fast-import: checkpoint: dump branches/tags/marks even if object_count==0
[git] / sha1_file.c
1 /*
2  * GIT - The information manager from hell
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This handles basic git sha1 object files - packing, unpacking,
7  * creation etc.
8  */
9 #include "cache.h"
10 #include "config.h"
11 #include "string-list.h"
12 #include "lockfile.h"
13 #include "delta.h"
14 #include "pack.h"
15 #include "blob.h"
16 #include "commit.h"
17 #include "run-command.h"
18 #include "tag.h"
19 #include "tree.h"
20 #include "tree-walk.h"
21 #include "refs.h"
22 #include "pack-revindex.h"
23 #include "sha1-lookup.h"
24 #include "bulk-checkin.h"
25 #include "streaming.h"
26 #include "dir.h"
27 #include "mru.h"
28 #include "list.h"
29 #include "mergesort.h"
30 #include "quote.h"
31
32 #define SZ_FMT PRIuMAX
33 static inline uintmax_t sz_fmt(size_t s) { return s; }
34
35 const unsigned char null_sha1[20];
36 const struct object_id null_oid;
37 const struct object_id empty_tree_oid = {
38         EMPTY_TREE_SHA1_BIN_LITERAL
39 };
40 const struct object_id empty_blob_oid = {
41         EMPTY_BLOB_SHA1_BIN_LITERAL
42 };
43
44 /*
45  * This is meant to hold a *small* number of objects that you would
46  * want read_sha1_file() to be able to return, but yet you do not want
47  * to write them into the object store (e.g. a browse-only
48  * application).
49  */
50 static struct cached_object {
51         unsigned char sha1[20];
52         enum object_type type;
53         void *buf;
54         unsigned long size;
55 } *cached_objects;
56 static int cached_object_nr, cached_object_alloc;
57
58 static struct cached_object empty_tree = {
59         EMPTY_TREE_SHA1_BIN_LITERAL,
60         OBJ_TREE,
61         "",
62         0
63 };
64
65 static struct cached_object *find_cached_object(const unsigned char *sha1)
66 {
67         int i;
68         struct cached_object *co = cached_objects;
69
70         for (i = 0; i < cached_object_nr; i++, co++) {
71                 if (!hashcmp(co->sha1, sha1))
72                         return co;
73         }
74         if (!hashcmp(sha1, empty_tree.sha1))
75                 return &empty_tree;
76         return NULL;
77 }
78
79 int mkdir_in_gitdir(const char *path)
80 {
81         if (mkdir(path, 0777)) {
82                 int saved_errno = errno;
83                 struct stat st;
84                 struct strbuf sb = STRBUF_INIT;
85
86                 if (errno != EEXIST)
87                         return -1;
88                 /*
89                  * Are we looking at a path in a symlinked worktree
90                  * whose original repository does not yet have it?
91                  * e.g. .git/rr-cache pointing at its original
92                  * repository in which the user hasn't performed any
93                  * conflict resolution yet?
94                  */
95                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
96                     strbuf_readlink(&sb, path, st.st_size) ||
97                     !is_absolute_path(sb.buf) ||
98                     mkdir(sb.buf, 0777)) {
99                         strbuf_release(&sb);
100                         errno = saved_errno;
101                         return -1;
102                 }
103                 strbuf_release(&sb);
104         }
105         return adjust_shared_perm(path);
106 }
107
108 enum scld_error safe_create_leading_directories(char *path)
109 {
110         char *next_component = path + offset_1st_component(path);
111         enum scld_error ret = SCLD_OK;
112
113         while (ret == SCLD_OK && next_component) {
114                 struct stat st;
115                 char *slash = next_component, slash_character;
116
117                 while (*slash && !is_dir_sep(*slash))
118                         slash++;
119
120                 if (!*slash)
121                         break;
122
123                 next_component = slash + 1;
124                 while (is_dir_sep(*next_component))
125                         next_component++;
126                 if (!*next_component)
127                         break;
128
129                 slash_character = *slash;
130                 *slash = '\0';
131                 if (!stat(path, &st)) {
132                         /* path exists */
133                         if (!S_ISDIR(st.st_mode)) {
134                                 errno = ENOTDIR;
135                                 ret = SCLD_EXISTS;
136                         }
137                 } else if (mkdir(path, 0777)) {
138                         if (errno == EEXIST &&
139                             !stat(path, &st) && S_ISDIR(st.st_mode))
140                                 ; /* somebody created it since we checked */
141                         else if (errno == ENOENT)
142                                 /*
143                                  * Either mkdir() failed because
144                                  * somebody just pruned the containing
145                                  * directory, or stat() failed because
146                                  * the file that was in our way was
147                                  * just removed.  Either way, inform
148                                  * the caller that it might be worth
149                                  * trying again:
150                                  */
151                                 ret = SCLD_VANISHED;
152                         else
153                                 ret = SCLD_FAILED;
154                 } else if (adjust_shared_perm(path)) {
155                         ret = SCLD_PERMS;
156                 }
157                 *slash = slash_character;
158         }
159         return ret;
160 }
161
162 enum scld_error safe_create_leading_directories_const(const char *path)
163 {
164         int save_errno;
165         /* path points to cache entries, so xstrdup before messing with it */
166         char *buf = xstrdup(path);
167         enum scld_error result = safe_create_leading_directories(buf);
168
169         save_errno = errno;
170         free(buf);
171         errno = save_errno;
172         return result;
173 }
174
175 int raceproof_create_file(const char *path, create_file_fn fn, void *cb)
176 {
177         /*
178          * The number of times we will try to remove empty directories
179          * in the way of path. This is only 1 because if another
180          * process is racily creating directories that conflict with
181          * us, we don't want to fight against them.
182          */
183         int remove_directories_remaining = 1;
184
185         /*
186          * The number of times that we will try to create the
187          * directories containing path. We are willing to attempt this
188          * more than once, because another process could be trying to
189          * clean up empty directories at the same time as we are
190          * trying to create them.
191          */
192         int create_directories_remaining = 3;
193
194         /* A scratch copy of path, filled lazily if we need it: */
195         struct strbuf path_copy = STRBUF_INIT;
196
197         int ret, save_errno;
198
199         /* Sanity check: */
200         assert(*path);
201
202 retry_fn:
203         ret = fn(path, cb);
204         save_errno = errno;
205         if (!ret)
206                 goto out;
207
208         if (errno == EISDIR && remove_directories_remaining-- > 0) {
209                 /*
210                  * A directory is in the way. Maybe it is empty; try
211                  * to remove it:
212                  */
213                 if (!path_copy.len)
214                         strbuf_addstr(&path_copy, path);
215
216                 if (!remove_dir_recursively(&path_copy, REMOVE_DIR_EMPTY_ONLY))
217                         goto retry_fn;
218         } else if (errno == ENOENT && create_directories_remaining-- > 0) {
219                 /*
220                  * Maybe the containing directory didn't exist, or
221                  * maybe it was just deleted by a process that is
222                  * racing with us to clean up empty directories. Try
223                  * to create it:
224                  */
225                 enum scld_error scld_result;
226
227                 if (!path_copy.len)
228                         strbuf_addstr(&path_copy, path);
229
230                 do {
231                         scld_result = safe_create_leading_directories(path_copy.buf);
232                         if (scld_result == SCLD_OK)
233                                 goto retry_fn;
234                 } while (scld_result == SCLD_VANISHED && create_directories_remaining-- > 0);
235         }
236
237 out:
238         strbuf_release(&path_copy);
239         errno = save_errno;
240         return ret;
241 }
242
243 static void fill_sha1_path(struct strbuf *buf, const unsigned char *sha1)
244 {
245         int i;
246         for (i = 0; i < 20; i++) {
247                 static char hex[] = "0123456789abcdef";
248                 unsigned int val = sha1[i];
249                 strbuf_addch(buf, hex[val >> 4]);
250                 strbuf_addch(buf, hex[val & 0xf]);
251                 if (!i)
252                         strbuf_addch(buf, '/');
253         }
254 }
255
256 const char *sha1_file_name(const unsigned char *sha1)
257 {
258         static struct strbuf buf = STRBUF_INIT;
259
260         strbuf_reset(&buf);
261         strbuf_addf(&buf, "%s/", get_object_directory());
262
263         fill_sha1_path(&buf, sha1);
264         return buf.buf;
265 }
266
267 struct strbuf *alt_scratch_buf(struct alternate_object_database *alt)
268 {
269         strbuf_setlen(&alt->scratch, alt->base_len);
270         return &alt->scratch;
271 }
272
273 static const char *alt_sha1_path(struct alternate_object_database *alt,
274                                  const unsigned char *sha1)
275 {
276         struct strbuf *buf = alt_scratch_buf(alt);
277         fill_sha1_path(buf, sha1);
278         return buf->buf;
279 }
280
281  char *odb_pack_name(struct strbuf *buf,
282                      const unsigned char *sha1,
283                      const char *ext)
284 {
285         strbuf_reset(buf);
286         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
287                     sha1_to_hex(sha1), ext);
288         return buf->buf;
289 }
290
291 char *sha1_pack_name(const unsigned char *sha1)
292 {
293         static struct strbuf buf = STRBUF_INIT;
294         return odb_pack_name(&buf, sha1, "pack");
295 }
296
297 char *sha1_pack_index_name(const unsigned char *sha1)
298 {
299         static struct strbuf buf = STRBUF_INIT;
300         return odb_pack_name(&buf, sha1, "idx");
301 }
302
303 struct alternate_object_database *alt_odb_list;
304 static struct alternate_object_database **alt_odb_tail;
305
306 /*
307  * Return non-zero iff the path is usable as an alternate object database.
308  */
309 static int alt_odb_usable(struct strbuf *path, const char *normalized_objdir)
310 {
311         struct alternate_object_database *alt;
312
313         /* Detect cases where alternate disappeared */
314         if (!is_directory(path->buf)) {
315                 error("object directory %s does not exist; "
316                       "check .git/objects/info/alternates.",
317                       path->buf);
318                 return 0;
319         }
320
321         /*
322          * Prevent the common mistake of listing the same
323          * thing twice, or object directory itself.
324          */
325         for (alt = alt_odb_list; alt; alt = alt->next) {
326                 if (!fspathcmp(path->buf, alt->path))
327                         return 0;
328         }
329         if (!fspathcmp(path->buf, normalized_objdir))
330                 return 0;
331
332         return 1;
333 }
334
335 /*
336  * Prepare alternate object database registry.
337  *
338  * The variable alt_odb_list points at the list of struct
339  * alternate_object_database.  The elements on this list come from
340  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
341  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
342  * whose contents is similar to that environment variable but can be
343  * LF separated.  Its base points at a statically allocated buffer that
344  * contains "/the/directory/corresponding/to/.git/objects/...", while
345  * its name points just after the slash at the end of ".git/objects/"
346  * in the example above, and has enough space to hold 40-byte hex
347  * SHA1, an extra slash for the first level indirection, and the
348  * terminating NUL.
349  */
350 static void read_info_alternates(const char * relative_base, int depth);
351 static int link_alt_odb_entry(const char *entry, const char *relative_base,
352         int depth, const char *normalized_objdir)
353 {
354         struct alternate_object_database *ent;
355         struct strbuf pathbuf = STRBUF_INIT;
356
357         if (!is_absolute_path(entry) && relative_base) {
358                 strbuf_realpath(&pathbuf, relative_base, 1);
359                 strbuf_addch(&pathbuf, '/');
360         }
361         strbuf_addstr(&pathbuf, entry);
362
363         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
364                 error("unable to normalize alternate object path: %s",
365                       pathbuf.buf);
366                 strbuf_release(&pathbuf);
367                 return -1;
368         }
369
370         /*
371          * The trailing slash after the directory name is given by
372          * this function at the end. Remove duplicates.
373          */
374         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
375                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
376
377         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
378                 strbuf_release(&pathbuf);
379                 return -1;
380         }
381
382         ent = alloc_alt_odb(pathbuf.buf);
383
384         /* add the alternate entry */
385         *alt_odb_tail = ent;
386         alt_odb_tail = &(ent->next);
387         ent->next = NULL;
388
389         /* recursively add alternates */
390         read_info_alternates(pathbuf.buf, depth + 1);
391
392         strbuf_release(&pathbuf);
393         return 0;
394 }
395
396 static const char *parse_alt_odb_entry(const char *string,
397                                        int sep,
398                                        struct strbuf *out)
399 {
400         const char *end;
401
402         strbuf_reset(out);
403
404         if (*string == '#') {
405                 /* comment; consume up to next separator */
406                 end = strchrnul(string, sep);
407         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
408                 /*
409                  * quoted path; unquote_c_style has copied the
410                  * data for us and set "end". Broken quoting (e.g.,
411                  * an entry that doesn't end with a quote) falls
412                  * back to the unquoted case below.
413                  */
414         } else {
415                 /* normal, unquoted path */
416                 end = strchrnul(string, sep);
417                 strbuf_add(out, string, end - string);
418         }
419
420         if (*end)
421                 end++;
422         return end;
423 }
424
425 static void link_alt_odb_entries(const char *alt, int len, int sep,
426                                  const char *relative_base, int depth)
427 {
428         struct strbuf objdirbuf = STRBUF_INIT;
429         struct strbuf entry = STRBUF_INIT;
430
431         if (depth > 5) {
432                 error("%s: ignoring alternate object stores, nesting too deep.",
433                                 relative_base);
434                 return;
435         }
436
437         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
438         if (strbuf_normalize_path(&objdirbuf) < 0)
439                 die("unable to normalize object directory: %s",
440                     objdirbuf.buf);
441
442         while (*alt) {
443                 alt = parse_alt_odb_entry(alt, sep, &entry);
444                 if (!entry.len)
445                         continue;
446                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
447         }
448         strbuf_release(&entry);
449         strbuf_release(&objdirbuf);
450 }
451
452 static void read_info_alternates(const char * relative_base, int depth)
453 {
454         char *map;
455         size_t mapsz;
456         struct stat st;
457         char *path;
458         int fd;
459
460         path = xstrfmt("%s/info/alternates", relative_base);
461         fd = git_open(path);
462         free(path);
463         if (fd < 0)
464                 return;
465         if (fstat(fd, &st) || (st.st_size == 0)) {
466                 close(fd);
467                 return;
468         }
469         mapsz = xsize_t(st.st_size);
470         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
471         close(fd);
472
473         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
474
475         munmap(map, mapsz);
476 }
477
478 struct alternate_object_database *alloc_alt_odb(const char *dir)
479 {
480         struct alternate_object_database *ent;
481
482         FLEX_ALLOC_STR(ent, path, dir);
483         strbuf_init(&ent->scratch, 0);
484         strbuf_addf(&ent->scratch, "%s/", dir);
485         ent->base_len = ent->scratch.len;
486
487         return ent;
488 }
489
490 void add_to_alternates_file(const char *reference)
491 {
492         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
493         char *alts = git_pathdup("objects/info/alternates");
494         FILE *in, *out;
495
496         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
497         out = fdopen_lock_file(lock, "w");
498         if (!out)
499                 die_errno("unable to fdopen alternates lockfile");
500
501         in = fopen(alts, "r");
502         if (in) {
503                 struct strbuf line = STRBUF_INIT;
504                 int found = 0;
505
506                 while (strbuf_getline(&line, in) != EOF) {
507                         if (!strcmp(reference, line.buf)) {
508                                 found = 1;
509                                 break;
510                         }
511                         fprintf_or_die(out, "%s\n", line.buf);
512                 }
513
514                 strbuf_release(&line);
515                 fclose(in);
516
517                 if (found) {
518                         rollback_lock_file(lock);
519                         lock = NULL;
520                 }
521         }
522         else if (errno != ENOENT)
523                 die_errno("unable to read alternates file");
524
525         if (lock) {
526                 fprintf_or_die(out, "%s\n", reference);
527                 if (commit_lock_file(lock))
528                         die_errno("unable to move new alternates file into place");
529                 if (alt_odb_tail)
530                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
531         }
532         free(alts);
533 }
534
535 void add_to_alternates_memory(const char *reference)
536 {
537         /*
538          * Make sure alternates are initialized, or else our entry may be
539          * overwritten when they are.
