Merge branch 'jk/c99'
[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 int link_alt_odb_entry(const char *entry, const char *relative_base,
351         int depth, const char *normalized_objdir)
352 {
353         struct alternate_object_database *ent;
354         struct strbuf pathbuf = STRBUF_INIT;
355
356         if (!is_absolute_path(entry) && relative_base) {
357                 strbuf_realpath(&pathbuf, relative_base, 1);
358                 strbuf_addch(&pathbuf, '/');
359         }
360         strbuf_addstr(&pathbuf, entry);
361
362         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
363                 error("unable to normalize alternate object path: %s",
364                       pathbuf.buf);
365                 strbuf_release(&pathbuf);
366                 return -1;
367         }
368
369         /*
370          * The trailing slash after the directory name is given by
371          * this function at the end. Remove duplicates.
372          */
373         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
374                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
375
376         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
377                 strbuf_release(&pathbuf);
378                 return -1;
379         }
380
381         ent = alloc_alt_odb(pathbuf.buf);
382
383         /* add the alternate entry */
384         *alt_odb_tail = ent;
385         alt_odb_tail = &(ent->next);
386         ent->next = NULL;
387
388         /* recursively add alternates */
389         read_info_alternates(pathbuf.buf, depth + 1);
390
391         strbuf_release(&pathbuf);
392         return 0;
393 }
394
395 static const char *parse_alt_odb_entry(const char *string,
396                                        int sep,
397                                        struct strbuf *out)
398 {
399         const char *end;
400
401         strbuf_reset(out);
402
403         if (*string == '#') {
404                 /* comment; consume up to next separator */
405                 end = strchrnul(string, sep);
406         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
407                 /*
408                  * quoted path; unquote_c_style has copied the
409                  * data for us and set "end". Broken quoting (e.g.,
410                  * an entry that doesn't end with a quote) falls
411                  * back to the unquoted case below.
412                  */
413         } else {
414                 /* normal, unquoted path */
415                 end = strchrnul(string, sep);
416                 strbuf_add(out, string, end - string);
417         }
418
419         if (*end)
420                 end++;
421         return end;
422 }
423
424 static void link_alt_odb_entries(const char *alt, int len, int sep,
425                                  const char *relative_base, int depth)
426 {
427         struct strbuf objdirbuf = STRBUF_INIT;
428         struct strbuf entry = STRBUF_INIT;
429
430         if (depth > 5) {
431                 error("%s: ignoring alternate object stores, nesting too deep.",
432                                 relative_base);
433                 return;
434         }
435
436         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
437         if (strbuf_normalize_path(&objdirbuf) < 0)
438                 die("unable to normalize object directory: %s",
439                     objdirbuf.buf);
440
441         while (*alt) {
442                 alt = parse_alt_odb_entry(alt, sep, &entry);
443                 if (!entry.len)
444                         continue;
445                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
446         }
447         strbuf_release(&entry);
448         strbuf_release(&objdirbuf);
449 }
450
451 void read_info_alternates(const char * relative_base, int depth)
452 {
453         char *map;
454         size_t mapsz;
455         struct stat st;
456         char *path;
457         int fd;
458
459         path = xstrfmt("%s/info/alternates", relative_base);
460         fd = git_open(path);
461         free(path);
462         if (fd < 0)
463                 return;
464         if (fstat(fd, &st) || (st.st_size == 0)) {
465                 close(fd);
466                 return;
467         }
468         mapsz = xsize_t(st.st_size);
469         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
470         close(fd);
471
472         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
473
474         munmap(map, mapsz);
475 }
476
477 struct alternate_object_database *alloc_alt_odb(const char *dir)
478 {
479         struct alternate_object_database *ent;
480
481         FLEX_ALLOC_STR(ent, path, dir);
482         strbuf_init(&ent->scratch, 0);
483         strbuf_addf(&ent->scratch, "%s/", dir);
484         ent->base_len = ent->scratch.len;
485
486         return ent;
487 }
488
489 void add_to_alternates_file(const char *reference)
490 {
491         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
492         char *alts = git_pathdup("objects/info/alternates");
493         FILE *in, *out;
494
495         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
496         out = fdopen_lock_file(lock, "w");
497         if (!out)
498                 die_errno("unable to fdopen alternates lockfile");
499
500         in = fopen(alts, "r");
501         if (in) {
502                 struct strbuf line = STRBUF_INIT;
503                 int found = 0;
504
505                 while (strbuf_getline(&line, in) != EOF) {
506                         if (!strcmp(reference, line.buf)) {
507                                 found = 1;
508                                 break;
509                         }
510                         fprintf_or_die(out, "%s\n", line.buf);
511                 }
512
513                 strbuf_release(&line);
514                 fclose(in);
515
516                 if (found) {
517                         rollback_lock_file(lock);
518                         lock = NULL;
519                 }
520         }
521         else if (errno != ENOENT)
522                 die_errno("unable to read alternates file");
523
524         if (lock) {
525                 fprintf_or_die(out, "%s\n", reference);
526                 if (commit_lock_file(lock))
527                         die_errno("unable to move new alternates file into place");
528                 if (alt_odb_tail)
529                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
530         }
531         free(alts);
532 }
533
534 void add_to_alternates_memory(const char *reference)
535 {
536         /*
537          * Make sure alternates are initialized, or else our entry may be
538          * overwritten when they are.
539          */
540         prepare_alt_odb();
541
542         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
543 }
544
545 /*
546  * Compute the exact path an alternate is at and returns it. In case of
547  * error NULL is returned and the human readable error is added to `err`
548  * `path` may be relative and should point to $GITDIR.
549  * `err` must not be null.
550  */
551 char *compute_alternate_path(const char *path, struct strbuf *err)
552 {
553         char *ref_git = NULL;
554         const char *repo, *ref_git_s;
555         int seen_error = 0;
556
557         ref_git_s = real_path_if_valid(path);
558         if (!ref_git_s) {
559                 seen_error = 1;
560                 strbuf_addf(err, _("path '%s' does not exist"), path);
561                 goto out;
562         } else
563                 /*
564                  * Beware: read_gitfile(), real_path() and mkpath()
565                  * return static buffer
566                  */
567                 ref_git = xstrdup(ref_git_s);
568
569         repo = read_gitfile(ref_git);
570         if (!repo)
571                 repo = read_gitfile(mkpath("%s/.git", ref_git));
572         if (repo) {
573                 free(ref_git);
574                 ref_git = xstrdup(repo);
575         }
576
577         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
578                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
579                 free(ref_git);
580                 ref_git = ref_git_git;
581         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
582                 struct strbuf sb = STRBUF_INIT;
583                 seen_error = 1;
584                 if (get_common_dir(&sb, ref_git)) {
585                         strbuf_addf(err,
586                                     _("reference repository '%s' as a linked "
587                                       "checkout is not supported yet."),
588                                     path);
589                         goto out;
590                 }
591
592                 strbuf_addf(err, _("reference repository '%s' is not a "
593                                         "local repository."), path);
594                 goto out;
595         }
596
597         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
598                 strbuf_addf(err, _("reference repository '%s' is shallow"),
599                             path);
600                 seen_error = 1;
601                 goto out;
602         }
603
604         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
605                 strbuf_addf(err,
606                             _("reference repository '%s' is grafted"),
607                             path);
608                 seen_error = 1;
609                 goto out;
610         }
611
612 out:
613         if (seen_error) {
614                 FREE_AND_NULL(ref_git);
615         }
616
617         return ref_git;
618 }
619
620 int foreach_alt_odb(alt_odb_fn fn, void *cb)
621 {
622         struct alternate_object_database *ent;
623         int r = 0;
624
625         prepare_alt_odb();
626         for (ent = alt_odb_list; ent; ent = ent->next) {
627                 r = fn(ent, cb);
628                 if (r)
629                         break;
630         }
631         return r;
632 }
633
634 void prepare_alt_odb(void)
635 {
636         const char *alt;
637
638         if (alt_odb_tail)
639                 return;
640
641         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
642         if (!alt) alt = "";
643
644         alt_odb_tail = &alt_odb_list;
645         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
646
647         read_info_alternates(get_object_directory(), 0);
648 }
649
650 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
651 static int freshen_file(const char *fn)
652 {
653         struct utimbuf t;
654         t.actime = t.modtime = time(NULL);
655         return !utime(fn, &t);
656 }
657
658 /*
659  * All of the check_and_freshen functions return 1 if the file exists and was
660  * freshened (if freshening was requested), 0 otherwise. If they return
661  * 0, you should not assume that it is safe to skip a write of the object (it
662  * either does not exist on disk, or has a stale mtime and may be subject to
663  * pruning).
664  */
665 int check_and_freshen_file(const char *fn, int freshen)
666 {
667         if (access(fn, F_OK))
668                 return 0;
669         if (freshen && !freshen_file(fn))
670                 return 0;
671         return 1;
672 }
673
674 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
675 {
676         return check_and_freshen_file(sha1_file_name(sha1), freshen);
677 }
678
679 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
680 {
681         struct alternate_object_database *alt;
682         prepare_alt_odb();
683         for (alt = alt_odb_list; alt; alt = alt->next) {
684                 const char *path = alt_sha1_path(alt, sha1);
685                 if (check_and_freshen_file(path, freshen))
686                         return 1;
687         }
688         return 0;
689 }
690
691 static int check_and_freshen(const unsigned char *sha1, int freshen)
692 {
693         return check_and_freshen_local(sha1, freshen) ||
694                check_and_freshen_nonlocal(sha1, freshen);
695 }
696
697 int has_loose_object_nonlocal(const unsigned char *sha1)
698 {
699         return check_and_freshen_nonlocal(sha1, 0);
700 }
701
702 static int has_loose_object(const unsigned char *sha1)
703 {
704         return check_and_freshen(sha1, 0);
705 }
706
707 static unsigned int pack_used_ctr;
708 static unsigned int pack_mmap_calls;
709 static unsigned int peak_pack_open_windows;
710 static unsigned int pack_open_windows;
711 static unsigned int pack_open_fds;
712 static unsigned int pack_max_fds;
713 static size_t peak_pack_mapped;
714 static size_t pack_mapped;
715 struct packed_git *packed_git;
716
717 static struct mru packed_git_mru_storage;
718 struct mru *packed_git_mru = &packed_git_mru_storage;
719
720 void pack_report(void)
721 {
722         fprintf(stderr,
723                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
724                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
725                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
726                 sz_fmt(getpagesize()),
727                 sz_fmt(packed_git_window_size),
728                 sz_fmt(packed_git_limit));
729         fprintf(stderr,
730                 "pack_report: pack_used_ctr            = %10u\n"
731                 "pack_report: pack_mmap_calls          = %10u\n"
732                 "pack_report: pack_open_windows        = %10u / %10u\n"
733                 "pack_report: pack_mapped              = "
734                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
735                 pack_used_ctr,
736                 pack_mmap_calls,
737                 pack_open_windows, peak_pack_open_windows,
738                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
739 }
740
741 /*
742  * Open and mmap the index file at path, perform a couple of
743  * consistency checks, then record its information to p.  Return 0 on
744  * success.
745  */
746 static int check_packed_git_idx(const char *path, struct packed_git *p)
747 {
748         void *idx_map;
749         struct pack_idx_header *hdr;
750         size_t idx_size;
751         uint32_t version, nr, i, *index;
752         int fd = git_open(path);
753         struct stat st;
754
755         if (fd < 0)
756                 return -1;
757         if (fstat(fd, &st)) {
758                 close(fd);
759                 return -1;
760         }
761         idx_size = xsize_t(st.st_size);
762         if (idx_size < 4 * 256 + 20 + 20) {
763                 close(fd);
764                 return error("index file %s is too small", path);
765         }
766         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
767         close(fd);
768
769         hdr = idx_map;
770         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
771                 version = ntohl(hdr->idx_version);
772                 if (version < 2 || version > 2) {
773                         munmap(idx_map, idx_size);
774                         return error("index file %s is version %"PRIu32
775                                      " and is not supported by this binary"
776                                      " (try upgrading GIT to a newer version)",
777                                      path, version);
778                 }
779         } else
780                 version = 1;
781
782         nr = 0;
783         index = idx_map;
784         if (version > 1)
785                 index += 2;  /* skip index header */
786         for (i = 0; i < 256; i++) {
787                 uint32_t n = ntohl(index[i]);
788                 if (n < nr) {
789                         munmap(idx_map, idx_size);
790                         return error("non-monotonic index %s", path);
791                 }
792                 nr = n;
793         }
794
795         if (version == 1) {
796                 /*
797                  * Total size:
798                  *  - 256 index entries 4 bytes each
799                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
800                  *  - 20-byte SHA1 of the packfile
801                  *  - 20-byte SHA1 file checksum
802                  */
803                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
804                         munmap(idx_map, idx_size);
805                         return error("wrong index v1 file size in %s", path);
806                 }
807         } else if (version == 2) {
808                 /*
809                  * Minimum size:
810                  *  - 8 bytes of header
811                  *  - 256 index entries 4 bytes each
812                  *  - 20-byte sha1 entry * nr
813                  *  - 4-byte crc entry * nr
814                  *  - 4-byte offset entry * nr
815                  *  - 20-byte SHA1 of the packfile
816                  *  - 20-byte SHA1 file checksum
817                  * And after the 4-byte offset table might be a
818                  * variable sized table containing 8-byte entries
819                  * for offsets larger than 2^31.
820                  */
821                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
822                 unsigned long max_size = min_size;
823                 if (nr)
824                         max_size += (nr - 1)*8;
825                 if (idx_size < min_size || idx_size > max_size) {
826                         munmap(idx_map, idx_size);
827                         return error("wrong index v2 file size in %s", path);
828                 }
829                 if (idx_size != min_size &&
830                     /*
831                      * make sure we can deal with large pack offsets.
832                      * 31-bit signed offset won't be enough, neither
833                      * 32-bit unsigned one will be.
