OSDN Git Service

Extract verify_pack_index for reuse from verify_pack
[git-core/git.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 void close_pack_index(struct packed_git *p)
610 {
611         if (p->index_data) {
612                 munmap((void *)p->index_data, p->index_size);
613                 p->index_data = NULL;
614         }
615 }
616
617 /*
618  * This is used by git-repack in case a newly created pack happens to
619  * contain the same set of objects as an existing one.  In that case
620  * the resulting file might be different even if its name would be the
621  * same.  It is best to close any reference to the old pack before it is
622  * replaced on disk.  Of course no index pointers nor windows for given pack
623  * must subsist at this point.  If ever objects from this pack are requested
624  * again, the new version of the pack will be reinitialized through
625  * reprepare_packed_git().
626  */
627 void free_pack_by_name(const char *pack_name)
628 {
629         struct packed_git *p, **pp = &packed_git;
630
631         while (*pp) {
632                 p = *pp;
633                 if (strcmp(pack_name, p->pack_name) == 0) {
634                         clear_delta_base_cache();
635                         close_pack_windows(p);
636                         if (p->pack_fd != -1)
637                                 close(p->pack_fd);
638                         close_pack_index(p);
639                         free(p->bad_object_sha1);
640                         *pp = p->next;
641                         free(p);
642                         return;
643                 }
644                 pp = &p->next;
645         }
646 }
647
648 /*
649  * Do not call this directly as this leaks p->pack_fd on error return;
650  * call open_packed_git() instead.
651  */
652 static int open_packed_git_1(struct packed_git *p)
653 {
654         struct stat st;
655         struct pack_header hdr;
656         unsigned char sha1[20];
657         unsigned char *idx_sha1;
658         long fd_flag;
659
660         if (!p->index_data && open_pack_index(p))
661                 return error("packfile %s index unavailable", p->pack_name);
662
663         p->pack_fd = open(p->pack_name, O_RDONLY);
664         while (p->pack_fd < 0 && errno == EMFILE && unuse_one_window(p, -1))
665                 p->pack_fd = open(p->pack_name, O_RDONLY);
666         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
667                 return -1;
668
669         /* If we created the struct before we had the pack we lack size. */
670         if (!p->pack_size) {
671                 if (!S_ISREG(st.st_mode))
672                         return error("packfile %s not a regular file", p->pack_name);
673                 p->pack_size = st.st_size;
674         } else if (p->pack_size != st.st_size)
675                 return error("packfile %s size changed", p->pack_name);
676
677         /* We leave these file descriptors open with sliding mmap;
678          * there is no point keeping them open across exec(), though.
679          */
680         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
681         if (fd_flag < 0)
682                 return error("cannot determine file descriptor flags");
683         fd_flag |= FD_CLOEXEC;
684         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
685                 return error("cannot set FD_CLOEXEC");
686
687         /* Verify we recognize this pack file format. */
688         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
689                 return error("file %s is far too short to be a packfile", p->pack_name);
690         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
691                 return error("file %s is not a GIT packfile", p->pack_name);
692         if (!pack_version_ok(hdr.hdr_version))
693                 return error("packfile %s is version %"PRIu32" and not"
694                         " supported (try upgrading GIT to a newer version)",
695                         p->pack_name, ntohl(hdr.hdr_version));
696
697         /* Verify the pack matches its index. */
698         if (p->num_objects != ntohl(hdr.hdr_entries))
699                 return error("packfile %s claims to have %"PRIu32" objects"
700                              " while index indicates %"PRIu32" objects",
701                              p->pack_name, ntohl(hdr.hdr_entries),
702                              p->num_objects);
703         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
704                 return error("end of packfile %s is unavailable", p->pack_name);
705         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
706                 return error("packfile %s signature is unavailable", p->pack_name);
707         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
708         if (hashcmp(sha1, idx_sha1))
709                 return error("packfile %s does not match index", p->pack_name);
710         return 0;
711 }
712
713 static int open_packed_git(struct packed_git *p)
714 {
715         if (!open_packed_git_1(p))
716                 return 0;
717         if (p->pack_fd != -1) {
718                 close(p->pack_fd);
719                 p->pack_fd = -1;
720         }
721         return -1;
722 }
723
724 static int in_window(struct pack_window *win, off_t offset)
725 {
726         /* We must promise at least 20 bytes (one hash) after the
727          * offset is available from this window, otherwise the offset
728          * is not actually in this window and a different window (which
729          * has that one hash excess) must be used.  This is to support
730          * the object header and delta base parsing routines below.
731          */
732         off_t win_off = win->offset;
733         return win_off <= offset
734                 && (offset + 20) <= (win_off + win->len);
735 }
736
737 unsigned char *use_pack(struct packed_git *p,
738                 struct pack_window **w_cursor,
739                 off_t offset,
740                 unsigned int *left)
741 {
742         struct pack_window *win = *w_cursor;
743
744         if (p->pack_fd == -1 && open_packed_git(p))
745                 die("packfile %s cannot be accessed", p->pack_name);
746
747         /* Since packfiles end in a hash of their content and it's
748          * pointless to ask for an offset into the middle of that
749          * hash, and the in_window function above wouldn't match
750          * don't allow an offset too close to the end of the file.
751          */
752         if (offset > (p->pack_size - 20))
753                 die("offset beyond end of packfile (truncated pack?)");
754
755         if (!win || !in_window(win, offset)) {
756                 if (win)
757                         win->inuse_cnt--;
758                 for (win = p->windows; win; win = win->next) {
759                         if (in_window(win, offset))
760                                 break;
761                 }
762                 if (!win) {
763                         size_t window_align = packed_git_window_size / 2;
764                         off_t len;
765                         win = xcalloc(1, sizeof(*win));
766                         win->offset = (offset / window_align) * window_align;
767                         len = p->pack_size - win->offset;
768                         if (len > packed_git_window_size)
769                                 len = packed_git_window_size;
770                         win->len = (size_t)len;
771                         pack_mapped += win->len;
772                         while (packed_git_limit < pack_mapped
773                                 && unuse_one_window(p, p->pack_fd))
774                                 ; /* nothing */
775                         win->base = xmmap(NULL, win->len,
776                                 PROT_READ, MAP_PRIVATE,
777                                 p->pack_fd, win->offset);
778                         if (win->base == MAP_FAILED)
779                                 die("packfile %s cannot be mapped: %s",
780                                         p->pack_name,
781                                         strerror(errno));
782                         pack_mmap_calls++;
783                         pack_open_windows++;
784                         if (pack_mapped > peak_pack_mapped)
785                                 peak_pack_mapped = pack_mapped;
786                         if (pack_open_windows > peak_pack_open_windows)
787                                 peak_pack_open_windows = pack_open_windows;
788                         win->next = p->windows;
789                         p->windows = win;
790                 }
791         }
792         if (win != *w_cursor) {
793                 win->last_used = pack_used_ctr++;
794                 win->inuse_cnt++;
795                 *w_cursor = win;
796         }
797         offset -= win->offset;
798         if (left)
799                 *left = win->len - xsize_t(offset);
800         return win->base + offset;
801 }
802
803 static struct packed_git *alloc_packed_git(int extra)
804 {
805         struct packed_git *p = xmalloc(sizeof(*p) + extra);
806         memset(p, 0, sizeof(*p));
807         p->pack_fd = -1;
808         return p;
809 }
810
811 struct packed_git *add_packed_git(const char *path, int path_len, int local)
812 {
813         struct stat st;
814         struct packed_git *p = alloc_packed_git(path_len + 2);
815
816         /*
817          * Make sure a corresponding .pack file exists and that
818          * the index looks sane.
