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
;
28 * This only updates the "non-critical" parts of the directory
29 * cache, ie the parts that aren't tracked by GIT, and only used
30 * to validate the cache.
32 void fill_stat_cache_info(struct cache_entry
*ce
, struct stat
*st
)
34 ce
->ce_ctime
.sec
= htonl(st
->st_ctime
);
35 ce
->ce_mtime
.sec
= htonl(st
->st_mtime
);
37 ce
->ce_ctime
.nsec
= htonl(st
->st_ctim
.tv_nsec
);
38 ce
->ce_mtime
.nsec
= htonl(st
->st_mtim
.tv_nsec
);
40 ce
->ce_dev
= htonl(st
->st_dev
);
41 ce
->ce_ino
= htonl(st
->st_ino
);
42 ce
->ce_uid
= htonl(st
->st_uid
);
43 ce
->ce_gid
= htonl(st
->st_gid
);
44 ce
->ce_size
= htonl(st
->st_size
);
47 ce
->ce_flags
|= htons(CE_VALID
);
50 static int ce_compare_data(struct cache_entry
*ce
, struct stat
*st
)
53 int fd
= open(ce
->name
, O_RDONLY
);
56 unsigned char sha1
[20];
57 if (!index_fd(sha1
, fd
, st
, 0, NULL
))
58 match
= memcmp(sha1
, ce
->sha1
, 20);
64 static int ce_compare_link(struct cache_entry
*ce
, unsigned long expected_size
)
73 target
= xmalloc(expected_size
);
74 len
= readlink(ce
->name
, target
, expected_size
);
75 if (len
!= expected_size
) {
79 buffer
= read_sha1_file(ce
->sha1
, type
, &size
);
84 if (size
== expected_size
)
85 match
= memcmp(buffer
, target
, size
);
91 static int ce_modified_check_fs(struct cache_entry
*ce
, struct stat
*st
)
93 switch (st
->st_mode
& S_IFMT
) {
95 if (ce_compare_data(ce
, st
))
99 if (ce_compare_link(ce
, st
->st_size
))
108 static int ce_match_stat_basic(struct cache_entry
*ce
, struct stat
*st
)
110 unsigned int changed
= 0;
112 switch (ntohl(ce
->ce_mode
) & S_IFMT
) {
114 changed
|= !S_ISREG(st
->st_mode
) ? TYPE_CHANGED
: 0;
115 /* We consider only the owner x bit to be relevant for
118 if (trust_executable_bit
&&
119 (0100 & (ntohl(ce
->ce_mode
) ^ st
->st_mode
)))
120 changed
|= MODE_CHANGED
;
123 changed
|= !S_ISLNK(st
->st_mode
) ? TYPE_CHANGED
: 0;
126 die("internal error: ce_mode is %o", ntohl(ce
->ce_mode
));
128 if (ce
->ce_mtime
.sec
!= htonl(st
->st_mtime
))
129 changed
|= MTIME_CHANGED
;
130 if (ce
->ce_ctime
.sec
!= htonl(st
->st_ctime
))
131 changed
|= CTIME_CHANGED
;
135 * nsec seems unreliable - not all filesystems support it, so
136 * as long as it is in the inode cache you get right nsec
137 * but after it gets flushed, you get zero nsec.
139 if (ce
->ce_mtime
.nsec
!= htonl(st
->st_mtim
.tv_nsec
))
140 changed
|= MTIME_CHANGED
;
141 if (ce
->ce_ctime
.nsec
!= htonl(st
->st_ctim
.tv_nsec
))
142 changed
|= CTIME_CHANGED
;
145 if (ce
->ce_uid
!= htonl(st
->st_uid
) ||
146 ce
->ce_gid
!= htonl(st
->st_gid
))
147 changed
|= OWNER_CHANGED
;
148 if (ce
->ce_ino
!= htonl(st
->st_ino
))
149 changed
|= INODE_CHANGED
;
153 * st_dev breaks on network filesystems where different
154 * clients will have different views of what "device"
155 * the filesystem is on
157 if (ce
->ce_dev
!= htonl(st
->st_dev
))
158 changed
|= INODE_CHANGED
;
161 if (ce
->ce_size
!= htonl(st
->st_size
))
162 changed
|= DATA_CHANGED
;
167 int ce_match_stat(struct cache_entry
*ce
, struct stat
*st
, int ignore_valid
)
169 unsigned int changed
;
172 * If it's marked as always valid in the index, it's
173 * valid whatever the checked-out copy says.
