read-cache.c

   1 /*
   2  * GIT - The information manager from hell
   3  *
   4  * Copyright (C) Linus Torvalds, 2005
   5  */
   6 #include "cache.h"
   7
   8 struct cache_entry **active_cache = NULL;
   9 static time_t index_file_timestamp;
  10 unsigned int active_nr = 0, active_alloc = 0, active_cache_changed = 0;
  11
  12 /*
  13  * This only updates the "non-critical" parts of the directory
  14  * cache, ie the parts that aren't tracked by GIT, and only used
  15  * to validate the cache.
  16  */
  17 void fill_stat_cache_info(struct cache_entry *ce, struct stat *st)
  18 {
  19         ce->ce_ctime.sec = htonl(st->st_ctime);
  20         ce->ce_mtime.sec = htonl(st->st_mtime);
  21 #ifdef USE_NSEC
  22         ce->ce_ctime.nsec = htonl(st->st_ctim.tv_nsec);
  23         ce->ce_mtime.nsec = htonl(st->st_mtim.tv_nsec);
  24 #endif
  25         ce->ce_dev = htonl(st->st_dev);
  26         ce->ce_ino = htonl(st->st_ino);
  27         ce->ce_uid = htonl(st->st_uid);
  28         ce->ce_gid = htonl(st->st_gid);
  29         ce->ce_size = htonl(st->st_size);
  30
  31         if (assume_unchanged)
  32                 ce->ce_flags |= htons(CE_VALID);
  33 }
  34
  35 static int ce_compare_data(struct cache_entry *ce, struct stat *st)
  36 {
  37         int match = -1;
  38         int fd = open(ce->name, O_RDONLY);
  39
  40         if (fd >= 0) {
  41                 unsigned char sha1[20];
  42                 if (!index_fd(sha1, fd, st, 0, NULL))
  43                         match = memcmp(sha1, ce->sha1, 20);
  44                 close(fd);
  45         }
  46         return match;
  47 }
  48
  49 static int ce_compare_link(struct cache_entry *ce, unsigned long expected_size)
  50 {
  51         int match = -1;
  52         char *target;
  53         void *buffer;
  54         unsigned long size;
  55         char type[10];
  56         int len;
  57
  58         target = xmalloc(expected_size);
  59         len = readlink(ce->name, target, expected_size);
  60         if (len != expected_size) {
  61                 free(target);
  62                 return -1;
  63         }
  64         buffer = read_sha1_file(ce->sha1, type, &size);
  65         if (!buffer) {
  66                 free(target);
  67                 return -1;
  68         }
  69         if (size == expected_size)
  70                 match = memcmp(buffer, target, size);
  71         free(buffer);
  72         free(target);
  73         return match;
  74 }
  75
  76 static int ce_modified_check_fs(struct cache_entry *ce, struct stat *st)
  77 {
  78         switch (st->st_mode & S_IFMT) {
  79         case S_IFREG:
  80                 if (ce_compare_data(ce, st))
  81                         return DATA_CHANGED;
  82                 break;
  83         case S_IFLNK:
  84                 if (ce_compare_link(ce, st->st_size))
  85                         return DATA_CHANGED;
  86                 break;
  87         default:
  88                 return TYPE_CHANGED;
  89         }
  90         return 0;
  91 }
  92
  93 static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
  94 {
  95         unsigned int changed = 0;
  96
  97         switch (ntohl(ce->ce_mode) & S_IFMT) {
  98         case S_IFREG:
  99                 changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
 100                 /* We consider only the owner x bit to be relevant for
 101                  * "mode changes"
 102                  */
 103                 if (trust_executable_bit &&
 104                     (0100 & (ntohl(ce->ce_mode) ^ st->st_mode)))
 105                         changed |= MODE_CHANGED;
 106                 break;
 107         case S_IFLNK:
 108                 changed |= !S_ISLNK(st->st_mode) ? TYPE_CHANGED : 0;
 109                 break;
 110         default:
 111                 die("internal error: ce_mode is %o", ntohl(ce->ce_mode));
 112         }
 113         if (ce->ce_mtime.sec != htonl(st->st_mtime))
 114                 changed |= MTIME_CHANGED;
 115         if (ce->ce_ctime.sec != htonl(st->st_ctime))
 116                 changed |= CTIME_CHANGED;
 117
 118 #ifdef USE_NSEC
 119         /*
 120          * nsec seems unreliable - not all filesystems support it, so
 121          * as long as it is in the inode cache you get right nsec
 122          * but after it gets flushed, you get zero nsec.
