Merge branch 'nd/maint-hide-checkout-index-from-error'
[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 int git_open_noatime(const char *name, struct packed_git *p);
39
40 int safe_create_leading_directories(char *path)
41 {
42         char *pos = path + offset_1st_component(path);
43         struct stat st;
44
45         while (pos) {
46                 pos = strchr(pos, '/');
47                 if (!pos)
48                         break;
49                 while (*++pos == '/')
50                         ;
51                 if (!*pos)
52                         break;
53                 *--pos = '\0';
54                 if (!stat(path, &st)) {
55                         /* path exists */
56                         if (!S_ISDIR(st.st_mode)) {
57                                 *pos = '/';
58                                 return -3;
59                         }
60                 }
61                 else if (mkdir(path, 0777)) {
62                         *pos = '/';
63                         return -1;
64                 }
65                 else if (adjust_shared_perm(path)) {
66                         *pos = '/';
67                         return -2;
68                 }
69                 *pos++ = '/';
70         }
71         return 0;
72 }
73
74 int safe_create_leading_directories_const(const char *path)
75 {
76         /* path points to cache entries, so xstrdup before messing with it */
77         char *buf = xstrdup(path);
78         int result = safe_create_leading_directories(buf);
79         free(buf);
80         return result;
81 }
82
83 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
84 {
85         int i;
86         for (i = 0; i < 20; i++) {
87                 static char hex[] = "0123456789abcdef";
88                 unsigned int val = sha1[i];
89                 char *pos = pathbuf + i*2 + (i > 0);
90                 *pos++ = hex[val >> 4];
91                 *pos = hex[val & 0xf];
92         }
93 }
94
95 /*
96  * NOTE! This returns a statically allocated buffer, so you have to be
97  * careful about using it. Do an "xstrdup()" if you need to save the
98  * filename.
99  *
100  * Also note that this returns the location for creating.  Reading
101  * SHA1 file can happen from any alternate directory listed in the
102  * DB_ENVIRONMENT environment variable if it is not found in
103  * the primary object database.
104  */
105 char *sha1_file_name(const unsigned char *sha1)
106 {
107         static char buf[PATH_MAX];
108         const char *objdir;
109         int len;
110
111         objdir = get_object_directory();
112         len = strlen(objdir);
113
114         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
115         if (len + 43 > PATH_MAX)
116                 die("insanely long object directory %s", objdir);
117         memcpy(buf, objdir, len);
118         buf[len] = '/';
119         buf[len+3] = '/';
120         buf[len+42] = '\0';
121         fill_sha1_path(buf + len + 1, sha1);
122         return buf;
123 }
124
125 static char *sha1_get_pack_name(const unsigned char *sha1,
126                                 char **name, char **base, const char *which)
127 {
128         static const char hex[] = "0123456789abcdef";
129         char *buf;
130         int i;
131
132         if (!*base) {
133                 const char *sha1_file_directory = get_object_directory();
134                 int len = strlen(sha1_file_directory);
135                 *base = xmalloc(len + 60);
136                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
137                         sha1_file_directory, which);
138                 *name = *base + len + 11;
139         }
140
141         buf = *name;
142
143         for (i = 0; i < 20; i++) {
144                 unsigned int val = *sha1++;
145                 *buf++ = hex[val >> 4];
146                 *buf++ = hex[val & 0xf];
147         }
148
149         return *base;
150 }
151
152 char *sha1_pack_name(const unsigned char *sha1)
153 {
154         static char *name, *base;
155
156         return sha1_get_pack_name(sha1, &name, &base, "pack");
157 }
158
159 char *sha1_pack_index_name(const unsigned char *sha1)
160 {
161         static char *name, *base;
162
163         return sha1_get_pack_name(sha1, &name, &base, "idx");
164 }
165
166 struct alternate_object_database *alt_odb_list;
167 static struct alternate_object_database **alt_odb_tail;
168
169 static void read_info_alternates(const char * alternates, int depth);
170
171 /*
172  * Prepare alternate object database registry.
173  *
174  * The variable alt_odb_list points at the list of struct
175  * alternate_object_database.  The elements on this list come from
176  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
177  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
178  * whose contents is similar to that environment variable but can be
179  * LF separated.  Its base points at a statically allocated buffer that
180  * contains "/the/directory/corresponding/to/.git/objects/...", while
181  * its name points just after the slash at the end of ".git/objects/"
182  * in the example above, and has enough space to hold 40-byte hex
183  * SHA1, an extra slash for the first level indirection, and the
184  * terminating NUL.
185  */
186 static int link_alt_odb_entry(const char * entry, int len, const char * relative_base, int depth)
187 {
188         const char *objdir = get_object_directory();
189         struct alternate_object_database *ent;
190         struct alternate_object_database *alt;
191         /* 43 = 40-byte + 2 '/' + terminating NUL */
192         int pfxlen = len;
193         int entlen = pfxlen + 43;
194         int base_len = -1;
195
196         if (!is_absolute_path(entry) && relative_base) {
197                 /* Relative alt-odb */
198                 if (base_len < 0)
199                         base_len = strlen(relative_base) + 1;
200                 entlen += base_len;
201                 pfxlen += base_len;
202         }
203         ent = xmalloc(sizeof(*ent) + entlen);
204
205         if (!is_absolute_path(entry) && relative_base) {
206                 memcpy(ent->base, relative_base, base_len - 1);
207                 ent->base[base_len - 1] = '/';
208                 memcpy(ent->base + base_len, entry, len);
209         }
210         else
211                 memcpy(ent->base, entry, pfxlen);
212
213         ent->name = ent->base + pfxlen + 1;
214         ent->base[pfxlen + 3] = '/';
215         ent->base[pfxlen] = ent->base[entlen-1] = 0;
216
217         /* Detect cases where alternate disappeared */
218         if (!is_directory(ent->base)) {
219                 error("object directory %s does not exist; "
220                       "check .git/objects/info/alternates.",
221                       ent->base);
222                 free(ent);
223                 return -1;
224         }
225
226         /* Prevent the common mistake of listing the same
227          * thing twice, or object directory itself.
228          */
229         for (alt = alt_odb_list; alt; alt = alt->next) {
230                 if (!memcmp(ent->base, alt->base, pfxlen)) {
231                         free(ent);
232                         return -1;
233                 }
234         }
235         if (!memcmp(ent->base, objdir, pfxlen)) {
236                 free(ent);
237                 return -1;
238         }
239
240         /* add the alternate entry */
241         *alt_odb_tail = ent;
242         alt_odb_tail = &(ent->next);
243         ent->next = NULL;
244
245         /* recursively add alternates */
246         read_info_alternates(ent->base, depth + 1);
247
248         ent->base[pfxlen] = '/';
249
250         return 0;
251 }
252
253 static void link_alt_odb_entries(const char *alt, const char *ep, int sep,
254                                  const char *relative_base, int depth)
255 {
256         const char *cp, *last;
257
258         if (depth > 5) {
259                 error("%s: ignoring alternate object stores, nesting too deep.",
260                                 relative_base);
261                 return;
262         }
263
264         last = alt;
265         while (last < ep) {
266                 cp = last;
267                 if (cp < ep && *cp == '#') {
268                         while (cp < ep && *cp != sep)
269                                 cp++;
270                         last = cp + 1;
271                         continue;
272                 }
273                 while (cp < ep && *cp != sep)
274                         cp++;
275                 if (last != cp) {
276                         if (!is_absolute_path(last) && depth) {
277                                 error("%s: ignoring relative alternate object store %s",
278                                                 relative_base, last);
279                         } else {
280                                 link_alt_odb_entry(last, cp - last,
281                                                 relative_base, depth);
282                         }
283                 }
284                 while (cp < ep && *cp == sep)
285                         cp++;
286                 last = cp;
287         }
288 }
289
290 static void read_info_alternates(const char * relative_base, int depth)
291 {
292         char *map;
293         size_t mapsz;
294         struct stat st;
295         const char alt_file_name[] = "info/alternates";
296         /* Given that relative_base is no longer than PATH_MAX,
297            ensure that "path" has enough space to append "/", the
298            file name, "info/alternates", and a trailing NUL.  */
299         char path[PATH_MAX + 1 + sizeof alt_file_name];
300         int fd;
301
302         sprintf(path, "%s/%s", relative_base, alt_file_name);
303         fd = git_open_noatime(path, NULL);
304         if (fd < 0)
305                 return;
306         if (fstat(fd, &st) || (st.st_size == 0)) {
307                 close(fd);
308                 return;
309         }
310         mapsz = xsize_t(st.st_size);
311         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
312         close(fd);
313
314         link_alt_odb_entries(map, map + mapsz, '\n', relative_base, depth);
315
316         munmap(map, mapsz);
317 }
318
319 void add_to_alternates_file(const char *reference)
320 {
321         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
322         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
323         char *alt = mkpath("%s/objects\n", reference);
324         write_or_die(fd, alt, strlen(alt));
325         if (commit_lock_file(lock))
326                 die("could not close alternates file");
327         if (alt_odb_tail)
328                 link_alt_odb_entries(alt, alt + strlen(alt), '\n', NULL, 0);
329 }
330
331 void foreach_alt_odb(alt_odb_fn fn, void *cb)
332 {
333         struct alternate_object_database *ent;
334
335         prepare_alt_odb();
336         for (ent = alt_odb_list; ent; ent = ent->next)
337                 if (fn(ent, cb))
338                         return;
339 }
340
341 void prepare_alt_odb(void)
342 {
343         const char *alt;
344
345         if (alt_odb_tail)
346                 return;
347
348         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
349         if (!alt) alt = "";
350
351         alt_odb_tail = &alt_odb_list;
352         link_alt_odb_entries(alt, alt + strlen(alt), PATH_SEP, NULL, 0);
353
354         read_info_alternates(get_object_directory(), 0);
355 }
356
357 static int has_loose_object_local(const unsigned char *sha1)
358 {
359         char *name = sha1_file_name(sha1);
360         return !access(name, F_OK);
361 }
362
363 int has_loose_object_nonlocal(const unsigned char *sha1)
364 {
365         struct alternate_object_database *alt;
366         prepare_alt_odb();
367         for (alt = alt_odb_list; alt; alt = alt->next) {
368                 fill_sha1_path(alt->name, sha1);
369                 if (!access(alt->base, F_OK))
370                         return 1;
371         }
372         return 0;
373 }
374
375 static int has_loose_object(const unsigned char *sha1)
376 {
