Merge branch 'bw/object-id'
[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                                 errno = ENOTDIR;
134                                 ret = SCLD_EXISTS;
135                         }
136                 } else if (mkdir(path, 0777)) {
137                         if (errno == EEXIST &&
138                             !stat(path, &st) && S_ISDIR(st.st_mode))
139                                 ; /* somebody created it since we checked */
140                         else if (errno == ENOENT)
141                                 /*
142                                  * Either mkdir() failed because
143                                  * somebody just pruned the containing
144                                  * directory, or stat() failed because
145                                  * the file that was in our way was
146                                  * just removed.  Either way, inform
147                                  * the caller that it might be worth
148                                  * trying again:
149                                  */
150                                 ret = SCLD_VANISHED;
151                         else
152                                 ret = SCLD_FAILED;
153                 } else if (adjust_shared_perm(path)) {
154                         ret = SCLD_PERMS;
155                 }
156                 *slash = slash_character;
157         }
158         return ret;
159 }
160
161 enum scld_error safe_create_leading_directories_const(const char *path)
162 {
163         int save_errno;
164         /* path points to cache entries, so xstrdup before messing with it */
165         char *buf = xstrdup(path);
166         enum scld_error result = safe_create_leading_directories(buf);
167
168         save_errno = errno;
169         free(buf);
170         errno = save_errno;
171         return result;
172 }
173
174 int raceproof_create_file(const char *path, create_file_fn fn, void *cb)
175 {
176         /*
177          * The number of times we will try to remove empty directories
178          * in the way of path. This is only 1 because if another
179          * process is racily creating directories that conflict with
180          * us, we don't want to fight against them.
181          */
182         int remove_directories_remaining = 1;
183
184         /*
185          * The number of times that we will try to create the
186          * directories containing path. We are willing to attempt this
187          * more than once, because another process could be trying to
188          * clean up empty directories at the same time as we are
189          * trying to create them.
190          */
191         int create_directories_remaining = 3;
192
193         /* A scratch copy of path, filled lazily if we need it: */
194         struct strbuf path_copy = STRBUF_INIT;
195
196         int ret, save_errno;
197
198         /* Sanity check: */
199         assert(*path);
200
201 retry_fn:
202         ret = fn(path, cb);
203         save_errno = errno;
204         if (!ret)
205                 goto out;
206
207         if (errno == EISDIR && remove_directories_remaining-- > 0) {
208                 /*
209                  * A directory is in the way. Maybe it is empty; try
210                  * to remove it:
211                  */
212                 if (!path_copy.len)
213                         strbuf_addstr(&path_copy, path);
214
215                 if (!remove_dir_recursively(&path_copy, REMOVE_DIR_EMPTY_ONLY))
216                         goto retry_fn;
217         } else if (errno == ENOENT && create_directories_remaining-- > 0) {
218                 /*
219                  * Maybe the containing directory didn't exist, or
220                  * maybe it was just deleted by a process that is
221                  * racing with us to clean up empty directories. Try
222                  * to create it:
223                  */
224                 enum scld_error scld_result;
225
226                 if (!path_copy.len)
227                         strbuf_addstr(&path_copy, path);
228
229                 do {
230                         scld_result = safe_create_leading_directories(path_copy.buf);
231                         if (scld_result == SCLD_OK)
232                                 goto retry_fn;
233                 } while (scld_result == SCLD_VANISHED && create_directories_remaining-- > 0);
234         }
235
236 out:
237         strbuf_release(&path_copy);
238         errno = save_errno;
239         return ret;
240 }
241
242 static void fill_sha1_path(struct strbuf *buf, const unsigned char *sha1)
243 {
244         int i;
245         for (i = 0; i < 20; i++) {
246                 static char hex[] = "0123456789abcdef";
247                 unsigned int val = sha1[i];
248                 strbuf_addch(buf, hex[val >> 4]);
249                 strbuf_addch(buf, hex[val & 0xf]);
250                 if (!i)
251                         strbuf_addch(buf, '/');
252         }
253 }
254
255 const char *sha1_file_name(const unsigned char *sha1)
256 {
257         static struct strbuf buf = STRBUF_INIT;
258
259         strbuf_reset(&buf);
260         strbuf_addf(&buf, "%s/", get_object_directory());
261
262         fill_sha1_path(&buf, sha1);
263         return buf.buf;
264 }
265
266 struct strbuf *alt_scratch_buf(struct alternate_object_database *alt)
267 {
268         strbuf_setlen(&alt->scratch, alt->base_len);
269         return &alt->scratch;
270 }
271
272 static const char *alt_sha1_path(struct alternate_object_database *alt,
273                                  const unsigned char *sha1)
274 {
275         struct strbuf *buf = alt_scratch_buf(alt);
276         fill_sha1_path(buf, sha1);
277         return buf->buf;
278 }
279
280  char *odb_pack_name(struct strbuf *buf,
281                      const unsigned char *sha1,
282                      const char *ext)
283 {
284         strbuf_reset(buf);
285         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
286                     sha1_to_hex(sha1), ext);
287         return buf->buf;
288 }
289
290 char *sha1_pack_name(const unsigned char *sha1)
291 {
292         static struct strbuf buf = STRBUF_INIT;
293         return odb_pack_name(&buf, sha1, "pack");
294 }
295
296 char *sha1_pack_index_name(const unsigned char *sha1)
297 {
298         static struct strbuf buf = STRBUF_INIT;
299         return odb_pack_name(&buf, sha1, "idx");
300 }
301
302 struct alternate_object_database *alt_odb_list;
303 static struct alternate_object_database **alt_odb_tail;
304
305 /*
306  * Return non-zero iff the path is usable as an alternate object database.
307  */
308 static int alt_odb_usable(struct strbuf *path, const char *normalized_objdir)
309 {
310         struct alternate_object_database *alt;
311
312         /* Detect cases where alternate disappeared */
313         if (!is_directory(path->buf)) {
314                 error("object directory %s does not exist; "
315                       "check .git/objects/info/alternates.",
316                       path->buf);
317                 return 0;
318         }
319
320         /*
321          * Prevent the common mistake of listing the same
322          * thing twice, or object directory itself.
323          */
324         for (alt = alt_odb_list; alt; alt = alt->next) {
325                 if (!fspathcmp(path->buf, alt->path))
326                         return 0;
327         }
328         if (!fspathcmp(path->buf, normalized_objdir))
329                 return 0;
330
331         return 1;
332 }
333
334 /*
335  * Prepare alternate object database registry.
336  *
337  * The variable alt_odb_list points at the list of struct
338  * alternate_object_database.  The elements on this list come from
339  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
340  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
341  * whose contents is similar to that environment variable but can be
342  * LF separated.  Its base points at a statically allocated buffer that
343  * contains "/the/directory/corresponding/to/.git/objects/...", while
344  * its name points just after the slash at the end of ".git/objects/"
345  * in the example above, and has enough space to hold 40-byte hex
346  * SHA1, an extra slash for the first level indirection, and the
347  * terminating NUL.
348  */
349 static int link_alt_odb_entry(const char *entry, const char *relative_base,
350         int depth, const char *normalized_objdir)
351 {
352         struct alternate_object_database *ent;
353         struct strbuf pathbuf = STRBUF_INIT;
354
355         if (!is_absolute_path(entry) && relative_base) {
356                 strbuf_realpath(&pathbuf, relative_base, 1);
357                 strbuf_addch(&pathbuf, '/');
358         }
359         strbuf_addstr(&pathbuf, entry);
360
361         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
362                 error("unable to normalize alternate object path: %s",
363                       pathbuf.buf);
364                 strbuf_release(&pathbuf);
365                 return -1;
366         }
367
368         /*
369          * The trailing slash after the directory name is given by
370          * this function at the end. Remove duplicates.
371          */
372         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
373                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
374
375         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
376                 strbuf_release(&pathbuf);
377                 return -1;
378         }
379
380         ent = alloc_alt_odb(pathbuf.buf);
381
382         /* add the alternate entry */
383         *alt_odb_tail = ent;
384         alt_odb_tail = &(ent->next);
385         ent->next = NULL;
386
387         /* recursively add alternates */
388         read_info_alternates(pathbuf.buf, depth + 1);
389
390         strbuf_release(&pathbuf);
391         return 0;
392 }
393
394 static const char *parse_alt_odb_entry(const char *string,
395                                        int sep,
396                                        struct strbuf *out)
397 {
398         const char *end;
399
400         strbuf_reset(out);
401
402         if (*string == '#') {
403                 /* comment; consume up to next separator */
404                 end = strchrnul(string, sep);
405         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
406                 /*
407                  * quoted path; unquote_c_style has copied the
408                  * data for us and set "end". Broken quoting (e.g.,
409                  * an entry that doesn't end with a quote) falls
410                  * back to the unquoted case below.
411                  */
412         } else {
413                 /* normal, unquoted path */
414                 end = strchrnul(string, sep);
415                 strbuf_add(out, string, end - string);
416         }
417
418         if (*end)
419                 end++;
420         return end;
421 }
422
423 static void link_alt_odb_entries(const char *alt, int len, int sep,
424                                  const char *relative_base, int depth)
425 {
426         struct strbuf objdirbuf = STRBUF_INIT;
427         struct strbuf entry = STRBUF_INIT;
428
429         if (depth > 5) {
430                 error("%s: ignoring alternate object stores, nesting too deep.",
431                                 relative_base);
432                 return;
433         }
434
435         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
436         if (strbuf_normalize_path(&objdirbuf) < 0)
437                 die("unable to normalize object directory: %s",
438                     objdirbuf.buf);
439
440         while (*alt) {
441                 alt = parse_alt_odb_entry(alt, sep, &entry);
442                 if (!entry.len)
443                         continue;
444                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
445         }
446         strbuf_release(&entry);
447         strbuf_release(&objdirbuf);
448 }
449
450 void read_info_alternates(const char * relative_base, int depth)
451 {
452         char *map;
453         size_t mapsz;
454         struct stat st;
455         char *path;
456         int fd;
457
458         path = xstrfmt("%s/info/alternates", relative_base);
459         fd = git_open(path);
460         free(path);
461         if (fd < 0)
462                 return;
463         if (fstat(fd, &st) || (st.st_size == 0)) {
464                 close(fd);
465                 return;
466         }
467         mapsz = xsize_t(st.st_size);
468         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
469         close(fd);
470
471         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
472
473         munmap(map, mapsz);
474 }
475
476 struct alternate_object_database *alloc_alt_odb(const char *dir)
477 {
478         struct alternate_object_database *ent;
479
480         FLEX_ALLOC_STR(ent, path, dir);
481         strbuf_init(&ent->scratch, 0);
482         strbuf_addf(&ent->scratch, "%s/", dir);
483         ent->base_len = ent->scratch.len;
484
485         return ent;
486 }
487
488 void add_to_alternates_file(const char *reference)
489 {
490         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
491         char *alts = git_pathdup("objects/info/alternates");
492         FILE *in, *out;
493
494         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
495         out = fdopen_lock_file(lock, "w");
496         if (!out)
497                 die_errno("unable to fdopen alternates lockfile");
498
499         in = fopen(alts, "r");
500         if (in) {
501                 struct strbuf line = STRBUF_INIT;
502                 int found = 0;
503
504                 while (strbuf_getline(&line, in) != EOF) {
505                         if (!strcmp(reference, line.buf)) {
506                                 found = 1;
507                                 break;
508                         }
509                         fprintf_or_die(out, "%s\n", line.buf);
510                 }
511
512                 strbuf_release(&line);
513                 fclose(in);
514
515                 if (found) {
516                         rollback_lock_file(lock);
517                         lock = NULL;
518                 }
519         }
520         else if (errno != ENOENT)
521                 die_errno("unable to read alternates file");
522
523         if (lock) {
524                 fprintf_or_die(out, "%s\n", reference);
525                 if (commit_lock_file(lock))
526                         die_errno("unable to move new alternates file into place");
527                 if (alt_odb_tail)
528                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
529         }
530         free(alts);
531 }
532
533 void add_to_alternates_memory(const char *reference)
534 {
535         /*
536          * Make sure alternates are initialized, or else our entry may be
537          * overwritten when they are.
538          */
539         prepare_alt_odb();
540
541         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
542 }
543
544 /*
545  * Compute the exact path an alternate is at and returns it. In case of
546  * error NULL is returned and the human readable error is added to `err`
547  * `path` may be relative and should point to $GITDIR.
548  * `err` must not be null.
549  */
550 char *compute_alternate_path(const char *path, struct strbuf *err)
551 {
552         char *ref_git = NULL;
553         const char *repo, *ref_git_s;
554         int seen_error = 0;
555
556         ref_git_s = real_path_if_valid(path);
557         if (!ref_git_s) {
558                 seen_error = 1;
559                 strbuf_addf(err, _("path '%s' does not exist"), path);
560                 goto out;
561         } else
562                 /*
563                  * Beware: read_gitfile(), real_path() and mkpath()
564                  * return static buffer
565                  */
566                 ref_git = xstrdup(ref_git_s);
567
568         repo = read_gitfile(ref_git);
569         if (!repo)
570                 repo = read_gitfile(mkpath("%s/.git", ref_git));
571         if (repo) {
572                 free(ref_git);
573                 ref_git = xstrdup(repo);
574         }
575
576         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
577                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
578                 free(ref_git);
579                 ref_git = ref_git_git;
580         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
581                 struct strbuf sb = STRBUF_INIT;
582                 seen_error = 1;
583                 if (get_common_dir(&sb, ref_git)) {
584                         strbuf_addf(err,
585                                     _("reference repository '%s' as a linked "
586                                       "checkout is not supported yet."),
587                                     path);
588                         goto out;
589                 }
590
591                 strbuf_addf(err, _("reference repository '%s' is not a "
592                                         "local repository."), path);
593                 goto out;
594         }
595
596         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
597                 strbuf_addf(err, _("reference repository '%s' is shallow"),
598                             path);
599                 seen_error = 1;
600                 goto out;
601         }
602
603         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
604                 strbuf_addf(err,
605                             _("reference repository '%s' is grafted"),
606                             path);
607                 seen_error = 1;
608                 goto out;
609         }
610
611 out:
612         if (seen_error) {
613                 free(ref_git);
614                 ref_git = NULL;
615         }
616
617         return ref_git;
618 }
619
620 int foreach_alt_odb(alt_odb_fn fn, void *cb)
621 {
622         struct alternate_object_database *ent;
623         int r = 0;
624
625         prepare_alt_odb();
626         for (ent = alt_odb_list; ent; ent = ent->next) {
627                 r = fn(ent, cb);
628                 if (r)
629                         break;
630         }
631         return r;
632 }
633
634 void prepare_alt_odb(void)
635 {
636         const char *alt;
637
638         if (alt_odb_tail)
639                 return;
640
641         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
642         if (!alt) alt = "";
643
644         alt_odb_tail = &alt_odb_list;
645         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
646
647         read_info_alternates(get_object_directory(), 0);
648 }
649
650 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
651 static int freshen_file(const char *fn)
652 {
653         struct utimbuf t;
654         t.actime = t.modtime = time(NULL);
655         return !utime(fn, &t);
656 }
657
658 /*
659  * All of the check_and_freshen functions return 1 if the file exists and was
660  * freshened (if freshening was requested), 0 otherwise. If they return
661  * 0, you should not assume that it is safe to skip a write of the object (it
662  * either does not exist on disk, or has a stale mtime and may be subject to
663  * pruning).
