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 static struct packed_git *last_found_pack;
  64
  65 static struct cached_object *find_cached_object(const unsigned char *sha1)
  66 {
  67         int i;
  68         struct cached_object *co = cached_objects;
  69
  70         for (i = 0; i < cached_object_nr; i++, co++) {
  71                 if (!hashcmp(co->sha1, sha1))
  72                         return co;
  73         }
  74         if (!hashcmp(sha1, empty_tree.sha1))
  75                 return &empty_tree;
  76         return NULL;
  77 }
  78
  79 int mkdir_in_gitdir(const char *path)
  80 {
  81         if (mkdir(path, 0777)) {
  82                 int saved_errno = errno;
  83                 struct stat st;
  84                 struct strbuf sb = STRBUF_INIT;
  85
  86                 if (errno != EEXIST)
  87                         return -1;
  88                 /*
  89                  * Are we looking at a path in a symlinked worktree
  90                  * whose original repository does not yet have it?
  91                  * e.g. .git/rr-cache pointing at its original
  92                  * repository in which the user hasn't performed any
  93                  * conflict resolution yet?
  94                  */
  95                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
  96                     strbuf_readlink(&sb, path, st.st_size) ||
  97                     !is_absolute_path(sb.buf) ||
  98                     mkdir(sb.buf, 0777)) {
  99                         strbuf_release(&sb);
 100                         errno = saved_errno;
 101                         return -1;
 102                 }
 103                 strbuf_release(&sb);
 104         }
 105         return adjust_shared_perm(path);
 106 }
 107
 108 int safe_create_leading_directories(char *path)
 109 {
 110         char *pos = path + offset_1st_component(path);
 111         struct stat st;
 112
 113         while (pos) {
 114                 pos = strchr(pos, '/');
 115                 if (!pos)
 116                         break;
 117                 while (*++pos == '/')
 118                         ;
 119                 if (!*pos)
 120                         break;
 121                 *--pos = '\0';
 122                 if (!stat(path, &st)) {
 123                         /* path exists */
 124                         if (!S_ISDIR(st.st_mode)) {
 125                                 *pos = '/';
 126                                 return -3;
 127                         }
 128                 }
 129                 else if (mkdir(path, 0777)) {
 130                         if (errno == EEXIST &&
 131                             !stat(path, &st) && S_ISDIR(st.st_mode)) {
 132                                 ; /* somebody created it since we checked */
 133                         } else {
 134                                 *pos = '/';
 135                                 return -1;
 136                         }
 137                 }
 138                 else if (adjust_shared_perm(path)) {
 139                         *pos = '/';
 140                         return -2;
 141                 }
 142                 *pos++ = '/';
 143         }
 144         return 0;
 145 }
 146
 147 int safe_create_leading_directories_const(const char *path)
 148 {
 149         /* path points to cache entries, so xstrdup before messing with it */
 150         char *buf = xstrdup(path);
 151         int result = safe_create_leading_directories(buf);
 152         free(buf);
 153         return result;
 154 }
 155
 156 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
 157 {
 158         int i;
 159         for (i = 0; i < 20; i++) {
 160                 static char hex[] = "0123456789abcdef";
 161                 unsigned int val = sha1[i];
 162                 char *pos = pathbuf + i*2 + (i > 0);
 163                 *pos++ = hex[val >> 4];
 164                 *pos = hex[val & 0xf];
 165         }
 166 }
 167
 168 /*
 169  * NOTE! This returns a statically allocated buffer, so you have to be
 170  * careful about using it. Do an "xstrdup()" if you need to save the
 171  * filename.
 172  *
 173  * Also note that this returns the location for creating.  Reading
 174  * SHA1 file can happen from any alternate directory listed in the
 175  * DB_ENVIRONMENT environment variable if it is not found in
 176  * the primary object database.
 177  */
 178 char *sha1_file_name(const unsigned char *sha1)
 179 {
 180         static char buf[PATH_MAX];
 181         const char *objdir;
 182         int len;
 183
 184         objdir = get_object_directory();
 185         len = strlen(objdir);
 186
 187         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
 188         if (len + 43 > PATH_MAX)
 189                 die("insanely long object directory %s", objdir);
 190         memcpy(buf, objdir, len);
 191         buf[len] = '/';
 192         buf[len+3] = '/';
 193         buf[len+42] = '\0';
 194         fill_sha1_path(buf + len + 1, sha1);
 195         return buf;
 196 }
 197
 198 static char *sha1_get_pack_name(const unsigned char *sha1,
 199                                 char **name, char **base, const char *which)
 200 {
 201         static const char hex[] = "0123456789abcdef";
 202         char *buf;
 203         int i;
 204
 205         if (!*base) {
 206                 const char *sha1_file_directory = get_object_directory();
 207                 int len = strlen(sha1_file_directory);
 208                 *base = xmalloc(len + 60);
 209                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
 210                         sha1_file_directory, which);
 211                 *name = *base + len + 11;
 212         }
 213
 214         buf = *name;
 215
 216         for (i = 0; i < 20; i++) {
 217                 unsigned int val = *sha1++;
 218                 *buf++ = hex[val >> 4];
 219                 *buf++ = hex[val & 0xf];
 220         }
 221
 222         return *base;
 223 }
 224
 225 char *sha1_pack_name(const unsigned char *sha1)
 226 {
 227         static char *name, *base;
 228
 229         return sha1_get_pack_name(sha1, &name, &base, "pack");
 230 }
 231
 232 char *sha1_pack_index_name(const unsigned char *sha1)
 233 {
 234         static char *name, *base;
 235
 236         return sha1_get_pack_name(sha1, &name, &base, "idx");
 237 }
 238
 239 struct alternate_object_database *alt_odb_list;
 240 static struct alternate_object_database **alt_odb_tail;
 241
 242 static int git_open_noatime(const char *name);
 243
 244 /*
 245  * Prepare alternate object database registry.
 246  *
 247  * The variable alt_odb_list points at the list of struct
 248  * alternate_object_database.  The elements on this list come from
 249  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
 250  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
 251  * whose contents is similar to that environment variable but can be
 252  * LF separated.  Its base points at a statically allocated buffer that
 253  * contains "/the/directory/corresponding/to/.git/objects/...", while
 254  * its name points just after the slash at the end of ".git/objects/"
 255  * in the example above, and has enough space to hold 40-byte hex
 256  * SHA1, an extra slash for the first level indirection, and the
 257  * terminating NUL.
 258  */
 259 static int link_alt_odb_entry(const char *entry, const char *relative_base, int depth)
 260 {
 261         const char *objdir = get_object_directory();
 262         struct alternate_object_database *ent;
 263         struct alternate_object_database *alt;
 264         int pfxlen, entlen;
 265         struct strbuf pathbuf = STRBUF_INIT;
 266
 267         if (!is_absolute_path(entry) && relative_base) {
 268                 strbuf_addstr(&pathbuf, real_path(relative_base));
 269                 strbuf_addch(&pathbuf, '/');
 270         }
 271         strbuf_addstr(&pathbuf, entry);
 272
 273         normalize_path_copy(pathbuf.buf, pathbuf.buf);
 274
 275         pfxlen = strlen(pathbuf.buf);
 276
 277         /*
 278          * The trailing slash after the directory name is given by
 279          * this function at the end. Remove duplicates.
 280          */
 281         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
 282                 pfxlen -= 1;
 283
 284         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
 285         ent = xmalloc(sizeof(*ent) + entlen);
 286         memcpy(ent->base, pathbuf.buf, pfxlen);
 287         strbuf_release(&pathbuf);
 288
 289         ent->name = ent->base + pfxlen + 1;
 290         ent->base[pfxlen + 3] = '/';
 291         ent->base[pfxlen] = ent->base[entlen-1] = 0;
 292
 293         /* Detect cases where alternate disappeared */
 294         if (!is_directory(ent->base)) {
 295                 error("object directory %s does not exist; "
 296                       "check .git/objects/info/alternates.",
 297                       ent->base);
 298                 free(ent);
 299                 return -1;
 300         }
 301
 302         /* Prevent the common mistake of listing the same
 303          * thing twice, or object directory itself.
 304          */
 305         for (alt = alt_odb_list; alt; alt = alt->next) {
 306                 if (!memcmp(ent->base, alt->base, pfxlen)) {
 307                         free(ent);
 308                         return -1;
 309                 }
 310         }
 311         if (!strcmp(ent->base, objdir)) {
 312                 free(ent);
 313                 return -1;
 314         }
 315
 316         /* add the alternate entry */
 317         *alt_odb_tail = ent;
 318         alt_odb_tail = &(ent->next);
 319         ent->next = NULL;
 320
 321         /* recursively add alternates */
 322         read_info_alternates(ent->base, depth + 1);
 323
 324         ent->base[pfxlen] = '/';
 325
 326         return 0;
 327 }
 328
 329 static void link_alt_odb_entries(const char *alt, int len, int sep,
 330                                  const char *relative_base, int depth)
 331 {
 332         struct string_list entries = STRING_LIST_INIT_NODUP;
 333         char *alt_copy;
 334         int i;
 335
 336         if (depth > 5) {
 337                 error("%s: ignoring alternate object stores, nesting too deep.",
 338                                 relative_base);
 339                 return;
 340         }
 341
 342         alt_copy = xmemdupz(alt, len);
 343         string_list_split_in_place(&entries, alt_copy, sep, -1);
 344         for (i = 0; i < entries.nr; i++) {
 345                 const char *entry = entries.items[i].string;
 346                 if (entry[0] == '\0' || entry[0] == '#')
 347                         continue;
 348                 if (!is_absolute_path(entry) && depth) {
 349                         error("%s: ignoring relative alternate object store %s",
 350                                         relative_base, entry);
 351                 } else {
 352                         link_alt_odb_entry(entry, relative_base, depth);
 353                 }
 354         }
 355         string_list_clear(&entries, 0);
 356         free(alt_copy);
 357 }
 358
 359 void read_info_alternates(const char * relative_base, int depth)
 360 {
 361         char *map;
 362         size_t mapsz;
 363         struct stat st;
 364         const char alt_file_name[] = "info/alternates";
 365         /* Given that relative_base is no longer than PATH_MAX,
 366            ensure that "path" has enough space to append "/", the
 367            file name, "info/alternates", and a trailing NUL.  */
 368         char path[PATH_MAX + 1 + sizeof alt_file_name];
 369         int fd;
 370
 371         sprintf(path, "%s/%s", relative_base, alt_file_name);
 372         fd = git_open_noatime(path);
 373         if (fd < 0)
 374                 return;
 375         if (fstat(fd, &st) || (st.st_size == 0)) {
 376                 close(fd);
 377                 return;
 378         }
 379         mapsz = xsize_t(st.st_size);
 380         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
 381         close(fd);
 382
 383         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
 384
 385         munmap(map, mapsz);
 386 }
 387
 388 void add_to_alternates_file(const char *reference)
 389 {
 390         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
 391         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
 392         char *alt = mkpath("%s\n", reference);
 393         write_or_die(fd, alt, strlen(alt));
 394         if (commit_lock_file(lock))
 395                 die("could not close alternates file");
 396         if (alt_odb_tail)
 397                 link_alt_odb_entries(alt, strlen(alt), '\n', NULL, 0);
 398 }
 399
 400 void foreach_alt_odb(alt_odb_fn fn, void *cb)
 401 {
 402         struct alternate_object_database *ent;
 403
 404         prepare_alt_odb();
 405         for (ent = alt_odb_list; ent; ent = ent->next)
 406                 if (fn(ent, cb))
 407                         return;
 408 }
 409
 410 void prepare_alt_odb(void)
 411 {
 412         const char *alt;
 413
 414         if (alt_odb_tail)
 415                 return;
 416
 417         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
 418         if (!alt) alt = "";
 419
 420         alt_odb_tail = &alt_odb_list;
 421         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
 422
 423         read_info_alternates(get_object_directory(), 0);
 424 }
 425
 426 static int has_loose_object_local(const unsigned char *sha1)
 427 {
 428         char *name = sha1_file_name(sha1);
 429         return !access(name, F_OK);
 430 }
 431
 432 int has_loose_object_nonlocal(const unsigned char *sha1)
 433 {
 434         struct alternate_object_database *alt;
 435         prepare_alt_odb();
 436         for (alt = alt_odb_list; alt; alt = alt->next) {
 437                 fill_sha1_path(alt->name, sha1);
 438                 if (!access(alt->base, F_OK))