834                      */
835                     (sizeof(off_t) <= 4)) {
836                         munmap(idx_map, idx_size);
837                         return error("pack too large for current definition of off_t in %s", path);
838                 }
839         }
840
841         p->index_version = version;
842         p->index_data = idx_map;
843         p->index_size = idx_size;
844         p->num_objects = nr;
845         return 0;
846 }
847
848 int open_pack_index(struct packed_git *p)
849 {
850         char *idx_name;
851         size_t len;
852         int ret;
853
854         if (p->index_data)
855                 return 0;
856
857         if (!strip_suffix(p->pack_name, ".pack", &len))
858                 die("BUG: pack_name does not end in .pack");
859         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
860         ret = check_packed_git_idx(idx_name, p);
861         free(idx_name);
862         return ret;
863 }
864
865 static void scan_windows(struct packed_git *p,
866         struct packed_git **lru_p,
867         struct pack_window **lru_w,
868         struct pack_window **lru_l)
869 {
870         struct pack_window *w, *w_l;
871
872         for (w_l = NULL, w = p->windows; w; w = w->next) {
873                 if (!w->inuse_cnt) {
874                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
875                                 *lru_p = p;
876                                 *lru_w = w;
877                                 *lru_l = w_l;
878                         }
879                 }
880                 w_l = w;
881         }
882 }
883
884 static int unuse_one_window(struct packed_git *current)
885 {
886         struct packed_git *p, *lru_p = NULL;
887         struct pack_window *lru_w = NULL, *lru_l = NULL;
888
889         if (current)
890                 scan_windows(current, &lru_p, &lru_w, &lru_l);
891         for (p = packed_git; p; p = p->next)
892                 scan_windows(p, &lru_p, &lru_w, &lru_l);
893         if (lru_p) {
894                 munmap(lru_w->base, lru_w->len);
895                 pack_mapped -= lru_w->len;
896                 if (lru_l)
897                         lru_l->next = lru_w->next;
898                 else
899                         lru_p->windows = lru_w->next;
900                 free(lru_w);
901                 pack_open_windows--;
902                 return 1;
903         }
904         return 0;
905 }
906
907 void release_pack_memory(size_t need)
908 {
909         size_t cur = pack_mapped;
910         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
911                 ; /* nothing */
912 }
913
914 static void mmap_limit_check(size_t length)
915 {
916         static size_t limit = 0;
917         if (!limit) {
918                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
919                 if (!limit)
920                         limit = SIZE_MAX;
921         }
922         if (length > limit)
923                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
924                     (uintmax_t)length, (uintmax_t)limit);
925 }
926
927 void *xmmap_gently(void *start, size_t length,
928                   int prot, int flags, int fd, off_t offset)
929 {
930         void *ret;
931
932         mmap_limit_check(length);
933         ret = mmap(start, length, prot, flags, fd, offset);
934         if (ret == MAP_FAILED) {
935                 if (!length)
936                         return NULL;
937                 release_pack_memory(length);
938                 ret = mmap(start, length, prot, flags, fd, offset);
939         }
940         return ret;
941 }
942
943 void *xmmap(void *start, size_t length,
944         int prot, int flags, int fd, off_t offset)
945 {
946         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
947         if (ret == MAP_FAILED)
948                 die_errno("mmap failed");
949         return ret;
950 }
951
952 void close_pack_windows(struct packed_git *p)
953 {
954         while (p->windows) {
955                 struct pack_window *w = p->windows;
956
957                 if (w->inuse_cnt)
958                         die("pack '%s' still has open windows to it",
959                             p->pack_name);
960                 munmap(w->base, w->len);
961                 pack_mapped -= w->len;
962                 pack_open_windows--;
963                 p->windows = w->next;
964                 free(w);
965         }
966 }
967
968 static int close_pack_fd(struct packed_git *p)
969 {
970         if (p->pack_fd < 0)
971                 return 0;
972
973         close(p->pack_fd);
974         pack_open_fds--;
975         p->pack_fd = -1;
976
977         return 1;
978 }
979
980 static void close_pack(struct packed_git *p)
981 {
982         close_pack_windows(p);
983         close_pack_fd(p);
984         close_pack_index(p);
985 }
986
987 void close_all_packs(void)
988 {
989         struct packed_git *p;
990
991         for (p = packed_git; p; p = p->next)
992                 if (p->do_not_close)
993                         die("BUG: want to close pack marked 'do-not-close'");
994                 else
995                         close_pack(p);
996 }
997
998
999 /*
1000  * The LRU pack is the one with the oldest MRU window, preferring packs
1001  * with no used windows, or the oldest mtime if it has no windows allocated.
1002  */
1003 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
1004 {
1005         struct pack_window *w, *this_mru_w;
1006         int has_windows_inuse = 0;
1007
1008         /*
1009          * Reject this pack if it has windows and the previously selected
1010          * one does not.  If this pack does not have windows, reject
1011          * it if the pack file is newer than the previously selected one.
1012          */
1013         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
1014                 return;
1015
1016         for (w = this_mru_w = p->windows; w; w = w->next) {
1017                 /*
1018                  * Reject this pack if any of its windows are in use,
1019                  * but the previously selected pack did not have any
1020                  * inuse windows.  Otherwise, record that this pack
1021                  * has windows in use.
1022                  */
1023                 if (w->inuse_cnt) {
1024                         if (*accept_windows_inuse)
1025                                 has_windows_inuse = 1;
1026                         else
1027                                 return;
1028                 }
1029
1030                 if (w->last_used > this_mru_w->last_used)
1031                         this_mru_w = w;
1032
1033                 /*
1034                  * Reject this pack if it has windows that have been
1035                  * used more recently than the previously selected pack.
1036                  * If the previously selected pack had windows inuse and
1037                  * we have not encountered a window in this pack that is
1038                  * inuse, skip this check since we prefer a pack with no
1039                  * inuse windows to one that has inuse windows.
1040                  */
1041                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
1042                     this_mru_w->last_used > (*mru_w)->last_used)
1043                         return;
1044         }
1045
1046         /*
1047          * Select this pack.
1048          */
1049         *mru_w = this_mru_w;
1050         *lru_p = p;
1051         *accept_windows_inuse = has_windows_inuse;
1052 }
1053
1054 static int close_one_pack(void)
1055 {
1056         struct packed_git *p, *lru_p = NULL;
1057         struct pack_window *mru_w = NULL;
1058         int accept_windows_inuse = 1;
1059
1060         for (p = packed_git; p; p = p->next) {
1061                 if (p->pack_fd == -1)
1062                         continue;
1063                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
1064         }
1065
1066         if (lru_p)
1067                 return close_pack_fd(lru_p);
1068
1069         return 0;
1070 }
1071
1072 void unuse_pack(struct pack_window **w_cursor)
1073 {
1074         struct pack_window *w = *w_cursor;
1075         if (w) {
1076                 w->inuse_cnt--;
1077                 *w_cursor = NULL;
1078         }
1079 }
1080
1081 void close_pack_index(struct packed_git *p)
1082 {
1083         if (p->index_data) {
1084                 munmap((void *)p->index_data, p->index_size);
1085                 p->index_data = NULL;
1086         }
1087 }
1088
1089 static unsigned int get_max_fd_limit(void)
1090 {
1091 #ifdef RLIMIT_NOFILE
1092         {
1093                 struct rlimit lim;
1094
1095                 if (!getrlimit(RLIMIT_NOFILE, &lim))
1096                         return lim.rlim_cur;
1097         }
1098 #endif
1099
1100 #ifdef _SC_OPEN_MAX
1101         {
1102                 long open_max = sysconf(_SC_OPEN_MAX);
1103                 if (0 < open_max)
1104                         return open_max;
1105                 /*
1106                  * Otherwise, we got -1 for one of the two
1107                  * reasons:
1108                  *
1109                  * (1) sysconf() did not understand _SC_OPEN_MAX
1110                  *     and signaled an error with -1; or
1111                  * (2) sysconf() said there is no limit.
1112                  *
1113                  * We _could_ clear errno before calling sysconf() to
1114                  * tell these two cases apart and return a huge number
1115                  * in the latter case to let the caller cap it to a
1116                  * value that is not so selfish, but letting the
1117                  * fallback OPEN_MAX codepath take care of these cases
1118                  * is a lot simpler.
1119                  */
1120         }
1121 #endif
1122
1123 #ifdef OPEN_MAX
1124         return OPEN_MAX;
1125 #else
1126         return 1; /* see the caller ;-) */
1127 #endif
1128 }
1129
1130 /*
1131  * Do not call this directly as this leaks p->pack_fd on error return;
1132  * call open_packed_git() instead.
1133  */
1134 static int open_packed_git_1(struct packed_git *p)
1135 {
1136         struct stat st;
1137         struct pack_header hdr;
1138         unsigned char sha1[20];
1139         unsigned char *idx_sha1;
1140         long fd_flag;
1141
1142         if (!p->index_data && open_pack_index(p))
1143                 return error("packfile %s index unavailable", p->pack_name);
1144
1145         if (!pack_max_fds) {
1146                 unsigned int max_fds = get_max_fd_limit();
1147
1148                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1149                 if (25 < max_fds)
1150                         pack_max_fds = max_fds - 25;
1151                 else
1152                         pack_max_fds = 1;
1153         }
1154
1155         while (pack_max_fds <= pack_open_fds && close_one_pack())
1156                 ; /* nothing */
1157
1158         p->pack_fd = git_open(p->pack_name);
1159         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1160                 return -1;
1161         pack_open_fds++;
1162
1163         /* If we created the struct before we had the pack we lack size. */
1164         if (!p->pack_size) {
1165                 if (!S_ISREG(st.st_mode))
1166                         return error("packfile %s not a regular file", p->pack_name);
1167                 p->pack_size = st.st_size;
1168         } else if (p->pack_size != st.st_size)
1169                 return error("packfile %s size changed", p->pack_name);
1170
1171         /* We leave these file descriptors open with sliding mmap;
1172          * there is no point keeping them open across exec(), though.
1173          */
1174         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1175         if (fd_flag < 0)
1176                 return error("cannot determine file descriptor flags");
1177         fd_flag |= FD_CLOEXEC;
1178         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1179                 return error("cannot set FD_CLOEXEC");
1180
1181         /* Verify we recognize this pack file format. */
1182         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1183                 return error("file %s is far too short to be a packfile", p->pack_name);
1184         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1185                 return error("file %s is not a GIT packfile", p->pack_name);
1186         if (!pack_version_ok(hdr.hdr_version))
1187                 return error("packfile %s is version %"PRIu32" and not"
1188                         " supported (try upgrading GIT to a newer version)",
1189                         p->pack_name, ntohl(hdr.hdr_version));
1190
1191         /* Verify the pack matches its index. */
1192         if (p->num_objects != ntohl(hdr.hdr_entries))
1193                 return error("packfile %s claims to have %"PRIu32" objects"
1194                              " while index indicates %"PRIu32" objects",
1195                              p->pack_name, ntohl(hdr.hdr_entries),
1196                              p->num_objects);
1197         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1198                 return error("end of packfile %s is unavailable", p->pack_name);
1199         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1200                 return error("packfile %s signature is unavailable", p->pack_name);
1201         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1202         if (hashcmp(sha1, idx_sha1))
1203                 return error("packfile %s does not match index", p->pack_name);
1204         return 0;
1205 }
1206
1207 static int open_packed_git(struct packed_git *p)
1208 {
1209         if (!open_packed_git_1(p))
1210                 return 0;
1211         close_pack_fd(p);
1212         return -1;
1213 }
1214
1215 static int in_window(struct pack_window *win, off_t offset)
1216 {
1217         /* We must promise at least 20 bytes (one hash) after the
1218          * offset is available from this window, otherwise the offset
1219          * is not actually in this window and a different window (which
1220          * has that one hash excess) must be used.  This is to support
1221          * the object header and delta base parsing routines below.
1222          */
1223         off_t win_off = win->offset;
1224         return win_off <= offset
1225                 && (offset + 20) <= (win_off + win->len);
1226 }
1227
1228 unsigned char *use_pack(struct packed_git *p,
1229                 struct pack_window **w_cursor,
1230                 off_t offset,
1231                 unsigned long *left)
1232 {
1233         struct pack_window *win = *w_cursor;
1234
1235         /* Since packfiles end in a hash of their content and it's
1236          * pointless to ask for an offset into the middle of that
1237          * hash, and the in_window function above wouldn't match
1238          * don't allow an offset too close to the end of the file.
1239          */
1240         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1241                 die("packfile %s cannot be accessed", p->pack_name);
1242         if (offset > (p->pack_size - 20))
1243                 die("offset beyond end of packfile (truncated pack?)");
1244         if (offset < 0)
1245                 die(_("offset before end of packfile (broken .idx?)"));
1246
1247         if (!win || !in_window(win, offset)) {
1248                 if (win)
1249                         win->inuse_cnt--;
1250                 for (win = p->windows; win; win = win->next) {
1251                         if (in_window(win, offset))
1252                                 break;
1253                 }
1254                 if (!win) {
1255                         size_t window_align = packed_git_window_size / 2;
1256                         off_t len;
1257
1258                         if (p->pack_fd == -1 && open_packed_git(p))
1259                                 die("packfile %s cannot be accessed", p->pack_name);
1260
1261                         win = xcalloc(1, sizeof(*win));
1262                         win->offset = (offset / window_align) * window_align;
1263                         len = p->pack_size - win->offset;
1264                         if (len > packed_git_window_size)
1265                                 len = packed_git_window_size;
1266                         win->len = (size_t)len;
1267                         pack_mapped += win->len;
1268                         while (packed_git_limit < pack_mapped
1269                                 && unuse_one_window(p))
1270                                 ; /* nothing */
1271                         win->base = xmmap(NULL, win->len,
1272                                 PROT_READ, MAP_PRIVATE,
1273                                 p->pack_fd, win->offset);
1274                         if (win->base == MAP_FAILED)
1275                                 die_errno("packfile %s cannot be mapped",
1276                                           p->pack_name);
1277                         if (!win->offset && win->len == p->pack_size
1278                                 && !p->do_not_close)
1279                                 close_pack_fd(p);
1280                         pack_mmap_calls++;
1281                         pack_open_windows++;
1282                         if (pack_mapped > peak_pack_mapped)
1283                                 peak_pack_mapped = pack_mapped;
1284                         if (pack_open_windows > peak_pack_open_windows)
1285                                 peak_pack_open_windows = pack_open_windows;
1286                         win->next = p->windows;
1287                         p->windows = win;
1288                 }
1289         }
1290         if (win != *w_cursor) {
1291                 win->last_used = pack_used_ctr++;
1292                 win->inuse_cnt++;
1293                 *w_cursor = win;
1294         }
1295         offset -= win->offset;
1296         if (left)
1297                 *left = win->len - xsize_t(offset);
1298         return win->base + offset;
1299 }
1300
1301 static struct packed_git *alloc_packed_git(int extra)
1302 {
1303         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1304         memset(p, 0, sizeof(*p));
1305         p->pack_fd = -1;
1306         return p;
1307 }
1308
1309 static void try_to_free_pack_memory(size_t size)
1310 {
1311         release_pack_memory(size);
1312 }
1313
1314 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1315 {
1316         static int have_set_try_to_free_routine;
1317         struct stat st;
1318         size_t alloc;
1319         struct packed_git *p;
1320
1321         if (!have_set_try_to_free_routine) {
1322                 have_set_try_to_free_routine = 1;
1323                 set_try_to_free_routine(try_to_free_pack_memory);
1324         }
1325
1326         /*
1327          * Make sure a corresponding .pack file exists and that
1328          * the index looks sane.