377         return has_loose_object_local(sha1) ||
378                has_loose_object_nonlocal(sha1);
379 }
380
381 static unsigned int pack_used_ctr;
382 static unsigned int pack_mmap_calls;
383 static unsigned int peak_pack_open_windows;
384 static unsigned int pack_open_windows;
385 static size_t peak_pack_mapped;
386 static size_t pack_mapped;
387 struct packed_git *packed_git;
388
389 void pack_report(void)
390 {
391         fprintf(stderr,
392                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
393                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
394                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
395                 sz_fmt(getpagesize()),
396                 sz_fmt(packed_git_window_size),
397                 sz_fmt(packed_git_limit));
398         fprintf(stderr,
399                 "pack_report: pack_used_ctr            = %10u\n"
400                 "pack_report: pack_mmap_calls          = %10u\n"
401                 "pack_report: pack_open_windows        = %10u / %10u\n"
402                 "pack_report: pack_mapped              = "
403                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
404                 pack_used_ctr,
405                 pack_mmap_calls,
406                 pack_open_windows, peak_pack_open_windows,
407                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
408 }
409
410 static int check_packed_git_idx(const char *path,  struct packed_git *p)
411 {
412         void *idx_map;
413         struct pack_idx_header *hdr;
414         size_t idx_size;
415         uint32_t version, nr, i, *index;
416         int fd = git_open_noatime(path, p);
417         struct stat st;
418
419         if (fd < 0)
420                 return -1;
421         if (fstat(fd, &st)) {
422                 close(fd);
423                 return -1;
424         }
425         idx_size = xsize_t(st.st_size);
426         if (idx_size < 4 * 256 + 20 + 20) {
427                 close(fd);
428                 return error("index file %s is too small", path);
429         }
430         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
431         close(fd);
432
433         hdr = idx_map;
434         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
435                 version = ntohl(hdr->idx_version);
436                 if (version < 2 || version > 2) {
437                         munmap(idx_map, idx_size);
438                         return error("index file %s is version %"PRIu32
439                                      " and is not supported by this binary"
440                                      " (try upgrading GIT to a newer version)",
441                                      path, version);
442                 }
443         } else
444                 version = 1;
445
446         nr = 0;
447         index = idx_map;
448         if (version > 1)
449                 index += 2;  /* skip index header */
450         for (i = 0; i < 256; i++) {
451                 uint32_t n = ntohl(index[i]);
452                 if (n < nr) {
453                         munmap(idx_map, idx_size);
454                         return error("non-monotonic index %s", path);
455                 }
456                 nr = n;
457         }
458
459         if (version == 1) {
460                 /*
461                  * Total size:
462                  *  - 256 index entries 4 bytes each
463                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
464                  *  - 20-byte SHA1 of the packfile
465                  *  - 20-byte SHA1 file checksum
466                  */
467                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
468                         munmap(idx_map, idx_size);
469                         return error("wrong index v1 file size in %s", path);
470                 }
471         } else if (version == 2) {
472                 /*
473                  * Minimum size:
474                  *  - 8 bytes of header
475                  *  - 256 index entries 4 bytes each
476                  *  - 20-byte sha1 entry * nr
477                  *  - 4-byte crc entry * nr
478                  *  - 4-byte offset entry * nr
479                  *  - 20-byte SHA1 of the packfile
480                  *  - 20-byte SHA1 file checksum
481                  * And after the 4-byte offset table might be a
482                  * variable sized table containing 8-byte entries
483                  * for offsets larger than 2^31.
484                  */
485                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
486                 unsigned long max_size = min_size;
487                 if (nr)
488                         max_size += (nr - 1)*8;
489                 if (idx_size < min_size || idx_size > max_size) {
490                         munmap(idx_map, idx_size);
491                         return error("wrong index v2 file size in %s", path);
492                 }
493                 if (idx_size != min_size &&
494                     /*
495                      * make sure we can deal with large pack offsets.
496                      * 31-bit signed offset won't be enough, neither
497                      * 32-bit unsigned one will be.
498                      */
499                     (sizeof(off_t) <= 4)) {
500                         munmap(idx_map, idx_size);
501                         return error("pack too large for current definition of off_t in %s", path);
502                 }
503         }
504
505         p->index_version = version;
506         p->index_data = idx_map;
507         p->index_size = idx_size;
508         p->num_objects = nr;
509         return 0;
510 }
511
512 int open_pack_index(struct packed_git *p)
513 {
514         char *idx_name;
515         int ret;
516
517         if (p->index_data)
518                 return 0;
519
520         idx_name = xstrdup(p->pack_name);
521         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
522         ret = check_packed_git_idx(idx_name, p);
523         free(idx_name);
524         return ret;
525 }
526
527 static void scan_windows(struct packed_git *p,
528         struct packed_git **lru_p,
529         struct pack_window **lru_w,
530         struct pack_window **lru_l)
531 {
532         struct pack_window *w, *w_l;
533
534         for (w_l = NULL, w = p->windows; w; w = w->next) {
535                 if (!w->inuse_cnt) {
536                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
537                                 *lru_p = p;
538                                 *lru_w = w;
539                                 *lru_l = w_l;
540                         }
541                 }
542                 w_l = w;
543         }
544 }
545
546 static int unuse_one_window(struct packed_git *current, int keep_fd)
547 {
548         struct packed_git *p, *lru_p = NULL;
549         struct pack_window *lru_w = NULL, *lru_l = NULL;
550
551         if (current)
552                 scan_windows(current, &lru_p, &lru_w, &lru_l);
553         for (p = packed_git; p; p = p->next)
554                 scan_windows(p, &lru_p, &lru_w, &lru_l);
555         if (lru_p) {
556                 munmap(lru_w->base, lru_w->len);
557                 pack_mapped -= lru_w->len;
558                 if (lru_l)
559                         lru_l->next = lru_w->next;
560                 else {
561                         lru_p->windows = lru_w->next;
562                         if (!lru_p->windows && lru_p->pack_fd != keep_fd) {
563                                 close(lru_p->pack_fd);
564                                 lru_p->pack_fd = -1;
565                         }
566                 }
567                 free(lru_w);
568                 pack_open_windows--;
569                 return 1;
570         }
571         return 0;
572 }
573
574 void release_pack_memory(size_t need, int fd)
575 {
576         size_t cur = pack_mapped;
577         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
578                 ; /* nothing */
579 }
580
581 void *xmmap(void *start, size_t length,
582         int prot, int flags, int fd, off_t offset)
583 {
584         void *ret = mmap(start, length, prot, flags, fd, offset);
585         if (ret == MAP_FAILED) {
586                 if (!length)
587                         return NULL;
588                 release_pack_memory(length, fd);
589                 ret = mmap(start, length, prot, flags, fd, offset);
590                 if (ret == MAP_FAILED)
591                         die_errno("Out of memory? mmap failed");
592         }
593         return ret;
594 }
595
596 void close_pack_windows(struct packed_git *p)
597 {
598         while (p->windows) {
599                 struct pack_window *w = p->windows;
600
601                 if (w->inuse_cnt)
602                         die("pack '%s' still has open windows to it",
603                             p->pack_name);
604                 munmap(w->base, w->len);
605                 pack_mapped -= w->len;
606                 pack_open_windows--;
607                 p->windows = w->next;
608                 free(w);
609         }
610 }
611
612 void unuse_pack(struct pack_window **w_cursor)
613 {
614         struct pack_window *w = *w_cursor;
615         if (w) {
616                 w->inuse_cnt--;
617                 *w_cursor = NULL;
618         }
619 }
620
621 void close_pack_index(struct packed_git *p)
622 {
623         if (p->index_data) {
624                 munmap((void *)p->index_data, p->index_size);
625                 p->index_data = NULL;
626         }
627 }
628
629 /*
630  * This is used by git-repack in case a newly created pack happens to
631  * contain the same set of objects as an existing one.  In that case
632  * the resulting file might be different even if its name would be the
633  * same.  It is best to close any reference to the old pack before it is
634  * replaced on disk.  Of course no index pointers nor windows for given pack
635  * must subsist at this point.  If ever objects from this pack are requested
636  * again, the new version of the pack will be reinitialized through
637  * reprepare_packed_git().
638  */
639 void free_pack_by_name(const char *pack_name)
640 {
641         struct packed_git *p, **pp = &packed_git;
642
643         while (*pp) {
644                 p = *pp;
645                 if (strcmp(pack_name, p->pack_name) == 0) {
646                         clear_delta_base_cache();
647                         close_pack_windows(p);
648                         if (p->pack_fd != -1)
649                                 close(p->pack_fd);
650                         close_pack_index(p);
651                         free(p->bad_object_sha1);
652                         *pp = p->next;
653                         free(p);
654                         return;
655                 }
656                 pp = &p->next;
657         }
658 }
659
660 /*
661  * Do not call this directly as this leaks p->pack_fd on error return;
662  * call open_packed_git() instead.
663  */
664 static int open_packed_git_1(struct packed_git *p)
665 {
666         struct stat st;
667         struct pack_header hdr;
668         unsigned char sha1[20];
669         unsigned char *idx_sha1;
670         long fd_flag;
671
672         if (!p->index_data && open_pack_index(p))
673                 return error("packfile %s index unavailable", p->pack_name);
674
675         p->pack_fd = git_open_noatime(p->pack_name, p);
676         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
677                 return -1;
678
679         /* If we created the struct before we had the pack we lack size. */
680         if (!p->pack_size) {
681                 if (!S_ISREG(st.st_mode))
682                         return error("packfile %s not a regular file", p->pack_name);
683                 p->pack_size = st.st_size;
684         } else if (p->pack_size != st.st_size)
685                 return error("packfile %s size changed", p->pack_name);
686
687         /* We leave these file descriptors open with sliding mmap;
688          * there is no point keeping them open across exec(), though.
