriptors open with sliding mmap;
 953          * there is no point keeping them open across exec(), though.
 954          */
 955         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 956         if (fd_flag < 0)
 957                 return error("cannot determine file descriptor flags");
 958         fd_flag |= FD_CLOEXEC;
 959         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 960                 return error("cannot set FD_CLOEXEC");
 961
 962         /* Verify we recognize this pack file format. */
 963         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 964                 return error("file %s is far too short to be a packfile", p->pack_name);
 965         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 966                 return error("file %s is not a GIT packfile", p->pack_name);
 967         if (!pack_version_ok(hdr.hdr_version))
 968                 return error("packfile %s is version %"PRIu32" and not"
 969                         " supported (try upgrading GIT to a newer version)",
 970                         p->pack_name, ntohl(hdr.hdr_version));
 971
 972         /* Verify the pack matches its index. */
 973         if (p->num_objects != ntohl(hdr.hdr_entries))
 974                 return error("packfile %s claims to have %"PRIu32" objects"
 975                              " while index indicates %"PRIu32" objects",
 976                              p->pack_name, ntohl(hdr.hdr_entries),
 977                              p->num_objects);
 978         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
 979                 return error("end of packfile %s is unavailable", p->pack_name);
 980         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
 981                 return error("packfile %s signature is unavailable", p->pack_name);
 982         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
 983         if (hashcmp(sha1, idx_sha1))
 984                 return error("packfile %s does not match index", p->pack_name);
 985         return 0;
 986 }
 987
 988 static int open_packed_git(struct packed_git *p)
 989 {
 990         if (!open_packed_git_1(p))
 991                 return 0;
 992         if (p->pack_fd != -1) {
 993                 close(p->pack_fd);
 994                 pack_open_fds--;
 995                 p->pack_fd = -1;
 996         }
 997         return -1;
 998 }
 999
1000 static int in_window(struct pack_window *win, off_t offset)
1001 {
1002         /* We must promise at least 20 bytes (one hash) after the
1003          * offset is available from this window, otherwise the offset
1004          * is not actually in this window and a different window (which
1005          * has that one hash excess) must be used.  This is to support
1006          * the object header and delta base parsing routines below.
1007          */
1008         off_t win_off = win->offset;
1009         return win_off <= offset
1010                 && (offset + 20) <= (win_off + win->len);
1011 }
1012
1013 unsigned char *use_pack(struct packed_git *p,
1014                 struct pack_window **w_cursor,
1015                 off_t offset,
1016                 unsigned long *left)
1017 {
1018         struct pack_window *win = *w_cursor;
1019
1020         /* Since packfiles end in a hash of their content and it's
1021          * pointless to ask for an offset into the middle of that
1022          * hash, and the in_window function above wouldn't match
1023          * don't allow an offset too close to the end of the file.
1024          */
1025         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1026                 die("packfile %s cannot be accessed", p->pack_name);
1027         if (offset > (p->pack_size - 20))
1028                 die("offset beyond end of packfile (truncated pack?)");
1029
1030         if (!win || !in_window(win, offset)) {
1031                 if (win)
1032                         win->inuse_cnt--;
1033                 for (win = p->windows; win; win = win->next) {
1034                         if (in_window(win, offset))
1035                                 break;
1036                 }
1037                 if (!win) {
1038                         size_t window_align = packed_git_window_size / 2;
1039                         off_t len;
1040
1041                         if (p->pack_fd == -1 && open_packed_git(p))
1042                                 die("packfile %s cannot be accessed", p->pack_name);
1043
1044                         win = xcalloc(1, sizeof(*win));
1045                         win->offset = (offset / window_align) * window_align;
1046                         len = p->pack_size - win->offset;
1047                         if (len > packed_git_window_size)
1048                                 len = packed_git_window_size;
1049                         win->len = (size_t)len;
1050                         pack_mapped += win->len;
1051                         while (packed_git_limit < pack_mapped
1052                                 && unuse_one_window(p))
1053                                 ; /* nothing */
1054                         win->base = xmmap(NULL, win->len,
1055                                 PROT_READ, MAP_PRIVATE,
1056                                 p->pack_fd, win->offset);
1057                         if (win->base == MAP_FAILED)
1058                                 die("packfile %s cannot be mapped: %s",
1059                                         p->pack_name,
1060                                         strerror(errno));
1061                         if (!win->offset && win->len == p->pack_size
1062                                 && !p->do_not_close) {
1063                                 close(p->pack_fd);
1064                                 pack_open_fds--;
1065                                 p->pack_fd = -1;
1066                         }
1067                         pack_mmap_calls++;
1068                         pack_open_windows++;
1069                         if (pack_mapped > peak_pack_mapped)
1070                                 peak_pack_mapped = pack_mapped;
1071                         if (pack_open_windows > peak_pack_open_windows)
1072                                 peak_pack_open_windows = pack_open_windows;
1073                         win->next = p->windows;
1074                         p->windows = win;
1075                 }
1076         }
1077         if (win != *w_cursor) {
1078                 win->last_used = pack_used_ctr++;
1079                 win->inuse_cnt++;
1080                 *w_cursor = win;
1081         }
1082         offset -= win->offset;
1083         if (left)
1084                 *left = win->len - xsize_t(offset);
1085         return win->base + offset;
1086 }
1087
1088 static struct packed_git *alloc_packed_git(int extra)
1089 {
1090         struct packed_git *p = xmalloc(sizeof(*p) + extra);
1091         memset(p, 0, sizeof(*p));
1092         p->pack_fd = -1;
1093         return p;
1094 }
1095
1096 static void try_to_free_pack_memory(size_t size)
1097 {
1098         release_pack_memory(size);
1099 }
1100
1101 struct packed_git *add_packed_git(const char *path, int path_len, int local)
1102 {
1103         static int have_set_try_to_free_routine;
1104         struct stat st;
1105         struct packed_git *p = alloc_packed_git(path_len + 2);
1106
1107         if (!have_set_try_to_free_routine) {
1108                 have_set_try_to_free_routine = 1;
1109                 set_try_to_free_routine(try_to_free_pack_memory);
1110         }
1111
1112         /*
1113          * Make sure a corresponding .pack file exists and that
1114          * the index looks sane.
1115          */
1116         path_len -= strlen(".idx");
1117         if (path_len < 1) {
1118                 free(p);
1119                 return NULL;
1120         }
1121         memcpy(p->pack_name, path, path_len);
1122
1123         strcpy(p->pack_name + path_len, ".keep");
1124         if (!access(p->pack_name, F_OK))
1125                 p->pack_keep = 1;
1126
1127         strcpy(p->pack_name + path_len, ".pack");
1128         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1129                 free(p);
1130                 return NULL;
1131         }
1132
1133         /* ok, it looks sane as far as we can check without
1134          * actually mapping the pack file.
1135          */
1136         p->pack_size = st.st_size;
1137         p->pack_local = local;
1138         p->mtime = st.st_mtime;
1139         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1140                 hashclr(p->sha1);
1141         return p;
1142 }
1143
1144 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1145 {
1146         const char *path = sha1_pack_name(sha1);
1147         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
1148
1149         strcpy(p->pack_name, path);
1150         hashcpy(p->sha1, sha1);
1151         if (check_packed_git_idx(idx_path, p)) {
1152                 free(p);
1153                 return NULL;
1154         }
1155
1156         return p;
1157 }
1158
1159 void install_packed_git(struct packed_git *pack)
1160 {
1161         if (pack->pack_fd != -1)
1162                 pack_open_fds++;
1163
1164         pack->next = packed_git;
1165         packed_git = pack;
1166 }
1167
1168 void (*report_garbage)(const char *desc, const char *path);
1169
1170 static void report_helper(const struct string_list *list,
1171                           int seen_bits, int first, int last)
1172 {
1173         const char *msg;
1174         switch (seen_bits) {
1175         case 0:
1176                 msg = "no corresponding .idx or .pack";
1177                 break;
1178         case 1:
1179                 msg = "no corresponding .idx";
1180                 break;
1181         case 2:
1182                 msg = "no corresponding .pack";
1183                 break;
1184         default:
1185                 return;
1186         }
1187         for (; first < last; first++)
1188                 report_garbage(msg, list->items[first].string);
1189 }
1190
1191 static void report_pack_garbage(struct string_list *list)
1192 {
1193         int i, baselen = -1, first = 0, seen_bits = 0;
1194
1195         if (!report_garbage)
1196                 return;
1197
1198         string_list_sort(list);
1199
1200         for (i = 0; i < list->nr; i++) {
1201                 const char *path = list->items[i].string;
1202                 if (baselen != -1 &&
1203                     strncmp(path, list->items[first].string, baselen)) {
1204                         report_helper(list, seen_bits, first, i);
1205                         baselen = -1;
1206                         seen_bits = 0;
1207                 }
1208                 if (baselen == -1) {
1209                         const char *dot = strrchr(path, '.');
1210                         if (!dot) {
1211                                 report_garbage("garbage found", path);
1212                                 continue;
1213                         }
1214                         baselen = dot - path + 1;
1215                         first = i;
1216                 }
1217                 if (!strcmp(path + baselen, "pack"))
1218                         seen_bits |= 1;
1219                 else if (!strcmp(path + baselen, "idx"))
1220                         seen_bits |= 2;
1221         }
1222         report_helper(list, seen_bits, first, list->nr);
1223 }
1224
1225 static void prepare_packed_git_one(char *objdir, int local)
1226 {
1227         struct strbuf path = STRBUF_INIT;
1228         size_t dirnamelen;
1229         DIR *dir;
1230         struct dirent *de;
1231         struct string_list garbage = STRING_LIST_INIT_DUP;
1232
1233         strbuf_addstr(&path, objdir);
1234         strbuf_addstr(&path, "/pack");
1235         dir = opendir(path.buf);
1236         if (!dir) {
1237                 if (errno != ENOENT)
1238                         error("unable to open object pack directory: %s: %s",
1239                               path.buf, strerror(errno));
1240                 strbuf_release(&path);
1241                 return;
1242         }
1243         strbuf_addch(&path, '/');
1244         dirnamelen = path.len;
1245         while ((de = readdir(dir)) != NULL) {
1246                 struct packed_git *p;
1247                 size_t base_len;
1248
1249                 if (is_dot_or_dotdot(de->d_name))
1250                         continue;
1251
1252                 strbuf_setlen(&path, dirnamelen);
1253                 strbuf_addstr(&path, de->d_name);
1254
1255                 base_len = path.len;
1256                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1257                         /* Don't reopen a pack we already have. */
1258                         for (p = packed_git; p; p = p->next) {
1259                                 size_t len;
1260                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1261                                     len == base_len &&
1262                                     !memcmp(p->pack_name, path.buf, len))
1263                                         break;
1264                         }
1265                         if (p == NULL &&
1266                             /*
1267                              * See if it really is a valid .idx file with
1268                              * corresponding .pack file that we can map.
