read-cache.c

   1 /*
   2  * GIT - The information manager from hell
   3  *
   4  * Copyright (C) Linus Torvalds, 2005
   5  */
   6 #include "cache.h"
   7 #include "cache-tree.h"
   8
   9 /* Index extensions.
  10  *
  11  * The first letter should be 'A'..'Z' for extensions that are not
  12  * necessary for a correct operation (i.e. optimization data).
  13  * When new extensions are added that _needs_ to be understood in
  14  * order to correctly interpret the index file, pick character that
  15  * is outside the range, to cause the reader to abort.
  16  */
  17
  18 #define CACHE_EXT(s) ( (s[0]<<24)|(s[1]<<16)|(s[2]<<8)|(s[3]) )
  19 #define CACHE_EXT_TREE 0x54524545       /* "TREE" */
  20
  21 struct cache_entry **active_cache = NULL;
  22 static time_t index_file_timestamp;
  23 unsigned int active_nr = 0, active_alloc = 0, active_cache_changed = 0;
  24
  25 struct cache_tree *active_cache_tree = NULL;
  26
  27 int cache_errno = 0;
  28
  29 static void *cache_mmap = NULL;
  30 static size_t cache_mmap_size = 0;
  31
  32 /*
  33  * This only updates the "non-critical" parts of the directory
  34  * cache, ie the parts that aren't tracked by GIT, and only used
  35  * to validate the cache.
  36  */
  37 void fill_stat_cache_info(struct cache_entry *ce, struct stat *st)
  38 {
  39         ce->ce_ctime.sec = htonl(st->st_ctime);
  40         ce->ce_mtime.sec = htonl(st->st_mtime);
  41 #ifdef USE_NSEC
  42         ce->ce_ctime.nsec = htonl(st->st_ctim.tv_nsec);
  43         ce->ce_mtime.nsec = htonl(st->st_mtim.tv_nsec);
  44 #endif
  45         ce->ce_dev = htonl(st->st_dev);
  46         ce->ce_ino = htonl(st->st_ino);
  47         ce->ce_uid = htonl(st->st_uid);
  48         ce->ce_gid = htonl(st->st_gid);
  49         ce->ce_size = htonl(st->st_size);
  50
  51         if (assume_unchanged)
  52                 ce->ce_flags |= htons(CE_VALID);
  53 }
  54
  55 static int ce_compare_data(struct cache_entry *ce, struct stat *st)
  56 {
  57         int match = -1;
  58         int fd = open(ce->name, O_RDONLY);
  59
  60         if (fd >= 0) {
  61                 unsigned char sha1[20];
  62                 if (!index_fd(sha1, fd, st, 0, NULL))
  63                         match = memcmp(sha1, ce->sha1, 20);
  64                 /* index_fd() closed the file descriptor already */
  65         }
  66         return match;
  67 }
  68
  69 static int ce_compare_link(struct cache_entry *ce, unsigned long expected_size)
  70 {
  71         int match = -1;
  72         char *target;
  73         void *buffer;
  74         unsigned long size;
  75         char type[10];
  76         int len;
  77
  78         target = xmalloc(expected_size);
  79         len = readlink(ce->name, target, expected_size);
  80         if (len != expected_size) {
  81                 free(target);
  82                 return -1;
  83         }
  84         buffer = read_sha1_file(ce->sha1, type, &size);
  85         if (!buffer) {
  86                 free(target);
  87                 return -1;
  88         }
  89         if (size == expected_size)
  90                 match = memcmp(buffer, target, size);
  91         free(buffer);
  92         free(target);
  93         return match;
  94 }
  95
  96 static int ce_modified_check_fs(struct cache_entry *ce, struct stat *st)
  97 {
  98         switch (st->st_mode & S_IFMT) {
  99         case S_IFREG:
 100                 if (ce_compare_data(ce, st))
 101                         return DATA_CHANGED;
 102                 break;
 103         case S_IFLNK:
 104                 if (ce_compare_link(ce, st->st_size))
 105                         return DATA_CHANGED;
 106                 break;
 107         default:
 108                 return TYPE_CHANGED;
 109         }
 110         return 0;
 111 }
 112
 113 static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
 114 {
 115         unsigned int changed = 0;
 116
 117         switch (ntohl(ce->ce_mode) & S_IFMT) {
 118         case S_IFREG:
 119                 changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
 120                 /* We consider only the owner x bit to be relevant for
 121                  * "mode changes"
