Merge branch 'maint'
[git] / sha1_file.c
1 /*
2  * GIT - The information manager from hell
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This handles basic git sha1 object files - packing, unpacking,
7  * creation etc.
8  */
9 #include "cache.h"
10 #include "delta.h"
11 #include "pack.h"
12 #include "blob.h"
13 #include "commit.h"
14 #include "tag.h"
15 #include "tree.h"
16 #include "refs.h"
17 #include "pack-revindex.h"
18 #include "sha1-lookup.h"
19
20 #ifndef O_NOATIME
21 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
22 #define O_NOATIME 01000000
23 #else
24 #define O_NOATIME 0
25 #endif
26 #endif
27
28 #ifdef NO_C99_FORMAT
29 #define SZ_FMT "lu"
30 static unsigned long sz_fmt(size_t s) { return (unsigned long)s; }
31 #else
32 #define SZ_FMT "zu"
33 static size_t sz_fmt(size_t s) { return s; }
34 #endif
35
36 const unsigned char null_sha1[20];
37
38 int safe_create_leading_directories(char *path)
39 {
40         char *pos = path + offset_1st_component(path);
41         struct stat st;
42
43         while (pos) {
44                 pos = strchr(pos, '/');
45                 if (!pos)
46                         break;
47                 while (*++pos == '/')
48                         ;
49                 if (!*pos)
50                         break;
51                 *--pos = '\0';
52                 if (!stat(path, &st)) {
53                         /* path exists */
54                         if (!S_ISDIR(st.st_mode)) {
55                                 *pos = '/';
56                                 return -3;
57                         }
58                 }
59                 else if (mkdir(path, 0777)) {
60                         *pos = '/';
61                         return -1;
62                 }
63                 else if (adjust_shared_perm(path)) {
64                         *pos = '/';
65                         return -2;
66                 }
67                 *pos++ = '/';
68         }
69         return 0;
70 }
71
72 int safe_create_leading_directories_const(const char *path)
73 {
74         /* path points to cache entries, so xstrdup before messing with it */
75         char *buf = xstrdup(path);
76         int result = safe_create_leading_directories(buf);
77         free(buf);
78         return result;
79 }
80
81 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
82 {
83         int i;
84         for (i = 0; i < 20; i++) {
85                 static char hex[] = "0123456789abcdef";
86                 unsigned int val = sha1[i];
87                 char *pos = pathbuf + i*2 + (i > 0);
88                 *pos++ = hex[val >> 4];
89                 *pos = hex[val & 0xf];
90         }
91 }
92
93 /*
94  * NOTE! This returns a statically allocated buffer, so you have to be
95  * careful about using it. Do an "xstrdup()" if you need to save the
96  * filename.
97  *
98  * Also note that this returns the location for creating.  Reading
99  * SHA1 file can happen from any alternate directory listed in the
100  * DB_ENVIRONMENT environment variable if it is not found in
101  * the primary object database.
102  */
103 char *sha1_file_name(const unsigned char *sha1)
104 {
105         static char *name, *base;
106
107         if (!base) {
108                 const char *sha1_file_directory = get_object_directory();
109                 int len = strlen(sha1_file_directory);
110                 base = xmalloc(len + 60);
111                 memcpy(base, sha1_file_directory, len);
112                 memset(base+len, 0, 60);
113                 base[len] = '/';
114                 base[len+3] = '/';
115                 name = base + len + 1;
116         }
117         fill_sha1_path(name, sha1);
118         return base;
119 }
120
121 static char *sha1_get_pack_name(const unsigned char *sha1,
122                                 char **name, char **base, const char *which)
123 {
124         static const char hex[] = "0123456789abcdef";
125         char *buf;
126         int i;
127
128         if (!*base) {
129                 const char *sha1_file_directory = get_object_directory();
130                 int len = strlen(sha1_file_directory);
131                 *base = xmalloc(len + 60);
132                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
133                         sha1_file_directory, which);
134                 *name = *base + len + 11;
135         }
136
137         buf = *name;
138
139         for (i = 0; i < 20; i++) {
140                 unsigned int val = *sha1++;
141                 *buf++ = hex[val >> 4];
142                 *buf++ = hex[val & 0xf];
143         }
144
145         return *base;
146 }
147
148 char *sha1_pack_name(const unsigned char *sha1)
149 {
150         static char *name, *base;
151
152         return sha1_get_pack_name(sha1, &name, &base, "pack");
153 }
154
155 char *sha1_pack_index_name(const unsigned char *sha1)
156 {
157         static char *name, *base;
158
159         return sha1_get_pack_name(sha1, &name, &base, "idx");
160 }
161
162 struct alternate_object_database *alt_odb_list;
163 static struct alternate_object_database **alt_odb_tail;
164
165 static void read_info_alternates(const char * alternates, int depth);
166
167 /*
168  * Prepare alternate object database registry.
169  *
170  * The variable alt_odb_list points at the list of struct
171  * alternate_object_database.  The elements on this list come from
172  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
173  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
174  * whose contents is similar to that environment variable but can be
175  * LF separated.  Its base points at a statically allocated buffer that
176  * contains "/the/directory/corresponding/to/.git/objects/...", while
177  * its name points just after the slash at the end of ".git/objects/"
178  * in the example above, and has enough space to hold 40-byte hex
179  * SHA1, an extra slash for the first level indirection, and the
180  * terminating NUL.
181  */
182 static int link_alt_odb_entry(const char * entry, int len, const char * relative_base, int depth)
183 {
184         const char *objdir = get_object_directory();
185         struct alternate_object_database *ent;
186         struct alternate_object_database *alt;
187         /* 43 = 40-byte + 2 '/' + terminating NUL */
188         int pfxlen = len;
189         int entlen = pfxlen + 43;
190         int base_len = -1;
191
192         if (!is_absolute_path(entry) && relative_base) {
193                 /* Relative alt-odb */
194                 if (base_len < 0)
195                         base_len = strlen(relative_base) + 1;
196                 entlen += base_len;
197                 pfxlen += base_len;
198         }
199         ent = xmalloc(sizeof(*ent) + entlen);
200
201         if (!is_absolute_path(entry) && relative_base) {
202                 memcpy(ent->base, relative_base, base_len - 1);
203                 ent->base[base_len - 1] = '/';
204                 memcpy(ent->base + base_len, entry, len);
205         }
206         else
207                 memcpy(ent->base, entry, pfxlen);
208
209         ent->name = ent->base + pfxlen + 1;
210         ent->base[pfxlen + 3] = '/';
211         ent->base[pfxlen] = ent->base[entlen-1] = 0;
212
213         /* Detect cases where alternate disappeared */
214         if (!is_directory(ent->base)) {
215                 error("object directory %s does not exist; "
216                       "check .git/objects/info/alternates.",
217                       ent->base);
218                 free(ent);
219                 return -1;
220         }
221
222         /* Prevent the common mistake of listing the same
223          * thing twice, or object directory itself.
