convert: make it safer to add conversion attributes
[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 "tree-walk.h"
17 #include "refs.h"
18 #include "pack-revindex.h"
19 #include "sha1-lookup.h"
20
21 #ifndef O_NOATIME
22 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
23 #define O_NOATIME 01000000
24 #else
25 #define O_NOATIME 0
26 #endif
27 #endif
28
29 #define SZ_FMT PRIuMAX
30 static inline uintmax_t sz_fmt(size_t s) { return s; }
31
32 const unsigned char null_sha1[20];
33
34 static int git_open_noatime(const char *name, struct packed_git *p);
35
36 /*
37  * This is meant to hold a *small* number of objects that you would
38  * want read_sha1_file() to be able to return, but yet you do not want
39  * to write them into the object store (e.g. a browse-only
40  * application).
41  */
42 static struct cached_object {
43         unsigned char sha1[20];
44         enum object_type type;
45         void *buf;
46         unsigned long size;
47 } *cached_objects;
48 static int cached_object_nr, cached_object_alloc;
49
50 static struct cached_object empty_tree = {
51         EMPTY_TREE_SHA1_BIN_LITERAL,
52         OBJ_TREE,
53         "",
54         0
55 };
56
57 static struct cached_object *find_cached_object(const unsigned char *sha1)
58 {
59         int i;
60         struct cached_object *co = cached_objects;
61
62         for (i = 0; i < cached_object_nr; i++, co++) {
63                 if (!hashcmp(co->sha1, sha1))
64                         return co;
65         }
66         if (!hashcmp(sha1, empty_tree.sha1))
67                 return &empty_tree;
68         return NULL;
69 }
70
71 int mkdir_in_gitdir(const char *path)
72 {
73         if (mkdir(path, 0777)) {
74                 int saved_errno = errno;
75                 struct stat st;
76                 struct strbuf sb = STRBUF_INIT;
77
78                 if (errno != EEXIST)
79                         return -1;
80                 /*
81                  * Are we looking at a path in a symlinked worktree
82                  * whose original repository does not yet have it?
83                  * e.g. .git/rr-cache pointing at its original
84                  * repository in which the user hasn't performed any
85                  * conflict resolution yet?
86                  */
87                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
88                     strbuf_readlink(&sb, path, st.st_size) ||
89                     !is_absolute_path(sb.buf) ||
90                     mkdir(sb.buf, 0777)) {
91                         strbuf_release(&sb);
92                         errno = saved_errno;
93                         return -1;
94                 }
95                 strbuf_release(&sb);
96         }
97         return adjust_shared_perm(path);
98 }
99
100 int safe_create_leading_directories(char *path)
101 {
102         char *pos = path + offset_1st_component(path);
103         struct stat st;
104
105         while (pos) {
106                 pos = strchr(pos, '/');
107                 if (!pos)
108                         break;
109                 while (*++pos == '/')
110                         ;
111                 if (!*pos)
112                         break;
113                 *--pos = '\0';
114                 if (!stat(path, &st)) {
115                         /* path exists */
116                         if (!S_ISDIR(st.st_mode)) {
117                                 *pos = '/';
118                                 return -3;
119                         }
120                 }
121                 else if (mkdir(path, 0777)) {
122                         *pos = '/';
123                         return -1;
124                 }
125                 else if (adjust_shared_perm(path)) {
126                         *pos = '/';
127                         return -2;
128                 }
129                 *pos++ = '/';
130         }
131         return 0;
132 }
133
134 int safe_create_leading_directories_const(const char *path)
135 {
136         /* path points to cache entries, so xstrdup before messing with it */
137         char *buf = xstrdup(path);
138         int result = safe_create_leading_directories(buf);
139         free(buf);
140         return result;
141 }
142
143 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
144 {
145         int i;
146         for (i = 0; i < 20; i++) {
147                 static char hex[] = "0123456789abcdef";
148                 unsigned int val = sha1[i];
149                 char *pos = pathbuf + i*2 + (i > 0);
150                 *pos++ = hex[val >> 4];
151                 *pos = hex[val & 0xf];
152         }
153 }
154
155 /*
156  * NOTE! This returns a statically allocated buffer, so you have to be
157  * careful about using it. Do an "xstrdup()" if you need to save the
158  * filename.
159  *
160  * Also note that this returns the location for creating.  Reading
161  * SHA1 file can happen from any alternate directory listed in the
162  * DB_ENVIRONMENT environment variable if it is not found in
163  * the primary object database.
164  */
165 char *sha1_file_name(const unsigned char *sha1)
166 {
167         static char buf[PATH_MAX];
168         const char *objdir;
169         int len;
170
171         objdir = get_object_directory();
172         len = strlen(objdir);
173
174         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
175         if (len + 43 > PATH_MAX)
176                 die("insanely long object directory %s", objdir);
177         memcpy(buf, objdir, len);
178         buf[len] = '/';
179         buf[len+3] = '/';
180         buf[len+42] = '\0';
181         fill_sha1_path(buf + len + 1, sha1);
182         return buf;
183 }
184
185 static char *sha1_get_pack_name(const unsigned char *sha1,
186                                 char **name, char **base, const char *which)
187 {
188         static const char hex[] = "0123456789abcdef";
189         char *buf;
190         int i;
191
192         if (!*base) {
193                 const char *sha1_file_directory = get_object_directory();
194                 int len = strlen(sha1_file_directory);
195                 *base = xmalloc(len + 60);
196                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
197                         sha1_file_directory, which);
198                 *name = *base + len + 11;
199         }
200
201         buf = *name;
202
203         for (i = 0; i < 20; i++) {
204                 unsigned int val = *sha1++;
205                 *buf++ = hex[val >> 4];
206                 *buf++ = hex[val & 0xf];
207         }
208
209         return *base;
210 }
211
212 char *sha1_pack_name(const unsigned char *sha1)
213 {
214         static char *name, *base;
215
216         return sha1_get_pack_name(sha1, &name, &base, "pack");
217 }
218
219 char *sha1_pack_index_name(const unsigned char *sha1)
220 {
221         static char *name, *base;
222
223         return sha1_get_pack_name(sha1, &name, &base, "idx");
224 }
225
226 struct alternate_object_database *alt_odb_list;
227 static struct alternate_object_database **alt_odb_tail;
228
229 static void read_info_alternates(const char * alternates, int depth);
230
231 /*
232  * Prepare alternate object database registry.
233  *
234  * The variable alt_odb_list points at the list of struct
235  * alternate_object_database.  The elements on this list come from
236  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
237  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
238  * whose contents is similar to that environment variable but can be
239  * LF separated.  Its base points at a statically allocated buffer that
240  * contains "/the/directory/corresponding/to/.git/objects/...", while
241  * its name points just after the slash at the end of ".git/objects/"
242  * in the example above, and has enough space to hold 40-byte hex
243  * SHA1, an extra slash for the first level indirection, and the
244  * terminating NUL.
245  */
246 static int link_alt_odb_entry(const char * entry, int len, const char * relative_base, int depth)
247 {
248         const char *objdir = get_object_directory();
249         struct alternate_object_database *ent;
250         struct alternate_object_database *alt;
251         /* 43 = 40-byte + 2 '/' + terminating NUL */
252         int pfxlen = len;
253         int entlen = pfxlen + 43;
254         int base_len = -1;
255
256         if (!is_absolute_path(entry) && relative_base) {
257                 /* Relative alt-odb */
258                 if (base_len < 0)
259                         base_len = strlen(relative_base) + 1;
260                 entlen += base_len;
261                 pfxlen += base_len;
262         }
263         ent = xmalloc(sizeof(*ent) + entlen);
264
265         if (!is_absolute_path(entry) && relative_base) {
266                 memcpy(ent->base, relative_base, base_len - 1);
267                 ent->base[base_len - 1] = '/';
268                 memcpy(ent->base + base_len, entry, len);
269         }
270         else
271                 memcpy(ent->base, entry, pfxlen);
272
273         ent->name = ent->base + pfxlen + 1;
274         ent->base[pfxlen + 3] = '/';
275         ent->base[pfxlen] = ent->base[entlen-1] = 0;
276
277         /* Detect cases where alternate disappeared */
278         if (!is_directory(ent->base)) {
279                 error("object directory %s does not exist; "
280                       "check .git/objects/info/alternates.",
281                       ent->base);
282                 free(ent);
283                 return -1;
284         }
285
286         /* Prevent the common mistake of listing the same
287          * thing twice, or object directory itself.