1329          */
1330         if (!strip_suffix_mem(path, &path_len, ".idx"))
1331                 return NULL;
1332
1333         /*
1334          * ".pack" is long enough to hold any suffix we're adding (and
1335          * the use xsnprintf double-checks that)
1336          */
1337         alloc = st_add3(path_len, strlen(".pack"), 1);
1338         p = alloc_packed_git(alloc);
1339         memcpy(p->pack_name, path, path_len);
1340
1341         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1342         if (!access(p->pack_name, F_OK))
1343                 p->pack_keep = 1;
1344
1345         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1346         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1347                 free(p);
1348                 return NULL;
1349         }
1350
1351         /* ok, it looks sane as far as we can check without
1352          * actually mapping the pack file.
1353          */
1354         p->pack_size = st.st_size;
1355         p->pack_local = local;
1356         p->mtime = st.st_mtime;
1357         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1358                 hashclr(p->sha1);
1359         return p;
1360 }
1361
1362 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1363 {
1364         const char *path = sha1_pack_name(sha1);
1365         size_t alloc = st_add(strlen(path), 1);
1366         struct packed_git *p = alloc_packed_git(alloc);
1367
1368         memcpy(p->pack_name, path, alloc); /* includes NUL */
1369         hashcpy(p->sha1, sha1);
1370         if (check_packed_git_idx(idx_path, p)) {
1371                 free(p);
1372                 return NULL;
1373         }
1374
1375         return p;
1376 }
1377
1378 void install_packed_git(struct packed_git *pack)
1379 {
1380         if (pack->pack_fd != -1)
1381                 pack_open_fds++;
1382
1383         pack->next = packed_git;
1384         packed_git = pack;
1385 }
1386
1387 void (*report_garbage)(unsigned seen_bits, const char *path);
1388
1389 static void report_helper(const struct string_list *list,
1390                           int seen_bits, int first, int last)
1391 {
1392         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1393                 return;
1394
1395         for (; first < last; first++)
1396                 report_garbage(seen_bits, list->items[first].string);
1397 }
1398
1399 static void report_pack_garbage(struct string_list *list)
1400 {
1401         int i, baselen = -1, first = 0, seen_bits = 0;
1402
1403         if (!report_garbage)
1404                 return;
1405
1406         string_list_sort(list);
1407
1408         for (i = 0; i < list->nr; i++) {
1409                 const char *path = list->items[i].string;
1410                 if (baselen != -1 &&
1411                     strncmp(path, list->items[first].string, baselen)) {
1412                         report_helper(list, seen_bits, first, i);
1413                         baselen = -1;
1414                         seen_bits = 0;
1415                 }
1416                 if (baselen == -1) {
1417                         const char *dot = strrchr(path, '.');
1418                         if (!dot) {
1419                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1420                                 continue;
1421                         }
1422                         baselen = dot - path + 1;
1423                         first = i;
1424                 }
1425                 if (!strcmp(path + baselen, "pack"))
1426                         seen_bits |= 1;
1427                 else if (!strcmp(path + baselen, "idx"))
1428                         seen_bits |= 2;
1429         }
1430         report_helper(list, seen_bits, first, list->nr);
1431 }
1432
1433 static void prepare_packed_git_one(char *objdir, int local)
1434 {
1435         struct strbuf path = STRBUF_INIT;
1436         size_t dirnamelen;
1437         DIR *dir;
1438         struct dirent *de;
1439         struct string_list garbage = STRING_LIST_INIT_DUP;
1440
1441         strbuf_addstr(&path, objdir);
1442         strbuf_addstr(&path, "/pack");
1443         dir = opendir(path.buf);
1444         if (!dir) {
1445                 if (errno != ENOENT)
1446                         error_errno("unable to open object pack directory: %s",
1447                                     path.buf);
1448                 strbuf_release(&path);
1449                 return;
1450         }
1451         strbuf_addch(&path, '/');
1452         dirnamelen = path.len;
1453         while ((de = readdir(dir)) != NULL) {
1454                 struct packed_git *p;
1455                 size_t base_len;
1456
1457                 if (is_dot_or_dotdot(de->d_name))
1458                         continue;
1459
1460                 strbuf_setlen(&path, dirnamelen);
1461                 strbuf_addstr(&path, de->d_name);
1462
1463                 base_len = path.len;
1464                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1465                         /* Don't reopen a pack we already have. */
1466                         for (p = packed_git; p; p = p->next) {
1467                                 size_t len;
1468                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1469                                     len == base_len &&
1470                                     !memcmp(p->pack_name, path.buf, len))
1471                                         break;
1472                         }
1473                         if (p == NULL &&
1474                             /*
1475                              * See if it really is a valid .idx file with
1476                              * corresponding .pack file that we can map.
1477                              */
1478                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1479                                 install_packed_git(p);
1480                 }
1481
1482                 if (!report_garbage)
1483                         continue;
1484
1485                 if (ends_with(de->d_name, ".idx") ||
1486                     ends_with(de->d_name, ".pack") ||
1487                     ends_with(de->d_name, ".bitmap") ||
1488                     ends_with(de->d_name, ".keep"))
1489                         string_list_append(&garbage, path.buf);
1490                 else
1491                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1492         }
1493         closedir(dir);
1494         report_pack_garbage(&garbage);
1495         string_list_clear(&garbage, 0);
1496         strbuf_release(&path);
1497 }
1498
1499 static int approximate_object_count_valid;
1500
1501 /*
1502  * Give a fast, rough count of the number of objects in the repository. This
1503  * ignores loose objects completely. If you have a lot of them, then either
1504  * you should repack because your performance will be awful, or they are
1505  * all unreachable objects about to be pruned, in which case they're not really
1506  * interesting as a measure of repo size in the first place.
1507  */
1508 unsigned long approximate_object_count(void)
1509 {
1510         static unsigned long count;
1511         if (!approximate_object_count_valid) {
1512                 struct packed_git *p;
1513
1514                 prepare_packed_git();
1515                 count = 0;
1516                 for (p = packed_git; p; p = p->next) {
1517                         if (open_pack_index(p))
1518                                 continue;
1519                         count += p->num_objects;
1520                 }
1521         }
1522         return count;
1523 }
1524
1525 static void *get_next_packed_git(const void *p)
1526 {
1527         return ((const struct packed_git *)p)->next;
1528 }
1529
1530 static void set_next_packed_git(void *p, void *next)
1531 {
1532         ((struct packed_git *)p)->next = next;
1533 }
1534
1535 static int sort_pack(const void *a_, const void *b_)
1536 {
1537         const struct packed_git *a = a_;
1538         const struct packed_git *b = b_;
1539         int st;
1540
1541         /*
1542          * Local packs tend to contain objects specific to our
1543          * variant of the project than remote ones.  In addition,
1544          * remote ones could be on a network mounted filesystem.
1545          * Favor local ones for these reasons.
1546          */
1547         st = a->pack_local - b->pack_local;
1548         if (st)
1549                 return -st;
1550
1551         /*
1552          * Younger packs tend to contain more recent objects,
1553          * and more recent objects tend to get accessed more
1554          * often.
1555          */
1556         if (a->mtime < b->mtime)
1557                 return 1;
1558         else if (a->mtime == b->mtime)
1559                 return 0;
1560         return -1;
1561 }
1562
1563 static void rearrange_packed_git(void)
1564 {
1565         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1566                                      set_next_packed_git, sort_pack);
1567 }
1568
1569 static void prepare_packed_git_mru(void)
1570 {
1571         struct packed_git *p;
1572
1573         mru_clear(packed_git_mru);
1574         for (p = packed_git; p; p = p->next)
1575                 mru_append(packed_git_mru, p);
1576 }
1577
1578 static int prepare_packed_git_run_once = 0;
1579 void prepare_packed_git(void)
1580 {
1581         struct alternate_object_database *alt;
1582
1583         if (prepare_packed_git_run_once)
1584                 return;
1585         prepare_packed_git_one(get_object_directory(), 1);
1586         prepare_alt_odb();
1587         for (alt = alt_odb_list; alt; alt = alt->next)
1588                 prepare_packed_git_one(alt->path, 0);
1589         rearrange_packed_git();
1590         prepare_packed_git_mru();
1591         prepare_packed_git_run_once = 1;
1592 }
1593
1594 void reprepare_packed_git(void)
1595 {
1596         approximate_object_count_valid = 0;
1597         prepare_packed_git_run_once = 0;
1598         prepare_packed_git();
1599 }
1600
1601 static void mark_bad_packed_object(struct packed_git *p,
1602                                    const unsigned char *sha1)
1603 {
1604         unsigned i;
1605         for (i = 0; i < p->num_bad_objects; i++)
1606                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1607                         return;
1608         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1609                                       st_mult(GIT_MAX_RAWSZ,
1610                                               st_add(p->num_bad_objects, 1)));
1611         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1612         p->num_bad_objects++;
1613 }
1614
1615 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1616 {
1617         struct packed_git *p;
1618         unsigned i;
1619
1620         for (p = packed_git; p; p = p->next)
1621                 for (i = 0; i < p->num_bad_objects; i++)
1622                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1623                                 return p;
1624         return NULL;
1625 }
1626
1627 /*
1628  * With an in-core object data in "map", rehash it to make sure the
1629  * object name actually matches "sha1" to detect object corruption.
1630  * With "map" == NULL, try reading the object named with "sha1" using
1631  * the streaming interface and rehash it to do the same.
1632  */
1633 int check_sha1_signature(const unsigned char *sha1, void *map,
1634                          unsigned long size, const char *type)
1635 {
1636         unsigned char real_sha1[20];
1637         enum object_type obj_type;
1638         struct git_istream *st;
1639         git_SHA_CTX c;
1640         char hdr[32];
1641         int hdrlen;
1642
1643         if (map) {
1644                 hash_sha1_file(map, size, type, real_sha1);
1645                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1646         }
1647
1648         st = open_istream(sha1, &obj_type, &size, NULL);
1649         if (!st)
1650                 return -1;
1651
1652         /* Generate the header */
1653         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1654
1655         /* Sha1.. */
1656         git_SHA1_Init(&c);
1657         git_SHA1_Update(&c, hdr, hdrlen);
1658         for (;;) {
1659                 char buf[1024 * 16];
1660                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1661
1662                 if (readlen < 0) {
1663                         close_istream(st);
1664                         return -1;
1665                 }
1666                 if (!readlen)
1667                         break;
1668                 git_SHA1_Update(&c, buf, readlen);
1669         }
1670         git_SHA1_Final(real_sha1, &c);
1671         close_istream(st);
1672         return hashcmp(sha1, real_sha1) ? -1 : 0;
1673 }
1674
1675 int git_open_cloexec(const char *name, int flags)
1676 {
1677         int fd;
1678         static int o_cloexec = O_CLOEXEC;
1679
1680         fd = open(name, flags | o_cloexec);
1681         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
1682                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
1683                 o_cloexec &= ~O_CLOEXEC;
1684                 fd = open(name, flags | o_cloexec);
1685         }
1686
1687 #if defined(F_GETFD) && defined(F_SETFD) && defined(FD_CLOEXEC)
1688         {
1689                 static int fd_cloexec = FD_CLOEXEC;
1690
1691                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
1692                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
1693                         int flags = fcntl(fd, F_GETFD);
1694                         if (fcntl(fd, F_SETFD, flags | fd_cloexec))
1695                                 fd_cloexec = 0;
1696                 }
1697         }
1698 #endif
1699         return fd;
1700 }
1701
1702 /*
1703  * Find "sha1" as a loose object in the local repository or in an alternate.
1704  * Returns 0 on success, negative on failure.
1705  *
1706  * The "path" out-parameter will give the path of the object we found (if any).
1707  * Note that it may point to static storage and is only valid until another
1708  * call to sha1_file_name(), etc.
1709  */
1710 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
1711                           const char **path)
1712 {
1713         struct alternate_object_database *alt;
1714
1715         *path = sha1_file_name(sha1);
1716         if (!lstat(*path, st))
1717                 return 0;
1718
1719         prepare_alt_odb();
1720         errno = ENOENT;
1721         for (alt = alt_odb_list; alt; alt = alt->next) {
1722                 *path = alt_sha1_path(alt, sha1);
1723                 if (!lstat(*path, st))
1724                         return 0;
1725         }
1726
1727         return -1;
1728 }
1729
1730 /*
1731  * Like stat_sha1_file(), but actually open the object and return the
1732  * descriptor. See the caveats on the "path" parameter above.
