Merge branch 'rs/apply-fuzzy-match-fix'
[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 (!alt || !*alt)
408                 return;
409
410         if (depth > 5) {
411                 error("%s: ignoring alternate object stores, nesting too deep.",
412                                 relative_base);
413                 return;
414         }
415
416         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
417         if (strbuf_normalize_path(&objdirbuf) < 0)
418                 die("unable to normalize object directory: %s",
419                     objdirbuf.buf);
420
421         while (*alt) {
422                 alt = parse_alt_odb_entry(alt, sep, &entry);
423                 if (!entry.len)
424                         continue;
425                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
426         }
427         strbuf_release(&entry);
428         strbuf_release(&objdirbuf);
429 }
430
431 static void read_info_alternates(const char * relative_base, int depth)
432 {
433         char *path;
434         struct strbuf buf = STRBUF_INIT;
435
436         path = xstrfmt("%s/info/alternates", relative_base);
437         if (strbuf_read_file(&buf, path, 1024) < 0) {
438                 warn_on_fopen_errors(path);
439                 free(path);
440                 return;
441         }
442
443         link_alt_odb_entries(buf.buf, '\n', relative_base, depth);
444         strbuf_release(&buf);
445         free(path);
446 }
447
448 struct alternate_object_database *alloc_alt_odb(const char *dir)
449 {
450         struct alternate_object_database *ent;
451
452         FLEX_ALLOC_STR(ent, path, dir);
453         strbuf_init(&ent->scratch, 0);
454         strbuf_addf(&ent->scratch, "%s/", dir);
455         ent->base_len = ent->scratch.len;
456
457         return ent;
458 }
459
460 void add_to_alternates_file(const char *reference)
461 {
462         struct lock_file lock = LOCK_INIT;
463         char *alts = git_pathdup("objects/info/alternates");
464         FILE *in, *out;
465         int found = 0;
466
467         hold_lock_file_for_update(&lock, alts, LOCK_DIE_ON_ERROR);
468         out = fdopen_lock_file(&lock, "w");
469         if (!out)
470                 die_errno("unable to fdopen alternates lockfile");
471
472         in = fopen(alts, "r");
473         if (in) {
474                 struct strbuf line = STRBUF_INIT;
475
476                 while (strbuf_getline(&line, in) != EOF) {
477                         if (!strcmp(reference, line.buf)) {
478                                 found = 1;
479                                 break;
480                         }
481                         fprintf_or_die(out, "%s\n", line.buf);
482                 }
483
484                 strbuf_release(&line);
485                 fclose(in);
486         }
487         else if (errno != ENOENT)
488                 die_errno("unable to read alternates file");
489
490         if (found) {
491                 rollback_lock_file(&lock);
492         } else {
493                 fprintf_or_die(out, "%s\n", reference);
494                 if (commit_lock_file(&lock))
495                         die_errno("unable to move new alternates file into place");
496                 if (alt_odb_tail)
497                         link_alt_odb_entries(reference, '\n', NULL, 0);
498         }
499         free(alts);
500 }
501
502 void add_to_alternates_memory(const char *reference)
503 {
504         /*
505          * Make sure alternates are initialized, or else our entry may be
506          * overwritten when they are.
507          */
508         prepare_alt_odb();
509
510         link_alt_odb_entries(reference, '\n', NULL, 0);
511 }
512
513 /*
514  * Compute the exact path an alternate is at and returns it. In case of
515  * error NULL is returned and the human readable error is added to `err`
516  * `path` may be relative and should point to $GITDIR.
517  * `err` must not be null.
518  */
519 char *compute_alternate_path(const char *path, struct strbuf *err)
520 {
521         char *ref_git = NULL;
522         const char *repo, *ref_git_s;
523         int seen_error = 0;
524
525         ref_git_s = real_path_if_valid(path);
526         if (!ref_git_s) {
527                 seen_error = 1;
528                 strbuf_addf(err, _("path '%s' does not exist"), path);
529                 goto out;
530         } else
531                 /*
532                  * Beware: read_gitfile(), real_path() and mkpath()
533                  * return static buffer
534                  */
535                 ref_git = xstrdup(ref_git_s);
536
537         repo = read_gitfile(ref_git);
538         if (!repo)
539                 repo = read_gitfile(mkpath("%s/.git", ref_git));
540         if (repo) {
541                 free(ref_git);
542                 ref_git = xstrdup(repo);
543         }
544
545         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
546                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
547                 free(ref_git);
548                 ref_git = ref_git_git;
549         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
550                 struct strbuf sb = STRBUF_INIT;
551                 seen_error = 1;
552                 if (get_common_dir(&sb, ref_git)) {
553                         strbuf_addf(err,
554                                     _("reference repository '%s' as a linked "
555                                       "checkout is not supported yet."),
556                                     path);
557                         goto out;
558                 }
559
560                 strbuf_addf(err, _("reference repository '%s' is not a "
561                                         "local repository."), path);
562                 goto out;
563         }
564
565         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
566                 strbuf_addf(err, _("reference repository '%s' is shallow"),
567                             path);
568                 seen_error = 1;
569                 goto out;
570         }
571
572         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
573                 strbuf_addf(err,
574                             _("reference repository '%s' is grafted"),
575                             path);
576                 seen_error = 1;
577                 goto out;
578         }
579
580 out:
581         if (seen_error) {
582                 FREE_AND_NULL(ref_git);
583         }
584
585         return ref_git;
586 }
587
588 int foreach_alt_odb(alt_odb_fn fn, void *cb)
589 {
590         struct alternate_object_database *ent;
591         int r = 0;
592
593         prepare_alt_odb();
594         for (ent = alt_odb_list; ent; ent = ent->next) {
595                 r = fn(ent, cb);
596                 if (r)
597                         break;
598         }
599         return r;
600 }
601
602 void prepare_alt_odb(void)
603 {
604         const char *alt;
605
606         if (alt_odb_tail)
607                 return;
608
609         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
610
611         alt_odb_tail = &alt_odb_list;
612         link_alt_odb_entries(alt, PATH_SEP, NULL, 0);
613
614         read_info_alternates(get_object_directory(), 0);
615 }
616
617 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
618 static int freshen_file(const char *fn)
619 {
620         struct utimbuf t;
621         t.actime = t.modtime = time(NULL);
622         return !utime(fn, &t);
623 }
624
625 /*
626  * All of the check_and_freshen functions return 1 if the file exists and was
627  * freshened (if freshening was requested), 0 otherwise. If they return
628  * 0, you should not assume that it is safe to skip a write of the object (it
629  * either does not exist on disk, or has a stale mtime and may be subject to
630  * pruning).
