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 "delta.h"
  12 #include "pack.h"
  13 #include "blob.h"
  14 #include "commit.h"
  15 #include "run-command.h"
  16 #include "tag.h"
  17 #include "tree.h"
  18 #include "tree-walk.h"
  19 #include "refs.h"
  20 #include "pack-revindex.h"
  21 #include "sha1-lookup.h"
  22 #include "bulk-checkin.h"
  23 #include "streaming.h"
  24 #include "dir.h"
  25
  26 #ifndef O_NOATIME
  27 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
  28 #define O_NOATIME 01000000
  29 #else
  30 #define O_NOATIME 0
  31 #endif
  32 #endif
  33
  34 #define SZ_FMT PRIuMAX
  35 static inline uintmax_t sz_fmt(size_t s) { return s; }
  36
  37 const unsigned char null_sha1[20];
  38
  39 /*
  40  * This is meant to hold a *small* number of objects that you would
  41  * want read_sha1_file() to be able to return, but yet you do not want
  42  * to write them into the object store (e.g. a browse-only
  43  * application).
  44  */
  45 static struct cached_object {
  46         unsigned char sha1[20];
  47         enum object_type type;
  48         void *buf;
  49         unsigned long size;
  50 } *cached_objects;
  51 static int cached_object_nr, cached_object_alloc;
  52
  53 static struct cached_object empty_tree = {
  54         EMPTY_TREE_SHA1_BIN_LITERAL,
  55         OBJ_TREE,
  56         "",
  57         0
  58 };
  59
  60 static struct packed_git *last_found_pack;
  61
  62 static struct cached_object *find_cached_object(const unsigned char *sha1)
  63 {
  64         int i;
  65         struct cached_object *co = cached_objects;
  66
  67         for (i = 0; i < cached_object_nr; i++, co++) {
  68                 if (!hashcmp(co->sha1, sha1))
  69                         return co;
  70         }
  71         if (!hashcmp(sha1, empty_tree.sha1))
  72                 return &empty_tree;
  73         return NULL;
  74 }
  75
  76 int mkdir_in_gitdir(const char *path)
  77 {
  78         if (mkdir(path, 0777)) {
  79                 int saved_errno = errno;
  80                 struct stat st;
  81                 struct strbuf sb = STRBUF_INIT;
  82
  83                 if (errno != EEXIST)
  84                         return -1;
  85                 /*
  86                  * Are we looking at a path in a symlinked worktree
  87                  * whose original repository does not yet have it?
  88                  * e.g. .git/rr-cache pointing at its original
  89                  * repository in which the user hasn't performed any
  90                  * conflict resolution yet?
  91                  */
  92                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
  93                     strbuf_readlink(&sb, path, st.st_size) ||
  94                     !is_absolute_path(sb.buf) ||
  95                     mkdir(sb.buf, 0777)) {
  96                         strbuf_release(&sb);
  97                         errno = saved_errno;
  98                         return -1;
  99                 }
 100                 strbuf_release(&sb);
 101         }
 102         return adjust_shared_perm(path);
 103 }
 104
 105 int safe_create_leading_directories(char *path)
 106 {
 107         char *pos = path + offset_1st_component(path);
 108         struct stat st;
 109
 110         while (pos) {
 111                 pos = strchr(pos, '/');
 112                 if (!pos)
 113                         break;
 114                 while (*++pos == '/')
 115                         ;
 116                 if (!*pos)
 117                         break;
 118                 *--pos = '\0';
 119                 if (!stat(path, &st)) {
 120                         /* path exists */
 121                         if (!S_ISDIR(st.st_mode)) {
 122                                 *pos = '/';
 123                                 return -3;
 124                         }
 125                 }
 126                 else if (mkdir(path, 0777)) {
 127                         if (errno == EEXIST &&
 128                             !stat(path, &st) && S_ISDIR(st.st_mode)) {
 129                                 ; /* somebody created it since we checked */
 130                         } else {
 131                                 *pos = '/';
 132                                 return -1;
 133                         }
 134                 }
 135                 else if (adjust_shared_perm(path)) {
 136                         *pos = '/';
 137                         return -2;
 138                 }
 139                 *pos++ = '/';
 140         }
 141         return 0;
 142 }
 143
 144 int safe_create_leading_directories_const(const char *path)
 145 {
 146         /* path points to cache entries, so xstrdup before messing with it */
 147         char *buf = xstrdup(path);
 148         int result = safe_create_leading_directories(buf);
 149         free(buf);
 150         return result;
 151 }
 152
 153 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
 154 {
 155         int i;
 156         for (i = 0; i < 20; i++) {
 157                 static char hex[] = "0123456789abcdef";
 158                 unsigned int val = sha1[i];
 159                 char *pos = pathbuf + i*2 + (i > 0);
 160                 *pos++ = hex[val >> 4];
 161                 *pos = hex[val & 0xf];
 162         }
 163 }
 164
 165 /*
 166  * NOTE! This returns a statically allocated buffer, so you have to be
 167  * careful about using it. Do an "xstrdup()" if you need to save the
 168  * filename.
 169  *
 170  * Also note that this returns the location for creating.  Reading
 171  * SHA1 file can happen from any alternate directory listed in the
 172  * DB_ENVIRONMENT environment variable if it is not found in
 173  * the primary object database.
 174  */
 175 char *sha1_file_name(const unsigned char *sha1)
 176 {
 177         static char buf[PATH_MAX];
 178         const char *objdir;
 179         int len;
 180
 181         objdir = get_object_directory();
 182         len = strlen(objdir);
 183
 184         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
 185         if (len + 43 > PATH_MAX)
 186                 die("insanely long object directory %s", objdir);
 187         memcpy(buf, objdir, len);
 188         buf[len] = '/';
 189         buf[len+3] = '/';
 190         buf[len+42] = '\0';
 191         fill_sha1_path(buf + len + 1, sha1);
 192         return buf;
 193 }
 194
 195 static char *sha1_get_pack_name(const unsigned char *sha1,
 196                                 char **name, char **base, const char *which)
 197 {
 198         static const char hex[] = "0123456789abcdef";
 199         char *buf;
 200         int i;
 201
 202         if (!*base) {
 203                 const char *sha1_file_directory = get_object_directory();
 204                 int len = strlen(sha1_file_directory);
 205                 *base = xmalloc(len + 60);
 206                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 207                         sha1_file_directory, which);
 208                 *name = *base + len + 11;
 209         }
 210
 211         buf = *name;
 212
 213         for (i = 0; i < 20; i++) {
 214                 unsigned int val = *sha1++;
 215                 *buf++ = hex[val >> 4];
 216                 *buf++ = hex[val & 0xf];
 217         }
 218
 219         return *base;
 220 }
 221
 222 char *sha1_pack_name(const unsigned char *sha1)
 223 {
 224         static char *name, *base;
 225
 226         return sha1_get_pack_name(sha1, &name, &base, "pack");
 227 }
 228
 229 char *sha1_pack_index_name(const unsigned char *sha1)
 230 {
 231         static char *name, *base;
 232
 233         return sha1_get_pack_name(sha1, &name, &base, "idx");
 234 }
 235
 236 struct alternate_object_database *alt_odb_list;
 237 static struct alternate_object_database **alt_odb_tail;
 238
 239 static int git_open_noatime(const char *name);
 240
 241 /*
 242  * Prepare alternate object database registry.
 243  *
 244  * The variable alt_odb_list points at the list of struct
 245  * alternate_object_database.  The elements on this list come from
 246  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 247  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 248  * whose contents is similar to that environment variable but can be
 249  * LF separated.  Its base points at a statically allocated buffer that
 250  * contains "/the/directory/corresponding/to/.git/objects/...", while
 251  * its name points just after the slash at the end of ".git/objects/"
 252  * in the example above, and has enough space to hold 40-byte hex
 253  * SHA1, an extra slash for the first level indirection, and the
 254  * terminating NUL.
 255  */
 256 static int link_alt_odb_entry(const char *entry, const char *relative_base, int depth)
 257 {
 258         const char *objdir = get_object_directory();
 259         struct alternate_object_database *ent;
 260         struct alternate_object_database *alt;
 261         int pfxlen, entlen;
 262         struct strbuf pathbuf = STRBUF_INIT;
 263
 264         if (!is_absolute_path(entry) && relative_base) {
 265                 strbuf_addstr(&pathbuf, real_path(relative_base));
 266                 strbuf_addch(&pathbuf, '/');
 267         }
 268         strbuf_addstr(&pathbuf, entry);
 269
 270         normalize_path_copy(pathbuf.buf, pathbuf.buf);
 271
 272         pfxlen = strlen(pathbuf.buf);
 273
 274         /*
 275          * The trailing slash after the directory name is given by
 276          * this function at the end. Remove duplicates.
 277          */
 278         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
 279                 pfxlen -= 1;
 280
 281         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
 282         ent = xmalloc(sizeof(*ent) + entlen);
 283         memcpy(ent->base, pathbuf.buf, pfxlen);
 284         strbuf_release(&pathbuf);
 285
 286         ent->name = ent->base + pfxlen + 1;
 287         ent->base[pfxlen + 3] = '/';
 288         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 289
 290         /* Detect cases where alternate disappeared */
 291         if (!is_directory(ent->base)) {
 292                 error("object directory %s does not exist; "
 293                       "check .git/objects/info/alternates.",
 294                       ent->base);
 295                 free(ent);
 296                 return -1;
 297         }
 298
 299         /* Prevent the common mistake of listing the same
 300          * thing twice, or object directory itself.
