Merge branch 'vs/prompt-avoid-unset-variable'
[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 (!fspathcmp(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_errno("packfile %s cannot be mapped",
1111                                           p->pack_name);
1112                         if (!win->offset && win->len == p->pack_size
1113                                 && !p->do_not_close)
1114                                 close_pack_fd(p);
1115                         pack_mmap_calls++;
1116                         pack_open_windows++;
1117                         if (pack_mapped > peak_pack_mapped)
1118                                 peak_pack_mapped = pack_mapped;
1119                         if (pack_open_windows > peak_pack_open_windows)
1120                                 peak_pack_open_windows = pack_open_windows;
1121                         win->next = p->windows;
1122                         p->windows = win;
1123                 }
1124         }
1125         if (win != *w_cursor) {
1126                 win->last_used = pack_used_ctr++;
1127                 win->inuse_cnt++;
1128                 *w_cursor = win;
1129         }
1130         offset -= win->offset;
1131         if (left)
1132                 *left = win->len - xsize_t(offset);
1133         return win->base + offset;
1134 }
1135
1136 static struct packed_git *alloc_packed_git(int extra)
1137 {
1138         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
1139         memset(p, 0, sizeof(*p));
1140         p->pack_fd = -1;
1141         return p;
1142 }
1143
1144 static void try_to_free_pack_memory(size_t size)
1145 {
1146         release_pack_memory(size);
1147 }
1148
1149 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
1150 {
1151         static int have_set_try_to_free_routine;
1152         struct stat st;
1153         size_t alloc;
1154         struct packed_git *p;
1155
1156         if (!have_set_try_to_free_routine) {
1157                 have_set_try_to_free_routine = 1;
1158                 set_try_to_free_routine(try_to_free_pack_memory);
1159         }
1160
1161         /*
1162          * Make sure a corresponding .pack file exists and that
1163          * the index looks sane.
1164          */
1165         if (!strip_suffix_mem(path, &path_len, ".idx"))
1166                 return NULL;
1167
1168         /*
1169          * ".pack" is long enough to hold any suffix we're adding (and
1170          * the use xsnprintf double-checks that)
1171          */
1172         alloc = st_add3(path_len, strlen(".pack"), 1);
1173         p = alloc_packed_git(alloc);
1174         memcpy(p->pack_name, path, path_len);
1175
1176         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
1177         if (!access(p->pack_name, F_OK))
1178                 p->pack_keep = 1;
1179
1180         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
1181         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1182                 free(p);
1183                 return NULL;
1184         }
1185
1186         /* ok, it looks sane as far as we can check without
1187          * actually mapping the pack file.
1188          */
1189         p->pack_size = st.st_size;
1190         p->pack_local = local;
1191         p->mtime = st.st_mtime;
1192         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1193                 hashclr(p->sha1);
1194         return p;
1195 }
1196
1197 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1198 {
1199         const char *path = sha1_pack_name(sha1);
1200         size_t alloc = st_add(strlen(path), 1);
1201         struct packed_git *p = alloc_packed_git(alloc);
1202
1203         memcpy(p->pack_name, path, alloc); /* includes NUL */
1204         hashcpy(p->sha1, sha1);
1205         if (check_packed_git_idx(idx_path, p)) {
1206                 free(p);
1207                 return NULL;
1208         }
1209
1210         return p;
1211 }
1212
1213 void install_packed_git(struct packed_git *pack)
1214 {
1215         if (pack->pack_fd != -1)
1216                 pack_open_fds++;
1217
1218         pack->next = packed_git;
1219         packed_git = pack;
1220 }
1221
1222 void (*report_garbage)(unsigned seen_bits, const char *path);
1223
1224 static void report_helper(const struct string_list *list,
1225                           int seen_bits, int first, int last)
1226 {
1227         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
1228                 return;
1229
1230         for (; first < last; first++)
1231                 report_garbage(seen_bits, list->items[first].string);
1232 }
1233
1234 static void report_pack_garbage(struct string_list *list)
1235 {
1236         int i, baselen = -1, first = 0, seen_bits = 0;
1237
1238         if (!report_garbage)
1239                 return;
1240
1241         string_list_sort(list);
1242
1243         for (i = 0; i < list->nr; i++) {
1244                 const char *path = list->items[i].string;
1245                 if (baselen != -1 &&
1246                     strncmp(path, list->items[first].string, baselen)) {
1247                         report_helper(list, seen_bits, first, i);
1248                         baselen = -1;
1249                         seen_bits = 0;
1250                 }
1251                 if (baselen == -1) {
1252                         const char *dot = strrchr(path, '.');
1253                         if (!dot) {
1254                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
1255                                 continue;
1256                         }
1257                         baselen = dot - path + 1;
1258                         first = i;
1259                 }
1260                 if (!strcmp(path + baselen, "pack"))
1261                         seen_bits |= 1;
1262                 else if (!strcmp(path + baselen, "idx"))
1263                         seen_bits |= 2;
1264         }
1265         report_helper(list, seen_bits, first, list->nr);
1266 }
1267
1268 static void prepare_packed_git_one(char *objdir, int local)
1269 {
1270         struct strbuf path = STRBUF_INIT;
1271         size_t dirnamelen;
1272         DIR *dir;
1273         struct dirent *de;
1274         struct string_list garbage = STRING_LIST_INIT_DUP;
1275
1276         strbuf_addstr(&path, objdir);
1277         strbuf_addstr(&path, "/pack");
1278         dir = opendir(path.buf);
1279         if (!dir) {
1280                 if (errno != ENOENT)
1281                         error_errno("unable to open object pack directory: %s",
1282                                     path.buf);
1283                 strbuf_release(&path);
1284                 return;
1285         }
1286         strbuf_addch(&path, '/');
1287         dirnamelen = path.len;
1288         while ((de = readdir(dir)) != NULL) {
1289                 struct packed_git *p;
1290                 size_t base_len;
1291
1292                 if (is_dot_or_dotdot(de->d_name))
1293                         continue;
1294
1295                 strbuf_setlen(&path, dirnamelen);
1296                 strbuf_addstr(&path, de->d_name);
1297
1298                 base_len = path.len;
1299                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
1300                         /* Don't reopen a pack we already have. */
1301                         for (p = packed_git; p; p = p->next) {
1302                                 size_t len;
1303                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
1304                                     len == base_len &&
1305                                     !memcmp(p->pack_name, path.buf, len))
1306                                         break;
1307                         }
1308                         if (p == NULL &&
1309                             /*
1310                              * See if it really is a valid .idx file with
1311                              * corresponding .pack file that we can map.
1312                              */
1313                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
1314                                 install_packed_git(p);
1315                 }
1316
1317                 if (!report_garbage)
1318                         continue;
1319
1320                 if (ends_with(de->d_name, ".idx") ||
1321                     ends_with(de->d_name, ".pack") ||
1322                     ends_with(de->d_name, ".bitmap") ||
1323                     ends_with(de->d_name, ".keep"))
1324                         string_list_append(&garbage, path.buf);
1325                 else
1326                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
1327         }
1328         closedir(dir);
1329         report_pack_garbage(&garbage);
1330         string_list_clear(&garbage, 0);
1331         strbuf_release(&path);
1332 }
1333
1334 static int sort_pack(const void *a_, const void *b_)
1335 {
1336         struct packed_git *a = *((struct packed_git **)a_);
1337         struct packed_git *b = *((struct packed_git **)b_);
1338         int st;
1339
1340         /*
1341          * Local packs tend to contain objects specific to our
1342          * variant of the project than remote ones.  In addition,
1343          * remote ones could be on a network mounted filesystem.
1344          * Favor local ones for these reasons.
1345          */
1346         st = a->pack_local - b->pack_local;
1347         if (st)
1348                 return -st;
1349
1350         /*
1351          * Younger packs tend to contain more recent objects,
1352          * and more recent objects tend to get accessed more
1353          * often.
1354          */
1355         if (a->mtime < b->mtime)
1356                 return 1;
1357         else if (a->mtime == b->mtime)
1358                 return 0;
1359         return -1;
1360 }
1361
1362 static void rearrange_packed_git(void)
1363 {
1364         struct packed_git **ary, *p;
1365         int i, n;
1366
1367         for (n = 0, p = packed_git; p; p = p->next)
1368                 n++;
1369         if (n < 2)
1370                 return;
1371
1372         /* prepare an array of packed_git for easier sorting */
1373         ary = xcalloc(n, sizeof(struct packed_git *));
1374         for (n = 0, p = packed_git; p; p = p->next)
1375                 ary[n++] = p;
1376
1377         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1378
1379         /* link them back again */
1380         for (i = 0; i < n - 1; i++)
1381                 ary[i]->next = ary[i + 1];
1382         ary[n - 1]->next = NULL;
1383         packed_git = ary[0];
1384
1385         free(ary);
1386 }
1387
1388 static int prepare_packed_git_run_once = 0;
1389 void prepare_packed_git(void)
1390 {
1391         struct alternate_object_database *alt;
1392
1393         if (prepare_packed_git_run_once)
1394                 return;
1395         prepare_packed_git_one(get_object_directory(), 1);
1396         prepare_alt_odb();
1397         for (alt = alt_odb_list; alt; alt = alt->next) {
1398                 alt->name[-1] = 0;
1399                 prepare_packed_git_one(alt->base, 0);
1400                 alt->name[-1] = '/';
1401         }
1402         rearrange_packed_git();
1403         prepare_packed_git_run_once = 1;
1404 }
1405
1406 void reprepare_packed_git(void)
1407 {
1408         prepare_packed_git_run_once = 0;
1409         prepare_packed_git();
1410 }
1411
1412 static void mark_bad_packed_object(struct packed_git *p,
1413                                    const unsigned char *sha1)
1414 {
1415         unsigned i;
1416         for (i = 0; i < p->num_bad_objects; i++)
1417                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1418                         return;
1419         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1420                                       st_mult(GIT_SHA1_RAWSZ,
1421                                               st_add(p->num_bad_objects, 1)));
1422         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1423         p->num_bad_objects++;
1424 }
1425
1426 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1427 {
1428         struct packed_git *p;
1429         unsigned i;
1430
1431         for (p = packed_git; p; p = p->next)
1432                 for (i = 0; i < p->num_bad_objects; i++)
1433                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1434                                 return p;
1435         return NULL;
1436 }
1437
1438 /*
1439  * With an in-core object data in "map", rehash it to make sure the
1440  * object name actually matches "sha1" to detect object corruption.