664  */
665 int check_and_freshen_file(const char *fn, int freshen)
666 {
667         if (access(fn, F_OK))
668                 return 0;
669         if (freshen && !freshen_file(fn))
670                 return 0;
671         return 1;
672 }
673
674 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
675 {
676         return check_and_freshen_file(sha1_file_name(sha1), freshen);
677 }
678
679 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
680 {
681         struct alternate_object_database *alt;
682         prepare_alt_odb();
683         for (alt = alt_odb_list; alt; alt = alt->next) {
684                 const char *path = alt_sha1_path(alt, sha1);
685                 if (check_and_freshen_file(path, freshen))
686                         return 1;
687         }
688         return 0;
689 }
690
691 static int check_and_freshen(const unsigned char *sha1, int freshen)
692 {
693         return check_and_freshen_local(sha1, freshen) ||
694                check_and_freshen_nonlocal(sha1, freshen);
695 }
696
697 int has_loose_object_nonlocal(const unsigned char *sha1)
698 {
699         return check_and_freshen_nonlocal(sha1, 0);
700 }
701
702 static int has_loose_object(const unsigned char *sha1)
703 {
704         return check_and_freshen(sha1, 0);
705 }
706
707 static unsigned int pack_used_ctr;
708 static unsigned int pack_mmap_calls;
709 static unsigned int peak_pack_open_windows;
710 static unsigned int pack_open_windows;
711 static unsigned int pack_open_fds;
712 static unsigned int pack_max_fds;
713 static size_t peak_pack_mapped;
714 static size_t pack_mapped;
715 struct packed_git *packed_git;
716
717 static struct mru packed_git_mru_storage;
718 struct mru *packed_git_mru = &packed_git_mru_storage;
719
720 void pack_report(void)
721 {
722         fprintf(stderr,
723                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
724                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
725                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
726                 sz_fmt(getpagesize()),
727                 sz_fmt(packed_git_window_size),
728                 sz_fmt(packed_git_limit));
729         fprintf(stderr,
730                 "pack_report: pack_used_ctr            = %10u\n"
731                 "pack_report: pack_mmap_calls          = %10u\n"
732                 "pack_report: pack_open_windows        = %10u / %10u\n"
733                 "pack_report: pack_mapped              = "
734                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
735                 pack_used_ctr,
736                 pack_mmap_calls,
737                 pack_open_windows, peak_pack_open_windows,
738                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
739 }
740
741 /*
742  * Open and mmap the index file at path, perform a couple of
743  * consistency checks, then record its information to p.  Return 0 on
744  * success.
745  */
746 static int check_packed_git_idx(const char *path, struct packed_git *p)
747 {
748         void *idx_map;
749         struct pack_idx_header *hdr;
750         size_t idx_size;
751         uint32_t version, nr, i, *index;
752         int fd = git_open(path);
753         struct stat st;
754
755         if (fd < 0)
756                 return -1;
757         if (fstat(fd, &st)) {
758                 close(fd);
759                 return -1;
760         }
761         idx_size = xsize_t(st.st_size);
762         if (idx_size < 4 * 256 + 20 + 20) {
763                 close(fd);
764                 return error("index file %s is too small", path);
765         }
766         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
767         close(fd);
768
769         hdr = idx_map;
770         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
771                 version = ntohl(hdr->idx_version);
772                 if (version < 2 || version > 2) {
773                         munmap(idx_map, idx_size);
774                         return error("index file %s is version %"PRIu32
775                                      " and is not supported by this binary"
776                                      " (try upgrading GIT to a newer version)",
777                                      path, version);
778                 }
779         } else
780                 version = 1;
781
782         nr = 0;
783         index = idx_map;
784         if (version > 1)
785                 index += 2;  /* skip index header */
786         for (i = 0; i < 256; i++) {
787                 uint32_t n = ntohl(index[i]);
788                 if (n < nr) {
789                         munmap(idx_map, idx_size);
790                         return error("non-monotonic index %s", path);
791                 }
792                 nr = n;
793         }
794
795         if (version == 1) {
796                 /*
797                  * Total size:
798                  *  - 256 index entries 4 bytes each
799                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
800                  *  - 20-byte SHA1 of the packfile
801                  *  - 20-byte SHA1 file checksum
802                  */
803                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
804                         munmap(idx_map, idx_size);
805                         return error("wrong index v1 file size in %s", path);
806                 }
807         } else if (version == 2) {
808                 /*
809                  * Minimum size:
810                  *  - 8 bytes of header
811                  *  - 256 index entries 4 bytes each
812                  *  - 20-byte sha1 entry * nr
813                  *  - 4-byte crc entry * nr
814                  *  - 4-byte offset entry * nr
815                  *  - 20-byte SHA1 of the packfile
816                  *  - 20-byte SHA1 file checksum
817                  * And after the 4-byte offset table might be a
818                  * variable sized table containing 8-byte entries
819                  * for offsets larger than 2^31.
820                  */
821                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
822                 unsigned long max_size = min_size;
823                 if (nr)
824                         max_size += (nr - 1)*8;
825                 if (idx_size < min_size || idx_size > max_size) {
826                         munmap(idx_map, idx_size);
827                         return error("wrong index v2 file size in %s", path);
828                 }
829                 if (idx_size != min_size &&
830                     /*
831                      * make sure we can deal with large pack offsets.
832                      * 31-bit signed offset won't be enough, neither
833                      * 32-bit unsigned one will be.
834                      */
835                     (sizeof(off_t) <= 4)) {
836                         munmap(idx_map, idx_size);
837                         return error("pack too large for current definition of off_t in %s", path);
838                 }
839         }
840
841         p->index_version = version;
842         p->index_data = idx_map;
843         p->index_size = idx_size;
844         p->num_objects = nr;
845         return 0;
846 }
847
848 int open_pack_index(struct packed_git *p)
849 {
850         char *idx_name;
851         size_t len;
852         int ret;
853
854         if (p->index_data)
855                 return 0;
856
857         if (!strip_suffix(p->pack_name, ".pack", &len))
858                 die("BUG: pack_name does not end in .pack");
859         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
860         ret = check_packed_git_idx(idx_name, p);
861         free(idx_name);
862         return ret;
863 }
864
865 static void scan_windows(struct packed_git *p,
866         struct packed_git **lru_p,
867         struct pack_window **lru_w,
868         struct pack_window **lru_l)
869 {
870         struct pack_window *w, *w_l;
871
872         for (w_l = NULL, w = p->windows; w; w = w->next) {
873                 if (!w->inuse_cnt) {
874                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
875                                 *lru_p = p;
876                                 *lru_w = w;
877                                 *lru_l = w_l;
878                         }
879                 }
880                 w_l = w;
881         }
882 }
883
884 static int unuse_one_window(struct packed_git *current)
885 {
886         struct packed_git *p, *lru_p = NULL;
887         struct pack_window *lru_w = NULL, *lru_l = NULL;
888
889         if (current)
890                 scan_windows(current, &lru_p, &lru_w, &lru_l);
891         for (p = packed_git; p; p = p->next)
892                 scan_windows(p, &lru_p, &lru_w, &lru_l);
893         if (lru_p) {
894                 munmap(lru_w->base, lru_w->len);
895                 pack_mapped -= lru_w->len;
896                 if (lru_l)
897                         lru_l->next = lru_w->next;
898                 else
899                         lru_p->windows = lru_w->next;
900                 free(lru_w);
901                 pack_open_windows--;
902                 return 1;
903         }
904         return 0;
905 }
906
907 void release_pack_memory(size_t need)
908 {
909         size_t cur = pack_mapped;
910         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
911                 ; /* nothing */
912 }
913
914 static void mmap_limit_check(size_t length)
915 {
916         static size_t limit = 0;
917         if (!limit) {
918                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
919                 if (!limit)
920                         limit = SIZE_MAX;
921         }
922         if (length > limit)
923                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
924                     (uintmax_t)length, (uintmax_t)limit);
925 }
926
927 void *xmmap_gently(void *start, size_t length,
928                   int prot, int flags, int fd, off_t offset)
929 {
930         void *ret;
931
932         mmap_limit_check(length);
933         ret = mmap(start, length, prot, flags, fd, offset);
934         if (ret == MAP_FAILED) {
935                 if (!length)
936                         return NULL;
937                 release_pack_memory(length);
938                 ret = mmap(start, length, prot, flags, fd, offset);
939         }
940         return ret;
941 }
942
943 void *xmmap(void *start, size_t length,
944         int prot, int flags, int fd, off_t offset)
945 {
946         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
947         if (ret == MAP_FAILED)
948                 die_errno("mmap failed");
949         return ret;
950 }
951
952 void close_pack_windows(struct packed_git *p)
953 {
954         while (p->windows) {
955                 struct pack_window *w = p->windows;
956
957                 if (w->inuse_cnt)
958                         die("pack '%s' still has open windows to it",
959                             p->pack_name);
960                 munmap(w->base, w->len);
961                 pack_mapped -= w->len;
962                 pack_open_windows--;
963                 p->windows = w->next;
964                 free(w);
965         }
966 }
967
968 static int close_pack_fd(struct packed_git *p)
969 {
970         if (p->pack_fd < 0)
971                 return 0;
972
973         close(p->pack_fd);
974         pack_open_fds--;
975         p->pack_fd = -1;
976
977         return 1;
978 }
979
980 static void close_pack(struct packed_git *p)
981 {
982         close_pack_windows(p);
983         close_pack_fd(p);
984         close_pack_index(p);
985 }
986
987 void close_all_packs(void)
988 {
989         struct packed_git *p;
990
991         for (p = packed_git; p; p = p->next)
992                 if (p->do_not_close)
993                         die("BUG: want to close pack marked 'do-not-close'");
994                 else
995                         close_pack(p);
996 }
997
998
999 /*
1000  * The LRU pack is the one with the oldest MRU window, preferring packs
1001  * with no used windows, or the oldest mtime if it has no windows allocated.
1002  */
1003 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
1004 {
1005         struct pack_window *w, *this_mru_w;
1006         int has_windows_inuse = 0;
1007
1008         /*
1009          * Reject this pack if it has windows and the previously selected
1010          * one does not.  If this pack does not have windows, reject
1011          * it if the pack file is newer than the previously selected one.
1012          */
1013         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
1014                 return;
1015
1016         for (w = this_mru_w = p->windows; w; w = w->next) {
1017                 /*
1018                  * Reject this pack if any of its windows are in use,
1019                  * but the previously selected pack did not have any
1020                  * inuse windows.  Otherwise, record that this pack
1021                  * has windows in use.
1022                  */
1023                 if (w->inuse_cnt) {
1024                         if (*accept_windows_inuse)
1025                                 has_windows_inuse = 1;
1026                         else
1027                                 return;
1028                 }
1029
1030                 if (w->last_used > this_mru_w->last_used)
1031                         this_mru_w = w;
1032
1033                 /*
1034                  * Reject this pack if it has windows that have been
1035                  * used more recently than the previously selected pack.
1036                  * If the previously selected pack had windows inuse and
1037                  * we have not encountered a window in this pack that is
1038                  * inuse, skip this check since we prefer a pack with no
1039                  * inuse windows to one that has inuse windows.
1040                  */
1041                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
1042                     this_mru_w->last_used > (*mru_w)->last_used)
1043                         return;
1044         }
1045
1046         /*
1047          * Select this pack.
1048          */
1049         *mru_w = this_mru_w;
1050         *lru_p = p;
1051         *accept_windows_inuse = has_windows_inuse;
1052 }
1053
1054 static int close_one_pack(void)
1055 {
1056         struct packed_git *p, *lru_p = NULL;
1057         struct pack_window *mru_w = NULL;
1058         int accept_windows_inuse = 1;
1059
1060         for (p = packed_git; p; p = p->next) {
1061                 if (p->pack_fd == -1)
1062                         continue;
1063                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
1064         }
1065
1066         if (lru_p)
1067                 return close_pack_fd(lru_p);
1068
1069         return 0;
1070 }
1071
1072 void unuse_pack(struct pack_window **w_cursor)
1073 {
1074         struct pack_window *w = *w_cursor;
1075         if (w) {
1076                 w->inuse_cnt--;
1077                 *w_cursor = NULL;
1078         }
1079 }
1080
1081 void close_pack_index(struct packed_git *p)
1082 {
1083         if (p->index_data) {
1084                 munmap((void *)p->index_data, p->index_size);
1085                 p->index_data = NULL;
1086         }
1087 }
1088
1089 static unsigned int get_max_fd_limit(void)
1090 {
1091 #ifdef RLIMIT_NOFILE
1092         {
1093                 struct rlimit lim;
1094
1095                 if (!getrlimit(RLIMIT_NOFILE, &lim))
1096                         return lim.rlim_cur;
1097         }
1098 #endif
1099
1100 #ifdef _SC_OPEN_MAX
1101         {
1102                 long open_max = sysconf(_SC_OPEN_MAX);
1103                 if (0 < open_max)
1104                         return open_max;
1105                 /*
1106                  * Otherwise, we got -1 for one of the two
1107                  * reasons:
1108                  *
1109                  * (1) sysconf() did not understand _SC_OPEN_MAX
1110                  *     and signaled an error with -1; or
1111                  * (2) sysconf() said there is no limit.
1112                  *
1113                  * We _could_ clear errno before calling sysconf() to
1114                  * tell these two cases apart and return a huge number
1115                  * in the latter case to let the caller cap it to a
1116                  * value that is not so selfish, but letting the
1117                  * fallback OPEN_MAX codepath take care of these cases
1118                  * is a lot simpler.
1119                  */
1120         }
1121 #endif
1122
1123 #ifdef OPEN_MAX
1124         return OPEN_MAX;
1125 #else
1126         return 1; /* see the caller ;-) */
1127 #endif
1128 }
1129
1130 /*
1131  * Do not call this directly as this leaks p->pack_fd on error return;
1132  * call open_packed_git() instead.