1269                              */
1270                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1271                                 install_packed_git(p);
1272                 }
1273
1274                 if (!report_garbage)
1275                         continue;
1276
1277                 if (ends_with(de->d_name, ".idx") ||
1278                     ends_with(de->d_name, ".pack") ||
1279                     ends_with(de->d_name, ".bitmap") ||
1280                     ends_with(de->d_name, ".keep"))
1281                         string_list_append(&garbage, path.buf);
1282                 else
1283                         report_garbage("garbage found", path.buf);
1284         }
1285         closedir(dir);
1286         report_pack_garbage(&garbage);
1287         string_list_clear(&garbage, 0);
1288         strbuf_release(&path);
1289 }
1290
1291 static int sort_pack(const void *a_, const void *b_)
1292 {
1293         struct packed_git *a = *((struct packed_git **)a_);
1294         struct packed_git *b = *((struct packed_git **)b_);
1295         int st;
1296
1297         /*
1298          * Local packs tend to contain objects specific to our
1299          * variant of the project than remote ones.  In addition,
1300          * remote ones could be on a network mounted filesystem.
1301          * Favor local ones for these reasons.
1302          */
1303         st = a->pack_local - b->pack_local;
1304         if (st)
1305                 return -st;
1306
1307         /*
1308          * Younger packs tend to contain more recent objects,
1309          * and more recent objects tend to get accessed more
1310          * often.
1311          */
1312         if (a->mtime < b->mtime)
1313                 return 1;
1314         else if (a->mtime == b->mtime)
1315                 return 0;
1316         return -1;
1317 }
1318
1319 static void rearrange_packed_git(void)
1320 {
1321         struct packed_git **ary, *p;
1322         int i, n;
1323
1324         for (n = 0, p = packed_git; p; p = p->next)
1325                 n++;
1326         if (n < 2)
1327                 return;
1328
1329         /* prepare an array of packed_git for easier sorting */
1330         ary = xcalloc(n, sizeof(struct packed_git *));
1331         for (n = 0, p = packed_git; p; p = p->next)
1332                 ary[n++] = p;
1333
1334         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1335
1336         /* link them back again */
1337         for (i = 0; i < n - 1; i++)
1338                 ary[i]->next = ary[i + 1];
1339         ary[n - 1]->next = NULL;
1340         packed_git = ary[0];
1341
1342         free(ary);
1343 }
1344
1345 static int prepare_packed_git_run_once = 0;
1346 void prepare_packed_git(void)
1347 {
1348         struct alternate_object_database *alt;
1349
1350         if (prepare_packed_git_run_once)
1351                 return;
1352         prepare_packed_git_one(get_object_directory(), 1);
1353         prepare_alt_odb();
1354         for (alt = alt_odb_list; alt; alt = alt->next) {
1355                 alt->name[-1] = 0;
1356                 prepare_packed_git_one(alt->base, 0);
1357                 alt->name[-1] = '/';
1358         }
1359         rearrange_packed_git();
1360         prepare_packed_git_run_once = 1;
1361 }
1362
1363 void reprepare_packed_git(void)
1364 {
1365         prepare_packed_git_run_once = 0;
1366         prepare_packed_git();
1367 }
1368
1369 static void mark_bad_packed_object(struct packed_git *p,
1370                                    const unsigned char *sha1)
1371 {
1372         unsigned i;
1373         for (i = 0; i < p->num_bad_objects; i++)
1374                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1375                         return;
1376         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1377         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1378         p->num_bad_objects++;
1379 }
1380
1381 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1382 {
1383         struct packed_git *p;
1384         unsigned i;
1385
1386         for (p = packed_git; p; p = p->next)
1387                 for (i = 0; i < p->num_bad_objects; i++)
1388                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1389                                 return p;
1390         return NULL;
1391 }
1392
1393 /*
1394  * With an in-core object data in "map", rehash it to make sure the
1395  * object name actually matches "sha1" to detect object corruption.
1396  * With "map" == NULL, try reading the object named with "sha1" using
1397  * the streaming interface and rehash it to do the same.
1398  */
1399 int check_sha1_signature(const unsigned char *sha1, void *map,
1400                          unsigned long size, const char *type)
1401 {
1402         unsigned char real_sha1[20];
1403         enum object_type obj_type;
1404         struct git_istream *st;
1405         git_SHA_CTX c;
1406         char hdr[32];
1407         int hdrlen;
1408
1409         if (map) {
1410                 hash_sha1_file(map, size, type, real_sha1);
1411                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1412         }
1413
1414         st = open_istream(sha1, &obj_type, &size, NULL);
1415         if (!st)
1416                 return -1;
1417
1418         /* Generate the header */
1419         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1420
1421         /* Sha1.. */
1422         git_SHA1_Init(&c);
1423         git_SHA1_Update(&c, hdr, hdrlen);
1424         for (;;) {
1425                 char buf[1024 * 16];
1426                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1427
1428                 if (readlen < 0) {
1429                         close_istream(st);
1430                         return -1;
1431                 }
1432                 if (!readlen)
1433                         break;
1434                 git_SHA1_Update(&c, buf, readlen);
1435         }
1436         git_SHA1_Final(real_sha1, &c);
1437         close_istream(st);
1438         return hashcmp(sha1, real_sha1) ? -1 : 0;
1439 }
1440
1441 int git_open_noatime(const char *name)
1442 {
1443         static int sha1_file_open_flag = O_NOATIME;
1444
1445         for (;;) {
1446                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1447                 if (fd >= 0)
1448                         return fd;
1449
1450                 /* Might the failure be due to O_NOATIME? */
1451                 if (errno != ENOENT && sha1_file_open_flag) {
1452                         sha1_file_open_flag = 0;
1453                         continue;
1454                 }
1455
1456                 return -1;
1457         }
1458 }
1459
1460 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1461 {
1462         struct alternate_object_database *alt;
1463
1464         if (!lstat(sha1_file_name(sha1), st))
1465                 return 0;
1466
1467         prepare_alt_odb();
1468         errno = ENOENT;
1469         for (alt = alt_odb_list; alt; alt = alt->next) {
1470                 fill_sha1_path(alt->name, sha1);
1471                 if (!lstat(alt->base, st))
1472                         return 0;
1473         }
1474
1475         return -1;
1476 }
1477
1478 static int open_sha1_file(const unsigned char *sha1)
1479 {
1480         int fd;
1481         struct alternate_object_database *alt;
1482         int most_interesting_errno;
1483
1484         fd = git_open_noatime(sha1_file_name(sha1));
1485         if (fd >= 0)
1486                 return fd;
1487         most_interesting_errno = errno;
1488
1489         prepare_alt_odb();
1490         for (alt = alt_odb_list; alt; alt = alt->next) {
1491                 fill_sha1_path(alt->name, sha1);
1492                 fd = git_open_noatime(alt->base);
1493                 if (fd >= 0)
1494                         return fd;
1495                 if (most_interesting_errno == ENOENT)
1496                         most_interesting_errno = errno;
1497         }
1498         errno = most_interesting_errno;
1499         return -1;
1500 }
1501
1502 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1503 {
1504         void *map;
1505         int fd;
1506
1507         fd = open_sha1_file(sha1);
1508         map = NULL;
1509         if (fd >= 0) {
1510                 struct stat st;
1511
1512                 if (!fstat(fd, &st)) {
1513                         *size = xsize_t(st.st_size);
1514                         if (!*size) {
1515                                 /* mmap() is forbidden on empty files */
1516                                 error("object file %s is empty", sha1_file_name(sha1));
1517                                 return NULL;
1518                         }
1519                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1520                 }
1521                 close(fd);
1522         }
1523         return map;
1524 }
1525
1526 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1527                 unsigned long len, enum object_type *type, unsigned long *sizep)
1528 {
1529         unsigned shift;
1530         unsigned long size, c;
1531         unsigned long used = 0;
1532
1533         c = buf[used++];
1534         *type = (c >> 4) & 7;
1535         size = c & 15;
1536         shift = 4;
1537         while (c & 0x80) {
1538                 if (len <= used || bitsizeof(long) <= shift) {
1539                         error("bad object header");
1540                         size = used = 0;
1541                         break;
1542                 }
1543                 c = buf[used++];
1544                 size += (c & 0x7f) << shift;
1545                 shift += 7;
1546         }
1547         *sizep = size;
1548         return used;
1549 }
1550
1551 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1552 {
1553         /* Get the data stream */
1554         memset(stream, 0, sizeof(*stream));
1555         stream->next_in = map;
1556         stream->avail_in = mapsize;
1557         stream->next_out = buffer;
1558         stream->avail_out = bufsiz;
1559
1560         git_inflate_init(stream);
1561         return git_inflate(stream, 0);
1562 }
1563
1564 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1565 {
1566         int bytes = strlen(buffer) + 1;
1567         unsigned char *buf = xmallocz(size);
1568         unsigned long n;
1569         int status = Z_OK;
1570
1571         n = stream->total_out - bytes;
1572         if (n > size)
1573                 n = size;
1574         memcpy(buf, (char *) buffer + bytes, n);
1575         bytes = n;
1576         if (bytes <= size) {
1577                 /*
1578                  * The above condition must be (bytes <= size), not
1579                  * (bytes < size).  In other words, even though we
1580                  * expect no more output and set avail_out to zero,
1581                  * the input zlib stream may have bytes that express
1582                  * "this concludes the stream", and we *do* want to
1583                  * eat that input.