1441  * With "map" == NULL, try reading the object named with "sha1" using
1442  * the streaming interface and rehash it to do the same.
1443  */
1444 int check_sha1_signature(const unsigned char *sha1, void *map,
1445                          unsigned long size, const char *type)
1446 {
1447         unsigned char real_sha1[20];
1448         enum object_type obj_type;
1449         struct git_istream *st;
1450         git_SHA_CTX c;
1451         char hdr[32];
1452         int hdrlen;
1453
1454         if (map) {
1455                 hash_sha1_file(map, size, type, real_sha1);
1456                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1457         }
1458
1459         st = open_istream(sha1, &obj_type, &size, NULL);
1460         if (!st)
1461                 return -1;
1462
1463         /* Generate the header */
1464         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(obj_type), size) + 1;
1465
1466         /* Sha1.. */
1467         git_SHA1_Init(&c);
1468         git_SHA1_Update(&c, hdr, hdrlen);
1469         for (;;) {
1470                 char buf[1024 * 16];
1471                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1472
1473                 if (readlen < 0) {
1474                         close_istream(st);
1475                         return -1;
1476                 }
1477                 if (!readlen)
1478                         break;
1479                 git_SHA1_Update(&c, buf, readlen);
1480         }
1481         git_SHA1_Final(real_sha1, &c);
1482         close_istream(st);
1483         return hashcmp(sha1, real_sha1) ? -1 : 0;
1484 }
1485
1486 int git_open_noatime(const char *name)
1487 {
1488         static int sha1_file_open_flag = O_NOATIME;
1489
1490         for (;;) {
1491                 int fd;
1492
1493                 errno = 0;
1494                 fd = open(name, O_RDONLY | sha1_file_open_flag);
1495                 if (fd >= 0)
1496                         return fd;
1497
1498                 /* Might the failure be due to O_NOATIME? */
1499                 if (errno != ENOENT && sha1_file_open_flag) {
1500                         sha1_file_open_flag = 0;
1501                         continue;
1502                 }
1503
1504                 return -1;
1505         }
1506 }
1507
1508 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1509 {
1510         struct alternate_object_database *alt;
1511
1512         if (!lstat(sha1_file_name(sha1), st))
1513                 return 0;
1514
1515         prepare_alt_odb();
1516         errno = ENOENT;
1517         for (alt = alt_odb_list; alt; alt = alt->next) {
1518                 fill_sha1_path(alt->name, sha1);
1519                 if (!lstat(alt->base, st))
1520                         return 0;
1521         }
1522
1523         return -1;
1524 }
1525
1526 static int open_sha1_file(const unsigned char *sha1)
1527 {
1528         int fd;
1529         struct alternate_object_database *alt;
1530         int most_interesting_errno;
1531
1532         fd = git_open_noatime(sha1_file_name(sha1));
1533         if (fd >= 0)
1534                 return fd;
1535         most_interesting_errno = errno;
1536
1537         prepare_alt_odb();
1538         for (alt = alt_odb_list; alt; alt = alt->next) {
1539                 fill_sha1_path(alt->name, sha1);
1540                 fd = git_open_noatime(alt->base);
1541                 if (fd >= 0)
1542                         return fd;
1543                 if (most_interesting_errno == ENOENT)
1544                         most_interesting_errno = errno;
1545         }
1546         errno = most_interesting_errno;
1547         return -1;
1548 }
1549
1550 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1551 {
1552         void *map;
1553         int fd;
1554
1555         fd = open_sha1_file(sha1);
1556         map = NULL;
1557         if (fd >= 0) {
1558                 struct stat st;
1559
1560                 if (!fstat(fd, &st)) {
1561                         *size = xsize_t(st.st_size);
1562                         if (!*size) {
1563                                 /* mmap() is forbidden on empty files */
1564                                 error("object file %s is empty", sha1_file_name(sha1));
1565                                 return NULL;
1566                         }
1567                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1568                 }
1569                 close(fd);
1570         }
1571         return map;
1572 }
1573
1574 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1575                 unsigned long len, enum object_type *type, unsigned long *sizep)
1576 {
1577         unsigned shift;
1578         unsigned long size, c;
1579         unsigned long used = 0;
1580
1581         c = buf[used++];
1582         *type = (c >> 4) & 7;
1583         size = c & 15;
1584         shift = 4;
1585         while (c & 0x80) {
1586                 if (len <= used || bitsizeof(long) <= shift) {
1587                         error("bad object header");
1588                         size = used = 0;
1589                         break;
1590                 }
1591                 c = buf[used++];
1592                 size += (c & 0x7f) << shift;
1593                 shift += 7;
1594         }
1595         *sizep = size;
1596         return used;
1597 }
1598
1599 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1600 {
1601         /* Get the data stream */
1602         memset(stream, 0, sizeof(*stream));
1603         stream->next_in = map;
1604         stream->avail_in = mapsize;
1605         stream->next_out = buffer;
1606         stream->avail_out = bufsiz;
1607
1608         git_inflate_init(stream);
1609         return git_inflate(stream, 0);
1610 }
1611
1612 static int unpack_sha1_header_to_strbuf(git_zstream *stream, unsigned char *map,
1613                                         unsigned long mapsize, void *buffer,
1614                                         unsigned long bufsiz, struct strbuf *header)
1615 {
1616         int status;
1617
1618         status = unpack_sha1_header(stream, map, mapsize, buffer, bufsiz);
1619
1620         /*
1621          * Check if entire header is unpacked in the first iteration.
1622          */
1623         if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1624                 return 0;
1625
1626         /*
1627          * buffer[0..bufsiz] was not large enough.  Copy the partial
1628          * result out to header, and then append the result of further
1629          * reading the stream.
1630          */
1631         strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1632         stream->next_out = buffer;
1633         stream->avail_out = bufsiz;
1634
1635         do {
1636                 status = git_inflate(stream, 0);
1637                 strbuf_add(header, buffer, stream->next_out - (unsigned char *)buffer);
1638                 if (memchr(buffer, '\0', stream->next_out - (unsigned char *)buffer))
1639                         return 0;
1640                 stream->next_out = buffer;
1641                 stream->avail_out = bufsiz;
1642         } while (status != Z_STREAM_END);
1643         return -1;
1644 }
1645
1646 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1647 {
1648         int bytes = strlen(buffer) + 1;
1649         unsigned char *buf = xmallocz(size);
1650         unsigned long n;
1651         int status = Z_OK;
1652
1653         n = stream->total_out - bytes;
1654         if (n > size)
1655                 n = size;
1656         memcpy(buf, (char *) buffer + bytes, n);
1657         bytes = n;
1658         if (bytes <= size) {
1659                 /*
1660                  * The above condition must be (bytes <= size), not
1661                  * (bytes < size).  In other words, even though we
1662                  * expect no more output and set avail_out to zero,
1663                  * the input zlib stream may have bytes that express
1664                  * "this concludes the stream", and we *do* want to
1665                  * eat that input.
1666                  *
1667                  * Otherwise we would not be able to test that we
1668                  * consumed all the input to reach the expected size;
1669                  * we also want to check that zlib tells us that all
1670                  * went well with status == Z_STREAM_END at the end.
1671                  */
1672                 stream->next_out = buf + bytes;
1673                 stream->avail_out = size - bytes;
1674                 while (status == Z_OK)
1675                         status = git_inflate(stream, Z_FINISH);
1676         }
1677         if (status == Z_STREAM_END && !stream->avail_in) {
1678                 git_inflate_end(stream);
1679                 return buf;
1680         }
1681
1682         if (status < 0)
1683                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1684         else if (stream->avail_in)
1685                 error("garbage at end of loose object '%s'",
1686                       sha1_to_hex(sha1));
1687         free(buf);
1688         return NULL;
1689 }
1690
1691 /*
1692  * We used to just use "sscanf()", but that's actually way
1693  * too permissive for what we want to check. So do an anal
1694  * object header parse by hand.
1695  */
1696 static int parse_sha1_header_extended(const char *hdr, struct object_info *oi,
1697                                unsigned int flags)
1698 {
1699         const char *type_buf = hdr;
1700         unsigned long size;
1701         int type, type_len = 0;
1702
1703         /*
1704          * The type can be of any size but is followed by
1705          * a space.
1706          */
1707         for (;;) {
1708                 char c = *hdr++;
1709                 if (c == ' ')
1710                         break;
1711                 type_len++;
1712         }
1713
1714         type = type_from_string_gently(type_buf, type_len, 1);
1715         if (oi->typename)
1716                 strbuf_add(oi->typename, type_buf, type_len);
1717         /*
1718          * Set type to 0 if its an unknown object and
1719          * we're obtaining the type using '--allow-unkown-type'
1720          * option.
1721          */
1722         if ((flags & LOOKUP_UNKNOWN_OBJECT) && (type < 0))
1723                 type = 0;
1724         else if (type < 0)
1725                 die("invalid object type");
1726         if (oi->typep)
1727                 *oi->typep = type;
1728
1729         /*
1730          * The length must follow immediately, and be in canonical
1731          * decimal format (ie "010" is not valid).