819          */
820         path_len -= strlen(".idx");
821         if (path_len < 1) {
822                 free(p);
823                 return NULL;
824         }
825         memcpy(p->pack_name, path, path_len);
826
827         strcpy(p->pack_name + path_len, ".keep");
828         if (!access(p->pack_name, F_OK))
829                 p->pack_keep = 1;
830
831         strcpy(p->pack_name + path_len, ".pack");
832         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
833                 free(p);
834                 return NULL;
835         }
836
837         /* ok, it looks sane as far as we can check without
838          * actually mapping the pack file.
839          */
840         p->pack_size = st.st_size;
841         p->pack_local = local;
842         p->mtime = st.st_mtime;
843         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
844                 hashclr(p->sha1);
845         return p;
846 }
847
848 struct packed_git *parse_pack_index(unsigned char *sha1)
849 {
850         const char *idx_path = sha1_pack_index_name(sha1);
851         const char *path = sha1_pack_name(sha1);
852         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
853
854         strcpy(p->pack_name, path);
855         hashcpy(p->sha1, sha1);
856         if (check_packed_git_idx(idx_path, p)) {
857                 free(p);
858                 return NULL;
859         }
860
861         return p;
862 }
863
864 void install_packed_git(struct packed_git *pack)
865 {
866         pack->next = packed_git;
867         packed_git = pack;
868 }
869
870 static void prepare_packed_git_one(char *objdir, int local)
871 {
872         /* Ensure that this buffer is large enough so that we can
873            append "/pack/" without clobbering the stack even if
874            strlen(objdir) were PATH_MAX.  */
875         char path[PATH_MAX + 1 + 4 + 1 + 1];
876         int len;
877         DIR *dir;
878         struct dirent *de;
879
880         sprintf(path, "%s/pack", objdir);
881         len = strlen(path);
882         dir = opendir(path);
883         while (!dir && errno == EMFILE && unuse_one_window(packed_git, -1))
884                 dir = opendir(path);
885         if (!dir) {
886                 if (errno != ENOENT)
887                         error("unable to open object pack directory: %s: %s",
888                               path, strerror(errno));
889                 return;
890         }
891         path[len++] = '/';
892         while ((de = readdir(dir)) != NULL) {
893                 int namelen = strlen(de->d_name);
894                 struct packed_git *p;
895
896                 if (!has_extension(de->d_name, ".idx"))
897                         continue;
898
899                 if (len + namelen + 1 > sizeof(path))
900                         continue;
901
902                 /* Don't reopen a pack we already have. */
903                 strcpy(path + len, de->d_name);
904                 for (p = packed_git; p; p = p->next) {
905                         if (!memcmp(path, p->pack_name, len + namelen - 4))
906                                 break;
907                 }
908                 if (p)
909                         continue;
910                 /* See if it really is a valid .idx file with corresponding
911                  * .pack file that we can map.
912                  */
913                 p = add_packed_git(path, len + namelen, local);
914                 if (!p)
915                         continue;
916                 install_packed_git(p);
917         }
918         closedir(dir);
919 }
920
921 static int sort_pack(const void *a_, const void *b_)
922 {
923         struct packed_git *a = *((struct packed_git **)a_);
924         struct packed_git *b = *((struct packed_git **)b_);
925         int st;
926
927         /*
928          * Local packs tend to contain objects specific to our
929          * variant of the project than remote ones.  In addition,
930          * remote ones could be on a network mounted filesystem.
931          * Favor local ones for these reasons.
932          */
933         st = a->pack_local - b->pack_local;
934         if (st)
935                 return -st;
936
937         /*
938          * Younger packs tend to contain more recent objects,
939          * and more recent objects tend to get accessed more
940          * often.
941          */
942         if (a->mtime < b->mtime)
943                 return 1;
944         else if (a->mtime == b->mtime)
945                 return 0;
946         return -1;
947 }
948
949 static void rearrange_packed_git(void)
950 {
951         struct packed_git **ary, *p;
952         int i, n;
953
954         for (n = 0, p = packed_git; p; p = p->next)
955                 n++;
956         if (n < 2)
957                 return;
958
959         /* prepare an array of packed_git for easier sorting */
960         ary = xcalloc(n, sizeof(struct packed_git *));
961         for (n = 0, p = packed_git; p; p = p->next)
962                 ary[n++] = p;
963
964         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
965
966         /* link them back again */
967         for (i = 0; i < n - 1; i++)
968                 ary[i]->next = ary[i + 1];
969         ary[n - 1]->next = NULL;
970         packed_git = ary[0];
971
972         free(ary);
973 }
974
975 static int prepare_packed_git_run_once = 0;
976 void prepare_packed_git(void)
977 {
978         struct alternate_object_database *alt;
979
980         if (prepare_packed_git_run_once)
981                 return;
982         prepare_packed_git_one(get_object_directory(), 1);
983         prepare_alt_odb();
984         for (alt = alt_odb_list; alt; alt = alt->next) {
985                 alt->name[-1] = 0;
986                 prepare_packed_git_one(alt->base, 0);
987                 alt->name[-1] = '/';
988         }
989         rearrange_packed_git();
990         prepare_packed_git_run_once = 1;
991 }
992
993 void reprepare_packed_git(void)
994 {
995         discard_revindex();
996         prepare_packed_git_run_once = 0;
997         prepare_packed_git();
998 }
999
1000 static void mark_bad_packed_object(struct packed_git *p,
1001                                    const unsigned char *sha1)
1002 {
1003         unsigned i;
1004         for (i = 0; i < p->num_bad_objects; i++)
1005                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1006                         return;
1007         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1008         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1009         p->num_bad_objects++;
1010 }
1011
1012 static int has_packed_and_bad(const unsigned char *sha1)
1013 {
1014         struct packed_git *p;
1015         unsigned i;
1016
1017         for (p = packed_git; p; p = p->next)
1018                 for (i = 0; i < p->num_bad_objects; i++)
1019                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1020                                 return 1;
1021         return 0;
1022 }
1023
1024 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1025 {
1026         unsigned char real_sha1[20];
1027         hash_sha1_file(map, size, type, real_sha1);
1028         return hashcmp(sha1, real_sha1) ? -1 : 0;
1029 }
1030
1031 static int git_open_noatime(const char *name)
1032 {
1033         static int sha1_file_open_flag = O_NOATIME;
1034         int fd = open(name, O_RDONLY | sha1_file_open_flag);
1035
1036         /* Might the failure be due to O_NOATIME? */
1037         if (fd < 0 && errno != ENOENT && sha1_file_open_flag) {
1038                 fd = open(name, O_RDONLY);
1039                 if (fd >= 0)
1040                         sha1_file_open_flag = 0;
1041         }
1042         return fd;
1043 }
1044
1045 static int open_sha1_file(const unsigned char *sha1)
1046 {
1047         int fd;
1048         char *name = sha1_file_name(sha1);
1049         struct alternate_object_database *alt;
1050
1051         fd = git_open_noatime(name);
1052         if (fd >= 0)
1053                 return fd;
1054
1055         prepare_alt_odb();
1056         errno = ENOENT;
1057         for (alt = alt_odb_list; alt; alt = alt->next) {
1058                 name = alt->name;
1059                 fill_sha1_path(name, sha1);
1060                 fd = git_open_noatime(alt->base);
1061                 if (fd >= 0)
1062                         return fd;
1063         }
1064         return -1;
1065 }
1066
1067 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1068 {
1069         void *map;
1070         int fd;
1071
1072         fd = open_sha1_file(sha1);
1073         map = NULL;
1074         if (fd >= 0) {
1075                 struct stat st;
1076
1077                 if (!fstat(fd, &st)) {
1078                         *size = xsize_t(st.st_size);
1079                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1080                 }
1081                 close(fd);
1082         }
1083         return map;
1084 }
1085
1086 static int legacy_loose_object(unsigned char *map)
1087 {
1088         unsigned int word;
1089
1090         /*
1091          * Is it a zlib-compressed buffer? If so, the first byte
