git diff: Don't test submodule dirtiness with --ignore-submodules
[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 = xmalloc(1+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         buf[size] = 0;
1198         if (status == Z_STREAM_END && !stream->avail_in) {
1199                 git_inflate_end(stream);
1200                 return buf;
1201         }
1202
1203         if (status < 0)
1204                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1205         else if (stream->avail_in)
1206                 error("garbage at end of loose object '%s'",
1207                       sha1_to_hex(sha1));
1208         free(buf);
1209         return NULL;
1210 }
1211
1212 /*
1213  * We used to just use "sscanf()", but that's actually way
1214  * too permissive for what we want to check. So do an anal
1215  * object header parse by hand.
1216  */
1217 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1218 {
1219         char type[10];
1220         int i;
1221         unsigned long size;
1222
1223         /*
1224          * The type can be at most ten bytes (including the
1225          * terminating '\0' that we add), and is followed by
1226          * a space.
1227          */
1228         i = 0;
1229         for (;;) {
1230                 char c = *hdr++;
1231                 if (c == ' ')
1232                         break;
1233                 type[i++] = c;
1234                 if (i >= sizeof(type))
1235                         return -1;
1236         }
1237         type[i] = 0;
1238
1239         /*
1240          * The length must follow immediately, and be in canonical
1241          * decimal format (ie "010" is not valid).
1242          */
1243         size = *hdr++ - '0';
1244         if (size > 9)
1245                 return -1;
1246         if (size) {
1247                 for (;;) {
1248                         unsigned long c = *hdr - '0';
1249                         if (c > 9)
1250                                 break;
1251                         hdr++;
1252                         size = size * 10 + c;
1253                 }
1254         }
1255         *sizep = size;
1256
1257         /*
1258          * The length must be followed by a zero byte
1259          */
1260         return *hdr ? -1 : type_from_string(type);
1261 }
1262
1263 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1264 {
1265         int ret;
1266         z_stream stream;
1267         char hdr[8192];
1268
1269         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1270         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1271                 return NULL;
1272
1273         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1274 }
1275
1276 unsigned long get_size_from_delta(struct packed_git *p,
1277                                   struct pack_window **w_curs,
1278                                   off_t curpos)
1279 {
1280         const unsigned char *data;
1281         unsigned char delta_head[20], *in;
1282         z_stream stream;
1283         int st;
1284
1285         memset(&stream, 0, sizeof(stream));
1286         stream.next_out = delta_head;
1287         stream.avail_out = sizeof(delta_head);
1288
1289         git_inflate_init(&stream);
1290         do {
1291                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1292                 stream.next_in = in;
1293                 st = git_inflate(&stream, Z_FINISH);
1294                 curpos += stream.next_in - in;
1295         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1296                  stream.total_out < sizeof(delta_head));
1297         git_inflate_end(&stream);
1298         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1299                 error("delta data unpack-initial failed");
1300                 return 0;
1301         }
1302
1303         /* Examine the initial part of the delta to figure out
1304          * the result size.
1305          */
1306         data = delta_head;
1307
1308         /* ignore base size */
1309         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1310
1311         /* Read the result size */
1312         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1313 }
1314
1315 static off_t get_delta_base(struct packed_git *p,
1316                                     struct pack_window **w_curs,
1317                                     off_t *curpos,
1318                                     enum object_type type,
1319                                     off_t delta_obj_offset)
1320 {
1321         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1322         off_t base_offset;
1323
1324         /* use_pack() assured us we have [base_info, base_info + 20)
1325          * as a range that we can look at without walking off the
1326          * end of the mapped window.  Its actually the hash size
1327          * that is assured.  An OFS_DELTA longer than the hash size
1328          * is stupid, as then a REF_DELTA would be smaller to store.
1329          */
1330         if (type == OBJ_OFS_DELTA) {
1331                 unsigned used = 0;
1332                 unsigned char c = base_info[used++];
1333                 base_offset = c & 127;
1334                 while (c & 128) {
1335                         base_offset += 1;
1336                         if (!base_offset || MSB(base_offset, 7))
1337                                 return 0;  /* overflow */
1338                         c = base_info[used++];
1339                         base_offset = (base_offset << 7) + (c & 127);
1340                 }
1341                 base_offset = delta_obj_offset - base_offset;
1342                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1343                         return 0;  /* out of bound */
1344                 *curpos += used;
1345         } else if (type == OBJ_REF_DELTA) {
1346                 /* The base entry _must_ be in the same pack */
1347                 base_offset = find_pack_entry_one(base_info, p);
1348                 *curpos += 20;
1349         } else
1350                 die("I am totally screwed");
1351         return base_offset;
1352 }
1353
1354 /* forward declaration for a mutually recursive function */
1355 static int packed_object_info(struct packed_git *p, off_t offset,
1356                               unsigned long *sizep);
1357
1358 static int packed_delta_info(struct packed_git *p,
1359                              struct pack_window **w_curs,
1360                              off_t curpos,
1361                              enum object_type type,
1362                              off_t obj_offset,
1363                              unsigned long *sizep)
1364 {
1365         off_t base_offset;
1366
1367         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1368         if (!base_offset)
1369                 return OBJ_BAD;
1370         type = packed_object_info(p, base_offset, NULL);
1371         if (type <= OBJ_NONE) {
1372                 struct revindex_entry *revidx;
1373                 const unsigned char *base_sha1;
1374                 revidx = find_pack_revindex(p, base_offset);
1375                 if (!revidx)
1376                         return OBJ_BAD;
1377                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1378                 mark_bad_packed_object(p, base_sha1);
1379                 type = sha1_object_info(base_sha1, NULL);
1380                 if (type <= OBJ_NONE)
1381                         return OBJ_BAD;
1382         }
1383
1384         /* We choose to only get the type of the base object and
1385          * ignore potentially corrupt pack file that expects the delta
1386          * based on a base with a wrong size.  This saves tons of
1387          * inflate() calls.
