remove unnecessary initializations
[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 #include <time.h>
9
10 /* Index extensions.
11 *
12 * The first letter should be 'A'..'Z' for extensions that are not
13 * necessary for a correct operation (i.e. optimization data).
14 * When new extensions are added that _needs_ to be understood in
15 * order to correctly interpret the index file, pick character that
16 * is outside the range, to cause the reader to abort.
17 */
18
19 #define CACHE_EXT(s) ( (s[0]<<24)|(s[1]<<16)|(s[2]<<8)|(s[3]) )
20 #define CACHE_EXT_TREE 0x54524545 /* "TREE" */
21
22 struct cache_entry **active_cache;
23 static time_t index_file_timestamp;
24 unsigned int active_nr, active_alloc, active_cache_changed;
25
26 struct cache_tree *active_cache_tree;
27
28 int cache_errno;
29
30 static void *cache_mmap;
31 static size_t cache_mmap_size;
32
33 /*
34 * This only updates the "non-critical" parts of the directory
35 * cache, ie the parts that aren't tracked by GIT, and only used
36 * to validate the cache.
37 */
38 void fill_stat_cache_info(struct cache_entry *ce, struct stat *st)
39 {
40 ce->ce_ctime.sec = htonl(st->st_ctime);
41 ce->ce_mtime.sec = htonl(st->st_mtime);
42 #ifdef USE_NSEC
43 ce->ce_ctime.nsec = htonl(st->st_ctim.tv_nsec);
44 ce->ce_mtime.nsec = htonl(st->st_mtim.tv_nsec);
45 #endif
46 ce->ce_dev = htonl(st->st_dev);
47 ce->ce_ino = htonl(st->st_ino);
48 ce->ce_uid = htonl(st->st_uid);
49 ce->ce_gid = htonl(st->st_gid);
50 ce->ce_size = htonl(st->st_size);
51
52 if (assume_unchanged)
53 ce->ce_flags |= htons(CE_VALID);
54 }
55
56 static int ce_compare_data(struct cache_entry *ce, struct stat *st)
57 {
58 int match = -1;
59 int fd = open(ce->name, O_RDONLY);
60
61 if (fd >= 0) {
62 unsigned char sha1[20];
63 if (!index_fd(sha1, fd, st, 0, NULL))
64 match = memcmp(sha1, ce->sha1, 20);
65 /* index_fd() closed the file descriptor already */
66 }
67 return match;
68 }
69
70 static int ce_compare_link(struct cache_entry *ce, unsigned long expected_size)
71 {
72 int match = -1;
73 char *target;
74 void *buffer;
75 unsigned long size;
76 char type[10];
77 int len;
78
79 target = xmalloc(expected_size);
80 len = readlink(ce->name, target, expected_size);
81 if (len != expected_size) {
82 free(target);
83 return -1;
84 }
85 buffer = read_sha1_file(ce->sha1, type, &size);
86 if (!buffer) {
87 free(target);
88 return -1;
89 }
90 if (size == expected_size)
91 match = memcmp(buffer, target, size);
92 free(buffer);
93 free(target);
94 return match;
95 }
96
97 static int ce_modified_check_fs(struct cache_entry *ce, struct stat *st)
98 {
99 switch (st->st_mode & S_IFMT) {
100 case S_IFREG:
101 if (ce_compare_data(ce, st))
102 return DATA_CHANGED;
103 break;
104 case S_IFLNK:
105 if (ce_compare_link(ce, st->st_size))
106 return DATA_CHANGED;
107 break;
108 default:
109 return TYPE_CHANGED;
110 }
111 return 0;
112 }
113
114 static int ce_match_stat_basic(struct cache_entry *ce, struct stat *st)
115 {
116 unsigned int changed = 0;
117
118 switch (ntohl(ce->ce_mode) & S_IFMT) {
119 case S_IFREG:
120 changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
121 /* We consider only the owner x bit to be relevant for
122 * "mode changes"
123 */
124 if (trust_executable_bit &&
125 (0100 & (ntohl(ce->ce_mode) ^ st->st_mode)))
126 changed |= MODE_CHANGED;
127 break;
128 case S_IFLNK:
129 changed |= !S_ISLNK(st->st_mode) ? TYPE_CHANGED : 0;
130 break;
131 default:
132 die("internal error: ce_mode is %o", ntohl(ce->ce_mode));
133 }
134 if (ce->ce_mtime.sec != htonl(st->st_mtime))
135 changed |= MTIME_CHANGED;
136 if (ce->ce_ctime.sec != htonl(st->st_ctime))
137 changed |= CTIME_CHANGED;
138
139 #ifdef USE_NSEC
140 /*
141 * nsec seems unreliable - not all filesystems support it, so
142 * as long as it is in the inode cache you get right nsec
143 * but after it gets flushed, you get zero nsec.