1732          */
1733         size = *hdr++ - '0';
1734         if (size > 9)
1735                 return -1;
1736         if (size) {
1737                 for (;;) {
1738                         unsigned long c = *hdr - '0';
1739                         if (c > 9)
1740                                 break;
1741                         hdr++;
1742                         size = size * 10 + c;
1743                 }
1744         }
1745
1746         if (oi->sizep)
1747                 *oi->sizep = size;
1748
1749         /*
1750          * The length must be followed by a zero byte
1751          */
1752         return *hdr ? -1 : type;
1753 }
1754
1755 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1756 {
1757         struct object_info oi;
1758
1759         oi.sizep = sizep;
1760         oi.typename = NULL;
1761         oi.typep = NULL;
1762         return parse_sha1_header_extended(hdr, &oi, LOOKUP_REPLACE_OBJECT);
1763 }
1764
1765 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1766 {
1767         int ret;
1768         git_zstream stream;
1769         char hdr[8192];
1770
1771         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1772         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1773                 return NULL;
1774
1775         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1776 }
1777
1778 unsigned long get_size_from_delta(struct packed_git *p,
1779                                   struct pack_window **w_curs,
1780                                   off_t curpos)
1781 {
1782         const unsigned char *data;
1783         unsigned char delta_head[20], *in;
1784         git_zstream stream;
1785         int st;
1786
1787         memset(&stream, 0, sizeof(stream));
1788         stream.next_out = delta_head;
1789         stream.avail_out = sizeof(delta_head);
1790
1791         git_inflate_init(&stream);
1792         do {
1793                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1794                 stream.next_in = in;
1795                 st = git_inflate(&stream, Z_FINISH);
1796                 curpos += stream.next_in - in;
1797         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1798                  stream.total_out < sizeof(delta_head));
1799         git_inflate_end(&stream);
1800         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1801                 error("delta data unpack-initial failed");
1802                 return 0;
1803         }
1804
1805         /* Examine the initial part of the delta to figure out
1806          * the result size.
1807          */
1808         data = delta_head;
1809
1810         /* ignore base size */
1811         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1812
1813         /* Read the result size */
1814         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1815 }
1816
1817 static off_t get_delta_base(struct packed_git *p,
1818                                     struct pack_window **w_curs,
1819                                     off_t *curpos,
1820                                     enum object_type type,
1821                                     off_t delta_obj_offset)
1822 {
1823         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1824         off_t base_offset;
1825
1826         /* use_pack() assured us we have [base_info, base_info + 20)
1827          * as a range that we can look at without walking off the
1828          * end of the mapped window.  Its actually the hash size
1829          * that is assured.  An OFS_DELTA longer than the hash size
1830          * is stupid, as then a REF_DELTA would be smaller to store.
1831          */
1832         if (type == OBJ_OFS_DELTA) {
1833                 unsigned used = 0;
1834                 unsigned char c = base_info[used++];
1835                 base_offset = c & 127;
1836                 while (c & 128) {
1837                         base_offset += 1;
1838                         if (!base_offset || MSB(base_offset, 7))
1839                                 return 0;  /* overflow */
1840                         c = base_info[used++];
1841                         base_offset = (base_offset << 7) + (c & 127);
1842                 }
1843                 base_offset = delta_obj_offset - base_offset;
1844                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1845                         return 0;  /* out of bound */
1846                 *curpos += used;
1847         } else if (type == OBJ_REF_DELTA) {
1848                 /* The base entry _must_ be in the same pack */
1849                 base_offset = find_pack_entry_one(base_info, p);
1850                 *curpos += 20;
1851         } else
1852                 die("I am totally screwed");
1853         return base_offset;
1854 }
1855
1856 /*
1857  * Like get_delta_base above, but we return the sha1 instead of the pack
1858  * offset. This means it is cheaper for REF deltas (we do not have to do
1859  * the final object lookup), but more expensive for OFS deltas (we
1860  * have to load the revidx to convert the offset back into a sha1).
1861  */
1862 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1863                                                 struct pack_window **w_curs,
1864                                                 off_t curpos,
1865                                                 enum object_type type,
1866                                                 off_t delta_obj_offset)
1867 {
1868         if (type == OBJ_REF_DELTA) {
1869                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1870                 return base;
1871         } else if (type == OBJ_OFS_DELTA) {
1872                 struct revindex_entry *revidx;
1873                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1874                                                    type, delta_obj_offset);
1875
1876                 if (!base_offset)
1877                         return NULL;
1878
1879                 revidx = find_pack_revindex(p, base_offset);
1880                 if (!revidx)
1881                         return NULL;
1882
1883                 return nth_packed_object_sha1(p, revidx->nr);
1884         } else
1885                 return NULL;
1886 }
1887
1888 int unpack_object_header(struct packed_git *p,
1889                          struct pack_window **w_curs,
1890                          off_t *curpos,
1891                          unsigned long *sizep)
1892 {
1893         unsigned char *base;
1894         unsigned long left;
1895         unsigned long used;
1896         enum object_type type;
1897
1898         /* use_pack() assures us we have [base, base + 20) available
1899          * as a range that we can look at.  (Its actually the hash
1900          * size that is assured.)  With our object header encoding
1901          * the maximum deflated object size is 2^137, which is just
1902          * insane, so we know won't exceed what we have been given.
1903          */
1904         base = use_pack(p, w_curs, *curpos, &left);
1905         used = unpack_object_header_buffer(base, left, &type, sizep);
1906         if (!used) {
1907                 type = OBJ_BAD;
1908         } else
1909                 *curpos += used;
1910
1911         return type;
1912 }
1913
1914 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1915 {
1916         int type;
1917         struct revindex_entry *revidx;
1918         const unsigned char *sha1;
1919         revidx = find_pack_revindex(p, obj_offset);
1920         if (!revidx)
1921                 return OBJ_BAD;
1922         sha1 = nth_packed_object_sha1(p, revidx->nr);
1923         mark_bad_packed_object(p, sha1);
1924         type = sha1_object_info(sha1, NULL);
1925         if (type <= OBJ_NONE)
1926                 return OBJ_BAD;
1927         return type;
1928 }
1929
1930 #define POI_STACK_PREALLOC 64
1931
1932 static enum object_type packed_to_object_type(struct packed_git *p,
1933                                               off_t obj_offset,
1934                                               enum object_type type,
1935                                               struct pack_window **w_curs,
1936                                               off_t curpos)
1937 {
1938         off_t small_poi_stack[POI_STACK_PREALLOC];
1939         off_t *poi_stack = small_poi_stack;
1940         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1941
1942         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1943                 off_t base_offset;
1944                 unsigned long size;
1945                 /* Push the object we're going to leave behind */
1946                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1947                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1948                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1949                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1950                 } else {
1951                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1952                 }
1953                 poi_stack[poi_stack_nr++] = obj_offset;
1954                 /* If parsing the base offset fails, just unwind */
1955                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1956                 if (!base_offset)
1957                         goto unwind;
1958                 curpos = obj_offset = base_offset;
1959                 type = unpack_object_header(p, w_curs, &curpos, &size);
1960                 if (type <= OBJ_NONE) {
1961                         /* If getting the base itself fails, we first
1962                          * retry the base, otherwise unwind */
1963                         type = retry_bad_packed_offset(p, base_offset);
1964                         if (type > OBJ_NONE)
1965                                 goto out;
1966                         goto unwind;
1967                 }
1968         }
1969
1970         switch (type) {
1971         case OBJ_BAD:
1972         case OBJ_COMMIT:
1973         case OBJ_TREE:
1974         case OBJ_BLOB:
1975         case OBJ_TAG:
1976                 break;
1977         default:
1978                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1979                       type, (uintmax_t)obj_offset, p->pack_name);
1980                 type = OBJ_BAD;
1981         }
1982
1983 out:
1984         if (poi_stack != small_poi_stack)
1985                 free(poi_stack);
1986         return type;
1987
1988 unwind:
1989         while (poi_stack_nr) {
1990                 obj_offset = poi_stack[--poi_stack_nr];
1991                 type = retry_bad_packed_offset(p, obj_offset);
1992                 if (type > OBJ_NONE)
1993                         goto out;
1994         }
1995         type = OBJ_BAD;
1996         goto out;
1997 }
1998
1999 static int packed_object_info(struct packed_git *p, off_t obj_offset,
2000                               struct object_info *oi)
2001 {
2002         struct pack_window *w_curs = NULL;
2003         unsigned long size;
2004         off_t curpos = obj_offset;
2005         enum object_type type;
2006
2007         /*
2008          * We always get the representation type, but only convert it to
2009          * a "real" type later if the caller is interested.