1733  */
1734 static int open_sha1_file(const unsigned char *sha1, const char **path)
1735 {
1736         int fd;
1737         struct alternate_object_database *alt;
1738         int most_interesting_errno;
1739
1740         *path = sha1_file_name(sha1);
1741         fd = git_open(*path);
1742         if (fd >= 0)
1743                 return fd;
1744         most_interesting_errno = errno;
1745
1746         prepare_alt_odb();
1747         for (alt = alt_odb_list; alt; alt = alt->next) {
1748                 *path = alt_sha1_path(alt, sha1);
1749                 fd = git_open(*path);
1750                 if (fd >= 0)
1751                         return fd;
1752                 if (most_interesting_errno == ENOENT)
1753                         most_interesting_errno = errno;
1754         }
1755         errno = most_interesting_errno;
1756         return -1;
1757 }
1758
1759 /*
1760  * Map the loose object at "path" if it is not NULL, or the path found by
1761  * searching for a loose object named "sha1".
1762  */
1763 static void *map_sha1_file_1(const char *path,
1764                              const unsigned char *sha1,
1765                              unsigned long *size)
1766 {
1767         void *map;
1768         int fd;
1769
1770         if (path)
1771                 fd = git_open(path);
1772         else
1773                 fd = open_sha1_file(sha1, &path);
1774         map = NULL;
1775         if (fd >= 0) {
1776                 struct stat st;
1777
1778                 if (!fstat(fd, &st)) {
1779                         *size = xsize_t(st.st_size);
1780                         if (!*size) {
1781                                 /* mmap() is forbidden on empty files */
1782                                 error("object file %s is empty", path);
1783                                 return NULL;
1784                         }
1785                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1786                 }
1787                 close(fd);
1788         }
1789         return map;
1790 }
1791
1792 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1793 {
1794         return map_sha1_file_1(NULL, sha1, size);
1795 }
1796
1797 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1798                 unsigned long len, enum object_type *type, unsigned long *sizep)
1799 {
1800         unsigned shift;
1801         unsigned long size, c;
1802         unsigned long used = 0;
1803
1804         c = buf[used++];
1805         *type = (c >> 4) & 7;
1806         size = c & 15;
1807         shift = 4;
1808         while (c & 0x80) {
1809                 if (len <= used || bitsizeof(long) <= shift) {
1810                         error("bad object header");
1811                         size = used = 0;
1812                         break;
1813                 }
1814                 c = buf[used++];
1815                 size += (c & 0x7f) << shift;
1816                 shift += 7;
1817         }
1818         *sizep = size;
1819         return used;
1820 }
1821
1822 static int unpack_sha1_short_header(git_zstream *stream,
1823                                     unsigned char *map, unsigned long mapsize,
1824                                     void *buffer, unsigned long bufsiz)
1825 {
1826         /* Get the data stream */
1827         memset(stream, 0, sizeof(*stream));
1828         stream->next_in = map;
1829         stream->avail_in = mapsize;
1830         stream->next_out = buffer;
1831         stream->avail_out = bufsiz;
1832
1833         git_inflate_init(stream);
1834         return git_inflate(stream, 0);
1835 }
1836
1837 int unpack_sha1_header(git_zstream *stream,
1838                        unsigned char *map, unsigned long mapsize,
1839                        void *buffer, unsigned long bufsiz)
1840 {
1841         int status = unpack_sha1_short_header(stream, map, mapsize,
1842                                               buffer, bufsiz);
1843
1844         if (status < Z_OK)
1845                 return status;
1846
1847         /* Make sure we have the terminating NUL */
1848         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1849                 return -1;
1850         return 0;
1851 }
1852
1853 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1854                                         unsigned long mapsize, void *buffer,
1855                                         unsigned long bufsiz, struct strbuf *header)
1856 {
1857         int status;
1858
1859         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
1860         if (status < Z_OK)
1861                 return -1;
1862
1863         /*
1864          * Check if entire header is unpacked in the first iteration.
1865          */
1866         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1867                 return 0;
1868
1869         /*
1870          * buffer[0..bufsiz] was not large enough.  Copy the partial
1871          * result out to header, and then append the result of further
1872          * reading the stream.
1873          */
1874         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1875         stream->next_out = buffer;
1876         stream->avail_out = bufsiz;
1877
1878         do {
1879                 status = git_inflate(stream, 0);
1880                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1881                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1882                         return 0;
1883                 stream->next_out = buffer;
1884                 stream->avail_out = bufsiz;
1885         } while (status != Z_STREAM_END);
1886         return -1;
1887 }
1888
1889 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1890 {
1891         int bytes = strlen(buffer) + 1;
1892         unsigned char *buf = xmallocz(size);
1893         unsigned long n;
1894         int status = Z_OK;
1895
1896         n = stream->total_out - bytes;
1897         if (n > size)
1898                 n = size;
1899         memcpy(buf, (char *) buffer + bytes, n);
1900         bytes = n;
1901         if (bytes <= size) {
1902                 /*
1903                  * The above condition must be (bytes <= size), not
1904                  * (bytes < size).  In other words, even though we
1905                  * expect no more output and set avail_out to zero,
1906                  * the input zlib stream may have bytes that express
1907                  * "this concludes the stream", and we *do* want to
1908                  * eat that input.
1909                  *
1910                  * Otherwise we would not be able to test that we
1911                  * consumed all the input to reach the expected size;
1912                  * we also want to check that zlib tells us that all
1913                  * went well with status == Z_STREAM_END at the end.
1914                  */
1915                 stream->next_out = buf + bytes;
1916                 stream->avail_out = size - bytes;
1917                 while (status == Z_OK)
1918                         status = git_inflate(stream, Z_FINISH);
1919         }
1920         if (status == Z_STREAM_END && !stream->avail_in) {
1921                 git_inflate_end(stream);
1922                 return buf;
1923         }
1924
1925         if (status < 0)
1926                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1927         else if (stream->avail_in)
1928                 error("garbage at end of loose object '%s'",
1929                       sha1_to_hex(sha1));
1930         free(buf);
1931         return NULL;
1932 }
1933
1934 /*
1935  * We used to just use "sscanf()", but that's actually way
1936  * too permissive for what we want to check. So do an anal
1937  * object header parse by hand.
1938  */
1939 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1940                                unsigned int flags)
1941 {
1942         const char *type_buf = hdr;
1943         unsigned long size;
1944         int type, type_len = 0;
1945
1946         /*
1947          * The type can be of any size but is followed by
1948          * a space.
1949          */
1950         for (;;) {
1951                 char c = *hdr++;
1952                 if (!c)
1953                         return -1;
1954                 if (c == ' ')
1955                         break;
1956                 type_len++;
1957         }
1958
1959         type = type_from_string_gently(type_buf, type_len, 1);
1960         if (oi->typename)
1961                 strbuf_add(oi->typename, type_buf, type_len);
1962         /*
1963          * Set type to 0 if its an unknown object and
1964          * we're obtaining the type using '--allow-unknown-type'
1965          * option.
1966          */
1967         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE) && (type < 0))
1968                 type = 0;
1969         else if (type < 0)
1970                 die("invalid object type");
1971         if (oi->typep)
1972                 *oi->typep = type;
1973
1974         /*
1975          * The length must follow immediately, and be in canonical
1976          * decimal format (ie "010" is not valid).
1977          */
1978         size = *hdr++ - '0';
1979         if (size > 9)
1980                 return -1;
1981         if (size) {
1982                 for (;;) {
1983                         unsigned long c = *hdr - '0';
1984                         if (c > 9)
1985                                 break;
1986                         hdr++;
1987                         size = size * 10 + c;
1988                 }
1989         }
1990
1991         if (oi->sizep)
1992                 *oi->sizep = size;
1993
1994         /*
1995          * The length must be followed by a zero byte
1996          */
1997         return *hdr ? -1 : type;
1998 }
1999
2000 int parse_sha1_header(const char *hdr, unsigned long *sizep)
2001 {
2002         struct object_info oi = OBJECT_INFO_INIT;
2003
2004         oi.sizep = sizep;
2005         return parse_sha1_header_extended(hdr, &oi, 0);
2006 }
2007
2008 unsigned long get_size_from_delta(struct packed_git *p,
2009                                   struct pack_window **w_curs,
2010                                   off_t curpos)
2011 {
2012         const unsigned char *data;
2013         unsigned char delta_head[20], *in;
2014         git_zstream stream;
2015         int st;
2016
2017         memset(&stream, 0, sizeof(stream));
2018         stream.next_out = delta_head;
2019         stream.avail_out = sizeof(delta_head);
2020
2021         git_inflate_init(&stream);
2022         do {
2023                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2024                 stream.next_in = in;
2025                 st = git_inflate(&stream, Z_FINISH);
2026                 curpos += stream.next_in - in;
2027         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
2028                  stream.total_out < sizeof(delta_head));
2029         git_inflate_end(&stream);
2030         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
2031                 error("delta data unpack-initial failed");
2032                 return 0;
2033         }
2034
2035         /* Examine the initial part of the delta to figure out
2036          * the result size.
2037          */
2038         data = delta_head;
2039
2040         /* ignore base size */
2041         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2042
2043         /* Read the result size */
2044         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2045 }
2046
2047 static off_t get_delta_base(struct packed_git *p,
2048                                     struct pack_window **w_curs,
2049                                     off_t *curpos,
2050                                     enum object_type type,
2051                                     off_t delta_obj_offset)
2052 {
2053         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
2054         off_t base_offset;
2055
2056         /* use_pack() assured us we have [base_info, base_info + 20)
2057          * as a range that we can look at without walking off the
2058          * end of the mapped window.  Its actually the hash size
2059          * that is assured.  An OFS_DELTA longer than the hash size
2060          * is stupid, as then a REF_DELTA would be smaller to store.
2061          */
2062         if (type == OBJ_OFS_DELTA) {
2063                 unsigned used = 0;
2064                 unsigned char c = base_info[used++];
2065                 base_offset = c & 127;
2066                 while (c & 128) {
2067                         base_offset += 1;
2068                         if (!base_offset || MSB(base_offset, 7))
2069                                 return 0;  /* overflow */
2070                         c = base_info[used++];
2071                         base_offset = (base_offset << 7) + (c & 127);
2072                 }
2073                 base_offset = delta_obj_offset - base_offset;
2074                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
2075                         return 0;  /* out of bound */
2076                 *curpos += used;
2077         } else if (type == OBJ_REF_DELTA) {
2078                 /* The base entry _must_ be in the same pack */
2079                 base_offset = find_pack_entry_one(base_info, p);
2080                 *curpos += 20;
2081         } else
2082                 die("I am totally screwed");
2083         return base_offset;
2084 }
2085
2086 /*
2087  * Like get_delta_base above, but we return the sha1 instead of the pack
2088  * offset. This means it is cheaper for REF deltas (we do not have to do
2089  * the final object lookup), but more expensive for OFS deltas (we
2090  * have to load the revidx to convert the offset back into a sha1).
2091  */
2092 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
2093                                                 struct pack_window **w_curs,
2094                                                 off_t curpos,
2095                                                 enum object_type type,
2096                                                 off_t delta_obj_offset)
2097 {
2098         if (type == OBJ_REF_DELTA) {
2099                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
2100                 return base;
2101         } else if (type == OBJ_OFS_DELTA) {
2102                 struct revindex_entry *revidx;
2103                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
2104                                                    type, delta_obj_offset);
2105
2106                 if (!base_offset)
2107                         return NULL;
2108
2109                 revidx = find_pack_revindex(p, base_offset);
2110                 if (!revidx)
2111                         return NULL;
2112
2113                 return nth_packed_object_sha1(p, revidx->nr);
2114         } else
2115                 return NULL;
2116 }
2117
2118 int unpack_object_header(struct packed_git *p,
2119                          struct pack_window **w_curs,
2120                          off_t *curpos,
2121                          unsigned long *sizep)
2122 {
2123         unsigned char *base;
2124         unsigned long left;
2125         unsigned long used;
2126         enum object_type type;
2127
2128         /* use_pack() assures us we have [base, base + 20) available
2129          * as a range that we can look at.  (Its actually the hash
2130          * size that is assured.)  With our object header encoding
2131          * the maximum deflated object size is 2^137, which is just
2132          * insane, so we know won't exceed what we have been given.
