Merge branch 'jk/maint-push-tracking-wo-remote' into 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 static inline int offset_1st_component(const char *path)
39 {
40         if (has_dos_drive_prefix(path))
41                 return 2 + (path[2] == '/');
42         return *path == '/';
43 }
44
45 int safe_create_leading_directories(char *path)
46 {
47         char *pos = path + offset_1st_component(path);
48         struct stat st;
49
50         while (pos) {
51                 pos = strchr(pos, '/');
52                 if (!pos)
53                         break;
54                 while (*++pos == '/')
55                         ;
56                 if (!*pos)
57                         break;
58                 *--pos = '\0';
59                 if (!stat(path, &st)) {
60                         /* path exists */
61                         if (!S_ISDIR(st.st_mode)) {
62                                 *pos = '/';
63                                 return -3;
64                         }
65                 }
66                 else if (mkdir(path, 0777)) {
67                         *pos = '/';
68                         return -1;
69                 }
70                 else if (adjust_shared_perm(path)) {
71                         *pos = '/';
72                         return -2;
73                 }
74                 *pos++ = '/';
75         }
76         return 0;
77 }
78
79 int safe_create_leading_directories_const(const char *path)
80 {
81         /* path points to cache entries, so xstrdup before messing with it */
82         char *buf = xstrdup(path);
83         int result = safe_create_leading_directories(buf);
84         free(buf);
85         return result;
86 }
87
88 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
89 {
90         int i;
91         for (i = 0; i < 20; i++) {
92                 static char hex[] = "0123456789abcdef";
93                 unsigned int val = sha1[i];
94                 char *pos = pathbuf + i*2 + (i > 0);
95                 *pos++ = hex[val >> 4];
96                 *pos = hex[val & 0xf];
97         }
98 }
99
100 /*
101  * NOTE! This returns a statically allocated buffer, so you have to be
102  * careful about using it. Do an "xstrdup()" if you need to save the
103  * filename.
104  *
105  * Also note that this returns the location for creating.  Reading
106  * SHA1 file can happen from any alternate directory listed in the
107  * DB_ENVIRONMENT environment variable if it is not found in
108  * the primary object database.
109  */
110 char *sha1_file_name(const unsigned char *sha1)
111 {
112         static char *name, *base;
113
114         if (!base) {
115                 const char *sha1_file_directory = get_object_directory();
116                 int len = strlen(sha1_file_directory);
117                 base = xmalloc(len + 60);
118                 memcpy(base, sha1_file_directory, len);
119                 memset(base+len, 0, 60);
120                 base[len] = '/';
121                 base[len+3] = '/';
122                 name = base + len + 1;
123         }
124         fill_sha1_path(name, sha1);
125         return base;
126 }
127
128 static char *sha1_get_pack_name(const unsigned char *sha1,
129                                 char **name, char **base, const char *which)
130 {
131         static const char hex[] = "0123456789abcdef";
132         char *buf;
133         int i;
134
135         if (!*base) {
136                 const char *sha1_file_directory = get_object_directory();
137                 int len = strlen(sha1_file_directory);
138                 *base = xmalloc(len + 60);
139                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
140                         sha1_file_directory, which);
141                 *name = *base + len + 11;
142         }
143
144         buf = *name;
145
146         for (i = 0; i < 20; i++) {
147                 unsigned int val = *sha1++;
148                 *buf++ = hex[val >> 4];
149                 *buf++ = hex[val & 0xf];
150         }
151
152         return *base;
153 }
154
155 char *sha1_pack_name(const unsigned char *sha1)
156 {
157         static char *name, *base;
158
159         return sha1_get_pack_name(sha1, &name, &base, "pack");
160 }
161
162 char *sha1_pack_index_name(const unsigned char *sha1)
163 {
164         static char *name, *base;
165
166         return sha1_get_pack_name(sha1, &name, &base, "idx");
167 }
168
169 struct alternate_object_database *alt_odb_list;
170 static struct alternate_object_database **alt_odb_tail;
171
172 static void read_info_alternates(const char * alternates, int depth);
173
174 /*
175  * Prepare alternate object database registry.
176  *
177  * The variable alt_odb_list points at the list of struct
178  * alternate_object_database.  The elements on this list come from
179  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
180  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
181  * whose contents is similar to that environment variable but can be
182  * LF separated.  Its base points at a statically allocated buffer that
183  * contains "/the/directory/corresponding/to/.git/objects/...", while
184  * its name points just after the slash at the end of ".git/objects/"
185  * in the example above, and has enough space to hold 40-byte hex
186  * SHA1, an extra slash for the first level indirection, and the
187  * terminating NUL.
188  */
189 static int link_alt_odb_entry(const char * entry, int len, const char * relative_base, int depth)
190 {
191         const char *objdir = get_object_directory();
192         struct alternate_object_database *ent;
193         struct alternate_object_database *alt;
194         /* 43 = 40-byte + 2 '/' + terminating NUL */
195         int pfxlen = len;
196         int entlen = pfxlen + 43;
197         int base_len = -1;
198
199         if (!is_absolute_path(entry) && relative_base) {
200                 /* Relative alt-odb */
201                 if (base_len < 0)
202                         base_len = strlen(relative_base) + 1;
203                 entlen += base_len;
204                 pfxlen += base_len;
205         }
206         ent = xmalloc(sizeof(*ent) + entlen);
207
208         if (!is_absolute_path(entry) && relative_base) {
209                 memcpy(ent->base, relative_base, base_len - 1);
210                 ent->base[base_len - 1] = '/';
211                 memcpy(ent->base + base_len, entry, len);
212         }
213         else
214                 memcpy(ent->base, entry, pfxlen);
215
216         ent->name = ent->base + pfxlen + 1;
217         ent->base[pfxlen + 3] = '/';
218         ent->base[pfxlen] = ent->base[entlen-1] = 0;
219
220         /* Detect cases where alternate disappeared */
221         if (!is_directory(ent->base)) {
222                 error("object directory %s does not exist; "
223                       "check .git/objects/info/alternates.",
224                       ent->base);
225                 free(ent);
226                 return -1;
227         }
228
229         /* Prevent the common mistake of listing the same
230          * thing twice, or object directory itself.
231          */
232         for (alt = alt_odb_list; alt; alt = alt->next) {
233                 if (!memcmp(ent->base, alt->base, pfxlen)) {
234                         free(ent);
235                         return -1;
236                 }
237         }
238         if (!memcmp(ent->base, objdir, pfxlen)) {
239                 free(ent);
240                 return -1;
241         }
242
243         /* add the alternate entry */
244         *alt_odb_tail = ent;
245         alt_odb_tail = &(ent->next);
246         ent->next = NULL;
247
248         /* recursively add alternates */
249         read_info_alternates(ent->base, depth + 1);
250
251         ent->base[pfxlen] = '/';
252
253         return 0;
254 }
255
256 static void link_alt_odb_entries(const char *alt, const char *ep, int sep,
257                                  const char *relative_base, int depth)
258 {
259         const char *cp, *last;
260
261         if (depth > 5) {
262                 error("%s: ignoring alternate object stores, nesting too deep.",
263                                 relative_base);
264                 return;
265         }
266
267         last = alt;
268         while (last < ep) {
269                 cp = last;
270                 if (cp < ep && *cp == '#') {
271                         while (cp < ep && *cp != sep)
272                                 cp++;
273                         last = cp + 1;
274                         continue;
275                 }
276                 while (cp < ep && *cp != sep)
277                         cp++;
278                 if (last != cp) {
279                         if (!is_absolute_path(last) && depth) {
280                                 error("%s: ignoring relative alternate object store %s",
281                                                 relative_base, last);
282                         } else {
283                                 link_alt_odb_entry(last, cp - last,
284                                                 relative_base, depth);
285                         }
286                 }
287                 while (cp < ep && *cp == sep)
288                         cp++;
289                 last = cp;
290         }
291 }
292
293 static void read_info_alternates(const char * relative_base, int depth)
294 {
295         char *map;
296         size_t mapsz;
297         struct stat st;
298         const char alt_file_name[] = "info/alternates";
299         /* Given that relative_base is no longer than PATH_MAX,
300            ensure that "path" has enough space to append "/", the
301            file name, "info/alternates", and a trailing NUL.  */
302         char path[PATH_MAX + 1 + sizeof alt_file_name];
303         int fd;
304
305         sprintf(path, "%s/%s", relative_base, alt_file_name);
306         fd = open(path, O_RDONLY);
307         if (fd < 0)
308                 return;
309         if (fstat(fd, &st) || (st.st_size == 0)) {
310                 close(fd);
311                 return;
312         }
313         mapsz = xsize_t(st.st_size);
314         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
315         close(fd);
316
317         link_alt_odb_entries(map, map + mapsz, '\n', relative_base, depth);
318
319         munmap(map, mapsz);
320 }
321
322 void add_to_alternates_file(const char *reference)
323 {
324         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
325         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
326         char *alt = mkpath("%s/objects\n", reference);
327         write_or_die(fd, alt, strlen(alt));
328         if (commit_lock_file(lock))
329                 die("could not close alternates file");
330         if (alt_odb_tail)
331                 link_alt_odb_entries(alt, alt + strlen(alt), '\n', NULL, 0);
332 }
333
334 void foreach_alt_odb(alt_odb_fn fn, void *cb)
335 {
336         struct alternate_object_database *ent;
337
338         prepare_alt_odb();
339         for (ent = alt_odb_list; ent; ent = ent->next)
340                 if (fn(ent, cb))
341                         return;
342 }
343
344 void prepare_alt_odb(void)
345 {
346         const char *alt;
347
348         if (alt_odb_tail)
349                 return;
350
351         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
352         if (!alt) alt = "";
353
354         alt_odb_tail = &alt_odb_list;
355         link_alt_odb_entries(alt, alt + strlen(alt), PATH_SEP, NULL, 0);
356
357         read_info_alternates(get_object_directory(), 0);
358 }
359
360 static int has_loose_object_local(const unsigned char *sha1)
361 {
362         char *name = sha1_file_name(sha1);
363         return !access(name, F_OK);
364 }
365
366 int has_loose_object_nonlocal(const unsigned char *sha1)
367 {
368         struct alternate_object_database *alt;