 301          */
 302         for (alt = alt_odb_list; alt; alt = alt->next) {
 303                 if (!memcmp(ent->base, alt->base, pfxlen)) {
 304                         free(ent);
 305                         return -1;
 306                 }
 307         }
 308         if (!strcmp(ent->base, objdir)) {
 309                 free(ent);
 310                 return -1;
 311         }
 312
 313         /* add the alternate entry */
 314         *alt_odb_tail = ent;
 315         alt_odb_tail = &(ent->next);
 316         ent->next = NULL;
 317
 318         /* recursively add alternates */
 319         read_info_alternates(ent->base, depth + 1);
 320
 321         ent->base[pfxlen] = '/';
 322
 323         return 0;
 324 }
 325
 326 static void link_alt_odb_entries(const char *alt, int len, int sep,
 327                                  const char *relative_base, int depth)
 328 {
 329         struct string_list entries = STRING_LIST_INIT_NODUP;
 330         char *alt_copy;
 331         int i;
 332
 333         if (depth > 5) {
 334                 error("%s: ignoring alternate object stores, nesting too deep.",
 335                                 relative_base);
 336                 return;
 337         }
 338
 339         alt_copy = xmemdupz(alt, len);
 340         string_list_split_in_place(&entries, alt_copy, sep, -1);
 341         for (i = 0; i < entries.nr; i++) {
 342                 const char *entry = entries.items[i].string;
 343                 if (entry[0] == '\0' || entry[0] == '#')
 344                         continue;
 345                 if (!is_absolute_path(entry) && depth) {
 346                         error("%s: ignoring relative alternate object store %s",
 347                                         relative_base, entry);
 348                 } else {
 349                         link_alt_odb_entry(entry, relative_base, depth);
 350                 }
 351         }
 352         string_list_clear(&entries, 0);
 353         free(alt_copy);
 354 }
 355
 356 void read_info_alternates(const char * relative_base, int depth)
 357 {
 358         char *map;
 359         size_t mapsz;
 360         struct stat st;
 361         const char alt_file_name[] = "info/alternates";
 362         /* Given that relative_base is no longer than PATH_MAX,
 363            ensure that "path" has enough space to append "/", the
 364            file name, "info/alternates", and a trailing NUL.  */
 365         char path[PATH_MAX + 1 + sizeof alt_file_name];
 366         int fd;
 367
 368         sprintf(path, "%s/%s", relative_base, alt_file_name);
 369         fd = git_open_noatime(path);
 370         if (fd < 0)
 371                 return;
 372         if (fstat(fd, &st) || (st.st_size == 0)) {
 373                 close(fd);
 374                 return;
 375         }
 376         mapsz = xsize_t(st.st_size);
 377         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 378         close(fd);
 379
 380         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
 381
 382         munmap(map, mapsz);
 383 }
 384
 385 void add_to_alternates_file(const char *reference)
 386 {
 387         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 388         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
 389         char *alt = mkpath("%s\n", reference);
 390         write_or_die(fd, alt, strlen(alt));
 391         if (commit_lock_file(lock))
 392                 die("could not close alternates file");
 393         if (alt_odb_tail)
 394                 link_alt_odb_entries(alt, strlen(alt), '\n', NULL, 0);
 395 }
 396
 397 void foreach_alt_odb(alt_odb_fn fn, void *cb)
 398 {
 399         struct alternate_object_database *ent;
 400
 401         prepare_alt_odb();
 402         for (ent = alt_odb_list; ent; ent = ent->next)
 403                 if (fn(ent, cb))
 404                         return;
 405 }
 406
 407 void prepare_alt_odb(void)
 408 {
 409         const char *alt;
 410
 411         if (alt_odb_tail)
 412                 return;
 413
 414         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 415         if (!alt) alt = "";
 416
 417         alt_odb_tail = &alt_odb_list;
 418         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
 419
 420         read_info_alternates(get_object_directory(), 0);
 421 }
 422
 423 static int has_loose_object_local(const unsigned char *sha1)
 424 {
 425         char *name = sha1_file_name(sha1);
 426         return !access(name, F_OK);
 427 }
 428
 429 int has_loose_object_nonlocal(const unsigned char *sha1)
 430 {
 431         struct alternate_object_database *alt;
 432         prepare_alt_odb();
 433         for (alt = alt_odb_list; alt; alt = alt->next) {
 434                 fill_sha1_path(alt->name, sha1);
 435                 if (!access(alt->base, F_OK))
 436                         return 1;
 437         }
 438         return 0;
 439 }
 440
 441 static int has_loose_object(const unsigned char *sha1)
 442 {
 443         return has_loose_object_local(sha1) ||
 444                has_loose_object_nonlocal(sha1);
 445 }
 446
 447 static unsigned int pack_used_ctr;
 448 static unsigned int pack_mmap_calls;
 449 static unsigned int peak_pack_open_windows;
 450 static unsigned int pack_open_windows;
 451 static unsigned int pack_open_fds;
 452 static unsigned int pack_max_fds;
 453 static size_t peak_pack_mapped;
 454 static size_t pack_mapped;
 455 struct packed_git *packed_git;
 456
 457 void pack_report(void)
 458 {
 459         fprintf(stderr,
 460                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 461                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 462                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 463                 sz_fmt(getpagesize()),
 464                 sz_fmt(packed_git_window_size),
 465                 sz_fmt(packed_git_limit));
 466         fprintf(stderr,
 467                 "pack_report: pack_used_ctr            = %10u\n"
 468                 "pack_report: pack_mmap_calls          = %10u\n"
 469                 "pack_report: pack_open_windows        = %10u / %10u\n"
 470                 "pack_report: pack_mapped              = "
 471                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 472                 pack_used_ctr,
 473                 pack_mmap_calls,
 474                 pack_open_windows, peak_pack_open_windows,
 475                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 476 }
 477
 478 static int check_packed_git_idx(const char *path,  struct packed_git *p)
 479 {
 480         void *idx_map;
 481         struct pack_idx_header *hdr;
 482         size_t idx_size;
 483         uint32_t version, nr, i, *index;
 484         int fd = git_open_noatime(path);
 485         struct stat st;
 486
 487         if (fd < 0)
 488                 return -1;
 489         if (fstat(fd, &st)) {
 490                 close(fd);
 491                 return -1;
 492         }
 493         idx_size = xsize_t(st.st_size);
 494         if (idx_size < 4 * 256 + 20 + 20) {
 495                 close(fd);
 496                 return error("index file %s is too small", path);
 497         }
 498         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 499         close(fd);
 500
 501         hdr = idx_map;
 502         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 503                 version = ntohl(hdr->idx_version);
 504                 if (version < 2 || version > 2) {
 505                         munmap(idx_map, idx_size);
 506                         return error("index file %s is version %"PRIu32
 507                                      " and is not supported by this binary"
 508                                      " (try upgrading GIT to a newer version)",
 509                                      path, version);
 510                 }
 511         } else
 512                 version = 1;
 513
 514         nr = 0;
 515         index = idx_map;
 516         if (version > 1)
 517                 index += 2;  /* skip index header */
 518         for (i = 0; i < 256; i++) {
 519                 uint32_t n = ntohl(index[i]);
 520                 if (n < nr) {
 521                         munmap(idx_map, idx_size);
 522                         return error("non-monotonic index %s", path);
 523                 }
 524                 nr = n;
 525         }
 526
 527         if (version == 1) {
 528                 /*
 529                  * Total size:
 530                  *  - 256 index entries 4 bytes each
 531                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 532                  *  - 20-byte SHA1 of the packfile
 533                  *  - 20-byte SHA1 file checksum
 534                  */
 535                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 536                         munmap(idx_map, idx_size);
 537                         return error("wrong index v1 file size in %s", path);
 538                 }
 539         } else if (version == 2) {
 540                 /*
 541                  * Minimum size:
 542                  *  - 8 bytes of header
 543                  *  - 256 index entries 4 bytes each
 544                  *  - 20-byte sha1 entry * nr
 545                  *  - 4-byte crc entry * nr
 546                  *  - 4-byte offset entry * nr
 547                  *  - 20-byte SHA1 of the packfile
 548                  *  - 20-byte SHA1 file checksum
 549                  * And after the 4-byte offset table might be a
 550                  * variable sized table containing 8-byte entries
 551                  * for offsets larger than 2^31.
 552                  */
 553                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 554                 unsigned long max_size = min_size;
 555                 if (nr)
 556                         max_size += (nr - 1)*8;
 557                 if (idx_size < min_size || idx_size > max_size) {
 558                         munmap(idx_map, idx_size);
 559                         return error("wrong index v2 file size in %s", path);
 560                 }
 561                 if (idx_size != min_size &&
 562                     /*
 563                      * make sure we can deal with large pack offsets.
 564                      * 31-bit signed offset won't be enough, neither
 565                      * 32-bit unsigned one will be.
 566                      */
 567                     (sizeof(off_t) <= 4)) {
 568                         munmap(idx_map, idx_size);
 569                         return error("pack too large for current definition of off_t in %s", path);
 570                 }
 571         }
 572
 573         p->index_version = version;
 574         p->index_data = idx_map;
 575         p->index_size = idx_size;
 576         p->num_objects = nr;
 577         return 0;
 578 }
 579
 580 int open_pack_index(struct packed_git *p)
 581 {
 582         char *idx_name;
 583         int ret;
 584
 585         if (p->index_data)
 586                 return 0;
 587
 588         idx_name = xstrdup(p->pack_name);
 589         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 590         ret = check_packed_git_idx(idx_name, p);
 591         free(idx_name);
 592         return ret;
 593 }
 594
 595 static void scan_windows(struct packed_git *p,
 596         struct packed_git **lru_p,
 597         struct pack_window **lru_w,
 598         struct pack_window **lru_l)
 599 {
 600         struct pack_window *w, *w_l;
 601
 602         for (w_l = NULL, w = p->windows; w; w = w->next) {
 603                 if (!w->inuse_cnt) {
 604                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 605                                 *lru_p = p;
 606                                 *lru_w = w;
 607                                 *lru_l = w_l;
 608                         }
 609                 }
 610                 w_l = w;
 611         }
 612 }
 613
 614 static int unuse_one_window(struct packed_git *current, int keep_fd)
 615 {
 616         struct packed_git *p, *lru_p = NULL;
 617         struct pack_window *lru_w = NULL, *lru_l = NULL;
 618
 619         if (current)
 620                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 621         for (p = packed_git; p; p = p->next)
 622                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 623         if (lru_p) {
 624                 munmap(lru_w->base, lru_w->len);
 625                 pack_mapped -= lru_w->len;
 626                 if (lru_l)
 627                         lru_l->next = lru_w->next;
 628                 else {
 629                         lru_p->windows = lru_w->next;
 630                         if (!lru_p->windows && lru_p->pack_fd != -1
 631                                 && lru_p->pack_fd != keep_fd) {
 632                                 close(lru_p->pack_fd);
 633                                 pack_open_fds--;
 634                                 lru_p->pack_fd = -1;
 635                         }
 636                 }
 637                 free(lru_w);
 638                 pack_open_windows--;
 639                 return 1;
 640         }
 641         return 0;
 642 }
 643
 644 void release_pack_memory(size_t need, int fd)
 645 {
 646         size_t cur = pack_mapped;
 647         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
 648                 ; /* nothing */
 649 }
 650
 651 void *xmmap(void *start, size_t length,
 652         int prot, int flags, int fd, off_t offset)
 653 {
 654         void *ret = mmap(start, length, prot, flags, fd, offset);
 655         if (ret == MAP_FAILED) {
 656                 if (!length)
 657                         return NULL;
 658                 release_pack_memory(length, fd);
 659                 ret = mmap(start, length, prot, flags, fd, offset);
 660                 if (ret == MAP_FAILED)
 661                         die_errno("Out of memory? mmap failed");
 662         }
 663         return ret;
 664 }
 665
 666 void close_pack_windows(struct packed_git *p)
 667 {
 668         while (p->windows) {
 669                 struct pack_window *w = p->windows;
 670
 671                 if (w->inuse_cnt)
 672                         die("pack '%s' still has open windows to it",
 673                             p->pack_name);
 674                 munmap(w->base, w->len);
 675                 pack_mapped -= w->len;
 676                 pack_open_windows--;
 677                 p->windows = w->next;
 678                 free(w);
 679         }
 680 }
 681
 682 void unuse_pack(struct pack_window **w_cursor)
 683 {
 684         struct pack_window *w = *w_cursor;
 685         if (w) {
 686                 w->inuse_cnt--;
 687                 *w_cursor = NULL;
 688         }
 689 }
 690
 691 void close_pack_index(struct packed_git *p)
 692 {
 693         if (p->index_data) {
 694                 munmap((void *)p->index_data, p->index_size);
 695                 p->index_data = NULL;
 696         }
 697 }
 698
 699 /*
 700  * This is used by git-repack in case a newly created pack happens to
 701  * contain the same set of objects as an existing one.  In that case
 702  * the resulting file might be different even if its name would be the
 703  * same.  It is best to close any reference to the old pack before it is
 704  * replaced on disk.  Of course no index pointers nor windows for given pack
 705  * must subsist at this point.  If ever objects from this pack are requested
 706  * again, the new version of the pack will be reinitialized through
 707  * reprepare_packed_git().
 708  */
 709 void free_pack_by_name(const char *pack_name)
 710 {
 711         struct packed_git *p, **pp = &packed_git;
 712
 713         while (*pp) {
 714                 p = *pp;
 715                 if (strcmp(pack_name, p->pack_name) == 0) {
 716                         clear_delta_base_cache();
 717                         close_pack_windows(p);
 718                         if (p->pack_fd != -1) {
 719                                 close(p->pack_fd);
 720                                 pack_open_fds--;
 721                         }
 722                         close_pack_index(p);
 723                         free(p->bad_object_sha1);
 724                         *pp = p->next;
 725                         if (last_found_pack == p)
 726                                 last_found_pack = NULL;
 727                         free(p);
 728                         return;
 729                 }
 730                 pp = &p->next;
 731         }
 732 }
 733
 734 static unsigned int get_max_fd_limit(void)
 735 {
 736 #ifdef RLIMIT_NOFILE
 737         struct rlimit lim;
 738
 739         if (getrlimit(RLIMIT_NOFILE, &lim))
 740                 die_errno("cannot get RLIMIT_NOFILE");
 741
 742         return lim.rlim_cur;
 743 #elif defined(_SC_OPEN_MAX)
 744         return sysconf(_SC_OPEN_MAX);
 745 #elif defined(OPEN_MAX)
 746         return OPEN_MAX;
 747 #else
 748         return 1; /* see the caller ;-) */
 749 #endif
 750 }
 751
 752 /*
 753  * Do not call this directly as this leaks p->pack_fd on error return;
 754  * call open_packed_git() instead.