2133          */
2134         base = use_pack(p, w_curs, *curpos, &left);
2135         used = unpack_object_header_buffer(base, left, &type, sizep);
2136         if (!used) {
2137                 type = OBJ_BAD;
2138         } else
2139                 *curpos += used;
2140
2141         return type;
2142 }
2143
2144 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
2145 {
2146         int type;
2147         struct revindex_entry *revidx;
2148         const unsigned char *sha1;
2149         revidx = find_pack_revindex(p, obj_offset);
2150         if (!revidx)
2151                 return OBJ_BAD;
2152         sha1 = nth_packed_object_sha1(p, revidx->nr);
2153         mark_bad_packed_object(p, sha1);
2154         type = sha1_object_info(sha1, NULL);
2155         if (type <= OBJ_NONE)
2156                 return OBJ_BAD;
2157         return type;
2158 }
2159
2160 #define POI_STACK_PREALLOC 64
2161
2162 static enum object_type packed_to_object_type(struct packed_git *p,
2163                                               off_t obj_offset,
2164                                               enum object_type type,
2165                                               struct pack_window **w_curs,
2166                                               off_t curpos)
2167 {
2168         off_t small_poi_stack[POI_STACK_PREALLOC];
2169         off_t *poi_stack = small_poi_stack;
2170         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
2171
2172         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2173                 off_t base_offset;
2174                 unsigned long size;
2175                 /* Push the object we're going to leave behind */
2176                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
2177                         poi_stack_alloc = alloc_nr(poi_stack_nr);
2178                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
2179                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2180                 } else {
2181                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2182                 }
2183                 poi_stack[poi_stack_nr++] = obj_offset;
2184                 /* If parsing the base offset fails, just unwind */
2185                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2186                 if (!base_offset)
2187                         goto unwind;
2188                 curpos = obj_offset = base_offset;
2189                 type = unpack_object_header(p, w_curs, &curpos, &size);
2190                 if (type <= OBJ_NONE) {
2191                         /* If getting the base itself fails, we first
2192                          * retry the base, otherwise unwind */
2193                         type = retry_bad_packed_offset(p, base_offset);
2194                         if (type > OBJ_NONE)
2195                                 goto out;
2196                         goto unwind;
2197                 }
2198         }
2199
2200         switch (type) {
2201         case OBJ_BAD:
2202         case OBJ_COMMIT:
2203         case OBJ_TREE:
2204         case OBJ_BLOB:
2205         case OBJ_TAG:
2206                 break;
2207         default:
2208                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2209                       type, (uintmax_t)obj_offset, p->pack_name);
2210                 type = OBJ_BAD;
2211         }
2212
2213 out:
2214         if (poi_stack != small_poi_stack)
2215                 free(poi_stack);
2216         return type;
2217
2218 unwind:
2219         while (poi_stack_nr) {
2220                 obj_offset = poi_stack[--poi_stack_nr];
2221                 type = retry_bad_packed_offset(p, obj_offset);
2222                 if (type > OBJ_NONE)
2223                         goto out;
2224         }
2225         type = OBJ_BAD;
2226         goto out;
2227 }
2228
2229 static struct hashmap delta_base_cache;
2230 static size_t delta_base_cached;
2231
2232 static LIST_HEAD(delta_base_cache_lru);
2233
2234 struct delta_base_cache_key {
2235         struct packed_git *p;
2236         off_t base_offset;
2237 };
2238
2239 struct delta_base_cache_entry {
2240         struct hashmap hash;
2241         struct delta_base_cache_key key;
2242         struct list_head lru;
2243         void *data;
2244         unsigned long size;
2245         enum object_type type;
2246 };
2247
2248 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2249 {
2250         unsigned int hash;
2251
2252         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2253         hash += (hash >> 8) + (hash >> 16);
2254         return hash;
2255 }
2256
2257 static struct delta_base_cache_entry *
2258 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2259 {
2260         struct hashmap_entry entry;
2261         struct delta_base_cache_key key;
2262
2263         if (!delta_base_cache.cmpfn)
2264                 return NULL;
2265
2266         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2267         key.p = p;
2268         key.base_offset = base_offset;
2269         return hashmap_get(&delta_base_cache, &entry, &key);
2270 }
2271
2272 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2273                                    const struct delta_base_cache_key *b)
2274 {
2275         return a->p == b->p && a->base_offset == b->base_offset;
2276 }
2277
2278 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
2279                                      const void *va, const void *vb,
2280                                      const void *vkey)
2281 {
2282         const struct delta_base_cache_entry *a = va, *b = vb;
2283         const struct delta_base_cache_key *key = vkey;
2284         if (key)
2285                 return !delta_base_cache_key_eq(&a->key, key);
2286         else
2287                 return !delta_base_cache_key_eq(&a->key, &b->key);
2288 }
2289
2290 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2291 {
2292         return !!get_delta_base_cache_entry(p, base_offset);
2293 }
2294
2295 /*
2296  * Remove the entry from the cache, but do _not_ free the associated
2297  * entry data. The caller takes ownership of the "data" buffer, and
2298  * should copy out any fields it wants before detaching.
2299  */
2300 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2301 {
2302         hashmap_remove(&delta_base_cache, ent, &ent->key);
2303         list_del(&ent->lru);
2304         delta_base_cached -= ent->size;
2305         free(ent);
2306 }
2307
2308 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2309         unsigned long *base_size, enum object_type *type)
2310 {
2311         struct delta_base_cache_entry *ent;
2312
2313         ent = get_delta_base_cache_entry(p, base_offset);
2314         if (!ent)
2315                 return unpack_entry(p, base_offset, type, base_size);
2316
2317         if (type)
2318                 *type = ent->type;
2319         if (base_size)
2320                 *base_size = ent->size;
2321         return xmemdupz(ent->data, ent->size);
2322 }
2323
2324 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2325 {
2326         free(ent->data);
2327         detach_delta_base_cache_entry(ent);
2328 }
2329
2330 void clear_delta_base_cache(void)
2331 {
2332         struct list_head *lru, *tmp;
2333         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2334                 struct delta_base_cache_entry *entry =
2335                         list_entry(lru, struct delta_base_cache_entry, lru);
2336                 release_delta_base_cache(entry);
2337         }
2338 }
2339
2340 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2341         void *base, unsigned long base_size, enum object_type type)
2342 {
2343         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2344         struct list_head *lru, *tmp;
2345
2346         delta_base_cached += base_size;
2347
2348         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2349                 struct delta_base_cache_entry *f =
2350                         list_entry(lru, struct delta_base_cache_entry, lru);
2351                 if (delta_base_cached <= delta_base_cache_limit)
2352                         break;
2353                 release_delta_base_cache(f);
2354         }
2355
2356         ent->key.p = p;
2357         ent->key.base_offset = base_offset;
2358         ent->type = type;
2359         ent->data = base;
2360         ent->size = base_size;
2361         list_add_tail(&ent->lru, &delta_base_cache_lru);
2362
2363         if (!delta_base_cache.cmpfn)
2364                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
2365         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2366         hashmap_add(&delta_base_cache, ent);
2367 }
2368
2369 int packed_object_info(struct packed_git *p, off_t obj_offset,
2370                        struct object_info *oi)
2371 {
2372         struct pack_window *w_curs = NULL;
2373         unsigned long size;
2374         off_t curpos = obj_offset;
2375         enum object_type type;
2376
2377         /*
2378          * We always get the representation type, but only convert it to
2379          * a "real" type later if the caller is interested.
2380          */
2381         if (oi->contentp) {
2382                 *oi->contentp = cache_or_unpack_entry(p, obj_offset, oi->sizep,
2383                                                       &type);
2384                 if (!*oi->contentp)
2385                         type = OBJ_BAD;
2386         } else {
2387                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2388         }
2389
2390         if (!oi->contentp && oi->sizep) {
2391                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2392                         off_t tmp_pos = curpos;
2393                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2394                                                            type, obj_offset);
2395                         if (!base_offset) {
2396                                 type = OBJ_BAD;
2397                                 goto out;
2398                         }
2399                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2400                         if (*oi->sizep == 0) {
2401                                 type = OBJ_BAD;
2402                                 goto out;
2403                         }
2404                 } else {
2405                         *oi->sizep = size;
2406                 }
2407         }
2408
2409         if (oi->disk_sizep) {
2410                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2411                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2412         }
2413
2414         if (oi->typep || oi->typename) {
2415                 enum object_type ptot;
2416                 ptot = packed_to_object_type(p, obj_offset, type, &w_curs,
2417                                              curpos);
2418                 if (oi->typep)
2419                         *oi->typep = ptot;
2420                 if (oi->typename) {
2421                         const char *tn = typename(ptot);
2422                         if (tn)
2423                                 strbuf_addstr(oi->typename, tn);
2424                 }
2425                 if (ptot < 0) {
2426                         type = OBJ_BAD;
2427                         goto out;
2428                 }
2429         }
2430
2431         if (oi->delta_base_sha1) {
2432                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2433                         const unsigned char *base;
2434
2435                         base = get_delta_base_sha1(p, &w_curs, curpos,
2436                                                    type, obj_offset);
2437                         if (!base) {
2438                                 type = OBJ_BAD;
2439                                 goto out;
2440                         }
2441
2442                         hashcpy(oi->delta_base_sha1, base);
2443                 } else
2444                         hashclr(oi->delta_base_sha1);
2445         }
2446
2447 out:
2448         unuse_pack(&w_curs);
2449         return type;
2450 }
2451
2452 static void *unpack_compressed_entry(struct packed_git *p,
2453                                     struct pack_window **w_curs,
2454                                     off_t curpos,
2455                                     unsigned long size)
2456 {
2457         int st;
2458         git_zstream stream;
2459         unsigned char *buffer, *in;
2460
2461         buffer = xmallocz_gently(size);
2462         if (!buffer)
2463                 return NULL;
2464         memset(&stream, 0, sizeof(stream));
2465         stream.next_out = buffer;
2466         stream.avail_out = size + 1;
2467
2468         git_inflate_init(&stream);
2469         do {
2470                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2471                 stream.next_in = in;
2472                 st = git_inflate(&stream, Z_FINISH);
2473                 if (!stream.avail_out)
2474                         break; /* the payload is larger than it should be */
2475                 curpos += stream.next_in - in;
2476         } while (st == Z_OK || st == Z_BUF_ERROR);
2477         git_inflate_end(&stream);
2478         if ((st != Z_STREAM_END) || stream.total_out != size) {
2479                 free(buffer);
2480                 return NULL;
2481         }
2482
2483         return buffer;
2484 }
2485
2486 static void *read_object(const unsigned char *sha1, enum object_type *type,
2487                          unsigned long *size);
2488
2489 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2490 {
2491         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2492         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2493                          p->pack_name, (uintmax_t)obj_offset);
2494 }
2495
2496 int do_check_packed_object_crc;
2497
2498 #define UNPACK_ENTRY_STACK_PREALLOC 64
2499 struct unpack_entry_stack_ent {
2500         off_t obj_offset;
2501         off_t curpos;
2502         unsigned long size;
2503 };
2504
2505 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2506                    enum object_type *final_type, unsigned long *final_size)
2507 {
2508         struct pack_window *w_curs = NULL;
2509         off_t curpos = obj_offset;
2510         void *data = NULL;
2511         unsigned long size;
2512         enum object_type type;
2513         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2514         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2515         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2516         int base_from_cache = 0;
2517
2518         write_pack_access_log(p, obj_offset);
2519
2520         /* PHASE 1: drill down to the innermost base object */
2521         for (;;) {
2522                 off_t base_offset;
2523                 int i;
2524                 struct delta_base_cache_entry *ent;
2525
2526                 ent = get_delta_base_cache_entry(p, curpos);
2527                 if (ent) {
2528                         type = ent->type;
2529                         data = ent->data;
2530                         size = ent->size;
2531                         detach_delta_base_cache_entry(ent);
2532                         base_from_cache = 1;
2533                         break;
2534                 }
2535
2536                 if (do_check_packed_object_crc && p->index_version > 1) {
2537                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2538                         off_t len = revidx[1].offset - obj_offset;
2539                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2540                                 const unsigned char *sha1 =
2541                                         nth_packed_object_sha1(p, revidx->nr);
2542                                 error("bad packed object CRC for %s",
2543                                       sha1_to_hex(sha1));
2544                                 mark_bad_packed_object(p, sha1);
2545                                 unuse_pack(&w_curs);
2546                                 return NULL;
2547                         }
2548                 }
2549
2550                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2551                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2552                         break;
2553
2554                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2555                 if (!base_offset) {
2556                         error("failed to validate delta base reference "
2557                               "at offset %"PRIuMAX" from %s",
2558                               (uintmax_t)curpos, p->pack_name);
2559                         /* bail to phase 2, in hopes of recovery */
2560                         data = NULL;
2561                         break;
2562                 }
2563
2564                 /* push object, proceed to base */
2565                 if (delta_stack_nr >= delta_stack_alloc
2566                     && delta_stack == small_delta_stack) {
2567                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2568                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2569                         memcpy(delta_stack, small_delta_stack,
2570                                sizeof(*delta_stack)*delta_stack_nr);
2571                 } else {
2572                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2573                 }
2574                 i = delta_stack_nr++;
2575                 delta_stack[i].obj_offset = obj_offset;
2576                 delta_stack[i].curpos = curpos;
2577                 delta_stack[i].size = size;
2578
2579                 curpos = obj_offset = base_offset;
2580         }
2581
2582         /* PHASE 2: handle the base */
2583         switch (type) {
2584         case OBJ_OFS_DELTA:
2585         case OBJ_REF_DELTA:
2586                 if (data)
2587                         die("BUG: unpack_entry: left loop at a valid delta");
2588                 break;
2589         case OBJ_COMMIT:
2590         case OBJ_TREE:
2591         case OBJ_BLOB:
2592         case OBJ_TAG:
2593                 if (!base_from_cache)
2594                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2595                 break;
2596         default:
2597                 data = NULL;
2598                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2599                       type, (uintmax_t)obj_offset, p->pack_name);
2600         }
2601
2602         /* PHASE 3: apply deltas in order */
2603
2604         /* invariants:
2605          *   'data' holds the base data, or NULL if there was corruption
2606          */
2607         while (delta_stack_nr) {
2608                 void *delta_data;
2609                 void *base = data;
2610                 void *external_base = NULL;
2611                 unsigned long delta_size, base_size = size;
2612                 int i;
2613
2614                 data = NULL;
2615
2616                 if (base)
2617                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2618
2619                 if (!base) {
2620                         /*
2621                          * We're probably in deep shit, but let's try to fetch
2622                          * the required base anyway from another pack or loose.
2623                          * This is costly but should happen only in the presence
2624                          * of a corrupted pack, and is better than failing outright.
2625                          */
2626                         struct revindex_entry *revidx;
2627                         const unsigned char *base_sha1;
2628                         revidx = find_pack_revindex(p, obj_offset);
2629                         if (revidx) {
2630                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2631                                 error("failed to read delta base object %s"
2632                                       " at offset %"PRIuMAX" from %s",
2633                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2634                                       p->pack_name);
2635                                 mark_bad_packed_object(p, base_sha1);
2636                                 base = read_object(base_sha1, &type, &base_size);
2637                                 external_base = base;
2638                         }
2639                 }
2640
2641                 i = --delta_stack_nr;
2642                 obj_offset = delta_stack[i].obj_offset;
2643                 curpos = delta_stack[i].curpos;
2644                 delta_size = delta_stack[i].size;
2645
2646                 if (!base)
2647                         continue;
2648
2649                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2650
2651                 if (!delta_data) {
2652                         error("failed to unpack compressed delta "
2653                               "at offset %"PRIuMAX" from %s",
2654                               (uintmax_t)curpos, p->pack_name);
2655                         data = NULL;
2656                         free(external_base);
2657                         continue;
2658                 }
2659
2660                 data = patch_delta(base, base_size,
2661                                    delta_data, delta_size,
2662                                    &size);
2663
2664                 /*
2665                  * We could not apply the delta; warn the user, but keep going.
2666                  * Our failure will be noticed either in the next iteration of
2667                  * the loop, or if this is the final delta, in the caller when
2668                  * we return NULL. Those code paths will take care of making
2669                  * a more explicit warning and retrying with another copy of
2670                  * the object.
