Imported Upstream version 2.14.3
[platform/upstream/git.git] / sha1_file.c
1 /*
2  * GIT - The information manager from hell
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This handles basic git sha1 object files - packing, unpacking,
7  * creation etc.
8  */
9 #include "cache.h"
10 #include "config.h"
11 #include "string-list.h"
12 #include "lockfile.h"
13 #include "delta.h"
14 #include "pack.h"
15 #include "blob.h"
16 #include "commit.h"
17 #include "run-command.h"
18 #include "tag.h"
19 #include "tree.h"
20 #include "tree-walk.h"
21 #include "refs.h"
22 #include "pack-revindex.h"
23 #include "sha1-lookup.h"
24 #include "bulk-checkin.h"
25 #include "streaming.h"
26 #include "dir.h"
27 #include "mru.h"
28 #include "list.h"
29 #include "mergesort.h"
30 #include "quote.h"
31
32 #define SZ_FMT PRIuMAX
33 static inline uintmax_t sz_fmt(size_t s) { return s; }
34
35 const unsigned char null_sha1[20];
36 const struct object_id null_oid;
37 const struct object_id empty_tree_oid = {
38         EMPTY_TREE_SHA1_BIN_LITERAL
39 };
40 const struct object_id empty_blob_oid = {
41         EMPTY_BLOB_SHA1_BIN_LITERAL
42 };
43
44 /*
45  * This is meant to hold a *small* number of objects that you would
46  * want read_sha1_file() to be able to return, but yet you do not want
47  * to write them into the object store (e.g. a browse-only
48  * application).
49  */
50 static struct cached_object {
51         unsigned char sha1[20];
52         enum object_type type;
53         void *buf;
54         unsigned long size;
55 } *cached_objects;
56 static int cached_object_nr, cached_object_alloc;
57
58 static struct cached_object empty_tree = {
59         EMPTY_TREE_SHA1_BIN_LITERAL,
60         OBJ_TREE,
61         "",
62         0
63 };
64
65 static struct cached_object *find_cached_object(const unsigned char *sha1)
66 {
67         int i;
68         struct cached_object *co = cached_objects;
69
70         for (i = 0; i < cached_object_nr; i++, co++) {
71                 if (!hashcmp(co->sha1, sha1))
72                         return co;
73         }
74         if (!hashcmp(sha1, empty_tree.sha1))
75                 return &empty_tree;
76         return NULL;
77 }
78
79 int mkdir_in_gitdir(const char *path)
80 {
81         if (mkdir(path, 0777)) {
82                 int saved_errno = errno;
83                 struct stat st;
84                 struct strbuf sb = STRBUF_INIT;
85
86                 if (errno != EEXIST)
87                         return -1;
88                 /*
89                  * Are we looking at a path in a symlinked worktree
90                  * whose original repository does not yet have it?
91                  * e.g. .git/rr-cache pointing at its original
92                  * repository in which the user hasn't performed any
93                  * conflict resolution yet?
94                  */
95                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
96                     strbuf_readlink(&sb, path, st.st_size) ||
97                     !is_absolute_path(sb.buf) ||
98                     mkdir(sb.buf, 0777)) {
99                         strbuf_release(&sb);
100                         errno = saved_errno;
101                         return -1;
102                 }
103                 strbuf_release(&sb);
104         }
105         return adjust_shared_perm(path);
106 }
107
108 enum scld_error safe_create_leading_directories(char *path)
109 {
110         char *next_component = path + offset_1st_component(path);
111         enum scld_error ret = SCLD_OK;
112
113         while (ret == SCLD_OK && next_component) {
114                 struct stat st;
115                 char *slash = next_component, slash_character;
116
117                 while (*slash && !is_dir_sep(*slash))
118                         slash++;
119
120                 if (!*slash)
121                         break;
122
123                 next_component = slash + 1;
124                 while (is_dir_sep(*next_component))
125                         next_component++;
126                 if (!*next_component)
127                         break;
128
129                 slash_character = *slash;
130                 *slash = '\0';
131                 if (!stat(path, &st)) {
132                         /* path exists */
133                         if (!S_ISDIR(st.st_mode)) {
134                                 errno = ENOTDIR;
135                                 ret = SCLD_EXISTS;
136                         }
137                 } else if (mkdir(path, 0777)) {
138                         if (errno == EEXIST &&
139                             !stat(path, &st) && S_ISDIR(st.st_mode))
140                                 ; /* somebody created it since we checked */
141                         else if (errno == ENOENT)
142                                 /*
143                                  * Either mkdir() failed because
144                                  * somebody just pruned the containing
145                                  * directory, or stat() failed because
146                                  * the file that was in our way was
147                                  * just removed.  Either way, inform
148                                  * the caller that it might be worth
149                                  * trying again:
150                                  */
151                                 ret = SCLD_VANISHED;
152                         else
153                                 ret = SCLD_FAILED;
154                 } else if (adjust_shared_perm(path)) {
155                         ret = SCLD_PERMS;
156                 }
157                 *slash = slash_character;
158         }
159         return ret;
160 }
161
162 enum scld_error safe_create_leading_directories_const(const char *path)
163 {
164         int save_errno;
165         /* path points to cache entries, so xstrdup before messing with it */
166         char *buf = xstrdup(path);
167         enum scld_error result = safe_create_leading_directories(buf);
168
169         save_errno = errno;
170         free(buf);
171         errno = save_errno;
172         return result;
173 }
174
175 int raceproof_create_file(const char *path, create_file_fn fn, void *cb)
176 {
177         /*
178          * The number of times we will try to remove empty directories
179          * in the way of path. This is only 1 because if another
180          * process is racily creating directories that conflict with
181          * us, we don't want to fight against them.
182          */
183         int remove_directories_remaining = 1;
184
185         /*
186          * The number of times that we will try to create the
187          * directories containing path. We are willing to attempt this
188          * more than once, because another process could be trying to
189          * clean up empty directories at the same time as we are
190          * trying to create them.
191          */
192         int create_directories_remaining = 3;
193
194         /* A scratch copy of path, filled lazily if we need it: */
195         struct strbuf path_copy = STRBUF_INIT;
196
197         int ret, save_errno;
198
199         /* Sanity check: */
200         assert(*path);
201
202 retry_fn:
203         ret = fn(path, cb);
204         save_errno = errno;
205         if (!ret)
206                 goto out;
207
208         if (errno == EISDIR && remove_directories_remaining-- > 0) {
209                 /*
210                  * A directory is in the way. Maybe it is empty; try
211                  * to remove it:
212                  */
213                 if (!path_copy.len)
214                         strbuf_addstr(&path_copy, path);
215
216                 if (!remove_dir_recursively(&path_copy, REMOVE_DIR_EMPTY_ONLY))
217                         goto retry_fn;
218         } else if (errno == ENOENT && create_directories_remaining-- > 0) {
219                 /*
220                  * Maybe the containing directory didn't exist, or
221                  * maybe it was just deleted by a process that is
222                  * racing with us to clean up empty directories. Try
223                  * to create it:
224                  */
225                 enum scld_error scld_result;
226
227                 if (!path_copy.len)
228                         strbuf_addstr(&path_copy, path);
229
230                 do {
231                         scld_result = safe_create_leading_directories(path_copy.buf);
232                         if (scld_result == SCLD_OK)
233                                 goto retry_fn;
234                 } while (scld_result == SCLD_VANISHED && create_directories_remaining-- > 0);
235         }
236
237 out:
238         strbuf_release(&path_copy);
239         errno = save_errno;
240         return ret;
241 }
242
243 static void fill_sha1_path(struct strbuf *buf, const unsigned char *sha1)
244 {
245         int i;
246         for (i = 0; i < 20; i++) {
247                 static char hex[] = "0123456789abcdef";
248                 unsigned int val = sha1[i];
249                 strbuf_addch(buf, hex[val >> 4]);
250                 strbuf_addch(buf, hex[val & 0xf]);
251                 if (!i)
252                         strbuf_addch(buf, '/');
253         }
254 }
255
256 const char *sha1_file_name(const unsigned char *sha1)
257 {
258         static struct strbuf buf = STRBUF_INIT;
259
260         strbuf_reset(&buf);
261         strbuf_addf(&buf, "%s/", get_object_directory());
262
263         fill_sha1_path(&buf, sha1);
264         return buf.buf;
265 }
266
267 struct strbuf *alt_scratch_buf(struct alternate_object_database *alt)
268 {
269         strbuf_setlen(&alt->scratch, alt->base_len);
270         return &alt->scratch;
271 }
272
273 static const char *alt_sha1_path(struct alternate_object_database *alt,
274                                  const unsigned char *sha1)
275 {
276         struct strbuf *buf = alt_scratch_buf(alt);
277         fill_sha1_path(buf, sha1);
278         return buf->buf;
279 }
280
281  char *odb_pack_name(struct strbuf *buf,
282                      const unsigned char *sha1,
283                      const char *ext)
284 {
285         strbuf_reset(buf);
286         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
287                     sha1_to_hex(sha1), ext);
288         return buf->buf;
289 }
290
291 char *sha1_pack_name(const unsigned char *sha1)
292 {
293         static struct strbuf buf = STRBUF_INIT;
294         return odb_pack_name(&buf, sha1, "pack");
295 }
296
297 char *sha1_pack_index_name(const unsigned char *sha1)
298 {
299         static struct strbuf buf = STRBUF_INIT;
300         return odb_pack_name(&buf, sha1, "idx");
301 }
302
303 struct alternate_object_database *alt_odb_list;
304 static struct alternate_object_database **alt_odb_tail;
305
306 /*
307  * Return non-zero iff the path is usable as an alternate object database.
308  */
309 static int alt_odb_usable(struct strbuf *path, const char *normalized_objdir)
310 {
311         struct alternate_object_database *alt;
312
313         /* Detect cases where alternate disappeared */
314         if (!is_directory(path->buf)) {
315                 error("object directory %s does not exist; "
316                       "check .git/objects/info/alternates.",
317                       path->buf);
318                 return 0;
319         }
320
321         /*
322          * Prevent the common mistake of listing the same
323          * thing twice, or object directory itself.
324          */
325         for (alt = alt_odb_list; alt; alt = alt->next) {
326                 if (!fspathcmp(path->buf, alt->path))
327                         return 0;
328         }
329         if (!fspathcmp(path->buf, normalized_objdir))
330                 return 0;
331
332         return 1;
333 }
334
335 /*
336  * Prepare alternate object database registry.
337  *
338  * The variable alt_odb_list points at the list of struct
339  * alternate_object_database.  The elements on this list come from
340  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
341  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
342  * whose contents is similar to that environment variable but can be
343  * LF separated.  Its base points at a statically allocated buffer that
344  * contains "/the/directory/corresponding/to/.git/objects/...", while
345  * its name points just after the slash at the end of ".git/objects/"
346  * in the example above, and has enough space to hold 40-byte hex
347  * SHA1, an extra slash for the first level indirection, and the
348  * terminating NUL.
349  */
350 static void read_info_alternates(const char * relative_base, int depth);
351 static int link_alt_odb_entry(const char *entry, const char *relative_base,
352         int depth, const char *normalized_objdir)
353 {
354         struct alternate_object_database *ent;
355         struct strbuf pathbuf = STRBUF_INIT;
356
357         if (!is_absolute_path(entry) && relative_base) {
358                 strbuf_realpath(&pathbuf, relative_base, 1);
359                 strbuf_addch(&pathbuf, '/');
360         }
361         strbuf_addstr(&pathbuf, entry);
362
363         if (strbuf_normalize_path(&pathbuf) < 0 && relative_base) {
364                 error("unable to normalize alternate object path: %s",
365                       pathbuf.buf);
366                 strbuf_release(&pathbuf);
367                 return -1;
368         }
369
370         /*
371          * The trailing slash after the directory name is given by
372          * this function at the end. Remove duplicates.
373          */
374         while (pathbuf.len && pathbuf.buf[pathbuf.len - 1] == '/')
375                 strbuf_setlen(&pathbuf, pathbuf.len - 1);
376
377         if (!alt_odb_usable(&pathbuf, normalized_objdir)) {
378                 strbuf_release(&pathbuf);
379                 return -1;
380         }
381
382         ent = alloc_alt_odb(pathbuf.buf);
383
384         /* add the alternate entry */
385         *alt_odb_tail = ent;
386         alt_odb_tail = &(ent->next);
387         ent->next = NULL;
388
389         /* recursively add alternates */
390         read_info_alternates(pathbuf.buf, depth + 1);
391
392         strbuf_release(&pathbuf);
393         return 0;
394 }
395
396 static const char *parse_alt_odb_entry(const char *string,
397                                        int sep,
398                                        struct strbuf *out)
399 {
400         const char *end;
401
402         strbuf_reset(out);
403
404         if (*string == '#') {
405                 /* comment; consume up to next separator */
406                 end = strchrnul(string, sep);
407         } else if (*string == '"' && !unquote_c_style(out, string, &end)) {
408                 /*
409                  * quoted path; unquote_c_style has copied the
410                  * data for us and set "end". Broken quoting (e.g.,
411                  * an entry that doesn't end with a quote) falls
412                  * back to the unquoted case below.
413                  */
414         } else {
415                 /* normal, unquoted path */
416                 end = strchrnul(string, sep);
417                 strbuf_add(out, string, end - string);
418         }
419
420         if (*end)
421                 end++;
422         return end;
423 }
424
425 static void link_alt_odb_entries(const char *alt, int sep,
426                                  const char *relative_base, int depth)
427 {
428         struct strbuf objdirbuf = STRBUF_INIT;
429         struct strbuf entry = STRBUF_INIT;
430
431         if (depth > 5) {
432                 error("%s: ignoring alternate object stores, nesting too deep.",
433                                 relative_base);
434                 return;
435         }
436
437         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
438         if (strbuf_normalize_path(&objdirbuf) < 0)
439                 die("unable to normalize object directory: %s",
440                     objdirbuf.buf);
441
442         while (*alt) {
443                 alt = parse_alt_odb_entry(alt, sep, &entry);
444                 if (!entry.len)
445                         continue;
446                 link_alt_odb_entry(entry.buf, relative_base, depth, objdirbuf.buf);
447         }
448         strbuf_release(&entry);
449         strbuf_release(&objdirbuf);
450 }
451
452 static void read_info_alternates(const char * relative_base, int depth)
453 {
454         char *path;
455         struct strbuf buf = STRBUF_INIT;
456
457         path = xstrfmt("%s/info/alternates", relative_base);
458         if (strbuf_read_file(&buf, path, 1024) < 0) {
459                 warn_on_fopen_errors(path);
460                 free(path);
461                 return;
462         }
463
464         link_alt_odb_entries(buf.buf, '\n', relative_base, depth);
465         strbuf_release(&buf);
466         free(path);
467 }
468
469 struct alternate_object_database *alloc_alt_odb(const char *dir)
470 {
471         struct alternate_object_database *ent;
472
473         FLEX_ALLOC_STR(ent, path, dir);
474         strbuf_init(&ent->scratch, 0);
475         strbuf_addf(&ent->scratch, "%s/", dir);
476         ent->base_len = ent->scratch.len;
477
478         return ent;
479 }
480
481 void add_to_alternates_file(const char *reference)
482 {
483         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
484         char *alts = git_pathdup("objects/info/alternates");
485         FILE *in, *out;
486
487         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
488         out = fdopen_lock_file(lock, "w");
489         if (!out)
490                 die_errno("unable to fdopen alternates lockfile");
491
492         in = fopen(alts, "r");
493         if (in) {
494                 struct strbuf line = STRBUF_INIT;
495                 int found = 0;
496
497                 while (strbuf_getline(&line, in) != EOF) {
498                         if (!strcmp(reference, line.buf)) {
499                                 found = 1;
500                                 break;
501                         }
502                         fprintf_or_die(out, "%s\n", line.buf);
503                 }
504
505                 strbuf_release(&line);
506                 fclose(in);
507
508                 if (found) {
509                         rollback_lock_file(lock);
510                         lock = NULL;
511                 }
512         }
513         else if (errno != ENOENT)
514                 die_errno("unable to read alternates file");
515
516         if (lock) {
517                 fprintf_or_die(out, "%s\n", reference);
518                 if (commit_lock_file(lock))
519                         die_errno("unable to move new alternates file into place");
520                 if (alt_odb_tail)
521                         link_alt_odb_entries(reference, '\n', NULL, 0);
522         }
523         free(alts);
524 }
525
526 void add_to_alternates_memory(const char *reference)
527 {
528         /*
529          * Make sure alternates are initialized, or else our entry may be
530          * overwritten when they are.
