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;
  22 static time_t index_file_timestamp;
  23 unsigned int active_nr, active_alloc, active_cache_changed;
  24
  25 struct cache_tree *active_cache_tree;
  26
  27 int cache_errno;
  28
  29 static void *cache_mmap;
  30 static size_t cache_mmap_size;
  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 = hashcmp(sha1, ce->sha1);
  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 options)
 173 {
 174         unsigned int changed;
 175         int ignore_valid = options & 01;
 176         int assume_racy_is_modified = options & 02;
 177
 178         /*
 179          * If it's marked as always valid in the index, it's
 180          * valid whatever the checked-out copy says.
 181          */
 182         if (!ignore_valid && (ce->ce_flags & htons(CE_VALID)))
 183                 return 0;
 184
 185         changed = ce_match_stat_basic(ce, st);
 186
 187         /*
 188          * Within 1 second of this sequence:
 189          *      echo xyzzy >file && git-update-index --add file
 190          * running this command:
 191          *      echo frotz >file
 192          * would give a falsely clean cache entry.  The mtime and
 193          * length match the cache, and other stat fields do not change.
 194          *
 195          * We could detect this at update-index time (the cache entry
 196          * being registered/updated records the same time as "now")
 197          * and delay the return from git-update-index, but that would
 198          * effectively mean we can make at most one commit per second,
 199          * which is not acceptable.  Instead, we check cache entries
 200          * whose mtime are the same as the index file timestamp more
 201          * carefully than others.
 202          */
 203         if (!changed &&
 204             index_file_timestamp &&
 205             index_file_timestamp <= ntohl(ce->ce_mtime.sec)) {
 206                 if (assume_racy_is_modified)
 207                         changed |= DATA_CHANGED;
 208                 else
 209                         changed |= ce_modified_check_fs(ce, st);
 210         }
 211
 212         return changed;
 213 }
 214
 215 int ce_modified(struct cache_entry *ce, struct stat *st, int really)
 216 {
 217         int changed, changed_fs;
 218         changed = ce_match_stat(ce, st, really);
 219         if (!changed)
 220                 return 0;
 221         /*
 222          * If the mode or type has changed, there's no point in trying
 223          * to refresh the entry - it's not going to match
 224          */
 225         if (changed & (MODE_CHANGED | TYPE_CHANGED))
 226                 return changed;
 227
 228         /* Immediately after read-tree or update-index --cacheinfo,
 229          * the length field is zero.  For other cases the ce_size
 230          * should match the SHA1 recorded in the index entry.
 231          */
 232         if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
 233                 return changed;
 234
 235         changed_fs = ce_modified_check_fs(ce, st);
 236         if (changed_fs)
 237                 return changed | changed_fs;
 238         return 0;
 239 }
 240
 241 int base_name_compare(const char *name1, int len1, int mode1,
 242                       const char *name2, int len2, int mode2)
 243 {
 244         unsigned char c1, c2;
 245         int len = len1 < len2 ? len1 : len2;
 246         int cmp;
 247
 248         cmp = memcmp(name1, name2, len);
 249         if (cmp)
 250                 return cmp;
 251         c1 = name1[len];
 252         c2 = name2[len];
 253         if (!c1 && S_ISDIR(mode1))
 254                 c1 = '/';
 255         if (!c2 && S_ISDIR(mode2))
 256                 c2 = '/';
 257         return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
 258 }
 259
 260 int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
 261 {
 262         int len1 = flags1 & CE_NAMEMASK;
 263         int len2 = flags2 & CE_NAMEMASK;
 264         int len = len1 < len2 ? len1 : len2;
 265         int cmp;
 266
 267         cmp = memcmp(name1, name2, len);
 268         if (cmp)
 269                 return cmp;
 270         if (len1 < len2)
 271                 return -1;
 272         if (len1 > len2)
 273                 return 1;
 274
 275         /* Compare stages  */
 276         flags1 &= CE_STAGEMASK;