540          */
541         prepare_alt_odb();
542
543         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
544 }
545
546 /*
547  * Compute the exact path an alternate is at and returns it. In case of
548  * error NULL is returned and the human readable error is added to `err`
549  * `path` may be relative and should point to $GITDIR.
550  * `err` must not be null.
551  */
552 char *compute_alternate_path(const char *path, struct strbuf *err)
553 {
554         char *ref_git = NULL;
555         const char *repo, *ref_git_s;
556         int seen_error = 0;
557
558         ref_git_s = real_path_if_valid(path);
559         if (!ref_git_s) {
560                 seen_error = 1;
561                 strbuf_addf(err, _("path '%s' does not exist"), path);
562                 goto out;
563         } else
564                 /*
565                  * Beware: read_gitfile(), real_path() and mkpath()
566                  * return static buffer
567                  */
568                 ref_git = xstrdup(ref_git_s);
569
570         repo = read_gitfile(ref_git);
571         if (!repo)
572                 repo = read_gitfile(mkpath("%s/.git", ref_git));
573         if (repo) {
574                 free(ref_git);
575                 ref_git = xstrdup(repo);
576         }
577
578         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
579                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
580                 free(ref_git);
581                 ref_git = ref_git_git;
582         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
583                 struct strbuf sb = STRBUF_INIT;
584                 seen_error = 1;
585                 if (get_common_dir(&sb, ref_git)) {
586                         strbuf_addf(err,
587                                     _("reference repository '%s' as a linked "
588                                       "checkout is not supported yet."),
589                                     path);
590                         goto out;
591                 }
592
593                 strbuf_addf(err, _("reference repository '%s' is not a "
594                                         "local repository."), path);
595                 goto out;
596         }
597
598         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
599                 strbuf_addf(err, _("reference repository '%s' is shallow"),
600                             path);
601                 seen_error = 1;
602                 goto out;
603         }
604
605         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
606                 strbuf_addf(err,
607                             _("reference repository '%s' is grafted"),
608                             path);
609                 seen_error = 1;
610                 goto out;
611         }
612
613 out:
614         if (seen_error) {
615                 FREE_AND_NULL(ref_git);
616         }
617
618         return ref_git;
619 }
620
621 int foreach_alt_odb(alt_odb_fn fn, void *cb)
622 {
623         struct alternate_object_database *ent;
624         int r = 0;
625
626         prepare_alt_odb();
627         for (ent = alt_odb_list; ent; ent = ent->next) {
628                 r = fn(ent, cb);
629                 if (r)
630                         break;
631         }
632         return r;
633 }
634
635 void prepare_alt_odb(void)
636 {
637         const char *alt;
638
639         if (alt_odb_tail)
640                 return;
641
642         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
643         if (!alt) alt = "";
644
645         alt_odb_tail = &alt_odb_list;
646         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
647
648         read_info_alternates(get_object_directory(), 0);
649 }
650
651 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
652 static int freshen_file(const char *fn)
653 {
654         struct utimbuf t;
655         t.actime = t.modtime = time(NULL);
656         return !utime(fn, &t);
657 }
658
659 /*
660  * All of the check_and_freshen functions return 1 if the file exists and was
661  * freshened (if freshening was requested), 0 otherwise. If they return
662  * 0, you should not assume that it is safe to skip a write of the object (it
663  * either does not exist on disk, or has a stale mtime and may be subject to
664  * pruning).
665  */
666 int check_and_freshen_file(const char *fn, int freshen)
667 {
668         if (access(fn, F_OK))
669                 return 0;
670         if (freshen && !freshen_file(fn))
671                 return 0;
672         return 1;
673 }
674
675 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
676 {
677         return check_and_freshen_file(sha1_file_name(sha1), freshen);
678 }
679
680 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
681 {
682         struct alternate_object_database *alt;
683         prepare_alt_odb();
684         for (alt = alt_odb_list; alt; alt = alt->next) {
685                 const char *path = alt_sha1_path(alt, sha1);
686                 if (check_and_freshen_file(path, freshen))
687                         return 1;
688         }
689         return 0;
690 }
691
692 static int check_and_freshen(const unsigned char *sha1, int freshen)
693 {
694         return check_and_freshen_local(sha1, freshen) ||
695                check_and_freshen_nonlocal(sha1, freshen);
696 }
697
698 int has_loose_object_nonlocal(const unsigned char *sha1)
699 {
700         return check_and_freshen_nonlocal(sha1, 0);
701 }
702
703 static int has_loose_object(const unsigned char *sha1)
704 {
705         return check_and_freshen(sha1, 0);
706 }
707
708 static unsigned int pack_used_ctr;
709 static unsigned int pack_mmap_calls;
710 static unsigned int peak_pack_open_windows;
711 static unsigned int pack_open_windows;
712 static unsigned int pack_open_fds;
713 static unsigned int pack_max_fds;
714 static size_t peak_pack_mapped;
715 static size_t pack_mapped;
716 struct packed_git *packed_git;
717
718 static struct mru packed_git_mru_storage;
719 struct mru *packed_git_mru = &packed_git_mru_storage;
720
721 void pack_report(void)
722 {
723         fprintf(stderr,
724                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
725                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
726                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
727                 sz_fmt(getpagesize()),
728                 sz_fmt(packed_git_window_size),
729                 sz_fmt(packed_git_limit));
730         fprintf(stderr,
731                 "pack_report: pack_used_ctr            = %10u\n"
732                 "pack_report: pack_mmap_calls          = %10u\n"
733                 "pack_report: pack_open_windows        = %10u / %10u\n"
734                 "pack_report: pack_mapped              = "
735                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
736                 pack_used_ctr,
737                 pack_mmap_calls,
738                 pack_open_windows, peak_pack_open_windows,
739                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
740 }
741
742 /*
743  * Open and mmap the index file at path, perform a couple of
744  * consistency checks, then record its information to p.  Return 0 on
745  * success.
746  */
747 static int check_packed_git_idx(const char *path, struct packed_git *p)
748 {
749         void *idx_map;
750         struct pack_idx_header *hdr;
751         size_t idx_size;
752         uint32_t version, nr, i, *index;
753         int fd = git_open(path);
754         struct stat st;
755
756         if (fd < 0)
757                 return -1;
758         if (fstat(fd, &st)) {
759                 close(fd);
760                 return -1;
761         }
762         idx_size = xsize_t(st.st_size);
763         if (idx_size < 4 * 256 + 20 + 20) {
764                 close(fd);
765                 return error("index file %s is too small", path);
766         }
767         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
768         close(fd);
769
770         hdr = idx_map;
771         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
772                 version = ntohl(hdr->idx_version);
773                 if (version < 2 || version > 2) {
774                         munmap(idx_map, idx_size);
775                         return error("index file %s is version %"PRIu32
776                                      " and is not supported by this binary"
777                                      " (try upgrading GIT to a newer version)",
778                                      path, version);
779                 }
780         } else
781                 version = 1;
782
783         nr = 0;
784         index = idx_map;
785         if (version > 1)
786                 index += 2;  /* skip index header */
787         for (i = 0; i < 256; i++) {
788                 uint32_t n = ntohl(index[i]);
789                 if (n < nr) {
790                         munmap(idx_map, idx_size);
791                         return error("non-monotonic index %s", path);
792                 }
793                 nr = n;
794         }
795
796         if (version == 1) {
797                 /*
798                  * Total size:
799                  *  - 256 index entries 4 bytes each
800                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
801                  *  - 20-byte SHA1 of the packfile
802                  *  - 20-byte SHA1 file checksum
803                  */
804                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
805                         munmap(idx_map, idx_size);
806                         return error("wrong index v1 file size in %s", path);
807                 }
808         } else if (version == 2) {
809                 /*
810                  * Minimum size:
811                  *  - 8 bytes of header
812                  *  - 256 index entries 4 bytes each
813                  *  - 20-byte sha1 entry * nr
814                  *  - 4-byte crc entry * nr
815                  *  - 4-byte offset entry * nr
816                  *  - 20-byte SHA1 of the packfile
817                  *  - 20-byte SHA1 file checksum
818                  * And after the 4-byte offset table might be a
819                  * variable sized table containing 8-byte entries
820                  * for offsets larger than 2^31.
821                  */
822                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
823                 unsigned long max_size = min_size;
824                 if (nr)
825                         max_size += (nr - 1)*8;
826                 if (idx_size < min_size || idx_size > max_size) {
827                         munmap(idx_map, idx_size);
828                         return error("wrong index v2 file size in %s", path);
829                 }
830                 if (idx_size != min_size &&
831                     /*
832                      * make sure we can deal with large pack offsets.
833                      * 31-bit signed offset won't be enough, neither
834                      * 32-bit unsigned one will be.
835                      */
836                     (sizeof(off_t) <= 4)) {
837                         munmap(idx_map, idx_size);
838                         return error("pack too large for current definition of off_t in %s", path);
839                 }
840         }
841
842         p->index_version = version;
843         p->index_data = idx_map;
844         p->index_size = idx_size;
845         p->num_objects = nr;
846         return 0;
847 }
848
849 int open_pack_index(struct packed_git *p)
850 {
851         char *idx_name;
852         size_t len;
853         int ret;
854
855         if (p->index_data)
856                 return 0;
857
858         if (!strip_suffix(p->pack_name, ".pack", &len))
859                 die("BUG: pack_name does not end in .pack");
860         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
861         ret = check_packed_git_idx(idx_name, p);
862         free(idx_name);
863         return ret;
864 }
865
866 static void scan_windows(struct packed_git *p,
867         struct packed_git **lru_p,
868         struct pack_window **lru_w,
869         struct pack_window **lru_l)
870 {
871         struct pack_window *w, *w_l;
872
873         for (w_l = NULL, w = p->windows; w; w = w->next) {
874                 if (!w->inuse_cnt) {
875                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
876                                 *lru_p = p;
877                                 *lru_w = w;
878                                 *lru_l = w_l;
879                         }
880                 }
881                 w_l = w;
882         }
883 }
884
885 static int unuse_one_window(struct packed_git *current)
886 {
887         struct packed_git *p, *lru_p = NULL;
888         struct pack_window *lru_w = NULL, *lru_l = NULL;
889
890         if (current)
891                 scan_windows(current, &lru_p, &lru_w, &lru_l);
892         for (p = packed_git; p; p = p->next)
893                 scan_windows(p, &lru_p, &lru_w, &lru_l);
894         if (lru_p) {
895                 munmap(lru_w->base, lru_w->len);
896                 pack_mapped -= lru_w->len;
897                 if (lru_l)
898                         lru_l->next = lru_w->next;
899                 else
900                         lru_p->windows = lru_w->next;
901                 free(lru_w);
902                 pack_open_windows--;
903                 return 1;
904         }
905         return 0;
906 }
907
908 void release_pack_memory(size_t need)
909 {
910         size_t cur = pack_mapped;
911         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
912                 ; /* nothing */
913 }
914
915 static void mmap_limit_check(size_t length)
916 {
917         static size_t limit = 0;
918         if (!limit) {
919                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
920                 if (!limit)
921                         limit = SIZE_MAX;
922         }
923         if (length > limit)
924                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
925                     (uintmax_t)length, (uintmax_t)limit);
926 }
927
928 void *xmmap_gently(void *start, size_t length,
929                   int prot, int flags, int fd, off_t offset)
930 {
931         void *ret;
932
933         mmap_limit_check(length);
934         ret = mmap(start, length, prot, flags, fd, offset);
935         if (ret == MAP_FAILED) {
936                 if (!length)
937                         return NULL;
938                 release_pack_memory(length);
939                 ret = mmap(start, length, prot, flags, fd, offset);
940         }
941         return ret;
942 }
943
944 void *xmmap(void *start, size_t length,
945         int prot, int flags, int fd, off_t offset)
946 {
947         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
948         if (ret == MAP_FAILED)
949                 die_errno("mmap failed");
950         return ret;
951 }
952
953 void close_pack_windows(struct packed_git *p)
954 {
955         while (p->windows) {
956                 struct pack_window *w = p->windows;
957
958                 if (w->inuse_cnt)
959                         die("pack '%s' still has open windows to it",
960                             p->pack_name);
961                 munmap(w->base, w->len);
962                 pack_mapped -= w->len;
963                 pack_open_windows--;
964                 p->windows = w->next;
965                 free(w);
966         }
967 }
968
969 static int close_pack_fd(struct packed_git *p)
970 {
971         if (p->pack_fd < 0)
972                 return 0;
973
974         close(p->pack_fd);
975         pack_open_fds--;
976         p->pack_fd = -1;
977
978         return 1;
979 }
980
981 static void close_pack(struct packed_git *p)
982 {
983         close_pack_windows(p);
984         close_pack_fd(p);
985         close_pack_index(p);
986 }
987
988 void close_all_packs(void)
989 {
990         struct packed_git *p;
991
992         for (p = packed_git; p; p = p->next)
993                 if (p->do_not_close)
994                         die("BUG: want to close pack marked 'do-not-close'");
995                 else
996                         close_pack(p);
997 }
998
999
1000 /*
1001  * The LRU pack is the one with the oldest MRU window, preferring packs
1002  * with no used windows, or the oldest mtime if it has no windows allocated.
1003  */
1004 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
1005 {
1006         struct pack_window *w, *this_mru_w;
1007         int has_windows_inuse = 0;
1008
1009         /*
1010          * Reject this pack if it has windows and the previously selected
1011          * one does not.  If this pack does not have windows, reject
1012          * it if the pack file is newer than the previously selected one.
1013          */
1014         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
1015                 return;
1016
1017         for (w = this_mru_w = p->windows; w; w = w->next) {
1018                 /*
1019                  * Reject this pack if any of its windows are in use,
1020                  * but the previously selected pack did not have any
1021                  * inuse windows.  Otherwise, record that this pack
1022                  * has windows in use.
1023                  */
1024                 if (w->inuse_cnt) {
1025                         if (*accept_windows_inuse)
1026                                 has_windows_inuse = 1;
1027                         else
1028                                 return;
1029                 }
1030
1031                 if (w->last_used > this_mru_w->last_used)
1032                         this_mru_w = w;
1033
1034                 /*
1035                  * Reject this pack if it has windows that have been
1036                  * used more recently than the previously selected pack.
1037                  * If the previously selected pack had windows inuse and
1038                  * we have not encountered a window in this pack that is
1039                  * inuse, skip this check since we prefer a pack with no
1040                  * inuse windows to one that has inuse windows.
1041                  */
1042                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
1043                     this_mru_w->last_used > (*mru_w)->last_used)
1044                         return;
1045         }
1046
1047         /*
1048          * Select this pack.
