1 /*
2 * GIT - The information manager from hell
3 *
4 * Copyright (C) Linus Torvalds, 2005
5 */
6 #include "cache.h"
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;
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 }
32 static int ce_compare_data(struct cache_entry *ce, struct stat *st)
33 {
34 int match = -1;
35 int fd = open(ce->name, O_RDONLY);
37 if (fd >= 0) {
38 unsigned char sha1[20];
39 if (!index_fd(sha1, fd, st, 0, NULL))
40 match = memcmp(sha1, ce->sha1, 20);
41 close(fd);
42 }
43 return match;
44 }
46 static int ce_compare_link(struct cache_entry *ce, unsigned long expected_size)
47 {
48 int match = -1;
49 char *target;
50 void *buffer;
51 unsigned long size;
52 char type[10];
53 int len;
55 target = xmalloc(expected_size);
56 len = readlink(ce->name, target, expected_size);
57 if (len != expected_size) {
58 free(target);
59 return -1;
60 }
61 buffer = read_sha1_file(ce->sha1, type, &size);
62 if (!buffer) {
63 free(target);
64 return -1;
65 }
66 if (size == expected_size)
67 match = memcmp(buffer, target, size);
68 free(buffer);
69 free(target);
70 return match;
71 }
73 static int ce_modified_check_fs(struct cache_entry *ce, struct stat *st)
74 {
75 switch (st->st_mode & S_IFMT) {
76 case S_IFREG:
77 if (ce_compare_data(ce, st))
78 return DATA_CHANGED;
79 break;
80 case S_IFLNK:
81 if (ce_compare_link(ce, st->st_size))
82 return DATA_CHANGED;
83 break;
84 default:
85 return TYPE_CHANGED;
86 }
87 return 0;
88 }
90 static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
91 {
92 unsigned int changed = 0;
94 switch (ntohl(ce->ce_mode) & S_IFMT) {
95 case S_IFREG:
96 changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
97 /* We consider only the owner x bit to be relevant for
98 * "mode changes"
99 */
100 if (trust_executable_bit &&
101 (0100 & (ntohl(ce->ce_mode) ^ st->st_mode)))
102 changed |= MODE_CHANGED;
103 break;
104 case S_IFLNK:
105 changed |= !S_ISLNK(st->st_mode) ? TYPE_CHANGED : 0;
106 break;
107 default:
108 die("internal error: ce_mode is %o", ntohl(ce->ce_mode));
109 }
110 if (ce->ce_mtime.sec != htonl(st->st_mtime))
111 changed |= MTIME_CHANGED;
112 if (ce->ce_ctime.sec != htonl(st->st_ctime))
113 changed |= CTIME_CHANGED;
115 #ifdef USE_NSEC
116 /*
117 * nsec seems unreliable - not all filesystems support it, so
118 * as long as it is in the inode cache you get right nsec
119 * but after it gets flushed, you get zero nsec.
120 */
121 if (ce->ce_mtime.nsec != htonl(st->st_mtim.tv_nsec))
122 changed |= MTIME_CHANGED;
123 if (ce->ce_ctime.nsec != htonl(st->st_ctim.tv_nsec))
124 changed |= CTIME_CHANGED;
125 #endif
127 if (ce->ce_uid != htonl(st->st_uid) ||
128 ce->ce_gid != htonl(st->st_gid))
129 changed |= OWNER_CHANGED;
130 if (ce->ce_ino != htonl(st->st_ino))
131 changed |= INODE_CHANGED;
133 #ifdef USE_STDEV
134 /*
135 * st_dev breaks on network filesystems where different
136 * clients will have different views of what "device"
137 * the filesystem is on
138 */
139 if (ce->ce_dev != htonl(st->st_dev))
140 changed |= INODE_CHANGED;
141 #endif
143 if (ce->ce_size != htonl(st->st_size))
144 changed |= DATA_CHANGED;
146 return changed;
147 }
149 int ce_match_stat(struct cache_entry *ce, struct stat *st)
150 {
151 unsigned int changed = ce_match_stat_basic(ce, st);
153 /*
154 * Within 1 second of this sequence:
155 * echo xyzzy >file && git-update-index --add file
156 * running this command:
157 * echo frotz >file
158 * would give a falsely clean cache entry. The mtime and
159 * length match the cache, and other stat fields do not change.
160 *
161 * We could detect this at update-index time (the cache entry
162 * being registered/updated records the same time as "now")
163 * and delay the return from git-update-index, but that would
164 * effectively mean we can make at most one commit per second,
165 * which is not acceptable. Instead, we check cache entries
166 * whose mtime are the same as the index file timestamp more
167 * careful than others.