 439                         return 1;
 440         }
 441         return 0;
 442 }
 443
 444 static int has_loose_object(const unsigned char *sha1)
 445 {
 446         return has_loose_object_local(sha1) ||
 447                has_loose_object_nonlocal(sha1);
 448 }
 449
 450 static unsigned int pack_used_ctr;
 451 static unsigned int pack_mmap_calls;
 452 static unsigned int peak_pack_open_windows;
 453 static unsigned int pack_open_windows;
 454 static unsigned int pack_open_fds;
 455 static unsigned int pack_max_fds;
 456 static size_t peak_pack_mapped;
 457 static size_t pack_mapped;
 458 struct packed_git *packed_git;
 459
 460 void pack_report(void)
 461 {
 462         fprintf(stderr,
 463                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
 464                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
 465                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
 466                 sz_fmt(getpagesize()),
 467                 sz_fmt(packed_git_window_size),
 468                 sz_fmt(packed_git_limit));
 469         fprintf(stderr,
 470                 "pack_report: pack_used_ctr            = %10u\n"
 471                 "pack_report: pack_mmap_calls          = %10u\n"
 472                 "pack_report: pack_open_windows        = %10u / %10u\n"
 473                 "pack_report: pack_mapped              = "
 474                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
 475                 pack_used_ctr,
 476                 pack_mmap_calls,
 477                 pack_open_windows, peak_pack_open_windows,
 478                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
 479 }
 480
 481 static int check_packed_git_idx(const char *path,  struct packed_git *p)
 482 {
 483         void *idx_map;
 484         struct pack_idx_header *hdr;
 485         size_t idx_size;
 486         uint32_t version, nr, i, *index;
 487         int fd = git_open_noatime(path);
 488         struct stat st;
 489
 490         if (fd < 0)
 491                 return -1;
 492         if (fstat(fd, &st)) {
 493                 close(fd);
 494                 return -1;
 495         }
 496         idx_size = xsize_t(st.st_size);
 497         if (idx_size < 4 * 256 + 20 + 20) {
 498                 close(fd);
 499                 return error("index file %s is too small", path);
 500         }
 501         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 502         close(fd);
 503
 504         hdr = idx_map;
 505         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 506                 version = ntohl(hdr->idx_version);
 507                 if (version < 2 || version > 2) {
 508                         munmap(idx_map, idx_size);
 509                         return error("index file %s is version %"PRIu32
 510                                      " and is not supported by this binary"
 511                                      " (try upgrading GIT to a newer version)",
 512                                      path, version);
 513                 }
 514         } else
 515                 version = 1;
 516
 517         nr = 0;
 518         index = idx_map;
 519         if (version > 1)
 520                 index += 2;  /* skip index header */
 521         for (i = 0; i < 256; i++) {
 522                 uint32_t n = ntohl(index[i]);
 523                 if (n < nr) {
 524                         munmap(idx_map, idx_size);
 525                         return error("non-monotonic index %s", path);
 526                 }
 527                 nr = n;
 528         }
 529
 530         if (version == 1) {
 531                 /*
 532                  * Total size:
 533                  *  - 256 index entries 4 bytes each
 534                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
 535                  *  - 20-byte SHA1 of the packfile
 536                  *  - 20-byte SHA1 file checksum
 537                  */
 538                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
 539                         munmap(idx_map, idx_size);
 540                         return error("wrong index v1 file size in %s", path);
 541                 }
 542         } else if (version == 2) {
 543                 /*
 544                  * Minimum size:
 545                  *  - 8 bytes of header
 546                  *  - 256 index entries 4 bytes each
 547                  *  - 20-byte sha1 entry * nr
 548                  *  - 4-byte crc entry * nr
 549                  *  - 4-byte offset entry * nr
 550                  *  - 20-byte SHA1 of the packfile
 551                  *  - 20-byte SHA1 file checksum
 552                  * And after the 4-byte offset table might be a
 553                  * variable sized table containing 8-byte entries
 554                  * for offsets larger than 2^31.
 555                  */
 556                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
 557                 unsigned long max_size = min_size;
 558                 if (nr)
 559                         max_size += (nr - 1)*8;
 560                 if (idx_size < min_size || idx_size > max_size) {
 561                         munmap(idx_map, idx_size);
 562                         return error("wrong index v2 file size in %s", path);
 563                 }
 564                 if (idx_size != min_size &&
 565                     /*
 566                      * make sure we can deal with large pack offsets.
 567                      * 31-bit signed offset won't be enough, neither
 568                      * 32-bit unsigned one will be.
 569                      */
 570                     (sizeof(off_t) <= 4)) {
 571                         munmap(idx_map, idx_size);
 572                         return error("pack too large for current definition of off_t in %s", path);
 573                 }
 574         }
 575
 576         p->index_version = version;
 577         p->index_data = idx_map;
 578         p->index_size = idx_size;
 579         p->num_objects = nr;
 580         return 0;
 581 }
 582
 583 int open_pack_index(struct packed_git *p)
 584 {
 585         char *idx_name;
 586         int ret;
 587
 588         if (p->index_data)
 589                 return 0;
 590
 591         idx_name = xstrdup(p->pack_name);
 592         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
 593         ret = check_packed_git_idx(idx_name, p);
 594         free(idx_name);
 595         return ret;
 596 }
 597
 598 static void scan_windows(struct packed_git *p,
 599         struct packed_git **lru_p,
 600         struct pack_window **lru_w,
 601         struct pack_window **lru_l)
 602 {
 603         struct pack_window *w, *w_l;
 604
 605         for (w_l = NULL, w = p->windows; w; w = w->next) {
 606                 if (!w->inuse_cnt) {
 607                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 608                                 *lru_p = p;
 609                                 *lru_w = w;
 610                                 *lru_l = w_l;
 611                         }
 612                 }
 613                 w_l = w;
 614         }
 615 }
 616
 617 static int unuse_one_window(struct packed_git *current, int keep_fd)
 618 {
 619         struct packed_git *p, *lru_p = NULL;
 620         struct pack_window *lru_w = NULL, *lru_l = NULL;
 621
 622         if (current)
 623                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 624         for (p = packed_git; p; p = p->next)
 625                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 626         if (lru_p) {
 627                 munmap(lru_w->base, lru_w->len);
 628                 pack_mapped -= lru_w->len;
 629                 if (lru_l)
 630                         lru_l->next = lru_w->next;
 631                 else {
 632                         lru_p->windows = lru_w->next;
 633                         if (!lru_p->windows && lru_p->pack_fd != -1
 634                                 && lru_p->pack_fd != keep_fd) {
 635                                 close(lru_p->pack_fd);
 636                                 pack_open_fds--;
 637                                 lru_p->pack_fd = -1;
 638                         }
 639                 }
 640                 free(lru_w);
 641                 pack_open_windows--;
 642                 return 1;
 643         }
 644         return 0;
 645 }
 646
 647 void release_pack_memory(size_t need, int fd)
 648 {
 649         size_t cur = pack_mapped;
 650         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
 651                 ; /* nothing */
 652 }
 653
 654 void *xmmap(void *start, size_t length,
 655         int prot, int flags, int fd, off_t offset)
 656 {
 657         void *ret = mmap(start, length, prot, flags, fd, offset);
 658         if (ret == MAP_FAILED) {
 659                 if (!length)
 660                         return NULL;
 661                 release_pack_memory(length, fd);
 662                 ret = mmap(start, length, prot, flags, fd, offset);
 663                 if (ret == MAP_FAILED)
 664                         die_errno("Out of memory? mmap failed");
 665         }
 666         return ret;
 667 }
 668
 669 void close_pack_windows(struct packed_git *p)
 670 {
 671         while (p->windows) {
 672                 struct pack_window *w = p->windows;
 673
 674                 if (w->inuse_cnt)
 675                         die("pack '%s' still has open windows to it",
 676                             p->pack_name);
 677                 munmap(w->base, w->len);
 678                 pack_mapped -= w->len;
 679                 pack_open_windows--;
 680                 p->windows = w->next;
 681                 free(w);
 682         }
 683 }
 684
 685 void unuse_pack(struct pack_window **w_cursor)
 686 {
 687         struct pack_window *w = *w_cursor;
 688         if (w) {
 689                 w->inuse_cnt--;
 690                 *w_cursor = NULL;
 691         }
 692 }
 693
 694 void close_pack_index(struct packed_git *p)
 695 {
 696         if (p->index_data) {
 697                 munmap((void *)p->index_data, p->index_size);
 698                 p->index_data = NULL;
 699         }
 700 }
 701
 702 /*
 703  * This is used by git-repack in case a newly created pack happens to
 704  * contain the same set of objects as an existing one.  In that case
 705  * the resulting file might be different even if its name would be the
 706  * same.  It is best to close any reference to the old pack before it is
 707  * replaced on disk.  Of course no index pointers nor windows for given pack
 708  * must subsist at this point.  If ever objects from this pack are requested
 709  * again, the new version of the pack will be reinitialized through
 710  * reprepare_packed_git().
 711  */
 712 void free_pack_by_name(const char *pack_name)
 713 {
 714         struct packed_git *p, **pp = &packed_git;
 715
 716         while (*pp) {
 717                 p = *pp;
 718                 if (strcmp(pack_name, p->pack_name) == 0) {
 719                         clear_delta_base_cache();
 720                         close_pack_windows(p);
 721                         if (p->pack_fd != -1) {
 722                                 close(p->pack_fd);
 723                                 pack_open_fds--;
 724                         }
 725                         close_pack_index(p);
 726                         free(p->bad_object_sha1);
 727                         *pp = p->next;
 728                         if (last_found_pack == p)
 729                                 last_found_pack = NULL;
 730                         free(p);
 731                         return;
 732                 }
 733                 pp = &p->next;
 734         }
 735 }
 736
 737 static unsigned int get_max_fd_limit(void)
 738 {
 739 #ifdef RLIMIT_NOFILE
 740         struct rlimit lim;
 741
 742         if (getrlimit(RLIMIT_NOFILE, &lim))
 743                 die_errno("cannot get RLIMIT_NOFILE");
 744
 745         return lim.rlim_cur;
 746 #elif defined(_SC_OPEN_MAX)
 747         return sysconf(_SC_OPEN_MAX);
 748 #elif defined(OPEN_MAX)
 749         return OPEN_MAX;
 750 #else
 751         return 1; /* see the caller ;-) */
 752 #endif
 753 }
 754
 755 /*
 756  * Do not call this directly as this leaks p->pack_fd on error return;
 757  * call open_packed_git() instead.
 758  */
 759 static int open_packed_git_1(struct packed_git *p)
 760 {
 761         struct stat st;
 762         struct pack_header hdr;
 763         unsigned char sha1[20];
 764         unsigned char *idx_sha1;
 765         long fd_flag;
 766
 767         if (!p->index_data && open_pack_index(p))
 768                 return error("packfile %s index unavailable", p->pack_name);
 769
 770         if (!pack_max_fds) {
 771                 unsigned int max_fds = get_max_fd_limit();
 772
 773                 /* Save 3 for stdin/stdout/stderr, 22 for work */
 774                 if (25 < max_fds)
 775                         pack_max_fds = max_fds - 25;
 776                 else
 777                         pack_max_fds = 1;
 778         }
 779
 780         while (pack_max_fds <= pack_open_fds && unuse_one_window(NULL, -1))
 781                 ; /* nothing */
 782
 783         p->pack_fd = git_open_noatime(p->pack_name);
 784         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 785                 return -1;
 786         pack_open_fds++;
 787
 788         /* If we created the struct before we had the pack we lack size. */
 789         if (!p->pack_size) {
 790                 if (!S_ISREG(st.st_mode))
 791                         return error("packfile %s not a regular file", p->pack_name);
 792                 p->pack_size = st.st_size;
 793         } else if (p->pack_size != st.st_size)
 794                 return error("packfile %s size changed", p->pack_name);
 795
 796         /* We leave these file descriptors open with sliding mmap;
 797          * there is no point keeping them open across exec(), though.