1049          */
1050         *mru_w = this_mru_w;
1051         *lru_p = p;
1052         *accept_windows_inuse = has_windows_inuse;
1053 }
1054
1055 static int close_one_pack(void)
1056 {
1057         struct packed_git *p, *lru_p = NULL;
1058         struct pack_window *mru_w = NULL;
1059         int accept_windows_inuse = 1;
1060
1061         for (p = packed_git; p; p = p->next) {
1062                 if (p->pack_fd == -1)
1063                         continue;
1064                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
1065         }
1066
1067         if (lru_p)
1068                 return close_pack_fd(lru_p);
1069
1070         return 0;
1071 }
1072
1073 void unuse_pack(struct pack_window **w_cursor)
1074 {
1075         struct pack_window *w = *w_cursor;
1076         if (w) {
1077                 w->inuse_cnt--;
1078                 *w_cursor = NULL;
1079         }
1080 }
1081
1082 void close_pack_index(struct packed_git *p)
1083 {
1084         if (p->index_data) {
1085                 munmap((void *)p->index_data, p->index_size);
1086                 p->index_data = NULL;
1087         }
1088 }
1089
1090 static unsigned int get_max_fd_limit(void)
1091 {
1092 #ifdef RLIMIT_NOFILE
1093         {
1094                 struct rlimit lim;
1095
1096                 if (!getrlimit(RLIMIT_NOFILE, &lim))
1097                         return lim.rlim_cur;
1098         }
1099 #endif
1100
1101 #ifdef _SC_OPEN_MAX
1102         {
1103                 long open_max = sysconf(_SC_OPEN_MAX);
1104                 if (0 < open_max)
1105                         return open_max;
1106                 /*
1107                  * Otherwise, we got -1 for one of the two
1108                  * reasons:
1109                  *
1110                  * (1) sysconf() did not understand _SC_OPEN_MAX
1111                  *     and signaled an error with -1; or
1112                  * (2) sysconf() said there is no limit.
1113                  *
1114                  * We _could_ clear errno before calling sysconf() to
1115                  * tell these two cases apart and return a huge number
1116                  * in the latter case to let the caller cap it to a
1117                  * value that is not so selfish, but letting the
1118                  * fallback OPEN_MAX codepath take care of these cases
1119                  * is a lot simpler.
1120                  */
1121         }
1122 #endif
1123
1124 #ifdef OPEN_MAX
1125         return OPEN_MAX;
1126 #else
1127         return 1; /* see the caller ;-) */
1128 #endif
1129 }
1130
1131 /*
1132  * Do not call this directly as this leaks p->pack_fd on error return;
1133  * call open_packed_git() instead.
1134  */
1135 static int open_packed_git_1(struct packed_git *p)
1136 {
1137         struct stat st;
1138         struct pack_header hdr;
1139         unsigned char sha1[20];
1140         unsigned char *idx_sha1;
1141         long fd_flag;
1142
1143         if (!p->index_data && open_pack_index(p))
1144                 return error("packfile %s index unavailable", p->pack_name);
1145
1146         if (!pack_max_fds) {
1147                 unsigned int max_fds = get_max_fd_limit();
1148
1149                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1150                 if (25 < max_fds)
1151                         pack_max_fds = max_fds - 25;
1152                 else
1153                         pack_max_fds = 1;
1154         }
1155
1156         while (pack_max_fds <= pack_open_fds && close_one_pack())
1157                 ; /* nothing */
1158
1159         p->pack_fd = git_open(p->pack_name);
1160         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1161                 return -1;
1162         pack_open_fds++;
1163
1164         /* If we created the struct before we had the pack we lack size. */
1165         if (!p->pack_size) {
1166                 if (!S_ISREG(st.st_mode))
1167                         return error("packfile %s not a regular file", p->pack_name);
1168                 p->pack_size = st.st_size;
1169         } else if (p->pack_size != st.st_size)
1170                 return error("packfile %s size changed", p->pack_name);
1171
1172         /* We leave these file descriptors open with sliding mmap;
1173          * there is no point keeping them open across exec(), though.
1174          */
1175         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1176         if (fd_flag < 0)
1177                 return error("cannot determine file descriptor flags");
1178         fd_flag |= FD_CLOEXEC;
1179         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1180                 return error("cannot set FD_CLOEXEC");
1181
1182         /* Verify we recognize this pack file format. */
1183         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1184                 return error("file %s is far too short to be a packfile", p->pack_name);
1185         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1186                 return error("file %s is not a GIT packfile", p->pack_name);
1187         if (!pack_version_ok(hdr.hdr_version))
1188                 return error("packfile %s is version %"PRIu32" and not"
1189                         " supported (try upgrading GIT to a newer version)",
1190                         p->pack_name, ntohl(hdr.hdr_version));
1191
1192         /* Verify the pack matches its index. */
1193         if (p->num_objects != ntohl(hdr.hdr_entries))
1194                 return error("packfile %s claims to have %"PRIu32" objects"
1195                              " while index indicates %"PRIu32" objects",
1196                              p->pack_name, ntohl(hdr.hdr_entries),
1197                              p->num_objects);
1198         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1199                 return error("end of packfile %s is unavailable", p->pack_name);
1200         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1201                 return error("packfile %s signature is unavailable", p->pack_name);
1202         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1203         if (hashcmp(sha1, idx_sha1))
1204                 return error("packfile %s does not match index", p->pack_name);
1205         return 0;
1206 }
1207
1208 static int open_packed_git(struct packed_git *p)
1209 {
1210         if (!open_packed_git_1(p))
1211                 return 0;
1212         close_pack_fd(p);
1213         return -1;
1214 }
1215
1216 static int in_window(struct pack_window *win, off_t offset)
1217 {
1218         /* We must promise at least 20 bytes (one hash) after the
1219          * offset is available from this window, otherwise the offset
1220          * is not actually in this window and a different window (which
1221          * has that one hash excess) must be used.  This is to support
1222          * the object header and delta base parsing routines below.
1223          */
1224         off_t win_off = win->offset;
1225         return win_off <= offset
1226                 && (offset + 20) <= (win_off + win->len);
1227 }
1228
1229 unsigned char *use_pack(struct packed_git *p,
1230                 struct pack_window **w_cursor,
1231                 off_t offset,
1232                 unsigned long *left)
1233 {
1234         struct pack_window *win = *w_cursor;
1235
1236         /* Since packfiles end in a hash of their content and it's
1237          * pointless to ask for an offset into the middle of that
1238          * hash, and the in_window function above wouldn't match
1239          * don't allow an offset too close to the end of the file.
1240          */
1241         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1242                 die("packfile %s cannot be accessed", p->pack_name);
1243         if (offset > (p->pack_size - 20))
1244                 die("offset beyond end of packfile (truncated pack?)");
1245         if (offset < 0)
1246                 die(_("offset before end of packfile (broken .idx?)"));
1247
1248         if (!win || !in_window(win, offset)) {
1249                 if (win)
1250                         win->inuse_cnt--;
1251                 for (win = p->windows; win; win = win->next) {
1252                         if (in_window(win, offset))
1253                                 break;
1254                 }
1255                 if (!win) {
1256                         size_t window_align = packed_git_window_size / 2;
1257                         off_t len;
1258
1259                         if (p->pack_fd == -1 && open_packed_git(p))
1260                                 die("packfile %s cannot be accessed", p->pack_name);
1261
1262                         win = xcalloc(1, sizeof(*win));
1263                         win->offset = (offset / window_align) * window_align;
1264                         len = p->pack_size - win->offset;
1265                         if (len > packed_git_window_size)
1266                                 len = packed_git_window_size;
1267                         win->len = (size_t)len;
1268                         pack_mapped += win->len;
1269                         while (packed_git_limit < pack_mapped
1270                                 && unuse_one_window(p))
1271                                 ; /* nothing */
1272                         win->base = xmmap(NULL, win->len,
1273                                 PROT_READ, MAP_PRIVATE,
1274                                 p->pack_fd, win->offset);
1275                         if (win->base == MAP_FAILED)
1276                                 die_errno("packfile %s cannot be mapped",
1277                                           p->pack_name);
1278                         if (!win->offset && win->len == p->pack_size
1279                                 && !p->do_not_close)
1280                                 close_pack_fd(p);
1281                         pack_mmap_calls++;
1282                         pack_open_windows++;
1283                         if (pack_mapped > peak_pack_mapped)
1284                                 peak_pack_mapped = pack_mapped;
1285                         if (pack_open_windows > peak_pack_open_windows)
1286                                 peak_pack_open_windows = pack_open_windows;
1287                         win->next = p->windows;
1288                         p->windows = win;
1289                 }
1290         }
1291         if (win != *w_cursor) {
1292                 win->last_used = pack_used_ctr++;
1293                 win->inuse_cnt++;
1294                 *w_cursor = win;
1295         }
1296         offset -= win->offset;
1297         if (left)
1298                 *left = win->len - xsize_t(offset);
1299         return win->base + offset;
1300 }
1301
1302 static struct packed_git *alloc_packed_git(int extra)
1303 {
1304         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1305         memset(p, 0, sizeof(*p));
1306         p->pack_fd = -1;
1307         return p;
1308 }
1309
1310 static void try_to_free_pack_memory(size_t size)
1311 {
1312         release_pack_memory(size);
1313 }
1314
1315 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1316 {
1317         static int have_set_try_to_free_routine;
1318         struct stat st;
1319         size_t alloc;
1320         struct packed_git *p;
1321
1322         if (!have_set_try_to_free_routine) {
1323                 have_set_try_to_free_routine = 1;
1324                 set_try_to_free_routine(try_to_free_pack_memory);
1325         }
1326
1327         /*
1328          * Make sure a corresponding .pack file exists and that
1329          * the index looks sane.
1330          */
1331         if (!strip_suffix_mem(path, &path_len, ".idx"))
1332                 return NULL;
1333
1334         /*
1335          * ".pack" is long enough to hold any suffix we're adding (and
1336          * the use xsnprintf double-checks that)
1337          */
1338         alloc = st_add3(path_len, strlen(".pack"), 1);
1339         p = alloc_packed_git(alloc);
1340         memcpy(p->pack_name, path, path_len);
1341
1342         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1343         if (!access(p->pack_name, F_OK))
1344                 p->pack_keep = 1;
1345
1346         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1347         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1348                 free(p);
1349                 return NULL;
1350         }
1351
1352         /* ok, it looks sane as far as we can check without
1353          * actually mapping the pack file.
1354          */
1355         p->pack_size = st.st_size;
1356         p->pack_local = local;
1357         p->mtime = st.st_mtime;
1358         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1359                 hashclr(p->sha1);
1360         return p;
1361 }
1362
1363 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1364 {
1365         const char *path = sha1_pack_name(sha1);
1366         size_t alloc = st_add(strlen(path), 1);
1367         struct packed_git *p = alloc_packed_git(alloc);
1368
1369         memcpy(p->pack_name, path, alloc); /* includes NUL */
1370         hashcpy(p->sha1, sha1);
1371         if (check_packed_git_idx(idx_path, p)) {
1372                 free(p);
1373                 return NULL;
1374         }
1375
1376         return p;
1377 }
1378
1379 void install_packed_git(struct packed_git *pack)
1380 {
1381         if (pack->pack_fd != -1)
1382                 pack_open_fds++;
1383
1384         pack->next = packed_git;
1385         packed_git = pack;
1386 }
1387
1388 void (*report_garbage)(unsigned seen_bits, const char *path);
1389
1390 static void report_helper(const struct string_list *list,
1391                           int seen_bits, int first, int last)
1392 {
1393         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1394                 return;
1395
1396         for (; first < last; first++)
1397                 report_garbage(seen_bits, list->items[first].string);
1398 }
1399
1400 static void report_pack_garbage(struct string_list *list)
1401 {
1402         int i, baselen = -1, first = 0, seen_bits = 0;
1403
1404         if (!report_garbage)
1405                 return;
1406
1407         string_list_sort(list);
1408
1409         for (i = 0; i < list->nr; i++) {
1410                 const char *path = list->items[i].string;
1411                 if (baselen != -1 &&
1412                     strncmp(path, list->items[first].string, baselen)) {
1413                         report_helper(list, seen_bits, first, i);
1414                         baselen = -1;
1415                         seen_bits = 0;
1416                 }
1417                 if (baselen == -1) {
1418                         const char *dot = strrchr(path, '.');
1419                         if (!dot) {
1420                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1421                                 continue;
1422                         }
1423                         baselen = dot - path + 1;
1424                         first = i;
1425                 }
1426                 if (!strcmp(path + baselen, "pack"))
1427                         seen_bits |= 1;
1428                 else if (!strcmp(path + baselen, "idx"))
1429                         seen_bits |= 2;
1430         }
1431         report_helper(list, seen_bits, first, list->nr);
1432 }
1433
1434 static void prepare_packed_git_one(char *objdir, int local)
1435 {
1436         struct strbuf path = STRBUF_INIT;
1437         size_t dirnamelen;
1438         DIR *dir;
1439         struct dirent *de;
1440         struct string_list garbage = STRING_LIST_INIT_DUP;
1441
1442         strbuf_addstr(&path, objdir);
1443         strbuf_addstr(&path, "/pack");
1444         dir = opendir(path.buf);
1445         if (!dir) {
1446                 if (errno != ENOENT)
1447                         error_errno("unable to open object pack directory: %s",
1448                                     path.buf);
1449                 strbuf_release(&path);
1450                 return;
1451         }
1452         strbuf_addch(&path, '/');
1453         dirnamelen = path.len;
1454         while ((de = readdir(dir)) != NULL) {
1455                 struct packed_git *p;
1456                 size_t base_len;
1457
1458                 if (is_dot_or_dotdot(de->d_name))
1459                         continue;
1460
1461                 strbuf_setlen(&path, dirnamelen);
1462                 strbuf_addstr(&path, de->d_name);
1463
1464                 base_len = path.len;
1465                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1466                         /* Don't reopen a pack we already have. */
1467                         for (p = packed_git; p; p = p->next) {
1468                                 size_t len;
1469                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1470                                     len == base_len &&
1471                                     !memcmp(p->pack_name, path.buf, len))
1472                                         break;
1473                         }
1474                         if (p == NULL &&
1475                             /*
1476                              * See if it really is a valid .idx file with
1477                              * corresponding .pack file that we can map.
1478                              */
1479                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1480                                 install_packed_git(p);
1481                 }
1482
1483                 if (!report_garbage)
1484                         continue;
1485
1486                 if (ends_with(de->d_name, ".idx") ||
1487                     ends_with(de->d_name, ".pack") ||
1488                     ends_with(de->d_name, ".bitmap") ||
1489                     ends_with(de->d_name, ".keep"))
1490                         string_list_append(&garbage, path.buf);
1491                 else
1492                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1493         }
1494         closedir(dir);
1495         report_pack_garbage(&garbage);
1496         string_list_clear(&garbage, 0);
1497         strbuf_release(&path);
1498 }
1499
1500 static int approximate_object_count_valid;
1501
1502 /*
1503  * Give a fast, rough count of the number of objects in the repository. This
1504  * ignores loose objects completely. If you have a lot of them, then either
1505  * you should repack because your performance will be awful, or they are
1506  * all unreachable objects about to be pruned, in which case they're not really
1507  * interesting as a measure of repo size in the first place.
