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  * Do we have another file that has the beginning components being a
 336  * proper superset of the name we're trying to add?
 337  */
 338 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 339 {
 340         int retval = 0;
 341         int len = ce_namelen(ce);
 342         int stage = ce_stage(ce);
 343         const char *name = ce->name;
 344
 345         while (pos < active_nr) {
 346                 struct cache_entry *p = active_cache[pos++];
 347
 348                 if (len >= ce_namelen(p))
 349                         break;
 350                 if (memcmp(name, p->name, len))
 351                         break;
 352                 if (ce_stage(p) != stage)
 353                         continue;
 354                 if (p->name[len] != '/')
 355                         continue;
 356                 retval = -1;
 357                 if (!ok_to_replace)
 358                         break;
 359                 remove_cache_entry_at(--pos);
 360         }
 361         return retval;
 362 }
 363
 364 /*
 365  * Do we have another file with a pathname that is a proper
 366  * subset of the name we're trying to add?
 367  */
 368 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
 369 {
 370         int retval = 0;
 371         int stage = ce_stage(ce);
 372         const char *name = ce->name;
 373         const char *slash = name + ce_namelen(ce);
 374
 375         for (;;) {
 376                 int len;
 377
 378                 for (;;) {
 379                         if (*--slash == '/')
 380                                 break;
 381                         if (slash <= ce->name)
 382                                 return retval;
 383                 }
 384                 len = slash - name;
 385
 386                 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
 387                 if (pos >= 0) {
 388                         retval = -1;
 389                         if (ok_to_replace)
 390                                 break;
 391                         remove_cache_entry_at(pos);
 392                         continue;
 393                 }
 394
 395                 /*
 396                  * Trivial optimization: if we find an entry that
 397                  * already matches the sub-directory, then we know
 398                  * we're ok, and we can exit.
 399                  */
 400                 pos = -pos-1;
 401                 while (pos < active_nr) {
 402                         struct cache_entry *p = active_cache[pos];
 403                         if ((ce_namelen(p) <= len) ||
 404                             (p->name[len] != '/') ||
 405                             memcmp(p->name, name, len))
 406                                 break; /* not our subdirectory */
 407                         if (ce_stage(p) == stage)
 408                                 /* p is at the same stage as our entry, and
 409                                  * is a subdirectory of what we are looking
 410                                  * at, so we cannot have conflicts at our
 411                                  * level or anything shorter.
 412                                  */
 413                                 return retval;
 414                         pos++;
 415                 }
 416         }
 417         return retval;
 418 }
 419
 420 /* We may be in a situation where we already have path/file and path
 421  * is being added, or we already have path and path/file is being
 422  * added.  Either one would result in a nonsense tree that has path
 423  * twice when git-write-tree tries to write it out.  Prevent it.
 424  *
 425  * If ok-to-replace is specified, we remove the conflicting entries
 426  * from the cache so the caller should recompute the insert position.
 427  * When this happens, we return non-zero.
 428  */
 429 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
 430 {
 431         /*
 432          * We check if the path is a sub-path of a subsequent pathname
 433          * first, since removing those will not change the position
 434          * in the array
 435          */
 436         int retval = has_file_name(ce, pos, ok_to_replace);
 437         /*
 438          * Then check if the path might have a clashing sub-directory
 439          * before it.
 440          */
 441         return retval + has_dir_name(ce, pos, ok_to_replace);
 442 }
 443
 444 int add_cache_entry(struct cache_entry *ce, int option)
 445 {
 446         int pos;
 447         int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
 448         int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
 449         int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
 450
 451         pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 452
 453         /* existing match? Just replace it. */
 454         if (pos >= 0) {
 455                 active_cache_changed = 1;
 456                 active_cache[pos] = ce;
 457                 return 0;
 458         }
 459         pos = -pos-1;
 460
 461         /*
 462          * Inserting a merged entry ("stage 0") into the index
 463          * will always replace all non-merged entries..
