convert object type handling from a string to a number
[git/git.git] / 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 enum object_type type;
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 if (trust_executable_bit)
348 ce->ce_mode = create_ce_mode(st.st_mode);
349 else {
350 /* If there is an existing entry, pick the mode bits
351 * from it, otherwise assume unexecutable.
352 */
353 struct cache_entry *ent;
354 int pos = cache_name_pos(path, namelen);
355
356 ent = (0 <= pos) ? active_cache[pos] : NULL;
357 ce->ce_mode = ce_mode_from_stat(ent, st.st_mode);
358 }
359
360 if (index_path(ce->sha1, path, &st, 1))
361 die("unable to index file %s", path);
362 if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE))
363 die("unable to add %s to index",path);
364 if (verbose)
365 printf("add '%s'\n", path);
366 cache_tree_invalidate_path(active_cache_tree, path);
367 return 0;
368 }
369
370 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
371 {
372 int len = ce_namelen(a);
373 return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
374 }
375
376 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
377 {
378 const char *match, *name;
379 int len;
380
381 if (!pathspec)
382 return 1;
383
384 len = ce_namelen(ce);
385 name = ce->name;
386 while ((match = *pathspec++) != NULL) {
387 int matchlen = strlen(match);
388 if (matchlen > len)
389 continue;
390 if (memcmp(name, match, matchlen))
391 continue;
392 if (matchlen && name[matchlen-1] == '/')
393 return 1;
394 if (name[matchlen] == '/' || !name[matchlen])
395 return 1;
396 if (!matchlen)
397 return 1;
398 }
399 return 0;
400 }
401
402 /*
403 * We fundamentally don't like some paths: we don't want
404 * dot or dot-dot anywhere, and for obvious reasons don't
405 * want to recurse into ".git" either.
406 *
407 * Also, we don't want double slashes or slashes at the
408 * end that can make pathnames ambiguous.
409 */
410 static int verify_dotfile(const char *rest)
411 {
412 /*
413 * The first character was '.', but that
414 * has already been discarded, we now test
415 * the rest.
416 */
417 switch (*rest) {
418 /* "." is not allowed */
419 case '\0': case '/':
420 return 0;
421
422 /*
423 * ".git" followed by NUL or slash is bad. This
424 * shares the path end test with the ".." case.
425 */
426 case 'g':
427 if (rest[1] != 'i')
428 break;
429 if (rest[2] != 't')
430 break;
431 rest += 2;
432 /* fallthrough */
433 case '.':
434 if (rest[1] == '\0' || rest[1] == '/')
435 return 0;
436 }
437 return 1;
438 }
439
440 int verify_path(const char *path)
441 {
442 char c;
443
444 goto inside;
445 for (;;) {
446 if (!c)
447 return 1;
448 if (c == '/') {
449 inside:
450 c = *path++;
451 switch (c) {
452 default:
453 continue;
454 case '/': case '\0':
455 break;
456 case '.':
457 if (verify_dotfile(path))
458 continue;
459 }
460 return 0;
461 }
462 c = *path++;
463 }
464 }
465
466 /*
467 * Do we have another file that has the beginning components being a
468 * proper superset of the name we're trying to add?
469 */
470 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
471 {
472 int retval = 0;
473 int len = ce_namelen(ce);
474 int stage = ce_stage(ce);
475 const char *name = ce->name;
476
477 while (pos < active_nr) {
478 struct cache_entry *p = active_cache[pos++];
479
480 if (len >= ce_namelen(p))
481 break;
482 if (memcmp(name, p->name, len))
483 break;
484 if (ce_stage(p) != stage)
485 continue;
486 if (p->name[len] != '/')
487 continue;
488 retval = -1;
489 if (!ok_to_replace)
490 break;
491 remove_cache_entry_at(--pos);
492 }
493 return retval;
494 }
495
496 /*
497 * Do we have another file with a pathname that is a proper
498 * subset of the name we're trying to add?
