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
= NULL
;
22 static time_t index_file_timestamp
;
23 unsigned int active_nr
= 0, active_alloc
= 0, active_cache_changed
= 0;
25 struct cache_tree
*active_cache_tree
= NULL
;
29 static void *cache_mmap
= NULL
;
30 static size_t cache_mmap_size
= 0;
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
= memcmp(sha1
, ce
->sha1
, 20);
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 ignore_valid
)
174 unsigned int changed
;
177 * If it's marked as always valid in the index, it's
178 * valid whatever the checked-out copy says.
180 if (!ignore_valid
&& (ce
->ce_flags
& htons(CE_VALID
)))
183 changed
= ce_match_stat_basic(ce
, st
);
186 * Within 1 second of this sequence:
187 * echo xyzzy >file && git-update-index --add file
188 * running this command:
190 * would give a falsely clean cache entry. The mtime and
191 * length match the cache, and other stat fields do not change.
193 * We could detect this at update-index time (the cache entry
194 * being registered/updated records the same time as "now")
195 * and delay the return from git-update-index, but that would
196 * effectively mean we can make at most one commit per second,
197 * which is not acceptable. Instead, we check cache entries
198 * whose mtime are the same as the index file timestamp more
199 * carefully than others.
202 index_file_timestamp
&&
203 index_file_timestamp
<= ntohl(ce
->ce_mtime
.sec
))
204 changed
|= ce_modified_check_fs(ce
, st
);
209 int ce_modified(struct cache_entry
*ce
, struct stat
*st
, int really
)
211 int changed
, changed_fs
;
212 changed
= ce_match_stat(ce
, st
, really
);
216 * If the mode or type has changed, there's no point in trying
217 * to refresh the entry - it's not going to match
219 if (changed
& (MODE_CHANGED
| TYPE_CHANGED
))
222 /* Immediately after read-tree or update-index --cacheinfo,
223 * the length field is zero. For other cases the ce_size
224 * should match the SHA1 recorded in the index entry.
226 if ((changed
& DATA_CHANGED
) && ce
->ce_size
!= htonl(0))
229 changed_fs
= ce_modified_check_fs(ce
, st
);
231 return changed
| changed_fs
;
235 int base_name_compare(const char *name1
, int len1
, int mode1
,
236 const char *name2
, int len2
, int mode2
)
238 unsigned char c1
, c2
;
239 int len
= len1
< len2
? len1
: len2
;
242 cmp
= memcmp(name1
, name2
, len
);
247 if (!c1
&& S_ISDIR(mode1
))
249 if (!c2
&& S_ISDIR(mode2
))
251 return (c1
< c2
) ? -1 : (c1
> c2
) ? 1 : 0;
254 int cache_name_compare(const char *name1
, int flags1
, const char *name2
, int flags2
)
256 int len1
= flags1
& CE_NAMEMASK
;
257 int len2
= flags2
& CE_NAMEMASK
;
258 int len
= len1
< len2
? len1
: len2
;
261 cmp
= memcmp(name1
, name2
, len
);
270 flags1
&= CE_STAGEMASK
;
271 flags2
&= CE_STAGEMASK
;
280 int cache_name_pos(const char *name
, int namelen
)
286 while (last
> first
) {
287 int next
= (last
+ first
) >> 1;
288 struct cache_entry
*ce
= active_cache
[next
];
289 int cmp
= cache_name_compare(name
, namelen
, ce
->name
, ntohs(ce
->ce_flags
));
301 /* Remove entry, return true if there are more entries to go.. */
302 int remove_cache_entry_at(int pos
)
304 active_cache_changed
= 1;
306 if (pos
>= active_nr
)
308 memmove(active_cache
+ pos
, active_cache
+ pos
+ 1, (active_nr
- pos
) * sizeof(struct cache_entry
*));
312 int remove_file_from_cache(const char *path
)
314 int pos
= cache_name_pos(path
, strlen(path
));
317 while (pos
< active_nr
&& !strcmp(active_cache
[pos
]->name
, path
))
318 remove_cache_entry_at(pos
);
322 int ce_same_name(struct cache_entry
*a
, struct cache_entry
*b
)
324 int len
= ce_namelen(a
);
325 return ce_namelen(b
) == len
&& !memcmp(a
->name
, b
->name
, len
);
328 int ce_path_match(const struct cache_entry
*ce
, const char **pathspec
)
330 const char *match
, *name
;
336 len
= ce_namelen(ce
);
338 while ((match
= *pathspec
++) != NULL
) {
339 int matchlen
= strlen(match
);
342 if (memcmp(name
, match
, matchlen
))
344 if (matchlen
&& name
[matchlen
-1] == '/')
346 if (name
[matchlen
] == '/' || !name
[matchlen
])
355 * We fundamentally don't like some paths: we don't want
356 * dot or dot-dot anywhere, and for obvious reasons don't
357 * want to recurse into ".git" either.