 755  */
 756 static int open_packed_git_1(struct packed_git *p)
 757 {
 758         struct stat st;
 759         struct pack_header hdr;
 760         unsigned char sha1[20];
 761         unsigned char *idx_sha1;
 762         long fd_flag;
 763
 764         if (!p->index_data && open_pack_index(p))
 765                 return error("packfile %s index unavailable", p->pack_name);
 766
 767         if (!pack_max_fds) {
 768                 unsigned int max_fds = get_max_fd_limit();
 769
 770                 /* Save 3 for stdin/stdout/stderr, 22 for work */
 771                 if (25 < max_fds)
 772                         pack_max_fds = max_fds - 25;
 773                 else
 774                         pack_max_fds = 1;
 775         }
 776
 777         while (pack_max_fds <= pack_open_fds && unuse_one_window(NULL, -1))
 778                 ; /* nothing */
 779
 780         p->pack_fd = git_open_noatime(p->pack_name);
 781         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 782                 return -1;
 783         pack_open_fds++;
 784
 785         /* If we created the struct before we had the pack we lack size. */
 786         if (!p->pack_size) {
 787                 if (!S_ISREG(st.st_mode))
 788                         return error("packfile %s not a regular file", p->pack_name);
 789                 p->pack_size = st.st_size;
 790         } else if (p->pack_size != st.st_size)
 791                 return error("packfile %s size changed", p->pack_name);
 792
 793         /* We leave these file descriptors open with sliding mmap;
 794          * there is no point keeping them open across exec(), though.
 795          */
 796         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 797         if (fd_flag < 0)
 798                 return error("cannot determine file descriptor flags");
 799         fd_flag |= FD_CLOEXEC;
 800         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 801                 return error("cannot set FD_CLOEXEC");
 802
 803         /* Verify we recognize this pack file format. */
 804         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 805                 return error("file %s is far too short to be a packfile", p->pack_name);
 806         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 807                 return error("file %s is not a GIT packfile", p->pack_name);
 808         if (!pack_version_ok(hdr.hdr_version))
 809                 return error("packfile %s is version %"PRIu32" and not"
 810                         " supported (try upgrading GIT to a newer version)",
 811                         p->pack_name, ntohl(hdr.hdr_version));
 812
 813         /* Verify the pack matches its index. */
 814         if (p->num_objects != ntohl(hdr.hdr_entries))
 815                 return error("packfile %s claims to have %"PRIu32" objects"
 816                              " while index indicates %"PRIu32" objects",
 817                              p->pack_name, ntohl(hdr.hdr_entries),
 818                              p->num_objects);
 819         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
 820                 return error("end of packfile %s is unavailable", p->pack_name);
 821         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
 822                 return error("packfile %s signature is unavailable", p->pack_name);
 823         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
 824         if (hashcmp(sha1, idx_sha1))
 825                 return error("packfile %s does not match index", p->pack_name);
 826         return 0;
 827 }
 828
 829 static int open_packed_git(struct packed_git *p)
 830 {
 831         if (!open_packed_git_1(p))
 832                 return 0;
 833         if (p->pack_fd != -1) {
 834                 close(p->pack_fd);
 835                 pack_open_fds--;
 836                 p->pack_fd = -1;
 837         }
 838         return -1;
 839 }
 840
 841 static int in_window(struct pack_window *win, off_t offset)
 842 {
 843         /* We must promise at least 20 bytes (one hash) after the
 844          * offset is available from this window, otherwise the offset
 845          * is not actually in this window and a different window (which
 846          * has that one hash excess) must be used.  This is to support
 847          * the object header and delta base parsing routines below.
 848          */
 849         off_t win_off = win->offset;
 850         return win_off <= offset
 851                 && (offset + 20) <= (win_off + win->len);
 852 }
 853
 854 unsigned char *use_pack(struct packed_git *p,
 855                 struct pack_window **w_cursor,
 856                 off_t offset,
 857                 unsigned long *left)
 858 {
 859         struct pack_window *win = *w_cursor;
 860
 861         /* Since packfiles end in a hash of their content and it's
 862          * pointless to ask for an offset into the middle of that
 863          * hash, and the in_window function above wouldn't match
 864          * don't allow an offset too close to the end of the file.
 865          */
 866         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
 867                 die("packfile %s cannot be accessed", p->pack_name);
 868         if (offset > (p->pack_size - 20))
 869                 die("offset beyond end of packfile (truncated pack?)");
 870
 871         if (!win || !in_window(win, offset)) {
 872                 if (win)
 873                         win->inuse_cnt--;
 874                 for (win = p->windows; win; win = win->next) {
 875                         if (in_window(win, offset))
 876                                 break;
 877                 }
 878                 if (!win) {
 879                         size_t window_align = packed_git_window_size / 2;
 880                         off_t len;
 881
 882                         if (p->pack_fd == -1 && open_packed_git(p))
 883                                 die("packfile %s cannot be accessed", p->pack_name);
 884
 885                         win = xcalloc(1, sizeof(*win));
 886                         win->offset = (offset / window_align) * window_align;
 887                         len = p->pack_size - win->offset;
 888                         if (len > packed_git_window_size)
 889                                 len = packed_git_window_size;
 890                         win->len = (size_t)len;
 891                         pack_mapped += win->len;
 892                         while (packed_git_limit < pack_mapped
 893                                 && unuse_one_window(p, p->pack_fd))
 894                                 ; /* nothing */
 895                         win->base = xmmap(NULL, win->len,
 896                                 PROT_READ, MAP_PRIVATE,
 897                                 p->pack_fd, win->offset);
 898                         if (win->base == MAP_FAILED)
 899                                 die("packfile %s cannot be mapped: %s",
 900                                         p->pack_name,
 901                                         strerror(errno));
 902                         if (!win->offset && win->len == p->pack_size
 903                                 && !p->do_not_close) {
 904                                 close(p->pack_fd);
 905                                 pack_open_fds--;
 906                                 p->pack_fd = -1;
 907                         }
 908                         pack_mmap_calls++;
 909                         pack_open_windows++;
 910                         if (pack_mapped > peak_pack_mapped)
 911                                 peak_pack_mapped = pack_mapped;
 912                         if (pack_open_windows > peak_pack_open_windows)
 913                                 peak_pack_open_windows = pack_open_windows;
 914                         win->next = p->windows;
 915                         p->windows = win;
 916                 }
 917         }
 918         if (win != *w_cursor) {
 919                 win->last_used = pack_used_ctr++;
 920                 win->inuse_cnt++;
 921                 *w_cursor = win;
 922         }
 923         offset -= win->offset;
 924         if (left)
 925                 *left = win->len - xsize_t(offset);
 926         return win->base + offset;
 927 }
 928
 929 static struct packed_git *alloc_packed_git(int extra)
 930 {
 931         struct packed_git *p = xmalloc(sizeof(*p) + extra);
 932         memset(p, 0, sizeof(*p));
 933         p->pack_fd = -1;
 934         return p;
 935 }
 936
 937 static void try_to_free_pack_memory(size_t size)
 938 {
 939         release_pack_memory(size, -1);
 940 }
 941
 942 struct packed_git *add_packed_git(const char *path, int path_len, int local)
 943 {
 944         static int have_set_try_to_free_routine;
 945         struct stat st;
 946         struct packed_git *p = alloc_packed_git(path_len + 2);
 947
 948         if (!have_set_try_to_free_routine) {
 949                 have_set_try_to_free_routine = 1;
 950                 set_try_to_free_routine(try_to_free_pack_memory);
 951         }
 952
 953         /*
 954          * Make sure a corresponding .pack file exists and that
 955          * the index looks sane.
 956          */
 957         path_len -= strlen(".idx");
 958         if (path_len < 1) {
 959                 free(p);
 960                 return NULL;
 961         }
 962         memcpy(p->pack_name, path, path_len);
 963
 964         strcpy(p->pack_name + path_len, ".keep");
 965         if (!access(p->pack_name, F_OK))
 966                 p->pack_keep = 1;
 967
 968         strcpy(p->pack_name + path_len, ".pack");
 969         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 970                 free(p);
 971                 return NULL;
 972         }
 973
 974         /* ok, it looks sane as far as we can check without
 975          * actually mapping the pack file.
 976          */
 977         p->pack_size = st.st_size;
 978         p->pack_local = local;
 979         p->mtime = st.st_mtime;
 980         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
 981                 hashclr(p->sha1);
 982         return p;
 983 }
 984
 985 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
 986 {
 987         const char *path = sha1_pack_name(sha1);
 988         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
 989
 990         strcpy(p->pack_name, path);
 991         hashcpy(p->sha1, sha1);
 992         if (check_packed_git_idx(idx_path, p)) {
 993                 free(p);
 994                 return NULL;
 995         }
 996
 997         return p;
 998 }
 999
1000 void install_packed_git(struct packed_git *pack)
1001 {
1002         if (pack->pack_fd != -1)
1003                 pack_open_fds++;
1004
1005         pack->next = packed_git;
1006         packed_git = pack;
1007 }
1008
1009 void (*report_garbage)(const char *desc, const char *path);
1010
1011 static void report_helper(const struct string_list *list,
1012                           int seen_bits, int first, int last)
1013 {
1014         const char *msg;
1015         switch (seen_bits) {
1016         case 0:
1017                 msg = "no corresponding .idx nor .pack";
1018                 break;
1019         case 1:
1020                 msg = "no corresponding .idx";
1021                 break;
1022         case 2:
1023                 msg = "no corresponding .pack";
1024                 break;
1025         default:
1026                 return;
1027         }
1028         for (; first < last; first++)
1029                 report_garbage(msg, list->items[first].string);
1030 }
1031
1032 static void report_pack_garbage(struct string_list *list)
1033 {
1034         int i, baselen = -1, first = 0, seen_bits = 0;
1035
1036         if (!report_garbage)
1037                 return;
1038
1039         sort_string_list(list);
1040
1041         for (i = 0; i < list->nr; i++) {
1042                 const char *path = list->items[i].string;
1043                 if (baselen != -1 &&
1044                     strncmp(path, list->items[first].string, baselen)) {
1045                         report_helper(list, seen_bits, first, i);
1046                         baselen = -1;
1047                         seen_bits = 0;
1048                 }
1049                 if (baselen == -1) {
1050                         const char *dot = strrchr(path, '.');
1051                         if (!dot) {
1052                                 report_garbage("garbage found", path);
1053                                 continue;
1054                         }
1055                         baselen = dot - path + 1;
1056                         first = i;
1057                 }
1058                 if (!strcmp(path + baselen, "pack"))
1059                         seen_bits |= 1;
1060                 else if (!strcmp(path + baselen, "idx"))
1061                         seen_bits |= 2;
1062         }
1063         report_helper(list, seen_bits, first, list->nr);
1064 }
1065
1066 static void prepare_packed_git_one(char *objdir, int local)
1067 {
1068         /* Ensure that this buffer is large enough so that we can
1069            append "/pack/" without clobbering the stack even if
1070            strlen(objdir) were PATH_MAX.  */
1071         char path[PATH_MAX + 1 + 4 + 1 + 1];
1072         int len;
1073         DIR *dir;
1074         struct dirent *de;
1075         struct string_list garbage = STRING_LIST_INIT_DUP;
1076
1077         sprintf(path, "%s/pack", objdir);
1078         len = strlen(path);
1079         dir = opendir(path);
1080         if (!dir) {
1081                 if (errno != ENOENT)
1082                         error("unable to open object pack directory: %s: %s",
1083                               path, strerror(errno));
1084                 return;
1085         }
1086         path[len++] = '/';
1087         while ((de = readdir(dir)) != NULL) {
1088                 int namelen = strlen(de->d_name);
1089                 struct packed_git *p;
1090
1091                 if (len + namelen + 1 > sizeof(path)) {
1092                         if (report_garbage) {
1093                                 struct strbuf sb = STRBUF_INIT;
1094                                 strbuf_addf(&sb, "%.*s/%s", len - 1, path, de->d_name);
1095                                 report_garbage("path too long", sb.buf);
1096                                 strbuf_release(&sb);
1097                         }
1098                         continue;
1099                 }
1100
1101                 if (is_dot_or_dotdot(de->d_name))
1102                         continue;
1103
1104                 strcpy(path + len, de->d_name);
1105
1106                 if (has_extension(de->d_name, ".idx")) {
1107                         /* Don't reopen a pack we already have. */
1108                         for (p = packed_git; p; p = p->next) {
1109                                 if (!memcmp(path, p->pack_name, len + namelen - 4))
1110                                         break;
1111                         }
1112                         if (p == NULL &&
1113                             /*
1114                              * See if it really is a valid .idx file with
1115                              * corresponding .pack file that we can map.