1388          */
1389         if (sizep) {
1390                 *sizep = get_size_from_delta(p, w_curs, curpos);
1391                 if (*sizep == 0)
1392                         type = OBJ_BAD;
1393         }
1394
1395         return type;
1396 }
1397
1398 static int unpack_object_header(struct packed_git *p,
1399                                 struct pack_window **w_curs,
1400                                 off_t *curpos,
1401                                 unsigned long *sizep)
1402 {
1403         unsigned char *base;
1404         unsigned int left;
1405         unsigned long used;
1406         enum object_type type;
1407
1408         /* use_pack() assures us we have [base, base + 20) available
1409          * as a range that we can look at at.  (Its actually the hash
1410          * size that is assured.)  With our object header encoding
1411          * the maximum deflated object size is 2^137, which is just
1412          * insane, so we know won't exceed what we have been given.
1413          */
1414         base = use_pack(p, w_curs, *curpos, &left);
1415         used = unpack_object_header_buffer(base, left, &type, sizep);
1416         if (!used) {
1417                 type = OBJ_BAD;
1418         } else
1419                 *curpos += used;
1420
1421         return type;
1422 }
1423
1424 const char *packed_object_info_detail(struct packed_git *p,
1425                                       off_t obj_offset,
1426                                       unsigned long *size,
1427                                       unsigned long *store_size,
1428                                       unsigned int *delta_chain_length,
1429                                       unsigned char *base_sha1)
1430 {
1431         struct pack_window *w_curs = NULL;
1432         off_t curpos;
1433         unsigned long dummy;
1434         unsigned char *next_sha1;
1435         enum object_type type;
1436         struct revindex_entry *revidx;
1437
1438         *delta_chain_length = 0;
1439         curpos = obj_offset;
1440         type = unpack_object_header(p, &w_curs, &curpos, size);
1441
1442         revidx = find_pack_revindex(p, obj_offset);
1443         *store_size = revidx[1].offset - obj_offset;
1444
1445         for (;;) {
1446                 switch (type) {
1447                 default:
1448                         die("pack %s contains unknown object type %d",
1449                             p->pack_name, type);
1450                 case OBJ_COMMIT:
1451                 case OBJ_TREE:
1452                 case OBJ_BLOB:
1453                 case OBJ_TAG:
1454                         unuse_pack(&w_curs);
1455                         return typename(type);
1456                 case OBJ_OFS_DELTA:
1457                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1458                         if (!obj_offset)
1459                                 die("pack %s contains bad delta base reference of type %s",
1460                                     p->pack_name, typename(type));
1461                         if (*delta_chain_length == 0) {
1462                                 revidx = find_pack_revindex(p, obj_offset);
1463                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1464                         }
1465                         break;
1466                 case OBJ_REF_DELTA:
1467                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1468                         if (*delta_chain_length == 0)
1469                                 hashcpy(base_sha1, next_sha1);
1470                         obj_offset = find_pack_entry_one(next_sha1, p);
1471                         break;
1472                 }
1473                 (*delta_chain_length)++;
1474                 curpos = obj_offset;
1475                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1476         }
1477 }
1478
1479 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1480                               unsigned long *sizep)
1481 {
1482         struct pack_window *w_curs = NULL;
1483         unsigned long size;
1484         off_t curpos = obj_offset;
1485         enum object_type type;
1486
1487         type = unpack_object_header(p, &w_curs, &curpos, &size);
1488
1489         switch (type) {
1490         case OBJ_OFS_DELTA:
1491         case OBJ_REF_DELTA:
1492                 type = packed_delta_info(p, &w_curs, curpos,
1493                                          type, obj_offset, sizep);
1494                 break;
1495         case OBJ_COMMIT:
1496         case OBJ_TREE:
1497         case OBJ_BLOB:
1498         case OBJ_TAG:
1499                 if (sizep)
1500                         *sizep = size;
1501                 break;
1502         default:
1503                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1504                       type, (uintmax_t)obj_offset, p->pack_name);
1505                 type = OBJ_BAD;
1506         }
1507         unuse_pack(&w_curs);
1508         return type;
1509 }
1510
1511 static void *unpack_compressed_entry(struct packed_git *p,
1512                                     struct pack_window **w_curs,
1513                                     off_t curpos,
1514                                     unsigned long size)
1515 {
1516         int st;
1517         z_stream stream;
1518         unsigned char *buffer, *in;
1519
1520         buffer = xmalloc(size + 1);
1521         buffer[size] = 0;
1522         memset(&stream, 0, sizeof(stream));
1523         stream.next_out = buffer;
1524         stream.avail_out = size + 1;
1525
1526         git_inflate_init(&stream);
1527         do {
1528                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1529                 stream.next_in = in;