 277         flags2 &= CE_STAGEMASK;
 278
 279         if (flags1 < flags2)
 280                 return -1;
 281         if (flags1 > flags2)
 282                 return 1;
 283         return 0;
 284 }
 285
 286 int cache_name_pos(const char *name, int namelen)
 287 {
 288         int first, last;
 289
 290         first = 0;
 291         last = active_nr;
 292         while (last > first) {
 293                 int next = (last + first) >> 1;
 294                 struct cache_entry *ce = active_cache[next];
 295                 int cmp = cache_name_compare(name, namelen, ce->name, ntohs(ce->ce_flags));
 296                 if (!cmp)
 297                         return next;
 298                 if (cmp < 0) {
 299                         last = next;
 300                         continue;
 301                 }
 302                 first = next+1;
 303         }
 304         return -first-1;
 305 }
 306
 307 /* Remove entry, return true if there are more entries to go.. */
 308 int remove_cache_entry_at(int pos)
 309 {
 310         active_cache_changed = 1;
 311         active_nr--;
 312         if (pos >= active_nr)
 313                 return 0;
 314         memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
 315         return 1;
 316 }
 317
 318 int remove_file_from_cache(const char *path)
 319 {
 320         int pos = cache_name_pos(path, strlen(path));
 321         if (pos < 0)
 322                 pos = -pos-1;
 323         while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
 324                 remove_cache_entry_at(pos);
 325         return 0;
 326 }
 327
 328 int add_file_to_index(const char *path, int verbose)
 329 {
 330         int size, namelen;
 331         struct stat st;
 332         struct cache_entry *ce;
 333
 334         if (lstat(path, &st))
 335                 die("%s: unable to stat (%s)", path, strerror(errno));
 336
 337         if (!S_ISREG(st.st_mode) && !S_ISLNK(st.st_mode))
 338                 die("%s: can only add regular files or symbolic links", path);
 339
 340         namelen = strlen(path);
 341         size = cache_entry_size(namelen);
 342         ce = xcalloc(1, size);
 343         memcpy(ce->name, path, namelen);
 344         ce->ce_flags = htons(namelen);
 345         fill_stat_cache_info(ce, &st);
 346
 347         ce->ce_mode = create_ce_mode(st.st_mode);
 348         if (!trust_executable_bit) {
 349                 /* If there is an existing entry, pick the mode bits
 350                  * from it.
 351                  */
 352                 int pos = cache_name_pos(path, namelen);
 353                 if (pos >= 0)
 354                         ce->ce_mode = active_cache[pos]->ce_mode;
 355         }
 356
 357         if (index_path(ce->sha1, path, &st, 1))
 358                 die("unable to index file %s", path);
 359         if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD))
 360                 die("unable to add %s to index",path);
 361         if (verbose)
 362                 printf("add '%s'\n", path);
 363         cache_tree_invalidate_path(active_cache_tree, path);
 364         return 0;
 365 }
 366
 367 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
 368 {
 369         int len = ce_namelen(a);
 370         return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
 371 }
 372
 373 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
 374 {
 375         const char *match, *name;
 376         int len;
 377
 378         if (!pathspec)
 379                 return 1;
 380
 381         len = ce_namelen(ce);
 382         name = ce->name;
 383         while ((match = *pathspec++) != NULL) {
 384                 int matchlen = strlen(match);
 385                 if (matchlen > len)
 386                         continue;
 387                 if (memcmp(name, match, matchlen))
 388                         continue;
 389                 if (matchlen && name[matchlen-1] == '/')
 390                         return 1;
 391                 if (name[matchlen] == '/' || !name[matchlen])
 392                         return 1;
 393                 if (!matchlen)
 394                         return 1;
 395         }
 396         return 0;
 397 }
 398
 399 /*
 400  * We fundamentally don't like some paths: we don't want
 401  * dot or dot-dot anywhere, and for obvious reasons don't
 402  * want to recurse into ".git" either.