369         prepare_alt_odb();
370         for (alt = alt_odb_list; alt; alt = alt->next) {
371                 fill_sha1_path(alt->name, sha1);
372                 if (!access(alt->base, F_OK))
373                         return 1;
374         }
375         return 0;
376 }
377
378 static int has_loose_object(const unsigned char *sha1)
379 {
380         return has_loose_object_local(sha1) ||
381                has_loose_object_nonlocal(sha1);
382 }
383
384 static unsigned int pack_used_ctr;
385 static unsigned int pack_mmap_calls;
386 static unsigned int peak_pack_open_windows;
387 static unsigned int pack_open_windows;
388 static size_t peak_pack_mapped;
389 static size_t pack_mapped;
390 struct packed_git *packed_git;
391
392 void pack_report(void)
393 {
394         fprintf(stderr,
395                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
396                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
397                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
398                 sz_fmt(getpagesize()),
399                 sz_fmt(packed_git_window_size),
400                 sz_fmt(packed_git_limit));
401         fprintf(stderr,
402                 "pack_report: pack_used_ctr            = %10u\n"
403                 "pack_report: pack_mmap_calls          = %10u\n"
404                 "pack_report: pack_open_windows        = %10u / %10u\n"
405                 "pack_report: pack_mapped              = "
406                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
407                 pack_used_ctr,
408                 pack_mmap_calls,
409                 pack_open_windows, peak_pack_open_windows,
410                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
411 }
412
413 static int check_packed_git_idx(const char *path,  struct packed_git *p)
414 {
415         void *idx_map;
416         struct pack_idx_header *hdr;
417         size_t idx_size;
418         uint32_t version, nr, i, *index;
419         int fd = open(path, O_RDONLY);
420         struct stat st;
421
422         if (fd < 0)
423                 return -1;
424         if (fstat(fd, &st)) {
425                 close(fd);
426                 return -1;
427         }
428         idx_size = xsize_t(st.st_size);
429         if (idx_size < 4 * 256 + 20 + 20) {
430                 close(fd);
431                 return error("index file %s is too small", path);
432         }
433         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
434         close(fd);
435
436         hdr = idx_map;
437         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
438                 version = ntohl(hdr->idx_version);
439                 if (version < 2 || version > 2) {
440                         munmap(idx_map, idx_size);
441                         return error("index file %s is version %"PRIu32
442                                      " and is not supported by this binary"
443                                      " (try upgrading GIT to a newer version)",
444                                      path, version);
445                 }
446         } else
447                 version = 1;
448
449         nr = 0;
450         index = idx_map;
451         if (version > 1)
452                 index += 2;  /* skip index header */
453         for (i = 0; i < 256; i++) {
454                 uint32_t n = ntohl(index[i]);
455                 if (n < nr) {
456                         munmap(idx_map, idx_size);
457                         return error("non-monotonic index %s", path);
458                 }
459                 nr = n;
460         }
461
462         if (version == 1) {
463                 /*
464                  * Total size:
465                  *  - 256 index entries 4 bytes each
466                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
467                  *  - 20-byte SHA1 of the packfile
468                  *  - 20-byte SHA1 file checksum
469                  */
470                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
471                         munmap(idx_map, idx_size);
472                         return error("wrong index v1 file size in %s", path);
473                 }
474         } else if (version == 2) {
475                 /*
476                  * Minimum size:
477                  *  - 8 bytes of header
478                  *  - 256 index entries 4 bytes each
479                  *  - 20-byte sha1 entry * nr
480                  *  - 4-byte crc entry * nr
481                  *  - 4-byte offset entry * nr
482                  *  - 20-byte SHA1 of the packfile
483                  *  - 20-byte SHA1 file checksum
484                  * And after the 4-byte offset table might be a
485                  * variable sized table containing 8-byte entries
486                  * for offsets larger than 2^31.
487                  */
488                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
489                 unsigned long max_size = min_size;
490                 if (nr)
491                         max_size += (nr - 1)*8;
492                 if (idx_size < min_size || idx_size > max_size) {
493                         munmap(idx_map, idx_size);
494                         return error("wrong index v2 file size in %s", path);
495                 }
496                 if (idx_size != min_size &&
497                     /*
498                      * make sure we can deal with large pack offsets.
499                      * 31-bit signed offset won't be enough, neither
500                      * 32-bit unsigned one will be.
501                      */
502                     (sizeof(off_t) <= 4)) {
503                         munmap(idx_map, idx_size);
504                         return error("pack too large for current definition of off_t in %s", path);
505                 }
506         }
507
508         p->index_version = version;
509         p->index_data = idx_map;
510         p->index_size = idx_size;
511         p->num_objects = nr;
512         return 0;
513 }
514
515 int open_pack_index(struct packed_git *p)
516 {
517         char *idx_name;
518         int ret;
519
520         if (p->index_data)
521                 return 0;
522
523         idx_name = xstrdup(p->pack_name);
524         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
525         ret = check_packed_git_idx(idx_name, p);
526         free(idx_name);
527         return ret;
528 }
529
530 static void scan_windows(struct packed_git *p,
531         struct packed_git **lru_p,
532         struct pack_window **lru_w,
533         struct pack_window **lru_l)
534 {
535         struct pack_window *w, *w_l;
536
537         for (w_l = NULL, w = p->windows; w; w = w->next) {
538                 if (!w->inuse_cnt) {
539                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
540                                 *lru_p = p;
541                                 *lru_w = w;
542                                 *lru_l = w_l;
543                         }
544                 }
545                 w_l = w;
546         }
547 }
548
549 static int unuse_one_window(struct packed_git *current, int keep_fd)
550 {
551         struct packed_git *p, *lru_p = NULL;
552         struct pack_window *lru_w = NULL, *lru_l = NULL;
553
554         if (current)
555                 scan_windows(current, &lru_p, &lru_w, &lru_l);
556         for (p = packed_git; p; p = p->next)
557                 scan_windows(p, &lru_p, &lru_w, &lru_l);
558         if (lru_p) {
559                 munmap(lru_w->base, lru_w->len);
560                 pack_mapped -= lru_w->len;
561                 if (lru_l)
562                         lru_l->next = lru_w->next;
563                 else {
564                         lru_p->windows = lru_w->next;
565                         if (!lru_p->windows && lru_p->pack_fd != keep_fd) {
566                                 close(lru_p->pack_fd);
567                                 lru_p->pack_fd = -1;
568                         }
569                 }
570                 free(lru_w);
571                 pack_open_windows--;
572                 return 1;
573         }
574         return 0;
575 }
576
577 void release_pack_memory(size_t need, int fd)
578 {
579         size_t cur = pack_mapped;
580         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
581                 ; /* nothing */
582 }
583
584 void close_pack_windows(struct packed_git *p)
585 {
586         while (p->windows) {
587                 struct pack_window *w = p->windows;
588
589                 if (w->inuse_cnt)
590                         die("pack '%s' still has open windows to it",
591                             p->pack_name);
592                 munmap(w->base, w->len);
593                 pack_mapped -= w->len;
594                 pack_open_windows--;
595                 p->windows = w->next;
596                 free(w);
597         }
598 }
599
600 void unuse_pack(struct pack_window **w_cursor)
601 {
602         struct pack_window *w = *w_cursor;
603         if (w) {
604                 w->inuse_cnt--;
605                 *w_cursor = NULL;
606         }
607 }
608
609 /*
610  * This is used by git-repack in case a newly created pack happens to
611  * contain the same set of objects as an existing one.  In that case
612  * the resulting file might be different even if its name would be the
613  * same.  It is best to close any reference to the old pack before it is
614  * replaced on disk.  Of course no index pointers nor windows for given pack
615  * must subsist at this point.  If ever objects from this pack are requested
616  * again, the new version of the pack will be reinitialized through
617  * reprepare_packed_git().
618  */
619 void free_pack_by_name(const char *pack_name)
620 {
621         struct packed_git *p, **pp = &packed_git;
622
623         while (*pp) {
624                 p = *pp;
625                 if (strcmp(pack_name, p->pack_name) == 0) {
626                         clear_delta_base_cache();
627                         close_pack_windows(p);
628                         if (p->pack_fd != -1)
629                                 close(p->pack_fd);
630                         if (p->index_data)
631                                 munmap((void *)p->index_data, p->index_size);
632                         free(p->bad_object_sha1);
633                         *pp = p->next;
634                         free(p);
635                         return;
636                 }
637                 pp = &p->next;
638         }
639 }
640
641 /*
642  * Do not call this directly as this leaks p->pack_fd on error return;
643  * call open_packed_git() instead.
644  */
645 static int open_packed_git_1(struct packed_git *p)
646 {
647         struct stat st;
648         struct pack_header hdr;
649         unsigned char sha1[20];
650         unsigned char *idx_sha1;
651         long fd_flag;
652
653         if (!p->index_data && open_pack_index(p))
654                 return error("packfile %s index unavailable", p->pack_name);
655
656         p->pack_fd = open(p->pack_name, O_RDONLY);
657         while (p->pack_fd < 0 && errno == EMFILE && unuse_one_window(p, -1))
658                 p->pack_fd = open(p->pack_name, O_RDONLY);
659         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
660                 return -1;
661
662         /* If we created the struct before we had the pack we lack size. */
663         if (!p->pack_size) {
664                 if (!S_ISREG(st.st_mode))
665                         return error("packfile %s not a regular file", p->pack_name);
666                 p->pack_size = st.st_size;
667         } else if (p->pack_size != st.st_size)
668                 return error("packfile %s size changed", p->pack_name);
669
670         /* We leave these file descriptors open with sliding mmap;
671          * there is no point keeping them open across exec(), though.