168 */
169 if (!changed &&
170 index_file_timestamp &&
171 index_file_timestamp <= ntohl(ce->ce_mtime.sec))
172 changed |= ce_modified_check_fs(ce, st);
174 return changed;
175 }
177 int ce_modified(struct cache_entry *ce, struct stat *st)
178 {
179 int changed, changed_fs;
180 changed = ce_match_stat(ce, st);
181 if (!changed)
182 return 0;
183 /*
184 * If the mode or type has changed, there's no point in trying
185 * to refresh the entry - it's not going to match
186 */
187 if (changed & (MODE_CHANGED | TYPE_CHANGED))
188 return changed;
190 /* Immediately after read-tree or update-index --cacheinfo,
191 * the length field is zero. For other cases the ce_size
192 * should match the SHA1 recorded in the index entry.
193 */
194 if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
195 return changed;
197 changed_fs = ce_modified_check_fs(ce, st);
198 if (changed_fs)
199 return changed | changed_fs;
200 return 0;
201 }
203 int base_name_compare(const char *name1, int len1, int mode1,
204 const char *name2, int len2, int mode2)
205 {
206 unsigned char c1, c2;
207 int len = len1 < len2 ? len1 : len2;
208 int cmp;
210 cmp = memcmp(name1, name2, len);
211 if (cmp)
212 return cmp;
213 c1 = name1[len];
214 c2 = name2[len];
215 if (!c1 && S_ISDIR(mode1))
216 c1 = '/';
217 if (!c2 && S_ISDIR(mode2))
218 c2 = '/';
219 return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
220 }
222 int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
223 {
224 int len1 = flags1 & CE_NAMEMASK;
225 int len2 = flags2 & CE_NAMEMASK;
226 int len = len1 < len2 ? len1 : len2;
227 int cmp;
229 cmp = memcmp(name1, name2, len);
230 if (cmp)
231 return cmp;
232 if (len1 < len2)
233 return -1;
234 if (len1 > len2)
235 return 1;
236 if (flags1 < flags2)
237 return -1;
238 if (flags1 > flags2)
239 return 1;
240 return 0;
241 }
243 int cache_name_pos(const char *name, int namelen)
244 {
245 int first, last;
247 first = 0;
248 last = active_nr;
249 while (last > first) {
250 int next = (last + first) >> 1;
251 struct cache_entry *ce = active_cache[next];
252 int cmp = cache_name_compare(name, namelen, ce->name, ntohs(ce->ce_flags));
253 if (!cmp)
254 return next;
255 if (cmp < 0) {
256 last = next;
257 continue;
258 }
259 first = next+1;
260 }
261 return -first-1;
262 }
264 /* Remove entry, return true if there are more entries to go.. */
265 int remove_cache_entry_at(int pos)
266 {
267 active_cache_changed = 1;
268 active_nr--;
269 if (pos >= active_nr)
270 return 0;
271 memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
272 return 1;
273 }
275 int remove_file_from_cache(const char *path)
276 {
277 int pos = cache_name_pos(path, strlen(path));
278 if (pos < 0)
279 pos = -pos-1;
280 while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
281 remove_cache_entry_at(pos);
282 return 0;
283 }
285 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
286 {
287 int len = ce_namelen(a);
288 return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
289 }
291 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
292 {
293 const char *match, *name;
294 int len;
296 if (!pathspec)
297 return 1;
299 len = ce_namelen(ce);
300 name = ce->name;
301 while ((match = *pathspec++) != NULL) {
302 int matchlen = strlen(match);
303 if (matchlen > len)
304 continue;
305 if (memcmp(name, match, matchlen))
306 continue;
307 if (matchlen && name[matchlen-1] == '/')
308 return 1;
309 if (name[matchlen] == '/' || !name[matchlen])
310 return 1;
311 if (!matchlen)
312 return 1;
313 }
314 return 0;
315 }
317 /*
318 * Do we have another file that has the beginning components being a
319 * proper superset of the name we're trying to add?
320 */
321 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
322 {
323 int retval = 0;
324 int len = ce_namelen(ce);
325 int stage = ce_stage(ce);
326 const char *name = ce->name;
328 while (pos < active_nr) {
329 struct cache_entry *p = active_cache[pos++];
331 if (len >= ce_namelen(p))
332 break;
333 if (memcmp(name, p->name, len))
334 break;
335 if (ce_stage(p) != stage)
336 continue;
337 if (p->name[len] != '/')
338 continue;
339 retval = -1;
340 if (!ok_to_replace)
341 break;
342 remove_cache_entry_at(--pos);
343 }
344 return retval;
345 }
347 /*
348 * Do we have another file with a pathname that is a proper
349 * subset of the name we're trying to add?
