Merge branch 'nd/config-misc-fixes'
[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 "string-list.h"
11 #include "lockfile.h"
12 #include "delta.h"
13 #include "pack.h"
14 #include "blob.h"
15 #include "commit.h"
16 #include "run-command.h"
17 #include "tag.h"
18 #include "tree.h"
19 #include "tree-walk.h"
20 #include "refs.h"
21 #include "pack-revindex.h"
22 #include "sha1-lookup.h"
23 #include "bulk-checkin.h"
24 #include "streaming.h"
25 #include "dir.h"
26 #include "mru.h"
27 #include "list.h"
28 #include "mergesort.h"
29 #include "quote.h"
30
31 #define SZ_FMT PRIuMAX
32 static inline uintmax_t sz_fmt(size_t s) { return s; }
33
34 const unsigned char null_sha1[20];
35 const struct object_id null_oid;
36 const struct object_id empty_tree_oid = {
37         EMPTY_TREE_SHA1_BIN_LITERAL
38 };
39 const struct object_id empty_blob_oid = {
40         EMPTY_BLOB_SHA1_BIN_LITERAL
41 };
42
43 /*
44  * This is meant to hold a *small* number of objects that you would
45  * want read_sha1_file() to be able to return, but yet you do not want
46  * to write them into the object store (e.g. a browse-only
47  * application).
48  */
49 static struct cached_object {
50         unsigned char sha1[20];
51         enum object_type type;
52         void *buf;
53         unsigned long size;
54 } *cached_objects;
55 static int cached_object_nr, cached_object_alloc;
56
57 static struct cached_object empty_tree = {
58         EMPTY_TREE_SHA1_BIN_LITERAL,
59         OBJ_TREE,
60         "",
61         0
62 };
63
64 static struct cached_object *find_cached_object(const unsigned char *sha1)
65 {
66         int i;
67         struct cached_object *co = cached_objects;
68
69         for (i = 0; i < cached_object_nr; i++, co++) {
70                 if (!hashcmp(co->sha1, sha1))
71                         return co;
72         }
73         if (!hashcmp(sha1, empty_tree.sha1))
74                 return &empty_tree;
75         return NULL;
76 }
77
78 int mkdir_in_gitdir(const char *path)
79 {
80         if (mkdir(path, 0777)) {
81                 int saved_errno = errno;
82                 struct stat st;
83                 struct strbuf sb = STRBUF_INIT;
84
85                 if (errno != EEXIST)
86                         return -1;
87                 /*
88                  * Are we looking at a path in a symlinked worktree
89                  * whose original repository does not yet have it?
90                  * e.g. .git/rr-cache pointing at its original
91                  * repository in which the user hasn't performed any
92                  * conflict resolution yet?
93                  */
94                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
95                     strbuf_readlink(&sb, path, st.st_size) ||
96                     !is_absolute_path(sb.buf) ||
97                     mkdir(sb.buf, 0777)) {
98                         strbuf_release(&sb);
99                         errno = saved_errno;
100                         return -1;
101                 }
102                 strbuf_release(&sb);
103         }
104         return adjust_shared_perm(path);
105 }
106
107 enum scld_error safe_create_leading_directories(char *path)
108 {
109         char *next_component = path + offset_1st_component(path);
110         enum scld_error ret = SCLD_OK;
111
112         while (ret == SCLD_OK && next_component) {
113                 struct stat st;
114                 char *slash = next_component, slash_character;
115
116                 while (*slash && !is_dir_sep(*slash))
117                         slash++;
118
119                 if (!*slash)
120                         break;
121
122                 next_component = slash + 1;
123                 while (is_dir_sep(*next_component))
124                         next_component++;
125                 if (!*next_component)
126                         break;
127
128                 slash_character = *slash;
129                 *slash = '\0';
130                 if (!stat(path, &st)) {
131                         /* path exists */
132                         if (!S_ISDIR(st.st_mode))
133                                 ret = SCLD_EXISTS;
134                 } else if (mkdir(path, 0777)) {
135                         if (errno == EEXIST &&
136                             !stat(path, &st) && S_ISDIR(st.st_mode))
137                                 ; /* somebody created it since we checked */
138                         else if (errno == ENOENT)
139                                 /*
140                                  * Either mkdir() failed because
141                                  * somebody just pruned the containing
142                                  * directory, or stat() failed because
143                                  * the file that was in our way was
144                                  * just removed.  Either way, inform
145                                  * the caller that it might be worth
146                                  * trying again:
147                                  */
148                                 ret = SCLD_VANISHED;
149                         else
150                                 ret = SCLD_FAILED;
151                 } else if (adjust_shared_perm(path)) {
152                         ret = SCLD_PERMS;
153                 }
154                 *slash = slash_character;
155         }
156         return ret;
157 }
158
159 enum scld_error safe_create_leading_directories_const(const char *path)
160 {
161         /* path points to cache entries, so xstrdup before messing with it */
162         char *buf = xstrdup(path);
163         enum scld_error result = safe_create_leading_directories(buf);
164         free(buf);
165         return result;
166 }
167
168 static void fill_sha1_path(struct strbuf *buf, const unsigned char *sha1)
169 {
170         int i;
171         for (i = 0; i < 20; i++) {
172                 static char hex[] = "0123456789abcdef";
173                 unsigned int val = sha1[i];
174                 strbuf_addch(buf, hex[val >> 4]);
175                 strbuf_addch(buf, hex[val & 0xf]);
176                 if (!i)
177                         strbuf_addch(buf, '/');
178         }
179 }
180
181 const char *sha1_file_name(const unsigned char *sha1)
182 {
183         static struct strbuf buf = STRBUF_INIT;
184
185         strbuf_reset(&buf);
186         strbuf_addf(&buf, "%s/", get_object_directory());
187
188         fill_sha1_path(&buf, sha1);
189         return buf.buf;
190 }
191
192 struct strbuf *alt_scratch_buf(struct alternate_object_database *alt)
193 {
194         strbuf_setlen(&alt->scratch, alt->base_len);
195         return &alt->scratch;
196 }
197
198 static const char *alt_sha1_path(struct alternate_object_database *alt,
199                                  const unsigned char *sha1)
200 {
201         struct strbuf *buf = alt_scratch_buf(alt);
202         fill_sha1_path(buf, sha1);
203         return buf->buf;
204 }
205
206 /*
207  * Return the name of the pack or index file with the specified sha1
208  * in its filename.  *base and *name are scratch space that must be
209  * provided by the caller.  which should be "pack" or "idx".
210  */
211 static char *sha1_get_pack_name(const unsigned char *sha1,
212                                 struct strbuf *buf,
213                                 const char *which)
214 {
215         strbuf_reset(buf);
216         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
217                     sha1_to_hex(sha1), which);
218         return buf->buf;
219 }
220
221 char *sha1_pack_name(const unsigned char *sha1)
222 {
223         static struct strbuf buf = STRBUF_INIT;
224         return sha1_get_pack_name(sha1, &buf, "pack");
225 }
226
227 char *sha1_pack_index_name(const unsigned char *sha1)
228 {
229         static struct strbuf buf = STRBUF_INIT;
230         return sha1_get_pack_name(sha1, &buf, "idx");
231 }
232
233 struct alternate_object_database *alt_odb_list;
234 static struct alternate_object_database **alt_odb_tail;
235
236 /*
237  * Return non-zero iff the path is usable as an alternate object database.
238  */
239 static int alt_odb_usable(struct strbuf *path, const char *normalized_objdir)
240 {
241         struct alternate_object_database *alt;
242
243         /* Detect cases where alternate disappeared */
244         if (!is_directory(path->buf)) {
245                 error("object directory %s does not exist; "
246                       "check .git/objects/info/alternates.",
247                       path->buf);
248                 return 0;
249         }
250
251         /*
252          * Prevent the common mistake of listing the same
253          * thing twice, or object directory itself.
254          */
255         for (alt = alt_odb_list; alt; alt = alt->next) {
256                 if (!fspathcmp(path->buf, alt->path))
257                         return 0;
258         }
259         if (!fspathcmp(path->buf, normalized_objdir))
260                 return 0;
261
262         return 1;
263 }
264
265 /*
266  * Prepare alternate object database registry.
267  *
268  * The variable alt_odb_list points at the list of struct
269  * alternate_object_database.  The elements on this list come from
270  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
271  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
272  * whose contents is similar to that environment variable but can be
273  * LF separated.  Its base points at a statically allocated buffer that
274  * contains "/the/directory/corresponding/to/.git/objects/...", while
275  * its name points just after the slash at the end of ".git/objects/"
276  * in the example above, and has enough space to hold 40-byte hex
277  * SHA1, an extra slash for the first level indirection, and the
278  * terminating NUL.
279  */
280 static int link_alt_odb_entry(const char *entry, const char *relative_base,
281         int depth, const char *normalized_objdir)
282 {
283         struct alternate_object_database *ent;
284         struct strbuf pathbuf = STRBUF_INIT;
285
286         if (!is_absolute_path(entry) && relative_base) {
287                 strbuf_addstr(&pathbuf, real_path(relative_base));
288                 strbuf_addch(&pathbuf, '/');
289         }
290         strbuf_addstr(&pathbuf, entry);
291
292         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
293                 error("unable to normalize alternate object path: %s",
294                       pathbuf.buf);
295                 strbuf_release(&pathbuf);
296                 return -1;
297         }
298
299         /*
300          * The trailing slash after the directory name is given by
301          * this function at the end. Remove duplicates.
302          */
303         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
304                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
305
306         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
307                 strbuf_release(&pathbuf);
308                 return -1;
309         }
310
311         ent = alloc_alt_odb(pathbuf.buf);
312
313         /* add the alternate entry */
314         *alt_odb_tail = ent;
315         alt_odb_tail = &(ent->next);
316         ent->next = NULL;
317
318         /* recursively add alternates */
319         read_info_alternates(pathbuf.buf, depth + 1);
320
321         strbuf_release(&pathbuf);
322         return 0;
323 }
324
325 static const char *parse_alt_odb_entry(const char *string,
326                                        int sep,
327                                        struct strbuf *out)
328 {
329         const char *end;
330
331         strbuf_reset(out);
332
333         if (*string == '#') {
334                 /* comment; consume up to next separator */
335                 end = strchrnul(string, sep);
336         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
337                 /*
338                  * quoted path; unquote_c_style has copied the
339                  * data for us and set "end". Broken quoting (e.g.,
340                  * an entry that doesn't end with a quote) falls
341                  * back to the unquoted case below.
342                  */
343         } else {
344                 /* normal, unquoted path */
345                 end = strchrnul(string, sep);
346                 strbuf_add(out, string, end - string);
347         }
348
349         if (*end)
350                 end++;
351         return end;
352 }
353
354 static void link_alt_odb_entries(const char *alt, int len, int sep,
355                                  const char *relative_base, int depth)
356 {
357         struct strbuf objdirbuf = STRBUF_INIT;
358         struct strbuf entry = STRBUF_INIT;
359
360         if (depth > 5) {
361                 error("%s: ignoring alternate object stores, nesting too deep.",
362                                 relative_base);
363                 return;
364         }
365
366         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
367         if (strbuf_normalize_path(&objdirbuf) < 0)
368                 die("unable to normalize object directory: %s",
369                     objdirbuf.buf);
370
371         while (*alt) {
372                 alt = parse_alt_odb_entry(alt, sep, &entry);
373                 if (!entry.len)
374                         continue;
375                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
376         }
377         strbuf_release(&entry);
378         strbuf_release(&objdirbuf);
379 }
380
381 void read_info_alternates(const char * relative_base, int depth)
382 {
383         char *map;
384         size_t mapsz;
385         struct stat st;
386         char *path;
387         int fd;
388
389         path = xstrfmt("%s/info/alternates", relative_base);
390         fd = git_open(path);
391         free(path);
392         if (fd < 0)
393                 return;
394         if (fstat(fd, &st) || (st.st_size == 0)) {
395                 close(fd);
396                 return;
397         }
398         mapsz = xsize_t(st.st_size);
399         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
400         close(fd);
401
402         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
403
404         munmap(map, mapsz);
405 }
406
407 struct alternate_object_database *alloc_alt_odb(const char *dir)
408 {
409         struct alternate_object_database *ent;
410
411         FLEX_ALLOC_STR(ent, path, dir);
412         strbuf_init(&ent->scratch, 0);
413         strbuf_addf(&ent->scratch, "%s/", dir);
414         ent->base_len = ent->scratch.len;
415
416         return ent;
417 }
418
419 void add_to_alternates_file(const char *reference)
420 {
421         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
422         char *alts = git_pathdup("objects/info/alternates");
423         FILE *in, *out;
424
425         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
426         out = fdopen_lock_file(lock, "w");
427         if (!out)
428                 die_errno("unable to fdopen alternates lockfile");
429
430         in = fopen(alts, "r");
431         if (in) {
432                 struct strbuf line = STRBUF_INIT;
433                 int found = 0;
434
435                 while (strbuf_getline(&line, in) != EOF) {
436                         if (!strcmp(reference, line.buf)) {
437                                 found = 1;
438                                 break;
439                         }
440                         fprintf_or_die(out, "%s\n", line.buf);
441                 }
442
443                 strbuf_release(&line);
444                 fclose(in);
445
446                 if (found) {
447                         rollback_lock_file(lock);
448                         lock = NULL;
449                 }
450         }
451         else if (errno != ENOENT)
452                 die_errno("unable to read alternates file");
453
454         if (lock) {
455                 fprintf_or_die(out, "%s\n", reference);
456                 if (commit_lock_file(lock))
457                         die_errno("unable to move new alternates file into place");
458                 if (alt_odb_tail)
459                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
460         }
461         free(alts);
462 }
463
464 void add_to_alternates_memory(const char *reference)
465 {
466         /*
467          * Make sure alternates are initialized, or else our entry may be
468          * overwritten when they are.
469          */
470         prepare_alt_odb();
471
472         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
473 }
474
475 /*
476  * Compute the exact path an alternate is at and returns it. In case of
477  * error NULL is returned and the human readable error is added to `err`
478  * `path` may be relative and should point to $GITDIR.
479  * `err` must not be null.