672          */
673         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
674         if (fd_flag < 0)
675                 return error("cannot determine file descriptor flags");
676         fd_flag |= FD_CLOEXEC;
677         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
678                 return error("cannot set FD_CLOEXEC");
679
680         /* Verify we recognize this pack file format. */
681         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
682                 return error("file %s is far too short to be a packfile", p->pack_name);
683         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
684                 return error("file %s is not a GIT packfile", p->pack_name);
685         if (!pack_version_ok(hdr.hdr_version))
686                 return error("packfile %s is version %"PRIu32" and not"
687                         " supported (try upgrading GIT to a newer version)",
688                         p->pack_name, ntohl(hdr.hdr_version));
689
690         /* Verify the pack matches its index. */
691         if (p->num_objects != ntohl(hdr.hdr_entries))
692                 return error("packfile %s claims to have %"PRIu32" objects"
693                              " while index indicates %"PRIu32" objects",
694                              p->pack_name, ntohl(hdr.hdr_entries),
695                              p->num_objects);
696         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
697                 return error("end of packfile %s is unavailable", p->pack_name);
698         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
699                 return error("packfile %s signature is unavailable", p->pack_name);
700         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
701         if (hashcmp(sha1, idx_sha1))
702                 return error("packfile %s does not match index", p->pack_name);
703         return 0;
704 }
705
706 static int open_packed_git(struct packed_git *p)
707 {
708         if (!open_packed_git_1(p))
709                 return 0;
710         if (p->pack_fd != -1) {
711                 close(p->pack_fd);
712                 p->pack_fd = -1;
713         }
714         return -1;
715 }
716
717 static int in_window(struct pack_window *win, off_t offset)
718 {
719         /* We must promise at least 20 bytes (one hash) after the
720          * offset is available from this window, otherwise the offset
721          * is not actually in this window and a different window (which
722          * has that one hash excess) must be used.  This is to support
723          * the object header and delta base parsing routines below.
724          */
725         off_t win_off = win->offset;
726         return win_off <= offset
727                 && (offset + 20) <= (win_off + win->len);
728 }
729
730 unsigned char *use_pack(struct packed_git *p,
731                 struct pack_window **w_cursor,
732                 off_t offset,
733                 unsigned int *left)
734 {
735         struct pack_window *win = *w_cursor;
736
737         if (p->pack_fd == -1 && open_packed_git(p))
738                 die("packfile %s cannot be accessed", p->pack_name);
739
740         /* Since packfiles end in a hash of their content and it's
741          * pointless to ask for an offset into the middle of that
742          * hash, and the in_window function above wouldn't match
743          * don't allow an offset too close to the end of the file.
744          */
745         if (offset > (p->pack_size - 20))
746                 die("offset beyond end of packfile (truncated pack?)");
747
748         if (!win || !in_window(win, offset)) {
749                 if (win)
750                         win->inuse_cnt--;
751                 for (win = p->windows; win; win = win->next) {
752                         if (in_window(win, offset))
753                                 break;
754                 }
755                 if (!win) {
756                         size_t window_align = packed_git_window_size / 2;
757                         off_t len;
758                         win = xcalloc(1, sizeof(*win));
759                         win->offset = (offset / window_align) * window_align;
760                         len = p->pack_size - win->offset;
761                         if (len > packed_git_window_size)
762                                 len = packed_git_window_size;
763                         win->len = (size_t)len;
764                         pack_mapped += win->len;
765                         while (packed_git_limit < pack_mapped
766                                 && unuse_one_window(p, p->pack_fd))
767                                 ; /* nothing */
768                         win->base = xmmap(NULL, win->len,
769                                 PROT_READ, MAP_PRIVATE,
770                                 p->pack_fd, win->offset);
771                         if (win->base == MAP_FAILED)
772                                 die("packfile %s cannot be mapped: %s",
773                                         p->pack_name,
774                                         strerror(errno));
775                         pack_mmap_calls++;
776                         pack_open_windows++;
777                         if (pack_mapped > peak_pack_mapped)
778                                 peak_pack_mapped = pack_mapped;
779                         if (pack_open_windows > peak_pack_open_windows)
780                                 peak_pack_open_windows = pack_open_windows;
781                         win->next = p->windows;
782                         p->windows = win;
783                 }
784         }
785         if (win != *w_cursor) {
786                 win->last_used = pack_used_ctr++;
787                 win->inuse_cnt++;
788                 *w_cursor = win;
789         }
790         offset -= win->offset;
791         if (left)
792                 *left = win->len - xsize_t(offset);
793         return win->base + offset;
794 }
795
796 static struct packed_git *alloc_packed_git(int extra)
797 {
798         struct packed_git *p = xmalloc(sizeof(*p) + extra);
799         memset(p, 0, sizeof(*p));
800         p->pack_fd = -1;
801         return p;
802 }
803
804 struct packed_git *add_packed_git(const char *path, int path_len, int local)
805 {
806         struct stat st;
807         struct packed_git *p = alloc_packed_git(path_len + 2);
808
809         /*
810          * Make sure a corresponding .pack file exists and that
811          * the index looks sane.
812          */
813         path_len -= strlen(".idx");
814         if (path_len < 1) {
815                 free(p);
816                 return NULL;
817         }
818         memcpy(p->pack_name, path, path_len);
819
820         strcpy(p->pack_name + path_len, ".keep");
821         if (!access(p->pack_name, F_OK))
822                 p->pack_keep = 1;
823
824         strcpy(p->pack_name + path_len, ".pack");
825         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
826                 free(p);
827                 return NULL;
828         }
829
830         /* ok, it looks sane as far as we can check without
831          * actually mapping the pack file.
832          */
833         p->pack_size = st.st_size;
834         p->pack_local = local;
835         p->mtime = st.st_mtime;
836         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
837                 hashclr(p->sha1);
838         return p;
839 }
840
841 struct packed_git *parse_pack_index(unsigned char *sha1)
842 {
843         const char *idx_path = sha1_pack_index_name(sha1);
844         const char *path = sha1_pack_name(sha1);
845         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
846
847         strcpy(p->pack_name, path);
848         hashcpy(p->sha1, sha1);
849         if (check_packed_git_idx(idx_path, p)) {
850                 free(p);
851                 return NULL;
852         }
853
854         return p;
855 }
856
857 void install_packed_git(struct packed_git *pack)
858 {
859         pack->next = packed_git;
860         packed_git = pack;
861 }
862
863 static void prepare_packed_git_one(char *objdir, int local)
864 {
865         /* Ensure that this buffer is large enough so that we can
866            append "/pack/" without clobbering the stack even if
867            strlen(objdir) were PATH_MAX.  */
868         char path[PATH_MAX + 1 + 4 + 1 + 1];
869         int len;
870         DIR *dir;
871         struct dirent *de;
872
873         sprintf(path, "%s/pack", objdir);
874         len = strlen(path);
875         dir = opendir(path);
876         while (!dir && errno == EMFILE && unuse_one_window(packed_git, -1))
877                 dir = opendir(path);
878         if (!dir) {
879                 if (errno != ENOENT)
880                         error("unable to open object pack directory: %s: %s",
881                               path, strerror(errno));
882                 return;
883         }
884         path[len++] = '/';
885         while ((de = readdir(dir)) != NULL) {
886                 int namelen = strlen(de->d_name);
887                 struct packed_git *p;
888
889                 if (!has_extension(de->d_name, ".idx"))
890                         continue;
891
892                 if (len + namelen + 1 > sizeof(path))
893                         continue;
894
895                 /* Don't reopen a pack we already have. */
896                 strcpy(path + len, de->d_name);
897                 for (p = packed_git; p; p = p->next) {
898                         if (!memcmp(path, p->pack_name, len + namelen - 4))
899                                 break;
900                 }
901                 if (p)
902                         continue;
903                 /* See if it really is a valid .idx file with corresponding
904                  * .pack file that we can map.
905                  */
906                 p = add_packed_git(path, len + namelen, local);
907                 if (!p)
908                         continue;
909                 install_packed_git(p);
910         }
911         closedir(dir);
912 }
913
914 static int sort_pack(const void *a_, const void *b_)
915 {
916         struct packed_git *a = *((struct packed_git **)a_);
917         struct packed_git *b = *((struct packed_git **)b_);
918         int st;
919
920         /*
921          * Local packs tend to contain objects specific to our
922          * variant of the project than remote ones.  In addition,
923          * remote ones could be on a network mounted filesystem.
924          * Favor local ones for these reasons.
925          */
926         st = a->pack_local - b->pack_local;
927         if (st)
928                 return -st;
929
930         /*
931          * Younger packs tend to contain more recent objects,
932          * and more recent objects tend to get accessed more
933          * often.