1133  */
1134 static int open_packed_git_1(struct packed_git *p)
1135 {
1136         struct stat st;
1137         struct pack_header hdr;
1138         unsigned char sha1[20];
1139         unsigned char *idx_sha1;
1140         long fd_flag;
1141
1142         if (!p->index_data && open_pack_index(p))
1143                 return error("packfile %s index unavailable", p->pack_name);
1144
1145         if (!pack_max_fds) {
1146                 unsigned int max_fds = get_max_fd_limit();
1147
1148                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1149                 if (25 < max_fds)
1150                         pack_max_fds = max_fds - 25;
1151                 else
1152                         pack_max_fds = 1;
1153         }
1154
1155         while (pack_max_fds <= pack_open_fds && close_one_pack())
1156                 ; /* nothing */
1157
1158         p->pack_fd = git_open(p->pack_name);
1159         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1160                 return -1;
1161         pack_open_fds++;
1162
1163         /* If we created the struct before we had the pack we lack size. */
1164         if (!p->pack_size) {
1165                 if (!S_ISREG(st.st_mode))
1166                         return error("packfile %s not a regular file", p->pack_name);
1167                 p->pack_size = st.st_size;
1168         } else if (p->pack_size != st.st_size)
1169                 return error("packfile %s size changed", p->pack_name);
1170
1171         /* We leave these file descriptors open with sliding mmap;
1172          * there is no point keeping them open across exec(), though.
1173          */
1174         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1175         if (fd_flag < 0)
1176                 return error("cannot determine file descriptor flags");
1177         fd_flag |= FD_CLOEXEC;
1178         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1179                 return error("cannot set FD_CLOEXEC");
1180
1181         /* Verify we recognize this pack file format. */
1182         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1183                 return error("file %s is far too short to be a packfile", p->pack_name);
1184         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1185                 return error("file %s is not a GIT packfile", p->pack_name);
1186         if (!pack_version_ok(hdr.hdr_version))
1187                 return error("packfile %s is version %"PRIu32" and not"
1188                         " supported (try upgrading GIT to a newer version)",
1189                         p->pack_name, ntohl(hdr.hdr_version));
1190
1191         /* Verify the pack matches its index. */
1192         if (p->num_objects != ntohl(hdr.hdr_entries))
1193                 return error("packfile %s claims to have %"PRIu32" objects"
1194                              " while index indicates %"PRIu32" objects",
1195                              p->pack_name, ntohl(hdr.hdr_entries),
1196                              p->num_objects);
1197         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1198                 return error("end of packfile %s is unavailable", p->pack_name);
1199         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1200                 return error("packfile %s signature is unavailable", p->pack_name);
1201         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1202         if (hashcmp(sha1, idx_sha1))
1203                 return error("packfile %s does not match index", p->pack_name);
1204         return 0;
1205 }
1206
1207 static int open_packed_git(struct packed_git *p)
1208 {
1209         if (!open_packed_git_1(p))
1210                 return 0;
1211         close_pack_fd(p);
1212         return -1;
1213 }
1214
1215 static int in_window(struct pack_window *win, off_t offset)
1216 {
1217         /* We must promise at least 20 bytes (one hash) after the
1218          * offset is available from this window, otherwise the offset
1219          * is not actually in this window and a different window (which
1220          * has that one hash excess) must be used.  This is to support
1221          * the object header and delta base parsing routines below.
1222          */
1223         off_t win_off = win->offset;
1224         return win_off <= offset
1225                 && (offset + 20) <= (win_off + win->len);
1226 }
1227
1228 unsigned char *use_pack(struct packed_git *p,
1229                 struct pack_window **w_cursor,
1230                 off_t offset,
1231                 unsigned long *left)
1232 {
1233         struct pack_window *win = *w_cursor;
1234
1235         /* Since packfiles end in a hash of their content and it's
1236          * pointless to ask for an offset into the middle of that
1237          * hash, and the in_window function above wouldn't match
1238          * don't allow an offset too close to the end of the file.
1239          */
1240         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1241                 die("packfile %s cannot be accessed", p->pack_name);
1242         if (offset > (p->pack_size - 20))
1243                 die("offset beyond end of packfile (truncated pack?)");
1244         if (offset < 0)
1245                 die(_("offset before end of packfile (broken .idx?)"));
1246
1247         if (!win || !in_window(win, offset)) {
1248                 if (win)
1249                         win->inuse_cnt--;
1250                 for (win = p->windows; win; win = win->next) {
1251                         if (in_window(win, offset))
1252                                 break;
1253                 }
1254                 if (!win) {
1255                         size_t window_align = packed_git_window_size / 2;
1256                         off_t len;
1257
1258                         if (p->pack_fd == -1 && open_packed_git(p))
1259                                 die("packfile %s cannot be accessed", p->pack_name);
1260
1261                         win = xcalloc(1, sizeof(*win));
1262                         win->offset = (offset / window_align) * window_align;
1263                         len = p->pack_size - win->offset;
1264                         if (len > packed_git_window_size)
1265                                 len = packed_git_window_size;
1266                         win->len = (size_t)len;
1267                         pack_mapped += win->len;
1268                         while (packed_git_limit < pack_mapped
1269                                 && unuse_one_window(p))
1270                                 ; /* nothing */
1271                         win->base = xmmap(NULL, win->len,
1272                                 PROT_READ, MAP_PRIVATE,
1273                                 p->pack_fd, win->offset);
1274                         if (win->base == MAP_FAILED)
1275                                 die_errno("packfile %s cannot be mapped",
1276                                           p->pack_name);
1277                         if (!win->offset && win->len == p->pack_size
1278                                 && !p->do_not_close)
1279                                 close_pack_fd(p);
1280                         pack_mmap_calls++;
1281                         pack_open_windows++;
1282                         if (pack_mapped > peak_pack_mapped)
1283                                 peak_pack_mapped = pack_mapped;
1284                         if (pack_open_windows > peak_pack_open_windows)
1285                                 peak_pack_open_windows = pack_open_windows;
1286                         win->next = p->windows;
1287                         p->windows = win;
1288                 }
1289         }
1290         if (win != *w_cursor) {
1291                 win->last_used = pack_used_ctr++;
1292                 win->inuse_cnt++;
1293                 *w_cursor = win;
1294         }
1295         offset -= win->offset;
1296         if (left)
1297                 *left = win->len - xsize_t(offset);
1298         return win->base + offset;
1299 }
1300
1301 static struct packed_git *alloc_packed_git(int extra)
1302 {
1303         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1304         memset(p, 0, sizeof(*p));
1305         p->pack_fd = -1;
1306         return p;
1307 }
1308
1309 static void try_to_free_pack_memory(size_t size)
1310 {
1311         release_pack_memory(size);
1312 }
1313
1314 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1315 {
1316         static int have_set_try_to_free_routine;
1317         struct stat st;
1318         size_t alloc;
1319         struct packed_git *p;
1320
1321         if (!have_set_try_to_free_routine) {
1322                 have_set_try_to_free_routine = 1;
1323                 set_try_to_free_routine(try_to_free_pack_memory);
1324         }
1325
1326         /*
1327          * Make sure a corresponding .pack file exists and that
1328          * the index looks sane.
1329          */
1330         if (!strip_suffix_mem(path, &path_len, ".idx"))
1331                 return NULL;
1332
1333         /*
1334          * ".pack" is long enough to hold any suffix we're adding (and
1335          * the use xsnprintf double-checks that)
1336          */
1337         alloc = st_add3(path_len, strlen(".pack"), 1);
1338         p = alloc_packed_git(alloc);
1339         memcpy(p->pack_name, path, path_len);
1340
1341         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1342         if (!access(p->pack_name, F_OK))
1343                 p->pack_keep = 1;
1344
1345         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1346         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1347                 free(p);
1348                 return NULL;
1349         }
1350
1351         /* ok, it looks sane as far as we can check without
1352          * actually mapping the pack file.
1353          */
1354         p->pack_size = st.st_size;
1355         p->pack_local = local;
1356         p->mtime = st.st_mtime;
1357         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1358                 hashclr(p->sha1);
1359         return p;
1360 }
1361
1362 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1363 {
1364         const char *path = sha1_pack_name(sha1);
1365         size_t alloc = st_add(strlen(path), 1);
1366         struct packed_git *p = alloc_packed_git(alloc);
1367
1368         memcpy(p->pack_name, path, alloc); /* includes NUL */
1369         hashcpy(p->sha1, sha1);
1370         if (check_packed_git_idx(idx_path, p)) {
1371                 free(p);
1372                 return NULL;
1373         }
1374
1375         return p;
1376 }
1377
1378 void install_packed_git(struct packed_git *pack)
1379 {
1380         if (pack->pack_fd != -1)
1381                 pack_open_fds++;
1382
1383         pack->next = packed_git;
1384         packed_git = pack;
1385 }
1386
1387 void (*report_garbage)(unsigned seen_bits, const char *path);
1388
1389 static void report_helper(const struct string_list *list,
1390                           int seen_bits, int first, int last)
1391 {
1392         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1393                 return;
1394
1395         for (; first < last; first++)
1396                 report_garbage(seen_bits, list->items[first].string);
1397 }
1398
1399 static void report_pack_garbage(struct string_list *list)
1400 {
1401         int i, baselen = -1, first = 0, seen_bits = 0;
1402
1403         if (!report_garbage)
1404                 return;
1405
1406         string_list_sort(list);
1407
1408         for (i = 0; i < list->nr; i++) {
1409                 const char *path = list->items[i].string;
1410                 if (baselen != -1 &&
1411                     strncmp(path, list->items[first].string, baselen)) {
1412                         report_helper(list, seen_bits, first, i);
1413                         baselen = -1;
1414                         seen_bits = 0;
1415                 }
1416                 if (baselen == -1) {
1417                         const char *dot = strrchr(path, '.');
1418                         if (!dot) {
1419                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1420                                 continue;
1421                         }
1422                         baselen = dot - path + 1;
1423                         first = i;
1424                 }
1425                 if (!strcmp(path + baselen, "pack"))
1426                         seen_bits |= 1;
1427                 else if (!strcmp(path + baselen, "idx"))
1428                         seen_bits |= 2;
1429         }
1430         report_helper(list, seen_bits, first, list->nr);
1431 }
1432
1433 static void prepare_packed_git_one(char *objdir, int local)
1434 {
1435         struct strbuf path = STRBUF_INIT;
1436         size_t dirnamelen;
1437         DIR *dir;
1438         struct dirent *de;
1439         struct string_list garbage = STRING_LIST_INIT_DUP;
1440
1441         strbuf_addstr(&path, objdir);
1442         strbuf_addstr(&path, "/pack");
1443         dir = opendir(path.buf);
1444         if (!dir) {
1445                 if (errno != ENOENT)
1446                         error_errno("unable to open object pack directory: %s",
1447                                     path.buf);
1448                 strbuf_release(&path);
1449                 return;
1450         }
1451         strbuf_addch(&path, '/');
1452         dirnamelen = path.len;
1453         while ((de = readdir(dir)) != NULL) {
1454                 struct packed_git *p;
1455                 size_t base_len;
1456
1457                 if (is_dot_or_dotdot(de->d_name))
1458                         continue;
1459
1460                 strbuf_setlen(&path, dirnamelen);
1461                 strbuf_addstr(&path, de->d_name);
1462
1463                 base_len = path.len;
1464                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1465                         /* Don't reopen a pack we already have. */
1466                         for (p = packed_git; p; p = p->next) {
1467                                 size_t len;
1468                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1469                                     len == base_len &&
1470                                     !memcmp(p->pack_name, path.buf, len))
1471                                         break;
1472                         }
1473                         if (p == NULL &&
1474                             /*
1475                              * See if it really is a valid .idx file with
1476                              * corresponding .pack file that we can map.
1477                              */
1478                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1479                                 install_packed_git(p);
1480                 }
1481
1482                 if (!report_garbage)
1483                         continue;
1484
1485                 if (ends_with(de->d_name, ".idx") ||
1486                     ends_with(de->d_name, ".pack") ||
1487                     ends_with(de->d_name, ".bitmap") ||
1488                     ends_with(de->d_name, ".keep"))
1489                         string_list_append(&garbage, path.buf);
1490                 else
1491                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1492         }
1493         closedir(dir);
1494         report_pack_garbage(&garbage);
1495         string_list_clear(&garbage, 0);
1496         strbuf_release(&path);
1497 }
1498
1499 static int approximate_object_count_valid;
1500
1501 /*
1502  * Give a fast, rough count of the number of objects in the repository. This
1503  * ignores loose objects completely. If you have a lot of them, then either
1504  * you should repack because your performance will be awful, or they are
1505  * all unreachable objects about to be pruned, in which case they're not really
1506  * interesting as a measure of repo size in the first place.
1507  */
1508 unsigned long approximate_object_count(void)
1509 {
1510         static unsigned long count;
1511         if (!approximate_object_count_valid) {
1512                 struct packed_git *p;
1513
1514                 prepare_packed_git();
1515                 count = 0;
1516                 for (p = packed_git; p; p = p->next) {
1517                         if (open_pack_index(p))
1518                                 continue;
1519                         count += p->num_objects;
1520                 }
1521         }
1522         return count;
1523 }
1524
1525 static void *get_next_packed_git(const void *p)
1526 {
1527         return ((const struct packed_git *)p)->next;
1528 }
1529
1530 static void set_next_packed_git(void *p, void *next)
1531 {
1532         ((struct packed_git *)p)->next = next;
1533 }
1534
1535 static int sort_pack(const void *a_, const void *b_)
1536 {
1537         const struct packed_git *a = a_;
1538         const struct packed_git *b = b_;
1539         int st;
1540
1541         /*
1542          * Local packs tend to contain objects specific to our
1543          * variant of the project than remote ones.  In addition,
1544          * remote ones could be on a network mounted filesystem.
1545          * Favor local ones for these reasons.
1546          */
1547         st = a->pack_local - b->pack_local;
1548         if (st)
1549                 return -st;
1550
1551         /*
1552          * Younger packs tend to contain more recent objects,
1553          * and more recent objects tend to get accessed more
1554          * often.