224          */
225         for (alt = alt_odb_list; alt; alt = alt->next) {
226                 if (!memcmp(ent->base, alt->base, pfxlen)) {
227                         free(ent);
228                         return -1;
229                 }
230         }
231         if (!memcmp(ent->base, objdir, pfxlen)) {
232                 free(ent);
233                 return -1;
234         }
235
236         /* add the alternate entry */
237         *alt_odb_tail = ent;
238         alt_odb_tail = &(ent->next);
239         ent->next = NULL;
240
241         /* recursively add alternates */
242         read_info_alternates(ent->base, depth + 1);
243
244         ent->base[pfxlen] = '/';
245
246         return 0;
247 }
248
249 static void link_alt_odb_entries(const char *alt, const char *ep, int sep,
250                                  const char *relative_base, int depth)
251 {
252         const char *cp, *last;
253
254         if (depth > 5) {
255                 error("%s: ignoring alternate object stores, nesting too deep.",
256                                 relative_base);
257                 return;
258         }
259
260         last = alt;
261         while (last < ep) {
262                 cp = last;
263                 if (cp < ep && *cp == '#') {
264                         while (cp < ep && *cp != sep)
265                                 cp++;
266                         last = cp + 1;
267                         continue;
268                 }
269                 while (cp < ep && *cp != sep)
270                         cp++;
271                 if (last != cp) {
272                         if (!is_absolute_path(last) && depth) {
273                                 error("%s: ignoring relative alternate object store %s",
274                                                 relative_base, last);
275                         } else {
276                                 link_alt_odb_entry(last, cp - last,
277                                                 relative_base, depth);
278                         }
279                 }
280                 while (cp < ep && *cp == sep)
281                         cp++;
282                 last = cp;
283         }
284 }
285
286 static void read_info_alternates(const char * relative_base, int depth)
287 {
288         char *map;
289         size_t mapsz;
290         struct stat st;
291         const char alt_file_name[] = "info/alternates";
292         /* Given that relative_base is no longer than PATH_MAX,
293            ensure that "path" has enough space to append "/", the
294            file name, "info/alternates", and a trailing NUL.  */
295         char path[PATH_MAX + 1 + sizeof alt_file_name];
296         int fd;
297
298         sprintf(path, "%s/%s", relative_base, alt_file_name);
299         fd = open(path, O_RDONLY);
300         if (fd < 0)
301                 return;
302         if (fstat(fd, &st) || (st.st_size == 0)) {
303                 close(fd);
304                 return;
305         }
306         mapsz = xsize_t(st.st_size);
307         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
308         close(fd);
309
310         link_alt_odb_entries(map, map + mapsz, '\n', relative_base, depth);
311
312         munmap(map, mapsz);
313 }
314
315 void add_to_alternates_file(const char *reference)
316 {
317         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
318         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
319         char *alt = mkpath("%s/objects\n", reference);
320         write_or_die(fd, alt, strlen(alt));
321         if (commit_lock_file(lock))
322                 die("could not close alternates file");
323         if (alt_odb_tail)
324                 link_alt_odb_entries(alt, alt + strlen(alt), '\n', NULL, 0);
325 }
326
327 void foreach_alt_odb(alt_odb_fn fn, void *cb)
328 {
329         struct alternate_object_database *ent;
330
331         prepare_alt_odb();
332         for (ent = alt_odb_list; ent; ent = ent->next)
333                 if (fn(ent, cb))
334                         return;
335 }
336
337 void prepare_alt_odb(void)
338 {
339         const char *alt;
340
341         if (alt_odb_tail)
342                 return;
343
344         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
345         if (!alt) alt = "";
346
347         alt_odb_tail = &alt_odb_list;
348         link_alt_odb_entries(alt, alt + strlen(alt), PATH_SEP, NULL, 0);
349
350         read_info_alternates(get_object_directory(), 0);
351 }
352
353 static int has_loose_object_local(const unsigned char *sha1)
354 {
355         char *name = sha1_file_name(sha1);
356         return !access(name, F_OK);
357 }
358
359 int has_loose_object_nonlocal(const unsigned char *sha1)
360 {
361         struct alternate_object_database *alt;
362         prepare_alt_odb();
363         for (alt = alt_odb_list; alt; alt = alt->next) {
364                 fill_sha1_path(alt->name, sha1);
365                 if (!access(alt->base, F_OK))
366                         return 1;
367         }
368         return 0;
369 }
370
371 static int has_loose_object(const unsigned char *sha1)
372 {
373         return has_loose_object_local(sha1) ||
374                has_loose_object_nonlocal(sha1);
375 }
376
377 static unsigned int pack_used_ctr;
378 static unsigned int pack_mmap_calls;
379 static unsigned int peak_pack_open_windows;
380 static unsigned int pack_open_windows;
381 static size_t peak_pack_mapped;
382 static size_t pack_mapped;
383 struct packed_git *packed_git;
384
385 void pack_report(void)
386 {
387         fprintf(stderr,
388                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
389                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
390                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
391                 sz_fmt(getpagesize()),
392                 sz_fmt(packed_git_window_size),
393                 sz_fmt(packed_git_limit));
394         fprintf(stderr,
395                 "pack_report: pack_used_ctr            = %10u\n"
396                 "pack_report: pack_mmap_calls          = %10u\n"
397                 "pack_report: pack_open_windows        = %10u / %10u\n"
398                 "pack_report: pack_mapped              = "
399                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
400                 pack_used_ctr,
401                 pack_mmap_calls,
402                 pack_open_windows, peak_pack_open_windows,
403                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
404 }
405
406 static int check_packed_git_idx(const char *path,  struct packed_git *p)
407 {
408         void *idx_map;
409         struct pack_idx_header *hdr;
410         size_t idx_size;
411         uint32_t version, nr, i, *index;
412         int fd = open(path, O_RDONLY);
413         struct stat st;
414
415         if (fd < 0)
416                 return -1;
417         if (fstat(fd, &st)) {
418                 close(fd);
419                 return -1;
420         }
421         idx_size = xsize_t(st.st_size);
422         if (idx_size < 4 * 256 + 20 + 20) {
423                 close(fd);
424                 return error("index file %s is too small", path);
425         }
426         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
427         close(fd);
428
429         hdr = idx_map;
430         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
431                 version = ntohl(hdr->idx_version);
432                 if (version < 2 || version > 2) {
433                         munmap(idx_map, idx_size);
434                         return error("index file %s is version %"PRIu32
435                                      " and is not supported by this binary"
436                                      " (try upgrading GIT to a newer version)",
437                                      path, version);
438                 }
439         } else
440                 version = 1;
441
442         nr = 0;
443         index = idx_map;
444         if (version > 1)
445                 index += 2;  /* skip index header */
446         for (i = 0; i < 256; i++) {
447                 uint32_t n = ntohl(index[i]);
448                 if (n < nr) {
449                         munmap(idx_map, idx_size);
450                         return error("non-monotonic index %s", path);
451                 }
452                 nr = n;
453         }
454
455         if (version == 1) {
456                 /*
457                  * Total size:
458                  *  - 256 index entries 4 bytes each
459                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
460                  *  - 20-byte SHA1 of the packfile
461                  *  - 20-byte SHA1 file checksum
462                  */
463                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
464                         munmap(idx_map, idx_size);
465                         return error("wrong index v1 file size in %s", path);
466                 }
467         } else if (version == 2) {
468                 /*
469                  * Minimum size:
470                  *  - 8 bytes of header
471                  *  - 256 index entries 4 bytes each
472                  *  - 20-byte sha1 entry * nr
473                  *  - 4-byte crc entry * nr
474                  *  - 4-byte offset entry * nr
475                  *  - 20-byte SHA1 of the packfile
476                  *  - 20-byte SHA1 file checksum
477                  * And after the 4-byte offset table might be a
478                  * variable sized table containing 8-byte entries
479                  * for offsets larger than 2^31.
480                  */
481                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
482                 unsigned long max_size = min_size;
483                 if (nr)
484                         max_size += (nr - 1)*8;
485                 if (idx_size < min_size || idx_size > max_size) {
486                         munmap(idx_map, idx_size);
487                         return error("wrong index v2 file size in %s", path);
488                 }
489                 if (idx_size != min_size &&
490                     /*
491                      * make sure we can deal with large pack offsets.
492                      * 31-bit signed offset won't be enough, neither
493                      * 32-bit unsigned one will be.
494                      */
495                     (sizeof(off_t) <= 4)) {
496                         munmap(idx_map, idx_size);
497                         return error("pack too large for current definition of off_t in %s", path);
498                 }
499         }
500
501         p->index_version = version;
502         p->index_data = idx_map;
503         p->index_size = idx_size;
504         p->num_objects = nr;
505         return 0;
506 }
507
508 int open_pack_index(struct packed_git *p)
509 {
510         char *idx_name;
511         int ret;
512
513         if (p->index_data)
514                 return 0;
515
516         idx_name = xstrdup(p->pack_name);
517         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
518         ret = check_packed_git_idx(idx_name, p);
519         free(idx_name);
520         return ret;
521 }
522
523 static void scan_windows(struct packed_git *p,
524         struct packed_git **lru_p,
525         struct pack_window **lru_w,
526         struct pack_window **lru_l)
527 {
528         struct pack_window *w, *w_l;
529
530         for (w_l = NULL, w = p->windows; w; w = w->next) {
531                 if (!w->inuse_cnt) {
532                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
533                                 *lru_p = p;
534                                 *lru_w = w;
535                                 *lru_l = w_l;
536                         }
537                 }
538                 w_l = w;
539         }
540 }
541
542 static int unuse_one_window(struct packed_git *current, int keep_fd)
543 {
544         struct packed_git *p, *lru_p = NULL;
545         struct pack_window *lru_w = NULL, *lru_l = NULL;
546
547         if (current)
548                 scan_windows(current, &lru_p, &lru_w, &lru_l);
549         for (p = packed_git; p; p = p->next)
550                 scan_windows(p, &lru_p, &lru_w, &lru_l);
551         if (lru_p) {
552                 munmap(lru_w->base, lru_w->len);
553                 pack_mapped -= lru_w->len;
554                 if (lru_l)
555                         lru_l->next = lru_w->next;
556                 else {
557                         lru_p->windows = lru_w->next;
558                         if (!lru_p->windows && lru_p->pack_fd != keep_fd) {
559                                 close(lru_p->pack_fd);
560                                 lru_p->pack_fd = -1;
561                         }
562                 }
563                 free(lru_w);
564                 pack_open_windows--;
565                 return 1;
566         }
567         return 0;
568 }
569
570 void release_pack_memory(size_t need, int fd)
571 {
572         size_t cur = pack_mapped;
573         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
574                 ; /* nothing */
575 }
576
577 void close_pack_windows(struct packed_git *p)
578 {
579         while (p->windows) {
580                 struct pack_window *w = p->windows;
581
582                 if (w->inuse_cnt)
583                         die("pack '%s' still has open windows to it",
584                             p->pack_name);
585                 munmap(w->base, w->len);
586                 pack_mapped -= w->len;
587                 pack_open_windows--;
588                 p->windows = w->next;
589                 free(w);
590         }
591 }
592
593 void unuse_pack(struct pack_window **w_cursor)
594 {
595         struct pack_window *w = *w_cursor;
596         if (w) {
597                 w->inuse_cnt--;
598                 *w_cursor = NULL;
599         }
600 }
601
602 /*
603  * This is used by git-repack in case a newly created pack happens to
604  * contain the same set of objects as an existing one.  In that case
605  * the resulting file might be different even if its name would be the
606  * same.  It is best to close any reference to the old pack before it is
607  * replaced on disk.  Of course no index pointers nor windows for given pack
608  * must subsist at this point.  If ever objects from this pack are requested
609  * again, the new version of the pack will be reinitialized through
610  * reprepare_packed_git().
611  */
612 void free_pack_by_name(const char *pack_name)
613 {
614         struct packed_git *p, **pp = &packed_git;
615
616         while (*pp) {
617                 p = *pp;
618                 if (strcmp(pack_name, p->pack_name) == 0) {
619                         clear_delta_base_cache();
620                         close_pack_windows(p);
621                         if (p->pack_fd != -1)
622                                 close(p->pack_fd);
623                         if (p->index_data)
624                                 munmap((void *)p->index_data, p->index_size);
625                         free(p->bad_object_sha1);
626                         *pp = p->next;
627                         free(p);
628                         return;
629                 }
630                 pp = &p->next;
631         }
632 }
633
634 /*
635  * Do not call this directly as this leaks p->pack_fd on error return;
636  * call open_packed_git() instead.