175 if (!ignore_valid
&& (ce
->ce_flags
& htons(CE_VALID
)))
178 changed
= ce_match_stat_basic(ce
, st
);
181 * Within 1 second of this sequence:
182 * echo xyzzy >file && git-update-index --add file
183 * running this command:
185 * would give a falsely clean cache entry. The mtime and
186 * length match the cache, and other stat fields do not change.
188 * We could detect this at update-index time (the cache entry
189 * being registered/updated records the same time as "now")
190 * and delay the return from git-update-index, but that would
191 * effectively mean we can make at most one commit per second,
192 * which is not acceptable. Instead, we check cache entries
193 * whose mtime are the same as the index file timestamp more
194 * carefully than others.
197 index_file_timestamp
&&
198 index_file_timestamp
<= ntohl(ce
->ce_mtime
.sec
))
199 changed
|= ce_modified_check_fs(ce
, st
);
204 int ce_modified(struct cache_entry
*ce
, struct stat
*st
, int really
)
206 int changed
, changed_fs
;
207 changed
= ce_match_stat(ce
, st
, really
);
211 * If the mode or type has changed, there's no point in trying
212 * to refresh the entry - it's not going to match
214 if (changed
& (MODE_CHANGED
| TYPE_CHANGED
))
217 /* Immediately after read-tree or update-index --cacheinfo,
218 * the length field is zero. For other cases the ce_size
219 * should match the SHA1 recorded in the index entry.
221 if ((changed
& DATA_CHANGED
) && ce
->ce_size
!= htonl(0))
224 changed_fs
= ce_modified_check_fs(ce
, st
);
226 return changed
| changed_fs
;
230 int base_name_compare(const char *name1
, int len1
, int mode1
,
231 const char *name2
, int len2
, int mode2
)
233 unsigned char c1
, c2
;
234 int len
= len1
< len2
? len1
: len2
;
237 cmp
= memcmp(name1
, name2
, len
);
242 if (!c1
&& S_ISDIR(mode1
))
244 if (!c2
&& S_ISDIR(mode2
))
246 return (c1
< c2
) ? -1 : (c1
> c2
) ? 1 : 0;
249 int cache_name_compare(const char *name1
, int flags1
, const char *name2
, int flags2
)
251 int len1
= flags1
& CE_NAMEMASK
;
252 int len2
= flags2
& CE_NAMEMASK
;
253 int len
= len1
< len2
? len1
: len2
;
256 cmp
= memcmp(name1
, name2
, len
);
265 flags1
&= CE_STAGEMASK
;
266 flags2
&= CE_STAGEMASK
;
275 int cache_name_pos(const char *name
, int namelen
)
281 while (last
> first
) {
282 int next
= (last
+ first
) >> 1;
283 struct cache_entry
*ce
= active_cache
[next
];
284 int cmp
= cache_name_compare(name
, namelen
, ce
->name
, ntohs(ce
->ce_flags
));
296 /* Remove entry, return true if there are more entries to go.. */
297 int remove_cache_entry_at(int pos
)
299 active_cache_changed
= 1;
301 if (pos
>= active_nr
)
303 memmove(active_cache
+ pos
, active_cache
+ pos
+ 1, (active_nr
- pos
) * sizeof(struct cache_entry
*));
307 int remove_file_from_cache(const char *path
)
309 int pos
= cache_name_pos(path
, strlen(path
));
312 while (pos
< active_nr
&& !strcmp(active_cache
[pos
]->name
, path
))
313 remove_cache_entry_at(pos
);
317 int ce_same_name(struct cache_entry
*a
, struct cache_entry
*b
)
319 int len
= ce_namelen(a
);
320 return ce_namelen(b
) == len
&& !memcmp(a
->name
, b
->name
, len
);
323 int ce_path_match(const struct cache_entry
*ce
, const char **pathspec
)
325 const char *match
, *name
;
331 len
= ce_namelen(ce
);
333 while ((match
= *pathspec
++) != NULL
) {
334 int matchlen
= strlen(match
);
337 if (memcmp(name
, match
, matchlen
))
339 if (matchlen
&& name
[matchlen
-1] == '/')
341 if (name
[matchlen
] == '/' || !name
[matchlen
])
350 * We fundamentally don't like some paths: we don't want
351 * dot or dot-dot anywhere, and for obvious reasons don't
352 * want to recurse into ".git" either.
