Merge branch 'mr/vcs-svn-printf-ulong'
[git] / sha1_file.c
1 /*
2  * GIT - The information manager from hell
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This handles basic git sha1 object files - packing, unpacking,
7  * creation etc.
8  */
9 #include "cache.h"
10 #include "string-list.h"
11 #include "lockfile.h"
12 #include "delta.h"
13 #include "pack.h"
14 #include "blob.h"
15 #include "commit.h"
16 #include "run-command.h"
17 #include "tag.h"
18 #include "tree.h"
19 #include "tree-walk.h"
20 #include "refs.h"
21 #include "pack-revindex.h"
22 #include "sha1-lookup.h"
23 #include "bulk-checkin.h"
24 #include "streaming.h"
25 #include "dir.h"
26 #include "mru.h"
27 #include "list.h"
28 #include "mergesort.h"
29
30 #ifndef O_NOATIME
31 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
32 #define O_NOATIME 01000000
33 #else
34 #define O_NOATIME 0
35 #endif
36 #endif
37
38 #define SZ_FMT PRIuMAX
39 static inline uintmax_t sz_fmt(size_t s) { return s; }
40
41 const unsigned char null_sha1[20];
42 const struct object_id null_oid;
43 const struct object_id empty_tree_oid = {
44         EMPTY_TREE_SHA1_BIN_LITERAL
45 };
46 const struct object_id empty_blob_oid = {
47         EMPTY_BLOB_SHA1_BIN_LITERAL
48 };
49
50 /*
51  * This is meant to hold a *small* number of objects that you would
52  * want read_sha1_file() to be able to return, but yet you do not want
53  * to write them into the object store (e.g. a browse-only
54  * application).
55  */
56 static struct cached_object {
57         unsigned char sha1[20];
58         enum object_type type;
59         void *buf;
60         unsigned long size;
61 } *cached_objects;
62 static int cached_object_nr, cached_object_alloc;
63
64 static struct cached_object empty_tree = {
65         EMPTY_TREE_SHA1_BIN_LITERAL,
66         OBJ_TREE,
67         "",
68         0
69 };
70
71 static struct cached_object *find_cached_object(const unsigned char *sha1)
72 {
73         int i;
74         struct cached_object *co = cached_objects;
75
76         for (i = 0; i < cached_object_nr; i++, co++) {
77                 if (!hashcmp(co->sha1, sha1))
78                         return co;
79         }
80         if (!hashcmp(sha1, empty_tree.sha1))
81                 return &empty_tree;
82         return NULL;
83 }
84
85 int mkdir_in_gitdir(const char *path)
86 {
87         if (mkdir(path, 0777)) {
88                 int saved_errno = errno;
89                 struct stat st;
90                 struct strbuf sb = STRBUF_INIT;
91
92                 if (errno != EEXIST)
93                         return -1;
94                 /*
95                  * Are we looking at a path in a symlinked worktree
96                  * whose original repository does not yet have it?
97                  * e.g. .git/rr-cache pointing at its original
98                  * repository in which the user hasn't performed any
99                  * conflict resolution yet?
100                  */
101                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
102                     strbuf_readlink(&sb, path, st.st_size) ||
103                     !is_absolute_path(sb.buf) ||
104                     mkdir(sb.buf, 0777)) {
105                         strbuf_release(&sb);
106                         errno = saved_errno;
107                         return -1;
108                 }
109                 strbuf_release(&sb);
110         }
111         return adjust_shared_perm(path);
112 }
113
114 enum scld_error safe_create_leading_directories(char *path)
115 {
116         char *next_component = path + offset_1st_component(path);
117         enum scld_error ret = SCLD_OK;
118
119         while (ret == SCLD_OK && next_component) {
120                 struct stat st;
121                 char *slash = next_component, slash_character;
122
123                 while (*slash && !is_dir_sep(*slash))
124                         slash++;
125
126                 if (!*slash)
127                         break;
128
129                 next_component = slash + 1;
130                 while (is_dir_sep(*next_component))
131                         next_component++;
132                 if (!*next_component)
133                         break;
134
135                 slash_character = *slash;
136                 *slash = '\0';
137                 if (!stat(path, &st)) {
138                         /* path exists */
139                         if (!S_ISDIR(st.st_mode))
140                                 ret = SCLD_EXISTS;
141                 } else if (mkdir(path, 0777)) {
142                         if (errno == EEXIST &&
143                             !stat(path, &st) && S_ISDIR(st.st_mode))
144                                 ; /* somebody created it since we checked */
145                         else if (errno == ENOENT)
146                                 /*
147                                  * Either mkdir() failed because
148                                  * somebody just pruned the containing
149                                  * directory, or stat() failed because
150                                  * the file that was in our way was
151                                  * just removed.  Either way, inform
152                                  * the caller that it might be worth
153                                  * trying again:
154                                  */
155                                 ret = SCLD_VANISHED;
156                         else
157                                 ret = SCLD_FAILED;
158                 } else if (adjust_shared_perm(path)) {
159                         ret = SCLD_PERMS;
160                 }
161                 *slash = slash_character;
162         }
163         return ret;
164 }
165
166 enum scld_error safe_create_leading_directories_const(const char *path)
167 {
168         /* path points to cache entries, so xstrdup before messing with it */
169         char *buf = xstrdup(path);
170         enum scld_error result = safe_create_leading_directories(buf);
171         free(buf);
172         return result;
173 }
174
175 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
176 {
177         int i;
178         for (i = 0; i < 20; i++) {
179                 static char hex[] = "0123456789abcdef";
180                 unsigned int val = sha1[i];
181                 char *pos = pathbuf + i*2 + (i > 0);
182                 *pos++ = hex[val >> 4];
183                 *pos = hex[val & 0xf];
184         }
185 }
186
187 const char *sha1_file_name(const unsigned char *sha1)
188 {
189         static char buf[PATH_MAX];
190         const char *objdir;
191         int len;
192
193         objdir = get_object_directory();
194         len = strlen(objdir);
195
196         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
197         if (len + 43 > PATH_MAX)
198                 die("insanely long object directory %s", objdir);
199         memcpy(buf, objdir, len);
200         buf[len] = '/';
201         buf[len+3] = '/';
202         buf[len+42] = '\0';
203         fill_sha1_path(buf + len + 1, sha1);
204         return buf;
205 }
206
207 /*
208  * Return the name of the pack or index file with the specified sha1
209  * in its filename.  *base and *name are scratch space that must be
210  * provided by the caller.  which should be "pack" or "idx".
211  */
212 static char *sha1_get_pack_name(const unsigned char *sha1,
213                                 struct strbuf *buf,
214                                 const char *which)
215 {
216         strbuf_reset(buf);
217         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
218                     sha1_to_hex(sha1), which);
219         return buf->buf;
220 }
221
222 char *sha1_pack_name(const unsigned char *sha1)
223 {
224         static struct strbuf buf = STRBUF_INIT;
225         return sha1_get_pack_name(sha1, &buf, "pack");
226 }
227
228 char *sha1_pack_index_name(const unsigned char *sha1)
229 {
230         static struct strbuf buf = STRBUF_INIT;
231         return sha1_get_pack_name(sha1, &buf, "idx");
232 }
233
234 struct alternate_object_database *alt_odb_list;
235 static struct alternate_object_database **alt_odb_tail;
236
237 /*
238  * Prepare alternate object database registry.
239  *
240  * The variable alt_odb_list points at the list of struct
241  * alternate_object_database.  The elements on this list come from
242  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
243  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
244  * whose contents is similar to that environment variable but can be
245  * LF separated.  Its base points at a statically allocated buffer that
246  * contains "/the/directory/corresponding/to/.git/objects/...", while
247  * its name points just after the slash at the end of ".git/objects/"
248  * in the example above, and has enough space to hold 40-byte hex
249  * SHA1, an extra slash for the first level indirection, and the
250  * terminating NUL.
251  */
252 static int link_alt_odb_entry(const char *entry, const char *relative_base,
253         int depth, const char *normalized_objdir)
254 {
255         struct alternate_object_database *ent;
256         struct alternate_object_database *alt;
257         size_t pfxlen, entlen;
258         struct strbuf pathbuf = STRBUF_INIT;
259
260         if (!is_absolute_path(entry) && relative_base) {
261                 strbuf_addstr(&pathbuf, real_path(relative_base));
262                 strbuf_addch(&pathbuf, '/');
263         }
264         strbuf_addstr(&pathbuf, entry);
265
266         normalize_path_copy(pathbuf.buf, pathbuf.buf);
267
268         pfxlen = strlen(pathbuf.buf);
269
270         /*
271          * The trailing slash after the directory name is given by
272          * this function at the end. Remove duplicates.
273          */
274         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
275                 pfxlen -= 1;
276
277         entlen = st_add(pfxlen, 43); /* '/' + 2 hex + '/' + 38 hex + NUL */
278         ent = xmalloc(st_add(sizeof(*ent), entlen));
279         memcpy(ent->base, pathbuf.buf, pfxlen);
280         strbuf_release(&pathbuf);
281
282         ent->name = ent->base + pfxlen + 1;
283         ent->base[pfxlen + 3] = '/';
284         ent->base[pfxlen] = ent->base[entlen-1] = 0;
285
286         /* Detect cases where alternate disappeared */
287         if (!is_directory(ent->base)) {
288                 error("object directory %s does not exist; "
289                       "check .git/objects/info/alternates.",
290                       ent->base);
291                 free(ent);
292                 return -1;
293         }
294
295         /* Prevent the common mistake of listing the same
296          * thing twice, or object directory itself.
297          */
298         for (alt = alt_odb_list; alt; alt = alt->next) {
299                 if (pfxlen == alt->name - alt->base - 1 &&
300                     !memcmp(ent->base, alt->base, pfxlen)) {
301                         free(ent);
302                         return -1;
303                 }
304         }
305         if (!fspathcmp(ent->base, normalized_objdir)) {
306                 free(ent);
307                 return -1;
308         }
309
310         /* add the alternate entry */
311         *alt_odb_tail = ent;
312         alt_odb_tail = &(ent->next);
313         ent->next = NULL;
314
315         /* recursively add alternates */
316         read_info_alternates(ent->base, depth + 1);
317
318         ent->base[pfxlen] = '/';
319
320         return 0;
321 }
322
323 static void link_alt_odb_entries(const char *alt, int len, int sep,
324                                  const char *relative_base, int depth)
325 {
326         struct string_list entries = STRING_LIST_INIT_NODUP;
327         char *alt_copy;
328         int i;
329         struct strbuf objdirbuf = STRBUF_INIT;
330
331         if (depth > 5) {
332                 error("%s: ignoring alternate object stores, nesting too deep.",
333                                 relative_base);
334                 return;
335         }
336
337         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
338         normalize_path_copy(objdirbuf.buf, objdirbuf.buf);
339
340         alt_copy = xmemdupz(alt, len);
341         string_list_split_in_place(&entries, alt_copy, sep, -1);
342         for (i = 0; i < entries.nr; i++) {
343                 const char *entry = entries.items[i].string;
344                 if (entry[0] == '\0' || entry[0] == '#')
345                         continue;
346                 if (!is_absolute_path(entry) && depth) {
347                         error("%s: ignoring relative alternate object store %s",
348                                         relative_base, entry);
349                 } else {
350                         link_alt_odb_entry(entry, relative_base, depth, objdirbuf.buf);
351                 }
352         }
353         string_list_clear(&entries, 0);
354         free(alt_copy);
355         strbuf_release(&objdirbuf);
356 }
357
358 void read_info_alternates(const char * relative_base, int depth)
359 {
360         char *map;
361         size_t mapsz;
362         struct stat st;
363         char *path;
364         int fd;
365
366         path = xstrfmt("%s/info/alternates", relative_base);
367         fd = git_open_noatime(path);
368         free(path);
369         if (fd < 0)
370                 return;
371         if (fstat(fd, &st) || (st.st_size == 0)) {
372                 close(fd);
373                 return;
374         }
375         mapsz = xsize_t(st.st_size);
376         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
377         close(fd);
378
379         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
380
381         munmap(map, mapsz);
382 }
383
384 void add_to_alternates_file(const char *reference)
385 {
386         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
387         char *alts = git_pathdup("objects/info/alternates");
388         FILE *in, *out;
389
390         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
391         out = fdopen_lock_file(lock, "w");
392         if (!out)
393                 die_errno("unable to fdopen alternates lockfile");
394
395         in = fopen(alts, "r");
396         if (in) {
397                 struct strbuf line = STRBUF_INIT;
398                 int found = 0;
399
400                 while (strbuf_getline(&line, in) != EOF) {
401                         if (!strcmp(reference, line.buf)) {
402                                 found = 1;
403                                 break;
404                         }
405                         fprintf_or_die(out, "%s\n", line.buf);
406                 }
407
408                 strbuf_release(&line);
409                 fclose(in);
410
411                 if (found) {
412                         rollback_lock_file(lock);
413                         lock = NULL;
414                 }
415         }
416         else if (errno != ENOENT)
417                 die_errno("unable to read alternates file");
418
419         if (lock) {
420                 fprintf_or_die(out, "%s\n", reference);
421                 if (commit_lock_file(lock))
422                         die_errno("unable to move new alternates file into place");
423                 if (alt_odb_tail)
424                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
425         }
426         free(alts);
427 }
428
429 /*
430  * Compute the exact path an alternate is at and returns it. In case of
431  * error NULL is returned and the human readable error is added to `err`
432  * `path` may be relative and should point to $GITDIR.
433  * `err` must not be null.
