Sync with maint
[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
27 #ifndef O_NOATIME
28 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
29 #define O_NOATIME 01000000
30 #else
31 #define O_NOATIME 0
32 #endif
33 #endif
34
35 #define SZ_FMT PRIuMAX
36 static inline uintmax_t sz_fmt(size_t s) { return s; }
37
38 const unsigned char null_sha1[20];
39
40 /*
41  * This is meant to hold a *small* number of objects that you would
42  * want read_sha1_file() to be able to return, but yet you do not want
43  * to write them into the object store (e.g. a browse-only
44  * application).
45  */
46 static struct cached_object {
47         unsigned char sha1[20];
48         enum object_type type;
49         void *buf;
50         unsigned long size;
51 } *cached_objects;
52 static int cached_object_nr, cached_object_alloc;
53
54 static struct cached_object empty_tree = {
55         EMPTY_TREE_SHA1_BIN_LITERAL,
56         OBJ_TREE,
57         "",
58         0
59 };
60
61 /*
62  * A pointer to the last packed_git in which an object was found.
63  * When an object is sought, we look in this packfile first, because
64  * objects that are looked up at similar times are often in the same
65  * packfile as one another.
66  */
67 static struct packed_git *last_found_pack;
68
69 static struct cached_object *find_cached_object(const unsigned char *sha1)
70 {
71         int i;
72         struct cached_object *co = cached_objects;
73
74         for (i = 0; i < cached_object_nr; i++, co++) {
75                 if (!hashcmp(co->sha1, sha1))
76                         return co;
77         }
78         if (!hashcmp(sha1, empty_tree.sha1))
79                 return &empty_tree;
80         return NULL;
81 }
82
83 int mkdir_in_gitdir(const char *path)
84 {
85         if (mkdir(path, 0777)) {
86                 int saved_errno = errno;
87                 struct stat st;
88                 struct strbuf sb = STRBUF_INIT;
89
90                 if (errno != EEXIST)
91                         return -1;
92                 /*
93                  * Are we looking at a path in a symlinked worktree
94                  * whose original repository does not yet have it?
95                  * e.g. .git/rr-cache pointing at its original
96                  * repository in which the user hasn't performed any
97                  * conflict resolution yet?
98                  */
99                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
100                     strbuf_readlink(&sb, path, st.st_size) ||
101                     !is_absolute_path(sb.buf) ||
102                     mkdir(sb.buf, 0777)) {
103                         strbuf_release(&sb);
104                         errno = saved_errno;
105                         return -1;
106                 }
107                 strbuf_release(&sb);
108         }
109         return adjust_shared_perm(path);
110 }
111
112 enum scld_error safe_create_leading_directories(char *path)
113 {
114         char *next_component = path + offset_1st_component(path);
115         enum scld_error ret = SCLD_OK;
116
117         while (ret == SCLD_OK && next_component) {
118                 struct stat st;
119                 char *slash = next_component, slash_character;
120
121                 while (*slash && !is_dir_sep(*slash))
122                         slash++;
123
124                 if (!*slash)
125                         break;
126
127                 next_component = slash + 1;
128                 while (is_dir_sep(*next_component))
129                         next_component++;
130                 if (!*next_component)
131                         break;
132
133                 slash_character = *slash;
134                 *slash = '\0';
135                 if (!stat(path, &st)) {
136                         /* path exists */
137                         if (!S_ISDIR(st.st_mode))
138                                 ret = SCLD_EXISTS;
139                 } else if (mkdir(path, 0777)) {
140                         if (errno == EEXIST &&
141                             !stat(path, &st) && S_ISDIR(st.st_mode))
142                                 ; /* somebody created it since we checked */
143                         else if (errno == ENOENT)
144                                 /*
145                                  * Either mkdir() failed because
146                                  * somebody just pruned the containing
147                                  * directory, or stat() failed because
148                                  * the file that was in our way was
149                                  * just removed.  Either way, inform
150                                  * the caller that it might be worth
151                                  * trying again:
152                                  */
153                                 ret = SCLD_VANISHED;
154                         else
155                                 ret = SCLD_FAILED;
156                 } else if (adjust_shared_perm(path)) {
157                         ret = SCLD_PERMS;
158                 }
159                 *slash = slash_character;
160         }
161         return ret;
162 }
163
164 enum scld_error safe_create_leading_directories_const(const char *path)
165 {
166         /* path points to cache entries, so xstrdup before messing with it */
167         char *buf = xstrdup(path);
168         enum scld_error result = safe_create_leading_directories(buf);
169         free(buf);
170         return result;
171 }
172
173 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
174 {
175         int i;
176         for (i = 0; i < 20; i++) {
177                 static char hex[] = "0123456789abcdef";
178                 unsigned int val = sha1[i];
179                 char *pos = pathbuf + i*2 + (i > 0);
180                 *pos++ = hex[val >> 4];
181                 *pos = hex[val & 0xf];
182         }
183 }
184
185 const char *sha1_file_name(const unsigned char *sha1)
186 {
187         static char buf[PATH_MAX];
188         const char *objdir;
189         int len;
190
191         objdir = get_object_directory();
192         len = strlen(objdir);
193
194         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
195         if (len + 43 > PATH_MAX)
196                 die("insanely long object directory %s", objdir);
197         memcpy(buf, objdir, len);
198         buf[len] = '/';
199         buf[len+3] = '/';
200         buf[len+42] = '\0';
201         fill_sha1_path(buf + len + 1, sha1);
202         return buf;
203 }
204
205 /*
206  * Return the name of the pack or index file with the specified sha1
207  * in its filename.  *base and *name are scratch space that must be
208  * provided by the caller.  which should be "pack" or "idx".
209  */
210 static char *sha1_get_pack_name(const unsigned char *sha1,
211                                 char **name, char **base, const char *which)
212 {
213         static const char hex[] = "0123456789abcdef";
214         char *buf;
215         int i;
216
217         if (!*base) {
218                 const char *sha1_file_directory = get_object_directory();
219                 int len = strlen(sha1_file_directory);
220                 *base = xmalloc(len + 60);
221                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
222                         sha1_file_directory, which);
223                 *name = *base + len + 11;
224         }
225
226         buf = *name;
227
228         for (i = 0; i < 20; i++) {
229                 unsigned int val = *sha1++;
230                 *buf++ = hex[val >> 4];
231                 *buf++ = hex[val & 0xf];
232         }
233
234         return *base;
235 }
236
237 char *sha1_pack_name(const unsigned char *sha1)
238 {
239         static char *name, *base;
240
241         return sha1_get_pack_name(sha1, &name, &base, "pack");
242 }
243
244 char *sha1_pack_index_name(const unsigned char *sha1)
245 {
246         static char *name, *base;
247
248         return sha1_get_pack_name(sha1, &name, &base, "idx");
249 }
250
251 struct alternate_object_database *alt_odb_list;
252 static struct alternate_object_database **alt_odb_tail;
253
254 /*
255  * Prepare alternate object database registry.
256  *
257  * The variable alt_odb_list points at the list of struct
258  * alternate_object_database.  The elements on this list come from
259  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
260  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
261  * whose contents is similar to that environment variable but can be
262  * LF separated.  Its base points at a statically allocated buffer that
263  * contains "/the/directory/corresponding/to/.git/objects/...", while
264  * its name points just after the slash at the end of ".git/objects/"
265  * in the example above, and has enough space to hold 40-byte hex
266  * SHA1, an extra slash for the first level indirection, and the
267  * terminating NUL.
268  */
269 static int link_alt_odb_entry(const char *entry, const char *relative_base,
270         int depth, const char *normalized_objdir)
271 {
272         struct alternate_object_database *ent;
273         struct alternate_object_database *alt;
274         int pfxlen, entlen;
275         struct strbuf pathbuf = STRBUF_INIT;
276
277         if (!is_absolute_path(entry) && relative_base) {
278                 strbuf_addstr(&pathbuf, real_path(relative_base));
279                 strbuf_addch(&pathbuf, '/');
280         }
281         strbuf_addstr(&pathbuf, entry);
282
283         normalize_path_copy(pathbuf.buf, pathbuf.buf);
284
285         pfxlen = strlen(pathbuf.buf);
286
287         /*
288          * The trailing slash after the directory name is given by
289          * this function at the end. Remove duplicates.
290          */
291         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
292                 pfxlen -= 1;
293
294         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
295         ent = xmalloc(sizeof(*ent) + entlen);
296         memcpy(ent->base, pathbuf.buf, pfxlen);
297         strbuf_release(&pathbuf);
298
299         ent->name = ent->base + pfxlen + 1;
300         ent->base[pfxlen + 3] = '/';
301         ent->base[pfxlen] = ent->base[entlen-1] = 0;
302
303         /* Detect cases where alternate disappeared */
304         if (!is_directory(ent->base)) {
305                 error("object directory %s does not exist; "
306                       "check .git/objects/info/alternates.",
307                       ent->base);
308                 free(ent);
309                 return -1;
310         }
311
312         /* Prevent the common mistake of listing the same
313          * thing twice, or object directory itself.
314          */
315         for (alt = alt_odb_list; alt; alt = alt->next) {
316                 if (pfxlen == alt->name - alt->base - 1 &&
317                     !memcmp(ent->base, alt->base, pfxlen)) {
318                         free(ent);
319                         return -1;
320                 }
321         }
322         if (!strcmp_icase(ent->base, normalized_objdir)) {
323                 free(ent);
324                 return -1;
325         }
326
327         /* add the alternate entry */
328         *alt_odb_tail = ent;
329         alt_odb_tail = &(ent->next);
330         ent->next = NULL;
331
332         /* recursively add alternates */
333         read_info_alternates(ent->base, depth + 1);
334
335         ent->base[pfxlen] = '/';
336
337         return 0;
338 }
339
340 static void link_alt_odb_entries(const char *alt, int len, int sep,
341                                  const char *relative_base, int depth)
342 {
343         struct string_list entries = STRING_LIST_INIT_NODUP;
344         char *alt_copy;
345         int i;
346         struct strbuf objdirbuf = STRBUF_INIT;
347
348         if (depth > 5) {
349                 error("%s: ignoring alternate object stores, nesting too deep.",
350                                 relative_base);
351                 return;
352         }
353
354         strbuf_add_absolute_path(&objdirbuf, get_object_directory());
355         normalize_path_copy(objdirbuf.buf, objdirbuf.buf);
356
357         alt_copy = xmemdupz(alt, len);
358         string_list_split_in_place(&entries, alt_copy, sep, -1);
359         for (i = 0; i < entries.nr; i++) {
360                 const char *entry = entries.items[i].string;
361                 if (entry[0] == '\0' || entry[0] == '#')
362                         continue;
363                 if (!is_absolute_path(entry) && depth) {
364                         error("%s: ignoring relative alternate object store %s",
365                                         relative_base, entry);
366                 } else {
367                         link_alt_odb_entry(entry, relative_base, depth, objdirbuf.buf);
368                 }
369         }
370         string_list_clear(&entries, 0);
371         free(alt_copy);
372         strbuf_release(&objdirbuf);
373 }
374
375 void read_info_alternates(const char * relative_base, int depth)
376 {
377         char *map;
378         size_t mapsz;
379         struct stat st;
380         const char alt_file_name[] = "info/alternates";
381         /* Given that relative_base is no longer than PATH_MAX,
382            ensure that "path" has enough space to append "/", the
383            file name, "info/alternates", and a trailing NUL.  */
384         char path[PATH_MAX + 1 + sizeof alt_file_name];
385         int fd;
386
387         sprintf(path, "%s/%s", relative_base, alt_file_name);
388         fd = git_open_noatime(path);
389         if (fd < 0)
390                 return;
391         if (fstat(fd, &st) || (st.st_size == 0)) {
392                 close(fd);
393                 return;
394         }
395         mapsz = xsize_t(st.st_size);
396         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
397         close(fd);
398
399         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
400
401         munmap(map, mapsz);
402 }
403
404 void add_to_alternates_file(const char *reference)
405 {
406         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
407         char *alts = git_pathdup("objects/info/alternates");
408         FILE *in, *out;
409
410         hold_lock_file_for_update(lock, alts, LOCK_DIE_ON_ERROR);
411         out = fdopen_lock_file(lock, "w");
412         if (!out)
413                 die_errno("unable to fdopen alternates lockfile");
414
415         in = fopen(alts, "r");
416         if (in) {
417                 struct strbuf line = STRBUF_INIT;
418                 int found = 0;
419
420                 while (strbuf_getline(&line, in, '\n') != EOF) {
421                         if (!strcmp(reference, line.buf)) {
422                                 found = 1;
423                                 break;
424                         }
425                         fprintf_or_die(out, "%s\n", line.buf);
426                 }
427
428                 strbuf_release(&line);
429                 fclose(in);
430
431                 if (found) {
432                         rollback_lock_file(lock);
433                         lock = NULL;
434                 }
435         }
436         else if (errno != ENOENT)
437                 die_errno("unable to read alternates file");
438
439         if (lock) {
440                 fprintf_or_die(out, "%s\n", reference);
441                 if (commit_lock_file(lock))
442                         die_errno("unable to move new alternates file into place");
443                 if (alt_odb_tail)
444                         link_alt_odb_entries(reference, strlen(reference), '\n', NULL, 0);
445         }
446         free(alts);
447 }
448
449 int foreach_alt_odb(alt_odb_fn fn, void *cb)
450 {
451         struct alternate_object_database *ent;
452         int r = 0;
453
454         prepare_alt_odb();
455         for (ent = alt_odb_list; ent; ent = ent->next) {
456                 r = fn(ent, cb);
457                 if (r)
458                         break;
459         }
460         return r;
461 }
462
463 void prepare_alt_odb(void)
464 {
465         const char *alt;
466
467         if (alt_odb_tail)
468                 return;
469
470         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
471         if (!alt) alt = "";
472
473         alt_odb_tail = &alt_odb_list;
474         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
475
476         read_info_alternates(get_object_directory(), 0);
477 }
478
479 /* Returns 1 if we have successfully freshened the file, 0 otherwise. */
480 static int freshen_file(const char *fn)
481 {
482         struct utimbuf t;
483         t.actime = t.modtime = time(NULL);
484         return !utime(fn, &t);
485 }
486
487 /*
488  * All of the check_and_freshen functions return 1 if the file exists and was
489  * freshened (if freshening was requested), 0 otherwise. If they return
490  * 0, you should not assume that it is safe to skip a write of the object (it
491  * either does not exist on disk, or has a stale mtime and may be subject to
492  * pruning).