354 * Also, we don't want double slashes or slashes at the
355 * end that can make pathnames ambiguous.
357 static int verify_dotfile(const char *rest
)
360 * The first character was '.', but that
361 * has already been discarded, we now test
365 /* "." is not allowed */
370 * ".git" followed by NUL or slash is bad. This
371 * shares the path end test with the ".." case.
381 if (rest
[1] == '\0' || rest
[1] == '/')
387 int verify_path(const char *path
)
404 if (verify_dotfile(path
))
414 * Do we have another file that has the beginning components being a
415 * proper superset of the name we're trying to add?
417 static int has_file_name(const struct cache_entry
*ce
, int pos
, int ok_to_replace
)
420 int len
= ce_namelen(ce
);
421 int stage
= ce_stage(ce
);
422 const char *name
= ce
->name
;
424 while (pos
< active_nr
) {
425 struct cache_entry
*p
= active_cache
[pos
++];
427 if (len
>= ce_namelen(p
))
429 if (memcmp(name
, p
->name
, len
))
431 if (ce_stage(p
) != stage
)
433 if (p
->name
[len
] != '/')
438 remove_cache_entry_at(--pos
);
444 * Do we have another file with a pathname that is a proper
445 * subset of the name we're trying to add?
447 static int has_dir_name(const struct cache_entry
*ce
, int pos
, int ok_to_replace
)
450 int stage
= ce_stage(ce
);
451 const char *name
= ce
->name
;
452 const char *slash
= name
+ ce_namelen(ce
);
460 if (slash
<= ce
->name
)
465 pos
= cache_name_pos(name
, ntohs(create_ce_flags(len
, stage
)));
470 remove_cache_entry_at(pos
);
475 * Trivial optimization: if we find an entry that
476 * already matches the sub-directory, then we know
477 * we're ok, and we can exit.
480 while (pos
< active_nr
) {
481 struct cache_entry
*p
= active_cache
[pos
];
482 if ((ce_namelen(p
) <= len
) ||
483 (p
->name
[len
] != '/') ||
484 memcmp(p
->name
, name
, len
))
485 break; /* not our subdirectory */
486 if (ce_stage(p
) == stage
)
487 /* p is at the same stage as our entry, and
488 * is a subdirectory of what we are looking
489 * at, so we cannot have conflicts at our
490 * level or anything shorter.
499 /* We may be in a situation where we already have path/file and path
500 * is being added, or we already have path and path/file is being
501 * added. Either one would result in a nonsense tree that has path
502 * twice when git-write-tree tries to write it out. Prevent it.
504 * If ok-to-replace is specified, we remove the conflicting entries
505 * from the cache so the caller should recompute the insert position.
506 * When this happens, we return non-zero.