288          */
289         for (alt = alt_odb_list; alt; alt = alt->next) {
290                 if (!memcmp(ent->base, alt->base, pfxlen)) {
291                         free(ent);
292                         return -1;
293                 }
294         }
295         if (!memcmp(ent->base, objdir, pfxlen)) {
296                 free(ent);
297                 return -1;
298         }
299
300         /* add the alternate entry */
301         *alt_odb_tail = ent;
302         alt_odb_tail = &(ent->next);
303         ent->next = NULL;
304
305         /* recursively add alternates */
306         read_info_alternates(ent->base, depth + 1);
307
308         ent->base[pfxlen] = '/';
309
310         return 0;
311 }
312
313 static void link_alt_odb_entries(const char *alt, const char *ep, int sep,
314                                  const char *relative_base, int depth)
315 {
316         const char *cp, *last;
317
318         if (depth > 5) {
319                 error("%s: ignoring alternate object stores, nesting too deep.",
320                                 relative_base);
321                 return;
322         }
323
324         last = alt;
325         while (last < ep) {
326                 cp = last;
327                 if (cp < ep && *cp == '#') {
328                         while (cp < ep && *cp != sep)
329                                 cp++;
330                         last = cp + 1;
331                         continue;
332                 }
333                 while (cp < ep && *cp != sep)
334                         cp++;
335                 if (last != cp) {
336                         if (!is_absolute_path(last) && depth) {
337                                 error("%s: ignoring relative alternate object store %s",
338                                                 relative_base, last);
339                         } else {
340                                 link_alt_odb_entry(last, cp - last,
341                                                 relative_base, depth);
342                         }
343                 }
344                 while (cp < ep && *cp == sep)
345                         cp++;
346                 last = cp;
347         }
348 }
349
350 static void read_info_alternates(const char * relative_base, int depth)
351 {
352         char *map;
353         size_t mapsz;
354         struct stat st;
355         const char alt_file_name[] = "info/alternates";
356         /* Given that relative_base is no longer than PATH_MAX,
357            ensure that "path" has enough space to append "/", the
358            file name, "info/alternates", and a trailing NUL.  */
359         char path[PATH_MAX + 1 + sizeof alt_file_name];
360         int fd;
361
362         sprintf(path, "%s/%s", relative_base, alt_file_name);
363         fd = git_open_noatime(path, NULL);
364         if (fd < 0)
365                 return;
366         if (fstat(fd, &st) || (st.st_size == 0)) {
367                 close(fd);
368                 return;
369         }
370         mapsz = xsize_t(st.st_size);
371         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
372         close(fd);
373
374         link_alt_odb_entries(map, map + mapsz, '\n', relative_base, depth);
375
376         munmap(map, mapsz);
377 }
378
379 void add_to_alternates_file(const char *reference)
380 {
381         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
382         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
383         char *alt = mkpath("%s/objects\n", reference);
384         write_or_die(fd, alt, strlen(alt));
385         if (commit_lock_file(lock))
386                 die("could not close alternates file");
387         if (alt_odb_tail)
388                 link_alt_odb_entries(alt, alt + strlen(alt), '\n', NULL, 0);
389 }
390
391 void foreach_alt_odb(alt_odb_fn fn, void *cb)
392 {
393         struct alternate_object_database *ent;
394
395         prepare_alt_odb();
396         for (ent = alt_odb_list; ent; ent = ent->next)
397                 if (fn(ent, cb))
398                         return;
399 }
400
401 void prepare_alt_odb(void)
402 {
403         const char *alt;
404
405         if (alt_odb_tail)
406                 return;
407
408         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
409         if (!alt) alt = "";
410
411         alt_odb_tail = &alt_odb_list;
412         link_alt_odb_entries(alt, alt + strlen(alt), PATH_SEP, NULL, 0);
413
414         read_info_alternates(get_object_directory(), 0);
415 }
416
417 static int has_loose_object_local(const unsigned char *sha1)
418 {
419         char *name = sha1_file_name(sha1);
420         return !access(name, F_OK);
421 }
422
423 int has_loose_object_nonlocal(const unsigned char *sha1)
424 {
425         struct alternate_object_database *alt;
426         prepare_alt_odb();
427         for (alt = alt_odb_list; alt; alt = alt->next) {
428                 fill_sha1_path(alt->name, sha1);
429                 if (!access(alt->base, F_OK))
430                         return 1;
431         }
432         return 0;
433 }
434
435 static int has_loose_object(const unsigned char *sha1)
436 {
437         return has_loose_object_local(sha1) ||
438                has_loose_object_nonlocal(sha1);
439 }
440
441 static unsigned int pack_used_ctr;
442 static unsigned int pack_mmap_calls;
443 static unsigned int peak_pack_open_windows;
444 static unsigned int pack_open_windows;
445 static unsigned int pack_open_fds;
446 static unsigned int pack_max_fds;
447 static size_t peak_pack_mapped;
448 static size_t pack_mapped;
449 struct packed_git *packed_git;
450
451 void pack_report(void)
452 {
453         fprintf(stderr,
454                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
455                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
456                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
457                 sz_fmt(getpagesize()),
458                 sz_fmt(packed_git_window_size),
459                 sz_fmt(packed_git_limit));
460         fprintf(stderr,
461                 "pack_report: pack_used_ctr            = %10u\n"
462                 "pack_report: pack_mmap_calls          = %10u\n"
463                 "pack_report: pack_open_windows        = %10u / %10u\n"
464                 "pack_report: pack_mapped              = "
465                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
466                 pack_used_ctr,
467                 pack_mmap_calls,
468                 pack_open_windows, peak_pack_open_windows,
469                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
470 }
471
472 static int check_packed_git_idx(const char *path,  struct packed_git *p)
473 {
474         void *idx_map;
475         struct pack_idx_header *hdr;
476         size_t idx_size;
477         uint32_t version, nr, i, *index;
478         int fd = git_open_noatime(path, p);
479         struct stat st;
480
481         if (fd < 0)
482                 return -1;
483         if (fstat(fd, &st)) {
484                 close(fd);
485                 return -1;
486         }
487         idx_size = xsize_t(st.st_size);
488         if (idx_size < 4 * 256 + 20 + 20) {
489                 close(fd);
490                 return error("index file %s is too small", path);
491         }
492         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
493         close(fd);
494
495         hdr = idx_map;
496         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
497                 version = ntohl(hdr->idx_version);
498                 if (version < 2 || version > 2) {
499                         munmap(idx_map, idx_size);
500                         return error("index file %s is version %"PRIu32
501                                      " and is not supported by this binary"
502                                      " (try upgrading GIT to a newer version)",
503                                      path, version);
504                 }
505         } else
506                 version = 1;
507
508         nr = 0;
509         index = idx_map;
510         if (version > 1)
511                 index += 2;  /* skip index header */
512         for (i = 0; i < 256; i++) {
513                 uint32_t n = ntohl(index[i]);
514                 if (n < nr) {
515                         munmap(idx_map, idx_size);
516                         return error("non-monotonic index %s", path);
517                 }
518                 nr = n;
519         }
520
521         if (version == 1) {
522                 /*
523                  * Total size:
524                  *  - 256 index entries 4 bytes each
525                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
526                  *  - 20-byte SHA1 of the packfile
527                  *  - 20-byte SHA1 file checksum
528                  */
529                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
530                         munmap(idx_map, idx_size);
531                         return error("wrong index v1 file size in %s", path);
532                 }
533         } else if (version == 2) {
534                 /*
535                  * Minimum size:
536                  *  - 8 bytes of header
537                  *  - 256 index entries 4 bytes each
538                  *  - 20-byte sha1 entry * nr
539                  *  - 4-byte crc entry * nr
540                  *  - 4-byte offset entry * nr
541                  *  - 20-byte SHA1 of the packfile
542                  *  - 20-byte SHA1 file checksum
543                  * And after the 4-byte offset table might be a
544                  * variable sized table containing 8-byte entries
545                  * for offsets larger than 2^31.
546                  */
547                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
548                 unsigned long max_size = min_size;
549                 if (nr)
550                         max_size += (nr - 1)*8;
551                 if (idx_size < min_size || idx_size > max_size) {
552                         munmap(idx_map, idx_size);
553                         return error("wrong index v2 file size in %s", path);
554                 }
555                 if (idx_size != min_size &&
556                     /*
557                      * make sure we can deal with large pack offsets.
558                      * 31-bit signed offset won't be enough, neither
559                      * 32-bit unsigned one will be.
560                      */
561                     (sizeof(off_t) <= 4)) {
562                         munmap(idx_map, idx_size);
563                         return error("pack too large for current definition of off_t in %s", path);
564                 }
565         }
566
567         p->index_version = version;
568         p->index_data = idx_map;
569         p->index_size = idx_size;
570         p->num_objects = nr;
571         return 0;
572 }
573
574 int open_pack_index(struct packed_git *p)
575 {
576         char *idx_name;
577         int ret;
578
579         if (p->index_data)
580                 return 0;
581
582         idx_name = xstrdup(p->pack_name);
583         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
584         ret = check_packed_git_idx(idx_name, p);
585         free(idx_name);
586         return ret;
587 }
588
589 static void scan_windows(struct packed_git *p,
590         struct packed_git **lru_p,
591         struct pack_window **lru_w,
592         struct pack_window **lru_l)
593 {
594         struct pack_window *w, *w_l;
595
596         for (w_l = NULL, w = p->windows; w; w = w->next) {
597                 if (!w->inuse_cnt) {
598                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
599                                 *lru_p = p;
600                                 *lru_w = w;
601                                 *lru_l = w_l;
602                         }
603                 }
604                 w_l = w;
605         }
606 }
607
608 static int unuse_one_window(struct packed_git *current, int keep_fd)
609 {
610         struct packed_git *p, *lru_p = NULL;
611         struct pack_window *lru_w = NULL, *lru_l = NULL;
612
613         if (current)
614                 scan_windows(current, &lru_p, &lru_w, &lru_l);
615         for (p = packed_git; p; p = p->next)
616                 scan_windows(p, &lru_p, &lru_w, &lru_l);
617         if (lru_p) {
618                 munmap(lru_w->base, lru_w->len);
619                 pack_mapped -= lru_w->len;
620                 if (lru_l)
621                         lru_l->next = lru_w->next;
622                 else {
623                         lru_p->windows = lru_w->next;
624                         if (!lru_p->windows && lru_p->pack_fd != -1
625                                 && lru_p->pack_fd != keep_fd) {
626                                 close(lru_p->pack_fd);
627                                 pack_open_fds--;
628                                 lru_p->pack_fd = -1;
629                         }
630                 }
631                 free(lru_w);
632                 pack_open_windows--;
633                 return 1;
634         }
635         return 0;
636 }
637
638 void release_pack_memory(size_t need, int fd)
639 {
640         size_t cur = pack_mapped;
641         while (need >= (cur - pack_mapped) && unuse_one_window(NULL, fd))
642                 ; /* nothing */
643 }
644
645 void *xmmap(void *start, size_t length,
646         int prot, int flags, int fd, off_t offset)
647 {
648         void *ret = mmap(start, length, prot, flags, fd, offset);
649         if (ret == MAP_FAILED) {
650                 if (!length)
651                         return NULL;
652                 release_pack_memory(length, fd);
653                 ret = mmap(start, length, prot, flags, fd, offset);
654                 if (ret == MAP_FAILED)
655                         die_errno("Out of memory? mmap failed");
656         }
657         return ret;
658 }
659
660 void close_pack_windows(struct packed_git *p)
661 {
662         while (p->windows) {
663                 struct pack_window *w = p->windows;
664
665                 if (w->inuse_cnt)
666                         die("pack '%s' still has open windows to it",
667                             p->pack_name);
668                 munmap(w->base, w->len);
669                 pack_mapped -= w->len;
670                 pack_open_windows--;
671                 p->windows = w->next;
672                 free(w);
673         }
674 }
675
676 void unuse_pack(struct pack_window **w_cursor)
677 {
678         struct pack_window *w = *w_cursor;
679         if (w) {
680                 w->inuse_cnt--;
681                 *w_cursor = NULL;
682         }
683 }
684
685 void close_pack_index(struct packed_git *p)
686 {
687         if (p->index_data) {
688                 munmap((void *)p->index_data, p->index_size);
689                 p->index_data = NULL;
690         }
691 }
692
693 /*
694  * This is used by git-repack in case a newly created pack happens to
695  * contain the same set of objects as an existing one.  In that case
696  * the resulting file might be different even if its name would be the
697  * same.  It is best to close any reference to the old pack before it is
698  * replaced on disk.  Of course no index pointers nor windows for given pack
699  * must subsist at this point.  If ever objects from this pack are requested
700  * again, the new version of the pack will be reinitialized through
701  * reprepare_packed_git().