493  */
494 static int check_and_freshen_file(const char *fn, int freshen)
495 {
496         if (access(fn, F_OK))
497                 return 0;
498         if (freshen && !freshen_file(fn))
499                 return 0;
500         return 1;
501 }
502
503 static int check_and_freshen_local(const unsigned char *sha1, int freshen)
504 {
505         return check_and_freshen_file(sha1_file_name(sha1), freshen);
506 }
507
508 static int check_and_freshen_nonlocal(const unsigned char *sha1, int freshen)
509 {
510         struct alternate_object_database *alt;
511         prepare_alt_odb();
512         for (alt = alt_odb_list; alt; alt = alt->next) {
513                 fill_sha1_path(alt->name, sha1);
514                 if (check_and_freshen_file(alt->base, freshen))
515                         return 1;
516         }
517         return 0;
518 }
519
520 static int check_and_freshen(const unsigned char *sha1, int freshen)
521 {
522         return check_and_freshen_local(sha1, freshen) ||
523                check_and_freshen_nonlocal(sha1, freshen);
524 }
525
526 int has_loose_object_nonlocal(const unsigned char *sha1)
527 {
528         return check_and_freshen_nonlocal(sha1, 0);
529 }
530
531 static int has_loose_object(const unsigned char *sha1)
532 {
533         return check_and_freshen(sha1, 0);
534 }
535
536 static unsigned int pack_used_ctr;
537 static unsigned int pack_mmap_calls;
538 static unsigned int peak_pack_open_windows;
539 static unsigned int pack_open_windows;
540 static unsigned int pack_open_fds;
541 static unsigned int pack_max_fds;
542 static size_t peak_pack_mapped;
543 static size_t pack_mapped;
544 struct packed_git *packed_git;
545
546 void pack_report(void)
547 {
548         fprintf(stderr,
549                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
550                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
551                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
552                 sz_fmt(getpagesize()),
553                 sz_fmt(packed_git_window_size),
554                 sz_fmt(packed_git_limit));
555         fprintf(stderr,
556                 "pack_report: pack_used_ctr            = %10u\n"
557                 "pack_report: pack_mmap_calls          = %10u\n"
558                 "pack_report: pack_open_windows        = %10u / %10u\n"
559                 "pack_report: pack_mapped              = "
560                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
561                 pack_used_ctr,
562                 pack_mmap_calls,
563                 pack_open_windows, peak_pack_open_windows,
564                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
565 }
566
567 /*
568  * Open and mmap the index file at path, perform a couple of
569  * consistency checks, then record its information to p.  Return 0 on
570  * success.
571  */
572 static int check_packed_git_idx(const char *path, struct packed_git *p)
573 {
574         void *idx_map;
575         struct pack_idx_header *hdr;
576         size_t idx_size;
577         uint32_t version, nr, i, *index;
578         int fd = git_open_noatime(path);
579         struct stat st;
580
581         if (fd < 0)
582                 return -1;
583         if (fstat(fd, &st)) {
584                 close(fd);
585                 return -1;
586         }
587         idx_size = xsize_t(st.st_size);
588         if (idx_size < 4 * 256 + 20 + 20) {
589                 close(fd);
590                 return error("index file %s is too small", path);
591         }
592         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
593         close(fd);
594
595         hdr = idx_map;
596         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
597                 version = ntohl(hdr->idx_version);
598                 if (version < 2 || version > 2) {
599                         munmap(idx_map, idx_size);
600                         return error("index file %s is version %"PRIu32
601                                      " and is not supported by this binary"
602                                      " (try upgrading GIT to a newer version)",
603                                      path, version);
604                 }
605         } else
606                 version = 1;
607
608         nr = 0;
609         index = idx_map;
610         if (version > 1)
611                 index += 2;  /* skip index header */
612         for (i = 0; i < 256; i++) {
613                 uint32_t n = ntohl(index[i]);
614                 if (n < nr) {
615                         munmap(idx_map, idx_size);
616                         return error("non-monotonic index %s", path);
617                 }
618                 nr = n;
619         }
620
621         if (version == 1) {
622                 /*
623                  * Total size:
624                  *  - 256 index entries 4 bytes each
625                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
626                  *  - 20-byte SHA1 of the packfile
627                  *  - 20-byte SHA1 file checksum
628                  */
629                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
630                         munmap(idx_map, idx_size);
631                         return error("wrong index v1 file size in %s", path);
632                 }
633         } else if (version == 2) {
634                 /*
635                  * Minimum size:
636                  *  - 8 bytes of header
637                  *  - 256 index entries 4 bytes each
638                  *  - 20-byte sha1 entry * nr
639                  *  - 4-byte crc entry * nr
640                  *  - 4-byte offset entry * nr
641                  *  - 20-byte SHA1 of the packfile
642                  *  - 20-byte SHA1 file checksum
643                  * And after the 4-byte offset table might be a
644                  * variable sized table containing 8-byte entries
645                  * for offsets larger than 2^31.
646                  */
647                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
648                 unsigned long max_size = min_size;
649                 if (nr)
650                         max_size += (nr - 1)*8;
651                 if (idx_size < min_size || idx_size > max_size) {
652                         munmap(idx_map, idx_size);
653                         return error("wrong index v2 file size in %s", path);
654                 }
655                 if (idx_size != min_size &&
656                     /*
657                      * make sure we can deal with large pack offsets.
658                      * 31-bit signed offset won't be enough, neither
659                      * 32-bit unsigned one will be.
660                      */
661                     (sizeof(off_t) <= 4)) {
662                         munmap(idx_map, idx_size);
663                         return error("pack too large for current definition of off_t in %s", path);
664                 }
665         }
666
667         p->index_version = version;
668         p->index_data = idx_map;
669         p->index_size = idx_size;
670         p->num_objects = nr;
671         return 0;
672 }
673
674 int open_pack_index(struct packed_git *p)
675 {
676         char *idx_name;
677         int ret;
678
679         if (p->index_data)
680                 return 0;
681
682         idx_name = xstrdup(p->pack_name);
683         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
684         ret = check_packed_git_idx(idx_name, p);
685         free(idx_name);
686         return ret;
687 }
688
689 static void scan_windows(struct packed_git *p,
690         struct packed_git **lru_p,
691         struct pack_window **lru_w,
692         struct pack_window **lru_l)
693 {
694         struct pack_window *w, *w_l;
695
696         for (w_l = NULL, w = p->windows; w; w = w->next) {
697                 if (!w->inuse_cnt) {
698                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
699                                 *lru_p = p;
700                                 *lru_w = w;
701                                 *lru_l = w_l;
702                         }
703                 }
704                 w_l = w;
705         }
706 }
707
708 static int unuse_one_window(struct packed_git *current)
709 {
710         struct packed_git *p, *lru_p = NULL;
711         struct pack_window *lru_w = NULL, *lru_l = NULL;
712
713         if (current)
714                 scan_windows(current, &lru_p, &lru_w, &lru_l);
715         for (p = packed_git; p; p = p->next)
716                 scan_windows(p, &lru_p, &lru_w, &lru_l);
717         if (lru_p) {
718                 munmap(lru_w->base, lru_w->len);
719                 pack_mapped -= lru_w->len;
720                 if (lru_l)
721                         lru_l->next = lru_w->next;
722                 else
723                         lru_p->windows = lru_w->next;
724                 free(lru_w);
725                 pack_open_windows--;
726                 return 1;
727         }
728         return 0;
729 }
730
731 void release_pack_memory(size_t need)
732 {
733         size_t cur = pack_mapped;
734         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
735                 ; /* nothing */
736 }
737
738 static void mmap_limit_check(size_t length)
739 {
740         static size_t limit = 0;
741         if (!limit) {
742                 limit = git_env_ulong("GIT_MMAP_LIMIT", 0);
743                 if (!limit)
744                         limit = SIZE_MAX;
745         }
746         if (length > limit)
747                 die("attempting to mmap %"PRIuMAX" over limit %"PRIuMAX,
748                     (uintmax_t)length, (uintmax_t)limit);
749 }
750
751 void *xmmap_gently(void *start, size_t length,
752                   int prot, int flags, int fd, off_t offset)
753 {
754         void *ret;
755
756         mmap_limit_check(length);
757         ret = mmap(start, length, prot, flags, fd, offset);
758         if (ret == MAP_FAILED) {
759                 if (!length)
760                         return NULL;
761                 release_pack_memory(length);
762                 ret = mmap(start, length, prot, flags, fd, offset);
763         }
764         return ret;
765 }
766
767 void *xmmap(void *start, size_t length,
768         int prot, int flags, int fd, off_t offset)
769 {
770         void *ret = xmmap_gently(start, length, prot, flags, fd, offset);
771         if (ret == MAP_FAILED)
772                 die_errno("mmap failed");
773         return ret;
774 }
775
776 void close_pack_windows(struct packed_git *p)
777 {
778         while (p->windows) {
779                 struct pack_window *w = p->windows;
780
781                 if (w->inuse_cnt)
782                         die("pack '%s' still has open windows to it",
783                             p->pack_name);
784                 munmap(w->base, w->len);
785                 pack_mapped -= w->len;
786                 pack_open_windows--;
787                 p->windows = w->next;
788                 free(w);
789         }
790 }
791
792 /*
793  * The LRU pack is the one with the oldest MRU window, preferring packs
794  * with no used windows, or the oldest mtime if it has no windows allocated.
795  */
796 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
797 {
798         struct pack_window *w, *this_mru_w;
799         int has_windows_inuse = 0;
800
801         /*
802          * Reject this pack if it has windows and the previously selected
803          * one does not.  If this pack does not have windows, reject
804          * it if the pack file is newer than the previously selected one.
805          */
806         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
807                 return;
808
809         for (w = this_mru_w = p->windows; w; w = w->next) {
810                 /*
811                  * Reject this pack if any of its windows are in use,
812                  * but the previously selected pack did not have any
813                  * inuse windows.  Otherwise, record that this pack
814                  * has windows in use.
815                  */
816                 if (w->inuse_cnt) {
817                         if (*accept_windows_inuse)
818                                 has_windows_inuse = 1;
819                         else
820                                 return;
821                 }
822
823                 if (w->last_used > this_mru_w->last_used)
824                         this_mru_w = w;
825
826                 /*
827                  * Reject this pack if it has windows that have been
828                  * used more recently than the previously selected pack.
829                  * If the previously selected pack had windows inuse and
830                  * we have not encountered a window in this pack that is
831                  * inuse, skip this check since we prefer a pack with no
832                  * inuse windows to one that has inuse windows.
833                  */
834                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
835                     this_mru_w->last_used > (*mru_w)->last_used)
836                         return;
837         }
838
839         /*
840          * Select this pack.
841          */
842         *mru_w = this_mru_w;
843         *lru_p = p;
844         *accept_windows_inuse = has_windows_inuse;
845 }
846
847 static int close_one_pack(void)
848 {
849         struct packed_git *p, *lru_p = NULL;
850         struct pack_window *mru_w = NULL;
851         int accept_windows_inuse = 1;
852
853         for (p = packed_git; p; p = p->next) {
854                 if (p->pack_fd == -1)
855                         continue;
856                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
857         }
858
859         if (lru_p) {
860                 close(lru_p->pack_fd);
861                 pack_open_fds--;
862                 lru_p->pack_fd = -1;
863                 return 1;
864         }
865
866         return 0;
867 }
868
869 void unuse_pack(struct pack_window **w_cursor)
870 {
871         struct pack_window *w = *w_cursor;
872         if (w) {
873                 w->inuse_cnt--;
874                 *w_cursor = NULL;
875         }
876 }
877
878 void close_pack_index(struct packed_git *p)
879 {
880         if (p->index_data) {
881                 munmap((void *)p->index_data, p->index_size);
882                 p->index_data = NULL;
883         }
884 }
885
886 /*
887  * This is used by git-repack in case a newly created pack happens to
888  * contain the same set of objects as an existing one.  In that case
889  * the resulting file might be different even if its name would be the
890  * same.  It is best to close any reference to the old pack before it is
891  * replaced on disk.  Of course no index pointers or windows for given pack
892  * must subsist at this point.  If ever objects from this pack are requested
893  * again, the new version of the pack will be reinitialized through
894  * reprepare_packed_git().
895  */
896 void free_pack_by_name(const char *pack_name)
897 {
898         struct packed_git *p, **pp = &packed_git;
899
900         while (*pp) {
901                 p = *pp;
902                 if (strcmp(pack_name, p->pack_name) == 0) {
903                         clear_delta_base_cache();
904                         close_pack_windows(p);
905                         if (p->pack_fd != -1) {
906                                 close(p->pack_fd);
907                                 pack_open_fds--;
908                         }
909                         close_pack_index(p);
910                         free(p->bad_object_sha1);
911                         *pp = p->next;
912                         if (last_found_pack == p)
913                                 last_found_pack = NULL;
914                         free(p);
915                         return;
916                 }
917                 pp = &p->next;
918         }
919 }
920
921 static unsigned int get_max_fd_limit(void)
922 {
923 #ifdef RLIMIT_NOFILE
924         {
925                 struct rlimit lim;
926
927                 if (!getrlimit(RLIMIT_NOFILE, &lim))
928                         return lim.rlim_cur;
929         }
930 #endif
931
932 #ifdef _SC_OPEN_MAX
933         {
934                 long open_max = sysconf(_SC_OPEN_MAX);
935                 if (0 < open_max)
936                         return open_max;
937                 /*
938                  * Otherwise, we got -1 for one of the two
939                  * reasons:
940                  *
941                  * (1) sysconf() did not understand _SC_OPEN_MAX
942                  *     and signaled an error with -1; or
943                  * (2) sysconf() said there is no limit.