531          */
532         prepare_alt_odb();
533
534         link_alt_odb_entries(reference, '\n', NULL, 0);
535 }
536
537 /*
538  * Compute the exact path an alternate is at and returns it. In case of
539  * error NULL is returned and the human readable error is added to `err`
540  * `path` may be relative and should point to $GITDIR.
541  * `err` must not be null.
542  */
543 char *compute_alternate_path(const char *path, struct strbuf *err)
544 {
545         char *ref_git = NULL;
546         const char *repo, *ref_git_s;
547         int seen_error = 0;
548
549         ref_git_s = real_path_if_valid(path);
550         if (!ref_git_s) {
551                 seen_error = 1;
552                 strbuf_addf(err, _("path '%s' does not exist"), path);
553                 goto out;
554         } else
555                 /*
556                  * Beware: read_gitfile(), real_path() and mkpath()
557                  * return static buffer
558                  */
559                 ref_git = xstrdup(ref_git_s);
560
561         repo = read_gitfile(ref_git);
562         if (!repo)
563                 repo = read_gitfile(mkpath("%s/.git", ref_git));
564         if (repo) {
565                 free(ref_git);
566                 ref_git = xstrdup(repo);
567         }
568
569         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
570                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
571                 free(ref_git);
572                 ref_git = ref_git_git;
573         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
574                 struct strbuf sb = STRBUF_INIT;
575                 seen_error = 1;
576                 if (get_common_dir(&sb, ref_git)) {
577                         strbuf_addf(err,
578                                     _("reference repository '%s' as a linked "
579                                       "checkout is not supported yet."),
580                                     path);
581                         goto out;
582                 }
583
584                 strbuf_addf(err, _("reference repository '%s' is not a "
585                                         "local repository."), path);
586                 goto out;
587         }
588
589         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
590                 strbuf_addf(err, _("reference repository '%s' is shallow"),
591                             path);
592                 seen_error = 1;
593                 goto out;
594         }
595
596         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
597                 strbuf_addf(err,
598                             _("reference repository '%s' is grafted"),
599                             path);
600                 seen_error = 1;
601                 goto out;
602         }
603
604 out:
605         if (seen_error) {
606                 FREE_AND_NULL(ref_git);
607         }
608
609         return ref_git;
610 }
611
612 int foreach_alt_odb(alt_odb_fn fn, void *cb)
613 {
614         struct alternate_object_database *ent;
615         int r = 0;
616
617         prepare_alt_odb();
618         for (ent = alt_odb_list; ent; ent = ent->next) {
619                 r = fn(ent, cb);
620                 if (r)
621                         break;
622         }
623         return r;
624 }
625
626 void prepare_alt_odb(void)
627 {
628         const char *alt;
629
630         if (alt_odb_tail)
631                 return;
632
633         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
634         if (!alt) alt = "";
635
636         alt_odb_tail = &alt_odb_list;
637         link_alt_odb_entries(alt, PATH_SEP, NULL, 0);
638
639         read_info_alternates(get_object_directory(), 0);
640 }
641
642 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
643 static int freshen_file(const char *fn)
644 {
645         struct utimbuf t;
646         t.actime = t.modtime = time(NULL);
647         return !utime(fn, &t);
648 }
649
650 /*
651  * All of the check_and_freshen functions return 1 if the file exists and was
652  * freshened (if freshening was requested), 0 otherwise. If they return
653  * 0, you should not assume that it is safe to skip a write of the object (it
654  * either does not exist on disk, or has a stale mtime and may be subject to
655  * pruning).
656  */
657 int check_and_freshen_file(const char *fn, int freshen)
658 {
659         if (access(fn, F_OK))
660                 return 0;
661         if (freshen && !freshen_file(fn))
662                 return 0;
663         return 1;
664 }
665
666 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
667 {
668         return check_and_freshen_file(sha1_file_name(sha1), freshen);
669 }
670
671 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
672 {
673         struct alternate_object_database *alt;
674         prepare_alt_odb();
675         for (alt = alt_odb_list; alt; alt = alt->next) {
676                 const char *path = alt_sha1_path(alt, sha1);
677                 if (check_and_freshen_file(path, freshen))
678                         return 1;
679         }
680         return 0;
681 }
682
683 static int check_and_freshen(const unsigned char *sha1, int freshen)
684 {
685         return check_and_freshen_local(sha1, freshen) ||
686                check_and_freshen_nonlocal(sha1, freshen);
687 }
688
689 int has_loose_object_nonlocal(const unsigned char *sha1)
690 {
691         return check_and_freshen_nonlocal(sha1, 0);
692 }
693
694 static int has_loose_object(const unsigned char *sha1)
695 {
696         return check_and_freshen(sha1, 0);
697 }
698
699 static unsigned int pack_used_ctr;
700 static unsigned int pack_mmap_calls;
701 static unsigned int peak_pack_open_windows;
702 static unsigned int pack_open_windows;
703 static unsigned int pack_open_fds;
704 static unsigned int pack_max_fds;
705 static size_t peak_pack_mapped;
706 static size_t pack_mapped;
707 struct packed_git *packed_git;
708
709 static struct mru packed_git_mru_storage;
710 struct mru *packed_git_mru = &packed_git_mru_storage;
711
712 void pack_report(void)
713 {
714         fprintf(stderr,
715                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
716                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
717                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
718                 sz_fmt(getpagesize()),
719                 sz_fmt(packed_git_window_size),
720                 sz_fmt(packed_git_limit));
721         fprintf(stderr,
722                 "pack_report: pack_used_ctr            = %10u\n"
723                 "pack_report: pack_mmap_calls          = %10u\n"
724                 "pack_report: pack_open_windows        = %10u / %10u\n"
725                 "pack_report: pack_mapped              = "
726                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
727                 pack_used_ctr,
728                 pack_mmap_calls,
729                 pack_open_windows, peak_pack_open_windows,
730                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
731 }
732
733 /*
734  * Open and mmap the index file at path, perform a couple of
735  * consistency checks, then record its information to p.  Return 0 on
736  * success.
737  */
738 static int check_packed_git_idx(const char *path, struct packed_git *p)
739 {
740         void *idx_map;
741         struct pack_idx_header *hdr;
742         size_t idx_size;
743         uint32_t version, nr, i, *index;
744         int fd = git_open(path);
745         struct stat st;
746
747         if (fd < 0)
748                 return -1;
749         if (fstat(fd, &st)) {
750                 close(fd);
751                 return -1;
752         }
753         idx_size = xsize_t(st.st_size);
754         if (idx_size < 4 * 256 + 20 + 20) {
755                 close(fd);
756                 return error("index file %s is too small", path);
757         }
758         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
759         close(fd);
760
761         hdr = idx_map;
762         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
763                 version = ntohl(hdr->idx_version);
764                 if (version < 2 || version > 2) {
765                         munmap(idx_map, idx_size);
766                         return error("index file %s is version %"PRIu32
767                                      " and is not supported by this binary"
768                                      " (try upgrading GIT to a newer version)",
769                                      path, version);
770                 }
771         } else
772                 version = 1;
773
774         nr = 0;
775         index = idx_map;
776         if (version > 1)
777                 index += 2;  /* skip index header */
778         for (i = 0; i < 256; i++) {
779                 uint32_t n = ntohl(index[i]);
780                 if (n < nr) {
781                         munmap(idx_map, idx_size);
782                         return error("non-monotonic index %s", path);
783                 }
784                 nr = n;
785         }
786
787         if (version == 1) {
788                 /*
789                  * Total size:
790                  *  - 256 index entries 4 bytes each
791                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
792                  *  - 20-byte SHA1 of the packfile
793                  *  - 20-byte SHA1 file checksum
794                  */
795                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
796                         munmap(idx_map, idx_size);
797                         return error("wrong index v1 file size in %s", path);
798                 }
799         } else if (version == 2) {
800                 /*
801                  * Minimum size:
802                  *  - 8 bytes of header
803                  *  - 256 index entries 4 bytes each
804                  *  - 20-byte sha1 entry * nr
805                  *  - 4-byte crc entry * nr
806                  *  - 4-byte offset entry * nr
807                  *  - 20-byte SHA1 of the packfile
808                  *  - 20-byte SHA1 file checksum
809                  * And after the 4-byte offset table might be a
810                  * variable sized table containing 8-byte entries
811                  * for offsets larger than 2^31.
812                  */
813                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
814                 unsigned long max_size = min_size;
815                 if (nr)
816                         max_size += (nr - 1)*8;
817                 if (idx_size < min_size || idx_size > max_size) {
818                         munmap(idx_map, idx_size);
819                         return error("wrong index v2 file size in %s", path);
820                 }
821                 if (idx_size != min_size &&
822                     /*
823                      * make sure we can deal with large pack offsets.
824                      * 31-bit signed offset won't be enough, neither
825                      * 32-bit unsigned one will be.
826                      */
827                     (sizeof(off_t) <= 4)) {
828                         munmap(idx_map, idx_size);
829                         return error("pack too large for current definition of off_t in %s", path);
830                 }
831         }
832
833         p->index_version = version;
834         p->index_data = idx_map;
835         p->index_size = idx_size;
836         p->num_objects = nr;
837         return 0;
838 }
839
840 int open_pack_index(struct packed_git *p)
841 {
842         char *idx_name;
843         size_t len;
844         int ret;
845
846         if (p->index_data)
847                 return 0;
848
849         if (!strip_suffix(p->pack_name, ".pack", &len))
850                 die("BUG: pack_name does not end in .pack");
851         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
852         ret = check_packed_git_idx(idx_name, p);
853         free(idx_name);
854         return ret;
855 }
856
857 static void scan_windows(struct packed_git *p,
858         struct packed_git **lru_p,
859         struct pack_window **lru_w,
860         struct pack_window **lru_l)
861 {
862         struct pack_window *w, *w_l;
863
864         for (w_l = NULL, w = p->windows; w; w = w->next) {
865                 if (!w->inuse_cnt) {
866                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
867                                 *lru_p = p;
868                                 *lru_w = w;
869                                 *lru_l = w_l;
870                         }
871                 }
872                 w_l = w;
873         }
874 }
875
876 static int unuse_one_window(struct packed_git *current)
877 {
878         struct packed_git *p, *lru_p = NULL;
879         struct pack_window *lru_w = NULL, *lru_l = NULL;
880
881         if (current)
882                 scan_windows(current, &lru_p, &lru_w, &lru_l);
883         for (p = packed_git; p; p = p->next)
884                 scan_windows(p, &lru_p, &lru_w, &lru_l);
885         if (lru_p) {
886                 munmap(lru_w->base, lru_w->len);
887                 pack_mapped -= lru_w->len;
888                 if (lru_l)
889                         lru_l->next = lru_w->next;
890                 else
891                         lru_p->windows = lru_w->next;
892                 free(lru_w);
893                 pack_open_windows--;
894                 return 1;
895         }
896         return 0;
897 }
898
899 void release_pack_memory(size_t need)
900 {
901         size_t cur = pack_mapped;
902         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
903                 ; /* nothing */
904 }
905
906 static void mmap_limit_check(size_t length)
907 {
908         static size_t limit = 0;
909         if (!limit) {
910                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
911                 if (!limit)
912                         limit = SIZE_MAX;
913         }
914         if (length > limit)
915                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
916                     (uintmax_t)length, (uintmax_t)limit);
917 }
918
919 void *xmmap_gently(void *start, size_t length,
920                   int prot, int flags, int fd, off_t offset)
921 {
922         void *ret;
923
924         mmap_limit_check(length);
925         ret = mmap(start, length, prot, flags, fd, offset);
926         if (ret == MAP_FAILED) {
927                 if (!length)
928                         return NULL;
929                 release_pack_memory(length);
930                 ret = mmap(start, length, prot, flags, fd, offset);
931         }
932         return ret;
933 }
934
935 void *xmmap(void *start, size_t length,
936         int prot, int flags, int fd, off_t offset)
937 {
938         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
939         if (ret == MAP_FAILED)
940                 die_errno("mmap failed");
941         return ret;
942 }
943
944 void close_pack_windows(struct packed_git *p)
945 {
946         while (p->windows) {
947                 struct pack_window *w = p->windows;
948
949                 if (w->inuse_cnt)
950                         die("pack '%s' still has open windows to it",
951                             p->pack_name);
952                 munmap(w->base, w->len);
953                 pack_mapped -= w->len;
954                 pack_open_windows--;
955                 p->windows = w->next;
956                 free(w);
957         }
958 }
959
960 static int close_pack_fd(struct packed_git *p)
961 {
962         if (p->pack_fd < 0)
963                 return 0;
964
965         close(p->pack_fd);
966         pack_open_fds--;
967         p->pack_fd = -1;
968
969         return 1;
970 }
971
972 static void close_pack(struct packed_git *p)
973 {
974         close_pack_windows(p);
975         close_pack_fd(p);
976         close_pack_index(p);
977 }
978
979 void close_all_packs(void)
980 {
981         struct packed_git *p;
982
983         for (p = packed_git; p; p = p->next)
984                 if (p->do_not_close)
985                         die("BUG: want to close pack marked 'do-not-close'");
986                 else
987                         close_pack(p);
988 }
989
990
991 /*
992  * The LRU pack is the one with the oldest MRU window, preferring packs
993  * with no used windows, or the oldest mtime if it has no windows allocated.
994  */
995 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
996 {
997         struct pack_window *w, *this_mru_w;
998         int has_windows_inuse = 0;
999
1000         /*
1001          * Reject this pack if it has windows and the previously selected
1002          * one does not.  If this pack does not have windows, reject
1003          * it if the pack file is newer than the previously selected one.
1004          */
1005         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
1006                 return;
1007
1008         for (w = this_mru_w = p->windows; w; w = w->next) {
1009                 /*
1010                  * Reject this pack if any of its windows are in use,
1011                  * but the previously selected pack did not have any
1012                  * inuse windows.  Otherwise, record that this pack
1013                  * has windows in use.
1014                  */
1015                 if (w->inuse_cnt) {
1016                         if (*accept_windows_inuse)
1017                                 has_windows_inuse = 1;
1018                         else
1019                                 return;
1020                 }
1021
1022                 if (w->last_used > this_mru_w->last_used)
1023                         this_mru_w = w;
1024
1025                 /*
1026                  * Reject this pack if it has windows that have been
1027                  * used more recently than the previously selected pack.
1028                  * If the previously selected pack had windows inuse and
1029                  * we have not encountered a window in this pack that is
1030                  * inuse, skip this check since we prefer a pack with no
1031                  * inuse windows to one that has inuse windows.