689          */
690         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
691         if (fd_flag < 0)
692                 return error("cannot determine file descriptor flags");
693         fd_flag |= FD_CLOEXEC;
694         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
695                 return error("cannot set FD_CLOEXEC");
696
697         /* Verify we recognize this pack file format. */
698         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
699                 return error("file %s is far too short to be a packfile", p->pack_name);
700         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
701                 return error("file %s is not a GIT packfile", p->pack_name);
702         if (!pack_version_ok(hdr.hdr_version))
703                 return error("packfile %s is version %"PRIu32" and not"
704                         " supported (try upgrading GIT to a newer version)",
705                         p->pack_name, ntohl(hdr.hdr_version));
706
707         /* Verify the pack matches its index. */
708         if (p->num_objects != ntohl(hdr.hdr_entries))
709                 return error("packfile %s claims to have %"PRIu32" objects"
710                              " while index indicates %"PRIu32" objects",
711                              p->pack_name, ntohl(hdr.hdr_entries),
712                              p->num_objects);
713         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
714                 return error("end of packfile %s is unavailable", p->pack_name);
715         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
716                 return error("packfile %s signature is unavailable", p->pack_name);
717         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
718         if (hashcmp(sha1, idx_sha1))
719                 return error("packfile %s does not match index", p->pack_name);
720         return 0;
721 }
722
723 static int open_packed_git(struct packed_git *p)
724 {
725         if (!open_packed_git_1(p))
726                 return 0;
727         if (p->pack_fd != -1) {
728                 close(p->pack_fd);
729                 p->pack_fd = -1;
730         }
731         return -1;
732 }
733
734 static int in_window(struct pack_window *win, off_t offset)
735 {
736         /* We must promise at least 20 bytes (one hash) after the
737          * offset is available from this window, otherwise the offset
738          * is not actually in this window and a different window (which
739          * has that one hash excess) must be used.  This is to support
740          * the object header and delta base parsing routines below.
741          */
742         off_t win_off = win->offset;
743         return win_off <= offset
744                 && (offset + 20) <= (win_off + win->len);
745 }
746
747 unsigned char *use_pack(struct packed_git *p,
748                 struct pack_window **w_cursor,
749                 off_t offset,
750                 unsigned int *left)
751 {
752         struct pack_window *win = *w_cursor;
753
754         if (p->pack_fd == -1 && open_packed_git(p))
755                 die("packfile %s cannot be accessed", p->pack_name);
756
757         /* Since packfiles end in a hash of their content and it's
758          * pointless to ask for an offset into the middle of that
759          * hash, and the in_window function above wouldn't match
760          * don't allow an offset too close to the end of the file.
761          */
762         if (offset > (p->pack_size - 20))
763                 die("offset beyond end of packfile (truncated pack?)");
764
765         if (!win || !in_window(win, offset)) {
766                 if (win)
767                         win->inuse_cnt--;
768                 for (win = p->windows; win; win = win->next) {
769                         if (in_window(win, offset))
770                                 break;
771                 }
772                 if (!win) {
773                         size_t window_align = packed_git_window_size / 2;
774                         off_t len;
775                         win = xcalloc(1, sizeof(*win));
776                         win->offset = (offset / window_align) * window_align;
777                         len = p->pack_size - win->offset;
778                         if (len > packed_git_window_size)
779                                 len = packed_git_window_size;
780                         win->len = (size_t)len;
781                         pack_mapped += win->len;
782                         while (packed_git_limit < pack_mapped
783                                 && unuse_one_window(p, p->pack_fd))
784                                 ; /* nothing */
785                         win->base = xmmap(NULL, win->len,
786                                 PROT_READ, MAP_PRIVATE,
787                                 p->pack_fd, win->offset);
788                         if (win->base == MAP_FAILED)
789                                 die("packfile %s cannot be mapped: %s",
790                                         p->pack_name,
791                                         strerror(errno));
792                         pack_mmap_calls++;
793                         pack_open_windows++;
794                         if (pack_mapped > peak_pack_mapped)
795                                 peak_pack_mapped = pack_mapped;
796                         if (pack_open_windows > peak_pack_open_windows)
797                                 peak_pack_open_windows = pack_open_windows;
798                         win->next = p->windows;
799                         p->windows = win;
800                 }
801         }
802         if (win != *w_cursor) {
803                 win->last_used = pack_used_ctr++;
804                 win->inuse_cnt++;
805                 *w_cursor = win;
806         }
807         offset -= win->offset;
808         if (left)
809                 *left = win->len - xsize_t(offset);
810         return win->base + offset;
811 }
812
813 static struct packed_git *alloc_packed_git(int extra)
814 {
815         struct packed_git *p = xmalloc(sizeof(*p) + extra);
816         memset(p, 0, sizeof(*p));
817         p->pack_fd = -1;
818         return p;
819 }
820
821 static void try_to_free_pack_memory(size_t size)
822 {
823         release_pack_memory(size, -1);
824 }
825
826 struct packed_git *add_packed_git(const char *path, int path_len, int local)
827 {
828         static int have_set_try_to_free_routine;
829         struct stat st;
830         struct packed_git *p = alloc_packed_git(path_len + 2);
831
832         if (!have_set_try_to_free_routine) {
833                 have_set_try_to_free_routine = 1;
834                 set_try_to_free_routine(try_to_free_pack_memory);
835         }
836
837         /*
838          * Make sure a corresponding .pack file exists and that
839          * the index looks sane.
840          */
841         path_len -= strlen(".idx");
842         if (path_len < 1) {
843                 free(p);
844                 return NULL;
845         }
846         memcpy(p->pack_name, path, path_len);
847
848         strcpy(p->pack_name + path_len, ".keep");
849         if (!access(p->pack_name, F_OK))
850                 p->pack_keep = 1;
851
852         strcpy(p->pack_name + path_len, ".pack");
853         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
854                 free(p);
855                 return NULL;
856         }
857
858         /* ok, it looks sane as far as we can check without
859          * actually mapping the pack file.
860          */
861         p->pack_size = st.st_size;
862         p->pack_local = local;
863         p->mtime = st.st_mtime;
864         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
865                 hashclr(p->sha1);
866         return p;
867 }
868
869 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
870 {
871         const char *path = sha1_pack_name(sha1);
872         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
873
874         strcpy(p->pack_name, path);
875         hashcpy(p->sha1, sha1);
876         if (check_packed_git_idx(idx_path, p)) {
877                 free(p);
878                 return NULL;
879         }
880
881         return p;
882 }
883
884 void install_packed_git(struct packed_git *pack)
885 {
886         pack->next = packed_git;
887         packed_git = pack;
888 }
889
890 static void prepare_packed_git_one(char *objdir, int local)
891 {
892         /* Ensure that this buffer is large enough so that we can
893            append "/pack/" without clobbering the stack even if
894            strlen(objdir) were PATH_MAX.  */
895         char path[PATH_MAX + 1 + 4 + 1 + 1];
896         int len;
897         DIR *dir;
898         struct dirent *de;
899
900         sprintf(path, "%s/pack", objdir);
901         len = strlen(path);
902         dir = opendir(path);
903         while (!dir && errno == EMFILE && unuse_one_window(NULL, -1))
904                 dir = opendir(path);
905         if (!dir) {
906                 if (errno != ENOENT)
907                         error("unable to open object pack directory: %s: %s",
908                               path, strerror(errno));
909                 return;
910         }
911         path[len++] = '/';
912         while ((de = readdir(dir)) != NULL) {
913                 int namelen = strlen(de->d_name);
914                 struct packed_git *p;
915
916                 if (!has_extension(de->d_name, ".idx"))
917                         continue;
918
919                 if (len + namelen + 1 > sizeof(path))
920                         continue;
921
922                 /* Don't reopen a pack we already have. */
923                 strcpy(path + len, de->d_name);
924                 for (p = packed_git; p; p = p->next) {
925                         if (!memcmp(path, p->pack_name, len + namelen - 4))
926                                 break;
927                 }
928                 if (p)
929                         continue;
930                 /* See if it really is a valid .idx file with corresponding
931                  * .pack file that we can map.
932                  */
933                 p = add_packed_git(path, len + namelen, local);
934                 if (!p)
935                         continue;
936                 install_packed_git(p);
937         }
938         closedir(dir);
939 }
940
941 static int sort_pack(const void *a_, const void *b_)
942 {
943         struct packed_git *a = *((struct packed_git **)a_);
944         struct packed_git *b = *((struct packed_git **)b_);
945         int st;
946
947         /*
948          * Local packs tend to contain objects specific to our
949          * variant of the project than remote ones.  In addition,
950          * remote ones could be on a network mounted filesystem.
951          * Favor local ones for these reasons.
952          */
953         st = a->pack_local - b->pack_local;
954         if (st)
955                 return -st;
956
957         /*
958          * Younger packs tend to contain more recent objects,
959          * and more recent objects tend to get accessed more
960          * often.