 123          */
 124         if (ce->ce_mtime.nsec != htonl(st->st_mtim.tv_nsec))
 125                 changed |= MTIME_CHANGED;
 126         if (ce->ce_ctime.nsec != htonl(st->st_ctim.tv_nsec))
 127                 changed |= CTIME_CHANGED;
 128 #endif
 129
 130         if (ce->ce_uid != htonl(st->st_uid) ||
 131             ce->ce_gid != htonl(st->st_gid))
 132                 changed |= OWNER_CHANGED;
 133         if (ce->ce_ino != htonl(st->st_ino))
 134                 changed |= INODE_CHANGED;
 135
 136 #ifdef USE_STDEV
 137         /*
 138          * st_dev breaks on network filesystems where different
 139          * clients will have different views of what "device"
 140          * the filesystem is on
 141          */
 142         if (ce->ce_dev != htonl(st->st_dev))
 143                 changed |= INODE_CHANGED;
 144 #endif
 145
 146         if (ce->ce_size != htonl(st->st_size))
 147                 changed |= DATA_CHANGED;
 148
 149         return changed;
 150 }
 151
 152 int ce_match_stat(struct cache_entry *ce, struct stat *st, int ignore_valid)
 153 {
 154         unsigned int changed;
 155
 156         /*
 157          * If it's marked as always valid in the index, it's
 158          * valid whatever the checked-out copy says.
 159          */
 160         if (!ignore_valid && (ce->ce_flags & htons(CE_VALID)))
 161                 return 0;
 162
 163         changed = ce_match_stat_basic(ce, st);
 164
 165         /*
 166          * Within 1 second of this sequence:
 167          *      echo xyzzy >file && git-update-index --add file
 168          * running this command:
 169          *      echo frotz >file
 170          * would give a falsely clean cache entry.  The mtime and
 171          * length match the cache, and other stat fields do not change.
 172          *
 173          * We could detect this at update-index time (the cache entry
 174          * being registered/updated records the same time as "now")
 175          * and delay the return from git-update-index, but that would
 176          * effectively mean we can make at most one commit per second,
 177          * which is not acceptable.  Instead, we check cache entries
 178          * whose mtime are the same as the index file timestamp more
 179          * carefully than others.
 180          */
 181         if (!changed &&
 182             index_file_timestamp &&
 183             index_file_timestamp <= ntohl(ce->ce_mtime.sec))
 184                 changed |= ce_modified_check_fs(ce, st);
 185
 186         return changed;
 187 }
 188
 189 int ce_modified(struct cache_entry *ce, struct stat *st, int really)
 190 {
 191         int changed, changed_fs;
 192         changed = ce_match_stat(ce, st, really);
 193         if (!changed)
 194                 return 0;
 195         /*
 196          * If the mode or type has changed, there's no point in trying
 197          * to refresh the entry - it's not going to match
 198          */
 199         if (changed & (MODE_CHANGED | TYPE_CHANGED))
 200                 return changed;
 201
 202         /* Immediately after read-tree or update-index --cacheinfo,
 203          * the length field is zero.  For other cases the ce_size
 204          * should match the SHA1 recorded in the index entry.