499 */
500 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
501 {
502 int retval = 0;
503 int stage = ce_stage(ce);
504 const char *name = ce->name;
505 const char *slash = name + ce_namelen(ce);
506
507 for (;;) {
508 int len;
509
510 for (;;) {
511 if (*--slash == '/')
512 break;
513 if (slash <= ce->name)
514 return retval;
515 }
516 len = slash - name;
517
518 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
519 if (pos >= 0) {
520 retval = -1;
521 if (!ok_to_replace)
522 break;
523 remove_cache_entry_at(pos);
524 continue;
525 }
526
527 /*
528 * Trivial optimization: if we find an entry that
529 * already matches the sub-directory, then we know
530 * we're ok, and we can exit.
531 */
532 pos = -pos-1;
533 while (pos < active_nr) {
534 struct cache_entry *p = active_cache[pos];
535 if ((ce_namelen(p) <= len) ||
536 (p->name[len] != '/') ||
537 memcmp(p->name, name, len))
538 break; /* not our subdirectory */
539 if (ce_stage(p) == stage)
540 /* p is at the same stage as our entry, and
541 * is a subdirectory of what we are looking
542 * at, so we cannot have conflicts at our
543 * level or anything shorter.
544 */
545 return retval;
546 pos++;
547 }
548 }
549 return retval;
550 }
551
552 /* We may be in a situation where we already have path/file and path
553 * is being added, or we already have path and path/file is being
554 * added. Either one would result in a nonsense tree that has path
555 * twice when git-write-tree tries to write it out. Prevent it.
556 *
557 * If ok-to-replace is specified, we remove the conflicting entries
558 * from the cache so the caller should recompute the insert position.
559 * When this happens, we return non-zero.
560 */
561 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
562 {
563 /*
564 * We check if the path is a sub-path of a subsequent pathname
565 * first, since removing those will not change the position
566 * in the array
567 */
568 int retval = has_file_name(ce, pos, ok_to_replace);
569 /*
570 * Then check if the path might have a clashing sub-directory
571 * before it.
572 */
573 return retval + has_dir_name(ce, pos, ok_to_replace);
574 }
575
576 int add_cache_entry(struct cache_entry *ce, int option)
577 {
578 int pos;
579 int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
580 int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
581 int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
582
583 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
584
585 /* existing match? Just replace it. */
586 if (pos >= 0) {
587 active_cache_changed = 1;
588 active_cache[pos] = ce;
589 return 0;
590 }
591 pos = -pos-1;
592
593 /*
594 * Inserting a merged entry ("stage 0") into the index
595 * will always replace all non-merged entries..
596 */
597 if (pos < active_nr && ce_stage(ce) == 0) {
598 while (ce_same_name(active_cache[pos], ce)) {
599 ok_to_add = 1;
600 if (!remove_cache_entry_at(pos))
601 break;
602 }
603 }
604
605 if (!ok_to_add)
606 return -1;
607 if (!verify_path(ce->name))
608 return -1;
609
610 if (!skip_df_check &&
611 check_file_directory_conflict(ce, pos, ok_to_replace)) {
612 if (!ok_to_replace)
613 return error("'%s' appears as both a file and as a directory", ce->name);
614 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
615 pos = -pos-1;
616 }
617
618 /* Make sure the array is big enough .. */
619 if (active_nr == active_alloc) {
620 active_alloc = alloc_nr(active_alloc);
621 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
622 }
623
624 /* Add it in.. */
625 active_nr++;
626 if (active_nr > pos)
627 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
628 active_cache[pos] = ce;
629 active_cache_changed = 1;
630 return 0;
631 }
632
633 /*
634 * "refresh" does not calculate a new sha1 file or bring the
635 * cache up-to-date for mode/content changes. But what it
636 * _does_ do is to "re-match" the stat information of a file
637 * with the cache, so that you can refresh the cache for a
638 * file that hasn't been changed but where the stat entry is
639 * out of date.