702  */
703 void free_pack_by_name(const char *pack_name)
704 {
705         struct packed_git *p, **pp = &packed_git;
706
707         while (*pp) {
708                 p = *pp;
709                 if (strcmp(pack_name, p->pack_name) == 0) {
710                         clear_delta_base_cache();
711                         close_pack_windows(p);
712                         if (p->pack_fd != -1) {
713                                 close(p->pack_fd);
714                                 pack_open_fds--;
715                         }
716                         close_pack_index(p);
717                         free(p->bad_object_sha1);
718                         *pp = p->next;
719                         free(p);
720                         return;
721                 }
722                 pp = &p->next;
723         }
724 }
725
726 /*
727  * Do not call this directly as this leaks p->pack_fd on error return;
728  * call open_packed_git() instead.
729  */
730 static int open_packed_git_1(struct packed_git *p)
731 {
732         struct stat st;
733         struct pack_header hdr;
734         unsigned char sha1[20];
735         unsigned char *idx_sha1;
736         long fd_flag;
737
738         if (!p->index_data && open_pack_index(p))
739                 return error("packfile %s index unavailable", p->pack_name);
740
741         if (!pack_max_fds) {
742                 struct rlimit lim;
743                 unsigned int max_fds;
744
745                 if (getrlimit(RLIMIT_NOFILE, &lim))
746                         die_errno("cannot get RLIMIT_NOFILE");
747
748                 max_fds = lim.rlim_cur;
749
750                 /* Save 3 for stdin/stdout/stderr, 22 for work */
751                 if (25 < max_fds)
752                         pack_max_fds = max_fds - 25;
753                 else
754                         pack_max_fds = 1;
755         }
756
757         while (pack_max_fds <= pack_open_fds && unuse_one_window(NULL, -1))
758                 ; /* nothing */
759
760         p->pack_fd = git_open_noatime(p->pack_name, p);
761         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
762                 return -1;
763         pack_open_fds++;
764
765         /* If we created the struct before we had the pack we lack size. */
766         if (!p->pack_size) {
767                 if (!S_ISREG(st.st_mode))
768                         return error("packfile %s not a regular file", p->pack_name);
769                 p->pack_size = st.st_size;
770         } else if (p->pack_size != st.st_size)
771                 return error("packfile %s size changed", p->pack_name);
772
773         /* We leave these file descriptors open with sliding mmap;
774          * there is no point keeping them open across exec(), though.
775          */
776         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
777         if (fd_flag < 0)
778                 return error("cannot determine file descriptor flags");
779         fd_flag |= FD_CLOEXEC;
780         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
781                 return error("cannot set FD_CLOEXEC");
782
783         /* Verify we recognize this pack file format. */
784         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
785                 return error("file %s is far too short to be a packfile", p->pack_name);
786         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
787                 return error("file %s is not a GIT packfile", p->pack_name);
788         if (!pack_version_ok(hdr.hdr_version))
789                 return error("packfile %s is version %"PRIu32" and not"
790                         " supported (try upgrading GIT to a newer version)",
791                         p->pack_name, ntohl(hdr.hdr_version));
792
793         /* Verify the pack matches its index. */
794         if (p->num_objects != ntohl(hdr.hdr_entries))
795                 return error("packfile %s claims to have %"PRIu32" objects"
796                              " while index indicates %"PRIu32" objects",
797                              p->pack_name, ntohl(hdr.hdr_entries),
798                              p->num_objects);
799         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
800                 return error("end of packfile %s is unavailable", p->pack_name);
801         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
802                 return error("packfile %s signature is unavailable", p->pack_name);
803         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
804         if (hashcmp(sha1, idx_sha1))
805                 return error("packfile %s does not match index", p->pack_name);
806         return 0;
807 }
808
809 static int open_packed_git(struct packed_git *p)
810 {
811         if (!open_packed_git_1(p))
812                 return 0;
813         if (p->pack_fd != -1) {
814                 close(p->pack_fd);
815                 pack_open_fds--;
816                 p->pack_fd = -1;
817         }
818         return -1;
819 }
820
821 static int in_window(struct pack_window *win, off_t offset)
822 {
823         /* We must promise at least 20 bytes (one hash) after the
824          * offset is available from this window, otherwise the offset
825          * is not actually in this window and a different window (which
826          * has that one hash excess) must be used.  This is to support
827          * the object header and delta base parsing routines below.
828          */
829         off_t win_off = win->offset;
830         return win_off <= offset
831                 && (offset + 20) <= (win_off + win->len);
832 }
833
834 unsigned char *use_pack(struct packed_git *p,
835                 struct pack_window **w_cursor,
836                 off_t offset,
837                 unsigned int *left)
838 {
839         struct pack_window *win = *w_cursor;
840
841         /* Since packfiles end in a hash of their content and it's
842          * pointless to ask for an offset into the middle of that
843          * hash, and the in_window function above wouldn't match
844          * don't allow an offset too close to the end of the file.
845          */
846         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
847                 die("packfile %s cannot be accessed", p->pack_name);
848         if (offset > (p->pack_size - 20))
849                 die("offset beyond end of packfile (truncated pack?)");
850
851         if (!win || !in_window(win, offset)) {
852                 if (win)
853                         win->inuse_cnt--;
854                 for (win = p->windows; win; win = win->next) {
855                         if (in_window(win, offset))
856                                 break;
857                 }
858                 if (!win) {
859                         size_t window_align = packed_git_window_size / 2;
860                         off_t len;
861
862                         if (p->pack_fd == -1 && open_packed_git(p))
863                                 die("packfile %s cannot be accessed", p->pack_name);
864
865                         win = xcalloc(1, sizeof(*win));
866                         win->offset = (offset / window_align) * window_align;
867                         len = p->pack_size - win->offset;
868                         if (len > packed_git_window_size)
869                                 len = packed_git_window_size;
870                         win->len = (size_t)len;
871                         pack_mapped += win->len;
872                         while (packed_git_limit < pack_mapped
873                                 && unuse_one_window(p, p->pack_fd))
874                                 ; /* nothing */
875                         win->base = xmmap(NULL, win->len,
876                                 PROT_READ, MAP_PRIVATE,
877                                 p->pack_fd, win->offset);
878                         if (win->base == MAP_FAILED)
879                                 die("packfile %s cannot be mapped: %s",
880                                         p->pack_name,
881                                         strerror(errno));
882                         if (!win->offset && win->len == p->pack_size
883                                 && !p->do_not_close) {
884                                 close(p->pack_fd);
885                                 pack_open_fds--;
886                                 p->pack_fd = -1;
887                         }
888                         pack_mmap_calls++;
889                         pack_open_windows++;
890                         if (pack_mapped > peak_pack_mapped)
891                                 peak_pack_mapped = pack_mapped;
892                         if (pack_open_windows > peak_pack_open_windows)
893                                 peak_pack_open_windows = pack_open_windows;
894                         win->next = p->windows;
895                         p->windows = win;
896                 }
897         }
898         if (win != *w_cursor) {
899                 win->last_used = pack_used_ctr++;
900                 win->inuse_cnt++;
901                 *w_cursor = win;
902         }
903         offset -= win->offset;
904         if (left)
905                 *left = win->len - xsize_t(offset);
906         return win->base + offset;
907 }
908
909 static struct packed_git *alloc_packed_git(int extra)
910 {
911         struct packed_git *p = xmalloc(sizeof(*p) + extra);
912         memset(p, 0, sizeof(*p));
913         p->pack_fd = -1;
914         return p;
915 }
916
917 static void try_to_free_pack_memory(size_t size)
918 {
919         release_pack_memory(size, -1);
920 }
921
922 struct packed_git *add_packed_git(const char *path, int path_len, int local)
923 {
924         static int have_set_try_to_free_routine;
925         struct stat st;
926         struct packed_git *p = alloc_packed_git(path_len + 2);
927
928         if (!have_set_try_to_free_routine) {
929                 have_set_try_to_free_routine = 1;
930                 set_try_to_free_routine(try_to_free_pack_memory);
931         }
932
933         /*
934          * Make sure a corresponding .pack file exists and that
935          * the index looks sane.
936          */
937         path_len -= strlen(".idx");
938         if (path_len < 1) {
939                 free(p);
940                 return NULL;
941         }
942         memcpy(p->pack_name, path, path_len);
943
944         strcpy(p->pack_name + path_len, ".keep");
945         if (!access(p->pack_name, F_OK))
946                 p->pack_keep = 1;
947
948         strcpy(p->pack_name + path_len, ".pack");
949         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
950                 free(p);
951                 return NULL;
952         }
953
954         /* ok, it looks sane as far as we can check without
955          * actually mapping the pack file.
956          */
957         p->pack_size = st.st_size;
958         p->pack_local = local;
959         p->mtime = st.st_mtime;
960         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
961                 hashclr(p->sha1);
962         return p;
963 }
964
965 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
966 {
967         const char *path = sha1_pack_name(sha1);
968         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
969
970         strcpy(p->pack_name, path);
971         hashcpy(p->sha1, sha1);
972         if (check_packed_git_idx(idx_path, p)) {
973                 free(p);
974                 return NULL;
975         }
976
977         return p;
978 }
979
980 void install_packed_git(struct packed_git *pack)
981 {
982         if (pack->pack_fd != -1)
983                 pack_open_fds++;
984
985         pack->next = packed_git;
986         packed_git = pack;
987 }
988
989 static void prepare_packed_git_one(char *objdir, int local)
990 {
991         /* Ensure that this buffer is large enough so that we can
992            append "/pack/" without clobbering the stack even if
993            strlen(objdir) were PATH_MAX.  */
994         char path[PATH_MAX + 1 + 4 + 1 + 1];
995         int len;
996         DIR *dir;
997         struct dirent *de;
998
999         sprintf(path, "%s/pack", objdir);
1000         len = strlen(path);
1001         dir = opendir(path);
1002         if (!dir) {
1003                 if (errno != ENOENT)
1004                         error("unable to open object pack directory: %s: %s",
1005                               path, strerror(errno));
1006                 return;
1007         }
1008         path[len++] = '/';
1009         while ((de = readdir(dir)) != NULL) {
1010                 int namelen = strlen(de->d_name);
1011                 struct packed_git *p;
1012
1013                 if (!has_extension(de->d_name, ".idx"))
1014                         continue;
1015
1016                 if (len + namelen + 1 > sizeof(path))
1017                         continue;
1018
1019                 /* Don't reopen a pack we already have. */
1020                 strcpy(path + len, de->d_name);
1021                 for (p = packed_git; p; p = p->next) {
1022                         if (!memcmp(path, p->pack_name, len + namelen - 4))
1023                                 break;
1024                 }
1025                 if (p)
1026                         continue;
1027                 /* See if it really is a valid .idx file with corresponding
1028                  * .pack file that we can map.