934          */
935         if (a->mtime < b->mtime)
936                 return 1;
937         else if (a->mtime == b->mtime)
938                 return 0;
939         return -1;
940 }
941
942 static void rearrange_packed_git(void)
943 {
944         struct packed_git **ary, *p;
945         int i, n;
946
947         for (n = 0, p = packed_git; p; p = p->next)
948                 n++;
949         if (n < 2)
950                 return;
951
952         /* prepare an array of packed_git for easier sorting */
953         ary = xcalloc(n, sizeof(struct packed_git *));
954         for (n = 0, p = packed_git; p; p = p->next)
955                 ary[n++] = p;
956
957         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
958
959         /* link them back again */
960         for (i = 0; i < n - 1; i++)
961                 ary[i]->next = ary[i + 1];
962         ary[n - 1]->next = NULL;
963         packed_git = ary[0];
964
965         free(ary);
966 }
967
968 static int prepare_packed_git_run_once = 0;
969 void prepare_packed_git(void)
970 {
971         struct alternate_object_database *alt;
972
973         if (prepare_packed_git_run_once)
974                 return;
975         prepare_packed_git_one(get_object_directory(), 1);
976         prepare_alt_odb();
977         for (alt = alt_odb_list; alt; alt = alt->next) {
978                 alt->name[-1] = 0;
979                 prepare_packed_git_one(alt->base, 0);
980                 alt->name[-1] = '/';
981         }
982         rearrange_packed_git();
983         prepare_packed_git_run_once = 1;
984 }
985
986 void reprepare_packed_git(void)
987 {
988         discard_revindex();
989         prepare_packed_git_run_once = 0;
990         prepare_packed_git();
991 }
992
993 static void mark_bad_packed_object(struct packed_git *p,
994                                    const unsigned char *sha1)
995 {
996         unsigned i;
997         for (i = 0; i < p->num_bad_objects; i++)
998                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
999                         return;
1000         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1001         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1002         p->num_bad_objects++;
1003 }
1004
1005 static int has_packed_and_bad(const unsigned char *sha1)
1006 {
1007         struct packed_git *p;
1008         unsigned i;
1009
1010         for (p = packed_git; p; p = p->next)
1011                 for (i = 0; i < p->num_bad_objects; i++)
1012                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1013                                 return 1;
1014         return 0;
1015 }
1016
1017 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1018 {
1019         unsigned char real_sha1[20];
1020         hash_sha1_file(map, size, type, real_sha1);
1021         return hashcmp(sha1, real_sha1) ? -1 : 0;
1022 }
1023
1024 static int git_open_noatime(const char *name)
1025 {
1026         static int sha1_file_open_flag = O_NOATIME;
1027         int fd = open(name, O_RDONLY | sha1_file_open_flag);
1028
1029         /* Might the failure be due to O_NOATIME? */
1030         if (fd < 0 && errno != ENOENT && sha1_file_open_flag) {
1031                 fd = open(name, O_RDONLY);
1032                 if (fd >= 0)
1033                         sha1_file_open_flag = 0;
1034         }
1035         return fd;
1036 }
1037
1038 static int open_sha1_file(const unsigned char *sha1)
1039 {
1040         int fd;
1041         char *name = sha1_file_name(sha1);
1042         struct alternate_object_database *alt;
1043
1044         fd = git_open_noatime(name);
1045         if (fd >= 0)
1046                 return fd;
1047
1048         prepare_alt_odb();
1049         errno = ENOENT;
1050         for (alt = alt_odb_list; alt; alt = alt->next) {
1051                 name = alt->name;
1052                 fill_sha1_path(name, sha1);
1053                 fd = git_open_noatime(alt->base);
1054                 if (fd >= 0)
1055                         return fd;
1056         }
1057         return -1;
1058 }
1059
1060 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1061 {
1062         void *map;
1063         int fd;
1064
1065         fd = open_sha1_file(sha1);
1066         map = NULL;
1067         if (fd >= 0) {
1068                 struct stat st;
1069
1070                 if (!fstat(fd, &st)) {
1071                         *size = xsize_t(st.st_size);
1072                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1073                 }
1074                 close(fd);
1075         }
1076         return map;
1077 }
1078
1079 static int legacy_loose_object(unsigned char *map)
1080 {
1081         unsigned int word;
1082
1083         /*
1084          * Is it a zlib-compressed buffer? If so, the first byte
1085          * must be 0x78 (15-bit window size, deflated), and the
1086          * first 16-bit word is evenly divisible by 31
1087          */
1088         word = (map[0] << 8) + map[1];
1089         if (map[0] == 0x78 && !(word % 31))
1090                 return 1;
1091         else
1092                 return 0;
1093 }
1094
1095 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1096                 unsigned long len, enum object_type *type, unsigned long *sizep)
1097 {
1098         unsigned shift;
1099         unsigned long size, c;
1100         unsigned long used = 0;
1101
1102         c = buf[used++];
1103         *type = (c >> 4) & 7;
1104         size = c & 15;
1105         shift = 4;
1106         while (c & 0x80) {
1107                 if (len <= used || bitsizeof(long) <= shift) {
1108                         error("bad object header");
1109                         return 0;
1110                 }
1111                 c = buf[used++];
1112                 size += (c & 0x7f) << shift;
1113                 shift += 7;
1114         }
1115         *sizep = size;
1116         return used;
1117 }
1118
1119 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1120 {
1121         unsigned long size, used;
1122         static const char valid_loose_object_type[8] = {
1123                 0, /* OBJ_EXT */
1124                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1125                 0, /* "delta" and others are invalid in a loose object */
1126         };
1127         enum object_type type;
1128
1129         /* Get the data stream */
1130         memset(stream, 0, sizeof(*stream));
1131         stream->next_in = map;
1132         stream->avail_in = mapsize;
1133         stream->next_out = buffer;
1134         stream->avail_out = bufsiz;
1135
1136         if (legacy_loose_object(map)) {
1137                 git_inflate_init(stream);
1138                 return git_inflate(stream, 0);
1139         }
1140
1141
1142         /*
1143          * There used to be a second loose object header format which
1144          * was meant to mimic the in-pack format, allowing for direct
1145          * copy of the object data.  This format turned up not to be
1146          * really worth it and we don't write it any longer.  But we
1147          * can still read it.
1148          */
1149         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1150         if (!used || !valid_loose_object_type[type])
1151                 return -1;
1152         map += used;
1153         mapsize -= used;
1154
1155         /* Set up the stream for the rest.. */
1156         stream->next_in = map;
1157         stream->avail_in = mapsize;
1158         git_inflate_init(stream);
1159
1160         /* And generate the fake traditional header */
1161         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1162                                          typename(type), size);
1163         return 0;
1164 }
1165
1166 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1167 {
1168         int bytes = strlen(buffer) + 1;
1169         unsigned char *buf = xmallocz(size);
1170         unsigned long n;
1171         int status = Z_OK;
1172
1173         n = stream->total_out - bytes;
1174         if (n > size)
1175                 n = size;
1176         memcpy(buf, (char *) buffer + bytes, n);
1177         bytes = n;
1178         if (bytes <= size) {
1179                 /*
1180                  * The above condition must be (bytes <= size), not
1181                  * (bytes < size).  In other words, even though we
1182                  * expect no more output and set avail_out to zer0,
1183                  * the input zlib stream may have bytes that express
1184                  * "this concludes the stream", and we *do* want to
1185                  * eat that input.
1186                  *
1187                  * Otherwise we would not be able to test that we
1188                  * consumed all the input to reach the expected size;
1189                  * we also want to check that zlib tells us that all
1190                  * went well with status == Z_STREAM_END at the end.
1191                  */
1192                 stream->next_out = buf + bytes;
1193                 stream->avail_out = size - bytes;
1194                 while (status == Z_OK)
1195                         status = git_inflate(stream, Z_FINISH);
1196         }
1197         if (status == Z_STREAM_END && !stream->avail_in) {
1198                 git_inflate_end(stream);
1199                 return buf;
1200         }
1201
1202         if (status < 0)
1203                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1204         else if (stream->avail_in)
1205                 error("garbage at end of loose object '%s'",
1206                       sha1_to_hex(sha1));
1207         free(buf);
1208         return NULL;
1209 }
1210
1211 /*
1212  * We used to just use "sscanf()", but that's actually way
1213  * too permissive for what we want to check. So do an anal
1214  * object header parse by hand.
1215  */
1216 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1217 {
1218         char type[10];
1219         int i;
1220         unsigned long size;
1221
1222         /*
1223          * The type can be at most ten bytes (including the
1224          * terminating '\0' that we add), and is followed by
1225          * a space.
1226          */
1227         i = 0;
1228         for (;;) {
1229                 char c = *hdr++;
1230                 if (c == ' ')
1231                         break;
1232                 type[i++] = c;
1233                 if (i >= sizeof(type))
1234                         return -1;
1235         }
1236         type[i] = 0;
1237
1238         /*
1239          * The length must follow immediately, and be in canonical
1240          * decimal format (ie "010" is not valid).
1241          */
1242         size = *hdr++ - '0';
1243         if (size > 9)
1244                 return -1;
1245         if (size) {
1246                 for (;;) {
1247                         unsigned long c = *hdr - '0';
1248                         if (c > 9)
1249                                 break;
1250                         hdr++;
1251                         size = size * 10 + c;
1252                 }
1253         }
1254         *sizep = size;
1255
1256         /*
1257          * The length must be followed by a zero byte
1258          */
1259         return *hdr ? -1 : type_from_string(type);
1260 }
1261
1262 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1263 {
1264         int ret;
1265         z_stream stream;
1266         char hdr[8192];
1267
1268         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1269         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1270                 return NULL;
1271
1272         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1273 }
1274
1275 unsigned long get_size_from_delta(struct packed_git *p,
1276                                   struct pack_window **w_curs,
1277                                   off_t curpos)
1278 {
1279         const unsigned char *data;
1280         unsigned char delta_head[20], *in;
1281         z_stream stream;
1282         int st;
1283
1284         memset(&stream, 0, sizeof(stream));
1285         stream.next_out = delta_head;
1286         stream.avail_out = sizeof(delta_head);
1287
1288         git_inflate_init(&stream);
1289         do {
1290                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1291                 stream.next_in = in;
1292                 st = git_inflate(&stream, Z_FINISH);
1293                 curpos += stream.next_in - in;
1294         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1295                  stream.total_out < sizeof(delta_head));
1296         git_inflate_end(&stream);
1297         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1298                 error("delta data unpack-initial failed");
1299                 return 0;
1300         }
1301
1302         /* Examine the initial part of the delta to figure out
1303          * the result size.