944                  *
945                  * We _could_ clear errno before calling sysconf() to
946                  * tell these two cases apart and return a huge number
947                  * in the latter case to let the caller cap it to a
948                  * value that is not so selfish, but letting the
949                  * fallback OPEN_MAX codepath take care of these cases
950                  * is a lot simpler.
951                  */
952         }
953 #endif
954
955 #ifdef OPEN_MAX
956         return OPEN_MAX;
957 #else
958         return 1; /* see the caller ;-) */
959 #endif
960 }
961
962 /*
963  * Do not call this directly as this leaks p->pack_fd on error return;
964  * call open_packed_git() instead.
965  */
966 static int open_packed_git_1(struct packed_git *p)
967 {
968         struct stat st;
969         struct pack_header hdr;
970         unsigned char sha1[20];
971         unsigned char *idx_sha1;
972         long fd_flag;
973
974         if (!p->index_data && open_pack_index(p))
975                 return error("packfile %s index unavailable", p->pack_name);
976
977         if (!pack_max_fds) {
978                 unsigned int max_fds = get_max_fd_limit();
979
980                 /* Save 3 for stdin/stdout/stderr, 22 for work */
981                 if (25 < max_fds)
982                         pack_max_fds = max_fds - 25;
983                 else
984                         pack_max_fds = 1;
985         }
986
987         while (pack_max_fds <= pack_open_fds && close_one_pack())
988                 ; /* nothing */
989
990         p->pack_fd = git_open_noatime(p->pack_name);
991         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
992                 return -1;
993         pack_open_fds++;
994
995         /* If we created the struct before we had the pack we lack size. */
996         if (!p->pack_size) {
997                 if (!S_ISREG(st.st_mode))
998                         return error("packfile %s not a regular file", p->pack_name);
999                 p->pack_size = st.st_size;
1000         } else if (p->pack_size != st.st_size)
1001                 return error("packfile %s size changed", p->pack_name);
1002
1003         /* We leave these file descriptors open with sliding mmap;
1004          * there is no point keeping them open across exec(), though.
1005          */
1006         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
1007         if (fd_flag < 0)
1008                 return error("cannot determine file descriptor flags");
1009         fd_flag |= FD_CLOEXEC;
1010         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
1011                 return error("cannot set FD_CLOEXEC");
1012
1013         /* Verify we recognize this pack file format. */
1014         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
1015                 return error("file %s is far too short to be a packfile", p->pack_name);
1016         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
1017                 return error("file %s is not a GIT packfile", p->pack_name);
1018         if (!pack_version_ok(hdr.hdr_version))
1019                 return error("packfile %s is version %"PRIu32" and not"
1020                         " supported (try upgrading GIT to a newer version)",
1021                         p->pack_name, ntohl(hdr.hdr_version));
1022
1023         /* Verify the pack matches its index. */
1024         if (p->num_objects != ntohl(hdr.hdr_entries))
1025                 return error("packfile %s claims to have %"PRIu32" objects"
1026                              " while index indicates %"PRIu32" objects",
1027                              p->pack_name, ntohl(hdr.hdr_entries),
1028                              p->num_objects);
1029         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
1030                 return error("end of packfile %s is unavailable", p->pack_name);
1031         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
1032                 return error("packfile %s signature is unavailable", p->pack_name);
1033         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
1034         if (hashcmp(sha1, idx_sha1))
1035                 return error("packfile %s does not match index", p->pack_name);
1036         return 0;
1037 }
1038
1039 static int open_packed_git(struct packed_git *p)
1040 {
1041         if (!open_packed_git_1(p))
1042                 return 0;
1043         if (p->pack_fd != -1) {
1044                 close(p->pack_fd);
1045                 pack_open_fds--;
1046                 p->pack_fd = -1;
1047         }
1048         return -1;
1049 }
1050
1051 static int in_window(struct pack_window *win, off_t offset)
1052 {
1053         /* We must promise at least 20 bytes (one hash) after the
1054          * offset is available from this window, otherwise the offset
1055          * is not actually in this window and a different window (which
1056          * has that one hash excess) must be used.  This is to support
1057          * the object header and delta base parsing routines below.
1058          */
1059         off_t win_off = win->offset;
1060         return win_off <= offset
1061                 && (offset + 20) <= (win_off + win->len);
1062 }
1063
1064 unsigned char *use_pack(struct packed_git *p,
1065                 struct pack_window **w_cursor,
1066                 off_t offset,
1067                 unsigned long *left)
1068 {
1069         struct pack_window *win = *w_cursor;
1070
1071         /* Since packfiles end in a hash of their content and it's
1072          * pointless to ask for an offset into the middle of that
1073          * hash, and the in_window function above wouldn't match
1074          * don't allow an offset too close to the end of the file.
1075          */
1076         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
1077                 die("packfile %s cannot be accessed", p->pack_name);
1078         if (offset > (p->pack_size - 20))
1079                 die("offset beyond end of packfile (truncated pack?)");
1080
1081         if (!win || !in_window(win, offset)) {
1082                 if (win)
1083                         win->inuse_cnt--;
1084                 for (win = p->windows; win; win = win->next) {
1085                         if (in_window(win, offset))
1086                                 break;
1087                 }
1088                 if (!win) {
1089                         size_t window_align = packed_git_window_size / 2;
1090                         off_t len;
1091
1092                         if (p->pack_fd == -1 && open_packed_git(p))
1093                                 die("packfile %s cannot be accessed", p->pack_name);
1094
1095                         win = xcalloc(1, sizeof(*win));
1096                         win->offset = (offset / window_align) * window_align;
1097                         len = p->pack_size - win->offset;
1098                         if (len > packed_git_window_size)
1099                                 len = packed_git_window_size;
1100                         win->len = (size_t)len;
1101                         pack_mapped += win->len;
1102                         while (packed_git_limit < pack_mapped
1103                                 && unuse_one_window(p))
1104                                 ; /* nothing */
1105                         win->base = xmmap(NULL, win->len,
1106                                 PROT_READ, MAP_PRIVATE,
1107                                 p->pack_fd, win->offset);
1108                         if (win->base == MAP_FAILED)
1109                                 die("packfile %s cannot be mapped: %s",
1110                                         p->pack_name,
1111                                         strerror(errno));
1112                         if (!win->offset && win->len == p->pack_size
1113                                 && !p->do_not_close) {
1114                                 close(p->pack_fd);
1115                                 pack_open_fds--;
1116                                 p->pack_fd = -1;
1117                         }
1118                         pack_mmap_calls++;
1119                         pack_open_windows++;
1120                         if (pack_mapped > peak_pack_mapped)
1121                                 peak_pack_mapped = pack_mapped;
1122                         if (pack_open_windows > peak_pack_open_windows)
1123                                 peak_pack_open_windows = pack_open_windows;
1124                         win->next = p->windows;
1125                         p->windows = win;
1126                 }
1127         }
1128         if (win != *w_cursor) {
1129                 win->last_used = pack_used_ctr++;
1130                 win->inuse_cnt++;
1131                 *w_cursor = win;
1132         }
1133         offset -= win->offset;
1134         if (left)
1135                 *left = win->len - xsize_t(offset);
1136         return win->base + offset;
1137 }
1138
1139 static struct packed_git *alloc_packed_git(int extra)
1140 {
1141         struct packed_git *p = xmalloc(sizeof(*p) + extra);
1142         memset(p, 0, sizeof(*p));
1143         p->pack_fd = -1;
1144         return p;
1145 }
1146
1147 static void try_to_free_pack_memory(size_t size)
1148 {
1149         release_pack_memory(size);
1150 }
1151
1152 struct packed_git *add_packed_git(const char *path, int path_len, int local)
1153 {
1154         static int have_set_try_to_free_routine;
1155         struct stat st;
1156         struct packed_git *p = alloc_packed_git(path_len + 2);
1157
1158         if (!have_set_try_to_free_routine) {
1159                 have_set_try_to_free_routine = 1;
1160                 set_try_to_free_routine(try_to_free_pack_memory);
1161         }
1162
1163         /*
1164          * Make sure a corresponding .pack file exists and that
1165          * the index looks sane.
1166          */
1167         path_len -= strlen(".idx");
1168         if (path_len < 1) {
1169                 free(p);
1170                 return NULL;
1171         }
1172         memcpy(p->pack_name, path, path_len);
1173
1174         strcpy(p->pack_name + path_len, ".keep");
1175         if (!access(p->pack_name, F_OK))
1176                 p->pack_keep = 1;
1177
1178         strcpy(p->pack_name + path_len, ".pack");
1179         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1180                 free(p);
1181                 return NULL;
1182         }
1183
1184         /* ok, it looks sane as far as we can check without
1185          * actually mapping the pack file.
1186          */
1187         p->pack_size = st.st_size;
1188         p->pack_local = local;
1189         p->mtime = st.st_mtime;
1190         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1191                 hashclr(p->sha1);
1192         return p;
1193 }
1194
1195 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1196 {
1197         const char *path = sha1_pack_name(sha1);
1198         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
1199
1200         strcpy(p->pack_name, path);
1201         hashcpy(p->sha1, sha1);
1202         if (check_packed_git_idx(idx_path, p)) {
1203                 free(p);
1204                 return NULL;
1205         }
1206
1207         return p;
1208 }
1209
1210 void install_packed_git(struct packed_git *pack)
1211 {
1212         if (pack->pack_fd != -1)
1213                 pack_open_fds++;
1214
1215         pack->next = packed_git;
1216         packed_git = pack;
1217 }
1218
1219 void (*report_garbage)(const char *desc, const char *path);
1220
1221 static void report_helper(const struct string_list *list,
1222                           int seen_bits, int first, int last)
1223 {
1224         const char *msg;
1225         switch (seen_bits) {
1226         case 0:
1227                 msg = "no corresponding .idx or .pack";
1228                 break;
1229         case 1:
1230                 msg = "no corresponding .idx";
1231                 break;
1232         case 2:
1233                 msg = "no corresponding .pack";
1234                 break;
1235         default:
1236                 return;
1237         }
1238         for (; first < last; first++)
1239                 report_garbage(msg, list->items[first].string);
1240 }
1241
1242 static void report_pack_garbage(struct string_list *list)
1243 {
1244         int i, baselen = -1, first = 0, seen_bits = 0;
1245
1246         if (!report_garbage)
1247                 return;
1248
1249         string_list_sort(list);
1250
1251         for (i = 0; i < list->nr; i++) {
1252                 const char *path = list->items[i].string;
1253                 if (baselen != -1 &&
1254                     strncmp(path, list->items[first].string, baselen)) {
1255                         report_helper(list, seen_bits, first, i);
1256                         baselen = -1;
1257                         seen_bits = 0;
1258                 }
1259                 if (baselen == -1) {
1260                         const char *dot = strrchr(path, '.');
1261                         if (!dot) {
1262                                 report_garbage("garbage found", path);
1263                                 continue;
1264                         }
1265                         baselen = dot - path + 1;
1266                         first = i;
1267                 }
1268                 if (!strcmp(path + baselen, "pack"))
1269                         seen_bits |= 1;
1270                 else if (!strcmp(path + baselen, "idx"))
1271                         seen_bits |= 2;
1272         }
1273         report_helper(list, seen_bits, first, list->nr);
1274 }
1275
1276 static void prepare_packed_git_one(char *objdir, int local)
1277 {
1278         struct strbuf path = STRBUF_INIT;
1279         size_t dirnamelen;
1280         DIR *dir;
1281         struct dirent *de;
1282         struct string_list garbage = STRING_LIST_INIT_DUP;
1283
1284         strbuf_addstr(&path, objdir);
1285         strbuf_addstr(&path, "/pack");
1286         dir = opendir(path.buf);
1287         if (!dir) {
1288                 if (errno != ENOENT)
1289                         error("unable to open object pack directory: %s: %s",
1290                               path.buf, strerror(errno));
1291                 strbuf_release(&path);
1292                 return;
1293         }
1294         strbuf_addch(&path, '/');
1295         dirnamelen = path.len;
1296         while ((de = readdir(dir)) != NULL) {
1297                 struct packed_git *p;
1298                 size_t base_len;
1299
1300                 if (is_dot_or_dotdot(de->d_name))
1301                         continue;
1302
1303                 strbuf_setlen(&path, dirnamelen);
1304                 strbuf_addstr(&path, de->d_name);
1305
1306                 base_len = path.len;
1307                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1308                         /* Don't reopen a pack we already have. */
1309                         for (p = packed_git; p; p = p->next) {
1310                                 size_t len;
1311                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1312                                     len == base_len &&
1313                                     !memcmp(p->pack_name, path.buf, len))
1314                                         break;
1315                         }
1316                         if (p == NULL &&
1317                             /*
1318                              * See if it really is a valid .idx file with
1319                              * corresponding .pack file that we can map.
