Merge branch 'cb/t4201-robustify'
[git] / sha1_file.c
1 /*
2  * GIT - The information manager from hell
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This handles basic git sha1 object files - packing, unpacking,
7  * creation etc.
8  */
9 #include "cache.h"
10 #include "config.h"
11 #include "string-list.h"
12 #include "lockfile.h"
13 #include "delta.h"
14 #include "pack.h"
15 #include "blob.h"
16 #include "commit.h"
17 #include "run-command.h"
18 #include "tag.h"
19 #include "tree.h"
20 #include "tree-walk.h"
21 #include "refs.h"
22 #include "pack-revindex.h"
23 #include "sha1-lookup.h"
24 #include "bulk-checkin.h"
25 #include "streaming.h"
26 #include "dir.h"
27 #include "mru.h"
28 #include "list.h"
29 #include "mergesort.h"
30 #include "quote.h"
31 #include "packfile.h"
32
33 const unsigned char null_sha1[GIT_MAX_RAWSZ];
34 const struct object_id null_oid;
35 const struct object_id empty_tree_oid = {
36         EMPTY_TREE_SHA1_BIN_LITERAL
37 };
38 const struct object_id empty_blob_oid = {
39         EMPTY_BLOB_SHA1_BIN_LITERAL
40 };
41
42 /*
43  * This is meant to hold a *small* number of objects that you would
44  * want read_sha1_file() to be able to return, but yet you do not want
45  * to write them into the object store (e.g. a browse-only
46  * application).
47  */
48 static struct cached_object {
49         unsigned char sha1[20];
50         enum object_type type;
51         void *buf;
52         unsigned long size;
53 } *cached_objects;
54 static int cached_object_nr, cached_object_alloc;
55
56 static struct cached_object empty_tree = {
57         EMPTY_TREE_SHA1_BIN_LITERAL,
58         OBJ_TREE,
59         "",
60         0
61 };
62
63 static struct cached_object *find_cached_object(const unsigned char *sha1)
64 {
65         int i;
66         struct cached_object *co = cached_objects;
67
68         for (i = 0; i < cached_object_nr; i++, co++) {
69                 if (!hashcmp(co->sha1, sha1))
70                         return co;
71         }
72         if (!hashcmp(sha1, empty_tree.sha1))
73                 return &empty_tree;
74         return NULL;
75 }
76
77 int mkdir_in_gitdir(const char *path)
78 {
79         if (mkdir(path, 0777)) {
80                 int saved_errno = errno;
81                 struct stat st;
82                 struct strbuf sb = STRBUF_INIT;
83
84                 if (errno != EEXIST)
85                         return -1;
86                 /*
87                  * Are we looking at a path in a symlinked worktree
88                  * whose original repository does not yet have it?
89                  * e.g. .git/rr-cache pointing at its original
90                  * repository in which the user hasn't performed any
91                  * conflict resolution yet?
92                  */
93                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
94                     strbuf_readlink(&sb, path, st.st_size) ||
95                     !is_absolute_path(sb.buf) ||
96                     mkdir(sb.buf, 0777)) {
97                         strbuf_release(&sb);
98                         errno = saved_errno;
99                         return -1;
100                 }
101                 strbuf_release(&sb);
102         }
103         return adjust_shared_perm(path);
104 }
105
106 enum scld_error safe_create_leading_directories(char *path)
107 {
108         char *next_component = path + offset_1st_component(path);
109         enum scld_error ret = SCLD_OK;
110
111         while (ret == SCLD_OK && next_component) {
112                 struct stat st;
113                 char *slash = next_component, slash_character;
114
115                 while (*slash && !is_dir_sep(*slash))
116                         slash++;
117
118                 if (!*slash)
119                         break;
120
121                 next_component = slash + 1;
122                 while (is_dir_sep(*next_component))
123                         next_component++;
124                 if (!*next_component)
125                         break;
126
127                 slash_character = *slash;
128                 *slash = '\0';
129                 if (!stat(path, &st)) {
130                         /* path exists */
131                         if (!S_ISDIR(st.st_mode)) {
132                                 errno = ENOTDIR;
133                                 ret = SCLD_EXISTS;
134                         }
135                 } else if (mkdir(path, 0777)) {
136                         if (errno == EEXIST &&
137                             !stat(path, &st) && S_ISDIR(st.st_mode))
138                                 ; /* somebody created it since we checked */
139                         else if (errno == ENOENT)
140                                 /*
141                                  * Either mkdir() failed because
142                                  * somebody just pruned the containing
143                                  * directory, or stat() failed because
144                                  * the file that was in our way was
145                                  * just removed.  Either way, inform
146                                  * the caller that it might be worth
147                                  * trying again:
148                                  */
149                                 ret = SCLD_VANISHED;
150                         else
151                                 ret = SCLD_FAILED;
152                 } else if (adjust_shared_perm(path)) {
153                         ret = SCLD_PERMS;
154                 }
155                 *slash = slash_character;
156         }
157         return ret;
158 }
159
160 enum scld_error safe_create_leading_directories_const(const char *path)
161 {
162         int save_errno;
163         /* path points to cache entries, so xstrdup before messing with it */
164         char *buf = xstrdup(path);
165         enum scld_error result = safe_create_leading_directories(buf);
166
167         save_errno = errno;
168         free(buf);
169         errno = save_errno;
170         return result;
171 }
172
173 int raceproof_create_file(const char *path, create_file_fn fn, void *cb)
174 {
175         /*
176          * The number of times we will try to remove empty directories
177          * in the way of path. This is only 1 because if another
178          * process is racily creating directories that conflict with
179          * us, we don't want to fight against them.
180          */
181         int remove_directories_remaining = 1;
182
183         /*
184          * The number of times that we will try to create the
185          * directories containing path. We are willing to attempt this
186          * more than once, because another process could be trying to
187          * clean up empty directories at the same time as we are
188          * trying to create them.
189          */
190         int create_directories_remaining = 3;
191
192         /* A scratch copy of path, filled lazily if we need it: */
193         struct strbuf path_copy = STRBUF_INIT;
194
195         int ret, save_errno;
196
197         /* Sanity check: */
198         assert(*path);
199
200 retry_fn:
201         ret = fn(path, cb);
202         save_errno = errno;
203         if (!ret)
204                 goto out;
205
206         if (errno == EISDIR && remove_directories_remaining-- > 0) {
207                 /*
208                  * A directory is in the way. Maybe it is empty; try
209                  * to remove it:
210                  */
211                 if (!path_copy.len)
212                         strbuf_addstr(&path_copy, path);
213
214                 if (!remove_dir_recursively(&path_copy, REMOVE_DIR_EMPTY_ONLY))
215                         goto retry_fn;
216         } else if (errno == ENOENT && create_directories_remaining-- > 0) {
217                 /*
218                  * Maybe the containing directory didn't exist, or
219                  * maybe it was just deleted by a process that is
220                  * racing with us to clean up empty directories. Try
221                  * to create it:
222                  */
223                 enum scld_error scld_result;
224
225                 if (!path_copy.len)
226                         strbuf_addstr(&path_copy, path);
227
228                 do {
229                         scld_result = safe_create_leading_directories(path_copy.buf);
230                         if (scld_result == SCLD_OK)
231                                 goto retry_fn;
232                 } while (scld_result == SCLD_VANISHED && create_directories_remaining-- > 0);
233         }
234
235 out:
236         strbuf_release(&path_copy);
237         errno = save_errno;
238         return ret;
239 }
240
241 static void fill_sha1_path(struct strbuf *buf, const unsigned char *sha1)
242 {
243         int i;
244         for (i = 0; i < 20; i++) {
245                 static char hex[] = "0123456789abcdef";
246                 unsigned int val = sha1[i];
247                 strbuf_addch(buf, hex[val >> 4]);
248                 strbuf_addch(buf, hex[val & 0xf]);
249                 if (!i)
250                         strbuf_addch(buf, '/');
251         }
252 }
253
254 const char *sha1_file_name(const unsigned char *sha1)
255 {
256         static struct strbuf buf = STRBUF_INIT;
257
258         strbuf_reset(&buf);
259         strbuf_addf(&buf, "%s/", get_object_directory());
260
261         fill_sha1_path(&buf, sha1);
262         return buf.buf;
263 }
264
265 struct strbuf *alt_scratch_buf(struct alternate_object_database *alt)
266 {
267         strbuf_setlen(&alt->scratch, alt->base_len);
268         return &alt->scratch;
269 }
270
271 static const char *alt_sha1_path(struct alternate_object_database *alt,
272                                  const unsigned char *sha1)
273 {
274         struct strbuf *buf = alt_scratch_buf(alt);
275         fill_sha1_path(buf, sha1);
276         return buf->buf;
277 }
278
279 struct alternate_object_database *alt_odb_list;
280 static struct alternate_object_database **alt_odb_tail;
281
282 /*
283  * Return non-zero iff the path is usable as an alternate object database.
