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