631  */
632 int check_and_freshen_file(const char *fn, int freshen)
633 {
634         if (access(fn, F_OK))
635                 return 0;
636         if (freshen && !freshen_file(fn))
637                 return 0;
638         return 1;
639 }
640
641 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
642 {
643         return check_and_freshen_file(sha1_file_name(sha1), freshen);
644 }
645
646 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
647 {
648         struct alternate_object_database *alt;
649         prepare_alt_odb();
650         for (alt = alt_odb_list; alt; alt = alt->next) {
651                 const char *path = alt_sha1_path(alt, sha1);
652                 if (check_and_freshen_file(path, freshen))
653                         return 1;
654         }
655         return 0;
656 }
657
658 static int check_and_freshen(const unsigned char *sha1, int freshen)
659 {
660         return check_and_freshen_local(sha1, freshen) ||
661                check_and_freshen_nonlocal(sha1, freshen);
662 }
663
664 int has_loose_object_nonlocal(const unsigned char *sha1)
665 {
666         return check_and_freshen_nonlocal(sha1, 0);
667 }
668
669 static int has_loose_object(const unsigned char *sha1)
670 {
671         return check_and_freshen(sha1, 0);
672 }
673
674 static void mmap_limit_check(size_t length)
675 {
676         static size_t limit = 0;
677         if (!limit) {
678                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
679                 if (!limit)
680                         limit = SIZE_MAX;
681         }
682         if (length > limit)
683                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
684                     (uintmax_t)length, (uintmax_t)limit);
685 }
686
687 void *xmmap_gently(void *start, size_t length,
688                   int prot, int flags, int fd, off_t offset)
689 {
690         void *ret;
691
692         mmap_limit_check(length);
693         ret = mmap(start, length, prot, flags, fd, offset);
694         if (ret == MAP_FAILED) {
695                 if (!length)
696                         return NULL;
697                 release_pack_memory(length);
698                 ret = mmap(start, length, prot, flags, fd, offset);
699         }
700         return ret;
701 }
702
703 void *xmmap(void *start, size_t length,
704         int prot, int flags, int fd, off_t offset)
705 {
706         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
707         if (ret == MAP_FAILED)
708                 die_errno("mmap failed");
709         return ret;
710 }
711
712 /*
713  * With an in-core object data in "map", rehash it to make sure the
714  * object name actually matches "sha1" to detect object corruption.
715  * With "map" == NULL, try reading the object named with "sha1" using
716  * the streaming interface and rehash it to do the same.
717  */
718 int check_sha1_signature(const unsigned char *sha1, void *map,
719                          unsigned long size, const char *type)
720 {
721         unsigned char real_sha1[20];
722         enum object_type obj_type;
723         struct git_istream *st;
724         git_SHA_CTX c;
725         char hdr[32];
726         int hdrlen;
727
728         if (map) {
729                 hash_sha1_file(map, size, type, real_sha1);
730                 return hashcmp(sha1, real_sha1) ? -1 : 0;
731         }
732
733         st = open_istream(sha1, &obj_type, &size, NULL);
734         if (!st)
735                 return -1;
736
737         /* Generate the header */
738         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
739
740         /* Sha1.. */
741         git_SHA1_Init(&c);
742         git_SHA1_Update(&c, hdr, hdrlen);
743         for (;;) {
744                 char buf[1024 * 16];
745                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
746
747                 if (readlen < 0) {
748                         close_istream(st);
749                         return -1;
750                 }
751                 if (!readlen)
752                         break;
753                 git_SHA1_Update(&c, buf, readlen);
754         }
755         git_SHA1_Final(real_sha1, &c);
756         close_istream(st);
757         return hashcmp(sha1, real_sha1) ? -1 : 0;
758 }
759
760 int git_open_cloexec(const char *name, int flags)
761 {
762         int fd;
763         static int o_cloexec = O_CLOEXEC;
764
765         fd = open(name, flags | o_cloexec);
766         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
767                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
768                 o_cloexec &= ~O_CLOEXEC;
769                 fd = open(name, flags | o_cloexec);
770         }
771
772 #if defined(F_GETFD) && defined(F_SETFD) && defined(FD_CLOEXEC)
773         {
774                 static int fd_cloexec = FD_CLOEXEC;
775
776                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
777                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
778                         int flags = fcntl(fd, F_GETFD);
779                         if (fcntl(fd, F_SETFD, flags | fd_cloexec))
780                                 fd_cloexec = 0;
781                 }
782         }
783 #endif
784         return fd;
785 }
786
787 /*
788  * Find "sha1" as a loose object in the local repository or in an alternate.
789  * Returns 0 on success, negative on failure.
790  *
791  * The "path" out-parameter will give the path of the object we found (if any).
792  * Note that it may point to static storage and is only valid until another
793  * call to sha1_file_name(), etc.
794  */
795 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
796                           const char **path)
797 {
798         struct alternate_object_database *alt;
799
800         *path = sha1_file_name(sha1);
801         if (!lstat(*path, st))
802                 return 0;
803
804         prepare_alt_odb();
805         errno = ENOENT;
806         for (alt = alt_odb_list; alt; alt = alt->next) {
807                 *path = alt_sha1_path(alt, sha1);
808                 if (!lstat(*path, st))
809                         return 0;
810         }
811
812         return -1;
813 }
814
815 /*
816  * Like stat_sha1_file(), but actually open the object and return the
817  * descriptor. See the caveats on the "path" parameter above.
818  */
819 static int open_sha1_file(const unsigned char *sha1, const char **path)
820 {
821         int fd;
822         struct alternate_object_database *alt;
823         int most_interesting_errno;
824
825         *path = sha1_file_name(sha1);
826         fd = git_open(*path);
827         if (fd >= 0)
828                 return fd;
829         most_interesting_errno = errno;
830
831         prepare_alt_odb();
832         for (alt = alt_odb_list; alt; alt = alt->next) {
833                 *path = alt_sha1_path(alt, sha1);
834                 fd = git_open(*path);
835                 if (fd >= 0)
836                         return fd;
837                 if (most_interesting_errno == ENOENT)
838                         most_interesting_errno = errno;
839         }
840         errno = most_interesting_errno;
841         return -1;
842 }
843
844 /*
845  * Map the loose object at "path" if it is not NULL, or the path found by
846  * searching for a loose object named "sha1".