637  */
638 static int open_packed_git_1(struct packed_git *p)
639 {
640         struct stat st;
641         struct pack_header hdr;
642         unsigned char sha1[20];
643         unsigned char *idx_sha1;
644         long fd_flag;
645
646         if (!p->index_data && open_pack_index(p))
647                 return error("packfile %s index unavailable", p->pack_name);
648
649         p->pack_fd = open(p->pack_name, O_RDONLY);
650         while (p->pack_fd < 0 && errno == EMFILE && unuse_one_window(p, -1))
651                 p->pack_fd = open(p->pack_name, O_RDONLY);
652         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
653                 return -1;
654
655         /* If we created the struct before we had the pack we lack size. */
656         if (!p->pack_size) {
657                 if (!S_ISREG(st.st_mode))
658                         return error("packfile %s not a regular file", p->pack_name);
659                 p->pack_size = st.st_size;
660         } else if (p->pack_size != st.st_size)
661                 return error("packfile %s size changed", p->pack_name);
662
663         /* We leave these file descriptors open with sliding mmap;
664          * there is no point keeping them open across exec(), though.
665          */
666         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
667         if (fd_flag < 0)
668                 return error("cannot determine file descriptor flags");
669         fd_flag |= FD_CLOEXEC;
670         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
671                 return error("cannot set FD_CLOEXEC");
672
673         /* Verify we recognize this pack file format. */
674         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
675                 return error("file %s is far too short to be a packfile", p->pack_name);
676         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
677                 return error("file %s is not a GIT packfile", p->pack_name);
678         if (!pack_version_ok(hdr.hdr_version))
679                 return error("packfile %s is version %"PRIu32" and not"
680                         " supported (try upgrading GIT to a newer version)",
681                         p->pack_name, ntohl(hdr.hdr_version));
682
683         /* Verify the pack matches its index. */
684         if (p->num_objects != ntohl(hdr.hdr_entries))
685                 return error("packfile %s claims to have %"PRIu32" objects"
686                              " while index indicates %"PRIu32" objects",
687                              p->pack_name, ntohl(hdr.hdr_entries),
688                              p->num_objects);
689         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
690                 return error("end of packfile %s is unavailable", p->pack_name);
691         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
692                 return error("packfile %s signature is unavailable", p->pack_name);
693         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
694         if (hashcmp(sha1, idx_sha1))
695                 return error("packfile %s does not match index", p->pack_name);
696         return 0;
697 }
698
699 static int open_packed_git(struct packed_git *p)
700 {
701         if (!open_packed_git_1(p))
702                 return 0;
703         if (p->pack_fd != -1) {
704                 close(p->pack_fd);
705                 p->pack_fd = -1;
706         }
707         return -1;
708 }
709
710 static int in_window(struct pack_window *win, off_t offset)
711 {
712         /* We must promise at least 20 bytes (one hash) after the
713          * offset is available from this window, otherwise the offset
714          * is not actually in this window and a different window (which
715          * has that one hash excess) must be used.  This is to support
716          * the object header and delta base parsing routines below.
717          */
718         off_t win_off = win->offset;
719         return win_off <= offset
720                 && (offset + 20) <= (win_off + win->len);
721 }
722
723 unsigned char *use_pack(struct packed_git *p,
724                 struct pack_window **w_cursor,
725                 off_t offset,
726                 unsigned int *left)
727 {
728         struct pack_window *win = *w_cursor;
729
730         if (p->pack_fd == -1 && open_packed_git(p))
731                 die("packfile %s cannot be accessed", p->pack_name);
732
733         /* Since packfiles end in a hash of their content and it's
734          * pointless to ask for an offset into the middle of that
735          * hash, and the in_window function above wouldn't match
736          * don't allow an offset too close to the end of the file.
737          */
738         if (offset > (p->pack_size - 20))
739                 die("offset beyond end of packfile (truncated pack?)");
740
741         if (!win || !in_window(win, offset)) {
742                 if (win)
743                         win->inuse_cnt--;
744                 for (win = p->windows; win; win = win->next) {
745                         if (in_window(win, offset))
746                                 break;
747                 }
748                 if (!win) {
749                         size_t window_align = packed_git_window_size / 2;
750                         off_t len;
751                         win = xcalloc(1, sizeof(*win));
752                         win->offset = (offset / window_align) * window_align;
753                         len = p->pack_size - win->offset;
754                         if (len > packed_git_window_size)
755                                 len = packed_git_window_size;
756                         win->len = (size_t)len;
757                         pack_mapped += win->len;
758                         while (packed_git_limit < pack_mapped
759                                 && unuse_one_window(p, p->pack_fd))
760                                 ; /* nothing */
761                         win->base = xmmap(NULL, win->len,
762                                 PROT_READ, MAP_PRIVATE,
763                                 p->pack_fd, win->offset);
764                         if (win->base == MAP_FAILED)
765                                 die("packfile %s cannot be mapped: %s",
766                                         p->pack_name,
767                                         strerror(errno));
768                         pack_mmap_calls++;
769                         pack_open_windows++;
770                         if (pack_mapped > peak_pack_mapped)
771                                 peak_pack_mapped = pack_mapped;
772                         if (pack_open_windows > peak_pack_open_windows)
773                                 peak_pack_open_windows = pack_open_windows;
774                         win->next = p->windows;
775                         p->windows = win;
776                 }
777         }
778         if (win != *w_cursor) {
779                 win->last_used = pack_used_ctr++;
780                 win->inuse_cnt++;
781                 *w_cursor = win;
782         }
783         offset -= win->offset;
784         if (left)
785                 *left = win->len - xsize_t(offset);
786         return win->base + offset;
787 }
788
789 static struct packed_git *alloc_packed_git(int extra)
790 {
791         struct packed_git *p = xmalloc(sizeof(*p) + extra);
792         memset(p, 0, sizeof(*p));
793         p->pack_fd = -1;
794         return p;
795 }
796
797 struct packed_git *add_packed_git(const char *path, int path_len, int local)
798 {
799         struct stat st;
800         struct packed_git *p = alloc_packed_git(path_len + 2);
801
802         /*
803          * Make sure a corresponding .pack file exists and that
804          * the index looks sane.
805          */
806         path_len -= strlen(".idx");
807         if (path_len < 1) {
808                 free(p);
809                 return NULL;
810         }
811         memcpy(p->pack_name, path, path_len);
812
813         strcpy(p->pack_name + path_len, ".keep");
814         if (!access(p->pack_name, F_OK))
815                 p->pack_keep = 1;
816
817         strcpy(p->pack_name + path_len, ".pack");
818         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
819                 free(p);
820                 return NULL;
821         }
822
823         /* ok, it looks sane as far as we can check without
824          * actually mapping the pack file.
825          */
826         p->pack_size = st.st_size;
827         p->pack_local = local;
828         p->mtime = st.st_mtime;
829         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
830                 hashclr(p->sha1);
831         return p;
832 }
833
834 struct packed_git *parse_pack_index(unsigned char *sha1)
835 {
836         const char *idx_path = sha1_pack_index_name(sha1);
837         const char *path = sha1_pack_name(sha1);
838         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
839
840         strcpy(p->pack_name, path);
841         hashcpy(p->sha1, sha1);
842         if (check_packed_git_idx(idx_path, p)) {
843                 free(p);
844                 return NULL;
845         }
846
847         return p;
848 }
849
850 void install_packed_git(struct packed_git *pack)
851 {
852         pack->next = packed_git;
853         packed_git = pack;
854 }
855
856 static void prepare_packed_git_one(char *objdir, int local)
857 {
858         /* Ensure that this buffer is large enough so that we can
859            append "/pack/" without clobbering the stack even if
860            strlen(objdir) were PATH_MAX.  */
861         char path[PATH_MAX + 1 + 4 + 1 + 1];
862         int len;
863         DIR *dir;
864         struct dirent *de;
865
866         sprintf(path, "%s/pack", objdir);
867         len = strlen(path);
868         dir = opendir(path);
869         while (!dir && errno == EMFILE && unuse_one_window(packed_git, -1))
870                 dir = opendir(path);
871         if (!dir) {
872                 if (errno != ENOENT)
873                         error("unable to open object pack directory: %s: %s",
874                               path, strerror(errno));
875                 return;
876         }
877         path[len++] = '/';
878         while ((de = readdir(dir)) != NULL) {
879                 int namelen = strlen(de->d_name);
880                 struct packed_git *p;
881
882                 if (!has_extension(de->d_name, ".idx"))
883                         continue;
884
885                 if (len + namelen + 1 > sizeof(path))
886                         continue;
887
888                 /* Don't reopen a pack we already have. */
889                 strcpy(path + len, de->d_name);
890                 for (p = packed_git; p; p = p->next) {
891                         if (!memcmp(path, p->pack_name, len + namelen - 4))
892                                 break;
893                 }
894                 if (p)
895                         continue;
896                 /* See if it really is a valid .idx file with corresponding
897                  * .pack file that we can map.
898                  */
899                 p = add_packed_git(path, len + namelen, local);
900                 if (!p)
901                         continue;
902                 install_packed_git(p);
903         }
904         closedir(dir);
905 }
906
907 static int sort_pack(const void *a_, const void *b_)
908 {
909         struct packed_git *a = *((struct packed_git **)a_);
910         struct packed_git *b = *((struct packed_git **)b_);
911         int st;
912
913         /*
914          * Local packs tend to contain objects specific to our
915          * variant of the project than remote ones.  In addition,
916          * remote ones could be on a network mounted filesystem.
917          * Favor local ones for these reasons.
918          */
919         st = a->pack_local - b->pack_local;
920         if (st)
921                 return -st;
922
923         /*
924          * Younger packs tend to contain more recent objects,
925          * and more recent objects tend to get accessed more
926          * often.