1584                  *
1585                  * Otherwise we would not be able to test that we
1586                  * consumed all the input to reach the expected size;
1587                  * we also want to check that zlib tells us that all
1588                  * went well with status == Z_STREAM_END at the end.
1589                  */
1590                 stream->next_out = buf + bytes;
1591                 stream->avail_out = size - bytes;
1592                 while (status == Z_OK)
1593                         status = git_inflate(stream, Z_FINISH);
1594         }
1595         if (status == Z_STREAM_END && !stream->avail_in) {
1596                 git_inflate_end(stream);
1597                 return buf;
1598         }
1599
1600         if (status < 0)
1601                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1602         else if (stream->avail_in)
1603                 error("garbage at end of loose object '%s'",
1604                       sha1_to_hex(sha1));
1605         free(buf);
1606         return NULL;
1607 }
1608
1609 /*
1610  * We used to just use "sscanf()", but that's actually way
1611  * too permissive for what we want to check. So do an anal
1612  * object header parse by hand.
1613  */
1614 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1615 {
1616         char type[10];
1617         int i;
1618         unsigned long size;
1619
1620         /*
1621          * The type can be at most ten bytes (including the
1622          * terminating '\0' that we add), and is followed by
1623          * a space.
1624          */
1625         i = 0;
1626         for (;;) {
1627                 char c = *hdr++;
1628                 if (c == ' ')
1629                         break;
1630                 type[i++] = c;
1631                 if (i >= sizeof(type))
1632                         return -1;
1633         }
1634         type[i] = 0;
1635
1636         /*
1637          * The length must follow immediately, and be in canonical
1638          * decimal format (ie "010" is not valid).
1639          */
1640         size = *hdr++ - '0';
1641         if (size > 9)
1642                 return -1;
1643         if (size) {
1644                 for (;;) {
1645                         unsigned long c = *hdr - '0';
1646                         if (c > 9)
1647                                 break;
1648                         hdr++;
1649                         size = size * 10 + c;
1650                 }
1651         }
1652         *sizep = size;
1653
1654         /*
1655          * The length must be followed by a zero byte
1656          */
1657         return *hdr ? -1 : type_from_string(type);
1658 }
1659
1660 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1661 {
1662         int ret;
1663         git_zstream stream;
1664         char hdr[8192];
1665
1666         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1667         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1668                 return NULL;
1669
1670         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1671 }
1672
1673 unsigned long get_size_from_delta(struct packed_git *p,
1674                                   struct pack_window **w_curs,
1675                                   off_t curpos)
1676 {
1677         const unsigned char *data;
1678         unsigned char delta_head[20], *in;
1679         git_zstream stream;
1680         int st;
1681
1682         memset(&stream, 0, sizeof(stream));
1683         stream.next_out = delta_head;
1684         stream.avail_out = sizeof(delta_head);
1685
1686         git_inflate_init(&stream);
1687         do {
1688                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1689                 stream.next_in = in;
1690                 st = git_inflate(&stream, Z_FINISH);
1691                 curpos += stream.next_in - in;
1692         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1693                  stream.total_out < sizeof(delta_head));
1694         git_inflate_end(&stream);
1695         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1696                 error("delta data unpack-initial failed");
1697                 return 0;
1698         }
1699
1700         /* Examine the initial part of the delta to figure out
1701          * the result size.
1702          */
1703         data = delta_head;
1704
1705         /* ignore base size */
1706         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1707
1708         /* Read the result size */
1709         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1710 }
1711
1712 static off_t get_delta_base(struct packed_git *p,
1713                                     struct pack_window **w_curs,
1714                                     off_t *curpos,
1715                                     enum object_type type,
1716                                     off_t delta_obj_offset)
1717 {
1718         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1719         off_t base_offset;
1720
1721         /* use_pack() assured us we have [base_info, base_info + 20)
1722          * as a range that we can look at without walking off the
1723          * end of the mapped window.  Its actually the hash size
1724          * that is assured.  An OFS_DELTA longer than the hash size
1725          * is stupid, as then a REF_DELTA would be smaller to store.
1726          */
1727         if (type == OBJ_OFS_DELTA) {
1728                 unsigned used = 0;
1729                 unsigned char c = base_info[used++];
1730                 base_offset = c & 127;
1731                 while (c & 128) {
1732                         base_offset += 1;
1733                         if (!base_offset || MSB(base_offset, 7))
1734                                 return 0;  /* overflow */
1735                         c = base_info[used++];
1736                         base_offset = (base_offset << 7) + (c & 127);
1737                 }
1738                 base_offset = delta_obj_offset - base_offset;
1739                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1740                         return 0;  /* out of bound */
1741                 *curpos += used;
1742         } else if (type == OBJ_REF_DELTA) {
1743                 /* The base entry _must_ be in the same pack */
1744                 base_offset = find_pack_entry_one(base_info, p);
1745                 *curpos += 20;
1746         } else
1747                 die("I am totally screwed");
1748         return base_offset;
1749 }
1750
1751 /*
1752  * Like get_delta_base above, but we return the sha1 instead of the pack
1753  * offset. This means it is cheaper for REF deltas (we do not have to do
1754  * the final object lookup), but more expensive for OFS deltas (we
1755  * have to load the revidx to convert the offset back into a sha1).
1756  */
1757 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1758                                                 struct pack_window **w_curs,
1759                                                 off_t curpos,
1760                                                 enum object_type type,
1761                                                 off_t delta_obj_offset)
1762 {
1763         if (type == OBJ_REF_DELTA) {
1764                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1765                 return base;
1766         } else if (type == OBJ_OFS_DELTA) {
1767                 struct revindex_entry *revidx;
1768                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1769                                                    type, delta_obj_offset);
1770
1771                 if (!base_offset)
1772                         return NULL;
1773
1774                 revidx = find_pack_revindex(p, base_offset);
1775                 if (!revidx)
1776                         return NULL;
1777
1778                 return nth_packed_object_sha1(p, revidx->nr);
1779         } else
1780                 return NULL;
1781 }
1782
1783 int unpack_object_header(struct packed_git *p,
1784                          struct pack_window **w_curs,
1785                          off_t *curpos,
1786                          unsigned long *sizep)
1787 {
1788         unsigned char *base;
1789         unsigned long left;
1790         unsigned long used;
1791         enum object_type type;
1792
1793         /* use_pack() assures us we have [base, base + 20) available
1794          * as a range that we can look at.  (Its actually the hash
1795          * size that is assured.)  With our object header encoding
1796          * the maximum deflated object size is 2^137, which is just
1797          * insane, so we know won't exceed what we have been given.
1798          */
1799         base = use_pack(p, w_curs, *curpos, &left);
1800         used = unpack_object_header_buffer(base, left, &type, sizep);
1801         if (!used) {
1802                 type = OBJ_BAD;
1803         } else
1804                 *curpos += used;
1805
1806         return type;
1807 }
1808
1809 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1810 {
1811         int type;
1812         struct revindex_entry *revidx;
1813         const unsigned char *sha1;
1814         revidx = find_pack_revindex(p, obj_offset);
1815         if (!revidx)
1816                 return OBJ_BAD;
1817         sha1 = nth_packed_object_sha1(p, revidx->nr);
1818         mark_bad_packed_object(p, sha1);
1819         type = sha1_object_info(sha1, NULL);
1820         if (type <= OBJ_NONE)
1821                 return OBJ_BAD;
1822         return type;
1823 }
1824
1825 #define POI_STACK_PREALLOC 64
1826
1827 static enum object_type packed_to_object_type(struct packed_git *p,
1828                                               off_t obj_offset,
1829                                               enum object_type type,
1830                                               struct pack_window **w_curs,
1831                                               off_t curpos)
1832 {
1833         off_t small_poi_stack[POI_STACK_PREALLOC];
1834         off_t *poi_stack = small_poi_stack;
1835         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1836
1837         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1838                 off_t base_offset;
1839                 unsigned long size;
1840                 /* Push the object we're going to leave behind */
1841                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1842                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1843                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1844                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1845                 } else {
1846                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1847                 }
1848                 poi_stack[poi_stack_nr++] = obj_offset;
1849                 /* If parsing the base offset fails, just unwind */
1850                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1851                 if (!base_offset)
1852                         goto unwind;
1853                 curpos = obj_offset = base_offset;
1854                 type = unpack_object_header(p, w_curs, &curpos, &size);
1855                 if (type <= OBJ_NONE) {
1856                         /* If getting the base itself fails, we first
1857                          * retry the base, otherwise unwind */
1858                         type = retry_bad_packed_offset(p, base_offset);
1859                         if (type > OBJ_NONE)
1860                                 goto out;
1861                         goto unwind;
1862                 }
1863         }
1864
1865         switch (type) {
1866         case OBJ_BAD:
1867         case OBJ_COMMIT:
1868         case OBJ_TREE:
1869         case OBJ_BLOB:
1870         case OBJ_TAG:
1871                 break;
1872         default:
1873                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1874                       type, (uintmax_t)obj_offset, p->pack_name);
1875                 type = OBJ_BAD;
1876         }
1877
1878 out:
1879         if (poi_stack != small_poi_stack)
1880                 free(poi_stack);
1881         return type;
1882
1883 unwind:
1884         while (poi_stack_nr) {
1885                 obj_offset = poi_stack[--poi_stack_nr];
1886                 type = retry_bad_packed_offset(p, obj_offset);
1887                 if (type > OBJ_NONE)
1888                         goto out;
1889         }
1890         type = OBJ_BAD;
1891         goto out;
1892 }
1893
1894 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1895                               struct object_info *oi)
1896 {
1897         struct pack_window *w_curs = NULL;
1898         unsigned long size;
1899         off_t curpos = obj_offset;
1900         enum object_type type;
1901
1902         /*
1903          * We always get the representation type, but only convert it to
1904          * a "real" type later if the caller is interested.