1304          */
1305         data = delta_head;
1306
1307         /* ignore base size */
1308         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1309
1310         /* Read the result size */
1311         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1312 }
1313
1314 static off_t get_delta_base(struct packed_git *p,
1315                                     struct pack_window **w_curs,
1316                                     off_t *curpos,
1317                                     enum object_type type,
1318                                     off_t delta_obj_offset)
1319 {
1320         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1321         off_t base_offset;
1322
1323         /* use_pack() assured us we have [base_info, base_info + 20)
1324          * as a range that we can look at without walking off the
1325          * end of the mapped window.  Its actually the hash size
1326          * that is assured.  An OFS_DELTA longer than the hash size
1327          * is stupid, as then a REF_DELTA would be smaller to store.
1328          */
1329         if (type == OBJ_OFS_DELTA) {
1330                 unsigned used = 0;
1331                 unsigned char c = base_info[used++];
1332                 base_offset = c & 127;
1333                 while (c & 128) {
1334                         base_offset += 1;
1335                         if (!base_offset || MSB(base_offset, 7))
1336                                 return 0;  /* overflow */
1337                         c = base_info[used++];
1338                         base_offset = (base_offset << 7) + (c & 127);
1339                 }
1340                 base_offset = delta_obj_offset - base_offset;
1341                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1342                         return 0;  /* out of bound */
1343                 *curpos += used;
1344         } else if (type == OBJ_REF_DELTA) {
1345                 /* The base entry _must_ be in the same pack */
1346                 base_offset = find_pack_entry_one(base_info, p);
1347                 *curpos += 20;
1348         } else
1349                 die("I am totally screwed");
1350         return base_offset;
1351 }
1352
1353 /* forward declaration for a mutually recursive function */
1354 static int packed_object_info(struct packed_git *p, off_t offset,
1355                               unsigned long *sizep);
1356
1357 static int packed_delta_info(struct packed_git *p,
1358                              struct pack_window **w_curs,
1359                              off_t curpos,
1360                              enum object_type type,
1361                              off_t obj_offset,
1362                              unsigned long *sizep)
1363 {
1364         off_t base_offset;
1365
1366         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1367         if (!base_offset)
1368                 return OBJ_BAD;
1369         type = packed_object_info(p, base_offset, NULL);
1370         if (type <= OBJ_NONE) {
1371                 struct revindex_entry *revidx;
1372                 const unsigned char *base_sha1;
1373                 revidx = find_pack_revindex(p, base_offset);
1374                 if (!revidx)
1375                         return OBJ_BAD;
1376                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1377                 mark_bad_packed_object(p, base_sha1);
1378                 type = sha1_object_info(base_sha1, NULL);
1379                 if (type <= OBJ_NONE)
1380                         return OBJ_BAD;
1381         }
1382
1383         /* We choose to only get the type of the base object and
1384          * ignore potentially corrupt pack file that expects the delta
1385          * based on a base with a wrong size.  This saves tons of
1386          * inflate() calls.
1387          */
1388         if (sizep) {
1389                 *sizep = get_size_from_delta(p, w_curs, curpos);
1390                 if (*sizep == 0)
1391                         type = OBJ_BAD;
1392         }
1393
1394         return type;
1395 }
1396
1397 static int unpack_object_header(struct packed_git *p,
1398                                 struct pack_window **w_curs,
1399                                 off_t *curpos,
1400                                 unsigned long *sizep)
1401 {
1402         unsigned char *base;
1403         unsigned int left;
1404         unsigned long used;
1405         enum object_type type;
1406
1407         /* use_pack() assures us we have [base, base + 20) available
1408          * as a range that we can look at at.  (Its actually the hash
1409          * size that is assured.)  With our object header encoding
1410          * the maximum deflated object size is 2^137, which is just
1411          * insane, so we know won't exceed what we have been given.
1412          */
1413         base = use_pack(p, w_curs, *curpos, &left);
1414         used = unpack_object_header_buffer(base, left, &type, sizep);
1415         if (!used) {
1416                 type = OBJ_BAD;
1417         } else
1418                 *curpos += used;
1419
1420         return type;
1421 }
1422
1423 const char *packed_object_info_detail(struct packed_git *p,
1424                                       off_t obj_offset,
1425                                       unsigned long *size,
1426                                       unsigned long *store_size,
1427                                       unsigned int *delta_chain_length,
1428                                       unsigned char *base_sha1)
1429 {
1430         struct pack_window *w_curs = NULL;
1431         off_t curpos;
1432         unsigned long dummy;
1433         unsigned char *next_sha1;
1434         enum object_type type;
1435         struct revindex_entry *revidx;
1436
1437         *delta_chain_length = 0;
1438         curpos = obj_offset;
1439         type = unpack_object_header(p, &w_curs, &curpos, size);
1440
1441         revidx = find_pack_revindex(p, obj_offset);
1442         *store_size = revidx[1].offset - obj_offset;
1443
1444         for (;;) {
1445                 switch (type) {
1446                 default:
1447                         die("pack %s contains unknown object type %d",
1448                             p->pack_name, type);
1449                 case OBJ_COMMIT:
1450                 case OBJ_TREE:
1451                 case OBJ_BLOB:
1452                 case OBJ_TAG:
1453                         unuse_pack(&w_curs);
1454                         return typename(type);
1455                 case OBJ_OFS_DELTA:
1456                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1457                         if (!obj_offset)
1458                                 die("pack %s contains bad delta base reference of type %s",
1459                                     p->pack_name, typename(type));
1460                         if (*delta_chain_length == 0) {
1461                                 revidx = find_pack_revindex(p, obj_offset);
1462                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1463                         }
1464                         break;
1465                 case OBJ_REF_DELTA:
1466                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1467                         if (*delta_chain_length == 0)
1468                                 hashcpy(base_sha1, next_sha1);
1469                         obj_offset = find_pack_entry_one(next_sha1, p);
1470                         break;
1471                 }
1472                 (*delta_chain_length)++;
1473                 curpos = obj_offset;
1474                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1475         }
1476 }
1477
1478 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1479                               unsigned long *sizep)
1480 {
1481         struct pack_window *w_curs = NULL;
1482         unsigned long size;
1483         off_t curpos = obj_offset;
1484         enum object_type type;
1485
1486         type = unpack_object_header(p, &w_curs, &curpos, &size);
1487
1488         switch (type) {
1489         case OBJ_OFS_DELTA:
1490         case OBJ_REF_DELTA:
1491                 type = packed_delta_info(p, &w_curs, curpos,
1492                                          type, obj_offset, sizep);
1493                 break;
1494         case OBJ_COMMIT:
1495         case OBJ_TREE:
1496         case OBJ_BLOB:
1497         case OBJ_TAG:
1498                 if (sizep)
1499                         *sizep = size;
1500                 break;
1501         default:
1502                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1503                       type, (uintmax_t)obj_offset, p->pack_name);
1504                 type = OBJ_BAD;
1505         }
1506         unuse_pack(&w_curs);
1507         return type;
1508 }
1509
1510 static void *unpack_compressed_entry(struct packed_git *p,
1511                                     struct pack_window **w_curs,
1512                                     off_t curpos,
1513                                     unsigned long size)
1514 {
1515         int st;
1516         z_stream stream;
1517         unsigned char *buffer, *in;
1518
1519         buffer = xmallocz(size);
1520         memset(&stream, 0, sizeof(stream));
1521         stream.next_out = buffer;
1522         stream.avail_out = size + 1;
1523
1524         git_inflate_init(&stream);
1525         do {
1526                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1527                 stream.next_in = in;
1528                 st = git_inflate(&stream, Z_FINISH);
1529                 if (!stream.avail_out)
1530                         break; /* the payload is larger than it should be */
1531                 curpos += stream.next_in - in;
1532         } while (st == Z_OK || st == Z_BUF_ERROR);
1533         git_inflate_end(&stream);
1534         if ((st != Z_STREAM_END) || stream.total_out != size) {
1535                 free(buffer);
1536                 return NULL;
1537         }
1538
1539         return buffer;
1540 }
1541
1542 #define MAX_DELTA_CACHE (256)
1543
1544 static size_t delta_base_cached;
1545
1546 static struct delta_base_cache_lru_list {
1547         struct delta_base_cache_lru_list *prev;
1548         struct delta_base_cache_lru_list *next;
1549 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1550
1551 static struct delta_base_cache_entry {
1552         struct delta_base_cache_lru_list lru;
1553         void *data;
1554         struct packed_git *p;
1555         off_t base_offset;
1556         unsigned long size;
1557         enum object_type type;
1558 } delta_base_cache[MAX_DELTA_CACHE];
1559
1560 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1561 {
1562         unsigned long hash;
1563
1564         hash = (unsigned long)p + (unsigned long)base_offset;
1565         hash += (hash >> 8) + (hash >> 16);
1566         return hash % MAX_DELTA_CACHE;
1567 }
1568
1569 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1570         unsigned long *base_size, enum object_type *type, int keep_cache)
1571 {
1572         void *ret;
1573         unsigned long hash = pack_entry_hash(p, base_offset);
1574         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1575
1576         ret = ent->data;
1577         if (!ret || ent->p != p || ent->base_offset != base_offset)
1578                 return unpack_entry(p, base_offset, type, base_size);
1579
1580         if (!keep_cache) {
1581                 ent->data = NULL;
1582                 ent->lru.next->prev = ent->lru.prev;
1583                 ent->lru.prev->next = ent->lru.next;
1584                 delta_base_cached -= ent->size;
1585         } else {
1586                 ret = xmemdupz(ent->data, ent->size);
1587         }
1588         *type = ent->type;
1589         *base_size = ent->size;
1590         return ret;
1591 }
1592
1593 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1594 {
1595         if (ent->data) {
1596                 free(ent->data);