350 */
351 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
352 {
353 int retval = 0;
354 int stage = ce_stage(ce);
355 const char *name = ce->name;
356 const char *slash = name + ce_namelen(ce);
358 for (;;) {
359 int len;
361 for (;;) {
362 if (*--slash == '/')
363 break;
364 if (slash <= ce->name)
365 return retval;
366 }
367 len = slash - name;
369 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
370 if (pos >= 0) {
371 retval = -1;
372 if (ok_to_replace)
373 break;
374 remove_cache_entry_at(pos);
375 continue;
376 }
378 /*
379 * Trivial optimization: if we find an entry that
380 * already matches the sub-directory, then we know
381 * we're ok, and we can exit.
382 */
383 pos = -pos-1;
384 while (pos < active_nr) {
385 struct cache_entry *p = active_cache[pos];
386 if ((ce_namelen(p) <= len) ||
387 (p->name[len] != '/') ||
388 memcmp(p->name, name, len))
389 break; /* not our subdirectory */
390 if (ce_stage(p) == stage)
391 /* p is at the same stage as our entry, and
392 * is a subdirectory of what we are looking
393 * at, so we cannot have conflicts at our
394 * level or anything shorter.
395 */
396 return retval;
397 pos++;
398 }
399 }
400 return retval;
401 }
403 /* We may be in a situation where we already have path/file and path
404 * is being added, or we already have path and path/file is being
405 * added. Either one would result in a nonsense tree that has path
406 * twice when git-write-tree tries to write it out. Prevent it.
407 *
408 * If ok-to-replace is specified, we remove the conflicting entries
409 * from the cache so the caller should recompute the insert position.
410 * When this happens, we return non-zero.
411 */
412 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
413 {
414 /*
415 * We check if the path is a sub-path of a subsequent pathname
416 * first, since removing those will not change the position
417 * in the array
418 */
419 int retval = has_file_name(ce, pos, ok_to_replace);
420 /*
421 * Then check if the path might have a clashing sub-directory
422 * before it.
423 */
424 return retval + has_dir_name(ce, pos, ok_to_replace);
425 }
427 int add_cache_entry(struct cache_entry *ce, int option)
428 {
429 int pos;
430 int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
431 int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
432 int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
433 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
435 /* existing match? Just replace it. */
436 if (pos >= 0) {
437 active_cache_changed = 1;
438 active_cache[pos] = ce;
439 return 0;
440 }
441 pos = -pos-1;
443 /*
444 * Inserting a merged entry ("stage 0") into the index
445 * will always replace all non-merged entries..
446 */
447 if (pos < active_nr && ce_stage(ce) == 0) {
448 while (ce_same_name(active_cache[pos], ce)) {
449 ok_to_add = 1;
450 if (!remove_cache_entry_at(pos))
451 break;
452 }
453 }
455 if (!ok_to_add)
456 return -1;
458 if (!skip_df_check &&
459 check_file_directory_conflict(ce, pos, ok_to_replace)) {
460 if (!ok_to_replace)
461 return -1;
462 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
463 pos = -pos-1;
464 }
466 /* Make sure the array is big enough .. */
467 if (active_nr == active_alloc) {
468 active_alloc = alloc_nr(active_alloc);
469 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
470 }
472 /* Add it in.. */
473 active_nr++;
474 if (active_nr > pos)
475 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
476 active_cache[pos] = ce;
477 active_cache_changed = 1;
478 return 0;
479 }
481 static int verify_hdr(struct cache_header *hdr, unsigned long size)
482 {
483 SHA_CTX c;
484 unsigned char sha1[20];
486 if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
487 return error("bad signature");
488 if (hdr->hdr_version != htonl(2))
489 return error("bad index version");
490 SHA1_Init(&c);
491 SHA1_Update(&c, hdr, size - 20);
492 SHA1_Final(sha1, &c);
493 if (memcmp(sha1, (void *)hdr + size - 20, 20))
494 return error("bad index file sha1 signature");
495 return 0;
496 }
498 int read_cache(void)
499 {
500 int fd, i;
501 struct stat st;
502 unsigned long size, offset;
503 void *map;
504 struct cache_header *hdr;
506 errno = EBUSY;
507 if (active_cache)
508 return active_nr;
510 errno = ENOENT;
511 index_file_timestamp = 0;
512 fd = open(get_index_file(), O_RDONLY);
513 if (fd < 0) {
514 if (errno == ENOENT)
515 return 0;
516 die("index file open failed (%s)", strerror(errno));
517 }
519 size = 0; // avoid gcc warning
520 map = MAP_FAILED;
521 if (!fstat(fd, &st)) {