640 *
641 * For example, you'd want to do this after doing a "git-read-tree",
642 * to link up the stat cache details with the proper files.
643 */
644 struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
645 {
646 struct stat st;
647 struct cache_entry *updated;
648 int changed, size;
649
650 if (lstat(ce->name, &st) < 0) {
651 cache_errno = errno;
652 return NULL;
653 }
654
655 changed = ce_match_stat(ce, &st, really);
656 if (!changed) {
657 if (really && assume_unchanged &&
658 !(ce->ce_flags & htons(CE_VALID)))
659 ; /* mark this one VALID again */
660 else
661 return ce;
662 }
663
664 if (ce_modified(ce, &st, really)) {
665 cache_errno = EINVAL;
666 return NULL;
667 }
668
669 size = ce_size(ce);
670 updated = xmalloc(size);
671 memcpy(updated, ce, size);
672 fill_stat_cache_info(updated, &st);
673
674 /* In this case, if really is not set, we should leave
675 * CE_VALID bit alone. Otherwise, paths marked with
676 * --no-assume-unchanged (i.e. things to be edited) will
677 * reacquire CE_VALID bit automatically, which is not
678 * really what we want.
679 */
680 if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
681 updated->ce_flags &= ~htons(CE_VALID);
682
683 return updated;
684 }
685
686 int refresh_cache(unsigned int flags)
687 {
688 int i;
689 int has_errors = 0;
690 int really = (flags & REFRESH_REALLY) != 0;
691 int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
692 int quiet = (flags & REFRESH_QUIET) != 0;
693 int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
694
695 for (i = 0; i < active_nr; i++) {
696 struct cache_entry *ce, *new;
697 ce = active_cache[i];
698 if (ce_stage(ce)) {
699 while ((i < active_nr) &&
700 ! strcmp(active_cache[i]->name, ce->name))
701 i++;
702 i--;
703 if (allow_unmerged)
704 continue;
705 printf("%s: needs merge\n", ce->name);
706 has_errors = 1;
707 continue;
708 }
709
710 new = refresh_cache_entry(ce, really);
711 if (new == ce)
712 continue;
713 if (!new) {
714 if (not_new && cache_errno == ENOENT)
715 continue;
716 if (really && cache_errno == EINVAL) {
717 /* If we are doing --really-refresh that
718 * means the index is not valid anymore.
719 */
720 ce->ce_flags &= ~htons(CE_VALID);
721 active_cache_changed = 1;
722 }
723 if (quiet)
724 continue;
725 printf("%s: needs update\n", ce->name);
726 has_errors = 1;
727 continue;
728 }
729 active_cache_changed = 1;
730 /* You can NOT just free active_cache[i] here, since it
731 * might not be necessarily malloc()ed but can also come
732 * from mmap(). */
733 active_cache[i] = new;
734 }
735 return has_errors;
736 }
737
738 static int verify_hdr(struct cache_header *hdr, unsigned long size)
739 {
740 SHA_CTX c;
741 unsigned char sha1[20];
742
743 if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
744 return error("bad signature");
745 if (hdr->hdr_version != htonl(2))
746 return error("bad index version");
747 SHA1_Init(&c);
748 SHA1_Update(&c, hdr, size - 20);
749 SHA1_Final(sha1, &c);
750 if (hashcmp(sha1, (unsigned char *)hdr + size - 20))
751 return error("bad index file sha1 signature");
752 return 0;
753 }
754
755 static int read_index_extension(const char *ext, void *data, unsigned long sz)
756 {
757 switch (CACHE_EXT(ext)) {
758 case CACHE_EXT_TREE:
759 active_cache_tree = cache_tree_read(data, sz);
760 break;
761 default:
762 if (*ext < 'A' || 'Z' < *ext)
763 return error("index uses %.4s extension, which we do not understand",
764 ext);
765 fprintf(stderr, "ignoring %.4s extension\n", ext);
766 break;
767 }
768 return 0;
769 }
770
771 int read_cache(void)
772 {