1320                              */
1321                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1322                                 install_packed_git(p);
1323                 }
1324
1325                 if (!report_garbage)
1326                         continue;
1327
1328                 if (ends_with(de->d_name, ".idx") ||
1329                     ends_with(de->d_name, ".pack") ||
1330                     ends_with(de->d_name, ".bitmap") ||
1331                     ends_with(de->d_name, ".keep"))
1332                         string_list_append(&garbage, path.buf);
1333                 else
1334                         report_garbage("garbage found", path.buf);
1335         }
1336         closedir(dir);
1337         report_pack_garbage(&garbage);
1338         string_list_clear(&garbage, 0);
1339         strbuf_release(&path);
1340 }
1341
1342 static int sort_pack(const void *a_, const void *b_)
1343 {
1344         struct packed_git *a = *((struct packed_git **)a_);
1345         struct packed_git *b = *((struct packed_git **)b_);
1346         int st;
1347
1348         /*
1349          * Local packs tend to contain objects specific to our
1350          * variant of the project than remote ones.  In addition,
1351          * remote ones could be on a network mounted filesystem.
1352          * Favor local ones for these reasons.
1353          */
1354         st = a->pack_local - b->pack_local;
1355         if (st)
1356                 return -st;
1357
1358         /*
1359          * Younger packs tend to contain more recent objects,
1360          * and more recent objects tend to get accessed more
1361          * often.
1362          */
1363         if (a->mtime < b->mtime)
1364                 return 1;
1365         else if (a->mtime == b->mtime)
1366                 return 0;
1367         return -1;
1368 }
1369
1370 static void rearrange_packed_git(void)
1371 {
1372         struct packed_git **ary, *p;
1373         int i, n;
1374
1375         for (n = 0, p = packed_git; p; p = p->next)
1376                 n++;
1377         if (n < 2)
1378                 return;
1379
1380         /* prepare an array of packed_git for easier sorting */
1381         ary = xcalloc(n, sizeof(struct packed_git *));
1382         for (n = 0, p = packed_git; p; p = p->next)
1383                 ary[n++] = p;
1384
1385         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1386
1387         /* link them back again */
1388         for (i = 0; i < n - 1; i++)
1389                 ary[i]->next = ary[i + 1];
1390         ary[n - 1]->next = NULL;
1391         packed_git = ary[0];
1392
1393         free(ary);
1394 }
1395
1396 static int prepare_packed_git_run_once = 0;
1397 void prepare_packed_git(void)
1398 {
1399         struct alternate_object_database *alt;
1400
1401         if (prepare_packed_git_run_once)
1402                 return;
1403         prepare_packed_git_one(get_object_directory(), 1);
1404         prepare_alt_odb();
1405         for (alt = alt_odb_list; alt; alt = alt->next) {
1406                 alt->name[-1] = 0;
1407                 prepare_packed_git_one(alt->base, 0);
1408                 alt->name[-1] = '/';
1409         }
1410         rearrange_packed_git();
1411         prepare_packed_git_run_once = 1;
1412 }
1413
1414 void reprepare_packed_git(void)
1415 {
1416         prepare_packed_git_run_once = 0;
1417         prepare_packed_git();
1418 }
1419
1420 static void mark_bad_packed_object(struct packed_git *p,
1421                                    const unsigned char *sha1)
1422 {
1423         unsigned i;
1424         for (i = 0; i < p->num_bad_objects; i++)
1425                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1426                         return;
1427         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1428         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1429         p->num_bad_objects++;
1430 }
1431
1432 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1433 {
1434         struct packed_git *p;
1435         unsigned i;
1436
1437         for (p = packed_git; p; p = p->next)
1438                 for (i = 0; i < p->num_bad_objects; i++)
1439                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1440                                 return p;
1441         return NULL;
1442 }
1443
1444 /*
1445  * With an in-core object data in "map", rehash it to make sure the
1446  * object name actually matches "sha1" to detect object corruption.
1447  * With "map" == NULL, try reading the object named with "sha1" using
1448  * the streaming interface and rehash it to do the same.
1449  */
1450 int check_sha1_signature(const unsigned char *sha1, void *map,
1451                          unsigned long size, const char *type)
1452 {
1453         unsigned char real_sha1[20];
1454         enum object_type obj_type;
1455         struct git_istream *st;
1456         git_SHA_CTX c;
1457         char hdr[32];
1458         int hdrlen;
1459
1460         if (map) {
1461                 hash_sha1_file(map, size, type, real_sha1);
1462                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1463         }
1464
1465         st = open_istream(sha1, &obj_type, &size, NULL);
1466         if (!st)
1467                 return -1;
1468
1469         /* Generate the header */
1470         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1471
1472         /* Sha1.. */
1473         git_SHA1_Init(&c);
1474         git_SHA1_Update(&c, hdr, hdrlen);
1475         for (;;) {
1476                 char buf[1024 * 16];
1477                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1478
1479                 if (readlen < 0) {
1480                         close_istream(st);
1481                         return -1;
1482                 }
1483                 if (!readlen)
1484                         break;
1485                 git_SHA1_Update(&c, buf, readlen);
1486         }
1487         git_SHA1_Final(real_sha1, &c);
1488         close_istream(st);
1489         return hashcmp(sha1, real_sha1) ? -1 : 0;
1490 }
1491
1492 int git_open_noatime(const char *name)
1493 {
1494         static int sha1_file_open_flag = O_NOATIME;
1495
1496         for (;;) {
1497                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1498                 if (fd >= 0)
1499                         return fd;
1500
1501                 /* Might the failure be due to O_NOATIME? */
1502                 if (errno != ENOENT && sha1_file_open_flag) {
1503                         sha1_file_open_flag = 0;
1504                         continue;
1505                 }
1506
1507                 return -1;
1508         }
1509 }
1510
1511 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1512 {
1513         struct alternate_object_database *alt;
1514
1515         if (!lstat(sha1_file_name(sha1), st))
1516                 return 0;
1517
1518         prepare_alt_odb();
1519         errno = ENOENT;
1520         for (alt = alt_odb_list; alt; alt = alt->next) {
1521                 fill_sha1_path(alt->name, sha1);
1522                 if (!lstat(alt->base, st))
1523                         return 0;
1524         }
1525
1526         return -1;
1527 }
1528
1529 static int open_sha1_file(const unsigned char *sha1)
1530 {
1531         int fd;
1532         struct alternate_object_database *alt;
1533         int most_interesting_errno;
1534
1535         fd = git_open_noatime(sha1_file_name(sha1));
1536         if (fd >= 0)
1537                 return fd;
1538         most_interesting_errno = errno;
1539
1540         prepare_alt_odb();
1541         for (alt = alt_odb_list; alt; alt = alt->next) {
1542                 fill_sha1_path(alt->name, sha1);
1543                 fd = git_open_noatime(alt->base);
1544                 if (fd >= 0)
1545                         return fd;
1546                 if (most_interesting_errno == ENOENT)
1547                         most_interesting_errno = errno;
1548         }
1549         errno = most_interesting_errno;
1550         return -1;
1551 }
1552
1553 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1554 {
1555         void *map;
1556         int fd;
1557
1558         fd = open_sha1_file(sha1);
1559         map = NULL;
1560         if (fd >= 0) {
1561                 struct stat st;
1562
1563                 if (!fstat(fd, &st)) {
1564                         *size = xsize_t(st.st_size);
1565                         if (!*size) {
1566                                 /* mmap() is forbidden on empty files */
1567                                 error("object file %s is empty", sha1_file_name(sha1));
1568                                 return NULL;
1569                         }
1570                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1571                 }
1572                 close(fd);
1573         }
1574         return map;
1575 }
1576
1577 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1578                 unsigned long len, enum object_type *type, unsigned long *sizep)
1579 {
1580         unsigned shift;
1581         unsigned long size, c;
1582         unsigned long used = 0;
1583
1584         c = buf[used++];
1585         *type = (c >> 4) & 7;
1586         size = c & 15;
1587         shift = 4;
1588         while (c & 0x80) {
1589                 if (len <= used || bitsizeof(long) <= shift) {
1590                         error("bad object header");
1591                         size = used = 0;
1592                         break;
1593                 }
1594                 c = buf[used++];
1595                 size += (c & 0x7f) << shift;
1596                 shift += 7;
1597         }
1598         *sizep = size;
1599         return used;
1600 }
1601
1602 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1603 {
1604         /* Get the data stream */
1605         memset(stream, 0, sizeof(*stream));
1606         stream->next_in = map;
1607         stream->avail_in = mapsize;
1608         stream->next_out = buffer;
1609         stream->avail_out = bufsiz;
1610
1611         git_inflate_init(stream);
1612         return git_inflate(stream, 0);
1613 }
1614
1615 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1616                                         unsigned long mapsize, void *buffer,
1617                                         unsigned long bufsiz, struct strbuf *header)
1618 {
1619         int status;
1620
1621         status = unpack_sha1_header(stream, map, mapsize, buffer, bufsiz);
1622
1623         /*
1624          * Check if entire header is unpacked in the first iteration.
1625          */
1626         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1627                 return 0;
1628
1629         /*
1630          * buffer[0..bufsiz] was not large enough.  Copy the partial
1631          * result out to header, and then append the result of further
1632          * reading the stream.
1633          */
1634         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1635         stream->next_out = buffer;
1636         stream->avail_out = bufsiz;
1637
1638         do {
1639                 status = git_inflate(stream, 0);
1640                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1641                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1642                         return 0;
1643                 stream->next_out = buffer;
1644                 stream->avail_out = bufsiz;
1645         } while (status != Z_STREAM_END);
1646         return -1;
1647 }
1648
1649 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1650 {
1651         int bytes = strlen(buffer) + 1;
1652         unsigned char *buf = xmallocz(size);
1653         unsigned long n;
1654         int status = Z_OK;
1655
1656         n = stream->total_out - bytes;
1657         if (n > size)
1658                 n = size;
1659         memcpy(buf, (char *) buffer + bytes, n);
1660         bytes = n;
1661         if (bytes <= size) {
1662                 /*
1663                  * The above condition must be (bytes <= size), not
1664                  * (bytes < size).  In other words, even though we
1665                  * expect no more output and set avail_out to zero,
1666                  * the input zlib stream may have bytes that express
1667                  * "this concludes the stream", and we *do* want to
1668                  * eat that input.
1669                  *
1670                  * Otherwise we would not be able to test that we
1671                  * consumed all the input to reach the expected size;
1672                  * we also want to check that zlib tells us that all
1673                  * went well with status == Z_STREAM_END at the end.
1674                  */
1675                 stream->next_out = buf + bytes;
1676                 stream->avail_out = size - bytes;
1677                 while (status == Z_OK)
1678                         status = git_inflate(stream, Z_FINISH);
1679         }
1680         if (status == Z_STREAM_END && !stream->avail_in) {
1681                 git_inflate_end(stream);
1682                 return buf;
1683         }
1684
1685         if (status < 0)
1686                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1687         else if (stream->avail_in)
1688                 error("garbage at end of loose object '%s'",
1689                       sha1_to_hex(sha1));
1690         free(buf);
1691         return NULL;
1692 }
1693
1694 /*
1695  * We used to just use "sscanf()", but that's actually way
1696  * too permissive for what we want to check. So do an anal
1697  * object header parse by hand.
1698  */
1699 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1700                                unsigned int flags)
1701 {
1702         const char *type_buf = hdr;
1703         unsigned long size;
1704         int type, type_len = 0;
1705
1706         /*
1707          * The type can be of any size but is followed by
1708          * a space.
1709          */
1710         for (;;) {
1711                 char c = *hdr++;
1712                 if (c == ' ')
1713                         break;
1714                 type_len++;
1715         }
1716
1717         type = type_from_string_gently(type_buf, type_len, 1);
1718         if (oi->typename)
1719                 strbuf_add(oi->typename, type_buf, type_len);
1720         /*
1721          * Set type to 0 if its an unknown object and
1722          * we're obtaining the type using '--allow-unkown-type'
1723          * option.
1724          */
1725         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1726                 type = 0;
1727         else if (type < 0)
1728                 die("invalid object type");
1729         if (oi->typep)
1730                 *oi->typep = type;
1731
1732         /*
1733          * The length must follow immediately, and be in canonical
1734          * decimal format (ie "010" is not valid).
1735          */
1736         size = *hdr++ - '0';
1737         if (size > 9)
1738                 return -1;
1739         if (size) {
1740                 for (;;) {
1741                         unsigned long c = *hdr - '0';
1742                         if (c > 9)
1743                                 break;
1744                         hdr++;
1745                         size = size * 10 + c;
1746                 }
1747         }
1748
1749         if (oi->sizep)
1750                 *oi->sizep = size;
1751
1752         /*
1753          * The length must be followed by a zero byte
1754          */
1755         return *hdr ? -1 : type;
1756 }
1757
1758 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1759 {
1760         struct object_info oi;
1761
1762         oi.sizep = sizep;
1763         oi.typename = NULL;
1764         oi.typep = NULL;
1765         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
1766 }
1767
1768 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1769 {
1770         int ret;
1771         git_zstream stream;
1772         char hdr[8192];
1773
1774         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1775         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1776                 return NULL;
1777
1778         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1779 }
1780
1781 unsigned long get_size_from_delta(struct packed_git *p,
1782                                   struct pack_window **w_curs,
1783                                   off_t curpos)
1784 {
1785         const unsigned char *data;
1786         unsigned char delta_head[20], *in;
1787         git_zstream stream;
1788         int st;
1789
1790         memset(&stream, 0, sizeof(stream));
1791         stream.next_out = delta_head;
1792         stream.avail_out = sizeof(delta_head);
1793
1794         git_inflate_init(&stream);
1795         do {
1796                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1797                 stream.next_in = in;
1798                 st = git_inflate(&stream, Z_FINISH);
1799                 curpos += stream.next_in - in;
1800         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1801                  stream.total_out < sizeof(delta_head));
1802         git_inflate_end(&stream);
1803         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1804                 error("delta data unpack-initial failed");
1805                 return 0;
1806         }
1807
1808         /* Examine the initial part of the delta to figure out
1809          * the result size.