284  */
285 static int alt_odb_usable(struct strbuf *path, const char *normalized_objdir)
286 {
287         struct alternate_object_database *alt;
288
289         /* Detect cases where alternate disappeared */
290         if (!is_directory(path->buf)) {
291                 error("object directory %s does not exist; "
292                       "check .git/objects/info/alternates.",
293                       path->buf);
294                 return 0;
295         }
296
297         /*
298          * Prevent the common mistake of listing the same
299          * thing twice, or object directory itself.
300          */
301         for (alt = alt_odb_list; alt; alt = alt->next) {
302                 if (!fspathcmp(path->buf, alt->path))
303                         return 0;
304         }
305         if (!fspathcmp(path->buf, normalized_objdir))
306                 return 0;
307
308         return 1;
309 }
310
311 /*
312  * Prepare alternate object database registry.
313  *
314  * The variable alt_odb_list points at the list of struct
315  * alternate_object_database.  The elements on this list come from
316  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
317  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
318  * whose contents is similar to that environment variable but can be
319  * LF separated.  Its base points at a statically allocated buffer that
320  * contains "/the/directory/corresponding/to/.git/objects/...", while
321  * its name points just after the slash at the end of ".git/objects/"
322  * in the example above, and has enough space to hold 40-byte hex
323  * SHA1, an extra slash for the first level indirection, and the
324  * terminating NUL.
325  */
326 static void read_info_alternates(const char * relative_base, int depth);
327 static int link_alt_odb_entry(const char *entry, const char *relative_base,
328         int depth, const char *normalized_objdir)
329 {
330         struct alternate_object_database *ent;
331         struct strbuf pathbuf = STRBUF_INIT;
332
333         if (!is_absolute_path(entry) && relative_base) {
334                 strbuf_realpath(&pathbuf, relative_base, 1);
335                 strbuf_addch(&pathbuf, '/');
336         }
337         strbuf_addstr(&pathbuf, entry);
338
339         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
340                 error("unable to normalize alternate object path: %s",
341                       pathbuf.buf);
342                 strbuf_release(&pathbuf);
343                 return -1;
344         }
345
346         /*
347          * The trailing slash after the directory name is given by
348          * this function at the end. Remove duplicates.
349          */
350         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
351                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
352
353         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
354                 strbuf_release(&pathbuf);
355                 return -1;
356         }
357
358         ent = alloc_alt_odb(pathbuf.buf);
359
360         /* add the alternate entry */
361         *alt_odb_tail = ent;
362         alt_odb_tail = &(ent->next);
363         ent->next = NULL;
364
365         /* recursively add alternates */
366         read_info_alternates(pathbuf.buf, depth + 1);
367
368         strbuf_release(&pathbuf);
369         return 0;
370 }
371
372 static const char *parse_alt_odb_entry(const char *string,
373                                        int sep,
374                                        struct strbuf *out)
375 {
376         const char *end;
377
378         strbuf_reset(out);
379
380         if (*string == '#') {
381                 /* comment; consume up to next separator */
382                 end = strchrnul(string, sep);
383         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
384                 /*
385                  * quoted path; unquote_c_style has copied the
386                  * data for us and set "end". Broken quoting (e.g.,
387                  * an entry that doesn't end with a quote) falls
388                  * back to the unquoted case below.
389                  */
390         } else {
391                 /* normal, unquoted path */
392                 end = strchrnul(string, sep);
393                 strbuf_add(out, string, end - string);
394         }
395
396         if (*end)
397                 end++;
398         return end;
399 }
400
401 static void link_alt_odb_entries(const char *alt, int sep,
402                                  const char *relative_base, int depth)
403 {
404         struct strbuf objdirbuf = STRBUF_INIT;
405         struct strbuf entry = STRBUF_INIT;
406
407         if (depth > 5) {
408                 error("%s: ignoring alternate object stores, nesting too deep.",
409                                 relative_base);
410                 return;
411         }
412
413         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
414         if (strbuf_normalize_path(&objdirbuf) < 0)
415                 die("unable to normalize object directory: %s",
416                     objdirbuf.buf);
417
418         while (*alt) {
419                 alt = parse_alt_odb_entry(alt, sep, &entry);
420                 if (!entry.len)
421                         continue;
422                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
423         }
424         strbuf_release(&entry);
425         strbuf_release(&objdirbuf);
426 }
427
428 static void read_info_alternates(const char * relative_base, int depth)
429 {
430         char *path;
431         struct strbuf buf = STRBUF_INIT;
432
433         path = xstrfmt("%s/info/alternates", relative_base);
434         if (strbuf_read_file(&buf, path, 1024) < 0) {
435                 warn_on_fopen_errors(path);
436                 free(path);
437                 return;
438         }
439
440         link_alt_odb_entries(buf.buf, '\n', relative_base, depth);
441         strbuf_release(&buf);
442         free(path);
443 }
444
445 struct alternate_object_database *alloc_alt_odb(const char *dir)
446 {
447         struct alternate_object_database *ent;
448
449         FLEX_ALLOC_STR(ent, path, dir);
450         strbuf_init(&ent->scratch, 0);
451         strbuf_addf(&ent->scratch, "%s/", dir);
452         ent->base_len = ent->scratch.len;
453
454         return ent;
455 }
456
457 void add_to_alternates_file(const char *reference)
458 {
459         struct lock_file lock = LOCK_INIT;
460         char *alts = git_pathdup("objects/info/alternates");
461         FILE *in, *out;
462         int found = 0;
463
464         hold_lock_file_for_update(&lock, alts, LOCK_DIE_ON_ERROR);
465         out = fdopen_lock_file(&lock, "w");
466         if (!out)
467                 die_errno("unable to fdopen alternates lockfile");
468
469         in = fopen(alts, "r");
470         if (in) {
471                 struct strbuf line = STRBUF_INIT;
472
473                 while (strbuf_getline(&line, in) != EOF) {
474                         if (!strcmp(reference, line.buf)) {
475                                 found = 1;
476                                 break;
477                         }
478                         fprintf_or_die(out, "%s\n", line.buf);
479                 }
480
481                 strbuf_release(&line);
482                 fclose(in);
483         }
484         else if (errno != ENOENT)
485                 die_errno("unable to read alternates file");
486
487         if (found) {
488                 rollback_lock_file(&lock);
489         } else {
490                 fprintf_or_die(out, "%s\n", reference);
491                 if (commit_lock_file(&lock))
492                         die_errno("unable to move new alternates file into place");
493                 if (alt_odb_tail)
494                         link_alt_odb_entries(reference, '\n', NULL, 0);
495         }
496         free(alts);
497 }
498
499 void add_to_alternates_memory(const char *reference)
500 {
501         /*
502          * Make sure alternates are initialized, or else our entry may be
503          * overwritten when they are.
504          */
505         prepare_alt_odb();
506
507         link_alt_odb_entries(reference, '\n', NULL, 0);
508 }
509
510 /*
511  * Compute the exact path an alternate is at and returns it. In case of
512  * error NULL is returned and the human readable error is added to `err`
513  * `path` may be relative and should point to $GITDIR.
514  * `err` must not be null.