927          */
928         if (a->mtime < b->mtime)
929                 return 1;
930         else if (a->mtime == b->mtime)
931                 return 0;
932         return -1;
933 }
934
935 static void rearrange_packed_git(void)
936 {
937         struct packed_git **ary, *p;
938         int i, n;
939
940         for (n = 0, p = packed_git; p; p = p->next)
941                 n++;
942         if (n < 2)
943                 return;
944
945         /* prepare an array of packed_git for easier sorting */
946         ary = xcalloc(n, sizeof(struct packed_git *));
947         for (n = 0, p = packed_git; p; p = p->next)
948                 ary[n++] = p;
949
950         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
951
952         /* link them back again */
953         for (i = 0; i < n - 1; i++)
954                 ary[i]->next = ary[i + 1];
955         ary[n - 1]->next = NULL;
956         packed_git = ary[0];
957
958         free(ary);
959 }
960
961 static int prepare_packed_git_run_once = 0;
962 void prepare_packed_git(void)
963 {
964         struct alternate_object_database *alt;
965
966         if (prepare_packed_git_run_once)
967                 return;
968         prepare_packed_git_one(get_object_directory(), 1);
969         prepare_alt_odb();
970         for (alt = alt_odb_list; alt; alt = alt->next) {
971                 alt->name[-1] = 0;
972                 prepare_packed_git_one(alt->base, 0);
973                 alt->name[-1] = '/';
974         }
975         rearrange_packed_git();
976         prepare_packed_git_run_once = 1;
977 }
978
979 void reprepare_packed_git(void)
980 {
981         discard_revindex();
982         prepare_packed_git_run_once = 0;
983         prepare_packed_git();
984 }
985
986 static void mark_bad_packed_object(struct packed_git *p,
987                                    const unsigned char *sha1)
988 {
989         unsigned i;
990         for (i = 0; i < p->num_bad_objects; i++)
991                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
992                         return;
993         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
994         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
995         p->num_bad_objects++;
996 }
997
998 static int has_packed_and_bad(const unsigned char *sha1)
999 {
1000         struct packed_git *p;
1001         unsigned i;
1002
1003         for (p = packed_git; p; p = p->next)
1004                 for (i = 0; i < p->num_bad_objects; i++)
1005                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1006                                 return 1;
1007         return 0;
1008 }
1009
1010 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1011 {
1012         unsigned char real_sha1[20];
1013         hash_sha1_file(map, size, type, real_sha1);
1014         return hashcmp(sha1, real_sha1) ? -1 : 0;
1015 }
1016
1017 static int git_open_noatime(const char *name)
1018 {
1019         static int sha1_file_open_flag = O_NOATIME;
1020         int fd = open(name, O_RDONLY | sha1_file_open_flag);
1021
1022         /* Might the failure be due to O_NOATIME? */
1023         if (fd < 0 && errno != ENOENT && sha1_file_open_flag) {
1024                 fd = open(name, O_RDONLY);
1025                 if (fd >= 0)
1026                         sha1_file_open_flag = 0;
1027         }
1028         return fd;
1029 }
1030
1031 static int open_sha1_file(const unsigned char *sha1)
1032 {
1033         int fd;
1034         char *name = sha1_file_name(sha1);
1035         struct alternate_object_database *alt;
1036
1037         fd = git_open_noatime(name);
1038         if (fd >= 0)
1039                 return fd;
1040
1041         prepare_alt_odb();
1042         errno = ENOENT;
1043         for (alt = alt_odb_list; alt; alt = alt->next) {
1044                 name = alt->name;
1045                 fill_sha1_path(name, sha1);
1046                 fd = git_open_noatime(alt->base);
1047                 if (fd >= 0)
1048                         return fd;
1049         }
1050         return -1;
1051 }
1052
1053 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1054 {
1055         void *map;
1056         int fd;
1057
1058         fd = open_sha1_file(sha1);
1059         map = NULL;
1060         if (fd >= 0) {
1061                 struct stat st;
1062
1063                 if (!fstat(fd, &st)) {
1064                         *size = xsize_t(st.st_size);
1065                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1066                 }
1067                 close(fd);
1068         }
1069         return map;
1070 }
1071
1072 static int legacy_loose_object(unsigned char *map)
1073 {
1074         unsigned int word;
1075
1076         /*
1077          * Is it a zlib-compressed buffer? If so, the first byte
1078          * must be 0x78 (15-bit window size, deflated), and the
1079          * first 16-bit word is evenly divisible by 31
1080          */
1081         word = (map[0] << 8) + map[1];
1082         if (map[0] == 0x78 && !(word % 31))
1083                 return 1;
1084         else
1085                 return 0;
1086 }
1087
1088 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1089                 unsigned long len, enum object_type *type, unsigned long *sizep)
1090 {
1091         unsigned shift;
1092         unsigned long size, c;
1093         unsigned long used = 0;
1094
1095         c = buf[used++];
1096         *type = (c >> 4) & 7;
1097         size = c & 15;
1098         shift = 4;
1099         while (c & 0x80) {
1100                 if (len <= used || bitsizeof(long) <= shift) {
1101                         error("bad object header");
1102                         return 0;
1103                 }
1104                 c = buf[used++];
1105                 size += (c & 0x7f) << shift;
1106                 shift += 7;
1107         }
1108         *sizep = size;
1109         return used;
1110 }
1111
1112 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1113 {
1114         unsigned long size, used;
1115         static const char valid_loose_object_type[8] = {
1116                 0, /* OBJ_EXT */
1117                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1118                 0, /* "delta" and others are invalid in a loose object */
1119         };
1120         enum object_type type;
1121
1122         /* Get the data stream */
1123         memset(stream, 0, sizeof(*stream));
1124         stream->next_in = map;
1125         stream->avail_in = mapsize;
1126         stream->next_out = buffer;
1127         stream->avail_out = bufsiz;
1128
1129         if (legacy_loose_object(map)) {
1130                 git_inflate_init(stream);
1131                 return git_inflate(stream, 0);
1132         }
1133
1134
1135         /*
1136          * There used to be a second loose object header format which
1137          * was meant to mimic the in-pack format, allowing for direct
1138          * copy of the object data.  This format turned up not to be
1139          * really worth it and we don't write it any longer.  But we
1140          * can still read it.
1141          */
1142         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1143         if (!used || !valid_loose_object_type[type])
1144                 return -1;
1145         map += used;
1146         mapsize -= used;
1147
1148         /* Set up the stream for the rest.. */
1149         stream->next_in = map;
1150         stream->avail_in = mapsize;
1151         git_inflate_init(stream);
1152
1153         /* And generate the fake traditional header */
1154         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1155                                          typename(type), size);
1156         return 0;
1157 }
1158
1159 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1160 {
1161         int bytes = strlen(buffer) + 1;
1162         unsigned char *buf = xmallocz(size);
1163         unsigned long n;
1164         int status = Z_OK;
1165
1166         n = stream->total_out - bytes;
1167         if (n > size)
1168                 n = size;
1169         memcpy(buf, (char *) buffer + bytes, n);
1170         bytes = n;
1171         if (bytes <= size) {
1172                 /*
1173                  * The above condition must be (bytes <= size), not
1174                  * (bytes < size).  In other words, even though we
1175                  * expect no more output and set avail_out to zer0,
1176                  * the input zlib stream may have bytes that express
1177                  * "this concludes the stream", and we *do* want to
1178                  * eat that input.
1179                  *
1180                  * Otherwise we would not be able to test that we
1181                  * consumed all the input to reach the expected size;
1182                  * we also want to check that zlib tells us that all
1183                  * went well with status == Z_STREAM_END at the end.
1184                  */
1185                 stream->next_out = buf + bytes;
1186                 stream->avail_out = size - bytes;
1187                 while (status == Z_OK)
1188                         status = git_inflate(stream, Z_FINISH);
1189         }
1190         if (status == Z_STREAM_END && !stream->avail_in) {
1191                 git_inflate_end(stream);
1192                 return buf;
1193         }
1194
1195         if (status < 0)
1196                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1197         else if (stream->avail_in)
1198                 error("garbage at end of loose object '%s'",
1199                       sha1_to_hex(sha1));
1200         free(buf);
1201         return NULL;
1202 }
1203
1204 /*
1205  * We used to just use "sscanf()", but that's actually way
1206  * too permissive for what we want to check. So do an anal
1207  * object header parse by hand.
1208  */
1209 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1210 {
1211         char type[10];
1212         int i;
1213         unsigned long size;
1214
1215         /*
1216          * The type can be at most ten bytes (including the
1217          * terminating '\0' that we add), and is followed by
1218          * a space.
1219          */
1220         i = 0;
1221         for (;;) {
1222                 char c = *hdr++;
1223                 if (c == ' ')
1224                         break;
1225                 type[i++] = c;
1226                 if (i >= sizeof(type))
1227                         return -1;
1228         }
1229         type[i] = 0;
1230
1231         /*
1232          * The length must follow immediately, and be in canonical
1233          * decimal format (ie "010" is not valid).
1234          */
1235         size = *hdr++ - '0';
1236         if (size > 9)
1237                 return -1;
1238         if (size) {
1239                 for (;;) {
1240                         unsigned long c = *hdr - '0';
1241                         if (c > 9)
1242                                 break;
1243                         hdr++;
1244                         size = size * 10 + c;
1245                 }
1246         }
1247         *sizep = size;
1248
1249         /*
1250          * The length must be followed by a zero byte
1251          */
1252         return *hdr ? -1 : type_from_string(type);
1253 }
1254
1255 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1256 {
1257         int ret;
1258         z_stream stream;
1259         char hdr[8192];
1260
1261         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1262         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1263                 return NULL;
1264
1265         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1266 }
1267
1268 unsigned long get_size_from_delta(struct packed_git *p,
1269                                   struct pack_window **w_curs,
1270                                   off_t curpos)
1271 {
1272         const unsigned char *data;
1273         unsigned char delta_head[20], *in;
1274         z_stream stream;
1275         int st;
1276
1277         memset(&stream, 0, sizeof(stream));
1278         stream.next_out = delta_head;
1279         stream.avail_out = sizeof(delta_head);
1280
1281         git_inflate_init(&stream);
1282         do {
1283                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1284                 stream.next_in = in;
1285                 st = git_inflate(&stream, Z_FINISH);
1286                 curpos += stream.next_in - in;
1287         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1288                  stream.total_out < sizeof(delta_head));
1289         git_inflate_end(&stream);
1290         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1291                 error("delta data unpack-initial failed");
1292                 return 0;
1293         }
1294
1295         /* Examine the initial part of the delta to figure out
1296          * the result size.