1032                  */
1033                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
1034                     this_mru_w->last_used > (*mru_w)->last_used)
1035                         return;
1036         }
1037
1038         /*
1039          * Select this pack.
1040          */
1041         *mru_w = this_mru_w;
1042         *lru_p = p;
1043         *accept_windows_inuse = has_windows_inuse;
1044 }
1045
1046 static int close_one_pack(void)
1047 {
1048         struct packed_git *p, *lru_p = NULL;
1049         struct pack_window *mru_w = NULL;
1050         int accept_windows_inuse = 1;
1051
1052         for (p = packed_git; p; p = p->next) {
1053                 if (p->pack_fd == -1)
1054                         continue;
1055                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
1056         }
1057
1058         if (lru_p)
1059                 return close_pack_fd(lru_p);
1060
1061         return 0;
1062 }
1063
1064 void unuse_pack(struct pack_window **w_cursor)
1065 {
1066         struct pack_window *w = *w_cursor;
1067         if (w) {
1068                 w->inuse_cnt--;
1069                 *w_cursor = NULL;
1070         }
1071 }
1072
1073 void close_pack_index(struct packed_git *p)
1074 {
1075         if (p->index_data) {
1076                 munmap((void *)p->index_data, p->index_size);
1077                 p->index_data = NULL;
1078         }
1079 }
1080
1081 static unsigned int get_max_fd_limit(void)
1082 {
1083 #ifdef RLIMIT_NOFILE
1084         {
1085                 struct rlimit lim;
1086
1087                 if (!getrlimit(RLIMIT_NOFILE, &lim))
1088                         return lim.rlim_cur;
1089         }
1090 #endif
1091
1092 #ifdef _SC_OPEN_MAX
1093         {
1094                 long open_max = sysconf(_SC_OPEN_MAX);
1095                 if (0 < open_max)
1096                         return open_max;
1097                 /*
1098                  * Otherwise, we got -1 for one of the two
1099                  * reasons:
1100                  *
1101                  * (1) sysconf() did not understand _SC_OPEN_MAX
1102                  *     and signaled an error with -1; or
1103                  * (2) sysconf() said there is no limit.
1104                  *
1105                  * We _could_ clear errno before calling sysconf() to
1106                  * tell these two cases apart and return a huge number
1107                  * in the latter case to let the caller cap it to a
1108                  * value that is not so selfish, but letting the
1109                  * fallback OPEN_MAX codepath take care of these cases
1110                  * is a lot simpler.
1111                  */
1112         }
1113 #endif
1114
1115 #ifdef OPEN_MAX
1116         return OPEN_MAX;
1117 #else
1118         return 1; /* see the caller ;-) */
1119 #endif
1120 }
1121
1122 /*
1123  * Do not call this directly as this leaks p->pack_fd on error return;
1124  * call open_packed_git() instead.
1125  */
1126 static int open_packed_git_1(struct packed_git *p)
1127 {
1128         struct stat st;
1129         struct pack_header hdr;
1130         unsigned char sha1[20];
1131         unsigned char *idx_sha1;
1132         long fd_flag;
1133
1134         if (!p->index_data && open_pack_index(p))
1135                 return error("packfile %s index unavailable", p->pack_name);
1136
1137         if (!pack_max_fds) {
1138                 unsigned int max_fds = get_max_fd_limit();
1139
1140                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1141                 if (25 < max_fds)
1142                         pack_max_fds = max_fds - 25;
1143                 else
1144                         pack_max_fds = 1;
1145         }
1146
1147         while (pack_max_fds <= pack_open_fds && close_one_pack())
1148                 ; /* nothing */
1149
1150         p->pack_fd = git_open(p->pack_name);
1151         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1152                 return -1;
1153         pack_open_fds++;
1154
1155         /* If we created the struct before we had the pack we lack size. */
1156         if (!p->pack_size) {
1157                 if (!S_ISREG(st.st_mode))
1158                         return error("packfile %s not a regular file", p->pack_name);
1159                 p->pack_size = st.st_size;
1160         } else if (p->pack_size != st.st_size)
1161                 return error("packfile %s size changed", p->pack_name);
1162
1163         /* We leave these file descriptors open with sliding mmap;
1164          * there is no point keeping them open across exec(), though.
1165          */
1166         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1167         if (fd_flag < 0)
1168                 return error("cannot determine file descriptor flags");
1169         fd_flag |= FD_CLOEXEC;
1170         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1171                 return error("cannot set FD_CLOEXEC");
1172
1173         /* Verify we recognize this pack file format. */
1174         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1175                 return error("file %s is far too short to be a packfile", p->pack_name);
1176         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1177                 return error("file %s is not a GIT packfile", p->pack_name);
1178         if (!pack_version_ok(hdr.hdr_version))
1179                 return error("packfile %s is version %"PRIu32" and not"
1180                         " supported (try upgrading GIT to a newer version)",
1181                         p->pack_name, ntohl(hdr.hdr_version));
1182
1183         /* Verify the pack matches its index. */
1184         if (p->num_objects != ntohl(hdr.hdr_entries))
1185                 return error("packfile %s claims to have %"PRIu32" objects"
1186                              " while index indicates %"PRIu32" objects",
1187                              p->pack_name, ntohl(hdr.hdr_entries),
1188                              p->num_objects);
1189         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1190                 return error("end of packfile %s is unavailable", p->pack_name);
1191         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1192                 return error("packfile %s signature is unavailable", p->pack_name);
1193         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1194         if (hashcmp(sha1, idx_sha1))
1195                 return error("packfile %s does not match index", p->pack_name);
1196         return 0;
1197 }
1198
1199 static int open_packed_git(struct packed_git *p)
1200 {
1201         if (!open_packed_git_1(p))
1202                 return 0;
1203         close_pack_fd(p);
1204         return -1;
1205 }
1206
1207 static int in_window(struct pack_window *win, off_t offset)
1208 {
1209         /* We must promise at least 20 bytes (one hash) after the
1210          * offset is available from this window, otherwise the offset
1211          * is not actually in this window and a different window (which
1212          * has that one hash excess) must be used.  This is to support
1213          * the object header and delta base parsing routines below.
1214          */
1215         off_t win_off = win->offset;
1216         return win_off <= offset
1217                 && (offset + 20) <= (win_off + win->len);
1218 }
1219
1220 unsigned char *use_pack(struct packed_git *p,
1221                 struct pack_window **w_cursor,
1222                 off_t offset,
1223                 unsigned long *left)
1224 {
1225         struct pack_window *win = *w_cursor;
1226
1227         /* Since packfiles end in a hash of their content and it's
1228          * pointless to ask for an offset into the middle of that
1229          * hash, and the in_window function above wouldn't match
1230          * don't allow an offset too close to the end of the file.
1231          */
1232         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1233                 die("packfile %s cannot be accessed", p->pack_name);
1234         if (offset > (p->pack_size - 20))
1235                 die("offset beyond end of packfile (truncated pack?)");
1236         if (offset < 0)
1237                 die(_("offset before end of packfile (broken .idx?)"));
1238
1239         if (!win || !in_window(win, offset)) {
1240                 if (win)
1241                         win->inuse_cnt--;
1242                 for (win = p->windows; win; win = win->next) {
1243                         if (in_window(win, offset))
1244                                 break;
1245                 }
1246                 if (!win) {
1247                         size_t window_align = packed_git_window_size / 2;
1248                         off_t len;
1249
1250                         if (p->pack_fd == -1 && open_packed_git(p))
1251                                 die("packfile %s cannot be accessed", p->pack_name);
1252
1253                         win = xcalloc(1, sizeof(*win));
1254                         win->offset = (offset / window_align) * window_align;
1255                         len = p->pack_size - win->offset;
1256                         if (len > packed_git_window_size)
1257                                 len = packed_git_window_size;
1258                         win->len = (size_t)len;
1259                         pack_mapped += win->len;
1260                         while (packed_git_limit < pack_mapped
1261                                 && unuse_one_window(p))
1262                                 ; /* nothing */
1263                         win->base = xmmap(NULL, win->len,
1264                                 PROT_READ, MAP_PRIVATE,
1265                                 p->pack_fd, win->offset);
1266                         if (win->base == MAP_FAILED)
1267                                 die_errno("packfile %s cannot be mapped",
1268                                           p->pack_name);
1269                         if (!win->offset && win->len == p->pack_size
1270                                 && !p->do_not_close)
1271                                 close_pack_fd(p);
1272                         pack_mmap_calls++;
1273                         pack_open_windows++;
1274                         if (pack_mapped > peak_pack_mapped)
1275                                 peak_pack_mapped = pack_mapped;
1276                         if (pack_open_windows > peak_pack_open_windows)
1277                                 peak_pack_open_windows = pack_open_windows;
1278                         win->next = p->windows;
1279                         p->windows = win;
1280                 }
1281         }
1282         if (win != *w_cursor) {
1283                 win->last_used = pack_used_ctr++;
1284                 win->inuse_cnt++;
1285                 *w_cursor = win;
1286         }
1287         offset -= win->offset;
1288         if (left)
1289                 *left = win->len - xsize_t(offset);
1290         return win->base + offset;
1291 }
1292
1293 static struct packed_git *alloc_packed_git(int extra)
1294 {
1295         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1296         memset(p, 0, sizeof(*p));
1297         p->pack_fd = -1;
1298         return p;
1299 }
1300
1301 static void try_to_free_pack_memory(size_t size)
1302 {
1303         release_pack_memory(size);
1304 }
1305
1306 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1307 {
1308         static int have_set_try_to_free_routine;
1309         struct stat st;
1310         size_t alloc;
1311         struct packed_git *p;
1312
1313         if (!have_set_try_to_free_routine) {
1314                 have_set_try_to_free_routine = 1;
1315                 set_try_to_free_routine(try_to_free_pack_memory);
1316         }
1317
1318         /*
1319          * Make sure a corresponding .pack file exists and that
1320          * the index looks sane.
1321          */
1322         if (!strip_suffix_mem(path, &path_len, ".idx"))
1323                 return NULL;
1324
1325         /*
1326          * ".pack" is long enough to hold any suffix we're adding (and
1327          * the use xsnprintf double-checks that)
1328          */
1329         alloc = st_add3(path_len, strlen(".pack"), 1);
1330         p = alloc_packed_git(alloc);
1331         memcpy(p->pack_name, path, path_len);
1332
1333         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1334         if (!access(p->pack_name, F_OK))
1335                 p->pack_keep = 1;
1336
1337         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1338         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1339                 free(p);
1340                 return NULL;
1341         }
1342
1343         /* ok, it looks sane as far as we can check without
1344          * actually mapping the pack file.
1345          */
1346         p->pack_size = st.st_size;
1347         p->pack_local = local;
1348         p->mtime = st.st_mtime;
1349         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1350                 hashclr(p->sha1);
1351         return p;
1352 }
1353
1354 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1355 {
1356         const char *path = sha1_pack_name(sha1);
1357         size_t alloc = st_add(strlen(path), 1);
1358         struct packed_git *p = alloc_packed_git(alloc);
1359
1360         memcpy(p->pack_name, path, alloc); /* includes NUL */
1361         hashcpy(p->sha1, sha1);
1362         if (check_packed_git_idx(idx_path, p)) {
1363                 free(p);
1364                 return NULL;
1365         }
1366
1367         return p;
1368 }
1369
1370 void install_packed_git(struct packed_git *pack)
1371 {
1372         if (pack->pack_fd != -1)
1373                 pack_open_fds++;
1374
1375         pack->next = packed_git;
1376         packed_git = pack;
1377 }
1378
1379 void (*report_garbage)(unsigned seen_bits, const char *path);
1380
1381 static void report_helper(const struct string_list *list,
1382                           int seen_bits, int first, int last)
1383 {
1384         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1385                 return;
1386
1387         for (; first < last; first++)
1388                 report_garbage(seen_bits, list->items[first].string);
1389 }
1390
1391 static void report_pack_garbage(struct string_list *list)
1392 {
1393         int i, baselen = -1, first = 0, seen_bits = 0;
1394
1395         if (!report_garbage)
1396                 return;
1397
1398         string_list_sort(list);
1399
1400         for (i = 0; i < list->nr; i++) {
1401                 const char *path = list->items[i].string;
1402                 if (baselen != -1 &&
1403                     strncmp(path, list->items[first].string, baselen)) {
1404                         report_helper(list, seen_bits, first, i);
1405                         baselen = -1;
1406                         seen_bits = 0;
1407                 }
1408                 if (baselen == -1) {
1409                         const char *dot = strrchr(path, '.');
1410                         if (!dot) {
1411                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1412                                 continue;
1413                         }
1414                         baselen = dot - path + 1;
1415                         first = i;
1416                 }
1417                 if (!strcmp(path + baselen, "pack"))
1418                         seen_bits |= 1;
1419                 else if (!strcmp(path + baselen, "idx"))
1420                         seen_bits |= 2;
1421         }
1422         report_helper(list, seen_bits, first, list->nr);
1423 }
1424
1425 static void prepare_packed_git_one(char *objdir, int local)
1426 {
1427         struct strbuf path = STRBUF_INIT;
1428         size_t dirnamelen;
1429         DIR *dir;
1430         struct dirent *de;
1431         struct string_list garbage = STRING_LIST_INIT_DUP;
1432
1433         strbuf_addstr(&path, objdir);
1434         strbuf_addstr(&path, "/pack");
1435         dir = opendir(path.buf);
1436         if (!dir) {
1437                 if (errno != ENOENT)
1438                         error_errno("unable to open object pack directory: %s",
1439                                     path.buf);
1440                 strbuf_release(&path);
1441                 return;
1442         }
1443         strbuf_addch(&path, '/');
1444         dirnamelen = path.len;
1445         while ((de = readdir(dir)) != NULL) {
1446                 struct packed_git *p;
1447                 size_t base_len;
1448
1449                 if (is_dot_or_dotdot(de->d_name))
1450                         continue;
1451
1452                 strbuf_setlen(&path, dirnamelen);
1453                 strbuf_addstr(&path, de->d_name);
1454
1455                 base_len = path.len;
1456                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1457                         /* Don't reopen a pack we already have. */
1458                         for (p = packed_git; p; p = p->next) {
1459                                 size_t len;
1460                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1461                                     len == base_len &&
1462                                     !memcmp(p->pack_name, path.buf, len))
1463                                         break;
1464                         }
1465                         if (p == NULL &&
1466                             /*
1467                              * See if it really is a valid .idx file with
1468                              * corresponding .pack file that we can map.
1469                              */
1470                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1471                                 install_packed_git(p);
1472                 }
1473
1474                 if (!report_garbage)
1475                         continue;
1476
1477                 if (ends_with(de->d_name, ".idx") ||
1478                     ends_with(de->d_name, ".pack") ||
1479                     ends_with(de->d_name, ".bitmap") ||
1480                     ends_with(de->d_name, ".keep"))
1481                         string_list_append(&garbage, path.buf);
1482                 else
1483                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1484         }
1485         closedir(dir);
1486         report_pack_garbage(&garbage);
1487         string_list_clear(&garbage, 0);
1488         strbuf_release(&path);
1489 }
1490
1491 static int approximate_object_count_valid;
1492
1493 /*
1494  * Give a fast, rough count of the number of objects in the repository. This
1495  * ignores loose objects completely. If you have a lot of them, then either
1496  * you should repack because your performance will be awful, or they are
1497  * all unreachable objects about to be pruned, in which case they're not really
1498  * interesting as a measure of repo size in the first place.