 403  *
 404  * Also, we don't want double slashes or slashes at the
 405  * end that can make pathnames ambiguous.
 406  */
 407 static int verify_dotfile(const char *rest)
 408 {
 409         /*
 410          * The first character was '.', but that
 411          * has already been discarded, we now test
 412          * the rest.
 413          */
 414         switch (*rest) {
 415         /* "." is not allowed */
 416         case '\0': case '/':
 417                 return 0;
 418
 419         /*
 420          * ".git" followed by  NUL or slash is bad. This
 421          * shares the path end test with the ".." case.
 422          */
 423         case 'g':
 424                 if (rest[1] != 'i')
 425                         break;
 426                 if (rest[2] != 't')
 427                         break;
 428                 rest += 2;
 429         /* fallthrough */
 430         case '.':
 431                 if (rest[1] == '\0' || rest[1] == '/')
 432                         return 0;
 433         }
 434         return 1;
 435 }
 436
 437 int verify_path(const char *path)
 438 {
 439         char c;
 440
 441         goto inside;
 442         for (;;) {
 443                 if (!c)
 444                         return 1;
 445                 if (c == '/') {
 446 inside:
 447                         c = *path++;
 448                         switch (c) {
 449                         default:
 450                                 continue;
 451                         case '/': case '\0':
 452                                 break;
 453                         case '.':
 454                                 if (verify_dotfile(path))
 455                                         continue;
 456                         }
 457                         return 0;
 458                 }
 459                 c = *path++;
 460         }
 461 }
 462
 463 /*
 464  * Do we have another file that has the beginning components being a
 465  * proper superset of the name we're trying to add?
 466  */
 467 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 468 {
 469         int retval = 0;
 470         int len = ce_namelen(ce);
 471         int stage = ce_stage(ce);
 472         const char *name = ce->name;
 473
 474         while (pos < active_nr) {
 475                 struct cache_entry *p = active_cache[pos++];
 476
 477                 if (len >= ce_namelen(p))
 478                         break;
 479                 if (memcmp(name, p->name, len))
 480                         break;
 481                 if (ce_stage(p) != stage)
 482                         continue;
 483                 if (p->name[len] != '/')
 484                         continue;
 485                 retval = -1;
 486                 if (!ok_to_replace)
 487                         break;
 488                 remove_cache_entry_at(--pos);
 489         }
 490         return retval;
 491 }
 492
 493 /*
 494  * Do we have another file with a pathname that is a proper
 495  * subset of the name we're trying to add?
 496  */
 497 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 498 {
 499         int retval = 0;
 500         int stage = ce_stage(ce);
 501         const char *name = ce->name;
 502         const char *slash = name + ce_namelen(ce);
 503
 504         for (;;) {
 505                 int len;
 506
 507                 for (;;) {
 508                         if (*--slash == '/')
 509                                 break;
 510                         if (slash <= ce->name)
 511                                 return retval;
 512                 }
 513                 len = slash - name;
 514
 515                 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
 516                 if (pos >= 0) {
 517                         retval = -1;
 518                         if (ok_to_replace)
 519                                 break;
 520                         remove_cache_entry_at(pos);
 521                         continue;
 522                 }
 523
 524                 /*
 525                  * Trivial optimization: if we find an entry that
 526                  * already matches the sub-directory, then we know
 527                  * we're ok, and we can exit.
 528                  */
 529                 pos = -pos-1;
 530                 while (pos < active_nr) {
 531                         struct cache_entry *p = active_cache[pos];
 532                         if ((ce_namelen(p) <= len) ||
 533                             (p->name[len] != '/') ||
 534                             memcmp(p->name, name, len))
 535                                 break; /* not our subdirectory */
 536                         if (ce_stage(p) == stage)
 537                                 /* p is at the same stage as our entry, and
 538                                  * is a subdirectory of what we are looking
 539                                  * at, so we cannot have conflicts at our
 540                                  * level or anything shorter.