1555          */
1556         if (a->mtime < b->mtime)
1557                 return 1;
1558         else if (a->mtime == b->mtime)
1559                 return 0;
1560         return -1;
1561 }
1562
1563 static void rearrange_packed_git(void)
1564 {
1565         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1566                                      set_next_packed_git, sort_pack);
1567 }
1568
1569 static void prepare_packed_git_mru(void)
1570 {
1571         struct packed_git *p;
1572
1573         mru_clear(packed_git_mru);
1574         for (p = packed_git; p; p = p->next)
1575                 mru_append(packed_git_mru, p);
1576 }
1577
1578 static int prepare_packed_git_run_once = 0;
1579 void prepare_packed_git(void)
1580 {
1581         struct alternate_object_database *alt;
1582
1583         if (prepare_packed_git_run_once)
1584                 return;
1585         prepare_packed_git_one(get_object_directory(), 1);
1586         prepare_alt_odb();
1587         for (alt = alt_odb_list; alt; alt = alt->next)
1588                 prepare_packed_git_one(alt->path, 0);
1589         rearrange_packed_git();
1590         prepare_packed_git_mru();
1591         prepare_packed_git_run_once = 1;
1592 }
1593
1594 void reprepare_packed_git(void)
1595 {
1596         approximate_object_count_valid = 0;
1597         prepare_packed_git_run_once = 0;
1598         prepare_packed_git();
1599 }
1600
1601 static void mark_bad_packed_object(struct packed_git *p,
1602                                    const unsigned char *sha1)
1603 {
1604         unsigned i;
1605         for (i = 0; i < p->num_bad_objects; i++)
1606                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1607                         return;
1608         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1609                                       st_mult(GIT_MAX_RAWSZ,
1610                                               st_add(p->num_bad_objects, 1)));
1611         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1612         p->num_bad_objects++;
1613 }
1614
1615 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1616 {
1617         struct packed_git *p;
1618         unsigned i;
1619
1620         for (p = packed_git; p; p = p->next)
1621                 for (i = 0; i < p->num_bad_objects; i++)
1622                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1623                                 return p;
1624         return NULL;
1625 }
1626
1627 /*
1628  * With an in-core object data in "map", rehash it to make sure the
1629  * object name actually matches "sha1" to detect object corruption.
1630  * With "map" == NULL, try reading the object named with "sha1" using
1631  * the streaming interface and rehash it to do the same.
1632  */
1633 int check_sha1_signature(const unsigned char *sha1, void *map,
1634                          unsigned long size, const char *type)
1635 {
1636         unsigned char real_sha1[20];
1637         enum object_type obj_type;
1638         struct git_istream *st;
1639         git_SHA_CTX c;
1640         char hdr[32];
1641         int hdrlen;
1642
1643         if (map) {
1644                 hash_sha1_file(map, size, type, real_sha1);
1645                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1646         }
1647
1648         st = open_istream(sha1, &obj_type, &size, NULL);
1649         if (!st)
1650                 return -1;
1651
1652         /* Generate the header */
1653         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1654
1655         /* Sha1.. */
1656         git_SHA1_Init(&c);
1657         git_SHA1_Update(&c, hdr, hdrlen);
1658         for (;;) {
1659                 char buf[1024 * 16];
1660                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1661
1662                 if (readlen < 0) {
1663                         close_istream(st);
1664                         return -1;
1665                 }
1666                 if (!readlen)
1667                         break;
1668                 git_SHA1_Update(&c, buf, readlen);
1669         }
1670         git_SHA1_Final(real_sha1, &c);
1671         close_istream(st);
1672         return hashcmp(sha1, real_sha1) ? -1 : 0;
1673 }
1674
1675 int git_open_cloexec(const char *name, int flags)
1676 {
1677         int fd;
1678         static int o_cloexec = O_CLOEXEC;
1679
1680         fd = open(name, flags | o_cloexec);
1681         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
1682                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
1683                 o_cloexec &= ~O_CLOEXEC;
1684                 fd = open(name, flags | o_cloexec);
1685         }
1686
1687 #if defined(F_GETFL) && defined(F_SETFL) && defined(FD_CLOEXEC)
1688         {
1689                 static int fd_cloexec = FD_CLOEXEC;
1690
1691                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
1692                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
1693                         int flags = fcntl(fd, F_GETFL);
1694                         if (fcntl(fd, F_SETFL, flags | fd_cloexec))
1695                                 fd_cloexec = 0;
1696                 }
1697         }
1698 #endif
1699         return fd;
1700 }
1701
1702 /*
1703  * Find "sha1" as a loose object in the local repository or in an alternate.
1704  * Returns 0 on success, negative on failure.
1705  *
1706  * The "path" out-parameter will give the path of the object we found (if any).
1707  * Note that it may point to static storage and is only valid until another
1708  * call to sha1_file_name(), etc.
1709  */
1710 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
1711                           const char **path)
1712 {
1713         struct alternate_object_database *alt;
1714
1715         *path = sha1_file_name(sha1);
1716         if (!lstat(*path, st))
1717                 return 0;
1718
1719         prepare_alt_odb();
1720         errno = ENOENT;
1721         for (alt = alt_odb_list; alt; alt = alt->next) {
1722                 *path = alt_sha1_path(alt, sha1);
1723                 if (!lstat(*path, st))
1724                         return 0;
1725         }
1726
1727         return -1;
1728 }
1729
1730 /*
1731  * Like stat_sha1_file(), but actually open the object and return the
1732  * descriptor. See the caveats on the "path" parameter above.
1733  */
1734 static int open_sha1_file(const unsigned char *sha1, const char **path)
1735 {
1736         int fd;
1737         struct alternate_object_database *alt;
1738         int most_interesting_errno;
1739
1740         *path = sha1_file_name(sha1);
1741         fd = git_open(*path);
1742         if (fd >= 0)
1743                 return fd;
1744         most_interesting_errno = errno;
1745
1746         prepare_alt_odb();
1747         for (alt = alt_odb_list; alt; alt = alt->next) {
1748                 *path = alt_sha1_path(alt, sha1);
1749                 fd = git_open(*path);
1750                 if (fd >= 0)
1751                         return fd;
1752                 if (most_interesting_errno == ENOENT)
1753                         most_interesting_errno = errno;
1754         }
1755         errno = most_interesting_errno;
1756         return -1;
1757 }
1758
1759 /*
1760  * Map the loose object at "path" if it is not NULL, or the path found by
1761  * searching for a loose object named "sha1".
1762  */
1763 static void *map_sha1_file_1(const char *path,
1764                              const unsigned char *sha1,
1765                              unsigned long *size)
1766 {
1767         void *map;
1768         int fd;
1769
1770         if (path)
1771                 fd = git_open(path);
1772         else
1773                 fd = open_sha1_file(sha1, &path);
1774         map = NULL;
1775         if (fd >= 0) {
1776                 struct stat st;
1777
1778                 if (!fstat(fd, &st)) {
1779                         *size = xsize_t(st.st_size);
1780                         if (!*size) {
1781                                 /* mmap() is forbidden on empty files */
1782                                 error("object file %s is empty", path);
1783                                 return NULL;
1784                         }
1785                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1786                 }
1787                 close(fd);
1788         }
1789         return map;
1790 }
1791
1792 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1793 {
1794         return map_sha1_file_1(NULL, sha1, size);
1795 }
1796
1797 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1798                 unsigned long len, enum object_type *type, unsigned long *sizep)
1799 {
1800         unsigned shift;
1801         unsigned long size, c;
1802         unsigned long used = 0;
1803
1804         c = buf[used++];
1805         *type = (c >> 4) & 7;
1806         size = c & 15;
1807         shift = 4;
1808         while (c & 0x80) {
1809                 if (len <= used || bitsizeof(long) <= shift) {
1810                         error("bad object header");
1811                         size = used = 0;
1812                         break;
1813                 }
1814                 c = buf[used++];
1815                 size += (c & 0x7f) << shift;
1816                 shift += 7;
1817         }
1818         *sizep = size;
1819         return used;
1820 }
1821
1822 static int unpack_sha1_short_header(git_zstream *stream,
1823                                     unsigned char *map, unsigned long mapsize,
1824                                     void *buffer, unsigned long bufsiz)
1825 {
1826         /* Get the data stream */
1827         memset(stream, 0, sizeof(*stream));
1828         stream->next_in = map;
1829         stream->avail_in = mapsize;
1830         stream->next_out = buffer;
1831         stream->avail_out = bufsiz;
1832
1833         git_inflate_init(stream);
1834         return git_inflate(stream, 0);
1835 }
1836
1837 int unpack_sha1_header(git_zstream *stream,
1838                        unsigned char *map, unsigned long mapsize,
1839                        void *buffer, unsigned long bufsiz)
1840 {
1841         int status = unpack_sha1_short_header(stream, map, mapsize,
1842                                               buffer, bufsiz);
1843
1844         if (status < Z_OK)
1845                 return status;
1846
1847         /* Make sure we have the terminating NUL */
1848         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1849                 return -1;
1850         return 0;
1851 }
1852
1853 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1854                                         unsigned long mapsize, void *buffer,
1855                                         unsigned long bufsiz, struct strbuf *header)
1856 {
1857         int status;
1858
1859         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
1860         if (status < Z_OK)
1861                 return -1;
1862
1863         /*
1864          * Check if entire header is unpacked in the first iteration.
1865          */
1866         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1867                 return 0;
1868
1869         /*
1870          * buffer[0..bufsiz] was not large enough.  Copy the partial
1871          * result out to header, and then append the result of further
1872          * reading the stream.
1873          */
1874         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1875         stream->next_out = buffer;
1876         stream->avail_out = bufsiz;
1877
1878         do {
1879                 status = git_inflate(stream, 0);
1880                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1881                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1882                         return 0;
1883                 stream->next_out = buffer;
1884                 stream->avail_out = bufsiz;
1885         } while (status != Z_STREAM_END);
1886         return -1;
1887 }
1888
1889 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1890 {
1891         int bytes = strlen(buffer) + 1;
1892         unsigned char *buf = xmallocz(size);
1893         unsigned long n;
1894         int status = Z_OK;
1895
1896         n = stream->total_out - bytes;
1897         if (n > size)
1898                 n = size;
1899         memcpy(buf, (char *) buffer + bytes, n);
1900         bytes = n;
1901         if (bytes <= size) {
1902                 /*
1903                  * The above condition must be (bytes <= size), not
1904                  * (bytes < size).  In other words, even though we
1905                  * expect no more output and set avail_out to zero,
1906                  * the input zlib stream may have bytes that express
1907                  * "this concludes the stream", and we *do* want to
1908                  * eat that input.
1909                  *
1910                  * Otherwise we would not be able to test that we
1911                  * consumed all the input to reach the expected size;
1912                  * we also want to check that zlib tells us that all
1913                  * went well with status == Z_STREAM_END at the end.
1914                  */
1915                 stream->next_out = buf + bytes;
1916                 stream->avail_out = size - bytes;
1917                 while (status == Z_OK)
1918                         status = git_inflate(stream, Z_FINISH);
1919         }
1920         if (status == Z_STREAM_END && !stream->avail_in) {
1921                 git_inflate_end(stream);
1922                 return buf;
1923         }
1924
1925         if (status < 0)
1926                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1927         else if (stream->avail_in)
1928                 error("garbage at end of loose object '%s'",
1929                       sha1_to_hex(sha1));
1930         free(buf);
1931         return NULL;
1932 }
1933
1934 /*
1935  * We used to just use "sscanf()", but that's actually way
1936  * too permissive for what we want to check. So do an anal
1937  * object header parse by hand.
1938  */
1939 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1940                                unsigned int flags)
1941 {
1942         const char *type_buf = hdr;
1943         unsigned long size;
1944         int type, type_len = 0;
1945
1946         /*
1947          * The type can be of any size but is followed by
1948          * a space.
1949          */
1950         for (;;) {
1951                 char c = *hdr++;
1952                 if (!c)
1953                         return -1;
1954                 if (c == ' ')
1955                         break;
1956                 type_len++;
1957         }
1958
1959         type = type_from_string_gently(type_buf, type_len, 1);
1960         if (oi->typename)
1961                 strbuf_add(oi->typename, type_buf, type_len);
1962         /*
1963          * Set type to 0 if its an unknown object and
1964          * we're obtaining the type using '--allow-unknown-type'
1965          * option.
1966          */
1967         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1968                 type = 0;
1969         else if (type < 0)
1970                 die("invalid object type");
1971         if (oi->typep)
1972                 *oi->typep = type;
1973
1974         /*
1975          * The length must follow immediately, and be in canonical
1976          * decimal format (ie "010" is not valid).
1977          */
1978         size = *hdr++ - '0';
1979         if (size > 9)
1980                 return -1;
1981         if (size) {
1982                 for (;;) {
1983                         unsigned long c = *hdr - '0';
1984                         if (c > 9)
1985                                 break;
1986                         hdr++;
1987                         size = size * 10 + c;
1988                 }
1989         }
1990
1991         if (oi->sizep)
1992                 *oi->sizep = size;
1993
1994         /*
1995          * The length must be followed by a zero byte
1996          */
1997         return *hdr ? -1 : type;
1998 }
1999
2000 int parse_sha1_header(const char *hdr, unsigned long *sizep)
2001 {
2002         struct object_info oi = OBJECT_INFO_INIT;
2003
2004         oi.sizep = sizep;
2005         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
2006 }
2007
2008 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
2009 {
2010         int ret;
2011         git_zstream stream;
2012         char hdr[8192];
2013
2014         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
2015         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
2016                 return NULL;
2017
2018         return unpack_sha1_rest(&stream, hdr, *size, sha1);
2019 }
2020
2021 unsigned long get_size_from_delta(struct packed_git *p,
2022                                   struct pack_window **w_curs,
2023                                   off_t curpos)
2024 {
2025         const unsigned char *data;
2026         unsigned char delta_head[20], *in;
2027         git_zstream stream;
2028         int st;
2029
2030         memset(&stream, 0, sizeof(stream));
2031         stream.next_out = delta_head;
2032         stream.avail_out = sizeof(delta_head);
2033
2034         git_inflate_init(&stream);
2035         do {
2036                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2037                 stream.next_in = in;
2038                 st = git_inflate(&stream, Z_FINISH);
2039                 curpos += stream.next_in - in;
2040         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
2041                  stream.total_out < sizeof(delta_head));
2042         git_inflate_end(&stream);
2043         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
2044                 error("delta data unpack-initial failed");
2045                 return 0;
2046         }
2047
2048         /* Examine the initial part of the delta to figure out
2049          * the result size.
2050          */
2051         data = delta_head;
2052
2053         /* ignore base size */
2054         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2055
2056         /* Read the result size */
2057         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2058 }
2059
2060 static off_t get_delta_base(struct packed_git *p,
2061                                     struct pack_window **w_curs,
2062                                     off_t *curpos,
2063                                     enum object_type type,
2064                                     off_t delta_obj_offset)
2065 {
2066         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
2067         off_t base_offset;
2068
2069         /* use_pack() assured us we have [base_info, base_info + 20)
2070          * as a range that we can look at without walking off the
2071          * end of the mapped window.  Its actually the hash size
2072          * that is assured.  An OFS_DELTA longer than the hash size
2073          * is stupid, as then a REF_DELTA would be smaller to store.