 122                  */
 123                 if (trust_executable_bit &&
 124                     (0100 & (ntohl(ce->ce_mode) ^ st->st_mode)))
 125                         changed |= MODE_CHANGED;
 126                 break;
 127         case S_IFLNK:
 128                 changed |= !S_ISLNK(st->st_mode) ? TYPE_CHANGED : 0;
 129                 break;
 130         default:
 131                 die("internal error: ce_mode is %o", ntohl(ce->ce_mode));
 132         }
 133         if (ce->ce_mtime.sec != htonl(st->st_mtime))
 134                 changed |= MTIME_CHANGED;
 135         if (ce->ce_ctime.sec != htonl(st->st_ctime))
 136                 changed |= CTIME_CHANGED;
 137
 138 #ifdef USE_NSEC
 139         /*
 140          * nsec seems unreliable - not all filesystems support it, so
 141          * as long as it is in the inode cache you get right nsec
 142          * but after it gets flushed, you get zero nsec.
 143          */
 144         if (ce->ce_mtime.nsec != htonl(st->st_mtim.tv_nsec))
 145                 changed |= MTIME_CHANGED;
 146         if (ce->ce_ctime.nsec != htonl(st->st_ctim.tv_nsec))
 147                 changed |= CTIME_CHANGED;
 148 #endif
 149
 150         if (ce->ce_uid != htonl(st->st_uid) ||
 151             ce->ce_gid != htonl(st->st_gid))
 152                 changed |= OWNER_CHANGED;
 153         if (ce->ce_ino != htonl(st->st_ino))
 154                 changed |= INODE_CHANGED;
 155
 156 #ifdef USE_STDEV
 157         /*
 158          * st_dev breaks on network filesystems where different
 159          * clients will have different views of what "device"
 160          * the filesystem is on
 161          */
 162         if (ce->ce_dev != htonl(st->st_dev))
 163                 changed |= INODE_CHANGED;
 164 #endif
 165
 166         if (ce->ce_size != htonl(st->st_size))
 167                 changed |= DATA_CHANGED;
 168
 169         return changed;
 170 }
 171
 172 int ce_match_stat(struct cache_entry *ce, struct stat *st, int ignore_valid)
 173 {
 174         unsigned int changed;
 175
 176         /*
 177          * If it's marked as always valid in the index, it's
 178          * valid whatever the checked-out copy says.
 179          */
 180         if (!ignore_valid && (ce->ce_flags & htons(CE_VALID)))
 181                 return 0;
 182
 183         changed = ce_match_stat_basic(ce, st);
 184
 185         /*
 186          * Within 1 second of this sequence:
 187          *      echo xyzzy >file && git-update-index --add file
 188          * running this command:
 189          *      echo frotz >file
 190          * would give a falsely clean cache entry.  The mtime and
 191          * length match the cache, and other stat fields do not change.
 192          *
 193          * We could detect this at update-index time (the cache entry
 194          * being registered/updated records the same time as "now")
 195          * and delay the return from git-update-index, but that would
 196          * effectively mean we can make at most one commit per second,
 197          * which is not acceptable.  Instead, we check cache entries
 198          * whose mtime are the same as the index file timestamp more
 199          * carefully than others.
 200          */
 201         if (!changed &&
 202             index_file_timestamp &&
 203             index_file_timestamp <= ntohl(ce->ce_mtime.sec))
 204                 changed |= ce_modified_check_fs(ce, st);
 205
 206         return changed;
 207 }
 208
 209 int ce_modified(struct cache_entry *ce, struct stat *st, int really)
 210 {
 211         int changed, changed_fs;
 212         changed = ce_match_stat(ce, st, really);
 213         if (!changed)
 214                 return 0;
 215         /*
 216          * If the mode or type has changed, there's no point in trying
 217          * to refresh the entry - it's not going to match
 218          */
 219         if (changed & (MODE_CHANGED | TYPE_CHANGED))
 220                 return changed;
 221
 222         /* Immediately after read-tree or update-index --cacheinfo,
 223          * the length field is zero.  For other cases the ce_size
 224          * should match the SHA1 recorded in the index entry.