961          */
962         if (a->mtime < b->mtime)
963                 return 1;
964         else if (a->mtime == b->mtime)
965                 return 0;
966         return -1;
967 }
968
969 static void rearrange_packed_git(void)
970 {
971         struct packed_git **ary, *p;
972         int i, n;
973
974         for (n = 0, p = packed_git; p; p = p->next)
975                 n++;
976         if (n < 2)
977                 return;
978
979         /* prepare an array of packed_git for easier sorting */
980         ary = xcalloc(n, sizeof(struct packed_git *));
981         for (n = 0, p = packed_git; p; p = p->next)
982                 ary[n++] = p;
983
984         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
985
986         /* link them back again */
987         for (i = 0; i < n - 1; i++)
988                 ary[i]->next = ary[i + 1];
989         ary[n - 1]->next = NULL;
990         packed_git = ary[0];
991
992         free(ary);
993 }
994
995 static int prepare_packed_git_run_once = 0;
996 void prepare_packed_git(void)
997 {
998         struct alternate_object_database *alt;
999
1000         if (prepare_packed_git_run_once)
1001                 return;
1002         prepare_packed_git_one(get_object_directory(), 1);
1003         prepare_alt_odb();
1004         for (alt = alt_odb_list; alt; alt = alt->next) {
1005                 alt->name[-1] = 0;
1006                 prepare_packed_git_one(alt->base, 0);
1007                 alt->name[-1] = '/';
1008         }
1009         rearrange_packed_git();
1010         prepare_packed_git_run_once = 1;
1011 }
1012
1013 void reprepare_packed_git(void)
1014 {
1015         discard_revindex();
1016         prepare_packed_git_run_once = 0;
1017         prepare_packed_git();
1018 }
1019
1020 static void mark_bad_packed_object(struct packed_git *p,
1021                                    const unsigned char *sha1)
1022 {
1023         unsigned i;
1024         for (i = 0; i < p->num_bad_objects; i++)
1025                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1026                         return;
1027         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1028         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1029         p->num_bad_objects++;
1030 }
1031
1032 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1033 {
1034         struct packed_git *p;
1035         unsigned i;
1036
1037         for (p = packed_git; p; p = p->next)
1038                 for (i = 0; i < p->num_bad_objects; i++)
1039                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1040                                 return p;
1041         return NULL;
1042 }
1043
1044 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1045 {
1046         unsigned char real_sha1[20];
1047         hash_sha1_file(map, size, type, real_sha1);
1048         return hashcmp(sha1, real_sha1) ? -1 : 0;
1049 }
1050
1051 static int git_open_noatime(const char *name, struct packed_git *p)
1052 {
1053         static int sha1_file_open_flag = O_NOATIME;
1054
1055         for (;;) {
1056                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1057                 if (fd >= 0)
1058                         return fd;
1059
1060                 /* Might the failure be insufficient file descriptors? */
1061                 if (errno == EMFILE) {
1062                         if (unuse_one_window(p, -1))
1063                                 continue;
1064                         else
1065                                 return -1;
1066                 }
1067
1068                 /* Might the failure be due to O_NOATIME? */
1069                 if (errno != ENOENT && sha1_file_open_flag) {
1070                         sha1_file_open_flag = 0;
1071                         continue;
1072                 }
1073
1074                 return -1;
1075         }
1076 }
1077
1078 static int open_sha1_file(const unsigned char *sha1)
1079 {
1080         int fd;
1081         char *name = sha1_file_name(sha1);
1082         struct alternate_object_database *alt;
1083
1084         fd = git_open_noatime(name, NULL);
1085         if (fd >= 0)
1086                 return fd;
1087
1088         prepare_alt_odb();
1089         errno = ENOENT;
1090         for (alt = alt_odb_list; alt; alt = alt->next) {
1091                 name = alt->name;
1092                 fill_sha1_path(name, sha1);
1093                 fd = git_open_noatime(alt->base, NULL);
1094                 if (fd >= 0)
1095                         return fd;
1096         }
1097         return -1;
1098 }
1099
1100 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1101 {
1102         void *map;
1103         int fd;
1104
1105         fd = open_sha1_file(sha1);
1106         map = NULL;
1107         if (fd >= 0) {
1108                 struct stat st;
1109
1110                 if (!fstat(fd, &st)) {
1111                         *size = xsize_t(st.st_size);
1112                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1113                 }
1114                 close(fd);
1115         }
1116         return map;
1117 }
1118
1119 static int legacy_loose_object(unsigned char *map)
1120 {
1121         unsigned int word;
1122
1123         /*
1124          * Is it a zlib-compressed buffer? If so, the first byte
1125          * must be 0x78 (15-bit window size, deflated), and the
1126          * first 16-bit word is evenly divisible by 31
1127          */
1128         word = (map[0] << 8) + map[1];
1129         if (map[0] == 0x78 && !(word % 31))
1130                 return 1;
1131         else
1132                 return 0;
1133 }
1134
1135 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1136                 unsigned long len, enum object_type *type, unsigned long *sizep)
1137 {
1138         unsigned shift;
1139         unsigned long size, c;
1140         unsigned long used = 0;
1141
1142         c = buf[used++];
1143         *type = (c >> 4) & 7;
1144         size = c & 15;
1145         shift = 4;
1146         while (c & 0x80) {
1147                 if (len <= used || bitsizeof(long) <= shift) {
1148                         error("bad object header");
1149                         return 0;
1150                 }
1151                 c = buf[used++];
1152                 size += (c & 0x7f) << shift;
1153                 shift += 7;
1154         }
1155         *sizep = size;
1156         return used;
1157 }
1158
1159 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1160 {
1161         unsigned long size, used;
1162         static const char valid_loose_object_type[8] = {
1163                 0, /* OBJ_EXT */
1164                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1165                 0, /* "delta" and others are invalid in a loose object */
1166         };
1167         enum object_type type;
1168
1169         /* Get the data stream */
1170         memset(stream, 0, sizeof(*stream));
1171         stream->next_in = map;
1172         stream->avail_in = mapsize;
1173         stream->next_out = buffer;
1174         stream->avail_out = bufsiz;
1175
1176         if (legacy_loose_object(map)) {
1177                 git_inflate_init(stream);
1178                 return git_inflate(stream, 0);
1179         }
1180
1181
1182         /*
1183          * There used to be a second loose object header format which
1184          * was meant to mimic the in-pack format, allowing for direct
1185          * copy of the object data.  This format turned up not to be
1186          * really worth it and we don't write it any longer.  But we
1187          * can still read it.
1188          */
1189         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1190         if (!used || !valid_loose_object_type[type])
1191                 return -1;
1192         map += used;
1193         mapsize -= used;
1194
1195         /* Set up the stream for the rest.. */
1196         stream->next_in = map;
1197         stream->avail_in = mapsize;
1198         git_inflate_init(stream);
1199
1200         /* And generate the fake traditional header */
1201         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1202                                          typename(type), size);
1203         return 0;
1204 }
1205
1206 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1207 {
1208         int bytes = strlen(buffer) + 1;
1209         unsigned char *buf = xmallocz(size);
1210         unsigned long n;
1211         int status = Z_OK;
1212
1213         n = stream->total_out - bytes;
1214         if (n > size)
1215                 n = size;
1216         memcpy(buf, (char *) buffer + bytes, n);
1217         bytes = n;
1218         if (bytes <= size) {
1219                 /*
1220                  * The above condition must be (bytes <= size), not
1221                  * (bytes < size).  In other words, even though we
1222                  * expect no more output and set avail_out to zer0,
1223                  * the input zlib stream may have bytes that express
1224                  * "this concludes the stream", and we *do* want to
1225                  * eat that input.
1226                  *
1227                  * Otherwise we would not be able to test that we
1228                  * consumed all the input to reach the expected size;
1229                  * we also want to check that zlib tells us that all
1230                  * went well with status == Z_STREAM_END at the end.
1231                  */
1232                 stream->next_out = buf + bytes;
1233                 stream->avail_out = size - bytes;
1234                 while (status == Z_OK)
1235                         status = git_inflate(stream, Z_FINISH);
1236         }
1237         if (status == Z_STREAM_END && !stream->avail_in) {
1238                 git_inflate_end(stream);
1239                 return buf;
1240         }
1241
1242         if (status < 0)
1243                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1244         else if (stream->avail_in)
1245                 error("garbage at end of loose object '%s'",
1246                       sha1_to_hex(sha1));
1247         free(buf);
1248         return NULL;
1249 }
1250
1251 /*
1252  * We used to just use "sscanf()", but that's actually way
1253  * too permissive for what we want to check. So do an anal
1254  * object header parse by hand.
1255  */
1256 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1257 {
1258         char type[10];
1259         int i;
1260         unsigned long size;
1261
1262         /*
1263          * The type can be at most ten bytes (including the
1264          * terminating '\0' that we add), and is followed by
1265          * a space.
1266          */
1267         i = 0;
1268         for (;;) {
1269                 char c = *hdr++;
1270                 if (c == ' ')
1271                         break;
1272                 type[i++] = c;
1273                 if (i >= sizeof(type))
1274                         return -1;
1275         }
1276         type[i] = 0;
1277
1278         /*
1279          * The length must follow immediately, and be in canonical
1280          * decimal format (ie "010" is not valid).
1281          */
1282         size = *hdr++ - '0';
1283         if (size > 9)
1284                 return -1;
1285         if (size) {
1286                 for (;;) {
1287                         unsigned long c = *hdr - '0';
1288                         if (c > 9)
1289                                 break;
1290                         hdr++;
1291                         size = size * 10 + c;
1292                 }
1293         }
1294         *sizep = size;
1295
1296         /*
1297          * The length must be followed by a zero byte
1298          */
1299         return *hdr ? -1 : type_from_string(type);
1300 }
1301
1302 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1303 {
1304         int ret;
1305         z_stream stream;
1306         char hdr[8192];
1307
1308         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1309         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1310                 return NULL;
1311
1312         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1313 }
1314
1315 unsigned long get_size_from_delta(struct packed_git *p,
1316                                   struct pack_window **w_curs,
1317                                   off_t curpos)
1318 {
1319         const unsigned char *data;
1320         unsigned char delta_head[20], *in;
1321         z_stream stream;
1322         int st;
1323
1324         memset(&stream, 0, sizeof(stream));
1325         stream.next_out = delta_head;
1326         stream.avail_out = sizeof(delta_head);
1327
1328         git_inflate_init(&stream);
1329         do {
1330                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1331                 stream.next_in = in;
1332                 st = git_inflate(&stream, Z_FINISH);
1333                 curpos += stream.next_in - in;
1334         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1335                  stream.total_out < sizeof(delta_head));
1336         git_inflate_end(&stream);
1337         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1338                 error("delta data unpack-initial failed");
1339                 return 0;
1340         }
1341
1342         /* Examine the initial part of the delta to figure out
1343          * the result size.