 541                                  */
 542                                 return retval;
 543                         pos++;
 544                 }
 545         }
 546         return retval;
 547 }
 548
 549 /* We may be in a situation where we already have path/file and path
 550  * is being added, or we already have path and path/file is being
 551  * added.  Either one would result in a nonsense tree that has path
 552  * twice when git-write-tree tries to write it out.  Prevent it.
 553  *
 554  * If ok-to-replace is specified, we remove the conflicting entries
 555  * from the cache so the caller should recompute the insert position.
 556  * When this happens, we return non-zero.
 557  */
 558 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
 559 {
 560         /*
 561          * We check if the path is a sub-path of a subsequent pathname
 562          * first, since removing those will not change the position
 563          * in the array
 564          */
 565         int retval = has_file_name(ce, pos, ok_to_replace);
 566         /*
 567          * Then check if the path might have a clashing sub-directory
 568          * before it.
 569          */
 570         return retval + has_dir_name(ce, pos, ok_to_replace);
 571 }
 572
 573 int add_cache_entry(struct cache_entry *ce, int option)
 574 {
 575         int pos;
 576         int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
 577         int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
 578         int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
 579
 580         pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 581
 582         /* existing match? Just replace it. */
 583         if (pos >= 0) {
 584                 active_cache_changed = 1;
 585                 active_cache[pos] = ce;
 586                 return 0;
 587         }
 588         pos = -pos-1;
 589
 590         /*
 591          * Inserting a merged entry ("stage 0") into the index
 592          * will always replace all non-merged entries..
 593          */
 594         if (pos < active_nr && ce_stage(ce) == 0) {
 595                 while (ce_same_name(active_cache[pos], ce)) {
 596                         ok_to_add = 1;
 597                         if (!remove_cache_entry_at(pos))
 598                                 break;
 599                 }
 600         }
 601
 602         if (!ok_to_add)
 603                 return -1;
 604         if (!verify_path(ce->name))
 605                 return -1;
 606
 607         if (!skip_df_check &&
 608             check_file_directory_conflict(ce, pos, ok_to_replace)) {
 609                 if (!ok_to_replace)
 610                         return -1;
 611                 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 612                 pos = -pos-1;
 613         }
 614
 615         /* Make sure the array is big enough .. */
 616         if (active_nr == active_alloc) {
 617                 active_alloc = alloc_nr(active_alloc);
 618                 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
 619         }
 620
 621         /* Add it in.. */
 622         active_nr++;
 623         if (active_nr > pos)
 624                 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
 625         active_cache[pos] = ce;
 626         active_cache_changed = 1;
 627         return 0;
 628 }
 629
 630 /*
 631  * "refresh" does not calculate a new sha1 file or bring the
 632  * cache up-to-date for mode/content changes. But what it
 633  * _does_ do is to "re-match" the stat information of a file
 634  * with the cache, so that you can refresh the cache for a
 635  * file that hasn't been changed but where the stat entry is
 636  * out of date.
 637  *
 638  * For example, you'd want to do this after doing a "git-read-tree",
 639  * to link up the stat cache details with the proper files.
 640  */
 641 struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
 642 {
 643         struct stat st;
 644         struct cache_entry *updated;
 645         int changed, size;
 646
 647         if (lstat(ce->name, &st) < 0) {
 648                 cache_errno = errno;
 649                 return NULL;
 650         }
 651
 652         changed = ce_match_stat(ce, &st, really);
 653         if (!changed) {
 654                 if (really && assume_unchanged &&
 655                     !(ce->ce_flags & htons(CE_VALID)))
 656                         ; /* mark this one VALID again */
 657                 else
 658                         return ce;
 659         }
 660
 661         if (ce_modified(ce, &st, really)) {
 662                 cache_errno = EINVAL;
 663                 return NULL;
 664         }
 665
 666         size = ce_size(ce);
 667         updated = xmalloc(size);
 668         memcpy(updated, ce, size);
 669         fill_stat_cache_info(updated, &st);
 670
 671         /* In this case, if really is not set, we should leave
 672          * CE_VALID bit alone.  Otherwise, paths marked with
 673          * --no-assume-unchanged (i.e. things to be edited) will
 674          * reacquire CE_VALID bit automatically, which is not
 675          * really what we want.