359 * Also, we don't want double slashes or slashes at the
360 * end that can make pathnames ambiguous.
362 static int verify_dotfile(const char *rest
)
365 * The first character was '.', but that
366 * has already been discarded, we now test
370 /* "." is not allowed */
375 * ".git" followed by NUL or slash is bad. This
376 * shares the path end test with the ".." case.
386 if (rest
[1] == '\0' || rest
[1] == '/')
392 int verify_path(const char *path
)
409 if (verify_dotfile(path
))
419 * Do we have another file that has the beginning components being a
420 * proper superset of the name we're trying to add?
422 static int has_file_name(const struct cache_entry
*ce
, int pos
, int ok_to_replace
)
425 int len
= ce_namelen(ce
);
426 int stage
= ce_stage(ce
);
427 const char *name
= ce
->name
;
429 while (pos
< active_nr
) {
430 struct cache_entry
*p
= active_cache
[pos
++];
432 if (len
>= ce_namelen(p
))
434 if (memcmp(name
, p
->name
, len
))
436 if (ce_stage(p
) != stage
)
438 if (p
->name
[len
] != '/')
443 remove_cache_entry_at(--pos
);
449 * Do we have another file with a pathname that is a proper
450 * subset of the name we're trying to add?
452 static int has_dir_name(const struct cache_entry
*ce
, int pos
, int ok_to_replace
)
455 int stage
= ce_stage(ce
);
456 const char *name
= ce
->name
;
457 const char *slash
= name
+ ce_namelen(ce
);
465 if (slash
<= ce
->name
)
470 pos
= cache_name_pos(name
, ntohs(create_ce_flags(len
, stage
)));
475 remove_cache_entry_at(pos
);
480 * Trivial optimization: if we find an entry that
481 * already matches the sub-directory, then we know
482 * we're ok, and we can exit.
485 while (pos
< active_nr
) {
486 struct cache_entry
*p
= active_cache
[pos
];
487 if ((ce_namelen(p
) <= len
) ||
488 (p
->name
[len
] != '/') ||
489 memcmp(p
->name
, name
, len
))
490 break; /* not our subdirectory */
491 if (ce_stage(p
) == stage
)
492 /* p is at the same stage as our entry, and
493 * is a subdirectory of what we are looking
494 * at, so we cannot have conflicts at our
495 * level or anything shorter.
504 /* We may be in a situation where we already have path/file and path
505 * is being added, or we already have path and path/file is being
506 * added. Either one would result in a nonsense tree that has path
507 * twice when git-write-tree tries to write it out. Prevent it.
509 * If ok-to-replace is specified, we remove the conflicting entries
510 * from the cache so the caller should recompute the insert position.
511 * When this happens, we return non-zero.
513 static int check_file_directory_conflict(const struct cache_entry
*ce
, int pos
, int ok_to_replace
)
516 * We check if the path is a sub-path of a subsequent pathname
517 * first, since removing those will not change the position
520 int retval
= has_file_name(ce
, pos
, ok_to_replace
);
522 * Then check if the path might have a clashing sub-directory
525 return retval
+ has_dir_name(ce
, pos
, ok_to_replace
);
528 int add_cache_entry(struct cache_entry
*ce
, int option
)
531 int ok_to_add
= option
& ADD_CACHE_OK_TO_ADD
;
532 int ok_to_replace
= option
& ADD_CACHE_OK_TO_REPLACE
;
533 int skip_df_check
= option
& ADD_CACHE_SKIP_DFCHECK
;
535 pos
= cache_name_pos(ce
->name
, ntohs(ce
->ce_flags
));
537 /* existing match? Just replace it. */
539 active_cache_changed
= 1;
540 active_cache
[pos
] = ce
;
546 * Inserting a merged entry ("stage 0") into the index
547 * will always replace all non-merged entries..