434  */
435 char *compute_alternate_path(const char *path, struct strbuf *err)
436 {
437         char *ref_git = NULL;
438         const char *repo, *ref_git_s;
439         int seen_error = 0;
440
441         ref_git_s = real_path_if_valid(path);
442         if (!ref_git_s) {
443                 seen_error = 1;
444                 strbuf_addf(err, _("path '%s' does not exist"), path);
445                 goto out;
446         } else
447                 /*
448                  * Beware: read_gitfile(), real_path() and mkpath()
449                  * return static buffer
450                  */
451                 ref_git = xstrdup(ref_git_s);
452
453         repo = read_gitfile(ref_git);
454         if (!repo)
455                 repo = read_gitfile(mkpath("%s/.git", ref_git));
456         if (repo) {
457                 free(ref_git);
458                 ref_git = xstrdup(repo);
459         }
460
461         if (!repo && is_directory(mkpath("%s/.git/objects", ref_git))) {
462                 char *ref_git_git = mkpathdup("%s/.git", ref_git);
463                 free(ref_git);
464                 ref_git = ref_git_git;
465         } else if (!is_directory(mkpath("%s/objects", ref_git))) {
466                 struct strbuf sb = STRBUF_INIT;
467                 seen_error = 1;
468                 if (get_common_dir(&sb, ref_git)) {
469                         strbuf_addf(err,
470                                     _("reference repository '%s' as a linked "
471                                       "checkout is not supported yet."),
472                                     path);
473                         goto out;
474                 }
475
476                 strbuf_addf(err, _("reference repository '%s' is not a "
477                                         "local repository."), path);
478                 goto out;
479         }
480
481         if (!access(mkpath("%s/shallow", ref_git), F_OK)) {
482                 strbuf_addf(err, _("reference repository '%s' is shallow"),
483                             path);
484                 seen_error = 1;
485                 goto out;
486         }
487
488         if (!access(mkpath("%s/info/grafts", ref_git), F_OK)) {
489                 strbuf_addf(err,
490                             _("reference repository '%s' is grafted"),
491                             path);
492                 seen_error = 1;
493                 goto out;
494         }
495
496 out:
497         if (seen_error) {
498                 free(ref_git);
499                 ref_git = NULL;
500         }
501
502         return ref_git;
503 }
504
505 int foreach_alt_odb(alt_odb_fn fn, void *cb)
506 {
507         struct alternate_object_database *ent;
508         int r = 0;
509
510         prepare_alt_odb();
511         for (ent = alt_odb_list; ent; ent = ent->next) {
512                 r = fn(ent, cb);
513                 if (r)
514                         break;
515         }
516         return r;
517 }
518
519 void prepare_alt_odb(void)
520 {
521         const char *alt;
522
523         if (alt_odb_tail)
524                 return;
525
526         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
527         if (!alt) alt = "";
528
529         alt_odb_tail = &alt_odb_list;
530         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
531
532         read_info_alternates(get_object_directory(), 0);
533 }
534
535 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
536 static int freshen_file(const char *fn)
537 {
538         struct utimbuf t;
539         t.actime = t.modtime = time(NULL);
540         return !utime(fn, &t);
541 }
542
543 /*
544  * All of the check_and_freshen functions return 1 if the file exists and was
545  * freshened (if freshening was requested), 0 otherwise. If they return
546  * 0, you should not assume that it is safe to skip a write of the object (it
547  * either does not exist on disk, or has a stale mtime and may be subject to
548  * pruning).
549  */
550 static int check_and_freshen_file(const char *fn, int freshen)
551 {
552         if (access(fn, F_OK))
553                 return 0;
554         if (freshen && !freshen_file(fn))
555                 return 0;
556         return 1;
557 }
558
559 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
560 {
561         return check_and_freshen_file(sha1_file_name(sha1), freshen);
562 }
563
564 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
565 {
566         struct alternate_object_database *alt;
567         prepare_alt_odb();
568         for (alt = alt_odb_list; alt; alt = alt->next) {
569                 fill_sha1_path(alt->name, sha1);
570                 if (check_and_freshen_file(alt->base, freshen))
571                         return 1;
572         }
573         return 0;
574 }
575
576 static int check_and_freshen(const unsigned char *sha1, int freshen)
577 {
578         return check_and_freshen_local(sha1, freshen) ||
579                check_and_freshen_nonlocal(sha1, freshen);
580 }
581
582 int has_loose_object_nonlocal(const unsigned char *sha1)
583 {
584         return check_and_freshen_nonlocal(sha1, 0);
585 }
586
587 static int has_loose_object(const unsigned char *sha1)
588 {
589         return check_and_freshen(sha1, 0);
590 }
591
592 static unsigned int pack_used_ctr;
593 static unsigned int pack_mmap_calls;
594 static unsigned int peak_pack_open_windows;
595 static unsigned int pack_open_windows;
596 static unsigned int pack_open_fds;
597 static unsigned int pack_max_fds;
598 static size_t peak_pack_mapped;
599 static size_t pack_mapped;
600 struct packed_git *packed_git;
601
602 static struct mru packed_git_mru_storage;
603 struct mru *packed_git_mru = &packed_git_mru_storage;
604
605 void pack_report(void)
606 {
607         fprintf(stderr,
608                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
609                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
610                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
611                 sz_fmt(getpagesize()),
612                 sz_fmt(packed_git_window_size),
613                 sz_fmt(packed_git_limit));
614         fprintf(stderr,
615                 "pack_report: pack_used_ctr            = %10u\n"
616                 "pack_report: pack_mmap_calls          = %10u\n"
617                 "pack_report: pack_open_windows        = %10u / %10u\n"
618                 "pack_report: pack_mapped              = "
619                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
620                 pack_used_ctr,
621                 pack_mmap_calls,
622                 pack_open_windows, peak_pack_open_windows,
623                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
624 }
625
626 /*
627  * Open and mmap the index file at path, perform a couple of
628  * consistency checks, then record its information to p.  Return 0 on
629  * success.
630  */
631 static int check_packed_git_idx(const char *path, struct packed_git *p)
632 {
633         void *idx_map;
634         struct pack_idx_header *hdr;
635         size_t idx_size;
636         uint32_t version, nr, i, *index;
637         int fd = git_open_noatime(path);
638         struct stat st;
639
640         if (fd < 0)
641                 return -1;
642         if (fstat(fd, &st)) {
643                 close(fd);
644                 return -1;
645         }
646         idx_size = xsize_t(st.st_size);
647         if (idx_size < 4 * 256 + 20 + 20) {
648                 close(fd);
649                 return error("index file %s is too small", path);
650         }
651         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
652         close(fd);
653
654         hdr = idx_map;
655         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
656                 version = ntohl(hdr->idx_version);
657                 if (version < 2 || version > 2) {
658                         munmap(idx_map, idx_size);
659                         return error("index file %s is version %"PRIu32
660                                      " and is not supported by this binary"
661                                      " (try upgrading GIT to a newer version)",
662                                      path, version);
663                 }
664         } else
665                 version = 1;
666
667         nr = 0;
668         index = idx_map;
669         if (version > 1)
670                 index += 2;  /* skip index header */
671         for (i = 0; i < 256; i++) {
672                 uint32_t n = ntohl(index[i]);
673                 if (n < nr) {
674                         munmap(idx_map, idx_size);
675                         return error("non-monotonic index %s", path);
676                 }
677                 nr = n;
678         }
679
680         if (version == 1) {
681                 /*
682                  * Total size:
683                  *  - 256 index entries 4 bytes each
684                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
685                  *  - 20-byte SHA1 of the packfile
686                  *  - 20-byte SHA1 file checksum
687                  */
688                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
689                         munmap(idx_map, idx_size);
690                         return error("wrong index v1 file size in %s", path);
691                 }
692         } else if (version == 2) {
693                 /*
694                  * Minimum size:
695                  *  - 8 bytes of header
696                  *  - 256 index entries 4 bytes each
697                  *  - 20-byte sha1 entry * nr
698                  *  - 4-byte crc entry * nr
699                  *  - 4-byte offset entry * nr
700                  *  - 20-byte SHA1 of the packfile
701                  *  - 20-byte SHA1 file checksum
702                  * And after the 4-byte offset table might be a
703                  * variable sized table containing 8-byte entries
704                  * for offsets larger than 2^31.
705                  */
706                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
707                 unsigned long max_size = min_size;
708                 if (nr)
709                         max_size += (nr - 1)*8;
710                 if (idx_size < min_size || idx_size > max_size) {
711                         munmap(idx_map, idx_size);
712                         return error("wrong index v2 file size in %s", path);
713                 }
714                 if (idx_size != min_size &&
715                     /*
716                      * make sure we can deal with large pack offsets.
717                      * 31-bit signed offset won't be enough, neither
718                      * 32-bit unsigned one will be.
719                      */
720                     (sizeof(off_t) <= 4)) {
721                         munmap(idx_map, idx_size);
722                         return error("pack too large for current definition of off_t in %s", path);
723                 }
724         }
725
726         p->index_version = version;
727         p->index_data = idx_map;
728         p->index_size = idx_size;
729         p->num_objects = nr;
730         return 0;
731 }
732
733 int open_pack_index(struct packed_git *p)
734 {
735         char *idx_name;
736         size_t len;
737         int ret;
738
739         if (p->index_data)
740                 return 0;
741
742         if (!strip_suffix(p->pack_name, ".pack", &len))
743                 die("BUG: pack_name does not end in .pack");
744         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
745         ret = check_packed_git_idx(idx_name, p);
746         free(idx_name);
747         return ret;
748 }
749
750 static void scan_windows(struct packed_git *p,
751         struct packed_git **lru_p,
752         struct pack_window **lru_w,
753         struct pack_window **lru_l)
754 {
755         struct pack_window *w, *w_l;
756
757         for (w_l = NULL, w = p->windows; w; w = w->next) {
758                 if (!w->inuse_cnt) {
759                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
760                                 *lru_p = p;
761                                 *lru_w = w;
762                                 *lru_l = w_l;
763                         }
764                 }
765                 w_l = w;
766         }
767 }
768
769 static int unuse_one_window(struct packed_git *current)
770 {
771         struct packed_git *p, *lru_p = NULL;
772         struct pack_window *lru_w = NULL, *lru_l = NULL;
773
774         if (current)
775                 scan_windows(current, &lru_p, &lru_w, &lru_l);
776         for (p = packed_git; p; p = p->next)
777                 scan_windows(p, &lru_p, &lru_w, &lru_l);
778         if (lru_p) {
779                 munmap(lru_w->base, lru_w->len);
780                 pack_mapped -= lru_w->len;
781                 if (lru_l)
782                         lru_l->next = lru_w->next;
783                 else
784                         lru_p->windows = lru_w->next;
785                 free(lru_w);
786                 pack_open_windows--;
787                 return 1;
788         }
789         return 0;
790 }
791
792 void release_pack_memory(size_t need)
793 {
794         size_t cur = pack_mapped;
795         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
796                 ; /* nothing */
797 }
798
799 static void mmap_limit_check(size_t length)
800 {
801         static size_t limit = 0;
802         if (!limit) {
803                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
804                 if (!limit)
805                         limit = SIZE_MAX;
806         }
807         if (length > limit)
808                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
809                     (uintmax_t)length, (uintmax_t)limit);
810 }
811
812 void *xmmap_gently(void *start, size_t length,
813                   int prot, int flags, int fd, off_t offset)
814 {
815         void *ret;
816
817         mmap_limit_check(length);
818         ret = mmap(start, length, prot, flags, fd, offset);
819         if (ret == MAP_FAILED) {
820                 if (!length)
821                         return NULL;
822                 release_pack_memory(length);
823                 ret = mmap(start, length, prot, flags, fd, offset);
824         }
825         return ret;
826 }
827
828 void *xmmap(void *start, size_t length,
829         int prot, int flags, int fd, off_t offset)
830 {
831         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
832         if (ret == MAP_FAILED)
833                 die_errno("mmap failed");
834         return ret;
835 }
836
837 void close_pack_windows(struct packed_git *p)
838 {
839         while (p->windows) {
840                 struct pack_window *w = p->windows;
841
842                 if (w->inuse_cnt)
843                         die("pack '%s' still has open windows to it",
844                             p->pack_name);
845                 munmap(w->base, w->len);
846                 pack_mapped -= w->len;
847                 pack_open_windows--;
848                 p->windows = w->next;
849                 free(w);
850         }
851 }
852
853 static int close_pack_fd(struct packed_git *p)
854 {
855         if (p->pack_fd < 0)
856                 return 0;
857
858         close(p->pack_fd);
859         pack_open_fds--;
860         p->pack_fd = -1;
861
862         return 1;
863 }
864
865 static void close_pack(struct packed_git *p)
866 {
867         close_pack_windows(p);
868         close_pack_fd(p);
869         close_pack_index(p);
870 }
871
872 void close_all_packs(void)
873 {
874         struct packed_git *p;
875
876         for (p = packed_git; p; p = p->next)
877                 if (p->do_not_close)
878                         die("BUG: want to close pack marked 'do-not-close'");
879                 else
880                         close_pack(p);
881 }
882
883
884 /*
885  * The LRU pack is the one with the oldest MRU window, preferring packs
886  * with no used windows, or the oldest mtime if it has no windows allocated.
887  */
888 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
889 {
890         struct pack_window *w, *this_mru_w;
891         int has_windows_inuse = 0;
892
893         /*
894          * Reject this pack if it has windows and the previously selected
895          * one does not.  If this pack does not have windows, reject
896          * it if the pack file is newer than the previously selected one.
897          */
898         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
899                 return;
900
901         for (w = this_mru_w = p->windows; w; w = w->next) {
902                 /*
903                  * Reject this pack if any of its windows are in use,
904                  * but the previously selected pack did not have any
905                  * inuse windows.  Otherwise, record that this pack
906                  * has windows in use.
907                  */
908                 if (w->inuse_cnt) {
909                         if (*accept_windows_inuse)
910                                 has_windows_inuse = 1;
911                         else
912                                 return;
913                 }
914
915                 if (w->last_used > this_mru_w->last_used)
916                         this_mru_w = w;
917
918                 /*
919                  * Reject this pack if it has windows that have been
920                  * used more recently than the previously selected pack.
921                  * If the previously selected pack had windows inuse and
922                  * we have not encountered a window in this pack that is
923                  * inuse, skip this check since we prefer a pack with no
924                  * inuse windows to one that has inuse windows.
925                  */
926                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
927                     this_mru_w->last_used > (*mru_w)->last_used)
928                         return;
929         }
930
931         /*
932          * Select this pack.