1508  */
1509 unsigned long approximate_object_count(void)
1510 {
1511         static unsigned long count;
1512         if (!approximate_object_count_valid) {
1513                 struct packed_git *p;
1514
1515                 prepare_packed_git();
1516                 count = 0;
1517                 for (p = packed_git; p; p = p->next) {
1518                         if (open_pack_index(p))
1519                                 continue;
1520                         count += p->num_objects;
1521                 }
1522         }
1523         return count;
1524 }
1525
1526 static void *get_next_packed_git(const void *p)
1527 {
1528         return ((const struct packed_git *)p)->next;
1529 }
1530
1531 static void set_next_packed_git(void *p, void *next)
1532 {
1533         ((struct packed_git *)p)->next = next;
1534 }
1535
1536 static int sort_pack(const void *a_, const void *b_)
1537 {
1538         const struct packed_git *a = a_;
1539         const struct packed_git *b = b_;
1540         int st;
1541
1542         /*
1543          * Local packs tend to contain objects specific to our
1544          * variant of the project than remote ones.  In addition,
1545          * remote ones could be on a network mounted filesystem.
1546          * Favor local ones for these reasons.
1547          */
1548         st = a->pack_local - b->pack_local;
1549         if (st)
1550                 return -st;
1551
1552         /*
1553          * Younger packs tend to contain more recent objects,
1554          * and more recent objects tend to get accessed more
1555          * often.
1556          */
1557         if (a->mtime < b->mtime)
1558                 return 1;
1559         else if (a->mtime == b->mtime)
1560                 return 0;
1561         return -1;
1562 }
1563
1564 static void rearrange_packed_git(void)
1565 {
1566         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1567                                      set_next_packed_git, sort_pack);
1568 }
1569
1570 static void prepare_packed_git_mru(void)
1571 {
1572         struct packed_git *p;
1573
1574         mru_clear(packed_git_mru);
1575         for (p = packed_git; p; p = p->next)
1576                 mru_append(packed_git_mru, p);
1577 }
1578
1579 static int prepare_packed_git_run_once = 0;
1580 void prepare_packed_git(void)
1581 {
1582         struct alternate_object_database *alt;
1583
1584         if (prepare_packed_git_run_once)
1585                 return;
1586         prepare_packed_git_one(get_object_directory(), 1);
1587         prepare_alt_odb();
1588         for (alt = alt_odb_list; alt; alt = alt->next)
1589                 prepare_packed_git_one(alt->path, 0);
1590         rearrange_packed_git();
1591         prepare_packed_git_mru();
1592         prepare_packed_git_run_once = 1;
1593 }
1594
1595 void reprepare_packed_git(void)
1596 {
1597         approximate_object_count_valid = 0;
1598         prepare_packed_git_run_once = 0;
1599         prepare_packed_git();
1600 }
1601
1602 static void mark_bad_packed_object(struct packed_git *p,
1603                                    const unsigned char *sha1)
1604 {
1605         unsigned i;
1606         for (i = 0; i < p->num_bad_objects; i++)
1607                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1608                         return;
1609         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1610                                       st_mult(GIT_MAX_RAWSZ,
1611                                               st_add(p->num_bad_objects, 1)));
1612         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1613         p->num_bad_objects++;
1614 }
1615
1616 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1617 {
1618         struct packed_git *p;
1619         unsigned i;
1620
1621         for (p = packed_git; p; p = p->next)
1622                 for (i = 0; i < p->num_bad_objects; i++)
1623                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1624                                 return p;
1625         return NULL;
1626 }
1627
1628 /*
1629  * With an in-core object data in "map", rehash it to make sure the
1630  * object name actually matches "sha1" to detect object corruption.
1631  * With "map" == NULL, try reading the object named with "sha1" using
1632  * the streaming interface and rehash it to do the same.
1633  */
1634 int check_sha1_signature(const unsigned char *sha1, void *map,
1635                          unsigned long size, const char *type)
1636 {
1637         unsigned char real_sha1[20];
1638         enum object_type obj_type;
1639         struct git_istream *st;
1640         git_SHA_CTX c;
1641         char hdr[32];
1642         int hdrlen;
1643
1644         if (map) {
1645                 hash_sha1_file(map, size, type, real_sha1);
1646                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1647         }
1648
1649         st = open_istream(sha1, &obj_type, &size, NULL);
1650         if (!st)
1651                 return -1;
1652
1653         /* Generate the header */
1654         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1655
1656         /* Sha1.. */
1657         git_SHA1_Init(&c);
1658         git_SHA1_Update(&c, hdr, hdrlen);
1659         for (;;) {
1660                 char buf[1024 * 16];
1661                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1662
1663                 if (readlen < 0) {
1664                         close_istream(st);
1665                         return -1;
1666                 }
1667                 if (!readlen)
1668                         break;
1669                 git_SHA1_Update(&c, buf, readlen);
1670         }
1671         git_SHA1_Final(real_sha1, &c);
1672         close_istream(st);
1673         return hashcmp(sha1, real_sha1) ? -1 : 0;
1674 }
1675
1676 int git_open_cloexec(const char *name, int flags)
1677 {
1678         int fd;
1679         static int o_cloexec = O_CLOEXEC;
1680
1681         fd = open(name, flags | o_cloexec);
1682         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
1683                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
1684                 o_cloexec &= ~O_CLOEXEC;
1685                 fd = open(name, flags | o_cloexec);
1686         }
1687
1688 #if defined(F_GETFD) && defined(F_SETFD) && defined(FD_CLOEXEC)
1689         {
1690                 static int fd_cloexec = FD_CLOEXEC;
1691
1692                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
1693                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
1694                         int flags = fcntl(fd, F_GETFD);
1695                         if (fcntl(fd, F_SETFD, flags | fd_cloexec))
1696                                 fd_cloexec = 0;
1697                 }
1698         }
1699 #endif
1700         return fd;
1701 }
1702
1703 /*
1704  * Find "sha1" as a loose object in the local repository or in an alternate.
1705  * Returns 0 on success, negative on failure.
1706  *
1707  * The "path" out-parameter will give the path of the object we found (if any).
1708  * Note that it may point to static storage and is only valid until another
1709  * call to sha1_file_name(), etc.
1710  */
1711 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
1712                           const char **path)
1713 {
1714         struct alternate_object_database *alt;
1715
1716         *path = sha1_file_name(sha1);
1717         if (!lstat(*path, st))
1718                 return 0;
1719
1720         prepare_alt_odb();
1721         errno = ENOENT;
1722         for (alt = alt_odb_list; alt; alt = alt->next) {
1723                 *path = alt_sha1_path(alt, sha1);
1724                 if (!lstat(*path, st))
1725                         return 0;
1726         }
1727
1728         return -1;
1729 }
1730
1731 /*
1732  * Like stat_sha1_file(), but actually open the object and return the
1733  * descriptor. See the caveats on the "path" parameter above.
1734  */
1735 static int open_sha1_file(const unsigned char *sha1, const char **path)
1736 {
1737         int fd;
1738         struct alternate_object_database *alt;
1739         int most_interesting_errno;
1740
1741         *path = sha1_file_name(sha1);
1742         fd = git_open(*path);
1743         if (fd >= 0)
1744                 return fd;
1745         most_interesting_errno = errno;
1746
1747         prepare_alt_odb();
1748         for (alt = alt_odb_list; alt; alt = alt->next) {
1749                 *path = alt_sha1_path(alt, sha1);
1750                 fd = git_open(*path);
1751                 if (fd >= 0)
1752                         return fd;
1753                 if (most_interesting_errno == ENOENT)
1754                         most_interesting_errno = errno;
1755         }
1756         errno = most_interesting_errno;
1757         return -1;
1758 }
1759
1760 /*
1761  * Map the loose object at "path" if it is not NULL, or the path found by
1762  * searching for a loose object named "sha1".
1763  */
1764 static void *map_sha1_file_1(const char *path,
1765                              const unsigned char *sha1,
1766                              unsigned long *size)
1767 {
1768         void *map;
1769         int fd;
1770
1771         if (path)
1772                 fd = git_open(path);
1773         else
1774                 fd = open_sha1_file(sha1, &path);
1775         map = NULL;
1776         if (fd >= 0) {
1777                 struct stat st;
1778
1779                 if (!fstat(fd, &st)) {
1780                         *size = xsize_t(st.st_size);
1781                         if (!*size) {
1782                                 /* mmap() is forbidden on empty files */
1783                                 error("object file %s is empty", path);
1784                                 return NULL;
1785                         }
1786                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1787                 }
1788                 close(fd);
1789         }
1790         return map;
1791 }
1792
1793 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1794 {
1795         return map_sha1_file_1(NULL, sha1, size);
1796 }
1797
1798 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1799                 unsigned long len, enum object_type *type, unsigned long *sizep)
1800 {
1801         unsigned shift;
1802         unsigned long size, c;
1803         unsigned long used = 0;
1804
1805         c = buf[used++];
1806         *type = (c >> 4) & 7;
1807         size = c & 15;
1808         shift = 4;
1809         while (c & 0x80) {
1810                 if (len <= used || bitsizeof(long) <= shift) {
1811                         error("bad object header");
1812                         size = used = 0;
1813                         break;
1814                 }
1815                 c = buf[used++];
1816                 size += (c & 0x7f) << shift;
1817                 shift += 7;
1818         }
1819         *sizep = size;
1820         return used;
1821 }
1822
1823 static int unpack_sha1_short_header(git_zstream *stream,
1824                                     unsigned char *map, unsigned long mapsize,
1825                                     void *buffer, unsigned long bufsiz)
1826 {
1827         /* Get the data stream */
1828         memset(stream, 0, sizeof(*stream));
1829         stream->next_in = map;
1830         stream->avail_in = mapsize;
1831         stream->next_out = buffer;
1832         stream->avail_out = bufsiz;
1833
1834         git_inflate_init(stream);
1835         return git_inflate(stream, 0);
1836 }
1837
1838 int unpack_sha1_header(git_zstream *stream,
1839                        unsigned char *map, unsigned long mapsize,
1840                        void *buffer, unsigned long bufsiz)
1841 {
1842         int status = unpack_sha1_short_header(stream, map, mapsize,
1843                                               buffer, bufsiz);
1844
1845         if (status < Z_OK)
1846                 return status;
1847
1848         /* Make sure we have the terminating NUL */
1849         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1850                 return -1;
1851         return 0;
1852 }
1853
1854 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1855                                         unsigned long mapsize, void *buffer,
1856                                         unsigned long bufsiz, struct strbuf *header)
1857 {
1858         int status;
1859
1860         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
1861         if (status < Z_OK)
1862                 return -1;
1863
1864         /*
1865          * Check if entire header is unpacked in the first iteration.
1866          */
1867         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1868                 return 0;
1869
1870         /*
1871          * buffer[0..bufsiz] was not large enough.  Copy the partial
1872          * result out to header, and then append the result of further
1873          * reading the stream.
1874          */
1875         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1876         stream->next_out = buffer;
1877         stream->avail_out = bufsiz;
1878
1879         do {
1880                 status = git_inflate(stream, 0);
1881                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1882                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1883                         return 0;
1884                 stream->next_out = buffer;
1885                 stream->avail_out = bufsiz;
1886         } while (status != Z_STREAM_END);
1887         return -1;
1888 }
1889
1890 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1891 {
1892         int bytes = strlen(buffer) + 1;
1893         unsigned char *buf = xmallocz(size);
1894         unsigned long n;
1895         int status = Z_OK;
1896
1897         n = stream->total_out - bytes;
1898         if (n > size)
1899                 n = size;
1900         memcpy(buf, (char *) buffer + bytes, n);
1901         bytes = n;
1902         if (bytes <= size) {
1903                 /*
1904                  * The above condition must be (bytes <= size), not
1905                  * (bytes < size).  In other words, even though we
1906                  * expect no more output and set avail_out to zero,
1907                  * the input zlib stream may have bytes that express
1908                  * "this concludes the stream", and we *do* want to
1909                  * eat that input.
1910                  *
1911                  * Otherwise we would not be able to test that we
1912                  * consumed all the input to reach the expected size;
1913                  * we also want to check that zlib tells us that all
1914                  * went well with status == Z_STREAM_END at the end.
1915                  */
1916                 stream->next_out = buf + bytes;
1917                 stream->avail_out = size - bytes;
1918                 while (status == Z_OK)
1919                         status = git_inflate(stream, Z_FINISH);
1920         }
1921         if (status == Z_STREAM_END && !stream->avail_in) {
1922                 git_inflate_end(stream);
1923                 return buf;
1924         }
1925
1926         if (status < 0)
1927                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1928         else if (stream->avail_in)
1929                 error("garbage at end of loose object '%s'",
1930                       sha1_to_hex(sha1));
1931         free(buf);
1932         return NULL;
1933 }
1934
1935 /*
1936  * We used to just use "sscanf()", but that's actually way
1937  * too permissive for what we want to check. So do an anal
1938  * object header parse by hand.
1939  */
1940 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1941                                unsigned int flags)
1942 {
1943         const char *type_buf = hdr;
1944         unsigned long size;
1945         int type, type_len = 0;
1946
1947         /*
1948          * The type can be of any size but is followed by
1949          * a space.
1950          */
1951         for (;;) {
1952                 char c = *hdr++;
1953                 if (!c)
1954                         return -1;
1955                 if (c == ' ')
1956                         break;
1957                 type_len++;
1958         }
1959
1960         type = type_from_string_gently(type_buf, type_len, 1);
1961         if (oi->typename)
1962                 strbuf_add(oi->typename, type_buf, type_len);
1963         /*
1964          * Set type to 0 if its an unknown object and
1965          * we're obtaining the type using '--allow-unknown-type'
1966          * option.
1967          */
1968         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE) && (type < 0))
1969                 type = 0;
1970         else if (type < 0)
1971                 die("invalid object type");
1972         if (oi->typep)
1973                 *oi->typep = type;
1974
1975         /*
1976          * The length must follow immediately, and be in canonical
1977          * decimal format (ie "010" is not valid).
1978          */
1979         size = *hdr++ - '0';
1980         if (size > 9)
1981                 return -1;
1982         if (size) {
1983                 for (;;) {
1984                         unsigned long c = *hdr - '0';
1985                         if (c > 9)
1986                                 break;
1987                         hdr++;
1988                         size = size * 10 + c;
1989                 }
1990         }
1991
1992         if (oi->sizep)
1993                 *oi->sizep = size;
1994
1995         /*
1996          * The length must be followed by a zero byte
1997          */
1998         return *hdr ? -1 : type;
1999 }
2000
2001 int parse_sha1_header(const char *hdr, unsigned long *sizep)
2002 {
2003         struct object_info oi = OBJECT_INFO_INIT;
2004
2005         oi.sizep = sizep;
2006         return parse_sha1_header_extended(hdr, &oi, 0);
2007 }
2008
2009 unsigned long get_size_from_delta(struct packed_git *p,
2010                                   struct pack_window **w_curs,
2011                                   off_t curpos)
2012 {
2013         const unsigned char *data;
2014         unsigned char delta_head[20], *in;
2015         git_zstream stream;
2016         int st;
2017
2018         memset(&stream, 0, sizeof(stream));
2019         stream.next_out = delta_head;
2020         stream.avail_out = sizeof(delta_head);
2021
2022         git_inflate_init(&stream);
2023         do {
2024                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2025                 stream.next_in = in;
2026                 st = git_inflate(&stream, Z_FINISH);
2027                 curpos += stream.next_in - in;
2028         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
2029                  stream.total_out < sizeof(delta_head));
2030         git_inflate_end(&stream);
2031         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
2032                 error("delta data unpack-initial failed");
2033                 return 0;
2034         }
2035
2036         /* Examine the initial part of the delta to figure out
2037          * the result size.