 225          */
 226         if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
 227                 return changed;
 228
 229         changed_fs = ce_modified_check_fs(ce, st);
 230         if (changed_fs)
 231                 return changed | changed_fs;
 232         return 0;
 233 }
 234
 235 int base_name_compare(const char *name1, int len1, int mode1,
 236                       const char *name2, int len2, int mode2)
 237 {
 238         unsigned char c1, c2;
 239         int len = len1 < len2 ? len1 : len2;
 240         int cmp;
 241
 242         cmp = memcmp(name1, name2, len);
 243         if (cmp)
 244                 return cmp;
 245         c1 = name1[len];
 246         c2 = name2[len];
 247         if (!c1 && S_ISDIR(mode1))
 248                 c1 = '/';
 249         if (!c2 && S_ISDIR(mode2))
 250                 c2 = '/';
 251         return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
 252 }
 253
 254 int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
 255 {
 256         int len1 = flags1 & CE_NAMEMASK;
 257         int len2 = flags2 & CE_NAMEMASK;
 258         int len = len1 < len2 ? len1 : len2;
 259         int cmp;
 260
 261         cmp = memcmp(name1, name2, len);
 262         if (cmp)
 263                 return cmp;
 264         if (len1 < len2)
 265                 return -1;
 266         if (len1 > len2)
 267                 return 1;
 268
 269         /* Compare stages  */
 270         flags1 &= CE_STAGEMASK;
 271         flags2 &= CE_STAGEMASK;
 272
 273         if (flags1 < flags2)
 274                 return -1;
 275         if (flags1 > flags2)
 276                 return 1;
 277         return 0;
 278 }
 279
 280 int cache_name_pos(const char *name, int namelen)
 281 {
 282         int first, last;
 283
 284         first = 0;
 285         last = active_nr;
 286         while (last > first) {
 287                 int next = (last + first) >> 1;
 288                 struct cache_entry *ce = active_cache[next];
 289                 int cmp = cache_name_compare(name, namelen, ce->name, ntohs(ce->ce_flags));
 290                 if (!cmp)
 291                         return next;
 292                 if (cmp < 0) {
 293                         last = next;
 294                         continue;
 295                 }
 296                 first = next+1;
 297         }
 298         return -first-1;
 299 }
 300
 301 /* Remove entry, return true if there are more entries to go.. */
 302 int remove_cache_entry_at(int pos)
 303 {
 304         active_cache_changed = 1;
 305         active_nr--;
 306         if (pos >= active_nr)
 307                 return 0;
 308         memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
 309         return 1;
 310 }
 311
 312 int remove_file_from_cache(const char *path)
 313 {
 314         int pos = cache_name_pos(path, strlen(path));
 315         if (pos < 0)
 316                 pos = -pos-1;
 317         while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
 318                 remove_cache_entry_at(pos);
 319         return 0;
 320 }
 321
 322 int add_file_to_index(const char *path, int verbose)
 323 {
 324         int size, namelen;
 325         struct stat st;
 326         struct cache_entry *ce;
 327
 328         if (lstat(path, &st))
 329                 die("%s: unable to stat (%s)", path, strerror(errno));
 330
 331         if (!S_ISREG(st.st_mode) && !S_ISLNK(st.st_mode))
 332                 die("%s: can only add regular files or symbolic links", path);
 333
 334         namelen = strlen(path);
 335         size = cache_entry_size(namelen);
 336         ce = xcalloc(1, size);
 337         memcpy(ce->name, path, namelen);
 338         ce->ce_flags = htons(namelen);
 339         fill_stat_cache_info(ce, &st);
 340
 341         ce->ce_mode = create_ce_mode(st.st_mode);
 342         if (!trust_executable_bit) {
 343                 /* If there is an existing entry, pick the mode bits
 344                  * from it.