515  */
516 char *compute_alternate_path(const char *path, struct strbuf *err)
517 {
518         char *ref_git = NULL;
519         const char *repo, *ref_git_s;
520         int seen_error = 0;
521
522         ref_git_s = real_path_if_valid(path);
523         if (!ref_git_s) {
524                 seen_error = 1;
525                 strbuf_addf(err, _("path '%s' does not exist"), path);
526                 goto out;
527         } else
528                 /*
529                  * Beware: read_gitfile(), real_path() and mkpath()
530                  * return static buffer
531                  */
532                 ref_git = xstrdup(ref_git_s);
533
534         repo = read_gitfile(ref_git);
535         if (!repo)
536                 repo = read_gitfile(mkpath("%s/.git", ref_git));
537         if (repo) {
538                 free(ref_git);
539                 ref_git = xstrdup(repo);
540         }
541
542         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
543                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
544                 free(ref_git);
545                 ref_git = ref_git_git;
546         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
547                 struct strbuf sb = STRBUF_INIT;
548                 seen_error = 1;
549                 if (get_common_dir(&sb, ref_git)) {
550                         strbuf_addf(err,
551                                     _("reference repository '%s' as a linked "
552                                       "checkout is not supported yet."),
553                                     path);
554                         goto out;
555                 }
556
557                 strbuf_addf(err, _("reference repository '%s' is not a "
558                                         "local repository."), path);
559                 goto out;
560         }
561
562         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
563                 strbuf_addf(err, _("reference repository '%s' is shallow"),
564                             path);
565                 seen_error = 1;
566                 goto out;
567         }
568
569         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
570                 strbuf_addf(err,
571                             _("reference repository '%s' is grafted"),
572                             path);
573                 seen_error = 1;
574                 goto out;
575         }
576
577 out:
578         if (seen_error) {
579                 FREE_AND_NULL(ref_git);
580         }
581
582         return ref_git;
583 }
584
585 int foreach_alt_odb(alt_odb_fn fn, void *cb)
586 {
587         struct alternate_object_database *ent;
588         int r = 0;
589
590         prepare_alt_odb();
591         for (ent = alt_odb_list; ent; ent = ent->next) {
592                 r = fn(ent, cb);
593                 if (r)
594                         break;
595         }
596         return r;
597 }
598
599 void prepare_alt_odb(void)
600 {
601         const char *alt;
602
603         if (alt_odb_tail)
604                 return;
605
606         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
607         if (!alt) alt = "";
608
609         alt_odb_tail = &alt_odb_list;
610         link_alt_odb_entries(alt, PATH_SEP, NULL, 0);
611
612         read_info_alternates(get_object_directory(), 0);
613 }
614
615 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
616 static int freshen_file(const char *fn)
617 {
618         struct utimbuf t;
619         t.actime = t.modtime = time(NULL);
620         return !utime(fn, &t);
621 }
622
623 /*
624  * All of the check_and_freshen functions return 1 if the file exists and was
625  * freshened (if freshening was requested), 0 otherwise. If they return
626  * 0, you should not assume that it is safe to skip a write of the object (it
627  * either does not exist on disk, or has a stale mtime and may be subject to
628  * pruning).
629  */
630 int check_and_freshen_file(const char *fn, int freshen)
631 {
632         if (access(fn, F_OK))
633                 return 0;
634         if (freshen && !freshen_file(fn))
635                 return 0;
636         return 1;
637 }
638
639 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
640 {
641         return check_and_freshen_file(sha1_file_name(sha1), freshen);
642 }
643
644 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
645 {
646         struct alternate_object_database *alt;
647         prepare_alt_odb();
648         for (alt = alt_odb_list; alt; alt = alt->next) {
649                 const char *path = alt_sha1_path(alt, sha1);
650                 if (check_and_freshen_file(path, freshen))
651                         return 1;
652         }
653         return 0;
654 }
655
656 static int check_and_freshen(const unsigned char *sha1, int freshen)
657 {
658         return check_and_freshen_local(sha1, freshen) ||
659                check_and_freshen_nonlocal(sha1, freshen);
660 }
661
662 int has_loose_object_nonlocal(const unsigned char *sha1)
663 {
664         return check_and_freshen_nonlocal(sha1, 0);
665 }
666
667 static int has_loose_object(const unsigned char *sha1)
668 {
669         return check_and_freshen(sha1, 0);
670 }
671
672 static void mmap_limit_check(size_t length)
673 {
674         static size_t limit = 0;
675         if (!limit) {
676                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
677                 if (!limit)
678                         limit = SIZE_MAX;
679         }
680         if (length > limit)
681                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
682                     (uintmax_t)length, (uintmax_t)limit);
683 }
684
685 void *xmmap_gently(void *start, size_t length,
686                   int prot, int flags, int fd, off_t offset)
687 {
688         void *ret;
689
690         mmap_limit_check(length);
691         ret = mmap(start, length, prot, flags, fd, offset);
692         if (ret == MAP_FAILED) {
693                 if (!length)
694                         return NULL;
695                 release_pack_memory(length);
696                 ret = mmap(start, length, prot, flags, fd, offset);
697         }
698         return ret;
699 }
700
701 void *xmmap(void *start, size_t length,
702         int prot, int flags, int fd, off_t offset)
703 {
704         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
705         if (ret == MAP_FAILED)
706                 die_errno("mmap failed");
707         return ret;
708 }
709
710 /*
711  * With an in-core object data in "map", rehash it to make sure the
712  * object name actually matches "sha1" to detect object corruption.
713  * With "map" == NULL, try reading the object named with "sha1" using
714  * the streaming interface and rehash it to do the same.
715  */
716 int check_sha1_signature(const unsigned char *sha1, void *map,
717                          unsigned long size, const char *type)
718 {
719         unsigned char real_sha1[20];
720         enum object_type obj_type;
721         struct git_istream *st;
722         git_SHA_CTX c;
723         char hdr[32];
724         int hdrlen;
725
726         if (map) {
727                 hash_sha1_file(map, size, type, real_sha1);
728                 return hashcmp(sha1, real_sha1) ? -1 : 0;
729         }
730
731         st = open_istream(sha1, &obj_type, &size, NULL);
732         if (!st)
733                 return -1;
734
735         /* Generate the header */
736         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
737
738         /* Sha1.. */
739         git_SHA1_Init(&c);
740         git_SHA1_Update(&c, hdr, hdrlen);
741         for (;;) {
742                 char buf[1024 * 16];
743                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
744
745                 if (readlen < 0) {
746                         close_istream(st);
747                         return -1;
748                 }
749                 if (!readlen)
750                         break;
751                 git_SHA1_Update(&c, buf, readlen);
752         }
753         git_SHA1_Final(real_sha1, &c);
754         close_istream(st);
755         return hashcmp(sha1, real_sha1) ? -1 : 0;
756 }
757
758 int git_open_cloexec(const char *name, int flags)
759 {
760         int fd;
761         static int o_cloexec = O_CLOEXEC;
762
763         fd = open(name, flags | o_cloexec);
764         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
765                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
766                 o_cloexec &= ~O_CLOEXEC;
767                 fd = open(name, flags | o_cloexec);
768         }
769
770 #if defined(F_GETFD) && defined(F_SETFD) && defined(FD_CLOEXEC)
771         {
772                 static int fd_cloexec = FD_CLOEXEC;
773
774                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
775                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
776                         int flags = fcntl(fd, F_GETFD);
777                         if (fcntl(fd, F_SETFD, flags | fd_cloexec))
778                                 fd_cloexec = 0;
779                 }
780         }
781 #endif
782         return fd;
783 }
784
785 /*
786  * Find "sha1" as a loose object in the local repository or in an alternate.
787  * Returns 0 on success, negative on failure.
788  *
789  * The "path" out-parameter will give the path of the object we found (if any).
790  * Note that it may point to static storage and is only valid until another
791  * call to sha1_file_name(), etc.
792  */
793 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
794                           const char **path)
795 {
796         struct alternate_object_database *alt;
797
798         *path = sha1_file_name(sha1);
799         if (!lstat(*path, st))
800                 return 0;
801
802         prepare_alt_odb();
803         errno = ENOENT;
804         for (alt = alt_odb_list; alt; alt = alt->next) {
805                 *path = alt_sha1_path(alt, sha1);
806                 if (!lstat(*path, st))
807                         return 0;
808         }
809
810         return -1;
811 }
812
813 /*
814  * Like stat_sha1_file(), but actually open the object and return the
815  * descriptor. See the caveats on the "path" parameter above.
816  */
817 static int open_sha1_file(const unsigned char *sha1, const char **path)
818 {
819         int fd;
820         struct alternate_object_database *alt;
821         int most_interesting_errno;
822
823         *path = sha1_file_name(sha1);
824         fd = git_open(*path);
825         if (fd >= 0)
826                 return fd;
827         most_interesting_errno = errno;
828
829         prepare_alt_odb();
830         for (alt = alt_odb_list; alt; alt = alt->next) {
831                 *path = alt_sha1_path(alt, sha1);
832                 fd = git_open(*path);
833                 if (fd >= 0)
834                         return fd;
835                 if (most_interesting_errno == ENOENT)
836                         most_interesting_errno = errno;
837         }
838         errno = most_interesting_errno;
839         return -1;
840 }
841
842 /*
843  * Map the loose object at "path" if it is not NULL, or the path found by
844  * searching for a loose object named "sha1".
