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