1905          */
1906         type = unpack_object_header(p, &w_curs, &curpos, &size);
1907
1908         if (oi->sizep) {
1909                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1910                         off_t tmp_pos = curpos;
1911                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1912                                                            type, obj_offset);
1913                         if (!base_offset) {
1914                                 type = OBJ_BAD;
1915                                 goto out;
1916                         }
1917                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1918                         if (*oi->sizep == 0) {
1919                                 type = OBJ_BAD;
1920                                 goto out;
1921                         }
1922                 } else {
1923                         *oi->sizep = size;
1924                 }
1925         }
1926
1927         if (oi->disk_sizep) {
1928                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1929                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1930         }
1931
1932         if (oi->typep) {
1933                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
1934                 if (*oi->typep < 0) {
1935                         type = OBJ_BAD;
1936                         goto out;
1937                 }
1938         }
1939
1940         if (oi->delta_base_sha1) {
1941                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1942                         const unsigned char *base;
1943
1944                         base = get_delta_base_sha1(p, &w_curs, curpos,
1945                                                    type, obj_offset);
1946                         if (!base) {
1947                                 type = OBJ_BAD;
1948                                 goto out;
1949                         }
1950
1951                         hashcpy(oi->delta_base_sha1, base);
1952                 } else
1953                         hashclr(oi->delta_base_sha1);
1954         }
1955
1956 out:
1957         unuse_pack(&w_curs);
1958         return type;
1959 }
1960
1961 static void *unpack_compressed_entry(struct packed_git *p,
1962                                     struct pack_window **w_curs,
1963                                     off_t curpos,
1964                                     unsigned long size)
1965 {
1966         int st;
1967         git_zstream stream;
1968         unsigned char *buffer, *in;
1969
1970         buffer = xmallocz_gently(size);
1971         if (!buffer)
1972                 return NULL;
1973         memset(&stream, 0, sizeof(stream));
1974         stream.next_out = buffer;
1975         stream.avail_out = size + 1;
1976
1977         git_inflate_init(&stream);
1978         do {
1979                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1980                 stream.next_in = in;
1981                 st = git_inflate(&stream, Z_FINISH);
1982                 if (!stream.avail_out)
1983                         break; /* the payload is larger than it should be */
1984                 curpos += stream.next_in - in;
1985         } while (st == Z_OK || st == Z_BUF_ERROR);
1986         git_inflate_end(&stream);
1987         if ((st != Z_STREAM_END) || stream.total_out != size) {
1988                 free(buffer);
1989                 return NULL;
1990         }
1991
1992         return buffer;
1993 }
1994
1995 #define MAX_DELTA_CACHE (256)
1996
1997 static size_t delta_base_cached;
1998
1999 static struct delta_base_cache_lru_list {
2000         struct delta_base_cache_lru_list *prev;
2001         struct delta_base_cache_lru_list *next;
2002 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
2003
2004 static struct delta_base_cache_entry {
2005         struct delta_base_cache_lru_list lru;
2006         void *data;
2007         struct packed_git *p;
2008         off_t base_offset;
2009         unsigned long size;
2010         enum object_type type;
2011 } delta_base_cache[MAX_DELTA_CACHE];
2012
2013 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
2014 {
2015         unsigned long hash;
2016
2017         hash = (unsigned long)p + (unsigned long)base_offset;
2018         hash += (hash >> 8) + (hash >> 16);
2019         return hash % MAX_DELTA_CACHE;
2020 }
2021
2022 static struct delta_base_cache_entry *
2023 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2024 {
2025         unsigned long hash = pack_entry_hash(p, base_offset);
2026         return delta_base_cache + hash;
2027 }
2028
2029 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
2030                                      struct packed_git *p, off_t base_offset)
2031 {
2032         return (ent->data && ent->p == p && ent->base_offset == base_offset);
2033 }
2034
2035 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2036 {
2037         struct delta_base_cache_entry *ent;
2038         ent = get_delta_base_cache_entry(p, base_offset);
2039         return eq_delta_base_cache_entry(ent, p, base_offset);
2040 }
2041
2042 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2043 {
2044         ent->data = NULL;
2045         ent->lru.next->prev = ent->lru.prev;
2046         ent->lru.prev->next = ent->lru.next;
2047         delta_base_cached -= ent->size;
2048 }
2049
2050 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2051         unsigned long *base_size, enum object_type *type, int keep_cache)
2052 {
2053         struct delta_base_cache_entry *ent;
2054         void *ret;
2055
2056         ent = get_delta_base_cache_entry(p, base_offset);
2057
2058         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2059                 return unpack_entry(p, base_offset, type, base_size);
2060
2061         ret = ent->data;
2062
2063         if (!keep_cache)
2064                 clear_delta_base_cache_entry(ent);
2065         else
2066                 ret = xmemdupz(ent->data, ent->size);
2067         *type = ent->type;
2068         *base_size = ent->size;
2069         return ret;
2070 }
2071
2072 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2073 {
2074         if (ent->data) {
2075                 free(ent->data);
2076                 ent->data = NULL;
2077                 ent->lru.next->prev = ent->lru.prev;
2078                 ent->lru.prev->next = ent->lru.next;
2079                 delta_base_cached -= ent->size;
2080         }
2081 }
2082
2083 void clear_delta_base_cache(void)
2084 {
2085         unsigned long p;
2086         for (p = 0; p < MAX_DELTA_CACHE; p++)
2087                 release_delta_base_cache(&delta_base_cache[p]);
2088 }
2089
2090 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2091         void *base, unsigned long base_size, enum object_type type)
2092 {
2093         unsigned long hash = pack_entry_hash(p, base_offset);
2094         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2095         struct delta_base_cache_lru_list *lru;
2096
2097         release_delta_base_cache(ent);
2098         delta_base_cached += base_size;
2099
2100         for (lru = delta_base_cache_lru.next;
2101              delta_base_cached > delta_base_cache_limit
2102              && lru != &delta_base_cache_lru;
2103              lru = lru->next) {
2104                 struct delta_base_cache_entry *f = (void *)lru;
2105                 if (f->type == OBJ_BLOB)
2106                         release_delta_base_cache(f);
2107         }
2108         for (lru = delta_base_cache_lru.next;
2109              delta_base_cached > delta_base_cache_limit
2110              && lru != &delta_base_cache_lru;
2111              lru = lru->next) {
2112                 struct delta_base_cache_entry *f = (void *)lru;
2113                 release_delta_base_cache(f);
2114         }
2115
2116         ent->p = p;
2117         ent->base_offset = base_offset;
2118         ent->type = type;
2119         ent->data = base;
2120         ent->size = base_size;
2121         ent->lru.next = &delta_base_cache_lru;
2122         ent->lru.prev = delta_base_cache_lru.prev;
2123         delta_base_cache_lru.prev->next = &ent->lru;
2124         delta_base_cache_lru.prev = &ent->lru;
2125 }
2126
2127 static void *read_object(const unsigned char *sha1, enum object_type *type,
2128                          unsigned long *size);
2129
2130 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2131 {
2132         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2133         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2134                          p->pack_name, (uintmax_t)obj_offset);
2135 }
2136
2137 int do_check_packed_object_crc;
2138
2139 #define UNPACK_ENTRY_STACK_PREALLOC 64
2140 struct unpack_entry_stack_ent {
2141         off_t obj_offset;
2142         off_t curpos;
2143         unsigned long size;
2144 };
2145
2146 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2147                    enum object_type *final_type, unsigned long *final_size)
2148 {
2149         struct pack_window *w_curs = NULL;
2150         off_t curpos = obj_offset;
2151         void *data = NULL;
2152         unsigned long size;
2153         enum object_type type;
2154         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2155         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2156         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2157         int base_from_cache = 0;
2158
2159         write_pack_access_log(p, obj_offset);
2160
2161         /* PHASE 1: drill down to the innermost base object */
2162         for (;;) {
2163                 off_t base_offset;
2164                 int i;
2165                 struct delta_base_cache_entry *ent;
2166
2167                 ent = get_delta_base_cache_entry(p, curpos);
2168                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2169                         type = ent->type;
2170                         data = ent->data;
2171                         size = ent->size;
2172                         clear_delta_base_cache_entry(ent);
2173                         base_from_cache = 1;
2174                         break;
2175                 }
2176
2177                 if (do_check_packed_object_crc && p->index_version > 1) {
2178                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2179                         unsigned long len = revidx[1].offset - obj_offset;
2180                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2181                                 const unsigned char *sha1 =
2182                                         nth_packed_object_sha1(p, revidx->nr);
2183                                 error("bad packed object CRC for %s",
2184                                       sha1_to_hex(sha1));
2185                                 mark_bad_packed_object(p, sha1);
2186                                 unuse_pack(&w_curs);
2187                                 return NULL;
2188                         }
2189                 }
2190
2191                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2192                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2193                         break;
2194
2195                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2196                 if (!base_offset) {
2197                         error("failed to validate delta base reference "
2198                               "at offset %"PRIuMAX" from %s",
2199                               (uintmax_t)curpos, p->pack_name);
2200                         /* bail to phase 2, in hopes of recovery */
2201                         data = NULL;
2202                         break;
2203                 }
2204
2205                 /* push object, proceed to base */
2206                 if (delta_stack_nr >= delta_stack_alloc
2207                     && delta_stack == small_delta_stack) {
2208                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2209                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2210                         memcpy(delta_stack, small_delta_stack,
2211                                sizeof(*delta_stack)*delta_stack_nr);
2212                 } else {
2213                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2214                 }
2215                 i = delta_stack_nr++;
2216                 delta_stack[i].obj_offset = obj_offset;
2217                 delta_stack[i].curpos = curpos;
2218                 delta_stack[i].size = size;
2219
2220                 curpos = obj_offset = base_offset;
2221         }
2222
2223         /* PHASE 2: handle the base */
2224         switch (type) {
2225         case OBJ_OFS_DELTA:
2226         case OBJ_REF_DELTA:
2227                 if (data)
2228                         die("BUG in unpack_entry: left loop at a valid delta");
2229                 break;
2230         case OBJ_COMMIT:
2231         case OBJ_TREE:
2232         case OBJ_BLOB:
2233         case OBJ_TAG:
2234                 if (!base_from_cache)
2235                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2236                 break;
2237         default:
2238                 data = NULL;
2239                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2240                       type, (uintmax_t)obj_offset, p->pack_name);
2241         }
2242
2243         /* PHASE 3: apply deltas in order */
2244
2245         /* invariants:
2246          *   'data' holds the base data, or NULL if there was corruption
2247          */
2248         while (delta_stack_nr) {
2249                 void *delta_data;
2250                 void *base = data;
2251                 unsigned long delta_size, base_size = size;
2252                 int i;
2253
2254                 data = NULL;
2255
2256                 if (base)
2257                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2258
2259                 if (!base) {
2260                         /*
2261                          * We're probably in deep shit, but let's try to fetch
2262                          * the required base anyway from another pack or loose.