1530                 st = git_inflate(&stream, Z_FINISH);
1531                 if (!stream.avail_out)
1532                         break; /* the payload is larger than it should be */
1533                 curpos += stream.next_in - in;
1534         } while (st == Z_OK || st == Z_BUF_ERROR);
1535         git_inflate_end(&stream);
1536         if ((st != Z_STREAM_END) || stream.total_out != size) {
1537                 free(buffer);
1538                 return NULL;
1539         }
1540
1541         return buffer;
1542 }
1543
1544 #define MAX_DELTA_CACHE (256)
1545
1546 static size_t delta_base_cached;
1547
1548 static struct delta_base_cache_lru_list {
1549         struct delta_base_cache_lru_list *prev;
1550         struct delta_base_cache_lru_list *next;
1551 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1552
1553 static struct delta_base_cache_entry {
1554         struct delta_base_cache_lru_list lru;
1555         void *data;
1556         struct packed_git *p;
1557         off_t base_offset;
1558         unsigned long size;
1559         enum object_type type;
1560 } delta_base_cache[MAX_DELTA_CACHE];
1561
1562 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1563 {
1564         unsigned long hash;
1565
1566         hash = (unsigned long)p + (unsigned long)base_offset;
1567         hash += (hash >> 8) + (hash >> 16);
1568         return hash % MAX_DELTA_CACHE;
1569 }
1570
1571 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1572         unsigned long *base_size, enum object_type *type, int keep_cache)
1573 {
1574         void *ret;
1575         unsigned long hash = pack_entry_hash(p, base_offset);
1576         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1577
1578         ret = ent->data;
1579         if (!ret || ent->p != p || ent->base_offset != base_offset)
1580                 return unpack_entry(p, base_offset, type, base_size);
1581
1582         if (!keep_cache) {
1583                 ent->data = NULL;
1584                 ent->lru.next->prev = ent->lru.prev;
1585                 ent->lru.prev->next = ent->lru.next;
1586                 delta_base_cached -= ent->size;
1587         } else {
1588                 ret = xmemdupz(ent->data, ent->size);
1589         }
1590         *type = ent->type;
1591         *base_size = ent->size;
1592         return ret;
1593 }
1594
1595 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1596 {
1597         if (ent->data) {
1598                 free(ent->data);
1599                 ent->data = NULL;
1600                 ent->lru.next->prev = ent->lru.prev;
1601                 ent->lru.prev->next = ent->lru.next;
1602                 delta_base_cached -= ent->size;
1603         }
1604 }
1605
1606 void clear_delta_base_cache(void)
1607 {
1608         unsigned long p;
1609         for (p = 0; p < MAX_DELTA_CACHE; p++)
1610                 release_delta_base_cache(&delta_base_cache[p]);
1611 }
1612
1613 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1614         void *base, unsigned long base_size, enum object_type type)
1615 {
1616         unsigned long hash = pack_entry_hash(p, base_offset);
1617         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1618         struct delta_base_cache_lru_list *lru;
1619
1620         release_delta_base_cache(ent);
1621         delta_base_cached += base_size;
1622
1623         for (lru = delta_base_cache_lru.next;
1624              delta_base_cached > delta_base_cache_limit
1625              && lru != &delta_base_cache_lru;
1626              lru = lru->next) {
1627                 struct delta_base_cache_entry *f = (void *)lru;
1628                 if (f->type == OBJ_BLOB)
1629                         release_delta_base_cache(f);
1630         }
1631         for (lru = delta_base_cache_lru.next;
1632              delta_base_cached > delta_base_cache_limit
1633              && lru != &delta_base_cache_lru;
1634              lru = lru->next) {
1635                 struct delta_base_cache_entry *f = (void *)lru;
1636                 release_delta_base_cache(f);
1637         }
1638
1639         ent->p = p;
1640         ent->base_offset = base_offset;
1641         ent->type = type;
1642         ent->data = base;
1643         ent->size = base_size;
1644         ent->lru.next = &delta_base_cache_lru;
1645         ent->lru.prev = delta_base_cache_lru.prev;
1646         delta_base_cache_lru.prev->next = &ent->lru;
1647         delta_base_cache_lru.prev = &ent->lru;
1648 }
1649
1650 static void *read_object(const unsigned char *sha1, enum object_type *type,
1651                          unsigned long *size);
1652
1653 static void *unpack_delta_entry(struct packed_git *p,
1654                                 struct pack_window **w_curs,
1655                                 off_t curpos,
1656                                 unsigned long delta_size,
1657                                 off_t obj_offset,
1658                                 enum object_type *type,
1659                                 unsigned long *sizep)
1660 {
1661         void *delta_data, *result, *base;
1662         unsigned long base_size;
1663         off_t base_offset;
1664
1665         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1666         if (!base_offset) {
1667                 error("failed to validate delta base reference "
1668                       "at offset %"PRIuMAX" from %s",
1669                       (uintmax_t)curpos, p->pack_name);
1670                 return NULL;
1671         }
1672         unuse_pack(w_curs);
1673         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1674         if (!base) {
1675                 /*
1676                  * We're probably in deep shit, but let's try to fetch
1677                  * the required base anyway from another pack or loose.