144 */
145 if (ce->ce_mtime.nsec != htonl(st->st_mtim.tv_nsec))
146 changed |= MTIME_CHANGED;
147 if (ce->ce_ctime.nsec != htonl(st->st_ctim.tv_nsec))
148 changed |= CTIME_CHANGED;
149 #endif
150
151 if (ce->ce_uid != htonl(st->st_uid) ||
152 ce->ce_gid != htonl(st->st_gid))
153 changed |= OWNER_CHANGED;
154 if (ce->ce_ino != htonl(st->st_ino))
155 changed |= INODE_CHANGED;
156
157 #ifdef USE_STDEV
158 /*
159 * st_dev breaks on network filesystems where different
160 * clients will have different views of what "device"
161 * the filesystem is on
162 */
163 if (ce->ce_dev != htonl(st->st_dev))
164 changed |= INODE_CHANGED;
165 #endif
166
167 if (ce->ce_size != htonl(st->st_size))
168 changed |= DATA_CHANGED;
169
170 return changed;
171 }
172
173 int ce_match_stat(struct cache_entry *ce, struct stat *st, int ignore_valid)
174 {
175 unsigned int changed;
176
177 /*
178 * If it's marked as always valid in the index, it's
179 * valid whatever the checked-out copy says.
180 */
181 if (!ignore_valid && (ce->ce_flags & htons(CE_VALID)))
182 return 0;
183
184 changed = ce_match_stat_basic(ce, st);
185
186 /*
187 * Within 1 second of this sequence:
188 * echo xyzzy >file && git-update-index --add file
189 * running this command:
190 * echo frotz >file
191 * would give a falsely clean cache entry. The mtime and
192 * length match the cache, and other stat fields do not change.
193 *
194 * We could detect this at update-index time (the cache entry
195 * being registered/updated records the same time as "now")
196 * and delay the return from git-update-index, but that would
197 * effectively mean we can make at most one commit per second,
198 * which is not acceptable. Instead, we check cache entries
199 * whose mtime are the same as the index file timestamp more
200 * carefully than others.
201 */
202 if (!changed &&
203 index_file_timestamp &&
204 index_file_timestamp <= ntohl(ce->ce_mtime.sec))
205 changed |= ce_modified_check_fs(ce, st);
206
207 return changed;
208 }
209
210 int ce_modified(struct cache_entry *ce, struct stat *st, int really)
211 {
212 int changed, changed_fs;
213 changed = ce_match_stat(ce, st, really);
214 if (!changed)
215 return 0;
216 /*
217 * If the mode or type has changed, there's no point in trying
218 * to refresh the entry - it's not going to match
219 */
220 if (changed & (MODE_CHANGED | TYPE_CHANGED))
221 return changed;
222
223 /* Immediately after read-tree or update-index --cacheinfo,
224 * the length field is zero. For other cases the ce_size
225 * should match the SHA1 recorded in the index entry.