1810          */
1811         data = delta_head;
1812
1813         /* ignore base size */
1814         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1815
1816         /* Read the result size */
1817         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1818 }
1819
1820 static off_t get_delta_base(struct packed_git *p,
1821                                     struct pack_window **w_curs,
1822                                     off_t *curpos,
1823                                     enum object_type type,
1824                                     off_t delta_obj_offset)
1825 {
1826         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1827         off_t base_offset;
1828
1829         /* use_pack() assured us we have [base_info, base_info + 20)
1830          * as a range that we can look at without walking off the
1831          * end of the mapped window.  Its actually the hash size
1832          * that is assured.  An OFS_DELTA longer than the hash size
1833          * is stupid, as then a REF_DELTA would be smaller to store.
1834          */
1835         if (type == OBJ_OFS_DELTA) {
1836                 unsigned used = 0;
1837                 unsigned char c = base_info[used++];
1838                 base_offset = c & 127;
1839                 while (c & 128) {
1840                         base_offset += 1;
1841                         if (!base_offset || MSB(base_offset, 7))
1842                                 return 0;  /* overflow */
1843                         c = base_info[used++];
1844                         base_offset = (base_offset << 7) + (c & 127);
1845                 }
1846                 base_offset = delta_obj_offset - base_offset;
1847                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1848                         return 0;  /* out of bound */
1849                 *curpos += used;
1850         } else if (type == OBJ_REF_DELTA) {
1851                 /* The base entry _must_ be in the same pack */
1852                 base_offset = find_pack_entry_one(base_info, p);
1853                 *curpos += 20;
1854         } else
1855                 die("I am totally screwed");
1856         return base_offset;
1857 }
1858
1859 /*
1860  * Like get_delta_base above, but we return the sha1 instead of the pack
1861  * offset. This means it is cheaper for REF deltas (we do not have to do
1862  * the final object lookup), but more expensive for OFS deltas (we
1863  * have to load the revidx to convert the offset back into a sha1).
1864  */
1865 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1866                                                 struct pack_window **w_curs,
1867                                                 off_t curpos,
1868                                                 enum object_type type,
1869                                                 off_t delta_obj_offset)
1870 {
1871         if (type == OBJ_REF_DELTA) {
1872                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1873                 return base;
1874         } else if (type == OBJ_OFS_DELTA) {
1875                 struct revindex_entry *revidx;
1876                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1877                                                    type, delta_obj_offset);
1878
1879                 if (!base_offset)
1880                         return NULL;
1881
1882                 revidx = find_pack_revindex(p, base_offset);
1883                 if (!revidx)
1884                         return NULL;
1885
1886                 return nth_packed_object_sha1(p, revidx->nr);
1887         } else
1888                 return NULL;
1889 }
1890
1891 int unpack_object_header(struct packed_git *p,
1892                          struct pack_window **w_curs,
1893                          off_t *curpos,
1894                          unsigned long *sizep)
1895 {
1896         unsigned char *base;
1897         unsigned long left;
1898         unsigned long used;
1899         enum object_type type;
1900
1901         /* use_pack() assures us we have [base, base + 20) available
1902          * as a range that we can look at.  (Its actually the hash
1903          * size that is assured.)  With our object header encoding
1904          * the maximum deflated object size is 2^137, which is just
1905          * insane, so we know won't exceed what we have been given.
1906          */
1907         base = use_pack(p, w_curs, *curpos, &left);
1908         used = unpack_object_header_buffer(base, left, &type, sizep);
1909         if (!used) {
1910                 type = OBJ_BAD;
1911         } else
1912                 *curpos += used;
1913
1914         return type;
1915 }
1916
1917 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1918 {
1919         int type;
1920         struct revindex_entry *revidx;
1921         const unsigned char *sha1;
1922         revidx = find_pack_revindex(p, obj_offset);
1923         if (!revidx)
1924                 return OBJ_BAD;
1925         sha1 = nth_packed_object_sha1(p, revidx->nr);
1926         mark_bad_packed_object(p, sha1);
1927         type = sha1_object_info(sha1, NULL);
1928         if (type <= OBJ_NONE)
1929                 return OBJ_BAD;
1930         return type;
1931 }
1932
1933 #define POI_STACK_PREALLOC 64
1934
1935 static enum object_type packed_to_object_type(struct packed_git *p,
1936                                               off_t obj_offset,
1937                                               enum object_type type,
1938                                               struct pack_window **w_curs,
1939                                               off_t curpos)
1940 {
1941         off_t small_poi_stack[POI_STACK_PREALLOC];
1942         off_t *poi_stack = small_poi_stack;
1943         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1944
1945         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1946                 off_t base_offset;
1947                 unsigned long size;
1948                 /* Push the object we're going to leave behind */
1949                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1950                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1951                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1952                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1953                 } else {
1954                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1955                 }
1956                 poi_stack[poi_stack_nr++] = obj_offset;
1957                 /* If parsing the base offset fails, just unwind */
1958                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1959                 if (!base_offset)
1960                         goto unwind;
1961                 curpos = obj_offset = base_offset;
1962                 type = unpack_object_header(p, w_curs, &curpos, &size);
1963                 if (type <= OBJ_NONE) {
1964                         /* If getting the base itself fails, we first
1965                          * retry the base, otherwise unwind */
1966                         type = retry_bad_packed_offset(p, base_offset);
1967                         if (type > OBJ_NONE)
1968                                 goto out;
1969                         goto unwind;
1970                 }
1971         }
1972
1973         switch (type) {
1974         case OBJ_BAD:
1975         case OBJ_COMMIT:
1976         case OBJ_TREE:
1977         case OBJ_BLOB:
1978         case OBJ_TAG:
1979                 break;
1980         default:
1981                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1982                       type, (uintmax_t)obj_offset, p->pack_name);
1983                 type = OBJ_BAD;
1984         }
1985
1986 out:
1987         if (poi_stack != small_poi_stack)
1988                 free(poi_stack);
1989         return type;
1990
1991 unwind:
1992         while (poi_stack_nr) {
1993                 obj_offset = poi_stack[--poi_stack_nr];
1994                 type = retry_bad_packed_offset(p, obj_offset);
1995                 if (type > OBJ_NONE)
1996                         goto out;
1997         }
1998         type = OBJ_BAD;
1999         goto out;
2000 }
2001
2002 static int packed_object_info(struct packed_git *p, off_t obj_offset,
2003                               struct object_info *oi)
2004 {
2005         struct pack_window *w_curs = NULL;
2006         unsigned long size;
2007         off_t curpos = obj_offset;
2008         enum object_type type;
2009
2010         /*
2011          * We always get the representation type, but only convert it to
2012          * a "real" type later if the caller is interested.
2013          */
2014         type = unpack_object_header(p, &w_curs, &curpos, &size);
2015
2016         if (oi->sizep) {
2017                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2018                         off_t tmp_pos = curpos;
2019                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2020                                                            type, obj_offset);
2021                         if (!base_offset) {
2022                                 type = OBJ_BAD;
2023                                 goto out;
2024                         }
2025                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2026                         if (*oi->sizep == 0) {
2027                                 type = OBJ_BAD;
2028                                 goto out;
2029                         }
2030                 } else {
2031                         *oi->sizep = size;
2032                 }
2033         }
2034
2035         if (oi->disk_sizep) {
2036                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2037                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2038         }
2039
2040         if (oi->typep) {
2041                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2042                 if (*oi->typep < 0) {
2043                         type = OBJ_BAD;
2044                         goto out;
2045                 }
2046         }
2047
2048         if (oi->delta_base_sha1) {
2049                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2050                         const unsigned char *base;
2051
2052                         base = get_delta_base_sha1(p, &w_curs, curpos,
2053                                                    type, obj_offset);
2054                         if (!base) {
2055                                 type = OBJ_BAD;
2056                                 goto out;
2057                         }
2058
2059                         hashcpy(oi->delta_base_sha1, base);
2060                 } else
2061                         hashclr(oi->delta_base_sha1);
2062         }
2063
2064 out:
2065         unuse_pack(&w_curs);
2066         return type;
2067 }
2068
2069 static void *unpack_compressed_entry(struct packed_git *p,
2070                                     struct pack_window **w_curs,
2071                                     off_t curpos,
2072                                     unsigned long size)
2073 {
2074         int st;
2075         git_zstream stream;
2076         unsigned char *buffer, *in;
2077
2078         buffer = xmallocz_gently(size);
2079         if (!buffer)
2080                 return NULL;
2081         memset(&stream, 0, sizeof(stream));
2082         stream.next_out = buffer;
2083         stream.avail_out = size + 1;
2084
2085         git_inflate_init(&stream);
2086         do {
2087                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2088                 stream.next_in = in;
2089                 st = git_inflate(&stream, Z_FINISH);
2090                 if (!stream.avail_out)
2091                         break; /* the payload is larger than it should be */
2092                 curpos += stream.next_in - in;
2093         } while (st == Z_OK || st == Z_BUF_ERROR);
2094         git_inflate_end(&stream);
2095         if ((st != Z_STREAM_END) || stream.total_out != size) {
2096                 free(buffer);
2097                 return NULL;
2098         }
2099
2100         return buffer;
2101 }
2102
2103 #define MAX_DELTA_CACHE (256)
2104
2105 static size_t delta_base_cached;
2106
2107 static struct delta_base_cache_lru_list {
2108         struct delta_base_cache_lru_list *prev;
2109         struct delta_base_cache_lru_list *next;
2110 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
2111
2112 static struct delta_base_cache_entry {
2113         struct delta_base_cache_lru_list lru;
2114         void *data;
2115         struct packed_git *p;
2116         off_t base_offset;
2117         unsigned long size;
2118         enum object_type type;
2119 } delta_base_cache[MAX_DELTA_CACHE];
2120
2121 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
2122 {
2123         unsigned long hash;
2124
2125         hash = (unsigned long)p + (unsigned long)base_offset;
2126         hash += (hash >> 8) + (hash >> 16);
2127         return hash % MAX_DELTA_CACHE;
2128 }
2129
2130 static struct delta_base_cache_entry *
2131 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2132 {
2133         unsigned long hash = pack_entry_hash(p, base_offset);
2134         return delta_base_cache + hash;
2135 }
2136
2137 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
2138                                      struct packed_git *p, off_t base_offset)
2139 {
2140         return (ent->data && ent->p == p && ent->base_offset == base_offset);
2141 }
2142
2143 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2144 {
2145         struct delta_base_cache_entry *ent;
2146         ent = get_delta_base_cache_entry(p, base_offset);
2147         return eq_delta_base_cache_entry(ent, p, base_offset);
2148 }
2149
2150 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2151 {
2152         ent->data = NULL;
2153         ent->lru.next->prev = ent->lru.prev;
2154         ent->lru.prev->next = ent->lru.next;
2155         delta_base_cached -= ent->size;
2156 }
2157
2158 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2159         unsigned long *base_size, enum object_type *type, int keep_cache)
2160 {
2161         struct delta_base_cache_entry *ent;
2162         void *ret;
2163
2164         ent = get_delta_base_cache_entry(p, base_offset);
2165
2166         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2167                 return unpack_entry(p, base_offset, type, base_size);
2168
2169         ret = ent->data;
2170
2171         if (!keep_cache)
2172                 clear_delta_base_cache_entry(ent);
2173         else
2174                 ret = xmemdupz(ent->data, ent->size);
2175         *type = ent->type;
2176         *base_size = ent->size;
2177         return ret;
2178 }
2179
2180 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2181 {
2182         if (ent->data) {
2183                 free(ent->data);
2184                 ent->data = NULL;
2185                 ent->lru.next->prev = ent->lru.prev;
2186                 ent->lru.prev->next = ent->lru.next;
2187                 delta_base_cached -= ent->size;
2188         }
2189 }
2190
2191 void clear_delta_base_cache(void)
2192 {
2193         unsigned long p;
2194         for (p = 0; p < MAX_DELTA_CACHE; p++)
2195                 release_delta_base_cache(&delta_base_cache[p]);
2196 }
2197
2198 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2199         void *base, unsigned long base_size, enum object_type type)
2200 {
2201         unsigned long hash = pack_entry_hash(p, base_offset);
2202         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2203         struct delta_base_cache_lru_list *lru;
2204
2205         release_delta_base_cache(ent);
2206         delta_base_cached += base_size;
2207
2208         for (lru = delta_base_cache_lru.next;
2209              delta_base_cached > delta_base_cache_limit
2210              && lru != &delta_base_cache_lru;
2211              lru = lru->next) {
2212                 struct delta_base_cache_entry *f = (void *)lru;
2213                 if (f->type == OBJ_BLOB)
2214                         release_delta_base_cache(f);
2215         }
2216         for (lru = delta_base_cache_lru.next;
2217              delta_base_cached > delta_base_cache_limit
2218              && lru != &delta_base_cache_lru;
2219              lru = lru->next) {
2220                 struct delta_base_cache_entry *f = (void *)lru;
2221                 release_delta_base_cache(f);
2222         }
2223
2224         ent->p = p;
2225         ent->base_offset = base_offset;
2226         ent->type = type;
2227         ent->data = base;
2228         ent->size = base_size;
2229         ent->lru.next = &delta_base_cache_lru;
2230         ent->lru.prev = delta_base_cache_lru.prev;
2231         delta_base_cache_lru.prev->next = &ent->lru;
2232         delta_base_cache_lru.prev = &ent->lru;
2233 }
2234
2235 static void *read_object(const unsigned char *sha1, enum object_type *type,
2236                          unsigned long *size);
2237
2238 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2239 {
2240         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2241         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2242                          p->pack_name, (uintmax_t)obj_offset);
2243 }
2244
2245 int do_check_packed_object_crc;
2246
2247 #define UNPACK_ENTRY_STACK_PREALLOC 64
2248 struct unpack_entry_stack_ent {
2249         off_t obj_offset;
2250         off_t curpos;
2251         unsigned long size;
2252 };
2253
2254 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2255                    enum object_type *final_type, unsigned long *final_size)
2256 {
2257         struct pack_window *w_curs = NULL;
2258         off_t curpos = obj_offset;
2259         void *data = NULL;
2260         unsigned long size;
2261         enum object_type type;
2262         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2263         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2264         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2265         int base_from_cache = 0;
2266
2267         write_pack_access_log(p, obj_offset);
2268
2269         /* PHASE 1: drill down to the innermost base object */
2270         for (;;) {
2271                 off_t base_offset;
2272                 int i;
2273                 struct delta_base_cache_entry *ent;
2274
2275                 ent = get_delta_base_cache_entry(p, curpos);
2276                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2277                         type = ent->type;
2278                         data = ent->data;
2279                         size = ent->size;
2280                         clear_delta_base_cache_entry(ent);
2281                         base_from_cache = 1;
2282                         break;
2283                 }
2284
2285                 if (do_check_packed_object_crc && p->index_version > 1) {
2286                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2287                         unsigned long len = revidx[1].offset - obj_offset;
2288                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2289                                 const unsigned char *sha1 =
2290                                         nth_packed_object_sha1(p, revidx->nr);
2291                                 error("bad packed object CRC for %s",
2292                                       sha1_to_hex(sha1));
2293                                 mark_bad_packed_object(p, sha1);
2294                                 unuse_pack(&w_curs);
2295                                 return NULL;
2296                         }
2297                 }
2298
2299                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2300                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2301                         break;
2302
2303                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2304                 if (!base_offset) {
2305                         error("failed to validate delta base reference "
2306                               "at offset %"PRIuMAX" from %s",
2307                               (uintmax_t)curpos, p->pack_name);
2308                         /* bail to phase 2, in hopes of recovery */
2309                         data = NULL;
2310                         break;
2311                 }
2312
2313                 /* push object, proceed to base */
2314                 if (delta_stack_nr >= delta_stack_alloc
2315                     && delta_stack == small_delta_stack) {
2316                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2317                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2318                         memcpy(delta_stack, small_delta_stack,
2319                                sizeof(*delta_stack)*delta_stack_nr);
2320                 } else {
2321                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2322                 }
2323                 i = delta_stack_nr++;
2324                 delta_stack[i].obj_offset = obj_offset;
2325                 delta_stack[i].curpos = curpos;
2326                 delta_stack[i].size = size;
2327
2328                 curpos = obj_offset = base_offset;
2329         }
2330
2331         /* PHASE 2: handle the base */
2332         switch (type) {
2333         case OBJ_OFS_DELTA:
2334         case OBJ_REF_DELTA:
2335                 if (data)
2336                         die("BUG in unpack_entry: left loop at a valid delta");
2337                 break;
2338         case OBJ_COMMIT:
2339         case OBJ_TREE:
2340         case OBJ_BLOB:
2341         case OBJ_TAG:
2342                 if (!base_from_cache)
2343                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2344                 break;
2345         default:
2346                 data = NULL;
2347                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2348                       type, (uintmax_t)obj_offset, p->pack_name);
2349         }
2350
2351         /* PHASE 3: apply deltas in order */
2352
2353         /* invariants:
2354          *   'data' holds the base data, or NULL if there was corruption
2355          */
2356         while (delta_stack_nr) {
2357                 void *delta_data;
2358                 void *base = data;
2359                 unsigned long delta_size, base_size = size;
2360                 int i;
2361
2362                 data = NULL;
2363
2364                 if (base)
2365                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2366
2367                 if (!base) {
2368                         /*
2369                          * We're probably in deep shit, but let's try to fetch
2370                          * the required base anyway from another pack or loose.