933          */
934         *mru_w = this_mru_w;
935         *lru_p = p;
936         *accept_windows_inuse = has_windows_inuse;
937 }
938
939 static int close_one_pack(void)
940 {
941         struct packed_git *p, *lru_p = NULL;
942         struct pack_window *mru_w = NULL;
943         int accept_windows_inuse = 1;
944
945         for (p = packed_git; p; p = p->next) {
946                 if (p->pack_fd == -1)
947                         continue;
948                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
949         }
950
951         if (lru_p)
952                 return close_pack_fd(lru_p);
953
954         return 0;
955 }
956
957 void unuse_pack(struct pack_window **w_cursor)
958 {
959         struct pack_window *w = *w_cursor;
960         if (w) {
961                 w->inuse_cnt--;
962                 *w_cursor = NULL;
963         }
964 }
965
966 void close_pack_index(struct packed_git *p)
967 {
968         if (p->index_data) {
969                 munmap((void *)p->index_data, p->index_size);
970                 p->index_data = NULL;
971         }
972 }
973
974 static unsigned int get_max_fd_limit(void)
975 {
976 #ifdef RLIMIT_NOFILE
977         {
978                 struct rlimit lim;
979
980                 if (!getrlimit(RLIMIT_NOFILE, &lim))
981                         return lim.rlim_cur;
982         }
983 #endif
984
985 #ifdef _SC_OPEN_MAX
986         {
987                 long open_max = sysconf(_SC_OPEN_MAX);
988                 if (0 < open_max)
989                         return open_max;
990                 /*
991                  * Otherwise, we got -1 for one of the two
992                  * reasons:
993                  *
994                  * (1) sysconf() did not understand _SC_OPEN_MAX
995                  *     and signaled an error with -1; or
996                  * (2) sysconf() said there is no limit.
997                  *
998                  * We _could_ clear errno before calling sysconf() to
999                  * tell these two cases apart and return a huge number
1000                  * in the latter case to let the caller cap it to a
1001                  * value that is not so selfish, but letting the
1002                  * fallback OPEN_MAX codepath take care of these cases
1003                  * is a lot simpler.
1004                  */
1005         }
1006 #endif
1007
1008 #ifdef OPEN_MAX
1009         return OPEN_MAX;
1010 #else
1011         return 1; /* see the caller ;-) */
1012 #endif
1013 }
1014
1015 /*
1016  * Do not call this directly as this leaks p->pack_fd on error return;
1017  * call open_packed_git() instead.
1018  */
1019 static int open_packed_git_1(struct packed_git *p)
1020 {
1021         struct stat st;
1022         struct pack_header hdr;
1023         unsigned char sha1[20];
1024         unsigned char *idx_sha1;
1025         long fd_flag;
1026
1027         if (!p->index_data && open_pack_index(p))
1028                 return error("packfile %s index unavailable", p->pack_name);
1029
1030         if (!pack_max_fds) {
1031                 unsigned int max_fds = get_max_fd_limit();
1032
1033                 /* Save 3 for stdin/stdout/stderr, 22 for work */
1034                 if (25 < max_fds)
1035                         pack_max_fds = max_fds - 25;
1036                 else
1037                         pack_max_fds = 1;
1038         }
1039
1040         while (pack_max_fds <= pack_open_fds && close_one_pack())
1041                 ; /* nothing */
1042
1043         p->pack_fd = git_open_noatime(p->pack_name);
1044         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
1045                 return -1;
1046         pack_open_fds++;
1047
1048         /* If we created the struct before we had the pack we lack size. */
1049         if (!p->pack_size) {
1050                 if (!S_ISREG(st.st_mode))
1051                         return error("packfile %s not a regular file", p->pack_name);
1052                 p->pack_size = st.st_size;
1053         } else if (p->pack_size != st.st_size)
1054                 return error("packfile %s size changed", p->pack_name);
1055
1056         /* We leave these file descriptors open with sliding mmap;
1057          * there is no point keeping them open across exec(), though.
1058          */
1059         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1060         if (fd_flag < 0)
1061                 return error("cannot determine file descriptor flags");
1062         fd_flag |= FD_CLOEXEC;
1063         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1064                 return error("cannot set FD_CLOEXEC");
1065
1066         /* Verify we recognize this pack file format. */
1067         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1068                 return error("file %s is far too short to be a packfile", p->pack_name);
1069         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1070                 return error("file %s is not a GIT packfile", p->pack_name);
1071         if (!pack_version_ok(hdr.hdr_version))
1072                 return error("packfile %s is version %"PRIu32" and not"
1073                         " supported (try upgrading GIT to a newer version)",
1074                         p->pack_name, ntohl(hdr.hdr_version));
1075
1076         /* Verify the pack matches its index. */
1077         if (p->num_objects != ntohl(hdr.hdr_entries))
1078                 return error("packfile %s claims to have %"PRIu32" objects"
1079                              " while index indicates %"PRIu32" objects",
1080                              p->pack_name, ntohl(hdr.hdr_entries),
1081                              p->num_objects);
1082         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1083                 return error("end of packfile %s is unavailable", p->pack_name);
1084         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1085                 return error("packfile %s signature is unavailable", p->pack_name);
1086         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1087         if (hashcmp(sha1, idx_sha1))
1088                 return error("packfile %s does not match index", p->pack_name);
1089         return 0;
1090 }
1091
1092 static int open_packed_git(struct packed_git *p)
1093 {
1094         if (!open_packed_git_1(p))
1095                 return 0;
1096         close_pack_fd(p);
1097         return -1;
1098 }
1099
1100 static int in_window(struct pack_window *win, off_t offset)
1101 {
1102         /* We must promise at least 20 bytes (one hash) after the
1103          * offset is available from this window, otherwise the offset
1104          * is not actually in this window and a different window (which
1105          * has that one hash excess) must be used.  This is to support
1106          * the object header and delta base parsing routines below.
1107          */
1108         off_t win_off = win->offset;
1109         return win_off <= offset
1110                 && (offset + 20) <= (win_off + win->len);
1111 }
1112
1113 unsigned char *use_pack(struct packed_git *p,
1114                 struct pack_window **w_cursor,
1115                 off_t offset,
1116                 unsigned long *left)
1117 {
1118         struct pack_window *win = *w_cursor;
1119
1120         /* Since packfiles end in a hash of their content and it's
1121          * pointless to ask for an offset into the middle of that
1122          * hash, and the in_window function above wouldn't match
1123          * don't allow an offset too close to the end of the file.
1124          */
1125         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1126                 die("packfile %s cannot be accessed", p->pack_name);
1127         if (offset > (p->pack_size - 20))
1128                 die("offset beyond end of packfile (truncated pack?)");
1129         if (offset < 0)
1130                 die(_("offset before end of packfile (broken .idx?)"));
1131
1132         if (!win || !in_window(win, offset)) {
1133                 if (win)
1134                         win->inuse_cnt--;
1135                 for (win = p->windows; win; win = win->next) {
1136                         if (in_window(win, offset))
1137                                 break;
1138                 }
1139                 if (!win) {
1140                         size_t window_align = packed_git_window_size / 2;
1141                         off_t len;
1142
1143                         if (p->pack_fd == -1 && open_packed_git(p))
1144                                 die("packfile %s cannot be accessed", p->pack_name);
1145
1146                         win = xcalloc(1, sizeof(*win));
1147                         win->offset = (offset / window_align) * window_align;
1148                         len = p->pack_size - win->offset;
1149                         if (len > packed_git_window_size)
1150                                 len = packed_git_window_size;
1151                         win->len = (size_t)len;
1152                         pack_mapped += win->len;
1153                         while (packed_git_limit < pack_mapped
1154                                 && unuse_one_window(p))
1155                                 ; /* nothing */
1156                         win->base = xmmap(NULL, win->len,
1157                                 PROT_READ, MAP_PRIVATE,
1158                                 p->pack_fd, win->offset);
1159                         if (win->base == MAP_FAILED)
1160                                 die_errno("packfile %s cannot be mapped",
1161                                           p->pack_name);
1162                         if (!win->offset && win->len == p->pack_size
1163                                 && !p->do_not_close)
1164                                 close_pack_fd(p);
1165                         pack_mmap_calls++;
1166                         pack_open_windows++;
1167                         if (pack_mapped > peak_pack_mapped)
1168                                 peak_pack_mapped = pack_mapped;
1169                         if (pack_open_windows > peak_pack_open_windows)
1170                                 peak_pack_open_windows = pack_open_windows;
1171                         win->next = p->windows;
1172                         p->windows = win;
1173                 }
1174         }
1175         if (win != *w_cursor) {
1176                 win->last_used = pack_used_ctr++;
1177                 win->inuse_cnt++;
1178                 *w_cursor = win;
1179         }
1180         offset -= win->offset;
1181         if (left)
1182                 *left = win->len - xsize_t(offset);
1183         return win->base + offset;
1184 }
1185
1186 static struct packed_git *alloc_packed_git(int extra)
1187 {
1188         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1189         memset(p, 0, sizeof(*p));
1190         p->pack_fd = -1;
1191         return p;
1192 }
1193
1194 static void try_to_free_pack_memory(size_t size)
1195 {
1196         release_pack_memory(size);
1197 }
1198
1199 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1200 {
1201         static int have_set_try_to_free_routine;
1202         struct stat st;
1203         size_t alloc;
1204         struct packed_git *p;
1205
1206         if (!have_set_try_to_free_routine) {
1207                 have_set_try_to_free_routine = 1;
1208                 set_try_to_free_routine(try_to_free_pack_memory);
1209         }
1210
1211         /*
1212          * Make sure a corresponding .pack file exists and that
1213          * the index looks sane.
1214          */
1215         if (!strip_suffix_mem(path, &path_len, ".idx"))
1216                 return NULL;
1217
1218         /*
1219          * ".pack" is long enough to hold any suffix we're adding (and
1220          * the use xsnprintf double-checks that)
1221          */
1222         alloc = st_add3(path_len, strlen(".pack"), 1);
1223         p = alloc_packed_git(alloc);
1224         memcpy(p->pack_name, path, path_len);
1225
1226         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1227         if (!access(p->pack_name, F_OK))
1228                 p->pack_keep = 1;
1229
1230         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1231         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1232                 free(p);
1233                 return NULL;
1234         }
1235
1236         /* ok, it looks sane as far as we can check without
1237          * actually mapping the pack file.
1238          */
1239         p->pack_size = st.st_size;
1240         p->pack_local = local;
1241         p->mtime = st.st_mtime;
1242         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1243                 hashclr(p->sha1);
1244         return p;
1245 }
1246
1247 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1248 {
1249         const char *path = sha1_pack_name(sha1);
1250         size_t alloc = st_add(strlen(path), 1);
1251         struct packed_git *p = alloc_packed_git(alloc);
1252
1253         memcpy(p->pack_name, path, alloc); /* includes NUL */
1254         hashcpy(p->sha1, sha1);
1255         if (check_packed_git_idx(idx_path, p)) {
1256                 free(p);
1257                 return NULL;
1258         }
1259
1260         return p;
1261 }
1262
1263 void install_packed_git(struct packed_git *pack)
1264 {
1265         if (pack->pack_fd != -1)
1266                 pack_open_fds++;
1267
1268         pack->next = packed_git;
1269         packed_git = pack;
1270 }
1271
1272 void (*report_garbage)(unsigned seen_bits, const char *path);
1273
1274 static void report_helper(const struct string_list *list,
1275                           int seen_bits, int first, int last)
1276 {
1277         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1278                 return;
1279
1280         for (; first < last; first++)
1281                 report_garbage(seen_bits, list->items[first].string);
1282 }
1283
1284 static void report_pack_garbage(struct string_list *list)
1285 {
1286         int i, baselen = -1, first = 0, seen_bits = 0;
1287
1288         if (!report_garbage)
1289                 return;
1290
1291         string_list_sort(list);
1292
1293         for (i = 0; i < list->nr; i++) {
1294                 const char *path = list->items[i].string;
1295                 if (baselen != -1 &&
1296                     strncmp(path, list->items[first].string, baselen)) {
1297                         report_helper(list, seen_bits, first, i);
1298                         baselen = -1;
1299                         seen_bits = 0;
1300                 }
1301                 if (baselen == -1) {
1302                         const char *dot = strrchr(path, '.');
1303                         if (!dot) {
1304                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1305                                 continue;
1306                         }
1307                         baselen = dot - path + 1;
1308                         first = i;
1309                 }
1310                 if (!strcmp(path + baselen, "pack"))
1311                         seen_bits |= 1;
1312                 else if (!strcmp(path + baselen, "idx"))
1313                         seen_bits |= 2;
1314         }
1315         report_helper(list, seen_bits, first, list->nr);
1316 }
1317
1318 static void prepare_packed_git_one(char *objdir, int local)
1319 {
1320         struct strbuf path = STRBUF_INIT;
1321         size_t dirnamelen;
1322         DIR *dir;
1323         struct dirent *de;
1324         struct string_list garbage = STRING_LIST_INIT_DUP;
1325
1326         strbuf_addstr(&path, objdir);
1327         strbuf_addstr(&path, "/pack");
1328         dir = opendir(path.buf);
1329         if (!dir) {
1330                 if (errno != ENOENT)
1331                         error_errno("unable to open object pack directory: %s",
1332                                     path.buf);
1333                 strbuf_release(&path);
1334                 return;
1335         }
1336         strbuf_addch(&path, '/');
1337         dirnamelen = path.len;
1338         while ((de = readdir(dir)) != NULL) {
1339                 struct packed_git *p;
1340                 size_t base_len;
1341
1342                 if (is_dot_or_dotdot(de->d_name))
1343                         continue;
1344
1345                 strbuf_setlen(&path, dirnamelen);
1346                 strbuf_addstr(&path, de->d_name);
1347
1348                 base_len = path.len;
1349                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1350                         /* Don't reopen a pack we already have. */
1351                         for (p = packed_git; p; p = p->next) {
1352                                 size_t len;
1353                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1354                                     len == base_len &&
1355                                     !memcmp(p->pack_name, path.buf, len))
1356                                         break;
1357                         }
1358                         if (p == NULL &&
1359                             /*
1360                              * See if it really is a valid .idx file with
1361                              * corresponding .pack file that we can map.