847  */
848 static void *map_sha1_file_1(const char *path,
849                              const unsigned char *sha1,
850                              unsigned long *size)
851 {
852         void *map;
853         int fd;
854
855         if (path)
856                 fd = git_open(path);
857         else
858                 fd = open_sha1_file(sha1, &path);
859         map = NULL;
860         if (fd >= 0) {
861                 struct stat st;
862
863                 if (!fstat(fd, &st)) {
864                         *size = xsize_t(st.st_size);
865                         if (!*size) {
866                                 /* mmap() is forbidden on empty files */
867                                 error("object file %s is empty", path);
868                                 return NULL;
869                         }
870                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
871                 }
872                 close(fd);
873         }
874         return map;
875 }
876
877 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
878 {
879         return map_sha1_file_1(NULL, sha1, size);
880 }
881
882 static int unpack_sha1_short_header(git_zstream *stream,
883                                     unsigned char *map, unsigned long mapsize,
884                                     void *buffer, unsigned long bufsiz)
885 {
886         /* Get the data stream */
887         memset(stream, 0, sizeof(*stream));
888         stream->next_in = map;
889         stream->avail_in = mapsize;
890         stream->next_out = buffer;
891         stream->avail_out = bufsiz;
892
893         git_inflate_init(stream);
894         return git_inflate(stream, 0);
895 }
896
897 int unpack_sha1_header(git_zstream *stream,
898                        unsigned char *map, unsigned long mapsize,
899                        void *buffer, unsigned long bufsiz)
900 {
901         int status = unpack_sha1_short_header(stream, map, mapsize,
902                                               buffer, bufsiz);
903
904         if (status < Z_OK)
905                 return status;
906
907         /* Make sure we have the terminating NUL */
908         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
909                 return -1;
910         return 0;
911 }
912
913 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
914                                         unsigned long mapsize, void *buffer,
915                                         unsigned long bufsiz, struct strbuf *header)
916 {
917         int status;
918
919         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
920         if (status < Z_OK)
921                 return -1;
922
923         /*
924          * Check if entire header is unpacked in the first iteration.
925          */
926         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
927                 return 0;
928
929         /*
930          * buffer[0..bufsiz] was not large enough.  Copy the partial
931          * result out to header, and then append the result of further
932          * reading the stream.
933          */
934         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
935         stream->next_out = buffer;
936         stream->avail_out = bufsiz;
937
938         do {
939                 status = git_inflate(stream, 0);
940                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
941                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
942                         return 0;
943                 stream->next_out = buffer;
944                 stream->avail_out = bufsiz;
945         } while (status != Z_STREAM_END);
946         return -1;
947 }
948
949 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
950 {
951         int bytes = strlen(buffer) + 1;
952         unsigned char *buf = xmallocz(size);
953         unsigned long n;
954         int status = Z_OK;
955
956         n = stream->total_out - bytes;
957         if (n > size)
958                 n = size;
959         memcpy(buf, (char *) buffer + bytes, n);
960         bytes = n;
961         if (bytes <= size) {
962                 /*
963                  * The above condition must be (bytes <= size), not
964                  * (bytes < size).  In other words, even though we
965                  * expect no more output and set avail_out to zero,
966                  * the input zlib stream may have bytes that express
967                  * "this concludes the stream", and we *do* want to
968                  * eat that input.
969                  *
970                  * Otherwise we would not be able to test that we
971                  * consumed all the input to reach the expected size;
972                  * we also want to check that zlib tells us that all
973                  * went well with status == Z_STREAM_END at the end.
974                  */
975                 stream->next_out = buf + bytes;
976                 stream->avail_out = size - bytes;
977                 while (status == Z_OK)
978                         status = git_inflate(stream, Z_FINISH);
979         }
980         if (status == Z_STREAM_END && !stream->avail_in) {
981                 git_inflate_end(stream);
982                 return buf;
983         }
984
985         if (status < 0)
986                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
987         else if (stream->avail_in)
988                 error("garbage at end of loose object '%s'",
989                       sha1_to_hex(sha1));
990         free(buf);
991         return NULL;
992 }
993
994 /*
995  * We used to just use "sscanf()", but that's actually way
996  * too permissive for what we want to check. So do an anal
997  * object header parse by hand.
998  */
999 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1000                                unsigned int flags)
1001 {
1002         const char *type_buf = hdr;
1003         unsigned long size;
1004         int type, type_len = 0;
1005
1006         /*
1007          * The type can be of any size but is followed by
1008          * a space.
1009          */
1010         for (;;) {
1011                 char c = *hdr++;
1012                 if (!c)
1013                         return -1;
1014                 if (c == ' ')
1015                         break;
1016                 type_len++;
1017         }
1018
1019         type = type_from_string_gently(type_buf, type_len, 1);
1020         if (oi->typename)
1021                 strbuf_add(oi->typename, type_buf, type_len);
1022         /*
1023          * Set type to 0 if its an unknown object and
1024          * we're obtaining the type using '--allow-unknown-type'
1025          * option.
1026          */
1027         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE) && (type < 0))
1028                 type = 0;
1029         else if (type < 0)
1030                 die("invalid object type");
1031         if (oi->typep)
1032                 *oi->typep = type;
1033
1034         /*
1035          * The length must follow immediately, and be in canonical
1036          * decimal format (ie "010" is not valid).
1037          */
1038         size = *hdr++ - '0';
1039         if (size > 9)
1040                 return -1;
1041         if (size) {
1042                 for (;;) {
1043                         unsigned long c = *hdr - '0';
1044                         if (c > 9)
1045                                 break;
1046                         hdr++;
1047                         size = size * 10 + c;
1048                 }
1049         }
1050
1051         if (oi->sizep)
1052                 *oi->sizep = size;
1053
1054         /*
1055          * The length must be followed by a zero byte
1056          */
1057         return *hdr ? -1 : type;
1058 }
1059
1060 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1061 {
1062         struct object_info oi = OBJECT_INFO_INIT;
1063
1064         oi.sizep = sizep;
1065         return parse_sha1_header_extended(hdr, &oi, 0);
1066 }
1067
1068 static int sha1_loose_object_info(const unsigned char *sha1,
1069                                   struct object_info *oi,
1070                                   int flags)
1071 {
1072         int status = 0;
1073         unsigned long mapsize;
1074         void *map;
1075         git_zstream stream;
1076         char hdr[32];
1077         struct strbuf hdrbuf = STRBUF_INIT;
1078         unsigned long size_scratch;
1079
1080         if (oi->delta_base_sha1)
1081                 hashclr(oi->delta_base_sha1);
1082
1083         /*
1084          * If we don't care about type or size, then we don't
1085          * need to look inside the object at all. Note that we
1086          * do not optimize out the stat call, even if the
1087          * caller doesn't care about the disk-size, since our
1088          * return value implicitly indicates whether the
1089          * object even exists.