 798          */
 799         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 800         if (fd_flag < 0)
 801                 return error("cannot determine file descriptor flags");
 802         fd_flag |= FD_CLOEXEC;
 803         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 804                 return error("cannot set FD_CLOEXEC");
 805
 806         /* Verify we recognize this pack file format. */
 807         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 808                 return error("file %s is far too short to be a packfile", p->pack_name);
 809         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 810                 return error("file %s is not a GIT packfile", p->pack_name);
 811         if (!pack_version_ok(hdr.hdr_version))
 812                 return error("packfile %s is version %"PRIu32" and not"
 813                         " supported (try upgrading GIT to a newer version)",
 814                         p->pack_name, ntohl(hdr.hdr_version));
 815
 816         /* Verify the pack matches its index. */
 817         if (p->num_objects != ntohl(hdr.hdr_entries))
 818                 return error("packfile %s claims to have %"PRIu32" objects"
 819                              " while index indicates %"PRIu32" objects",
 820                              p->pack_name, ntohl(hdr.hdr_entries),
 821                              p->num_objects);
 822         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
 823                 return error("end of packfile %s is unavailable", p->pack_name);
 824         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
 825                 return error("packfile %s signature is unavailable", p->pack_name);
 826         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
 827         if (hashcmp(sha1, idx_sha1))
 828                 return error("packfile %s does not match index", p->pack_name);
 829         return 0;
 830 }
 831
 832 static int open_packed_git(struct packed_git *p)
 833 {
 834         if (!open_packed_git_1(p))
 835                 return 0;
 836         if (p->pack_fd != -1) {
 837                 close(p->pack_fd);
 838                 pack_open_fds--;
 839                 p->pack_fd = -1;
 840         }
 841         return -1;
 842 }
 843
 844 static int in_window(struct pack_window *win, off_t offset)
 845 {
 846         /* We must promise at least 20 bytes (one hash) after the
 847          * offset is available from this window, otherwise the offset
 848          * is not actually in this window and a different window (which
 849          * has that one hash excess) must be used.  This is to support
 850          * the object header and delta base parsing routines below.
 851          */
 852         off_t win_off = win->offset;
 853         return win_off <= offset
 854                 && (offset + 20) <= (win_off + win->len);
 855 }
 856
 857 unsigned char *use_pack(struct packed_git *p,
 858                 struct pack_window **w_cursor,
 859                 off_t offset,
 860                 unsigned long *left)
 861 {
 862         struct pack_window *win = *w_cursor;
 863
 864         /* Since packfiles end in a hash of their content and it's
 865          * pointless to ask for an offset into the middle of that
 866          * hash, and the in_window function above wouldn't match
 867          * don't allow an offset too close to the end of the file.
 868          */
 869         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
 870                 die("packfile %s cannot be accessed", p->pack_name);
 871         if (offset > (p->pack_size - 20))
 872                 die("offset beyond end of packfile (truncated pack?)");
 873
 874         if (!win || !in_window(win, offset)) {
 875                 if (win)
 876                         win->inuse_cnt--;
 877                 for (win = p->windows; win; win = win->next) {
 878                         if (in_window(win, offset))
 879                                 break;
 880                 }
 881                 if (!win) {
 882                         size_t window_align = packed_git_window_size / 2;
 883                         off_t len;
 884
 885                         if (p->pack_fd == -1 && open_packed_git(p))
 886                                 die("packfile %s cannot be accessed", p->pack_name);
 887
 888                         win = xcalloc(1, sizeof(*win));
 889                         win->offset = (offset / window_align) * window_align;
 890                         len = p->pack_size - win->offset;
 891                         if (len > packed_git_window_size)
 892                                 len = packed_git_window_size;
 893                         win->len = (size_t)len;
 894                         pack_mapped += win->len;
 895                         while (packed_git_limit < pack_mapped
 896                                 && unuse_one_window(p, p->pack_fd))
 897                                 ; /* nothing */
 898                         win->base = xmmap(NULL, win->len,
 899                                 PROT_READ, MAP_PRIVATE,
 900                                 p->pack_fd, win->offset);
 901                         if (win->base == MAP_FAILED)
 902                                 die("packfile %s cannot be mapped: %s",
 903                                         p->pack_name,
 904                                         strerror(errno));
 905                         if (!win->offset && win->len == p->pack_size
 906                                 && !p->do_not_close) {
 907                                 close(p->pack_fd);
 908                                 pack_open_fds--;
 909                                 p->pack_fd = -1;
 910                         }
 911                         pack_mmap_calls++;
 912                         pack_open_windows++;
 913                         if (pack_mapped > peak_pack_mapped)
 914                                 peak_pack_mapped = pack_mapped;
 915                         if (pack_open_windows > peak_pack_open_windows)
 916                                 peak_pack_open_windows = pack_open_windows;
 917                         win->next = p->windows;
 918                         p->windows = win;
 919                 }
 920         }
 921         if (win != *w_cursor) {
 922                 win->last_used = pack_used_ctr++;
 923                 win->inuse_cnt++;
 924                 *w_cursor = win;
 925         }
 926         offset -= win->offset;
 927         if (left)
 928                 *left = win->len - xsize_t(offset);
 929         return win->base + offset;
 930 }
 931
 932 static struct packed_git *alloc_packed_git(int extra)
 933 {
 934         struct packed_git *p = xmalloc(sizeof(*p) + extra);
 935         memset(p, 0, sizeof(*p));
 936         p->pack_fd = -1;
 937         return p;
 938 }
 939
 940 static void try_to_free_pack_memory(size_t size)
 941 {
 942         release_pack_memory(size, -1);
 943 }
 944
 945 struct packed_git *add_packed_git(const char *path, int path_len, int local)
 946 {
 947         static int have_set_try_to_free_routine;
 948         struct stat st;
 949         struct packed_git *p = alloc_packed_git(path_len + 2);
 950
 951         if (!have_set_try_to_free_routine) {
 952                 have_set_try_to_free_routine = 1;
 953                 set_try_to_free_routine(try_to_free_pack_memory);
 954         }
 955
 956         /*
 957          * Make sure a corresponding .pack file exists and that
 958          * the index looks sane.
 959          */
 960         path_len -= strlen(".idx");
 961         if (path_len < 1) {
 962                 free(p);
 963                 return NULL;
 964         }
 965         memcpy(p->pack_name, path, path_len);
 966
 967         strcpy(p->pack_name + path_len, ".keep");
 968         if (!access(p->pack_name, F_OK))
 969                 p->pack_keep = 1;
 970
 971         strcpy(p->pack_name + path_len, ".pack");
 972         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 973                 free(p);
 974                 return NULL;
 975         }
 976
 977         /* ok, it looks sane as far as we can check without
 978          * actually mapping the pack file.
 979          */
 980         p->pack_size = st.st_size;
 981         p->pack_local = local;
 982         p->mtime = st.st_mtime;
 983         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
 984                 hashclr(p->sha1);
 985         return p;
 986 }
 987
 988 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
 989 {
 990         const char *path = sha1_pack_name(sha1);
 991         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
 992
 993         strcpy(p->pack_name, path);
 994         hashcpy(p->sha1, sha1);
 995         if (check_packed_git_idx(idx_path, p)) {
 996                 free(p);
 997                 return NULL;
 998         }
 999
1000         return p;
1001 }
1002
1003 void install_packed_git(struct packed_git *pack)
1004 {
1005         if (pack->pack_fd != -1)
1006                 pack_open_fds++;
1007
1008         pack->next = packed_git;
1009         packed_git = pack;
1010 }
1011
1012 void (*report_garbage)(const char *desc, const char *path);
1013
1014 static void report_helper(const struct string_list *list,
1015                           int seen_bits, int first, int last)
1016 {
1017         const char *msg;
1018         switch (seen_bits) {
1019         case 0:
1020                 msg = "no corresponding .idx nor .pack";
1021                 break;
1022         case 1:
1023                 msg = "no corresponding .idx";
1024                 break;
1025         case 2:
1026                 msg = "no corresponding .pack";
1027                 break;
1028         default:
1029                 return;
1030         }
1031         for (; first < last; first++)
1032                 report_garbage(msg, list->items[first].string);
1033 }
1034
1035 static void report_pack_garbage(struct string_list *list)
1036 {
1037         int i, baselen = -1, first = 0, seen_bits = 0;
1038
1039         if (!report_garbage)
1040                 return;
1041
1042         sort_string_list(list);
1043
1044         for (i = 0; i < list->nr; i++) {
1045                 const char *path = list->items[i].string;
1046                 if (baselen != -1 &&
1047                     strncmp(path, list->items[first].string, baselen)) {
1048                         report_helper(list, seen_bits, first, i);
1049                         baselen = -1;
1050                         seen_bits = 0;
1051                 }
1052                 if (baselen == -1) {
1053                         const char *dot = strrchr(path, '.');
1054                         if (!dot) {
1055                                 report_garbage("garbage found", path);
1056                                 continue;
1057                         }
1058                         baselen = dot - path + 1;
1059                         first = i;
1060                 }
1061                 if (!strcmp(path + baselen, "pack"))
1062                         seen_bits |= 1;
1063                 else if (!strcmp(path + baselen, "idx"))
1064                         seen_bits |= 2;
1065         }
1066         report_helper(list, seen_bits, first, list->nr);
1067 }
1068
1069 static void prepare_packed_git_one(char *objdir, int local)
1070 {
1071         /* Ensure that this buffer is large enough so that we can
1072            append "/pack/" without clobbering the stack even if
1073            strlen(objdir) were PATH_MAX.  */
1074         char path[PATH_MAX + 1 + 4 + 1 + 1];
1075         int len;
1076         DIR *dir;
1077         struct dirent *de;
1078         struct string_list garbage = STRING_LIST_INIT_DUP;
1079
1080         sprintf(path, "%s/pack", objdir);
1081         len = strlen(path);
1082         dir = opendir(path);
1083         if (!dir) {
1084                 if (errno != ENOENT)
1085                         error("unable to open object pack directory: %s: %s",
1086                               path, strerror(errno));
1087                 return;
1088         }
1089         path[len++] = '/';
1090         while ((de = readdir(dir)) != NULL) {
1091                 int namelen = strlen(de->d_name);
1092                 struct packed_git *p;
1093
1094                 if (len + namelen + 1 > sizeof(path)) {
1095                         if (report_garbage) {
1096                                 struct strbuf sb = STRBUF_INIT;
1097                                 strbuf_addf(&sb, "%.*s/%s", len - 1, path, de->d_name);
1098                                 report_garbage("path too long", sb.buf);
1099                                 strbuf_release(&sb);
1100                         }
1101                         continue;
1102                 }
1103
1104                 if (is_dot_or_dotdot(de->d_name))
1105                         continue;
1106
1107                 strcpy(path + len, de->d_name);
1108
1109                 if (has_extension(de->d_name, ".idx")) {
1110                         /* Don't reopen a pack we already have. */
1111                         for (p = packed_git; p; p = p->next) {
1112                                 if (!memcmp(path, p->pack_name, len + namelen - 4))
1113                                         break;
1114                         }
1115                         if (p == NULL &&
1116                             /*
1117                              * See if it really is a valid .idx file with
1118                              * corresponding .pack file that we can map.
1119                              */
1120                             (p = add_packed_git(path, len + namelen, local)) != NULL)
1121                                 install_packed_git(p);
1122                 }
1123
1124                 if (!report_garbage)
1125                         continue;
1126
1127                 if (has_extension(de->d_name, ".idx") ||
1128                     has_extension(de->d_name, ".pack") ||
1129                     has_extension(de->d_name, ".keep"))
1130                         string_list_append(&garbage, path);
1131                 else
1132                         report_garbage("garbage found", path);
1133         }
1134         closedir(dir);
1135         report_pack_garbage(&garbage);
1136         string_list_clear(&garbage, 0);
1137 }
1138
1139 static int sort_pack(const void *a_, const void *b_)
1140 {
1141         struct packed_git *a = *((struct packed_git **)a_);
1142         struct packed_git *b = *((struct packed_git **)b_);
1143         int st;
1144
1145         /*
1146          * Local packs tend to contain objects specific to our
1147          * variant of the project than remote ones.  In addition,
1148          * remote ones could be on a network mounted filesystem.