 676          */
 677         if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
 678                 updated->ce_flags &= ~htons(CE_VALID);
 679
 680         return updated;
 681 }
 682
 683 int refresh_cache(unsigned int flags)
 684 {
 685         int i;
 686         int has_errors = 0;
 687         int really = (flags & REFRESH_REALLY) != 0;
 688         int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
 689         int quiet = (flags & REFRESH_QUIET) != 0;
 690         int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
 691
 692         for (i = 0; i < active_nr; i++) {
 693                 struct cache_entry *ce, *new;
 694                 ce = active_cache[i];
 695                 if (ce_stage(ce)) {
 696                         while ((i < active_nr) &&
 697                                ! strcmp(active_cache[i]->name, ce->name))
 698                                 i++;
 699                         i--;
 700                         if (allow_unmerged)
 701                                 continue;
 702                         printf("%s: needs merge\n", ce->name);
 703                         has_errors = 1;
 704                         continue;
 705                 }
 706
 707                 new = refresh_cache_entry(ce, really);
 708                 if (new == ce)
 709                         continue;
 710                 if (!new) {
 711                         if (not_new && cache_errno == ENOENT)
 712                                 continue;
 713                         if (really && cache_errno == EINVAL) {
 714                                 /* If we are doing --really-refresh that
 715                                  * means the index is not valid anymore.
 716                                  */
 717                                 ce->ce_flags &= ~htons(CE_VALID);
 718                                 active_cache_changed = 1;
 719                         }
 720                         if (quiet)
 721                                 continue;
 722                         printf("%s: needs update\n", ce->name);
 723                         has_errors = 1;
 724                         continue;
 725                 }
 726                 active_cache_changed = 1;
 727                 /* You can NOT just free active_cache[i] here, since it
 728                  * might not be necessarily malloc()ed but can also come
 729                  * from mmap(). */
 730                 active_cache[i] = new;
 731         }
 732         return has_errors;
 733 }
 734
 735 static int verify_hdr(struct cache_header *hdr, unsigned long size)
 736 {
 737         SHA_CTX c;
 738         unsigned char sha1[20];
 739
 740         if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
 741                 return error("bad signature");
 742         if (hdr->hdr_version != htonl(2))
 743                 return error("bad index version");
 744         SHA1_Init(&c);
 745         SHA1_Update(&c, hdr, size - 20);
 746         SHA1_Final(sha1, &c);
 747         if (hashcmp(sha1, (unsigned char *)hdr + size - 20))
 748                 return error("bad index file sha1 signature");
 749         return 0;
 750 }
 751
 752 static int read_index_extension(const char *ext, void *data, unsigned long sz)
 753 {
 754         switch (CACHE_EXT(ext)) {
 755         case CACHE_EXT_TREE:
 756                 active_cache_tree = cache_tree_read(data, sz);
 757                 break;
 758         default:
 759                 if (*ext < 'A' || 'Z' < *ext)
 760                         return error("index uses %.4s extension, which we do not understand",
 761                                      ext);
 762                 fprintf(stderr, "ignoring %.4s extension\n", ext);
 763                 break;
 764         }
 765         return 0;
 766 }
 767
 768 int read_cache(void)
 769 {
 770         return read_cache_from(get_index_file());
 771 }
 772
 773 /* remember to discard_cache() before reading a different cache! */
 774 int read_cache_from(const char *path)
 775 {
 776         int fd, i;
 777         struct stat st;
 778         unsigned long offset;
 779         struct cache_header *hdr;
 780
 781         errno = EBUSY;
 782         if (cache_mmap)
 783                 return active_nr;
 784
 785         errno = ENOENT;
 786         index_file_timestamp = 0;
 787         fd = open(path, O_RDONLY);
 788         if (fd < 0) {
 789                 if (errno == ENOENT)
 790                         return 0;
 791                 die("index file open failed (%s)", strerror(errno));
 792         }
 793
 794         cache_mmap = MAP_FAILED;
 795         if (!fstat(fd, &st)) {
 796                 cache_mmap_size = st.st_size;
 797                 errno = EINVAL;
 798                 if (cache_mmap_size >= sizeof(struct cache_header) + 20)
 799                         cache_mmap = mmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
