2 * GIT - The information manager from hell
4 * Copyright (C) Linus Torvalds, 2005
7 #include "cache-tree.h"
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.
18 #define CACHE_EXT(s) ( (s[0]<<24)|(s[1]<<16)|(s[2]<<8)|(s[3]) )
19 #define CACHE_EXT_TREE 0x54524545 /* "TREE" */
21 struct cache_entry **active_cache;
22 static time_t index_file_timestamp;
23 unsigned int active_nr, active_alloc, active_cache_changed;
25 struct cache_tree *active_cache_tree;
29 static void *cache_mmap;
30 static size_t cache_mmap_size;
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.
37 void fill_stat_cache_info(struct cache_entry *ce, struct stat *st)
39 ce->ce_ctime.sec = htonl(st->st_ctime);
40 ce->ce_mtime.sec = htonl(st->st_mtime);
42 ce->ce_ctime.nsec = htonl(st->st_ctim.tv_nsec);
43 ce->ce_mtime.nsec = htonl(st->st_mtim.tv_nsec);
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);
52 ce->ce_flags |= htons(CE_VALID);
55 static int ce_compare_data(struct cache_entry *ce, struct stat *st)
58 int fd = open(ce->name, O_RDONLY);
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 */
69 static int ce_compare_link(struct cache_entry *ce, unsigned long expected_size)
78 target = xmalloc(expected_size);
79 len = readlink(ce->name, target, expected_size);
80 if (len != expected_size) {
84 buffer = read_sha1_file(ce->sha1, type, &size);
89 if (size == expected_size)
90 match = memcmp(buffer, target, size);
96 static int ce_modified_check_fs(struct cache_entry *ce, struct stat *st)
98 switch (st->st_mode & S_IFMT) {
100 if (ce_compare_data(ce, st))
104 if (ce_compare_link(ce, st->st_size))
113 static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
115 unsigned int changed = 0;
117 switch (ntohl(ce->ce_mode) & S_IFMT) {
119 changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
120 /* We consider only the owner x bit to be relevant for
123 if (trust_executable_bit &&
124 (0100 & (ntohl(ce->ce_mode) ^ st->st_mode)))
125 changed |= MODE_CHANGED;
128 changed |= !S_ISLNK(st->st_mode) ? TYPE_CHANGED : 0;
131 die("internal error: ce_mode is %o", ntohl(ce->ce_mode));
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;
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.
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;
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;
158 * st_dev breaks on network filesystems where different
159 * clients will have different views of what "device"
160 * the filesystem is on
162 if (ce->ce_dev != htonl(st->st_dev))
163 changed |= INODE_CHANGED;
166 if (ce->ce_size != htonl(st->st_size))
167 changed |= DATA_CHANGED;
172 int ce_match_stat(struct cache_entry *ce, struct stat *st, int options)
174 unsigned int changed;
175 int ignore_valid = options & 01;
176 int assume_racy_is_modified = options & 02;
179 * If it's marked as always valid in the index, it's
180 * valid whatever the checked-out copy says.
182 if (!ignore_valid && (ce->ce_flags & htons(CE_VALID)))
185 changed = ce_match_stat_basic(ce, st);
188 * Within 1 second of this sequence:
189 * echo xyzzy >file && git-update-index --add file
190 * running this command:
192 * would give a falsely clean cache entry. The mtime and
193 * length match the cache, and other stat fields do not change.
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.
204 index_file_timestamp &&
205 index_file_timestamp <= ntohl(ce->ce_mtime.sec)) {
206 if (assume_racy_is_modified)
207 changed |= DATA_CHANGED;
209 changed |= ce_modified_check_fs(ce, st);
215 int ce_modified(struct cache_entry *ce, struct stat *st, int really)
217 int changed, changed_fs;
218 changed = ce_match_stat(ce, st, really);
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
225 if (changed & (MODE_CHANGED | TYPE_CHANGED))
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.
232 if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
235 changed_fs = ce_modified_check_fs(ce, st);
237 return changed | changed_fs;
241 int base_name_compare(const char *name1, int len1, int mode1,
242 const char *name2, int len2, int mode2)
244 unsigned char c1, c2;
245 int len = len1 < len2 ? len1 : len2;
248 cmp = memcmp(name1, name2, len);
253 if (!c1 && S_ISDIR(mode1))
255 if (!c2 && S_ISDIR(mode2))
257 return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
260 int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
262 int len1 = flags1 & CE_NAMEMASK;
263 int len2 = flags2 & CE_NAMEMASK;
264 int len = len1 < len2 ? len1 : len2;
267 cmp = memcmp(name1, name2, len);
276 flags1 &= CE_STAGEMASK;
277 flags2 &= CE_STAGEMASK;
286 int cache_name_pos(const char *name, int namelen)
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));
307 /* Remove entry, return true if there are more entries to go.. */
308 int remove_cache_entry_at(int pos)
310 active_cache_changed = 1;
312 if (pos >= active_nr)
314 memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
318 int remove_file_from_cache(const char *path)
320 int pos = cache_name_pos(path, strlen(path));
323 while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
324 remove_cache_entry_at(pos);
328 int add_file_to_index(const char *path, int verbose)
332 struct cache_entry *ce;
334 if (lstat(path, &st))
335 die("%s: unable to stat (%s)", path, strerror(errno));
337 if (!S_ISREG(st.st_mode) && !S_ISLNK(st.st_mode))
338 die("%s: can only add regular files or symbolic links", path);
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);
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, otherwise assume unexecutable.