845  */
846 static void *map_sha1_file_1(const char *path,
847                              const unsigned char *sha1,
848                              unsigned long *size)
849 {
850         void *map;
851         int fd;
852
853         if (path)
854                 fd = git_open(path);
855         else
856                 fd = open_sha1_file(sha1, &path);
857         map = NULL;
858         if (fd >= 0) {
859                 struct stat st;
860
861                 if (!fstat(fd, &st)) {
862                         *size = xsize_t(st.st_size);
863                         if (!*size) {
864                                 /* mmap() is forbidden on empty files */
865                                 error("object file %s is empty", path);
866                                 return NULL;
867                         }
868                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
869                 }
870                 close(fd);
871         }
872         return map;
873 }
874
875 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
876 {
877         return map_sha1_file_1(NULL, sha1, size);
878 }
879
880 static int unpack_sha1_short_header(git_zstream *stream,
881                                     unsigned char *map, unsigned long mapsize,
882                                     void *buffer, unsigned long bufsiz)
883 {
884         /* Get the data stream */
885         memset(stream, 0, sizeof(*stream));
886         stream->next_in = map;
887         stream->avail_in = mapsize;
888         stream->next_out = buffer;
889         stream->avail_out = bufsiz;
890
891         git_inflate_init(stream);
892         return git_inflate(stream, 0);
893 }
894
895 int unpack_sha1_header(git_zstream *stream,
896                        unsigned char *map, unsigned long mapsize,
897                        void *buffer, unsigned long bufsiz)
898 {
899         int status = unpack_sha1_short_header(stream, map, mapsize,
900                                               buffer, bufsiz);
901
902         if (status < Z_OK)
903                 return status;
904
905         /* Make sure we have the terminating NUL */
906         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
907                 return -1;
908         return 0;
909 }
910
911 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
912                                         unsigned long mapsize, void *buffer,
913                                         unsigned long bufsiz, struct strbuf *header)
914 {
915         int status;
916
917         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
918         if (status < Z_OK)
919                 return -1;
920
921         /*
922          * Check if entire header is unpacked in the first iteration.
923          */
924         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
925                 return 0;
926
927         /*
928          * buffer[0..bufsiz] was not large enough.  Copy the partial
929          * result out to header, and then append the result of further
930          * reading the stream.
931          */
932         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
933         stream->next_out = buffer;
934         stream->avail_out = bufsiz;
935
936         do {
937                 status = git_inflate(stream, 0);
938                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
939                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
940                         return 0;
941                 stream->next_out = buffer;
942                 stream->avail_out = bufsiz;
943         } while (status != Z_STREAM_END);
944         return -1;
945 }
946
947 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
948 {
949         int bytes = strlen(buffer) + 1;
950         unsigned char *buf = xmallocz(size);
951         unsigned long n;
952         int status = Z_OK;
953
954         n = stream->total_out - bytes;
955         if (n > size)
956                 n = size;
957         memcpy(buf, (char *) buffer + bytes, n);
958         bytes = n;
959         if (bytes <= size) {
960                 /*
961                  * The above condition must be (bytes <= size), not
962                  * (bytes < size).  In other words, even though we
963                  * expect no more output and set avail_out to zero,
964                  * the input zlib stream may have bytes that express
965                  * "this concludes the stream", and we *do* want to
966                  * eat that input.
967                  *
968                  * Otherwise we would not be able to test that we
969                  * consumed all the input to reach the expected size;
970                  * we also want to check that zlib tells us that all
971                  * went well with status == Z_STREAM_END at the end.
972                  */
973                 stream->next_out = buf + bytes;
974                 stream->avail_out = size - bytes;
975                 while (status == Z_OK)
976                         status = git_inflate(stream, Z_FINISH);
977         }
978         if (status == Z_STREAM_END && !stream->avail_in) {
979                 git_inflate_end(stream);
980                 return buf;
981         }
982
983         if (status < 0)
984                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
985         else if (stream->avail_in)
986                 error("garbage at end of loose object '%s'",
987                       sha1_to_hex(sha1));
988         free(buf);
989         return NULL;
990 }
991
992 /*
993  * We used to just use "sscanf()", but that's actually way
994  * too permissive for what we want to check. So do an anal
995  * object header parse by hand.
996  */
997 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
998                                unsigned int flags)
999 {
1000         const char *type_buf = hdr;
1001         unsigned long size;
1002         int type, type_len = 0;
1003
1004         /*
1005          * The type can be of any size but is followed by
1006          * a space.
1007          */
1008         for (;;) {
1009                 char c = *hdr++;
1010                 if (!c)
1011                         return -1;
1012                 if (c == ' ')
1013                         break;
1014                 type_len++;
1015         }
1016
1017         type = type_from_string_gently(type_buf, type_len, 1);
1018         if (oi->typename)
1019                 strbuf_add(oi->typename, type_buf, type_len);
1020         /*
1021          * Set type to 0 if its an unknown object and
1022          * we're obtaining the type using '--allow-unknown-type'
1023          * option.
1024          */
1025         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE) && (type < 0))
1026                 type = 0;
1027         else if (type < 0)
1028                 die("invalid object type");
1029         if (oi->typep)
1030                 *oi->typep = type;
1031
1032         /*
1033          * The length must follow immediately, and be in canonical
1034          * decimal format (ie "010" is not valid).
1035          */
1036         size = *hdr++ - '0';
1037         if (size > 9)
1038                 return -1;
1039         if (size) {
1040                 for (;;) {
1041                         unsigned long c = *hdr - '0';
1042                         if (c > 9)
1043                                 break;
1044                         hdr++;
1045                         size = size * 10 + c;
1046                 }
1047         }
1048
1049         if (oi->sizep)
1050                 *oi->sizep = size;
1051
1052         /*
1053          * The length must be followed by a zero byte
1054          */
1055         return *hdr ? -1 : type;
1056 }
1057
1058 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1059 {
1060         struct object_info oi = OBJECT_INFO_INIT;
1061
1062         oi.sizep = sizep;
1063         return parse_sha1_header_extended(hdr, &oi, 0);
1064 }
1065
1066 static int sha1_loose_object_info(const unsigned char *sha1,
1067                                   struct object_info *oi,
1068                                   int flags)
1069 {
1070         int status = 0;
1071         unsigned long mapsize;
1072         void *map;
1073         git_zstream stream;
1074         char hdr[32];
1075         struct strbuf hdrbuf = STRBUF_INIT;
1076         unsigned long size_scratch;
1077
1078         if (oi->delta_base_sha1)
1079                 hashclr(oi->delta_base_sha1);
1080
1081         /*
1082          * If we don't care about type or size, then we don't
1083          * need to look inside the object at all. Note that we
1084          * do not optimize out the stat call, even if the
1085          * caller doesn't care about the disk-size, since our
1086          * return value implicitly indicates whether the
1087          * object even exists.
1088          */
1089         if (!oi->typep && !oi->typename && !oi->sizep && !oi->contentp) {
1090                 const char *path;
1091                 struct stat st;
1092                 if (stat_sha1_file(sha1, &st, &path) < 0)
1093                         return -1;
1094                 if (oi->disk_sizep)
1095                         *oi->disk_sizep = st.st_size;
1096                 return 0;
1097         }
1098
1099         map = map_sha1_file(sha1, &mapsize);
1100         if (!map)
1101                 return -1;
1102
1103         if (!oi->sizep)
1104                 oi->sizep = &size_scratch;
1105
1106         if (oi->disk_sizep)
1107                 *oi->disk_sizep = mapsize;
1108         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE)) {
1109                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
1110                         status = error("unable to unpack %s header with --allow-unknown-type",
1111                                        sha1_to_hex(sha1));
1112         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1113                 status = error("unable to unpack %s header",
1114                                sha1_to_hex(sha1));
1115         if (status < 0)
1116                 ; /* Do nothing */
1117         else if (hdrbuf.len) {
1118                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
1119                         status = error("unable to parse %s header with --allow-unknown-type",
1120                                        sha1_to_hex(sha1));
1121         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
1122                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1123
1124         if (status >= 0 && oi->contentp) {
1125                 *oi->contentp = unpack_sha1_rest(&stream, hdr,
1126                                                  *oi->sizep, sha1);
1127                 if (!*oi->contentp) {
1128                         git_inflate_end(&stream);
1129                         status = -1;
1130                 }
1131         } else
1132                 git_inflate_end(&stream);
1133
1134         munmap(map, mapsize);
1135         if (status && oi->typep)
1136                 *oi->typep = status;
1137         if (oi->sizep == &size_scratch)
1138                 oi->sizep = NULL;
1139         strbuf_release(&hdrbuf);
1140         oi->whence = OI_LOOSE;
1141         return (status < 0) ? status : 0;
1142 }
1143
1144 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
1145 {
1146         static struct object_info blank_oi = OBJECT_INFO_INIT;
1147         struct pack_entry e;
1148         int rtype;
1149         const unsigned char *real = (flags & OBJECT_INFO_LOOKUP_REPLACE) ?