508 static int check_file_directory_conflict(const struct cache_entry
*ce
, int pos
, int ok_to_replace
)
511 * We check if the path is a sub-path of a subsequent pathname
512 * first, since removing those will not change the position
515 int retval
= has_file_name(ce
, pos
, ok_to_replace
);
517 * Then check if the path might have a clashing sub-directory
520 return retval
+ has_dir_name(ce
, pos
, ok_to_replace
);
523 int add_cache_entry(struct cache_entry
*ce
, int option
)
526 int ok_to_add
= option
& ADD_CACHE_OK_TO_ADD
;
527 int ok_to_replace
= option
& ADD_CACHE_OK_TO_REPLACE
;
528 int skip_df_check
= option
& ADD_CACHE_SKIP_DFCHECK
;
530 pos
= cache_name_pos(ce
->name
, ntohs(ce
->ce_flags
));
532 /* existing match? Just replace it. */
534 active_cache_changed
= 1;
535 active_cache
[pos
] = ce
;
541 * Inserting a merged entry ("stage 0") into the index
542 * will always replace all non-merged entries..
544 if (pos
< active_nr
&& ce_stage(ce
) == 0) {
545 while (ce_same_name(active_cache
[pos
], ce
)) {
547 if (!remove_cache_entry_at(pos
))
554 if (!verify_path(ce
->name
))
557 if (!skip_df_check
&&
558 check_file_directory_conflict(ce
, pos
, ok_to_replace
)) {
561 pos
= cache_name_pos(ce
->name
, ntohs(ce
->ce_flags
));
565 /* Make sure the array is big enough .. */
566 if (active_nr
== active_alloc
) {
567 active_alloc
= alloc_nr(active_alloc
);
568 active_cache
= xrealloc(active_cache
, active_alloc
* sizeof(struct cache_entry
*));
574 memmove(active_cache
+ pos
+ 1, active_cache
+ pos
, (active_nr
- pos
- 1) * sizeof(ce
));
575 active_cache
[pos
] = ce
;
576 active_cache_changed
= 1;
580 /* Three functions to allow overloaded pointer return; see linux/err.h */
581 static inline void *ERR_PTR(long error
)
583 return (void *) error
;
586 static inline long PTR_ERR(const void *ptr
)
591 static inline long IS_ERR(const void *ptr
)
593 return (unsigned long)ptr
> (unsigned long)-1000L;
597 * "refresh" does not calculate a new sha1 file or bring the
598 * cache up-to-date for mode/content changes. But what it
599 * _does_ do is to "re-match" the stat information of a file
600 * with the cache, so that you can refresh the cache for a
601 * file that hasn't been changed but where the stat entry is
604 * For example, you'd want to do this after doing a "git-read-tree",
605 * to link up the stat cache details with the proper files.
607 static struct cache_entry
*refresh_entry(struct cache_entry
*ce
, int really
)
610 struct cache_entry
*updated
;
613 if (lstat(ce
->name
, &st
) < 0)
614 return ERR_PTR(-errno
);
616 changed
= ce_match_stat(ce
, &st
, really
);
618 if (really
&& assume_unchanged
&&
619 !(ce
->ce_flags
& htons(CE_VALID
)))
620 ; /* mark this one VALID again */
625 if (ce_modified(ce
, &st
, really
))
626 return ERR_PTR(-EINVAL
);
629 updated
= xmalloc(size
);
630 memcpy(updated
, ce
, size
);
631 fill_stat_cache_info(updated
, &st
);
633 /* In this case, if really is not set, we should leave
634 * CE_VALID bit alone. Otherwise, paths marked with
635 * --no-assume-unchanged (i.e. things to be edited) will
636 * reacquire CE_VALID bit automatically, which is not
637 * really what we want.
639 if (!really
&& assume_unchanged
&& !(ce
->ce_flags
& htons(CE_VALID
)))
640 updated
->ce_flags
&= ~htons(CE_VALID
);
645 int refresh_cache(unsigned int flags
)
649 int really
= (flags
& REFRESH_REALLY
) != 0;
650 int allow_unmerged
= (flags
& REFRESH_UNMERGED
) != 0;
651 int quiet
= (flags
& REFRESH_QUIET
) != 0;
652 int not_new
= (flags
& REFRESH_IGNORE_MISSING
) != 0;
654 for (i
= 0; i
< active_nr
; i
++) {
655 struct cache_entry
*ce
, *new;
656 ce
= active_cache
[i
];
658 while ((i
< active_nr
) &&
659 ! strcmp(active_cache
[i
]->name
, ce
->name
))
664 printf("%s: needs merge\n", ce
->name
);
669 new = refresh_entry(ce
, really
);
673 if (not_new
&& PTR_ERR(new) == -ENOENT
)
675 if (really
&& PTR_ERR(new) == -EINVAL
) {
676 /* If we are doing --really-refresh that
677 * means the index is not valid anymore.