1597                 ent->data = NULL;
1598                 ent->lru.next->prev = ent->lru.prev;
1599                 ent->lru.prev->next = ent->lru.next;
1600                 delta_base_cached -= ent->size;
1601         }
1602 }
1603
1604 void clear_delta_base_cache(void)
1605 {
1606         unsigned long p;
1607         for (p = 0; p < MAX_DELTA_CACHE; p++)
1608                 release_delta_base_cache(&delta_base_cache[p]);
1609 }
1610
1611 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1612         void *base, unsigned long base_size, enum object_type type)
1613 {
1614         unsigned long hash = pack_entry_hash(p, base_offset);
1615         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1616         struct delta_base_cache_lru_list *lru;
1617
1618         release_delta_base_cache(ent);
1619         delta_base_cached += base_size;
1620
1621         for (lru = delta_base_cache_lru.next;
1622              delta_base_cached > delta_base_cache_limit
1623              && lru != &delta_base_cache_lru;
1624              lru = lru->next) {
1625                 struct delta_base_cache_entry *f = (void *)lru;
1626                 if (f->type == OBJ_BLOB)
1627                         release_delta_base_cache(f);
1628         }
1629         for (lru = delta_base_cache_lru.next;
1630              delta_base_cached > delta_base_cache_limit
1631              && lru != &delta_base_cache_lru;
1632              lru = lru->next) {
1633                 struct delta_base_cache_entry *f = (void *)lru;
1634                 release_delta_base_cache(f);
1635         }
1636
1637         ent->p = p;
1638         ent->base_offset = base_offset;
1639         ent->type = type;
1640         ent->data = base;
1641         ent->size = base_size;
1642         ent->lru.next = &delta_base_cache_lru;
1643         ent->lru.prev = delta_base_cache_lru.prev;
1644         delta_base_cache_lru.prev->next = &ent->lru;
1645         delta_base_cache_lru.prev = &ent->lru;
1646 }
1647
1648 static void *read_object(const unsigned char *sha1, enum object_type *type,
1649                          unsigned long *size);
1650
1651 static void *unpack_delta_entry(struct packed_git *p,
1652                                 struct pack_window **w_curs,
1653                                 off_t curpos,
1654                                 unsigned long delta_size,
1655                                 off_t obj_offset,
1656                                 enum object_type *type,
1657                                 unsigned long *sizep)
1658 {
1659         void *delta_data, *result, *base;
1660         unsigned long base_size;
1661         off_t base_offset;
1662
1663         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1664         if (!base_offset) {
1665                 error("failed to validate delta base reference "
1666                       "at offset %"PRIuMAX" from %s",
1667                       (uintmax_t)curpos, p->pack_name);
1668                 return NULL;
1669         }
1670         unuse_pack(w_curs);
1671         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1672         if (!base) {
1673                 /*
1674                  * We're probably in deep shit, but let's try to fetch
1675                  * the required base anyway from another pack or loose.
1676                  * This is costly but should happen only in the presence
1677                  * of a corrupted pack, and is better than failing outright.
1678                  */
1679                 struct revindex_entry *revidx;
1680                 const unsigned char *base_sha1;
1681                 revidx = find_pack_revindex(p, base_offset);
1682                 if (!revidx)
1683                         return NULL;
1684                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1685                 error("failed to read delta base object %s"
1686                       " at offset %"PRIuMAX" from %s",
1687                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1688                       p->pack_name);
1689                 mark_bad_packed_object(p, base_sha1);
1690                 base = read_object(base_sha1, type, &base_size);
1691                 if (!base)
1692                         return NULL;
1693         }
1694
1695         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1696         if (!delta_data) {
1697                 error("failed to unpack compressed delta "
1698                       "at offset %"PRIuMAX" from %s",
1699                       (uintmax_t)curpos, p->pack_name);
1700                 free(base);
1701                 return NULL;
1702         }
1703         result = patch_delta(base, base_size,
1704                              delta_data, delta_size,
1705                              sizep);
1706         if (!result)
1707                 die("failed to apply delta");
1708         free(delta_data);
1709         add_delta_base_cache(p, base_offset, base, base_size, *type);
1710         return result;
1711 }
1712
1713 int do_check_packed_object_crc;
1714
1715 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1716                    enum object_type *type, unsigned long *sizep)
1717 {
1718         struct pack_window *w_curs = NULL;
1719         off_t curpos = obj_offset;
1720         void *data;
1721
1722         if (do_check_packed_object_crc && p->index_version > 1) {
1723                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1724                 unsigned long len = revidx[1].offset - obj_offset;
1725                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1726                         const unsigned char *sha1 =
1727                                 nth_packed_object_sha1(p, revidx->nr);
1728                         error("bad packed object CRC for %s",
1729                               sha1_to_hex(sha1));
1730                         mark_bad_packed_object(p, sha1);
1731                         unuse_pack(&w_curs);
1732                         return NULL;
1733                 }
1734         }
1735
1736         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1737         switch (*type) {
1738         case OBJ_OFS_DELTA:
1739         case OBJ_REF_DELTA:
1740                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1741                                           obj_offset, type, sizep);
1742                 break;
1743         case OBJ_COMMIT:
1744         case OBJ_TREE:
1745         case OBJ_BLOB:
1746         case OBJ_TAG:
1747                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1748                 break;
1749         default:
1750                 data = NULL;
1751                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1752                       *type, (uintmax_t)obj_offset, p->pack_name);
1753         }
1754         unuse_pack(&w_curs);
1755         return data;
1756 }
1757
1758 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1759                                             uint32_t n)
1760 {
1761         const unsigned char *index = p->index_data;
1762         if (!index) {
1763                 if (open_pack_index(p))
1764                         return NULL;
1765                 index = p->index_data;
1766         }
1767         if (n >= p->num_objects)
1768                 return NULL;
1769         index += 4 * 256;
1770         if (p->index_version == 1) {
1771                 return index + 24 * n + 4;
1772         } else {
1773                 index += 8;
1774                 return index + 20 * n;
1775         }
1776 }
1777
1778 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1779 {
1780         const unsigned char *index = p->index_data;
1781         index += 4 * 256;
1782         if (p->index_version == 1) {
1783                 return ntohl(*((uint32_t *)(index + 24 * n)));
1784         } else {
1785                 uint32_t off;
1786                 index += 8 + p->num_objects * (20 + 4);
1787                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1788                 if (!(off & 0x80000000))
1789                         return off;
1790                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1791                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1792                                    ntohl(*((uint32_t *)(index + 4)));
1793         }
1794 }
1795
1796 off_t find_pack_entry_one(const unsigned char *sha1,
1797                                   struct packed_git *p)
1798 {
1799         const uint32_t *level1_ofs = p->index_data;
1800         const unsigned char *index = p->index_data;
1801         unsigned hi, lo, stride;
1802         static int use_lookup = -1;
1803         static int debug_lookup = -1;
1804
1805         if (debug_lookup < 0)
1806                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1807
1808         if (!index) {
1809                 if (open_pack_index(p))
1810                         return 0;
1811                 level1_ofs = p->index_data;
1812                 index = p->index_data;
1813         }
1814         if (p->index_version > 1) {
1815                 level1_ofs += 2;
1816                 index += 8;
1817         }
1818         index += 4 * 256;
1819         hi = ntohl(level1_ofs[*sha1]);
1820         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1821         if (p->index_version > 1) {
1822                 stride = 20;
1823         } else {
1824                 stride = 24;
1825                 index += 4;
1826         }
1827
1828         if (debug_lookup)
1829                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1830                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1831
1832         if (use_lookup < 0)
1833                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1834         if (use_lookup) {
1835                 int pos = sha1_entry_pos(index, stride, 0,
1836                                          lo, hi, p->num_objects, sha1);
1837                 if (pos < 0)
1838                         return 0;
1839                 return nth_packed_object_offset(p, pos);
1840         }
1841
1842         do {
1843                 unsigned mi = (lo + hi) / 2;
1844                 int cmp = hashcmp(index + mi * stride, sha1);
1845
1846                 if (debug_lookup)
1847                         printf("lo %u hi %u rg %u mi %u\n",
1848                                lo, hi, hi - lo, mi);
1849                 if (!cmp)
1850                         return nth_packed_object_offset(p, mi);
1851                 if (cmp > 0)
1852                         hi = mi;
1853                 else
1854                         lo = mi+1;
1855         } while (lo < hi);
1856         return 0;
1857 }
1858
1859 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1860 {
1861         static struct packed_git *last_found = (void *)1;
1862         struct packed_git *p;
1863         off_t offset;
1864
1865         prepare_packed_git();
1866         if (!packed_git)
1867                 return 0;
1868         p = (last_found == (void *)1) ? packed_git : last_found;
1869
1870         do {
1871                 if (p->num_bad_objects) {
1872                         unsigned i;
1873                         for (i = 0; i < p->num_bad_objects; i++)
1874                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1875                                         goto next;
1876                 }
1877
1878                 offset = find_pack_entry_one(sha1, p);
1879                 if (offset) {
1880                         /*
1881                          * We are about to tell the caller where they can
1882                          * locate the requested object.  We better make
1883                          * sure the packfile is still here and can be
1884                          * accessed before supplying that answer, as
1885                          * it may have been deleted since the index
1886                          * was loaded!