1149          * Favor local ones for these reasons.
1150          */
1151         st = a->pack_local - b->pack_local;
1152         if (st)
1153                 return -st;
1154
1155         /*
1156          * Younger packs tend to contain more recent objects,
1157          * and more recent objects tend to get accessed more
1158          * often.
1159          */
1160         if (a->mtime < b->mtime)
1161                 return 1;
1162         else if (a->mtime == b->mtime)
1163                 return 0;
1164         return -1;
1165 }
1166
1167 static void rearrange_packed_git(void)
1168 {
1169         struct packed_git **ary, *p;
1170         int i, n;
1171
1172         for (n = 0, p = packed_git; p; p = p->next)
1173                 n++;
1174         if (n < 2)
1175                 return;
1176
1177         /* prepare an array of packed_git for easier sorting */
1178         ary = xcalloc(n, sizeof(struct packed_git *));
1179         for (n = 0, p = packed_git; p; p = p->next)
1180                 ary[n++] = p;
1181
1182         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1183
1184         /* link them back again */
1185         for (i = 0; i < n - 1; i++)
1186                 ary[i]->next = ary[i + 1];
1187         ary[n - 1]->next = NULL;
1188         packed_git = ary[0];
1189
1190         free(ary);
1191 }
1192
1193 static int prepare_packed_git_run_once = 0;
1194 void prepare_packed_git(void)
1195 {
1196         struct alternate_object_database *alt;
1197
1198         if (prepare_packed_git_run_once)
1199                 return;
1200         prepare_packed_git_one(get_object_directory(), 1);
1201         prepare_alt_odb();
1202         for (alt = alt_odb_list; alt; alt = alt->next) {
1203                 alt->name[-1] = 0;
1204                 prepare_packed_git_one(alt->base, 0);
1205                 alt->name[-1] = '/';
1206         }
1207         rearrange_packed_git();
1208         prepare_packed_git_run_once = 1;
1209 }
1210
1211 void reprepare_packed_git(void)
1212 {
1213         discard_revindex();
1214         prepare_packed_git_run_once = 0;
1215         prepare_packed_git();
1216 }
1217
1218 static void mark_bad_packed_object(struct packed_git *p,
1219                                    const unsigned char *sha1)
1220 {
1221         unsigned i;
1222         for (i = 0; i < p->num_bad_objects; i++)
1223                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1224                         return;
1225         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1226         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1227         p->num_bad_objects++;
1228 }
1229
1230 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1231 {
1232         struct packed_git *p;
1233         unsigned i;
1234
1235         for (p = packed_git; p; p = p->next)
1236                 for (i = 0; i < p->num_bad_objects; i++)
1237                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1238                                 return p;
1239         return NULL;
1240 }
1241
1242 /*
1243  * With an in-core object data in "map", rehash it to make sure the
1244  * object name actually matches "sha1" to detect object corruption.
1245  * With "map" == NULL, try reading the object named with "sha1" using
1246  * the streaming interface and rehash it to do the same.
1247  */
1248 int check_sha1_signature(const unsigned char *sha1, void *map,
1249                          unsigned long size, const char *type)
1250 {
1251         unsigned char real_sha1[20];
1252         enum object_type obj_type;
1253         struct git_istream *st;
1254         git_SHA_CTX c;
1255         char hdr[32];
1256         int hdrlen;
1257
1258         if (map) {
1259                 hash_sha1_file(map, size, type, real_sha1);
1260                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1261         }
1262
1263         st = open_istream(sha1, &obj_type, &size, NULL);
1264         if (!st)
1265                 return -1;
1266
1267         /* Generate the header */
1268         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1269
1270         /* Sha1.. */
1271         git_SHA1_Init(&c);
1272         git_SHA1_Update(&c, hdr, hdrlen);
1273         for (;;) {
1274                 char buf[1024 * 16];
1275                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1276
1277                 if (readlen < 0) {
1278                         close_istream(st);
1279                         return -1;
1280                 }
1281                 if (!readlen)
1282                         break;
1283                 git_SHA1_Update(&c, buf, readlen);
1284         }
1285         git_SHA1_Final(real_sha1, &c);
1286         close_istream(st);
1287         return hashcmp(sha1, real_sha1) ? -1 : 0;
1288 }
1289
1290 static int git_open_noatime(const char *name)
1291 {
1292         static int sha1_file_open_flag = O_NOATIME;
1293
1294         for (;;) {
1295                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1296                 if (fd >= 0)
1297                         return fd;
1298
1299                 /* Might the failure be due to O_NOATIME? */
1300                 if (errno != ENOENT && sha1_file_open_flag) {
1301                         sha1_file_open_flag = 0;
1302                         continue;
1303                 }
1304
1305                 return -1;
1306         }
1307 }
1308
1309 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1310 {
1311         char *name = sha1_file_name(sha1);
1312         struct alternate_object_database *alt;
1313
1314         if (!lstat(name, st))
1315                 return 0;
1316
1317         prepare_alt_odb();
1318         errno = ENOENT;
1319         for (alt = alt_odb_list; alt; alt = alt->next) {
1320                 name = alt->name;
1321                 fill_sha1_path(name, sha1);
1322                 if (!lstat(alt->base, st))
1323                         return 0;
1324         }
1325
1326         return -1;
1327 }
1328
1329 static int open_sha1_file(const unsigned char *sha1)
1330 {
1331         int fd;
1332         char *name = sha1_file_name(sha1);
1333         struct alternate_object_database *alt;
1334
1335         fd = git_open_noatime(name);
1336         if (fd >= 0)
1337                 return fd;
1338
1339         prepare_alt_odb();
1340         errno = ENOENT;
1341         for (alt = alt_odb_list; alt; alt = alt->next) {
1342                 name = alt->name;
1343                 fill_sha1_path(name, sha1);
1344                 fd = git_open_noatime(alt->base);
1345                 if (fd >= 0)
1346                         return fd;
1347         }
1348         return -1;
1349 }
1350
1351 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1352 {
1353         void *map;
1354         int fd;
1355
1356         fd = open_sha1_file(sha1);
1357         map = NULL;
1358         if (fd >= 0) {
1359                 struct stat st;
1360
1361                 if (!fstat(fd, &st)) {
1362                         *size = xsize_t(st.st_size);
1363                         if (!*size) {
1364                                 /* mmap() is forbidden on empty files */
1365                                 error("object file %s is empty", sha1_file_name(sha1));
1366                                 return NULL;
1367                         }
1368                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1369                 }
1370                 close(fd);
1371         }
1372         return map;
1373 }
1374
1375 /*
1376  * There used to be a second loose object header format which
1377  * was meant to mimic the in-pack format, allowing for direct
1378  * copy of the object data.  This format turned up not to be
1379  * really worth it and we no longer write loose objects in that
1380  * format.
1381  */
1382 static int experimental_loose_object(unsigned char *map)
1383 {
1384         unsigned int word;
1385
1386         /*
1387          * We must determine if the buffer contains the standard
1388          * zlib-deflated stream or the experimental format based
1389          * on the in-pack object format. Compare the header byte
1390          * for each format:
1391          *
1392          * RFC1950 zlib w/ deflate : 0www1000 : 0 <= www <= 7
1393          * Experimental pack-based : Stttssss : ttt = 1,2,3,4
1394          *
1395          * If bit 7 is clear and bits 0-3 equal 8, the buffer MUST be
1396          * in standard loose-object format, UNLESS it is a Git-pack
1397          * format object *exactly* 8 bytes in size when inflated.
1398          *
1399          * However, RFC1950 also specifies that the 1st 16-bit word
1400          * must be divisible by 31 - this checksum tells us our buffer
1401          * is in the standard format, giving a false positive only if
1402          * the 1st word of the Git-pack format object happens to be
1403          * divisible by 31, ie:
1404          *      ((byte0 * 256) + byte1) % 31 = 0
1405          *   =>        0ttt10000www1000 % 31 = 0
1406          *
1407          * As it happens, this case can only arise for www=3 & ttt=1
1408          * - ie, a Commit object, which would have to be 8 bytes in
1409          * size. As no Commit can be that small, we find that the
1410          * combination of these two criteria (bitmask & checksum)
1411          * can always correctly determine the buffer format.
1412          */
1413         word = (map[0] << 8) + map[1];
1414         if ((map[0] & 0x8F) == 0x08 && !(word % 31))
1415                 return 0;
1416         else
1417                 return 1;
1418 }
1419
1420 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1421                 unsigned long len, enum object_type *type, unsigned long *sizep)
1422 {
1423         unsigned shift;
1424         unsigned long size, c;
1425         unsigned long used = 0;
1426
1427         c = buf[used++];
1428         *type = (c >> 4) & 7;
1429         size = c & 15;
1430         shift = 4;
1431         while (c & 0x80) {
1432                 if (len <= used || bitsizeof(long) <= shift) {
1433                         error("bad object header");
1434                         size = used = 0;
1435                         break;
1436                 }
1437                 c = buf[used++];
1438                 size += (c & 0x7f) << shift;
1439                 shift += 7;
1440         }
1441         *sizep = size;
1442         return used;
1443 }
1444
1445 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1446 {
1447         unsigned long size, used;
1448         static const char valid_loose_object_type[8] = {
1449                 0, /* OBJ_EXT */
1450                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1451                 0, /* "delta" and others are invalid in a loose object */
1452         };
1453         enum object_type type;
1454
1455         /* Get the data stream */
1456         memset(stream, 0, sizeof(*stream));
1457         stream->next_in = map;
1458         stream->avail_in = mapsize;
1459         stream->next_out = buffer;
1460         stream->avail_out = bufsiz;
1461
1462         if (experimental_loose_object(map)) {
1463                 /*
1464                  * The old experimental format we no longer produce;
1465                  * we can still read it.
1466                  */
1467                 used = unpack_object_header_buffer(map, mapsize, &type, &size);
1468                 if (!used || !valid_loose_object_type[type])
1469                         return -1;
1470                 map += used;
1471                 mapsize -= used;
1472
1473                 /* Set up the stream for the rest.. */
1474                 stream->next_in = map;
1475                 stream->avail_in = mapsize;
1476                 git_inflate_init(stream);
1477
1478                 /* And generate the fake traditional header */
1479                 stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1480                                                  typename(type), size);
1481                 return 0;
1482         }
1483         git_inflate_init(stream);
1484         return git_inflate(stream, 0);
1485 }
1486
1487 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1488 {
1489         int bytes = strlen(buffer) + 1;
1490         unsigned char *buf = xmallocz(size);
1491         unsigned long n;
1492         int status = Z_OK;
1493
1494         n = stream->total_out - bytes;
1495         if (n > size)
1496                 n = size;
1497         memcpy(buf, (char *) buffer + bytes, n);
1498         bytes = n;
1499         if (bytes <= size) {
1500                 /*
1501                  * The above condition must be (bytes <= size), not
1502                  * (bytes < size).  In other words, even though we
1503                  * expect no more output and set avail_out to zero,
1504                  * the input zlib stream may have bytes that express
1505                  * "this concludes the stream", and we *do* want to
1506                  * eat that input.
1507                  *
1508                  * Otherwise we would not be able to test that we
1509                  * consumed all the input to reach the expected size;
1510                  * we also want to check that zlib tells us that all
1511                  * went well with status == Z_STREAM_END at the end.
1512                  */
1513                 stream->next_out = buf + bytes;
1514                 stream->avail_out = size - bytes;
1515                 while (status == Z_OK)
1516                         status = git_inflate(stream, Z_FINISH);
1517         }
1518         if (status == Z_STREAM_END && !stream->avail_in) {
1519                 git_inflate_end(stream);
1520                 return buf;
1521         }
1522
1523         if (status < 0)
1524                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1525         else if (stream->avail_in)
1526                 error("garbage at end of loose object '%s'",
1527                       sha1_to_hex(sha1));
1528         free(buf);
1529         return NULL;
1530 }
1531
1532 /*
1533  * We used to just use "sscanf()", but that's actually way
1534  * too permissive for what we want to check. So do an anal
1535  * object header parse by hand.
1536  */
1537 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1538 {
1539         char type[10];
1540         int i;
1541         unsigned long size;
1542
1543         /*
1544          * The type can be at most ten bytes (including the
1545          * terminating '\0' that we add), and is followed by
1546          * a space.