2263                          * This is costly but should happen only in the presence
2264                          * of a corrupted pack, and is better than failing outright.
2265                          */
2266                         struct revindex_entry *revidx;
2267                         const unsigned char *base_sha1;
2268                         revidx = find_pack_revindex(p, obj_offset);
2269                         if (revidx) {
2270                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2271                                 error("failed to read delta base object %s"
2272                                       " at offset %"PRIuMAX" from %s",
2273                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2274                                       p->pack_name);
2275                                 mark_bad_packed_object(p, base_sha1);
2276                                 base = read_object(base_sha1, &type, &base_size);
2277                         }
2278                 }
2279
2280                 i = --delta_stack_nr;
2281                 obj_offset = delta_stack[i].obj_offset;
2282                 curpos = delta_stack[i].curpos;
2283                 delta_size = delta_stack[i].size;
2284
2285                 if (!base)
2286                         continue;
2287
2288                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2289
2290                 if (!delta_data) {
2291                         error("failed to unpack compressed delta "
2292                               "at offset %"PRIuMAX" from %s",
2293                               (uintmax_t)curpos, p->pack_name);
2294                         data = NULL;
2295                         continue;
2296                 }
2297
2298                 data = patch_delta(base, base_size,
2299                                    delta_data, delta_size,
2300                                    &size);
2301
2302                 /*
2303                  * We could not apply the delta; warn the user, but keep going.
2304                  * Our failure will be noticed either in the next iteration of
2305                  * the loop, or if this is the final delta, in the caller when
2306                  * we return NULL. Those code paths will take care of making
2307                  * a more explicit warning and retrying with another copy of
2308                  * the object.
2309                  */
2310                 if (!data)
2311                         error("failed to apply delta");
2312
2313                 free(delta_data);
2314         }
2315
2316         *final_type = type;
2317         *final_size = size;
2318
2319         unuse_pack(&w_curs);
2320
2321         if (delta_stack != small_delta_stack)
2322                 free(delta_stack);
2323
2324         return data;
2325 }
2326
2327 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2328                                             uint32_t n)
2329 {
2330         const unsigned char *index = p->index_data;
2331         if (!index) {
2332                 if (open_pack_index(p))
2333                         return NULL;
2334                 index = p->index_data;
2335         }
2336         if (n >= p->num_objects)
2337                 return NULL;
2338         index += 4 * 256;
2339         if (p->index_version == 1) {
2340                 return index + 24 * n + 4;
2341         } else {
2342                 index += 8;
2343                 return index + 20 * n;
2344         }
2345 }
2346
2347 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2348 {
2349         const unsigned char *index = p->index_data;
2350         index += 4 * 256;
2351         if (p->index_version == 1) {
2352                 return ntohl(*((uint32_t *)(index + 24 * n)));
2353         } else {
2354                 uint32_t off;
2355                 index += 8 + p->num_objects * (20 + 4);
2356                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2357                 if (!(off & 0x80000000))
2358                         return off;
2359                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2360                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2361                                    ntohl(*((uint32_t *)(index + 4)));
2362         }
2363 }
2364
2365 off_t find_pack_entry_one(const unsigned char *sha1,
2366                                   struct packed_git *p)
2367 {
2368         const uint32_t *level1_ofs = p->index_data;
2369         const unsigned char *index = p->index_data;
2370         unsigned hi, lo, stride;
2371         static int use_lookup = -1;
2372         static int debug_lookup = -1;
2373
2374         if (debug_lookup < 0)
2375                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2376
2377         if (!index) {
2378                 if (open_pack_index(p))
2379                         return 0;
2380                 level1_ofs = p->index_data;
2381                 index = p->index_data;
2382         }
2383         if (p->index_version > 1) {
2384                 level1_ofs += 2;
2385                 index += 8;
2386         }
2387         index += 4 * 256;
2388         hi = ntohl(level1_ofs[*sha1]);
2389         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2390         if (p->index_version > 1) {
2391                 stride = 20;
2392         } else {
2393                 stride = 24;
2394                 index += 4;
2395         }
2396
2397         if (debug_lookup)
2398                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2399                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2400
2401         if (use_lookup < 0)
2402                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2403         if (use_lookup) {
2404                 int pos = sha1_entry_pos(index, stride, 0,
2405                                          lo, hi, p->num_objects, sha1);
2406                 if (pos < 0)
2407                         return 0;
2408                 return nth_packed_object_offset(p, pos);
2409         }
2410
2411         do {
2412                 unsigned mi = (lo + hi) / 2;
2413                 int cmp = hashcmp(index + mi * stride, sha1);
2414
2415                 if (debug_lookup)
2416                         printf("lo %u hi %u rg %u mi %u\n",
2417                                lo, hi, hi - lo, mi);
2418                 if (!cmp)
2419                         return nth_packed_object_offset(p, mi);
2420                 if (cmp > 0)
2421                         hi = mi;
2422                 else
2423                         lo = mi+1;
2424         } while (lo < hi);
2425         return 0;
2426 }
2427
2428 int is_pack_valid(struct packed_git *p)
2429 {
2430         /* An already open pack is known to be valid. */
2431         if (p->pack_fd != -1)
2432                 return 1;
2433
2434         /* If the pack has one window completely covering the
2435          * file size, the pack is known to be valid even if
2436          * the descriptor is not currently open.
2437          */
2438         if (p->windows) {
2439                 struct pack_window *w = p->windows;
2440
2441                 if (!w->offset && w->len == p->pack_size)
2442                         return 1;
2443         }
2444
2445         /* Force the pack to open to prove its valid. */
2446         return !open_packed_git(p);
2447 }
2448
2449 static int fill_pack_entry(const unsigned char *sha1,
2450                            struct pack_entry *e,
2451                            struct packed_git *p)
2452 {
2453         off_t offset;
2454
2455         if (p->num_bad_objects) {
2456                 unsigned i;
2457                 for (i = 0; i < p->num_bad_objects; i++)
2458                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2459                                 return 0;
2460         }
2461
2462         offset = find_pack_entry_one(sha1, p);
2463         if (!offset)
2464                 return 0;
2465
2466         /*
2467          * We are about to tell the caller where they can locate the
2468          * requested object.  We better make sure the packfile is
2469          * still here and can be accessed before supplying that
2470          * answer, as it may have been deleted since the index was
2471          * loaded!
2472          */
2473         if (!is_pack_valid(p)) {
2474                 warning("packfile %s cannot be accessed", p->pack_name);
2475                 return 0;
2476         }
2477         e->offset = offset;
2478         e->p = p;
2479         hashcpy(e->sha1, sha1);
2480         return 1;
2481 }
2482
2483 /*
2484  * Iff a pack file contains the object named by sha1, return true and
2485  * store its location to e.