352 int pos = cache_name_pos(path, namelen);
354 ce->ce_mode = active_cache[pos]->ce_mode;
355 else if (S_ISREG(st.st_mode))
356 ce->ce_mode = create_ce_mode(S_IFREG | 0666);
359 if (index_path(ce->sha1, path, &st, 1))
360 die("unable to index file %s", path);
361 if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE))
362 die("unable to add %s to index",path);
364 printf("add '%s'\n", path);
365 cache_tree_invalidate_path(active_cache_tree, path);
369 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
371 int len = ce_namelen(a);
372 return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
375 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
377 const char *match, *name;
383 len = ce_namelen(ce);
385 while ((match = *pathspec++) != NULL) {
386 int matchlen = strlen(match);
389 if (memcmp(name, match, matchlen))
391 if (matchlen && name[matchlen-1] == '/')
393 if (name[matchlen] == '/' || !name[matchlen])
402 * We fundamentally don't like some paths: we don't want
403 * dot or dot-dot anywhere, and for obvious reasons don't
404 * want to recurse into ".git" either.
406 * Also, we don't want double slashes or slashes at the
407 * end that can make pathnames ambiguous.
409 static int verify_dotfile(const char *rest)
412 * The first character was '.', but that
413 * has already been discarded, we now test
417 /* "." is not allowed */
422 * ".git" followed by NUL or slash is bad. This
423 * shares the path end test with the ".." case.
433 if (rest[1] == '\0' || rest[1] == '/')
439 int verify_path(const char *path)
456 if (verify_dotfile(path))
466 * Do we have another file that has the beginning components being a
467 * proper superset of the name we're trying to add?
469 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
472 int len = ce_namelen(ce);
473 int stage = ce_stage(ce);
474 const char *name = ce->name;
476 while (pos < active_nr) {
477 struct cache_entry *p = active_cache[pos++];
479 if (len >= ce_namelen(p))
481 if (memcmp(name, p->name, len))
483 if (ce_stage(p) != stage)
485 if (p->name[len] != '/')
490 remove_cache_entry_at(--pos);
496 * Do we have another file with a pathname that is a proper
497 * subset of the name we're trying to add?
499 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
502 int stage = ce_stage(ce);
503 const char *name = ce->name;
504 const char *slash = name + ce_namelen(ce);
512 if (slash <= ce->name)
517 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
522 remove_cache_entry_at(pos);
527 * Trivial optimization: if we find an entry that
528 * already matches the sub-directory, then we know
529 * we're ok, and we can exit.
532 while (pos < active_nr) {
533 struct cache_entry *p = active_cache[pos];
534 if ((ce_namelen(p) <= len) ||
535 (p->name[len] != '/') ||
536 memcmp(p->name, name, len))
537 break; /* not our subdirectory */
538 if (ce_stage(p) == stage)
539 /* p is at the same stage as our entry, and
540 * is a subdirectory of what we are looking
541 * at, so we cannot have conflicts at our
542 * level or anything shorter.
551 /* We may be in a situation where we already have path/file and path
552 * is being added, or we already have path and path/file is being
553 * added. Either one would result in a nonsense tree that has path
554 * twice when git-write-tree tries to write it out. Prevent it.
556 * If ok-to-replace is specified, we remove the conflicting entries
557 * from the cache so the caller should recompute the insert position.
558 * When this happens, we return non-zero.
560 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
563 * We check if the path is a sub-path of a subsequent pathname
564 * first, since removing those will not change the position
567 int retval = has_file_name(ce, pos, ok_to_replace);
569 * Then check if the path might have a clashing sub-directory
572 return retval + has_dir_name(ce, pos, ok_to_replace);
575 int add_cache_entry(struct cache_entry *ce, int option)
578 int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
579 int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
580 int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
582 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
584 /* existing match? Just replace it. */
586 active_cache_changed = 1;
587 active_cache[pos] = ce;
593 * Inserting a merged entry ("stage 0") into the index
594 * will always replace all non-merged entries..