2074          */
2075         if (type == OBJ_OFS_DELTA) {
2076                 unsigned used = 0;
2077                 unsigned char c = base_info[used++];
2078                 base_offset = c & 127;
2079                 while (c & 128) {
2080                         base_offset += 1;
2081                         if (!base_offset || MSB(base_offset, 7))
2082                                 return 0;  /* overflow */
2083                         c = base_info[used++];
2084                         base_offset = (base_offset << 7) + (c & 127);
2085                 }
2086                 base_offset = delta_obj_offset - base_offset;
2087                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
2088                         return 0;  /* out of bound */
2089                 *curpos += used;
2090         } else if (type == OBJ_REF_DELTA) {
2091                 /* The base entry _must_ be in the same pack */
2092                 base_offset = find_pack_entry_one(base_info, p);
2093                 *curpos += 20;
2094         } else
2095                 die("I am totally screwed");
2096         return base_offset;
2097 }
2098
2099 /*
2100  * Like get_delta_base above, but we return the sha1 instead of the pack
2101  * offset. This means it is cheaper for REF deltas (we do not have to do
2102  * the final object lookup), but more expensive for OFS deltas (we
2103  * have to load the revidx to convert the offset back into a sha1).
2104  */
2105 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
2106                                                 struct pack_window **w_curs,
2107                                                 off_t curpos,
2108                                                 enum object_type type,
2109                                                 off_t delta_obj_offset)
2110 {
2111         if (type == OBJ_REF_DELTA) {
2112                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
2113                 return base;
2114         } else if (type == OBJ_OFS_DELTA) {
2115                 struct revindex_entry *revidx;
2116                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
2117                                                    type, delta_obj_offset);
2118
2119                 if (!base_offset)
2120                         return NULL;
2121
2122                 revidx = find_pack_revindex(p, base_offset);
2123                 if (!revidx)
2124                         return NULL;
2125
2126                 return nth_packed_object_sha1(p, revidx->nr);
2127         } else
2128                 return NULL;
2129 }
2130
2131 int unpack_object_header(struct packed_git *p,
2132                          struct pack_window **w_curs,
2133                          off_t *curpos,
2134                          unsigned long *sizep)
2135 {
2136         unsigned char *base;
2137         unsigned long left;
2138         unsigned long used;
2139         enum object_type type;
2140
2141         /* use_pack() assures us we have [base, base + 20) available
2142          * as a range that we can look at.  (Its actually the hash
2143          * size that is assured.)  With our object header encoding
2144          * the maximum deflated object size is 2^137, which is just
2145          * insane, so we know won't exceed what we have been given.
2146          */
2147         base = use_pack(p, w_curs, *curpos, &left);
2148         used = unpack_object_header_buffer(base, left, &type, sizep);
2149         if (!used) {
2150                 type = OBJ_BAD;
2151         } else
2152                 *curpos += used;
2153
2154         return type;
2155 }
2156
2157 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
2158 {
2159         int type;
2160         struct revindex_entry *revidx;
2161         const unsigned char *sha1;
2162         revidx = find_pack_revindex(p, obj_offset);
2163         if (!revidx)
2164                 return OBJ_BAD;
2165         sha1 = nth_packed_object_sha1(p, revidx->nr);
2166         mark_bad_packed_object(p, sha1);
2167         type = sha1_object_info(sha1, NULL);
2168         if (type <= OBJ_NONE)
2169                 return OBJ_BAD;
2170         return type;
2171 }
2172
2173 #define POI_STACK_PREALLOC 64
2174
2175 static enum object_type packed_to_object_type(struct packed_git *p,
2176                                               off_t obj_offset,
2177                                               enum object_type type,
2178                                               struct pack_window **w_curs,
2179                                               off_t curpos)
2180 {
2181         off_t small_poi_stack[POI_STACK_PREALLOC];
2182         off_t *poi_stack = small_poi_stack;
2183         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
2184
2185         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2186                 off_t base_offset;
2187                 unsigned long size;
2188                 /* Push the object we're going to leave behind */
2189                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
2190                         poi_stack_alloc = alloc_nr(poi_stack_nr);
2191                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
2192                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2193                 } else {
2194                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2195                 }
2196                 poi_stack[poi_stack_nr++] = obj_offset;
2197                 /* If parsing the base offset fails, just unwind */
2198                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2199                 if (!base_offset)
2200                         goto unwind;
2201                 curpos = obj_offset = base_offset;
2202                 type = unpack_object_header(p, w_curs, &curpos, &size);
2203                 if (type <= OBJ_NONE) {
2204                         /* If getting the base itself fails, we first
2205                          * retry the base, otherwise unwind */
2206                         type = retry_bad_packed_offset(p, base_offset);
2207                         if (type > OBJ_NONE)
2208                                 goto out;
2209                         goto unwind;
2210                 }
2211         }
2212
2213         switch (type) {
2214         case OBJ_BAD:
2215         case OBJ_COMMIT:
2216         case OBJ_TREE:
2217         case OBJ_BLOB:
2218         case OBJ_TAG:
2219                 break;
2220         default:
2221                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2222                       type, (uintmax_t)obj_offset, p->pack_name);
2223                 type = OBJ_BAD;
2224         }
2225
2226 out:
2227         if (poi_stack != small_poi_stack)
2228                 free(poi_stack);
2229         return type;
2230
2231 unwind:
2232         while (poi_stack_nr) {
2233                 obj_offset = poi_stack[--poi_stack_nr];
2234                 type = retry_bad_packed_offset(p, obj_offset);
2235                 if (type > OBJ_NONE)
2236                         goto out;
2237         }
2238         type = OBJ_BAD;
2239         goto out;
2240 }
2241
2242 int packed_object_info(struct packed_git *p, off_t obj_offset,
2243                        struct object_info *oi)
2244 {
2245         struct pack_window *w_curs = NULL;
2246         unsigned long size;
2247         off_t curpos = obj_offset;
2248         enum object_type type;
2249
2250         /*
2251          * We always get the representation type, but only convert it to
2252          * a "real" type later if the caller is interested.
2253          */
2254         type = unpack_object_header(p, &w_curs, &curpos, &size);
2255
2256         if (oi->sizep) {
2257                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2258                         off_t tmp_pos = curpos;
2259                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2260                                                            type, obj_offset);
2261                         if (!base_offset) {
2262                                 type = OBJ_BAD;
2263                                 goto out;
2264                         }
2265                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2266                         if (*oi->sizep == 0) {
2267                                 type = OBJ_BAD;
2268                                 goto out;
2269                         }
2270                 } else {
2271                         *oi->sizep = size;
2272                 }
2273         }
2274
2275         if (oi->disk_sizep) {
2276                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2277                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2278         }
2279
2280         if (oi->typep) {
2281                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2282                 if (*oi->typep < 0) {
2283                         type = OBJ_BAD;
2284                         goto out;
2285                 }
2286         }
2287
2288         if (oi->delta_base_sha1) {
2289                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2290                         const unsigned char *base;
2291
2292                         base = get_delta_base_sha1(p, &w_curs, curpos,
2293                                                    type, obj_offset);
2294                         if (!base) {
2295                                 type = OBJ_BAD;
2296                                 goto out;
2297                         }
2298
2299                         hashcpy(oi->delta_base_sha1, base);
2300                 } else
2301                         hashclr(oi->delta_base_sha1);
2302         }
2303
2304 out:
2305         unuse_pack(&w_curs);
2306         return type;
2307 }
2308
2309 static void *unpack_compressed_entry(struct packed_git *p,
2310                                     struct pack_window **w_curs,
2311                                     off_t curpos,
2312                                     unsigned long size)
2313 {
2314         int st;
2315         git_zstream stream;
2316         unsigned char *buffer, *in;
2317
2318         buffer = xmallocz_gently(size);
2319         if (!buffer)
2320                 return NULL;
2321         memset(&stream, 0, sizeof(stream));
2322         stream.next_out = buffer;
2323         stream.avail_out = size + 1;
2324
2325         git_inflate_init(&stream);
2326         do {
2327                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2328                 stream.next_in = in;
2329                 st = git_inflate(&stream, Z_FINISH);
2330                 if (!stream.avail_out)
2331                         break; /* the payload is larger than it should be */
2332                 curpos += stream.next_in - in;
2333         } while (st == Z_OK || st == Z_BUF_ERROR);
2334         git_inflate_end(&stream);
2335         if ((st != Z_STREAM_END) || stream.total_out != size) {
2336                 free(buffer);
2337                 return NULL;
2338         }
2339
2340         return buffer;
2341 }
2342
2343 static struct hashmap delta_base_cache;
2344 static size_t delta_base_cached;
2345
2346 static LIST_HEAD(delta_base_cache_lru);
2347
2348 struct delta_base_cache_key {
2349         struct packed_git *p;
2350         off_t base_offset;
2351 };
2352
2353 struct delta_base_cache_entry {
2354         struct hashmap hash;
2355         struct delta_base_cache_key key;
2356         struct list_head lru;
2357         void *data;
2358         unsigned long size;
2359         enum object_type type;
2360 };
2361
2362 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2363 {
2364         unsigned int hash;
2365
2366         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2367         hash += (hash >> 8) + (hash >> 16);
2368         return hash;
2369 }
2370
2371 static struct delta_base_cache_entry *
2372 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2373 {
2374         struct hashmap_entry entry;
2375         struct delta_base_cache_key key;
2376
2377         if (!delta_base_cache.cmpfn)
2378                 return NULL;
2379
2380         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2381         key.p = p;
2382         key.base_offset = base_offset;
2383         return hashmap_get(&delta_base_cache, &entry, &key);
2384 }
2385
2386 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2387                                    const struct delta_base_cache_key *b)
2388 {
2389         return a->p == b->p && a->base_offset == b->base_offset;
2390 }
2391
2392 static int delta_base_cache_hash_cmp(const void *va, const void *vb,
2393                                      const void *vkey)
2394 {
2395         const struct delta_base_cache_entry *a = va, *b = vb;
2396         const struct delta_base_cache_key *key = vkey;
2397         if (key)
2398                 return !delta_base_cache_key_eq(&a->key, key);
2399         else
2400                 return !delta_base_cache_key_eq(&a->key, &b->key);
2401 }
2402
2403 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2404 {
2405         return !!get_delta_base_cache_entry(p, base_offset);
2406 }
2407
2408 /*
2409  * Remove the entry from the cache, but do _not_ free the associated
2410  * entry data. The caller takes ownership of the "data" buffer, and
2411  * should copy out any fields it wants before detaching.
2412  */
2413 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2414 {
2415         hashmap_remove(&delta_base_cache, ent, &ent->key);
2416         list_del(&ent->lru);
2417         delta_base_cached -= ent->size;
2418         free(ent);
2419 }
2420
2421 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2422         unsigned long *base_size, enum object_type *type)
2423 {
2424         struct delta_base_cache_entry *ent;
2425
2426         ent = get_delta_base_cache_entry(p, base_offset);
2427         if (!ent)
2428                 return unpack_entry(p, base_offset, type, base_size);
2429
2430         *type = ent->type;
2431         *base_size = ent->size;
2432         return xmemdupz(ent->data, ent->size);
2433 }
2434
2435 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2436 {
2437         free(ent->data);
2438         detach_delta_base_cache_entry(ent);
2439 }
2440
2441 void clear_delta_base_cache(void)
2442 {
2443         struct list_head *lru, *tmp;
2444         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2445                 struct delta_base_cache_entry *entry =
2446                         list_entry(lru, struct delta_base_cache_entry, lru);
2447                 release_delta_base_cache(entry);
2448         }
2449 }
2450
2451 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2452         void *base, unsigned long base_size, enum object_type type)
2453 {
2454         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2455         struct list_head *lru, *tmp;
2456
2457         delta_base_cached += base_size;
2458
2459         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2460                 struct delta_base_cache_entry *f =
2461                         list_entry(lru, struct delta_base_cache_entry, lru);
2462                 if (delta_base_cached <= delta_base_cache_limit)
2463                         break;
2464                 release_delta_base_cache(f);
2465         }
2466
2467         ent->key.p = p;
2468         ent->key.base_offset = base_offset;
2469         ent->type = type;
2470         ent->data = base;
2471         ent->size = base_size;
2472         list_add_tail(&ent->lru, &delta_base_cache_lru);
2473
2474         if (!delta_base_cache.cmpfn)
2475                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, 0);
2476         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2477         hashmap_add(&delta_base_cache, ent);
2478 }
2479
2480 static void *read_object(const unsigned char *sha1, enum object_type *type,
2481                          unsigned long *size);
2482
2483 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2484 {
2485         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2486         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2487                          p->pack_name, (uintmax_t)obj_offset);
2488 }
2489
2490 int do_check_packed_object_crc;
2491
2492 #define UNPACK_ENTRY_STACK_PREALLOC 64
2493 struct unpack_entry_stack_ent {
2494         off_t obj_offset;
2495         off_t curpos;
2496         unsigned long size;
2497 };
2498
2499 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2500                    enum object_type *final_type, unsigned long *final_size)
2501 {
2502         struct pack_window *w_curs = NULL;
2503         off_t curpos = obj_offset;
2504         void *data = NULL;
2505         unsigned long size;
2506         enum object_type type;
2507         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2508         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2509         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2510         int base_from_cache = 0;
2511
2512         write_pack_access_log(p, obj_offset);
2513
2514         /* PHASE 1: drill down to the innermost base object */
2515         for (;;) {
2516                 off_t base_offset;
2517                 int i;
2518                 struct delta_base_cache_entry *ent;
2519
2520                 ent = get_delta_base_cache_entry(p, curpos);
2521                 if (ent) {
2522                         type = ent->type;
2523                         data = ent->data;
2524                         size = ent->size;
2525                         detach_delta_base_cache_entry(ent);
2526                         base_from_cache = 1;
2527                         break;
2528                 }
2529
2530                 if (do_check_packed_object_crc && p->index_version > 1) {
2531                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2532                         off_t len = revidx[1].offset - obj_offset;
2533                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2534                                 const unsigned char *sha1 =
2535                                         nth_packed_object_sha1(p, revidx->nr);
2536                                 error("bad packed object CRC for %s",
2537                                       sha1_to_hex(sha1));
2538                                 mark_bad_packed_object(p, sha1);
2539                                 unuse_pack(&w_curs);
2540                                 return NULL;
2541                         }
2542                 }
2543
2544                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2545                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2546                         break;
2547
2548                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2549                 if (!base_offset) {
2550                         error("failed to validate delta base reference "
2551                               "at offset %"PRIuMAX" from %s",
2552                               (uintmax_t)curpos, p->pack_name);
2553                         /* bail to phase 2, in hopes of recovery */
2554                         data = NULL;
2555                         break;
2556                 }
2557
2558                 /* push object, proceed to base */
2559                 if (delta_stack_nr >= delta_stack_alloc
2560                     && delta_stack == small_delta_stack) {
2561                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2562                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2563                         memcpy(delta_stack, small_delta_stack,
2564                                sizeof(*delta_stack)*delta_stack_nr);
2565                 } else {
2566                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2567                 }
2568                 i = delta_stack_nr++;
2569                 delta_stack[i].obj_offset = obj_offset;
2570                 delta_stack[i].curpos = curpos;
2571                 delta_stack[i].size = size;
2572
2573                 curpos = obj_offset = base_offset;
2574         }
2575
2576         /* PHASE 2: handle the base */
2577         switch (type) {
2578         case OBJ_OFS_DELTA:
2579         case OBJ_REF_DELTA:
2580                 if (data)
2581                         die("BUG: unpack_entry: left loop at a valid delta");
2582                 break;
2583         case OBJ_COMMIT:
2584         case OBJ_TREE:
2585         case OBJ_BLOB:
2586         case OBJ_TAG:
2587                 if (!base_from_cache)
2588                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2589                 break;
2590         default:
2591                 data = NULL;
2592                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2593                       type, (uintmax_t)obj_offset, p->pack_name);
2594         }
2595
2596         /* PHASE 3: apply deltas in order */
2597
2598         /* invariants:
2599          *   'data' holds the base data, or NULL if there was corruption
2600          */
2601         while (delta_stack_nr) {
2602                 void *delta_data;
2603                 void *base = data;
2604                 void *external_base = NULL;
2605                 unsigned long delta_size, base_size = size;
2606                 int i;
2607
2608                 data = NULL;
2609
2610                 if (base)
2611                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2612
2613                 if (!base) {
2614                         /*
2615                          * We're probably in deep shit, but let's try to fetch
2616                          * the required base anyway from another pack or loose.