1887                          */
1888                         if (p->pack_fd == -1 && open_packed_git(p)) {
1889                                 error("packfile %s cannot be accessed", p->pack_name);
1890                                 goto next;
1891                         }
1892                         e->offset = offset;
1893                         e->p = p;
1894                         hashcpy(e->sha1, sha1);
1895                         last_found = p;
1896                         return 1;
1897                 }
1898
1899                 next:
1900                 if (p == last_found)
1901                         p = packed_git;
1902                 else
1903                         p = p->next;
1904                 if (p == last_found)
1905                         p = p->next;
1906         } while (p);
1907         return 0;
1908 }
1909
1910 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1911                                   struct packed_git *packs)
1912 {
1913         struct packed_git *p;
1914
1915         for (p = packs; p; p = p->next) {
1916                 if (find_pack_entry_one(sha1, p))
1917                         return p;
1918         }
1919         return NULL;
1920
1921 }
1922
1923 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1924 {
1925         int status;
1926         unsigned long mapsize, size;
1927         void *map;
1928         z_stream stream;
1929         char hdr[32];
1930
1931         map = map_sha1_file(sha1, &mapsize);
1932         if (!map)
1933                 return error("unable to find %s", sha1_to_hex(sha1));
1934         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1935                 status = error("unable to unpack %s header",
1936                                sha1_to_hex(sha1));
1937         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1938                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1939         else if (sizep)
1940                 *sizep = size;
1941         git_inflate_end(&stream);
1942         munmap(map, mapsize);
1943         return status;
1944 }
1945
1946 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1947 {
1948         struct pack_entry e;
1949         int status;
1950
1951         if (!find_pack_entry(sha1, &e)) {
1952                 /* Most likely it's a loose object. */
1953                 status = sha1_loose_object_info(sha1, sizep);
1954                 if (status >= 0)
1955                         return status;
1956
1957                 /* Not a loose object; someone else may have just packed it. */
1958                 reprepare_packed_git();
1959                 if (!find_pack_entry(sha1, &e))
1960                         return status;
1961         }
1962
1963         status = packed_object_info(e.p, e.offset, sizep);
1964         if (status < 0) {
1965                 mark_bad_packed_object(e.p, sha1);
1966                 status = sha1_object_info(sha1, sizep);
1967         }
1968
1969         return status;
1970 }
1971
1972 static void *read_packed_sha1(const unsigned char *sha1,
1973                               enum object_type *type, unsigned long *size)
1974 {
1975         struct pack_entry e;
1976         void *data;
1977
1978         if (!find_pack_entry(sha1, &e))
1979                 return NULL;
1980         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
1981         if (!data) {
1982                 /*
1983                  * We're probably in deep shit, but let's try to fetch
1984                  * the required object anyway from another pack or loose.
1985                  * This should happen only in the presence of a corrupted
1986                  * pack, and is better than failing outright.
1987                  */
1988                 error("failed to read object %s at offset %"PRIuMAX" from %s",
1989                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
1990                 mark_bad_packed_object(e.p, sha1);
1991                 data = read_object(sha1, type, size);
1992         }
1993         return data;
1994 }
1995
1996 /*
1997  * This is meant to hold a *small* number of objects that you would
1998  * want read_sha1_file() to be able to return, but yet you do not want
1999  * to write them into the object store (e.g. a browse-only
2000  * application).
2001  */
2002 static struct cached_object {
2003         unsigned char sha1[20];
2004         enum object_type type;
2005         void *buf;
2006         unsigned long size;
2007 } *cached_objects;
2008 static int cached_object_nr, cached_object_alloc;
2009
2010 static struct cached_object empty_tree = {
2011         EMPTY_TREE_SHA1_BIN,
2012         OBJ_TREE,
2013         "",
2014         0
2015 };
2016
2017 static struct cached_object *find_cached_object(const unsigned char *sha1)
2018 {
2019         int i;
2020         struct cached_object *co = cached_objects;
2021
2022         for (i = 0; i < cached_object_nr; i++, co++) {
2023                 if (!hashcmp(co->sha1, sha1))
2024                         return co;
2025         }
2026         if (!hashcmp(sha1, empty_tree.sha1))
2027                 return &empty_tree;
2028         return NULL;
2029 }
2030
2031 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2032                       unsigned char *sha1)
2033 {
2034         struct cached_object *co;
2035
2036         hash_sha1_file(buf, len, typename(type), sha1);
2037         if (has_sha1_file(sha1) || find_cached_object(sha1))
2038                 return 0;
2039         if (cached_object_alloc <= cached_object_nr) {
2040                 cached_object_alloc = alloc_nr(cached_object_alloc);
2041                 cached_objects = xrealloc(cached_objects,
2042                                           sizeof(*cached_objects) *
2043                                           cached_object_alloc);
2044         }
2045         co = &cached_objects[cached_object_nr++];
2046         co->size = len;
2047         co->type = type;
2048         co->buf = xmalloc(len);
2049         memcpy(co->buf, buf, len);
2050         hashcpy(co->sha1, sha1);
2051         return 0;
2052 }
2053
2054 static void *read_object(const unsigned char *sha1, enum object_type *type,
2055                          unsigned long *size)
2056 {
2057         unsigned long mapsize;
2058         void *map, *buf;
2059         struct cached_object *co;
2060
2061         co = find_cached_object(sha1);
2062         if (co) {
2063                 *type = co->type;
2064                 *size = co->size;
2065                 return xmemdupz(co->buf, co->size);
2066         }
2067
2068         buf = read_packed_sha1(sha1, type, size);
2069         if (buf)
2070                 return buf;
2071         map = map_sha1_file(sha1, &mapsize);
2072         if (map) {
2073                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2074                 munmap(map, mapsize);
2075                 return buf;
2076         }
2077         reprepare_packed_git();
2078         return read_packed_sha1(sha1, type, size);
2079 }
2080
2081 void *read_sha1_file_repl(const unsigned char *sha1,
2082                           enum object_type *type,
2083                           unsigned long *size,
2084                           const unsigned char **replacement)
2085 {
2086         const unsigned char *repl = lookup_replace_object(sha1);
2087         void *data = read_object(repl, type, size);
2088
2089         /* die if we replaced an object with one that does not exist */
2090         if (!data && repl != sha1)
2091                 die("replacement %s not found for %s",
2092                     sha1_to_hex(repl), sha1_to_hex(sha1));
2093
2094         /* legacy behavior is to die on corrupted objects */
2095         if (!data && (has_loose_object(repl) || has_packed_and_bad(repl)))
2096                 die("object %s is corrupted", sha1_to_hex(repl));
2097
2098         if (replacement)
2099                 *replacement = repl;
2100
2101         return data;
2102 }
2103
2104 void *read_object_with_reference(const unsigned char *sha1,
2105                                  const char *required_type_name,
2106                                  unsigned long *size,
2107                                  unsigned char *actual_sha1_return)
2108 {
2109         enum object_type type, required_type;
2110         void *buffer;
2111         unsigned long isize;
2112         unsigned char actual_sha1[20];
2113
2114         required_type = type_from_string(required_type_name);
2115         hashcpy(actual_sha1, sha1);
2116         while (1) {
2117                 int ref_length = -1;
2118                 const char *ref_type = NULL;
2119
2120                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2121                 if (!buffer)
2122                         return NULL;
2123                 if (type == required_type) {
2124                         *size = isize;
2125                         if (actual_sha1_return)
2126                                 hashcpy(actual_sha1_return, actual_sha1);
2127                         return buffer;
2128                 }
2129                 /* Handle references */
2130                 else if (type == OBJ_COMMIT)
2131                         ref_type = "tree ";
2132                 else if (type == OBJ_TAG)
2133                         ref_type = "object ";
2134                 else {
2135                         free(buffer);
2136                         return NULL;
2137                 }
2138                 ref_length = strlen(ref_type);
2139
2140                 if (ref_length + 40 > isize ||
2141                     memcmp(buffer, ref_type, ref_length) ||
2142                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2143                         free(buffer);
2144                         return NULL;
2145                 }
2146                 free(buffer);
2147                 /* Now we have the ID of the referred-to object in
2148                  * actual_sha1.  Check again. */
2149         }
2150 }
2151
2152 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2153                                     const char *type, unsigned char *sha1,
2154                                     char *hdr, int *hdrlen)
2155 {
2156         git_SHA_CTX c;
2157
2158         /* Generate the header */
2159         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2160
2161         /* Sha1.. */
2162         git_SHA1_Init(&c);
2163         git_SHA1_Update(&c, hdr, *hdrlen);
2164         git_SHA1_Update(&c, buf, len);
2165         git_SHA1_Final(sha1, &c);
2166 }
2167
2168 /*
2169  * Move the just written object into its final resting place.
2170  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2171  * "moving" is only a part of what it does, when no patch between
2172  * master to pu changes the call sites of this function.
2173  */
2174 int move_temp_to_file(const char *tmpfile, const char *filename)
2175 {
2176         int ret = 0;
2177
2178         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2179                 goto try_rename;
2180         else if (link(tmpfile, filename))
2181                 ret = errno;
2182
2183         /*
2184          * Coda hack - coda doesn't like cross-directory links,
2185          * so we fall back to a rename, which will mean that it
2186          * won't be able to check collisions, but that's not a
2187          * big deal.
2188          *
2189          * The same holds for FAT formatted media.
2190          *
2191          * When this succeeds, we just return.  We have nothing
2192          * left to unlink.