1297          */
1298         data = delta_head;
1299
1300         /* ignore base size */
1301         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1302
1303         /* Read the result size */
1304         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1305 }
1306
1307 static off_t get_delta_base(struct packed_git *p,
1308                                     struct pack_window **w_curs,
1309                                     off_t *curpos,
1310                                     enum object_type type,
1311                                     off_t delta_obj_offset)
1312 {
1313         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1314         off_t base_offset;
1315
1316         /* use_pack() assured us we have [base_info, base_info + 20)
1317          * as a range that we can look at without walking off the
1318          * end of the mapped window.  Its actually the hash size
1319          * that is assured.  An OFS_DELTA longer than the hash size
1320          * is stupid, as then a REF_DELTA would be smaller to store.
1321          */
1322         if (type == OBJ_OFS_DELTA) {
1323                 unsigned used = 0;
1324                 unsigned char c = base_info[used++];
1325                 base_offset = c & 127;
1326                 while (c & 128) {
1327                         base_offset += 1;
1328                         if (!base_offset || MSB(base_offset, 7))
1329                                 return 0;  /* overflow */
1330                         c = base_info[used++];
1331                         base_offset = (base_offset << 7) + (c & 127);
1332                 }
1333                 base_offset = delta_obj_offset - base_offset;
1334                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1335                         return 0;  /* out of bound */
1336                 *curpos += used;
1337         } else if (type == OBJ_REF_DELTA) {
1338                 /* The base entry _must_ be in the same pack */
1339                 base_offset = find_pack_entry_one(base_info, p);
1340                 *curpos += 20;
1341         } else
1342                 die("I am totally screwed");
1343         return base_offset;
1344 }
1345
1346 /* forward declaration for a mutually recursive function */
1347 static int packed_object_info(struct packed_git *p, off_t offset,
1348                               unsigned long *sizep);
1349
1350 static int packed_delta_info(struct packed_git *p,
1351                              struct pack_window **w_curs,
1352                              off_t curpos,
1353                              enum object_type type,
1354                              off_t obj_offset,
1355                              unsigned long *sizep)
1356 {
1357         off_t base_offset;
1358
1359         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1360         if (!base_offset)
1361                 return OBJ_BAD;
1362         type = packed_object_info(p, base_offset, NULL);
1363         if (type <= OBJ_NONE) {
1364                 struct revindex_entry *revidx;
1365                 const unsigned char *base_sha1;
1366                 revidx = find_pack_revindex(p, base_offset);
1367                 if (!revidx)
1368                         return OBJ_BAD;
1369                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1370                 mark_bad_packed_object(p, base_sha1);
1371                 type = sha1_object_info(base_sha1, NULL);
1372                 if (type <= OBJ_NONE)
1373                         return OBJ_BAD;
1374         }
1375
1376         /* We choose to only get the type of the base object and
1377          * ignore potentially corrupt pack file that expects the delta
1378          * based on a base with a wrong size.  This saves tons of
1379          * inflate() calls.
1380          */
1381         if (sizep) {
1382                 *sizep = get_size_from_delta(p, w_curs, curpos);
1383                 if (*sizep == 0)
1384                         type = OBJ_BAD;
1385         }
1386
1387         return type;
1388 }
1389
1390 static int unpack_object_header(struct packed_git *p,
1391                                 struct pack_window **w_curs,
1392                                 off_t *curpos,
1393                                 unsigned long *sizep)
1394 {
1395         unsigned char *base;
1396         unsigned int left;
1397         unsigned long used;
1398         enum object_type type;
1399
1400         /* use_pack() assures us we have [base, base + 20) available
1401          * as a range that we can look at at.  (Its actually the hash
1402          * size that is assured.)  With our object header encoding
1403          * the maximum deflated object size is 2^137, which is just
1404          * insane, so we know won't exceed what we have been given.
1405          */
1406         base = use_pack(p, w_curs, *curpos, &left);
1407         used = unpack_object_header_buffer(base, left, &type, sizep);
1408         if (!used) {
1409                 type = OBJ_BAD;
1410         } else
1411                 *curpos += used;
1412
1413         return type;
1414 }
1415
1416 const char *packed_object_info_detail(struct packed_git *p,
1417                                       off_t obj_offset,
1418                                       unsigned long *size,
1419                                       unsigned long *store_size,
1420                                       unsigned int *delta_chain_length,
1421                                       unsigned char *base_sha1)
1422 {
1423         struct pack_window *w_curs = NULL;
1424         off_t curpos;
1425         unsigned long dummy;
1426         unsigned char *next_sha1;
1427         enum object_type type;
1428         struct revindex_entry *revidx;
1429
1430         *delta_chain_length = 0;
1431         curpos = obj_offset;
1432         type = unpack_object_header(p, &w_curs, &curpos, size);
1433
1434         revidx = find_pack_revindex(p, obj_offset);
1435         *store_size = revidx[1].offset - obj_offset;
1436
1437         for (;;) {
1438                 switch (type) {
1439                 default:
1440                         die("pack %s contains unknown object type %d",
1441                             p->pack_name, type);
1442                 case OBJ_COMMIT:
1443                 case OBJ_TREE:
1444                 case OBJ_BLOB:
1445                 case OBJ_TAG:
1446                         unuse_pack(&w_curs);
1447                         return typename(type);
1448                 case OBJ_OFS_DELTA:
1449                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1450                         if (!obj_offset)
1451                                 die("pack %s contains bad delta base reference of type %s",
1452                                     p->pack_name, typename(type));
1453                         if (*delta_chain_length == 0) {
1454                                 revidx = find_pack_revindex(p, obj_offset);
1455                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1456                         }
1457                         break;
1458                 case OBJ_REF_DELTA:
1459                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1460                         if (*delta_chain_length == 0)
1461                                 hashcpy(base_sha1, next_sha1);
1462                         obj_offset = find_pack_entry_one(next_sha1, p);
1463                         break;
1464                 }
1465                 (*delta_chain_length)++;
1466                 curpos = obj_offset;
1467                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1468         }
1469 }
1470
1471 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1472                               unsigned long *sizep)
1473 {
1474         struct pack_window *w_curs = NULL;
1475         unsigned long size;
1476         off_t curpos = obj_offset;
1477         enum object_type type;
1478
1479         type = unpack_object_header(p, &w_curs, &curpos, &size);
1480
1481         switch (type) {
1482         case OBJ_OFS_DELTA:
1483         case OBJ_REF_DELTA:
1484                 type = packed_delta_info(p, &w_curs, curpos,
1485                                          type, obj_offset, sizep);
1486                 break;
1487         case OBJ_COMMIT:
1488         case OBJ_TREE:
1489         case OBJ_BLOB:
1490         case OBJ_TAG:
1491                 if (sizep)
1492                         *sizep = size;
1493                 break;
1494         default:
1495                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1496                       type, (uintmax_t)obj_offset, p->pack_name);
1497                 type = OBJ_BAD;
1498         }
1499         unuse_pack(&w_curs);
1500         return type;
1501 }
1502
1503 static void *unpack_compressed_entry(struct packed_git *p,
1504                                     struct pack_window **w_curs,
1505                                     off_t curpos,
1506                                     unsigned long size)
1507 {
1508         int st;
1509         z_stream stream;
1510         unsigned char *buffer, *in;
1511
1512         buffer = xmallocz(size);
1513         memset(&stream, 0, sizeof(stream));
1514         stream.next_out = buffer;
1515         stream.avail_out = size + 1;
1516
1517         git_inflate_init(&stream);
1518         do {
1519                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1520                 stream.next_in = in;
1521                 st = git_inflate(&stream, Z_FINISH);
1522                 if (!stream.avail_out)
1523                         break; /* the payload is larger than it should be */
1524                 curpos += stream.next_in - in;
1525         } while (st == Z_OK || st == Z_BUF_ERROR);
1526         git_inflate_end(&stream);
1527         if ((st != Z_STREAM_END) || stream.total_out != size) {
1528                 free(buffer);
1529                 return NULL;
1530         }
1531
1532         return buffer;
1533 }
1534
1535 #define MAX_DELTA_CACHE (256)
1536
1537 static size_t delta_base_cached;
1538
1539 static struct delta_base_cache_lru_list {
1540         struct delta_base_cache_lru_list *prev;
1541         struct delta_base_cache_lru_list *next;
1542 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1543
1544 static struct delta_base_cache_entry {
1545         struct delta_base_cache_lru_list lru;
1546         void *data;
1547         struct packed_git *p;
1548         off_t base_offset;
1549         unsigned long size;
1550         enum object_type type;
1551 } delta_base_cache[MAX_DELTA_CACHE];
1552
1553 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1554 {
1555         unsigned long hash;
1556
1557         hash = (unsigned long)p + (unsigned long)base_offset;
1558         hash += (hash >> 8) + (hash >> 16);
1559         return hash % MAX_DELTA_CACHE;
1560 }
1561
1562 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1563         unsigned long *base_size, enum object_type *type, int keep_cache)
1564 {
1565         void *ret;
1566         unsigned long hash = pack_entry_hash(p, base_offset);
1567         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1568
1569         ret = ent->data;
1570         if (!ret || ent->p != p || ent->base_offset != base_offset)
1571                 return unpack_entry(p, base_offset, type, base_size);
1572
1573         if (!keep_cache) {
1574                 ent->data = NULL;
1575                 ent->lru.next->prev = ent->lru.prev;
1576                 ent->lru.prev->next = ent->lru.next;
1577                 delta_base_cached -= ent->size;
1578         } else {
1579                 ret = xmemdupz(ent->data, ent->size);
1580         }
1581         *type = ent->type;
1582         *base_size = ent->size;
1583         return ret;
1584 }
1585
1586 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1587 {
1588         if (ent->data) {
1589                 free(ent->data);