1362                              */
1363                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1364                                 install_packed_git(p);
1365                 }
1366
1367                 if (!report_garbage)
1368                         continue;
1369
1370                 if (ends_with(de->d_name, ".idx") ||
1371                     ends_with(de->d_name, ".pack") ||
1372                     ends_with(de->d_name, ".bitmap") ||
1373                     ends_with(de->d_name, ".keep"))
1374                         string_list_append(&garbage, path.buf);
1375                 else
1376                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1377         }
1378         closedir(dir);
1379         report_pack_garbage(&garbage);
1380         string_list_clear(&garbage, 0);
1381         strbuf_release(&path);
1382 }
1383
1384 static void *get_next_packed_git(const void *p)
1385 {
1386         return ((const struct packed_git *)p)->next;
1387 }
1388
1389 static void set_next_packed_git(void *p, void *next)
1390 {
1391         ((struct packed_git *)p)->next = next;
1392 }
1393
1394 static int sort_pack(const void *a_, const void *b_)
1395 {
1396         const struct packed_git *a = a_;
1397         const struct packed_git *b = b_;
1398         int st;
1399
1400         /*
1401          * Local packs tend to contain objects specific to our
1402          * variant of the project than remote ones.  In addition,
1403          * remote ones could be on a network mounted filesystem.
1404          * Favor local ones for these reasons.
1405          */
1406         st = a->pack_local - b->pack_local;
1407         if (st)
1408                 return -st;
1409
1410         /*
1411          * Younger packs tend to contain more recent objects,
1412          * and more recent objects tend to get accessed more
1413          * often.
1414          */
1415         if (a->mtime < b->mtime)
1416                 return 1;
1417         else if (a->mtime == b->mtime)
1418                 return 0;
1419         return -1;
1420 }
1421
1422 static void rearrange_packed_git(void)
1423 {
1424         packed_git = llist_mergesort(packed_git, get_next_packed_git,
1425                                      set_next_packed_git, sort_pack);
1426 }
1427
1428 static void prepare_packed_git_mru(void)
1429 {
1430         struct packed_git *p;
1431
1432         mru_clear(packed_git_mru);
1433         for (p = packed_git; p; p = p->next)
1434                 mru_append(packed_git_mru, p);
1435 }
1436
1437 static int prepare_packed_git_run_once = 0;
1438 void prepare_packed_git(void)
1439 {
1440         struct alternate_object_database *alt;
1441
1442         if (prepare_packed_git_run_once)
1443                 return;
1444         prepare_packed_git_one(get_object_directory(), 1);
1445         prepare_alt_odb();
1446         for (alt = alt_odb_list; alt; alt = alt->next) {
1447                 alt->name[-1] = 0;
1448                 prepare_packed_git_one(alt->base, 0);
1449                 alt->name[-1] = '/';
1450         }
1451         rearrange_packed_git();
1452         prepare_packed_git_mru();
1453         prepare_packed_git_run_once = 1;
1454 }
1455
1456 void reprepare_packed_git(void)
1457 {
1458         prepare_packed_git_run_once = 0;
1459         prepare_packed_git();
1460 }
1461
1462 static void mark_bad_packed_object(struct packed_git *p,
1463                                    const unsigned char *sha1)
1464 {
1465         unsigned i;
1466         for (i = 0; i < p->num_bad_objects; i++)
1467                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1468                         return;
1469         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1470                                       st_mult(GIT_SHA1_RAWSZ,
1471                                               st_add(p->num_bad_objects, 1)));
1472         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1473         p->num_bad_objects++;
1474 }
1475
1476 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1477 {
1478         struct packed_git *p;
1479         unsigned i;
1480
1481         for (p = packed_git; p; p = p->next)
1482                 for (i = 0; i < p->num_bad_objects; i++)
1483                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1484                                 return p;
1485         return NULL;
1486 }
1487
1488 /*
1489  * With an in-core object data in "map", rehash it to make sure the
1490  * object name actually matches "sha1" to detect object corruption.
1491  * With "map" == NULL, try reading the object named with "sha1" using
1492  * the streaming interface and rehash it to do the same.
1493  */
1494 int check_sha1_signature(const unsigned char *sha1, void *map,
1495                          unsigned long size, const char *type)
1496 {
1497         unsigned char real_sha1[20];
1498         enum object_type obj_type;
1499         struct git_istream *st;
1500         git_SHA_CTX c;
1501         char hdr[32];
1502         int hdrlen;
1503
1504         if (map) {
1505                 hash_sha1_file(map, size, type, real_sha1);
1506                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1507         }
1508
1509         st = open_istream(sha1, &obj_type, &size, NULL);
1510         if (!st)
1511                 return -1;
1512
1513         /* Generate the header */
1514         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1515
1516         /* Sha1.. */
1517         git_SHA1_Init(&c);
1518         git_SHA1_Update(&c, hdr, hdrlen);
1519         for (;;) {
1520                 char buf[1024 * 16];
1521                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1522
1523                 if (readlen < 0) {
1524                         close_istream(st);
1525                         return -1;
1526                 }
1527                 if (!readlen)
1528                         break;
1529                 git_SHA1_Update(&c, buf, readlen);
1530         }
1531         git_SHA1_Final(real_sha1, &c);
1532         close_istream(st);
1533         return hashcmp(sha1, real_sha1) ? -1 : 0;
1534 }
1535
1536 int git_open_noatime(const char *name)
1537 {
1538         static int sha1_file_open_flag = O_NOATIME;
1539
1540         for (;;) {
1541                 int fd;
1542
1543                 errno = 0;
1544                 fd = open(name, O_RDONLY | sha1_file_open_flag);
1545                 if (fd >= 0)
1546                         return fd;
1547
1548                 /* Might the failure be due to O_NOATIME? */
1549                 if (errno != ENOENT && sha1_file_open_flag) {
1550                         sha1_file_open_flag = 0;
1551                         continue;
1552                 }
1553
1554                 return -1;
1555         }
1556 }
1557
1558 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1559 {
1560         struct alternate_object_database *alt;
1561
1562         if (!lstat(sha1_file_name(sha1), st))
1563                 return 0;
1564
1565         prepare_alt_odb();
1566         errno = ENOENT;
1567         for (alt = alt_odb_list; alt; alt = alt->next) {
1568                 fill_sha1_path(alt->name, sha1);
1569                 if (!lstat(alt->base, st))
1570                         return 0;
1571         }
1572
1573         return -1;
1574 }
1575
1576 static int open_sha1_file(const unsigned char *sha1)
1577 {
1578         int fd;
1579         struct alternate_object_database *alt;
1580         int most_interesting_errno;
1581
1582         fd = git_open_noatime(sha1_file_name(sha1));
1583         if (fd >= 0)
1584                 return fd;
1585         most_interesting_errno = errno;
1586
1587         prepare_alt_odb();
1588         for (alt = alt_odb_list; alt; alt = alt->next) {
1589                 fill_sha1_path(alt->name, sha1);
1590                 fd = git_open_noatime(alt->base);
1591                 if (fd >= 0)
1592                         return fd;
1593                 if (most_interesting_errno == ENOENT)
1594                         most_interesting_errno = errno;
1595         }
1596         errno = most_interesting_errno;
1597         return -1;
1598 }
1599
1600 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1601 {
1602         void *map;
1603         int fd;
1604
1605         fd = open_sha1_file(sha1);
1606         map = NULL;
1607         if (fd >= 0) {
1608                 struct stat st;
1609
1610                 if (!fstat(fd, &st)) {
1611                         *size = xsize_t(st.st_size);
1612                         if (!*size) {
1613                                 /* mmap() is forbidden on empty files */
1614                                 error("object file %s is empty", sha1_file_name(sha1));
1615                                 return NULL;
1616                         }
1617                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1618                 }
1619                 close(fd);
1620         }
1621         return map;
1622 }
1623
1624 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1625                 unsigned long len, enum object_type *type, unsigned long *sizep)
1626 {
1627         unsigned shift;
1628         unsigned long size, c;
1629         unsigned long used = 0;
1630
1631         c = buf[used++];
1632         *type = (c >> 4) & 7;
1633         size = c & 15;
1634         shift = 4;
1635         while (c & 0x80) {
1636                 if (len <= used || bitsizeof(long) <= shift) {
1637                         error("bad object header");
1638                         size = used = 0;
1639                         break;
1640                 }
1641                 c = buf[used++];
1642                 size += (c & 0x7f) << shift;
1643                 shift += 7;
1644         }
1645         *sizep = size;
1646         return used;
1647 }
1648
1649 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1650 {
1651         /* Get the data stream */
1652         memset(stream, 0, sizeof(*stream));
1653         stream->next_in = map;
1654         stream->avail_in = mapsize;
1655         stream->next_out = buffer;
1656         stream->avail_out = bufsiz;
1657
1658         git_inflate_init(stream);
1659         return git_inflate(stream, 0);
1660 }
1661
1662 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1663                                         unsigned long mapsize, void *buffer,
1664                                         unsigned long bufsiz, struct strbuf *header)
1665 {
1666         int status;
1667
1668         status = unpack_sha1_header(stream, map, mapsize, buffer, bufsiz);
1669
1670         /*
1671          * Check if entire header is unpacked in the first iteration.
1672          */
1673         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1674                 return 0;
1675
1676         /*
1677          * buffer[0..bufsiz] was not large enough.  Copy the partial
1678          * result out to header, and then append the result of further
1679          * reading the stream.
1680          */
1681         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1682         stream->next_out = buffer;
1683         stream->avail_out = bufsiz;
1684
1685         do {
1686                 status = git_inflate(stream, 0);
1687                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1688                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1689                         return 0;
1690                 stream->next_out = buffer;
1691                 stream->avail_out = bufsiz;
1692         } while (status != Z_STREAM_END);
1693         return -1;
1694 }
1695
1696 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1697 {
1698         int bytes = strlen(buffer) + 1;
1699         unsigned char *buf = xmallocz(size);
1700         unsigned long n;
1701         int status = Z_OK;
1702
1703         n = stream->total_out - bytes;
1704         if (n > size)
1705                 n = size;
1706         memcpy(buf, (char *) buffer + bytes, n);
1707         bytes = n;
1708         if (bytes <= size) {
1709                 /*
1710                  * The above condition must be (bytes <= size), not
1711                  * (bytes < size).  In other words, even though we
1712                  * expect no more output and set avail_out to zero,
1713                  * the input zlib stream may have bytes that express
1714                  * "this concludes the stream", and we *do* want to
1715                  * eat that input.
1716                  *
1717                  * Otherwise we would not be able to test that we
1718                  * consumed all the input to reach the expected size;
1719                  * we also want to check that zlib tells us that all
1720                  * went well with status == Z_STREAM_END at the end.
1721                  */
1722                 stream->next_out = buf + bytes;
1723                 stream->avail_out = size - bytes;
1724                 while (status == Z_OK)
1725                         status = git_inflate(stream, Z_FINISH);
1726         }
1727         if (status == Z_STREAM_END && !stream->avail_in) {
1728                 git_inflate_end(stream);
1729                 return buf;
1730         }
1731
1732         if (status < 0)
1733                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1734         else if (stream->avail_in)
1735                 error("garbage at end of loose object '%s'",
1736                       sha1_to_hex(sha1));
1737         free(buf);
1738         return NULL;
1739 }
1740
1741 /*
1742  * We used to just use "sscanf()", but that's actually way
1743  * too permissive for what we want to check. So do an anal
1744  * object header parse by hand.
1745  */
1746 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1747                                unsigned int flags)
1748 {
1749         const char *type_buf = hdr;
1750         unsigned long size;
1751         int type, type_len = 0;
1752
1753         /*
1754          * The type can be of any size but is followed by
1755          * a space.
1756          */
1757         for (;;) {
1758                 char c = *hdr++;
1759                 if (c == ' ')
1760                         break;
1761                 type_len++;
1762         }
1763
1764         type = type_from_string_gently(type_buf, type_len, 1);
1765         if (oi->typename)
1766                 strbuf_add(oi->typename, type_buf, type_len);
1767         /*
1768          * Set type to 0 if its an unknown object and
1769          * we're obtaining the type using '--allow-unknown-type'
1770          * option.
1771          */
1772         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1773                 type = 0;
1774         else if (type < 0)
1775                 die("invalid object type");
1776         if (oi->typep)
1777                 *oi->typep = type;
1778
1779         /*
1780          * The length must follow immediately, and be in canonical
1781          * decimal format (ie "010" is not valid).
1782          */
1783         size = *hdr++ - '0';
1784         if (size > 9)
1785                 return -1;
1786         if (size) {
1787                 for (;;) {
1788                         unsigned long c = *hdr - '0';
1789                         if (c > 9)
1790                                 break;
1791                         hdr++;
1792                         size = size * 10 + c;
1793                 }
1794         }
1795
1796         if (oi->sizep)
1797                 *oi->sizep = size;
1798
1799         /*
1800          * The length must be followed by a zero byte
1801          */
1802         return *hdr ? -1 : type;
1803 }
1804
1805 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1806 {
1807         struct object_info oi;
1808
1809         oi.sizep = sizep;
1810         oi.typename = NULL;
1811         oi.typep = NULL;
1812         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
1813 }
1814
1815 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1816 {
1817         int ret;
1818         git_zstream stream;
1819         char hdr[8192];
1820
1821         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1822         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1823                 return NULL;
1824
1825         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1826 }
1827
1828 unsigned long get_size_from_delta(struct packed_git *p,
1829                                   struct pack_window **w_curs,
1830                                   off_t curpos)
1831 {
1832         const unsigned char *data;
1833         unsigned char delta_head[20], *in;
1834         git_zstream stream;
1835         int st;
1836
1837         memset(&stream, 0, sizeof(stream));
1838         stream.next_out = delta_head;
1839         stream.avail_out = sizeof(delta_head);
1840
1841         git_inflate_init(&stream);
1842         do {
1843                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1844                 stream.next_in = in;
1845                 st = git_inflate(&stream, Z_FINISH);
1846                 curpos += stream.next_in - in;
1847         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1848                  stream.total_out < sizeof(delta_head));
1849         git_inflate_end(&stream);
1850         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1851                 error("delta data unpack-initial failed");
1852                 return 0;
1853         }
1854
1855         /* Examine the initial part of the delta to figure out
1856          * the result size.