1547          */
1548         i = 0;
1549         for (;;) {
1550                 char c = *hdr++;
1551                 if (c == ' ')
1552                         break;
1553                 type[i++] = c;
1554                 if (i >= sizeof(type))
1555                         return -1;
1556         }
1557         type[i] = 0;
1558
1559         /*
1560          * The length must follow immediately, and be in canonical
1561          * decimal format (ie "010" is not valid).
1562          */
1563         size = *hdr++ - '0';
1564         if (size > 9)
1565                 return -1;
1566         if (size) {
1567                 for (;;) {
1568                         unsigned long c = *hdr - '0';
1569                         if (c > 9)
1570                                 break;
1571                         hdr++;
1572                         size = size * 10 + c;
1573                 }
1574         }
1575         *sizep = size;
1576
1577         /*
1578          * The length must be followed by a zero byte
1579          */
1580         return *hdr ? -1 : type_from_string(type);
1581 }
1582
1583 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1584 {
1585         int ret;
1586         git_zstream stream;
1587         char hdr[8192];
1588
1589         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1590         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1591                 return NULL;
1592
1593         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1594 }
1595
1596 unsigned long get_size_from_delta(struct packed_git *p,
1597                                   struct pack_window **w_curs,
1598                                   off_t curpos)
1599 {
1600         const unsigned char *data;
1601         unsigned char delta_head[20], *in;
1602         git_zstream stream;
1603         int st;
1604
1605         memset(&stream, 0, sizeof(stream));
1606         stream.next_out = delta_head;
1607         stream.avail_out = sizeof(delta_head);
1608
1609         git_inflate_init(&stream);
1610         do {
1611                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1612                 stream.next_in = in;
1613                 st = git_inflate(&stream, Z_FINISH);
1614                 curpos += stream.next_in - in;
1615         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1616                  stream.total_out < sizeof(delta_head));
1617         git_inflate_end(&stream);
1618         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1619                 error("delta data unpack-initial failed");
1620                 return 0;
1621         }
1622
1623         /* Examine the initial part of the delta to figure out
1624          * the result size.
1625          */
1626         data = delta_head;
1627
1628         /* ignore base size */
1629         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1630
1631         /* Read the result size */
1632         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1633 }
1634
1635 static off_t get_delta_base(struct packed_git *p,
1636                                     struct pack_window **w_curs,
1637                                     off_t *curpos,
1638                                     enum object_type type,
1639                                     off_t delta_obj_offset)
1640 {
1641         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1642         off_t base_offset;
1643
1644         /* use_pack() assured us we have [base_info, base_info + 20)
1645          * as a range that we can look at without walking off the
1646          * end of the mapped window.  Its actually the hash size
1647          * that is assured.  An OFS_DELTA longer than the hash size
1648          * is stupid, as then a REF_DELTA would be smaller to store.
1649          */
1650         if (type == OBJ_OFS_DELTA) {
1651                 unsigned used = 0;
1652                 unsigned char c = base_info[used++];
1653                 base_offset = c & 127;
1654                 while (c & 128) {
1655                         base_offset += 1;
1656                         if (!base_offset || MSB(base_offset, 7))
1657                                 return 0;  /* overflow */
1658                         c = base_info[used++];
1659                         base_offset = (base_offset << 7) + (c & 127);
1660                 }
1661                 base_offset = delta_obj_offset - base_offset;
1662                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1663                         return 0;  /* out of bound */
1664                 *curpos += used;
1665         } else if (type == OBJ_REF_DELTA) {
1666                 /* The base entry _must_ be in the same pack */
1667                 base_offset = find_pack_entry_one(base_info, p);
1668                 *curpos += 20;
1669         } else
1670                 die("I am totally screwed");
1671         return base_offset;
1672 }
1673
1674 int unpack_object_header(struct packed_git *p,
1675                          struct pack_window **w_curs,
1676                          off_t *curpos,
1677                          unsigned long *sizep)
1678 {
1679         unsigned char *base;
1680         unsigned long left;
1681         unsigned long used;
1682         enum object_type type;
1683
1684         /* use_pack() assures us we have [base, base + 20) available
1685          * as a range that we can look at.  (Its actually the hash
1686          * size that is assured.)  With our object header encoding
1687          * the maximum deflated object size is 2^137, which is just
1688          * insane, so we know won't exceed what we have been given.
1689          */
1690         base = use_pack(p, w_curs, *curpos, &left);
1691         used = unpack_object_header_buffer(base, left, &type, sizep);
1692         if (!used) {
1693                 type = OBJ_BAD;
1694         } else
1695                 *curpos += used;
1696
1697         return type;
1698 }
1699
1700 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1701 {
1702         int type;
1703         struct revindex_entry *revidx;
1704         const unsigned char *sha1;
1705         revidx = find_pack_revindex(p, obj_offset);
1706         if (!revidx)
1707                 return OBJ_BAD;
1708         sha1 = nth_packed_object_sha1(p, revidx->nr);
1709         mark_bad_packed_object(p, sha1);
1710         type = sha1_object_info(sha1, NULL);
1711         if (type <= OBJ_NONE)
1712                 return OBJ_BAD;
1713         return type;
1714 }
1715
1716 #define POI_STACK_PREALLOC 64
1717
1718 static enum object_type packed_to_object_type(struct packed_git *p,
1719                                               off_t obj_offset,
1720                                               enum object_type type,
1721                                               struct pack_window **w_curs,
1722                                               off_t curpos)
1723 {
1724         off_t small_poi_stack[POI_STACK_PREALLOC];
1725         off_t *poi_stack = small_poi_stack;
1726         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1727
1728         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1729                 off_t base_offset;
1730                 unsigned long size;
1731                 /* Push the object we're going to leave behind */
1732                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1733                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1734                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1735                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1736                 } else {
1737                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1738                 }
1739                 poi_stack[poi_stack_nr++] = obj_offset;
1740                 /* If parsing the base offset fails, just unwind */
1741                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1742                 if (!base_offset)
1743                         goto unwind;
1744                 curpos = obj_offset = base_offset;
1745                 type = unpack_object_header(p, w_curs, &curpos, &size);
1746                 if (type <= OBJ_NONE) {
1747                         /* If getting the base itself fails, we first
1748                          * retry the base, otherwise unwind */
1749                         type = retry_bad_packed_offset(p, base_offset);
1750                         if (type > OBJ_NONE)
1751                                 goto out;
1752                         goto unwind;
1753                 }
1754         }
1755
1756         switch (type) {
1757         case OBJ_BAD:
1758         case OBJ_COMMIT:
1759         case OBJ_TREE:
1760         case OBJ_BLOB:
1761         case OBJ_TAG:
1762                 break;
1763         default:
1764                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1765                       type, (uintmax_t)obj_offset, p->pack_name);
1766                 type = OBJ_BAD;
1767         }
1768
1769 out:
1770         if (poi_stack != small_poi_stack)
1771                 free(poi_stack);
1772         return type;
1773
1774 unwind:
1775         while (poi_stack_nr) {
1776                 obj_offset = poi_stack[--poi_stack_nr];
1777                 type = retry_bad_packed_offset(p, obj_offset);
1778                 if (type > OBJ_NONE)
1779                         goto out;
1780         }
1781         type = OBJ_BAD;
1782         goto out;
1783 }
1784
1785 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1786                               struct object_info *oi)
1787 {
1788         struct pack_window *w_curs = NULL;
1789         unsigned long size;
1790         off_t curpos = obj_offset;
1791         enum object_type type;
1792
1793         /*
1794          * We always get the representation type, but only convert it to
1795          * a "real" type later if the caller is interested.
1796          */
1797         type = unpack_object_header(p, &w_curs, &curpos, &size);
1798
1799         if (oi->sizep) {
1800                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1801                         off_t tmp_pos = curpos;
1802                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1803                                                            type, obj_offset);
1804                         if (!base_offset) {
1805                                 type = OBJ_BAD;
1806                                 goto out;
1807                         }
1808                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1809                         if (*oi->sizep == 0) {
1810                                 type = OBJ_BAD;
1811                                 goto out;
1812                         }
1813                 } else {
1814                         *oi->sizep = size;
1815                 }
1816         }
1817
1818         if (oi->disk_sizep) {
1819                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1820                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1821         }
1822
1823         if (oi->typep) {
1824                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
1825                 if (*oi->typep < 0) {
1826                         type = OBJ_BAD;
1827                         goto out;
1828                 }
1829         }
1830
1831 out:
1832         unuse_pack(&w_curs);
1833         return type;
1834 }
1835
1836 static void *unpack_compressed_entry(struct packed_git *p,
1837                                     struct pack_window **w_curs,
1838                                     off_t curpos,
1839                                     unsigned long size)
1840 {
1841         int st;
1842         git_zstream stream;
1843         unsigned char *buffer, *in;
1844
1845         buffer = xmallocz(size);
1846         memset(&stream, 0, sizeof(stream));
1847         stream.next_out = buffer;
1848         stream.avail_out = size + 1;
1849
1850         git_inflate_init(&stream);
1851         do {
1852                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1853                 stream.next_in = in;
1854                 st = git_inflate(&stream, Z_FINISH);
1855                 if (!stream.avail_out)
1856                         break; /* the payload is larger than it should be */
1857                 curpos += stream.next_in - in;
1858         } while (st == Z_OK || st == Z_BUF_ERROR);
1859         git_inflate_end(&stream);
1860         if ((st != Z_STREAM_END) || stream.total_out != size) {
1861                 free(buffer);
1862                 return NULL;
1863         }
1864
1865         return buffer;
1866 }
1867
1868 #define MAX_DELTA_CACHE (256)
1869
1870 static size_t delta_base_cached;
1871
1872 static struct delta_base_cache_lru_list {
1873         struct delta_base_cache_lru_list *prev;
1874         struct delta_base_cache_lru_list *next;
1875 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1876
1877 static struct delta_base_cache_entry {
1878         struct delta_base_cache_lru_list lru;
1879         void *data;
1880         struct packed_git *p;
1881         off_t base_offset;
1882         unsigned long size;
1883         enum object_type type;
1884 } delta_base_cache[MAX_DELTA_CACHE];
1885
1886 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1887 {
1888         unsigned long hash;
1889
1890         hash = (unsigned long)p + (unsigned long)base_offset;
1891         hash += (hash >> 8) + (hash >> 16);
1892         return hash % MAX_DELTA_CACHE;
1893 }
1894
1895 static struct delta_base_cache_entry *
1896 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1897 {
1898         unsigned long hash = pack_entry_hash(p, base_offset);
1899         return delta_base_cache + hash;
1900 }
1901
1902 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
1903                                      struct packed_git *p, off_t base_offset)
1904 {
1905         return (ent->data && ent->p == p && ent->base_offset == base_offset);
1906 }
1907
1908 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1909 {
1910         struct delta_base_cache_entry *ent;
1911         ent = get_delta_base_cache_entry(p, base_offset);
1912         return eq_delta_base_cache_entry(ent, p, base_offset);
1913 }
1914
1915 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1916 {
1917         ent->data = NULL;
1918         ent->lru.next->prev = ent->lru.prev;
1919         ent->lru.prev->next = ent->lru.next;
1920         delta_base_cached -= ent->size;
1921 }
1922
1923 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1924         unsigned long *base_size, enum object_type *type, int keep_cache)
1925 {
1926         struct delta_base_cache_entry *ent;
1927         void *ret;
1928
1929         ent = get_delta_base_cache_entry(p, base_offset);
1930
1931         if (!eq_delta_base_cache_entry(ent, p, base_offset))
1932                 return unpack_entry(p, base_offset, type, base_size);
1933
1934         ret = ent->data;
1935
1936         if (!keep_cache)
1937                 clear_delta_base_cache_entry(ent);
1938         else
1939                 ret = xmemdupz(ent->data, ent->size);
1940         *type = ent->type;
1941         *base_size = ent->size;
1942         return ret;
1943 }
1944
1945 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1946 {
1947         if (ent->data) {
1948                 free(ent->data);
1949                 ent->data = NULL;
1950                 ent->lru.next->prev = ent->lru.prev;
1951                 ent->lru.prev->next = ent->lru.next;
1952                 delta_base_cached -= ent->size;
1953         }
1954 }
1955
1956 void clear_delta_base_cache(void)
1957 {
1958         unsigned long p;
1959         for (p = 0; p < MAX_DELTA_CACHE; p++)
1960                 release_delta_base_cache(&delta_base_cache[p]);
1961 }
1962
1963 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1964         void *base, unsigned long base_size, enum object_type type)
1965 {
1966         unsigned long hash = pack_entry_hash(p, base_offset);
1967         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1968         struct delta_base_cache_lru_list *lru;