1590                 ent->data = NULL;
1591                 ent->lru.next->prev = ent->lru.prev;
1592                 ent->lru.prev->next = ent->lru.next;
1593                 delta_base_cached -= ent->size;
1594         }
1595 }
1596
1597 void clear_delta_base_cache(void)
1598 {
1599         unsigned long p;
1600         for (p = 0; p < MAX_DELTA_CACHE; p++)
1601                 release_delta_base_cache(&delta_base_cache[p]);
1602 }
1603
1604 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1605         void *base, unsigned long base_size, enum object_type type)
1606 {
1607         unsigned long hash = pack_entry_hash(p, base_offset);
1608         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1609         struct delta_base_cache_lru_list *lru;
1610
1611         release_delta_base_cache(ent);
1612         delta_base_cached += base_size;
1613
1614         for (lru = delta_base_cache_lru.next;
1615              delta_base_cached > delta_base_cache_limit
1616              && lru != &delta_base_cache_lru;
1617              lru = lru->next) {
1618                 struct delta_base_cache_entry *f = (void *)lru;
1619                 if (f->type == OBJ_BLOB)
1620                         release_delta_base_cache(f);
1621         }
1622         for (lru = delta_base_cache_lru.next;
1623              delta_base_cached > delta_base_cache_limit
1624              && lru != &delta_base_cache_lru;
1625              lru = lru->next) {
1626                 struct delta_base_cache_entry *f = (void *)lru;
1627                 release_delta_base_cache(f);
1628         }
1629
1630         ent->p = p;
1631         ent->base_offset = base_offset;
1632         ent->type = type;
1633         ent->data = base;
1634         ent->size = base_size;
1635         ent->lru.next = &delta_base_cache_lru;
1636         ent->lru.prev = delta_base_cache_lru.prev;
1637         delta_base_cache_lru.prev->next = &ent->lru;
1638         delta_base_cache_lru.prev = &ent->lru;
1639 }
1640
1641 static void *read_object(const unsigned char *sha1, enum object_type *type,
1642                          unsigned long *size);
1643
1644 static void *unpack_delta_entry(struct packed_git *p,
1645                                 struct pack_window **w_curs,
1646                                 off_t curpos,
1647                                 unsigned long delta_size,
1648                                 off_t obj_offset,
1649                                 enum object_type *type,
1650                                 unsigned long *sizep)
1651 {
1652         void *delta_data, *result, *base;
1653         unsigned long base_size;
1654         off_t base_offset;
1655
1656         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1657         if (!base_offset) {
1658                 error("failed to validate delta base reference "
1659                       "at offset %"PRIuMAX" from %s",
1660                       (uintmax_t)curpos, p->pack_name);
1661                 return NULL;
1662         }
1663         unuse_pack(w_curs);
1664         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1665         if (!base) {
1666                 /*
1667                  * We're probably in deep shit, but let's try to fetch
1668                  * the required base anyway from another pack or loose.
1669                  * This is costly but should happen only in the presence
1670                  * of a corrupted pack, and is better than failing outright.
1671                  */
1672                 struct revindex_entry *revidx;
1673                 const unsigned char *base_sha1;
1674                 revidx = find_pack_revindex(p, base_offset);
1675                 if (!revidx)
1676                         return NULL;
1677                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1678                 error("failed to read delta base object %s"
1679                       " at offset %"PRIuMAX" from %s",
1680                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1681                       p->pack_name);
1682                 mark_bad_packed_object(p, base_sha1);
1683                 base = read_object(base_sha1, type, &base_size);
1684                 if (!base)
1685                         return NULL;
1686         }
1687
1688         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1689         if (!delta_data) {
1690                 error("failed to unpack compressed delta "
1691                       "at offset %"PRIuMAX" from %s",
1692                       (uintmax_t)curpos, p->pack_name);
1693                 free(base);
1694                 return NULL;
1695         }
1696         result = patch_delta(base, base_size,
1697                              delta_data, delta_size,
1698                              sizep);
1699         if (!result)
1700                 die("failed to apply delta");
1701         free(delta_data);
1702         add_delta_base_cache(p, base_offset, base, base_size, *type);
1703         return result;
1704 }
1705
1706 int do_check_packed_object_crc;
1707
1708 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1709                    enum object_type *type, unsigned long *sizep)
1710 {
1711         struct pack_window *w_curs = NULL;
1712         off_t curpos = obj_offset;
1713         void *data;
1714
1715         if (do_check_packed_object_crc && p->index_version > 1) {
1716                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1717                 unsigned long len = revidx[1].offset - obj_offset;
1718                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1719                         const unsigned char *sha1 =
1720                                 nth_packed_object_sha1(p, revidx->nr);
1721                         error("bad packed object CRC for %s",
1722                               sha1_to_hex(sha1));
1723                         mark_bad_packed_object(p, sha1);
1724                         unuse_pack(&w_curs);
1725                         return NULL;
1726                 }
1727         }
1728
1729         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1730         switch (*type) {
1731         case OBJ_OFS_DELTA:
1732         case OBJ_REF_DELTA:
1733                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1734                                           obj_offset, type, sizep);
1735                 break;
1736         case OBJ_COMMIT:
1737         case OBJ_TREE:
1738         case OBJ_BLOB:
1739         case OBJ_TAG:
1740                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1741                 break;
1742         default:
1743                 data = NULL;
1744                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1745                       *type, (uintmax_t)obj_offset, p->pack_name);
1746         }
1747         unuse_pack(&w_curs);
1748         return data;
1749 }
1750
1751 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1752                                             uint32_t n)
1753 {
1754         const unsigned char *index = p->index_data;
1755         if (!index) {
1756                 if (open_pack_index(p))
1757                         return NULL;
1758                 index = p->index_data;
1759         }
1760         if (n >= p->num_objects)
1761                 return NULL;
1762         index += 4 * 256;
1763         if (p->index_version == 1) {
1764                 return index + 24 * n + 4;
1765         } else {
1766                 index += 8;
1767                 return index + 20 * n;
1768         }
1769 }
1770
1771 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1772 {
1773         const unsigned char *index = p->index_data;
1774         index += 4 * 256;
1775         if (p->index_version == 1) {
1776                 return ntohl(*((uint32_t *)(index + 24 * n)));
1777         } else {
1778                 uint32_t off;
1779                 index += 8 + p->num_objects * (20 + 4);
1780                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1781                 if (!(off & 0x80000000))
1782                         return off;
1783                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1784                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1785                                    ntohl(*((uint32_t *)(index + 4)));
1786         }
1787 }
1788
1789 off_t find_pack_entry_one(const unsigned char *sha1,
1790                                   struct packed_git *p)
1791 {
1792         const uint32_t *level1_ofs = p->index_data;
1793         const unsigned char *index = p->index_data;
1794         unsigned hi, lo, stride;
1795         static int use_lookup = -1;
1796         static int debug_lookup = -1;
1797
1798         if (debug_lookup < 0)
1799                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1800
1801         if (!index) {
1802                 if (open_pack_index(p))
1803                         return 0;
1804                 level1_ofs = p->index_data;
1805                 index = p->index_data;
1806         }
1807         if (p->index_version > 1) {
1808                 level1_ofs += 2;
1809                 index += 8;
1810         }
1811         index += 4 * 256;
1812         hi = ntohl(level1_ofs[*sha1]);
1813         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1814         if (p->index_version > 1) {
1815                 stride = 20;
1816         } else {
1817                 stride = 24;
1818                 index += 4;
1819         }
1820
1821         if (debug_lookup)
1822                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1823                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1824
1825         if (use_lookup < 0)
1826                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1827         if (use_lookup) {
1828                 int pos = sha1_entry_pos(index, stride, 0,
1829                                          lo, hi, p->num_objects, sha1);
1830                 if (pos < 0)
1831                         return 0;
1832                 return nth_packed_object_offset(p, pos);
1833         }
1834
1835         do {
1836                 unsigned mi = (lo + hi) / 2;
1837                 int cmp = hashcmp(index + mi * stride, sha1);
1838
1839                 if (debug_lookup)
1840                         printf("lo %u hi %u rg %u mi %u\n",
1841                                lo, hi, hi - lo, mi);
1842                 if (!cmp)
1843                         return nth_packed_object_offset(p, mi);
1844                 if (cmp > 0)
1845                         hi = mi;
1846                 else
1847                         lo = mi+1;
1848         } while (lo < hi);
1849         return 0;
1850 }
1851
1852 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1853 {
1854         static struct packed_git *last_found = (void *)1;
1855         struct packed_git *p;
1856         off_t offset;
1857
1858         prepare_packed_git();
1859         if (!packed_git)
1860                 return 0;
1861         p = (last_found == (void *)1) ? packed_git : last_found;
1862
1863         do {
1864                 if (p->num_bad_objects) {
1865                         unsigned i;
1866                         for (i = 0; i < p->num_bad_objects; i++)
1867                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1868                                         goto next;
1869                 }
1870
1871                 offset = find_pack_entry_one(sha1, p);
1872                 if (offset) {
1873                         /*
1874                          * We are about to tell the caller where they can
1875                          * locate the requested object.  We better make
1876                          * sure the packfile is still here and can be
1877                          * accessed before supplying that answer, as
1878                          * it may have been deleted since the index
1879                          * was loaded!