 345                  */
 346                 int pos = cache_name_pos(path, namelen);
 347                 if (pos >= 0)
 348                         ce->ce_mode = active_cache[pos]->ce_mode;
 349         }
 350
 351         if (index_path(ce->sha1, path, &st, 1))
 352                 die("unable to index file %s", path);
 353         if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD))
 354                 die("unable to add %s to index",path);
 355         if (verbose)
 356                 printf("add '%s'\n", path);
 357         cache_tree_invalidate_path(active_cache_tree, path);
 358         return 0;
 359 }
 360
 361 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
 362 {
 363         int len = ce_namelen(a);
 364         return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
 365 }
 366
 367 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
 368 {
 369         const char *match, *name;
 370         int len;
 371
 372         if (!pathspec)
 373                 return 1;
 374
 375         len = ce_namelen(ce);
 376         name = ce->name;
 377         while ((match = *pathspec++) != NULL) {
 378                 int matchlen = strlen(match);
 379                 if (matchlen > len)
 380                         continue;
 381                 if (memcmp(name, match, matchlen))
 382                         continue;
 383                 if (matchlen && name[matchlen-1] == '/')
 384                         return 1;
 385                 if (name[matchlen] == '/' || !name[matchlen])
 386                         return 1;
 387                 if (!matchlen)
 388                         return 1;
 389         }
 390         return 0;
 391 }
 392
 393 /*
 394  * We fundamentally don't like some paths: we don't want
 395  * dot or dot-dot anywhere, and for obvious reasons don't
 396  * want to recurse into ".git" either.
 397  *
 398  * Also, we don't want double slashes or slashes at the
 399  * end that can make pathnames ambiguous.
 400  */
 401 static int verify_dotfile(const char *rest)
 402 {
 403         /*
 404          * The first character was '.', but that
 405          * has already been discarded, we now test
 406          * the rest.
 407          */
 408         switch (*rest) {
 409         /* "." is not allowed */
 410         case '\0': case '/':
 411                 return 0;
 412
 413         /*
 414          * ".git" followed by  NUL or slash is bad. This
 415          * shares the path end test with the ".." case.
 416          */
 417         case 'g':
 418                 if (rest[1] != 'i')
 419                         break;
 420                 if (rest[2] != 't')
 421                         break;
 422                 rest += 2;
 423         /* fallthrough */
 424         case '.':
 425                 if (rest[1] == '\0' || rest[1] == '/')
 426                         return 0;
 427         }
 428         return 1;
 429 }
 430
 431 int verify_path(const char *path)
 432 {
 433         char c;
 434
 435         goto inside;
 436         for (;;) {
 437                 if (!c)
 438                         return 1;
 439                 if (c == '/') {
 440 inside:
 441                         c = *path++;
 442                         switch (c) {
 443                         default:
 444                                 continue;
 445                         case '/': case '\0':
 446                                 break;
 447                         case '.':
 448                                 if (verify_dotfile(path))
 449                                         continue;
 450                         }
 451                         return 0;
 452                 }
 453                 c = *path++;
 454         }
 455 }
 456
 457 /*
 458  * Do we have another file that has the beginning components being a
 459  * proper superset of the name we're trying to add?
 460  */
 461 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 462 {
 463         int retval = 0;
 464         int len = ce_namelen(ce);
 465         int stage = ce_stage(ce);
 466         const char *name = ce->name;
 467
 468         while (pos < active_nr) {
 469                 struct cache_entry *p = active_cache[pos++];
 470
 471                 if (len >= ce_namelen(p))
 472                         break;
 473                 if (memcmp(name, p->name, len))
 474                         break;
 475                 if (ce_stage(p) != stage)
 476                         continue;
 477                 if (p->name[len] != '/')
 478                         continue;
 479                 retval = -1;
 480                 if (!ok_to_replace)
 481                         break;
 482                 remove_cache_entry_at(--pos);
 483         }
 484         return retval;
 485 }
 486
 487 /*
 488  * Do we have another file with a pathname that is a proper
 489  * subset of the name we're trying to add?