1116                              */
1117                             (p = add_packed_git(path, len + namelen, local)) != NULL)
1118                                 install_packed_git(p);
1119                 }
1120
1121                 if (!report_garbage)
1122                         continue;
1123
1124                 if (has_extension(de->d_name, ".idx") ||
1125                     has_extension(de->d_name, ".pack") ||
1126                     has_extension(de->d_name, ".keep"))
1127                         string_list_append(&garbage, path);
1128                 else
1129                         report_garbage("garbage found", path);
1130         }
1131         closedir(dir);
1132         report_pack_garbage(&garbage);
1133         string_list_clear(&garbage, 0);
1134 }
1135
1136 static int sort_pack(const void *a_, const void *b_)
1137 {
1138         struct packed_git *a = *((struct packed_git **)a_);
1139         struct packed_git *b = *((struct packed_git **)b_);
1140         int st;
1141
1142         /*
1143          * Local packs tend to contain objects specific to our
1144          * variant of the project than remote ones.  In addition,
1145          * remote ones could be on a network mounted filesystem.
1146          * Favor local ones for these reasons.
1147          */
1148         st = a->pack_local - b->pack_local;
1149         if (st)
1150                 return -st;
1151
1152         /*
1153          * Younger packs tend to contain more recent objects,
1154          * and more recent objects tend to get accessed more
1155          * often.
1156          */
1157         if (a->mtime < b->mtime)
1158                 return 1;
1159         else if (a->mtime == b->mtime)
1160                 return 0;
1161         return -1;
1162 }
1163
1164 static void rearrange_packed_git(void)
1165 {
1166         struct packed_git **ary, *p;
1167         int i, n;
1168
1169         for (n = 0, p = packed_git; p; p = p->next)
1170                 n++;
1171         if (n < 2)
1172                 return;
1173
1174         /* prepare an array of packed_git for easier sorting */
1175         ary = xcalloc(n, sizeof(struct packed_git *));
1176         for (n = 0, p = packed_git; p; p = p->next)
1177                 ary[n++] = p;
1178
1179         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1180
1181         /* link them back again */
1182         for (i = 0; i < n - 1; i++)
1183                 ary[i]->next = ary[i + 1];
1184         ary[n - 1]->next = NULL;
1185         packed_git = ary[0];
1186
1187         free(ary);
1188 }
1189
1190 static int prepare_packed_git_run_once = 0;
1191 void prepare_packed_git(void)
1192 {
1193         struct alternate_object_database *alt;
1194
1195         if (prepare_packed_git_run_once)
1196                 return;
1197         prepare_packed_git_one(get_object_directory(), 1);
1198         prepare_alt_odb();
1199         for (alt = alt_odb_list; alt; alt = alt->next) {
1200                 alt->name[-1] = 0;
1201                 prepare_packed_git_one(alt->base, 0);
1202                 alt->name[-1] = '/';
1203         }
1204         rearrange_packed_git();
1205         prepare_packed_git_run_once = 1;
1206 }
1207
1208 void reprepare_packed_git(void)
1209 {
1210         discard_revindex();
1211         prepare_packed_git_run_once = 0;
1212         prepare_packed_git();
1213 }
1214
1215 static void mark_bad_packed_object(struct packed_git *p,
1216                                    const unsigned char *sha1)
1217 {
1218         unsigned i;
1219         for (i = 0; i < p->num_bad_objects; i++)
1220                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1221                         return;
1222         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1223         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1224         p->num_bad_objects++;
1225 }
1226
1227 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1228 {
1229         struct packed_git *p;
1230         unsigned i;
1231
1232         for (p = packed_git; p; p = p->next)
1233                 for (i = 0; i < p->num_bad_objects; i++)
1234                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1235                                 return p;
1236         return NULL;
1237 }
1238
1239 /*
1240  * With an in-core object data in "map", rehash it to make sure the
1241  * object name actually matches "sha1" to detect object corruption.
1242  * With "map" == NULL, try reading the object named with "sha1" using
1243  * the streaming interface and rehash it to do the same.
1244  */
1245 int check_sha1_signature(const unsigned char *sha1, void *map,
1246                          unsigned long size, const char *type)
1247 {
1248         unsigned char real_sha1[20];
1249         enum object_type obj_type;
1250         struct git_istream *st;
1251         git_SHA_CTX c;
1252         char hdr[32];
1253         int hdrlen;
1254
1255         if (map) {
1256                 hash_sha1_file(map, size, type, real_sha1);
1257                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1258         }
1259
1260         st = open_istream(sha1, &obj_type, &size, NULL);
1261         if (!st)
1262                 return -1;
1263
1264         /* Generate the header */
1265         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1266
1267         /* Sha1.. */
1268         git_SHA1_Init(&c);
1269         git_SHA1_Update(&c, hdr, hdrlen);
1270         for (;;) {
1271                 char buf[1024 * 16];
1272                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1273
1274                 if (readlen < 0) {
1275                         close_istream(st);
1276                         return -1;
1277                 }
1278                 if (!readlen)
1279                         break;
1280                 git_SHA1_Update(&c, buf, readlen);
1281         }
1282         git_SHA1_Final(real_sha1, &c);
1283         close_istream(st);
1284         return hashcmp(sha1, real_sha1) ? -1 : 0;
1285 }
1286
1287 static int git_open_noatime(const char *name)
1288 {
1289         static int sha1_file_open_flag = O_NOATIME;
1290
1291         for (;;) {
1292                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1293                 if (fd >= 0)
1294                         return fd;
1295
1296                 /* Might the failure be due to O_NOATIME? */
1297                 if (errno != ENOENT && sha1_file_open_flag) {
1298                         sha1_file_open_flag = 0;
1299                         continue;
1300                 }
1301
1302                 return -1;
1303         }
1304 }
1305
1306 static int open_sha1_file(const unsigned char *sha1)
1307 {
1308         int fd;
1309         char *name = sha1_file_name(sha1);
1310         struct alternate_object_database *alt;
1311
1312         fd = git_open_noatime(name);
1313         if (fd >= 0)
1314                 return fd;
1315
1316         prepare_alt_odb();
1317         errno = ENOENT;
1318         for (alt = alt_odb_list; alt; alt = alt->next) {
1319                 name = alt->name;
1320                 fill_sha1_path(name, sha1);
1321                 fd = git_open_noatime(alt->base);
1322                 if (fd >= 0)
1323                         return fd;
1324         }
1325         return -1;
1326 }
1327
1328 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1329 {
1330         void *map;
1331         int fd;
1332
1333         fd = open_sha1_file(sha1);
1334         map = NULL;
1335         if (fd >= 0) {
1336                 struct stat st;
1337
1338                 if (!fstat(fd, &st)) {
1339                         *size = xsize_t(st.st_size);
1340                         if (!*size) {
1341                                 /* mmap() is forbidden on empty files */
1342                                 error("object file %s is empty", sha1_file_name(sha1));
1343                                 return NULL;
1344                         }
1345                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1346                 }
1347                 close(fd);
1348         }
1349         return map;
1350 }
1351
1352 /*
1353  * There used to be a second loose object header format which
1354  * was meant to mimic the in-pack format, allowing for direct
1355  * copy of the object data.  This format turned up not to be
1356  * really worth it and we no longer write loose objects in that
1357  * format.
1358  */
1359 static int experimental_loose_object(unsigned char *map)
1360 {
1361         unsigned int word;
1362
1363         /*
1364          * We must determine if the buffer contains the standard
1365          * zlib-deflated stream or the experimental format based
1366          * on the in-pack object format. Compare the header byte
1367          * for each format:
1368          *
1369          * RFC1950 zlib w/ deflate : 0www1000 : 0 <= www <= 7
1370          * Experimental pack-based : Stttssss : ttt = 1,2,3,4
1371          *
1372          * If bit 7 is clear and bits 0-3 equal 8, the buffer MUST be
1373          * in standard loose-object format, UNLESS it is a Git-pack
1374          * format object *exactly* 8 bytes in size when inflated.
1375          *
1376          * However, RFC1950 also specifies that the 1st 16-bit word
1377          * must be divisible by 31 - this checksum tells us our buffer
1378          * is in the standard format, giving a false positive only if
1379          * the 1st word of the Git-pack format object happens to be
1380          * divisible by 31, ie:
1381          *      ((byte0 * 256) + byte1) % 31 = 0
1382          *   =>        0ttt10000www1000 % 31 = 0
1383          *
1384          * As it happens, this case can only arise for www=3 & ttt=1
1385          * - ie, a Commit object, which would have to be 8 bytes in
1386          * size. As no Commit can be that small, we find that the
1387          * combination of these two criteria (bitmask & checksum)
1388          * can always correctly determine the buffer format.
1389          */
1390         word = (map[0] << 8) + map[1];
1391         if ((map[0] & 0x8F) == 0x08 && !(word % 31))
1392                 return 0;
1393         else
1394                 return 1;
1395 }
1396
1397 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1398                 unsigned long len, enum object_type *type, unsigned long *sizep)
1399 {
1400         unsigned shift;
1401         unsigned long size, c;
1402         unsigned long used = 0;
1403
1404         c = buf[used++];
1405         *type = (c >> 4) & 7;
1406         size = c & 15;
1407         shift = 4;
1408         while (c & 0x80) {
1409                 if (len <= used || bitsizeof(long) <= shift) {
1410                         error("bad object header");
1411                         size = used = 0;
1412                         break;
1413                 }
1414                 c = buf[used++];
1415                 size += (c & 0x7f) << shift;
1416                 shift += 7;
1417         }
1418         *sizep = size;
1419         return used;
1420 }
1421
1422 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1423 {
1424         unsigned long size, used;
1425         static const char valid_loose_object_type[8] = {
1426                 0, /* OBJ_EXT */
1427                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1428                 0, /* "delta" and others are invalid in a loose object */
1429         };
1430         enum object_type type;
1431
1432         /* Get the data stream */
1433         memset(stream, 0, sizeof(*stream));
1434         stream->next_in = map;
1435         stream->avail_in = mapsize;
1436         stream->next_out = buffer;
1437         stream->avail_out = bufsiz;
1438
1439         if (experimental_loose_object(map)) {
1440                 /*
1441                  * The old experimental format we no longer produce;
1442                  * we can still read it.
1443                  */
1444                 used = unpack_object_header_buffer(map, mapsize, &type, &size);
1445                 if (!used || !valid_loose_object_type[type])
1446                         return -1;
1447                 map += used;
1448                 mapsize -= used;
1449
1450                 /* Set up the stream for the rest.. */
1451                 stream->next_in = map;
1452                 stream->avail_in = mapsize;
1453                 git_inflate_init(stream);
1454
1455                 /* And generate the fake traditional header */
1456                 stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1457                                                  typename(type), size);
1458                 return 0;
1459         }
1460         git_inflate_init(stream);
1461         return git_inflate(stream, 0);
1462 }
1463
1464 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1465 {
1466         int bytes = strlen(buffer) + 1;
1467         unsigned char *buf = xmallocz(size);
1468         unsigned long n;
1469         int status = Z_OK;
1470
1471         n = stream->total_out - bytes;
1472         if (n > size)
1473                 n = size;
1474         memcpy(buf, (char *) buffer + bytes, n);
1475         bytes = n;
1476         if (bytes <= size) {
1477                 /*
1478                  * The above condition must be (bytes <= size), not
1479                  * (bytes < size).  In other words, even though we
1480                  * expect no more output and set avail_out to zero,
1481                  * the input zlib stream may have bytes that express
1482                  * "this concludes the stream", and we *do* want to
1483                  * eat that input.