2038          */
2039         data = delta_head;
2040
2041         /* ignore base size */
2042         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2043
2044         /* Read the result size */
2045         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2046 }
2047
2048 static off_t get_delta_base(struct packed_git *p,
2049                                     struct pack_window **w_curs,
2050                                     off_t *curpos,
2051                                     enum object_type type,
2052                                     off_t delta_obj_offset)
2053 {
2054         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
2055         off_t base_offset;
2056
2057         /* use_pack() assured us we have [base_info, base_info + 20)
2058          * as a range that we can look at without walking off the
2059          * end of the mapped window.  Its actually the hash size
2060          * that is assured.  An OFS_DELTA longer than the hash size
2061          * is stupid, as then a REF_DELTA would be smaller to store.
2062          */
2063         if (type == OBJ_OFS_DELTA) {
2064                 unsigned used = 0;
2065                 unsigned char c = base_info[used++];
2066                 base_offset = c & 127;
2067                 while (c & 128) {
2068                         base_offset += 1;
2069                         if (!base_offset || MSB(base_offset, 7))
2070                                 return 0;  /* overflow */
2071                         c = base_info[used++];
2072                         base_offset = (base_offset << 7) + (c & 127);
2073                 }
2074                 base_offset = delta_obj_offset - base_offset;
2075                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
2076                         return 0;  /* out of bound */
2077                 *curpos += used;
2078         } else if (type == OBJ_REF_DELTA) {
2079                 /* The base entry _must_ be in the same pack */
2080                 base_offset = find_pack_entry_one(base_info, p);
2081                 *curpos += 20;
2082         } else
2083                 die("I am totally screwed");
2084         return base_offset;
2085 }
2086
2087 /*
2088  * Like get_delta_base above, but we return the sha1 instead of the pack
2089  * offset. This means it is cheaper for REF deltas (we do not have to do
2090  * the final object lookup), but more expensive for OFS deltas (we
2091  * have to load the revidx to convert the offset back into a sha1).
2092  */
2093 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
2094                                                 struct pack_window **w_curs,
2095                                                 off_t curpos,
2096                                                 enum object_type type,
2097                                                 off_t delta_obj_offset)
2098 {
2099         if (type == OBJ_REF_DELTA) {
2100                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
2101                 return base;
2102         } else if (type == OBJ_OFS_DELTA) {
2103                 struct revindex_entry *revidx;
2104                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
2105                                                    type, delta_obj_offset);
2106
2107                 if (!base_offset)
2108                         return NULL;
2109
2110                 revidx = find_pack_revindex(p, base_offset);
2111                 if (!revidx)
2112                         return NULL;
2113
2114                 return nth_packed_object_sha1(p, revidx->nr);
2115         } else
2116                 return NULL;
2117 }
2118
2119 int unpack_object_header(struct packed_git *p,
2120                          struct pack_window **w_curs,
2121                          off_t *curpos,
2122                          unsigned long *sizep)
2123 {
2124         unsigned char *base;
2125         unsigned long left;
2126         unsigned long used;
2127         enum object_type type;
2128
2129         /* use_pack() assures us we have [base, base + 20) available
2130          * as a range that we can look at.  (Its actually the hash
2131          * size that is assured.)  With our object header encoding
2132          * the maximum deflated object size is 2^137, which is just
2133          * insane, so we know won't exceed what we have been given.
2134          */
2135         base = use_pack(p, w_curs, *curpos, &left);
2136         used = unpack_object_header_buffer(base, left, &type, sizep);
2137         if (!used) {
2138                 type = OBJ_BAD;
2139         } else
2140                 *curpos += used;
2141
2142         return type;
2143 }
2144
2145 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
2146 {
2147         int type;
2148         struct revindex_entry *revidx;
2149         const unsigned char *sha1;
2150         revidx = find_pack_revindex(p, obj_offset);
2151         if (!revidx)
2152                 return OBJ_BAD;
2153         sha1 = nth_packed_object_sha1(p, revidx->nr);
2154         mark_bad_packed_object(p, sha1);
2155         type = sha1_object_info(sha1, NULL);
2156         if (type <= OBJ_NONE)
2157                 return OBJ_BAD;
2158         return type;
2159 }
2160
2161 #define POI_STACK_PREALLOC 64
2162
2163 static enum object_type packed_to_object_type(struct packed_git *p,
2164                                               off_t obj_offset,
2165                                               enum object_type type,
2166                                               struct pack_window **w_curs,
2167                                               off_t curpos)
2168 {
2169         off_t small_poi_stack[POI_STACK_PREALLOC];
2170         off_t *poi_stack = small_poi_stack;
2171         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
2172
2173         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2174                 off_t base_offset;
2175                 unsigned long size;
2176                 /* Push the object we're going to leave behind */
2177                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
2178                         poi_stack_alloc = alloc_nr(poi_stack_nr);
2179                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
2180                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2181                 } else {
2182                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2183                 }
2184                 poi_stack[poi_stack_nr++] = obj_offset;
2185                 /* If parsing the base offset fails, just unwind */
2186                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2187                 if (!base_offset)
2188                         goto unwind;
2189                 curpos = obj_offset = base_offset;
2190                 type = unpack_object_header(p, w_curs, &curpos, &size);
2191                 if (type <= OBJ_NONE) {
2192                         /* If getting the base itself fails, we first
2193                          * retry the base, otherwise unwind */
2194                         type = retry_bad_packed_offset(p, base_offset);
2195                         if (type > OBJ_NONE)
2196                                 goto out;
2197                         goto unwind;
2198                 }
2199         }
2200
2201         switch (type) {
2202         case OBJ_BAD:
2203         case OBJ_COMMIT:
2204         case OBJ_TREE:
2205         case OBJ_BLOB:
2206         case OBJ_TAG:
2207                 break;
2208         default:
2209                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2210                       type, (uintmax_t)obj_offset, p->pack_name);
2211                 type = OBJ_BAD;
2212         }
2213
2214 out:
2215         if (poi_stack != small_poi_stack)
2216                 free(poi_stack);
2217         return type;
2218
2219 unwind:
2220         while (poi_stack_nr) {
2221                 obj_offset = poi_stack[--poi_stack_nr];
2222                 type = retry_bad_packed_offset(p, obj_offset);
2223                 if (type > OBJ_NONE)
2224                         goto out;
2225         }
2226         type = OBJ_BAD;
2227         goto out;
2228 }
2229
2230 static struct hashmap delta_base_cache;
2231 static size_t delta_base_cached;
2232
2233 static LIST_HEAD(delta_base_cache_lru);
2234
2235 struct delta_base_cache_key {
2236         struct packed_git *p;
2237         off_t base_offset;
2238 };
2239
2240 struct delta_base_cache_entry {
2241         struct hashmap hash;
2242         struct delta_base_cache_key key;
2243         struct list_head lru;
2244         void *data;
2245         unsigned long size;
2246         enum object_type type;
2247 };
2248
2249 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2250 {
2251         unsigned int hash;
2252
2253         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2254         hash += (hash >> 8) + (hash >> 16);
2255         return hash;
2256 }
2257
2258 static struct delta_base_cache_entry *
2259 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2260 {
2261         struct hashmap_entry entry;
2262         struct delta_base_cache_key key;
2263
2264         if (!delta_base_cache.cmpfn)
2265                 return NULL;
2266
2267         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2268         key.p = p;
2269         key.base_offset = base_offset;
2270         return hashmap_get(&delta_base_cache, &entry, &key);
2271 }
2272
2273 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2274                                    const struct delta_base_cache_key *b)
2275 {
2276         return a->p == b->p && a->base_offset == b->base_offset;
2277 }
2278
2279 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
2280                                      const void *va, const void *vb,
2281                                      const void *vkey)
2282 {
2283         const struct delta_base_cache_entry *a = va, *b = vb;
2284         const struct delta_base_cache_key *key = vkey;
2285         if (key)
2286                 return !delta_base_cache_key_eq(&a->key, key);
2287         else
2288                 return !delta_base_cache_key_eq(&a->key, &b->key);
2289 }
2290
2291 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2292 {
2293         return !!get_delta_base_cache_entry(p, base_offset);
2294 }
2295
2296 /*
2297  * Remove the entry from the cache, but do _not_ free the associated
2298  * entry data. The caller takes ownership of the "data" buffer, and
2299  * should copy out any fields it wants before detaching.
2300  */
2301 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2302 {
2303         hashmap_remove(&delta_base_cache, ent, &ent->key);
2304         list_del(&ent->lru);
2305         delta_base_cached -= ent->size;
2306         free(ent);
2307 }
2308
2309 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2310         unsigned long *base_size, enum object_type *type)
2311 {
2312         struct delta_base_cache_entry *ent;
2313
2314         ent = get_delta_base_cache_entry(p, base_offset);
2315         if (!ent)
2316                 return unpack_entry(p, base_offset, type, base_size);
2317
2318         if (type)
2319                 *type = ent->type;
2320         if (base_size)
2321                 *base_size = ent->size;
2322         return xmemdupz(ent->data, ent->size);
2323 }
2324
2325 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2326 {
2327         free(ent->data);
2328         detach_delta_base_cache_entry(ent);
2329 }
2330
2331 void clear_delta_base_cache(void)
2332 {
2333         struct list_head *lru, *tmp;
2334         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2335                 struct delta_base_cache_entry *entry =
2336                         list_entry(lru, struct delta_base_cache_entry, lru);
2337                 release_delta_base_cache(entry);
2338         }
2339 }
2340
2341 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2342         void *base, unsigned long base_size, enum object_type type)
2343 {
2344         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2345         struct list_head *lru, *tmp;
2346
2347         delta_base_cached += base_size;
2348
2349         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2350                 struct delta_base_cache_entry *f =
2351                         list_entry(lru, struct delta_base_cache_entry, lru);
2352                 if (delta_base_cached <= delta_base_cache_limit)
2353                         break;
2354                 release_delta_base_cache(f);
2355         }
2356
2357         ent->key.p = p;
2358         ent->key.base_offset = base_offset;
2359         ent->type = type;
2360         ent->data = base;
2361         ent->size = base_size;
2362         list_add_tail(&ent->lru, &delta_base_cache_lru);
2363
2364         if (!delta_base_cache.cmpfn)
2365                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
2366         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2367         hashmap_add(&delta_base_cache, ent);
2368 }
2369
2370 int packed_object_info(struct packed_git *p, off_t obj_offset,
2371                        struct object_info *oi)
2372 {
2373         struct pack_window *w_curs = NULL;
2374         unsigned long size;
2375         off_t curpos = obj_offset;
2376         enum object_type type;
2377
2378         /*
2379          * We always get the representation type, but only convert it to
2380          * a "real" type later if the caller is interested.
2381          */
2382         if (oi->contentp) {
2383                 *oi->contentp = cache_or_unpack_entry(p, obj_offset, oi->sizep,
2384                                                       &type);
2385                 if (!*oi->contentp)
2386                         type = OBJ_BAD;
2387         } else {
2388                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2389         }
2390
2391         if (!oi->contentp && oi->sizep) {
2392                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2393                         off_t tmp_pos = curpos;
2394                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2395                                                            type, obj_offset);
2396                         if (!base_offset) {
2397                                 type = OBJ_BAD;
2398                                 goto out;
2399                         }
2400                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2401                         if (*oi->sizep == 0) {
2402                                 type = OBJ_BAD;
2403                                 goto out;
2404                         }
2405                 } else {
2406                         *oi->sizep = size;
2407                 }
2408         }
2409
2410         if (oi->disk_sizep) {
2411                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2412                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2413         }
2414
2415         if (oi->typep || oi->typename) {
2416                 enum object_type ptot;
2417                 ptot = packed_to_object_type(p, obj_offset, type, &w_curs,
2418                                              curpos);
2419                 if (oi->typep)
2420                         *oi->typep = ptot;
2421                 if (oi->typename) {
2422                         const char *tn = typename(ptot);
2423                         if (tn)
2424                                 strbuf_addstr(oi->typename, tn);
2425                 }
2426                 if (ptot < 0) {
2427                         type = OBJ_BAD;
2428                         goto out;
2429                 }
2430         }
2431
2432         if (oi->delta_base_sha1) {
2433                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2434                         const unsigned char *base;
2435
2436                         base = get_delta_base_sha1(p, &w_curs, curpos,
2437                                                    type, obj_offset);
2438                         if (!base) {
2439                                 type = OBJ_BAD;
2440                                 goto out;
2441                         }
2442
2443                         hashcpy(oi->delta_base_sha1, base);
2444                 } else
2445                         hashclr(oi->delta_base_sha1);
2446         }
2447
2448 out:
2449         unuse_pack(&w_curs);
2450         return type;
2451 }
2452
2453 static void *unpack_compressed_entry(struct packed_git *p,
2454                                     struct pack_window **w_curs,
2455                                     off_t curpos,
2456                                     unsigned long size)
2457 {
2458         int st;
2459         git_zstream stream;
2460         unsigned char *buffer, *in;
2461
2462         buffer = xmallocz_gently(size);
2463         if (!buffer)
2464                 return NULL;
2465         memset(&stream, 0, sizeof(stream));
2466         stream.next_out = buffer;
2467         stream.avail_out = size + 1;
2468
2469         git_inflate_init(&stream);
2470         do {
2471                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2472                 stream.next_in = in;
2473                 st = git_inflate(&stream, Z_FINISH);
2474                 if (!stream.avail_out)
2475                         break; /* the payload is larger than it should be */
2476                 curpos += stream.next_in - in;
2477         } while (st == Z_OK || st == Z_BUF_ERROR);
2478         git_inflate_end(&stream);
2479         if ((st != Z_STREAM_END) || stream.total_out != size) {
2480                 free(buffer);
2481                 return NULL;
2482         }
2483
2484         return buffer;
2485 }
2486
2487 static void *read_object(const unsigned char *sha1, enum object_type *type,
2488                          unsigned long *size);
2489
2490 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2491 {
2492         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2493         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2494                          p->pack_name, (uintmax_t)obj_offset);
2495 }
2496
2497 int do_check_packed_object_crc;
2498
2499 #define UNPACK_ENTRY_STACK_PREALLOC 64
2500 struct unpack_entry_stack_ent {
2501         off_t obj_offset;
2502         off_t curpos;
2503         unsigned long size;
2504 };
2505
2506 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2507                    enum object_type *final_type, unsigned long *final_size)
2508 {
2509         struct pack_window *w_curs = NULL;
2510         off_t curpos = obj_offset;
2511         void *data = NULL;
2512         unsigned long size;
2513         enum object_type type;
2514         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2515         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2516         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2517         int base_from_cache = 0;
2518
2519         write_pack_access_log(p, obj_offset);
2520
2521         /* PHASE 1: drill down to the innermost base object */
2522         for (;;) {
2523                 off_t base_offset;
2524                 int i;
2525                 struct delta_base_cache_entry *ent;
2526
2527                 ent = get_delta_base_cache_entry(p, curpos);
2528                 if (ent) {
2529                         type = ent->type;
2530                         data = ent->data;
2531                         size = ent->size;
2532                         detach_delta_base_cache_entry(ent);
2533                         base_from_cache = 1;
2534                         break;
2535                 }
2536
2537                 if (do_check_packed_object_crc && p->index_version > 1) {
2538                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2539                         off_t len = revidx[1].offset - obj_offset;
2540                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2541                                 const unsigned char *sha1 =
2542                                         nth_packed_object_sha1(p, revidx->nr);
2543                                 error("bad packed object CRC for %s",
2544                                       sha1_to_hex(sha1));
2545                                 mark_bad_packed_object(p, sha1);
2546                                 data = NULL;
2547                                 goto out;
2548                         }
2549                 }
2550
2551                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2552                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2553                         break;
2554
2555                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2556                 if (!base_offset) {
2557                         error("failed to validate delta base reference "
2558                               "at offset %"PRIuMAX" from %s",
2559                               (uintmax_t)curpos, p->pack_name);
2560                         /* bail to phase 2, in hopes of recovery */
2561                         data = NULL;
2562                         break;
2563                 }
2564
2565                 /* push object, proceed to base */
2566                 if (delta_stack_nr >= delta_stack_alloc
2567                     && delta_stack == small_delta_stack) {
2568                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2569                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2570                         memcpy(delta_stack, small_delta_stack,
2571                                sizeof(*delta_stack)*delta_stack_nr);
2572                 } else {
2573                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2574                 }
2575                 i = delta_stack_nr++;
2576                 delta_stack[i].obj_offset = obj_offset;
2577                 delta_stack[i].curpos = curpos;
2578                 delta_stack[i].size = size;
2579
2580                 curpos = obj_offset = base_offset;
2581         }
2582
2583         /* PHASE 2: handle the base */
2584         switch (type) {
2585         case OBJ_OFS_DELTA:
2586         case OBJ_REF_DELTA:
2587                 if (data)
2588                         die("BUG: unpack_entry: left loop at a valid delta");
2589                 break;
2590         case OBJ_COMMIT:
2591         case OBJ_TREE:
2592         case OBJ_BLOB:
2593         case OBJ_TAG:
2594                 if (!base_from_cache)
2595                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2596                 break;
2597         default:
2598                 data = NULL;
2599                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2600                       type, (uintmax_t)obj_offset, p->pack_name);
2601         }
2602
2603         /* PHASE 3: apply deltas in order */
2604
2605         /* invariants:
2606          *   'data' holds the base data, or NULL if there was corruption
2607          */
2608         while (delta_stack_nr) {
2609                 void *delta_data;
2610                 void *base = data;
2611                 void *external_base = NULL;
2612                 unsigned long delta_size, base_size = size;
2613                 int i;
2614
2615                 data = NULL;
2616
2617                 if (base)
2618                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2619
2620                 if (!base) {
2621                         /*
2622                          * We're probably in deep shit, but let's try to fetch
2623                          * the required base anyway from another pack or loose.