226 */
227 if ((changed & DATA_CHANGED) && ce->ce_size != htonl(0))
228 return changed;
229
230 changed_fs = ce_modified_check_fs(ce, st);
231 if (changed_fs)
232 return changed | changed_fs;
233 return 0;
234 }
235
236 int base_name_compare(const char *name1, int len1, int mode1,
237 const char *name2, int len2, int mode2)
238 {
239 unsigned char c1, c2;
240 int len = len1 < len2 ? len1 : len2;
241 int cmp;
242
243 cmp = memcmp(name1, name2, len);
244 if (cmp)
245 return cmp;
246 c1 = name1[len];
247 c2 = name2[len];
248 if (!c1 && S_ISDIR(mode1))
249 c1 = '/';
250 if (!c2 && S_ISDIR(mode2))
251 c2 = '/';
252 return (c1 < c2) ? -1 : (c1 > c2) ? 1 : 0;
253 }
254
255 int cache_name_compare(const char *name1, int flags1, const char *name2, int flags2)
256 {
257 int len1 = flags1 & CE_NAMEMASK;
258 int len2 = flags2 & CE_NAMEMASK;
259 int len = len1 < len2 ? len1 : len2;
260 int cmp;
261
262 cmp = memcmp(name1, name2, len);
263 if (cmp)
264 return cmp;
265 if (len1 < len2)
266 return -1;
267 if (len1 > len2)
268 return 1;
269
270 /* Compare stages */
271 flags1 &= CE_STAGEMASK;
272 flags2 &= CE_STAGEMASK;
273
274 if (flags1 < flags2)
275 return -1;
276 if (flags1 > flags2)
277 return 1;
278 return 0;
279 }
280
281 int cache_name_pos(const char *name, int namelen)
282 {
283 int first, last;
284
285 first = 0;
286 last = active_nr;
287 while (last > first) {
288 int next = (last + first) >> 1;
289 struct cache_entry *ce = active_cache[next];
290 int cmp = cache_name_compare(name, namelen, ce->name, ntohs(ce->ce_flags));
291 if (!cmp)
292 return next;
293 if (cmp < 0) {
294 last = next;
295 continue;
296 }
297 first = next+1;
298 }
299 return -first-1;
300 }
301
302 /* Remove entry, return true if there are more entries to go.. */
303 int remove_cache_entry_at(int pos)
304 {
305 active_cache_changed = 1;
306 active_nr--;
307 if (pos >= active_nr)
308 return 0;
309 memmove(active_cache + pos, active_cache + pos + 1, (active_nr - pos) * sizeof(struct cache_entry *));
310 return 1;
311 }
312
313 int remove_file_from_cache(const char *path)
314 {
315 int pos = cache_name_pos(path, strlen(path));
316 if (pos < 0)
317 pos = -pos-1;
318 while (pos < active_nr && !strcmp(active_cache[pos]->name, path))
319 remove_cache_entry_at(pos);
320 return 0;
321 }
322
323 int add_file_to_index(const char *path, int verbose)
324 {
325 int size, namelen;
326 struct stat st;
327 struct cache_entry *ce;
328
329 if (lstat(path, &st))
330 die("%s: unable to stat (%s)", path, strerror(errno));
331
332 if (!S_ISREG(st.st_mode) && !S_ISLNK(st.st_mode))
333 die("%s: can only add regular files or symbolic links", path);
334
335 namelen = strlen(path);
336 size = cache_entry_size(namelen);
337 ce = xcalloc(1, size);
338 memcpy(ce->name, path, namelen);
339 ce->ce_flags = htons(namelen);
340 fill_stat_cache_info(ce, &st);
341
342 ce->ce_mode = create_ce_mode(st.st_mode);
343 if (!trust_executable_bit) {
344 /* If there is an existing entry, pick the mode bits
345 * from it.
346 */
347 int pos = cache_name_pos(path, namelen);
348 if (pos >= 0)
349 ce->ce_mode = active_cache[pos]->ce_mode;
350 }
351
352 if (index_path(ce->sha1, path, &st, 1))
353 die("unable to index file %s", path);
354 if (add_cache_entry(ce, ADD_CACHE_OK_TO_ADD))
355 die("unable to add %s to index",path);
356 if (verbose)
357 printf("add '%s'\n", path);
358 cache_tree_invalidate_path(active_cache_tree, path);
359 return 0;
360 }
361
362 int ce_same_name(struct cache_entry *a, struct cache_entry *b)
363 {
364 int len = ce_namelen(a);
365 return ce_namelen(b) == len && !memcmp(a->name, b->name, len);
366 }
367
368 int ce_path_match(const struct cache_entry *ce, const char **pathspec)
369 {
370 const char *match, *name;
371 int len;
372
373 if (!pathspec)
374 return 1;
375
376 len = ce_namelen(ce);
377 name = ce->name;
378 while ((match = *pathspec++) != NULL) {
379 int matchlen = strlen(match);
380 if (matchlen > len)
381 continue;
382 if (memcmp(name, match, matchlen))
383 continue;
384 if (matchlen && name[matchlen-1] == '/')
385 return 1;
386 if (name[matchlen] == '/' || !name[matchlen])
387 return 1;
388 if (!matchlen)
389 return 1;
390 }
391 return 0;
392 }
393
394 /*
395 * We fundamentally don't like some paths: we don't want
396 * dot or dot-dot anywhere, and for obvious reasons don't
397 * want to recurse into ".git" either.