 205          */
 206         if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
 207                 return changed;
 208
 209         changed_fs = ce_modified_check_fs(ce, st);
 210         if (changed_fs)
 211                 return changed | changed_fs;
 212         return 0;
 213 }
 214
 215 int base_name_compare(const char *name1, int len1, int mode1,
 216                       const char *name2, int len2, int mode2)
 217 {
 218         unsigned char c1, c2;
 219         int len = len1 < len2 ? len1 : len2;
 220         int cmp;
 221
 222         cmp = memcmp(name1, name2, len);
 223         if (cmp)
 224                 return cmp;
 225         c1 = name1[len];
 226         c2 = name2[len];
 227         if (!c1 && S_ISDIR(mode1))
 228                 c1 = '/';
 229         if (!c2 && S_ISDIR(mode2))
 230                 c2 = '/';
 231         return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
 232 }
 233
 234 int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
 235 {
 236         int len1 = flags1 & CE_NAMEMASK;
 237         int len2 = flags2 & CE_NAMEMASK;
 238         int len = len1 < len2 ? len1 : len2;
 239         int cmp;
 240
 241         cmp = memcmp(name1, name2, len);
 242         if (cmp)
 243                 return cmp;
 244         if (len1 < len2)
 245                 return -1;
 246         if (len1 > len2)
 247                 return 1;
 248
 249         /* Compare stages  */
 250         flags1 &= CE_STAGEMASK;
 251         flags2 &= CE_STAGEMASK;
 252
 253         if (flags1 < flags2)
 254                 return -1;
 255         if (flags1 > flags2)
 256                 return 1;
 257         return 0;
 258 }
 259
 260 int cache_name_pos(const char *name, int namelen)
 261 {
 262         int first, last;
 263
 264         first = 0;
 265         last = active_nr;
 266         while (last > first) {
 267                 int next = (last + first) >> 1;
 268                 struct cache_entry *ce = active_cache[next];
 269                 int cmp = cache_name_compare(name, namelen, ce->name, ntohs(ce->ce_flags));
 270                 if (!cmp)
 271                         return next;
 272                 if (cmp < 0) {
 273                         last = next;
 274                         continue;
 275                 }
 276                 first = next+1;
 277         }
 278         return -first-1;
 279 }
 280
 281 /* Remove entry, return true if there are more entries to go.. */
 282 int remove_cache_entry_at(int pos)
 283 {
 284         active_cache_changed = 1;
 285         active_nr--;
 286         if (pos >= active_nr)
 287                 return 0;
 288         memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
 289         return 1;
 290 }
 291
 292 int remove_file_from_cache(const char *path)
 293 {
 294         int pos = cache_name_pos(path, strlen(path));
 295         if (pos < 0)
 296                 pos = -pos-1;
 297         while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
 298                 remove_cache_entry_at(pos);
 299         return 0;
 300 }
 301
 302 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
 303 {
 304         int len = ce_namelen(a);
 305         return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
 306 }
 307
 308 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
 309 {
 310         const char *match, *name;
 311         int len;
 312
 313         if (!pathspec)
 314                 return 1;
 315
 316         len = ce_namelen(ce);
 317         name = ce->name;
 318         while ((match = *pathspec++) != NULL) {
 319                 int matchlen = strlen(match);
 320                 if (matchlen > len)
 321                         continue;
 322                 if (memcmp(name, match, matchlen))
 323                         continue;
 324                 if (matchlen && name[matchlen-1] == '/')
 325                         return 1;
 326                 if (name[matchlen] == '/' || !name[matchlen])
 327                         return 1;
 328                 if (!matchlen)
 329                         return 1;
 330         }
 331         return 0;
 332 }
 333
 334 /*
 335  * We fundamentally don't like some paths: we don't want
 336  * dot or dot-dot anywhere, and for obvious reasons don't
 337  * want to recurse into ".git" either.
 338  *
 339  * Also, we don't want double slashes or slashes at the
 340  * end that can make pathnames ambiguous.
 341  */
 342 static int verify_dotfile(const char *rest)
 343 {
 344         /*
 345          * The first character was '.', but that
 346          * has already been discarded, we now test
 347          * the rest.
 348          */
 349         switch (*rest) {
 350         /* "." is not allowed */
 351         case '\0': case '/':
 352                 return 0;
 353
 354         /*
 355          * ".git" followed by  NUL or slash is bad. This
 356          * shares the path end test with the ".." case.