1150                                     lookup_replace_object(sha1) :
1151                                     sha1;
1152
1153         if (!oi)
1154                 oi = &blank_oi;
1155
1156         if (!(flags & OBJECT_INFO_SKIP_CACHED)) {
1157                 struct cached_object *co = find_cached_object(real);
1158                 if (co) {
1159                         if (oi->typep)
1160                                 *(oi->typep) = co->type;
1161                         if (oi->sizep)
1162                                 *(oi->sizep) = co->size;
1163                         if (oi->disk_sizep)
1164                                 *(oi->disk_sizep) = 0;
1165                         if (oi->delta_base_sha1)
1166                                 hashclr(oi->delta_base_sha1);
1167                         if (oi->typename)
1168                                 strbuf_addstr(oi->typename, typename(co->type));
1169                         if (oi->contentp)
1170                                 *oi->contentp = xmemdupz(co->buf, co->size);
1171                         oi->whence = OI_CACHED;
1172                         return 0;
1173                 }
1174         }
1175
1176         if (!find_pack_entry(real, &e)) {
1177                 /* Most likely it's a loose object. */
1178                 if (!sha1_loose_object_info(real, oi, flags))
1179                         return 0;
1180
1181                 /* Not a loose object; someone else may have just packed it. */
1182                 if (flags & OBJECT_INFO_QUICK) {
1183                         return -1;
1184                 } else {
1185                         reprepare_packed_git();
1186                         if (!find_pack_entry(real, &e))
1187                                 return -1;
1188                 }
1189         }
1190
1191         if (oi == &blank_oi)
1192                 /*
1193                  * We know that the caller doesn't actually need the
1194                  * information below, so return early.
1195                  */
1196                 return 0;
1197
1198         rtype = packed_object_info(e.p, e.offset, oi);
1199         if (rtype < 0) {
1200                 mark_bad_packed_object(e.p, real);
1201                 return sha1_object_info_extended(real, oi, 0);
1202         } else if (oi->whence == OI_PACKED) {
1203                 oi->u.packed.offset = e.offset;
1204                 oi->u.packed.pack = e.p;
1205                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
1206                                          rtype == OBJ_OFS_DELTA);
1207         }
1208
1209         return 0;
1210 }
1211
1212 /* returns enum object_type or negative */
1213 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1214 {
1215         enum object_type type;
1216         struct object_info oi = OBJECT_INFO_INIT;
1217
1218         oi.typep = &type;
1219         oi.sizep = sizep;
1220         if (sha1_object_info_extended(sha1, &oi,
1221                                       OBJECT_INFO_LOOKUP_REPLACE) < 0)
1222                 return -1;
1223         return type;
1224 }
1225
1226 static void *read_object(const unsigned char *sha1, enum object_type *type,
1227                          unsigned long *size)
1228 {
1229         struct object_info oi = OBJECT_INFO_INIT;
1230         void *content;
1231         oi.typep = type;
1232         oi.sizep = size;
1233         oi.contentp = &content;
1234
1235         if (sha1_object_info_extended(sha1, &oi, 0) < 0)
1236                 return NULL;
1237         return content;
1238 }
1239
1240 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
1241                       unsigned char *sha1)
1242 {
1243         struct cached_object *co;
1244
1245         hash_sha1_file(buf, len, typename(type), sha1);
1246         if (has_sha1_file(sha1) || find_cached_object(sha1))
1247                 return 0;
1248         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
1249         co = &cached_objects[cached_object_nr++];
1250         co->size = len;
1251         co->type = type;
1252         co->buf = xmalloc(len);
1253         memcpy(co->buf, buf, len);
1254         hashcpy(co->sha1, sha1);
1255         return 0;
1256 }
1257
1258 /*
1259  * This function dies on corrupt objects; the callers who want to
1260  * deal with them should arrange to call read_object() and give error
1261  * messages themselves.
1262  */
1263 void *read_sha1_file_extended(const unsigned char *sha1,
1264                               enum object_type *type,
1265                               unsigned long *size,
1266                               int lookup_replace)
1267 {
1268         void *data;
1269         const struct packed_git *p;
1270         const char *path;
1271         struct stat st;
1272         const unsigned char *repl = lookup_replace ? lookup_replace_object(sha1)
1273                                                    : sha1;
1274
1275         errno = 0;
1276         data = read_object(repl, type, size);
1277         if (data)
1278                 return data;
1279
1280         if (errno && errno != ENOENT)
1281                 die_errno("failed to read object %s", sha1_to_hex(sha1));
1282
1283         /* die if we replaced an object with one that does not exist */
1284         if (repl != sha1)
1285                 die("replacement %s not found for %s",
1286                     sha1_to_hex(repl), sha1_to_hex(sha1));
1287
1288         if (!stat_sha1_file(repl, &st, &path))
1289                 die("loose object %s (stored in %s) is corrupt",
1290                     sha1_to_hex(repl), path);
1291
1292         if ((p = has_packed_and_bad(repl)) != NULL)
1293                 die("packed object %s (stored in %s) is corrupt",
1294                     sha1_to_hex(repl), p->pack_name);
1295
1296         return NULL;
1297 }
1298
1299 void *read_object_with_reference(const unsigned char *sha1,
1300                                  const char *required_type_name,
1301                                  unsigned long *size,
1302                                  unsigned char *actual_sha1_return)
1303 {
1304         enum object_type type, required_type;
1305         void *buffer;
1306         unsigned long isize;
1307         unsigned char actual_sha1[20];
1308
1309         required_type = type_from_string(required_type_name);
1310         hashcpy(actual_sha1, sha1);
1311         while (1) {
1312                 int ref_length = -1;
1313                 const char *ref_type = NULL;
1314
1315                 buffer = read_sha1_file(actual_sha1, &type, &isize);
1316                 if (!buffer)
1317                         return NULL;
1318                 if (type == required_type) {
1319                         *size = isize;
1320                         if (actual_sha1_return)
1321                                 hashcpy(actual_sha1_return, actual_sha1);
1322                         return buffer;
1323                 }
1324                 /* Handle references */
1325                 else if (type == OBJ_COMMIT)
1326                         ref_type = "tree ";
1327                 else if (type == OBJ_TAG)
1328                         ref_type = "object ";
1329                 else {
1330                         free(buffer);
1331                         return NULL;
1332                 }
1333                 ref_length = strlen(ref_type);
1334
1335                 if (ref_length + 40 > isize ||
1336                     memcmp(buffer, ref_type, ref_length) ||
1337                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
1338                         free(buffer);
1339                         return NULL;
1340                 }
1341                 free(buffer);
1342                 /* Now we have the ID of the referred-to object in
1343                  * actual_sha1.  Check again. */
1344         }
1345 }
1346
1347 static void write_sha1_file_prepare(const void *buf, unsigned long len,
1348                                     const char *type, unsigned char *sha1,
1349                                     char *hdr, int *hdrlen)
1350 {
1351         git_SHA_CTX c;
1352
1353         /* Generate the header */
1354         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
1355
1356         /* Sha1.. */
1357         git_SHA1_Init(&c);
1358         git_SHA1_Update(&c, hdr, *hdrlen);
1359         git_SHA1_Update(&c, buf, len);
1360         git_SHA1_Final(sha1, &c);
1361 }
1362
1363 /*
1364  * Move the just written object into its final resting place.