 490  */
 491 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 492 {
 493         int retval = 0;
 494         int stage = ce_stage(ce);
 495         const char *name = ce->name;
 496         const char *slash = name + ce_namelen(ce);
 497
 498         for (;;) {
 499                 int len;
 500
 501                 for (;;) {
 502                         if (*--slash == '/')
 503                                 break;
 504                         if (slash <= ce->name)
 505                                 return retval;
 506                 }
 507                 len = slash - name;
 508
 509                 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
 510                 if (pos >= 0) {
 511                         retval = -1;
 512                         if (ok_to_replace)
 513                                 break;
 514                         remove_cache_entry_at(pos);
 515                         continue;
 516                 }
 517
 518                 /*
 519                  * Trivial optimization: if we find an entry that
 520                  * already matches the sub-directory, then we know
 521                  * we're ok, and we can exit.
 522                  */
 523                 pos = -pos-1;
 524                 while (pos < active_nr) {
 525                         struct cache_entry *p = active_cache[pos];
 526                         if ((ce_namelen(p) <= len) ||
 527                             (p->name[len] != '/') ||
 528                             memcmp(p->name, name, len))
 529                                 break; /* not our subdirectory */
 530                         if (ce_stage(p) == stage)
 531                                 /* p is at the same stage as our entry, and
 532                                  * is a subdirectory of what we are looking
 533                                  * at, so we cannot have conflicts at our
 534                                  * level or anything shorter.
 535                                  */
 536                                 return retval;
 537                         pos++;
 538                 }
 539         }
 540         return retval;
 541 }
 542
 543 /* We may be in a situation where we already have path/file and path
 544  * is being added, or we already have path and path/file is being
 545  * added.  Either one would result in a nonsense tree that has path
 546  * twice when git-write-tree tries to write it out.  Prevent it.
 547  *
 548  * If ok-to-replace is specified, we remove the conflicting entries
 549  * from the cache so the caller should recompute the insert position.
 550  * When this happens, we return non-zero.
 551  */
 552 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
 553 {
 554         /*
 555          * We check if the path is a sub-path of a subsequent pathname
 556          * first, since removing those will not change the position
 557          * in the array
 558          */
 559         int retval = has_file_name(ce, pos, ok_to_replace);
 560         /*
 561          * Then check if the path might have a clashing sub-directory
 562          * before it.
 563          */
 564         return retval + has_dir_name(ce, pos, ok_to_replace);
 565 }
 566
 567 int add_cache_entry(struct cache_entry *ce, int option)
 568 {
 569         int pos;
 570         int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
 571         int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
 572         int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
 573
 574         pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 575
 576         /* existing match? Just replace it. */
 577         if (pos >= 0) {
 578                 active_cache_changed = 1;
 579                 active_cache[pos] = ce;
 580                 return 0;
 581         }
 582         pos = -pos-1;
 583
 584         /*
 585          * Inserting a merged entry ("stage 0") into the index
 586          * will always replace all non-merged entries..
 587          */
 588         if (pos < active_nr && ce_stage(ce) == 0) {
 589                 while (ce_same_name(active_cache[pos], ce)) {
 590                         ok_to_add = 1;
 591                         if (!remove_cache_entry_at(pos))
 592                                 break;
 593                 }
 594         }
 595
 596         if (!ok_to_add)
 597                 return -1;
 598         if (!verify_path(ce->name))
 599                 return -1;
 600
 601         if (!skip_df_check &&
 602             check_file_directory_conflict(ce, pos, ok_to_replace)) {
 603                 if (!ok_to_replace)
 604                         return -1;
 605                 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 606                 pos = -pos-1;
 607         }
 608
 609         /* Make sure the array is big enough .. */
 610         if (active_nr == active_alloc) {
 611                 active_alloc = alloc_nr(active_alloc);
 612                 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
 613         }
 614
 615         /* Add it in.. */
 616         active_nr++;
 617         if (active_nr > pos)
 618                 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
 619         active_cache[pos] = ce;
 620         active_cache_changed = 1;
 621         return 0;
 622 }
 623
 624 /*
 625  * "refresh" does not calculate a new sha1 file or bring the
 626  * cache up-to-date for mode/content changes. But what it
 627  * _does_ do is to "re-match" the stat information of a file
 628  * with the cache, so that you can refresh the cache for a
 629  * file that hasn't been changed but where the stat entry is
 630  * out of date.