773 return read_cache_from(get_index_file());
774 }
775
776 /* remember to discard_cache() before reading a different cache! */
777 int read_cache_from(const char *path)
778 {
779 int fd, i;
780 struct stat st;
781 unsigned long offset;
782 struct cache_header *hdr;
783
784 errno = EBUSY;
785 if (cache_mmap)
786 return active_nr;
787
788 errno = ENOENT;
789 index_file_timestamp = 0;
790 fd = open(path, O_RDONLY);
791 if (fd < 0) {
792 if (errno == ENOENT)
793 return 0;
794 die("index file open failed (%s)", strerror(errno));
795 }
796
797 if (!fstat(fd, &st)) {
798 cache_mmap_size = st.st_size;
799 errno = EINVAL;
800 if (cache_mmap_size >= sizeof(struct cache_header) + 20)
801 cache_mmap = xmmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
802 else
803 die("index file smaller than expected");
804 } else
805 die("cannot stat the open index (%s)", strerror(errno));
806 close(fd);
807
808 hdr = cache_mmap;
809 if (verify_hdr(hdr, cache_mmap_size) < 0)
810 goto unmap;
811
812 active_nr = ntohl(hdr->hdr_entries);
813 active_alloc = alloc_nr(active_nr);
814 active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
815
816 offset = sizeof(*hdr);
817 for (i = 0; i < active_nr; i++) {
818 struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
819 offset = offset + ce_size(ce);
820 active_cache[i] = ce;
821 }
822 index_file_timestamp = st.st_mtime;
823 while (offset <= cache_mmap_size - 20 - 8) {
824 /* After an array of active_nr index entries,
825 * there can be arbitrary number of extended
826 * sections, each of which is prefixed with
827 * extension name (4-byte) and section length
828 * in 4-byte network byte order.
829 */
830 unsigned long extsize;
831 memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
832 extsize = ntohl(extsize);
833 if (read_index_extension(((const char *) cache_mmap) + offset,
834 (char *) cache_mmap + offset + 8,
835 extsize) < 0)
836 goto unmap;
837 offset += 8;
838 offset += extsize;
839 }
840 return active_nr;
841
842 unmap:
843 munmap(cache_mmap, cache_mmap_size);
844 errno = EINVAL;
845 die("index file corrupt");
846 }
847
848 int discard_cache(void)
849 {
850 int ret;
851
852 active_nr = active_cache_changed = 0;
853 index_file_timestamp = 0;
854 cache_tree_free(&active_cache_tree);
855 if (cache_mmap == NULL)
856 return 0;
857 ret = munmap(cache_mmap, cache_mmap_size);
858 cache_mmap = NULL;
859 cache_mmap_size = 0;
860
861 /* no need to throw away allocated active_cache */
862 return ret;
863 }
864
865 #define WRITE_BUFFER_SIZE 8192
866 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
867 static unsigned long write_buffer_len;
868
869 static int ce_write_flush(SHA_CTX *context, int fd)
870 {
871 unsigned int buffered = write_buffer_len;
872 if (buffered) {
873 SHA1_Update(context, write_buffer, buffered);
874 if (write_in_full(fd, write_buffer, buffered) != buffered)
875 return -1;
876 write_buffer_len = 0;
877 }
878 return 0;
879 }
880
881 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
882 {
883 while (len) {
884 unsigned int buffered = write_buffer_len;
885 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
886 if (partial > len)
887 partial = len;
888 memcpy(write_buffer + buffered, data, partial);
889 buffered += partial;
890 if (buffered == WRITE_BUFFER_SIZE) {
891 write_buffer_len = buffered;
892 if (ce_write_flush(context, fd))
893 return -1;
894 buffered = 0;
895 }
896 write_buffer_len = buffered;
897 len -= partial;
898 data = (char *) data + partial;
899 }
900 return 0;
901 }
902
903 static int write_index_ext_header(SHA_CTX *context, int fd,