2671                  */
2672                 if (!data)
2673                         error("failed to apply delta");
2674
2675                 free(delta_data);
2676                 free(external_base);
2677         }
2678
2679         if (final_type)
2680                 *final_type = type;
2681         if (final_size)
2682                 *final_size = size;
2683
2684         unuse_pack(&w_curs);
2685
2686         if (delta_stack != small_delta_stack)
2687                 free(delta_stack);
2688
2689         return data;
2690 }
2691
2692 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2693                                             uint32_t n)
2694 {
2695         const unsigned char *index = p->index_data;
2696         if (!index) {
2697                 if (open_pack_index(p))
2698                         return NULL;
2699                 index = p->index_data;
2700         }
2701         if (n >= p->num_objects)
2702                 return NULL;
2703         index += 4 * 256;
2704         if (p->index_version == 1) {
2705                 return index + 24 * n + 4;
2706         } else {
2707                 index += 8;
2708                 return index + 20 * n;
2709         }
2710 }
2711
2712 const struct object_id *nth_packed_object_oid(struct object_id *oid,
2713                                               struct packed_git *p,
2714                                               uint32_t n)
2715 {
2716         const unsigned char *hash = nth_packed_object_sha1(p, n);
2717         if (!hash)
2718                 return NULL;
2719         hashcpy(oid->hash, hash);
2720         return oid;
2721 }
2722
2723 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2724 {
2725         const unsigned char *ptr = vptr;
2726         const unsigned char *start = p->index_data;
2727         const unsigned char *end = start + p->index_size;
2728         if (ptr < start)
2729                 die(_("offset before start of pack index for %s (corrupt index?)"),
2730                     p->pack_name);
2731         /* No need to check for underflow; .idx files must be at least 8 bytes */
2732         if (ptr >= end - 8)
2733                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2734                     p->pack_name);
2735 }
2736
2737 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2738 {
2739         const unsigned char *index = p->index_data;
2740         index += 4 * 256;
2741         if (p->index_version == 1) {
2742                 return ntohl(*((uint32_t *)(index + 24 * n)));
2743         } else {
2744                 uint32_t off;
2745                 index += 8 + p->num_objects * (20 + 4);
2746                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2747                 if (!(off & 0x80000000))
2748                         return off;
2749                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2750                 check_pack_index_ptr(p, index);
2751                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2752                                    ntohl(*((uint32_t *)(index + 4)));
2753         }
2754 }
2755
2756 off_t find_pack_entry_one(const unsigned char *sha1,
2757                                   struct packed_git *p)
2758 {
2759         const uint32_t *level1_ofs = p->index_data;
2760         const unsigned char *index = p->index_data;
2761         unsigned hi, lo, stride;
2762         static int use_lookup = -1;
2763         static int debug_lookup = -1;
2764
2765         if (debug_lookup < 0)
2766                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2767
2768         if (!index) {
2769                 if (open_pack_index(p))
2770                         return 0;
2771                 level1_ofs = p->index_data;
2772                 index = p->index_data;
2773         }
2774         if (p->index_version > 1) {
2775                 level1_ofs += 2;
2776                 index += 8;
2777         }
2778         index += 4 * 256;
2779         hi = ntohl(level1_ofs[*sha1]);
2780         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2781         if (p->index_version > 1) {
2782                 stride = 20;
2783         } else {
2784                 stride = 24;
2785                 index += 4;
2786         }
2787
2788         if (debug_lookup)
2789                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2790                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2791
2792         if (use_lookup < 0)
2793                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2794         if (use_lookup) {
2795                 int pos = sha1_entry_pos(index, stride, 0,
2796                                          lo, hi, p->num_objects, sha1);
2797                 if (pos < 0)
2798                         return 0;
2799                 return nth_packed_object_offset(p, pos);
2800         }
2801
2802         do {
2803                 unsigned mi = (lo + hi) / 2;
2804                 int cmp = hashcmp(index + mi * stride, sha1);
2805
2806                 if (debug_lookup)
2807                         printf("lo %u hi %u rg %u mi %u\n",
2808                                lo, hi, hi - lo, mi);
2809                 if (!cmp)
2810                         return nth_packed_object_offset(p, mi);
2811                 if (cmp > 0)
2812                         hi = mi;
2813                 else
2814                         lo = mi+1;
2815         } while (lo < hi);
2816         return 0;
2817 }
2818
2819 int is_pack_valid(struct packed_git *p)
2820 {
2821         /* An already open pack is known to be valid. */
2822         if (p->pack_fd != -1)
2823                 return 1;
2824
2825         /* If the pack has one window completely covering the
2826          * file size, the pack is known to be valid even if
2827          * the descriptor is not currently open.
2828          */
2829         if (p->windows) {
2830                 struct pack_window *w = p->windows;
2831
2832                 if (!w->offset && w->len == p->pack_size)
2833                         return 1;
2834         }
2835
2836         /* Force the pack to open to prove its valid. */
2837         return !open_packed_git(p);
2838 }
2839
2840 static int fill_pack_entry(const unsigned char *sha1,
2841                            struct pack_entry *e,
2842                            struct packed_git *p)
2843 {
2844         off_t offset;
2845
2846         if (p->num_bad_objects) {
2847                 unsigned i;
2848                 for (i = 0; i < p->num_bad_objects; i++)
2849                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2850                                 return 0;
2851         }
2852
2853         offset = find_pack_entry_one(sha1, p);
2854         if (!offset)
2855                 return 0;
2856
2857         /*
2858          * We are about to tell the caller where they can locate the
2859          * requested object.  We better make sure the packfile is
2860          * still here and can be accessed before supplying that
2861          * answer, as it may have been deleted since the index was
2862          * loaded!
2863          */
2864         if (!is_pack_valid(p))
2865                 return 0;
2866         e->offset = offset;
2867         e->p = p;
2868         hashcpy(e->sha1, sha1);
2869         return 1;
2870 }
2871
2872 /*
2873  * Iff a pack file contains the object named by sha1, return true and
2874  * store its location to e.
2875  */
2876 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2877 {
2878         struct mru_entry *p;
2879
2880         prepare_packed_git();
2881         if (!packed_git)
2882                 return 0;
2883
2884         for (p = packed_git_mru->head; p; p = p->next) {
2885                 if (fill_pack_entry(sha1, e, p->item)) {
2886                         mru_mark(packed_git_mru, p);
2887                         return 1;
2888                 }
2889         }
2890         return 0;
2891 }
2892
2893 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2894                                   struct packed_git *packs)
2895 {
2896         struct packed_git *p;
2897
2898         for (p = packs; p; p = p->next) {
2899                 if (find_pack_entry_one(sha1, p))
2900                         return p;
2901         }
2902         return NULL;
2903
2904 }
2905
2906 static int sha1_loose_object_info(const unsigned char *sha1,
2907                                   struct object_info *oi,
2908                                   int flags)
2909 {
2910         int status = 0;
2911         unsigned long mapsize;
2912         void *map;
2913         git_zstream stream;
2914         char hdr[32];
2915         struct strbuf hdrbuf = STRBUF_INIT;
2916         unsigned long size_scratch;
2917
2918         if (oi->delta_base_sha1)
2919                 hashclr(oi->delta_base_sha1);
2920
2921         /*
2922          * If we don't care about type or size, then we don't
2923          * need to look inside the object at all. Note that we
2924          * do not optimize out the stat call, even if the
2925          * caller doesn't care about the disk-size, since our
2926          * return value implicitly indicates whether the
2927          * object even exists.
2928          */
2929         if (!oi->typep && !oi->typename && !oi->sizep && !oi->contentp) {
2930                 const char *path;
2931                 struct stat st;
2932                 if (stat_sha1_file(sha1, &st, &path) < 0)
2933                         return -1;
2934                 if (oi->disk_sizep)
2935                         *oi->disk_sizep = st.st_size;
2936                 return 0;
2937         }
2938
2939         map = map_sha1_file(sha1, &mapsize);
2940         if (!map)
2941                 return -1;
2942
2943         if (!oi->sizep)
2944                 oi->sizep = &size_scratch;
2945
2946         if (oi->disk_sizep)
2947                 *oi->disk_sizep = mapsize;
2948         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE)) {
2949                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2950                         status = error("unable to unpack %s header with --allow-unknown-type",
2951                                        sha1_to_hex(sha1));
2952         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2953                 status = error("unable to unpack %s header",
2954                                sha1_to_hex(sha1));
2955         if (status < 0)
2956                 ; /* Do nothing */
2957         else if (hdrbuf.len) {
2958                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2959                         status = error("unable to parse %s header with --allow-unknown-type",
2960                                        sha1_to_hex(sha1));
2961         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2962                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2963
2964         if (status >= 0 && oi->contentp)
2965                 *oi->contentp = unpack_sha1_rest(&stream, hdr,
2966                                                  *oi->sizep, sha1);
2967         else
2968                 git_inflate_end(&stream);
2969
2970         munmap(map, mapsize);
2971         if (status && oi->typep)
2972                 *oi->typep = status;
2973         if (oi->sizep == &size_scratch)
2974                 oi->sizep = NULL;
2975         strbuf_release(&hdrbuf);
2976         return (status < 0) ? status : 0;
2977 }
2978
2979 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2980 {
2981         static struct object_info blank_oi = OBJECT_INFO_INIT;
2982         struct pack_entry e;
2983         int rtype;
2984         const unsigned char *real = (flags & OBJECT_INFO_LOOKUP_REPLACE) ?
2985                                     lookup_replace_object(sha1) :
2986                                     sha1;
2987
2988         if (!oi)
2989                 oi = &blank_oi;
2990
2991         if (!(flags & OBJECT_INFO_SKIP_CACHED)) {
2992                 struct cached_object *co = find_cached_object(real);
2993                 if (co) {
2994                         if (oi->typep)
2995                                 *(oi->typep) = co->type;
2996                         if (oi->sizep)
2997                                 *(oi->sizep) = co->size;
2998                         if (oi->disk_sizep)
2999                                 *(oi->disk_sizep) = 0;
3000                         if (oi->delta_base_sha1)
3001                                 hashclr(oi->delta_base_sha1);
3002                         if (oi->typename)
3003                                 strbuf_addstr(oi->typename, typename(co->type));
3004                         if (oi->contentp)
3005                                 *oi->contentp = xmemdupz(co->buf, co->size);
3006                         oi->whence = OI_CACHED;
3007                         return 0;
3008                 }
3009         }
3010
3011         if (!find_pack_entry(real, &e)) {
3012                 /* Most likely it's a loose object. */
3013                 if (!sha1_loose_object_info(real, oi, flags)) {
3014                         oi->whence = OI_LOOSE;
3015                         return 0;
3016                 }
3017
3018                 /* Not a loose object; someone else may have just packed it. */
3019                 if (flags & OBJECT_INFO_QUICK) {
3020                         return -1;
3021                 } else {
3022                         reprepare_packed_git();
3023                         if (!find_pack_entry(real, &e))
3024                                 return -1;
3025                 }
3026         }
3027
3028         if (oi == &blank_oi)
3029                 /*
3030                  * We know that the caller doesn't actually need the
3031                  * information below, so return early.
3032                  */
3033                 return 0;
3034
3035         rtype = packed_object_info(e.p, e.offset, oi);
3036         if (rtype < 0) {
3037                 mark_bad_packed_object(e.p, real);
3038                 return sha1_object_info_extended(real, oi, 0);
3039         } else if (in_delta_base_cache(e.p, e.offset)) {
3040                 oi->whence = OI_DBCACHED;
3041         } else {
3042                 oi->whence = OI_PACKED;
3043                 oi->u.packed.offset = e.offset;
3044                 oi->u.packed.pack = e.p;
3045                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
3046                                          rtype == OBJ_OFS_DELTA);
3047         }
3048
3049         return 0;
3050 }
3051
3052 /* returns enum object_type or negative */
3053 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
3054 {
3055         enum object_type type;
3056         struct object_info oi = OBJECT_INFO_INIT;
3057
3058         oi.typep = &type;
3059         oi.sizep = sizep;
3060         if (sha1_object_info_extended(sha1, &oi,
3061                                       OBJECT_INFO_LOOKUP_REPLACE) < 0)
3062                 return -1;
3063         return type;
3064 }
3065
3066 static void *read_packed_sha1(const unsigned char *sha1,
3067                               enum object_type *type, unsigned long *size)
3068 {
3069         struct pack_entry e;
3070         void *data;
3071
3072         if (!find_pack_entry(sha1, &e))
3073                 return NULL;
3074         data = cache_or_unpack_entry(e.p, e.offset, size, type);
3075         if (!data) {
3076                 /*
3077                  * We're probably in deep shit, but let's try to fetch
3078                  * the required object anyway from another pack or loose.
3079                  * This should happen only in the presence of a corrupted
3080                  * pack, and is better than failing outright.
3081                  */
3082                 error("failed to read object %s at offset %"PRIuMAX" from %s",
3083                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
3084                 mark_bad_packed_object(e.p, sha1);
3085                 data = read_object(sha1, type, size);
3086         }
3087         return data;
3088 }
3089
3090 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
3091                       unsigned char *sha1)
3092 {
3093         struct cached_object *co;
3094
3095         hash_sha1_file(buf, len, typename(type), sha1);
3096         if (has_sha1_file(sha1) || find_cached_object(sha1))
3097                 return 0;
3098         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
3099         co = &cached_objects[cached_object_nr++];
3100         co->size = len;
3101         co->type = type;
3102         co->buf = xmalloc(len);
3103         memcpy(co->buf, buf, len);
3104         hashcpy(co->sha1, sha1);
3105         return 0;
3106 }
3107
3108 static void *read_object(const unsigned char *sha1, enum object_type *type,
3109                          unsigned long *size)
3110 {
3111         struct object_info oi = OBJECT_INFO_INIT;
3112         void *content;
3113         oi.typep = type;
3114         oi.sizep = size;
3115         oi.contentp = &content;
3116
3117         if (sha1_object_info_extended(sha1, &oi, 0) < 0)
3118                 return NULL;
3119         return content;
3120 }
3121
3122 /*
3123  * This function dies on corrupt objects; the callers who want to
3124  * deal with them should arrange to call read_object() and give error
3125  * messages themselves.