1029                  */
1030                 p = add_packed_git(path, len + namelen, local);
1031                 if (!p)
1032                         continue;
1033                 install_packed_git(p);
1034         }
1035         closedir(dir);
1036 }
1037
1038 static int sort_pack(const void *a_, const void *b_)
1039 {
1040         struct packed_git *a = *((struct packed_git **)a_);
1041         struct packed_git *b = *((struct packed_git **)b_);
1042         int st;
1043
1044         /*
1045          * Local packs tend to contain objects specific to our
1046          * variant of the project than remote ones.  In addition,
1047          * remote ones could be on a network mounted filesystem.
1048          * Favor local ones for these reasons.
1049          */
1050         st = a->pack_local - b->pack_local;
1051         if (st)
1052                 return -st;
1053
1054         /*
1055          * Younger packs tend to contain more recent objects,
1056          * and more recent objects tend to get accessed more
1057          * often.
1058          */
1059         if (a->mtime < b->mtime)
1060                 return 1;
1061         else if (a->mtime == b->mtime)
1062                 return 0;
1063         return -1;
1064 }
1065
1066 static void rearrange_packed_git(void)
1067 {
1068         struct packed_git **ary, *p;
1069         int i, n;
1070
1071         for (n = 0, p = packed_git; p; p = p->next)
1072                 n++;
1073         if (n < 2)
1074                 return;
1075
1076         /* prepare an array of packed_git for easier sorting */
1077         ary = xcalloc(n, sizeof(struct packed_git *));
1078         for (n = 0, p = packed_git; p; p = p->next)
1079                 ary[n++] = p;
1080
1081         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1082
1083         /* link them back again */
1084         for (i = 0; i < n - 1; i++)
1085                 ary[i]->next = ary[i + 1];
1086         ary[n - 1]->next = NULL;
1087         packed_git = ary[0];
1088
1089         free(ary);
1090 }
1091
1092 static int prepare_packed_git_run_once = 0;
1093 void prepare_packed_git(void)
1094 {
1095         struct alternate_object_database *alt;
1096
1097         if (prepare_packed_git_run_once)
1098                 return;
1099         prepare_packed_git_one(get_object_directory(), 1);
1100         prepare_alt_odb();
1101         for (alt = alt_odb_list; alt; alt = alt->next) {
1102                 alt->name[-1] = 0;
1103                 prepare_packed_git_one(alt->base, 0);
1104                 alt->name[-1] = '/';
1105         }
1106         rearrange_packed_git();
1107         prepare_packed_git_run_once = 1;
1108 }
1109
1110 void reprepare_packed_git(void)
1111 {
1112         discard_revindex();
1113         prepare_packed_git_run_once = 0;
1114         prepare_packed_git();
1115 }
1116
1117 static void mark_bad_packed_object(struct packed_git *p,
1118                                    const unsigned char *sha1)
1119 {
1120         unsigned i;
1121         for (i = 0; i < p->num_bad_objects; i++)
1122                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1123                         return;
1124         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1125         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1126         p->num_bad_objects++;
1127 }
1128
1129 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1130 {
1131         struct packed_git *p;
1132         unsigned i;
1133
1134         for (p = packed_git; p; p = p->next)
1135                 for (i = 0; i < p->num_bad_objects; i++)
1136                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1137                                 return p;
1138         return NULL;
1139 }
1140
1141 int check_sha1_signature(const unsigned char *sha1, void *map, unsigned long size, const char *type)
1142 {
1143         unsigned char real_sha1[20];
1144         hash_sha1_file(map, size, type, real_sha1);
1145         return hashcmp(sha1, real_sha1) ? -1 : 0;
1146 }
1147
1148 static int git_open_noatime(const char *name, struct packed_git *p)
1149 {
1150         static int sha1_file_open_flag = O_NOATIME;
1151
1152         for (;;) {
1153                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1154                 if (fd >= 0)
1155                         return fd;
1156
1157                 /* Might the failure be due to O_NOATIME? */
1158                 if (errno != ENOENT && sha1_file_open_flag) {
1159                         sha1_file_open_flag = 0;
1160                         continue;
1161                 }
1162
1163                 return -1;
1164         }
1165 }
1166
1167 static int open_sha1_file(const unsigned char *sha1)
1168 {
1169         int fd;
1170         char *name = sha1_file_name(sha1);
1171         struct alternate_object_database *alt;
1172
1173         fd = git_open_noatime(name, NULL);
1174         if (fd >= 0)
1175                 return fd;
1176
1177         prepare_alt_odb();
1178         errno = ENOENT;
1179         for (alt = alt_odb_list; alt; alt = alt->next) {
1180                 name = alt->name;
1181                 fill_sha1_path(name, sha1);
1182                 fd = git_open_noatime(alt->base, NULL);
1183                 if (fd >= 0)
1184                         return fd;
1185         }
1186         return -1;
1187 }
1188
1189 static void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1190 {
1191         void *map;
1192         int fd;
1193
1194         fd = open_sha1_file(sha1);
1195         map = NULL;
1196         if (fd >= 0) {
1197                 struct stat st;
1198
1199                 if (!fstat(fd, &st)) {
1200                         *size = xsize_t(st.st_size);
1201                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1202                 }
1203                 close(fd);
1204         }
1205         return map;
1206 }
1207
1208 static int legacy_loose_object(unsigned char *map)
1209 {
1210         unsigned int word;
1211
1212         /*
1213          * Is it a zlib-compressed buffer? If so, the first byte
1214          * must be 0x78 (15-bit window size, deflated), and the
1215          * first 16-bit word is evenly divisible by 31
1216          */
1217         word = (map[0] << 8) + map[1];
1218         if (map[0] == 0x78 && !(word % 31))
1219                 return 1;
1220         else
1221                 return 0;
1222 }
1223
1224 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1225                 unsigned long len, enum object_type *type, unsigned long *sizep)
1226 {
1227         unsigned shift;
1228         unsigned long size, c;
1229         unsigned long used = 0;
1230
1231         c = buf[used++];
1232         *type = (c >> 4) & 7;
1233         size = c & 15;
1234         shift = 4;
1235         while (c & 0x80) {
1236                 if (len <= used || bitsizeof(long) <= shift) {
1237                         error("bad object header");
1238                         return 0;
1239                 }
1240                 c = buf[used++];
1241                 size += (c & 0x7f) << shift;
1242                 shift += 7;
1243         }
1244         *sizep = size;
1245         return used;
1246 }
1247
1248 static int unpack_sha1_header(z_stream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1249 {
1250         unsigned long size, used;
1251         static const char valid_loose_object_type[8] = {
1252                 0, /* OBJ_EXT */
1253                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1254                 0, /* "delta" and others are invalid in a loose object */
1255         };
1256         enum object_type type;
1257
1258         /* Get the data stream */
1259         memset(stream, 0, sizeof(*stream));
1260         stream->next_in = map;
1261         stream->avail_in = mapsize;
1262         stream->next_out = buffer;
1263         stream->avail_out = bufsiz;
1264
1265         if (legacy_loose_object(map)) {
1266                 git_inflate_init(stream);
1267                 return git_inflate(stream, 0);
1268         }
1269
1270
1271         /*
1272          * There used to be a second loose object header format which
1273          * was meant to mimic the in-pack format, allowing for direct
1274          * copy of the object data.  This format turned up not to be
1275          * really worth it and we don't write it any longer.  But we
1276          * can still read it.
1277          */
1278         used = unpack_object_header_buffer(map, mapsize, &type, &size);
1279         if (!used || !valid_loose_object_type[type])
1280                 return -1;
1281         map += used;
1282         mapsize -= used;
1283
1284         /* Set up the stream for the rest.. */
1285         stream->next_in = map;
1286         stream->avail_in = mapsize;
1287         git_inflate_init(stream);
1288
1289         /* And generate the fake traditional header */
1290         stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1291                                          typename(type), size);
1292         return 0;
1293 }
1294
1295 static void *unpack_sha1_rest(z_stream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1296 {
1297         int bytes = strlen(buffer) + 1;
1298         unsigned char *buf = xmallocz(size);
1299         unsigned long n;
1300         int status = Z_OK;
1301
1302         n = stream->total_out - bytes;
1303         if (n > size)
1304                 n = size;
1305         memcpy(buf, (char *) buffer + bytes, n);
1306         bytes = n;
1307         if (bytes <= size) {
1308                 /*
1309                  * The above condition must be (bytes <= size), not
1310                  * (bytes < size).  In other words, even though we
1311                  * expect no more output and set avail_out to zer0,
1312                  * the input zlib stream may have bytes that express
1313                  * "this concludes the stream", and we *do* want to
1314                  * eat that input.
1315                  *
1316                  * Otherwise we would not be able to test that we
1317                  * consumed all the input to reach the expected size;
1318                  * we also want to check that zlib tells us that all
1319                  * went well with status == Z_STREAM_END at the end.
1320                  */
1321                 stream->next_out = buf + bytes;
1322                 stream->avail_out = size - bytes;
1323                 while (status == Z_OK)
1324                         status = git_inflate(stream, Z_FINISH);
1325         }
1326         if (status == Z_STREAM_END && !stream->avail_in) {
1327                 git_inflate_end(stream);
1328                 return buf;
1329         }
1330
1331         if (status < 0)
1332                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1333         else if (stream->avail_in)
1334                 error("garbage at end of loose object '%s'",
1335                       sha1_to_hex(sha1));
1336         free(buf);
1337         return NULL;
1338 }
1339
1340 /*
1341  * We used to just use "sscanf()", but that's actually way
1342  * too permissive for what we want to check. So do an anal
1343  * object header parse by hand.
1344  */
1345 static int parse_sha1_header(const char *hdr, unsigned long *sizep)
1346 {
1347         char type[10];
1348         int i;
1349         unsigned long size;
1350
1351         /*
1352          * The type can be at most ten bytes (including the
1353          * terminating '\0' that we add), and is followed by
1354          * a space.
1355          */
1356         i = 0;
1357         for (;;) {
1358                 char c = *hdr++;
1359                 if (c == ' ')
1360                         break;
1361                 type[i++] = c;
1362                 if (i >= sizeof(type))
1363                         return -1;
1364         }
1365         type[i] = 0;
1366
1367         /*
1368          * The length must follow immediately, and be in canonical
1369          * decimal format (ie "010" is not valid).