480  */
481 char *compute_alternate_path(const char *path, struct strbuf *err)
482 {
483         char *ref_git = NULL;
484         const char *repo, *ref_git_s;
485         int seen_error = 0;
486
487         ref_git_s = real_path_if_valid(path);
488         if (!ref_git_s) {
489                 seen_error = 1;
490                 strbuf_addf(err, _("path '%s' does not exist"), path);
491                 goto out;
492         } else
493                 /*
494                  * Beware: read_gitfile(), real_path() and mkpath()
495                  * return static buffer
496                  */
497                 ref_git = xstrdup(ref_git_s);
498
499         repo = read_gitfile(ref_git);
500         if (!repo)
501                 repo = read_gitfile(mkpath("%s/.git", ref_git));
502         if (repo) {
503                 free(ref_git);
504                 ref_git = xstrdup(repo);
505         }
506
507         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
508                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
509                 free(ref_git);
510                 ref_git = ref_git_git;
511         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
512                 struct strbuf sb = STRBUF_INIT;
513                 seen_error = 1;
514                 if (get_common_dir(&sb, ref_git)) {
515                         strbuf_addf(err,
516                                     _("reference repository '%s' as a linked "
517                                       "checkout is not supported yet."),
518                                     path);
519                         goto out;
520                 }
521
522                 strbuf_addf(err, _("reference repository '%s' is not a "
523                                         "local repository."), path);
524                 goto out;
525         }
526
527         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
528                 strbuf_addf(err, _("reference repository '%s' is shallow"),
529                             path);
530                 seen_error = 1;
531                 goto out;
532         }
533
534         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
535                 strbuf_addf(err,
536                             _("reference repository '%s' is grafted"),
537                             path);
538                 seen_error = 1;
539                 goto out;
540         }
541
542 out:
543         if (seen_error) {
544                 free(ref_git);
545                 ref_git = NULL;
546         }
547
548         return ref_git;
549 }
550
551 int foreach_alt_odb(alt_odb_fn fn, void *cb)
552 {
553         struct alternate_object_database *ent;
554         int r = 0;
555
556         prepare_alt_odb();
557         for (ent = alt_odb_list; ent; ent = ent->next) {
558                 r = fn(ent, cb);
559                 if (r)
560                         break;
561         }
562         return r;
563 }
564
565 void prepare_alt_odb(void)
566 {
567         const char *alt;
568
569         if (alt_odb_tail)
570                 return;
571
572         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
573         if (!alt) alt = "";
574
575         alt_odb_tail = &alt_odb_list;
576         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
577
578         read_info_alternates(get_object_directory(), 0);
579 }
580
581 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
582 static int freshen_file(const char *fn)
583 {
584         struct utimbuf t;
585         t.actime = t.modtime = time(NULL);
586         return !utime(fn, &t);
587 }
588
589 /*
590  * All of the check_and_freshen functions return 1 if the file exists and was
591  * freshened (if freshening was requested), 0 otherwise. If they return
592  * 0, you should not assume that it is safe to skip a write of the object (it
593  * either does not exist on disk, or has a stale mtime and may be subject to
594  * pruning).
595  */
596 static int check_and_freshen_file(const char *fn, int freshen)
597 {
598         if (access(fn, F_OK))
599                 return 0;
600         if (freshen && !freshen_file(fn))
601                 return 0;
602         return 1;
603 }
604
605 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
606 {
607         return check_and_freshen_file(sha1_file_name(sha1), freshen);
608 }
609
610 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
611 {
612         struct alternate_object_database *alt;
613         prepare_alt_odb();
614         for (alt = alt_odb_list; alt; alt = alt->next) {
615                 const char *path = alt_sha1_path(alt, sha1);
616                 if (check_and_freshen_file(path, freshen))
617                         return 1;
618         }
619         return 0;
620 }
621
622 static int check_and_freshen(const unsigned char *sha1, int freshen)
623 {
624         return check_and_freshen_local(sha1, freshen) ||
625                check_and_freshen_nonlocal(sha1, freshen);
626 }
627
628 int has_loose_object_nonlocal(const unsigned char *sha1)
629 {
630         return check_and_freshen_nonlocal(sha1, 0);
631 }
632
633 static int has_loose_object(const unsigned char *sha1)
634 {
635         return check_and_freshen(sha1, 0);
636 }
637
638 static unsigned int pack_used_ctr;
639 static unsigned int pack_mmap_calls;
640 static unsigned int peak_pack_open_windows;
641 static unsigned int pack_open_windows;
642 static unsigned int pack_open_fds;
643 static unsigned int pack_max_fds;
644 static size_t peak_pack_mapped;
645 static size_t pack_mapped;
646 struct packed_git *packed_git;
647
648 static struct mru packed_git_mru_storage;
649 struct mru *packed_git_mru = &packed_git_mru_storage;
650
651 void pack_report(void)
652 {
653         fprintf(stderr,
654                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
655                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
656                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
657                 sz_fmt(getpagesize()),
658                 sz_fmt(packed_git_window_size),
659                 sz_fmt(packed_git_limit));
660         fprintf(stderr,
661                 "pack_report: pack_used_ctr            = %10u\n"
662                 "pack_report: pack_mmap_calls          = %10u\n"
663                 "pack_report: pack_open_windows        = %10u / %10u\n"
664                 "pack_report: pack_mapped              = "
665                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
666                 pack_used_ctr,
667                 pack_mmap_calls,
668                 pack_open_windows, peak_pack_open_windows,
669                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
670 }
671
672 /*
673  * Open and mmap the index file at path, perform a couple of
674  * consistency checks, then record its information to p.  Return 0 on
675  * success.
676  */
677 static int check_packed_git_idx(const char *path, struct packed_git *p)
678 {
679         void *idx_map;
680         struct pack_idx_header *hdr;
681         size_t idx_size;
682         uint32_t version, nr, i, *index;
683         int fd = git_open(path);
684         struct stat st;
685
686         if (fd < 0)
687                 return -1;
688         if (fstat(fd, &st)) {
689                 close(fd);
690                 return -1;
691         }
692         idx_size = xsize_t(st.st_size);
693         if (idx_size < 4 * 256 + 20 + 20) {
694                 close(fd);
695                 return error("index file %s is too small", path);
696         }
697         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
698         close(fd);
699
700         hdr = idx_map;
701         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
702                 version = ntohl(hdr->idx_version);
703                 if (version < 2 || version > 2) {
704                         munmap(idx_map, idx_size);
705                         return error("index file %s is version %"PRIu32
706                                      " and is not supported by this binary"
707                                      " (try upgrading GIT to a newer version)",
708                                      path, version);
709                 }
710         } else
711                 version = 1;
712
713         nr = 0;
714         index = idx_map;
715         if (version > 1)
716                 index += 2;  /* skip index header */
717         for (i = 0; i < 256; i++) {
718                 uint32_t n = ntohl(index[i]);
719                 if (n < nr) {
720                         munmap(idx_map, idx_size);
721                         return error("non-monotonic index %s", path);
722                 }
723                 nr = n;
724         }
725
726         if (version == 1) {
727                 /*
728                  * Total size:
729                  *  - 256 index entries 4 bytes each
730                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
731                  *  - 20-byte SHA1 of the packfile
732                  *  - 20-byte SHA1 file checksum
733                  */
734                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
735                         munmap(idx_map, idx_size);
736                         return error("wrong index v1 file size in %s", path);
737                 }
738         } else if (version == 2) {
739                 /*
740                  * Minimum size:
741                  *  - 8 bytes of header
742                  *  - 256 index entries 4 bytes each
743                  *  - 20-byte sha1 entry * nr
744                  *  - 4-byte crc entry * nr
745                  *  - 4-byte offset entry * nr
746                  *  - 20-byte SHA1 of the packfile
747                  *  - 20-byte SHA1 file checksum
748                  * And after the 4-byte offset table might be a
749                  * variable sized table containing 8-byte entries
750                  * for offsets larger than 2^31.
751                  */
752                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
753                 unsigned long max_size = min_size;
754                 if (nr)
755                         max_size += (nr - 1)*8;
756                 if (idx_size < min_size || idx_size > max_size) {
757                         munmap(idx_map, idx_size);
758                         return error("wrong index v2 file size in %s", path);
759                 }
760                 if (idx_size != min_size &&
761                     /*
762                      * make sure we can deal with large pack offsets.
763                      * 31-bit signed offset won't be enough, neither
764                      * 32-bit unsigned one will be.
765                      */
766                     (sizeof(off_t) <= 4)) {
767                         munmap(idx_map, idx_size);
768                         return error("pack too large for current definition of off_t in %s", path);
769                 }
770         }
771
772         p->index_version = version;
773         p->index_data = idx_map;
774         p->index_size = idx_size;
775         p->num_objects = nr;
776         return 0;
777 }
778
779 int open_pack_index(struct packed_git *p)
780 {
781         char *idx_name;
782         size_t len;
783         int ret;
784
785         if (p->index_data)
786                 return 0;
787
788         if (!strip_suffix(p->pack_name, ".pack", &len))
789                 die("BUG: pack_name does not end in .pack");
790         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
791         ret = check_packed_git_idx(idx_name, p);
792         free(idx_name);
793         return ret;
794 }
795
796 static void scan_windows(struct packed_git *p,
797         struct packed_git **lru_p,
798         struct pack_window **lru_w,
799         struct pack_window **lru_l)
800 {
801         struct pack_window *w, *w_l;
802
803         for (w_l = NULL, w = p->windows; w; w = w->next) {
804                 if (!w->inuse_cnt) {
805                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
806                                 *lru_p = p;
807                                 *lru_w = w;
808                                 *lru_l = w_l;
809                         }
810                 }
811                 w_l = w;
812         }
813 }
814
815 static int unuse_one_window(struct packed_git *current)
816 {
817         struct packed_git *p, *lru_p = NULL;
818         struct pack_window *lru_w = NULL, *lru_l = NULL;
819
820         if (current)
821                 scan_windows(current, &lru_p, &lru_w, &lru_l);
822         for (p = packed_git; p; p = p->next)
823                 scan_windows(p, &lru_p, &lru_w, &lru_l);
824         if (lru_p) {
825                 munmap(lru_w->base, lru_w->len);
826                 pack_mapped -= lru_w->len;
827                 if (lru_l)
828                         lru_l->next = lru_w->next;
829                 else
830                         lru_p->windows = lru_w->next;
831                 free(lru_w);
832                 pack_open_windows--;
833                 return 1;
834         }
835         return 0;
836 }
837
838 void release_pack_memory(size_t need)
839 {
840         size_t cur = pack_mapped;
841         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
842                 ; /* nothing */
843 }
844
845 static void mmap_limit_check(size_t length)
846 {
847         static size_t limit = 0;
848         if (!limit) {
849                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
850                 if (!limit)
851                         limit = SIZE_MAX;
852         }
853         if (length > limit)
854                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
855                     (uintmax_t)length, (uintmax_t)limit);
856 }
857
858 void *xmmap_gently(void *start, size_t length,
859                   int prot, int flags, int fd, off_t offset)
860 {
861         void *ret;
862
863         mmap_limit_check(length);
864         ret = mmap(start, length, prot, flags, fd, offset);
865         if (ret == MAP_FAILED) {
866                 if (!length)
867                         return NULL;
868                 release_pack_memory(length);
869                 ret = mmap(start, length, prot, flags, fd, offset);
870         }
871         return ret;
872 }
873
874 void *xmmap(void *start, size_t length,
875         int prot, int flags, int fd, off_t offset)
876 {
877         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
878         if (ret == MAP_FAILED)
879                 die_errno("mmap failed");
880         return ret;
881 }
882
883 void close_pack_windows(struct packed_git *p)
884 {
885         while (p->windows) {
886                 struct pack_window *w = p->windows;
887
888                 if (w->inuse_cnt)
889                         die("pack '%s' still has open windows to it",
890                             p->pack_name);
891                 munmap(w->base, w->len);
892                 pack_mapped -= w->len;
893                 pack_open_windows--;
894                 p->windows = w->next;
895                 free(w);
896         }
897 }
898
899 static int close_pack_fd(struct packed_git *p)
900 {
901         if (p->pack_fd < 0)
902                 return 0;
903
904         close(p->pack_fd);
905         pack_open_fds--;
906         p->pack_fd = -1;
907
908         return 1;
909 }
910
911 static void close_pack(struct packed_git *p)
912 {
913         close_pack_windows(p);
914         close_pack_fd(p);
915         close_pack_index(p);
916 }
917
918 void close_all_packs(void)
919 {
920         struct packed_git *p;
921
922         for (p = packed_git; p; p = p->next)
923                 if (p->do_not_close)
924                         die("BUG: want to close pack marked 'do-not-close'");
925                 else
926                         close_pack(p);
927 }
928
929
930 /*
931  * The LRU pack is the one with the oldest MRU window, preferring packs
932  * with no used windows, or the oldest mtime if it has no windows allocated.
933  */
934 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
935 {
936         struct pack_window *w, *this_mru_w;
937         int has_windows_inuse = 0;
938
939         /*
940          * Reject this pack if it has windows and the previously selected
941          * one does not.  If this pack does not have windows, reject
942          * it if the pack file is newer than the previously selected one.
943          */
944         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
945                 return;
946
947         for (w = this_mru_w = p->windows; w; w = w->next) {
948                 /*
949                  * Reject this pack if any of its windows are in use,
950                  * but the previously selected pack did not have any
951                  * inuse windows.  Otherwise, record that this pack
952                  * has windows in use.
953                  */
954                 if (w->inuse_cnt) {
955                         if (*accept_windows_inuse)
956                                 has_windows_inuse = 1;
957                         else
958                                 return;
959                 }
960
961                 if (w->last_used > this_mru_w->last_used)
962                         this_mru_w = w;
963
964                 /*
965                  * Reject this pack if it has windows that have been
966                  * used more recently than the previously selected pack.
967                  * If the previously selected pack had windows inuse and
968                  * we have not encountered a window in this pack that is
969                  * inuse, skip this check since we prefer a pack with no
970                  * inuse windows to one that has inuse windows.
971                  */
972                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
973                     this_mru_w->last_used > (*mru_w)->last_used)
974                         return;
975         }
976
977         /*
978          * Select this pack.