2617                          * This is costly but should happen only in the presence
2618                          * of a corrupted pack, and is better than failing outright.
2619                          */
2620                         struct revindex_entry *revidx;
2621                         const unsigned char *base_sha1;
2622                         revidx = find_pack_revindex(p, obj_offset);
2623                         if (revidx) {
2624                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2625                                 error("failed to read delta base object %s"
2626                                       " at offset %"PRIuMAX" from %s",
2627                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2628                                       p->pack_name);
2629                                 mark_bad_packed_object(p, base_sha1);
2630                                 base = read_object(base_sha1, &type, &base_size);
2631                                 external_base = base;
2632                         }
2633                 }
2634
2635                 i = --delta_stack_nr;
2636                 obj_offset = delta_stack[i].obj_offset;
2637                 curpos = delta_stack[i].curpos;
2638                 delta_size = delta_stack[i].size;
2639
2640                 if (!base)
2641                         continue;
2642
2643                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2644
2645                 if (!delta_data) {
2646                         error("failed to unpack compressed delta "
2647                               "at offset %"PRIuMAX" from %s",
2648                               (uintmax_t)curpos, p->pack_name);
2649                         data = NULL;
2650                         free(external_base);
2651                         continue;
2652                 }
2653
2654                 data = patch_delta(base, base_size,
2655                                    delta_data, delta_size,
2656                                    &size);
2657
2658                 /*
2659                  * We could not apply the delta; warn the user, but keep going.
2660                  * Our failure will be noticed either in the next iteration of
2661                  * the loop, or if this is the final delta, in the caller when
2662                  * we return NULL. Those code paths will take care of making
2663                  * a more explicit warning and retrying with another copy of
2664                  * the object.
2665                  */
2666                 if (!data)
2667                         error("failed to apply delta");
2668
2669                 free(delta_data);
2670                 free(external_base);
2671         }
2672
2673         *final_type = type;
2674         *final_size = size;
2675
2676         unuse_pack(&w_curs);
2677
2678         if (delta_stack != small_delta_stack)
2679                 free(delta_stack);
2680
2681         return data;
2682 }
2683
2684 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2685                                             uint32_t n)
2686 {
2687         const unsigned char *index = p->index_data;
2688         if (!index) {
2689                 if (open_pack_index(p))
2690                         return NULL;
2691                 index = p->index_data;
2692         }
2693         if (n >= p->num_objects)
2694                 return NULL;
2695         index += 4 * 256;
2696         if (p->index_version == 1) {
2697                 return index + 24 * n + 4;
2698         } else {
2699                 index += 8;
2700                 return index + 20 * n;
2701         }
2702 }
2703
2704 const struct object_id *nth_packed_object_oid(struct object_id *oid,
2705                                               struct packed_git *p,
2706                                               uint32_t n)
2707 {
2708         const unsigned char *hash = nth_packed_object_sha1(p, n);
2709         if (!hash)
2710                 return NULL;
2711         hashcpy(oid->hash, hash);
2712         return oid;
2713 }
2714
2715 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2716 {
2717         const unsigned char *ptr = vptr;
2718         const unsigned char *start = p->index_data;
2719         const unsigned char *end = start + p->index_size;
2720         if (ptr < start)
2721                 die(_("offset before start of pack index for %s (corrupt index?)"),
2722                     p->pack_name);
2723         /* No need to check for underflow; .idx files must be at least 8 bytes */
2724         if (ptr >= end - 8)
2725                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2726                     p->pack_name);
2727 }
2728
2729 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2730 {
2731         const unsigned char *index = p->index_data;
2732         index += 4 * 256;
2733         if (p->index_version == 1) {
2734                 return ntohl(*((uint32_t *)(index + 24 * n)));
2735         } else {
2736                 uint32_t off;
2737                 index += 8 + p->num_objects * (20 + 4);
2738                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2739                 if (!(off & 0x80000000))
2740                         return off;
2741                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2742                 check_pack_index_ptr(p, index);
2743                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2744                                    ntohl(*((uint32_t *)(index + 4)));
2745         }
2746 }
2747
2748 off_t find_pack_entry_one(const unsigned char *sha1,
2749                                   struct packed_git *p)
2750 {
2751         const uint32_t *level1_ofs = p->index_data;
2752         const unsigned char *index = p->index_data;
2753         unsigned hi, lo, stride;
2754         static int use_lookup = -1;
2755         static int debug_lookup = -1;
2756
2757         if (debug_lookup < 0)
2758                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2759
2760         if (!index) {
2761                 if (open_pack_index(p))
2762                         return 0;
2763                 level1_ofs = p->index_data;
2764                 index = p->index_data;
2765         }
2766         if (p->index_version > 1) {
2767                 level1_ofs += 2;
2768                 index += 8;
2769         }
2770         index += 4 * 256;
2771         hi = ntohl(level1_ofs[*sha1]);
2772         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2773         if (p->index_version > 1) {
2774                 stride = 20;
2775         } else {
2776                 stride = 24;
2777                 index += 4;
2778         }
2779
2780         if (debug_lookup)
2781                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2782                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2783
2784         if (use_lookup < 0)
2785                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2786         if (use_lookup) {
2787                 int pos = sha1_entry_pos(index, stride, 0,
2788                                          lo, hi, p->num_objects, sha1);
2789                 if (pos < 0)
2790                         return 0;
2791                 return nth_packed_object_offset(p, pos);
2792         }
2793
2794         do {
2795                 unsigned mi = (lo + hi) / 2;
2796                 int cmp = hashcmp(index + mi * stride, sha1);
2797
2798                 if (debug_lookup)
2799                         printf("lo %u hi %u rg %u mi %u\n",
2800                                lo, hi, hi - lo, mi);
2801                 if (!cmp)
2802                         return nth_packed_object_offset(p, mi);
2803                 if (cmp > 0)
2804                         hi = mi;
2805                 else
2806                         lo = mi+1;
2807         } while (lo < hi);
2808         return 0;
2809 }
2810
2811 int is_pack_valid(struct packed_git *p)
2812 {
2813         /* An already open pack is known to be valid. */
2814         if (p->pack_fd != -1)
2815                 return 1;
2816
2817         /* If the pack has one window completely covering the
2818          * file size, the pack is known to be valid even if
2819          * the descriptor is not currently open.
2820          */
2821         if (p->windows) {
2822                 struct pack_window *w = p->windows;
2823
2824                 if (!w->offset && w->len == p->pack_size)
2825                         return 1;
2826         }
2827
2828         /* Force the pack to open to prove its valid. */
2829         return !open_packed_git(p);
2830 }
2831
2832 static int fill_pack_entry(const unsigned char *sha1,
2833                            struct pack_entry *e,
2834                            struct packed_git *p)
2835 {
2836         off_t offset;
2837
2838         if (p->num_bad_objects) {
2839                 unsigned i;
2840                 for (i = 0; i < p->num_bad_objects; i++)
2841                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2842                                 return 0;
2843         }
2844
2845         offset = find_pack_entry_one(sha1, p);
2846         if (!offset)
2847                 return 0;
2848
2849         /*
2850          * We are about to tell the caller where they can locate the
2851          * requested object.  We better make sure the packfile is
2852          * still here and can be accessed before supplying that
2853          * answer, as it may have been deleted since the index was
2854          * loaded!
2855          */
2856         if (!is_pack_valid(p))
2857                 return 0;
2858         e->offset = offset;
2859         e->p = p;
2860         hashcpy(e->sha1, sha1);
2861         return 1;
2862 }
2863
2864 /*
2865  * Iff a pack file contains the object named by sha1, return true and
2866  * store its location to e.
2867  */
2868 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2869 {
2870         struct mru_entry *p;
2871
2872         prepare_packed_git();
2873         if (!packed_git)
2874                 return 0;
2875
2876         for (p = packed_git_mru->head; p; p = p->next) {
2877                 if (fill_pack_entry(sha1, e, p->item)) {
2878                         mru_mark(packed_git_mru, p);
2879                         return 1;
2880                 }
2881         }
2882         return 0;
2883 }
2884
2885 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2886                                   struct packed_git *packs)
2887 {
2888         struct packed_git *p;
2889
2890         for (p = packs; p; p = p->next) {
2891                 if (find_pack_entry_one(sha1, p))
2892                         return p;
2893         }
2894         return NULL;
2895
2896 }
2897
2898 static int sha1_loose_object_info(const unsigned char *sha1,
2899                                   struct object_info *oi,
2900                                   int flags)
2901 {
2902         int status = 0;
2903         unsigned long mapsize;
2904         void *map;
2905         git_zstream stream;
2906         char hdr[32];
2907         struct strbuf hdrbuf = STRBUF_INIT;
2908
2909         if (oi->delta_base_sha1)
2910                 hashclr(oi->delta_base_sha1);
2911
2912         /*
2913          * If we don't care about type or size, then we don't
2914          * need to look inside the object at all. Note that we
2915          * do not optimize out the stat call, even if the
2916          * caller doesn't care about the disk-size, since our
2917          * return value implicitly indicates whether the
2918          * object even exists.
2919          */
2920         if (!oi->typep && !oi->typename && !oi->sizep) {
2921                 const char *path;
2922                 struct stat st;
2923                 if (stat_sha1_file(sha1, &st, &path) < 0)
2924                         return -1;
2925                 if (oi->disk_sizep)
2926                         *oi->disk_sizep = st.st_size;
2927                 return 0;
2928         }
2929
2930         map = map_sha1_file(sha1, &mapsize);
2931         if (!map)
2932                 return -1;
2933         if (oi->disk_sizep)
2934                 *oi->disk_sizep = mapsize;
2935         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2936                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2937                         status = error("unable to unpack %s header with --allow-unknown-type",
2938                                        sha1_to_hex(sha1));
2939         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2940                 status = error("unable to unpack %s header",
2941                                sha1_to_hex(sha1));
2942         if (status < 0)
2943                 ; /* Do nothing */
2944         else if (hdrbuf.len) {
2945                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2946                         status = error("unable to parse %s header with --allow-unknown-type",
2947                                        sha1_to_hex(sha1));
2948         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2949                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2950         git_inflate_end(&stream);
2951         munmap(map, mapsize);
2952         if (status && oi->typep)
2953                 *oi->typep = status;
2954         strbuf_release(&hdrbuf);
2955         return (status < 0) ? status : 0;
2956 }
2957
2958 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2959 {
2960         struct cached_object *co;
2961         struct pack_entry e;
2962         int rtype;
2963         enum object_type real_type;
2964         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2965
2966         co = find_cached_object(real);
2967         if (co) {
2968                 if (oi->typep)
2969                         *(oi->typep) = co->type;
2970                 if (oi->sizep)
2971                         *(oi->sizep) = co->size;
2972                 if (oi->disk_sizep)
2973                         *(oi->disk_sizep) = 0;
2974                 if (oi->delta_base_sha1)
2975                         hashclr(oi->delta_base_sha1);
2976                 if (oi->typename)
2977                         strbuf_addstr(oi->typename, typename(co->type));
2978                 oi->whence = OI_CACHED;
2979                 return 0;
2980         }
2981
2982         if (!find_pack_entry(real, &e)) {
2983                 /* Most likely it's a loose object. */
2984                 if (!sha1_loose_object_info(real, oi, flags)) {
2985                         oi->whence = OI_LOOSE;
2986                         return 0;
2987                 }
2988
2989                 /* Not a loose object; someone else may have just packed it. */
2990                 reprepare_packed_git();
2991                 if (!find_pack_entry(real, &e))
2992                         return -1;
2993         }
2994
2995         /*
2996          * packed_object_info() does not follow the delta chain to
2997          * find out the real type, unless it is given oi->typep.
2998          */
2999         if (oi->typename && !oi->typep)
3000                 oi->typep = &real_type;
3001
3002         rtype = packed_object_info(e.p, e.offset, oi);
3003         if (rtype < 0) {
3004                 mark_bad_packed_object(e.p, real);
3005                 if (oi->typep == &real_type)
3006                         oi->typep = NULL;
3007                 return sha1_object_info_extended(real, oi, 0);
3008         } else if (in_delta_base_cache(e.p, e.offset)) {
3009                 oi->whence = OI_DBCACHED;
3010         } else {
3011                 oi->whence = OI_PACKED;
3012                 oi->u.packed.offset = e.offset;
3013                 oi->u.packed.pack = e.p;
3014                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
3015                                          rtype == OBJ_OFS_DELTA);
3016         }
3017         if (oi->typename)
3018                 strbuf_addstr(oi->typename, typename(*oi->typep));
3019         if (oi->typep == &real_type)
3020                 oi->typep = NULL;
3021
3022         return 0;
3023 }
3024
3025 /* returns enum object_type or negative */
3026 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
3027 {
3028         enum object_type type;
3029         struct object_info oi = OBJECT_INFO_INIT;
3030
3031         oi.typep = &type;
3032         oi.sizep = sizep;
3033         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
3034                 return -1;
3035         return type;
3036 }
3037
3038 static void *read_packed_sha1(const unsigned char *sha1,
3039                               enum object_type *type, unsigned long *size)
3040 {
3041         struct pack_entry e;
3042         void *data;
3043
3044         if (!find_pack_entry(sha1, &e))
3045                 return NULL;
3046         data = cache_or_unpack_entry(e.p, e.offset, size, type);
3047         if (!data) {
3048                 /*
3049                  * We're probably in deep shit, but let's try to fetch
3050                  * the required object anyway from another pack or loose.
3051                  * This should happen only in the presence of a corrupted
3052                  * pack, and is better than failing outright.