 631  *
 632  * For example, you'd want to do this after doing a "git-read-tree",
 633  * to link up the stat cache details with the proper files.
 634  */
 635 struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
 636 {
 637         struct stat st;
 638         struct cache_entry *updated;
 639         int changed, size;
 640
 641         if (lstat(ce->name, &st) < 0) {
 642                 cache_errno = errno;
 643                 return NULL;
 644         }
 645
 646         changed = ce_match_stat(ce, &st, really);
 647         if (!changed) {
 648                 if (really && assume_unchanged &&
 649                     !(ce->ce_flags & htons(CE_VALID)))
 650                         ; /* mark this one VALID again */
 651                 else
 652                         return ce;
 653         }
 654
 655         if (ce_modified(ce, &st, really)) {
 656                 cache_errno = EINVAL;
 657                 return NULL;
 658         }
 659
 660         size = ce_size(ce);
 661         updated = xmalloc(size);
 662         memcpy(updated, ce, size);
 663         fill_stat_cache_info(updated, &st);
 664
 665         /* In this case, if really is not set, we should leave
 666          * CE_VALID bit alone.  Otherwise, paths marked with
 667          * --no-assume-unchanged (i.e. things to be edited) will
 668          * reacquire CE_VALID bit automatically, which is not
 669          * really what we want.
 670          */
 671         if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
 672                 updated->ce_flags &= ~htons(CE_VALID);
 673
 674         return updated;
 675 }
 676
 677 int refresh_cache(unsigned int flags)
 678 {
 679         int i;
 680         int has_errors = 0;
 681         int really = (flags & REFRESH_REALLY) != 0;
 682         int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
 683         int quiet = (flags & REFRESH_QUIET) != 0;
 684         int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
 685
 686         for (i = 0; i < active_nr; i++) {
 687                 struct cache_entry *ce, *new;
 688                 ce = active_cache[i];
 689                 if (ce_stage(ce)) {
 690                         while ((i < active_nr) &&
 691                                ! strcmp(active_cache[i]->name, ce->name))
 692                                 i++;
 693                         i--;
 694                         if (allow_unmerged)
 695                                 continue;
 696                         printf("%s: needs merge\n", ce->name);
 697                         has_errors = 1;
 698                         continue;
 699                 }
 700
 701                 new = refresh_cache_entry(ce, really);
 702                 if (new == ce)
 703                         continue;
 704                 if (!new) {
 705                         if (not_new && cache_errno == ENOENT)
 706                                 continue;
 707                         if (really && cache_errno == EINVAL) {
 708                                 /* If we are doing --really-refresh that
 709                                  * means the index is not valid anymore.
 710                                  */
 711                                 ce->ce_flags &= ~htons(CE_VALID);
 712                                 active_cache_changed = 1;
 713                         }
 714                         if (quiet)
 715                                 continue;
 716                         printf("%s: needs update\n", ce->name);
 717                         has_errors = 1;
 718                         continue;
 719                 }
 720                 active_cache_changed = 1;
 721                 /* You can NOT just free active_cache[i] here, since it
 722                  * might not be necessarily malloc()ed but can also come
 723                  * from mmap(). */
 724                 active_cache[i] = new;
 725         }
 726         return has_errors;
 727 }
 728
 729 static int verify_hdr(struct cache_header *hdr, unsigned long size)
 730 {
 731         SHA_CTX c;
 732         unsigned char sha1[20];
 733
 734         if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
 735                 return error("bad signature");
 736         if (hdr->hdr_version != htonl(2))
 737                 return error("bad index version");
 738         SHA1_Init(&c);
 739         SHA1_Update(&c, hdr, size - 20);
 740         SHA1_Final(sha1, &c);
 741         if (memcmp(sha1, (char *) hdr + size - 20, 20))
 742                 return error("bad index file sha1 signature");
 743         return 0;
 744 }
 745
 746 static int read_index_extension(const char *ext, void *data, unsigned long sz)
 747 {
 748         switch (CACHE_EXT(ext)) {
 749         case CACHE_EXT_TREE:
 750                 active_cache_tree = cache_tree_read(data, sz);
 751                 break;
 752         default:
 753                 if (*ext < 'A' || 'Z' < *ext)
 754                         return error("index uses %.4s extension, which we do not understand",
 755                                      ext);
 756                 fprintf(stderr, "ignoring %.4s extension\n", ext);
 757                 break;
 758         }
 759         return 0;
 760 }
 761
 762 int read_cache(void)
 763 {
 764         return read_cache_from(get_index_file());
 765 }
 766
 767 /* remember to discard_cache() before reading a different cache! */
 768 int read_cache_from(const char *path)