1092          * must be 0x78 (15-bit window size, deflated), and the
1093          * first 16-bit word is evenly divisible by 31
1094          */
1095         word = (map[0] << 8) + map[1];
1096         if (map[0] == 0x78 && !(word % 31))
1097                 return 1;
1098         else
1099                 return 0;
1100 }
1101
1102 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1103                 unsigned long len, enum object_type *type, unsigned long *sizep)
1104 {
1105         unsigned shift;
1106         unsigned long size, c;
1107         unsigned long used = 0;
1108
1109         c = buf[used++];
1110         *type = (c >> 4) & 7;
1111         size = c & 15;
1112         shift = 4;
1113         while (c & 0x80) {
1114                 if (len <= used || bitsizeof(long) <= shift) {
1115                         error("bad object header");
1116                         return 0;
1117                 }
1118                 c = buf[used++];
1119                 size += (c & 0x7f) << shift;
1120                 shift += 7;
1121         }
1122         *sizep = size;
1123         return used;
1124 }
1125
1126 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1127 {
1128         unsigned long size, used;
1129         static const char valid_loose_object_type[8] = {
1130                 0, /* OBJ_EXT */
1131                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1132                 0, /* "delta" and others are invalid in a loose object */
1133         };
1134         enum object_type type;
1135
1136         /* Get the data stream */
1137         memset(stream, 0, sizeof(*stream));
1138         stream->next_in = map;
1139         stream->avail_in = mapsize;
1140         stream->next_out = buffer;
1141         stream->avail_out = bufsiz;
1142
1143         if (legacy_loose_object(map)) {
1144                 git_inflate_init(stream);
1145                 return git_inflate(stream, 0);
1146         }
1147
1148
1149         /*
1150          * There used to be a second loose object header format which
1151          * was meant to mimic the in-pack format, allowing for direct
1152          * copy of the object data.  This format turned up not to be
1153          * really worth it and we don't write it any longer.  But we
1154          * can still read it.
1155          */
1156         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1157         if (!used || !valid_loose_object_type[type])
1158                 return -1;
1159         map += used;
1160         mapsize -= used;
1161
1162         /* Set up the stream for the rest.. */
1163         stream->next_in = map;
1164         stream->avail_in = mapsize;
1165         git_inflate_init(stream);
1166
1167         /* And generate the fake traditional header */
1168         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1169                                          typename(type), size);
1170         return 0;
1171 }
1172
1173 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1174 {
1175         int bytes = strlen(buffer) + 1;
1176         unsigned char *buf = xmallocz(size);
1177         unsigned long n;
1178         int status = Z_OK;
1179
1180         n = stream->total_out - bytes;
1181         if (n > size)
1182                 n = size;
1183         memcpy(buf, (char *) buffer + bytes, n);
1184         bytes = n;
1185         if (bytes <= size) {
1186                 /*
1187                  * The above condition must be (bytes <= size), not
1188                  * (bytes < size).  In other words, even though we
1189                  * expect no more output and set avail_out to zer0,
1190                  * the input zlib stream may have bytes that express
1191                  * "this concludes the stream", and we *do* want to
1192                  * eat that input.
1193                  *
1194                  * Otherwise we would not be able to test that we
1195                  * consumed all the input to reach the expected size;
1196                  * we also want to check that zlib tells us that all
1197                  * went well with status == Z_STREAM_END at the end.
1198                  */
1199                 stream->next_out = buf + bytes;
1200                 stream->avail_out = size - bytes;
1201                 while (status == Z_OK)
1202                         status = git_inflate(stream, Z_FINISH);
1203         }
1204         if (status == Z_STREAM_END && !stream->avail_in) {
1205                 git_inflate_end(stream);
1206                 return buf;
1207         }
1208
1209         if (status < 0)
1210                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1211         else if (stream->avail_in)
1212                 error("garbage at end of loose object '%s'",
1213                       sha1_to_hex(sha1));
1214         free(buf);
1215         return NULL;
1216 }
1217
1218 /*
1219  * We used to just use "sscanf()", but that's actually way
1220  * too permissive for what we want to check. So do an anal
1221  * object header parse by hand.
1222  */
1223 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1224 {
1225         char type[10];
1226         int i;
1227         unsigned long size;
1228
1229         /*
1230          * The type can be at most ten bytes (including the
1231          * terminating '\0' that we add), and is followed by
1232          * a space.
1233          */
1234         i = 0;
1235         for (;;) {
1236                 char c = *hdr++;
1237                 if (c == ' ')
1238                         break;
1239                 type[i++] = c;
1240                 if (i >= sizeof(type))
1241                         return -1;
1242         }
1243         type[i] = 0;
1244
1245         /*
1246          * The length must follow immediately, and be in canonical
1247          * decimal format (ie "010" is not valid).
1248          */
1249         size = *hdr++ - '0';
1250         if (size > 9)
1251                 return -1;
1252         if (size) {
1253                 for (;;) {
1254                         unsigned long c = *hdr - '0';
1255                         if (c > 9)
1256                                 break;
1257                         hdr++;
1258                         size = size * 10 + c;
1259                 }
1260         }
1261         *sizep = size;
1262
1263         /*
1264          * The length must be followed by a zero byte
1265          */
1266         return *hdr ? -1 : type_from_string(type);
1267 }
1268
1269 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1270 {
1271         int ret;
1272         z_stream stream;
1273         char hdr[8192];
1274
1275         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1276         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1277                 return NULL;
1278
1279         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1280 }
1281
1282 unsigned long get_size_from_delta(struct packed_git *p,
1283                                   struct pack_window **w_curs,
1284                                   off_t curpos)
1285 {
1286         const unsigned char *data;
1287         unsigned char delta_head[20], *in;
1288         z_stream stream;
1289         int st;
1290
1291         memset(&stream, 0, sizeof(stream));
1292         stream.next_out = delta_head;
1293         stream.avail_out = sizeof(delta_head);
1294
1295         git_inflate_init(&stream);
1296         do {
1297                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1298                 stream.next_in = in;
1299                 st = git_inflate(&stream, Z_FINISH);
1300                 curpos += stream.next_in - in;
1301         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1302                  stream.total_out < sizeof(delta_head));
1303         git_inflate_end(&stream);
1304         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1305                 error("delta data unpack-initial failed");
1306                 return 0;
1307         }
1308
1309         /* Examine the initial part of the delta to figure out
1310          * the result size.
1311          */
1312         data = delta_head;
1313
1314         /* ignore base size */
1315         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1316
1317         /* Read the result size */
1318         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1319 }
1320
1321 static off_t get_delta_base(struct packed_git *p,
1322                                     struct pack_window **w_curs,
1323                                     off_t *curpos,
1324                                     enum object_type type,
1325                                     off_t delta_obj_offset)
1326 {
1327         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1328         off_t base_offset;
1329
1330         /* use_pack() assured us we have [base_info, base_info + 20)
1331          * as a range that we can look at without walking off the
1332          * end of the mapped window.  Its actually the hash size
1333          * that is assured.  An OFS_DELTA longer than the hash size
1334          * is stupid, as then a REF_DELTA would be smaller to store.