1678                  * This is costly but should happen only in the presence
1679                  * of a corrupted pack, and is better than failing outright.
1680                  */
1681                 struct revindex_entry *revidx;
1682                 const unsigned char *base_sha1;
1683                 revidx = find_pack_revindex(p, base_offset);
1684                 if (!revidx)
1685                         return NULL;
1686                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1687                 error("failed to read delta base object %s"
1688                       " at offset %"PRIuMAX" from %s",
1689                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1690                       p->pack_name);
1691                 mark_bad_packed_object(p, base_sha1);
1692                 base = read_object(base_sha1, type, &base_size);
1693                 if (!base)
1694                         return NULL;
1695         }
1696
1697         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1698         if (!delta_data) {
1699                 error("failed to unpack compressed delta "
1700                       "at offset %"PRIuMAX" from %s",
1701                       (uintmax_t)curpos, p->pack_name);
1702                 free(base);
1703                 return NULL;
1704         }
1705         result = patch_delta(base, base_size,
1706                              delta_data, delta_size,
1707                              sizep);
1708         if (!result)
1709                 die("failed to apply delta");
1710         free(delta_data);
1711         add_delta_base_cache(p, base_offset, base, base_size, *type);
1712         return result;
1713 }
1714
1715 int do_check_packed_object_crc;
1716
1717 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1718                    enum object_type *type, unsigned long *sizep)
1719 {
1720         struct pack_window *w_curs = NULL;
1721         off_t curpos = obj_offset;
1722         void *data;
1723
1724         if (do_check_packed_object_crc && p->index_version > 1) {
1725                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1726                 unsigned long len = revidx[1].offset - obj_offset;
1727                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1728                         const unsigned char *sha1 =
1729                                 nth_packed_object_sha1(p, revidx->nr);
1730                         error("bad packed object CRC for %s",
1731                               sha1_to_hex(sha1));
1732                         mark_bad_packed_object(p, sha1);
1733                         unuse_pack(&w_curs);
1734                         return NULL;
1735                 }
1736         }
1737
1738         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1739         switch (*type) {
1740         case OBJ_OFS_DELTA:
1741         case OBJ_REF_DELTA:
1742                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1743                                           obj_offset, type, sizep);
1744                 break;
1745         case OBJ_COMMIT:
1746         case OBJ_TREE:
1747         case OBJ_BLOB:
1748         case OBJ_TAG:
1749                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1750                 break;
1751         default:
1752                 data = NULL;
1753                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1754                       *type, (uintmax_t)obj_offset, p->pack_name);
1755         }
1756         unuse_pack(&w_curs);
1757         return data;
1758 }
1759
1760 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1761                                             uint32_t n)
1762 {
1763         const unsigned char *index = p->index_data;
1764         if (!index) {
1765                 if (open_pack_index(p))
1766                         return NULL;
1767                 index = p->index_data;
1768         }
1769         if (n >= p->num_objects)
1770                 return NULL;
1771         index += 4 * 256;
1772         if (p->index_version == 1) {
1773                 return index + 24 * n + 4;
1774         } else {
1775                 index += 8;
1776                 return index + 20 * n;
1777         }
1778 }
1779
1780 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1781 {
1782         const unsigned char *index = p->index_data;
1783         index += 4 * 256;
1784         if (p->index_version == 1) {
1785                 return ntohl(*((uint32_t *)(index + 24 * n)));
1786         } else {
1787                 uint32_t off;
1788                 index += 8 + p->num_objects * (20 + 4);
1789                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1790                 if (!(off & 0x80000000))
1791                         return off;
1792                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1793                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1794                                    ntohl(*((uint32_t *)(index + 4)));
1795         }
1796 }
1797
1798 off_t find_pack_entry_one(const unsigned char *sha1,
1799                                   struct packed_git *p)
1800 {
1801         const uint32_t *level1_ofs = p->index_data;
1802         const unsigned char *index = p->index_data;
1803         unsigned hi, lo, stride;
1804         static int use_lookup = -1;
1805         static int debug_lookup = -1;
1806
1807         if (debug_lookup < 0)
1808                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1809
1810         if (!index) {
1811                 if (open_pack_index(p))
1812                         return 0;
1813                 level1_ofs = p->index_data;
1814                 index = p->index_data;
1815         }
1816         if (p->index_version > 1) {
1817                 level1_ofs += 2;
1818                 index += 8;
1819         }
1820         index += 4 * 256;
1821         hi = ntohl(level1_ofs[*sha1]);
1822         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1823         if (p->index_version > 1) {
1824                 stride = 20;
1825         } else {
1826                 stride = 24;
1827                 index += 4;
1828         }
1829
1830         if (debug_lookup)
1831                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1832                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1833
1834         if (use_lookup < 0)
1835                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1836         if (use_lookup) {