1370          */
1371         size = *hdr++ - '0';
1372         if (size > 9)
1373                 return -1;
1374         if (size) {
1375                 for (;;) {
1376                         unsigned long c = *hdr - '0';
1377                         if (c > 9)
1378                                 break;
1379                         hdr++;
1380                         size = size * 10 + c;
1381                 }
1382         }
1383         *sizep = size;
1384
1385         /*
1386          * The length must be followed by a zero byte
1387          */
1388         return *hdr ? -1 : type_from_string(type);
1389 }
1390
1391 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1392 {
1393         int ret;
1394         z_stream stream;
1395         char hdr[8192];
1396
1397         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1398         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1399                 return NULL;
1400
1401         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1402 }
1403
1404 unsigned long get_size_from_delta(struct packed_git *p,
1405                                   struct pack_window **w_curs,
1406                                   off_t curpos)
1407 {
1408         const unsigned char *data;
1409         unsigned char delta_head[20], *in;
1410         z_stream stream;
1411         int st;
1412
1413         memset(&stream, 0, sizeof(stream));
1414         stream.next_out = delta_head;
1415         stream.avail_out = sizeof(delta_head);
1416
1417         git_inflate_init(&stream);
1418         do {
1419                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1420                 stream.next_in = in;
1421                 st = git_inflate(&stream, Z_FINISH);
1422                 curpos += stream.next_in - in;
1423         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1424                  stream.total_out < sizeof(delta_head));
1425         git_inflate_end(&stream);
1426         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1427                 error("delta data unpack-initial failed");
1428                 return 0;
1429         }
1430
1431         /* Examine the initial part of the delta to figure out
1432          * the result size.
1433          */
1434         data = delta_head;
1435
1436         /* ignore base size */
1437         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1438
1439         /* Read the result size */
1440         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1441 }
1442
1443 static off_t get_delta_base(struct packed_git *p,
1444                                     struct pack_window **w_curs,
1445                                     off_t *curpos,
1446                                     enum object_type type,
1447                                     off_t delta_obj_offset)
1448 {
1449         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1450         off_t base_offset;
1451
1452         /* use_pack() assured us we have [base_info, base_info + 20)
1453          * as a range that we can look at without walking off the
1454          * end of the mapped window.  Its actually the hash size
1455          * that is assured.  An OFS_DELTA longer than the hash size
1456          * is stupid, as then a REF_DELTA would be smaller to store.
1457          */
1458         if (type == OBJ_OFS_DELTA) {
1459                 unsigned used = 0;
1460                 unsigned char c = base_info[used++];
1461                 base_offset = c & 127;
1462                 while (c & 128) {
1463                         base_offset += 1;
1464                         if (!base_offset || MSB(base_offset, 7))
1465                                 return 0;  /* overflow */
1466                         c = base_info[used++];
1467                         base_offset = (base_offset << 7) + (c & 127);
1468                 }
1469                 base_offset = delta_obj_offset - base_offset;
1470                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1471                         return 0;  /* out of bound */
1472                 *curpos += used;
1473         } else if (type == OBJ_REF_DELTA) {
1474                 /* The base entry _must_ be in the same pack */
1475                 base_offset = find_pack_entry_one(base_info, p);
1476                 *curpos += 20;
1477         } else
1478                 die("I am totally screwed");
1479         return base_offset;
1480 }
1481
1482 /* forward declaration for a mutually recursive function */
1483 static int packed_object_info(struct packed_git *p, off_t offset,
1484                               unsigned long *sizep);
1485
1486 static int packed_delta_info(struct packed_git *p,
1487                              struct pack_window **w_curs,
1488                              off_t curpos,
1489                              enum object_type type,
1490                              off_t obj_offset,
1491                              unsigned long *sizep)
1492 {
1493         off_t base_offset;
1494
1495         base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1496         if (!base_offset)
1497                 return OBJ_BAD;
1498         type = packed_object_info(p, base_offset, NULL);
1499         if (type <= OBJ_NONE) {
1500                 struct revindex_entry *revidx;
1501                 const unsigned char *base_sha1;
1502                 revidx = find_pack_revindex(p, base_offset);
1503                 if (!revidx)
1504                         return OBJ_BAD;
1505                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1506                 mark_bad_packed_object(p, base_sha1);
1507                 type = sha1_object_info(base_sha1, NULL);
1508                 if (type <= OBJ_NONE)
1509                         return OBJ_BAD;
1510         }
1511
1512         /* We choose to only get the type of the base object and
1513          * ignore potentially corrupt pack file that expects the delta
1514          * based on a base with a wrong size.  This saves tons of
1515          * inflate() calls.
1516          */
1517         if (sizep) {
1518                 *sizep = get_size_from_delta(p, w_curs, curpos);
1519                 if (*sizep == 0)
1520                         type = OBJ_BAD;
1521         }
1522
1523         return type;
1524 }
1525
1526 static int unpack_object_header(struct packed_git *p,
1527                                 struct pack_window **w_curs,
1528                                 off_t *curpos,
1529                                 unsigned long *sizep)
1530 {
1531         unsigned char *base;
1532         unsigned int left;
1533         unsigned long used;
1534         enum object_type type;
1535
1536         /* use_pack() assures us we have [base, base + 20) available
1537          * as a range that we can look at.  (Its actually the hash
1538          * size that is assured.)  With our object header encoding
1539          * the maximum deflated object size is 2^137, which is just
1540          * insane, so we know won't exceed what we have been given.
1541          */
1542         base = use_pack(p, w_curs, *curpos, &left);
1543         used = unpack_object_header_buffer(base, left, &type, sizep);
1544         if (!used) {
1545                 type = OBJ_BAD;
1546         } else
1547                 *curpos += used;
1548
1549         return type;
1550 }
1551
1552 const char *packed_object_info_detail(struct packed_git *p,
1553                                       off_t obj_offset,
1554                                       unsigned long *size,
1555                                       unsigned long *store_size,
1556                                       unsigned int *delta_chain_length,
1557                                       unsigned char *base_sha1)
1558 {
1559         struct pack_window *w_curs = NULL;
1560         off_t curpos;
1561         unsigned long dummy;
1562         unsigned char *next_sha1;
1563         enum object_type type;
1564         struct revindex_entry *revidx;
1565
1566         *delta_chain_length = 0;
1567         curpos = obj_offset;
1568         type = unpack_object_header(p, &w_curs, &curpos, size);
1569
1570         revidx = find_pack_revindex(p, obj_offset);
1571         *store_size = revidx[1].offset - obj_offset;
1572
1573         for (;;) {
1574                 switch (type) {
1575                 default:
1576                         die("pack %s contains unknown object type %d",
1577                             p->pack_name, type);
1578                 case OBJ_COMMIT:
1579                 case OBJ_TREE:
1580                 case OBJ_BLOB:
1581                 case OBJ_TAG:
1582                         unuse_pack(&w_curs);
1583                         return typename(type);
1584                 case OBJ_OFS_DELTA:
1585                         obj_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1586                         if (!obj_offset)
1587                                 die("pack %s contains bad delta base reference of type %s",
1588                                     p->pack_name, typename(type));
1589                         if (*delta_chain_length == 0) {
1590                                 revidx = find_pack_revindex(p, obj_offset);
1591                                 hashcpy(base_sha1, nth_packed_object_sha1(p, revidx->nr));
1592                         }
1593                         break;
1594                 case OBJ_REF_DELTA:
1595                         next_sha1 = use_pack(p, &w_curs, curpos, NULL);
1596                         if (*delta_chain_length == 0)
1597                                 hashcpy(base_sha1, next_sha1);
1598                         obj_offset = find_pack_entry_one(next_sha1, p);
1599                         break;
1600                 }
1601                 (*delta_chain_length)++;
1602                 curpos = obj_offset;
1603                 type = unpack_object_header(p, &w_curs, &curpos, &dummy);
1604         }
1605 }
1606
1607 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1608                               unsigned long *sizep)
1609 {
1610         struct pack_window *w_curs = NULL;
1611         unsigned long size;
1612         off_t curpos = obj_offset;
1613         enum object_type type;
1614
1615         type = unpack_object_header(p, &w_curs, &curpos, &size);
1616
1617         switch (type) {
1618         case OBJ_OFS_DELTA:
1619         case OBJ_REF_DELTA:
1620                 type = packed_delta_info(p, &w_curs, curpos,
1621                                          type, obj_offset, sizep);
1622                 break;
1623         case OBJ_COMMIT:
1624         case OBJ_TREE:
1625         case OBJ_BLOB:
1626         case OBJ_TAG:
1627                 if (sizep)
1628                         *sizep = size;
1629                 break;
1630         default:
1631                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1632                       type, (uintmax_t)obj_offset, p->pack_name);
1633                 type = OBJ_BAD;
1634         }
1635         unuse_pack(&w_curs);
1636         return type;
1637 }
1638
1639 static void *unpack_compressed_entry(struct packed_git *p,
1640                                     struct pack_window **w_curs,
1641                                     off_t curpos,
1642                                     unsigned long size)
1643 {
1644         int st;
1645         z_stream stream;
1646         unsigned char *buffer, *in;
1647
1648         buffer = xmallocz(size);
1649         memset(&stream, 0, sizeof(stream));
1650         stream.next_out = buffer;
1651         stream.avail_out = size + 1;
1652
1653         git_inflate_init(&stream);
1654         do {
1655                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1656                 stream.next_in = in;
1657                 st = git_inflate(&stream, Z_FINISH);
1658                 if (!stream.avail_out)
1659                         break; /* the payload is larger than it should be */
1660                 curpos += stream.next_in - in;
1661         } while (st == Z_OK || st == Z_BUF_ERROR);
1662         git_inflate_end(&stream);
1663         if ((st != Z_STREAM_END) || stream.total_out != size) {
1664                 free(buffer);
1665                 return NULL;
1666         }
1667
1668         return buffer;
1669 }
1670
1671 #define MAX_DELTA_CACHE (256)
1672
1673 static size_t delta_base_cached;
1674
1675 static struct delta_base_cache_lru_list {
1676         struct delta_base_cache_lru_list *prev;
1677         struct delta_base_cache_lru_list *next;
1678 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1679
1680 static struct delta_base_cache_entry {
1681         struct delta_base_cache_lru_list lru;
1682         void *data;
1683         struct packed_git *p;
1684         off_t base_offset;
1685         unsigned long size;
1686         enum object_type type;
1687 } delta_base_cache[MAX_DELTA_CACHE];
1688
1689 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1690 {
1691         unsigned long hash;
1692
1693         hash = (unsigned long)p + (unsigned long)base_offset;
1694         hash += (hash >> 8) + (hash >> 16);
1695         return hash % MAX_DELTA_CACHE;
1696 }
1697
1698 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1699         unsigned long *base_size, enum object_type *type, int keep_cache)
1700 {
1701         void *ret;
1702         unsigned long hash = pack_entry_hash(p, base_offset);
1703         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1704