1335          */
1336         if (type == OBJ_OFS_DELTA) {
1337                 unsigned used = 0;
1338                 unsigned char c = base_info[used++];
1339                 base_offset = c & 127;
1340                 while (c & 128) {
1341                         base_offset += 1;
1342                         if (!base_offset || MSB(base_offset, 7))
1343                                 return 0;  /* overflow */
1344                         c = base_info[used++];
1345                         base_offset = (base_offset << 7) + (c & 127);
1346                 }
1347                 base_offset = delta_obj_offset - base_offset;
1348                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1349                         return 0;  /* out of bound */
1350                 *curpos += used;
1351         } else if (type == OBJ_REF_DELTA) {
1352                 /* The base entry _must_ be in the same pack */
1353                 base_offset = find_pack_entry_one(base_info, p);
1354                 *curpos += 20;
1355         } else
1356                 die("I am totally screwed");
1357         return base_offset;
1358 }
1359
1360 /* forward declaration for a mutually recursive function */
1361 static int packed_object_info(struct packed_git *p, off_t offset,
1362                               unsigned long *sizep);
1363
1364 static int packed_delta_info(struct packed_git *p,
1365                              struct pack_window **w_curs,
1366                              off_t curpos,
1367                              enum object_type type,
1368                              off_t obj_offset,
1369                              unsigned long *sizep)
1370 {
1371         off_t base_offset;
1372
1373         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1374         if (!base_offset)
1375                 return OBJ_BAD;
1376         type = packed_object_info(p, base_offset, NULL);
1377         if (type <= OBJ_NONE) {
1378                 struct revindex_entry *revidx;
1379                 const unsigned char *base_sha1;
1380                 revidx = find_pack_revindex(p, base_offset);
1381                 if (!revidx)
1382                         return OBJ_BAD;
1383                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1384                 mark_bad_packed_object(p, base_sha1);
1385                 type = sha1_object_info(base_sha1, NULL);
1386                 if (type <= OBJ_NONE)
1387                         return OBJ_BAD;
1388         }
1389
1390         /* We choose to only get the type of the base object and
1391          * ignore potentially corrupt pack file that expects the delta
1392          * based on a base with a wrong size.  This saves tons of
1393          * inflate() calls.
1394          */
1395         if (sizep) {
1396                 *sizep = get_size_from_delta(p, w_curs, curpos);
1397                 if (*sizep == 0)
1398                         type = OBJ_BAD;
1399         }
1400
1401         return type;
1402 }
1403
1404 static int unpack_object_header(struct packed_git *p,
1405                                 struct pack_window **w_curs,
1406                                 off_t *curpos,
1407                                 unsigned long *sizep)
1408 {
1409         unsigned char *base;
1410         unsigned int left;
1411         unsigned long used;
1412         enum object_type type;
1413
1414         /* use_pack() assures us we have [base, base + 20) available
1415          * as a range that we can look at at.  (Its actually the hash
1416          * size that is assured.)  With our object header encoding
1417          * the maximum deflated object size is 2^137, which is just
1418          * insane, so we know won't exceed what we have been given.
1419          */
1420         base = use_pack(p, w_curs, *curpos, &left);
1421         used = unpack_object_header_buffer(base, left, &type, sizep);
1422         if (!used) {
1423                 type = OBJ_BAD;
1424         } else
1425                 *curpos += used;
1426
1427         return type;
1428 }
1429
1430 const char *packed_object_info_detail(struct packed_git *p,
1431                                       off_t obj_offset,
1432                                       unsigned long *size,
1433                                       unsigned long *store_size,
1434                                       unsigned int *delta_chain_length,
1435                                       unsigned char *base_sha1)
1436 {
1437         struct pack_window *w_curs = NULL;
1438         off_t curpos;
1439         unsigned long dummy;
1440         unsigned char *next_sha1;
1441         enum object_type type;
1442         struct revindex_entry *revidx;
1443
1444         *delta_chain_length = 0;
1445         curpos = obj_offset;
1446         type = unpack_object_header(p, &w_curs, &curpos, size);
1447
1448         revidx = find_pack_revindex(p, obj_offset);
1449         *store_size = revidx[1].offset - obj_offset;
1450
1451         for (;;) {
1452                 switch (type) {
1453                 default:
1454                         die("pack %s contains unknown object type %d",
1455                             p->pack_name, type);
1456                 case OBJ_COMMIT:
1457                 case OBJ_TREE:
1458                 case OBJ_BLOB:
1459                 case OBJ_TAG:
1460                         unuse_pack(&w_curs);
1461                         return typename(type);
1462                 case OBJ_OFS_DELTA:
1463                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1464                         if (!obj_offset)
1465                                 die("pack %s contains bad delta base reference of type %s",
1466                                     p->pack_name, typename(type));
1467                         if (*delta_chain_length == 0) {
1468                                 revidx = find_pack_revindex(p, obj_offset);
1469                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1470                         }
1471                         break;
1472                 case OBJ_REF_DELTA:
1473                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1474                         if (*delta_chain_length == 0)
1475                                 hashcpy(base_sha1, next_sha1);
1476                         obj_offset = find_pack_entry_one(next_sha1, p);
1477                         break;
1478                 }
1479                 (*delta_chain_length)++;
1480                 curpos = obj_offset;
1481                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1482         }
1483 }
1484
1485 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1486                               unsigned long *sizep)
1487 {
1488         struct pack_window *w_curs = NULL;
1489         unsigned long size;
1490         off_t curpos = obj_offset;
1491         enum object_type type;
1492
1493         type = unpack_object_header(p, &w_curs, &curpos, &size);
1494
1495         switch (type) {
1496         case OBJ_OFS_DELTA:
1497         case OBJ_REF_DELTA:
1498                 type = packed_delta_info(p, &w_curs, curpos,
1499                                          type, obj_offset, sizep);
1500                 break;
1501         case OBJ_COMMIT:
1502         case OBJ_TREE:
1503         case OBJ_BLOB:
1504         case OBJ_TAG:
1505                 if (sizep)
1506                         *sizep = size;
1507                 break;
1508         default:
1509                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1510                       type, (uintmax_t)obj_offset, p->pack_name);
1511                 type = OBJ_BAD;
1512         }
1513         unuse_pack(&w_curs);
1514         return type;
1515 }
1516
1517 static void *unpack_compressed_entry(struct packed_git *p,
1518                                     struct pack_window **w_curs,
1519                                     off_t curpos,
1520                                     unsigned long size)
1521 {
1522         int st;
1523         z_stream stream;
1524         unsigned char *buffer, *in;
1525
1526         buffer = xmallocz(size);
1527         memset(&stream, 0, sizeof(stream));
1528         stream.next_out = buffer;
1529         stream.avail_out = size + 1;
1530
1531         git_inflate_init(&stream);
1532         do {
1533                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1534                 stream.next_in = in;
1535                 st = git_inflate(&stream, Z_FINISH);
1536                 if (!stream.avail_out)
1537                         break; /* the payload is larger than it should be */
1538                 curpos += stream.next_in - in;
1539         } while (st == Z_OK || st == Z_BUF_ERROR);
1540         git_inflate_end(&stream);
1541         if ((st != Z_STREAM_END) || stream.total_out != size) {
1542                 free(buffer);
1543                 return NULL;
1544         }
1545
1546         return buffer;
1547 }
1548
1549 #define MAX_DELTA_CACHE (256)
1550
1551 static size_t delta_base_cached;
1552