1090          */
1091         if (!oi->typep && !oi->typename && !oi->sizep && !oi->contentp) {
1092                 const char *path;
1093                 struct stat st;
1094                 if (stat_sha1_file(sha1, &st, &path) < 0)
1095                         return -1;
1096                 if (oi->disk_sizep)
1097                         *oi->disk_sizep = st.st_size;
1098                 return 0;
1099         }
1100
1101         map = map_sha1_file(sha1, &mapsize);
1102         if (!map)
1103                 return -1;
1104
1105         if (!oi->sizep)
1106                 oi->sizep = &size_scratch;
1107
1108         if (oi->disk_sizep)
1109                 *oi->disk_sizep = mapsize;
1110         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE)) {
1111                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
1112                         status = error("unable to unpack %s header with --allow-unknown-type",
1113                                        sha1_to_hex(sha1));
1114         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1115                 status = error("unable to unpack %s header",
1116                                sha1_to_hex(sha1));
1117         if (status < 0)
1118                 ; /* Do nothing */
1119         else if (hdrbuf.len) {
1120                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
1121                         status = error("unable to parse %s header with --allow-unknown-type",
1122                                        sha1_to_hex(sha1));
1123         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
1124                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1125
1126         if (status >= 0 && oi->contentp) {
1127                 *oi->contentp = unpack_sha1_rest(&stream, hdr,
1128                                                  *oi->sizep, sha1);
1129                 if (!*oi->contentp) {
1130                         git_inflate_end(&stream);
1131                         status = -1;
1132                 }
1133         } else
1134                 git_inflate_end(&stream);
1135
1136         munmap(map, mapsize);
1137         if (status && oi->typep)
1138                 *oi->typep = status;
1139         if (oi->sizep == &size_scratch)
1140                 oi->sizep = NULL;
1141         strbuf_release(&hdrbuf);
1142         oi->whence = OI_LOOSE;
1143         return (status < 0) ? status : 0;
1144 }
1145
1146 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
1147 {
1148         static struct object_info blank_oi = OBJECT_INFO_INIT;
1149         struct pack_entry e;
1150         int rtype;
1151         const unsigned char *real = (flags & OBJECT_INFO_LOOKUP_REPLACE) ?
1152                                     lookup_replace_object(sha1) :
1153                                     sha1;
1154
1155         if (!oi)
1156                 oi = &blank_oi;
1157
1158         if (!(flags & OBJECT_INFO_SKIP_CACHED)) {
1159                 struct cached_object *co = find_cached_object(real);
1160                 if (co) {
1161                         if (oi->typep)
1162                                 *(oi->typep) = co->type;
1163                         if (oi->sizep)
1164                                 *(oi->sizep) = co->size;
1165                         if (oi->disk_sizep)
1166                                 *(oi->disk_sizep) = 0;
1167                         if (oi->delta_base_sha1)
1168                                 hashclr(oi->delta_base_sha1);
1169                         if (oi->typename)
1170                                 strbuf_addstr(oi->typename, typename(co->type));
1171                         if (oi->contentp)
1172                                 *oi->contentp = xmemdupz(co->buf, co->size);
1173                         oi->whence = OI_CACHED;
1174                         return 0;
1175                 }
1176         }
1177
1178         if (!find_pack_entry(real, &e)) {
1179                 /* Most likely it's a loose object. */
1180                 if (!sha1_loose_object_info(real, oi, flags))
1181                         return 0;
1182
1183                 /* Not a loose object; someone else may have just packed it. */
1184                 if (flags & OBJECT_INFO_QUICK) {
1185                         return -1;
1186                 } else {
1187                         reprepare_packed_git();
1188                         if (!find_pack_entry(real, &e))
1189                                 return -1;
1190                 }
1191         }
1192
1193         if (oi == &blank_oi)
1194                 /*
1195                  * We know that the caller doesn't actually need the
1196                  * information below, so return early.
1197                  */
1198                 return 0;
1199
1200         rtype = packed_object_info(e.p, e.offset, oi);
1201         if (rtype < 0) {
1202                 mark_bad_packed_object(e.p, real);
1203                 return sha1_object_info_extended(real, oi, 0);
1204         } else if (oi->whence == OI_PACKED) {
1205                 oi->u.packed.offset = e.offset;
1206                 oi->u.packed.pack = e.p;
1207                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
1208                                          rtype == OBJ_OFS_DELTA);
1209         }
1210
1211         return 0;
1212 }
1213
1214 /* returns enum object_type or negative */
1215 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1216 {
1217         enum object_type type;
1218         struct object_info oi = OBJECT_INFO_INIT;
1219
1220         oi.typep = &type;
1221         oi.sizep = sizep;
1222         if (sha1_object_info_extended(sha1, &oi,
1223                                       OBJECT_INFO_LOOKUP_REPLACE) < 0)
1224                 return -1;
1225         return type;
1226 }
1227
1228 static void *read_object(const unsigned char *sha1, enum object_type *type,
1229                          unsigned long *size)
1230 {
1231         struct object_info oi = OBJECT_INFO_INIT;
1232         void *content;
1233         oi.typep = type;
1234         oi.sizep = size;
1235         oi.contentp = &content;
1236
1237         if (sha1_object_info_extended(sha1, &oi, 0) < 0)
1238                 return NULL;
1239         return content;
1240 }
1241
1242 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
1243                       unsigned char *sha1)
1244 {
1245         struct cached_object *co;
1246
1247         hash_sha1_file(buf, len, typename(type), sha1);
1248         if (has_sha1_file(sha1) || find_cached_object(sha1))
1249                 return 0;
1250         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
1251         co = &cached_objects[cached_object_nr++];
1252         co->size = len;
1253         co->type = type;
1254         co->buf = xmalloc(len);
1255         memcpy(co->buf, buf, len);
1256         hashcpy(co->sha1, sha1);
1257         return 0;
1258 }
1259
1260 /*
1261  * This function dies on corrupt objects; the callers who want to
1262  * deal with them should arrange to call read_object() and give error
1263  * messages themselves.