1365  */
1366 int finalize_object_file(const char *tmpfile, const char *filename)
1367 {
1368         int ret = 0;
1369
1370         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
1371                 goto try_rename;
1372         else if (link(tmpfile, filename))
1373                 ret = errno;
1374
1375         /*
1376          * Coda hack - coda doesn't like cross-directory links,
1377          * so we fall back to a rename, which will mean that it
1378          * won't be able to check collisions, but that's not a
1379          * big deal.
1380          *
1381          * The same holds for FAT formatted media.
1382          *
1383          * When this succeeds, we just return.  We have nothing
1384          * left to unlink.
1385          */
1386         if (ret && ret != EEXIST) {
1387         try_rename:
1388                 if (!rename(tmpfile, filename))
1389                         goto out;
1390                 ret = errno;
1391         }
1392         unlink_or_warn(tmpfile);
1393         if (ret) {
1394                 if (ret != EEXIST) {
1395                         return error_errno("unable to write sha1 filename %s", filename);
1396                 }
1397                 /* FIXME!!! Collision check here ? */
1398         }
1399
1400 out:
1401         if (adjust_shared_perm(filename))
1402                 return error("unable to set permission to '%s'", filename);
1403         return 0;
1404 }
1405
1406 static int write_buffer(int fd, const void *buf, size_t len)
1407 {
1408         if (write_in_full(fd, buf, len) < 0)
1409                 return error_errno("file write error");
1410         return 0;
1411 }
1412
1413 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
1414                    unsigned char *sha1)
1415 {
1416         char hdr[32];
1417         int hdrlen = sizeof(hdr);
1418         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
1419         return 0;
1420 }
1421
1422 /* Finalize a file on disk, and close it. */
1423 static void close_sha1_file(int fd)
1424 {
1425         if (fsync_object_files)
1426                 fsync_or_die(fd, "sha1 file");
1427         if (close(fd) != 0)
1428                 die_errno("error when closing sha1 file");
1429 }
1430
1431 /* Size of directory component, including the ending '/' */
1432 static inline int directory_size(const char *filename)
1433 {
1434         const char *s = strrchr(filename, '/');
1435         if (!s)
1436                 return 0;
1437         return s - filename + 1;
1438 }
1439
1440 /*
1441  * This creates a temporary file in the same directory as the final
1442  * 'filename'
1443  *
1444  * We want to avoid cross-directory filename renames, because those
1445  * can have problems on various filesystems (FAT, NFS, Coda).
1446  */
1447 static int create_tmpfile(struct strbuf *tmp, const char *filename)
1448 {
1449         int fd, dirlen = directory_size(filename);
1450
1451         strbuf_reset(tmp);
1452         strbuf_add(tmp, filename, dirlen);
1453         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
1454         fd = git_mkstemp_mode(tmp->buf, 0444);
1455         if (fd < 0 && dirlen && errno == ENOENT) {
1456                 /*
1457                  * Make sure the directory exists; note that the contents
1458                  * of the buffer are undefined after mkstemp returns an
1459                  * error, so we have to rewrite the whole buffer from
1460                  * scratch.
1461                  */
1462                 strbuf_reset(tmp);
1463                 strbuf_add(tmp, filename, dirlen - 1);
1464                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
1465                         return -1;
1466                 if (adjust_shared_perm(tmp->buf))
1467                         return -1;
1468
1469                 /* Try again */
1470                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
1471                 fd = git_mkstemp_mode(tmp->buf, 0444);
1472         }
1473         return fd;
1474 }
1475
1476 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
1477                               const void *buf, unsigned long len, time_t mtime)
1478 {
1479         int fd, ret;
1480         unsigned char compressed[4096];
1481         git_zstream stream;
1482         git_SHA_CTX c;
1483         unsigned char parano_sha1[20];
1484         static struct strbuf tmp_file = STRBUF_INIT;
1485         const char *filename = sha1_file_name(sha1);
1486
1487         fd = create_tmpfile(&tmp_file, filename);
1488         if (fd < 0) {
1489                 if (errno == EACCES)
1490                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
1491                 else
1492                         return error_errno("unable to create temporary file");
1493         }
1494
1495         /* Set it up */
1496         git_deflate_init(&stream, zlib_compression_level);
1497         stream.next_out = compressed;
1498         stream.avail_out = sizeof(compressed);
1499         git_SHA1_Init(&c);
1500
1501         /* First header.. */
1502         stream.next_in = (unsigned char *)hdr;
1503         stream.avail_in = hdrlen;
1504         while (git_deflate(&stream, 0) == Z_OK)
1505                 ; /* nothing */
1506         git_SHA1_Update(&c, hdr, hdrlen);
1507
1508         /* Then the data itself.. */
1509         stream.next_in = (void *)buf;
1510         stream.avail_in = len;
1511         do {
1512                 unsigned char *in0 = stream.next_in;
1513                 ret = git_deflate(&stream, Z_FINISH);
1514                 git_SHA1_Update(&c, in0, stream.next_in - in0);
1515                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
1516                         die("unable to write sha1 file");
1517                 stream.next_out = compressed;
1518                 stream.avail_out = sizeof(compressed);
1519         } while (ret == Z_OK);
1520
1521         if (ret != Z_STREAM_END)
1522                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
1523         ret = git_deflate_end_gently(&stream);
1524         if (ret != Z_OK)
1525                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
1526         git_SHA1_Final(parano_sha1, &c);
1527         if (hashcmp(sha1, parano_sha1) != 0)
1528                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
1529
1530         close_sha1_file(fd);
1531
1532         if (mtime) {
1533                 struct utimbuf utb;
1534                 utb.actime = mtime;
1535                 utb.modtime = mtime;
1536                 if (utime(tmp_file.buf, &utb) < 0)
1537                         warning_errno("failed utime() on %s", tmp_file.buf);
1538         }
1539
1540         return finalize_object_file(tmp_file.buf, filename);
1541 }
1542
1543 static int freshen_loose_object(const unsigned char *sha1)
1544 {
1545         return check_and_freshen(sha1, 1);
1546 }
1547
1548 static int freshen_packed_object(const unsigned char *sha1)
1549 {
1550         struct pack_entry e;
1551         if (!find_pack_entry(sha1, &e))
1552                 return 0;
1553         if (e.p->freshened)
1554                 return 1;
1555         if (!freshen_file(e.p->pack_name))
1556                 return 0;
1557         e.p->freshened = 1;
1558         return 1;
1559 }
1560
1561 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
1562 {
1563         char hdr[32];
1564         int hdrlen = sizeof(hdr);
1565
1566         /* Normally if we have it in the pack then we do not bother writing
1567          * it out into .git/objects/??/?{38} file.