1857          */
1858         data = delta_head;
1859
1860         /* ignore base size */
1861         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1862
1863         /* Read the result size */
1864         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1865 }
1866
1867 static off_t get_delta_base(struct packed_git *p,
1868                                     struct pack_window **w_curs,
1869                                     off_t *curpos,
1870                                     enum object_type type,
1871                                     off_t delta_obj_offset)
1872 {
1873         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1874         off_t base_offset;
1875
1876         /* use_pack() assured us we have [base_info, base_info + 20)
1877          * as a range that we can look at without walking off the
1878          * end of the mapped window.  Its actually the hash size
1879          * that is assured.  An OFS_DELTA longer than the hash size
1880          * is stupid, as then a REF_DELTA would be smaller to store.
1881          */
1882         if (type == OBJ_OFS_DELTA) {
1883                 unsigned used = 0;
1884                 unsigned char c = base_info[used++];
1885                 base_offset = c & 127;
1886                 while (c & 128) {
1887                         base_offset += 1;
1888                         if (!base_offset || MSB(base_offset, 7))
1889                                 return 0;  /* overflow */
1890                         c = base_info[used++];
1891                         base_offset = (base_offset << 7) + (c & 127);
1892                 }
1893                 base_offset = delta_obj_offset - base_offset;
1894                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1895                         return 0;  /* out of bound */
1896                 *curpos += used;
1897         } else if (type == OBJ_REF_DELTA) {
1898                 /* The base entry _must_ be in the same pack */
1899                 base_offset = find_pack_entry_one(base_info, p);
1900                 *curpos += 20;
1901         } else
1902                 die("I am totally screwed");
1903         return base_offset;
1904 }
1905
1906 /*
1907  * Like get_delta_base above, but we return the sha1 instead of the pack
1908  * offset. This means it is cheaper for REF deltas (we do not have to do
1909  * the final object lookup), but more expensive for OFS deltas (we
1910  * have to load the revidx to convert the offset back into a sha1).
1911  */
1912 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1913                                                 struct pack_window **w_curs,
1914                                                 off_t curpos,
1915                                                 enum object_type type,
1916                                                 off_t delta_obj_offset)
1917 {
1918         if (type == OBJ_REF_DELTA) {
1919                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1920                 return base;
1921         } else if (type == OBJ_OFS_DELTA) {
1922                 struct revindex_entry *revidx;
1923                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1924                                                    type, delta_obj_offset);
1925
1926                 if (!base_offset)
1927                         return NULL;
1928
1929                 revidx = find_pack_revindex(p, base_offset);
1930                 if (!revidx)
1931                         return NULL;
1932
1933                 return nth_packed_object_sha1(p, revidx->nr);
1934         } else
1935                 return NULL;
1936 }
1937
1938 int unpack_object_header(struct packed_git *p,
1939                          struct pack_window **w_curs,
1940                          off_t *curpos,
1941                          unsigned long *sizep)
1942 {
1943         unsigned char *base;
1944         unsigned long left;
1945         unsigned long used;
1946         enum object_type type;
1947
1948         /* use_pack() assures us we have [base, base + 20) available
1949          * as a range that we can look at.  (Its actually the hash
1950          * size that is assured.)  With our object header encoding
1951          * the maximum deflated object size is 2^137, which is just
1952          * insane, so we know won't exceed what we have been given.
1953          */
1954         base = use_pack(p, w_curs, *curpos, &left);
1955         used = unpack_object_header_buffer(base, left, &type, sizep);
1956         if (!used) {
1957                 type = OBJ_BAD;
1958         } else
1959                 *curpos += used;
1960
1961         return type;
1962 }
1963
1964 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1965 {
1966         int type;
1967         struct revindex_entry *revidx;
1968         const unsigned char *sha1;
1969         revidx = find_pack_revindex(p, obj_offset);
1970         if (!revidx)
1971                 return OBJ_BAD;
1972         sha1 = nth_packed_object_sha1(p, revidx->nr);
1973         mark_bad_packed_object(p, sha1);
1974         type = sha1_object_info(sha1, NULL);
1975         if (type <= OBJ_NONE)
1976                 return OBJ_BAD;
1977         return type;
1978 }
1979
1980 #define POI_STACK_PREALLOC 64
1981
1982 static enum object_type packed_to_object_type(struct packed_git *p,
1983                                               off_t obj_offset,
1984                                               enum object_type type,
1985                                               struct pack_window **w_curs,
1986                                               off_t curpos)
1987 {
1988         off_t small_poi_stack[POI_STACK_PREALLOC];
1989         off_t *poi_stack = small_poi_stack;
1990         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1991
1992         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1993                 off_t base_offset;
1994                 unsigned long size;
1995                 /* Push the object we're going to leave behind */
1996                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1997                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1998                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1999                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
2000                 } else {
2001                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
2002                 }
2003                 poi_stack[poi_stack_nr++] = obj_offset;
2004                 /* If parsing the base offset fails, just unwind */
2005                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
2006                 if (!base_offset)
2007                         goto unwind;
2008                 curpos = obj_offset = base_offset;
2009                 type = unpack_object_header(p, w_curs, &curpos, &size);
2010                 if (type <= OBJ_NONE) {
2011                         /* If getting the base itself fails, we first
2012                          * retry the base, otherwise unwind */
2013                         type = retry_bad_packed_offset(p, base_offset);
2014                         if (type > OBJ_NONE)
2015                                 goto out;
2016                         goto unwind;
2017                 }
2018         }
2019
2020         switch (type) {
2021         case OBJ_BAD:
2022         case OBJ_COMMIT:
2023         case OBJ_TREE:
2024         case OBJ_BLOB:
2025         case OBJ_TAG:
2026                 break;
2027         default:
2028                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2029                       type, (uintmax_t)obj_offset, p->pack_name);
2030                 type = OBJ_BAD;
2031         }
2032
2033 out:
2034         if (poi_stack != small_poi_stack)
2035                 free(poi_stack);
2036         return type;
2037
2038 unwind:
2039         while (poi_stack_nr) {
2040                 obj_offset = poi_stack[--poi_stack_nr];
2041                 type = retry_bad_packed_offset(p, obj_offset);
2042                 if (type > OBJ_NONE)
2043                         goto out;
2044         }
2045         type = OBJ_BAD;
2046         goto out;
2047 }
2048
2049 static int packed_object_info(struct packed_git *p, off_t obj_offset,
2050                               struct object_info *oi)
2051 {
2052         struct pack_window *w_curs = NULL;
2053         unsigned long size;
2054         off_t curpos = obj_offset;
2055         enum object_type type;
2056
2057         /*
2058          * We always get the representation type, but only convert it to
2059          * a "real" type later if the caller is interested.
2060          */
2061         type = unpack_object_header(p, &w_curs, &curpos, &size);
2062
2063         if (oi->sizep) {
2064                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2065                         off_t tmp_pos = curpos;
2066                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2067                                                            type, obj_offset);
2068                         if (!base_offset) {
2069                                 type = OBJ_BAD;
2070                                 goto out;
2071                         }
2072                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2073                         if (*oi->sizep == 0) {
2074                                 type = OBJ_BAD;
2075                                 goto out;
2076                         }
2077                 } else {
2078                         *oi->sizep = size;
2079                 }
2080         }
2081
2082         if (oi->disk_sizep) {
2083                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2084                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2085         }
2086
2087         if (oi->typep) {
2088                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2089                 if (*oi->typep < 0) {
2090                         type = OBJ_BAD;
2091                         goto out;
2092                 }
2093         }
2094
2095         if (oi->delta_base_sha1) {
2096                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2097                         const unsigned char *base;
2098
2099                         base = get_delta_base_sha1(p, &w_curs, curpos,
2100                                                    type, obj_offset);
2101                         if (!base) {
2102                                 type = OBJ_BAD;
2103                                 goto out;
2104                         }
2105
2106                         hashcpy(oi->delta_base_sha1, base);
2107                 } else
2108                         hashclr(oi->delta_base_sha1);
2109         }
2110
2111 out:
2112         unuse_pack(&w_curs);
2113         return type;
2114 }
2115
2116 static void *unpack_compressed_entry(struct packed_git *p,
2117                                     struct pack_window **w_curs,
2118                                     off_t curpos,
2119                                     unsigned long size)
2120 {
2121         int st;
2122         git_zstream stream;
2123         unsigned char *buffer, *in;
2124
2125         buffer = xmallocz_gently(size);
2126         if (!buffer)
2127                 return NULL;
2128         memset(&stream, 0, sizeof(stream));
2129         stream.next_out = buffer;
2130         stream.avail_out = size + 1;
2131
2132         git_inflate_init(&stream);
2133         do {
2134                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2135                 stream.next_in = in;
2136                 st = git_inflate(&stream, Z_FINISH);
2137                 if (!stream.avail_out)
2138                         break; /* the payload is larger than it should be */
2139                 curpos += stream.next_in - in;
2140         } while (st == Z_OK || st == Z_BUF_ERROR);
2141         git_inflate_end(&stream);
2142         if ((st != Z_STREAM_END) || stream.total_out != size) {
2143                 free(buffer);
2144                 return NULL;
2145         }
2146
2147         return buffer;
2148 }
2149
2150 static struct hashmap delta_base_cache;
2151 static size_t delta_base_cached;
2152
2153 static LIST_HEAD(delta_base_cache_lru);
2154
2155 struct delta_base_cache_key {
2156         struct packed_git *p;
2157         off_t base_offset;
2158 };
2159
2160 struct delta_base_cache_entry {
2161         struct hashmap hash;
2162         struct delta_base_cache_key key;
2163         struct list_head lru;
2164         void *data;
2165         unsigned long size;
2166         enum object_type type;
2167 };
2168
2169 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
2170 {
2171         unsigned int hash;
2172
2173         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
2174         hash += (hash >> 8) + (hash >> 16);
2175         return hash;
2176 }
2177
2178 static struct delta_base_cache_entry *
2179 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2180 {
2181         struct hashmap_entry entry;
2182         struct delta_base_cache_key key;
2183
2184         if (!delta_base_cache.cmpfn)
2185                 return NULL;
2186
2187         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
2188         key.p = p;
2189         key.base_offset = base_offset;
2190         return hashmap_get(&delta_base_cache, &entry, &key);
2191 }
2192
2193 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
2194                                    const struct delta_base_cache_key *b)
2195 {
2196         return a->p == b->p && a->base_offset == b->base_offset;
2197 }
2198
2199 static int delta_base_cache_hash_cmp(const void *va, const void *vb,
2200                                      const void *vkey)
2201 {
2202         const struct delta_base_cache_entry *a = va, *b = vb;
2203         const struct delta_base_cache_key *key = vkey;
2204         if (key)
2205                 return !delta_base_cache_key_eq(&a->key, key);
2206         else
2207                 return !delta_base_cache_key_eq(&a->key, &b->key);
2208 }
2209
2210 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2211 {
2212         return !!get_delta_base_cache_entry(p, base_offset);
2213 }
2214
2215 /*
2216  * Remove the entry from the cache, but do _not_ free the associated
2217  * entry data. The caller takes ownership of the "data" buffer, and
2218  * should copy out any fields it wants before detaching.