398 *
399 * Also, we don't want double slashes or slashes at the
400 * end that can make pathnames ambiguous.
401 */
402 static int verify_dotfile(const char *rest)
403 {
404 /*
405 * The first character was '.', but that
406 * has already been discarded, we now test
407 * the rest.
408 */
409 switch (*rest) {
410 /* "." is not allowed */
411 case '\0': case '/':
412 return 0;
413
414 /*
415 * ".git" followed by NUL or slash is bad. This
416 * shares the path end test with the ".." case.
417 */
418 case 'g':
419 if (rest[1] != 'i')
420 break;
421 if (rest[2] != 't')
422 break;
423 rest += 2;
424 /* fallthrough */
425 case '.':
426 if (rest[1] == '\0' || rest[1] == '/')
427 return 0;
428 }
429 return 1;
430 }
431
432 int verify_path(const char *path)
433 {
434 char c;
435
436 goto inside;
437 for (;;) {
438 if (!c)
439 return 1;
440 if (c == '/') {
441 inside:
442 c = *path++;
443 switch (c) {
444 default:
445 continue;
446 case '/': case '\0':
447 break;
448 case '.':
449 if (verify_dotfile(path))
450 continue;
451 }
452 return 0;
453 }
454 c = *path++;
455 }
456 }
457
458 /*
459 * Do we have another file that has the beginning components being a
460 * proper superset of the name we're trying to add?
461 */
462 static int has_file_name(const struct cache_entry *ce, int pos, int ok_to_replace)
463 {
464 int retval = 0;
465 int len = ce_namelen(ce);
466 int stage = ce_stage(ce);
467 const char *name = ce->name;
468
469 while (pos < active_nr) {
470 struct cache_entry *p = active_cache[pos++];
471
472 if (len >= ce_namelen(p))
473 break;
474 if (memcmp(name, p->name, len))
475 break;
476 if (ce_stage(p) != stage)
477 continue;
478 if (p->name[len] != '/')
479 continue;
480 retval = -1;
481 if (!ok_to_replace)
482 break;
483 remove_cache_entry_at(--pos);
484 }
485 return retval;
486 }
487
488 /*
489 * Do we have another file with a pathname that is a proper
490 * subset of the name we're trying to add?
491 */
492 static int has_dir_name(const struct cache_entry *ce, int pos, int ok_to_replace)
493 {
494 int retval = 0;
495 int stage = ce_stage(ce);
496 const char *name = ce->name;
497 const char *slash = name + ce_namelen(ce);
498
499 for (;;) {
500 int len;
501
502 for (;;) {
503 if (*--slash == '/')
504 break;
505 if (slash <= ce->name)
506 return retval;
507 }
508 len = slash - name;
509
510 pos = cache_name_pos(name, ntohs(create_ce_flags(len, stage)));
511 if (pos >= 0) {
512 retval = -1;
513 if (ok_to_replace)
514 break;
515 remove_cache_entry_at(pos);
516 continue;
517 }
518
519 /*
520 * Trivial optimization: if we find an entry that
521 * already matches the sub-directory, then we know
522 * we're ok, and we can exit.
523 */
524 pos = -pos-1;
525 while (pos < active_nr) {
526 struct cache_entry *p = active_cache[pos];
527 if ((ce_namelen(p) <= len) ||
528 (p->name[len] != '/') ||
529 memcmp(p->name, name, len))
530 break; /* not our subdirectory */
531 if (ce_stage(p) == stage)
532 /* p is at the same stage as our entry, and
533 * is a subdirectory of what we are looking
534 * at, so we cannot have conflicts at our
535 * level or anything shorter.
536 */
537 return retval;
538 pos++;
539 }
540 }
541 return retval;
542 }
543
544 /* We may be in a situation where we already have path/file and path
545 * is being added, or we already have path and path/file is being
546 * added. Either one would result in a nonsense tree that has path
547 * twice when git-write-tree tries to write it out. Prevent it.
548 *
549 * If ok-to-replace is specified, we remove the conflicting entries
550 * from the cache so the caller should recompute the insert position.
551 * When this happens, we return non-zero.