1499  */
1500 unsigned long approximate_object_count(void)
1501 {
1502         static unsigned long count;
1503         if (!approximate_object_count_valid) {
1504                 struct packed_git *p;
1505
1506                 prepare_packed_git();
1507                 count = 0;
1508                 for (p = packed_git; p; p = p->next) {
1509                         if (open_pack_index(p))
1510                                 continue;
1511                         count += p->num_objects;
1512                 }
1513         }
1514         return count;
1515 }
1516
1517 static void *get_next_packed_git(const void *p)
1518 {
1519         return ((const struct packed_git *)p)->next;
1520 }
1521
1522 static void set_next_packed_git(void *p, void *next)
1523 {
1524         ((struct packed_git *)p)->next = next;
1525 }
1526
1527 static int sort_pack(const void *a_, const void *b_)
1528 {
1529         const struct packed_git *a = a_;
1530         const struct packed_git *b = b_;
1531         int st;
1532
1533         /*
1534          * Local packs tend to contain objects specific to our
1535          * variant of the project than remote ones.  In addition,
1536          * remote ones could be on a network mounted filesystem.
1537          * Favor local ones for these reasons.
1538          */
1539         st = a->pack_local - b->pack_local;
1540         if (st)
1541                 return -st;
1542
1543         /*
1544          * Younger packs tend to contain more recent objects,
1545          * and more recent objects tend to get accessed more
1546          * often.
1547          */
1548         if (a->mtime < b->mtime)
1549                 return 1;
1550         else if (a->mtime == b->mtime)
1551                 return 0;
1552         return -1;
1553 }
1554
1555 static void rearrange_packed_git(void)
1556 {
1557         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1558                                      set_next_packed_git, sort_pack);
1559 }
1560
1561 static void prepare_packed_git_mru(void)
1562 {
1563         struct packed_git *p;
1564
1565         mru_clear(packed_git_mru);
1566         for (p = packed_git; p; p = p->next)
1567                 mru_append(packed_git_mru, p);
1568 }
1569
1570 static int prepare_packed_git_run_once = 0;
1571 void prepare_packed_git(void)
1572 {
1573         struct alternate_object_database *alt;
1574
1575         if (prepare_packed_git_run_once)
1576                 return;
1577         prepare_packed_git_one(get_object_directory(), 1);
1578         prepare_alt_odb();
1579         for (alt = alt_odb_list; alt; alt = alt->next)
1580                 prepare_packed_git_one(alt->path, 0);
1581         rearrange_packed_git();
1582         prepare_packed_git_mru();
1583         prepare_packed_git_run_once = 1;
1584 }
1585
1586 void reprepare_packed_git(void)
1587 {
1588         approximate_object_count_valid = 0;
1589         prepare_packed_git_run_once = 0;
1590         prepare_packed_git();
1591 }
1592
1593 static void mark_bad_packed_object(struct packed_git *p,
1594                                    const unsigned char *sha1)
1595 {
1596         unsigned i;
1597         for (i = 0; i < p->num_bad_objects; i++)
1598                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1599                         return;
1600         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1601                                       st_mult(GIT_MAX_RAWSZ,
1602                                               st_add(p->num_bad_objects, 1)));
1603         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1604         p->num_bad_objects++;
1605 }
1606
1607 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1608 {
1609         struct packed_git *p;
1610         unsigned i;
1611
1612         for (p = packed_git; p; p = p->next)
1613                 for (i = 0; i < p->num_bad_objects; i++)
1614                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1615                                 return p;
1616         return NULL;
1617 }
1618
1619 /*
1620  * With an in-core object data in "map", rehash it to make sure the
1621  * object name actually matches "sha1" to detect object corruption.
1622  * With "map" == NULL, try reading the object named with "sha1" using
1623  * the streaming interface and rehash it to do the same.
1624  */
1625 int check_sha1_signature(const unsigned char *sha1, void *map,
1626                          unsigned long size, const char *type)
1627 {
1628         unsigned char real_sha1[20];
1629         enum object_type obj_type;
1630         struct git_istream *st;
1631         git_SHA_CTX c;
1632         char hdr[32];
1633         int hdrlen;
1634
1635         if (map) {
1636                 hash_sha1_file(map, size, type, real_sha1);
1637                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1638         }
1639
1640         st = open_istream(sha1, &obj_type, &size, NULL);
1641         if (!st)
1642                 return -1;
1643
1644         /* Generate the header */
1645         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1646
1647         /* Sha1.. */
1648         git_SHA1_Init(&c);
1649         git_SHA1_Update(&c, hdr, hdrlen);
1650         for (;;) {
1651                 char buf[1024 * 16];
1652                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1653
1654                 if (readlen < 0) {
1655                         close_istream(st);
1656                         return -1;
1657                 }
1658                 if (!readlen)
1659                         break;
1660                 git_SHA1_Update(&c, buf, readlen);
1661         }
1662         git_SHA1_Final(real_sha1, &c);
1663         close_istream(st);
1664         return hashcmp(sha1, real_sha1) ? -1 : 0;
1665 }
1666
1667 int git_open_cloexec(const char *name, int flags)
1668 {
1669         int fd;
1670         static int o_cloexec = O_CLOEXEC;
1671
1672         fd = open(name, flags | o_cloexec);
1673         if ((o_cloexec & O_CLOEXEC) && fd < 0 && errno == EINVAL) {
1674                 /* Try again w/o O_CLOEXEC: the kernel might not support it */
1675                 o_cloexec &= ~O_CLOEXEC;
1676                 fd = open(name, flags | o_cloexec);
1677         }
1678
1679 #if defined(F_GETFD) && defined(F_SETFD) && defined(FD_CLOEXEC)
1680         {
1681                 static int fd_cloexec = FD_CLOEXEC;
1682
1683                 if (!o_cloexec && 0 <= fd && fd_cloexec) {
1684                         /* Opened w/o O_CLOEXEC?  try with fcntl(2) to add it */
1685                         int flags = fcntl(fd, F_GETFD);
1686                         if (fcntl(fd, F_SETFD, flags | fd_cloexec))
1687                                 fd_cloexec = 0;
1688                 }
1689         }
1690 #endif
1691         return fd;
1692 }
1693
1694 /*
1695  * Find "sha1" as a loose object in the local repository or in an alternate.
1696  * Returns 0 on success, negative on failure.
1697  *
1698  * The "path" out-parameter will give the path of the object we found (if any).
1699  * Note that it may point to static storage and is only valid until another
1700  * call to sha1_file_name(), etc.
1701  */
1702 static int stat_sha1_file(const unsigned char *sha1, struct stat *st,
1703                           const char **path)
1704 {
1705         struct alternate_object_database *alt;
1706
1707         *path = sha1_file_name(sha1);
1708         if (!lstat(*path, st))
1709                 return 0;
1710
1711         prepare_alt_odb();
1712         errno = ENOENT;
1713         for (alt = alt_odb_list; alt; alt = alt->next) {
1714                 *path = alt_sha1_path(alt, sha1);
1715                 if (!lstat(*path, st))
1716                         return 0;
1717         }
1718
1719         return -1;
1720 }
1721
1722 /*
1723  * Like stat_sha1_file(), but actually open the object and return the
1724  * descriptor. See the caveats on the "path" parameter above.
1725  */
1726 static int open_sha1_file(const unsigned char *sha1, const char **path)
1727 {
1728         int fd;
1729         struct alternate_object_database *alt;
1730         int most_interesting_errno;
1731
1732         *path = sha1_file_name(sha1);
1733         fd = git_open(*path);
1734         if (fd >= 0)
1735                 return fd;
1736         most_interesting_errno = errno;
1737
1738         prepare_alt_odb();
1739         for (alt = alt_odb_list; alt; alt = alt->next) {
1740                 *path = alt_sha1_path(alt, sha1);
1741                 fd = git_open(*path);
1742                 if (fd >= 0)
1743                         return fd;
1744                 if (most_interesting_errno == ENOENT)
1745                         most_interesting_errno = errno;
1746         }
1747         errno = most_interesting_errno;
1748         return -1;
1749 }
1750
1751 /*
1752  * Map the loose object at "path" if it is not NULL, or the path found by
1753  * searching for a loose object named "sha1".
1754  */
1755 static void *map_sha1_file_1(const char *path,
1756                              const unsigned char *sha1,
1757                              unsigned long *size)
1758 {
1759         void *map;
1760         int fd;
1761
1762         if (path)
1763                 fd = git_open(path);
1764         else
1765                 fd = open_sha1_file(sha1, &path);
1766         map = NULL;
1767         if (fd >= 0) {
1768                 struct stat st;
1769
1770                 if (!fstat(fd, &st)) {
1771                         *size = xsize_t(st.st_size);
1772                         if (!*size) {
1773                                 /* mmap() is forbidden on empty files */
1774                                 error("object file %s is empty", path);
1775                                 return NULL;
1776                         }
1777                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1778                 }
1779                 close(fd);
1780         }
1781         return map;
1782 }
1783
1784 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1785 {
1786         return map_sha1_file_1(NULL, sha1, size);
1787 }
1788
1789 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1790                 unsigned long len, enum object_type *type, unsigned long *sizep)
1791 {
1792         unsigned shift;
1793         unsigned long size, c;
1794         unsigned long used = 0;
1795
1796         c = buf[used++];
1797         *type = (c >> 4) & 7;
1798         size = c & 15;
1799         shift = 4;
1800         while (c & 0x80) {
1801                 if (len <= used || bitsizeof(long) <= shift) {
1802                         error("bad object header");
1803                         size = used = 0;
1804                         break;
1805                 }
1806                 c = buf[used++];
1807                 size += (c & 0x7f) << shift;
1808                 shift += 7;
1809         }
1810         *sizep = size;
1811         return used;
1812 }
1813
1814 static int unpack_sha1_short_header(git_zstream *stream,
1815                                     unsigned char *map, unsigned long mapsize,
1816                                     void *buffer, unsigned long bufsiz)
1817 {
1818         /* Get the data stream */
1819         memset(stream, 0, sizeof(*stream));
1820         stream->next_in = map;
1821         stream->avail_in = mapsize;
1822         stream->next_out = buffer;
1823         stream->avail_out = bufsiz;
1824
1825         git_inflate_init(stream);
1826         return git_inflate(stream, 0);
1827 }
1828
1829 int unpack_sha1_header(git_zstream *stream,
1830                        unsigned char *map, unsigned long mapsize,
1831                        void *buffer, unsigned long bufsiz)
1832 {
1833         int status = unpack_sha1_short_header(stream, map, mapsize,
1834                                               buffer, bufsiz);
1835
1836         if (status < Z_OK)
1837                 return status;
1838
1839         /* Make sure we have the terminating NUL */
1840         if (!memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1841                 return -1;
1842         return 0;
1843 }
1844
1845 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1846                                         unsigned long mapsize, void *buffer,
1847                                         unsigned long bufsiz, struct strbuf *header)
1848 {
1849         int status;
1850
1851         status = unpack_sha1_short_header(stream, map, mapsize, buffer, bufsiz);
1852         if (status < Z_OK)
1853                 return -1;
1854
1855         /*
1856          * Check if entire header is unpacked in the first iteration.
1857          */
1858         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1859                 return 0;
1860
1861         /*
1862          * buffer[0..bufsiz] was not large enough.  Copy the partial
1863          * result out to header, and then append the result of further
1864          * reading the stream.
1865          */
1866         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1867         stream->next_out = buffer;
1868         stream->avail_out = bufsiz;
1869
1870         do {
1871                 status = git_inflate(stream, 0);
1872                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1873                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1874                         return 0;
1875                 stream->next_out = buffer;
1876                 stream->avail_out = bufsiz;
1877         } while (status != Z_STREAM_END);
1878         return -1;
1879 }
1880
1881 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1882 {
1883         int bytes = strlen(buffer) + 1;
1884         unsigned char *buf = xmallocz(size);
1885         unsigned long n;
1886         int status = Z_OK;
1887
1888         n = stream->total_out - bytes;
1889         if (n > size)
1890                 n = size;
1891         memcpy(buf, (char *) buffer + bytes, n);
1892         bytes = n;
1893         if (bytes <= size) {
1894                 /*
1895                  * The above condition must be (bytes <= size), not
1896                  * (bytes < size).  In other words, even though we
1897                  * expect no more output and set avail_out to zero,
1898                  * the input zlib stream may have bytes that express
1899                  * "this concludes the stream", and we *do* want to
1900                  * eat that input.
1901                  *
1902                  * Otherwise we would not be able to test that we
1903                  * consumed all the input to reach the expected size;
1904                  * we also want to check that zlib tells us that all
1905                  * went well with status == Z_STREAM_END at the end.
1906                  */
1907                 stream->next_out = buf + bytes;
1908                 stream->avail_out = size - bytes;
1909                 while (status == Z_OK)
1910                         status = git_inflate(stream, Z_FINISH);
1911         }
1912         if (status == Z_STREAM_END && !stream->avail_in) {
1913                 git_inflate_end(stream);
1914                 return buf;
1915         }
1916
1917         if (status < 0)
1918                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1919         else if (stream->avail_in)
1920                 error("garbage at end of loose object '%s'",
1921                       sha1_to_hex(sha1));
1922         free(buf);
1923         return NULL;
1924 }
1925
1926 /*
1927  * We used to just use "sscanf()", but that's actually way
1928  * too permissive for what we want to check. So do an anal
1929  * object header parse by hand.
1930  */
1931 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1932                                unsigned int flags)
1933 {
1934         const char *type_buf = hdr;
1935         unsigned long size;
1936         int type, type_len = 0;
1937
1938         /*
1939          * The type can be of any size but is followed by
1940          * a space.
1941          */
1942         for (;;) {
1943                 char c = *hdr++;
1944                 if (!c)
1945                         return -1;
1946                 if (c == ' ')
1947                         break;
1948                 type_len++;
1949         }
1950
1951         type = type_from_string_gently(type_buf, type_len, 1);
1952         if (oi->typename)
1953                 strbuf_add(oi->typename, type_buf, type_len);
1954         /*
1955          * Set type to 0 if its an unknown object and
1956          * we're obtaining the type using '--allow-unknown-type'
1957          * option.
1958          */
1959         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE) && (type < 0))
1960                 type = 0;
1961         else if (type < 0)
1962                 die("invalid object type");
1963         if (oi->typep)
1964                 *oi->typep = type;
1965
1966         /*
1967          * The length must follow immediately, and be in canonical
1968          * decimal format (ie "010" is not valid).
1969          */
1970         size = *hdr++ - '0';
1971         if (size > 9)
1972                 return -1;
1973         if (size) {
1974                 for (;;) {
1975                         unsigned long c = *hdr - '0';
1976                         if (c > 9)
1977                                 break;
1978                         hdr++;
1979                         size = size * 10 + c;
1980                 }
1981         }
1982
1983         if (oi->sizep)
1984                 *oi->sizep = size;
1985
1986         /*
1987          * The length must be followed by a zero byte
1988          */
1989         return *hdr ? -1 : type;
1990 }
1991
1992 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1993 {
1994         struct object_info oi = OBJECT_INFO_INIT;
1995
1996         oi.sizep = sizep;
1997         return parse_sha1_header_extended(hdr, &oi, 0);
1998 }
1999
2000 unsigned long get_size_from_delta(struct packed_git *p,
2001                                   struct pack_window **w_curs,
2002                                   off_t curpos)
2003 {
2004         const unsigned char *data;
2005         unsigned char delta_head[20], *in;
2006         git_zstream stream;
2007         int st;
2008
2009         memset(&stream, 0, sizeof(stream));
2010         stream.next_out = delta_head;
2011         stream.avail_out = sizeof(delta_head);
2012
2013         git_inflate_init(&stream);
2014         do {
2015                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2016                 stream.next_in = in;
2017                 st = git_inflate(&stream, Z_FINISH);
2018                 curpos += stream.next_in - in;
2019         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
2020                  stream.total_out < sizeof(delta_head));
2021         git_inflate_end(&stream);
2022         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
2023                 error("delta data unpack-initial failed");
2024                 return 0;
2025         }
2026
2027         /* Examine the initial part of the delta to figure out
2028          * the result size.