 464          */
 465         if (pos < active_nr && ce_stage(ce) == 0) {
 466                 while (ce_same_name(active_cache[pos], ce)) {
 467                         ok_to_add = 1;
 468                         if (!remove_cache_entry_at(pos))
 469                                 break;
 470                 }
 471         }
 472
 473         if (!ok_to_add)
 474                 return -1;
 475
 476         if (!skip_df_check &&
 477             check_file_directory_conflict(ce, pos, ok_to_replace)) {
 478                 if (!ok_to_replace)
 479                         return -1;
 480                 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
 481                 pos = -pos-1;
 482         }
 483
 484         /* Make sure the array is big enough .. */
 485         if (active_nr == active_alloc) {
 486                 active_alloc = alloc_nr(active_alloc);
 487                 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
 488         }
 489
 490         /* Add it in.. */
 491         active_nr++;
 492         if (active_nr > pos)
 493                 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
 494         active_cache[pos] = ce;
 495         active_cache_changed = 1;
 496         return 0;
 497 }
 498
 499 static int verify_hdr(struct cache_header *hdr, unsigned long size)
 500 {
 501         SHA_CTX c;
 502         unsigned char sha1[20];
 503
 504         if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
 505                 return error("bad signature");
 506         if (hdr->hdr_version != htonl(2))
 507                 return error("bad index version");
 508         SHA1_Init(&c);
 509         SHA1_Update(&c, hdr, size - 20);
 510         SHA1_Final(sha1, &c);
 511         if (memcmp(sha1, (void *)hdr + size - 20, 20))
 512                 return error("bad index file sha1 signature");
 513         return 0;
 514 }
 515
 516 int read_cache(void)
 517 {
 518         int fd, i;
 519         struct stat st;
 520         unsigned long size, offset;
 521         void *map;
 522         struct cache_header *hdr;
 523
 524         errno = EBUSY;
 525         if (active_cache)
 526                 return active_nr;
 527
 528         errno = ENOENT;
 529         index_file_timestamp = 0;
 530         fd = open(get_index_file(), O_RDONLY);
 531         if (fd < 0) {
 532                 if (errno == ENOENT)
 533                         return 0;
 534                 die("index file open failed (%s)", strerror(errno));
 535         }
 536
 537         size = 0; // avoid gcc warning
 538         map = MAP_FAILED;
 539         if (!fstat(fd, &st)) {
 540                 size = st.st_size;
 541                 errno = EINVAL;
 542                 if (size >= sizeof(struct cache_header) + 20)
 543                         map = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
 544         }
 545         close(fd);
 546         if (map == MAP_FAILED)
 547                 die("index file mmap failed (%s)", strerror(errno));
 548
 549         hdr = map;
 550         if (verify_hdr(hdr, size) < 0)
 551                 goto unmap;
 552
 553         active_nr = ntohl(hdr->hdr_entries);
 554         active_alloc = alloc_nr(active_nr);
 555         active_cache = calloc(active_alloc, sizeof(struct cache_entry *));
 556
 557         offset = sizeof(*hdr);
 558         for (i = 0; i < active_nr; i++) {
 559                 struct cache_entry *ce = map + offset;
 560                 offset = offset + ce_size(ce);
 561                 active_cache[i] = ce;
 562         }
 563         index_file_timestamp = st.st_mtime;
 564         return active_nr;
 565
 566 unmap:
 567         munmap(map, size);
 568         errno = EINVAL;
 569         die("index file corrupt");
 570 }
 571
 572 #define WRITE_BUFFER_SIZE 8192
 573 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
 574 static unsigned long write_buffer_len;
 575
 576 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
 577 {
 578         while (len) {
 579                 unsigned int buffered = write_buffer_len;
 580                 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
 581                 if (partial > len)
 582                         partial = len;
 583                 memcpy(write_buffer + buffered, data, partial);
 584                 buffered += partial;
 585                 if (buffered == WRITE_BUFFER_SIZE) {
 586                         SHA1_Update(context, write_buffer, WRITE_BUFFER_SIZE);