1484                  *
1485                  * Otherwise we would not be able to test that we
1486                  * consumed all the input to reach the expected size;
1487                  * we also want to check that zlib tells us that all
1488                  * went well with status == Z_STREAM_END at the end.
1489                  */
1490                 stream->next_out = buf + bytes;
1491                 stream->avail_out = size - bytes;
1492                 while (status == Z_OK)
1493                         status = git_inflate(stream, Z_FINISH);
1494         }
1495         if (status == Z_STREAM_END && !stream->avail_in) {
1496                 git_inflate_end(stream);
1497                 return buf;
1498         }
1499
1500         if (status < 0)
1501                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1502         else if (stream->avail_in)
1503                 error("garbage at end of loose object '%s'",
1504                       sha1_to_hex(sha1));
1505         free(buf);
1506         return NULL;
1507 }
1508
1509 /*
1510  * We used to just use "sscanf()", but that's actually way
1511  * too permissive for what we want to check. So do an anal
1512  * object header parse by hand.
1513  */
1514 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1515 {
1516         char type[10];
1517         int i;
1518         unsigned long size;
1519
1520         /*
1521          * The type can be at most ten bytes (including the
1522          * terminating '\0' that we add), and is followed by
1523          * a space.
1524          */
1525         i = 0;
1526         for (;;) {
1527                 char c = *hdr++;
1528                 if (c == ' ')
1529                         break;
1530                 type[i++] = c;
1531                 if (i >= sizeof(type))
1532                         return -1;
1533         }
1534         type[i] = 0;
1535
1536         /*
1537          * The length must follow immediately, and be in canonical
1538          * decimal format (ie "010" is not valid).
1539          */
1540         size = *hdr++ - '0';
1541         if (size > 9)
1542                 return -1;
1543         if (size) {
1544                 for (;;) {
1545                         unsigned long c = *hdr - '0';
1546                         if (c > 9)
1547                                 break;
1548                         hdr++;
1549                         size = size * 10 + c;
1550                 }
1551         }
1552         *sizep = size;
1553
1554         /*
1555          * The length must be followed by a zero byte
1556          */
1557         return *hdr ? -1 : type_from_string(type);
1558 }
1559
1560 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1561 {
1562         int ret;
1563         git_zstream stream;
1564         char hdr[8192];
1565
1566         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1567         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1568                 return NULL;
1569
1570         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1571 }
1572
1573 unsigned long get_size_from_delta(struct packed_git *p,
1574                                   struct pack_window **w_curs,
1575                                   off_t curpos)
1576 {
1577         const unsigned char *data;
1578         unsigned char delta_head[20], *in;
1579         git_zstream stream;
1580         int st;
1581
1582         memset(&stream, 0, sizeof(stream));
1583         stream.next_out = delta_head;
1584         stream.avail_out = sizeof(delta_head);
1585
1586         git_inflate_init(&stream);
1587         do {
1588                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1589                 stream.next_in = in;
1590                 st = git_inflate(&stream, Z_FINISH);
1591                 curpos += stream.next_in - in;
1592         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1593                  stream.total_out < sizeof(delta_head));
1594         git_inflate_end(&stream);
1595         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1596                 error("delta data unpack-initial failed");
1597                 return 0;
1598         }
1599
1600         /* Examine the initial part of the delta to figure out
1601          * the result size.
1602          */
1603         data = delta_head;
1604
1605         /* ignore base size */
1606         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1607
1608         /* Read the result size */
1609         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1610 }
1611
1612 static off_t get_delta_base(struct packed_git *p,
1613                                     struct pack_window **w_curs,
1614                                     off_t *curpos,
1615                                     enum object_type type,
1616                                     off_t delta_obj_offset)
1617 {
1618         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1619         off_t base_offset;
1620
1621         /* use_pack() assured us we have [base_info, base_info + 20)
1622          * as a range that we can look at without walking off the
1623          * end of the mapped window.  Its actually the hash size
1624          * that is assured.  An OFS_DELTA longer than the hash size
1625          * is stupid, as then a REF_DELTA would be smaller to store.
1626          */
1627         if (type == OBJ_OFS_DELTA) {
1628                 unsigned used = 0;
1629                 unsigned char c = base_info[used++];
1630                 base_offset = c & 127;
1631                 while (c & 128) {
1632                         base_offset += 1;
1633                         if (!base_offset || MSB(base_offset, 7))
1634                                 return 0;  /* overflow */
1635                         c = base_info[used++];
1636                         base_offset = (base_offset << 7) + (c & 127);
1637                 }
1638                 base_offset = delta_obj_offset - base_offset;
1639                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1640                         return 0;  /* out of bound */
1641                 *curpos += used;
1642         } else if (type == OBJ_REF_DELTA) {
1643                 /* The base entry _must_ be in the same pack */
1644                 base_offset = find_pack_entry_one(base_info, p);
1645                 *curpos += 20;
1646         } else
1647                 die("I am totally screwed");
1648         return base_offset;
1649 }
1650
1651 int unpack_object_header(struct packed_git *p,
1652                          struct pack_window **w_curs,
1653                          off_t *curpos,
1654                          unsigned long *sizep)
1655 {
1656         unsigned char *base;
1657         unsigned long left;
1658         unsigned long used;
1659         enum object_type type;
1660
1661         /* use_pack() assures us we have [base, base + 20) available
1662          * as a range that we can look at.  (Its actually the hash
1663          * size that is assured.)  With our object header encoding
1664          * the maximum deflated object size is 2^137, which is just
1665          * insane, so we know won't exceed what we have been given.
1666          */
1667         base = use_pack(p, w_curs, *curpos, &left);
1668         used = unpack_object_header_buffer(base, left, &type, sizep);
1669         if (!used) {
1670                 type = OBJ_BAD;
1671         } else
1672                 *curpos += used;
1673
1674         return type;
1675 }
1676
1677 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1678 {
1679         int type;
1680         struct revindex_entry *revidx;
1681         const unsigned char *sha1;
1682         revidx = find_pack_revindex(p, obj_offset);
1683         if (!revidx)
1684                 return OBJ_BAD;
1685         sha1 = nth_packed_object_sha1(p, revidx->nr);
1686         mark_bad_packed_object(p, sha1);
1687         type = sha1_object_info(sha1, NULL);
1688         if (type <= OBJ_NONE)
1689                 return OBJ_BAD;
1690         return type;
1691 }
1692
1693
1694 #define POI_STACK_PREALLOC 64
1695
1696 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1697                               unsigned long *sizep, int *rtype,
1698                               unsigned long *disk_sizep)
1699 {
1700         struct pack_window *w_curs = NULL;
1701         unsigned long size;
1702         off_t curpos = obj_offset;
1703         enum object_type type;
1704         off_t small_poi_stack[POI_STACK_PREALLOC];
1705         off_t *poi_stack = small_poi_stack;
1706         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1707
1708         type = unpack_object_header(p, &w_curs, &curpos, &size);
1709
1710         if (rtype)
1711                 *rtype = type; /* representation type */
1712
1713         if (sizep) {
1714                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1715                         off_t tmp_pos = curpos;
1716                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1717                                                            type, obj_offset);
1718                         if (!base_offset) {
1719                                 type = OBJ_BAD;
1720                                 goto out;
1721                         }
1722                         *sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1723                         if (*sizep == 0) {
1724                                 type = OBJ_BAD;
1725                                 goto out;
1726                         }
1727                 } else {
1728                         *sizep = size;
1729                 }
1730         }
1731
1732         if (disk_sizep) {
1733                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1734                 *disk_sizep = revidx[1].offset - obj_offset;
1735         }
1736
1737         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1738                 off_t base_offset;
1739                 /* Push the object we're going to leave behind */
1740                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1741                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1742                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1743                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1744                 } else {
1745                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1746                 }
1747                 poi_stack[poi_stack_nr++] = obj_offset;
1748                 /* If parsing the base offset fails, just unwind */
1749                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1750                 if (!base_offset)
1751                         goto unwind;
1752                 curpos = obj_offset = base_offset;
1753                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1754                 if (type <= OBJ_NONE) {
1755                         /* If getting the base itself fails, we first
1756                          * retry the base, otherwise unwind */
1757                         type = retry_bad_packed_offset(p, base_offset);
1758                         if (type > OBJ_NONE)
1759                                 goto out;
1760                         goto unwind;
1761                 }
1762         }
1763
1764         switch (type) {
1765         case OBJ_BAD:
1766         case OBJ_COMMIT:
1767         case OBJ_TREE:
1768         case OBJ_BLOB:
1769         case OBJ_TAG:
1770                 break;
1771         default:
1772                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1773                       type, (uintmax_t)obj_offset, p->pack_name);
1774                 type = OBJ_BAD;
1775         }
1776
1777 out:
1778         if (poi_stack != small_poi_stack)
1779                 free(poi_stack);
1780         unuse_pack(&w_curs);
1781         return type;
1782
1783 unwind:
1784         while (poi_stack_nr) {
1785                 obj_offset = poi_stack[--poi_stack_nr];
1786                 type = retry_bad_packed_offset(p, obj_offset);
1787                 if (type > OBJ_NONE)
1788                         goto out;
1789         }
1790         type = OBJ_BAD;
1791         goto out;
1792 }
1793
1794 static void *unpack_compressed_entry(struct packed_git *p,
1795                                     struct pack_window **w_curs,
1796                                     off_t curpos,
1797                                     unsigned long size)
1798 {
1799         int st;
1800         git_zstream stream;
1801         unsigned char *buffer, *in;
1802
1803         buffer = xmallocz(size);
1804         memset(&stream, 0, sizeof(stream));
1805         stream.next_out = buffer;
1806         stream.avail_out = size + 1;
1807
1808         git_inflate_init(&stream);
1809         do {
1810                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1811                 stream.next_in = in;
1812                 st = git_inflate(&stream, Z_FINISH);
1813                 if (!stream.avail_out)
1814                         break; /* the payload is larger than it should be */
1815                 curpos += stream.next_in - in;
1816         } while (st == Z_OK || st == Z_BUF_ERROR);
1817         git_inflate_end(&stream);
1818         if ((st != Z_STREAM_END) || stream.total_out != size) {
1819                 free(buffer);
1820                 return NULL;
1821         }
1822
1823         return buffer;
1824 }
1825
1826 #define MAX_DELTA_CACHE (256)
1827
1828 static size_t delta_base_cached;
1829
1830 static struct delta_base_cache_lru_list {
1831         struct delta_base_cache_lru_list *prev;
1832         struct delta_base_cache_lru_list *next;
1833 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1834
1835 static struct delta_base_cache_entry {
1836         struct delta_base_cache_lru_list lru;
1837         void *data;
1838         struct packed_git *p;
1839         off_t base_offset;
1840         unsigned long size;
1841         enum object_type type;
1842 } delta_base_cache[MAX_DELTA_CACHE];
1843
1844 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1845 {
1846         unsigned long hash;
1847
1848         hash = (unsigned long)p + (unsigned long)base_offset;
1849         hash += (hash >> 8) + (hash >> 16);
1850         return hash % MAX_DELTA_CACHE;
1851 }
1852
1853 static struct delta_base_cache_entry *
1854 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1855 {
1856         unsigned long hash = pack_entry_hash(p, base_offset);
1857         return delta_base_cache + hash;
1858 }
1859
1860 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
1861                                      struct packed_git *p, off_t base_offset)
1862 {
1863         return (ent->data && ent->p == p && ent->base_offset == base_offset);
1864 }
1865
1866 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1867 {
1868         struct delta_base_cache_entry *ent;
1869         ent = get_delta_base_cache_entry(p, base_offset);
1870         return eq_delta_base_cache_entry(ent, p, base_offset);
1871 }
1872
1873 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1874 {
1875         ent->data = NULL;
1876         ent->lru.next->prev = ent->lru.prev;
1877         ent->lru.prev->next = ent->lru.next;
1878         delta_base_cached -= ent->size;
1879 }
1880
1881 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1882         unsigned long *base_size, enum object_type *type, int keep_cache)