549 if (pos
< active_nr
&& ce_stage(ce
) == 0) {
550 while (ce_same_name(active_cache
[pos
], ce
)) {
552 if (!remove_cache_entry_at(pos
))
559 if (!verify_path(ce
->name
))
562 if (!skip_df_check
&&
563 check_file_directory_conflict(ce
, pos
, ok_to_replace
)) {
566 pos
= cache_name_pos(ce
->name
, ntohs(ce
->ce_flags
));
570 /* Make sure the array is big enough .. */
571 if (active_nr
== active_alloc
) {
572 active_alloc
= alloc_nr(active_alloc
);
573 active_cache
= xrealloc(active_cache
, active_alloc
* sizeof(struct cache_entry
*));
579 memmove(active_cache
+ pos
+ 1, active_cache
+ pos
, (active_nr
- pos
- 1) * sizeof(ce
));
580 active_cache
[pos
] = ce
;
581 active_cache_changed
= 1;
586 * "refresh" does not calculate a new sha1 file or bring the
587 * cache up-to-date for mode/content changes. But what it
588 * _does_ do is to "re-match" the stat information of a file
589 * with the cache, so that you can refresh the cache for a
590 * file that hasn't been changed but where the stat entry is
593 * For example, you'd want to do this after doing a "git-read-tree",
594 * to link up the stat cache details with the proper files.
596 struct cache_entry
*refresh_cache_entry(struct cache_entry
*ce
, int really
)
599 struct cache_entry
*updated
;
602 if (lstat(ce
->name
, &st
) < 0) {
607 changed
= ce_match_stat(ce
, &st
, really
);
609 if (really
&& assume_unchanged
&&
610 !(ce
->ce_flags
& htons(CE_VALID
)))
611 ; /* mark this one VALID again */
616 if (ce_modified(ce
, &st
, really
)) {
617 cache_errno
= EINVAL
;
622 updated
= xmalloc(size
);
623 memcpy(updated
, ce
, size
);
624 fill_stat_cache_info(updated
, &st
);
626 /* In this case, if really is not set, we should leave
627 * CE_VALID bit alone. Otherwise, paths marked with
628 * --no-assume-unchanged (i.e. things to be edited) will
629 * reacquire CE_VALID bit automatically, which is not
630 * really what we want.
632 if (!really
&& assume_unchanged
&& !(ce
->ce_flags
& htons(CE_VALID
)))
633 updated
->ce_flags
&= ~htons(CE_VALID
);
638 int refresh_cache(unsigned int flags
)
642 int really
= (flags
& REFRESH_REALLY
) != 0;
643 int allow_unmerged
= (flags
& REFRESH_UNMERGED
) != 0;
644 int quiet
= (flags
& REFRESH_QUIET
) != 0;
645 int not_new
= (flags
& REFRESH_IGNORE_MISSING
) != 0;
647 for (i
= 0; i
< active_nr
; i
++) {
648 struct cache_entry
*ce
, *new;
649 ce
= active_cache
[i
];
651 while ((i
< active_nr
) &&
652 ! strcmp(active_cache
[i
]->name
, ce
->name
))
657 printf("%s: needs merge\n", ce
->name
);
662 new = refresh_cache_entry(ce
, really
);
666 if (not_new
&& cache_errno
== ENOENT
)
668 if (really
&& cache_errno
== EINVAL
) {
669 /* If we are doing --really-refresh that
670 * means the index is not valid anymore.