 587                         if (write(fd, write_buffer, WRITE_BUFFER_SIZE) != WRITE_BUFFER_SIZE)
 588                                 return -1;
 589                         buffered = 0;
 590                 }
 591                 write_buffer_len = buffered;
 592                 len -= partial;
 593                 data += partial;
 594         }
 595         return 0;
 596 }
 597
 598 static int ce_flush(SHA_CTX *context, int fd)
 599 {
 600         unsigned int left = write_buffer_len;
 601
 602         if (left) {
 603                 write_buffer_len = 0;
 604                 SHA1_Update(context, write_buffer, left);
 605         }
 606
 607         /* Flush first if not enough space for SHA1 signature */
 608         if (left + 20 > WRITE_BUFFER_SIZE) {
 609                 if (write(fd, write_buffer, left) != left)
 610                         return -1;
 611                 left = 0;
 612         }
 613
 614         /* Append the SHA1 signature at the end */
 615         SHA1_Final(write_buffer + left, context);
 616         left += 20;
 617         if (write(fd, write_buffer, left) != left)
 618                 return -1;
 619         return 0;
 620 }
 621
 622 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
 623 {
 624         /*
 625          * The only thing we care about in this function is to smudge the
 626          * falsely clean entry due to touch-update-touch race, so we leave
 627          * everything else as they are.  We are called for entries whose
 628          * ce_mtime match the index file mtime.
 629          */
 630         struct stat st;
 631
 632         if (lstat(ce->name, &st) < 0)
 633                 return;
 634         if (ce_match_stat_basic(ce, &st))
 635                 return;
 636         if (ce_modified_check_fs(ce, &st)) {
 637                 /* This is "racily clean"; smudge it.  Note that this
 638                  * is a tricky code.  At first glance, it may appear
 639                  * that it can break with this sequence:
 640                  *
 641                  * $ echo xyzzy >frotz
 642                  * $ git-update-index --add frotz
 643                  * $ : >frotz
 644                  * $ sleep 3
 645                  * $ echo filfre >nitfol
 646                  * $ git-update-index --add nitfol
 647                  *
 648                  * but it does not.  Whe the second update-index runs,
 649                  * it notices that the entry "frotz" has the same timestamp
 650                  * as index, and if we were to smudge it by resetting its
 651                  * size to zero here, then the object name recorded
 652                  * in index is the 6-byte file but the cached stat information
 653                  * becomes zero --- which would then match what we would
 654                  * obtain from the filesystem next time we stat("frotz").
 655                  *
 656                  * However, the second update-index, before calling
 657                  * this function, notices that the cached size is 6
 658                  * bytes and what is on the filesystem is an empty
 659                  * file, and never calls us, so the cached size information
 660                  * for "frotz" stays 6 which does not match the filesystem.
 661                  */
 662                 ce->ce_size = htonl(0);
 663         }
 664 }
 665
 666 int write_cache(int newfd, struct cache_entry **cache, int entries)
 667 {
 668         SHA_CTX c;
 669         struct cache_header hdr;
 670         int i, removed;
 671
 672         for (i = removed = 0; i < entries; i++)
 673                 if (!cache[i]->ce_mode)
 674                         removed++;
 675
 676         hdr.hdr_signature = htonl(CACHE_SIGNATURE);
 677         hdr.hdr_version = htonl(2);
 678         hdr.hdr_entries = htonl(entries - removed);
 679
 680         SHA1_Init(&c);
 681         if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
 682                 return -1;
 683
 684         for (i = 0; i < entries; i++) {
 685                 struct cache_entry *ce = cache[i];
 686                 if (!ce->ce_mode)
 687                         continue;
 688                 if (index_file_timestamp &&
 689                     index_file_timestamp <= ntohl(ce->ce_mtime.sec))
 690                         ce_smudge_racily_clean_entry(ce);
 691                 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
 692                         return -1;
 693         }
 694         return ce_flush(&c, newfd);
 695 }