2010          */
2011         type = unpack_object_header(p, &w_curs, &curpos, &size);
2012
2013         if (oi->sizep) {
2014                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2015                         off_t tmp_pos = curpos;
2016                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
2017                                                            type, obj_offset);
2018                         if (!base_offset) {
2019                                 type = OBJ_BAD;
2020                                 goto out;
2021                         }
2022                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
2023                         if (*oi->sizep == 0) {
2024                                 type = OBJ_BAD;
2025                                 goto out;
2026                         }
2027                 } else {
2028                         *oi->sizep = size;
2029                 }
2030         }
2031
2032         if (oi->disk_sizep) {
2033                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2034                 *oi->disk_sizep = revidx[1].offset - obj_offset;
2035         }
2036
2037         if (oi->typep) {
2038                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
2039                 if (*oi->typep < 0) {
2040                         type = OBJ_BAD;
2041                         goto out;
2042                 }
2043         }
2044
2045         if (oi->delta_base_sha1) {
2046                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
2047                         const unsigned char *base;
2048
2049                         base = get_delta_base_sha1(p, &w_curs, curpos,
2050                                                    type, obj_offset);
2051                         if (!base) {
2052                                 type = OBJ_BAD;
2053                                 goto out;
2054                         }
2055
2056                         hashcpy(oi->delta_base_sha1, base);
2057                 } else
2058                         hashclr(oi->delta_base_sha1);
2059         }
2060
2061 out:
2062         unuse_pack(&w_curs);
2063         return type;
2064 }
2065
2066 static void *unpack_compressed_entry(struct packed_git *p,
2067                                     struct pack_window **w_curs,
2068                                     off_t curpos,
2069                                     unsigned long size)
2070 {
2071         int st;
2072         git_zstream stream;
2073         unsigned char *buffer, *in;
2074
2075         buffer = xmallocz_gently(size);
2076         if (!buffer)
2077                 return NULL;
2078         memset(&stream, 0, sizeof(stream));
2079         stream.next_out = buffer;
2080         stream.avail_out = size + 1;
2081
2082         git_inflate_init(&stream);
2083         do {
2084                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
2085                 stream.next_in = in;
2086                 st = git_inflate(&stream, Z_FINISH);
2087                 if (!stream.avail_out)
2088                         break; /* the payload is larger than it should be */
2089                 curpos += stream.next_in - in;
2090         } while (st == Z_OK || st == Z_BUF_ERROR);
2091         git_inflate_end(&stream);
2092         if ((st != Z_STREAM_END) || stream.total_out != size) {
2093                 free(buffer);
2094                 return NULL;
2095         }
2096
2097         return buffer;
2098 }
2099
2100 #define MAX_DELTA_CACHE (256)
2101
2102 static size_t delta_base_cached;
2103
2104 static struct delta_base_cache_lru_list {
2105         struct delta_base_cache_lru_list *prev;
2106         struct delta_base_cache_lru_list *next;
2107 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
2108
2109 static struct delta_base_cache_entry {
2110         struct delta_base_cache_lru_list lru;
2111         void *data;
2112         struct packed_git *p;
2113         off_t base_offset;
2114         unsigned long size;
2115         enum object_type type;
2116 } delta_base_cache[MAX_DELTA_CACHE];
2117
2118 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
2119 {
2120         unsigned long hash;
2121
2122         hash = (unsigned long)(intptr_t)p + (unsigned long)base_offset;
2123         hash += (hash >> 8) + (hash >> 16);
2124         return hash % MAX_DELTA_CACHE;
2125 }
2126
2127 static struct delta_base_cache_entry *
2128 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
2129 {
2130         unsigned long hash = pack_entry_hash(p, base_offset);
2131         return delta_base_cache + hash;
2132 }
2133
2134 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
2135                                      struct packed_git *p, off_t base_offset)
2136 {
2137         return (ent->data && ent->p == p && ent->base_offset == base_offset);
2138 }
2139
2140 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
2141 {
2142         struct delta_base_cache_entry *ent;
2143         ent = get_delta_base_cache_entry(p, base_offset);
2144         return eq_delta_base_cache_entry(ent, p, base_offset);
2145 }
2146
2147 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
2148 {
2149         ent->data = NULL;
2150         ent->lru.next->prev = ent->lru.prev;
2151         ent->lru.prev->next = ent->lru.next;
2152         delta_base_cached -= ent->size;
2153 }
2154
2155 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2156         unsigned long *base_size, enum object_type *type, int keep_cache)
2157 {
2158         struct delta_base_cache_entry *ent;
2159         void *ret;
2160
2161         ent = get_delta_base_cache_entry(p, base_offset);
2162
2163         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2164                 return unpack_entry(p, base_offset, type, base_size);
2165
2166         ret = ent->data;
2167
2168         if (!keep_cache)
2169                 clear_delta_base_cache_entry(ent);
2170         else
2171                 ret = xmemdupz(ent->data, ent->size);
2172         *type = ent->type;
2173         *base_size = ent->size;
2174         return ret;
2175 }
2176
2177 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2178 {
2179         if (ent->data) {
2180                 free(ent->data);
2181                 ent->data = NULL;
2182                 ent->lru.next->prev = ent->lru.prev;
2183                 ent->lru.prev->next = ent->lru.next;
2184                 delta_base_cached -= ent->size;
2185         }
2186 }
2187
2188 void clear_delta_base_cache(void)
2189 {
2190         unsigned long p;
2191         for (p = 0; p < MAX_DELTA_CACHE; p++)
2192                 release_delta_base_cache(&delta_base_cache[p]);
2193 }
2194
2195 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2196         void *base, unsigned long base_size, enum object_type type)
2197 {
2198         unsigned long hash = pack_entry_hash(p, base_offset);
2199         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2200         struct delta_base_cache_lru_list *lru;
2201
2202         release_delta_base_cache(ent);
2203         delta_base_cached += base_size;
2204
2205         for (lru = delta_base_cache_lru.next;
2206              delta_base_cached > delta_base_cache_limit
2207              && lru != &delta_base_cache_lru;
2208              lru = lru->next) {
2209                 struct delta_base_cache_entry *f = (void *)lru;
2210                 if (f->type == OBJ_BLOB)
2211                         release_delta_base_cache(f);
2212         }
2213         for (lru = delta_base_cache_lru.next;
2214              delta_base_cached > delta_base_cache_limit
2215              && lru != &delta_base_cache_lru;
2216              lru = lru->next) {
2217                 struct delta_base_cache_entry *f = (void *)lru;
2218                 release_delta_base_cache(f);
2219         }
2220
2221         ent->p = p;
2222         ent->base_offset = base_offset;
2223         ent->type = type;
2224         ent->data = base;
2225         ent->size = base_size;
2226         ent->lru.next = &delta_base_cache_lru;
2227         ent->lru.prev = delta_base_cache_lru.prev;
2228         delta_base_cache_lru.prev->next = &ent->lru;
2229         delta_base_cache_lru.prev = &ent->lru;
2230 }
2231
2232 static void *read_object(const unsigned char *sha1, enum object_type *type,
2233                          unsigned long *size);
2234
2235 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2236 {
2237         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
2238         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
2239                          p->pack_name, (uintmax_t)obj_offset);
2240 }
2241
2242 int do_check_packed_object_crc;
2243
2244 #define UNPACK_ENTRY_STACK_PREALLOC 64
2245 struct unpack_entry_stack_ent {
2246         off_t obj_offset;
2247         off_t curpos;
2248         unsigned long size;
2249 };
2250
2251 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2252                    enum object_type *final_type, unsigned long *final_size)
2253 {
2254         struct pack_window *w_curs = NULL;
2255         off_t curpos = obj_offset;
2256         void *data = NULL;
2257         unsigned long size;
2258         enum object_type type;
2259         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2260         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2261         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2262         int base_from_cache = 0;
2263
2264         write_pack_access_log(p, obj_offset);
2265
2266         /* PHASE 1: drill down to the innermost base object */
2267         for (;;) {
2268                 off_t base_offset;
2269                 int i;
2270                 struct delta_base_cache_entry *ent;
2271
2272                 ent = get_delta_base_cache_entry(p, curpos);
2273                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2274                         type = ent->type;
2275                         data = ent->data;
2276                         size = ent->size;
2277                         clear_delta_base_cache_entry(ent);
2278                         base_from_cache = 1;
2279                         break;
2280                 }
2281
2282                 if (do_check_packed_object_crc && p->index_version > 1) {
2283                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2284                         unsigned long len = revidx[1].offset - obj_offset;
2285                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2286                                 const unsigned char *sha1 =
2287                                         nth_packed_object_sha1(p, revidx->nr);
2288                                 error("bad packed object CRC for %s",
2289                                       sha1_to_hex(sha1));
2290                                 mark_bad_packed_object(p, sha1);
2291                                 unuse_pack(&w_curs);
2292                                 return NULL;
2293                         }
2294                 }
2295
2296                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2297                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2298                         break;
2299
2300                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2301                 if (!base_offset) {
2302                         error("failed to validate delta base reference "
2303                               "at offset %"PRIuMAX" from %s",
2304                               (uintmax_t)curpos, p->pack_name);
2305                         /* bail to phase 2, in hopes of recovery */
2306                         data = NULL;
2307                         break;
2308                 }
2309
2310                 /* push object, proceed to base */
2311                 if (delta_stack_nr >= delta_stack_alloc
2312                     && delta_stack == small_delta_stack) {
2313                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2314                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
2315                         memcpy(delta_stack, small_delta_stack,
2316                                sizeof(*delta_stack)*delta_stack_nr);
2317                 } else {
2318                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2319                 }
2320                 i = delta_stack_nr++;
2321                 delta_stack[i].obj_offset = obj_offset;
2322                 delta_stack[i].curpos = curpos;
2323                 delta_stack[i].size = size;
2324
2325                 curpos = obj_offset = base_offset;
2326         }
2327
2328         /* PHASE 2: handle the base */
2329         switch (type) {
2330         case OBJ_OFS_DELTA:
2331         case OBJ_REF_DELTA:
2332                 if (data)
2333                         die("BUG in unpack_entry: left loop at a valid delta");
2334                 break;
2335         case OBJ_COMMIT:
2336         case OBJ_TREE:
2337         case OBJ_BLOB:
2338         case OBJ_TAG:
2339                 if (!base_from_cache)
2340                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2341                 break;
2342         default:
2343                 data = NULL;
2344                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2345                       type, (uintmax_t)obj_offset, p->pack_name);
2346         }
2347
2348         /* PHASE 3: apply deltas in order */
2349
2350         /* invariants:
2351          *   'data' holds the base data, or NULL if there was corruption
2352          */
2353         while (delta_stack_nr) {
2354                 void *delta_data;
2355                 void *base = data;
2356                 unsigned long delta_size, base_size = size;
2357                 int i;
2358
2359                 data = NULL;
2360
2361                 if (base)
2362                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2363
2364                 if (!base) {
2365                         /*
2366                          * We're probably in deep shit, but let's try to fetch
2367                          * the required base anyway from another pack or loose.
2368                          * This is costly but should happen only in the presence
2369                          * of a corrupted pack, and is better than failing outright.
2370                          */
2371                         struct revindex_entry *revidx;
2372                         const unsigned char *base_sha1;
2373                         revidx = find_pack_revindex(p, obj_offset);
2374                         if (revidx) {
2375                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2376                                 error("failed to read delta base object %s"
2377                                       " at offset %"PRIuMAX" from %s",
2378                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2379                                       p->pack_name);
2380                                 mark_bad_packed_object(p, base_sha1);
2381                                 base = read_object(base_sha1, &type, &base_size);
2382                         }
2383                 }
2384
2385                 i = --delta_stack_nr;
2386                 obj_offset = delta_stack[i].obj_offset;
2387                 curpos = delta_stack[i].curpos;
2388                 delta_size = delta_stack[i].size;
2389
2390                 if (!base)
2391                         continue;
2392
2393                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2394
2395                 if (!delta_data) {
2396                         error("failed to unpack compressed delta "
2397                               "at offset %"PRIuMAX" from %s",
2398                               (uintmax_t)curpos, p->pack_name);
2399                         data = NULL;
2400                         continue;
2401                 }
2402
2403                 data = patch_delta(base, base_size,
2404                                    delta_data, delta_size,
2405                                    &size);
2406
2407                 /*
2408                  * We could not apply the delta; warn the user, but keep going.