979          */
980         *mru_w = this_mru_w;
981         *lru_p = p;
982         *accept_windows_inuse = has_windows_inuse;
983 }
984
985 static int close_one_pack(void)
986 {
987         struct packed_git *p, *lru_p = NULL;
988         struct pack_window *mru_w = NULL;
989         int accept_windows_inuse = 1;
990
991         for (p = packed_git; p; p = p->next) {
992                 if (p->pack_fd == -1)
993                         continue;
994                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
995         }
996
997         if (lru_p)
998                 return close_pack_fd(lru_p);
999
1000         return 0;
1001 }
1002
1003 void unuse_pack(struct pack_window **w_cursor)
1004 {
1005         struct pack_window *w = *w_cursor;
1006         if (w) {
1007                 w->inuse_cnt--;
1008                 *w_cursor = NULL;
1009         }
1010 }
1011
1012 void close_pack_index(struct packed_git *p)
1013 {
1014         if (p->index_data) {
1015                 munmap((void *)p->index_data, p->index_size);
1016                 p->index_data = NULL;
1017         }
1018 }
1019
1020 static unsigned int get_max_fd_limit(void)
1021 {
1022 #ifdef RLIMIT_NOFILE
1023         {
1024                 struct rlimit lim;
1025
1026                 if (!getrlimit(RLIMIT_NOFILE, &lim))
1027                         return lim.rlim_cur;
1028         }
1029 #endif
1030
1031 #ifdef _SC_OPEN_MAX
1032         {
1033                 long open_max = sysconf(_SC_OPEN_MAX);
1034                 if (0 < open_max)
1035                         return open_max;
1036                 /*
1037                  * Otherwise, we got -1 for one of the two
1038                  * reasons:
1039                  *
1040                  * (1) sysconf() did not understand _SC_OPEN_MAX
1041                  *     and signaled an error with -1; or
1042                  * (2) sysconf() said there is no limit.
1043                  *
1044                  * We _could_ clear errno before calling sysconf() to
1045                  * tell these two cases apart and return a huge number
1046                  * in the latter case to let the caller cap it to a
1047                  * value that is not so selfish, but letting the
1048                  * fallback OPEN_MAX codepath take care of these cases
1049                  * is a lot simpler.
1050                  */
1051         }
1052 #endif
1053
1054 #ifdef OPEN_MAX
1055         return OPEN_MAX;
1056 #else
1057         return 1; /* see the caller ;-) */
1058 #endif
1059 }
1060
1061 /*
1062  * Do not call this directly as this leaks p->pack_fd on error return;
1063  * call open_packed_git() instead.
1064  */
1065 static int open_packed_git_1(struct packed_git *p)
1066 {
1067         struct stat st;
1068         struct pack_header hdr;
1069         unsigned char sha1[20];
1070         unsigned char *idx_sha1;
1071         long fd_flag;
1072
1073         if (!p->index_data && open_pack_index(p))
1074                 return error("packfile %s index unavailable", p->pack_name);
1075
1076         if (!pack_max_fds) {
1077                 unsigned int max_fds = get_max_fd_limit();
1078
1079                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1080                 if (25 < max_fds)
1081                         pack_max_fds = max_fds - 25;
1082                 else
1083                         pack_max_fds = 1;
1084         }
1085
1086         while (pack_max_fds <= pack_open_fds && close_one_pack())
1087                 ; /* nothing */
1088
1089         p->pack_fd = git_open(p->pack_name);
1090         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1091                 return -1;
1092         pack_open_fds++;
1093
1094         /* If we created the struct before we had the pack we lack size. */
1095         if (!p->pack_size) {
1096                 if (!S_ISREG(st.st_mode))
1097                         return error("packfile %s not a regular file", p->pack_name);
1098                 p->pack_size = st.st_size;
1099         } else if (p->pack_size != st.st_size)
1100                 return error("packfile %s size changed", p->pack_name);
1101
1102         /* We leave these file descriptors open with sliding mmap;
1103          * there is no point keeping them open across exec(), though.
1104          */
1105         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1106         if (fd_flag < 0)
1107                 return error("cannot determine file descriptor flags");
1108         fd_flag |= FD_CLOEXEC;
1109         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1110                 return error("cannot set FD_CLOEXEC");
1111
1112         /* Verify we recognize this pack file format. */
1113         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1114                 return error("file %s is far too short to be a packfile", p->pack_name);
1115         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1116                 return error("file %s is not a GIT packfile", p->pack_name);
1117         if (!pack_version_ok(hdr.hdr_version))
1118                 return error("packfile %s is version %"PRIu32" and not"
1119                         " supported (try upgrading GIT to a newer version)",
1120                         p->pack_name, ntohl(hdr.hdr_version));
1121
1122         /* Verify the pack matches its index. */
1123         if (p->num_objects != ntohl(hdr.hdr_entries))
1124                 return error("packfile %s claims to have %"PRIu32" objects"
1125                              " while index indicates %"PRIu32" objects",
1126                              p->pack_name, ntohl(hdr.hdr_entries),
1127                              p->num_objects);
1128         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1129                 return error("end of packfile %s is unavailable", p->pack_name);
1130         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1131                 return error("packfile %s signature is unavailable", p->pack_name);
1132         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1133         if (hashcmp(sha1, idx_sha1))
1134                 return error("packfile %s does not match index", p->pack_name);
1135         return 0;
1136 }
1137
1138 static int open_packed_git(struct packed_git *p)
1139 {
1140         if (!open_packed_git_1(p))
1141                 return 0;
1142         close_pack_fd(p);
1143         return -1;
1144 }
1145
1146 static int in_window(struct pack_window *win, off_t offset)
1147 {
1148         /* We must promise at least 20 bytes (one hash) after the
1149          * offset is available from this window, otherwise the offset
1150          * is not actually in this window and a different window (which
1151          * has that one hash excess) must be used.  This is to support
1152          * the object header and delta base parsing routines below.
1153          */
1154         off_t win_off = win->offset;
1155         return win_off <= offset
1156                 && (offset + 20) <= (win_off + win->len);
1157 }
1158
1159 unsigned char *use_pack(struct packed_git *p,
1160                 struct pack_window **w_cursor,
1161                 off_t offset,
1162                 unsigned long *left)
1163 {
1164         struct pack_window *win = *w_cursor;
1165
1166         /* Since packfiles end in a hash of their content and it's
1167          * pointless to ask for an offset into the middle of that
1168          * hash, and the in_window function above wouldn't match
1169          * don't allow an offset too close to the end of the file.
1170          */
1171         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1172                 die("packfile %s cannot be accessed", p->pack_name);
1173         if (offset > (p->pack_size - 20))
1174                 die("offset beyond end of packfile (truncated pack?)");
1175         if (offset < 0)
1176                 die(_("offset before end of packfile (broken .idx?)"));
1177
1178         if (!win || !in_window(win, offset)) {
1179                 if (win)
1180                         win->inuse_cnt--;
1181                 for (win = p->windows; win; win = win->next) {
1182                         if (in_window(win, offset))
1183                                 break;
1184                 }
1185                 if (!win) {
1186                         size_t window_align = packed_git_window_size / 2;
1187                         off_t len;
1188
1189                         if (p->pack_fd == -1 && open_packed_git(p))
1190                                 die("packfile %s cannot be accessed", p->pack_name);
1191
1192                         win = xcalloc(1, sizeof(*win));
1193                         win->offset = (offset / window_align) * window_align;
1194                         len = p->pack_size - win->offset;
1195                         if (len > packed_git_window_size)
1196                                 len = packed_git_window_size;
1197                         win->len = (size_t)len;
1198                         pack_mapped += win->len;
1199                         while (packed_git_limit < pack_mapped
1200                                 && unuse_one_window(p))
1201                                 ; /* nothing */
1202                         win->base = xmmap(NULL, win->len,
1203                                 PROT_READ, MAP_PRIVATE,
1204                                 p->pack_fd, win->offset);
1205                         if (win->base == MAP_FAILED)
1206                                 die_errno("packfile %s cannot be mapped",
1207                                           p->pack_name);
1208                         if (!win->offset && win->len == p->pack_size
1209                                 && !p->do_not_close)
1210                                 close_pack_fd(p);
1211                         pack_mmap_calls++;
1212                         pack_open_windows++;
1213                         if (pack_mapped > peak_pack_mapped)
1214                                 peak_pack_mapped = pack_mapped;
1215                         if (pack_open_windows > peak_pack_open_windows)
1216                                 peak_pack_open_windows = pack_open_windows;
1217                         win->next = p->windows;
1218                         p->windows = win;
1219                 }
1220         }
1221         if (win != *w_cursor) {
1222                 win->last_used = pack_used_ctr++;
1223                 win->inuse_cnt++;
1224                 *w_cursor = win;
1225         }
1226         offset -= win->offset;
1227         if (left)
1228                 *left = win->len - xsize_t(offset);
1229         return win->base + offset;
1230 }
1231
1232 static struct packed_git *alloc_packed_git(int extra)
1233 {
1234         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1235         memset(p, 0, sizeof(*p));
1236         p->pack_fd = -1;
1237         return p;
1238 }
1239
1240 static void try_to_free_pack_memory(size_t size)
1241 {
1242         release_pack_memory(size);
1243 }
1244
1245 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1246 {
1247         static int have_set_try_to_free_routine;
1248         struct stat st;
1249         size_t alloc;
1250         struct packed_git *p;
1251
1252         if (!have_set_try_to_free_routine) {
1253                 have_set_try_to_free_routine = 1;
1254                 set_try_to_free_routine(try_to_free_pack_memory);
1255         }
1256
1257         /*
1258          * Make sure a corresponding .pack file exists and that
1259          * the index looks sane.
1260          */
1261         if (!strip_suffix_mem(path, &path_len, ".idx"))
1262                 return NULL;
1263
1264         /*
1265          * ".pack" is long enough to hold any suffix we're adding (and
1266          * the use xsnprintf double-checks that)
1267          */
1268         alloc = st_add3(path_len, strlen(".pack"), 1);
1269         p = alloc_packed_git(alloc);
1270         memcpy(p->pack_name, path, path_len);
1271
1272         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1273         if (!access(p->pack_name, F_OK))
1274                 p->pack_keep = 1;
1275
1276         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1277         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1278                 free(p);
1279                 return NULL;
1280         }
1281
1282         /* ok, it looks sane as far as we can check without
1283          * actually mapping the pack file.
1284          */
1285         p->pack_size = st.st_size;
1286         p->pack_local = local;
1287         p->mtime = st.st_mtime;
1288         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1289                 hashclr(p->sha1);
1290         return p;
1291 }
1292
1293 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1294 {
1295         const char *path = sha1_pack_name(sha1);
1296         size_t alloc = st_add(strlen(path), 1);
1297         struct packed_git *p = alloc_packed_git(alloc);
1298
1299         memcpy(p->pack_name, path, alloc); /* includes NUL */
1300         hashcpy(p->sha1, sha1);
1301         if (check_packed_git_idx(idx_path, p)) {
1302                 free(p);
1303                 return NULL;
1304         }
1305
1306         return p;
1307 }
1308
1309 void install_packed_git(struct packed_git *pack)
1310 {
1311         if (pack->pack_fd != -1)
1312                 pack_open_fds++;
1313
1314         pack->next = packed_git;
1315         packed_git = pack;
1316 }
1317
1318 void (*report_garbage)(unsigned seen_bits, const char *path);
1319
1320 static void report_helper(const struct string_list *list,
1321                           int seen_bits, int first, int last)
1322 {
1323         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1324                 return;
1325
1326         for (; first < last; first++)
1327                 report_garbage(seen_bits, list->items[first].string);
1328 }
1329
1330 static void report_pack_garbage(struct string_list *list)
1331 {
1332         int i, baselen = -1, first = 0, seen_bits = 0;
1333
1334         if (!report_garbage)
1335                 return;
1336
1337         string_list_sort(list);
1338
1339         for (i = 0; i < list->nr; i++) {
1340                 const char *path = list->items[i].string;
1341                 if (baselen != -1 &&
1342                     strncmp(path, list->items[first].string, baselen)) {
1343                         report_helper(list, seen_bits, first, i);
1344                         baselen = -1;
1345                         seen_bits = 0;
1346                 }
1347                 if (baselen == -1) {
1348                         const char *dot = strrchr(path, '.');
1349                         if (!dot) {
1350                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1351                                 continue;
1352                         }
1353                         baselen = dot - path + 1;
1354                         first = i;
1355                 }
1356                 if (!strcmp(path + baselen, "pack"))
1357                         seen_bits |= 1;
1358                 else if (!strcmp(path + baselen, "idx"))
1359                         seen_bits |= 2;
1360         }
1361         report_helper(list, seen_bits, first, list->nr);
1362 }
1363
1364 static void prepare_packed_git_one(char *objdir, int local)
1365 {
1366         struct strbuf path = STRBUF_INIT;
1367         size_t dirnamelen;
1368         DIR *dir;
1369         struct dirent *de;
1370         struct string_list garbage = STRING_LIST_INIT_DUP;
1371
1372         strbuf_addstr(&path, objdir);
1373         strbuf_addstr(&path, "/pack");
1374         dir = opendir(path.buf);
1375         if (!dir) {
1376                 if (errno != ENOENT)
1377                         error_errno("unable to open object pack directory: %s",
1378                                     path.buf);
1379                 strbuf_release(&path);
1380                 return;
1381         }
1382         strbuf_addch(&path, '/');
1383         dirnamelen = path.len;
1384         while ((de = readdir(dir)) != NULL) {
1385                 struct packed_git *p;
1386                 size_t base_len;
1387
1388                 if (is_dot_or_dotdot(de->d_name))
1389                         continue;
1390
1391                 strbuf_setlen(&path, dirnamelen);
1392                 strbuf_addstr(&path, de->d_name);
1393
1394                 base_len = path.len;
1395                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1396                         /* Don't reopen a pack we already have. */
1397                         for (p = packed_git; p; p = p->next) {
1398                                 size_t len;
1399                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1400                                     len == base_len &&
1401                                     !memcmp(p->pack_name, path.buf, len))
1402                                         break;
1403                         }
1404                         if (p == NULL &&
1405                             /*
1406                              * See if it really is a valid .idx file with
1407                              * corresponding .pack file that we can map.