679 ce
->ce_flags
&= ~htons(CE_VALID
);
680 active_cache_changed
= 1;
684 printf("%s: needs update\n", ce
->name
);
688 active_cache_changed
= 1;
689 /* You can NOT just free active_cache[i] here, since it
690 * might not be necessarily malloc()ed but can also come
692 active_cache
[i
] = new;
697 static int verify_hdr(struct cache_header
*hdr
, unsigned long size
)
700 unsigned char sha1
[20];
702 if (hdr
->hdr_signature
!= htonl(CACHE_SIGNATURE
))
703 return error("bad signature");
704 if (hdr
->hdr_version
!= htonl(2))
705 return error("bad index version");
707 SHA1_Update(&c
, hdr
, size
- 20);
708 SHA1_Final(sha1
, &c
);
709 if (memcmp(sha1
, (char *) hdr
+ size
- 20, 20))
710 return error("bad index file sha1 signature");
714 static int read_index_extension(const char *ext
, void *data
, unsigned long sz
)
716 switch (CACHE_EXT(ext
)) {
718 active_cache_tree
= cache_tree_read(data
, sz
);
721 if (*ext
< 'A' || 'Z' < *ext
)
722 return error("index uses %.4s extension, which we do not understand",
724 fprintf(stderr
, "ignoring %.4s extension\n", ext
);
734 unsigned long size
, offset
;
736 struct cache_header
*hdr
;
743 index_file_timestamp
= 0;
744 fd
= open(get_index_file(), O_RDONLY
);
748 die("index file open failed (%s)", strerror(errno
));
751 size
= 0; /* avoid gcc warning */
753 if (!fstat(fd
, &st
)) {
756 if (size
>= sizeof(struct cache_header
) + 20)
757 map
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
, fd
, 0);
760 if (map
== MAP_FAILED
)
761 die("index file mmap failed (%s)", strerror(errno
));
764 if (verify_hdr(hdr
, size
) < 0)
767 active_nr
= ntohl(hdr
->hdr_entries
);
768 active_alloc
= alloc_nr(active_nr
);
769 active_cache
= xcalloc(active_alloc
, sizeof(struct cache_entry
*));
771 offset
= sizeof(*hdr
);
772 for (i
= 0; i
< active_nr
; i
++) {
773 struct cache_entry
*ce
= (struct cache_entry
*) ((char *) map
+ offset
);
774 offset
= offset
+ ce_size(ce
);
775 active_cache
[i
] = ce
;
777 index_file_timestamp
= st
.st_mtime
;
778 while (offset
<= size
- 20 - 8) {
779 /* After an array of active_nr index entries,
780 * there can be arbitrary number of extended
781 * sections, each of which is prefixed with
782 * extension name (4-byte) and section length
783 * in 4-byte network byte order.