 357          */
 358         case 'g':
 359                 if (rest[1] != 'i')
 360                         break;
 361                 if (rest[2] != 't')
 362                         break;
 363                 rest += 2;
 364         /* fallthrough */
 365         case '.':
 366                 if (rest[1] == '\0' || rest[1] == '/')
 367                         return 0;
 368         }
 369         return 1;
 370 }
 371
 372 int verify_path(const char *path)
 373 {
 374         char c;
 375
 376         goto inside;
 377         for (;;) {
 378                 if (!c)
 379                         return 1;
 380                 if (c == '/') {
 381 inside:
 382                         c = *path++;
 383                         switch (c) {
 384                         default:
 385                                 continue;
 386                         case '/': case '\0':
 387                                 break;
 388                         case '.':
 389                                 if (verify_dotfile(path))
 390                                         continue;
 391                         }
 392                         return 0;
 393                 }
 394                 c = *path++;
 395         }
 396 }
 397
 398 /*
 399  * Do we have another file that has the beginning components being a
 400  * proper superset of the name we're trying to add?
 401  */
 402 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 403 {
 404         int retval = 0;
 405         int len = ce_namelen(ce);
 406         int stage = ce_stage(ce);
 407         const char *name = ce->name;
 408
 409         while (pos < active_nr) {
 410                 struct cache_entry *p = active_cache[pos++];
 411
 412                 if (len >= ce_namelen(p))
 413                         break;
 414                 if (memcmp(name, p->name, len))
 415                         break;
 416                 if (ce_stage(p) != stage)
 417                         continue;
 418                 if (p->name[len] != '/')
 419                         continue;
 420                 retval = -1;
 421                 if (!ok_to_replace)
 422                         break;
 423                 remove_cache_entry_at(--pos);
 424         }
 425         return retval;
 426 }
 427
 428 /*
 429  * Do we have another file with a pathname that is a proper
 430  * subset of the name we're trying to add?
 431  */
 432 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 433 {
 434         int retval = 0;
 435         int stage = ce_stage(ce);
 436         const char *name = ce->name;
 437         const char *slash = name + ce_namelen(ce);
 438
 439         for (;;) {
 440                 int len;
 441
 442                 for (;;) {
 443                         if (*--slash == '/')
 444                                 break;
 445                         if (slash <= ce->name)
 446                                 return retval;
 447                 }
 448                 len = slash - name;
 449
 450                 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
 451                 if (pos >= 0) {
 452                         retval = -1;
 453                         if (ok_to_replace)
 454                                 break;
 455                         remove_cache_entry_at(pos);
 456                         continue;
 457                 }
 458
 459                 /*
 460                  * Trivial optimization: if we find an entry that
 461                  * already matches the sub-directory, then we know
 462                  * we're ok, and we can exit.
 463                  */
 464                 pos = -pos-1;
 465                 while (pos < active_nr) {
 466                         struct cache_entry *p = active_cache[pos];
 467                         if ((ce_namelen(p) <= len) ||
 468                             (p->name[len] != '/') ||
 469                             memcmp(p->name, name, len))
 470                                 break; /* not our subdirectory */
 471                         if (ce_stage(p) == stage)
 472                                 /* p is at the same stage as our entry, and
 473                                  * is a subdirectory of what we are looking
 474                                  * at, so we cannot have conflicts at our
 475                                  * level or anything shorter.
 476                                  */
 477                                 return retval;
 478                         pos++;
 479                 }
 480         }
 481         return retval;
 482 }
 483
 484 /* We may be in a situation where we already have path/file and path
 485  * is being added, or we already have path and path/file is being
 486  * added.  Either one would result in a nonsense tree that has path
 487  * twice when git-write-tree tries to write it out.  Prevent it.
 488  *
 489  * If ok-to-replace is specified, we remove the conflicting entries
 490  * from the cache so the caller should recompute the insert position.
 491  * When this happens, we return non-zero.
 492  */
 493 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
 494 {
 495         /*
 496          * We check if the path is a sub-path of a subsequent pathname
 497          * first, since removing those will not change the position
 498          * in the array
 499          */
 500         int retval = has_file_name(ce, pos, ok_to_replace);
 501         /*
 502          * Then check if the path might have a clashing sub-directory
 503          * before it.