1705         ret = ent->data;
1706         if (!ret || ent->p != p || ent->base_offset != base_offset)
1707                 return unpack_entry(p, base_offset, type, base_size);
1708
1709         if (!keep_cache) {
1710                 ent->data = NULL;
1711                 ent->lru.next->prev = ent->lru.prev;
1712                 ent->lru.prev->next = ent->lru.next;
1713                 delta_base_cached -= ent->size;
1714         } else {
1715                 ret = xmemdupz(ent->data, ent->size);
1716         }
1717         *type = ent->type;
1718         *base_size = ent->size;
1719         return ret;
1720 }
1721
1722 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1723 {
1724         if (ent->data) {
1725                 free(ent->data);
1726                 ent->data = NULL;
1727                 ent->lru.next->prev = ent->lru.prev;
1728                 ent->lru.prev->next = ent->lru.next;
1729                 delta_base_cached -= ent->size;
1730         }
1731 }
1732
1733 void clear_delta_base_cache(void)
1734 {
1735         unsigned long p;
1736         for (p = 0; p < MAX_DELTA_CACHE; p++)
1737                 release_delta_base_cache(&delta_base_cache[p]);
1738 }
1739
1740 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1741         void *base, unsigned long base_size, enum object_type type)
1742 {
1743         unsigned long hash = pack_entry_hash(p, base_offset);
1744         struct delta_base_cache_entry *ent = delta_base_cache + hash;
1745         struct delta_base_cache_lru_list *lru;
1746
1747         release_delta_base_cache(ent);
1748         delta_base_cached += base_size;
1749
1750         for (lru = delta_base_cache_lru.next;
1751              delta_base_cached > delta_base_cache_limit
1752              && lru != &delta_base_cache_lru;
1753              lru = lru->next) {
1754                 struct delta_base_cache_entry *f = (void *)lru;
1755                 if (f->type == OBJ_BLOB)
1756                         release_delta_base_cache(f);
1757         }
1758         for (lru = delta_base_cache_lru.next;
1759              delta_base_cached > delta_base_cache_limit
1760              && lru != &delta_base_cache_lru;
1761              lru = lru->next) {
1762                 struct delta_base_cache_entry *f = (void *)lru;
1763                 release_delta_base_cache(f);
1764         }
1765
1766         ent->p = p;
1767         ent->base_offset = base_offset;
1768         ent->type = type;
1769         ent->data = base;
1770         ent->size = base_size;
1771         ent->lru.next = &delta_base_cache_lru;
1772         ent->lru.prev = delta_base_cache_lru.prev;
1773         delta_base_cache_lru.prev->next = &ent->lru;
1774         delta_base_cache_lru.prev = &ent->lru;
1775 }
1776
1777 static void *read_object(const unsigned char *sha1, enum object_type *type,
1778                          unsigned long *size);
1779
1780 static void *unpack_delta_entry(struct packed_git *p,
1781                                 struct pack_window **w_curs,
1782                                 off_t curpos,
1783                                 unsigned long delta_size,
1784                                 off_t obj_offset,
1785                                 enum object_type *type,
1786                                 unsigned long *sizep)
1787 {
1788         void *delta_data, *result, *base;
1789         unsigned long base_size;
1790         off_t base_offset;
1791
1792         base_offset = get_delta_base(p, w_curs, &curpos, *type, obj_offset);
1793         if (!base_offset) {
1794                 error("failed to validate delta base reference "
1795                       "at offset %"PRIuMAX" from %s",
1796                       (uintmax_t)curpos, p->pack_name);
1797                 return NULL;
1798         }
1799         unuse_pack(w_curs);
1800         base = cache_or_unpack_entry(p, base_offset, &base_size, type, 0);
1801         if (!base) {
1802                 /*
1803                  * We're probably in deep shit, but let's try to fetch
1804                  * the required base anyway from another pack or loose.
1805                  * This is costly but should happen only in the presence
1806                  * of a corrupted pack, and is better than failing outright.
1807                  */
1808                 struct revindex_entry *revidx;
1809                 const unsigned char *base_sha1;
1810                 revidx = find_pack_revindex(p, base_offset);
1811                 if (!revidx)
1812                         return NULL;
1813                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
1814                 error("failed to read delta base object %s"
1815                       " at offset %"PRIuMAX" from %s",
1816                       sha1_to_hex(base_sha1), (uintmax_t)base_offset,
1817                       p->pack_name);
1818                 mark_bad_packed_object(p, base_sha1);
1819                 base = read_object(base_sha1, type, &base_size);
1820                 if (!base)
1821                         return NULL;
1822         }
1823
1824         delta_data = unpack_compressed_entry(p, w_curs, curpos, delta_size);
1825         if (!delta_data) {
1826                 error("failed to unpack compressed delta "
1827                       "at offset %"PRIuMAX" from %s",
1828                       (uintmax_t)curpos, p->pack_name);
1829                 free(base);
1830                 return NULL;
1831         }
1832         result = patch_delta(base, base_size,
1833                              delta_data, delta_size,
1834                              sizep);
1835         if (!result)
1836                 die("failed to apply delta");
1837         free(delta_data);
1838         add_delta_base_cache(p, base_offset, base, base_size, *type);
1839         return result;
1840 }
1841
1842 int do_check_packed_object_crc;
1843
1844 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1845                    enum object_type *type, unsigned long *sizep)
1846 {
1847         struct pack_window *w_curs = NULL;
1848         off_t curpos = obj_offset;
1849         void *data;
1850
1851         if (do_check_packed_object_crc && p->index_version > 1) {
1852                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1853                 unsigned long len = revidx[1].offset - obj_offset;
1854                 if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1855                         const unsigned char *sha1 =
1856                                 nth_packed_object_sha1(p, revidx->nr);
1857                         error("bad packed object CRC for %s",
1858                               sha1_to_hex(sha1));
1859                         mark_bad_packed_object(p, sha1);
1860                         unuse_pack(&w_curs);
1861                         return NULL;
1862                 }
1863         }
1864
1865         *type = unpack_object_header(p, &w_curs, &curpos, sizep);
1866         switch (*type) {
1867         case OBJ_OFS_DELTA:
1868         case OBJ_REF_DELTA:
1869                 data = unpack_delta_entry(p, &w_curs, curpos, *sizep,
1870                                           obj_offset, type, sizep);
1871                 break;
1872         case OBJ_COMMIT:
1873         case OBJ_TREE:
1874         case OBJ_BLOB:
1875         case OBJ_TAG:
1876                 data = unpack_compressed_entry(p, &w_curs, curpos, *sizep);
1877                 break;
1878         default:
1879                 data = NULL;
1880                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1881                       *type, (uintmax_t)obj_offset, p->pack_name);
1882         }
1883         unuse_pack(&w_curs);
1884         return data;
1885 }
1886
1887 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1888                                             uint32_t n)
1889 {
1890         const unsigned char *index = p->index_data;
1891         if (!index) {
1892                 if (open_pack_index(p))
1893                         return NULL;
1894                 index = p->index_data;
1895         }
1896         if (n >= p->num_objects)
1897                 return NULL;
1898         index += 4 * 256;
1899         if (p->index_version == 1) {
1900                 return index + 24 * n + 4;
1901         } else {
1902                 index += 8;
1903                 return index + 20 * n;
1904         }
1905 }
1906
1907 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1908 {
1909         const unsigned char *index = p->index_data;
1910         index += 4 * 256;
1911         if (p->index_version == 1) {
1912                 return ntohl(*((uint32_t *)(index + 24 * n)));
1913         } else {
1914                 uint32_t off;
1915                 index += 8 + p->num_objects * (20 + 4);
1916                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1917                 if (!(off & 0x80000000))
1918                         return off;
1919                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1920                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
1921                                    ntohl(*((uint32_t *)(index + 4)));
1922         }
1923 }
1924
1925 off_t find_pack_entry_one(const unsigned char *sha1,
1926                                   struct packed_git *p)
1927 {
1928         const uint32_t *level1_ofs = p->index_data;
1929         const unsigned char *index = p->index_data;
1930         unsigned hi, lo, stride;
1931         static int use_lookup = -1;
1932         static int debug_lookup = -1;
1933
1934         if (debug_lookup < 0)
1935                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
1936
1937         if (!index) {
1938                 if (open_pack_index(p))
1939                         return 0;
1940                 level1_ofs = p->index_data;
1941                 index = p->index_data;
1942         }
1943         if (p->index_version > 1) {
1944                 level1_ofs += 2;
1945                 index += 8;
1946         }
1947         index += 4 * 256;
1948         hi = ntohl(level1_ofs[*sha1]);
1949         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
1950         if (p->index_version > 1) {
1951                 stride = 20;
1952         } else {
1953                 stride = 24;
1954                 index += 4;
1955         }
1956
1957         if (debug_lookup)
1958                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
1959                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
1960
1961         if (use_lookup < 0)
1962                 use_lookup = !!getenv("GIT_USE_LOOKUP");
1963         if (use_lookup) {
1964                 int pos = sha1_entry_pos(index, stride, 0,
1965                                          lo, hi, p->num_objects, sha1);
1966                 if (pos < 0)
1967                         return 0;
1968                 return nth_packed_object_offset(p, pos);
1969         }
1970
1971         do {
1972                 unsigned mi = (lo + hi) / 2;
1973                 int cmp = hashcmp(index + mi * stride, sha1);
1974
1975                 if (debug_lookup)
1976                         printf("lo %u hi %u rg %u mi %u\n",
1977                                lo, hi, hi - lo, mi);
1978                 if (!cmp)
1979                         return nth_packed_object_offset(p, mi);
1980                 if (cmp > 0)
1981                         hi = mi;
1982                 else
1983                         lo = mi+1;
1984         } while (lo < hi);
1985         return 0;
1986 }
1987
1988 static int is_pack_valid(struct packed_git *p)
1989 {
1990         /* An already open pack is known to be valid. */
1991         if (p->pack_fd != -1)
1992                 return 1;
1993
1994         /* If the pack has one window completely covering the
1995          * file size, the pack is known to be valid even if
1996          * the descriptor is not currently open.
1997          */
1998         if (p->windows) {
1999                 struct pack_window *w = p->windows;
2000
2001                 if (!w->offset && w->len == p->pack_size)
2002                         return 1;
2003         }
2004
2005         /* Force the pack to open to prove its valid. */
2006         return !open_packed_git(p);
2007 }
2008
2009 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2010 {
2011         static struct packed_git *last_found = (void *)1;
2012         struct packed_git *p;
2013         off_t offset;
2014
2015         prepare_packed_git();
2016         if (!packed_git)
2017                 return 0;
2018         p = (last_found == (void *)1) ? packed_git : last_found;
2019
2020         do {
2021                 if (p->num_bad_objects) {
2022                         unsigned i;
2023                         for (i = 0; i < p->num_bad_objects; i++)
2024                                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2025                                         goto next;
2026                 }
2027
2028                 offset = find_pack_entry_one(sha1, p);
2029                 if (offset) {
2030                         /*
2031                          * We are about to tell the caller where they can
2032                          * locate the requested object.  We better make
2033                          * sure the packfile is still here and can be
2034                          * accessed before supplying that answer, as
2035                          * it may have been deleted since the index
2036                          * was loaded!