 800         }
 801         close(fd);
 802         if (cache_mmap == MAP_FAILED)
 803                 die("index file mmap failed (%s)", strerror(errno));
 804
 805         hdr = cache_mmap;
 806         if (verify_hdr(hdr, cache_mmap_size) < 0)
 807                 goto unmap;
 808
 809         active_nr = ntohl(hdr->hdr_entries);
 810         active_alloc = alloc_nr(active_nr);
 811         active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
 812
 813         offset = sizeof(*hdr);
 814         for (i = 0; i < active_nr; i++) {
 815                 struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
 816                 offset = offset + ce_size(ce);
 817                 active_cache[i] = ce;
 818         }
 819         index_file_timestamp = st.st_mtime;
 820         while (offset <= cache_mmap_size - 20 - 8) {
 821                 /* After an array of active_nr index entries,
 822                  * there can be arbitrary number of extended
 823                  * sections, each of which is prefixed with
 824                  * extension name (4-byte) and section length
 825                  * in 4-byte network byte order.
 826                  */
 827                 unsigned long extsize;
 828                 memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
 829                 extsize = ntohl(extsize);
 830                 if (read_index_extension(((const char *) cache_mmap) + offset,
 831                                          (char *) cache_mmap + offset + 8,
 832                                          extsize) < 0)
 833                         goto unmap;
 834                 offset += 8;
 835                 offset += extsize;
 836         }
 837         return active_nr;
 838
 839 unmap:
 840         munmap(cache_mmap, cache_mmap_size);
 841         errno = EINVAL;
 842         die("index file corrupt");
 843 }
 844
 845 #define WRITE_BUFFER_SIZE 8192
 846 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
 847 static unsigned long write_buffer_len;
 848
 849 static int ce_write_flush(SHA_CTX *context, int fd)
 850 {
 851         unsigned int buffered = write_buffer_len;
 852         if (buffered) {
 853                 SHA1_Update(context, write_buffer, buffered);
 854                 if (write(fd, write_buffer, buffered) != buffered)
 855                         return -1;
 856                 write_buffer_len = 0;
 857         }
 858         return 0;
 859 }
 860
 861 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
 862 {
 863         while (len) {
 864                 unsigned int buffered = write_buffer_len;
 865                 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
 866                 if (partial > len)
 867                         partial = len;
 868                 memcpy(write_buffer + buffered, data, partial);
 869                 buffered += partial;
 870                 if (buffered == WRITE_BUFFER_SIZE) {
 871                         write_buffer_len = buffered;
 872                         if (ce_write_flush(context, fd))
 873                                 return -1;
 874                         buffered = 0;
 875                 }
 876                 write_buffer_len = buffered;
 877                 len -= partial;
 878                 data = (char *) data + partial;
 879         }
 880         return 0;
 881 }
 882
 883 static int write_index_ext_header(SHA_CTX *context, int fd,
 884                                   unsigned int ext, unsigned int sz)
 885 {
 886         ext = htonl(ext);
 887         sz = htonl(sz);
 888         return ((ce_write(context, fd, &ext, 4) < 0) ||
 889                 (ce_write(context, fd, &sz, 4) < 0)) ? -1 : 0;
 890 }
 891
 892 static int ce_flush(SHA_CTX *context, int fd)
 893 {
 894         unsigned int left = write_buffer_len;
 895
 896         if (left) {
 897                 write_buffer_len = 0;
 898                 SHA1_Update(context, write_buffer, left);
 899         }
 900
 901         /* Flush first if not enough space for SHA1 signature */
 902         if (left + 20 > WRITE_BUFFER_SIZE) {
 903                 if (write(fd, write_buffer, left) != left)
 904                         return -1;
 905                 left = 0;
 906         }
 907
 908         /* Append the SHA1 signature at the end */
 909         SHA1_Final(write_buffer + left, context);
 910         left += 20;
 911         return (write(fd, write_buffer, left) != left) ? -1 : 0;
 912 }
 913
 914 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
 915 {
 916         /*
 917          * The only thing we care about in this function is to smudge the
 918          * falsely clean entry due to touch-update-touch race, so we leave
 919          * everything else as they are.  We are called for entries whose
 920          * ce_mtime match the index file mtime.