2409                  * Our failure will be noticed either in the next iteration of
2410                  * the loop, or if this is the final delta, in the caller when
2411                  * we return NULL. Those code paths will take care of making
2412                  * a more explicit warning and retrying with another copy of
2413                  * the object.
2414                  */
2415                 if (!data)
2416                         error("failed to apply delta");
2417
2418                 free(delta_data);
2419         }
2420
2421         *final_type = type;
2422         *final_size = size;
2423
2424         unuse_pack(&w_curs);
2425
2426         if (delta_stack != small_delta_stack)
2427                 free(delta_stack);
2428
2429         return data;
2430 }
2431
2432 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2433                                             uint32_t n)
2434 {
2435         const unsigned char *index = p->index_data;
2436         if (!index) {
2437                 if (open_pack_index(p))
2438                         return NULL;
2439                 index = p->index_data;
2440         }
2441         if (n >= p->num_objects)
2442                 return NULL;
2443         index += 4 * 256;
2444         if (p->index_version == 1) {
2445                 return index + 24 * n + 4;
2446         } else {
2447                 index += 8;
2448                 return index + 20 * n;
2449         }
2450 }
2451
2452 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
2453 {
2454         const unsigned char *ptr = vptr;
2455         const unsigned char *start = p->index_data;
2456         const unsigned char *end = start + p->index_size;
2457         if (ptr < start)
2458                 die(_("offset before start of pack index for %s (corrupt index?)"),
2459                     p->pack_name);
2460         /* No need to check for underflow; .idx files must be at least 8 bytes */
2461         if (ptr >= end - 8)
2462                 die(_("offset beyond end of pack index for %s (truncated index?)"),
2463                     p->pack_name);
2464 }
2465
2466 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2467 {
2468         const unsigned char *index = p->index_data;
2469         index += 4 * 256;
2470         if (p->index_version == 1) {
2471                 return ntohl(*((uint32_t *)(index + 24 * n)));
2472         } else {
2473                 uint32_t off;
2474                 index += 8 + p->num_objects * (20 + 4);
2475                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2476                 if (!(off & 0x80000000))
2477                         return off;
2478                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2479                 check_pack_index_ptr(p, index);
2480                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2481                                    ntohl(*((uint32_t *)(index + 4)));
2482         }
2483 }
2484
2485 off_t find_pack_entry_one(const unsigned char *sha1,
2486                                   struct packed_git *p)
2487 {
2488         const uint32_t *level1_ofs = p->index_data;
2489         const unsigned char *index = p->index_data;
2490         unsigned hi, lo, stride;
2491         static int use_lookup = -1;
2492         static int debug_lookup = -1;
2493
2494         if (debug_lookup < 0)
2495                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2496
2497         if (!index) {
2498                 if (open_pack_index(p))
2499                         return 0;
2500                 level1_ofs = p->index_data;
2501                 index = p->index_data;
2502         }
2503         if (p->index_version > 1) {
2504                 level1_ofs += 2;
2505                 index += 8;
2506         }
2507         index += 4 * 256;
2508         hi = ntohl(level1_ofs[*sha1]);
2509         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2510         if (p->index_version > 1) {
2511                 stride = 20;
2512         } else {
2513                 stride = 24;
2514                 index += 4;
2515         }
2516
2517         if (debug_lookup)
2518                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2519                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2520
2521         if (use_lookup < 0)
2522                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2523         if (use_lookup) {
2524                 int pos = sha1_entry_pos(index, stride, 0,
2525                                          lo, hi, p->num_objects, sha1);
2526                 if (pos < 0)
2527                         return 0;
2528                 return nth_packed_object_offset(p, pos);
2529         }
2530
2531         do {
2532                 unsigned mi = (lo + hi) / 2;
2533                 int cmp = hashcmp(index + mi * stride, sha1);
2534
2535                 if (debug_lookup)
2536                         printf("lo %u hi %u rg %u mi %u\n",
2537                                lo, hi, hi - lo, mi);
2538                 if (!cmp)
2539                         return nth_packed_object_offset(p, mi);
2540                 if (cmp > 0)
2541                         hi = mi;
2542                 else
2543                         lo = mi+1;
2544         } while (lo < hi);
2545         return 0;
2546 }
2547
2548 int is_pack_valid(struct packed_git *p)
2549 {
2550         /* An already open pack is known to be valid. */
2551         if (p->pack_fd != -1)
2552                 return 1;
2553
2554         /* If the pack has one window completely covering the
2555          * file size, the pack is known to be valid even if
2556          * the descriptor is not currently open.
2557          */
2558         if (p->windows) {
2559                 struct pack_window *w = p->windows;
2560
2561                 if (!w->offset && w->len == p->pack_size)
2562                         return 1;
2563         }
2564
2565         /* Force the pack to open to prove its valid. */
2566         return !open_packed_git(p);
2567 }
2568
2569 static int fill_pack_entry(const unsigned char *sha1,
2570                            struct pack_entry *e,
2571                            struct packed_git *p)
2572 {
2573         off_t offset;
2574
2575         if (p->num_bad_objects) {
2576                 unsigned i;
2577                 for (i = 0; i < p->num_bad_objects; i++)
2578                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2579                                 return 0;
2580         }
2581
2582         offset = find_pack_entry_one(sha1, p);
2583         if (!offset)
2584                 return 0;
2585
2586         /*
2587          * We are about to tell the caller where they can locate the
2588          * requested object.  We better make sure the packfile is
2589          * still here and can be accessed before supplying that
2590          * answer, as it may have been deleted since the index was
2591          * loaded!
2592          */
2593         if (!is_pack_valid(p))
2594                 return 0;
2595         e->offset = offset;
2596         e->p = p;
2597         hashcpy(e->sha1, sha1);
2598         return 1;
2599 }
2600
2601 /*
2602  * Iff a pack file contains the object named by sha1, return true and
2603  * store its location to e.
2604  */
2605 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2606 {
2607         struct packed_git *p;
2608
2609         prepare_packed_git();
2610         if (!packed_git)
2611                 return 0;
2612
2613         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2614                 return 1;
2615
2616         for (p = packed_git; p; p = p->next) {
2617                 if (p == last_found_pack)
2618                         continue; /* we already checked this one */
2619
2620                 if (fill_pack_entry(sha1, e, p)) {
2621                         last_found_pack = p;
2622                         return 1;
2623                 }
2624         }
2625         return 0;
2626 }
2627
2628 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2629                                   struct packed_git *packs)
2630 {
2631         struct packed_git *p;
2632
2633         for (p = packs; p; p = p->next) {
2634                 if (find_pack_entry_one(sha1, p))
2635                         return p;
2636         }
2637         return NULL;
2638
2639 }
2640
2641 static int sha1_loose_object_info(const unsigned char *sha1,
2642                                   struct object_info *oi,
2643                                   int flags)
2644 {
2645         int status = 0;
2646         unsigned long mapsize;
2647         void *map;
2648         git_zstream stream;
2649         char hdr[32];
2650         struct strbuf hdrbuf = STRBUF_INIT;
2651
2652         if (oi->delta_base_sha1)
2653                 hashclr(oi->delta_base_sha1);
2654
2655         /*
2656          * If we don't care about type or size, then we don't
2657          * need to look inside the object at all. Note that we
2658          * do not optimize out the stat call, even if the
2659          * caller doesn't care about the disk-size, since our
2660          * return value implicitly indicates whether the
2661          * object even exists.
2662          */
2663         if (!oi->typep && !oi->typename && !oi->sizep) {
2664                 struct stat st;
2665                 if (stat_sha1_file(sha1, &st) < 0)
2666                         return -1;
2667                 if (oi->disk_sizep)
2668                         *oi->disk_sizep = st.st_size;
2669                 return 0;
2670         }
2671
2672         map = map_sha1_file(sha1, &mapsize);
2673         if (!map)
2674                 return -1;
2675         if (oi->disk_sizep)
2676                 *oi->disk_sizep = mapsize;
2677         if ((flags & LOOKUP_UNKNOWN_OBJECT)) {
2678                 if (unpack_sha1_header_to_strbuf(&stream, map, mapsize, hdr, sizeof(hdr), &hdrbuf) < 0)
2679                         status = error("unable to unpack %s header with --allow-unknown-type",
2680                                        sha1_to_hex(sha1));
2681         } else if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2682                 status = error("unable to unpack %s header",
2683                                sha1_to_hex(sha1));
2684         if (status < 0)
2685                 ; /* Do nothing */
2686         else if (hdrbuf.len) {
2687                 if ((status = parse_sha1_header_extended(hdrbuf.buf, oi, flags)) < 0)
2688                         status = error("unable to parse %s header with --allow-unknown-type",
2689                                        sha1_to_hex(sha1));
2690         } else if ((status = parse_sha1_header_extended(hdr, oi, flags)) < 0)
2691                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2692         git_inflate_end(&stream);
2693         munmap(map, mapsize);
2694         if (status && oi->typep)
2695                 *oi->typep = status;
2696         strbuf_release(&hdrbuf);
2697         return 0;
2698 }
2699
2700 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2701 {
2702         struct cached_object *co;
2703         struct pack_entry e;
2704         int rtype;
2705         enum object_type real_type;
2706         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2707
2708         co = find_cached_object(real);
2709         if (co) {
2710                 if (oi->typep)
2711                         *(oi->typep) = co->type;
2712                 if (oi->sizep)
2713                         *(oi->sizep) = co->size;
2714                 if (oi->disk_sizep)
2715                         *(oi->disk_sizep) = 0;
2716                 if (oi->delta_base_sha1)
2717                         hashclr(oi->delta_base_sha1);
2718                 if (oi->typename)
2719                         strbuf_addstr(oi->typename, typename(co->type));
2720                 oi->whence = OI_CACHED;
2721                 return 0;
2722         }
2723
2724         if (!find_pack_entry(real, &e)) {
2725                 /* Most likely it's a loose object. */
2726                 if (!sha1_loose_object_info(real, oi, flags)) {
2727                         oi->whence = OI_LOOSE;
2728                         return 0;
2729                 }
2730
2731                 /* Not a loose object; someone else may have just packed it. */
2732                 reprepare_packed_git();
2733                 if (!find_pack_entry(real, &e))
2734                         return -1;
2735         }
2736
2737         /*
2738          * packed_object_info() does not follow the delta chain to
2739          * find out the real type, unless it is given oi->typep.