522 size = st.st_size;
523 errno = EINVAL;
524 if (size >= sizeof(struct cache_header) + 20)
525 map = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
526 }
527 close(fd);
528 if (map == MAP_FAILED)
529 die("index file mmap failed (%s)", strerror(errno));
531 hdr = map;
532 if (verify_hdr(hdr, size) < 0)
533 goto unmap;
535 active_nr = ntohl(hdr->hdr_entries);
536 active_alloc = alloc_nr(active_nr);
537 active_cache = calloc(active_alloc, sizeof(struct cache_entry *));
539 offset = sizeof(*hdr);
540 for (i = 0; i < active_nr; i++) {
541 struct cache_entry *ce = map + offset;
542 offset = offset + ce_size(ce);
543 active_cache[i] = ce;
544 }
545 index_file_timestamp = st.st_mtime;
546 return active_nr;
548 unmap:
549 munmap(map, size);
550 errno = EINVAL;
551 die("index file corrupt");
552 }
554 #define WRITE_BUFFER_SIZE 8192
555 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
556 static unsigned long write_buffer_len;
558 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
559 {
560 while (len) {
561 unsigned int buffered = write_buffer_len;
562 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
563 if (partial > len)
564 partial = len;
565 memcpy(write_buffer + buffered, data, partial);
566 buffered += partial;
567 if (buffered == WRITE_BUFFER_SIZE) {
568 SHA1_Update(context, write_buffer, WRITE_BUFFER_SIZE);
569 if (write(fd, write_buffer, WRITE_BUFFER_SIZE) != WRITE_BUFFER_SIZE)
570 return -1;
571 buffered = 0;
572 }
573 write_buffer_len = buffered;
574 len -= partial;
575 data += partial;
576 }
577 return 0;
578 }
580 static int ce_flush(SHA_CTX *context, int fd)
581 {
582 unsigned int left = write_buffer_len;
584 if (left) {
585 write_buffer_len = 0;
586 SHA1_Update(context, write_buffer, left);
587 }
589 /* Flush first if not enough space for SHA1 signature */
590 if (left + 20 > WRITE_BUFFER_SIZE) {
591 if (write(fd, write_buffer, left) != left)
592 return -1;
593 left = 0;
594 }
596 /* Append the SHA1 signature at the end */
597 SHA1_Final(write_buffer + left, context);
598 left += 20;
599 if (write(fd, write_buffer, left) != left)
600 return -1;
601 return 0;
602 }
604 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
605 {
606 /*
607 * The only thing we care about in this function is to smudge the
608 * falsely clean entry due to touch-update-touch race, so we leave
609 * everything else as they are. We are called for entries whose
610 * ce_mtime match the index file mtime.
611 */
612 struct stat st;
614 if (lstat(ce->name, &st) < 0)
615 return;
616 if (ce_match_stat_basic(ce, &st))
617 return;
618 if (ce_modified_check_fs(ce, &st)) {
619 /* This is "racily clean"; smudge it. Note that this
620 * is a tricky code. At first glance, it may appear
621 * that it can break with this sequence:
622 *
623 * $ echo xyzzy >frotz
624 * $ git-update-index --add frotz
625 * $ : >frotz
626 * $ sleep 3
627 * $ echo filfre >nitfol
628 * $ git-update-index --add nitfol
629 *
630 * but it does not. Whe the second update-index runs,
631 * it notices that the entry "frotz" has the same timestamp
632 * as index, and if we were to smudge it by resetting its
633 * size to zero here, then the object name recorded
634 * in index is the 6-byte file but the cached stat information
635 * becomes zero --- which would then match what we would
636 * obtain from the filesystem next time we stat("frotz").
637 *
638 * However, the second update-index, before calling
639 * this function, notices that the cached size is 6
640 * bytes and what is on the filesystem is an empty
641 * file, and never calls us, so the cached size information
642 * for "frotz" stays 6 which does not match the filesystem.
643 */
644 ce->ce_size = htonl(0);
645 }
646 }
648 int write_cache(int newfd, struct cache_entry **cache, int entries)
649 {
650 SHA_CTX c;
651 struct cache_header hdr;
652 int i, removed;
654 for (i = removed = 0; i < entries; i++)
655 if (!cache[i]->ce_mode)
656 removed++;
658 hdr.hdr_signature = htonl(CACHE_SIGNATURE);
659 hdr.hdr_version = htonl(2);
660 hdr.hdr_entries = htonl(entries - removed);
662 SHA1_Init(&c);
663 if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
664 return -1;
666 for (i = 0; i < entries; i++) {
667 struct cache_entry *ce = cache[i];
668 if (!ce->ce_mode)
669 continue;
670 if (index_file_timestamp &&
671 index_file_timestamp <= ntohl(ce->ce_mtime.sec))
672 ce_smudge_racily_clean_entry(ce);
673 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
674 return -1;
675 }
676 return ce_flush(&c, newfd);
677 }