2486  */
2487 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2488 {
2489         struct packed_git *p;
2490
2491         prepare_packed_git();
2492         if (!packed_git)
2493                 return 0;
2494
2495         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2496                 return 1;
2497
2498         for (p = packed_git; p; p = p->next) {
2499                 if (p == last_found_pack)
2500                         continue; /* we already checked this one */
2501
2502                 if (fill_pack_entry(sha1, e, p)) {
2503                         last_found_pack = p;
2504                         return 1;
2505                 }
2506         }
2507         return 0;
2508 }
2509
2510 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2511                                   struct packed_git *packs)
2512 {
2513         struct packed_git *p;
2514
2515         for (p = packs; p; p = p->next) {
2516                 if (find_pack_entry_one(sha1, p))
2517                         return p;
2518         }
2519         return NULL;
2520
2521 }
2522
2523 static int sha1_loose_object_info(const unsigned char *sha1,
2524                                   struct object_info *oi)
2525 {
2526         int status;
2527         unsigned long mapsize, size;
2528         void *map;
2529         git_zstream stream;
2530         char hdr[32];
2531
2532         if (oi->delta_base_sha1)
2533                 hashclr(oi->delta_base_sha1);
2534
2535         /*
2536          * If we don't care about type or size, then we don't
2537          * need to look inside the object at all. Note that we
2538          * do not optimize out the stat call, even if the
2539          * caller doesn't care about the disk-size, since our
2540          * return value implicitly indicates whether the
2541          * object even exists.
2542          */
2543         if (!oi->typep && !oi->sizep) {
2544                 struct stat st;
2545                 if (stat_sha1_file(sha1, &st) < 0)
2546                         return -1;
2547                 if (oi->disk_sizep)
2548                         *oi->disk_sizep = st.st_size;
2549                 return 0;
2550         }
2551
2552         map = map_sha1_file(sha1, &mapsize);
2553         if (!map)
2554                 return -1;
2555         if (oi->disk_sizep)
2556                 *oi->disk_sizep = mapsize;
2557         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2558                 status = error("unable to unpack %s header",
2559                                sha1_to_hex(sha1));
2560         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2561                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2562         else if (oi->sizep)
2563                 *oi->sizep = size;
2564         git_inflate_end(&stream);
2565         munmap(map, mapsize);
2566         if (oi->typep)
2567                 *oi->typep = status;
2568         return 0;
2569 }
2570
2571 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2572 {
2573         struct cached_object *co;
2574         struct pack_entry e;
2575         int rtype;
2576         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2577
2578         co = find_cached_object(real);
2579         if (co) {
2580                 if (oi->typep)
2581                         *(oi->typep) = co->type;
2582                 if (oi->sizep)
2583                         *(oi->sizep) = co->size;
2584                 if (oi->disk_sizep)
2585                         *(oi->disk_sizep) = 0;
2586                 if (oi->delta_base_sha1)
2587                         hashclr(oi->delta_base_sha1);
2588                 oi->whence = OI_CACHED;
2589                 return 0;
2590         }
2591
2592         if (!find_pack_entry(real, &e)) {
2593                 /* Most likely it's a loose object. */
2594                 if (!sha1_loose_object_info(real, oi)) {
2595                         oi->whence = OI_LOOSE;
2596                         return 0;
2597                 }
2598
2599                 /* Not a loose object; someone else may have just packed it. */
2600                 reprepare_packed_git();
2601                 if (!find_pack_entry(real, &e))
2602                         return -1;
2603         }
2604
2605         rtype = packed_object_info(e.p, e.offset, oi);
2606         if (rtype < 0) {
2607                 mark_bad_packed_object(e.p, real);
2608                 return sha1_object_info_extended(real, oi, 0);
2609         } else if (in_delta_base_cache(e.p, e.offset)) {
2610                 oi->whence = OI_DBCACHED;
2611         } else {
2612                 oi->whence = OI_PACKED;
2613                 oi->u.packed.offset = e.offset;
2614                 oi->u.packed.pack = e.p;
2615                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2616                                          rtype == OBJ_OFS_DELTA);
2617         }
2618
2619         return 0;
2620 }
2621
2622 /* returns enum object_type or negative */
2623 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2624 {
2625         enum object_type type;
2626         struct object_info oi = {NULL};
2627
2628         oi.typep = &type;
2629         oi.sizep = sizep;
2630         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2631                 return -1;
2632         return type;
2633 }
2634
2635 static void *read_packed_sha1(const unsigned char *sha1,
2636                               enum object_type *type, unsigned long *size)
2637 {
2638         struct pack_entry e;
2639         void *data;
2640
2641         if (!find_pack_entry(sha1, &e))
2642                 return NULL;
2643         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2644         if (!data) {
2645                 /*
2646                  * We're probably in deep shit, but let's try to fetch
2647                  * the required object anyway from another pack or loose.
2648                  * This should happen only in the presence of a corrupted
2649                  * pack, and is better than failing outright.
2650                  */
2651                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2652                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2653                 mark_bad_packed_object(e.p, sha1);
2654                 data = read_object(sha1, type, size);
2655         }
2656         return data;
2657 }
2658
2659 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2660                       unsigned char *sha1)
2661 {
2662         struct cached_object *co;
2663
2664         hash_sha1_file(buf, len, typename(type), sha1);
2665         if (has_sha1_file(sha1) || find_cached_object(sha1))
2666                 return 0;
2667         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2668         co = &cached_objects[cached_object_nr++];
2669         co->size = len;
2670         co->type = type;
2671         co->buf = xmalloc(len);
2672         memcpy(co->buf, buf, len);
2673         hashcpy(co->sha1, sha1);
2674         return 0;
2675 }
2676
2677 static void *read_object(const unsigned char *sha1, enum object_type *type,
2678                          unsigned long *size)
2679 {
2680         unsigned long mapsize;
2681         void *map, *buf;
2682         struct cached_object *co;
2683
2684         co = find_cached_object(sha1);
2685         if (co) {
2686                 *type = co->type;
2687                 *size = co->size;
2688                 return xmemdupz(co->buf, co->size);
2689         }
2690
2691         buf = read_packed_sha1(sha1, type, size);
2692         if (buf)
2693                 return buf;
2694         map = map_sha1_file(sha1, &mapsize);
2695         if (map) {
2696                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2697                 munmap(map, mapsize);
2698                 return buf;
2699         }
2700         reprepare_packed_git();
2701         return read_packed_sha1(sha1, type, size);
2702 }
2703
2704 /*
2705  * This function dies on corrupt objects; the callers who want to
2706  * deal with them should arrange to call read_object() and give error
2707  * messages themselves.
2708  */
2709 void *read_sha1_file_extended(const unsigned char *sha1,
2710                               enum object_type *type,
2711                               unsigned long *size,
2712                               unsigned flag)
2713 {
2714         void *data;
2715         const struct packed_git *p;
2716         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2717
2718         errno = 0;
2719         data = read_object(repl, type, size);
2720         if (data)
2721                 return data;
2722
2723         if (errno && errno != ENOENT)
2724                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2725
2726         /* die if we replaced an object with one that does not exist */
2727         if (repl != sha1)
2728                 die("replacement %s not found for %s",
2729                     sha1_to_hex(repl), sha1_to_hex(sha1));
2730
2731         if (has_loose_object(repl)) {
2732                 const char *path = sha1_file_name(sha1);
2733
2734                 die("loose object %s (stored in %s) is corrupt",
2735                     sha1_to_hex(repl), path);
2736         }
2737
2738         if ((p = has_packed_and_bad(repl)) != NULL)
2739                 die("packed object %s (stored in %s) is corrupt",
2740                     sha1_to_hex(repl), p->pack_name);
2741
2742         return NULL;
2743 }
2744
2745 void *read_object_with_reference(const unsigned char *sha1,
2746                                  const char *required_type_name,
2747                                  unsigned long *size,
2748                                  unsigned char *actual_sha1_return)
2749 {
2750         enum object_type type, required_type;
2751         void *buffer;
2752         unsigned long isize;
2753         unsigned char actual_sha1[20];
2754
2755         required_type = type_from_string(required_type_name);
2756         hashcpy(actual_sha1, sha1);
2757         while (1) {
2758                 int ref_length = -1;
2759                 const char *ref_type = NULL;
2760
2761                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2762                 if (!buffer)
2763                         return NULL;
2764                 if (type == required_type) {
2765                         *size = isize;
2766                         if (actual_sha1_return)
2767                                 hashcpy(actual_sha1_return, actual_sha1);
2768                         return buffer;
2769                 }
2770                 /* Handle references */
2771                 else if (type == OBJ_COMMIT)
2772                         ref_type = "tree ";
2773                 else if (type == OBJ_TAG)
2774                         ref_type = "object ";
2775                 else {
2776                         free(buffer);
2777                         return NULL;
2778                 }
2779                 ref_length = strlen(ref_type);
2780
2781                 if (ref_length + 40 > isize ||
2782                     memcmp(buffer, ref_type, ref_length) ||
2783                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2784                         free(buffer);
2785                         return NULL;
2786                 }
2787                 free(buffer);
2788                 /* Now we have the ID of the referred-to object in
2789                  * actual_sha1.  Check again. */
2790         }
2791 }
2792
2793 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2794                                     const char *type, unsigned char *sha1,
2795                                     char *hdr, int *hdrlen)
2796 {
2797         git_SHA_CTX c;
2798
2799         /* Generate the header */
2800         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2801
2802         /* Sha1.. */
2803         git_SHA1_Init(&c);
2804         git_SHA1_Update(&c, hdr, *hdrlen);
2805         git_SHA1_Update(&c, buf, len);
2806         git_SHA1_Final(sha1, &c);
2807 }
2808
2809 /*
2810  * Move the just written object into its final resting place.
2811  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2812  * "moving" is only a part of what it does, when no patch between
2813  * master to pu changes the call sites of this function.
2814  */
2815 int move_temp_to_file(const char *tmpfile, const char *filename)
2816 {
2817         int ret = 0;
2818
2819         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2820                 goto try_rename;
2821         else if (link(tmpfile, filename))
2822                 ret = errno;
2823
2824         /*
2825          * Coda hack - coda doesn't like cross-directory links,
2826          * so we fall back to a rename, which will mean that it
2827          * won't be able to check collisions, but that's not a
2828          * big deal.