2740          */
2741         if (oi->typename && !oi->typep)
2742                 oi->typep = &real_type;
2743
2744         rtype = packed_object_info(e.p, e.offset, oi);
2745         if (rtype < 0) {
2746                 mark_bad_packed_object(e.p, real);
2747                 if (oi->typep == &real_type)
2748                         oi->typep = NULL;
2749                 return sha1_object_info_extended(real, oi, 0);
2750         } else if (in_delta_base_cache(e.p, e.offset)) {
2751                 oi->whence = OI_DBCACHED;
2752         } else {
2753                 oi->whence = OI_PACKED;
2754                 oi->u.packed.offset = e.offset;
2755                 oi->u.packed.pack = e.p;
2756                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2757                                          rtype == OBJ_OFS_DELTA);
2758         }
2759         if (oi->typename)
2760                 strbuf_addstr(oi->typename, typename(*oi->typep));
2761         if (oi->typep == &real_type)
2762                 oi->typep = NULL;
2763
2764         return 0;
2765 }
2766
2767 /* returns enum object_type or negative */
2768 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2769 {
2770         enum object_type type;
2771         struct object_info oi = {NULL};
2772
2773         oi.typep = &type;
2774         oi.sizep = sizep;
2775         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2776                 return -1;
2777         return type;
2778 }
2779
2780 static void *read_packed_sha1(const unsigned char *sha1,
2781                               enum object_type *type, unsigned long *size)
2782 {
2783         struct pack_entry e;
2784         void *data;
2785
2786         if (!find_pack_entry(sha1, &e))
2787                 return NULL;
2788         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2789         if (!data) {
2790                 /*
2791                  * We're probably in deep shit, but let's try to fetch
2792                  * the required object anyway from another pack or loose.
2793                  * This should happen only in the presence of a corrupted
2794                  * pack, and is better than failing outright.
2795                  */
2796                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2797                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2798                 mark_bad_packed_object(e.p, sha1);
2799                 data = read_object(sha1, type, size);
2800         }
2801         return data;
2802 }
2803
2804 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2805                       unsigned char *sha1)
2806 {
2807         struct cached_object *co;
2808
2809         hash_sha1_file(buf, len, typename(type), sha1);
2810         if (has_sha1_file(sha1) || find_cached_object(sha1))
2811                 return 0;
2812         ALLOC_GROW(cached_objects, cached_object_nr + 1, cached_object_alloc);
2813         co = &cached_objects[cached_object_nr++];
2814         co->size = len;
2815         co->type = type;
2816         co->buf = xmalloc(len);
2817         memcpy(co->buf, buf, len);
2818         hashcpy(co->sha1, sha1);
2819         return 0;
2820 }
2821
2822 static void *read_object(const unsigned char *sha1, enum object_type *type,
2823                          unsigned long *size)
2824 {
2825         unsigned long mapsize;
2826         void *map, *buf;
2827         struct cached_object *co;
2828
2829         co = find_cached_object(sha1);
2830         if (co) {
2831                 *type = co->type;
2832                 *size = co->size;
2833                 return xmemdupz(co->buf, co->size);
2834         }
2835
2836         buf = read_packed_sha1(sha1, type, size);
2837         if (buf)
2838                 return buf;
2839         map = map_sha1_file(sha1, &mapsize);
2840         if (map) {
2841                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2842                 munmap(map, mapsize);
2843                 return buf;
2844         }
2845         reprepare_packed_git();
2846         return read_packed_sha1(sha1, type, size);
2847 }
2848
2849 /*
2850  * This function dies on corrupt objects; the callers who want to
2851  * deal with them should arrange to call read_object() and give error
2852  * messages themselves.
2853  */
2854 void *read_sha1_file_extended(const unsigned char *sha1,
2855                               enum object_type *type,
2856                               unsigned long *size,
2857                               unsigned flag)
2858 {
2859         void *data;
2860         const struct packed_git *p;
2861         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2862
2863         errno = 0;
2864         data = read_object(repl, type, size);
2865         if (data)
2866                 return data;
2867
2868         if (errno && errno != ENOENT)
2869                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2870
2871         /* die if we replaced an object with one that does not exist */
2872         if (repl != sha1)
2873                 die("replacement %s not found for %s",
2874                     sha1_to_hex(repl), sha1_to_hex(sha1));
2875
2876         if (has_loose_object(repl)) {
2877                 const char *path = sha1_file_name(sha1);
2878
2879                 die("loose object %s (stored in %s) is corrupt",
2880                     sha1_to_hex(repl), path);
2881         }
2882
2883         if ((p = has_packed_and_bad(repl)) != NULL)
2884                 die("packed object %s (stored in %s) is corrupt",
2885                     sha1_to_hex(repl), p->pack_name);
2886
2887         return NULL;
2888 }
2889
2890 void *read_object_with_reference(const unsigned char *sha1,
2891                                  const char *required_type_name,
2892                                  unsigned long *size,
2893                                  unsigned char *actual_sha1_return)
2894 {
2895         enum object_type type, required_type;
2896         void *buffer;
2897         unsigned long isize;
2898         unsigned char actual_sha1[20];
2899
2900         required_type = type_from_string(required_type_name);
2901         hashcpy(actual_sha1, sha1);
2902         while (1) {
2903                 int ref_length = -1;
2904                 const char *ref_type = NULL;
2905
2906                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2907                 if (!buffer)
2908                         return NULL;
2909                 if (type == required_type) {
2910                         *size = isize;
2911                         if (actual_sha1_return)
2912                                 hashcpy(actual_sha1_return, actual_sha1);
2913                         return buffer;
2914                 }
2915                 /* Handle references */
2916                 else if (type == OBJ_COMMIT)
2917                         ref_type = "tree ";
2918                 else if (type == OBJ_TAG)
2919                         ref_type = "object ";
2920                 else {
2921                         free(buffer);
2922                         return NULL;
2923                 }
2924                 ref_length = strlen(ref_type);
2925
2926                 if (ref_length + 40 > isize ||
2927                     memcmp(buffer, ref_type, ref_length) ||
2928                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2929                         free(buffer);
2930                         return NULL;
2931                 }
2932                 free(buffer);
2933                 /* Now we have the ID of the referred-to object in
2934                  * actual_sha1.  Check again. */
2935         }
2936 }
2937
2938 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2939                                     const char *type, unsigned char *sha1,
2940                                     char *hdr, int *hdrlen)
2941 {
2942         git_SHA_CTX c;
2943
2944         /* Generate the header */
2945         *hdrlen = xsnprintf(hdr, *hdrlen, "%s %lu", type, len)+1;
2946
2947         /* Sha1.. */
2948         git_SHA1_Init(&c);
2949         git_SHA1_Update(&c, hdr, *hdrlen);
2950         git_SHA1_Update(&c, buf, len);
2951         git_SHA1_Final(sha1, &c);
2952 }
2953
2954 /*
2955  * Move the just written object into its final resting place.
2956  */
2957 int finalize_object_file(const char *tmpfile, const char *filename)
2958 {
2959         int ret = 0;
2960
2961         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2962                 goto try_rename;
2963         else if (link(tmpfile, filename))
2964                 ret = errno;
2965
2966         /*
2967          * Coda hack - coda doesn't like cross-directory links,
2968          * so we fall back to a rename, which will mean that it
2969          * won't be able to check collisions, but that's not a
2970          * big deal.
2971          *
2972          * The same holds for FAT formatted media.
2973          *
2974          * When this succeeds, we just return.  We have nothing
2975          * left to unlink.
2976          */
2977         if (ret && ret != EEXIST) {
2978         try_rename:
2979                 if (!rename(tmpfile, filename))
2980                         goto out;
2981                 ret = errno;
2982         }
2983         unlink_or_warn(tmpfile);
2984         if (ret) {
2985                 if (ret != EEXIST) {
2986                         return error_errno("unable to write sha1 filename %s", filename);
2987                 }
2988                 /* FIXME!!! Collision check here ? */
2989         }
2990
2991 out:
2992         if (adjust_shared_perm(filename))
2993                 return error("unable to set permission to '%s'", filename);
2994         return 0;
2995 }
2996
2997 static int write_buffer(int fd, const void *buf, size_t len)
2998 {
2999         if (write_in_full(fd, buf, len) < 0)
3000                 return error_errno("file write error");
3001         return 0;
3002 }
3003
3004 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
3005                    unsigned char *sha1)
3006 {
3007         char hdr[32];
3008         int hdrlen = sizeof(hdr);
3009         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3010         return 0;
3011 }
3012
3013 /* Finalize a file on disk, and close it. */
3014 static void close_sha1_file(int fd)
3015 {
3016         if (fsync_object_files)
3017                 fsync_or_die(fd, "sha1 file");
3018         if (close(fd) != 0)
3019                 die_errno("error when closing sha1 file");
3020 }
3021
3022 /* Size of directory component, including the ending '/' */
3023 static inline int directory_size(const char *filename)
3024 {
3025         const char *s = strrchr(filename, '/');
3026         if (!s)
3027                 return 0;
3028         return s - filename + 1;
3029 }
3030
3031 /*
3032  * This creates a temporary file in the same directory as the final
3033  * 'filename'
3034  *
3035  * We want to avoid cross-directory filename renames, because those
3036  * can have problems on various filesystems (FAT, NFS, Coda).