1344          */
1345         data = delta_head;
1346
1347         /* ignore base size */
1348         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1349
1350         /* Read the result size */
1351         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1352 }
1353
1354 static off_t get_delta_base(struct packed_git *p,
1355                                     struct pack_window **w_curs,
1356                                     off_t *curpos,
1357                                     enum object_type type,
1358                                     off_t delta_obj_offset)
1359 {
1360         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1361         off_t base_offset;
1362
1363         /* use_pack() assured us we have [base_info, base_info + 20)
1364          * as a range that we can look at without walking off the
1365          * end of the mapped window.  Its actually the hash size
1366          * that is assured.  An OFS_DELTA longer than the hash size
1367          * is stupid, as then a REF_DELTA would be smaller to store.
1368          */
1369         if (type == OBJ_OFS_DELTA) {
1370                 unsigned used = 0;
1371                 unsigned char c = base_info[used++];
1372                 base_offset = c & 127;
1373                 while (c & 128) {
1374                         base_offset += 1;
1375                         if (!base_offset || MSB(base_offset, 7))
1376                                 return 0;  /* overflow */
1377                         c = base_info[used++];
1378                         base_offset = (base_offset << 7) + (c & 127);
1379                 }
1380                 base_offset = delta_obj_offset - base_offset;
1381                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1382                         return 0;  /* out of bound */
1383                 *curpos += used;
1384         } else if (type == OBJ_REF_DELTA) {
1385                 /* The base entry _must_ be in the same pack */
1386                 base_offset = find_pack_entry_one(base_info, p);
1387                 *curpos += 20;
1388         } else
1389                 die("I am totally screwed");
1390         return base_offset;
1391 }
1392
1393 /* forward declaration for a mutually recursive function */
1394 static int packed_object_info(struct packed_git *p, off_t offset,
1395                               unsigned long *sizep);
1396
1397 static int packed_delta_info(struct packed_git *p,
1398                              struct pack_window **w_curs,
1399                              off_t curpos,
1400                              enum object_type type,
1401                              off_t obj_offset,
1402                              unsigned long *sizep)
1403 {
1404         off_t base_offset;
1405
1406         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1407         if (!base_offset)
1408                 return OBJ_BAD;
1409         type = packed_object_info(p, base_offset, NULL);
1410         if (type <= OBJ_NONE) {
1411                 struct revindex_entry *revidx;
1412                 const unsigned char *base_sha1;
1413                 revidx = find_pack_revindex(p, base_offset);
1414                 if (!revidx)
1415                         return OBJ_BAD;
1416                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1417                 mark_bad_packed_object(p, base_sha1);
1418                 type = sha1_object_info(base_sha1, NULL);
1419                 if (type <= OBJ_NONE)
1420                         return OBJ_BAD;
1421         }
1422
1423         /* We choose to only get the type of the base object and
1424          * ignore potentially corrupt pack file that expects the delta
1425          * based on a base with a wrong size.  This saves tons of
1426          * inflate() calls.
1427          */
1428         if (sizep) {
1429                 *sizep = get_size_from_delta(p, w_curs, curpos);
1430                 if (*sizep == 0)
1431                         type = OBJ_BAD;
1432         }
1433
1434         return type;
1435 }
1436
1437 static int unpack_object_header(struct packed_git *p,
1438                                 struct pack_window **w_curs,
1439                                 off_t *curpos,
1440                                 unsigned long *sizep)
1441 {
1442         unsigned char *base;
1443         unsigned int left;
1444         unsigned long used;
1445         enum object_type type;
1446
1447         /* use_pack() assures us we have [base, base + 20) available
1448          * as a range that we can look at at.  (Its actually the hash
1449          * size that is assured.)  With our object header encoding
1450          * the maximum deflated object size is 2^137, which is just
1451          * insane, so we know won't exceed what we have been given.
1452          */
1453         base = use_pack(p, w_curs, *curpos, &left);
1454         used = unpack_object_header_buffer(base, left, &type, sizep);
1455         if (!used) {
1456                 type = OBJ_BAD;
1457         } else
1458                 *curpos += used;
1459
1460         return type;
1461 }
1462
1463 const char *packed_object_info_detail(struct packed_git *p,
1464                                       off_t obj_offset,
1465                                       unsigned long *size,
1466                                       unsigned long *store_size,
1467                                       unsigned int *delta_chain_length,
1468                                       unsigned char *base_sha1)
1469 {
1470         struct pack_window *w_curs = NULL;
1471         off_t curpos;
1472         unsigned long dummy;
1473         unsigned char *next_sha1;
1474         enum object_type type;
1475         struct revindex_entry *revidx;
1476
1477         *delta_chain_length = 0;
1478         curpos = obj_offset;
1479         type = unpack_object_header(p, &w_curs, &curpos, size);
1480
1481         revidx = find_pack_revindex(p, obj_offset);
1482         *store_size = revidx[1].offset - obj_offset;
1483
1484         for (;;) {
1485                 switch (type) {
1486                 default:
1487                         die("pack %s contains unknown object type %d",
1488                             p->pack_name, type);
1489                 case OBJ_COMMIT:
1490                 case OBJ_TREE:
1491                 case OBJ_BLOB:
1492                 case OBJ_TAG:
1493                         unuse_pack(&w_curs);
1494                         return typename(type);
1495                 case OBJ_OFS_DELTA:
1496                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1497                         if (!obj_offset)
1498                                 die("pack %s contains bad delta base reference of type %s",
1499                                     p->pack_name, typename(type));
1500                         if (*delta_chain_length == 0) {
1501                                 revidx = find_pack_revindex(p, obj_offset);
1502                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1503                         }
1504                         break;
1505                 case OBJ_REF_DELTA:
1506                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1507                         if (*delta_chain_length == 0)
1508                                 hashcpy(base_sha1, next_sha1);
1509                         obj_offset = find_pack_entry_one(next_sha1, p);
1510                         break;
1511                 }
1512                 (*delta_chain_length)++;
1513                 curpos = obj_offset;
1514                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1515         }
1516 }
1517
1518 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1519                               unsigned long *sizep)
1520 {
1521         struct pack_window *w_curs = NULL;
1522         unsigned long size;
1523         off_t curpos = obj_offset;
1524         enum object_type type;
1525
1526         type = unpack_object_header(p, &w_curs, &curpos, &size);
1527
1528         switch (type) {
1529         case OBJ_OFS_DELTA:
1530         case OBJ_REF_DELTA:
1531                 type = packed_delta_info(p, &w_curs, curpos,
1532                                          type, obj_offset, sizep);
1533                 break;
1534         case OBJ_COMMIT:
1535         case OBJ_TREE:
1536         case OBJ_BLOB:
1537         case OBJ_TAG:
1538                 if (sizep)
1539                         *sizep = size;
1540                 break;
1541         default:
1542                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1543                       type, (uintmax_t)obj_offset, p->pack_name);
1544                 type = OBJ_BAD;
1545         }
1546         unuse_pack(&w_curs);
1547         return type;
1548 }
1549
1550 static void *unpack_compressed_entry(struct packed_git *p,
1551                                     struct pack_window **w_curs,
1552                                     off_t curpos,
1553                                     unsigned long size)
1554 {
1555         int st;
1556         z_stream stream;
1557         unsigned char *buffer, *in;
1558
1559         buffer = xmallocz(size);
1560         memset(&stream, 0, sizeof(stream));
1561         stream.next_out = buffer;
1562         stream.avail_out = size + 1;
1563
1564         git_inflate_init(&stream);
1565         do {
1566                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1567                 stream.next_in = in;
1568                 st = git_inflate(&stream, Z_FINISH);
1569                 if (!stream.avail_out)
1570                         break; /* the payload is larger than it should be */
1571                 curpos += stream.next_in - in;
1572         } while (st == Z_OK || st == Z_BUF_ERROR);
1573         git_inflate_end(&stream);
1574         if ((st != Z_STREAM_END) || stream.total_out != size) {
1575                 free(buffer);
1576                 return NULL;
1577         }
1578
1579         return buffer;
1580 }
1581
1582 #define MAX_DELTA_CACHE (256)
1583
1584 static size_t delta_base_cached;
1585
1586 static struct delta_base_cache_lru_list {
1587         struct delta_base_cache_lru_list *prev;
1588         struct delta_base_cache_lru_list *next;
1589 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1590
1591 static struct delta_base_cache_entry {
1592         struct delta_base_cache_lru_list lru;
1593         void *data;
1594         struct packed_git *p;
1595         off_t base_offset;
1596         unsigned long size;
1597         enum object_type type;
1598 } delta_base_cache[MAX_DELTA_CACHE];
1599
1600 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1601 {
1602         unsigned long hash;
1603
1604         hash = (unsigned long)p + (unsigned long)base_offset;
1605         hash += (hash >> 8) + (hash >> 16);
1606         return hash % MAX_DELTA_CACHE;
1607 }
1608
1609 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1610         unsigned long *base_size, enum object_type *type, int keep_cache)
1611 {
1612         void *ret;
1613         unsigned long hash = pack_entry_hash(p, base_offset);
1614         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1615
1616         ret = ent->data;
1617         if (!ret || ent->p != p || ent->base_offset != base_offset)
1618                 return unpack_entry(p, base_offset, type, base_size);
1619
1620         if (!keep_cache) {
1621                 ent->data = NULL;
1622                 ent->lru.next->prev = ent->lru.prev;
1623                 ent->lru.prev->next = ent->lru.next;
1624                 delta_base_cached -= ent->size;
1625         } else {
1626                 ret = xmemdupz(ent->data, ent->size);
1627         }
1628         *type = ent->type;
1629         *base_size = ent->size;
1630         return ret;
1631 }
1632
1633 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1634 {
1635         if (ent->data) {
1636                 free(ent->data);