1837                 int pos = sha1_entry_pos(index, stride, 0,
1838                                          lo, hi, p->num_objects, sha1);
1839                 if (pos < 0)
1840                         return 0;
1841                 return nth_packed_object_offset(p, pos);
1842         }
1843
1844         do {
1845                 unsigned mi = (lo + hi) / 2;
1846                 int cmp = hashcmp(index + mi * stride, sha1);
1847
1848                 if (debug_lookup)
1849                         printf("lo %u hi %u rg %u mi %u\n",
1850                                lo, hi, hi - lo, mi);
1851                 if (!cmp)
1852                         return nth_packed_object_offset(p, mi);
1853                 if (cmp > 0)
1854                         hi = mi;
1855                 else
1856                         lo = mi+1;
1857         } while (lo < hi);
1858         return 0;
1859 }
1860
1861 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1862 {
1863         static struct packed_git *last_found = (void *)1;
1864         struct packed_git *p;
1865         off_t offset;
1866
1867         prepare_packed_git();
1868         if (!packed_git)
1869                 return 0;
1870         p = (last_found == (void *)1) ? packed_git : last_found;
1871
1872         do {
1873                 if (p->num_bad_objects) {
1874                         unsigned i;
1875                         for (i = 0; i < p->num_bad_objects; i++)
1876                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1877                                         goto next;
1878                 }
1879
1880                 offset = find_pack_entry_one(sha1, p);
1881                 if (offset) {
1882                         /*
1883                          * We are about to tell the caller where they can
1884                          * locate the requested object.  We better make
1885                          * sure the packfile is still here and can be
1886                          * accessed before supplying that answer, as
1887                          * it may have been deleted since the index
1888                          * was loaded!
1889                          */
1890                         if (p->pack_fd == -1 && open_packed_git(p)) {
1891                                 error("packfile %s cannot be accessed", p->pack_name);
1892                                 goto next;
1893                         }
1894                         e->offset = offset;
1895                         e->p = p;
1896                         hashcpy(e->sha1, sha1);
1897                         last_found = p;
1898                         return 1;
1899                 }
1900
1901                 next:
1902                 if (p == last_found)
1903                         p = packed_git;
1904                 else
1905                         p = p->next;
1906                 if (p == last_found)
1907                         p = p->next;
1908         } while (p);
1909         return 0;
1910 }
1911
1912 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1913                                   struct packed_git *packs)
1914 {
1915         struct packed_git *p;
1916
1917         for (p = packs; p; p = p->next) {
1918                 if (find_pack_entry_one(sha1, p))
1919                         return p;
1920         }
1921         return NULL;
1922
1923 }
1924
1925 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1926 {
1927         int status;
1928         unsigned long mapsize, size;
1929         void *map;
1930         z_stream stream;
1931         char hdr[32];
1932
1933         map = map_sha1_file(sha1, &mapsize);
1934         if (!map)
1935                 return error("unable to find %s", sha1_to_hex(sha1));
1936         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1937                 status = error("unable to unpack %s header",
1938                                sha1_to_hex(sha1));
1939         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1940                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1941         else if (sizep)
1942                 *sizep = size;
1943         git_inflate_end(&stream);
1944         munmap(map, mapsize);
1945         return status;
1946 }
1947
1948 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1949 {
1950         struct pack_entry e;
1951         int status;
1952
1953         if (!find_pack_entry(sha1, &e)) {
1954                 /* Most likely it's a loose object. */
1955                 status = sha1_loose_object_info(sha1, sizep);
1956                 if (status >= 0)
1957                         return status;
1958
1959                 /* Not a loose object; someone else may have just packed it. */
1960                 reprepare_packed_git();
1961                 if (!find_pack_entry(sha1, &e))
1962                         return status;
1963         }
1964
1965         status = packed_object_info(e.p, e.offset, sizep);
1966         if (status < 0) {
1967                 mark_bad_packed_object(e.p, sha1);
1968                 status = sha1_object_info(sha1, sizep);
1969         }
1970
1971         return status;
1972 }
1973
1974 static void *read_packed_sha1(const unsigned char *sha1,
1975                               enum object_type *type, unsigned long *size)
1976 {
1977         struct pack_entry e;
1978         void *data;
1979
1980         if (!find_pack_entry(sha1, &e))
1981                 return NULL;
1982         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
1983         if (!data) {
1984                 /*
1985                  * We're probably in deep shit, but let's try to fetch
1986                  * the required object anyway from another pack or loose.
1987                  * This should happen only in the presence of a corrupted
1988                  * pack, and is better than failing outright.
1989                  */
1990                 error("failed to read object %s at offset %"PRIuMAX" from %s",
1991                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
1992                 mark_bad_packed_object(e.p, sha1);
1993                 data = read_object(sha1, type, size);
1994         }
1995         return data;
1996 }
1997
1998 /*
1999  * This is meant to hold a *small* number of objects that you would
2000  * want read_sha1_file() to be able to return, but yet you do not want
2001  * to write them into the object store (e.g. a browse-only
2002  * application).