672 ce
->ce_flags
&= ~htons(CE_VALID
);
673 active_cache_changed
= 1;
677 printf("%s: needs update\n", ce
->name
);
681 active_cache_changed
= 1;
682 /* You can NOT just free active_cache[i] here, since it
683 * might not be necessarily malloc()ed but can also come
685 active_cache
[i
] = new;
690 static int verify_hdr(struct cache_header
*hdr
, unsigned long size
)
693 unsigned char sha1
[20];
695 if (hdr
->hdr_signature
!= htonl(CACHE_SIGNATURE
))
696 return error("bad signature");
697 if (hdr
->hdr_version
!= htonl(2))
698 return error("bad index version");
700 SHA1_Update(&c
, hdr
, size
- 20);
701 SHA1_Final(sha1
, &c
);
702 if (memcmp(sha1
, (char *) hdr
+ size
- 20, 20))
703 return error("bad index file sha1 signature");
707 static int read_index_extension(const char *ext
, void *data
, unsigned long sz
)
709 switch (CACHE_EXT(ext
)) {
711 active_cache_tree
= cache_tree_read(data
, sz
);
714 if (*ext
< 'A' || 'Z' < *ext
)
715 return error("index uses %.4s extension, which we do not understand",
717 fprintf(stderr
, "ignoring %.4s extension\n", ext
);
725 return read_cache_from(get_index_file());
728 /* remember to discard_cache() before reading a different cache! */
729 int read_cache_from(const char *path
)
733 unsigned long offset
;
734 struct cache_header
*hdr
;
741 index_file_timestamp
= 0;
742 fd
= open(path
, O_RDONLY
);
746 die("index file open failed (%s)", strerror(errno
));
749 cache_mmap
= MAP_FAILED
;
750 if (!fstat(fd
, &st
)) {
751 cache_mmap_size
= st
.st_size
;
753 if (cache_mmap_size
>= sizeof(struct cache_header
) + 20)
754 cache_mmap
= mmap(NULL
, cache_mmap_size
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
, fd
, 0);
757 if (cache_mmap
== MAP_FAILED
)
758 die("index file mmap failed (%s)", strerror(errno
));
761 if (verify_hdr(hdr
, cache_mmap_size
) < 0)
764 active_nr
= ntohl(hdr
->hdr_entries
);
765 active_alloc
= alloc_nr(active_nr
);
766 active_cache
= xcalloc(active_alloc
, sizeof(struct cache_entry
*));
768 offset
= sizeof(*hdr
);
769 for (i
= 0; i
< active_nr
; i
++) {
770 struct cache_entry
*ce
= (struct cache_entry
*) ((char *) cache_mmap
+ offset
);
771 offset
= offset
+ ce_size(ce
);
772 active_cache
[i
] = ce
;
774 index_file_timestamp
= st
.st_mtime
;
775 while (offset
<= cache_mmap_size
- 20 - 8) {
776 /* After an array of active_nr index entries,
777 * there can be arbitrary number of extended
778 * sections, each of which is prefixed with
779 * extension name (4-byte) and section length
780 * in 4-byte network byte order.