 769 {
 770         int fd, i;
 771         struct stat st;
 772         unsigned long offset;
 773         struct cache_header *hdr;
 774
 775         errno = EBUSY;
 776         if (cache_mmap)
 777                 return active_nr;
 778
 779         errno = ENOENT;
 780         index_file_timestamp = 0;
 781         fd = open(path, O_RDONLY);
 782         if (fd < 0) {
 783                 if (errno == ENOENT)
 784                         return 0;
 785                 die("index file open failed (%s)", strerror(errno));
 786         }
 787
 788         cache_mmap = MAP_FAILED;
 789         if (!fstat(fd, &st)) {
 790                 cache_mmap_size = st.st_size;
 791                 errno = EINVAL;
 792                 if (cache_mmap_size >= sizeof(struct cache_header) + 20)
 793                         cache_mmap = mmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
 794         }
 795         close(fd);
 796         if (cache_mmap == MAP_FAILED)
 797                 die("index file mmap failed (%s)", strerror(errno));
 798
 799         hdr = cache_mmap;
 800         if (verify_hdr(hdr, cache_mmap_size) < 0)
 801                 goto unmap;
 802
 803         active_nr = ntohl(hdr->hdr_entries);
 804         active_alloc = alloc_nr(active_nr);
 805         active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
 806
 807         offset = sizeof(*hdr);
 808         for (i = 0; i < active_nr; i++) {
 809                 struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
 810                 offset = offset + ce_size(ce);
 811                 active_cache[i] = ce;
 812         }
 813         index_file_timestamp = st.st_mtime;
 814         while (offset <= cache_mmap_size - 20 - 8) {
 815                 /* After an array of active_nr index entries,
 816                  * there can be arbitrary number of extended
 817                  * sections, each of which is prefixed with
 818                  * extension name (4-byte) and section length
 819                  * in 4-byte network byte order.
 820                  */
 821                 unsigned long extsize;
 822                 memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
 823                 extsize = ntohl(extsize);
 824                 if (read_index_extension(((const char *) cache_mmap) + offset,
 825                                          (char *) cache_mmap + offset + 8,
 826                                          extsize) < 0)
 827                         goto unmap;
 828                 offset += 8;
 829                 offset += extsize;
 830         }
 831         return active_nr;
 832
 833 unmap:
 834         munmap(cache_mmap, cache_mmap_size);
 835         errno = EINVAL;
 836         die("index file corrupt");
 837 }
 838
 839 #define WRITE_BUFFER_SIZE 8192
 840 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
 841 static unsigned long write_buffer_len;
 842
 843 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
 844 {
 845         while (len) {
 846                 unsigned int buffered = write_buffer_len;
 847                 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
 848                 if (partial > len)
 849                         partial = len;
 850                 memcpy(write_buffer + buffered, data, partial);
 851                 buffered += partial;
 852                 if (buffered == WRITE_BUFFER_SIZE) {
 853                         SHA1_Update(context, write_buffer, WRITE_BUFFER_SIZE);
 854                         if (write(fd, write_buffer, WRITE_BUFFER_SIZE) != WRITE_BUFFER_SIZE)
 855                                 return -1;
 856                         buffered = 0;
 857                 }
 858                 write_buffer_len = buffered;
 859                 len -= partial;
 860                 data = (char *) data + partial;
 861         }
 862         return 0;
 863 }
 864
 865 static int write_index_ext_header(SHA_CTX *context, int fd,
 866                                   unsigned int ext, unsigned int sz)
 867 {
 868         ext = htonl(ext);
 869         sz = htonl(sz);
 870         if ((ce_write(context, fd, &ext, 4) < 0) ||
 871             (ce_write(context, fd, &sz, 4) < 0))
 872                 return -1;
 873         return 0;
 874 }
 875
 876 static int ce_flush(SHA_CTX *context, int fd)
 877 {
 878         unsigned int left = write_buffer_len;
 879
 880         if (left) {
 881                 write_buffer_len = 0;
 882                 SHA1_Update(context, write_buffer, left);
 883         }
 884
 885         /* Flush first if not enough space for SHA1 signature */
 886         if (left + 20 > WRITE_BUFFER_SIZE) {
 887                 if (write(fd, write_buffer, left) != left)
 888                         return -1;
 889                 left = 0;
 890         }
 891
 892         /* Append the SHA1 signature at the end */
 893         SHA1_Final(write_buffer + left, context);
 894         left += 20;
 895         if (write(fd, write_buffer, left) != left)
 896                 return -1;
 897         return 0;
 898 }
 899
 900 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
 901 {
 902         /*
 903          * The only thing we care about in this function is to smudge the
 904          * falsely clean entry due to touch-update-touch race, so we leave
 905          * everything else as they are.  We are called for entries whose
 906          * ce_mtime match the index file mtime.