 504          */
 505         return retval + has_dir_name(ce, pos, ok_to_replace);
 506 }
 507
 508 int add_cache_entry(struct cache_entry *ce, int option)
 509 {
 510         int pos;
 511         int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
 512         int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
 513         int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
 514
 515         pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 516
 517         /* existing match? Just replace it. */
 518         if (pos >= 0) {
 519                 active_cache_changed = 1;
 520                 active_cache[pos] = ce;
 521                 return 0;
 522         }
 523         pos = -pos-1;
 524
 525         /*
 526          * Inserting a merged entry ("stage 0") into the index
 527          * will always replace all non-merged entries..
 528          */
 529         if (pos < active_nr && ce_stage(ce) == 0) {
 530                 while (ce_same_name(active_cache[pos], ce)) {
 531                         ok_to_add = 1;
 532                         if (!remove_cache_entry_at(pos))
 533                                 break;
 534                 }
 535         }
 536
 537         if (!ok_to_add)
 538                 return -1;
 539         if (!verify_path(ce->name))
 540                 return -1;
 541
 542         if (!skip_df_check &&
 543             check_file_directory_conflict(ce, pos, ok_to_replace)) {
 544                 if (!ok_to_replace)
 545                         return -1;
 546                 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 547                 pos = -pos-1;
 548         }
 549
 550         /* Make sure the array is big enough .. */
 551         if (active_nr == active_alloc) {
 552                 active_alloc = alloc_nr(active_alloc);
 553                 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
 554         }
 555
 556         /* Add it in.. */
 557         active_nr++;
 558         if (active_nr > pos)
 559                 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
 560         active_cache[pos] = ce;
 561         active_cache_changed = 1;
 562         return 0;
 563 }
 564
 565 /* Three functions to allow overloaded pointer return; see linux/err.h */
 566 static inline void *ERR_PTR(long error)
 567 {
 568         return (void *) error;
 569 }
 570
 571 static inline long PTR_ERR(const void *ptr)
 572 {
 573         return (long) ptr;
 574 }
 575
 576 static inline long IS_ERR(const void *ptr)
 577 {
 578         return (unsigned long)ptr > (unsigned long)-1000L;
 579 }
 580
 581 /*
 582  * "refresh" does not calculate a new sha1 file or bring the
 583  * cache up-to-date for mode/content changes. But what it
 584  * _does_ do is to "re-match" the stat information of a file
 585  * with the cache, so that you can refresh the cache for a
 586  * file that hasn't been changed but where the stat entry is
 587  * out of date.
 588  *
 589  * For example, you'd want to do this after doing a "git-read-tree",
 590  * to link up the stat cache details with the proper files.
 591  */
 592 static struct cache_entry *refresh_entry(struct cache_entry *ce, int really)
 593 {
 594         struct stat st;
 595         struct cache_entry *updated;
 596         int changed, size;
 597
 598         if (lstat(ce->name, &st) < 0)
 599                 return ERR_PTR(-errno);
 600
 601         changed = ce_match_stat(ce, &st, really);
 602         if (!changed) {
 603                 if (really && assume_unchanged &&
 604                     !(ce->ce_flags & htons(CE_VALID)))
 605                         ; /* mark this one VALID again */
 606                 else
 607                         return NULL;
 608         }
 609
 610         if (ce_modified(ce, &st, really))
 611                 return ERR_PTR(-EINVAL);
 612
 613         size = ce_size(ce);
 614         updated = xmalloc(size);
 615         memcpy(updated, ce, size);
 616         fill_stat_cache_info(updated, &st);
 617
 618         /* In this case, if really is not set, we should leave
 619          * CE_VALID bit alone.  Otherwise, paths marked with
 620          * --no-assume-unchanged (i.e. things to be edited) will
 621          * reacquire CE_VALID bit automatically, which is not
 622          * really what we want.