2829          *
2830          * The same holds for FAT formatted media.
2831          *
2832          * When this succeeds, we just return.  We have nothing
2833          * left to unlink.
2834          */
2835         if (ret && ret != EEXIST) {
2836         try_rename:
2837                 if (!rename(tmpfile, filename))
2838                         goto out;
2839                 ret = errno;
2840         }
2841         unlink_or_warn(tmpfile);
2842         if (ret) {
2843                 if (ret != EEXIST) {
2844                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2845                 }
2846                 /* FIXME!!! Collision check here ? */
2847         }
2848
2849 out:
2850         if (adjust_shared_perm(filename))
2851                 return error("unable to set permission to '%s'", filename);
2852         return 0;
2853 }
2854
2855 static int write_buffer(int fd, const void *buf, size_t len)
2856 {
2857         if (write_in_full(fd, buf, len) < 0)
2858                 return error("file write error (%s)", strerror(errno));
2859         return 0;
2860 }
2861
2862 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2863                    unsigned char *sha1)
2864 {
2865         char hdr[32];
2866         int hdrlen;
2867         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2868         return 0;
2869 }
2870
2871 /* Finalize a file on disk, and close it. */
2872 static void close_sha1_file(int fd)
2873 {
2874         if (fsync_object_files)
2875                 fsync_or_die(fd, "sha1 file");
2876         if (close(fd) != 0)
2877                 die_errno("error when closing sha1 file");
2878 }
2879
2880 /* Size of directory component, including the ending '/' */
2881 static inline int directory_size(const char *filename)
2882 {
2883         const char *s = strrchr(filename, '/');
2884         if (!s)
2885                 return 0;
2886         return s - filename + 1;
2887 }
2888
2889 /*
2890  * This creates a temporary file in the same directory as the final
2891  * 'filename'
2892  *
2893  * We want to avoid cross-directory filename renames, because those
2894  * can have problems on various filesystems (FAT, NFS, Coda).
2895  */
2896 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2897 {
2898         int fd, dirlen = directory_size(filename);
2899
2900         if (dirlen + 20 > bufsiz) {
2901                 errno = ENAMETOOLONG;
2902                 return -1;
2903         }
2904         memcpy(buffer, filename, dirlen);
2905         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2906         fd = git_mkstemp_mode(buffer, 0444);
2907         if (fd < 0 && dirlen && errno == ENOENT) {
2908                 /* Make sure the directory exists */
2909                 memcpy(buffer, filename, dirlen);
2910                 buffer[dirlen-1] = 0;
2911                 if (mkdir(buffer, 0777) && errno != EEXIST)
2912                         return -1;
2913                 if (adjust_shared_perm(buffer))
2914                         return -1;
2915
2916                 /* Try again */
2917                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2918                 fd = git_mkstemp_mode(buffer, 0444);
2919         }
2920         return fd;
2921 }
2922
2923 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2924                               const void *buf, unsigned long len, time_t mtime)
2925 {
2926         int fd, ret;
2927         unsigned char compressed[4096];
2928         git_zstream stream;
2929         git_SHA_CTX c;
2930         unsigned char parano_sha1[20];
2931         static char tmp_file[PATH_MAX];
2932         const char *filename = sha1_file_name(sha1);
2933
2934         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
2935         if (fd < 0) {
2936                 if (errno == EACCES)
2937                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
2938                 else
2939                         return error("unable to create temporary file: %s", strerror(errno));
2940         }
2941
2942         /* Set it up */
2943         git_deflate_init(&stream, zlib_compression_level);
2944         stream.next_out = compressed;
2945         stream.avail_out = sizeof(compressed);
2946         git_SHA1_Init(&c);
2947
2948         /* First header.. */
2949         stream.next_in = (unsigned char *)hdr;
2950         stream.avail_in = hdrlen;
2951         while (git_deflate(&stream, 0) == Z_OK)
2952                 ; /* nothing */
2953         git_SHA1_Update(&c, hdr, hdrlen);
2954
2955         /* Then the data itself.. */
2956         stream.next_in = (void *)buf;
2957         stream.avail_in = len;
2958         do {
2959                 unsigned char *in0 = stream.next_in;
2960                 ret = git_deflate(&stream, Z_FINISH);
2961                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2962                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2963                         die("unable to write sha1 file");
2964                 stream.next_out = compressed;
2965                 stream.avail_out = sizeof(compressed);
2966         } while (ret == Z_OK);
2967
2968         if (ret != Z_STREAM_END)
2969                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2970         ret = git_deflate_end_gently(&stream);
2971         if (ret != Z_OK)
2972                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2973         git_SHA1_Final(parano_sha1, &c);
2974         if (hashcmp(sha1, parano_sha1) != 0)
2975                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2976
2977         close_sha1_file(fd);
2978
2979         if (mtime) {
2980                 struct utimbuf utb;
2981                 utb.actime = mtime;
2982                 utb.modtime = mtime;
2983                 if (utime(tmp_file, &utb) < 0)
2984                         warning("failed utime() on %s: %s",
2985                                 tmp_file, strerror(errno));
2986         }
2987
2988         return move_temp_to_file(tmp_file, filename);
2989 }
2990
2991 static int freshen_loose_object(const unsigned char *sha1)
2992 {
2993         return check_and_freshen(sha1, 1);
2994 }
2995
2996 static int freshen_packed_object(const unsigned char *sha1)
2997 {
2998         struct pack_entry e;
2999         return find_pack_entry(sha1, &e) && freshen_file(e.p->pack_name);
3000 }
3001
3002 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
3003 {
3004         unsigned char sha1[20];
3005         char hdr[32];
3006         int hdrlen;
3007
3008         /* Normally if we have it in the pack then we do not bother writing
3009          * it out into .git/objects/??/?{38} file.
3010          */
3011         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3012         if (returnsha1)
3013                 hashcpy(returnsha1, sha1);
3014         if (freshen_loose_object(sha1) || freshen_packed_object(sha1))
3015                 return 0;
3016         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3017 }
3018
3019 int force_object_loose(const unsigned char *sha1, time_t mtime)
3020 {
3021         void *buf;
3022         unsigned long len;
3023         enum object_type type;
3024         char hdr[32];
3025         int hdrlen;
3026         int ret;
3027
3028         if (has_loose_object(sha1))
3029                 return 0;
3030         buf = read_packed_sha1(sha1, &type, &len);
3031         if (!buf)
3032                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3033         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
3034         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3035         free(buf);
3036
3037         return ret;
3038 }
3039
3040 int has_pack_index(const unsigned char *sha1)
3041 {
3042         struct stat st;
3043         if (stat(sha1_pack_index_name(sha1), &st))
3044                 return 0;
3045         return 1;
3046 }
3047
3048 int has_sha1_pack(const unsigned char *sha1)
3049 {
3050         struct pack_entry e;
3051         return find_pack_entry(sha1, &e);
3052 }
3053
3054 int has_sha1_file(const unsigned char *sha1)
3055 {
3056         struct pack_entry e;
3057
3058         if (find_pack_entry(sha1, &e))
3059                 return 1;
3060         if (has_loose_object(sha1))
3061                 return 1;
3062         reprepare_packed_git();
3063         return find_pack_entry(sha1, &e);
3064 }
3065
3066 static void check_tree(const void *buf, size_t size)
3067 {
3068         struct tree_desc desc;
3069         struct name_entry entry;
3070
3071         init_tree_desc(&desc, buf, size);
3072         while (tree_entry(&desc, &entry))
3073                 /* do nothing
3074                  * tree_entry() will die() on malformed entries */
3075                 ;
3076 }
3077
3078 static void check_commit(const void *buf, size_t size)
3079 {
3080         struct commit c;
3081         memset(&c, 0, sizeof(c));
3082         if (parse_commit_buffer(&c, buf, size))
3083                 die("corrupt commit");
3084 }
3085
3086 static void check_tag(const void *buf, size_t size)
3087 {
3088         struct tag t;
3089         memset(&t, 0, sizeof(t));
3090         if (parse_tag_buffer(&t, buf, size))
3091                 die("corrupt tag");
3092 }
3093
3094 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3095                      enum object_type type,
3096                      const char *path, unsigned flags)
3097 {
3098         int ret, re_allocated = 0;
3099         int write_object = flags & HASH_WRITE_OBJECT;
3100
3101         if (!type)
3102                 type = OBJ_BLOB;
3103
3104         /*
3105          * Convert blobs to git internal format
3106          */
3107         if ((type == OBJ_BLOB) && path) {
3108                 struct strbuf nbuf = STRBUF_INIT;
3109                 if (convert_to_git(path, buf, size, &nbuf,
3110                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3111                         buf = strbuf_detach(&nbuf, &size);
3112                         re_allocated = 1;
3113                 }
3114         }
3115         if (flags & HASH_FORMAT_CHECK) {
3116                 if (type == OBJ_TREE)
3117                         check_tree(buf, size);
3118                 if (type == OBJ_COMMIT)
3119                         check_commit(buf, size);
3120                 if (type == OBJ_TAG)
3121                         check_tag(buf, size);
3122         }
3123
3124         if (write_object)
3125                 ret = write_sha1_file(buf, size, typename(type), sha1);
3126         else
3127                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3128         if (re_allocated)
3129                 free(buf);
3130         return ret;
3131 }
3132
3133 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3134                                      const char *path, unsigned flags)
3135 {
3136         int ret;
3137         const int write_object = flags & HASH_WRITE_OBJECT;
3138         struct strbuf sbuf = STRBUF_INIT;
3139
3140         assert(path);
3141         assert(would_convert_to_git_filter_fd(path));
3142
3143         convert_to_git_filter_fd(path, fd, &sbuf,
3144                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3145
3146         if (write_object)
3147                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3148                                       sha1);
3149         else
3150                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3151                                      sha1);
3152         strbuf_release(&sbuf);
3153         return ret;
3154 }
3155
3156 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3157                       const char *path, unsigned flags)
3158 {
3159         struct strbuf sbuf = STRBUF_INIT;
3160         int ret;
3161
3162         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3163                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3164         else
3165                 ret = -1;
3166         strbuf_release(&sbuf);
3167         return ret;
3168 }
3169
3170 #define SMALL_FILE_SIZE (32*1024)
3171
3172 static int index_core(unsigned char *sha1, int fd, size_t size,
3173                       enum object_type type, const char *path,
3174                       unsigned flags)
3175 {
3176         int ret;
3177
3178         if (!size) {
3179                 ret = index_mem(sha1, NULL, size, type, path, flags);
3180         } else if (size <= SMALL_FILE_SIZE) {
3181                 char *buf = xmalloc(size);
3182                 if (size == read_in_full(fd, buf, size))
3183                         ret = index_mem(sha1, buf, size, type, path, flags);
3184                 else
3185                         ret = error("short read %s", strerror(errno));
3186                 free(buf);
3187         } else {
3188                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3189                 ret = index_mem(sha1, buf, size, type, path, flags);
3190                 munmap(buf, size);
3191         }
3192         return ret;
3193 }
3194
3195 /*
3196  * This creates one packfile per large blob unless bulk-checkin
3197  * machinery is "plugged".