1264  */
1265 void *read_sha1_file_extended(const unsigned char *sha1,
1266                               enum object_type *type,
1267                               unsigned long *size,
1268                               int lookup_replace)
1269 {
1270         void *data;
1271         const struct packed_git *p;
1272         const char *path;
1273         struct stat st;
1274         const unsigned char *repl = lookup_replace ? lookup_replace_object(sha1)
1275                                                    : sha1;
1276
1277         errno = 0;
1278         data = read_object(repl, type, size);
1279         if (data)
1280                 return data;
1281
1282         if (errno && errno != ENOENT)
1283                 die_errno("failed to read object %s", sha1_to_hex(sha1));
1284
1285         /* die if we replaced an object with one that does not exist */
1286         if (repl != sha1)
1287                 die("replacement %s not found for %s",
1288                     sha1_to_hex(repl), sha1_to_hex(sha1));
1289
1290         if (!stat_sha1_file(repl, &st, &path))
1291                 die("loose object %s (stored in %s) is corrupt",
1292                     sha1_to_hex(repl), path);
1293
1294         if ((p = has_packed_and_bad(repl)) != NULL)
1295                 die("packed object %s (stored in %s) is corrupt",
1296                     sha1_to_hex(repl), p->pack_name);
1297
1298         return NULL;
1299 }
1300
1301 void *read_object_with_reference(const unsigned char *sha1,
1302                                  const char *required_type_name,
1303                                  unsigned long *size,
1304                                  unsigned char *actual_sha1_return)
1305 {
1306         enum object_type type, required_type;
1307         void *buffer;
1308         unsigned long isize;
1309         unsigned char actual_sha1[20];
1310
1311         required_type = type_from_string(required_type_name);
1312         hashcpy(actual_sha1, sha1);
1313         while (1) {
1314                 int ref_length = -1;
1315                 const char *ref_type = NULL;
1316
1317                 buffer = read_sha1_file(actual_sha1, &type, &isize);
1318                 if (!buffer)
1319                         return NULL;
1320                 if (type == required_type) {
1321                         *size = isize;
1322                         if (actual_sha1_return)
1323                                 hashcpy(actual_sha1_return, actual_sha1);
1324                         return buffer;
1325                 }
1326                 /* Handle references */
1327                 else if (type == OBJ_COMMIT)
1328                         ref_type = "tree ";
1329                 else if (type == OBJ_TAG)
1330                         ref_type = "object ";
1331                 else {
1332                         free(buffer);
1333                         return NULL;
1334                 }
1335                 ref_length = strlen(ref_type);
1336
1337                 if (ref_length + 40 > isize ||
1338                     memcmp(buffer, ref_type, ref_length) ||
1339                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
1340                         free(buffer);
1341                         return NULL;
1342                 }
1343                 free(buffer);
1344                 /* Now we have the ID of the referred-to object in
1345                  * actual_sha1.  Check again. */
1346         }
1347 }
1348
1349 static void write_sha1_file_prepare(const void *buf, unsigned long len,
1350                                     const char *type, unsigned char *sha1,
1351                                     char *hdr, int *hdrlen)
1352 {
1353         git_SHA_CTX c;
1354
1355         /* Generate the header */
1356         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
1357
1358         /* Sha1.. */
1359         git_SHA1_Init(&c);
1360         git_SHA1_Update(&c, hdr, *hdrlen);
1361         git_SHA1_Update(&c, buf, len);
1362         git_SHA1_Final(sha1, &c);
1363 }
1364
1365 /*
1366  * Move the just written object into its final resting place.
1367  */
1368 int finalize_object_file(const char *tmpfile, const char *filename)
1369 {
1370         int ret = 0;
1371
1372         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
1373                 goto try_rename;
1374         else if (link(tmpfile, filename))
1375                 ret = errno;
1376
1377         /*
1378          * Coda hack - coda doesn't like cross-directory links,
1379          * so we fall back to a rename, which will mean that it
1380          * won't be able to check collisions, but that's not a
1381          * big deal.
1382          *
1383          * The same holds for FAT formatted media.
1384          *
1385          * When this succeeds, we just return.  We have nothing
1386          * left to unlink.
1387          */
1388         if (ret && ret != EEXIST) {
1389         try_rename:
1390                 if (!rename(tmpfile, filename))
1391                         goto out;
1392                 ret = errno;
1393         }
1394         unlink_or_warn(tmpfile);
1395         if (ret) {
1396                 if (ret != EEXIST) {
1397                         return error_errno("unable to write sha1 filename %s", filename);
1398                 }
1399                 /* FIXME!!! Collision check here ? */
1400         }
1401
1402 out:
1403         if (adjust_shared_perm(filename))
1404                 return error("unable to set permission to '%s'", filename);
1405         return 0;
1406 }
1407
1408 static int write_buffer(int fd, const void *buf, size_t len)
1409 {
1410         if (write_in_full(fd, buf, len) < 0)
1411                 return error_errno("file write error");
1412         return 0;
1413 }
1414
1415 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
1416                    unsigned char *sha1)
1417 {
1418         char hdr[32];
1419         int hdrlen = sizeof(hdr);
1420         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
1421         return 0;
1422 }
1423
1424 /* Finalize a file on disk, and close it. */
1425 static void close_sha1_file(int fd)
1426 {
1427         if (fsync_object_files)
1428                 fsync_or_die(fd, "sha1 file");
1429         if (close(fd) != 0)
1430                 die_errno("error when closing sha1 file");
1431 }
1432
1433 /* Size of directory component, including the ending '/' */
1434 static inline int directory_size(const char *filename)
1435 {
1436         const char *s = strrchr(filename, '/');
1437         if (!s)
1438                 return 0;
1439         return s - filename + 1;
1440 }
1441
1442 /*
1443  * This creates a temporary file in the same directory as the final
1444  * 'filename'
1445  *
1446  * We want to avoid cross-directory filename renames, because those
1447  * can have problems on various filesystems (FAT, NFS, Coda).
1448  */
1449 static int create_tmpfile(struct strbuf *tmp, const char *filename)
1450 {
1451         int fd, dirlen = directory_size(filename);
1452
1453         strbuf_reset(tmp);
1454         strbuf_add(tmp, filename, dirlen);
1455         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
1456         fd = git_mkstemp_mode(tmp->buf, 0444);
1457         if (fd < 0 && dirlen && errno == ENOENT) {
1458                 /*
1459                  * Make sure the directory exists; note that the contents
1460                  * of the buffer are undefined after mkstemp returns an
1461                  * error, so we have to rewrite the whole buffer from
1462                  * scratch.
1463                  */
1464                 strbuf_reset(tmp);
1465                 strbuf_add(tmp, filename, dirlen - 1);
1466                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
1467                         return -1;
1468                 if (adjust_shared_perm(tmp->buf))
1469                         return -1;
1470
1471                 /* Try again */
1472                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
1473                 fd = git_mkstemp_mode(tmp->buf, 0444);
1474         }
1475         return fd;
1476 }
1477
1478 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
1479                               const void *buf, unsigned long len, time_t mtime)
1480 {
1481         int fd, ret;
1482         unsigned char compressed[4096];
1483         git_zstream stream;
1484         git_SHA_CTX c;
1485         unsigned char parano_sha1[20];
1486         static struct strbuf tmp_file = STRBUF_INIT;
1487         const char *filename = sha1_file_name(sha1);
1488
1489         fd = create_tmpfile(&tmp_file, filename);
1490         if (fd < 0) {
1491                 if (errno == EACCES)
1492                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
1493                 else
1494                         return error_errno("unable to create temporary file");
1495         }
1496
1497         /* Set it up */
1498         git_deflate_init(&stream, zlib_compression_level);
1499         stream.next_out = compressed;
1500         stream.avail_out = sizeof(compressed);
1501         git_SHA1_Init(&c);
1502
1503         /* First header.. */
1504         stream.next_in = (unsigned char *)hdr;
1505         stream.avail_in = hdrlen;
1506         while (git_deflate(&stream, 0) == Z_OK)
1507                 ; /* nothing */
1508         git_SHA1_Update(&c, hdr, hdrlen);
1509
1510         /* Then the data itself.. */
1511         stream.next_in = (void *)buf;
1512         stream.avail_in = len;
1513         do {
1514                 unsigned char *in0 = stream.next_in;
1515                 ret = git_deflate(&stream, Z_FINISH);
1516                 git_SHA1_Update(&c, in0, stream.next_in - in0);
1517                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
1518                         die("unable to write sha1 file");
1519                 stream.next_out = compressed;
1520                 stream.avail_out = sizeof(compressed);
1521         } while (ret == Z_OK);
1522
1523         if (ret != Z_STREAM_END)
1524                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
1525         ret = git_deflate_end_gently(&stream);
1526         if (ret != Z_OK)
1527                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
1528         git_SHA1_Final(parano_sha1, &c);
1529         if (hashcmp(sha1, parano_sha1) != 0)
1530                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
1531
1532         close_sha1_file(fd);
1533
1534         if (mtime) {
1535                 struct utimbuf utb;
1536                 utb.actime = mtime;
1537                 utb.modtime = mtime;
1538                 if (utime(tmp_file.buf, &utb) < 0)
1539                         warning_errno("failed utime() on %s", tmp_file.buf);
1540         }
1541
1542         return finalize_object_file(tmp_file.buf, filename);
1543 }
1544
1545 static int freshen_loose_object(const unsigned char *sha1)
1546 {
1547         return check_and_freshen(sha1, 1);
1548 }
1549
1550 static int freshen_packed_object(const unsigned char *sha1)
1551 {
1552         struct pack_entry e;
1553         if (!find_pack_entry(sha1, &e))
1554                 return 0;
1555         if (e.p->freshened)
1556                 return 1;
1557         if (!freshen_file(e.p->pack_name))
1558                 return 0;
1559         e.p->freshened = 1;
1560         return 1;
1561 }
1562
1563 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
1564 {
1565         char hdr[32];
1566         int hdrlen = sizeof(hdr);
1567
1568         /* Normally if we have it in the pack then we do not bother writing
1569          * it out into .git/objects/??/?{38} file.