2003  */
2004 static struct cached_object {
2005         unsigned char sha1[20];
2006         enum object_type type;
2007         void *buf;
2008         unsigned long size;
2009 } *cached_objects;
2010 static int cached_object_nr, cached_object_alloc;
2011
2012 static struct cached_object empty_tree = {
2013         EMPTY_TREE_SHA1_BIN,
2014         OBJ_TREE,
2015         "",
2016         0
2017 };
2018
2019 static struct cached_object *find_cached_object(const unsigned char *sha1)
2020 {
2021         int i;
2022         struct cached_object *co = cached_objects;
2023
2024         for (i = 0; i < cached_object_nr; i++, co++) {
2025                 if (!hashcmp(co->sha1, sha1))
2026                         return co;
2027         }
2028         if (!hashcmp(sha1, empty_tree.sha1))
2029                 return &empty_tree;
2030         return NULL;
2031 }
2032
2033 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2034                       unsigned char *sha1)
2035 {
2036         struct cached_object *co;
2037
2038         hash_sha1_file(buf, len, typename(type), sha1);
2039         if (has_sha1_file(sha1) || find_cached_object(sha1))
2040                 return 0;
2041         if (cached_object_alloc <= cached_object_nr) {
2042                 cached_object_alloc = alloc_nr(cached_object_alloc);
2043                 cached_objects = xrealloc(cached_objects,
2044                                           sizeof(*cached_objects) *
2045                                           cached_object_alloc);
2046         }
2047         co = &cached_objects[cached_object_nr++];
2048         co->size = len;
2049         co->type = type;
2050         co->buf = xmalloc(len);
2051         memcpy(co->buf, buf, len);
2052         hashcpy(co->sha1, sha1);
2053         return 0;
2054 }
2055
2056 static void *read_object(const unsigned char *sha1, enum object_type *type,
2057                          unsigned long *size)
2058 {
2059         unsigned long mapsize;
2060         void *map, *buf;
2061         struct cached_object *co;
2062
2063         co = find_cached_object(sha1);
2064         if (co) {
2065                 *type = co->type;
2066                 *size = co->size;
2067                 return xmemdupz(co->buf, co->size);
2068         }
2069
2070         buf = read_packed_sha1(sha1, type, size);
2071         if (buf)
2072                 return buf;
2073         map = map_sha1_file(sha1, &mapsize);
2074         if (map) {
2075                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2076                 munmap(map, mapsize);
2077                 return buf;
2078         }
2079         reprepare_packed_git();
2080         return read_packed_sha1(sha1, type, size);
2081 }
2082
2083 void *read_sha1_file_repl(const unsigned char *sha1,
2084                           enum object_type *type,
2085                           unsigned long *size,
2086                           const unsigned char **replacement)
2087 {
2088         const unsigned char *repl = lookup_replace_object(sha1);
2089         void *data = read_object(repl, type, size);
2090
2091         /* die if we replaced an object with one that does not exist */
2092         if (!data && repl != sha1)
2093                 die("replacement %s not found for %s",
2094                     sha1_to_hex(repl), sha1_to_hex(sha1));
2095
2096         /* legacy behavior is to die on corrupted objects */
2097         if (!data && (has_loose_object(repl) || has_packed_and_bad(repl)))
2098                 die("object %s is corrupted", sha1_to_hex(repl));
2099
2100         if (replacement)
2101                 *replacement = repl;
2102
2103         return data;
2104 }
2105
2106 void *read_object_with_reference(const unsigned char *sha1,
2107                                  const char *required_type_name,
2108                                  unsigned long *size,
2109                                  unsigned char *actual_sha1_return)
2110 {
2111         enum object_type type, required_type;
2112         void *buffer;
2113         unsigned long isize;
2114         unsigned char actual_sha1[20];
2115
2116         required_type = type_from_string(required_type_name);
2117         hashcpy(actual_sha1, sha1);
2118         while (1) {
2119                 int ref_length = -1;
2120                 const char *ref_type = NULL;
2121
2122                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2123                 if (!buffer)
2124                         return NULL;
2125                 if (type == required_type) {
2126                         *size = isize;
2127                         if (actual_sha1_return)
2128                                 hashcpy(actual_sha1_return, actual_sha1);
2129                         return buffer;
2130                 }
2131                 /* Handle references */
2132                 else if (type == OBJ_COMMIT)
2133                         ref_type = "tree ";
2134                 else if (type == OBJ_TAG)
2135                         ref_type = "object ";
2136                 else {
2137                         free(buffer);
2138                         return NULL;
2139                 }
2140                 ref_length = strlen(ref_type);
2141
2142                 if (ref_length + 40 > isize ||
2143                     memcmp(buffer, ref_type, ref_length) ||
2144                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2145                         free(buffer);
2146                         return NULL;
2147                 }
2148                 free(buffer);
2149                 /* Now we have the ID of the referred-to object in
2150                  * actual_sha1.  Check again. */
2151         }
2152 }
2153
2154 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2155                                     const char *type, unsigned char *sha1,
2156                                     char *hdr, int *hdrlen)
2157 {
2158         git_SHA_CTX c;
2159
2160         /* Generate the header */
2161         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2162
2163         /* Sha1.. */
2164         git_SHA1_Init(&c);
2165         git_SHA1_Update(&c, hdr, *hdrlen);
2166         git_SHA1_Update(&c, buf, len);
2167         git_SHA1_Final(sha1, &c);
2168 }
2169
2170 /*
2171  * Move the just written object into its final resting place.