1880                          */
1881                         if (p->pack_fd == -1 && open_packed_git(p)) {
1882                                 error("packfile %s cannot be accessed", p->pack_name);
1883                                 goto next;
1884                         }
1885                         e->offset = offset;
1886                         e->p = p;
1887                         hashcpy(e->sha1, sha1);
1888                         last_found = p;
1889                         return 1;
1890                 }
1891
1892                 next:
1893                 if (p == last_found)
1894                         p = packed_git;
1895                 else
1896                         p = p->next;
1897                 if (p == last_found)
1898                         p = p->next;
1899         } while (p);
1900         return 0;
1901 }
1902
1903 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1904                                   struct packed_git *packs)
1905 {
1906         struct packed_git *p;
1907
1908         for (p = packs; p; p = p->next) {
1909                 if (find_pack_entry_one(sha1, p))
1910                         return p;
1911         }
1912         return NULL;
1913
1914 }
1915
1916 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1917 {
1918         int status;
1919         unsigned long mapsize, size;
1920         void *map;
1921         z_stream stream;
1922         char hdr[32];
1923
1924         map = map_sha1_file(sha1, &mapsize);
1925         if (!map)
1926                 return error("unable to find %s", sha1_to_hex(sha1));
1927         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1928                 status = error("unable to unpack %s header",
1929                                sha1_to_hex(sha1));
1930         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1931                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1932         else if (sizep)
1933                 *sizep = size;
1934         git_inflate_end(&stream);
1935         munmap(map, mapsize);
1936         return status;
1937 }
1938
1939 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1940 {
1941         struct pack_entry e;
1942         int status;
1943
1944         if (!find_pack_entry(sha1, &e)) {
1945                 /* Most likely it's a loose object. */
1946                 status = sha1_loose_object_info(sha1, sizep);
1947                 if (status >= 0)
1948                         return status;
1949
1950                 /* Not a loose object; someone else may have just packed it. */
1951                 reprepare_packed_git();
1952                 if (!find_pack_entry(sha1, &e))
1953                         return status;
1954         }
1955
1956         status = packed_object_info(e.p, e.offset, sizep);
1957         if (status < 0) {
1958                 mark_bad_packed_object(e.p, sha1);
1959                 status = sha1_object_info(sha1, sizep);
1960         }
1961
1962         return status;
1963 }
1964
1965 static void *read_packed_sha1(const unsigned char *sha1,
1966                               enum object_type *type, unsigned long *size)
1967 {
1968         struct pack_entry e;
1969         void *data;
1970
1971         if (!find_pack_entry(sha1, &e))
1972                 return NULL;
1973         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
1974         if (!data) {
1975                 /*
1976                  * We're probably in deep shit, but let's try to fetch
1977                  * the required object anyway from another pack or loose.
1978                  * This should happen only in the presence of a corrupted
1979                  * pack, and is better than failing outright.
1980                  */
1981                 error("failed to read object %s at offset %"PRIuMAX" from %s",
1982                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
1983                 mark_bad_packed_object(e.p, sha1);
1984                 data = read_object(sha1, type, size);
1985         }
1986         return data;
1987 }
1988
1989 /*
1990  * This is meant to hold a *small* number of objects that you would
1991  * want read_sha1_file() to be able to return, but yet you do not want
1992  * to write them into the object store (e.g. a browse-only
1993  * application).
1994  */
1995 static struct cached_object {
1996         unsigned char sha1[20];
1997         enum object_type type;
1998         void *buf;
1999         unsigned long size;
2000 } *cached_objects;
2001 static int cached_object_nr, cached_object_alloc;
2002
2003 static struct cached_object empty_tree = {
2004         EMPTY_TREE_SHA1_BIN,
2005         OBJ_TREE,
2006         "",
2007         0
2008 };
2009
2010 static struct cached_object *find_cached_object(const unsigned char *sha1)
2011 {
2012         int i;
2013         struct cached_object *co = cached_objects;
2014
2015         for (i = 0; i < cached_object_nr; i++, co++) {
2016                 if (!hashcmp(co->sha1, sha1))
2017                         return co;
2018         }
2019         if (!hashcmp(sha1, empty_tree.sha1))
2020                 return &empty_tree;
2021         return NULL;
2022 }
2023
2024 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2025                       unsigned char *sha1)
2026 {
2027         struct cached_object *co;
2028
2029         hash_sha1_file(buf, len, typename(type), sha1);
2030         if (has_sha1_file(sha1) || find_cached_object(sha1))
2031                 return 0;
2032         if (cached_object_alloc <= cached_object_nr) {
2033                 cached_object_alloc = alloc_nr(cached_object_alloc);
2034                 cached_objects = xrealloc(cached_objects,
2035                                           sizeof(*cached_objects) *
2036                                           cached_object_alloc);
2037         }
2038         co = &cached_objects[cached_object_nr++];
2039         co->size = len;
2040         co->type = type;
2041         co->buf = xmalloc(len);
2042         memcpy(co->buf, buf, len);
2043         hashcpy(co->sha1, sha1);
2044         return 0;
2045 }
2046
2047 static void *read_object(const unsigned char *sha1, enum object_type *type,
2048                          unsigned long *size)
2049 {
2050         unsigned long mapsize;
2051         void *map, *buf;
2052         struct cached_object *co;
2053
2054         co = find_cached_object(sha1);
2055         if (co) {
2056                 *type = co->type;
2057                 *size = co->size;
2058                 return xmemdupz(co->buf, co->size);
2059         }
2060
2061         buf = read_packed_sha1(sha1, type, size);
2062         if (buf)
2063                 return buf;
2064         map = map_sha1_file(sha1, &mapsize);
2065         if (map) {
2066                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2067                 munmap(map, mapsize);
2068                 return buf;
2069         }
2070         reprepare_packed_git();
2071         return read_packed_sha1(sha1, type, size);
2072 }
2073
2074 void *read_sha1_file_repl(const unsigned char *sha1,
2075                           enum object_type *type,
2076                           unsigned long *size,
2077                           const unsigned char **replacement)
2078 {
2079         const unsigned char *repl = lookup_replace_object(sha1);
2080         void *data = read_object(repl, type, size);
2081
2082         /* die if we replaced an object with one that does not exist */
2083         if (!data && repl != sha1)
2084                 die("replacement %s not found for %s",
2085                     sha1_to_hex(repl), sha1_to_hex(sha1));
2086
2087         /* legacy behavior is to die on corrupted objects */
2088         if (!data && (has_loose_object(repl) || has_packed_and_bad(repl)))
2089                 die("object %s is corrupted", sha1_to_hex(repl));
2090
2091         if (replacement)
2092                 *replacement = repl;
2093
2094         return data;
2095 }
2096
2097 void *read_object_with_reference(const unsigned char *sha1,
2098                                  const char *required_type_name,
2099                                  unsigned long *size,
2100                                  unsigned char *actual_sha1_return)
2101 {
2102         enum object_type type, required_type;
2103         void *buffer;
2104         unsigned long isize;
2105         unsigned char actual_sha1[20];
2106
2107         required_type = type_from_string(required_type_name);
2108         hashcpy(actual_sha1, sha1);
2109         while (1) {
2110                 int ref_length = -1;
2111                 const char *ref_type = NULL;
2112
2113                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2114                 if (!buffer)
2115                         return NULL;
2116                 if (type == required_type) {
2117                         *size = isize;
2118                         if (actual_sha1_return)
2119                                 hashcpy(actual_sha1_return, actual_sha1);
2120                         return buffer;
2121                 }
2122                 /* Handle references */
2123                 else if (type == OBJ_COMMIT)
2124                         ref_type = "tree ";
2125                 else if (type == OBJ_TAG)
2126                         ref_type = "object ";
2127                 else {
2128                         free(buffer);
2129                         return NULL;
2130                 }
2131                 ref_length = strlen(ref_type);
2132
2133                 if (ref_length + 40 > isize ||
2134                     memcmp(buffer, ref_type, ref_length) ||
2135                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2136                         free(buffer);
2137                         return NULL;
2138                 }
2139                 free(buffer);
2140                 /* Now we have the ID of the referred-to object in
2141                  * actual_sha1.  Check again. */
2142         }
2143 }
2144
2145 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2146                                     const char *type, unsigned char *sha1,
2147                                     char *hdr, int *hdrlen)
2148 {
2149         git_SHA_CTX c;
2150
2151         /* Generate the header */
2152         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2153
2154         /* Sha1.. */
2155         git_SHA1_Init(&c);
2156         git_SHA1_Update(&c, hdr, *hdrlen);
2157         git_SHA1_Update(&c, buf, len);
2158         git_SHA1_Final(sha1, &c);
2159 }
2160
2161 /*
2162  * Move the just written object into its final resting place.
2163  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2164  * "moving" is only a part of what it does, when no patch between
2165  * master to pu changes the call sites of this function.
2166  */
2167 int move_temp_to_file(const char *tmpfile, const char *filename)
2168 {
2169         int ret = 0;
2170
2171         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2172                 goto try_rename;
2173         else if (link(tmpfile, filename))
2174                 ret = errno;
2175
2176         /*
2177          * Coda hack - coda doesn't like cross-directory links,
2178          * so we fall back to a rename, which will mean that it
2179          * won't be able to check collisions, but that's not a
2180          * big deal.
2181          *
2182          * The same holds for FAT formatted media.
2183          *
2184          * When this succeeds, we just return.  We have nothing
2185          * left to unlink.