2624                          * This is costly but should happen only in the presence
2625                          * of a corrupted pack, and is better than failing outright.
2626                          */
2627                         struct revindex_entry *revidx;
2628                         const unsigned char *base_sha1;
2629                         revidx = find_pack_revindex(p, obj_offset);
2630                         if (revidx) {
2631                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2632                                 error("failed to read delta base object %s"
2633                                       " at offset %"PRIuMAX" from %s",
2634                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2635                                       p->pack_name);
2636                                 mark_bad_packed_object(p, base_sha1);
2637                                 base = read_object(base_sha1, &type, &base_size);
2638                                 external_base = base;
2639                         }
2640                 }
2641
2642                 i = --delta_stack_nr;
2643                 obj_offset = delta_stack[i].obj_offset;
2644                 curpos = delta_stack[i].curpos;
2645                 delta_size = delta_stack[i].size;
2646
2647                 if (!base)
2648                         continue;
2649
2650                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2651
2652                 if (!delta_data) {
2653                         error("failed to unpack compressed delta "
2654                               "at offset %"PRIuMAX" from %s",
2655                               (uintmax_t)curpos, p->pack_name);
2656                         data = NULL;
2657                         free(external_base);
2658                         continue;
2659                 }
2660
2661                 data = patch_delta(base, base_size,
2662                                    delta_data, delta_size,
2663                                    &size);
2664
2665                 /*
2666                  * We could not apply the delta; warn the user, but keep going.
2667                  * Our failure will be noticed either in the next iteration of
2668                  * the loop, or if this is the final delta, in the caller when
2669                  * we return NULL. Those code paths will take care of making
2670                  * a more explicit warning and retrying with another copy of
2671                  * the object.
2672                  */
2673                 if (!data)
2674                         error("failed to apply delta");
2675
2676                 free(delta_data);
2677                 free(external_base);
2678         }
2679
2680         if (final_type)
2681                 *final_type = type;
2682         if (final_size)
2683                 *final_size = size;
2684
2685 out:
2686         unuse_pack(&w_curs);
2687
2688         if (delta_stack != small_delta_stack)
2689                 free(delta_stack);
2690
2691         return data;
2692 }
2693
2694 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2695                                             uint32_t n)
2696 {
2697         const unsigned char *index = p->index_data;
2698         if (!index) {
2699                 if (open_pack_index(p))
2700                         return NULL;
2701                 index = p->index_data;
2702         }
2703         if (n >= p->num_objects)
2704                 return NULL;
2705         index += 4 * 256;
2706         if (p->index_version == 1) {
2707                 return index + 24 * n + 4;
2708         } else {
2709                 index += 8;
2710                 return index + 20 * n;
2711         }
2712 }
2713
2714 const struct object_id *nth_packed_object_oid(struct object_id *oid,
2715                                               struct packed_git *p,
2716                                               uint32_t n)
2717 {
2718         const unsigned char *hash = nth_packed_object_sha1(p, n);
2719         if (!hash)
2720                 return NULL;
2721         hashcpy(oid->hash, hash);
2722         return oid;
2723 }
2724
2725 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2726 {
2727         const unsigned char *ptr = vptr;
2728         const unsigned char *start = p->index_data;
2729         const unsigned char *end = start + p->index_size;
2730         if (ptr < start)
2731                 die(_("offset before start of pack index for %s (corrupt index?)"),
2732                     p->pack_name);
2733         /* No need to check for underflow; .idx files must be at least 8 bytes */
2734         if (ptr >= end - 8)
2735                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2736                     p->pack_name);
2737 }
2738
2739 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2740 {
2741         const unsigned char *index = p->index_data;
2742         index += 4 * 256;
2743         if (p->index_version == 1) {
2744                 return ntohl(*((uint32_t *)(index + 24 * n)));
2745         } else {
2746                 uint32_t off;
2747                 index += 8 + p->num_objects * (20 + 4);
2748                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2749                 if (!(off & 0x80000000))
2750                         return off;
2751                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2752                 check_pack_index_ptr(p, index);
2753                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2754                                    ntohl(*((uint32_t *)(index + 4)));
2755         }
2756 }
2757
2758 off_t find_pack_entry_one(const unsigned char *sha1,
2759                                   struct packed_git *p)
2760 {
2761         const uint32_t *level1_ofs = p->index_data;
2762         const unsigned char *index = p->index_data;
2763         unsigned hi, lo, stride;
2764         static int use_lookup = -1;
2765         static int debug_lookup = -1;
2766
2767         if (debug_lookup < 0)
2768                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2769
2770         if (!index) {
2771                 if (open_pack_index(p))
2772                         return 0;
2773                 level1_ofs = p->index_data;
2774                 index = p->index_data;
2775         }
2776         if (p->index_version > 1) {
2777                 level1_ofs += 2;
2778                 index += 8;
2779         }
2780         index += 4 * 256;
2781         hi = ntohl(level1_ofs[*sha1]);
2782         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2783         if (p->index_version > 1) {
2784                 stride = 20;
2785         } else {
2786                 stride = 24;
2787                 index += 4;
2788         }
2789
2790         if (debug_lookup)
2791                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2792                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2793
2794         if (use_lookup < 0)
2795                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2796         if (use_lookup) {
2797                 int pos = sha1_entry_pos(index, stride, 0,
2798                                          lo, hi, p->num_objects, sha1);
2799                 if (pos < 0)
2800                         return 0;
2801                 return nth_packed_object_offset(p, pos);
2802         }
2803
2804         while (lo < hi) {
2805                 unsigned mi = (lo + hi) / 2;
2806                 int cmp = hashcmp(index + mi * stride, sha1);
2807
2808                 if (debug_lookup)
2809                         printf("lo %u hi %u rg %u mi %u\n",
2810                                lo, hi, hi - lo, mi);
2811                 if (!cmp)
2812                         return nth_packed_object_offset(p, mi);
2813                 if (cmp > 0)
2814                         hi = mi;
2815                 else
2816                         lo = mi+1;
2817         }
2818         return 0;
2819 }
2820
2821 int is_pack_valid(struct packed_git *p)
2822 {
2823         /* An already open pack is known to be valid. */
2824         if (p->pack_fd != -1)
2825                 return 1;
2826
2827         /* If the pack has one window completely covering the
2828          * file size, the pack is known to be valid even if
2829          * the descriptor is not currently open.
2830          */
2831         if (p->windows) {
2832                 struct pack_window *w = p->windows;
2833
2834                 if (!w->offset && w->len == p->pack_size)
2835                         return 1;
2836         }
2837
2838         /* Force the pack to open to prove its valid. */
2839         return !open_packed_git(p);
2840 }
2841
2842 static int fill_pack_entry(const unsigned char *sha1,
2843                            struct pack_entry *e,
2844                            struct packed_git *p)
2845 {
2846         off_t offset;
2847
2848         if (p->num_bad_objects) {
2849                 unsigned i;
2850                 for (i = 0; i < p->num_bad_objects; i++)
2851                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2852                                 return 0;
2853         }
2854
2855         offset = find_pack_entry_one(sha1, p);
2856         if (!offset)
2857                 return 0;
2858
2859         /*
2860          * We are about to tell the caller where they can locate the
2861          * requested object.  We better make sure the packfile is
2862          * still here and can be accessed before supplying that
2863          * answer, as it may have been deleted since the index was
2864          * loaded!
2865          */
2866         if (!is_pack_valid(p))
2867                 return 0;
2868         e->offset = offset;
2869         e->p = p;
2870         hashcpy(e->sha1, sha1);
2871         return 1;
2872 }
2873
2874 /*
2875  * Iff a pack file contains the object named by sha1, return true and
2876  * store its location to e.
2877  */
2878 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2879 {
2880         struct mru_entry *p;
2881
2882         prepare_packed_git();
2883         if (!packed_git)
2884                 return 0;
2885
2886         for (p = packed_git_mru->head; p; p = p->next) {
2887                 if (fill_pack_entry(sha1, e, p->item)) {
2888                         mru_mark(packed_git_mru, p);
2889                         return 1;
2890                 }
2891         }
2892         return 0;
2893 }
2894
2895 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2896                                   struct packed_git *packs)
2897 {
2898         struct packed_git *p;
2899
2900         for (p = packs; p; p = p->next) {
2901                 if (find_pack_entry_one(sha1, p))
2902                         return p;
2903         }
2904         return NULL;
2905
2906 }
2907
2908 static int sha1_loose_object_info(const unsigned char *sha1,
2909                                   struct object_info *oi,
2910                                   int flags)
2911 {
2912         int status = 0;
2913         unsigned long mapsize;
2914         void *map;
2915         git_zstream stream;
2916         char hdr[32];
2917         struct strbuf hdrbuf = STRBUF_INIT;
2918         unsigned long size_scratch;
2919
2920         if (oi->delta_base_sha1)
2921                 hashclr(oi->delta_base_sha1);
2922
2923         /*
2924          * If we don't care about type or size, then we don't
2925          * need to look inside the object at all. Note that we
2926          * do not optimize out the stat call, even if the
2927          * caller doesn't care about the disk-size, since our
2928          * return value implicitly indicates whether the
2929          * object even exists.
2930          */
2931         if (!oi->typep && !oi->typename && !oi->sizep && !oi->contentp) {
2932                 const char *path;
2933                 struct stat st;
2934                 if (stat_sha1_file(sha1, &st, &path) < 0)
2935                         return -1;
2936                 if (oi->disk_sizep)
2937                         *oi->disk_sizep = st.st_size;
2938                 return 0;
2939         }
2940
2941         map = map_sha1_file(sha1, &mapsize);
2942         if (!map)
2943                 return -1;
2944
2945         if (!oi->sizep)
2946                 oi->sizep = &size_scratch;
2947
2948         if (oi->disk_sizep)
2949                 *oi->disk_sizep = mapsize;
2950         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE)) {
2951                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2952                         status = error("unable to unpack %s header with --allow-unknown-type",
2953                                        sha1_to_hex(sha1));
2954         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2955                 status = error("unable to unpack %s header",
2956                                sha1_to_hex(sha1));
2957         if (status < 0)
2958                 ; /* Do nothing */
2959         else if (hdrbuf.len) {
2960                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2961                         status = error("unable to parse %s header with --allow-unknown-type",
2962                                        sha1_to_hex(sha1));
2963         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2964                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2965
2966         if (status >= 0 && oi->contentp)
2967                 *oi->contentp = unpack_sha1_rest(&stream, hdr,
2968                                                  *oi->sizep, sha1);
2969         else
2970                 git_inflate_end(&stream);
2971
2972         munmap(map, mapsize);
2973         if (status && oi->typep)
2974                 *oi->typep = status;
2975         if (oi->sizep == &size_scratch)
2976                 oi->sizep = NULL;
2977         strbuf_release(&hdrbuf);
2978         return (status < 0) ? status : 0;
2979 }
2980
2981 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2982 {
2983         static struct object_info blank_oi = OBJECT_INFO_INIT;
2984         struct pack_entry e;
2985         int rtype;
2986         const unsigned char *real = (flags & OBJECT_INFO_LOOKUP_REPLACE) ?
2987                                     lookup_replace_object(sha1) :
2988                                     sha1;
2989
2990         if (!oi)
2991                 oi = &blank_oi;
2992
2993         if (!(flags & OBJECT_INFO_SKIP_CACHED)) {
2994                 struct cached_object *co = find_cached_object(real);
2995                 if (co) {
2996                         if (oi->typep)
2997                                 *(oi->typep) = co->type;
2998                         if (oi->sizep)
2999                                 *(oi->sizep) = co->size;
3000                         if (oi->disk_sizep)
3001                                 *(oi->disk_sizep) = 0;
3002                         if (oi->delta_base_sha1)
3003                                 hashclr(oi->delta_base_sha1);
3004                         if (oi->typename)
3005                                 strbuf_addstr(oi->typename, typename(co->type));
3006                         if (oi->contentp)
3007                                 *oi->contentp = xmemdupz(co->buf, co->size);
3008                         oi->whence = OI_CACHED;
3009                         return 0;
3010                 }
3011         }
3012
3013         if (!find_pack_entry(real, &e)) {
3014                 /* Most likely it's a loose object. */
3015                 if (!sha1_loose_object_info(real, oi, flags)) {
3016                         oi->whence = OI_LOOSE;
3017                         return 0;
3018                 }
3019
3020                 /* Not a loose object; someone else may have just packed it. */
3021                 if (flags & OBJECT_INFO_QUICK) {
3022                         return -1;
3023                 } else {
3024                         reprepare_packed_git();
3025                         if (!find_pack_entry(real, &e))
3026                                 return -1;
3027                 }
3028         }
3029
3030         if (oi == &blank_oi)
3031                 /*
3032                  * We know that the caller doesn't actually need the
3033                  * information below, so return early.