552 */
553 static int check_file_directory_conflict(const struct cache_entry *ce, int pos, int ok_to_replace)
554 {
555 /*
556 * We check if the path is a sub-path of a subsequent pathname
557 * first, since removing those will not change the position
558 * in the array
559 */
560 int retval = has_file_name(ce, pos, ok_to_replace);
561 /*
562 * Then check if the path might have a clashing sub-directory
563 * before it.
564 */
565 return retval + has_dir_name(ce, pos, ok_to_replace);
566 }
567
568 int add_cache_entry(struct cache_entry *ce, int option)
569 {
570 int pos;
571 int ok_to_add = option & ADD_CACHE_OK_TO_ADD;
572 int ok_to_replace = option & ADD_CACHE_OK_TO_REPLACE;
573 int skip_df_check = option & ADD_CACHE_SKIP_DFCHECK;
574
575 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
576
577 /* existing match? Just replace it. */
578 if (pos >= 0) {
579 active_cache_changed = 1;
580 active_cache[pos] = ce;
581 return 0;
582 }
583 pos = -pos-1;
584
585 /*
586 * Inserting a merged entry ("stage 0") into the index
587 * will always replace all non-merged entries..
588 */
589 if (pos < active_nr && ce_stage(ce) == 0) {
590 while (ce_same_name(active_cache[pos], ce)) {
591 ok_to_add = 1;
592 if (!remove_cache_entry_at(pos))
593 break;
594 }
595 }
596
597 if (!ok_to_add)
598 return -1;
599 if (!verify_path(ce->name))
600 return -1;
601
602 if (!skip_df_check &&
603 check_file_directory_conflict(ce, pos, ok_to_replace)) {
604 if (!ok_to_replace)
605 return -1;
606 pos = cache_name_pos(ce->name, ntohs(ce->ce_flags));
607 pos = -pos-1;
608 }
609
610 /* Make sure the array is big enough .. */
611 if (active_nr == active_alloc) {
612 active_alloc = alloc_nr(active_alloc);
613 active_cache = xrealloc(active_cache, active_alloc * sizeof(struct cache_entry *));
614 }
615
616 /* Add it in.. */
617 active_nr++;
618 if (active_nr > pos)
619 memmove(active_cache + pos + 1, active_cache + pos, (active_nr - pos - 1) * sizeof(ce));
620 active_cache[pos] = ce;
621 active_cache_changed = 1;
622 return 0;
623 }
624
625 /*
626 * "refresh" does not calculate a new sha1 file or bring the
627 * cache up-to-date for mode/content changes. But what it
628 * _does_ do is to "re-match" the stat information of a file
629 * with the cache, so that you can refresh the cache for a
630 * file that hasn't been changed but where the stat entry is
631 * out of date.
632 *
633 * For example, you'd want to do this after doing a "git-read-tree",
634 * to link up the stat cache details with the proper files.
635 */
636 struct cache_entry *refresh_cache_entry(struct cache_entry *ce, int really)
637 {
638 struct stat st;
639 struct cache_entry *updated;
640 int changed, size;
641
642 if (lstat(ce->name, &st) < 0) {
643 cache_errno = errno;
644 return NULL;
645 }
646
647 changed = ce_match_stat(ce, &st, really);
648 if (!changed) {
649 if (really && assume_unchanged &&
650 !(ce->ce_flags & htons(CE_VALID)))
651 ; /* mark this one VALID again */
652 else
653 return ce;
654 }
655
656 if (ce_modified(ce, &st, really)) {
657 cache_errno = EINVAL;
658 return NULL;
659 }
660
661 size = ce_size(ce);
662 updated = xmalloc(size);
663 memcpy(updated, ce, size);
664 fill_stat_cache_info(updated, &st);
665
666 /* In this case, if really is not set, we should leave
667 * CE_VALID bit alone. Otherwise, paths marked with
668 * --no-assume-unchanged (i.e. things to be edited) will
669 * reacquire CE_VALID bit automatically, which is not
670 * really what we want.