2029          */
2030         data = delta_head;
2031
2032         /* ignore base size */
2033         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2034
2035         /* Read the result size */
2036         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
2037 }
2038
2039 static off_t get_delta_base(struct packed_git *p,
2040                                     struct pack_window **w_curs,
2041                                     off_t *curpos,
2042                                     enum object_type type,
2043                                     off_t delta_obj_offset)
2044 {
2045         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
2046         off_t base_offset;
2047
2048         /* use_pack() assured us we have [base_info, base_info + 20)
2049          * as a range that we can look at without walking off the
2050          * end of the mapped window.  Its actually the hash size
2051          * that is assured.  An OFS_DELTA longer than the hash size
2052          * is stupid, as then a REF_DELTA would be smaller to store.
2053          */
2054         if (type == OBJ_OFS_DELTA) {
2055                 unsigned used = 0;
2056                 unsigned char c = base_info[used++];
2057                 base_offset = c & 127;
2058                 while (c & 128) {
2059                         base_offset += 1;
2060                         if (!base_offset || MSB(base_offset, 7))
2061                                 return 0;  /* overflow */
2062                         c = base_info[used++];
2063                         base_offset = (base_offset << 7) + (c & 127);
2064                 }
2065                 base_offset = delta_obj_offset - base_offset;
2066                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
2067                         return 0;  /* out of bound */
2068                 *curpos += used;
2069         } else if (type == OBJ_REF_DELTA) {
2070                 /* The base entry _must_ be in the same pack */
2071                 base_offset = find_pack_entry_one(base_info, p);
2072                 *curpos += 20;
2073         } else
2074                 die("I am totally screwed");
2075         return base_offset;
2076 }
2077
2078 /*
2079  * Like get_delta_base above, but we return the sha1 instead of the pack
2080  * offset. This means it is cheaper for REF deltas (we do not have to do
2081  * the final object lookup), but more expensive for OFS deltas (we
2082  * have to load the revidx to convert the offset back into a sha1).
2083  */
2084 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
2085                                                 struct pack_window **w_curs,
2086                                                 off_t curpos,
2087                                                 enum object_type type,
2088                                                 off_t delta_obj_offset)
2089 {
2090         if (type == OBJ_REF_DELTA) {
2091                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
2092                 return base;
2093         } else if (type == OBJ_OFS_DELTA) {
2094                 struct revindex_entry *revidx;
2095                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
2096                                                    type, delta_obj_offset);
2097
2098                 if (!base_offset)
2099                         return NULL;
2100
2101                 revidx = find_pack_revindex(p, base_offset);
2102                 if (!revidx)
2103                         return NULL;
2104
2105                 return nth_packed_object_sha1(p, revidx->nr);
2106         } else
2107                 return NULL;
2108 }
2109
2110 int unpack_object_header(struct packed_git *p,
2111                          struct pack_window **w_curs,
2112                          off_t *curpos,
2113                          unsigned long *sizep)
2114 {
2115         unsigned char *base;
2116         unsigned long left;
2117         unsigned long used;
2118         enum object_type type;
2119
2120         /* use_pack() assures us we have [base, base + 20) available
2121          * as a range that we can look at.  (Its actually the hash
2122          * size that is assured.)  With our object header encoding
2123          * the maximum deflated object size is 2^137, which is just
2124          * insane, so we know won't exceed what we have been given.
2125          */
2126         base = use_pack(p, w_curs, *curpos, &left);
2127         used = unpack_object_header_buffer(base, left, &type, sizep);
2128         if (!used) {
2129                 type = OBJ_BAD;
2130         } else
2131                 *curpos += used;
2132
2133         return type;
2134 }
2135
2136 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
2137 {
2138         int type;
2139         struct revindex_entry *revidx;
2140         const unsigned char *sha1;
2141         revidx = find_pack_revindex(p, obj_offset);
2142         if (!revidx)
2143                 return OBJ_BAD;
2144         sha1 = nth_packed_object_sha1(p, revidx->nr);
2145         mark_bad_packed_object(p, sha1);
2146         type = sha1_object_info(sha1, NULL);
2147         if (type <= OBJ_NONE)
2148                 return OBJ_BAD;
2149         return type;
2150 }
2151
2152 #define POI_STACK_PREALLOC 64
2153
2154 static enum object_type packed_to_object_type(struct packed_git *p,
2155                                               off_t obj_offset,
2156                                               enum object_type type,
2157                                               struct pack_window **w_curs,
2158                                               off_t curpos)
2159 {
2160         off_t small_poi_stack[POI_STACK_PREALLOC];
2161         off_t *poi_stack = small_poi_stack;
2162         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
2163
2164         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2165                 off_t base_offset;
2166                 unsigned long size;
2167                 /* Push the object we're going to leave behind */
2168                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
2169                         poi_stack_alloc = alloc_nr(poi_stack_nr);
2170                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
2171                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2172                 } else {
2173                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2174                 }
2175                 poi_stack[poi_stack_nr++] = obj_offset;
2176                 /* If parsing the base offset fails, just unwind */
2177                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2178                 if (!base_offset)
2179                         goto unwind;
2180                 curpos = obj_offset = base_offset;
2181                 type = unpack_object_header(p, w_curs, &curpos, &size);
2182                 if (type <= OBJ_NONE) {
2183                         /* If getting the base itself fails, we first
2184                          * retry the base, otherwise unwind */
2185                         type = retry_bad_packed_offset(p, base_offset);
2186                         if (type > OBJ_NONE)
2187                                 goto out;
2188                         goto unwind;
2189                 }
2190         }
2191
2192         switch (type) {
2193         case OBJ_BAD:
2194         case OBJ_COMMIT:
2195         case OBJ_TREE:
2196         case OBJ_BLOB:
2197         case OBJ_TAG:
2198                 break;
2199         default:
2200                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2201                       type, (uintmax_t)obj_offset, p->pack_name);
2202                 type = OBJ_BAD;
2203         }
2204
2205 out:
2206         if (poi_stack != small_poi_stack)
2207                 free(poi_stack);
2208         return type;
2209
2210 unwind:
2211         while (poi_stack_nr) {
2212                 obj_offset = poi_stack[--poi_stack_nr];
2213                 type = retry_bad_packed_offset(p, obj_offset);
2214                 if (type > OBJ_NONE)
2215                         goto out;
2216         }
2217         type = OBJ_BAD;
2218         goto out;
2219 }
2220
2221 static struct hashmap delta_base_cache;
2222 static size_t delta_base_cached;
2223
2224 static LIST_HEAD(delta_base_cache_lru);
2225
2226 struct delta_base_cache_key {
2227         struct packed_git *p;
2228         off_t base_offset;
2229 };
2230
2231 struct delta_base_cache_entry {
2232         struct hashmap hash;
2233         struct delta_base_cache_key key;
2234         struct list_head lru;
2235         void *data;
2236         unsigned long size;
2237         enum object_type type;
2238 };
2239
2240 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2241 {
2242         unsigned int hash;
2243
2244         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2245         hash += (hash >> 8) + (hash >> 16);
2246         return hash;
2247 }
2248
2249 static struct delta_base_cache_entry *
2250 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2251 {
2252         struct hashmap_entry entry;
2253         struct delta_base_cache_key key;
2254
2255         if (!delta_base_cache.cmpfn)
2256                 return NULL;
2257
2258         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2259         key.p = p;
2260         key.base_offset = base_offset;
2261         return hashmap_get(&delta_base_cache, &entry, &key);
2262 }
2263
2264 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2265                                    const struct delta_base_cache_key *b)
2266 {
2267         return a->p == b->p && a->base_offset == b->base_offset;
2268 }
2269
2270 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
2271                                      const void *va, const void *vb,
2272                                      const void *vkey)
2273 {
2274         const struct delta_base_cache_entry *a = va, *b = vb;
2275         const struct delta_base_cache_key *key = vkey;
2276         if (key)
2277                 return !delta_base_cache_key_eq(&a->key, key);
2278         else
2279                 return !delta_base_cache_key_eq(&a->key, &b->key);
2280 }
2281
2282 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2283 {
2284         return !!get_delta_base_cache_entry(p, base_offset);
2285 }
2286
2287 /*
2288  * Remove the entry from the cache, but do _not_ free the associated
2289  * entry data. The caller takes ownership of the "data" buffer, and
2290  * should copy out any fields it wants before detaching.
2291  */
2292 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2293 {
2294         hashmap_remove(&delta_base_cache, ent, &ent->key);
2295         list_del(&ent->lru);
2296         delta_base_cached -= ent->size;
2297         free(ent);
2298 }
2299
2300 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2301         unsigned long *base_size, enum object_type *type)
2302 {
2303         struct delta_base_cache_entry *ent;
2304
2305         ent = get_delta_base_cache_entry(p, base_offset);
2306         if (!ent)
2307                 return unpack_entry(p, base_offset, type, base_size);
2308
2309         if (type)
2310                 *type = ent->type;
2311         if (base_size)
2312                 *base_size = ent->size;
2313         return xmemdupz(ent->data, ent->size);
2314 }
2315
2316 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2317 {
2318         free(ent->data);
2319         detach_delta_base_cache_entry(ent);
2320 }
2321
2322 void clear_delta_base_cache(void)
2323 {
2324         struct list_head *lru, *tmp;
2325         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2326                 struct delta_base_cache_entry *entry =
2327                         list_entry(lru, struct delta_base_cache_entry, lru);
2328                 release_delta_base_cache(entry);
2329         }
2330 }
2331
2332 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2333         void *base, unsigned long base_size, enum object_type type)
2334 {
2335         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2336         struct list_head *lru, *tmp;
2337
2338         delta_base_cached += base_size;
2339
2340         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2341                 struct delta_base_cache_entry *f =
2342                         list_entry(lru, struct delta_base_cache_entry, lru);
2343                 if (delta_base_cached <= delta_base_cache_limit)
2344                         break;
2345                 release_delta_base_cache(f);
2346         }
2347
2348         ent->key.p = p;
2349         ent->key.base_offset = base_offset;
2350         ent->type = type;
2351         ent->data = base;
2352         ent->size = base_size;
2353         list_add_tail(&ent->lru, &delta_base_cache_lru);
2354
2355         if (!delta_base_cache.cmpfn)
2356                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
2357         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2358         hashmap_add(&delta_base_cache, ent);
2359 }
2360
2361 int packed_object_info(struct packed_git *p, off_t obj_offset,
2362                        struct object_info *oi)
2363 {
2364         struct pack_window *w_curs = NULL;
2365         unsigned long size;
2366         off_t curpos = obj_offset;
2367         enum object_type type;
2368
2369         /*
2370          * We always get the representation type, but only convert it to
2371          * a "real" type later if the caller is interested.
2372          */
2373         if (oi->contentp) {
2374                 *oi->contentp = cache_or_unpack_entry(p, obj_offset, oi->sizep,
2375                                                       &type);
2376                 if (!*oi->contentp)
2377                         type = OBJ_BAD;
2378         } else {
2379                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2380         }
2381
2382         if (!oi->contentp && oi->sizep) {
2383                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2384                         off_t tmp_pos = curpos;
2385                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2386                                                            type, obj_offset);
2387                         if (!base_offset) {
2388                                 type = OBJ_BAD;
2389                                 goto out;
2390                         }
2391                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2392                         if (*oi->sizep == 0) {
2393                                 type = OBJ_BAD;
2394                                 goto out;
2395                         }
2396                 } else {
2397                         *oi->sizep = size;
2398                 }
2399         }
2400
2401         if (oi->disk_sizep) {
2402                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2403                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2404         }
2405
2406         if (oi->typep || oi->typename) {
2407                 enum object_type ptot;
2408                 ptot = packed_to_object_type(p, obj_offset, type, &w_curs,
2409                                              curpos);
2410                 if (oi->typep)
2411                         *oi->typep = ptot;
2412                 if (oi->typename) {
2413                         const char *tn = typename(ptot);
2414                         if (tn)
2415                                 strbuf_addstr(oi->typename, tn);
2416                 }
2417                 if (ptot < 0) {
2418                         type = OBJ_BAD;
2419                         goto out;
2420                 }
2421         }
2422
2423         if (oi->delta_base_sha1) {
2424                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2425                         const unsigned char *base;
2426
2427                         base = get_delta_base_sha1(p, &w_curs, curpos,
2428                                                    type, obj_offset);
2429                         if (!base) {
2430                                 type = OBJ_BAD;
2431                                 goto out;
2432                         }
2433
2434                         hashcpy(oi->delta_base_sha1, base);
2435                 } else
2436                         hashclr(oi->delta_base_sha1);
2437         }
2438
2439 out:
2440         unuse_pack(&w_curs);
2441         return type;
2442 }
2443
2444 static void *unpack_compressed_entry(struct packed_git *p,
2445                                     struct pack_window **w_curs,
2446                                     off_t curpos,
2447                                     unsigned long size)
2448 {
2449         int st;
2450         git_zstream stream;
2451         unsigned char *buffer, *in;
2452
2453         buffer = xmallocz_gently(size);
2454         if (!buffer)
2455                 return NULL;
2456         memset(&stream, 0, sizeof(stream));
2457         stream.next_out = buffer;
2458         stream.avail_out = size + 1;
2459
2460         git_inflate_init(&stream);
2461         do {
2462                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2463                 stream.next_in = in;
2464                 st = git_inflate(&stream, Z_FINISH);
2465                 if (!stream.avail_out)
2466                         break; /* the payload is larger than it should be */
2467                 curpos += stream.next_in - in;
2468         } while (st == Z_OK || st == Z_BUF_ERROR);
2469         git_inflate_end(&stream);
2470         if ((st != Z_STREAM_END) || stream.total_out != size) {
2471                 free(buffer);
2472                 return NULL;
2473         }
2474
2475         return buffer;
2476 }
2477
2478 static void *read_object(const unsigned char *sha1, enum object_type *type,
2479                          unsigned long *size);
2480
2481 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2482 {
2483         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2484         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2485                          p->pack_name, (uintmax_t)obj_offset);
2486 }
2487
2488 int do_check_packed_object_crc;
2489
2490 #define UNPACK_ENTRY_STACK_PREALLOC 64
2491 struct unpack_entry_stack_ent {
2492         off_t obj_offset;
2493         off_t curpos;
2494         unsigned long size;
2495 };
2496
2497 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2498                    enum object_type *final_type, unsigned long *final_size)
2499 {
2500         struct pack_window *w_curs = NULL;
2501         off_t curpos = obj_offset;
2502         void *data = NULL;
2503         unsigned long size;
2504         enum object_type type;
2505         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2506         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2507         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2508         int base_from_cache = 0;
2509
2510         write_pack_access_log(p, obj_offset);
2511
2512         /* PHASE 1: drill down to the innermost base object */
2513         for (;;) {
2514                 off_t base_offset;
2515                 int i;
2516                 struct delta_base_cache_entry *ent;
2517
2518                 ent = get_delta_base_cache_entry(p, curpos);
2519                 if (ent) {
2520                         type = ent->type;
2521                         data = ent->data;
2522                         size = ent->size;
2523                         detach_delta_base_cache_entry(ent);
2524                         base_from_cache = 1;
2525                         break;
2526                 }
2527
2528                 if (do_check_packed_object_crc && p->index_version > 1) {
2529                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2530                         off_t len = revidx[1].offset - obj_offset;
2531                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2532                                 const unsigned char *sha1 =
2533                                         nth_packed_object_sha1(p, revidx->nr);
2534                                 error("bad packed object CRC for %s",
2535                                       sha1_to_hex(sha1));
2536                                 mark_bad_packed_object(p, sha1);
2537                                 data = NULL;
2538                                 goto out;
2539                         }
2540                 }
2541
2542                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2543                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2544                         break;
2545
2546                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2547                 if (!base_offset) {
2548                         error("failed to validate delta base reference "
2549                               "at offset %"PRIuMAX" from %s",
2550                               (uintmax_t)curpos, p->pack_name);
2551                         /* bail to phase 2, in hopes of recovery */
2552                         data = NULL;
2553                         break;
2554                 }
2555
2556                 /* push object, proceed to base */
2557                 if (delta_stack_nr >= delta_stack_alloc
2558                     && delta_stack == small_delta_stack) {
2559                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2560                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2561                         memcpy(delta_stack, small_delta_stack,
2562                                sizeof(*delta_stack)*delta_stack_nr);
2563                 } else {
2564                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2565                 }
2566                 i = delta_stack_nr++;
2567                 delta_stack[i].obj_offset = obj_offset;
2568                 delta_stack[i].curpos = curpos;
2569                 delta_stack[i].size = size;
2570
2571                 curpos = obj_offset = base_offset;
2572         }
2573
2574         /* PHASE 2: handle the base */
2575         switch (type) {
2576         case OBJ_OFS_DELTA:
2577         case OBJ_REF_DELTA:
2578                 if (data)
2579                         die("BUG: unpack_entry: left loop at a valid delta");
2580                 break;
2581         case OBJ_COMMIT:
2582         case OBJ_TREE:
2583         case OBJ_BLOB:
2584         case OBJ_TAG:
2585                 if (!base_from_cache)
2586                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2587                 break;
2588         default:
2589                 data = NULL;
2590                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2591                       type, (uintmax_t)obj_offset, p->pack_name);
2592         }
2593
2594         /* PHASE 3: apply deltas in order */
2595
2596         /* invariants:
2597          *   'data' holds the base data, or NULL if there was corruption
2598          */
2599         while (delta_stack_nr) {
2600                 void *delta_data;
2601                 void *base = data;
2602                 void *external_base = NULL;
2603                 unsigned long delta_size, base_size = size;
2604                 int i;
2605
2606                 data = NULL;
2607
2608                 if (base)
2609                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2610
2611                 if (!base) {
2612                         /*
2613                          * We're probably in deep shit, but let's try to fetch
2614                          * the required base anyway from another pack or loose.