3126  */
3127 void *read_sha1_file_extended(const unsigned char *sha1,
3128                               enum object_type *type,
3129                               unsigned long *size,
3130                               int lookup_replace)
3131 {
3132         void *data;
3133         const struct packed_git *p;
3134         const char *path;
3135         struct stat st;
3136         const unsigned char *repl = lookup_replace ? lookup_replace_object(sha1)
3137                                                    : sha1;
3138
3139         errno = 0;
3140         data = read_object(repl, type, size);
3141         if (data)
3142                 return data;
3143
3144         if (errno && errno != ENOENT)
3145                 die_errno("failed to read object %s", sha1_to_hex(sha1));
3146
3147         /* die if we replaced an object with one that does not exist */
3148         if (repl != sha1)
3149                 die("replacement %s not found for %s",
3150                     sha1_to_hex(repl), sha1_to_hex(sha1));
3151
3152         if (!stat_sha1_file(repl, &st, &path))
3153                 die("loose object %s (stored in %s) is corrupt",
3154                     sha1_to_hex(repl), path);
3155
3156         if ((p = has_packed_and_bad(repl)) != NULL)
3157                 die("packed object %s (stored in %s) is corrupt",
3158                     sha1_to_hex(repl), p->pack_name);
3159
3160         return NULL;
3161 }
3162
3163 void *read_object_with_reference(const unsigned char *sha1,
3164                                  const char *required_type_name,
3165                                  unsigned long *size,
3166                                  unsigned char *actual_sha1_return)
3167 {
3168         enum object_type type, required_type;
3169         void *buffer;
3170         unsigned long isize;
3171         unsigned char actual_sha1[20];
3172
3173         required_type = type_from_string(required_type_name);
3174         hashcpy(actual_sha1, sha1);
3175         while (1) {
3176                 int ref_length = -1;
3177                 const char *ref_type = NULL;
3178
3179                 buffer = read_sha1_file(actual_sha1, &type, &isize);
3180                 if (!buffer)
3181                         return NULL;
3182                 if (type == required_type) {
3183                         *size = isize;
3184                         if (actual_sha1_return)
3185                                 hashcpy(actual_sha1_return, actual_sha1);
3186                         return buffer;
3187                 }
3188                 /* Handle references */
3189                 else if (type == OBJ_COMMIT)
3190                         ref_type = "tree ";
3191                 else if (type == OBJ_TAG)
3192                         ref_type = "object ";
3193                 else {
3194                         free(buffer);
3195                         return NULL;
3196                 }
3197                 ref_length = strlen(ref_type);
3198
3199                 if (ref_length + 40 > isize ||
3200                     memcmp(buffer, ref_type, ref_length) ||
3201                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
3202                         free(buffer);
3203                         return NULL;
3204                 }
3205                 free(buffer);
3206                 /* Now we have the ID of the referred-to object in
3207                  * actual_sha1.  Check again. */
3208         }
3209 }
3210
3211 static void write_sha1_file_prepare(const void *buf, unsigned long len,
3212                                     const char *type, unsigned char *sha1,
3213                                     char *hdr, int *hdrlen)
3214 {
3215         git_SHA_CTX c;
3216
3217         /* Generate the header */
3218         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
3219
3220         /* Sha1.. */
3221         git_SHA1_Init(&c);
3222         git_SHA1_Update(&c, hdr, *hdrlen);
3223         git_SHA1_Update(&c, buf, len);
3224         git_SHA1_Final(sha1, &c);
3225 }
3226
3227 /*
3228  * Move the just written object into its final resting place.
3229  */
3230 int finalize_object_file(const char *tmpfile, const char *filename)
3231 {
3232         int ret = 0;
3233
3234         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3235                 goto try_rename;
3236         else if (link(tmpfile, filename))
3237                 ret = errno;
3238
3239         /*
3240          * Coda hack - coda doesn't like cross-directory links,
3241          * so we fall back to a rename, which will mean that it
3242          * won't be able to check collisions, but that's not a
3243          * big deal.
3244          *
3245          * The same holds for FAT formatted media.
3246          *
3247          * When this succeeds, we just return.  We have nothing
3248          * left to unlink.
3249          */
3250         if (ret && ret != EEXIST) {
3251         try_rename:
3252                 if (!rename(tmpfile, filename))
3253                         goto out;
3254                 ret = errno;
3255         }
3256         unlink_or_warn(tmpfile);
3257         if (ret) {
3258                 if (ret != EEXIST) {
3259                         return error_errno("unable to write sha1 filename %s", filename);
3260                 }
3261                 /* FIXME!!! Collision check here ? */
3262         }
3263
3264 out:
3265         if (adjust_shared_perm(filename))
3266                 return error("unable to set permission to '%s'", filename);
3267         return 0;
3268 }
3269
3270 static int write_buffer(int fd, const void *buf, size_t len)
3271 {
3272         if (write_in_full(fd, buf, len) < 0)
3273                 return error_errno("file write error");
3274         return 0;
3275 }
3276
3277 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3278                    unsigned char *sha1)
3279 {
3280         char hdr[32];
3281         int hdrlen = sizeof(hdr);
3282         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3283         return 0;
3284 }
3285
3286 /* Finalize a file on disk, and close it. */
3287 static void close_sha1_file(int fd)
3288 {
3289         if (fsync_object_files)
3290                 fsync_or_die(fd, "sha1 file");
3291         if (close(fd) != 0)
3292                 die_errno("error when closing sha1 file");
3293 }
3294
3295 /* Size of directory component, including the ending '/' */
3296 static inline int directory_size(const char *filename)
3297 {
3298         const char *s = strrchr(filename, '/');
3299         if (!s)
3300                 return 0;
3301         return s - filename + 1;
3302 }
3303
3304 /*
3305  * This creates a temporary file in the same directory as the final
3306  * 'filename'
3307  *
3308  * We want to avoid cross-directory filename renames, because those
3309  * can have problems on various filesystems (FAT, NFS, Coda).
3310  */
3311 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3312 {
3313         int fd, dirlen = directory_size(filename);
3314
3315         strbuf_reset(tmp);
3316         strbuf_add(tmp, filename, dirlen);
3317         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3318         fd = git_mkstemp_mode(tmp->buf, 0444);
3319         if (fd < 0 && dirlen && errno == ENOENT) {
3320                 /*
3321                  * Make sure the directory exists; note that the contents
3322                  * of the buffer are undefined after mkstemp returns an
3323                  * error, so we have to rewrite the whole buffer from
3324                  * scratch.
3325                  */
3326                 strbuf_reset(tmp);
3327                 strbuf_add(tmp, filename, dirlen - 1);
3328                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3329                         return -1;
3330                 if (adjust_shared_perm(tmp->buf))
3331                         return -1;
3332
3333                 /* Try again */
3334                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3335                 fd = git_mkstemp_mode(tmp->buf, 0444);
3336         }
3337         return fd;
3338 }
3339
3340 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3341                               const void *buf, unsigned long len, time_t mtime)
3342 {
3343         int fd, ret;
3344         unsigned char compressed[4096];
3345         git_zstream stream;
3346         git_SHA_CTX c;
3347         unsigned char parano_sha1[20];
3348         static struct strbuf tmp_file = STRBUF_INIT;
3349         const char *filename = sha1_file_name(sha1);
3350
3351         fd = create_tmpfile(&tmp_file, filename);
3352         if (fd < 0) {
3353                 if (errno == EACCES)
3354                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3355                 else
3356                         return error_errno("unable to create temporary file");
3357         }
3358
3359         /* Set it up */
3360         git_deflate_init(&stream, zlib_compression_level);
3361         stream.next_out = compressed;
3362         stream.avail_out = sizeof(compressed);
3363         git_SHA1_Init(&c);
3364
3365         /* First header.. */
3366         stream.next_in = (unsigned char *)hdr;
3367         stream.avail_in = hdrlen;
3368         while (git_deflate(&stream, 0) == Z_OK)
3369                 ; /* nothing */
3370         git_SHA1_Update(&c, hdr, hdrlen);
3371
3372         /* Then the data itself.. */
3373         stream.next_in = (void *)buf;
3374         stream.avail_in = len;
3375         do {
3376                 unsigned char *in0 = stream.next_in;
3377                 ret = git_deflate(&stream, Z_FINISH);
3378                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3379                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3380                         die("unable to write sha1 file");
3381                 stream.next_out = compressed;
3382                 stream.avail_out = sizeof(compressed);
3383         } while (ret == Z_OK);
3384
3385         if (ret != Z_STREAM_END)
3386                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3387         ret = git_deflate_end_gently(&stream);
3388         if (ret != Z_OK)
3389                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3390         git_SHA1_Final(parano_sha1, &c);
3391         if (hashcmp(sha1, parano_sha1) != 0)
3392                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3393
3394         close_sha1_file(fd);
3395
3396         if (mtime) {
3397                 struct utimbuf utb;
3398                 utb.actime = mtime;
3399                 utb.modtime = mtime;
3400                 if (utime(tmp_file.buf, &utb) < 0)
3401                         warning_errno("failed utime() on %s", tmp_file.buf);
3402         }
3403
3404         return finalize_object_file(tmp_file.buf, filename);
3405 }
3406
3407 static int freshen_loose_object(const unsigned char *sha1)
3408 {
3409         return check_and_freshen(sha1, 1);
3410 }
3411
3412 static int freshen_packed_object(const unsigned char *sha1)
3413 {
3414         struct pack_entry e;
3415         if (!find_pack_entry(sha1, &e))
3416                 return 0;
3417         if (e.p->freshened)
3418                 return 1;
3419         if (!freshen_file(e.p->pack_name))
3420                 return 0;
3421         e.p->freshened = 1;
3422         return 1;
3423 }
3424
3425 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3426 {
3427         char hdr[32];
3428         int hdrlen = sizeof(hdr);
3429
3430         /* Normally if we have it in the pack then we do not bother writing
3431          * it out into .git/objects/??/?{38} file.
3432          */
3433         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3434         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3435                 return 0;
3436         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3437 }
3438
3439 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3440                              unsigned char *sha1, unsigned flags)
3441 {
3442         char *header;
3443         int hdrlen, status = 0;
3444
3445         /* type string, SP, %lu of the length plus NUL must fit this */
3446         hdrlen = strlen(type) + 32;
3447         header = xmalloc(hdrlen);
3448         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3449
3450         if (!(flags & HASH_WRITE_OBJECT))
3451                 goto cleanup;
3452         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3453                 goto cleanup;
3454         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3455
3456 cleanup:
3457         free(header);
3458         return status;
3459 }
3460
3461 int force_object_loose(const unsigned char *sha1, time_t mtime)
3462 {
3463         void *buf;
3464         unsigned long len;
3465         enum object_type type;
3466         char hdr[32];
3467         int hdrlen;
3468         int ret;
3469
3470         if (has_loose_object(sha1))
3471                 return 0;
3472         buf = read_packed_sha1(sha1, &type, &len);
3473         if (!buf)
3474                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3475         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3476         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3477         free(buf);
3478
3479         return ret;
3480 }
3481
3482 int has_pack_index(const unsigned char *sha1)
3483 {
3484         struct stat st;
3485         if (stat(sha1_pack_index_name(sha1), &st))
3486                 return 0;
3487         return 1;
3488 }
3489
3490 int has_sha1_pack(const unsigned char *sha1)
3491 {
3492         struct pack_entry e;
3493         return find_pack_entry(sha1, &e);
3494 }
3495
3496 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3497 {
3498         if (!startup_info->have_repository)
3499                 return 0;
3500         return sha1_object_info_extended(sha1, NULL,
3501                                          flags | OBJECT_INFO_SKIP_CACHED) >= 0;
3502 }
3503
3504 int has_object_file(const struct object_id *oid)
3505 {
3506         return has_sha1_file(oid->hash);
3507 }
3508
3509 int has_object_file_with_flags(const struct object_id *oid, int flags)
3510 {
3511         return has_sha1_file_with_flags(oid->hash, flags);
3512 }
3513
3514 static void check_tree(const void *buf, size_t size)
3515 {
3516         struct tree_desc desc;
3517         struct name_entry entry;
3518
3519         init_tree_desc(&desc, buf, size);
3520         while (tree_entry(&desc, &entry))
3521                 /* do nothing
3522                  * tree_entry() will die() on malformed entries */
3523                 ;
3524 }
3525
3526 static void check_commit(const void *buf, size_t size)
3527 {
3528         struct commit c;
3529         memset(&c, 0, sizeof(c));
3530         if (parse_commit_buffer(&c, buf, size))
3531                 die("corrupt commit");
3532 }
3533
3534 static void check_tag(const void *buf, size_t size)
3535 {
3536         struct tag t;
3537         memset(&t, 0, sizeof(t));
3538         if (parse_tag_buffer(&t, buf, size))
3539                 die("corrupt tag");
3540 }
3541
3542 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3543                      enum object_type type,
3544                      const char *path, unsigned flags)
3545 {
3546         int ret, re_allocated = 0;
3547         int write_object = flags & HASH_WRITE_OBJECT;
3548
3549         if (!type)
3550                 type = OBJ_BLOB;
3551
3552         /*
3553          * Convert blobs to git internal format
3554          */
3555         if ((type == OBJ_BLOB) && path) {
3556                 struct strbuf nbuf = STRBUF_INIT;
3557                 if (convert_to_git(&the_index, path, buf, size, &nbuf,
3558                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3559                         buf = strbuf_detach(&nbuf, &size);
3560                         re_allocated = 1;
3561                 }
3562         }
3563         if (flags & HASH_FORMAT_CHECK) {
3564                 if (type == OBJ_TREE)
3565                         check_tree(buf, size);
3566                 if (type == OBJ_COMMIT)
3567                         check_commit(buf, size);
3568                 if (type == OBJ_TAG)
3569                         check_tag(buf, size);
3570         }
3571
3572         if (write_object)
3573                 ret = write_sha1_file(buf, size, typename(type), sha1);
3574         else
3575                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3576         if (re_allocated)
3577                 free(buf);
3578         return ret;
3579 }
3580
3581 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3582                                      const char *path, unsigned flags)
3583 {
3584         int ret;
3585         const int write_object = flags & HASH_WRITE_OBJECT;
3586         struct strbuf sbuf = STRBUF_INIT;
3587
3588         assert(path);
3589         assert(would_convert_to_git_filter_fd(path));
3590
3591         convert_to_git_filter_fd(&the_index, path, fd, &sbuf,
3592                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3593
3594         if (write_object)
3595                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3596                                       sha1);
3597         else
3598                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3599                                      sha1);
3600         strbuf_release(&sbuf);
3601         return ret;
3602 }
3603
3604 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3605                       const char *path, unsigned flags)
3606 {
3607         struct strbuf sbuf = STRBUF_INIT;
3608         int ret;
3609
3610         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3611                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3612         else
3613                 ret = -1;
3614         strbuf_release(&sbuf);
3615         return ret;
3616 }
3617
3618 #define SMALL_FILE_SIZE (32*1024)
3619
3620 static int index_core(unsigned char *sha1, int fd, size_t size,
3621                       enum object_type type, const char *path,
3622                       unsigned flags)
3623 {
3624         int ret;
3625
3626         if (!size) {
3627                 ret = index_mem(sha1, "", size, type, path, flags);
3628         } else if (size <= SMALL_FILE_SIZE) {
3629                 char *buf = xmalloc(size);
3630                 if (size == read_in_full(fd, buf, size))
3631                         ret = index_mem(sha1, buf, size, type, path, flags);
3632                 else
3633                         ret = error_errno("short read");
3634                 free(buf);
3635         } else {
3636                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3637                 ret = index_mem(sha1, buf, size, type, path, flags);
3638                 munmap(buf, size);
3639         }
3640         return ret;
3641 }
3642
3643 /*
3644  * This creates one packfile per large blob unless bulk-checkin
3645  * machinery is "plugged".