1969
1970         release_delta_base_cache(ent);
1971         delta_base_cached += base_size;
1972
1973         for (lru = delta_base_cache_lru.next;
1974              delta_base_cached > delta_base_cache_limit
1975              && lru != &delta_base_cache_lru;
1976              lru = lru->next) {
1977                 struct delta_base_cache_entry *f = (void *)lru;
1978                 if (f->type == OBJ_BLOB)
1979                         release_delta_base_cache(f);
1980         }
1981         for (lru = delta_base_cache_lru.next;
1982              delta_base_cached > delta_base_cache_limit
1983              && lru != &delta_base_cache_lru;
1984              lru = lru->next) {
1985                 struct delta_base_cache_entry *f = (void *)lru;
1986                 release_delta_base_cache(f);
1987         }
1988
1989         ent->p = p;
1990         ent->base_offset = base_offset;
1991         ent->type = type;
1992         ent->data = base;
1993         ent->size = base_size;
1994         ent->lru.next = &delta_base_cache_lru;
1995         ent->lru.prev = delta_base_cache_lru.prev;
1996         delta_base_cache_lru.prev->next = &ent->lru;
1997         delta_base_cache_lru.prev = &ent->lru;
1998 }
1999
2000 static void *read_object(const unsigned char *sha1, enum object_type *type,
2001                          unsigned long *size);
2002
2003 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2004 {
2005         static FILE *log_file;
2006
2007         if (!log_pack_access)
2008                 log_pack_access = getenv("GIT_TRACE_PACK_ACCESS");
2009         if (!log_pack_access)
2010                 log_pack_access = no_log_pack_access;
2011         if (log_pack_access == no_log_pack_access)
2012                 return;
2013
2014         if (!log_file) {
2015                 log_file = fopen(log_pack_access, "w");
2016                 if (!log_file) {
2017                         error("cannot open pack access log '%s' for writing: %s",
2018                               log_pack_access, strerror(errno));
2019                         log_pack_access = no_log_pack_access;
2020                         return;
2021                 }
2022         }
2023         fprintf(log_file, "%s %"PRIuMAX"\n",
2024                 p->pack_name, (uintmax_t)obj_offset);
2025         fflush(log_file);
2026 }
2027
2028 int do_check_packed_object_crc;
2029
2030 #define UNPACK_ENTRY_STACK_PREALLOC 64
2031 struct unpack_entry_stack_ent {
2032         off_t obj_offset;
2033         off_t curpos;
2034         unsigned long size;
2035 };
2036
2037 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2038                    enum object_type *final_type, unsigned long *final_size)
2039 {
2040         struct pack_window *w_curs = NULL;
2041         off_t curpos = obj_offset;
2042         void *data = NULL;
2043         unsigned long size;
2044         enum object_type type;
2045         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2046         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2047         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2048         int base_from_cache = 0;
2049
2050         if (log_pack_access != no_log_pack_access)
2051                 write_pack_access_log(p, obj_offset);
2052
2053         /* PHASE 1: drill down to the innermost base object */
2054         for (;;) {
2055                 off_t base_offset;
2056                 int i;
2057                 struct delta_base_cache_entry *ent;
2058
2059                 if (do_check_packed_object_crc && p->index_version > 1) {
2060                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2061                         unsigned long len = revidx[1].offset - obj_offset;
2062                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2063                                 const unsigned char *sha1 =
2064                                         nth_packed_object_sha1(p, revidx->nr);
2065                                 error("bad packed object CRC for %s",
2066                                       sha1_to_hex(sha1));
2067                                 mark_bad_packed_object(p, sha1);
2068                                 unuse_pack(&w_curs);
2069                                 return NULL;
2070                         }
2071                 }
2072
2073                 ent = get_delta_base_cache_entry(p, curpos);
2074                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2075                         type = ent->type;
2076                         data = ent->data;
2077                         size = ent->size;
2078                         clear_delta_base_cache_entry(ent);
2079                         base_from_cache = 1;
2080                         break;
2081                 }
2082
2083                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2084                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2085                         break;
2086
2087                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2088                 if (!base_offset) {
2089                         error("failed to validate delta base reference "
2090                               "at offset %"PRIuMAX" from %s",
2091                               (uintmax_t)curpos, p->pack_name);
2092                         /* bail to phase 2, in hopes of recovery */
2093                         data = NULL;
2094                         break;
2095                 }
2096
2097                 /* push object, proceed to base */
2098                 if (delta_stack_nr >= delta_stack_alloc
2099                     && delta_stack == small_delta_stack) {
2100                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2101                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2102                         memcpy(delta_stack, small_delta_stack,
2103                                sizeof(*delta_stack)*delta_stack_nr);
2104                 } else {
2105                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2106                 }
2107                 i = delta_stack_nr++;
2108                 delta_stack[i].obj_offset = obj_offset;
2109                 delta_stack[i].curpos = curpos;
2110                 delta_stack[i].size = size;
2111
2112                 curpos = obj_offset = base_offset;
2113         }
2114
2115         /* PHASE 2: handle the base */
2116         switch (type) {
2117         case OBJ_OFS_DELTA:
2118         case OBJ_REF_DELTA:
2119                 if (data)
2120                         die("BUG in unpack_entry: left loop at a valid delta");
2121                 break;
2122         case OBJ_COMMIT:
2123         case OBJ_TREE:
2124         case OBJ_BLOB:
2125         case OBJ_TAG:
2126                 if (!base_from_cache)
2127                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2128                 break;
2129         default:
2130                 data = NULL;
2131                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2132                       type, (uintmax_t)obj_offset, p->pack_name);
2133         }
2134
2135         /* PHASE 3: apply deltas in order */
2136
2137         /* invariants:
2138          *   'data' holds the base data, or NULL if there was corruption
2139          */
2140         while (delta_stack_nr) {
2141                 void *delta_data;
2142                 void *base = data;
2143                 unsigned long delta_size, base_size = size;
2144                 int i;
2145
2146                 data = NULL;
2147
2148                 if (base)
2149                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2150
2151                 if (!base) {
2152                         /*
2153                          * We're probably in deep shit, but let's try to fetch
2154                          * the required base anyway from another pack or loose.
2155                          * This is costly but should happen only in the presence
2156                          * of a corrupted pack, and is better than failing outright.
2157                          */
2158                         struct revindex_entry *revidx;
2159                         const unsigned char *base_sha1;
2160                         revidx = find_pack_revindex(p, obj_offset);
2161                         if (revidx) {
2162                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2163                                 error("failed to read delta base object %s"
2164                                       " at offset %"PRIuMAX" from %s",
2165                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2166                                       p->pack_name);
2167                                 mark_bad_packed_object(p, base_sha1);
2168                                 base = read_object(base_sha1, &type, &base_size);
2169                         }
2170                 }
2171
2172                 i = --delta_stack_nr;
2173                 obj_offset = delta_stack[i].obj_offset;
2174                 curpos = delta_stack[i].curpos;
2175                 delta_size = delta_stack[i].size;
2176
2177                 if (!base)
2178                         continue;
2179
2180                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2181
2182                 if (!delta_data) {
2183                         error("failed to unpack compressed delta "
2184                               "at offset %"PRIuMAX" from %s",
2185                               (uintmax_t)curpos, p->pack_name);
2186                         data = NULL;
2187                         continue;
2188                 }
2189
2190                 data = patch_delta(base, base_size,
2191                                    delta_data, delta_size,
2192                                    &size);
2193
2194                 /*
2195                  * We could not apply the delta; warn the user, but keep going.
2196                  * Our failure will be noticed either in the next iteration of
2197                  * the loop, or if this is the final delta, in the caller when
2198                  * we return NULL. Those code paths will take care of making
2199                  * a more explicit warning and retrying with another copy of
2200                  * the object.
2201                  */
2202                 if (!data)
2203                         error("failed to apply delta");
2204
2205                 free(delta_data);
2206         }
2207
2208         *final_type = type;
2209         *final_size = size;
2210
2211         unuse_pack(&w_curs);
2212         return data;
2213 }
2214
2215 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2216                                             uint32_t n)
2217 {
2218         const unsigned char *index = p->index_data;
2219         if (!index) {
2220                 if (open_pack_index(p))
2221                         return NULL;
2222                 index = p->index_data;
2223         }
2224         if (n >= p->num_objects)
2225                 return NULL;
2226         index += 4 * 256;
2227         if (p->index_version == 1) {
2228                 return index + 24 * n + 4;
2229         } else {
2230                 index += 8;
2231                 return index + 20 * n;
2232         }
2233 }
2234
2235 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2236 {
2237         const unsigned char *index = p->index_data;
2238         index += 4 * 256;
2239         if (p->index_version == 1) {
2240                 return ntohl(*((uint32_t *)(index + 24 * n)));
2241         } else {
2242                 uint32_t off;
2243                 index += 8 + p->num_objects * (20 + 4);
2244                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2245                 if (!(off & 0x80000000))
2246                         return off;
2247                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2248                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2249                                    ntohl(*((uint32_t *)(index + 4)));
2250         }
2251 }
2252
2253 off_t find_pack_entry_one(const unsigned char *sha1,
2254                                   struct packed_git *p)
2255 {
2256         const uint32_t *level1_ofs = p->index_data;
2257         const unsigned char *index = p->index_data;
2258         unsigned hi, lo, stride;
2259         static int use_lookup = -1;
2260         static int debug_lookup = -1;
2261
2262         if (debug_lookup < 0)
2263                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2264
2265         if (!index) {
2266                 if (open_pack_index(p))
2267                         return 0;
2268                 level1_ofs = p->index_data;
2269                 index = p->index_data;
2270         }
2271         if (p->index_version > 1) {
2272                 level1_ofs += 2;
2273                 index += 8;
2274         }
2275         index += 4 * 256;
2276         hi = ntohl(level1_ofs[*sha1]);
2277         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2278         if (p->index_version > 1) {
2279                 stride = 20;
2280         } else {
2281                 stride = 24;
2282                 index += 4;
2283         }
2284
2285         if (debug_lookup)
2286                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2287                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2288
2289         if (use_lookup < 0)
2290                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2291         if (use_lookup) {
2292                 int pos = sha1_entry_pos(index, stride, 0,
2293                                          lo, hi, p->num_objects, sha1);
2294                 if (pos < 0)
2295                         return 0;
2296                 return nth_packed_object_offset(p, pos);
2297         }
2298
2299         do {
2300                 unsigned mi = (lo + hi) / 2;
2301                 int cmp = hashcmp(index + mi * stride, sha1);
2302
2303                 if (debug_lookup)
2304                         printf("lo %u hi %u rg %u mi %u\n",
2305                                lo, hi, hi - lo, mi);
2306                 if (!cmp)
2307                         return nth_packed_object_offset(p, mi);
2308                 if (cmp > 0)
2309                         hi = mi;
2310                 else
2311                         lo = mi+1;
2312         } while (lo < hi);
2313         return 0;
2314 }
2315
2316 int is_pack_valid(struct packed_git *p)
2317 {
2318         /* An already open pack is known to be valid. */
2319         if (p->pack_fd != -1)
2320                 return 1;
2321
2322         /* If the pack has one window completely covering the
2323          * file size, the pack is known to be valid even if
2324          * the descriptor is not currently open.
2325          */
2326         if (p->windows) {
2327                 struct pack_window *w = p->windows;
2328
2329                 if (!w->offset && w->len == p->pack_size)
2330                         return 1;
2331         }
2332
2333         /* Force the pack to open to prove its valid. */
2334         return !open_packed_git(p);
2335 }
2336
2337 static int fill_pack_entry(const unsigned char *sha1,
2338                            struct pack_entry *e,
2339                            struct packed_git *p)
2340 {
2341         off_t offset;
2342
2343         if (p->num_bad_objects) {
2344                 unsigned i;
2345                 for (i = 0; i < p->num_bad_objects; i++)
2346                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2347                                 return 0;
2348         }
2349
2350         offset = find_pack_entry_one(sha1, p);
2351         if (!offset)
2352                 return 0;
2353
2354         /*
2355          * We are about to tell the caller where they can locate the
2356          * requested object.  We better make sure the packfile is
2357          * still here and can be accessed before supplying that
2358          * answer, as it may have been deleted since the index was
2359          * loaded!