2219  */
2220 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2221 {
2222         hashmap_remove(&delta_base_cache, ent, &ent->key);
2223         list_del(&ent->lru);
2224         delta_base_cached -= ent->size;
2225         free(ent);
2226 }
2227
2228 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2229         unsigned long *base_size, enum object_type *type)
2230 {
2231         struct delta_base_cache_entry *ent;
2232
2233         ent = get_delta_base_cache_entry(p, base_offset);
2234         if (!ent)
2235                 return unpack_entry(p, base_offset, type, base_size);
2236
2237         *type = ent->type;
2238         *base_size = ent->size;
2239         return xmemdupz(ent->data, ent->size);
2240 }
2241
2242 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2243 {
2244         free(ent->data);
2245         detach_delta_base_cache_entry(ent);
2246 }
2247
2248 void clear_delta_base_cache(void)
2249 {
2250         struct hashmap_iter iter;
2251         struct delta_base_cache_entry *entry;
2252         for (entry = hashmap_iter_first(&delta_base_cache, &iter);
2253              entry;
2254              entry = hashmap_iter_next(&iter)) {
2255                 release_delta_base_cache(entry);
2256         }
2257 }
2258
2259 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2260         void *base, unsigned long base_size, enum object_type type)
2261 {
2262         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
2263         struct list_head *lru, *tmp;
2264
2265         delta_base_cached += base_size;
2266
2267         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
2268                 struct delta_base_cache_entry *f =
2269                         list_entry(lru, struct delta_base_cache_entry, lru);
2270                 if (delta_base_cached <= delta_base_cache_limit)
2271                         break;
2272                 release_delta_base_cache(f);
2273         }
2274
2275         ent->key.p = p;
2276         ent->key.base_offset = base_offset;
2277         ent->type = type;
2278         ent->data = base;
2279         ent->size = base_size;
2280         list_add_tail(&ent->lru, &delta_base_cache_lru);
2281
2282         if (!delta_base_cache.cmpfn)
2283                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, 0);
2284         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
2285         hashmap_add(&delta_base_cache, ent);
2286 }
2287
2288 static void *read_object(const unsigned char *sha1, enum object_type *type,
2289                          unsigned long *size);
2290
2291 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2292 {
2293         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2294         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2295                          p->pack_name, (uintmax_t)obj_offset);
2296 }
2297
2298 int do_check_packed_object_crc;
2299
2300 #define UNPACK_ENTRY_STACK_PREALLOC 64
2301 struct unpack_entry_stack_ent {
2302         off_t obj_offset;
2303         off_t curpos;
2304         unsigned long size;
2305 };
2306
2307 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2308                    enum object_type *final_type, unsigned long *final_size)
2309 {
2310         struct pack_window *w_curs = NULL;
2311         off_t curpos = obj_offset;
2312         void *data = NULL;
2313         unsigned long size;
2314         enum object_type type;
2315         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2316         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2317         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2318         int base_from_cache = 0;
2319
2320         write_pack_access_log(p, obj_offset);
2321
2322         /* PHASE 1: drill down to the innermost base object */
2323         for (;;) {
2324                 off_t base_offset;
2325                 int i;
2326                 struct delta_base_cache_entry *ent;
2327
2328                 ent = get_delta_base_cache_entry(p, curpos);
2329                 if (ent) {
2330                         type = ent->type;
2331                         data = ent->data;
2332                         size = ent->size;
2333                         detach_delta_base_cache_entry(ent);
2334                         base_from_cache = 1;
2335                         break;
2336                 }
2337
2338                 if (do_check_packed_object_crc && p->index_version > 1) {
2339                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2340                         off_t len = revidx[1].offset - obj_offset;
2341                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2342                                 const unsigned char *sha1 =
2343                                         nth_packed_object_sha1(p, revidx->nr);
2344                                 error("bad packed object CRC for %s",
2345                                       sha1_to_hex(sha1));
2346                                 mark_bad_packed_object(p, sha1);
2347                                 unuse_pack(&w_curs);
2348                                 return NULL;
2349                         }
2350                 }
2351
2352                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2353                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2354                         break;
2355
2356                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2357                 if (!base_offset) {
2358                         error("failed to validate delta base reference "
2359                               "at offset %"PRIuMAX" from %s",
2360                               (uintmax_t)curpos, p->pack_name);
2361                         /* bail to phase 2, in hopes of recovery */
2362                         data = NULL;
2363                         break;
2364                 }
2365
2366                 /* push object, proceed to base */
2367                 if (delta_stack_nr >= delta_stack_alloc
2368                     && delta_stack == small_delta_stack) {
2369                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2370                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2371                         memcpy(delta_stack, small_delta_stack,
2372                                sizeof(*delta_stack)*delta_stack_nr);
2373                 } else {
2374                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2375                 }
2376                 i = delta_stack_nr++;
2377                 delta_stack[i].obj_offset = obj_offset;
2378                 delta_stack[i].curpos = curpos;
2379                 delta_stack[i].size = size;
2380
2381                 curpos = obj_offset = base_offset;
2382         }
2383
2384         /* PHASE 2: handle the base */
2385         switch (type) {
2386         case OBJ_OFS_DELTA:
2387         case OBJ_REF_DELTA:
2388                 if (data)
2389                         die("BUG: unpack_entry: left loop at a valid delta");
2390                 break;
2391         case OBJ_COMMIT:
2392         case OBJ_TREE:
2393         case OBJ_BLOB:
2394         case OBJ_TAG:
2395                 if (!base_from_cache)
2396                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2397                 break;
2398         default:
2399                 data = NULL;
2400                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2401                       type, (uintmax_t)obj_offset, p->pack_name);
2402         }
2403
2404         /* PHASE 3: apply deltas in order */
2405
2406         /* invariants:
2407          *   'data' holds the base data, or NULL if there was corruption
2408          */
2409         while (delta_stack_nr) {
2410                 void *delta_data;
2411                 void *base = data;
2412                 unsigned long delta_size, base_size = size;
2413                 int i;
2414
2415                 data = NULL;
2416
2417                 if (base)
2418                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2419
2420                 if (!base) {
2421                         /*
2422                          * We're probably in deep shit, but let's try to fetch
2423                          * the required base anyway from another pack or loose.
2424                          * This is costly but should happen only in the presence
2425                          * of a corrupted pack, and is better than failing outright.
2426                          */
2427                         struct revindex_entry *revidx;
2428                         const unsigned char *base_sha1;
2429                         revidx = find_pack_revindex(p, obj_offset);
2430                         if (revidx) {
2431                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2432                                 error("failed to read delta base object %s"
2433                                       " at offset %"PRIuMAX" from %s",
2434                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2435                                       p->pack_name);
2436                                 mark_bad_packed_object(p, base_sha1);
2437                                 base = read_object(base_sha1, &type, &base_size);
2438                         }
2439                 }
2440
2441                 i = --delta_stack_nr;
2442                 obj_offset = delta_stack[i].obj_offset;
2443                 curpos = delta_stack[i].curpos;
2444                 delta_size = delta_stack[i].size;
2445
2446                 if (!base)
2447                         continue;
2448
2449                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2450
2451                 if (!delta_data) {
2452                         error("failed to unpack compressed delta "
2453                               "at offset %"PRIuMAX" from %s",
2454                               (uintmax_t)curpos, p->pack_name);
2455                         data = NULL;
2456                         continue;
2457                 }
2458
2459                 data = patch_delta(base, base_size,
2460                                    delta_data, delta_size,
2461                                    &size);
2462
2463                 /*
2464                  * We could not apply the delta; warn the user, but keep going.
2465                  * Our failure will be noticed either in the next iteration of
2466                  * the loop, or if this is the final delta, in the caller when
2467                  * we return NULL. Those code paths will take care of making
2468                  * a more explicit warning and retrying with another copy of
2469                  * the object.
2470                  */
2471                 if (!data)
2472                         error("failed to apply delta");
2473
2474                 free(delta_data);
2475         }
2476
2477         *final_type = type;
2478         *final_size = size;
2479
2480         unuse_pack(&w_curs);
2481
2482         if (delta_stack != small_delta_stack)
2483                 free(delta_stack);
2484
2485         return data;
2486 }
2487
2488 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2489                                             uint32_t n)
2490 {
2491         const unsigned char *index = p->index_data;
2492         if (!index) {
2493                 if (open_pack_index(p))
2494                         return NULL;
2495                 index = p->index_data;
2496         }
2497         if (n >= p->num_objects)
2498                 return NULL;
2499         index += 4 * 256;
2500         if (p->index_version == 1) {
2501                 return index + 24 * n + 4;
2502         } else {
2503                 index += 8;
2504                 return index + 20 * n;
2505         }
2506 }
2507
2508 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2509 {
2510         const unsigned char *ptr = vptr;
2511         const unsigned char *start = p->index_data;
2512         const unsigned char *end = start + p->index_size;
2513         if (ptr < start)
2514                 die(_("offset before start of pack index for %s (corrupt index?)"),
2515                     p->pack_name);
2516         /* No need to check for underflow; .idx files must be at least 8 bytes */
2517         if (ptr >= end - 8)
2518                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2519                     p->pack_name);
2520 }
2521
2522 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2523 {
2524         const unsigned char *index = p->index_data;
2525         index += 4 * 256;
2526         if (p->index_version == 1) {
2527                 return ntohl(*((uint32_t *)(index + 24 * n)));
2528         } else {
2529                 uint32_t off;
2530                 index += 8 + p->num_objects * (20 + 4);
2531                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2532                 if (!(off & 0x80000000))
2533                         return off;
2534                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2535                 check_pack_index_ptr(p, index);
2536                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2537                                    ntohl(*((uint32_t *)(index + 4)));
2538         }
2539 }
2540
2541 off_t find_pack_entry_one(const unsigned char *sha1,
2542                                   struct packed_git *p)
2543 {
2544         const uint32_t *level1_ofs = p->index_data;
2545         const unsigned char *index = p->index_data;
2546         unsigned hi, lo, stride;
2547         static int use_lookup = -1;
2548         static int debug_lookup = -1;
2549
2550         if (debug_lookup < 0)
2551                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2552
2553         if (!index) {
2554                 if (open_pack_index(p))
2555                         return 0;
2556                 level1_ofs = p->index_data;
2557                 index = p->index_data;
2558         }
2559         if (p->index_version > 1) {
2560                 level1_ofs += 2;
2561                 index += 8;
2562         }
2563         index += 4 * 256;
2564         hi = ntohl(level1_ofs[*sha1]);
2565         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2566         if (p->index_version > 1) {
2567                 stride = 20;
2568         } else {
2569                 stride = 24;
2570                 index += 4;
2571         }
2572
2573         if (debug_lookup)
2574                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2575                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2576
2577         if (use_lookup < 0)
2578                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2579         if (use_lookup) {
2580                 int pos = sha1_entry_pos(index, stride, 0,
2581                                          lo, hi, p->num_objects, sha1);
2582                 if (pos < 0)
2583                         return 0;
2584                 return nth_packed_object_offset(p, pos);
2585         }
2586
2587         do {
2588                 unsigned mi = (lo + hi) / 2;
2589                 int cmp = hashcmp(index + mi * stride, sha1);
2590
2591                 if (debug_lookup)
2592                         printf("lo %u hi %u rg %u mi %u\n",
2593                                lo, hi, hi - lo, mi);
2594                 if (!cmp)
2595                         return nth_packed_object_offset(p, mi);
2596                 if (cmp > 0)
2597                         hi = mi;
2598                 else
2599                         lo = mi+1;
2600         } while (lo < hi);
2601         return 0;
2602 }
2603
2604 int is_pack_valid(struct packed_git *p)
2605 {
2606         /* An already open pack is known to be valid. */
2607         if (p->pack_fd != -1)
2608                 return 1;
2609
2610         /* If the pack has one window completely covering the
2611          * file size, the pack is known to be valid even if
2612          * the descriptor is not currently open.
2613          */
2614         if (p->windows) {
2615                 struct pack_window *w = p->windows;
2616
2617                 if (!w->offset && w->len == p->pack_size)
2618                         return 1;
2619         }
2620
2621         /* Force the pack to open to prove its valid. */
2622         return !open_packed_git(p);
2623 }
2624
2625 static int fill_pack_entry(const unsigned char *sha1,
2626                            struct pack_entry *e,
2627                            struct packed_git *p)
2628 {
2629         off_t offset;
2630
2631         if (p->num_bad_objects) {
2632                 unsigned i;
2633                 for (i = 0; i < p->num_bad_objects; i++)
2634                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2635                                 return 0;
2636         }
2637
2638         offset = find_pack_entry_one(sha1, p);
2639         if (!offset)
2640                 return 0;
2641
2642         /*
2643          * We are about to tell the caller where they can locate the
2644          * requested object.  We better make sure the packfile is
2645          * still here and can be accessed before supplying that
2646          * answer, as it may have been deleted since the index was
2647          * loaded!
2648          */
2649         if (!is_pack_valid(p))
2650                 return 0;
2651         e->offset = offset;
2652         e->p = p;
2653         hashcpy(e->sha1, sha1);
2654         return 1;
2655 }
2656
2657 /*
2658  * Iff a pack file contains the object named by sha1, return true and
2659  * store its location to e.
2660  */
2661 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2662 {
2663         struct mru_entry *p;
2664
2665         prepare_packed_git();
2666         if (!packed_git)
2667                 return 0;
2668
2669         for (p = packed_git_mru->head; p; p = p->next) {
2670                 if (fill_pack_entry(sha1, e, p->item)) {
2671                         mru_mark(packed_git_mru, p);
2672                         return 1;
2673                 }
2674         }
2675         return 0;
2676 }
2677
2678 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2679                                   struct packed_git *packs)
2680 {
2681         struct packed_git *p;
2682
2683         for (p = packs; p; p = p->next) {
2684                 if (find_pack_entry_one(sha1, p))
2685                         return p;
2686         }
2687         return NULL;
2688
2689 }
2690
2691 static int sha1_loose_object_info(const unsigned char *sha1,
2692                                   struct object_info *oi,
2693                                   int flags)
2694 {
2695         int status = 0;
2696         unsigned long mapsize;
2697         void *map;
2698         git_zstream stream;
2699         char hdr[32];
2700         struct strbuf hdrbuf = STRBUF_INIT;
2701
2702         if (oi->delta_base_sha1)
2703                 hashclr(oi->delta_base_sha1);
2704
2705         /*
2706          * If we don't care about type or size, then we don't
2707          * need to look inside the object at all. Note that we
2708          * do not optimize out the stat call, even if the
2709          * caller doesn't care about the disk-size, since our
2710          * return value implicitly indicates whether the
2711          * object even exists.
2712          */
2713         if (!oi->typep && !oi->typename && !oi->sizep) {
2714                 struct stat st;
2715                 if (stat_sha1_file(sha1, &st) < 0)
2716                         return -1;
2717                 if (oi->disk_sizep)
2718                         *oi->disk_sizep = st.st_size;
2719                 return 0;
2720         }
2721
2722         map = map_sha1_file(sha1, &mapsize);
2723         if (!map)
2724                 return -1;
2725         if (oi->disk_sizep)
2726                 *oi->disk_sizep = mapsize;
2727         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2728                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2729                         status = error("unable to unpack %s header with --allow-unknown-type",
2730                                        sha1_to_hex(sha1));
2731         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2732                 status = error("unable to unpack %s header",
2733                                sha1_to_hex(sha1));
2734         if (status < 0)
2735                 ; /* Do nothing */
2736         else if (hdrbuf.len) {
2737                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2738                         status = error("unable to parse %s header with --allow-unknown-type",
2739                                        sha1_to_hex(sha1));
2740         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2741                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2742         git_inflate_end(&stream);
2743         munmap(map, mapsize);
2744         if (status && oi->typep)
2745                 *oi->typep = status;
2746         strbuf_release(&hdrbuf);
2747         return 0;
2748 }
2749
2750 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2751 {
2752         struct cached_object *co;
2753         struct pack_entry e;
2754         int rtype;
2755         enum object_type real_type;
2756         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2757
2758         co = find_cached_object(real);
2759         if (co) {
2760                 if (oi->typep)
2761                         *(oi->typep) = co->type;
2762                 if (oi->sizep)
2763                         *(oi->sizep) = co->size;
2764                 if (oi->disk_sizep)
2765                         *(oi->disk_sizep) = 0;
2766                 if (oi->delta_base_sha1)
2767                         hashclr(oi->delta_base_sha1);
2768                 if (oi->typename)
2769                         strbuf_addstr(oi->typename, typename(co->type));
2770                 oi->whence = OI_CACHED;
2771                 return 0;
2772         }
2773
2774         if (!find_pack_entry(real, &e)) {
2775                 /* Most likely it's a loose object. */
2776                 if (!sha1_loose_object_info(real, oi, flags)) {
2777                         oi->whence = OI_LOOSE;
2778                         return 0;
2779                 }
2780
2781                 /* Not a loose object; someone else may have just packed it. */
2782                 reprepare_packed_git();
2783                 if (!find_pack_entry(real, &e))
2784                         return -1;
2785         }
2786
2787         /*
2788          * packed_object_info() does not follow the delta chain to
2789          * find out the real type, unless it is given oi->typep.