1553 static struct delta_base_cache_lru_list {
1554         struct delta_base_cache_lru_list *prev;
1555         struct delta_base_cache_lru_list *next;
1556 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1557
1558 static struct delta_base_cache_entry {
1559         struct delta_base_cache_lru_list lru;
1560         void *data;
1561         struct packed_git *p;
1562         off_t base_offset;
1563         unsigned long size;
1564         enum object_type type;
1565 } delta_base_cache[MAX_DELTA_CACHE];
1566
1567 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1568 {
1569         unsigned long hash;
1570
1571         hash = (unsigned long)p + (unsigned long)base_offset;
1572         hash += (hash >> 8) + (hash >> 16);
1573         return hash % MAX_DELTA_CACHE;
1574 }
1575
1576 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1577         unsigned long *base_size, enum object_type *type, int keep_cache)
1578 {
1579         void *ret;
1580         unsigned long hash = pack_entry_hash(p, base_offset);
1581         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1582
1583         ret = ent->data;
1584         if (!ret || ent->p != p || ent->base_offset != base_offset)
1585                 return unpack_entry(p, base_offset, type, base_size);
1586
1587         if (!keep_cache) {
1588                 ent->data = NULL;
1589                 ent->lru.next->prev = ent->lru.prev;
1590                 ent->lru.prev->next = ent->lru.next;
1591                 delta_base_cached -= ent->size;
1592         } else {
1593                 ret = xmemdupz(ent->data, ent->size);
1594         }
1595         *type = ent->type;
1596         *base_size = ent->size;
1597         return ret;
1598 }
1599
1600 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1601 {
1602         if (ent->data) {
1603                 free(ent->data);
1604                 ent->data = NULL;
1605                 ent->lru.next->prev = ent->lru.prev;
1606                 ent->lru.prev->next = ent->lru.next;
1607                 delta_base_cached -= ent->size;
1608         }
1609 }
1610
1611 void clear_delta_base_cache(void)
1612 {
1613         unsigned long p;
1614         for (p = 0; p < MAX_DELTA_CACHE; p++)
1615                 release_delta_base_cache(&delta_base_cache[p]);
1616 }
1617
1618 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1619         void *base, unsigned long base_size, enum object_type type)
1620 {
1621         unsigned long hash = pack_entry_hash(p, base_offset);
1622         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1623         struct delta_base_cache_lru_list *lru;
1624
1625         release_delta_base_cache(ent);
1626         delta_base_cached += base_size;
1627
1628         for (lru = delta_base_cache_lru.next;
1629              delta_base_cached > delta_base_cache_limit
1630              && lru != &delta_base_cache_lru;
1631              lru = lru->next) {
1632                 struct delta_base_cache_entry *f = (void *)lru;
1633                 if (f->type == OBJ_BLOB)
1634                         release_delta_base_cache(f);
1635         }
1636         for (lru = delta_base_cache_lru.next;
1637              delta_base_cached > delta_base_cache_limit
1638              && lru != &delta_base_cache_lru;
1639              lru = lru->next) {
1640                 struct delta_base_cache_entry *f = (void *)lru;
1641                 release_delta_base_cache(f);
1642         }
1643
1644         ent->p = p;
1645         ent->base_offset = base_offset;
1646         ent->type = type;
1647         ent->data = base;
1648         ent->size = base_size;
1649         ent->lru.next = &delta_base_cache_lru;
1650         ent->lru.prev = delta_base_cache_lru.prev;
1651         delta_base_cache_lru.prev->next = &ent->lru;
1652         delta_base_cache_lru.prev = &ent->lru;
1653 }
1654
1655 static void *read_object(const unsigned char *sha1, enum object_type *type,
1656                          unsigned long *size);
1657
1658 static void *unpack_delta_entry(struct packed_git *p,
1659                                 struct pack_window **w_curs,
1660                                 off_t curpos,
1661                                 unsigned long delta_size,
1662                                 off_t obj_offset,
1663                                 enum object_type *type,
1664                                 unsigned long *sizep)
1665 {
1666         void *delta_data, *result, *base;
1667         unsigned long base_size;
1668         off_t base_offset;
1669
1670         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1671         if (!base_offset) {
1672                 error("failed to validate delta base reference "
1673                       "at offset %"PRIuMAX" from %s",
1674                       (uintmax_t)curpos, p->pack_name);
1675                 return NULL;
1676         }
1677         unuse_pack(w_curs);
1678         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1679         if (!base) {
1680                 /*
1681                  * We're probably in deep shit, but let's try to fetch
1682                  * the required base anyway from another pack or loose.
1683                  * This is costly but should happen only in the presence
1684                  * of a corrupted pack, and is better than failing outright.
1685                  */
1686                 struct revindex_entry *revidx;
1687                 const unsigned char *base_sha1;
1688                 revidx = find_pack_revindex(p, base_offset);
1689                 if (!revidx)
1690                         return NULL;
1691                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1692                 error("failed to read delta base object %s"
1693                       " at offset %"PRIuMAX" from %s",
1694                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1695                       p->pack_name);
1696                 mark_bad_packed_object(p, base_sha1);
1697                 base = read_object(base_sha1, type, &base_size);
1698                 if (!base)
1699                         return NULL;
1700         }
1701
1702         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1703         if (!delta_data) {
1704                 error("failed to unpack compressed delta "
1705                       "at offset %"PRIuMAX" from %s",
1706                       (uintmax_t)curpos, p->pack_name);
1707                 free(base);
1708                 return NULL;
1709         }
1710         result = patch_delta(base, base_size,
1711                              delta_data, delta_size,
1712                              sizep);
1713         if (!result)
1714                 die("failed to apply delta");
1715         free(delta_data);
1716         add_delta_base_cache(p, base_offset, base, base_size, *type);
1717         return result;
1718 }
1719
1720 int do_check_packed_object_crc;
1721
1722 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1723                    enum object_type *type, unsigned long *sizep)
1724 {
1725         struct pack_window *w_curs = NULL;
1726         off_t curpos = obj_offset;
1727         void *data;
1728
1729         if (do_check_packed_object_crc && p->index_version > 1) {
1730                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1731                 unsigned long len = revidx[1].offset - obj_offset;
1732                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1733                         const unsigned char *sha1 =
1734                                 nth_packed_object_sha1(p, revidx->nr);
1735                         error("bad packed object CRC for %s",
1736                               sha1_to_hex(sha1));
1737                         mark_bad_packed_object(p, sha1);
1738                         unuse_pack(&w_curs);
1739                         return NULL;
1740                 }
1741         }
1742
1743         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1744         switch (*type) {
1745         case OBJ_OFS_DELTA:
1746         case OBJ_REF_DELTA:
1747                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1748                                           obj_offset, type, sizep);
1749                 break;
1750         case OBJ_COMMIT:
1751         case OBJ_TREE:
1752         case OBJ_BLOB:
1753         case OBJ_TAG:
1754                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1755                 break;
1756         default:
1757                 data = NULL;
1758                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1759                       *type, (uintmax_t)obj_offset, p->pack_name);
1760         }
1761         unuse_pack(&w_curs);
1762         return data;
1763 }
1764
1765 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1766                                             uint32_t n)
1767 {
1768         const unsigned char *index = p->index_data;
1769         if (!index) {
1770                 if (open_pack_index(p))
1771                         return NULL;
1772                 index = p->index_data;