3198  *
3199  * This also bypasses the usual "convert-to-git" dance, and that is on
3200  * purpose. We could write a streaming version of the converting
3201  * functions and insert that before feeding the data to fast-import
3202  * (or equivalent in-core API described above). However, that is
3203  * somewhat complicated, as we do not know the size of the filter
3204  * result, which we need to know beforehand when writing a git object.
3205  * Since the primary motivation for trying to stream from the working
3206  * tree file and to avoid mmaping it in core is to deal with large
3207  * binary blobs, they generally do not want to get any conversion, and
3208  * callers should avoid this code path when filters are requested.
3209  */
3210 static int index_stream(unsigned char *sha1, int fd, size_t size,
3211                         enum object_type type, const char *path,
3212                         unsigned flags)
3213 {
3214         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3215 }
3216
3217 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3218              enum object_type type, const char *path, unsigned flags)
3219 {
3220         int ret;
3221
3222         /*
3223          * Call xsize_t() only when needed to avoid potentially unnecessary
3224          * die() for large files.
3225          */
3226         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3227                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3228         else if (!S_ISREG(st->st_mode))
3229                 ret = index_pipe(sha1, fd, type, path, flags);
3230         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3231                  (path && would_convert_to_git(path)))
3232                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3233                                  flags);
3234         else
3235                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3236                                    flags);
3237         close(fd);
3238         return ret;
3239 }
3240
3241 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3242 {
3243         int fd;
3244         struct strbuf sb = STRBUF_INIT;
3245
3246         switch (st->st_mode & S_IFMT) {
3247         case S_IFREG:
3248                 fd = open(path, O_RDONLY);
3249                 if (fd < 0)
3250                         return error("open(\"%s\"): %s", path,
3251                                      strerror(errno));
3252                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3253                         return error("%s: failed to insert into database",
3254                                      path);
3255                 break;
3256         case S_IFLNK:
3257                 if (strbuf_readlink(&sb, path, st->st_size)) {
3258                         char *errstr = strerror(errno);
3259                         return error("readlink(\"%s\"): %s", path,
3260                                      errstr);
3261                 }
3262                 if (!(flags & HASH_WRITE_OBJECT))
3263                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3264                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3265                         return error("%s: failed to insert into database",
3266                                      path);
3267                 strbuf_release(&sb);
3268                 break;
3269         case S_IFDIR:
3270                 return resolve_gitlink_ref(path, "HEAD", sha1);
3271         default:
3272                 return error("%s: unsupported file type", path);
3273         }
3274         return 0;
3275 }
3276
3277 int read_pack_header(int fd, struct pack_header *header)
3278 {
3279         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3280                 /* "eof before pack header was fully read" */
3281                 return PH_ERROR_EOF;
3282
3283         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3284                 /* "protocol error (pack signature mismatch detected)" */
3285                 return PH_ERROR_PACK_SIGNATURE;
3286         if (!pack_version_ok(header->hdr_version))
3287                 /* "protocol error (pack version unsupported)" */
3288                 return PH_ERROR_PROTOCOL;
3289         return 0;
3290 }
3291
3292 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3293 {
3294         enum object_type type = sha1_object_info(sha1, NULL);
3295         if (type < 0)
3296                 die("%s is not a valid object", sha1_to_hex(sha1));
3297         if (type != expect)
3298                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3299                     typename(expect));
3300 }
3301
3302 static int for_each_file_in_obj_subdir(int subdir_nr,
3303                                        struct strbuf *path,
3304                                        each_loose_object_fn obj_cb,
3305                                        each_loose_cruft_fn cruft_cb,
3306                                        each_loose_subdir_fn subdir_cb,
3307                                        void *data)
3308 {
3309         size_t baselen = path->len;
3310         DIR *dir = opendir(path->buf);
3311         struct dirent *de;
3312         int r = 0;
3313
3314         if (!dir) {
3315                 if (errno == ENOENT)
3316                         return 0;
3317                 return error("unable to open %s: %s", path->buf, strerror(errno));
3318         }
3319
3320         while ((de = readdir(dir))) {
3321                 if (is_dot_or_dotdot(de->d_name))
3322                         continue;
3323
3324                 strbuf_setlen(path, baselen);
3325                 strbuf_addf(path, "/%s", de->d_name);
3326
3327                 if (strlen(de->d_name) == 38)  {
3328                         char hex[41];
3329                         unsigned char sha1[20];
3330
3331                         snprintf(hex, sizeof(hex), "%02x%s",
3332                                  subdir_nr, de->d_name);
3333                         if (!get_sha1_hex(hex, sha1)) {
3334                                 if (obj_cb) {
3335                                         r = obj_cb(sha1, path->buf, data);
3336                                         if (r)
3337                                                 break;
3338                                 }
3339                                 continue;
3340                         }
3341                 }
3342
3343                 if (cruft_cb) {
3344                         r = cruft_cb(de->d_name, path->buf, data);
3345                         if (r)
3346                                 break;
3347                 }
3348         }
3349         strbuf_setlen(path, baselen);
3350
3351         if (!r && subdir_cb)
3352                 r = subdir_cb(subdir_nr, path->buf, data);
3353
3354         closedir(dir);
3355         return r;
3356 }
3357
3358 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3359                             each_loose_object_fn obj_cb,
3360                             each_loose_cruft_fn cruft_cb,
3361                             each_loose_subdir_fn subdir_cb,
3362                             void *data)
3363 {
3364         size_t baselen = path->len;
3365         int r = 0;
3366         int i;
3367
3368         for (i = 0; i < 256; i++) {
3369                 strbuf_addf(path, "/%02x", i);
3370                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3371                                                 subdir_cb, data);
3372                 strbuf_setlen(path, baselen);
3373                 if (r)
3374                         break;
3375         }
3376
3377         return r;
3378 }
3379
3380 int for_each_loose_file_in_objdir(const char *path,
3381                                   each_loose_object_fn obj_cb,
3382                                   each_loose_cruft_fn cruft_cb,
3383                                   each_loose_subdir_fn subdir_cb,
3384                                   void *data)
3385 {
3386         struct strbuf buf = STRBUF_INIT;
3387         int r;
3388
3389         strbuf_addstr(&buf, path);
3390         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3391                                               subdir_cb, data);
3392         strbuf_release(&buf);
3393
3394         return r;
3395 }
3396
3397 struct loose_alt_odb_data {
3398         each_loose_object_fn *cb;
3399         void *data;
3400 };
3401
3402 static int loose_from_alt_odb(struct alternate_object_database *alt,
3403                               void *vdata)
3404 {
3405         struct loose_alt_odb_data *data = vdata;
3406         struct strbuf buf = STRBUF_INIT;
3407         int r;
3408
3409         /* copy base not including trailing '/' */
3410         strbuf_add(&buf, alt->base, alt->name - alt->base - 1);
3411         r = for_each_loose_file_in_objdir_buf(&buf,
3412                                               data->cb, NULL, NULL,
3413                                               data->data);
3414         strbuf_release(&buf);
3415         return r;
3416 }
3417
3418 int for_each_loose_object(each_loose_object_fn cb, void *data)
3419 {
3420         struct loose_alt_odb_data alt;
3421         int r;
3422
3423         r = for_each_loose_file_in_objdir(get_object_directory(),
3424                                           cb, NULL, NULL, data);
3425         if (r)
3426                 return r;
3427
3428         alt.cb = cb;
3429         alt.data = data;
3430         return foreach_alt_odb(loose_from_alt_odb, &alt);
3431 }
3432
3433 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3434 {
3435         uint32_t i;
3436         int r = 0;
3437
3438         for (i = 0; i < p->num_objects; i++) {
3439                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3440
3441                 if (!sha1)
3442                         return error("unable to get sha1 of object %u in %s",
3443                                      i, p->pack_name);
3444
3445                 r = cb(sha1, p, i, data);
3446                 if (r)
3447                         break;
3448         }
3449         return r;
3450 }
3451
3452 int for_each_packed_object(each_packed_object_fn cb, void *data)
3453 {
3454         struct packed_git *p;
3455         int r = 0;
3456
3457         prepare_packed_git();
3458         for (p = packed_git; p; p = p->next) {
3459                 r = for_each_object_in_pack(p, cb, data);
3460                 if (r)
3461                         break;
3462         }
3463         return r;
3464 }