596 if (pos < active_nr && ce_stage(ce) == 0) {
597 while (ce_same_name(active_cache[pos], ce)) {
599 if (!remove_cache_entry_at(pos))
606 if (!verify_path(ce->name))
609 if (!skip_df_check &&
610 check_file_directory_conflict(ce, pos, ok_to_replace)) {
612 return error("'%s' appears as both a file and as a directory", ce->name);
613 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
617 /* Make sure the array is big enough .. */
618 if (active_nr == active_alloc) {
619 active_alloc = alloc_nr(active_alloc);
620 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
626 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
627 active_cache[pos] = ce;
628 active_cache_changed = 1;
633 * "refresh" does not calculate a new sha1 file or bring the
634 * cache up-to-date for mode/content changes. But what it
635 * _does_ do is to "re-match" the stat information of a file
636 * with the cache, so that you can refresh the cache for a
637 * file that hasn't been changed but where the stat entry is
640 * For example, you'd want to do this after doing a "git-read-tree",
641 * to link up the stat cache details with the proper files.
643 struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
646 struct cache_entry *updated;
649 if (lstat(ce->name, &st) < 0) {
654 changed = ce_match_stat(ce, &st, really);
656 if (really && assume_unchanged &&
657 !(ce->ce_flags & htons(CE_VALID)))
658 ; /* mark this one VALID again */
663 if (ce_modified(ce, &st, really)) {
664 cache_errno = EINVAL;
669 updated = xmalloc(size);
670 memcpy(updated, ce, size);
671 fill_stat_cache_info(updated, &st);
673 /* In this case, if really is not set, we should leave
674 * CE_VALID bit alone. Otherwise, paths marked with
675 * --no-assume-unchanged (i.e. things to be edited) will
676 * reacquire CE_VALID bit automatically, which is not
677 * really what we want.
679 if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
680 updated->ce_flags &= ~htons(CE_VALID);
685 int refresh_cache(unsigned int flags)
689 int really = (flags & REFRESH_REALLY) != 0;
690 int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
691 int quiet = (flags & REFRESH_QUIET) != 0;
692 int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
694 for (i = 0; i < active_nr; i++) {
695 struct cache_entry *ce, *new;
696 ce = active_cache[i];
698 while ((i < active_nr) &&
699 ! strcmp(active_cache[i]->name, ce->name))
704 printf("%s: needs merge\n", ce->name);
709 new = refresh_cache_entry(ce, really);
713 if (not_new && cache_errno == ENOENT)
715 if (really && cache_errno == EINVAL) {
716 /* If we are doing --really-refresh that
717 * means the index is not valid anymore.
719 ce->ce_flags &= ~htons(CE_VALID);
720 active_cache_changed = 1;
724 printf("%s: needs update\n", ce->name);
728 active_cache_changed = 1;
729 /* You can NOT just free active_cache[i] here, since it
730 * might not be necessarily malloc()ed but can also come
732 active_cache[i] = new;
737 static int verify_hdr(struct cache_header *hdr, unsigned long size)
740 unsigned char sha1[20];
742 if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
743 return error("bad signature");
744 if (hdr->hdr_version != htonl(2))
745 return error("bad index version");
747 SHA1_Update(&c, hdr, size - 20);
748 SHA1_Final(sha1, &c);
749 if (hashcmp(sha1, (unsigned char *)hdr + size - 20))
750 return error("bad index file sha1 signature");
754 static int read_index_extension(const char *ext, void *data, unsigned long sz)
756 switch (CACHE_EXT(ext)) {
758 active_cache_tree = cache_tree_read(data, sz);
761 if (*ext < 'A' || 'Z' < *ext)
762 return error("index uses %.4s extension, which we do not understand",
764 fprintf(stderr, "ignoring %.4s extension\n", ext);
772 return read_cache_from(get_index_file());
775 /* remember to discard_cache() before reading a different cache! */
776 int read_cache_from(const char *path)
780 unsigned long offset;
781 struct cache_header *hdr;
788 index_file_timestamp = 0;
789 fd = open(path, O_RDONLY);
793 die("index file open failed (%s)", strerror(errno));
796 cache_mmap = MAP_FAILED;
797 if (!fstat(fd, &st)) {
798 cache_mmap_size = st.st_size;
800 if (cache_mmap_size >= sizeof(struct cache_header) + 20)
801 cache_mmap = mmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
804 if (cache_mmap == MAP_FAILED)
805 die("index file mmap failed (%s)", strerror(errno));
808 if (verify_hdr(hdr, cache_mmap_size) < 0)
811 active_nr = ntohl(hdr->hdr_entries);
812 active_alloc = alloc_nr(active_nr);
813 active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
815 offset = sizeof(*hdr);
816 for (i = 0; i < active_nr; i++) {
817 struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
818 offset = offset + ce_size(ce);
819 active_cache[i] = ce;
821 index_file_timestamp = st.st_mtime;
822 while (offset <= cache_mmap_size - 20 - 8) {
823 /* After an array of active_nr index entries,
824 * there can be arbitrary number of extended
825 * sections, each of which is prefixed with
826 * extension name (4-byte) and section length
827 * in 4-byte network byte order.