 623          */
 624         if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
 625                 updated->ce_flags &= ~htons(CE_VALID);
 626
 627         return updated;
 628 }
 629
 630 int refresh_cache(unsigned int flags)
 631 {
 632         int i;
 633         int has_errors = 0;
 634         int really = (flags & REFRESH_REALLY) != 0;
 635         int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
 636         int quiet = (flags & REFRESH_QUIET) != 0;
 637         int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
 638
 639         for (i = 0; i < active_nr; i++) {
 640                 struct cache_entry *ce, *new;
 641                 ce = active_cache[i];
 642                 if (ce_stage(ce)) {
 643                         while ((i < active_nr) &&
 644                                ! strcmp(active_cache[i]->name, ce->name))
 645                                 i++;
 646                         i--;
 647                         if (allow_unmerged)
 648                                 continue;
 649                         printf("%s: needs merge\n", ce->name);
 650                         has_errors = 1;
 651                         continue;
 652                 }
 653
 654                 new = refresh_entry(ce, really);
 655                 if (!new)
 656                         continue;
 657                 if (IS_ERR(new)) {
 658                         if (not_new && PTR_ERR(new) == -ENOENT)
 659                                 continue;
 660                         if (really && PTR_ERR(new) == -EINVAL) {
 661                                 /* If we are doing --really-refresh that
 662                                  * means the index is not valid anymore.
 663                                  */
 664                                 ce->ce_flags &= ~htons(CE_VALID);
 665                                 active_cache_changed = 1;
 666                         }
 667                         if (quiet)
 668                                 continue;
 669                         printf("%s: needs update\n", ce->name);
 670                         has_errors = 1;
 671                         continue;
 672                 }
 673                 active_cache_changed = 1;
 674                 /* You can NOT just free active_cache[i] here, since it
 675                  * might not be necessarily malloc()ed but can also come
 676                  * from mmap(). */
 677                 active_cache[i] = new;
 678         }
 679         return has_errors;
 680 }
 681
 682 static int verify_hdr(struct cache_header *hdr, unsigned long size)
 683 {
 684         SHA_CTX c;
 685         unsigned char sha1[20];
 686
 687         if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
 688                 return error("bad signature");
 689         if (hdr->hdr_version != htonl(2))
 690                 return error("bad index version");
 691         SHA1_Init(&c);
 692         SHA1_Update(&c, hdr, size - 20);
 693         SHA1_Final(sha1, &c);
 694         if (memcmp(sha1, (void *)hdr + size - 20, 20))
 695                 return error("bad index file sha1 signature");
 696         return 0;
 697 }
 698
 699 int read_cache(void)
 700 {
 701         int fd, i;
 702         struct stat st;
 703         unsigned long size, offset;
 704         void *map;
 705         struct cache_header *hdr;
 706
 707         errno = EBUSY;
 708         if (active_cache)
 709                 return active_nr;
 710
 711         errno = ENOENT;
 712         index_file_timestamp = 0;
 713         fd = open(get_index_file(), O_RDONLY);
 714         if (fd < 0) {
 715                 if (errno == ENOENT)
 716                         return 0;
 717                 die("index file open failed (%s)", strerror(errno));
 718         }
 719
 720         size = 0; // avoid gcc warning
 721         map = MAP_FAILED;
 722         if (!fstat(fd, &st)) {
 723                 size = st.st_size;
 724                 errno = EINVAL;
 725                 if (size >= sizeof(struct cache_header) + 20)
 726                         map = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
 727         }
 728         close(fd);
 729         if (map == MAP_FAILED)
 730                 die("index file mmap failed (%s)", strerror(errno));
 731
 732         hdr = map;
 733         if (verify_hdr(hdr, size) < 0)
 734                 goto unmap;
 735
 736         active_nr = ntohl(hdr->hdr_entries);
 737         active_alloc = alloc_nr(active_nr);
 738         active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
 739
 740         offset = sizeof(*hdr);
 741         for (i = 0; i < active_nr; i++) {
 742                 struct cache_entry *ce = map + offset;
 743                 offset = offset + ce_size(ce);
 744                 active_cache[i] = ce;
 745         }
 746         index_file_timestamp = st.st_mtime;
 747         return active_nr;
 748
 749 unmap:
 750         munmap(map, size);
 751         errno = EINVAL;
 752         die("index file corrupt");
 753 }
 754
 755 #define WRITE_BUFFER_SIZE 8192
 756 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