1883 {
1884         struct delta_base_cache_entry *ent;
1885         void *ret;
1886
1887         ent = get_delta_base_cache_entry(p, base_offset);
1888
1889         if (!eq_delta_base_cache_entry(ent, p, base_offset))
1890                 return unpack_entry(p, base_offset, type, base_size);
1891
1892         ret = ent->data;
1893
1894         if (!keep_cache)
1895                 clear_delta_base_cache_entry(ent);
1896         else
1897                 ret = xmemdupz(ent->data, ent->size);
1898         *type = ent->type;
1899         *base_size = ent->size;
1900         return ret;
1901 }
1902
1903 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1904 {
1905         if (ent->data) {
1906                 free(ent->data);
1907                 ent->data = NULL;
1908                 ent->lru.next->prev = ent->lru.prev;
1909                 ent->lru.prev->next = ent->lru.next;
1910                 delta_base_cached -= ent->size;
1911         }
1912 }
1913
1914 void clear_delta_base_cache(void)
1915 {
1916         unsigned long p;
1917         for (p = 0; p < MAX_DELTA_CACHE; p++)
1918                 release_delta_base_cache(&delta_base_cache[p]);
1919 }
1920
1921 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1922         void *base, unsigned long base_size, enum object_type type)
1923 {
1924         unsigned long hash = pack_entry_hash(p, base_offset);
1925         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1926         struct delta_base_cache_lru_list *lru;
1927
1928         release_delta_base_cache(ent);
1929         delta_base_cached += base_size;
1930
1931         for (lru = delta_base_cache_lru.next;
1932              delta_base_cached > delta_base_cache_limit
1933              && lru != &delta_base_cache_lru;
1934              lru = lru->next) {
1935                 struct delta_base_cache_entry *f = (void *)lru;
1936                 if (f->type == OBJ_BLOB)
1937                         release_delta_base_cache(f);
1938         }
1939         for (lru = delta_base_cache_lru.next;
1940              delta_base_cached > delta_base_cache_limit
1941              && lru != &delta_base_cache_lru;
1942              lru = lru->next) {
1943                 struct delta_base_cache_entry *f = (void *)lru;
1944                 release_delta_base_cache(f);
1945         }
1946
1947         ent->p = p;
1948         ent->base_offset = base_offset;
1949         ent->type = type;
1950         ent->data = base;
1951         ent->size = base_size;
1952         ent->lru.next = &delta_base_cache_lru;
1953         ent->lru.prev = delta_base_cache_lru.prev;
1954         delta_base_cache_lru.prev->next = &ent->lru;
1955         delta_base_cache_lru.prev = &ent->lru;
1956 }
1957
1958 static void *read_object(const unsigned char *sha1, enum object_type *type,
1959                          unsigned long *size);
1960
1961 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1962 {
1963         static FILE *log_file;
1964
1965         if (!log_file) {
1966                 log_file = fopen(log_pack_access, "w");
1967                 if (!log_file) {
1968                         error("cannot open pack access log '%s' for writing: %s",
1969                               log_pack_access, strerror(errno));
1970                         log_pack_access = NULL;
1971                         return;
1972                 }
1973         }
1974         fprintf(log_file, "%s %"PRIuMAX"\n",
1975                 p->pack_name, (uintmax_t)obj_offset);
1976         fflush(log_file);
1977 }
1978
1979 int do_check_packed_object_crc;
1980
1981 #define UNPACK_ENTRY_STACK_PREALLOC 64
1982 struct unpack_entry_stack_ent {
1983         off_t obj_offset;
1984         off_t curpos;
1985         unsigned long size;
1986 };
1987
1988 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1989                    enum object_type *final_type, unsigned long *final_size)
1990 {
1991         struct pack_window *w_curs = NULL;
1992         off_t curpos = obj_offset;
1993         void *data = NULL;
1994         unsigned long size;
1995         enum object_type type;
1996         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1997         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1998         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1999         int base_from_cache = 0;
2000
2001         if (log_pack_access)
2002                 write_pack_access_log(p, obj_offset);
2003
2004         /* PHASE 1: drill down to the innermost base object */
2005         for (;;) {
2006                 off_t base_offset;
2007                 int i;
2008                 struct delta_base_cache_entry *ent;
2009
2010                 if (do_check_packed_object_crc && p->index_version > 1) {
2011                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2012                         unsigned long len = revidx[1].offset - obj_offset;
2013                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2014                                 const unsigned char *sha1 =
2015                                         nth_packed_object_sha1(p, revidx->nr);
2016                                 error("bad packed object CRC for %s",
2017                                       sha1_to_hex(sha1));
2018                                 mark_bad_packed_object(p, sha1);
2019                                 unuse_pack(&w_curs);
2020                                 return NULL;
2021                         }
2022                 }
2023
2024                 ent = get_delta_base_cache_entry(p, curpos);
2025                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2026                         type = ent->type;
2027                         data = ent->data;
2028                         size = ent->size;
2029                         clear_delta_base_cache_entry(ent);
2030                         base_from_cache = 1;
2031                         break;
2032                 }
2033
2034                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2035                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2036                         break;
2037
2038                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2039                 if (!base_offset) {
2040                         error("failed to validate delta base reference "
2041                               "at offset %"PRIuMAX" from %s",
2042                               (uintmax_t)curpos, p->pack_name);
2043                         /* bail to phase 2, in hopes of recovery */
2044                         data = NULL;
2045                         break;
2046                 }
2047
2048                 /* push object, proceed to base */
2049                 if (delta_stack_nr >= delta_stack_alloc
2050                     && delta_stack == small_delta_stack) {
2051                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2052                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2053                         memcpy(delta_stack, small_delta_stack,
2054                                sizeof(*delta_stack)*delta_stack_nr);
2055                 } else {
2056                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2057                 }
2058                 i = delta_stack_nr++;
2059                 delta_stack[i].obj_offset = obj_offset;
2060                 delta_stack[i].curpos = curpos;
2061                 delta_stack[i].size = size;
2062
2063                 curpos = obj_offset = base_offset;
2064         }
2065
2066         /* PHASE 2: handle the base */
2067         switch (type) {
2068         case OBJ_OFS_DELTA:
2069         case OBJ_REF_DELTA:
2070                 if (data)
2071                         die("BUG in unpack_entry: left loop at a valid delta");
2072                 break;
2073         case OBJ_COMMIT:
2074         case OBJ_TREE:
2075         case OBJ_BLOB:
2076         case OBJ_TAG:
2077                 if (!base_from_cache)
2078                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2079                 break;
2080         default:
2081                 data = NULL;
2082                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2083                       type, (uintmax_t)obj_offset, p->pack_name);
2084         }
2085
2086         /* PHASE 3: apply deltas in order */
2087
2088         /* invariants:
2089          *   'data' holds the base data, or NULL if there was corruption
2090          */
2091         while (delta_stack_nr) {
2092                 void *delta_data;
2093                 void *base = data;
2094                 unsigned long delta_size, base_size = size;
2095                 int i;
2096
2097                 data = NULL;
2098
2099                 if (base)
2100                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2101
2102                 if (!base) {
2103                         /*
2104                          * We're probably in deep shit, but let's try to fetch
2105                          * the required base anyway from another pack or loose.
2106                          * This is costly but should happen only in the presence
2107                          * of a corrupted pack, and is better than failing outright.
2108                          */
2109                         struct revindex_entry *revidx;
2110                         const unsigned char *base_sha1;
2111                         revidx = find_pack_revindex(p, obj_offset);
2112                         if (revidx) {
2113                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2114                                 error("failed to read delta base object %s"
2115                                       " at offset %"PRIuMAX" from %s",
2116                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2117                                       p->pack_name);
2118                                 mark_bad_packed_object(p, base_sha1);
2119                                 base = read_object(base_sha1, &type, &base_size);
2120                         }
2121                 }
2122
2123                 i = --delta_stack_nr;
2124                 obj_offset = delta_stack[i].obj_offset;
2125                 curpos = delta_stack[i].curpos;
2126                 delta_size = delta_stack[i].size;
2127
2128                 if (!base)
2129                         continue;
2130
2131                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2132
2133                 if (!delta_data) {
2134                         error("failed to unpack compressed delta "
2135                               "at offset %"PRIuMAX" from %s",
2136                               (uintmax_t)curpos, p->pack_name);
2137                         data = NULL;
2138                         continue;
2139                 }
2140
2141                 data = patch_delta(base, base_size,
2142                                    delta_data, delta_size,
2143                                    &size);
2144                 if (!data)
2145                         die("failed to apply delta");
2146
2147                 free (delta_data);
2148         }
2149
2150         *final_type = type;
2151         *final_size = size;
2152
2153         unuse_pack(&w_curs);
2154         return data;
2155 }
2156
2157 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2158                                             uint32_t n)
2159 {
2160         const unsigned char *index = p->index_data;
2161         if (!index) {
2162                 if (open_pack_index(p))
2163                         return NULL;
2164                 index = p->index_data;
2165         }
2166         if (n >= p->num_objects)
2167                 return NULL;
2168         index += 4 * 256;
2169         if (p->index_version == 1) {
2170                 return index + 24 * n + 4;
2171         } else {
2172                 index += 8;
2173                 return index + 20 * n;
2174         }
2175 }
2176
2177 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2178 {
2179         const unsigned char *index = p->index_data;
2180         index += 4 * 256;
2181         if (p->index_version == 1) {
2182                 return ntohl(*((uint32_t *)(index + 24 * n)));
2183         } else {
2184                 uint32_t off;
2185                 index += 8 + p->num_objects * (20 + 4);
2186                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2187                 if (!(off & 0x80000000))
2188                         return off;
2189                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2190                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2191                                    ntohl(*((uint32_t *)(index + 4)));
2192         }
2193 }
2194
2195 off_t find_pack_entry_one(const unsigned char *sha1,
2196                                   struct packed_git *p)
2197 {
2198         const uint32_t *level1_ofs = p->index_data;
2199         const unsigned char *index = p->index_data;
2200         unsigned hi, lo, stride;
2201         static int use_lookup = -1;
2202         static int debug_lookup = -1;
2203
2204         if (debug_lookup < 0)
2205                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2206
2207         if (!index) {
2208                 if (open_pack_index(p))
2209                         return 0;
2210                 level1_ofs = p->index_data;
2211                 index = p->index_data;
2212         }
2213         if (p->index_version > 1) {
2214                 level1_ofs += 2;
2215                 index += 8;
2216         }
2217         index += 4 * 256;
2218         hi = ntohl(level1_ofs[*sha1]);
2219         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2220         if (p->index_version > 1) {
2221                 stride = 20;
2222         } else {
2223                 stride = 24;
2224                 index += 4;
2225         }
2226
2227         if (debug_lookup)
2228                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2229                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2230
2231         if (use_lookup < 0)
2232                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2233         if (use_lookup) {
2234                 int pos = sha1_entry_pos(index, stride, 0,
2235                                          lo, hi, p->num_objects, sha1);
2236                 if (pos < 0)
2237                         return 0;
2238                 return nth_packed_object_offset(p, pos);
2239         }
2240
2241         do {
2242                 unsigned mi = (lo + hi) / 2;
2243                 int cmp = hashcmp(index + mi * stride, sha1);
2244
2245                 if (debug_lookup)
2246                         printf("lo %u hi %u rg %u mi %u\n",
2247                                lo, hi, hi - lo, mi);
2248                 if (!cmp)
2249                         return nth_packed_object_offset(p, mi);
2250                 if (cmp > 0)
2251                         hi = mi;
2252                 else
2253                         lo = mi+1;
2254         } while (lo < hi);
2255         return 0;
2256 }
2257
2258 int is_pack_valid(struct packed_git *p)
2259 {
2260         /* An already open pack is known to be valid. */
2261         if (p->pack_fd != -1)
2262                 return 1;
2263
2264         /* If the pack has one window completely covering the
2265          * file size, the pack is known to be valid even if
2266          * the descriptor is not currently open.