1408                              */
1409                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1410                                 install_packed_git(p);
1411                 }
1412
1413                 if (!report_garbage)
1414                         continue;
1415
1416                 if (ends_with(de->d_name, ".idx") ||
1417                     ends_with(de->d_name, ".pack") ||
1418                     ends_with(de->d_name, ".bitmap") ||
1419                     ends_with(de->d_name, ".keep"))
1420                         string_list_append(&garbage, path.buf);
1421                 else
1422                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1423         }
1424         closedir(dir);
1425         report_pack_garbage(&garbage);
1426         string_list_clear(&garbage, 0);
1427         strbuf_release(&path);
1428 }
1429
1430 static int approximate_object_count_valid;
1431
1432 /*
1433  * Give a fast, rough count of the number of objects in the repository. This
1434  * ignores loose objects completely. If you have a lot of them, then either
1435  * you should repack because your performance will be awful, or they are
1436  * all unreachable objects about to be pruned, in which case they're not really
1437  * interesting as a measure of repo size in the first place.
1438  */
1439 unsigned long approximate_object_count(void)
1440 {
1441         static unsigned long count;
1442         if (!approximate_object_count_valid) {
1443                 struct packed_git *p;
1444
1445                 prepare_packed_git();
1446                 count = 0;
1447                 for (p = packed_git; p; p = p->next) {
1448                         if (open_pack_index(p))
1449                                 continue;
1450                         count += p->num_objects;
1451                 }
1452         }
1453         return count;
1454 }
1455
1456 static void *get_next_packed_git(const void *p)
1457 {
1458         return ((const struct packed_git *)p)->next;
1459 }
1460
1461 static void set_next_packed_git(void *p, void *next)
1462 {
1463         ((struct packed_git *)p)->next = next;
1464 }
1465
1466 static int sort_pack(const void *a_, const void *b_)
1467 {
1468         const struct packed_git *a = a_;
1469         const struct packed_git *b = b_;
1470         int st;
1471
1472         /*
1473          * Local packs tend to contain objects specific to our
1474          * variant of the project than remote ones.  In addition,
1475          * remote ones could be on a network mounted filesystem.
1476          * Favor local ones for these reasons.
1477          */
1478         st = a->pack_local - b->pack_local;
1479         if (st)
1480                 return -st;
1481
1482         /*
1483          * Younger packs tend to contain more recent objects,
1484          * and more recent objects tend to get accessed more
1485          * often.
1486          */
1487         if (a->mtime < b->mtime)
1488                 return 1;
1489         else if (a->mtime == b->mtime)
1490                 return 0;
1491         return -1;
1492 }
1493
1494 static void rearrange_packed_git(void)
1495 {
1496         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1497                                      set_next_packed_git, sort_pack);
1498 }
1499
1500 static void prepare_packed_git_mru(void)
1501 {
1502         struct packed_git *p;
1503
1504         mru_clear(packed_git_mru);
1505         for (p = packed_git; p; p = p->next)
1506                 mru_append(packed_git_mru, p);
1507 }
1508
1509 static int prepare_packed_git_run_once = 0;
1510 void prepare_packed_git(void)
1511 {
1512         struct alternate_object_database *alt;
1513
1514         if (prepare_packed_git_run_once)
1515                 return;
1516         prepare_packed_git_one(get_object_directory(), 1);
1517         prepare_alt_odb();
1518         for (alt = alt_odb_list; alt; alt = alt->next)
1519                 prepare_packed_git_one(alt->path, 0);
1520         rearrange_packed_git();
1521         prepare_packed_git_mru();
1522         prepare_packed_git_run_once = 1;
1523 }
1524
1525 void reprepare_packed_git(void)
1526 {
1527         approximate_object_count_valid = 0;
1528         prepare_packed_git_run_once = 0;
1529         prepare_packed_git();
1530 }
1531
1532 static void mark_bad_packed_object(struct packed_git *p,
1533                                    const unsigned char *sha1)
1534 {
1535         unsigned i;
1536         for (i = 0; i < p->num_bad_objects; i++)
1537                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1538                         return;
1539         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1540                                       st_mult(GIT_SHA1_RAWSZ,
1541                                               st_add(p->num_bad_objects, 1)));
1542         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1543         p->num_bad_objects++;
1544 }
1545
1546 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1547 {
1548         struct packed_git *p;
1549         unsigned i;
1550
1551         for (p = packed_git; p; p = p->next)
1552                 for (i = 0; i < p->num_bad_objects; i++)
1553                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1554                                 return p;
1555         return NULL;
1556 }
1557
1558 /*
1559  * With an in-core object data in "map", rehash it to make sure the
1560  * object name actually matches "sha1" to detect object corruption.
1561  * With "map" == NULL, try reading the object named with "sha1" using
1562  * the streaming interface and rehash it to do the same.
1563  */
1564 int check_sha1_signature(const unsigned char *sha1, void *map,
1565                          unsigned long size, const char *type)
1566 {
1567         unsigned char real_sha1[20];
1568         enum object_type obj_type;
1569         struct git_istream *st;
1570         git_SHA_CTX c;
1571         char hdr[32];
1572         int hdrlen;
1573
1574         if (map) {
1575                 hash_sha1_file(map, size, type, real_sha1);
1576                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1577         }
1578
1579         st = open_istream(sha1, &obj_type, &size, NULL);
1580         if (!st)
1581                 return -1;
1582
1583         /* Generate the header */
1584         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1585
1586         /* Sha1.. */
1587         git_SHA1_Init(&c);
1588         git_SHA1_Update(&c, hdr, hdrlen);
1589         for (;;) {
1590                 char buf[1024 * 16];
1591                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1592
1593                 if (readlen < 0) {
1594                         close_istream(st);
1595                         return -1;
1596                 }
1597                 if (!readlen)
1598                         break;
1599                 git_SHA1_Update(&c, buf, readlen);
1600         }
1601         git_SHA1_Final(real_sha1, &c);
1602         close_istream(st);
1603         return hashcmp(sha1, real_sha1) ? -1 : 0;
1604 }
1605
1606 int git_open_cloexec(const char *name, int flags)
1607 {
1608         int fd;
1609         static int o_cloexec = O_CLOEXEC;
1610
1611         fd = open(name, flags | o_cloexec);
1612         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
1613                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
1614                 o_cloexec &= ~O_CLOEXEC;
1615                 fd = open(name, flags | o_cloexec);
1616         }
1617
1618 #if defined(F_GETFL) && defined(F_SETFL) && defined(FD_CLOEXEC)
1619         {
1620                 static int fd_cloexec = FD_CLOEXEC;
1621
1622                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
1623                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
1624                         int flags = fcntl(fd, F_GETFL);
1625                         if (fcntl(fd, F_SETFL, flags | fd_cloexec))
1626                                 fd_cloexec = 0;
1627                 }
1628         }
1629 #endif
1630         return fd;
1631 }
1632
1633 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1634 {
1635         struct alternate_object_database *alt;
1636
1637         if (!lstat(sha1_file_name(sha1), st))
1638                 return 0;
1639
1640         prepare_alt_odb();
1641         errno = ENOENT;
1642         for (alt = alt_odb_list; alt; alt = alt->next) {
1643                 const char *path = alt_sha1_path(alt, sha1);
1644                 if (!lstat(path, st))
1645                         return 0;
1646         }
1647
1648         return -1;
1649 }
1650
1651 static int open_sha1_file(const unsigned char *sha1)
1652 {
1653         int fd;
1654         struct alternate_object_database *alt;
1655         int most_interesting_errno;
1656
1657         fd = git_open(sha1_file_name(sha1));
1658         if (fd >= 0)
1659                 return fd;
1660         most_interesting_errno = errno;
1661
1662         prepare_alt_odb();
1663         for (alt = alt_odb_list; alt; alt = alt->next) {
1664                 const char *path = alt_sha1_path(alt, sha1);
1665                 fd = git_open(path);
1666                 if (fd >= 0)
1667                         return fd;
1668                 if (most_interesting_errno == ENOENT)
1669                         most_interesting_errno = errno;
1670         }
1671         errno = most_interesting_errno;
1672         return -1;
1673 }
1674
1675 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1676 {
1677         void *map;
1678         int fd;
1679
1680         fd = open_sha1_file(sha1);
1681         map = NULL;
1682         if (fd >= 0) {
1683                 struct stat st;
1684
1685                 if (!fstat(fd, &st)) {
1686                         *size = xsize_t(st.st_size);
1687                         if (!*size) {
1688                                 /* mmap() is forbidden on empty files */
1689                                 error("object file %s is empty", sha1_file_name(sha1));
1690                                 return NULL;
1691                         }
1692                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1693                 }
1694                 close(fd);
1695         }
1696         return map;
1697 }
1698
1699 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1700                 unsigned long len, enum object_type *type, unsigned long *sizep)
1701 {
1702         unsigned shift;
1703         unsigned long size, c;
1704         unsigned long used = 0;
1705
1706         c = buf[used++];
1707         *type = (c >> 4) & 7;
1708         size = c & 15;
1709         shift = 4;
1710         while (c & 0x80) {
1711                 if (len <= used || bitsizeof(long) <= shift) {
1712                         error("bad object header");
1713                         size = used = 0;
1714                         break;
1715                 }
1716                 c = buf[used++];
1717                 size += (c & 0x7f) << shift;
1718                 shift += 7;
1719         }
1720         *sizep = size;
1721         return used;
1722 }
1723
1724 static int unpack_sha1_short_header(git_zstream *stream,
1725                                     unsigned char *map, unsigned long mapsize,
1726                                     void *buffer, unsigned long bufsiz)
1727 {
1728         /* Get the data stream */
1729         memset(stream, 0, sizeof(*stream));
1730         stream->next_in = map;
1731         stream->avail_in = mapsize;
1732         stream->next_out = buffer;
1733         stream->avail_out = bufsiz;
1734
1735         git_inflate_init(stream);
1736         return git_inflate(stream, 0);
1737 }
1738
1739 int unpack_sha1_header(git_zstream *stream,
1740                        unsigned char *map, unsigned long mapsize,
1741                        void *buffer, unsigned long bufsiz)
1742 {
1743         int status = unpack_sha1_short_header(stream, map, mapsize,
1744                                               buffer, bufsiz);
1745
1746         if (status < Z_OK)
1747                 return status;
1748
1749         /* Make sure we have the terminating NUL */
1750         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1751                 return -1;
1752         return 0;
1753 }
1754
1755 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1756                                         unsigned long mapsize, void *buffer,
1757                                         unsigned long bufsiz, struct strbuf *header)
1758 {
1759         int status;
1760
1761         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
1762         if (status < Z_OK)
1763                 return -1;
1764
1765         /*
1766          * Check if entire header is unpacked in the first iteration.
1767          */
1768         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1769                 return 0;
1770
1771         /*
1772          * buffer[0..bufsiz] was not large enough.  Copy the partial
1773          * result out to header, and then append the result of further
1774          * reading the stream.
1775          */
1776         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1777         stream->next_out = buffer;
1778         stream->avail_out = bufsiz;
1779
1780         do {
1781                 status = git_inflate(stream, 0);
1782                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1783                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1784                         return 0;
1785                 stream->next_out = buffer;
1786                 stream->avail_out = bufsiz;
1787         } while (status != Z_STREAM_END);
1788         return -1;
1789 }
1790
1791 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1792 {
1793         int bytes = strlen(buffer) + 1;
1794         unsigned char *buf = xmallocz(size);
1795         unsigned long n;
1796         int status = Z_OK;
1797
1798         n = stream->total_out - bytes;
1799         if (n > size)
1800                 n = size;
1801         memcpy(buf, (char *) buffer + bytes, n);
1802         bytes = n;
1803         if (bytes <= size) {
1804                 /*
1805                  * The above condition must be (bytes <= size), not
1806                  * (bytes < size).  In other words, even though we
1807                  * expect no more output and set avail_out to zero,
1808                  * the input zlib stream may have bytes that express
1809                  * "this concludes the stream", and we *do* want to
1810                  * eat that input.
1811                  *
1812                  * Otherwise we would not be able to test that we
1813                  * consumed all the input to reach the expected size;
1814                  * we also want to check that zlib tells us that all
1815                  * went well with status == Z_STREAM_END at the end.
1816                  */
1817                 stream->next_out = buf + bytes;
1818                 stream->avail_out = size - bytes;
1819                 while (status == Z_OK)
1820                         status = git_inflate(stream, Z_FINISH);
1821         }
1822         if (status == Z_STREAM_END && !stream->avail_in) {
1823                 git_inflate_end(stream);
1824                 return buf;
1825         }
1826
1827         if (status < 0)
1828                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1829         else if (stream->avail_in)
1830                 error("garbage at end of loose object '%s'",
1831                       sha1_to_hex(sha1));
1832         free(buf);
1833         return NULL;
1834 }
1835
1836 /*
1837  * We used to just use "sscanf()", but that's actually way
1838  * too permissive for what we want to check. So do an anal
1839  * object header parse by hand.
1840  */
1841 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1842                                unsigned int flags)
1843 {
1844         const char *type_buf = hdr;
1845         unsigned long size;
1846         int type, type_len = 0;
1847
1848         /*
1849          * The type can be of any size but is followed by
1850          * a space.
1851          */
1852         for (;;) {
1853                 char c = *hdr++;
1854                 if (!c)
1855                         return -1;
1856                 if (c == ' ')
1857                         break;
1858                 type_len++;
1859         }
1860
1861         type = type_from_string_gently(type_buf, type_len, 1);
1862         if (oi->typename)
1863                 strbuf_add(oi->typename, type_buf, type_len);
1864         /*
1865          * Set type to 0 if its an unknown object and
1866          * we're obtaining the type using '--allow-unknown-type'
1867          * option.
1868          */
1869         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1870                 type = 0;
1871         else if (type < 0)
1872                 die("invalid object type");
1873         if (oi->typep)
1874                 *oi->typep = type;
1875
1876         /*
1877          * The length must follow immediately, and be in canonical
1878          * decimal format (ie "010" is not valid).