3037  */
3038 static int create_tmpfile(struct strbuf *tmp, const char *filename)
3039 {
3040         int fd, dirlen = directory_size(filename);
3041
3042         strbuf_reset(tmp);
3043         strbuf_add(tmp, filename, dirlen);
3044         strbuf_addstr(tmp, "tmp_obj_XXXXXX");
3045         fd = git_mkstemp_mode(tmp->buf, 0444);
3046         if (fd < 0 && dirlen && errno == ENOENT) {
3047                 /*
3048                  * Make sure the directory exists; note that the contents
3049                  * of the buffer are undefined after mkstemp returns an
3050                  * error, so we have to rewrite the whole buffer from
3051                  * scratch.
3052                  */
3053                 strbuf_reset(tmp);
3054                 strbuf_add(tmp, filename, dirlen - 1);
3055                 if (mkdir(tmp->buf, 0777) && errno != EEXIST)
3056                         return -1;
3057                 if (adjust_shared_perm(tmp->buf))
3058                         return -1;
3059
3060                 /* Try again */
3061                 strbuf_addstr(tmp, "/tmp_obj_XXXXXX");
3062                 fd = git_mkstemp_mode(tmp->buf, 0444);
3063         }
3064         return fd;
3065 }
3066
3067 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
3068                               const void *buf, unsigned long len, time_t mtime)
3069 {
3070         int fd, ret;
3071         unsigned char compressed[4096];
3072         git_zstream stream;
3073         git_SHA_CTX c;
3074         unsigned char parano_sha1[20];
3075         static struct strbuf tmp_file = STRBUF_INIT;
3076         const char *filename = sha1_file_name(sha1);
3077
3078         fd = create_tmpfile(&tmp_file, filename);
3079         if (fd < 0) {
3080                 if (errno == EACCES)
3081                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
3082                 else
3083                         return error_errno("unable to create temporary file");
3084         }
3085
3086         /* Set it up */
3087         git_deflate_init(&stream, zlib_compression_level);
3088         stream.next_out = compressed;
3089         stream.avail_out = sizeof(compressed);
3090         git_SHA1_Init(&c);
3091
3092         /* First header.. */
3093         stream.next_in = (unsigned char *)hdr;
3094         stream.avail_in = hdrlen;
3095         while (git_deflate(&stream, 0) == Z_OK)
3096                 ; /* nothing */
3097         git_SHA1_Update(&c, hdr, hdrlen);
3098
3099         /* Then the data itself.. */
3100         stream.next_in = (void *)buf;
3101         stream.avail_in = len;
3102         do {
3103                 unsigned char *in0 = stream.next_in;
3104                 ret = git_deflate(&stream, Z_FINISH);
3105                 git_SHA1_Update(&c, in0, stream.next_in - in0);
3106                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
3107                         die("unable to write sha1 file");
3108                 stream.next_out = compressed;
3109                 stream.avail_out = sizeof(compressed);
3110         } while (ret == Z_OK);
3111
3112         if (ret != Z_STREAM_END)
3113                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
3114         ret = git_deflate_end_gently(&stream);
3115         if (ret != Z_OK)
3116                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
3117         git_SHA1_Final(parano_sha1, &c);
3118         if (hashcmp(sha1, parano_sha1) != 0)
3119                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
3120
3121         close_sha1_file(fd);
3122
3123         if (mtime) {
3124                 struct utimbuf utb;
3125                 utb.actime = mtime;
3126                 utb.modtime = mtime;
3127                 if (utime(tmp_file.buf, &utb) < 0)
3128                         warning_errno("failed utime() on %s", tmp_file.buf);
3129         }
3130
3131         return finalize_object_file(tmp_file.buf, filename);
3132 }
3133
3134 static int freshen_loose_object(const unsigned char *sha1)
3135 {
3136         return check_and_freshen(sha1, 1);
3137 }
3138
3139 static int freshen_packed_object(const unsigned char *sha1)
3140 {
3141         struct pack_entry e;
3142         if (!find_pack_entry(sha1, &e))
3143                 return 0;
3144         if (e.p->freshened)
3145                 return 1;
3146         if (!freshen_file(e.p->pack_name))
3147                 return 0;
3148         e.p->freshened = 1;
3149         return 1;
3150 }
3151
3152 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *sha1)
3153 {
3154         char hdr[32];
3155         int hdrlen = sizeof(hdr);
3156
3157         /* Normally if we have it in the pack then we do not bother writing
3158          * it out into .git/objects/??/?{38} file.
3159          */
3160         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
3161         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3162                 return 0;
3163         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
3164 }
3165
3166 int hash_sha1_file_literally(const void *buf, unsigned long len, const char *type,
3167                              unsigned char *sha1, unsigned flags)
3168 {
3169         char *header;
3170         int hdrlen, status = 0;
3171
3172         /* type string, SP, %lu of the length plus NUL must fit this */
3173         hdrlen = strlen(type) + 32;
3174         header = xmalloc(hdrlen);
3175         write_sha1_file_prepare(buf, len, type, sha1, header, &hdrlen);
3176
3177         if (!(flags & HASH_WRITE_OBJECT))
3178                 goto cleanup;
3179         if (freshen_packed_object(sha1) || freshen_loose_object(sha1))
3180                 goto cleanup;
3181         status = write_loose_object(sha1, header, hdrlen, buf, len, 0);
3182
3183 cleanup:
3184         free(header);
3185         return status;
3186 }
3187
3188 int force_object_loose(const unsigned char *sha1, time_t mtime)
3189 {
3190         void *buf;
3191         unsigned long len;
3192         enum object_type type;
3193         char hdr[32];
3194         int hdrlen;
3195         int ret;
3196
3197         if (has_loose_object(sha1))
3198                 return 0;
3199         buf = read_packed_sha1(sha1, &type, &len);
3200         if (!buf)
3201                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
3202         hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %lu", typename(type), len) + 1;
3203         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
3204         free(buf);
3205
3206         return ret;
3207 }
3208
3209 int has_pack_index(const unsigned char *sha1)
3210 {
3211         struct stat st;
3212         if (stat(sha1_pack_index_name(sha1), &st))
3213                 return 0;
3214         return 1;
3215 }
3216
3217 int has_sha1_pack(const unsigned char *sha1)
3218 {
3219         struct pack_entry e;
3220         return find_pack_entry(sha1, &e);
3221 }
3222
3223 int has_sha1_file_with_flags(const unsigned char *sha1, int flags)
3224 {
3225         struct pack_entry e;
3226
3227         if (find_pack_entry(sha1, &e))
3228                 return 1;
3229         if (has_loose_object(sha1))
3230                 return 1;
3231         if (flags & HAS_SHA1_QUICK)
3232                 return 0;
3233         reprepare_packed_git();
3234         return find_pack_entry(sha1, &e);
3235 }
3236
3237 int has_object_file(const struct object_id *oid)
3238 {
3239         return has_sha1_file(oid->hash);
3240 }
3241
3242 static void check_tree(const void *buf, size_t size)
3243 {
3244         struct tree_desc desc;
3245         struct name_entry entry;
3246
3247         init_tree_desc(&desc, buf, size);
3248         while (tree_entry(&desc, &entry))
3249                 /* do nothing
3250                  * tree_entry() will die() on malformed entries */
3251                 ;
3252 }
3253
3254 static void check_commit(const void *buf, size_t size)
3255 {
3256         struct commit c;
3257         memset(&c, 0, sizeof(c));
3258         if (parse_commit_buffer(&c, buf, size))
3259                 die("corrupt commit");
3260 }
3261
3262 static void check_tag(const void *buf, size_t size)
3263 {
3264         struct tag t;
3265         memset(&t, 0, sizeof(t));
3266         if (parse_tag_buffer(&t, buf, size))
3267                 die("corrupt tag");
3268 }
3269
3270 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3271                      enum object_type type,
3272                      const char *path, unsigned flags)
3273 {
3274         int ret, re_allocated = 0;
3275         int write_object = flags & HASH_WRITE_OBJECT;
3276
3277         if (!type)
3278                 type = OBJ_BLOB;
3279
3280         /*
3281          * Convert blobs to git internal format
3282          */
3283         if ((type == OBJ_BLOB) && path) {
3284                 struct strbuf nbuf = STRBUF_INIT;
3285                 if (convert_to_git(path, buf, size, &nbuf,
3286                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3287                         buf = strbuf_detach(&nbuf, &size);
3288                         re_allocated = 1;
3289                 }
3290         }
3291         if (flags & HASH_FORMAT_CHECK) {
3292                 if (type == OBJ_TREE)
3293                         check_tree(buf, size);
3294                 if (type == OBJ_COMMIT)
3295                         check_commit(buf, size);
3296                 if (type == OBJ_TAG)
3297                         check_tag(buf, size);
3298         }
3299
3300         if (write_object)
3301                 ret = write_sha1_file(buf, size, typename(type), sha1);
3302         else
3303                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3304         if (re_allocated)
3305                 free(buf);
3306         return ret;
3307 }
3308
3309 static int index_stream_convert_blob(unsigned char *sha1, int fd,
3310                                      const char *path, unsigned flags)
3311 {
3312         int ret;
3313         const int write_object = flags & HASH_WRITE_OBJECT;
3314         struct strbuf sbuf = STRBUF_INIT;
3315
3316         assert(path);
3317         assert(would_convert_to_git_filter_fd(path));
3318
3319         convert_to_git_filter_fd(path, fd, &sbuf,
3320                                  write_object ? safe_crlf : SAFE_CRLF_FALSE);
3321
3322         if (write_object)
3323                 ret = write_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3324                                       sha1);
3325         else
3326                 ret = hash_sha1_file(sbuf.buf, sbuf.len, typename(OBJ_BLOB),
3327                                      sha1);
3328         strbuf_release(&sbuf);
3329         return ret;
3330 }
3331
3332 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3333                       const char *path, unsigned flags)
3334 {
3335         struct strbuf sbuf = STRBUF_INIT;
3336         int ret;
3337
3338         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3339                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3340         else
3341                 ret = -1;
3342         strbuf_release(&sbuf);
3343         return ret;
3344 }
3345
3346 #define SMALL_FILE_SIZE (32*1024)
3347
3348 static int index_core(unsigned char *sha1, int fd, size_t size,
3349                       enum object_type type, const char *path,
3350                       unsigned flags)
3351 {
3352         int ret;
3353
3354         if (!size) {
3355                 ret = index_mem(sha1, "", size, type, path, flags);
3356         } else if (size <= SMALL_FILE_SIZE) {
3357                 char *buf = xmalloc(size);
3358                 if (size == read_in_full(fd, buf, size))
3359                         ret = index_mem(sha1, buf, size, type, path, flags);
3360                 else
3361                         ret = error_errno("short read");
3362                 free(buf);
3363         } else {
3364                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3365                 ret = index_mem(sha1, buf, size, type, path, flags);
3366                 munmap(buf, size);
3367         }
3368         return ret;
3369 }
3370
3371 /*
3372  * This creates one packfile per large blob unless bulk-checkin
3373  * machinery is "plugged".