2267          */
2268         if (p->windows) {
2269                 struct pack_window *w = p->windows;
2270
2271                 if (!w->offset && w->len == p->pack_size)
2272                         return 1;
2273         }
2274
2275         /* Force the pack to open to prove its valid. */
2276         return !open_packed_git(p);
2277 }
2278
2279 static int fill_pack_entry(const unsigned char *sha1,
2280                            struct pack_entry *e,
2281                            struct packed_git *p)
2282 {
2283         off_t offset;
2284
2285         if (p->num_bad_objects) {
2286                 unsigned i;
2287                 for (i = 0; i < p->num_bad_objects; i++)
2288                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2289                                 return 0;
2290         }
2291
2292         offset = find_pack_entry_one(sha1, p);
2293         if (!offset)
2294                 return 0;
2295
2296         /*
2297          * We are about to tell the caller where they can locate the
2298          * requested object.  We better make sure the packfile is
2299          * still here and can be accessed before supplying that
2300          * answer, as it may have been deleted since the index was
2301          * loaded!
2302          */
2303         if (!is_pack_valid(p)) {
2304                 warning("packfile %s cannot be accessed", p->pack_name);
2305                 return 0;
2306         }
2307         e->offset = offset;
2308         e->p = p;
2309         hashcpy(e->sha1, sha1);
2310         return 1;
2311 }
2312
2313 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2314 {
2315         struct packed_git *p;
2316
2317         prepare_packed_git();
2318         if (!packed_git)
2319                 return 0;
2320
2321         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2322                 return 1;
2323
2324         for (p = packed_git; p; p = p->next) {
2325                 if (p == last_found_pack || !fill_pack_entry(sha1, e, p))
2326                         continue;
2327
2328                 last_found_pack = p;
2329                 return 1;
2330         }
2331         return 0;
2332 }
2333
2334 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2335                                   struct packed_git *packs)
2336 {
2337         struct packed_git *p;
2338
2339         for (p = packs; p; p = p->next) {
2340                 if (find_pack_entry_one(sha1, p))
2341                         return p;
2342         }
2343         return NULL;
2344
2345 }
2346
2347 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep,
2348                                   unsigned long *disk_sizep)
2349 {
2350         int status;
2351         unsigned long mapsize, size;
2352         void *map;
2353         git_zstream stream;
2354         char hdr[32];
2355
2356         map = map_sha1_file(sha1, &mapsize);
2357         if (!map)
2358                 return error("unable to find %s", sha1_to_hex(sha1));
2359         if (disk_sizep)
2360                 *disk_sizep = mapsize;
2361         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2362                 status = error("unable to unpack %s header",
2363                                sha1_to_hex(sha1));
2364         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2365                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2366         else if (sizep)
2367                 *sizep = size;
2368         git_inflate_end(&stream);
2369         munmap(map, mapsize);
2370         return status;
2371 }
2372
2373 /* returns enum object_type or negative */
2374 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi)
2375 {
2376         struct cached_object *co;
2377         struct pack_entry e;
2378         int status, rtype;
2379
2380         co = find_cached_object(sha1);
2381         if (co) {
2382                 if (oi->sizep)
2383                         *(oi->sizep) = co->size;
2384                 if (oi->disk_sizep)
2385                         *(oi->disk_sizep) = 0;
2386                 oi->whence = OI_CACHED;
2387                 return co->type;
2388         }
2389
2390         if (!find_pack_entry(sha1, &e)) {
2391                 /* Most likely it's a loose object. */
2392                 status = sha1_loose_object_info(sha1, oi->sizep, oi->disk_sizep);
2393                 if (status >= 0) {
2394                         oi->whence = OI_LOOSE;
2395                         return status;
2396                 }
2397
2398                 /* Not a loose object; someone else may have just packed it. */
2399                 reprepare_packed_git();
2400                 if (!find_pack_entry(sha1, &e))
2401                         return status;
2402         }
2403
2404         status = packed_object_info(e.p, e.offset, oi->sizep, &rtype,
2405                                     oi->disk_sizep);
2406         if (status < 0) {
2407                 mark_bad_packed_object(e.p, sha1);
2408                 status = sha1_object_info_extended(sha1, oi);
2409         } else if (in_delta_base_cache(e.p, e.offset)) {
2410                 oi->whence = OI_DBCACHED;
2411         } else {
2412                 oi->whence = OI_PACKED;
2413                 oi->u.packed.offset = e.offset;
2414                 oi->u.packed.pack = e.p;
2415                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2416                                          rtype == OBJ_OFS_DELTA);
2417         }
2418
2419         return status;
2420 }
2421
2422 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2423 {
2424         struct object_info oi = {0};
2425
2426         oi.sizep = sizep;
2427         return sha1_object_info_extended(sha1, &oi);
2428 }
2429
2430 static void *read_packed_sha1(const unsigned char *sha1,
2431                               enum object_type *type, unsigned long *size)
2432 {
2433         struct pack_entry e;
2434         void *data;
2435
2436         if (!find_pack_entry(sha1, &e))
2437                 return NULL;
2438         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2439         if (!data) {
2440                 /*
2441                  * We're probably in deep shit, but let's try to fetch
2442                  * the required object anyway from another pack or loose.
2443                  * This should happen only in the presence of a corrupted
2444                  * pack, and is better than failing outright.
2445                  */
2446                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2447                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2448                 mark_bad_packed_object(e.p, sha1);
2449                 data = read_object(sha1, type, size);
2450         }
2451         return data;
2452 }
2453
2454 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2455                       unsigned char *sha1)
2456 {
2457         struct cached_object *co;
2458
2459         hash_sha1_file(buf, len, typename(type), sha1);
2460         if (has_sha1_file(sha1) || find_cached_object(sha1))
2461                 return 0;
2462         if (cached_object_alloc <= cached_object_nr) {
2463                 cached_object_alloc = alloc_nr(cached_object_alloc);
2464                 cached_objects = xrealloc(cached_objects,
2465                                           sizeof(*cached_objects) *
2466                                           cached_object_alloc);
2467         }
2468         co = &cached_objects[cached_object_nr++];
2469         co->size = len;
2470         co->type = type;
2471         co->buf = xmalloc(len);
2472         memcpy(co->buf, buf, len);
2473         hashcpy(co->sha1, sha1);
2474         return 0;
2475 }
2476
2477 static void *read_object(const unsigned char *sha1, enum object_type *type,
2478                          unsigned long *size)
2479 {
2480         unsigned long mapsize;
2481         void *map, *buf;
2482         struct cached_object *co;
2483
2484         co = find_cached_object(sha1);
2485         if (co) {
2486                 *type = co->type;
2487                 *size = co->size;
2488                 return xmemdupz(co->buf, co->size);
2489         }
2490
2491         buf = read_packed_sha1(sha1, type, size);
2492         if (buf)
2493                 return buf;
2494         map = map_sha1_file(sha1, &mapsize);
2495         if (map) {
2496                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2497                 munmap(map, mapsize);
2498                 return buf;
2499         }
2500         reprepare_packed_git();
2501         return read_packed_sha1(sha1, type, size);
2502 }
2503
2504 /*
2505  * This function dies on corrupt objects; the callers who want to
2506  * deal with them should arrange to call read_object() and give error
2507  * messages themselves.
2508  */
2509 void *read_sha1_file_extended(const unsigned char *sha1,
2510                               enum object_type *type,
2511                               unsigned long *size,
2512                               unsigned flag)
2513 {
2514         void *data;
2515         char *path;
2516         const struct packed_git *p;
2517         const unsigned char *repl = (flag & READ_SHA1_FILE_REPLACE)
2518                 ? lookup_replace_object(sha1) : sha1;
2519
2520         errno = 0;
2521         data = read_object(repl, type, size);
2522         if (data)
2523                 return data;
2524
2525         if (errno && errno != ENOENT)
2526                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2527
2528         /* die if we replaced an object with one that does not exist */
2529         if (repl != sha1)
2530                 die("replacement %s not found for %s",
2531                     sha1_to_hex(repl), sha1_to_hex(sha1));
2532
2533         if (has_loose_object(repl)) {
2534                 path = sha1_file_name(sha1);
2535                 die("loose object %s (stored in %s) is corrupt",
2536                     sha1_to_hex(repl), path);
2537         }
2538
2539         if ((p = has_packed_and_bad(repl)) != NULL)
2540                 die("packed object %s (stored in %s) is corrupt",
2541                     sha1_to_hex(repl), p->pack_name);
2542
2543         return NULL;
2544 }
2545
2546 void *read_object_with_reference(const unsigned char *sha1,
2547                                  const char *required_type_name,
2548                                  unsigned long *size,
2549                                  unsigned char *actual_sha1_return)
2550 {
2551         enum object_type type, required_type;
2552         void *buffer;
2553         unsigned long isize;
2554         unsigned char actual_sha1[20];
2555
2556         required_type = type_from_string(required_type_name);
2557         hashcpy(actual_sha1, sha1);
2558         while (1) {
2559                 int ref_length = -1;
2560                 const char *ref_type = NULL;
2561
2562                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2563                 if (!buffer)
2564                         return NULL;
2565                 if (type == required_type) {
2566                         *size = isize;
2567                         if (actual_sha1_return)
2568                                 hashcpy(actual_sha1_return, actual_sha1);
2569                         return buffer;
2570                 }
2571                 /* Handle references */
2572                 else if (type == OBJ_COMMIT)
2573                         ref_type = "tree ";
2574                 else if (type == OBJ_TAG)
2575                         ref_type = "object ";
2576                 else {
2577                         free(buffer);
2578                         return NULL;
2579                 }
2580                 ref_length = strlen(ref_type);
2581
2582                 if (ref_length + 40 > isize ||
2583                     memcmp(buffer, ref_type, ref_length) ||
2584                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2585                         free(buffer);
2586                         return NULL;
2587                 }
2588                 free(buffer);
2589                 /* Now we have the ID of the referred-to object in
2590                  * actual_sha1.  Check again. */
2591         }
2592 }
2593
2594 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2595                                     const char *type, unsigned char *sha1,
2596                                     char *hdr, int *hdrlen)
2597 {
2598         git_SHA_CTX c;
2599
2600         /* Generate the header */
2601         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2602
2603         /* Sha1.. */
2604         git_SHA1_Init(&c);
2605         git_SHA1_Update(&c, hdr, *hdrlen);
2606         git_SHA1_Update(&c, buf, len);
2607         git_SHA1_Final(sha1, &c);
2608 }
2609
2610 /*
2611  * Move the just written object into its final resting place.
2612  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2613  * "moving" is only a part of what it does, when no patch between
2614  * master to pu changes the call sites of this function.
2615  */
2616 int move_temp_to_file(const char *tmpfile, const char *filename)
2617 {
2618         int ret = 0;
2619
2620         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2621                 goto try_rename;
2622         else if (link(tmpfile, filename))
2623                 ret = errno;
2624
2625         /*
2626          * Coda hack - coda doesn't like cross-directory links,
2627          * so we fall back to a rename, which will mean that it
2628          * won't be able to check collisions, but that's not a
2629          * big deal.
2630          *
2631          * The same holds for FAT formatted media.
2632          *
2633          * When this succeeds, we just return.  We have nothing
2634          * left to unlink.
2635          */
2636         if (ret && ret != EEXIST) {
2637         try_rename:
2638                 if (!rename(tmpfile, filename))
2639                         goto out;
2640                 ret = errno;
2641         }
2642         unlink_or_warn(tmpfile);
2643         if (ret) {
2644                 if (ret != EEXIST) {
2645                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2646                 }
2647                 /* FIXME!!! Collision check here ? */
2648         }
2649
2650 out:
2651         if (adjust_shared_perm(filename))
2652                 return error("unable to set permission to '%s'", filename);
2653         return 0;
2654 }
2655
2656 static int write_buffer(int fd, const void *buf, size_t len)
2657 {
2658         if (write_in_full(fd, buf, len) < 0)
2659                 return error("file write error (%s)", strerror(errno));
2660         return 0;
2661 }
2662
2663 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2664                    unsigned char *sha1)
2665 {
2666         char hdr[32];
2667         int hdrlen;
2668         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2669         return 0;
2670 }
2671
2672 /* Finalize a file on disk, and close it. */
2673 static void close_sha1_file(int fd)
2674 {
2675         if (fsync_object_files)
2676                 fsync_or_die(fd, "sha1 file");
2677         if (close(fd) != 0)
2678                 die_errno("error when closing sha1 file");
2679 }
2680
2681 /* Size of directory component, including the ending '/' */
2682 static inline int directory_size(const char *filename)
2683 {
2684         const char *s = strrchr(filename, '/');
2685         if (!s)
2686                 return 0;
2687         return s - filename + 1;
2688 }
2689
2690 /*
2691  * This creates a temporary file in the same directory as the final
2692  * 'filename'
2693  *
2694  * We want to avoid cross-directory filename renames, because those
2695  * can have problems on various filesystems (FAT, NFS, Coda).