1637                 ent->data = NULL;
1638                 ent->lru.next->prev = ent->lru.prev;
1639                 ent->lru.prev->next = ent->lru.next;
1640                 delta_base_cached -= ent->size;
1641         }
1642 }
1643
1644 void clear_delta_base_cache(void)
1645 {
1646         unsigned long p;
1647         for (p = 0; p < MAX_DELTA_CACHE; p++)
1648                 release_delta_base_cache(&delta_base_cache[p]);
1649 }
1650
1651 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1652         void *base, unsigned long base_size, enum object_type type)
1653 {
1654         unsigned long hash = pack_entry_hash(p, base_offset);
1655         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1656         struct delta_base_cache_lru_list *lru;
1657
1658         release_delta_base_cache(ent);
1659         delta_base_cached += base_size;
1660
1661         for (lru = delta_base_cache_lru.next;
1662              delta_base_cached > delta_base_cache_limit
1663              && lru != &delta_base_cache_lru;
1664              lru = lru->next) {
1665                 struct delta_base_cache_entry *f = (void *)lru;
1666                 if (f->type == OBJ_BLOB)
1667                         release_delta_base_cache(f);
1668         }
1669         for (lru = delta_base_cache_lru.next;
1670              delta_base_cached > delta_base_cache_limit
1671              && lru != &delta_base_cache_lru;
1672              lru = lru->next) {
1673                 struct delta_base_cache_entry *f = (void *)lru;
1674                 release_delta_base_cache(f);
1675         }
1676
1677         ent->p = p;
1678         ent->base_offset = base_offset;
1679         ent->type = type;
1680         ent->data = base;
1681         ent->size = base_size;
1682         ent->lru.next = &delta_base_cache_lru;
1683         ent->lru.prev = delta_base_cache_lru.prev;
1684         delta_base_cache_lru.prev->next = &ent->lru;
1685         delta_base_cache_lru.prev = &ent->lru;
1686 }
1687
1688 static void *read_object(const unsigned char *sha1, enum object_type *type,
1689                          unsigned long *size);
1690
1691 static void *unpack_delta_entry(struct packed_git *p,
1692                                 struct pack_window **w_curs,
1693                                 off_t curpos,
1694                                 unsigned long delta_size,
1695                                 off_t obj_offset,
1696                                 enum object_type *type,
1697                                 unsigned long *sizep)
1698 {
1699         void *delta_data, *result, *base;
1700         unsigned long base_size;
1701         off_t base_offset;
1702
1703         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1704         if (!base_offset) {
1705                 error("failed to validate delta base reference "
1706                       "at offset %"PRIuMAX" from %s",
1707                       (uintmax_t)curpos, p->pack_name);
1708                 return NULL;
1709         }
1710         unuse_pack(w_curs);
1711         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1712         if (!base) {
1713                 /*
1714                  * We're probably in deep shit, but let's try to fetch
1715                  * the required base anyway from another pack or loose.
1716                  * This is costly but should happen only in the presence
1717                  * of a corrupted pack, and is better than failing outright.
1718                  */
1719                 struct revindex_entry *revidx;
1720                 const unsigned char *base_sha1;
1721                 revidx = find_pack_revindex(p, base_offset);
1722                 if (!revidx)
1723                         return NULL;
1724                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1725                 error("failed to read delta base object %s"
1726                       " at offset %"PRIuMAX" from %s",
1727                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1728                       p->pack_name);
1729                 mark_bad_packed_object(p, base_sha1);
1730                 base = read_object(base_sha1, type, &base_size);
1731                 if (!base)
1732                         return NULL;
1733         }
1734
1735         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1736         if (!delta_data) {
1737                 error("failed to unpack compressed delta "
1738                       "at offset %"PRIuMAX" from %s",
1739                       (uintmax_t)curpos, p->pack_name);
1740                 free(base);
1741                 return NULL;
1742         }
1743         result = patch_delta(base, base_size,
1744                              delta_data, delta_size,
1745                              sizep);
1746         if (!result)
1747                 die("failed to apply delta");
1748         free(delta_data);
1749         add_delta_base_cache(p, base_offset, base, base_size, *type);
1750         return result;
1751 }
1752
1753 int do_check_packed_object_crc;
1754
1755 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1756                    enum object_type *type, unsigned long *sizep)
1757 {
1758         struct pack_window *w_curs = NULL;
1759         off_t curpos = obj_offset;
1760         void *data;
1761
1762         if (do_check_packed_object_crc && p->index_version > 1) {
1763                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1764                 unsigned long len = revidx[1].offset - obj_offset;
1765                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1766                         const unsigned char *sha1 =
1767                                 nth_packed_object_sha1(p, revidx->nr);
1768                         error("bad packed object CRC for %s",
1769                               sha1_to_hex(sha1));
1770                         mark_bad_packed_object(p, sha1);
1771                         unuse_pack(&w_curs);
1772                         return NULL;
1773                 }
1774         }
1775
1776         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1777         switch (*type) {
1778         case OBJ_OFS_DELTA:
1779         case OBJ_REF_DELTA:
1780                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1781                                           obj_offset, type, sizep);
1782                 break;
1783         case OBJ_COMMIT:
1784         case OBJ_TREE:
1785         case OBJ_BLOB:
1786         case OBJ_TAG:
1787                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1788                 break;
1789         default:
1790                 data = NULL;
1791                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1792                       *type, (uintmax_t)obj_offset, p->pack_name);
1793         }
1794         unuse_pack(&w_curs);
1795         return data;
1796 }
1797
1798 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1799                                             uint32_t n)
1800 {
1801         const unsigned char *index = p->index_data;
1802         if (!index) {
1803                 if (open_pack_index(p))
1804                         return NULL;
1805                 index = p->index_data;
1806         }
1807         if (n >= p->num_objects)
1808                 return NULL;
1809         index += 4 * 256;
1810         if (p->index_version == 1) {
1811                 return index + 24 * n + 4;
1812         } else {
1813                 index += 8;
1814                 return index + 20 * n;
1815         }
1816 }
1817
1818 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1819 {
1820         const unsigned char *index = p->index_data;
1821         index += 4 * 256;
1822         if (p->index_version == 1) {
1823                 return ntohl(*((uint32_t *)(index + 24 * n)));
1824         } else {
1825                 uint32_t off;
1826                 index += 8 + p->num_objects * (20 + 4);
1827                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1828                 if (!(off & 0x80000000))
1829                         return off;
1830                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1831                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1832                                    ntohl(*((uint32_t *)(index + 4)));
1833         }
1834 }
1835
1836 off_t find_pack_entry_one(const unsigned char *sha1,
1837                                   struct packed_git *p)
1838 {
1839         const uint32_t *level1_ofs = p->index_data;
1840         const unsigned char *index = p->index_data;
1841         unsigned hi, lo, stride;
1842         static int use_lookup = -1;
1843         static int debug_lookup = -1;
1844
1845         if (debug_lookup < 0)
1846                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1847
1848         if (!index) {
1849                 if (open_pack_index(p))
1850                         return 0;
1851                 level1_ofs = p->index_data;
1852                 index = p->index_data;
1853         }
1854         if (p->index_version > 1) {
1855                 level1_ofs += 2;
1856                 index += 8;
1857         }
1858         index += 4 * 256;
1859         hi = ntohl(level1_ofs[*sha1]);
1860         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1861         if (p->index_version > 1) {
1862                 stride = 20;
1863         } else {
1864                 stride = 24;
1865                 index += 4;
1866         }
1867
1868         if (debug_lookup)
1869                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1870                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1871
1872         if (use_lookup < 0)
1873                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1874         if (use_lookup) {
1875                 int pos = sha1_entry_pos(index, stride, 0,
1876                                          lo, hi, p->num_objects, sha1);
1877                 if (pos < 0)
1878                         return 0;
1879                 return nth_packed_object_offset(p, pos);
1880         }
1881
1882         do {
1883                 unsigned mi = (lo + hi) / 2;
1884                 int cmp = hashcmp(index + mi * stride, sha1);
1885
1886                 if (debug_lookup)
1887                         printf("lo %u hi %u rg %u mi %u\n",
1888                                lo, hi, hi - lo, mi);
1889                 if (!cmp)
1890                         return nth_packed_object_offset(p, mi);
1891                 if (cmp > 0)
1892                         hi = mi;
1893                 else
1894                         lo = mi+1;
1895         } while (lo < hi);
1896         return 0;
1897 }
1898
1899 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
1900 {
1901         static struct packed_git *last_found = (void *)1;
1902         struct packed_git *p;
1903         off_t offset;
1904
1905         prepare_packed_git();
1906         if (!packed_git)
1907                 return 0;
1908         p = (last_found == (void *)1) ? packed_git : last_found;
1909
1910         do {
1911                 if (p->num_bad_objects) {
1912                         unsigned i;
1913                         for (i = 0; i < p->num_bad_objects; i++)
1914                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1915                                         goto next;
1916                 }
1917
1918                 offset = find_pack_entry_one(sha1, p);
1919                 if (offset) {
1920                         /*
1921                          * We are about to tell the caller where they can
1922                          * locate the requested object.  We better make
1923                          * sure the packfile is still here and can be
1924                          * accessed before supplying that answer, as
1925                          * it may have been deleted since the index
1926                          * was loaded!