3053                  */
3054                 error("failed to read object %s at offset %"PRIuMAX" from %s",
3055                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
3056                 mark_bad_packed_object(e.p, sha1);
3057                 data = read_object(sha1, type, size);
3058         }
3059         return data;
3060 }
3061
3062 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
3063                       unsigned char *sha1)
3064 {
3065         struct cached_object *co;
3066
3067         hash_sha1_file(buf, len, typename(type), sha1);
3068         if (has_sha1_file(sha1) || find_cached_object(sha1))
3069                 return 0;
3070         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
3071         co = &cached_objects[cached_object_nr++];
3072         co->size = len;
3073         co->type = type;
3074         co->buf = xmalloc(len);
3075         memcpy(co->buf, buf, len);
3076         hashcpy(co->sha1, sha1);
3077         return 0;
3078 }
3079
3080 static void *read_object(const unsigned char *sha1, enum object_type *type,
3081                          unsigned long *size)
3082 {
3083         unsigned long mapsize;
3084         void *map, *buf;
3085         struct cached_object *co;
3086
3087         co = find_cached_object(sha1);
3088         if (co) {
3089                 *type = co->type;
3090                 *size = co->size;
3091                 return xmemdupz(co->buf, co->size);
3092         }
3093
3094         buf = read_packed_sha1(sha1, type, size);
3095         if (buf)
3096                 return buf;
3097         map = map_sha1_file(sha1, &mapsize);
3098         if (map) {
3099                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
3100                 munmap(map, mapsize);
3101                 return buf;
3102         }
3103         reprepare_packed_git();
3104         return read_packed_sha1(sha1, type, size);
3105 }
3106
3107 /*
3108  * This function dies on corrupt objects; the callers who want to
3109  * deal with them should arrange to call read_object() and give error
3110  * messages themselves.
3111  */
3112 void *read_sha1_file_extended(const unsigned char *sha1,
3113                               enum object_type *type,
3114                               unsigned long *size,
3115                               unsigned flag)
3116 {
3117         void *data;
3118         const struct packed_git *p;
3119         const char *path;
3120         struct stat st;
3121         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
3122
3123         errno = 0;
3124         data = read_object(repl, type, size);
3125         if (data)
3126                 return data;
3127
3128         if (errno && errno != ENOENT)
3129                 die_errno("failed to read object %s", sha1_to_hex(sha1));
3130
3131         /* die if we replaced an object with one that does not exist */
3132         if (repl != sha1)
3133                 die("replacement %s not found for %s",
3134                     sha1_to_hex(repl), sha1_to_hex(sha1));
3135
3136         if (!stat_sha1_file(repl, &st, &path))
3137                 die("loose object %s (stored in %s) is corrupt",
3138                     sha1_to_hex(repl), path);
3139
3140         if ((p = has_packed_and_bad(repl)) != NULL)
3141                 die("packed object %s (stored in %s) is corrupt",
3142                     sha1_to_hex(repl), p->pack_name);
3143
3144         return NULL;
3145 }
3146
3147 void *read_object_with_reference(const unsigned char *sha1,
3148                                  const char *required_type_name,
3149                                  unsigned long *size,
3150                                  unsigned char *actual_sha1_return)
3151 {
3152         enum object_type type, required_type;
3153         void *buffer;
3154         unsigned long isize;
3155         unsigned char actual_sha1[20];
3156
3157         required_type = type_from_string(required_type_name);
3158         hashcpy(actual_sha1, sha1);
3159         while (1) {
3160                 int ref_length = -1;
3161                 const char *ref_type = NULL;
3162
3163                 buffer = read_sha1_file(actual_sha1, &type, &isize);
3164                 if (!buffer)
3165                         return NULL;
3166                 if (type == required_type) {
3167                         *size = isize;
3168                         if (actual_sha1_return)
3169                                 hashcpy(actual_sha1_return, actual_sha1);
3170                         return buffer;
3171                 }
3172                 /* Handle references */
3173                 else if (type == OBJ_COMMIT)
3174                         ref_type = "tree ";
3175                 else if (type == OBJ_TAG)
3176                         ref_type = "object ";
3177                 else {
3178                         free(buffer);
3179                         return NULL;
3180                 }
3181                 ref_length = strlen(ref_type);
3182
3183                 if (ref_length + 40 > isize ||
3184                     memcmp(buffer, ref_type, ref_length) ||
3185                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
3186                         free(buffer);
3187                         return NULL;
3188                 }
3189                 free(buffer);
3190                 /* Now we have the ID of the referred-to object in
3191                  * actual_sha1.  Check again. */
3192         }
3193 }
3194
3195 static void write_sha1_file_prepare(const void *buf, unsigned long len,
3196                                     const char *type, unsigned char *sha1,
3197                                     char *hdr, int *hdrlen)
3198 {
3199         git_SHA_CTX c;
3200
3201         /* Generate the header */
3202         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
3203
3204         /* Sha1.. */
3205         git_SHA1_Init(&c);
3206         git_SHA1_Update(&c, hdr, *hdrlen);
3207         git_SHA1_Update(&c, buf, len);
3208         git_SHA1_Final(sha1, &c);
3209 }
3210
3211 /*
3212  * Move the just written object into its final resting place.
3213  */
3214 int finalize_object_file(const char *tmpfile, const char *filename)
3215 {
3216         int ret = 0;
3217
3218         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3219                 goto try_rename;
3220         else if (link(tmpfile, filename))
3221                 ret = errno;
3222
3223         /*
3224          * Coda hack - coda doesn't like cross-directory links,
3225          * so we fall back to a rename, which will mean that it
3226          * won't be able to check collisions, but that's not a
3227          * big deal.
3228          *
3229          * The same holds for FAT formatted media.
3230          *
3231          * When this succeeds, we just return.  We have nothing
3232          * left to unlink.
3233          */
3234         if (ret && ret != EEXIST) {
3235         try_rename:
3236                 if (!rename(tmpfile, filename))
3237                         goto out;
3238                 ret = errno;
3239         }
3240         unlink_or_warn(tmpfile);
3241         if (ret) {
3242                 if (ret != EEXIST) {
3243                         return error_errno("unable to write sha1 filename %s", filename);
3244                 }
3245                 /* FIXME!!! Collision check here ? */
3246         }
3247
3248 out:
3249         if (adjust_shared_perm(filename))
3250                 return error("unable to set permission to '%s'", filename);
3251         return 0;
3252 }
3253
3254 static int write_buffer(int fd, const void *buf, size_t len)
3255 {
3256         if (write_in_full(fd, buf, len) < 0)
3257                 return error_errno("file write error");
3258         return 0;
3259 }
3260
3261 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3262                    unsigned char *sha1)
3263 {
3264         char hdr[32];
3265         int hdrlen = sizeof(hdr);
3266         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3267         return 0;
3268 }
3269
3270 /* Finalize a file on disk, and close it. */
3271 static void close_sha1_file(int fd)
3272 {
3273         if (fsync_object_files)
3274                 fsync_or_die(fd, "sha1 file");
3275         if (close(fd) != 0)
3276                 die_errno("error when closing sha1 file");
3277 }
3278
3279 /* Size of directory component, including the ending '/' */
3280 static inline int directory_size(const char *filename)
3281 {
3282         const char *s = strrchr(filename, '/');
3283         if (!s)
3284                 return 0;
3285         return s - filename + 1;
3286 }
3287
3288 /*
3289  * This creates a temporary file in the same directory as the final
3290  * 'filename'
3291  *
3292  * We want to avoid cross-directory filename renames, because those
3293  * can have problems on various filesystems (FAT, NFS, Coda).
3294  */
3295 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3296 {
3297         int fd, dirlen = directory_size(filename);
3298
3299         strbuf_reset(tmp);
3300         strbuf_add(tmp, filename, dirlen);
3301         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3302         fd = git_mkstemp_mode(tmp->buf, 0444);
3303         if (fd < 0 && dirlen && errno == ENOENT) {
3304                 /*
3305                  * Make sure the directory exists; note that the contents
3306                  * of the buffer are undefined after mkstemp returns an
3307                  * error, so we have to rewrite the whole buffer from
3308                  * scratch.
3309                  */
3310                 strbuf_reset(tmp);
3311                 strbuf_add(tmp, filename, dirlen - 1);
3312                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3313                         return -1;
3314                 if (adjust_shared_perm(tmp->buf))
3315                         return -1;
3316
3317                 /* Try again */
3318                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3319                 fd = git_mkstemp_mode(tmp->buf, 0444);
3320         }
3321         return fd;
3322 }
3323
3324 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3325                               const void *buf, unsigned long len, time_t mtime)
3326 {
3327         int fd, ret;
3328         unsigned char compressed[4096];
3329         git_zstream stream;
3330         git_SHA_CTX c;
3331         unsigned char parano_sha1[20];
3332         static struct strbuf tmp_file = STRBUF_INIT;
3333         const char *filename = sha1_file_name(sha1);
3334
3335         fd = create_tmpfile(&tmp_file, filename);
3336         if (fd < 0) {
3337                 if (errno == EACCES)
3338                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3339                 else
3340                         return error_errno("unable to create temporary file");
3341         }
3342
3343         /* Set it up */
3344         git_deflate_init(&stream, zlib_compression_level);
3345         stream.next_out = compressed;
3346         stream.avail_out = sizeof(compressed);
3347         git_SHA1_Init(&c);
3348
3349         /* First header.. */
3350         stream.next_in = (unsigned char *)hdr;
3351         stream.avail_in = hdrlen;
3352         while (git_deflate(&stream, 0) == Z_OK)
3353                 ; /* nothing */
3354         git_SHA1_Update(&c, hdr, hdrlen);
3355
3356         /* Then the data itself.. */
3357         stream.next_in = (void *)buf;
3358         stream.avail_in = len;
3359         do {
3360                 unsigned char *in0 = stream.next_in;
3361                 ret = git_deflate(&stream, Z_FINISH);
3362                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3363                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3364                         die("unable to write sha1 file");
3365                 stream.next_out = compressed;
3366                 stream.avail_out = sizeof(compressed);
3367         } while (ret == Z_OK);
3368
3369         if (ret != Z_STREAM_END)
3370                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3371         ret = git_deflate_end_gently(&stream);
3372         if (ret != Z_OK)
3373                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3374         git_SHA1_Final(parano_sha1, &c);
3375         if (hashcmp(sha1, parano_sha1) != 0)
3376                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3377
3378         close_sha1_file(fd);
3379
3380         if (mtime) {
3381                 struct utimbuf utb;
3382                 utb.actime = mtime;
3383                 utb.modtime = mtime;
3384                 if (utime(tmp_file.buf, &utb) < 0)
3385                         warning_errno("failed utime() on %s", tmp_file.buf);
3386         }
3387
3388         return finalize_object_file(tmp_file.buf, filename);
3389 }
3390
3391 static int freshen_loose_object(const unsigned char *sha1)
3392 {
3393         return check_and_freshen(sha1, 1);
3394 }
3395
3396 static int freshen_packed_object(const unsigned char *sha1)
3397 {
3398         struct pack_entry e;
3399         if (!find_pack_entry(sha1, &e))
3400                 return 0;
3401         if (e.p->freshened)
3402                 return 1;
3403         if (!freshen_file(e.p->pack_name))
3404                 return 0;
3405         e.p->freshened = 1;
3406         return 1;
3407 }
3408
3409 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3410 {
3411         char hdr[32];
3412         int hdrlen = sizeof(hdr);
3413
3414         /* Normally if we have it in the pack then we do not bother writing
3415          * it out into .git/objects/??/?{38} file.
3416          */
3417         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3418         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3419                 return 0;
3420         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3421 }
3422
3423 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3424                              unsigned char *sha1, unsigned flags)
3425 {
3426         char *header;
3427         int hdrlen, status = 0;
3428
3429         /* type string, SP, %lu of the length plus NUL must fit this */
3430         hdrlen = strlen(type) + 32;
3431         header = xmalloc(hdrlen);
3432         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3433
3434         if (!(flags & HASH_WRITE_OBJECT))
3435                 goto cleanup;
3436         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3437                 goto cleanup;
3438         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3439
3440 cleanup:
3441         free(header);
3442         return status;
3443 }
3444
3445 int force_object_loose(const unsigned char *sha1, time_t mtime)
3446 {
3447         void *buf;
3448         unsigned long len;
3449         enum object_type type;
3450         char hdr[32];
3451         int hdrlen;
3452         int ret;
3453
3454         if (has_loose_object(sha1))
3455                 return 0;
3456         buf = read_packed_sha1(sha1, &type, &len);
3457         if (!buf)
3458                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3459         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3460         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3461         free(buf);
3462
3463         return ret;
3464 }
3465
3466 int has_pack_index(const unsigned char *sha1)
3467 {
3468         struct stat st;
3469         if (stat(sha1_pack_index_name(sha1), &st))
3470                 return 0;
3471         return 1;
3472 }
3473
3474 int has_sha1_pack(const unsigned char *sha1)
3475 {
3476         struct pack_entry e;
3477         return find_pack_entry(sha1, &e);
3478 }
3479
3480 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3481 {
3482         struct pack_entry e;
3483
3484         if (!startup_info->have_repository)
3485                 return 0;
3486         if (find_pack_entry(sha1, &e))
3487                 return 1;
3488         if (has_loose_object(sha1))
3489                 return 1;
3490         if (flags & HAS_SHA1_QUICK)
3491                 return 0;
3492         reprepare_packed_git();
3493         return find_pack_entry(sha1, &e);
3494 }
3495
3496 int has_object_file(const struct object_id *oid)
3497 {
3498         return has_sha1_file(oid->hash);
3499 }
3500
3501 int has_object_file_with_flags(const struct object_id *oid, int flags)
3502 {
3503         return has_sha1_file_with_flags(oid->hash, flags);
3504 }
3505
3506 static void check_tree(const void *buf, size_t size)
3507 {
3508         struct tree_desc desc;
3509         struct name_entry entry;
3510
3511         init_tree_desc(&desc, buf, size);
3512         while (tree_entry(&desc, &entry))
3513                 /* do nothing
3514                  * tree_entry() will die() on malformed entries */
3515                 ;
3516 }
3517
3518 static void check_commit(const void *buf, size_t size)
3519 {
3520         struct commit c;
3521         memset(&c, 0, sizeof(c));
3522         if (parse_commit_buffer(&c, buf, size))
3523                 die("corrupt commit");
3524 }
3525
3526 static void check_tag(const void *buf, size_t size)
3527 {
3528         struct tag t;
3529         memset(&t, 0, sizeof(t));
3530         if (parse_tag_buffer(&t, buf, size))
3531                 die("corrupt tag");
3532 }
3533
3534 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3535                      enum object_type type,
3536                      const char *path, unsigned flags)
3537 {
3538         int ret, re_allocated = 0;
3539         int write_object = flags & HASH_WRITE_OBJECT;
3540
3541         if (!type)
3542                 type = OBJ_BLOB;
3543
3544         /*
3545          * Convert blobs to git internal format
3546          */
3547         if ((type == OBJ_BLOB) && path) {
3548                 struct strbuf nbuf = STRBUF_INIT;
3549                 if (convert_to_git(path, buf, size, &nbuf,
3550                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3551                         buf = strbuf_detach(&nbuf, &size);
3552                         re_allocated = 1;
3553                 }
3554         }
3555         if (flags & HASH_FORMAT_CHECK) {
3556                 if (type == OBJ_TREE)
3557                         check_tree(buf, size);
3558                 if (type == OBJ_COMMIT)
3559                         check_commit(buf, size);
3560                 if (type == OBJ_TAG)
3561                         check_tag(buf, size);
3562         }
3563
3564         if (write_object)
3565                 ret = write_sha1_file(buf, size, typename(type), sha1);
3566         else
3567                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3568         if (re_allocated)
3569                 free(buf);
3570         return ret;
3571 }
3572
3573 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3574                                      const char *path, unsigned flags)
3575 {
3576         int ret;
3577         const int write_object = flags & HASH_WRITE_OBJECT;
3578         struct strbuf sbuf = STRBUF_INIT;
3579
3580         assert(path);
3581         assert(would_convert_to_git_filter_fd(path));
3582
3583         convert_to_git_filter_fd(path, fd, &sbuf,
3584                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3585
3586         if (write_object)
3587                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3588                                       sha1);
3589         else
3590                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3591                                      sha1);
3592         strbuf_release(&sbuf);
3593         return ret;
3594 }
3595
3596 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3597                       const char *path, unsigned flags)
3598 {
3599         struct strbuf sbuf = STRBUF_INIT;
3600         int ret;
3601
3602         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3603                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3604         else
3605                 ret = -1;
3606         strbuf_release(&sbuf);
3607         return ret;
3608 }
3609
3610 #define SMALL_FILE_SIZE (32*1024)
3611
3612 static int index_core(unsigned char *sha1, int fd, size_t size,
3613                       enum object_type type, const char *path,
3614                       unsigned flags)
3615 {
3616         int ret;
3617
3618         if (!size) {
3619                 ret = index_mem(sha1, "", size, type, path, flags);
3620         } else if (size <= SMALL_FILE_SIZE) {
3621                 char *buf = xmalloc(size);
3622                 if (size == read_in_full(fd, buf, size))
3623                         ret = index_mem(sha1, buf, size, type, path, flags);
3624                 else
3625                         ret = error_errno("short read");
3626                 free(buf);
3627         } else {
3628                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3629                 ret = index_mem(sha1, buf, size, type, path, flags);
3630                 munmap(buf, size);
3631         }
3632         return ret;
3633 }
3634
3635 /*
3636  * This creates one packfile per large blob unless bulk-checkin
3637  * machinery is "plugged".