1568          */
1569         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
1570         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
1571                 return 0;
1572         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
1573 }
1574
1575 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
1576                              struct object_id *oid, unsigned flags)
1577 {
1578         char *header;
1579         int hdrlen, status = 0;
1580
1581         /* type string, SP, %lu of the length plus NUL must fit this */
1582         hdrlen = strlen(type) + 32;
1583         header = xmalloc(hdrlen);
1584         write_sha1_file_prepare(buf, len, type, oid->hash, header, &hdrlen);
1585
1586         if (!(flags & HASH_WRITE_OBJECT))
1587                 goto cleanup;
1588         if (freshen_packed_object(oid->hash) || freshen_loose_object(oid->hash))
1589                 goto cleanup;
1590         status = write_loose_object(oid->hash, header, hdrlen, buf, len, 0);
1591
1592 cleanup:
1593         free(header);
1594         return status;
1595 }
1596
1597 int force_object_loose(const unsigned char *sha1, time_t mtime)
1598 {
1599         void *buf;
1600         unsigned long len;
1601         enum object_type type;
1602         char hdr[32];
1603         int hdrlen;
1604         int ret;
1605
1606         if (has_loose_object(sha1))
1607                 return 0;
1608         buf = read_object(sha1, &type, &len);
1609         if (!buf)
1610                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
1611         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
1612         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
1613         free(buf);
1614
1615         return ret;
1616 }
1617
1618 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
1619 {
1620         if (!startup_info->have_repository)
1621                 return 0;
1622         return sha1_object_info_extended(sha1, NULL,
1623                                          flags | OBJECT_INFO_SKIP_CACHED) >= 0;
1624 }
1625
1626 int has_object_file(const struct object_id *oid)
1627 {
1628         return has_sha1_file(oid->hash);
1629 }
1630
1631 int has_object_file_with_flags(const struct object_id *oid, int flags)
1632 {
1633         return has_sha1_file_with_flags(oid->hash, flags);
1634 }
1635
1636 static void check_tree(const void *buf, size_t size)
1637 {
1638         struct tree_desc desc;
1639         struct name_entry entry;
1640
1641         init_tree_desc(&desc, buf, size);
1642         while (tree_entry(&desc, &entry))
1643                 /* do nothing
1644                  * tree_entry() will die() on malformed entries */
1645                 ;
1646 }
1647
1648 static void check_commit(const void *buf, size_t size)
1649 {
1650         struct commit c;
1651         memset(&c, 0, sizeof(c));
1652         if (parse_commit_buffer(&c, buf, size))
1653                 die("corrupt commit");
1654 }
1655
1656 static void check_tag(const void *buf, size_t size)
1657 {
1658         struct tag t;
1659         memset(&t, 0, sizeof(t));
1660         if (parse_tag_buffer(&t, buf, size))
1661                 die("corrupt tag");
1662 }
1663
1664 static int index_mem(struct object_id *oid, void *buf, size_t size,
1665                      enum object_type type,
1666                      const char *path, unsigned flags)
1667 {
1668         int ret, re_allocated = 0;
1669         int write_object = flags & HASH_WRITE_OBJECT;
1670
1671         if (!type)
1672                 type = OBJ_BLOB;
1673
1674         /*
1675          * Convert blobs to git internal format
1676          */
1677         if ((type == OBJ_BLOB) && path) {
1678                 struct strbuf nbuf = STRBUF_INIT;
1679                 if (convert_to_git(&the_index, path, buf, size, &nbuf,
1680                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
1681                         buf = strbuf_detach(&nbuf, &size);
1682                         re_allocated = 1;
1683                 }
1684         }
1685         if (flags & HASH_FORMAT_CHECK) {
1686                 if (type == OBJ_TREE)
1687                         check_tree(buf, size);
1688                 if (type == OBJ_COMMIT)
1689                         check_commit(buf, size);
1690                 if (type == OBJ_TAG)
1691                         check_tag(buf, size);
1692         }
1693
1694         if (write_object)
1695                 ret = write_sha1_file(buf, size, typename(type), oid->hash);
1696         else
1697                 ret = hash_sha1_file(buf, size, typename(type), oid->hash);
1698         if (re_allocated)
1699                 free(buf);
1700         return ret;
1701 }
1702
1703 static int index_stream_convert_blob(struct object_id *oid, int fd,
1704                                      const char *path, unsigned flags)
1705 {
1706         int ret;
1707         const int write_object = flags & HASH_WRITE_OBJECT;
1708         struct strbuf sbuf = STRBUF_INIT;
1709
1710         assert(path);
1711         assert(would_convert_to_git_filter_fd(path));
1712
1713         convert_to_git_filter_fd(&the_index, path, fd, &sbuf,
1714                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
1715
1716         if (write_object)
1717                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
1718                                       oid->hash);
1719         else
1720                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
1721                                      oid->hash);
1722         strbuf_release(&sbuf);
1723         return ret;
1724 }
1725
1726 static int index_pipe(struct object_id *oid, int fd, enum object_type type,
1727                       const char *path, unsigned flags)
1728 {
1729         struct strbuf sbuf = STRBUF_INIT;
1730         int ret;
1731
1732         if (strbuf_read(&sbuf, fd, 4096) >= 0)
1733                 ret = index_mem(oid, sbuf.buf, sbuf.len, type, path, flags);
1734         else
1735                 ret = -1;
1736         strbuf_release(&sbuf);
1737         return ret;
1738 }
1739
1740 #define SMALL_FILE_SIZE (32*1024)
1741
1742 static int index_core(struct object_id *oid, int fd, size_t size,
1743                       enum object_type type, const char *path,
1744                       unsigned flags)
1745 {
1746         int ret;
1747
1748         if (!size) {
1749                 ret = index_mem(oid, "", size, type, path, flags);
1750         } else if (size <= SMALL_FILE_SIZE) {
1751                 char *buf = xmalloc(size);
1752                 ssize_t read_result = read_in_full(fd, buf, size);
1753                 if (read_result < 0)
1754                         ret = error_errno("read error while indexing %s",
1755                                           path ? path : "<unknown>");
1756                 else if (read_result != size)
1757                         ret = error("short read while indexing %s",
1758                                     path ? path : "<unknown>");
1759                 else
1760                         ret = index_mem(oid, buf, size, type, path, flags);
1761                 free(buf);
1762         } else {
1763                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
1764                 ret = index_mem(oid, buf, size, type, path, flags);
1765                 munmap(buf, size);
1766         }
1767         return ret;
1768 }
1769
1770 /*
1771  * This creates one packfile per large blob unless bulk-checkin
1772  * machinery is "plugged".
1773  *
1774  * This also bypasses the usual "convert-to-git" dance, and that is on
1775  * purpose. We could write a streaming version of the converting
1776  * functions and insert that before feeding the data to fast-import
1777  * (or equivalent in-core API described above). However, that is
1778  * somewhat complicated, as we do not know the size of the filter
1779  * result, which we need to know beforehand when writing a git object.
1780  * Since the primary motivation for trying to stream from the working
1781  * tree file and to avoid mmaping it in core is to deal with large
1782  * binary blobs, they generally do not want to get any conversion, and
1783  * callers should avoid this code path when filters are requested.
1784  */
1785 static int index_stream(struct object_id *oid, int fd, size_t size,
1786                         enum object_type type, const char *path,
1787                         unsigned flags)
1788 {
1789         return index_bulk_checkin(oid->hash, fd, size, type, path, flags);
1790 }
1791
1792 int index_fd(struct object_id *oid, int fd, struct stat *st,
1793              enum object_type type, const char *path, unsigned flags)
1794 {
1795         int ret;
1796
1797         /*
1798          * Call xsize_t() only when needed to avoid potentially unnecessary
1799          * die() for large files.
1800          */
1801         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
1802                 ret = index_stream_convert_blob(oid, fd, path, flags);
1803         else if (!S_ISREG(st->st_mode))
1804                 ret = index_pipe(oid, fd, type, path, flags);
1805         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
1806                  (path && would_convert_to_git(&the_index, path)))
1807                 ret = index_core(oid, fd, xsize_t(st->st_size), type, path,
1808                                  flags);
1809         else
1810                 ret = index_stream(oid, fd, xsize_t(st->st_size), type, path,
1811                                    flags);
1812         close(fd);
1813         return ret;
1814 }
1815
1816 int index_path(struct object_id *oid, const char *path, struct stat *st, unsigned flags)
1817 {
1818         int fd;
1819         struct strbuf sb = STRBUF_INIT;
1820         int rc = 0;
1821
1822         switch (st->st_mode & S_IFMT) {
1823         case S_IFREG:
1824                 fd = open(path, O_RDONLY);
1825                 if (fd < 0)
1826                         return error_errno("open(\"%s\")", path);
1827                 if (index_fd(oid, fd, st, OBJ_BLOB, path, flags) < 0)
1828                         return error("%s: failed to insert into database",
1829                                      path);
1830                 break;
1831         case S_IFLNK:
1832                 if (strbuf_readlink(&sb, path, st->st_size))
1833                         return error_errno("readlink(\"%s\")", path);
1834                 if (!(flags & HASH_WRITE_OBJECT))
1835                         hash_sha1_file(sb.buf, sb.len, blob_type, oid->hash);
1836                 else if (write_sha1_file(sb.buf, sb.len, blob_type, oid->hash))
1837                         rc = error("%s: failed to insert into database", path);