1927                          */
1928                         if (p->pack_fd == -1 && open_packed_git(p)) {
1929                                 error("packfile %s cannot be accessed", p->pack_name);
1930                                 goto next;
1931                         }
1932                         e->offset = offset;
1933                         e->p = p;
1934                         hashcpy(e->sha1, sha1);
1935                         last_found = p;
1936                         return 1;
1937                 }
1938
1939                 next:
1940                 if (p == last_found)
1941                         p = packed_git;
1942                 else
1943                         p = p->next;
1944                 if (p == last_found)
1945                         p = p->next;
1946         } while (p);
1947         return 0;
1948 }
1949
1950 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1951                                   struct packed_git *packs)
1952 {
1953         struct packed_git *p;
1954
1955         for (p = packs; p; p = p->next) {
1956                 if (find_pack_entry_one(sha1, p))
1957                         return p;
1958         }
1959         return NULL;
1960
1961 }
1962
1963 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
1964 {
1965         int status;
1966         unsigned long mapsize, size;
1967         void *map;
1968         z_stream stream;
1969         char hdr[32];
1970
1971         map = map_sha1_file(sha1, &mapsize);
1972         if (!map)
1973                 return error("unable to find %s", sha1_to_hex(sha1));
1974         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
1975                 status = error("unable to unpack %s header",
1976                                sha1_to_hex(sha1));
1977         else if ((status = parse_sha1_header(hdr, &size)) < 0)
1978                 status = error("unable to parse %s header", sha1_to_hex(sha1));
1979         else if (sizep)
1980                 *sizep = size;
1981         git_inflate_end(&stream);
1982         munmap(map, mapsize);
1983         return status;
1984 }
1985
1986 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
1987 {
1988         struct pack_entry e;
1989         int status;
1990
1991         if (!find_pack_entry(sha1, &e)) {
1992                 /* Most likely it's a loose object. */
1993                 status = sha1_loose_object_info(sha1, sizep);
1994                 if (status >= 0)
1995                         return status;
1996
1997                 /* Not a loose object; someone else may have just packed it. */
1998                 reprepare_packed_git();
1999                 if (!find_pack_entry(sha1, &e))
2000                         return status;
2001         }
2002
2003         status = packed_object_info(e.p, e.offset, sizep);
2004         if (status < 0) {
2005                 mark_bad_packed_object(e.p, sha1);
2006                 status = sha1_object_info(sha1, sizep);
2007         }
2008
2009         return status;
2010 }
2011
2012 static void *read_packed_sha1(const unsigned char *sha1,
2013                               enum object_type *type, unsigned long *size)
2014 {
2015         struct pack_entry e;
2016         void *data;
2017
2018         if (!find_pack_entry(sha1, &e))
2019                 return NULL;
2020         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2021         if (!data) {
2022                 /*
2023                  * We're probably in deep shit, but let's try to fetch
2024                  * the required object anyway from another pack or loose.
2025                  * This should happen only in the presence of a corrupted
2026                  * pack, and is better than failing outright.
2027                  */
2028                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2029                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2030                 mark_bad_packed_object(e.p, sha1);
2031                 data = read_object(sha1, type, size);
2032         }
2033         return data;
2034 }
2035
2036 /*
2037  * This is meant to hold a *small* number of objects that you would
2038  * want read_sha1_file() to be able to return, but yet you do not want
2039  * to write them into the object store (e.g. a browse-only
2040  * application).
2041  */
2042 static struct cached_object {
2043         unsigned char sha1[20];
2044         enum object_type type;
2045         void *buf;
2046         unsigned long size;
2047 } *cached_objects;
2048 static int cached_object_nr, cached_object_alloc;
2049
2050 static struct cached_object empty_tree = {
2051         EMPTY_TREE_SHA1_BIN,
2052         OBJ_TREE,
2053         "",
2054         0
2055 };
2056
2057 static struct cached_object *find_cached_object(const unsigned char *sha1)
2058 {
2059         int i;
2060         struct cached_object *co = cached_objects;
2061
2062         for (i = 0; i < cached_object_nr; i++, co++) {
2063                 if (!hashcmp(co->sha1, sha1))
2064                         return co;
2065         }
2066         if (!hashcmp(sha1, empty_tree.sha1))
2067                 return &empty_tree;
2068         return NULL;
2069 }
2070
2071 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2072                       unsigned char *sha1)
2073 {
2074         struct cached_object *co;
2075
2076         hash_sha1_file(buf, len, typename(type), sha1);
2077         if (has_sha1_file(sha1) || find_cached_object(sha1))
2078                 return 0;
2079         if (cached_object_alloc <= cached_object_nr) {
2080                 cached_object_alloc = alloc_nr(cached_object_alloc);
2081                 cached_objects = xrealloc(cached_objects,
2082                                           sizeof(*cached_objects) *
2083                                           cached_object_alloc);
2084         }
2085         co = &cached_objects[cached_object_nr++];
2086         co->size = len;
2087         co->type = type;
2088         co->buf = xmalloc(len);
2089         memcpy(co->buf, buf, len);
2090         hashcpy(co->sha1, sha1);
2091         return 0;
2092 }
2093
2094 static void *read_object(const unsigned char *sha1, enum object_type *type,
2095                          unsigned long *size)
2096 {
2097         unsigned long mapsize;
2098         void *map, *buf;
2099         struct cached_object *co;
2100
2101         co = find_cached_object(sha1);
2102         if (co) {
2103                 *type = co->type;
2104                 *size = co->size;
2105                 return xmemdupz(co->buf, co->size);
2106         }
2107
2108         buf = read_packed_sha1(sha1, type, size);
2109         if (buf)
2110                 return buf;
2111         map = map_sha1_file(sha1, &mapsize);
2112         if (map) {
2113                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2114                 munmap(map, mapsize);
2115                 return buf;
2116         }
2117         reprepare_packed_git();
2118         return read_packed_sha1(sha1, type, size);
2119 }
2120
2121 /*
2122  * This function dies on corrupt objects; the callers who want to
2123  * deal with them should arrange to call read_object() and give error
2124  * messages themselves.
2125  */
2126 void *read_sha1_file_repl(const unsigned char *sha1,
2127                           enum object_type *type,
2128                           unsigned long *size,
2129                           const unsigned char **replacement)
2130 {
2131         const unsigned char *repl = lookup_replace_object(sha1);
2132         void *data;
2133         char *path;
2134         const struct packed_git *p;
2135
2136         errno = 0;
2137         data = read_object(repl, type, size);
2138         if (data) {
2139                 if (replacement)
2140                         *replacement = repl;
2141                 return data;
2142         }
2143
2144         if (errno != ENOENT)
2145                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2146
2147         /* die if we replaced an object with one that does not exist */
2148         if (repl != sha1)
2149                 die("replacement %s not found for %s",
2150                     sha1_to_hex(repl), sha1_to_hex(sha1));
2151
2152         if (has_loose_object(repl)) {
2153                 path = sha1_file_name(sha1);
2154                 die("loose object %s (stored in %s) is corrupt",
2155                     sha1_to_hex(repl), path);
2156         }
2157
2158         if ((p = has_packed_and_bad(repl)) != NULL)
2159                 die("packed object %s (stored in %s) is corrupt",
2160                     sha1_to_hex(repl), p->pack_name);
2161
2162         return NULL;
2163 }
2164
2165 void *read_object_with_reference(const unsigned char *sha1,
2166                                  const char *required_type_name,
2167                                  unsigned long *size,
2168                                  unsigned char *actual_sha1_return)
2169 {
2170         enum object_type type, required_type;
2171         void *buffer;
2172         unsigned long isize;
2173         unsigned char actual_sha1[20];
2174
2175         required_type = type_from_string(required_type_name);
2176         hashcpy(actual_sha1, sha1);
2177         while (1) {
2178                 int ref_length = -1;
2179                 const char *ref_type = NULL;
2180
2181                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2182                 if (!buffer)
2183                         return NULL;
2184                 if (type == required_type) {
2185                         *size = isize;
2186                         if (actual_sha1_return)
2187                                 hashcpy(actual_sha1_return, actual_sha1);
2188                         return buffer;
2189                 }
2190                 /* Handle references */
2191                 else if (type == OBJ_COMMIT)
2192                         ref_type = "tree ";
2193                 else if (type == OBJ_TAG)
2194                         ref_type = "object ";
2195                 else {
2196                         free(buffer);
2197                         return NULL;
2198                 }
2199                 ref_length = strlen(ref_type);
2200
2201                 if (ref_length + 40 > isize ||
2202                     memcmp(buffer, ref_type, ref_length) ||
2203                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2204                         free(buffer);
2205                         return NULL;
2206                 }
2207                 free(buffer);
2208                 /* Now we have the ID of the referred-to object in
2209                  * actual_sha1.  Check again. */
2210         }
2211 }
2212
2213 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2214                                     const char *type, unsigned char *sha1,
2215                                     char *hdr, int *hdrlen)
2216 {
2217         git_SHA_CTX c;
2218
2219         /* Generate the header */
2220         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2221
2222         /* Sha1.. */
2223         git_SHA1_Init(&c);
2224         git_SHA1_Update(&c, hdr, *hdrlen);
2225         git_SHA1_Update(&c, buf, len);
2226         git_SHA1_Final(sha1, &c);
2227 }
2228
2229 /*
2230  * Move the just written object into its final resting place.
2231  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2232  * "moving" is only a part of what it does, when no patch between
2233  * master to pu changes the call sites of this function.
2234  */
2235 int move_temp_to_file(const char *tmpfile, const char *filename)
2236 {
2237         int ret = 0;
2238
2239         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2240                 goto try_rename;
2241         else if (link(tmpfile, filename))
2242                 ret = errno;
2243
2244         /*
2245          * Coda hack - coda doesn't like cross-directory links,
2246          * so we fall back to a rename, which will mean that it
2247          * won't be able to check collisions, but that's not a
2248          * big deal.
2249          *
2250          * The same holds for FAT formatted media.
2251          *
2252          * When this succeeds, we just return.  We have nothing
2253          * left to unlink.