2615                          * This is costly but should happen only in the presence
2616                          * of a corrupted pack, and is better than failing outright.
2617                          */
2618                         struct revindex_entry *revidx;
2619                         const unsigned char *base_sha1;
2620                         revidx = find_pack_revindex(p, obj_offset);
2621                         if (revidx) {
2622                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2623                                 error("failed to read delta base object %s"
2624                                       " at offset %"PRIuMAX" from %s",
2625                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2626                                       p->pack_name);
2627                                 mark_bad_packed_object(p, base_sha1);
2628                                 base = read_object(base_sha1, &type, &base_size);
2629                                 external_base = base;
2630                         }
2631                 }
2632
2633                 i = --delta_stack_nr;
2634                 obj_offset = delta_stack[i].obj_offset;
2635                 curpos = delta_stack[i].curpos;
2636                 delta_size = delta_stack[i].size;
2637
2638                 if (!base)
2639                         continue;
2640
2641                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2642
2643                 if (!delta_data) {
2644                         error("failed to unpack compressed delta "
2645                               "at offset %"PRIuMAX" from %s",
2646                               (uintmax_t)curpos, p->pack_name);
2647                         data = NULL;
2648                         free(external_base);
2649                         continue;
2650                 }
2651
2652                 data = patch_delta(base, base_size,
2653                                    delta_data, delta_size,
2654                                    &size);
2655
2656                 /*
2657                  * We could not apply the delta; warn the user, but keep going.
2658                  * Our failure will be noticed either in the next iteration of
2659                  * the loop, or if this is the final delta, in the caller when
2660                  * we return NULL. Those code paths will take care of making
2661                  * a more explicit warning and retrying with another copy of
2662                  * the object.
2663                  */
2664                 if (!data)
2665                         error("failed to apply delta");
2666
2667                 free(delta_data);
2668                 free(external_base);
2669         }
2670
2671         if (final_type)
2672                 *final_type = type;
2673         if (final_size)
2674                 *final_size = size;
2675
2676 out:
2677         unuse_pack(&w_curs);
2678
2679         if (delta_stack != small_delta_stack)
2680                 free(delta_stack);
2681
2682         return data;
2683 }
2684
2685 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2686                                             uint32_t n)
2687 {
2688         const unsigned char *index = p->index_data;
2689         if (!index) {
2690                 if (open_pack_index(p))
2691                         return NULL;
2692                 index = p->index_data;
2693         }
2694         if (n >= p->num_objects)
2695                 return NULL;
2696         index += 4 * 256;
2697         if (p->index_version == 1) {
2698                 return index + 24 * n + 4;
2699         } else {
2700                 index += 8;
2701                 return index + 20 * n;
2702         }
2703 }
2704
2705 const struct object_id *nth_packed_object_oid(struct object_id *oid,
2706                                               struct packed_git *p,
2707                                               uint32_t n)
2708 {
2709         const unsigned char *hash = nth_packed_object_sha1(p, n);
2710         if (!hash)
2711                 return NULL;
2712         hashcpy(oid->hash, hash);
2713         return oid;
2714 }
2715
2716 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2717 {
2718         const unsigned char *ptr = vptr;
2719         const unsigned char *start = p->index_data;
2720         const unsigned char *end = start + p->index_size;
2721         if (ptr < start)
2722                 die(_("offset before start of pack index for %s (corrupt index?)"),
2723                     p->pack_name);
2724         /* No need to check for underflow; .idx files must be at least 8 bytes */
2725         if (ptr >= end - 8)
2726                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2727                     p->pack_name);
2728 }
2729
2730 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2731 {
2732         const unsigned char *index = p->index_data;
2733         index += 4 * 256;
2734         if (p->index_version == 1) {
2735                 return ntohl(*((uint32_t *)(index + 24 * n)));
2736         } else {
2737                 uint32_t off;
2738                 index += 8 + p->num_objects * (20 + 4);
2739                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2740                 if (!(off & 0x80000000))
2741                         return off;
2742                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2743                 check_pack_index_ptr(p, index);
2744                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2745                                    ntohl(*((uint32_t *)(index + 4)));
2746         }
2747 }
2748
2749 off_t find_pack_entry_one(const unsigned char *sha1,
2750                                   struct packed_git *p)
2751 {
2752         const uint32_t *level1_ofs = p->index_data;
2753         const unsigned char *index = p->index_data;
2754         unsigned hi, lo, stride;
2755         static int debug_lookup = -1;
2756
2757         if (debug_lookup < 0)
2758                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2759
2760         if (!index) {
2761                 if (open_pack_index(p))
2762                         return 0;
2763                 level1_ofs = p->index_data;
2764                 index = p->index_data;
2765         }
2766         if (p->index_version > 1) {
2767                 level1_ofs += 2;
2768                 index += 8;
2769         }
2770         index += 4 * 256;
2771         hi = ntohl(level1_ofs[*sha1]);
2772         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2773         if (p->index_version > 1) {
2774                 stride = 20;
2775         } else {
2776                 stride = 24;
2777                 index += 4;
2778         }
2779
2780         if (debug_lookup)
2781                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2782                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2783
2784         while (lo < hi) {
2785                 unsigned mi = (lo + hi) / 2;
2786                 int cmp = hashcmp(index + mi * stride, sha1);
2787
2788                 if (debug_lookup)
2789                         printf("lo %u hi %u rg %u mi %u\n",
2790                                lo, hi, hi - lo, mi);
2791                 if (!cmp)
2792                         return nth_packed_object_offset(p, mi);
2793                 if (cmp > 0)
2794                         hi = mi;
2795                 else
2796                         lo = mi+1;
2797         }
2798         return 0;
2799 }
2800
2801 int is_pack_valid(struct packed_git *p)
2802 {
2803         /* An already open pack is known to be valid. */
2804         if (p->pack_fd != -1)
2805                 return 1;
2806
2807         /* If the pack has one window completely covering the
2808          * file size, the pack is known to be valid even if
2809          * the descriptor is not currently open.
2810          */
2811         if (p->windows) {
2812                 struct pack_window *w = p->windows;
2813
2814                 if (!w->offset && w->len == p->pack_size)
2815                         return 1;
2816         }
2817
2818         /* Force the pack to open to prove its valid. */
2819         return !open_packed_git(p);
2820 }
2821
2822 static int fill_pack_entry(const unsigned char *sha1,
2823                            struct pack_entry *e,
2824                            struct packed_git *p)
2825 {
2826         off_t offset;
2827
2828         if (p->num_bad_objects) {
2829                 unsigned i;
2830                 for (i = 0; i < p->num_bad_objects; i++)
2831                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2832                                 return 0;
2833         }
2834
2835         offset = find_pack_entry_one(sha1, p);
2836         if (!offset)
2837                 return 0;
2838
2839         /*
2840          * We are about to tell the caller where they can locate the
2841          * requested object.  We better make sure the packfile is
2842          * still here and can be accessed before supplying that
2843          * answer, as it may have been deleted since the index was
2844          * loaded!
2845          */
2846         if (!is_pack_valid(p))
2847                 return 0;
2848         e->offset = offset;
2849         e->p = p;
2850         hashcpy(e->sha1, sha1);
2851         return 1;
2852 }
2853
2854 /*
2855  * Iff a pack file contains the object named by sha1, return true and
2856  * store its location to e.
2857  */
2858 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2859 {
2860         struct mru_entry *p;
2861
2862         prepare_packed_git();
2863         if (!packed_git)
2864                 return 0;
2865
2866         for (p = packed_git_mru->head; p; p = p->next) {
2867                 if (fill_pack_entry(sha1, e, p->item)) {
2868                         mru_mark(packed_git_mru, p);
2869                         return 1;
2870                 }
2871         }
2872         return 0;
2873 }
2874
2875 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2876                                   struct packed_git *packs)
2877 {
2878         struct packed_git *p;
2879
2880         for (p = packs; p; p = p->next) {
2881                 if (find_pack_entry_one(sha1, p))
2882                         return p;
2883         }
2884         return NULL;
2885
2886 }
2887
2888 static int sha1_loose_object_info(const unsigned char *sha1,
2889                                   struct object_info *oi,
2890                                   int flags)
2891 {
2892         int status = 0;
2893         unsigned long mapsize;
2894         void *map;
2895         git_zstream stream;
2896         char hdr[32];
2897         struct strbuf hdrbuf = STRBUF_INIT;
2898         unsigned long size_scratch;
2899
2900         if (oi->delta_base_sha1)
2901                 hashclr(oi->delta_base_sha1);
2902
2903         /*
2904          * If we don't care about type or size, then we don't
2905          * need to look inside the object at all. Note that we
2906          * do not optimize out the stat call, even if the
2907          * caller doesn't care about the disk-size, since our
2908          * return value implicitly indicates whether the
2909          * object even exists.
2910          */
2911         if (!oi->typep && !oi->typename && !oi->sizep && !oi->contentp) {
2912                 const char *path;
2913                 struct stat st;
2914                 if (stat_sha1_file(sha1, &st, &path) < 0)
2915                         return -1;
2916                 if (oi->disk_sizep)
2917                         *oi->disk_sizep = st.st_size;
2918                 return 0;
2919         }
2920
2921         map = map_sha1_file(sha1, &mapsize);
2922         if (!map)
2923                 return -1;
2924
2925         if (!oi->sizep)
2926                 oi->sizep = &size_scratch;
2927
2928         if (oi->disk_sizep)
2929                 *oi->disk_sizep = mapsize;
2930         if ((flags & OBJECT_INFO_ALLOW_UNKNOWN_TYPE)) {
2931                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2932                         status = error("unable to unpack %s header with --allow-unknown-type",
2933                                        sha1_to_hex(sha1));
2934         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2935                 status = error("unable to unpack %s header",
2936                                sha1_to_hex(sha1));
2937         if (status < 0)
2938                 ; /* Do nothing */
2939         else if (hdrbuf.len) {
2940                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2941                         status = error("unable to parse %s header with --allow-unknown-type",
2942                                        sha1_to_hex(sha1));
2943         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2944                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2945
2946         if (status >= 0 && oi->contentp) {
2947                 *oi->contentp = unpack_sha1_rest(&stream, hdr,
2948                                                  *oi->sizep, sha1);
2949                 if (!*oi->contentp) {
2950                         git_inflate_end(&stream);
2951                         status = -1;
2952                 }
2953         } else
2954                 git_inflate_end(&stream);
2955
2956         munmap(map, mapsize);
2957         if (status && oi->typep)
2958                 *oi->typep = status;
2959         if (oi->sizep == &size_scratch)
2960                 oi->sizep = NULL;
2961         strbuf_release(&hdrbuf);
2962         return (status < 0) ? status : 0;
2963 }
2964
2965 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2966 {
2967         static struct object_info blank_oi = OBJECT_INFO_INIT;
2968         struct pack_entry e;
2969         int rtype;
2970         const unsigned char *real = (flags & OBJECT_INFO_LOOKUP_REPLACE) ?
2971                                     lookup_replace_object(sha1) :
2972                                     sha1;
2973
2974         if (!oi)
2975                 oi = &blank_oi;
2976
2977         if (!(flags & OBJECT_INFO_SKIP_CACHED)) {
2978                 struct cached_object *co = find_cached_object(real);
2979                 if (co) {
2980                         if (oi->typep)
2981                                 *(oi->typep) = co->type;
2982                         if (oi->sizep)
2983                                 *(oi->sizep) = co->size;
2984                         if (oi->disk_sizep)
2985                                 *(oi->disk_sizep) = 0;
2986                         if (oi->delta_base_sha1)
2987                                 hashclr(oi->delta_base_sha1);
2988                         if (oi->typename)
2989                                 strbuf_addstr(oi->typename, typename(co->type));
2990                         if (oi->contentp)
2991                                 *oi->contentp = xmemdupz(co->buf, co->size);
2992                         oi->whence = OI_CACHED;
2993                         return 0;
2994                 }
2995         }
2996
2997         if (!find_pack_entry(real, &e)) {
2998                 /* Most likely it's a loose object. */
2999                 if (!sha1_loose_object_info(real, oi, flags)) {
3000                         oi->whence = OI_LOOSE;
3001                         return 0;
3002                 }
3003
3004                 /* Not a loose object; someone else may have just packed it. */
3005                 if (flags & OBJECT_INFO_QUICK) {
3006                         return -1;
3007                 } else {
3008                         reprepare_packed_git();
3009                         if (!find_pack_entry(real, &e))
3010                                 return -1;
3011                 }
3012         }
3013
3014         if (oi == &blank_oi)
3015                 /*
3016                  * We know that the caller doesn't actually need the
3017                  * information below, so return early.