1879          */
1880         size = *hdr++ - '0';
1881         if (size > 9)
1882                 return -1;
1883         if (size) {
1884                 for (;;) {
1885                         unsigned long c = *hdr - '0';
1886                         if (c > 9)
1887                                 break;
1888                         hdr++;
1889                         size = size * 10 + c;
1890                 }
1891         }
1892
1893         if (oi->sizep)
1894                 *oi->sizep = size;
1895
1896         /*
1897          * The length must be followed by a zero byte
1898          */
1899         return *hdr ? -1 : type;
1900 }
1901
1902 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1903 {
1904         struct object_info oi = OBJECT_INFO_INIT;
1905
1906         oi.sizep = sizep;
1907         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
1908 }
1909
1910 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1911 {
1912         int ret;
1913         git_zstream stream;
1914         char hdr[8192];
1915
1916         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1917         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1918                 return NULL;
1919
1920         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1921 }
1922
1923 unsigned long get_size_from_delta(struct packed_git *p,
1924                                   struct pack_window **w_curs,
1925                                   off_t curpos)
1926 {
1927         const unsigned char *data;
1928         unsigned char delta_head[20], *in;
1929         git_zstream stream;
1930         int st;
1931
1932         memset(&stream, 0, sizeof(stream));
1933         stream.next_out = delta_head;
1934         stream.avail_out = sizeof(delta_head);
1935
1936         git_inflate_init(&stream);
1937         do {
1938                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1939                 stream.next_in = in;
1940                 st = git_inflate(&stream, Z_FINISH);
1941                 curpos += stream.next_in - in;
1942         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1943                  stream.total_out < sizeof(delta_head));
1944         git_inflate_end(&stream);
1945         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1946                 error("delta data unpack-initial failed");
1947                 return 0;
1948         }
1949
1950         /* Examine the initial part of the delta to figure out
1951          * the result size.
1952          */
1953         data = delta_head;
1954
1955         /* ignore base size */
1956         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1957
1958         /* Read the result size */
1959         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1960 }
1961
1962 static off_t get_delta_base(struct packed_git *p,
1963                                     struct pack_window **w_curs,
1964                                     off_t *curpos,
1965                                     enum object_type type,
1966                                     off_t delta_obj_offset)
1967 {
1968         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1969         off_t base_offset;
1970
1971         /* use_pack() assured us we have [base_info, base_info + 20)
1972          * as a range that we can look at without walking off the
1973          * end of the mapped window.  Its actually the hash size
1974          * that is assured.  An OFS_DELTA longer than the hash size
1975          * is stupid, as then a REF_DELTA would be smaller to store.
1976          */
1977         if (type == OBJ_OFS_DELTA) {
1978                 unsigned used = 0;
1979                 unsigned char c = base_info[used++];
1980                 base_offset = c & 127;
1981                 while (c & 128) {
1982                         base_offset += 1;
1983                         if (!base_offset || MSB(base_offset, 7))
1984                                 return 0;  /* overflow */
1985                         c = base_info[used++];
1986                         base_offset = (base_offset << 7) + (c & 127);
1987                 }
1988                 base_offset = delta_obj_offset - base_offset;
1989                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1990                         return 0;  /* out of bound */
1991                 *curpos += used;
1992         } else if (type == OBJ_REF_DELTA) {
1993                 /* The base entry _must_ be in the same pack */
1994                 base_offset = find_pack_entry_one(base_info, p);
1995                 *curpos += 20;
1996         } else
1997                 die("I am totally screwed");
1998         return base_offset;
1999 }
2000
2001 /*
2002  * Like get_delta_base above, but we return the sha1 instead of the pack
2003  * offset. This means it is cheaper for REF deltas (we do not have to do
2004  * the final object lookup), but more expensive for OFS deltas (we
2005  * have to load the revidx to convert the offset back into a sha1).
2006  */
2007 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
2008                                                 struct pack_window **w_curs,
2009                                                 off_t curpos,
2010                                                 enum object_type type,
2011                                                 off_t delta_obj_offset)
2012 {
2013         if (type == OBJ_REF_DELTA) {
2014                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
2015                 return base;
2016         } else if (type == OBJ_OFS_DELTA) {
2017                 struct revindex_entry *revidx;
2018                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
2019                                                    type, delta_obj_offset);
2020
2021                 if (!base_offset)
2022                         return NULL;
2023
2024                 revidx = find_pack_revindex(p, base_offset);
2025                 if (!revidx)
2026                         return NULL;
2027
2028                 return nth_packed_object_sha1(p, revidx->nr);
2029         } else
2030                 return NULL;
2031 }
2032
2033 int unpack_object_header(struct packed_git *p,
2034                          struct pack_window **w_curs,
2035                          off_t *curpos,
2036                          unsigned long *sizep)
2037 {
2038         unsigned char *base;
2039         unsigned long left;
2040         unsigned long used;
2041         enum object_type type;
2042
2043         /* use_pack() assures us we have [base, base + 20) available
2044          * as a range that we can look at.  (Its actually the hash
2045          * size that is assured.)  With our object header encoding
2046          * the maximum deflated object size is 2^137, which is just
2047          * insane, so we know won't exceed what we have been given.
2048          */
2049         base = use_pack(p, w_curs, *curpos, &left);
2050         used = unpack_object_header_buffer(base, left, &type, sizep);
2051         if (!used) {
2052                 type = OBJ_BAD;
2053         } else
2054                 *curpos += used;
2055
2056         return type;
2057 }
2058
2059 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
2060 {
2061         int type;
2062         struct revindex_entry *revidx;
2063         const unsigned char *sha1;
2064         revidx = find_pack_revindex(p, obj_offset);
2065         if (!revidx)
2066                 return OBJ_BAD;
2067         sha1 = nth_packed_object_sha1(p, revidx->nr);
2068         mark_bad_packed_object(p, sha1);
2069         type = sha1_object_info(sha1, NULL);
2070         if (type <= OBJ_NONE)
2071                 return OBJ_BAD;
2072         return type;
2073 }
2074
2075 #define POI_STACK_PREALLOC 64
2076
2077 static enum object_type packed_to_object_type(struct packed_git *p,
2078                                               off_t obj_offset,
2079                                               enum object_type type,
2080                                               struct pack_window **w_curs,
2081                                               off_t curpos)
2082 {
2083         off_t small_poi_stack[POI_STACK_PREALLOC];
2084         off_t *poi_stack = small_poi_stack;
2085         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
2086
2087         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2088                 off_t base_offset;
2089                 unsigned long size;
2090                 /* Push the object we're going to leave behind */
2091                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
2092                         poi_stack_alloc = alloc_nr(poi_stack_nr);
2093                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
2094                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2095                 } else {
2096                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2097                 }
2098                 poi_stack[poi_stack_nr++] = obj_offset;
2099                 /* If parsing the base offset fails, just unwind */
2100                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2101                 if (!base_offset)
2102                         goto unwind;
2103                 curpos = obj_offset = base_offset;
2104                 type = unpack_object_header(p, w_curs, &curpos, &size);
2105                 if (type <= OBJ_NONE) {
2106                         /* If getting the base itself fails, we first
2107                          * retry the base, otherwise unwind */
2108                         type = retry_bad_packed_offset(p, base_offset);
2109                         if (type > OBJ_NONE)
2110                                 goto out;
2111                         goto unwind;
2112                 }
2113         }
2114
2115         switch (type) {
2116         case OBJ_BAD:
2117         case OBJ_COMMIT:
2118         case OBJ_TREE:
2119         case OBJ_BLOB:
2120         case OBJ_TAG:
2121                 break;
2122         default:
2123                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2124                       type, (uintmax_t)obj_offset, p->pack_name);
2125                 type = OBJ_BAD;
2126         }
2127
2128 out:
2129         if (poi_stack != small_poi_stack)
2130                 free(poi_stack);
2131         return type;
2132
2133 unwind:
2134         while (poi_stack_nr) {
2135                 obj_offset = poi_stack[--poi_stack_nr];
2136                 type = retry_bad_packed_offset(p, obj_offset);
2137                 if (type > OBJ_NONE)
2138                         goto out;
2139         }
2140         type = OBJ_BAD;
2141         goto out;
2142 }
2143
2144 int packed_object_info(struct packed_git *p, off_t obj_offset,
2145                        struct object_info *oi)
2146 {
2147         struct pack_window *w_curs = NULL;
2148         unsigned long size;
2149         off_t curpos = obj_offset;
2150         enum object_type type;
2151
2152         /*
2153          * We always get the representation type, but only convert it to
2154          * a "real" type later if the caller is interested.
2155          */
2156         type = unpack_object_header(p, &w_curs, &curpos, &size);
2157
2158         if (oi->sizep) {
2159                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2160                         off_t tmp_pos = curpos;
2161                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2162                                                            type, obj_offset);
2163                         if (!base_offset) {
2164                                 type = OBJ_BAD;
2165                                 goto out;
2166                         }
2167                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2168                         if (*oi->sizep == 0) {
2169                                 type = OBJ_BAD;
2170                                 goto out;
2171                         }
2172                 } else {
2173                         *oi->sizep = size;
2174                 }
2175         }
2176
2177         if (oi->disk_sizep) {
2178                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2179                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2180         }
2181
2182         if (oi->typep) {
2183                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2184                 if (*oi->typep < 0) {
2185                         type = OBJ_BAD;
2186                         goto out;
2187                 }
2188         }
2189
2190         if (oi->delta_base_sha1) {
2191                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2192                         const unsigned char *base;
2193
2194                         base = get_delta_base_sha1(p, &w_curs, curpos,
2195                                                    type, obj_offset);
2196                         if (!base) {
2197                                 type = OBJ_BAD;
2198                                 goto out;
2199                         }
2200
2201                         hashcpy(oi->delta_base_sha1, base);
2202                 } else
2203                         hashclr(oi->delta_base_sha1);
2204         }
2205
2206 out:
2207         unuse_pack(&w_curs);
2208         return type;
2209 }
2210
2211 static void *unpack_compressed_entry(struct packed_git *p,
2212                                     struct pack_window **w_curs,
2213                                     off_t curpos,
2214                                     unsigned long size)
2215 {
2216         int st;
2217         git_zstream stream;
2218         unsigned char *buffer, *in;
2219
2220         buffer = xmallocz_gently(size);
2221         if (!buffer)
2222                 return NULL;
2223         memset(&stream, 0, sizeof(stream));
2224         stream.next_out = buffer;
2225         stream.avail_out = size + 1;
2226
2227         git_inflate_init(&stream);
2228         do {
2229                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2230                 stream.next_in = in;
2231                 st = git_inflate(&stream, Z_FINISH);
2232                 if (!stream.avail_out)
2233                         break; /* the payload is larger than it should be */
2234                 curpos += stream.next_in - in;
2235         } while (st == Z_OK || st == Z_BUF_ERROR);
2236         git_inflate_end(&stream);
2237         if ((st != Z_STREAM_END) || stream.total_out != size) {
2238                 free(buffer);
2239                 return NULL;
2240         }
2241
2242         return buffer;
2243 }
2244
2245 static struct hashmap delta_base_cache;
2246 static size_t delta_base_cached;
2247
2248 static LIST_HEAD(delta_base_cache_lru);
2249
2250 struct delta_base_cache_key {
2251         struct packed_git *p;
2252         off_t base_offset;
2253 };
2254
2255 struct delta_base_cache_entry {
2256         struct hashmap hash;
2257         struct delta_base_cache_key key;
2258         struct list_head lru;
2259         void *data;
2260         unsigned long size;
2261         enum object_type type;
2262 };
2263
2264 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2265 {
2266         unsigned int hash;
2267
2268         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2269         hash += (hash >> 8) + (hash >> 16);
2270         return hash;
2271 }
2272
2273 static struct delta_base_cache_entry *
2274 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2275 {
2276         struct hashmap_entry entry;
2277         struct delta_base_cache_key key;
2278
2279         if (!delta_base_cache.cmpfn)
2280                 return NULL;
2281
2282         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2283         key.p = p;
2284         key.base_offset = base_offset;
2285         return hashmap_get(&delta_base_cache, &entry, &key);
2286 }
2287
2288 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2289                                    const struct delta_base_cache_key *b)
2290 {
2291         return a->p == b->p && a->base_offset == b->base_offset;
2292 }
2293
2294 static int delta_base_cache_hash_cmp(const void *va, const void *vb,
2295                                      const void *vkey)
2296 {
2297         const struct delta_base_cache_entry *a = va, *b = vb;
2298         const struct delta_base_cache_key *key = vkey;
2299         if (key)
2300                 return !delta_base_cache_key_eq(&a->key, key);
2301         else
2302                 return !delta_base_cache_key_eq(&a->key, &b->key);
2303 }
2304
2305 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2306 {
2307         return !!get_delta_base_cache_entry(p, base_offset);
2308 }
2309
2310 /*
2311  * Remove the entry from the cache, but do _not_ free the associated
2312  * entry data. The caller takes ownership of the "data" buffer, and
2313  * should copy out any fields it wants before detaching.
