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