3638  *
3639  * This also bypasses the usual "convert-to-git" dance, and that is on
3640  * purpose. We could write a streaming version of the converting
3641  * functions and insert that before feeding the data to fast-import
3642  * (or equivalent in-core API described above). However, that is
3643  * somewhat complicated, as we do not know the size of the filter
3644  * result, which we need to know beforehand when writing a git object.
3645  * Since the primary motivation for trying to stream from the working
3646  * tree file and to avoid mmaping it in core is to deal with large
3647  * binary blobs, they generally do not want to get any conversion, and
3648  * callers should avoid this code path when filters are requested.
3649  */
3650 static int index_stream(unsigned char *sha1, int fd, size_t size,
3651                         enum object_type type, const char *path,
3652                         unsigned flags)
3653 {
3654         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3655 }
3656
3657 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3658              enum object_type type, const char *path, unsigned flags)
3659 {
3660         int ret;
3661
3662         /*
3663          * Call xsize_t() only when needed to avoid potentially unnecessary
3664          * die() for large files.
3665          */
3666         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3667                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3668         else if (!S_ISREG(st->st_mode))
3669                 ret = index_pipe(sha1, fd, type, path, flags);
3670         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3671                  (path && would_convert_to_git(path)))
3672                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3673                                  flags);
3674         else
3675                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3676                                    flags);
3677         close(fd);
3678         return ret;
3679 }
3680
3681 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3682 {
3683         int fd;
3684         struct strbuf sb = STRBUF_INIT;
3685
3686         switch (st->st_mode & S_IFMT) {
3687         case S_IFREG:
3688                 fd = open(path, O_RDONLY);
3689                 if (fd < 0)
3690                         return error_errno("open(\"%s\")", path);
3691                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3692                         return error("%s: failed to insert into database",
3693                                      path);
3694                 break;
3695         case S_IFLNK:
3696                 if (strbuf_readlink(&sb, path, st->st_size))
3697                         return error_errno("readlink(\"%s\")", path);
3698                 if (!(flags & HASH_WRITE_OBJECT))
3699                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3700                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3701                         return error("%s: failed to insert into database",
3702                                      path);
3703                 strbuf_release(&sb);
3704                 break;
3705         case S_IFDIR:
3706                 return resolve_gitlink_ref(path, "HEAD", sha1);
3707         default:
3708                 return error("%s: unsupported file type", path);
3709         }
3710         return 0;
3711 }
3712
3713 int read_pack_header(int fd, struct pack_header *header)
3714 {
3715         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3716                 /* "eof before pack header was fully read" */
3717                 return PH_ERROR_EOF;
3718
3719         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3720                 /* "protocol error (pack signature mismatch detected)" */
3721                 return PH_ERROR_PACK_SIGNATURE;
3722         if (!pack_version_ok(header->hdr_version))
3723                 /* "protocol error (pack version unsupported)" */
3724                 return PH_ERROR_PROTOCOL;
3725         return 0;
3726 }
3727
3728 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3729 {
3730         enum object_type type = sha1_object_info(sha1, NULL);
3731         if (type < 0)
3732                 die("%s is not a valid object", sha1_to_hex(sha1));
3733         if (type != expect)
3734                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3735                     typename(expect));
3736 }
3737
3738 static int for_each_file_in_obj_subdir(int subdir_nr,
3739                                        struct strbuf *path,
3740                                        each_loose_object_fn obj_cb,
3741                                        each_loose_cruft_fn cruft_cb,
3742                                        each_loose_subdir_fn subdir_cb,
3743                                        void *data)
3744 {
3745         size_t baselen = path->len;
3746         DIR *dir = opendir(path->buf);
3747         struct dirent *de;
3748         int r = 0;
3749
3750         if (!dir) {
3751                 if (errno == ENOENT)
3752                         return 0;
3753                 return error_errno("unable to open %s", path->buf);
3754         }
3755
3756         while ((de = readdir(dir))) {
3757                 if (is_dot_or_dotdot(de->d_name))
3758                         continue;
3759
3760                 strbuf_setlen(path, baselen);
3761                 strbuf_addf(path, "/%s", de->d_name);
3762
3763                 if (strlen(de->d_name) == GIT_SHA1_HEXSZ - 2)  {
3764                         char hex[GIT_MAX_HEXSZ+1];
3765                         struct object_id oid;
3766
3767                         xsnprintf(hex, sizeof(hex), "%02x%s",
3768                                   subdir_nr, de->d_name);
3769                         if (!get_oid_hex(hex, &oid)) {
3770                                 if (obj_cb) {
3771                                         r = obj_cb(&oid, path->buf, data);
3772                                         if (r)
3773                                                 break;
3774                                 }
3775                                 continue;
3776                         }
3777                 }
3778
3779                 if (cruft_cb) {
3780                         r = cruft_cb(de->d_name, path->buf, data);
3781                         if (r)
3782                                 break;
3783                 }
3784         }
3785         closedir(dir);
3786
3787         strbuf_setlen(path, baselen);
3788         if (!r && subdir_cb)
3789                 r = subdir_cb(subdir_nr, path->buf, data);
3790
3791         return r;
3792 }
3793
3794 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3795                             each_loose_object_fn obj_cb,
3796                             each_loose_cruft_fn cruft_cb,
3797                             each_loose_subdir_fn subdir_cb,
3798                             void *data)
3799 {
3800         size_t baselen = path->len;
3801         int r = 0;
3802         int i;
3803
3804         for (i = 0; i < 256; i++) {
3805                 strbuf_addf(path, "/%02x", i);
3806                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3807                                                 subdir_cb, data);
3808                 strbuf_setlen(path, baselen);
3809                 if (r)
3810                         break;
3811         }
3812
3813         return r;
3814 }
3815
3816 int for_each_loose_file_in_objdir(const char *path,
3817                                   each_loose_object_fn obj_cb,
3818                                   each_loose_cruft_fn cruft_cb,
3819                                   each_loose_subdir_fn subdir_cb,
3820                                   void *data)
3821 {
3822         struct strbuf buf = STRBUF_INIT;
3823         int r;
3824
3825         strbuf_addstr(&buf, path);
3826         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3827                                               subdir_cb, data);
3828         strbuf_release(&buf);
3829
3830         return r;
3831 }
3832
3833 struct loose_alt_odb_data {
3834         each_loose_object_fn *cb;
3835         void *data;
3836 };
3837
3838 static int loose_from_alt_odb(struct alternate_object_database *alt,
3839                               void *vdata)
3840 {
3841         struct loose_alt_odb_data *data = vdata;
3842         struct strbuf buf = STRBUF_INIT;
3843         int r;
3844
3845         strbuf_addstr(&buf, alt->path);
3846         r = for_each_loose_file_in_objdir_buf(&buf,
3847                                               data->cb, NULL, NULL,
3848                                               data->data);
3849         strbuf_release(&buf);
3850         return r;
3851 }
3852
3853 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3854 {
3855         struct loose_alt_odb_data alt;
3856         int r;
3857
3858         r = for_each_loose_file_in_objdir(get_object_directory(),
3859                                           cb, NULL, NULL, data);
3860         if (r)
3861                 return r;
3862
3863         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3864                 return 0;
3865
3866         alt.cb = cb;
3867         alt.data = data;
3868         return foreach_alt_odb(loose_from_alt_odb, &alt);
3869 }
3870
3871 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3872 {
3873         uint32_t i;
3874         int r = 0;
3875
3876         for (i = 0; i < p->num_objects; i++) {
3877                 struct object_id oid;
3878
3879                 if (!nth_packed_object_oid(&oid, p, i))
3880                         return error("unable to get sha1 of object %u in %s",
3881                                      i, p->pack_name);
3882
3883                 r = cb(&oid, p, i, data);
3884                 if (r)
3885                         break;
3886         }
3887         return r;
3888 }
3889
3890 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3891 {
3892         struct packed_git *p;
3893         int r = 0;
3894         int pack_errors = 0;
3895
3896         prepare_packed_git();
3897         for (p = packed_git; p; p = p->next) {
3898                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3899                         continue;
3900                 if (open_pack_index(p)) {
3901                         pack_errors = 1;
3902                         continue;
3903                 }
3904                 r = for_each_object_in_pack(p, cb, data);
3905                 if (r)
3906                         break;
3907         }
3908         return r ? r : pack_errors;
3909 }
3910
3911 static int check_stream_sha1(git_zstream *stream,
3912                              const char *hdr,
3913                              unsigned long size,
3914                              const char *path,
3915                              const unsigned char *expected_sha1)
3916 {
3917         git_SHA_CTX c;
3918         unsigned char real_sha1[GIT_MAX_RAWSZ];
3919         unsigned char buf[4096];
3920         unsigned long total_read;
3921         int status = Z_OK;
3922
3923         git_SHA1_Init(&c);
3924         git_SHA1_Update(&c, hdr, stream->total_out);
3925
3926         /*
3927          * We already read some bytes into hdr, but the ones up to the NUL
3928          * do not count against the object's content size.
3929          */
3930         total_read = stream->total_out - strlen(hdr) - 1;
3931
3932         /*
3933          * This size comparison must be "<=" to read the final zlib packets;
3934          * see the comment in unpack_sha1_rest for details.
3935          */
3936         while (total_read <= size &&
3937                (status == Z_OK || status == Z_BUF_ERROR)) {
3938                 stream->next_out = buf;
3939                 stream->avail_out = sizeof(buf);
3940                 if (size - total_read < stream->avail_out)
3941                         stream->avail_out = size - total_read;
3942                 status = git_inflate(stream, Z_FINISH);
3943                 git_SHA1_Update(&c, buf, stream->next_out - buf);
3944                 total_read += stream->next_out - buf;
3945         }
3946         git_inflate_end(stream);
3947
3948         if (status != Z_STREAM_END) {
3949                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
3950                 return -1;
3951         }
3952         if (stream->avail_in) {
3953                 error("garbage at end of loose object '%s'",
3954                       sha1_to_hex(expected_sha1));
3955                 return -1;
3956         }
3957
3958         git_SHA1_Final(real_sha1, &c);
3959         if (hashcmp(expected_sha1, real_sha1)) {
3960                 error("sha1 mismatch for %s (expected %s)", path,
3961                       sha1_to_hex(expected_sha1));
3962                 return -1;
3963         }
3964
3965         return 0;
3966 }
3967
3968 int read_loose_object(const char *path,
3969                       const unsigned char *expected_sha1,
3970                       enum object_type *type,
3971                       unsigned long *size,
3972                       void **contents)
3973 {
3974         int ret = -1;
3975         void *map = NULL;
3976         unsigned long mapsize;
3977         git_zstream stream;
3978         char hdr[32];
3979
3980         *contents = NULL;
3981
3982         map = map_sha1_file_1(path, NULL, &mapsize);
3983         if (!map) {
3984                 error_errno("unable to mmap %s", path);
3985                 goto out;
3986         }
3987
3988         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
3989                 error("unable to unpack header of %s", path);
3990                 goto out;
3991         }
3992
3993         *type = parse_sha1_header(hdr, size);
3994         if (*type < 0) {
3995                 error("unable to parse header of %s", path);
3996                 git_inflate_end(&stream);
3997                 goto out;
3998         }
3999
4000         if (*type == OBJ_BLOB) {
4001                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
4002                         goto out;
4003         } else {
4004                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
4005                 if (!*contents) {
4006                         error("unable to unpack contents of %s", path);
4007                         git_inflate_end(&stream);
4008                         goto out;
4009                 }
4010                 if (check_sha1_signature(expected_sha1, *contents,
4011                                          *size, typename(*type))) {
4012                         error("sha1 mismatch for %s (expected %s)", path,
4013                               sha1_to_hex(expected_sha1));
4014                         free(*contents);
4015                         goto out;
4016                 }
4017         }
4018
4019         ret = 0; /* everything checks out */
4020
4021 out:
4022         if (map)
4023                 munmap(map, mapsize);
4024         return ret;
4025 }