2314  */
2315 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2316 {
2317         hashmap_remove(&delta_base_cache, ent, &ent->key);
2318         list_del(&ent->lru);
2319         delta_base_cached -= ent->size;
2320         free(ent);
2321 }
2322
2323 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2324         unsigned long *base_size, enum object_type *type)
2325 {
2326         struct delta_base_cache_entry *ent;
2327
2328         ent = get_delta_base_cache_entry(p, base_offset);
2329         if (!ent)
2330                 return unpack_entry(p, base_offset, type, base_size);
2331
2332         *type = ent->type;
2333         *base_size = ent->size;
2334         return xmemdupz(ent->data, ent->size);
2335 }
2336
2337 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2338 {
2339         free(ent->data);
2340         detach_delta_base_cache_entry(ent);
2341 }
2342
2343 void clear_delta_base_cache(void)
2344 {
2345         struct hashmap_iter iter;
2346         struct delta_base_cache_entry *entry;
2347         for (entry = hashmap_iter_first(&delta_base_cache, &iter);
2348              entry;
2349              entry = hashmap_iter_next(&iter)) {
2350                 release_delta_base_cache(entry);
2351         }
2352 }
2353
2354 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2355         void *base, unsigned long base_size, enum object_type type)
2356 {
2357         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2358         struct list_head *lru, *tmp;
2359
2360         delta_base_cached += base_size;
2361
2362         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2363                 struct delta_base_cache_entry *f =
2364                         list_entry(lru, struct delta_base_cache_entry, lru);
2365                 if (delta_base_cached <= delta_base_cache_limit)
2366                         break;
2367                 release_delta_base_cache(f);
2368         }
2369
2370         ent->key.p = p;
2371         ent->key.base_offset = base_offset;
2372         ent->type = type;
2373         ent->data = base;
2374         ent->size = base_size;
2375         list_add_tail(&ent->lru, &delta_base_cache_lru);
2376
2377         if (!delta_base_cache.cmpfn)
2378                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, 0);
2379         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2380         hashmap_add(&delta_base_cache, ent);
2381 }
2382
2383 static void *read_object(const unsigned char *sha1, enum object_type *type,
2384                          unsigned long *size);
2385
2386 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2387 {
2388         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2389         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2390                          p->pack_name, (uintmax_t)obj_offset);
2391 }
2392
2393 int do_check_packed_object_crc;
2394
2395 #define UNPACK_ENTRY_STACK_PREALLOC 64
2396 struct unpack_entry_stack_ent {
2397         off_t obj_offset;
2398         off_t curpos;
2399         unsigned long size;
2400 };
2401
2402 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2403                    enum object_type *final_type, unsigned long *final_size)
2404 {
2405         struct pack_window *w_curs = NULL;
2406         off_t curpos = obj_offset;
2407         void *data = NULL;
2408         unsigned long size;
2409         enum object_type type;
2410         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2411         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2412         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2413         int base_from_cache = 0;
2414
2415         write_pack_access_log(p, obj_offset);
2416
2417         /* PHASE 1: drill down to the innermost base object */
2418         for (;;) {
2419                 off_t base_offset;
2420                 int i;
2421                 struct delta_base_cache_entry *ent;
2422
2423                 ent = get_delta_base_cache_entry(p, curpos);
2424                 if (ent) {
2425                         type = ent->type;
2426                         data = ent->data;
2427                         size = ent->size;
2428                         detach_delta_base_cache_entry(ent);
2429                         base_from_cache = 1;
2430                         break;
2431                 }
2432
2433                 if (do_check_packed_object_crc && p->index_version > 1) {
2434                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2435                         off_t len = revidx[1].offset - obj_offset;
2436                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2437                                 const unsigned char *sha1 =
2438                                         nth_packed_object_sha1(p, revidx->nr);
2439                                 error("bad packed object CRC for %s",
2440                                       sha1_to_hex(sha1));
2441                                 mark_bad_packed_object(p, sha1);
2442                                 unuse_pack(&w_curs);
2443                                 return NULL;
2444                         }
2445                 }
2446
2447                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2448                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2449                         break;
2450
2451                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2452                 if (!base_offset) {
2453                         error("failed to validate delta base reference "
2454                               "at offset %"PRIuMAX" from %s",
2455                               (uintmax_t)curpos, p->pack_name);
2456                         /* bail to phase 2, in hopes of recovery */
2457                         data = NULL;
2458                         break;
2459                 }
2460
2461                 /* push object, proceed to base */
2462                 if (delta_stack_nr >= delta_stack_alloc
2463                     && delta_stack == small_delta_stack) {
2464                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2465                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2466                         memcpy(delta_stack, small_delta_stack,
2467                                sizeof(*delta_stack)*delta_stack_nr);
2468                 } else {
2469                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2470                 }
2471                 i = delta_stack_nr++;
2472                 delta_stack[i].obj_offset = obj_offset;
2473                 delta_stack[i].curpos = curpos;
2474                 delta_stack[i].size = size;
2475
2476                 curpos = obj_offset = base_offset;
2477         }
2478
2479         /* PHASE 2: handle the base */
2480         switch (type) {
2481         case OBJ_OFS_DELTA:
2482         case OBJ_REF_DELTA:
2483                 if (data)
2484                         die("BUG: unpack_entry: left loop at a valid delta");
2485                 break;
2486         case OBJ_COMMIT:
2487         case OBJ_TREE:
2488         case OBJ_BLOB:
2489         case OBJ_TAG:
2490                 if (!base_from_cache)
2491                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2492                 break;
2493         default:
2494                 data = NULL;
2495                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2496                       type, (uintmax_t)obj_offset, p->pack_name);
2497         }
2498
2499         /* PHASE 3: apply deltas in order */
2500
2501         /* invariants:
2502          *   'data' holds the base data, or NULL if there was corruption
2503          */
2504         while (delta_stack_nr) {
2505                 void *delta_data;
2506                 void *base = data;
2507                 unsigned long delta_size, base_size = size;
2508                 int i;
2509
2510                 data = NULL;
2511
2512                 if (base)
2513                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2514
2515                 if (!base) {
2516                         /*
2517                          * We're probably in deep shit, but let's try to fetch
2518                          * the required base anyway from another pack or loose.
2519                          * This is costly but should happen only in the presence
2520                          * of a corrupted pack, and is better than failing outright.
2521                          */
2522                         struct revindex_entry *revidx;
2523                         const unsigned char *base_sha1;
2524                         revidx = find_pack_revindex(p, obj_offset);
2525                         if (revidx) {
2526                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2527                                 error("failed to read delta base object %s"
2528                                       " at offset %"PRIuMAX" from %s",
2529                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2530                                       p->pack_name);
2531                                 mark_bad_packed_object(p, base_sha1);
2532                                 base = read_object(base_sha1, &type, &base_size);
2533                         }
2534                 }
2535
2536                 i = --delta_stack_nr;
2537                 obj_offset = delta_stack[i].obj_offset;
2538                 curpos = delta_stack[i].curpos;
2539                 delta_size = delta_stack[i].size;
2540
2541                 if (!base)
2542                         continue;
2543
2544                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2545
2546                 if (!delta_data) {
2547                         error("failed to unpack compressed delta "
2548                               "at offset %"PRIuMAX" from %s",
2549                               (uintmax_t)curpos, p->pack_name);
2550                         data = NULL;
2551                         continue;
2552                 }
2553
2554                 data = patch_delta(base, base_size,
2555                                    delta_data, delta_size,
2556                                    &size);
2557
2558                 /*
2559                  * We could not apply the delta; warn the user, but keep going.
2560                  * Our failure will be noticed either in the next iteration of
2561                  * the loop, or if this is the final delta, in the caller when
2562                  * we return NULL. Those code paths will take care of making
2563                  * a more explicit warning and retrying with another copy of
2564                  * the object.
2565                  */
2566                 if (!data)
2567                         error("failed to apply delta");
2568
2569                 free(delta_data);
2570         }
2571
2572         *final_type = type;
2573         *final_size = size;
2574
2575         unuse_pack(&w_curs);
2576
2577         if (delta_stack != small_delta_stack)
2578                 free(delta_stack);
2579
2580         return data;
2581 }
2582
2583 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2584                                             uint32_t n)
2585 {
2586         const unsigned char *index = p->index_data;
2587         if (!index) {
2588                 if (open_pack_index(p))
2589                         return NULL;
2590                 index = p->index_data;
2591         }
2592         if (n >= p->num_objects)
2593                 return NULL;
2594         index += 4 * 256;
2595         if (p->index_version == 1) {
2596                 return index + 24 * n + 4;
2597         } else {
2598                 index += 8;
2599                 return index + 20 * n;
2600         }
2601 }
2602
2603 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2604 {
2605         const unsigned char *ptr = vptr;
2606         const unsigned char *start = p->index_data;
2607         const unsigned char *end = start + p->index_size;
2608         if (ptr < start)
2609                 die(_("offset before start of pack index for %s (corrupt index?)"),
2610                     p->pack_name);
2611         /* No need to check for underflow; .idx files must be at least 8 bytes */
2612         if (ptr >= end - 8)
2613                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2614                     p->pack_name);
2615 }
2616
2617 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2618 {
2619         const unsigned char *index = p->index_data;
2620         index += 4 * 256;
2621         if (p->index_version == 1) {
2622                 return ntohl(*((uint32_t *)(index + 24 * n)));
2623         } else {
2624                 uint32_t off;
2625                 index += 8 + p->num_objects * (20 + 4);
2626                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2627                 if (!(off & 0x80000000))
2628                         return off;
2629                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2630                 check_pack_index_ptr(p, index);
2631                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2632                                    ntohl(*((uint32_t *)(index + 4)));
2633         }
2634 }
2635
2636 off_t find_pack_entry_one(const unsigned char *sha1,
2637                                   struct packed_git *p)
2638 {
2639         const uint32_t *level1_ofs = p->index_data;
2640         const unsigned char *index = p->index_data;
2641         unsigned hi, lo, stride;
2642         static int use_lookup = -1;
2643         static int debug_lookup = -1;
2644
2645         if (debug_lookup < 0)
2646                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2647
2648         if (!index) {
2649                 if (open_pack_index(p))
2650                         return 0;
2651                 level1_ofs = p->index_data;
2652                 index = p->index_data;
2653         }
2654         if (p->index_version > 1) {
2655                 level1_ofs += 2;
2656                 index += 8;
2657         }
2658         index += 4 * 256;
2659         hi = ntohl(level1_ofs[*sha1]);
2660         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2661         if (p->index_version > 1) {
2662                 stride = 20;
2663         } else {
2664                 stride = 24;
2665                 index += 4;
2666         }
2667
2668         if (debug_lookup)
2669                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2670                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2671
2672         if (use_lookup < 0)
2673                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2674         if (use_lookup) {
2675                 int pos = sha1_entry_pos(index, stride, 0,
2676                                          lo, hi, p->num_objects, sha1);
2677                 if (pos < 0)
2678                         return 0;
2679                 return nth_packed_object_offset(p, pos);
2680         }
2681
2682         do {
2683                 unsigned mi = (lo + hi) / 2;
2684                 int cmp = hashcmp(index + mi * stride, sha1);
2685
2686                 if (debug_lookup)
2687                         printf("lo %u hi %u rg %u mi %u\n",
2688                                lo, hi, hi - lo, mi);
2689                 if (!cmp)
2690                         return nth_packed_object_offset(p, mi);
2691                 if (cmp > 0)
2692                         hi = mi;
2693                 else
2694                         lo = mi+1;
2695         } while (lo < hi);
2696         return 0;
2697 }
2698
2699 int is_pack_valid(struct packed_git *p)
2700 {
2701         /* An already open pack is known to be valid. */
2702         if (p->pack_fd != -1)
2703                 return 1;
2704
2705         /* If the pack has one window completely covering the
2706          * file size, the pack is known to be valid even if
2707          * the descriptor is not currently open.
2708          */
2709         if (p->windows) {
2710                 struct pack_window *w = p->windows;
2711
2712                 if (!w->offset && w->len == p->pack_size)
2713                         return 1;
2714         }
2715
2716         /* Force the pack to open to prove its valid. */
2717         return !open_packed_git(p);
2718 }
2719
2720 static int fill_pack_entry(const unsigned char *sha1,
2721                            struct pack_entry *e,
2722                            struct packed_git *p)
2723 {
2724         off_t offset;
2725
2726         if (p->num_bad_objects) {
2727                 unsigned i;
2728                 for (i = 0; i < p->num_bad_objects; i++)
2729                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2730                                 return 0;
2731         }
2732
2733         offset = find_pack_entry_one(sha1, p);
2734         if (!offset)
2735                 return 0;
2736
2737         /*
2738          * We are about to tell the caller where they can locate the
2739          * requested object.  We better make sure the packfile is
2740          * still here and can be accessed before supplying that
2741          * answer, as it may have been deleted since the index was
2742          * loaded!
2743          */
2744         if (!is_pack_valid(p))
2745                 return 0;
2746         e->offset = offset;
2747         e->p = p;
2748         hashcpy(e->sha1, sha1);
2749         return 1;
2750 }
2751
2752 /*
2753  * Iff a pack file contains the object named by sha1, return true and
2754  * store its location to e.
2755  */
2756 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2757 {
2758         struct mru_entry *p;
2759
2760         prepare_packed_git();
2761         if (!packed_git)
2762                 return 0;
2763
2764         for (p = packed_git_mru->head; p; p = p->next) {
2765                 if (fill_pack_entry(sha1, e, p->item)) {
2766                         mru_mark(packed_git_mru, p);
2767                         return 1;
2768                 }
2769         }
2770         return 0;
2771 }
2772
2773 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2774                                   struct packed_git *packs)
2775 {
2776         struct packed_git *p;
2777
2778         for (p = packs; p; p = p->next) {
2779                 if (find_pack_entry_one(sha1, p))
2780                         return p;
2781         }
2782         return NULL;
2783
2784 }
2785
2786 static int sha1_loose_object_info(const unsigned char *sha1,
2787                                   struct object_info *oi,
2788                                   int flags)
2789 {
2790         int status = 0;
2791         unsigned long mapsize;
2792         void *map;
2793         git_zstream stream;
2794         char hdr[32];
2795         struct strbuf hdrbuf = STRBUF_INIT;
2796
2797         if (oi->delta_base_sha1)
2798                 hashclr(oi->delta_base_sha1);
2799
2800         /*
2801          * If we don't care about type or size, then we don't
2802          * need to look inside the object at all. Note that we
2803          * do not optimize out the stat call, even if the
2804          * caller doesn't care about the disk-size, since our
2805          * return value implicitly indicates whether the
2806          * object even exists.
2807          */
2808         if (!oi->typep && !oi->typename && !oi->sizep) {
2809                 struct stat st;
2810                 if (stat_sha1_file(sha1, &st) < 0)
2811                         return -1;
2812                 if (oi->disk_sizep)
2813                         *oi->disk_sizep = st.st_size;
2814                 return 0;
2815         }
2816
2817         map = map_sha1_file(sha1, &mapsize);
2818         if (!map)
2819                 return -1;
2820         if (oi->disk_sizep)
2821                 *oi->disk_sizep = mapsize;
2822         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2823                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2824                         status = error("unable to unpack %s header with --allow-unknown-type",
2825                                        sha1_to_hex(sha1));
2826         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2827                 status = error("unable to unpack %s header",
2828                                sha1_to_hex(sha1));
2829         if (status < 0)
2830                 ; /* Do nothing */
2831         else if (hdrbuf.len) {
2832                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2833                         status = error("unable to parse %s header with --allow-unknown-type",
2834                                        sha1_to_hex(sha1));
2835         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2836                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2837         git_inflate_end(&stream);
2838         munmap(map, mapsize);
2839         if (status && oi->typep)
2840                 *oi->typep = status;
2841         strbuf_release(&hdrbuf);
2842         return 0;
2843 }
2844
2845 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2846 {
2847         struct cached_object *co;
2848         struct pack_entry e;
2849         int rtype;
2850         enum object_type real_type;
2851         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2852
2853         co = find_cached_object(real);
2854         if (co) {
2855                 if (oi->typep)
2856                         *(oi->typep) = co->type;
2857                 if (oi->sizep)
2858                         *(oi->sizep) = co->size;
2859                 if (oi->disk_sizep)
2860                         *(oi->disk_sizep) = 0;
2861                 if (oi->delta_base_sha1)
2862                         hashclr(oi->delta_base_sha1);
2863                 if (oi->typename)
2864                         strbuf_addstr(oi->typename, typename(co->type));
2865                 oi->whence = OI_CACHED;
2866                 return 0;
2867         }
2868
2869         if (!find_pack_entry(real, &e)) {
2870                 /* Most likely it's a loose object. */
2871                 if (!sha1_loose_object_info(real, oi, flags)) {
2872                         oi->whence = OI_LOOSE;
2873                         return 0;
2874                 }
2875
2876                 /* Not a loose object; someone else may have just packed it. */
2877                 reprepare_packed_git();
2878                 if (!find_pack_entry(real, &e))
2879                         return -1;
2880         }
2881
2882         /*
2883          * packed_object_info() does not follow the delta chain to
2884          * find out the real type, unless it is given oi->typep.