829 unsigned long extsize;
830 memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
831 extsize = ntohl(extsize);
832 if (read_index_extension(((const char *) cache_mmap) + offset,
833 (char *) cache_mmap + offset + 8,
842 munmap(cache_mmap, cache_mmap_size);
844 die("index file corrupt");
847 int discard_cache(void)
851 active_nr = active_cache_changed = 0;
852 index_file_timestamp = 0;
853 cache_tree_free(&active_cache_tree);
854 if (cache_mmap == NULL)
856 ret = munmap(cache_mmap, cache_mmap_size);
860 /* no need to throw away allocated active_cache */
864 #define WRITE_BUFFER_SIZE 8192
865 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
866 static unsigned long write_buffer_len;
868 static int ce_write_flush(SHA_CTX *context, int fd)
870 unsigned int buffered = write_buffer_len;
872 SHA1_Update(context, write_buffer, buffered);
873 if (write(fd, write_buffer, buffered) != buffered)
875 write_buffer_len = 0;
880 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
883 unsigned int buffered = write_buffer_len;
884 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
887 memcpy(write_buffer + buffered, data, partial);
889 if (buffered == WRITE_BUFFER_SIZE) {
890 write_buffer_len = buffered;
891 if (ce_write_flush(context, fd))
895 write_buffer_len = buffered;
897 data = (char *) data + partial;
902 static int write_index_ext_header(SHA_CTX *context, int fd,
903 unsigned int ext, unsigned int sz)
907 return ((ce_write(context, fd, &ext, 4) < 0) ||
908 (ce_write(context, fd, &sz, 4) < 0)) ? -1 : 0;
911 static int ce_flush(SHA_CTX *context, int fd)
913 unsigned int left = write_buffer_len;
916 write_buffer_len = 0;
917 SHA1_Update(context, write_buffer, left);
920 /* Flush first if not enough space for SHA1 signature */
921 if (left + 20 > WRITE_BUFFER_SIZE) {
922 if (write(fd, write_buffer, left) != left)
927 /* Append the SHA1 signature at the end */
928 SHA1_Final(write_buffer + left, context);
930 return (write(fd, write_buffer, left) != left) ? -1 : 0;
933 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
936 * The only thing we care about in this function is to smudge the
937 * falsely clean entry due to touch-update-touch race, so we leave
938 * everything else as they are. We are called for entries whose
939 * ce_mtime match the index file mtime.
943 if (lstat(ce->name, &st) < 0)
945 if (ce_match_stat_basic(ce, &st))
947 if (ce_modified_check_fs(ce, &st)) {
948 /* This is "racily clean"; smudge it. Note that this
949 * is a tricky code. At first glance, it may appear
950 * that it can break with this sequence:
952 * $ echo xyzzy >frotz
953 * $ git-update-index --add frotz
956 * $ echo filfre >nitfol
957 * $ git-update-index --add nitfol
959 * but it does not. When the second update-index runs,
960 * it notices that the entry "frotz" has the same timestamp
961 * as index, and if we were to smudge it by resetting its
962 * size to zero here, then the object name recorded
963 * in index is the 6-byte file but the cached stat information
964 * becomes zero --- which would then match what we would
965 * obtain from the filesystem next time we stat("frotz").
967 * However, the second update-index, before calling
968 * this function, notices that the cached size is 6
969 * bytes and what is on the filesystem is an empty
970 * file, and never calls us, so the cached size information
971 * for "frotz" stays 6 which does not match the filesystem.
973 ce->ce_size = htonl(0);
977 int write_cache(int newfd, struct cache_entry **cache, int entries)
980 struct cache_header hdr;
983 for (i = removed = 0; i < entries; i++)
984 if (!cache[i]->ce_mode)
987 hdr.hdr_signature = htonl(CACHE_SIGNATURE);
988 hdr.hdr_version = htonl(2);
989 hdr.hdr_entries = htonl(entries - removed);
992 if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
995 for (i = 0; i < entries; i++) {
996 struct cache_entry *ce = cache[i];
999 if (index_file_timestamp &&
1000 index_file_timestamp <= ntohl(ce->ce_mtime.sec))
1001 ce_smudge_racily_clean_entry(ce);
1002 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
1006 /* Write extension data here */
1007 if (active_cache_tree) {
1009 void *data = cache_tree_write(active_cache_tree, &sz);
1011 !write_index_ext_header(&c, newfd, CACHE_EXT_TREE, sz) &&
1012 !ce_write(&c, newfd, data, sz))
1019 return ce_flush(&c, newfd);