2360          */
2361         if (!is_pack_valid(p)) {
2362                 warning("packfile %s cannot be accessed", p->pack_name);
2363                 return 0;
2364         }
2365         e->offset = offset;
2366         e->p = p;
2367         hashcpy(e->sha1, sha1);
2368         return 1;
2369 }
2370
2371 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2372 {
2373         struct packed_git *p;
2374
2375         prepare_packed_git();
2376         if (!packed_git)
2377                 return 0;
2378
2379         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2380                 return 1;
2381
2382         for (p = packed_git; p; p = p->next) {
2383                 if (p == last_found_pack || !fill_pack_entry(sha1, e, p))
2384                         continue;
2385
2386                 last_found_pack = p;
2387                 return 1;
2388         }
2389         return 0;
2390 }
2391
2392 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2393                                   struct packed_git *packs)
2394 {
2395         struct packed_git *p;
2396
2397         for (p = packs; p; p = p->next) {
2398                 if (find_pack_entry_one(sha1, p))
2399                         return p;
2400         }
2401         return NULL;
2402
2403 }
2404
2405 static int sha1_loose_object_info(const unsigned char *sha1,
2406                                   struct object_info *oi)
2407 {
2408         int status;
2409         unsigned long mapsize, size;
2410         void *map;
2411         git_zstream stream;
2412         char hdr[32];
2413
2414         /*
2415          * If we don't care about type or size, then we don't
2416          * need to look inside the object at all.
2417          */
2418         if (!oi->typep && !oi->sizep) {
2419                 if (oi->disk_sizep) {
2420                         struct stat st;
2421                         if (stat_sha1_file(sha1, &st) < 0)
2422                                 return -1;
2423                         *oi->disk_sizep = st.st_size;
2424                 }
2425                 return 0;
2426         }
2427
2428         map = map_sha1_file(sha1, &mapsize);
2429         if (!map)
2430                 return -1;
2431         if (oi->disk_sizep)
2432                 *oi->disk_sizep = mapsize;
2433         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2434                 status = error("unable to unpack %s header",
2435                                sha1_to_hex(sha1));
2436         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2437                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2438         else if (oi->sizep)
2439                 *oi->sizep = size;
2440         git_inflate_end(&stream);
2441         munmap(map, mapsize);
2442         if (oi->typep)
2443                 *oi->typep = status;
2444         return 0;
2445 }
2446
2447 /* returns enum object_type or negative */
2448 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi)
2449 {
2450         struct cached_object *co;
2451         struct pack_entry e;
2452         int rtype;
2453
2454         co = find_cached_object(sha1);
2455         if (co) {
2456                 if (oi->typep)
2457                         *(oi->typep) = co->type;
2458                 if (oi->sizep)
2459                         *(oi->sizep) = co->size;
2460                 if (oi->disk_sizep)
2461                         *(oi->disk_sizep) = 0;
2462                 oi->whence = OI_CACHED;
2463                 return 0;
2464         }
2465
2466         if (!find_pack_entry(sha1, &e)) {
2467                 /* Most likely it's a loose object. */
2468                 if (!sha1_loose_object_info(sha1, oi)) {
2469                         oi->whence = OI_LOOSE;
2470                         return 0;
2471                 }
2472
2473                 /* Not a loose object; someone else may have just packed it. */
2474                 reprepare_packed_git();
2475                 if (!find_pack_entry(sha1, &e))
2476                         return -1;
2477         }
2478
2479         rtype = packed_object_info(e.p, e.offset, oi);
2480         if (rtype < 0) {
2481                 mark_bad_packed_object(e.p, sha1);
2482                 return sha1_object_info_extended(sha1, oi);
2483         } else if (in_delta_base_cache(e.p, e.offset)) {
2484                 oi->whence = OI_DBCACHED;
2485         } else {
2486                 oi->whence = OI_PACKED;
2487                 oi->u.packed.offset = e.offset;
2488                 oi->u.packed.pack = e.p;
2489                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2490                                          rtype == OBJ_OFS_DELTA);
2491         }
2492
2493         return 0;
2494 }
2495
2496 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2497 {
2498         enum object_type type;
2499         struct object_info oi = {NULL};
2500
2501         oi.typep = &type;
2502         oi.sizep = sizep;
2503         if (sha1_object_info_extended(sha1, &oi) < 0)
2504                 return -1;
2505         return type;
2506 }
2507
2508 static void *read_packed_sha1(const unsigned char *sha1,
2509                               enum object_type *type, unsigned long *size)
2510 {
2511         struct pack_entry e;
2512         void *data;
2513
2514         if (!find_pack_entry(sha1, &e))
2515                 return NULL;
2516         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2517         if (!data) {
2518                 /*
2519                  * We're probably in deep shit, but let's try to fetch
2520                  * the required object anyway from another pack or loose.
2521                  * This should happen only in the presence of a corrupted
2522                  * pack, and is better than failing outright.
2523                  */
2524                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2525                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2526                 mark_bad_packed_object(e.p, sha1);
2527                 data = read_object(sha1, type, size);
2528         }
2529         return data;
2530 }
2531
2532 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2533                       unsigned char *sha1)
2534 {
2535         struct cached_object *co;
2536
2537         hash_sha1_file(buf, len, typename(type), sha1);
2538         if (has_sha1_file(sha1) || find_cached_object(sha1))
2539                 return 0;
2540         if (cached_object_alloc <= cached_object_nr) {
2541                 cached_object_alloc = alloc_nr(cached_object_alloc);
2542                 cached_objects = xrealloc(cached_objects,
2543                                           sizeof(*cached_objects) *
2544                                           cached_object_alloc);
2545         }
2546         co = &cached_objects[cached_object_nr++];
2547         co->size = len;
2548         co->type = type;
2549         co->buf = xmalloc(len);
2550         memcpy(co->buf, buf, len);
2551         hashcpy(co->sha1, sha1);
2552         return 0;
2553 }
2554
2555 static void *read_object(const unsigned char *sha1, enum object_type *type,
2556                          unsigned long *size)
2557 {
2558         unsigned long mapsize;
2559         void *map, *buf;
2560         struct cached_object *co;
2561
2562         co = find_cached_object(sha1);
2563         if (co) {
2564                 *type = co->type;
2565                 *size = co->size;
2566                 return xmemdupz(co->buf, co->size);
2567         }
2568
2569         buf = read_packed_sha1(sha1, type, size);
2570         if (buf)
2571                 return buf;
2572         map = map_sha1_file(sha1, &mapsize);
2573         if (map) {
2574                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2575                 munmap(map, mapsize);
2576                 return buf;
2577         }
2578         reprepare_packed_git();
2579         return read_packed_sha1(sha1, type, size);
2580 }
2581
2582 /*
2583  * This function dies on corrupt objects; the callers who want to
2584  * deal with them should arrange to call read_object() and give error
2585  * messages themselves.
2586  */
2587 void *read_sha1_file_extended(const unsigned char *sha1,
2588                               enum object_type *type,
2589                               unsigned long *size,
2590                               unsigned flag)
2591 {
2592         void *data;
2593         char *path;
2594         const struct packed_git *p;
2595         const unsigned char *repl = (flag & READ_SHA1_FILE_REPLACE)
2596                 ? lookup_replace_object(sha1) : sha1;
2597
2598         errno = 0;
2599         data = read_object(repl, type, size);
2600         if (data)
2601                 return data;
2602
2603         if (errno && errno != ENOENT)
2604                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2605
2606         /* die if we replaced an object with one that does not exist */
2607         if (repl != sha1)
2608                 die("replacement %s not found for %s",
2609                     sha1_to_hex(repl), sha1_to_hex(sha1));
2610
2611         if (has_loose_object(repl)) {
2612                 path = sha1_file_name(sha1);
2613                 die("loose object %s (stored in %s) is corrupt",
2614                     sha1_to_hex(repl), path);
2615         }
2616
2617         if ((p = has_packed_and_bad(repl)) != NULL)
2618                 die("packed object %s (stored in %s) is corrupt",
2619                     sha1_to_hex(repl), p->pack_name);
2620
2621         return NULL;
2622 }
2623
2624 void *read_object_with_reference(const unsigned char *sha1,
2625                                  const char *required_type_name,
2626                                  unsigned long *size,
2627                                  unsigned char *actual_sha1_return)
2628 {
2629         enum object_type type, required_type;
2630         void *buffer;
2631         unsigned long isize;
2632         unsigned char actual_sha1[20];
2633
2634         required_type = type_from_string(required_type_name);
2635         hashcpy(actual_sha1, sha1);
2636         while (1) {
2637                 int ref_length = -1;
2638                 const char *ref_type = NULL;
2639
2640                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2641                 if (!buffer)
2642                         return NULL;
2643                 if (type == required_type) {
2644                         *size = isize;
2645                         if (actual_sha1_return)
2646                                 hashcpy(actual_sha1_return, actual_sha1);
2647                         return buffer;
2648                 }
2649                 /* Handle references */
2650                 else if (type == OBJ_COMMIT)
2651                         ref_type = "tree ";
2652                 else if (type == OBJ_TAG)
2653                         ref_type = "object ";
2654                 else {
2655                         free(buffer);
2656                         return NULL;
2657                 }
2658                 ref_length = strlen(ref_type);
2659
2660                 if (ref_length + 40 > isize ||
2661                     memcmp(buffer, ref_type, ref_length) ||
2662                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2663                         free(buffer);
2664                         return NULL;
2665                 }
2666                 free(buffer);
2667                 /* Now we have the ID of the referred-to object in
2668                  * actual_sha1.  Check again. */
2669         }
2670 }
2671
2672 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2673                                     const char *type, unsigned char *sha1,
2674                                     char *hdr, int *hdrlen)
2675 {
2676         git_SHA_CTX c;
2677
2678         /* Generate the header */
2679         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2680
2681         /* Sha1.. */
2682         git_SHA1_Init(&c);
2683         git_SHA1_Update(&c, hdr, *hdrlen);
2684         git_SHA1_Update(&c, buf, len);
2685         git_SHA1_Final(sha1, &c);
2686 }
2687
2688 /*
2689  * Move the just written object into its final resting place.
2690  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2691  * "moving" is only a part of what it does, when no patch between
2692  * master to pu changes the call sites of this function.
2693  */
2694 int move_temp_to_file(const char *tmpfile, const char *filename)
2695 {
2696         int ret = 0;
2697
2698         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2699                 goto try_rename;
2700         else if (link(tmpfile, filename))
2701                 ret = errno;
2702
2703         /*
2704          * Coda hack - coda doesn't like cross-directory links,
2705          * so we fall back to a rename, which will mean that it
2706          * won't be able to check collisions, but that's not a
2707          * big deal.
2708          *
2709          * The same holds for FAT formatted media.
2710          *
2711          * When this succeeds, we just return.  We have nothing
2712          * left to unlink.
2713          */
2714         if (ret && ret != EEXIST) {
2715         try_rename:
2716                 if (!rename(tmpfile, filename))
2717                         goto out;
2718                 ret = errno;
2719         }
2720         unlink_or_warn(tmpfile);
2721         if (ret) {
2722                 if (ret != EEXIST) {
2723                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2724                 }
2725                 /* FIXME!!! Collision check here ? */
2726         }
2727
2728 out:
2729         if (adjust_shared_perm(filename))
2730                 return error("unable to set permission to '%s'", filename);
2731         return 0;
2732 }
2733
2734 static int write_buffer(int fd, const void *buf, size_t len)
2735 {
2736         if (write_in_full(fd, buf, len) < 0)
2737                 return error("file write error (%s)", strerror(errno));
2738         return 0;
2739 }
2740
2741 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2742                    unsigned char *sha1)
2743 {
2744         char hdr[32];
2745         int hdrlen;
2746         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2747         return 0;
2748 }
2749
2750 /* Finalize a file on disk, and close it. */
2751 static void close_sha1_file(int fd)
2752 {
2753         if (fsync_object_files)
2754                 fsync_or_die(fd, "sha1 file");
2755         if (close(fd) != 0)
2756                 die_errno("error when closing sha1 file");
2757 }
2758
2759 /* Size of directory component, including the ending '/' */
2760 static inline int directory_size(const char *filename)
2761 {
2762         const char *s = strrchr(filename, '/');
2763         if (!s)
2764                 return 0;
2765         return s - filename + 1;
2766 }
2767
2768 /*
2769  * This creates a temporary file in the same directory as the final
2770  * 'filename'
2771  *
2772  * We want to avoid cross-directory filename renames, because those
2773  * can have problems on various filesystems (FAT, NFS, Coda).