 907          */
 908         struct stat st;
 909
 910         if (lstat(ce->name, &st) < 0)
 911                 return;
 912         if (ce_match_stat_basic(ce, &st))
 913                 return;
 914         if (ce_modified_check_fs(ce, &st)) {
 915                 /* This is "racily clean"; smudge it.  Note that this
 916                  * is a tricky code.  At first glance, it may appear
 917                  * that it can break with this sequence:
 918                  *
 919                  * $ echo xyzzy >frotz
 920                  * $ git-update-index --add frotz
 921                  * $ : >frotz
 922                  * $ sleep 3
 923                  * $ echo filfre >nitfol
 924                  * $ git-update-index --add nitfol
 925                  *
 926                  * but it does not.  Whe the second update-index runs,
 927                  * it notices that the entry "frotz" has the same timestamp
 928                  * as index, and if we were to smudge it by resetting its
 929                  * size to zero here, then the object name recorded
 930                  * in index is the 6-byte file but the cached stat information
 931                  * becomes zero --- which would then match what we would
 932                  * obtain from the filesystem next time we stat("frotz").
 933                  *
 934                  * However, the second update-index, before calling
 935                  * this function, notices that the cached size is 6
 936                  * bytes and what is on the filesystem is an empty
 937                  * file, and never calls us, so the cached size information
 938                  * for "frotz" stays 6 which does not match the filesystem.
 939                  */
 940                 ce->ce_size = htonl(0);
 941         }
 942 }
 943
 944 int write_cache(int newfd, struct cache_entry **cache, int entries)
 945 {
 946         SHA_CTX c;
 947         struct cache_header hdr;
 948         int i, removed;
 949
 950         for (i = removed = 0; i < entries; i++)
 951                 if (!cache[i]->ce_mode)
 952                         removed++;
 953
 954         hdr.hdr_signature = htonl(CACHE_SIGNATURE);
 955         hdr.hdr_version = htonl(2);
 956         hdr.hdr_entries = htonl(entries - removed);
 957
 958         SHA1_Init(&c);
 959         if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
 960                 return -1;
 961
 962         for (i = 0; i < entries; i++) {
 963                 struct cache_entry *ce = cache[i];
 964                 if (!ce->ce_mode)
 965                         continue;
 966                 if (index_file_timestamp &&
 967                     index_file_timestamp <= ntohl(ce->ce_mtime.sec))
 968                         ce_smudge_racily_clean_entry(ce);
 969                 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
 970                         return -1;
 971         }
 972
 973         /* Write extension data here */
 974         if (active_cache_tree) {
 975                 unsigned long sz;
 976                 void *data = cache_tree_write(active_cache_tree, &sz);
 977                 if (data &&
 978                     !write_index_ext_header(&c, newfd, CACHE_EXT_TREE, sz) &&
 979                     !ce_write(&c, newfd, data, sz))
 980                         ;
 981                 else {
 982                         free(data);
 983                         return -1;
 984                 }
 985         }
 986         return ce_flush(&c, newfd);
 987 }