2172  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2173  * "moving" is only a part of what it does, when no patch between
2174  * master to pu changes the call sites of this function.
2175  */
2176 int move_temp_to_file(const char *tmpfile, const char *filename)
2177 {
2178         int ret = 0;
2179
2180         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2181                 goto try_rename;
2182         else if (link(tmpfile, filename))
2183                 ret = errno;
2184
2185         /*
2186          * Coda hack - coda doesn't like cross-directory links,
2187          * so we fall back to a rename, which will mean that it
2188          * won't be able to check collisions, but that's not a
2189          * big deal.
2190          *
2191          * The same holds for FAT formatted media.
2192          *
2193          * When this succeeds, we just return.  We have nothing
2194          * left to unlink.
2195          */
2196         if (ret && ret != EEXIST) {
2197         try_rename:
2198                 if (!rename(tmpfile, filename))
2199                         goto out;
2200                 ret = errno;
2201         }
2202         unlink_or_warn(tmpfile);
2203         if (ret) {
2204                 if (ret != EEXIST) {
2205                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2206                 }
2207                 /* FIXME!!! Collision check here ? */
2208         }
2209
2210 out:
2211         if (set_shared_perm(filename, (S_IFREG|0444)))
2212                 return error("unable to set permission to '%s'", filename);
2213         return 0;
2214 }
2215
2216 static int write_buffer(int fd, const void *buf, size_t len)
2217 {
2218         if (write_in_full(fd, buf, len) < 0)
2219                 return error("file write error (%s)", strerror(errno));
2220         return 0;
2221 }
2222
2223 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2224                    unsigned char *sha1)
2225 {
2226         char hdr[32];
2227         int hdrlen;
2228         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2229         return 0;
2230 }
2231
2232 /* Finalize a file on disk, and close it. */
2233 static void close_sha1_file(int fd)
2234 {
2235         if (fsync_object_files)
2236                 fsync_or_die(fd, "sha1 file");
2237         if (close(fd) != 0)
2238                 die_errno("error when closing sha1 file");
2239 }
2240
2241 /* Size of directory component, including the ending '/' */
2242 static inline int directory_size(const char *filename)
2243 {
2244         const char *s = strrchr(filename, '/');
2245         if (!s)
2246                 return 0;
2247         return s - filename + 1;
2248 }
2249
2250 /*
2251  * This creates a temporary file in the same directory as the final
2252  * 'filename'
2253  *
2254  * We want to avoid cross-directory filename renames, because those
2255  * can have problems on various filesystems (FAT, NFS, Coda).
2256  */
2257 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2258 {
2259         int fd, dirlen = directory_size(filename);
2260
2261         if (dirlen + 20 > bufsiz) {
2262                 errno = ENAMETOOLONG;
2263                 return -1;
2264         }
2265         memcpy(buffer, filename, dirlen);
2266         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2267         fd = mkstemp(buffer);
2268         if (fd < 0 && dirlen && errno == ENOENT) {
2269                 /* Make sure the directory exists */
2270                 memcpy(buffer, filename, dirlen);
2271                 buffer[dirlen-1] = 0;
2272                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2273                         return -1;
2274
2275                 /* Try again */
2276                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2277                 fd = mkstemp(buffer);
2278         }
2279         return fd;
2280 }
2281
2282 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2283                               void *buf, unsigned long len, time_t mtime)
2284 {
2285         int fd, ret;
2286         size_t size;
2287         unsigned char *compressed;
2288         z_stream stream;
2289         char *filename;
2290         static char tmpfile[PATH_MAX];
2291
2292         filename = sha1_file_name(sha1);
2293         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2294         while (fd < 0 && errno == EMFILE && unuse_one_window(packed_git, -1))
2295                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2296         if (fd < 0) {
2297                 if (errno == EACCES)
2298                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2299                 else
2300                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2301         }
2302
2303         /* Set it up */
2304         memset(&stream, 0, sizeof(stream));
2305         deflateInit(&stream, zlib_compression_level);
2306         size = 8 + deflateBound(&stream, len+hdrlen);
2307         compressed = xmalloc(size);
2308
2309         /* Compress it */
2310         stream.next_out = compressed;
2311         stream.avail_out = size;
2312
2313         /* First header.. */
2314         stream.next_in = (unsigned char *)hdr;
2315         stream.avail_in = hdrlen;
2316         while (deflate(&stream, 0) == Z_OK)
2317                 /* nothing */;
2318
2319         /* Then the data itself.. */
2320         stream.next_in = buf;
2321         stream.avail_in = len;
2322         ret = deflate(&stream, Z_FINISH);
2323         if (ret != Z_STREAM_END)
2324                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2325
2326         ret = deflateEnd(&stream);
2327         if (ret != Z_OK)
2328                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2329
2330         size = stream.total_out;
2331
2332         if (write_buffer(fd, compressed, size) < 0)
2333                 die("unable to write sha1 file");
2334         close_sha1_file(fd);
2335         free(compressed);
2336
2337         if (mtime) {
2338                 struct utimbuf utb;
2339                 utb.actime = mtime;
2340                 utb.modtime = mtime;
2341                 if (utime(tmpfile, &utb) < 0)
2342                         warning("failed utime() on %s: %s",
2343                                 tmpfile, strerror(errno));
2344         }
2345
2346         return move_temp_to_file(tmpfile, filename);
2347 }
2348
2349 int write_sha1_file(void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2350 {
2351         unsigned char sha1[20];
2352         char hdr[32];
2353         int hdrlen;
2354
2355         /* Normally if we have it in the pack then we do not bother writing
2356          * it out into .git/objects/??/?{38} file.