2037                          */
2038                         if (!is_pack_valid(p)) {
2039                                 error("packfile %s cannot be accessed", p->pack_name);
2040                                 goto next;
2041                         }
2042                         e->offset = offset;
2043                         e->p = p;
2044                         hashcpy(e->sha1, sha1);
2045                         last_found = p;
2046                         return 1;
2047                 }
2048
2049                 next:
2050                 if (p == last_found)
2051                         p = packed_git;
2052                 else
2053                         p = p->next;
2054                 if (p == last_found)
2055                         p = p->next;
2056         } while (p);
2057         return 0;
2058 }
2059
2060 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2061                                   struct packed_git *packs)
2062 {
2063         struct packed_git *p;
2064
2065         for (p = packs; p; p = p->next) {
2066                 if (find_pack_entry_one(sha1, p))
2067                         return p;
2068         }
2069         return NULL;
2070
2071 }
2072
2073 static int sha1_loose_object_info(const unsigned char *sha1, unsigned long *sizep)
2074 {
2075         int status;
2076         unsigned long mapsize, size;
2077         void *map;
2078         z_stream stream;
2079         char hdr[32];
2080
2081         map = map_sha1_file(sha1, &mapsize);
2082         if (!map)
2083                 return error("unable to find %s", sha1_to_hex(sha1));
2084         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2085                 status = error("unable to unpack %s header",
2086                                sha1_to_hex(sha1));
2087         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2088                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2089         else if (sizep)
2090                 *sizep = size;
2091         git_inflate_end(&stream);
2092         munmap(map, mapsize);
2093         return status;
2094 }
2095
2096 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2097 {
2098         struct cached_object *co;
2099         struct pack_entry e;
2100         int status;
2101
2102         co = find_cached_object(sha1);
2103         if (co) {
2104                 if (sizep)
2105                         *sizep = co->size;
2106                 return co->type;
2107         }
2108
2109         if (!find_pack_entry(sha1, &e)) {
2110                 /* Most likely it's a loose object. */
2111                 status = sha1_loose_object_info(sha1, sizep);
2112                 if (status >= 0)
2113                         return status;
2114
2115                 /* Not a loose object; someone else may have just packed it. */
2116                 reprepare_packed_git();
2117                 if (!find_pack_entry(sha1, &e))
2118                         return status;
2119         }
2120
2121         status = packed_object_info(e.p, e.offset, sizep);
2122         if (status < 0) {
2123                 mark_bad_packed_object(e.p, sha1);
2124                 status = sha1_object_info(sha1, sizep);
2125         }
2126
2127         return status;
2128 }
2129
2130 static void *read_packed_sha1(const unsigned char *sha1,
2131                               enum object_type *type, unsigned long *size)
2132 {
2133         struct pack_entry e;
2134         void *data;
2135
2136         if (!find_pack_entry(sha1, &e))
2137                 return NULL;
2138         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2139         if (!data) {
2140                 /*
2141                  * We're probably in deep shit, but let's try to fetch
2142                  * the required object anyway from another pack or loose.
2143                  * This should happen only in the presence of a corrupted
2144                  * pack, and is better than failing outright.
2145                  */
2146                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2147                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2148                 mark_bad_packed_object(e.p, sha1);
2149                 data = read_object(sha1, type, size);
2150         }
2151         return data;
2152 }
2153
2154 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2155                       unsigned char *sha1)
2156 {
2157         struct cached_object *co;
2158
2159         hash_sha1_file(buf, len, typename(type), sha1);
2160         if (has_sha1_file(sha1) || find_cached_object(sha1))
2161                 return 0;
2162         if (cached_object_alloc <= cached_object_nr) {
2163                 cached_object_alloc = alloc_nr(cached_object_alloc);
2164                 cached_objects = xrealloc(cached_objects,
2165                                           sizeof(*cached_objects) *
2166                                           cached_object_alloc);
2167         }
2168         co = &cached_objects[cached_object_nr++];
2169         co->size = len;
2170         co->type = type;
2171         co->buf = xmalloc(len);
2172         memcpy(co->buf, buf, len);
2173         hashcpy(co->sha1, sha1);
2174         return 0;
2175 }
2176
2177 static void *read_object(const unsigned char *sha1, enum object_type *type,
2178                          unsigned long *size)
2179 {
2180         unsigned long mapsize;
2181         void *map, *buf;
2182         struct cached_object *co;
2183
2184         co = find_cached_object(sha1);
2185         if (co) {
2186                 *type = co->type;
2187                 *size = co->size;
2188                 return xmemdupz(co->buf, co->size);
2189         }
2190
2191         buf = read_packed_sha1(sha1, type, size);
2192         if (buf)
2193                 return buf;
2194         map = map_sha1_file(sha1, &mapsize);
2195         if (map) {
2196                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2197                 munmap(map, mapsize);
2198                 return buf;
2199         }
2200         reprepare_packed_git();
2201         return read_packed_sha1(sha1, type, size);
2202 }
2203
2204 /*
2205  * This function dies on corrupt objects; the callers who want to
2206  * deal with them should arrange to call read_object() and give error
2207  * messages themselves.
2208  */
2209 void *read_sha1_file_repl(const unsigned char *sha1,
2210                           enum object_type *type,
2211                           unsigned long *size,
2212                           const unsigned char **replacement)
2213 {
2214         const unsigned char *repl = lookup_replace_object(sha1);
2215         void *data;
2216         char *path;
2217         const struct packed_git *p;
2218
2219         errno = 0;
2220         data = read_object(repl, type, size);
2221         if (data) {
2222                 if (replacement)
2223                         *replacement = repl;
2224                 return data;
2225         }
2226
2227         if (errno && errno != ENOENT)
2228                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2229
2230         /* die if we replaced an object with one that does not exist */
2231         if (repl != sha1)
2232                 die("replacement %s not found for %s",
2233                     sha1_to_hex(repl), sha1_to_hex(sha1));
2234
2235         if (has_loose_object(repl)) {
2236                 path = sha1_file_name(sha1);
2237                 die("loose object %s (stored in %s) is corrupt",
2238                     sha1_to_hex(repl), path);
2239         }
2240
2241         if ((p = has_packed_and_bad(repl)) != NULL)
2242                 die("packed object %s (stored in %s) is corrupt",
2243                     sha1_to_hex(repl), p->pack_name);
2244
2245         return NULL;
2246 }
2247
2248 void *read_object_with_reference(const unsigned char *sha1,
2249                                  const char *required_type_name,
2250                                  unsigned long *size,
2251                                  unsigned char *actual_sha1_return)
2252 {
2253         enum object_type type, required_type;
2254         void *buffer;
2255         unsigned long isize;
2256         unsigned char actual_sha1[20];
2257
2258         required_type = type_from_string(required_type_name);
2259         hashcpy(actual_sha1, sha1);
2260         while (1) {
2261                 int ref_length = -1;
2262                 const char *ref_type = NULL;
2263
2264                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2265                 if (!buffer)
2266                         return NULL;
2267                 if (type == required_type) {
2268                         *size = isize;
2269                         if (actual_sha1_return)
2270                                 hashcpy(actual_sha1_return, actual_sha1);
2271                         return buffer;
2272                 }
2273                 /* Handle references */
2274                 else if (type == OBJ_COMMIT)
2275                         ref_type = "tree ";
2276                 else if (type == OBJ_TAG)
2277                         ref_type = "object ";
2278                 else {
2279                         free(buffer);
2280                         return NULL;
2281                 }
2282                 ref_length = strlen(ref_type);
2283
2284                 if (ref_length + 40 > isize ||
2285                     memcmp(buffer, ref_type, ref_length) ||
2286                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2287                         free(buffer);
2288                         return NULL;
2289                 }
2290                 free(buffer);
2291                 /* Now we have the ID of the referred-to object in
2292                  * actual_sha1.  Check again. */
2293         }
2294 }
2295
2296 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2297                                     const char *type, unsigned char *sha1,
2298                                     char *hdr, int *hdrlen)
2299 {
2300         git_SHA_CTX c;
2301
2302         /* Generate the header */
2303         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2304
2305         /* Sha1.. */
2306         git_SHA1_Init(&c);
2307         git_SHA1_Update(&c, hdr, *hdrlen);
2308         git_SHA1_Update(&c, buf, len);
2309         git_SHA1_Final(sha1, &c);
2310 }
2311
2312 /*
2313  * Move the just written object into its final resting place.
2314  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2315  * "moving" is only a part of what it does, when no patch between
2316  * master to pu changes the call sites of this function.
2317  */
2318 int move_temp_to_file(const char *tmpfile, const char *filename)
2319 {
2320         int ret = 0;
2321
2322         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2323                 goto try_rename;
2324         else if (link(tmpfile, filename))
2325                 ret = errno;
2326
2327         /*
2328          * Coda hack - coda doesn't like cross-directory links,
2329          * so we fall back to a rename, which will mean that it
2330          * won't be able to check collisions, but that's not a
2331          * big deal.
2332          *
2333          * The same holds for FAT formatted media.
2334          *
2335          * When this succeeds, we just return.  We have nothing
2336          * left to unlink.
2337          */
2338         if (ret && ret != EEXIST) {
2339         try_rename:
2340                 if (!rename(tmpfile, filename))
2341                         goto out;
2342                 ret = errno;
2343         }
2344         unlink_or_warn(tmpfile);
2345         if (ret) {
2346                 if (ret != EEXIST) {
2347                         return error("unable to write sha1 filename %s: %s\n", filename, strerror(ret));
2348                 }
2349                 /* FIXME!!! Collision check here ? */
2350         }
2351
2352 out:
2353         if (adjust_shared_perm(filename))
2354                 return error("unable to set permission to '%s'", filename);
2355         return 0;
2356 }
2357
2358 static int write_buffer(int fd, const void *buf, size_t len)
2359 {
2360         if (write_in_full(fd, buf, len) < 0)
2361                 return error("file write error (%s)", strerror(errno));
2362         return 0;
2363 }
2364
2365 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2366                    unsigned char *sha1)
2367 {
2368         char hdr[32];
2369         int hdrlen;
2370         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2371         return 0;
2372 }
2373
2374 /* Finalize a file on disk, and close it. */
2375 static void close_sha1_file(int fd)
2376 {
2377         if (fsync_object_files)
2378                 fsync_or_die(fd, "sha1 file");
2379         if (close(fd) != 0)
2380                 die_errno("error when closing sha1 file");
2381 }
2382
2383 /* Size of directory component, including the ending '/' */
2384 static inline int directory_size(const char *filename)
2385 {
2386         const char *s = strrchr(filename, '/');
2387         if (!s)
2388                 return 0;
2389         return s - filename + 1;
2390 }
2391
2392 /*
2393  * This creates a temporary file in the same directory as the final
2394  * 'filename'
2395  *
2396  * We want to avoid cross-directory filename renames, because those
2397  * can have problems on various filesystems (FAT, NFS, Coda).