1773         }
1774         if (n >= p->num_objects)
1775                 return NULL;
1776         index += 4 * 256;
1777         if (p->index_version == 1) {
1778                 return index + 24 * n + 4;
1779         } else {
1780                 index += 8;
1781                 return index + 20 * n;
1782         }
1783 }
1784
1785 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1786 {
1787         const unsigned char *index = p->index_data;
1788         index += 4 * 256;
1789         if (p->index_version == 1) {
1790                 return ntohl(*((uint32_t *)(index + 24 * n)));
1791         } else {
1792                 uint32_t off;
1793                 index += 8 + p->num_objects * (20 + 4);
1794                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1795                 if (!(off & 0x80000000))
1796                         return off;
1797                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1798                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1799                                    ntohl(*((uint32_t *)(index + 4)));
1800         }
1801 }
1802
1803 off_t find_pack_entry_one(const unsigned char *sha1,
1804                                   struct packed_git *p)
1805 {
1806         const uint32_t *level1_ofs = p->index_data;
1807         const unsigned char *index = p->index_data;
1808         unsigned hi, lo, stride;
1809         static int use_lookup = -1;
1810         static int debug_lookup = -1;
1811
1812         if (debug_lookup < 0)
1813                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1814
1815         if (!index) {
1816                 if (open_pack_index(p))
1817                         return 0;
1818                 level1_ofs = p->index_data;
1819                 index = p->index_data;
1820         }
1821         if (p->index_version > 1) {
1822                 level1_ofs += 2;
1823                 index += 8;
1824         }
1825         index += 4 * 256;
1826         hi = ntohl(level1_ofs[*sha1]);
1827         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1828         if (p->index_version > 1) {
1829                 stride = 20;
1830         } else {
1831                 stride = 24;
1832                 index += 4;
1833         }
1834
1835         if (debug_lookup)
1836                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1837                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1838
1839         if (use_lookup < 0)
1840                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1841         if (use_lookup) {
1842                 int pos = sha1_entry_pos(index, stride, 0,
1843                                          lo, hi, p->num_objects, sha1);
1844                 if (pos < 0)
1845                         return 0;
1846                 return nth_packed_object_offset(p, pos);
1847         }
1848
1849         do {
1850                 unsigned mi = (lo + hi) / 2;
1851                 int cmp = hashcmp(index + mi * stride, sha1);
1852
1853                 if (debug_lookup)
1854                         printf("lo %u hi %u rg %u mi %u\n",
1855                                lo, hi, hi - lo, mi);
1856                 if (!cmp)
1857                         return nth_packed_object_offset(p, mi);
1858                 if (cmp > 0)
1859                         hi = mi;
1860                 else
1861                         lo = mi+1;
1862         } while (lo < hi);
1863         return 0;
1864 }
1865
1866 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1867 {
1868         static struct packed_git *last_found = (void *)1;
1869         struct packed_git *p;
1870         off_t offset;
1871
1872         prepare_packed_git();
1873         if (!packed_git)
1874                 return 0;
1875         p = (last_found == (void *)1) ? packed_git : last_found;
1876
1877         do {
1878                 if (p->num_bad_objects) {
1879                         unsigned i;
1880                         for (i = 0; i < p->num_bad_objects; i++)
1881                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1882                                         goto next;
1883                 }
1884
1885                 offset = find_pack_entry_one(sha1, p);
1886                 if (offset) {
1887                         /*
1888                          * We are about to tell the caller where they can
1889                          * locate the requested object.  We better make
1890                          * sure the packfile is still here and can be
1891                          * accessed before supplying that answer, as
1892                          * it may have been deleted since the index
1893                          * was loaded!
1894                          */
1895                         if (p->pack_fd == -1 && open_packed_git(p)) {
1896                                 error("packfile %s cannot be accessed", p->pack_name);
1897                                 goto next;
1898                         }
1899                         e->offset = offset;
1900                         e->p = p;
1901                         hashcpy(e->sha1, sha1);
1902                         last_found = p;
1903                         return 1;
1904                 }
1905
1906                 next:
1907                 if (p == last_found)
1908                         p = packed_git;
1909                 else
1910                         p = p->next;
1911                 if (p == last_found)
1912                         p = p->next;
1913         } while (p);
1914         return 0;
1915 }
1916
1917 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1918                                   struct packed_git *packs)
1919 {
1920         struct packed_git *p;
1921
1922         for (p = packs; p; p = p->next) {
1923                 if (find_pack_entry_one(sha1, p))
1924                         return p;
1925         }
1926         return NULL;
1927
1928 }
1929
1930 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1931 {
1932         int status;
1933         unsigned long mapsize, size;
1934         void *map;
1935         z_stream stream;
1936         char hdr[32];
1937
1938         map = map_sha1_file(sha1, &mapsize);
1939         if (!map)
1940                 return error("unable to find %s", sha1_to_hex(sha1));
1941         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1942                 status = error("unable to unpack %s header",
1943                                sha1_to_hex(sha1));
1944         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1945                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1946         else if (sizep)
1947                 *sizep = size;
1948         git_inflate_end(&stream);
1949         munmap(map, mapsize);
1950         return status;
1951 }
1952
1953 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1954 {
1955         struct pack_entry e;
1956         int status;
1957
1958         if (!find_pack_entry(sha1, &e)) {
1959                 /* Most likely it's a loose object. */
1960                 status = sha1_loose_object_info(sha1, sizep);
1961                 if (status >= 0)
1962                         return status;
1963
1964                 /* Not a loose object; someone else may have just packed it. */
1965                 reprepare_packed_git();
1966                 if (!find_pack_entry(sha1, &e))
1967                         return status;
1968         }
1969
1970         status = packed_object_info(e.p, e.offset, sizep);
1971         if (status < 0) {
1972                 mark_bad_packed_object(e.p, sha1);
1973                 status = sha1_object_info(sha1, sizep);
1974         }
1975
1976         return status;
1977 }
1978
1979 static void *read_packed_sha1(const unsigned char *sha1,
1980                               enum object_type *type, unsigned long *size)
1981 {
1982         struct pack_entry e;
1983         void *data;
1984
1985         if (!find_pack_entry(sha1, &e))
1986                 return NULL;
1987         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
1988         if (!data) {
1989                 /*
1990                  * We're probably in deep shit, but let's try to fetch
1991                  * the required object anyway from another pack or loose.
1992                  * This should happen only in the presence of a corrupted
1993                  * pack, and is better than failing outright.
1994                  */
1995                 error("failed to read object %s at offset %"PRIuMAX" from %s",
1996                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
1997                 mark_bad_packed_object(e.p, sha1);
1998                 data = read_object(sha1, type, size);
1999         }
2000         return data;
2001 }
2002
2003 /*
2004  * This is meant to hold a *small* number of objects that you would
2005  * want read_sha1_file() to be able to return, but yet you do not want
2006  * to write them into the object store (e.g. a browse-only
2007  * application).