2186          */
2187         if (ret && ret != EEXIST) {
2188         try_rename:
2189                 if (!rename(tmpfile, filename))
2190                         goto out;
2191                 ret = errno;
2192         }
2193         unlink_or_warn(tmpfile);
2194         if (ret) {
2195                 if (ret != EEXIST) {
2196                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2197                 }
2198                 /* FIXME!!! Collision check here ? */
2199         }
2200
2201 out:
2202         if (adjust_shared_perm(filename))
2203                 return error("unable to set permission to '%s'", filename);
2204         return 0;
2205 }
2206
2207 static int write_buffer(int fd, const void *buf, size_t len)
2208 {
2209         if (write_in_full(fd, buf, len) < 0)
2210                 return error("file write error (%s)", strerror(errno));
2211         return 0;
2212 }
2213
2214 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2215                    unsigned char *sha1)
2216 {
2217         char hdr[32];
2218         int hdrlen;
2219         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2220         return 0;
2221 }
2222
2223 /* Finalize a file on disk, and close it. */
2224 static void close_sha1_file(int fd)
2225 {
2226         if (fsync_object_files)
2227                 fsync_or_die(fd, "sha1 file");
2228         if (close(fd) != 0)
2229                 die_errno("error when closing sha1 file");
2230 }
2231
2232 /* Size of directory component, including the ending '/' */
2233 static inline int directory_size(const char *filename)
2234 {
2235         const char *s = strrchr(filename, '/');
2236         if (!s)
2237                 return 0;
2238         return s - filename + 1;
2239 }
2240
2241 /*
2242  * This creates a temporary file in the same directory as the final
2243  * 'filename'
2244  *
2245  * We want to avoid cross-directory filename renames, because those
2246  * can have problems on various filesystems (FAT, NFS, Coda).
2247  */
2248 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2249 {
2250         int fd, dirlen = directory_size(filename);
2251
2252         if (dirlen + 20 > bufsiz) {
2253                 errno = ENAMETOOLONG;
2254                 return -1;
2255         }
2256         memcpy(buffer, filename, dirlen);
2257         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2258         fd = git_mkstemp_mode(buffer, 0444);
2259         if (fd < 0 && dirlen && errno == ENOENT) {
2260                 /* Make sure the directory exists */
2261                 memcpy(buffer, filename, dirlen);
2262                 buffer[dirlen-1] = 0;
2263                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2264                         return -1;
2265
2266                 /* Try again */
2267                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2268                 fd = git_mkstemp_mode(buffer, 0444);
2269         }
2270         return fd;
2271 }
2272
2273 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2274                               const void *buf, unsigned long len, time_t mtime)
2275 {
2276         int fd, ret;
2277         unsigned char compressed[4096];
2278         z_stream stream;
2279         git_SHA_CTX c;
2280         unsigned char parano_sha1[20];
2281         char *filename;
2282         static char tmpfile[PATH_MAX];
2283
2284         filename = sha1_file_name(sha1);
2285         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2286         while (fd < 0 && errno == EMFILE && unuse_one_window(packed_git, -1))
2287                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2288         if (fd < 0) {
2289                 if (errno == EACCES)
2290                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2291                 else
2292                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2293         }
2294
2295         /* Set it up */
2296         memset(&stream, 0, sizeof(stream));
2297         deflateInit(&stream, zlib_compression_level);
2298         stream.next_out = compressed;
2299         stream.avail_out = sizeof(compressed);
2300         git_SHA1_Init(&c);
2301
2302         /* First header.. */
2303         stream.next_in = (unsigned char *)hdr;
2304         stream.avail_in = hdrlen;
2305         while (deflate(&stream, 0) == Z_OK)
2306                 /* nothing */;
2307         git_SHA1_Update(&c, hdr, hdrlen);
2308
2309         /* Then the data itself.. */
2310         stream.next_in = (void *)buf;
2311         stream.avail_in = len;
2312         do {
2313                 unsigned char *in0 = stream.next_in;
2314                 ret = deflate(&stream, Z_FINISH);
2315                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2316                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2317                         die("unable to write sha1 file");
2318                 stream.next_out = compressed;
2319                 stream.avail_out = sizeof(compressed);
2320         } while (ret == Z_OK);
2321
2322         if (ret != Z_STREAM_END)
2323                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2324         ret = deflateEnd(&stream);
2325         if (ret != Z_OK)
2326                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2327         git_SHA1_Final(parano_sha1, &c);
2328         if (hashcmp(sha1, parano_sha1) != 0)
2329                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2330
2331         close_sha1_file(fd);
2332
2333         if (mtime) {
2334                 struct utimbuf utb;
2335                 utb.actime = mtime;
2336                 utb.modtime = mtime;
2337                 if (utime(tmpfile, &utb) < 0)
2338                         warning("failed utime() on %s: %s",
2339                                 tmpfile, strerror(errno));
2340         }
2341
2342         return move_temp_to_file(tmpfile, filename);
2343 }
2344
2345 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2346 {
2347         unsigned char sha1[20];
2348         char hdr[32];
2349         int hdrlen;
2350
2351         /* Normally if we have it in the pack then we do not bother writing
2352          * it out into .git/objects/??/?{38} file.
2353          */
2354         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2355         if (returnsha1)
2356                 hashcpy(returnsha1, sha1);
2357         if (has_sha1_file(sha1))
2358                 return 0;
2359         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2360 }
2361
2362 int force_object_loose(const unsigned char *sha1, time_t mtime)
2363 {
2364         void *buf;
2365         unsigned long len;
2366         enum object_type type;
2367         char hdr[32];
2368         int hdrlen;
2369         int ret;
2370
2371         if (has_loose_object(sha1))
2372                 return 0;
2373         buf = read_packed_sha1(sha1, &type, &len);
2374         if (!buf)
2375                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2376         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2377         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2378         free(buf);
2379
2380         return ret;
2381 }
2382
2383 int has_pack_index(const unsigned char *sha1)
2384 {
2385         struct stat st;
2386         if (stat(sha1_pack_index_name(sha1), &st))
2387                 return 0;
2388         return 1;
2389 }
2390
2391 int has_sha1_pack(const unsigned char *sha1)
2392 {
2393         struct pack_entry e;
2394         return find_pack_entry(sha1, &e);
2395 }
2396
2397 int has_sha1_file(const unsigned char *sha1)
2398 {
2399         struct pack_entry e;
2400
2401         if (find_pack_entry(sha1, &e))
2402                 return 1;
2403         return has_loose_object(sha1);
2404 }
2405
2406 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2407                      int write_object, enum object_type type, const char *path)
2408 {
2409         int ret, re_allocated = 0;
2410
2411         if (!type)
2412                 type = OBJ_BLOB;
2413
2414         /*
2415          * Convert blobs to git internal format
2416          */
2417         if ((type == OBJ_BLOB) && path) {
2418                 struct strbuf nbuf = STRBUF_INIT;
2419                 if (convert_to_git(path, buf, size, &nbuf,
2420                                    write_object ? safe_crlf : 0)) {
2421                         buf = strbuf_detach(&nbuf, &size);
2422                         re_allocated = 1;
2423                 }
2424         }
2425
2426         if (write_object)
2427                 ret = write_sha1_file(buf, size, typename(type), sha1);
2428         else
2429                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2430         if (re_allocated)
2431                 free(buf);
2432         return ret;
2433 }
2434
2435 #define SMALL_FILE_SIZE (32*1024)
2436
2437 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2438              enum object_type type, const char *path)
2439 {
2440         int ret;
2441         size_t size = xsize_t(st->st_size);
2442
2443         if (!S_ISREG(st->st_mode)) {
2444                 struct strbuf sbuf = STRBUF_INIT;
2445                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2446                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2447                                         type, path);
2448                 else
2449                         ret = -1;
2450                 strbuf_release(&sbuf);
2451         } else if (size <= SMALL_FILE_SIZE) {
2452                 char *buf = xmalloc(size);
2453                 if (size == read_in_full(fd, buf, size))
2454                         ret = index_mem(sha1, buf, size, write_object, type,
2455                                         path);
2456                 else
2457                         ret = error("short read %s", strerror(errno));
2458                 free(buf);
2459         } else if (size) {
2460                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2461                 ret = index_mem(sha1, buf, size, write_object, type, path);
2462                 munmap(buf, size);
2463         } else
2464                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2465         close(fd);
2466         return ret;
2467 }
2468
2469 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2470 {
2471         int fd;
2472         struct strbuf sb = STRBUF_INIT;
2473
2474         switch (st->st_mode & S_IFMT) {
2475         case S_IFREG:
2476                 fd = open(path, O_RDONLY);
2477                 if (fd < 0)
2478                         return error("open(\"%s\"): %s", path,
2479                                      strerror(errno));
2480                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2481                         return error("%s: failed to insert into database",
2482                                      path);
2483                 break;
2484         case S_IFLNK:
2485                 if (strbuf_readlink(&sb, path, st->st_size)) {
2486                         char *errstr = strerror(errno);
2487                         return error("readlink(\"%s\"): %s", path,
2488                                      errstr);
2489                 }
2490                 if (!write_object)
2491                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2492                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2493                         return error("%s: failed to insert into database",
2494                                      path);
2495                 strbuf_release(&sb);
2496                 break;
2497         case S_IFDIR:
2498                 return resolve_gitlink_ref(path, "HEAD", sha1);
2499         default:
2500                 return error("%s: unsupported file type", path);
2501         }
2502         return 0;
2503 }
2504
2505 int read_pack_header(int fd, struct pack_header *header)
2506 {
2507         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2508                 /* "eof before pack header was fully read" */
2509                 return PH_ERROR_EOF;
2510
2511         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2512                 /* "protocol error (pack signature mismatch detected)" */
2513                 return PH_ERROR_PACK_SIGNATURE;
2514         if (!pack_version_ok(header->hdr_version))
2515                 /* "protocol error (pack version unsupported)" */
2516                 return PH_ERROR_PROTOCOL;
2517         return 0;
2518 }