2790          */
2791         if (oi->typename && !oi->typep)
2792                 oi->typep = &real_type;
2793
2794         rtype = packed_object_info(e.p, e.offset, oi);
2795         if (rtype < 0) {
2796                 mark_bad_packed_object(e.p, real);
2797                 if (oi->typep == &real_type)
2798                         oi->typep = NULL;
2799                 return sha1_object_info_extended(real, oi, 0);
2800         } else if (in_delta_base_cache(e.p, e.offset)) {
2801                 oi->whence = OI_DBCACHED;
2802         } else {
2803                 oi->whence = OI_PACKED;
2804                 oi->u.packed.offset = e.offset;
2805                 oi->u.packed.pack = e.p;
2806                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2807                                          rtype == OBJ_OFS_DELTA);
2808         }
2809         if (oi->typename)
2810                 strbuf_addstr(oi->typename, typename(*oi->typep));
2811         if (oi->typep == &real_type)
2812                 oi->typep = NULL;
2813
2814         return 0;
2815 }
2816
2817 /* returns enum object_type or negative */
2818 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2819 {
2820         enum object_type type;
2821         struct object_info oi = {NULL};
2822
2823         oi.typep = &type;
2824         oi.sizep = sizep;
2825         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2826                 return -1;
2827         return type;
2828 }
2829
2830 static void *read_packed_sha1(const unsigned char *sha1,
2831                               enum object_type *type, unsigned long *size)
2832 {
2833         struct pack_entry e;
2834         void *data;
2835
2836         if (!find_pack_entry(sha1, &e))
2837                 return NULL;
2838         data = cache_or_unpack_entry(e.p, e.offset, size, type);
2839         if (!data) {
2840                 /*
2841                  * We're probably in deep shit, but let's try to fetch
2842                  * the required object anyway from another pack or loose.
2843                  * This should happen only in the presence of a corrupted
2844                  * pack, and is better than failing outright.
2845                  */
2846                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2847                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2848                 mark_bad_packed_object(e.p, sha1);
2849                 data = read_object(sha1, type, size);
2850         }
2851         return data;
2852 }
2853
2854 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2855                       unsigned char *sha1)
2856 {
2857         struct cached_object *co;
2858
2859         hash_sha1_file(buf, len, typename(type), sha1);
2860         if (has_sha1_file(sha1) || find_cached_object(sha1))
2861                 return 0;
2862         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2863         co = &cached_objects[cached_object_nr++];
2864         co->size = len;
2865         co->type = type;
2866         co->buf = xmalloc(len);
2867         memcpy(co->buf, buf, len);
2868         hashcpy(co->sha1, sha1);
2869         return 0;
2870 }
2871
2872 static void *read_object(const unsigned char *sha1, enum object_type *type,
2873                          unsigned long *size)
2874 {
2875         unsigned long mapsize;
2876         void *map, *buf;
2877         struct cached_object *co;
2878
2879         co = find_cached_object(sha1);
2880         if (co) {
2881                 *type = co->type;
2882                 *size = co->size;
2883                 return xmemdupz(co->buf, co->size);
2884         }
2885
2886         buf = read_packed_sha1(sha1, type, size);
2887         if (buf)
2888                 return buf;
2889         map = map_sha1_file(sha1, &mapsize);
2890         if (map) {
2891                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2892                 munmap(map, mapsize);
2893                 return buf;
2894         }
2895         reprepare_packed_git();
2896         return read_packed_sha1(sha1, type, size);
2897 }
2898
2899 /*
2900  * This function dies on corrupt objects; the callers who want to
2901  * deal with them should arrange to call read_object() and give error
2902  * messages themselves.
2903  */
2904 void *read_sha1_file_extended(const unsigned char *sha1,
2905                               enum object_type *type,
2906                               unsigned long *size,
2907                               unsigned flag)
2908 {
2909         void *data;
2910         const struct packed_git *p;
2911         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2912
2913         errno = 0;
2914         data = read_object(repl, type, size);
2915         if (data)
2916                 return data;
2917
2918         if (errno && errno != ENOENT)
2919                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2920
2921         /* die if we replaced an object with one that does not exist */
2922         if (repl != sha1)
2923                 die("replacement %s not found for %s",
2924                     sha1_to_hex(repl), sha1_to_hex(sha1));
2925
2926         if (has_loose_object(repl)) {
2927                 const char *path = sha1_file_name(sha1);
2928
2929                 die("loose object %s (stored in %s) is corrupt",
2930                     sha1_to_hex(repl), path);
2931         }
2932
2933         if ((p = has_packed_and_bad(repl)) != NULL)
2934                 die("packed object %s (stored in %s) is corrupt",
2935                     sha1_to_hex(repl), p->pack_name);
2936
2937         return NULL;
2938 }
2939
2940 void *read_object_with_reference(const unsigned char *sha1,
2941                                  const char *required_type_name,
2942                                  unsigned long *size,
2943                                  unsigned char *actual_sha1_return)
2944 {
2945         enum object_type type, required_type;
2946         void *buffer;
2947         unsigned long isize;
2948         unsigned char actual_sha1[20];
2949
2950         required_type = type_from_string(required_type_name);
2951         hashcpy(actual_sha1, sha1);
2952         while (1) {
2953                 int ref_length = -1;
2954                 const char *ref_type = NULL;
2955
2956                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2957                 if (!buffer)
2958                         return NULL;
2959                 if (type == required_type) {
2960                         *size = isize;
2961                         if (actual_sha1_return)
2962                                 hashcpy(actual_sha1_return, actual_sha1);
2963                         return buffer;
2964                 }
2965                 /* Handle references */
2966                 else if (type == OBJ_COMMIT)
2967                         ref_type = "tree ";
2968                 else if (type == OBJ_TAG)
2969                         ref_type = "object ";
2970                 else {
2971                         free(buffer);
2972                         return NULL;
2973                 }
2974                 ref_length = strlen(ref_type);
2975
2976                 if (ref_length + 40 > isize ||
2977                     memcmp(buffer, ref_type, ref_length) ||
2978                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2979                         free(buffer);
2980                         return NULL;
2981                 }
2982                 free(buffer);
2983                 /* Now we have the ID of the referred-to object in
2984                  * actual_sha1.  Check again. */
2985         }
2986 }
2987
2988 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2989                                     const char *type, unsigned char *sha1,
2990                                     char *hdr, int *hdrlen)
2991 {
2992         git_SHA_CTX c;
2993
2994         /* Generate the header */
2995         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
2996
2997         /* Sha1.. */
2998         git_SHA1_Init(&c);
2999         git_SHA1_Update(&c, hdr, *hdrlen);
3000         git_SHA1_Update(&c, buf, len);
3001         git_SHA1_Final(sha1, &c);
3002 }
3003
3004 /*
3005  * Move the just written object into its final resting place.
3006  */
3007 int finalize_object_file(const char *tmpfile, const char *filename)
3008 {
3009         int ret = 0;
3010
3011         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
3012                 goto try_rename;
3013         else if (link(tmpfile, filename))
3014                 ret = errno;
3015
3016         /*
3017          * Coda hack - coda doesn't like cross-directory links,
3018          * so we fall back to a rename, which will mean that it
3019          * won't be able to check collisions, but that's not a
3020          * big deal.
3021          *
3022          * The same holds for FAT formatted media.
3023          *
3024          * When this succeeds, we just return.  We have nothing
3025          * left to unlink.
3026          */
3027         if (ret && ret != EEXIST) {
3028         try_rename:
3029                 if (!rename(tmpfile, filename))
3030                         goto out;
3031                 ret = errno;
3032         }
3033         unlink_or_warn(tmpfile);
3034         if (ret) {
3035                 if (ret != EEXIST) {
3036                         return error_errno("unable to write sha1 filename %s", filename);
3037                 }
3038                 /* FIXME!!! Collision check here ? */
3039         }
3040
3041 out:
3042         if (adjust_shared_perm(filename))
3043                 return error("unable to set permission to '%s'", filename);
3044         return 0;
3045 }
3046
3047 static int write_buffer(int fd, const void *buf, size_t len)
3048 {
3049         if (write_in_full(fd, buf, len) < 0)
3050                 return error_errno("file write error");
3051         return 0;
3052 }
3053
3054 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3055                    unsigned char *sha1)
3056 {
3057         char hdr[32];
3058         int hdrlen = sizeof(hdr);
3059         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3060         return 0;
3061 }
3062
3063 /* Finalize a file on disk, and close it. */
3064 static void close_sha1_file(int fd)
3065 {
3066         if (fsync_object_files)
3067                 fsync_or_die(fd, "sha1 file");
3068         if (close(fd) != 0)
3069                 die_errno("error when closing sha1 file");
3070 }
3071
3072 /* Size of directory component, including the ending '/' */
3073 static inline int directory_size(const char *filename)
3074 {
3075         const char *s = strrchr(filename, '/');
3076         if (!s)
3077                 return 0;
3078         return s - filename + 1;
3079 }
3080
3081 /*
3082  * This creates a temporary file in the same directory as the final
3083  * 'filename'
3084  *
3085  * We want to avoid cross-directory filename renames, because those
3086  * can have problems on various filesystems (FAT, NFS, Coda).
3087  */
3088 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3089 {
3090         int fd, dirlen = directory_size(filename);
3091
3092         strbuf_reset(tmp);
3093         strbuf_add(tmp, filename, dirlen);
3094         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3095         fd = git_mkstemp_mode(tmp->buf, 0444);
3096         if (fd < 0 && dirlen && errno == ENOENT) {
3097                 /*
3098                  * Make sure the directory exists; note that the contents
3099                  * of the buffer are undefined after mkstemp returns an
3100                  * error, so we have to rewrite the whole buffer from
3101                  * scratch.
3102                  */
3103                 strbuf_reset(tmp);
3104                 strbuf_add(tmp, filename, dirlen - 1);
3105                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3106                         return -1;
3107                 if (adjust_shared_perm(tmp->buf))
3108                         return -1;
3109
3110                 /* Try again */
3111                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3112                 fd = git_mkstemp_mode(tmp->buf, 0444);
3113         }
3114         return fd;
3115 }
3116
3117 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3118                               const void *buf, unsigned long len, time_t mtime)
3119 {
3120         int fd, ret;
3121         unsigned char compressed[4096];
3122         git_zstream stream;
3123         git_SHA_CTX c;
3124         unsigned char parano_sha1[20];
3125         static struct strbuf tmp_file = STRBUF_INIT;
3126         const char *filename = sha1_file_name(sha1);
3127
3128         fd = create_tmpfile(&tmp_file, filename);
3129         if (fd < 0) {
3130                 if (errno == EACCES)
3131                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3132                 else
3133                         return error_errno("unable to create temporary file");
3134         }
3135
3136         /* Set it up */
3137         git_deflate_init(&stream, zlib_compression_level);
3138         stream.next_out = compressed;
3139         stream.avail_out = sizeof(compressed);
3140         git_SHA1_Init(&c);
3141
3142         /* First header.. */
3143         stream.next_in = (unsigned char *)hdr;
3144         stream.avail_in = hdrlen;
3145         while (git_deflate(&stream, 0) == Z_OK)
3146                 ; /* nothing */
3147         git_SHA1_Update(&c, hdr, hdrlen);
3148
3149         /* Then the data itself.. */
3150         stream.next_in = (void *)buf;
3151         stream.avail_in = len;
3152         do {
3153                 unsigned char *in0 = stream.next_in;
3154                 ret = git_deflate(&stream, Z_FINISH);
3155                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3156                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3157                         die("unable to write sha1 file");
3158                 stream.next_out = compressed;
3159                 stream.avail_out = sizeof(compressed);
3160         } while (ret == Z_OK);
3161
3162         if (ret != Z_STREAM_END)
3163                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3164         ret = git_deflate_end_gently(&stream);
3165         if (ret != Z_OK)
3166                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3167         git_SHA1_Final(parano_sha1, &c);
3168         if (hashcmp(sha1, parano_sha1) != 0)
3169                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3170
3171         close_sha1_file(fd);
3172
3173         if (mtime) {
3174                 struct utimbuf utb;
3175                 utb.actime = mtime;
3176                 utb.modtime = mtime;
3177                 if (utime(tmp_file.buf, &utb) < 0)
3178                         warning_errno("failed utime() on %s", tmp_file.buf);
3179         }
3180
3181         return finalize_object_file(tmp_file.buf, filename);
3182 }
3183
3184 static int freshen_loose_object(const unsigned char *sha1)
3185 {
3186         return check_and_freshen(sha1, 1);
3187 }
3188
3189 static int freshen_packed_object(const unsigned char *sha1)
3190 {
3191         struct pack_entry e;
3192         if (!find_pack_entry(sha1, &e))
3193                 return 0;
3194         if (e.p->freshened)
3195                 return 1;
3196         if (!freshen_file(e.p->pack_name))
3197                 return 0;
3198         e.p->freshened = 1;
3199         return 1;
3200 }
3201
3202 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3203 {
3204         char hdr[32];
3205         int hdrlen = sizeof(hdr);
3206
3207         /* Normally if we have it in the pack then we do not bother writing
3208          * it out into .git/objects/??/?{38} file.