2008  */
2009 static struct cached_object {
2010         unsigned char sha1[20];
2011         enum object_type type;
2012         void *buf;
2013         unsigned long size;
2014 } *cached_objects;
2015 static int cached_object_nr, cached_object_alloc;
2016
2017 static struct cached_object empty_tree = {
2018         EMPTY_TREE_SHA1_BIN,
2019         OBJ_TREE,
2020         "",
2021         0
2022 };
2023
2024 static struct cached_object *find_cached_object(const unsigned char *sha1)
2025 {
2026         int i;
2027         struct cached_object *co = cached_objects;
2028
2029         for (i = 0; i < cached_object_nr; i++, co++) {
2030                 if (!hashcmp(co->sha1, sha1))
2031                         return co;
2032         }
2033         if (!hashcmp(sha1, empty_tree.sha1))
2034                 return &empty_tree;
2035         return NULL;
2036 }
2037
2038 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2039                       unsigned char *sha1)
2040 {
2041         struct cached_object *co;
2042
2043         hash_sha1_file(buf, len, typename(type), sha1);
2044         if (has_sha1_file(sha1) || find_cached_object(sha1))
2045                 return 0;
2046         if (cached_object_alloc <= cached_object_nr) {
2047                 cached_object_alloc = alloc_nr(cached_object_alloc);
2048                 cached_objects = xrealloc(cached_objects,
2049                                           sizeof(*cached_objects) *
2050                                           cached_object_alloc);
2051         }
2052         co = &cached_objects[cached_object_nr++];
2053         co->size = len;
2054         co->type = type;
2055         co->buf = xmalloc(len);
2056         memcpy(co->buf, buf, len);
2057         hashcpy(co->sha1, sha1);
2058         return 0;
2059 }
2060
2061 static void *read_object(const unsigned char *sha1, enum object_type *type,
2062                          unsigned long *size)
2063 {
2064         unsigned long mapsize;
2065         void *map, *buf;
2066         struct cached_object *co;
2067
2068         co = find_cached_object(sha1);
2069         if (co) {
2070                 *type = co->type;
2071                 *size = co->size;
2072                 return xmemdupz(co->buf, co->size);
2073         }
2074
2075         buf = read_packed_sha1(sha1, type, size);
2076         if (buf)
2077                 return buf;
2078         map = map_sha1_file(sha1, &mapsize);
2079         if (map) {
2080                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2081                 munmap(map, mapsize);
2082                 return buf;
2083         }
2084         reprepare_packed_git();
2085         return read_packed_sha1(sha1, type, size);
2086 }
2087
2088 void *read_sha1_file_repl(const unsigned char *sha1,
2089                           enum object_type *type,
2090                           unsigned long *size,
2091                           const unsigned char **replacement)
2092 {
2093         const unsigned char *repl = lookup_replace_object(sha1);
2094         void *data = read_object(repl, type, size);
2095
2096         /* die if we replaced an object with one that does not exist */
2097         if (!data && repl != sha1)
2098                 die("replacement %s not found for %s",
2099                     sha1_to_hex(repl), sha1_to_hex(sha1));
2100
2101         /* legacy behavior is to die on corrupted objects */
2102         if (!data && (has_loose_object(repl) || has_packed_and_bad(repl)))
2103                 die("object %s is corrupted", sha1_to_hex(repl));
2104
2105         if (replacement)
2106                 *replacement = repl;
2107
2108         return data;
2109 }
2110
2111 void *read_object_with_reference(const unsigned char *sha1,
2112                                  const char *required_type_name,
2113                                  unsigned long *size,
2114                                  unsigned char *actual_sha1_return)
2115 {
2116         enum object_type type, required_type;
2117         void *buffer;
2118         unsigned long isize;
2119         unsigned char actual_sha1[20];
2120
2121         required_type = type_from_string(required_type_name);
2122         hashcpy(actual_sha1, sha1);
2123         while (1) {
2124                 int ref_length = -1;
2125                 const char *ref_type = NULL;
2126
2127                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2128                 if (!buffer)
2129                         return NULL;
2130                 if (type == required_type) {
2131                         *size = isize;
2132                         if (actual_sha1_return)
2133                                 hashcpy(actual_sha1_return, actual_sha1);
2134                         return buffer;
2135                 }
2136                 /* Handle references */
2137                 else if (type == OBJ_COMMIT)
2138                         ref_type = "tree ";
2139                 else if (type == OBJ_TAG)
2140                         ref_type = "object ";
2141                 else {
2142                         free(buffer);
2143                         return NULL;
2144                 }
2145                 ref_length = strlen(ref_type);
2146
2147                 if (ref_length + 40 > isize ||
2148                     memcmp(buffer, ref_type, ref_length) ||
2149                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2150                         free(buffer);
2151                         return NULL;
2152                 }
2153                 free(buffer);
2154                 /* Now we have the ID of the referred-to object in
2155                  * actual_sha1.  Check again. */
2156         }
2157 }
2158
2159 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2160                                     const char *type, unsigned char *sha1,
2161                                     char *hdr, int *hdrlen)
2162 {
2163         git_SHA_CTX c;
2164
2165         /* Generate the header */
2166         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2167
2168         /* Sha1.. */
2169         git_SHA1_Init(&c);
2170         git_SHA1_Update(&c, hdr, *hdrlen);
2171         git_SHA1_Update(&c, buf, len);
2172         git_SHA1_Final(sha1, &c);
2173 }
2174
2175 /*
2176  * Move the just written object into its final resting place.
2177  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2178  * "moving" is only a part of what it does, when no patch between
2179  * master to pu changes the call sites of this function.
2180  */
2181 int move_temp_to_file(const char *tmpfile, const char *filename)
2182 {
2183         int ret = 0;
2184
2185         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2186                 goto try_rename;
2187         else if (link(tmpfile, filename))
2188                 ret = errno;
2189
2190         /*
2191          * Coda hack - coda doesn't like cross-directory links,
2192          * so we fall back to a rename, which will mean that it
2193          * won't be able to check collisions, but that's not a
2194          * big deal.
2195          *
2196          * The same holds for FAT formatted media.
2197          *
2198          * When this succeeds, we just return.  We have nothing
2199          * left to unlink.
2200          */
2201         if (ret && ret != EEXIST) {
2202         try_rename:
2203                 if (!rename(tmpfile, filename))
2204                         goto out;
2205                 ret = errno;
2206         }
2207         unlink_or_warn(tmpfile);
2208         if (ret) {
2209                 if (ret != EEXIST) {
2210                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2211                 }
2212                 /* FIXME!!! Collision check here ? */
2213         }
2214
2215 out:
2216         if (adjust_shared_perm(filename))
2217                 return error("unable to set permission to '%s'", filename);
2218         return 0;
2219 }
2220
2221 static int write_buffer(int fd, const void *buf, size_t len)
2222 {
2223         if (write_in_full(fd, buf, len) < 0)
2224                 return error("file write error (%s)", strerror(errno));
2225         return 0;
2226 }
2227
2228 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2229                    unsigned char *sha1)
2230 {
2231         char hdr[32];
2232         int hdrlen;
2233         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2234         return 0;
2235 }
2236
2237 /* Finalize a file on disk, and close it. */
2238 static void close_sha1_file(int fd)
2239 {
2240         if (fsync_object_files)
2241                 fsync_or_die(fd, "sha1 file");
2242         if (close(fd) != 0)
2243                 die_errno("error when closing sha1 file");
2244 }
2245
2246 /* Size of directory component, including the ending '/' */
2247 static inline int directory_size(const char *filename)
2248 {
2249         const char *s = strrchr(filename, '/');
2250         if (!s)
2251                 return 0;
2252         return s - filename + 1;
2253 }
2254
2255 /*
2256  * This creates a temporary file in the same directory as the final
2257  * 'filename'
2258  *
2259  * We want to avoid cross-directory filename renames, because those
2260  * can have problems on various filesystems (FAT, NFS, Coda).