1570          */
1571         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
1572         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
1573                 return 0;
1574         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
1575 }
1576
1577 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
1578                              struct object_id *oid, unsigned flags)
1579 {
1580         char *header;
1581         int hdrlen, status = 0;
1582
1583         /* type string, SP, %lu of the length plus NUL must fit this */
1584         hdrlen = strlen(type) + 32;
1585         header = xmalloc(hdrlen);
1586         write_sha1_file_prepare(buf, len, type, oid->hash, header, &hdrlen);
1587
1588         if (!(flags & HASH_WRITE_OBJECT))
1589                 goto cleanup;
1590         if (freshen_packed_object(oid->hash) || freshen_loose_object(oid->hash))
1591                 goto cleanup;
1592         status = write_loose_object(oid->hash, header, hdrlen, buf, len, 0);
1593
1594 cleanup:
1595         free(header);
1596         return status;
1597 }
1598
1599 int force_object_loose(const unsigned char *sha1, time_t mtime)
1600 {
1601         void *buf;
1602         unsigned long len;
1603         enum object_type type;
1604         char hdr[32];
1605         int hdrlen;
1606         int ret;
1607
1608         if (has_loose_object(sha1))
1609                 return 0;
1610         buf = read_object(sha1, &type, &len);
1611         if (!buf)
1612                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
1613         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
1614         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
1615         free(buf);
1616
1617         return ret;
1618 }
1619
1620 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
1621 {
1622         if (!startup_info->have_repository)
1623                 return 0;
1624         return sha1_object_info_extended(sha1, NULL,
1625                                          flags | OBJECT_INFO_SKIP_CACHED) >= 0;
1626 }
1627
1628 int has_object_file(const struct object_id *oid)
1629 {
1630         return has_sha1_file(oid->hash);
1631 }
1632
1633 int has_object_file_with_flags(const struct object_id *oid, int flags)
1634 {
1635         return has_sha1_file_with_flags(oid->hash, flags);
1636 }
1637
1638 static void check_tree(const void *buf, size_t size)
1639 {
1640         struct tree_desc desc;
1641         struct name_entry entry;
1642
1643         init_tree_desc(&desc, buf, size);
1644         while (tree_entry(&desc, &entry))
1645                 /* do nothing
1646                  * tree_entry() will die() on malformed entries */
1647                 ;
1648 }
1649
1650 static void check_commit(const void *buf, size_t size)
1651 {
1652         struct commit c;
1653         memset(&c, 0, sizeof(c));
1654         if (parse_commit_buffer(&c, buf, size))
1655                 die("corrupt commit");
1656 }
1657
1658 static void check_tag(const void *buf, size_t size)
1659 {
1660         struct tag t;
1661         memset(&t, 0, sizeof(t));
1662         if (parse_tag_buffer(&t, buf, size))
1663                 die("corrupt tag");
1664 }
1665
1666 static int index_mem(struct object_id *oid, void *buf, size_t size,
1667                      enum object_type type,
1668                      const char *path, unsigned flags)
1669 {
1670         int ret, re_allocated = 0;
1671         int write_object = flags & HASH_WRITE_OBJECT;
1672
1673         if (!type)
1674                 type = OBJ_BLOB;
1675
1676         /*
1677          * Convert blobs to git internal format
1678          */
1679         if ((type == OBJ_BLOB) && path) {
1680                 struct strbuf nbuf = STRBUF_INIT;
1681                 if (convert_to_git(&the_index, path, buf, size, &nbuf,
1682                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
1683                         buf = strbuf_detach(&nbuf, &size);
1684                         re_allocated = 1;
1685                 }
1686         }
1687         if (flags & HASH_FORMAT_CHECK) {
1688                 if (type == OBJ_TREE)
1689                         check_tree(buf, size);
1690                 if (type == OBJ_COMMIT)
1691                         check_commit(buf, size);
1692                 if (type == OBJ_TAG)
1693                         check_tag(buf, size);
1694         }
1695
1696         if (write_object)
1697                 ret = write_sha1_file(buf, size, typename(type), oid->hash);
1698         else
1699                 ret = hash_sha1_file(buf, size, typename(type), oid->hash);
1700         if (re_allocated)
1701                 free(buf);
1702         return ret;
1703 }
1704
1705 static int index_stream_convert_blob(struct object_id *oid, int fd,
1706                                      const char *path, unsigned flags)
1707 {
1708         int ret;
1709         const int write_object = flags & HASH_WRITE_OBJECT;
1710         struct strbuf sbuf = STRBUF_INIT;
1711
1712         assert(path);
1713         assert(would_convert_to_git_filter_fd(path));
1714
1715         convert_to_git_filter_fd(&the_index, path, fd, &sbuf,
1716                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
1717
1718         if (write_object)
1719                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
1720                                       oid->hash);
1721         else
1722                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
1723                                      oid->hash);
1724         strbuf_release(&sbuf);
1725         return ret;
1726 }
1727
1728 static int index_pipe(struct object_id *oid, int fd, enum object_type type,
1729                       const char *path, unsigned flags)
1730 {
1731         struct strbuf sbuf = STRBUF_INIT;
1732         int ret;
1733
1734         if (strbuf_read(&sbuf, fd, 4096) >= 0)
1735                 ret = index_mem(oid, sbuf.buf, sbuf.len, type, path, flags);
1736         else
1737                 ret = -1;
1738         strbuf_release(&sbuf);
1739         return ret;
1740 }
1741
1742 #define SMALL_FILE_SIZE (32*1024)
1743
1744 static int index_core(struct object_id *oid, int fd, size_t size,
1745                       enum object_type type, const char *path,
1746                       unsigned flags)
1747 {
1748         int ret;
1749
1750         if (!size) {
1751                 ret = index_mem(oid, "", size, type, path, flags);
1752         } else if (size <= SMALL_FILE_SIZE) {
1753                 char *buf = xmalloc(size);
1754                 ssize_t read_result = read_in_full(fd, buf, size);
1755                 if (read_result < 0)
1756                         ret = error_errno("read error while indexing %s",
1757                                           path ? path : "<unknown>");
1758                 else if (read_result != size)
1759                         ret = error("short read while indexing %s",
1760                                     path ? path : "<unknown>");
1761                 else
1762                         ret = index_mem(oid, buf, size, type, path, flags);
1763                 free(buf);
1764         } else {
1765                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
1766                 ret = index_mem(oid, buf, size, type, path, flags);
1767                 munmap(buf, size);
1768         }
1769         return ret;
1770 }
1771
1772 /*
1773  * This creates one packfile per large blob unless bulk-checkin
1774  * machinery is "plugged".