671 */
672 if (!really && assume_unchanged && !(ce->ce_flags & htons(CE_VALID)))
673 updated->ce_flags &= ~htons(CE_VALID);
674
675 return updated;
676 }
677
678 int refresh_cache(unsigned int flags)
679 {
680 int i;
681 int has_errors = 0;
682 int really = (flags & REFRESH_REALLY) != 0;
683 int allow_unmerged = (flags & REFRESH_UNMERGED) != 0;
684 int quiet = (flags & REFRESH_QUIET) != 0;
685 int not_new = (flags & REFRESH_IGNORE_MISSING) != 0;
686
687 for (i = 0; i < active_nr; i++) {
688 struct cache_entry *ce, *new;
689 ce = active_cache[i];
690 if (ce_stage(ce)) {
691 while ((i < active_nr) &&
692 ! strcmp(active_cache[i]->name, ce->name))
693 i++;
694 i--;
695 if (allow_unmerged)
696 continue;
697 printf("%s: needs merge\n", ce->name);
698 has_errors = 1;
699 continue;
700 }
701
702 new = refresh_cache_entry(ce, really);
703 if (new == ce)
704 continue;
705 if (!new) {
706 if (not_new && cache_errno == ENOENT)
707 continue;
708 if (really && cache_errno == EINVAL) {
709 /* If we are doing --really-refresh that
710 * means the index is not valid anymore.
711 */
712 ce->ce_flags &= ~htons(CE_VALID);
713 active_cache_changed = 1;
714 }
715 if (quiet)
716 continue;
717 printf("%s: needs update\n", ce->name);
718 has_errors = 1;
719 continue;
720 }
721 active_cache_changed = 1;
722 /* You can NOT just free active_cache[i] here, since it
723 * might not be necessarily malloc()ed but can also come
724 * from mmap(). */
725 active_cache[i] = new;
726 }
727 return has_errors;
728 }
729
730 static int verify_hdr(struct cache_header *hdr, unsigned long size)
731 {
732 SHA_CTX c;
733 unsigned char sha1[20];
734
735 if (hdr->hdr_signature != htonl(CACHE_SIGNATURE))
736 return error("bad signature");
737 if (hdr->hdr_version != htonl(2))
738 return error("bad index version");
739 SHA1_Init(&c);
740 SHA1_Update(&c, hdr, size - 20);
741 SHA1_Final(sha1, &c);
742 if (memcmp(sha1, (char *) hdr + size - 20, 20))
743 return error("bad index file sha1 signature");
744 return 0;
745 }
746
747 static int read_index_extension(const char *ext, void *data, unsigned long sz)
748 {
749 switch (CACHE_EXT(ext)) {
750 case CACHE_EXT_TREE:
751 active_cache_tree = cache_tree_read(data, sz);
752 break;
753 default:
754 if (*ext < 'A' || 'Z' < *ext)
755 return error("index uses %.4s extension, which we do not understand",
756 ext);
757 fprintf(stderr, "ignoring %.4s extension\n", ext);
758 break;
759 }
760 return 0;
761 }
762
763 int read_cache(void)
764 {
765 return read_cache_from(get_index_file());
766 }
767
768 /* remember to discard_cache() before reading a different cache! */
769 int read_cache_from(const char *path)
770 {
771 int fd, i;
772 struct stat st;
773 unsigned long offset;
774 struct cache_header *hdr;
775
776 errno = EBUSY;
777 if (cache_mmap)
778 return active_nr;
779
780 errno = ENOENT;
781 index_file_timestamp = 0;
782 fd = open(path, O_RDONLY);
783 if (fd < 0) {
784 if (errno == ENOENT)
785 return 0;
786 die("index file open failed (%s)", strerror(errno));
787 }
788
789 cache_mmap = MAP_FAILED;
790 if (!fstat(fd, &st)) {
791 cache_mmap_size = st.st_size;
792 errno = EINVAL;
793 if (cache_mmap_size >= sizeof(struct cache_header) + 20)
794 cache_mmap = mmap(NULL, cache_mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
795 }
796 close(fd);
797 if (cache_mmap == MAP_FAILED)
798 die("index file mmap failed (%s)", strerror(errno));
799
800 hdr = cache_mmap;
801 if (verify_hdr(hdr, cache_mmap_size) < 0)
802 goto unmap;
803
804 active_nr = ntohl(hdr->hdr_entries);
805 active_alloc = alloc_nr(active_nr);
806 active_cache = xcalloc(active_alloc, sizeof(struct cache_entry *));
807
808 offset = sizeof(*hdr);
809 for (i = 0; i < active_nr; i++) {
810 struct cache_entry *ce = (struct cache_entry *) ((char *) cache_mmap + offset);
811 offset = offset + ce_size(ce);
812 active_cache[i] = ce;
813 }
814 index_file_timestamp = st.st_mtime;
815 while (offset <= cache_mmap_size - 20 - 8) {
816 /* After an array of active_nr index entries,
817 * there can be arbitrary number of extended
818 * sections, each of which is prefixed with
819 * extension name (4-byte) and section length
820 * in 4-byte network byte order.