2261  */
2262 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2263 {
2264         int fd, dirlen = directory_size(filename);
2265
2266         if (dirlen + 20 > bufsiz) {
2267                 errno = ENAMETOOLONG;
2268                 return -1;
2269         }
2270         memcpy(buffer, filename, dirlen);
2271         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2272         fd = git_mkstemp_mode(buffer, 0444);
2273         if (fd < 0 && dirlen && errno == ENOENT) {
2274                 /* Make sure the directory exists */
2275                 memcpy(buffer, filename, dirlen);
2276                 buffer[dirlen-1] = 0;
2277                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2278                         return -1;
2279
2280                 /* Try again */
2281                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2282                 fd = git_mkstemp_mode(buffer, 0444);
2283         }
2284         return fd;
2285 }
2286
2287 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2288                               void *buf, unsigned long len, time_t mtime)
2289 {
2290         int fd, ret;
2291         unsigned char compressed[4096];
2292         z_stream stream;
2293         git_SHA_CTX c;
2294         unsigned char parano_sha1[20];
2295         char *filename;
2296         static char tmpfile[PATH_MAX];
2297
2298         filename = sha1_file_name(sha1);
2299         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2300         while (fd < 0 && errno == EMFILE && unuse_one_window(packed_git, -1))
2301                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2302         if (fd < 0) {
2303                 if (errno == EACCES)
2304                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2305                 else
2306                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2307         }
2308
2309         /* Set it up */
2310         memset(&stream, 0, sizeof(stream));
2311         deflateInit(&stream, zlib_compression_level);
2312         stream.next_out = compressed;
2313         stream.avail_out = sizeof(compressed);
2314         git_SHA1_Init(&c);
2315
2316         /* First header.. */
2317         stream.next_in = (unsigned char *)hdr;
2318         stream.avail_in = hdrlen;
2319         while (deflate(&stream, 0) == Z_OK)
2320                 /* nothing */;
2321         git_SHA1_Update(&c, hdr, hdrlen);
2322
2323         /* Then the data itself.. */
2324         stream.next_in = buf;
2325         stream.avail_in = len;
2326         do {
2327                 unsigned char *in0 = stream.next_in;
2328                 ret = deflate(&stream, Z_FINISH);
2329                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2330                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2331                         die("unable to write sha1 file");
2332                 stream.next_out = compressed;
2333                 stream.avail_out = sizeof(compressed);
2334         } while (ret == Z_OK);
2335
2336         if (ret != Z_STREAM_END)
2337                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2338         ret = deflateEnd(&stream);
2339         if (ret != Z_OK)
2340                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2341         git_SHA1_Final(parano_sha1, &c);
2342         if (hashcmp(sha1, parano_sha1) != 0)
2343                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2344
2345         close_sha1_file(fd);
2346
2347         if (mtime) {
2348                 struct utimbuf utb;
2349                 utb.actime = mtime;
2350                 utb.modtime = mtime;
2351                 if (utime(tmpfile, &utb) < 0)
2352                         warning("failed utime() on %s: %s",
2353                                 tmpfile, strerror(errno));
2354         }
2355
2356         return move_temp_to_file(tmpfile, filename);
2357 }
2358
2359 int write_sha1_file(void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2360 {
2361         unsigned char sha1[20];
2362         char hdr[32];
2363         int hdrlen;
2364
2365         /* Normally if we have it in the pack then we do not bother writing
2366          * it out into .git/objects/??/?{38} file.
2367          */
2368         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2369         if (returnsha1)
2370                 hashcpy(returnsha1, sha1);
2371         if (has_sha1_file(sha1))
2372                 return 0;
2373         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2374 }
2375
2376 int force_object_loose(const unsigned char *sha1, time_t mtime)
2377 {
2378         void *buf;
2379         unsigned long len;
2380         enum object_type type;
2381         char hdr[32];
2382         int hdrlen;
2383         int ret;
2384
2385         if (has_loose_object(sha1))
2386                 return 0;
2387         buf = read_packed_sha1(sha1, &type, &len);
2388         if (!buf)
2389                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2390         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2391         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2392         free(buf);
2393
2394         return ret;
2395 }
2396
2397 int has_pack_index(const unsigned char *sha1)
2398 {
2399         struct stat st;
2400         if (stat(sha1_pack_index_name(sha1), &st))
2401                 return 0;
2402         return 1;
2403 }
2404
2405 int has_sha1_pack(const unsigned char *sha1)
2406 {
2407         struct pack_entry e;
2408         return find_pack_entry(sha1, &e);
2409 }
2410
2411 int has_sha1_file(const unsigned char *sha1)
2412 {
2413         struct pack_entry e;
2414
2415         if (find_pack_entry(sha1, &e))
2416                 return 1;
2417         return has_loose_object(sha1);
2418 }
2419
2420 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2421                      int write_object, enum object_type type, const char *path)
2422 {
2423         int ret, re_allocated = 0;
2424
2425         if (!type)
2426                 type = OBJ_BLOB;
2427
2428         /*
2429          * Convert blobs to git internal format
2430          */
2431         if ((type == OBJ_BLOB) && path) {
2432                 struct strbuf nbuf = STRBUF_INIT;
2433                 if (convert_to_git(path, buf, size, &nbuf,
2434                                    write_object ? safe_crlf : 0)) {
2435                         buf = strbuf_detach(&nbuf, &size);
2436                         re_allocated = 1;
2437                 }
2438         }
2439
2440         if (write_object)
2441                 ret = write_sha1_file(buf, size, typename(type), sha1);
2442         else
2443                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2444         if (re_allocated)
2445                 free(buf);
2446         return ret;
2447 }
2448
2449 #define SMALL_FILE_SIZE (32*1024)
2450
2451 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2452              enum object_type type, const char *path)
2453 {
2454         int ret;
2455         size_t size = xsize_t(st->st_size);
2456
2457         if (!S_ISREG(st->st_mode)) {
2458                 struct strbuf sbuf = STRBUF_INIT;
2459                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2460                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2461                                         type, path);
2462                 else
2463                         ret = -1;
2464                 strbuf_release(&sbuf);
2465         } else if (size <= SMALL_FILE_SIZE) {
2466                 char *buf = xmalloc(size);
2467                 if (size == read_in_full(fd, buf, size))
2468                         ret = index_mem(sha1, buf, size, write_object, type,
2469                                         path);
2470                 else
2471                         ret = error("short read %s", strerror(errno));
2472                 free(buf);
2473         } else if (size) {
2474                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2475                 ret = index_mem(sha1, buf, size, write_object, type, path);
2476                 munmap(buf, size);
2477         } else
2478                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2479         close(fd);
2480         return ret;
2481 }
2482
2483 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2484 {
2485         int fd;
2486         struct strbuf sb = STRBUF_INIT;
2487
2488         switch (st->st_mode & S_IFMT) {
2489         case S_IFREG:
2490                 fd = open(path, O_RDONLY);
2491                 if (fd < 0)
2492                         return error("open(\"%s\"): %s", path,
2493                                      strerror(errno));
2494                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2495                         return error("%s: failed to insert into database",
2496                                      path);
2497                 break;
2498         case S_IFLNK:
2499                 if (strbuf_readlink(&sb, path, st->st_size)) {
2500                         char *errstr = strerror(errno);
2501                         return error("readlink(\"%s\"): %s", path,
2502                                      errstr);
2503                 }
2504                 if (!write_object)
2505                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2506                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2507                         return error("%s: failed to insert into database",
2508                                      path);
2509                 strbuf_release(&sb);
2510                 break;
2511         case S_IFDIR:
2512                 return resolve_gitlink_ref(path, "HEAD", sha1);
2513         default:
2514                 return error("%s: unsupported file type", path);
2515         }
2516         return 0;
2517 }
2518
2519 int read_pack_header(int fd, struct pack_header *header)
2520 {
2521         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2522                 /* "eof before pack header was fully read" */
2523                 return PH_ERROR_EOF;
2524
2525         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2526                 /* "protocol error (pack signature mismatch detected)" */
2527                 return PH_ERROR_PACK_SIGNATURE;
2528         if (!pack_version_ok(header->hdr_version))
2529                 /* "protocol error (pack version unsupported)" */
2530                 return PH_ERROR_PROTOCOL;
2531         return 0;
2532 }