 757 static unsigned long write_buffer_len;
 758
 759 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
 760 {
 761         while (len) {
 762                 unsigned int buffered = write_buffer_len;
 763                 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
 764                 if (partial > len)
 765                         partial = len;
 766                 memcpy(write_buffer + buffered, data, partial);
 767                 buffered += partial;
 768                 if (buffered == WRITE_BUFFER_SIZE) {
 769                         SHA1_Update(context, write_buffer, WRITE_BUFFER_SIZE);
 770                         if (write(fd, write_buffer, WRITE_BUFFER_SIZE) != WRITE_BUFFER_SIZE)
 771                                 return -1;
 772                         buffered = 0;
 773                 }
 774                 write_buffer_len = buffered;
 775                 len -= partial;
 776                 data += partial;
 777         }
 778         return 0;
 779 }
 780
 781 static int ce_flush(SHA_CTX *context, int fd)
 782 {
 783         unsigned int left = write_buffer_len;
 784
 785         if (left) {
 786                 write_buffer_len = 0;
 787                 SHA1_Update(context, write_buffer, left);
 788         }
 789
 790         /* Flush first if not enough space for SHA1 signature */
 791         if (left + 20 > WRITE_BUFFER_SIZE) {
 792                 if (write(fd, write_buffer, left) != left)
 793                         return -1;
 794                 left = 0;
 795         }
 796
 797         /* Append the SHA1 signature at the end */
 798         SHA1_Final(write_buffer + left, context);
 799         left += 20;
 800         if (write(fd, write_buffer, left) != left)
 801                 return -1;
 802         return 0;
 803 }
 804
 805 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
 806 {
 807         /*
 808          * The only thing we care about in this function is to smudge the
 809          * falsely clean entry due to touch-update-touch race, so we leave
 810          * everything else as they are.  We are called for entries whose
 811          * ce_mtime match the index file mtime.
 812          */
 813         struct stat st;
 814
 815         if (lstat(ce->name, &st) < 0)
 816                 return;
 817         if (ce_match_stat_basic(ce, &st))
 818                 return;
 819         if (ce_modified_check_fs(ce, &st)) {
 820                 /* This is "racily clean"; smudge it.  Note that this
 821                  * is a tricky code.  At first glance, it may appear
 822                  * that it can break with this sequence:
 823                  *
 824                  * $ echo xyzzy >frotz
 825                  * $ git-update-index --add frotz
 826                  * $ : >frotz
 827                  * $ sleep 3
 828                  * $ echo filfre >nitfol
 829                  * $ git-update-index --add nitfol
 830                  *
 831                  * but it does not.  Whe the second update-index runs,
 832                  * it notices that the entry "frotz" has the same timestamp
 833                  * as index, and if we were to smudge it by resetting its
 834                  * size to zero here, then the object name recorded
 835                  * in index is the 6-byte file but the cached stat information
 836                  * becomes zero --- which would then match what we would
 837                  * obtain from the filesystem next time we stat("frotz").
 838                  *
 839                  * However, the second update-index, before calling
 840                  * this function, notices that the cached size is 6
 841                  * bytes and what is on the filesystem is an empty
 842                  * file, and never calls us, so the cached size information
 843                  * for "frotz" stays 6 which does not match the filesystem.
 844                  */
 845                 ce->ce_size = htonl(0);
 846         }
 847 }
 848
 849 int write_cache(int newfd, struct cache_entry **cache, int entries)
 850 {
 851         SHA_CTX c;
 852         struct cache_header hdr;
 853         int i, removed;
 854
 855         for (i = removed = 0; i < entries; i++)
 856                 if (!cache[i]->ce_mode)
 857                         removed++;
 858
 859         hdr.hdr_signature = htonl(CACHE_SIGNATURE);
 860         hdr.hdr_version = htonl(2);
 861         hdr.hdr_entries = htonl(entries - removed);
 862
 863         SHA1_Init(&c);
 864         if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
 865                 return -1;
 866
 867         for (i = 0; i < entries; i++) {
 868                 struct cache_entry *ce = cache[i];
 869                 if (!ce->ce_mode)
 870                         continue;
 871                 if (index_file_timestamp &&
 872                     index_file_timestamp <= ntohl(ce->ce_mtime.sec))
 873                         ce_smudge_racily_clean_entry(ce);
 874                 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
 875                         return -1;
 876         }
 877         return ce_flush(&c, newfd);
 878 }