3034                  */
3035                 return 0;
3036
3037         rtype = packed_object_info(e.p, e.offset, oi);
3038         if (rtype < 0) {
3039                 mark_bad_packed_object(e.p, real);
3040                 return sha1_object_info_extended(real, oi, 0);
3041         } else if (in_delta_base_cache(e.p, e.offset)) {
3042                 oi->whence = OI_DBCACHED;
3043         } else {
3044                 oi->whence = OI_PACKED;
3045                 oi->u.packed.offset = e.offset;
3046                 oi->u.packed.pack = e.p;
3047                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
3048                                          rtype == OBJ_OFS_DELTA);
3049         }
3050
3051         return 0;
3052 }
3053
3054 /* returns enum object_type or negative */
3055 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
3056 {
3057         enum object_type type;
3058         struct object_info oi = OBJECT_INFO_INIT;
3059
3060         oi.typep = &type;
3061         oi.sizep = sizep;
3062         if (sha1_object_info_extended(sha1, &oi,
3063                                       OBJECT_INFO_LOOKUP_REPLACE) < 0)
3064                 return -1;
3065         return type;
3066 }
3067
3068 static void *read_packed_sha1(const unsigned char *sha1,
3069                               enum object_type *type, unsigned long *size)
3070 {
3071         struct pack_entry e;
3072         void *data;
3073
3074         if (!find_pack_entry(sha1, &e))
3075                 return NULL;
3076         data = cache_or_unpack_entry(e.p, e.offset, size, type);
3077         if (!data) {
3078                 /*
3079                  * We're probably in deep shit, but let's try to fetch
3080                  * the required object anyway from another pack or loose.
3081                  * This should happen only in the presence of a corrupted
3082                  * pack, and is better than failing outright.
3083                  */
3084                 error("failed to read object %s at offset %"PRIuMAX" from %s",
3085                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
3086                 mark_bad_packed_object(e.p, sha1);
3087                 data = read_object(sha1, type, size);
3088         }
3089         return data;
3090 }
3091
3092 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
3093                       unsigned char *sha1)
3094 {
3095         struct cached_object *co;
3096
3097         hash_sha1_file(buf, len, typename(type), sha1);
3098         if (has_sha1_file(sha1) || find_cached_object(sha1))
3099                 return 0;
3100         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
3101         co = &cached_objects[cached_object_nr++];
3102         co->size = len;
3103         co->type = type;
3104         co->buf = xmalloc(len);
3105         memcpy(co->buf, buf, len);
3106         hashcpy(co->sha1, sha1);
3107         return 0;
3108 }
3109
3110 static void *read_object(const unsigned char *sha1, enum object_type *type,
3111                          unsigned long *size)
3112 {
3113         struct object_info oi = OBJECT_INFO_INIT;
3114         void *content;
3115         oi.typep = type;
3116         oi.sizep = size;
3117         oi.contentp = &content;
3118
3119         if (sha1_object_info_extended(sha1, &oi, 0) < 0)
3120                 return NULL;
3121         return content;
3122 }
3123
3124 /*
3125  * This function dies on corrupt objects; the callers who want to
3126  * deal with them should arrange to call read_object() and give error
3127  * messages themselves.
3128  */
3129 void *read_sha1_file_extended(const unsigned char *sha1,
3130                               enum object_type *type,
3131                               unsigned long *size,
3132                               int lookup_replace)
3133 {
3134         void *data;
3135         const struct packed_git *p;
3136         const char *path;
3137         struct stat st;
3138         const unsigned char *repl = lookup_replace ? lookup_replace_object(sha1)
3139                                                    : sha1;
3140
3141         errno = 0;
3142         data = read_object(repl, type, size);
3143         if (data)
3144                 return data;
3145
3146         if (errno && errno != ENOENT)
3147                 die_errno("failed to read object %s", sha1_to_hex(sha1));
3148
3149         /* die if we replaced an object with one that does not exist */
3150         if (repl != sha1)
3151                 die("replacement %s not found for %s",
3152                     sha1_to_hex(repl), sha1_to_hex(sha1));
3153
3154         if (!stat_sha1_file(repl, &st, &path))
3155                 die("loose object %s (stored in %s) is corrupt",
3156                     sha1_to_hex(repl), path);
3157
3158         if ((p = has_packed_and_bad(repl)) != NULL)
3159                 die("packed object %s (stored in %s) is corrupt",
3160                     sha1_to_hex(repl), p->pack_name);
3161
3162         return NULL;
3163 }
3164
3165 void *read_object_with_reference(const unsigned char *sha1,
3166                                  const char *required_type_name,
3167                                  unsigned long *size,
3168                                  unsigned char *actual_sha1_return)
3169 {
3170         enum object_type type, required_type;
3171         void *buffer;
3172         unsigned long isize;
3173         unsigned char actual_sha1[20];
3174
3175         required_type = type_from_string(required_type_name);
3176         hashcpy(actual_sha1, sha1);
3177         while (1) {
3178                 int ref_length = -1;
3179                 const char *ref_type = NULL;
3180
3181                 buffer = read_sha1_file(actual_sha1, &type, &isize);
3182                 if (!buffer)
3183                         return NULL;
3184                 if (type == required_type) {
3185                         *size = isize;
3186                         if (actual_sha1_return)
3187                                 hashcpy(actual_sha1_return, actual_sha1);
3188                         return buffer;
3189                 }
3190                 /* Handle references */
3191                 else if (type == OBJ_COMMIT)
3192                         ref_type = "tree ";
3193                 else if (type == OBJ_TAG)
3194                         ref_type = "object ";
3195                 else {
3196                         free(buffer);
3197                         return NULL;
3198                 }
3199                 ref_length = strlen(ref_type);
3200
3201                 if (ref_length + 40 > isize ||
3202                     memcmp(buffer, ref_type, ref_length) ||
3203                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
3204                         free(buffer);
3205                         return NULL;
3206                 }
3207                 free(buffer);
3208                 /* Now we have the ID of the referred-to object in
3209                  * actual_sha1.  Check again. */
3210         }
3211 }
3212
3213 static void write_sha1_file_prepare(const void *buf, unsigned long len,
3214                                     const char *type, unsigned char *sha1,
3215                                     char *hdr, int *hdrlen)
3216 {
3217         git_SHA_CTX c;
3218
3219         /* Generate the header */
3220         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
3221
3222         /* Sha1.. */
3223         git_SHA1_Init(&c);
3224         git_SHA1_Update(&c, hdr, *hdrlen);
3225         git_SHA1_Update(&c, buf, len);
3226         git_SHA1_Final(sha1, &c);
3227 }
3228
3229 /*
3230  * Move the just written object into its final resting place.
3231  */
3232 int finalize_object_file(const char *tmpfile, const char *filename)
3233 {
3234         int ret = 0;
3235
3236         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3237                 goto try_rename;
3238         else if (link(tmpfile, filename))
3239                 ret = errno;
3240
3241         /*
3242          * Coda hack - coda doesn't like cross-directory links,
3243          * so we fall back to a rename, which will mean that it
3244          * won't be able to check collisions, but that's not a
3245          * big deal.
3246          *
3247          * The same holds for FAT formatted media.
3248          *
3249          * When this succeeds, we just return.  We have nothing
3250          * left to unlink.
3251          */
3252         if (ret && ret != EEXIST) {
3253         try_rename:
3254                 if (!rename(tmpfile, filename))
3255                         goto out;
3256                 ret = errno;
3257         }
3258         unlink_or_warn(tmpfile);
3259         if (ret) {
3260                 if (ret != EEXIST) {
3261                         return error_errno("unable to write sha1 filename %s", filename);
3262                 }
3263                 /* FIXME!!! Collision check here ? */
3264         }
3265
3266 out:
3267         if (adjust_shared_perm(filename))
3268                 return error("unable to set permission to '%s'", filename);
3269         return 0;
3270 }
3271
3272 static int write_buffer(int fd, const void *buf, size_t len)
3273 {
3274         if (write_in_full(fd, buf, len) < 0)
3275                 return error_errno("file write error");
3276         return 0;
3277 }
3278
3279 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3280                    unsigned char *sha1)
3281 {
3282         char hdr[32];
3283         int hdrlen = sizeof(hdr);
3284         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3285         return 0;
3286 }
3287
3288 /* Finalize a file on disk, and close it. */
3289 static void close_sha1_file(int fd)
3290 {
3291         if (fsync_object_files)
3292                 fsync_or_die(fd, "sha1 file");
3293         if (close(fd) != 0)
3294                 die_errno("error when closing sha1 file");
3295 }
3296
3297 /* Size of directory component, including the ending '/' */
3298 static inline int directory_size(const char *filename)
3299 {
3300         const char *s = strrchr(filename, '/');
3301         if (!s)
3302                 return 0;
3303         return s - filename + 1;
3304 }
3305
3306 /*
3307  * This creates a temporary file in the same directory as the final
3308  * 'filename'
3309  *
3310  * We want to avoid cross-directory filename renames, because those
3311  * can have problems on various filesystems (FAT, NFS, Coda).
3312  */
3313 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3314 {
3315         int fd, dirlen = directory_size(filename);
3316
3317         strbuf_reset(tmp);
3318         strbuf_add(tmp, filename, dirlen);
3319         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3320         fd = git_mkstemp_mode(tmp->buf, 0444);
3321         if (fd < 0 && dirlen && errno == ENOENT) {
3322                 /*
3323                  * Make sure the directory exists; note that the contents
3324                  * of the buffer are undefined after mkstemp returns an
3325                  * error, so we have to rewrite the whole buffer from
3326                  * scratch.
3327                  */
3328                 strbuf_reset(tmp);
3329                 strbuf_add(tmp, filename, dirlen - 1);
3330                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3331                         return -1;
3332                 if (adjust_shared_perm(tmp->buf))
3333                         return -1;
3334
3335                 /* Try again */
3336                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3337                 fd = git_mkstemp_mode(tmp->buf, 0444);
3338         }
3339         return fd;
3340 }
3341
3342 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3343                               const void *buf, unsigned long len, time_t mtime)
3344 {
3345         int fd, ret;
3346         unsigned char compressed[4096];
3347         git_zstream stream;
3348         git_SHA_CTX c;
3349         unsigned char parano_sha1[20];
3350         static struct strbuf tmp_file = STRBUF_INIT;
3351         const char *filename = sha1_file_name(sha1);
3352
3353         fd = create_tmpfile(&tmp_file, filename);
3354         if (fd < 0) {
3355                 if (errno == EACCES)
3356                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3357                 else
3358                         return error_errno("unable to create temporary file");
3359         }
3360
3361         /* Set it up */
3362         git_deflate_init(&stream, zlib_compression_level);
3363         stream.next_out = compressed;
3364         stream.avail_out = sizeof(compressed);
3365         git_SHA1_Init(&c);
3366
3367         /* First header.. */
3368         stream.next_in = (unsigned char *)hdr;
3369         stream.avail_in = hdrlen;
3370         while (git_deflate(&stream, 0) == Z_OK)
3371                 ; /* nothing */
3372         git_SHA1_Update(&c, hdr, hdrlen);
3373
3374         /* Then the data itself.. */
3375         stream.next_in = (void *)buf;
3376         stream.avail_in = len;
3377         do {
3378                 unsigned char *in0 = stream.next_in;
3379                 ret = git_deflate(&stream, Z_FINISH);
3380                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3381                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3382                         die("unable to write sha1 file");
3383                 stream.next_out = compressed;
3384                 stream.avail_out = sizeof(compressed);
3385         } while (ret == Z_OK);
3386
3387         if (ret != Z_STREAM_END)
3388                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3389         ret = git_deflate_end_gently(&stream);
3390         if (ret != Z_OK)
3391                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3392         git_SHA1_Final(parano_sha1, &c);
3393         if (hashcmp(sha1, parano_sha1) != 0)
3394                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3395
3396         close_sha1_file(fd);
3397
3398         if (mtime) {
3399                 struct utimbuf utb;
3400                 utb.actime = mtime;
3401                 utb.modtime = mtime;
3402                 if (utime(tmp_file.buf, &utb) < 0)
3403                         warning_errno("failed utime() on %s", tmp_file.buf);
3404         }
3405
3406         return finalize_object_file(tmp_file.buf, filename);
3407 }
3408
3409 static int freshen_loose_object(const unsigned char *sha1)
3410 {
3411         return check_and_freshen(sha1, 1);
3412 }
3413
3414 static int freshen_packed_object(const unsigned char *sha1)
3415 {
3416         struct pack_entry e;
3417         if (!find_pack_entry(sha1, &e))
3418                 return 0;
3419         if (e.p->freshened)
3420                 return 1;
3421         if (!freshen_file(e.p->pack_name))
3422                 return 0;
3423         e.p->freshened = 1;
3424         return 1;
3425 }
3426
3427 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3428 {
3429         char hdr[32];
3430         int hdrlen = sizeof(hdr);
3431
3432         /* Normally if we have it in the pack then we do not bother writing
3433          * it out into .git/objects/??/?{38} file.