821 */
822 unsigned long extsize;
823 memcpy(&extsize, (char *) cache_mmap + offset + 4, 4);
824 extsize = ntohl(extsize);
825 if (read_index_extension(((const char *) cache_mmap) + offset,
826 (char *) cache_mmap + offset + 8,
827 extsize) < 0)
828 goto unmap;
829 offset += 8;
830 offset += extsize;
831 }
832 return active_nr;
833
834 unmap:
835 munmap(cache_mmap, cache_mmap_size);
836 errno = EINVAL;
837 die("index file corrupt");
838 }
839
840 #define WRITE_BUFFER_SIZE 8192
841 static unsigned char write_buffer[WRITE_BUFFER_SIZE];
842 static unsigned long write_buffer_len;
843
844 static int ce_write_flush(SHA_CTX *context, int fd)
845 {
846 unsigned int buffered = write_buffer_len;
847 if (buffered) {
848 SHA1_Update(context, write_buffer, buffered);
849 if (write(fd, write_buffer, buffered) != buffered)
850 return -1;
851 write_buffer_len = 0;
852 }
853 return 0;
854 }
855
856 static int ce_write(SHA_CTX *context, int fd, void *data, unsigned int len)
857 {
858 while (len) {
859 unsigned int buffered = write_buffer_len;
860 unsigned int partial = WRITE_BUFFER_SIZE - buffered;
861 if (partial > len)
862 partial = len;
863 memcpy(write_buffer + buffered, data, partial);
864 buffered += partial;
865 if (buffered == WRITE_BUFFER_SIZE) {
866 write_buffer_len = buffered;
867 if (ce_write_flush(context, fd))
868 return -1;
869 buffered = 0;
870 }
871 write_buffer_len = buffered;
872 len -= partial;
873 data = (char *) data + partial;
874 }
875 return 0;
876 }
877
878 static int write_index_ext_header(SHA_CTX *context, int fd,
879 unsigned int ext, unsigned int sz)
880 {
881 ext = htonl(ext);
882 sz = htonl(sz);
883 return ((ce_write(context, fd, &ext, 4) < 0) ||
884 (ce_write(context, fd, &sz, 4) < 0)) ? -1 : 0;
885 }
886
887 static int ce_flush(SHA_CTX *context, int fd)
888 {
889 unsigned int left = write_buffer_len;
890
891 if (left) {
892 write_buffer_len = 0;
893 SHA1_Update(context, write_buffer, left);
894 }
895
896 /* Flush first if not enough space for SHA1 signature */
897 if (left + 20 > WRITE_BUFFER_SIZE) {
898 if (write(fd, write_buffer, left) != left)
899 return -1;
900 left = 0;
901 }
902
903 /* Append the SHA1 signature at the end */
904 SHA1_Final(write_buffer + left, context);
905 left += 20;
906 return (write(fd, write_buffer, left) != left) ? -1 : 0;
907 }
908
909 static void ce_smudge_racily_clean_entry(struct cache_entry *ce)
910 {
911 /*
912 * The only thing we care about in this function is to smudge the
913 * falsely clean entry due to touch-update-touch race, so we leave
914 * everything else as they are. We are called for entries whose
915 * ce_mtime match the index file mtime.
916 */
917 struct stat st;
918
919 if (lstat(ce->name, &st) < 0)
920 return;
921 if (ce_match_stat_basic(ce, &st))
922 return;
923 if (ce_modified_check_fs(ce, &st)) {
924 /* This is "racily clean"; smudge it. Note that this
925 * is a tricky code. At first glance, it may appear
926 * that it can break with this sequence:
927 *
928 * $ echo xyzzy >frotz
929 * $ git-update-index --add frotz
930 * $ : >frotz
931 * $ sleep 3
932 * $ echo filfre >nitfol
933 * $ git-update-index --add nitfol
934 *
935 * but it does not. When the second update-index runs,
936 * it notices that the entry "frotz" has the same timestamp
937 * as index, and if we were to smudge it by resetting its
938 * size to zero here, then the object name recorded
939 * in index is the 6-byte file but the cached stat information
940 * becomes zero --- which would then match what we would
941 * obtain from the filesystem next time we stat("frotz").