1838                 strbuf_release(&sb);
1839                 break;
1840         case S_IFDIR:
1841                 return resolve_gitlink_ref(path, "HEAD", oid);
1842         default:
1843                 return error("%s: unsupported file type", path);
1844         }
1845         return rc;
1846 }
1847
1848 int read_pack_header(int fd, struct pack_header *header)
1849 {
1850         if (read_in_full(fd, header, sizeof(*header)) != sizeof(*header))
1851                 /* "eof before pack header was fully read" */
1852                 return PH_ERROR_EOF;
1853
1854         if (header->hdr_signature != htonl(PACK_SIGNATURE))
1855                 /* "protocol error (pack signature mismatch detected)" */
1856                 return PH_ERROR_PACK_SIGNATURE;
1857         if (!pack_version_ok(header->hdr_version))
1858                 /* "protocol error (pack version unsupported)" */
1859                 return PH_ERROR_PROTOCOL;
1860         return 0;
1861 }
1862
1863 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
1864 {
1865         enum object_type type = sha1_object_info(sha1, NULL);
1866         if (type < 0)
1867                 die("%s is not a valid object", sha1_to_hex(sha1));
1868         if (type != expect)
1869                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
1870                     typename(expect));
1871 }
1872
1873 int for_each_file_in_obj_subdir(unsigned int subdir_nr,
1874                                 struct strbuf *path,
1875                                 each_loose_object_fn obj_cb,
1876                                 each_loose_cruft_fn cruft_cb,
1877                                 each_loose_subdir_fn subdir_cb,
1878                                 void *data)
1879 {
1880         size_t origlen, baselen;
1881         DIR *dir;
1882         struct dirent *de;
1883         int r = 0;
1884         struct object_id oid;
1885
1886         if (subdir_nr > 0xff)
1887                 BUG("invalid loose object subdirectory: %x", subdir_nr);
1888
1889         origlen = path->len;
1890         strbuf_complete(path, '/');
1891         strbuf_addf(path, "%02x", subdir_nr);
1892         baselen = path->len;
1893
1894         dir = opendir(path->buf);
1895         if (!dir) {
1896                 if (errno != ENOENT)
1897                         r = error_errno("unable to open %s", path->buf);
1898                 strbuf_setlen(path, origlen);
1899                 return r;
1900         }
1901
1902         oid.hash[0] = subdir_nr;
1903
1904         while ((de = readdir(dir))) {
1905                 if (is_dot_or_dotdot(de->d_name))
1906                         continue;
1907
1908                 strbuf_setlen(path, baselen);
1909                 strbuf_addf(path, "/%s", de->d_name);
1910
1911                 if (strlen(de->d_name) == GIT_SHA1_HEXSZ - 2 &&
1912                     !hex_to_bytes(oid.hash + 1, de->d_name,
1913                                   GIT_SHA1_RAWSZ - 1)) {
1914                         if (obj_cb) {
1915                                 r = obj_cb(&oid, path->buf, data);
1916                                 if (r)
1917                                         break;
1918                         }
1919                         continue;
1920                 }
1921
1922                 if (cruft_cb) {
1923                         r = cruft_cb(de->d_name, path->buf, data);
1924                         if (r)
1925                                 break;
1926                 }
1927         }
1928         closedir(dir);
1929
1930         strbuf_setlen(path, baselen);
1931         if (!r && subdir_cb)
1932                 r = subdir_cb(subdir_nr, path->buf, data);
1933
1934         strbuf_setlen(path, origlen);
1935
1936         return r;
1937 }
1938
1939 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
1940                             each_loose_object_fn obj_cb,
1941                             each_loose_cruft_fn cruft_cb,
1942                             each_loose_subdir_fn subdir_cb,
1943                             void *data)
1944 {
1945         int r = 0;
1946         int i;
1947
1948         for (i = 0; i < 256; i++) {
1949                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
1950                                                 subdir_cb, data);
1951                 if (r)
1952                         break;
1953         }
1954
1955         return r;
1956 }
1957
1958 int for_each_loose_file_in_objdir(const char *path,
1959                                   each_loose_object_fn obj_cb,
1960                                   each_loose_cruft_fn cruft_cb,
1961                                   each_loose_subdir_fn subdir_cb,
1962                                   void *data)
1963 {
1964         struct strbuf buf = STRBUF_INIT;
1965         int r;
1966
1967         strbuf_addstr(&buf, path);
1968         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
1969                                               subdir_cb, data);
1970         strbuf_release(&buf);
1971
1972         return r;
1973 }
1974
1975 struct loose_alt_odb_data {
1976         each_loose_object_fn *cb;
1977         void *data;
1978 };
1979
1980 static int loose_from_alt_odb(struct alternate_object_database *alt,
1981                               void *vdata)
1982 {
1983         struct loose_alt_odb_data *data = vdata;
1984         struct strbuf buf = STRBUF_INIT;
1985         int r;
1986
1987         strbuf_addstr(&buf, alt->path);
1988         r = for_each_loose_file_in_objdir_buf(&buf,
1989                                               data->cb, NULL, NULL,
1990                                               data->data);
1991         strbuf_release(&buf);
1992         return r;
1993 }
1994
1995 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
1996 {
1997         struct loose_alt_odb_data alt;
1998         int r;
1999
2000         r = for_each_loose_file_in_objdir(get_object_directory(),
2001                                           cb, NULL, NULL, data);
2002         if (r)
2003                 return r;
2004
2005         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
2006                 return 0;
2007
2008         alt.cb = cb;
2009         alt.data = data;
2010         return foreach_alt_odb(loose_from_alt_odb, &alt);
2011 }
2012
2013 static int check_stream_sha1(git_zstream *stream,
2014                              const char *hdr,
2015                              unsigned long size,
2016                              const char *path,
2017                              const unsigned char *expected_sha1)
2018 {
2019         git_SHA_CTX c;
2020         unsigned char real_sha1[GIT_MAX_RAWSZ];
2021         unsigned char buf[4096];
2022         unsigned long total_read;
2023         int status = Z_OK;
2024
2025         git_SHA1_Init(&c);
2026         git_SHA1_Update(&c, hdr, stream->total_out);
2027
2028         /*
2029          * We already read some bytes into hdr, but the ones up to the NUL
2030          * do not count against the object's content size.
2031          */
2032         total_read = stream->total_out - strlen(hdr) - 1;
2033
2034         /*
2035          * This size comparison must be "<=" to read the final zlib packets;
2036          * see the comment in unpack_sha1_rest for details.
2037          */
2038         while (total_read <= size &&
2039                (status == Z_OK || status == Z_BUF_ERROR)) {
2040                 stream->next_out = buf;
2041                 stream->avail_out = sizeof(buf);
2042                 if (size - total_read < stream->avail_out)
2043                         stream->avail_out = size - total_read;
2044                 status = git_inflate(stream, Z_FINISH);
2045                 git_SHA1_Update(&c, buf, stream->next_out - buf);
2046                 total_read += stream->next_out - buf;
2047         }
2048         git_inflate_end(stream);
2049
2050         if (status != Z_STREAM_END) {
2051                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
2052                 return -1;
2053         }
2054         if (stream->avail_in) {
2055                 error("garbage at end of loose object '%s'",
2056                       sha1_to_hex(expected_sha1));
2057                 return -1;
2058         }
2059
2060         git_SHA1_Final(real_sha1, &c);
2061         if (hashcmp(expected_sha1, real_sha1)) {
2062                 error("sha1 mismatch for %s (expected %s)", path,
2063                       sha1_to_hex(expected_sha1));
2064                 return -1;
2065         }
2066
2067         return 0;
2068 }
2069
2070 int read_loose_object(const char *path,
2071                       const unsigned char *expected_sha1,
2072                       enum object_type *type,
2073                       unsigned long *size,
2074                       void **contents)
2075 {
2076         int ret = -1;
2077         void *map = NULL;
2078         unsigned long mapsize;
2079         git_zstream stream;
2080         char hdr[32];
2081
2082         *contents = NULL;
2083
2084         map = map_sha1_file_1(path, NULL, &mapsize);
2085         if (!map) {
2086                 error_errno("unable to mmap %s", path);
2087                 goto out;
2088         }
2089
2090         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
2091                 error("unable to unpack header of %s", path);
2092                 goto out;
2093         }
2094
2095         *type = parse_sha1_header(hdr, size);
2096         if (*type < 0) {
2097                 error("unable to parse header of %s", path);
2098                 git_inflate_end(&stream);
2099                 goto out;
2100         }
2101
2102         if (*type == OBJ_BLOB) {
2103                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
2104                         goto out;
2105         } else {
2106                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
2107                 if (!*contents) {
2108                         error("unable to unpack contents of %s", path);
2109                         git_inflate_end(&stream);
2110                         goto out;
2111                 }
2112                 if (check_sha1_signature(expected_sha1, *contents,
2113                                          *size, typename(*type))) {
2114                         error("sha1 mismatch for %s (expected %s)", path,
2115                               sha1_to_hex(expected_sha1));
2116                         free(*contents);
2117                         goto out;
2118                 }
2119         }
2120
2121         ret = 0; /* everything checks out */
2122
2123 out:
2124         if (map)
2125                 munmap(map, mapsize);
2126         return ret;
2127 }