3018                  */
3019                 return 0;
3020
3021         rtype = packed_object_info(e.p, e.offset, oi);
3022         if (rtype < 0) {
3023                 mark_bad_packed_object(e.p, real);
3024                 return sha1_object_info_extended(real, oi, 0);
3025         } else if (in_delta_base_cache(e.p, e.offset)) {
3026                 oi->whence = OI_DBCACHED;
3027         } else {
3028                 oi->whence = OI_PACKED;
3029                 oi->u.packed.offset = e.offset;
3030                 oi->u.packed.pack = e.p;
3031                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
3032                                          rtype == OBJ_OFS_DELTA);
3033         }
3034
3035         return 0;
3036 }
3037
3038 /* returns enum object_type or negative */
3039 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
3040 {
3041         enum object_type type;
3042         struct object_info oi = OBJECT_INFO_INIT;
3043
3044         oi.typep = &type;
3045         oi.sizep = sizep;
3046         if (sha1_object_info_extended(sha1, &oi,
3047                                       OBJECT_INFO_LOOKUP_REPLACE) < 0)
3048                 return -1;
3049         return type;
3050 }
3051
3052 static void *read_packed_sha1(const unsigned char *sha1,
3053                               enum object_type *type, unsigned long *size)
3054 {
3055         struct pack_entry e;
3056         void *data;
3057
3058         if (!find_pack_entry(sha1, &e))
3059                 return NULL;
3060         data = cache_or_unpack_entry(e.p, e.offset, size, type);
3061         if (!data) {
3062                 /*
3063                  * We're probably in deep shit, but let's try to fetch
3064                  * the required object anyway from another pack or loose.
3065                  * This should happen only in the presence of a corrupted
3066                  * pack, and is better than failing outright.
3067                  */
3068                 error("failed to read object %s at offset %"PRIuMAX" from %s",
3069                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
3070                 mark_bad_packed_object(e.p, sha1);
3071                 data = read_object(sha1, type, size);
3072         }
3073         return data;
3074 }
3075
3076 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
3077                       unsigned char *sha1)
3078 {
3079         struct cached_object *co;
3080
3081         hash_sha1_file(buf, len, typename(type), sha1);
3082         if (has_sha1_file(sha1) || find_cached_object(sha1))
3083                 return 0;
3084         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
3085         co = &cached_objects[cached_object_nr++];
3086         co->size = len;
3087         co->type = type;
3088         co->buf = xmalloc(len);
3089         memcpy(co->buf, buf, len);
3090         hashcpy(co->sha1, sha1);
3091         return 0;
3092 }
3093
3094 static void *read_object(const unsigned char *sha1, enum object_type *type,
3095                          unsigned long *size)
3096 {
3097         struct object_info oi = OBJECT_INFO_INIT;
3098         void *content;
3099         oi.typep = type;
3100         oi.sizep = size;
3101         oi.contentp = &content;
3102
3103         if (sha1_object_info_extended(sha1, &oi, 0) < 0)
3104                 return NULL;
3105         return content;
3106 }
3107
3108 /*
3109  * This function dies on corrupt objects; the callers who want to
3110  * deal with them should arrange to call read_object() and give error
3111  * messages themselves.
3112  */
3113 void *read_sha1_file_extended(const unsigned char *sha1,
3114                               enum object_type *type,
3115                               unsigned long *size,
3116                               int lookup_replace)
3117 {
3118         void *data;
3119         const struct packed_git *p;
3120         const char *path;
3121         struct stat st;
3122         const unsigned char *repl = lookup_replace ? lookup_replace_object(sha1)
3123                                                    : sha1;
3124
3125         errno = 0;
3126         data = read_object(repl, type, size);
3127         if (data)
3128                 return data;
3129
3130         if (errno && errno != ENOENT)
3131                 die_errno("failed to read object %s", sha1_to_hex(sha1));
3132
3133         /* die if we replaced an object with one that does not exist */
3134         if (repl != sha1)
3135                 die("replacement %s not found for %s",
3136                     sha1_to_hex(repl), sha1_to_hex(sha1));
3137
3138         if (!stat_sha1_file(repl, &st, &path))
3139                 die("loose object %s (stored in %s) is corrupt",
3140                     sha1_to_hex(repl), path);
3141
3142         if ((p = has_packed_and_bad(repl)) != NULL)
3143                 die("packed object %s (stored in %s) is corrupt",
3144                     sha1_to_hex(repl), p->pack_name);
3145
3146         return NULL;
3147 }
3148
3149 void *read_object_with_reference(const unsigned char *sha1,
3150                                  const char *required_type_name,
3151                                  unsigned long *size,
3152                                  unsigned char *actual_sha1_return)
3153 {
3154         enum object_type type, required_type;
3155         void *buffer;
3156         unsigned long isize;
3157         unsigned char actual_sha1[20];
3158
3159         required_type = type_from_string(required_type_name);
3160         hashcpy(actual_sha1, sha1);
3161         while (1) {
3162                 int ref_length = -1;
3163                 const char *ref_type = NULL;
3164
3165                 buffer = read_sha1_file(actual_sha1, &type, &isize);
3166                 if (!buffer)
3167                         return NULL;
3168                 if (type == required_type) {
3169                         *size = isize;
3170                         if (actual_sha1_return)
3171                                 hashcpy(actual_sha1_return, actual_sha1);
3172                         return buffer;
3173                 }
3174                 /* Handle references */
3175                 else if (type == OBJ_COMMIT)
3176                         ref_type = "tree ";
3177                 else if (type == OBJ_TAG)
3178                         ref_type = "object ";
3179                 else {
3180                         free(buffer);
3181                         return NULL;
3182                 }
3183                 ref_length = strlen(ref_type);
3184
3185                 if (ref_length + 40 > isize ||
3186                     memcmp(buffer, ref_type, ref_length) ||
3187                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
3188                         free(buffer);
3189                         return NULL;
3190                 }
3191                 free(buffer);
3192                 /* Now we have the ID of the referred-to object in
3193                  * actual_sha1.  Check again. */
3194         }
3195 }
3196
3197 static void write_sha1_file_prepare(const void *buf, unsigned long len,
3198                                     const char *type, unsigned char *sha1,
3199                                     char *hdr, int *hdrlen)
3200 {
3201         git_SHA_CTX c;
3202
3203         /* Generate the header */
3204         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
3205
3206         /* Sha1.. */
3207         git_SHA1_Init(&c);
3208         git_SHA1_Update(&c, hdr, *hdrlen);
3209         git_SHA1_Update(&c, buf, len);
3210         git_SHA1_Final(sha1, &c);
3211 }
3212
3213 /*
3214  * Move the just written object into its final resting place.
3215  */
3216 int finalize_object_file(const char *tmpfile, const char *filename)
3217 {
3218         int ret = 0;
3219
3220         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3221                 goto try_rename;
3222         else if (link(tmpfile, filename))
3223                 ret = errno;
3224
3225         /*
3226          * Coda hack - coda doesn't like cross-directory links,
3227          * so we fall back to a rename, which will mean that it
3228          * won't be able to check collisions, but that's not a
3229          * big deal.
3230          *
3231          * The same holds for FAT formatted media.
3232          *
3233          * When this succeeds, we just return.  We have nothing
3234          * left to unlink.
3235          */
3236         if (ret && ret != EEXIST) {
3237         try_rename:
3238                 if (!rename(tmpfile, filename))
3239                         goto out;
3240                 ret = errno;
3241         }
3242         unlink_or_warn(tmpfile);
3243         if (ret) {
3244                 if (ret != EEXIST) {
3245                         return error_errno("unable to write sha1 filename %s", filename);
3246                 }
3247                 /* FIXME!!! Collision check here ? */
3248         }
3249
3250 out:
3251         if (adjust_shared_perm(filename))
3252                 return error("unable to set permission to '%s'", filename);
3253         return 0;
3254 }
3255
3256 static int write_buffer(int fd, const void *buf, size_t len)
3257 {
3258         if (write_in_full(fd, buf, len) < 0)
3259                 return error_errno("file write error");
3260         return 0;
3261 }
3262
3263 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3264                    unsigned char *sha1)
3265 {
3266         char hdr[32];
3267         int hdrlen = sizeof(hdr);
3268         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3269         return 0;
3270 }
3271
3272 /* Finalize a file on disk, and close it. */
3273 static void close_sha1_file(int fd)
3274 {
3275         if (fsync_object_files)
3276                 fsync_or_die(fd, "sha1 file");
3277         if (close(fd) != 0)
3278                 die_errno("error when closing sha1 file");
3279 }
3280
3281 /* Size of directory component, including the ending '/' */
3282 static inline int directory_size(const char *filename)
3283 {
3284         const char *s = strrchr(filename, '/');
3285         if (!s)
3286                 return 0;
3287         return s - filename + 1;
3288 }
3289
3290 /*
3291  * This creates a temporary file in the same directory as the final
3292  * 'filename'
3293  *
3294  * We want to avoid cross-directory filename renames, because those
3295  * can have problems on various filesystems (FAT, NFS, Coda).
3296  */
3297 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3298 {
3299         int fd, dirlen = directory_size(filename);
3300
3301         strbuf_reset(tmp);
3302         strbuf_add(tmp, filename, dirlen);
3303         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3304         fd = git_mkstemp_mode(tmp->buf, 0444);
3305         if (fd < 0 && dirlen && errno == ENOENT) {
3306                 /*
3307                  * Make sure the directory exists; note that the contents
3308                  * of the buffer are undefined after mkstemp returns an
3309                  * error, so we have to rewrite the whole buffer from
3310                  * scratch.
3311                  */
3312                 strbuf_reset(tmp);
3313                 strbuf_add(tmp, filename, dirlen - 1);
3314                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3315                         return -1;
3316                 if (adjust_shared_perm(tmp->buf))
3317                         return -1;
3318
3319                 /* Try again */
3320                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3321                 fd = git_mkstemp_mode(tmp->buf, 0444);
3322         }
3323         return fd;
3324 }
3325
3326 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3327                               const void *buf, unsigned long len, time_t mtime)
3328 {
3329         int fd, ret;
3330         unsigned char compressed[4096];
3331         git_zstream stream;
3332         git_SHA_CTX c;
3333         unsigned char parano_sha1[20];
3334         static struct strbuf tmp_file = STRBUF_INIT;
3335         const char *filename = sha1_file_name(sha1);
3336
3337         fd = create_tmpfile(&tmp_file, filename);
3338         if (fd < 0) {
3339                 if (errno == EACCES)
3340                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3341                 else
3342                         return error_errno("unable to create temporary file");
3343         }
3344
3345         /* Set it up */
3346         git_deflate_init(&stream, zlib_compression_level);
3347         stream.next_out = compressed;
3348         stream.avail_out = sizeof(compressed);
3349         git_SHA1_Init(&c);
3350
3351         /* First header.. */
3352         stream.next_in = (unsigned char *)hdr;
3353         stream.avail_in = hdrlen;
3354         while (git_deflate(&stream, 0) == Z_OK)
3355                 ; /* nothing */
3356         git_SHA1_Update(&c, hdr, hdrlen);
3357
3358         /* Then the data itself.. */
3359         stream.next_in = (void *)buf;
3360         stream.avail_in = len;
3361         do {
3362                 unsigned char *in0 = stream.next_in;
3363                 ret = git_deflate(&stream, Z_FINISH);
3364                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3365                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3366                         die("unable to write sha1 file");
3367                 stream.next_out = compressed;
3368                 stream.avail_out = sizeof(compressed);
3369         } while (ret == Z_OK);
3370
3371         if (ret != Z_STREAM_END)
3372                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3373         ret = git_deflate_end_gently(&stream);
3374         if (ret != Z_OK)
3375                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3376         git_SHA1_Final(parano_sha1, &c);
3377         if (hashcmp(sha1, parano_sha1) != 0)
3378                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3379
3380         close_sha1_file(fd);
3381
3382         if (mtime) {
3383                 struct utimbuf utb;
3384                 utb.actime = mtime;
3385                 utb.modtime = mtime;
3386                 if (utime(tmp_file.buf, &utb) < 0)
3387                         warning_errno("failed utime() on %s", tmp_file.buf);
3388         }
3389
3390         return finalize_object_file(tmp_file.buf, filename);
3391 }
3392
3393 static int freshen_loose_object(const unsigned char *sha1)
3394 {
3395         return check_and_freshen(sha1, 1);
3396 }
3397
3398 static int freshen_packed_object(const unsigned char *sha1)
3399 {
3400         struct pack_entry e;
3401         if (!find_pack_entry(sha1, &e))
3402                 return 0;
3403         if (e.p->freshened)
3404                 return 1;
3405         if (!freshen_file(e.p->pack_name))
3406                 return 0;
3407         e.p->freshened = 1;
3408         return 1;
3409 }
3410
3411 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3412 {
3413         char hdr[32];
3414         int hdrlen = sizeof(hdr);
3415
3416         /* Normally if we have it in the pack then we do not bother writing
3417          * it out into .git/objects/??/?{38} file.