 921          */
 922         struct stat st;
 923
 924         if (lstat(ce->name, &st) < 0)
 925                 return;
 926         if (ce_match_stat_basic(ce, &st))
 927                 return;
 928         if (ce_modified_check_fs(ce, &st)) {
 929                 /* This is "racily clean"; smudge it.  Note that this
 930                  * is a tricky code.  At first glance, it may appear
 931                  * that it can break with this sequence:
 932                  *
 933                  * $ echo xyzzy >frotz
 934                  * $ git-update-index --add frotz
 935                  * $ : >frotz
 936                  * $ sleep 3
 937                  * $ echo filfre >nitfol
 938                  * $ git-update-index --add nitfol
 939                  *
 940                  * but it does not.  When the second update-index runs,
 941                  * it notices that the entry "frotz" has the same timestamp
 942                  * as index, and if we were to smudge it by resetting its
 943                  * size to zero here, then the object name recorded
 944                  * in index is the 6-byte file but the cached stat information
 945                  * becomes zero --- which would then match what we would
 946                  * obtain from the filesystem next time we stat("frotz").
 947                  *
 948                  * However, the second update-index, before calling
 949                  * this function, notices that the cached size is 6
 950                  * bytes and what is on the filesystem is an empty
 951                  * file, and never calls us, so the cached size information
 952                  * for "frotz" stays 6 which does not match the filesystem.
 953                  */
 954                 ce->ce_size = htonl(0);
 955         }
 956 }
 957
 958 int write_cache(int newfd, struct cache_entry **cache, int entries)
 959 {
 960         SHA_CTX c;
 961         struct cache_header hdr;
 962         int i, removed;
 963
 964         for (i = removed = 0; i < entries; i++)
 965                 if (!cache[i]->ce_mode)
 966                         removed++;
 967
 968         hdr.hdr_signature = htonl(CACHE_SIGNATURE);
 969         hdr.hdr_version = htonl(2);
 970         hdr.hdr_entries = htonl(entries - removed);
 971
 972         SHA1_Init(&c);
 973         if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
 974                 return -1;
 975
 976         for (i = 0; i < entries; i++) {
 977                 struct cache_entry *ce = cache[i];
 978                 if (!ce->ce_mode)
 979                         continue;
 980                 if (index_file_timestamp &&
 981                     index_file_timestamp <= ntohl(ce->ce_mtime.sec))
 982                         ce_smudge_racily_clean_entry(ce);
 983                 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
 984                         return -1;
 985         }
 986
 987         /* Write extension data here */
 988         if (active_cache_tree) {
 989                 unsigned long sz;
 990                 void *data = cache_tree_write(active_cache_tree, &sz);
 991                 if (data &&
 992                     !write_index_ext_header(&c, newfd, CACHE_EXT_TREE, sz) &&
 993                     !ce_write(&c, newfd, data, sz))
 994                         ;
 995                 else {
 996                         free(data);
 997                         return -1;
 998                 }
 999         }
1000         return ce_flush(&c, newfd);
1001 }