904 unsigned int ext, unsigned int sz)
905 {
906 ext = htonl(ext);
907 sz = htonl(sz);
908 return ((ce_write(context, fd, &ext, 4) < 0) ||
909 (ce_write(context, fd, &sz, 4) < 0)) ? -1 : 0;
910 }
911
912 static int ce_flush(SHA_CTX *context, int fd)
913 {
914 unsigned int left = write_buffer_len;
915
916 if (left) {
917 write_buffer_len = 0;
918 SHA1_Update(context, write_buffer, left);
919 }
920
921 /* Flush first if not enough space for SHA1 signature */
922 if (left + 20 > WRITE_BUFFER_SIZE) {
923 if (write_in_full(fd, write_buffer, left) != left)
924 return -1;
925 left = 0;
926 }
927
928 /* Append the SHA1 signature at the end */
929 SHA1_Final(write_buffer + left, context);
930 left += 20;
931 return (write_in_full(fd, write_buffer, left) != left) ? -1 : 0;
932 }
933
934 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
935 {
936 /*
937 * The only thing we care about in this function is to smudge the
938 * falsely clean entry due to touch-update-touch race, so we leave
939 * everything else as they are. We are called for entries whose
940 * ce_mtime match the index file mtime.
941 */
942 struct stat st;
943
944 if (lstat(ce->name, &st) < 0)
945 return;
946 if (ce_match_stat_basic(ce, &st))
947 return;
948 if (ce_modified_check_fs(ce, &st)) {
949 /* This is "racily clean"; smudge it. Note that this
950 * is a tricky code. At first glance, it may appear
951 * that it can break with this sequence:
952 *
953 * $ echo xyzzy >frotz
954 * $ git-update-index --add frotz
955 * $ : >frotz
956 * $ sleep 3
957 * $ echo filfre >nitfol
958 * $ git-update-index --add nitfol
959 *
960 * but it does not. When the second update-index runs,
961 * it notices that the entry "frotz" has the same timestamp
962 * as index, and if we were to smudge it by resetting its
963 * size to zero here, then the object name recorded
964 * in index is the 6-byte file but the cached stat information
965 * becomes zero --- which would then match what we would
966 * obtain from the filesystem next time we stat("frotz").
967 *
968 * However, the second update-index, before calling
969 * this function, notices that the cached size is 6
970 * bytes and what is on the filesystem is an empty
971 * file, and never calls us, so the cached size information
972 * for "frotz" stays 6 which does not match the filesystem.
973 */
974 ce->ce_size = htonl(0);
975 }
976 }
977
978 int write_cache(int newfd, struct cache_entry **cache, int entries)
979 {
980 SHA_CTX c;
981 struct cache_header hdr;
982 int i, removed;
983
984 for (i = removed = 0; i < entries; i++)
985 if (!cache[i]->ce_mode)
986 removed++;
987
988 hdr.hdr_signature = htonl(CACHE_SIGNATURE);
989 hdr.hdr_version = htonl(2);
990 hdr.hdr_entries = htonl(entries - removed);
991
992 SHA1_Init(&c);
993 if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
994 return -1;
995
996 for (i = 0; i < entries; i++) {
997 struct cache_entry *ce = cache[i];
998 if (!ce->ce_mode)
999 continue;
1000 if (index_file_timestamp &&
1001 index_file_timestamp <= ntohl(ce->ce_mtime.sec))
1002 ce_smudge_racily_clean_entry(ce);
1003 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
1004 return -1;
1005 }
1006
1007 /* Write extension data here */
1008 if (active_cache_tree) {
1009 unsigned long sz;
1010 void *data = cache_tree_write(active_cache_tree, &sz);
1011 if (data &&
1012 !write_index_ext_header(&c, newfd, CACHE_EXT_TREE, sz) &&
1013 !ce_write(&c, newfd, data, sz))
1014 free(data);
1015 else {
1016 free(data);
1017 return -1;
1018 }
1019 }
1020 return ce_flush(&c, newfd);
1021 }