3434          */
3435         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3436         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3437                 return 0;
3438         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3439 }
3440
3441 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3442                              unsigned char *sha1, unsigned flags)
3443 {
3444         char *header;
3445         int hdrlen, status = 0;
3446
3447         /* type string, SP, %lu of the length plus NUL must fit this */
3448         hdrlen = strlen(type) + 32;
3449         header = xmalloc(hdrlen);
3450         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3451
3452         if (!(flags & HASH_WRITE_OBJECT))
3453                 goto cleanup;
3454         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3455                 goto cleanup;
3456         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3457
3458 cleanup:
3459         free(header);
3460         return status;
3461 }
3462
3463 int force_object_loose(const unsigned char *sha1, time_t mtime)
3464 {
3465         void *buf;
3466         unsigned long len;
3467         enum object_type type;
3468         char hdr[32];
3469         int hdrlen;
3470         int ret;
3471
3472         if (has_loose_object(sha1))
3473                 return 0;
3474         buf = read_packed_sha1(sha1, &type, &len);
3475         if (!buf)
3476                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3477         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3478         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3479         free(buf);
3480
3481         return ret;
3482 }
3483
3484 int has_pack_index(const unsigned char *sha1)
3485 {
3486         struct stat st;
3487         if (stat(sha1_pack_index_name(sha1), &st))
3488                 return 0;
3489         return 1;
3490 }
3491
3492 int has_sha1_pack(const unsigned char *sha1)
3493 {
3494         struct pack_entry e;
3495         return find_pack_entry(sha1, &e);
3496 }
3497
3498 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3499 {
3500         if (!startup_info->have_repository)
3501                 return 0;
3502         return sha1_object_info_extended(sha1, NULL,
3503                                          flags | OBJECT_INFO_SKIP_CACHED) >= 0;
3504 }
3505
3506 int has_object_file(const struct object_id *oid)
3507 {
3508         return has_sha1_file(oid->hash);
3509 }
3510
3511 int has_object_file_with_flags(const struct object_id *oid, int flags)
3512 {
3513         return has_sha1_file_with_flags(oid->hash, flags);
3514 }
3515
3516 static void check_tree(const void *buf, size_t size)
3517 {
3518         struct tree_desc desc;
3519         struct name_entry entry;
3520
3521         init_tree_desc(&desc, buf, size);
3522         while (tree_entry(&desc, &entry))
3523                 /* do nothing
3524                  * tree_entry() will die() on malformed entries */
3525                 ;
3526 }
3527
3528 static void check_commit(const void *buf, size_t size)
3529 {
3530         struct commit c;
3531         memset(&c, 0, sizeof(c));
3532         if (parse_commit_buffer(&c, buf, size))
3533                 die("corrupt commit");
3534 }
3535
3536 static void check_tag(const void *buf, size_t size)
3537 {
3538         struct tag t;
3539         memset(&t, 0, sizeof(t));
3540         if (parse_tag_buffer(&t, buf, size))
3541                 die("corrupt tag");
3542 }
3543
3544 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3545                      enum object_type type,
3546                      const char *path, unsigned flags)
3547 {
3548         int ret, re_allocated = 0;
3549         int write_object = flags & HASH_WRITE_OBJECT;
3550
3551         if (!type)
3552                 type = OBJ_BLOB;
3553
3554         /*
3555          * Convert blobs to git internal format
3556          */
3557         if ((type == OBJ_BLOB) && path) {
3558                 struct strbuf nbuf = STRBUF_INIT;
3559                 if (convert_to_git(&the_index, path, buf, size, &nbuf,
3560                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3561                         buf = strbuf_detach(&nbuf, &size);
3562                         re_allocated = 1;
3563                 }
3564         }
3565         if (flags & HASH_FORMAT_CHECK) {
3566                 if (type == OBJ_TREE)
3567                         check_tree(buf, size);
3568                 if (type == OBJ_COMMIT)
3569                         check_commit(buf, size);
3570                 if (type == OBJ_TAG)
3571                         check_tag(buf, size);
3572         }
3573
3574         if (write_object)
3575                 ret = write_sha1_file(buf, size, typename(type), sha1);
3576         else
3577                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3578         if (re_allocated)
3579                 free(buf);
3580         return ret;
3581 }
3582
3583 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3584                                      const char *path, unsigned flags)
3585 {
3586         int ret;
3587         const int write_object = flags & HASH_WRITE_OBJECT;
3588         struct strbuf sbuf = STRBUF_INIT;
3589
3590         assert(path);
3591         assert(would_convert_to_git_filter_fd(path));
3592
3593         convert_to_git_filter_fd(&the_index, path, fd, &sbuf,
3594                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3595
3596         if (write_object)
3597                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3598                                       sha1);
3599         else
3600                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3601                                      sha1);
3602         strbuf_release(&sbuf);
3603         return ret;
3604 }
3605
3606 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3607                       const char *path, unsigned flags)
3608 {
3609         struct strbuf sbuf = STRBUF_INIT;
3610         int ret;
3611
3612         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3613                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3614         else
3615                 ret = -1;
3616         strbuf_release(&sbuf);
3617         return ret;
3618 }
3619
3620 #define SMALL_FILE_SIZE (32*1024)
3621
3622 static int index_core(unsigned char *sha1, int fd, size_t size,
3623                       enum object_type type, const char *path,
3624                       unsigned flags)
3625 {
3626         int ret;
3627
3628         if (!size) {
3629                 ret = index_mem(sha1, "", size, type, path, flags);
3630         } else if (size <= SMALL_FILE_SIZE) {
3631                 char *buf = xmalloc(size);
3632                 if (size == read_in_full(fd, buf, size))
3633                         ret = index_mem(sha1, buf, size, type, path, flags);
3634                 else
3635                         ret = error_errno("short read");
3636                 free(buf);
3637         } else {
3638                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3639                 ret = index_mem(sha1, buf, size, type, path, flags);
3640                 munmap(buf, size);
3641         }
3642         return ret;
3643 }
3644
3645 /*
3646  * This creates one packfile per large blob unless bulk-checkin
3647  * machinery is "plugged".
3648  *
3649  * This also bypasses the usual "convert-to-git" dance, and that is on
3650  * purpose. We could write a streaming version of the converting
3651  * functions and insert that before feeding the data to fast-import
3652  * (or equivalent in-core API described above). However, that is
3653  * somewhat complicated, as we do not know the size of the filter
3654  * result, which we need to know beforehand when writing a git object.
3655  * Since the primary motivation for trying to stream from the working
3656  * tree file and to avoid mmaping it in core is to deal with large
3657  * binary blobs, they generally do not want to get any conversion, and
3658  * callers should avoid this code path when filters are requested.
3659  */
3660 static int index_stream(unsigned char *sha1, int fd, size_t size,
3661                         enum object_type type, const char *path,
3662                         unsigned flags)
3663 {
3664         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3665 }
3666
3667 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3668              enum object_type type, const char *path, unsigned flags)
3669 {
3670         int ret;
3671
3672         /*
3673          * Call xsize_t() only when needed to avoid potentially unnecessary
3674          * die() for large files.
3675          */
3676         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3677                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3678         else if (!S_ISREG(st->st_mode))
3679                 ret = index_pipe(sha1, fd, type, path, flags);
3680         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3681                  (path && would_convert_to_git(&the_index, path)))
3682                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3683                                  flags);
3684         else
3685                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3686                                    flags);
3687         close(fd);
3688         return ret;
3689 }
3690
3691 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3692 {
3693         int fd;
3694         struct strbuf sb = STRBUF_INIT;
3695
3696         switch (st->st_mode & S_IFMT) {
3697         case S_IFREG:
3698                 fd = open(path, O_RDONLY);
3699                 if (fd < 0)
3700                         return error_errno("open(\"%s\")", path);
3701                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3702                         return error("%s: failed to insert into database",
3703                                      path);
3704                 break;
3705         case S_IFLNK:
3706                 if (strbuf_readlink(&sb, path, st->st_size))
3707                         return error_errno("readlink(\"%s\")", path);
3708                 if (!(flags & HASH_WRITE_OBJECT))
3709                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3710                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3711                         return error("%s: failed to insert into database",
3712                                      path);
3713                 strbuf_release(&sb);
3714                 break;
3715         case S_IFDIR:
3716                 return resolve_gitlink_ref(path, "HEAD", sha1);
3717         default:
3718                 return error("%s: unsupported file type", path);
3719         }
3720         return 0;
3721 }
3722
3723 int read_pack_header(int fd, struct pack_header *header)
3724 {
3725         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3726                 /* "eof before pack header was fully read" */
3727                 return PH_ERROR_EOF;
3728
3729         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3730                 /* "protocol error (pack signature mismatch detected)" */
3731                 return PH_ERROR_PACK_SIGNATURE;
3732         if (!pack_version_ok(header->hdr_version))
3733                 /* "protocol error (pack version unsupported)" */
3734                 return PH_ERROR_PROTOCOL;
3735         return 0;
3736 }
3737
3738 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3739 {
3740         enum object_type type = sha1_object_info(sha1, NULL);
3741         if (type < 0)
3742                 die("%s is not a valid object", sha1_to_hex(sha1));
3743         if (type != expect)
3744                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3745                     typename(expect));
3746 }
3747
3748 int for_each_file_in_obj_subdir(unsigned int subdir_nr,
3749                                 struct strbuf *path,
3750                                 each_loose_object_fn obj_cb,
3751                                 each_loose_cruft_fn cruft_cb,
3752                                 each_loose_subdir_fn subdir_cb,
3753                                 void *data)
3754 {
3755         size_t origlen, baselen;
3756         DIR *dir;
3757         struct dirent *de;
3758         int r = 0;
3759
3760         if (subdir_nr > 0xff)
3761                 BUG("invalid loose object subdirectory: %x", subdir_nr);
3762
3763         origlen = path->len;
3764         strbuf_complete(path, '/');
3765         strbuf_addf(path, "%02x", subdir_nr);
3766         baselen = path->len;
3767
3768         dir = opendir(path->buf);
3769         if (!dir) {
3770                 if (errno != ENOENT)
3771                         r = error_errno("unable to open %s", path->buf);
3772                 strbuf_setlen(path, origlen);
3773                 return r;
3774         }
3775
3776         while ((de = readdir(dir))) {
3777                 if (is_dot_or_dotdot(de->d_name))
3778                         continue;
3779
3780                 strbuf_setlen(path, baselen);
3781                 strbuf_addf(path, "/%s", de->d_name);
3782
3783                 if (strlen(de->d_name) == GIT_SHA1_HEXSZ - 2)  {
3784                         char hex[GIT_MAX_HEXSZ+1];
3785                         struct object_id oid;
3786
3787                         xsnprintf(hex, sizeof(hex), "%02x%s",
3788                                   subdir_nr, de->d_name);
3789                         if (!get_oid_hex(hex, &oid)) {
3790                                 if (obj_cb) {
3791                                         r = obj_cb(&oid, path->buf, data);
3792                                         if (r)
3793                                                 break;
3794                                 }
3795                                 continue;
3796                         }
3797                 }
3798
3799                 if (cruft_cb) {
3800                         r = cruft_cb(de->d_name, path->buf, data);
3801                         if (r)
3802                                 break;
3803                 }
3804         }
3805         closedir(dir);
3806
3807         strbuf_setlen(path, baselen);
3808         if (!r && subdir_cb)
3809                 r = subdir_cb(subdir_nr, path->buf, data);
3810
3811         strbuf_setlen(path, origlen);
3812
3813         return r;
3814 }
3815
3816 int for_each_loose_file_in_objdir_buf(struct strbuf *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         int r = 0;
3823         int i;
3824
3825         for (i = 0; i < 256; i++) {
3826                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3827                                                 subdir_cb, data);
3828                 if (r)
3829                         break;
3830         }
3831
3832         return r;
3833 }
3834
3835 int for_each_loose_file_in_objdir(const char *path,
3836                                   each_loose_object_fn obj_cb,
3837                                   each_loose_cruft_fn cruft_cb,
3838                                   each_loose_subdir_fn subdir_cb,
3839                                   void *data)
3840 {
3841         struct strbuf buf = STRBUF_INIT;
3842         int r;
3843
3844         strbuf_addstr(&buf, path);
3845         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3846                                               subdir_cb, data);
3847         strbuf_release(&buf);
3848
3849         return r;
3850 }
3851
3852 struct loose_alt_odb_data {
3853         each_loose_object_fn *cb;
3854         void *data;
3855 };
3856
3857 static int loose_from_alt_odb(struct alternate_object_database *alt,
3858                               void *vdata)
3859 {
3860         struct loose_alt_odb_data *data = vdata;
3861         struct strbuf buf = STRBUF_INIT;
3862         int r;
3863
3864         strbuf_addstr(&buf, alt->path);
3865         r = for_each_loose_file_in_objdir_buf(&buf,
3866                                               data->cb, NULL, NULL,
3867                                               data->data);
3868         strbuf_release(&buf);
3869         return r;
3870 }
3871
3872 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3873 {
3874         struct loose_alt_odb_data alt;
3875         int r;
3876
3877         r = for_each_loose_file_in_objdir(get_object_directory(),
3878                                           cb, NULL, NULL, data);
3879         if (r)
3880                 return r;
3881
3882         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3883                 return 0;
3884
3885         alt.cb = cb;
3886         alt.data = data;
3887         return foreach_alt_odb(loose_from_alt_odb, &alt);
3888 }
3889
3890 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3891 {
3892         uint32_t i;
3893         int r = 0;
3894
3895         for (i = 0; i < p->num_objects; i++) {
3896                 struct object_id oid;
3897
3898                 if (!nth_packed_object_oid(&oid, p, i))
3899                         return error("unable to get sha1 of object %u in %s",
3900                                      i, p->pack_name);
3901
3902                 r = cb(&oid, p, i, data);
3903                 if (r)
3904                         break;
3905         }
3906         return r;
3907 }
3908
3909 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3910 {
3911         struct packed_git *p;
3912         int r = 0;
3913         int pack_errors = 0;
3914
3915         prepare_packed_git();
3916         for (p = packed_git; p; p = p->next) {
3917                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3918                         continue;
3919                 if (open_pack_index(p)) {
3920                         pack_errors = 1;
3921                         continue;
3922                 }
3923                 r = for_each_object_in_pack(p, cb, data);
3924                 if (r)
3925                         break;
3926         }
3927         return r ? r : pack_errors;
3928 }
3929
3930 static int check_stream_sha1(git_zstream *stream,
3931                              const char *hdr,
3932                              unsigned long size,
3933                              const char *path,
3934                              const unsigned char *expected_sha1)
3935 {
3936         git_SHA_CTX c;
3937         unsigned char real_sha1[GIT_MAX_RAWSZ];
3938         unsigned char buf[4096];
3939         unsigned long total_read;
3940         int status = Z_OK;
3941
3942         git_SHA1_Init(&c);
3943         git_SHA1_Update(&c, hdr, stream->total_out);
3944
3945         /*
3946          * We already read some bytes into hdr, but the ones up to the NUL
3947          * do not count against the object's content size.
3948          */
3949         total_read = stream->total_out - strlen(hdr) - 1;
3950
3951         /*
3952          * This size comparison must be "<=" to read the final zlib packets;
3953          * see the comment in unpack_sha1_rest for details.
3954          */
3955         while (total_read <= size &&
3956                (status == Z_OK || status == Z_BUF_ERROR)) {
3957                 stream->next_out = buf;
3958                 stream->avail_out = sizeof(buf);
3959                 if (size - total_read < stream->avail_out)
3960                         stream->avail_out = size - total_read;
3961                 status = git_inflate(stream, Z_FINISH);
3962                 git_SHA1_Update(&c, buf, stream->next_out - buf);
3963                 total_read += stream->next_out - buf;
3964         }
3965         git_inflate_end(stream);
3966
3967         if (status != Z_STREAM_END) {
3968                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
3969                 return -1;
3970         }
3971         if (stream->avail_in) {
3972                 error("garbage at end of loose object '%s'",
3973                       sha1_to_hex(expected_sha1));
3974                 return -1;
3975         }
3976
3977         git_SHA1_Final(real_sha1, &c);
3978         if (hashcmp(expected_sha1, real_sha1)) {
3979                 error("sha1 mismatch for %s (expected %s)", path,
3980                       sha1_to_hex(expected_sha1));
3981                 return -1;
3982         }
3983
3984         return 0;
3985 }
3986
3987 int read_loose_object(const char *path,
3988                       const unsigned char *expected_sha1,
3989                       enum object_type *type,
3990                       unsigned long *size,
3991                       void **contents)
3992 {
3993         int ret = -1;
3994         void *map = NULL;
3995         unsigned long mapsize;
3996         git_zstream stream;
3997         char hdr[32];
3998
3999         *contents = NULL;
4000
4001         map = map_sha1_file_1(path, NULL, &mapsize);
4002         if (!map) {
4003                 error_errno("unable to mmap %s", path);
4004                 goto out;
4005         }
4006
4007         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
4008                 error("unable to unpack header of %s", path);
4009                 goto out;
4010         }
4011
4012         *type = parse_sha1_header(hdr, size);
4013         if (*type < 0) {
4014                 error("unable to parse header of %s", path);
4015                 git_inflate_end(&stream);
4016                 goto out;
4017         }
4018
4019         if (*type == OBJ_BLOB) {
4020                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
4021                         goto out;
4022         } else {
4023                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
4024                 if (!*contents) {
4025                         error("unable to unpack contents of %s", path);
4026                         git_inflate_end(&stream);
4027                         goto out;
4028                 }
4029                 if (check_sha1_signature(expected_sha1, *contents,
4030                                          *size, typename(*type))) {
4031                         error("sha1 mismatch for %s (expected %s)", path,
4032                               sha1_to_hex(expected_sha1));
4033                         free(*contents);
4034                         goto out;
4035                 }
4036         }
4037
4038         ret = 0; /* everything checks out */
4039
4040 out:
4041         if (map)
4042                 munmap(map, mapsize);
4043         return ret;
4044 }