1775  *
1776  * This also bypasses the usual "convert-to-git" dance, and that is on
1777  * purpose. We could write a streaming version of the converting
1778  * functions and insert that before feeding the data to fast-import
1779  * (or equivalent in-core API described above). However, that is
1780  * somewhat complicated, as we do not know the size of the filter
1781  * result, which we need to know beforehand when writing a git object.
1782  * Since the primary motivation for trying to stream from the working
1783  * tree file and to avoid mmaping it in core is to deal with large
1784  * binary blobs, they generally do not want to get any conversion, and
1785  * callers should avoid this code path when filters are requested.
1786  */
1787 static int index_stream(struct object_id *oid, int fd, size_t size,
1788                         enum object_type type, const char *path,
1789                         unsigned flags)
1790 {
1791         return index_bulk_checkin(oid->hash, fd, size, type, path, flags);
1792 }
1793
1794 int index_fd(struct object_id *oid, int fd, struct stat *st,
1795              enum object_type type, const char *path, unsigned flags)
1796 {
1797         int ret;
1798
1799         /*
1800          * Call xsize_t() only when needed to avoid potentially unnecessary
1801          * die() for large files.
1802          */
1803         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
1804                 ret = index_stream_convert_blob(oid, fd, path, flags);
1805         else if (!S_ISREG(st->st_mode))
1806                 ret = index_pipe(oid, fd, type, path, flags);
1807         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
1808                  (path && would_convert_to_git(&the_index, path)))
1809                 ret = index_core(oid, fd, xsize_t(st->st_size), type, path,
1810                                  flags);
1811         else
1812                 ret = index_stream(oid, fd, xsize_t(st->st_size), type, path,
1813                                    flags);
1814         close(fd);
1815         return ret;
1816 }
1817
1818 int index_path(struct object_id *oid, const char *path, struct stat *st, unsigned flags)
1819 {
1820         int fd;
1821         struct strbuf sb = STRBUF_INIT;
1822         int rc = 0;
1823
1824         switch (st->st_mode & S_IFMT) {
1825         case S_IFREG:
1826                 fd = open(path, O_RDONLY);
1827                 if (fd < 0)
1828                         return error_errno("open(\"%s\")", path);
1829                 if (index_fd(oid, fd, st, OBJ_BLOB, path, flags) < 0)
1830                         return error("%s: failed to insert into database",
1831                                      path);
1832                 break;
1833         case S_IFLNK:
1834                 if (strbuf_readlink(&sb, path, st->st_size))
1835                         return error_errno("readlink(\"%s\")", path);
1836                 if (!(flags & HASH_WRITE_OBJECT))
1837                         hash_sha1_file(sb.buf, sb.len, blob_type, oid->hash);
1838                 else if (write_sha1_file(sb.buf, sb.len, blob_type, oid->hash))
1839                         rc = error("%s: failed to insert into database", path);
1840                 strbuf_release(&sb);
1841                 break;
1842         case S_IFDIR:
1843                 return resolve_gitlink_ref(path, "HEAD", oid);
1844         default:
1845                 return error("%s: unsupported file type", path);
1846         }
1847         return rc;
1848 }
1849
1850 int read_pack_header(int fd, struct pack_header *header)
1851 {
1852         if (read_in_full(fd, header, sizeof(*header)) != sizeof(*header))
1853                 /* "eof before pack header was fully read" */
1854                 return PH_ERROR_EOF;
1855
1856         if (header->hdr_signature != htonl(PACK_SIGNATURE))
1857                 /* "protocol error (pack signature mismatch detected)" */
1858                 return PH_ERROR_PACK_SIGNATURE;
1859         if (!pack_version_ok(header->hdr_version))
1860                 /* "protocol error (pack version unsupported)" */
1861                 return PH_ERROR_PROTOCOL;
1862         return 0;
1863 }
1864
1865 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
1866 {
1867         enum object_type type = sha1_object_info(sha1, NULL);
1868         if (type < 0)
1869                 die("%s is not a valid object", sha1_to_hex(sha1));
1870         if (type != expect)
1871                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
1872                     typename(expect));
1873 }
1874
1875 int for_each_file_in_obj_subdir(unsigned int subdir_nr,
1876                                 struct strbuf *path,
1877                                 each_loose_object_fn obj_cb,
1878                                 each_loose_cruft_fn cruft_cb,
1879                                 each_loose_subdir_fn subdir_cb,
1880                                 void *data)
1881 {
1882         size_t origlen, baselen;
1883         DIR *dir;
1884         struct dirent *de;
1885         int r = 0;
1886         struct object_id oid;
1887
1888         if (subdir_nr > 0xff)
1889                 BUG("invalid loose object subdirectory: %x", subdir_nr);
1890
1891         origlen = path->len;
1892         strbuf_complete(path, '/');
1893         strbuf_addf(path, "%02x", subdir_nr);
1894         baselen = path->len;
1895
1896         dir = opendir(path->buf);
1897         if (!dir) {
1898                 if (errno != ENOENT)
1899                         r = error_errno("unable to open %s", path->buf);
1900                 strbuf_setlen(path, origlen);
1901                 return r;
1902         }
1903
1904         oid.hash[0] = subdir_nr;
1905
1906         while ((de = readdir(dir))) {
1907                 if (is_dot_or_dotdot(de->d_name))
1908                         continue;
1909
1910                 strbuf_setlen(path, baselen);
1911                 strbuf_addf(path, "/%s", de->d_name);
1912
1913                 if (strlen(de->d_name) == GIT_SHA1_HEXSZ - 2 &&
1914                     !hex_to_bytes(oid.hash + 1, de->d_name,
1915                                   GIT_SHA1_RAWSZ - 1)) {
1916                         if (obj_cb) {
1917                                 r = obj_cb(&oid, path->buf, data);
1918                                 if (r)
1919                                         break;
1920                         }
1921                         continue;
1922                 }
1923
1924                 if (cruft_cb) {
1925                         r = cruft_cb(de->d_name, path->buf, data);
1926                         if (r)
1927                                 break;
1928                 }
1929         }
1930         closedir(dir);
1931
1932         strbuf_setlen(path, baselen);
1933         if (!r && subdir_cb)
1934                 r = subdir_cb(subdir_nr, path->buf, data);
1935
1936         strbuf_setlen(path, origlen);
1937
1938         return r;