2254          */
2255         if (ret && ret != EEXIST) {
2256         try_rename:
2257                 if (!rename(tmpfile, filename))
2258                         goto out;
2259                 ret = errno;
2260         }
2261         unlink_or_warn(tmpfile);
2262         if (ret) {
2263                 if (ret != EEXIST) {
2264                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2265                 }
2266                 /* FIXME!!! Collision check here ? */
2267         }
2268
2269 out:
2270         if (adjust_shared_perm(filename))
2271                 return error("unable to set permission to '%s'", filename);
2272         return 0;
2273 }
2274
2275 static int write_buffer(int fd, const void *buf, size_t len)
2276 {
2277         if (write_in_full(fd, buf, len) < 0)
2278                 return error("file write error (%s)", strerror(errno));
2279         return 0;
2280 }
2281
2282 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2283                    unsigned char *sha1)
2284 {
2285         char hdr[32];
2286         int hdrlen;
2287         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2288         return 0;
2289 }
2290
2291 /* Finalize a file on disk, and close it. */
2292 static void close_sha1_file(int fd)
2293 {
2294         if (fsync_object_files)
2295                 fsync_or_die(fd, "sha1 file");
2296         if (close(fd) != 0)
2297                 die_errno("error when closing sha1 file");
2298 }
2299
2300 /* Size of directory component, including the ending '/' */
2301 static inline int directory_size(const char *filename)
2302 {
2303         const char *s = strrchr(filename, '/');
2304         if (!s)
2305                 return 0;
2306         return s - filename + 1;
2307 }
2308
2309 /*
2310  * This creates a temporary file in the same directory as the final
2311  * 'filename'
2312  *
2313  * We want to avoid cross-directory filename renames, because those
2314  * can have problems on various filesystems (FAT, NFS, Coda).
2315  */
2316 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2317 {
2318         int fd, dirlen = directory_size(filename);
2319
2320         if (dirlen + 20 > bufsiz) {
2321                 errno = ENAMETOOLONG;
2322                 return -1;
2323         }
2324         memcpy(buffer, filename, dirlen);
2325         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2326         fd = git_mkstemp_mode(buffer, 0444);
2327         if (fd < 0 && dirlen && errno == ENOENT) {
2328                 /* Make sure the directory exists */
2329                 memcpy(buffer, filename, dirlen);
2330                 buffer[dirlen-1] = 0;
2331                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2332                         return -1;
2333
2334                 /* Try again */
2335                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2336                 fd = git_mkstemp_mode(buffer, 0444);
2337         }
2338         return fd;
2339 }
2340
2341 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2342                               const void *buf, unsigned long len, time_t mtime)
2343 {
2344         int fd, ret;
2345         unsigned char compressed[4096];
2346         z_stream stream;
2347         git_SHA_CTX c;
2348         unsigned char parano_sha1[20];
2349         char *filename;
2350         static char tmpfile[PATH_MAX];
2351
2352         filename = sha1_file_name(sha1);
2353         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2354         while (fd < 0 && errno == EMFILE && unuse_one_window(NULL, -1))
2355                 fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2356         if (fd < 0) {
2357                 if (errno == EACCES)
2358                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2359                 else
2360                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2361         }
2362
2363         /* Set it up */
2364         memset(&stream, 0, sizeof(stream));
2365         deflateInit(&stream, zlib_compression_level);
2366         stream.next_out = compressed;
2367         stream.avail_out = sizeof(compressed);
2368         git_SHA1_Init(&c);
2369
2370         /* First header.. */
2371         stream.next_in = (unsigned char *)hdr;
2372         stream.avail_in = hdrlen;
2373         while (deflate(&stream, 0) == Z_OK)
2374                 /* nothing */;
2375         git_SHA1_Update(&c, hdr, hdrlen);
2376
2377         /* Then the data itself.. */
2378         stream.next_in = (void *)buf;
2379         stream.avail_in = len;
2380         do {
2381                 unsigned char *in0 = stream.next_in;
2382                 ret = deflate(&stream, Z_FINISH);
2383                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2384                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2385                         die("unable to write sha1 file");
2386                 stream.next_out = compressed;
2387                 stream.avail_out = sizeof(compressed);
2388         } while (ret == Z_OK);
2389
2390         if (ret != Z_STREAM_END)
2391                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2392         ret = deflateEnd(&stream);
2393         if (ret != Z_OK)
2394                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2395         git_SHA1_Final(parano_sha1, &c);
2396         if (hashcmp(sha1, parano_sha1) != 0)
2397                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2398
2399         close_sha1_file(fd);
2400
2401         if (mtime) {
2402                 struct utimbuf utb;
2403                 utb.actime = mtime;
2404                 utb.modtime = mtime;
2405                 if (utime(tmpfile, &utb) < 0)
2406                         warning("failed utime() on %s: %s",
2407                                 tmpfile, strerror(errno));
2408         }
2409
2410         return move_temp_to_file(tmpfile, filename);
2411 }
2412
2413 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2414 {
2415         unsigned char sha1[20];
2416         char hdr[32];
2417         int hdrlen;
2418
2419         /* Normally if we have it in the pack then we do not bother writing
2420          * it out into .git/objects/??/?{38} file.
2421          */
2422         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2423         if (returnsha1)
2424                 hashcpy(returnsha1, sha1);
2425         if (has_sha1_file(sha1))
2426                 return 0;
2427         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2428 }
2429
2430 int force_object_loose(const unsigned char *sha1, time_t mtime)
2431 {
2432         void *buf;
2433         unsigned long len;
2434         enum object_type type;
2435         char hdr[32];
2436         int hdrlen;
2437         int ret;
2438
2439         if (has_loose_object(sha1))
2440                 return 0;
2441         buf = read_packed_sha1(sha1, &type, &len);
2442         if (!buf)
2443                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2444         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2445         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2446         free(buf);
2447
2448         return ret;
2449 }
2450
2451 int has_pack_index(const unsigned char *sha1)
2452 {
2453         struct stat st;
2454         if (stat(sha1_pack_index_name(sha1), &st))
2455                 return 0;
2456         return 1;
2457 }
2458
2459 int has_sha1_pack(const unsigned char *sha1)
2460 {
2461         struct pack_entry e;
2462         return find_pack_entry(sha1, &e);
2463 }
2464
2465 int has_sha1_file(const unsigned char *sha1)
2466 {
2467         struct pack_entry e;
2468
2469         if (find_pack_entry(sha1, &e))
2470                 return 1;
2471         return has_loose_object(sha1);
2472 }
2473
2474 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2475                      int write_object, enum object_type type, const char *path)
2476 {
2477         int ret, re_allocated = 0;
2478
2479         if (!type)
2480                 type = OBJ_BLOB;
2481
2482         /*
2483          * Convert blobs to git internal format
2484          */
2485         if ((type == OBJ_BLOB) && path) {
2486                 struct strbuf nbuf = STRBUF_INIT;
2487                 if (convert_to_git(path, buf, size, &nbuf,
2488                                    write_object ? safe_crlf : 0)) {
2489                         buf = strbuf_detach(&nbuf, &size);
2490                         re_allocated = 1;
2491                 }
2492         }
2493
2494         if (write_object)
2495                 ret = write_sha1_file(buf, size, typename(type), sha1);
2496         else
2497                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2498         if (re_allocated)
2499                 free(buf);
2500         return ret;
2501 }
2502
2503 #define SMALL_FILE_SIZE (32*1024)
2504
2505 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2506              enum object_type type, const char *path)
2507 {
2508         int ret;
2509         size_t size = xsize_t(st->st_size);
2510
2511         if (!S_ISREG(st->st_mode)) {
2512                 struct strbuf sbuf = STRBUF_INIT;
2513                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2514                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2515                                         type, path);
2516                 else
2517                         ret = -1;
2518                 strbuf_release(&sbuf);
2519         } else if (!size) {
2520                 ret = index_mem(sha1, NULL, size, write_object, type, path);
2521         } else if (size <= SMALL_FILE_SIZE) {
2522                 char *buf = xmalloc(size);
2523                 if (size == read_in_full(fd, buf, size))
2524                         ret = index_mem(sha1, buf, size, write_object, type,
2525                                         path);
2526                 else
2527                         ret = error("short read %s", strerror(errno));
2528                 free(buf);
2529         } else {
2530                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2531                 ret = index_mem(sha1, buf, size, write_object, type, path);
2532                 munmap(buf, size);
2533         }
2534         close(fd);
2535         return ret;
2536 }
2537
2538 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2539 {
2540         int fd;
2541         struct strbuf sb = STRBUF_INIT;
2542
2543         switch (st->st_mode & S_IFMT) {
2544         case S_IFREG:
2545                 fd = open(path, O_RDONLY);
2546                 if (fd < 0)
2547                         return error("open(\"%s\"): %s", path,
2548                                      strerror(errno));
2549                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path) < 0)
2550                         return error("%s: failed to insert into database",
2551                                      path);
2552                 break;
2553         case S_IFLNK:
2554                 if (strbuf_readlink(&sb, path, st->st_size)) {
2555                         char *errstr = strerror(errno);
2556                         return error("readlink(\"%s\"): %s", path,
2557                                      errstr);
2558                 }
2559                 if (!write_object)
2560                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2561                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2562                         return error("%s: failed to insert into database",
2563                                      path);
2564                 strbuf_release(&sb);
2565                 break;
2566         case S_IFDIR:
2567                 return resolve_gitlink_ref(path, "HEAD", sha1);
2568         default:
2569                 return error("%s: unsupported file type", path);
2570         }
2571         return 0;
2572 }
2573
2574 int read_pack_header(int fd, struct pack_header *header)
2575 {
2576         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2577                 /* "eof before pack header was fully read" */
2578                 return PH_ERROR_EOF;
2579
2580         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2581                 /* "protocol error (pack signature mismatch detected)" */
2582                 return PH_ERROR_PACK_SIGNATURE;
2583         if (!pack_version_ok(header->hdr_version))
2584                 /* "protocol error (pack version unsupported)" */
2585                 return PH_ERROR_PROTOCOL;
2586         return 0;
2587 }
2588
2589 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
2590 {
2591         enum object_type type = sha1_object_info(sha1, NULL);
2592         if (type < 0)
2593                 die("%s is not a valid object", sha1_to_hex(sha1));
2594         if (type != expect)
2595                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
2596                     typename(expect));
2597 }