785 unsigned long extsize
;
786 memcpy(&extsize
, (char *) map
+ offset
+ 4, 4);
787 extsize
= ntohl(extsize
);
788 if (read_index_extension(((const char *) map
) + offset
,
789 (char *) map
+ offset
+ 8,
800 die("index file corrupt");
803 #define WRITE_BUFFER_SIZE 8192
804 static unsigned char write_buffer
[WRITE_BUFFER_SIZE
];
805 static unsigned long write_buffer_len
;
807 static int ce_write(SHA_CTX
*context
, int fd
, void *data
, unsigned int len
)
810 unsigned int buffered
= write_buffer_len
;
811 unsigned int partial
= WRITE_BUFFER_SIZE
- buffered
;
814 memcpy(write_buffer
+ buffered
, data
, partial
);
816 if (buffered
== WRITE_BUFFER_SIZE
) {
817 SHA1_Update(context
, write_buffer
, WRITE_BUFFER_SIZE
);
818 if (write(fd
, write_buffer
, WRITE_BUFFER_SIZE
) != WRITE_BUFFER_SIZE
)
822 write_buffer_len
= buffered
;
824 data
= (char *) data
+ partial
;
829 static int write_index_ext_header(SHA_CTX
*context
, int fd
,
830 unsigned int ext
, unsigned int sz
)
834 if ((ce_write(context
, fd
, &ext
, 4) < 0) ||
835 (ce_write(context
, fd
, &sz
, 4) < 0))
840 static int ce_flush(SHA_CTX
*context
, int fd
)
842 unsigned int left
= write_buffer_len
;
845 write_buffer_len
= 0;
846 SHA1_Update(context
, write_buffer
, left
);
849 /* Flush first if not enough space for SHA1 signature */
850 if (left
+ 20 > WRITE_BUFFER_SIZE
) {
851 if (write(fd
, write_buffer
, left
) != left
)
856 /* Append the SHA1 signature at the end */
857 SHA1_Final(write_buffer
+ left
, context
);
859 if (write(fd
, write_buffer
, left
) != left
)
864 static void ce_smudge_racily_clean_entry(struct cache_entry
*ce
)
867 * The only thing we care about in this function is to smudge the
868 * falsely clean entry due to touch-update-touch race, so we leave
869 * everything else as they are. We are called for entries whose
870 * ce_mtime match the index file mtime.
874 if (lstat(ce
->name
, &st
) < 0)
876 if (ce_match_stat_basic(ce
, &st
))
878 if (ce_modified_check_fs(ce
, &st
)) {
879 /* This is "racily clean"; smudge it. Note that this
880 * is a tricky code. At first glance, it may appear
881 * that it can break with this sequence:
883 * $ echo xyzzy >frotz
884 * $ git-update-index --add frotz
887 * $ echo filfre >nitfol
888 * $ git-update-index --add nitfol
890 * but it does not. Whe the second update-index runs,
891 * it notices that the entry "frotz" has the same timestamp
892 * as index, and if we were to smudge it by resetting its
893 * size to zero here, then the object name recorded
894 * in index is the 6-byte file but the cached stat information
895 * becomes zero --- which would then match what we would
896 * obtain from the filesystem next time we stat("frotz").
898 * However, the second update-index, before calling
899 * this function, notices that the cached size is 6
900 * bytes and what is on the filesystem is an empty
901 * file, and never calls us, so the cached size information
902 * for "frotz" stays 6 which does not match the filesystem.
904 ce
->ce_size
= htonl(0);
908 int write_cache(int newfd
, struct cache_entry
**cache
, int entries
)
911 struct cache_header hdr
;
914 for (i
= removed
= 0; i
< entries
; i
++)
915 if (!cache
[i
]->ce_mode
)
918 hdr
.hdr_signature
= htonl(CACHE_SIGNATURE
);
919 hdr
.hdr_version
= htonl(2);
920 hdr
.hdr_entries
= htonl(entries
- removed
);
923 if (ce_write(&c
, newfd
, &hdr
, sizeof(hdr
)) < 0)
926 for (i
= 0; i
< entries
; i
++) {
927 struct cache_entry
*ce
= cache
[i
];
930 if (index_file_timestamp
&&
931 index_file_timestamp
<= ntohl(ce
->ce_mtime
.sec
))
932 ce_smudge_racily_clean_entry(ce
);
933 if (ce_write(&c
, newfd
, ce
, ce_size(ce
)) < 0)
937 /* Write extension data here */
938 if (active_cache_tree
) {
940 void *data
= cache_tree_write(active_cache_tree
, &sz
);
942 !write_index_ext_header(&c
, newfd
, CACHE_EXT_TREE
, sz
) &&
943 !ce_write(&c
, newfd
, data
, sz
))
950 return ce_flush(&c
, newfd
);