2696  */
2697 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2698 {
2699         int fd, dirlen = directory_size(filename);
2700
2701         if (dirlen + 20 > bufsiz) {
2702                 errno = ENAMETOOLONG;
2703                 return -1;
2704         }
2705         memcpy(buffer, filename, dirlen);
2706         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2707         fd = git_mkstemp_mode(buffer, 0444);
2708         if (fd < 0 && dirlen && errno == ENOENT) {
2709                 /* Make sure the directory exists */
2710                 memcpy(buffer, filename, dirlen);
2711                 buffer[dirlen-1] = 0;
2712                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2713                         return -1;
2714
2715                 /* Try again */
2716                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2717                 fd = git_mkstemp_mode(buffer, 0444);
2718         }
2719         return fd;
2720 }
2721
2722 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2723                               const void *buf, unsigned long len, time_t mtime)
2724 {
2725         int fd, ret;
2726         unsigned char compressed[4096];
2727         git_zstream stream;
2728         git_SHA_CTX c;
2729         unsigned char parano_sha1[20];
2730         char *filename;
2731         static char tmp_file[PATH_MAX];
2732
2733         filename = sha1_file_name(sha1);
2734         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
2735         if (fd < 0) {
2736                 if (errno == EACCES)
2737                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
2738                 else
2739                         return error("unable to create temporary file: %s", strerror(errno));
2740         }
2741
2742         /* Set it up */
2743         memset(&stream, 0, sizeof(stream));
2744         git_deflate_init(&stream, zlib_compression_level);
2745         stream.next_out = compressed;
2746         stream.avail_out = sizeof(compressed);
2747         git_SHA1_Init(&c);
2748
2749         /* First header.. */
2750         stream.next_in = (unsigned char *)hdr;
2751         stream.avail_in = hdrlen;
2752         while (git_deflate(&stream, 0) == Z_OK)
2753                 ; /* nothing */
2754         git_SHA1_Update(&c, hdr, hdrlen);
2755
2756         /* Then the data itself.. */
2757         stream.next_in = (void *)buf;
2758         stream.avail_in = len;
2759         do {
2760                 unsigned char *in0 = stream.next_in;
2761                 ret = git_deflate(&stream, Z_FINISH);
2762                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2763                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2764                         die("unable to write sha1 file");
2765                 stream.next_out = compressed;
2766                 stream.avail_out = sizeof(compressed);
2767         } while (ret == Z_OK);
2768
2769         if (ret != Z_STREAM_END)
2770                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2771         ret = git_deflate_end_gently(&stream);
2772         if (ret != Z_OK)
2773                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2774         git_SHA1_Final(parano_sha1, &c);
2775         if (hashcmp(sha1, parano_sha1) != 0)
2776                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2777
2778         close_sha1_file(fd);
2779
2780         if (mtime) {
2781                 struct utimbuf utb;
2782                 utb.actime = mtime;
2783                 utb.modtime = mtime;
2784                 if (utime(tmp_file, &utb) < 0)
2785                         warning("failed utime() on %s: %s",
2786                                 tmp_file, strerror(errno));
2787         }
2788
2789         return move_temp_to_file(tmp_file, filename);
2790 }
2791
2792 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2793 {
2794         unsigned char sha1[20];
2795         char hdr[32];
2796         int hdrlen;
2797
2798         /* Normally if we have it in the pack then we do not bother writing
2799          * it out into .git/objects/??/?{38} file.
2800          */
2801         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2802         if (returnsha1)
2803                 hashcpy(returnsha1, sha1);
2804         if (has_sha1_file(sha1))
2805                 return 0;
2806         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2807 }
2808
2809 int force_object_loose(const unsigned char *sha1, time_t mtime)
2810 {
2811         void *buf;
2812         unsigned long len;
2813         enum object_type type;
2814         char hdr[32];
2815         int hdrlen;
2816         int ret;
2817
2818         if (has_loose_object(sha1))
2819                 return 0;
2820         buf = read_packed_sha1(sha1, &type, &len);
2821         if (!buf)
2822                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2823         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2824         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2825         free(buf);
2826
2827         return ret;
2828 }
2829
2830 int has_pack_index(const unsigned char *sha1)
2831 {
2832         struct stat st;
2833         if (stat(sha1_pack_index_name(sha1), &st))
2834                 return 0;
2835         return 1;
2836 }
2837
2838 int has_sha1_pack(const unsigned char *sha1)
2839 {
2840         struct pack_entry e;
2841         return find_pack_entry(sha1, &e);
2842 }
2843
2844 int has_sha1_file(const unsigned char *sha1)
2845 {
2846         struct pack_entry e;
2847
2848         if (find_pack_entry(sha1, &e))
2849                 return 1;
2850         return has_loose_object(sha1);
2851 }
2852
2853 static void check_tree(const void *buf, size_t size)
2854 {
2855         struct tree_desc desc;
2856         struct name_entry entry;
2857
2858         init_tree_desc(&desc, buf, size);
2859         while (tree_entry(&desc, &entry))
2860                 /* do nothing
2861                  * tree_entry() will die() on malformed entries */
2862                 ;
2863 }
2864
2865 static void check_commit(const void *buf, size_t size)
2866 {
2867         struct commit c;
2868         memset(&c, 0, sizeof(c));
2869         if (parse_commit_buffer(&c, buf, size))
2870                 die("corrupt commit");
2871 }
2872
2873 static void check_tag(const void *buf, size_t size)
2874 {
2875         struct tag t;
2876         memset(&t, 0, sizeof(t));
2877         if (parse_tag_buffer(&t, buf, size))
2878                 die("corrupt tag");
2879 }
2880
2881 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2882                      enum object_type type,
2883                      const char *path, unsigned flags)
2884 {
2885         int ret, re_allocated = 0;
2886         int write_object = flags & HASH_WRITE_OBJECT;
2887
2888         if (!type)
2889                 type = OBJ_BLOB;
2890
2891         /*
2892          * Convert blobs to git internal format
2893          */
2894         if ((type == OBJ_BLOB) && path) {
2895                 struct strbuf nbuf = STRBUF_INIT;
2896                 if (convert_to_git(path, buf, size, &nbuf,
2897                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
2898                         buf = strbuf_detach(&nbuf, &size);
2899                         re_allocated = 1;
2900                 }
2901         }
2902         if (flags & HASH_FORMAT_CHECK) {
2903                 if (type == OBJ_TREE)
2904                         check_tree(buf, size);
2905                 if (type == OBJ_COMMIT)
2906                         check_commit(buf, size);
2907                 if (type == OBJ_TAG)
2908                         check_tag(buf, size);
2909         }
2910
2911         if (write_object)
2912                 ret = write_sha1_file(buf, size, typename(type), sha1);
2913         else
2914                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2915         if (re_allocated)
2916                 free(buf);
2917         return ret;
2918 }
2919
2920 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
2921                       const char *path, unsigned flags)
2922 {
2923         struct strbuf sbuf = STRBUF_INIT;
2924         int ret;
2925
2926         if (strbuf_read(&sbuf, fd, 4096) >= 0)
2927                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
2928         else
2929                 ret = -1;
2930         strbuf_release(&sbuf);
2931         return ret;
2932 }
2933
2934 #define SMALL_FILE_SIZE (32*1024)
2935
2936 static int index_core(unsigned char *sha1, int fd, size_t size,
2937                       enum object_type type, const char *path,
2938                       unsigned flags)
2939 {
2940         int ret;
2941
2942         if (!size) {
2943                 ret = index_mem(sha1, NULL, size, type, path, flags);
2944         } else if (size <= SMALL_FILE_SIZE) {
2945                 char *buf = xmalloc(size);
2946                 if (size == read_in_full(fd, buf, size))
2947                         ret = index_mem(sha1, buf, size, type, path, flags);
2948                 else
2949                         ret = error("short read %s", strerror(errno));
2950                 free(buf);
2951         } else {
2952                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2953                 ret = index_mem(sha1, buf, size, type, path, flags);
2954                 munmap(buf, size);
2955         }
2956         return ret;
2957 }
2958
2959 /*
2960  * This creates one packfile per large blob unless bulk-checkin
2961  * machinery is "plugged".
2962  *
2963  * This also bypasses the usual "convert-to-git" dance, and that is on
2964  * purpose. We could write a streaming version of the converting
2965  * functions and insert that before feeding the data to fast-import
2966  * (or equivalent in-core API described above). However, that is
2967  * somewhat complicated, as we do not know the size of the filter
2968  * result, which we need to know beforehand when writing a git object.
2969  * Since the primary motivation for trying to stream from the working
2970  * tree file and to avoid mmaping it in core is to deal with large
2971  * binary blobs, they generally do not want to get any conversion, and
2972  * callers should avoid this code path when filters are requested.
2973  */
2974 static int index_stream(unsigned char *sha1, int fd, size_t size,
2975                         enum object_type type, const char *path,
2976                         unsigned flags)
2977 {
2978         return index_bulk_checkin(sha1, fd, size, type, path, flags);
2979 }
2980
2981 int index_fd(unsigned char *sha1, int fd, struct stat *st,
2982              enum object_type type, const char *path, unsigned flags)
2983 {
2984         int ret;
2985         size_t size = xsize_t(st->st_size);
2986
2987         if (!S_ISREG(st->st_mode))
2988                 ret = index_pipe(sha1, fd, type, path, flags);
2989         else if (size <= big_file_threshold || type != OBJ_BLOB ||
2990                  (path && would_convert_to_git(path, NULL, 0, 0)))
2991                 ret = index_core(sha1, fd, size, type, path, flags);
2992         else
2993                 ret = index_stream(sha1, fd, size, type, path, flags);
2994         close(fd);
2995         return ret;
2996 }
2997
2998 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
2999 {
3000         int fd;
3001         struct strbuf sb = STRBUF_INIT;
3002
3003         switch (st->st_mode & S_IFMT) {
3004         case S_IFREG:
3005                 fd = open(path, O_RDONLY);
3006                 if (fd < 0)
3007                         return error("open(\"%s\"): %s", path,
3008                                      strerror(errno));
3009                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3010                         return error("%s: failed to insert into database",
3011                                      path);
3012                 break;
3013         case S_IFLNK:
3014                 if (strbuf_readlink(&sb, path, st->st_size)) {
3015                         char *errstr = strerror(errno);
3016                         return error("readlink(\"%s\"): %s", path,
3017                                      errstr);
3018                 }
3019                 if (!(flags & HASH_WRITE_OBJECT))
3020                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3021                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3022                         return error("%s: failed to insert into database",
3023                                      path);
3024                 strbuf_release(&sb);
3025                 break;
3026         case S_IFDIR:
3027                 return resolve_gitlink_ref(path, "HEAD", sha1);
3028         default:
3029                 return error("%s: unsupported file type", path);
3030         }
3031         return 0;
3032 }
3033
3034 int read_pack_header(int fd, struct pack_header *header)
3035 {
3036         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3037                 /* "eof before pack header was fully read" */
3038                 return PH_ERROR_EOF;
3039
3040         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3041                 /* "protocol error (pack signature mismatch detected)" */
3042                 return PH_ERROR_PACK_SIGNATURE;
3043         if (!pack_version_ok(header->hdr_version))
3044                 /* "protocol error (pack version unsupported)" */
3045                 return PH_ERROR_PROTOCOL;
3046         return 0;
3047 }
3048
3049 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3050 {
3051         enum object_type type = sha1_object_info(sha1, NULL);
3052         if (type < 0)
3053                 die("%s is not a valid object", sha1_to_hex(sha1));
3054         if (type != expect)
3055                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3056                     typename(expect));
3057 }