3418          */
3419         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3420         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3421                 return 0;
3422         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3423 }
3424
3425 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3426                              unsigned char *sha1, unsigned flags)
3427 {
3428         char *header;
3429         int hdrlen, status = 0;
3430
3431         /* type string, SP, %lu of the length plus NUL must fit this */
3432         hdrlen = strlen(type) + 32;
3433         header = xmalloc(hdrlen);
3434         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3435
3436         if (!(flags & HASH_WRITE_OBJECT))
3437                 goto cleanup;
3438         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3439                 goto cleanup;
3440         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3441
3442 cleanup:
3443         free(header);
3444         return status;
3445 }
3446
3447 int force_object_loose(const unsigned char *sha1, time_t mtime)
3448 {
3449         void *buf;
3450         unsigned long len;
3451         enum object_type type;
3452         char hdr[32];
3453         int hdrlen;
3454         int ret;
3455
3456         if (has_loose_object(sha1))
3457                 return 0;
3458         buf = read_packed_sha1(sha1, &type, &len);
3459         if (!buf)
3460                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3461         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3462         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3463         free(buf);
3464
3465         return ret;
3466 }
3467
3468 int has_pack_index(const unsigned char *sha1)
3469 {
3470         struct stat st;
3471         if (stat(sha1_pack_index_name(sha1), &st))
3472                 return 0;
3473         return 1;
3474 }
3475
3476 int has_sha1_pack(const unsigned char *sha1)
3477 {
3478         struct pack_entry e;
3479         return find_pack_entry(sha1, &e);
3480 }
3481
3482 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3483 {
3484         if (!startup_info->have_repository)
3485                 return 0;
3486         return sha1_object_info_extended(sha1, NULL,
3487                                          flags | OBJECT_INFO_SKIP_CACHED) >= 0;
3488 }
3489
3490 int has_object_file(const struct object_id *oid)
3491 {
3492         return has_sha1_file(oid->hash);
3493 }
3494
3495 int has_object_file_with_flags(const struct object_id *oid, int flags)
3496 {
3497         return has_sha1_file_with_flags(oid->hash, flags);
3498 }
3499
3500 static void check_tree(const void *buf, size_t size)
3501 {
3502         struct tree_desc desc;
3503         struct name_entry entry;
3504
3505         init_tree_desc(&desc, buf, size);
3506         while (tree_entry(&desc, &entry))
3507                 /* do nothing
3508                  * tree_entry() will die() on malformed entries */
3509                 ;
3510 }
3511
3512 static void check_commit(const void *buf, size_t size)
3513 {
3514         struct commit c;
3515         memset(&c, 0, sizeof(c));
3516         if (parse_commit_buffer(&c, buf, size))
3517                 die("corrupt commit");
3518 }
3519
3520 static void check_tag(const void *buf, size_t size)
3521 {
3522         struct tag t;
3523         memset(&t, 0, sizeof(t));
3524         if (parse_tag_buffer(&t, buf, size))
3525                 die("corrupt tag");
3526 }
3527
3528 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3529                      enum object_type type,
3530                      const char *path, unsigned flags)
3531 {
3532         int ret, re_allocated = 0;
3533         int write_object = flags & HASH_WRITE_OBJECT;
3534
3535         if (!type)
3536                 type = OBJ_BLOB;
3537
3538         /*
3539          * Convert blobs to git internal format
3540          */
3541         if ((type == OBJ_BLOB) && path) {
3542                 struct strbuf nbuf = STRBUF_INIT;
3543                 if (convert_to_git(&the_index, path, buf, size, &nbuf,
3544                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3545                         buf = strbuf_detach(&nbuf, &size);
3546                         re_allocated = 1;
3547                 }
3548         }
3549         if (flags & HASH_FORMAT_CHECK) {
3550                 if (type == OBJ_TREE)
3551                         check_tree(buf, size);
3552                 if (type == OBJ_COMMIT)
3553                         check_commit(buf, size);
3554                 if (type == OBJ_TAG)
3555                         check_tag(buf, size);
3556         }
3557
3558         if (write_object)
3559                 ret = write_sha1_file(buf, size, typename(type), sha1);
3560         else
3561                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3562         if (re_allocated)
3563                 free(buf);
3564         return ret;
3565 }
3566
3567 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3568                                      const char *path, unsigned flags)
3569 {
3570         int ret;
3571         const int write_object = flags & HASH_WRITE_OBJECT;
3572         struct strbuf sbuf = STRBUF_INIT;
3573
3574         assert(path);
3575         assert(would_convert_to_git_filter_fd(path));
3576
3577         convert_to_git_filter_fd(&the_index, path, fd, &sbuf,
3578                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3579
3580         if (write_object)
3581                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3582                                       sha1);
3583         else
3584                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3585                                      sha1);
3586         strbuf_release(&sbuf);
3587         return ret;
3588 }
3589
3590 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3591                       const char *path, unsigned flags)
3592 {
3593         struct strbuf sbuf = STRBUF_INIT;
3594         int ret;
3595
3596         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3597                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3598         else
3599                 ret = -1;
3600         strbuf_release(&sbuf);
3601         return ret;
3602 }
3603
3604 #define SMALL_FILE_SIZE (32*1024)
3605
3606 static int index_core(unsigned char *sha1, int fd, size_t size,
3607                       enum object_type type, const char *path,
3608                       unsigned flags)
3609 {
3610         int ret;
3611
3612         if (!size) {
3613                 ret = index_mem(sha1, "", size, type, path, flags);
3614         } else if (size <= SMALL_FILE_SIZE) {
3615                 char *buf = xmalloc(size);
3616                 if (size == read_in_full(fd, buf, size))
3617                         ret = index_mem(sha1, buf, size, type, path, flags);
3618                 else
3619                         ret = error_errno("short read");
3620                 free(buf);
3621         } else {
3622                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3623                 ret = index_mem(sha1, buf, size, type, path, flags);
3624                 munmap(buf, size);
3625         }
3626         return ret;
3627 }
3628
3629 /*
3630  * This creates one packfile per large blob unless bulk-checkin
3631  * machinery is "plugged".
3632  *
3633  * This also bypasses the usual "convert-to-git" dance, and that is on
3634  * purpose. We could write a streaming version of the converting
3635  * functions and insert that before feeding the data to fast-import
3636  * (or equivalent in-core API described above). However, that is
3637  * somewhat complicated, as we do not know the size of the filter
3638  * result, which we need to know beforehand when writing a git object.
3639  * Since the primary motivation for trying to stream from the working
3640  * tree file and to avoid mmaping it in core is to deal with large
3641  * binary blobs, they generally do not want to get any conversion, and
3642  * callers should avoid this code path when filters are requested.
3643  */
3644 static int index_stream(unsigned char *sha1, int fd, size_t size,
3645                         enum object_type type, const char *path,
3646                         unsigned flags)
3647 {
3648         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3649 }
3650
3651 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3652              enum object_type type, const char *path, unsigned flags)
3653 {
3654         int ret;
3655
3656         /*
3657          * Call xsize_t() only when needed to avoid potentially unnecessary
3658          * die() for large files.
3659          */
3660         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3661                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3662         else if (!S_ISREG(st->st_mode))
3663                 ret = index_pipe(sha1, fd, type, path, flags);
3664         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3665                  (path && would_convert_to_git(&the_index, path)))
3666                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3667                                  flags);
3668         else
3669                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3670                                    flags);
3671         close(fd);
3672         return ret;
3673 }
3674
3675 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3676 {
3677         int fd;
3678         struct strbuf sb = STRBUF_INIT;
3679
3680         switch (st->st_mode & S_IFMT) {
3681         case S_IFREG:
3682                 fd = open(path, O_RDONLY);
3683                 if (fd < 0)
3684                         return error_errno("open(\"%s\")", path);
3685                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3686                         return error("%s: failed to insert into database",
3687                                      path);
3688                 break;
3689         case S_IFLNK:
3690                 if (strbuf_readlink(&sb, path, st->st_size))
3691                         return error_errno("readlink(\"%s\")", path);
3692                 if (!(flags & HASH_WRITE_OBJECT))
3693                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3694                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3695                         return error("%s: failed to insert into database",
3696                                      path);
3697                 strbuf_release(&sb);
3698                 break;
3699         case S_IFDIR:
3700                 return resolve_gitlink_ref(path, "HEAD", sha1);
3701         default:
3702                 return error("%s: unsupported file type", path);
3703         }
3704         return 0;
3705 }
3706
3707 int read_pack_header(int fd, struct pack_header *header)
3708 {
3709         if (read_in_full(fd, header, sizeof(*header)) != sizeof(*header))
3710                 /* "eof before pack header was fully read" */
3711                 return PH_ERROR_EOF;
3712
3713         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3714                 /* "protocol error (pack signature mismatch detected)" */
3715                 return PH_ERROR_PACK_SIGNATURE;
3716         if (!pack_version_ok(header->hdr_version))
3717                 /* "protocol error (pack version unsupported)" */
3718                 return PH_ERROR_PROTOCOL;
3719         return 0;
3720 }
3721
3722 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3723 {
3724         enum object_type type = sha1_object_info(sha1, NULL);
3725         if (type < 0)
3726                 die("%s is not a valid object", sha1_to_hex(sha1));
3727         if (type != expect)
3728                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3729                     typename(expect));
3730 }
3731
3732 int for_each_file_in_obj_subdir(unsigned int subdir_nr,
3733                                 struct strbuf *path,
3734                                 each_loose_object_fn obj_cb,
3735                                 each_loose_cruft_fn cruft_cb,
3736                                 each_loose_subdir_fn subdir_cb,
3737                                 void *data)
3738 {
3739         size_t origlen, baselen;
3740         DIR *dir;
3741         struct dirent *de;
3742         int r = 0;
3743
3744         if (subdir_nr > 0xff)
3745                 BUG("invalid loose object subdirectory: %x", subdir_nr);
3746
3747         origlen = path->len;
3748         strbuf_complete(path, '/');
3749         strbuf_addf(path, "%02x", subdir_nr);
3750         baselen = path->len;
3751
3752         dir = opendir(path->buf);
3753         if (!dir) {
3754                 if (errno != ENOENT)
3755                         r = error_errno("unable to open %s", path->buf);
3756                 strbuf_setlen(path, origlen);
3757                 return r;
3758         }
3759
3760         while ((de = readdir(dir))) {
3761                 if (is_dot_or_dotdot(de->d_name))
3762                         continue;
3763
3764                 strbuf_setlen(path, baselen);
3765                 strbuf_addf(path, "/%s", de->d_name);
3766
3767                 if (strlen(de->d_name) == GIT_SHA1_HEXSZ - 2)  {
3768                         char hex[GIT_MAX_HEXSZ+1];
3769                         struct object_id oid;
3770
3771                         xsnprintf(hex, sizeof(hex), "%02x%s",
3772                                   subdir_nr, de->d_name);
3773                         if (!get_oid_hex(hex, &oid)) {
3774                                 if (obj_cb) {
3775                                         r = obj_cb(&oid, path->buf, data);
3776                                         if (r)
3777                                                 break;
3778                                 }
3779                                 continue;
3780                         }
3781                 }
3782
3783                 if (cruft_cb) {
3784                         r = cruft_cb(de->d_name, path->buf, data);
3785                         if (r)
3786                                 break;
3787                 }
3788         }
3789         closedir(dir);
3790
3791         strbuf_setlen(path, baselen);
3792         if (!r && subdir_cb)
3793                 r = subdir_cb(subdir_nr, path->buf, data);
3794
3795         strbuf_setlen(path, origlen);
3796
3797         return r;
3798 }
3799
3800 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3801                             each_loose_object_fn obj_cb,
3802                             each_loose_cruft_fn cruft_cb,
3803                             each_loose_subdir_fn subdir_cb,
3804                             void *data)
3805 {
3806         int r = 0;
3807         int i;
3808
3809         for (i = 0; i < 256; i++) {
3810                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3811                                                 subdir_cb, data);
3812                 if (r)
3813                         break;
3814         }
3815
3816         return r;
3817 }
3818
3819 int for_each_loose_file_in_objdir(const char *path,
3820                                   each_loose_object_fn obj_cb,
3821                                   each_loose_cruft_fn cruft_cb,
3822                                   each_loose_subdir_fn subdir_cb,
3823                                   void *data)
3824 {
3825         struct strbuf buf = STRBUF_INIT;
3826         int r;
3827
3828         strbuf_addstr(&buf, path);
3829         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3830                                               subdir_cb, data);
3831         strbuf_release(&buf);
3832
3833         return r;
3834 }
3835
3836 struct loose_alt_odb_data {
3837         each_loose_object_fn *cb;
3838         void *data;
3839 };
3840
3841 static int loose_from_alt_odb(struct alternate_object_database *alt,
3842                               void *vdata)
3843 {
3844         struct loose_alt_odb_data *data = vdata;
3845         struct strbuf buf = STRBUF_INIT;
3846         int r;
3847
3848         strbuf_addstr(&buf, alt->path);
3849         r = for_each_loose_file_in_objdir_buf(&buf,
3850                                               data->cb, NULL, NULL,
3851                                               data->data);
3852         strbuf_release(&buf);
3853         return r;
3854 }
3855
3856 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3857 {
3858         struct loose_alt_odb_data alt;
3859         int r;
3860
3861         r = for_each_loose_file_in_objdir(get_object_directory(),
3862                                           cb, NULL, NULL, data);
3863         if (r)
3864                 return r;
3865
3866         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3867                 return 0;
3868
3869         alt.cb = cb;
3870         alt.data = data;
3871         return foreach_alt_odb(loose_from_alt_odb, &alt);
3872 }
3873
3874 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3875 {
3876         uint32_t i;
3877         int r = 0;
3878
3879         for (i = 0; i < p->num_objects; i++) {
3880                 struct object_id oid;
3881
3882                 if (!nth_packed_object_oid(&oid, p, i))
3883                         return error("unable to get sha1 of object %u in %s",
3884                                      i, p->pack_name);
3885
3886                 r = cb(&oid, p, i, data);
3887                 if (r)
3888                         break;
3889         }
3890         return r;
3891 }
3892
3893 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3894 {
3895         struct packed_git *p;
3896         int r = 0;
3897         int pack_errors = 0;
3898
3899         prepare_packed_git();
3900         for (p = packed_git; p; p = p->next) {
3901                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3902                         continue;
3903                 if (open_pack_index(p)) {
3904                         pack_errors = 1;
3905                         continue;
3906                 }
3907                 r = for_each_object_in_pack(p, cb, data);
3908                 if (r)
3909                         break;
3910         }
3911         return r ? r : pack_errors;
3912 }
3913
3914 static int check_stream_sha1(git_zstream *stream,
3915                              const char *hdr,
3916                              unsigned long size,
3917                              const char *path,
3918                              const unsigned char *expected_sha1)
3919 {
3920         git_SHA_CTX c;
3921         unsigned char real_sha1[GIT_MAX_RAWSZ];
3922         unsigned char buf[4096];
3923         unsigned long total_read;
3924         int status = Z_OK;
3925
3926         git_SHA1_Init(&c);
3927         git_SHA1_Update(&c, hdr, stream->total_out);
3928
3929         /*
3930          * We already read some bytes into hdr, but the ones up to the NUL
3931          * do not count against the object's content size.
3932          */
3933         total_read = stream->total_out - strlen(hdr) - 1;
3934
3935         /*
3936          * This size comparison must be "<=" to read the final zlib packets;
3937          * see the comment in unpack_sha1_rest for details.
3938          */
3939         while (total_read <= size &&
3940                (status == Z_OK || status == Z_BUF_ERROR)) {
3941                 stream->next_out = buf;
3942                 stream->avail_out = sizeof(buf);
3943                 if (size - total_read < stream->avail_out)
3944                         stream->avail_out = size - total_read;
3945                 status = git_inflate(stream, Z_FINISH);
3946                 git_SHA1_Update(&c, buf, stream->next_out - buf);
3947                 total_read += stream->next_out - buf;
3948         }
3949         git_inflate_end(stream);
3950
3951         if (status != Z_STREAM_END) {
3952                 error("corrupt loose object '%s'", sha1_to_hex(expected_sha1));
3953                 return -1;
3954         }
3955         if (stream->avail_in) {
3956                 error("garbage at end of loose object '%s'",
3957                       sha1_to_hex(expected_sha1));
3958                 return -1;
3959         }
3960
3961         git_SHA1_Final(real_sha1, &c);
3962         if (hashcmp(expected_sha1, real_sha1)) {
3963                 error("sha1 mismatch for %s (expected %s)", path,
3964                       sha1_to_hex(expected_sha1));
3965                 return -1;
3966         }
3967
3968         return 0;
3969 }
3970
3971 int read_loose_object(const char *path,
3972                       const unsigned char *expected_sha1,
3973                       enum object_type *type,
3974                       unsigned long *size,
3975                       void **contents)
3976 {
3977         int ret = -1;
3978         void *map = NULL;
3979         unsigned long mapsize;
3980         git_zstream stream;
3981         char hdr[32];
3982
3983         *contents = NULL;
3984
3985         map = map_sha1_file_1(path, NULL, &mapsize);
3986         if (!map) {
3987                 error_errno("unable to mmap %s", path);
3988                 goto out;
3989         }
3990
3991         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0) {
3992                 error("unable to unpack header of %s", path);
3993                 goto out;
3994         }
3995
3996         *type = parse_sha1_header(hdr, size);
3997         if (*type < 0) {
3998                 error("unable to parse header of %s", path);
3999                 git_inflate_end(&stream);
4000                 goto out;
4001         }
4002
4003         if (*type == OBJ_BLOB) {
4004                 if (check_stream_sha1(&stream, hdr, *size, path, expected_sha1) < 0)
4005                         goto out;
4006         } else {
4007                 *contents = unpack_sha1_rest(&stream, hdr, *size, expected_sha1);
4008                 if (!*contents) {
4009                         error("unable to unpack contents of %s", path);
4010                         git_inflate_end(&stream);
4011                         goto out;
4012                 }
4013                 if (check_sha1_signature(expected_sha1, *contents,
4014                                          *size, typename(*type))) {
4015                         error("sha1 mismatch for %s (expected %s)", path,
4016                               sha1_to_hex(expected_sha1));
4017                         free(*contents);
4018                         goto out;
4019                 }
4020         }
4021
4022         ret = 0; /* everything checks out */
4023
4024 out:
4025         if (map)
4026                 munmap(map, mapsize);
4027         return ret;
4028 }