2371                          * This is costly but should happen only in the presence
2372                          * of a corrupted pack, and is better than failing outright.
2373                          */
2374                         struct revindex_entry *revidx;
2375                         const unsigned char *base_sha1;
2376                         revidx = find_pack_revindex(p, obj_offset);
2377                         if (revidx) {
2378                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2379                                 error("failed to read delta base object %s"
2380                                       " at offset %"PRIuMAX" from %s",
2381                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2382                                       p->pack_name);
2383                                 mark_bad_packed_object(p, base_sha1);
2384                                 base = read_object(base_sha1, &type, &base_size);
2385                         }
2386                 }
2387
2388                 i = --delta_stack_nr;
2389                 obj_offset = delta_stack[i].obj_offset;
2390                 curpos = delta_stack[i].curpos;
2391                 delta_size = delta_stack[i].size;
2392
2393                 if (!base)
2394                         continue;
2395
2396                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2397
2398                 if (!delta_data) {
2399                         error("failed to unpack compressed delta "
2400                               "at offset %"PRIuMAX" from %s",
2401                               (uintmax_t)curpos, p->pack_name);
2402                         data = NULL;
2403                         continue;
2404                 }
2405
2406                 data = patch_delta(base, base_size,
2407                                    delta_data, delta_size,
2408                                    &size);
2409
2410                 /*
2411                  * We could not apply the delta; warn the user, but keep going.
2412                  * Our failure will be noticed either in the next iteration of
2413                  * the loop, or if this is the final delta, in the caller when
2414                  * we return NULL. Those code paths will take care of making
2415                  * a more explicit warning and retrying with another copy of
2416                  * the object.
2417                  */
2418                 if (!data)
2419                         error("failed to apply delta");
2420
2421                 free(delta_data);
2422         }
2423
2424         *final_type = type;
2425         *final_size = size;
2426
2427         unuse_pack(&w_curs);
2428
2429         if (delta_stack != small_delta_stack)
2430                 free(delta_stack);
2431
2432         return data;
2433 }
2434
2435 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2436                                             uint32_t n)
2437 {
2438         const unsigned char *index = p->index_data;
2439         if (!index) {
2440                 if (open_pack_index(p))
2441                         return NULL;
2442                 index = p->index_data;
2443         }
2444         if (n >= p->num_objects)
2445                 return NULL;
2446         index += 4 * 256;
2447         if (p->index_version == 1) {
2448                 return index + 24 * n + 4;
2449         } else {
2450                 index += 8;
2451                 return index + 20 * n;
2452         }
2453 }
2454
2455 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2456 {
2457         const unsigned char *index = p->index_data;
2458         index += 4 * 256;
2459         if (p->index_version == 1) {
2460                 return ntohl(*((uint32_t *)(index + 24 * n)));
2461         } else {
2462                 uint32_t off;
2463                 index += 8 + p->num_objects * (20 + 4);
2464                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2465                 if (!(off & 0x80000000))
2466                         return off;
2467                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2468                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2469                                    ntohl(*((uint32_t *)(index + 4)));
2470         }
2471 }
2472
2473 off_t find_pack_entry_one(const unsigned char *sha1,
2474                                   struct packed_git *p)
2475 {
2476         const uint32_t *level1_ofs = p->index_data;
2477         const unsigned char *index = p->index_data;
2478         unsigned hi, lo, stride;
2479         static int use_lookup = -1;
2480         static int debug_lookup = -1;
2481
2482         if (debug_lookup < 0)
2483                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2484
2485         if (!index) {
2486                 if (open_pack_index(p))
2487                         return 0;
2488                 level1_ofs = p->index_data;
2489                 index = p->index_data;
2490         }
2491         if (p->index_version > 1) {
2492                 level1_ofs += 2;
2493                 index += 8;
2494         }
2495         index += 4 * 256;
2496         hi = ntohl(level1_ofs[*sha1]);
2497         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2498         if (p->index_version > 1) {
2499                 stride = 20;
2500         } else {
2501                 stride = 24;
2502                 index += 4;
2503         }
2504
2505         if (debug_lookup)
2506                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2507                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2508
2509         if (use_lookup < 0)
2510                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2511         if (use_lookup) {
2512                 int pos = sha1_entry_pos(index, stride, 0,
2513                                          lo, hi, p->num_objects, sha1);
2514                 if (pos < 0)
2515                         return 0;
2516                 return nth_packed_object_offset(p, pos);
2517         }
2518
2519         do {
2520                 unsigned mi = (lo + hi) / 2;
2521                 int cmp = hashcmp(index + mi * stride, sha1);
2522
2523                 if (debug_lookup)
2524                         printf("lo %u hi %u rg %u mi %u\n",
2525                                lo, hi, hi - lo, mi);
2526                 if (!cmp)
2527                         return nth_packed_object_offset(p, mi);
2528                 if (cmp > 0)
2529                         hi = mi;
2530                 else
2531                         lo = mi+1;
2532         } while (lo < hi);
2533         return 0;
2534 }
2535
2536 int is_pack_valid(struct packed_git *p)
2537 {
2538         /* An already open pack is known to be valid. */
2539         if (p->pack_fd != -1)
2540                 return 1;
2541
2542         /* If the pack has one window completely covering the
2543          * file size, the pack is known to be valid even if
2544          * the descriptor is not currently open.
2545          */
2546         if (p->windows) {
2547                 struct pack_window *w = p->windows;
2548
2549                 if (!w->offset && w->len == p->pack_size)
2550                         return 1;
2551         }
2552
2553         /* Force the pack to open to prove its valid. */
2554         return !open_packed_git(p);
2555 }
2556
2557 static int fill_pack_entry(const unsigned char *sha1,
2558                            struct pack_entry *e,
2559                            struct packed_git *p)
2560 {
2561         off_t offset;
2562
2563         if (p->num_bad_objects) {
2564                 unsigned i;
2565                 for (i = 0; i < p->num_bad_objects; i++)
2566                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2567                                 return 0;
2568         }
2569
2570         offset = find_pack_entry_one(sha1, p);
2571         if (!offset)
2572                 return 0;
2573
2574         /*
2575          * We are about to tell the caller where they can locate the
2576          * requested object.  We better make sure the packfile is
2577          * still here and can be accessed before supplying that
2578          * answer, as it may have been deleted since the index was
2579          * loaded!
2580          */
2581         if (!is_pack_valid(p))
2582                 return 0;
2583         e->offset = offset;
2584         e->p = p;
2585         hashcpy(e->sha1, sha1);
2586         return 1;
2587 }
2588
2589 /*
2590  * Iff a pack file contains the object named by sha1, return true and
2591  * store its location to e.
2592  */
2593 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2594 {
2595         struct packed_git *p;
2596
2597         prepare_packed_git();
2598         if (!packed_git)
2599                 return 0;
2600
2601         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2602                 return 1;
2603
2604         for (p = packed_git; p; p = p->next) {
2605                 if (p == last_found_pack)
2606                         continue; /* we already checked this one */
2607
2608                 if (fill_pack_entry(sha1, e, p)) {
2609                         last_found_pack = p;
2610                         return 1;
2611                 }
2612         }
2613         return 0;
2614 }
2615
2616 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2617                                   struct packed_git *packs)
2618 {
2619         struct packed_git *p;
2620
2621         for (p = packs; p; p = p->next) {
2622                 if (find_pack_entry_one(sha1, p))
2623                         return p;
2624         }
2625         return NULL;
2626
2627 }
2628
2629 static int sha1_loose_object_info(const unsigned char *sha1,
2630                                   struct object_info *oi,
2631                                   int flags)
2632 {
2633         int status = 0;
2634         unsigned long mapsize;
2635         void *map;
2636         git_zstream stream;
2637         char hdr[32];
2638         struct strbuf hdrbuf = STRBUF_INIT;
2639
2640         if (oi->delta_base_sha1)
2641                 hashclr(oi->delta_base_sha1);
2642
2643         /*
2644          * If we don't care about type or size, then we don't
2645          * need to look inside the object at all. Note that we
2646          * do not optimize out the stat call, even if the
2647          * caller doesn't care about the disk-size, since our
2648          * return value implicitly indicates whether the
2649          * object even exists.
2650          */
2651         if (!oi->typep && !oi->typename && !oi->sizep) {
2652                 struct stat st;
2653                 if (stat_sha1_file(sha1, &st) < 0)
2654                         return -1;
2655                 if (oi->disk_sizep)
2656                         *oi->disk_sizep = st.st_size;
2657                 return 0;
2658         }
2659
2660         map = map_sha1_file(sha1, &mapsize);
2661         if (!map)
2662                 return -1;
2663         if (oi->disk_sizep)
2664                 *oi->disk_sizep = mapsize;
2665         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2666                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2667                         status = error("unable to unpack %s header with --allow-unknown-type",
2668                                        sha1_to_hex(sha1));
2669         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2670                 status = error("unable to unpack %s header",
2671                                sha1_to_hex(sha1));
2672         if (status < 0)
2673                 ; /* Do nothing */
2674         else if (hdrbuf.len) {
2675                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2676                         status = error("unable to parse %s header with --allow-unknown-type",
2677                                        sha1_to_hex(sha1));
2678         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2679                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2680         git_inflate_end(&stream);
2681         munmap(map, mapsize);
2682         if (status && oi->typep)
2683                 *oi->typep = status;
2684         strbuf_release(&hdrbuf);
2685         return 0;
2686 }
2687
2688 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2689 {
2690         struct cached_object *co;
2691         struct pack_entry e;
2692         int rtype;
2693         enum object_type real_type;
2694         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2695
2696         co = find_cached_object(real);
2697         if (co) {
2698                 if (oi->typep)
2699                         *(oi->typep) = co->type;
2700                 if (oi->sizep)
2701                         *(oi->sizep) = co->size;
2702                 if (oi->disk_sizep)
2703                         *(oi->disk_sizep) = 0;
2704                 if (oi->delta_base_sha1)
2705                         hashclr(oi->delta_base_sha1);
2706                 if (oi->typename)
2707                         strbuf_addstr(oi->typename, typename(co->type));
2708                 oi->whence = OI_CACHED;
2709                 return 0;
2710         }
2711
2712         if (!find_pack_entry(real, &e)) {
2713                 /* Most likely it's a loose object. */
2714                 if (!sha1_loose_object_info(real, oi, flags)) {
2715                         oi->whence = OI_LOOSE;
2716                         return 0;
2717                 }
2718
2719                 /* Not a loose object; someone else may have just packed it. */
2720                 reprepare_packed_git();
2721                 if (!find_pack_entry(real, &e))
2722                         return -1;
2723         }
2724
2725         /*
2726          * packed_object_info() does not follow the delta chain to
2727          * find out the real type, unless it is given oi->typep.