2774  */
2775 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2776 {
2777         int fd, dirlen = directory_size(filename);
2778
2779         if (dirlen + 20 > bufsiz) {
2780                 errno = ENAMETOOLONG;
2781                 return -1;
2782         }
2783         memcpy(buffer, filename, dirlen);
2784         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2785         fd = git_mkstemp_mode(buffer, 0444);
2786         if (fd < 0 && dirlen && errno == ENOENT) {
2787                 /* Make sure the directory exists */
2788                 memcpy(buffer, filename, dirlen);
2789                 buffer[dirlen-1] = 0;
2790                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2791                         return -1;
2792
2793                 /* Try again */
2794                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2795                 fd = git_mkstemp_mode(buffer, 0444);
2796         }
2797         return fd;
2798 }
2799
2800 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2801                               const void *buf, unsigned long len, time_t mtime)
2802 {
2803         int fd, ret;
2804         unsigned char compressed[4096];
2805         git_zstream stream;
2806         git_SHA_CTX c;
2807         unsigned char parano_sha1[20];
2808         char *filename;
2809         static char tmp_file[PATH_MAX];
2810
2811         filename = sha1_file_name(sha1);
2812         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
2813         if (fd < 0) {
2814                 if (errno == EACCES)
2815                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
2816                 else
2817                         return error("unable to create temporary file: %s", strerror(errno));
2818         }
2819
2820         /* Set it up */
2821         memset(&stream, 0, sizeof(stream));
2822         git_deflate_init(&stream, zlib_compression_level);
2823         stream.next_out = compressed;
2824         stream.avail_out = sizeof(compressed);
2825         git_SHA1_Init(&c);
2826
2827         /* First header.. */
2828         stream.next_in = (unsigned char *)hdr;
2829         stream.avail_in = hdrlen;
2830         while (git_deflate(&stream, 0) == Z_OK)
2831                 ; /* nothing */
2832         git_SHA1_Update(&c, hdr, hdrlen);
2833
2834         /* Then the data itself.. */
2835         stream.next_in = (void *)buf;
2836         stream.avail_in = len;
2837         do {
2838                 unsigned char *in0 = stream.next_in;
2839                 ret = git_deflate(&stream, Z_FINISH);
2840                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2841                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2842                         die("unable to write sha1 file");
2843                 stream.next_out = compressed;
2844                 stream.avail_out = sizeof(compressed);
2845         } while (ret == Z_OK);
2846
2847         if (ret != Z_STREAM_END)
2848                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2849         ret = git_deflate_end_gently(&stream);
2850         if (ret != Z_OK)
2851                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2852         git_SHA1_Final(parano_sha1, &c);
2853         if (hashcmp(sha1, parano_sha1) != 0)
2854                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2855
2856         close_sha1_file(fd);
2857
2858         if (mtime) {
2859                 struct utimbuf utb;
2860                 utb.actime = mtime;
2861                 utb.modtime = mtime;
2862                 if (utime(tmp_file, &utb) < 0)
2863                         warning("failed utime() on %s: %s",
2864                                 tmp_file, strerror(errno));
2865         }
2866
2867         return move_temp_to_file(tmp_file, filename);
2868 }
2869
2870 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2871 {
2872         unsigned char sha1[20];
2873         char hdr[32];
2874         int hdrlen;
2875
2876         /* Normally if we have it in the pack then we do not bother writing
2877          * it out into .git/objects/??/?{38} file.
2878          */
2879         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2880         if (returnsha1)
2881                 hashcpy(returnsha1, sha1);
2882         if (has_sha1_file(sha1))
2883                 return 0;
2884         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2885 }
2886
2887 int force_object_loose(const unsigned char *sha1, time_t mtime)
2888 {
2889         void *buf;
2890         unsigned long len;
2891         enum object_type type;
2892         char hdr[32];
2893         int hdrlen;
2894         int ret;
2895
2896         if (has_loose_object(sha1))
2897                 return 0;
2898         buf = read_packed_sha1(sha1, &type, &len);
2899         if (!buf)
2900                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2901         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2902         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2903         free(buf);
2904
2905         return ret;
2906 }
2907
2908 int has_pack_index(const unsigned char *sha1)
2909 {
2910         struct stat st;
2911         if (stat(sha1_pack_index_name(sha1), &st))
2912                 return 0;
2913         return 1;
2914 }
2915
2916 int has_sha1_pack(const unsigned char *sha1)
2917 {
2918         struct pack_entry e;
2919         return find_pack_entry(sha1, &e);
2920 }
2921
2922 int has_sha1_file(const unsigned char *sha1)
2923 {
2924         struct pack_entry e;
2925
2926         if (find_pack_entry(sha1, &e))
2927                 return 1;
2928         return has_loose_object(sha1);
2929 }
2930
2931 static void check_tree(const void *buf, size_t size)
2932 {
2933         struct tree_desc desc;
2934         struct name_entry entry;
2935
2936         init_tree_desc(&desc, buf, size);
2937         while (tree_entry(&desc, &entry))
2938                 /* do nothing
2939                  * tree_entry() will die() on malformed entries */
2940                 ;
2941 }
2942
2943 static void check_commit(const void *buf, size_t size)
2944 {
2945         struct commit c;
2946         memset(&c, 0, sizeof(c));
2947         if (parse_commit_buffer(&c, buf, size))
2948                 die("corrupt commit");
2949 }
2950
2951 static void check_tag(const void *buf, size_t size)
2952 {
2953         struct tag t;
2954         memset(&t, 0, sizeof(t));
2955         if (parse_tag_buffer(&t, buf, size))
2956                 die("corrupt tag");
2957 }
2958
2959 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2960                      enum object_type type,
2961                      const char *path, unsigned flags)
2962 {
2963         int ret, re_allocated = 0;
2964         int write_object = flags & HASH_WRITE_OBJECT;
2965
2966         if (!type)
2967                 type = OBJ_BLOB;
2968
2969         /*
2970          * Convert blobs to git internal format
2971          */
2972         if ((type == OBJ_BLOB) && path) {
2973                 struct strbuf nbuf = STRBUF_INIT;
2974                 if (convert_to_git(path, buf, size, &nbuf,
2975                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
2976                         buf = strbuf_detach(&nbuf, &size);
2977                         re_allocated = 1;
2978                 }
2979         }
2980         if (flags & HASH_FORMAT_CHECK) {
2981                 if (type == OBJ_TREE)
2982                         check_tree(buf, size);
2983                 if (type == OBJ_COMMIT)
2984                         check_commit(buf, size);
2985                 if (type == OBJ_TAG)
2986                         check_tag(buf, size);
2987         }
2988
2989         if (write_object)
2990                 ret = write_sha1_file(buf, size, typename(type), sha1);
2991         else
2992                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2993         if (re_allocated)
2994                 free(buf);
2995         return ret;
2996 }
2997
2998 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
2999                       const char *path, unsigned flags)
3000 {
3001         struct strbuf sbuf = STRBUF_INIT;
3002         int ret;
3003
3004         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3005                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3006         else
3007                 ret = -1;
3008         strbuf_release(&sbuf);
3009         return ret;
3010 }
3011
3012 #define SMALL_FILE_SIZE (32*1024)
3013
3014 static int index_core(unsigned char *sha1, int fd, size_t size,
3015                       enum object_type type, const char *path,
3016                       unsigned flags)
3017 {
3018         int ret;
3019
3020         if (!size) {
3021                 ret = index_mem(sha1, NULL, size, type, path, flags);
3022         } else if (size <= SMALL_FILE_SIZE) {
3023                 char *buf = xmalloc(size);
3024                 if (size == read_in_full(fd, buf, size))
3025                         ret = index_mem(sha1, buf, size, type, path, flags);
3026                 else
3027                         ret = error("short read %s", strerror(errno));
3028                 free(buf);
3029         } else {
3030                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3031                 ret = index_mem(sha1, buf, size, type, path, flags);
3032                 munmap(buf, size);
3033         }
3034         return ret;
3035 }
3036
3037 /*
3038  * This creates one packfile per large blob unless bulk-checkin
3039  * machinery is "plugged".
3040  *
3041  * This also bypasses the usual "convert-to-git" dance, and that is on
3042  * purpose. We could write a streaming version of the converting
3043  * functions and insert that before feeding the data to fast-import
3044  * (or equivalent in-core API described above). However, that is
3045  * somewhat complicated, as we do not know the size of the filter
3046  * result, which we need to know beforehand when writing a git object.
3047  * Since the primary motivation for trying to stream from the working
3048  * tree file and to avoid mmaping it in core is to deal with large
3049  * binary blobs, they generally do not want to get any conversion, and
3050  * callers should avoid this code path when filters are requested.
3051  */
3052 static int index_stream(unsigned char *sha1, int fd, size_t size,
3053                         enum object_type type, const char *path,
3054                         unsigned flags)
3055 {
3056         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3057 }
3058
3059 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3060              enum object_type type, const char *path, unsigned flags)
3061 {
3062         int ret;
3063         size_t size = xsize_t(st->st_size);
3064
3065         if (!S_ISREG(st->st_mode))
3066                 ret = index_pipe(sha1, fd, type, path, flags);
3067         else if (size <= big_file_threshold || type != OBJ_BLOB ||
3068                  (path && would_convert_to_git(path, NULL, 0, 0)))
3069                 ret = index_core(sha1, fd, size, type, path, flags);
3070         else
3071                 ret = index_stream(sha1, fd, size, type, path, flags);
3072         close(fd);
3073         return ret;
3074 }
3075
3076 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3077 {
3078         int fd;
3079         struct strbuf sb = STRBUF_INIT;
3080
3081         switch (st->st_mode & S_IFMT) {
3082         case S_IFREG:
3083                 fd = open(path, O_RDONLY);
3084                 if (fd < 0)
3085                         return error("open(\"%s\"): %s", path,
3086                                      strerror(errno));
3087                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3088                         return error("%s: failed to insert into database",
3089                                      path);
3090                 break;
3091         case S_IFLNK:
3092                 if (strbuf_readlink(&sb, path, st->st_size)) {
3093                         char *errstr = strerror(errno);
3094                         return error("readlink(\"%s\"): %s", path,
3095                                      errstr);
3096                 }
3097                 if (!(flags & HASH_WRITE_OBJECT))
3098                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3099                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3100                         return error("%s: failed to insert into database",
3101                                      path);
3102                 strbuf_release(&sb);
3103                 break;
3104         case S_IFDIR:
3105                 return resolve_gitlink_ref(path, "HEAD", sha1);
3106         default:
3107                 return error("%s: unsupported file type", path);
3108         }
3109         return 0;
3110 }
3111
3112 int read_pack_header(int fd, struct pack_header *header)
3113 {
3114         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3115                 /* "eof before pack header was fully read" */
3116                 return PH_ERROR_EOF;
3117
3118         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3119                 /* "protocol error (pack signature mismatch detected)" */
3120                 return PH_ERROR_PACK_SIGNATURE;
3121         if (!pack_version_ok(header->hdr_version))
3122                 /* "protocol error (pack version unsupported)" */
3123                 return PH_ERROR_PROTOCOL;
3124         return 0;
3125 }
3126
3127 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3128 {
3129         enum object_type type = sha1_object_info(sha1, NULL);
3130         if (type < 0)
3131                 die("%s is not a valid object", sha1_to_hex(sha1));
3132         if (type != expect)
3133                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3134                     typename(expect));
3135 }