2357          */
2358         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2359         if (returnsha1)
2360                 hashcpy(returnsha1, sha1);
2361         if (has_sha1_file(sha1))
2362                 return 0;
2363         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2364 }
2365
2366 int force_object_loose(const unsigned char *sha1, time_t mtime)
2367 {
2368         void *buf;
2369         unsigned long len;
2370         enum object_type type;
2371         char hdr[32];
2372         int hdrlen;
2373         int ret;
2374
2375         if (has_loose_object(sha1))
2376                 return 0;
2377         buf = read_packed_sha1(sha1, &type, &len);
2378         if (!buf)
2379                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2380         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2381         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2382         free(buf);
2383
2384         return ret;
2385 }
2386
2387 int has_pack_index(const unsigned char *sha1)
2388 {
2389         struct stat st;
2390         if (stat(sha1_pack_index_name(sha1), &st))
2391                 return 0;
2392         return 1;
2393 }
2394
2395 int has_sha1_pack(const unsigned char *sha1)
2396 {
2397         struct pack_entry e;
2398         return find_pack_entry(sha1, &e);
2399 }
2400
2401 int has_sha1_file(const unsigned char *sha1)
2402 {
2403         struct pack_entry e;
2404
2405         if (find_pack_entry(sha1, &e))
2406                 return 1;
2407         return has_loose_object(sha1);
2408 }
2409
2410 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2411                      int write_object, enum object_type type, const char *path)
2412 {
2413         int ret, re_allocated = 0;
2414
2415         if (!type)
2416                 type = OBJ_BLOB;
2417
2418         /*
2419          * Convert blobs to git internal format
2420          */
2421         if ((type == OBJ_BLOB) && path) {
2422                 struct strbuf nbuf = STRBUF_INIT;
2423                 if (convert_to_git(path, buf, size, &nbuf,
2424                                    write_object ? safe_crlf : 0)) {
2425                         buf = strbuf_detach(&nbuf, &size);
2426                         re_allocated = 1;
2427                 }
2428         }
2429
2430         if (write_object)
2431                 ret = write_sha1_file(buf, size, typename(type), sha1);
2432         else
2433                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2434         if (re_allocated)
2435                 free(buf);
2436         return ret;
2437 }
2438
2439 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2440              enum object_type type, const char *path)
2441 {
2442         int ret;
2443         size_t size = xsize_t(st->st_size);
2444
2445         if (!S_ISREG(st->st_mode)) {
2446                 struct strbuf sbuf = STRBUF_INIT;
2447                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2448                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2449                                         type, path);
2450                 else
2451                         ret = -1;
2452                 strbuf_release(&sbuf);
2453         } else if (size) {
2454                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2455                 ret = index_mem(sha1, buf, size, write_object, type, path);
2456                 munmap(buf, size);
2457         } else
2458                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2459         close(fd);
2460         return ret;
2461 }
2462
2463 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2464 {
2465         int fd;
2466         struct strbuf sb = STRBUF_INIT;
2467
2468         switch (st->st_mode & S_IFMT) {
2469         case S_IFREG:
2470                 fd = open(path, O_RDONLY);
2471                 if (fd < 0)
2472                         return error("open(\"%s\"): %s", path,
2473                                      strerror(errno));
2474                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2475                         return error("%s: failed to insert into database",
2476                                      path);
2477                 break;
2478         case S_IFLNK:
2479                 if (strbuf_readlink(&sb, path, st->st_size)) {
2480                         char *errstr = strerror(errno);
2481                         return error("readlink(\"%s\"): %s", path,
2482                                      errstr);
2483                 }
2484                 if (!write_object)
2485                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2486                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2487                         return error("%s: failed to insert into database",
2488                                      path);
2489                 strbuf_release(&sb);
2490                 break;
2491         case S_IFDIR:
2492                 return resolve_gitlink_ref(path, "HEAD", sha1);
2493         default:
2494                 return error("%s: unsupported file type", path);
2495         }
2496         return 0;
2497 }
2498
2499 int read_pack_header(int fd, struct pack_header *header)
2500 {
2501         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2502                 /* "eof before pack header was fully read" */
2503                 return PH_ERROR_EOF;
2504
2505         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2506                 /* "protocol error (pack signature mismatch detected)" */
2507                 return PH_ERROR_PACK_SIGNATURE;
2508         if (!pack_version_ok(header->hdr_version))
2509                 /* "protocol error (pack version unsupported)" */
2510                 return PH_ERROR_PROTOCOL;
2511         return 0;
2512 }