2728          */
2729         if (oi->typename && !oi->typep)
2730                 oi->typep = &real_type;
2731
2732         rtype = packed_object_info(e.p, e.offset, oi);
2733         if (rtype < 0) {
2734                 mark_bad_packed_object(e.p, real);
2735                 if (oi->typep == &real_type)
2736                         oi->typep = NULL;
2737                 return sha1_object_info_extended(real, oi, 0);
2738         } else if (in_delta_base_cache(e.p, e.offset)) {
2739                 oi->whence = OI_DBCACHED;
2740         } else {
2741                 oi->whence = OI_PACKED;
2742                 oi->u.packed.offset = e.offset;
2743                 oi->u.packed.pack = e.p;
2744                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2745                                          rtype == OBJ_OFS_DELTA);
2746         }
2747         if (oi->typename)
2748                 strbuf_addstr(oi->typename, typename(*oi->typep));
2749         if (oi->typep == &real_type)
2750                 oi->typep = NULL;
2751
2752         return 0;
2753 }
2754
2755 /* returns enum object_type or negative */
2756 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2757 {
2758         enum object_type type;
2759         struct object_info oi = {NULL};
2760
2761         oi.typep = &type;
2762         oi.sizep = sizep;
2763         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2764                 return -1;
2765         return type;
2766 }
2767
2768 static void *read_packed_sha1(const unsigned char *sha1,
2769                               enum object_type *type, unsigned long *size)
2770 {
2771         struct pack_entry e;
2772         void *data;
2773
2774         if (!find_pack_entry(sha1, &e))
2775                 return NULL;
2776         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2777         if (!data) {
2778                 /*
2779                  * We're probably in deep shit, but let's try to fetch
2780                  * the required object anyway from another pack or loose.
2781                  * This should happen only in the presence of a corrupted
2782                  * pack, and is better than failing outright.
2783                  */
2784                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2785                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2786                 mark_bad_packed_object(e.p, sha1);
2787                 data = read_object(sha1, type, size);
2788         }
2789         return data;
2790 }
2791
2792 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2793                       unsigned char *sha1)
2794 {
2795         struct cached_object *co;
2796
2797         hash_sha1_file(buf, len, typename(type), sha1);
2798         if (has_sha1_file(sha1) || find_cached_object(sha1))
2799                 return 0;
2800         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2801         co = &cached_objects[cached_object_nr++];
2802         co->size = len;
2803         co->type = type;
2804         co->buf = xmalloc(len);
2805         memcpy(co->buf, buf, len);
2806         hashcpy(co->sha1, sha1);
2807         return 0;
2808 }
2809
2810 static void *read_object(const unsigned char *sha1, enum object_type *type,
2811                          unsigned long *size)
2812 {
2813         unsigned long mapsize;
2814         void *map, *buf;
2815         struct cached_object *co;
2816
2817         co = find_cached_object(sha1);
2818         if (co) {
2819                 *type = co->type;
2820                 *size = co->size;
2821                 return xmemdupz(co->buf, co->size);
2822         }
2823
2824         buf = read_packed_sha1(sha1, type, size);
2825         if (buf)
2826                 return buf;
2827         map = map_sha1_file(sha1, &mapsize);
2828         if (map) {
2829                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2830                 munmap(map, mapsize);
2831                 return buf;
2832         }
2833         reprepare_packed_git();
2834         return read_packed_sha1(sha1, type, size);
2835 }
2836
2837 /*
2838  * This function dies on corrupt objects; the callers who want to
2839  * deal with them should arrange to call read_object() and give error
2840  * messages themselves.
2841  */
2842 void *read_sha1_file_extended(const unsigned char *sha1,
2843                               enum object_type *type,
2844                               unsigned long *size,
2845                               unsigned flag)
2846 {
2847         void *data;
2848         const struct packed_git *p;
2849         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2850
2851         errno = 0;
2852         data = read_object(repl, type, size);
2853         if (data)
2854                 return data;
2855
2856         if (errno && errno != ENOENT)
2857                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2858
2859         /* die if we replaced an object with one that does not exist */
2860         if (repl != sha1)
2861                 die("replacement %s not found for %s",
2862                     sha1_to_hex(repl), sha1_to_hex(sha1));
2863
2864         if (has_loose_object(repl)) {
2865                 const char *path = sha1_file_name(sha1);
2866
2867                 die("loose object %s (stored in %s) is corrupt",
2868                     sha1_to_hex(repl), path);
2869         }
2870
2871         if ((p = has_packed_and_bad(repl)) != NULL)
2872                 die("packed object %s (stored in %s) is corrupt",
2873                     sha1_to_hex(repl), p->pack_name);
2874
2875         return NULL;
2876 }
2877
2878 void *read_object_with_reference(const unsigned char *sha1,
2879                                  const char *required_type_name,
2880                                  unsigned long *size,
2881                                  unsigned char *actual_sha1_return)
2882 {
2883         enum object_type type, required_type;
2884         void *buffer;
2885         unsigned long isize;
2886         unsigned char actual_sha1[20];
2887
2888         required_type = type_from_string(required_type_name);
2889         hashcpy(actual_sha1, sha1);
2890         while (1) {
2891                 int ref_length = -1;
2892                 const char *ref_type = NULL;
2893
2894                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2895                 if (!buffer)
2896                         return NULL;
2897                 if (type == required_type) {
2898                         *size = isize;
2899                         if (actual_sha1_return)
2900                                 hashcpy(actual_sha1_return, actual_sha1);
2901                         return buffer;
2902                 }
2903                 /* Handle references */
2904                 else if (type == OBJ_COMMIT)
2905                         ref_type = "tree ";
2906                 else if (type == OBJ_TAG)
2907                         ref_type = "object ";
2908                 else {
2909                         free(buffer);
2910                         return NULL;
2911                 }
2912                 ref_length = strlen(ref_type);
2913
2914                 if (ref_length + 40 > isize ||
2915                     memcmp(buffer, ref_type, ref_length) ||
2916                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2917                         free(buffer);
2918                         return NULL;
2919                 }
2920                 free(buffer);
2921                 /* Now we have the ID of the referred-to object in
2922                  * actual_sha1.  Check again. */
2923         }
2924 }
2925
2926 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2927                                     const char *type, unsigned char *sha1,
2928                                     char *hdr, int *hdrlen)
2929 {
2930         git_SHA_CTX c;
2931
2932         /* Generate the header */
2933         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2934
2935         /* Sha1.. */
2936         git_SHA1_Init(&c);
2937         git_SHA1_Update(&c, hdr, *hdrlen);
2938         git_SHA1_Update(&c, buf, len);
2939         git_SHA1_Final(sha1, &c);
2940 }
2941
2942 /*
2943  * Move the just written object into its final resting place.
2944  */
2945 int finalize_object_file(const char *tmpfile, const char *filename)
2946 {
2947         int ret = 0;
2948
2949         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2950                 goto try_rename;
2951         else if (link(tmpfile, filename))
2952                 ret = errno;
2953
2954         /*
2955          * Coda hack - coda doesn't like cross-directory links,
2956          * so we fall back to a rename, which will mean that it
2957          * won't be able to check collisions, but that's not a
2958          * big deal.
2959          *
2960          * The same holds for FAT formatted media.
2961          *
2962          * When this succeeds, we just return.  We have nothing
2963          * left to unlink.
2964          */
2965         if (ret && ret != EEXIST) {
2966         try_rename:
2967                 if (!rename(tmpfile, filename))
2968                         goto out;
2969                 ret = errno;
2970         }
2971         unlink_or_warn(tmpfile);
2972         if (ret) {
2973                 if (ret != EEXIST) {
2974                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2975                 }
2976                 /* FIXME!!! Collision check here ? */
2977         }
2978
2979 out:
2980         if (adjust_shared_perm(filename))
2981                 return error("unable to set permission to '%s'", filename);
2982         return 0;
2983 }
2984
2985 static int write_buffer(int fd, const void *buf, size_t len)
2986 {
2987         if (write_in_full(fd, buf, len) < 0)
2988                 return error("file write error (%s)", strerror(errno));
2989         return 0;
2990 }
2991
2992 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2993                    unsigned char *sha1)
2994 {
2995         char hdr[32];
2996         int hdrlen;
2997         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2998         return 0;
2999 }
3000
3001 /* Finalize a file on disk, and close it. */
3002 static void close_sha1_file(int fd)
3003 {
3004         if (fsync_object_files)
3005                 fsync_or_die(fd, "sha1 file");
3006         if (close(fd) != 0)
3007                 die_errno("error when closing sha1 file");
3008 }
3009
3010 /* Size of directory component, including the ending '/' */
3011 static inline int directory_size(const char *filename)
3012 {
3013         const char *s = strrchr(filename, '/');
3014         if (!s)
3015                 return 0;
3016         return s - filename + 1;
3017 }
3018
3019 /*
3020  * This creates a temporary file in the same directory as the final
3021  * 'filename'
3022  *
3023  * We want to avoid cross-directory filename renames, because those
3024  * can have problems on various filesystems (FAT, NFS, Coda).
3025  */
3026 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
3027 {
3028         int fd, dirlen = directory_size(filename);
3029
3030         if (dirlen + 20 > bufsiz) {
3031                 errno = ENAMETOOLONG;
3032                 return -1;
3033         }
3034         memcpy(buffer, filename, dirlen);
3035         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
3036         fd = git_mkstemp_mode(buffer, 0444);
3037         if (fd < 0 && dirlen && errno == ENOENT) {
3038                 /* Make sure the directory exists */
3039                 memcpy(buffer, filename, dirlen);
3040                 buffer[dirlen-1] = 0;
3041                 if (mkdir(buffer, 0777) && errno != EEXIST)
3042                         return -1;
3043                 if (adjust_shared_perm(buffer))
3044                         return -1;
3045
3046                 /* Try again */
3047                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
3048                 fd = git_mkstemp_mode(buffer, 0444);
3049         }
3050         return fd;
3051 }
3052
3053 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3054                               const void *buf, unsigned long len, time_t mtime)
3055 {
3056         int fd, ret;
3057         unsigned char compressed[4096];
3058         git_zstream stream;
3059         git_SHA_CTX c;
3060         unsigned char parano_sha1[20];
3061         static char tmp_file[PATH_MAX];
3062         const char *filename = sha1_file_name(sha1);
3063
3064         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
3065         if (fd < 0) {
3066                 if (errno == EACCES)
3067                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3068                 else
3069                         return error("unable to create temporary file: %s", strerror(errno));
3070         }
3071
3072         /* Set it up */
3073         git_deflate_init(&stream, zlib_compression_level);
3074         stream.next_out = compressed;
3075         stream.avail_out = sizeof(compressed);
3076         git_SHA1_Init(&c);
3077
3078         /* First header.. */
3079         stream.next_in = (unsigned char *)hdr;
3080         stream.avail_in = hdrlen;
3081         while (git_deflate(&stream, 0) == Z_OK)
3082                 ; /* nothing */
3083         git_SHA1_Update(&c, hdr, hdrlen);
3084
3085         /* Then the data itself.. */
3086         stream.next_in = (void *)buf;
3087         stream.avail_in = len;
3088         do {
3089                 unsigned char *in0 = stream.next_in;
3090                 ret = git_deflate(&stream, Z_FINISH);
3091                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3092                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3093                         die("unable to write sha1 file");
3094                 stream.next_out = compressed;
3095                 stream.avail_out = sizeof(compressed);
3096         } while (ret == Z_OK);
3097
3098         if (ret != Z_STREAM_END)
3099                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3100         ret = git_deflate_end_gently(&stream);
3101         if (ret != Z_OK)
3102                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3103         git_SHA1_Final(parano_sha1, &c);
3104         if (hashcmp(sha1, parano_sha1) != 0)
3105                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3106
3107         close_sha1_file(fd);
3108
3109         if (mtime) {
3110                 struct utimbuf utb;
3111                 utb.actime = mtime;
3112                 utb.modtime = mtime;
3113                 if (utime(tmp_file, &utb) < 0)
3114                         warning("failed utime() on %s: %s",
3115                                 tmp_file, strerror(errno));
3116         }
3117
3118         return finalize_object_file(tmp_file, filename);
3119 }
3120
3121 static int freshen_loose_object(const unsigned char *sha1)
3122 {
3123         return check_and_freshen(sha1, 1);
3124 }
3125
3126 static int freshen_packed_object(const unsigned char *sha1)
3127 {
3128         struct pack_entry e;
3129         if (!find_pack_entry(sha1, &e))
3130                 return 0;
3131         if (e.p->freshened)
3132                 return 1;
3133         if (!freshen_file(e.p->pack_name))
3134                 return 0;
3135         e.p->freshened = 1;
3136         return 1;
3137 }
3138
3139 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3140 {
3141         char hdr[32];
3142         int hdrlen;
3143
3144         /* Normally if we have it in the pack then we do not bother writing
3145          * it out into .git/objects/??/?{38} file.