1939 }
1940
1941 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
1942                             each_loose_object_fn obj_cb,
1943                             each_loose_cruft_fn cruft_cb,
1944                             each_loose_subdir_fn subdir_cb,
1945                             void *data)
1946 {
1947         int r = 0;
1948         int i;
1949
1950         for (i = 0; i < 256; i++) {
1951                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
1952                                                 subdir_cb, data);
1953                 if (r)
1954                         break;
1955         }
1956
1957         return r;
1958 }
1959
1960 int for_each_loose_file_in_objdir(const char *path,
1961                                   each_loose_object_fn obj_cb,
1962                                   each_loose_cruft_fn cruft_cb,
1963                                   each_loose_subdir_fn subdir_cb,
1964                                   void *data)
1965 {
1966         struct strbuf buf = STRBUF_INIT;
1967         int r;
1968
1969         strbuf_addstr(&buf, path);
1970         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
1971                                               subdir_cb, data);
1972         strbuf_release(&buf);
1973
1974         return r;
1975 }
1976
1977 struct loose_alt_odb_data {
1978         each_loose_object_fn *cb;
1979         void *data;
1980 };
1981
1982 static int loose_from_alt_odb(struct alternate_object_database *alt,
1983                               void *vdata)
1984 {
1985         struct loose_alt_odb_data *data = vdata;
1986         struct strbuf buf = STRBUF_INIT;
1987         int r;
1988
1989         strbuf_addstr(&buf, alt->path);
1990         r = for_each_loose_file_in_objdir_buf(&buf,
1991                                               data->cb, NULL, NULL,
1992                                               data->data);
1993         strbuf_release(&buf);
1994         return r;
1995 }
1996
1997 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
1998 {
1999         struct loose_alt_odb_data alt;
2000         int r;
2001
2002         r = for_each_loose_file_in_objdir(get_object_directory(),
2003                                           cb, NULL, NULL, data);
2004         if (r)
2005                 return r;
2006
2007         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
2008                 return 0;
2009
2010         alt.cb = cb;
2011         alt.data = data;
2012         return foreach_alt_odb(loose_from_alt_odb, &alt);
2013 }
2014
2015 static int check_stream_sha1(git_zstream *stream,
2016                              const char *hdr,
2017                              unsigned long size,
2018                              const char *path,
2019                              const unsigned char *expected_sha1)
2020 {
2021         git_SHA_CTX c;
2022         unsigned char real_sha1[GIT_MAX_RAWSZ];
2023         unsigned char buf[4096];
2024         unsigned long total_read;
2025         int status = Z_OK;
2026
2027         git_SHA1_Init(&c);
2028         git_SHA1_Update(&c, hdr, stream->total_out);
2029
2030         /*
2031          * We already read some bytes into hdr, but the ones up to the NUL
2032          * do not count against the object's content size.
2033          */
2034         total_read = stream->total_out - strlen(hdr) - 1;
2035
2036         /*
2037          * This size comparison must be "<=" to read the final zlib packets;
2038          * see the comment in unpack_sha1_rest for details.
2039          */
2040         while (total_read <= size &&
2041                (status == Z_OK || status == Z_BUF_ERROR)) {
2042                 stream->next_out = buf;
2043                 stream->avail_out = sizeof(buf);
2044                 if (size - total_read < stream->avail_out)
2045                         stream->avail_out = size - total_read;
2046                 status = git_inflate(stream, Z_FINISH);
2047                 git_SHA1_Update(&c, buf, stream->next_out - buf);
2048                 total_read += stream->next_out - buf;
2049         }
2050         git_inflate_end(stream);
2051
2052         if (status != Z_STREAM_END) {
2053                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
2054                 return -1;
2055         }
2056         if (stream->avail_in) {
2057                 error("garbage at end of loose object '%s'",
2058                       sha1_to_hex(expected_sha1));
2059                 return -1;
2060         }
2061
2062         git_SHA1_Final(real_sha1, &c);
2063         if (hashcmp(expected_sha1, real_sha1)) {
2064                 error("sha1 mismatch for %s (expected %s)", path,
2065                       sha1_to_hex(expected_sha1));
2066                 return -1;
2067         }
2068
2069         return 0;
2070 }
2071
2072 int read_loose_object(const char *path,
2073                       const unsigned char *expected_sha1,
2074                       enum object_type *type,
2075                       unsigned long *size,
2076                       void **contents)
2077 {
2078         int ret = -1;
2079         void *map = NULL;
2080         unsigned long mapsize;
2081         git_zstream stream;
2082         char hdr[32];
2083
2084         *contents = NULL;
2085
2086         map = map_sha1_file_1(path, NULL, &mapsize);
2087         if (!map) {
2088                 error_errno("unable to mmap %s", path);
2089                 goto out;
2090         }
2091
2092         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
2093                 error("unable to unpack header of %s", path);
2094                 goto out;
2095         }
2096
2097         *type = parse_sha1_header(hdr, size);
2098         if (*type < 0) {
2099                 error("unable to parse header of %s", path);
2100                 git_inflate_end(&stream);
2101                 goto out;
2102         }
2103
2104         if (*type == OBJ_BLOB) {
2105                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
2106                         goto out;
2107         } else {
2108                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
2109                 if (!*contents) {
2110                         error("unable to unpack contents of %s", path);
2111                         git_inflate_end(&stream);
2112                         goto out;
2113                 }
2114                 if (check_sha1_signature(expected_sha1, *contents,
2115                                          *size, typename(*type))) {
2116                         error("sha1 mismatch for %s (expected %s)", path,
2117                               sha1_to_hex(expected_sha1));
2118                         free(*contents);
2119                         goto out;
2120                 }
2121         }
2122
2123         ret = 0; /* everything checks out */
2124
2125 out:
2126         if (map)
2127                 munmap(map, mapsize);
2128         return ret;
2129 }