2885          */
2886         if (oi->typename && !oi->typep)
2887                 oi->typep = &real_type;
2888
2889         rtype = packed_object_info(e.p, e.offset, oi);
2890         if (rtype < 0) {
2891                 mark_bad_packed_object(e.p, real);
2892                 if (oi->typep == &real_type)
2893                         oi->typep = NULL;
2894                 return sha1_object_info_extended(real, oi, 0);
2895         } else if (in_delta_base_cache(e.p, e.offset)) {
2896                 oi->whence = OI_DBCACHED;
2897         } else {
2898                 oi->whence = OI_PACKED;
2899                 oi->u.packed.offset = e.offset;
2900                 oi->u.packed.pack = e.p;
2901                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2902                                          rtype == OBJ_OFS_DELTA);
2903         }
2904         if (oi->typename)
2905                 strbuf_addstr(oi->typename, typename(*oi->typep));
2906         if (oi->typep == &real_type)
2907                 oi->typep = NULL;
2908
2909         return 0;
2910 }
2911
2912 /* returns enum object_type or negative */
2913 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2914 {
2915         enum object_type type;
2916         struct object_info oi = OBJECT_INFO_INIT;
2917
2918         oi.typep = &type;
2919         oi.sizep = sizep;
2920         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2921                 return -1;
2922         return type;
2923 }
2924
2925 static void *read_packed_sha1(const unsigned char *sha1,
2926                               enum object_type *type, unsigned long *size)
2927 {
2928         struct pack_entry e;
2929         void *data;
2930
2931         if (!find_pack_entry(sha1, &e))
2932                 return NULL;
2933         data = cache_or_unpack_entry(e.p, e.offset, size, type);
2934         if (!data) {
2935                 /*
2936                  * We're probably in deep shit, but let's try to fetch
2937                  * the required object anyway from another pack or loose.
2938                  * This should happen only in the presence of a corrupted
2939                  * pack, and is better than failing outright.
2940                  */
2941                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2942                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2943                 mark_bad_packed_object(e.p, sha1);
2944                 data = read_object(sha1, type, size);
2945         }
2946         return data;
2947 }
2948
2949 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2950                       unsigned char *sha1)
2951 {
2952         struct cached_object *co;
2953
2954         hash_sha1_file(buf, len, typename(type), sha1);
2955         if (has_sha1_file(sha1) || find_cached_object(sha1))
2956                 return 0;
2957         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2958         co = &cached_objects[cached_object_nr++];
2959         co->size = len;
2960         co->type = type;
2961         co->buf = xmalloc(len);
2962         memcpy(co->buf, buf, len);
2963         hashcpy(co->sha1, sha1);
2964         return 0;
2965 }
2966
2967 static void *read_object(const unsigned char *sha1, enum object_type *type,
2968                          unsigned long *size)
2969 {
2970         unsigned long mapsize;
2971         void *map, *buf;
2972         struct cached_object *co;
2973
2974         co = find_cached_object(sha1);
2975         if (co) {
2976                 *type = co->type;
2977                 *size = co->size;
2978                 return xmemdupz(co->buf, co->size);
2979         }
2980
2981         buf = read_packed_sha1(sha1, type, size);
2982         if (buf)
2983                 return buf;
2984         map = map_sha1_file(sha1, &mapsize);
2985         if (map) {
2986                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2987                 munmap(map, mapsize);
2988                 return buf;
2989         }
2990         reprepare_packed_git();
2991         return read_packed_sha1(sha1, type, size);
2992 }
2993
2994 /*
2995  * This function dies on corrupt objects; the callers who want to
2996  * deal with them should arrange to call read_object() and give error
2997  * messages themselves.
2998  */
2999 void *read_sha1_file_extended(const unsigned char *sha1,
3000                               enum object_type *type,
3001                               unsigned long *size,
3002                               unsigned flag)
3003 {
3004         void *data;
3005         const struct packed_git *p;
3006         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
3007
3008         errno = 0;
3009         data = read_object(repl, type, size);
3010         if (data)
3011                 return data;
3012
3013         if (errno && errno != ENOENT)
3014                 die_errno("failed to read object %s", sha1_to_hex(sha1));
3015
3016         /* die if we replaced an object with one that does not exist */
3017         if (repl != sha1)
3018                 die("replacement %s not found for %s",
3019                     sha1_to_hex(repl), sha1_to_hex(sha1));
3020
3021         if (has_loose_object(repl)) {
3022                 const char *path = sha1_file_name(sha1);
3023
3024                 die("loose object %s (stored in %s) is corrupt",
3025                     sha1_to_hex(repl), path);
3026         }
3027
3028         if ((p = has_packed_and_bad(repl)) != NULL)
3029                 die("packed object %s (stored in %s) is corrupt",
3030                     sha1_to_hex(repl), p->pack_name);
3031
3032         return NULL;
3033 }
3034
3035 void *read_object_with_reference(const unsigned char *sha1,
3036                                  const char *required_type_name,
3037                                  unsigned long *size,
3038                                  unsigned char *actual_sha1_return)
3039 {
3040         enum object_type type, required_type;
3041         void *buffer;
3042         unsigned long isize;
3043         unsigned char actual_sha1[20];
3044
3045         required_type = type_from_string(required_type_name);
3046         hashcpy(actual_sha1, sha1);
3047         while (1) {
3048                 int ref_length = -1;
3049                 const char *ref_type = NULL;
3050
3051                 buffer = read_sha1_file(actual_sha1, &type, &isize);
3052                 if (!buffer)
3053                         return NULL;
3054                 if (type == required_type) {
3055                         *size = isize;
3056                         if (actual_sha1_return)
3057                                 hashcpy(actual_sha1_return, actual_sha1);
3058                         return buffer;
3059                 }
3060                 /* Handle references */
3061                 else if (type == OBJ_COMMIT)
3062                         ref_type = "tree ";
3063                 else if (type == OBJ_TAG)
3064                         ref_type = "object ";
3065                 else {
3066                         free(buffer);
3067                         return NULL;
3068                 }
3069                 ref_length = strlen(ref_type);
3070
3071                 if (ref_length + 40 > isize ||
3072                     memcmp(buffer, ref_type, ref_length) ||
3073                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
3074                         free(buffer);
3075                         return NULL;
3076                 }
3077                 free(buffer);
3078                 /* Now we have the ID of the referred-to object in
3079                  * actual_sha1.  Check again. */
3080         }
3081 }
3082
3083 static void write_sha1_file_prepare(const void *buf, unsigned long len,
3084                                     const char *type, unsigned char *sha1,
3085                                     char *hdr, int *hdrlen)
3086 {
3087         git_SHA_CTX c;
3088
3089         /* Generate the header */
3090         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
3091
3092         /* Sha1.. */
3093         git_SHA1_Init(&c);
3094         git_SHA1_Update(&c, hdr, *hdrlen);
3095         git_SHA1_Update(&c, buf, len);
3096         git_SHA1_Final(sha1, &c);
3097 }
3098
3099 /*
3100  * Move the just written object into its final resting place.
3101  */
3102 int finalize_object_file(const char *tmpfile, const char *filename)
3103 {
3104         int ret = 0;
3105
3106         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3107                 goto try_rename;
3108         else if (link(tmpfile, filename))
3109                 ret = errno;
3110
3111         /*
3112          * Coda hack - coda doesn't like cross-directory links,
3113          * so we fall back to a rename, which will mean that it
3114          * won't be able to check collisions, but that's not a
3115          * big deal.
3116          *
3117          * The same holds for FAT formatted media.
3118          *
3119          * When this succeeds, we just return.  We have nothing
3120          * left to unlink.
3121          */
3122         if (ret && ret != EEXIST) {
3123         try_rename:
3124                 if (!rename(tmpfile, filename))
3125                         goto out;
3126                 ret = errno;
3127         }
3128         unlink_or_warn(tmpfile);
3129         if (ret) {
3130                 if (ret != EEXIST) {
3131                         return error_errno("unable to write sha1 filename %s", filename);
3132                 }
3133                 /* FIXME!!! Collision check here ? */
3134         }
3135
3136 out:
3137         if (adjust_shared_perm(filename))
3138                 return error("unable to set permission to '%s'", filename);
3139         return 0;
3140 }
3141
3142 static int write_buffer(int fd, const void *buf, size_t len)
3143 {
3144         if (write_in_full(fd, buf, len) < 0)
3145                 return error_errno("file write error");
3146         return 0;
3147 }
3148
3149 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3150                    unsigned char *sha1)
3151 {
3152         char hdr[32];
3153         int hdrlen = sizeof(hdr);
3154         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3155         return 0;
3156 }
3157
3158 /* Finalize a file on disk, and close it. */
3159 static void close_sha1_file(int fd)
3160 {
3161         if (fsync_object_files)
3162                 fsync_or_die(fd, "sha1 file");
3163         if (close(fd) != 0)
3164                 die_errno("error when closing sha1 file");
3165 }
3166
3167 /* Size of directory component, including the ending '/' */
3168 static inline int directory_size(const char *filename)
3169 {
3170         const char *s = strrchr(filename, '/');
3171         if (!s)
3172                 return 0;
3173         return s - filename + 1;
3174 }
3175
3176 /*
3177  * This creates a temporary file in the same directory as the final
3178  * 'filename'
3179  *
3180  * We want to avoid cross-directory filename renames, because those
3181  * can have problems on various filesystems (FAT, NFS, Coda).
3182  */
3183 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3184 {
3185         int fd, dirlen = directory_size(filename);
3186
3187         strbuf_reset(tmp);
3188         strbuf_add(tmp, filename, dirlen);
3189         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3190         fd = git_mkstemp_mode(tmp->buf, 0444);
3191         if (fd < 0 && dirlen && errno == ENOENT) {
3192                 /*
3193                  * Make sure the directory exists; note that the contents
3194                  * of the buffer are undefined after mkstemp returns an
3195                  * error, so we have to rewrite the whole buffer from
3196                  * scratch.
3197                  */
3198                 strbuf_reset(tmp);
3199                 strbuf_add(tmp, filename, dirlen - 1);
3200                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3201                         return -1;
3202                 if (adjust_shared_perm(tmp->buf))
3203                         return -1;
3204
3205                 /* Try again */
3206                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3207                 fd = git_mkstemp_mode(tmp->buf, 0444);
3208         }
3209         return fd;
3210 }
3211
3212 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3213                               const void *buf, unsigned long len, time_t mtime)
3214 {
3215         int fd, ret;
3216         unsigned char compressed[4096];
3217         git_zstream stream;
3218         git_SHA_CTX c;
3219         unsigned char parano_sha1[20];
3220         static struct strbuf tmp_file = STRBUF_INIT;
3221         const char *filename = sha1_file_name(sha1);
3222
3223         fd = create_tmpfile(&tmp_file, filename);
3224         if (fd < 0) {
3225                 if (errno == EACCES)
3226                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3227                 else
3228                         return error_errno("unable to create temporary file");
3229         }
3230
3231         /* Set it up */
3232         git_deflate_init(&stream, zlib_compression_level);
3233         stream.next_out = compressed;
3234         stream.avail_out = sizeof(compressed);
3235         git_SHA1_Init(&c);
3236
3237         /* First header.. */
3238         stream.next_in = (unsigned char *)hdr;
3239         stream.avail_in = hdrlen;
3240         while (git_deflate(&stream, 0) == Z_OK)
3241                 ; /* nothing */
3242         git_SHA1_Update(&c, hdr, hdrlen);
3243
3244         /* Then the data itself.. */
3245         stream.next_in = (void *)buf;
3246         stream.avail_in = len;
3247         do {
3248                 unsigned char *in0 = stream.next_in;
3249                 ret = git_deflate(&stream, Z_FINISH);
3250                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3251                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3252                         die("unable to write sha1 file");
3253                 stream.next_out = compressed;
3254                 stream.avail_out = sizeof(compressed);
3255         } while (ret == Z_OK);
3256
3257         if (ret != Z_STREAM_END)
3258                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3259         ret = git_deflate_end_gently(&stream);
3260         if (ret != Z_OK)
3261                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3262         git_SHA1_Final(parano_sha1, &c);
3263         if (hashcmp(sha1, parano_sha1) != 0)
3264                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3265
3266         close_sha1_file(fd);
3267
3268         if (mtime) {
3269                 struct utimbuf utb;
3270                 utb.actime = mtime;
3271                 utb.modtime = mtime;
3272                 if (utime(tmp_file.buf, &utb) < 0)
3273                         warning_errno("failed utime() on %s", tmp_file.buf);
3274         }
3275
3276         return finalize_object_file(tmp_file.buf, filename);
3277 }
3278
3279 static int freshen_loose_object(const unsigned char *sha1)
3280 {
3281         return check_and_freshen(sha1, 1);
3282 }
3283
3284 static int freshen_packed_object(const unsigned char *sha1)
3285 {
3286         struct pack_entry e;
3287         if (!find_pack_entry(sha1, &e))
3288                 return 0;
3289         if (e.p->freshened)
3290                 return 1;
3291         if (!freshen_file(e.p->pack_name))
3292                 return 0;
3293         e.p->freshened = 1;
3294         return 1;
3295 }
3296
3297 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3298 {
3299         char hdr[32];
3300         int hdrlen = sizeof(hdr);
3301
3302         /* Normally if we have it in the pack then we do not bother writing
3303          * it out into .git/objects/??/?{38} file.