942 *
943 * However, the second update-index, before calling
944 * this function, notices that the cached size is 6
945 * bytes and what is on the filesystem is an empty
946 * file, and never calls us, so the cached size information
947 * for "frotz" stays 6 which does not match the filesystem.
948 */
949 ce->ce_size = htonl(0);
950 }
951 }
952
953 int write_cache(int newfd, struct cache_entry **cache, int entries)
954 {
955 SHA_CTX c;
956 struct cache_header hdr;
957 int i, removed, recent;
958 struct stat st;
959 time_t now;
960
961 for (i = removed = 0; i < entries; i++)
962 if (!cache[i]->ce_mode)
963 removed++;
964
965 hdr.hdr_signature = htonl(CACHE_SIGNATURE);
966 hdr.hdr_version = htonl(2);
967 hdr.hdr_entries = htonl(entries - removed);
968
969 SHA1_Init(&c);
970 if (ce_write(&c, newfd, &hdr, sizeof(hdr)) < 0)
971 return -1;
972
973 for (i = 0; i < entries; i++) {
974 struct cache_entry *ce = cache[i];
975 if (!ce->ce_mode)
976 continue;
977 if (index_file_timestamp &&
978 index_file_timestamp <= ntohl(ce->ce_mtime.sec))
979 ce_smudge_racily_clean_entry(ce);
980 if (ce_write(&c, newfd, ce, ce_size(ce)) < 0)
981 return -1;
982 }
983
984 /* Write extension data here */
985 if (active_cache_tree) {
986 unsigned long sz;
987 void *data = cache_tree_write(active_cache_tree, &sz);
988 if (data &&
989 !write_index_ext_header(&c, newfd, CACHE_EXT_TREE, sz) &&
990 !ce_write(&c, newfd, data, sz))
991 ;
992 else {
993 free(data);
994 return -1;
995 }
996 }
997
998 /*
999 * To prevent later ce_match_stat() from always falling into
1000 * check_fs(), if we have too many entries that can trigger
1001 * racily clean check, we are better off delaying the return.
1002 * We arbitrarily say if more than 20 paths or 25% of total
1003 * paths are very new, we delay the return until the index
1004 * file gets a new timestamp.
1005 *
1006 * NOTE! NOTE! NOTE!
1007 *
1008 * This assumes that nobody is touching the working tree while
1009 * we are updating the index.
1010 */
1011
1012 /* Make sure that the new index file has st_mtime
1013 * that is current enough -- ce_write() batches the data
1014 * so it might not have written anything yet.
1015 */
1016 ce_write_flush(&c, newfd);
1017
1018 now = fstat(newfd, &st) ? 0 : st.st_mtime;
1019 if (now) {
1020 recent = 0;
1021 for (i = 0; i < entries; i++) {
1022 struct cache_entry *ce = cache[i];
1023 time_t entry_time = (time_t) ntohl(ce->ce_mtime.sec);
1024 if (!ce->ce_mode)
1025 continue;
1026 if (now && now <= entry_time)
1027 recent++;
1028 }
1029 if (20 < recent && entries <= recent * 4) {
1030 #if 0
1031 fprintf(stderr, "entries %d\n", entries);
1032 fprintf(stderr, "recent %d\n", recent);
1033 fprintf(stderr, "now %lu\n", now);
1034 #endif
1035 while (!fstat(newfd, &st) && st.st_mtime <= now) {
1036 struct timespec rq, rm;
1037 off_t where = lseek(newfd, 0, SEEK_CUR);
1038 rq.tv_sec = 0;
1039 rq.tv_nsec = 250000000;
1040 nanosleep(&rq, &rm);
1041 if ((where == (off_t) -1) ||
1042 (write(newfd, "", 1) != 1) ||
1043 (lseek(newfd, -1, SEEK_CUR) != where) ||
1044 ftruncate(newfd, where))
1045 break;
1046 }
1047 }
1048 }
1049 return ce_flush(&c, newfd);
1050 }