3209          */
3210         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3211         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3212                 return 0;
3213         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3214 }
3215
3216 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3217                              unsigned char *sha1, unsigned flags)
3218 {
3219         char *header;
3220         int hdrlen, status = 0;
3221
3222         /* type string, SP, %lu of the length plus NUL must fit this */
3223         hdrlen = strlen(type) + 32;
3224         header = xmalloc(hdrlen);
3225         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3226
3227         if (!(flags & HASH_WRITE_OBJECT))
3228                 goto cleanup;
3229         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3230                 goto cleanup;
3231         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3232
3233 cleanup:
3234         free(header);
3235         return status;
3236 }
3237
3238 int force_object_loose(const unsigned char *sha1, time_t mtime)
3239 {
3240         void *buf;
3241         unsigned long len;
3242         enum object_type type;
3243         char hdr[32];
3244         int hdrlen;
3245         int ret;
3246
3247         if (has_loose_object(sha1))
3248                 return 0;
3249         buf = read_packed_sha1(sha1, &type, &len);
3250         if (!buf)
3251                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3252         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3253         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3254         free(buf);
3255
3256         return ret;
3257 }
3258
3259 int has_pack_index(const unsigned char *sha1)
3260 {
3261         struct stat st;
3262         if (stat(sha1_pack_index_name(sha1), &st))
3263                 return 0;
3264         return 1;
3265 }
3266
3267 int has_sha1_pack(const unsigned char *sha1)
3268 {
3269         struct pack_entry e;
3270         return find_pack_entry(sha1, &e);
3271 }
3272
3273 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3274 {
3275         struct pack_entry e;
3276
3277         if (find_pack_entry(sha1, &e))
3278                 return 1;
3279         if (has_loose_object(sha1))
3280                 return 1;
3281         if (flags & HAS_SHA1_QUICK)
3282                 return 0;
3283         reprepare_packed_git();
3284         return find_pack_entry(sha1, &e);
3285 }
3286
3287 int has_object_file(const struct object_id *oid)
3288 {
3289         return has_sha1_file(oid->hash);
3290 }
3291
3292 static void check_tree(const void *buf, size_t size)
3293 {
3294         struct tree_desc desc;
3295         struct name_entry entry;
3296
3297         init_tree_desc(&desc, buf, size);
3298         while (tree_entry(&desc, &entry))
3299                 /* do nothing
3300                  * tree_entry() will die() on malformed entries */
3301                 ;
3302 }
3303
3304 static void check_commit(const void *buf, size_t size)
3305 {
3306         struct commit c;
3307         memset(&c, 0, sizeof(c));
3308         if (parse_commit_buffer(&c, buf, size))
3309                 die("corrupt commit");
3310 }
3311
3312 static void check_tag(const void *buf, size_t size)
3313 {
3314         struct tag t;
3315         memset(&t, 0, sizeof(t));
3316         if (parse_tag_buffer(&t, buf, size))
3317                 die("corrupt tag");
3318 }
3319
3320 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3321                      enum object_type type,
3322                      const char *path, unsigned flags)
3323 {
3324         int ret, re_allocated = 0;
3325         int write_object = flags & HASH_WRITE_OBJECT;
3326
3327         if (!type)
3328                 type = OBJ_BLOB;
3329
3330         /*
3331          * Convert blobs to git internal format
3332          */
3333         if ((type == OBJ_BLOB) && path) {
3334                 struct strbuf nbuf = STRBUF_INIT;
3335                 if (convert_to_git(path, buf, size, &nbuf,
3336                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3337                         buf = strbuf_detach(&nbuf, &size);
3338                         re_allocated = 1;
3339                 }
3340         }
3341         if (flags & HASH_FORMAT_CHECK) {
3342                 if (type == OBJ_TREE)
3343                         check_tree(buf, size);
3344                 if (type == OBJ_COMMIT)
3345                         check_commit(buf, size);
3346                 if (type == OBJ_TAG)
3347                         check_tag(buf, size);
3348         }
3349
3350         if (write_object)
3351                 ret = write_sha1_file(buf, size, typename(type), sha1);
3352         else
3353                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3354         if (re_allocated)
3355                 free(buf);
3356         return ret;
3357 }
3358
3359 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3360                                      const char *path, unsigned flags)
3361 {
3362         int ret;
3363         const int write_object = flags & HASH_WRITE_OBJECT;
3364         struct strbuf sbuf = STRBUF_INIT;
3365
3366         assert(path);
3367         assert(would_convert_to_git_filter_fd(path));
3368
3369         convert_to_git_filter_fd(path, fd, &sbuf,
3370                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3371
3372         if (write_object)
3373                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3374                                       sha1);
3375         else
3376                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3377                                      sha1);
3378         strbuf_release(&sbuf);
3379         return ret;
3380 }
3381
3382 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3383                       const char *path, unsigned flags)
3384 {
3385         struct strbuf sbuf = STRBUF_INIT;
3386         int ret;
3387
3388         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3389                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3390         else
3391                 ret = -1;
3392         strbuf_release(&sbuf);
3393         return ret;
3394 }
3395
3396 #define SMALL_FILE_SIZE (32*1024)
3397
3398 static int index_core(unsigned char *sha1, int fd, size_t size,
3399                       enum object_type type, const char *path,
3400                       unsigned flags)
3401 {
3402         int ret;
3403
3404         if (!size) {
3405                 ret = index_mem(sha1, "", size, type, path, flags);
3406         } else if (size <= SMALL_FILE_SIZE) {
3407                 char *buf = xmalloc(size);
3408                 if (size == read_in_full(fd, buf, size))
3409                         ret = index_mem(sha1, buf, size, type, path, flags);
3410                 else
3411                         ret = error_errno("short read");
3412                 free(buf);
3413         } else {
3414                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3415                 ret = index_mem(sha1, buf, size, type, path, flags);
3416                 munmap(buf, size);
3417         }
3418         return ret;
3419 }
3420
3421 /*
3422  * This creates one packfile per large blob unless bulk-checkin
3423  * machinery is "plugged".
3424  *
3425  * This also bypasses the usual "convert-to-git" dance, and that is on
3426  * purpose. We could write a streaming version of the converting
3427  * functions and insert that before feeding the data to fast-import
3428  * (or equivalent in-core API described above). However, that is
3429  * somewhat complicated, as we do not know the size of the filter
3430  * result, which we need to know beforehand when writing a git object.
3431  * Since the primary motivation for trying to stream from the working
3432  * tree file and to avoid mmaping it in core is to deal with large
3433  * binary blobs, they generally do not want to get any conversion, and
3434  * callers should avoid this code path when filters are requested.
3435  */
3436 static int index_stream(unsigned char *sha1, int fd, size_t size,
3437                         enum object_type type, const char *path,
3438                         unsigned flags)
3439 {
3440         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3441 }
3442
3443 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3444              enum object_type type, const char *path, unsigned flags)
3445 {
3446         int ret;
3447
3448         /*
3449          * Call xsize_t() only when needed to avoid potentially unnecessary
3450          * die() for large files.
3451          */
3452         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3453                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3454         else if (!S_ISREG(st->st_mode))
3455                 ret = index_pipe(sha1, fd, type, path, flags);
3456         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3457                  (path && would_convert_to_git(path)))
3458                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3459                                  flags);
3460         else
3461                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3462                                    flags);
3463         close(fd);
3464         return ret;
3465 }
3466
3467 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3468 {
3469         int fd;
3470         struct strbuf sb = STRBUF_INIT;
3471
3472         switch (st->st_mode & S_IFMT) {
3473         case S_IFREG:
3474                 fd = open(path, O_RDONLY);
3475                 if (fd < 0)
3476                         return error_errno("open(\"%s\")", path);
3477                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3478                         return error("%s: failed to insert into database",
3479                                      path);
3480                 break;
3481         case S_IFLNK:
3482                 if (strbuf_readlink(&sb, path, st->st_size))
3483                         return error_errno("readlink(\"%s\")", path);
3484                 if (!(flags & HASH_WRITE_OBJECT))
3485                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3486                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3487                         return error("%s: failed to insert into database",
3488                                      path);
3489                 strbuf_release(&sb);
3490                 break;
3491         case S_IFDIR:
3492                 return resolve_gitlink_ref(path, "HEAD", sha1);
3493         default:
3494                 return error("%s: unsupported file type", path);
3495         }
3496         return 0;
3497 }
3498
3499 int read_pack_header(int fd, struct pack_header *header)
3500 {
3501         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3502                 /* "eof before pack header was fully read" */
3503                 return PH_ERROR_EOF;
3504
3505         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3506                 /* "protocol error (pack signature mismatch detected)" */
3507                 return PH_ERROR_PACK_SIGNATURE;
3508         if (!pack_version_ok(header->hdr_version))
3509                 /* "protocol error (pack version unsupported)" */
3510                 return PH_ERROR_PROTOCOL;
3511         return 0;
3512 }
3513
3514 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3515 {
3516         enum object_type type = sha1_object_info(sha1, NULL);
3517         if (type < 0)
3518                 die("%s is not a valid object", sha1_to_hex(sha1));
3519         if (type != expect)
3520                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3521                     typename(expect));
3522 }
3523
3524 static int for_each_file_in_obj_subdir(int subdir_nr,
3525                                        struct strbuf *path,
3526                                        each_loose_object_fn obj_cb,
3527                                        each_loose_cruft_fn cruft_cb,
3528                                        each_loose_subdir_fn subdir_cb,
3529                                        void *data)
3530 {
3531         size_t baselen = path->len;
3532         DIR *dir = opendir(path->buf);
3533         struct dirent *de;
3534         int r = 0;
3535
3536         if (!dir) {
3537                 if (errno == ENOENT)
3538                         return 0;
3539                 return error_errno("unable to open %s", path->buf);
3540         }
3541
3542         while ((de = readdir(dir))) {
3543                 if (is_dot_or_dotdot(de->d_name))
3544                         continue;
3545
3546                 strbuf_setlen(path, baselen);
3547                 strbuf_addf(path, "/%s", de->d_name);
3548
3549                 if (strlen(de->d_name) == 38)  {
3550                         char hex[41];
3551                         unsigned char sha1[20];
3552
3553                         snprintf(hex, sizeof(hex), "%02x%s",
3554                                  subdir_nr, de->d_name);
3555                         if (!get_sha1_hex(hex, sha1)) {
3556                                 if (obj_cb) {
3557                                         r = obj_cb(sha1, path->buf, data);
3558                                         if (r)
3559                                                 break;
3560                                 }
3561                                 continue;
3562                         }
3563                 }
3564
3565                 if (cruft_cb) {
3566                         r = cruft_cb(de->d_name, path->buf, data);
3567                         if (r)
3568                                 break;
3569                 }
3570         }
3571         closedir(dir);
3572
3573         strbuf_setlen(path, baselen);
3574         if (!r && subdir_cb)
3575                 r = subdir_cb(subdir_nr, path->buf, data);
3576
3577         return r;
3578 }
3579
3580 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3581                             each_loose_object_fn obj_cb,
3582                             each_loose_cruft_fn cruft_cb,
3583                             each_loose_subdir_fn subdir_cb,
3584                             void *data)
3585 {
3586         size_t baselen = path->len;
3587         int r = 0;
3588         int i;
3589
3590         for (i = 0; i < 256; i++) {
3591                 strbuf_addf(path, "/%02x", i);
3592                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3593                                                 subdir_cb, data);
3594                 strbuf_setlen(path, baselen);
3595                 if (r)
3596                         break;
3597         }
3598
3599         return r;
3600 }
3601
3602 int for_each_loose_file_in_objdir(const char *path,
3603                                   each_loose_object_fn obj_cb,
3604                                   each_loose_cruft_fn cruft_cb,
3605                                   each_loose_subdir_fn subdir_cb,
3606                                   void *data)
3607 {
3608         struct strbuf buf = STRBUF_INIT;
3609         int r;
3610
3611         strbuf_addstr(&buf, path);
3612         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3613                                               subdir_cb, data);
3614         strbuf_release(&buf);
3615
3616         return r;
3617 }
3618
3619 struct loose_alt_odb_data {
3620         each_loose_object_fn *cb;
3621         void *data;
3622 };
3623
3624 static int loose_from_alt_odb(struct alternate_object_database *alt,
3625                               void *vdata)
3626 {
3627         struct loose_alt_odb_data *data = vdata;
3628         struct strbuf buf = STRBUF_INIT;
3629         int r;
3630
3631         /* copy base not including trailing '/' */
3632         strbuf_add(&buf, alt->base, alt->name - alt->base - 1);
3633         r = for_each_loose_file_in_objdir_buf(&buf,
3634                                               data->cb, NULL, NULL,
3635                                               data->data);
3636         strbuf_release(&buf);
3637         return r;
3638 }
3639
3640 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3641 {
3642         struct loose_alt_odb_data alt;
3643         int r;
3644
3645         r = for_each_loose_file_in_objdir(get_object_directory(),
3646                                           cb, NULL, NULL, data);
3647         if (r)
3648                 return r;
3649
3650         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3651                 return 0;
3652
3653         alt.cb = cb;
3654         alt.data = data;
3655         return foreach_alt_odb(loose_from_alt_odb, &alt);
3656 }
3657
3658 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3659 {
3660         uint32_t i;
3661         int r = 0;
3662
3663         for (i = 0; i < p->num_objects; i++) {
3664                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3665
3666                 if (!sha1)
3667                         return error("unable to get sha1 of object %u in %s",
3668                                      i, p->pack_name);
3669
3670                 r = cb(sha1, p, i, data);
3671                 if (r)
3672                         break;
3673         }
3674         return r;
3675 }
3676
3677 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3678 {
3679         struct packed_git *p;
3680         int r = 0;
3681         int pack_errors = 0;
3682
3683         prepare_packed_git();
3684         for (p = packed_git; p; p = p->next) {
3685                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3686                         continue;
3687                 if (open_pack_index(p)) {
3688                         pack_errors = 1;
3689                         continue;
3690                 }
3691                 r = for_each_object_in_pack(p, cb, data);
3692                 if (r)
3693                         break;
3694         }
3695         return r ? r : pack_errors;
3696 }