3304          */
3305         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3306         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3307                 return 0;
3308         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3309 }
3310
3311 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3312                              unsigned char *sha1, unsigned flags)
3313 {
3314         char *header;
3315         int hdrlen, status = 0;
3316
3317         /* type string, SP, %lu of the length plus NUL must fit this */
3318         hdrlen = strlen(type) + 32;
3319         header = xmalloc(hdrlen);
3320         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3321
3322         if (!(flags & HASH_WRITE_OBJECT))
3323                 goto cleanup;
3324         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3325                 goto cleanup;
3326         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3327
3328 cleanup:
3329         free(header);
3330         return status;
3331 }
3332
3333 int force_object_loose(const unsigned char *sha1, time_t mtime)
3334 {
3335         void *buf;
3336         unsigned long len;
3337         enum object_type type;
3338         char hdr[32];
3339         int hdrlen;
3340         int ret;
3341
3342         if (has_loose_object(sha1))
3343                 return 0;
3344         buf = read_packed_sha1(sha1, &type, &len);
3345         if (!buf)
3346                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3347         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3348         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3349         free(buf);
3350
3351         return ret;
3352 }
3353
3354 int has_pack_index(const unsigned char *sha1)
3355 {
3356         struct stat st;
3357         if (stat(sha1_pack_index_name(sha1), &st))
3358                 return 0;
3359         return 1;
3360 }
3361
3362 int has_sha1_pack(const unsigned char *sha1)
3363 {
3364         struct pack_entry e;
3365         return find_pack_entry(sha1, &e);
3366 }
3367
3368 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3369 {
3370         struct pack_entry e;
3371
3372         if (find_pack_entry(sha1, &e))
3373                 return 1;
3374         if (has_loose_object(sha1))
3375                 return 1;
3376         if (flags & HAS_SHA1_QUICK)
3377                 return 0;
3378         reprepare_packed_git();
3379         return find_pack_entry(sha1, &e);
3380 }
3381
3382 int has_object_file(const struct object_id *oid)
3383 {
3384         return has_sha1_file(oid->hash);
3385 }
3386
3387 int has_object_file_with_flags(const struct object_id *oid, int flags)
3388 {
3389         return has_sha1_file_with_flags(oid->hash, flags);
3390 }
3391
3392 static void check_tree(const void *buf, size_t size)
3393 {
3394         struct tree_desc desc;
3395         struct name_entry entry;
3396
3397         init_tree_desc(&desc, buf, size);
3398         while (tree_entry(&desc, &entry))
3399                 /* do nothing
3400                  * tree_entry() will die() on malformed entries */
3401                 ;
3402 }
3403
3404 static void check_commit(const void *buf, size_t size)
3405 {
3406         struct commit c;
3407         memset(&c, 0, sizeof(c));
3408         if (parse_commit_buffer(&c, buf, size))
3409                 die("corrupt commit");
3410 }
3411
3412 static void check_tag(const void *buf, size_t size)
3413 {
3414         struct tag t;
3415         memset(&t, 0, sizeof(t));
3416         if (parse_tag_buffer(&t, buf, size))
3417                 die("corrupt tag");
3418 }
3419
3420 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3421                      enum object_type type,
3422                      const char *path, unsigned flags)
3423 {
3424         int ret, re_allocated = 0;
3425         int write_object = flags & HASH_WRITE_OBJECT;
3426
3427         if (!type)
3428                 type = OBJ_BLOB;
3429
3430         /*
3431          * Convert blobs to git internal format
3432          */
3433         if ((type == OBJ_BLOB) && path) {
3434                 struct strbuf nbuf = STRBUF_INIT;
3435                 if (convert_to_git(path, buf, size, &nbuf,
3436                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3437                         buf = strbuf_detach(&nbuf, &size);
3438                         re_allocated = 1;
3439                 }
3440         }
3441         if (flags & HASH_FORMAT_CHECK) {
3442                 if (type == OBJ_TREE)
3443                         check_tree(buf, size);
3444                 if (type == OBJ_COMMIT)
3445                         check_commit(buf, size);
3446                 if (type == OBJ_TAG)
3447                         check_tag(buf, size);
3448         }
3449
3450         if (write_object)
3451                 ret = write_sha1_file(buf, size, typename(type), sha1);
3452         else
3453                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3454         if (re_allocated)
3455                 free(buf);
3456         return ret;
3457 }
3458
3459 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3460                                      const char *path, unsigned flags)
3461 {
3462         int ret;
3463         const int write_object = flags & HASH_WRITE_OBJECT;
3464         struct strbuf sbuf = STRBUF_INIT;
3465
3466         assert(path);
3467         assert(would_convert_to_git_filter_fd(path));
3468
3469         convert_to_git_filter_fd(path, fd, &sbuf,
3470                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3471
3472         if (write_object)
3473                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3474                                       sha1);
3475         else
3476                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3477                                      sha1);
3478         strbuf_release(&sbuf);
3479         return ret;
3480 }
3481
3482 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3483                       const char *path, unsigned flags)
3484 {
3485         struct strbuf sbuf = STRBUF_INIT;
3486         int ret;
3487
3488         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3489                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3490         else
3491                 ret = -1;
3492         strbuf_release(&sbuf);
3493         return ret;
3494 }
3495
3496 #define SMALL_FILE_SIZE (32*1024)
3497
3498 static int index_core(unsigned char *sha1, int fd, size_t size,
3499                       enum object_type type, const char *path,
3500                       unsigned flags)
3501 {
3502         int ret;
3503
3504         if (!size) {
3505                 ret = index_mem(sha1, "", size, type, path, flags);
3506         } else if (size <= SMALL_FILE_SIZE) {
3507                 char *buf = xmalloc(size);
3508                 if (size == read_in_full(fd, buf, size))
3509                         ret = index_mem(sha1, buf, size, type, path, flags);
3510                 else
3511                         ret = error_errno("short read");
3512                 free(buf);
3513         } else {
3514                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3515                 ret = index_mem(sha1, buf, size, type, path, flags);
3516                 munmap(buf, size);
3517         }
3518         return ret;
3519 }
3520
3521 /*
3522  * This creates one packfile per large blob unless bulk-checkin
3523  * machinery is "plugged".
3524  *
3525  * This also bypasses the usual "convert-to-git" dance, and that is on
3526  * purpose. We could write a streaming version of the converting
3527  * functions and insert that before feeding the data to fast-import
3528  * (or equivalent in-core API described above). However, that is
3529  * somewhat complicated, as we do not know the size of the filter
3530  * result, which we need to know beforehand when writing a git object.
3531  * Since the primary motivation for trying to stream from the working
3532  * tree file and to avoid mmaping it in core is to deal with large
3533  * binary blobs, they generally do not want to get any conversion, and
3534  * callers should avoid this code path when filters are requested.
3535  */
3536 static int index_stream(unsigned char *sha1, int fd, size_t size,
3537                         enum object_type type, const char *path,
3538                         unsigned flags)
3539 {
3540         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3541 }
3542
3543 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3544              enum object_type type, const char *path, unsigned flags)
3545 {
3546         int ret;
3547
3548         /*
3549          * Call xsize_t() only when needed to avoid potentially unnecessary
3550          * die() for large files.
3551          */
3552         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3553                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3554         else if (!S_ISREG(st->st_mode))
3555                 ret = index_pipe(sha1, fd, type, path, flags);
3556         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3557                  (path && would_convert_to_git(path)))
3558                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3559                                  flags);
3560         else
3561                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3562                                    flags);
3563         close(fd);
3564         return ret;
3565 }
3566
3567 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3568 {
3569         int fd;
3570         struct strbuf sb = STRBUF_INIT;
3571
3572         switch (st->st_mode & S_IFMT) {
3573         case S_IFREG:
3574                 fd = open(path, O_RDONLY);
3575                 if (fd < 0)
3576                         return error_errno("open(\"%s\")", path);
3577                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3578                         return error("%s: failed to insert into database",
3579                                      path);
3580                 break;
3581         case S_IFLNK:
3582                 if (strbuf_readlink(&sb, path, st->st_size))
3583                         return error_errno("readlink(\"%s\")", path);
3584                 if (!(flags & HASH_WRITE_OBJECT))
3585                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3586                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3587                         return error("%s: failed to insert into database",
3588                                      path);
3589                 strbuf_release(&sb);
3590                 break;
3591         case S_IFDIR:
3592                 return resolve_gitlink_ref(path, "HEAD", sha1);
3593         default:
3594                 return error("%s: unsupported file type", path);
3595         }
3596         return 0;
3597 }
3598
3599 int read_pack_header(int fd, struct pack_header *header)
3600 {
3601         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3602                 /* "eof before pack header was fully read" */
3603                 return PH_ERROR_EOF;
3604
3605         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3606                 /* "protocol error (pack signature mismatch detected)" */
3607                 return PH_ERROR_PACK_SIGNATURE;
3608         if (!pack_version_ok(header->hdr_version))
3609                 /* "protocol error (pack version unsupported)" */
3610                 return PH_ERROR_PROTOCOL;
3611         return 0;
3612 }
3613
3614 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3615 {
3616         enum object_type type = sha1_object_info(sha1, NULL);
3617         if (type < 0)
3618                 die("%s is not a valid object", sha1_to_hex(sha1));
3619         if (type != expect)
3620                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3621                     typename(expect));
3622 }
3623
3624 static int for_each_file_in_obj_subdir(int subdir_nr,
3625                                        struct strbuf *path,
3626                                        each_loose_object_fn obj_cb,
3627                                        each_loose_cruft_fn cruft_cb,
3628                                        each_loose_subdir_fn subdir_cb,
3629                                        void *data)
3630 {
3631         size_t baselen = path->len;
3632         DIR *dir = opendir(path->buf);
3633         struct dirent *de;
3634         int r = 0;
3635
3636         if (!dir) {
3637                 if (errno == ENOENT)
3638                         return 0;
3639                 return error_errno("unable to open %s", path->buf);
3640         }
3641
3642         while ((de = readdir(dir))) {
3643                 if (is_dot_or_dotdot(de->d_name))
3644                         continue;
3645
3646                 strbuf_setlen(path, baselen);
3647                 strbuf_addf(path, "/%s", de->d_name);
3648
3649                 if (strlen(de->d_name) == 38)  {
3650                         char hex[41];
3651                         unsigned char sha1[20];
3652
3653                         snprintf(hex, sizeof(hex), "%02x%s",
3654                                  subdir_nr, de->d_name);
3655                         if (!get_sha1_hex(hex, sha1)) {
3656                                 if (obj_cb) {
3657                                         r = obj_cb(sha1, path->buf, data);
3658                                         if (r)
3659                                                 break;
3660                                 }
3661                                 continue;
3662                         }
3663                 }
3664
3665                 if (cruft_cb) {
3666                         r = cruft_cb(de->d_name, path->buf, data);
3667                         if (r)
3668                                 break;
3669                 }
3670         }
3671         closedir(dir);
3672
3673         strbuf_setlen(path, baselen);
3674         if (!r && subdir_cb)
3675                 r = subdir_cb(subdir_nr, path->buf, data);
3676
3677         return r;
3678 }
3679
3680 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3681                             each_loose_object_fn obj_cb,
3682                             each_loose_cruft_fn cruft_cb,
3683                             each_loose_subdir_fn subdir_cb,
3684                             void *data)
3685 {
3686         size_t baselen = path->len;
3687         int r = 0;
3688         int i;
3689
3690         for (i = 0; i < 256; i++) {
3691                 strbuf_addf(path, "/%02x", i);
3692                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3693                                                 subdir_cb, data);
3694                 strbuf_setlen(path, baselen);
3695                 if (r)
3696                         break;
3697         }
3698
3699         return r;
3700 }
3701
3702 int for_each_loose_file_in_objdir(const char *path,
3703                                   each_loose_object_fn obj_cb,
3704                                   each_loose_cruft_fn cruft_cb,
3705                                   each_loose_subdir_fn subdir_cb,
3706                                   void *data)
3707 {
3708         struct strbuf buf = STRBUF_INIT;
3709         int r;
3710
3711         strbuf_addstr(&buf, path);
3712         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3713                                               subdir_cb, data);
3714         strbuf_release(&buf);
3715
3716         return r;
3717 }
3718
3719 struct loose_alt_odb_data {
3720         each_loose_object_fn *cb;
3721         void *data;
3722 };
3723
3724 static int loose_from_alt_odb(struct alternate_object_database *alt,
3725                               void *vdata)
3726 {
3727         struct loose_alt_odb_data *data = vdata;
3728         struct strbuf buf = STRBUF_INIT;
3729         int r;
3730
3731         strbuf_addstr(&buf, alt->path);
3732         r = for_each_loose_file_in_objdir_buf(&buf,
3733                                               data->cb, NULL, NULL,
3734                                               data->data);
3735         strbuf_release(&buf);
3736         return r;
3737 }
3738
3739 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3740 {
3741         struct loose_alt_odb_data alt;
3742         int r;
3743
3744         r = for_each_loose_file_in_objdir(get_object_directory(),
3745                                           cb, NULL, NULL, data);
3746         if (r)
3747                 return r;
3748
3749         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3750                 return 0;
3751
3752         alt.cb = cb;
3753         alt.data = data;
3754         return foreach_alt_odb(loose_from_alt_odb, &alt);
3755 }
3756
3757 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3758 {
3759         uint32_t i;
3760         int r = 0;
3761
3762         for (i = 0; i < p->num_objects; i++) {
3763                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3764
3765                 if (!sha1)
3766                         return error("unable to get sha1 of object %u in %s",
3767                                      i, p->pack_name);
3768
3769                 r = cb(sha1, p, i, data);
3770                 if (r)
3771                         break;
3772         }
3773         return r;
3774 }
3775
3776 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3777 {
3778         struct packed_git *p;
3779         int r = 0;
3780         int pack_errors = 0;
3781
3782         prepare_packed_git();
3783         for (p = packed_git; p; p = p->next) {
3784                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3785                         continue;
3786                 if (open_pack_index(p)) {
3787                         pack_errors = 1;
3788                         continue;
3789                 }
3790                 r = for_each_object_in_pack(p, cb, data);
3791                 if (r)
3792                         break;
3793         }
3794         return r ? r : pack_errors;
3795 }