3374  *
3375  * This also bypasses the usual "convert-to-git" dance, and that is on
3376  * purpose. We could write a streaming version of the converting
3377  * functions and insert that before feeding the data to fast-import
3378  * (or equivalent in-core API described above). However, that is
3379  * somewhat complicated, as we do not know the size of the filter
3380  * result, which we need to know beforehand when writing a git object.
3381  * Since the primary motivation for trying to stream from the working
3382  * tree file and to avoid mmaping it in core is to deal with large
3383  * binary blobs, they generally do not want to get any conversion, and
3384  * callers should avoid this code path when filters are requested.
3385  */
3386 static int index_stream(unsigned char *sha1, int fd, size_t size,
3387                         enum object_type type, const char *path,
3388                         unsigned flags)
3389 {
3390         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3391 }
3392
3393 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3394              enum object_type type, const char *path, unsigned flags)
3395 {
3396         int ret;
3397
3398         /*
3399          * Call xsize_t() only when needed to avoid potentially unnecessary
3400          * die() for large files.
3401          */
3402         if (type == OBJ_BLOB && path && would_convert_to_git_filter_fd(path))
3403                 ret = index_stream_convert_blob(sha1, fd, path, flags);
3404         else if (!S_ISREG(st->st_mode))
3405                 ret = index_pipe(sha1, fd, type, path, flags);
3406         else if (st->st_size <= big_file_threshold || type != OBJ_BLOB ||
3407                  (path && would_convert_to_git(path)))
3408                 ret = index_core(sha1, fd, xsize_t(st->st_size), type, path,
3409                                  flags);
3410         else
3411                 ret = index_stream(sha1, fd, xsize_t(st->st_size), type, path,
3412                                    flags);
3413         close(fd);
3414         return ret;
3415 }
3416
3417 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3418 {
3419         int fd;
3420         struct strbuf sb = STRBUF_INIT;
3421
3422         switch (st->st_mode & S_IFMT) {
3423         case S_IFREG:
3424                 fd = open(path, O_RDONLY);
3425                 if (fd < 0)
3426                         return error_errno("open(\"%s\")", path);
3427                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3428                         return error("%s: failed to insert into database",
3429                                      path);
3430                 break;
3431         case S_IFLNK:
3432                 if (strbuf_readlink(&sb, path, st->st_size))
3433                         return error_errno("readlink(\"%s\")", path);
3434                 if (!(flags & HASH_WRITE_OBJECT))
3435                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3436                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3437                         return error("%s: failed to insert into database",
3438                                      path);
3439                 strbuf_release(&sb);
3440                 break;
3441         case S_IFDIR:
3442                 return resolve_gitlink_ref(path, "HEAD", sha1);
3443         default:
3444                 return error("%s: unsupported file type", path);
3445         }
3446         return 0;
3447 }
3448
3449 int read_pack_header(int fd, struct pack_header *header)
3450 {
3451         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3452                 /* "eof before pack header was fully read" */
3453                 return PH_ERROR_EOF;
3454
3455         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3456                 /* "protocol error (pack signature mismatch detected)" */
3457                 return PH_ERROR_PACK_SIGNATURE;
3458         if (!pack_version_ok(header->hdr_version))
3459                 /* "protocol error (pack version unsupported)" */
3460                 return PH_ERROR_PROTOCOL;
3461         return 0;
3462 }
3463
3464 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3465 {
3466         enum object_type type = sha1_object_info(sha1, NULL);
3467         if (type < 0)
3468                 die("%s is not a valid object", sha1_to_hex(sha1));
3469         if (type != expect)
3470                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3471                     typename(expect));
3472 }
3473
3474 static int for_each_file_in_obj_subdir(int subdir_nr,
3475                                        struct strbuf *path,
3476                                        each_loose_object_fn obj_cb,
3477                                        each_loose_cruft_fn cruft_cb,
3478                                        each_loose_subdir_fn subdir_cb,
3479                                        void *data)
3480 {
3481         size_t baselen = path->len;
3482         DIR *dir = opendir(path->buf);
3483         struct dirent *de;
3484         int r = 0;
3485
3486         if (!dir) {
3487                 if (errno == ENOENT)
3488                         return 0;
3489                 return error_errno("unable to open %s", path->buf);
3490         }
3491
3492         while ((de = readdir(dir))) {
3493                 if (is_dot_or_dotdot(de->d_name))
3494                         continue;
3495
3496                 strbuf_setlen(path, baselen);
3497                 strbuf_addf(path, "/%s", de->d_name);
3498
3499                 if (strlen(de->d_name) == 38)  {
3500                         char hex[41];
3501                         unsigned char sha1[20];
3502
3503                         snprintf(hex, sizeof(hex), "%02x%s",
3504                                  subdir_nr, de->d_name);
3505                         if (!get_sha1_hex(hex, sha1)) {
3506                                 if (obj_cb) {
3507                                         r = obj_cb(sha1, path->buf, data);
3508                                         if (r)
3509                                                 break;
3510                                 }
3511                                 continue;
3512                         }
3513                 }
3514
3515                 if (cruft_cb) {
3516                         r = cruft_cb(de->d_name, path->buf, data);
3517                         if (r)
3518                                 break;
3519                 }
3520         }
3521         closedir(dir);
3522
3523         strbuf_setlen(path, baselen);
3524         if (!r && subdir_cb)
3525                 r = subdir_cb(subdir_nr, path->buf, data);
3526
3527         return r;
3528 }
3529
3530 int for_each_loose_file_in_objdir_buf(struct strbuf *path,
3531                             each_loose_object_fn obj_cb,
3532                             each_loose_cruft_fn cruft_cb,
3533                             each_loose_subdir_fn subdir_cb,
3534                             void *data)
3535 {
3536         size_t baselen = path->len;
3537         int r = 0;
3538         int i;
3539
3540         for (i = 0; i < 256; i++) {
3541                 strbuf_addf(path, "/%02x", i);
3542                 r = for_each_file_in_obj_subdir(i, path, obj_cb, cruft_cb,
3543                                                 subdir_cb, data);
3544                 strbuf_setlen(path, baselen);
3545                 if (r)
3546                         break;
3547         }
3548
3549         return r;
3550 }
3551
3552 int for_each_loose_file_in_objdir(const char *path,
3553                                   each_loose_object_fn obj_cb,
3554                                   each_loose_cruft_fn cruft_cb,
3555                                   each_loose_subdir_fn subdir_cb,
3556                                   void *data)
3557 {
3558         struct strbuf buf = STRBUF_INIT;
3559         int r;
3560
3561         strbuf_addstr(&buf, path);
3562         r = for_each_loose_file_in_objdir_buf(&buf, obj_cb, cruft_cb,
3563                                               subdir_cb, data);
3564         strbuf_release(&buf);
3565
3566         return r;
3567 }
3568
3569 struct loose_alt_odb_data {
3570         each_loose_object_fn *cb;
3571         void *data;
3572 };
3573
3574 static int loose_from_alt_odb(struct alternate_object_database *alt,
3575                               void *vdata)
3576 {
3577         struct loose_alt_odb_data *data = vdata;
3578         struct strbuf buf = STRBUF_INIT;
3579         int r;
3580
3581         /* copy base not including trailing '/' */
3582         strbuf_add(&buf, alt->base, alt->name - alt->base - 1);
3583         r = for_each_loose_file_in_objdir_buf(&buf,
3584                                               data->cb, NULL, NULL,
3585                                               data->data);
3586         strbuf_release(&buf);
3587         return r;
3588 }
3589
3590 int for_each_loose_object(each_loose_object_fn cb, void *data, unsigned flags)
3591 {
3592         struct loose_alt_odb_data alt;
3593         int r;
3594
3595         r = for_each_loose_file_in_objdir(get_object_directory(),
3596                                           cb, NULL, NULL, data);
3597         if (r)
3598                 return r;
3599
3600         if (flags & FOR_EACH_OBJECT_LOCAL_ONLY)
3601                 return 0;
3602
3603         alt.cb = cb;
3604         alt.data = data;
3605         return foreach_alt_odb(loose_from_alt_odb, &alt);
3606 }
3607
3608 static int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
3609 {
3610         uint32_t i;
3611         int r = 0;
3612
3613         for (i = 0; i < p->num_objects; i++) {
3614                 const unsigned char *sha1 = nth_packed_object_sha1(p, i);
3615
3616                 if (!sha1)
3617                         return error("unable to get sha1 of object %u in %s",
3618                                      i, p->pack_name);
3619
3620                 r = cb(sha1, p, i, data);
3621                 if (r)
3622                         break;
3623         }
3624         return r;
3625 }
3626
3627 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
3628 {
3629         struct packed_git *p;
3630         int r = 0;
3631         int pack_errors = 0;
3632
3633         prepare_packed_git();
3634         for (p = packed_git; p; p = p->next) {
3635                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
3636                         continue;
3637                 if (open_pack_index(p)) {
3638                         pack_errors = 1;
3639                         continue;
3640                 }
3641                 r = for_each_object_in_pack(p, cb, data);
3642                 if (r)
3643                         break;
3644         }
3645         return r ? r : pack_errors;
3646 }