2398  */
2399 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2400 {
2401         int fd, dirlen = directory_size(filename);
2402
2403         if (dirlen + 20 > bufsiz) {
2404                 errno = ENAMETOOLONG;
2405                 return -1;
2406         }
2407         memcpy(buffer, filename, dirlen);
2408         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2409         fd = git_mkstemp_mode(buffer, 0444);
2410         if (fd < 0 && dirlen && errno == ENOENT) {
2411                 /* Make sure the directory exists */
2412                 memcpy(buffer, filename, dirlen);
2413                 buffer[dirlen-1] = 0;
2414                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2415                         return -1;
2416
2417                 /* Try again */
2418                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2419                 fd = git_mkstemp_mode(buffer, 0444);
2420         }
2421         return fd;
2422 }
2423
2424 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2425                               const void *buf, unsigned long len, time_t mtime)
2426 {
2427         int fd, ret;
2428         unsigned char compressed[4096];
2429         z_stream stream;
2430         git_SHA_CTX c;
2431         unsigned char parano_sha1[20];
2432         char *filename;
2433         static char tmpfile[PATH_MAX];
2434
2435         filename = sha1_file_name(sha1);
2436         fd = create_tmpfile(tmpfile, sizeof(tmpfile), filename);
2437         if (fd < 0) {
2438                 if (errno == EACCES)
2439                         return error("insufficient permission for adding an object to repository database %s\n", get_object_directory());
2440                 else
2441                         return error("unable to create temporary sha1 filename %s: %s\n", tmpfile, strerror(errno));
2442         }
2443
2444         /* Set it up */
2445         memset(&stream, 0, sizeof(stream));
2446         deflateInit(&stream, zlib_compression_level);
2447         stream.next_out = compressed;
2448         stream.avail_out = sizeof(compressed);
2449         git_SHA1_Init(&c);
2450
2451         /* First header.. */
2452         stream.next_in = (unsigned char *)hdr;
2453         stream.avail_in = hdrlen;
2454         while (deflate(&stream, 0) == Z_OK)
2455                 /* nothing */;
2456         git_SHA1_Update(&c, hdr, hdrlen);
2457
2458         /* Then the data itself.. */
2459         stream.next_in = (void *)buf;
2460         stream.avail_in = len;
2461         do {
2462                 unsigned char *in0 = stream.next_in;
2463                 ret = deflate(&stream, Z_FINISH);
2464                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2465                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2466                         die("unable to write sha1 file");
2467                 stream.next_out = compressed;
2468                 stream.avail_out = sizeof(compressed);
2469         } while (ret == Z_OK);
2470
2471         if (ret != Z_STREAM_END)
2472                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2473         ret = deflateEnd(&stream);
2474         if (ret != Z_OK)
2475                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2476         git_SHA1_Final(parano_sha1, &c);
2477         if (hashcmp(sha1, parano_sha1) != 0)
2478                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2479
2480         close_sha1_file(fd);
2481
2482         if (mtime) {
2483                 struct utimbuf utb;
2484                 utb.actime = mtime;
2485                 utb.modtime = mtime;
2486                 if (utime(tmpfile, &utb) < 0)
2487                         warning("failed utime() on %s: %s",
2488                                 tmpfile, strerror(errno));
2489         }
2490
2491         return move_temp_to_file(tmpfile, filename);
2492 }
2493
2494 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2495 {
2496         unsigned char sha1[20];
2497         char hdr[32];
2498         int hdrlen;
2499
2500         /* Normally if we have it in the pack then we do not bother writing
2501          * it out into .git/objects/??/?{38} file.
2502          */
2503         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2504         if (returnsha1)
2505                 hashcpy(returnsha1, sha1);
2506         if (has_sha1_file(sha1))
2507                 return 0;
2508         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2509 }
2510
2511 int force_object_loose(const unsigned char *sha1, time_t mtime)
2512 {
2513         void *buf;
2514         unsigned long len;
2515         enum object_type type;
2516         char hdr[32];
2517         int hdrlen;
2518         int ret;
2519
2520         if (has_loose_object(sha1))
2521                 return 0;
2522         buf = read_packed_sha1(sha1, &type, &len);
2523         if (!buf)
2524                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2525         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2526         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2527         free(buf);
2528
2529         return ret;
2530 }
2531
2532 int has_pack_index(const unsigned char *sha1)
2533 {
2534         struct stat st;
2535         if (stat(sha1_pack_index_name(sha1), &st))
2536                 return 0;
2537         return 1;
2538 }
2539
2540 int has_sha1_pack(const unsigned char *sha1)
2541 {
2542         struct pack_entry e;
2543         return find_pack_entry(sha1, &e);
2544 }
2545
2546 int has_sha1_file(const unsigned char *sha1)
2547 {
2548         struct pack_entry e;
2549
2550         if (find_pack_entry(sha1, &e))
2551                 return 1;
2552         return has_loose_object(sha1);
2553 }
2554
2555 static void check_tree(const void *buf, size_t size)
2556 {
2557         struct tree_desc desc;
2558         struct name_entry entry;
2559
2560         init_tree_desc(&desc, buf, size);
2561         while (tree_entry(&desc, &entry))
2562                 /* do nothing
2563                  * tree_entry() will die() on malformed entries */
2564                 ;
2565 }
2566
2567 static void check_commit(const void *buf, size_t size)
2568 {
2569         struct commit c;
2570         memset(&c, 0, sizeof(c));
2571         if (parse_commit_buffer(&c, buf, size))
2572                 die("corrupt commit");
2573 }
2574
2575 static void check_tag(const void *buf, size_t size)
2576 {
2577         struct tag t;
2578         memset(&t, 0, sizeof(t));
2579         if (parse_tag_buffer(&t, buf, size))
2580                 die("corrupt tag");
2581 }
2582
2583 static int index_mem(unsigned char *sha1, void *buf, size_t size,
2584                      int write_object, enum object_type type,
2585                      const char *path, int format_check)
2586 {
2587         int ret, re_allocated = 0;
2588
2589         if (!type)
2590                 type = OBJ_BLOB;
2591
2592         /*
2593          * Convert blobs to git internal format
2594          */
2595         if ((type == OBJ_BLOB) && path) {
2596                 struct strbuf nbuf = STRBUF_INIT;
2597                 if (convert_to_git(path, buf, size, &nbuf,
2598                                    write_object ? safe_crlf : 0)) {
2599                         buf = strbuf_detach(&nbuf, &size);
2600                         re_allocated = 1;
2601                 }
2602         }
2603         if (format_check) {
2604                 if (type == OBJ_TREE)
2605                         check_tree(buf, size);
2606                 if (type == OBJ_COMMIT)
2607                         check_commit(buf, size);
2608                 if (type == OBJ_TAG)
2609                         check_tag(buf, size);
2610         }
2611
2612         if (write_object)
2613                 ret = write_sha1_file(buf, size, typename(type), sha1);
2614         else
2615                 ret = hash_sha1_file(buf, size, typename(type), sha1);
2616         if (re_allocated)
2617                 free(buf);
2618         return ret;
2619 }
2620
2621 #define SMALL_FILE_SIZE (32*1024)
2622
2623 int index_fd(unsigned char *sha1, int fd, struct stat *st, int write_object,
2624              enum object_type type, const char *path, int format_check)
2625 {
2626         int ret;
2627         size_t size = xsize_t(st->st_size);
2628
2629         if (!S_ISREG(st->st_mode)) {
2630                 struct strbuf sbuf = STRBUF_INIT;
2631                 if (strbuf_read(&sbuf, fd, 4096) >= 0)
2632                         ret = index_mem(sha1, sbuf.buf, sbuf.len, write_object,
2633                                         type, path, format_check);
2634                 else
2635                         ret = -1;
2636                 strbuf_release(&sbuf);
2637         } else if (!size) {
2638                 ret = index_mem(sha1, NULL, size, write_object, type, path,
2639                                 format_check);
2640         } else if (size <= SMALL_FILE_SIZE) {
2641                 char *buf = xmalloc(size);
2642                 if (size == read_in_full(fd, buf, size))
2643                         ret = index_mem(sha1, buf, size, write_object, type,
2644                                         path, format_check);
2645                 else
2646                         ret = error("short read %s", strerror(errno));
2647                 free(buf);
2648         } else {
2649                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
2650                 ret = index_mem(sha1, buf, size, write_object, type, path,
2651                                 format_check);
2652                 munmap(buf, size);
2653         }
2654         close(fd);
2655         return ret;
2656 }
2657
2658 int index_path(unsigned char *sha1, const char *path, struct stat *st, int write_object)
2659 {
2660         int fd;
2661         struct strbuf sb = STRBUF_INIT;
2662
2663         switch (st->st_mode & S_IFMT) {
2664         case S_IFREG:
2665                 fd = open(path, O_RDONLY);
2666                 if (fd < 0)
2667                         return error("open(\"%s\"): %s", path,
2668                                      strerror(errno));
2669                 if (index_fd(sha1, fd, st, write_object, OBJ_BLOB, path, 0) < 0)
2670                         return error("%s: failed to insert into database",
2671                                      path);
2672                 break;
2673         case S_IFLNK:
2674                 if (strbuf_readlink(&sb, path, st->st_size)) {
2675                         char *errstr = strerror(errno);
2676                         return error("readlink(\"%s\"): %s", path,
2677                                      errstr);
2678                 }
2679                 if (!write_object)
2680                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
2681                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
2682                         return error("%s: failed to insert into database",
2683                                      path);
2684                 strbuf_release(&sb);
2685                 break;
2686         case S_IFDIR:
2687                 return resolve_gitlink_ref(path, "HEAD", sha1);
2688         default:
2689                 return error("%s: unsupported file type", path);
2690         }
2691         return 0;
2692 }
2693
2694 int read_pack_header(int fd, struct pack_header *header)
2695 {
2696         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
2697                 /* "eof before pack header was fully read" */
2698                 return PH_ERROR_EOF;
2699
2700         if (header->hdr_signature != htonl(PACK_SIGNATURE))
2701                 /* "protocol error (pack signature mismatch detected)" */
2702                 return PH_ERROR_PACK_SIGNATURE;
2703         if (!pack_version_ok(header->hdr_version))
2704                 /* "protocol error (pack version unsupported)" */
2705                 return PH_ERROR_PROTOCOL;
2706         return 0;
2707 }
2708
2709 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
2710 {
2711         enum object_type type = sha1_object_info(sha1, NULL);
2712         if (type < 0)
2713                 die("%s is not a valid object", sha1_to_hex(sha1));
2714         if (type != expect)
2715                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
2716                     typename(expect));
2717 }