3646  *
3647  * This also bypasses the usual "convert-to-git" dance, and that is on
3648  * purpose. We could write a streaming version of the converting
3649  * functions and insert that before feeding the data to fast-import
3650  * (or equivalent in-core API described above). However, that is
3651  * somewhat complicated, as we do not know the size of the filter
3652  * result, which we need to know beforehand when writing a git object.
3653  * Since the primary motivation for trying to stream from the working
3654  * tree file and to avoid mmaping it in core is to deal with large
3655  * binary blobs, they generally do not want to get any conversion, and
3656  * callers should avoid this code path when filters are requested.
3657  */
3658 static int index_stream(unsigned char *sha1, int fd, size_t size,
3659                         enum object_type type, const char *path,
3660                         unsigned flags)
3661 {
3662         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3663 }
3664
3665 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3666              enum object_type type, const char *path, unsigned flags)
3667 {
3668         int ret;
3669
3670         /*
3671          * Call xsize_t() only when needed to avoid potentially unnecessary
3672          * die() for large files.
3673          */
3674         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3675                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3676         else if (!S_ISREG(st->st_mode))
3677                 ret = index_pipe(sha1, fd, type, path, flags);
3678         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3679                  (path && would_convert_to_git(&the_index, path)))
3680                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3681                                  flags);
3682         else
3683                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3684                                    flags);
3685         close(fd);
3686         return ret;
3687 }
3688
3689 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3690 {
3691         int fd;
3692         struct strbuf sb = STRBUF_INIT;
3693
3694         switch (st->st_mode & S_IFMT) {
3695         case S_IFREG:
3696                 fd = open(path, O_RDONLY);
3697                 if (fd < 0)
3698                         return error_errno("open(\"%s\")", path);
3699                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3700                         return error("%s: failed to insert into database",
3701                                      path);
3702                 break;
3703         case S_IFLNK:
3704                 if (strbuf_readlink(&sb, path, st->st_size))
3705                         return error_errno("readlink(\"%s\")", path);
3706                 if (!(flags & HASH_WRITE_OBJECT))
3707                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3708                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3709                         return error("%s: failed to insert into database",
3710                                      path);
3711                 strbuf_release(&sb);
3712                 break;
3713         case S_IFDIR:
3714                 return resolve_gitlink_ref(path, "HEAD", sha1);
3715         default:
3716                 return error("%s: unsupported file type", path);
3717         }
3718         return 0;
3719 }
3720
3721 int read_pack_header(int fd, struct pack_header *header)
3722 {
3723         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3724                 /* "eof before pack header was fully read" */
3725                 return PH_ERROR_EOF;
3726
3727         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3728                 /* "protocol error (pack signature mismatch detected)" */
3729                 return PH_ERROR_PACK_SIGNATURE;
3730         if (!pack_version_ok(header->hdr_version))
3731                 /* "protocol error (pack version unsupported)" */
3732                 return PH_ERROR_PROTOCOL;
3733         return 0;
3734 }
3735
3736 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3737 {
3738         enum object_type type = sha1_object_info(sha1, NULL);
3739         if (type < 0)
3740                 die("%s is not a valid object", sha1_to_hex(sha1));
3741         if (type != expect)
3742                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3743                     typename(expect));
3744 }
3745
3746 int for_each_file_in_obj_subdir(unsigned int subdir_nr,
3747                                 struct strbuf *path,
3748                                 each_loose_object_fn obj_cb,
3749                                 each_loose_cruft_fn cruft_cb,
3750                                 each_loose_subdir_fn subdir_cb,
3751                                 void *data)
3752 {
3753         size_t origlen, baselen;
3754         DIR *dir;
3755         struct dirent *de;
3756         int r = 0;
3757
3758         if (subdir_nr > 0xff)
3759                 BUG("invalid loose object subdirectory: %x", subdir_nr);
3760
3761         origlen = path->len;
3762         strbuf_complete(path, '/');
3763         strbuf_addf(path, "%02x", subdir_nr);
3764         baselen = path->len;
3765
3766         dir = opendir(path->buf);
3767         if (!dir) {
3768                 if (errno != ENOENT)
3769                         r = error_errno("unable to open %s", path->buf);
3770                 strbuf_setlen(path, origlen);
3771                 return r;
3772         }
3773
3774         while ((de = readdir(dir))) {
3775                 if (is_dot_or_dotdot(de->d_name))
3776                         continue;
3777
3778                 strbuf_setlen(path, baselen);
3779                 strbuf_addf(path, "/%s", de->d_name);
3780
3781                 if (strlen(de->d_name) == GIT_SHA1_HEXSZ - 2)  {
3782                         char hex[GIT_MAX_HEXSZ+1];
3783                         struct object_id oid;
3784
3785                         xsnprintf(hex, sizeof(hex), "%02x%s",
3786                                   subdir_nr, de->d_name);
3787                         if (!get_oid_hex(hex, &oid)) {
3788                                 if (obj_cb) {
3789                                         r = obj_cb(&oid, path->buf, data);
3790                                         if (r)
3791                                                 break;
3792                                 }
3793                                 continue;
3794                         }
3795                 }
3796
3797                 if (cruft_cb) {
3798                         r = cruft_cb(de->d_name, path->buf, data);
3799                         if (r)
3800                                 break;
3801                 }
3802         }
3803         closedir(dir);
3804
3805         strbuf_setlen(path, baselen);
3806         if (!r && subdir_cb)
3807                 r = subdir_cb(subdir_nr, path->buf, data);
3808
3809         strbuf_setlen(path, origlen);
3810
3811         return r;
3812 }
3813
3814 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3815                             each_loose_object_fn obj_cb,
3816                             each_loose_cruft_fn cruft_cb,
3817                             each_loose_subdir_fn subdir_cb,
3818                             void *data)
3819 {
3820         int r = 0;
3821         int i;
3822
3823         for (i = 0; i < 256; i++) {
3824                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3825                                                 subdir_cb, data);
3826                 if (r)
3827                         break;
3828         }
3829
3830         return r;
3831 }
3832
3833 int for_each_loose_file_in_objdir(const char *path,
3834                                   each_loose_object_fn obj_cb,
3835                                   each_loose_cruft_fn cruft_cb,
3836                                   each_loose_subdir_fn subdir_cb,
3837                                   void *data)
3838 {
3839         struct strbuf buf = STRBUF_INIT;
3840         int r;
3841
3842         strbuf_addstr(&buf, path);
3843         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3844                                               subdir_cb, data);
3845         strbuf_release(&buf);
3846
3847         return r;
3848 }
3849
3850 struct loose_alt_odb_data {
3851         each_loose_object_fn *cb;
3852         void *data;
3853 };
3854
3855 static int loose_from_alt_odb(struct alternate_object_database *alt,
3856                               void *vdata)
3857 {
3858         struct loose_alt_odb_data *data = vdata;
3859         struct strbuf buf = STRBUF_INIT;
3860         int r;
3861
3862         strbuf_addstr(&buf, alt->path);
3863         r = for_each_loose_file_in_objdir_buf(&buf,
3864                                               data->cb, NULL, NULL,
3865                                               data->data);
3866         strbuf_release(&buf);
3867         return r;
3868 }
3869
3870 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3871 {
3872         struct loose_alt_odb_data alt;
3873         int r;
3874
3875         r = for_each_loose_file_in_objdir(get_object_directory(),
3876                                           cb, NULL, NULL, data);
3877         if (r)
3878                 return r;
3879
3880         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3881                 return 0;
3882
3883         alt.cb = cb;
3884         alt.data = data;
3885         return foreach_alt_odb(loose_from_alt_odb, &alt);
3886 }
3887
3888 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3889 {
3890         uint32_t i;
3891         int r = 0;
3892
3893         for (i = 0; i < p->num_objects; i++) {
3894                 struct object_id oid;
3895
3896                 if (!nth_packed_object_oid(&oid, p, i))
3897                         return error("unable to get sha1 of object %u in %s",
3898                                      i, p->pack_name);
3899
3900                 r = cb(&oid, p, i, data);
3901                 if (r)
3902                         break;
3903         }
3904         return r;
3905 }
3906
3907 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3908 {
3909         struct packed_git *p;
3910         int r = 0;
3911         int pack_errors = 0;
3912
3913         prepare_packed_git();
3914         for (p = packed_git; p; p = p->next) {
3915                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3916                         continue;
3917                 if (open_pack_index(p)) {
3918                         pack_errors = 1;
3919                         continue;
3920                 }
3921                 r = for_each_object_in_pack(p, cb, data);
3922                 if (r)
3923                         break;
3924         }
3925         return r ? r : pack_errors;
3926 }
3927
3928 static int check_stream_sha1(git_zstream *stream,
3929                              const char *hdr,
3930                              unsigned long size,
3931                              const char *path,
3932                              const unsigned char *expected_sha1)
3933 {
3934         git_SHA_CTX c;
3935         unsigned char real_sha1[GIT_MAX_RAWSZ];
3936         unsigned char buf[4096];
3937         unsigned long total_read;
3938         int status = Z_OK;
3939
3940         git_SHA1_Init(&c);
3941         git_SHA1_Update(&c, hdr, stream->total_out);
3942
3943         /*
3944          * We already read some bytes into hdr, but the ones up to the NUL
3945          * do not count against the object's content size.
3946          */
3947         total_read = stream->total_out - strlen(hdr) - 1;
3948
3949         /*
3950          * This size comparison must be "<=" to read the final zlib packets;
3951          * see the comment in unpack_sha1_rest for details.
3952          */
3953         while (total_read <= size &&
3954                (status == Z_OK || status == Z_BUF_ERROR)) {
3955                 stream->next_out = buf;
3956                 stream->avail_out = sizeof(buf);
3957                 if (size - total_read < stream->avail_out)
3958                         stream->avail_out = size - total_read;
3959                 status = git_inflate(stream, Z_FINISH);
3960                 git_SHA1_Update(&c, buf, stream->next_out - buf);
3961                 total_read += stream->next_out - buf;
3962         }
3963         git_inflate_end(stream);
3964
3965         if (status != Z_STREAM_END) {
3966                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
3967                 return -1;
3968         }
3969         if (stream->avail_in) {
3970                 error("garbage at end of loose object '%s'",
3971                       sha1_to_hex(expected_sha1));
3972                 return -1;
3973         }
3974
3975         git_SHA1_Final(real_sha1, &c);
3976         if (hashcmp(expected_sha1, real_sha1)) {
3977                 error("sha1 mismatch for %s (expected %s)", path,
3978                       sha1_to_hex(expected_sha1));
3979                 return -1;
3980         }
3981
3982         return 0;
3983 }
3984
3985 int read_loose_object(const char *path,
3986                       const unsigned char *expected_sha1,
3987                       enum object_type *type,
3988                       unsigned long *size,
3989                       void **contents)
3990 {
3991         int ret = -1;
3992         void *map = NULL;
3993         unsigned long mapsize;
3994         git_zstream stream;
3995         char hdr[32];
3996
3997         *contents = NULL;
3998
3999         map = map_sha1_file_1(path, NULL, &mapsize);
4000         if (!map) {
4001                 error_errno("unable to mmap %s", path);
4002                 goto out;
4003         }
4004
4005         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
4006                 error("unable to unpack header of %s", path);
4007                 goto out;
4008         }
4009
4010         *type = parse_sha1_header(hdr, size);
4011         if (*type < 0) {
4012                 error("unable to parse header of %s", path);
4013                 git_inflate_end(&stream);
4014                 goto out;
4015         }
4016
4017         if (*type == OBJ_BLOB) {
4018                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
4019                         goto out;
4020         } else {
4021                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
4022                 if (!*contents) {
4023                         error("unable to unpack contents of %s", path);
4024                         git_inflate_end(&stream);
4025                         goto out;
4026                 }
4027                 if (check_sha1_signature(expected_sha1, *contents,
4028                                          *size, typename(*type))) {
4029                         error("sha1 mismatch for %s (expected %s)", path,
4030                               sha1_to_hex(expected_sha1));
4031                         free(*contents);
4032                         goto out;
4033                 }
4034         }
4035
4036         ret = 0; /* everything checks out */
4037
4038 out:
4039         if (map)
4040                 munmap(map, mapsize);
4041         return ret;
4042 }