782 unsigned long extsize
;
783 memcpy(&extsize
, (char *) cache_mmap
+ offset
+ 4, 4);
784 extsize
= ntohl(extsize
);
785 if (read_index_extension(((const char *) cache_mmap
) + offset
,
786 (char *) cache_mmap
+ offset
+ 8,
795 munmap(cache_mmap
, cache_mmap_size
);
797 die("index file corrupt");
804 if (cache_mmap
== NULL
)
806 ret
= munmap(cache_mmap
, cache_mmap_size
);
809 active_nr
= active_cache_changed
= 0;
810 index_file_timestamp
= 0;
811 cache_tree_free(&active_cache_tree
);
813 /* no need to throw away allocated active_cache */
817 #define WRITE_BUFFER_SIZE 8192
818 static unsigned char write_buffer
[WRITE_BUFFER_SIZE
];
819 static unsigned long write_buffer_len
;
821 static int ce_write(SHA_CTX
*context
, int fd
, void *data
, unsigned int len
)
824 unsigned int buffered
= write_buffer_len
;
825 unsigned int partial
= WRITE_BUFFER_SIZE
- buffered
;
828 memcpy(write_buffer
+ buffered
, data
, partial
);
830 if (buffered
== WRITE_BUFFER_SIZE
) {
831 SHA1_Update(context
, write_buffer
, WRITE_BUFFER_SIZE
);
832 if (write(fd
, write_buffer
, WRITE_BUFFER_SIZE
) != WRITE_BUFFER_SIZE
)
836 write_buffer_len
= buffered
;
838 data
= (char *) data
+ partial
;
843 static int write_index_ext_header(SHA_CTX
*context
, int fd
,
844 unsigned int ext
, unsigned int sz
)
848 if ((ce_write(context
, fd
, &ext
, 4) < 0) ||
849 (ce_write(context
, fd
, &sz
, 4) < 0))
854 static int ce_flush(SHA_CTX
*context
, int fd
)
856 unsigned int left
= write_buffer_len
;
859 write_buffer_len
= 0;
860 SHA1_Update(context
, write_buffer
, left
);
863 /* Flush first if not enough space for SHA1 signature */
864 if (left
+ 20 > WRITE_BUFFER_SIZE
) {
865 if (write(fd
, write_buffer
, left
) != left
)
870 /* Append the SHA1 signature at the end */
871 SHA1_Final(write_buffer
+ left
, context
);
873 if (write(fd
, write_buffer
, left
) != left
)
878 static void ce_smudge_racily_clean_entry(struct cache_entry
*ce
)
881 * The only thing we care about in this function is to smudge the
882 * falsely clean entry due to touch-update-touch race, so we leave
883 * everything else as they are. We are called for entries whose
884 * ce_mtime match the index file mtime.
888 if (lstat(ce
->name
, &st
) < 0)
890 if (ce_match_stat_basic(ce
, &st
))
892 if (ce_modified_check_fs(ce
, &st
)) {
893 /* This is "racily clean"; smudge it. Note that this
894 * is a tricky code. At first glance, it may appear
895 * that it can break with this sequence:
897 * $ echo xyzzy >frotz
898 * $ git-update-index --add frotz
901 * $ echo filfre >nitfol
902 * $ git-update-index --add nitfol
904 * but it does not. Whe the second update-index runs,
905 * it notices that the entry "frotz" has the same timestamp
906 * as index, and if we were to smudge it by resetting its
907 * size to zero here, then the object name recorded
908 * in index is the 6-byte file but the cached stat information
909 * becomes zero --- which would then match what we would
910 * obtain from the filesystem next time we stat("frotz").
912 * However, the second update-index, before calling
913 * this function, notices that the cached size is 6
914 * bytes and what is on the filesystem is an empty
915 * file, and never calls us, so the cached size information
916 * for "frotz" stays 6 which does not match the filesystem.
918 ce
->ce_size
= htonl(0);
922 int write_cache(int newfd
, struct cache_entry
**cache
, int entries
)
925 struct cache_header hdr
;
928 for (i
= removed
= 0; i
< entries
; i
++)
929 if (!cache
[i
]->ce_mode
)
932 hdr
.hdr_signature
= htonl(CACHE_SIGNATURE
);
933 hdr
.hdr_version
= htonl(2);
934 hdr
.hdr_entries
= htonl(entries
- removed
);
937 if (ce_write(&c
, newfd
, &hdr
, sizeof(hdr
)) < 0)
940 for (i
= 0; i
< entries
; i
++) {
941 struct cache_entry
*ce
= cache
[i
];
944 if (index_file_timestamp
&&
945 index_file_timestamp
<= ntohl(ce
->ce_mtime
.sec
))
946 ce_smudge_racily_clean_entry(ce
);
947 if (ce_write(&c
, newfd
, ce
, ce_size(ce
)) < 0)
951 /* Write extension data here */
952 if (active_cache_tree
) {
954 void *data
= cache_tree_write(active_cache_tree
, &sz
);
956 !write_index_ext_header(&c
, newfd
, CACHE_EXT_TREE
, sz
) &&
957 !ce_write(&c
, newfd
, data
, sz
))
964 return ce_flush(&c
, newfd
);