2193          */
2194         if (ret && ret != EEXIST) {
2195         try_rename:
2196                 if (!rename(tmpfile, filename))
2197                         goto out;
2198                 ret = errno;
2199         }
2200         unlink_or_warn(tmpfile);
2201         if (ret) {
2202                 if (ret != EEXIST) {
2203                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2204                 }
2205                 /* FIXME!!! Collision check here ? */
2206         }
2207
2208 out:
2209         if (adjust_shared_perm(filename))
2210                 return error("unable to set permission to '%s'", filename);
2211         return 0;
2212 }
2213
2214 static int write_buffer(int fd, const void *buf, size_t len)
2215 {
2216         if (write_in_full(fd, buf, len) < 0)
2217                 return error("file write error (%s)", strerror(errno));
2218         return 0;
2219 }
2220
2221 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2222                    unsigned char *sha1)
2223 {
2224         char hdr[32];
2225         int hdrlen;
2226         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2227         return 0;
2228 }
2229
2230 /* Finalize a file on disk, and close it. */
2231 static void close_sha1_file(int fd)
2232 {
2233         if (fsync_object_files)
2234                 fsync_or_die(fd, "sha1 file");
2235         if (close(fd) != 0)
2236                 die_errno("error when closing sha1 file");
2237 }
2238
2239 /* Size of directory component, including the ending '/' */
2240 static inline int directory_size(const char *filename)
2241 {
2242         const char *s = strrchr(filename, '/');
2243         if (!s)
2244                 return 0;
2245         return s - filename + 1;
2246 }
2247
2248 /*
2249  * This creates a temporary file in the same directory as the final
2250  * 'filename'
2251  *
2252  * We want to avoid cross-directory filename renames, because those
2253  * can have problems on various filesystems (FAT, NFS, Coda).
2254  */
2255 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2256 {
2257         int fd, dirlen = directory_size(filename);
2258
2259         if (dirlen + 20 > bufsiz) {
2260                 errno = ENAMETOOLONG;
2261                 return -1;
2262         }
2263         memcpy(buffer, filename, dirlen);
2264         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2265         fd = git_mkstemp_mode(buffer, 0444);
2266         if (fd < 0 && dirlen && errno == ENOENT) {
2267                 /* Make sure the directory exists */
2268                 memcpy(buffer, filename, dirlen);
2269                 buffer[dirlen-1] = 0;
2270                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2271                         return -1;
2272
2273                 /* Try again */
2274                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2275                 fd = git_mkstemp_mode(buffer, 0444);
2276         }
2277         return fd;
2278 }
2279
2280 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2281                               void *buf, unsigned long len, time_t mtime)
2282 {
2283         int fd, ret;
2284         unsigned char compressed[4096];
2285         z_stream stream;
2286         git_SHA_CTX c;
2287         unsigned char parano_sha1[20];
2288         char *filename;
2289         static char tmpfile[PATH_MAX];
2290
2291         filename = sha1_file_name(sha1);
2292         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2293         while (fd < 0 && errno == EMFILE && unuse_one_window(packed_git, -1))
2294                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2295         if (fd < 0) {
2296                 if (errno == EACCES)
2297                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2298                 else
2299                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2300         }
2301
2302         /* Set it up */
2303         memset(&stream, 0, sizeof(stream));
2304         deflateInit(&stream, zlib_compression_level);
2305         stream.next_out = compressed;
2306         stream.avail_out = sizeof(compressed);
2307         git_SHA1_Init(&c);
2308
2309         /* First header.. */
2310         stream.next_in = (unsigned char *)hdr;
2311         stream.avail_in = hdrlen;
2312         while (deflate(&stream, 0) == Z_OK)
2313                 /* nothing */;
2314         git_SHA1_Update(&c, hdr, hdrlen);
2315
2316         /* Then the data itself.. */
2317         stream.next_in = buf;
2318         stream.avail_in = len;
2319         do {
2320                 unsigned char *in0 = stream.next_in;
2321                 ret = deflate(&stream, Z_FINISH);
2322                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2323                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2324                         die("unable to write sha1 file");
2325                 stream.next_out = compressed;
2326                 stream.avail_out = sizeof(compressed);
2327         } while (ret == Z_OK);
2328
2329         if (ret != Z_STREAM_END)
2330                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2331         ret = deflateEnd(&stream);
2332         if (ret != Z_OK)
2333                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2334         git_SHA1_Final(parano_sha1, &c);
2335         if (hashcmp(sha1, parano_sha1) != 0)
2336                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2337
2338         close_sha1_file(fd);
2339
2340         if (mtime) {
2341                 struct utimbuf utb;
2342                 utb.actime = mtime;
2343                 utb.modtime = mtime;
2344                 if (utime(tmpfile, &utb) < 0)
2345                         warning("failed utime() on %s: %s",
2346                                 tmpfile, strerror(errno));
2347         }
2348
2349         return move_temp_to_file(tmpfile, filename);
2350 }
2351
2352 int write_sha1_file(void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2353 {
2354         unsigned char sha1[20];
2355         char hdr[32];
2356         int hdrlen;
2357
2358         /* Normally if we have it in the pack then we do not bother writing
2359          * it out into .git/objects/??/?{38} file.
2360          */
2361         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2362         if (returnsha1)
2363                 hashcpy(returnsha1, sha1);
2364         if (has_sha1_file(sha1))
2365                 return 0;
2366         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2367 }
2368
2369 int force_object_loose(const unsigned char *sha1, time_t mtime)
2370 {
2371         void *buf;
2372         unsigned long len;
2373         enum object_type type;
2374         char hdr[32];
2375         int hdrlen;
2376         int ret;
2377
2378         if (has_loose_object(sha1))
2379                 return 0;
2380         buf = read_packed_sha1(sha1, &type, &len);
2381         if (!buf)
2382                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2383         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2384         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2385         free(buf);
2386
2387         return ret;
2388 }
2389
2390 int has_pack_index(const unsigned char *sha1)
2391 {
2392         struct stat st;
2393         if (stat(sha1_pack_index_name(sha1), &st))
2394                 return 0;
2395         return 1;
2396 }
2397
2398 int has_sha1_pack(const unsigned char *sha1)
2399 {
2400         struct pack_entry e;
2401         return find_pack_entry(sha1, &e);
2402 }
2403
2404 int has_sha1_file(const unsigned char *sha1)
2405 {
2406         struct pack_entry e;
2407
2408         if (find_pack_entry(sha1, &e))
2409                 return 1;
2410         return has_loose_object(sha1);
2411 }
2412
2413 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2414                      int write_object, enum object_type type, const char *path)
2415 {
2416         int ret, re_allocated = 0;
2417
2418         if (!type)
2419                 type = OBJ_BLOB;
2420
2421         /*
2422          * Convert blobs to git internal format
2423          */
2424         if ((type == OBJ_BLOB) && path) {
2425                 struct strbuf nbuf = STRBUF_INIT;
2426                 if (convert_to_git(path, buf, size, &nbuf,
2427                                    write_object ? safe_crlf : 0)) {
2428                         buf = strbuf_detach(&nbuf, &size);
2429                         re_allocated = 1;
2430                 }
2431         }
2432
2433         if (write_object)
2434                 ret = write_sha1_file(buf, size, typename(type), sha1);
2435         else
2436                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2437         if (re_allocated)
2438                 free(buf);
2439         return ret;
2440 }
2441
2442 #define SMALL_FILE_SIZE (32*1024)
2443
2444 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2445              enum object_type type, const char *path)
2446 {
2447         int ret;
2448         size_t size = xsize_t(st->st_size);
2449
2450         if (!S_ISREG(st->st_mode)) {
2451                 struct strbuf sbuf = STRBUF_INIT;
2452                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2453                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2454                                         type, path);
2455                 else
2456                         ret = -1;
2457                 strbuf_release(&sbuf);
2458         } else if (size <= SMALL_FILE_SIZE) {
2459                 char *buf = xmalloc(size);
2460                 if (size == read_in_full(fd, buf, size))
2461                         ret = index_mem(sha1, buf, size, write_object, type,
2462                                         path);
2463                 else
2464                         ret = error("short read %s", strerror(errno));
2465                 free(buf);
2466         } else if (size) {
2467                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2468                 ret = index_mem(sha1, buf, size, write_object, type, path);
2469                 munmap(buf, size);
2470         } else
2471                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2472         close(fd);
2473         return ret;
2474 }
2475
2476 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2477 {
2478         int fd;
2479         struct strbuf sb = STRBUF_INIT;
2480
2481         switch (st->st_mode & S_IFMT) {
2482         case S_IFREG:
2483                 fd = open(path, O_RDONLY);
2484                 if (fd < 0)
2485                         return error("open(\"%s\"): %s", path,
2486                                      strerror(errno));
2487                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2488                         return error("%s: failed to insert into database",
2489                                      path);
2490                 break;
2491         case S_IFLNK:
2492                 if (strbuf_readlink(&sb, path, st->st_size)) {
2493                         char *errstr = strerror(errno);
2494                         return error("readlink(\"%s\"): %s", path,
2495                                      errstr);
2496                 }
2497                 if (!write_object)
2498                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2499                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2500                         return error("%s: failed to insert into database",
2501                                      path);
2502                 strbuf_release(&sb);
2503                 break;
2504         case S_IFDIR:
2505                 return resolve_gitlink_ref(path, "HEAD", sha1);
2506         default:
2507                 return error("%s: unsupported file type", path);
2508         }
2509         return 0;
2510 }
2511
2512 int read_pack_header(int fd, struct pack_header *header)
2513 {
2514         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2515                 /* "eof before pack header was fully read" */
2516                 return PH_ERROR_EOF;
2517
2518         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2519                 /* "protocol error (pack signature mismatch detected)" */
2520                 return PH_ERROR_PACK_SIGNATURE;
2521         if (!pack_version_ok(header->hdr_version))
2522                 /* "protocol error (pack version unsupported)" */
2523                 return PH_ERROR_PROTOCOL;
2524         return 0;
2525 }