3146          */
3147         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3148         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3149                 return 0;
3150         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3151 }
3152
3153 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3154                              unsigned char *sha1, unsigned flags)
3155 {
3156         char *header;
3157         int hdrlen, status = 0;
3158
3159         /* type string, SP, %lu of the length plus NUL must fit this */
3160         header = xmalloc(strlen(type) + 32);
3161         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3162
3163         if (!(flags & HASH_WRITE_OBJECT))
3164                 goto cleanup;
3165         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3166                 goto cleanup;
3167         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3168
3169 cleanup:
3170         free(header);
3171         return status;
3172 }
3173
3174 int force_object_loose(const unsigned char *sha1, time_t mtime)
3175 {
3176         void *buf;
3177         unsigned long len;
3178         enum object_type type;
3179         char hdr[32];
3180         int hdrlen;
3181         int ret;
3182
3183         if (has_loose_object(sha1))
3184                 return 0;
3185         buf = read_packed_sha1(sha1, &type, &len);
3186         if (!buf)
3187                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3188         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
3189         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3190         free(buf);
3191
3192         return ret;
3193 }
3194
3195 int has_pack_index(const unsigned char *sha1)
3196 {
3197         struct stat st;
3198         if (stat(sha1_pack_index_name(sha1), &st))
3199                 return 0;
3200         return 1;
3201 }
3202
3203 int has_sha1_pack(const unsigned char *sha1)
3204 {
3205         struct pack_entry e;
3206         return find_pack_entry(sha1, &e);
3207 }
3208
3209 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3210 {
3211         struct pack_entry e;
3212
3213         if (find_pack_entry(sha1, &e))
3214                 return 1;
3215         if (has_loose_object(sha1))
3216                 return 1;
3217         if (flags & HAS_SHA1_QUICK)
3218                 return 0;
3219         reprepare_packed_git();
3220         return find_pack_entry(sha1, &e);
3221 }
3222
3223 static void check_tree(const void *buf, size_t size)
3224 {
3225         struct tree_desc desc;
3226         struct name_entry entry;
3227
3228         init_tree_desc(&desc, buf, size);
3229         while (tree_entry(&desc, &entry))
3230                 /* do nothing
3231                  * tree_entry() will die() on malformed entries */
3232                 ;
3233 }
3234
3235 static void check_commit(const void *buf, size_t size)
3236 {
3237         struct commit c;
3238         memset(&c, 0, sizeof(c));
3239         if (parse_commit_buffer(&c, buf, size))
3240                 die("corrupt commit");
3241 }
3242
3243 static void check_tag(const void *buf, size_t size)
3244 {
3245         struct tag t;
3246         memset(&t, 0, sizeof(t));
3247         if (parse_tag_buffer(&t, buf, size))
3248                 die("corrupt tag");
3249 }
3250
3251 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3252                      enum object_type type,
3253                      const char *path, unsigned flags)
3254 {
3255         int ret, re_allocated = 0;
3256         int write_object = flags & HASH_WRITE_OBJECT;
3257
3258         if (!type)
3259                 type = OBJ_BLOB;
3260
3261         /*
3262          * Convert blobs to git internal format
3263          */
3264         if ((type == OBJ_BLOB) && path) {
3265                 struct strbuf nbuf = STRBUF_INIT;
3266                 if (convert_to_git(path, buf, size, &nbuf,
3267                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3268                         buf = strbuf_detach(&nbuf, &size);
3269                         re_allocated = 1;
3270                 }
3271         }
3272         if (flags & HASH_FORMAT_CHECK) {
3273                 if (type == OBJ_TREE)
3274                         check_tree(buf, size);
3275                 if (type == OBJ_COMMIT)
3276                         check_commit(buf, size);
3277                 if (type == OBJ_TAG)
3278                         check_tag(buf, size);
3279         }
3280
3281         if (write_object)
3282                 ret = write_sha1_file(buf, size, typename(type), sha1);
3283         else
3284                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3285         if (re_allocated)
3286                 free(buf);
3287         return ret;
3288 }
3289
3290 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3291                                      const char *path, unsigned flags)
3292 {
3293         int ret;
3294         const int write_object = flags & HASH_WRITE_OBJECT;
3295         struct strbuf sbuf = STRBUF_INIT;
3296
3297         assert(path);
3298         assert(would_convert_to_git_filter_fd(path));
3299
3300         convert_to_git_filter_fd(path, fd, &sbuf,
3301                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3302
3303         if (write_object)
3304                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3305                                       sha1);
3306         else
3307                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3308                                      sha1);
3309         strbuf_release(&sbuf);
3310         return ret;
3311 }
3312
3313 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3314                       const char *path, unsigned flags)
3315 {
3316         struct strbuf sbuf = STRBUF_INIT;
3317         int ret;
3318
3319         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3320                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3321         else
3322                 ret = -1;
3323         strbuf_release(&sbuf);
3324         return ret;
3325 }
3326
3327 #define SMALL_FILE_SIZE (32*1024)
3328
3329 static int index_core(unsigned char *sha1, int fd, size_t size,
3330                       enum object_type type, const char *path,
3331                       unsigned flags)
3332 {
3333         int ret;
3334
3335         if (!size) {
3336                 ret = index_mem(sha1, "", size, type, path, flags);
3337         } else if (size <= SMALL_FILE_SIZE) {
3338                 char *buf = xmalloc(size);
3339                 if (size == read_in_full(fd, buf, size))
3340                         ret = index_mem(sha1, buf, size, type, path, flags);
3341                 else
3342                         ret = error("short read %s", strerror(errno));
3343                 free(buf);
3344         } else {
3345                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3346                 ret = index_mem(sha1, buf, size, type, path, flags);
3347                 munmap(buf, size);
3348         }
3349         return ret;
3350 }
3351
3352 /*
3353  * This creates one packfile per large blob unless bulk-checkin
3354  * machinery is "plugged".
3355  *
3356  * This also bypasses the usual "convert-to-git" dance, and that is on
3357  * purpose. We could write a streaming version of the converting
3358  * functions and insert that before feeding the data to fast-import
3359  * (or equivalent in-core API described above). However, that is
3360  * somewhat complicated, as we do not know the size of the filter
3361  * result, which we need to know beforehand when writing a git object.
3362  * Since the primary motivation for trying to stream from the working
3363  * tree file and to avoid mmaping it in core is to deal with large
3364  * binary blobs, they generally do not want to get any conversion, and
3365  * callers should avoid this code path when filters are requested.
3366  */
3367 static int index_stream(unsigned char *sha1, int fd, size_t size,
3368                         enum object_type type, const char *path,
3369                         unsigned flags)
3370 {
3371         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3372 }
3373
3374 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3375              enum object_type type, const char *path, unsigned flags)
3376 {
3377         int ret;
3378
3379         /*
3380          * Call xsize_t() only when needed to avoid potentially unnecessary
3381          * die() for large files.
3382          */
3383         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3384                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3385         else if (!S_ISREG(st->st_mode))
3386                 ret = index_pipe(sha1, fd, type, path, flags);
3387         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3388                  (path && would_convert_to_git(path)))
3389                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3390                                  flags);
3391         else
3392                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3393                                    flags);
3394         close(fd);
3395         return ret;
3396 }
3397
3398 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3399 {
3400         int fd;
3401         struct strbuf sb = STRBUF_INIT;
3402
3403         switch (st->st_mode & S_IFMT) {
3404         case S_IFREG:
3405                 fd = open(path, O_RDONLY);
3406                 if (fd < 0)
3407                         return error("open(\"%s\"): %s", path,
3408                                      strerror(errno));
3409                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3410                         return error("%s: failed to insert into database",
3411                                      path);
3412                 break;
3413         case S_IFLNK:
3414                 if (strbuf_readlink(&sb, path, st->st_size)) {
3415                         char *errstr = strerror(errno);
3416                         return error("readlink(\"%s\"): %s", path,
3417                                      errstr);
3418                 }
3419                 if (!(flags & HASH_WRITE_OBJECT))
3420                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3421                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3422                         return error("%s: failed to insert into database",
3423                                      path);
3424                 strbuf_release(&sb);
3425                 break;
3426         case S_IFDIR:
3427                 return resolve_gitlink_ref(path, "HEAD", sha1);
3428         default:
3429                 return error("%s: unsupported file type", path);
3430         }
3431         return 0;
3432 }
3433
3434 int read_pack_header(int fd, struct pack_header *header)
3435 {
3436         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3437                 /* "eof before pack header was fully read" */
3438                 return PH_ERROR_EOF;
3439
3440         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3441                 /* "protocol error (pack signature mismatch detected)" */
3442                 return PH_ERROR_PACK_SIGNATURE;
3443         if (!pack_version_ok(header->hdr_version))
3444                 /* "protocol error (pack version unsupported)" */
3445                 return PH_ERROR_PROTOCOL;
3446         return 0;
3447 }
3448
3449 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3450 {
3451         enum object_type type = sha1_object_info(sha1, NULL);
3452         if (type < 0)
3453                 die("%s is not a valid object", sha1_to_hex(sha1));
3454         if (type != expect)
3455                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3456                     typename(expect));
3457 }
3458
3459 static int for_each_file_in_obj_subdir(int subdir_nr,
3460                                        struct strbuf *path,
3461                                        each_loose_object_fn obj_cb,
3462                                        each_loose_cruft_fn cruft_cb,
3463                                        each_loose_subdir_fn subdir_cb,
3464                                        void *data)
3465 {
3466         size_t baselen = path->len;
3467         DIR *dir = opendir(path->buf);
3468         struct dirent *de;
3469         int r = 0;
3470
3471         if (!dir) {
3472                 if (errno == ENOENT)
3473                         return 0;
3474                 return error("unable to open %s: %s", path->buf, strerror(errno));
3475         }
3476
3477         while ((de = readdir(dir))) {
3478                 if (is_dot_or_dotdot(de->d_name))
3479                         continue;
3480
3481                 strbuf_setlen(path, baselen);
3482                 strbuf_addf(path, "/%s", de->d_name);
3483
3484                 if (strlen(de->d_name) == 38)  {
3485                         char hex[41];
3486                         unsigned char sha1[20];
3487
3488                         snprintf(hex, sizeof(hex), "%02x%s",
3489                                  subdir_nr, de->d_name);
3490                         if (!get_sha1_hex(hex, sha1)) {
3491                                 if (obj_cb) {
3492                                         r = obj_cb(sha1, path->buf, data);
3493                                         if (r)
3494                                                 break;
3495                                 }
3496                                 continue;
3497                         }
3498                 }
3499
3500                 if (cruft_cb) {
3501                         r = cruft_cb(de->d_name, path->buf, data);
3502                         if (r)
3503                                 break;
3504                 }
3505         }
3506         strbuf_setlen(path, baselen);
3507
3508         if (!r && subdir_cb)
3509                 r = subdir_cb(subdir_nr, path->buf, data);
3510
3511         closedir(dir);
3512         return r;
3513 }
3514
3515 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3516                             each_loose_object_fn obj_cb,
3517                             each_loose_cruft_fn cruft_cb,
3518                             each_loose_subdir_fn subdir_cb,
3519                             void *data)
3520 {
3521         size_t baselen = path->len;
3522         int r = 0;
3523         int i;
3524
3525         for (i = 0; i < 256; i++) {
3526                 strbuf_addf(path, "/%02x", i);
3527                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3528                                                 subdir_cb, data);
3529                 strbuf_setlen(path, baselen);
3530                 if (r)
3531                         break;
3532         }
3533
3534         return r;
3535 }
3536
3537 int for_each_loose_file_in_objdir(const char *path,
3538                                   each_loose_object_fn obj_cb,
3539                                   each_loose_cruft_fn cruft_cb,
3540                                   each_loose_subdir_fn subdir_cb,
3541                                   void *data)
3542 {
3543         struct strbuf buf = STRBUF_INIT;
3544         int r;
3545
3546         strbuf_addstr(&buf, path);
3547         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3548                                               subdir_cb, data);
3549         strbuf_release(&buf);
3550
3551         return r;
3552 }
3553
3554 struct loose_alt_odb_data {
3555         each_loose_object_fn *cb;
3556         void *data;
3557 };
3558
3559 static int loose_from_alt_odb(struct alternate_object_database *alt,
3560                               void *vdata)
3561 {
3562         struct loose_alt_odb_data *data = vdata;
3563         struct strbuf buf = STRBUF_INIT;
3564         int r;
3565
3566         /* copy base not including trailing '/' */
3567         strbuf_add(&buf, alt->base, alt->name - alt->base - 1);
3568         r = for_each_loose_file_in_objdir_buf(&buf,
3569                                               data->cb, NULL, NULL,
3570                                               data->data);
3571         strbuf_release(&buf);
3572         return r;
3573 }
3574
3575 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3576 {
3577         struct loose_alt_odb_data alt;
3578         int r;
3579
3580         r = for_each_loose_file_in_objdir(get_object_directory(),
3581                                           cb, NULL, NULL, data);
3582         if (r)
3583                 return r;
3584
3585         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3586                 return 0;
3587
3588         alt.cb = cb;
3589         alt.data = data;
3590         return foreach_alt_odb(loose_from_alt_odb, &alt);
3591 }
3592
3593 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3594 {
3595         uint32_t i;
3596         int r = 0;
3597
3598         for (i = 0; i < p->num_objects; i++) {
3599                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3600
3601                 if (!sha1)
3602                         return error("unable to get sha1 of object %u in %s",
3603                                      i, p->pack_name);
3604
3605                 r = cb(sha1, p, i, data);
3606                 if (r)
3607                         break;
3608         }
3609         return r;
3610 }
3611
3612 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3613 {
3614         struct packed_git *p;
3615         int r = 0;
3616         int pack_errors = 0;
3617
3618         prepare_packed_git();
3619         for (p = packed_git; p; p = p->next) {
3620                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3621                         continue;
3622                 if (open_pack_index(p)) {
3623                         pack_errors = 1;
3624                         continue;
3625                 }
3626                 r = for_each_object_in_pack(p, cb, data);
3627                 if (r)
3628                         break;
3629         }
3630         return r ? r : pack_errors;
3631 }